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%

1 files changed, 399 insertions, 510 deletions

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;
 }