Add a rewrite rule to collapse constant casts

If, e.g., we know from bounds analysis that the result of an operation fits in the range r[0~>0], we now just replace it with the literal constant. Fixes #493 After | File Name | Before || Change | % Change -------------------------------------------------------------------------------------------- 21m22.14s | Total | 21m22.79s || -0m00.65s | -0.05% -------------------------------------------------------------------------------------------- 4m09.97s | PushButtonSynthesis.vo | 4m10.56s || -0m00.59s | -0.23% 3m09.12s | p384_32.c | 3m08.91s || +0m00.21s | +0.11% 2m05.94s | Rewriter.vo | 2m06.30s || -0m00.35s | -0.28% 1m56.58s | RewriterWf2.vo | 1m56.09s || +0m00.48s | +0.42% 1m52.39s | RewriterRulesGood.vo | 1m52.04s || +0m00.35s | +0.31% 1m46.01s | RewriterRulesInterpGood.vo | 1m45.79s || +0m00.21s | +0.20% 0m46.44s | RewriterInterpProofs1.vo | 0m46.47s || -0m00.03s | -0.06% 0m44.96s | ExtractionHaskell/word_by_word_montgomery | 0m45.59s || -0m00.63s | -1.38% 0m39.18s | p521_32.c | 0m39.33s || -0m00.14s | -0.38% 0m32.41s | p521_64.c | 0m32.54s || -0m00.13s | -0.39% 0m30.87s | ExtractionHaskell/unsaturated_solinas | 0m30.67s || +0m00.19s | +0.65% 0m24.32s | ExtractionHaskell/saturated_solinas | 0m24.44s || -0m00.12s | -0.49% 0m23.59s | RewriterWf1.vo | 0m24.10s || -0m00.51s | -2.11% 0m17.01s | ExtractionOCaml/word_by_word_montgomery | 0m17.14s || -0m00.12s | -0.75% 0m13.48s | secp256k1_32.c | 0m13.30s || +0m00.17s | +1.35% 0m13.11s | p256_32.c | 0m13.37s || -0m00.25s | -1.94% 0m11.34s | p484_64.c | 0m11.34s || +0m00.00s | +0.00% 0m10.78s | ExtractionOCaml/unsaturated_solinas | 0m10.79s || -0m00.00s | -0.09% 0m10.27s | ExtractionOCaml/word_by_word_montgomery.ml | 0m10.06s || +0m00.20s | +2.08% 0m08.11s | ExtractionOCaml/saturated_solinas | 0m07.92s || +0m00.18s | +2.39% 0m06.92s | ExtractionOCaml/unsaturated_solinas.ml | 0m07.02s || -0m00.09s | -1.42% 0m06.18s | ExtractionHaskell/word_by_word_montgomery.hs | 0m06.28s || -0m00.10s | -1.59% 0m06.13s | BoundsPipeline.vo | 0m05.98s || +0m00.14s | +2.50% 0m05.90s | p224_32.c | 0m05.92s || -0m00.01s | -0.33% 0m05.29s | p384_64.c | 0m05.33s || -0m00.04s | -0.75% 0m05.17s | ExtractionOCaml/saturated_solinas.ml | 0m05.20s || -0m00.03s | -0.57% 0m04.91s | ExtractionHaskell/unsaturated_solinas.hs | 0m04.93s || -0m00.01s | -0.40% 0m04.06s | ExtractionHaskell/saturated_solinas.hs | 0m04.00s || +0m00.05s | +1.49% 0m02.21s | curve25519_32.c | 0m02.22s || -0m00.01s | -0.45% 0m01.52s | curve25519_64.c | 0m01.50s || +0m00.02s | +1.33% 0m01.38s | CLI.vo | 0m01.42s || -0m00.04s | -2.81% 0m01.14s | RewriterProofs.vo | 0m01.13s || +0m00.01s | +0.88% 0m01.14s | StandaloneOCamlMain.vo | 0m00.96s || +0m00.17s | +18.74% 0m01.12s | StandaloneHaskellMain.vo | 0m01.03s || +0m00.09s | +8.73% 0m01.12s | secp256k1_64.c | 0m01.00s || +0m00.12s | +12.00% 0m01.05s | p256_64.c | 0m00.98s || +0m00.07s | +7.14% 0m01.03s | p224_64.c | 0m01.15s || -0m00.11s | -10.43%
author: Jason Gross <jgross@mit.edu> 2019-01-15 15:53:34 -0500
committer: Jason Gross <jasongross9@gmail.com> 2019-01-16 16:51:11 -0500
commit: 60bade02ccd577550bfcd5974d3c62a3d40e751a (patch)
tree: 46a9e9ef505704bdff47f0c794c163ecd5d7fd76
parent: 538e541709d70caaa78104739cb965c6523d89a8 (diff)
16 files changed, 5974 insertions, 6397 deletions
diff --git a/curve25519_32.c b/curve25519_32.c
index 5366ec40c..508141786 100644
--- a/curve25519_32.c
+++ b/curve25519_32.c
@@ -748,43 +748,42 @@ static void fiat_25519_to_bytes(uint8_t out1[32], const uint32_t arg1[10]) {
   uint8_t x81 = (uint8_t)(x78 & UINT8_C(0xff));
   uint8_t x82 = (uint8_t)(x80 >> 8);
   uint8_t x83 = (uint8_t)(x80 & UINT8_C(0xff));
-  fiat_25519_uint1 x84 = (fiat_25519_uint1)(x82 >> 8);
-  uint8_t x85 = (uint8_t)(x82 & UINT8_C(0xff));
-  uint32_t x86 = (x84 + x32);
-  uint32_t x87 = (x86 >> 8);
-  uint8_t x88 = (uint8_t)(x86 & UINT8_C(0xff));
-  uint32_t x89 = (x87 >> 8);
-  uint8_t x90 = (uint8_t)(x87 & UINT8_C(0xff));
-  fiat_25519_uint1 x91 = (fiat_25519_uint1)(x89 >> 8);
-  uint8_t x92 = (uint8_t)(x89 & UINT8_C(0xff));
-  uint32_t x93 = (x91 + x45);
-  uint32_t x94 = (x93 >> 8);
-  uint8_t x95 = (uint8_t)(x93 & UINT8_C(0xff));
-  uint32_t x96 = (x94 >> 8);
-  uint8_t x97 = (uint8_t)(x94 & UINT8_C(0xff));
-  uint8_t x98 = (uint8_t)(x96 >> 8);
-  uint8_t x99 = (uint8_t)(x96 & UINT8_C(0xff));
-  uint32_t x100 = (x98 + x44);
-  uint32_t x101 = (x100 >> 8);
-  uint8_t x102 = (uint8_t)(x100 & UINT8_C(0xff));
-  uint32_t x103 = (x101 >> 8);
-  uint8_t x104 = (uint8_t)(x101 & UINT8_C(0xff));
-  uint8_t x105 = (uint8_t)(x103 >> 8);
-  uint8_t x106 = (uint8_t)(x103 & UINT8_C(0xff));
-  uint32_t x107 = (x105 + x43);
-  uint32_t x108 = (x107 >> 8);
-  uint8_t x109 = (uint8_t)(x107 & UINT8_C(0xff));
-  uint32_t x110 = (x108 >> 8);
-  uint8_t x111 = (uint8_t)(x108 & UINT8_C(0xff));
-  uint8_t x112 = (uint8_t)(x110 >> 8);
-  uint8_t x113 = (uint8_t)(x110 & UINT8_C(0xff));
-  uint32_t x114 = (x112 + x42);
-  uint32_t x115 = (x114 >> 8);
-  uint8_t x116 = (uint8_t)(x114 & UINT8_C(0xff));
-  uint32_t x117 = (x115 >> 8);
-  uint8_t x118 = (uint8_t)(x115 & UINT8_C(0xff));
-  uint8_t x119 = (uint8_t)(x117 >> 8);
-  uint8_t x120 = (uint8_t)(x117 & UINT8_C(0xff));
+  uint8_t x84 = (uint8_t)(x82 & UINT8_C(0xff));
+  uint32_t x85 = (0x0 + x32);
+  uint32_t x86 = (x85 >> 8);
+  uint8_t x87 = (uint8_t)(x85 & UINT8_C(0xff));
+  uint32_t x88 = (x86 >> 8);
+  uint8_t x89 = (uint8_t)(x86 & UINT8_C(0xff));
+  fiat_25519_uint1 x90 = (fiat_25519_uint1)(x88 >> 8);
+  uint8_t x91 = (uint8_t)(x88 & UINT8_C(0xff));
+  uint32_t x92 = (x90 + x45);
+  uint32_t x93 = (x92 >> 8);
+  uint8_t x94 = (uint8_t)(x92 & UINT8_C(0xff));
+  uint32_t x95 = (x93 >> 8);
+  uint8_t x96 = (uint8_t)(x93 & UINT8_C(0xff));
+  uint8_t x97 = (uint8_t)(x95 >> 8);
+  uint8_t x98 = (uint8_t)(x95 & UINT8_C(0xff));
+  uint32_t x99 = (x97 + x44);
+  uint32_t x100 = (x99 >> 8);
+  uint8_t x101 = (uint8_t)(x99 & UINT8_C(0xff));
+  uint32_t x102 = (x100 >> 8);
+  uint8_t x103 = (uint8_t)(x100 & UINT8_C(0xff));
+  uint8_t x104 = (uint8_t)(x102 >> 8);
+  uint8_t x105 = (uint8_t)(x102 & UINT8_C(0xff));
+  uint32_t x106 = (x104 + x43);
+  uint32_t x107 = (x106 >> 8);
+  uint8_t x108 = (uint8_t)(x106 & UINT8_C(0xff));
+  uint32_t x109 = (x107 >> 8);
+  uint8_t x110 = (uint8_t)(x107 & UINT8_C(0xff));
+  uint8_t x111 = (uint8_t)(x109 >> 8);
+  uint8_t x112 = (uint8_t)(x109 & UINT8_C(0xff));
+  uint32_t x113 = (x111 + x42);
+  uint32_t x114 = (x113 >> 8);
+  uint8_t x115 = (uint8_t)(x113 & UINT8_C(0xff));
+  uint32_t x116 = (x114 >> 8);
+  uint8_t x117 = (uint8_t)(x114 & UINT8_C(0xff));
+  uint8_t x118 = (uint8_t)(x116 >> 8);
+  uint8_t x119 = (uint8_t)(x116 & UINT8_C(0xff));
   out1[0] = x51;
   out1[1] = x53;
   out1[2] = x55;
@@ -800,23 +799,23 @@ static void fiat_25519_to_bytes(uint8_t out1[32], const uint32_t arg1[10]) {
   out1[12] = x79;
   out1[13] = x81;
   out1[14] = x83;
-  out1[15] = x85;
-  out1[16] = x88;
-  out1[17] = x90;
-  out1[18] = x92;
-  out1[19] = x95;
-  out1[20] = x97;
-  out1[21] = x99;
-  out1[22] = x102;
-  out1[23] = x104;
-  out1[24] = x106;
-  out1[25] = x109;
-  out1[26] = x111;
-  out1[27] = x113;
-  out1[28] = x116;
-  out1[29] = x118;
-  out1[30] = x120;
-  out1[31] = x119;
+  out1[15] = x84;
+  out1[16] = x87;
+  out1[17] = x89;
+  out1[18] = x91;
+  out1[19] = x94;
+  out1[20] = x96;
+  out1[21] = x98;
+  out1[22] = x101;
+  out1[23] = x103;
+  out1[24] = x105;
+  out1[25] = x108;
+  out1[26] = x110;
+  out1[27] = x112;
+  out1[28] = x115;
+  out1[29] = x117;
+  out1[30] = x119;
+  out1[31] = x118;
 }
 
 /*
@@ -880,30 +879,29 @@ static void fiat_25519_from_bytes(uint32_t out1[10], const uint8_t arg1[32]) {
   uint8_t x52 = (uint8_t)(x51 >> 25);
   uint32_t x53 = (x51 & UINT32_C(0x1ffffff));
   uint32_t x54 = (x52 + x41);
-  fiat_25519_uint1 x55 = (fiat_25519_uint1)(x54 >> 26);
-  uint32_t x56 = (x54 & UINT32_C(0x3ffffff));
-  uint32_t x57 = (x55 + x40);
-  uint8_t x58 = (uint8_t)(x57 >> 25);
-  uint32_t x59 = (x57 & UINT32_C(0x1ffffff));
-  uint32_t x60 = (x58 + x39);
-  uint8_t x61 = (uint8_t)(x60 >> 26);
-  uint32_t x62 = (x60 & UINT32_C(0x3ffffff));
-  uint32_t x63 = (x61 + x38);
-  uint8_t x64 = (uint8_t)(x63 >> 25);
-  uint32_t x65 = (x63 & UINT32_C(0x1ffffff));
-  uint32_t x66 = (x64 + x37);
-  uint8_t x67 = (uint8_t)(x66 >> 26);
-  uint32_t x68 = (x66 & UINT32_C(0x3ffffff));
-  uint32_t x69 = (x67 + x36);
+  uint32_t x55 = (x54 & UINT32_C(0x3ffffff));
+  uint32_t x56 = (0x0 + x40);
+  uint8_t x57 = (uint8_t)(x56 >> 25);
+  uint32_t x58 = (x56 & UINT32_C(0x1ffffff));
+  uint32_t x59 = (x57 + x39);
+  uint8_t x60 = (uint8_t)(x59 >> 26);
+  uint32_t x61 = (x59 & UINT32_C(0x3ffffff));
+  uint32_t x62 = (x60 + x38);
+  uint8_t x63 = (uint8_t)(x62 >> 25);
+  uint32_t x64 = (x62 & UINT32_C(0x1ffffff));
+  uint32_t x65 = (x63 + x37);
+  uint8_t x66 = (uint8_t)(x65 >> 26);
+  uint32_t x67 = (x65 & UINT32_C(0x3ffffff));
+  uint32_t x68 = (x66 + x36);
   out1[0] = x35;
   out1[1] = x47;
   out1[2] = x50;
   out1[3] = x53;
-  out1[4] = x56;
-  out1[5] = x59;
-  out1[6] = x62;
-  out1[7] = x65;
-  out1[8] = x68;
-  out1[9] = x69;
+  out1[4] = x55;
+  out1[5] = x58;
+  out1[6] = x61;
+  out1[7] = x64;
+  out1[8] = x67;
+  out1[9] = x68;
 }
 
diff --git a/p224_32.c b/p224_32.c
index 65d26dff1..eda2b7e5f 100644
--- a/p224_32.c
+++ b/p224_32.c
@@ -2100,475 +2100,472 @@ static void fiat_p224_from_montgomery(uint32_t out1[7], const uint32_t arg1[7])
   fiat_p224_addcarryx_u32(&x38, &x39, x37, x26, x29);
   uint32_t x40;
   fiat_p224_uint1 x41;
-  fiat_p224_addcarryx_u32(&x40, &x41, x23, 0x0, 0x0);
+  fiat_p224_addcarryx_u32(&x40, &x41, 0x0, x32, x12);
   uint32_t x42;
   fiat_p224_uint1 x43;
-  fiat_p224_addcarryx_u32(&x42, &x43, 0x0, x24, x22);
+  fiat_p224_addcarryx_u32(&x42, &x43, x41, x34, x14);
   uint32_t x44;
   fiat_p224_uint1 x45;
-  fiat_p224_addcarryx_u32(&x44, &x45, x43, 0x0, (fiat_p224_uint1)x40);
+  fiat_p224_addcarryx_u32(&x44, &x45, x43, x36, x16);
   uint32_t x46;
   fiat_p224_uint1 x47;
-  fiat_p224_addcarryx_u32(&x46, &x47, x45, 0x0, x10);
+  fiat_p224_addcarryx_u32(&x46, &x47, x17, 0x0, x5);
   uint32_t x48;
   fiat_p224_uint1 x49;
-  fiat_p224_addcarryx_u32(&x48, &x49, x47, x32, x12);
+  fiat_p224_addcarryx_u32(&x48, &x49, x45, x38, x46);
   uint32_t x50;
   fiat_p224_uint1 x51;
-  fiat_p224_addcarryx_u32(&x50, &x51, x49, x34, x14);
+  fiat_p224_addcarryx_u32(&x50, &x51, x39, 0x0, x27);
   uint32_t x52;
   fiat_p224_uint1 x53;
-  fiat_p224_addcarryx_u32(&x52, &x53, x51, x36, x16);
+  fiat_p224_addcarryx_u32(&x52, &x53, x49, x50, 0x0);
   uint32_t x54;
   fiat_p224_uint1 x55;
-  fiat_p224_addcarryx_u32(&x54, &x55, x17, 0x0, x5);
+  fiat_p224_addcarryx_u32(&x54, &x55, x23, 0x0, 0x0);
   uint32_t x56;
   fiat_p224_uint1 x57;
-  fiat_p224_addcarryx_u32(&x56, &x57, x53, x38, x54);
+  fiat_p224_addcarryx_u32(&x56, &x57, 0x0, x24, x22);
   uint32_t x58;
   fiat_p224_uint1 x59;
-  fiat_p224_addcarryx_u32(&x58, &x59, x39, 0x0, x27);
+  fiat_p224_addcarryx_u32(&x58, &x59, x57, 0x0, (fiat_p224_uint1)x54);
   uint32_t x60;
   fiat_p224_uint1 x61;
-  fiat_p224_addcarryx_u32(&x60, &x61, x57, x58, 0x0);
+  fiat_p224_addcarryx_u32(&x60, &x61, 0x0, (arg1[2]), x58);
   uint32_t x62;
   fiat_p224_uint1 x63;
-  fiat_p224_addcarryx_u32(&x62, &x63, 0x0, (arg1[2]), x44);
+  fiat_p224_addcarryx_u32(&x62, &x63, x61, 0x0, x10);
   uint32_t x64;
   fiat_p224_uint1 x65;
-  fiat_p224_addcarryx_u32(&x64, &x65, x63, 0x0, x46);
+  fiat_p224_addcarryx_u32(&x64, &x65, x63, 0x0, x40);
   uint32_t x66;
   fiat_p224_uint1 x67;
-  fiat_p224_addcarryx_u32(&x66, &x67, x65, 0x0, x48);
+  fiat_p224_addcarryx_u32(&x66, &x67, x65, 0x0, x42);
   uint32_t x68;
   fiat_p224_uint1 x69;
-  fiat_p224_addcarryx_u32(&x68, &x69, x67, 0x0, x50);
+  fiat_p224_addcarryx_u32(&x68, &x69, x67, 0x0, x44);
   uint32_t x70;
   fiat_p224_uint1 x71;
-  fiat_p224_addcarryx_u32(&x70, &x71, x69, 0x0, x52);
+  fiat_p224_addcarryx_u32(&x70, &x71, x69, 0x0, x48);
   uint32_t x72;
   fiat_p224_uint1 x73;
-  fiat_p224_addcarryx_u32(&x72, &x73, x71, 0x0, x56);
+  fiat_p224_addcarryx_u32(&x72, &x73, x71, 0x0, x52);
   uint32_t x74;
-  fiat_p224_uint1 x75;
-  fiat_p224_addcarryx_u32(&x74, &x75, x73, 0x0, x60);
+  uint32_t x75;
+  fiat_p224_mulx_u32(&x74, &x75, x60, UINT32_C(0xffffffff));
   uint32_t x76;
-  fiat_p224_uint1 x77;
-  fiat_p224_addcarryx_u32(&x76, &x77, x61, 0x0, 0x0);
+  uint32_t x77;
+  fiat_p224_mulx_u32(&x76, &x77, x74, UINT32_C(0xffffffff));
   uint32_t x78;
-  fiat_p224_uint1 x79;
-  fiat_p224_addcarryx_u32(&x78, &x79, x75, 0x0, (fiat_p224_uint1)x76);
+  uint32_t x79;
+  fiat_p224_mulx_u32(&x78, &x79, x74, UINT32_C(0xffffffff));
   uint32_t x80;
   uint32_t x81;
-  fiat_p224_mulx_u32(&x80, &x81, x62, UINT32_C(0xffffffff));
+  fiat_p224_mulx_u32(&x80, &x81, x74, UINT32_C(0xffffffff));
   uint32_t x82;
   uint32_t x83;
-  fiat_p224_mulx_u32(&x82, &x83, x80, UINT32_C(0xffffffff));
+  fiat_p224_mulx_u32(&x82, &x83, x74, UINT32_C(0xffffffff));
   uint32_t x84;
-  uint32_t x85;
-  fiat_p224_mulx_u32(&x84, &x85, x80, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x85;
+  fiat_p224_addcarryx_u32(&x84, &x85, 0x0, x80, x83);
   uint32_t x86;
-  uint32_t x87;
-  fiat_p224_mulx_u32(&x86, &x87, x80, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x87;
+  fiat_p224_addcarryx_u32(&x86, &x87, x85, x78, x81);
   uint32_t x88;
-  uint32_t x89;
-  fiat_p224_mulx_u32(&x88, &x89, x80, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x89;
+  fiat_p224_addcarryx_u32(&x88, &x89, x87, x76, x79);
   uint32_t x90;
   fiat_p224_uint1 x91;
-  fiat_p224_addcarryx_u32(&x90, &x91, 0x0, x86, x89);
+  fiat_p224_addcarryx_u32(&x90, &x91, 0x0, x74, x60);
   uint32_t x92;
   fiat_p224_uint1 x93;
-  fiat_p224_addcarryx_u32(&x92, &x93, x91, x84, x87);
+  fiat_p224_addcarryx_u32(&x92, &x93, x91, 0x0, x62);
   uint32_t x94;
   fiat_p224_uint1 x95;
-  fiat_p224_addcarryx_u32(&x94, &x95, x93, x82, x85);
+  fiat_p224_addcarryx_u32(&x94, &x95, x93, 0x0, x64);
   uint32_t x96;
   fiat_p224_uint1 x97;
-  fiat_p224_addcarryx_u32(&x96, &x97, 0x0, x80, x62);
+  fiat_p224_addcarryx_u32(&x96, &x97, x95, x82, x66);
   uint32_t x98;
   fiat_p224_uint1 x99;
-  fiat_p224_addcarryx_u32(&x98, &x99, x97, 0x0, x64);
+  fiat_p224_addcarryx_u32(&x98, &x99, x97, x84, x68);
   uint32_t x100;
   fiat_p224_uint1 x101;
-  fiat_p224_addcarryx_u32(&x100, &x101, x99, 0x0, x66);
+  fiat_p224_addcarryx_u32(&x100, &x101, x99, x86, x70);
   uint32_t x102;
   fiat_p224_uint1 x103;
-  fiat_p224_addcarryx_u32(&x102, &x103, x101, x88, x68);
+  fiat_p224_addcarryx_u32(&x102, &x103, x101, x88, x72);
   uint32_t x104;
   fiat_p224_uint1 x105;
-  fiat_p224_addcarryx_u32(&x104, &x105, x103, x90, x70);
+  fiat_p224_addcarryx_u32(&x104, &x105, x53, 0x0, 0x0);
   uint32_t x106;
   fiat_p224_uint1 x107;
-  fiat_p224_addcarryx_u32(&x106, &x107, x105, x92, x72);
+  fiat_p224_addcarryx_u32(&x106, &x107, x73, 0x0, (fiat_p224_uint1)x104);
   uint32_t x108;
   fiat_p224_uint1 x109;
-  fiat_p224_addcarryx_u32(&x108, &x109, x107, x94, x74);
+  fiat_p224_addcarryx_u32(&x108, &x109, x89, 0x0, x77);
   uint32_t x110;
   fiat_p224_uint1 x111;
-  fiat_p224_addcarryx_u32(&x110, &x111, x95, 0x0, x83);
+  fiat_p224_addcarryx_u32(&x110, &x111, x103, x108, x106);
   uint32_t x112;
   fiat_p224_uint1 x113;
-  fiat_p224_addcarryx_u32(&x112, &x113, x109, x110, x78);
+  fiat_p224_addcarryx_u32(&x112, &x113, 0x0, (arg1[3]), x92);
   uint32_t x114;
   fiat_p224_uint1 x115;
-  fiat_p224_addcarryx_u32(&x114, &x115, 0x0, (arg1[3]), x98);
+  fiat_p224_addcarryx_u32(&x114, &x115, x113, 0x0, x94);
   uint32_t x116;
   fiat_p224_uint1 x117;
-  fiat_p224_addcarryx_u32(&x116, &x117, x115, 0x0, x100);
+  fiat_p224_addcarryx_u32(&x116, &x117, x115, 0x0, x96);
   uint32_t x118;
   fiat_p224_uint1 x119;
-  fiat_p224_addcarryx_u32(&x118, &x119, x117, 0x0, x102);
+  fiat_p224_addcarryx_u32(&x118, &x119, x117, 0x0, x98);
   uint32_t x120;
   fiat_p224_uint1 x121;
-  fiat_p224_addcarryx_u32(&x120, &x121, x119, 0x0, x104);
+  fiat_p224_addcarryx_u32(&x120, &x121, x119, 0x0, x100);
   uint32_t x122;
   fiat_p224_uint1 x123;
-  fiat_p224_addcarryx_u32(&x122, &x123, x121, 0x0, x106);
+  fiat_p224_addcarryx_u32(&x122, &x123, x121, 0x0, x102);
   uint32_t x124;
   fiat_p224_uint1 x125;
-  fiat_p224_addcarryx_u32(&x124, &x125, x123, 0x0, x108);
+  fiat_p224_addcarryx_u32(&x124, &x125, x123, 0x0, x110);
   uint32_t x126;
-  fiat_p224_uint1 x127;
-  fiat_p224_addcarryx_u32(&x126, &x127, x125, 0x0, x112);
+  uint32_t x127;
+  fiat_p224_mulx_u32(&x126, &x127, x112, UINT32_C(0xffffffff));
   uint32_t x128;
-  fiat_p224_uint1 x129;
-  fiat_p224_addcarryx_u32(&x128, &x129, x113, 0x0, x79);
+  uint32_t x129;
+  fiat_p224_mulx_u32(&x128, &x129, x126, UINT32_C(0xffffffff));
   uint32_t x130;
-  fiat_p224_uint1 x131;
-  fiat_p224_addcarryx_u32(&x130, &x131, x127, 0x0, (fiat_p224_uint1)x128);
+  uint32_t x131;
+  fiat_p224_mulx_u32(&x130, &x131, x126, UINT32_C(0xffffffff));
   uint32_t x132;
   uint32_t x133;
-  fiat_p224_mulx_u32(&x132, &x133, x114, UINT32_C(0xffffffff));
+  fiat_p224_mulx_u32(&x132, &x133, x126, UINT32_C(0xffffffff));
   uint32_t x134;
   uint32_t x135;
-  fiat_p224_mulx_u32(&x134, &x135, x132, UINT32_C(0xffffffff));
+  fiat_p224_mulx_u32(&x134, &x135, x126, UINT32_C(0xffffffff));
   uint32_t x136;
-  uint32_t x137;
-  fiat_p224_mulx_u32(&x136, &x137, x132, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x137;
+  fiat_p224_addcarryx_u32(&x136, &x137, 0x0, x132, x135);
   uint32_t x138;
-  uint32_t x139;
-  fiat_p224_mulx_u32(&x138, &x139, x132, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x139;
+  fiat_p224_addcarryx_u32(&x138, &x139, x137, x130, x133);
   uint32_t x140;
-  uint32_t x141;
-  fiat_p224_mulx_u32(&x140, &x141, x132, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x141;
+  fiat_p224_addcarryx_u32(&x140, &x141, x139, x128, x131);
   uint32_t x142;
   fiat_p224_uint1 x143;
-  fiat_p224_addcarryx_u32(&x142, &x143, 0x0, x138, x141);
+  fiat_p224_addcarryx_u32(&x142, &x143, 0x0, x126, x112);
   uint32_t x144;
   fiat_p224_uint1 x145;
-  fiat_p224_addcarryx_u32(&x144, &x145, x143, x136, x139);
+  fiat_p224_addcarryx_u32(&x144, &x145, x143, 0x0, x114);
   uint32_t x146;
   fiat_p224_uint1 x147;
-  fiat_p224_addcarryx_u32(&x146, &x147, x145, x134, x137);
+  fiat_p224_addcarryx_u32(&x146, &x147, x145, 0x0, x116);
   uint32_t x148;
   fiat_p224_uint1 x149;
-  fiat_p224_addcarryx_u32(&x148, &x149, 0x0, x132, x114);
+  fiat_p224_addcarryx_u32(&x148, &x149, x147, x134, x118);
   uint32_t x150;
   fiat_p224_uint1 x151;
-  fiat_p224_addcarryx_u32(&x150, &x151, x149, 0x0, x116);
+  fiat_p224_addcarryx_u32(&x150, &x151, x149, x136, x120);
   uint32_t x152;
   fiat_p224_uint1 x153;
-  fiat_p224_addcarryx_u32(&x152, &x153, x151, 0x0, x118);
+  fiat_p224_addcarryx_u32(&x152, &x153, x151, x138, x122);
   uint32_t x154;
   fiat_p224_uint1 x155;
-  fiat_p224_addcarryx_u32(&x154, &x155, x153, x140, x120);
+  fiat_p224_addcarryx_u32(&x154, &x155, x153, x140, x124);
   uint32_t x156;
   fiat_p224_uint1 x157;
-  fiat_p224_addcarryx_u32(&x156, &x157, x155, x142, x122);
+  fiat_p224_addcarryx_u32(&x156, &x157, x111, 0x0, 0x0);
   uint32_t x158;
   fiat_p224_uint1 x159;
-  fiat_p224_addcarryx_u32(&x158, &x159, x157, x144, x124);
+  fiat_p224_addcarryx_u32(&x158, &x159, x125, 0x0, (fiat_p224_uint1)x156);
   uint32_t x160;
   fiat_p224_uint1 x161;
-  fiat_p224_addcarryx_u32(&x160, &x161, x159, x146, x126);
+  fiat_p224_addcarryx_u32(&x160, &x161, x141, 0x0, x129);
   uint32_t x162;
   fiat_p224_uint1 x163;
-  fiat_p224_addcarryx_u32(&x162, &x163, x147, 0x0, x135);
+  fiat_p224_addcarryx_u32(&x162, &x163, x155, x160, x158);
   uint32_t x164;
   fiat_p224_uint1 x165;
-  fiat_p224_addcarryx_u32(&x164, &x165, x161, x162, x130);
+  fiat_p224_addcarryx_u32(&x164, &x165, 0x0, (arg1[4]), x144);
   uint32_t x166;
   fiat_p224_uint1 x167;
-  fiat_p224_addcarryx_u32(&x166, &x167, 0x0, (arg1[4]), x150);
+  fiat_p224_addcarryx_u32(&x166, &x167, x165, 0x0, x146);
   uint32_t x168;
   fiat_p224_uint1 x169;
-  fiat_p224_addcarryx_u32(&x168, &x169, x167, 0x0, x152);
+  fiat_p224_addcarryx_u32(&x168, &x169, x167, 0x0, x148);
   uint32_t x170;
   fiat_p224_uint1 x171;
-  fiat_p224_addcarryx_u32(&x170, &x171, x169, 0x0, x154);
+  fiat_p224_addcarryx_u32(&x170, &x171, x169, 0x0, x150);
   uint32_t x172;
   fiat_p224_uint1 x173;
-  fiat_p224_addcarryx_u32(&x172, &x173, x171, 0x0, x156);
+  fiat_p224_addcarryx_u32(&x172, &x173, x171, 0x0, x152);
   uint32_t x174;
   fiat_p224_uint1 x175;
-  fiat_p224_addcarryx_u32(&x174, &x175, x173, 0x0, x158);
+  fiat_p224_addcarryx_u32(&x174, &x175, x173, 0x0, x154);
   uint32_t x176;
   fiat_p224_uint1 x177;
-  fiat_p224_addcarryx_u32(&x176, &x177, x175, 0x0, x160);
+  fiat_p224_addcarryx_u32(&x176, &x177, x175, 0x0, x162);
   uint32_t x178;
-  fiat_p224_uint1 x179;
-  fiat_p224_addcarryx_u32(&x178, &x179, x177, 0x0, x164);
+  uint32_t x179;
+  fiat_p224_mulx_u32(&x178, &x179, x164, UINT32_C(0xffffffff));
   uint32_t x180;
-  fiat_p224_uint1 x181;
-  fiat_p224_addcarryx_u32(&x180, &x181, x165, 0x0, x131);
+  uint32_t x181;
+  fiat_p224_mulx_u32(&x180, &x181, x178, UINT32_C(0xffffffff));
   uint32_t x182;
-  fiat_p224_uint1 x183;
-  fiat_p224_addcarryx_u32(&x182, &x183, x179, 0x0, (fiat_p224_uint1)x180);
+  uint32_t x183;
+  fiat_p224_mulx_u32(&x182, &x183, x178, UINT32_C(0xffffffff));
   uint32_t x184;
   uint32_t x185;
-  fiat_p224_mulx_u32(&x184, &x185, x166, UINT32_C(0xffffffff));
+  fiat_p224_mulx_u32(&x184, &x185, x178, UINT32_C(0xffffffff));
   uint32_t x186;
   uint32_t x187;
-  fiat_p224_mulx_u32(&x186, &x187, x184, UINT32_C(0xffffffff));
+  fiat_p224_mulx_u32(&x186, &x187, x178, UINT32_C(0xffffffff));
   uint32_t x188;
-  uint32_t x189;
-  fiat_p224_mulx_u32(&x188, &x189, x184, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x189;
+  fiat_p224_addcarryx_u32(&x188, &x189, 0x0, x184, x187);
   uint32_t x190;
-  uint32_t x191;
-  fiat_p224_mulx_u32(&x190, &x191, x184, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x191;
+  fiat_p224_addcarryx_u32(&x190, &x191, x189, x182, x185);
   uint32_t x192;
-  uint32_t x193;
-  fiat_p224_mulx_u32(&x192, &x193, x184, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x193;
+  fiat_p224_addcarryx_u32(&x192, &x193, x191, x180, x183);
   uint32_t x194;
   fiat_p224_uint1 x195;
-  fiat_p224_addcarryx_u32(&x194, &x195, 0x0, x190, x193);
+  fiat_p224_addcarryx_u32(&x194, &x195, 0x0, x178, x164);
   uint32_t x196;
   fiat_p224_uint1 x197;
-  fiat_p224_addcarryx_u32(&x196, &x197, x195, x188, x191);
+  fiat_p224_addcarryx_u32(&x196, &x197, x195, 0x0, x166);
   uint32_t x198;
   fiat_p224_uint1 x199;
-  fiat_p224_addcarryx_u32(&x198, &x199, x197, x186, x189);
+  fiat_p224_addcarryx_u32(&x198, &x199, x197, 0x0, x168);
   uint32_t x200;
   fiat_p224_uint1 x201;
-  fiat_p224_addcarryx_u32(&x200, &x201, 0x0, x184, x166);
+  fiat_p224_addcarryx_u32(&x200, &x201, x199, x186, x170);
   uint32_t x202;
   fiat_p224_uint1 x203;
-  fiat_p224_addcarryx_u32(&x202, &x203, x201, 0x0, x168);
+  fiat_p224_addcarryx_u32(&x202, &x203, x201, x188, x172);
   uint32_t x204;
   fiat_p224_uint1 x205;
-  fiat_p224_addcarryx_u32(&x204, &x205, x203, 0x0, x170);
+  fiat_p224_addcarryx_u32(&x204, &x205, x203, x190, x174);
   uint32_t x206;
   fiat_p224_uint1 x207;
-  fiat_p224_addcarryx_u32(&x206, &x207, x205, x192, x172);
+  fiat_p224_addcarryx_u32(&x206, &x207, x205, x192, x176);
   uint32_t x208;
   fiat_p224_uint1 x209;
-  fiat_p224_addcarryx_u32(&x208, &x209, x207, x194, x174);
+  fiat_p224_addcarryx_u32(&x208, &x209, x163, 0x0, 0x0);
   uint32_t x210;
   fiat_p224_uint1 x211;
-  fiat_p224_addcarryx_u32(&x210, &x211, x209, x196, x176);
+  fiat_p224_addcarryx_u32(&x210, &x211, x177, 0x0, (fiat_p224_uint1)x208);
   uint32_t x212;
   fiat_p224_uint1 x213;
-  fiat_p224_addcarryx_u32(&x212, &x213, x211, x198, x178);
+  fiat_p224_addcarryx_u32(&x212, &x213, x193, 0x0, x181);
   uint32_t x214;
   fiat_p224_uint1 x215;
-  fiat_p224_addcarryx_u32(&x214, &x215, x199, 0x0, x187);
+  fiat_p224_addcarryx_u32(&x214, &x215, x207, x212, x210);
   uint32_t x216;
   fiat_p224_uint1 x217;
-  fiat_p224_addcarryx_u32(&x216, &x217, x213, x214, x182);
+  fiat_p224_addcarryx_u32(&x216, &x217, 0x0, (arg1[5]), x196);
   uint32_t x218;
   fiat_p224_uint1 x219;
-  fiat_p224_addcarryx_u32(&x218, &x219, 0x0, (arg1[5]), x202);
+  fiat_p224_addcarryx_u32(&x218, &x219, x217, 0x0, x198);
   uint32_t x220;
   fiat_p224_uint1 x221;
-  fiat_p224_addcarryx_u32(&x220, &x221, x219, 0x0, x204);
+  fiat_p224_addcarryx_u32(&x220, &x221, x219, 0x0, x200);
   uint32_t x222;
   fiat_p224_uint1 x223;
-  fiat_p224_addcarryx_u32(&x222, &x223, x221, 0x0, x206);
+  fiat_p224_addcarryx_u32(&x222, &x223, x221, 0x0, x202);
   uint32_t x224;
   fiat_p224_uint1 x225;
-  fiat_p224_addcarryx_u32(&x224, &x225, x223, 0x0, x208);
+  fiat_p224_addcarryx_u32(&x224, &x225, x223, 0x0, x204);
   uint32_t x226;
   fiat_p224_uint1 x227;
-  fiat_p224_addcarryx_u32(&x226, &x227, x225, 0x0, x210);
+  fiat_p224_addcarryx_u32(&x226, &x227, x225, 0x0, x206);
   uint32_t x228;
   fiat_p224_uint1 x229;
-  fiat_p224_addcarryx_u32(&x228, &x229, x227, 0x0, x212);
+  fiat_p224_addcarryx_u32(&x228, &x229, x227, 0x0, x214);
   uint32_t x230;
-  fiat_p224_uint1 x231;
-  fiat_p224_addcarryx_u32(&x230, &x231, x229, 0x0, x216);
+  uint32_t x231;
+  fiat_p224_mulx_u32(&x230, &x231, x216, UINT32_C(0xffffffff));
   uint32_t x232;
-  fiat_p224_uint1 x233;
-  fiat_p224_addcarryx_u32(&x232, &x233, x217, 0x0, x183);
+  uint32_t x233;
+  fiat_p224_mulx_u32(&x232, &x233, x230, UINT32_C(0xffffffff));
   uint32_t x234;
-  fiat_p224_uint1 x235;
-  fiat_p224_addcarryx_u32(&x234, &x235, x231, 0x0, (fiat_p224_uint1)x232);
+  uint32_t x235;
+  fiat_p224_mulx_u32(&x234, &x235, x230, UINT32_C(0xffffffff));
   uint32_t x236;
   uint32_t x237;
-  fiat_p224_mulx_u32(&x236, &x237, x218, UINT32_C(0xffffffff));
+  fiat_p224_mulx_u32(&x236, &x237, x230, UINT32_C(0xffffffff));
   uint32_t x238;
   uint32_t x239;
-  fiat_p224_mulx_u32(&x238, &x239, x236, UINT32_C(0xffffffff));
+  fiat_p224_mulx_u32(&x238, &x239, x230, UINT32_C(0xffffffff));
   uint32_t x240;
-  uint32_t x241;
-  fiat_p224_mulx_u32(&x240, &x241, x236, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x241;
+  fiat_p224_addcarryx_u32(&x240, &x241, 0x0, x236, x239);
   uint32_t x242;
-  uint32_t x243;
-  fiat_p224_mulx_u32(&x242, &x243, x236, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x243;
+  fiat_p224_addcarryx_u32(&x242, &x243, x241, x234, x237);
   uint32_t x244;
-  uint32_t x245;
-  fiat_p224_mulx_u32(&x244, &x245, x236, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x245;
+  fiat_p224_addcarryx_u32(&x244, &x245, x243, x232, x235);
   uint32_t x246;
   fiat_p224_uint1 x247;
-  fiat_p224_addcarryx_u32(&x246, &x247, 0x0, x242, x245);
+  fiat_p224_addcarryx_u32(&x246, &x247, 0x0, x230, x216);
   uint32_t x248;
   fiat_p224_uint1 x249;
-  fiat_p224_addcarryx_u32(&x248, &x249, x247, x240, x243);
+  fiat_p224_addcarryx_u32(&x248, &x249, x247, 0x0, x218);
   uint32_t x250;
   fiat_p224_uint1 x251;
-  fiat_p224_addcarryx_u32(&x250, &x251, x249, x238, x241);
+  fiat_p224_addcarryx_u32(&x250, &x251, x249, 0x0, x220);
   uint32_t x252;
   fiat_p224_uint1 x253;
-  fiat_p224_addcarryx_u32(&x252, &x253, 0x0, x236, x218);
+  fiat_p224_addcarryx_u32(&x252, &x253, x251, x238, x222);
   uint32_t x254;
   fiat_p224_uint1 x255;
-  fiat_p224_addcarryx_u32(&x254, &x255, x253, 0x0, x220);
+  fiat_p224_addcarryx_u32(&x254, &x255, x253, x240, x224);
   uint32_t x256;
   fiat_p224_uint1 x257;
-  fiat_p224_addcarryx_u32(&x256, &x257, x255, 0x0, x222);
+  fiat_p224_addcarryx_u32(&x256, &x257, x255, x242, x226);
   uint32_t x258;
   fiat_p224_uint1 x259;
-  fiat_p224_addcarryx_u32(&x258, &x259, x257, x244, x224);
+  fiat_p224_addcarryx_u32(&x258, &x259, x257, x244, x228);
   uint32_t x260;
   fiat_p224_uint1 x261;
-  fiat_p224_addcarryx_u32(&x260, &x261, x259, x246, x226);
+  fiat_p224_addcarryx_u32(&x260, &x261, x215, 0x0, 0x0);
   uint32_t x262;
   fiat_p224_uint1 x263;
-  fiat_p224_addcarryx_u32(&x262, &x263, x261, x248, x228);
+  fiat_p224_addcarryx_u32(&x262, &x263, x229, 0x0, (fiat_p224_uint1)x260);
   uint32_t x264;
   fiat_p224_uint1 x265;
-  fiat_p224_addcarryx_u32(&x264, &x265, x263, x250, x230);
+  fiat_p224_addcarryx_u32(&x264, &x265, x245, 0x0, x233);
   uint32_t x266;
   fiat_p224_uint1 x267;
-  fiat_p224_addcarryx_u32(&x266, &x267, x251, 0x0, x239);
+  fiat_p224_addcarryx_u32(&x266, &x267, x259, x264, x262);
   uint32_t x268;
   fiat_p224_uint1 x269;
-  fiat_p224_addcarryx_u32(&x268, &x269, x265, x266, x234);
+  fiat_p224_addcarryx_u32(&x268, &x269, 0x0, (arg1[6]), x248);
   uint32_t x270;
   fiat_p224_uint1 x271;
-  fiat_p224_addcarryx_u32(&x270, &x271, 0x0, (arg1[6]), x254);
+  fiat_p224_addcarryx_u32(&x270, &x271, x269, 0x0, x250);
   uint32_t x272;
   fiat_p224_uint1 x273;
-  fiat_p224_addcarryx_u32(&x272, &x273, x271, 0x0, x256);
+  fiat_p224_addcarryx_u32(&x272, &x273, x271, 0x0, x252);
   uint32_t x274;
   fiat_p224_uint1 x275;
-  fiat_p224_addcarryx_u32(&x274, &x275, x273, 0x0, x258);
+  fiat_p224_addcarryx_u32(&x274, &x275, x273, 0x0, x254);
   uint32_t x276;
   fiat_p224_uint1 x277;
-  fiat_p224_addcarryx_u32(&x276, &x277, x275, 0x0, x260);
+  fiat_p224_addcarryx_u32(&x276, &x277, x275, 0x0, x256);
   uint32_t x278;
   fiat_p224_uint1 x279;
-  fiat_p224_addcarryx_u32(&x278, &x279, x277, 0x0, x262);
+  fiat_p224_addcarryx_u32(&x278, &x279, x277, 0x0, x258);
   uint32_t x280;
   fiat_p224_uint1 x281;
-  fiat_p224_addcarryx_u32(&x280, &x281, x279, 0x0, x264);
+  fiat_p224_addcarryx_u32(&x280, &x281, x279, 0x0, x266);
   uint32_t x282;
-  fiat_p224_uint1 x283;
-  fiat_p224_addcarryx_u32(&x282, &x283, x281, 0x0, x268);
+  uint32_t x283;
+  fiat_p224_mulx_u32(&x282, &x283, x268, UINT32_C(0xffffffff));
   uint32_t x284;
-  fiat_p224_uint1 x285;
-  fiat_p224_addcarryx_u32(&x284, &x285, x269, 0x0, x235);
+  uint32_t x285;
+  fiat_p224_mulx_u32(&x284, &x285, x282, UINT32_C(0xffffffff));
   uint32_t x286;
-  fiat_p224_uint1 x287;
-  fiat_p224_addcarryx_u32(&x286, &x287, x283, 0x0, (fiat_p224_uint1)x284);
+  uint32_t x287;
+  fiat_p224_mulx_u32(&x286, &x287, x282, UINT32_C(0xffffffff));
   uint32_t x288;
   uint32_t x289;
-  fiat_p224_mulx_u32(&x288, &x289, x270, UINT32_C(0xffffffff));
+  fiat_p224_mulx_u32(&x288, &x289, x282, UINT32_C(0xffffffff));
   uint32_t x290;
   uint32_t x291;
-  fiat_p224_mulx_u32(&x290, &x291, x288, UINT32_C(0xffffffff));
+  fiat_p224_mulx_u32(&x290, &x291, x282, UINT32_C(0xffffffff));
   uint32_t x292;
-  uint32_t x293;
-  fiat_p224_mulx_u32(&x292, &x293, x288, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x293;
+  fiat_p224_addcarryx_u32(&x292, &x293, 0x0, x288, x291);
   uint32_t x294;
-  uint32_t x295;
-  fiat_p224_mulx_u32(&x294, &x295, x288, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x295;
+  fiat_p224_addcarryx_u32(&x294, &x295, x293, x286, x289);
   uint32_t x296;
-  uint32_t x297;
-  fiat_p224_mulx_u32(&x296, &x297, x288, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x297;
+  fiat_p224_addcarryx_u32(&x296, &x297, x295, x284, x287);
   uint32_t x298;
   fiat_p224_uint1 x299;
-  fiat_p224_addcarryx_u32(&x298, &x299, 0x0, x294, x297);
+  fiat_p224_addcarryx_u32(&x298, &x299, 0x0, x282, x268);
   uint32_t x300;
   fiat_p224_uint1 x301;
-  fiat_p224_addcarryx_u32(&x300, &x301, x299, x292, x295);
+  fiat_p224_addcarryx_u32(&x300, &x301, x299, 0x0, x270);
   uint32_t x302;
   fiat_p224_uint1 x303;
-  fiat_p224_addcarryx_u32(&x302, &x303, x301, x290, x293);
+  fiat_p224_addcarryx_u32(&x302, &x303, x301, 0x0, x272);
   uint32_t x304;
   fiat_p224_uint1 x305;
-  fiat_p224_addcarryx_u32(&x304, &x305, 0x0, x288, x270);
+  fiat_p224_addcarryx_u32(&x304, &x305, x303, x290, x274);
   uint32_t x306;
   fiat_p224_uint1 x307;
-  fiat_p224_addcarryx_u32(&x306, &x307, x305, 0x0, x272);
+  fiat_p224_addcarryx_u32(&x306, &x307, x305, x292, x276);
   uint32_t x308;
   fiat_p224_uint1 x309;
-  fiat_p224_addcarryx_u32(&x308, &x309, x307, 0x0, x274);
+  fiat_p224_addcarryx_u32(&x308, &x309, x307, x294, x278);
   uint32_t x310;
   fiat_p224_uint1 x311;
-  fiat_p224_addcarryx_u32(&x310, &x311, x309, x296, x276);
+  fiat_p224_addcarryx_u32(&x310, &x311, x309, x296, x280);
   uint32_t x312;
   fiat_p224_uint1 x313;
-  fiat_p224_addcarryx_u32(&x312, &x313, x311, x298, x278);
+  fiat_p224_addcarryx_u32(&x312, &x313, x267, 0x0, 0x0);
   uint32_t x314;
   fiat_p224_uint1 x315;
-  fiat_p224_addcarryx_u32(&x314, &x315, x313, x300, x280);
+  fiat_p224_addcarryx_u32(&x314, &x315, x281, 0x0, (fiat_p224_uint1)x312);
   uint32_t x316;
   fiat_p224_uint1 x317;
-  fiat_p224_addcarryx_u32(&x316, &x317, x315, x302, x282);
+  fiat_p224_addcarryx_u32(&x316, &x317, x297, 0x0, x285);
   uint32_t x318;
   fiat_p224_uint1 x319;
-  fiat_p224_addcarryx_u32(&x318, &x319, x303, 0x0, x291);
+  fiat_p224_addcarryx_u32(&x318, &x319, x311, x316, x314);
   uint32_t x320;
   fiat_p224_uint1 x321;
-  fiat_p224_addcarryx_u32(&x320, &x321, x317, x318, x286);
+  fiat_p224_subborrowx_u32(&x320, &x321, 0x0, x300, 0x1);
   uint32_t x322;
   fiat_p224_uint1 x323;
-  fiat_p224_subborrowx_u32(&x322, &x323, 0x0, x306, 0x1);
+  fiat_p224_subborrowx_u32(&x322, &x323, x321, x302, 0x0);
   uint32_t x324;
   fiat_p224_uint1 x325;
-  fiat_p224_subborrowx_u32(&x324, &x325, x323, x308, 0x0);
+  fiat_p224_subborrowx_u32(&x324, &x325, x323, x304, 0x0);
   uint32_t x326;
   fiat_p224_uint1 x327;
-  fiat_p224_subborrowx_u32(&x326, &x327, x325, x310, 0x0);
+  fiat_p224_subborrowx_u32(&x326, &x327, x325, x306, UINT32_C(0xffffffff));
   uint32_t x328;
   fiat_p224_uint1 x329;
-  fiat_p224_subborrowx_u32(&x328, &x329, x327, x312, UINT32_C(0xffffffff));
+  fiat_p224_subborrowx_u32(&x328, &x329, x327, x308, UINT32_C(0xffffffff));
   uint32_t x330;
   fiat_p224_uint1 x331;
-  fiat_p224_subborrowx_u32(&x330, &x331, x329, x314, UINT32_C(0xffffffff));
+  fiat_p224_subborrowx_u32(&x330, &x331, x329, x310, UINT32_C(0xffffffff));
   uint32_t x332;
   fiat_p224_uint1 x333;
-  fiat_p224_subborrowx_u32(&x332, &x333, x331, x316, UINT32_C(0xffffffff));
+  fiat_p224_subborrowx_u32(&x332, &x333, x331, x318, UINT32_C(0xffffffff));
   uint32_t x334;
   fiat_p224_uint1 x335;
-  fiat_p224_subborrowx_u32(&x334, &x335, x333, x320, UINT32_C(0xffffffff));
+  fiat_p224_addcarryx_u32(&x334, &x335, x319, 0x0, 0x0);
   uint32_t x336;
   fiat_p224_uint1 x337;
-  fiat_p224_addcarryx_u32(&x336, &x337, x321, 0x0, x287);
+  fiat_p224_subborrowx_u32(&x336, &x337, x333, (fiat_p224_uint1)x334, 0x0);
   uint32_t x338;
-  fiat_p224_uint1 x339;
-  fiat_p224_subborrowx_u32(&x338, &x339, x335, (fiat_p224_uint1)x336, 0x0);
+  fiat_p224_cmovznz_u32(&x338, x337, x320, x300);
+  uint32_t x339;
+  fiat_p224_cmovznz_u32(&x339, x337, x322, x302);
   uint32_t x340;
-  fiat_p224_cmovznz_u32(&x340, x339, x322, x306);
+  fiat_p224_cmovznz_u32(&x340, x337, x324, x304);
   uint32_t x341;
-  fiat_p224_cmovznz_u32(&x341, x339, x324, x308);
+  fiat_p224_cmovznz_u32(&x341, x337, x326, x306);
   uint32_t x342;
-  fiat_p224_cmovznz_u32(&x342, x339, x326, x310);
+  fiat_p224_cmovznz_u32(&x342, x337, x328, x308);
   uint32_t x343;
-  fiat_p224_cmovznz_u32(&x343, x339, x328, x312);
+  fiat_p224_cmovznz_u32(&x343, x337, x330, x310);
   uint32_t x344;
-  fiat_p224_cmovznz_u32(&x344, x339, x330, x314);
-  uint32_t x345;
-  fiat_p224_cmovznz_u32(&x345, x339, x332, x316);
-  uint32_t x346;
-  fiat_p224_cmovznz_u32(&x346, x339, x334, x320);
-  out1[0] = x340;
-  out1[1] = x341;
-  out1[2] = x342;
-  out1[3] = x343;
-  out1[4] = x344;
-  out1[5] = x345;
-  out1[6] = x346;
+  fiat_p224_cmovznz_u32(&x344, x337, x332, x318);
+  out1[0] = x338;
+  out1[1] = x339;
+  out1[2] = x340;
+  out1[3] = x341;
+  out1[4] = x342;
+  out1[5] = x343;
+  out1[6] = x344;
 }
 
 /*
@@ -2634,88 +2631,82 @@ static void fiat_p224_to_bytes(uint8_t out1[28], const uint32_t arg1[7]) {
   uint8_t x11 = (uint8_t)(x8 & UINT8_C(0xff));
   uint8_t x12 = (uint8_t)(x10 >> 8);
   uint8_t x13 = (uint8_t)(x10 & UINT8_C(0xff));
-  fiat_p224_uint1 x14 = (fiat_p224_uint1)(x12 >> 8);
-  uint8_t x15 = (uint8_t)(x12 & UINT8_C(0xff));
-  uint32_t x16 = (x14 + x6);
-  uint32_t x17 = (x16 >> 8);
-  uint8_t x18 = (uint8_t)(x16 & UINT8_C(0xff));
-  uint32_t x19 = (x17 >> 8);
-  uint8_t x20 = (uint8_t)(x17 & UINT8_C(0xff));
-  uint8_t x21 = (uint8_t)(x19 >> 8);
-  uint8_t x22 = (uint8_t)(x19 & UINT8_C(0xff));
-  fiat_p224_uint1 x23 = (fiat_p224_uint1)(x21 >> 8);
-  uint8_t x24 = (uint8_t)(x21 & UINT8_C(0xff));
-  uint32_t x25 = (x23 + x5);
-  uint32_t x26 = (x25 >> 8);
-  uint8_t x27 = (uint8_t)(x25 & UINT8_C(0xff));
-  uint32_t x28 = (x26 >> 8);
+  uint8_t x14 = (uint8_t)(x12 & UINT8_C(0xff));
+  uint32_t x15 = (0x0 + x6);
+  uint32_t x16 = (x15 >> 8);
+  uint8_t x17 = (uint8_t)(x15 & UINT8_C(0xff));
+  uint32_t x18 = (x16 >> 8);
+  uint8_t x19 = (uint8_t)(x16 & UINT8_C(0xff));
+  uint8_t x20 = (uint8_t)(x18 >> 8);
+  uint8_t x21 = (uint8_t)(x18 & UINT8_C(0xff));
+  uint8_t x22 = (uint8_t)(x20 & UINT8_C(0xff));
+  uint32_t x23 = (0x0 + x5);
+  uint32_t x24 = (x23 >> 8);
+  uint8_t x25 = (uint8_t)(x23 & UINT8_C(0xff));
+  uint32_t x26 = (x24 >> 8);
+  uint8_t x27 = (uint8_t)(x24 & UINT8_C(0xff));
+  uint8_t x28 = (uint8_t)(x26 >> 8);
   uint8_t x29 = (uint8_t)(x26 & UINT8_C(0xff));
-  uint8_t x30 = (uint8_t)(x28 >> 8);
-  uint8_t x31 = (uint8_t)(x28 & UINT8_C(0xff));
-  fiat_p224_uint1 x32 = (fiat_p224_uint1)(x30 >> 8);
-  uint8_t x33 = (uint8_t)(x30 & UINT8_C(0xff));
-  uint32_t x34 = (x32 + x4);
-  uint32_t x35 = (x34 >> 8);
-  uint8_t x36 = (uint8_t)(x34 & UINT8_C(0xff));
-  uint32_t x37 = (x35 >> 8);
-  uint8_t x38 = (uint8_t)(x35 & UINT8_C(0xff));
-  uint8_t x39 = (uint8_t)(x37 >> 8);
-  uint8_t x40 = (uint8_t)(x37 & UINT8_C(0xff));
-  fiat_p224_uint1 x41 = (fiat_p224_uint1)(x39 >> 8);
-  uint8_t x42 = (uint8_t)(x39 & UINT8_C(0xff));
-  uint32_t x43 = (x41 + x3);
-  uint32_t x44 = (x43 >> 8);
-  uint8_t x45 = (uint8_t)(x43 & UINT8_C(0xff));
-  uint32_t x46 = (x44 >> 8);
-  uint8_t x47 = (uint8_t)(x44 & UINT8_C(0xff));
-  uint8_t x48 = (uint8_t)(x46 >> 8);
-  uint8_t x49 = (uint8_t)(x46 & UINT8_C(0xff));
-  fiat_p224_uint1 x50 = (fiat_p224_uint1)(x48 >> 8);
+  uint8_t x30 = (uint8_t)(x28 & UINT8_C(0xff));
+  uint32_t x31 = (0x0 + x4);
+  uint32_t x32 = (x31 >> 8);
+  uint8_t x33 = (uint8_t)(x31 & UINT8_C(0xff));
+  uint32_t x34 = (x32 >> 8);
+  uint8_t x35 = (uint8_t)(x32 & UINT8_C(0xff));
+  uint8_t x36 = (uint8_t)(x34 >> 8);
+  uint8_t x37 = (uint8_t)(x34 & UINT8_C(0xff));
+  uint8_t x38 = (uint8_t)(x36 & UINT8_C(0xff));
+  uint32_t x39 = (0x0 + x3);
+  uint32_t x40 = (x39 >> 8);
+  uint8_t x41 = (uint8_t)(x39 & UINT8_C(0xff));
+  uint32_t x42 = (x40 >> 8);
+  uint8_t x43 = (uint8_t)(x40 & UINT8_C(0xff));
+  uint8_t x44 = (uint8_t)(x42 >> 8);
+  uint8_t x45 = (uint8_t)(x42 & UINT8_C(0xff));
+  uint8_t x46 = (uint8_t)(x44 & UINT8_C(0xff));
+  uint32_t x47 = (0x0 + x2);
+  uint32_t x48 = (x47 >> 8);
+  uint8_t x49 = (uint8_t)(x47 & UINT8_C(0xff));
+  uint32_t x50 = (x48 >> 8);
   uint8_t x51 = (uint8_t)(x48 & UINT8_C(0xff));
-  uint32_t x52 = (x50 + x2);
-  uint32_t x53 = (x52 >> 8);
+  uint8_t x52 = (uint8_t)(x50 >> 8);
+  uint8_t x53 = (uint8_t)(x50 & UINT8_C(0xff));
   uint8_t x54 = (uint8_t)(x52 & UINT8_C(0xff));
-  uint32_t x55 = (x53 >> 8);
-  uint8_t x56 = (uint8_t)(x53 & UINT8_C(0xff));
-  uint8_t x57 = (uint8_t)(x55 >> 8);
-  uint8_t x58 = (uint8_t)(x55 & UINT8_C(0xff));
-  fiat_p224_uint1 x59 = (fiat_p224_uint1)(x57 >> 8);
-  uint8_t x60 = (uint8_t)(x57 & UINT8_C(0xff));
-  uint32_t x61 = (x59 + x1);
-  uint32_t x62 = (x61 >> 8);
-  uint8_t x63 = (uint8_t)(x61 & UINT8_C(0xff));
-  uint32_t x64 = (x62 >> 8);
-  uint8_t x65 = (uint8_t)(x62 & UINT8_C(0xff));
-  uint8_t x66 = (uint8_t)(x64 >> 8);
-  uint8_t x67 = (uint8_t)(x64 & UINT8_C(0xff));
+  uint32_t x55 = (0x0 + x1);
+  uint32_t x56 = (x55 >> 8);
+  uint8_t x57 = (uint8_t)(x55 & UINT8_C(0xff));
+  uint32_t x58 = (x56 >> 8);
+  uint8_t x59 = (uint8_t)(x56 & UINT8_C(0xff));
+  uint8_t x60 = (uint8_t)(x58 >> 8);
+  uint8_t x61 = (uint8_t)(x58 & UINT8_C(0xff));
   out1[0] = x9;
   out1[1] = x11;
   out1[2] = x13;
-  out1[3] = x15;
-  out1[4] = x18;
-  out1[5] = x20;
-  out1[6] = x22;
-  out1[7] = x24;
-  out1[8] = x27;
-  out1[9] = x29;
-  out1[10] = x31;
-  out1[11] = x33;
-  out1[12] = x36;
-  out1[13] = x38;
-  out1[14] = x40;
-  out1[15] = x42;
-  out1[16] = x45;
-  out1[17] = x47;
-  out1[18] = x49;
-  out1[19] = x51;
-  out1[20] = x54;
-  out1[21] = x56;
-  out1[22] = x58;
-  out1[23] = x60;
-  out1[24] = x63;
-  out1[25] = x65;
-  out1[26] = x67;
-  out1[27] = x66;
+  out1[3] = x14;
+  out1[4] = x17;
+  out1[5] = x19;
+  out1[6] = x21;
+  out1[7] = x22;
+  out1[8] = x25;
+  out1[9] = x27;
+  out1[10] = x29;
+  out1[11] = x30;
+  out1[12] = x33;
+  out1[13] = x35;
+  out1[14] = x37;
+  out1[15] = x38;
+  out1[16] = x41;
+  out1[17] = x43;
+  out1[18] = x45;
+  out1[19] = x46;
+  out1[20] = x49;
+  out1[21] = x51;
+  out1[22] = x53;
+  out1[23] = x54;
+  out1[24] = x57;
+  out1[25] = x59;
+  out1[26] = x61;
+  out1[27] = x60;
 }
 
 /*
@@ -2754,36 +2745,30 @@ static void fiat_p224_from_bytes(uint32_t out1[7], const uint8_t arg1[28]) {
   uint32_t x27 = ((uint32_t)(arg1[1]) << 8);
   uint8_t x28 = (arg1[0]);
   uint32_t x29 = (x28 + (x27 + (x26 + x25)));
-  fiat_p224_uint1 x30 = (fiat_p224_uint1)((uint64_t)x29 >> 32);
-  uint32_t x31 = (x29 & UINT32_C(0xffffffff));
-  uint32_t x32 = (x4 + (x3 + (x2 + x1)));
-  uint32_t x33 = (x8 + (x7 + (x6 + x5)));
-  uint32_t x34 = (x12 + (x11 + (x10 + x9)));
-  uint32_t x35 = (x16 + (x15 + (x14 + x13)));
-  uint32_t x36 = (x20 + (x19 + (x18 + x17)));
-  uint32_t x37 = (x24 + (x23 + (x22 + x21)));
-  uint32_t x38 = (x30 + x37);
-  fiat_p224_uint1 x39 = (fiat_p224_uint1)((uint64_t)x38 >> 32);
-  uint32_t x40 = (x38 & UINT32_C(0xffffffff));
-  uint32_t x41 = (x39 + x36);
-  fiat_p224_uint1 x42 = (fiat_p224_uint1)((uint64_t)x41 >> 32);
-  uint32_t x43 = (x41 & UINT32_C(0xffffffff));
-  uint32_t x44 = (x42 + x35);
-  fiat_p224_uint1 x45 = (fiat_p224_uint1)((uint64_t)x44 >> 32);
-  uint32_t x46 = (x44 & UINT32_C(0xffffffff));
-  uint32_t x47 = (x45 + x34);
-  fiat_p224_uint1 x48 = (fiat_p224_uint1)((uint64_t)x47 >> 32);
-  uint32_t x49 = (x47 & UINT32_C(0xffffffff));
-  uint32_t x50 = (x48 + x33);
-  fiat_p224_uint1 x51 = (fiat_p224_uint1)((uint64_t)x50 >> 32);
-  uint32_t x52 = (x50 & UINT32_C(0xffffffff));
-  uint32_t x53 = (x51 + x32);
-  out1[0] = x31;
-  out1[1] = x40;
-  out1[2] = x43;
-  out1[3] = x46;
-  out1[4] = x49;
-  out1[5] = x52;
-  out1[6] = x53;
+  uint32_t x30 = (x29 & UINT32_C(0xffffffff));
+  uint32_t x31 = (x4 + (x3 + (x2 + x1)));
+  uint32_t x32 = (x8 + (x7 + (x6 + x5)));
+  uint32_t x33 = (x12 + (x11 + (x10 + x9)));
+  uint32_t x34 = (x16 + (x15 + (x14 + x13)));
+  uint32_t x35 = (x20 + (x19 + (x18 + x17)));
+  uint32_t x36 = (x24 + (x23 + (x22 + x21)));
+  uint32_t x37 = (0x0 + x36);
+  uint32_t x38 = (x37 & UINT32_C(0xffffffff));
+  uint32_t x39 = (0x0 + x35);
+  uint32_t x40 = (x39 & UINT32_C(0xffffffff));
+  uint32_t x41 = (0x0 + x34);
+  uint32_t x42 = (x41 & UINT32_C(0xffffffff));
+  uint32_t x43 = (0x0 + x33);
+  uint32_t x44 = (x43 & UINT32_C(0xffffffff));
+  uint32_t x45 = (0x0 + x32);
+  uint32_t x46 = (x45 & UINT32_C(0xffffffff));
+  uint32_t x47 = (0x0 + x31);
+  out1[0] = x30;
+  out1[1] = x38;
+  out1[2] = x40;
+  out1[3] = x42;
+  out1[4] = x44;
+  out1[5] = x46;
+  out1[6] = x47;
 }
 
diff --git a/p224_64.c b/p224_64.c
index d67ee56ca..05e0e02af 100644
--- a/p224_64.c
+++ b/p224_64.c
@@ -926,202 +926,181 @@ static void fiat_p224_from_montgomery(uint64_t out1[4], const uint64_t arg1[4])
   fiat_p224_addcarryx_u64(&x20, &x21, x19, x12, 0x0);
   uint64_t x22;
   fiat_p224_uint1 x23;
-  fiat_p224_addcarryx_u64(&x22, &x23, x13, 0x0, x5);
+  fiat_p224_addcarryx_u64(&x22, &x23, 0x0, (arg1[1]), x16);
   uint64_t x24;
   fiat_p224_uint1 x25;
-  fiat_p224_addcarryx_u64(&x24, &x25, x21, x22, 0x0);
+  fiat_p224_addcarryx_u64(&x24, &x25, x23, 0x0, x18);
   uint64_t x26;
   fiat_p224_uint1 x27;
-  fiat_p224_addcarryx_u64(&x26, &x27, 0x0, (arg1[1]), x16);
+  fiat_p224_addcarryx_u64(&x26, &x27, x25, 0x0, x20);
   uint64_t x28;
-  fiat_p224_uint1 x29;
-  fiat_p224_addcarryx_u64(&x28, &x29, x27, 0x0, x18);
+  uint64_t x29;
+  fiat_p224_mulx_u64(&x28, &x29, x22, UINT64_C(0xffffffffffffffff));
   uint64_t x30;
-  fiat_p224_uint1 x31;
-  fiat_p224_addcarryx_u64(&x30, &x31, x29, 0x0, x20);
+  uint64_t x31;
+  fiat_p224_mulx_u64(&x30, &x31, x28, UINT32_C(0xffffffff));
   uint64_t x32;
-  fiat_p224_uint1 x33;
-  fiat_p224_addcarryx_u64(&x32, &x33, x31, 0x0, x24);
+  uint64_t x33;
+  fiat_p224_mulx_u64(&x32, &x33, x28, UINT64_C(0xffffffffffffffff));
   uint64_t x34;
-  fiat_p224_uint1 x35;
-  fiat_p224_addcarryx_u64(&x34, &x35, x25, 0x0, 0x0);
+  uint64_t x35;
+  fiat_p224_mulx_u64(&x34, &x35, x28, UINT64_C(0xffffffff00000000));
   uint64_t x36;
   fiat_p224_uint1 x37;
-  fiat_p224_addcarryx_u64(&x36, &x37, x33, 0x0, (fiat_p224_uint1)x34);
+  fiat_p224_addcarryx_u64(&x36, &x37, 0x0, x32, x35);
   uint64_t x38;
-  uint64_t x39;
-  fiat_p224_mulx_u64(&x38, &x39, x26, UINT64_C(0xffffffffffffffff));
+  fiat_p224_uint1 x39;
+  fiat_p224_addcarryx_u64(&x38, &x39, x37, x30, x33);
   uint64_t x40;
-  uint64_t x41;
-  fiat_p224_mulx_u64(&x40, &x41, x38, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x41;
+  fiat_p224_addcarryx_u64(&x40, &x41, 0x0, x28, x22);
   uint64_t x42;
-  uint64_t x43;
-  fiat_p224_mulx_u64(&x42, &x43, x38, UINT64_C(0xffffffffffffffff));
+  fiat_p224_uint1 x43;
+  fiat_p224_addcarryx_u64(&x42, &x43, x41, x34, x24);
   uint64_t x44;
-  uint64_t x45;
-  fiat_p224_mulx_u64(&x44, &x45, x38, UINT64_C(0xffffffff00000000));
+  fiat_p224_uint1 x45;
+  fiat_p224_addcarryx_u64(&x44, &x45, x43, x36, x26);
   uint64_t x46;
   fiat_p224_uint1 x47;
-  fiat_p224_addcarryx_u64(&x46, &x47, 0x0, x42, x45);
+  fiat_p224_addcarryx_u64(&x46, &x47, x13, 0x0, x5);
   uint64_t x48;
   fiat_p224_uint1 x49;
-  fiat_p224_addcarryx_u64(&x48, &x49, x47, x40, x43);
+  fiat_p224_addcarryx_u64(&x48, &x49, x21, x46, 0x0);
   uint64_t x50;
   fiat_p224_uint1 x51;
-  fiat_p224_addcarryx_u64(&x50, &x51, 0x0, x38, x26);
+  fiat_p224_addcarryx_u64(&x50, &x51, x27, 0x0, x48);
   uint64_t x52;
   fiat_p224_uint1 x53;
-  fiat_p224_addcarryx_u64(&x52, &x53, x51, x44, x28);
+  fiat_p224_addcarryx_u64(&x52, &x53, x45, x38, x50);
   uint64_t x54;
   fiat_p224_uint1 x55;
-  fiat_p224_addcarryx_u64(&x54, &x55, x53, x46, x30);
+  fiat_p224_addcarryx_u64(&x54, &x55, 0x0, (arg1[2]), x42);
   uint64_t x56;
   fiat_p224_uint1 x57;
-  fiat_p224_addcarryx_u64(&x56, &x57, x55, x48, x32);
+  fiat_p224_addcarryx_u64(&x56, &x57, x55, 0x0, x44);
   uint64_t x58;
   fiat_p224_uint1 x59;
-  fiat_p224_addcarryx_u64(&x58, &x59, x49, 0x0, x41);
+  fiat_p224_addcarryx_u64(&x58, &x59, x57, 0x0, x52);
   uint64_t x60;
-  fiat_p224_uint1 x61;
-  fiat_p224_addcarryx_u64(&x60, &x61, x57, x58, (fiat_p224_uint1)x36);
+  uint64_t x61;
+  fiat_p224_mulx_u64(&x60, &x61, x54, UINT64_C(0xffffffffffffffff));
   uint64_t x62;
-  fiat_p224_uint1 x63;
-  fiat_p224_addcarryx_u64(&x62, &x63, 0x0, (arg1[2]), x52);
+  uint64_t x63;
+  fiat_p224_mulx_u64(&x62, &x63, x60, UINT32_C(0xffffffff));
   uint64_t x64;
-  fiat_p224_uint1 x65;
-  fiat_p224_addcarryx_u64(&x64, &x65, x63, 0x0, x54);
+  uint64_t x65;
+  fiat_p224_mulx_u64(&x64, &x65, x60, UINT64_C(0xffffffffffffffff));
   uint64_t x66;
-  fiat_p224_uint1 x67;
-  fiat_p224_addcarryx_u64(&x66, &x67, x65, 0x0, x56);
+  uint64_t x67;
+  fiat_p224_mulx_u64(&x66, &x67, x60, UINT64_C(0xffffffff00000000));
   uint64_t x68;
   fiat_p224_uint1 x69;
-  fiat_p224_addcarryx_u64(&x68, &x69, x67, 0x0, x60);
+  fiat_p224_addcarryx_u64(&x68, &x69, 0x0, x64, x67);
   uint64_t x70;
   fiat_p224_uint1 x71;
-  fiat_p224_addcarryx_u64(&x70, &x71, x61, 0x0, x37);
+  fiat_p224_addcarryx_u64(&x70, &x71, x69, x62, x65);
   uint64_t x72;
   fiat_p224_uint1 x73;
-  fiat_p224_addcarryx_u64(&x72, &x73, x69, 0x0, (fiat_p224_uint1)x70);
+  fiat_p224_addcarryx_u64(&x72, &x73, 0x0, x60, x54);
   uint64_t x74;
-  uint64_t x75;
-  fiat_p224_mulx_u64(&x74, &x75, x62, UINT64_C(0xffffffffffffffff));
+  fiat_p224_uint1 x75;
+  fiat_p224_addcarryx_u64(&x74, &x75, x73, x66, x56);
   uint64_t x76;
-  uint64_t x77;
-  fiat_p224_mulx_u64(&x76, &x77, x74, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x77;
+  fiat_p224_addcarryx_u64(&x76, &x77, x75, x68, x58);
   uint64_t x78;
-  uint64_t x79;
-  fiat_p224_mulx_u64(&x78, &x79, x74, UINT64_C(0xffffffffffffffff));
+  fiat_p224_uint1 x79;
+  fiat_p224_addcarryx_u64(&x78, &x79, x39, 0x0, x31);
   uint64_t x80;
-  uint64_t x81;
-  fiat_p224_mulx_u64(&x80, &x81, x74, UINT64_C(0xffffffff00000000));
+  fiat_p224_uint1 x81;
+  fiat_p224_addcarryx_u64(&x80, &x81, x53, x78, 0x0);
   uint64_t x82;
   fiat_p224_uint1 x83;
-  fiat_p224_addcarryx_u64(&x82, &x83, 0x0, x78, x81);
+  fiat_p224_addcarryx_u64(&x82, &x83, x59, 0x0, x80);
   uint64_t x84;
   fiat_p224_uint1 x85;
-  fiat_p224_addcarryx_u64(&x84, &x85, x83, x76, x79);
+  fiat_p224_addcarryx_u64(&x84, &x85, x77, x70, x82);
   uint64_t x86;
   fiat_p224_uint1 x87;
-  fiat_p224_addcarryx_u64(&x86, &x87, 0x0, x74, x62);
+  fiat_p224_addcarryx_u64(&x86, &x87, 0x0, (arg1[3]), x74);
   uint64_t x88;
   fiat_p224_uint1 x89;
-  fiat_p224_addcarryx_u64(&x88, &x89, x87, x80, x64);
+  fiat_p224_addcarryx_u64(&x88, &x89, x87, 0x0, x76);
   uint64_t x90;
   fiat_p224_uint1 x91;
-  fiat_p224_addcarryx_u64(&x90, &x91, x89, x82, x66);
+  fiat_p224_addcarryx_u64(&x90, &x91, x89, 0x0, x84);
   uint64_t x92;
-  fiat_p224_uint1 x93;
-  fiat_p224_addcarryx_u64(&x92, &x93, x91, x84, x68);
+  uint64_t x93;
+  fiat_p224_mulx_u64(&x92, &x93, x86, UINT64_C(0xffffffffffffffff));
   uint64_t x94;
-  fiat_p224_uint1 x95;
-  fiat_p224_addcarryx_u64(&x94, &x95, x85, 0x0, x77);
+  uint64_t x95;
+  fiat_p224_mulx_u64(&x94, &x95, x92, UINT32_C(0xffffffff));
   uint64_t x96;
-  fiat_p224_uint1 x97;
-  fiat_p224_addcarryx_u64(&x96, &x97, x93, x94, (fiat_p224_uint1)x72);
+  uint64_t x97;
+  fiat_p224_mulx_u64(&x96, &x97, x92, UINT64_C(0xffffffffffffffff));
   uint64_t x98;
-  fiat_p224_uint1 x99;
-  fiat_p224_addcarryx_u64(&x98, &x99, 0x0, (arg1[3]), x88);
+  uint64_t x99;
+  fiat_p224_mulx_u64(&x98, &x99, x92, UINT64_C(0xffffffff00000000));
   uint64_t x100;
   fiat_p224_uint1 x101;
-  fiat_p224_addcarryx_u64(&x100, &x101, x99, 0x0, x90);
+  fiat_p224_addcarryx_u64(&x100, &x101, 0x0, x96, x99);
   uint64_t x102;
   fiat_p224_uint1 x103;
-  fiat_p224_addcarryx_u64(&x102, &x103, x101, 0x0, x92);
+  fiat_p224_addcarryx_u64(&x102, &x103, x101, x94, x97);
   uint64_t x104;
   fiat_p224_uint1 x105;
-  fiat_p224_addcarryx_u64(&x104, &x105, x103, 0x0, x96);
+  fiat_p224_addcarryx_u64(&x104, &x105, 0x0, x92, x86);
   uint64_t x106;
   fiat_p224_uint1 x107;
-  fiat_p224_addcarryx_u64(&x106, &x107, x97, 0x0, x73);
+  fiat_p224_addcarryx_u64(&x106, &x107, x105, x98, x88);
   uint64_t x108;
   fiat_p224_uint1 x109;
-  fiat_p224_addcarryx_u64(&x108, &x109, x105, 0x0, (fiat_p224_uint1)x106);
+  fiat_p224_addcarryx_u64(&x108, &x109, x107, x100, x90);
   uint64_t x110;
-  uint64_t x111;
-  fiat_p224_mulx_u64(&x110, &x111, x98, UINT64_C(0xffffffffffffffff));
+  fiat_p224_uint1 x111;
+  fiat_p224_addcarryx_u64(&x110, &x111, x71, 0x0, x63);
   uint64_t x112;
-  uint64_t x113;
-  fiat_p224_mulx_u64(&x112, &x113, x110, UINT32_C(0xffffffff));
+  fiat_p224_uint1 x113;
+  fiat_p224_addcarryx_u64(&x112, &x113, x85, x110, 0x0);
   uint64_t x114;
-  uint64_t x115;
-  fiat_p224_mulx_u64(&x114, &x115, x110, UINT64_C(0xffffffffffffffff));
+  fiat_p224_uint1 x115;
+  fiat_p224_addcarryx_u64(&x114, &x115, x91, 0x0, x112);
   uint64_t x116;
-  uint64_t x117;
-  fiat_p224_mulx_u64(&x116, &x117, x110, UINT64_C(0xffffffff00000000));
+  fiat_p224_uint1 x117;
+  fiat_p224_addcarryx_u64(&x116, &x117, x109, x102, x114);
   uint64_t x118;
   fiat_p224_uint1 x119;
-  fiat_p224_addcarryx_u64(&x118, &x119, 0x0, x114, x117);
+  fiat_p224_addcarryx_u64(&x118, &x119, x103, 0x0, x95);
   uint64_t x120;
   fiat_p224_uint1 x121;
-  fiat_p224_addcarryx_u64(&x120, &x121, x119, x112, x115);
+  fiat_p224_addcarryx_u64(&x120, &x121, x117, x118, 0x0);
   uint64_t x122;
   fiat_p224_uint1 x123;
-  fiat_p224_addcarryx_u64(&x122, &x123, 0x0, x110, x98);
+  fiat_p224_subborrowx_u64(&x122, &x123, 0x0, x106, 0x1);
   uint64_t x124;
   fiat_p224_uint1 x125;
-  fiat_p224_addcarryx_u64(&x124, &x125, x123, x116, x100);
+  fiat_p224_subborrowx_u64(&x124, &x125, x123, x108, UINT64_C(0xffffffff00000000));
   uint64_t x126;
   fiat_p224_uint1 x127;
-  fiat_p224_addcarryx_u64(&x126, &x127, x125, x118, x102);
+  fiat_p224_subborrowx_u64(&x126, &x127, x125, x116, UINT64_C(0xffffffffffffffff));
   uint64_t x128;
   fiat_p224_uint1 x129;
-  fiat_p224_addcarryx_u64(&x128, &x129, x127, x120, x104);
+  fiat_p224_subborrowx_u64(&x128, &x129, x127, x120, UINT32_C(0xffffffff));
   uint64_t x130;
   fiat_p224_uint1 x131;
-  fiat_p224_addcarryx_u64(&x130, &x131, x121, 0x0, x113);
+  fiat_p224_subborrowx_u64(&x130, &x131, x129, 0x0, 0x0);
   uint64_t x132;
-  fiat_p224_uint1 x133;
-  fiat_p224_addcarryx_u64(&x132, &x133, x129, x130, (fiat_p224_uint1)x108);
+  fiat_p224_cmovznz_u64(&x132, x131, x122, x106);
+  uint64_t x133;
+  fiat_p224_cmovznz_u64(&x133, x131, x124, x108);
   uint64_t x134;
-  fiat_p224_uint1 x135;
-  fiat_p224_subborrowx_u64(&x134, &x135, 0x0, x124, 0x1);
-  uint64_t x136;
-  fiat_p224_uint1 x137;
-  fiat_p224_subborrowx_u64(&x136, &x137, x135, x126, UINT64_C(0xffffffff00000000));
-  uint64_t x138;
-  fiat_p224_uint1 x139;
-  fiat_p224_subborrowx_u64(&x138, &x139, x137, x128, UINT64_C(0xffffffffffffffff));
-  uint64_t x140;
-  fiat_p224_uint1 x141;
-  fiat_p224_subborrowx_u64(&x140, &x141, x139, x132, UINT32_C(0xffffffff));
-  uint64_t x142;
-  fiat_p224_uint1 x143;
-  fiat_p224_addcarryx_u64(&x142, &x143, x133, 0x0, x109);
-  uint64_t x144;
-  fiat_p224_uint1 x145;
-  fiat_p224_subborrowx_u64(&x144, &x145, x141, (fiat_p224_uint1)x142, 0x0);
-  uint64_t x146;
-  fiat_p224_cmovznz_u64(&x146, x145, x134, x124);
-  uint64_t x147;
-  fiat_p224_cmovznz_u64(&x147, x145, x136, x126);
-  uint64_t x148;
-  fiat_p224_cmovznz_u64(&x148, x145, x138, x128);
-  uint64_t x149;
-  fiat_p224_cmovznz_u64(&x149, x145, x140, x132);
-  out1[0] = x146;
-  out1[1] = x147;
-  out1[2] = x148;
-  out1[3] = x149;
+  fiat_p224_cmovznz_u64(&x134, x131, x126, x116);
+  uint64_t x135;
+  fiat_p224_cmovznz_u64(&x135, x131, x128, x120);
+  out1[0] = x132;
+  out1[1] = x133;
+  out1[2] = x134;
+  out1[3] = x135;
 }
 
 /*
@@ -1183,28 +1162,28 @@ static void fiat_p224_to_bytes(uint8_t out1[32], const uint64_t arg1[4]) {
   uint8_t x16 = (uint8_t)(x13 & UINT8_C(0xff));
   uint8_t x17 = (uint8_t)(x15 >> 8);
   uint8_t x18 = (uint8_t)(x15 & UINT8_C(0xff));
-  fiat_p224_uint1 x19 = (fiat_p224_uint1)(x17 >> 8);
-  uint8_t x20 = (uint8_t)(x17 & UINT8_C(0xff));
-  uint64_t x21 = (x19 + x3);
-  uint64_t x22 = (x21 >> 8);
-  uint8_t x23 = (uint8_t)(x21 & UINT8_C(0xff));
-  uint64_t x24 = (x22 >> 8);
-  uint8_t x25 = (uint8_t)(x22 & UINT8_C(0xff));
-  uint64_t x26 = (x24 >> 8);
-  uint8_t x27 = (uint8_t)(x24 & UINT8_C(0xff));
-  uint64_t x28 = (x26 >> 8);
-  uint8_t x29 = (uint8_t)(x26 & UINT8_C(0xff));
-  uint64_t x30 = (x28 >> 8);
-  uint8_t x31 = (uint8_t)(x28 & UINT8_C(0xff));
-  uint64_t x32 = (x30 >> 8);
-  uint8_t x33 = (uint8_t)(x30 & UINT8_C(0xff));
-  uint8_t x34 = (uint8_t)(x32 >> 8);
-  uint8_t x35 = (uint8_t)(x32 & UINT8_C(0xff));
-  fiat_p224_uint1 x36 = (fiat_p224_uint1)(x34 >> 8);
-  uint8_t x37 = (uint8_t)(x34 & UINT8_C(0xff));
-  uint64_t x38 = (x36 + x2);
-  uint64_t x39 = (x38 >> 8);
-  uint8_t x40 = (uint8_t)(x38 & UINT8_C(0xff));
+  uint8_t x19 = (uint8_t)(x17 & UINT8_C(0xff));
+  uint64_t x20 = (0x0 + x3);
+  uint64_t x21 = (x20 >> 8);
+  uint8_t x22 = (uint8_t)(x20 & UINT8_C(0xff));
+  uint64_t x23 = (x21 >> 8);
+  uint8_t x24 = (uint8_t)(x21 & UINT8_C(0xff));
+  uint64_t x25 = (x23 >> 8);
+  uint8_t x26 = (uint8_t)(x23 & UINT8_C(0xff));
+  uint64_t x27 = (x25 >> 8);
+  uint8_t x28 = (uint8_t)(x25 & UINT8_C(0xff));
+  uint64_t x29 = (x27 >> 8);
+  uint8_t x30 = (uint8_t)(x27 & UINT8_C(0xff));
+  uint64_t x31 = (x29 >> 8);
+  uint8_t x32 = (uint8_t)(x29 & UINT8_C(0xff));
+  uint8_t x33 = (uint8_t)(x31 >> 8);
+  uint8_t x34 = (uint8_t)(x31 & UINT8_C(0xff));
+  uint8_t x35 = (uint8_t)(x33 & UINT8_C(0xff));
+  uint64_t x36 = (0x0 + x2);
+  uint64_t x37 = (x36 >> 8);
+  uint8_t x38 = (uint8_t)(x36 & UINT8_C(0xff));
+  uint64_t x39 = (x37 >> 8);
+  uint8_t x40 = (uint8_t)(x37 & UINT8_C(0xff));
   uint64_t x41 = (x39 >> 8);
   uint8_t x42 = (uint8_t)(x39 & UINT8_C(0xff));
   uint64_t x43 = (x41 >> 8);
@@ -1213,27 +1192,17 @@ static void fiat_p224_to_bytes(uint8_t out1[32], const uint64_t arg1[4]) {
   uint8_t x46 = (uint8_t)(x43 & UINT8_C(0xff));
   uint64_t x47 = (x45 >> 8);
   uint8_t x48 = (uint8_t)(x45 & UINT8_C(0xff));
-  uint64_t x49 = (x47 >> 8);
+  uint8_t x49 = (uint8_t)(x47 >> 8);
   uint8_t x50 = (uint8_t)(x47 & UINT8_C(0xff));
-  uint8_t x51 = (uint8_t)(x49 >> 8);
-  uint8_t x52 = (uint8_t)(x49 & UINT8_C(0xff));
-  fiat_p224_uint1 x53 = (fiat_p224_uint1)(x51 >> 8);
-  uint8_t x54 = (uint8_t)(x51 & UINT8_C(0xff));
-  uint64_t x55 = (x53 + x1);
-  uint64_t x56 = (x55 >> 8);
-  uint8_t x57 = (uint8_t)(x55 & UINT8_C(0xff));
-  uint64_t x58 = (x56 >> 8);
-  uint8_t x59 = (uint8_t)(x56 & UINT8_C(0xff));
-  uint8_t x60 = (uint8_t)(x58 >> 8);
-  uint8_t x61 = (uint8_t)(x58 & UINT8_C(0xff));
-  fiat_p224_uint1 x62 = (fiat_p224_uint1)(x60 >> 8);
-  uint8_t x63 = (uint8_t)(x60 & UINT8_C(0xff));
-  fiat_p224_uint1 x64 = (fiat_p224_uint1)(x62 >> 8);
-  fiat_p224_uint1 x65 = (fiat_p224_uint1)(x62 & UINT8_C(0xff));
-  fiat_p224_uint1 x66 = (fiat_p224_uint1)(x64 >> 8);
-  fiat_p224_uint1 x67 = (fiat_p224_uint1)(x64 & UINT8_C(0xff));
-  fiat_p224_uint1 x68 = (fiat_p224_uint1)(x66 >> 8);
-  fiat_p224_uint1 x69 = (fiat_p224_uint1)(x66 & UINT8_C(0xff));
+  uint8_t x51 = (uint8_t)(x49 & UINT8_C(0xff));
+  uint64_t x52 = (0x0 + x1);
+  uint64_t x53 = (x52 >> 8);
+  uint8_t x54 = (uint8_t)(x52 & UINT8_C(0xff));
+  uint64_t x55 = (x53 >> 8);
+  uint8_t x56 = (uint8_t)(x53 & UINT8_C(0xff));
+  uint8_t x57 = (uint8_t)(x55 >> 8);
+  uint8_t x58 = (uint8_t)(x55 & UINT8_C(0xff));
+  uint8_t x59 = (uint8_t)(x57 & UINT8_C(0xff));
   out1[0] = x6;
   out1[1] = x8;
   out1[2] = x10;
@@ -1241,31 +1210,31 @@ static void fiat_p224_to_bytes(uint8_t out1[32], const uint64_t arg1[4]) {
   out1[4] = x14;
   out1[5] = x16;
   out1[6] = x18;
-  out1[7] = x20;
-  out1[8] = x23;
-  out1[9] = x25;
-  out1[10] = x27;
-  out1[11] = x29;
-  out1[12] = x31;
-  out1[13] = x33;
-  out1[14] = x35;
-  out1[15] = x37;
-  out1[16] = x40;
-  out1[17] = x42;
-  out1[18] = x44;
-  out1[19] = x46;
-  out1[20] = x48;
-  out1[21] = x50;
-  out1[22] = x52;
-  out1[23] = x54;
-  out1[24] = x57;
-  out1[25] = x59;
-  out1[26] = x61;
-  out1[27] = x63;
-  out1[28] = x65;
-  out1[29] = x67;
-  out1[30] = x69;
-  out1[31] = x68;
+  out1[7] = x19;
+  out1[8] = x22;
+  out1[9] = x24;
+  out1[10] = x26;
+  out1[11] = x28;
+  out1[12] = x30;
+  out1[13] = x32;
+  out1[14] = x34;
+  out1[15] = x35;
+  out1[16] = x38;
+  out1[17] = x40;
+  out1[18] = x42;
+  out1[19] = x44;
+  out1[20] = x46;
+  out1[21] = x48;
+  out1[22] = x50;
+  out1[23] = x51;
+  out1[24] = x54;
+  out1[25] = x56;
+  out1[26] = x58;
+  out1[27] = x59;
+  out1[28] = 0x0;
+  out1[29] = 0x0;
+  out1[30] = 0x0;
+  out1[31] = 0x0;
 }
 
 /*
@@ -1275,54 +1244,47 @@ static void fiat_p224_to_bytes(uint8_t out1[32], const uint64_t arg1[4]) {
  *   out1: [[0x0 ~> 0xffffffffffffffff], [0x0 ~> 0xffffffffffffffff], [0x0 ~> 0xffffffffffffffff], [0x0 ~> 0xffffffff]]
  */
 static void fiat_p224_from_bytes(uint64_t out1[4], const uint8_t arg1[32]) {
-  fiat_p224_uint1 x1 = (fiat_p224_uint1)((uint64_t)(fiat_p224_uint1)(arg1[31]) << 56);
-  fiat_p224_uint1 x2 = (fiat_p224_uint1)((uint64_t)(fiat_p224_uint1)(arg1[30]) << 48);
-  fiat_p224_uint1 x3 = (fiat_p224_uint1)((uint64_t)(fiat_p224_uint1)(arg1[29]) << 40);
-  fiat_p224_uint1 x4 = (fiat_p224_uint1)((uint64_t)(fiat_p224_uint1)(arg1[28]) << 32);
-  uint64_t x5 = ((uint64_t)(arg1[27]) << 24);
-  uint64_t x6 = ((uint64_t)(arg1[26]) << 16);
-  uint64_t x7 = ((uint64_t)(arg1[25]) << 8);
-  uint8_t x8 = (arg1[24]);
-  uint64_t x9 = ((uint64_t)(arg1[23]) << 56);
-  uint64_t x10 = ((uint64_t)(arg1[22]) << 48);
-  uint64_t x11 = ((uint64_t)(arg1[21]) << 40);
-  uint64_t x12 = ((uint64_t)(arg1[20]) << 32);
-  uint64_t x13 = ((uint64_t)(arg1[19]) << 24);
-  uint64_t x14 = ((uint64_t)(arg1[18]) << 16);
-  uint64_t x15 = ((uint64_t)(arg1[17]) << 8);
-  uint8_t x16 = (arg1[16]);
-  uint64_t x17 = ((uint64_t)(arg1[15]) << 56);
-  uint64_t x18 = ((uint64_t)(arg1[14]) << 48);
-  uint64_t x19 = ((uint64_t)(arg1[13]) << 40);
-  uint64_t x20 = ((uint64_t)(arg1[12]) << 32);
-  uint64_t x21 = ((uint64_t)(arg1[11]) << 24);
-  uint64_t x22 = ((uint64_t)(arg1[10]) << 16);
-  uint64_t x23 = ((uint64_t)(arg1[9]) << 8);
-  uint8_t x24 = (arg1[8]);
-  uint64_t x25 = ((uint64_t)(arg1[7]) << 56);
-  uint64_t x26 = ((uint64_t)(arg1[6]) << 48);
-  uint64_t x27 = ((uint64_t)(arg1[5]) << 40);
-  uint64_t x28 = ((uint64_t)(arg1[4]) << 32);
-  uint64_t x29 = ((uint64_t)(arg1[3]) << 24);
-  uint64_t x30 = ((uint64_t)(arg1[2]) << 16);
-  uint64_t x31 = ((uint64_t)(arg1[1]) << 8);
-  uint8_t x32 = (arg1[0]);
-  uint64_t x33 = (x32 + (x31 + (x30 + (x29 + (x28 + (x27 + (x26 + x25)))))));
-  fiat_p224_uint1 x34 = (fiat_p224_uint1)((fiat_p224_uint128)x33 >> 64);
-  uint64_t x35 = (x33 & UINT64_C(0xffffffffffffffff));
-  uint64_t x36 = (x8 + (x7 + (x6 + (x5 + (uint64_t)(fiat_p224_uint1)(x4 + (fiat_p224_uint1)(x3 + (fiat_p224_uint1)(x2 + x1)))))));
-  uint64_t x37 = (x16 + (x15 + (x14 + (x13 + (x12 + (x11 + (x10 + x9)))))));
-  uint64_t x38 = (x24 + (x23 + (x22 + (x21 + (x20 + (x19 + (x18 + x17)))))));
-  uint64_t x39 = (x34 + x38);
-  fiat_p224_uint1 x40 = (fiat_p224_uint1)((fiat_p224_uint128)x39 >> 64);
-  uint64_t x41 = (x39 & UINT64_C(0xffffffffffffffff));
-  uint64_t x42 = (x40 + x37);
-  fiat_p224_uint1 x43 = (fiat_p224_uint1)((fiat_p224_uint128)x42 >> 64);
-  uint64_t x44 = (x42 & UINT64_C(0xffffffffffffffff));
-  uint64_t x45 = (x43 + x36);
-  out1[0] = x35;
-  out1[1] = x41;
-  out1[2] = x44;
-  out1[3] = x45;
+  uint64_t x1 = ((uint64_t)(arg1[27]) << 24);
+  uint64_t x2 = ((uint64_t)(arg1[26]) << 16);
+  uint64_t x3 = ((uint64_t)(arg1[25]) << 8);
+  uint8_t x4 = (arg1[24]);
+  uint64_t x5 = ((uint64_t)(arg1[23]) << 56);
+  uint64_t x6 = ((uint64_t)(arg1[22]) << 48);
+  uint64_t x7 = ((uint64_t)(arg1[21]) << 40);
+  uint64_t x8 = ((uint64_t)(arg1[20]) << 32);
+  uint64_t x9 = ((uint64_t)(arg1[19]) << 24);
+  uint64_t x10 = ((uint64_t)(arg1[18]) << 16);
+  uint64_t x11 = ((uint64_t)(arg1[17]) << 8);
+  uint8_t x12 = (arg1[16]);
+  uint64_t x13 = ((uint64_t)(arg1[15]) << 56);
+  uint64_t x14 = ((uint64_t)(arg1[14]) << 48);
+  uint64_t x15 = ((uint64_t)(arg1[13]) << 40);
+  uint64_t x16 = ((uint64_t)(arg1[12]) << 32);
+  uint64_t x17 = ((uint64_t)(arg1[11]) << 24);
+  uint64_t x18 = ((uint64_t)(arg1[10]) << 16);
+  uint64_t x19 = ((uint64_t)(arg1[9]) << 8);
+  uint8_t x20 = (arg1[8]);
+  uint64_t x21 = ((uint64_t)(arg1[7]) << 56);
+  uint64_t x22 = ((uint64_t)(arg1[6]) << 48);
+  uint64_t x23 = ((uint64_t)(arg1[5]) << 40);
+  uint64_t x24 = ((uint64_t)(arg1[4]) << 32);
+  uint64_t x25 = ((uint64_t)(arg1[3]) << 24);
+  uint64_t x26 = ((uint64_t)(arg1[2]) << 16);
+  uint64_t x27 = ((uint64_t)(arg1[1]) << 8);
+  uint8_t x28 = (arg1[0]);
+  uint64_t x29 = (x28 + (x27 + (x26 + (x25 + (x24 + (x23 + (x22 + x21)))))));
+  uint64_t x30 = (x29 & UINT64_C(0xffffffffffffffff));
+  uint64_t x31 = (x4 + (x3 + (x2 + (x1 + (uint64_t)0x0))));
+  uint64_t x32 = (x12 + (x11 + (x10 + (x9 + (x8 + (x7 + (x6 + x5)))))));
+  uint64_t x33 = (x20 + (x19 + (x18 + (x17 + (x16 + (x15 + (x14 + x13)))))));
+  uint64_t x34 = (0x0 + x33);
+  uint64_t x35 = (x34 & UINT64_C(0xffffffffffffffff));
+  uint64_t x36 = (0x0 + x32);
+  uint64_t x37 = (x36 & UINT64_C(0xffffffffffffffff));
+  uint64_t x38 = (0x0 + x31);
+  out1[0] = x30;
+  out1[1] = x35;
+  out1[2] = x37;
+  out1[3] = x38;
 }
 
diff --git a/p256_32.c b/p256_32.c
index 825a10ca2..fe0bf4cbc 100644
--- a/p256_32.c
+++ b/p256_32.c
@@ -168,88 +168,88 @@ static void fiat_p256_mul(uint32_t out1[8], const uint32_t arg1[8], const uint32
   fiat_p256_addcarryx_u32(&x53, &x54, x52, 0x0, x44);
   uint32_t x55;
   fiat_p256_uint1 x56;
-  fiat_p256_addcarryx_u32(&x55, &x56, x54, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x55, &x56, 0x0, x47, x23);
   uint32_t x57;
   fiat_p256_uint1 x58;
-  fiat_p256_addcarryx_u32(&x57, &x58, 0x0, x47, x23);
+  fiat_p256_addcarryx_u32(&x57, &x58, x56, x49, x25);
   uint32_t x59;
   fiat_p256_uint1 x60;
-  fiat_p256_addcarryx_u32(&x59, &x60, x58, x49, x25);
+  fiat_p256_addcarryx_u32(&x59, &x60, x58, x51, x27);
   uint32_t x61;
   fiat_p256_uint1 x62;
-  fiat_p256_addcarryx_u32(&x61, &x62, x60, x51, x27);
+  fiat_p256_addcarryx_u32(&x61, &x62, x60, x53, x29);
   uint32_t x63;
   fiat_p256_uint1 x64;
-  fiat_p256_addcarryx_u32(&x63, &x64, x62, x53, x29);
+  fiat_p256_addcarryx_u32(&x63, &x64, x62, 0x0, x31);
   uint32_t x65;
   fiat_p256_uint1 x66;
-  fiat_p256_addcarryx_u32(&x65, &x66, x64, (fiat_p256_uint1)x55, x31);
+  fiat_p256_addcarryx_u32(&x65, &x66, x64, 0x0, x33);
   uint32_t x67;
   fiat_p256_uint1 x68;
-  fiat_p256_addcarryx_u32(&x67, &x68, x66, 0x0, x33);
+  fiat_p256_addcarryx_u32(&x67, &x68, x66, x23, x35);
   uint32_t x69;
   fiat_p256_uint1 x70;
-  fiat_p256_addcarryx_u32(&x69, &x70, x68, x23, x35);
+  fiat_p256_addcarryx_u32(&x69, &x70, x68, x41, x37);
   uint32_t x71;
   fiat_p256_uint1 x72;
-  fiat_p256_addcarryx_u32(&x71, &x72, x70, x41, x37);
+  fiat_p256_addcarryx_u32(&x71, &x72, x70, x42, x39);
   uint32_t x73;
   fiat_p256_uint1 x74;
-  fiat_p256_addcarryx_u32(&x73, &x74, x72, x42, x39);
+  fiat_p256_addcarryx_u32(&x73, &x74, x72, 0x0, 0x0);
   uint32_t x75;
-  fiat_p256_uint1 x76;
-  fiat_p256_addcarryx_u32(&x75, &x76, x74, 0x0, 0x0);
+  uint32_t x76;
+  fiat_p256_mulx_u32(&x75, &x76, x1, (arg2[7]));
   uint32_t x77;
   uint32_t x78;
-  fiat_p256_mulx_u32(&x77, &x78, x1, (arg2[7]));
+  fiat_p256_mulx_u32(&x77, &x78, x1, (arg2[6]));
   uint32_t x79;
   uint32_t x80;
-  fiat_p256_mulx_u32(&x79, &x80, x1, (arg2[6]));
+  fiat_p256_mulx_u32(&x79, &x80, x1, (arg2[5]));
   uint32_t x81;
   uint32_t x82;
-  fiat_p256_mulx_u32(&x81, &x82, x1, (arg2[5]));
+  fiat_p256_mulx_u32(&x81, &x82, x1, (arg2[4]));
   uint32_t x83;
   uint32_t x84;
-  fiat_p256_mulx_u32(&x83, &x84, x1, (arg2[4]));
+  fiat_p256_mulx_u32(&x83, &x84, x1, (arg2[3]));
   uint32_t x85;
   uint32_t x86;
-  fiat_p256_mulx_u32(&x85, &x86, x1, (arg2[3]));
+  fiat_p256_mulx_u32(&x85, &x86, x1, (arg2[2]));
   uint32_t x87;
   uint32_t x88;
-  fiat_p256_mulx_u32(&x87, &x88, x1, (arg2[2]));
+  fiat_p256_mulx_u32(&x87, &x88, x1, (arg2[1]));
   uint32_t x89;
   uint32_t x90;
-  fiat_p256_mulx_u32(&x89, &x90, x1, (arg2[1]));
+  fiat_p256_mulx_u32(&x89, &x90, x1, (arg2[0]));
   uint32_t x91;
-  uint32_t x92;
-  fiat_p256_mulx_u32(&x91, &x92, x1, (arg2[0]));
+  fiat_p256_uint1 x92;
+  fiat_p256_addcarryx_u32(&x91, &x92, 0x0, x87, x90);
   uint32_t x93;
   fiat_p256_uint1 x94;
-  fiat_p256_addcarryx_u32(&x93, &x94, 0x0, x89, x92);
+  fiat_p256_addcarryx_u32(&x93, &x94, x92, x85, x88);
   uint32_t x95;
   fiat_p256_uint1 x96;
-  fiat_p256_addcarryx_u32(&x95, &x96, x94, x87, x90);
+  fiat_p256_addcarryx_u32(&x95, &x96, x94, x83, x86);
   uint32_t x97;
   fiat_p256_uint1 x98;
-  fiat_p256_addcarryx_u32(&x97, &x98, x96, x85, x88);
+  fiat_p256_addcarryx_u32(&x97, &x98, x96, x81, x84);
   uint32_t x99;
   fiat_p256_uint1 x100;
-  fiat_p256_addcarryx_u32(&x99, &x100, x98, x83, x86);
+  fiat_p256_addcarryx_u32(&x99, &x100, x98, x79, x82);
   uint32_t x101;
   fiat_p256_uint1 x102;
-  fiat_p256_addcarryx_u32(&x101, &x102, x100, x81, x84);
+  fiat_p256_addcarryx_u32(&x101, &x102, x100, x77, x80);
   uint32_t x103;
   fiat_p256_uint1 x104;
-  fiat_p256_addcarryx_u32(&x103, &x104, x102, x79, x82);
+  fiat_p256_addcarryx_u32(&x103, &x104, x102, x75, x78);
   uint32_t x105;
   fiat_p256_uint1 x106;
-  fiat_p256_addcarryx_u32(&x105, &x106, x104, x77, x80);
+  fiat_p256_addcarryx_u32(&x105, &x106, x104, 0x0, x76);
   uint32_t x107;
   fiat_p256_uint1 x108;
-  fiat_p256_addcarryx_u32(&x107, &x108, x106, 0x0, x78);
+  fiat_p256_addcarryx_u32(&x107, &x108, 0x0, x89, x57);
   uint32_t x109;
   fiat_p256_uint1 x110;
-  fiat_p256_addcarryx_u32(&x109, &x110, 0x0, x91, x59);
+  fiat_p256_addcarryx_u32(&x109, &x110, x108, x91, x59);
   uint32_t x111;
   fiat_p256_uint1 x112;
   fiat_p256_addcarryx_u32(&x111, &x112, x110, x93, x61);
@@ -270,115 +270,115 @@ static void fiat_p256_mul(uint32_t out1[8], const uint32_t arg1[8], const uint32
   fiat_p256_addcarryx_u32(&x121, &x122, x120, x103, x71);
   uint32_t x123;
   fiat_p256_uint1 x124;
-  fiat_p256_addcarryx_u32(&x123, &x124, x122, x105, x73);
+  fiat_p256_addcarryx_u32(&x123, &x124, x122, x105, (fiat_p256_uint1)x73);
   uint32_t x125;
-  fiat_p256_uint1 x126;
-  fiat_p256_addcarryx_u32(&x125, &x126, x124, x107, (fiat_p256_uint1)x75);
+  uint32_t x126;
+  fiat_p256_mulx_u32(&x125, &x126, x107, UINT32_C(0xffffffff));
   uint32_t x127;
   uint32_t x128;
-  fiat_p256_mulx_u32(&x127, &x128, x109, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u32(&x127, &x128, x107, UINT32_C(0xffffffff));
   uint32_t x129;
   uint32_t x130;
-  fiat_p256_mulx_u32(&x129, &x130, x109, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u32(&x129, &x130, x107, UINT32_C(0xffffffff));
   uint32_t x131;
   uint32_t x132;
-  fiat_p256_mulx_u32(&x131, &x132, x109, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u32(&x131, &x132, x107, UINT32_C(0xffffffff));
   uint32_t x133;
-  uint32_t x134;
-  fiat_p256_mulx_u32(&x133, &x134, x109, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x134;
+  fiat_p256_addcarryx_u32(&x133, &x134, 0x0, x129, x132);
   uint32_t x135;
   fiat_p256_uint1 x136;
-  fiat_p256_addcarryx_u32(&x135, &x136, 0x0, x131, x134);
+  fiat_p256_addcarryx_u32(&x135, &x136, x134, x127, x130);
   uint32_t x137;
   fiat_p256_uint1 x138;
-  fiat_p256_addcarryx_u32(&x137, &x138, x136, x129, x132);
+  fiat_p256_addcarryx_u32(&x137, &x138, x136, 0x0, x128);
   uint32_t x139;
   fiat_p256_uint1 x140;
-  fiat_p256_addcarryx_u32(&x139, &x140, x138, 0x0, x130);
+  fiat_p256_addcarryx_u32(&x139, &x140, 0x0, x131, x107);
   uint32_t x141;
   fiat_p256_uint1 x142;
-  fiat_p256_addcarryx_u32(&x141, &x142, x140, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x141, &x142, x140, x133, x109);
   uint32_t x143;
   fiat_p256_uint1 x144;
-  fiat_p256_addcarryx_u32(&x143, &x144, 0x0, x133, x109);
+  fiat_p256_addcarryx_u32(&x143, &x144, x142, x135, x111);
   uint32_t x145;
   fiat_p256_uint1 x146;
-  fiat_p256_addcarryx_u32(&x145, &x146, x144, x135, x111);
+  fiat_p256_addcarryx_u32(&x145, &x146, x144, x137, x113);
   uint32_t x147;
   fiat_p256_uint1 x148;
-  fiat_p256_addcarryx_u32(&x147, &x148, x146, x137, x113);
+  fiat_p256_addcarryx_u32(&x147, &x148, x146, 0x0, x115);
   uint32_t x149;
   fiat_p256_uint1 x150;
-  fiat_p256_addcarryx_u32(&x149, &x150, x148, x139, x115);
+  fiat_p256_addcarryx_u32(&x149, &x150, x148, 0x0, x117);
   uint32_t x151;
   fiat_p256_uint1 x152;
-  fiat_p256_addcarryx_u32(&x151, &x152, x150, (fiat_p256_uint1)x141, x117);
+  fiat_p256_addcarryx_u32(&x151, &x152, x150, x107, x119);
   uint32_t x153;
   fiat_p256_uint1 x154;
-  fiat_p256_addcarryx_u32(&x153, &x154, x152, 0x0, x119);
+  fiat_p256_addcarryx_u32(&x153, &x154, x152, x125, x121);
   uint32_t x155;
   fiat_p256_uint1 x156;
-  fiat_p256_addcarryx_u32(&x155, &x156, x154, x109, x121);
+  fiat_p256_addcarryx_u32(&x155, &x156, x154, x126, x123);
   uint32_t x157;
   fiat_p256_uint1 x158;
-  fiat_p256_addcarryx_u32(&x157, &x158, x156, x127, x123);
+  fiat_p256_addcarryx_u32(&x157, &x158, x156, 0x0, x124);
   uint32_t x159;
-  fiat_p256_uint1 x160;
-  fiat_p256_addcarryx_u32(&x159, &x160, x158, x128, x125);
+  uint32_t x160;
+  fiat_p256_mulx_u32(&x159, &x160, x2, (arg2[7]));
   uint32_t x161;
-  fiat_p256_uint1 x162;
-  fiat_p256_addcarryx_u32(&x161, &x162, x160, 0x0, x126);
+  uint32_t x162;
+  fiat_p256_mulx_u32(&x161, &x162, x2, (arg2[6]));
   uint32_t x163;
   uint32_t x164;
-  fiat_p256_mulx_u32(&x163, &x164, x2, (arg2[7]));
+  fiat_p256_mulx_u32(&x163, &x164, x2, (arg2[5]));
   uint32_t x165;
   uint32_t x166;
-  fiat_p256_mulx_u32(&x165, &x166, x2, (arg2[6]));
+  fiat_p256_mulx_u32(&x165, &x166, x2, (arg2[4]));
   uint32_t x167;
   uint32_t x168;
-  fiat_p256_mulx_u32(&x167, &x168, x2, (arg2[5]));
+  fiat_p256_mulx_u32(&x167, &x168, x2, (arg2[3]));
   uint32_t x169;
   uint32_t x170;
-  fiat_p256_mulx_u32(&x169, &x170, x2, (arg2[4]));
+  fiat_p256_mulx_u32(&x169, &x170, x2, (arg2[2]));
   uint32_t x171;
   uint32_t x172;
-  fiat_p256_mulx_u32(&x171, &x172, x2, (arg2[3]));
+  fiat_p256_mulx_u32(&x171, &x172, x2, (arg2[1]));
   uint32_t x173;
   uint32_t x174;
-  fiat_p256_mulx_u32(&x173, &x174, x2, (arg2[2]));
+  fiat_p256_mulx_u32(&x173, &x174, x2, (arg2[0]));
   uint32_t x175;
-  uint32_t x176;
-  fiat_p256_mulx_u32(&x175, &x176, x2, (arg2[1]));
+  fiat_p256_uint1 x176;
+  fiat_p256_addcarryx_u32(&x175, &x176, 0x0, x171, x174);
   uint32_t x177;
-  uint32_t x178;
-  fiat_p256_mulx_u32(&x177, &x178, x2, (arg2[0]));
+  fiat_p256_uint1 x178;
+  fiat_p256_addcarryx_u32(&x177, &x178, x176, x169, x172);
   uint32_t x179;
   fiat_p256_uint1 x180;
-  fiat_p256_addcarryx_u32(&x179, &x180, 0x0, x175, x178);
+  fiat_p256_addcarryx_u32(&x179, &x180, x178, x167, x170);
   uint32_t x181;
   fiat_p256_uint1 x182;
-  fiat_p256_addcarryx_u32(&x181, &x182, x180, x173, x176);
+  fiat_p256_addcarryx_u32(&x181, &x182, x180, x165, x168);
   uint32_t x183;
   fiat_p256_uint1 x184;
-  fiat_p256_addcarryx_u32(&x183, &x184, x182, x171, x174);
+  fiat_p256_addcarryx_u32(&x183, &x184, x182, x163, x166);
   uint32_t x185;
   fiat_p256_uint1 x186;
-  fiat_p256_addcarryx_u32(&x185, &x186, x184, x169, x172);
+  fiat_p256_addcarryx_u32(&x185, &x186, x184, x161, x164);
   uint32_t x187;
   fiat_p256_uint1 x188;
-  fiat_p256_addcarryx_u32(&x187, &x188, x186, x167, x170);
+  fiat_p256_addcarryx_u32(&x187, &x188, x186, x159, x162);
   uint32_t x189;
   fiat_p256_uint1 x190;
-  fiat_p256_addcarryx_u32(&x189, &x190, x188, x165, x168);
+  fiat_p256_addcarryx_u32(&x189, &x190, x188, 0x0, x160);
   uint32_t x191;
   fiat_p256_uint1 x192;
-  fiat_p256_addcarryx_u32(&x191, &x192, x190, x163, x166);
+  fiat_p256_addcarryx_u32(&x191, &x192, 0x0, x173, x141);
   uint32_t x193;
   fiat_p256_uint1 x194;
-  fiat_p256_addcarryx_u32(&x193, &x194, x192, 0x0, x164);
+  fiat_p256_addcarryx_u32(&x193, &x194, x192, x175, x143);
   uint32_t x195;
   fiat_p256_uint1 x196;
-  fiat_p256_addcarryx_u32(&x195, &x196, 0x0, x177, x145);
+  fiat_p256_addcarryx_u32(&x195, &x196, x194, x177, x145);
   uint32_t x197;
   fiat_p256_uint1 x198;
   fiat_p256_addcarryx_u32(&x197, &x198, x196, x179, x147);
@@ -398,116 +398,116 @@ static void fiat_p256_mul(uint32_t out1[8], const uint32_t arg1[8], const uint32
   fiat_p256_uint1 x208;
   fiat_p256_addcarryx_u32(&x207, &x208, x206, x189, x157);
   uint32_t x209;
-  fiat_p256_uint1 x210;
-  fiat_p256_addcarryx_u32(&x209, &x210, x208, x191, x159);
+  uint32_t x210;
+  fiat_p256_mulx_u32(&x209, &x210, x191, UINT32_C(0xffffffff));
   uint32_t x211;
-  fiat_p256_uint1 x212;
-  fiat_p256_addcarryx_u32(&x211, &x212, x210, x193, x161);
+  uint32_t x212;
+  fiat_p256_mulx_u32(&x211, &x212, x191, UINT32_C(0xffffffff));
   uint32_t x213;
   uint32_t x214;
-  fiat_p256_mulx_u32(&x213, &x214, x195, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u32(&x213, &x214, x191, UINT32_C(0xffffffff));
   uint32_t x215;
   uint32_t x216;
-  fiat_p256_mulx_u32(&x215, &x216, x195, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u32(&x215, &x216, x191, UINT32_C(0xffffffff));
   uint32_t x217;
-  uint32_t x218;
-  fiat_p256_mulx_u32(&x217, &x218, x195, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x218;
+  fiat_p256_addcarryx_u32(&x217, &x218, 0x0, x213, x216);
   uint32_t x219;
-  uint32_t x220;
-  fiat_p256_mulx_u32(&x219, &x220, x195, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x220;
+  fiat_p256_addcarryx_u32(&x219, &x220, x218, x211, x214);
   uint32_t x221;
   fiat_p256_uint1 x222;
-  fiat_p256_addcarryx_u32(&x221, &x222, 0x0, x217, x220);
+  fiat_p256_addcarryx_u32(&x221, &x222, x220, 0x0, x212);
   uint32_t x223;
   fiat_p256_uint1 x224;
-  fiat_p256_addcarryx_u32(&x223, &x224, x222, x215, x218);
+  fiat_p256_addcarryx_u32(&x223, &x224, 0x0, x215, x191);
   uint32_t x225;
   fiat_p256_uint1 x226;
-  fiat_p256_addcarryx_u32(&x225, &x226, x224, 0x0, x216);
+  fiat_p256_addcarryx_u32(&x225, &x226, x224, x217, x193);
   uint32_t x227;
   fiat_p256_uint1 x228;
-  fiat_p256_addcarryx_u32(&x227, &x228, x226, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x227, &x228, x226, x219, x195);
   uint32_t x229;
   fiat_p256_uint1 x230;
-  fiat_p256_addcarryx_u32(&x229, &x230, 0x0, x219, x195);
+  fiat_p256_addcarryx_u32(&x229, &x230, x228, x221, x197);
   uint32_t x231;
   fiat_p256_uint1 x232;
-  fiat_p256_addcarryx_u32(&x231, &x232, x230, x221, x197);
+  fiat_p256_addcarryx_u32(&x231, &x232, x230, 0x0, x199);
   uint32_t x233;
   fiat_p256_uint1 x234;
-  fiat_p256_addcarryx_u32(&x233, &x234, x232, x223, x199);
+  fiat_p256_addcarryx_u32(&x233, &x234, x232, 0x0, x201);
   uint32_t x235;
   fiat_p256_uint1 x236;
-  fiat_p256_addcarryx_u32(&x235, &x236, x234, x225, x201);
+  fiat_p256_addcarryx_u32(&x235, &x236, x234, x191, x203);
   uint32_t x237;
   fiat_p256_uint1 x238;
-  fiat_p256_addcarryx_u32(&x237, &x238, x236, (fiat_p256_uint1)x227, x203);
+  fiat_p256_addcarryx_u32(&x237, &x238, x236, x209, x205);
   uint32_t x239;
   fiat_p256_uint1 x240;
-  fiat_p256_addcarryx_u32(&x239, &x240, x238, 0x0, x205);
+  fiat_p256_addcarryx_u32(&x239, &x240, x238, x210, x207);
   uint32_t x241;
   fiat_p256_uint1 x242;
-  fiat_p256_addcarryx_u32(&x241, &x242, x240, x195, x207);
+  fiat_p256_addcarryx_u32(&x241, &x242, x240, 0x0, x208);
   uint32_t x243;
-  fiat_p256_uint1 x244;
-  fiat_p256_addcarryx_u32(&x243, &x244, x242, x213, x209);
+  uint32_t x244;
+  fiat_p256_mulx_u32(&x243, &x244, x3, (arg2[7]));
   uint32_t x245;
-  fiat_p256_uint1 x246;
-  fiat_p256_addcarryx_u32(&x245, &x246, x244, x214, x211);
+  uint32_t x246;
+  fiat_p256_mulx_u32(&x245, &x246, x3, (arg2[6]));
   uint32_t x247;
-  fiat_p256_uint1 x248;
-  fiat_p256_addcarryx_u32(&x247, &x248, x246, 0x0, x212);
+  uint32_t x248;
+  fiat_p256_mulx_u32(&x247, &x248, x3, (arg2[5]));
   uint32_t x249;
   uint32_t x250;
-  fiat_p256_mulx_u32(&x249, &x250, x3, (arg2[7]));
+  fiat_p256_mulx_u32(&x249, &x250, x3, (arg2[4]));
   uint32_t x251;
   uint32_t x252;
-  fiat_p256_mulx_u32(&x251, &x252, x3, (arg2[6]));
+  fiat_p256_mulx_u32(&x251, &x252, x3, (arg2[3]));
   uint32_t x253;
   uint32_t x254;
-  fiat_p256_mulx_u32(&x253, &x254, x3, (arg2[5]));
+  fiat_p256_mulx_u32(&x253, &x254, x3, (arg2[2]));
   uint32_t x255;
   uint32_t x256;
-  fiat_p256_mulx_u32(&x255, &x256, x3, (arg2[4]));
+  fiat_p256_mulx_u32(&x255, &x256, x3, (arg2[1]));
   uint32_t x257;
   uint32_t x258;
-  fiat_p256_mulx_u32(&x257, &x258, x3, (arg2[3]));
+  fiat_p256_mulx_u32(&x257, &x258, x3, (arg2[0]));
   uint32_t x259;
-  uint32_t x260;
-  fiat_p256_mulx_u32(&x259, &x260, x3, (arg2[2]));
+  fiat_p256_uint1 x260;
+  fiat_p256_addcarryx_u32(&x259, &x260, 0x0, x255, x258);
   uint32_t x261;
-  uint32_t x262;
-  fiat_p256_mulx_u32(&x261, &x262, x3, (arg2[1]));
+  fiat_p256_uint1 x262;
+  fiat_p256_addcarryx_u32(&x261, &x262, x260, x253, x256);
   uint32_t x263;
-  uint32_t x264;
-  fiat_p256_mulx_u32(&x263, &x264, x3, (arg2[0]));
+  fiat_p256_uint1 x264;
+  fiat_p256_addcarryx_u32(&x263, &x264, x262, x251, x254);
   uint32_t x265;
   fiat_p256_uint1 x266;
-  fiat_p256_addcarryx_u32(&x265, &x266, 0x0, x261, x264);
+  fiat_p256_addcarryx_u32(&x265, &x266, x264, x249, x252);
   uint32_t x267;
   fiat_p256_uint1 x268;
-  fiat_p256_addcarryx_u32(&x267, &x268, x266, x259, x262);
+  fiat_p256_addcarryx_u32(&x267, &x268, x266, x247, x250);
   uint32_t x269;
   fiat_p256_uint1 x270;
-  fiat_p256_addcarryx_u32(&x269, &x270, x268, x257, x260);
+  fiat_p256_addcarryx_u32(&x269, &x270, x268, x245, x248);
   uint32_t x271;
   fiat_p256_uint1 x272;
-  fiat_p256_addcarryx_u32(&x271, &x272, x270, x255, x258);
+  fiat_p256_addcarryx_u32(&x271, &x272, x270, x243, x246);
   uint32_t x273;
   fiat_p256_uint1 x274;
-  fiat_p256_addcarryx_u32(&x273, &x274, x272, x253, x256);
+  fiat_p256_addcarryx_u32(&x273, &x274, x272, 0x0, x244);
   uint32_t x275;
   fiat_p256_uint1 x276;
-  fiat_p256_addcarryx_u32(&x275, &x276, x274, x251, x254);
+  fiat_p256_addcarryx_u32(&x275, &x276, 0x0, x257, x225);
   uint32_t x277;
   fiat_p256_uint1 x278;
-  fiat_p256_addcarryx_u32(&x277, &x278, x276, x249, x252);
+  fiat_p256_addcarryx_u32(&x277, &x278, x276, x259, x227);
   uint32_t x279;
   fiat_p256_uint1 x280;
-  fiat_p256_addcarryx_u32(&x279, &x280, x278, 0x0, x250);
+  fiat_p256_addcarryx_u32(&x279, &x280, x278, x261, x229);
   uint32_t x281;
   fiat_p256_uint1 x282;
-  fiat_p256_addcarryx_u32(&x281, &x282, 0x0, x263, x231);
+  fiat_p256_addcarryx_u32(&x281, &x282, x280, x263, x231);
   uint32_t x283;
   fiat_p256_uint1 x284;
   fiat_p256_addcarryx_u32(&x283, &x284, x282, x265, x233);
@@ -524,119 +524,119 @@ static void fiat_p256_mul(uint32_t out1[8], const uint32_t arg1[8], const uint32
   fiat_p256_uint1 x292;
   fiat_p256_addcarryx_u32(&x291, &x292, x290, x273, x241);
   uint32_t x293;
-  fiat_p256_uint1 x294;
-  fiat_p256_addcarryx_u32(&x293, &x294, x292, x275, x243);
+  uint32_t x294;
+  fiat_p256_mulx_u32(&x293, &x294, x275, UINT32_C(0xffffffff));
   uint32_t x295;
-  fiat_p256_uint1 x296;
-  fiat_p256_addcarryx_u32(&x295, &x296, x294, x277, x245);
+  uint32_t x296;
+  fiat_p256_mulx_u32(&x295, &x296, x275, UINT32_C(0xffffffff));
   uint32_t x297;
-  fiat_p256_uint1 x298;
-  fiat_p256_addcarryx_u32(&x297, &x298, x296, x279, x247);
+  uint32_t x298;
+  fiat_p256_mulx_u32(&x297, &x298, x275, UINT32_C(0xffffffff));
   uint32_t x299;
   uint32_t x300;
-  fiat_p256_mulx_u32(&x299, &x300, x281, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u32(&x299, &x300, x275, UINT32_C(0xffffffff));
   uint32_t x301;
-  uint32_t x302;
-  fiat_p256_mulx_u32(&x301, &x302, x281, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x302;
+  fiat_p256_addcarryx_u32(&x301, &x302, 0x0, x297, x300);
   uint32_t x303;
-  uint32_t x304;
-  fiat_p256_mulx_u32(&x303, &x304, x281, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x304;
+  fiat_p256_addcarryx_u32(&x303, &x304, x302, x295, x298);
   uint32_t x305;
-  uint32_t x306;
-  fiat_p256_mulx_u32(&x305, &x306, x281, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x306;
+  fiat_p256_addcarryx_u32(&x305, &x306, x304, 0x0, x296);
   uint32_t x307;
   fiat_p256_uint1 x308;
-  fiat_p256_addcarryx_u32(&x307, &x308, 0x0, x303, x306);
+  fiat_p256_addcarryx_u32(&x307, &x308, 0x0, x299, x275);
   uint32_t x309;
   fiat_p256_uint1 x310;
-  fiat_p256_addcarryx_u32(&x309, &x310, x308, x301, x304);
+  fiat_p256_addcarryx_u32(&x309, &x310, x308, x301, x277);
   uint32_t x311;
   fiat_p256_uint1 x312;
-  fiat_p256_addcarryx_u32(&x311, &x312, x310, 0x0, x302);
+  fiat_p256_addcarryx_u32(&x311, &x312, x310, x303, x279);
   uint32_t x313;
   fiat_p256_uint1 x314;
-  fiat_p256_addcarryx_u32(&x313, &x314, x312, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x313, &x314, x312, x305, x281);
   uint32_t x315;
   fiat_p256_uint1 x316;
-  fiat_p256_addcarryx_u32(&x315, &x316, 0x0, x305, x281);
+  fiat_p256_addcarryx_u32(&x315, &x316, x314, 0x0, x283);
   uint32_t x317;
   fiat_p256_uint1 x318;
-  fiat_p256_addcarryx_u32(&x317, &x318, x316, x307, x283);
+  fiat_p256_addcarryx_u32(&x317, &x318, x316, 0x0, x285);
   uint32_t x319;
   fiat_p256_uint1 x320;
-  fiat_p256_addcarryx_u32(&x319, &x320, x318, x309, x285);
+  fiat_p256_addcarryx_u32(&x319, &x320, x318, x275, x287);
   uint32_t x321;
   fiat_p256_uint1 x322;
-  fiat_p256_addcarryx_u32(&x321, &x322, x320, x311, x287);
+  fiat_p256_addcarryx_u32(&x321, &x322, x320, x293, x289);
   uint32_t x323;
   fiat_p256_uint1 x324;
-  fiat_p256_addcarryx_u32(&x323, &x324, x322, (fiat_p256_uint1)x313, x289);
+  fiat_p256_addcarryx_u32(&x323, &x324, x322, x294, x291);
   uint32_t x325;
   fiat_p256_uint1 x326;
-  fiat_p256_addcarryx_u32(&x325, &x326, x324, 0x0, x291);
+  fiat_p256_addcarryx_u32(&x325, &x326, x324, 0x0, x292);
   uint32_t x327;
-  fiat_p256_uint1 x328;
-  fiat_p256_addcarryx_u32(&x327, &x328, x326, x281, x293);
+  uint32_t x328;
+  fiat_p256_mulx_u32(&x327, &x328, x4, (arg2[7]));
   uint32_t x329;
-  fiat_p256_uint1 x330;
-  fiat_p256_addcarryx_u32(&x329, &x330, x328, x299, x295);
+  uint32_t x330;
+  fiat_p256_mulx_u32(&x329, &x330, x4, (arg2[6]));
   uint32_t x331;
-  fiat_p256_uint1 x332;
-  fiat_p256_addcarryx_u32(&x331, &x332, x330, x300, x297);
+  uint32_t x332;
+  fiat_p256_mulx_u32(&x331, &x332, x4, (arg2[5]));
   uint32_t x333;
-  fiat_p256_uint1 x334;
-  fiat_p256_addcarryx_u32(&x333, &x334, x332, 0x0, x298);
+  uint32_t x334;
+  fiat_p256_mulx_u32(&x333, &x334, x4, (arg2[4]));
   uint32_t x335;
   uint32_t x336;
-  fiat_p256_mulx_u32(&x335, &x336, x4, (arg2[7]));
+  fiat_p256_mulx_u32(&x335, &x336, x4, (arg2[3]));
   uint32_t x337;
   uint32_t x338;
-  fiat_p256_mulx_u32(&x337, &x338, x4, (arg2[6]));
+  fiat_p256_mulx_u32(&x337, &x338, x4, (arg2[2]));
   uint32_t x339;
   uint32_t x340;
-  fiat_p256_mulx_u32(&x339, &x340, x4, (arg2[5]));
+  fiat_p256_mulx_u32(&x339, &x340, x4, (arg2[1]));
   uint32_t x341;
   uint32_t x342;
-  fiat_p256_mulx_u32(&x341, &x342, x4, (arg2[4]));
+  fiat_p256_mulx_u32(&x341, &x342, x4, (arg2[0]));
   uint32_t x343;
-  uint32_t x344;
-  fiat_p256_mulx_u32(&x343, &x344, x4, (arg2[3]));
+  fiat_p256_uint1 x344;
+  fiat_p256_addcarryx_u32(&x343, &x344, 0x0, x339, x342);
   uint32_t x345;
-  uint32_t x346;
-  fiat_p256_mulx_u32(&x345, &x346, x4, (arg2[2]));
+  fiat_p256_uint1 x346;
+  fiat_p256_addcarryx_u32(&x345, &x346, x344, x337, x340);
   uint32_t x347;
-  uint32_t x348;
-  fiat_p256_mulx_u32(&x347, &x348, x4, (arg2[1]));
+  fiat_p256_uint1 x348;
+  fiat_p256_addcarryx_u32(&x347, &x348, x346, x335, x338);
   uint32_t x349;
-  uint32_t x350;
-  fiat_p256_mulx_u32(&x349, &x350, x4, (arg2[0]));
+  fiat_p256_uint1 x350;
+  fiat_p256_addcarryx_u32(&x349, &x350, x348, x333, x336);
   uint32_t x351;
   fiat_p256_uint1 x352;
-  fiat_p256_addcarryx_u32(&x351, &x352, 0x0, x347, x350);
+  fiat_p256_addcarryx_u32(&x351, &x352, x350, x331, x334);
   uint32_t x353;
   fiat_p256_uint1 x354;
-  fiat_p256_addcarryx_u32(&x353, &x354, x352, x345, x348);
+  fiat_p256_addcarryx_u32(&x353, &x354, x352, x329, x332);
   uint32_t x355;
   fiat_p256_uint1 x356;
-  fiat_p256_addcarryx_u32(&x355, &x356, x354, x343, x346);
+  fiat_p256_addcarryx_u32(&x355, &x356, x354, x327, x330);
   uint32_t x357;
   fiat_p256_uint1 x358;
-  fiat_p256_addcarryx_u32(&x357, &x358, x356, x341, x344);
+  fiat_p256_addcarryx_u32(&x357, &x358, x356, 0x0, x328);
   uint32_t x359;
   fiat_p256_uint1 x360;
-  fiat_p256_addcarryx_u32(&x359, &x360, x358, x339, x342);
+  fiat_p256_addcarryx_u32(&x359, &x360, 0x0, x341, x309);
   uint32_t x361;
   fiat_p256_uint1 x362;
-  fiat_p256_addcarryx_u32(&x361, &x362, x360, x337, x340);
+  fiat_p256_addcarryx_u32(&x361, &x362, x360, x343, x311);
   uint32_t x363;
   fiat_p256_uint1 x364;
-  fiat_p256_addcarryx_u32(&x363, &x364, x362, x335, x338);
+  fiat_p256_addcarryx_u32(&x363, &x364, x362, x345, x313);
   uint32_t x365;
   fiat_p256_uint1 x366;
-  fiat_p256_addcarryx_u32(&x365, &x366, x364, 0x0, x336);
+  fiat_p256_addcarryx_u32(&x365, &x366, x364, x347, x315);
   uint32_t x367;
   fiat_p256_uint1 x368;
-  fiat_p256_addcarryx_u32(&x367, &x368, 0x0, x349, x317);
+  fiat_p256_addcarryx_u32(&x367, &x368, x366, x349, x317);
   uint32_t x369;
   fiat_p256_uint1 x370;
   fiat_p256_addcarryx_u32(&x369, &x370, x368, x351, x319);
@@ -650,122 +650,122 @@ static void fiat_p256_mul(uint32_t out1[8], const uint32_t arg1[8], const uint32
   fiat_p256_uint1 x376;
   fiat_p256_addcarryx_u32(&x375, &x376, x374, x357, x325);
   uint32_t x377;
-  fiat_p256_uint1 x378;
-  fiat_p256_addcarryx_u32(&x377, &x378, x376, x359, x327);
+  uint32_t x378;
+  fiat_p256_mulx_u32(&x377, &x378, x359, UINT32_C(0xffffffff));
   uint32_t x379;
-  fiat_p256_uint1 x380;
-  fiat_p256_addcarryx_u32(&x379, &x380, x378, x361, x329);
+  uint32_t x380;
+  fiat_p256_mulx_u32(&x379, &x380, x359, UINT32_C(0xffffffff));
   uint32_t x381;
-  fiat_p256_uint1 x382;
-  fiat_p256_addcarryx_u32(&x381, &x382, x380, x363, x331);
+  uint32_t x382;
+  fiat_p256_mulx_u32(&x381, &x382, x359, UINT32_C(0xffffffff));
   uint32_t x383;
-  fiat_p256_uint1 x384;
-  fiat_p256_addcarryx_u32(&x383, &x384, x382, x365, x333);
+  uint32_t x384;
+  fiat_p256_mulx_u32(&x383, &x384, x359, UINT32_C(0xffffffff));
   uint32_t x385;
-  uint32_t x386;
-  fiat_p256_mulx_u32(&x385, &x386, x367, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x386;
+  fiat_p256_addcarryx_u32(&x385, &x386, 0x0, x381, x384);
   uint32_t x387;
-  uint32_t x388;
-  fiat_p256_mulx_u32(&x387, &x388, x367, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x388;
+  fiat_p256_addcarryx_u32(&x387, &x388, x386, x379, x382);
   uint32_t x389;
-  uint32_t x390;
-  fiat_p256_mulx_u32(&x389, &x390, x367, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x390;
+  fiat_p256_addcarryx_u32(&x389, &x390, x388, 0x0, x380);
   uint32_t x391;
-  uint32_t x392;
-  fiat_p256_mulx_u32(&x391, &x392, x367, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x392;
+  fiat_p256_addcarryx_u32(&x391, &x392, 0x0, x383, x359);
   uint32_t x393;
   fiat_p256_uint1 x394;
-  fiat_p256_addcarryx_u32(&x393, &x394, 0x0, x389, x392);
+  fiat_p256_addcarryx_u32(&x393, &x394, x392, x385, x361);
   uint32_t x395;
   fiat_p256_uint1 x396;
-  fiat_p256_addcarryx_u32(&x395, &x396, x394, x387, x390);
+  fiat_p256_addcarryx_u32(&x395, &x396, x394, x387, x363);
   uint32_t x397;
   fiat_p256_uint1 x398;
-  fiat_p256_addcarryx_u32(&x397, &x398, x396, 0x0, x388);
+  fiat_p256_addcarryx_u32(&x397, &x398, x396, x389, x365);
   uint32_t x399;
   fiat_p256_uint1 x400;
-  fiat_p256_addcarryx_u32(&x399, &x400, x398, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x399, &x400, x398, 0x0, x367);
   uint32_t x401;
   fiat_p256_uint1 x402;
-  fiat_p256_addcarryx_u32(&x401, &x402, 0x0, x391, x367);
+  fiat_p256_addcarryx_u32(&x401, &x402, x400, 0x0, x369);
   uint32_t x403;
   fiat_p256_uint1 x404;
-  fiat_p256_addcarryx_u32(&x403, &x404, x402, x393, x369);
+  fiat_p256_addcarryx_u32(&x403, &x404, x402, x359, x371);
   uint32_t x405;
   fiat_p256_uint1 x406;
-  fiat_p256_addcarryx_u32(&x405, &x406, x404, x395, x371);
+  fiat_p256_addcarryx_u32(&x405, &x406, x404, x377, x373);
   uint32_t x407;
   fiat_p256_uint1 x408;
-  fiat_p256_addcarryx_u32(&x407, &x408, x406, x397, x373);
+  fiat_p256_addcarryx_u32(&x407, &x408, x406, x378, x375);
   uint32_t x409;
   fiat_p256_uint1 x410;
-  fiat_p256_addcarryx_u32(&x409, &x410, x408, (fiat_p256_uint1)x399, x375);
+  fiat_p256_addcarryx_u32(&x409, &x410, x408, 0x0, x376);
   uint32_t x411;
-  fiat_p256_uint1 x412;
-  fiat_p256_addcarryx_u32(&x411, &x412, x410, 0x0, x377);
+  uint32_t x412;
+  fiat_p256_mulx_u32(&x411, &x412, x5, (arg2[7]));
   uint32_t x413;
-  fiat_p256_uint1 x414;
-  fiat_p256_addcarryx_u32(&x413, &x414, x412, x367, x379);
+  uint32_t x414;
+  fiat_p256_mulx_u32(&x413, &x414, x5, (arg2[6]));
   uint32_t x415;
-  fiat_p256_uint1 x416;
-  fiat_p256_addcarryx_u32(&x415, &x416, x414, x385, x381);
+  uint32_t x416;
+  fiat_p256_mulx_u32(&x415, &x416, x5, (arg2[5]));
   uint32_t x417;
-  fiat_p256_uint1 x418;
-  fiat_p256_addcarryx_u32(&x417, &x418, x416, x386, x383);
+  uint32_t x418;
+  fiat_p256_mulx_u32(&x417, &x418, x5, (arg2[4]));
   uint32_t x419;
-  fiat_p256_uint1 x420;
-  fiat_p256_addcarryx_u32(&x419, &x420, x418, 0x0, x384);
+  uint32_t x420;
+  fiat_p256_mulx_u32(&x419, &x420, x5, (arg2[3]));
   uint32_t x421;
   uint32_t x422;
-  fiat_p256_mulx_u32(&x421, &x422, x5, (arg2[7]));
+  fiat_p256_mulx_u32(&x421, &x422, x5, (arg2[2]));
   uint32_t x423;
   uint32_t x424;
-  fiat_p256_mulx_u32(&x423, &x424, x5, (arg2[6]));
+  fiat_p256_mulx_u32(&x423, &x424, x5, (arg2[1]));
   uint32_t x425;
   uint32_t x426;
-  fiat_p256_mulx_u32(&x425, &x426, x5, (arg2[5]));
+  fiat_p256_mulx_u32(&x425, &x426, x5, (arg2[0]));
   uint32_t x427;
-  uint32_t x428;
-  fiat_p256_mulx_u32(&x427, &x428, x5, (arg2[4]));
+  fiat_p256_uint1 x428;
+  fiat_p256_addcarryx_u32(&x427, &x428, 0x0, x423, x426);
   uint32_t x429;
-  uint32_t x430;
-  fiat_p256_mulx_u32(&x429, &x430, x5, (arg2[3]));
+  fiat_p256_uint1 x430;
+  fiat_p256_addcarryx_u32(&x429, &x430, x428, x421, x424);
   uint32_t x431;
-  uint32_t x432;
-  fiat_p256_mulx_u32(&x431, &x432, x5, (arg2[2]));
+  fiat_p256_uint1 x432;
+  fiat_p256_addcarryx_u32(&x431, &x432, x430, x419, x422);
   uint32_t x433;
-  uint32_t x434;
-  fiat_p256_mulx_u32(&x433, &x434, x5, (arg2[1]));
+  fiat_p256_uint1 x434;
+  fiat_p256_addcarryx_u32(&x433, &x434, x432, x417, x420);
   uint32_t x435;
-  uint32_t x436;
-  fiat_p256_mulx_u32(&x435, &x436, x5, (arg2[0]));
+  fiat_p256_uint1 x436;
+  fiat_p256_addcarryx_u32(&x435, &x436, x434, x415, x418);
   uint32_t x437;
   fiat_p256_uint1 x438;
-  fiat_p256_addcarryx_u32(&x437, &x438, 0x0, x433, x436);
+  fiat_p256_addcarryx_u32(&x437, &x438, x436, x413, x416);
   uint32_t x439;
   fiat_p256_uint1 x440;
-  fiat_p256_addcarryx_u32(&x439, &x440, x438, x431, x434);
+  fiat_p256_addcarryx_u32(&x439, &x440, x438, x411, x414);
   uint32_t x441;
   fiat_p256_uint1 x442;
-  fiat_p256_addcarryx_u32(&x441, &x442, x440, x429, x432);
+  fiat_p256_addcarryx_u32(&x441, &x442, x440, 0x0, x412);
   uint32_t x443;
   fiat_p256_uint1 x444;
-  fiat_p256_addcarryx_u32(&x443, &x444, x442, x427, x430);
+  fiat_p256_addcarryx_u32(&x443, &x444, 0x0, x425, x393);
   uint32_t x445;
   fiat_p256_uint1 x446;
-  fiat_p256_addcarryx_u32(&x445, &x446, x444, x425, x428);
+  fiat_p256_addcarryx_u32(&x445, &x446, x444, x427, x395);
   uint32_t x447;
   fiat_p256_uint1 x448;
-  fiat_p256_addcarryx_u32(&x447, &x448, x446, x423, x426);
+  fiat_p256_addcarryx_u32(&x447, &x448, x446, x429, x397);
   uint32_t x449;
   fiat_p256_uint1 x450;
-  fiat_p256_addcarryx_u32(&x449, &x450, x448, x421, x424);
+  fiat_p256_addcarryx_u32(&x449, &x450, x448, x431, x399);
   uint32_t x451;
   fiat_p256_uint1 x452;
-  fiat_p256_addcarryx_u32(&x451, &x452, x450, 0x0, x422);
+  fiat_p256_addcarryx_u32(&x451, &x452, x450, x433, x401);
   uint32_t x453;
   fiat_p256_uint1 x454;
-  fiat_p256_addcarryx_u32(&x453, &x454, 0x0, x435, x403);
+  fiat_p256_addcarryx_u32(&x453, &x454, x452, x435, x403);
   uint32_t x455;
   fiat_p256_uint1 x456;
   fiat_p256_addcarryx_u32(&x455, &x456, x454, x437, x405);
@@ -776,125 +776,125 @@ static void fiat_p256_mul(uint32_t out1[8], const uint32_t arg1[8], const uint32
   fiat_p256_uint1 x460;
   fiat_p256_addcarryx_u32(&x459, &x460, x458, x441, x409);
   uint32_t x461;
-  fiat_p256_uint1 x462;
-  fiat_p256_addcarryx_u32(&x461, &x462, x460, x443, x411);
+  uint32_t x462;
+  fiat_p256_mulx_u32(&x461, &x462, x443, UINT32_C(0xffffffff));
   uint32_t x463;
-  fiat_p256_uint1 x464;
-  fiat_p256_addcarryx_u32(&x463, &x464, x462, x445, x413);
+  uint32_t x464;
+  fiat_p256_mulx_u32(&x463, &x464, x443, UINT32_C(0xffffffff));
   uint32_t x465;
-  fiat_p256_uint1 x466;
-  fiat_p256_addcarryx_u32(&x465, &x466, x464, x447, x415);
+  uint32_t x466;
+  fiat_p256_mulx_u32(&x465, &x466, x443, UINT32_C(0xffffffff));
   uint32_t x467;
-  fiat_p256_uint1 x468;
-  fiat_p256_addcarryx_u32(&x467, &x468, x466, x449, x417);
+  uint32_t x468;
+  fiat_p256_mulx_u32(&x467, &x468, x443, UINT32_C(0xffffffff));
   uint32_t x469;
   fiat_p256_uint1 x470;
-  fiat_p256_addcarryx_u32(&x469, &x470, x468, x451, x419);
+  fiat_p256_addcarryx_u32(&x469, &x470, 0x0, x465, x468);
   uint32_t x471;
-  uint32_t x472;
-  fiat_p256_mulx_u32(&x471, &x472, x453, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x472;
+  fiat_p256_addcarryx_u32(&x471, &x472, x470, x463, x466);
   uint32_t x473;
-  uint32_t x474;
-  fiat_p256_mulx_u32(&x473, &x474, x453, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x474;
+  fiat_p256_addcarryx_u32(&x473, &x474, x472, 0x0, x464);
   uint32_t x475;
-  uint32_t x476;
-  fiat_p256_mulx_u32(&x475, &x476, x453, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x476;
+  fiat_p256_addcarryx_u32(&x475, &x476, 0x0, x467, x443);
   uint32_t x477;
-  uint32_t x478;
-  fiat_p256_mulx_u32(&x477, &x478, x453, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x478;
+  fiat_p256_addcarryx_u32(&x477, &x478, x476, x469, x445);
   uint32_t x479;
   fiat_p256_uint1 x480;
-  fiat_p256_addcarryx_u32(&x479, &x480, 0x0, x475, x478);
+  fiat_p256_addcarryx_u32(&x479, &x480, x478, x471, x447);
   uint32_t x481;
   fiat_p256_uint1 x482;
-  fiat_p256_addcarryx_u32(&x481, &x482, x480, x473, x476);
+  fiat_p256_addcarryx_u32(&x481, &x482, x480, x473, x449);
   uint32_t x483;
   fiat_p256_uint1 x484;
-  fiat_p256_addcarryx_u32(&x483, &x484, x482, 0x0, x474);
+  fiat_p256_addcarryx_u32(&x483, &x484, x482, 0x0, x451);
   uint32_t x485;
   fiat_p256_uint1 x486;
-  fiat_p256_addcarryx_u32(&x485, &x486, x484, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x485, &x486, x484, 0x0, x453);
   uint32_t x487;
   fiat_p256_uint1 x488;
-  fiat_p256_addcarryx_u32(&x487, &x488, 0x0, x477, x453);
+  fiat_p256_addcarryx_u32(&x487, &x488, x486, x443, x455);
   uint32_t x489;
   fiat_p256_uint1 x490;
-  fiat_p256_addcarryx_u32(&x489, &x490, x488, x479, x455);
+  fiat_p256_addcarryx_u32(&x489, &x490, x488, x461, x457);
   uint32_t x491;
   fiat_p256_uint1 x492;
-  fiat_p256_addcarryx_u32(&x491, &x492, x490, x481, x457);
+  fiat_p256_addcarryx_u32(&x491, &x492, x490, x462, x459);
   uint32_t x493;
   fiat_p256_uint1 x494;
-  fiat_p256_addcarryx_u32(&x493, &x494, x492, x483, x459);
+  fiat_p256_addcarryx_u32(&x493, &x494, x492, 0x0, x460);
   uint32_t x495;
-  fiat_p256_uint1 x496;
-  fiat_p256_addcarryx_u32(&x495, &x496, x494, (fiat_p256_uint1)x485, x461);
+  uint32_t x496;
+  fiat_p256_mulx_u32(&x495, &x496, x6, (arg2[7]));
   uint32_t x497;
-  fiat_p256_uint1 x498;
-  fiat_p256_addcarryx_u32(&x497, &x498, x496, 0x0, x463);
+  uint32_t x498;
+  fiat_p256_mulx_u32(&x497, &x498, x6, (arg2[6]));
   uint32_t x499;
-  fiat_p256_uint1 x500;
-  fiat_p256_addcarryx_u32(&x499, &x500, x498, x453, x465);
+  uint32_t x500;
+  fiat_p256_mulx_u32(&x499, &x500, x6, (arg2[5]));
   uint32_t x501;
-  fiat_p256_uint1 x502;
-  fiat_p256_addcarryx_u32(&x501, &x502, x500, x471, x467);
+  uint32_t x502;
+  fiat_p256_mulx_u32(&x501, &x502, x6, (arg2[4]));
   uint32_t x503;
-  fiat_p256_uint1 x504;
-  fiat_p256_addcarryx_u32(&x503, &x504, x502, x472, x469);
+  uint32_t x504;
+  fiat_p256_mulx_u32(&x503, &x504, x6, (arg2[3]));
   uint32_t x505;
-  fiat_p256_uint1 x506;
-  fiat_p256_addcarryx_u32(&x505, &x506, x504, 0x0, x470);
+  uint32_t x506;
+  fiat_p256_mulx_u32(&x505, &x506, x6, (arg2[2]));
   uint32_t x507;
   uint32_t x508;
-  fiat_p256_mulx_u32(&x507, &x508, x6, (arg2[7]));
+  fiat_p256_mulx_u32(&x507, &x508, x6, (arg2[1]));
   uint32_t x509;
   uint32_t x510;
-  fiat_p256_mulx_u32(&x509, &x510, x6, (arg2[6]));
+  fiat_p256_mulx_u32(&x509, &x510, x6, (arg2[0]));
   uint32_t x511;
-  uint32_t x512;
-  fiat_p256_mulx_u32(&x511, &x512, x6, (arg2[5]));
+  fiat_p256_uint1 x512;
+  fiat_p256_addcarryx_u32(&x511, &x512, 0x0, x507, x510);
   uint32_t x513;
-  uint32_t x514;
-  fiat_p256_mulx_u32(&x513, &x514, x6, (arg2[4]));
+  fiat_p256_uint1 x514;
+  fiat_p256_addcarryx_u32(&x513, &x514, x512, x505, x508);
   uint32_t x515;
-  uint32_t x516;
-  fiat_p256_mulx_u32(&x515, &x516, x6, (arg2[3]));
+  fiat_p256_uint1 x516;
+  fiat_p256_addcarryx_u32(&x515, &x516, x514, x503, x506);
   uint32_t x517;
-  uint32_t x518;
-  fiat_p256_mulx_u32(&x517, &x518, x6, (arg2[2]));
+  fiat_p256_uint1 x518;
+  fiat_p256_addcarryx_u32(&x517, &x518, x516, x501, x504);
   uint32_t x519;
-  uint32_t x520;
-  fiat_p256_mulx_u32(&x519, &x520, x6, (arg2[1]));
+  fiat_p256_uint1 x520;
+  fiat_p256_addcarryx_u32(&x519, &x520, x518, x499, x502);
   uint32_t x521;
-  uint32_t x522;
-  fiat_p256_mulx_u32(&x521, &x522, x6, (arg2[0]));
+  fiat_p256_uint1 x522;
+  fiat_p256_addcarryx_u32(&x521, &x522, x520, x497, x500);
   uint32_t x523;
   fiat_p256_uint1 x524;
-  fiat_p256_addcarryx_u32(&x523, &x524, 0x0, x519, x522);
+  fiat_p256_addcarryx_u32(&x523, &x524, x522, x495, x498);
   uint32_t x525;
   fiat_p256_uint1 x526;
-  fiat_p256_addcarryx_u32(&x525, &x526, x524, x517, x520);
+  fiat_p256_addcarryx_u32(&x525, &x526, x524, 0x0, x496);
   uint32_t x527;
   fiat_p256_uint1 x528;
-  fiat_p256_addcarryx_u32(&x527, &x528, x526, x515, x518);
+  fiat_p256_addcarryx_u32(&x527, &x528, 0x0, x509, x477);
   uint32_t x529;
   fiat_p256_uint1 x530;
-  fiat_p256_addcarryx_u32(&x529, &x530, x528, x513, x516);
+  fiat_p256_addcarryx_u32(&x529, &x530, x528, x511, x479);
   uint32_t x531;
   fiat_p256_uint1 x532;
-  fiat_p256_addcarryx_u32(&x531, &x532, x530, x511, x514);
+  fiat_p256_addcarryx_u32(&x531, &x532, x530, x513, x481);
   uint32_t x533;
   fiat_p256_uint1 x534;
-  fiat_p256_addcarryx_u32(&x533, &x534, x532, x509, x512);
+  fiat_p256_addcarryx_u32(&x533, &x534, x532, x515, x483);
   uint32_t x535;
   fiat_p256_uint1 x536;
-  fiat_p256_addcarryx_u32(&x535, &x536, x534, x507, x510);
+  fiat_p256_addcarryx_u32(&x535, &x536, x534, x517, x485);
   uint32_t x537;
   fiat_p256_uint1 x538;
-  fiat_p256_addcarryx_u32(&x537, &x538, x536, 0x0, x508);
+  fiat_p256_addcarryx_u32(&x537, &x538, x536, x519, x487);
   uint32_t x539;
   fiat_p256_uint1 x540;
-  fiat_p256_addcarryx_u32(&x539, &x540, 0x0, x521, x489);
+  fiat_p256_addcarryx_u32(&x539, &x540, x538, x521, x489);
   uint32_t x541;
   fiat_p256_uint1 x542;
   fiat_p256_addcarryx_u32(&x541, &x542, x540, x523, x491);
@@ -902,257 +902,233 @@ static void fiat_p256_mul(uint32_t out1[8], const uint32_t arg1[8], const uint32
   fiat_p256_uint1 x544;
   fiat_p256_addcarryx_u32(&x543, &x544, x542, x525, x493);
   uint32_t x545;
-  fiat_p256_uint1 x546;
-  fiat_p256_addcarryx_u32(&x545, &x546, x544, x527, x495);
+  uint32_t x546;
+  fiat_p256_mulx_u32(&x545, &x546, x527, UINT32_C(0xffffffff));
   uint32_t x547;
-  fiat_p256_uint1 x548;
-  fiat_p256_addcarryx_u32(&x547, &x548, x546, x529, x497);
+  uint32_t x548;
+  fiat_p256_mulx_u32(&x547, &x548, x527, UINT32_C(0xffffffff));
   uint32_t x549;
-  fiat_p256_uint1 x550;
-  fiat_p256_addcarryx_u32(&x549, &x550, x548, x531, x499);
+  uint32_t x550;
+  fiat_p256_mulx_u32(&x549, &x550, x527, UINT32_C(0xffffffff));
   uint32_t x551;
-  fiat_p256_uint1 x552;
-  fiat_p256_addcarryx_u32(&x551, &x552, x550, x533, x501);
+  uint32_t x552;
+  fiat_p256_mulx_u32(&x551, &x552, x527, UINT32_C(0xffffffff));
   uint32_t x553;
   fiat_p256_uint1 x554;
-  fiat_p256_addcarryx_u32(&x553, &x554, x552, x535, x503);
+  fiat_p256_addcarryx_u32(&x553, &x554, 0x0, x549, x552);
   uint32_t x555;
   fiat_p256_uint1 x556;
-  fiat_p256_addcarryx_u32(&x555, &x556, x554, x537, x505);
+  fiat_p256_addcarryx_u32(&x555, &x556, x554, x547, x550);
   uint32_t x557;
-  uint32_t x558;
-  fiat_p256_mulx_u32(&x557, &x558, x539, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x558;
+  fiat_p256_addcarryx_u32(&x557, &x558, x556, 0x0, x548);
   uint32_t x559;
-  uint32_t x560;
-  fiat_p256_mulx_u32(&x559, &x560, x539, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x560;
+  fiat_p256_addcarryx_u32(&x559, &x560, 0x0, x551, x527);
   uint32_t x561;
-  uint32_t x562;
-  fiat_p256_mulx_u32(&x561, &x562, x539, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x562;
+  fiat_p256_addcarryx_u32(&x561, &x562, x560, x553, x529);
   uint32_t x563;
-  uint32_t x564;
-  fiat_p256_mulx_u32(&x563, &x564, x539, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x564;
+  fiat_p256_addcarryx_u32(&x563, &x564, x562, x555, x531);
   uint32_t x565;
   fiat_p256_uint1 x566;
-  fiat_p256_addcarryx_u32(&x565, &x566, 0x0, x561, x564);
+  fiat_p256_addcarryx_u32(&x565, &x566, x564, x557, x533);
   uint32_t x567;
   fiat_p256_uint1 x568;
-  fiat_p256_addcarryx_u32(&x567, &x568, x566, x559, x562);
+  fiat_p256_addcarryx_u32(&x567, &x568, x566, 0x0, x535);
   uint32_t x569;
   fiat_p256_uint1 x570;
-  fiat_p256_addcarryx_u32(&x569, &x570, x568, 0x0, x560);
+  fiat_p256_addcarryx_u32(&x569, &x570, x568, 0x0, x537);
   uint32_t x571;
   fiat_p256_uint1 x572;
-  fiat_p256_addcarryx_u32(&x571, &x572, x570, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x571, &x572, x570, x527, x539);
   uint32_t x573;
   fiat_p256_uint1 x574;
-  fiat_p256_addcarryx_u32(&x573, &x574, 0x0, x563, x539);
+  fiat_p256_addcarryx_u32(&x573, &x574, x572, x545, x541);
   uint32_t x575;
   fiat_p256_uint1 x576;
-  fiat_p256_addcarryx_u32(&x575, &x576, x574, x565, x541);
+  fiat_p256_addcarryx_u32(&x575, &x576, x574, x546, x543);
   uint32_t x577;
   fiat_p256_uint1 x578;
-  fiat_p256_addcarryx_u32(&x577, &x578, x576, x567, x543);
+  fiat_p256_addcarryx_u32(&x577, &x578, x576, 0x0, x544);
   uint32_t x579;
-  fiat_p256_uint1 x580;
-  fiat_p256_addcarryx_u32(&x579, &x580, x578, x569, x545);
+  uint32_t x580;
+  fiat_p256_mulx_u32(&x579, &x580, x7, (arg2[7]));
   uint32_t x581;
-  fiat_p256_uint1 x582;
-  fiat_p256_addcarryx_u32(&x581, &x582, x580, (fiat_p256_uint1)x571, x547);
+  uint32_t x582;
+  fiat_p256_mulx_u32(&x581, &x582, x7, (arg2[6]));
   uint32_t x583;
-  fiat_p256_uint1 x584;
-  fiat_p256_addcarryx_u32(&x583, &x584, x582, 0x0, x549);
+  uint32_t x584;
+  fiat_p256_mulx_u32(&x583, &x584, x7, (arg2[5]));
   uint32_t x585;
-  fiat_p256_uint1 x586;
-  fiat_p256_addcarryx_u32(&x585, &x586, x584, x539, x551);
+  uint32_t x586;
+  fiat_p256_mulx_u32(&x585, &x586, x7, (arg2[4]));
   uint32_t x587;
-  fiat_p256_uint1 x588;
-  fiat_p256_addcarryx_u32(&x587, &x588, x586, x557, x553);
+  uint32_t x588;
+  fiat_p256_mulx_u32(&x587, &x588, x7, (arg2[3]));
   uint32_t x589;
-  fiat_p256_uint1 x590;
-  fiat_p256_addcarryx_u32(&x589, &x590, x588, x558, x555);
+  uint32_t x590;
+  fiat_p256_mulx_u32(&x589, &x590, x7, (arg2[2]));
   uint32_t x591;
-  fiat_p256_uint1 x592;
-  fiat_p256_addcarryx_u32(&x591, &x592, x590, 0x0, x556);
+  uint32_t x592;
+  fiat_p256_mulx_u32(&x591, &x592, x7, (arg2[1]));
   uint32_t x593;
   uint32_t x594;
-  fiat_p256_mulx_u32(&x593, &x594, x7, (arg2[7]));
+  fiat_p256_mulx_u32(&x593, &x594, x7, (arg2[0]));
   uint32_t x595;
-  uint32_t x596;
-  fiat_p256_mulx_u32(&x595, &x596, x7, (arg2[6]));
+  fiat_p256_uint1 x596;
+  fiat_p256_addcarryx_u32(&x595, &x596, 0x0, x591, x594);
   uint32_t x597;
-  uint32_t x598;
-  fiat_p256_mulx_u32(&x597, &x598, x7, (arg2[5]));
+  fiat_p256_uint1 x598;
+  fiat_p256_addcarryx_u32(&x597, &x598, x596, x589, x592);
   uint32_t x599;
-  uint32_t x600;
-  fiat_p256_mulx_u32(&x599, &x600, x7, (arg2[4]));
+  fiat_p256_uint1 x600;
+  fiat_p256_addcarryx_u32(&x599, &x600, x598, x587, x590);
   uint32_t x601;
-  uint32_t x602;
-  fiat_p256_mulx_u32(&x601, &x602, x7, (arg2[3]));
+  fiat_p256_uint1 x602;
+  fiat_p256_addcarryx_u32(&x601, &x602, x600, x585, x588);
   uint32_t x603;
-  uint32_t x604;
-  fiat_p256_mulx_u32(&x603, &x604, x7, (arg2[2]));
+  fiat_p256_uint1 x604;
+  fiat_p256_addcarryx_u32(&x603, &x604, x602, x583, x586);
   uint32_t x605;
-  uint32_t x606;
-  fiat_p256_mulx_u32(&x605, &x606, x7, (arg2[1]));
+  fiat_p256_uint1 x606;
+  fiat_p256_addcarryx_u32(&x605, &x606, x604, x581, x584);
   uint32_t x607;
-  uint32_t x608;
-  fiat_p256_mulx_u32(&x607, &x608, x7, (arg2[0]));
+  fiat_p256_uint1 x608;
+  fiat_p256_addcarryx_u32(&x607, &x608, x606, x579, x582);
   uint32_t x609;
   fiat_p256_uint1 x610;
-  fiat_p256_addcarryx_u32(&x609, &x610, 0x0, x605, x608);
+  fiat_p256_addcarryx_u32(&x609, &x610, x608, 0x0, x580);
   uint32_t x611;
   fiat_p256_uint1 x612;
-  fiat_p256_addcarryx_u32(&x611, &x612, x610, x603, x606);
+  fiat_p256_addcarryx_u32(&x611, &x612, 0x0, x593, x561);
   uint32_t x613;
   fiat_p256_uint1 x614;
-  fiat_p256_addcarryx_u32(&x613, &x614, x612, x601, x604);
+  fiat_p256_addcarryx_u32(&x613, &x614, x612, x595, x563);
   uint32_t x615;
   fiat_p256_uint1 x616;
-  fiat_p256_addcarryx_u32(&x615, &x616, x614, x599, x602);
+  fiat_p256_addcarryx_u32(&x615, &x616, x614, x597, x565);
   uint32_t x617;
   fiat_p256_uint1 x618;
-  fiat_p256_addcarryx_u32(&x617, &x618, x616, x597, x600);
+  fiat_p256_addcarryx_u32(&x617, &x618, x616, x599, x567);
   uint32_t x619;
   fiat_p256_uint1 x620;
-  fiat_p256_addcarryx_u32(&x619, &x620, x618, x595, x598);
+  fiat_p256_addcarryx_u32(&x619, &x620, x618, x601, x569);
   uint32_t x621;
   fiat_p256_uint1 x622;
-  fiat_p256_addcarryx_u32(&x621, &x622, x620, x593, x596);
+  fiat_p256_addcarryx_u32(&x621, &x622, x620, x603, x571);
   uint32_t x623;
   fiat_p256_uint1 x624;
-  fiat_p256_addcarryx_u32(&x623, &x624, x622, 0x0, x594);
+  fiat_p256_addcarryx_u32(&x623, &x624, x622, x605, x573);
   uint32_t x625;
   fiat_p256_uint1 x626;
-  fiat_p256_addcarryx_u32(&x625, &x626, 0x0, x607, x575);
+  fiat_p256_addcarryx_u32(&x625, &x626, x624, x607, x575);
   uint32_t x627;
   fiat_p256_uint1 x628;
   fiat_p256_addcarryx_u32(&x627, &x628, x626, x609, x577);
   uint32_t x629;
-  fiat_p256_uint1 x630;
-  fiat_p256_addcarryx_u32(&x629, &x630, x628, x611, x579);
+  uint32_t x630;
+  fiat_p256_mulx_u32(&x629, &x630, x611, UINT32_C(0xffffffff));
   uint32_t x631;
-  fiat_p256_uint1 x632;
-  fiat_p256_addcarryx_u32(&x631, &x632, x630, x613, x581);
+  uint32_t x632;
+  fiat_p256_mulx_u32(&x631, &x632, x611, UINT32_C(0xffffffff));
   uint32_t x633;
-  fiat_p256_uint1 x634;
-  fiat_p256_addcarryx_u32(&x633, &x634, x632, x615, x583);
+  uint32_t x634;
+  fiat_p256_mulx_u32(&x633, &x634, x611, UINT32_C(0xffffffff));
   uint32_t x635;
-  fiat_p256_uint1 x636;
-  fiat_p256_addcarryx_u32(&x635, &x636, x634, x617, x585);
+  uint32_t x636;
+  fiat_p256_mulx_u32(&x635, &x636, x611, UINT32_C(0xffffffff));
   uint32_t x637;
   fiat_p256_uint1 x638;
-  fiat_p256_addcarryx_u32(&x637, &x638, x636, x619, x587);
+  fiat_p256_addcarryx_u32(&x637, &x638, 0x0, x633, x636);
   uint32_t x639;
   fiat_p256_uint1 x640;
-  fiat_p256_addcarryx_u32(&x639, &x640, x638, x621, x589);
+  fiat_p256_addcarryx_u32(&x639, &x640, x638, x631, x634);
   uint32_t x641;
   fiat_p256_uint1 x642;
-  fiat_p256_addcarryx_u32(&x641, &x642, x640, x623, x591);
+  fiat_p256_addcarryx_u32(&x641, &x642, x640, 0x0, x632);
   uint32_t x643;
-  uint32_t x644;
-  fiat_p256_mulx_u32(&x643, &x644, x625, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x644;
+  fiat_p256_addcarryx_u32(&x643, &x644, 0x0, x635, x611);
   uint32_t x645;
-  uint32_t x646;
-  fiat_p256_mulx_u32(&x645, &x646, x625, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x646;
+  fiat_p256_addcarryx_u32(&x645, &x646, x644, x637, x613);
   uint32_t x647;
-  uint32_t x648;
-  fiat_p256_mulx_u32(&x647, &x648, x625, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x648;
+  fiat_p256_addcarryx_u32(&x647, &x648, x646, x639, x615);
   uint32_t x649;
-  uint32_t x650;
-  fiat_p256_mulx_u32(&x649, &x650, x625, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x650;
+  fiat_p256_addcarryx_u32(&x649, &x650, x648, x641, x617);
   uint32_t x651;
   fiat_p256_uint1 x652;
-  fiat_p256_addcarryx_u32(&x651, &x652, 0x0, x647, x650);
+  fiat_p256_addcarryx_u32(&x651, &x652, x650, 0x0, x619);
   uint32_t x653;
   fiat_p256_uint1 x654;
-  fiat_p256_addcarryx_u32(&x653, &x654, x652, x645, x648);
+  fiat_p256_addcarryx_u32(&x653, &x654, x652, 0x0, x621);
   uint32_t x655;
   fiat_p256_uint1 x656;
-  fiat_p256_addcarryx_u32(&x655, &x656, x654, 0x0, x646);
+  fiat_p256_addcarryx_u32(&x655, &x656, x654, x611, x623);
   uint32_t x657;
   fiat_p256_uint1 x658;
-  fiat_p256_addcarryx_u32(&x657, &x658, x656, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x657, &x658, x656, x629, x625);
   uint32_t x659;
   fiat_p256_uint1 x660;
-  fiat_p256_addcarryx_u32(&x659, &x660, 0x0, x649, x625);
+  fiat_p256_addcarryx_u32(&x659, &x660, x658, x630, x627);
   uint32_t x661;
   fiat_p256_uint1 x662;
-  fiat_p256_addcarryx_u32(&x661, &x662, x660, x651, x627);
+  fiat_p256_addcarryx_u32(&x661, &x662, x660, 0x0, x628);
   uint32_t x663;
   fiat_p256_uint1 x664;
-  fiat_p256_addcarryx_u32(&x663, &x664, x662, x653, x629);
+  fiat_p256_subborrowx_u32(&x663, &x664, 0x0, x645, UINT32_C(0xffffffff));
   uint32_t x665;
   fiat_p256_uint1 x666;
-  fiat_p256_addcarryx_u32(&x665, &x666, x664, x655, x631);
+  fiat_p256_subborrowx_u32(&x665, &x666, x664, x647, UINT32_C(0xffffffff));
   uint32_t x667;
   fiat_p256_uint1 x668;
-  fiat_p256_addcarryx_u32(&x667, &x668, x666, (fiat_p256_uint1)x657, x633);
+  fiat_p256_subborrowx_u32(&x667, &x668, x666, x649, UINT32_C(0xffffffff));
   uint32_t x669;
   fiat_p256_uint1 x670;
-  fiat_p256_addcarryx_u32(&x669, &x670, x668, 0x0, x635);
+  fiat_p256_subborrowx_u32(&x669, &x670, x668, x651, 0x0);
   uint32_t x671;
   fiat_p256_uint1 x672;
-  fiat_p256_addcarryx_u32(&x671, &x672, x670, x625, x637);
+  fiat_p256_subborrowx_u32(&x671, &x672, x670, x653, 0x0);
   uint32_t x673;
   fiat_p256_uint1 x674;
-  fiat_p256_addcarryx_u32(&x673, &x674, x672, x643, x639);
+  fiat_p256_subborrowx_u32(&x673, &x674, x672, x655, 0x0);
   uint32_t x675;
   fiat_p256_uint1 x676;
-  fiat_p256_addcarryx_u32(&x675, &x676, x674, x644, x641);
+  fiat_p256_subborrowx_u32(&x675, &x676, x674, x657, 0x1);
   uint32_t x677;
   fiat_p256_uint1 x678;
-  fiat_p256_addcarryx_u32(&x677, &x678, x676, 0x0, x642);
+  fiat_p256_subborrowx_u32(&x677, &x678, x676, x659, UINT32_C(0xffffffff));
   uint32_t x679;
   fiat_p256_uint1 x680;
-  fiat_p256_subborrowx_u32(&x679, &x680, 0x0, x661, UINT32_C(0xffffffff));
+  fiat_p256_subborrowx_u32(&x679, &x680, x678, x661, 0x0);
   uint32_t x681;
-  fiat_p256_uint1 x682;
-  fiat_p256_subborrowx_u32(&x681, &x682, x680, x663, UINT32_C(0xffffffff));
+  fiat_p256_cmovznz_u32(&x681, x680, x663, x645);
+  uint32_t x682;
+  fiat_p256_cmovznz_u32(&x682, x680, x665, x647);
   uint32_t x683;
-  fiat_p256_uint1 x684;
-  fiat_p256_subborrowx_u32(&x683, &x684, x682, x665, UINT32_C(0xffffffff));
+  fiat_p256_cmovznz_u32(&x683, x680, x667, x649);
+  uint32_t x684;
+  fiat_p256_cmovznz_u32(&x684, x680, x669, x651);
   uint32_t x685;
-  fiat_p256_uint1 x686;
-  fiat_p256_subborrowx_u32(&x685, &x686, x684, x667, 0x0);
+  fiat_p256_cmovznz_u32(&x685, x680, x671, x653);
+  uint32_t x686;
+  fiat_p256_cmovznz_u32(&x686, x680, x673, x655);
   uint32_t x687;
-  fiat_p256_uint1 x688;
-  fiat_p256_subborrowx_u32(&x687, &x688, x686, x669, 0x0);
-  uint32_t x689;
-  fiat_p256_uint1 x690;
-  fiat_p256_subborrowx_u32(&x689, &x690, x688, x671, 0x0);
-  uint32_t x691;
-  fiat_p256_uint1 x692;
-  fiat_p256_subborrowx_u32(&x691, &x692, x690, x673, 0x1);
-  uint32_t x693;
-  fiat_p256_uint1 x694;
-  fiat_p256_subborrowx_u32(&x693, &x694, x692, x675, UINT32_C(0xffffffff));
-  uint32_t x695;
-  fiat_p256_uint1 x696;
-  fiat_p256_subborrowx_u32(&x695, &x696, x694, x677, 0x0);
-  uint32_t x697;
-  fiat_p256_cmovznz_u32(&x697, x696, x679, x661);
-  uint32_t x698;
-  fiat_p256_cmovznz_u32(&x698, x696, x681, x663);
-  uint32_t x699;
-  fiat_p256_cmovznz_u32(&x699, x696, x683, x665);
-  uint32_t x700;
-  fiat_p256_cmovznz_u32(&x700, x696, x685, x667);
-  uint32_t x701;
-  fiat_p256_cmovznz_u32(&x701, x696, x687, x669);
-  uint32_t x702;
-  fiat_p256_cmovznz_u32(&x702, x696, x689, x671);
-  uint32_t x703;
-  fiat_p256_cmovznz_u32(&x703, x696, x691, x673);
-  uint32_t x704;
-  fiat_p256_cmovznz_u32(&x704, x696, x693, x675);
-  out1[0] = x697;
-  out1[1] = x698;
-  out1[2] = x699;
-  out1[3] = x700;
-  out1[4] = x701;
-  out1[5] = x702;
-  out1[6] = x703;
-  out1[7] = x704;
+  fiat_p256_cmovznz_u32(&x687, x680, x675, x657);
+  uint32_t x688;
+  fiat_p256_cmovznz_u32(&x688, x680, x677, x659);
+  out1[0] = x681;
+  out1[1] = x682;
+  out1[2] = x683;
+  out1[3] = x684;
+  out1[4] = x685;
+  out1[5] = x686;
+  out1[6] = x687;
+  out1[7] = x688;
 }
 
 /*
@@ -1241,88 +1217,88 @@ static void fiat_p256_square(uint32_t out1[8], const uint32_t arg1[8]) {
   fiat_p256_addcarryx_u32(&x53, &x54, x52, 0x0, x44);
   uint32_t x55;
   fiat_p256_uint1 x56;
-  fiat_p256_addcarryx_u32(&x55, &x56, x54, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x55, &x56, 0x0, x47, x23);
   uint32_t x57;
   fiat_p256_uint1 x58;
-  fiat_p256_addcarryx_u32(&x57, &x58, 0x0, x47, x23);
+  fiat_p256_addcarryx_u32(&x57, &x58, x56, x49, x25);
   uint32_t x59;
   fiat_p256_uint1 x60;
-  fiat_p256_addcarryx_u32(&x59, &x60, x58, x49, x25);
+  fiat_p256_addcarryx_u32(&x59, &x60, x58, x51, x27);
   uint32_t x61;
   fiat_p256_uint1 x62;
-  fiat_p256_addcarryx_u32(&x61, &x62, x60, x51, x27);
+  fiat_p256_addcarryx_u32(&x61, &x62, x60, x53, x29);
   uint32_t x63;
   fiat_p256_uint1 x64;
-  fiat_p256_addcarryx_u32(&x63, &x64, x62, x53, x29);
+  fiat_p256_addcarryx_u32(&x63, &x64, x62, 0x0, x31);
   uint32_t x65;
   fiat_p256_uint1 x66;
-  fiat_p256_addcarryx_u32(&x65, &x66, x64, (fiat_p256_uint1)x55, x31);
+  fiat_p256_addcarryx_u32(&x65, &x66, x64, 0x0, x33);
   uint32_t x67;
   fiat_p256_uint1 x68;
-  fiat_p256_addcarryx_u32(&x67, &x68, x66, 0x0, x33);
+  fiat_p256_addcarryx_u32(&x67, &x68, x66, x23, x35);
   uint32_t x69;
   fiat_p256_uint1 x70;
-  fiat_p256_addcarryx_u32(&x69, &x70, x68, x23, x35);
+  fiat_p256_addcarryx_u32(&x69, &x70, x68, x41, x37);
   uint32_t x71;
   fiat_p256_uint1 x72;
-  fiat_p256_addcarryx_u32(&x71, &x72, x70, x41, x37);
+  fiat_p256_addcarryx_u32(&x71, &x72, x70, x42, x39);
   uint32_t x73;
   fiat_p256_uint1 x74;
-  fiat_p256_addcarryx_u32(&x73, &x74, x72, x42, x39);
+  fiat_p256_addcarryx_u32(&x73, &x74, x72, 0x0, 0x0);
   uint32_t x75;
-  fiat_p256_uint1 x76;
-  fiat_p256_addcarryx_u32(&x75, &x76, x74, 0x0, 0x0);
+  uint32_t x76;
+  fiat_p256_mulx_u32(&x75, &x76, x1, (arg1[7]));
   uint32_t x77;
   uint32_t x78;
-  fiat_p256_mulx_u32(&x77, &x78, x1, (arg1[7]));
+  fiat_p256_mulx_u32(&x77, &x78, x1, (arg1[6]));
   uint32_t x79;
   uint32_t x80;
-  fiat_p256_mulx_u32(&x79, &x80, x1, (arg1[6]));
+  fiat_p256_mulx_u32(&x79, &x80, x1, (arg1[5]));
   uint32_t x81;
   uint32_t x82;
-  fiat_p256_mulx_u32(&x81, &x82, x1, (arg1[5]));
+  fiat_p256_mulx_u32(&x81, &x82, x1, (arg1[4]));
   uint32_t x83;
   uint32_t x84;
-  fiat_p256_mulx_u32(&x83, &x84, x1, (arg1[4]));
+  fiat_p256_mulx_u32(&x83, &x84, x1, (arg1[3]));
   uint32_t x85;
   uint32_t x86;
-  fiat_p256_mulx_u32(&x85, &x86, x1, (arg1[3]));
+  fiat_p256_mulx_u32(&x85, &x86, x1, (arg1[2]));
   uint32_t x87;
   uint32_t x88;
-  fiat_p256_mulx_u32(&x87, &x88, x1, (arg1[2]));
+  fiat_p256_mulx_u32(&x87, &x88, x1, (arg1[1]));
   uint32_t x89;
   uint32_t x90;
-  fiat_p256_mulx_u32(&x89, &x90, x1, (arg1[1]));
+  fiat_p256_mulx_u32(&x89, &x90, x1, (arg1[0]));
   uint32_t x91;
-  uint32_t x92;
-  fiat_p256_mulx_u32(&x91, &x92, x1, (arg1[0]));
+  fiat_p256_uint1 x92;
+  fiat_p256_addcarryx_u32(&x91, &x92, 0x0, x87, x90);
   uint32_t x93;
   fiat_p256_uint1 x94;
-  fiat_p256_addcarryx_u32(&x93, &x94, 0x0, x89, x92);
+  fiat_p256_addcarryx_u32(&x93, &x94, x92, x85, x88);
   uint32_t x95;
   fiat_p256_uint1 x96;
-  fiat_p256_addcarryx_u32(&x95, &x96, x94, x87, x90);
+  fiat_p256_addcarryx_u32(&x95, &x96, x94, x83, x86);
   uint32_t x97;
   fiat_p256_uint1 x98;
-  fiat_p256_addcarryx_u32(&x97, &x98, x96, x85, x88);
+  fiat_p256_addcarryx_u32(&x97, &x98, x96, x81, x84);
   uint32_t x99;
   fiat_p256_uint1 x100;
-  fiat_p256_addcarryx_u32(&x99, &x100, x98, x83, x86);
+  fiat_p256_addcarryx_u32(&x99, &x100, x98, x79, x82);
   uint32_t x101;
   fiat_p256_uint1 x102;
-  fiat_p256_addcarryx_u32(&x101, &x102, x100, x81, x84);
+  fiat_p256_addcarryx_u32(&x101, &x102, x100, x77, x80);
   uint32_t x103;
   fiat_p256_uint1 x104;
-  fiat_p256_addcarryx_u32(&x103, &x104, x102, x79, x82);
+  fiat_p256_addcarryx_u32(&x103, &x104, x102, x75, x78);
   uint32_t x105;
   fiat_p256_uint1 x106;
-  fiat_p256_addcarryx_u32(&x105, &x106, x104, x77, x80);
+  fiat_p256_addcarryx_u32(&x105, &x106, x104, 0x0, x76);
   uint32_t x107;
   fiat_p256_uint1 x108;
-  fiat_p256_addcarryx_u32(&x107, &x108, x106, 0x0, x78);
+  fiat_p256_addcarryx_u32(&x107, &x108, 0x0, x89, x57);
   uint32_t x109;
   fiat_p256_uint1 x110;
-  fiat_p256_addcarryx_u32(&x109, &x110, 0x0, x91, x59);
+  fiat_p256_addcarryx_u32(&x109, &x110, x108, x91, x59);
   uint32_t x111;
   fiat_p256_uint1 x112;
   fiat_p256_addcarryx_u32(&x111, &x112, x110, x93, x61);
@@ -1343,115 +1319,115 @@ static void fiat_p256_square(uint32_t out1[8], const uint32_t arg1[8]) {
   fiat_p256_addcarryx_u32(&x121, &x122, x120, x103, x71);
   uint32_t x123;
   fiat_p256_uint1 x124;
-  fiat_p256_addcarryx_u32(&x123, &x124, x122, x105, x73);
+  fiat_p256_addcarryx_u32(&x123, &x124, x122, x105, (fiat_p256_uint1)x73);
   uint32_t x125;
-  fiat_p256_uint1 x126;
-  fiat_p256_addcarryx_u32(&x125, &x126, x124, x107, (fiat_p256_uint1)x75);
+  uint32_t x126;
+  fiat_p256_mulx_u32(&x125, &x126, x107, UINT32_C(0xffffffff));
   uint32_t x127;
   uint32_t x128;
-  fiat_p256_mulx_u32(&x127, &x128, x109, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u32(&x127, &x128, x107, UINT32_C(0xffffffff));
   uint32_t x129;
   uint32_t x130;
-  fiat_p256_mulx_u32(&x129, &x130, x109, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u32(&x129, &x130, x107, UINT32_C(0xffffffff));
   uint32_t x131;
   uint32_t x132;
-  fiat_p256_mulx_u32(&x131, &x132, x109, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u32(&x131, &x132, x107, UINT32_C(0xffffffff));
   uint32_t x133;
-  uint32_t x134;
-  fiat_p256_mulx_u32(&x133, &x134, x109, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x134;
+  fiat_p256_addcarryx_u32(&x133, &x134, 0x0, x129, x132);
   uint32_t x135;
   fiat_p256_uint1 x136;
-  fiat_p256_addcarryx_u32(&x135, &x136, 0x0, x131, x134);
+  fiat_p256_addcarryx_u32(&x135, &x136, x134, x127, x130);
   uint32_t x137;
   fiat_p256_uint1 x138;
-  fiat_p256_addcarryx_u32(&x137, &x138, x136, x129, x132);
+  fiat_p256_addcarryx_u32(&x137, &x138, x136, 0x0, x128);
   uint32_t x139;
   fiat_p256_uint1 x140;
-  fiat_p256_addcarryx_u32(&x139, &x140, x138, 0x0, x130);
+  fiat_p256_addcarryx_u32(&x139, &x140, 0x0, x131, x107);
   uint32_t x141;
   fiat_p256_uint1 x142;
-  fiat_p256_addcarryx_u32(&x141, &x142, x140, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x141, &x142, x140, x133, x109);
   uint32_t x143;
   fiat_p256_uint1 x144;
-  fiat_p256_addcarryx_u32(&x143, &x144, 0x0, x133, x109);
+  fiat_p256_addcarryx_u32(&x143, &x144, x142, x135, x111);
   uint32_t x145;
   fiat_p256_uint1 x146;
-  fiat_p256_addcarryx_u32(&x145, &x146, x144, x135, x111);
+  fiat_p256_addcarryx_u32(&x145, &x146, x144, x137, x113);
   uint32_t x147;
   fiat_p256_uint1 x148;
-  fiat_p256_addcarryx_u32(&x147, &x148, x146, x137, x113);
+  fiat_p256_addcarryx_u32(&x147, &x148, x146, 0x0, x115);
   uint32_t x149;
   fiat_p256_uint1 x150;
-  fiat_p256_addcarryx_u32(&x149, &x150, x148, x139, x115);
+  fiat_p256_addcarryx_u32(&x149, &x150, x148, 0x0, x117);
   uint32_t x151;
   fiat_p256_uint1 x152;
-  fiat_p256_addcarryx_u32(&x151, &x152, x150, (fiat_p256_uint1)x141, x117);
+  fiat_p256_addcarryx_u32(&x151, &x152, x150, x107, x119);
   uint32_t x153;
   fiat_p256_uint1 x154;
-  fiat_p256_addcarryx_u32(&x153, &x154, x152, 0x0, x119);
+  fiat_p256_addcarryx_u32(&x153, &x154, x152, x125, x121);
   uint32_t x155;
   fiat_p256_uint1 x156;
-  fiat_p256_addcarryx_u32(&x155, &x156, x154, x109, x121);
+  fiat_p256_addcarryx_u32(&x155, &x156, x154, x126, x123);
   uint32_t x157;
   fiat_p256_uint1 x158;
-  fiat_p256_addcarryx_u32(&x157, &x158, x156, x127, x123);
+  fiat_p256_addcarryx_u32(&x157, &x158, x156, 0x0, x124);
   uint32_t x159;
-  fiat_p256_uint1 x160;
-  fiat_p256_addcarryx_u32(&x159, &x160, x158, x128, x125);
+  uint32_t x160;
+  fiat_p256_mulx_u32(&x159, &x160, x2, (arg1[7]));
   uint32_t x161;
-  fiat_p256_uint1 x162;
-  fiat_p256_addcarryx_u32(&x161, &x162, x160, 0x0, x126);
+  uint32_t x162;
+  fiat_p256_mulx_u32(&x161, &x162, x2, (arg1[6]));
   uint32_t x163;
   uint32_t x164;
-  fiat_p256_mulx_u32(&x163, &x164, x2, (arg1[7]));
+  fiat_p256_mulx_u32(&x163, &x164, x2, (arg1[5]));
   uint32_t x165;
   uint32_t x166;
-  fiat_p256_mulx_u32(&x165, &x166, x2, (arg1[6]));
+  fiat_p256_mulx_u32(&x165, &x166, x2, (arg1[4]));
   uint32_t x167;
   uint32_t x168;
-  fiat_p256_mulx_u32(&x167, &x168, x2, (arg1[5]));
+  fiat_p256_mulx_u32(&x167, &x168, x2, (arg1[3]));
   uint32_t x169;
   uint32_t x170;
-  fiat_p256_mulx_u32(&x169, &x170, x2, (arg1[4]));
+  fiat_p256_mulx_u32(&x169, &x170, x2, (arg1[2]));
   uint32_t x171;
   uint32_t x172;
-  fiat_p256_mulx_u32(&x171, &x172, x2, (arg1[3]));
+  fiat_p256_mulx_u32(&x171, &x172, x2, (arg1[1]));
   uint32_t x173;
   uint32_t x174;
-  fiat_p256_mulx_u32(&x173, &x174, x2, (arg1[2]));
+  fiat_p256_mulx_u32(&x173, &x174, x2, (arg1[0]));
   uint32_t x175;
-  uint32_t x176;
-  fiat_p256_mulx_u32(&x175, &x176, x2, (arg1[1]));
+  fiat_p256_uint1 x176;
+  fiat_p256_addcarryx_u32(&x175, &x176, 0x0, x171, x174);
   uint32_t x177;
-  uint32_t x178;
-  fiat_p256_mulx_u32(&x177, &x178, x2, (arg1[0]));
+  fiat_p256_uint1 x178;
+  fiat_p256_addcarryx_u32(&x177, &x178, x176, x169, x172);
   uint32_t x179;
   fiat_p256_uint1 x180;
-  fiat_p256_addcarryx_u32(&x179, &x180, 0x0, x175, x178);
+  fiat_p256_addcarryx_u32(&x179, &x180, x178, x167, x170);
   uint32_t x181;
   fiat_p256_uint1 x182;
-  fiat_p256_addcarryx_u32(&x181, &x182, x180, x173, x176);
+  fiat_p256_addcarryx_u32(&x181, &x182, x180, x165, x168);
   uint32_t x183;
   fiat_p256_uint1 x184;
-  fiat_p256_addcarryx_u32(&x183, &x184, x182, x171, x174);
+  fiat_p256_addcarryx_u32(&x183, &x184, x182, x163, x166);
   uint32_t x185;
   fiat_p256_uint1 x186;
-  fiat_p256_addcarryx_u32(&x185, &x186, x184, x169, x172);
+  fiat_p256_addcarryx_u32(&x185, &x186, x184, x161, x164);
   uint32_t x187;
   fiat_p256_uint1 x188;
-  fiat_p256_addcarryx_u32(&x187, &x188, x186, x167, x170);
+  fiat_p256_addcarryx_u32(&x187, &x188, x186, x159, x162);
   uint32_t x189;
   fiat_p256_uint1 x190;
-  fiat_p256_addcarryx_u32(&x189, &x190, x188, x165, x168);
+  fiat_p256_addcarryx_u32(&x189, &x190, x188, 0x0, x160);
   uint32_t x191;
   fiat_p256_uint1 x192;
-  fiat_p256_addcarryx_u32(&x191, &x192, x190, x163, x166);
+  fiat_p256_addcarryx_u32(&x191, &x192, 0x0, x173, x141);
   uint32_t x193;
   fiat_p256_uint1 x194;
-  fiat_p256_addcarryx_u32(&x193, &x194, x192, 0x0, x164);
+  fiat_p256_addcarryx_u32(&x193, &x194, x192, x175, x143);
   uint32_t x195;
   fiat_p256_uint1 x196;
-  fiat_p256_addcarryx_u32(&x195, &x196, 0x0, x177, x145);
+  fiat_p256_addcarryx_u32(&x195, &x196, x194, x177, x145);
   uint32_t x197;
   fiat_p256_uint1 x198;
   fiat_p256_addcarryx_u32(&x197, &x198, x196, x179, x147);
@@ -1471,116 +1447,116 @@ static void fiat_p256_square(uint32_t out1[8], const uint32_t arg1[8]) {
   fiat_p256_uint1 x208;
   fiat_p256_addcarryx_u32(&x207, &x208, x206, x189, x157);
   uint32_t x209;
-  fiat_p256_uint1 x210;
-  fiat_p256_addcarryx_u32(&x209, &x210, x208, x191, x159);
+  uint32_t x210;
+  fiat_p256_mulx_u32(&x209, &x210, x191, UINT32_C(0xffffffff));
   uint32_t x211;
-  fiat_p256_uint1 x212;
-  fiat_p256_addcarryx_u32(&x211, &x212, x210, x193, x161);
+  uint32_t x212;
+  fiat_p256_mulx_u32(&x211, &x212, x191, UINT32_C(0xffffffff));
   uint32_t x213;
   uint32_t x214;
-  fiat_p256_mulx_u32(&x213, &x214, x195, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u32(&x213, &x214, x191, UINT32_C(0xffffffff));
   uint32_t x215;
   uint32_t x216;
-  fiat_p256_mulx_u32(&x215, &x216, x195, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u32(&x215, &x216, x191, UINT32_C(0xffffffff));
   uint32_t x217;
-  uint32_t x218;
-  fiat_p256_mulx_u32(&x217, &x218, x195, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x218;
+  fiat_p256_addcarryx_u32(&x217, &x218, 0x0, x213, x216);
   uint32_t x219;
-  uint32_t x220;
-  fiat_p256_mulx_u32(&x219, &x220, x195, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x220;
+  fiat_p256_addcarryx_u32(&x219, &x220, x218, x211, x214);
   uint32_t x221;
   fiat_p256_uint1 x222;
-  fiat_p256_addcarryx_u32(&x221, &x222, 0x0, x217, x220);
+  fiat_p256_addcarryx_u32(&x221, &x222, x220, 0x0, x212);
   uint32_t x223;
   fiat_p256_uint1 x224;
-  fiat_p256_addcarryx_u32(&x223, &x224, x222, x215, x218);
+  fiat_p256_addcarryx_u32(&x223, &x224, 0x0, x215, x191);
   uint32_t x225;
   fiat_p256_uint1 x226;
-  fiat_p256_addcarryx_u32(&x225, &x226, x224, 0x0, x216);
+  fiat_p256_addcarryx_u32(&x225, &x226, x224, x217, x193);
   uint32_t x227;
   fiat_p256_uint1 x228;
-  fiat_p256_addcarryx_u32(&x227, &x228, x226, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x227, &x228, x226, x219, x195);
   uint32_t x229;
   fiat_p256_uint1 x230;
-  fiat_p256_addcarryx_u32(&x229, &x230, 0x0, x219, x195);
+  fiat_p256_addcarryx_u32(&x229, &x230, x228, x221, x197);
   uint32_t x231;
   fiat_p256_uint1 x232;
-  fiat_p256_addcarryx_u32(&x231, &x232, x230, x221, x197);
+  fiat_p256_addcarryx_u32(&x231, &x232, x230, 0x0, x199);
   uint32_t x233;
   fiat_p256_uint1 x234;
-  fiat_p256_addcarryx_u32(&x233, &x234, x232, x223, x199);
+  fiat_p256_addcarryx_u32(&x233, &x234, x232, 0x0, x201);
   uint32_t x235;
   fiat_p256_uint1 x236;
-  fiat_p256_addcarryx_u32(&x235, &x236, x234, x225, x201);
+  fiat_p256_addcarryx_u32(&x235, &x236, x234, x191, x203);
   uint32_t x237;
   fiat_p256_uint1 x238;
-  fiat_p256_addcarryx_u32(&x237, &x238, x236, (fiat_p256_uint1)x227, x203);
+  fiat_p256_addcarryx_u32(&x237, &x238, x236, x209, x205);
   uint32_t x239;
   fiat_p256_uint1 x240;
-  fiat_p256_addcarryx_u32(&x239, &x240, x238, 0x0, x205);
+  fiat_p256_addcarryx_u32(&x239, &x240, x238, x210, x207);
   uint32_t x241;
   fiat_p256_uint1 x242;
-  fiat_p256_addcarryx_u32(&x241, &x242, x240, x195, x207);
+  fiat_p256_addcarryx_u32(&x241, &x242, x240, 0x0, x208);
   uint32_t x243;
-  fiat_p256_uint1 x244;
-  fiat_p256_addcarryx_u32(&x243, &x244, x242, x213, x209);
+  uint32_t x244;
+  fiat_p256_mulx_u32(&x243, &x244, x3, (arg1[7]));
   uint32_t x245;
-  fiat_p256_uint1 x246;
-  fiat_p256_addcarryx_u32(&x245, &x246, x244, x214, x211);
+  uint32_t x246;
+  fiat_p256_mulx_u32(&x245, &x246, x3, (arg1[6]));
   uint32_t x247;
-  fiat_p256_uint1 x248;
-  fiat_p256_addcarryx_u32(&x247, &x248, x246, 0x0, x212);
+  uint32_t x248;
+  fiat_p256_mulx_u32(&x247, &x248, x3, (arg1[5]));
   uint32_t x249;
   uint32_t x250;
-  fiat_p256_mulx_u32(&x249, &x250, x3, (arg1[7]));
+  fiat_p256_mulx_u32(&x249, &x250, x3, (arg1[4]));
   uint32_t x251;
   uint32_t x252;
-  fiat_p256_mulx_u32(&x251, &x252, x3, (arg1[6]));
+  fiat_p256_mulx_u32(&x251, &x252, x3, (arg1[3]));
   uint32_t x253;
   uint32_t x254;
-  fiat_p256_mulx_u32(&x253, &x254, x3, (arg1[5]));
+  fiat_p256_mulx_u32(&x253, &x254, x3, (arg1[2]));
   uint32_t x255;
   uint32_t x256;
-  fiat_p256_mulx_u32(&x255, &x256, x3, (arg1[4]));
+  fiat_p256_mulx_u32(&x255, &x256, x3, (arg1[1]));
   uint32_t x257;
   uint32_t x258;
-  fiat_p256_mulx_u32(&x257, &x258, x3, (arg1[3]));
+  fiat_p256_mulx_u32(&x257, &x258, x3, (arg1[0]));
   uint32_t x259;
-  uint32_t x260;
-  fiat_p256_mulx_u32(&x259, &x260, x3, (arg1[2]));
+  fiat_p256_uint1 x260;
+  fiat_p256_addcarryx_u32(&x259, &x260, 0x0, x255, x258);
   uint32_t x261;
-  uint32_t x262;
-  fiat_p256_mulx_u32(&x261, &x262, x3, (arg1[1]));
+  fiat_p256_uint1 x262;
+  fiat_p256_addcarryx_u32(&x261, &x262, x260, x253, x256);
   uint32_t x263;
-  uint32_t x264;
-  fiat_p256_mulx_u32(&x263, &x264, x3, (arg1[0]));
+  fiat_p256_uint1 x264;
+  fiat_p256_addcarryx_u32(&x263, &x264, x262, x251, x254);
   uint32_t x265;
   fiat_p256_uint1 x266;
-  fiat_p256_addcarryx_u32(&x265, &x266, 0x0, x261, x264);
+  fiat_p256_addcarryx_u32(&x265, &x266, x264, x249, x252);
   uint32_t x267;
   fiat_p256_uint1 x268;
-  fiat_p256_addcarryx_u32(&x267, &x268, x266, x259, x262);
+  fiat_p256_addcarryx_u32(&x267, &x268, x266, x247, x250);
   uint32_t x269;
   fiat_p256_uint1 x270;
-  fiat_p256_addcarryx_u32(&x269, &x270, x268, x257, x260);
+  fiat_p256_addcarryx_u32(&x269, &x270, x268, x245, x248);
   uint32_t x271;
   fiat_p256_uint1 x272;
-  fiat_p256_addcarryx_u32(&x271, &x272, x270, x255, x258);
+  fiat_p256_addcarryx_u32(&x271, &x272, x270, x243, x246);
   uint32_t x273;
   fiat_p256_uint1 x274;
-  fiat_p256_addcarryx_u32(&x273, &x274, x272, x253, x256);
+  fiat_p256_addcarryx_u32(&x273, &x274, x272, 0x0, x244);
   uint32_t x275;
   fiat_p256_uint1 x276;
-  fiat_p256_addcarryx_u32(&x275, &x276, x274, x251, x254);
+  fiat_p256_addcarryx_u32(&x275, &x276, 0x0, x257, x225);
   uint32_t x277;
   fiat_p256_uint1 x278;
-  fiat_p256_addcarryx_u32(&x277, &x278, x276, x249, x252);
+  fiat_p256_addcarryx_u32(&x277, &x278, x276, x259, x227);
   uint32_t x279;
   fiat_p256_uint1 x280;
-  fiat_p256_addcarryx_u32(&x279, &x280, x278, 0x0, x250);
+  fiat_p256_addcarryx_u32(&x279, &x280, x278, x261, x229);
   uint32_t x281;
   fiat_p256_uint1 x282;
-  fiat_p256_addcarryx_u32(&x281, &x282, 0x0, x263, x231);
+  fiat_p256_addcarryx_u32(&x281, &x282, x280, x263, x231);
   uint32_t x283;
   fiat_p256_uint1 x284;
   fiat_p256_addcarryx_u32(&x283, &x284, x282, x265, x233);
@@ -1597,119 +1573,119 @@ static void fiat_p256_square(uint32_t out1[8], const uint32_t arg1[8]) {
   fiat_p256_uint1 x292;
   fiat_p256_addcarryx_u32(&x291, &x292, x290, x273, x241);
   uint32_t x293;
-  fiat_p256_uint1 x294;
-  fiat_p256_addcarryx_u32(&x293, &x294, x292, x275, x243);
+  uint32_t x294;
+  fiat_p256_mulx_u32(&x293, &x294, x275, UINT32_C(0xffffffff));
   uint32_t x295;
-  fiat_p256_uint1 x296;
-  fiat_p256_addcarryx_u32(&x295, &x296, x294, x277, x245);
+  uint32_t x296;
+  fiat_p256_mulx_u32(&x295, &x296, x275, UINT32_C(0xffffffff));
   uint32_t x297;
-  fiat_p256_uint1 x298;
-  fiat_p256_addcarryx_u32(&x297, &x298, x296, x279, x247);
+  uint32_t x298;
+  fiat_p256_mulx_u32(&x297, &x298, x275, UINT32_C(0xffffffff));
   uint32_t x299;
   uint32_t x300;
-  fiat_p256_mulx_u32(&x299, &x300, x281, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u32(&x299, &x300, x275, UINT32_C(0xffffffff));
   uint32_t x301;
-  uint32_t x302;
-  fiat_p256_mulx_u32(&x301, &x302, x281, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x302;
+  fiat_p256_addcarryx_u32(&x301, &x302, 0x0, x297, x300);
   uint32_t x303;
-  uint32_t x304;
-  fiat_p256_mulx_u32(&x303, &x304, x281, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x304;
+  fiat_p256_addcarryx_u32(&x303, &x304, x302, x295, x298);
   uint32_t x305;
-  uint32_t x306;
-  fiat_p256_mulx_u32(&x305, &x306, x281, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x306;
+  fiat_p256_addcarryx_u32(&x305, &x306, x304, 0x0, x296);
   uint32_t x307;
   fiat_p256_uint1 x308;
-  fiat_p256_addcarryx_u32(&x307, &x308, 0x0, x303, x306);
+  fiat_p256_addcarryx_u32(&x307, &x308, 0x0, x299, x275);
   uint32_t x309;
   fiat_p256_uint1 x310;
-  fiat_p256_addcarryx_u32(&x309, &x310, x308, x301, x304);
+  fiat_p256_addcarryx_u32(&x309, &x310, x308, x301, x277);
   uint32_t x311;
   fiat_p256_uint1 x312;
-  fiat_p256_addcarryx_u32(&x311, &x312, x310, 0x0, x302);
+  fiat_p256_addcarryx_u32(&x311, &x312, x310, x303, x279);
   uint32_t x313;
   fiat_p256_uint1 x314;
-  fiat_p256_addcarryx_u32(&x313, &x314, x312, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x313, &x314, x312, x305, x281);
   uint32_t x315;
   fiat_p256_uint1 x316;
-  fiat_p256_addcarryx_u32(&x315, &x316, 0x0, x305, x281);
+  fiat_p256_addcarryx_u32(&x315, &x316, x314, 0x0, x283);
   uint32_t x317;
   fiat_p256_uint1 x318;
-  fiat_p256_addcarryx_u32(&x317, &x318, x316, x307, x283);
+  fiat_p256_addcarryx_u32(&x317, &x318, x316, 0x0, x285);
   uint32_t x319;
   fiat_p256_uint1 x320;
-  fiat_p256_addcarryx_u32(&x319, &x320, x318, x309, x285);
+  fiat_p256_addcarryx_u32(&x319, &x320, x318, x275, x287);
   uint32_t x321;
   fiat_p256_uint1 x322;
-  fiat_p256_addcarryx_u32(&x321, &x322, x320, x311, x287);
+  fiat_p256_addcarryx_u32(&x321, &x322, x320, x293, x289);
   uint32_t x323;
   fiat_p256_uint1 x324;
-  fiat_p256_addcarryx_u32(&x323, &x324, x322, (fiat_p256_uint1)x313, x289);
+  fiat_p256_addcarryx_u32(&x323, &x324, x322, x294, x291);
   uint32_t x325;
   fiat_p256_uint1 x326;
-  fiat_p256_addcarryx_u32(&x325, &x326, x324, 0x0, x291);
+  fiat_p256_addcarryx_u32(&x325, &x326, x324, 0x0, x292);
   uint32_t x327;
-  fiat_p256_uint1 x328;
-  fiat_p256_addcarryx_u32(&x327, &x328, x326, x281, x293);
+  uint32_t x328;
+  fiat_p256_mulx_u32(&x327, &x328, x4, (arg1[7]));
   uint32_t x329;
-  fiat_p256_uint1 x330;
-  fiat_p256_addcarryx_u32(&x329, &x330, x328, x299, x295);
+  uint32_t x330;
+  fiat_p256_mulx_u32(&x329, &x330, x4, (arg1[6]));
   uint32_t x331;
-  fiat_p256_uint1 x332;
-  fiat_p256_addcarryx_u32(&x331, &x332, x330, x300, x297);
+  uint32_t x332;
+  fiat_p256_mulx_u32(&x331, &x332, x4, (arg1[5]));
   uint32_t x333;
-  fiat_p256_uint1 x334;
-  fiat_p256_addcarryx_u32(&x333, &x334, x332, 0x0, x298);
+  uint32_t x334;
+  fiat_p256_mulx_u32(&x333, &x334, x4, (arg1[4]));
   uint32_t x335;
   uint32_t x336;
-  fiat_p256_mulx_u32(&x335, &x336, x4, (arg1[7]));
+  fiat_p256_mulx_u32(&x335, &x336, x4, (arg1[3]));
   uint32_t x337;
   uint32_t x338;
-  fiat_p256_mulx_u32(&x337, &x338, x4, (arg1[6]));
+  fiat_p256_mulx_u32(&x337, &x338, x4, (arg1[2]));
   uint32_t x339;
   uint32_t x340;
-  fiat_p256_mulx_u32(&x339, &x340, x4, (arg1[5]));
+  fiat_p256_mulx_u32(&x339, &x340, x4, (arg1[1]));
   uint32_t x341;
   uint32_t x342;
-  fiat_p256_mulx_u32(&x341, &x342, x4, (arg1[4]));
+  fiat_p256_mulx_u32(&x341, &x342, x4, (arg1[0]));
   uint32_t x343;
-  uint32_t x344;
-  fiat_p256_mulx_u32(&x343, &x344, x4, (arg1[3]));
+  fiat_p256_uint1 x344;
+  fiat_p256_addcarryx_u32(&x343, &x344, 0x0, x339, x342);
   uint32_t x345;
-  uint32_t x346;
-  fiat_p256_mulx_u32(&x345, &x346, x4, (arg1[2]));
+  fiat_p256_uint1 x346;
+  fiat_p256_addcarryx_u32(&x345, &x346, x344, x337, x340);
   uint32_t x347;
-  uint32_t x348;
-  fiat_p256_mulx_u32(&x347, &x348, x4, (arg1[1]));
+  fiat_p256_uint1 x348;
+  fiat_p256_addcarryx_u32(&x347, &x348, x346, x335, x338);
   uint32_t x349;
-  uint32_t x350;
-  fiat_p256_mulx_u32(&x349, &x350, x4, (arg1[0]));
+  fiat_p256_uint1 x350;
+  fiat_p256_addcarryx_u32(&x349, &x350, x348, x333, x336);
   uint32_t x351;
   fiat_p256_uint1 x352;
-  fiat_p256_addcarryx_u32(&x351, &x352, 0x0, x347, x350);
+  fiat_p256_addcarryx_u32(&x351, &x352, x350, x331, x334);
   uint32_t x353;
   fiat_p256_uint1 x354;
-  fiat_p256_addcarryx_u32(&x353, &x354, x352, x345, x348);
+  fiat_p256_addcarryx_u32(&x353, &x354, x352, x329, x332);
   uint32_t x355;
   fiat_p256_uint1 x356;
-  fiat_p256_addcarryx_u32(&x355, &x356, x354, x343, x346);
+  fiat_p256_addcarryx_u32(&x355, &x356, x354, x327, x330);
   uint32_t x357;
   fiat_p256_uint1 x358;
-  fiat_p256_addcarryx_u32(&x357, &x358, x356, x341, x344);
+  fiat_p256_addcarryx_u32(&x357, &x358, x356, 0x0, x328);
   uint32_t x359;
   fiat_p256_uint1 x360;
-  fiat_p256_addcarryx_u32(&x359, &x360, x358, x339, x342);
+  fiat_p256_addcarryx_u32(&x359, &x360, 0x0, x341, x309);
   uint32_t x361;
   fiat_p256_uint1 x362;
-  fiat_p256_addcarryx_u32(&x361, &x362, x360, x337, x340);
+  fiat_p256_addcarryx_u32(&x361, &x362, x360, x343, x311);
   uint32_t x363;
   fiat_p256_uint1 x364;
-  fiat_p256_addcarryx_u32(&x363, &x364, x362, x335, x338);
+  fiat_p256_addcarryx_u32(&x363, &x364, x362, x345, x313);
   uint32_t x365;
   fiat_p256_uint1 x366;
-  fiat_p256_addcarryx_u32(&x365, &x366, x364, 0x0, x336);
+  fiat_p256_addcarryx_u32(&x365, &x366, x364, x347, x315);
   uint32_t x367;
   fiat_p256_uint1 x368;
-  fiat_p256_addcarryx_u32(&x367, &x368, 0x0, x349, x317);
+  fiat_p256_addcarryx_u32(&x367, &x368, x366, x349, x317);
   uint32_t x369;
   fiat_p256_uint1 x370;
   fiat_p256_addcarryx_u32(&x369, &x370, x368, x351, x319);
@@ -1723,122 +1699,122 @@ static void fiat_p256_square(uint32_t out1[8], const uint32_t arg1[8]) {
   fiat_p256_uint1 x376;
   fiat_p256_addcarryx_u32(&x375, &x376, x374, x357, x325);
   uint32_t x377;
-  fiat_p256_uint1 x378;
-  fiat_p256_addcarryx_u32(&x377, &x378, x376, x359, x327);
+  uint32_t x378;
+  fiat_p256_mulx_u32(&x377, &x378, x359, UINT32_C(0xffffffff));
   uint32_t x379;
-  fiat_p256_uint1 x380;
-  fiat_p256_addcarryx_u32(&x379, &x380, x378, x361, x329);
+  uint32_t x380;
+  fiat_p256_mulx_u32(&x379, &x380, x359, UINT32_C(0xffffffff));
   uint32_t x381;
-  fiat_p256_uint1 x382;
-  fiat_p256_addcarryx_u32(&x381, &x382, x380, x363, x331);
+  uint32_t x382;
+  fiat_p256_mulx_u32(&x381, &x382, x359, UINT32_C(0xffffffff));
   uint32_t x383;
-  fiat_p256_uint1 x384;
-  fiat_p256_addcarryx_u32(&x383, &x384, x382, x365, x333);
+  uint32_t x384;
+  fiat_p256_mulx_u32(&x383, &x384, x359, UINT32_C(0xffffffff));
   uint32_t x385;
-  uint32_t x386;
-  fiat_p256_mulx_u32(&x385, &x386, x367, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x386;
+  fiat_p256_addcarryx_u32(&x385, &x386, 0x0, x381, x384);
   uint32_t x387;
-  uint32_t x388;
-  fiat_p256_mulx_u32(&x387, &x388, x367, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x388;
+  fiat_p256_addcarryx_u32(&x387, &x388, x386, x379, x382);
   uint32_t x389;
-  uint32_t x390;
-  fiat_p256_mulx_u32(&x389, &x390, x367, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x390;
+  fiat_p256_addcarryx_u32(&x389, &x390, x388, 0x0, x380);
   uint32_t x391;
-  uint32_t x392;
-  fiat_p256_mulx_u32(&x391, &x392, x367, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x392;
+  fiat_p256_addcarryx_u32(&x391, &x392, 0x0, x383, x359);
   uint32_t x393;
   fiat_p256_uint1 x394;
-  fiat_p256_addcarryx_u32(&x393, &x394, 0x0, x389, x392);
+  fiat_p256_addcarryx_u32(&x393, &x394, x392, x385, x361);
   uint32_t x395;
   fiat_p256_uint1 x396;
-  fiat_p256_addcarryx_u32(&x395, &x396, x394, x387, x390);
+  fiat_p256_addcarryx_u32(&x395, &x396, x394, x387, x363);
   uint32_t x397;
   fiat_p256_uint1 x398;
-  fiat_p256_addcarryx_u32(&x397, &x398, x396, 0x0, x388);
+  fiat_p256_addcarryx_u32(&x397, &x398, x396, x389, x365);
   uint32_t x399;
   fiat_p256_uint1 x400;
-  fiat_p256_addcarryx_u32(&x399, &x400, x398, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x399, &x400, x398, 0x0, x367);
   uint32_t x401;
   fiat_p256_uint1 x402;
-  fiat_p256_addcarryx_u32(&x401, &x402, 0x0, x391, x367);
+  fiat_p256_addcarryx_u32(&x401, &x402, x400, 0x0, x369);
   uint32_t x403;
   fiat_p256_uint1 x404;
-  fiat_p256_addcarryx_u32(&x403, &x404, x402, x393, x369);
+  fiat_p256_addcarryx_u32(&x403, &x404, x402, x359, x371);
   uint32_t x405;
   fiat_p256_uint1 x406;
-  fiat_p256_addcarryx_u32(&x405, &x406, x404, x395, x371);
+  fiat_p256_addcarryx_u32(&x405, &x406, x404, x377, x373);
   uint32_t x407;
   fiat_p256_uint1 x408;
-  fiat_p256_addcarryx_u32(&x407, &x408, x406, x397, x373);
+  fiat_p256_addcarryx_u32(&x407, &x408, x406, x378, x375);
   uint32_t x409;
   fiat_p256_uint1 x410;
-  fiat_p256_addcarryx_u32(&x409, &x410, x408, (fiat_p256_uint1)x399, x375);
+  fiat_p256_addcarryx_u32(&x409, &x410, x408, 0x0, x376);
   uint32_t x411;
-  fiat_p256_uint1 x412;
-  fiat_p256_addcarryx_u32(&x411, &x412, x410, 0x0, x377);
+  uint32_t x412;
+  fiat_p256_mulx_u32(&x411, &x412, x5, (arg1[7]));
   uint32_t x413;
-  fiat_p256_uint1 x414;
-  fiat_p256_addcarryx_u32(&x413, &x414, x412, x367, x379);
+  uint32_t x414;
+  fiat_p256_mulx_u32(&x413, &x414, x5, (arg1[6]));
   uint32_t x415;
-  fiat_p256_uint1 x416;
-  fiat_p256_addcarryx_u32(&x415, &x416, x414, x385, x381);
+  uint32_t x416;
+  fiat_p256_mulx_u32(&x415, &x416, x5, (arg1[5]));
   uint32_t x417;
-  fiat_p256_uint1 x418;
-  fiat_p256_addcarryx_u32(&x417, &x418, x416, x386, x383);
+  uint32_t x418;
+  fiat_p256_mulx_u32(&x417, &x418, x5, (arg1[4]));
   uint32_t x419;
-  fiat_p256_uint1 x420;
-  fiat_p256_addcarryx_u32(&x419, &x420, x418, 0x0, x384);
+  uint32_t x420;
+  fiat_p256_mulx_u32(&x419, &x420, x5, (arg1[3]));
   uint32_t x421;
   uint32_t x422;
-  fiat_p256_mulx_u32(&x421, &x422, x5, (arg1[7]));
+  fiat_p256_mulx_u32(&x421, &x422, x5, (arg1[2]));
   uint32_t x423;
   uint32_t x424;
-  fiat_p256_mulx_u32(&x423, &x424, x5, (arg1[6]));
+  fiat_p256_mulx_u32(&x423, &x424, x5, (arg1[1]));
   uint32_t x425;
   uint32_t x426;
-  fiat_p256_mulx_u32(&x425, &x426, x5, (arg1[5]));
+  fiat_p256_mulx_u32(&x425, &x426, x5, (arg1[0]));
   uint32_t x427;
-  uint32_t x428;
-  fiat_p256_mulx_u32(&x427, &x428, x5, (arg1[4]));
+  fiat_p256_uint1 x428;
+  fiat_p256_addcarryx_u32(&x427, &x428, 0x0, x423, x426);
   uint32_t x429;
-  uint32_t x430;
-  fiat_p256_mulx_u32(&x429, &x430, x5, (arg1[3]));
+  fiat_p256_uint1 x430;
+  fiat_p256_addcarryx_u32(&x429, &x430, x428, x421, x424);
   uint32_t x431;
-  uint32_t x432;
-  fiat_p256_mulx_u32(&x431, &x432, x5, (arg1[2]));
+  fiat_p256_uint1 x432;
+  fiat_p256_addcarryx_u32(&x431, &x432, x430, x419, x422);
   uint32_t x433;
-  uint32_t x434;
-  fiat_p256_mulx_u32(&x433, &x434, x5, (arg1[1]));
+  fiat_p256_uint1 x434;
+  fiat_p256_addcarryx_u32(&x433, &x434, x432, x417, x420);
   uint32_t x435;
-  uint32_t x436;
-  fiat_p256_mulx_u32(&x435, &x436, x5, (arg1[0]));
+  fiat_p256_uint1 x436;
+  fiat_p256_addcarryx_u32(&x435, &x436, x434, x415, x418);
   uint32_t x437;
   fiat_p256_uint1 x438;
-  fiat_p256_addcarryx_u32(&x437, &x438, 0x0, x433, x436);
+  fiat_p256_addcarryx_u32(&x437, &x438, x436, x413, x416);
   uint32_t x439;
   fiat_p256_uint1 x440;
-  fiat_p256_addcarryx_u32(&x439, &x440, x438, x431, x434);
+  fiat_p256_addcarryx_u32(&x439, &x440, x438, x411, x414);
   uint32_t x441;
   fiat_p256_uint1 x442;
-  fiat_p256_addcarryx_u32(&x441, &x442, x440, x429, x432);
+  fiat_p256_addcarryx_u32(&x441, &x442, x440, 0x0, x412);
   uint32_t x443;
   fiat_p256_uint1 x444;
-  fiat_p256_addcarryx_u32(&x443, &x444, x442, x427, x430);
+  fiat_p256_addcarryx_u32(&x443, &x444, 0x0, x425, x393);
   uint32_t x445;
   fiat_p256_uint1 x446;
-  fiat_p256_addcarryx_u32(&x445, &x446, x444, x425, x428);
+  fiat_p256_addcarryx_u32(&x445, &x446, x444, x427, x395);
   uint32_t x447;
   fiat_p256_uint1 x448;
-  fiat_p256_addcarryx_u32(&x447, &x448, x446, x423, x426);
+  fiat_p256_addcarryx_u32(&x447, &x448, x446, x429, x397);
   uint32_t x449;
   fiat_p256_uint1 x450;
-  fiat_p256_addcarryx_u32(&x449, &x450, x448, x421, x424);
+  fiat_p256_addcarryx_u32(&x449, &x450, x448, x431, x399);
   uint32_t x451;
   fiat_p256_uint1 x452;
-  fiat_p256_addcarryx_u32(&x451, &x452, x450, 0x0, x422);
+  fiat_p256_addcarryx_u32(&x451, &x452, x450, x433, x401);
   uint32_t x453;
   fiat_p256_uint1 x454;
-  fiat_p256_addcarryx_u32(&x453, &x454, 0x0, x435, x403);
+  fiat_p256_addcarryx_u32(&x453, &x454, x452, x435, x403);
   uint32_t x455;
   fiat_p256_uint1 x456;
   fiat_p256_addcarryx_u32(&x455, &x456, x454, x437, x405);
@@ -1849,125 +1825,125 @@ static void fiat_p256_square(uint32_t out1[8], const uint32_t arg1[8]) {
   fiat_p256_uint1 x460;
   fiat_p256_addcarryx_u32(&x459, &x460, x458, x441, x409);
   uint32_t x461;
-  fiat_p256_uint1 x462;
-  fiat_p256_addcarryx_u32(&x461, &x462, x460, x443, x411);
+  uint32_t x462;
+  fiat_p256_mulx_u32(&x461, &x462, x443, UINT32_C(0xffffffff));
   uint32_t x463;
-  fiat_p256_uint1 x464;
-  fiat_p256_addcarryx_u32(&x463, &x464, x462, x445, x413);
+  uint32_t x464;
+  fiat_p256_mulx_u32(&x463, &x464, x443, UINT32_C(0xffffffff));
   uint32_t x465;
-  fiat_p256_uint1 x466;
-  fiat_p256_addcarryx_u32(&x465, &x466, x464, x447, x415);
+  uint32_t x466;
+  fiat_p256_mulx_u32(&x465, &x466, x443, UINT32_C(0xffffffff));
   uint32_t x467;
-  fiat_p256_uint1 x468;
-  fiat_p256_addcarryx_u32(&x467, &x468, x466, x449, x417);
+  uint32_t x468;
+  fiat_p256_mulx_u32(&x467, &x468, x443, UINT32_C(0xffffffff));
   uint32_t x469;
   fiat_p256_uint1 x470;
-  fiat_p256_addcarryx_u32(&x469, &x470, x468, x451, x419);
+  fiat_p256_addcarryx_u32(&x469, &x470, 0x0, x465, x468);
   uint32_t x471;
-  uint32_t x472;
-  fiat_p256_mulx_u32(&x471, &x472, x453, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x472;
+  fiat_p256_addcarryx_u32(&x471, &x472, x470, x463, x466);
   uint32_t x473;
-  uint32_t x474;
-  fiat_p256_mulx_u32(&x473, &x474, x453, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x474;
+  fiat_p256_addcarryx_u32(&x473, &x474, x472, 0x0, x464);
   uint32_t x475;
-  uint32_t x476;
-  fiat_p256_mulx_u32(&x475, &x476, x453, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x476;
+  fiat_p256_addcarryx_u32(&x475, &x476, 0x0, x467, x443);
   uint32_t x477;
-  uint32_t x478;
-  fiat_p256_mulx_u32(&x477, &x478, x453, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x478;
+  fiat_p256_addcarryx_u32(&x477, &x478, x476, x469, x445);
   uint32_t x479;
   fiat_p256_uint1 x480;
-  fiat_p256_addcarryx_u32(&x479, &x480, 0x0, x475, x478);
+  fiat_p256_addcarryx_u32(&x479, &x480, x478, x471, x447);
   uint32_t x481;
   fiat_p256_uint1 x482;
-  fiat_p256_addcarryx_u32(&x481, &x482, x480, x473, x476);
+  fiat_p256_addcarryx_u32(&x481, &x482, x480, x473, x449);
   uint32_t x483;
   fiat_p256_uint1 x484;
-  fiat_p256_addcarryx_u32(&x483, &x484, x482, 0x0, x474);
+  fiat_p256_addcarryx_u32(&x483, &x484, x482, 0x0, x451);
   uint32_t x485;
   fiat_p256_uint1 x486;
-  fiat_p256_addcarryx_u32(&x485, &x486, x484, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x485, &x486, x484, 0x0, x453);
   uint32_t x487;
   fiat_p256_uint1 x488;
-  fiat_p256_addcarryx_u32(&x487, &x488, 0x0, x477, x453);
+  fiat_p256_addcarryx_u32(&x487, &x488, x486, x443, x455);
   uint32_t x489;
   fiat_p256_uint1 x490;
-  fiat_p256_addcarryx_u32(&x489, &x490, x488, x479, x455);
+  fiat_p256_addcarryx_u32(&x489, &x490, x488, x461, x457);
   uint32_t x491;
   fiat_p256_uint1 x492;
-  fiat_p256_addcarryx_u32(&x491, &x492, x490, x481, x457);
+  fiat_p256_addcarryx_u32(&x491, &x492, x490, x462, x459);
   uint32_t x493;
   fiat_p256_uint1 x494;
-  fiat_p256_addcarryx_u32(&x493, &x494, x492, x483, x459);
+  fiat_p256_addcarryx_u32(&x493, &x494, x492, 0x0, x460);
   uint32_t x495;
-  fiat_p256_uint1 x496;
-  fiat_p256_addcarryx_u32(&x495, &x496, x494, (fiat_p256_uint1)x485, x461);
+  uint32_t x496;
+  fiat_p256_mulx_u32(&x495, &x496, x6, (arg1[7]));
   uint32_t x497;
-  fiat_p256_uint1 x498;
-  fiat_p256_addcarryx_u32(&x497, &x498, x496, 0x0, x463);
+  uint32_t x498;
+  fiat_p256_mulx_u32(&x497, &x498, x6, (arg1[6]));
   uint32_t x499;
-  fiat_p256_uint1 x500;
-  fiat_p256_addcarryx_u32(&x499, &x500, x498, x453, x465);
+  uint32_t x500;
+  fiat_p256_mulx_u32(&x499, &x500, x6, (arg1[5]));
   uint32_t x501;
-  fiat_p256_uint1 x502;
-  fiat_p256_addcarryx_u32(&x501, &x502, x500, x471, x467);
+  uint32_t x502;
+  fiat_p256_mulx_u32(&x501, &x502, x6, (arg1[4]));
   uint32_t x503;
-  fiat_p256_uint1 x504;
-  fiat_p256_addcarryx_u32(&x503, &x504, x502, x472, x469);
+  uint32_t x504;
+  fiat_p256_mulx_u32(&x503, &x504, x6, (arg1[3]));
   uint32_t x505;
-  fiat_p256_uint1 x506;
-  fiat_p256_addcarryx_u32(&x505, &x506, x504, 0x0, x470);
+  uint32_t x506;
+  fiat_p256_mulx_u32(&x505, &x506, x6, (arg1[2]));
   uint32_t x507;
   uint32_t x508;
-  fiat_p256_mulx_u32(&x507, &x508, x6, (arg1[7]));
+  fiat_p256_mulx_u32(&x507, &x508, x6, (arg1[1]));
   uint32_t x509;
   uint32_t x510;
-  fiat_p256_mulx_u32(&x509, &x510, x6, (arg1[6]));
+  fiat_p256_mulx_u32(&x509, &x510, x6, (arg1[0]));
   uint32_t x511;
-  uint32_t x512;
-  fiat_p256_mulx_u32(&x511, &x512, x6, (arg1[5]));
+  fiat_p256_uint1 x512;
+  fiat_p256_addcarryx_u32(&x511, &x512, 0x0, x507, x510);
   uint32_t x513;
-  uint32_t x514;
-  fiat_p256_mulx_u32(&x513, &x514, x6, (arg1[4]));
+  fiat_p256_uint1 x514;
+  fiat_p256_addcarryx_u32(&x513, &x514, x512, x505, x508);
   uint32_t x515;
-  uint32_t x516;
-  fiat_p256_mulx_u32(&x515, &x516, x6, (arg1[3]));
+  fiat_p256_uint1 x516;
+  fiat_p256_addcarryx_u32(&x515, &x516, x514, x503, x506);
   uint32_t x517;
-  uint32_t x518;
-  fiat_p256_mulx_u32(&x517, &x518, x6, (arg1[2]));
+  fiat_p256_uint1 x518;
+  fiat_p256_addcarryx_u32(&x517, &x518, x516, x501, x504);
   uint32_t x519;
-  uint32_t x520;
-  fiat_p256_mulx_u32(&x519, &x520, x6, (arg1[1]));
+  fiat_p256_uint1 x520;
+  fiat_p256_addcarryx_u32(&x519, &x520, x518, x499, x502);
   uint32_t x521;
-  uint32_t x522;
-  fiat_p256_mulx_u32(&x521, &x522, x6, (arg1[0]));
+  fiat_p256_uint1 x522;
+  fiat_p256_addcarryx_u32(&x521, &x522, x520, x497, x500);
   uint32_t x523;
   fiat_p256_uint1 x524;
-  fiat_p256_addcarryx_u32(&x523, &x524, 0x0, x519, x522);
+  fiat_p256_addcarryx_u32(&x523, &x524, x522, x495, x498);
   uint32_t x525;
   fiat_p256_uint1 x526;
-  fiat_p256_addcarryx_u32(&x525, &x526, x524, x517, x520);
+  fiat_p256_addcarryx_u32(&x525, &x526, x524, 0x0, x496);
   uint32_t x527;
   fiat_p256_uint1 x528;
-  fiat_p256_addcarryx_u32(&x527, &x528, x526, x515, x518);
+  fiat_p256_addcarryx_u32(&x527, &x528, 0x0, x509, x477);
   uint32_t x529;
   fiat_p256_uint1 x530;
-  fiat_p256_addcarryx_u32(&x529, &x530, x528, x513, x516);
+  fiat_p256_addcarryx_u32(&x529, &x530, x528, x511, x479);
   uint32_t x531;
   fiat_p256_uint1 x532;
-  fiat_p256_addcarryx_u32(&x531, &x532, x530, x511, x514);
+  fiat_p256_addcarryx_u32(&x531, &x532, x530, x513, x481);
   uint32_t x533;
   fiat_p256_uint1 x534;
-  fiat_p256_addcarryx_u32(&x533, &x534, x532, x509, x512);
+  fiat_p256_addcarryx_u32(&x533, &x534, x532, x515, x483);
   uint32_t x535;
   fiat_p256_uint1 x536;
-  fiat_p256_addcarryx_u32(&x535, &x536, x534, x507, x510);
+  fiat_p256_addcarryx_u32(&x535, &x536, x534, x517, x485);
   uint32_t x537;
   fiat_p256_uint1 x538;
-  fiat_p256_addcarryx_u32(&x537, &x538, x536, 0x0, x508);
+  fiat_p256_addcarryx_u32(&x537, &x538, x536, x519, x487);
   uint32_t x539;
   fiat_p256_uint1 x540;
-  fiat_p256_addcarryx_u32(&x539, &x540, 0x0, x521, x489);
+  fiat_p256_addcarryx_u32(&x539, &x540, x538, x521, x489);
   uint32_t x541;
   fiat_p256_uint1 x542;
   fiat_p256_addcarryx_u32(&x541, &x542, x540, x523, x491);
@@ -1975,257 +1951,233 @@ static void fiat_p256_square(uint32_t out1[8], const uint32_t arg1[8]) {
   fiat_p256_uint1 x544;
   fiat_p256_addcarryx_u32(&x543, &x544, x542, x525, x493);
   uint32_t x545;
-  fiat_p256_uint1 x546;
-  fiat_p256_addcarryx_u32(&x545, &x546, x544, x527, x495);
+  uint32_t x546;
+  fiat_p256_mulx_u32(&x545, &x546, x527, UINT32_C(0xffffffff));
   uint32_t x547;
-  fiat_p256_uint1 x548;
-  fiat_p256_addcarryx_u32(&x547, &x548, x546, x529, x497);
+  uint32_t x548;
+  fiat_p256_mulx_u32(&x547, &x548, x527, UINT32_C(0xffffffff));
   uint32_t x549;
-  fiat_p256_uint1 x550;
-  fiat_p256_addcarryx_u32(&x549, &x550, x548, x531, x499);
+  uint32_t x550;
+  fiat_p256_mulx_u32(&x549, &x550, x527, UINT32_C(0xffffffff));
   uint32_t x551;
-  fiat_p256_uint1 x552;
-  fiat_p256_addcarryx_u32(&x551, &x552, x550, x533, x501);
+  uint32_t x552;
+  fiat_p256_mulx_u32(&x551, &x552, x527, UINT32_C(0xffffffff));
   uint32_t x553;
   fiat_p256_uint1 x554;
-  fiat_p256_addcarryx_u32(&x553, &x554, x552, x535, x503);
+  fiat_p256_addcarryx_u32(&x553, &x554, 0x0, x549, x552);
   uint32_t x555;
   fiat_p256_uint1 x556;
-  fiat_p256_addcarryx_u32(&x555, &x556, x554, x537, x505);
+  fiat_p256_addcarryx_u32(&x555, &x556, x554, x547, x550);
   uint32_t x557;
-  uint32_t x558;
-  fiat_p256_mulx_u32(&x557, &x558, x539, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x558;
+  fiat_p256_addcarryx_u32(&x557, &x558, x556, 0x0, x548);
   uint32_t x559;
-  uint32_t x560;
-  fiat_p256_mulx_u32(&x559, &x560, x539, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x560;
+  fiat_p256_addcarryx_u32(&x559, &x560, 0x0, x551, x527);
   uint32_t x561;
-  uint32_t x562;
-  fiat_p256_mulx_u32(&x561, &x562, x539, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x562;
+  fiat_p256_addcarryx_u32(&x561, &x562, x560, x553, x529);
   uint32_t x563;
-  uint32_t x564;
-  fiat_p256_mulx_u32(&x563, &x564, x539, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x564;
+  fiat_p256_addcarryx_u32(&x563, &x564, x562, x555, x531);
   uint32_t x565;
   fiat_p256_uint1 x566;
-  fiat_p256_addcarryx_u32(&x565, &x566, 0x0, x561, x564);
+  fiat_p256_addcarryx_u32(&x565, &x566, x564, x557, x533);
   uint32_t x567;
   fiat_p256_uint1 x568;
-  fiat_p256_addcarryx_u32(&x567, &x568, x566, x559, x562);
+  fiat_p256_addcarryx_u32(&x567, &x568, x566, 0x0, x535);
   uint32_t x569;
   fiat_p256_uint1 x570;
-  fiat_p256_addcarryx_u32(&x569, &x570, x568, 0x0, x560);
+  fiat_p256_addcarryx_u32(&x569, &x570, x568, 0x0, x537);
   uint32_t x571;
   fiat_p256_uint1 x572;
-  fiat_p256_addcarryx_u32(&x571, &x572, x570, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x571, &x572, x570, x527, x539);
   uint32_t x573;
   fiat_p256_uint1 x574;
-  fiat_p256_addcarryx_u32(&x573, &x574, 0x0, x563, x539);
+  fiat_p256_addcarryx_u32(&x573, &x574, x572, x545, x541);
   uint32_t x575;
   fiat_p256_uint1 x576;
-  fiat_p256_addcarryx_u32(&x575, &x576, x574, x565, x541);
+  fiat_p256_addcarryx_u32(&x575, &x576, x574, x546, x543);
   uint32_t x577;
   fiat_p256_uint1 x578;
-  fiat_p256_addcarryx_u32(&x577, &x578, x576, x567, x543);
+  fiat_p256_addcarryx_u32(&x577, &x578, x576, 0x0, x544);
   uint32_t x579;
-  fiat_p256_uint1 x580;
-  fiat_p256_addcarryx_u32(&x579, &x580, x578, x569, x545);
+  uint32_t x580;
+  fiat_p256_mulx_u32(&x579, &x580, x7, (arg1[7]));
   uint32_t x581;
-  fiat_p256_uint1 x582;
-  fiat_p256_addcarryx_u32(&x581, &x582, x580, (fiat_p256_uint1)x571, x547);
+  uint32_t x582;
+  fiat_p256_mulx_u32(&x581, &x582, x7, (arg1[6]));
   uint32_t x583;
-  fiat_p256_uint1 x584;
-  fiat_p256_addcarryx_u32(&x583, &x584, x582, 0x0, x549);
+  uint32_t x584;
+  fiat_p256_mulx_u32(&x583, &x584, x7, (arg1[5]));
   uint32_t x585;
-  fiat_p256_uint1 x586;
-  fiat_p256_addcarryx_u32(&x585, &x586, x584, x539, x551);
+  uint32_t x586;
+  fiat_p256_mulx_u32(&x585, &x586, x7, (arg1[4]));
   uint32_t x587;
-  fiat_p256_uint1 x588;
-  fiat_p256_addcarryx_u32(&x587, &x588, x586, x557, x553);
+  uint32_t x588;
+  fiat_p256_mulx_u32(&x587, &x588, x7, (arg1[3]));
   uint32_t x589;
-  fiat_p256_uint1 x590;
-  fiat_p256_addcarryx_u32(&x589, &x590, x588, x558, x555);
+  uint32_t x590;
+  fiat_p256_mulx_u32(&x589, &x590, x7, (arg1[2]));
   uint32_t x591;
-  fiat_p256_uint1 x592;
-  fiat_p256_addcarryx_u32(&x591, &x592, x590, 0x0, x556);
+  uint32_t x592;
+  fiat_p256_mulx_u32(&x591, &x592, x7, (arg1[1]));
   uint32_t x593;
   uint32_t x594;
-  fiat_p256_mulx_u32(&x593, &x594, x7, (arg1[7]));
+  fiat_p256_mulx_u32(&x593, &x594, x7, (arg1[0]));
   uint32_t x595;
-  uint32_t x596;
-  fiat_p256_mulx_u32(&x595, &x596, x7, (arg1[6]));
+  fiat_p256_uint1 x596;
+  fiat_p256_addcarryx_u32(&x595, &x596, 0x0, x591, x594);
   uint32_t x597;
-  uint32_t x598;
-  fiat_p256_mulx_u32(&x597, &x598, x7, (arg1[5]));
+  fiat_p256_uint1 x598;
+  fiat_p256_addcarryx_u32(&x597, &x598, x596, x589, x592);
   uint32_t x599;
-  uint32_t x600;
-  fiat_p256_mulx_u32(&x599, &x600, x7, (arg1[4]));
+  fiat_p256_uint1 x600;
+  fiat_p256_addcarryx_u32(&x599, &x600, x598, x587, x590);
   uint32_t x601;
-  uint32_t x602;
-  fiat_p256_mulx_u32(&x601, &x602, x7, (arg1[3]));
+  fiat_p256_uint1 x602;
+  fiat_p256_addcarryx_u32(&x601, &x602, x600, x585, x588);
   uint32_t x603;
-  uint32_t x604;
-  fiat_p256_mulx_u32(&x603, &x604, x7, (arg1[2]));
+  fiat_p256_uint1 x604;
+  fiat_p256_addcarryx_u32(&x603, &x604, x602, x583, x586);
   uint32_t x605;
-  uint32_t x606;
-  fiat_p256_mulx_u32(&x605, &x606, x7, (arg1[1]));
+  fiat_p256_uint1 x606;
+  fiat_p256_addcarryx_u32(&x605, &x606, x604, x581, x584);
   uint32_t x607;
-  uint32_t x608;
-  fiat_p256_mulx_u32(&x607, &x608, x7, (arg1[0]));
+  fiat_p256_uint1 x608;
+  fiat_p256_addcarryx_u32(&x607, &x608, x606, x579, x582);
   uint32_t x609;
   fiat_p256_uint1 x610;
-  fiat_p256_addcarryx_u32(&x609, &x610, 0x0, x605, x608);
+  fiat_p256_addcarryx_u32(&x609, &x610, x608, 0x0, x580);
   uint32_t x611;
   fiat_p256_uint1 x612;
-  fiat_p256_addcarryx_u32(&x611, &x612, x610, x603, x606);
+  fiat_p256_addcarryx_u32(&x611, &x612, 0x0, x593, x561);
   uint32_t x613;
   fiat_p256_uint1 x614;
-  fiat_p256_addcarryx_u32(&x613, &x614, x612, x601, x604);
+  fiat_p256_addcarryx_u32(&x613, &x614, x612, x595, x563);
   uint32_t x615;
   fiat_p256_uint1 x616;
-  fiat_p256_addcarryx_u32(&x615, &x616, x614, x599, x602);
+  fiat_p256_addcarryx_u32(&x615, &x616, x614, x597, x565);
   uint32_t x617;
   fiat_p256_uint1 x618;
-  fiat_p256_addcarryx_u32(&x617, &x618, x616, x597, x600);
+  fiat_p256_addcarryx_u32(&x617, &x618, x616, x599, x567);
   uint32_t x619;
   fiat_p256_uint1 x620;
-  fiat_p256_addcarryx_u32(&x619, &x620, x618, x595, x598);
+  fiat_p256_addcarryx_u32(&x619, &x620, x618, x601, x569);
   uint32_t x621;
   fiat_p256_uint1 x622;
-  fiat_p256_addcarryx_u32(&x621, &x622, x620, x593, x596);
+  fiat_p256_addcarryx_u32(&x621, &x622, x620, x603, x571);
   uint32_t x623;
   fiat_p256_uint1 x624;
-  fiat_p256_addcarryx_u32(&x623, &x624, x622, 0x0, x594);
+  fiat_p256_addcarryx_u32(&x623, &x624, x622, x605, x573);
   uint32_t x625;
   fiat_p256_uint1 x626;
-  fiat_p256_addcarryx_u32(&x625, &x626, 0x0, x607, x575);
+  fiat_p256_addcarryx_u32(&x625, &x626, x624, x607, x575);
   uint32_t x627;
   fiat_p256_uint1 x628;
   fiat_p256_addcarryx_u32(&x627, &x628, x626, x609, x577);
   uint32_t x629;
-  fiat_p256_uint1 x630;
-  fiat_p256_addcarryx_u32(&x629, &x630, x628, x611, x579);
+  uint32_t x630;
+  fiat_p256_mulx_u32(&x629, &x630, x611, UINT32_C(0xffffffff));
   uint32_t x631;
-  fiat_p256_uint1 x632;
-  fiat_p256_addcarryx_u32(&x631, &x632, x630, x613, x581);
+  uint32_t x632;
+  fiat_p256_mulx_u32(&x631, &x632, x611, UINT32_C(0xffffffff));
   uint32_t x633;
-  fiat_p256_uint1 x634;
-  fiat_p256_addcarryx_u32(&x633, &x634, x632, x615, x583);
+  uint32_t x634;
+  fiat_p256_mulx_u32(&x633, &x634, x611, UINT32_C(0xffffffff));
   uint32_t x635;
-  fiat_p256_uint1 x636;
-  fiat_p256_addcarryx_u32(&x635, &x636, x634, x617, x585);
+  uint32_t x636;
+  fiat_p256_mulx_u32(&x635, &x636, x611, UINT32_C(0xffffffff));
   uint32_t x637;
   fiat_p256_uint1 x638;
-  fiat_p256_addcarryx_u32(&x637, &x638, x636, x619, x587);
+  fiat_p256_addcarryx_u32(&x637, &x638, 0x0, x633, x636);
   uint32_t x639;
   fiat_p256_uint1 x640;
-  fiat_p256_addcarryx_u32(&x639, &x640, x638, x621, x589);
+  fiat_p256_addcarryx_u32(&x639, &x640, x638, x631, x634);
   uint32_t x641;
   fiat_p256_uint1 x642;
-  fiat_p256_addcarryx_u32(&x641, &x642, x640, x623, x591);
+  fiat_p256_addcarryx_u32(&x641, &x642, x640, 0x0, x632);
   uint32_t x643;
-  uint32_t x644;
-  fiat_p256_mulx_u32(&x643, &x644, x625, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x644;
+  fiat_p256_addcarryx_u32(&x643, &x644, 0x0, x635, x611);
   uint32_t x645;
-  uint32_t x646;
-  fiat_p256_mulx_u32(&x645, &x646, x625, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x646;
+  fiat_p256_addcarryx_u32(&x645, &x646, x644, x637, x613);
   uint32_t x647;
-  uint32_t x648;
-  fiat_p256_mulx_u32(&x647, &x648, x625, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x648;
+  fiat_p256_addcarryx_u32(&x647, &x648, x646, x639, x615);
   uint32_t x649;
-  uint32_t x650;
-  fiat_p256_mulx_u32(&x649, &x650, x625, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x650;
+  fiat_p256_addcarryx_u32(&x649, &x650, x648, x641, x617);
   uint32_t x651;
   fiat_p256_uint1 x652;
-  fiat_p256_addcarryx_u32(&x651, &x652, 0x0, x647, x650);
+  fiat_p256_addcarryx_u32(&x651, &x652, x650, 0x0, x619);
   uint32_t x653;
   fiat_p256_uint1 x654;
-  fiat_p256_addcarryx_u32(&x653, &x654, x652, x645, x648);
+  fiat_p256_addcarryx_u32(&x653, &x654, x652, 0x0, x621);
   uint32_t x655;
   fiat_p256_uint1 x656;
-  fiat_p256_addcarryx_u32(&x655, &x656, x654, 0x0, x646);
+  fiat_p256_addcarryx_u32(&x655, &x656, x654, x611, x623);
   uint32_t x657;
   fiat_p256_uint1 x658;
-  fiat_p256_addcarryx_u32(&x657, &x658, x656, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x657, &x658, x656, x629, x625);
   uint32_t x659;
   fiat_p256_uint1 x660;
-  fiat_p256_addcarryx_u32(&x659, &x660, 0x0, x649, x625);
+  fiat_p256_addcarryx_u32(&x659, &x660, x658, x630, x627);
   uint32_t x661;
   fiat_p256_uint1 x662;
-  fiat_p256_addcarryx_u32(&x661, &x662, x660, x651, x627);
+  fiat_p256_addcarryx_u32(&x661, &x662, x660, 0x0, x628);
   uint32_t x663;
   fiat_p256_uint1 x664;
-  fiat_p256_addcarryx_u32(&x663, &x664, x662, x653, x629);
+  fiat_p256_subborrowx_u32(&x663, &x664, 0x0, x645, UINT32_C(0xffffffff));
   uint32_t x665;
   fiat_p256_uint1 x666;
-  fiat_p256_addcarryx_u32(&x665, &x666, x664, x655, x631);
+  fiat_p256_subborrowx_u32(&x665, &x666, x664, x647, UINT32_C(0xffffffff));
   uint32_t x667;
   fiat_p256_uint1 x668;
-  fiat_p256_addcarryx_u32(&x667, &x668, x666, (fiat_p256_uint1)x657, x633);
+  fiat_p256_subborrowx_u32(&x667, &x668, x666, x649, UINT32_C(0xffffffff));
   uint32_t x669;
   fiat_p256_uint1 x670;
-  fiat_p256_addcarryx_u32(&x669, &x670, x668, 0x0, x635);
+  fiat_p256_subborrowx_u32(&x669, &x670, x668, x651, 0x0);
   uint32_t x671;
   fiat_p256_uint1 x672;
-  fiat_p256_addcarryx_u32(&x671, &x672, x670, x625, x637);
+  fiat_p256_subborrowx_u32(&x671, &x672, x670, x653, 0x0);
   uint32_t x673;
   fiat_p256_uint1 x674;
-  fiat_p256_addcarryx_u32(&x673, &x674, x672, x643, x639);
+  fiat_p256_subborrowx_u32(&x673, &x674, x672, x655, 0x0);
   uint32_t x675;
   fiat_p256_uint1 x676;
-  fiat_p256_addcarryx_u32(&x675, &x676, x674, x644, x641);
+  fiat_p256_subborrowx_u32(&x675, &x676, x674, x657, 0x1);
   uint32_t x677;
   fiat_p256_uint1 x678;
-  fiat_p256_addcarryx_u32(&x677, &x678, x676, 0x0, x642);
+  fiat_p256_subborrowx_u32(&x677, &x678, x676, x659, UINT32_C(0xffffffff));
   uint32_t x679;
   fiat_p256_uint1 x680;
-  fiat_p256_subborrowx_u32(&x679, &x680, 0x0, x661, UINT32_C(0xffffffff));
+  fiat_p256_subborrowx_u32(&x679, &x680, x678, x661, 0x0);
   uint32_t x681;
-  fiat_p256_uint1 x682;
-  fiat_p256_subborrowx_u32(&x681, &x682, x680, x663, UINT32_C(0xffffffff));
+  fiat_p256_cmovznz_u32(&x681, x680, x663, x645);
+  uint32_t x682;
+  fiat_p256_cmovznz_u32(&x682, x680, x665, x647);
   uint32_t x683;
-  fiat_p256_uint1 x684;
-  fiat_p256_subborrowx_u32(&x683, &x684, x682, x665, UINT32_C(0xffffffff));
+  fiat_p256_cmovznz_u32(&x683, x680, x667, x649);
+  uint32_t x684;
+  fiat_p256_cmovznz_u32(&x684, x680, x669, x651);
   uint32_t x685;
-  fiat_p256_uint1 x686;
-  fiat_p256_subborrowx_u32(&x685, &x686, x684, x667, 0x0);
+  fiat_p256_cmovznz_u32(&x685, x680, x671, x653);
+  uint32_t x686;
+  fiat_p256_cmovznz_u32(&x686, x680, x673, x655);
   uint32_t x687;
-  fiat_p256_uint1 x688;
-  fiat_p256_subborrowx_u32(&x687, &x688, x686, x669, 0x0);
-  uint32_t x689;
-  fiat_p256_uint1 x690;
-  fiat_p256_subborrowx_u32(&x689, &x690, x688, x671, 0x0);
-  uint32_t x691;
-  fiat_p256_uint1 x692;
-  fiat_p256_subborrowx_u32(&x691, &x692, x690, x673, 0x1);
-  uint32_t x693;
-  fiat_p256_uint1 x694;
-  fiat_p256_subborrowx_u32(&x693, &x694, x692, x675, UINT32_C(0xffffffff));
-  uint32_t x695;
-  fiat_p256_uint1 x696;
-  fiat_p256_subborrowx_u32(&x695, &x696, x694, x677, 0x0);
-  uint32_t x697;
-  fiat_p256_cmovznz_u32(&x697, x696, x679, x661);
-  uint32_t x698;
-  fiat_p256_cmovznz_u32(&x698, x696, x681, x663);
-  uint32_t x699;
-  fiat_p256_cmovznz_u32(&x699, x696, x683, x665);
-  uint32_t x700;
-  fiat_p256_cmovznz_u32(&x700, x696, x685, x667);
-  uint32_t x701;
-  fiat_p256_cmovznz_u32(&x701, x696, x687, x669);
-  uint32_t x702;
-  fiat_p256_cmovznz_u32(&x702, x696, x689, x671);
-  uint32_t x703;
-  fiat_p256_cmovznz_u32(&x703, x696, x691, x673);
-  uint32_t x704;
-  fiat_p256_cmovznz_u32(&x704, x696, x693, x675);
-  out1[0] = x697;
-  out1[1] = x698;
-  out1[2] = x699;
-  out1[3] = x700;
-  out1[4] = x701;
-  out1[5] = x702;
-  out1[6] = x703;
-  out1[7] = x704;
+  fiat_p256_cmovznz_u32(&x687, x680, x675, x657);
+  uint32_t x688;
+  fiat_p256_cmovznz_u32(&x688, x680, x677, x659);
+  out1[0] = x681;
+  out1[1] = x682;
+  out1[2] = x683;
+  out1[3] = x684;
+  out1[4] = x685;
+  out1[5] = x686;
+  out1[6] = x687;
+  out1[7] = x688;
 }
 
 /*
@@ -2476,646 +2428,580 @@ static void fiat_p256_from_montgomery(uint32_t out1[8], const uint32_t arg1[8])
   fiat_p256_addcarryx_u32(&x12, &x13, x11, x4, x7);
   uint32_t x14;
   fiat_p256_uint1 x15;
-  fiat_p256_addcarryx_u32(&x14, &x15, x13, 0x0, x5);
+  fiat_p256_addcarryx_u32(&x14, &x15, 0x0, x8, x1);
   uint32_t x16;
   fiat_p256_uint1 x17;
-  fiat_p256_addcarryx_u32(&x16, &x17, 0x0, x8, x1);
+  fiat_p256_addcarryx_u32(&x16, &x17, x15, x10, 0x0);
   uint32_t x18;
   fiat_p256_uint1 x19;
-  fiat_p256_addcarryx_u32(&x18, &x19, x17, x10, 0x0);
+  fiat_p256_addcarryx_u32(&x18, &x19, x17, x12, 0x0);
   uint32_t x20;
   fiat_p256_uint1 x21;
-  fiat_p256_addcarryx_u32(&x20, &x21, x19, x12, 0x0);
+  fiat_p256_addcarryx_u32(&x20, &x21, x13, 0x0, x5);
   uint32_t x22;
   fiat_p256_uint1 x23;
-  fiat_p256_addcarryx_u32(&x22, &x23, x21, x14, 0x0);
+  fiat_p256_addcarryx_u32(&x22, &x23, x19, x20, 0x0);
   uint32_t x24;
   fiat_p256_uint1 x25;
-  fiat_p256_addcarryx_u32(&x24, &x25, x15, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x24, &x25, 0x0, (arg1[1]), x16);
   uint32_t x26;
   fiat_p256_uint1 x27;
-  fiat_p256_addcarryx_u32(&x26, &x27, x23, (fiat_p256_uint1)x24, 0x0);
+  fiat_p256_addcarryx_u32(&x26, &x27, x25, 0x0, x18);
   uint32_t x28;
   fiat_p256_uint1 x29;
-  fiat_p256_addcarryx_u32(&x28, &x29, 0x0, (arg1[1]), x18);
+  fiat_p256_addcarryx_u32(&x28, &x29, x27, 0x0, x22);
   uint32_t x30;
-  fiat_p256_uint1 x31;
-  fiat_p256_addcarryx_u32(&x30, &x31, x29, 0x0, x20);
+  uint32_t x31;
+  fiat_p256_mulx_u32(&x30, &x31, x24, UINT32_C(0xffffffff));
   uint32_t x32;
-  fiat_p256_uint1 x33;
-  fiat_p256_addcarryx_u32(&x32, &x33, x31, 0x0, x22);
+  uint32_t x33;
+  fiat_p256_mulx_u32(&x32, &x33, x24, UINT32_C(0xffffffff));
   uint32_t x34;
-  fiat_p256_uint1 x35;
-  fiat_p256_addcarryx_u32(&x34, &x35, x33, 0x0, (fiat_p256_uint1)x26);
+  uint32_t x35;
+  fiat_p256_mulx_u32(&x34, &x35, x24, UINT32_C(0xffffffff));
   uint32_t x36;
-  fiat_p256_uint1 x37;
-  fiat_p256_addcarryx_u32(&x36, &x37, x27, 0x0, 0x0);
+  uint32_t x37;
+  fiat_p256_mulx_u32(&x36, &x37, x24, UINT32_C(0xffffffff));
   uint32_t x38;
   fiat_p256_uint1 x39;
-  fiat_p256_addcarryx_u32(&x38, &x39, x35, 0x0, (fiat_p256_uint1)x36);
+  fiat_p256_addcarryx_u32(&x38, &x39, 0x0, x34, x37);
   uint32_t x40;
   fiat_p256_uint1 x41;
-  fiat_p256_addcarryx_u32(&x40, &x41, x39, 0x0, x1);
+  fiat_p256_addcarryx_u32(&x40, &x41, x39, x32, x35);
   uint32_t x42;
   fiat_p256_uint1 x43;
-  fiat_p256_addcarryx_u32(&x42, &x43, x41, 0x0, x2);
+  fiat_p256_addcarryx_u32(&x42, &x43, 0x0, x36, x24);
   uint32_t x44;
   fiat_p256_uint1 x45;
-  fiat_p256_addcarryx_u32(&x44, &x45, x43, 0x0, x3);
+  fiat_p256_addcarryx_u32(&x44, &x45, x43, x38, x26);
   uint32_t x46;
-  uint32_t x47;
-  fiat_p256_mulx_u32(&x46, &x47, x28, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x47;
+  fiat_p256_addcarryx_u32(&x46, &x47, x45, x40, x28);
   uint32_t x48;
-  uint32_t x49;
-  fiat_p256_mulx_u32(&x48, &x49, x28, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x49;
+  fiat_p256_addcarryx_u32(&x48, &x49, x23, 0x0, 0x0);
   uint32_t x50;
-  uint32_t x51;
-  fiat_p256_mulx_u32(&x50, &x51, x28, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x51;
+  fiat_p256_addcarryx_u32(&x50, &x51, x29, 0x0, (fiat_p256_uint1)x48);
   uint32_t x52;
-  uint32_t x53;
-  fiat_p256_mulx_u32(&x52, &x53, x28, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x53;
+  fiat_p256_addcarryx_u32(&x52, &x53, x41, 0x0, x33);
   uint32_t x54;
   fiat_p256_uint1 x55;
-  fiat_p256_addcarryx_u32(&x54, &x55, 0x0, x50, x53);
+  fiat_p256_addcarryx_u32(&x54, &x55, x47, x52, x50);
   uint32_t x56;
   fiat_p256_uint1 x57;
-  fiat_p256_addcarryx_u32(&x56, &x57, x55, x48, x51);
+  fiat_p256_addcarryx_u32(&x56, &x57, 0x0, x24, x2);
   uint32_t x58;
   fiat_p256_uint1 x59;
-  fiat_p256_addcarryx_u32(&x58, &x59, x57, 0x0, x49);
+  fiat_p256_addcarryx_u32(&x58, &x59, x57, x30, x3);
   uint32_t x60;
   fiat_p256_uint1 x61;
-  fiat_p256_addcarryx_u32(&x60, &x61, 0x0, x52, x28);
+  fiat_p256_addcarryx_u32(&x60, &x61, 0x0, (arg1[2]), x44);
   uint32_t x62;
   fiat_p256_uint1 x63;
-  fiat_p256_addcarryx_u32(&x62, &x63, x61, x54, x30);
+  fiat_p256_addcarryx_u32(&x62, &x63, x61, 0x0, x46);
   uint32_t x64;
   fiat_p256_uint1 x65;
-  fiat_p256_addcarryx_u32(&x64, &x65, x63, x56, x32);
+  fiat_p256_addcarryx_u32(&x64, &x65, x63, 0x0, x54);
   uint32_t x66;
-  fiat_p256_uint1 x67;
-  fiat_p256_addcarryx_u32(&x66, &x67, x65, x58, x34);
+  uint32_t x67;
+  fiat_p256_mulx_u32(&x66, &x67, x60, UINT32_C(0xffffffff));
   uint32_t x68;
-  fiat_p256_uint1 x69;
-  fiat_p256_addcarryx_u32(&x68, &x69, x59, 0x0, 0x0);
+  uint32_t x69;
+  fiat_p256_mulx_u32(&x68, &x69, x60, UINT32_C(0xffffffff));
   uint32_t x70;
-  fiat_p256_uint1 x71;
-  fiat_p256_addcarryx_u32(&x70, &x71, x67, (fiat_p256_uint1)x68, (fiat_p256_uint1)x38);
+  uint32_t x71;
+  fiat_p256_mulx_u32(&x70, &x71, x60, UINT32_C(0xffffffff));
   uint32_t x72;
-  fiat_p256_uint1 x73;
-  fiat_p256_addcarryx_u32(&x72, &x73, x71, 0x0, x40);
+  uint32_t x73;
+  fiat_p256_mulx_u32(&x72, &x73, x60, UINT32_C(0xffffffff));
   uint32_t x74;
   fiat_p256_uint1 x75;
-  fiat_p256_addcarryx_u32(&x74, &x75, x73, x28, x42);
+  fiat_p256_addcarryx_u32(&x74, &x75, 0x0, x70, x73);
   uint32_t x76;
   fiat_p256_uint1 x77;
-  fiat_p256_addcarryx_u32(&x76, &x77, x75, x46, x44);
+  fiat_p256_addcarryx_u32(&x76, &x77, x75, x68, x71);
   uint32_t x78;
   fiat_p256_uint1 x79;
-  fiat_p256_addcarryx_u32(&x78, &x79, x45, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x78, &x79, 0x0, x72, x60);
   uint32_t x80;
   fiat_p256_uint1 x81;
-  fiat_p256_addcarryx_u32(&x80, &x81, x77, x47, (fiat_p256_uint1)x78);
+  fiat_p256_addcarryx_u32(&x80, &x81, x79, x74, x62);
   uint32_t x82;
   fiat_p256_uint1 x83;
-  fiat_p256_addcarryx_u32(&x82, &x83, 0x0, (arg1[2]), x62);
+  fiat_p256_addcarryx_u32(&x82, &x83, x81, x76, x64);
   uint32_t x84;
   fiat_p256_uint1 x85;
-  fiat_p256_addcarryx_u32(&x84, &x85, x83, 0x0, x64);
+  fiat_p256_addcarryx_u32(&x84, &x85, x55, 0x0, 0x0);
   uint32_t x86;
   fiat_p256_uint1 x87;
-  fiat_p256_addcarryx_u32(&x86, &x87, x85, 0x0, x66);
+  fiat_p256_addcarryx_u32(&x86, &x87, x65, 0x0, (fiat_p256_uint1)x84);
   uint32_t x88;
   fiat_p256_uint1 x89;
-  fiat_p256_addcarryx_u32(&x88, &x89, x87, 0x0, (fiat_p256_uint1)x70);
+  fiat_p256_addcarryx_u32(&x88, &x89, x77, 0x0, x69);
   uint32_t x90;
   fiat_p256_uint1 x91;
-  fiat_p256_addcarryx_u32(&x90, &x91, x89, 0x0, x72);
+  fiat_p256_addcarryx_u32(&x90, &x91, x83, x88, x86);
   uint32_t x92;
   fiat_p256_uint1 x93;
-  fiat_p256_addcarryx_u32(&x92, &x93, x91, 0x0, x74);
+  fiat_p256_addcarryx_u32(&x92, &x93, x91, 0x0, x1);
   uint32_t x94;
   fiat_p256_uint1 x95;
-  fiat_p256_addcarryx_u32(&x94, &x95, x93, 0x0, x76);
+  fiat_p256_addcarryx_u32(&x94, &x95, x93, 0x0, x56);
   uint32_t x96;
   fiat_p256_uint1 x97;
-  fiat_p256_addcarryx_u32(&x96, &x97, x95, 0x0, x80);
+  fiat_p256_addcarryx_u32(&x96, &x97, x95, x60, x58);
   uint32_t x98;
   fiat_p256_uint1 x99;
-  fiat_p256_addcarryx_u32(&x98, &x99, x81, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x98, &x99, x59, x31, 0x0);
   uint32_t x100;
   fiat_p256_uint1 x101;
-  fiat_p256_addcarryx_u32(&x100, &x101, x97, 0x0, (fiat_p256_uint1)x98);
+  fiat_p256_addcarryx_u32(&x100, &x101, x97, x66, x98);
   uint32_t x102;
-  uint32_t x103;
-  fiat_p256_mulx_u32(&x102, &x103, x82, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x103;
+  fiat_p256_addcarryx_u32(&x102, &x103, 0x0, (arg1[3]), x80);
   uint32_t x104;
-  uint32_t x105;
-  fiat_p256_mulx_u32(&x104, &x105, x82, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x105;
+  fiat_p256_addcarryx_u32(&x104, &x105, x103, 0x0, x82);
   uint32_t x106;
-  uint32_t x107;
-  fiat_p256_mulx_u32(&x106, &x107, x82, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x107;
+  fiat_p256_addcarryx_u32(&x106, &x107, x105, 0x0, x90);
   uint32_t x108;
-  uint32_t x109;
-  fiat_p256_mulx_u32(&x108, &x109, x82, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x109;
+  fiat_p256_addcarryx_u32(&x108, &x109, x107, 0x0, x92);
   uint32_t x110;
   fiat_p256_uint1 x111;
-  fiat_p256_addcarryx_u32(&x110, &x111, 0x0, x106, x109);
+  fiat_p256_addcarryx_u32(&x110, &x111, x109, 0x0, x94);
   uint32_t x112;
   fiat_p256_uint1 x113;
-  fiat_p256_addcarryx_u32(&x112, &x113, x111, x104, x107);
+  fiat_p256_addcarryx_u32(&x112, &x113, x111, 0x0, x96);
   uint32_t x114;
   fiat_p256_uint1 x115;
-  fiat_p256_addcarryx_u32(&x114, &x115, x113, 0x0, x105);
+  fiat_p256_addcarryx_u32(&x114, &x115, x113, 0x0, x100);
   uint32_t x116;
   fiat_p256_uint1 x117;
-  fiat_p256_addcarryx_u32(&x116, &x117, 0x0, x108, x82);
+  fiat_p256_addcarryx_u32(&x116, &x117, x101, x67, 0x0);
   uint32_t x118;
   fiat_p256_uint1 x119;
-  fiat_p256_addcarryx_u32(&x118, &x119, x117, x110, x84);
+  fiat_p256_addcarryx_u32(&x118, &x119, x115, 0x0, x116);
   uint32_t x120;
-  fiat_p256_uint1 x121;
-  fiat_p256_addcarryx_u32(&x120, &x121, x119, x112, x86);
+  uint32_t x121;
+  fiat_p256_mulx_u32(&x120, &x121, x102, UINT32_C(0xffffffff));
   uint32_t x122;
-  fiat_p256_uint1 x123;
-  fiat_p256_addcarryx_u32(&x122, &x123, x121, x114, x88);
+  uint32_t x123;
+  fiat_p256_mulx_u32(&x122, &x123, x102, UINT32_C(0xffffffff));
   uint32_t x124;
-  fiat_p256_uint1 x125;
-  fiat_p256_addcarryx_u32(&x124, &x125, x115, 0x0, 0x0);
+  uint32_t x125;
+  fiat_p256_mulx_u32(&x124, &x125, x102, UINT32_C(0xffffffff));
   uint32_t x126;
-  fiat_p256_uint1 x127;
-  fiat_p256_addcarryx_u32(&x126, &x127, x123, (fiat_p256_uint1)x124, x90);
+  uint32_t x127;
+  fiat_p256_mulx_u32(&x126, &x127, x102, UINT32_C(0xffffffff));
   uint32_t x128;
   fiat_p256_uint1 x129;
-  fiat_p256_addcarryx_u32(&x128, &x129, x127, 0x0, x92);
+  fiat_p256_addcarryx_u32(&x128, &x129, 0x0, x124, x127);
   uint32_t x130;
   fiat_p256_uint1 x131;
-  fiat_p256_addcarryx_u32(&x130, &x131, x129, x82, x94);
+  fiat_p256_addcarryx_u32(&x130, &x131, x129, x122, x125);
   uint32_t x132;
   fiat_p256_uint1 x133;
-  fiat_p256_addcarryx_u32(&x132, &x133, x131, x102, x96);
+  fiat_p256_addcarryx_u32(&x132, &x133, 0x0, x126, x102);
   uint32_t x134;
   fiat_p256_uint1 x135;
-  fiat_p256_addcarryx_u32(&x134, &x135, x133, x103, (fiat_p256_uint1)x100);
+  fiat_p256_addcarryx_u32(&x134, &x135, x133, x128, x104);
   uint32_t x136;
   fiat_p256_uint1 x137;
-  fiat_p256_addcarryx_u32(&x136, &x137, 0x0, (arg1[3]), x118);
+  fiat_p256_addcarryx_u32(&x136, &x137, x135, x130, x106);
   uint32_t x138;
   fiat_p256_uint1 x139;
-  fiat_p256_addcarryx_u32(&x138, &x139, x137, 0x0, x120);
+  fiat_p256_addcarryx_u32(&x138, &x139, x131, 0x0, x123);
   uint32_t x140;
   fiat_p256_uint1 x141;
-  fiat_p256_addcarryx_u32(&x140, &x141, x139, 0x0, x122);
+  fiat_p256_addcarryx_u32(&x140, &x141, x137, x138, x108);
   uint32_t x142;
   fiat_p256_uint1 x143;
-  fiat_p256_addcarryx_u32(&x142, &x143, x141, 0x0, x126);
+  fiat_p256_addcarryx_u32(&x142, &x143, x141, 0x0, x110);
   uint32_t x144;
   fiat_p256_uint1 x145;
-  fiat_p256_addcarryx_u32(&x144, &x145, x143, 0x0, x128);
+  fiat_p256_addcarryx_u32(&x144, &x145, x143, 0x0, x112);
   uint32_t x146;
   fiat_p256_uint1 x147;
-  fiat_p256_addcarryx_u32(&x146, &x147, x145, 0x0, x130);
+  fiat_p256_addcarryx_u32(&x146, &x147, x145, x102, x114);
   uint32_t x148;
   fiat_p256_uint1 x149;
-  fiat_p256_addcarryx_u32(&x148, &x149, x147, 0x0, x132);
+  fiat_p256_addcarryx_u32(&x148, &x149, x147, x120, x118);
   uint32_t x150;
   fiat_p256_uint1 x151;
-  fiat_p256_addcarryx_u32(&x150, &x151, x149, 0x0, x134);
+  fiat_p256_addcarryx_u32(&x150, &x151, x119, 0x0, 0x0);
   uint32_t x152;
   fiat_p256_uint1 x153;
-  fiat_p256_addcarryx_u32(&x152, &x153, x135, 0x0, x101);
+  fiat_p256_addcarryx_u32(&x152, &x153, x149, x121, (fiat_p256_uint1)x150);
   uint32_t x154;
   fiat_p256_uint1 x155;
-  fiat_p256_addcarryx_u32(&x154, &x155, x151, 0x0, (fiat_p256_uint1)x152);
+  fiat_p256_addcarryx_u32(&x154, &x155, 0x0, (arg1[4]), x134);
   uint32_t x156;
-  uint32_t x157;
-  fiat_p256_mulx_u32(&x156, &x157, x136, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x157;
+  fiat_p256_addcarryx_u32(&x156, &x157, x155, 0x0, x136);
   uint32_t x158;
-  uint32_t x159;
-  fiat_p256_mulx_u32(&x158, &x159, x136, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x159;
+  fiat_p256_addcarryx_u32(&x158, &x159, x157, 0x0, x140);
   uint32_t x160;
-  uint32_t x161;
-  fiat_p256_mulx_u32(&x160, &x161, x136, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x161;
+  fiat_p256_addcarryx_u32(&x160, &x161, x159, 0x0, x142);
   uint32_t x162;
-  uint32_t x163;
-  fiat_p256_mulx_u32(&x162, &x163, x136, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x163;
+  fiat_p256_addcarryx_u32(&x162, &x163, x161, 0x0, x144);
   uint32_t x164;
   fiat_p256_uint1 x165;
-  fiat_p256_addcarryx_u32(&x164, &x165, 0x0, x160, x163);
+  fiat_p256_addcarryx_u32(&x164, &x165, x163, 0x0, x146);
   uint32_t x166;
   fiat_p256_uint1 x167;
-  fiat_p256_addcarryx_u32(&x166, &x167, x165, x158, x161);
+  fiat_p256_addcarryx_u32(&x166, &x167, x165, 0x0, x148);
   uint32_t x168;
   fiat_p256_uint1 x169;
-  fiat_p256_addcarryx_u32(&x168, &x169, x167, 0x0, x159);
+  fiat_p256_addcarryx_u32(&x168, &x169, x167, 0x0, x152);
   uint32_t x170;
-  fiat_p256_uint1 x171;
-  fiat_p256_addcarryx_u32(&x170, &x171, 0x0, x162, x136);
+  uint32_t x171;
+  fiat_p256_mulx_u32(&x170, &x171, x154, UINT32_C(0xffffffff));
   uint32_t x172;
-  fiat_p256_uint1 x173;
-  fiat_p256_addcarryx_u32(&x172, &x173, x171, x164, x138);
+  uint32_t x173;
+  fiat_p256_mulx_u32(&x172, &x173, x154, UINT32_C(0xffffffff));
   uint32_t x174;
-  fiat_p256_uint1 x175;
-  fiat_p256_addcarryx_u32(&x174, &x175, x173, x166, x140);
+  uint32_t x175;
+  fiat_p256_mulx_u32(&x174, &x175, x154, UINT32_C(0xffffffff));
   uint32_t x176;
-  fiat_p256_uint1 x177;
-  fiat_p256_addcarryx_u32(&x176, &x177, x175, x168, x142);
+  uint32_t x177;
+  fiat_p256_mulx_u32(&x176, &x177, x154, UINT32_C(0xffffffff));
   uint32_t x178;
   fiat_p256_uint1 x179;
-  fiat_p256_addcarryx_u32(&x178, &x179, x169, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x178, &x179, 0x0, x174, x177);
   uint32_t x180;
   fiat_p256_uint1 x181;
-  fiat_p256_addcarryx_u32(&x180, &x181, x177, (fiat_p256_uint1)x178, x144);
+  fiat_p256_addcarryx_u32(&x180, &x181, x179, x172, x175);
   uint32_t x182;
   fiat_p256_uint1 x183;
-  fiat_p256_addcarryx_u32(&x182, &x183, x181, 0x0, x146);
+  fiat_p256_addcarryx_u32(&x182, &x183, 0x0, x176, x154);
   uint32_t x184;
   fiat_p256_uint1 x185;
-  fiat_p256_addcarryx_u32(&x184, &x185, x183, x136, x148);
+  fiat_p256_addcarryx_u32(&x184, &x185, x183, x178, x156);
   uint32_t x186;
   fiat_p256_uint1 x187;
-  fiat_p256_addcarryx_u32(&x186, &x187, x185, x156, x150);
+  fiat_p256_addcarryx_u32(&x186, &x187, x185, x180, x158);
   uint32_t x188;
   fiat_p256_uint1 x189;
-  fiat_p256_addcarryx_u32(&x188, &x189, x187, x157, (fiat_p256_uint1)x154);
+  fiat_p256_addcarryx_u32(&x188, &x189, x181, 0x0, x173);
   uint32_t x190;
   fiat_p256_uint1 x191;
-  fiat_p256_addcarryx_u32(&x190, &x191, 0x0, (arg1[4]), x172);
+  fiat_p256_addcarryx_u32(&x190, &x191, x187, x188, x160);
   uint32_t x192;
   fiat_p256_uint1 x193;
-  fiat_p256_addcarryx_u32(&x192, &x193, x191, 0x0, x174);
+  fiat_p256_addcarryx_u32(&x192, &x193, x191, 0x0, x162);
   uint32_t x194;
   fiat_p256_uint1 x195;
-  fiat_p256_addcarryx_u32(&x194, &x195, x193, 0x0, x176);
+  fiat_p256_addcarryx_u32(&x194, &x195, x193, 0x0, x164);
   uint32_t x196;
   fiat_p256_uint1 x197;
-  fiat_p256_addcarryx_u32(&x196, &x197, x195, 0x0, x180);
+  fiat_p256_addcarryx_u32(&x196, &x197, x195, x154, x166);
   uint32_t x198;
   fiat_p256_uint1 x199;
-  fiat_p256_addcarryx_u32(&x198, &x199, x197, 0x0, x182);
+  fiat_p256_addcarryx_u32(&x198, &x199, x197, x170, x168);
   uint32_t x200;
   fiat_p256_uint1 x201;
-  fiat_p256_addcarryx_u32(&x200, &x201, x199, 0x0, x184);
+  fiat_p256_addcarryx_u32(&x200, &x201, x153, 0x0, 0x0);
   uint32_t x202;
   fiat_p256_uint1 x203;
-  fiat_p256_addcarryx_u32(&x202, &x203, x201, 0x0, x186);
+  fiat_p256_addcarryx_u32(&x202, &x203, x169, 0x0, (fiat_p256_uint1)x200);
   uint32_t x204;
   fiat_p256_uint1 x205;
-  fiat_p256_addcarryx_u32(&x204, &x205, x203, 0x0, x188);
+  fiat_p256_addcarryx_u32(&x204, &x205, x199, x171, x202);
   uint32_t x206;
   fiat_p256_uint1 x207;
-  fiat_p256_addcarryx_u32(&x206, &x207, x189, 0x0, x155);
+  fiat_p256_addcarryx_u32(&x206, &x207, 0x0, (arg1[5]), x184);
   uint32_t x208;
   fiat_p256_uint1 x209;
-  fiat_p256_addcarryx_u32(&x208, &x209, x205, 0x0, (fiat_p256_uint1)x206);
+  fiat_p256_addcarryx_u32(&x208, &x209, x207, 0x0, x186);
   uint32_t x210;
-  uint32_t x211;
-  fiat_p256_mulx_u32(&x210, &x211, x190, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x211;
+  fiat_p256_addcarryx_u32(&x210, &x211, x209, 0x0, x190);
   uint32_t x212;
-  uint32_t x213;
-  fiat_p256_mulx_u32(&x212, &x213, x190, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x213;
+  fiat_p256_addcarryx_u32(&x212, &x213, x211, 0x0, x192);
   uint32_t x214;
-  uint32_t x215;
-  fiat_p256_mulx_u32(&x214, &x215, x190, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x215;
+  fiat_p256_addcarryx_u32(&x214, &x215, x213, 0x0, x194);
   uint32_t x216;
-  uint32_t x217;
-  fiat_p256_mulx_u32(&x216, &x217, x190, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x217;
+  fiat_p256_addcarryx_u32(&x216, &x217, x215, 0x0, x196);
   uint32_t x218;
   fiat_p256_uint1 x219;
-  fiat_p256_addcarryx_u32(&x218, &x219, 0x0, x214, x217);
+  fiat_p256_addcarryx_u32(&x218, &x219, x217, 0x0, x198);
   uint32_t x220;
   fiat_p256_uint1 x221;
-  fiat_p256_addcarryx_u32(&x220, &x221, x219, x212, x215);
+  fiat_p256_addcarryx_u32(&x220, &x221, x219, 0x0, x204);
   uint32_t x222;
-  fiat_p256_uint1 x223;
-  fiat_p256_addcarryx_u32(&x222, &x223, x221, 0x0, x213);
+  uint32_t x223;
+  fiat_p256_mulx_u32(&x222, &x223, x206, UINT32_C(0xffffffff));
   uint32_t x224;
-  fiat_p256_uint1 x225;
-  fiat_p256_addcarryx_u32(&x224, &x225, 0x0, x216, x190);
+  uint32_t x225;
+  fiat_p256_mulx_u32(&x224, &x225, x206, UINT32_C(0xffffffff));
   uint32_t x226;
-  fiat_p256_uint1 x227;
-  fiat_p256_addcarryx_u32(&x226, &x227, x225, x218, x192);
+  uint32_t x227;
+  fiat_p256_mulx_u32(&x226, &x227, x206, UINT32_C(0xffffffff));
   uint32_t x228;
-  fiat_p256_uint1 x229;
-  fiat_p256_addcarryx_u32(&x228, &x229, x227, x220, x194);
+  uint32_t x229;
+  fiat_p256_mulx_u32(&x228, &x229, x206, UINT32_C(0xffffffff));
   uint32_t x230;
   fiat_p256_uint1 x231;
-  fiat_p256_addcarryx_u32(&x230, &x231, x229, x222, x196);
+  fiat_p256_addcarryx_u32(&x230, &x231, 0x0, x226, x229);
   uint32_t x232;
   fiat_p256_uint1 x233;
-  fiat_p256_addcarryx_u32(&x232, &x233, x223, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x232, &x233, x231, x224, x227);
   uint32_t x234;
   fiat_p256_uint1 x235;
-  fiat_p256_addcarryx_u32(&x234, &x235, x231, (fiat_p256_uint1)x232, x198);
+  fiat_p256_addcarryx_u32(&x234, &x235, 0x0, x228, x206);
   uint32_t x236;
   fiat_p256_uint1 x237;
-  fiat_p256_addcarryx_u32(&x236, &x237, x235, 0x0, x200);
+  fiat_p256_addcarryx_u32(&x236, &x237, x235, x230, x208);
   uint32_t x238;
   fiat_p256_uint1 x239;
-  fiat_p256_addcarryx_u32(&x238, &x239, x237, x190, x202);
+  fiat_p256_addcarryx_u32(&x238, &x239, x237, x232, x210);
   uint32_t x240;
   fiat_p256_uint1 x241;
-  fiat_p256_addcarryx_u32(&x240, &x241, x239, x210, x204);
+  fiat_p256_addcarryx_u32(&x240, &x241, x233, 0x0, x225);
   uint32_t x242;
   fiat_p256_uint1 x243;
-  fiat_p256_addcarryx_u32(&x242, &x243, x241, x211, x208);
+  fiat_p256_addcarryx_u32(&x242, &x243, x239, x240, x212);
   uint32_t x244;
   fiat_p256_uint1 x245;
-  fiat_p256_addcarryx_u32(&x244, &x245, 0x0, (arg1[5]), x226);
+  fiat_p256_addcarryx_u32(&x244, &x245, x243, 0x0, x214);
   uint32_t x246;
   fiat_p256_uint1 x247;
-  fiat_p256_addcarryx_u32(&x246, &x247, x245, 0x0, x228);
+  fiat_p256_addcarryx_u32(&x246, &x247, x245, 0x0, x216);
   uint32_t x248;
   fiat_p256_uint1 x249;
-  fiat_p256_addcarryx_u32(&x248, &x249, x247, 0x0, x230);
+  fiat_p256_addcarryx_u32(&x248, &x249, x247, x206, x218);
   uint32_t x250;
   fiat_p256_uint1 x251;
-  fiat_p256_addcarryx_u32(&x250, &x251, x249, 0x0, x234);
+  fiat_p256_addcarryx_u32(&x250, &x251, x249, x222, x220);
   uint32_t x252;
   fiat_p256_uint1 x253;
-  fiat_p256_addcarryx_u32(&x252, &x253, x251, 0x0, x236);
+  fiat_p256_addcarryx_u32(&x252, &x253, x205, 0x0, 0x0);
   uint32_t x254;
   fiat_p256_uint1 x255;
-  fiat_p256_addcarryx_u32(&x254, &x255, x253, 0x0, x238);
+  fiat_p256_addcarryx_u32(&x254, &x255, x221, 0x0, (fiat_p256_uint1)x252);
   uint32_t x256;
   fiat_p256_uint1 x257;
-  fiat_p256_addcarryx_u32(&x256, &x257, x255, 0x0, x240);
+  fiat_p256_addcarryx_u32(&x256, &x257, x251, x223, x254);
   uint32_t x258;
   fiat_p256_uint1 x259;
-  fiat_p256_addcarryx_u32(&x258, &x259, x257, 0x0, x242);
+  fiat_p256_addcarryx_u32(&x258, &x259, 0x0, (arg1[6]), x236);
   uint32_t x260;
   fiat_p256_uint1 x261;
-  fiat_p256_addcarryx_u32(&x260, &x261, x243, 0x0, x209);
+  fiat_p256_addcarryx_u32(&x260, &x261, x259, 0x0, x238);
   uint32_t x262;
   fiat_p256_uint1 x263;
-  fiat_p256_addcarryx_u32(&x262, &x263, x259, 0x0, (fiat_p256_uint1)x260);
+  fiat_p256_addcarryx_u32(&x262, &x263, x261, 0x0, x242);
   uint32_t x264;
-  uint32_t x265;
-  fiat_p256_mulx_u32(&x264, &x265, x244, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x265;
+  fiat_p256_addcarryx_u32(&x264, &x265, x263, 0x0, x244);
   uint32_t x266;
-  uint32_t x267;
-  fiat_p256_mulx_u32(&x266, &x267, x244, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x267;
+  fiat_p256_addcarryx_u32(&x266, &x267, x265, 0x0, x246);
   uint32_t x268;
-  uint32_t x269;
-  fiat_p256_mulx_u32(&x268, &x269, x244, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x269;
+  fiat_p256_addcarryx_u32(&x268, &x269, x267, 0x0, x248);
   uint32_t x270;
-  uint32_t x271;
-  fiat_p256_mulx_u32(&x270, &x271, x244, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x271;
+  fiat_p256_addcarryx_u32(&x270, &x271, x269, 0x0, x250);
   uint32_t x272;
   fiat_p256_uint1 x273;
-  fiat_p256_addcarryx_u32(&x272, &x273, 0x0, x268, x271);
+  fiat_p256_addcarryx_u32(&x272, &x273, x271, 0x0, x256);
   uint32_t x274;
-  fiat_p256_uint1 x275;
-  fiat_p256_addcarryx_u32(&x274, &x275, x273, x266, x269);
+  uint32_t x275;
+  fiat_p256_mulx_u32(&x274, &x275, x258, UINT32_C(0xffffffff));
   uint32_t x276;
-  fiat_p256_uint1 x277;
-  fiat_p256_addcarryx_u32(&x276, &x277, x275, 0x0, x267);
+  uint32_t x277;
+  fiat_p256_mulx_u32(&x276, &x277, x258, UINT32_C(0xffffffff));
   uint32_t x278;
-  fiat_p256_uint1 x279;
-  fiat_p256_addcarryx_u32(&x278, &x279, 0x0, x270, x244);
+  uint32_t x279;
+  fiat_p256_mulx_u32(&x278, &x279, x258, UINT32_C(0xffffffff));
   uint32_t x280;
-  fiat_p256_uint1 x281;
-  fiat_p256_addcarryx_u32(&x280, &x281, x279, x272, x246);
+  uint32_t x281;
+  fiat_p256_mulx_u32(&x280, &x281, x258, UINT32_C(0xffffffff));
   uint32_t x282;
   fiat_p256_uint1 x283;
-  fiat_p256_addcarryx_u32(&x282, &x283, x281, x274, x248);
+  fiat_p256_addcarryx_u32(&x282, &x283, 0x0, x278, x281);
   uint32_t x284;
   fiat_p256_uint1 x285;
-  fiat_p256_addcarryx_u32(&x284, &x285, x283, x276, x250);
+  fiat_p256_addcarryx_u32(&x284, &x285, x283, x276, x279);
   uint32_t x286;
   fiat_p256_uint1 x287;
-  fiat_p256_addcarryx_u32(&x286, &x287, x277, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x286, &x287, 0x0, x280, x258);
   uint32_t x288;
   fiat_p256_uint1 x289;
-  fiat_p256_addcarryx_u32(&x288, &x289, x285, (fiat_p256_uint1)x286, x252);
+  fiat_p256_addcarryx_u32(&x288, &x289, x287, x282, x260);
   uint32_t x290;
   fiat_p256_uint1 x291;
-  fiat_p256_addcarryx_u32(&x290, &x291, x289, 0x0, x254);
+  fiat_p256_addcarryx_u32(&x290, &x291, x289, x284, x262);
   uint32_t x292;
   fiat_p256_uint1 x293;
-  fiat_p256_addcarryx_u32(&x292, &x293, x291, x244, x256);
+  fiat_p256_addcarryx_u32(&x292, &x293, x285, 0x0, x277);
   uint32_t x294;
   fiat_p256_uint1 x295;
-  fiat_p256_addcarryx_u32(&x294, &x295, x293, x264, x258);
+  fiat_p256_addcarryx_u32(&x294, &x295, x291, x292, x264);
   uint32_t x296;
   fiat_p256_uint1 x297;
-  fiat_p256_addcarryx_u32(&x296, &x297, x295, x265, x262);
+  fiat_p256_addcarryx_u32(&x296, &x297, x295, 0x0, x266);
   uint32_t x298;
   fiat_p256_uint1 x299;
-  fiat_p256_addcarryx_u32(&x298, &x299, 0x0, (arg1[6]), x280);
+  fiat_p256_addcarryx_u32(&x298, &x299, x297, 0x0, x268);
   uint32_t x300;
   fiat_p256_uint1 x301;
-  fiat_p256_addcarryx_u32(&x300, &x301, x299, 0x0, x282);
+  fiat_p256_addcarryx_u32(&x300, &x301, x299, x258, x270);
   uint32_t x302;
   fiat_p256_uint1 x303;
-  fiat_p256_addcarryx_u32(&x302, &x303, x301, 0x0, x284);
+  fiat_p256_addcarryx_u32(&x302, &x303, x301, x274, x272);
   uint32_t x304;
   fiat_p256_uint1 x305;
-  fiat_p256_addcarryx_u32(&x304, &x305, x303, 0x0, x288);
+  fiat_p256_addcarryx_u32(&x304, &x305, x257, 0x0, 0x0);
   uint32_t x306;
   fiat_p256_uint1 x307;
-  fiat_p256_addcarryx_u32(&x306, &x307, x305, 0x0, x290);
+  fiat_p256_addcarryx_u32(&x306, &x307, x273, 0x0, (fiat_p256_uint1)x304);
   uint32_t x308;
   fiat_p256_uint1 x309;
-  fiat_p256_addcarryx_u32(&x308, &x309, x307, 0x0, x292);
+  fiat_p256_addcarryx_u32(&x308, &x309, x303, x275, x306);
   uint32_t x310;
   fiat_p256_uint1 x311;
-  fiat_p256_addcarryx_u32(&x310, &x311, x309, 0x0, x294);
+  fiat_p256_addcarryx_u32(&x310, &x311, 0x0, (arg1[7]), x288);
   uint32_t x312;
   fiat_p256_uint1 x313;
-  fiat_p256_addcarryx_u32(&x312, &x313, x311, 0x0, x296);
+  fiat_p256_addcarryx_u32(&x312, &x313, x311, 0x0, x290);
   uint32_t x314;
   fiat_p256_uint1 x315;
-  fiat_p256_addcarryx_u32(&x314, &x315, x297, 0x0, x263);
+  fiat_p256_addcarryx_u32(&x314, &x315, x313, 0x0, x294);
   uint32_t x316;
   fiat_p256_uint1 x317;
-  fiat_p256_addcarryx_u32(&x316, &x317, x313, 0x0, (fiat_p256_uint1)x314);
+  fiat_p256_addcarryx_u32(&x316, &x317, x315, 0x0, x296);
   uint32_t x318;
-  uint32_t x319;
-  fiat_p256_mulx_u32(&x318, &x319, x298, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x319;
+  fiat_p256_addcarryx_u32(&x318, &x319, x317, 0x0, x298);
   uint32_t x320;
-  uint32_t x321;
-  fiat_p256_mulx_u32(&x320, &x321, x298, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x321;
+  fiat_p256_addcarryx_u32(&x320, &x321, x319, 0x0, x300);
   uint32_t x322;
-  uint32_t x323;
-  fiat_p256_mulx_u32(&x322, &x323, x298, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x323;
+  fiat_p256_addcarryx_u32(&x322, &x323, x321, 0x0, x302);
   uint32_t x324;
-  uint32_t x325;
-  fiat_p256_mulx_u32(&x324, &x325, x298, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x325;
+  fiat_p256_addcarryx_u32(&x324, &x325, x323, 0x0, x308);
   uint32_t x326;
-  fiat_p256_uint1 x327;
-  fiat_p256_addcarryx_u32(&x326, &x327, 0x0, x322, x325);
+  uint32_t x327;
+  fiat_p256_mulx_u32(&x326, &x327, x310, UINT32_C(0xffffffff));
   uint32_t x328;
-  fiat_p256_uint1 x329;
-  fiat_p256_addcarryx_u32(&x328, &x329, x327, x320, x323);
+  uint32_t x329;
+  fiat_p256_mulx_u32(&x328, &x329, x310, UINT32_C(0xffffffff));
   uint32_t x330;
-  fiat_p256_uint1 x331;
-  fiat_p256_addcarryx_u32(&x330, &x331, x329, 0x0, x321);
+  uint32_t x331;
+  fiat_p256_mulx_u32(&x330, &x331, x310, UINT32_C(0xffffffff));
   uint32_t x332;
-  fiat_p256_uint1 x333;
-  fiat_p256_addcarryx_u32(&x332, &x333, 0x0, x324, x298);
+  uint32_t x333;
+  fiat_p256_mulx_u32(&x332, &x333, x310, UINT32_C(0xffffffff));
   uint32_t x334;
   fiat_p256_uint1 x335;
-  fiat_p256_addcarryx_u32(&x334, &x335, x333, x326, x300);
+  fiat_p256_addcarryx_u32(&x334, &x335, 0x0, x330, x333);
   uint32_t x336;
   fiat_p256_uint1 x337;
-  fiat_p256_addcarryx_u32(&x336, &x337, x335, x328, x302);
+  fiat_p256_addcarryx_u32(&x336, &x337, x335, x328, x331);
   uint32_t x338;
   fiat_p256_uint1 x339;
-  fiat_p256_addcarryx_u32(&x338, &x339, x337, x330, x304);
+  fiat_p256_addcarryx_u32(&x338, &x339, 0x0, x332, x310);
   uint32_t x340;
   fiat_p256_uint1 x341;
-  fiat_p256_addcarryx_u32(&x340, &x341, x331, 0x0, 0x0);
+  fiat_p256_addcarryx_u32(&x340, &x341, x339, x334, x312);
   uint32_t x342;
   fiat_p256_uint1 x343;
-  fiat_p256_addcarryx_u32(&x342, &x343, x339, (fiat_p256_uint1)x340, x306);
+  fiat_p256_addcarryx_u32(&x342, &x343, x341, x336, x314);
   uint32_t x344;
   fiat_p256_uint1 x345;
-  fiat_p256_addcarryx_u32(&x344, &x345, x343, 0x0, x308);
+  fiat_p256_addcarryx_u32(&x344, &x345, x337, 0x0, x329);
   uint32_t x346;
   fiat_p256_uint1 x347;
-  fiat_p256_addcarryx_u32(&x346, &x347, x345, x298, x310);
+  fiat_p256_addcarryx_u32(&x346, &x347, x343, x344, x316);
   uint32_t x348;
   fiat_p256_uint1 x349;
-  fiat_p256_addcarryx_u32(&x348, &x349, x347, x318, x312);
+  fiat_p256_addcarryx_u32(&x348, &x349, x347, 0x0, x318);
   uint32_t x350;
   fiat_p256_uint1 x351;
-  fiat_p256_addcarryx_u32(&x350, &x351, x349, x319, x316);
+  fiat_p256_addcarryx_u32(&x350, &x351, x349, 0x0, x320);
   uint32_t x352;
   fiat_p256_uint1 x353;
-  fiat_p256_addcarryx_u32(&x352, &x353, 0x0, (arg1[7]), x334);
+  fiat_p256_addcarryx_u32(&x352, &x353, x351, x310, x322);
   uint32_t x354;
   fiat_p256_uint1 x355;
-  fiat_p256_addcarryx_u32(&x354, &x355, x353, 0x0, x336);
+  fiat_p256_addcarryx_u32(&x354, &x355, x353, x326, x324);
   uint32_t x356;
   fiat_p256_uint1 x357;
-  fiat_p256_addcarryx_u32(&x356, &x357, x355, 0x0, x338);
+  fiat_p256_addcarryx_u32(&x356, &x357, x309, 0x0, 0x0);
   uint32_t x358;
   fiat_p256_uint1 x359;
-  fiat_p256_addcarryx_u32(&x358, &x359, x357, 0x0, x342);
+  fiat_p256_addcarryx_u32(&x358, &x359, x325, 0x0, (fiat_p256_uint1)x356);
   uint32_t x360;
   fiat_p256_uint1 x361;
-  fiat_p256_addcarryx_u32(&x360, &x361, x359, 0x0, x344);
+  fiat_p256_addcarryx_u32(&x360, &x361, x355, x327, x358);
   uint32_t x362;
   fiat_p256_uint1 x363;
-  fiat_p256_addcarryx_u32(&x362, &x363, x361, 0x0, x346);
+  fiat_p256_subborrowx_u32(&x362, &x363, 0x0, x340, UINT32_C(0xffffffff));
   uint32_t x364;
   fiat_p256_uint1 x365;
-  fiat_p256_addcarryx_u32(&x364, &x365, x363, 0x0, x348);
+  fiat_p256_subborrowx_u32(&x364, &x365, x363, x342, UINT32_C(0xffffffff));
   uint32_t x366;
   fiat_p256_uint1 x367;
-  fiat_p256_addcarryx_u32(&x366, &x367, x365, 0x0, x350);
+  fiat_p256_subborrowx_u32(&x366, &x367, x365, x346, UINT32_C(0xffffffff));
   uint32_t x368;
   fiat_p256_uint1 x369;
-  fiat_p256_addcarryx_u32(&x368, &x369, x351, 0x0, x317);
+  fiat_p256_subborrowx_u32(&x368, &x369, x367, x348, 0x0);
   uint32_t x370;
   fiat_p256_uint1 x371;
-  fiat_p256_addcarryx_u32(&x370, &x371, x367, 0x0, (fiat_p256_uint1)x368);
+  fiat_p256_subborrowx_u32(&x370, &x371, x369, x350, 0x0);
   uint32_t x372;
-  uint32_t x373;
-  fiat_p256_mulx_u32(&x372, &x373, x352, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x373;
+  fiat_p256_subborrowx_u32(&x372, &x373, x371, x352, 0x0);
   uint32_t x374;
-  uint32_t x375;
-  fiat_p256_mulx_u32(&x374, &x375, x352, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x375;
+  fiat_p256_subborrowx_u32(&x374, &x375, x373, x354, 0x1);
   uint32_t x376;
-  uint32_t x377;
-  fiat_p256_mulx_u32(&x376, &x377, x352, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x377;
+  fiat_p256_subborrowx_u32(&x376, &x377, x375, x360, UINT32_C(0xffffffff));
   uint32_t x378;
-  uint32_t x379;
-  fiat_p256_mulx_u32(&x378, &x379, x352, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x379;
+  fiat_p256_addcarryx_u32(&x378, &x379, x361, 0x0, 0x0);
   uint32_t x380;
   fiat_p256_uint1 x381;
-  fiat_p256_addcarryx_u32(&x380, &x381, 0x0, x376, x379);
+  fiat_p256_subborrowx_u32(&x380, &x381, x377, (fiat_p256_uint1)x378, 0x0);
   uint32_t x382;
-  fiat_p256_uint1 x383;
-  fiat_p256_addcarryx_u32(&x382, &x383, x381, x374, x377);
+  fiat_p256_cmovznz_u32(&x382, x381, x362, x340);
+  uint32_t x383;
+  fiat_p256_cmovznz_u32(&x383, x381, x364, x342);
   uint32_t x384;
-  fiat_p256_uint1 x385;
-  fiat_p256_addcarryx_u32(&x384, &x385, x383, 0x0, x375);
+  fiat_p256_cmovznz_u32(&x384, x381, x366, x346);
+  uint32_t x385;
+  fiat_p256_cmovznz_u32(&x385, x381, x368, x348);
   uint32_t x386;
-  fiat_p256_uint1 x387;
-  fiat_p256_addcarryx_u32(&x386, &x387, 0x0, x378, x352);
+  fiat_p256_cmovznz_u32(&x386, x381, x370, x350);
+  uint32_t x387;
+  fiat_p256_cmovznz_u32(&x387, x381, x372, x352);
   uint32_t x388;
-  fiat_p256_uint1 x389;
-  fiat_p256_addcarryx_u32(&x388, &x389, x387, x380, x354);
-  uint32_t x390;
-  fiat_p256_uint1 x391;
-  fiat_p256_addcarryx_u32(&x390, &x391, x389, x382, x356);
-  uint32_t x392;
-  fiat_p256_uint1 x393;
-  fiat_p256_addcarryx_u32(&x392, &x393, x391, x384, x358);
-  uint32_t x394;
-  fiat_p256_uint1 x395;
-  fiat_p256_addcarryx_u32(&x394, &x395, x385, 0x0, 0x0);
-  uint32_t x396;
-  fiat_p256_uint1 x397;
-  fiat_p256_addcarryx_u32(&x396, &x397, x393, (fiat_p256_uint1)x394, x360);
-  uint32_t x398;
-  fiat_p256_uint1 x399;
-  fiat_p256_addcarryx_u32(&x398, &x399, x397, 0x0, x362);
-  uint32_t x400;
-  fiat_p256_uint1 x401;
-  fiat_p256_addcarryx_u32(&x400, &x401, x399, x352, x364);
-  uint32_t x402;
-  fiat_p256_uint1 x403;
-  fiat_p256_addcarryx_u32(&x402, &x403, x401, x372, x366);
-  uint32_t x404;
-  fiat_p256_uint1 x405;
-  fiat_p256_addcarryx_u32(&x404, &x405, x403, x373, x370);
-  uint32_t x406;
-  fiat_p256_uint1 x407;
-  fiat_p256_subborrowx_u32(&x406, &x407, 0x0, x388, UINT32_C(0xffffffff));
-  uint32_t x408;
-  fiat_p256_uint1 x409;
-  fiat_p256_subborrowx_u32(&x408, &x409, x407, x390, UINT32_C(0xffffffff));
-  uint32_t x410;
-  fiat_p256_uint1 x411;
-  fiat_p256_subborrowx_u32(&x410, &x411, x409, x392, UINT32_C(0xffffffff));
-  uint32_t x412;
-  fiat_p256_uint1 x413;
-  fiat_p256_subborrowx_u32(&x412, &x413, x411, x396, 0x0);
-  uint32_t x414;
-  fiat_p256_uint1 x415;
-  fiat_p256_subborrowx_u32(&x414, &x415, x413, x398, 0x0);
-  uint32_t x416;
-  fiat_p256_uint1 x417;
-  fiat_p256_subborrowx_u32(&x416, &x417, x415, x400, 0x0);
-  uint32_t x418;
-  fiat_p256_uint1 x419;
-  fiat_p256_subborrowx_u32(&x418, &x419, x417, x402, 0x1);
-  uint32_t x420;
-  fiat_p256_uint1 x421;
-  fiat_p256_subborrowx_u32(&x420, &x421, x419, x404, UINT32_C(0xffffffff));
-  uint32_t x422;
-  fiat_p256_uint1 x423;
-  fiat_p256_addcarryx_u32(&x422, &x423, x405, 0x0, x371);
-  uint32_t x424;
-  fiat_p256_uint1 x425;
-  fiat_p256_subborrowx_u32(&x424, &x425, x421, (fiat_p256_uint1)x422, 0x0);
-  uint32_t x426;
-  fiat_p256_cmovznz_u32(&x426, x425, x406, x388);
-  uint32_t x427;
-  fiat_p256_cmovznz_u32(&x427, x425, x408, x390);
-  uint32_t x428;
-  fiat_p256_cmovznz_u32(&x428, x425, x410, x392);
-  uint32_t x429;
-  fiat_p256_cmovznz_u32(&x429, x425, x412, x396);
-  uint32_t x430;
-  fiat_p256_cmovznz_u32(&x430, x425, x414, x398);
-  uint32_t x431;
-  fiat_p256_cmovznz_u32(&x431, x425, x416, x400);
-  uint32_t x432;
-  fiat_p256_cmovznz_u32(&x432, x425, x418, x402);
-  uint32_t x433;
-  fiat_p256_cmovznz_u32(&x433, x425, x420, x404);
-  out1[0] = x426;
-  out1[1] = x427;
-  out1[2] = x428;
-  out1[3] = x429;
-  out1[4] = x430;
-  out1[5] = x431;
-  out1[6] = x432;
-  out1[7] = x433;
+  fiat_p256_cmovznz_u32(&x388, x381, x374, x354);
+  uint32_t x389;
+  fiat_p256_cmovznz_u32(&x389, x381, x376, x360);
+  out1[0] = x382;
+  out1[1] = x383;
+  out1[2] = x384;
+  out1[3] = x385;
+  out1[4] = x386;
+  out1[5] = x387;
+  out1[6] = x388;
+  out1[7] = x389;
 }
 
 /*
@@ -3185,101 +3071,94 @@ static void fiat_p256_to_bytes(uint8_t out1[32], const uint32_t arg1[8]) {
   uint8_t x12 = (uint8_t)(x9 & UINT8_C(0xff));
   uint8_t x13 = (uint8_t)(x11 >> 8);
   uint8_t x14 = (uint8_t)(x11 & UINT8_C(0xff));
-  fiat_p256_uint1 x15 = (fiat_p256_uint1)(x13 >> 8);
-  uint8_t x16 = (uint8_t)(x13 & UINT8_C(0xff));
-  uint32_t x17 = (x15 + x7);
-  uint32_t x18 = (x17 >> 8);
-  uint8_t x19 = (uint8_t)(x17 & UINT8_C(0xff));
-  uint32_t x20 = (x18 >> 8);
-  uint8_t x21 = (uint8_t)(x18 & UINT8_C(0xff));
-  uint8_t x22 = (uint8_t)(x20 >> 8);
-  uint8_t x23 = (uint8_t)(x20 & UINT8_C(0xff));
-  fiat_p256_uint1 x24 = (fiat_p256_uint1)(x22 >> 8);
-  uint8_t x25 = (uint8_t)(x22 & UINT8_C(0xff));
-  uint32_t x26 = (x24 + x6);
-  uint32_t x27 = (x26 >> 8);
-  uint8_t x28 = (uint8_t)(x26 & UINT8_C(0xff));
-  uint32_t x29 = (x27 >> 8);
+  uint8_t x15 = (uint8_t)(x13 & UINT8_C(0xff));
+  uint32_t x16 = (0x0 + x7);
+  uint32_t x17 = (x16 >> 8);
+  uint8_t x18 = (uint8_t)(x16 & UINT8_C(0xff));
+  uint32_t x19 = (x17 >> 8);
+  uint8_t x20 = (uint8_t)(x17 & UINT8_C(0xff));
+  uint8_t x21 = (uint8_t)(x19 >> 8);
+  uint8_t x22 = (uint8_t)(x19 & UINT8_C(0xff));
+  uint8_t x23 = (uint8_t)(x21 & UINT8_C(0xff));
+  uint32_t x24 = (0x0 + x6);
+  uint32_t x25 = (x24 >> 8);
+  uint8_t x26 = (uint8_t)(x24 & UINT8_C(0xff));
+  uint32_t x27 = (x25 >> 8);
+  uint8_t x28 = (uint8_t)(x25 & UINT8_C(0xff));
+  uint8_t x29 = (uint8_t)(x27 >> 8);
   uint8_t x30 = (uint8_t)(x27 & UINT8_C(0xff));
-  uint8_t x31 = (uint8_t)(x29 >> 8);
-  uint8_t x32 = (uint8_t)(x29 & UINT8_C(0xff));
-  fiat_p256_uint1 x33 = (fiat_p256_uint1)(x31 >> 8);
-  uint8_t x34 = (uint8_t)(x31 & UINT8_C(0xff));
-  uint32_t x35 = (x33 + x5);
-  uint32_t x36 = (x35 >> 8);
-  uint8_t x37 = (uint8_t)(x35 & UINT8_C(0xff));
-  uint32_t x38 = (x36 >> 8);
-  uint8_t x39 = (uint8_t)(x36 & UINT8_C(0xff));
-  uint8_t x40 = (uint8_t)(x38 >> 8);
-  uint8_t x41 = (uint8_t)(x38 & UINT8_C(0xff));
-  fiat_p256_uint1 x42 = (fiat_p256_uint1)(x40 >> 8);
-  uint8_t x43 = (uint8_t)(x40 & UINT8_C(0xff));
-  uint32_t x44 = (x42 + x4);
-  uint32_t x45 = (x44 >> 8);
-  uint8_t x46 = (uint8_t)(x44 & UINT8_C(0xff));
-  uint32_t x47 = (x45 >> 8);
-  uint8_t x48 = (uint8_t)(x45 & UINT8_C(0xff));
-  uint8_t x49 = (uint8_t)(x47 >> 8);
-  uint8_t x50 = (uint8_t)(x47 & UINT8_C(0xff));
-  fiat_p256_uint1 x51 = (fiat_p256_uint1)(x49 >> 8);
+  uint8_t x31 = (uint8_t)(x29 & UINT8_C(0xff));
+  uint32_t x32 = (0x0 + x5);
+  uint32_t x33 = (x32 >> 8);
+  uint8_t x34 = (uint8_t)(x32 & UINT8_C(0xff));
+  uint32_t x35 = (x33 >> 8);
+  uint8_t x36 = (uint8_t)(x33 & UINT8_C(0xff));
+  uint8_t x37 = (uint8_t)(x35 >> 8);
+  uint8_t x38 = (uint8_t)(x35 & UINT8_C(0xff));
+  uint8_t x39 = (uint8_t)(x37 & UINT8_C(0xff));
+  uint32_t x40 = (0x0 + x4);
+  uint32_t x41 = (x40 >> 8);
+  uint8_t x42 = (uint8_t)(x40 & UINT8_C(0xff));
+  uint32_t x43 = (x41 >> 8);
+  uint8_t x44 = (uint8_t)(x41 & UINT8_C(0xff));
+  uint8_t x45 = (uint8_t)(x43 >> 8);
+  uint8_t x46 = (uint8_t)(x43 & UINT8_C(0xff));
+  uint8_t x47 = (uint8_t)(x45 & UINT8_C(0xff));
+  uint32_t x48 = (0x0 + x3);
+  uint32_t x49 = (x48 >> 8);
+  uint8_t x50 = (uint8_t)(x48 & UINT8_C(0xff));
+  uint32_t x51 = (x49 >> 8);
   uint8_t x52 = (uint8_t)(x49 & UINT8_C(0xff));
-  uint32_t x53 = (x51 + x3);
-  uint32_t x54 = (x53 >> 8);
+  uint8_t x53 = (uint8_t)(x51 >> 8);
+  uint8_t x54 = (uint8_t)(x51 & UINT8_C(0xff));
   uint8_t x55 = (uint8_t)(x53 & UINT8_C(0xff));
-  uint32_t x56 = (x54 >> 8);
-  uint8_t x57 = (uint8_t)(x54 & UINT8_C(0xff));
-  uint8_t x58 = (uint8_t)(x56 >> 8);
-  uint8_t x59 = (uint8_t)(x56 & UINT8_C(0xff));
-  fiat_p256_uint1 x60 = (fiat_p256_uint1)(x58 >> 8);
-  uint8_t x61 = (uint8_t)(x58 & UINT8_C(0xff));
-  uint32_t x62 = (x60 + x2);
-  uint32_t x63 = (x62 >> 8);
-  uint8_t x64 = (uint8_t)(x62 & UINT8_C(0xff));
-  uint32_t x65 = (x63 >> 8);
-  uint8_t x66 = (uint8_t)(x63 & UINT8_C(0xff));
-  uint8_t x67 = (uint8_t)(x65 >> 8);
+  uint32_t x56 = (0x0 + x2);
+  uint32_t x57 = (x56 >> 8);
+  uint8_t x58 = (uint8_t)(x56 & UINT8_C(0xff));
+  uint32_t x59 = (x57 >> 8);
+  uint8_t x60 = (uint8_t)(x57 & UINT8_C(0xff));
+  uint8_t x61 = (uint8_t)(x59 >> 8);
+  uint8_t x62 = (uint8_t)(x59 & UINT8_C(0xff));
+  uint8_t x63 = (uint8_t)(x61 & UINT8_C(0xff));
+  uint32_t x64 = (0x0 + x1);
+  uint32_t x65 = (x64 >> 8);
+  uint8_t x66 = (uint8_t)(x64 & UINT8_C(0xff));
+  uint32_t x67 = (x65 >> 8);
   uint8_t x68 = (uint8_t)(x65 & UINT8_C(0xff));
-  fiat_p256_uint1 x69 = (fiat_p256_uint1)(x67 >> 8);
+  uint8_t x69 = (uint8_t)(x67 >> 8);
   uint8_t x70 = (uint8_t)(x67 & UINT8_C(0xff));
-  uint32_t x71 = (x69 + x1);
-  uint32_t x72 = (x71 >> 8);
-  uint8_t x73 = (uint8_t)(x71 & UINT8_C(0xff));
-  uint32_t x74 = (x72 >> 8);
-  uint8_t x75 = (uint8_t)(x72 & UINT8_C(0xff));
-  uint8_t x76 = (uint8_t)(x74 >> 8);
-  uint8_t x77 = (uint8_t)(x74 & UINT8_C(0xff));
   out1[0] = x10;
   out1[1] = x12;
   out1[2] = x14;
-  out1[3] = x16;
-  out1[4] = x19;
-  out1[5] = x21;
-  out1[6] = x23;
-  out1[7] = x25;
-  out1[8] = x28;
-  out1[9] = x30;
-  out1[10] = x32;
-  out1[11] = x34;
-  out1[12] = x37;
-  out1[13] = x39;
-  out1[14] = x41;
-  out1[15] = x43;
-  out1[16] = x46;
-  out1[17] = x48;
-  out1[18] = x50;
-  out1[19] = x52;
-  out1[20] = x55;
-  out1[21] = x57;
-  out1[22] = x59;
-  out1[23] = x61;
-  out1[24] = x64;
-  out1[25] = x66;
-  out1[26] = x68;
-  out1[27] = x70;
-  out1[28] = x73;
-  out1[29] = x75;
-  out1[30] = x77;
-  out1[31] = x76;
+  out1[3] = x15;
+  out1[4] = x18;
+  out1[5] = x20;
+  out1[6] = x22;
+  out1[7] = x23;
+  out1[8] = x26;
+  out1[9] = x28;
+  out1[10] = x30;
+  out1[11] = x31;
+  out1[12] = x34;
+  out1[13] = x36;
+  out1[14] = x38;
+  out1[15] = x39;
+  out1[16] = x42;
+  out1[17] = x44;
+  out1[18] = x46;
+  out1[19] = x47;
+  out1[20] = x50;
+  out1[21] = x52;
+  out1[22] = x54;
+  out1[23] = x55;
+  out1[24] = x58;
+  out1[25] = x60;
+  out1[26] = x62;
+  out1[27] = x63;
+  out1[28] = x66;
+  out1[29] = x68;
+  out1[30] = x70;
+  out1[31] = x69;
 }
 
 /*
@@ -3322,41 +3201,34 @@ static void fiat_p256_from_bytes(uint32_t out1[8], const uint8_t arg1[32]) {
   uint32_t x31 = ((uint32_t)(arg1[1]) << 8);
   uint8_t x32 = (arg1[0]);
   uint32_t x33 = (x32 + (x31 + (x30 + x29)));
-  fiat_p256_uint1 x34 = (fiat_p256_uint1)((uint64_t)x33 >> 32);
-  uint32_t x35 = (x33 & UINT32_C(0xffffffff));
-  uint32_t x36 = (x4 + (x3 + (x2 + x1)));
-  uint32_t x37 = (x8 + (x7 + (x6 + x5)));
-  uint32_t x38 = (x12 + (x11 + (x10 + x9)));
-  uint32_t x39 = (x16 + (x15 + (x14 + x13)));
-  uint32_t x40 = (x20 + (x19 + (x18 + x17)));
-  uint32_t x41 = (x24 + (x23 + (x22 + x21)));
-  uint32_t x42 = (x28 + (x27 + (x26 + x25)));
-  uint32_t x43 = (x34 + x42);
-  fiat_p256_uint1 x44 = (fiat_p256_uint1)((uint64_t)x43 >> 32);
-  uint32_t x45 = (x43 & UINT32_C(0xffffffff));
-  uint32_t x46 = (x44 + x41);
-  fiat_p256_uint1 x47 = (fiat_p256_uint1)((uint64_t)x46 >> 32);
-  uint32_t x48 = (x46 & UINT32_C(0xffffffff));
-  uint32_t x49 = (x47 + x40);
-  fiat_p256_uint1 x50 = (fiat_p256_uint1)((uint64_t)x49 >> 32);
-  uint32_t x51 = (x49 & UINT32_C(0xffffffff));
-  uint32_t x52 = (x50 + x39);
-  fiat_p256_uint1 x53 = (fiat_p256_uint1)((uint64_t)x52 >> 32);
-  uint32_t x54 = (x52 & UINT32_C(0xffffffff));
-  uint32_t x55 = (x53 + x38);
-  fiat_p256_uint1 x56 = (fiat_p256_uint1)((uint64_t)x55 >> 32);
-  uint32_t x57 = (x55 & UINT32_C(0xffffffff));
-  uint32_t x58 = (x56 + x37);
-  fiat_p256_uint1 x59 = (fiat_p256_uint1)((uint64_t)x58 >> 32);
-  uint32_t x60 = (x58 & UINT32_C(0xffffffff));
-  uint32_t x61 = (x59 + x36);
-  out1[0] = x35;
-  out1[1] = x45;
-  out1[2] = x48;
-  out1[3] = x51;
-  out1[4] = x54;
-  out1[5] = x57;
-  out1[6] = x60;
-  out1[7] = x61;
+  uint32_t x34 = (x33 & UINT32_C(0xffffffff));
+  uint32_t x35 = (x4 + (x3 + (x2 + x1)));
+  uint32_t x36 = (x8 + (x7 + (x6 + x5)));
+  uint32_t x37 = (x12 + (x11 + (x10 + x9)));
+  uint32_t x38 = (x16 + (x15 + (x14 + x13)));
+  uint32_t x39 = (x20 + (x19 + (x18 + x17)));
+  uint32_t x40 = (x24 + (x23 + (x22 + x21)));
+  uint32_t x41 = (x28 + (x27 + (x26 + x25)));
+  uint32_t x42 = (0x0 + x41);
+  uint32_t x43 = (x42 & UINT32_C(0xffffffff));
+  uint32_t x44 = (0x0 + x40);
+  uint32_t x45 = (x44 & UINT32_C(0xffffffff));
+  uint32_t x46 = (0x0 + x39);
+  uint32_t x47 = (x46 & UINT32_C(0xffffffff));
+  uint32_t x48 = (0x0 + x38);
+  uint32_t x49 = (x48 & UINT32_C(0xffffffff));
+  uint32_t x50 = (0x0 + x37);
+  uint32_t x51 = (x50 & UINT32_C(0xffffffff));
+  uint32_t x52 = (0x0 + x36);
+  uint32_t x53 = (x52 & UINT32_C(0xffffffff));
+  uint32_t x54 = (0x0 + x35);
+  out1[0] = x34;
+  out1[1] = x43;
+  out1[2] = x45;
+  out1[3] = x47;
+  out1[4] = x49;
+  out1[5] = x51;
+  out1[6] = x53;
+  out1[7] = x54;
 }
 
diff --git a/p256_64.c b/p256_64.c
index 2bd9b8d51..018092b3f 100644
--- a/p256_64.c
+++ b/p256_64.c
@@ -136,289 +136,265 @@ static void fiat_p256_mul(uint64_t out1[4], const uint64_t arg1[4], const uint64
   fiat_p256_addcarryx_u64(&x29, &x30, x28, 0x0, x24);
   uint64_t x31;
   fiat_p256_uint1 x32;
-  fiat_p256_addcarryx_u64(&x31, &x32, x30, x21, 0x0);
+  fiat_p256_addcarryx_u64(&x31, &x32, 0x0, x25, x11);
   uint64_t x33;
   fiat_p256_uint1 x34;
-  fiat_p256_addcarryx_u64(&x33, &x34, x32, 0x0, x22);
+  fiat_p256_addcarryx_u64(&x33, &x34, x32, x27, x13);
   uint64_t x35;
   fiat_p256_uint1 x36;
-  fiat_p256_addcarryx_u64(&x35, &x36, 0x0, x25, x11);
+  fiat_p256_addcarryx_u64(&x35, &x36, x34, x29, x15);
   uint64_t x37;
   fiat_p256_uint1 x38;
-  fiat_p256_addcarryx_u64(&x37, &x38, x36, x27, x13);
+  fiat_p256_addcarryx_u64(&x37, &x38, x36, x21, x17);
   uint64_t x39;
   fiat_p256_uint1 x40;
-  fiat_p256_addcarryx_u64(&x39, &x40, x38, x29, x15);
+  fiat_p256_addcarryx_u64(&x39, &x40, x38, x22, x19);
   uint64_t x41;
   fiat_p256_uint1 x42;
-  fiat_p256_addcarryx_u64(&x41, &x42, x40, x31, x17);
+  fiat_p256_addcarryx_u64(&x41, &x42, x40, 0x0, 0x0);
   uint64_t x43;
-  fiat_p256_uint1 x44;
-  fiat_p256_addcarryx_u64(&x43, &x44, x42, x33, x19);
+  uint64_t x44;
+  fiat_p256_mulx_u64(&x43, &x44, x1, (arg2[3]));
   uint64_t x45;
-  fiat_p256_uint1 x46;
-  fiat_p256_addcarryx_u64(&x45, &x46, x44, 0x0, 0x0);
+  uint64_t x46;
+  fiat_p256_mulx_u64(&x45, &x46, x1, (arg2[2]));
   uint64_t x47;
   uint64_t x48;
-  fiat_p256_mulx_u64(&x47, &x48, x1, (arg2[3]));
+  fiat_p256_mulx_u64(&x47, &x48, x1, (arg2[1]));
   uint64_t x49;
   uint64_t x50;
-  fiat_p256_mulx_u64(&x49, &x50, x1, (arg2[2]));
+  fiat_p256_mulx_u64(&x49, &x50, x1, (arg2[0]));
   uint64_t x51;
-  uint64_t x52;
-  fiat_p256_mulx_u64(&x51, &x52, x1, (arg2[1]));
+  fiat_p256_uint1 x52;
+  fiat_p256_addcarryx_u64(&x51, &x52, 0x0, x47, x50);
   uint64_t x53;
-  uint64_t x54;
-  fiat_p256_mulx_u64(&x53, &x54, x1, (arg2[0]));
+  fiat_p256_uint1 x54;
+  fiat_p256_addcarryx_u64(&x53, &x54, x52, x45, x48);
   uint64_t x55;
   fiat_p256_uint1 x56;
-  fiat_p256_addcarryx_u64(&x55, &x56, 0x0, x51, x54);
+  fiat_p256_addcarryx_u64(&x55, &x56, x54, x43, x46);
   uint64_t x57;
   fiat_p256_uint1 x58;
-  fiat_p256_addcarryx_u64(&x57, &x58, x56, x49, x52);
+  fiat_p256_addcarryx_u64(&x57, &x58, x56, 0x0, x44);
   uint64_t x59;
   fiat_p256_uint1 x60;
-  fiat_p256_addcarryx_u64(&x59, &x60, x58, x47, x50);
+  fiat_p256_addcarryx_u64(&x59, &x60, 0x0, x49, x33);
   uint64_t x61;
   fiat_p256_uint1 x62;
-  fiat_p256_addcarryx_u64(&x61, &x62, x60, 0x0, x48);
+  fiat_p256_addcarryx_u64(&x61, &x62, x60, x51, x35);
   uint64_t x63;
   fiat_p256_uint1 x64;
-  fiat_p256_addcarryx_u64(&x63, &x64, 0x0, x53, x37);
+  fiat_p256_addcarryx_u64(&x63, &x64, x62, x53, x37);
   uint64_t x65;
   fiat_p256_uint1 x66;
   fiat_p256_addcarryx_u64(&x65, &x66, x64, x55, x39);
   uint64_t x67;
   fiat_p256_uint1 x68;
-  fiat_p256_addcarryx_u64(&x67, &x68, x66, x57, x41);
+  fiat_p256_addcarryx_u64(&x67, &x68, x66, x57, (fiat_p256_uint1)x41);
   uint64_t x69;
-  fiat_p256_uint1 x70;
-  fiat_p256_addcarryx_u64(&x69, &x70, x68, x59, x43);
+  uint64_t x70;
+  fiat_p256_mulx_u64(&x69, &x70, x59, UINT64_C(0xffffffff00000001));
   uint64_t x71;
-  fiat_p256_uint1 x72;
-  fiat_p256_addcarryx_u64(&x71, &x72, x70, x61, (fiat_p256_uint1)x45);
+  uint64_t x72;
+  fiat_p256_mulx_u64(&x71, &x72, x59, UINT32_C(0xffffffff));
   uint64_t x73;
   uint64_t x74;
-  fiat_p256_mulx_u64(&x73, &x74, x63, UINT64_C(0xffffffff00000001));
+  fiat_p256_mulx_u64(&x73, &x74, x59, UINT64_C(0xffffffffffffffff));
   uint64_t x75;
-  uint64_t x76;
-  fiat_p256_mulx_u64(&x75, &x76, x63, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x76;
+  fiat_p256_addcarryx_u64(&x75, &x76, 0x0, x71, x74);
   uint64_t x77;
-  uint64_t x78;
-  fiat_p256_mulx_u64(&x77, &x78, x63, UINT64_C(0xffffffffffffffff));
+  fiat_p256_uint1 x78;
+  fiat_p256_addcarryx_u64(&x77, &x78, x76, 0x0, x72);
   uint64_t x79;
   fiat_p256_uint1 x80;
-  fiat_p256_addcarryx_u64(&x79, &x80, 0x0, x75, x78);
+  fiat_p256_addcarryx_u64(&x79, &x80, 0x0, x73, x59);
   uint64_t x81;
   fiat_p256_uint1 x82;
-  fiat_p256_addcarryx_u64(&x81, &x82, x80, 0x0, x76);
+  fiat_p256_addcarryx_u64(&x81, &x82, x80, x75, x61);
   uint64_t x83;
   fiat_p256_uint1 x84;
-  fiat_p256_addcarryx_u64(&x83, &x84, x82, x73, 0x0);
+  fiat_p256_addcarryx_u64(&x83, &x84, x82, x77, x63);
   uint64_t x85;
   fiat_p256_uint1 x86;
-  fiat_p256_addcarryx_u64(&x85, &x86, x84, 0x0, x74);
+  fiat_p256_addcarryx_u64(&x85, &x86, x84, x69, x65);
   uint64_t x87;
   fiat_p256_uint1 x88;
-  fiat_p256_addcarryx_u64(&x87, &x88, 0x0, x77, x63);
+  fiat_p256_addcarryx_u64(&x87, &x88, x86, x70, x67);
   uint64_t x89;
   fiat_p256_uint1 x90;
-  fiat_p256_addcarryx_u64(&x89, &x90, x88, x79, x65);
+  fiat_p256_addcarryx_u64(&x89, &x90, x88, 0x0, x68);
   uint64_t x91;
-  fiat_p256_uint1 x92;
-  fiat_p256_addcarryx_u64(&x91, &x92, x90, x81, x67);
+  uint64_t x92;
+  fiat_p256_mulx_u64(&x91, &x92, x2, (arg2[3]));
   uint64_t x93;
-  fiat_p256_uint1 x94;
-  fiat_p256_addcarryx_u64(&x93, &x94, x92, x83, x69);
+  uint64_t x94;
+  fiat_p256_mulx_u64(&x93, &x94, x2, (arg2[2]));
   uint64_t x95;
-  fiat_p256_uint1 x96;
-  fiat_p256_addcarryx_u64(&x95, &x96, x94, x85, x71);
+  uint64_t x96;
+  fiat_p256_mulx_u64(&x95, &x96, x2, (arg2[1]));
   uint64_t x97;
-  fiat_p256_uint1 x98;
-  fiat_p256_addcarryx_u64(&x97, &x98, x96, 0x0, x72);
+  uint64_t x98;
+  fiat_p256_mulx_u64(&x97, &x98, x2, (arg2[0]));
   uint64_t x99;
-  uint64_t x100;
-  fiat_p256_mulx_u64(&x99, &x100, x2, (arg2[3]));
+  fiat_p256_uint1 x100;
+  fiat_p256_addcarryx_u64(&x99, &x100, 0x0, x95, x98);
   uint64_t x101;
-  uint64_t x102;
-  fiat_p256_mulx_u64(&x101, &x102, x2, (arg2[2]));
+  fiat_p256_uint1 x102;
+  fiat_p256_addcarryx_u64(&x101, &x102, x100, x93, x96);
   uint64_t x103;
-  uint64_t x104;
-  fiat_p256_mulx_u64(&x103, &x104, x2, (arg2[1]));
+  fiat_p256_uint1 x104;
+  fiat_p256_addcarryx_u64(&x103, &x104, x102, x91, x94);
   uint64_t x105;
-  uint64_t x106;
-  fiat_p256_mulx_u64(&x105, &x106, x2, (arg2[0]));
+  fiat_p256_uint1 x106;
+  fiat_p256_addcarryx_u64(&x105, &x106, x104, 0x0, x92);
   uint64_t x107;
   fiat_p256_uint1 x108;
-  fiat_p256_addcarryx_u64(&x107, &x108, 0x0, x103, x106);
+  fiat_p256_addcarryx_u64(&x107, &x108, 0x0, x97, x81);
   uint64_t x109;
   fiat_p256_uint1 x110;
-  fiat_p256_addcarryx_u64(&x109, &x110, x108, x101, x104);
+  fiat_p256_addcarryx_u64(&x109, &x110, x108, x99, x83);
   uint64_t x111;
   fiat_p256_uint1 x112;
-  fiat_p256_addcarryx_u64(&x111, &x112, x110, x99, x102);
+  fiat_p256_addcarryx_u64(&x111, &x112, x110, x101, x85);
   uint64_t x113;
   fiat_p256_uint1 x114;
-  fiat_p256_addcarryx_u64(&x113, &x114, x112, 0x0, x100);
+  fiat_p256_addcarryx_u64(&x113, &x114, x112, x103, x87);
   uint64_t x115;
   fiat_p256_uint1 x116;
-  fiat_p256_addcarryx_u64(&x115, &x116, 0x0, x105, x89);
+  fiat_p256_addcarryx_u64(&x115, &x116, x114, x105, x89);
   uint64_t x117;
-  fiat_p256_uint1 x118;
-  fiat_p256_addcarryx_u64(&x117, &x118, x116, x107, x91);
+  uint64_t x118;
+  fiat_p256_mulx_u64(&x117, &x118, x107, UINT64_C(0xffffffff00000001));
   uint64_t x119;
-  fiat_p256_uint1 x120;
-  fiat_p256_addcarryx_u64(&x119, &x120, x118, x109, x93);
+  uint64_t x120;
+  fiat_p256_mulx_u64(&x119, &x120, x107, UINT32_C(0xffffffff));
   uint64_t x121;
-  fiat_p256_uint1 x122;
-  fiat_p256_addcarryx_u64(&x121, &x122, x120, x111, x95);
+  uint64_t x122;
+  fiat_p256_mulx_u64(&x121, &x122, x107, UINT64_C(0xffffffffffffffff));
   uint64_t x123;
   fiat_p256_uint1 x124;
-  fiat_p256_addcarryx_u64(&x123, &x124, x122, x113, x97);
+  fiat_p256_addcarryx_u64(&x123, &x124, 0x0, x119, x122);
   uint64_t x125;
-  uint64_t x126;
-  fiat_p256_mulx_u64(&x125, &x126, x115, UINT64_C(0xffffffff00000001));
+  fiat_p256_uint1 x126;
+  fiat_p256_addcarryx_u64(&x125, &x126, x124, 0x0, x120);
   uint64_t x127;
-  uint64_t x128;
-  fiat_p256_mulx_u64(&x127, &x128, x115, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x128;
+  fiat_p256_addcarryx_u64(&x127, &x128, 0x0, x121, x107);
   uint64_t x129;
-  uint64_t x130;
-  fiat_p256_mulx_u64(&x129, &x130, x115, UINT64_C(0xffffffffffffffff));
+  fiat_p256_uint1 x130;
+  fiat_p256_addcarryx_u64(&x129, &x130, x128, x123, x109);
   uint64_t x131;
   fiat_p256_uint1 x132;
-  fiat_p256_addcarryx_u64(&x131, &x132, 0x0, x127, x130);
+  fiat_p256_addcarryx_u64(&x131, &x132, x130, x125, x111);
   uint64_t x133;
   fiat_p256_uint1 x134;
-  fiat_p256_addcarryx_u64(&x133, &x134, x132, 0x0, x128);
+  fiat_p256_addcarryx_u64(&x133, &x134, x132, x117, x113);
   uint64_t x135;
   fiat_p256_uint1 x136;
-  fiat_p256_addcarryx_u64(&x135, &x136, x134, x125, 0x0);
+  fiat_p256_addcarryx_u64(&x135, &x136, x134, x118, x115);
   uint64_t x137;
   fiat_p256_uint1 x138;
-  fiat_p256_addcarryx_u64(&x137, &x138, x136, 0x0, x126);
+  fiat_p256_addcarryx_u64(&x137, &x138, x136, 0x0, x116);
   uint64_t x139;
-  fiat_p256_uint1 x140;
-  fiat_p256_addcarryx_u64(&x139, &x140, 0x0, x129, x115);
+  uint64_t x140;
+  fiat_p256_mulx_u64(&x139, &x140, x3, (arg2[3]));
   uint64_t x141;
-  fiat_p256_uint1 x142;
-  fiat_p256_addcarryx_u64(&x141, &x142, x140, x131, x117);
+  uint64_t x142;
+  fiat_p256_mulx_u64(&x141, &x142, x3, (arg2[2]));
   uint64_t x143;
-  fiat_p256_uint1 x144;
-  fiat_p256_addcarryx_u64(&x143, &x144, x142, x133, x119);
+  uint64_t x144;
+  fiat_p256_mulx_u64(&x143, &x144, x3, (arg2[1]));
   uint64_t x145;
-  fiat_p256_uint1 x146;
-  fiat_p256_addcarryx_u64(&x145, &x146, x144, x135, x121);
+  uint64_t x146;
+  fiat_p256_mulx_u64(&x145, &x146, x3, (arg2[0]));
   uint64_t x147;
   fiat_p256_uint1 x148;
-  fiat_p256_addcarryx_u64(&x147, &x148, x146, x137, x123);
+  fiat_p256_addcarryx_u64(&x147, &x148, 0x0, x143, x146);
   uint64_t x149;
   fiat_p256_uint1 x150;
-  fiat_p256_addcarryx_u64(&x149, &x150, x148, 0x0, x124);
+  fiat_p256_addcarryx_u64(&x149, &x150, x148, x141, x144);
   uint64_t x151;
-  uint64_t x152;
-  fiat_p256_mulx_u64(&x151, &x152, x3, (arg2[3]));
+  fiat_p256_uint1 x152;
+  fiat_p256_addcarryx_u64(&x151, &x152, x150, x139, x142);
   uint64_t x153;
-  uint64_t x154;
-  fiat_p256_mulx_u64(&x153, &x154, x3, (arg2[2]));
+  fiat_p256_uint1 x154;
+  fiat_p256_addcarryx_u64(&x153, &x154, x152, 0x0, x140);
   uint64_t x155;
-  uint64_t x156;
-  fiat_p256_mulx_u64(&x155, &x156, x3, (arg2[1]));
+  fiat_p256_uint1 x156;
+  fiat_p256_addcarryx_u64(&x155, &x156, 0x0, x145, x129);
   uint64_t x157;
-  uint64_t x158;
-  fiat_p256_mulx_u64(&x157, &x158, x3, (arg2[0]));
+  fiat_p256_uint1 x158;
+  fiat_p256_addcarryx_u64(&x157, &x158, x156, x147, x131);
   uint64_t x159;
   fiat_p256_uint1 x160;
-  fiat_p256_addcarryx_u64(&x159, &x160, 0x0, x155, x158);
+  fiat_p256_addcarryx_u64(&x159, &x160, x158, x149, x133);
   uint64_t x161;
   fiat_p256_uint1 x162;
-  fiat_p256_addcarryx_u64(&x161, &x162, x160, x153, x156);
+  fiat_p256_addcarryx_u64(&x161, &x162, x160, x151, x135);
   uint64_t x163;
   fiat_p256_uint1 x164;
-  fiat_p256_addcarryx_u64(&x163, &x164, x162, x151, x154);
+  fiat_p256_addcarryx_u64(&x163, &x164, x162, x153, x137);
   uint64_t x165;
-  fiat_p256_uint1 x166;
-  fiat_p256_addcarryx_u64(&x165, &x166, x164, 0x0, x152);
+  uint64_t x166;
+  fiat_p256_mulx_u64(&x165, &x166, x155, UINT64_C(0xffffffff00000001));
   uint64_t x167;
-  fiat_p256_uint1 x168;
-  fiat_p256_addcarryx_u64(&x167, &x168, 0x0, x157, x141);
+  uint64_t x168;
+  fiat_p256_mulx_u64(&x167, &x168, x155, UINT32_C(0xffffffff));
   uint64_t x169;
-  fiat_p256_uint1 x170;
-  fiat_p256_addcarryx_u64(&x169, &x170, x168, x159, x143);
+  uint64_t x170;
+  fiat_p256_mulx_u64(&x169, &x170, x155, UINT64_C(0xffffffffffffffff));
   uint64_t x171;
   fiat_p256_uint1 x172;
-  fiat_p256_addcarryx_u64(&x171, &x172, x170, x161, x145);
+  fiat_p256_addcarryx_u64(&x171, &x172, 0x0, x167, x170);
   uint64_t x173;
   fiat_p256_uint1 x174;
-  fiat_p256_addcarryx_u64(&x173, &x174, x172, x163, x147);
+  fiat_p256_addcarryx_u64(&x173, &x174, x172, 0x0, x168);
   uint64_t x175;
   fiat_p256_uint1 x176;
-  fiat_p256_addcarryx_u64(&x175, &x176, x174, x165, x149);
+  fiat_p256_addcarryx_u64(&x175, &x176, 0x0, x169, x155);
   uint64_t x177;
-  uint64_t x178;
-  fiat_p256_mulx_u64(&x177, &x178, x167, UINT64_C(0xffffffff00000001));
+  fiat_p256_uint1 x178;
+  fiat_p256_addcarryx_u64(&x177, &x178, x176, x171, x157);
   uint64_t x179;
-  uint64_t x180;
-  fiat_p256_mulx_u64(&x179, &x180, x167, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x180;
+  fiat_p256_addcarryx_u64(&x179, &x180, x178, x173, x159);
   uint64_t x181;
-  uint64_t x182;
-  fiat_p256_mulx_u64(&x181, &x182, x167, UINT64_C(0xffffffffffffffff));
+  fiat_p256_uint1 x182;
+  fiat_p256_addcarryx_u64(&x181, &x182, x180, x165, x161);
   uint64_t x183;
   fiat_p256_uint1 x184;
-  fiat_p256_addcarryx_u64(&x183, &x184, 0x0, x179, x182);
+  fiat_p256_addcarryx_u64(&x183, &x184, x182, x166, x163);
   uint64_t x185;
   fiat_p256_uint1 x186;
-  fiat_p256_addcarryx_u64(&x185, &x186, x184, 0x0, x180);
+  fiat_p256_addcarryx_u64(&x185, &x186, x184, 0x0, x164);
   uint64_t x187;
   fiat_p256_uint1 x188;
-  fiat_p256_addcarryx_u64(&x187, &x188, x186, x177, 0x0);
+  fiat_p256_subborrowx_u64(&x187, &x188, 0x0, x177, UINT64_C(0xffffffffffffffff));
   uint64_t x189;
   fiat_p256_uint1 x190;
-  fiat_p256_addcarryx_u64(&x189, &x190, x188, 0x0, x178);
+  fiat_p256_subborrowx_u64(&x189, &x190, x188, x179, UINT32_C(0xffffffff));
   uint64_t x191;
   fiat_p256_uint1 x192;
-  fiat_p256_addcarryx_u64(&x191, &x192, 0x0, x181, x167);
+  fiat_p256_subborrowx_u64(&x191, &x192, x190, x181, 0x0);
   uint64_t x193;
   fiat_p256_uint1 x194;
-  fiat_p256_addcarryx_u64(&x193, &x194, x192, x183, x169);
+  fiat_p256_subborrowx_u64(&x193, &x194, x192, x183, UINT64_C(0xffffffff00000001));
   uint64_t x195;
   fiat_p256_uint1 x196;
-  fiat_p256_addcarryx_u64(&x195, &x196, x194, x185, x171);
+  fiat_p256_subborrowx_u64(&x195, &x196, x194, x185, 0x0);
   uint64_t x197;
-  fiat_p256_uint1 x198;
-  fiat_p256_addcarryx_u64(&x197, &x198, x196, x187, x173);
+  fiat_p256_cmovznz_u64(&x197, x196, x187, x177);
+  uint64_t x198;
+  fiat_p256_cmovznz_u64(&x198, x196, x189, x179);
   uint64_t x199;
-  fiat_p256_uint1 x200;
-  fiat_p256_addcarryx_u64(&x199, &x200, x198, x189, x175);
-  uint64_t x201;
-  fiat_p256_uint1 x202;
-  fiat_p256_addcarryx_u64(&x201, &x202, x200, 0x0, x176);
-  uint64_t x203;
-  fiat_p256_uint1 x204;
-  fiat_p256_subborrowx_u64(&x203, &x204, 0x0, x193, UINT64_C(0xffffffffffffffff));
-  uint64_t x205;
-  fiat_p256_uint1 x206;
-  fiat_p256_subborrowx_u64(&x205, &x206, x204, x195, UINT32_C(0xffffffff));
-  uint64_t x207;
-  fiat_p256_uint1 x208;
-  fiat_p256_subborrowx_u64(&x207, &x208, x206, x197, 0x0);
-  uint64_t x209;
-  fiat_p256_uint1 x210;
-  fiat_p256_subborrowx_u64(&x209, &x210, x208, x199, UINT64_C(0xffffffff00000001));
-  uint64_t x211;
-  fiat_p256_uint1 x212;
-  fiat_p256_subborrowx_u64(&x211, &x212, x210, x201, 0x0);
-  uint64_t x213;
-  fiat_p256_cmovznz_u64(&x213, x212, x203, x193);
-  uint64_t x214;
-  fiat_p256_cmovznz_u64(&x214, x212, x205, x195);
-  uint64_t x215;
-  fiat_p256_cmovznz_u64(&x215, x212, x207, x197);
-  uint64_t x216;
-  fiat_p256_cmovznz_u64(&x216, x212, x209, x199);
-  out1[0] = x213;
-  out1[1] = x214;
-  out1[2] = x215;
-  out1[3] = x216;
+  fiat_p256_cmovznz_u64(&x199, x196, x191, x181);
+  uint64_t x200;
+  fiat_p256_cmovznz_u64(&x200, x196, x193, x183);
+  out1[0] = x197;
+  out1[1] = x198;
+  out1[2] = x199;
+  out1[3] = x200;
 }
 
 /*
@@ -473,289 +449,265 @@ static void fiat_p256_square(uint64_t out1[4], const uint64_t arg1[4]) {
   fiat_p256_addcarryx_u64(&x29, &x30, x28, 0x0, x24);
   uint64_t x31;
   fiat_p256_uint1 x32;
-  fiat_p256_addcarryx_u64(&x31, &x32, x30, x21, 0x0);
+  fiat_p256_addcarryx_u64(&x31, &x32, 0x0, x25, x11);
   uint64_t x33;
   fiat_p256_uint1 x34;
-  fiat_p256_addcarryx_u64(&x33, &x34, x32, 0x0, x22);
+  fiat_p256_addcarryx_u64(&x33, &x34, x32, x27, x13);
   uint64_t x35;
   fiat_p256_uint1 x36;
-  fiat_p256_addcarryx_u64(&x35, &x36, 0x0, x25, x11);
+  fiat_p256_addcarryx_u64(&x35, &x36, x34, x29, x15);
   uint64_t x37;
   fiat_p256_uint1 x38;
-  fiat_p256_addcarryx_u64(&x37, &x38, x36, x27, x13);
+  fiat_p256_addcarryx_u64(&x37, &x38, x36, x21, x17);
   uint64_t x39;
   fiat_p256_uint1 x40;
-  fiat_p256_addcarryx_u64(&x39, &x40, x38, x29, x15);
+  fiat_p256_addcarryx_u64(&x39, &x40, x38, x22, x19);
   uint64_t x41;
   fiat_p256_uint1 x42;
-  fiat_p256_addcarryx_u64(&x41, &x42, x40, x31, x17);
+  fiat_p256_addcarryx_u64(&x41, &x42, x40, 0x0, 0x0);
   uint64_t x43;
-  fiat_p256_uint1 x44;
-  fiat_p256_addcarryx_u64(&x43, &x44, x42, x33, x19);
+  uint64_t x44;
+  fiat_p256_mulx_u64(&x43, &x44, x1, (arg1[3]));
   uint64_t x45;
-  fiat_p256_uint1 x46;
-  fiat_p256_addcarryx_u64(&x45, &x46, x44, 0x0, 0x0);
+  uint64_t x46;
+  fiat_p256_mulx_u64(&x45, &x46, x1, (arg1[2]));
   uint64_t x47;
   uint64_t x48;
-  fiat_p256_mulx_u64(&x47, &x48, x1, (arg1[3]));
+  fiat_p256_mulx_u64(&x47, &x48, x1, (arg1[1]));
   uint64_t x49;
   uint64_t x50;
-  fiat_p256_mulx_u64(&x49, &x50, x1, (arg1[2]));
+  fiat_p256_mulx_u64(&x49, &x50, x1, (arg1[0]));
   uint64_t x51;
-  uint64_t x52;
-  fiat_p256_mulx_u64(&x51, &x52, x1, (arg1[1]));
+  fiat_p256_uint1 x52;
+  fiat_p256_addcarryx_u64(&x51, &x52, 0x0, x47, x50);
   uint64_t x53;
-  uint64_t x54;
-  fiat_p256_mulx_u64(&x53, &x54, x1, (arg1[0]));
+  fiat_p256_uint1 x54;
+  fiat_p256_addcarryx_u64(&x53, &x54, x52, x45, x48);
   uint64_t x55;
   fiat_p256_uint1 x56;
-  fiat_p256_addcarryx_u64(&x55, &x56, 0x0, x51, x54);
+  fiat_p256_addcarryx_u64(&x55, &x56, x54, x43, x46);
   uint64_t x57;
   fiat_p256_uint1 x58;
-  fiat_p256_addcarryx_u64(&x57, &x58, x56, x49, x52);
+  fiat_p256_addcarryx_u64(&x57, &x58, x56, 0x0, x44);
   uint64_t x59;
   fiat_p256_uint1 x60;
-  fiat_p256_addcarryx_u64(&x59, &x60, x58, x47, x50);
+  fiat_p256_addcarryx_u64(&x59, &x60, 0x0, x49, x33);
   uint64_t x61;
   fiat_p256_uint1 x62;
-  fiat_p256_addcarryx_u64(&x61, &x62, x60, 0x0, x48);
+  fiat_p256_addcarryx_u64(&x61, &x62, x60, x51, x35);
   uint64_t x63;
   fiat_p256_uint1 x64;
-  fiat_p256_addcarryx_u64(&x63, &x64, 0x0, x53, x37);
+  fiat_p256_addcarryx_u64(&x63, &x64, x62, x53, x37);
   uint64_t x65;
   fiat_p256_uint1 x66;
   fiat_p256_addcarryx_u64(&x65, &x66, x64, x55, x39);
   uint64_t x67;
   fiat_p256_uint1 x68;
-  fiat_p256_addcarryx_u64(&x67, &x68, x66, x57, x41);
+  fiat_p256_addcarryx_u64(&x67, &x68, x66, x57, (fiat_p256_uint1)x41);
   uint64_t x69;
-  fiat_p256_uint1 x70;
-  fiat_p256_addcarryx_u64(&x69, &x70, x68, x59, x43);
+  uint64_t x70;
+  fiat_p256_mulx_u64(&x69, &x70, x59, UINT64_C(0xffffffff00000001));
   uint64_t x71;
-  fiat_p256_uint1 x72;
-  fiat_p256_addcarryx_u64(&x71, &x72, x70, x61, (fiat_p256_uint1)x45);
+  uint64_t x72;
+  fiat_p256_mulx_u64(&x71, &x72, x59, UINT32_C(0xffffffff));
   uint64_t x73;
   uint64_t x74;
-  fiat_p256_mulx_u64(&x73, &x74, x63, UINT64_C(0xffffffff00000001));
+  fiat_p256_mulx_u64(&x73, &x74, x59, UINT64_C(0xffffffffffffffff));
   uint64_t x75;
-  uint64_t x76;
-  fiat_p256_mulx_u64(&x75, &x76, x63, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x76;
+  fiat_p256_addcarryx_u64(&x75, &x76, 0x0, x71, x74);
   uint64_t x77;
-  uint64_t x78;
-  fiat_p256_mulx_u64(&x77, &x78, x63, UINT64_C(0xffffffffffffffff));
+  fiat_p256_uint1 x78;
+  fiat_p256_addcarryx_u64(&x77, &x78, x76, 0x0, x72);
   uint64_t x79;
   fiat_p256_uint1 x80;
-  fiat_p256_addcarryx_u64(&x79, &x80, 0x0, x75, x78);
+  fiat_p256_addcarryx_u64(&x79, &x80, 0x0, x73, x59);
   uint64_t x81;
   fiat_p256_uint1 x82;
-  fiat_p256_addcarryx_u64(&x81, &x82, x80, 0x0, x76);
+  fiat_p256_addcarryx_u64(&x81, &x82, x80, x75, x61);
   uint64_t x83;
   fiat_p256_uint1 x84;
-  fiat_p256_addcarryx_u64(&x83, &x84, x82, x73, 0x0);
+  fiat_p256_addcarryx_u64(&x83, &x84, x82, x77, x63);
   uint64_t x85;
   fiat_p256_uint1 x86;
-  fiat_p256_addcarryx_u64(&x85, &x86, x84, 0x0, x74);
+  fiat_p256_addcarryx_u64(&x85, &x86, x84, x69, x65);
   uint64_t x87;
   fiat_p256_uint1 x88;
-  fiat_p256_addcarryx_u64(&x87, &x88, 0x0, x77, x63);
+  fiat_p256_addcarryx_u64(&x87, &x88, x86, x70, x67);
   uint64_t x89;
   fiat_p256_uint1 x90;
-  fiat_p256_addcarryx_u64(&x89, &x90, x88, x79, x65);
+  fiat_p256_addcarryx_u64(&x89, &x90, x88, 0x0, x68);
   uint64_t x91;
-  fiat_p256_uint1 x92;
-  fiat_p256_addcarryx_u64(&x91, &x92, x90, x81, x67);
+  uint64_t x92;
+  fiat_p256_mulx_u64(&x91, &x92, x2, (arg1[3]));
   uint64_t x93;
-  fiat_p256_uint1 x94;
-  fiat_p256_addcarryx_u64(&x93, &x94, x92, x83, x69);
+  uint64_t x94;
+  fiat_p256_mulx_u64(&x93, &x94, x2, (arg1[2]));
   uint64_t x95;
-  fiat_p256_uint1 x96;
-  fiat_p256_addcarryx_u64(&x95, &x96, x94, x85, x71);
+  uint64_t x96;
+  fiat_p256_mulx_u64(&x95, &x96, x2, (arg1[1]));
   uint64_t x97;
-  fiat_p256_uint1 x98;
-  fiat_p256_addcarryx_u64(&x97, &x98, x96, 0x0, x72);
+  uint64_t x98;
+  fiat_p256_mulx_u64(&x97, &x98, x2, (arg1[0]));
   uint64_t x99;
-  uint64_t x100;
-  fiat_p256_mulx_u64(&x99, &x100, x2, (arg1[3]));
+  fiat_p256_uint1 x100;
+  fiat_p256_addcarryx_u64(&x99, &x100, 0x0, x95, x98);
   uint64_t x101;
-  uint64_t x102;
-  fiat_p256_mulx_u64(&x101, &x102, x2, (arg1[2]));
+  fiat_p256_uint1 x102;
+  fiat_p256_addcarryx_u64(&x101, &x102, x100, x93, x96);
   uint64_t x103;
-  uint64_t x104;
-  fiat_p256_mulx_u64(&x103, &x104, x2, (arg1[1]));
+  fiat_p256_uint1 x104;
+  fiat_p256_addcarryx_u64(&x103, &x104, x102, x91, x94);
   uint64_t x105;
-  uint64_t x106;
-  fiat_p256_mulx_u64(&x105, &x106, x2, (arg1[0]));
+  fiat_p256_uint1 x106;
+  fiat_p256_addcarryx_u64(&x105, &x106, x104, 0x0, x92);
   uint64_t x107;
   fiat_p256_uint1 x108;
-  fiat_p256_addcarryx_u64(&x107, &x108, 0x0, x103, x106);
+  fiat_p256_addcarryx_u64(&x107, &x108, 0x0, x97, x81);
   uint64_t x109;
   fiat_p256_uint1 x110;
-  fiat_p256_addcarryx_u64(&x109, &x110, x108, x101, x104);
+  fiat_p256_addcarryx_u64(&x109, &x110, x108, x99, x83);
   uint64_t x111;
   fiat_p256_uint1 x112;
-  fiat_p256_addcarryx_u64(&x111, &x112, x110, x99, x102);
+  fiat_p256_addcarryx_u64(&x111, &x112, x110, x101, x85);
   uint64_t x113;
   fiat_p256_uint1 x114;
-  fiat_p256_addcarryx_u64(&x113, &x114, x112, 0x0, x100);
+  fiat_p256_addcarryx_u64(&x113, &x114, x112, x103, x87);
   uint64_t x115;
   fiat_p256_uint1 x116;
-  fiat_p256_addcarryx_u64(&x115, &x116, 0x0, x105, x89);
+  fiat_p256_addcarryx_u64(&x115, &x116, x114, x105, x89);
   uint64_t x117;
-  fiat_p256_uint1 x118;
-  fiat_p256_addcarryx_u64(&x117, &x118, x116, x107, x91);
+  uint64_t x118;
+  fiat_p256_mulx_u64(&x117, &x118, x107, UINT64_C(0xffffffff00000001));
   uint64_t x119;
-  fiat_p256_uint1 x120;
-  fiat_p256_addcarryx_u64(&x119, &x120, x118, x109, x93);
+  uint64_t x120;
+  fiat_p256_mulx_u64(&x119, &x120, x107, UINT32_C(0xffffffff));
   uint64_t x121;
-  fiat_p256_uint1 x122;
-  fiat_p256_addcarryx_u64(&x121, &x122, x120, x111, x95);
+  uint64_t x122;
+  fiat_p256_mulx_u64(&x121, &x122, x107, UINT64_C(0xffffffffffffffff));
   uint64_t x123;
   fiat_p256_uint1 x124;
-  fiat_p256_addcarryx_u64(&x123, &x124, x122, x113, x97);
+  fiat_p256_addcarryx_u64(&x123, &x124, 0x0, x119, x122);
   uint64_t x125;
-  uint64_t x126;
-  fiat_p256_mulx_u64(&x125, &x126, x115, UINT64_C(0xffffffff00000001));
+  fiat_p256_uint1 x126;
+  fiat_p256_addcarryx_u64(&x125, &x126, x124, 0x0, x120);
   uint64_t x127;
-  uint64_t x128;
-  fiat_p256_mulx_u64(&x127, &x128, x115, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x128;
+  fiat_p256_addcarryx_u64(&x127, &x128, 0x0, x121, x107);
   uint64_t x129;
-  uint64_t x130;
-  fiat_p256_mulx_u64(&x129, &x130, x115, UINT64_C(0xffffffffffffffff));
+  fiat_p256_uint1 x130;
+  fiat_p256_addcarryx_u64(&x129, &x130, x128, x123, x109);
   uint64_t x131;
   fiat_p256_uint1 x132;
-  fiat_p256_addcarryx_u64(&x131, &x132, 0x0, x127, x130);
+  fiat_p256_addcarryx_u64(&x131, &x132, x130, x125, x111);
   uint64_t x133;
   fiat_p256_uint1 x134;
-  fiat_p256_addcarryx_u64(&x133, &x134, x132, 0x0, x128);
+  fiat_p256_addcarryx_u64(&x133, &x134, x132, x117, x113);
   uint64_t x135;
   fiat_p256_uint1 x136;
-  fiat_p256_addcarryx_u64(&x135, &x136, x134, x125, 0x0);
+  fiat_p256_addcarryx_u64(&x135, &x136, x134, x118, x115);
   uint64_t x137;
   fiat_p256_uint1 x138;
-  fiat_p256_addcarryx_u64(&x137, &x138, x136, 0x0, x126);
+  fiat_p256_addcarryx_u64(&x137, &x138, x136, 0x0, x116);
   uint64_t x139;
-  fiat_p256_uint1 x140;
-  fiat_p256_addcarryx_u64(&x139, &x140, 0x0, x129, x115);
+  uint64_t x140;
+  fiat_p256_mulx_u64(&x139, &x140, x3, (arg1[3]));
   uint64_t x141;
-  fiat_p256_uint1 x142;
-  fiat_p256_addcarryx_u64(&x141, &x142, x140, x131, x117);
+  uint64_t x142;
+  fiat_p256_mulx_u64(&x141, &x142, x3, (arg1[2]));
   uint64_t x143;
-  fiat_p256_uint1 x144;
-  fiat_p256_addcarryx_u64(&x143, &x144, x142, x133, x119);
+  uint64_t x144;
+  fiat_p256_mulx_u64(&x143, &x144, x3, (arg1[1]));
   uint64_t x145;
-  fiat_p256_uint1 x146;
-  fiat_p256_addcarryx_u64(&x145, &x146, x144, x135, x121);
+  uint64_t x146;
+  fiat_p256_mulx_u64(&x145, &x146, x3, (arg1[0]));
   uint64_t x147;
   fiat_p256_uint1 x148;
-  fiat_p256_addcarryx_u64(&x147, &x148, x146, x137, x123);
+  fiat_p256_addcarryx_u64(&x147, &x148, 0x0, x143, x146);
   uint64_t x149;
   fiat_p256_uint1 x150;
-  fiat_p256_addcarryx_u64(&x149, &x150, x148, 0x0, x124);
+  fiat_p256_addcarryx_u64(&x149, &x150, x148, x141, x144);
   uint64_t x151;
-  uint64_t x152;
-  fiat_p256_mulx_u64(&x151, &x152, x3, (arg1[3]));
+  fiat_p256_uint1 x152;
+  fiat_p256_addcarryx_u64(&x151, &x152, x150, x139, x142);
   uint64_t x153;
-  uint64_t x154;
-  fiat_p256_mulx_u64(&x153, &x154, x3, (arg1[2]));
+  fiat_p256_uint1 x154;
+  fiat_p256_addcarryx_u64(&x153, &x154, x152, 0x0, x140);
   uint64_t x155;
-  uint64_t x156;
-  fiat_p256_mulx_u64(&x155, &x156, x3, (arg1[1]));
+  fiat_p256_uint1 x156;
+  fiat_p256_addcarryx_u64(&x155, &x156, 0x0, x145, x129);
   uint64_t x157;
-  uint64_t x158;
-  fiat_p256_mulx_u64(&x157, &x158, x3, (arg1[0]));
+  fiat_p256_uint1 x158;
+  fiat_p256_addcarryx_u64(&x157, &x158, x156, x147, x131);
   uint64_t x159;
   fiat_p256_uint1 x160;
-  fiat_p256_addcarryx_u64(&x159, &x160, 0x0, x155, x158);
+  fiat_p256_addcarryx_u64(&x159, &x160, x158, x149, x133);
   uint64_t x161;
   fiat_p256_uint1 x162;
-  fiat_p256_addcarryx_u64(&x161, &x162, x160, x153, x156);
+  fiat_p256_addcarryx_u64(&x161, &x162, x160, x151, x135);
   uint64_t x163;
   fiat_p256_uint1 x164;
-  fiat_p256_addcarryx_u64(&x163, &x164, x162, x151, x154);
+  fiat_p256_addcarryx_u64(&x163, &x164, x162, x153, x137);
   uint64_t x165;
-  fiat_p256_uint1 x166;
-  fiat_p256_addcarryx_u64(&x165, &x166, x164, 0x0, x152);
+  uint64_t x166;
+  fiat_p256_mulx_u64(&x165, &x166, x155, UINT64_C(0xffffffff00000001));
   uint64_t x167;
-  fiat_p256_uint1 x168;
-  fiat_p256_addcarryx_u64(&x167, &x168, 0x0, x157, x141);
+  uint64_t x168;
+  fiat_p256_mulx_u64(&x167, &x168, x155, UINT32_C(0xffffffff));
   uint64_t x169;
-  fiat_p256_uint1 x170;
-  fiat_p256_addcarryx_u64(&x169, &x170, x168, x159, x143);
+  uint64_t x170;
+  fiat_p256_mulx_u64(&x169, &x170, x155, UINT64_C(0xffffffffffffffff));
   uint64_t x171;
   fiat_p256_uint1 x172;
-  fiat_p256_addcarryx_u64(&x171, &x172, x170, x161, x145);
+  fiat_p256_addcarryx_u64(&x171, &x172, 0x0, x167, x170);
   uint64_t x173;
   fiat_p256_uint1 x174;
-  fiat_p256_addcarryx_u64(&x173, &x174, x172, x163, x147);
+  fiat_p256_addcarryx_u64(&x173, &x174, x172, 0x0, x168);
   uint64_t x175;
   fiat_p256_uint1 x176;
-  fiat_p256_addcarryx_u64(&x175, &x176, x174, x165, x149);
+  fiat_p256_addcarryx_u64(&x175, &x176, 0x0, x169, x155);
   uint64_t x177;
-  uint64_t x178;
-  fiat_p256_mulx_u64(&x177, &x178, x167, UINT64_C(0xffffffff00000001));
+  fiat_p256_uint1 x178;
+  fiat_p256_addcarryx_u64(&x177, &x178, x176, x171, x157);
   uint64_t x179;
-  uint64_t x180;
-  fiat_p256_mulx_u64(&x179, &x180, x167, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x180;
+  fiat_p256_addcarryx_u64(&x179, &x180, x178, x173, x159);
   uint64_t x181;
-  uint64_t x182;
-  fiat_p256_mulx_u64(&x181, &x182, x167, UINT64_C(0xffffffffffffffff));
+  fiat_p256_uint1 x182;
+  fiat_p256_addcarryx_u64(&x181, &x182, x180, x165, x161);
   uint64_t x183;
   fiat_p256_uint1 x184;
-  fiat_p256_addcarryx_u64(&x183, &x184, 0x0, x179, x182);
+  fiat_p256_addcarryx_u64(&x183, &x184, x182, x166, x163);
   uint64_t x185;
   fiat_p256_uint1 x186;
-  fiat_p256_addcarryx_u64(&x185, &x186, x184, 0x0, x180);
+  fiat_p256_addcarryx_u64(&x185, &x186, x184, 0x0, x164);
   uint64_t x187;
   fiat_p256_uint1 x188;
-  fiat_p256_addcarryx_u64(&x187, &x188, x186, x177, 0x0);
+  fiat_p256_subborrowx_u64(&x187, &x188, 0x0, x177, UINT64_C(0xffffffffffffffff));
   uint64_t x189;
   fiat_p256_uint1 x190;
-  fiat_p256_addcarryx_u64(&x189, &x190, x188, 0x0, x178);
+  fiat_p256_subborrowx_u64(&x189, &x190, x188, x179, UINT32_C(0xffffffff));
   uint64_t x191;
   fiat_p256_uint1 x192;
-  fiat_p256_addcarryx_u64(&x191, &x192, 0x0, x181, x167);
+  fiat_p256_subborrowx_u64(&x191, &x192, x190, x181, 0x0);
   uint64_t x193;
   fiat_p256_uint1 x194;
-  fiat_p256_addcarryx_u64(&x193, &x194, x192, x183, x169);
+  fiat_p256_subborrowx_u64(&x193, &x194, x192, x183, UINT64_C(0xffffffff00000001));
   uint64_t x195;
   fiat_p256_uint1 x196;
-  fiat_p256_addcarryx_u64(&x195, &x196, x194, x185, x171);
+  fiat_p256_subborrowx_u64(&x195, &x196, x194, x185, 0x0);
   uint64_t x197;
-  fiat_p256_uint1 x198;
-  fiat_p256_addcarryx_u64(&x197, &x198, x196, x187, x173);
+  fiat_p256_cmovznz_u64(&x197, x196, x187, x177);
+  uint64_t x198;
+  fiat_p256_cmovznz_u64(&x198, x196, x189, x179);
   uint64_t x199;
-  fiat_p256_uint1 x200;
-  fiat_p256_addcarryx_u64(&x199, &x200, x198, x189, x175);
-  uint64_t x201;
-  fiat_p256_uint1 x202;
-  fiat_p256_addcarryx_u64(&x201, &x202, x200, 0x0, x176);
-  uint64_t x203;
-  fiat_p256_uint1 x204;
-  fiat_p256_subborrowx_u64(&x203, &x204, 0x0, x193, UINT64_C(0xffffffffffffffff));
-  uint64_t x205;
-  fiat_p256_uint1 x206;
-  fiat_p256_subborrowx_u64(&x205, &x206, x204, x195, UINT32_C(0xffffffff));
-  uint64_t x207;
-  fiat_p256_uint1 x208;
-  fiat_p256_subborrowx_u64(&x207, &x208, x206, x197, 0x0);
-  uint64_t x209;
-  fiat_p256_uint1 x210;
-  fiat_p256_subborrowx_u64(&x209, &x210, x208, x199, UINT64_C(0xffffffff00000001));
-  uint64_t x211;
-  fiat_p256_uint1 x212;
-  fiat_p256_subborrowx_u64(&x211, &x212, x210, x201, 0x0);
-  uint64_t x213;
-  fiat_p256_cmovznz_u64(&x213, x212, x203, x193);
-  uint64_t x214;
-  fiat_p256_cmovznz_u64(&x214, x212, x205, x195);
-  uint64_t x215;
-  fiat_p256_cmovznz_u64(&x215, x212, x207, x197);
-  uint64_t x216;
-  fiat_p256_cmovznz_u64(&x216, x212, x209, x199);
-  out1[0] = x213;
-  out1[1] = x214;
-  out1[2] = x215;
-  out1[3] = x216;
+  fiat_p256_cmovznz_u64(&x199, x196, x191, x181);
+  uint64_t x200;
+  fiat_p256_cmovznz_u64(&x200, x196, x193, x183);
+  out1[0] = x197;
+  out1[1] = x198;
+  out1[2] = x199;
+  out1[3] = x200;
 }
 
 /*
@@ -908,220 +860,163 @@ static void fiat_p256_from_montgomery(uint64_t out1[4], const uint64_t arg1[4])
   fiat_p256_addcarryx_u64(&x8, &x9, 0x0, x4, x7);
   uint64_t x10;
   fiat_p256_uint1 x11;
-  fiat_p256_addcarryx_u64(&x10, &x11, x9, 0x0, x5);
+  fiat_p256_addcarryx_u64(&x10, &x11, 0x0, x6, x1);
   uint64_t x12;
   fiat_p256_uint1 x13;
-  fiat_p256_addcarryx_u64(&x12, &x13, x11, x2, 0x0);
+  fiat_p256_addcarryx_u64(&x12, &x13, x11, x8, 0x0);
   uint64_t x14;
   fiat_p256_uint1 x15;
-  fiat_p256_addcarryx_u64(&x14, &x15, 0x0, x6, x1);
+  fiat_p256_addcarryx_u64(&x14, &x15, 0x0, (arg1[1]), x12);
   uint64_t x16;
-  fiat_p256_uint1 x17;
-  fiat_p256_addcarryx_u64(&x16, &x17, x15, x8, 0x0);
+  uint64_t x17;
+  fiat_p256_mulx_u64(&x16, &x17, x14, UINT64_C(0xffffffff00000001));
   uint64_t x18;
-  fiat_p256_uint1 x19;
-  fiat_p256_addcarryx_u64(&x18, &x19, x17, x10, 0x0);
+  uint64_t x19;
+  fiat_p256_mulx_u64(&x18, &x19, x14, UINT32_C(0xffffffff));
   uint64_t x20;
-  fiat_p256_uint1 x21;
-  fiat_p256_addcarryx_u64(&x20, &x21, x19, x12, 0x0);
+  uint64_t x21;
+  fiat_p256_mulx_u64(&x20, &x21, x14, UINT64_C(0xffffffffffffffff));
   uint64_t x22;
   fiat_p256_uint1 x23;
-  fiat_p256_addcarryx_u64(&x22, &x23, x13, 0x0, x3);
+  fiat_p256_addcarryx_u64(&x22, &x23, 0x0, x18, x21);
   uint64_t x24;
   fiat_p256_uint1 x25;
-  fiat_p256_addcarryx_u64(&x24, &x25, x21, x22, 0x0);
+  fiat_p256_addcarryx_u64(&x24, &x25, x9, 0x0, x5);
   uint64_t x26;
   fiat_p256_uint1 x27;
-  fiat_p256_addcarryx_u64(&x26, &x27, 0x0, (arg1[1]), x16);
+  fiat_p256_addcarryx_u64(&x26, &x27, x13, x24, 0x0);
   uint64_t x28;
   fiat_p256_uint1 x29;
-  fiat_p256_addcarryx_u64(&x28, &x29, x27, 0x0, x18);
+  fiat_p256_addcarryx_u64(&x28, &x29, x15, 0x0, x26);
   uint64_t x30;
   fiat_p256_uint1 x31;
-  fiat_p256_addcarryx_u64(&x30, &x31, x29, 0x0, x20);
+  fiat_p256_addcarryx_u64(&x30, &x31, 0x0, x20, x14);
   uint64_t x32;
   fiat_p256_uint1 x33;
-  fiat_p256_addcarryx_u64(&x32, &x33, x31, 0x0, x24);
+  fiat_p256_addcarryx_u64(&x32, &x33, x31, x22, x28);
   uint64_t x34;
   fiat_p256_uint1 x35;
-  fiat_p256_addcarryx_u64(&x34, &x35, x25, 0x0, 0x0);
+  fiat_p256_addcarryx_u64(&x34, &x35, x23, 0x0, x19);
   uint64_t x36;
   fiat_p256_uint1 x37;
-  fiat_p256_addcarryx_u64(&x36, &x37, x33, 0x0, (fiat_p256_uint1)x34);
+  fiat_p256_addcarryx_u64(&x36, &x37, x33, x34, x2);
   uint64_t x38;
-  uint64_t x39;
-  fiat_p256_mulx_u64(&x38, &x39, x26, UINT64_C(0xffffffff00000001));
+  fiat_p256_uint1 x39;
+  fiat_p256_addcarryx_u64(&x38, &x39, x37, x16, x3);
   uint64_t x40;
-  uint64_t x41;
-  fiat_p256_mulx_u64(&x40, &x41, x26, UINT32_C(0xffffffff));
+  fiat_p256_uint1 x41;
+  fiat_p256_addcarryx_u64(&x40, &x41, 0x0, (arg1[2]), x32);
   uint64_t x42;
-  uint64_t x43;
-  fiat_p256_mulx_u64(&x42, &x43, x26, UINT64_C(0xffffffffffffffff));
+  fiat_p256_uint1 x43;
+  fiat_p256_addcarryx_u64(&x42, &x43, x41, 0x0, x36);
   uint64_t x44;
   fiat_p256_uint1 x45;
-  fiat_p256_addcarryx_u64(&x44, &x45, 0x0, x40, x43);
+  fiat_p256_addcarryx_u64(&x44, &x45, x43, 0x0, x38);
   uint64_t x46;
-  fiat_p256_uint1 x47;
-  fiat_p256_addcarryx_u64(&x46, &x47, x45, 0x0, x41);
+  uint64_t x47;
+  fiat_p256_mulx_u64(&x46, &x47, x40, UINT64_C(0xffffffff00000001));
   uint64_t x48;
-  fiat_p256_uint1 x49;
-  fiat_p256_addcarryx_u64(&x48, &x49, x47, x38, 0x0);
+  uint64_t x49;
+  fiat_p256_mulx_u64(&x48, &x49, x40, UINT32_C(0xffffffff));
   uint64_t x50;
-  fiat_p256_uint1 x51;
-  fiat_p256_addcarryx_u64(&x50, &x51, 0x0, x42, x26);
+  uint64_t x51;
+  fiat_p256_mulx_u64(&x50, &x51, x40, UINT64_C(0xffffffffffffffff));
   uint64_t x52;
   fiat_p256_uint1 x53;
-  fiat_p256_addcarryx_u64(&x52, &x53, x51, x44, x28);
+  fiat_p256_addcarryx_u64(&x52, &x53, 0x0, x48, x51);
   uint64_t x54;
   fiat_p256_uint1 x55;
-  fiat_p256_addcarryx_u64(&x54, &x55, x53, x46, x30);
+  fiat_p256_addcarryx_u64(&x54, &x55, 0x0, x50, x40);
   uint64_t x56;
   fiat_p256_uint1 x57;
-  fiat_p256_addcarryx_u64(&x56, &x57, x55, x48, x32);
+  fiat_p256_addcarryx_u64(&x56, &x57, x55, x52, x42);
   uint64_t x58;
   fiat_p256_uint1 x59;
-  fiat_p256_addcarryx_u64(&x58, &x59, x49, 0x0, x39);
+  fiat_p256_addcarryx_u64(&x58, &x59, x53, 0x0, x49);
   uint64_t x60;
   fiat_p256_uint1 x61;
-  fiat_p256_addcarryx_u64(&x60, &x61, x57, x58, (fiat_p256_uint1)x36);
+  fiat_p256_addcarryx_u64(&x60, &x61, x57, x58, x44);
   uint64_t x62;
   fiat_p256_uint1 x63;
-  fiat_p256_addcarryx_u64(&x62, &x63, 0x0, (arg1[2]), x52);
+  fiat_p256_addcarryx_u64(&x62, &x63, x39, x17, 0x0);
   uint64_t x64;
   fiat_p256_uint1 x65;
-  fiat_p256_addcarryx_u64(&x64, &x65, x63, 0x0, x54);
+  fiat_p256_addcarryx_u64(&x64, &x65, x45, 0x0, x62);
   uint64_t x66;
   fiat_p256_uint1 x67;
-  fiat_p256_addcarryx_u64(&x66, &x67, x65, 0x0, x56);
+  fiat_p256_addcarryx_u64(&x66, &x67, x61, x46, x64);
   uint64_t x68;
   fiat_p256_uint1 x69;
-  fiat_p256_addcarryx_u64(&x68, &x69, x67, 0x0, x60);
+  fiat_p256_addcarryx_u64(&x68, &x69, 0x0, (arg1[3]), x56);
   uint64_t x70;
   fiat_p256_uint1 x71;
-  fiat_p256_addcarryx_u64(&x70, &x71, x61, 0x0, x37);
+  fiat_p256_addcarryx_u64(&x70, &x71, x69, 0x0, x60);
   uint64_t x72;
   fiat_p256_uint1 x73;
-  fiat_p256_addcarryx_u64(&x72, &x73, x69, 0x0, (fiat_p256_uint1)x70);
+  fiat_p256_addcarryx_u64(&x72, &x73, x71, 0x0, x66);
   uint64_t x74;
   uint64_t x75;
-  fiat_p256_mulx_u64(&x74, &x75, x62, UINT64_C(0xffffffff00000001));
+  fiat_p256_mulx_u64(&x74, &x75, x68, UINT64_C(0xffffffff00000001));
   uint64_t x76;
   uint64_t x77;
-  fiat_p256_mulx_u64(&x76, &x77, x62, UINT32_C(0xffffffff));
+  fiat_p256_mulx_u64(&x76, &x77, x68, UINT32_C(0xffffffff));
   uint64_t x78;
   uint64_t x79;
-  fiat_p256_mulx_u64(&x78, &x79, x62, UINT64_C(0xffffffffffffffff));
+  fiat_p256_mulx_u64(&x78, &x79, x68, UINT64_C(0xffffffffffffffff));
   uint64_t x80;
   fiat_p256_uint1 x81;
   fiat_p256_addcarryx_u64(&x80, &x81, 0x0, x76, x79);
   uint64_t x82;
   fiat_p256_uint1 x83;
-  fiat_p256_addcarryx_u64(&x82, &x83, x81, 0x0, x77);
+  fiat_p256_addcarryx_u64(&x82, &x83, 0x0, x78, x68);
   uint64_t x84;
   fiat_p256_uint1 x85;
-  fiat_p256_addcarryx_u64(&x84, &x85, x83, x74, 0x0);
+  fiat_p256_addcarryx_u64(&x84, &x85, x83, x80, x70);
   uint64_t x86;
   fiat_p256_uint1 x87;
-  fiat_p256_addcarryx_u64(&x86, &x87, 0x0, x78, x62);
+  fiat_p256_addcarryx_u64(&x86, &x87, x81, 0x0, x77);
   uint64_t x88;
   fiat_p256_uint1 x89;
-  fiat_p256_addcarryx_u64(&x88, &x89, x87, x80, x64);
+  fiat_p256_addcarryx_u64(&x88, &x89, x85, x86, x72);
   uint64_t x90;
   fiat_p256_uint1 x91;
-  fiat_p256_addcarryx_u64(&x90, &x91, x89, x82, x66);
+  fiat_p256_addcarryx_u64(&x90, &x91, x67, x47, 0x0);
   uint64_t x92;
   fiat_p256_uint1 x93;
-  fiat_p256_addcarryx_u64(&x92, &x93, x91, x84, x68);
+  fiat_p256_addcarryx_u64(&x92, &x93, x73, 0x0, x90);
   uint64_t x94;
   fiat_p256_uint1 x95;
-  fiat_p256_addcarryx_u64(&x94, &x95, x85, 0x0, x75);
+  fiat_p256_addcarryx_u64(&x94, &x95, x89, x74, x92);
   uint64_t x96;
   fiat_p256_uint1 x97;
-  fiat_p256_addcarryx_u64(&x96, &x97, x93, x94, (fiat_p256_uint1)x72);
+  fiat_p256_addcarryx_u64(&x96, &x97, x95, x75, 0x0);
   uint64_t x98;
   fiat_p256_uint1 x99;
-  fiat_p256_addcarryx_u64(&x98, &x99, 0x0, (arg1[3]), x88);
+  fiat_p256_subborrowx_u64(&x98, &x99, 0x0, x84, UINT64_C(0xffffffffffffffff));
   uint64_t x100;
   fiat_p256_uint1 x101;
-  fiat_p256_addcarryx_u64(&x100, &x101, x99, 0x0, x90);
+  fiat_p256_subborrowx_u64(&x100, &x101, x99, x88, UINT32_C(0xffffffff));
   uint64_t x102;
   fiat_p256_uint1 x103;
-  fiat_p256_addcarryx_u64(&x102, &x103, x101, 0x0, x92);
+  fiat_p256_subborrowx_u64(&x102, &x103, x101, x94, 0x0);
   uint64_t x104;
   fiat_p256_uint1 x105;
-  fiat_p256_addcarryx_u64(&x104, &x105, x103, 0x0, x96);
+  fiat_p256_subborrowx_u64(&x104, &x105, x103, x96, UINT64_C(0xffffffff00000001));
   uint64_t x106;
   fiat_p256_uint1 x107;
-  fiat_p256_addcarryx_u64(&x106, &x107, x97, 0x0, x73);
+  fiat_p256_subborrowx_u64(&x106, &x107, x105, 0x0, 0x0);
   uint64_t x108;
-  fiat_p256_uint1 x109;
-  fiat_p256_addcarryx_u64(&x108, &x109, x105, 0x0, (fiat_p256_uint1)x106);
+  fiat_p256_cmovznz_u64(&x108, x107, x98, x84);
+  uint64_t x109;
+  fiat_p256_cmovznz_u64(&x109, x107, x100, x88);
   uint64_t x110;
+  fiat_p256_cmovznz_u64(&x110, x107, x102, x94);
   uint64_t x111;
-  fiat_p256_mulx_u64(&x110, &x111, x98, UINT64_C(0xffffffff00000001));
-  uint64_t x112;
-  uint64_t x113;
-  fiat_p256_mulx_u64(&x112, &x113, x98, UINT32_C(0xffffffff));
-  uint64_t x114;
-  uint64_t x115;
-  fiat_p256_mulx_u64(&x114, &x115, x98, UINT64_C(0xffffffffffffffff));
-  uint64_t x116;
-  fiat_p256_uint1 x117;
-  fiat_p256_addcarryx_u64(&x116, &x117, 0x0, x112, x115);
-  uint64_t x118;
-  fiat_p256_uint1 x119;
-  fiat_p256_addcarryx_u64(&x118, &x119, x117, 0x0, x113);
-  uint64_t x120;
-  fiat_p256_uint1 x121;
-  fiat_p256_addcarryx_u64(&x120, &x121, x119, x110, 0x0);
-  uint64_t x122;
-  fiat_p256_uint1 x123;
-  fiat_p256_addcarryx_u64(&x122, &x123, 0x0, x114, x98);
-  uint64_t x124;
-  fiat_p256_uint1 x125;
-  fiat_p256_addcarryx_u64(&x124, &x125, x123, x116, x100);
-  uint64_t x126;
-  fiat_p256_uint1 x127;
-  fiat_p256_addcarryx_u64(&x126, &x127, x125, x118, x102);
-  uint64_t x128;
-  fiat_p256_uint1 x129;
-  fiat_p256_addcarryx_u64(&x128, &x129, x127, x120, x104);
-  uint64_t x130;
-  fiat_p256_uint1 x131;
-  fiat_p256_addcarryx_u64(&x130, &x131, x121, 0x0, x111);
-  uint64_t x132;
-  fiat_p256_uint1 x133;
-  fiat_p256_addcarryx_u64(&x132, &x133, x129, x130, (fiat_p256_uint1)x108);
-  uint64_t x134;
-  fiat_p256_uint1 x135;
-  fiat_p256_subborrowx_u64(&x134, &x135, 0x0, x124, UINT64_C(0xffffffffffffffff));
-  uint64_t x136;
-  fiat_p256_uint1 x137;
-  fiat_p256_subborrowx_u64(&x136, &x137, x135, x126, UINT32_C(0xffffffff));
-  uint64_t x138;
-  fiat_p256_uint1 x139;
-  fiat_p256_subborrowx_u64(&x138, &x139, x137, x128, 0x0);
-  uint64_t x140;
-  fiat_p256_uint1 x141;
-  fiat_p256_subborrowx_u64(&x140, &x141, x139, x132, UINT64_C(0xffffffff00000001));
-  uint64_t x142;
-  fiat_p256_uint1 x143;
-  fiat_p256_addcarryx_u64(&x142, &x143, x133, 0x0, x109);
-  uint64_t x144;
-  fiat_p256_uint1 x145;
-  fiat_p256_subborrowx_u64(&x144, &x145, x141, (fiat_p256_uint1)x142, 0x0);
-  uint64_t x146;
-  fiat_p256_cmovznz_u64(&x146, x145, x134, x124);
-  uint64_t x147;
-  fiat_p256_cmovznz_u64(&x147, x145, x136, x126);
-  uint64_t x148;
-  fiat_p256_cmovznz_u64(&x148, x145, x138, x128);
-  uint64_t x149;
-  fiat_p256_cmovznz_u64(&x149, x145, x140, x132);
-  out1[0] = x146;
-  out1[1] = x147;
-  out1[2] = x148;
-  out1[3] = x149;
+  fiat_p256_cmovznz_u64(&x111, x107, x104, x96);
+  out1[0] = x108;
+  out1[1] = x109;
+  out1[2] = x110;
+  out1[3] = x111;
 }
 
 /*
@@ -1183,28 +1078,28 @@ static void fiat_p256_to_bytes(uint8_t out1[32], const uint64_t arg1[4]) {
   uint8_t x16 = (uint8_t)(x13 & UINT8_C(0xff));
   uint8_t x17 = (uint8_t)(x15 >> 8);
   uint8_t x18 = (uint8_t)(x15 & UINT8_C(0xff));
-  fiat_p256_uint1 x19 = (fiat_p256_uint1)(x17 >> 8);
-  uint8_t x20 = (uint8_t)(x17 & UINT8_C(0xff));
-  uint64_t x21 = (x19 + x3);
-  uint64_t x22 = (x21 >> 8);
-  uint8_t x23 = (uint8_t)(x21 & UINT8_C(0xff));
-  uint64_t x24 = (x22 >> 8);
-  uint8_t x25 = (uint8_t)(x22 & UINT8_C(0xff));
-  uint64_t x26 = (x24 >> 8);
-  uint8_t x27 = (uint8_t)(x24 & UINT8_C(0xff));
-  uint64_t x28 = (x26 >> 8);
-  uint8_t x29 = (uint8_t)(x26 & UINT8_C(0xff));
-  uint64_t x30 = (x28 >> 8);
-  uint8_t x31 = (uint8_t)(x28 & UINT8_C(0xff));
-  uint64_t x32 = (x30 >> 8);
-  uint8_t x33 = (uint8_t)(x30 & UINT8_C(0xff));
-  uint8_t x34 = (uint8_t)(x32 >> 8);
-  uint8_t x35 = (uint8_t)(x32 & UINT8_C(0xff));
-  fiat_p256_uint1 x36 = (fiat_p256_uint1)(x34 >> 8);
-  uint8_t x37 = (uint8_t)(x34 & UINT8_C(0xff));
-  uint64_t x38 = (x36 + x2);
-  uint64_t x39 = (x38 >> 8);
-  uint8_t x40 = (uint8_t)(x38 & UINT8_C(0xff));
+  uint8_t x19 = (uint8_t)(x17 & UINT8_C(0xff));
+  uint64_t x20 = (0x0 + x3);
+  uint64_t x21 = (x20 >> 8);
+  uint8_t x22 = (uint8_t)(x20 & UINT8_C(0xff));
+  uint64_t x23 = (x21 >> 8);
+  uint8_t x24 = (uint8_t)(x21 & UINT8_C(0xff));
+  uint64_t x25 = (x23 >> 8);
+  uint8_t x26 = (uint8_t)(x23 & UINT8_C(0xff));
+  uint64_t x27 = (x25 >> 8);
+  uint8_t x28 = (uint8_t)(x25 & UINT8_C(0xff));
+  uint64_t x29 = (x27 >> 8);
+  uint8_t x30 = (uint8_t)(x27 & UINT8_C(0xff));
+  uint64_t x31 = (x29 >> 8);
+  uint8_t x32 = (uint8_t)(x29 & UINT8_C(0xff));
+  uint8_t x33 = (uint8_t)(x31 >> 8);
+  uint8_t x34 = (uint8_t)(x31 & UINT8_C(0xff));
+  uint8_t x35 = (uint8_t)(x33 & UINT8_C(0xff));
+  uint64_t x36 = (0x0 + x2);
+  uint64_t x37 = (x36 >> 8);
+  uint8_t x38 = (uint8_t)(x36 & UINT8_C(0xff));
+  uint64_t x39 = (x37 >> 8);
+  uint8_t x40 = (uint8_t)(x37 & UINT8_C(0xff));
   uint64_t x41 = (x39 >> 8);
   uint8_t x42 = (uint8_t)(x39 & UINT8_C(0xff));
   uint64_t x43 = (x41 >> 8);
@@ -1213,27 +1108,24 @@ static void fiat_p256_to_bytes(uint8_t out1[32], const uint64_t arg1[4]) {
   uint8_t x46 = (uint8_t)(x43 & UINT8_C(0xff));
   uint64_t x47 = (x45 >> 8);
   uint8_t x48 = (uint8_t)(x45 & UINT8_C(0xff));
-  uint64_t x49 = (x47 >> 8);
+  uint8_t x49 = (uint8_t)(x47 >> 8);
   uint8_t x50 = (uint8_t)(x47 & UINT8_C(0xff));
-  uint8_t x51 = (uint8_t)(x49 >> 8);
-  uint8_t x52 = (uint8_t)(x49 & UINT8_C(0xff));
-  fiat_p256_uint1 x53 = (fiat_p256_uint1)(x51 >> 8);
-  uint8_t x54 = (uint8_t)(x51 & UINT8_C(0xff));
-  uint64_t x55 = (x53 + x1);
-  uint64_t x56 = (x55 >> 8);
-  uint8_t x57 = (uint8_t)(x55 & UINT8_C(0xff));
-  uint64_t x58 = (x56 >> 8);
-  uint8_t x59 = (uint8_t)(x56 & UINT8_C(0xff));
-  uint64_t x60 = (x58 >> 8);
-  uint8_t x61 = (uint8_t)(x58 & UINT8_C(0xff));
-  uint64_t x62 = (x60 >> 8);
-  uint8_t x63 = (uint8_t)(x60 & UINT8_C(0xff));
-  uint64_t x64 = (x62 >> 8);
-  uint8_t x65 = (uint8_t)(x62 & UINT8_C(0xff));
-  uint64_t x66 = (x64 >> 8);
-  uint8_t x67 = (uint8_t)(x64 & UINT8_C(0xff));
-  uint8_t x68 = (uint8_t)(x66 >> 8);
-  uint8_t x69 = (uint8_t)(x66 & UINT8_C(0xff));
+  uint8_t x51 = (uint8_t)(x49 & UINT8_C(0xff));
+  uint64_t x52 = (0x0 + x1);
+  uint64_t x53 = (x52 >> 8);
+  uint8_t x54 = (uint8_t)(x52 & UINT8_C(0xff));
+  uint64_t x55 = (x53 >> 8);
+  uint8_t x56 = (uint8_t)(x53 & UINT8_C(0xff));
+  uint64_t x57 = (x55 >> 8);
+  uint8_t x58 = (uint8_t)(x55 & UINT8_C(0xff));
+  uint64_t x59 = (x57 >> 8);
+  uint8_t x60 = (uint8_t)(x57 & UINT8_C(0xff));
+  uint64_t x61 = (x59 >> 8);
+  uint8_t x62 = (uint8_t)(x59 & UINT8_C(0xff));
+  uint64_t x63 = (x61 >> 8);
+  uint8_t x64 = (uint8_t)(x61 & UINT8_C(0xff));
+  uint8_t x65 = (uint8_t)(x63 >> 8);
+  uint8_t x66 = (uint8_t)(x63 & UINT8_C(0xff));
   out1[0] = x6;
   out1[1] = x8;
   out1[2] = x10;
@@ -1241,31 +1133,31 @@ static void fiat_p256_to_bytes(uint8_t out1[32], const uint64_t arg1[4]) {
   out1[4] = x14;
   out1[5] = x16;
   out1[6] = x18;
-  out1[7] = x20;
-  out1[8] = x23;
-  out1[9] = x25;
-  out1[10] = x27;
-  out1[11] = x29;
-  out1[12] = x31;
-  out1[13] = x33;
-  out1[14] = x35;
-  out1[15] = x37;
-  out1[16] = x40;
-  out1[17] = x42;
-  out1[18] = x44;
-  out1[19] = x46;
-  out1[20] = x48;
-  out1[21] = x50;
-  out1[22] = x52;
-  out1[23] = x54;
-  out1[24] = x57;
-  out1[25] = x59;
-  out1[26] = x61;
-  out1[27] = x63;
-  out1[28] = x65;
-  out1[29] = x67;
-  out1[30] = x69;
-  out1[31] = x68;
+  out1[7] = x19;
+  out1[8] = x22;
+  out1[9] = x24;
+  out1[10] = x26;
+  out1[11] = x28;
+  out1[12] = x30;
+  out1[13] = x32;
+  out1[14] = x34;
+  out1[15] = x35;
+  out1[16] = x38;
+  out1[17] = x40;
+  out1[18] = x42;
+  out1[19] = x44;
+  out1[20] = x46;
+  out1[21] = x48;
+  out1[22] = x50;
+  out1[23] = x51;
+  out1[24] = x54;
+  out1[25] = x56;
+  out1[26] = x58;
+  out1[27] = x60;
+  out1[28] = x62;
+  out1[29] = x64;
+  out1[30] = x66;
+  out1[31] = x65;
 }
 
 /*
@@ -1308,21 +1200,18 @@ static void fiat_p256_from_bytes(uint64_t out1[4], const uint8_t arg1[32]) {
   uint64_t x31 = ((uint64_t)(arg1[1]) << 8);
   uint8_t x32 = (arg1[0]);
   uint64_t x33 = (x32 + (x31 + (x30 + (x29 + (x28 + (x27 + (x26 + x25)))))));
-  fiat_p256_uint1 x34 = (fiat_p256_uint1)((fiat_p256_uint128)x33 >> 64);
-  uint64_t x35 = (x33 & UINT64_C(0xffffffffffffffff));
-  uint64_t x36 = (x8 + (x7 + (x6 + (x5 + (x4 + (x3 + (x2 + x1)))))));
-  uint64_t x37 = (x16 + (x15 + (x14 + (x13 + (x12 + (x11 + (x10 + x9)))))));
-  uint64_t x38 = (x24 + (x23 + (x22 + (x21 + (x20 + (x19 + (x18 + x17)))))));
-  uint64_t x39 = (x34 + x38);
-  fiat_p256_uint1 x40 = (fiat_p256_uint1)((fiat_p256_uint128)x39 >> 64);
-  uint64_t x41 = (x39 & UINT64_C(0xffffffffffffffff));
-  uint64_t x42 = (x40 + x37);
-  fiat_p256_uint1 x43 = (fiat_p256_uint1)((fiat_p256_uint128)x42 >> 64);
-  uint64_t x44 = (x42 & UINT64_C(0xffffffffffffffff));
-  uint64_t x45 = (x43 + x36);
-  out1[0] = x35;
-  out1[1] = x41;
-  out1[2] = x44;
-  out1[3] = x45;
+  uint64_t x34 = (x33 & UINT64_C(0xffffffffffffffff));
+  uint64_t x35 = (x8 + (x7 + (x6 + (x5 + (x4 + (x3 + (x2 + x1)))))));
+  uint64_t x36 = (x16 + (x15 + (x14 + (x13 + (x12 + (x11 + (x10 + x9)))))));
+  uint64_t x37 = (x24 + (x23 + (x22 + (x21 + (x20 + (x19 + (x18 + x17)))))));
+  uint64_t x38 = (0x0 + x37);
+  uint64_t x39 = (x38 & UINT64_C(0xffffffffffffffff));
+  uint64_t x40 = (0x0 + x36);
+  uint64_t x41 = (x40 & UINT64_C(0xffffffffffffffff));
+  uint64_t x42 = (0x0 + x35);
+  out1[0] = x34;
+  out1[1] = x39;
+  out1[2] = x41;
+  out1[3] = x42;
 }
 
diff --git a/p384_32.c b/p384_32.c
index 3c8f908af..9dc2368dd 100644
--- a/p384_32.c
+++ b/p384_32.c
@@ -5620,1601 +5620,1565 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_uint1 x43;
   fiat_p384_addcarryx_u32(&x42, &x43, 0x0, (arg1[1]), x40);
   uint32_t x44;
-  fiat_p384_uint1 x45;
-  fiat_p384_addcarryx_u32(&x44, &x45, x41, 0x0, 0x0);
+  uint32_t x45;
+  fiat_p384_mulx_u32(&x44, &x45, x42, UINT32_C(0xffffffff));
   uint32_t x46;
-  fiat_p384_uint1 x47;
-  fiat_p384_addcarryx_u32(&x46, &x47, x43, 0x0, (fiat_p384_uint1)x44);
+  uint32_t x47;
+  fiat_p384_mulx_u32(&x46, &x47, x42, UINT32_C(0xffffffff));
   uint32_t x48;
-  fiat_p384_uint1 x49;
-  fiat_p384_addcarryx_u32(&x48, &x49, x47, 0x0, x18);
+  uint32_t x49;
+  fiat_p384_mulx_u32(&x48, &x49, x42, UINT32_C(0xffffffff));
   uint32_t x50;
-  fiat_p384_uint1 x51;
-  fiat_p384_addcarryx_u32(&x50, &x51, x49, 0x0, x22);
+  uint32_t x51;
+  fiat_p384_mulx_u32(&x50, &x51, x42, UINT32_C(0xffffffff));
   uint32_t x52;
-  fiat_p384_uint1 x53;
-  fiat_p384_addcarryx_u32(&x52, &x53, x51, 0x0, x24);
+  uint32_t x53;
+  fiat_p384_mulx_u32(&x52, &x53, x42, UINT32_C(0xffffffff));
   uint32_t x54;
-  fiat_p384_uint1 x55;
-  fiat_p384_addcarryx_u32(&x54, &x55, x53, 0x0, x26);
+  uint32_t x55;
+  fiat_p384_mulx_u32(&x54, &x55, x42, UINT32_C(0xffffffff));
   uint32_t x56;
-  fiat_p384_uint1 x57;
-  fiat_p384_addcarryx_u32(&x56, &x57, x55, 0x0, x28);
+  uint32_t x57;
+  fiat_p384_mulx_u32(&x56, &x57, x42, UINT32_C(0xffffffff));
   uint32_t x58;
-  fiat_p384_uint1 x59;
-  fiat_p384_addcarryx_u32(&x58, &x59, x57, 0x0, x30);
+  uint32_t x59;
+  fiat_p384_mulx_u32(&x58, &x59, x42, UINT32_C(0xfffffffe));
   uint32_t x60;
-  fiat_p384_uint1 x61;
-  fiat_p384_addcarryx_u32(&x60, &x61, x59, 0x0, x32);
+  uint32_t x61;
+  fiat_p384_mulx_u32(&x60, &x61, x42, UINT32_C(0xffffffff));
   uint32_t x62;
-  fiat_p384_uint1 x63;
-  fiat_p384_addcarryx_u32(&x62, &x63, x61, 0x0, x34);
+  uint32_t x63;
+  fiat_p384_mulx_u32(&x62, &x63, x42, UINT32_C(0xffffffff));
   uint32_t x64;
   fiat_p384_uint1 x65;
-  fiat_p384_addcarryx_u32(&x64, &x65, x63, 0x0, x36);
+  fiat_p384_addcarryx_u32(&x64, &x65, 0x0, x58, x61);
   uint32_t x66;
   fiat_p384_uint1 x67;
-  fiat_p384_addcarryx_u32(&x66, &x67, x37, 0x0, x3);
+  fiat_p384_addcarryx_u32(&x66, &x67, x65, x56, x59);
   uint32_t x68;
   fiat_p384_uint1 x69;
-  fiat_p384_addcarryx_u32(&x68, &x69, x65, 0x0, x66);
+  fiat_p384_addcarryx_u32(&x68, &x69, x67, x54, x57);
   uint32_t x70;
-  uint32_t x71;
-  fiat_p384_mulx_u32(&x70, &x71, x42, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x71;
+  fiat_p384_addcarryx_u32(&x70, &x71, x69, x52, x55);
   uint32_t x72;
-  uint32_t x73;
-  fiat_p384_mulx_u32(&x72, &x73, x42, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x73;
+  fiat_p384_addcarryx_u32(&x72, &x73, x71, x50, x53);
   uint32_t x74;
-  uint32_t x75;
-  fiat_p384_mulx_u32(&x74, &x75, x42, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x75;
+  fiat_p384_addcarryx_u32(&x74, &x75, x73, x48, x51);
   uint32_t x76;
-  uint32_t x77;
-  fiat_p384_mulx_u32(&x76, &x77, x42, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x77;
+  fiat_p384_addcarryx_u32(&x76, &x77, x75, x46, x49);
   uint32_t x78;
-  uint32_t x79;
-  fiat_p384_mulx_u32(&x78, &x79, x42, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x79;
+  fiat_p384_addcarryx_u32(&x78, &x79, x77, x44, x47);
   uint32_t x80;
-  uint32_t x81;
-  fiat_p384_mulx_u32(&x80, &x81, x42, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x81;
+  fiat_p384_addcarryx_u32(&x80, &x81, x43, 0x0, 0x0);
   uint32_t x82;
-  uint32_t x83;
-  fiat_p384_mulx_u32(&x82, &x83, x42, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x83;
+  fiat_p384_addcarryx_u32(&x82, &x83, 0x0, x62, x42);
   uint32_t x84;
-  uint32_t x85;
-  fiat_p384_mulx_u32(&x84, &x85, x42, UINT32_C(0xfffffffe));
+  fiat_p384_uint1 x85;
+  fiat_p384_addcarryx_u32(&x84, &x85, x83, x63, (fiat_p384_uint1)x80);
   uint32_t x86;
-  uint32_t x87;
-  fiat_p384_mulx_u32(&x86, &x87, x42, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x87;
+  fiat_p384_addcarryx_u32(&x86, &x87, x85, 0x0, x18);
   uint32_t x88;
-  uint32_t x89;
-  fiat_p384_mulx_u32(&x88, &x89, x42, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x89;
+  fiat_p384_addcarryx_u32(&x88, &x89, x87, x60, x22);
   uint32_t x90;
   fiat_p384_uint1 x91;
-  fiat_p384_addcarryx_u32(&x90, &x91, 0x0, x84, x87);
+  fiat_p384_addcarryx_u32(&x90, &x91, x89, x64, x24);
   uint32_t x92;
   fiat_p384_uint1 x93;
-  fiat_p384_addcarryx_u32(&x92, &x93, x91, x82, x85);
+  fiat_p384_addcarryx_u32(&x92, &x93, x91, x66, x26);
   uint32_t x94;
   fiat_p384_uint1 x95;
-  fiat_p384_addcarryx_u32(&x94, &x95, x93, x80, x83);
+  fiat_p384_addcarryx_u32(&x94, &x95, x93, x68, x28);
   uint32_t x96;
   fiat_p384_uint1 x97;
-  fiat_p384_addcarryx_u32(&x96, &x97, x95, x78, x81);
+  fiat_p384_addcarryx_u32(&x96, &x97, x95, x70, x30);
   uint32_t x98;
   fiat_p384_uint1 x99;
-  fiat_p384_addcarryx_u32(&x98, &x99, x97, x76, x79);
+  fiat_p384_addcarryx_u32(&x98, &x99, x97, x72, x32);
   uint32_t x100;
   fiat_p384_uint1 x101;
-  fiat_p384_addcarryx_u32(&x100, &x101, x99, x74, x77);
+  fiat_p384_addcarryx_u32(&x100, &x101, x99, x74, x34);
   uint32_t x102;
   fiat_p384_uint1 x103;
-  fiat_p384_addcarryx_u32(&x102, &x103, x101, x72, x75);
+  fiat_p384_addcarryx_u32(&x102, &x103, x101, x76, x36);
   uint32_t x104;
   fiat_p384_uint1 x105;
-  fiat_p384_addcarryx_u32(&x104, &x105, x103, x70, x73);
+  fiat_p384_addcarryx_u32(&x104, &x105, x37, 0x0, x3);
   uint32_t x106;
   fiat_p384_uint1 x107;
-  fiat_p384_addcarryx_u32(&x106, &x107, 0x0, x88, x42);
+  fiat_p384_addcarryx_u32(&x106, &x107, x103, x78, x104);
   uint32_t x108;
   fiat_p384_uint1 x109;
-  fiat_p384_addcarryx_u32(&x108, &x109, x107, x89, (fiat_p384_uint1)x46);
+  fiat_p384_addcarryx_u32(&x108, &x109, x79, 0x0, x45);
   uint32_t x110;
   fiat_p384_uint1 x111;
-  fiat_p384_addcarryx_u32(&x110, &x111, x109, 0x0, x48);
+  fiat_p384_addcarryx_u32(&x110, &x111, x107, x108, 0x0);
   uint32_t x112;
   fiat_p384_uint1 x113;
-  fiat_p384_addcarryx_u32(&x112, &x113, x111, x86, x50);
+  fiat_p384_addcarryx_u32(&x112, &x113, 0x0, (arg1[2]), x84);
   uint32_t x114;
   fiat_p384_uint1 x115;
-  fiat_p384_addcarryx_u32(&x114, &x115, x113, x90, x52);
+  fiat_p384_addcarryx_u32(&x114, &x115, x113, 0x0, x86);
   uint32_t x116;
   fiat_p384_uint1 x117;
-  fiat_p384_addcarryx_u32(&x116, &x117, x115, x92, x54);
+  fiat_p384_addcarryx_u32(&x116, &x117, x115, 0x0, x88);
   uint32_t x118;
   fiat_p384_uint1 x119;
-  fiat_p384_addcarryx_u32(&x118, &x119, x117, x94, x56);
+  fiat_p384_addcarryx_u32(&x118, &x119, x117, 0x0, x90);
   uint32_t x120;
   fiat_p384_uint1 x121;
-  fiat_p384_addcarryx_u32(&x120, &x121, x119, x96, x58);
+  fiat_p384_addcarryx_u32(&x120, &x121, x119, 0x0, x92);
   uint32_t x122;
   fiat_p384_uint1 x123;
-  fiat_p384_addcarryx_u32(&x122, &x123, x121, x98, x60);
+  fiat_p384_addcarryx_u32(&x122, &x123, x121, 0x0, x94);
   uint32_t x124;
   fiat_p384_uint1 x125;
-  fiat_p384_addcarryx_u32(&x124, &x125, x123, x100, x62);
+  fiat_p384_addcarryx_u32(&x124, &x125, x123, 0x0, x96);
   uint32_t x126;
   fiat_p384_uint1 x127;
-  fiat_p384_addcarryx_u32(&x126, &x127, x125, x102, x64);
+  fiat_p384_addcarryx_u32(&x126, &x127, x125, 0x0, x98);
   uint32_t x128;
   fiat_p384_uint1 x129;
-  fiat_p384_addcarryx_u32(&x128, &x129, x127, x104, x68);
+  fiat_p384_addcarryx_u32(&x128, &x129, x127, 0x0, x100);
   uint32_t x130;
   fiat_p384_uint1 x131;
-  fiat_p384_addcarryx_u32(&x130, &x131, x69, 0x0, 0x0);
+  fiat_p384_addcarryx_u32(&x130, &x131, x129, 0x0, x102);
   uint32_t x132;
   fiat_p384_uint1 x133;
-  fiat_p384_addcarryx_u32(&x132, &x133, x105, 0x0, x71);
+  fiat_p384_addcarryx_u32(&x132, &x133, x131, 0x0, x106);
   uint32_t x134;
   fiat_p384_uint1 x135;
-  fiat_p384_addcarryx_u32(&x134, &x135, x129, x132, (fiat_p384_uint1)x130);
+  fiat_p384_addcarryx_u32(&x134, &x135, x133, 0x0, x110);
   uint32_t x136;
-  fiat_p384_uint1 x137;
-  fiat_p384_addcarryx_u32(&x136, &x137, 0x0, (arg1[2]), x108);
+  uint32_t x137;
+  fiat_p384_mulx_u32(&x136, &x137, x112, UINT32_C(0xffffffff));
   uint32_t x138;
-  fiat_p384_uint1 x139;
-  fiat_p384_addcarryx_u32(&x138, &x139, x137, 0x0, x110);
+  uint32_t x139;
+  fiat_p384_mulx_u32(&x138, &x139, x112, UINT32_C(0xffffffff));
   uint32_t x140;
-  fiat_p384_uint1 x141;
-  fiat_p384_addcarryx_u32(&x140, &x141, x139, 0x0, x112);
+  uint32_t x141;
+  fiat_p384_mulx_u32(&x140, &x141, x112, UINT32_C(0xffffffff));
   uint32_t x142;
-  fiat_p384_uint1 x143;
-  fiat_p384_addcarryx_u32(&x142, &x143, x141, 0x0, x114);
+  uint32_t x143;
+  fiat_p384_mulx_u32(&x142, &x143, x112, UINT32_C(0xffffffff));
   uint32_t x144;
-  fiat_p384_uint1 x145;
-  fiat_p384_addcarryx_u32(&x144, &x145, x143, 0x0, x116);
+  uint32_t x145;
+  fiat_p384_mulx_u32(&x144, &x145, x112, UINT32_C(0xffffffff));
   uint32_t x146;
-  fiat_p384_uint1 x147;
-  fiat_p384_addcarryx_u32(&x146, &x147, x145, 0x0, x118);
+  uint32_t x147;
+  fiat_p384_mulx_u32(&x146, &x147, x112, UINT32_C(0xffffffff));
   uint32_t x148;
-  fiat_p384_uint1 x149;
-  fiat_p384_addcarryx_u32(&x148, &x149, x147, 0x0, x120);
+  uint32_t x149;
+  fiat_p384_mulx_u32(&x148, &x149, x112, UINT32_C(0xffffffff));
   uint32_t x150;
-  fiat_p384_uint1 x151;
-  fiat_p384_addcarryx_u32(&x150, &x151, x149, 0x0, x122);
+  uint32_t x151;
+  fiat_p384_mulx_u32(&x150, &x151, x112, UINT32_C(0xfffffffe));
   uint32_t x152;
-  fiat_p384_uint1 x153;
-  fiat_p384_addcarryx_u32(&x152, &x153, x151, 0x0, x124);
+  uint32_t x153;
+  fiat_p384_mulx_u32(&x152, &x153, x112, UINT32_C(0xffffffff));
   uint32_t x154;
-  fiat_p384_uint1 x155;
-  fiat_p384_addcarryx_u32(&x154, &x155, x153, 0x0, x126);
+  uint32_t x155;
+  fiat_p384_mulx_u32(&x154, &x155, x112, UINT32_C(0xffffffff));
   uint32_t x156;
   fiat_p384_uint1 x157;
-  fiat_p384_addcarryx_u32(&x156, &x157, x155, 0x0, x128);
+  fiat_p384_addcarryx_u32(&x156, &x157, 0x0, x150, x153);
   uint32_t x158;
   fiat_p384_uint1 x159;
-  fiat_p384_addcarryx_u32(&x158, &x159, x157, 0x0, x134);
+  fiat_p384_addcarryx_u32(&x158, &x159, x157, x148, x151);
   uint32_t x160;
   fiat_p384_uint1 x161;
-  fiat_p384_addcarryx_u32(&x160, &x161, x135, 0x0, 0x0);
+  fiat_p384_addcarryx_u32(&x160, &x161, x159, x146, x149);
   uint32_t x162;
   fiat_p384_uint1 x163;
-  fiat_p384_addcarryx_u32(&x162, &x163, x159, 0x0, (fiat_p384_uint1)x160);
+  fiat_p384_addcarryx_u32(&x162, &x163, x161, x144, x147);
   uint32_t x164;
-  uint32_t x165;
-  fiat_p384_mulx_u32(&x164, &x165, x136, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x165;
+  fiat_p384_addcarryx_u32(&x164, &x165, x163, x142, x145);
   uint32_t x166;
-  uint32_t x167;
-  fiat_p384_mulx_u32(&x166, &x167, x136, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x167;
+  fiat_p384_addcarryx_u32(&x166, &x167, x165, x140, x143);
   uint32_t x168;
-  uint32_t x169;
-  fiat_p384_mulx_u32(&x168, &x169, x136, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x169;
+  fiat_p384_addcarryx_u32(&x168, &x169, x167, x138, x141);
   uint32_t x170;
-  uint32_t x171;
-  fiat_p384_mulx_u32(&x170, &x171, x136, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x171;
+  fiat_p384_addcarryx_u32(&x170, &x171, x169, x136, x139);
   uint32_t x172;
-  uint32_t x173;
-  fiat_p384_mulx_u32(&x172, &x173, x136, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x173;
+  fiat_p384_addcarryx_u32(&x172, &x173, 0x0, x154, x112);
   uint32_t x174;
-  uint32_t x175;
-  fiat_p384_mulx_u32(&x174, &x175, x136, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x175;
+  fiat_p384_addcarryx_u32(&x174, &x175, x173, x155, x114);
   uint32_t x176;
-  uint32_t x177;
-  fiat_p384_mulx_u32(&x176, &x177, x136, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x177;
+  fiat_p384_addcarryx_u32(&x176, &x177, x175, 0x0, x116);
   uint32_t x178;
-  uint32_t x179;
-  fiat_p384_mulx_u32(&x178, &x179, x136, UINT32_C(0xfffffffe));
+  fiat_p384_uint1 x179;
+  fiat_p384_addcarryx_u32(&x178, &x179, x177, x152, x118);
   uint32_t x180;
-  uint32_t x181;
-  fiat_p384_mulx_u32(&x180, &x181, x136, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x181;
+  fiat_p384_addcarryx_u32(&x180, &x181, x179, x156, x120);
   uint32_t x182;
-  uint32_t x183;
-  fiat_p384_mulx_u32(&x182, &x183, x136, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x183;
+  fiat_p384_addcarryx_u32(&x182, &x183, x181, x158, x122);
   uint32_t x184;
   fiat_p384_uint1 x185;
-  fiat_p384_addcarryx_u32(&x184, &x185, 0x0, x178, x181);
+  fiat_p384_addcarryx_u32(&x184, &x185, x183, x160, x124);
   uint32_t x186;
   fiat_p384_uint1 x187;
-  fiat_p384_addcarryx_u32(&x186, &x187, x185, x176, x179);
+  fiat_p384_addcarryx_u32(&x186, &x187, x185, x162, x126);
   uint32_t x188;
   fiat_p384_uint1 x189;
-  fiat_p384_addcarryx_u32(&x188, &x189, x187, x174, x177);
+  fiat_p384_addcarryx_u32(&x188, &x189, x187, x164, x128);
   uint32_t x190;
   fiat_p384_uint1 x191;
-  fiat_p384_addcarryx_u32(&x190, &x191, x189, x172, x175);
+  fiat_p384_addcarryx_u32(&x190, &x191, x189, x166, x130);
   uint32_t x192;
   fiat_p384_uint1 x193;
-  fiat_p384_addcarryx_u32(&x192, &x193, x191, x170, x173);
+  fiat_p384_addcarryx_u32(&x192, &x193, x191, x168, x132);
   uint32_t x194;
   fiat_p384_uint1 x195;
-  fiat_p384_addcarryx_u32(&x194, &x195, x193, x168, x171);
+  fiat_p384_addcarryx_u32(&x194, &x195, x193, x170, x134);
   uint32_t x196;
   fiat_p384_uint1 x197;
-  fiat_p384_addcarryx_u32(&x196, &x197, x195, x166, x169);
+  fiat_p384_addcarryx_u32(&x196, &x197, x111, 0x0, 0x0);
   uint32_t x198;
   fiat_p384_uint1 x199;
-  fiat_p384_addcarryx_u32(&x198, &x199, x197, x164, x167);
+  fiat_p384_addcarryx_u32(&x198, &x199, x135, 0x0, (fiat_p384_uint1)x196);
   uint32_t x200;
   fiat_p384_uint1 x201;
-  fiat_p384_addcarryx_u32(&x200, &x201, 0x0, x182, x136);
+  fiat_p384_addcarryx_u32(&x200, &x201, x171, 0x0, x137);
   uint32_t x202;
   fiat_p384_uint1 x203;
-  fiat_p384_addcarryx_u32(&x202, &x203, x201, x183, x138);
+  fiat_p384_addcarryx_u32(&x202, &x203, x195, x200, x198);
   uint32_t x204;
   fiat_p384_uint1 x205;
-  fiat_p384_addcarryx_u32(&x204, &x205, x203, 0x0, x140);
+  fiat_p384_addcarryx_u32(&x204, &x205, 0x0, (arg1[3]), x174);
   uint32_t x206;
   fiat_p384_uint1 x207;
-  fiat_p384_addcarryx_u32(&x206, &x207, x205, x180, x142);
+  fiat_p384_addcarryx_u32(&x206, &x207, x205, 0x0, x176);
   uint32_t x208;
   fiat_p384_uint1 x209;
-  fiat_p384_addcarryx_u32(&x208, &x209, x207, x184, x144);
+  fiat_p384_addcarryx_u32(&x208, &x209, x207, 0x0, x178);
   uint32_t x210;
   fiat_p384_uint1 x211;
-  fiat_p384_addcarryx_u32(&x210, &x211, x209, x186, x146);
+  fiat_p384_addcarryx_u32(&x210, &x211, x209, 0x0, x180);
   uint32_t x212;
   fiat_p384_uint1 x213;
-  fiat_p384_addcarryx_u32(&x212, &x213, x211, x188, x148);
+  fiat_p384_addcarryx_u32(&x212, &x213, x211, 0x0, x182);
   uint32_t x214;
   fiat_p384_uint1 x215;
-  fiat_p384_addcarryx_u32(&x214, &x215, x213, x190, x150);
+  fiat_p384_addcarryx_u32(&x214, &x215, x213, 0x0, x184);
   uint32_t x216;
   fiat_p384_uint1 x217;
-  fiat_p384_addcarryx_u32(&x216, &x217, x215, x192, x152);
+  fiat_p384_addcarryx_u32(&x216, &x217, x215, 0x0, x186);
   uint32_t x218;
   fiat_p384_uint1 x219;
-  fiat_p384_addcarryx_u32(&x218, &x219, x217, x194, x154);
+  fiat_p384_addcarryx_u32(&x218, &x219, x217, 0x0, x188);
   uint32_t x220;
   fiat_p384_uint1 x221;
-  fiat_p384_addcarryx_u32(&x220, &x221, x219, x196, x156);
+  fiat_p384_addcarryx_u32(&x220, &x221, x219, 0x0, x190);
   uint32_t x222;
   fiat_p384_uint1 x223;
-  fiat_p384_addcarryx_u32(&x222, &x223, x221, x198, x158);
+  fiat_p384_addcarryx_u32(&x222, &x223, x221, 0x0, x192);
   uint32_t x224;
   fiat_p384_uint1 x225;
-  fiat_p384_addcarryx_u32(&x224, &x225, x199, 0x0, x165);
+  fiat_p384_addcarryx_u32(&x224, &x225, x223, 0x0, x194);
   uint32_t x226;
   fiat_p384_uint1 x227;
-  fiat_p384_addcarryx_u32(&x226, &x227, x223, x224, x162);
+  fiat_p384_addcarryx_u32(&x226, &x227, x225, 0x0, x202);
   uint32_t x228;
-  fiat_p384_uint1 x229;
-  fiat_p384_addcarryx_u32(&x228, &x229, 0x0, (arg1[3]), x202);
+  uint32_t x229;
+  fiat_p384_mulx_u32(&x228, &x229, x204, UINT32_C(0xffffffff));
   uint32_t x230;
-  fiat_p384_uint1 x231;
-  fiat_p384_addcarryx_u32(&x230, &x231, x229, 0x0, x204);
+  uint32_t x231;
+  fiat_p384_mulx_u32(&x230, &x231, x204, UINT32_C(0xffffffff));
   uint32_t x232;
-  fiat_p384_uint1 x233;
-  fiat_p384_addcarryx_u32(&x232, &x233, x231, 0x0, x206);
+  uint32_t x233;
+  fiat_p384_mulx_u32(&x232, &x233, x204, UINT32_C(0xffffffff));
   uint32_t x234;
-  fiat_p384_uint1 x235;
-  fiat_p384_addcarryx_u32(&x234, &x235, x233, 0x0, x208);
+  uint32_t x235;
+  fiat_p384_mulx_u32(&x234, &x235, x204, UINT32_C(0xffffffff));
   uint32_t x236;
-  fiat_p384_uint1 x237;
-  fiat_p384_addcarryx_u32(&x236, &x237, x235, 0x0, x210);
+  uint32_t x237;
+  fiat_p384_mulx_u32(&x236, &x237, x204, UINT32_C(0xffffffff));
   uint32_t x238;
-  fiat_p384_uint1 x239;
-  fiat_p384_addcarryx_u32(&x238, &x239, x237, 0x0, x212);
+  uint32_t x239;
+  fiat_p384_mulx_u32(&x238, &x239, x204, UINT32_C(0xffffffff));
   uint32_t x240;
-  fiat_p384_uint1 x241;
-  fiat_p384_addcarryx_u32(&x240, &x241, x239, 0x0, x214);
+  uint32_t x241;
+  fiat_p384_mulx_u32(&x240, &x241, x204, UINT32_C(0xffffffff));
   uint32_t x242;
-  fiat_p384_uint1 x243;
-  fiat_p384_addcarryx_u32(&x242, &x243, x241, 0x0, x216);
+  uint32_t x243;
+  fiat_p384_mulx_u32(&x242, &x243, x204, UINT32_C(0xfffffffe));
   uint32_t x244;
-  fiat_p384_uint1 x245;
-  fiat_p384_addcarryx_u32(&x244, &x245, x243, 0x0, x218);
+  uint32_t x245;
+  fiat_p384_mulx_u32(&x244, &x245, x204, UINT32_C(0xffffffff));
   uint32_t x246;
-  fiat_p384_uint1 x247;
-  fiat_p384_addcarryx_u32(&x246, &x247, x245, 0x0, x220);
+  uint32_t x247;
+  fiat_p384_mulx_u32(&x246, &x247, x204, UINT32_C(0xffffffff));
   uint32_t x248;
   fiat_p384_uint1 x249;
-  fiat_p384_addcarryx_u32(&x248, &x249, x247, 0x0, x222);
+  fiat_p384_addcarryx_u32(&x248, &x249, 0x0, x242, x245);
   uint32_t x250;
   fiat_p384_uint1 x251;
-  fiat_p384_addcarryx_u32(&x250, &x251, x249, 0x0, x226);
+  fiat_p384_addcarryx_u32(&x250, &x251, x249, x240, x243);
   uint32_t x252;
   fiat_p384_uint1 x253;
-  fiat_p384_addcarryx_u32(&x252, &x253, x227, 0x0, x163);
+  fiat_p384_addcarryx_u32(&x252, &x253, x251, x238, x241);
   uint32_t x254;
   fiat_p384_uint1 x255;
-  fiat_p384_addcarryx_u32(&x254, &x255, x251, 0x0, (fiat_p384_uint1)x252);
+  fiat_p384_addcarryx_u32(&x254, &x255, x253, x236, x239);
   uint32_t x256;
-  uint32_t x257;
-  fiat_p384_mulx_u32(&x256, &x257, x228, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x257;
+  fiat_p384_addcarryx_u32(&x256, &x257, x255, x234, x237);
   uint32_t x258;
-  uint32_t x259;
-  fiat_p384_mulx_u32(&x258, &x259, x228, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x259;
+  fiat_p384_addcarryx_u32(&x258, &x259, x257, x232, x235);
   uint32_t x260;
-  uint32_t x261;
-  fiat_p384_mulx_u32(&x260, &x261, x228, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x261;
+  fiat_p384_addcarryx_u32(&x260, &x261, x259, x230, x233);
   uint32_t x262;
-  uint32_t x263;
-  fiat_p384_mulx_u32(&x262, &x263, x228, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x263;
+  fiat_p384_addcarryx_u32(&x262, &x263, x261, x228, x231);
   uint32_t x264;
-  uint32_t x265;
-  fiat_p384_mulx_u32(&x264, &x265, x228, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x265;
+  fiat_p384_addcarryx_u32(&x264, &x265, 0x0, x246, x204);
   uint32_t x266;
-  uint32_t x267;
-  fiat_p384_mulx_u32(&x266, &x267, x228, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x267;
+  fiat_p384_addcarryx_u32(&x266, &x267, x265, x247, x206);
   uint32_t x268;
-  uint32_t x269;
-  fiat_p384_mulx_u32(&x268, &x269, x228, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x269;
+  fiat_p384_addcarryx_u32(&x268, &x269, x267, 0x0, x208);
   uint32_t x270;
-  uint32_t x271;
-  fiat_p384_mulx_u32(&x270, &x271, x228, UINT32_C(0xfffffffe));
+  fiat_p384_uint1 x271;
+  fiat_p384_addcarryx_u32(&x270, &x271, x269, x244, x210);
   uint32_t x272;
-  uint32_t x273;
-  fiat_p384_mulx_u32(&x272, &x273, x228, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x273;
+  fiat_p384_addcarryx_u32(&x272, &x273, x271, x248, x212);
   uint32_t x274;
-  uint32_t x275;
-  fiat_p384_mulx_u32(&x274, &x275, x228, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x275;
+  fiat_p384_addcarryx_u32(&x274, &x275, x273, x250, x214);
   uint32_t x276;
   fiat_p384_uint1 x277;
-  fiat_p384_addcarryx_u32(&x276, &x277, 0x0, x270, x273);
+  fiat_p384_addcarryx_u32(&x276, &x277, x275, x252, x216);
   uint32_t x278;
   fiat_p384_uint1 x279;
-  fiat_p384_addcarryx_u32(&x278, &x279, x277, x268, x271);
+  fiat_p384_addcarryx_u32(&x278, &x279, x277, x254, x218);
   uint32_t x280;
   fiat_p384_uint1 x281;
-  fiat_p384_addcarryx_u32(&x280, &x281, x279, x266, x269);
+  fiat_p384_addcarryx_u32(&x280, &x281, x279, x256, x220);
   uint32_t x282;
   fiat_p384_uint1 x283;
-  fiat_p384_addcarryx_u32(&x282, &x283, x281, x264, x267);
+  fiat_p384_addcarryx_u32(&x282, &x283, x281, x258, x222);
   uint32_t x284;
   fiat_p384_uint1 x285;
-  fiat_p384_addcarryx_u32(&x284, &x285, x283, x262, x265);
+  fiat_p384_addcarryx_u32(&x284, &x285, x283, x260, x224);
   uint32_t x286;
   fiat_p384_uint1 x287;
-  fiat_p384_addcarryx_u32(&x286, &x287, x285, x260, x263);
+  fiat_p384_addcarryx_u32(&x286, &x287, x285, x262, x226);
   uint32_t x288;
   fiat_p384_uint1 x289;
-  fiat_p384_addcarryx_u32(&x288, &x289, x287, x258, x261);
+  fiat_p384_addcarryx_u32(&x288, &x289, x203, 0x0, 0x0);
   uint32_t x290;
   fiat_p384_uint1 x291;
-  fiat_p384_addcarryx_u32(&x290, &x291, x289, x256, x259);
+  fiat_p384_addcarryx_u32(&x290, &x291, x227, 0x0, (fiat_p384_uint1)x288);
   uint32_t x292;
   fiat_p384_uint1 x293;
-  fiat_p384_addcarryx_u32(&x292, &x293, 0x0, x274, x228);
+  fiat_p384_addcarryx_u32(&x292, &x293, x263, 0x0, x229);
   uint32_t x294;
   fiat_p384_uint1 x295;
-  fiat_p384_addcarryx_u32(&x294, &x295, x293, x275, x230);
+  fiat_p384_addcarryx_u32(&x294, &x295, x287, x292, x290);
   uint32_t x296;
   fiat_p384_uint1 x297;
-  fiat_p384_addcarryx_u32(&x296, &x297, x295, 0x0, x232);
+  fiat_p384_addcarryx_u32(&x296, &x297, 0x0, (arg1[4]), x266);
   uint32_t x298;
   fiat_p384_uint1 x299;
-  fiat_p384_addcarryx_u32(&x298, &x299, x297, x272, x234);
+  fiat_p384_addcarryx_u32(&x298, &x299, x297, 0x0, x268);
   uint32_t x300;
   fiat_p384_uint1 x301;
-  fiat_p384_addcarryx_u32(&x300, &x301, x299, x276, x236);
+  fiat_p384_addcarryx_u32(&x300, &x301, x299, 0x0, x270);
   uint32_t x302;
   fiat_p384_uint1 x303;
-  fiat_p384_addcarryx_u32(&x302, &x303, x301, x278, x238);
+  fiat_p384_addcarryx_u32(&x302, &x303, x301, 0x0, x272);
   uint32_t x304;
   fiat_p384_uint1 x305;
-  fiat_p384_addcarryx_u32(&x304, &x305, x303, x280, x240);
+  fiat_p384_addcarryx_u32(&x304, &x305, x303, 0x0, x274);
   uint32_t x306;
   fiat_p384_uint1 x307;
-  fiat_p384_addcarryx_u32(&x306, &x307, x305, x282, x242);
+  fiat_p384_addcarryx_u32(&x306, &x307, x305, 0x0, x276);
   uint32_t x308;
   fiat_p384_uint1 x309;
-  fiat_p384_addcarryx_u32(&x308, &x309, x307, x284, x244);
+  fiat_p384_addcarryx_u32(&x308, &x309, x307, 0x0, x278);
   uint32_t x310;
   fiat_p384_uint1 x311;
-  fiat_p384_addcarryx_u32(&x310, &x311, x309, x286, x246);
+  fiat_p384_addcarryx_u32(&x310, &x311, x309, 0x0, x280);
   uint32_t x312;
   fiat_p384_uint1 x313;
-  fiat_p384_addcarryx_u32(&x312, &x313, x311, x288, x248);
+  fiat_p384_addcarryx_u32(&x312, &x313, x311, 0x0, x282);
   uint32_t x314;
   fiat_p384_uint1 x315;
-  fiat_p384_addcarryx_u32(&x314, &x315, x313, x290, x250);
+  fiat_p384_addcarryx_u32(&x314, &x315, x313, 0x0, x284);
   uint32_t x316;
   fiat_p384_uint1 x317;
-  fiat_p384_addcarryx_u32(&x316, &x317, x291, 0x0, x257);
+  fiat_p384_addcarryx_u32(&x316, &x317, x315, 0x0, x286);
   uint32_t x318;
   fiat_p384_uint1 x319;
-  fiat_p384_addcarryx_u32(&x318, &x319, x315, x316, x254);
+  fiat_p384_addcarryx_u32(&x318, &x319, x317, 0x0, x294);
   uint32_t x320;
-  fiat_p384_uint1 x321;
-  fiat_p384_addcarryx_u32(&x320, &x321, 0x0, (arg1[4]), x294);
+  uint32_t x321;
+  fiat_p384_mulx_u32(&x320, &x321, x296, UINT32_C(0xffffffff));
   uint32_t x322;
-  fiat_p384_uint1 x323;
-  fiat_p384_addcarryx_u32(&x322, &x323, x321, 0x0, x296);
+  uint32_t x323;
+  fiat_p384_mulx_u32(&x322, &x323, x296, UINT32_C(0xffffffff));
   uint32_t x324;
-  fiat_p384_uint1 x325;
-  fiat_p384_addcarryx_u32(&x324, &x325, x323, 0x0, x298);
+  uint32_t x325;
+  fiat_p384_mulx_u32(&x324, &x325, x296, UINT32_C(0xffffffff));
   uint32_t x326;
-  fiat_p384_uint1 x327;
-  fiat_p384_addcarryx_u32(&x326, &x327, x325, 0x0, x300);
+  uint32_t x327;
+  fiat_p384_mulx_u32(&x326, &x327, x296, UINT32_C(0xffffffff));
   uint32_t x328;
-  fiat_p384_uint1 x329;
-  fiat_p384_addcarryx_u32(&x328, &x329, x327, 0x0, x302);
+  uint32_t x329;
+  fiat_p384_mulx_u32(&x328, &x329, x296, UINT32_C(0xffffffff));
   uint32_t x330;
-  fiat_p384_uint1 x331;
-  fiat_p384_addcarryx_u32(&x330, &x331, x329, 0x0, x304);
+  uint32_t x331;
+  fiat_p384_mulx_u32(&x330, &x331, x296, UINT32_C(0xffffffff));
   uint32_t x332;
-  fiat_p384_uint1 x333;
-  fiat_p384_addcarryx_u32(&x332, &x333, x331, 0x0, x306);
+  uint32_t x333;
+  fiat_p384_mulx_u32(&x332, &x333, x296, UINT32_C(0xffffffff));
   uint32_t x334;
-  fiat_p384_uint1 x335;
-  fiat_p384_addcarryx_u32(&x334, &x335, x333, 0x0, x308);
+  uint32_t x335;
+  fiat_p384_mulx_u32(&x334, &x335, x296, UINT32_C(0xfffffffe));
   uint32_t x336;
-  fiat_p384_uint1 x337;
-  fiat_p384_addcarryx_u32(&x336, &x337, x335, 0x0, x310);
+  uint32_t x337;
+  fiat_p384_mulx_u32(&x336, &x337, x296, UINT32_C(0xffffffff));
   uint32_t x338;
-  fiat_p384_uint1 x339;
-  fiat_p384_addcarryx_u32(&x338, &x339, x337, 0x0, x312);
+  uint32_t x339;
+  fiat_p384_mulx_u32(&x338, &x339, x296, UINT32_C(0xffffffff));
   uint32_t x340;
   fiat_p384_uint1 x341;
-  fiat_p384_addcarryx_u32(&x340, &x341, x339, 0x0, x314);
+  fiat_p384_addcarryx_u32(&x340, &x341, 0x0, x334, x337);
   uint32_t x342;
   fiat_p384_uint1 x343;
-  fiat_p384_addcarryx_u32(&x342, &x343, x341, 0x0, x318);
+  fiat_p384_addcarryx_u32(&x342, &x343, x341, x332, x335);
   uint32_t x344;
   fiat_p384_uint1 x345;
-  fiat_p384_addcarryx_u32(&x344, &x345, x319, 0x0, x255);
+  fiat_p384_addcarryx_u32(&x344, &x345, x343, x330, x333);
   uint32_t x346;
   fiat_p384_uint1 x347;
-  fiat_p384_addcarryx_u32(&x346, &x347, x343, 0x0, (fiat_p384_uint1)x344);
+  fiat_p384_addcarryx_u32(&x346, &x347, x345, x328, x331);
   uint32_t x348;
-  uint32_t x349;
-  fiat_p384_mulx_u32(&x348, &x349, x320, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x349;
+  fiat_p384_addcarryx_u32(&x348, &x349, x347, x326, x329);
   uint32_t x350;
-  uint32_t x351;
-  fiat_p384_mulx_u32(&x350, &x351, x320, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x351;
+  fiat_p384_addcarryx_u32(&x350, &x351, x349, x324, x327);
   uint32_t x352;
-  uint32_t x353;
-  fiat_p384_mulx_u32(&x352, &x353, x320, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x353;
+  fiat_p384_addcarryx_u32(&x352, &x353, x351, x322, x325);
   uint32_t x354;
-  uint32_t x355;
-  fiat_p384_mulx_u32(&x354, &x355, x320, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x355;
+  fiat_p384_addcarryx_u32(&x354, &x355, x353, x320, x323);
   uint32_t x356;
-  uint32_t x357;
-  fiat_p384_mulx_u32(&x356, &x357, x320, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x357;
+  fiat_p384_addcarryx_u32(&x356, &x357, 0x0, x338, x296);
   uint32_t x358;
-  uint32_t x359;
-  fiat_p384_mulx_u32(&x358, &x359, x320, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x359;
+  fiat_p384_addcarryx_u32(&x358, &x359, x357, x339, x298);
   uint32_t x360;
-  uint32_t x361;
-  fiat_p384_mulx_u32(&x360, &x361, x320, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x361;
+  fiat_p384_addcarryx_u32(&x360, &x361, x359, 0x0, x300);
   uint32_t x362;
-  uint32_t x363;
-  fiat_p384_mulx_u32(&x362, &x363, x320, UINT32_C(0xfffffffe));
+  fiat_p384_uint1 x363;
+  fiat_p384_addcarryx_u32(&x362, &x363, x361, x336, x302);
   uint32_t x364;
-  uint32_t x365;
-  fiat_p384_mulx_u32(&x364, &x365, x320, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x365;
+  fiat_p384_addcarryx_u32(&x364, &x365, x363, x340, x304);
   uint32_t x366;
-  uint32_t x367;
-  fiat_p384_mulx_u32(&x366, &x367, x320, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x367;
+  fiat_p384_addcarryx_u32(&x366, &x367, x365, x342, x306);
   uint32_t x368;
   fiat_p384_uint1 x369;
-  fiat_p384_addcarryx_u32(&x368, &x369, 0x0, x362, x365);
+  fiat_p384_addcarryx_u32(&x368, &x369, x367, x344, x308);
   uint32_t x370;
   fiat_p384_uint1 x371;
-  fiat_p384_addcarryx_u32(&x370, &x371, x369, x360, x363);
+  fiat_p384_addcarryx_u32(&x370, &x371, x369, x346, x310);
   uint32_t x372;
   fiat_p384_uint1 x373;
-  fiat_p384_addcarryx_u32(&x372, &x373, x371, x358, x361);
+  fiat_p384_addcarryx_u32(&x372, &x373, x371, x348, x312);
   uint32_t x374;
   fiat_p384_uint1 x375;
-  fiat_p384_addcarryx_u32(&x374, &x375, x373, x356, x359);
+  fiat_p384_addcarryx_u32(&x374, &x375, x373, x350, x314);
   uint32_t x376;
   fiat_p384_uint1 x377;
-  fiat_p384_addcarryx_u32(&x376, &x377, x375, x354, x357);
+  fiat_p384_addcarryx_u32(&x376, &x377, x375, x352, x316);
   uint32_t x378;
   fiat_p384_uint1 x379;
-  fiat_p384_addcarryx_u32(&x378, &x379, x377, x352, x355);
+  fiat_p384_addcarryx_u32(&x378, &x379, x377, x354, x318);
   uint32_t x380;
   fiat_p384_uint1 x381;
-  fiat_p384_addcarryx_u32(&x380, &x381, x379, x350, x353);
+  fiat_p384_addcarryx_u32(&x380, &x381, x295, 0x0, 0x0);
   uint32_t x382;
   fiat_p384_uint1 x383;
-  fiat_p384_addcarryx_u32(&x382, &x383, x381, x348, x351);
+  fiat_p384_addcarryx_u32(&x382, &x383, x319, 0x0, (fiat_p384_uint1)x380);
   uint32_t x384;
   fiat_p384_uint1 x385;
-  fiat_p384_addcarryx_u32(&x384, &x385, 0x0, x366, x320);
+  fiat_p384_addcarryx_u32(&x384, &x385, x355, 0x0, x321);
   uint32_t x386;
   fiat_p384_uint1 x387;
-  fiat_p384_addcarryx_u32(&x386, &x387, x385, x367, x322);
+  fiat_p384_addcarryx_u32(&x386, &x387, x379, x384, x382);
   uint32_t x388;
   fiat_p384_uint1 x389;
-  fiat_p384_addcarryx_u32(&x388, &x389, x387, 0x0, x324);
+  fiat_p384_addcarryx_u32(&x388, &x389, 0x0, (arg1[5]), x358);
   uint32_t x390;
   fiat_p384_uint1 x391;
-  fiat_p384_addcarryx_u32(&x390, &x391, x389, x364, x326);
+  fiat_p384_addcarryx_u32(&x390, &x391, x389, 0x0, x360);
   uint32_t x392;
   fiat_p384_uint1 x393;
-  fiat_p384_addcarryx_u32(&x392, &x393, x391, x368, x328);
+  fiat_p384_addcarryx_u32(&x392, &x393, x391, 0x0, x362);
   uint32_t x394;
   fiat_p384_uint1 x395;
-  fiat_p384_addcarryx_u32(&x394, &x395, x393, x370, x330);
+  fiat_p384_addcarryx_u32(&x394, &x395, x393, 0x0, x364);
   uint32_t x396;
   fiat_p384_uint1 x397;
-  fiat_p384_addcarryx_u32(&x396, &x397, x395, x372, x332);
+  fiat_p384_addcarryx_u32(&x396, &x397, x395, 0x0, x366);
   uint32_t x398;
   fiat_p384_uint1 x399;
-  fiat_p384_addcarryx_u32(&x398, &x399, x397, x374, x334);
+  fiat_p384_addcarryx_u32(&x398, &x399, x397, 0x0, x368);
   uint32_t x400;
   fiat_p384_uint1 x401;
-  fiat_p384_addcarryx_u32(&x400, &x401, x399, x376, x336);
+  fiat_p384_addcarryx_u32(&x400, &x401, x399, 0x0, x370);
   uint32_t x402;
   fiat_p384_uint1 x403;
-  fiat_p384_addcarryx_u32(&x402, &x403, x401, x378, x338);
+  fiat_p384_addcarryx_u32(&x402, &x403, x401, 0x0, x372);
   uint32_t x404;
   fiat_p384_uint1 x405;
-  fiat_p384_addcarryx_u32(&x404, &x405, x403, x380, x340);
+  fiat_p384_addcarryx_u32(&x404, &x405, x403, 0x0, x374);
   uint32_t x406;
   fiat_p384_uint1 x407;
-  fiat_p384_addcarryx_u32(&x406, &x407, x405, x382, x342);
+  fiat_p384_addcarryx_u32(&x406, &x407, x405, 0x0, x376);
   uint32_t x408;
   fiat_p384_uint1 x409;
-  fiat_p384_addcarryx_u32(&x408, &x409, x383, 0x0, x349);
+  fiat_p384_addcarryx_u32(&x408, &x409, x407, 0x0, x378);
   uint32_t x410;
   fiat_p384_uint1 x411;
-  fiat_p384_addcarryx_u32(&x410, &x411, x407, x408, x346);
+  fiat_p384_addcarryx_u32(&x410, &x411, x409, 0x0, x386);
   uint32_t x412;
-  fiat_p384_uint1 x413;
-  fiat_p384_addcarryx_u32(&x412, &x413, 0x0, (arg1[5]), x386);
+  uint32_t x413;
+  fiat_p384_mulx_u32(&x412, &x413, x388, UINT32_C(0xffffffff));
   uint32_t x414;
-  fiat_p384_uint1 x415;
-  fiat_p384_addcarryx_u32(&x414, &x415, x413, 0x0, x388);
+  uint32_t x415;
+  fiat_p384_mulx_u32(&x414, &x415, x388, UINT32_C(0xffffffff));
   uint32_t x416;
-  fiat_p384_uint1 x417;
-  fiat_p384_addcarryx_u32(&x416, &x417, x415, 0x0, x390);
+  uint32_t x417;
+  fiat_p384_mulx_u32(&x416, &x417, x388, UINT32_C(0xffffffff));
   uint32_t x418;
-  fiat_p384_uint1 x419;
-  fiat_p384_addcarryx_u32(&x418, &x419, x417, 0x0, x392);
+  uint32_t x419;
+  fiat_p384_mulx_u32(&x418, &x419, x388, UINT32_C(0xffffffff));
   uint32_t x420;
-  fiat_p384_uint1 x421;
-  fiat_p384_addcarryx_u32(&x420, &x421, x419, 0x0, x394);
+  uint32_t x421;
+  fiat_p384_mulx_u32(&x420, &x421, x388, UINT32_C(0xffffffff));
   uint32_t x422;
-  fiat_p384_uint1 x423;
-  fiat_p384_addcarryx_u32(&x422, &x423, x421, 0x0, x396);
+  uint32_t x423;
+  fiat_p384_mulx_u32(&x422, &x423, x388, UINT32_C(0xffffffff));
   uint32_t x424;
-  fiat_p384_uint1 x425;
-  fiat_p384_addcarryx_u32(&x424, &x425, x423, 0x0, x398);
+  uint32_t x425;
+  fiat_p384_mulx_u32(&x424, &x425, x388, UINT32_C(0xffffffff));
   uint32_t x426;
-  fiat_p384_uint1 x427;
-  fiat_p384_addcarryx_u32(&x426, &x427, x425, 0x0, x400);
+  uint32_t x427;
+  fiat_p384_mulx_u32(&x426, &x427, x388, UINT32_C(0xfffffffe));
   uint32_t x428;
-  fiat_p384_uint1 x429;
-  fiat_p384_addcarryx_u32(&x428, &x429, x427, 0x0, x402);
+  uint32_t x429;
+  fiat_p384_mulx_u32(&x428, &x429, x388, UINT32_C(0xffffffff));
   uint32_t x430;
-  fiat_p384_uint1 x431;
-  fiat_p384_addcarryx_u32(&x430, &x431, x429, 0x0, x404);
+  uint32_t x431;
+  fiat_p384_mulx_u32(&x430, &x431, x388, UINT32_C(0xffffffff));
   uint32_t x432;
   fiat_p384_uint1 x433;
-  fiat_p384_addcarryx_u32(&x432, &x433, x431, 0x0, x406);
+  fiat_p384_addcarryx_u32(&x432, &x433, 0x0, x426, x429);
   uint32_t x434;
   fiat_p384_uint1 x435;
-  fiat_p384_addcarryx_u32(&x434, &x435, x433, 0x0, x410);
+  fiat_p384_addcarryx_u32(&x434, &x435, x433, x424, x427);
   uint32_t x436;
   fiat_p384_uint1 x437;
-  fiat_p384_addcarryx_u32(&x436, &x437, x411, 0x0, x347);
+  fiat_p384_addcarryx_u32(&x436, &x437, x435, x422, x425);
   uint32_t x438;
   fiat_p384_uint1 x439;
-  fiat_p384_addcarryx_u32(&x438, &x439, x435, 0x0, (fiat_p384_uint1)x436);
+  fiat_p384_addcarryx_u32(&x438, &x439, x437, x420, x423);
   uint32_t x440;
-  uint32_t x441;
-  fiat_p384_mulx_u32(&x440, &x441, x412, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x441;
+  fiat_p384_addcarryx_u32(&x440, &x441, x439, x418, x421);
   uint32_t x442;
-  uint32_t x443;
-  fiat_p384_mulx_u32(&x442, &x443, x412, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x443;
+  fiat_p384_addcarryx_u32(&x442, &x443, x441, x416, x419);
   uint32_t x444;
-  uint32_t x445;
-  fiat_p384_mulx_u32(&x444, &x445, x412, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x445;
+  fiat_p384_addcarryx_u32(&x444, &x445, x443, x414, x417);
   uint32_t x446;
-  uint32_t x447;
-  fiat_p384_mulx_u32(&x446, &x447, x412, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x447;
+  fiat_p384_addcarryx_u32(&x446, &x447, x445, x412, x415);
   uint32_t x448;
-  uint32_t x449;
-  fiat_p384_mulx_u32(&x448, &x449, x412, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x449;
+  fiat_p384_addcarryx_u32(&x448, &x449, 0x0, x430, x388);
   uint32_t x450;
-  uint32_t x451;
-  fiat_p384_mulx_u32(&x450, &x451, x412, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x451;
+  fiat_p384_addcarryx_u32(&x450, &x451, x449, x431, x390);
   uint32_t x452;
-  uint32_t x453;
-  fiat_p384_mulx_u32(&x452, &x453, x412, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x453;
+  fiat_p384_addcarryx_u32(&x452, &x453, x451, 0x0, x392);
   uint32_t x454;
-  uint32_t x455;
-  fiat_p384_mulx_u32(&x454, &x455, x412, UINT32_C(0xfffffffe));
+  fiat_p384_uint1 x455;
+  fiat_p384_addcarryx_u32(&x454, &x455, x453, x428, x394);
   uint32_t x456;
-  uint32_t x457;
-  fiat_p384_mulx_u32(&x456, &x457, x412, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x457;
+  fiat_p384_addcarryx_u32(&x456, &x457, x455, x432, x396);
   uint32_t x458;
-  uint32_t x459;
-  fiat_p384_mulx_u32(&x458, &x459, x412, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x459;
+  fiat_p384_addcarryx_u32(&x458, &x459, x457, x434, x398);
   uint32_t x460;
   fiat_p384_uint1 x461;
-  fiat_p384_addcarryx_u32(&x460, &x461, 0x0, x454, x457);
+  fiat_p384_addcarryx_u32(&x460, &x461, x459, x436, x400);
   uint32_t x462;
   fiat_p384_uint1 x463;
-  fiat_p384_addcarryx_u32(&x462, &x463, x461, x452, x455);
+  fiat_p384_addcarryx_u32(&x462, &x463, x461, x438, x402);
   uint32_t x464;
   fiat_p384_uint1 x465;
-  fiat_p384_addcarryx_u32(&x464, &x465, x463, x450, x453);
+  fiat_p384_addcarryx_u32(&x464, &x465, x463, x440, x404);
   uint32_t x466;
   fiat_p384_uint1 x467;
-  fiat_p384_addcarryx_u32(&x466, &x467, x465, x448, x451);
+  fiat_p384_addcarryx_u32(&x466, &x467, x465, x442, x406);
   uint32_t x468;
   fiat_p384_uint1 x469;
-  fiat_p384_addcarryx_u32(&x468, &x469, x467, x446, x449);
+  fiat_p384_addcarryx_u32(&x468, &x469, x467, x444, x408);
   uint32_t x470;
   fiat_p384_uint1 x471;
-  fiat_p384_addcarryx_u32(&x470, &x471, x469, x444, x447);
+  fiat_p384_addcarryx_u32(&x470, &x471, x469, x446, x410);
   uint32_t x472;
   fiat_p384_uint1 x473;
-  fiat_p384_addcarryx_u32(&x472, &x473, x471, x442, x445);
+  fiat_p384_addcarryx_u32(&x472, &x473, x387, 0x0, 0x0);
   uint32_t x474;
   fiat_p384_uint1 x475;
-  fiat_p384_addcarryx_u32(&x474, &x475, x473, x440, x443);
+  fiat_p384_addcarryx_u32(&x474, &x475, x411, 0x0, (fiat_p384_uint1)x472);
   uint32_t x476;
   fiat_p384_uint1 x477;
-  fiat_p384_addcarryx_u32(&x476, &x477, 0x0, x458, x412);
+  fiat_p384_addcarryx_u32(&x476, &x477, x447, 0x0, x413);
   uint32_t x478;
   fiat_p384_uint1 x479;
-  fiat_p384_addcarryx_u32(&x478, &x479, x477, x459, x414);
+  fiat_p384_addcarryx_u32(&x478, &x479, x471, x476, x474);
   uint32_t x480;
   fiat_p384_uint1 x481;
-  fiat_p384_addcarryx_u32(&x480, &x481, x479, 0x0, x416);
+  fiat_p384_addcarryx_u32(&x480, &x481, 0x0, (arg1[6]), x450);
   uint32_t x482;
   fiat_p384_uint1 x483;
-  fiat_p384_addcarryx_u32(&x482, &x483, x481, x456, x418);
+  fiat_p384_addcarryx_u32(&x482, &x483, x481, 0x0, x452);
   uint32_t x484;
   fiat_p384_uint1 x485;
-  fiat_p384_addcarryx_u32(&x484, &x485, x483, x460, x420);
+  fiat_p384_addcarryx_u32(&x484, &x485, x483, 0x0, x454);
   uint32_t x486;
   fiat_p384_uint1 x487;
-  fiat_p384_addcarryx_u32(&x486, &x487, x485, x462, x422);
+  fiat_p384_addcarryx_u32(&x486, &x487, x485, 0x0, x456);
   uint32_t x488;
   fiat_p384_uint1 x489;
-  fiat_p384_addcarryx_u32(&x488, &x489, x487, x464, x424);
+  fiat_p384_addcarryx_u32(&x488, &x489, x487, 0x0, x458);
   uint32_t x490;
   fiat_p384_uint1 x491;
-  fiat_p384_addcarryx_u32(&x490, &x491, x489, x466, x426);
+  fiat_p384_addcarryx_u32(&x490, &x491, x489, 0x0, x460);
   uint32_t x492;
   fiat_p384_uint1 x493;
-  fiat_p384_addcarryx_u32(&x492, &x493, x491, x468, x428);
+  fiat_p384_addcarryx_u32(&x492, &x493, x491, 0x0, x462);
   uint32_t x494;
   fiat_p384_uint1 x495;
-  fiat_p384_addcarryx_u32(&x494, &x495, x493, x470, x430);
+  fiat_p384_addcarryx_u32(&x494, &x495, x493, 0x0, x464);
   uint32_t x496;
   fiat_p384_uint1 x497;
-  fiat_p384_addcarryx_u32(&x496, &x497, x495, x472, x432);
+  fiat_p384_addcarryx_u32(&x496, &x497, x495, 0x0, x466);
   uint32_t x498;
   fiat_p384_uint1 x499;
-  fiat_p384_addcarryx_u32(&x498, &x499, x497, x474, x434);
+  fiat_p384_addcarryx_u32(&x498, &x499, x497, 0x0, x468);
   uint32_t x500;
   fiat_p384_uint1 x501;
-  fiat_p384_addcarryx_u32(&x500, &x501, x475, 0x0, x441);
+  fiat_p384_addcarryx_u32(&x500, &x501, x499, 0x0, x470);
   uint32_t x502;
   fiat_p384_uint1 x503;
-  fiat_p384_addcarryx_u32(&x502, &x503, x499, x500, x438);
+  fiat_p384_addcarryx_u32(&x502, &x503, x501, 0x0, x478);
   uint32_t x504;
-  fiat_p384_uint1 x505;
-  fiat_p384_addcarryx_u32(&x504, &x505, 0x0, (arg1[6]), x478);
+  uint32_t x505;
+  fiat_p384_mulx_u32(&x504, &x505, x480, UINT32_C(0xffffffff));
   uint32_t x506;
-  fiat_p384_uint1 x507;
-  fiat_p384_addcarryx_u32(&x506, &x507, x505, 0x0, x480);
+  uint32_t x507;
+  fiat_p384_mulx_u32(&x506, &x507, x480, UINT32_C(0xffffffff));
   uint32_t x508;
-  fiat_p384_uint1 x509;
-  fiat_p384_addcarryx_u32(&x508, &x509, x507, 0x0, x482);
+  uint32_t x509;
+  fiat_p384_mulx_u32(&x508, &x509, x480, UINT32_C(0xffffffff));
   uint32_t x510;
-  fiat_p384_uint1 x511;
-  fiat_p384_addcarryx_u32(&x510, &x511, x509, 0x0, x484);
+  uint32_t x511;
+  fiat_p384_mulx_u32(&x510, &x511, x480, UINT32_C(0xffffffff));
   uint32_t x512;
-  fiat_p384_uint1 x513;
-  fiat_p384_addcarryx_u32(&x512, &x513, x511, 0x0, x486);
+  uint32_t x513;
+  fiat_p384_mulx_u32(&x512, &x513, x480, UINT32_C(0xffffffff));
   uint32_t x514;
-  fiat_p384_uint1 x515;
-  fiat_p384_addcarryx_u32(&x514, &x515, x513, 0x0, x488);
+  uint32_t x515;
+  fiat_p384_mulx_u32(&x514, &x515, x480, UINT32_C(0xffffffff));
   uint32_t x516;
-  fiat_p384_uint1 x517;
-  fiat_p384_addcarryx_u32(&x516, &x517, x515, 0x0, x490);
+  uint32_t x517;
+  fiat_p384_mulx_u32(&x516, &x517, x480, UINT32_C(0xffffffff));
   uint32_t x518;
-  fiat_p384_uint1 x519;
-  fiat_p384_addcarryx_u32(&x518, &x519, x517, 0x0, x492);
+  uint32_t x519;
+  fiat_p384_mulx_u32(&x518, &x519, x480, UINT32_C(0xfffffffe));
   uint32_t x520;
-  fiat_p384_uint1 x521;
-  fiat_p384_addcarryx_u32(&x520, &x521, x519, 0x0, x494);
+  uint32_t x521;
+  fiat_p384_mulx_u32(&x520, &x521, x480, UINT32_C(0xffffffff));
   uint32_t x522;
-  fiat_p384_uint1 x523;
-  fiat_p384_addcarryx_u32(&x522, &x523, x521, 0x0, x496);
+  uint32_t x523;
+  fiat_p384_mulx_u32(&x522, &x523, x480, UINT32_C(0xffffffff));
   uint32_t x524;
   fiat_p384_uint1 x525;
-  fiat_p384_addcarryx_u32(&x524, &x525, x523, 0x0, x498);
+  fiat_p384_addcarryx_u32(&x524, &x525, 0x0, x518, x521);
   uint32_t x526;
   fiat_p384_uint1 x527;
-  fiat_p384_addcarryx_u32(&x526, &x527, x525, 0x0, x502);
+  fiat_p384_addcarryx_u32(&x526, &x527, x525, x516, x519);
   uint32_t x528;
   fiat_p384_uint1 x529;
-  fiat_p384_addcarryx_u32(&x528, &x529, x503, 0x0, x439);
+  fiat_p384_addcarryx_u32(&x528, &x529, x527, x514, x517);
   uint32_t x530;
   fiat_p384_uint1 x531;
-  fiat_p384_addcarryx_u32(&x530, &x531, x527, 0x0, (fiat_p384_uint1)x528);
+  fiat_p384_addcarryx_u32(&x530, &x531, x529, x512, x515);
   uint32_t x532;
-  uint32_t x533;
-  fiat_p384_mulx_u32(&x532, &x533, x504, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x533;
+  fiat_p384_addcarryx_u32(&x532, &x533, x531, x510, x513);
   uint32_t x534;
-  uint32_t x535;
-  fiat_p384_mulx_u32(&x534, &x535, x504, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x535;
+  fiat_p384_addcarryx_u32(&x534, &x535, x533, x508, x511);
   uint32_t x536;
-  uint32_t x537;
-  fiat_p384_mulx_u32(&x536, &x537, x504, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x537;
+  fiat_p384_addcarryx_u32(&x536, &x537, x535, x506, x509);
   uint32_t x538;
-  uint32_t x539;
-  fiat_p384_mulx_u32(&x538, &x539, x504, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x539;
+  fiat_p384_addcarryx_u32(&x538, &x539, x537, x504, x507);
   uint32_t x540;
-  uint32_t x541;
-  fiat_p384_mulx_u32(&x540, &x541, x504, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x541;
+  fiat_p384_addcarryx_u32(&x540, &x541, 0x0, x522, x480);
   uint32_t x542;
-  uint32_t x543;
-  fiat_p384_mulx_u32(&x542, &x543, x504, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x543;
+  fiat_p384_addcarryx_u32(&x542, &x543, x541, x523, x482);
   uint32_t x544;
-  uint32_t x545;
-  fiat_p384_mulx_u32(&x544, &x545, x504, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x545;
+  fiat_p384_addcarryx_u32(&x544, &x545, x543, 0x0, x484);
   uint32_t x546;
-  uint32_t x547;
-  fiat_p384_mulx_u32(&x546, &x547, x504, UINT32_C(0xfffffffe));
+  fiat_p384_uint1 x547;
+  fiat_p384_addcarryx_u32(&x546, &x547, x545, x520, x486);
   uint32_t x548;
-  uint32_t x549;
-  fiat_p384_mulx_u32(&x548, &x549, x504, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x549;
+  fiat_p384_addcarryx_u32(&x548, &x549, x547, x524, x488);
   uint32_t x550;
-  uint32_t x551;
-  fiat_p384_mulx_u32(&x550, &x551, x504, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x551;
+  fiat_p384_addcarryx_u32(&x550, &x551, x549, x526, x490);
   uint32_t x552;
   fiat_p384_uint1 x553;
-  fiat_p384_addcarryx_u32(&x552, &x553, 0x0, x546, x549);
+  fiat_p384_addcarryx_u32(&x552, &x553, x551, x528, x492);
   uint32_t x554;
   fiat_p384_uint1 x555;
-  fiat_p384_addcarryx_u32(&x554, &x555, x553, x544, x547);
+  fiat_p384_addcarryx_u32(&x554, &x555, x553, x530, x494);
   uint32_t x556;
   fiat_p384_uint1 x557;
-  fiat_p384_addcarryx_u32(&x556, &x557, x555, x542, x545);
+  fiat_p384_addcarryx_u32(&x556, &x557, x555, x532, x496);
   uint32_t x558;
   fiat_p384_uint1 x559;
-  fiat_p384_addcarryx_u32(&x558, &x559, x557, x540, x543);
+  fiat_p384_addcarryx_u32(&x558, &x559, x557, x534, x498);
   uint32_t x560;
   fiat_p384_uint1 x561;
-  fiat_p384_addcarryx_u32(&x560, &x561, x559, x538, x541);
+  fiat_p384_addcarryx_u32(&x560, &x561, x559, x536, x500);
   uint32_t x562;
   fiat_p384_uint1 x563;
-  fiat_p384_addcarryx_u32(&x562, &x563, x561, x536, x539);
+  fiat_p384_addcarryx_u32(&x562, &x563, x561, x538, x502);
   uint32_t x564;
   fiat_p384_uint1 x565;
-  fiat_p384_addcarryx_u32(&x564, &x565, x563, x534, x537);
+  fiat_p384_addcarryx_u32(&x564, &x565, x479, 0x0, 0x0);
   uint32_t x566;
   fiat_p384_uint1 x567;
-  fiat_p384_addcarryx_u32(&x566, &x567, x565, x532, x535);
+  fiat_p384_addcarryx_u32(&x566, &x567, x503, 0x0, (fiat_p384_uint1)x564);
   uint32_t x568;
   fiat_p384_uint1 x569;
-  fiat_p384_addcarryx_u32(&x568, &x569, 0x0, x550, x504);
+  fiat_p384_addcarryx_u32(&x568, &x569, x539, 0x0, x505);
   uint32_t x570;
   fiat_p384_uint1 x571;
-  fiat_p384_addcarryx_u32(&x570, &x571, x569, x551, x506);
+  fiat_p384_addcarryx_u32(&x570, &x571, x563, x568, x566);
   uint32_t x572;
   fiat_p384_uint1 x573;
-  fiat_p384_addcarryx_u32(&x572, &x573, x571, 0x0, x508);
+  fiat_p384_addcarryx_u32(&x572, &x573, 0x0, (arg1[7]), x542);
   uint32_t x574;
   fiat_p384_uint1 x575;
-  fiat_p384_addcarryx_u32(&x574, &x575, x573, x548, x510);
+  fiat_p384_addcarryx_u32(&x574, &x575, x573, 0x0, x544);
   uint32_t x576;
   fiat_p384_uint1 x577;
-  fiat_p384_addcarryx_u32(&x576, &x577, x575, x552, x512);
+  fiat_p384_addcarryx_u32(&x576, &x577, x575, 0x0, x546);
   uint32_t x578;
   fiat_p384_uint1 x579;
-  fiat_p384_addcarryx_u32(&x578, &x579, x577, x554, x514);
+  fiat_p384_addcarryx_u32(&x578, &x579, x577, 0x0, x548);
   uint32_t x580;
   fiat_p384_uint1 x581;
-  fiat_p384_addcarryx_u32(&x580, &x581, x579, x556, x516);
+  fiat_p384_addcarryx_u32(&x580, &x581, x579, 0x0, x550);
   uint32_t x582;
   fiat_p384_uint1 x583;
-  fiat_p384_addcarryx_u32(&x582, &x583, x581, x558, x518);
+  fiat_p384_addcarryx_u32(&x582, &x583, x581, 0x0, x552);
   uint32_t x584;
   fiat_p384_uint1 x585;
-  fiat_p384_addcarryx_u32(&x584, &x585, x583, x560, x520);
+  fiat_p384_addcarryx_u32(&x584, &x585, x583, 0x0, x554);
   uint32_t x586;
   fiat_p384_uint1 x587;
-  fiat_p384_addcarryx_u32(&x586, &x587, x585, x562, x522);
+  fiat_p384_addcarryx_u32(&x586, &x587, x585, 0x0, x556);
   uint32_t x588;
   fiat_p384_uint1 x589;
-  fiat_p384_addcarryx_u32(&x588, &x589, x587, x564, x524);
+  fiat_p384_addcarryx_u32(&x588, &x589, x587, 0x0, x558);
   uint32_t x590;
   fiat_p384_uint1 x591;
-  fiat_p384_addcarryx_u32(&x590, &x591, x589, x566, x526);
+  fiat_p384_addcarryx_u32(&x590, &x591, x589, 0x0, x560);
   uint32_t x592;
   fiat_p384_uint1 x593;
-  fiat_p384_addcarryx_u32(&x592, &x593, x567, 0x0, x533);
+  fiat_p384_addcarryx_u32(&x592, &x593, x591, 0x0, x562);
   uint32_t x594;
   fiat_p384_uint1 x595;
-  fiat_p384_addcarryx_u32(&x594, &x595, x591, x592, x530);
+  fiat_p384_addcarryx_u32(&x594, &x595, x593, 0x0, x570);
   uint32_t x596;
-  fiat_p384_uint1 x597;
-  fiat_p384_addcarryx_u32(&x596, &x597, 0x0, (arg1[7]), x570);
+  uint32_t x597;
+  fiat_p384_mulx_u32(&x596, &x597, x572, UINT32_C(0xffffffff));
   uint32_t x598;
-  fiat_p384_uint1 x599;
-  fiat_p384_addcarryx_u32(&x598, &x599, x597, 0x0, x572);
+  uint32_t x599;
+  fiat_p384_mulx_u32(&x598, &x599, x572, UINT32_C(0xffffffff));
   uint32_t x600;
-  fiat_p384_uint1 x601;
-  fiat_p384_addcarryx_u32(&x600, &x601, x599, 0x0, x574);
+  uint32_t x601;
+  fiat_p384_mulx_u32(&x600, &x601, x572, UINT32_C(0xffffffff));
   uint32_t x602;
-  fiat_p384_uint1 x603;
-  fiat_p384_addcarryx_u32(&x602, &x603, x601, 0x0, x576);
+  uint32_t x603;
+  fiat_p384_mulx_u32(&x602, &x603, x572, UINT32_C(0xffffffff));
   uint32_t x604;
-  fiat_p384_uint1 x605;
-  fiat_p384_addcarryx_u32(&x604, &x605, x603, 0x0, x578);
+  uint32_t x605;
+  fiat_p384_mulx_u32(&x604, &x605, x572, UINT32_C(0xffffffff));
   uint32_t x606;
-  fiat_p384_uint1 x607;
-  fiat_p384_addcarryx_u32(&x606, &x607, x605, 0x0, x580);
+  uint32_t x607;
+  fiat_p384_mulx_u32(&x606, &x607, x572, UINT32_C(0xffffffff));
   uint32_t x608;
-  fiat_p384_uint1 x609;
-  fiat_p384_addcarryx_u32(&x608, &x609, x607, 0x0, x582);
+  uint32_t x609;
+  fiat_p384_mulx_u32(&x608, &x609, x572, UINT32_C(0xffffffff));
   uint32_t x610;
-  fiat_p384_uint1 x611;
-  fiat_p384_addcarryx_u32(&x610, &x611, x609, 0x0, x584);
+  uint32_t x611;
+  fiat_p384_mulx_u32(&x610, &x611, x572, UINT32_C(0xfffffffe));
   uint32_t x612;
-  fiat_p384_uint1 x613;
-  fiat_p384_addcarryx_u32(&x612, &x613, x611, 0x0, x586);
+  uint32_t x613;
+  fiat_p384_mulx_u32(&x612, &x613, x572, UINT32_C(0xffffffff));
   uint32_t x614;
-  fiat_p384_uint1 x615;
-  fiat_p384_addcarryx_u32(&x614, &x615, x613, 0x0, x588);
+  uint32_t x615;
+  fiat_p384_mulx_u32(&x614, &x615, x572, UINT32_C(0xffffffff));
   uint32_t x616;
   fiat_p384_uint1 x617;
-  fiat_p384_addcarryx_u32(&x616, &x617, x615, 0x0, x590);
+  fiat_p384_addcarryx_u32(&x616, &x617, 0x0, x610, x613);
   uint32_t x618;
   fiat_p384_uint1 x619;
-  fiat_p384_addcarryx_u32(&x618, &x619, x617, 0x0, x594);
+  fiat_p384_addcarryx_u32(&x618, &x619, x617, x608, x611);
   uint32_t x620;
   fiat_p384_uint1 x621;
-  fiat_p384_addcarryx_u32(&x620, &x621, x595, 0x0, x531);
+  fiat_p384_addcarryx_u32(&x620, &x621, x619, x606, x609);
   uint32_t x622;
   fiat_p384_uint1 x623;
-  fiat_p384_addcarryx_u32(&x622, &x623, x619, 0x0, (fiat_p384_uint1)x620);
+  fiat_p384_addcarryx_u32(&x622, &x623, x621, x604, x607);
   uint32_t x624;
-  uint32_t x625;
-  fiat_p384_mulx_u32(&x624, &x625, x596, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x625;
+  fiat_p384_addcarryx_u32(&x624, &x625, x623, x602, x605);
   uint32_t x626;
-  uint32_t x627;
-  fiat_p384_mulx_u32(&x626, &x627, x596, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x627;
+  fiat_p384_addcarryx_u32(&x626, &x627, x625, x600, x603);
   uint32_t x628;
-  uint32_t x629;
-  fiat_p384_mulx_u32(&x628, &x629, x596, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x629;
+  fiat_p384_addcarryx_u32(&x628, &x629, x627, x598, x601);
   uint32_t x630;
-  uint32_t x631;
-  fiat_p384_mulx_u32(&x630, &x631, x596, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x631;
+  fiat_p384_addcarryx_u32(&x630, &x631, x629, x596, x599);
   uint32_t x632;
-  uint32_t x633;
-  fiat_p384_mulx_u32(&x632, &x633, x596, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x633;
+  fiat_p384_addcarryx_u32(&x632, &x633, 0x0, x614, x572);
   uint32_t x634;
-  uint32_t x635;
-  fiat_p384_mulx_u32(&x634, &x635, x596, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x635;
+  fiat_p384_addcarryx_u32(&x634, &x635, x633, x615, x574);
   uint32_t x636;
-  uint32_t x637;
-  fiat_p384_mulx_u32(&x636, &x637, x596, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x637;
+  fiat_p384_addcarryx_u32(&x636, &x637, x635, 0x0, x576);
   uint32_t x638;
-  uint32_t x639;
-  fiat_p384_mulx_u32(&x638, &x639, x596, UINT32_C(0xfffffffe));
+  fiat_p384_uint1 x639;
+  fiat_p384_addcarryx_u32(&x638, &x639, x637, x612, x578);
   uint32_t x640;
-  uint32_t x641;
-  fiat_p384_mulx_u32(&x640, &x641, x596, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x641;
+  fiat_p384_addcarryx_u32(&x640, &x641, x639, x616, x580);
   uint32_t x642;
-  uint32_t x643;
-  fiat_p384_mulx_u32(&x642, &x643, x596, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x643;
+  fiat_p384_addcarryx_u32(&x642, &x643, x641, x618, x582);
   uint32_t x644;
   fiat_p384_uint1 x645;
-  fiat_p384_addcarryx_u32(&x644, &x645, 0x0, x638, x641);
+  fiat_p384_addcarryx_u32(&x644, &x645, x643, x620, x584);
   uint32_t x646;
   fiat_p384_uint1 x647;
-  fiat_p384_addcarryx_u32(&x646, &x647, x645, x636, x639);
+  fiat_p384_addcarryx_u32(&x646, &x647, x645, x622, x586);
   uint32_t x648;
   fiat_p384_uint1 x649;
-  fiat_p384_addcarryx_u32(&x648, &x649, x647, x634, x637);
+  fiat_p384_addcarryx_u32(&x648, &x649, x647, x624, x588);
   uint32_t x650;
   fiat_p384_uint1 x651;
-  fiat_p384_addcarryx_u32(&x650, &x651, x649, x632, x635);
+  fiat_p384_addcarryx_u32(&x650, &x651, x649, x626, x590);
   uint32_t x652;
   fiat_p384_uint1 x653;
-  fiat_p384_addcarryx_u32(&x652, &x653, x651, x630, x633);
+  fiat_p384_addcarryx_u32(&x652, &x653, x651, x628, x592);
   uint32_t x654;
   fiat_p384_uint1 x655;
-  fiat_p384_addcarryx_u32(&x654, &x655, x653, x628, x631);
+  fiat_p384_addcarryx_u32(&x654, &x655, x653, x630, x594);
   uint32_t x656;
   fiat_p384_uint1 x657;
-  fiat_p384_addcarryx_u32(&x656, &x657, x655, x626, x629);
+  fiat_p384_addcarryx_u32(&x656, &x657, x571, 0x0, 0x0);
   uint32_t x658;
   fiat_p384_uint1 x659;
-  fiat_p384_addcarryx_u32(&x658, &x659, x657, x624, x627);
+  fiat_p384_addcarryx_u32(&x658, &x659, x595, 0x0, (fiat_p384_uint1)x656);
   uint32_t x660;
   fiat_p384_uint1 x661;
-  fiat_p384_addcarryx_u32(&x660, &x661, 0x0, x642, x596);
+  fiat_p384_addcarryx_u32(&x660, &x661, x631, 0x0, x597);
   uint32_t x662;
   fiat_p384_uint1 x663;
-  fiat_p384_addcarryx_u32(&x662, &x663, x661, x643, x598);
+  fiat_p384_addcarryx_u32(&x662, &x663, x655, x660, x658);
   uint32_t x664;
   fiat_p384_uint1 x665;
-  fiat_p384_addcarryx_u32(&x664, &x665, x663, 0x0, x600);
+  fiat_p384_addcarryx_u32(&x664, &x665, 0x0, (arg1[8]), x634);
   uint32_t x666;
   fiat_p384_uint1 x667;
-  fiat_p384_addcarryx_u32(&x666, &x667, x665, x640, x602);
+  fiat_p384_addcarryx_u32(&x666, &x667, x665, 0x0, x636);
   uint32_t x668;
   fiat_p384_uint1 x669;
-  fiat_p384_addcarryx_u32(&x668, &x669, x667, x644, x604);
+  fiat_p384_addcarryx_u32(&x668, &x669, x667, 0x0, x638);
   uint32_t x670;
   fiat_p384_uint1 x671;
-  fiat_p384_addcarryx_u32(&x670, &x671, x669, x646, x606);
+  fiat_p384_addcarryx_u32(&x670, &x671, x669, 0x0, x640);
   uint32_t x672;
   fiat_p384_uint1 x673;
-  fiat_p384_addcarryx_u32(&x672, &x673, x671, x648, x608);
+  fiat_p384_addcarryx_u32(&x672, &x673, x671, 0x0, x642);
   uint32_t x674;
   fiat_p384_uint1 x675;
-  fiat_p384_addcarryx_u32(&x674, &x675, x673, x650, x610);
+  fiat_p384_addcarryx_u32(&x674, &x675, x673, 0x0, x644);
   uint32_t x676;
   fiat_p384_uint1 x677;
-  fiat_p384_addcarryx_u32(&x676, &x677, x675, x652, x612);
+  fiat_p384_addcarryx_u32(&x676, &x677, x675, 0x0, x646);
   uint32_t x678;
   fiat_p384_uint1 x679;
-  fiat_p384_addcarryx_u32(&x678, &x679, x677, x654, x614);
+  fiat_p384_addcarryx_u32(&x678, &x679, x677, 0x0, x648);
   uint32_t x680;
   fiat_p384_uint1 x681;
-  fiat_p384_addcarryx_u32(&x680, &x681, x679, x656, x616);
+  fiat_p384_addcarryx_u32(&x680, &x681, x679, 0x0, x650);
   uint32_t x682;
   fiat_p384_uint1 x683;
-  fiat_p384_addcarryx_u32(&x682, &x683, x681, x658, x618);
+  fiat_p384_addcarryx_u32(&x682, &x683, x681, 0x0, x652);
   uint32_t x684;
   fiat_p384_uint1 x685;
-  fiat_p384_addcarryx_u32(&x684, &x685, x659, 0x0, x625);
+  fiat_p384_addcarryx_u32(&x684, &x685, x683, 0x0, x654);
   uint32_t x686;
   fiat_p384_uint1 x687;
-  fiat_p384_addcarryx_u32(&x686, &x687, x683, x684, x622);
+  fiat_p384_addcarryx_u32(&x686, &x687, x685, 0x0, x662);
   uint32_t x688;
-  fiat_p384_uint1 x689;
-  fiat_p384_addcarryx_u32(&x688, &x689, 0x0, (arg1[8]), x662);
+  uint32_t x689;
+  fiat_p384_mulx_u32(&x688, &x689, x664, UINT32_C(0xffffffff));
   uint32_t x690;
-  fiat_p384_uint1 x691;
-  fiat_p384_addcarryx_u32(&x690, &x691, x689, 0x0, x664);
+  uint32_t x691;
+  fiat_p384_mulx_u32(&x690, &x691, x664, UINT32_C(0xffffffff));
   uint32_t x692;
-  fiat_p384_uint1 x693;
-  fiat_p384_addcarryx_u32(&x692, &x693, x691, 0x0, x666);
+  uint32_t x693;
+  fiat_p384_mulx_u32(&x692, &x693, x664, UINT32_C(0xffffffff));
   uint32_t x694;
-  fiat_p384_uint1 x695;
-  fiat_p384_addcarryx_u32(&x694, &x695, x693, 0x0, x668);
+  uint32_t x695;
+  fiat_p384_mulx_u32(&x694, &x695, x664, UINT32_C(0xffffffff));
   uint32_t x696;
-  fiat_p384_uint1 x697;
-  fiat_p384_addcarryx_u32(&x696, &x697, x695, 0x0, x670);
+  uint32_t x697;
+  fiat_p384_mulx_u32(&x696, &x697, x664, UINT32_C(0xffffffff));
   uint32_t x698;
-  fiat_p384_uint1 x699;
-  fiat_p384_addcarryx_u32(&x698, &x699, x697, 0x0, x672);
+  uint32_t x699;
+  fiat_p384_mulx_u32(&x698, &x699, x664, UINT32_C(0xffffffff));
   uint32_t x700;
-  fiat_p384_uint1 x701;
-  fiat_p384_addcarryx_u32(&x700, &x701, x699, 0x0, x674);
+  uint32_t x701;
+  fiat_p384_mulx_u32(&x700, &x701, x664, UINT32_C(0xffffffff));
   uint32_t x702;
-  fiat_p384_uint1 x703;
-  fiat_p384_addcarryx_u32(&x702, &x703, x701, 0x0, x676);
+  uint32_t x703;
+  fiat_p384_mulx_u32(&x702, &x703, x664, UINT32_C(0xfffffffe));
   uint32_t x704;
-  fiat_p384_uint1 x705;
-  fiat_p384_addcarryx_u32(&x704, &x705, x703, 0x0, x678);
+  uint32_t x705;
+  fiat_p384_mulx_u32(&x704, &x705, x664, UINT32_C(0xffffffff));
   uint32_t x706;
-  fiat_p384_uint1 x707;
-  fiat_p384_addcarryx_u32(&x706, &x707, x705, 0x0, x680);
+  uint32_t x707;
+  fiat_p384_mulx_u32(&x706, &x707, x664, UINT32_C(0xffffffff));
   uint32_t x708;
   fiat_p384_uint1 x709;
-  fiat_p384_addcarryx_u32(&x708, &x709, x707, 0x0, x682);
+  fiat_p384_addcarryx_u32(&x708, &x709, 0x0, x702, x705);
   uint32_t x710;
   fiat_p384_uint1 x711;
-  fiat_p384_addcarryx_u32(&x710, &x711, x709, 0x0, x686);
+  fiat_p384_addcarryx_u32(&x710, &x711, x709, x700, x703);
   uint32_t x712;
   fiat_p384_uint1 x713;
-  fiat_p384_addcarryx_u32(&x712, &x713, x687, 0x0, x623);
+  fiat_p384_addcarryx_u32(&x712, &x713, x711, x698, x701);
   uint32_t x714;
   fiat_p384_uint1 x715;
-  fiat_p384_addcarryx_u32(&x714, &x715, x711, 0x0, (fiat_p384_uint1)x712);
+  fiat_p384_addcarryx_u32(&x714, &x715, x713, x696, x699);
   uint32_t x716;
-  uint32_t x717;
-  fiat_p384_mulx_u32(&x716, &x717, x688, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x717;
+  fiat_p384_addcarryx_u32(&x716, &x717, x715, x694, x697);
   uint32_t x718;
-  uint32_t x719;
-  fiat_p384_mulx_u32(&x718, &x719, x688, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x719;
+  fiat_p384_addcarryx_u32(&x718, &x719, x717, x692, x695);
   uint32_t x720;
-  uint32_t x721;
-  fiat_p384_mulx_u32(&x720, &x721, x688, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x721;
+  fiat_p384_addcarryx_u32(&x720, &x721, x719, x690, x693);
   uint32_t x722;
-  uint32_t x723;
-  fiat_p384_mulx_u32(&x722, &x723, x688, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x723;
+  fiat_p384_addcarryx_u32(&x722, &x723, x721, x688, x691);
   uint32_t x724;
-  uint32_t x725;
-  fiat_p384_mulx_u32(&x724, &x725, x688, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x725;
+  fiat_p384_addcarryx_u32(&x724, &x725, 0x0, x706, x664);
   uint32_t x726;
-  uint32_t x727;
-  fiat_p384_mulx_u32(&x726, &x727, x688, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x727;
+  fiat_p384_addcarryx_u32(&x726, &x727, x725, x707, x666);
   uint32_t x728;
-  uint32_t x729;
-  fiat_p384_mulx_u32(&x728, &x729, x688, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x729;
+  fiat_p384_addcarryx_u32(&x728, &x729, x727, 0x0, x668);
   uint32_t x730;
-  uint32_t x731;
-  fiat_p384_mulx_u32(&x730, &x731, x688, UINT32_C(0xfffffffe));
+  fiat_p384_uint1 x731;
+  fiat_p384_addcarryx_u32(&x730, &x731, x729, x704, x670);
   uint32_t x732;
-  uint32_t x733;
-  fiat_p384_mulx_u32(&x732, &x733, x688, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x733;
+  fiat_p384_addcarryx_u32(&x732, &x733, x731, x708, x672);
   uint32_t x734;
-  uint32_t x735;
-  fiat_p384_mulx_u32(&x734, &x735, x688, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x735;
+  fiat_p384_addcarryx_u32(&x734, &x735, x733, x710, x674);
   uint32_t x736;
   fiat_p384_uint1 x737;
-  fiat_p384_addcarryx_u32(&x736, &x737, 0x0, x730, x733);
+  fiat_p384_addcarryx_u32(&x736, &x737, x735, x712, x676);
   uint32_t x738;
   fiat_p384_uint1 x739;
-  fiat_p384_addcarryx_u32(&x738, &x739, x737, x728, x731);
+  fiat_p384_addcarryx_u32(&x738, &x739, x737, x714, x678);
   uint32_t x740;
   fiat_p384_uint1 x741;
-  fiat_p384_addcarryx_u32(&x740, &x741, x739, x726, x729);
+  fiat_p384_addcarryx_u32(&x740, &x741, x739, x716, x680);
   uint32_t x742;
   fiat_p384_uint1 x743;
-  fiat_p384_addcarryx_u32(&x742, &x743, x741, x724, x727);
+  fiat_p384_addcarryx_u32(&x742, &x743, x741, x718, x682);
   uint32_t x744;
   fiat_p384_uint1 x745;
-  fiat_p384_addcarryx_u32(&x744, &x745, x743, x722, x725);
+  fiat_p384_addcarryx_u32(&x744, &x745, x743, x720, x684);
   uint32_t x746;
   fiat_p384_uint1 x747;
-  fiat_p384_addcarryx_u32(&x746, &x747, x745, x720, x723);
+  fiat_p384_addcarryx_u32(&x746, &x747, x745, x722, x686);
   uint32_t x748;
   fiat_p384_uint1 x749;
-  fiat_p384_addcarryx_u32(&x748, &x749, x747, x718, x721);
+  fiat_p384_addcarryx_u32(&x748, &x749, x663, 0x0, 0x0);
   uint32_t x750;
   fiat_p384_uint1 x751;
-  fiat_p384_addcarryx_u32(&x750, &x751, x749, x716, x719);
+  fiat_p384_addcarryx_u32(&x750, &x751, x687, 0x0, (fiat_p384_uint1)x748);
   uint32_t x752;
   fiat_p384_uint1 x753;
-  fiat_p384_addcarryx_u32(&x752, &x753, 0x0, x734, x688);
+  fiat_p384_addcarryx_u32(&x752, &x753, x723, 0x0, x689);
   uint32_t x754;
   fiat_p384_uint1 x755;
-  fiat_p384_addcarryx_u32(&x754, &x755, x753, x735, x690);
+  fiat_p384_addcarryx_u32(&x754, &x755, x747, x752, x750);
   uint32_t x756;
   fiat_p384_uint1 x757;
-  fiat_p384_addcarryx_u32(&x756, &x757, x755, 0x0, x692);
+  fiat_p384_addcarryx_u32(&x756, &x757, 0x0, (arg1[9]), x726);
   uint32_t x758;
   fiat_p384_uint1 x759;
-  fiat_p384_addcarryx_u32(&x758, &x759, x757, x732, x694);
+  fiat_p384_addcarryx_u32(&x758, &x759, x757, 0x0, x728);
   uint32_t x760;
   fiat_p384_uint1 x761;
-  fiat_p384_addcarryx_u32(&x760, &x761, x759, x736, x696);
+  fiat_p384_addcarryx_u32(&x760, &x761, x759, 0x0, x730);
   uint32_t x762;
   fiat_p384_uint1 x763;
-  fiat_p384_addcarryx_u32(&x762, &x763, x761, x738, x698);
+  fiat_p384_addcarryx_u32(&x762, &x763, x761, 0x0, x732);
   uint32_t x764;
   fiat_p384_uint1 x765;
-  fiat_p384_addcarryx_u32(&x764, &x765, x763, x740, x700);
+  fiat_p384_addcarryx_u32(&x764, &x765, x763, 0x0, x734);
   uint32_t x766;
   fiat_p384_uint1 x767;
-  fiat_p384_addcarryx_u32(&x766, &x767, x765, x742, x702);
+  fiat_p384_addcarryx_u32(&x766, &x767, x765, 0x0, x736);
   uint32_t x768;
   fiat_p384_uint1 x769;
-  fiat_p384_addcarryx_u32(&x768, &x769, x767, x744, x704);
+  fiat_p384_addcarryx_u32(&x768, &x769, x767, 0x0, x738);
   uint32_t x770;
   fiat_p384_uint1 x771;
-  fiat_p384_addcarryx_u32(&x770, &x771, x769, x746, x706);
+  fiat_p384_addcarryx_u32(&x770, &x771, x769, 0x0, x740);
   uint32_t x772;
   fiat_p384_uint1 x773;
-  fiat_p384_addcarryx_u32(&x772, &x773, x771, x748, x708);
+  fiat_p384_addcarryx_u32(&x772, &x773, x771, 0x0, x742);
   uint32_t x774;
   fiat_p384_uint1 x775;
-  fiat_p384_addcarryx_u32(&x774, &x775, x773, x750, x710);
+  fiat_p384_addcarryx_u32(&x774, &x775, x773, 0x0, x744);
   uint32_t x776;
   fiat_p384_uint1 x777;
-  fiat_p384_addcarryx_u32(&x776, &x777, x751, 0x0, x717);
+  fiat_p384_addcarryx_u32(&x776, &x777, x775, 0x0, x746);
   uint32_t x778;
   fiat_p384_uint1 x779;
-  fiat_p384_addcarryx_u32(&x778, &x779, x775, x776, x714);
+  fiat_p384_addcarryx_u32(&x778, &x779, x777, 0x0, x754);
   uint32_t x780;
-  fiat_p384_uint1 x781;
-  fiat_p384_addcarryx_u32(&x780, &x781, 0x0, (arg1[9]), x754);
+  uint32_t x781;
+  fiat_p384_mulx_u32(&x780, &x781, x756, UINT32_C(0xffffffff));
   uint32_t x782;
-  fiat_p384_uint1 x783;
-  fiat_p384_addcarryx_u32(&x782, &x783, x781, 0x0, x756);
+  uint32_t x783;
+  fiat_p384_mulx_u32(&x782, &x783, x756, UINT32_C(0xffffffff));
   uint32_t x784;
-  fiat_p384_uint1 x785;
-  fiat_p384_addcarryx_u32(&x784, &x785, x783, 0x0, x758);
+  uint32_t x785;
+  fiat_p384_mulx_u32(&x784, &x785, x756, UINT32_C(0xffffffff));
   uint32_t x786;
-  fiat_p384_uint1 x787;
-  fiat_p384_addcarryx_u32(&x786, &x787, x785, 0x0, x760);
+  uint32_t x787;
+  fiat_p384_mulx_u32(&x786, &x787, x756, UINT32_C(0xffffffff));
   uint32_t x788;
-  fiat_p384_uint1 x789;
-  fiat_p384_addcarryx_u32(&x788, &x789, x787, 0x0, x762);
+  uint32_t x789;
+  fiat_p384_mulx_u32(&x788, &x789, x756, UINT32_C(0xffffffff));
   uint32_t x790;
-  fiat_p384_uint1 x791;
-  fiat_p384_addcarryx_u32(&x790, &x791, x789, 0x0, x764);
+  uint32_t x791;
+  fiat_p384_mulx_u32(&x790, &x791, x756, UINT32_C(0xffffffff));
   uint32_t x792;
-  fiat_p384_uint1 x793;
-  fiat_p384_addcarryx_u32(&x792, &x793, x791, 0x0, x766);
+  uint32_t x793;
+  fiat_p384_mulx_u32(&x792, &x793, x756, UINT32_C(0xffffffff));
   uint32_t x794;
-  fiat_p384_uint1 x795;
-  fiat_p384_addcarryx_u32(&x794, &x795, x793, 0x0, x768);
+  uint32_t x795;
+  fiat_p384_mulx_u32(&x794, &x795, x756, UINT32_C(0xfffffffe));
   uint32_t x796;
-  fiat_p384_uint1 x797;
-  fiat_p384_addcarryx_u32(&x796, &x797, x795, 0x0, x770);
+  uint32_t x797;
+  fiat_p384_mulx_u32(&x796, &x797, x756, UINT32_C(0xffffffff));
   uint32_t x798;
-  fiat_p384_uint1 x799;
-  fiat_p384_addcarryx_u32(&x798, &x799, x797, 0x0, x772);
+  uint32_t x799;
+  fiat_p384_mulx_u32(&x798, &x799, x756, UINT32_C(0xffffffff));
   uint32_t x800;
   fiat_p384_uint1 x801;
-  fiat_p384_addcarryx_u32(&x800, &x801, x799, 0x0, x774);
+  fiat_p384_addcarryx_u32(&x800, &x801, 0x0, x794, x797);
   uint32_t x802;
   fiat_p384_uint1 x803;
-  fiat_p384_addcarryx_u32(&x802, &x803, x801, 0x0, x778);
+  fiat_p384_addcarryx_u32(&x802, &x803, x801, x792, x795);
   uint32_t x804;
   fiat_p384_uint1 x805;
-  fiat_p384_addcarryx_u32(&x804, &x805, x779, 0x0, x715);
+  fiat_p384_addcarryx_u32(&x804, &x805, x803, x790, x793);
   uint32_t x806;
   fiat_p384_uint1 x807;
-  fiat_p384_addcarryx_u32(&x806, &x807, x803, 0x0, (fiat_p384_uint1)x804);
+  fiat_p384_addcarryx_u32(&x806, &x807, x805, x788, x791);
   uint32_t x808;
-  uint32_t x809;
-  fiat_p384_mulx_u32(&x808, &x809, x780, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x809;
+  fiat_p384_addcarryx_u32(&x808, &x809, x807, x786, x789);
   uint32_t x810;
-  uint32_t x811;
-  fiat_p384_mulx_u32(&x810, &x811, x780, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x811;
+  fiat_p384_addcarryx_u32(&x810, &x811, x809, x784, x787);
   uint32_t x812;
-  uint32_t x813;
-  fiat_p384_mulx_u32(&x812, &x813, x780, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x813;
+  fiat_p384_addcarryx_u32(&x812, &x813, x811, x782, x785);
   uint32_t x814;
-  uint32_t x815;
-  fiat_p384_mulx_u32(&x814, &x815, x780, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x815;
+  fiat_p384_addcarryx_u32(&x814, &x815, x813, x780, x783);
   uint32_t x816;
-  uint32_t x817;
-  fiat_p384_mulx_u32(&x816, &x817, x780, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x817;
+  fiat_p384_addcarryx_u32(&x816, &x817, 0x0, x798, x756);
   uint32_t x818;
-  uint32_t x819;
-  fiat_p384_mulx_u32(&x818, &x819, x780, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x819;
+  fiat_p384_addcarryx_u32(&x818, &x819, x817, x799, x758);
   uint32_t x820;
-  uint32_t x821;
-  fiat_p384_mulx_u32(&x820, &x821, x780, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x821;
+  fiat_p384_addcarryx_u32(&x820, &x821, x819, 0x0, x760);
   uint32_t x822;
-  uint32_t x823;
-  fiat_p384_mulx_u32(&x822, &x823, x780, UINT32_C(0xfffffffe));
+  fiat_p384_uint1 x823;
+  fiat_p384_addcarryx_u32(&x822, &x823, x821, x796, x762);
   uint32_t x824;
-  uint32_t x825;
-  fiat_p384_mulx_u32(&x824, &x825, x780, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x825;
+  fiat_p384_addcarryx_u32(&x824, &x825, x823, x800, x764);
   uint32_t x826;
-  uint32_t x827;
-  fiat_p384_mulx_u32(&x826, &x827, x780, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x827;
+  fiat_p384_addcarryx_u32(&x826, &x827, x825, x802, x766);
   uint32_t x828;
   fiat_p384_uint1 x829;
-  fiat_p384_addcarryx_u32(&x828, &x829, 0x0, x822, x825);
+  fiat_p384_addcarryx_u32(&x828, &x829, x827, x804, x768);
   uint32_t x830;
   fiat_p384_uint1 x831;
-  fiat_p384_addcarryx_u32(&x830, &x831, x829, x820, x823);
+  fiat_p384_addcarryx_u32(&x830, &x831, x829, x806, x770);
   uint32_t x832;
   fiat_p384_uint1 x833;
-  fiat_p384_addcarryx_u32(&x832, &x833, x831, x818, x821);
+  fiat_p384_addcarryx_u32(&x832, &x833, x831, x808, x772);
   uint32_t x834;
   fiat_p384_uint1 x835;
-  fiat_p384_addcarryx_u32(&x834, &x835, x833, x816, x819);
+  fiat_p384_addcarryx_u32(&x834, &x835, x833, x810, x774);
   uint32_t x836;
   fiat_p384_uint1 x837;
-  fiat_p384_addcarryx_u32(&x836, &x837, x835, x814, x817);
+  fiat_p384_addcarryx_u32(&x836, &x837, x835, x812, x776);
   uint32_t x838;
   fiat_p384_uint1 x839;
-  fiat_p384_addcarryx_u32(&x838, &x839, x837, x812, x815);
+  fiat_p384_addcarryx_u32(&x838, &x839, x837, x814, x778);
   uint32_t x840;
   fiat_p384_uint1 x841;
-  fiat_p384_addcarryx_u32(&x840, &x841, x839, x810, x813);
+  fiat_p384_addcarryx_u32(&x840, &x841, x755, 0x0, 0x0);
   uint32_t x842;
   fiat_p384_uint1 x843;
-  fiat_p384_addcarryx_u32(&x842, &x843, x841, x808, x811);
+  fiat_p384_addcarryx_u32(&x842, &x843, x779, 0x0, (fiat_p384_uint1)x840);
   uint32_t x844;
   fiat_p384_uint1 x845;
-  fiat_p384_addcarryx_u32(&x844, &x845, 0x0, x826, x780);
+  fiat_p384_addcarryx_u32(&x844, &x845, x815, 0x0, x781);
   uint32_t x846;
   fiat_p384_uint1 x847;
-  fiat_p384_addcarryx_u32(&x846, &x847, x845, x827, x782);
+  fiat_p384_addcarryx_u32(&x846, &x847, x839, x844, x842);
   uint32_t x848;
   fiat_p384_uint1 x849;
-  fiat_p384_addcarryx_u32(&x848, &x849, x847, 0x0, x784);
+  fiat_p384_addcarryx_u32(&x848, &x849, 0x0, (arg1[10]), x818);
   uint32_t x850;
   fiat_p384_uint1 x851;
-  fiat_p384_addcarryx_u32(&x850, &x851, x849, x824, x786);
+  fiat_p384_addcarryx_u32(&x850, &x851, x849, 0x0, x820);
   uint32_t x852;
   fiat_p384_uint1 x853;
-  fiat_p384_addcarryx_u32(&x852, &x853, x851, x828, x788);
+  fiat_p384_addcarryx_u32(&x852, &x853, x851, 0x0, x822);
   uint32_t x854;
   fiat_p384_uint1 x855;
-  fiat_p384_addcarryx_u32(&x854, &x855, x853, x830, x790);
+  fiat_p384_addcarryx_u32(&x854, &x855, x853, 0x0, x824);
   uint32_t x856;
   fiat_p384_uint1 x857;
-  fiat_p384_addcarryx_u32(&x856, &x857, x855, x832, x792);
+  fiat_p384_addcarryx_u32(&x856, &x857, x855, 0x0, x826);
   uint32_t x858;
   fiat_p384_uint1 x859;
-  fiat_p384_addcarryx_u32(&x858, &x859, x857, x834, x794);
+  fiat_p384_addcarryx_u32(&x858, &x859, x857, 0x0, x828);
   uint32_t x860;
   fiat_p384_uint1 x861;
-  fiat_p384_addcarryx_u32(&x860, &x861, x859, x836, x796);
+  fiat_p384_addcarryx_u32(&x860, &x861, x859, 0x0, x830);
   uint32_t x862;
   fiat_p384_uint1 x863;
-  fiat_p384_addcarryx_u32(&x862, &x863, x861, x838, x798);
+  fiat_p384_addcarryx_u32(&x862, &x863, x861, 0x0, x832);
   uint32_t x864;
   fiat_p384_uint1 x865;
-  fiat_p384_addcarryx_u32(&x864, &x865, x863, x840, x800);
+  fiat_p384_addcarryx_u32(&x864, &x865, x863, 0x0, x834);
   uint32_t x866;
   fiat_p384_uint1 x867;
-  fiat_p384_addcarryx_u32(&x866, &x867, x865, x842, x802);
+  fiat_p384_addcarryx_u32(&x866, &x867, x865, 0x0, x836);
   uint32_t x868;
   fiat_p384_uint1 x869;
-  fiat_p384_addcarryx_u32(&x868, &x869, x843, 0x0, x809);
+  fiat_p384_addcarryx_u32(&x868, &x869, x867, 0x0, x838);
   uint32_t x870;
   fiat_p384_uint1 x871;
-  fiat_p384_addcarryx_u32(&x870, &x871, x867, x868, x806);
+  fiat_p384_addcarryx_u32(&x870, &x871, x869, 0x0, x846);
   uint32_t x872;
-  fiat_p384_uint1 x873;
-  fiat_p384_addcarryx_u32(&x872, &x873, 0x0, (arg1[10]), x846);
+  uint32_t x873;
+  fiat_p384_mulx_u32(&x872, &x873, x848, UINT32_C(0xffffffff));
   uint32_t x874;
-  fiat_p384_uint1 x875;
-  fiat_p384_addcarryx_u32(&x874, &x875, x873, 0x0, x848);
+  uint32_t x875;
+  fiat_p384_mulx_u32(&x874, &x875, x848, UINT32_C(0xffffffff));
   uint32_t x876;
-  fiat_p384_uint1 x877;
-  fiat_p384_addcarryx_u32(&x876, &x877, x875, 0x0, x850);
+  uint32_t x877;
+  fiat_p384_mulx_u32(&x876, &x877, x848, UINT32_C(0xffffffff));
   uint32_t x878;
-  fiat_p384_uint1 x879;
-  fiat_p384_addcarryx_u32(&x878, &x879, x877, 0x0, x852);
+  uint32_t x879;
+  fiat_p384_mulx_u32(&x878, &x879, x848, UINT32_C(0xffffffff));
   uint32_t x880;
-  fiat_p384_uint1 x881;
-  fiat_p384_addcarryx_u32(&x880, &x881, x879, 0x0, x854);
+  uint32_t x881;
+  fiat_p384_mulx_u32(&x880, &x881, x848, UINT32_C(0xffffffff));
   uint32_t x882;
-  fiat_p384_uint1 x883;
-  fiat_p384_addcarryx_u32(&x882, &x883, x881, 0x0, x856);
+  uint32_t x883;
+  fiat_p384_mulx_u32(&x882, &x883, x848, UINT32_C(0xffffffff));
   uint32_t x884;
-  fiat_p384_uint1 x885;
-  fiat_p384_addcarryx_u32(&x884, &x885, x883, 0x0, x858);
+  uint32_t x885;
+  fiat_p384_mulx_u32(&x884, &x885, x848, UINT32_C(0xffffffff));
   uint32_t x886;
-  fiat_p384_uint1 x887;
-  fiat_p384_addcarryx_u32(&x886, &x887, x885, 0x0, x860);
+  uint32_t x887;
+  fiat_p384_mulx_u32(&x886, &x887, x848, UINT32_C(0xfffffffe));
   uint32_t x888;
-  fiat_p384_uint1 x889;
-  fiat_p384_addcarryx_u32(&x888, &x889, x887, 0x0, x862);
+  uint32_t x889;
+  fiat_p384_mulx_u32(&x888, &x889, x848, UINT32_C(0xffffffff));
   uint32_t x890;
-  fiat_p384_uint1 x891;
-  fiat_p384_addcarryx_u32(&x890, &x891, x889, 0x0, x864);
+  uint32_t x891;
+  fiat_p384_mulx_u32(&x890, &x891, x848, UINT32_C(0xffffffff));
   uint32_t x892;
   fiat_p384_uint1 x893;
-  fiat_p384_addcarryx_u32(&x892, &x893, x891, 0x0, x866);
+  fiat_p384_addcarryx_u32(&x892, &x893, 0x0, x886, x889);
   uint32_t x894;
   fiat_p384_uint1 x895;
-  fiat_p384_addcarryx_u32(&x894, &x895, x893, 0x0, x870);
+  fiat_p384_addcarryx_u32(&x894, &x895, x893, x884, x887);
   uint32_t x896;
   fiat_p384_uint1 x897;
-  fiat_p384_addcarryx_u32(&x896, &x897, x871, 0x0, x807);
+  fiat_p384_addcarryx_u32(&x896, &x897, x895, x882, x885);
   uint32_t x898;
   fiat_p384_uint1 x899;
-  fiat_p384_addcarryx_u32(&x898, &x899, x895, 0x0, (fiat_p384_uint1)x896);
+  fiat_p384_addcarryx_u32(&x898, &x899, x897, x880, x883);
   uint32_t x900;
-  uint32_t x901;
-  fiat_p384_mulx_u32(&x900, &x901, x872, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x901;
+  fiat_p384_addcarryx_u32(&x900, &x901, x899, x878, x881);
   uint32_t x902;
-  uint32_t x903;
-  fiat_p384_mulx_u32(&x902, &x903, x872, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x903;
+  fiat_p384_addcarryx_u32(&x902, &x903, x901, x876, x879);
   uint32_t x904;
-  uint32_t x905;
-  fiat_p384_mulx_u32(&x904, &x905, x872, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x905;
+  fiat_p384_addcarryx_u32(&x904, &x905, x903, x874, x877);
   uint32_t x906;
-  uint32_t x907;
-  fiat_p384_mulx_u32(&x906, &x907, x872, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x907;
+  fiat_p384_addcarryx_u32(&x906, &x907, x905, x872, x875);
   uint32_t x908;
-  uint32_t x909;
-  fiat_p384_mulx_u32(&x908, &x909, x872, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x909;
+  fiat_p384_addcarryx_u32(&x908, &x909, 0x0, x890, x848);
   uint32_t x910;
-  uint32_t x911;
-  fiat_p384_mulx_u32(&x910, &x911, x872, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x911;
+  fiat_p384_addcarryx_u32(&x910, &x911, x909, x891, x850);
   uint32_t x912;
-  uint32_t x913;
-  fiat_p384_mulx_u32(&x912, &x913, x872, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x913;
+  fiat_p384_addcarryx_u32(&x912, &x913, x911, 0x0, x852);
   uint32_t x914;
-  uint32_t x915;
-  fiat_p384_mulx_u32(&x914, &x915, x872, UINT32_C(0xfffffffe));
+  fiat_p384_uint1 x915;
+  fiat_p384_addcarryx_u32(&x914, &x915, x913, x888, x854);
   uint32_t x916;
-  uint32_t x917;
-  fiat_p384_mulx_u32(&x916, &x917, x872, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x917;
+  fiat_p384_addcarryx_u32(&x916, &x917, x915, x892, x856);
   uint32_t x918;
-  uint32_t x919;
-  fiat_p384_mulx_u32(&x918, &x919, x872, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x919;
+  fiat_p384_addcarryx_u32(&x918, &x919, x917, x894, x858);
   uint32_t x920;
   fiat_p384_uint1 x921;
-  fiat_p384_addcarryx_u32(&x920, &x921, 0x0, x914, x917);
+  fiat_p384_addcarryx_u32(&x920, &x921, x919, x896, x860);
   uint32_t x922;
   fiat_p384_uint1 x923;
-  fiat_p384_addcarryx_u32(&x922, &x923, x921, x912, x915);
+  fiat_p384_addcarryx_u32(&x922, &x923, x921, x898, x862);
   uint32_t x924;
   fiat_p384_uint1 x925;
-  fiat_p384_addcarryx_u32(&x924, &x925, x923, x910, x913);
+  fiat_p384_addcarryx_u32(&x924, &x925, x923, x900, x864);
   uint32_t x926;
   fiat_p384_uint1 x927;
-  fiat_p384_addcarryx_u32(&x926, &x927, x925, x908, x911);
+  fiat_p384_addcarryx_u32(&x926, &x927, x925, x902, x866);
   uint32_t x928;
   fiat_p384_uint1 x929;
-  fiat_p384_addcarryx_u32(&x928, &x929, x927, x906, x909);
+  fiat_p384_addcarryx_u32(&x928, &x929, x927, x904, x868);
   uint32_t x930;
   fiat_p384_uint1 x931;
-  fiat_p384_addcarryx_u32(&x930, &x931, x929, x904, x907);
+  fiat_p384_addcarryx_u32(&x930, &x931, x929, x906, x870);
   uint32_t x932;
   fiat_p384_uint1 x933;
-  fiat_p384_addcarryx_u32(&x932, &x933, x931, x902, x905);
+  fiat_p384_addcarryx_u32(&x932, &x933, x847, 0x0, 0x0);
   uint32_t x934;
   fiat_p384_uint1 x935;
-  fiat_p384_addcarryx_u32(&x934, &x935, x933, x900, x903);
+  fiat_p384_addcarryx_u32(&x934, &x935, x871, 0x0, (fiat_p384_uint1)x932);
   uint32_t x936;
   fiat_p384_uint1 x937;
-  fiat_p384_addcarryx_u32(&x936, &x937, 0x0, x918, x872);
+  fiat_p384_addcarryx_u32(&x936, &x937, x907, 0x0, x873);
   uint32_t x938;
   fiat_p384_uint1 x939;
-  fiat_p384_addcarryx_u32(&x938, &x939, x937, x919, x874);
+  fiat_p384_addcarryx_u32(&x938, &x939, x931, x936, x934);
   uint32_t x940;
   fiat_p384_uint1 x941;
-  fiat_p384_addcarryx_u32(&x940, &x941, x939, 0x0, x876);
+  fiat_p384_addcarryx_u32(&x940, &x941, 0x0, (arg1[11]), x910);
   uint32_t x942;
   fiat_p384_uint1 x943;
-  fiat_p384_addcarryx_u32(&x942, &x943, x941, x916, x878);
+  fiat_p384_addcarryx_u32(&x942, &x943, x941, 0x0, x912);
   uint32_t x944;
   fiat_p384_uint1 x945;
-  fiat_p384_addcarryx_u32(&x944, &x945, x943, x920, x880);
+  fiat_p384_addcarryx_u32(&x944, &x945, x943, 0x0, x914);
   uint32_t x946;
   fiat_p384_uint1 x947;
-  fiat_p384_addcarryx_u32(&x946, &x947, x945, x922, x882);
+  fiat_p384_addcarryx_u32(&x946, &x947, x945, 0x0, x916);
   uint32_t x948;
   fiat_p384_uint1 x949;
-  fiat_p384_addcarryx_u32(&x948, &x949, x947, x924, x884);
+  fiat_p384_addcarryx_u32(&x948, &x949, x947, 0x0, x918);
   uint32_t x950;
   fiat_p384_uint1 x951;
-  fiat_p384_addcarryx_u32(&x950, &x951, x949, x926, x886);
+  fiat_p384_addcarryx_u32(&x950, &x951, x949, 0x0, x920);
   uint32_t x952;
   fiat_p384_uint1 x953;
-  fiat_p384_addcarryx_u32(&x952, &x953, x951, x928, x888);
+  fiat_p384_addcarryx_u32(&x952, &x953, x951, 0x0, x922);
   uint32_t x954;
   fiat_p384_uint1 x955;
-  fiat_p384_addcarryx_u32(&x954, &x955, x953, x930, x890);
+  fiat_p384_addcarryx_u32(&x954, &x955, x953, 0x0, x924);
   uint32_t x956;
   fiat_p384_uint1 x957;
-  fiat_p384_addcarryx_u32(&x956, &x957, x955, x932, x892);
+  fiat_p384_addcarryx_u32(&x956, &x957, x955, 0x0, x926);
   uint32_t x958;
   fiat_p384_uint1 x959;
-  fiat_p384_addcarryx_u32(&x958, &x959, x957, x934, x894);
+  fiat_p384_addcarryx_u32(&x958, &x959, x957, 0x0, x928);
   uint32_t x960;
   fiat_p384_uint1 x961;
-  fiat_p384_addcarryx_u32(&x960, &x961, x935, 0x0, x901);
+  fiat_p384_addcarryx_u32(&x960, &x961, x959, 0x0, x930);
   uint32_t x962;
   fiat_p384_uint1 x963;
-  fiat_p384_addcarryx_u32(&x962, &x963, x959, x960, x898);
+  fiat_p384_addcarryx_u32(&x962, &x963, x961, 0x0, x938);
   uint32_t x964;
-  fiat_p384_uint1 x965;
-  fiat_p384_addcarryx_u32(&x964, &x965, 0x0, (arg1[11]), x938);
+  uint32_t x965;
+  fiat_p384_mulx_u32(&x964, &x965, x940, UINT32_C(0xffffffff));
   uint32_t x966;
-  fiat_p384_uint1 x967;
-  fiat_p384_addcarryx_u32(&x966, &x967, x965, 0x0, x940);
+  uint32_t x967;
+  fiat_p384_mulx_u32(&x966, &x967, x940, UINT32_C(0xffffffff));
   uint32_t x968;
-  fiat_p384_uint1 x969;
-  fiat_p384_addcarryx_u32(&x968, &x969, x967, 0x0, x942);
+  uint32_t x969;
+  fiat_p384_mulx_u32(&x968, &x969, x940, UINT32_C(0xffffffff));
   uint32_t x970;
-  fiat_p384_uint1 x971;
-  fiat_p384_addcarryx_u32(&x970, &x971, x969, 0x0, x944);
+  uint32_t x971;
+  fiat_p384_mulx_u32(&x970, &x971, x940, UINT32_C(0xffffffff));
   uint32_t x972;
-  fiat_p384_uint1 x973;
-  fiat_p384_addcarryx_u32(&x972, &x973, x971, 0x0, x946);
+  uint32_t x973;
+  fiat_p384_mulx_u32(&x972, &x973, x940, UINT32_C(0xffffffff));
   uint32_t x974;
-  fiat_p384_uint1 x975;
-  fiat_p384_addcarryx_u32(&x974, &x975, x973, 0x0, x948);
+  uint32_t x975;
+  fiat_p384_mulx_u32(&x974, &x975, x940, UINT32_C(0xffffffff));
   uint32_t x976;
-  fiat_p384_uint1 x977;
-  fiat_p384_addcarryx_u32(&x976, &x977, x975, 0x0, x950);
+  uint32_t x977;
+  fiat_p384_mulx_u32(&x976, &x977, x940, UINT32_C(0xffffffff));
   uint32_t x978;
-  fiat_p384_uint1 x979;
-  fiat_p384_addcarryx_u32(&x978, &x979, x977, 0x0, x952);
+  uint32_t x979;
+  fiat_p384_mulx_u32(&x978, &x979, x940, UINT32_C(0xfffffffe));
   uint32_t x980;
-  fiat_p384_uint1 x981;
-  fiat_p384_addcarryx_u32(&x980, &x981, x979, 0x0, x954);
+  uint32_t x981;
+  fiat_p384_mulx_u32(&x980, &x981, x940, UINT32_C(0xffffffff));
   uint32_t x982;
-  fiat_p384_uint1 x983;
-  fiat_p384_addcarryx_u32(&x982, &x983, x981, 0x0, x956);
+  uint32_t x983;
+  fiat_p384_mulx_u32(&x982, &x983, x940, UINT32_C(0xffffffff));
   uint32_t x984;
   fiat_p384_uint1 x985;
-  fiat_p384_addcarryx_u32(&x984, &x985, x983, 0x0, x958);
+  fiat_p384_addcarryx_u32(&x984, &x985, 0x0, x978, x981);
   uint32_t x986;
   fiat_p384_uint1 x987;
-  fiat_p384_addcarryx_u32(&x986, &x987, x985, 0x0, x962);
+  fiat_p384_addcarryx_u32(&x986, &x987, x985, x976, x979);
   uint32_t x988;
   fiat_p384_uint1 x989;
-  fiat_p384_addcarryx_u32(&x988, &x989, x963, 0x0, x899);
+  fiat_p384_addcarryx_u32(&x988, &x989, x987, x974, x977);
   uint32_t x990;
   fiat_p384_uint1 x991;
-  fiat_p384_addcarryx_u32(&x990, &x991, x987, 0x0, (fiat_p384_uint1)x988);
+  fiat_p384_addcarryx_u32(&x990, &x991, x989, x972, x975);
   uint32_t x992;
-  uint32_t x993;
-  fiat_p384_mulx_u32(&x992, &x993, x964, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x993;
+  fiat_p384_addcarryx_u32(&x992, &x993, x991, x970, x973);
   uint32_t x994;
-  uint32_t x995;
-  fiat_p384_mulx_u32(&x994, &x995, x964, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x995;
+  fiat_p384_addcarryx_u32(&x994, &x995, x993, x968, x971);
   uint32_t x996;
-  uint32_t x997;
-  fiat_p384_mulx_u32(&x996, &x997, x964, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x997;
+  fiat_p384_addcarryx_u32(&x996, &x997, x995, x966, x969);
   uint32_t x998;
-  uint32_t x999;
-  fiat_p384_mulx_u32(&x998, &x999, x964, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x999;
+  fiat_p384_addcarryx_u32(&x998, &x999, x997, x964, x967);
   uint32_t x1000;
-  uint32_t x1001;
-  fiat_p384_mulx_u32(&x1000, &x1001, x964, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x1001;
+  fiat_p384_addcarryx_u32(&x1000, &x1001, 0x0, x982, x940);
   uint32_t x1002;
-  uint32_t x1003;
-  fiat_p384_mulx_u32(&x1002, &x1003, x964, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x1003;
+  fiat_p384_addcarryx_u32(&x1002, &x1003, x1001, x983, x942);
   uint32_t x1004;
-  uint32_t x1005;
-  fiat_p384_mulx_u32(&x1004, &x1005, x964, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x1005;
+  fiat_p384_addcarryx_u32(&x1004, &x1005, x1003, 0x0, x944);
   uint32_t x1006;
-  uint32_t x1007;
-  fiat_p384_mulx_u32(&x1006, &x1007, x964, UINT32_C(0xfffffffe));
+  fiat_p384_uint1 x1007;
+  fiat_p384_addcarryx_u32(&x1006, &x1007, x1005, x980, x946);
   uint32_t x1008;
-  uint32_t x1009;
-  fiat_p384_mulx_u32(&x1008, &x1009, x964, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x1009;
+  fiat_p384_addcarryx_u32(&x1008, &x1009, x1007, x984, x948);
   uint32_t x1010;
-  uint32_t x1011;
-  fiat_p384_mulx_u32(&x1010, &x1011, x964, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x1011;
+  fiat_p384_addcarryx_u32(&x1010, &x1011, x1009, x986, x950);
   uint32_t x1012;
   fiat_p384_uint1 x1013;
-  fiat_p384_addcarryx_u32(&x1012, &x1013, 0x0, x1006, x1009);
+  fiat_p384_addcarryx_u32(&x1012, &x1013, x1011, x988, x952);
   uint32_t x1014;
   fiat_p384_uint1 x1015;
-  fiat_p384_addcarryx_u32(&x1014, &x1015, x1013, x1004, x1007);
+  fiat_p384_addcarryx_u32(&x1014, &x1015, x1013, x990, x954);
   uint32_t x1016;
   fiat_p384_uint1 x1017;
-  fiat_p384_addcarryx_u32(&x1016, &x1017, x1015, x1002, x1005);
+  fiat_p384_addcarryx_u32(&x1016, &x1017, x1015, x992, x956);
   uint32_t x1018;
   fiat_p384_uint1 x1019;
-  fiat_p384_addcarryx_u32(&x1018, &x1019, x1017, x1000, x1003);
+  fiat_p384_addcarryx_u32(&x1018, &x1019, x1017, x994, x958);
   uint32_t x1020;
   fiat_p384_uint1 x1021;
-  fiat_p384_addcarryx_u32(&x1020, &x1021, x1019, x998, x1001);
+  fiat_p384_addcarryx_u32(&x1020, &x1021, x1019, x996, x960);
   uint32_t x1022;
   fiat_p384_uint1 x1023;
-  fiat_p384_addcarryx_u32(&x1022, &x1023, x1021, x996, x999);
+  fiat_p384_addcarryx_u32(&x1022, &x1023, x1021, x998, x962);
   uint32_t x1024;
   fiat_p384_uint1 x1025;
-  fiat_p384_addcarryx_u32(&x1024, &x1025, x1023, x994, x997);
+  fiat_p384_addcarryx_u32(&x1024, &x1025, x939, 0x0, 0x0);
   uint32_t x1026;
   fiat_p384_uint1 x1027;
-  fiat_p384_addcarryx_u32(&x1026, &x1027, x1025, x992, x995);
+  fiat_p384_addcarryx_u32(&x1026, &x1027, x963, 0x0, (fiat_p384_uint1)x1024);
   uint32_t x1028;
   fiat_p384_uint1 x1029;
-  fiat_p384_addcarryx_u32(&x1028, &x1029, 0x0, x1010, x964);
+  fiat_p384_addcarryx_u32(&x1028, &x1029, x999, 0x0, x965);
   uint32_t x1030;
   fiat_p384_uint1 x1031;
-  fiat_p384_addcarryx_u32(&x1030, &x1031, x1029, x1011, x966);
+  fiat_p384_addcarryx_u32(&x1030, &x1031, x1023, x1028, x1026);
   uint32_t x1032;
   fiat_p384_uint1 x1033;
-  fiat_p384_addcarryx_u32(&x1032, &x1033, x1031, 0x0, x968);
+  fiat_p384_subborrowx_u32(&x1032, &x1033, 0x0, x1002, UINT32_C(0xffffffff));
   uint32_t x1034;
   fiat_p384_uint1 x1035;
-  fiat_p384_addcarryx_u32(&x1034, &x1035, x1033, x1008, x970);
+  fiat_p384_subborrowx_u32(&x1034, &x1035, x1033, x1004, 0x0);
   uint32_t x1036;
   fiat_p384_uint1 x1037;
-  fiat_p384_addcarryx_u32(&x1036, &x1037, x1035, x1012, x972);
+  fiat_p384_subborrowx_u32(&x1036, &x1037, x1035, x1006, 0x0);
   uint32_t x1038;
   fiat_p384_uint1 x1039;
-  fiat_p384_addcarryx_u32(&x1038, &x1039, x1037, x1014, x974);
+  fiat_p384_subborrowx_u32(&x1038, &x1039, x1037, x1008, UINT32_C(0xffffffff));
   uint32_t x1040;
   fiat_p384_uint1 x1041;
-  fiat_p384_addcarryx_u32(&x1040, &x1041, x1039, x1016, x976);
+  fiat_p384_subborrowx_u32(&x1040, &x1041, x1039, x1010, UINT32_C(0xfffffffe));
   uint32_t x1042;
   fiat_p384_uint1 x1043;
-  fiat_p384_addcarryx_u32(&x1042, &x1043, x1041, x1018, x978);
+  fiat_p384_subborrowx_u32(&x1042, &x1043, x1041, x1012, UINT32_C(0xffffffff));
   uint32_t x1044;
   fiat_p384_uint1 x1045;
-  fiat_p384_addcarryx_u32(&x1044, &x1045, x1043, x1020, x980);
+  fiat_p384_subborrowx_u32(&x1044, &x1045, x1043, x1014, UINT32_C(0xffffffff));
   uint32_t x1046;
   fiat_p384_uint1 x1047;
-  fiat_p384_addcarryx_u32(&x1046, &x1047, x1045, x1022, x982);
+  fiat_p384_subborrowx_u32(&x1046, &x1047, x1045, x1016, UINT32_C(0xffffffff));
   uint32_t x1048;
   fiat_p384_uint1 x1049;
-  fiat_p384_addcarryx_u32(&x1048, &x1049, x1047, x1024, x984);
+  fiat_p384_subborrowx_u32(&x1048, &x1049, x1047, x1018, UINT32_C(0xffffffff));
   uint32_t x1050;
   fiat_p384_uint1 x1051;
-  fiat_p384_addcarryx_u32(&x1050, &x1051, x1049, x1026, x986);
+  fiat_p384_subborrowx_u32(&x1050, &x1051, x1049, x1020, UINT32_C(0xffffffff));
   uint32_t x1052;
   fiat_p384_uint1 x1053;
-  fiat_p384_addcarryx_u32(&x1052, &x1053, x1027, 0x0, x993);
+  fiat_p384_subborrowx_u32(&x1052, &x1053, x1051, x1022, UINT32_C(0xffffffff));
   uint32_t x1054;
   fiat_p384_uint1 x1055;
-  fiat_p384_addcarryx_u32(&x1054, &x1055, x1051, x1052, x990);
+  fiat_p384_subborrowx_u32(&x1054, &x1055, x1053, x1030, UINT32_C(0xffffffff));
   uint32_t x1056;
   fiat_p384_uint1 x1057;
-  fiat_p384_subborrowx_u32(&x1056, &x1057, 0x0, x1030, UINT32_C(0xffffffff));
+  fiat_p384_addcarryx_u32(&x1056, &x1057, x1031, 0x0, 0x0);
   uint32_t x1058;
   fiat_p384_uint1 x1059;
-  fiat_p384_subborrowx_u32(&x1058, &x1059, x1057, x1032, 0x0);
+  fiat_p384_subborrowx_u32(&x1058, &x1059, x1055, (fiat_p384_uint1)x1056, 0x0);
   uint32_t x1060;
-  fiat_p384_uint1 x1061;
-  fiat_p384_subborrowx_u32(&x1060, &x1061, x1059, x1034, 0x0);
+  fiat_p384_cmovznz_u32(&x1060, x1059, x1032, x1002);
+  uint32_t x1061;
+  fiat_p384_cmovznz_u32(&x1061, x1059, x1034, x1004);
   uint32_t x1062;
-  fiat_p384_uint1 x1063;
-  fiat_p384_subborrowx_u32(&x1062, &x1063, x1061, x1036, UINT32_C(0xffffffff));
+  fiat_p384_cmovznz_u32(&x1062, x1059, x1036, x1006);
+  uint32_t x1063;
+  fiat_p384_cmovznz_u32(&x1063, x1059, x1038, x1008);
   uint32_t x1064;
-  fiat_p384_uint1 x1065;
-  fiat_p384_subborrowx_u32(&x1064, &x1065, x1063, x1038, UINT32_C(0xfffffffe));
+  fiat_p384_cmovznz_u32(&x1064, x1059, x1040, x1010);
+  uint32_t x1065;
+  fiat_p384_cmovznz_u32(&x1065, x1059, x1042, x1012);
   uint32_t x1066;
-  fiat_p384_uint1 x1067;
-  fiat_p384_subborrowx_u32(&x1066, &x1067, x1065, x1040, UINT32_C(0xffffffff));
+  fiat_p384_cmovznz_u32(&x1066, x1059, x1044, x1014);
+  uint32_t x1067;
+  fiat_p384_cmovznz_u32(&x1067, x1059, x1046, x1016);
   uint32_t x1068;
-  fiat_p384_uint1 x1069;
-  fiat_p384_subborrowx_u32(&x1068, &x1069, x1067, x1042, UINT32_C(0xffffffff));
+  fiat_p384_cmovznz_u32(&x1068, x1059, x1048, x1018);
+  uint32_t x1069;
+  fiat_p384_cmovznz_u32(&x1069, x1059, x1050, x1020);
   uint32_t x1070;
-  fiat_p384_uint1 x1071;
-  fiat_p384_subborrowx_u32(&x1070, &x1071, x1069, x1044, UINT32_C(0xffffffff));
-  uint32_t x1072;
-  fiat_p384_uint1 x1073;
-  fiat_p384_subborrowx_u32(&x1072, &x1073, x1071, x1046, UINT32_C(0xffffffff));
-  uint32_t x1074;
-  fiat_p384_uint1 x1075;
-  fiat_p384_subborrowx_u32(&x1074, &x1075, x1073, x1048, UINT32_C(0xffffffff));
-  uint32_t x1076;
-  fiat_p384_uint1 x1077;
-  fiat_p384_subborrowx_u32(&x1076, &x1077, x1075, x1050, UINT32_C(0xffffffff));
-  uint32_t x1078;
-  fiat_p384_uint1 x1079;
-  fiat_p384_subborrowx_u32(&x1078, &x1079, x1077, x1054, UINT32_C(0xffffffff));
-  uint32_t x1080;
-  fiat_p384_uint1 x1081;
-  fiat_p384_addcarryx_u32(&x1080, &x1081, x1055, 0x0, x991);
-  uint32_t x1082;
-  fiat_p384_uint1 x1083;
-  fiat_p384_subborrowx_u32(&x1082, &x1083, x1079, (fiat_p384_uint1)x1080, 0x0);
-  uint32_t x1084;
-  fiat_p384_cmovznz_u32(&x1084, x1083, x1056, x1030);
-  uint32_t x1085;
-  fiat_p384_cmovznz_u32(&x1085, x1083, x1058, x1032);
-  uint32_t x1086;
-  fiat_p384_cmovznz_u32(&x1086, x1083, x1060, x1034);
-  uint32_t x1087;
-  fiat_p384_cmovznz_u32(&x1087, x1083, x1062, x1036);
-  uint32_t x1088;
-  fiat_p384_cmovznz_u32(&x1088, x1083, x1064, x1038);
-  uint32_t x1089;
-  fiat_p384_cmovznz_u32(&x1089, x1083, x1066, x1040);
-  uint32_t x1090;
-  fiat_p384_cmovznz_u32(&x1090, x1083, x1068, x1042);
-  uint32_t x1091;
-  fiat_p384_cmovznz_u32(&x1091, x1083, x1070, x1044);
-  uint32_t x1092;
-  fiat_p384_cmovznz_u32(&x1092, x1083, x1072, x1046);
-  uint32_t x1093;
-  fiat_p384_cmovznz_u32(&x1093, x1083, x1074, x1048);
-  uint32_t x1094;
-  fiat_p384_cmovznz_u32(&x1094, x1083, x1076, x1050);
-  uint32_t x1095;
-  fiat_p384_cmovznz_u32(&x1095, x1083, x1078, x1054);
-  out1[0] = x1084;
-  out1[1] = x1085;
-  out1[2] = x1086;
-  out1[3] = x1087;
-  out1[4] = x1088;
-  out1[5] = x1089;
-  out1[6] = x1090;
-  out1[7] = x1091;
-  out1[8] = x1092;
-  out1[9] = x1093;
-  out1[10] = x1094;
-  out1[11] = x1095;
+  fiat_p384_cmovznz_u32(&x1070, x1059, x1052, x1022);
+  uint32_t x1071;
+  fiat_p384_cmovznz_u32(&x1071, x1059, x1054, x1030);
+  out1[0] = x1060;
+  out1[1] = x1061;
+  out1[2] = x1062;
+  out1[3] = x1063;
+  out1[4] = x1064;
+  out1[5] = x1065;
+  out1[6] = x1066;
+  out1[7] = x1067;
+  out1[8] = x1068;
+  out1[9] = x1069;
+  out1[10] = x1070;
+  out1[11] = x1071;
 }
 
 /*
@@ -7300,153 +7264,142 @@ static void fiat_p384_to_bytes(uint8_t out1[48], const uint32_t arg1[12]) {
   uint8_t x16 = (uint8_t)(x13 & UINT8_C(0xff));
   uint8_t x17 = (uint8_t)(x15 >> 8);
   uint8_t x18 = (uint8_t)(x15 & UINT8_C(0xff));
-  fiat_p384_uint1 x19 = (fiat_p384_uint1)(x17 >> 8);
-  uint8_t x20 = (uint8_t)(x17 & UINT8_C(0xff));
-  uint32_t x21 = (x19 + x11);
-  uint32_t x22 = (x21 >> 8);
-  uint8_t x23 = (uint8_t)(x21 & UINT8_C(0xff));
-  uint32_t x24 = (x22 >> 8);
-  uint8_t x25 = (uint8_t)(x22 & UINT8_C(0xff));
-  uint8_t x26 = (uint8_t)(x24 >> 8);
-  uint8_t x27 = (uint8_t)(x24 & UINT8_C(0xff));
-  fiat_p384_uint1 x28 = (fiat_p384_uint1)(x26 >> 8);
-  uint8_t x29 = (uint8_t)(x26 & UINT8_C(0xff));
-  uint32_t x30 = (x28 + x10);
-  uint32_t x31 = (x30 >> 8);
-  uint8_t x32 = (uint8_t)(x30 & UINT8_C(0xff));
-  uint32_t x33 = (x31 >> 8);
+  uint8_t x19 = (uint8_t)(x17 & UINT8_C(0xff));
+  uint32_t x20 = (0x0 + x11);
+  uint32_t x21 = (x20 >> 8);
+  uint8_t x22 = (uint8_t)(x20 & UINT8_C(0xff));
+  uint32_t x23 = (x21 >> 8);
+  uint8_t x24 = (uint8_t)(x21 & UINT8_C(0xff));
+  uint8_t x25 = (uint8_t)(x23 >> 8);
+  uint8_t x26 = (uint8_t)(x23 & UINT8_C(0xff));
+  uint8_t x27 = (uint8_t)(x25 & UINT8_C(0xff));
+  uint32_t x28 = (0x0 + x10);
+  uint32_t x29 = (x28 >> 8);
+  uint8_t x30 = (uint8_t)(x28 & UINT8_C(0xff));
+  uint32_t x31 = (x29 >> 8);
+  uint8_t x32 = (uint8_t)(x29 & UINT8_C(0xff));
+  uint8_t x33 = (uint8_t)(x31 >> 8);
   uint8_t x34 = (uint8_t)(x31 & UINT8_C(0xff));
-  uint8_t x35 = (uint8_t)(x33 >> 8);
-  uint8_t x36 = (uint8_t)(x33 & UINT8_C(0xff));
-  fiat_p384_uint1 x37 = (fiat_p384_uint1)(x35 >> 8);
-  uint8_t x38 = (uint8_t)(x35 & UINT8_C(0xff));
-  uint32_t x39 = (x37 + x9);
-  uint32_t x40 = (x39 >> 8);
-  uint8_t x41 = (uint8_t)(x39 & UINT8_C(0xff));
-  uint32_t x42 = (x40 >> 8);
-  uint8_t x43 = (uint8_t)(x40 & UINT8_C(0xff));
-  uint8_t x44 = (uint8_t)(x42 >> 8);
-  uint8_t x45 = (uint8_t)(x42 & UINT8_C(0xff));
-  fiat_p384_uint1 x46 = (fiat_p384_uint1)(x44 >> 8);
-  uint8_t x47 = (uint8_t)(x44 & UINT8_C(0xff));
-  uint32_t x48 = (x46 + x8);
-  uint32_t x49 = (x48 >> 8);
-  uint8_t x50 = (uint8_t)(x48 & UINT8_C(0xff));
-  uint32_t x51 = (x49 >> 8);
-  uint8_t x52 = (uint8_t)(x49 & UINT8_C(0xff));
-  uint8_t x53 = (uint8_t)(x51 >> 8);
-  uint8_t x54 = (uint8_t)(x51 & UINT8_C(0xff));
-  fiat_p384_uint1 x55 = (fiat_p384_uint1)(x53 >> 8);
+  uint8_t x35 = (uint8_t)(x33 & UINT8_C(0xff));
+  uint32_t x36 = (0x0 + x9);
+  uint32_t x37 = (x36 >> 8);
+  uint8_t x38 = (uint8_t)(x36 & UINT8_C(0xff));
+  uint32_t x39 = (x37 >> 8);
+  uint8_t x40 = (uint8_t)(x37 & UINT8_C(0xff));
+  uint8_t x41 = (uint8_t)(x39 >> 8);
+  uint8_t x42 = (uint8_t)(x39 & UINT8_C(0xff));
+  uint8_t x43 = (uint8_t)(x41 & UINT8_C(0xff));
+  uint32_t x44 = (0x0 + x8);
+  uint32_t x45 = (x44 >> 8);
+  uint8_t x46 = (uint8_t)(x44 & UINT8_C(0xff));
+  uint32_t x47 = (x45 >> 8);
+  uint8_t x48 = (uint8_t)(x45 & UINT8_C(0xff));
+  uint8_t x49 = (uint8_t)(x47 >> 8);
+  uint8_t x50 = (uint8_t)(x47 & UINT8_C(0xff));
+  uint8_t x51 = (uint8_t)(x49 & UINT8_C(0xff));
+  uint32_t x52 = (0x0 + x7);
+  uint32_t x53 = (x52 >> 8);
+  uint8_t x54 = (uint8_t)(x52 & UINT8_C(0xff));
+  uint32_t x55 = (x53 >> 8);
   uint8_t x56 = (uint8_t)(x53 & UINT8_C(0xff));
-  uint32_t x57 = (x55 + x7);
-  uint32_t x58 = (x57 >> 8);
+  uint8_t x57 = (uint8_t)(x55 >> 8);
+  uint8_t x58 = (uint8_t)(x55 & UINT8_C(0xff));
   uint8_t x59 = (uint8_t)(x57 & UINT8_C(0xff));
-  uint32_t x60 = (x58 >> 8);
-  uint8_t x61 = (uint8_t)(x58 & UINT8_C(0xff));
-  uint8_t x62 = (uint8_t)(x60 >> 8);
-  uint8_t x63 = (uint8_t)(x60 & UINT8_C(0xff));
-  fiat_p384_uint1 x64 = (fiat_p384_uint1)(x62 >> 8);
-  uint8_t x65 = (uint8_t)(x62 & UINT8_C(0xff));
-  uint32_t x66 = (x64 + x6);
-  uint32_t x67 = (x66 >> 8);
-  uint8_t x68 = (uint8_t)(x66 & UINT8_C(0xff));
-  uint32_t x69 = (x67 >> 8);
-  uint8_t x70 = (uint8_t)(x67 & UINT8_C(0xff));
-  uint8_t x71 = (uint8_t)(x69 >> 8);
+  uint32_t x60 = (0x0 + x6);
+  uint32_t x61 = (x60 >> 8);
+  uint8_t x62 = (uint8_t)(x60 & UINT8_C(0xff));
+  uint32_t x63 = (x61 >> 8);
+  uint8_t x64 = (uint8_t)(x61 & UINT8_C(0xff));
+  uint8_t x65 = (uint8_t)(x63 >> 8);
+  uint8_t x66 = (uint8_t)(x63 & UINT8_C(0xff));
+  uint8_t x67 = (uint8_t)(x65 & UINT8_C(0xff));
+  uint32_t x68 = (0x0 + x5);
+  uint32_t x69 = (x68 >> 8);
+  uint8_t x70 = (uint8_t)(x68 & UINT8_C(0xff));
+  uint32_t x71 = (x69 >> 8);
   uint8_t x72 = (uint8_t)(x69 & UINT8_C(0xff));
-  fiat_p384_uint1 x73 = (fiat_p384_uint1)(x71 >> 8);
+  uint8_t x73 = (uint8_t)(x71 >> 8);
   uint8_t x74 = (uint8_t)(x71 & UINT8_C(0xff));
-  uint32_t x75 = (x73 + x5);
-  uint32_t x76 = (x75 >> 8);
-  uint8_t x77 = (uint8_t)(x75 & UINT8_C(0xff));
-  uint32_t x78 = (x76 >> 8);
-  uint8_t x79 = (uint8_t)(x76 & UINT8_C(0xff));
-  uint8_t x80 = (uint8_t)(x78 >> 8);
-  uint8_t x81 = (uint8_t)(x78 & UINT8_C(0xff));
-  fiat_p384_uint1 x82 = (fiat_p384_uint1)(x80 >> 8);
-  uint8_t x83 = (uint8_t)(x80 & UINT8_C(0xff));
-  uint32_t x84 = (x82 + x4);
+  uint8_t x75 = (uint8_t)(x73 & UINT8_C(0xff));
+  uint32_t x76 = (0x0 + x4);
+  uint32_t x77 = (x76 >> 8);
+  uint8_t x78 = (uint8_t)(x76 & UINT8_C(0xff));
+  uint32_t x79 = (x77 >> 8);
+  uint8_t x80 = (uint8_t)(x77 & UINT8_C(0xff));
+  uint8_t x81 = (uint8_t)(x79 >> 8);
+  uint8_t x82 = (uint8_t)(x79 & UINT8_C(0xff));
+  uint8_t x83 = (uint8_t)(x81 & UINT8_C(0xff));
+  uint32_t x84 = (0x0 + x3);
   uint32_t x85 = (x84 >> 8);
   uint8_t x86 = (uint8_t)(x84 & UINT8_C(0xff));
   uint32_t x87 = (x85 >> 8);
   uint8_t x88 = (uint8_t)(x85 & UINT8_C(0xff));
   uint8_t x89 = (uint8_t)(x87 >> 8);
   uint8_t x90 = (uint8_t)(x87 & UINT8_C(0xff));
-  fiat_p384_uint1 x91 = (fiat_p384_uint1)(x89 >> 8);
-  uint8_t x92 = (uint8_t)(x89 & UINT8_C(0xff));
-  uint32_t x93 = (x91 + x3);
-  uint32_t x94 = (x93 >> 8);
-  uint8_t x95 = (uint8_t)(x93 & UINT8_C(0xff));
-  uint32_t x96 = (x94 >> 8);
-  uint8_t x97 = (uint8_t)(x94 & UINT8_C(0xff));
-  uint8_t x98 = (uint8_t)(x96 >> 8);
-  uint8_t x99 = (uint8_t)(x96 & UINT8_C(0xff));
-  fiat_p384_uint1 x100 = (fiat_p384_uint1)(x98 >> 8);
-  uint8_t x101 = (uint8_t)(x98 & UINT8_C(0xff));
-  uint32_t x102 = (x100 + x2);
-  uint32_t x103 = (x102 >> 8);
-  uint8_t x104 = (uint8_t)(x102 & UINT8_C(0xff));
-  uint32_t x105 = (x103 >> 8);
+  uint8_t x91 = (uint8_t)(x89 & UINT8_C(0xff));
+  uint32_t x92 = (0x0 + x2);
+  uint32_t x93 = (x92 >> 8);
+  uint8_t x94 = (uint8_t)(x92 & UINT8_C(0xff));
+  uint32_t x95 = (x93 >> 8);
+  uint8_t x96 = (uint8_t)(x93 & UINT8_C(0xff));
+  uint8_t x97 = (uint8_t)(x95 >> 8);
+  uint8_t x98 = (uint8_t)(x95 & UINT8_C(0xff));
+  uint8_t x99 = (uint8_t)(x97 & UINT8_C(0xff));
+  uint32_t x100 = (0x0 + x1);
+  uint32_t x101 = (x100 >> 8);
+  uint8_t x102 = (uint8_t)(x100 & UINT8_C(0xff));
+  uint32_t x103 = (x101 >> 8);
+  uint8_t x104 = (uint8_t)(x101 & UINT8_C(0xff));
+  uint8_t x105 = (uint8_t)(x103 >> 8);
   uint8_t x106 = (uint8_t)(x103 & UINT8_C(0xff));
-  uint8_t x107 = (uint8_t)(x105 >> 8);
-  uint8_t x108 = (uint8_t)(x105 & UINT8_C(0xff));
-  fiat_p384_uint1 x109 = (fiat_p384_uint1)(x107 >> 8);
-  uint8_t x110 = (uint8_t)(x107 & UINT8_C(0xff));
-  uint32_t x111 = (x109 + x1);
-  uint32_t x112 = (x111 >> 8);
-  uint8_t x113 = (uint8_t)(x111 & UINT8_C(0xff));
-  uint32_t x114 = (x112 >> 8);
-  uint8_t x115 = (uint8_t)(x112 & UINT8_C(0xff));
-  uint8_t x116 = (uint8_t)(x114 >> 8);
-  uint8_t x117 = (uint8_t)(x114 & UINT8_C(0xff));
   out1[0] = x14;
   out1[1] = x16;
   out1[2] = x18;
-  out1[3] = x20;
-  out1[4] = x23;
-  out1[5] = x25;
-  out1[6] = x27;
-  out1[7] = x29;
-  out1[8] = x32;
-  out1[9] = x34;
-  out1[10] = x36;
-  out1[11] = x38;
-  out1[12] = x41;
-  out1[13] = x43;
-  out1[14] = x45;
-  out1[15] = x47;
-  out1[16] = x50;
-  out1[17] = x52;
-  out1[18] = x54;
-  out1[19] = x56;
-  out1[20] = x59;
-  out1[21] = x61;
-  out1[22] = x63;
-  out1[23] = x65;
-  out1[24] = x68;
-  out1[25] = x70;
-  out1[26] = x72;
-  out1[27] = x74;
-  out1[28] = x77;
-  out1[29] = x79;
-  out1[30] = x81;
-  out1[31] = x83;
-  out1[32] = x86;
-  out1[33] = x88;
-  out1[34] = x90;
-  out1[35] = x92;
-  out1[36] = x95;
-  out1[37] = x97;
-  out1[38] = x99;
-  out1[39] = x101;
-  out1[40] = x104;
-  out1[41] = x106;
-  out1[42] = x108;
-  out1[43] = x110;
-  out1[44] = x113;
-  out1[45] = x115;
-  out1[46] = x117;
-  out1[47] = x116;
+  out1[3] = x19;
+  out1[4] = x22;
+  out1[5] = x24;
+  out1[6] = x26;
+  out1[7] = x27;
+  out1[8] = x30;
+  out1[9] = x32;
+  out1[10] = x34;
+  out1[11] = x35;
+  out1[12] = x38;
+  out1[13] = x40;
+  out1[14] = x42;
+  out1[15] = x43;
+  out1[16] = x46;
+  out1[17] = x48;
+  out1[18] = x50;
+  out1[19] = x51;
+  out1[20] = x54;
+  out1[21] = x56;
+  out1[22] = x58;
+  out1[23] = x59;
+  out1[24] = x62;
+  out1[25] = x64;
+  out1[26] = x66;
+  out1[27] = x67;
+  out1[28] = x70;
+  out1[29] = x72;
+  out1[30] = x74;
+  out1[31] = x75;
+  out1[32] = x78;
+  out1[33] = x80;
+  out1[34] = x82;
+  out1[35] = x83;
+  out1[36] = x86;
+  out1[37] = x88;
+  out1[38] = x90;
+  out1[39] = x91;
+  out1[40] = x94;
+  out1[41] = x96;
+  out1[42] = x98;
+  out1[43] = x99;
+  out1[44] = x102;
+  out1[45] = x104;
+  out1[46] = x106;
+  out1[47] = x105;
 }
 
 /*
@@ -7505,61 +7458,50 @@ static void fiat_p384_from_bytes(uint32_t out1[12], const uint8_t arg1[48]) {
   uint32_t x47 = ((uint32_t)(arg1[1]) << 8);
   uint8_t x48 = (arg1[0]);
   uint32_t x49 = (x48 + (x47 + (x46 + x45)));
-  fiat_p384_uint1 x50 = (fiat_p384_uint1)((uint64_t)x49 >> 32);
-  uint32_t x51 = (x49 & UINT32_C(0xffffffff));
-  uint32_t x52 = (x4 + (x3 + (x2 + x1)));
-  uint32_t x53 = (x8 + (x7 + (x6 + x5)));
-  uint32_t x54 = (x12 + (x11 + (x10 + x9)));
-  uint32_t x55 = (x16 + (x15 + (x14 + x13)));
-  uint32_t x56 = (x20 + (x19 + (x18 + x17)));
-  uint32_t x57 = (x24 + (x23 + (x22 + x21)));
-  uint32_t x58 = (x28 + (x27 + (x26 + x25)));
-  uint32_t x59 = (x32 + (x31 + (x30 + x29)));
-  uint32_t x60 = (x36 + (x35 + (x34 + x33)));
-  uint32_t x61 = (x40 + (x39 + (x38 + x37)));
-  uint32_t x62 = (x44 + (x43 + (x42 + x41)));
-  uint32_t x63 = (x50 + x62);
-  fiat_p384_uint1 x64 = (fiat_p384_uint1)((uint64_t)x63 >> 32);
-  uint32_t x65 = (x63 & UINT32_C(0xffffffff));
-  uint32_t x66 = (x64 + x61);
-  fiat_p384_uint1 x67 = (fiat_p384_uint1)((uint64_t)x66 >> 32);
-  uint32_t x68 = (x66 & UINT32_C(0xffffffff));
-  uint32_t x69 = (x67 + x60);
-  fiat_p384_uint1 x70 = (fiat_p384_uint1)((uint64_t)x69 >> 32);
-  uint32_t x71 = (x69 & UINT32_C(0xffffffff));
-  uint32_t x72 = (x70 + x59);
-  fiat_p384_uint1 x73 = (fiat_p384_uint1)((uint64_t)x72 >> 32);
-  uint32_t x74 = (x72 & UINT32_C(0xffffffff));
-  uint32_t x75 = (x73 + x58);
-  fiat_p384_uint1 x76 = (fiat_p384_uint1)((uint64_t)x75 >> 32);
-  uint32_t x77 = (x75 & UINT32_C(0xffffffff));
-  uint32_t x78 = (x76 + x57);
-  fiat_p384_uint1 x79 = (fiat_p384_uint1)((uint64_t)x78 >> 32);
-  uint32_t x80 = (x78 & UINT32_C(0xffffffff));
-  uint32_t x81 = (x79 + x56);
-  fiat_p384_uint1 x82 = (fiat_p384_uint1)((uint64_t)x81 >> 32);
-  uint32_t x83 = (x81 & UINT32_C(0xffffffff));
-  uint32_t x84 = (x82 + x55);
-  fiat_p384_uint1 x85 = (fiat_p384_uint1)((uint64_t)x84 >> 32);
-  uint32_t x86 = (x84 & UINT32_C(0xffffffff));
-  uint32_t x87 = (x85 + x54);
-  fiat_p384_uint1 x88 = (fiat_p384_uint1)((uint64_t)x87 >> 32);
-  uint32_t x89 = (x87 & UINT32_C(0xffffffff));
-  uint32_t x90 = (x88 + x53);
-  fiat_p384_uint1 x91 = (fiat_p384_uint1)((uint64_t)x90 >> 32);
-  uint32_t x92 = (x90 & UINT32_C(0xffffffff));
-  uint32_t x93 = (x91 + x52);
-  out1[0] = x51;
-  out1[1] = x65;
-  out1[2] = x68;
-  out1[3] = x71;
-  out1[4] = x74;
-  out1[5] = x77;
-  out1[6] = x80;
-  out1[7] = x83;
-  out1[8] = x86;
-  out1[9] = x89;
-  out1[10] = x92;
-  out1[11] = x93;
+  uint32_t x50 = (x49 & UINT32_C(0xffffffff));
+  uint32_t x51 = (x4 + (x3 + (x2 + x1)));
+  uint32_t x52 = (x8 + (x7 + (x6 + x5)));
+  uint32_t x53 = (x12 + (x11 + (x10 + x9)));
+  uint32_t x54 = (x16 + (x15 + (x14 + x13)));
+  uint32_t x55 = (x20 + (x19 + (x18 + x17)));
+  uint32_t x56 = (x24 + (x23 + (x22 + x21)));
+  uint32_t x57 = (x28 + (x27 + (x26 + x25)));
+  uint32_t x58 = (x32 + (x31 + (x30 + x29)));
+  uint32_t x59 = (x36 + (x35 + (x34 + x33)));
+  uint32_t x60 = (x40 + (x39 + (x38 + x37)));
+  uint32_t x61 = (x44 + (x43 + (x42 + x41)));
+  uint32_t x62 = (0x0 + x61);
+  uint32_t x63 = (x62 & UINT32_C(0xffffffff));
+  uint32_t x64 = (0x0 + x60);
+  uint32_t x65 = (x64 & UINT32_C(0xffffffff));
+  uint32_t x66 = (0x0 + x59);
+  uint32_t x67 = (x66 & UINT32_C(0xffffffff));
+  uint32_t x68 = (0x0 + x58);
+  uint32_t x69 = (x68 & UINT32_C(0xffffffff));
+  uint32_t x70 = (0x0 + x57);
+  uint32_t x71 = (x70 & UINT32_C(0xffffffff));
+  uint32_t x72 = (0x0 + x56);
+  uint32_t x73 = (x72 & UINT32_C(0xffffffff));
+  uint32_t x74 = (0x0 + x55);
+  uint32_t x75 = (x74 & UINT32_C(0xffffffff));
+  uint32_t x76 = (0x0 + x54);
+  uint32_t x77 = (x76 & UINT32_C(0xffffffff));
+  uint32_t x78 = (0x0 + x53);
+  uint32_t x79 = (x78 & UINT32_C(0xffffffff));
+  uint32_t x80 = (0x0 + x52);
+  uint32_t x81 = (x80 & UINT32_C(0xffffffff));
+  uint32_t x82 = (0x0 + x51);
+  out1[0] = x50;
+  out1[1] = x63;
+  out1[2] = x65;
+  out1[3] = x67;
+  out1[4] = x69;
+  out1[5] = x71;
+  out1[6] = x73;
+  out1[7] = x75;
+  out1[8] = x77;
+  out1[9] = x79;
+  out1[10] = x81;
+  out1[11] = x82;
 }
 
diff --git a/p384_64.c b/p384_64.c
index 094f27a92..906996746 100644
--- a/p384_64.c
+++ b/p384_64.c
@@ -1851,441 +1851,441 @@ static void fiat_p384_from_montgomery(uint64_t out1[6], const uint64_t arg1[6])
   fiat_p384_uint1 x53;
   fiat_p384_addcarryx_u64(&x52, &x53, x51, 0x0, x40);
   uint64_t x54;
-  fiat_p384_uint1 x55;
-  fiat_p384_addcarryx_u64(&x54, &x55, x41, 0x0, 0x0);
+  uint64_t x55;
+  fiat_p384_mulx_u64(&x54, &x55, x42, UINT64_C(0x100000001));
   uint64_t x56;
-  fiat_p384_uint1 x57;
-  fiat_p384_addcarryx_u64(&x56, &x57, x53, 0x0, (fiat_p384_uint1)x54);
+  uint64_t x57;
+  fiat_p384_mulx_u64(&x56, &x57, x54, UINT64_C(0xffffffffffffffff));
   uint64_t x58;
   uint64_t x59;
-  fiat_p384_mulx_u64(&x58, &x59, x42, UINT64_C(0x100000001));
+  fiat_p384_mulx_u64(&x58, &x59, x54, UINT64_C(0xffffffffffffffff));
   uint64_t x60;
   uint64_t x61;
-  fiat_p384_mulx_u64(&x60, &x61, x58, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x60, &x61, x54, UINT64_C(0xffffffffffffffff));
   uint64_t x62;
   uint64_t x63;
-  fiat_p384_mulx_u64(&x62, &x63, x58, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x62, &x63, x54, UINT64_C(0xfffffffffffffffe));
   uint64_t x64;
   uint64_t x65;
-  fiat_p384_mulx_u64(&x64, &x65, x58, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x64, &x65, x54, UINT64_C(0xffffffff00000000));
   uint64_t x66;
   uint64_t x67;
-  fiat_p384_mulx_u64(&x66, &x67, x58, UINT64_C(0xfffffffffffffffe));
+  fiat_p384_mulx_u64(&x66, &x67, x54, UINT32_C(0xffffffff));
   uint64_t x68;
-  uint64_t x69;
-  fiat_p384_mulx_u64(&x68, &x69, x58, UINT64_C(0xffffffff00000000));
+  fiat_p384_uint1 x69;
+  fiat_p384_addcarryx_u64(&x68, &x69, 0x0, x64, x67);
   uint64_t x70;
-  uint64_t x71;
-  fiat_p384_mulx_u64(&x70, &x71, x58, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x71;
+  fiat_p384_addcarryx_u64(&x70, &x71, x69, x62, x65);
   uint64_t x72;
   fiat_p384_uint1 x73;
-  fiat_p384_addcarryx_u64(&x72, &x73, 0x0, x68, x71);
+  fiat_p384_addcarryx_u64(&x72, &x73, x71, x60, x63);
   uint64_t x74;
   fiat_p384_uint1 x75;
-  fiat_p384_addcarryx_u64(&x74, &x75, x73, x66, x69);
+  fiat_p384_addcarryx_u64(&x74, &x75, x73, x58, x61);
   uint64_t x76;
   fiat_p384_uint1 x77;
-  fiat_p384_addcarryx_u64(&x76, &x77, x75, x64, x67);
+  fiat_p384_addcarryx_u64(&x76, &x77, x75, x56, x59);
   uint64_t x78;
   fiat_p384_uint1 x79;
-  fiat_p384_addcarryx_u64(&x78, &x79, x77, x62, x65);
+  fiat_p384_addcarryx_u64(&x78, &x79, 0x0, x66, x42);
   uint64_t x80;
   fiat_p384_uint1 x81;
-  fiat_p384_addcarryx_u64(&x80, &x81, x79, x60, x63);
+  fiat_p384_addcarryx_u64(&x80, &x81, x79, x68, x44);
   uint64_t x82;
   fiat_p384_uint1 x83;
-  fiat_p384_addcarryx_u64(&x82, &x83, 0x0, x70, x42);
+  fiat_p384_addcarryx_u64(&x82, &x83, x81, x70, x46);
   uint64_t x84;
   fiat_p384_uint1 x85;
-  fiat_p384_addcarryx_u64(&x84, &x85, x83, x72, x44);
+  fiat_p384_addcarryx_u64(&x84, &x85, x83, x72, x48);
   uint64_t x86;
   fiat_p384_uint1 x87;
-  fiat_p384_addcarryx_u64(&x86, &x87, x85, x74, x46);
+  fiat_p384_addcarryx_u64(&x86, &x87, x85, x74, x50);
   uint64_t x88;
   fiat_p384_uint1 x89;
-  fiat_p384_addcarryx_u64(&x88, &x89, x87, x76, x48);
+  fiat_p384_addcarryx_u64(&x88, &x89, x87, x76, x52);
   uint64_t x90;
   fiat_p384_uint1 x91;
-  fiat_p384_addcarryx_u64(&x90, &x91, x89, x78, x50);
+  fiat_p384_addcarryx_u64(&x90, &x91, x41, 0x0, 0x0);
   uint64_t x92;
   fiat_p384_uint1 x93;
-  fiat_p384_addcarryx_u64(&x92, &x93, x91, x80, x52);
+  fiat_p384_addcarryx_u64(&x92, &x93, x53, 0x0, (fiat_p384_uint1)x90);
   uint64_t x94;
   fiat_p384_uint1 x95;
-  fiat_p384_addcarryx_u64(&x94, &x95, x81, 0x0, x61);
+  fiat_p384_addcarryx_u64(&x94, &x95, x77, 0x0, x57);
   uint64_t x96;
   fiat_p384_uint1 x97;
-  fiat_p384_addcarryx_u64(&x96, &x97, x93, x94, x56);
+  fiat_p384_addcarryx_u64(&x96, &x97, x89, x94, x92);
   uint64_t x98;
   fiat_p384_uint1 x99;
-  fiat_p384_addcarryx_u64(&x98, &x99, 0x0, (arg1[2]), x84);
+  fiat_p384_addcarryx_u64(&x98, &x99, 0x0, (arg1[2]), x80);
   uint64_t x100;
   fiat_p384_uint1 x101;
-  fiat_p384_addcarryx_u64(&x100, &x101, x99, 0x0, x86);
+  fiat_p384_addcarryx_u64(&x100, &x101, x99, 0x0, x82);
   uint64_t x102;
   fiat_p384_uint1 x103;
-  fiat_p384_addcarryx_u64(&x102, &x103, x101, 0x0, x88);
+  fiat_p384_addcarryx_u64(&x102, &x103, x101, 0x0, x84);
   uint64_t x104;
   fiat_p384_uint1 x105;
-  fiat_p384_addcarryx_u64(&x104, &x105, x103, 0x0, x90);
+  fiat_p384_addcarryx_u64(&x104, &x105, x103, 0x0, x86);
   uint64_t x106;
   fiat_p384_uint1 x107;
-  fiat_p384_addcarryx_u64(&x106, &x107, x105, 0x0, x92);
+  fiat_p384_addcarryx_u64(&x106, &x107, x105, 0x0, x88);
   uint64_t x108;
   fiat_p384_uint1 x109;
   fiat_p384_addcarryx_u64(&x108, &x109, x107, 0x0, x96);
   uint64_t x110;
-  fiat_p384_uint1 x111;
-  fiat_p384_addcarryx_u64(&x110, &x111, x97, 0x0, x57);
+  uint64_t x111;
+  fiat_p384_mulx_u64(&x110, &x111, x98, UINT64_C(0x100000001));
   uint64_t x112;
-  fiat_p384_uint1 x113;
-  fiat_p384_addcarryx_u64(&x112, &x113, x109, 0x0, (fiat_p384_uint1)x110);
+  uint64_t x113;
+  fiat_p384_mulx_u64(&x112, &x113, x110, UINT64_C(0xffffffffffffffff));
   uint64_t x114;
   uint64_t x115;
-  fiat_p384_mulx_u64(&x114, &x115, x98, UINT64_C(0x100000001));
+  fiat_p384_mulx_u64(&x114, &x115, x110, UINT64_C(0xffffffffffffffff));
   uint64_t x116;
   uint64_t x117;
-  fiat_p384_mulx_u64(&x116, &x117, x114, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x116, &x117, x110, UINT64_C(0xffffffffffffffff));
   uint64_t x118;
   uint64_t x119;
-  fiat_p384_mulx_u64(&x118, &x119, x114, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x118, &x119, x110, UINT64_C(0xfffffffffffffffe));
   uint64_t x120;
   uint64_t x121;
-  fiat_p384_mulx_u64(&x120, &x121, x114, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x120, &x121, x110, UINT64_C(0xffffffff00000000));
   uint64_t x122;
   uint64_t x123;
-  fiat_p384_mulx_u64(&x122, &x123, x114, UINT64_C(0xfffffffffffffffe));
+  fiat_p384_mulx_u64(&x122, &x123, x110, UINT32_C(0xffffffff));
   uint64_t x124;
-  uint64_t x125;
-  fiat_p384_mulx_u64(&x124, &x125, x114, UINT64_C(0xffffffff00000000));
+  fiat_p384_uint1 x125;
+  fiat_p384_addcarryx_u64(&x124, &x125, 0x0, x120, x123);
   uint64_t x126;
-  uint64_t x127;
-  fiat_p384_mulx_u64(&x126, &x127, x114, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x127;
+  fiat_p384_addcarryx_u64(&x126, &x127, x125, x118, x121);
   uint64_t x128;
   fiat_p384_uint1 x129;
-  fiat_p384_addcarryx_u64(&x128, &x129, 0x0, x124, x127);
+  fiat_p384_addcarryx_u64(&x128, &x129, x127, x116, x119);
   uint64_t x130;
   fiat_p384_uint1 x131;
-  fiat_p384_addcarryx_u64(&x130, &x131, x129, x122, x125);
+  fiat_p384_addcarryx_u64(&x130, &x131, x129, x114, x117);
   uint64_t x132;
   fiat_p384_uint1 x133;
-  fiat_p384_addcarryx_u64(&x132, &x133, x131, x120, x123);
+  fiat_p384_addcarryx_u64(&x132, &x133, x131, x112, x115);
   uint64_t x134;
   fiat_p384_uint1 x135;
-  fiat_p384_addcarryx_u64(&x134, &x135, x133, x118, x121);
+  fiat_p384_addcarryx_u64(&x134, &x135, 0x0, x122, x98);
   uint64_t x136;
   fiat_p384_uint1 x137;
-  fiat_p384_addcarryx_u64(&x136, &x137, x135, x116, x119);
+  fiat_p384_addcarryx_u64(&x136, &x137, x135, x124, x100);
   uint64_t x138;
   fiat_p384_uint1 x139;
-  fiat_p384_addcarryx_u64(&x138, &x139, 0x0, x126, x98);
+  fiat_p384_addcarryx_u64(&x138, &x139, x137, x126, x102);
   uint64_t x140;
   fiat_p384_uint1 x141;
-  fiat_p384_addcarryx_u64(&x140, &x141, x139, x128, x100);
+  fiat_p384_addcarryx_u64(&x140, &x141, x139, x128, x104);
   uint64_t x142;
   fiat_p384_uint1 x143;
-  fiat_p384_addcarryx_u64(&x142, &x143, x141, x130, x102);
+  fiat_p384_addcarryx_u64(&x142, &x143, x141, x130, x106);
   uint64_t x144;
   fiat_p384_uint1 x145;
-  fiat_p384_addcarryx_u64(&x144, &x145, x143, x132, x104);
+  fiat_p384_addcarryx_u64(&x144, &x145, x143, x132, x108);
   uint64_t x146;
   fiat_p384_uint1 x147;
-  fiat_p384_addcarryx_u64(&x146, &x147, x145, x134, x106);
+  fiat_p384_addcarryx_u64(&x146, &x147, x97, 0x0, 0x0);
   uint64_t x148;
   fiat_p384_uint1 x149;
-  fiat_p384_addcarryx_u64(&x148, &x149, x147, x136, x108);
+  fiat_p384_addcarryx_u64(&x148, &x149, x109, 0x0, (fiat_p384_uint1)x146);
   uint64_t x150;
   fiat_p384_uint1 x151;
-  fiat_p384_addcarryx_u64(&x150, &x151, x137, 0x0, x117);
+  fiat_p384_addcarryx_u64(&x150, &x151, x133, 0x0, x113);
   uint64_t x152;
   fiat_p384_uint1 x153;
-  fiat_p384_addcarryx_u64(&x152, &x153, x149, x150, x112);
+  fiat_p384_addcarryx_u64(&x152, &x153, x145, x150, x148);
   uint64_t x154;
   fiat_p384_uint1 x155;
-  fiat_p384_addcarryx_u64(&x154, &x155, 0x0, (arg1[3]), x140);
+  fiat_p384_addcarryx_u64(&x154, &x155, 0x0, (arg1[3]), x136);
   uint64_t x156;
   fiat_p384_uint1 x157;
-  fiat_p384_addcarryx_u64(&x156, &x157, x155, 0x0, x142);
+  fiat_p384_addcarryx_u64(&x156, &x157, x155, 0x0, x138);
   uint64_t x158;
   fiat_p384_uint1 x159;
-  fiat_p384_addcarryx_u64(&x158, &x159, x157, 0x0, x144);
+  fiat_p384_addcarryx_u64(&x158, &x159, x157, 0x0, x140);
   uint64_t x160;
   fiat_p384_uint1 x161;
-  fiat_p384_addcarryx_u64(&x160, &x161, x159, 0x0, x146);
+  fiat_p384_addcarryx_u64(&x160, &x161, x159, 0x0, x142);
   uint64_t x162;
   fiat_p384_uint1 x163;
-  fiat_p384_addcarryx_u64(&x162, &x163, x161, 0x0, x148);
+  fiat_p384_addcarryx_u64(&x162, &x163, x161, 0x0, x144);
   uint64_t x164;
   fiat_p384_uint1 x165;
   fiat_p384_addcarryx_u64(&x164, &x165, x163, 0x0, x152);
   uint64_t x166;
-  fiat_p384_uint1 x167;
-  fiat_p384_addcarryx_u64(&x166, &x167, x153, 0x0, x113);
+  uint64_t x167;
+  fiat_p384_mulx_u64(&x166, &x167, x154, UINT64_C(0x100000001));
   uint64_t x168;
-  fiat_p384_uint1 x169;
-  fiat_p384_addcarryx_u64(&x168, &x169, x165, 0x0, (fiat_p384_uint1)x166);
+  uint64_t x169;
+  fiat_p384_mulx_u64(&x168, &x169, x166, UINT64_C(0xffffffffffffffff));
   uint64_t x170;
   uint64_t x171;
-  fiat_p384_mulx_u64(&x170, &x171, x154, UINT64_C(0x100000001));
+  fiat_p384_mulx_u64(&x170, &x171, x166, UINT64_C(0xffffffffffffffff));
   uint64_t x172;
   uint64_t x173;
-  fiat_p384_mulx_u64(&x172, &x173, x170, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x172, &x173, x166, UINT64_C(0xffffffffffffffff));
   uint64_t x174;
   uint64_t x175;
-  fiat_p384_mulx_u64(&x174, &x175, x170, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x174, &x175, x166, UINT64_C(0xfffffffffffffffe));
   uint64_t x176;
   uint64_t x177;
-  fiat_p384_mulx_u64(&x176, &x177, x170, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x176, &x177, x166, UINT64_C(0xffffffff00000000));
   uint64_t x178;
   uint64_t x179;
-  fiat_p384_mulx_u64(&x178, &x179, x170, UINT64_C(0xfffffffffffffffe));
+  fiat_p384_mulx_u64(&x178, &x179, x166, UINT32_C(0xffffffff));
   uint64_t x180;
-  uint64_t x181;
-  fiat_p384_mulx_u64(&x180, &x181, x170, UINT64_C(0xffffffff00000000));
+  fiat_p384_uint1 x181;
+  fiat_p384_addcarryx_u64(&x180, &x181, 0x0, x176, x179);
   uint64_t x182;
-  uint64_t x183;
-  fiat_p384_mulx_u64(&x182, &x183, x170, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x183;
+  fiat_p384_addcarryx_u64(&x182, &x183, x181, x174, x177);
   uint64_t x184;
   fiat_p384_uint1 x185;
-  fiat_p384_addcarryx_u64(&x184, &x185, 0x0, x180, x183);
+  fiat_p384_addcarryx_u64(&x184, &x185, x183, x172, x175);
   uint64_t x186;
   fiat_p384_uint1 x187;
-  fiat_p384_addcarryx_u64(&x186, &x187, x185, x178, x181);
+  fiat_p384_addcarryx_u64(&x186, &x187, x185, x170, x173);
   uint64_t x188;
   fiat_p384_uint1 x189;
-  fiat_p384_addcarryx_u64(&x188, &x189, x187, x176, x179);
+  fiat_p384_addcarryx_u64(&x188, &x189, x187, x168, x171);
   uint64_t x190;
   fiat_p384_uint1 x191;
-  fiat_p384_addcarryx_u64(&x190, &x191, x189, x174, x177);
+  fiat_p384_addcarryx_u64(&x190, &x191, 0x0, x178, x154);
   uint64_t x192;
   fiat_p384_uint1 x193;
-  fiat_p384_addcarryx_u64(&x192, &x193, x191, x172, x175);
+  fiat_p384_addcarryx_u64(&x192, &x193, x191, x180, x156);
   uint64_t x194;
   fiat_p384_uint1 x195;
-  fiat_p384_addcarryx_u64(&x194, &x195, 0x0, x182, x154);
+  fiat_p384_addcarryx_u64(&x194, &x195, x193, x182, x158);
   uint64_t x196;
   fiat_p384_uint1 x197;
-  fiat_p384_addcarryx_u64(&x196, &x197, x195, x184, x156);
+  fiat_p384_addcarryx_u64(&x196, &x197, x195, x184, x160);
   uint64_t x198;
   fiat_p384_uint1 x199;
-  fiat_p384_addcarryx_u64(&x198, &x199, x197, x186, x158);
+  fiat_p384_addcarryx_u64(&x198, &x199, x197, x186, x162);
   uint64_t x200;
   fiat_p384_uint1 x201;
-  fiat_p384_addcarryx_u64(&x200, &x201, x199, x188, x160);
+  fiat_p384_addcarryx_u64(&x200, &x201, x199, x188, x164);
   uint64_t x202;
   fiat_p384_uint1 x203;
-  fiat_p384_addcarryx_u64(&x202, &x203, x201, x190, x162);
+  fiat_p384_addcarryx_u64(&x202, &x203, x153, 0x0, 0x0);
   uint64_t x204;
   fiat_p384_uint1 x205;
-  fiat_p384_addcarryx_u64(&x204, &x205, x203, x192, x164);
+  fiat_p384_addcarryx_u64(&x204, &x205, x165, 0x0, (fiat_p384_uint1)x202);
   uint64_t x206;
   fiat_p384_uint1 x207;
-  fiat_p384_addcarryx_u64(&x206, &x207, x193, 0x0, x173);
+  fiat_p384_addcarryx_u64(&x206, &x207, x189, 0x0, x169);
   uint64_t x208;
   fiat_p384_uint1 x209;
-  fiat_p384_addcarryx_u64(&x208, &x209, x205, x206, x168);
+  fiat_p384_addcarryx_u64(&x208, &x209, x201, x206, x204);
   uint64_t x210;
   fiat_p384_uint1 x211;
-  fiat_p384_addcarryx_u64(&x210, &x211, 0x0, (arg1[4]), x196);
+  fiat_p384_addcarryx_u64(&x210, &x211, 0x0, (arg1[4]), x192);
   uint64_t x212;
   fiat_p384_uint1 x213;
-  fiat_p384_addcarryx_u64(&x212, &x213, x211, 0x0, x198);
+  fiat_p384_addcarryx_u64(&x212, &x213, x211, 0x0, x194);
   uint64_t x214;
   fiat_p384_uint1 x215;
-  fiat_p384_addcarryx_u64(&x214, &x215, x213, 0x0, x200);
+  fiat_p384_addcarryx_u64(&x214, &x215, x213, 0x0, x196);
   uint64_t x216;
   fiat_p384_uint1 x217;
-  fiat_p384_addcarryx_u64(&x216, &x217, x215, 0x0, x202);
+  fiat_p384_addcarryx_u64(&x216, &x217, x215, 0x0, x198);
   uint64_t x218;
   fiat_p384_uint1 x219;
-  fiat_p384_addcarryx_u64(&x218, &x219, x217, 0x0, x204);
+  fiat_p384_addcarryx_u64(&x218, &x219, x217, 0x0, x200);
   uint64_t x220;
   fiat_p384_uint1 x221;
   fiat_p384_addcarryx_u64(&x220, &x221, x219, 0x0, x208);
   uint64_t x222;
-  fiat_p384_uint1 x223;
-  fiat_p384_addcarryx_u64(&x222, &x223, x209, 0x0, x169);
+  uint64_t x223;
+  fiat_p384_mulx_u64(&x222, &x223, x210, UINT64_C(0x100000001));
   uint64_t x224;
-  fiat_p384_uint1 x225;
-  fiat_p384_addcarryx_u64(&x224, &x225, x221, 0x0, (fiat_p384_uint1)x222);
+  uint64_t x225;
+  fiat_p384_mulx_u64(&x224, &x225, x222, UINT64_C(0xffffffffffffffff));
   uint64_t x226;
   uint64_t x227;
-  fiat_p384_mulx_u64(&x226, &x227, x210, UINT64_C(0x100000001));
+  fiat_p384_mulx_u64(&x226, &x227, x222, UINT64_C(0xffffffffffffffff));
   uint64_t x228;
   uint64_t x229;
-  fiat_p384_mulx_u64(&x228, &x229, x226, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x228, &x229, x222, UINT64_C(0xffffffffffffffff));
   uint64_t x230;
   uint64_t x231;
-  fiat_p384_mulx_u64(&x230, &x231, x226, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x230, &x231, x222, UINT64_C(0xfffffffffffffffe));
   uint64_t x232;
   uint64_t x233;
-  fiat_p384_mulx_u64(&x232, &x233, x226, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x232, &x233, x222, UINT64_C(0xffffffff00000000));
   uint64_t x234;
   uint64_t x235;
-  fiat_p384_mulx_u64(&x234, &x235, x226, UINT64_C(0xfffffffffffffffe));
+  fiat_p384_mulx_u64(&x234, &x235, x222, UINT32_C(0xffffffff));
   uint64_t x236;
-  uint64_t x237;
-  fiat_p384_mulx_u64(&x236, &x237, x226, UINT64_C(0xffffffff00000000));
+  fiat_p384_uint1 x237;
+  fiat_p384_addcarryx_u64(&x236, &x237, 0x0, x232, x235);
   uint64_t x238;
-  uint64_t x239;
-  fiat_p384_mulx_u64(&x238, &x239, x226, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x239;
+  fiat_p384_addcarryx_u64(&x238, &x239, x237, x230, x233);
   uint64_t x240;
   fiat_p384_uint1 x241;
-  fiat_p384_addcarryx_u64(&x240, &x241, 0x0, x236, x239);
+  fiat_p384_addcarryx_u64(&x240, &x241, x239, x228, x231);
   uint64_t x242;
   fiat_p384_uint1 x243;
-  fiat_p384_addcarryx_u64(&x242, &x243, x241, x234, x237);
+  fiat_p384_addcarryx_u64(&x242, &x243, x241, x226, x229);
   uint64_t x244;
   fiat_p384_uint1 x245;
-  fiat_p384_addcarryx_u64(&x244, &x245, x243, x232, x235);
+  fiat_p384_addcarryx_u64(&x244, &x245, x243, x224, x227);
   uint64_t x246;
   fiat_p384_uint1 x247;
-  fiat_p384_addcarryx_u64(&x246, &x247, x245, x230, x233);
+  fiat_p384_addcarryx_u64(&x246, &x247, 0x0, x234, x210);
   uint64_t x248;
   fiat_p384_uint1 x249;
-  fiat_p384_addcarryx_u64(&x248, &x249, x247, x228, x231);
+  fiat_p384_addcarryx_u64(&x248, &x249, x247, x236, x212);
   uint64_t x250;
   fiat_p384_uint1 x251;
-  fiat_p384_addcarryx_u64(&x250, &x251, 0x0, x238, x210);
+  fiat_p384_addcarryx_u64(&x250, &x251, x249, x238, x214);
   uint64_t x252;
   fiat_p384_uint1 x253;
-  fiat_p384_addcarryx_u64(&x252, &x253, x251, x240, x212);
+  fiat_p384_addcarryx_u64(&x252, &x253, x251, x240, x216);
   uint64_t x254;
   fiat_p384_uint1 x255;
-  fiat_p384_addcarryx_u64(&x254, &x255, x253, x242, x214);
+  fiat_p384_addcarryx_u64(&x254, &x255, x253, x242, x218);
   uint64_t x256;
   fiat_p384_uint1 x257;
-  fiat_p384_addcarryx_u64(&x256, &x257, x255, x244, x216);
+  fiat_p384_addcarryx_u64(&x256, &x257, x255, x244, x220);
   uint64_t x258;
   fiat_p384_uint1 x259;
-  fiat_p384_addcarryx_u64(&x258, &x259, x257, x246, x218);
+  fiat_p384_addcarryx_u64(&x258, &x259, x209, 0x0, 0x0);
   uint64_t x260;
   fiat_p384_uint1 x261;
-  fiat_p384_addcarryx_u64(&x260, &x261, x259, x248, x220);
+  fiat_p384_addcarryx_u64(&x260, &x261, x221, 0x0, (fiat_p384_uint1)x258);
   uint64_t x262;
   fiat_p384_uint1 x263;
-  fiat_p384_addcarryx_u64(&x262, &x263, x249, 0x0, x229);
+  fiat_p384_addcarryx_u64(&x262, &x263, x245, 0x0, x225);
   uint64_t x264;
   fiat_p384_uint1 x265;
-  fiat_p384_addcarryx_u64(&x264, &x265, x261, x262, x224);
+  fiat_p384_addcarryx_u64(&x264, &x265, x257, x262, x260);
   uint64_t x266;
   fiat_p384_uint1 x267;
-  fiat_p384_addcarryx_u64(&x266, &x267, 0x0, (arg1[5]), x252);
+  fiat_p384_addcarryx_u64(&x266, &x267, 0x0, (arg1[5]), x248);
   uint64_t x268;
   fiat_p384_uint1 x269;
-  fiat_p384_addcarryx_u64(&x268, &x269, x267, 0x0, x254);
+  fiat_p384_addcarryx_u64(&x268, &x269, x267, 0x0, x250);
   uint64_t x270;
   fiat_p384_uint1 x271;
-  fiat_p384_addcarryx_u64(&x270, &x271, x269, 0x0, x256);
+  fiat_p384_addcarryx_u64(&x270, &x271, x269, 0x0, x252);
   uint64_t x272;
   fiat_p384_uint1 x273;
-  fiat_p384_addcarryx_u64(&x272, &x273, x271, 0x0, x258);
+  fiat_p384_addcarryx_u64(&x272, &x273, x271, 0x0, x254);
   uint64_t x274;
   fiat_p384_uint1 x275;
-  fiat_p384_addcarryx_u64(&x274, &x275, x273, 0x0, x260);
+  fiat_p384_addcarryx_u64(&x274, &x275, x273, 0x0, x256);
   uint64_t x276;
   fiat_p384_uint1 x277;
   fiat_p384_addcarryx_u64(&x276, &x277, x275, 0x0, x264);
   uint64_t x278;
-  fiat_p384_uint1 x279;
-  fiat_p384_addcarryx_u64(&x278, &x279, x265, 0x0, x225);
+  uint64_t x279;
+  fiat_p384_mulx_u64(&x278, &x279, x266, UINT64_C(0x100000001));
   uint64_t x280;
-  fiat_p384_uint1 x281;
-  fiat_p384_addcarryx_u64(&x280, &x281, x277, 0x0, (fiat_p384_uint1)x278);
+  uint64_t x281;
+  fiat_p384_mulx_u64(&x280, &x281, x278, UINT64_C(0xffffffffffffffff));
   uint64_t x282;
   uint64_t x283;
-  fiat_p384_mulx_u64(&x282, &x283, x266, UINT64_C(0x100000001));
+  fiat_p384_mulx_u64(&x282, &x283, x278, UINT64_C(0xffffffffffffffff));
   uint64_t x284;
   uint64_t x285;
-  fiat_p384_mulx_u64(&x284, &x285, x282, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x284, &x285, x278, UINT64_C(0xffffffffffffffff));
   uint64_t x286;
   uint64_t x287;
-  fiat_p384_mulx_u64(&x286, &x287, x282, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x286, &x287, x278, UINT64_C(0xfffffffffffffffe));
   uint64_t x288;
   uint64_t x289;
-  fiat_p384_mulx_u64(&x288, &x289, x282, UINT64_C(0xffffffffffffffff));
+  fiat_p384_mulx_u64(&x288, &x289, x278, UINT64_C(0xffffffff00000000));
   uint64_t x290;
   uint64_t x291;
-  fiat_p384_mulx_u64(&x290, &x291, x282, UINT64_C(0xfffffffffffffffe));
+  fiat_p384_mulx_u64(&x290, &x291, x278, UINT32_C(0xffffffff));
   uint64_t x292;
-  uint64_t x293;
-  fiat_p384_mulx_u64(&x292, &x293, x282, UINT64_C(0xffffffff00000000));
+  fiat_p384_uint1 x293;
+  fiat_p384_addcarryx_u64(&x292, &x293, 0x0, x288, x291);
   uint64_t x294;
-  uint64_t x295;
-  fiat_p384_mulx_u64(&x294, &x295, x282, UINT32_C(0xffffffff));
+  fiat_p384_uint1 x295;
+  fiat_p384_addcarryx_u64(&x294, &x295, x293, x286, x289);
   uint64_t x296;
   fiat_p384_uint1 x297;
-  fiat_p384_addcarryx_u64(&x296, &x297, 0x0, x292, x295);
+  fiat_p384_addcarryx_u64(&x296, &x297, x295, x284, x287);
   uint64_t x298;
   fiat_p384_uint1 x299;
-  fiat_p384_addcarryx_u64(&x298, &x299, x297, x290, x293);
+  fiat_p384_addcarryx_u64(&x298, &x299, x297, x282, x285);
   uint64_t x300;
   fiat_p384_uint1 x301;
-  fiat_p384_addcarryx_u64(&x300, &x301, x299, x288, x291);
+  fiat_p384_addcarryx_u64(&x300, &x301, x299, x280, x283);
   uint64_t x302;
   fiat_p384_uint1 x303;
-  fiat_p384_addcarryx_u64(&x302, &x303, x301, x286, x289);
+  fiat_p384_addcarryx_u64(&x302, &x303, 0x0, x290, x266);
   uint64_t x304;
   fiat_p384_uint1 x305;
-  fiat_p384_addcarryx_u64(&x304, &x305, x303, x284, x287);
+  fiat_p384_addcarryx_u64(&x304, &x305, x303, x292, x268);
   uint64_t x306;
   fiat_p384_uint1 x307;
-  fiat_p384_addcarryx_u64(&x306, &x307, 0x0, x294, x266);
+  fiat_p384_addcarryx_u64(&x306, &x307, x305, x294, x270);
   uint64_t x308;
   fiat_p384_uint1 x309;
-  fiat_p384_addcarryx_u64(&x308, &x309, x307, x296, x268);
+  fiat_p384_addcarryx_u64(&x308, &x309, x307, x296, x272);
   uint64_t x310;
   fiat_p384_uint1 x311;
-  fiat_p384_addcarryx_u64(&x310, &x311, x309, x298, x270);
+  fiat_p384_addcarryx_u64(&x310, &x311, x309, x298, x274);
   uint64_t x312;
   fiat_p384_uint1 x313;
-  fiat_p384_addcarryx_u64(&x312, &x313, x311, x300, x272);
+  fiat_p384_addcarryx_u64(&x312, &x313, x311, x300, x276);
   uint64_t x314;
   fiat_p384_uint1 x315;
-  fiat_p384_addcarryx_u64(&x314, &x315, x313, x302, x274);
+  fiat_p384_addcarryx_u64(&x314, &x315, x265, 0x0, 0x0);
   uint64_t x316;
   fiat_p384_uint1 x317;
-  fiat_p384_addcarryx_u64(&x316, &x317, x315, x304, x276);
+  fiat_p384_addcarryx_u64(&x316, &x317, x277, 0x0, (fiat_p384_uint1)x314);
   uint64_t x318;
   fiat_p384_uint1 x319;
-  fiat_p384_addcarryx_u64(&x318, &x319, x305, 0x0, x285);
+  fiat_p384_addcarryx_u64(&x318, &x319, x301, 0x0, x281);
   uint64_t x320;
   fiat_p384_uint1 x321;
-  fiat_p384_addcarryx_u64(&x320, &x321, x317, x318, x280);
+  fiat_p384_addcarryx_u64(&x320, &x321, x313, x318, x316);
   uint64_t x322;
   fiat_p384_uint1 x323;
-  fiat_p384_subborrowx_u64(&x322, &x323, 0x0, x308, UINT32_C(0xffffffff));
+  fiat_p384_subborrowx_u64(&x322, &x323, 0x0, x304, UINT32_C(0xffffffff));
   uint64_t x324;
   fiat_p384_uint1 x325;
-  fiat_p384_subborrowx_u64(&x324, &x325, x323, x310, UINT64_C(0xffffffff00000000));
+  fiat_p384_subborrowx_u64(&x324, &x325, x323, x306, UINT64_C(0xffffffff00000000));
   uint64_t x326;
   fiat_p384_uint1 x327;
-  fiat_p384_subborrowx_u64(&x326, &x327, x325, x312, UINT64_C(0xfffffffffffffffe));
+  fiat_p384_subborrowx_u64(&x326, &x327, x325, x308, UINT64_C(0xfffffffffffffffe));
   uint64_t x328;
   fiat_p384_uint1 x329;
-  fiat_p384_subborrowx_u64(&x328, &x329, x327, x314, UINT64_C(0xffffffffffffffff));
+  fiat_p384_subborrowx_u64(&x328, &x329, x327, x310, UINT64_C(0xffffffffffffffff));
   uint64_t x330;
   fiat_p384_uint1 x331;
-  fiat_p384_subborrowx_u64(&x330, &x331, x329, x316, UINT64_C(0xffffffffffffffff));
+  fiat_p384_subborrowx_u64(&x330, &x331, x329, x312, UINT64_C(0xffffffffffffffff));
   uint64_t x332;
   fiat_p384_uint1 x333;
   fiat_p384_subborrowx_u64(&x332, &x333, x331, x320, UINT64_C(0xffffffffffffffff));
   uint64_t x334;
   fiat_p384_uint1 x335;
-  fiat_p384_addcarryx_u64(&x334, &x335, x321, 0x0, x281);
+  fiat_p384_addcarryx_u64(&x334, &x335, x321, 0x0, 0x0);
   uint64_t x336;
   fiat_p384_uint1 x337;
   fiat_p384_subborrowx_u64(&x336, &x337, x333, (fiat_p384_uint1)x334, 0x0);
   uint64_t x338;
-  fiat_p384_cmovznz_u64(&x338, x337, x322, x308);
+  fiat_p384_cmovznz_u64(&x338, x337, x322, x304);
   uint64_t x339;
-  fiat_p384_cmovznz_u64(&x339, x337, x324, x310);
+  fiat_p384_cmovznz_u64(&x339, x337, x324, x306);
   uint64_t x340;
-  fiat_p384_cmovznz_u64(&x340, x337, x326, x312);
+  fiat_p384_cmovznz_u64(&x340, x337, x326, x308);
   uint64_t x341;
-  fiat_p384_cmovznz_u64(&x341, x337, x328, x314);
+  fiat_p384_cmovznz_u64(&x341, x337, x328, x310);
   uint64_t x342;
-  fiat_p384_cmovznz_u64(&x342, x337, x330, x316);
+  fiat_p384_cmovznz_u64(&x342, x337, x330, x312);
   uint64_t x343;
   fiat_p384_cmovznz_u64(&x343, x337, x332, x320);
   out1[0] = x338;
@@ -2363,28 +2363,28 @@ static void fiat_p384_to_bytes(uint8_t out1[48], const uint64_t arg1[6]) {
   uint8_t x18 = (uint8_t)(x15 & UINT8_C(0xff));
   uint8_t x19 = (uint8_t)(x17 >> 8);
   uint8_t x20 = (uint8_t)(x17 & UINT8_C(0xff));
-  fiat_p384_uint1 x21 = (fiat_p384_uint1)(x19 >> 8);
-  uint8_t x22 = (uint8_t)(x19 & UINT8_C(0xff));
-  uint64_t x23 = (x21 + x5);
-  uint64_t x24 = (x23 >> 8);
-  uint8_t x25 = (uint8_t)(x23 & UINT8_C(0xff));
-  uint64_t x26 = (x24 >> 8);
-  uint8_t x27 = (uint8_t)(x24 & UINT8_C(0xff));
-  uint64_t x28 = (x26 >> 8);
-  uint8_t x29 = (uint8_t)(x26 & UINT8_C(0xff));
-  uint64_t x30 = (x28 >> 8);
-  uint8_t x31 = (uint8_t)(x28 & UINT8_C(0xff));
-  uint64_t x32 = (x30 >> 8);
-  uint8_t x33 = (uint8_t)(x30 & UINT8_C(0xff));
-  uint64_t x34 = (x32 >> 8);
-  uint8_t x35 = (uint8_t)(x32 & UINT8_C(0xff));
-  uint8_t x36 = (uint8_t)(x34 >> 8);
-  uint8_t x37 = (uint8_t)(x34 & UINT8_C(0xff));
-  fiat_p384_uint1 x38 = (fiat_p384_uint1)(x36 >> 8);
-  uint8_t x39 = (uint8_t)(x36 & UINT8_C(0xff));
-  uint64_t x40 = (x38 + x4);
-  uint64_t x41 = (x40 >> 8);
-  uint8_t x42 = (uint8_t)(x40 & UINT8_C(0xff));
+  uint8_t x21 = (uint8_t)(x19 & UINT8_C(0xff));
+  uint64_t x22 = (0x0 + x5);
+  uint64_t x23 = (x22 >> 8);
+  uint8_t x24 = (uint8_t)(x22 & UINT8_C(0xff));
+  uint64_t x25 = (x23 >> 8);
+  uint8_t x26 = (uint8_t)(x23 & UINT8_C(0xff));
+  uint64_t x27 = (x25 >> 8);
+  uint8_t x28 = (uint8_t)(x25 & UINT8_C(0xff));
+  uint64_t x29 = (x27 >> 8);
+  uint8_t x30 = (uint8_t)(x27 & UINT8_C(0xff));
+  uint64_t x31 = (x29 >> 8);
+  uint8_t x32 = (uint8_t)(x29 & UINT8_C(0xff));
+  uint64_t x33 = (x31 >> 8);
+  uint8_t x34 = (uint8_t)(x31 & UINT8_C(0xff));
+  uint8_t x35 = (uint8_t)(x33 >> 8);
+  uint8_t x36 = (uint8_t)(x33 & UINT8_C(0xff));
+  uint8_t x37 = (uint8_t)(x35 & UINT8_C(0xff));
+  uint64_t x38 = (0x0 + x4);
+  uint64_t x39 = (x38 >> 8);
+  uint8_t x40 = (uint8_t)(x38 & UINT8_C(0xff));
+  uint64_t x41 = (x39 >> 8);
+  uint8_t x42 = (uint8_t)(x39 & UINT8_C(0xff));
   uint64_t x43 = (x41 >> 8);
   uint8_t x44 = (uint8_t)(x41 & UINT8_C(0xff));
   uint64_t x45 = (x43 >> 8);
@@ -2393,61 +2393,56 @@ static void fiat_p384_to_bytes(uint8_t out1[48], const uint64_t arg1[6]) {
   uint8_t x48 = (uint8_t)(x45 & UINT8_C(0xff));
   uint64_t x49 = (x47 >> 8);
   uint8_t x50 = (uint8_t)(x47 & UINT8_C(0xff));
-  uint64_t x51 = (x49 >> 8);
+  uint8_t x51 = (uint8_t)(x49 >> 8);
   uint8_t x52 = (uint8_t)(x49 & UINT8_C(0xff));
-  uint8_t x53 = (uint8_t)(x51 >> 8);
-  uint8_t x54 = (uint8_t)(x51 & UINT8_C(0xff));
-  fiat_p384_uint1 x55 = (fiat_p384_uint1)(x53 >> 8);
-  uint8_t x56 = (uint8_t)(x53 & UINT8_C(0xff));
-  uint64_t x57 = (x55 + x3);
-  uint64_t x58 = (x57 >> 8);
-  uint8_t x59 = (uint8_t)(x57 & UINT8_C(0xff));
-  uint64_t x60 = (x58 >> 8);
-  uint8_t x61 = (uint8_t)(x58 & UINT8_C(0xff));
-  uint64_t x62 = (x60 >> 8);
-  uint8_t x63 = (uint8_t)(x60 & UINT8_C(0xff));
-  uint64_t x64 = (x62 >> 8);
-  uint8_t x65 = (uint8_t)(x62 & UINT8_C(0xff));
-  uint64_t x66 = (x64 >> 8);
-  uint8_t x67 = (uint8_t)(x64 & UINT8_C(0xff));
-  uint64_t x68 = (x66 >> 8);
-  uint8_t x69 = (uint8_t)(x66 & UINT8_C(0xff));
-  uint8_t x70 = (uint8_t)(x68 >> 8);
-  uint8_t x71 = (uint8_t)(x68 & UINT8_C(0xff));
-  fiat_p384_uint1 x72 = (fiat_p384_uint1)(x70 >> 8);
-  uint8_t x73 = (uint8_t)(x70 & UINT8_C(0xff));
-  uint64_t x74 = (x72 + x2);
-  uint64_t x75 = (x74 >> 8);
-  uint8_t x76 = (uint8_t)(x74 & UINT8_C(0xff));
+  uint8_t x53 = (uint8_t)(x51 & UINT8_C(0xff));
+  uint64_t x54 = (0x0 + x3);
+  uint64_t x55 = (x54 >> 8);
+  uint8_t x56 = (uint8_t)(x54 & UINT8_C(0xff));
+  uint64_t x57 = (x55 >> 8);
+  uint8_t x58 = (uint8_t)(x55 & UINT8_C(0xff));
+  uint64_t x59 = (x57 >> 8);
+  uint8_t x60 = (uint8_t)(x57 & UINT8_C(0xff));
+  uint64_t x61 = (x59 >> 8);
+  uint8_t x62 = (uint8_t)(x59 & UINT8_C(0xff));
+  uint64_t x63 = (x61 >> 8);
+  uint8_t x64 = (uint8_t)(x61 & UINT8_C(0xff));
+  uint64_t x65 = (x63 >> 8);
+  uint8_t x66 = (uint8_t)(x63 & UINT8_C(0xff));
+  uint8_t x67 = (uint8_t)(x65 >> 8);
+  uint8_t x68 = (uint8_t)(x65 & UINT8_C(0xff));
+  uint8_t x69 = (uint8_t)(x67 & UINT8_C(0xff));
+  uint64_t x70 = (0x0 + x2);
+  uint64_t x71 = (x70 >> 8);
+  uint8_t x72 = (uint8_t)(x70 & UINT8_C(0xff));
+  uint64_t x73 = (x71 >> 8);
+  uint8_t x74 = (uint8_t)(x71 & UINT8_C(0xff));
+  uint64_t x75 = (x73 >> 8);
+  uint8_t x76 = (uint8_t)(x73 & UINT8_C(0xff));
   uint64_t x77 = (x75 >> 8);
   uint8_t x78 = (uint8_t)(x75 & UINT8_C(0xff));
   uint64_t x79 = (x77 >> 8);
   uint8_t x80 = (uint8_t)(x77 & UINT8_C(0xff));
   uint64_t x81 = (x79 >> 8);
   uint8_t x82 = (uint8_t)(x79 & UINT8_C(0xff));
-  uint64_t x83 = (x81 >> 8);
+  uint8_t x83 = (uint8_t)(x81 >> 8);
   uint8_t x84 = (uint8_t)(x81 & UINT8_C(0xff));
-  uint64_t x85 = (x83 >> 8);
-  uint8_t x86 = (uint8_t)(x83 & UINT8_C(0xff));
-  uint8_t x87 = (uint8_t)(x85 >> 8);
-  uint8_t x88 = (uint8_t)(x85 & UINT8_C(0xff));
-  fiat_p384_uint1 x89 = (fiat_p384_uint1)(x87 >> 8);
+  uint8_t x85 = (uint8_t)(x83 & UINT8_C(0xff));
+  uint64_t x86 = (0x0 + x1);
+  uint64_t x87 = (x86 >> 8);
+  uint8_t x88 = (uint8_t)(x86 & UINT8_C(0xff));
+  uint64_t x89 = (x87 >> 8);
   uint8_t x90 = (uint8_t)(x87 & UINT8_C(0xff));
-  uint64_t x91 = (x89 + x1);
-  uint64_t x92 = (x91 >> 8);
-  uint8_t x93 = (uint8_t)(x91 & UINT8_C(0xff));
-  uint64_t x94 = (x92 >> 8);
-  uint8_t x95 = (uint8_t)(x92 & UINT8_C(0xff));
-  uint64_t x96 = (x94 >> 8);
-  uint8_t x97 = (uint8_t)(x94 & UINT8_C(0xff));
-  uint64_t x98 = (x96 >> 8);
-  uint8_t x99 = (uint8_t)(x96 & UINT8_C(0xff));
-  uint64_t x100 = (x98 >> 8);
-  uint8_t x101 = (uint8_t)(x98 & UINT8_C(0xff));
-  uint64_t x102 = (x100 >> 8);
-  uint8_t x103 = (uint8_t)(x100 & UINT8_C(0xff));
-  uint8_t x104 = (uint8_t)(x102 >> 8);
-  uint8_t x105 = (uint8_t)(x102 & UINT8_C(0xff));
+  uint64_t x91 = (x89 >> 8);
+  uint8_t x92 = (uint8_t)(x89 & UINT8_C(0xff));
+  uint64_t x93 = (x91 >> 8);
+  uint8_t x94 = (uint8_t)(x91 & UINT8_C(0xff));
+  uint64_t x95 = (x93 >> 8);
+  uint8_t x96 = (uint8_t)(x93 & UINT8_C(0xff));
+  uint64_t x97 = (x95 >> 8);
+  uint8_t x98 = (uint8_t)(x95 & UINT8_C(0xff));
+  uint8_t x99 = (uint8_t)(x97 >> 8);
+  uint8_t x100 = (uint8_t)(x97 & UINT8_C(0xff));
   out1[0] = x8;
   out1[1] = x10;
   out1[2] = x12;
@@ -2455,47 +2450,47 @@ static void fiat_p384_to_bytes(uint8_t out1[48], const uint64_t arg1[6]) {
   out1[4] = x16;
   out1[5] = x18;
   out1[6] = x20;
-  out1[7] = x22;
-  out1[8] = x25;
-  out1[9] = x27;
-  out1[10] = x29;
-  out1[11] = x31;
-  out1[12] = x33;
-  out1[13] = x35;
-  out1[14] = x37;
-  out1[15] = x39;
-  out1[16] = x42;
-  out1[17] = x44;
-  out1[18] = x46;
-  out1[19] = x48;
-  out1[20] = x50;
-  out1[21] = x52;
-  out1[22] = x54;
-  out1[23] = x56;
-  out1[24] = x59;
-  out1[25] = x61;
-  out1[26] = x63;
-  out1[27] = x65;
-  out1[28] = x67;
-  out1[29] = x69;
-  out1[30] = x71;
-  out1[31] = x73;
-  out1[32] = x76;
-  out1[33] = x78;
-  out1[34] = x80;
-  out1[35] = x82;
-  out1[36] = x84;
-  out1[37] = x86;
-  out1[38] = x88;
-  out1[39] = x90;
-  out1[40] = x93;
-  out1[41] = x95;
-  out1[42] = x97;
-  out1[43] = x99;
-  out1[44] = x101;
-  out1[45] = x103;
-  out1[46] = x105;
-  out1[47] = x104;
+  out1[7] = x21;
+  out1[8] = x24;
+  out1[9] = x26;
+  out1[10] = x28;
+  out1[11] = x30;
+  out1[12] = x32;
+  out1[13] = x34;
+  out1[14] = x36;
+  out1[15] = x37;
+  out1[16] = x40;
+  out1[17] = x42;
+  out1[18] = x44;
+  out1[19] = x46;
+  out1[20] = x48;
+  out1[21] = x50;
+  out1[22] = x52;
+  out1[23] = x53;
+  out1[24] = x56;
+  out1[25] = x58;
+  out1[26] = x60;
+  out1[27] = x62;
+  out1[28] = x64;
+  out1[29] = x66;
+  out1[30] = x68;
+  out1[31] = x69;
+  out1[32] = x72;
+  out1[33] = x74;
+  out1[34] = x76;
+  out1[35] = x78;
+  out1[36] = x80;
+  out1[37] = x82;
+  out1[38] = x84;
+  out1[39] = x85;
+  out1[40] = x88;
+  out1[41] = x90;
+  out1[42] = x92;
+  out1[43] = x94;
+  out1[44] = x96;
+  out1[45] = x98;
+  out1[46] = x100;
+  out1[47] = x99;
 }
 
 /*
@@ -2554,31 +2549,26 @@ static void fiat_p384_from_bytes(uint64_t out1[6], const uint8_t arg1[48]) {
   uint64_t x47 = ((uint64_t)(arg1[1]) << 8);
   uint8_t x48 = (arg1[0]);
   uint64_t x49 = (x48 + (x47 + (x46 + (x45 + (x44 + (x43 + (x42 + x41)))))));
-  fiat_p384_uint1 x50 = (fiat_p384_uint1)((fiat_p384_uint128)x49 >> 64);
-  uint64_t x51 = (x49 & UINT64_C(0xffffffffffffffff));
-  uint64_t x52 = (x8 + (x7 + (x6 + (x5 + (x4 + (x3 + (x2 + x1)))))));
-  uint64_t x53 = (x16 + (x15 + (x14 + (x13 + (x12 + (x11 + (x10 + x9)))))));
-  uint64_t x54 = (x24 + (x23 + (x22 + (x21 + (x20 + (x19 + (x18 + x17)))))));
-  uint64_t x55 = (x32 + (x31 + (x30 + (x29 + (x28 + (x27 + (x26 + x25)))))));
-  uint64_t x56 = (x40 + (x39 + (x38 + (x37 + (x36 + (x35 + (x34 + x33)))))));
-  uint64_t x57 = (x50 + x56);
-  fiat_p384_uint1 x58 = (fiat_p384_uint1)((fiat_p384_uint128)x57 >> 64);
-  uint64_t x59 = (x57 & UINT64_C(0xffffffffffffffff));
-  uint64_t x60 = (x58 + x55);
-  fiat_p384_uint1 x61 = (fiat_p384_uint1)((fiat_p384_uint128)x60 >> 64);
-  uint64_t x62 = (x60 & UINT64_C(0xffffffffffffffff));
-  uint64_t x63 = (x61 + x54);
-  fiat_p384_uint1 x64 = (fiat_p384_uint1)((fiat_p384_uint128)x63 >> 64);
-  uint64_t x65 = (x63 & UINT64_C(0xffffffffffffffff));
-  uint64_t x66 = (x64 + x53);
-  fiat_p384_uint1 x67 = (fiat_p384_uint1)((fiat_p384_uint128)x66 >> 64);
-  uint64_t x68 = (x66 & UINT64_C(0xffffffffffffffff));
-  uint64_t x69 = (x67 + x52);
-  out1[0] = x51;
-  out1[1] = x59;
-  out1[2] = x62;
-  out1[3] = x65;
-  out1[4] = x68;
-  out1[5] = x69;
+  uint64_t x50 = (x49 & UINT64_C(0xffffffffffffffff));
+  uint64_t x51 = (x8 + (x7 + (x6 + (x5 + (x4 + (x3 + (x2 + x1)))))));
+  uint64_t x52 = (x16 + (x15 + (x14 + (x13 + (x12 + (x11 + (x10 + x9)))))));
+  uint64_t x53 = (x24 + (x23 + (x22 + (x21 + (x20 + (x19 + (x18 + x17)))))));
+  uint64_t x54 = (x32 + (x31 + (x30 + (x29 + (x28 + (x27 + (x26 + x25)))))));
+  uint64_t x55 = (x40 + (x39 + (x38 + (x37 + (x36 + (x35 + (x34 + x33)))))));
+  uint64_t x56 = (0x0 + x55);
+  uint64_t x57 = (x56 & UINT64_C(0xffffffffffffffff));
+  uint64_t x58 = (0x0 + x54);
+  uint64_t x59 = (x58 & UINT64_C(0xffffffffffffffff));
+  uint64_t x60 = (0x0 + x53);
+  uint64_t x61 = (x60 & UINT64_C(0xffffffffffffffff));
+  uint64_t x62 = (0x0 + x52);
+  uint64_t x63 = (x62 & UINT64_C(0xffffffffffffffff));
+  uint64_t x64 = (0x0 + x51);
+  out1[0] = x50;
+  out1[1] = x57;
+  out1[2] = x59;
+  out1[3] = x61;
+  out1[4] = x63;
+  out1[5] = x64;
 }
 
diff --git a/p484_64.c b/p484_64.c
index f3c66f1ff..399ada567 100644
--- a/p484_64.c
+++ b/p484_64.c
@@ -2315,616 +2315,577 @@ static void fiat_p484_from_montgomery(uint64_t out1[7], const uint64_t arg1[7])
   fiat_p484_addcarryx_u64(&x40, &x41, x39, x26, 0x0);
   uint64_t x42;
   fiat_p484_uint1 x43;
-  fiat_p484_addcarryx_u64(&x42, &x43, x27, 0x0, x3);
+  fiat_p484_addcarryx_u64(&x42, &x43, 0x0, (arg1[1]), x30);
   uint64_t x44;
   fiat_p484_uint1 x45;
-  fiat_p484_addcarryx_u64(&x44, &x45, x41, x42, 0x0);
+  fiat_p484_addcarryx_u64(&x44, &x45, x43, 0x0, x32);
   uint64_t x46;
   fiat_p484_uint1 x47;
-  fiat_p484_addcarryx_u64(&x46, &x47, 0x0, (arg1[1]), x30);
+  fiat_p484_addcarryx_u64(&x46, &x47, x45, 0x0, x34);
   uint64_t x48;
   fiat_p484_uint1 x49;
-  fiat_p484_addcarryx_u64(&x48, &x49, x47, 0x0, x32);
+  fiat_p484_addcarryx_u64(&x48, &x49, x47, 0x0, x36);
   uint64_t x50;
   fiat_p484_uint1 x51;
-  fiat_p484_addcarryx_u64(&x50, &x51, x49, 0x0, x34);
+  fiat_p484_addcarryx_u64(&x50, &x51, x49, 0x0, x38);
   uint64_t x52;
   fiat_p484_uint1 x53;
-  fiat_p484_addcarryx_u64(&x52, &x53, x51, 0x0, x36);
+  fiat_p484_addcarryx_u64(&x52, &x53, x51, 0x0, x40);
   uint64_t x54;
-  fiat_p484_uint1 x55;
-  fiat_p484_addcarryx_u64(&x54, &x55, x53, 0x0, x38);
+  uint64_t x55;
+  fiat_p484_mulx_u64(&x54, &x55, x42, UINT64_C(0x2341f27177344));
   uint64_t x56;
-  fiat_p484_uint1 x57;
-  fiat_p484_addcarryx_u64(&x56, &x57, x55, 0x0, x40);
+  uint64_t x57;
+  fiat_p484_mulx_u64(&x56, &x57, x42, UINT64_C(0x6cfc5fd681c52056));
   uint64_t x58;
-  fiat_p484_uint1 x59;
-  fiat_p484_addcarryx_u64(&x58, &x59, x57, 0x0, x44);
+  uint64_t x59;
+  fiat_p484_mulx_u64(&x58, &x59, x42, UINT64_C(0x7bc65c783158aea3));
   uint64_t x60;
-  fiat_p484_uint1 x61;
-  fiat_p484_addcarryx_u64(&x60, &x61, x45, 0x0, 0x0);
+  uint64_t x61;
+  fiat_p484_mulx_u64(&x60, &x61, x42, UINT64_C(0xfdc1767ae2ffffff));
   uint64_t x62;
-  fiat_p484_uint1 x63;
-  fiat_p484_addcarryx_u64(&x62, &x63, x59, 0x0, (fiat_p484_uint1)x60);
+  uint64_t x63;
+  fiat_p484_mulx_u64(&x62, &x63, x42, UINT64_C(0xffffffffffffffff));
   uint64_t x64;
   uint64_t x65;
-  fiat_p484_mulx_u64(&x64, &x65, x46, UINT64_C(0x2341f27177344));
+  fiat_p484_mulx_u64(&x64, &x65, x42, UINT64_C(0xffffffffffffffff));
   uint64_t x66;
   uint64_t x67;
-  fiat_p484_mulx_u64(&x66, &x67, x46, UINT64_C(0x6cfc5fd681c52056));
+  fiat_p484_mulx_u64(&x66, &x67, x42, UINT64_C(0xffffffffffffffff));
   uint64_t x68;
-  uint64_t x69;
-  fiat_p484_mulx_u64(&x68, &x69, x46, UINT64_C(0x7bc65c783158aea3));
+  fiat_p484_uint1 x69;
+  fiat_p484_addcarryx_u64(&x68, &x69, 0x0, x64, x67);
   uint64_t x70;
-  uint64_t x71;
-  fiat_p484_mulx_u64(&x70, &x71, x46, UINT64_C(0xfdc1767ae2ffffff));
+  fiat_p484_uint1 x71;
+  fiat_p484_addcarryx_u64(&x70, &x71, x69, x62, x65);
   uint64_t x72;
-  uint64_t x73;
-  fiat_p484_mulx_u64(&x72, &x73, x46, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x73;
+  fiat_p484_addcarryx_u64(&x72, &x73, x71, x60, x63);
   uint64_t x74;
-  uint64_t x75;
-  fiat_p484_mulx_u64(&x74, &x75, x46, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x75;
+  fiat_p484_addcarryx_u64(&x74, &x75, x73, x58, x61);
   uint64_t x76;
-  uint64_t x77;
-  fiat_p484_mulx_u64(&x76, &x77, x46, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x77;
+  fiat_p484_addcarryx_u64(&x76, &x77, x75, x56, x59);
   uint64_t x78;
   fiat_p484_uint1 x79;
-  fiat_p484_addcarryx_u64(&x78, &x79, 0x0, x74, x77);
+  fiat_p484_addcarryx_u64(&x78, &x79, x77, x54, x57);
   uint64_t x80;
   fiat_p484_uint1 x81;
-  fiat_p484_addcarryx_u64(&x80, &x81, x79, x72, x75);
+  fiat_p484_addcarryx_u64(&x80, &x81, 0x0, x66, x42);
   uint64_t x82;
   fiat_p484_uint1 x83;
-  fiat_p484_addcarryx_u64(&x82, &x83, x81, x70, x73);
+  fiat_p484_addcarryx_u64(&x82, &x83, x81, x68, x44);
   uint64_t x84;
   fiat_p484_uint1 x85;
-  fiat_p484_addcarryx_u64(&x84, &x85, x83, x68, x71);
+  fiat_p484_addcarryx_u64(&x84, &x85, x83, x70, x46);
   uint64_t x86;
   fiat_p484_uint1 x87;
-  fiat_p484_addcarryx_u64(&x86, &x87, x85, x66, x69);
+  fiat_p484_addcarryx_u64(&x86, &x87, x85, x72, x48);
   uint64_t x88;
   fiat_p484_uint1 x89;
-  fiat_p484_addcarryx_u64(&x88, &x89, x87, x64, x67);
+  fiat_p484_addcarryx_u64(&x88, &x89, x87, x74, x50);
   uint64_t x90;
   fiat_p484_uint1 x91;
-  fiat_p484_addcarryx_u64(&x90, &x91, 0x0, x76, x46);
+  fiat_p484_addcarryx_u64(&x90, &x91, x89, x76, x52);
   uint64_t x92;
   fiat_p484_uint1 x93;
-  fiat_p484_addcarryx_u64(&x92, &x93, x91, x78, x48);
+  fiat_p484_addcarryx_u64(&x92, &x93, x27, 0x0, x3);
   uint64_t x94;
   fiat_p484_uint1 x95;
-  fiat_p484_addcarryx_u64(&x94, &x95, x93, x80, x50);
+  fiat_p484_addcarryx_u64(&x94, &x95, x41, x92, 0x0);
   uint64_t x96;
   fiat_p484_uint1 x97;
-  fiat_p484_addcarryx_u64(&x96, &x97, x95, x82, x52);
+  fiat_p484_addcarryx_u64(&x96, &x97, x53, 0x0, x94);
   uint64_t x98;
   fiat_p484_uint1 x99;
-  fiat_p484_addcarryx_u64(&x98, &x99, x97, x84, x54);
+  fiat_p484_addcarryx_u64(&x98, &x99, x91, x78, x96);
   uint64_t x100;
   fiat_p484_uint1 x101;
-  fiat_p484_addcarryx_u64(&x100, &x101, x99, x86, x56);
+  fiat_p484_addcarryx_u64(&x100, &x101, 0x0, (arg1[2]), x82);
   uint64_t x102;
   fiat_p484_uint1 x103;
-  fiat_p484_addcarryx_u64(&x102, &x103, x101, x88, x58);
+  fiat_p484_addcarryx_u64(&x102, &x103, x101, 0x0, x84);
   uint64_t x104;
   fiat_p484_uint1 x105;
-  fiat_p484_addcarryx_u64(&x104, &x105, x89, 0x0, x65);
+  fiat_p484_addcarryx_u64(&x104, &x105, x103, 0x0, x86);
   uint64_t x106;
   fiat_p484_uint1 x107;
-  fiat_p484_addcarryx_u64(&x106, &x107, x103, x104, (fiat_p484_uint1)x62);
+  fiat_p484_addcarryx_u64(&x106, &x107, x105, 0x0, x88);
   uint64_t x108;
   fiat_p484_uint1 x109;
-  fiat_p484_addcarryx_u64(&x108, &x109, 0x0, (arg1[2]), x92);
+  fiat_p484_addcarryx_u64(&x108, &x109, x107, 0x0, x90);
   uint64_t x110;
   fiat_p484_uint1 x111;
-  fiat_p484_addcarryx_u64(&x110, &x111, x109, 0x0, x94);
+  fiat_p484_addcarryx_u64(&x110, &x111, x109, 0x0, x98);
   uint64_t x112;
-  fiat_p484_uint1 x113;
-  fiat_p484_addcarryx_u64(&x112, &x113, x111, 0x0, x96);
+  uint64_t x113;
+  fiat_p484_mulx_u64(&x112, &x113, x100, UINT64_C(0x2341f27177344));
   uint64_t x114;
-  fiat_p484_uint1 x115;
-  fiat_p484_addcarryx_u64(&x114, &x115, x113, 0x0, x98);
+  uint64_t x115;
+  fiat_p484_mulx_u64(&x114, &x115, x100, UINT64_C(0x6cfc5fd681c52056));
   uint64_t x116;
-  fiat_p484_uint1 x117;
-  fiat_p484_addcarryx_u64(&x116, &x117, x115, 0x0, x100);
+  uint64_t x117;
+  fiat_p484_mulx_u64(&x116, &x117, x100, UINT64_C(0x7bc65c783158aea3));
   uint64_t x118;
-  fiat_p484_uint1 x119;
-  fiat_p484_addcarryx_u64(&x118, &x119, x117, 0x0, x102);
+  uint64_t x119;
+  fiat_p484_mulx_u64(&x118, &x119, x100, UINT64_C(0xfdc1767ae2ffffff));
   uint64_t x120;
-  fiat_p484_uint1 x121;
-  fiat_p484_addcarryx_u64(&x120, &x121, x119, 0x0, x106);
+  uint64_t x121;
+  fiat_p484_mulx_u64(&x120, &x121, x100, UINT64_C(0xffffffffffffffff));
   uint64_t x122;
-  fiat_p484_uint1 x123;
-  fiat_p484_addcarryx_u64(&x122, &x123, x107, 0x0, x63);
+  uint64_t x123;
+  fiat_p484_mulx_u64(&x122, &x123, x100, UINT64_C(0xffffffffffffffff));
   uint64_t x124;
-  fiat_p484_uint1 x125;
-  fiat_p484_addcarryx_u64(&x124, &x125, x121, 0x0, (fiat_p484_uint1)x122);
+  uint64_t x125;
+  fiat_p484_mulx_u64(&x124, &x125, x100, UINT64_C(0xffffffffffffffff));
   uint64_t x126;
-  uint64_t x127;
-  fiat_p484_mulx_u64(&x126, &x127, x108, UINT64_C(0x2341f27177344));
+  fiat_p484_uint1 x127;
+  fiat_p484_addcarryx_u64(&x126, &x127, 0x0, x122, x125);
   uint64_t x128;
-  uint64_t x129;
-  fiat_p484_mulx_u64(&x128, &x129, x108, UINT64_C(0x6cfc5fd681c52056));
+  fiat_p484_uint1 x129;
+  fiat_p484_addcarryx_u64(&x128, &x129, x127, x120, x123);
   uint64_t x130;
-  uint64_t x131;
-  fiat_p484_mulx_u64(&x130, &x131, x108, UINT64_C(0x7bc65c783158aea3));
+  fiat_p484_uint1 x131;
+  fiat_p484_addcarryx_u64(&x130, &x131, x129, x118, x121);
   uint64_t x132;
-  uint64_t x133;
-  fiat_p484_mulx_u64(&x132, &x133, x108, UINT64_C(0xfdc1767ae2ffffff));
+  fiat_p484_uint1 x133;
+  fiat_p484_addcarryx_u64(&x132, &x133, x131, x116, x119);
   uint64_t x134;
-  uint64_t x135;
-  fiat_p484_mulx_u64(&x134, &x135, x108, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x135;
+  fiat_p484_addcarryx_u64(&x134, &x135, x133, x114, x117);
   uint64_t x136;
-  uint64_t x137;
-  fiat_p484_mulx_u64(&x136, &x137, x108, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x137;
+  fiat_p484_addcarryx_u64(&x136, &x137, x135, x112, x115);
   uint64_t x138;
-  uint64_t x139;
-  fiat_p484_mulx_u64(&x138, &x139, x108, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x139;
+  fiat_p484_addcarryx_u64(&x138, &x139, 0x0, x124, x100);
   uint64_t x140;
   fiat_p484_uint1 x141;
-  fiat_p484_addcarryx_u64(&x140, &x141, 0x0, x136, x139);
+  fiat_p484_addcarryx_u64(&x140, &x141, x139, x126, x102);
   uint64_t x142;
   fiat_p484_uint1 x143;
-  fiat_p484_addcarryx_u64(&x142, &x143, x141, x134, x137);
+  fiat_p484_addcarryx_u64(&x142, &x143, x141, x128, x104);
   uint64_t x144;
   fiat_p484_uint1 x145;
-  fiat_p484_addcarryx_u64(&x144, &x145, x143, x132, x135);
+  fiat_p484_addcarryx_u64(&x144, &x145, x143, x130, x106);
   uint64_t x146;
   fiat_p484_uint1 x147;
-  fiat_p484_addcarryx_u64(&x146, &x147, x145, x130, x133);
+  fiat_p484_addcarryx_u64(&x146, &x147, x145, x132, x108);
   uint64_t x148;
   fiat_p484_uint1 x149;
-  fiat_p484_addcarryx_u64(&x148, &x149, x147, x128, x131);
+  fiat_p484_addcarryx_u64(&x148, &x149, x147, x134, x110);
   uint64_t x150;
   fiat_p484_uint1 x151;
-  fiat_p484_addcarryx_u64(&x150, &x151, x149, x126, x129);
+  fiat_p484_addcarryx_u64(&x150, &x151, x79, 0x0, x55);
   uint64_t x152;
   fiat_p484_uint1 x153;
-  fiat_p484_addcarryx_u64(&x152, &x153, 0x0, x138, x108);
+  fiat_p484_addcarryx_u64(&x152, &x153, x99, x150, 0x0);
   uint64_t x154;
   fiat_p484_uint1 x155;
-  fiat_p484_addcarryx_u64(&x154, &x155, x153, x140, x110);
+  fiat_p484_addcarryx_u64(&x154, &x155, x111, 0x0, x152);
   uint64_t x156;
   fiat_p484_uint1 x157;
-  fiat_p484_addcarryx_u64(&x156, &x157, x155, x142, x112);
+  fiat_p484_addcarryx_u64(&x156, &x157, x149, x136, x154);
   uint64_t x158;
   fiat_p484_uint1 x159;
-  fiat_p484_addcarryx_u64(&x158, &x159, x157, x144, x114);
+  fiat_p484_addcarryx_u64(&x158, &x159, 0x0, (arg1[3]), x140);
   uint64_t x160;
   fiat_p484_uint1 x161;
-  fiat_p484_addcarryx_u64(&x160, &x161, x159, x146, x116);
+  fiat_p484_addcarryx_u64(&x160, &x161, x159, 0x0, x142);
   uint64_t x162;
   fiat_p484_uint1 x163;
-  fiat_p484_addcarryx_u64(&x162, &x163, x161, x148, x118);
+  fiat_p484_addcarryx_u64(&x162, &x163, x161, 0x0, x144);
   uint64_t x164;
   fiat_p484_uint1 x165;
-  fiat_p484_addcarryx_u64(&x164, &x165, x163, x150, x120);
+  fiat_p484_addcarryx_u64(&x164, &x165, x163, 0x0, x146);
   uint64_t x166;
   fiat_p484_uint1 x167;
-  fiat_p484_addcarryx_u64(&x166, &x167, x151, 0x0, x127);
+  fiat_p484_addcarryx_u64(&x166, &x167, x165, 0x0, x148);
   uint64_t x168;
   fiat_p484_uint1 x169;
-  fiat_p484_addcarryx_u64(&x168, &x169, x165, x166, (fiat_p484_uint1)x124);
+  fiat_p484_addcarryx_u64(&x168, &x169, x167, 0x0, x156);
   uint64_t x170;
-  fiat_p484_uint1 x171;
-  fiat_p484_addcarryx_u64(&x170, &x171, 0x0, (arg1[3]), x154);
+  uint64_t x171;
+  fiat_p484_mulx_u64(&x170, &x171, x158, UINT64_C(0x2341f27177344));
   uint64_t x172;
-  fiat_p484_uint1 x173;
-  fiat_p484_addcarryx_u64(&x172, &x173, x171, 0x0, x156);
+  uint64_t x173;
+  fiat_p484_mulx_u64(&x172, &x173, x158, UINT64_C(0x6cfc5fd681c52056));
   uint64_t x174;
-  fiat_p484_uint1 x175;
-  fiat_p484_addcarryx_u64(&x174, &x175, x173, 0x0, x158);
+  uint64_t x175;
+  fiat_p484_mulx_u64(&x174, &x175, x158, UINT64_C(0x7bc65c783158aea3));
   uint64_t x176;
-  fiat_p484_uint1 x177;
-  fiat_p484_addcarryx_u64(&x176, &x177, x175, 0x0, x160);
+  uint64_t x177;
+  fiat_p484_mulx_u64(&x176, &x177, x158, UINT64_C(0xfdc1767ae2ffffff));
   uint64_t x178;
-  fiat_p484_uint1 x179;
-  fiat_p484_addcarryx_u64(&x178, &x179, x177, 0x0, x162);
+  uint64_t x179;
+  fiat_p484_mulx_u64(&x178, &x179, x158, UINT64_C(0xffffffffffffffff));
   uint64_t x180;
-  fiat_p484_uint1 x181;
-  fiat_p484_addcarryx_u64(&x180, &x181, x179, 0x0, x164);
+  uint64_t x181;
+  fiat_p484_mulx_u64(&x180, &x181, x158, UINT64_C(0xffffffffffffffff));
   uint64_t x182;
-  fiat_p484_uint1 x183;
-  fiat_p484_addcarryx_u64(&x182, &x183, x181, 0x0, x168);
+  uint64_t x183;
+  fiat_p484_mulx_u64(&x182, &x183, x158, UINT64_C(0xffffffffffffffff));
   uint64_t x184;
   fiat_p484_uint1 x185;
-  fiat_p484_addcarryx_u64(&x184, &x185, x169, 0x0, x125);
+  fiat_p484_addcarryx_u64(&x184, &x185, 0x0, x180, x183);
   uint64_t x186;
   fiat_p484_uint1 x187;
-  fiat_p484_addcarryx_u64(&x186, &x187, x183, 0x0, (fiat_p484_uint1)x184);
+  fiat_p484_addcarryx_u64(&x186, &x187, x185, x178, x181);
   uint64_t x188;
-  uint64_t x189;
-  fiat_p484_mulx_u64(&x188, &x189, x170, UINT64_C(0x2341f27177344));
+  fiat_p484_uint1 x189;
+  fiat_p484_addcarryx_u64(&x188, &x189, x187, x176, x179);
   uint64_t x190;
-  uint64_t x191;
-  fiat_p484_mulx_u64(&x190, &x191, x170, UINT64_C(0x6cfc5fd681c52056));
+  fiat_p484_uint1 x191;
+  fiat_p484_addcarryx_u64(&x190, &x191, x189, x174, x177);
   uint64_t x192;
-  uint64_t x193;
-  fiat_p484_mulx_u64(&x192, &x193, x170, UINT64_C(0x7bc65c783158aea3));
+  fiat_p484_uint1 x193;
+  fiat_p484_addcarryx_u64(&x192, &x193, x191, x172, x175);
   uint64_t x194;
-  uint64_t x195;
-  fiat_p484_mulx_u64(&x194, &x195, x170, UINT64_C(0xfdc1767ae2ffffff));
+  fiat_p484_uint1 x195;
+  fiat_p484_addcarryx_u64(&x194, &x195, x193, x170, x173);
   uint64_t x196;
-  uint64_t x197;
-  fiat_p484_mulx_u64(&x196, &x197, x170, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x197;
+  fiat_p484_addcarryx_u64(&x196, &x197, 0x0, x182, x158);
   uint64_t x198;
-  uint64_t x199;
-  fiat_p484_mulx_u64(&x198, &x199, x170, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x199;
+  fiat_p484_addcarryx_u64(&x198, &x199, x197, x184, x160);
   uint64_t x200;
-  uint64_t x201;
-  fiat_p484_mulx_u64(&x200, &x201, x170, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x201;
+  fiat_p484_addcarryx_u64(&x200, &x201, x199, x186, x162);
   uint64_t x202;
   fiat_p484_uint1 x203;
-  fiat_p484_addcarryx_u64(&x202, &x203, 0x0, x198, x201);
+  fiat_p484_addcarryx_u64(&x202, &x203, x201, x188, x164);
   uint64_t x204;
   fiat_p484_uint1 x205;
-  fiat_p484_addcarryx_u64(&x204, &x205, x203, x196, x199);
+  fiat_p484_addcarryx_u64(&x204, &x205, x203, x190, x166);
   uint64_t x206;
   fiat_p484_uint1 x207;
-  fiat_p484_addcarryx_u64(&x206, &x207, x205, x194, x197);
+  fiat_p484_addcarryx_u64(&x206, &x207, x205, x192, x168);
   uint64_t x208;
   fiat_p484_uint1 x209;
-  fiat_p484_addcarryx_u64(&x208, &x209, x207, x192, x195);
+  fiat_p484_addcarryx_u64(&x208, &x209, x137, 0x0, x113);
   uint64_t x210;
   fiat_p484_uint1 x211;
-  fiat_p484_addcarryx_u64(&x210, &x211, x209, x190, x193);
+  fiat_p484_addcarryx_u64(&x210, &x211, x157, x208, 0x0);
   uint64_t x212;
   fiat_p484_uint1 x213;
-  fiat_p484_addcarryx_u64(&x212, &x213, x211, x188, x191);
+  fiat_p484_addcarryx_u64(&x212, &x213, x169, 0x0, x210);
   uint64_t x214;
   fiat_p484_uint1 x215;
-  fiat_p484_addcarryx_u64(&x214, &x215, 0x0, x200, x170);
+  fiat_p484_addcarryx_u64(&x214, &x215, x207, x194, x212);
   uint64_t x216;
   fiat_p484_uint1 x217;
-  fiat_p484_addcarryx_u64(&x216, &x217, x215, x202, x172);
+  fiat_p484_addcarryx_u64(&x216, &x217, 0x0, (arg1[4]), x198);
   uint64_t x218;
   fiat_p484_uint1 x219;
-  fiat_p484_addcarryx_u64(&x218, &x219, x217, x204, x174);
+  fiat_p484_addcarryx_u64(&x218, &x219, x217, 0x0, x200);
   uint64_t x220;
   fiat_p484_uint1 x221;
-  fiat_p484_addcarryx_u64(&x220, &x221, x219, x206, x176);
+  fiat_p484_addcarryx_u64(&x220, &x221, x219, 0x0, x202);
   uint64_t x222;
   fiat_p484_uint1 x223;
-  fiat_p484_addcarryx_u64(&x222, &x223, x221, x208, x178);
+  fiat_p484_addcarryx_u64(&x222, &x223, x221, 0x0, x204);
   uint64_t x224;
   fiat_p484_uint1 x225;
-  fiat_p484_addcarryx_u64(&x224, &x225, x223, x210, x180);
+  fiat_p484_addcarryx_u64(&x224, &x225, x223, 0x0, x206);
   uint64_t x226;
   fiat_p484_uint1 x227;
-  fiat_p484_addcarryx_u64(&x226, &x227, x225, x212, x182);
+  fiat_p484_addcarryx_u64(&x226, &x227, x225, 0x0, x214);
   uint64_t x228;
-  fiat_p484_uint1 x229;
-  fiat_p484_addcarryx_u64(&x228, &x229, x213, 0x0, x189);
+  uint64_t x229;
+  fiat_p484_mulx_u64(&x228, &x229, x216, UINT64_C(0x2341f27177344));
   uint64_t x230;
-  fiat_p484_uint1 x231;
-  fiat_p484_addcarryx_u64(&x230, &x231, x227, x228, (fiat_p484_uint1)x186);
+  uint64_t x231;
+  fiat_p484_mulx_u64(&x230, &x231, x216, UINT64_C(0x6cfc5fd681c52056));
   uint64_t x232;
-  fiat_p484_uint1 x233;
-  fiat_p484_addcarryx_u64(&x232, &x233, 0x0, (arg1[4]), x216);
+  uint64_t x233;
+  fiat_p484_mulx_u64(&x232, &x233, x216, UINT64_C(0x7bc65c783158aea3));
   uint64_t x234;
-  fiat_p484_uint1 x235;
-  fiat_p484_addcarryx_u64(&x234, &x235, x233, 0x0, x218);
+  uint64_t x235;
+  fiat_p484_mulx_u64(&x234, &x235, x216, UINT64_C(0xfdc1767ae2ffffff));
   uint64_t x236;
-  fiat_p484_uint1 x237;
-  fiat_p484_addcarryx_u64(&x236, &x237, x235, 0x0, x220);
+  uint64_t x237;
+  fiat_p484_mulx_u64(&x236, &x237, x216, UINT64_C(0xffffffffffffffff));
   uint64_t x238;
-  fiat_p484_uint1 x239;
-  fiat_p484_addcarryx_u64(&x238, &x239, x237, 0x0, x222);
+  uint64_t x239;
+  fiat_p484_mulx_u64(&x238, &x239, x216, UINT64_C(0xffffffffffffffff));
   uint64_t x240;
-  fiat_p484_uint1 x241;
-  fiat_p484_addcarryx_u64(&x240, &x241, x239, 0x0, x224);
+  uint64_t x241;
+  fiat_p484_mulx_u64(&x240, &x241, x216, UINT64_C(0xffffffffffffffff));
   uint64_t x242;
   fiat_p484_uint1 x243;
-  fiat_p484_addcarryx_u64(&x242, &x243, x241, 0x0, x226);
+  fiat_p484_addcarryx_u64(&x242, &x243, 0x0, x238, x241);
   uint64_t x244;
   fiat_p484_uint1 x245;
-  fiat_p484_addcarryx_u64(&x244, &x245, x243, 0x0, x230);
+  fiat_p484_addcarryx_u64(&x244, &x245, x243, x236, x239);
   uint64_t x246;
   fiat_p484_uint1 x247;
-  fiat_p484_addcarryx_u64(&x246, &x247, x231, 0x0, x187);
+  fiat_p484_addcarryx_u64(&x246, &x247, x245, x234, x237);
   uint64_t x248;
   fiat_p484_uint1 x249;
-  fiat_p484_addcarryx_u64(&x248, &x249, x245, 0x0, (fiat_p484_uint1)x246);
+  fiat_p484_addcarryx_u64(&x248, &x249, x247, x232, x235);
   uint64_t x250;
-  uint64_t x251;
-  fiat_p484_mulx_u64(&x250, &x251, x232, UINT64_C(0x2341f27177344));
+  fiat_p484_uint1 x251;
+  fiat_p484_addcarryx_u64(&x250, &x251, x249, x230, x233);
   uint64_t x252;
-  uint64_t x253;
-  fiat_p484_mulx_u64(&x252, &x253, x232, UINT64_C(0x6cfc5fd681c52056));
+  fiat_p484_uint1 x253;
+  fiat_p484_addcarryx_u64(&x252, &x253, x251, x228, x231);
   uint64_t x254;
-  uint64_t x255;
-  fiat_p484_mulx_u64(&x254, &x255, x232, UINT64_C(0x7bc65c783158aea3));
+  fiat_p484_uint1 x255;
+  fiat_p484_addcarryx_u64(&x254, &x255, 0x0, x240, x216);
   uint64_t x256;
-  uint64_t x257;
-  fiat_p484_mulx_u64(&x256, &x257, x232, UINT64_C(0xfdc1767ae2ffffff));
+  fiat_p484_uint1 x257;
+  fiat_p484_addcarryx_u64(&x256, &x257, x255, x242, x218);
   uint64_t x258;
-  uint64_t x259;
-  fiat_p484_mulx_u64(&x258, &x259, x232, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x259;
+  fiat_p484_addcarryx_u64(&x258, &x259, x257, x244, x220);
   uint64_t x260;
-  uint64_t x261;
-  fiat_p484_mulx_u64(&x260, &x261, x232, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x261;
+  fiat_p484_addcarryx_u64(&x260, &x261, x259, x246, x222);
   uint64_t x262;
-  uint64_t x263;
-  fiat_p484_mulx_u64(&x262, &x263, x232, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x263;
+  fiat_p484_addcarryx_u64(&x262, &x263, x261, x248, x224);
   uint64_t x264;
   fiat_p484_uint1 x265;
-  fiat_p484_addcarryx_u64(&x264, &x265, 0x0, x260, x263);
+  fiat_p484_addcarryx_u64(&x264, &x265, x263, x250, x226);
   uint64_t x266;
   fiat_p484_uint1 x267;
-  fiat_p484_addcarryx_u64(&x266, &x267, x265, x258, x261);
+  fiat_p484_addcarryx_u64(&x266, &x267, x195, 0x0, x171);
   uint64_t x268;
   fiat_p484_uint1 x269;
-  fiat_p484_addcarryx_u64(&x268, &x269, x267, x256, x259);
+  fiat_p484_addcarryx_u64(&x268, &x269, x215, x266, 0x0);
   uint64_t x270;
   fiat_p484_uint1 x271;
-  fiat_p484_addcarryx_u64(&x270, &x271, x269, x254, x257);
+  fiat_p484_addcarryx_u64(&x270, &x271, x227, 0x0, x268);
   uint64_t x272;
   fiat_p484_uint1 x273;
-  fiat_p484_addcarryx_u64(&x272, &x273, x271, x252, x255);
+  fiat_p484_addcarryx_u64(&x272, &x273, x265, x252, x270);
   uint64_t x274;
   fiat_p484_uint1 x275;
-  fiat_p484_addcarryx_u64(&x274, &x275, x273, x250, x253);
+  fiat_p484_addcarryx_u64(&x274, &x275, 0x0, (arg1[5]), x256);
   uint64_t x276;
   fiat_p484_uint1 x277;
-  fiat_p484_addcarryx_u64(&x276, &x277, 0x0, x262, x232);
+  fiat_p484_addcarryx_u64(&x276, &x277, x275, 0x0, x258);
   uint64_t x278;
   fiat_p484_uint1 x279;
-  fiat_p484_addcarryx_u64(&x278, &x279, x277, x264, x234);
+  fiat_p484_addcarryx_u64(&x278, &x279, x277, 0x0, x260);
   uint64_t x280;
   fiat_p484_uint1 x281;
-  fiat_p484_addcarryx_u64(&x280, &x281, x279, x266, x236);
+  fiat_p484_addcarryx_u64(&x280, &x281, x279, 0x0, x262);
   uint64_t x282;
   fiat_p484_uint1 x283;
-  fiat_p484_addcarryx_u64(&x282, &x283, x281, x268, x238);
+  fiat_p484_addcarryx_u64(&x282, &x283, x281, 0x0, x264);
   uint64_t x284;
   fiat_p484_uint1 x285;
-  fiat_p484_addcarryx_u64(&x284, &x285, x283, x270, x240);
+  fiat_p484_addcarryx_u64(&x284, &x285, x283, 0x0, x272);
   uint64_t x286;
-  fiat_p484_uint1 x287;
-  fiat_p484_addcarryx_u64(&x286, &x287, x285, x272, x242);
+  uint64_t x287;
+  fiat_p484_mulx_u64(&x286, &x287, x274, UINT64_C(0x2341f27177344));
   uint64_t x288;
-  fiat_p484_uint1 x289;
-  fiat_p484_addcarryx_u64(&x288, &x289, x287, x274, x244);
+  uint64_t x289;
+  fiat_p484_mulx_u64(&x288, &x289, x274, UINT64_C(0x6cfc5fd681c52056));
   uint64_t x290;
-  fiat_p484_uint1 x291;
-  fiat_p484_addcarryx_u64(&x290, &x291, x275, 0x0, x251);
+  uint64_t x291;
+  fiat_p484_mulx_u64(&x290, &x291, x274, UINT64_C(0x7bc65c783158aea3));
   uint64_t x292;
-  fiat_p484_uint1 x293;
-  fiat_p484_addcarryx_u64(&x292, &x293, x289, x290, (fiat_p484_uint1)x248);
+  uint64_t x293;
+  fiat_p484_mulx_u64(&x292, &x293, x274, UINT64_C(0xfdc1767ae2ffffff));
   uint64_t x294;
-  fiat_p484_uint1 x295;
-  fiat_p484_addcarryx_u64(&x294, &x295, 0x0, (arg1[5]), x278);
+  uint64_t x295;
+  fiat_p484_mulx_u64(&x294, &x295, x274, UINT64_C(0xffffffffffffffff));
   uint64_t x296;
-  fiat_p484_uint1 x297;
-  fiat_p484_addcarryx_u64(&x296, &x297, x295, 0x0, x280);
+  uint64_t x297;
+  fiat_p484_mulx_u64(&x296, &x297, x274, UINT64_C(0xffffffffffffffff));
   uint64_t x298;
-  fiat_p484_uint1 x299;
-  fiat_p484_addcarryx_u64(&x298, &x299, x297, 0x0, x282);
+  uint64_t x299;
+  fiat_p484_mulx_u64(&x298, &x299, x274, UINT64_C(0xffffffffffffffff));
   uint64_t x300;
   fiat_p484_uint1 x301;
-  fiat_p484_addcarryx_u64(&x300, &x301, x299, 0x0, x284);
+  fiat_p484_addcarryx_u64(&x300, &x301, 0x0, x296, x299);
   uint64_t x302;
   fiat_p484_uint1 x303;
-  fiat_p484_addcarryx_u64(&x302, &x303, x301, 0x0, x286);
+  fiat_p484_addcarryx_u64(&x302, &x303, x301, x294, x297);
   uint64_t x304;
   fiat_p484_uint1 x305;
-  fiat_p484_addcarryx_u64(&x304, &x305, x303, 0x0, x288);
+  fiat_p484_addcarryx_u64(&x304, &x305, x303, x292, x295);
   uint64_t x306;
   fiat_p484_uint1 x307;
-  fiat_p484_addcarryx_u64(&x306, &x307, x305, 0x0, x292);
+  fiat_p484_addcarryx_u64(&x306, &x307, x305, x290, x293);
   uint64_t x308;
   fiat_p484_uint1 x309;
-  fiat_p484_addcarryx_u64(&x308, &x309, x293, 0x0, x249);
+  fiat_p484_addcarryx_u64(&x308, &x309, x307, x288, x291);
   uint64_t x310;
   fiat_p484_uint1 x311;
-  fiat_p484_addcarryx_u64(&x310, &x311, x307, 0x0, (fiat_p484_uint1)x308);
+  fiat_p484_addcarryx_u64(&x310, &x311, x309, x286, x289);
   uint64_t x312;
-  uint64_t x313;
-  fiat_p484_mulx_u64(&x312, &x313, x294, UINT64_C(0x2341f27177344));
+  fiat_p484_uint1 x313;
+  fiat_p484_addcarryx_u64(&x312, &x313, 0x0, x298, x274);
   uint64_t x314;
-  uint64_t x315;
-  fiat_p484_mulx_u64(&x314, &x315, x294, UINT64_C(0x6cfc5fd681c52056));
+  fiat_p484_uint1 x315;
+  fiat_p484_addcarryx_u64(&x314, &x315, x313, x300, x276);
   uint64_t x316;
-  uint64_t x317;
-  fiat_p484_mulx_u64(&x316, &x317, x294, UINT64_C(0x7bc65c783158aea3));
+  fiat_p484_uint1 x317;
+  fiat_p484_addcarryx_u64(&x316, &x317, x315, x302, x278);
   uint64_t x318;
-  uint64_t x319;
-  fiat_p484_mulx_u64(&x318, &x319, x294, UINT64_C(0xfdc1767ae2ffffff));
+  fiat_p484_uint1 x319;
+  fiat_p484_addcarryx_u64(&x318, &x319, x317, x304, x280);
   uint64_t x320;
-  uint64_t x321;
-  fiat_p484_mulx_u64(&x320, &x321, x294, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x321;
+  fiat_p484_addcarryx_u64(&x320, &x321, x319, x306, x282);
   uint64_t x322;
-  uint64_t x323;
-  fiat_p484_mulx_u64(&x322, &x323, x294, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x323;
+  fiat_p484_addcarryx_u64(&x322, &x323, x321, x308, x284);
   uint64_t x324;
-  uint64_t x325;
-  fiat_p484_mulx_u64(&x324, &x325, x294, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x325;
+  fiat_p484_addcarryx_u64(&x324, &x325, x253, 0x0, x229);
   uint64_t x326;
   fiat_p484_uint1 x327;
-  fiat_p484_addcarryx_u64(&x326, &x327, 0x0, x322, x325);
+  fiat_p484_addcarryx_u64(&x326, &x327, x273, x324, 0x0);
   uint64_t x328;
   fiat_p484_uint1 x329;
-  fiat_p484_addcarryx_u64(&x328, &x329, x327, x320, x323);
+  fiat_p484_addcarryx_u64(&x328, &x329, x285, 0x0, x326);
   uint64_t x330;
   fiat_p484_uint1 x331;
-  fiat_p484_addcarryx_u64(&x330, &x331, x329, x318, x321);
+  fiat_p484_addcarryx_u64(&x330, &x331, x323, x310, x328);
   uint64_t x332;
   fiat_p484_uint1 x333;
-  fiat_p484_addcarryx_u64(&x332, &x333, x331, x316, x319);
+  fiat_p484_addcarryx_u64(&x332, &x333, 0x0, (arg1[6]), x314);
   uint64_t x334;
   fiat_p484_uint1 x335;
-  fiat_p484_addcarryx_u64(&x334, &x335, x333, x314, x317);
+  fiat_p484_addcarryx_u64(&x334, &x335, x333, 0x0, x316);
   uint64_t x336;
   fiat_p484_uint1 x337;
-  fiat_p484_addcarryx_u64(&x336, &x337, x335, x312, x315);
+  fiat_p484_addcarryx_u64(&x336, &x337, x335, 0x0, x318);
   uint64_t x338;
   fiat_p484_uint1 x339;
-  fiat_p484_addcarryx_u64(&x338, &x339, 0x0, x324, x294);
+  fiat_p484_addcarryx_u64(&x338, &x339, x337, 0x0, x320);
   uint64_t x340;
   fiat_p484_uint1 x341;
-  fiat_p484_addcarryx_u64(&x340, &x341, x339, x326, x296);
+  fiat_p484_addcarryx_u64(&x340, &x341, x339, 0x0, x322);
   uint64_t x342;
   fiat_p484_uint1 x343;
-  fiat_p484_addcarryx_u64(&x342, &x343, x341, x328, x298);
+  fiat_p484_addcarryx_u64(&x342, &x343, x341, 0x0, x330);
   uint64_t x344;
-  fiat_p484_uint1 x345;
-  fiat_p484_addcarryx_u64(&x344, &x345, x343, x330, x300);
+  uint64_t x345;
+  fiat_p484_mulx_u64(&x344, &x345, x332, UINT64_C(0x2341f27177344));
   uint64_t x346;
-  fiat_p484_uint1 x347;
-  fiat_p484_addcarryx_u64(&x346, &x347, x345, x332, x302);
+  uint64_t x347;
+  fiat_p484_mulx_u64(&x346, &x347, x332, UINT64_C(0x6cfc5fd681c52056));
   uint64_t x348;
-  fiat_p484_uint1 x349;
-  fiat_p484_addcarryx_u64(&x348, &x349, x347, x334, x304);
+  uint64_t x349;
+  fiat_p484_mulx_u64(&x348, &x349, x332, UINT64_C(0x7bc65c783158aea3));
   uint64_t x350;
-  fiat_p484_uint1 x351;
-  fiat_p484_addcarryx_u64(&x350, &x351, x349, x336, x306);
+  uint64_t x351;
+  fiat_p484_mulx_u64(&x350, &x351, x332, UINT64_C(0xfdc1767ae2ffffff));
   uint64_t x352;
-  fiat_p484_uint1 x353;
-  fiat_p484_addcarryx_u64(&x352, &x353, x337, 0x0, x313);
+  uint64_t x353;
+  fiat_p484_mulx_u64(&x352, &x353, x332, UINT64_C(0xffffffffffffffff));
   uint64_t x354;
-  fiat_p484_uint1 x355;
-  fiat_p484_addcarryx_u64(&x354, &x355, x351, x352, (fiat_p484_uint1)x310);
+  uint64_t x355;
+  fiat_p484_mulx_u64(&x354, &x355, x332, UINT64_C(0xffffffffffffffff));
   uint64_t x356;
-  fiat_p484_uint1 x357;
-  fiat_p484_addcarryx_u64(&x356, &x357, 0x0, (arg1[6]), x340);
+  uint64_t x357;
+  fiat_p484_mulx_u64(&x356, &x357, x332, UINT64_C(0xffffffffffffffff));
   uint64_t x358;
   fiat_p484_uint1 x359;
-  fiat_p484_addcarryx_u64(&x358, &x359, x357, 0x0, x342);
+  fiat_p484_addcarryx_u64(&x358, &x359, 0x0, x354, x357);
   uint64_t x360;
   fiat_p484_uint1 x361;
-  fiat_p484_addcarryx_u64(&x360, &x361, x359, 0x0, x344);
+  fiat_p484_addcarryx_u64(&x360, &x361, x359, x352, x355);
   uint64_t x362;
   fiat_p484_uint1 x363;
-  fiat_p484_addcarryx_u64(&x362, &x363, x361, 0x0, x346);
+  fiat_p484_addcarryx_u64(&x362, &x363, x361, x350, x353);
   uint64_t x364;
   fiat_p484_uint1 x365;
-  fiat_p484_addcarryx_u64(&x364, &x365, x363, 0x0, x348);
+  fiat_p484_addcarryx_u64(&x364, &x365, x363, x348, x351);
   uint64_t x366;
   fiat_p484_uint1 x367;
-  fiat_p484_addcarryx_u64(&x366, &x367, x365, 0x0, x350);
+  fiat_p484_addcarryx_u64(&x366, &x367, x365, x346, x349);
   uint64_t x368;
   fiat_p484_uint1 x369;
-  fiat_p484_addcarryx_u64(&x368, &x369, x367, 0x0, x354);
+  fiat_p484_addcarryx_u64(&x368, &x369, x367, x344, x347);
   uint64_t x370;
   fiat_p484_uint1 x371;
-  fiat_p484_addcarryx_u64(&x370, &x371, x355, 0x0, x311);
+  fiat_p484_addcarryx_u64(&x370, &x371, 0x0, x356, x332);
   uint64_t x372;
   fiat_p484_uint1 x373;
-  fiat_p484_addcarryx_u64(&x372, &x373, x369, 0x0, (fiat_p484_uint1)x370);
+  fiat_p484_addcarryx_u64(&x372, &x373, x371, x358, x334);
   uint64_t x374;
-  uint64_t x375;
-  fiat_p484_mulx_u64(&x374, &x375, x356, UINT64_C(0x2341f27177344));
+  fiat_p484_uint1 x375;
+  fiat_p484_addcarryx_u64(&x374, &x375, x373, x360, x336);
   uint64_t x376;
-  uint64_t x377;
-  fiat_p484_mulx_u64(&x376, &x377, x356, UINT64_C(0x6cfc5fd681c52056));
+  fiat_p484_uint1 x377;
+  fiat_p484_addcarryx_u64(&x376, &x377, x375, x362, x338);
   uint64_t x378;
-  uint64_t x379;
-  fiat_p484_mulx_u64(&x378, &x379, x356, UINT64_C(0x7bc65c783158aea3));
+  fiat_p484_uint1 x379;
+  fiat_p484_addcarryx_u64(&x378, &x379, x377, x364, x340);
   uint64_t x380;
-  uint64_t x381;
-  fiat_p484_mulx_u64(&x380, &x381, x356, UINT64_C(0xfdc1767ae2ffffff));
+  fiat_p484_uint1 x381;
+  fiat_p484_addcarryx_u64(&x380, &x381, x379, x366, x342);
   uint64_t x382;
-  uint64_t x383;
-  fiat_p484_mulx_u64(&x382, &x383, x356, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x383;
+  fiat_p484_addcarryx_u64(&x382, &x383, x311, 0x0, x287);
   uint64_t x384;
-  uint64_t x385;
-  fiat_p484_mulx_u64(&x384, &x385, x356, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x385;
+  fiat_p484_addcarryx_u64(&x384, &x385, x331, x382, 0x0);
   uint64_t x386;
-  uint64_t x387;
-  fiat_p484_mulx_u64(&x386, &x387, x356, UINT64_C(0xffffffffffffffff));
+  fiat_p484_uint1 x387;
+  fiat_p484_addcarryx_u64(&x386, &x387, x343, 0x0, x384);
   uint64_t x388;
   fiat_p484_uint1 x389;
-  fiat_p484_addcarryx_u64(&x388, &x389, 0x0, x384, x387);
+  fiat_p484_addcarryx_u64(&x388, &x389, x381, x368, x386);
   uint64_t x390;
   fiat_p484_uint1 x391;
-  fiat_p484_addcarryx_u64(&x390, &x391, x389, x382, x385);
+  fiat_p484_addcarryx_u64(&x390, &x391, x369, 0x0, x345);
   uint64_t x392;
   fiat_p484_uint1 x393;
-  fiat_p484_addcarryx_u64(&x392, &x393, x391, x380, x383);
+  fiat_p484_addcarryx_u64(&x392, &x393, x389, x390, 0x0);
   uint64_t x394;
   fiat_p484_uint1 x395;
-  fiat_p484_addcarryx_u64(&x394, &x395, x393, x378, x381);
+  fiat_p484_subborrowx_u64(&x394, &x395, 0x0, x372, UINT64_C(0xffffffffffffffff));
   uint64_t x396;
   fiat_p484_uint1 x397;
-  fiat_p484_addcarryx_u64(&x396, &x397, x395, x376, x379);
+  fiat_p484_subborrowx_u64(&x396, &x397, x395, x374, UINT64_C(0xffffffffffffffff));
   uint64_t x398;
   fiat_p484_uint1 x399;
-  fiat_p484_addcarryx_u64(&x398, &x399, x397, x374, x377);
+  fiat_p484_subborrowx_u64(&x398, &x399, x397, x376, UINT64_C(0xffffffffffffffff));
   uint64_t x400;
   fiat_p484_uint1 x401;
-  fiat_p484_addcarryx_u64(&x400, &x401, 0x0, x386, x356);
+  fiat_p484_subborrowx_u64(&x400, &x401, x399, x378, UINT64_C(0xfdc1767ae2ffffff));
   uint64_t x402;
   fiat_p484_uint1 x403;
-  fiat_p484_addcarryx_u64(&x402, &x403, x401, x388, x358);
+  fiat_p484_subborrowx_u64(&x402, &x403, x401, x380, UINT64_C(0x7bc65c783158aea3));
   uint64_t x404;
   fiat_p484_uint1 x405;
-  fiat_p484_addcarryx_u64(&x404, &x405, x403, x390, x360);
+  fiat_p484_subborrowx_u64(&x404, &x405, x403, x388, UINT64_C(0x6cfc5fd681c52056));
   uint64_t x406;
   fiat_p484_uint1 x407;
-  fiat_p484_addcarryx_u64(&x406, &x407, x405, x392, x362);
+  fiat_p484_subborrowx_u64(&x406, &x407, x405, x392, UINT64_C(0x2341f27177344));
   uint64_t x408;
   fiat_p484_uint1 x409;
-  fiat_p484_addcarryx_u64(&x408, &x409, x407, x394, x364);
+  fiat_p484_subborrowx_u64(&x408, &x409, x407, 0x0, 0x0);
   uint64_t x410;
-  fiat_p484_uint1 x411;
-  fiat_p484_addcarryx_u64(&x410, &x411, x409, x396, x366);
+  fiat_p484_cmovznz_u64(&x410, x409, x394, x372);
+  uint64_t x411;
+  fiat_p484_cmovznz_u64(&x411, x409, x396, x374);
   uint64_t x412;
-  fiat_p484_uint1 x413;
-  fiat_p484_addcarryx_u64(&x412, &x413, x411, x398, x368);
+  fiat_p484_cmovznz_u64(&x412, x409, x398, x376);
+  uint64_t x413;
+  fiat_p484_cmovznz_u64(&x413, x409, x400, x378);
   uint64_t x414;
-  fiat_p484_uint1 x415;
-  fiat_p484_addcarryx_u64(&x414, &x415, x399, 0x0, x375);
+  fiat_p484_cmovznz_u64(&x414, x409, x402, x380);
+  uint64_t x415;
+  fiat_p484_cmovznz_u64(&x415, x409, x404, x388);
   uint64_t x416;
-  fiat_p484_uint1 x417;
-  fiat_p484_addcarryx_u64(&x416, &x417, x413, x414, (fiat_p484_uint1)x372);
-  uint64_t x418;
-  fiat_p484_uint1 x419;
-  fiat_p484_subborrowx_u64(&x418, &x419, 0x0, x402, UINT64_C(0xffffffffffffffff));
-  uint64_t x420;
-  fiat_p484_uint1 x421;
-  fiat_p484_subborrowx_u64(&x420, &x421, x419, x404, UINT64_C(0xffffffffffffffff));
-  uint64_t x422;
-  fiat_p484_uint1 x423;
-  fiat_p484_subborrowx_u64(&x422, &x423, x421, x406, UINT64_C(0xffffffffffffffff));
-  uint64_t x424;
-  fiat_p484_uint1 x425;
-  fiat_p484_subborrowx_u64(&x424, &x425, x423, x408, UINT64_C(0xfdc1767ae2ffffff));
-  uint64_t x426;
-  fiat_p484_uint1 x427;
-  fiat_p484_subborrowx_u64(&x426, &x427, x425, x410, UINT64_C(0x7bc65c783158aea3));
-  uint64_t x428;
-  fiat_p484_uint1 x429;
-  fiat_p484_subborrowx_u64(&x428, &x429, x427, x412, UINT64_C(0x6cfc5fd681c52056));
-  uint64_t x430;
-  fiat_p484_uint1 x431;
-  fiat_p484_subborrowx_u64(&x430, &x431, x429, x416, UINT64_C(0x2341f27177344));
-  uint64_t x432;
-  fiat_p484_uint1 x433;
-  fiat_p484_addcarryx_u64(&x432, &x433, x417, 0x0, x373);
-  uint64_t x434;
-  fiat_p484_uint1 x435;
-  fiat_p484_subborrowx_u64(&x434, &x435, x431, (fiat_p484_uint1)x432, 0x0);
-  uint64_t x436;
-  fiat_p484_cmovznz_u64(&x436, x435, x418, x402);
-  uint64_t x437;
-  fiat_p484_cmovznz_u64(&x437, x435, x420, x404);
-  uint64_t x438;
-  fiat_p484_cmovznz_u64(&x438, x435, x422, x406);
-  uint64_t x439;
-  fiat_p484_cmovznz_u64(&x439, x435, x424, x408);
-  uint64_t x440;
-  fiat_p484_cmovznz_u64(&x440, x435, x426, x410);
-  uint64_t x441;
-  fiat_p484_cmovznz_u64(&x441, x435, x428, x412);
-  uint64_t x442;
-  fiat_p484_cmovznz_u64(&x442, x435, x430, x416);
-  out1[0] = x436;
-  out1[1] = x437;
-  out1[2] = x438;
-  out1[3] = x439;
-  out1[4] = x440;
-  out1[5] = x441;
-  out1[6] = x442;
+  fiat_p484_cmovznz_u64(&x416, x409, x406, x392);
+  out1[0] = x410;
+  out1[1] = x411;
+  out1[2] = x412;
+  out1[3] = x413;
+  out1[4] = x414;
+  out1[5] = x415;
+  out1[6] = x416;
 }
 
 /*
@@ -2998,28 +2959,28 @@ static void fiat_p484_to_bytes(uint8_t out1[56], const uint64_t arg1[7]) {
   uint8_t x19 = (uint8_t)(x16 & UINT8_C(0xff));
   uint8_t x20 = (uint8_t)(x18 >> 8);
   uint8_t x21 = (uint8_t)(x18 & UINT8_C(0xff));
-  fiat_p484_uint1 x22 = (fiat_p484_uint1)(x20 >> 8);
-  uint8_t x23 = (uint8_t)(x20 & UINT8_C(0xff));
-  uint64_t x24 = (x22 + x6);
-  uint64_t x25 = (x24 >> 8);
-  uint8_t x26 = (uint8_t)(x24 & UINT8_C(0xff));
-  uint64_t x27 = (x25 >> 8);
-  uint8_t x28 = (uint8_t)(x25 & UINT8_C(0xff));
-  uint64_t x29 = (x27 >> 8);
-  uint8_t x30 = (uint8_t)(x27 & UINT8_C(0xff));
-  uint64_t x31 = (x29 >> 8);
-  uint8_t x32 = (uint8_t)(x29 & UINT8_C(0xff));
-  uint64_t x33 = (x31 >> 8);
-  uint8_t x34 = (uint8_t)(x31 & UINT8_C(0xff));
-  uint64_t x35 = (x33 >> 8);
-  uint8_t x36 = (uint8_t)(x33 & UINT8_C(0xff));
-  uint8_t x37 = (uint8_t)(x35 >> 8);
-  uint8_t x38 = (uint8_t)(x35 & UINT8_C(0xff));
-  fiat_p484_uint1 x39 = (fiat_p484_uint1)(x37 >> 8);
-  uint8_t x40 = (uint8_t)(x37 & UINT8_C(0xff));
-  uint64_t x41 = (x39 + x5);
-  uint64_t x42 = (x41 >> 8);
-  uint8_t x43 = (uint8_t)(x41 & UINT8_C(0xff));
+  uint8_t x22 = (uint8_t)(x20 & UINT8_C(0xff));
+  uint64_t x23 = (0x0 + x6);
+  uint64_t x24 = (x23 >> 8);
+  uint8_t x25 = (uint8_t)(x23 & UINT8_C(0xff));
+  uint64_t x26 = (x24 >> 8);
+  uint8_t x27 = (uint8_t)(x24 & UINT8_C(0xff));
+  uint64_t x28 = (x26 >> 8);
+  uint8_t x29 = (uint8_t)(x26 & UINT8_C(0xff));
+  uint64_t x30 = (x28 >> 8);
+  uint8_t x31 = (uint8_t)(x28 & UINT8_C(0xff));
+  uint64_t x32 = (x30 >> 8);
+  uint8_t x33 = (uint8_t)(x30 & UINT8_C(0xff));
+  uint64_t x34 = (x32 >> 8);
+  uint8_t x35 = (uint8_t)(x32 & UINT8_C(0xff));
+  uint8_t x36 = (uint8_t)(x34 >> 8);
+  uint8_t x37 = (uint8_t)(x34 & UINT8_C(0xff));
+  uint8_t x38 = (uint8_t)(x36 & UINT8_C(0xff));
+  uint64_t x39 = (0x0 + x5);
+  uint64_t x40 = (x39 >> 8);
+  uint8_t x41 = (uint8_t)(x39 & UINT8_C(0xff));
+  uint64_t x42 = (x40 >> 8);
+  uint8_t x43 = (uint8_t)(x40 & UINT8_C(0xff));
   uint64_t x44 = (x42 >> 8);
   uint8_t x45 = (uint8_t)(x42 & UINT8_C(0xff));
   uint64_t x46 = (x44 >> 8);
@@ -3028,78 +2989,71 @@ static void fiat_p484_to_bytes(uint8_t out1[56], const uint64_t arg1[7]) {
   uint8_t x49 = (uint8_t)(x46 & UINT8_C(0xff));
   uint64_t x50 = (x48 >> 8);
   uint8_t x51 = (uint8_t)(x48 & UINT8_C(0xff));
-  uint64_t x52 = (x50 >> 8);
+  uint8_t x52 = (uint8_t)(x50 >> 8);
   uint8_t x53 = (uint8_t)(x50 & UINT8_C(0xff));
-  uint8_t x54 = (uint8_t)(x52 >> 8);
-  uint8_t x55 = (uint8_t)(x52 & UINT8_C(0xff));
-  fiat_p484_uint1 x56 = (fiat_p484_uint1)(x54 >> 8);
-  uint8_t x57 = (uint8_t)(x54 & UINT8_C(0xff));
-  uint64_t x58 = (x56 + x4);
-  uint64_t x59 = (x58 >> 8);
-  uint8_t x60 = (uint8_t)(x58 & UINT8_C(0xff));
-  uint64_t x61 = (x59 >> 8);
-  uint8_t x62 = (uint8_t)(x59 & UINT8_C(0xff));
-  uint64_t x63 = (x61 >> 8);
-  uint8_t x64 = (uint8_t)(x61 & UINT8_C(0xff));
-  uint64_t x65 = (x63 >> 8);
-  uint8_t x66 = (uint8_t)(x63 & UINT8_C(0xff));
-  uint64_t x67 = (x65 >> 8);
-  uint8_t x68 = (uint8_t)(x65 & UINT8_C(0xff));
-  uint64_t x69 = (x67 >> 8);
-  uint8_t x70 = (uint8_t)(x67 & UINT8_C(0xff));
-  uint8_t x71 = (uint8_t)(x69 >> 8);
-  uint8_t x72 = (uint8_t)(x69 & UINT8_C(0xff));
-  fiat_p484_uint1 x73 = (fiat_p484_uint1)(x71 >> 8);
-  uint8_t x74 = (uint8_t)(x71 & UINT8_C(0xff));
-  uint64_t x75 = (x73 + x3);
-  uint64_t x76 = (x75 >> 8);
-  uint8_t x77 = (uint8_t)(x75 & UINT8_C(0xff));
+  uint8_t x54 = (uint8_t)(x52 & UINT8_C(0xff));
+  uint64_t x55 = (0x0 + x4);
+  uint64_t x56 = (x55 >> 8);
+  uint8_t x57 = (uint8_t)(x55 & UINT8_C(0xff));
+  uint64_t x58 = (x56 >> 8);
+  uint8_t x59 = (uint8_t)(x56 & UINT8_C(0xff));
+  uint64_t x60 = (x58 >> 8);
+  uint8_t x61 = (uint8_t)(x58 & UINT8_C(0xff));
+  uint64_t x62 = (x60 >> 8);
+  uint8_t x63 = (uint8_t)(x60 & UINT8_C(0xff));
+  uint64_t x64 = (x62 >> 8);
+  uint8_t x65 = (uint8_t)(x62 & UINT8_C(0xff));
+  uint64_t x66 = (x64 >> 8);
+  uint8_t x67 = (uint8_t)(x64 & UINT8_C(0xff));
+  uint8_t x68 = (uint8_t)(x66 >> 8);
+  uint8_t x69 = (uint8_t)(x66 & UINT8_C(0xff));
+  uint8_t x70 = (uint8_t)(x68 & UINT8_C(0xff));
+  uint64_t x71 = (0x0 + x3);
+  uint64_t x72 = (x71 >> 8);
+  uint8_t x73 = (uint8_t)(x71 & UINT8_C(0xff));
+  uint64_t x74 = (x72 >> 8);
+  uint8_t x75 = (uint8_t)(x72 & UINT8_C(0xff));
+  uint64_t x76 = (x74 >> 8);
+  uint8_t x77 = (uint8_t)(x74 & UINT8_C(0xff));
   uint64_t x78 = (x76 >> 8);
   uint8_t x79 = (uint8_t)(x76 & UINT8_C(0xff));
   uint64_t x80 = (x78 >> 8);
   uint8_t x81 = (uint8_t)(x78 & UINT8_C(0xff));
   uint64_t x82 = (x80 >> 8);
   uint8_t x83 = (uint8_t)(x80 & UINT8_C(0xff));
-  uint64_t x84 = (x82 >> 8);
+  uint8_t x84 = (uint8_t)(x82 >> 8);
   uint8_t x85 = (uint8_t)(x82 & UINT8_C(0xff));
-  uint64_t x86 = (x84 >> 8);
-  uint8_t x87 = (uint8_t)(x84 & UINT8_C(0xff));
-  uint8_t x88 = (uint8_t)(x86 >> 8);
-  uint8_t x89 = (uint8_t)(x86 & UINT8_C(0xff));
-  fiat_p484_uint1 x90 = (fiat_p484_uint1)(x88 >> 8);
+  uint8_t x86 = (uint8_t)(x84 & UINT8_C(0xff));
+  uint64_t x87 = (0x0 + x2);
+  uint64_t x88 = (x87 >> 8);
+  uint8_t x89 = (uint8_t)(x87 & UINT8_C(0xff));
+  uint64_t x90 = (x88 >> 8);
   uint8_t x91 = (uint8_t)(x88 & UINT8_C(0xff));
-  uint64_t x92 = (x90 + x2);
-  uint64_t x93 = (x92 >> 8);
-  uint8_t x94 = (uint8_t)(x92 & UINT8_C(0xff));
-  uint64_t x95 = (x93 >> 8);
-  uint8_t x96 = (uint8_t)(x93 & UINT8_C(0xff));
-  uint64_t x97 = (x95 >> 8);
-  uint8_t x98 = (uint8_t)(x95 & UINT8_C(0xff));
-  uint64_t x99 = (x97 >> 8);
-  uint8_t x100 = (uint8_t)(x97 & UINT8_C(0xff));
-  uint64_t x101 = (x99 >> 8);
-  uint8_t x102 = (uint8_t)(x99 & UINT8_C(0xff));
-  uint64_t x103 = (x101 >> 8);
-  uint8_t x104 = (uint8_t)(x101 & UINT8_C(0xff));
-  uint8_t x105 = (uint8_t)(x103 >> 8);
-  uint8_t x106 = (uint8_t)(x103 & UINT8_C(0xff));
-  fiat_p484_uint1 x107 = (fiat_p484_uint1)(x105 >> 8);
-  uint8_t x108 = (uint8_t)(x105 & UINT8_C(0xff));
-  uint64_t x109 = (x107 + x1);
-  uint64_t x110 = (x109 >> 8);
-  uint8_t x111 = (uint8_t)(x109 & UINT8_C(0xff));
+  uint64_t x92 = (x90 >> 8);
+  uint8_t x93 = (uint8_t)(x90 & UINT8_C(0xff));
+  uint64_t x94 = (x92 >> 8);
+  uint8_t x95 = (uint8_t)(x92 & UINT8_C(0xff));
+  uint64_t x96 = (x94 >> 8);
+  uint8_t x97 = (uint8_t)(x94 & UINT8_C(0xff));
+  uint64_t x98 = (x96 >> 8);
+  uint8_t x99 = (uint8_t)(x96 & UINT8_C(0xff));
+  uint8_t x100 = (uint8_t)(x98 >> 8);
+  uint8_t x101 = (uint8_t)(x98 & UINT8_C(0xff));
+  uint8_t x102 = (uint8_t)(x100 & UINT8_C(0xff));
+  uint64_t x103 = (0x0 + x1);
+  uint64_t x104 = (x103 >> 8);
+  uint8_t x105 = (uint8_t)(x103 & UINT8_C(0xff));
+  uint64_t x106 = (x104 >> 8);
+  uint8_t x107 = (uint8_t)(x104 & UINT8_C(0xff));
+  uint64_t x108 = (x106 >> 8);
+  uint8_t x109 = (uint8_t)(x106 & UINT8_C(0xff));
+  uint64_t x110 = (x108 >> 8);
+  uint8_t x111 = (uint8_t)(x108 & UINT8_C(0xff));
   uint64_t x112 = (x110 >> 8);
   uint8_t x113 = (uint8_t)(x110 & UINT8_C(0xff));
-  uint64_t x114 = (x112 >> 8);
+  uint8_t x114 = (uint8_t)(x112 >> 8);
   uint8_t x115 = (uint8_t)(x112 & UINT8_C(0xff));
-  uint64_t x116 = (x114 >> 8);
-  uint8_t x117 = (uint8_t)(x114 & UINT8_C(0xff));
-  uint64_t x118 = (x116 >> 8);
-  uint8_t x119 = (uint8_t)(x116 & UINT8_C(0xff));
-  uint8_t x120 = (uint8_t)(x118 >> 8);
-  uint8_t x121 = (uint8_t)(x118 & UINT8_C(0xff));
-  fiat_p484_uint1 x122 = (fiat_p484_uint1)(x120 >> 8);
-  uint8_t x123 = (uint8_t)(x120 & UINT8_C(0xff));
+  uint8_t x116 = (uint8_t)(x114 & UINT8_C(0xff));
   out1[0] = x9;
   out1[1] = x11;
   out1[2] = x13;
@@ -3107,55 +3061,55 @@ static void fiat_p484_to_bytes(uint8_t out1[56], const uint64_t arg1[7]) {
   out1[4] = x17;
   out1[5] = x19;
   out1[6] = x21;
-  out1[7] = x23;
-  out1[8] = x26;
-  out1[9] = x28;
-  out1[10] = x30;
-  out1[11] = x32;
-  out1[12] = x34;
-  out1[13] = x36;
-  out1[14] = x38;
-  out1[15] = x40;
-  out1[16] = x43;
-  out1[17] = x45;
-  out1[18] = x47;
-  out1[19] = x49;
-  out1[20] = x51;
-  out1[21] = x53;
-  out1[22] = x55;
-  out1[23] = x57;
-  out1[24] = x60;
-  out1[25] = x62;
-  out1[26] = x64;
-  out1[27] = x66;
-  out1[28] = x68;
-  out1[29] = x70;
-  out1[30] = x72;
-  out1[31] = x74;
-  out1[32] = x77;
-  out1[33] = x79;
-  out1[34] = x81;
-  out1[35] = x83;
-  out1[36] = x85;
-  out1[37] = x87;
-  out1[38] = x89;
-  out1[39] = x91;
-  out1[40] = x94;
-  out1[41] = x96;
-  out1[42] = x98;
-  out1[43] = x100;
-  out1[44] = x102;
-  out1[45] = x104;
-  out1[46] = x106;
-  out1[47] = x108;
-  out1[48] = x111;
-  out1[49] = x113;
-  out1[50] = x115;
-  out1[51] = x117;
-  out1[52] = x119;
-  out1[53] = x121;
-  out1[54] = x123;
-  out1[55] = x122;
+  out1[7] = x22;
+  out1[8] = x25;
+  out1[9] = x27;
+  out1[10] = x29;
+  out1[11] = x31;
+  out1[12] = x33;
+  out1[13] = x35;
+  out1[14] = x37;
+  out1[15] = x38;
+  out1[16] = x41;
+  out1[17] = x43;
+  out1[18] = x45;
+  out1[19] = x47;
+  out1[20] = x49;
+  out1[21] = x51;
+  out1[22] = x53;
+  out1[23] = x54;
+  out1[24] = x57;
+  out1[25] = x59;
+  out1[26] = x61;
+  out1[27] = x63;
+  out1[28] = x65;
+  out1[29] = x67;
+  out1[30] = x69;
+  out1[31] = x70;
+  out1[32] = x73;
+  out1[33] = x75;
+  out1[34] = x77;
+  out1[35] = x79;
+  out1[36] = x81;
+  out1[37] = x83;
+  out1[38] = x85;
+  out1[39] = x86;
+  out1[40] = x89;
+  out1[41] = x91;
+  out1[42] = x93;
+  out1[43] = x95;
+  out1[44] = x97;
+  out1[45] = x99;
+  out1[46] = x101;
+  out1[47] = x102;
+  out1[48] = x105;
+  out1[49] = x107;
+  out1[50] = x109;
+  out1[51] = x111;
+  out1[52] = x113;
+  out1[53] = x115;
+  out1[54] = x116;
+  out1[55] = 0x0;
 }
 
 /*
@@ -3165,93 +3119,86 @@ static void fiat_p484_to_bytes(uint8_t out1[56], const uint64_t arg1[7]) {
  *   out1: [[0x0 ~> 0xffffffffffffffff], [0x0 ~> 0xffffffffffffffff], [0x0 ~> 0xffffffffffffffff], [0x0 ~> 0xffffffffffffffff], [0x0 ~> 0xffffffffffffffff], [0x0 ~> 0xffffffffffffffff], [0x0 ~> 0x3ffffffffffff]]
  */
 static void fiat_p484_from_bytes(uint64_t out1[7], const uint8_t arg1[56]) {
-  fiat_p484_uint1 x1 = (fiat_p484_uint1)((uint64_t)(fiat_p484_uint1)(arg1[55]) << 56);
-  uint64_t x2 = ((uint64_t)(arg1[54]) << 48);
-  uint64_t x3 = ((uint64_t)(arg1[53]) << 40);
-  uint64_t x4 = ((uint64_t)(arg1[52]) << 32);
-  uint64_t x5 = ((uint64_t)(arg1[51]) << 24);
-  uint64_t x6 = ((uint64_t)(arg1[50]) << 16);
-  uint64_t x7 = ((uint64_t)(arg1[49]) << 8);
-  uint8_t x8 = (arg1[48]);
-  uint64_t x9 = ((uint64_t)(arg1[47]) << 56);
-  uint64_t x10 = ((uint64_t)(arg1[46]) << 48);
-  uint64_t x11 = ((uint64_t)(arg1[45]) << 40);
-  uint64_t x12 = ((uint64_t)(arg1[44]) << 32);
-  uint64_t x13 = ((uint64_t)(arg1[43]) << 24);
-  uint64_t x14 = ((uint64_t)(arg1[42]) << 16);
-  uint64_t x15 = ((uint64_t)(arg1[41]) << 8);
-  uint8_t x16 = (arg1[40]);
-  uint64_t x17 = ((uint64_t)(arg1[39]) << 56);
-  uint64_t x18 = ((uint64_t)(arg1[38]) << 48);
-  uint64_t x19 = ((uint64_t)(arg1[37]) << 40);
-  uint64_t x20 = ((uint64_t)(arg1[36]) << 32);
-  uint64_t x21 = ((uint64_t)(arg1[35]) << 24);
-  uint64_t x22 = ((uint64_t)(arg1[34]) << 16);
-  uint64_t x23 = ((uint64_t)(arg1[33]) << 8);
-  uint8_t x24 = (arg1[32]);
-  uint64_t x25 = ((uint64_t)(arg1[31]) << 56);
-  uint64_t x26 = ((uint64_t)(arg1[30]) << 48);
-  uint64_t x27 = ((uint64_t)(arg1[29]) << 40);
-  uint64_t x28 = ((uint64_t)(arg1[28]) << 32);
-  uint64_t x29 = ((uint64_t)(arg1[27]) << 24);
-  uint64_t x30 = ((uint64_t)(arg1[26]) << 16);
-  uint64_t x31 = ((uint64_t)(arg1[25]) << 8);
-  uint8_t x32 = (arg1[24]);
-  uint64_t x33 = ((uint64_t)(arg1[23]) << 56);
-  uint64_t x34 = ((uint64_t)(arg1[22]) << 48);
-  uint64_t x35 = ((uint64_t)(arg1[21]) << 40);
-  uint64_t x36 = ((uint64_t)(arg1[20]) << 32);
-  uint64_t x37 = ((uint64_t)(arg1[19]) << 24);
-  uint64_t x38 = ((uint64_t)(arg1[18]) << 16);
-  uint64_t x39 = ((uint64_t)(arg1[17]) << 8);
-  uint8_t x40 = (arg1[16]);
-  uint64_t x41 = ((uint64_t)(arg1[15]) << 56);
-  uint64_t x42 = ((uint64_t)(arg1[14]) << 48);
-  uint64_t x43 = ((uint64_t)(arg1[13]) << 40);
-  uint64_t x44 = ((uint64_t)(arg1[12]) << 32);
-  uint64_t x45 = ((uint64_t)(arg1[11]) << 24);
-  uint64_t x46 = ((uint64_t)(arg1[10]) << 16);
-  uint64_t x47 = ((uint64_t)(arg1[9]) << 8);
-  uint8_t x48 = (arg1[8]);
-  uint64_t x49 = ((uint64_t)(arg1[7]) << 56);
-  uint64_t x50 = ((uint64_t)(arg1[6]) << 48);
-  uint64_t x51 = ((uint64_t)(arg1[5]) << 40);
-  uint64_t x52 = ((uint64_t)(arg1[4]) << 32);
-  uint64_t x53 = ((uint64_t)(arg1[3]) << 24);
-  uint64_t x54 = ((uint64_t)(arg1[2]) << 16);
-  uint64_t x55 = ((uint64_t)(arg1[1]) << 8);
-  uint8_t x56 = (arg1[0]);
-  uint64_t x57 = (x56 + (x55 + (x54 + (x53 + (x52 + (x51 + (x50 + x49)))))));
-  fiat_p484_uint1 x58 = (fiat_p484_uint1)((fiat_p484_uint128)x57 >> 64);
-  uint64_t x59 = (x57 & UINT64_C(0xffffffffffffffff));
-  uint64_t x60 = (x8 + (x7 + (x6 + (x5 + (x4 + (x3 + (x2 + (uint64_t)x1)))))));
-  uint64_t x61 = (x16 + (x15 + (x14 + (x13 + (x12 + (x11 + (x10 + x9)))))));
-  uint64_t x62 = (x24 + (x23 + (x22 + (x21 + (x20 + (x19 + (x18 + x17)))))));
-  uint64_t x63 = (x32 + (x31 + (x30 + (x29 + (x28 + (x27 + (x26 + x25)))))));
-  uint64_t x64 = (x40 + (x39 + (x38 + (x37 + (x36 + (x35 + (x34 + x33)))))));
-  uint64_t x65 = (x48 + (x47 + (x46 + (x45 + (x44 + (x43 + (x42 + x41)))))));
-  uint64_t x66 = (x58 + x65);
-  fiat_p484_uint1 x67 = (fiat_p484_uint1)((fiat_p484_uint128)x66 >> 64);
-  uint64_t x68 = (x66 & UINT64_C(0xffffffffffffffff));
-  uint64_t x69 = (x67 + x64);
-  fiat_p484_uint1 x70 = (fiat_p484_uint1)((fiat_p484_uint128)x69 >> 64);
-  uint64_t x71 = (x69 & UINT64_C(0xffffffffffffffff));
-  uint64_t x72 = (x70 + x63);
-  fiat_p484_uint1 x73 = (fiat_p484_uint1)((fiat_p484_uint128)x72 >> 64);
-  uint64_t x74 = (x72 & UINT64_C(0xffffffffffffffff));
-  uint64_t x75 = (x73 + x62);
-  fiat_p484_uint1 x76 = (fiat_p484_uint1)((fiat_p484_uint128)x75 >> 64);
-  uint64_t x77 = (x75 & UINT64_C(0xffffffffffffffff));
-  uint64_t x78 = (x76 + x61);
-  fiat_p484_uint1 x79 = (fiat_p484_uint1)((fiat_p484_uint128)x78 >> 64);
-  uint64_t x80 = (x78 & UINT64_C(0xffffffffffffffff));
-  uint64_t x81 = (x79 + x60);
-  out1[0] = x59;
-  out1[1] = x68;
-  out1[2] = x71;
-  out1[3] = x74;
-  out1[4] = x77;
-  out1[5] = x80;
-  out1[6] = x81;
+  uint64_t x1 = ((uint64_t)(arg1[54]) << 48);
+  uint64_t x2 = ((uint64_t)(arg1[53]) << 40);
+  uint64_t x3 = ((uint64_t)(arg1[52]) << 32);
+  uint64_t x4 = ((uint64_t)(arg1[51]) << 24);
+  uint64_t x5 = ((uint64_t)(arg1[50]) << 16);
+  uint64_t x6 = ((uint64_t)(arg1[49]) << 8);
+  uint8_t x7 = (arg1[48]);
+  uint64_t x8 = ((uint64_t)(arg1[47]) << 56);
+  uint64_t x9 = ((uint64_t)(arg1[46]) << 48);
+  uint64_t x10 = ((uint64_t)(arg1[45]) << 40);
+  uint64_t x11 = ((uint64_t)(arg1[44]) << 32);
+  uint64_t x12 = ((uint64_t)(arg1[43]) << 24);
+  uint64_t x13 = ((uint64_t)(arg1[42]) << 16);
+  uint64_t x14 = ((uint64_t)(arg1[41]) << 8);
+  uint8_t x15 = (arg1[40]);
+  uint64_t x16 = ((uint64_t)(arg1[39]) << 56);
+  uint64_t x17 = ((uint64_t)(arg1[38]) << 48);
+  uint64_t x18 = ((uint64_t)(arg1[37]) << 40);
+  uint64_t x19 = ((uint64_t)(arg1[36]) << 32);
+  uint64_t x20 = ((uint64_t)(arg1[35]) << 24);
+  uint64_t x21 = ((uint64_t)(arg1[34]) << 16);
+  uint64_t x22 = ((uint64_t)(arg1[33]) << 8);
+  uint8_t x23 = (arg1[32]);
+  uint64_t x24 = ((uint64_t)(arg1[31]) << 56);
+  uint64_t x25 = ((uint64_t)(arg1[30]) << 48);
+  uint64_t x26 = ((uint64_t)(arg1[29]) << 40);
+  uint64_t x27 = ((uint64_t)(arg1[28]) << 32);
+  uint64_t x28 = ((uint64_t)(arg1[27]) << 24);
+  uint64_t x29 = ((uint64_t)(arg1[26]) << 16);
+  uint64_t x30 = ((uint64_t)(arg1[25]) << 8);
+  uint8_t x31 = (arg1[24]);
+  uint64_t x32 = ((uint64_t)(arg1[23]) << 56);
+  uint64_t x33 = ((uint64_t)(arg1[22]) << 48);
+  uint64_t x34 = ((uint64_t)(arg1[21]) << 40);
+  uint64_t x35 = ((uint64_t)(arg1[20]) << 32);
+  uint64_t x36 = ((uint64_t)(arg1[19]) << 24);
+  uint64_t x37 = ((uint64_t)(arg1[18]) << 16);
+  uint64_t x38 = ((uint64_t)(arg1[17]) << 8);
+  uint8_t x39 = (arg1[16]);
+  uint64_t x40 = ((uint64_t)(arg1[15]) << 56);
+  uint64_t x41 = ((uint64_t)(arg1[14]) << 48);
+  uint64_t x42 = ((uint64_t)(arg1[13]) << 40);
+  uint64_t x43 = ((uint64_t)(arg1[12]) << 32);
+  uint64_t x44 = ((uint64_t)(arg1[11]) << 24);
+  uint64_t x45 = ((uint64_t)(arg1[10]) << 16);
+  uint64_t x46 = ((uint64_t)(arg1[9]) << 8);
+  uint8_t x47 = (arg1[8]);
+  uint64_t x48 = ((uint64_t)(arg1[7]) << 56);
+  uint64_t x49 = ((uint64_t)(arg1[6]) << 48);
+  uint64_t x50 = ((uint64_t)(arg1[5]) << 40);
+  uint64_t x51 = ((uint64_t)(arg1[4]) << 32);
+  uint64_t x52 = ((uint64_t)(arg1[3]) << 24);
+  uint64_t x53 = ((uint64_t)(arg1[2]) << 16);
+  uint64_t x54 = ((uint64_t)(arg1[1]) << 8);
+  uint8_t x55 = (arg1[0]);
+  uint64_t x56 = (x55 + (x54 + (x53 + (x52 + (x51 + (x50 + (x49 + x48)))))));
+  uint64_t x57 = (x56 & UINT64_C(0xffffffffffffffff));
+  uint64_t x58 = (x7 + (x6 + (x5 + (x4 + (x3 + (x2 + (x1 + (uint64_t)0x0)))))));
+  uint64_t x59 = (x15 + (x14 + (x13 + (x12 + (x11 + (x10 + (x9 + x8)))))));
+  uint64_t x60 = (x23 + (x22 + (x21 + (x20 + (x19 + (x18 + (x17 + x16)))))));
+  uint64_t x61 = (x31 + (x30 + (x29 + (x28 + (x27 + (x26 + (x25 + x24)))))));
+  uint64_t x62 = (x39 + (x38 + (x37 + (x36 + (x35 + (x34 + (x33 + x32)))))));
+  uint64_t x63 = (x47 + (x46 + (x45 + (x44 + (x43 + (x42 + (x41 + x40)))))));
+  uint64_t x64 = (0x0 + x63);
+  uint64_t x65 = (x64 & UINT64_C(0xffffffffffffffff));
+  uint64_t x66 = (0x0 + x62);
+  uint64_t x67 = (x66 & UINT64_C(0xffffffffffffffff));
+  uint64_t x68 = (0x0 + x61);
+  uint64_t x69 = (x68 & UINT64_C(0xffffffffffffffff));
+  uint64_t x70 = (0x0 + x60);
+  uint64_t x71 = (x70 & UINT64_C(0xffffffffffffffff));
+  uint64_t x72 = (0x0 + x59);
+  uint64_t x73 = (x72 & UINT64_C(0xffffffffffffffff));
+  uint64_t x74 = (0x0 + x58);
+  out1[0] = x57;
+  out1[1] = x65;
+  out1[2] = x67;
+  out1[3] = x69;
+  out1[4] = x71;
+  out1[5] = x73;
+  out1[6] = x74;
 }
 
diff --git a/p521_32.c b/p521_32.c
index d1ae368c9..b044dc804 100644
--- a/p521_32.c
+++ b/p521_32.c
@@ -1200,103 +1200,101 @@ static void fiat_p521_to_bytes(uint8_t out1[66], const uint32_t arg1[17]) {
   uint8_t x130 = (uint8_t)(x127 & UINT8_C(0xff));
   uint8_t x131 = (uint8_t)(x129 >> 8);
   uint8_t x132 = (uint8_t)(x129 & UINT8_C(0xff));
-  fiat_p521_uint1 x133 = (fiat_p521_uint1)(x131 >> 8);
-  uint8_t x134 = (uint8_t)(x131 & UINT8_C(0xff));
-  uint32_t x135 = (x133 + x48);
-  uint32_t x136 = (x135 >> 8);
-  uint8_t x137 = (uint8_t)(x135 & UINT8_C(0xff));
-  uint32_t x138 = (x136 >> 8);
-  uint8_t x139 = (uint8_t)(x136 & UINT8_C(0xff));
-  uint8_t x140 = (uint8_t)(x138 >> 8);
-  uint8_t x141 = (uint8_t)(x138 & UINT8_C(0xff));
-  uint64_t x142 = (x140 + x78);
-  uint32_t x143 = (uint32_t)(x142 >> 8);
-  uint8_t x144 = (uint8_t)(x142 & UINT8_C(0xff));
-  uint32_t x145 = (x143 >> 8);
-  uint8_t x146 = (uint8_t)(x143 & UINT8_C(0xff));
-  uint32_t x147 = (x145 >> 8);
-  uint8_t x148 = (uint8_t)(x145 & UINT8_C(0xff));
-  uint8_t x149 = (uint8_t)(x147 >> 8);
-  uint8_t x150 = (uint8_t)(x147 & UINT8_C(0xff));
-  uint64_t x151 = (x149 + x77);
-  uint32_t x152 = (uint32_t)(x151 >> 8);
-  uint8_t x153 = (uint8_t)(x151 & UINT8_C(0xff));
-  uint32_t x154 = (x152 >> 8);
-  uint8_t x155 = (uint8_t)(x152 & UINT8_C(0xff));
-  uint32_t x156 = (x154 >> 8);
-  uint8_t x157 = (uint8_t)(x154 & UINT8_C(0xff));
-  uint8_t x158 = (uint8_t)(x156 >> 8);
-  uint8_t x159 = (uint8_t)(x156 & UINT8_C(0xff));
-  uint64_t x160 = (x158 + x76);
-  uint32_t x161 = (uint32_t)(x160 >> 8);
-  uint8_t x162 = (uint8_t)(x160 & UINT8_C(0xff));
-  uint32_t x163 = (x161 >> 8);
-  uint8_t x164 = (uint8_t)(x161 & UINT8_C(0xff));
-  uint32_t x165 = (x163 >> 8);
-  uint8_t x166 = (uint8_t)(x163 & UINT8_C(0xff));
-  uint8_t x167 = (uint8_t)(x165 >> 8);
-  uint8_t x168 = (uint8_t)(x165 & UINT8_C(0xff));
-  uint64_t x169 = (x167 + x75);
-  uint32_t x170 = (uint32_t)(x169 >> 8);
-  uint8_t x171 = (uint8_t)(x169 & UINT8_C(0xff));
-  uint32_t x172 = (x170 >> 8);
-  uint8_t x173 = (uint8_t)(x170 & UINT8_C(0xff));
-  uint32_t x174 = (x172 >> 8);
-  uint8_t x175 = (uint8_t)(x172 & UINT8_C(0xff));
-  uint8_t x176 = (uint8_t)(x174 >> 8);
-  uint8_t x177 = (uint8_t)(x174 & UINT8_C(0xff));
-  uint32_t x178 = (x176 + x74);
-  uint32_t x179 = (x178 >> 8);
-  uint8_t x180 = (uint8_t)(x178 & UINT8_C(0xff));
-  uint32_t x181 = (x179 >> 8);
-  uint8_t x182 = (uint8_t)(x179 & UINT8_C(0xff));
-  uint8_t x183 = (uint8_t)(x181 >> 8);
-  uint8_t x184 = (uint8_t)(x181 & UINT8_C(0xff));
-  fiat_p521_uint1 x185 = (fiat_p521_uint1)(x183 >> 8);
-  uint8_t x186 = (uint8_t)(x183 & UINT8_C(0xff));
-  uint32_t x187 = (x185 + x60);
-  uint32_t x188 = (x187 >> 8);
-  uint8_t x189 = (uint8_t)(x187 & UINT8_C(0xff));
-  uint32_t x190 = (x188 >> 8);
+  uint8_t x133 = (uint8_t)(x131 & UINT8_C(0xff));
+  uint32_t x134 = (0x0 + x48);
+  uint32_t x135 = (x134 >> 8);
+  uint8_t x136 = (uint8_t)(x134 & UINT8_C(0xff));
+  uint32_t x137 = (x135 >> 8);
+  uint8_t x138 = (uint8_t)(x135 & UINT8_C(0xff));
+  uint8_t x139 = (uint8_t)(x137 >> 8);
+  uint8_t x140 = (uint8_t)(x137 & UINT8_C(0xff));
+  uint64_t x141 = (x139 + x78);
+  uint32_t x142 = (uint32_t)(x141 >> 8);
+  uint8_t x143 = (uint8_t)(x141 & UINT8_C(0xff));
+  uint32_t x144 = (x142 >> 8);
+  uint8_t x145 = (uint8_t)(x142 & UINT8_C(0xff));
+  uint32_t x146 = (x144 >> 8);
+  uint8_t x147 = (uint8_t)(x144 & UINT8_C(0xff));
+  uint8_t x148 = (uint8_t)(x146 >> 8);
+  uint8_t x149 = (uint8_t)(x146 & UINT8_C(0xff));
+  uint64_t x150 = (x148 + x77);
+  uint32_t x151 = (uint32_t)(x150 >> 8);
+  uint8_t x152 = (uint8_t)(x150 & UINT8_C(0xff));
+  uint32_t x153 = (x151 >> 8);
+  uint8_t x154 = (uint8_t)(x151 & UINT8_C(0xff));
+  uint32_t x155 = (x153 >> 8);
+  uint8_t x156 = (uint8_t)(x153 & UINT8_C(0xff));
+  uint8_t x157 = (uint8_t)(x155 >> 8);
+  uint8_t x158 = (uint8_t)(x155 & UINT8_C(0xff));
+  uint64_t x159 = (x157 + x76);
+  uint32_t x160 = (uint32_t)(x159 >> 8);
+  uint8_t x161 = (uint8_t)(x159 & UINT8_C(0xff));
+  uint32_t x162 = (x160 >> 8);
+  uint8_t x163 = (uint8_t)(x160 & UINT8_C(0xff));
+  uint32_t x164 = (x162 >> 8);
+  uint8_t x165 = (uint8_t)(x162 & UINT8_C(0xff));
+  uint8_t x166 = (uint8_t)(x164 >> 8);
+  uint8_t x167 = (uint8_t)(x164 & UINT8_C(0xff));
+  uint64_t x168 = (x166 + x75);
+  uint32_t x169 = (uint32_t)(x168 >> 8);
+  uint8_t x170 = (uint8_t)(x168 & UINT8_C(0xff));
+  uint32_t x171 = (x169 >> 8);
+  uint8_t x172 = (uint8_t)(x169 & UINT8_C(0xff));
+  uint32_t x173 = (x171 >> 8);
+  uint8_t x174 = (uint8_t)(x171 & UINT8_C(0xff));
+  uint8_t x175 = (uint8_t)(x173 >> 8);
+  uint8_t x176 = (uint8_t)(x173 & UINT8_C(0xff));
+  uint32_t x177 = (x175 + x74);
+  uint32_t x178 = (x177 >> 8);
+  uint8_t x179 = (uint8_t)(x177 & UINT8_C(0xff));
+  uint32_t x180 = (x178 >> 8);
+  uint8_t x181 = (uint8_t)(x178 & UINT8_C(0xff));
+  uint8_t x182 = (uint8_t)(x180 >> 8);
+  uint8_t x183 = (uint8_t)(x180 & UINT8_C(0xff));
+  uint8_t x184 = (uint8_t)(x182 & UINT8_C(0xff));
+  uint32_t x185 = (0x0 + x60);
+  uint32_t x186 = (x185 >> 8);
+  uint8_t x187 = (uint8_t)(x185 & UINT8_C(0xff));
+  uint32_t x188 = (x186 >> 8);
+  uint8_t x189 = (uint8_t)(x186 & UINT8_C(0xff));
+  uint8_t x190 = (uint8_t)(x188 >> 8);
   uint8_t x191 = (uint8_t)(x188 & UINT8_C(0xff));
-  uint8_t x192 = (uint8_t)(x190 >> 8);
-  uint8_t x193 = (uint8_t)(x190 & UINT8_C(0xff));
-  uint64_t x194 = (x192 + x73);
-  uint32_t x195 = (uint32_t)(x194 >> 8);
-  uint8_t x196 = (uint8_t)(x194 & UINT8_C(0xff));
+  uint64_t x192 = (x190 + x73);
+  uint32_t x193 = (uint32_t)(x192 >> 8);
+  uint8_t x194 = (uint8_t)(x192 & UINT8_C(0xff));
+  uint32_t x195 = (x193 >> 8);
+  uint8_t x196 = (uint8_t)(x193 & UINT8_C(0xff));
   uint32_t x197 = (x195 >> 8);
   uint8_t x198 = (uint8_t)(x195 & UINT8_C(0xff));
-  uint32_t x199 = (x197 >> 8);
+  uint8_t x199 = (uint8_t)(x197 >> 8);
   uint8_t x200 = (uint8_t)(x197 & UINT8_C(0xff));
-  uint8_t x201 = (uint8_t)(x199 >> 8);
-  uint8_t x202 = (uint8_t)(x199 & UINT8_C(0xff));
-  uint64_t x203 = (x201 + x72);
-  uint32_t x204 = (uint32_t)(x203 >> 8);
-  uint8_t x205 = (uint8_t)(x203 & UINT8_C(0xff));
+  uint64_t x201 = (x199 + x72);
+  uint32_t x202 = (uint32_t)(x201 >> 8);
+  uint8_t x203 = (uint8_t)(x201 & UINT8_C(0xff));
+  uint32_t x204 = (x202 >> 8);
+  uint8_t x205 = (uint8_t)(x202 & UINT8_C(0xff));
   uint32_t x206 = (x204 >> 8);
   uint8_t x207 = (uint8_t)(x204 & UINT8_C(0xff));
-  uint32_t x208 = (x206 >> 8);
+  uint8_t x208 = (uint8_t)(x206 >> 8);
   uint8_t x209 = (uint8_t)(x206 & UINT8_C(0xff));
-  uint8_t x210 = (uint8_t)(x208 >> 8);
-  uint8_t x211 = (uint8_t)(x208 & UINT8_C(0xff));
-  uint64_t x212 = (x210 + x71);
-  uint32_t x213 = (uint32_t)(x212 >> 8);
-  uint8_t x214 = (uint8_t)(x212 & UINT8_C(0xff));
+  uint64_t x210 = (x208 + x71);
+  uint32_t x211 = (uint32_t)(x210 >> 8);
+  uint8_t x212 = (uint8_t)(x210 & UINT8_C(0xff));
+  uint32_t x213 = (x211 >> 8);
+  uint8_t x214 = (uint8_t)(x211 & UINT8_C(0xff));
   uint32_t x215 = (x213 >> 8);
   uint8_t x216 = (uint8_t)(x213 & UINT8_C(0xff));
-  uint32_t x217 = (x215 >> 8);
+  uint8_t x217 = (uint8_t)(x215 >> 8);
   uint8_t x218 = (uint8_t)(x215 & UINT8_C(0xff));
-  uint8_t x219 = (uint8_t)(x217 >> 8);
-  uint8_t x220 = (uint8_t)(x217 & UINT8_C(0xff));
-  uint64_t x221 = (x219 + x70);
-  uint32_t x222 = (uint32_t)(x221 >> 8);
-  uint8_t x223 = (uint8_t)(x221 & UINT8_C(0xff));
+  uint64_t x219 = (x217 + x70);
+  uint32_t x220 = (uint32_t)(x219 >> 8);
+  uint8_t x221 = (uint8_t)(x219 & UINT8_C(0xff));
+  uint32_t x222 = (x220 >> 8);
+  uint8_t x223 = (uint8_t)(x220 & UINT8_C(0xff));
   uint32_t x224 = (x222 >> 8);
   uint8_t x225 = (uint8_t)(x222 & UINT8_C(0xff));
-  uint32_t x226 = (x224 >> 8);
+  fiat_p521_uint1 x226 = (fiat_p521_uint1)(x224 >> 8);
   uint8_t x227 = (uint8_t)(x224 & UINT8_C(0xff));
-  fiat_p521_uint1 x228 = (fiat_p521_uint1)(x226 >> 8);
-  uint8_t x229 = (uint8_t)(x226 & UINT8_C(0xff));
   out1[0] = x85;
   out1[1] = x87;
   out1[2] = x89;
@@ -1319,50 +1317,50 @@ static void fiat_p521_to_bytes(uint8_t out1[66], const uint32_t arg1[17]) {
   out1[19] = x128;
   out1[20] = x130;
   out1[21] = x132;
-  out1[22] = x134;
-  out1[23] = x137;
-  out1[24] = x139;
-  out1[25] = x141;
-  out1[26] = x144;
-  out1[27] = x146;
-  out1[28] = x148;
-  out1[29] = x150;
-  out1[30] = x153;
-  out1[31] = x155;
-  out1[32] = x157;
-  out1[33] = x159;
-  out1[34] = x162;
-  out1[35] = x164;
-  out1[36] = x166;
-  out1[37] = x168;
-  out1[38] = x171;
-  out1[39] = x173;
-  out1[40] = x175;
-  out1[41] = x177;
-  out1[42] = x180;
-  out1[43] = x182;
-  out1[44] = x184;
-  out1[45] = x186;
-  out1[46] = x189;
-  out1[47] = x191;
-  out1[48] = x193;
-  out1[49] = x196;
-  out1[50] = x198;
-  out1[51] = x200;
-  out1[52] = x202;
-  out1[53] = x205;
-  out1[54] = x207;
-  out1[55] = x209;
-  out1[56] = x211;
-  out1[57] = x214;
-  out1[58] = x216;
-  out1[59] = x218;
-  out1[60] = x220;
-  out1[61] = x223;
-  out1[62] = x225;
-  out1[63] = x227;
-  out1[64] = x229;
-  out1[65] = x228;
+  out1[22] = x133;
+  out1[23] = x136;
+  out1[24] = x138;
+  out1[25] = x140;
+  out1[26] = x143;
+  out1[27] = x145;
+  out1[28] = x147;
+  out1[29] = x149;
+  out1[30] = x152;
+  out1[31] = x154;
+  out1[32] = x156;
+  out1[33] = x158;
+  out1[34] = x161;
+  out1[35] = x163;
+  out1[36] = x165;
+  out1[37] = x167;
+  out1[38] = x170;
+  out1[39] = x172;
+  out1[40] = x174;
+  out1[41] = x176;
+  out1[42] = x179;
+  out1[43] = x181;
+  out1[44] = x183;
+  out1[45] = x184;
+  out1[46] = x187;
+  out1[47] = x189;
+  out1[48] = x191;
+  out1[49] = x194;
+  out1[50] = x196;
+  out1[51] = x198;
+  out1[52] = x200;
+  out1[53] = x203;
+  out1[54] = x205;
+  out1[55] = x207;
+  out1[56] = x209;
+  out1[57] = x212;
+  out1[58] = x214;
+  out1[59] = x216;
+  out1[60] = x218;
+  out1[61] = x221;
+  out1[62] = x223;
+  out1[63] = x225;
+  out1[64] = x227;
+  out1[65] = x226;
 }
 
 /*
@@ -1470,55 +1468,53 @@ static void fiat_p521_from_bytes(uint32_t out1[17], const uint8_t arg1[66]) {
   uint8_t x96 = (uint8_t)(x95 >> 31);
   uint32_t x97 = (uint32_t)(x95 & UINT32_C(0x7fffffff));
   uint32_t x98 = (x96 + x81);
-  fiat_p521_uint1 x99 = (fiat_p521_uint1)(x98 >> 30);
-  uint32_t x100 = (x98 & UINT32_C(0x3fffffff));
-  uint32_t x101 = (x99 + x80);
-  fiat_p521_uint1 x102 = (fiat_p521_uint1)(x101 >> 31);
-  uint32_t x103 = (x101 & UINT32_C(0x7fffffff));
-  uint64_t x104 = (x102 + x79);
-  uint8_t x105 = (uint8_t)(x104 >> 31);
-  uint32_t x106 = (uint32_t)(x104 & UINT32_C(0x7fffffff));
-  uint64_t x107 = (x105 + x78);
-  uint8_t x108 = (uint8_t)(x107 >> 30);
-  uint32_t x109 = (uint32_t)(x107 & UINT32_C(0x3fffffff));
-  uint64_t x110 = (x108 + x77);
-  uint8_t x111 = (uint8_t)(x110 >> 31);
-  uint32_t x112 = (uint32_t)(x110 & UINT32_C(0x7fffffff));
-  uint64_t x113 = (x111 + x76);
-  uint8_t x114 = (uint8_t)(x113 >> 31);
-  uint32_t x115 = (uint32_t)(x113 & UINT32_C(0x7fffffff));
-  uint32_t x116 = (x114 + x75);
-  fiat_p521_uint1 x117 = (fiat_p521_uint1)(x116 >> 30);
-  uint32_t x118 = (x116 & UINT32_C(0x3fffffff));
-  uint32_t x119 = (x117 + x74);
-  fiat_p521_uint1 x120 = (fiat_p521_uint1)(x119 >> 31);
-  uint32_t x121 = (x119 & UINT32_C(0x7fffffff));
-  uint64_t x122 = (x120 + x73);
-  uint8_t x123 = (uint8_t)(x122 >> 31);
-  uint32_t x124 = (uint32_t)(x122 & UINT32_C(0x7fffffff));
-  uint64_t x125 = (x123 + x72);
-  uint8_t x126 = (uint8_t)(x125 >> 30);
-  uint32_t x127 = (uint32_t)(x125 & UINT32_C(0x3fffffff));
-  uint64_t x128 = (x126 + x71);
-  uint8_t x129 = (uint8_t)(x128 >> 31);
-  uint32_t x130 = (uint32_t)(x128 & UINT32_C(0x7fffffff));
-  uint32_t x131 = (x129 + x70);
+  uint32_t x99 = (x98 & UINT32_C(0x3fffffff));
+  uint32_t x100 = (0x0 + x80);
+  fiat_p521_uint1 x101 = (fiat_p521_uint1)(x100 >> 31);
+  uint32_t x102 = (x100 & UINT32_C(0x7fffffff));
+  uint64_t x103 = (x101 + x79);
+  uint8_t x104 = (uint8_t)(x103 >> 31);
+  uint32_t x105 = (uint32_t)(x103 & UINT32_C(0x7fffffff));
+  uint64_t x106 = (x104 + x78);
+  uint8_t x107 = (uint8_t)(x106 >> 30);
+  uint32_t x108 = (uint32_t)(x106 & UINT32_C(0x3fffffff));
+  uint64_t x109 = (x107 + x77);
+  uint8_t x110 = (uint8_t)(x109 >> 31);
+  uint32_t x111 = (uint32_t)(x109 & UINT32_C(0x7fffffff));
+  uint64_t x112 = (x110 + x76);
+  uint8_t x113 = (uint8_t)(x112 >> 31);
+  uint32_t x114 = (uint32_t)(x112 & UINT32_C(0x7fffffff));
+  uint32_t x115 = (x113 + x75);
+  uint32_t x116 = (x115 & UINT32_C(0x3fffffff));
+  uint32_t x117 = (0x0 + x74);
+  fiat_p521_uint1 x118 = (fiat_p521_uint1)(x117 >> 31);
+  uint32_t x119 = (x117 & UINT32_C(0x7fffffff));
+  uint64_t x120 = (x118 + x73);
+  uint8_t x121 = (uint8_t)(x120 >> 31);
+  uint32_t x122 = (uint32_t)(x120 & UINT32_C(0x7fffffff));
+  uint64_t x123 = (x121 + x72);
+  uint8_t x124 = (uint8_t)(x123 >> 30);
+  uint32_t x125 = (uint32_t)(x123 & UINT32_C(0x3fffffff));
+  uint64_t x126 = (x124 + x71);
+  uint8_t x127 = (uint8_t)(x126 >> 31);
+  uint32_t x128 = (uint32_t)(x126 & UINT32_C(0x7fffffff));
+  uint32_t x129 = (x127 + x70);
   out1[0] = x69;
   out1[1] = x88;
   out1[2] = x91;
   out1[3] = x94;
   out1[4] = x97;
-  out1[5] = x100;
-  out1[6] = x103;
-  out1[7] = x106;
-  out1[8] = x109;
-  out1[9] = x112;
-  out1[10] = x115;
-  out1[11] = x118;
-  out1[12] = x121;
-  out1[13] = x124;
-  out1[14] = x127;
-  out1[15] = x130;
-  out1[16] = x131;
+  out1[5] = x99;
+  out1[6] = x102;
+  out1[7] = x105;
+  out1[8] = x108;
+  out1[9] = x111;
+  out1[10] = x114;
+  out1[11] = x116;
+  out1[12] = x119;
+  out1[13] = x122;
+  out1[14] = x125;
+  out1[15] = x128;
+  out1[16] = x129;
 }
 
diff --git a/p521_64.c b/p521_64.c
index f81fa7967..08ba9e1b7 100644
--- a/p521_64.c
+++ b/p521_64.c
@@ -646,73 +646,73 @@ static void fiat_p521_to_bytes(uint8_t out1[66], const uint64_t arg1[9]) {
   uint8_t x100 = (uint8_t)(x97 & UINT8_C(0xff));
   uint8_t x101 = (uint8_t)(x99 >> 8);
   uint8_t x102 = (uint8_t)(x99 & UINT8_C(0xff));
-  fiat_p521_uint1 x103 = (fiat_p521_uint1)(x101 >> 8);
-  uint8_t x104 = (uint8_t)(x101 & UINT8_C(0xff));
-  uint64_t x105 = (x103 + x28);
-  uint64_t x106 = (x105 >> 8);
-  uint8_t x107 = (uint8_t)(x105 & UINT8_C(0xff));
-  uint64_t x108 = (x106 >> 8);
-  uint8_t x109 = (uint8_t)(x106 & UINT8_C(0xff));
-  uint64_t x110 = (x108 >> 8);
-  uint8_t x111 = (uint8_t)(x108 & UINT8_C(0xff));
-  uint64_t x112 = (x110 >> 8);
-  uint8_t x113 = (uint8_t)(x110 & UINT8_C(0xff));
-  uint64_t x114 = (x112 >> 8);
-  uint8_t x115 = (uint8_t)(x112 & UINT8_C(0xff));
-  uint64_t x116 = (x114 >> 8);
-  uint8_t x117 = (uint8_t)(x114 & UINT8_C(0xff));
-  uint8_t x118 = (uint8_t)(x116 >> 8);
-  uint8_t x119 = (uint8_t)(x116 & UINT8_C(0xff));
-  uint64_t x120 = (x118 + x40);
-  uint64_t x121 = (x120 >> 8);
-  uint8_t x122 = (uint8_t)(x120 & UINT8_C(0xff));
-  uint64_t x123 = (x121 >> 8);
-  uint8_t x124 = (uint8_t)(x121 & UINT8_C(0xff));
-  uint64_t x125 = (x123 >> 8);
-  uint8_t x126 = (uint8_t)(x123 & UINT8_C(0xff));
-  uint64_t x127 = (x125 >> 8);
-  uint8_t x128 = (uint8_t)(x125 & UINT8_C(0xff));
-  uint64_t x129 = (x127 >> 8);
-  uint8_t x130 = (uint8_t)(x127 & UINT8_C(0xff));
-  uint64_t x131 = (x129 >> 8);
-  uint8_t x132 = (uint8_t)(x129 & UINT8_C(0xff));
-  uint8_t x133 = (uint8_t)(x131 >> 8);
-  uint8_t x134 = (uint8_t)(x131 & UINT8_C(0xff));
-  uint64_t x135 = (x133 + x39);
-  uint64_t x136 = (x135 >> 8);
-  uint8_t x137 = (uint8_t)(x135 & UINT8_C(0xff));
-  uint64_t x138 = (x136 >> 8);
-  uint8_t x139 = (uint8_t)(x136 & UINT8_C(0xff));
-  uint64_t x140 = (x138 >> 8);
-  uint8_t x141 = (uint8_t)(x138 & UINT8_C(0xff));
-  uint64_t x142 = (x140 >> 8);
-  uint8_t x143 = (uint8_t)(x140 & UINT8_C(0xff));
-  uint64_t x144 = (x142 >> 8);
-  uint8_t x145 = (uint8_t)(x142 & UINT8_C(0xff));
-  uint64_t x146 = (x144 >> 8);
-  uint8_t x147 = (uint8_t)(x144 & UINT8_C(0xff));
-  uint8_t x148 = (uint8_t)(x146 >> 8);
-  uint8_t x149 = (uint8_t)(x146 & UINT8_C(0xff));
-  uint64_t x150 = (x148 + x38);
-  uint64_t x151 = (x150 >> 8);
-  uint8_t x152 = (uint8_t)(x150 & UINT8_C(0xff));
-  uint64_t x153 = (x151 >> 8);
-  uint8_t x154 = (uint8_t)(x151 & UINT8_C(0xff));
-  uint64_t x155 = (x153 >> 8);
-  uint8_t x156 = (uint8_t)(x153 & UINT8_C(0xff));
-  uint64_t x157 = (x155 >> 8);
-  uint8_t x158 = (uint8_t)(x155 & UINT8_C(0xff));
-  uint64_t x159 = (x157 >> 8);
-  uint8_t x160 = (uint8_t)(x157 & UINT8_C(0xff));
-  uint64_t x161 = (x159 >> 8);
-  uint8_t x162 = (uint8_t)(x159 & UINT8_C(0xff));
-  uint8_t x163 = (uint8_t)(x161 >> 8);
-  uint8_t x164 = (uint8_t)(x161 & UINT8_C(0xff));
-  fiat_p521_uint1 x165 = (fiat_p521_uint1)(x163 >> 8);
-  uint8_t x166 = (uint8_t)(x163 & UINT8_C(0xff));
-  uint64_t x167 = (x165 + x36);
-  uint64_t x168 = (x167 >> 8);
-  uint8_t x169 = (uint8_t)(x167 & UINT8_C(0xff));
+  uint8_t x103 = (uint8_t)(x101 & UINT8_C(0xff));
+  uint64_t x104 = (0x0 + x28);
+  uint64_t x105 = (x104 >> 8);
+  uint8_t x106 = (uint8_t)(x104 & UINT8_C(0xff));
+  uint64_t x107 = (x105 >> 8);
+  uint8_t x108 = (uint8_t)(x105 & UINT8_C(0xff));
+  uint64_t x109 = (x107 >> 8);
+  uint8_t x110 = (uint8_t)(x107 & UINT8_C(0xff));
+  uint64_t x111 = (x109 >> 8);
+  uint8_t x112 = (uint8_t)(x109 & UINT8_C(0xff));
+  uint64_t x113 = (x111 >> 8);
+  uint8_t x114 = (uint8_t)(x111 & UINT8_C(0xff));
+  uint64_t x115 = (x113 >> 8);
+  uint8_t x116 = (uint8_t)(x113 & UINT8_C(0xff));
+  uint8_t x117 = (uint8_t)(x115 >> 8);
+  uint8_t x118 = (uint8_t)(x115 & UINT8_C(0xff));
+  uint64_t x119 = (x117 + x40);
+  uint64_t x120 = (x119 >> 8);
+  uint8_t x121 = (uint8_t)(x119 & UINT8_C(0xff));
+  uint64_t x122 = (x120 >> 8);
+  uint8_t x123 = (uint8_t)(x120 & UINT8_C(0xff));
+  uint64_t x124 = (x122 >> 8);
+  uint8_t x125 = (uint8_t)(x122 & UINT8_C(0xff));
+  uint64_t x126 = (x124 >> 8);
+  uint8_t x127 = (uint8_t)(x124 & UINT8_C(0xff));
+  uint64_t x128 = (x126 >> 8);
+  uint8_t x129 = (uint8_t)(x126 & UINT8_C(0xff));
+  uint64_t x130 = (x128 >> 8);
+  uint8_t x131 = (uint8_t)(x128 & UINT8_C(0xff));
+  uint8_t x132 = (uint8_t)(x130 >> 8);
+  uint8_t x133 = (uint8_t)(x130 & UINT8_C(0xff));
+  uint64_t x134 = (x132 + x39);
+  uint64_t x135 = (x134 >> 8);
+  uint8_t x136 = (uint8_t)(x134 & UINT8_C(0xff));
+  uint64_t x137 = (x135 >> 8);
+  uint8_t x138 = (uint8_t)(x135 & UINT8_C(0xff));
+  uint64_t x139 = (x137 >> 8);
+  uint8_t x140 = (uint8_t)(x137 & UINT8_C(0xff));
+  uint64_t x141 = (x139 >> 8);
+  uint8_t x142 = (uint8_t)(x139 & UINT8_C(0xff));
+  uint64_t x143 = (x141 >> 8);
+  uint8_t x144 = (uint8_t)(x141 & UINT8_C(0xff));
+  uint64_t x145 = (x143 >> 8);
+  uint8_t x146 = (uint8_t)(x143 & UINT8_C(0xff));
+  uint8_t x147 = (uint8_t)(x145 >> 8);
+  uint8_t x148 = (uint8_t)(x145 & UINT8_C(0xff));
+  uint64_t x149 = (x147 + x38);
+  uint64_t x150 = (x149 >> 8);
+  uint8_t x151 = (uint8_t)(x149 & UINT8_C(0xff));
+  uint64_t x152 = (x150 >> 8);
+  uint8_t x153 = (uint8_t)(x150 & UINT8_C(0xff));
+  uint64_t x154 = (x152 >> 8);
+  uint8_t x155 = (uint8_t)(x152 & UINT8_C(0xff));
+  uint64_t x156 = (x154 >> 8);
+  uint8_t x157 = (uint8_t)(x154 & UINT8_C(0xff));
+  uint64_t x158 = (x156 >> 8);
+  uint8_t x159 = (uint8_t)(x156 & UINT8_C(0xff));
+  uint64_t x160 = (x158 >> 8);
+  uint8_t x161 = (uint8_t)(x158 & UINT8_C(0xff));
+  uint8_t x162 = (uint8_t)(x160 >> 8);
+  uint8_t x163 = (uint8_t)(x160 & UINT8_C(0xff));
+  uint8_t x164 = (uint8_t)(x162 & UINT8_C(0xff));
+  uint64_t x165 = (0x0 + x36);
+  uint64_t x166 = (x165 >> 8);
+  uint8_t x167 = (uint8_t)(x165 & UINT8_C(0xff));
+  uint64_t x168 = (x166 >> 8);
+  uint8_t x169 = (uint8_t)(x166 & UINT8_C(0xff));
   uint64_t x170 = (x168 >> 8);
   uint8_t x171 = (uint8_t)(x168 & UINT8_C(0xff));
   uint64_t x172 = (x170 >> 8);
@@ -721,10 +721,8 @@ static void fiat_p521_to_bytes(uint8_t out1[66], const uint64_t arg1[9]) {
   uint8_t x175 = (uint8_t)(x172 & UINT8_C(0xff));
   uint64_t x176 = (x174 >> 8);
   uint8_t x177 = (uint8_t)(x174 & UINT8_C(0xff));
-  uint64_t x178 = (x176 >> 8);
+  fiat_p521_uint1 x178 = (fiat_p521_uint1)(x176 >> 8);
   uint8_t x179 = (uint8_t)(x176 & UINT8_C(0xff));
-  fiat_p521_uint1 x180 = (fiat_p521_uint1)(x178 >> 8);
-  uint8_t x181 = (uint8_t)(x178 & UINT8_C(0xff));
   out1[0] = x45;
   out1[1] = x47;
   out1[2] = x49;
@@ -753,44 +751,44 @@ static void fiat_p521_to_bytes(uint8_t out1[66], const uint64_t arg1[9]) {
   out1[25] = x98;
   out1[26] = x100;
   out1[27] = x102;
-  out1[28] = x104;
-  out1[29] = x107;
-  out1[30] = x109;
-  out1[31] = x111;
-  out1[32] = x113;
-  out1[33] = x115;
-  out1[34] = x117;
-  out1[35] = x119;
-  out1[36] = x122;
-  out1[37] = x124;
-  out1[38] = x126;
-  out1[39] = x128;
-  out1[40] = x130;
-  out1[41] = x132;
-  out1[42] = x134;
-  out1[43] = x137;
-  out1[44] = x139;
-  out1[45] = x141;
-  out1[46] = x143;
-  out1[47] = x145;
-  out1[48] = x147;
-  out1[49] = x149;
-  out1[50] = x152;
-  out1[51] = x154;
-  out1[52] = x156;
-  out1[53] = x158;
-  out1[54] = x160;
-  out1[55] = x162;
-  out1[56] = x164;
-  out1[57] = x166;
-  out1[58] = x169;
-  out1[59] = x171;
-  out1[60] = x173;
-  out1[61] = x175;
-  out1[62] = x177;
-  out1[63] = x179;
-  out1[64] = x181;
-  out1[65] = x180;
+  out1[28] = x103;
+  out1[29] = x106;
+  out1[30] = x108;
+  out1[31] = x110;
+  out1[32] = x112;
+  out1[33] = x114;
+  out1[34] = x116;
+  out1[35] = x118;
+  out1[36] = x121;
+  out1[37] = x123;
+  out1[38] = x125;
+  out1[39] = x127;
+  out1[40] = x129;
+  out1[41] = x131;
+  out1[42] = x133;
+  out1[43] = x136;
+  out1[44] = x138;
+  out1[45] = x140;
+  out1[46] = x142;
+  out1[47] = x144;
+  out1[48] = x146;
+  out1[49] = x148;
+  out1[50] = x151;
+  out1[51] = x153;
+  out1[52] = x155;
+  out1[53] = x157;
+  out1[54] = x159;
+  out1[55] = x161;
+  out1[56] = x163;
+  out1[57] = x164;
+  out1[58] = x167;
+  out1[59] = x169;
+  out1[60] = x171;
+  out1[61] = x173;
+  out1[62] = x175;
+  out1[63] = x177;
+  out1[64] = x179;
+  out1[65] = x178;
 }
 
 /*
@@ -884,29 +882,27 @@ static void fiat_p521_from_bytes(uint64_t out1[9], const uint8_t arg1[66]) {
   uint8_t x82 = (uint8_t)(x81 >> 58);
   uint64_t x83 = (x81 & UINT64_C(0x3ffffffffffffff));
   uint64_t x84 = (x82 + x75);
-  fiat_p521_uint1 x85 = (fiat_p521_uint1)(x84 >> 58);
-  uint64_t x86 = (x84 & UINT64_C(0x3ffffffffffffff));
-  uint64_t x87 = (x85 + x74);
-  uint8_t x88 = (uint8_t)(x87 >> 58);
-  uint64_t x89 = (x87 & UINT64_C(0x3ffffffffffffff));
-  uint64_t x90 = (x88 + x73);
-  uint8_t x91 = (uint8_t)(x90 >> 58);
-  uint64_t x92 = (x90 & UINT64_C(0x3ffffffffffffff));
-  uint64_t x93 = (x91 + x72);
-  uint8_t x94 = (uint8_t)(x93 >> 58);
-  uint64_t x95 = (x93 & UINT64_C(0x3ffffffffffffff));
-  uint64_t x96 = (x94 + x71);
-  fiat_p521_uint1 x97 = (fiat_p521_uint1)(x96 >> 58);
-  uint64_t x98 = (x96 & UINT64_C(0x3ffffffffffffff));
-  uint64_t x99 = (x97 + x70);
+  uint64_t x85 = (x84 & UINT64_C(0x3ffffffffffffff));
+  uint64_t x86 = (0x0 + x74);
+  uint8_t x87 = (uint8_t)(x86 >> 58);
+  uint64_t x88 = (x86 & UINT64_C(0x3ffffffffffffff));
+  uint64_t x89 = (x87 + x73);
+  uint8_t x90 = (uint8_t)(x89 >> 58);
+  uint64_t x91 = (x89 & UINT64_C(0x3ffffffffffffff));
+  uint64_t x92 = (x90 + x72);
+  uint8_t x93 = (uint8_t)(x92 >> 58);
+  uint64_t x94 = (x92 & UINT64_C(0x3ffffffffffffff));
+  uint64_t x95 = (x93 + x71);
+  uint64_t x96 = (x95 & UINT64_C(0x3ffffffffffffff));
+  uint64_t x97 = (0x0 + x70);
   out1[0] = x69;
   out1[1] = x80;
   out1[2] = x83;
-  out1[3] = x86;
-  out1[4] = x89;
-  out1[5] = x92;
-  out1[6] = x95;
-  out1[7] = x98;
-  out1[8] = x99;
+  out1[3] = x85;
+  out1[4] = x88;
+  out1[5] = x91;
+  out1[6] = x94;
+  out1[7] = x96;
+  out1[8] = x97;
 }
 
diff --git a/secp256k1_32.c b/secp256k1_32.c
index 94744d26a..a4015f134 100644
--- a/secp256k1_32.c
+++ b/secp256k1_32.c
@@ -2991,781 +2991,781 @@ static void fiat_secp256k1_from_montgomery(uint32_t out1[8], const uint32_t arg1
   fiat_secp256k1_uint1 x69;
   fiat_secp256k1_addcarryx_u32(&x68, &x69, x67, 0x0, x52);
   uint32_t x70;
-  fiat_secp256k1_uint1 x71;
-  fiat_secp256k1_addcarryx_u32(&x70, &x71, x53, 0x0, 0x0);
+  uint32_t x71;
+  fiat_secp256k1_mulx_u32(&x70, &x71, x54, UINT32_C(0xd2253531));
   uint32_t x72;
-  fiat_secp256k1_uint1 x73;
-  fiat_secp256k1_addcarryx_u32(&x72, &x73, x69, 0x0, (fiat_secp256k1_uint1)x70);
+  uint32_t x73;
+  fiat_secp256k1_mulx_u32(&x72, &x73, x70, UINT32_C(0xffffffff));
   uint32_t x74;
   uint32_t x75;
-  fiat_secp256k1_mulx_u32(&x74, &x75, x54, UINT32_C(0xd2253531));
+  fiat_secp256k1_mulx_u32(&x74, &x75, x70, UINT32_C(0xffffffff));
   uint32_t x76;
   uint32_t x77;
-  fiat_secp256k1_mulx_u32(&x76, &x77, x74, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x76, &x77, x70, UINT32_C(0xffffffff));
   uint32_t x78;
   uint32_t x79;
-  fiat_secp256k1_mulx_u32(&x78, &x79, x74, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x78, &x79, x70, UINT32_C(0xffffffff));
   uint32_t x80;
   uint32_t x81;
-  fiat_secp256k1_mulx_u32(&x80, &x81, x74, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x80, &x81, x70, UINT32_C(0xffffffff));
   uint32_t x82;
   uint32_t x83;
-  fiat_secp256k1_mulx_u32(&x82, &x83, x74, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x82, &x83, x70, UINT32_C(0xffffffff));
   uint32_t x84;
   uint32_t x85;
-  fiat_secp256k1_mulx_u32(&x84, &x85, x74, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x84, &x85, x70, UINT32_C(0xfffffffe));
   uint32_t x86;
   uint32_t x87;
-  fiat_secp256k1_mulx_u32(&x86, &x87, x74, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x86, &x87, x70, UINT32_C(0xfffffc2f));
   uint32_t x88;
-  uint32_t x89;
-  fiat_secp256k1_mulx_u32(&x88, &x89, x74, UINT32_C(0xfffffffe));
+  fiat_secp256k1_uint1 x89;
+  fiat_secp256k1_addcarryx_u32(&x88, &x89, 0x0, x84, x87);
   uint32_t x90;
-  uint32_t x91;
-  fiat_secp256k1_mulx_u32(&x90, &x91, x74, UINT32_C(0xfffffc2f));
+  fiat_secp256k1_uint1 x91;
+  fiat_secp256k1_addcarryx_u32(&x90, &x91, x89, x82, x85);
   uint32_t x92;
   fiat_secp256k1_uint1 x93;
-  fiat_secp256k1_addcarryx_u32(&x92, &x93, 0x0, x88, x91);
+  fiat_secp256k1_addcarryx_u32(&x92, &x93, x91, x80, x83);
   uint32_t x94;
   fiat_secp256k1_uint1 x95;
-  fiat_secp256k1_addcarryx_u32(&x94, &x95, x93, x86, x89);
+  fiat_secp256k1_addcarryx_u32(&x94, &x95, x93, x78, x81);
   uint32_t x96;
   fiat_secp256k1_uint1 x97;
-  fiat_secp256k1_addcarryx_u32(&x96, &x97, x95, x84, x87);
+  fiat_secp256k1_addcarryx_u32(&x96, &x97, x95, x76, x79);
   uint32_t x98;
   fiat_secp256k1_uint1 x99;
-  fiat_secp256k1_addcarryx_u32(&x98, &x99, x97, x82, x85);
+  fiat_secp256k1_addcarryx_u32(&x98, &x99, x97, x74, x77);
   uint32_t x100;
   fiat_secp256k1_uint1 x101;
-  fiat_secp256k1_addcarryx_u32(&x100, &x101, x99, x80, x83);
+  fiat_secp256k1_addcarryx_u32(&x100, &x101, x99, x72, x75);
   uint32_t x102;
   fiat_secp256k1_uint1 x103;
-  fiat_secp256k1_addcarryx_u32(&x102, &x103, x101, x78, x81);
+  fiat_secp256k1_addcarryx_u32(&x102, &x103, 0x0, x86, x54);
   uint32_t x104;
   fiat_secp256k1_uint1 x105;
-  fiat_secp256k1_addcarryx_u32(&x104, &x105, x103, x76, x79);
+  fiat_secp256k1_addcarryx_u32(&x104, &x105, x103, x88, x56);
   uint32_t x106;
   fiat_secp256k1_uint1 x107;
-  fiat_secp256k1_addcarryx_u32(&x106, &x107, 0x0, x90, x54);
+  fiat_secp256k1_addcarryx_u32(&x106, &x107, x105, x90, x58);
   uint32_t x108;
   fiat_secp256k1_uint1 x109;
-  fiat_secp256k1_addcarryx_u32(&x108, &x109, x107, x92, x56);
+  fiat_secp256k1_addcarryx_u32(&x108, &x109, x107, x92, x60);
   uint32_t x110;
   fiat_secp256k1_uint1 x111;
-  fiat_secp256k1_addcarryx_u32(&x110, &x111, x109, x94, x58);
+  fiat_secp256k1_addcarryx_u32(&x110, &x111, x109, x94, x62);
   uint32_t x112;
   fiat_secp256k1_uint1 x113;
-  fiat_secp256k1_addcarryx_u32(&x112, &x113, x111, x96, x60);
+  fiat_secp256k1_addcarryx_u32(&x112, &x113, x111, x96, x64);
   uint32_t x114;
   fiat_secp256k1_uint1 x115;
-  fiat_secp256k1_addcarryx_u32(&x114, &x115, x113, x98, x62);
+  fiat_secp256k1_addcarryx_u32(&x114, &x115, x113, x98, x66);
   uint32_t x116;
   fiat_secp256k1_uint1 x117;
-  fiat_secp256k1_addcarryx_u32(&x116, &x117, x115, x100, x64);
+  fiat_secp256k1_addcarryx_u32(&x116, &x117, x115, x100, x68);
   uint32_t x118;
   fiat_secp256k1_uint1 x119;
-  fiat_secp256k1_addcarryx_u32(&x118, &x119, x117, x102, x66);
+  fiat_secp256k1_addcarryx_u32(&x118, &x119, x53, 0x0, 0x0);
   uint32_t x120;
   fiat_secp256k1_uint1 x121;
-  fiat_secp256k1_addcarryx_u32(&x120, &x121, x119, x104, x68);
+  fiat_secp256k1_addcarryx_u32(&x120, &x121, x69, 0x0, (fiat_secp256k1_uint1)x118);
   uint32_t x122;
   fiat_secp256k1_uint1 x123;
-  fiat_secp256k1_addcarryx_u32(&x122, &x123, x105, 0x0, x77);
+  fiat_secp256k1_addcarryx_u32(&x122, &x123, x101, 0x0, x73);
   uint32_t x124;
   fiat_secp256k1_uint1 x125;
-  fiat_secp256k1_addcarryx_u32(&x124, &x125, x121, x122, x72);
+  fiat_secp256k1_addcarryx_u32(&x124, &x125, x117, x122, x120);
   uint32_t x126;
   fiat_secp256k1_uint1 x127;
-  fiat_secp256k1_addcarryx_u32(&x126, &x127, 0x0, (arg1[2]), x108);
+  fiat_secp256k1_addcarryx_u32(&x126, &x127, 0x0, (arg1[2]), x104);
   uint32_t x128;
   fiat_secp256k1_uint1 x129;
-  fiat_secp256k1_addcarryx_u32(&x128, &x129, x127, 0x0, x110);
+  fiat_secp256k1_addcarryx_u32(&x128, &x129, x127, 0x0, x106);
   uint32_t x130;
   fiat_secp256k1_uint1 x131;
-  fiat_secp256k1_addcarryx_u32(&x130, &x131, x129, 0x0, x112);
+  fiat_secp256k1_addcarryx_u32(&x130, &x131, x129, 0x0, x108);
   uint32_t x132;
   fiat_secp256k1_uint1 x133;
-  fiat_secp256k1_addcarryx_u32(&x132, &x133, x131, 0x0, x114);
+  fiat_secp256k1_addcarryx_u32(&x132, &x133, x131, 0x0, x110);
   uint32_t x134;
   fiat_secp256k1_uint1 x135;
-  fiat_secp256k1_addcarryx_u32(&x134, &x135, x133, 0x0, x116);
+  fiat_secp256k1_addcarryx_u32(&x134, &x135, x133, 0x0, x112);
   uint32_t x136;
   fiat_secp256k1_uint1 x137;
-  fiat_secp256k1_addcarryx_u32(&x136, &x137, x135, 0x0, x118);
+  fiat_secp256k1_addcarryx_u32(&x136, &x137, x135, 0x0, x114);
   uint32_t x138;
   fiat_secp256k1_uint1 x139;
-  fiat_secp256k1_addcarryx_u32(&x138, &x139, x137, 0x0, x120);
+  fiat_secp256k1_addcarryx_u32(&x138, &x139, x137, 0x0, x116);
   uint32_t x140;
   fiat_secp256k1_uint1 x141;
   fiat_secp256k1_addcarryx_u32(&x140, &x141, x139, 0x0, x124);
   uint32_t x142;
-  fiat_secp256k1_uint1 x143;
-  fiat_secp256k1_addcarryx_u32(&x142, &x143, x125, 0x0, x73);
+  uint32_t x143;
+  fiat_secp256k1_mulx_u32(&x142, &x143, x126, UINT32_C(0xd2253531));
   uint32_t x144;
-  fiat_secp256k1_uint1 x145;
-  fiat_secp256k1_addcarryx_u32(&x144, &x145, x141, 0x0, (fiat_secp256k1_uint1)x142);
+  uint32_t x145;
+  fiat_secp256k1_mulx_u32(&x144, &x145, x142, UINT32_C(0xffffffff));
   uint32_t x146;
   uint32_t x147;
-  fiat_secp256k1_mulx_u32(&x146, &x147, x126, UINT32_C(0xd2253531));
+  fiat_secp256k1_mulx_u32(&x146, &x147, x142, UINT32_C(0xffffffff));
   uint32_t x148;
   uint32_t x149;
-  fiat_secp256k1_mulx_u32(&x148, &x149, x146, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x148, &x149, x142, UINT32_C(0xffffffff));
   uint32_t x150;
   uint32_t x151;
-  fiat_secp256k1_mulx_u32(&x150, &x151, x146, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x150, &x151, x142, UINT32_C(0xffffffff));
   uint32_t x152;
   uint32_t x153;
-  fiat_secp256k1_mulx_u32(&x152, &x153, x146, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x152, &x153, x142, UINT32_C(0xffffffff));
   uint32_t x154;
   uint32_t x155;
-  fiat_secp256k1_mulx_u32(&x154, &x155, x146, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x154, &x155, x142, UINT32_C(0xffffffff));
   uint32_t x156;
   uint32_t x157;
-  fiat_secp256k1_mulx_u32(&x156, &x157, x146, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x156, &x157, x142, UINT32_C(0xfffffffe));
   uint32_t x158;
   uint32_t x159;
-  fiat_secp256k1_mulx_u32(&x158, &x159, x146, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x158, &x159, x142, UINT32_C(0xfffffc2f));
   uint32_t x160;
-  uint32_t x161;
-  fiat_secp256k1_mulx_u32(&x160, &x161, x146, UINT32_C(0xfffffffe));
+  fiat_secp256k1_uint1 x161;
+  fiat_secp256k1_addcarryx_u32(&x160, &x161, 0x0, x156, x159);
   uint32_t x162;
-  uint32_t x163;
-  fiat_secp256k1_mulx_u32(&x162, &x163, x146, UINT32_C(0xfffffc2f));
+  fiat_secp256k1_uint1 x163;
+  fiat_secp256k1_addcarryx_u32(&x162, &x163, x161, x154, x157);
   uint32_t x164;
   fiat_secp256k1_uint1 x165;
-  fiat_secp256k1_addcarryx_u32(&x164, &x165, 0x0, x160, x163);
+  fiat_secp256k1_addcarryx_u32(&x164, &x165, x163, x152, x155);
   uint32_t x166;
   fiat_secp256k1_uint1 x167;
-  fiat_secp256k1_addcarryx_u32(&x166, &x167, x165, x158, x161);
+  fiat_secp256k1_addcarryx_u32(&x166, &x167, x165, x150, x153);
   uint32_t x168;
   fiat_secp256k1_uint1 x169;
-  fiat_secp256k1_addcarryx_u32(&x168, &x169, x167, x156, x159);
+  fiat_secp256k1_addcarryx_u32(&x168, &x169, x167, x148, x151);
   uint32_t x170;
   fiat_secp256k1_uint1 x171;
-  fiat_secp256k1_addcarryx_u32(&x170, &x171, x169, x154, x157);
+  fiat_secp256k1_addcarryx_u32(&x170, &x171, x169, x146, x149);
   uint32_t x172;
   fiat_secp256k1_uint1 x173;
-  fiat_secp256k1_addcarryx_u32(&x172, &x173, x171, x152, x155);
+  fiat_secp256k1_addcarryx_u32(&x172, &x173, x171, x144, x147);
   uint32_t x174;
   fiat_secp256k1_uint1 x175;
-  fiat_secp256k1_addcarryx_u32(&x174, &x175, x173, x150, x153);
+  fiat_secp256k1_addcarryx_u32(&x174, &x175, 0x0, x158, x126);
   uint32_t x176;
   fiat_secp256k1_uint1 x177;
-  fiat_secp256k1_addcarryx_u32(&x176, &x177, x175, x148, x151);
+  fiat_secp256k1_addcarryx_u32(&x176, &x177, x175, x160, x128);
   uint32_t x178;
   fiat_secp256k1_uint1 x179;
-  fiat_secp256k1_addcarryx_u32(&x178, &x179, 0x0, x162, x126);
+  fiat_secp256k1_addcarryx_u32(&x178, &x179, x177, x162, x130);
   uint32_t x180;
   fiat_secp256k1_uint1 x181;
-  fiat_secp256k1_addcarryx_u32(&x180, &x181, x179, x164, x128);
+  fiat_secp256k1_addcarryx_u32(&x180, &x181, x179, x164, x132);
   uint32_t x182;
   fiat_secp256k1_uint1 x183;
-  fiat_secp256k1_addcarryx_u32(&x182, &x183, x181, x166, x130);
+  fiat_secp256k1_addcarryx_u32(&x182, &x183, x181, x166, x134);
   uint32_t x184;
   fiat_secp256k1_uint1 x185;
-  fiat_secp256k1_addcarryx_u32(&x184, &x185, x183, x168, x132);
+  fiat_secp256k1_addcarryx_u32(&x184, &x185, x183, x168, x136);
   uint32_t x186;
   fiat_secp256k1_uint1 x187;
-  fiat_secp256k1_addcarryx_u32(&x186, &x187, x185, x170, x134);
+  fiat_secp256k1_addcarryx_u32(&x186, &x187, x185, x170, x138);
   uint32_t x188;
   fiat_secp256k1_uint1 x189;
-  fiat_secp256k1_addcarryx_u32(&x188, &x189, x187, x172, x136);
+  fiat_secp256k1_addcarryx_u32(&x188, &x189, x187, x172, x140);
   uint32_t x190;
   fiat_secp256k1_uint1 x191;
-  fiat_secp256k1_addcarryx_u32(&x190, &x191, x189, x174, x138);
+  fiat_secp256k1_addcarryx_u32(&x190, &x191, x125, 0x0, 0x0);
   uint32_t x192;
   fiat_secp256k1_uint1 x193;
-  fiat_secp256k1_addcarryx_u32(&x192, &x193, x191, x176, x140);
+  fiat_secp256k1_addcarryx_u32(&x192, &x193, x141, 0x0, (fiat_secp256k1_uint1)x190);
   uint32_t x194;
   fiat_secp256k1_uint1 x195;
-  fiat_secp256k1_addcarryx_u32(&x194, &x195, x177, 0x0, x149);
+  fiat_secp256k1_addcarryx_u32(&x194, &x195, x173, 0x0, x145);
   uint32_t x196;
   fiat_secp256k1_uint1 x197;
-  fiat_secp256k1_addcarryx_u32(&x196, &x197, x193, x194, x144);
+  fiat_secp256k1_addcarryx_u32(&x196, &x197, x189, x194, x192);
   uint32_t x198;
   fiat_secp256k1_uint1 x199;
-  fiat_secp256k1_addcarryx_u32(&x198, &x199, 0x0, (arg1[3]), x180);
+  fiat_secp256k1_addcarryx_u32(&x198, &x199, 0x0, (arg1[3]), x176);
   uint32_t x200;
   fiat_secp256k1_uint1 x201;
-  fiat_secp256k1_addcarryx_u32(&x200, &x201, x199, 0x0, x182);
+  fiat_secp256k1_addcarryx_u32(&x200, &x201, x199, 0x0, x178);
   uint32_t x202;
   fiat_secp256k1_uint1 x203;
-  fiat_secp256k1_addcarryx_u32(&x202, &x203, x201, 0x0, x184);
+  fiat_secp256k1_addcarryx_u32(&x202, &x203, x201, 0x0, x180);
   uint32_t x204;
   fiat_secp256k1_uint1 x205;
-  fiat_secp256k1_addcarryx_u32(&x204, &x205, x203, 0x0, x186);
+  fiat_secp256k1_addcarryx_u32(&x204, &x205, x203, 0x0, x182);
   uint32_t x206;
   fiat_secp256k1_uint1 x207;
-  fiat_secp256k1_addcarryx_u32(&x206, &x207, x205, 0x0, x188);
+  fiat_secp256k1_addcarryx_u32(&x206, &x207, x205, 0x0, x184);
   uint32_t x208;
   fiat_secp256k1_uint1 x209;
-  fiat_secp256k1_addcarryx_u32(&x208, &x209, x207, 0x0, x190);
+  fiat_secp256k1_addcarryx_u32(&x208, &x209, x207, 0x0, x186);
   uint32_t x210;
   fiat_secp256k1_uint1 x211;
-  fiat_secp256k1_addcarryx_u32(&x210, &x211, x209, 0x0, x192);
+  fiat_secp256k1_addcarryx_u32(&x210, &x211, x209, 0x0, x188);
   uint32_t x212;
   fiat_secp256k1_uint1 x213;
   fiat_secp256k1_addcarryx_u32(&x212, &x213, x211, 0x0, x196);
   uint32_t x214;
-  fiat_secp256k1_uint1 x215;
-  fiat_secp256k1_addcarryx_u32(&x214, &x215, x197, 0x0, x145);
+  uint32_t x215;
+  fiat_secp256k1_mulx_u32(&x214, &x215, x198, UINT32_C(0xd2253531));
   uint32_t x216;
-  fiat_secp256k1_uint1 x217;
-  fiat_secp256k1_addcarryx_u32(&x216, &x217, x213, 0x0, (fiat_secp256k1_uint1)x214);
+  uint32_t x217;
+  fiat_secp256k1_mulx_u32(&x216, &x217, x214, UINT32_C(0xffffffff));
   uint32_t x218;
   uint32_t x219;
-  fiat_secp256k1_mulx_u32(&x218, &x219, x198, UINT32_C(0xd2253531));
+  fiat_secp256k1_mulx_u32(&x218, &x219, x214, UINT32_C(0xffffffff));
   uint32_t x220;
   uint32_t x221;
-  fiat_secp256k1_mulx_u32(&x220, &x221, x218, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x220, &x221, x214, UINT32_C(0xffffffff));
   uint32_t x222;
   uint32_t x223;
-  fiat_secp256k1_mulx_u32(&x222, &x223, x218, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x222, &x223, x214, UINT32_C(0xffffffff));
   uint32_t x224;
   uint32_t x225;
-  fiat_secp256k1_mulx_u32(&x224, &x225, x218, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x224, &x225, x214, UINT32_C(0xffffffff));
   uint32_t x226;
   uint32_t x227;
-  fiat_secp256k1_mulx_u32(&x226, &x227, x218, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x226, &x227, x214, UINT32_C(0xffffffff));
   uint32_t x228;
   uint32_t x229;
-  fiat_secp256k1_mulx_u32(&x228, &x229, x218, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x228, &x229, x214, UINT32_C(0xfffffffe));
   uint32_t x230;
   uint32_t x231;
-  fiat_secp256k1_mulx_u32(&x230, &x231, x218, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x230, &x231, x214, UINT32_C(0xfffffc2f));
   uint32_t x232;
-  uint32_t x233;
-  fiat_secp256k1_mulx_u32(&x232, &x233, x218, UINT32_C(0xfffffffe));
+  fiat_secp256k1_uint1 x233;
+  fiat_secp256k1_addcarryx_u32(&x232, &x233, 0x0, x228, x231);
   uint32_t x234;
-  uint32_t x235;
-  fiat_secp256k1_mulx_u32(&x234, &x235, x218, UINT32_C(0xfffffc2f));
+  fiat_secp256k1_uint1 x235;
+  fiat_secp256k1_addcarryx_u32(&x234, &x235, x233, x226, x229);
   uint32_t x236;
   fiat_secp256k1_uint1 x237;
-  fiat_secp256k1_addcarryx_u32(&x236, &x237, 0x0, x232, x235);
+  fiat_secp256k1_addcarryx_u32(&x236, &x237, x235, x224, x227);
   uint32_t x238;
   fiat_secp256k1_uint1 x239;
-  fiat_secp256k1_addcarryx_u32(&x238, &x239, x237, x230, x233);
+  fiat_secp256k1_addcarryx_u32(&x238, &x239, x237, x222, x225);
   uint32_t x240;
   fiat_secp256k1_uint1 x241;
-  fiat_secp256k1_addcarryx_u32(&x240, &x241, x239, x228, x231);
+  fiat_secp256k1_addcarryx_u32(&x240, &x241, x239, x220, x223);
   uint32_t x242;
   fiat_secp256k1_uint1 x243;
-  fiat_secp256k1_addcarryx_u32(&x242, &x243, x241, x226, x229);
+  fiat_secp256k1_addcarryx_u32(&x242, &x243, x241, x218, x221);
   uint32_t x244;
   fiat_secp256k1_uint1 x245;
-  fiat_secp256k1_addcarryx_u32(&x244, &x245, x243, x224, x227);
+  fiat_secp256k1_addcarryx_u32(&x244, &x245, x243, x216, x219);
   uint32_t x246;
   fiat_secp256k1_uint1 x247;
-  fiat_secp256k1_addcarryx_u32(&x246, &x247, x245, x222, x225);
+  fiat_secp256k1_addcarryx_u32(&x246, &x247, 0x0, x230, x198);
   uint32_t x248;
   fiat_secp256k1_uint1 x249;
-  fiat_secp256k1_addcarryx_u32(&x248, &x249, x247, x220, x223);
+  fiat_secp256k1_addcarryx_u32(&x248, &x249, x247, x232, x200);
   uint32_t x250;
   fiat_secp256k1_uint1 x251;
-  fiat_secp256k1_addcarryx_u32(&x250, &x251, 0x0, x234, x198);
+  fiat_secp256k1_addcarryx_u32(&x250, &x251, x249, x234, x202);
   uint32_t x252;
   fiat_secp256k1_uint1 x253;
-  fiat_secp256k1_addcarryx_u32(&x252, &x253, x251, x236, x200);
+  fiat_secp256k1_addcarryx_u32(&x252, &x253, x251, x236, x204);
   uint32_t x254;
   fiat_secp256k1_uint1 x255;
-  fiat_secp256k1_addcarryx_u32(&x254, &x255, x253, x238, x202);
+  fiat_secp256k1_addcarryx_u32(&x254, &x255, x253, x238, x206);
   uint32_t x256;
   fiat_secp256k1_uint1 x257;
-  fiat_secp256k1_addcarryx_u32(&x256, &x257, x255, x240, x204);
+  fiat_secp256k1_addcarryx_u32(&x256, &x257, x255, x240, x208);
   uint32_t x258;
   fiat_secp256k1_uint1 x259;
-  fiat_secp256k1_addcarryx_u32(&x258, &x259, x257, x242, x206);
+  fiat_secp256k1_addcarryx_u32(&x258, &x259, x257, x242, x210);
   uint32_t x260;
   fiat_secp256k1_uint1 x261;
-  fiat_secp256k1_addcarryx_u32(&x260, &x261, x259, x244, x208);
+  fiat_secp256k1_addcarryx_u32(&x260, &x261, x259, x244, x212);
   uint32_t x262;
   fiat_secp256k1_uint1 x263;
-  fiat_secp256k1_addcarryx_u32(&x262, &x263, x261, x246, x210);
+  fiat_secp256k1_addcarryx_u32(&x262, &x263, x197, 0x0, 0x0);
   uint32_t x264;
   fiat_secp256k1_uint1 x265;
-  fiat_secp256k1_addcarryx_u32(&x264, &x265, x263, x248, x212);
+  fiat_secp256k1_addcarryx_u32(&x264, &x265, x213, 0x0, (fiat_secp256k1_uint1)x262);
   uint32_t x266;
   fiat_secp256k1_uint1 x267;
-  fiat_secp256k1_addcarryx_u32(&x266, &x267, x249, 0x0, x221);
+  fiat_secp256k1_addcarryx_u32(&x266, &x267, x245, 0x0, x217);
   uint32_t x268;
   fiat_secp256k1_uint1 x269;
-  fiat_secp256k1_addcarryx_u32(&x268, &x269, x265, x266, x216);
+  fiat_secp256k1_addcarryx_u32(&x268, &x269, x261, x266, x264);
   uint32_t x270;
   fiat_secp256k1_uint1 x271;
-  fiat_secp256k1_addcarryx_u32(&x270, &x271, 0x0, (arg1[4]), x252);
+  fiat_secp256k1_addcarryx_u32(&x270, &x271, 0x0, (arg1[4]), x248);
   uint32_t x272;
   fiat_secp256k1_uint1 x273;
-  fiat_secp256k1_addcarryx_u32(&x272, &x273, x271, 0x0, x254);
+  fiat_secp256k1_addcarryx_u32(&x272, &x273, x271, 0x0, x250);
   uint32_t x274;
   fiat_secp256k1_uint1 x275;
-  fiat_secp256k1_addcarryx_u32(&x274, &x275, x273, 0x0, x256);
+  fiat_secp256k1_addcarryx_u32(&x274, &x275, x273, 0x0, x252);
   uint32_t x276;
   fiat_secp256k1_uint1 x277;
-  fiat_secp256k1_addcarryx_u32(&x276, &x277, x275, 0x0, x258);
+  fiat_secp256k1_addcarryx_u32(&x276, &x277, x275, 0x0, x254);
   uint32_t x278;
   fiat_secp256k1_uint1 x279;
-  fiat_secp256k1_addcarryx_u32(&x278, &x279, x277, 0x0, x260);
+  fiat_secp256k1_addcarryx_u32(&x278, &x279, x277, 0x0, x256);
   uint32_t x280;
   fiat_secp256k1_uint1 x281;
-  fiat_secp256k1_addcarryx_u32(&x280, &x281, x279, 0x0, x262);
+  fiat_secp256k1_addcarryx_u32(&x280, &x281, x279, 0x0, x258);
   uint32_t x282;
   fiat_secp256k1_uint1 x283;
-  fiat_secp256k1_addcarryx_u32(&x282, &x283, x281, 0x0, x264);
+  fiat_secp256k1_addcarryx_u32(&x282, &x283, x281, 0x0, x260);
   uint32_t x284;
   fiat_secp256k1_uint1 x285;
   fiat_secp256k1_addcarryx_u32(&x284, &x285, x283, 0x0, x268);
   uint32_t x286;
-  fiat_secp256k1_uint1 x287;
-  fiat_secp256k1_addcarryx_u32(&x286, &x287, x269, 0x0, x217);
+  uint32_t x287;
+  fiat_secp256k1_mulx_u32(&x286, &x287, x270, UINT32_C(0xd2253531));
   uint32_t x288;
-  fiat_secp256k1_uint1 x289;
-  fiat_secp256k1_addcarryx_u32(&x288, &x289, x285, 0x0, (fiat_secp256k1_uint1)x286);
+  uint32_t x289;
+  fiat_secp256k1_mulx_u32(&x288, &x289, x286, UINT32_C(0xffffffff));
   uint32_t x290;
   uint32_t x291;
-  fiat_secp256k1_mulx_u32(&x290, &x291, x270, UINT32_C(0xd2253531));
+  fiat_secp256k1_mulx_u32(&x290, &x291, x286, UINT32_C(0xffffffff));
   uint32_t x292;
   uint32_t x293;
-  fiat_secp256k1_mulx_u32(&x292, &x293, x290, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x292, &x293, x286, UINT32_C(0xffffffff));
   uint32_t x294;
   uint32_t x295;
-  fiat_secp256k1_mulx_u32(&x294, &x295, x290, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x294, &x295, x286, UINT32_C(0xffffffff));
   uint32_t x296;
   uint32_t x297;
-  fiat_secp256k1_mulx_u32(&x296, &x297, x290, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x296, &x297, x286, UINT32_C(0xffffffff));
   uint32_t x298;
   uint32_t x299;
-  fiat_secp256k1_mulx_u32(&x298, &x299, x290, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x298, &x299, x286, UINT32_C(0xffffffff));
   uint32_t x300;
   uint32_t x301;
-  fiat_secp256k1_mulx_u32(&x300, &x301, x290, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x300, &x301, x286, UINT32_C(0xfffffffe));
   uint32_t x302;
   uint32_t x303;
-  fiat_secp256k1_mulx_u32(&x302, &x303, x290, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x302, &x303, x286, UINT32_C(0xfffffc2f));
   uint32_t x304;
-  uint32_t x305;
-  fiat_secp256k1_mulx_u32(&x304, &x305, x290, UINT32_C(0xfffffffe));
+  fiat_secp256k1_uint1 x305;
+  fiat_secp256k1_addcarryx_u32(&x304, &x305, 0x0, x300, x303);
   uint32_t x306;
-  uint32_t x307;
-  fiat_secp256k1_mulx_u32(&x306, &x307, x290, UINT32_C(0xfffffc2f));
+  fiat_secp256k1_uint1 x307;
+  fiat_secp256k1_addcarryx_u32(&x306, &x307, x305, x298, x301);
   uint32_t x308;
   fiat_secp256k1_uint1 x309;
-  fiat_secp256k1_addcarryx_u32(&x308, &x309, 0x0, x304, x307);
+  fiat_secp256k1_addcarryx_u32(&x308, &x309, x307, x296, x299);
   uint32_t x310;
   fiat_secp256k1_uint1 x311;
-  fiat_secp256k1_addcarryx_u32(&x310, &x311, x309, x302, x305);
+  fiat_secp256k1_addcarryx_u32(&x310, &x311, x309, x294, x297);
   uint32_t x312;
   fiat_secp256k1_uint1 x313;
-  fiat_secp256k1_addcarryx_u32(&x312, &x313, x311, x300, x303);
+  fiat_secp256k1_addcarryx_u32(&x312, &x313, x311, x292, x295);
   uint32_t x314;
   fiat_secp256k1_uint1 x315;
-  fiat_secp256k1_addcarryx_u32(&x314, &x315, x313, x298, x301);
+  fiat_secp256k1_addcarryx_u32(&x314, &x315, x313, x290, x293);
   uint32_t x316;
   fiat_secp256k1_uint1 x317;
-  fiat_secp256k1_addcarryx_u32(&x316, &x317, x315, x296, x299);
+  fiat_secp256k1_addcarryx_u32(&x316, &x317, x315, x288, x291);
   uint32_t x318;
   fiat_secp256k1_uint1 x319;
-  fiat_secp256k1_addcarryx_u32(&x318, &x319, x317, x294, x297);
+  fiat_secp256k1_addcarryx_u32(&x318, &x319, 0x0, x302, x270);
   uint32_t x320;
   fiat_secp256k1_uint1 x321;
-  fiat_secp256k1_addcarryx_u32(&x320, &x321, x319, x292, x295);
+  fiat_secp256k1_addcarryx_u32(&x320, &x321, x319, x304, x272);
   uint32_t x322;
   fiat_secp256k1_uint1 x323;
-  fiat_secp256k1_addcarryx_u32(&x322, &x323, 0x0, x306, x270);
+  fiat_secp256k1_addcarryx_u32(&x322, &x323, x321, x306, x274);
   uint32_t x324;
   fiat_secp256k1_uint1 x325;
-  fiat_secp256k1_addcarryx_u32(&x324, &x325, x323, x308, x272);
+  fiat_secp256k1_addcarryx_u32(&x324, &x325, x323, x308, x276);
   uint32_t x326;
   fiat_secp256k1_uint1 x327;
-  fiat_secp256k1_addcarryx_u32(&x326, &x327, x325, x310, x274);
+  fiat_secp256k1_addcarryx_u32(&x326, &x327, x325, x310, x278);
   uint32_t x328;
   fiat_secp256k1_uint1 x329;
-  fiat_secp256k1_addcarryx_u32(&x328, &x329, x327, x312, x276);
+  fiat_secp256k1_addcarryx_u32(&x328, &x329, x327, x312, x280);
   uint32_t x330;
   fiat_secp256k1_uint1 x331;
-  fiat_secp256k1_addcarryx_u32(&x330, &x331, x329, x314, x278);
+  fiat_secp256k1_addcarryx_u32(&x330, &x331, x329, x314, x282);
   uint32_t x332;
   fiat_secp256k1_uint1 x333;
-  fiat_secp256k1_addcarryx_u32(&x332, &x333, x331, x316, x280);
+  fiat_secp256k1_addcarryx_u32(&x332, &x333, x331, x316, x284);
   uint32_t x334;
   fiat_secp256k1_uint1 x335;
-  fiat_secp256k1_addcarryx_u32(&x334, &x335, x333, x318, x282);
+  fiat_secp256k1_addcarryx_u32(&x334, &x335, x269, 0x0, 0x0);
   uint32_t x336;
   fiat_secp256k1_uint1 x337;
-  fiat_secp256k1_addcarryx_u32(&x336, &x337, x335, x320, x284);
+  fiat_secp256k1_addcarryx_u32(&x336, &x337, x285, 0x0, (fiat_secp256k1_uint1)x334);
   uint32_t x338;
   fiat_secp256k1_uint1 x339;
-  fiat_secp256k1_addcarryx_u32(&x338, &x339, x321, 0x0, x293);
+  fiat_secp256k1_addcarryx_u32(&x338, &x339, x317, 0x0, x289);
   uint32_t x340;
   fiat_secp256k1_uint1 x341;
-  fiat_secp256k1_addcarryx_u32(&x340, &x341, x337, x338, x288);
+  fiat_secp256k1_addcarryx_u32(&x340, &x341, x333, x338, x336);
   uint32_t x342;
   fiat_secp256k1_uint1 x343;
-  fiat_secp256k1_addcarryx_u32(&x342, &x343, 0x0, (arg1[5]), x324);
+  fiat_secp256k1_addcarryx_u32(&x342, &x343, 0x0, (arg1[5]), x320);
   uint32_t x344;
   fiat_secp256k1_uint1 x345;
-  fiat_secp256k1_addcarryx_u32(&x344, &x345, x343, 0x0, x326);
+  fiat_secp256k1_addcarryx_u32(&x344, &x345, x343, 0x0, x322);
   uint32_t x346;
   fiat_secp256k1_uint1 x347;
-  fiat_secp256k1_addcarryx_u32(&x346, &x347, x345, 0x0, x328);
+  fiat_secp256k1_addcarryx_u32(&x346, &x347, x345, 0x0, x324);
   uint32_t x348;
   fiat_secp256k1_uint1 x349;
-  fiat_secp256k1_addcarryx_u32(&x348, &x349, x347, 0x0, x330);
+  fiat_secp256k1_addcarryx_u32(&x348, &x349, x347, 0x0, x326);
   uint32_t x350;
   fiat_secp256k1_uint1 x351;
-  fiat_secp256k1_addcarryx_u32(&x350, &x351, x349, 0x0, x332);
+  fiat_secp256k1_addcarryx_u32(&x350, &x351, x349, 0x0, x328);
   uint32_t x352;
   fiat_secp256k1_uint1 x353;
-  fiat_secp256k1_addcarryx_u32(&x352, &x353, x351, 0x0, x334);
+  fiat_secp256k1_addcarryx_u32(&x352, &x353, x351, 0x0, x330);
   uint32_t x354;
   fiat_secp256k1_uint1 x355;
-  fiat_secp256k1_addcarryx_u32(&x354, &x355, x353, 0x0, x336);
+  fiat_secp256k1_addcarryx_u32(&x354, &x355, x353, 0x0, x332);
   uint32_t x356;
   fiat_secp256k1_uint1 x357;
   fiat_secp256k1_addcarryx_u32(&x356, &x357, x355, 0x0, x340);
   uint32_t x358;
-  fiat_secp256k1_uint1 x359;
-  fiat_secp256k1_addcarryx_u32(&x358, &x359, x341, 0x0, x289);
+  uint32_t x359;
+  fiat_secp256k1_mulx_u32(&x358, &x359, x342, UINT32_C(0xd2253531));
   uint32_t x360;
-  fiat_secp256k1_uint1 x361;
-  fiat_secp256k1_addcarryx_u32(&x360, &x361, x357, 0x0, (fiat_secp256k1_uint1)x358);
+  uint32_t x361;
+  fiat_secp256k1_mulx_u32(&x360, &x361, x358, UINT32_C(0xffffffff));
   uint32_t x362;
   uint32_t x363;
-  fiat_secp256k1_mulx_u32(&x362, &x363, x342, UINT32_C(0xd2253531));
+  fiat_secp256k1_mulx_u32(&x362, &x363, x358, UINT32_C(0xffffffff));
   uint32_t x364;
   uint32_t x365;
-  fiat_secp256k1_mulx_u32(&x364, &x365, x362, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x364, &x365, x358, UINT32_C(0xffffffff));
   uint32_t x366;
   uint32_t x367;
-  fiat_secp256k1_mulx_u32(&x366, &x367, x362, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x366, &x367, x358, UINT32_C(0xffffffff));
   uint32_t x368;
   uint32_t x369;
-  fiat_secp256k1_mulx_u32(&x368, &x369, x362, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x368, &x369, x358, UINT32_C(0xffffffff));
   uint32_t x370;
   uint32_t x371;
-  fiat_secp256k1_mulx_u32(&x370, &x371, x362, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x370, &x371, x358, UINT32_C(0xffffffff));
   uint32_t x372;
   uint32_t x373;
-  fiat_secp256k1_mulx_u32(&x372, &x373, x362, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x372, &x373, x358, UINT32_C(0xfffffffe));
   uint32_t x374;
   uint32_t x375;
-  fiat_secp256k1_mulx_u32(&x374, &x375, x362, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x374, &x375, x358, UINT32_C(0xfffffc2f));
   uint32_t x376;
-  uint32_t x377;
-  fiat_secp256k1_mulx_u32(&x376, &x377, x362, UINT32_C(0xfffffffe));
+  fiat_secp256k1_uint1 x377;
+  fiat_secp256k1_addcarryx_u32(&x376, &x377, 0x0, x372, x375);
   uint32_t x378;
-  uint32_t x379;
-  fiat_secp256k1_mulx_u32(&x378, &x379, x362, UINT32_C(0xfffffc2f));
+  fiat_secp256k1_uint1 x379;
+  fiat_secp256k1_addcarryx_u32(&x378, &x379, x377, x370, x373);
   uint32_t x380;
   fiat_secp256k1_uint1 x381;
-  fiat_secp256k1_addcarryx_u32(&x380, &x381, 0x0, x376, x379);
+  fiat_secp256k1_addcarryx_u32(&x380, &x381, x379, x368, x371);
   uint32_t x382;
   fiat_secp256k1_uint1 x383;
-  fiat_secp256k1_addcarryx_u32(&x382, &x383, x381, x374, x377);
+  fiat_secp256k1_addcarryx_u32(&x382, &x383, x381, x366, x369);
   uint32_t x384;
   fiat_secp256k1_uint1 x385;
-  fiat_secp256k1_addcarryx_u32(&x384, &x385, x383, x372, x375);
+  fiat_secp256k1_addcarryx_u32(&x384, &x385, x383, x364, x367);
   uint32_t x386;
   fiat_secp256k1_uint1 x387;
-  fiat_secp256k1_addcarryx_u32(&x386, &x387, x385, x370, x373);
+  fiat_secp256k1_addcarryx_u32(&x386, &x387, x385, x362, x365);
   uint32_t x388;
   fiat_secp256k1_uint1 x389;
-  fiat_secp256k1_addcarryx_u32(&x388, &x389, x387, x368, x371);
+  fiat_secp256k1_addcarryx_u32(&x388, &x389, x387, x360, x363);
   uint32_t x390;
   fiat_secp256k1_uint1 x391;
-  fiat_secp256k1_addcarryx_u32(&x390, &x391, x389, x366, x369);
+  fiat_secp256k1_addcarryx_u32(&x390, &x391, 0x0, x374, x342);
   uint32_t x392;
   fiat_secp256k1_uint1 x393;
-  fiat_secp256k1_addcarryx_u32(&x392, &x393, x391, x364, x367);
+  fiat_secp256k1_addcarryx_u32(&x392, &x393, x391, x376, x344);
   uint32_t x394;
   fiat_secp256k1_uint1 x395;
-  fiat_secp256k1_addcarryx_u32(&x394, &x395, 0x0, x378, x342);
+  fiat_secp256k1_addcarryx_u32(&x394, &x395, x393, x378, x346);
   uint32_t x396;
   fiat_secp256k1_uint1 x397;
-  fiat_secp256k1_addcarryx_u32(&x396, &x397, x395, x380, x344);
+  fiat_secp256k1_addcarryx_u32(&x396, &x397, x395, x380, x348);
   uint32_t x398;
   fiat_secp256k1_uint1 x399;
-  fiat_secp256k1_addcarryx_u32(&x398, &x399, x397, x382, x346);
+  fiat_secp256k1_addcarryx_u32(&x398, &x399, x397, x382, x350);
   uint32_t x400;
   fiat_secp256k1_uint1 x401;
-  fiat_secp256k1_addcarryx_u32(&x400, &x401, x399, x384, x348);
+  fiat_secp256k1_addcarryx_u32(&x400, &x401, x399, x384, x352);
   uint32_t x402;
   fiat_secp256k1_uint1 x403;
-  fiat_secp256k1_addcarryx_u32(&x402, &x403, x401, x386, x350);
+  fiat_secp256k1_addcarryx_u32(&x402, &x403, x401, x386, x354);
   uint32_t x404;
   fiat_secp256k1_uint1 x405;
-  fiat_secp256k1_addcarryx_u32(&x404, &x405, x403, x388, x352);
+  fiat_secp256k1_addcarryx_u32(&x404, &x405, x403, x388, x356);
   uint32_t x406;
   fiat_secp256k1_uint1 x407;
-  fiat_secp256k1_addcarryx_u32(&x406, &x407, x405, x390, x354);
+  fiat_secp256k1_addcarryx_u32(&x406, &x407, x341, 0x0, 0x0);
   uint32_t x408;
   fiat_secp256k1_uint1 x409;
-  fiat_secp256k1_addcarryx_u32(&x408, &x409, x407, x392, x356);
+  fiat_secp256k1_addcarryx_u32(&x408, &x409, x357, 0x0, (fiat_secp256k1_uint1)x406);
   uint32_t x410;
   fiat_secp256k1_uint1 x411;
-  fiat_secp256k1_addcarryx_u32(&x410, &x411, x393, 0x0, x365);
+  fiat_secp256k1_addcarryx_u32(&x410, &x411, x389, 0x0, x361);
   uint32_t x412;
   fiat_secp256k1_uint1 x413;
-  fiat_secp256k1_addcarryx_u32(&x412, &x413, x409, x410, x360);
+  fiat_secp256k1_addcarryx_u32(&x412, &x413, x405, x410, x408);
   uint32_t x414;
   fiat_secp256k1_uint1 x415;
-  fiat_secp256k1_addcarryx_u32(&x414, &x415, 0x0, (arg1[6]), x396);
+  fiat_secp256k1_addcarryx_u32(&x414, &x415, 0x0, (arg1[6]), x392);
   uint32_t x416;
   fiat_secp256k1_uint1 x417;
-  fiat_secp256k1_addcarryx_u32(&x416, &x417, x415, 0x0, x398);
+  fiat_secp256k1_addcarryx_u32(&x416, &x417, x415, 0x0, x394);
   uint32_t x418;
   fiat_secp256k1_uint1 x419;
-  fiat_secp256k1_addcarryx_u32(&x418, &x419, x417, 0x0, x400);
+  fiat_secp256k1_addcarryx_u32(&x418, &x419, x417, 0x0, x396);
   uint32_t x420;
   fiat_secp256k1_uint1 x421;
-  fiat_secp256k1_addcarryx_u32(&x420, &x421, x419, 0x0, x402);
+  fiat_secp256k1_addcarryx_u32(&x420, &x421, x419, 0x0, x398);
   uint32_t x422;
   fiat_secp256k1_uint1 x423;
-  fiat_secp256k1_addcarryx_u32(&x422, &x423, x421, 0x0, x404);
+  fiat_secp256k1_addcarryx_u32(&x422, &x423, x421, 0x0, x400);
   uint32_t x424;
   fiat_secp256k1_uint1 x425;
-  fiat_secp256k1_addcarryx_u32(&x424, &x425, x423, 0x0, x406);
+  fiat_secp256k1_addcarryx_u32(&x424, &x425, x423, 0x0, x402);
   uint32_t x426;
   fiat_secp256k1_uint1 x427;
-  fiat_secp256k1_addcarryx_u32(&x426, &x427, x425, 0x0, x408);
+  fiat_secp256k1_addcarryx_u32(&x426, &x427, x425, 0x0, x404);
   uint32_t x428;
   fiat_secp256k1_uint1 x429;
   fiat_secp256k1_addcarryx_u32(&x428, &x429, x427, 0x0, x412);
   uint32_t x430;
-  fiat_secp256k1_uint1 x431;
-  fiat_secp256k1_addcarryx_u32(&x430, &x431, x413, 0x0, x361);
+  uint32_t x431;
+  fiat_secp256k1_mulx_u32(&x430, &x431, x414, UINT32_C(0xd2253531));
   uint32_t x432;
-  fiat_secp256k1_uint1 x433;
-  fiat_secp256k1_addcarryx_u32(&x432, &x433, x429, 0x0, (fiat_secp256k1_uint1)x430);
+  uint32_t x433;
+  fiat_secp256k1_mulx_u32(&x432, &x433, x430, UINT32_C(0xffffffff));
   uint32_t x434;
   uint32_t x435;
-  fiat_secp256k1_mulx_u32(&x434, &x435, x414, UINT32_C(0xd2253531));
+  fiat_secp256k1_mulx_u32(&x434, &x435, x430, UINT32_C(0xffffffff));
   uint32_t x436;
   uint32_t x437;
-  fiat_secp256k1_mulx_u32(&x436, &x437, x434, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x436, &x437, x430, UINT32_C(0xffffffff));
   uint32_t x438;
   uint32_t x439;
-  fiat_secp256k1_mulx_u32(&x438, &x439, x434, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x438, &x439, x430, UINT32_C(0xffffffff));
   uint32_t x440;
   uint32_t x441;
-  fiat_secp256k1_mulx_u32(&x440, &x441, x434, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x440, &x441, x430, UINT32_C(0xffffffff));
   uint32_t x442;
   uint32_t x443;
-  fiat_secp256k1_mulx_u32(&x442, &x443, x434, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x442, &x443, x430, UINT32_C(0xffffffff));
   uint32_t x444;
   uint32_t x445;
-  fiat_secp256k1_mulx_u32(&x444, &x445, x434, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x444, &x445, x430, UINT32_C(0xfffffffe));
   uint32_t x446;
   uint32_t x447;
-  fiat_secp256k1_mulx_u32(&x446, &x447, x434, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x446, &x447, x430, UINT32_C(0xfffffc2f));
   uint32_t x448;
-  uint32_t x449;
-  fiat_secp256k1_mulx_u32(&x448, &x449, x434, UINT32_C(0xfffffffe));
+  fiat_secp256k1_uint1 x449;
+  fiat_secp256k1_addcarryx_u32(&x448, &x449, 0x0, x444, x447);
   uint32_t x450;
-  uint32_t x451;
-  fiat_secp256k1_mulx_u32(&x450, &x451, x434, UINT32_C(0xfffffc2f));
+  fiat_secp256k1_uint1 x451;
+  fiat_secp256k1_addcarryx_u32(&x450, &x451, x449, x442, x445);
   uint32_t x452;
   fiat_secp256k1_uint1 x453;
-  fiat_secp256k1_addcarryx_u32(&x452, &x453, 0x0, x448, x451);
+  fiat_secp256k1_addcarryx_u32(&x452, &x453, x451, x440, x443);
   uint32_t x454;
   fiat_secp256k1_uint1 x455;
-  fiat_secp256k1_addcarryx_u32(&x454, &x455, x453, x446, x449);
+  fiat_secp256k1_addcarryx_u32(&x454, &x455, x453, x438, x441);
   uint32_t x456;
   fiat_secp256k1_uint1 x457;
-  fiat_secp256k1_addcarryx_u32(&x456, &x457, x455, x444, x447);
+  fiat_secp256k1_addcarryx_u32(&x456, &x457, x455, x436, x439);
   uint32_t x458;
   fiat_secp256k1_uint1 x459;
-  fiat_secp256k1_addcarryx_u32(&x458, &x459, x457, x442, x445);
+  fiat_secp256k1_addcarryx_u32(&x458, &x459, x457, x434, x437);
   uint32_t x460;
   fiat_secp256k1_uint1 x461;
-  fiat_secp256k1_addcarryx_u32(&x460, &x461, x459, x440, x443);
+  fiat_secp256k1_addcarryx_u32(&x460, &x461, x459, x432, x435);
   uint32_t x462;
   fiat_secp256k1_uint1 x463;
-  fiat_secp256k1_addcarryx_u32(&x462, &x463, x461, x438, x441);
+  fiat_secp256k1_addcarryx_u32(&x462, &x463, 0x0, x446, x414);
   uint32_t x464;
   fiat_secp256k1_uint1 x465;
-  fiat_secp256k1_addcarryx_u32(&x464, &x465, x463, x436, x439);
+  fiat_secp256k1_addcarryx_u32(&x464, &x465, x463, x448, x416);
   uint32_t x466;
   fiat_secp256k1_uint1 x467;
-  fiat_secp256k1_addcarryx_u32(&x466, &x467, 0x0, x450, x414);
+  fiat_secp256k1_addcarryx_u32(&x466, &x467, x465, x450, x418);
   uint32_t x468;
   fiat_secp256k1_uint1 x469;
-  fiat_secp256k1_addcarryx_u32(&x468, &x469, x467, x452, x416);
+  fiat_secp256k1_addcarryx_u32(&x468, &x469, x467, x452, x420);
   uint32_t x470;
   fiat_secp256k1_uint1 x471;
-  fiat_secp256k1_addcarryx_u32(&x470, &x471, x469, x454, x418);
+  fiat_secp256k1_addcarryx_u32(&x470, &x471, x469, x454, x422);
   uint32_t x472;
   fiat_secp256k1_uint1 x473;
-  fiat_secp256k1_addcarryx_u32(&x472, &x473, x471, x456, x420);
+  fiat_secp256k1_addcarryx_u32(&x472, &x473, x471, x456, x424);
   uint32_t x474;
   fiat_secp256k1_uint1 x475;
-  fiat_secp256k1_addcarryx_u32(&x474, &x475, x473, x458, x422);
+  fiat_secp256k1_addcarryx_u32(&x474, &x475, x473, x458, x426);
   uint32_t x476;
   fiat_secp256k1_uint1 x477;
-  fiat_secp256k1_addcarryx_u32(&x476, &x477, x475, x460, x424);
+  fiat_secp256k1_addcarryx_u32(&x476, &x477, x475, x460, x428);
   uint32_t x478;
   fiat_secp256k1_uint1 x479;
-  fiat_secp256k1_addcarryx_u32(&x478, &x479, x477, x462, x426);
+  fiat_secp256k1_addcarryx_u32(&x478, &x479, x413, 0x0, 0x0);
   uint32_t x480;
   fiat_secp256k1_uint1 x481;
-  fiat_secp256k1_addcarryx_u32(&x480, &x481, x479, x464, x428);
+  fiat_secp256k1_addcarryx_u32(&x480, &x481, x429, 0x0, (fiat_secp256k1_uint1)x478);
   uint32_t x482;
   fiat_secp256k1_uint1 x483;
-  fiat_secp256k1_addcarryx_u32(&x482, &x483, x465, 0x0, x437);
+  fiat_secp256k1_addcarryx_u32(&x482, &x483, x461, 0x0, x433);
   uint32_t x484;
   fiat_secp256k1_uint1 x485;
-  fiat_secp256k1_addcarryx_u32(&x484, &x485, x481, x482, x432);
+  fiat_secp256k1_addcarryx_u32(&x484, &x485, x477, x482, x480);
   uint32_t x486;
   fiat_secp256k1_uint1 x487;
-  fiat_secp256k1_addcarryx_u32(&x486, &x487, 0x0, (arg1[7]), x468);
+  fiat_secp256k1_addcarryx_u32(&x486, &x487, 0x0, (arg1[7]), x464);
   uint32_t x488;
   fiat_secp256k1_uint1 x489;
-  fiat_secp256k1_addcarryx_u32(&x488, &x489, x487, 0x0, x470);
+  fiat_secp256k1_addcarryx_u32(&x488, &x489, x487, 0x0, x466);
   uint32_t x490;
   fiat_secp256k1_uint1 x491;
-  fiat_secp256k1_addcarryx_u32(&x490, &x491, x489, 0x0, x472);
+  fiat_secp256k1_addcarryx_u32(&x490, &x491, x489, 0x0, x468);
   uint32_t x492;
   fiat_secp256k1_uint1 x493;
-  fiat_secp256k1_addcarryx_u32(&x492, &x493, x491, 0x0, x474);
+  fiat_secp256k1_addcarryx_u32(&x492, &x493, x491, 0x0, x470);
   uint32_t x494;
   fiat_secp256k1_uint1 x495;
-  fiat_secp256k1_addcarryx_u32(&x494, &x495, x493, 0x0, x476);
+  fiat_secp256k1_addcarryx_u32(&x494, &x495, x493, 0x0, x472);
   uint32_t x496;
   fiat_secp256k1_uint1 x497;
-  fiat_secp256k1_addcarryx_u32(&x496, &x497, x495, 0x0, x478);
+  fiat_secp256k1_addcarryx_u32(&x496, &x497, x495, 0x0, x474);
   uint32_t x498;
   fiat_secp256k1_uint1 x499;
-  fiat_secp256k1_addcarryx_u32(&x498, &x499, x497, 0x0, x480);
+  fiat_secp256k1_addcarryx_u32(&x498, &x499, x497, 0x0, x476);
   uint32_t x500;
   fiat_secp256k1_uint1 x501;
   fiat_secp256k1_addcarryx_u32(&x500, &x501, x499, 0x0, x484);
   uint32_t x502;
-  fiat_secp256k1_uint1 x503;
-  fiat_secp256k1_addcarryx_u32(&x502, &x503, x485, 0x0, x433);
+  uint32_t x503;
+  fiat_secp256k1_mulx_u32(&x502, &x503, x486, UINT32_C(0xd2253531));
   uint32_t x504;
-  fiat_secp256k1_uint1 x505;
-  fiat_secp256k1_addcarryx_u32(&x504, &x505, x501, 0x0, (fiat_secp256k1_uint1)x502);
+  uint32_t x505;
+  fiat_secp256k1_mulx_u32(&x504, &x505, x502, UINT32_C(0xffffffff));
   uint32_t x506;
   uint32_t x507;
-  fiat_secp256k1_mulx_u32(&x506, &x507, x486, UINT32_C(0xd2253531));
+  fiat_secp256k1_mulx_u32(&x506, &x507, x502, UINT32_C(0xffffffff));
   uint32_t x508;
   uint32_t x509;
-  fiat_secp256k1_mulx_u32(&x508, &x509, x506, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x508, &x509, x502, UINT32_C(0xffffffff));
   uint32_t x510;
   uint32_t x511;
-  fiat_secp256k1_mulx_u32(&x510, &x511, x506, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x510, &x511, x502, UINT32_C(0xffffffff));
   uint32_t x512;
   uint32_t x513;
-  fiat_secp256k1_mulx_u32(&x512, &x513, x506, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x512, &x513, x502, UINT32_C(0xffffffff));
   uint32_t x514;
   uint32_t x515;
-  fiat_secp256k1_mulx_u32(&x514, &x515, x506, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x514, &x515, x502, UINT32_C(0xffffffff));
   uint32_t x516;
   uint32_t x517;
-  fiat_secp256k1_mulx_u32(&x516, &x517, x506, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x516, &x517, x502, UINT32_C(0xfffffffe));
   uint32_t x518;
   uint32_t x519;
-  fiat_secp256k1_mulx_u32(&x518, &x519, x506, UINT32_C(0xffffffff));
+  fiat_secp256k1_mulx_u32(&x518, &x519, x502, UINT32_C(0xfffffc2f));
   uint32_t x520;
-  uint32_t x521;
-  fiat_secp256k1_mulx_u32(&x520, &x521, x506, UINT32_C(0xfffffffe));
+  fiat_secp256k1_uint1 x521;
+  fiat_secp256k1_addcarryx_u32(&x520, &x521, 0x0, x516, x519);
   uint32_t x522;
-  uint32_t x523;
-  fiat_secp256k1_mulx_u32(&x522, &x523, x506, UINT32_C(0xfffffc2f));
+  fiat_secp256k1_uint1 x523;
+  fiat_secp256k1_addcarryx_u32(&x522, &x523, x521, x514, x517);
   uint32_t x524;
   fiat_secp256k1_uint1 x525;
-  fiat_secp256k1_addcarryx_u32(&x524, &x525, 0x0, x520, x523);
+  fiat_secp256k1_addcarryx_u32(&x524, &x525, x523, x512, x515);
   uint32_t x526;
   fiat_secp256k1_uint1 x527;
-  fiat_secp256k1_addcarryx_u32(&x526, &x527, x525, x518, x521);
+  fiat_secp256k1_addcarryx_u32(&x526, &x527, x525, x510, x513);
   uint32_t x528;
   fiat_secp256k1_uint1 x529;
-  fiat_secp256k1_addcarryx_u32(&x528, &x529, x527, x516, x519);
+  fiat_secp256k1_addcarryx_u32(&x528, &x529, x527, x508, x511);
   uint32_t x530;
   fiat_secp256k1_uint1 x531;
-  fiat_secp256k1_addcarryx_u32(&x530, &x531, x529, x514, x517);
+  fiat_secp256k1_addcarryx_u32(&x530, &x531, x529, x506, x509);
   uint32_t x532;
   fiat_secp256k1_uint1 x533;
-  fiat_secp256k1_addcarryx_u32(&x532, &x533, x531, x512, x515);
+  fiat_secp256k1_addcarryx_u32(&x532, &x533, x531, x504, x507);
   uint32_t x534;
   fiat_secp256k1_uint1 x535;
-  fiat_secp256k1_addcarryx_u32(&x534, &x535, x533, x510, x513);
+  fiat_secp256k1_addcarryx_u32(&x534, &x535, 0x0, x518, x486);
   uint32_t x536;
   fiat_secp256k1_uint1 x537;
-  fiat_secp256k1_addcarryx_u32(&x536, &x537, x535, x508, x511);
+  fiat_secp256k1_addcarryx_u32(&x536, &x537, x535, x520, x488);
   uint32_t x538;
   fiat_secp256k1_uint1 x539;
-  fiat_secp256k1_addcarryx_u32(&x538, &x539, 0x0, x522, x486);
+  fiat_secp256k1_addcarryx_u32(&x538, &x539, x537, x522, x490);
   uint32_t x540;
   fiat_secp256k1_uint1 x541;
-  fiat_secp256k1_addcarryx_u32(&x540, &x541, x539, x524, x488);
+  fiat_secp256k1_addcarryx_u32(&x540, &x541, x539, x524, x492);
   uint32_t x542;
   fiat_secp256k1_uint1 x543;
-  fiat_secp256k1_addcarryx_u32(&x542, &x543, x541, x526, x490);
+  fiat_secp256k1_addcarryx_u32(&x542, &x543, x541, x526, x494);
   uint32_t x544;
   fiat_secp256k1_uint1 x545;
-  fiat_secp256k1_addcarryx_u32(&x544, &x545, x543, x528, x492);
+  fiat_secp256k1_addcarryx_u32(&x544, &x545, x543, x528, x496);
   uint32_t x546;
   fiat_secp256k1_uint1 x547;
-  fiat_secp256k1_addcarryx_u32(&x546, &x547, x545, x530, x494);
+  fiat_secp256k1_addcarryx_u32(&x546, &x547, x545, x530, x498);
   uint32_t x548;
   fiat_secp256k1_uint1 x549;
-  fiat_secp256k1_addcarryx_u32(&x548, &x549, x547, x532, x496);
+  fiat_secp256k1_addcarryx_u32(&x548, &x549, x547, x532, x500);
   uint32_t x550;
   fiat_secp256k1_uint1 x551;
-  fiat_secp256k1_addcarryx_u32(&x550, &x551, x549, x534, x498);
+  fiat_secp256k1_addcarryx_u32(&x550, &x551, x485, 0x0, 0x0);
   uint32_t x552;
   fiat_secp256k1_uint1 x553;
-  fiat_secp256k1_addcarryx_u32(&x552, &x553, x551, x536, x500);
+  fiat_secp256k1_addcarryx_u32(&x552, &x553, x501, 0x0, (fiat_secp256k1_uint1)x550);
   uint32_t x554;
   fiat_secp256k1_uint1 x555;
-  fiat_secp256k1_addcarryx_u32(&x554, &x555, x537, 0x0, x509);
+  fiat_secp256k1_addcarryx_u32(&x554, &x555, x533, 0x0, x505);
   uint32_t x556;
   fiat_secp256k1_uint1 x557;
-  fiat_secp256k1_addcarryx_u32(&x556, &x557, x553, x554, x504);
+  fiat_secp256k1_addcarryx_u32(&x556, &x557, x549, x554, x552);
   uint32_t x558;
   fiat_secp256k1_uint1 x559;
-  fiat_secp256k1_subborrowx_u32(&x558, &x559, 0x0, x540, UINT32_C(0xfffffc2f));
+  fiat_secp256k1_subborrowx_u32(&x558, &x559, 0x0, x536, UINT32_C(0xfffffc2f));
   uint32_t x560;
   fiat_secp256k1_uint1 x561;
-  fiat_secp256k1_subborrowx_u32(&x560, &x561, x559, x542, UINT32_C(0xfffffffe));
+  fiat_secp256k1_subborrowx_u32(&x560, &x561, x559, x538, UINT32_C(0xfffffffe));
   uint32_t x562;
   fiat_secp256k1_uint1 x563;
-  fiat_secp256k1_subborrowx_u32(&x562, &x563, x561, x544, UINT32_C(0xffffffff));
+  fiat_secp256k1_subborrowx_u32(&x562, &x563, x561, x540, UINT32_C(0xffffffff));
   uint32_t x564;
   fiat_secp256k1_uint1 x565;
-  fiat_secp256k1_subborrowx_u32(&x564, &x565, x563, x546, UINT32_C(0xffffffff));
+  fiat_secp256k1_subborrowx_u32(&x564, &x565, x563, x542, UINT32_C(0xffffffff));
   uint32_t x566;
   fiat_secp256k1_uint1 x567;
-  fiat_secp256k1_subborrowx_u32(&x566, &x567, x565, x548, UINT32_C(0xffffffff));
+  fiat_secp256k1_subborrowx_u32(&x566, &x567, x565, x544, UINT32_C(0xffffffff));
   uint32_t x568;
   fiat_secp256k1_uint1 x569;
-  fiat_secp256k1_subborrowx_u32(&x568, &x569, x567, x550, UINT32_C(0xffffffff));
+  fiat_secp256k1_subborrowx_u32(&x568, &x569, x567, x546, UINT32_C(0xffffffff));
   uint32_t x570;
   fiat_secp256k1_uint1 x571;
-  fiat_secp256k1_subborrowx_u32(&x570, &x571, x569, x552, UINT32_C(0xffffffff));
+  fiat_secp256k1_subborrowx_u32(&x570, &x571, x569, x548, UINT32_C(0xffffffff));
   uint32_t x572;
   fiat_secp256k1_uint1 x573;
   fiat_secp256k1_subborrowx_u32(&x572, &x573, x571, x556, UINT32_C(0xffffffff));
   uint32_t x574;
   fiat_secp256k1_uint1 x575;
-  fiat_secp256k1_addcarryx_u32(&x574, &x575, x557, 0x0, x505);
+  fiat_secp256k1_addcarryx_u32(&x574, &x575, x557, 0x0, 0x0);
   uint32_t x576;
   fiat_secp256k1_uint1 x577;
   fiat_secp256k1_subborrowx_u32(&x576, &x577, x573, (fiat_secp256k1_uint1)x574, 0x0);
   uint32_t x578;
-  fiat_secp256k1_cmovznz_u32(&x578, x577, x558, x540);
+  fiat_secp256k1_cmovznz_u32(&x578, x577, x558, x536);
   uint32_t x579;
-  fiat_secp256k1_cmovznz_u32(&x579, x577, x560, x542);
+  fiat_secp256k1_cmovznz_u32(&x579, x577, x560, x538);
   uint32_t x580;
-  fiat_secp256k1_cmovznz_u32(&x580, x577, x562, x544);
+  fiat_secp256k1_cmovznz_u32(&x580, x577, x562, x540);
   uint32_t x581;
-  fiat_secp256k1_cmovznz_u32(&x581, x577, x564, x546);
+  fiat_secp256k1_cmovznz_u32(&x581, x577, x564, x542);
   uint32_t x582;
-  fiat_secp256k1_cmovznz_u32(&x582, x577, x566, x548);
+  fiat_secp256k1_cmovznz_u32(&x582, x577, x566, x544);
   uint32_t x583;
-  fiat_secp256k1_cmovznz_u32(&x583, x577, x568, x550);
+  fiat_secp256k1_cmovznz_u32(&x583, x577, x568, x546);
   uint32_t x584;
-  fiat_secp256k1_cmovznz_u32(&x584, x577, x570, x552);
+  fiat_secp256k1_cmovznz_u32(&x584, x577, x570, x548);
   uint32_t x585;
   fiat_secp256k1_cmovznz_u32(&x585, x577, x572, x556);
   out1[0] = x578;
@@ -3845,101 +3845,94 @@ static void fiat_secp256k1_to_bytes(uint8_t out1[32], const uint32_t arg1[8]) {
   uint8_t x12 = (uint8_t)(x9 & UINT8_C(0xff));
   uint8_t x13 = (uint8_t)(x11 >> 8);
   uint8_t x14 = (uint8_t)(x11 & UINT8_C(0xff));
-  fiat_secp256k1_uint1 x15 = (fiat_secp256k1_uint1)(x13 >> 8);
-  uint8_t x16 = (uint8_t)(x13 & UINT8_C(0xff));
-  uint32_t x17 = (x15 + x7);
-  uint32_t x18 = (x17 >> 8);
-  uint8_t x19 = (uint8_t)(x17 & UINT8_C(0xff));
-  uint32_t x20 = (x18 >> 8);
-  uint8_t x21 = (uint8_t)(x18 & UINT8_C(0xff));
-  uint8_t x22 = (uint8_t)(x20 >> 8);
-  uint8_t x23 = (uint8_t)(x20 & UINT8_C(0xff));
-  fiat_secp256k1_uint1 x24 = (fiat_secp256k1_uint1)(x22 >> 8);
-  uint8_t x25 = (uint8_t)(x22 & UINT8_C(0xff));
-  uint32_t x26 = (x24 + x6);
-  uint32_t x27 = (x26 >> 8);
-  uint8_t x28 = (uint8_t)(x26 & UINT8_C(0xff));
-  uint32_t x29 = (x27 >> 8);
+  uint8_t x15 = (uint8_t)(x13 & UINT8_C(0xff));
+  uint32_t x16 = (0x0 + x7);
+  uint32_t x17 = (x16 >> 8);
+  uint8_t x18 = (uint8_t)(x16 & UINT8_C(0xff));
+  uint32_t x19 = (x17 >> 8);
+  uint8_t x20 = (uint8_t)(x17 & UINT8_C(0xff));
+  uint8_t x21 = (uint8_t)(x19 >> 8);
+  uint8_t x22 = (uint8_t)(x19 & UINT8_C(0xff));
+  uint8_t x23 = (uint8_t)(x21 & UINT8_C(0xff));
+  uint32_t x24 = (0x0 + x6);
+  uint32_t x25 = (x24 >> 8);
+  uint8_t x26 = (uint8_t)(x24 & UINT8_C(0xff));
+  uint32_t x27 = (x25 >> 8);
+  uint8_t x28 = (uint8_t)(x25 & UINT8_C(0xff));
+  uint8_t x29 = (uint8_t)(x27 >> 8);
   uint8_t x30 = (uint8_t)(x27 & UINT8_C(0xff));
-  uint8_t x31 = (uint8_t)(x29 >> 8);
-  uint8_t x32 = (uint8_t)(x29 & UINT8_C(0xff));
-  fiat_secp256k1_uint1 x33 = (fiat_secp256k1_uint1)(x31 >> 8);
-  uint8_t x34 = (uint8_t)(x31 & UINT8_C(0xff));
-  uint32_t x35 = (x33 + x5);
-  uint32_t x36 = (x35 >> 8);
-  uint8_t x37 = (uint8_t)(x35 & UINT8_C(0xff));
-  uint32_t x38 = (x36 >> 8);
-  uint8_t x39 = (uint8_t)(x36 & UINT8_C(0xff));
-  uint8_t x40 = (uint8_t)(x38 >> 8);
-  uint8_t x41 = (uint8_t)(x38 & UINT8_C(0xff));
-  fiat_secp256k1_uint1 x42 = (fiat_secp256k1_uint1)(x40 >> 8);
-  uint8_t x43 = (uint8_t)(x40 & UINT8_C(0xff));
-  uint32_t x44 = (x42 + x4);
-  uint32_t x45 = (x44 >> 8);
-  uint8_t x46 = (uint8_t)(x44 & UINT8_C(0xff));
-  uint32_t x47 = (x45 >> 8);
-  uint8_t x48 = (uint8_t)(x45 & UINT8_C(0xff));
-  uint8_t x49 = (uint8_t)(x47 >> 8);
-  uint8_t x50 = (uint8_t)(x47 & UINT8_C(0xff));
-  fiat_secp256k1_uint1 x51 = (fiat_secp256k1_uint1)(x49 >> 8);
+  uint8_t x31 = (uint8_t)(x29 & UINT8_C(0xff));
+  uint32_t x32 = (0x0 + x5);
+  uint32_t x33 = (x32 >> 8);
+  uint8_t x34 = (uint8_t)(x32 & UINT8_C(0xff));
+  uint32_t x35 = (x33 >> 8);
+  uint8_t x36 = (uint8_t)(x33 & UINT8_C(0xff));
+  uint8_t x37 = (uint8_t)(x35 >> 8);
+  uint8_t x38 = (uint8_t)(x35 & UINT8_C(0xff));
+  uint8_t x39 = (uint8_t)(x37 & UINT8_C(0xff));
+  uint32_t x40 = (0x0 + x4);
+  uint32_t x41 = (x40 >> 8);
+  uint8_t x42 = (uint8_t)(x40 & UINT8_C(0xff));
+  uint32_t x43 = (x41 >> 8);
+  uint8_t x44 = (uint8_t)(x41 & UINT8_C(0xff));
+  uint8_t x45 = (uint8_t)(x43 >> 8);
+  uint8_t x46 = (uint8_t)(x43 & UINT8_C(0xff));
+  uint8_t x47 = (uint8_t)(x45 & UINT8_C(0xff));
+  uint32_t x48 = (0x0 + x3);
+  uint32_t x49 = (x48 >> 8);
+  uint8_t x50 = (uint8_t)(x48 & UINT8_C(0xff));
+  uint32_t x51 = (x49 >> 8);
   uint8_t x52 = (uint8_t)(x49 & UINT8_C(0xff));
-  uint32_t x53 = (x51 + x3);
-  uint32_t x54 = (x53 >> 8);
+  uint8_t x53 = (uint8_t)(x51 >> 8);
+  uint8_t x54 = (uint8_t)(x51 & UINT8_C(0xff));
   uint8_t x55 = (uint8_t)(x53 & UINT8_C(0xff));
-  uint32_t x56 = (x54 >> 8);
-  uint8_t x57 = (uint8_t)(x54 & UINT8_C(0xff));
-  uint8_t x58 = (uint8_t)(x56 >> 8);
-  uint8_t x59 = (uint8_t)(x56 & UINT8_C(0xff));
-  fiat_secp256k1_uint1 x60 = (fiat_secp256k1_uint1)(x58 >> 8);
-  uint8_t x61 = (uint8_t)(x58 & UINT8_C(0xff));
-  uint32_t x62 = (x60 + x2);
-  uint32_t x63 = (x62 >> 8);
-  uint8_t x64 = (uint8_t)(x62 & UINT8_C(0xff));
-  uint32_t x65 = (x63 >> 8);
-  uint8_t x66 = (uint8_t)(x63 & UINT8_C(0xff));
-  uint8_t x67 = (uint8_t)(x65 >> 8);
+  uint32_t x56 = (0x0 + x2);
+  uint32_t x57 = (x56 >> 8);
+  uint8_t x58 = (uint8_t)(x56 & UINT8_C(0xff));
+  uint32_t x59 = (x57 >> 8);
+  uint8_t x60 = (uint8_t)(x57 & UINT8_C(0xff));
+  uint8_t x61 = (uint8_t)(x59 >> 8);
+  uint8_t x62 = (uint8_t)(x59 & UINT8_C(0xff));
+  uint8_t x63 = (uint8_t)(x61 & UINT8_C(0xff));
+  uint32_t x64 = (0x0 + x1);
+  uint32_t x65 = (x64 >> 8);
+  uint8_t x66 = (uint8_t)(x64 & UINT8_C(0xff));
+  uint32_t x67 = (x65 >> 8);
   uint8_t x68 = (uint8_t)(x65 & UINT8_C(0xff));
-  fiat_secp256k1_uint1 x69 = (fiat_secp256k1_uint1)(x67 >> 8);
+  uint8_t x69 = (uint8_t)(x67 >> 8);
   uint8_t x70 = (uint8_t)(x67 & UINT8_C(0xff));
-  uint32_t x71 = (x69 + x1);
-  uint32_t x72 = (x71 >> 8);
-  uint8_t x73 = (uint8_t)(x71 & UINT8_C(0xff));
-  uint32_t x74 = (x72 >> 8);
-  uint8_t x75 = (uint8_t)(x72 & UINT8_C(0xff));
-  uint8_t x76 = (uint8_t)(x74 >> 8);
-  uint8_t x77 = (uint8_t)(x74 & UINT8_C(0xff));
   out1[0] = x10;
   out1[1] = x12;
   out1[2] = x14;
-  out1[3] = x16;
-  out1[4] = x19;
-  out1[5] = x21;
-  out1[6] = x23;
-  out1[7] = x25;
-  out1[8] = x28;
-  out1[9] = x30;
-  out1[10] = x32;
-  out1[11] = x34;
-  out1[12] = x37;
-  out1[13] = x39;
-  out1[14] = x41;
-  out1[15] = x43;
-  out1[16] = x46;
-  out1[17] = x48;
-  out1[18] = x50;
-  out1[19] = x52;
-  out1[20] = x55;
-  out1[21] = x57;
-  out1[22] = x59;
-  out1[23] = x61;
-  out1[24] = x64;
-  out1[25] = x66;
-  out1[26] = x68;
-  out1[27] = x70;
-  out1[28] = x73;
-  out1[29] = x75;
-  out1[30] = x77;
-  out1[31] = x76;
+  out1[3] = x15;
+  out1[4] = x18;
+  out1[5] = x20;
+  out1[6] = x22;
+  out1[7] = x23;
+  out1[8] = x26;
+  out1[9] = x28;
+  out1[10] = x30;
+  out1[11] = x31;
+  out1[12] = x34;
+  out1[13] = x36;
+  out1[14] = x38;
+  out1[15] = x39;
+  out1[16] = x42;
+  out1[17] = x44;
+  out1[18] = x46;
+  out1[19] = x47;
+  out1[20] = x50;
+  out1[21] = x52;
+  out1[22] = x54;
+  out1[23] = x55;
+  out1[24] = x58;
+  out1[25] = x60;
+  out1[26] = x62;
+  out1[27] = x63;
+  out1[28] = x66;
+  out1[29] = x68;
+  out1[30] = x70;
+  out1[31] = x69;
 }
 
 /*
@@ -3982,41 +3975,34 @@ static void fiat_secp256k1_from_bytes(uint32_t out1[8], const uint8_t arg1[32])
   uint32_t x31 = ((uint32_t)(arg1[1]) << 8);
   uint8_t x32 = (arg1[0]);
   uint32_t x33 = (x32 + (x31 + (x30 + x29)));
-  fiat_secp256k1_uint1 x34 = (fiat_secp256k1_uint1)((uint64_t)x33 >> 32);
-  uint32_t x35 = (x33 & UINT32_C(0xffffffff));
-  uint32_t x36 = (x4 + (x3 + (x2 + x1)));
-  uint32_t x37 = (x8 + (x7 + (x6 + x5)));
-  uint32_t x38 = (x12 + (x11 + (x10 + x9)));
-  uint32_t x39 = (x16 + (x15 + (x14 + x13)));
-  uint32_t x40 = (x20 + (x19 + (x18 + x17)));
-  uint32_t x41 = (x24 + (x23 + (x22 + x21)));
-  uint32_t x42 = (x28 + (x27 + (x26 + x25)));
-  uint32_t x43 = (x34 + x42);
-  fiat_secp256k1_uint1 x44 = (fiat_secp256k1_uint1)((uint64_t)x43 >> 32);
-  uint32_t x45 = (x43 & UINT32_C(0xffffffff));
-  uint32_t x46 = (x44 + x41);
-  fiat_secp256k1_uint1 x47 = (fiat_secp256k1_uint1)((uint64_t)x46 >> 32);
-  uint32_t x48 = (x46 & UINT32_C(0xffffffff));
-  uint32_t x49 = (x47 + x40);
-  fiat_secp256k1_uint1 x50 = (fiat_secp256k1_uint1)((uint64_t)x49 >> 32);
-  uint32_t x51 = (x49 & UINT32_C(0xffffffff));
-  uint32_t x52 = (x50 + x39);
-  fiat_secp256k1_uint1 x53 = (fiat_secp256k1_uint1)((uint64_t)x52 >> 32);
-  uint32_t x54 = (x52 & UINT32_C(0xffffffff));
-  uint32_t x55 = (x53 + x38);
-  fiat_secp256k1_uint1 x56 = (fiat_secp256k1_uint1)((uint64_t)x55 >> 32);
-  uint32_t x57 = (x55 & UINT32_C(0xffffffff));
-  uint32_t x58 = (x56 + x37);
-  fiat_secp256k1_uint1 x59 = (fiat_secp256k1_uint1)((uint64_t)x58 >> 32);
-  uint32_t x60 = (x58 & UINT32_C(0xffffffff));
-  uint32_t x61 = (x59 + x36);
-  out1[0] = x35;
-  out1[1] = x45;
-  out1[2] = x48;
-  out1[3] = x51;
-  out1[4] = x54;
-  out1[5] = x57;
-  out1[6] = x60;
-  out1[7] = x61;
+  uint32_t x34 = (x33 & UINT32_C(0xffffffff));
+  uint32_t x35 = (x4 + (x3 + (x2 + x1)));
+  uint32_t x36 = (x8 + (x7 + (x6 + x5)));
+  uint32_t x37 = (x12 + (x11 + (x10 + x9)));
+  uint32_t x38 = (x16 + (x15 + (x14 + x13)));
+  uint32_t x39 = (x20 + (x19 + (x18 + x17)));
+  uint32_t x40 = (x24 + (x23 + (x22 + x21)));
+  uint32_t x41 = (x28 + (x27 + (x26 + x25)));
+  uint32_t x42 = (0x0 + x41);
+  uint32_t x43 = (x42 & UINT32_C(0xffffffff));
+  uint32_t x44 = (0x0 + x40);
+  uint32_t x45 = (x44 & UINT32_C(0xffffffff));
+  uint32_t x46 = (0x0 + x39);
+  uint32_t x47 = (x46 & UINT32_C(0xffffffff));
+  uint32_t x48 = (0x0 + x38);
+  uint32_t x49 = (x48 & UINT32_C(0xffffffff));
+  uint32_t x50 = (0x0 + x37);
+  uint32_t x51 = (x50 & UINT32_C(0xffffffff));
+  uint32_t x52 = (0x0 + x36);
+  uint32_t x53 = (x52 & UINT32_C(0xffffffff));
+  uint32_t x54 = (0x0 + x35);
+  out1[0] = x34;
+  out1[1] = x43;
+  out1[2] = x45;
+  out1[3] = x47;
+  out1[4] = x49;
+  out1[5] = x51;
+  out1[6] = x53;
+  out1[7] = x54;
 }
 
diff --git a/secp256k1_64.c b/secp256k1_64.c
index 749e9bc40..a80485a4f 100644
--- a/secp256k1_64.c
+++ b/secp256k1_64.c
@@ -997,197 +997,197 @@ static void fiat_secp256k1_from_montgomery(uint64_t out1[4], const uint64_t arg1
   fiat_secp256k1_uint1 x37;
   fiat_secp256k1_addcarryx_u64(&x36, &x37, x35, 0x0, x28);
   uint64_t x38;
-  fiat_secp256k1_uint1 x39;
-  fiat_secp256k1_addcarryx_u64(&x38, &x39, x29, 0x0, 0x0);
+  uint64_t x39;
+  fiat_secp256k1_mulx_u64(&x38, &x39, x30, UINT64_C(0xd838091dd2253531));
   uint64_t x40;
-  fiat_secp256k1_uint1 x41;
-  fiat_secp256k1_addcarryx_u64(&x40, &x41, x37, 0x0, (fiat_secp256k1_uint1)x38);
+  uint64_t x41;
+  fiat_secp256k1_mulx_u64(&x40, &x41, x38, UINT64_C(0xffffffffffffffff));
   uint64_t x42;
   uint64_t x43;
-  fiat_secp256k1_mulx_u64(&x42, &x43, x30, UINT64_C(0xd838091dd2253531));
+  fiat_secp256k1_mulx_u64(&x42, &x43, x38, UINT64_C(0xffffffffffffffff));
   uint64_t x44;
   uint64_t x45;
-  fiat_secp256k1_mulx_u64(&x44, &x45, x42, UINT64_C(0xffffffffffffffff));
+  fiat_secp256k1_mulx_u64(&x44, &x45, x38, UINT64_C(0xffffffffffffffff));
   uint64_t x46;
   uint64_t x47;
-  fiat_secp256k1_mulx_u64(&x46, &x47, x42, UINT64_C(0xffffffffffffffff));
+  fiat_secp256k1_mulx_u64(&x46, &x47, x38, UINT64_C(0xfffffffefffffc2f));
   uint64_t x48;
-  uint64_t x49;
-  fiat_secp256k1_mulx_u64(&x48, &x49, x42, UINT64_C(0xffffffffffffffff));
+  fiat_secp256k1_uint1 x49;
+  fiat_secp256k1_addcarryx_u64(&x48, &x49, 0x0, x44, x47);
   uint64_t x50;
-  uint64_t x51;
-  fiat_secp256k1_mulx_u64(&x50, &x51, x42, UINT64_C(0xfffffffefffffc2f));
+  fiat_secp256k1_uint1 x51;
+  fiat_secp256k1_addcarryx_u64(&x50, &x51, x49, x42, x45);
   uint64_t x52;
   fiat_secp256k1_uint1 x53;
-  fiat_secp256k1_addcarryx_u64(&x52, &x53, 0x0, x48, x51);
+  fiat_secp256k1_addcarryx_u64(&x52, &x53, x51, x40, x43);
   uint64_t x54;
   fiat_secp256k1_uint1 x55;
-  fiat_secp256k1_addcarryx_u64(&x54, &x55, x53, x46, x49);
+  fiat_secp256k1_addcarryx_u64(&x54, &x55, 0x0, x46, x30);
   uint64_t x56;
   fiat_secp256k1_uint1 x57;
-  fiat_secp256k1_addcarryx_u64(&x56, &x57, x55, x44, x47);
+  fiat_secp256k1_addcarryx_u64(&x56, &x57, x55, x48, x32);
   uint64_t x58;
   fiat_secp256k1_uint1 x59;
-  fiat_secp256k1_addcarryx_u64(&x58, &x59, 0x0, x50, x30);
+  fiat_secp256k1_addcarryx_u64(&x58, &x59, x57, x50, x34);
   uint64_t x60;
   fiat_secp256k1_uint1 x61;
-  fiat_secp256k1_addcarryx_u64(&x60, &x61, x59, x52, x32);
+  fiat_secp256k1_addcarryx_u64(&x60, &x61, x59, x52, x36);
   uint64_t x62;
   fiat_secp256k1_uint1 x63;
-  fiat_secp256k1_addcarryx_u64(&x62, &x63, x61, x54, x34);
+  fiat_secp256k1_addcarryx_u64(&x62, &x63, x29, 0x0, 0x0);
   uint64_t x64;
   fiat_secp256k1_uint1 x65;
-  fiat_secp256k1_addcarryx_u64(&x64, &x65, x63, x56, x36);
+  fiat_secp256k1_addcarryx_u64(&x64, &x65, x37, 0x0, (fiat_secp256k1_uint1)x62);
   uint64_t x66;
   fiat_secp256k1_uint1 x67;
-  fiat_secp256k1_addcarryx_u64(&x66, &x67, x57, 0x0, x45);
+  fiat_secp256k1_addcarryx_u64(&x66, &x67, x53, 0x0, x41);
   uint64_t x68;
   fiat_secp256k1_uint1 x69;
-  fiat_secp256k1_addcarryx_u64(&x68, &x69, x65, x66, x40);
+  fiat_secp256k1_addcarryx_u64(&x68, &x69, x61, x66, x64);
   uint64_t x70;
   fiat_secp256k1_uint1 x71;
-  fiat_secp256k1_addcarryx_u64(&x70, &x71, 0x0, (arg1[2]), x60);
+  fiat_secp256k1_addcarryx_u64(&x70, &x71, 0x0, (arg1[2]), x56);
   uint64_t x72;
   fiat_secp256k1_uint1 x73;
-  fiat_secp256k1_addcarryx_u64(&x72, &x73, x71, 0x0, x62);
+  fiat_secp256k1_addcarryx_u64(&x72, &x73, x71, 0x0, x58);
   uint64_t x74;
   fiat_secp256k1_uint1 x75;
-  fiat_secp256k1_addcarryx_u64(&x74, &x75, x73, 0x0, x64);
+  fiat_secp256k1_addcarryx_u64(&x74, &x75, x73, 0x0, x60);
   uint64_t x76;
   fiat_secp256k1_uint1 x77;
   fiat_secp256k1_addcarryx_u64(&x76, &x77, x75, 0x0, x68);
   uint64_t x78;
-  fiat_secp256k1_uint1 x79;
-  fiat_secp256k1_addcarryx_u64(&x78, &x79, x69, 0x0, x41);
+  uint64_t x79;
+  fiat_secp256k1_mulx_u64(&x78, &x79, x70, UINT64_C(0xd838091dd2253531));
   uint64_t x80;
-  fiat_secp256k1_uint1 x81;
-  fiat_secp256k1_addcarryx_u64(&x80, &x81, x77, 0x0, (fiat_secp256k1_uint1)x78);
+  uint64_t x81;
+  fiat_secp256k1_mulx_u64(&x80, &x81, x78, UINT64_C(0xffffffffffffffff));
   uint64_t x82;
   uint64_t x83;
-  fiat_secp256k1_mulx_u64(&x82, &x83, x70, UINT64_C(0xd838091dd2253531));
+  fiat_secp256k1_mulx_u64(&x82, &x83, x78, UINT64_C(0xffffffffffffffff));
   uint64_t x84;
   uint64_t x85;
-  fiat_secp256k1_mulx_u64(&x84, &x85, x82, UINT64_C(0xffffffffffffffff));
+  fiat_secp256k1_mulx_u64(&x84, &x85, x78, UINT64_C(0xffffffffffffffff));
   uint64_t x86;
   uint64_t x87;
-  fiat_secp256k1_mulx_u64(&x86, &x87, x82, UINT64_C(0xffffffffffffffff));
+  fiat_secp256k1_mulx_u64(&x86, &x87, x78, UINT64_C(0xfffffffefffffc2f));
   uint64_t x88;
-  uint64_t x89;
-  fiat_secp256k1_mulx_u64(&x88, &x89, x82, UINT64_C(0xffffffffffffffff));
+  fiat_secp256k1_uint1 x89;
+  fiat_secp256k1_addcarryx_u64(&x88, &x89, 0x0, x84, x87);
   uint64_t x90;
-  uint64_t x91;
-  fiat_secp256k1_mulx_u64(&x90, &x91, x82, UINT64_C(0xfffffffefffffc2f));
+  fiat_secp256k1_uint1 x91;
+  fiat_secp256k1_addcarryx_u64(&x90, &x91, x89, x82, x85);
   uint64_t x92;
   fiat_secp256k1_uint1 x93;
-  fiat_secp256k1_addcarryx_u64(&x92, &x93, 0x0, x88, x91);
+  fiat_secp256k1_addcarryx_u64(&x92, &x93, x91, x80, x83);
   uint64_t x94;
   fiat_secp256k1_uint1 x95;
-  fiat_secp256k1_addcarryx_u64(&x94, &x95, x93, x86, x89);
+  fiat_secp256k1_addcarryx_u64(&x94, &x95, 0x0, x86, x70);
   uint64_t x96;
   fiat_secp256k1_uint1 x97;
-  fiat_secp256k1_addcarryx_u64(&x96, &x97, x95, x84, x87);
+  fiat_secp256k1_addcarryx_u64(&x96, &x97, x95, x88, x72);
   uint64_t x98;
   fiat_secp256k1_uint1 x99;
-  fiat_secp256k1_addcarryx_u64(&x98, &x99, 0x0, x90, x70);
+  fiat_secp256k1_addcarryx_u64(&x98, &x99, x97, x90, x74);
   uint64_t x100;
   fiat_secp256k1_uint1 x101;
-  fiat_secp256k1_addcarryx_u64(&x100, &x101, x99, x92, x72);
+  fiat_secp256k1_addcarryx_u64(&x100, &x101, x99, x92, x76);
   uint64_t x102;
   fiat_secp256k1_uint1 x103;
-  fiat_secp256k1_addcarryx_u64(&x102, &x103, x101, x94, x74);
+  fiat_secp256k1_addcarryx_u64(&x102, &x103, x69, 0x0, 0x0);
   uint64_t x104;
   fiat_secp256k1_uint1 x105;
-  fiat_secp256k1_addcarryx_u64(&x104, &x105, x103, x96, x76);
+  fiat_secp256k1_addcarryx_u64(&x104, &x105, x77, 0x0, (fiat_secp256k1_uint1)x102);
   uint64_t x106;
   fiat_secp256k1_uint1 x107;
-  fiat_secp256k1_addcarryx_u64(&x106, &x107, x97, 0x0, x85);
+  fiat_secp256k1_addcarryx_u64(&x106, &x107, x93, 0x0, x81);
   uint64_t x108;
   fiat_secp256k1_uint1 x109;
-  fiat_secp256k1_addcarryx_u64(&x108, &x109, x105, x106, x80);
+  fiat_secp256k1_addcarryx_u64(&x108, &x109, x101, x106, x104);
   uint64_t x110;
   fiat_secp256k1_uint1 x111;
-  fiat_secp256k1_addcarryx_u64(&x110, &x111, 0x0, (arg1[3]), x100);
+  fiat_secp256k1_addcarryx_u64(&x110, &x111, 0x0, (arg1[3]), x96);
   uint64_t x112;
   fiat_secp256k1_uint1 x113;
-  fiat_secp256k1_addcarryx_u64(&x112, &x113, x111, 0x0, x102);
+  fiat_secp256k1_addcarryx_u64(&x112, &x113, x111, 0x0, x98);
   uint64_t x114;
   fiat_secp256k1_uint1 x115;
-  fiat_secp256k1_addcarryx_u64(&x114, &x115, x113, 0x0, x104);
+  fiat_secp256k1_addcarryx_u64(&x114, &x115, x113, 0x0, x100);
   uint64_t x116;
   fiat_secp256k1_uint1 x117;
   fiat_secp256k1_addcarryx_u64(&x116, &x117, x115, 0x0, x108);
   uint64_t x118;
-  fiat_secp256k1_uint1 x119;
-  fiat_secp256k1_addcarryx_u64(&x118, &x119, x109, 0x0, x81);
+  uint64_t x119;
+  fiat_secp256k1_mulx_u64(&x118, &x119, x110, UINT64_C(0xd838091dd2253531));
   uint64_t x120;
-  fiat_secp256k1_uint1 x121;
-  fiat_secp256k1_addcarryx_u64(&x120, &x121, x117, 0x0, (fiat_secp256k1_uint1)x118);
+  uint64_t x121;
+  fiat_secp256k1_mulx_u64(&x120, &x121, x118, UINT64_C(0xffffffffffffffff));
   uint64_t x122;
   uint64_t x123;
-  fiat_secp256k1_mulx_u64(&x122, &x123, x110, UINT64_C(0xd838091dd2253531));
+  fiat_secp256k1_mulx_u64(&x122, &x123, x118, UINT64_C(0xffffffffffffffff));
   uint64_t x124;
   uint64_t x125;
-  fiat_secp256k1_mulx_u64(&x124, &x125, x122, UINT64_C(0xffffffffffffffff));
+  fiat_secp256k1_mulx_u64(&x124, &x125, x118, UINT64_C(0xffffffffffffffff));
   uint64_t x126;
   uint64_t x127;
-  fiat_secp256k1_mulx_u64(&x126, &x127, x122, UINT64_C(0xffffffffffffffff));
+  fiat_secp256k1_mulx_u64(&x126, &x127, x118, UINT64_C(0xfffffffefffffc2f));
   uint64_t x128;
-  uint64_t x129;
-  fiat_secp256k1_mulx_u64(&x128, &x129, x122, UINT64_C(0xffffffffffffffff));
+  fiat_secp256k1_uint1 x129;
+  fiat_secp256k1_addcarryx_u64(&x128, &x129, 0x0, x124, x127);
   uint64_t x130;
-  uint64_t x131;
-  fiat_secp256k1_mulx_u64(&x130, &x131, x122, UINT64_C(0xfffffffefffffc2f));
+  fiat_secp256k1_uint1 x131;
+  fiat_secp256k1_addcarryx_u64(&x130, &x131, x129, x122, x125);
   uint64_t x132;
   fiat_secp256k1_uint1 x133;
-  fiat_secp256k1_addcarryx_u64(&x132, &x133, 0x0, x128, x131);
+  fiat_secp256k1_addcarryx_u64(&x132, &x133, x131, x120, x123);
   uint64_t x134;
   fiat_secp256k1_uint1 x135;
-  fiat_secp256k1_addcarryx_u64(&x134, &x135, x133, x126, x129);
+  fiat_secp256k1_addcarryx_u64(&x134, &x135, 0x0, x126, x110);
   uint64_t x136;
   fiat_secp256k1_uint1 x137;
-  fiat_secp256k1_addcarryx_u64(&x136, &x137, x135, x124, x127);
+  fiat_secp256k1_addcarryx_u64(&x136, &x137, x135, x128, x112);
   uint64_t x138;
   fiat_secp256k1_uint1 x139;
-  fiat_secp256k1_addcarryx_u64(&x138, &x139, 0x0, x130, x110);
+  fiat_secp256k1_addcarryx_u64(&x138, &x139, x137, x130, x114);
   uint64_t x140;
   fiat_secp256k1_uint1 x141;
-  fiat_secp256k1_addcarryx_u64(&x140, &x141, x139, x132, x112);
+  fiat_secp256k1_addcarryx_u64(&x140, &x141, x139, x132, x116);
   uint64_t x142;
   fiat_secp256k1_uint1 x143;
-  fiat_secp256k1_addcarryx_u64(&x142, &x143, x141, x134, x114);
+  fiat_secp256k1_addcarryx_u64(&x142, &x143, x109, 0x0, 0x0);
   uint64_t x144;
   fiat_secp256k1_uint1 x145;
-  fiat_secp256k1_addcarryx_u64(&x144, &x145, x143, x136, x116);
+  fiat_secp256k1_addcarryx_u64(&x144, &x145, x117, 0x0, (fiat_secp256k1_uint1)x142);
   uint64_t x146;
   fiat_secp256k1_uint1 x147;
-  fiat_secp256k1_addcarryx_u64(&x146, &x147, x137, 0x0, x125);
+  fiat_secp256k1_addcarryx_u64(&x146, &x147, x133, 0x0, x121);
   uint64_t x148;
   fiat_secp256k1_uint1 x149;
-  fiat_secp256k1_addcarryx_u64(&x148, &x149, x145, x146, x120);
+  fiat_secp256k1_addcarryx_u64(&x148, &x149, x141, x146, x144);
   uint64_t x150;
   fiat_secp256k1_uint1 x151;
-  fiat_secp256k1_subborrowx_u64(&x150, &x151, 0x0, x140, UINT64_C(0xfffffffefffffc2f));
+  fiat_secp256k1_subborrowx_u64(&x150, &x151, 0x0, x136, UINT64_C(0xfffffffefffffc2f));
   uint64_t x152;
   fiat_secp256k1_uint1 x153;
-  fiat_secp256k1_subborrowx_u64(&x152, &x153, x151, x142, UINT64_C(0xffffffffffffffff));
+  fiat_secp256k1_subborrowx_u64(&x152, &x153, x151, x138, UINT64_C(0xffffffffffffffff));
   uint64_t x154;
   fiat_secp256k1_uint1 x155;
-  fiat_secp256k1_subborrowx_u64(&x154, &x155, x153, x144, UINT64_C(0xffffffffffffffff));
+  fiat_secp256k1_subborrowx_u64(&x154, &x155, x153, x140, UINT64_C(0xffffffffffffffff));
   uint64_t x156;
   fiat_secp256k1_uint1 x157;
   fiat_secp256k1_subborrowx_u64(&x156, &x157, x155, x148, UINT64_C(0xffffffffffffffff));
   uint64_t x158;
   fiat_secp256k1_uint1 x159;
-  fiat_secp256k1_addcarryx_u64(&x158, &x159, x149, 0x0, x121);
+  fiat_secp256k1_addcarryx_u64(&x158, &x159, x149, 0x0, 0x0);
   uint64_t x160;
   fiat_secp256k1_uint1 x161;
   fiat_secp256k1_subborrowx_u64(&x160, &x161, x157, (fiat_secp256k1_uint1)x158, 0x0);
   uint64_t x162;
-  fiat_secp256k1_cmovznz_u64(&x162, x161, x150, x140);
+  fiat_secp256k1_cmovznz_u64(&x162, x161, x150, x136);
   uint64_t x163;
-  fiat_secp256k1_cmovznz_u64(&x163, x161, x152, x142);
+  fiat_secp256k1_cmovznz_u64(&x163, x161, x152, x138);
   uint64_t x164;
-  fiat_secp256k1_cmovznz_u64(&x164, x161, x154, x144);
+  fiat_secp256k1_cmovznz_u64(&x164, x161, x154, x140);
   uint64_t x165;
   fiat_secp256k1_cmovznz_u64(&x165, x161, x156, x148);
   out1[0] = x162;
@@ -1255,28 +1255,28 @@ static void fiat_secp256k1_to_bytes(uint8_t out1[32], const uint64_t arg1[4]) {
   uint8_t x16 = (uint8_t)(x13 & UINT8_C(0xff));
   uint8_t x17 = (uint8_t)(x15 >> 8);
   uint8_t x18 = (uint8_t)(x15 & UINT8_C(0xff));
-  fiat_secp256k1_uint1 x19 = (fiat_secp256k1_uint1)(x17 >> 8);
-  uint8_t x20 = (uint8_t)(x17 & UINT8_C(0xff));
-  uint64_t x21 = (x19 + x3);
-  uint64_t x22 = (x21 >> 8);
-  uint8_t x23 = (uint8_t)(x21 & UINT8_C(0xff));
-  uint64_t x24 = (x22 >> 8);
-  uint8_t x25 = (uint8_t)(x22 & UINT8_C(0xff));
-  uint64_t x26 = (x24 >> 8);
-  uint8_t x27 = (uint8_t)(x24 & UINT8_C(0xff));
-  uint64_t x28 = (x26 >> 8);
-  uint8_t x29 = (uint8_t)(x26 & UINT8_C(0xff));
-  uint64_t x30 = (x28 >> 8);
-  uint8_t x31 = (uint8_t)(x28 & UINT8_C(0xff));
-  uint64_t x32 = (x30 >> 8);
-  uint8_t x33 = (uint8_t)(x30 & UINT8_C(0xff));
-  uint8_t x34 = (uint8_t)(x32 >> 8);
-  uint8_t x35 = (uint8_t)(x32 & UINT8_C(0xff));
-  fiat_secp256k1_uint1 x36 = (fiat_secp256k1_uint1)(x34 >> 8);
-  uint8_t x37 = (uint8_t)(x34 & UINT8_C(0xff));
-  uint64_t x38 = (x36 + x2);
-  uint64_t x39 = (x38 >> 8);
-  uint8_t x40 = (uint8_t)(x38 & UINT8_C(0xff));
+  uint8_t x19 = (uint8_t)(x17 & UINT8_C(0xff));
+  uint64_t x20 = (0x0 + x3);
+  uint64_t x21 = (x20 >> 8);
+  uint8_t x22 = (uint8_t)(x20 & UINT8_C(0xff));
+  uint64_t x23 = (x21 >> 8);
+  uint8_t x24 = (uint8_t)(x21 & UINT8_C(0xff));
+  uint64_t x25 = (x23 >> 8);
+  uint8_t x26 = (uint8_t)(x23 & UINT8_C(0xff));
+  uint64_t x27 = (x25 >> 8);
+  uint8_t x28 = (uint8_t)(x25 & UINT8_C(0xff));
+  uint64_t x29 = (x27 >> 8);
+  uint8_t x30 = (uint8_t)(x27 & UINT8_C(0xff));
+  uint64_t x31 = (x29 >> 8);
+  uint8_t x32 = (uint8_t)(x29 & UINT8_C(0xff));
+  uint8_t x33 = (uint8_t)(x31 >> 8);
+  uint8_t x34 = (uint8_t)(x31 & UINT8_C(0xff));
+  uint8_t x35 = (uint8_t)(x33 & UINT8_C(0xff));
+  uint64_t x36 = (0x0 + x2);
+  uint64_t x37 = (x36 >> 8);
+  uint8_t x38 = (uint8_t)(x36 & UINT8_C(0xff));
+  uint64_t x39 = (x37 >> 8);
+  uint8_t x40 = (uint8_t)(x37 & UINT8_C(0xff));
   uint64_t x41 = (x39 >> 8);
   uint8_t x42 = (uint8_t)(x39 & UINT8_C(0xff));
   uint64_t x43 = (x41 >> 8);
@@ -1285,27 +1285,24 @@ static void fiat_secp256k1_to_bytes(uint8_t out1[32], const uint64_t arg1[4]) {
   uint8_t x46 = (uint8_t)(x43 & UINT8_C(0xff));
   uint64_t x47 = (x45 >> 8);
   uint8_t x48 = (uint8_t)(x45 & UINT8_C(0xff));
-  uint64_t x49 = (x47 >> 8);
+  uint8_t x49 = (uint8_t)(x47 >> 8);
   uint8_t x50 = (uint8_t)(x47 & UINT8_C(0xff));
-  uint8_t x51 = (uint8_t)(x49 >> 8);
-  uint8_t x52 = (uint8_t)(x49 & UINT8_C(0xff));
-  fiat_secp256k1_uint1 x53 = (fiat_secp256k1_uint1)(x51 >> 8);
-  uint8_t x54 = (uint8_t)(x51 & UINT8_C(0xff));
-  uint64_t x55 = (x53 + x1);
-  uint64_t x56 = (x55 >> 8);
-  uint8_t x57 = (uint8_t)(x55 & UINT8_C(0xff));
-  uint64_t x58 = (x56 >> 8);
-  uint8_t x59 = (uint8_t)(x56 & UINT8_C(0xff));
-  uint64_t x60 = (x58 >> 8);
-  uint8_t x61 = (uint8_t)(x58 & UINT8_C(0xff));
-  uint64_t x62 = (x60 >> 8);
-  uint8_t x63 = (uint8_t)(x60 & UINT8_C(0xff));
-  uint64_t x64 = (x62 >> 8);
-  uint8_t x65 = (uint8_t)(x62 & UINT8_C(0xff));
-  uint64_t x66 = (x64 >> 8);
-  uint8_t x67 = (uint8_t)(x64 & UINT8_C(0xff));
-  uint8_t x68 = (uint8_t)(x66 >> 8);
-  uint8_t x69 = (uint8_t)(x66 & UINT8_C(0xff));
+  uint8_t x51 = (uint8_t)(x49 & UINT8_C(0xff));
+  uint64_t x52 = (0x0 + x1);
+  uint64_t x53 = (x52 >> 8);
+  uint8_t x54 = (uint8_t)(x52 & UINT8_C(0xff));
+  uint64_t x55 = (x53 >> 8);
+  uint8_t x56 = (uint8_t)(x53 & UINT8_C(0xff));
+  uint64_t x57 = (x55 >> 8);
+  uint8_t x58 = (uint8_t)(x55 & UINT8_C(0xff));
+  uint64_t x59 = (x57 >> 8);
+  uint8_t x60 = (uint8_t)(x57 & UINT8_C(0xff));
+  uint64_t x61 = (x59 >> 8);
+  uint8_t x62 = (uint8_t)(x59 & UINT8_C(0xff));
+  uint64_t x63 = (x61 >> 8);
+  uint8_t x64 = (uint8_t)(x61 & UINT8_C(0xff));
+  uint8_t x65 = (uint8_t)(x63 >> 8);
+  uint8_t x66 = (uint8_t)(x63 & UINT8_C(0xff));
   out1[0] = x6;
   out1[1] = x8;
   out1[2] = x10;
@@ -1313,31 +1310,31 @@ static void fiat_secp256k1_to_bytes(uint8_t out1[32], const uint64_t arg1[4]) {
   out1[4] = x14;
   out1[5] = x16;
   out1[6] = x18;
-  out1[7] = x20;
-  out1[8] = x23;
-  out1[9] = x25;
-  out1[10] = x27;
-  out1[11] = x29;
-  out1[12] = x31;
-  out1[13] = x33;
-  out1[14] = x35;
-  out1[15] = x37;
-  out1[16] = x40;
-  out1[17] = x42;
-  out1[18] = x44;
-  out1[19] = x46;
-  out1[20] = x48;
-  out1[21] = x50;
-  out1[22] = x52;
-  out1[23] = x54;
-  out1[24] = x57;
-  out1[25] = x59;
-  out1[26] = x61;
-  out1[27] = x63;
-  out1[28] = x65;
-  out1[29] = x67;
-  out1[30] = x69;
-  out1[31] = x68;
+  out1[7] = x19;
+  out1[8] = x22;
+  out1[9] = x24;
+  out1[10] = x26;
+  out1[11] = x28;
+  out1[12] = x30;
+  out1[13] = x32;
+  out1[14] = x34;
+  out1[15] = x35;
+  out1[16] = x38;
+  out1[17] = x40;
+  out1[18] = x42;
+  out1[19] = x44;
+  out1[20] = x46;
+  out1[21] = x48;
+  out1[22] = x50;
+  out1[23] = x51;
+  out1[24] = x54;
+  out1[25] = x56;
+  out1[26] = x58;
+  out1[27] = x60;
+  out1[28] = x62;
+  out1[29] = x64;
+  out1[30] = x66;
+  out1[31] = x65;
 }
 
 /*
@@ -1380,21 +1377,18 @@ static void fiat_secp256k1_from_bytes(uint64_t out1[4], const uint8_t arg1[32])
   uint64_t x31 = ((uint64_t)(arg1[1]) << 8);
   uint8_t x32 = (arg1[0]);
   uint64_t x33 = (x32 + (x31 + (x30 + (x29 + (x28 + (x27 + (x26 + x25)))))));
-  fiat_secp256k1_uint1 x34 = (fiat_secp256k1_uint1)((fiat_secp256k1_uint128)x33 >> 64);
-  uint64_t x35 = (x33 & UINT64_C(0xffffffffffffffff));
-  uint64_t x36 = (x8 + (x7 + (x6 + (x5 + (x4 + (x3 + (x2 + x1)))))));
-  uint64_t x37 = (x16 + (x15 + (x14 + (x13 + (x12 + (x11 + (x10 + x9)))))));
-  uint64_t x38 = (x24 + (x23 + (x22 + (x21 + (x20 + (x19 + (x18 + x17)))))));
-  uint64_t x39 = (x34 + x38);
-  fiat_secp256k1_uint1 x40 = (fiat_secp256k1_uint1)((fiat_secp256k1_uint128)x39 >> 64);
-  uint64_t x41 = (x39 & UINT64_C(0xffffffffffffffff));
-  uint64_t x42 = (x40 + x37);
-  fiat_secp256k1_uint1 x43 = (fiat_secp256k1_uint1)((fiat_secp256k1_uint128)x42 >> 64);
-  uint64_t x44 = (x42 & UINT64_C(0xffffffffffffffff));
-  uint64_t x45 = (x43 + x36);
-  out1[0] = x35;
-  out1[1] = x41;
-  out1[2] = x44;
-  out1[3] = x45;
+  uint64_t x34 = (x33 & UINT64_C(0xffffffffffffffff));
+  uint64_t x35 = (x8 + (x7 + (x6 + (x5 + (x4 + (x3 + (x2 + x1)))))));
+  uint64_t x36 = (x16 + (x15 + (x14 + (x13 + (x12 + (x11 + (x10 + x9)))))));
+  uint64_t x37 = (x24 + (x23 + (x22 + (x21 + (x20 + (x19 + (x18 + x17)))))));
+  uint64_t x38 = (0x0 + x37);
+  uint64_t x39 = (x38 & UINT64_C(0xffffffffffffffff));
+  uint64_t x40 = (0x0 + x36);
+  uint64_t x41 = (x40 & UINT64_C(0xffffffffffffffff));
+  uint64_t x42 = (0x0 + x35);
+  out1[0] = x34;
+  out1[1] = x39;
+  out1[2] = x41;
+  out1[3] = x42;
 }
 
diff --git a/src/PushButtonSynthesis.v b/src/PushButtonSynthesis.v
index 0c7f0c994..e20904f63 100644
--- a/src/PushButtonSynthesis.v
+++ b/src/PushButtonSynthesis.v
@@ -212,11 +212,13 @@ Local Opaque reified_%s_gen. (* needed for making [autorewrite] not take a very
     Definition saturated_bounds : ZRange.type.option.interp (base.type.list (base.type.Z))
       := Some saturated_bounds_list.
 
+    (* We include [0], so that even after bounds relaxation, we can
+       notice where the constant 0s are, and remove them. *)
     Definition possible_values_of_machine_wordsize
-      := [machine_wordsize; 2 * machine_wordsize]%Z.
+      := [0; machine_wordsize; 2 * machine_wordsize]%Z.
 
     Definition possible_values_of_machine_wordsize_with_bytes
-      := [1; 8; machine_wordsize; 2 * machine_wordsize]%Z.
+      := [0; 1; 8; machine_wordsize; 2 * machine_wordsize]%Z.
 
     Let possible_values := possible_values_of_machine_wordsize.
     Let possible_values_with_bytes := possible_values_of_machine_wordsize_with_bytes.
@@ -600,11 +602,13 @@ Local Opaque reified_%s_gen. (* needed for making [autorewrite] not take a very
     Definition m_enc : list Z
       := encode (weight (Qnum limbwidth) (Qden limbwidth)) n s c m.
 
+    (* We include [0], so that even after bounds relaxation, we can
+       notice where the constant 0s are, and remove them. *)
     Definition possible_values_of_machine_wordsize
-      := [machine_wordsize; 2 * machine_wordsize]%Z.
+      := [0; machine_wordsize; 2 * machine_wordsize]%Z.
 
     Definition possible_values_of_machine_wordsize_with_bytes
-      := [1; 8; machine_wordsize; 2 * machine_wordsize]%Z.
+      := [0; 1; 8; machine_wordsize; 2 * machine_wordsize]%Z.
 
     Let possible_values := possible_values_of_machine_wordsize.
     Let possible_values_with_bytes := possible_values_of_machine_wordsize_with_bytes.
@@ -1131,8 +1135,10 @@ Module SaturatedSolinas.
             (c : list (Z * Z))
             (machine_wordsize : Z).
 
+    (* We include [0], so that even after bounds relaxation, we can
+       notice where the constant 0s are, and remove them. *)
     Definition possible_values_of_machine_wordsize
-      := [1; machine_wordsize]%Z.
+      := [0; 1; machine_wordsize]%Z.
 
     Let n : nat := Z.to_nat (Qceiling (Z.log2_up s / machine_wordsize)).
     Let m := s - Associational.eval c.
@@ -1518,11 +1524,13 @@ Local Opaque reified_%s_gen. (* needed for making [autorewrite] not take a very
     Local Notation saturated_bounds_list := (saturated_bounds_list n machine_wordsize).
     Local Notation saturated_bounds := (saturated_bounds n machine_wordsize).
 
+    (* We include [0], so that even after bounds relaxation, we can
+       notice where the constant 0s are, and remove them. *)
     Definition possible_values_of_machine_wordsize
-      := [1; machine_wordsize; 2 * machine_wordsize]%Z.
+      := [0; 1; machine_wordsize; 2 * machine_wordsize]%Z.
 
     Definition possible_values_of_machine_wordsize_with_bytes
-      := [1; 8; machine_wordsize; 2 * machine_wordsize]%Z.
+      := [0; 1; 8; machine_wordsize; 2 * machine_wordsize]%Z.
 
     Let possible_values := possible_values_of_machine_wordsize.
     Let possible_values_with_bytes := possible_values_of_machine_wordsize_with_bytes.
@@ -2825,8 +2833,10 @@ Module BarrettReduction.
     Let muLow := mu mod (2 ^ machine_wordsize).
     Let consts_list := [M; muLow].
 
+    (* We include [0], so that even after bounds relaxation, we can
+       notice where the constant 0s are, and remove them. *)
     Definition possible_values_of_machine_wordsize
-      := [1; machine_wordsize / 2; machine_wordsize; 2 * machine_wordsize]%Z.
+      := [0; 1; machine_wordsize / 2; machine_wordsize; 2 * machine_wordsize]%Z.
     Let possible_values := possible_values_of_machine_wordsize.
 
     Definition check_args {T} (res : Pipeline.ErrorT T)
@@ -3026,8 +3036,10 @@ Module MontgomeryReduction.
     Let bound := Some value_range.
     Let consts_list := [N; N'].
 
+    (* We include [0], so that even after bounds relaxation, we can
+       notice where the constant 0s are, and remove them. *)
     Definition possible_values_of_machine_wordsize
-      := [1; machine_wordsize / 2; machine_wordsize; 2 * machine_wordsize]%Z.
+      := [0; 1; machine_wordsize / 2; machine_wordsize; 2 * machine_wordsize]%Z.
     Local Arguments possible_values_of_machine_wordsize / .
 
     Let possible_values := possible_values_of_machine_wordsize.
diff --git a/src/Rewriter.v b/src/Rewriter.v
index 76071b866..8bbf13173 100644
--- a/src/Rewriter.v
+++ b/src/Rewriter.v
@@ -1790,6 +1790,8 @@ Module Compilers.
         := [make_rewrite (#(@pattern.ident.fst '1 '2) @ (??, ??)) (fun _ _ x y => x)
             ; make_rewrite (#(@pattern.ident.snd '1 '2) @ (??, ??)) (fun _ x _ y => y)
 
+            ; make_rewriteo (??') (fun r v => ##(lower r)  when  lower r =? upper r)
+
             ; make_rewriteo
                 (#?ℤ   - (-'??'))
                 (fun z rnv rv v => cst rv v  when  (z =? 0) && (ZRange.normalize rv <=? -ZRange.normalize rnv)%zrange)
diff --git a/src/RewriterRulesInterpGood.v b/src/RewriterRulesInterpGood.v
index 40177e3a0..469f0cb31 100644
--- a/src/RewriterRulesInterpGood.v
+++ b/src/RewriterRulesInterpGood.v
@@ -502,7 +502,11 @@ Module Compilers.
                   => rewrite (@ident.cast_idempotent _ _ r)
                 | [ H : is_bounded_by_bool _ ?r = true |- _]
                   => is_var r; unique pose proof (ZRange.is_bounded_by_normalize _ _ H)
-
+                | [ H : lower ?x = upper ?x |- _ ] => is_var x; destruct x; cbn [upper lower] in *
+                | [ H : is_bounded_by_bool ?x (ZRange.normalize r[?y~>?y]) = true |- _ ]
+                  => apply ZRange.is_bounded_by_bool_normalize_constant_iff in H
+                | [ H : is_bounded_by_bool ?x r[?y~>?y] = true |- _ ]
+                  => apply ZRange.is_bounded_by_bool_constant_iff in H
                 end
               | progress intros
               | progress subst
diff --git a/src/arith_with_casts_rewrite_head.out b/src/arith_with_casts_rewrite_head.out
index 2baf1c5d2..053110d89 100644
--- a/src/arith_with_casts_rewrite_head.out
+++ b/src/arith_with_casts_rewrite_head.out
@@ -888,440 +888,71 @@ match idc in (Compilers.ident t) return (Compile.value' true t) with
 | Z_cc_m => fun x x0 : expr ℤ => Base (#(Z_cc_m)%expr @ x @ x0)%expr_pat
 | Z_cast range =>
     fun x : expr ℤ =>
-    (match x with
-     | @expr.Ident _ _ _ t idc =>
-         args <- invert_bind_args idc Raw.ident.Literal;
-         match
-           pattern.type.unify_extracted_cps ℤ (projT1 args) option
-             (fun x0 : option => x0)
-         with
-         | Some _ =>
-             if type.type_beq base.type base.type.type_beq ℤ (projT1 args)
-             then
-              xv <- ident.unify pattern.ident.Literal ##(projT2 args);
-              fv <- (x0 <- (if
-                             is_bounded_by_bool (let (x0, _) := xv in x0)
-                               range
-                            then Some (##(let (x0, _) := xv in x0))%expr
-                            else None);
-                     Some (Base x0));
-              Some (fv0 <-- fv;
-                    Base fv0)%under_lets
-             else None
-         | None => None
-         end
-     | @expr.App _ _ _ s _ (@expr.Ident _ _ _ t idc) x0 =>
-         args <- invert_bind_args idc Raw.ident.Z_cast;
-         match
-           pattern.type.unify_extracted_cps ℤ s option
-             (fun x1 : option => x1)
-         with
-         | Some _ =>
-             if type.type_beq base.type base.type.type_beq ℤ s
-             then
-              v <- type.try_make_transport_cps s ℤ;
-              fv <- (x1 <- (if
-                             (ZRange.normalize args <=?
-                              ZRange.normalize range)%zrange
-                            then
-                             Some
-                               (#(Z_cast args)%expr @ v (Compile.reflect x0))%expr_pat
-                            else None);
-                     Some (Base x1));
-              Some (fv0 <-- fv;
-                    Base fv0)%under_lets
-             else None
-         | None => None
-         end
-     | @expr.App _ _ _ s _
-       (@expr.App _ _ _ s0 _
-        (@expr.App _ _ _ s1 _ (@expr.Ident _ _ _ t idc) x2) x1) x0 =>
-         _ <- invert_bind_args idc Raw.ident.Z_add_with_carry;
-         match
-           pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
-             ((s1 -> s0) -> s)%ptype option (fun x3 : option => x3)
-         with
-         | Some (_, _, _)%zrange =>
-             if
-              type.type_beq base.type base.type.type_beq
-                ((ℤ -> ℤ) -> ℤ)%ptype ((s1 -> s0) -> s)%ptype
-             then
-              v <- type.try_make_transport_cps s1 ℤ;
-              v0 <- type.try_make_transport_cps s0 ℤ;
-              v1 <- type.try_make_transport_cps s ℤ;
-              Some
-                (UnderLet
-                   (#(Z_cast range)%expr @
-                    (#(Z_add_with_carry)%expr @ v (Compile.reflect x2) @
-                     v0 (Compile.reflect x1) @ v1 (Compile.reflect x0)))%expr_pat
-                   (fun v2 : var ℤ =>
-                    Base (#(Z_cast range)%expr @ ($v2)%expr)%expr_pat))
-             else None
-         | None => None
-         end
-     | @expr.App _ _ _ s _
-       (@expr.App _ _ _ s0 _ (@expr.App _ _ _ s1 _ ($_)%expr _) _) _ |
-       @expr.App _ _ _ s _
-       (@expr.App _ _ _ s0 _ (@expr.App _ _ _ s1 _ (@expr.Abs _ _ _ _ _ _) _)
-        _) _ | @expr.App _ _ _ s _
-       (@expr.App _ _ _ s0 _ (@expr.App _ _ _ s1 _ (_ @ _)%expr_pat _) _) _ |
-       @expr.App _ _ _ s _
-       (@expr.App _ _ _ s0 _
-        (@expr.App _ _ _ s1 _ (@expr.LetIn _ _ _ _ _ _ _) _) _) _ => None
-     | @expr.App _ _ _ s _ (@expr.App _ _ _ s0 _ #(_)%expr_pat _) _ |
-       @expr.App _ _ _ s _ (@expr.App _ _ _ s0 _ ($_)%expr _) _ | @expr.App _
-       _ _ s _ (@expr.App _ _ _ s0 _ (@expr.Abs _ _ _ _ _ _) _) _ | @expr.App
-       _ _ _ s _ (@expr.App _ _ _ s0 _ (@expr.LetIn _ _ _ _ _ _ _) _) _ =>
-         None
-     | @expr.App _ _ _ s _ ($_)%expr _ | @expr.App _ _ _ s _
-       (@expr.Abs _ _ _ _ _ _) _ | @expr.App _ _ _ s _
-       (@expr.LetIn _ _ _ _ _ _ _) _ => None
-     | _ => None
-     end;;;
-     Base (#(Z_cast range)%expr @ x)%expr_pat)%option
-| Z_cast2 range =>
-    fun x : expr (ℤ * ℤ)%etype =>
-    ((match x with
-      | @expr.App _ _ _ s _
-        (@expr.App _ _ _ s0 _ (@expr.Ident _ _ _ t idc) x1) x0 =>
-          args <- invert_bind_args idc Raw.ident.pair;
-          match
-            pattern.type.unify_extracted_cps
-              (((ℤ -> ℤ -> (ℤ * ℤ)%pbtype) -> ℤ) -> ℤ)%ptype
-              ((((let (x2, _) := args in x2) ->
-                 (let (_, y) := args in y) ->
-                 ((let (x2, _) := args in x2) * (let (_, y) := args in y))%etype) ->
-                s0) -> s)%ptype option (fun x2 : option => x2)
-          with
-          | Some (_, (_, (_, _)), _, _)%zrange =>
-              if
-               type.type_beq base.type base.type.type_beq
-                 (((ℤ -> ℤ -> (ℤ * ℤ)%etype) -> ℤ) -> ℤ)%ptype
-                 ((((let (x2, _) := args in x2) ->
-                    (let (_, y) := args in y) ->
-                    ((let (x2, _) := args in x2) * (let (_, y) := args in y))%etype) ->
-                   s0) -> s)%ptype
-              then
-               _ <- ident.unify pattern.ident.pair pair;
-               v <- type.try_make_transport_cps s0 ℤ;
-               v0 <- type.try_make_transport_cps s ℤ;
-               Some
-                 (fv0 <-- do_again (ℤ * ℤ)
-                            (#(Z_cast (Datatypes.fst range))%expr @
-                             ($(v (Compile.reflect x1)))%expr,
-                            #(Z_cast (Datatypes.snd range))%expr @
-                            ($(v0 (Compile.reflect x0)))%expr)%expr_pat;
+    (((match
+         pattern.type.unify_extracted_cps ℤ ℤ option (fun x0 : option => x0)
+       with
+       | Some _ =>
+           if type.type_beq base.type base.type.type_beq ℤ ℤ
+           then
+            fv <- (x0 <- (if lower range =? upper range
+                          then Some (##(lower range))%expr
+                          else None);
+                   Some (Base x0));
+            Some (fv0 <-- fv;
                   Base fv0)%under_lets
-              else None
-          | None => None
-          end
-      | @expr.App _ _ _ s _
-        (@expr.App _ _ _ s0 _
-         (@expr.App _ _ _ s1 _ (@expr.Ident _ _ _ t idc) x2) x1) x0 =>
-          (match x1 with
-           | (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _
-               (@expr.Ident _ _ _ t2 idc2) x5))%expr_pat =>
-               args <- invert_bind_args idc2 Raw.ident.Z_cast;
-               _ <- invert_bind_args idc1 Raw.ident.Z_opp;
-               args1 <- invert_bind_args idc0 Raw.ident.Z_cast;
-               _ <- invert_bind_args idc Raw.ident.Z_add_get_carry;
-               match
-                 pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
-                   ((s1 -> s4) -> s)%ptype option (fun x6 : option => x6)
-               with
-               | Some (_, _, _)%zrange =>
-                   if
-                    type.type_beq base.type base.type.type_beq
-                      ((ℤ -> ℤ) -> ℤ)%ptype ((s1 -> s4) -> s)%ptype
-                   then
-                    v <- type.try_make_transport_cps s1 ℤ;
-                    v0 <- type.try_make_transport_cps s4 ℤ;
-                    v1 <- type.try_make_transport_cps s ℤ;
-                    fv <- (if
-                            (ZRange.normalize args <=?
-                             - ZRange.normalize args1)%zrange
-                           then
-                            Some
-                              (UnderLet
-                                 (#(Z_cast2
-                                      (Datatypes.fst range,
-                                      - Datatypes.snd range))%expr @
-                                  (#(Z_sub_get_borrow)%expr @
-                                   v (Compile.reflect x2) @
-                                   v1 (Compile.reflect x0) @
-                                   (#(Z_cast args)%expr @
-                                    v0 (Compile.reflect x5))))%expr_pat
-                                 (fun vc : var (ℤ * ℤ)%etype =>
-                                  Base
-                                    (#(Z_cast (Datatypes.fst range))%expr @
-                                     (#(fst)%expr @
-                                      (#(Z_cast2
-                                           (Datatypes.fst range,
-                                           - Datatypes.snd range))%expr @
-                                       ($vc)%expr)),
-                                    #(Z_cast (Datatypes.snd range))%expr @
-                                    (-
-                                     (#(Z_cast (- Datatypes.snd range))%expr @
-                                      (#(snd)%expr @
-                                       (#(Z_cast2
-                                            (Datatypes.fst range,
-                                            - Datatypes.snd range))%expr @
-                                        $vc)))%expr_pat)%expr)%expr_pat))
-                           else None);
-                    Some (fv0 <-- fv;
-                          Base fv0)%under_lets
-                   else None
-               | None => None
-               end
-           | (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _ ($_)%expr _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _
-               (@expr.Abs _ _ _ _ _ _) _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _ (_ @ _) _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _
-               (@expr.LetIn _ _ _ _ _ _ _) _))%expr_pat => None
-           | (@expr.Ident _ _ _ t0 idc0 @ (@expr.Ident _ _ _ t1 idc1 @ #(_)))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ ($_)%expr))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.Abs _ _ _ _ _ _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.LetIn _ _ _ _ _ _ _))%expr_pat =>
-               None
-           | (@expr.Ident _ _ _ t0 idc0 @ #(_))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ ($_)%expr)%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ (($_)%expr @ _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ (@expr.Abs _ _ _ _ _ _ @ _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ (_ @ _ @ _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ (@expr.LetIn _ _ _ _ _ _ _ @ _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
-               None
-           | _ => None
-           end;;
-           match x0 with
-           | (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _
-               (@expr.Ident _ _ _ t2 idc2) x5))%expr_pat =>
-               args <- invert_bind_args idc2 Raw.ident.Z_cast;
-               _ <- invert_bind_args idc1 Raw.ident.Z_opp;
-               args1 <- invert_bind_args idc0 Raw.ident.Z_cast;
-               _ <- invert_bind_args idc Raw.ident.Z_add_get_carry;
-               match
-                 pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
-                   ((s1 -> s0) -> s4)%ptype option (fun x6 : option => x6)
-               with
-               | Some (_, _, _)%zrange =>
-                   if
-                    type.type_beq base.type base.type.type_beq
-                      ((ℤ -> ℤ) -> ℤ)%ptype ((s1 -> s0) -> s4)%ptype
-                   then
-                    v <- type.try_make_transport_cps s1 ℤ;
-                    v0 <- type.try_make_transport_cps s0 ℤ;
-                    v1 <- type.try_make_transport_cps s4 ℤ;
-                    fv <- (if
-                            (ZRange.normalize args <=?
-                             - ZRange.normalize args1)%zrange
-                           then
-                            Some
-                              (UnderLet
-                                 (#(Z_cast2
-                                      (Datatypes.fst range,
-                                      - Datatypes.snd range))%expr @
-                                  (#(Z_sub_get_borrow)%expr @
-                                   v (Compile.reflect x2) @
-                                   v0 (Compile.reflect x1) @
-                                   (#(Z_cast args)%expr @
-                                    v1 (Compile.reflect x5))))%expr_pat
-                                 (fun vc : var (ℤ * ℤ)%etype =>
-                                  Base
-                                    (#(Z_cast (Datatypes.fst range))%expr @
-                                     (#(fst)%expr @
-                                      (#(Z_cast2
-                                           (Datatypes.fst range,
-                                           - Datatypes.snd range))%expr @
-                                       ($vc)%expr)),
-                                    #(Z_cast (Datatypes.snd range))%expr @
-                                    (-
-                                     (#(Z_cast (- Datatypes.snd range))%expr @
-                                      (#(snd)%expr @
-                                       (#(Z_cast2
-                                            (Datatypes.fst range,
-                                            - Datatypes.snd range))%expr @
-                                        $vc)))%expr_pat)%expr)%expr_pat))
-                           else None);
-                    Some (fv0 <-- fv;
-                          Base fv0)%under_lets
-                   else None
-               | None => None
-               end
-           | (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _ ($_)%expr _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _
-               (@expr.Abs _ _ _ _ _ _) _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _ (_ @ _) _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _
-               (@expr.LetIn _ _ _ _ _ _ _) _))%expr_pat => None
-           | (@expr.Ident _ _ _ t0 idc0 @ (@expr.Ident _ _ _ t1 idc1 @ #(_)))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ ($_)%expr))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.Abs _ _ _ _ _ _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @
-              (@expr.Ident _ _ _ t1 idc1 @ @expr.LetIn _ _ _ _ _ _ _))%expr_pat =>
-               None
-           | (@expr.Ident _ _ _ t0 idc0 @ #(_))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ ($_)%expr)%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ (($_)%expr @ _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ (@expr.Abs _ _ _ _ _ _ @ _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ (_ @ _ @ _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ (@expr.LetIn _ _ _ _ _ _ _ @ _))%expr_pat |
-             (@expr.Ident _ _ _ t0 idc0 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
-               None
-           | _ => None
-           end;;
-           match x1 with
-           | @expr.Ident _ _ _ t0 idc0 =>
-               args <- invert_bind_args idc0 Raw.ident.Literal;
-               _ <- invert_bind_args idc Raw.ident.Z_add_get_carry;
-               match
-                 pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
-                   ((s1 -> (projT1 args)) -> s)%ptype option
-                   (fun x3 : option => x3)
-               with
-               | Some (_, _, _)%zrange =>
-                   if
-                    type.type_beq base.type base.type.type_beq
-                      ((ℤ -> ℤ) -> ℤ)%ptype
-                      ((s1 -> (projT1 args)) -> s)%ptype
-                   then
-                    v <- type.try_make_transport_cps s1 ℤ;
-                    xv <- ident.unify pattern.ident.Literal ##(projT2 args);
-                    v0 <- type.try_make_transport_cps s ℤ;
-                    fv <- (if (let (x3, _) := xv in x3) <? 0
-                           then
-                            Some
-                              (UnderLet
-                                 (#(Z_cast2
-                                      (Datatypes.fst range,
-                                      - Datatypes.snd range))%expr @
-                                  (#(Z_sub_get_borrow)%expr @
-                                   v (Compile.reflect x2) @
-                                   v0 (Compile.reflect x0) @
-                                   (##(- (let (x3, _) := xv in x3))%Z)%expr))%expr_pat
-                                 (fun vc : var (ℤ * ℤ)%etype =>
-                                  Base
-                                    (#(Z_cast (Datatypes.fst range))%expr @
-                                     (#(fst)%expr @
-                                      (#(Z_cast2
-                                           (Datatypes.fst range,
-                                           - Datatypes.snd range))%expr @
-                                       ($vc)%expr)),
-                                    #(Z_cast (Datatypes.snd range))%expr @
-                                    (-
-                                     (#(Z_cast (- Datatypes.snd range))%expr @
-                                      (#(snd)%expr @
-                                       (#(Z_cast2
-                                            (Datatypes.fst range,
-                                            - Datatypes.snd range))%expr @
-                                        $vc)))%expr_pat)%expr)%expr_pat))
-                           else None);
-                    Some (fv0 <-- fv;
-                          Base fv0)%under_lets
-                   else None
-               | None => None
-               end
-           | _ => None
-           end;;
-           match x0 with
-           | @expr.Ident _ _ _ t0 idc0 =>
-               args <- invert_bind_args idc0 Raw.ident.Literal;
-               _ <- invert_bind_args idc Raw.ident.Z_add_get_carry;
-               match
-                 pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
-                   ((s1 -> s0) -> (projT1 args))%ptype option
-                   (fun x3 : option => x3)
-               with
-               | Some (_, _, _)%zrange =>
-                   if
-                    type.type_beq base.type base.type.type_beq
-                      ((ℤ -> ℤ) -> ℤ)%ptype
-                      ((s1 -> s0) -> (projT1 args))%ptype
-                   then
-                    v <- type.try_make_transport_cps s1 ℤ;
-                    v0 <- type.try_make_transport_cps s0 ℤ;
-                    xv <- ident.unify pattern.ident.Literal ##(projT2 args);
-                    fv <- (if (let (x3, _) := xv in x3) <? 0
-                           then
-                            Some
-                              (UnderLet
-                                 (#(Z_cast2
-                                      (Datatypes.fst range,
-                                      - Datatypes.snd range))%expr @
-                                  (#(Z_sub_get_borrow)%expr @
-                                   v (Compile.reflect x2) @
-                                   v0 (Compile.reflect x1) @
-                                   (##(- (let (x3, _) := xv in x3))%Z)%expr))%expr_pat
-                                 (fun vc : var (ℤ * ℤ)%etype =>
-                                  Base
-                                    (#(Z_cast (Datatypes.fst range))%expr @
-                                     (#(fst)%expr @
-                                      (#(Z_cast2
-                                           (Datatypes.fst range,
-                                           - Datatypes.snd range))%expr @
-                                       ($vc)%expr)),
-                                    #(Z_cast (Datatypes.snd range))%expr @
-                                    (-
-                                     (#(Z_cast (- Datatypes.snd range))%expr @
-                                      (#(snd)%expr @
-                                       (#(Z_cast2
-                                            (Datatypes.fst range,
-                                            - Datatypes.snd range))%expr @
-                                        $vc)))%expr_pat)%expr)%expr_pat))
-                           else None);
-                    Some (fv0 <-- fv;
-                          Base fv0)%under_lets
-                   else None
-               | None => None
-               end
-           | _ => None
-           end;;
-           _ <- invert_bind_args idc Raw.ident.Z_add_get_carry;
+           else None
+       | None => None
+       end;;
+       match x with
+       | @expr.Ident _ _ _ t idc =>
+           args <- invert_bind_args idc Raw.ident.Literal;
            match
-             pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
-               ((s1 -> s0) -> s)%ptype option (fun x3 : option => x3)
+             pattern.type.unify_extracted_cps ℤ (projT1 args) option
+               (fun x0 : option => x0)
            with
-           | Some (_, _, _)%zrange =>
-               if
-                type.type_beq base.type base.type.type_beq
-                  ((ℤ -> ℤ) -> ℤ)%ptype ((s1 -> s0) -> s)%ptype
+           | Some _ =>
+               if type.type_beq base.type base.type.type_beq ℤ (projT1 args)
                then
-                v <- type.try_make_transport_cps s1 ℤ;
-                v0 <- type.try_make_transport_cps s0 ℤ;
-                v1 <- type.try_make_transport_cps s ℤ;
-                Some
-                  (UnderLet
-                     (#(Z_cast2 range)%expr @
-                      (#(Z_add_get_carry)%expr @ v (Compile.reflect x2) @
-                       v0 (Compile.reflect x1) @ v1 (Compile.reflect x0)))%expr_pat
-                     (fun v2 : var (ℤ * ℤ)%etype =>
-                      Base
-                        (#(Z_cast (Datatypes.fst range))%expr @
-                         (#(fst)%expr @ (#(Z_cast2 range)%expr @ ($v2)%expr)),
-                        #(Z_cast (Datatypes.snd range))%expr @
-                        (#(snd)%expr @ (#(Z_cast2 range)%expr @ ($v2)%expr)))%expr_pat))
+                xv <- ident.unify pattern.ident.Literal ##(projT2 args);
+                fv <- (x0 <- (if
+                               is_bounded_by_bool (let (x0, _) := xv in x0)
+                                 range
+                              then Some (##(let (x0, _) := xv in x0))%expr
+                              else None);
+                       Some (Base x0));
+                Some (fv0 <-- fv;
+                      Base fv0)%under_lets
+               else None
+           | None => None
+           end
+       | @expr.App _ _ _ s _ (@expr.Ident _ _ _ t idc) x0 =>
+           args <- invert_bind_args idc Raw.ident.Z_cast;
+           match
+             pattern.type.unify_extracted_cps ℤ s option
+               (fun x1 : option => x1)
+           with
+           | Some _ =>
+               if type.type_beq base.type base.type.type_beq ℤ s
+               then
+                v <- type.try_make_transport_cps s ℤ;
+                fv <- (x1 <- (if
+                               (ZRange.normalize args <=?
+                                ZRange.normalize range)%zrange
+                              then
+                               Some
+                                 (#(Z_cast args)%expr @
+                                  v (Compile.reflect x0))%expr_pat
+                              else None);
+                       Some (Base x1));
+                Some (fv0 <-- fv;
+                      Base fv0)%under_lets
                else None
            | None => None
-           end);;
-          (_ <- invert_bind_args idc Raw.ident.Z_sub_get_borrow;
+           end
+       | @expr.App _ _ _ s _
+         (@expr.App _ _ _ s0 _
+          (@expr.App _ _ _ s1 _ (@expr.Ident _ _ _ t idc) x2) x1) x0 =>
+           _ <- invert_bind_args idc Raw.ident.Z_add_with_carry;
            match
              pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
                ((s1 -> s0) -> s)%ptype option (fun x3 : option => x3)
@@ -1336,19 +967,347 @@ match idc in (Compilers.ident t) return (Compile.value' true t) with
                 v1 <- type.try_make_transport_cps s ℤ;
                 Some
                   (UnderLet
-                     (#(Z_cast2 range)%expr @
-                      (#(Z_sub_get_borrow)%expr @ v (Compile.reflect x2) @
+                     (#(Z_cast range)%expr @
+                      (#(Z_add_with_carry)%expr @ v (Compile.reflect x2) @
                        v0 (Compile.reflect x1) @ v1 (Compile.reflect x0)))%expr_pat
-                     (fun v2 : var (ℤ * ℤ)%etype =>
-                      Base
-                        (#(Z_cast (Datatypes.fst range))%expr @
-                         (#(fst)%expr @ (#(Z_cast2 range)%expr @ ($v2)%expr)),
-                        #(Z_cast (Datatypes.snd range))%expr @
-                        (#(snd)%expr @ (#(Z_cast2 range)%expr @ ($v2)%expr)))%expr_pat))
+                     (fun v2 : var ℤ =>
+                      Base (#(Z_cast range)%expr @ ($v2)%expr)%expr_pat))
                else None
            | None => None
-           end);;
-          _ <- invert_bind_args idc Raw.ident.Z_mul_split;
+           end
+       | @expr.App _ _ _ s _
+         (@expr.App _ _ _ s0 _ (@expr.App _ _ _ s1 _ ($_)%expr _) _) _ |
+         @expr.App _ _ _ s _
+         (@expr.App _ _ _ s0 _
+          (@expr.App _ _ _ s1 _ (@expr.Abs _ _ _ _ _ _) _) _) _ | @expr.App _
+         _ _ s _
+         (@expr.App _ _ _ s0 _ (@expr.App _ _ _ s1 _ (_ @ _)%expr_pat _) _)
+         _ | @expr.App _ _ _ s _
+         (@expr.App _ _ _ s0 _
+          (@expr.App _ _ _ s1 _ (@expr.LetIn _ _ _ _ _ _ _) _) _) _ => None
+       | @expr.App _ _ _ s _ (@expr.App _ _ _ s0 _ #(_)%expr_pat _) _ |
+         @expr.App _ _ _ s _ (@expr.App _ _ _ s0 _ ($_)%expr _) _ | @expr.App
+         _ _ _ s _ (@expr.App _ _ _ s0 _ (@expr.Abs _ _ _ _ _ _) _) _ |
+         @expr.App _ _ _ s _
+         (@expr.App _ _ _ s0 _ (@expr.LetIn _ _ _ _ _ _ _) _) _ => None
+       | @expr.App _ _ _ s _ ($_)%expr _ | @expr.App _ _ _ s _
+         (@expr.Abs _ _ _ _ _ _) _ | @expr.App _ _ _ s _
+         (@expr.LetIn _ _ _ _ _ _ _) _ => None
+       | _ => None
+       end);;
+      None);;;
+     Base (#(Z_cast range)%expr @ x)%expr_pat)%option
+| Z_cast2 range =>
+    fun x : expr (ℤ * ℤ)%etype =>
+    (match x with
+     | @expr.App _ _ _ s _
+       (@expr.App _ _ _ s0 _ (@expr.Ident _ _ _ t idc) x1) x0 =>
+         args <- invert_bind_args idc Raw.ident.pair;
+         match
+           pattern.type.unify_extracted_cps
+             (((ℤ -> ℤ -> (ℤ * ℤ)%pbtype) -> ℤ) -> ℤ)%ptype
+             ((((let (x2, _) := args in x2) ->
+                (let (_, y) := args in y) ->
+                ((let (x2, _) := args in x2) * (let (_, y) := args in y))%etype) ->
+               s0) -> s)%ptype option (fun x2 : option => x2)
+         with
+         | Some (_, (_, (_, _)), _, _)%zrange =>
+             if
+              type.type_beq base.type base.type.type_beq
+                (((ℤ -> ℤ -> (ℤ * ℤ)%etype) -> ℤ) -> ℤ)%ptype
+                ((((let (x2, _) := args in x2) ->
+                   (let (_, y) := args in y) ->
+                   ((let (x2, _) := args in x2) * (let (_, y) := args in y))%etype) ->
+                  s0) -> s)%ptype
+             then
+              _ <- ident.unify pattern.ident.pair pair;
+              v <- type.try_make_transport_cps s0 ℤ;
+              v0 <- type.try_make_transport_cps s ℤ;
+              Some
+                (fv0 <-- do_again (ℤ * ℤ)
+                           (#(Z_cast (Datatypes.fst range))%expr @
+                            ($(v (Compile.reflect x1)))%expr,
+                           #(Z_cast (Datatypes.snd range))%expr @
+                           ($(v0 (Compile.reflect x0)))%expr)%expr_pat;
+                 Base fv0)%under_lets
+             else None
+         | None => None
+         end
+     | @expr.App _ _ _ s _
+       (@expr.App _ _ _ s0 _
+        (@expr.App _ _ _ s1 _ (@expr.Ident _ _ _ t idc) x2) x1) x0 =>
+         (match x1 with
+          | (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _
+              (@expr.Ident _ _ _ t2 idc2) x5))%expr_pat =>
+              args <- invert_bind_args idc2 Raw.ident.Z_cast;
+              _ <- invert_bind_args idc1 Raw.ident.Z_opp;
+              args1 <- invert_bind_args idc0 Raw.ident.Z_cast;
+              _ <- invert_bind_args idc Raw.ident.Z_add_get_carry;
+              match
+                pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
+                  ((s1 -> s4) -> s)%ptype option (fun x6 : option => x6)
+              with
+              | Some (_, _, _)%zrange =>
+                  if
+                   type.type_beq base.type base.type.type_beq
+                     ((ℤ -> ℤ) -> ℤ)%ptype ((s1 -> s4) -> s)%ptype
+                  then
+                   v <- type.try_make_transport_cps s1 ℤ;
+                   v0 <- type.try_make_transport_cps s4 ℤ;
+                   v1 <- type.try_make_transport_cps s ℤ;
+                   fv <- (if
+                           (ZRange.normalize args <=?
+                            - ZRange.normalize args1)%zrange
+                          then
+                           Some
+                             (UnderLet
+                                (#(Z_cast2
+                                     (Datatypes.fst range,
+                                     - Datatypes.snd range))%expr @
+                                 (#(Z_sub_get_borrow)%expr @
+                                  v (Compile.reflect x2) @
+                                  v1 (Compile.reflect x0) @
+                                  (#(Z_cast args)%expr @
+                                   v0 (Compile.reflect x5))))%expr_pat
+                                (fun vc : var (ℤ * ℤ)%etype =>
+                                 Base
+                                   (#(Z_cast (Datatypes.fst range))%expr @
+                                    (#(fst)%expr @
+                                     (#(Z_cast2
+                                          (Datatypes.fst range,
+                                          - Datatypes.snd range))%expr @
+                                      ($vc)%expr)),
+                                   #(Z_cast (Datatypes.snd range))%expr @
+                                   (-
+                                    (#(Z_cast (- Datatypes.snd range))%expr @
+                                     (#(snd)%expr @
+                                      (#(Z_cast2
+                                           (Datatypes.fst range,
+                                           - Datatypes.snd range))%expr @ $vc)))%expr_pat)%expr)%expr_pat))
+                          else None);
+                   Some (fv0 <-- fv;
+                         Base fv0)%under_lets
+                  else None
+              | None => None
+              end
+          | (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _ ($_)%expr _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _
+              (@expr.Abs _ _ _ _ _ _) _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _ (_ @ _) _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _
+              (@expr.LetIn _ _ _ _ _ _ _) _))%expr_pat => None
+          | (@expr.Ident _ _ _ t0 idc0 @ (@expr.Ident _ _ _ t1 idc1 @ #(_)))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ ($_)%expr))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.Abs _ _ _ _ _ _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.LetIn _ _ _ _ _ _ _))%expr_pat =>
+              None
+          | (@expr.Ident _ _ _ t0 idc0 @ #(_))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ ($_)%expr)%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ (($_)%expr @ _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ (@expr.Abs _ _ _ _ _ _ @ _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ (_ @ _ @ _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ (@expr.LetIn _ _ _ _ _ _ _ @ _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
+              None
+          | _ => None
+          end;;
+          match x0 with
+          | (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _
+              (@expr.Ident _ _ _ t2 idc2) x5))%expr_pat =>
+              args <- invert_bind_args idc2 Raw.ident.Z_cast;
+              _ <- invert_bind_args idc1 Raw.ident.Z_opp;
+              args1 <- invert_bind_args idc0 Raw.ident.Z_cast;
+              _ <- invert_bind_args idc Raw.ident.Z_add_get_carry;
+              match
+                pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
+                  ((s1 -> s0) -> s4)%ptype option (fun x6 : option => x6)
+              with
+              | Some (_, _, _)%zrange =>
+                  if
+                   type.type_beq base.type base.type.type_beq
+                     ((ℤ -> ℤ) -> ℤ)%ptype ((s1 -> s0) -> s4)%ptype
+                  then
+                   v <- type.try_make_transport_cps s1 ℤ;
+                   v0 <- type.try_make_transport_cps s0 ℤ;
+                   v1 <- type.try_make_transport_cps s4 ℤ;
+                   fv <- (if
+                           (ZRange.normalize args <=?
+                            - ZRange.normalize args1)%zrange
+                          then
+                           Some
+                             (UnderLet
+                                (#(Z_cast2
+                                     (Datatypes.fst range,
+                                     - Datatypes.snd range))%expr @
+                                 (#(Z_sub_get_borrow)%expr @
+                                  v (Compile.reflect x2) @
+                                  v0 (Compile.reflect x1) @
+                                  (#(Z_cast args)%expr @
+                                   v1 (Compile.reflect x5))))%expr_pat
+                                (fun vc : var (ℤ * ℤ)%etype =>
+                                 Base
+                                   (#(Z_cast (Datatypes.fst range))%expr @
+                                    (#(fst)%expr @
+                                     (#(Z_cast2
+                                          (Datatypes.fst range,
+                                          - Datatypes.snd range))%expr @
+                                      ($vc)%expr)),
+                                   #(Z_cast (Datatypes.snd range))%expr @
+                                   (-
+                                    (#(Z_cast (- Datatypes.snd range))%expr @
+                                     (#(snd)%expr @
+                                      (#(Z_cast2
+                                           (Datatypes.fst range,
+                                           - Datatypes.snd range))%expr @ $vc)))%expr_pat)%expr)%expr_pat))
+                          else None);
+                   Some (fv0 <-- fv;
+                         Base fv0)%under_lets
+                  else None
+              | None => None
+              end
+          | (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _ ($_)%expr _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _
+              (@expr.Abs _ _ _ _ _ _) _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _ (_ @ _) _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s4 _
+              (@expr.LetIn _ _ _ _ _ _ _) _))%expr_pat => None
+          | (@expr.Ident _ _ _ t0 idc0 @ (@expr.Ident _ _ _ t1 idc1 @ #(_)))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ ($_)%expr))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.Abs _ _ _ _ _ _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @
+             (@expr.Ident _ _ _ t1 idc1 @ @expr.LetIn _ _ _ _ _ _ _))%expr_pat =>
+              None
+          | (@expr.Ident _ _ _ t0 idc0 @ #(_))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ ($_)%expr)%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ (($_)%expr @ _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ (@expr.Abs _ _ _ _ _ _ @ _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ (_ @ _ @ _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ (@expr.LetIn _ _ _ _ _ _ _ @ _))%expr_pat |
+            (@expr.Ident _ _ _ t0 idc0 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
+              None
+          | _ => None
+          end;;
+          match x1 with
+          | @expr.Ident _ _ _ t0 idc0 =>
+              args <- invert_bind_args idc0 Raw.ident.Literal;
+              _ <- invert_bind_args idc Raw.ident.Z_add_get_carry;
+              match
+                pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
+                  ((s1 -> (projT1 args)) -> s)%ptype option
+                  (fun x3 : option => x3)
+              with
+              | Some (_, _, _)%zrange =>
+                  if
+                   type.type_beq base.type base.type.type_beq
+                     ((ℤ -> ℤ) -> ℤ)%ptype ((s1 -> (projT1 args)) -> s)%ptype
+                  then
+                   v <- type.try_make_transport_cps s1 ℤ;
+                   xv <- ident.unify pattern.ident.Literal ##(projT2 args);
+                   v0 <- type.try_make_transport_cps s ℤ;
+                   fv <- (if (let (x3, _) := xv in x3) <? 0
+                          then
+                           Some
+                             (UnderLet
+                                (#(Z_cast2
+                                     (Datatypes.fst range,
+                                     - Datatypes.snd range))%expr @
+                                 (#(Z_sub_get_borrow)%expr @
+                                  v (Compile.reflect x2) @
+                                  v0 (Compile.reflect x0) @
+                                  (##(- (let (x3, _) := xv in x3))%Z)%expr))%expr_pat
+                                (fun vc : var (ℤ * ℤ)%etype =>
+                                 Base
+                                   (#(Z_cast (Datatypes.fst range))%expr @
+                                    (#(fst)%expr @
+                                     (#(Z_cast2
+                                          (Datatypes.fst range,
+                                          - Datatypes.snd range))%expr @
+                                      ($vc)%expr)),
+                                   #(Z_cast (Datatypes.snd range))%expr @
+                                   (-
+                                    (#(Z_cast (- Datatypes.snd range))%expr @
+                                     (#(snd)%expr @
+                                      (#(Z_cast2
+                                           (Datatypes.fst range,
+                                           - Datatypes.snd range))%expr @ $vc)))%expr_pat)%expr)%expr_pat))
+                          else None);
+                   Some (fv0 <-- fv;
+                         Base fv0)%under_lets
+                  else None
+              | None => None
+              end
+          | _ => None
+          end;;
+          match x0 with
+          | @expr.Ident _ _ _ t0 idc0 =>
+              args <- invert_bind_args idc0 Raw.ident.Literal;
+              _ <- invert_bind_args idc Raw.ident.Z_add_get_carry;
+              match
+                pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
+                  ((s1 -> s0) -> (projT1 args))%ptype option
+                  (fun x3 : option => x3)
+              with
+              | Some (_, _, _)%zrange =>
+                  if
+                   type.type_beq base.type base.type.type_beq
+                     ((ℤ -> ℤ) -> ℤ)%ptype
+                     ((s1 -> s0) -> (projT1 args))%ptype
+                  then
+                   v <- type.try_make_transport_cps s1 ℤ;
+                   v0 <- type.try_make_transport_cps s0 ℤ;
+                   xv <- ident.unify pattern.ident.Literal ##(projT2 args);
+                   fv <- (if (let (x3, _) := xv in x3) <? 0
+                          then
+                           Some
+                             (UnderLet
+                                (#(Z_cast2
+                                     (Datatypes.fst range,
+                                     - Datatypes.snd range))%expr @
+                                 (#(Z_sub_get_borrow)%expr @
+                                  v (Compile.reflect x2) @
+                                  v0 (Compile.reflect x1) @
+                                  (##(- (let (x3, _) := xv in x3))%Z)%expr))%expr_pat
+                                (fun vc : var (ℤ * ℤ)%etype =>
+                                 Base
+                                   (#(Z_cast (Datatypes.fst range))%expr @
+                                    (#(fst)%expr @
+                                     (#(Z_cast2
+                                          (Datatypes.fst range,
+                                          - Datatypes.snd range))%expr @
+                                      ($vc)%expr)),
+                                   #(Z_cast (Datatypes.snd range))%expr @
+                                   (-
+                                    (#(Z_cast (- Datatypes.snd range))%expr @
+                                     (#(snd)%expr @
+                                      (#(Z_cast2
+                                           (Datatypes.fst range,
+                                           - Datatypes.snd range))%expr @ $vc)))%expr_pat)%expr)%expr_pat))
+                          else None);
+                   Some (fv0 <-- fv;
+                         Base fv0)%under_lets
+                  else None
+              | None => None
+              end
+          | _ => None
+          end;;
+          _ <- invert_bind_args idc Raw.ident.Z_add_get_carry;
           match
             pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
               ((s1 -> s0) -> s)%ptype option (fun x3 : option => x3)
@@ -1364,7 +1323,7 @@ match idc in (Compilers.ident t) return (Compile.value' true t) with
                Some
                  (UnderLet
                     (#(Z_cast2 range)%expr @
-                     (#(Z_mul_split)%expr @ v (Compile.reflect x2) @
+                     (#(Z_add_get_carry)%expr @ v (Compile.reflect x2) @
                       v0 (Compile.reflect x1) @ v1 (Compile.reflect x0)))%expr_pat
                     (fun v2 : var (ℤ * ℤ)%etype =>
                      Base
@@ -1374,77 +1333,72 @@ match idc in (Compilers.ident t) return (Compile.value' true t) with
                        (#(snd)%expr @ (#(Z_cast2 range)%expr @ ($v2)%expr)))%expr_pat))
               else None
           | None => None
-          end
-      | @expr.App _ _ _ s _
-        (@expr.App _ _ _ s0 _
-         (@expr.App _ _ _ s1 _
-          (@expr.App _ _ _ s2 _ (@expr.Ident _ _ _ t idc) x3) x2) x1) x0 =>
-          (match x2 with
-           | @expr.Ident _ _ _ t0 idc0 =>
-               match x1 with
-               | (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _
-                   (@expr.Ident _ _ _ t3 idc3) x6))%expr_pat =>
-                   (args <- invert_bind_args idc3 Raw.ident.Z_cast;
-                    _ <- invert_bind_args idc2 Raw.ident.Z_opp;
-                    args1 <- invert_bind_args idc1 Raw.ident.Z_cast;
-                    args2 <- invert_bind_args idc0 Raw.ident.Literal;
-                    _ <- invert_bind_args idc Raw.ident.Z_add_with_get_carry;
-                    match
-                      pattern.type.unify_extracted_cps
-                        (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                        (((s2 -> (projT1 args2)) -> s5) -> s)%ptype option
-                        (fun x7 : option => x7)
-                    with
-                    | Some (_, _, _, _)%zrange =>
-                        if
-                         type.type_beq base.type base.type.type_beq
-                           (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                           (((s2 -> (projT1 args2)) -> s5) -> s)%ptype
-                        then
-                         v <- type.try_make_transport_cps s2 ℤ;
-                         xv <- ident.unify pattern.ident.Literal
-                                 ##(projT2 args2);
-                         v0 <- type.try_make_transport_cps s5 ℤ;
-                         v1 <- type.try_make_transport_cps s ℤ;
-                         fv <- (if
-                                 ((let (x7, _) := xv in x7) =? 0) &&
-                                 (ZRange.normalize args <=?
-                                  - ZRange.normalize args1)%zrange
-                                then
-                                 Some
-                                   (UnderLet
-                                      (#(Z_cast2
-                                           (Datatypes.fst range,
-                                           - Datatypes.snd range))%expr @
-                                       (#(Z_sub_get_borrow)%expr @
-                                        v (Compile.reflect x3) @
-                                        v1 (Compile.reflect x0) @
-                                        (#(Z_cast args)%expr @
-                                         v0 (Compile.reflect x6))))%expr_pat
-                                      (fun vc : var (ℤ * ℤ)%etype =>
-                                       Base
-                                         (#(Z_cast (Datatypes.fst range))%expr @
-                                          (#(fst)%expr @
-                                           (#(Z_cast2
-                                                (Datatypes.fst range,
-                                                - Datatypes.snd range))%expr @
-                                            ($vc)%expr)),
-                                         #(Z_cast (Datatypes.snd range))%expr @
-                                         (-
-                                          (#(Z_cast (- Datatypes.snd range))%expr @
-                                           (#(snd)%expr @
-                                            (#(Z_cast2
-                                                 (Datatypes.fst range,
-                                                 - Datatypes.snd range))%expr @
-                                             $vc)))%expr_pat)%expr)%expr_pat))
-                                else None);
-                         Some (fv0 <-- fv;
-                               Base fv0)%under_lets
-                        else None
-                    | None => None
-                    end);;
-                   args <- invert_bind_args idc3 Raw.ident.Z_cast;
+          end);;
+         (_ <- invert_bind_args idc Raw.ident.Z_sub_get_borrow;
+          match
+            pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
+              ((s1 -> s0) -> s)%ptype option (fun x3 : option => x3)
+          with
+          | Some (_, _, _)%zrange =>
+              if
+               type.type_beq base.type base.type.type_beq
+                 ((ℤ -> ℤ) -> ℤ)%ptype ((s1 -> s0) -> s)%ptype
+              then
+               v <- type.try_make_transport_cps s1 ℤ;
+               v0 <- type.try_make_transport_cps s0 ℤ;
+               v1 <- type.try_make_transport_cps s ℤ;
+               Some
+                 (UnderLet
+                    (#(Z_cast2 range)%expr @
+                     (#(Z_sub_get_borrow)%expr @ v (Compile.reflect x2) @
+                      v0 (Compile.reflect x1) @ v1 (Compile.reflect x0)))%expr_pat
+                    (fun v2 : var (ℤ * ℤ)%etype =>
+                     Base
+                       (#(Z_cast (Datatypes.fst range))%expr @
+                        (#(fst)%expr @ (#(Z_cast2 range)%expr @ ($v2)%expr)),
+                       #(Z_cast (Datatypes.snd range))%expr @
+                       (#(snd)%expr @ (#(Z_cast2 range)%expr @ ($v2)%expr)))%expr_pat))
+              else None
+          | None => None
+          end);;
+         _ <- invert_bind_args idc Raw.ident.Z_mul_split;
+         match
+           pattern.type.unify_extracted_cps ((ℤ -> ℤ) -> ℤ)%ptype
+             ((s1 -> s0) -> s)%ptype option (fun x3 : option => x3)
+         with
+         | Some (_, _, _)%zrange =>
+             if
+              type.type_beq base.type base.type.type_beq
+                ((ℤ -> ℤ) -> ℤ)%ptype ((s1 -> s0) -> s)%ptype
+             then
+              v <- type.try_make_transport_cps s1 ℤ;
+              v0 <- type.try_make_transport_cps s0 ℤ;
+              v1 <- type.try_make_transport_cps s ℤ;
+              Some
+                (UnderLet
+                   (#(Z_cast2 range)%expr @
+                    (#(Z_mul_split)%expr @ v (Compile.reflect x2) @
+                     v0 (Compile.reflect x1) @ v1 (Compile.reflect x0)))%expr_pat
+                   (fun v2 : var (ℤ * ℤ)%etype =>
+                    Base
+                      (#(Z_cast (Datatypes.fst range))%expr @
+                       (#(fst)%expr @ (#(Z_cast2 range)%expr @ ($v2)%expr)),
+                      #(Z_cast (Datatypes.snd range))%expr @
+                      (#(snd)%expr @ (#(Z_cast2 range)%expr @ ($v2)%expr)))%expr_pat))
+             else None
+         | None => None
+         end
+     | @expr.App _ _ _ s _
+       (@expr.App _ _ _ s0 _
+        (@expr.App _ _ _ s1 _
+         (@expr.App _ _ _ s2 _ (@expr.Ident _ _ _ t idc) x3) x2) x1) x0 =>
+         (match x2 with
+          | @expr.Ident _ _ _ t0 idc0 =>
+              match x1 with
+              | (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _
+                  (@expr.Ident _ _ _ t3 idc3) x6))%expr_pat =>
+                  (args <- invert_bind_args idc3 Raw.ident.Z_cast;
                    _ <- invert_bind_args idc2 Raw.ident.Z_opp;
                    args1 <- invert_bind_args idc1 Raw.ident.Z_cast;
                    args2 <- invert_bind_args idc0 Raw.ident.Literal;
@@ -1467,7 +1421,7 @@ match idc in (Compilers.ident t) return (Compile.value' true t) with
                         v0 <- type.try_make_transport_cps s5 ℤ;
                         v1 <- type.try_make_transport_cps s ℤ;
                         fv <- (if
-                                ((let (x7, _) := xv in x7) <? 0) &&
+                                ((let (x7, _) := xv in x7) =? 0) &&
                                 (ZRange.normalize args <=?
                                  - ZRange.normalize args1)%zrange
                                then
@@ -1476,9 +1430,8 @@ match idc in (Compilers.ident t) return (Compile.value' true t) with
                                      (#(Z_cast2
                                           (Datatypes.fst range,
                                           - Datatypes.snd range))%expr @
-                                      (#(Z_sub_with_get_borrow)%expr @
+                                      (#(Z_sub_get_borrow)%expr @
                                        v (Compile.reflect x3) @
-                                       (##(- (let (x7, _) := xv in x7))%Z)%expr @
                                        v1 (Compile.reflect x0) @
                                        (#(Z_cast args)%expr @
                                         v0 (Compile.reflect x6))))%expr_pat
@@ -1503,102 +1456,103 @@ match idc in (Compilers.ident t) return (Compile.value' true t) with
                               Base fv0)%under_lets
                        else None
                    | None => None
-                   end
-               | (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _ ($_)%expr
-                   _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _
-                   (@expr.Abs _ _ _ _ _ _) _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _ (_ @ _) _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _
-                   (@expr.LetIn _ _ _ _ _ _ _) _))%expr_pat => None
-               | (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ #(_)))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ ($_)%expr))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.Abs _ _ _ _ _ _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.LetIn _ _ _ _ _ _ _))%expr_pat =>
-                   None
-               | (@expr.Ident _ _ _ t1 idc1 @ #(_))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ ($_)%expr)%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ (($_)%expr @ _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ (@expr.Abs _ _ _ _ _ _ @ _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ (_ @ _ @ _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ (@expr.LetIn _ _ _ _ _ _ _ @ _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
-                   None
-               | _ => None
-               end;;
-               match x0 with
-               | (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _
-                   (@expr.Ident _ _ _ t3 idc3) x6))%expr_pat =>
-                   (args <- invert_bind_args idc3 Raw.ident.Z_cast;
-                    _ <- invert_bind_args idc2 Raw.ident.Z_opp;
-                    args1 <- invert_bind_args idc1 Raw.ident.Z_cast;
-                    args2 <- invert_bind_args idc0 Raw.ident.Literal;
-                    _ <- invert_bind_args idc Raw.ident.Z_add_with_get_carry;
-                    match
-                      pattern.type.unify_extracted_cps
-                        (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                        (((s2 -> (projT1 args2)) -> s0) -> s5)%ptype option
-                        (fun x7 : option => x7)
-                    with
-                    | Some (_, _, _, _)%zrange =>
-                        if
-                         type.type_beq base.type base.type.type_beq
-                           (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                           (((s2 -> (projT1 args2)) -> s0) -> s5)%ptype
-                        then
-                         v <- type.try_make_transport_cps s2 ℤ;
-                         xv <- ident.unify pattern.ident.Literal
-                                 ##(projT2 args2);
-                         v0 <- type.try_make_transport_cps s0 ℤ;
-                         v1 <- type.try_make_transport_cps s5 ℤ;
-                         fv <- (if
-                                 ((let (x7, _) := xv in x7) =? 0) &&
-                                 (ZRange.normalize args <=?
-                                  - ZRange.normalize args1)%zrange
-                                then
-                                 Some
-                                   (UnderLet
-                                      (#(Z_cast2
-                                           (Datatypes.fst range,
-                                           - Datatypes.snd range))%expr @
-                                       (#(Z_sub_get_borrow)%expr @
-                                        v (Compile.reflect x3) @
-                                        v0 (Compile.reflect x1) @
-                                        (#(Z_cast args)%expr @
-                                         v1 (Compile.reflect x6))))%expr_pat
-                                      (fun vc : var (ℤ * ℤ)%etype =>
-                                       Base
-                                         (#(Z_cast (Datatypes.fst range))%expr @
-                                          (#(fst)%expr @
-                                           (#(Z_cast2
-                                                (Datatypes.fst range,
-                                                - Datatypes.snd range))%expr @
-                                            ($vc)%expr)),
-                                         #(Z_cast (Datatypes.snd range))%expr @
-                                         (-
-                                          (#(Z_cast (- Datatypes.snd range))%expr @
-                                           (#(snd)%expr @
-                                            (#(Z_cast2
-                                                 (Datatypes.fst range,
-                                                 - Datatypes.snd range))%expr @
-                                             $vc)))%expr_pat)%expr)%expr_pat))
-                                else None);
-                         Some (fv0 <-- fv;
-                               Base fv0)%under_lets
-                        else None
-                    | None => None
-                    end);;
-                   args <- invert_bind_args idc3 Raw.ident.Z_cast;
+                   end);;
+                  args <- invert_bind_args idc3 Raw.ident.Z_cast;
+                  _ <- invert_bind_args idc2 Raw.ident.Z_opp;
+                  args1 <- invert_bind_args idc1 Raw.ident.Z_cast;
+                  args2 <- invert_bind_args idc0 Raw.ident.Literal;
+                  _ <- invert_bind_args idc Raw.ident.Z_add_with_get_carry;
+                  match
+                    pattern.type.unify_extracted_cps
+                      (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                      (((s2 -> (projT1 args2)) -> s5) -> s)%ptype option
+                      (fun x7 : option => x7)
+                  with
+                  | Some (_, _, _, _)%zrange =>
+                      if
+                       type.type_beq base.type base.type.type_beq
+                         (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                         (((s2 -> (projT1 args2)) -> s5) -> s)%ptype
+                      then
+                       v <- type.try_make_transport_cps s2 ℤ;
+                       xv <- ident.unify pattern.ident.Literal
+                               ##(projT2 args2);
+                       v0 <- type.try_make_transport_cps s5 ℤ;
+                       v1 <- type.try_make_transport_cps s ℤ;
+                       fv <- (if
+                               ((let (x7, _) := xv in x7) <? 0) &&
+                               (ZRange.normalize args <=?
+                                - ZRange.normalize args1)%zrange
+                              then
+                               Some
+                                 (UnderLet
+                                    (#(Z_cast2
+                                         (Datatypes.fst range,
+                                         - Datatypes.snd range))%expr @
+                                     (#(Z_sub_with_get_borrow)%expr @
+                                      v (Compile.reflect x3) @
+                                      (##(- (let (x7, _) := xv in x7))%Z)%expr @
+                                      v1 (Compile.reflect x0) @
+                                      (#(Z_cast args)%expr @
+                                       v0 (Compile.reflect x6))))%expr_pat
+                                    (fun vc : var (ℤ * ℤ)%etype =>
+                                     Base
+                                       (#(Z_cast (Datatypes.fst range))%expr @
+                                        (#(fst)%expr @
+                                         (#(Z_cast2
+                                              (Datatypes.fst range,
+                                              - Datatypes.snd range))%expr @
+                                          ($vc)%expr)),
+                                       #(Z_cast (Datatypes.snd range))%expr @
+                                       (-
+                                        (#(Z_cast (- Datatypes.snd range))%expr @
+                                         (#(snd)%expr @
+                                          (#(Z_cast2
+                                               (Datatypes.fst range,
+                                               - Datatypes.snd range))%expr @
+                                           $vc)))%expr_pat)%expr)%expr_pat))
+                              else None);
+                       Some (fv0 <-- fv;
+                             Base fv0)%under_lets
+                      else None
+                  | None => None
+                  end
+              | (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _ ($_)%expr
+                  _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _
+                  (@expr.Abs _ _ _ _ _ _) _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _ (_ @ _) _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _
+                  (@expr.LetIn _ _ _ _ _ _ _) _))%expr_pat => None
+              | (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ #(_)))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ ($_)%expr))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.Abs _ _ _ _ _ _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.LetIn _ _ _ _ _ _ _))%expr_pat =>
+                  None
+              | (@expr.Ident _ _ _ t1 idc1 @ #(_))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ ($_)%expr)%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ (($_)%expr @ _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ (@expr.Abs _ _ _ _ _ _ @ _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ (_ @ _ @ _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ (@expr.LetIn _ _ _ _ _ _ _ @ _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
+                  None
+              | _ => None
+              end;;
+              match x0 with
+              | (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _
+                  (@expr.Ident _ _ _ t3 idc3) x6))%expr_pat =>
+                  (args <- invert_bind_args idc3 Raw.ident.Z_cast;
                    _ <- invert_bind_args idc2 Raw.ident.Z_opp;
                    args1 <- invert_bind_args idc1 Raw.ident.Z_cast;
                    args2 <- invert_bind_args idc0 Raw.ident.Literal;
@@ -1621,7 +1575,7 @@ match idc in (Compilers.ident t) return (Compile.value' true t) with
                         v0 <- type.try_make_transport_cps s0 ℤ;
                         v1 <- type.try_make_transport_cps s5 ℤ;
                         fv <- (if
-                                ((let (x7, _) := xv in x7) <? 0) &&
+                                ((let (x7, _) := xv in x7) =? 0) &&
                                 (ZRange.normalize args <=?
                                  - ZRange.normalize args1)%zrange
                                then
@@ -1630,9 +1584,8 @@ match idc in (Compilers.ident t) return (Compile.value' true t) with
                                      (#(Z_cast2
                                           (Datatypes.fst range,
                                           - Datatypes.snd range))%expr @
-                                      (#(Z_sub_with_get_borrow)%expr @
+                                      (#(Z_sub_get_borrow)%expr @
                                        v (Compile.reflect x3) @
-                                       (##(- (let (x7, _) := xv in x7))%Z)%expr @
                                        v0 (Compile.reflect x1) @
                                        (#(Z_cast args)%expr @
                                         v1 (Compile.reflect x6))))%expr_pat
@@ -1657,665 +1610,688 @@ match idc in (Compilers.ident t) return (Compile.value' true t) with
                               Base fv0)%under_lets
                        else None
                    | None => None
-                   end
-               | (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _ ($_)%expr
-                   _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _
-                   (@expr.Abs _ _ _ _ _ _) _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _ (_ @ _) _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _
-                   (@expr.LetIn _ _ _ _ _ _ _) _))%expr_pat => None
-               | (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ #(_)))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ ($_)%expr))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.Abs _ _ _ _ _ _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @
-                  (@expr.Ident _ _ _ t2 idc2 @ @expr.LetIn _ _ _ _ _ _ _))%expr_pat =>
-                   None
-               | (@expr.Ident _ _ _ t1 idc1 @ #(_))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ ($_)%expr)%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ (($_)%expr @ _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ (@expr.Abs _ _ _ _ _ _ @ _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ (_ @ _ @ _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ (@expr.LetIn _ _ _ _ _ _ _ @ _))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
-                   None
-               | _ => None
-               end;;
-               match x1 with
-               | @expr.Ident _ _ _ t1 idc1 =>
-                   args <- invert_bind_args idc1 Raw.ident.Literal;
-                   args0 <- invert_bind_args idc0 Raw.ident.Literal;
-                   _ <- invert_bind_args idc Raw.ident.Z_add_with_get_carry;
-                   match
-                     pattern.type.unify_extracted_cps
-                       (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                       (((s2 -> (projT1 args0)) -> (projT1 args)) -> s)%ptype
-                       option (fun x4 : option => x4)
-                   with
-                   | Some (_, _, _, _)%zrange =>
-                       if
-                        type.type_beq base.type base.type.type_beq
-                          (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                          (((s2 -> (projT1 args0)) -> (projT1 args)) -> s)%ptype
-                       then
-                        v <- type.try_make_transport_cps s2 ℤ;
-                        xv <- ident.unify pattern.ident.Literal
-                                ##(projT2 args0);
-                        xv0 <- ident.unify pattern.ident.Literal
-                                 ##(projT2 args);
-                        v0 <- type.try_make_transport_cps s ℤ;
-                        fv <- (if
-                                ((let (x4, _) := xv0 in x4) <=? 0) &&
-                                ((let (x4, _) := xv in x4) <=? 0) &&
-                                ((let (x4, _) := xv0 in x4) +
-                                 (let (x4, _) := xv in x4) <? 0)
-                               then
-                                Some
-                                  (UnderLet
-                                     (#(Z_cast2
-                                          (Datatypes.fst range,
-                                          - Datatypes.snd range))%expr @
-                                      (#(Z_sub_with_get_borrow)%expr @
-                                       v (Compile.reflect x3) @
-                                       (##(- (let (x4, _) := xv in x4))%Z)%expr @
-                                       v0 (Compile.reflect x0) @
-                                       (##(- (let (x4, _) := xv0 in x4))%Z)%expr))%expr_pat
-                                     (fun vc : var (ℤ * ℤ)%etype =>
-                                      Base
-                                        (#(Z_cast (Datatypes.fst range))%expr @
-                                         (#(fst)%expr @
+                   end);;
+                  args <- invert_bind_args idc3 Raw.ident.Z_cast;
+                  _ <- invert_bind_args idc2 Raw.ident.Z_opp;
+                  args1 <- invert_bind_args idc1 Raw.ident.Z_cast;
+                  args2 <- invert_bind_args idc0 Raw.ident.Literal;
+                  _ <- invert_bind_args idc Raw.ident.Z_add_with_get_carry;
+                  match
+                    pattern.type.unify_extracted_cps
+                      (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                      (((s2 -> (projT1 args2)) -> s0) -> s5)%ptype option
+                      (fun x7 : option => x7)
+                  with
+                  | Some (_, _, _, _)%zrange =>
+                      if
+                       type.type_beq base.type base.type.type_beq
+                         (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                         (((s2 -> (projT1 args2)) -> s0) -> s5)%ptype
+                      then
+                       v <- type.try_make_transport_cps s2 ℤ;
+                       xv <- ident.unify pattern.ident.Literal
+                               ##(projT2 args2);
+                       v0 <- type.try_make_transport_cps s0 ℤ;
+                       v1 <- type.try_make_transport_cps s5 ℤ;
+                       fv <- (if
+                               ((let (x7, _) := xv in x7) <? 0) &&
+                               (ZRange.normalize args <=?
+                                - ZRange.normalize args1)%zrange
+                              then
+                               Some
+                                 (UnderLet
+                                    (#(Z_cast2
+                                         (Datatypes.fst range,
+                                         - Datatypes.snd range))%expr @
+                                     (#(Z_sub_with_get_borrow)%expr @
+                                      v (Compile.reflect x3) @
+                                      (##(- (let (x7, _) := xv in x7))%Z)%expr @
+                                      v0 (Compile.reflect x1) @
+                                      (#(Z_cast args)%expr @
+                                       v1 (Compile.reflect x6))))%expr_pat
+                                    (fun vc : var (ℤ * ℤ)%etype =>
+                                     Base
+                                       (#(Z_cast (Datatypes.fst range))%expr @
+                                        (#(fst)%expr @
+                                         (#(Z_cast2
+                                              (Datatypes.fst range,
+                                              - Datatypes.snd range))%expr @
+                                          ($vc)%expr)),
+                                       #(Z_cast (Datatypes.snd range))%expr @
+                                       (-
+                                        (#(Z_cast (- Datatypes.snd range))%expr @
+                                         (#(snd)%expr @
                                           (#(Z_cast2
                                                (Datatypes.fst range,
                                                - Datatypes.snd range))%expr @
-                                           ($vc)%expr)),
-                                        #(Z_cast (Datatypes.snd range))%expr @
-                                        (-
-                                         (#(Z_cast (- Datatypes.snd range))%expr @
-                                          (#(snd)%expr @
-                                           (#(Z_cast2
-                                                (Datatypes.fst range,
-                                                - Datatypes.snd range))%expr @
-                                            $vc)))%expr_pat)%expr)%expr_pat))
-                               else None);
-                        Some (fv0 <-- fv;
-                              Base fv0)%under_lets
-                       else None
-                   | None => None
-                   end
-               | _ => None
-               end;;
-               match x0 with
-               | @expr.Ident _ _ _ t1 idc1 =>
-                   args <- invert_bind_args idc1 Raw.ident.Literal;
-                   args0 <- invert_bind_args idc0 Raw.ident.Literal;
-                   _ <- invert_bind_args idc Raw.ident.Z_add_with_get_carry;
-                   match
-                     pattern.type.unify_extracted_cps
-                       (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                       (((s2 -> (projT1 args0)) -> s0) -> (projT1 args))%ptype
-                       option (fun x4 : option => x4)
-                   with
-                   | Some (_, _, _, _)%zrange =>
-                       if
-                        type.type_beq base.type base.type.type_beq
-                          (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                          (((s2 -> (projT1 args0)) -> s0) -> (projT1 args))%ptype
-                       then
-                        v <- type.try_make_transport_cps s2 ℤ;
-                        xv <- ident.unify pattern.ident.Literal
-                                ##(projT2 args0);
-                        v0 <- type.try_make_transport_cps s0 ℤ;
-                        xv0 <- ident.unify pattern.ident.Literal
-                                 ##(projT2 args);
-                        fv <- (if
-                                ((let (x4, _) := xv0 in x4) <=? 0) &&
-                                ((let (x4, _) := xv in x4) <=? 0) &&
-                                ((let (x4, _) := xv0 in x4) +
-                                 (let (x4, _) := xv in x4) <? 0)
-                               then
-                                Some
-                                  (UnderLet
-                                     (#(Z_cast2
-                                          (Datatypes.fst range,
-                                          - Datatypes.snd range))%expr @
-                                      (#(Z_sub_with_get_borrow)%expr @
-                                       v (Compile.reflect x3) @
-                                       (##(- (let (x4, _) := xv in x4))%Z)%expr @
-                                       v0 (Compile.reflect x1) @
-                                       (##(- (let (x4, _) := xv0 in x4))%Z)%expr))%expr_pat
-                                     (fun vc : var (ℤ * ℤ)%etype =>
-                                      Base
-                                        (#(Z_cast (Datatypes.fst range))%expr @
-                                         (#(fst)%expr @
+                                           $vc)))%expr_pat)%expr)%expr_pat))
+                              else None);
+                       Some (fv0 <-- fv;
+                             Base fv0)%under_lets
+                      else None
+                  | None => None
+                  end
+              | (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _ ($_)%expr
+                  _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _
+                  (@expr.Abs _ _ _ _ _ _) _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _ (_ @ _) _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.App _ _ _ s5 _
+                  (@expr.LetIn _ _ _ _ _ _ _) _))%expr_pat => None
+              | (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ #(_)))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ ($_)%expr))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.Abs _ _ _ _ _ _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @
+                 (@expr.Ident _ _ _ t2 idc2 @ @expr.LetIn _ _ _ _ _ _ _))%expr_pat =>
+                  None
+              | (@expr.Ident _ _ _ t1 idc1 @ #(_))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ ($_)%expr)%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ (($_)%expr @ _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ (@expr.Abs _ _ _ _ _ _ @ _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ (_ @ _ @ _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ (@expr.LetIn _ _ _ _ _ _ _ @ _))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
+                  None
+              | _ => None
+              end;;
+              match x1 with
+              | @expr.Ident _ _ _ t1 idc1 =>
+                  args <- invert_bind_args idc1 Raw.ident.Literal;
+                  args0 <- invert_bind_args idc0 Raw.ident.Literal;
+                  _ <- invert_bind_args idc Raw.ident.Z_add_with_get_carry;
+                  match
+                    pattern.type.unify_extracted_cps
+                      (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                      (((s2 -> (projT1 args0)) -> (projT1 args)) -> s)%ptype
+                      option (fun x4 : option => x4)
+                  with
+                  | Some (_, _, _, _)%zrange =>
+                      if
+                       type.type_beq base.type base.type.type_beq
+                         (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                         (((s2 -> (projT1 args0)) -> (projT1 args)) -> s)%ptype
+                      then
+                       v <- type.try_make_transport_cps s2 ℤ;
+                       xv <- ident.unify pattern.ident.Literal
+                               ##(projT2 args0);
+                       xv0 <- ident.unify pattern.ident.Literal
+                                ##(projT2 args);
+                       v0 <- type.try_make_transport_cps s ℤ;
+                       fv <- (if
+                               ((let (x4, _) := xv0 in x4) <=? 0) &&
+                               ((let (x4, _) := xv in x4) <=? 0) &&
+                               ((let (x4, _) := xv0 in x4) +
+                                (let (x4, _) := xv in x4) <? 0)
+                              then
+                               Some
+                                 (UnderLet
+                                    (#(Z_cast2
+                                         (Datatypes.fst range,
+                                         - Datatypes.snd range))%expr @
+                                     (#(Z_sub_with_get_borrow)%expr @
+                                      v (Compile.reflect x3) @
+                                      (##(- (let (x4, _) := xv in x4))%Z)%expr @
+                                      v0 (Compile.reflect x0) @
+                                      (##(- (let (x4, _) := xv0 in x4))%Z)%expr))%expr_pat
+                                    (fun vc : var (ℤ * ℤ)%etype =>
+                                     Base
+                                       (#(Z_cast (Datatypes.fst range))%expr @
+                                        (#(fst)%expr @
+                                         (#(Z_cast2
+                                              (Datatypes.fst range,
+                                              - Datatypes.snd range))%expr @
+                                          ($vc)%expr)),
+                                       #(Z_cast (Datatypes.snd range))%expr @
+                                       (-
+                                        (#(Z_cast (- Datatypes.snd range))%expr @
+                                         (#(snd)%expr @
                                           (#(Z_cast2
                                                (Datatypes.fst range,
                                                - Datatypes.snd range))%expr @
-                                           ($vc)%expr)),
-                                        #(Z_cast (Datatypes.snd range))%expr @
-                                        (-
-                                         (#(Z_cast (- Datatypes.snd range))%expr @
-                                          (#(snd)%expr @
-                                           (#(Z_cast2
-                                                (Datatypes.fst range,
-                                                - Datatypes.snd range))%expr @
-                                            $vc)))%expr_pat)%expr)%expr_pat))
-                               else None);
-                        Some (fv0 <-- fv;
-                              Base fv0)%under_lets
-                       else None
-                   | None => None
-                   end
-               | _ => None
-               end;;
-               args <- invert_bind_args idc0 Raw.ident.Literal;
-               _ <- invert_bind_args idc Raw.ident.Z_add_with_get_carry;
-               match
-                 pattern.type.unify_extracted_cps
-                   (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                   (((s2 -> (projT1 args)) -> s0) -> s)%ptype option
-                   (fun x4 : option => x4)
-               with
-               | Some (_, _, _, _)%zrange =>
-                   if
-                    type.type_beq base.type base.type.type_beq
+                                           $vc)))%expr_pat)%expr)%expr_pat))
+                              else None);
+                       Some (fv0 <-- fv;
+                             Base fv0)%under_lets
+                      else None
+                  | None => None
+                  end
+              | _ => None
+              end;;
+              match x0 with
+              | @expr.Ident _ _ _ t1 idc1 =>
+                  args <- invert_bind_args idc1 Raw.ident.Literal;
+                  args0 <- invert_bind_args idc0 Raw.ident.Literal;
+                  _ <- invert_bind_args idc Raw.ident.Z_add_with_get_carry;
+                  match
+                    pattern.type.unify_extracted_cps
                       (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                      (((s2 -> (projT1 args)) -> s0) -> s)%ptype
-                   then
-                    v <- type.try_make_transport_cps s2 ℤ;
-                    xv <- ident.unify pattern.ident.Literal ##(projT2 args);
-                    v0 <- type.try_make_transport_cps s0 ℤ;
-                    v1 <- type.try_make_transport_cps s ℤ;
-                    fv <- (if (let (x4, _) := xv in x4) =? 0
-                           then
-                            Some
-                              (UnderLet
-                                 (#(Z_cast2 range)%expr @
-                                  (#(Z_add_get_carry)%expr @
-                                   v (Compile.reflect x3) @
-                                   v0 (Compile.reflect x1) @
-                                   v1 (Compile.reflect x0)))%expr_pat
-                                 (fun vc : var (ℤ * ℤ)%etype =>
-                                  Base
-                                    (#(Z_cast (Datatypes.fst range))%expr @
-                                     (#(fst)%expr @
-                                      (#(Z_cast2 range)%expr @ ($vc)%expr)),
-                                    #(Z_cast (Datatypes.snd range))%expr @
-                                    (#(snd)%expr @
-                                     (#(Z_cast2 range)%expr @ ($vc)%expr)))%expr_pat))
-                           else None);
-                    Some (fv0 <-- fv;
-                          Base fv0)%under_lets
-                   else None
-               | None => None
-               end
-           | @expr.App _ _ _ s3 _ (@expr.Ident _ _ _ t0 idc0) x4 =>
-               match x4 with
-               | (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s5 _
-                  (@expr.Ident _ _ _ t2 idc2) x6)%expr_pat =>
-                   match x1 with
-                   | (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
-                       (@expr.Ident _ _ _ t5 idc5) x9))%expr_pat =>
-                       args <- invert_bind_args idc5 Raw.ident.Z_cast;
-                       _ <- invert_bind_args idc4 Raw.ident.Z_opp;
-                       args1 <- invert_bind_args idc3 Raw.ident.Z_cast;
-                       args2 <- invert_bind_args idc2 Raw.ident.Z_cast;
-                       _ <- invert_bind_args idc1 Raw.ident.Z_opp;
-                       args4 <- invert_bind_args idc0 Raw.ident.Z_cast;
-                       _ <- invert_bind_args idc
-                              Raw.ident.Z_add_with_get_carry;
-                       match
-                         pattern.type.unify_extracted_cps
-                           (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                           (((s2 -> s5) -> s8) -> s)%ptype option
-                           (fun x10 : option => x10)
-                       with
-                       | Some (_, _, _, _)%zrange =>
-                           if
-                            type.type_beq base.type base.type.type_beq
-                              (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                              (((s2 -> s5) -> s8) -> s)%ptype
-                           then
-                            v <- type.try_make_transport_cps s2 ℤ;
-                            v0 <- type.try_make_transport_cps s5 ℤ;
-                            v1 <- type.try_make_transport_cps s8 ℤ;
-                            v2 <- type.try_make_transport_cps s ℤ;
-                            fv <- (if
-                                    (ZRange.normalize args <=?
-                                     - ZRange.normalize args1)%zrange &&
-                                    (ZRange.normalize args2 <=?
-                                     - ZRange.normalize args4)%zrange
-                                   then
-                                    Some
-                                      (UnderLet
+                      (((s2 -> (projT1 args0)) -> s0) -> (projT1 args))%ptype
+                      option (fun x4 : option => x4)
+                  with
+                  | Some (_, _, _, _)%zrange =>
+                      if
+                       type.type_beq base.type base.type.type_beq
+                         (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                         (((s2 -> (projT1 args0)) -> s0) -> (projT1 args))%ptype
+                      then
+                       v <- type.try_make_transport_cps s2 ℤ;
+                       xv <- ident.unify pattern.ident.Literal
+                               ##(projT2 args0);
+                       v0 <- type.try_make_transport_cps s0 ℤ;
+                       xv0 <- ident.unify pattern.ident.Literal
+                                ##(projT2 args);
+                       fv <- (if
+                               ((let (x4, _) := xv0 in x4) <=? 0) &&
+                               ((let (x4, _) := xv in x4) <=? 0) &&
+                               ((let (x4, _) := xv0 in x4) +
+                                (let (x4, _) := xv in x4) <? 0)
+                              then
+                               Some
+                                 (UnderLet
+                                    (#(Z_cast2
+                                         (Datatypes.fst range,
+                                         - Datatypes.snd range))%expr @
+                                     (#(Z_sub_with_get_borrow)%expr @
+                                      v (Compile.reflect x3) @
+                                      (##(- (let (x4, _) := xv in x4))%Z)%expr @
+                                      v0 (Compile.reflect x1) @
+                                      (##(- (let (x4, _) := xv0 in x4))%Z)%expr))%expr_pat
+                                    (fun vc : var (ℤ * ℤ)%etype =>
+                                     Base
+                                       (#(Z_cast (Datatypes.fst range))%expr @
+                                        (#(fst)%expr @
                                          (#(Z_cast2
                                               (Datatypes.fst range,
                                               - Datatypes.snd range))%expr @
-                                          (#(Z_sub_with_get_borrow)%expr @
-                                           v (Compile.reflect x3) @
-                                           (#(Z_cast args2)%expr @
-                                            v0 (Compile.reflect x6)) @
-                                           v2 (Compile.reflect x0) @
-                                           (#(Z_cast args)%expr @
-                                            v1 (Compile.reflect x9))))%expr_pat
-                                         (fun vc : var (ℤ * ℤ)%etype =>
-                                          Base
-                                            (#(Z_cast (Datatypes.fst range))%expr @
-                                             (#(fst)%expr @
+                                          ($vc)%expr)),
+                                       #(Z_cast (Datatypes.snd range))%expr @
+                                       (-
+                                        (#(Z_cast (- Datatypes.snd range))%expr @
+                                         (#(snd)%expr @
+                                          (#(Z_cast2
+                                               (Datatypes.fst range,
+                                               - Datatypes.snd range))%expr @
+                                           $vc)))%expr_pat)%expr)%expr_pat))
+                              else None);
+                       Some (fv0 <-- fv;
+                             Base fv0)%under_lets
+                      else None
+                  | None => None
+                  end
+              | _ => None
+              end;;
+              args <- invert_bind_args idc0 Raw.ident.Literal;
+              _ <- invert_bind_args idc Raw.ident.Z_add_with_get_carry;
+              match
+                pattern.type.unify_extracted_cps (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                  (((s2 -> (projT1 args)) -> s0) -> s)%ptype option
+                  (fun x4 : option => x4)
+              with
+              | Some (_, _, _, _)%zrange =>
+                  if
+                   type.type_beq base.type base.type.type_beq
+                     (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                     (((s2 -> (projT1 args)) -> s0) -> s)%ptype
+                  then
+                   v <- type.try_make_transport_cps s2 ℤ;
+                   xv <- ident.unify pattern.ident.Literal ##(projT2 args);
+                   v0 <- type.try_make_transport_cps s0 ℤ;
+                   v1 <- type.try_make_transport_cps s ℤ;
+                   fv <- (if (let (x4, _) := xv in x4) =? 0
+                          then
+                           Some
+                             (UnderLet
+                                (#(Z_cast2 range)%expr @
+                                 (#(Z_add_get_carry)%expr @
+                                  v (Compile.reflect x3) @
+                                  v0 (Compile.reflect x1) @
+                                  v1 (Compile.reflect x0)))%expr_pat
+                                (fun vc : var (ℤ * ℤ)%etype =>
+                                 Base
+                                   (#(Z_cast (Datatypes.fst range))%expr @
+                                    (#(fst)%expr @
+                                     (#(Z_cast2 range)%expr @ ($vc)%expr)),
+                                   #(Z_cast (Datatypes.snd range))%expr @
+                                   (#(snd)%expr @
+                                    (#(Z_cast2 range)%expr @ ($vc)%expr)))%expr_pat))
+                          else None);
+                   Some (fv0 <-- fv;
+                         Base fv0)%under_lets
+                  else None
+              | None => None
+              end
+          | @expr.App _ _ _ s3 _ (@expr.Ident _ _ _ t0 idc0) x4 =>
+              match x4 with
+              | (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s5 _
+                 (@expr.Ident _ _ _ t2 idc2) x6)%expr_pat =>
+                  match x1 with
+                  | (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
+                      (@expr.Ident _ _ _ t5 idc5) x9))%expr_pat =>
+                      args <- invert_bind_args idc5 Raw.ident.Z_cast;
+                      _ <- invert_bind_args idc4 Raw.ident.Z_opp;
+                      args1 <- invert_bind_args idc3 Raw.ident.Z_cast;
+                      args2 <- invert_bind_args idc2 Raw.ident.Z_cast;
+                      _ <- invert_bind_args idc1 Raw.ident.Z_opp;
+                      args4 <- invert_bind_args idc0 Raw.ident.Z_cast;
+                      _ <- invert_bind_args idc
+                             Raw.ident.Z_add_with_get_carry;
+                      match
+                        pattern.type.unify_extracted_cps
+                          (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                          (((s2 -> s5) -> s8) -> s)%ptype option
+                          (fun x10 : option => x10)
+                      with
+                      | Some (_, _, _, _)%zrange =>
+                          if
+                           type.type_beq base.type base.type.type_beq
+                             (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                             (((s2 -> s5) -> s8) -> s)%ptype
+                          then
+                           v <- type.try_make_transport_cps s2 ℤ;
+                           v0 <- type.try_make_transport_cps s5 ℤ;
+                           v1 <- type.try_make_transport_cps s8 ℤ;
+                           v2 <- type.try_make_transport_cps s ℤ;
+                           fv <- (if
+                                   (ZRange.normalize args <=?
+                                    - ZRange.normalize args1)%zrange &&
+                                   (ZRange.normalize args2 <=?
+                                    - ZRange.normalize args4)%zrange
+                                  then
+                                   Some
+                                     (UnderLet
+                                        (#(Z_cast2
+                                             (Datatypes.fst range,
+                                             - Datatypes.snd range))%expr @
+                                         (#(Z_sub_with_get_borrow)%expr @
+                                          v (Compile.reflect x3) @
+                                          (#(Z_cast args2)%expr @
+                                           v0 (Compile.reflect x6)) @
+                                          v2 (Compile.reflect x0) @
+                                          (#(Z_cast args)%expr @
+                                           v1 (Compile.reflect x9))))%expr_pat
+                                        (fun vc : var (ℤ * ℤ)%etype =>
+                                         Base
+                                           (#(Z_cast (Datatypes.fst range))%expr @
+                                            (#(fst)%expr @
+                                             (#(Z_cast2
+                                                  (Datatypes.fst range,
+                                                  - Datatypes.snd range))%expr @
+                                              ($vc)%expr)),
+                                           #(Z_cast (Datatypes.snd range))%expr @
+                                           (-
+                                            (#(Z_cast (- Datatypes.snd range))%expr @
+                                             (#(snd)%expr @
                                               (#(Z_cast2
                                                    (Datatypes.fst range,
                                                    - Datatypes.snd range))%expr @
-                                               ($vc)%expr)),
-                                            #(Z_cast (Datatypes.snd range))%expr @
-                                            (-
-                                             (#(Z_cast
-                                                  (- Datatypes.snd range))%expr @
-                                              (#(snd)%expr @
-                                               (#(Z_cast2
-                                                    (Datatypes.fst range,
-                                                    - Datatypes.snd range))%expr @
-                                                $vc)))%expr_pat)%expr)%expr_pat))
-                                   else None);
-                            Some (fv0 <-- fv;
-                                  Base fv0)%under_lets
-                           else None
-                       | None => None
-                       end
-                   | (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
-                       ($_)%expr _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
-                       (@expr.Abs _ _ _ _ _ _) _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
-                       (_ @ _) _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
-                       (@expr.LetIn _ _ _ _ _ _ _) _))%expr_pat => None
-                   | (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ #(_)))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ ($_)%expr))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.Abs _ _ _ _ _ _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.LetIn _ _ _ _ _ _ _))%expr_pat =>
-                       None
-                   | (@expr.Ident _ _ _ t3 idc3 @ #(_))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @ ($_)%expr)%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @ (($_)%expr @ _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @ (@expr.Abs _ _ _ _ _ _ @ _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @ (_ @ _ @ _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.LetIn _ _ _ _ _ _ _ @ _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
-                       None
-                   | _ => None
-                   end;;
-                   match x0 with
-                   | (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
-                       (@expr.Ident _ _ _ t5 idc5) x9))%expr_pat =>
-                       args <- invert_bind_args idc5 Raw.ident.Z_cast;
-                       _ <- invert_bind_args idc4 Raw.ident.Z_opp;
-                       args1 <- invert_bind_args idc3 Raw.ident.Z_cast;
-                       args2 <- invert_bind_args idc2 Raw.ident.Z_cast;
-                       _ <- invert_bind_args idc1 Raw.ident.Z_opp;
-                       args4 <- invert_bind_args idc0 Raw.ident.Z_cast;
-                       _ <- invert_bind_args idc
-                              Raw.ident.Z_add_with_get_carry;
-                       match
-                         pattern.type.unify_extracted_cps
-                           (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                           (((s2 -> s5) -> s0) -> s8)%ptype option
-                           (fun x10 : option => x10)
-                       with
-                       | Some (_, _, _, _)%zrange =>
-                           if
-                            type.type_beq base.type base.type.type_beq
-                              (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                              (((s2 -> s5) -> s0) -> s8)%ptype
-                           then
-                            v <- type.try_make_transport_cps s2 ℤ;
-                            v0 <- type.try_make_transport_cps s5 ℤ;
-                            v1 <- type.try_make_transport_cps s0 ℤ;
-                            v2 <- type.try_make_transport_cps s8 ℤ;
-                            fv <- (if
-                                    (ZRange.normalize args <=?
-                                     - ZRange.normalize args1)%zrange &&
-                                    (ZRange.normalize args2 <=?
-                                     - ZRange.normalize args4)%zrange
-                                   then
-                                    Some
-                                      (UnderLet
-                                         (#(Z_cast2
-                                              (Datatypes.fst range,
-                                              - Datatypes.snd range))%expr @
-                                          (#(Z_sub_with_get_borrow)%expr @
-                                           v (Compile.reflect x3) @
-                                           (#(Z_cast args2)%expr @
-                                            v0 (Compile.reflect x6)) @
-                                           v1 (Compile.reflect x1) @
-                                           (#(Z_cast args)%expr @
-                                            v2 (Compile.reflect x9))))%expr_pat
-                                         (fun vc : var (ℤ * ℤ)%etype =>
-                                          Base
-                                            (#(Z_cast (Datatypes.fst range))%expr @
-                                             (#(fst)%expr @
+                                               $vc)))%expr_pat)%expr)%expr_pat))
+                                  else None);
+                           Some (fv0 <-- fv;
+                                 Base fv0)%under_lets
+                          else None
+                      | None => None
+                      end
+                  | (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
+                      ($_)%expr _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
+                      (@expr.Abs _ _ _ _ _ _) _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
+                      (_ @ _) _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
+                      (@expr.LetIn _ _ _ _ _ _ _) _))%expr_pat => None
+                  | (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ #(_)))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ ($_)%expr))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.Abs _ _ _ _ _ _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.LetIn _ _ _ _ _ _ _))%expr_pat =>
+                      None
+                  | (@expr.Ident _ _ _ t3 idc3 @ #(_))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @ ($_)%expr)%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @ (($_)%expr @ _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @ (@expr.Abs _ _ _ _ _ _ @ _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @ (_ @ _ @ _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.LetIn _ _ _ _ _ _ _ @ _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
+                      None
+                  | _ => None
+                  end;;
+                  match x0 with
+                  | (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
+                      (@expr.Ident _ _ _ t5 idc5) x9))%expr_pat =>
+                      args <- invert_bind_args idc5 Raw.ident.Z_cast;
+                      _ <- invert_bind_args idc4 Raw.ident.Z_opp;
+                      args1 <- invert_bind_args idc3 Raw.ident.Z_cast;
+                      args2 <- invert_bind_args idc2 Raw.ident.Z_cast;
+                      _ <- invert_bind_args idc1 Raw.ident.Z_opp;
+                      args4 <- invert_bind_args idc0 Raw.ident.Z_cast;
+                      _ <- invert_bind_args idc
+                             Raw.ident.Z_add_with_get_carry;
+                      match
+                        pattern.type.unify_extracted_cps
+                          (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                          (((s2 -> s5) -> s0) -> s8)%ptype option
+                          (fun x10 : option => x10)
+                      with
+                      | Some (_, _, _, _)%zrange =>
+                          if
+                           type.type_beq base.type base.type.type_beq
+                             (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                             (((s2 -> s5) -> s0) -> s8)%ptype
+                          then
+                           v <- type.try_make_transport_cps s2 ℤ;
+                           v0 <- type.try_make_transport_cps s5 ℤ;
+                           v1 <- type.try_make_transport_cps s0 ℤ;
+                           v2 <- type.try_make_transport_cps s8 ℤ;
+                           fv <- (if
+                                   (ZRange.normalize args <=?
+                                    - ZRange.normalize args1)%zrange &&
+                                   (ZRange.normalize args2 <=?
+                                    - ZRange.normalize args4)%zrange
+                                  then
+                                   Some
+                                     (UnderLet
+                                        (#(Z_cast2
+                                             (Datatypes.fst range,
+                                             - Datatypes.snd range))%expr @
+                                         (#(Z_sub_with_get_borrow)%expr @
+                                          v (Compile.reflect x3) @
+                                          (#(Z_cast args2)%expr @
+                                           v0 (Compile.reflect x6)) @
+                                          v1 (Compile.reflect x1) @
+                                          (#(Z_cast args)%expr @
+                                           v2 (Compile.reflect x9))))%expr_pat
+                                        (fun vc : var (ℤ * ℤ)%etype =>
+                                         Base
+                                           (#(Z_cast (Datatypes.fst range))%expr @
+                                            (#(fst)%expr @
+                                             (#(Z_cast2
+                                                  (Datatypes.fst range,
+                                                  - Datatypes.snd range))%expr @
+                                              ($vc)%expr)),
+                                           #(Z_cast (Datatypes.snd range))%expr @
+                                           (-
+                                            (#(Z_cast (- Datatypes.snd range))%expr @
+                                             (#(snd)%expr @
                                               (#(Z_cast2
                                                    (Datatypes.fst range,
                                                    - Datatypes.snd range))%expr @
-                                               ($vc)%expr)),
-                                            #(Z_cast (Datatypes.snd range))%expr @
-                                            (-
-                                             (#(Z_cast
-                                                  (- Datatypes.snd range))%expr @
-                                              (#(snd)%expr @
-                                               (#(Z_cast2
-                                                    (Datatypes.fst range,
-                                                    - Datatypes.snd range))%expr @
-                                                $vc)))%expr_pat)%expr)%expr_pat))
-                                   else None);
-                            Some (fv0 <-- fv;
-                                  Base fv0)%under_lets
-                           else None
-                       | None => None
-                       end
-                   | (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
-                       ($_)%expr _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
-                       (@expr.Abs _ _ _ _ _ _) _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
-                       (_ @ _) _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
-                       (@expr.LetIn _ _ _ _ _ _ _) _))%expr_pat => None
-                   | (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ #(_)))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ ($_)%expr))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.Abs _ _ _ _ _ _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.Ident _ _ _ t4 idc4 @ @expr.LetIn _ _ _ _ _ _ _))%expr_pat =>
-                       None
-                   | (@expr.Ident _ _ _ t3 idc3 @ #(_))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @ ($_)%expr)%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @ (($_)%expr @ _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @ (@expr.Abs _ _ _ _ _ _ @ _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @ (_ @ _ @ _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @
-                      (@expr.LetIn _ _ _ _ _ _ _ @ _))%expr_pat |
-                     (@expr.Ident _ _ _ t3 idc3 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
-                       None
-                   | _ => None
-                   end;;
-                   match x1 with
-                   | @expr.Ident _ _ _ t3 idc3 =>
-                       args <- invert_bind_args idc3 Raw.ident.Literal;
-                       args0 <- invert_bind_args idc2 Raw.ident.Z_cast;
-                       _ <- invert_bind_args idc1 Raw.ident.Z_opp;
-                       args2 <- invert_bind_args idc0 Raw.ident.Z_cast;
-                       _ <- invert_bind_args idc
-                              Raw.ident.Z_add_with_get_carry;
-                       match
-                         pattern.type.unify_extracted_cps
-                           (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                           (((s2 -> s5) -> (projT1 args)) -> s)%ptype option
-                           (fun x7 : option => x7)
-                       with
-                       | Some (_, _, _, _)%zrange =>
-                           if
-                            type.type_beq base.type base.type.type_beq
-                              (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                              (((s2 -> s5) -> (projT1 args)) -> s)%ptype
-                           then
-                            v <- type.try_make_transport_cps s2 ℤ;
-                            v0 <- type.try_make_transport_cps s5 ℤ;
-                            xv <- ident.unify pattern.ident.Literal
-                                    ##(projT2 args);
-                            v1 <- type.try_make_transport_cps s ℤ;
-                            fv <- (if
-                                    ((let (x7, _) := xv in x7) <=? 0) &&
-                                    (ZRange.normalize args0 <=?
-                                     - ZRange.normalize args2)%zrange
-                                   then
-                                    Some
-                                      (UnderLet
-                                         (#(Z_cast2
-                                              (Datatypes.fst range,
-                                              - Datatypes.snd range))%expr @
-                                          (#(Z_sub_with_get_borrow)%expr @
-                                           v (Compile.reflect x3) @
-                                           (#(Z_cast args0)%expr @
-                                            v0 (Compile.reflect x6)) @
-                                           v1 (Compile.reflect x0) @
-                                           (##(- (let (x7, _) := xv in x7))%Z)%expr))%expr_pat
-                                         (fun vc : var (ℤ * ℤ)%etype =>
-                                          Base
-                                            (#(Z_cast (Datatypes.fst range))%expr @
-                                             (#(fst)%expr @
+                                               $vc)))%expr_pat)%expr)%expr_pat))
+                                  else None);
+                           Some (fv0 <-- fv;
+                                 Base fv0)%under_lets
+                          else None
+                      | None => None
+                      end
+                  | (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
+                      ($_)%expr _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
+                      (@expr.Abs _ _ _ _ _ _) _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
+                      (_ @ _) _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.App _ _ _ s8 _
+                      (@expr.LetIn _ _ _ _ _ _ _) _))%expr_pat => None
+                  | (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ #(_)))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ ($_)%expr))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.Abs _ _ _ _ _ _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.Ident _ _ _ t4 idc4 @ @expr.LetIn _ _ _ _ _ _ _))%expr_pat =>
+                      None
+                  | (@expr.Ident _ _ _ t3 idc3 @ #(_))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @ ($_)%expr)%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @ (($_)%expr @ _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @ (@expr.Abs _ _ _ _ _ _ @ _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @ (_ @ _ @ _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @
+                     (@expr.LetIn _ _ _ _ _ _ _ @ _))%expr_pat |
+                    (@expr.Ident _ _ _ t3 idc3 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
+                      None
+                  | _ => None
+                  end;;
+                  match x1 with
+                  | @expr.Ident _ _ _ t3 idc3 =>
+                      args <- invert_bind_args idc3 Raw.ident.Literal;
+                      args0 <- invert_bind_args idc2 Raw.ident.Z_cast;
+                      _ <- invert_bind_args idc1 Raw.ident.Z_opp;
+                      args2 <- invert_bind_args idc0 Raw.ident.Z_cast;
+                      _ <- invert_bind_args idc
+                             Raw.ident.Z_add_with_get_carry;
+                      match
+                        pattern.type.unify_extracted_cps
+                          (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                          (((s2 -> s5) -> (projT1 args)) -> s)%ptype option
+                          (fun x7 : option => x7)
+                      with
+                      | Some (_, _, _, _)%zrange =>
+                          if
+                           type.type_beq base.type base.type.type_beq
+                             (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                             (((s2 -> s5) -> (projT1 args)) -> s)%ptype
+                          then
+                           v <- type.try_make_transport_cps s2 ℤ;
+                           v0 <- type.try_make_transport_cps s5 ℤ;
+                           xv <- ident.unify pattern.ident.Literal
+                                   ##(projT2 args);
+                           v1 <- type.try_make_transport_cps s ℤ;
+                           fv <- (if
+                                   ((let (x7, _) := xv in x7) <=? 0) &&
+                                   (ZRange.normalize args0 <=?
+                                    - ZRange.normalize args2)%zrange
+                                  then
+                                   Some
+                                     (UnderLet
+                                        (#(Z_cast2
+                                             (Datatypes.fst range,
+                                             - Datatypes.snd range))%expr @
+                                         (#(Z_sub_with_get_borrow)%expr @
+                                          v (Compile.reflect x3) @
+                                          (#(Z_cast args0)%expr @
+                                           v0 (Compile.reflect x6)) @
+                                          v1 (Compile.reflect x0) @
+                                          (##(- (let (x7, _) := xv in x7))%Z)%expr))%expr_pat
+                                        (fun vc : var (ℤ * ℤ)%etype =>
+                                         Base
+                                           (#(Z_cast (Datatypes.fst range))%expr @
+                                            (#(fst)%expr @
+                                             (#(Z_cast2
+                                                  (Datatypes.fst range,
+                                                  - Datatypes.snd range))%expr @
+                                              ($vc)%expr)),
+                                           #(Z_cast (Datatypes.snd range))%expr @
+                                           (-
+                                            (#(Z_cast (- Datatypes.snd range))%expr @
+                                             (#(snd)%expr @
                                               (#(Z_cast2
                                                    (Datatypes.fst range,
                                                    - Datatypes.snd range))%expr @
-                                               ($vc)%expr)),
-                                            #(Z_cast (Datatypes.snd range))%expr @
-                                            (-
-                                             (#(Z_cast
-                                                  (- Datatypes.snd range))%expr @
-                                              (#(snd)%expr @
-                                               (#(Z_cast2
-                                                    (Datatypes.fst range,
-                                                    - Datatypes.snd range))%expr @
-                                                $vc)))%expr_pat)%expr)%expr_pat))
-                                   else None);
-                            Some (fv0 <-- fv;
-                                  Base fv0)%under_lets
-                           else None
-                       | None => None
-                       end
-                   | _ => None
-                   end;;
-                   match x0 with
-                   | @expr.Ident _ _ _ t3 idc3 =>
-                       args <- invert_bind_args idc3 Raw.ident.Literal;
-                       args0 <- invert_bind_args idc2 Raw.ident.Z_cast;
-                       _ <- invert_bind_args idc1 Raw.ident.Z_opp;
-                       args2 <- invert_bind_args idc0 Raw.ident.Z_cast;
-                       _ <- invert_bind_args idc
-                              Raw.ident.Z_add_with_get_carry;
-                       match
-                         pattern.type.unify_extracted_cps
-                           (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                           (((s2 -> s5) -> s0) -> (projT1 args))%ptype option
-                           (fun x7 : option => x7)
-                       with
-                       | Some (_, _, _, _)%zrange =>
-                           if
-                            type.type_beq base.type base.type.type_beq
-                              (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                              (((s2 -> s5) -> s0) -> (projT1 args))%ptype
-                           then
-                            v <- type.try_make_transport_cps s2 ℤ;
-                            v0 <- type.try_make_transport_cps s5 ℤ;
-                            v1 <- type.try_make_transport_cps s0 ℤ;
-                            xv <- ident.unify pattern.ident.Literal
-                                    ##(projT2 args);
-                            fv <- (if
-                                    ((let (x7, _) := xv in x7) <=? 0) &&
-                                    (ZRange.normalize args0 <=?
-                                     - ZRange.normalize args2)%zrange
-                                   then
-                                    Some
-                                      (UnderLet
-                                         (#(Z_cast2
-                                              (Datatypes.fst range,
-                                              - Datatypes.snd range))%expr @
-                                          (#(Z_sub_with_get_borrow)%expr @
-                                           v (Compile.reflect x3) @
-                                           (#(Z_cast args0)%expr @
-                                            v0 (Compile.reflect x6)) @
-                                           v1 (Compile.reflect x1) @
-                                           (##(- (let (x7, _) := xv in x7))%Z)%expr))%expr_pat
-                                         (fun vc : var (ℤ * ℤ)%etype =>
-                                          Base
-                                            (#(Z_cast (Datatypes.fst range))%expr @
-                                             (#(fst)%expr @
+                                               $vc)))%expr_pat)%expr)%expr_pat))
+                                  else None);
+                           Some (fv0 <-- fv;
+                                 Base fv0)%under_lets
+                          else None
+                      | None => None
+                      end
+                  | _ => None
+                  end;;
+                  match x0 with
+                  | @expr.Ident _ _ _ t3 idc3 =>
+                      args <- invert_bind_args idc3 Raw.ident.Literal;
+                      args0 <- invert_bind_args idc2 Raw.ident.Z_cast;
+                      _ <- invert_bind_args idc1 Raw.ident.Z_opp;
+                      args2 <- invert_bind_args idc0 Raw.ident.Z_cast;
+                      _ <- invert_bind_args idc
+                             Raw.ident.Z_add_with_get_carry;
+                      match
+                        pattern.type.unify_extracted_cps
+                          (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                          (((s2 -> s5) -> s0) -> (projT1 args))%ptype option
+                          (fun x7 : option => x7)
+                      with
+                      | Some (_, _, _, _)%zrange =>
+                          if
+                           type.type_beq base.type base.type.type_beq
+                             (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                             (((s2 -> s5) -> s0) -> (projT1 args))%ptype
+                          then
+                           v <- type.try_make_transport_cps s2 ℤ;
+                           v0 <- type.try_make_transport_cps s5 ℤ;
+                           v1 <- type.try_make_transport_cps s0 ℤ;
+                           xv <- ident.unify pattern.ident.Literal
+                                   ##(projT2 args);
+                           fv <- (if
+                                   ((let (x7, _) := xv in x7) <=? 0) &&
+                                   (ZRange.normalize args0 <=?
+                                    - ZRange.normalize args2)%zrange
+                                  then
+                                   Some
+                                     (UnderLet
+                                        (#(Z_cast2
+                                             (Datatypes.fst range,
+                                             - Datatypes.snd range))%expr @
+                                         (#(Z_sub_with_get_borrow)%expr @
+                                          v (Compile.reflect x3) @
+                                          (#(Z_cast args0)%expr @
+                                           v0 (Compile.reflect x6)) @
+                                          v1 (Compile.reflect x1) @
+                                          (##(- (let (x7, _) := xv in x7))%Z)%expr))%expr_pat
+                                        (fun vc : var (ℤ * ℤ)%etype =>
+                                         Base
+                                           (#(Z_cast (Datatypes.fst range))%expr @
+                                            (#(fst)%expr @
+                                             (#(Z_cast2
+                                                  (Datatypes.fst range,
+                                                  - Datatypes.snd range))%expr @
+                                              ($vc)%expr)),
+                                           #(Z_cast (Datatypes.snd range))%expr @
+                                           (-
+                                            (#(Z_cast (- Datatypes.snd range))%expr @
+                                             (#(snd)%expr @
                                               (#(Z_cast2
                                                    (Datatypes.fst range,
                                                    - Datatypes.snd range))%expr @
-                                               ($vc)%expr)),
-                                            #(Z_cast (Datatypes.snd range))%expr @
-                                            (-
-                                             (#(Z_cast
-                                                  (- Datatypes.snd range))%expr @
-                                              (#(snd)%expr @
-                                               (#(Z_cast2
-                                                    (Datatypes.fst range,
-                                                    - Datatypes.snd range))%expr @
-                                                $vc)))%expr_pat)%expr)%expr_pat))
-                                   else None);
-                            Some (fv0 <-- fv;
-                                  Base fv0)%under_lets
-                           else None
-                       | None => None
-                       end
-                   | _ => None
-                   end
-               | (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s5 _ ($_)%expr
-                  _)%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s5 _
-                  (@expr.Abs _ _ _ _ _ _) _)%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s5 _ (_ @ _) _)%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s5 _
-                  (@expr.LetIn _ _ _ _ _ _ _) _)%expr_pat => None
-               | (@expr.Ident _ _ _ t1 idc1 @ #(_))%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ ($_)%expr)%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
-                 (@expr.Ident _ _ _ t1 idc1 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
-                   None
-               | _ => None
-               end;;
-               match x3 with
-               | @expr.Ident _ _ _ t1 idc1 =>
-                   match x1 with
-                   | @expr.Ident _ _ _ t2 idc2 =>
-                       match x0 with
-                       | @expr.Ident _ _ _ t3 idc3 =>
-                           args <- invert_bind_args idc3 Raw.ident.Literal;
-                           args0 <- invert_bind_args idc2 Raw.ident.Literal;
-                           args1 <- invert_bind_args idc1 Raw.ident.Literal;
-                           args2 <- invert_bind_args idc0 Raw.ident.Z_cast;
-                           _ <- invert_bind_args idc
-                                  Raw.ident.Z_add_with_get_carry;
-                           match
-                             pattern.type.unify_extracted_cps
-                               (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                               ((((projT1 args1) -> s3) -> (projT1 args0)) ->
-                                (projT1 args))%ptype option
-                               (fun x5 : option => x5)
-                           with
-                           | Some (_, _, _, _)%zrange =>
-                               if
-                                type.type_beq base.type base.type.type_beq
-                                  (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                                  ((((projT1 args1) -> s3) -> (projT1 args0)) ->
-                                   (projT1 args))%ptype
-                               then
-                                xv <- ident.unify pattern.ident.Literal
-                                        ##(projT2 args1);
-                                v <- type.try_make_transport_cps s3 ℤ;
-                                xv0 <- ident.unify pattern.ident.Literal
-                                         ##(projT2 args0);
-                                xv1 <- ident.unify pattern.ident.Literal
-                                         ##(projT2 args);
-                                fv <- (if
-                                        ((let (x5, _) := xv0 in x5) =? 0) &&
-                                        ((let (x5, _) := xv1 in x5) =? 0) &&
-                                        (ZRange.normalize args2 <=?
-                                         r[0 ~> (let (x5, _) := xv in x5) - 1])%zrange &&
-                                        is_bounded_by_bool 0
-                                          (Datatypes.snd range)
-                                       then
-                                        Some
-                                          (UnderLet
-                                             (#(Z_cast2 range)%expr @
-                                              (#(Z_add_with_get_carry)%expr @
-                                               (##(let (x5, _) := xv in x5))%expr @
-                                               (#(Z_cast args2)%expr @
-                                                v (Compile.reflect x4)) @
-                                               (##(let (x5, _) := xv0 in x5))%expr @
-                                               (##(let (x5, _) := xv1 in x5))%expr))%expr_pat
-                                             (fun vc : var (ℤ * ℤ)%etype =>
-                                              Base
-                                                (#(Z_cast
-                                                     (Datatypes.fst range))%expr @
-                                                 (#(fst)%expr @
-                                                  (#(Z_cast2 range)%expr @
-                                                   ($vc)%expr)), (##0)%expr)%expr_pat))
-                                       else None);
-                                Some (fv0 <-- fv;
-                                      Base fv0)%under_lets
-                               else None
-                           | None => None
-                           end
-                       | _ => None
-                       end
-                   | _ => None
-                   end
-               | _ => None
-               end
-           | @expr.App _ _ _ s3 _ ($_)%expr _ | @expr.App _ _ _ s3 _
-             (@expr.Abs _ _ _ _ _ _) _ | @expr.App _ _ _ s3 _
-             (_ @ _)%expr_pat _ | @expr.App _ _ _ s3 _
-             (@expr.LetIn _ _ _ _ _ _ _) _ => None
-           | _ => None
-           end;;
-           _ <- invert_bind_args idc Raw.ident.Z_add_with_get_carry;
-           match
-             pattern.type.unify_extracted_cps (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-               (((s2 -> s1) -> s0) -> s)%ptype option (fun x4 : option => x4)
-           with
-           | Some (_, _, _, _)%zrange =>
-               if
-                type.type_beq base.type base.type.type_beq
-                  (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
-                  (((s2 -> s1) -> s0) -> s)%ptype
-               then
-                v <- type.try_make_transport_cps s2 ℤ;
-                v0 <- type.try_make_transport_cps s1 ℤ;
-                v1 <- type.try_make_transport_cps s0 ℤ;
-                v2 <- type.try_make_transport_cps s ℤ;
-                Some
-                  (UnderLet
-                     (#(Z_cast2 range)%expr @
-                      (#(Z_add_with_get_carry)%expr @ v (Compile.reflect x3) @
-                       v0 (Compile.reflect x2) @ v1 (Compile.reflect x1) @
-                       v2 (Compile.reflect x0)))%expr_pat
-                     (fun v3 : var (ℤ * ℤ)%etype =>
-                      Base
-                        (#(Z_cast (Datatypes.fst range))%expr @
-                         (#(fst)%expr @ (#(Z_cast2 range)%expr @ ($v3)%expr)),
-                        #(Z_cast (Datatypes.snd range))%expr @
-                        (#(snd)%expr @ (#(Z_cast2 range)%expr @ ($v3)%expr)))%expr_pat))
-               else None
-           | None => None
-           end);;
-          _ <- invert_bind_args idc Raw.ident.Z_sub_with_get_borrow;
+                                               $vc)))%expr_pat)%expr)%expr_pat))
+                                  else None);
+                           Some (fv0 <-- fv;
+                                 Base fv0)%under_lets
+                          else None
+                      | None => None
+                      end
+                  | _ => None
+                  end
+              | (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s5 _ ($_)%expr _)%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s5 _
+                 (@expr.Abs _ _ _ _ _ _) _)%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s5 _ (_ @ _) _)%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ @expr.App _ _ _ s5 _
+                 (@expr.LetIn _ _ _ _ _ _ _) _)%expr_pat => None
+              | (@expr.Ident _ _ _ t1 idc1 @ #(_))%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ ($_)%expr)%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ @expr.Abs _ _ _ _ _ _)%expr_pat |
+                (@expr.Ident _ _ _ t1 idc1 @ @expr.LetIn _ _ _ _ _ _ _)%expr_pat =>
+                  None
+              | _ => None
+              end;;
+              match x3 with
+              | @expr.Ident _ _ _ t1 idc1 =>
+                  match x1 with
+                  | @expr.Ident _ _ _ t2 idc2 =>
+                      match x0 with
+                      | @expr.Ident _ _ _ t3 idc3 =>
+                          args <- invert_bind_args idc3 Raw.ident.Literal;
+                          args0 <- invert_bind_args idc2 Raw.ident.Literal;
+                          args1 <- invert_bind_args idc1 Raw.ident.Literal;
+                          args2 <- invert_bind_args idc0 Raw.ident.Z_cast;
+                          _ <- invert_bind_args idc
+                                 Raw.ident.Z_add_with_get_carry;
+                          match
+                            pattern.type.unify_extracted_cps
+                              (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                              ((((projT1 args1) -> s3) -> (projT1 args0)) ->
+                               (projT1 args))%ptype option
+                              (fun x5 : option => x5)
+                          with
+                          | Some (_, _, _, _)%zrange =>
+                              if
+                               type.type_beq base.type base.type.type_beq
+                                 (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+                                 ((((projT1 args1) -> s3) -> (projT1 args0)) ->
+                                  (projT1 args))%ptype
+                              then
+                               xv <- ident.unify pattern.ident.Literal
+                                       ##(projT2 args1);
+                               v <- type.try_make_transport_cps s3 ℤ;
+                               xv0 <- ident.unify pattern.ident.Literal
+                                        ##(projT2 args0);
+                               xv1 <- ident.unify pattern.ident.Literal
+                                        ##(projT2 args);
+                               fv <- (if
+                                       ((let (x5, _) := xv0 in x5) =? 0) &&
+                                       ((let (x5, _) := xv1 in x5) =? 0) &&
+                                       (ZRange.normalize args2 <=?
+                                        r[0 ~> (let (x5, _) := xv in x5) - 1])%zrange &&
+                                       is_bounded_by_bool 0
+                                         (Datatypes.snd range)
+                                      then
+                                       Some
+                                         (UnderLet
+                                            (#(Z_cast2 range)%expr @
+                                             (#(Z_add_with_get_carry)%expr @
+                                              (##(let (x5, _) := xv in x5))%expr @
+                                              (#(Z_cast args2)%expr @
+                                               v (Compile.reflect x4)) @
+                                              (##(let (x5, _) := xv0 in x5))%expr @
+                                              (##(let (x5, _) := xv1 in x5))%expr))%expr_pat
+                                            (fun vc : var (ℤ * ℤ)%etype =>
+                                             Base
+                                               (#(Z_cast
+                                                    (Datatypes.fst range))%expr @
+                                                (#(fst)%expr @
+                                                 (#(Z_cast2 range)%expr @
+                                                  ($vc)%expr)), (##0)%expr)%expr_pat))
+                                      else None);
+                               Some (fv0 <-- fv;
+                                     Base fv0)%under_lets
+                              else None
+                          | None => None
+                          end
+                      | _ => None
+                      end
+                  | _ => None
+                  end
+              | _ => None
+              end
+          | @expr.App _ _ _ s3 _ ($_)%expr _ | @expr.App _ _ _ s3 _
+            (@expr.Abs _ _ _ _ _ _) _ | @expr.App _ _ _ s3 _ (_ @ _)%expr_pat
+            _ | @expr.App _ _ _ s3 _ (@expr.LetIn _ _ _ _ _ _ _) _ => None
+          | _ => None
+          end;;
+          _ <- invert_bind_args idc Raw.ident.Z_add_with_get_carry;
           match
             pattern.type.unify_extracted_cps (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
               (((s2 -> s1) -> s0) -> s)%ptype option (fun x4 : option => x4)
@@ -2332,7 +2308,7 @@ match idc in (Compilers.ident t) return (Compile.value' true t) with
                Some
                  (UnderLet
                     (#(Z_cast2 range)%expr @
-                     (#(Z_sub_with_get_borrow)%expr @ v (Compile.reflect x3) @
+                     (#(Z_add_with_get_carry)%expr @ v (Compile.reflect x3) @
                       v0 (Compile.reflect x2) @ v1 (Compile.reflect x1) @
                       v2 (Compile.reflect x0)))%expr_pat
                     (fun v3 : var (ℤ * ℤ)%etype =>
@@ -2343,40 +2319,66 @@ match idc in (Compilers.ident t) return (Compile.value' true t) with
                        (#(snd)%expr @ (#(Z_cast2 range)%expr @ ($v3)%expr)))%expr_pat))
               else None
           | None => None
-          end
-      | @expr.App _ _ _ s _
-        (@expr.App _ _ _ s0 _
-         (@expr.App _ _ _ s1 _ (@expr.App _ _ _ s2 _ ($_)%expr _) _) _) _ |
-        @expr.App _ _ _ s _
-        (@expr.App _ _ _ s0 _
-         (@expr.App _ _ _ s1 _
-          (@expr.App _ _ _ s2 _ (@expr.Abs _ _ _ _ _ _) _) _) _) _ |
-        @expr.App _ _ _ s _
-        (@expr.App _ _ _ s0 _
-         (@expr.App _ _ _ s1 _ (@expr.App _ _ _ s2 _ (_ @ _)%expr_pat _) _)
-         _) _ | @expr.App _ _ _ s _
-        (@expr.App _ _ _ s0 _
-         (@expr.App _ _ _ s1 _
-          (@expr.App _ _ _ s2 _ (@expr.LetIn _ _ _ _ _ _ _) _) _) _) _ =>
-          None
-      | @expr.App _ _ _ s _
-        (@expr.App _ _ _ s0 _ (@expr.App _ _ _ s1 _ ($_)%expr _) _) _ |
-        @expr.App _ _ _ s _
-        (@expr.App _ _ _ s0 _
-         (@expr.App _ _ _ s1 _ (@expr.Abs _ _ _ _ _ _) _) _) _ | @expr.App _
-        _ _ s _
-        (@expr.App _ _ _ s0 _
-         (@expr.App _ _ _ s1 _ (@expr.LetIn _ _ _ _ _ _ _) _) _) _ => None
-      | @expr.App _ _ _ s _ (@expr.App _ _ _ s0 _ ($_)%expr _) _ | @expr.App
-        _ _ _ s _ (@expr.App _ _ _ s0 _ (@expr.Abs _ _ _ _ _ _) _) _ |
-        @expr.App _ _ _ s _
-        (@expr.App _ _ _ s0 _ (@expr.LetIn _ _ _ _ _ _ _) _) _ => None
-      | @expr.App _ _ _ s _ #(_)%expr_pat _ | @expr.App _ _ _ s _ ($_)%expr
-        _ | @expr.App _ _ _ s _ (@expr.Abs _ _ _ _ _ _) _ | @expr.App _ _ _ s
-        _ (@expr.LetIn _ _ _ _ _ _ _) _ => None
-      | _ => None
-      end;;
-      None);;;
+          end);;
+         _ <- invert_bind_args idc Raw.ident.Z_sub_with_get_borrow;
+         match
+           pattern.type.unify_extracted_cps (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype
+             (((s2 -> s1) -> s0) -> s)%ptype option (fun x4 : option => x4)
+         with
+         | Some (_, _, _, _)%zrange =>
+             if
+              type.type_beq base.type base.type.type_beq
+                (((ℤ -> ℤ) -> ℤ) -> ℤ)%ptype (((s2 -> s1) -> s0) -> s)%ptype
+             then
+              v <- type.try_make_transport_cps s2 ℤ;
+              v0 <- type.try_make_transport_cps s1 ℤ;
+              v1 <- type.try_make_transport_cps s0 ℤ;
+              v2 <- type.try_make_transport_cps s ℤ;
+              Some
+                (UnderLet
+                   (#(Z_cast2 range)%expr @
+                    (#(Z_sub_with_get_borrow)%expr @ v (Compile.reflect x3) @
+                     v0 (Compile.reflect x2) @ v1 (Compile.reflect x1) @
+                     v2 (Compile.reflect x0)))%expr_pat
+                   (fun v3 : var (ℤ * ℤ)%etype =>
+                    Base
+                      (#(Z_cast (Datatypes.fst range))%expr @
+                       (#(fst)%expr @ (#(Z_cast2 range)%expr @ ($v3)%expr)),
+                      #(Z_cast (Datatypes.snd range))%expr @
+                      (#(snd)%expr @ (#(Z_cast2 range)%expr @ ($v3)%expr)))%expr_pat))
+             else None
+         | None => None
+         end
+     | @expr.App _ _ _ s _
+       (@expr.App _ _ _ s0 _
+        (@expr.App _ _ _ s1 _ (@expr.App _ _ _ s2 _ ($_)%expr _) _) _) _ |
+       @expr.App _ _ _ s _
+       (@expr.App _ _ _ s0 _
+        (@expr.App _ _ _ s1 _
+         (@expr.App _ _ _ s2 _ (@expr.Abs _ _ _ _ _ _) _) _) _) _ | @expr.App
+       _ _ _ s _
+       (@expr.App _ _ _ s0 _
+        (@expr.App _ _ _ s1 _ (@expr.App _ _ _ s2 _ (_ @ _)%expr_pat _) _) _)
+       _ | @expr.App _ _ _ s _
+       (@expr.App _ _ _ s0 _
+        (@expr.App _ _ _ s1 _
+         (@expr.App _ _ _ s2 _ (@expr.LetIn _ _ _ _ _ _ _) _) _) _) _ => None
+     | @expr.App _ _ _ s _
+       (@expr.App _ _ _ s0 _ (@expr.App _ _ _ s1 _ ($_)%expr _) _) _ |
+       @expr.App _ _ _ s _
+       (@expr.App _ _ _ s0 _ (@expr.App _ _ _ s1 _ (@expr.Abs _ _ _ _ _ _) _)
+        _) _ | @expr.App _ _ _ s _
+       (@expr.App _ _ _ s0 _
+        (@expr.App _ _ _ s1 _ (@expr.LetIn _ _ _ _ _ _ _) _) _) _ => None
+     | @expr.App _ _ _ s _ (@expr.App _ _ _ s0 _ ($_)%expr _) _ | @expr.App _
+       _ _ s _ (@expr.App _ _ _ s0 _ (@expr.Abs _ _ _ _ _ _) _) _ | @expr.App
+       _ _ _ s _ (@expr.App _ _ _ s0 _ (@expr.LetIn _ _ _ _ _ _ _) _) _ =>
+         None
+     | @expr.App _ _ _ s _ #(_)%expr_pat _ | @expr.App _ _ _ s _ ($_)%expr
+       _ | @expr.App _ _ _ s _ (@expr.Abs _ _ _ _ _ _) _ | @expr.App _ _ _ s
+       _ (@expr.LetIn _ _ _ _ _ _ _) _ => None
+     | _ => None
+     end;;;
      Base (#(Z_cast2 range)%expr @ x)%expr_pat)%option
 | fancy_add log2wordmax imm =>
     fun x : expr (ℤ * ℤ)%etype =>
author	Jason Gross <jgross@mit.edu>	2019-01-15 15:53:34 -0500
committer	Jason Gross <jasongross9@gmail.com>	2019-01-16 16:51:11 -0500
commit	60bade02ccd577550bfcd5974d3c62a3d40e751a (patch)
tree	46a9e9ef505704bdff47f0c794c163ecd5d7fd76
parent	538e541709d70caaa78104739cb965c6523d89a8 (diff)