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, 286 insertions, 296 deletions

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;
 }