Rebuild .c files

It seems some commutativity has changed

1 files changed, 1554 insertions, 1554 deletions

diff --git a/p384_32.c b/p384_32.c
index c9591bdfc..b959e6389 100644
--- a/p384_32.c
+++ b/p384_32.c
@@ -166,40 +166,40 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x35, &x36, x12, (arg2[0]));
   uint32_t x37;
   fiat_p384_uint1 x38;
-  fiat_p384_addcarryx_u32(&x37, &x38, 0x0, x33, x36);
+  fiat_p384_addcarryx_u32(&x37, &x38, 0x0, x36, x33);
   uint32_t x39;
   fiat_p384_uint1 x40;
-  fiat_p384_addcarryx_u32(&x39, &x40, x38, x31, x34);
+  fiat_p384_addcarryx_u32(&x39, &x40, x38, x34, x31);
   uint32_t x41;
   fiat_p384_uint1 x42;
-  fiat_p384_addcarryx_u32(&x41, &x42, x40, x29, x32);
+  fiat_p384_addcarryx_u32(&x41, &x42, x40, x32, x29);
   uint32_t x43;
   fiat_p384_uint1 x44;
-  fiat_p384_addcarryx_u32(&x43, &x44, x42, x27, x30);
+  fiat_p384_addcarryx_u32(&x43, &x44, x42, x30, x27);
   uint32_t x45;
   fiat_p384_uint1 x46;
-  fiat_p384_addcarryx_u32(&x45, &x46, x44, x25, x28);
+  fiat_p384_addcarryx_u32(&x45, &x46, x44, x28, x25);
   uint32_t x47;
   fiat_p384_uint1 x48;
-  fiat_p384_addcarryx_u32(&x47, &x48, x46, x23, x26);
+  fiat_p384_addcarryx_u32(&x47, &x48, x46, x26, x23);
   uint32_t x49;
   fiat_p384_uint1 x50;
-  fiat_p384_addcarryx_u32(&x49, &x50, x48, x21, x24);
+  fiat_p384_addcarryx_u32(&x49, &x50, x48, x24, x21);
   uint32_t x51;
   fiat_p384_uint1 x52;
-  fiat_p384_addcarryx_u32(&x51, &x52, x50, x19, x22);
+  fiat_p384_addcarryx_u32(&x51, &x52, x50, x22, x19);
   uint32_t x53;
   fiat_p384_uint1 x54;
-  fiat_p384_addcarryx_u32(&x53, &x54, x52, x17, x20);
+  fiat_p384_addcarryx_u32(&x53, &x54, x52, x20, x17);
   uint32_t x55;
   fiat_p384_uint1 x56;
-  fiat_p384_addcarryx_u32(&x55, &x56, x54, x15, x18);
+  fiat_p384_addcarryx_u32(&x55, &x56, x54, x18, x15);
   uint32_t x57;
   fiat_p384_uint1 x58;
-  fiat_p384_addcarryx_u32(&x57, &x58, x56, x13, x16);
+  fiat_p384_addcarryx_u32(&x57, &x58, x56, x16, x13);
   uint32_t x59;
   fiat_p384_uint1 x60;
-  fiat_p384_addcarryx_u32(&x59, &x60, x58, 0x0, x14);
+  fiat_p384_addcarryx_u32(&x59, &x60, x58, x14, 0x0);
   uint32_t x61;
   uint32_t x62;
   fiat_p384_mulx_u32(&x61, &x62, x35, UINT32_C(0xffffffff));
@@ -232,70 +232,70 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x79, &x80, x35, UINT32_C(0xffffffff));
   uint32_t x81;
   fiat_p384_uint1 x82;
-  fiat_p384_addcarryx_u32(&x81, &x82, 0x0, x75, x78);
+  fiat_p384_addcarryx_u32(&x81, &x82, 0x0, x78, x75);
   uint32_t x83;
   fiat_p384_uint1 x84;
-  fiat_p384_addcarryx_u32(&x83, &x84, x82, x73, x76);
+  fiat_p384_addcarryx_u32(&x83, &x84, x82, x76, x73);
   uint32_t x85;
   fiat_p384_uint1 x86;
-  fiat_p384_addcarryx_u32(&x85, &x86, x84, x71, x74);
+  fiat_p384_addcarryx_u32(&x85, &x86, x84, x74, x71);
   uint32_t x87;
   fiat_p384_uint1 x88;
-  fiat_p384_addcarryx_u32(&x87, &x88, x86, x69, x72);
+  fiat_p384_addcarryx_u32(&x87, &x88, x86, x72, x69);
   uint32_t x89;
   fiat_p384_uint1 x90;
-  fiat_p384_addcarryx_u32(&x89, &x90, x88, x67, x70);
+  fiat_p384_addcarryx_u32(&x89, &x90, x88, x70, x67);
   uint32_t x91;
   fiat_p384_uint1 x92;
-  fiat_p384_addcarryx_u32(&x91, &x92, x90, x65, x68);
+  fiat_p384_addcarryx_u32(&x91, &x92, x90, x68, x65);
   uint32_t x93;
   fiat_p384_uint1 x94;
-  fiat_p384_addcarryx_u32(&x93, &x94, x92, x63, x66);
+  fiat_p384_addcarryx_u32(&x93, &x94, x92, x66, x63);
   uint32_t x95;
   fiat_p384_uint1 x96;
-  fiat_p384_addcarryx_u32(&x95, &x96, x94, x61, x64);
+  fiat_p384_addcarryx_u32(&x95, &x96, x94, x64, x61);
   uint32_t x97;
   fiat_p384_uint1 x98;
-  fiat_p384_addcarryx_u32(&x97, &x98, x96, 0x0, x62);
+  fiat_p384_addcarryx_u32(&x97, &x98, x96, x62, 0x0);
   uint32_t x99;
   fiat_p384_uint1 x100;
-  fiat_p384_addcarryx_u32(&x99, &x100, 0x0, x79, x35);
+  fiat_p384_addcarryx_u32(&x99, &x100, 0x0, x35, x79);
   uint32_t x101;
   fiat_p384_uint1 x102;
-  fiat_p384_addcarryx_u32(&x101, &x102, x100, x80, x37);
+  fiat_p384_addcarryx_u32(&x101, &x102, x100, x37, x80);
   uint32_t x103;
   fiat_p384_uint1 x104;
-  fiat_p384_addcarryx_u32(&x103, &x104, x102, 0x0, x39);
+  fiat_p384_addcarryx_u32(&x103, &x104, x102, x39, 0x0);
   uint32_t x105;
   fiat_p384_uint1 x106;
-  fiat_p384_addcarryx_u32(&x105, &x106, x104, x77, x41);
+  fiat_p384_addcarryx_u32(&x105, &x106, x104, x41, x77);
   uint32_t x107;
   fiat_p384_uint1 x108;
-  fiat_p384_addcarryx_u32(&x107, &x108, x106, x81, x43);
+  fiat_p384_addcarryx_u32(&x107, &x108, x106, x43, x81);
   uint32_t x109;
   fiat_p384_uint1 x110;
-  fiat_p384_addcarryx_u32(&x109, &x110, x108, x83, x45);
+  fiat_p384_addcarryx_u32(&x109, &x110, x108, x45, x83);
   uint32_t x111;
   fiat_p384_uint1 x112;
-  fiat_p384_addcarryx_u32(&x111, &x112, x110, x85, x47);
+  fiat_p384_addcarryx_u32(&x111, &x112, x110, x47, x85);
   uint32_t x113;
   fiat_p384_uint1 x114;
-  fiat_p384_addcarryx_u32(&x113, &x114, x112, x87, x49);
+  fiat_p384_addcarryx_u32(&x113, &x114, x112, x49, x87);
   uint32_t x115;
   fiat_p384_uint1 x116;
-  fiat_p384_addcarryx_u32(&x115, &x116, x114, x89, x51);
+  fiat_p384_addcarryx_u32(&x115, &x116, x114, x51, x89);
   uint32_t x117;
   fiat_p384_uint1 x118;
-  fiat_p384_addcarryx_u32(&x117, &x118, x116, x91, x53);
+  fiat_p384_addcarryx_u32(&x117, &x118, x116, x53, x91);
   uint32_t x119;
   fiat_p384_uint1 x120;
-  fiat_p384_addcarryx_u32(&x119, &x120, x118, x93, x55);
+  fiat_p384_addcarryx_u32(&x119, &x120, x118, x55, x93);
   uint32_t x121;
   fiat_p384_uint1 x122;
-  fiat_p384_addcarryx_u32(&x121, &x122, x120, x95, x57);
+  fiat_p384_addcarryx_u32(&x121, &x122, x120, x57, x95);
   uint32_t x123;
   fiat_p384_uint1 x124;
-  fiat_p384_addcarryx_u32(&x123, &x124, x122, x97, x59);
+  fiat_p384_addcarryx_u32(&x123, &x124, x122, x59, x97);
   uint32_t x125;
   fiat_p384_uint1 x126;
   fiat_p384_addcarryx_u32(&x125, &x126, x124, 0x0, 0x0);
@@ -337,79 +337,79 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x149, &x150, x1, (arg2[0]));
   uint32_t x151;
   fiat_p384_uint1 x152;
-  fiat_p384_addcarryx_u32(&x151, &x152, 0x0, x147, x150);
+  fiat_p384_addcarryx_u32(&x151, &x152, 0x0, x150, x147);
   uint32_t x153;
   fiat_p384_uint1 x154;
-  fiat_p384_addcarryx_u32(&x153, &x154, x152, x145, x148);
+  fiat_p384_addcarryx_u32(&x153, &x154, x152, x148, x145);
   uint32_t x155;
   fiat_p384_uint1 x156;
-  fiat_p384_addcarryx_u32(&x155, &x156, x154, x143, x146);
+  fiat_p384_addcarryx_u32(&x155, &x156, x154, x146, x143);
   uint32_t x157;
   fiat_p384_uint1 x158;
-  fiat_p384_addcarryx_u32(&x157, &x158, x156, x141, x144);
+  fiat_p384_addcarryx_u32(&x157, &x158, x156, x144, x141);
   uint32_t x159;
   fiat_p384_uint1 x160;
-  fiat_p384_addcarryx_u32(&x159, &x160, x158, x139, x142);
+  fiat_p384_addcarryx_u32(&x159, &x160, x158, x142, x139);
   uint32_t x161;
   fiat_p384_uint1 x162;
-  fiat_p384_addcarryx_u32(&x161, &x162, x160, x137, x140);
+  fiat_p384_addcarryx_u32(&x161, &x162, x160, x140, x137);
   uint32_t x163;
   fiat_p384_uint1 x164;
-  fiat_p384_addcarryx_u32(&x163, &x164, x162, x135, x138);
+  fiat_p384_addcarryx_u32(&x163, &x164, x162, x138, x135);
   uint32_t x165;
   fiat_p384_uint1 x166;
-  fiat_p384_addcarryx_u32(&x165, &x166, x164, x133, x136);
+  fiat_p384_addcarryx_u32(&x165, &x166, x164, x136, x133);
   uint32_t x167;
   fiat_p384_uint1 x168;
-  fiat_p384_addcarryx_u32(&x167, &x168, x166, x131, x134);
+  fiat_p384_addcarryx_u32(&x167, &x168, x166, x134, x131);
   uint32_t x169;
   fiat_p384_uint1 x170;
-  fiat_p384_addcarryx_u32(&x169, &x170, x168, x129, x132);
+  fiat_p384_addcarryx_u32(&x169, &x170, x168, x132, x129);
   uint32_t x171;
   fiat_p384_uint1 x172;
-  fiat_p384_addcarryx_u32(&x171, &x172, x170, x127, x130);
+  fiat_p384_addcarryx_u32(&x171, &x172, x170, x130, x127);
   uint32_t x173;
   fiat_p384_uint1 x174;
-  fiat_p384_addcarryx_u32(&x173, &x174, x172, 0x0, x128);
+  fiat_p384_addcarryx_u32(&x173, &x174, x172, x128, 0x0);
   uint32_t x175;
   fiat_p384_uint1 x176;
-  fiat_p384_addcarryx_u32(&x175, &x176, 0x0, x149, x101);
+  fiat_p384_addcarryx_u32(&x175, &x176, 0x0, x101, x149);
   uint32_t x177;
   fiat_p384_uint1 x178;
-  fiat_p384_addcarryx_u32(&x177, &x178, x176, x151, x103);
+  fiat_p384_addcarryx_u32(&x177, &x178, x176, x103, x151);
   uint32_t x179;
   fiat_p384_uint1 x180;
-  fiat_p384_addcarryx_u32(&x179, &x180, x178, x153, x105);
+  fiat_p384_addcarryx_u32(&x179, &x180, x178, x105, x153);
   uint32_t x181;
   fiat_p384_uint1 x182;
-  fiat_p384_addcarryx_u32(&x181, &x182, x180, x155, x107);
+  fiat_p384_addcarryx_u32(&x181, &x182, x180, x107, x155);
   uint32_t x183;
   fiat_p384_uint1 x184;
-  fiat_p384_addcarryx_u32(&x183, &x184, x182, x157, x109);
+  fiat_p384_addcarryx_u32(&x183, &x184, x182, x109, x157);
   uint32_t x185;
   fiat_p384_uint1 x186;
-  fiat_p384_addcarryx_u32(&x185, &x186, x184, x159, x111);
+  fiat_p384_addcarryx_u32(&x185, &x186, x184, x111, x159);
   uint32_t x187;
   fiat_p384_uint1 x188;
-  fiat_p384_addcarryx_u32(&x187, &x188, x186, x161, x113);
+  fiat_p384_addcarryx_u32(&x187, &x188, x186, x113, x161);
   uint32_t x189;
   fiat_p384_uint1 x190;
-  fiat_p384_addcarryx_u32(&x189, &x190, x188, x163, x115);
+  fiat_p384_addcarryx_u32(&x189, &x190, x188, x115, x163);
   uint32_t x191;
   fiat_p384_uint1 x192;
-  fiat_p384_addcarryx_u32(&x191, &x192, x190, x165, x117);
+  fiat_p384_addcarryx_u32(&x191, &x192, x190, x117, x165);
   uint32_t x193;
   fiat_p384_uint1 x194;
-  fiat_p384_addcarryx_u32(&x193, &x194, x192, x167, x119);
+  fiat_p384_addcarryx_u32(&x193, &x194, x192, x119, x167);
   uint32_t x195;
   fiat_p384_uint1 x196;
-  fiat_p384_addcarryx_u32(&x195, &x196, x194, x169, x121);
+  fiat_p384_addcarryx_u32(&x195, &x196, x194, x121, x169);
   uint32_t x197;
   fiat_p384_uint1 x198;
-  fiat_p384_addcarryx_u32(&x197, &x198, x196, x171, x123);
+  fiat_p384_addcarryx_u32(&x197, &x198, x196, x123, x171);
   uint32_t x199;
   fiat_p384_uint1 x200;
-  fiat_p384_addcarryx_u32(&x199, &x200, x198, x173, (fiat_p384_uint1)x125);
+  fiat_p384_addcarryx_u32(&x199, &x200, x198, (fiat_p384_uint1)x125, x173);
   uint32_t x201;
   uint32_t x202;
   fiat_p384_mulx_u32(&x201, &x202, x175, UINT32_C(0xffffffff));
@@ -442,73 +442,73 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x219, &x220, x175, UINT32_C(0xffffffff));
   uint32_t x221;
   fiat_p384_uint1 x222;
-  fiat_p384_addcarryx_u32(&x221, &x222, 0x0, x215, x218);
+  fiat_p384_addcarryx_u32(&x221, &x222, 0x0, x218, x215);
   uint32_t x223;
   fiat_p384_uint1 x224;
-  fiat_p384_addcarryx_u32(&x223, &x224, x222, x213, x216);
+  fiat_p384_addcarryx_u32(&x223, &x224, x222, x216, x213);
   uint32_t x225;
   fiat_p384_uint1 x226;
-  fiat_p384_addcarryx_u32(&x225, &x226, x224, x211, x214);
+  fiat_p384_addcarryx_u32(&x225, &x226, x224, x214, x211);
   uint32_t x227;
   fiat_p384_uint1 x228;
-  fiat_p384_addcarryx_u32(&x227, &x228, x226, x209, x212);
+  fiat_p384_addcarryx_u32(&x227, &x228, x226, x212, x209);
   uint32_t x229;
   fiat_p384_uint1 x230;
-  fiat_p384_addcarryx_u32(&x229, &x230, x228, x207, x210);
+  fiat_p384_addcarryx_u32(&x229, &x230, x228, x210, x207);
   uint32_t x231;
   fiat_p384_uint1 x232;
-  fiat_p384_addcarryx_u32(&x231, &x232, x230, x205, x208);
+  fiat_p384_addcarryx_u32(&x231, &x232, x230, x208, x205);
   uint32_t x233;
   fiat_p384_uint1 x234;
-  fiat_p384_addcarryx_u32(&x233, &x234, x232, x203, x206);
+  fiat_p384_addcarryx_u32(&x233, &x234, x232, x206, x203);
   uint32_t x235;
   fiat_p384_uint1 x236;
-  fiat_p384_addcarryx_u32(&x235, &x236, x234, x201, x204);
+  fiat_p384_addcarryx_u32(&x235, &x236, x234, x204, x201);
   uint32_t x237;
   fiat_p384_uint1 x238;
-  fiat_p384_addcarryx_u32(&x237, &x238, x236, 0x0, x202);
+  fiat_p384_addcarryx_u32(&x237, &x238, x236, x202, 0x0);
   uint32_t x239;
   fiat_p384_uint1 x240;
-  fiat_p384_addcarryx_u32(&x239, &x240, 0x0, x219, x175);
+  fiat_p384_addcarryx_u32(&x239, &x240, 0x0, x175, x219);
   uint32_t x241;
   fiat_p384_uint1 x242;
-  fiat_p384_addcarryx_u32(&x241, &x242, x240, x220, x177);
+  fiat_p384_addcarryx_u32(&x241, &x242, x240, x177, x220);
   uint32_t x243;
   fiat_p384_uint1 x244;
-  fiat_p384_addcarryx_u32(&x243, &x244, x242, 0x0, x179);
+  fiat_p384_addcarryx_u32(&x243, &x244, x242, x179, 0x0);
   uint32_t x245;
   fiat_p384_uint1 x246;
-  fiat_p384_addcarryx_u32(&x245, &x246, x244, x217, x181);
+  fiat_p384_addcarryx_u32(&x245, &x246, x244, x181, x217);
   uint32_t x247;
   fiat_p384_uint1 x248;
-  fiat_p384_addcarryx_u32(&x247, &x248, x246, x221, x183);
+  fiat_p384_addcarryx_u32(&x247, &x248, x246, x183, x221);
   uint32_t x249;
   fiat_p384_uint1 x250;
-  fiat_p384_addcarryx_u32(&x249, &x250, x248, x223, x185);
+  fiat_p384_addcarryx_u32(&x249, &x250, x248, x185, x223);
   uint32_t x251;
   fiat_p384_uint1 x252;
-  fiat_p384_addcarryx_u32(&x251, &x252, x250, x225, x187);
+  fiat_p384_addcarryx_u32(&x251, &x252, x250, x187, x225);
   uint32_t x253;
   fiat_p384_uint1 x254;
-  fiat_p384_addcarryx_u32(&x253, &x254, x252, x227, x189);
+  fiat_p384_addcarryx_u32(&x253, &x254, x252, x189, x227);
   uint32_t x255;
   fiat_p384_uint1 x256;
-  fiat_p384_addcarryx_u32(&x255, &x256, x254, x229, x191);
+  fiat_p384_addcarryx_u32(&x255, &x256, x254, x191, x229);
   uint32_t x257;
   fiat_p384_uint1 x258;
-  fiat_p384_addcarryx_u32(&x257, &x258, x256, x231, x193);
+  fiat_p384_addcarryx_u32(&x257, &x258, x256, x193, x231);
   uint32_t x259;
   fiat_p384_uint1 x260;
-  fiat_p384_addcarryx_u32(&x259, &x260, x258, x233, x195);
+  fiat_p384_addcarryx_u32(&x259, &x260, x258, x195, x233);
   uint32_t x261;
   fiat_p384_uint1 x262;
-  fiat_p384_addcarryx_u32(&x261, &x262, x260, x235, x197);
+  fiat_p384_addcarryx_u32(&x261, &x262, x260, x197, x235);
   uint32_t x263;
   fiat_p384_uint1 x264;
-  fiat_p384_addcarryx_u32(&x263, &x264, x262, x237, x199);
+  fiat_p384_addcarryx_u32(&x263, &x264, x262, x199, x237);
   uint32_t x265;
   fiat_p384_uint1 x266;
-  fiat_p384_addcarryx_u32(&x265, &x266, x264, 0x0, x200);
+  fiat_p384_addcarryx_u32(&x265, &x266, x264, x200, 0x0);
   uint32_t x267;
   uint32_t x268;
   fiat_p384_mulx_u32(&x267, &x268, x2, (arg2[11]));
@@ -547,79 +547,79 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x289, &x290, x2, (arg2[0]));
   uint32_t x291;
   fiat_p384_uint1 x292;
-  fiat_p384_addcarryx_u32(&x291, &x292, 0x0, x287, x290);
+  fiat_p384_addcarryx_u32(&x291, &x292, 0x0, x290, x287);
   uint32_t x293;
   fiat_p384_uint1 x294;
-  fiat_p384_addcarryx_u32(&x293, &x294, x292, x285, x288);
+  fiat_p384_addcarryx_u32(&x293, &x294, x292, x288, x285);
   uint32_t x295;
   fiat_p384_uint1 x296;
-  fiat_p384_addcarryx_u32(&x295, &x296, x294, x283, x286);
+  fiat_p384_addcarryx_u32(&x295, &x296, x294, x286, x283);
   uint32_t x297;
   fiat_p384_uint1 x298;
-  fiat_p384_addcarryx_u32(&x297, &x298, x296, x281, x284);
+  fiat_p384_addcarryx_u32(&x297, &x298, x296, x284, x281);
   uint32_t x299;
   fiat_p384_uint1 x300;
-  fiat_p384_addcarryx_u32(&x299, &x300, x298, x279, x282);
+  fiat_p384_addcarryx_u32(&x299, &x300, x298, x282, x279);
   uint32_t x301;
   fiat_p384_uint1 x302;
-  fiat_p384_addcarryx_u32(&x301, &x302, x300, x277, x280);
+  fiat_p384_addcarryx_u32(&x301, &x302, x300, x280, x277);
   uint32_t x303;
   fiat_p384_uint1 x304;
-  fiat_p384_addcarryx_u32(&x303, &x304, x302, x275, x278);
+  fiat_p384_addcarryx_u32(&x303, &x304, x302, x278, x275);
   uint32_t x305;
   fiat_p384_uint1 x306;
-  fiat_p384_addcarryx_u32(&x305, &x306, x304, x273, x276);
+  fiat_p384_addcarryx_u32(&x305, &x306, x304, x276, x273);
   uint32_t x307;
   fiat_p384_uint1 x308;
-  fiat_p384_addcarryx_u32(&x307, &x308, x306, x271, x274);
+  fiat_p384_addcarryx_u32(&x307, &x308, x306, x274, x271);
   uint32_t x309;
   fiat_p384_uint1 x310;
-  fiat_p384_addcarryx_u32(&x309, &x310, x308, x269, x272);
+  fiat_p384_addcarryx_u32(&x309, &x310, x308, x272, x269);
   uint32_t x311;
   fiat_p384_uint1 x312;
-  fiat_p384_addcarryx_u32(&x311, &x312, x310, x267, x270);
+  fiat_p384_addcarryx_u32(&x311, &x312, x310, x270, x267);
   uint32_t x313;
   fiat_p384_uint1 x314;
-  fiat_p384_addcarryx_u32(&x313, &x314, x312, 0x0, x268);
+  fiat_p384_addcarryx_u32(&x313, &x314, x312, x268, 0x0);
   uint32_t x315;
   fiat_p384_uint1 x316;
-  fiat_p384_addcarryx_u32(&x315, &x316, 0x0, x289, x241);
+  fiat_p384_addcarryx_u32(&x315, &x316, 0x0, x241, x289);
   uint32_t x317;
   fiat_p384_uint1 x318;
-  fiat_p384_addcarryx_u32(&x317, &x318, x316, x291, x243);
+  fiat_p384_addcarryx_u32(&x317, &x318, x316, x243, x291);
   uint32_t x319;
   fiat_p384_uint1 x320;
-  fiat_p384_addcarryx_u32(&x319, &x320, x318, x293, x245);
+  fiat_p384_addcarryx_u32(&x319, &x320, x318, x245, x293);
   uint32_t x321;
   fiat_p384_uint1 x322;
-  fiat_p384_addcarryx_u32(&x321, &x322, x320, x295, x247);
+  fiat_p384_addcarryx_u32(&x321, &x322, x320, x247, x295);
   uint32_t x323;
   fiat_p384_uint1 x324;
-  fiat_p384_addcarryx_u32(&x323, &x324, x322, x297, x249);
+  fiat_p384_addcarryx_u32(&x323, &x324, x322, x249, x297);
   uint32_t x325;
   fiat_p384_uint1 x326;
-  fiat_p384_addcarryx_u32(&x325, &x326, x324, x299, x251);
+  fiat_p384_addcarryx_u32(&x325, &x326, x324, x251, x299);
   uint32_t x327;
   fiat_p384_uint1 x328;
-  fiat_p384_addcarryx_u32(&x327, &x328, x326, x301, x253);
+  fiat_p384_addcarryx_u32(&x327, &x328, x326, x253, x301);
   uint32_t x329;
   fiat_p384_uint1 x330;
-  fiat_p384_addcarryx_u32(&x329, &x330, x328, x303, x255);
+  fiat_p384_addcarryx_u32(&x329, &x330, x328, x255, x303);
   uint32_t x331;
   fiat_p384_uint1 x332;
-  fiat_p384_addcarryx_u32(&x331, &x332, x330, x305, x257);
+  fiat_p384_addcarryx_u32(&x331, &x332, x330, x257, x305);
   uint32_t x333;
   fiat_p384_uint1 x334;
-  fiat_p384_addcarryx_u32(&x333, &x334, x332, x307, x259);
+  fiat_p384_addcarryx_u32(&x333, &x334, x332, x259, x307);
   uint32_t x335;
   fiat_p384_uint1 x336;
-  fiat_p384_addcarryx_u32(&x335, &x336, x334, x309, x261);
+  fiat_p384_addcarryx_u32(&x335, &x336, x334, x261, x309);
   uint32_t x337;
   fiat_p384_uint1 x338;
-  fiat_p384_addcarryx_u32(&x337, &x338, x336, x311, x263);
+  fiat_p384_addcarryx_u32(&x337, &x338, x336, x263, x311);
   uint32_t x339;
   fiat_p384_uint1 x340;
-  fiat_p384_addcarryx_u32(&x339, &x340, x338, x313, x265);
+  fiat_p384_addcarryx_u32(&x339, &x340, x338, x265, x313);
   uint32_t x341;
   uint32_t x342;
   fiat_p384_mulx_u32(&x341, &x342, x315, UINT32_C(0xffffffff));
@@ -652,73 +652,73 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x359, &x360, x315, UINT32_C(0xffffffff));
   uint32_t x361;
   fiat_p384_uint1 x362;
-  fiat_p384_addcarryx_u32(&x361, &x362, 0x0, x355, x358);
+  fiat_p384_addcarryx_u32(&x361, &x362, 0x0, x358, x355);
   uint32_t x363;
   fiat_p384_uint1 x364;
-  fiat_p384_addcarryx_u32(&x363, &x364, x362, x353, x356);
+  fiat_p384_addcarryx_u32(&x363, &x364, x362, x356, x353);
   uint32_t x365;
   fiat_p384_uint1 x366;
-  fiat_p384_addcarryx_u32(&x365, &x366, x364, x351, x354);
+  fiat_p384_addcarryx_u32(&x365, &x366, x364, x354, x351);
   uint32_t x367;
   fiat_p384_uint1 x368;
-  fiat_p384_addcarryx_u32(&x367, &x368, x366, x349, x352);
+  fiat_p384_addcarryx_u32(&x367, &x368, x366, x352, x349);
   uint32_t x369;
   fiat_p384_uint1 x370;
-  fiat_p384_addcarryx_u32(&x369, &x370, x368, x347, x350);
+  fiat_p384_addcarryx_u32(&x369, &x370, x368, x350, x347);
   uint32_t x371;
   fiat_p384_uint1 x372;
-  fiat_p384_addcarryx_u32(&x371, &x372, x370, x345, x348);
+  fiat_p384_addcarryx_u32(&x371, &x372, x370, x348, x345);
   uint32_t x373;
   fiat_p384_uint1 x374;
-  fiat_p384_addcarryx_u32(&x373, &x374, x372, x343, x346);
+  fiat_p384_addcarryx_u32(&x373, &x374, x372, x346, x343);
   uint32_t x375;
   fiat_p384_uint1 x376;
-  fiat_p384_addcarryx_u32(&x375, &x376, x374, x341, x344);
+  fiat_p384_addcarryx_u32(&x375, &x376, x374, x344, x341);
   uint32_t x377;
   fiat_p384_uint1 x378;
-  fiat_p384_addcarryx_u32(&x377, &x378, x376, 0x0, x342);
+  fiat_p384_addcarryx_u32(&x377, &x378, x376, x342, 0x0);
   uint32_t x379;
   fiat_p384_uint1 x380;
-  fiat_p384_addcarryx_u32(&x379, &x380, 0x0, x359, x315);
+  fiat_p384_addcarryx_u32(&x379, &x380, 0x0, x315, x359);
   uint32_t x381;
   fiat_p384_uint1 x382;
-  fiat_p384_addcarryx_u32(&x381, &x382, x380, x360, x317);
+  fiat_p384_addcarryx_u32(&x381, &x382, x380, x317, x360);
   uint32_t x383;
   fiat_p384_uint1 x384;
-  fiat_p384_addcarryx_u32(&x383, &x384, x382, 0x0, x319);
+  fiat_p384_addcarryx_u32(&x383, &x384, x382, x319, 0x0);
   uint32_t x385;
   fiat_p384_uint1 x386;
-  fiat_p384_addcarryx_u32(&x385, &x386, x384, x357, x321);
+  fiat_p384_addcarryx_u32(&x385, &x386, x384, x321, x357);
   uint32_t x387;
   fiat_p384_uint1 x388;
-  fiat_p384_addcarryx_u32(&x387, &x388, x386, x361, x323);
+  fiat_p384_addcarryx_u32(&x387, &x388, x386, x323, x361);
   uint32_t x389;
   fiat_p384_uint1 x390;
-  fiat_p384_addcarryx_u32(&x389, &x390, x388, x363, x325);
+  fiat_p384_addcarryx_u32(&x389, &x390, x388, x325, x363);
   uint32_t x391;
   fiat_p384_uint1 x392;
-  fiat_p384_addcarryx_u32(&x391, &x392, x390, x365, x327);
+  fiat_p384_addcarryx_u32(&x391, &x392, x390, x327, x365);
   uint32_t x393;
   fiat_p384_uint1 x394;
-  fiat_p384_addcarryx_u32(&x393, &x394, x392, x367, x329);
+  fiat_p384_addcarryx_u32(&x393, &x394, x392, x329, x367);
   uint32_t x395;
   fiat_p384_uint1 x396;
-  fiat_p384_addcarryx_u32(&x395, &x396, x394, x369, x331);
+  fiat_p384_addcarryx_u32(&x395, &x396, x394, x331, x369);
   uint32_t x397;
   fiat_p384_uint1 x398;
-  fiat_p384_addcarryx_u32(&x397, &x398, x396, x371, x333);
+  fiat_p384_addcarryx_u32(&x397, &x398, x396, x333, x371);
   uint32_t x399;
   fiat_p384_uint1 x400;
-  fiat_p384_addcarryx_u32(&x399, &x400, x398, x373, x335);
+  fiat_p384_addcarryx_u32(&x399, &x400, x398, x335, x373);
   uint32_t x401;
   fiat_p384_uint1 x402;
-  fiat_p384_addcarryx_u32(&x401, &x402, x400, x375, x337);
+  fiat_p384_addcarryx_u32(&x401, &x402, x400, x337, x375);
   uint32_t x403;
   fiat_p384_uint1 x404;
-  fiat_p384_addcarryx_u32(&x403, &x404, x402, x377, x339);
+  fiat_p384_addcarryx_u32(&x403, &x404, x402, x339, x377);
   uint32_t x405;
   fiat_p384_uint1 x406;
-  fiat_p384_addcarryx_u32(&x405, &x406, x404, 0x0, x340);
+  fiat_p384_addcarryx_u32(&x405, &x406, x404, x340, 0x0);
   uint32_t x407;
   uint32_t x408;
   fiat_p384_mulx_u32(&x407, &x408, x3, (arg2[11]));
@@ -757,79 +757,79 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x429, &x430, x3, (arg2[0]));
   uint32_t x431;
   fiat_p384_uint1 x432;
-  fiat_p384_addcarryx_u32(&x431, &x432, 0x0, x427, x430);
+  fiat_p384_addcarryx_u32(&x431, &x432, 0x0, x430, x427);
   uint32_t x433;
   fiat_p384_uint1 x434;
-  fiat_p384_addcarryx_u32(&x433, &x434, x432, x425, x428);
+  fiat_p384_addcarryx_u32(&x433, &x434, x432, x428, x425);
   uint32_t x435;
   fiat_p384_uint1 x436;
-  fiat_p384_addcarryx_u32(&x435, &x436, x434, x423, x426);
+  fiat_p384_addcarryx_u32(&x435, &x436, x434, x426, x423);
   uint32_t x437;
   fiat_p384_uint1 x438;
-  fiat_p384_addcarryx_u32(&x437, &x438, x436, x421, x424);
+  fiat_p384_addcarryx_u32(&x437, &x438, x436, x424, x421);
   uint32_t x439;
   fiat_p384_uint1 x440;
-  fiat_p384_addcarryx_u32(&x439, &x440, x438, x419, x422);
+  fiat_p384_addcarryx_u32(&x439, &x440, x438, x422, x419);
   uint32_t x441;
   fiat_p384_uint1 x442;
-  fiat_p384_addcarryx_u32(&x441, &x442, x440, x417, x420);
+  fiat_p384_addcarryx_u32(&x441, &x442, x440, x420, x417);
   uint32_t x443;
   fiat_p384_uint1 x444;
-  fiat_p384_addcarryx_u32(&x443, &x444, x442, x415, x418);
+  fiat_p384_addcarryx_u32(&x443, &x444, x442, x418, x415);
   uint32_t x445;
   fiat_p384_uint1 x446;
-  fiat_p384_addcarryx_u32(&x445, &x446, x444, x413, x416);
+  fiat_p384_addcarryx_u32(&x445, &x446, x444, x416, x413);
   uint32_t x447;
   fiat_p384_uint1 x448;
-  fiat_p384_addcarryx_u32(&x447, &x448, x446, x411, x414);
+  fiat_p384_addcarryx_u32(&x447, &x448, x446, x414, x411);
   uint32_t x449;
   fiat_p384_uint1 x450;
-  fiat_p384_addcarryx_u32(&x449, &x450, x448, x409, x412);
+  fiat_p384_addcarryx_u32(&x449, &x450, x448, x412, x409);
   uint32_t x451;
   fiat_p384_uint1 x452;
-  fiat_p384_addcarryx_u32(&x451, &x452, x450, x407, x410);
+  fiat_p384_addcarryx_u32(&x451, &x452, x450, x410, x407);
   uint32_t x453;
   fiat_p384_uint1 x454;
-  fiat_p384_addcarryx_u32(&x453, &x454, x452, 0x0, x408);
+  fiat_p384_addcarryx_u32(&x453, &x454, x452, x408, 0x0);
   uint32_t x455;
   fiat_p384_uint1 x456;
-  fiat_p384_addcarryx_u32(&x455, &x456, 0x0, x429, x381);
+  fiat_p384_addcarryx_u32(&x455, &x456, 0x0, x381, x429);
   uint32_t x457;
   fiat_p384_uint1 x458;
-  fiat_p384_addcarryx_u32(&x457, &x458, x456, x431, x383);
+  fiat_p384_addcarryx_u32(&x457, &x458, x456, x383, x431);
   uint32_t x459;
   fiat_p384_uint1 x460;
-  fiat_p384_addcarryx_u32(&x459, &x460, x458, x433, x385);
+  fiat_p384_addcarryx_u32(&x459, &x460, x458, x385, x433);
   uint32_t x461;
   fiat_p384_uint1 x462;
-  fiat_p384_addcarryx_u32(&x461, &x462, x460, x435, x387);
+  fiat_p384_addcarryx_u32(&x461, &x462, x460, x387, x435);
   uint32_t x463;
   fiat_p384_uint1 x464;
-  fiat_p384_addcarryx_u32(&x463, &x464, x462, x437, x389);
+  fiat_p384_addcarryx_u32(&x463, &x464, x462, x389, x437);
   uint32_t x465;
   fiat_p384_uint1 x466;
-  fiat_p384_addcarryx_u32(&x465, &x466, x464, x439, x391);
+  fiat_p384_addcarryx_u32(&x465, &x466, x464, x391, x439);
   uint32_t x467;
   fiat_p384_uint1 x468;
-  fiat_p384_addcarryx_u32(&x467, &x468, x466, x441, x393);
+  fiat_p384_addcarryx_u32(&x467, &x468, x466, x393, x441);
   uint32_t x469;
   fiat_p384_uint1 x470;
-  fiat_p384_addcarryx_u32(&x469, &x470, x468, x443, x395);
+  fiat_p384_addcarryx_u32(&x469, &x470, x468, x395, x443);
   uint32_t x471;
   fiat_p384_uint1 x472;
-  fiat_p384_addcarryx_u32(&x471, &x472, x470, x445, x397);
+  fiat_p384_addcarryx_u32(&x471, &x472, x470, x397, x445);
   uint32_t x473;
   fiat_p384_uint1 x474;
-  fiat_p384_addcarryx_u32(&x473, &x474, x472, x447, x399);
+  fiat_p384_addcarryx_u32(&x473, &x474, x472, x399, x447);
   uint32_t x475;
   fiat_p384_uint1 x476;
-  fiat_p384_addcarryx_u32(&x475, &x476, x474, x449, x401);
+  fiat_p384_addcarryx_u32(&x475, &x476, x474, x401, x449);
   uint32_t x477;
   fiat_p384_uint1 x478;
-  fiat_p384_addcarryx_u32(&x477, &x478, x476, x451, x403);
+  fiat_p384_addcarryx_u32(&x477, &x478, x476, x403, x451);
   uint32_t x479;
   fiat_p384_uint1 x480;
-  fiat_p384_addcarryx_u32(&x479, &x480, x478, x453, x405);
+  fiat_p384_addcarryx_u32(&x479, &x480, x478, x405, x453);
   uint32_t x481;
   uint32_t x482;
   fiat_p384_mulx_u32(&x481, &x482, x455, UINT32_C(0xffffffff));
@@ -862,73 +862,73 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x499, &x500, x455, UINT32_C(0xffffffff));
   uint32_t x501;
   fiat_p384_uint1 x502;
-  fiat_p384_addcarryx_u32(&x501, &x502, 0x0, x495, x498);
+  fiat_p384_addcarryx_u32(&x501, &x502, 0x0, x498, x495);
   uint32_t x503;
   fiat_p384_uint1 x504;
-  fiat_p384_addcarryx_u32(&x503, &x504, x502, x493, x496);
+  fiat_p384_addcarryx_u32(&x503, &x504, x502, x496, x493);
   uint32_t x505;
   fiat_p384_uint1 x506;
-  fiat_p384_addcarryx_u32(&x505, &x506, x504, x491, x494);
+  fiat_p384_addcarryx_u32(&x505, &x506, x504, x494, x491);
   uint32_t x507;
   fiat_p384_uint1 x508;
-  fiat_p384_addcarryx_u32(&x507, &x508, x506, x489, x492);
+  fiat_p384_addcarryx_u32(&x507, &x508, x506, x492, x489);
   uint32_t x509;
   fiat_p384_uint1 x510;
-  fiat_p384_addcarryx_u32(&x509, &x510, x508, x487, x490);
+  fiat_p384_addcarryx_u32(&x509, &x510, x508, x490, x487);
   uint32_t x511;
   fiat_p384_uint1 x512;
-  fiat_p384_addcarryx_u32(&x511, &x512, x510, x485, x488);
+  fiat_p384_addcarryx_u32(&x511, &x512, x510, x488, x485);
   uint32_t x513;
   fiat_p384_uint1 x514;
-  fiat_p384_addcarryx_u32(&x513, &x514, x512, x483, x486);
+  fiat_p384_addcarryx_u32(&x513, &x514, x512, x486, x483);
   uint32_t x515;
   fiat_p384_uint1 x516;
-  fiat_p384_addcarryx_u32(&x515, &x516, x514, x481, x484);
+  fiat_p384_addcarryx_u32(&x515, &x516, x514, x484, x481);
   uint32_t x517;
   fiat_p384_uint1 x518;
-  fiat_p384_addcarryx_u32(&x517, &x518, x516, 0x0, x482);
+  fiat_p384_addcarryx_u32(&x517, &x518, x516, x482, 0x0);
   uint32_t x519;
   fiat_p384_uint1 x520;
-  fiat_p384_addcarryx_u32(&x519, &x520, 0x0, x499, x455);
+  fiat_p384_addcarryx_u32(&x519, &x520, 0x0, x455, x499);
   uint32_t x521;
   fiat_p384_uint1 x522;
-  fiat_p384_addcarryx_u32(&x521, &x522, x520, x500, x457);
+  fiat_p384_addcarryx_u32(&x521, &x522, x520, x457, x500);
   uint32_t x523;
   fiat_p384_uint1 x524;
-  fiat_p384_addcarryx_u32(&x523, &x524, x522, 0x0, x459);
+  fiat_p384_addcarryx_u32(&x523, &x524, x522, x459, 0x0);
   uint32_t x525;
   fiat_p384_uint1 x526;
-  fiat_p384_addcarryx_u32(&x525, &x526, x524, x497, x461);
+  fiat_p384_addcarryx_u32(&x525, &x526, x524, x461, x497);
   uint32_t x527;
   fiat_p384_uint1 x528;
-  fiat_p384_addcarryx_u32(&x527, &x528, x526, x501, x463);
+  fiat_p384_addcarryx_u32(&x527, &x528, x526, x463, x501);
   uint32_t x529;
   fiat_p384_uint1 x530;
-  fiat_p384_addcarryx_u32(&x529, &x530, x528, x503, x465);
+  fiat_p384_addcarryx_u32(&x529, &x530, x528, x465, x503);
   uint32_t x531;
   fiat_p384_uint1 x532;
-  fiat_p384_addcarryx_u32(&x531, &x532, x530, x505, x467);
+  fiat_p384_addcarryx_u32(&x531, &x532, x530, x467, x505);
   uint32_t x533;
   fiat_p384_uint1 x534;
-  fiat_p384_addcarryx_u32(&x533, &x534, x532, x507, x469);
+  fiat_p384_addcarryx_u32(&x533, &x534, x532, x469, x507);
   uint32_t x535;
   fiat_p384_uint1 x536;
-  fiat_p384_addcarryx_u32(&x535, &x536, x534, x509, x471);
+  fiat_p384_addcarryx_u32(&x535, &x536, x534, x471, x509);
   uint32_t x537;
   fiat_p384_uint1 x538;
-  fiat_p384_addcarryx_u32(&x537, &x538, x536, x511, x473);
+  fiat_p384_addcarryx_u32(&x537, &x538, x536, x473, x511);
   uint32_t x539;
   fiat_p384_uint1 x540;
-  fiat_p384_addcarryx_u32(&x539, &x540, x538, x513, x475);
+  fiat_p384_addcarryx_u32(&x539, &x540, x538, x475, x513);
   uint32_t x541;
   fiat_p384_uint1 x542;
-  fiat_p384_addcarryx_u32(&x541, &x542, x540, x515, x477);
+  fiat_p384_addcarryx_u32(&x541, &x542, x540, x477, x515);
   uint32_t x543;
   fiat_p384_uint1 x544;
-  fiat_p384_addcarryx_u32(&x543, &x544, x542, x517, x479);
+  fiat_p384_addcarryx_u32(&x543, &x544, x542, x479, x517);
   uint32_t x545;
   fiat_p384_uint1 x546;
-  fiat_p384_addcarryx_u32(&x545, &x546, x544, 0x0, x480);
+  fiat_p384_addcarryx_u32(&x545, &x546, x544, x480, 0x0);
   uint32_t x547;
   uint32_t x548;
   fiat_p384_mulx_u32(&x547, &x548, x4, (arg2[11]));
@@ -967,79 +967,79 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x569, &x570, x4, (arg2[0]));
   uint32_t x571;
   fiat_p384_uint1 x572;
-  fiat_p384_addcarryx_u32(&x571, &x572, 0x0, x567, x570);
+  fiat_p384_addcarryx_u32(&x571, &x572, 0x0, x570, x567);
   uint32_t x573;
   fiat_p384_uint1 x574;
-  fiat_p384_addcarryx_u32(&x573, &x574, x572, x565, x568);
+  fiat_p384_addcarryx_u32(&x573, &x574, x572, x568, x565);
   uint32_t x575;
   fiat_p384_uint1 x576;
-  fiat_p384_addcarryx_u32(&x575, &x576, x574, x563, x566);
+  fiat_p384_addcarryx_u32(&x575, &x576, x574, x566, x563);
   uint32_t x577;
   fiat_p384_uint1 x578;
-  fiat_p384_addcarryx_u32(&x577, &x578, x576, x561, x564);
+  fiat_p384_addcarryx_u32(&x577, &x578, x576, x564, x561);
   uint32_t x579;
   fiat_p384_uint1 x580;
-  fiat_p384_addcarryx_u32(&x579, &x580, x578, x559, x562);
+  fiat_p384_addcarryx_u32(&x579, &x580, x578, x562, x559);
   uint32_t x581;
   fiat_p384_uint1 x582;
-  fiat_p384_addcarryx_u32(&x581, &x582, x580, x557, x560);
+  fiat_p384_addcarryx_u32(&x581, &x582, x580, x560, x557);
   uint32_t x583;
   fiat_p384_uint1 x584;
-  fiat_p384_addcarryx_u32(&x583, &x584, x582, x555, x558);
+  fiat_p384_addcarryx_u32(&x583, &x584, x582, x558, x555);
   uint32_t x585;
   fiat_p384_uint1 x586;
-  fiat_p384_addcarryx_u32(&x585, &x586, x584, x553, x556);
+  fiat_p384_addcarryx_u32(&x585, &x586, x584, x556, x553);
   uint32_t x587;
   fiat_p384_uint1 x588;
-  fiat_p384_addcarryx_u32(&x587, &x588, x586, x551, x554);
+  fiat_p384_addcarryx_u32(&x587, &x588, x586, x554, x551);
   uint32_t x589;
   fiat_p384_uint1 x590;
-  fiat_p384_addcarryx_u32(&x589, &x590, x588, x549, x552);
+  fiat_p384_addcarryx_u32(&x589, &x590, x588, x552, x549);
   uint32_t x591;
   fiat_p384_uint1 x592;
-  fiat_p384_addcarryx_u32(&x591, &x592, x590, x547, x550);
+  fiat_p384_addcarryx_u32(&x591, &x592, x590, x550, x547);
   uint32_t x593;
   fiat_p384_uint1 x594;
-  fiat_p384_addcarryx_u32(&x593, &x594, x592, 0x0, x548);
+  fiat_p384_addcarryx_u32(&x593, &x594, x592, x548, 0x0);
   uint32_t x595;
   fiat_p384_uint1 x596;
-  fiat_p384_addcarryx_u32(&x595, &x596, 0x0, x569, x521);
+  fiat_p384_addcarryx_u32(&x595, &x596, 0x0, x521, x569);
   uint32_t x597;
   fiat_p384_uint1 x598;
-  fiat_p384_addcarryx_u32(&x597, &x598, x596, x571, x523);
+  fiat_p384_addcarryx_u32(&x597, &x598, x596, x523, x571);
   uint32_t x599;
   fiat_p384_uint1 x600;
-  fiat_p384_addcarryx_u32(&x599, &x600, x598, x573, x525);
+  fiat_p384_addcarryx_u32(&x599, &x600, x598, x525, x573);
   uint32_t x601;
   fiat_p384_uint1 x602;
-  fiat_p384_addcarryx_u32(&x601, &x602, x600, x575, x527);
+  fiat_p384_addcarryx_u32(&x601, &x602, x600, x527, x575);
   uint32_t x603;
   fiat_p384_uint1 x604;
-  fiat_p384_addcarryx_u32(&x603, &x604, x602, x577, x529);
+  fiat_p384_addcarryx_u32(&x603, &x604, x602, x529, x577);
   uint32_t x605;
   fiat_p384_uint1 x606;
-  fiat_p384_addcarryx_u32(&x605, &x606, x604, x579, x531);
+  fiat_p384_addcarryx_u32(&x605, &x606, x604, x531, x579);
   uint32_t x607;
   fiat_p384_uint1 x608;
-  fiat_p384_addcarryx_u32(&x607, &x608, x606, x581, x533);
+  fiat_p384_addcarryx_u32(&x607, &x608, x606, x533, x581);
   uint32_t x609;
   fiat_p384_uint1 x610;
-  fiat_p384_addcarryx_u32(&x609, &x610, x608, x583, x535);
+  fiat_p384_addcarryx_u32(&x609, &x610, x608, x535, x583);
   uint32_t x611;
   fiat_p384_uint1 x612;
-  fiat_p384_addcarryx_u32(&x611, &x612, x610, x585, x537);
+  fiat_p384_addcarryx_u32(&x611, &x612, x610, x537, x585);
   uint32_t x613;
   fiat_p384_uint1 x614;
-  fiat_p384_addcarryx_u32(&x613, &x614, x612, x587, x539);
+  fiat_p384_addcarryx_u32(&x613, &x614, x612, x539, x587);
   uint32_t x615;
   fiat_p384_uint1 x616;
-  fiat_p384_addcarryx_u32(&x615, &x616, x614, x589, x541);
+  fiat_p384_addcarryx_u32(&x615, &x616, x614, x541, x589);
   uint32_t x617;
   fiat_p384_uint1 x618;
-  fiat_p384_addcarryx_u32(&x617, &x618, x616, x591, x543);
+  fiat_p384_addcarryx_u32(&x617, &x618, x616, x543, x591);
   uint32_t x619;
   fiat_p384_uint1 x620;
-  fiat_p384_addcarryx_u32(&x619, &x620, x618, x593, x545);
+  fiat_p384_addcarryx_u32(&x619, &x620, x618, x545, x593);
   uint32_t x621;
   uint32_t x622;
   fiat_p384_mulx_u32(&x621, &x622, x595, UINT32_C(0xffffffff));
@@ -1072,73 +1072,73 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x639, &x640, x595, UINT32_C(0xffffffff));
   uint32_t x641;
   fiat_p384_uint1 x642;
-  fiat_p384_addcarryx_u32(&x641, &x642, 0x0, x635, x638);
+  fiat_p384_addcarryx_u32(&x641, &x642, 0x0, x638, x635);
   uint32_t x643;
   fiat_p384_uint1 x644;
-  fiat_p384_addcarryx_u32(&x643, &x644, x642, x633, x636);
+  fiat_p384_addcarryx_u32(&x643, &x644, x642, x636, x633);
   uint32_t x645;
   fiat_p384_uint1 x646;
-  fiat_p384_addcarryx_u32(&x645, &x646, x644, x631, x634);
+  fiat_p384_addcarryx_u32(&x645, &x646, x644, x634, x631);
   uint32_t x647;
   fiat_p384_uint1 x648;
-  fiat_p384_addcarryx_u32(&x647, &x648, x646, x629, x632);
+  fiat_p384_addcarryx_u32(&x647, &x648, x646, x632, x629);
   uint32_t x649;
   fiat_p384_uint1 x650;
-  fiat_p384_addcarryx_u32(&x649, &x650, x648, x627, x630);
+  fiat_p384_addcarryx_u32(&x649, &x650, x648, x630, x627);
   uint32_t x651;
   fiat_p384_uint1 x652;
-  fiat_p384_addcarryx_u32(&x651, &x652, x650, x625, x628);
+  fiat_p384_addcarryx_u32(&x651, &x652, x650, x628, x625);
   uint32_t x653;
   fiat_p384_uint1 x654;
-  fiat_p384_addcarryx_u32(&x653, &x654, x652, x623, x626);
+  fiat_p384_addcarryx_u32(&x653, &x654, x652, x626, x623);
   uint32_t x655;
   fiat_p384_uint1 x656;
-  fiat_p384_addcarryx_u32(&x655, &x656, x654, x621, x624);
+  fiat_p384_addcarryx_u32(&x655, &x656, x654, x624, x621);
   uint32_t x657;
   fiat_p384_uint1 x658;
-  fiat_p384_addcarryx_u32(&x657, &x658, x656, 0x0, x622);
+  fiat_p384_addcarryx_u32(&x657, &x658, x656, x622, 0x0);
   uint32_t x659;
   fiat_p384_uint1 x660;
-  fiat_p384_addcarryx_u32(&x659, &x660, 0x0, x639, x595);
+  fiat_p384_addcarryx_u32(&x659, &x660, 0x0, x595, x639);
   uint32_t x661;
   fiat_p384_uint1 x662;
-  fiat_p384_addcarryx_u32(&x661, &x662, x660, x640, x597);
+  fiat_p384_addcarryx_u32(&x661, &x662, x660, x597, x640);
   uint32_t x663;
   fiat_p384_uint1 x664;
-  fiat_p384_addcarryx_u32(&x663, &x664, x662, 0x0, x599);
+  fiat_p384_addcarryx_u32(&x663, &x664, x662, x599, 0x0);
   uint32_t x665;
   fiat_p384_uint1 x666;
-  fiat_p384_addcarryx_u32(&x665, &x666, x664, x637, x601);
+  fiat_p384_addcarryx_u32(&x665, &x666, x664, x601, x637);
   uint32_t x667;
   fiat_p384_uint1 x668;
-  fiat_p384_addcarryx_u32(&x667, &x668, x666, x641, x603);
+  fiat_p384_addcarryx_u32(&x667, &x668, x666, x603, x641);
   uint32_t x669;
   fiat_p384_uint1 x670;
-  fiat_p384_addcarryx_u32(&x669, &x670, x668, x643, x605);
+  fiat_p384_addcarryx_u32(&x669, &x670, x668, x605, x643);
   uint32_t x671;
   fiat_p384_uint1 x672;
-  fiat_p384_addcarryx_u32(&x671, &x672, x670, x645, x607);
+  fiat_p384_addcarryx_u32(&x671, &x672, x670, x607, x645);
   uint32_t x673;
   fiat_p384_uint1 x674;
-  fiat_p384_addcarryx_u32(&x673, &x674, x672, x647, x609);
+  fiat_p384_addcarryx_u32(&x673, &x674, x672, x609, x647);
   uint32_t x675;
   fiat_p384_uint1 x676;
-  fiat_p384_addcarryx_u32(&x675, &x676, x674, x649, x611);
+  fiat_p384_addcarryx_u32(&x675, &x676, x674, x611, x649);
   uint32_t x677;
   fiat_p384_uint1 x678;
-  fiat_p384_addcarryx_u32(&x677, &x678, x676, x651, x613);
+  fiat_p384_addcarryx_u32(&x677, &x678, x676, x613, x651);
   uint32_t x679;
   fiat_p384_uint1 x680;
-  fiat_p384_addcarryx_u32(&x679, &x680, x678, x653, x615);
+  fiat_p384_addcarryx_u32(&x679, &x680, x678, x615, x653);
   uint32_t x681;
   fiat_p384_uint1 x682;
-  fiat_p384_addcarryx_u32(&x681, &x682, x680, x655, x617);
+  fiat_p384_addcarryx_u32(&x681, &x682, x680, x617, x655);
   uint32_t x683;
   fiat_p384_uint1 x684;
-  fiat_p384_addcarryx_u32(&x683, &x684, x682, x657, x619);
+  fiat_p384_addcarryx_u32(&x683, &x684, x682, x619, x657);
   uint32_t x685;
   fiat_p384_uint1 x686;
-  fiat_p384_addcarryx_u32(&x685, &x686, x684, 0x0, x620);
+  fiat_p384_addcarryx_u32(&x685, &x686, x684, x620, 0x0);
   uint32_t x687;
   uint32_t x688;
   fiat_p384_mulx_u32(&x687, &x688, x5, (arg2[11]));
@@ -1177,79 +1177,79 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x709, &x710, x5, (arg2[0]));
   uint32_t x711;
   fiat_p384_uint1 x712;
-  fiat_p384_addcarryx_u32(&x711, &x712, 0x0, x707, x710);
+  fiat_p384_addcarryx_u32(&x711, &x712, 0x0, x710, x707);
   uint32_t x713;
   fiat_p384_uint1 x714;
-  fiat_p384_addcarryx_u32(&x713, &x714, x712, x705, x708);
+  fiat_p384_addcarryx_u32(&x713, &x714, x712, x708, x705);
   uint32_t x715;
   fiat_p384_uint1 x716;
-  fiat_p384_addcarryx_u32(&x715, &x716, x714, x703, x706);
+  fiat_p384_addcarryx_u32(&x715, &x716, x714, x706, x703);
   uint32_t x717;
   fiat_p384_uint1 x718;
-  fiat_p384_addcarryx_u32(&x717, &x718, x716, x701, x704);
+  fiat_p384_addcarryx_u32(&x717, &x718, x716, x704, x701);
   uint32_t x719;
   fiat_p384_uint1 x720;
-  fiat_p384_addcarryx_u32(&x719, &x720, x718, x699, x702);
+  fiat_p384_addcarryx_u32(&x719, &x720, x718, x702, x699);
   uint32_t x721;
   fiat_p384_uint1 x722;
-  fiat_p384_addcarryx_u32(&x721, &x722, x720, x697, x700);
+  fiat_p384_addcarryx_u32(&x721, &x722, x720, x700, x697);
   uint32_t x723;
   fiat_p384_uint1 x724;
-  fiat_p384_addcarryx_u32(&x723, &x724, x722, x695, x698);
+  fiat_p384_addcarryx_u32(&x723, &x724, x722, x698, x695);
   uint32_t x725;
   fiat_p384_uint1 x726;
-  fiat_p384_addcarryx_u32(&x725, &x726, x724, x693, x696);
+  fiat_p384_addcarryx_u32(&x725, &x726, x724, x696, x693);
   uint32_t x727;
   fiat_p384_uint1 x728;
-  fiat_p384_addcarryx_u32(&x727, &x728, x726, x691, x694);
+  fiat_p384_addcarryx_u32(&x727, &x728, x726, x694, x691);
   uint32_t x729;
   fiat_p384_uint1 x730;
-  fiat_p384_addcarryx_u32(&x729, &x730, x728, x689, x692);
+  fiat_p384_addcarryx_u32(&x729, &x730, x728, x692, x689);
   uint32_t x731;
   fiat_p384_uint1 x732;
-  fiat_p384_addcarryx_u32(&x731, &x732, x730, x687, x690);
+  fiat_p384_addcarryx_u32(&x731, &x732, x730, x690, x687);
   uint32_t x733;
   fiat_p384_uint1 x734;
-  fiat_p384_addcarryx_u32(&x733, &x734, x732, 0x0, x688);
+  fiat_p384_addcarryx_u32(&x733, &x734, x732, x688, 0x0);
   uint32_t x735;
   fiat_p384_uint1 x736;
-  fiat_p384_addcarryx_u32(&x735, &x736, 0x0, x709, x661);
+  fiat_p384_addcarryx_u32(&x735, &x736, 0x0, x661, x709);
   uint32_t x737;
   fiat_p384_uint1 x738;
-  fiat_p384_addcarryx_u32(&x737, &x738, x736, x711, x663);
+  fiat_p384_addcarryx_u32(&x737, &x738, x736, x663, x711);
   uint32_t x739;
   fiat_p384_uint1 x740;
-  fiat_p384_addcarryx_u32(&x739, &x740, x738, x713, x665);
+  fiat_p384_addcarryx_u32(&x739, &x740, x738, x665, x713);
   uint32_t x741;
   fiat_p384_uint1 x742;
-  fiat_p384_addcarryx_u32(&x741, &x742, x740, x715, x667);
+  fiat_p384_addcarryx_u32(&x741, &x742, x740, x667, x715);
   uint32_t x743;
   fiat_p384_uint1 x744;
-  fiat_p384_addcarryx_u32(&x743, &x744, x742, x717, x669);
+  fiat_p384_addcarryx_u32(&x743, &x744, x742, x669, x717);
   uint32_t x745;
   fiat_p384_uint1 x746;
-  fiat_p384_addcarryx_u32(&x745, &x746, x744, x719, x671);
+  fiat_p384_addcarryx_u32(&x745, &x746, x744, x671, x719);
   uint32_t x747;
   fiat_p384_uint1 x748;
-  fiat_p384_addcarryx_u32(&x747, &x748, x746, x721, x673);
+  fiat_p384_addcarryx_u32(&x747, &x748, x746, x673, x721);
   uint32_t x749;
   fiat_p384_uint1 x750;
-  fiat_p384_addcarryx_u32(&x749, &x750, x748, x723, x675);
+  fiat_p384_addcarryx_u32(&x749, &x750, x748, x675, x723);
   uint32_t x751;
   fiat_p384_uint1 x752;
-  fiat_p384_addcarryx_u32(&x751, &x752, x750, x725, x677);
+  fiat_p384_addcarryx_u32(&x751, &x752, x750, x677, x725);
   uint32_t x753;
   fiat_p384_uint1 x754;
-  fiat_p384_addcarryx_u32(&x753, &x754, x752, x727, x679);
+  fiat_p384_addcarryx_u32(&x753, &x754, x752, x679, x727);
   uint32_t x755;
   fiat_p384_uint1 x756;
-  fiat_p384_addcarryx_u32(&x755, &x756, x754, x729, x681);
+  fiat_p384_addcarryx_u32(&x755, &x756, x754, x681, x729);
   uint32_t x757;
   fiat_p384_uint1 x758;
-  fiat_p384_addcarryx_u32(&x757, &x758, x756, x731, x683);
+  fiat_p384_addcarryx_u32(&x757, &x758, x756, x683, x731);
   uint32_t x759;
   fiat_p384_uint1 x760;
-  fiat_p384_addcarryx_u32(&x759, &x760, x758, x733, x685);
+  fiat_p384_addcarryx_u32(&x759, &x760, x758, x685, x733);
   uint32_t x761;
   uint32_t x762;
   fiat_p384_mulx_u32(&x761, &x762, x735, UINT32_C(0xffffffff));
@@ -1282,73 +1282,73 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x779, &x780, x735, UINT32_C(0xffffffff));
   uint32_t x781;
   fiat_p384_uint1 x782;
-  fiat_p384_addcarryx_u32(&x781, &x782, 0x0, x775, x778);
+  fiat_p384_addcarryx_u32(&x781, &x782, 0x0, x778, x775);
   uint32_t x783;
   fiat_p384_uint1 x784;
-  fiat_p384_addcarryx_u32(&x783, &x784, x782, x773, x776);
+  fiat_p384_addcarryx_u32(&x783, &x784, x782, x776, x773);
   uint32_t x785;
   fiat_p384_uint1 x786;
-  fiat_p384_addcarryx_u32(&x785, &x786, x784, x771, x774);
+  fiat_p384_addcarryx_u32(&x785, &x786, x784, x774, x771);
   uint32_t x787;
   fiat_p384_uint1 x788;
-  fiat_p384_addcarryx_u32(&x787, &x788, x786, x769, x772);
+  fiat_p384_addcarryx_u32(&x787, &x788, x786, x772, x769);
   uint32_t x789;
   fiat_p384_uint1 x790;
-  fiat_p384_addcarryx_u32(&x789, &x790, x788, x767, x770);
+  fiat_p384_addcarryx_u32(&x789, &x790, x788, x770, x767);
   uint32_t x791;
   fiat_p384_uint1 x792;
-  fiat_p384_addcarryx_u32(&x791, &x792, x790, x765, x768);
+  fiat_p384_addcarryx_u32(&x791, &x792, x790, x768, x765);
   uint32_t x793;
   fiat_p384_uint1 x794;
-  fiat_p384_addcarryx_u32(&x793, &x794, x792, x763, x766);
+  fiat_p384_addcarryx_u32(&x793, &x794, x792, x766, x763);
   uint32_t x795;
   fiat_p384_uint1 x796;
-  fiat_p384_addcarryx_u32(&x795, &x796, x794, x761, x764);
+  fiat_p384_addcarryx_u32(&x795, &x796, x794, x764, x761);
   uint32_t x797;
   fiat_p384_uint1 x798;
-  fiat_p384_addcarryx_u32(&x797, &x798, x796, 0x0, x762);
+  fiat_p384_addcarryx_u32(&x797, &x798, x796, x762, 0x0);
   uint32_t x799;
   fiat_p384_uint1 x800;
-  fiat_p384_addcarryx_u32(&x799, &x800, 0x0, x779, x735);
+  fiat_p384_addcarryx_u32(&x799, &x800, 0x0, x735, x779);
   uint32_t x801;
   fiat_p384_uint1 x802;
-  fiat_p384_addcarryx_u32(&x801, &x802, x800, x780, x737);
+  fiat_p384_addcarryx_u32(&x801, &x802, x800, x737, x780);
   uint32_t x803;
   fiat_p384_uint1 x804;
-  fiat_p384_addcarryx_u32(&x803, &x804, x802, 0x0, x739);
+  fiat_p384_addcarryx_u32(&x803, &x804, x802, x739, 0x0);
   uint32_t x805;
   fiat_p384_uint1 x806;
-  fiat_p384_addcarryx_u32(&x805, &x806, x804, x777, x741);
+  fiat_p384_addcarryx_u32(&x805, &x806, x804, x741, x777);
   uint32_t x807;
   fiat_p384_uint1 x808;
-  fiat_p384_addcarryx_u32(&x807, &x808, x806, x781, x743);
+  fiat_p384_addcarryx_u32(&x807, &x808, x806, x743, x781);
   uint32_t x809;
   fiat_p384_uint1 x810;
-  fiat_p384_addcarryx_u32(&x809, &x810, x808, x783, x745);
+  fiat_p384_addcarryx_u32(&x809, &x810, x808, x745, x783);
   uint32_t x811;
   fiat_p384_uint1 x812;
-  fiat_p384_addcarryx_u32(&x811, &x812, x810, x785, x747);
+  fiat_p384_addcarryx_u32(&x811, &x812, x810, x747, x785);
   uint32_t x813;
   fiat_p384_uint1 x814;
-  fiat_p384_addcarryx_u32(&x813, &x814, x812, x787, x749);
+  fiat_p384_addcarryx_u32(&x813, &x814, x812, x749, x787);
   uint32_t x815;
   fiat_p384_uint1 x816;
-  fiat_p384_addcarryx_u32(&x815, &x816, x814, x789, x751);
+  fiat_p384_addcarryx_u32(&x815, &x816, x814, x751, x789);
   uint32_t x817;
   fiat_p384_uint1 x818;
-  fiat_p384_addcarryx_u32(&x817, &x818, x816, x791, x753);
+  fiat_p384_addcarryx_u32(&x817, &x818, x816, x753, x791);
   uint32_t x819;
   fiat_p384_uint1 x820;
-  fiat_p384_addcarryx_u32(&x819, &x820, x818, x793, x755);
+  fiat_p384_addcarryx_u32(&x819, &x820, x818, x755, x793);
   uint32_t x821;
   fiat_p384_uint1 x822;
-  fiat_p384_addcarryx_u32(&x821, &x822, x820, x795, x757);
+  fiat_p384_addcarryx_u32(&x821, &x822, x820, x757, x795);
   uint32_t x823;
   fiat_p384_uint1 x824;
-  fiat_p384_addcarryx_u32(&x823, &x824, x822, x797, x759);
+  fiat_p384_addcarryx_u32(&x823, &x824, x822, x759, x797);
   uint32_t x825;
   fiat_p384_uint1 x826;
-  fiat_p384_addcarryx_u32(&x825, &x826, x824, 0x0, x760);
+  fiat_p384_addcarryx_u32(&x825, &x826, x824, x760, 0x0);
   uint32_t x827;
   uint32_t x828;
   fiat_p384_mulx_u32(&x827, &x828, x6, (arg2[11]));
@@ -1387,79 +1387,79 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x849, &x850, x6, (arg2[0]));
   uint32_t x851;
   fiat_p384_uint1 x852;
-  fiat_p384_addcarryx_u32(&x851, &x852, 0x0, x847, x850);
+  fiat_p384_addcarryx_u32(&x851, &x852, 0x0, x850, x847);
   uint32_t x853;
   fiat_p384_uint1 x854;
-  fiat_p384_addcarryx_u32(&x853, &x854, x852, x845, x848);
+  fiat_p384_addcarryx_u32(&x853, &x854, x852, x848, x845);
   uint32_t x855;
   fiat_p384_uint1 x856;
-  fiat_p384_addcarryx_u32(&x855, &x856, x854, x843, x846);
+  fiat_p384_addcarryx_u32(&x855, &x856, x854, x846, x843);
   uint32_t x857;
   fiat_p384_uint1 x858;
-  fiat_p384_addcarryx_u32(&x857, &x858, x856, x841, x844);
+  fiat_p384_addcarryx_u32(&x857, &x858, x856, x844, x841);
   uint32_t x859;
   fiat_p384_uint1 x860;
-  fiat_p384_addcarryx_u32(&x859, &x860, x858, x839, x842);
+  fiat_p384_addcarryx_u32(&x859, &x860, x858, x842, x839);
   uint32_t x861;
   fiat_p384_uint1 x862;
-  fiat_p384_addcarryx_u32(&x861, &x862, x860, x837, x840);
+  fiat_p384_addcarryx_u32(&x861, &x862, x860, x840, x837);
   uint32_t x863;
   fiat_p384_uint1 x864;
-  fiat_p384_addcarryx_u32(&x863, &x864, x862, x835, x838);
+  fiat_p384_addcarryx_u32(&x863, &x864, x862, x838, x835);
   uint32_t x865;
   fiat_p384_uint1 x866;
-  fiat_p384_addcarryx_u32(&x865, &x866, x864, x833, x836);
+  fiat_p384_addcarryx_u32(&x865, &x866, x864, x836, x833);
   uint32_t x867;
   fiat_p384_uint1 x868;
-  fiat_p384_addcarryx_u32(&x867, &x868, x866, x831, x834);
+  fiat_p384_addcarryx_u32(&x867, &x868, x866, x834, x831);
   uint32_t x869;
   fiat_p384_uint1 x870;
-  fiat_p384_addcarryx_u32(&x869, &x870, x868, x829, x832);
+  fiat_p384_addcarryx_u32(&x869, &x870, x868, x832, x829);
   uint32_t x871;
   fiat_p384_uint1 x872;
-  fiat_p384_addcarryx_u32(&x871, &x872, x870, x827, x830);
+  fiat_p384_addcarryx_u32(&x871, &x872, x870, x830, x827);
   uint32_t x873;
   fiat_p384_uint1 x874;
-  fiat_p384_addcarryx_u32(&x873, &x874, x872, 0x0, x828);
+  fiat_p384_addcarryx_u32(&x873, &x874, x872, x828, 0x0);
   uint32_t x875;
   fiat_p384_uint1 x876;
-  fiat_p384_addcarryx_u32(&x875, &x876, 0x0, x849, x801);
+  fiat_p384_addcarryx_u32(&x875, &x876, 0x0, x801, x849);
   uint32_t x877;
   fiat_p384_uint1 x878;
-  fiat_p384_addcarryx_u32(&x877, &x878, x876, x851, x803);
+  fiat_p384_addcarryx_u32(&x877, &x878, x876, x803, x851);
   uint32_t x879;
   fiat_p384_uint1 x880;
-  fiat_p384_addcarryx_u32(&x879, &x880, x878, x853, x805);
+  fiat_p384_addcarryx_u32(&x879, &x880, x878, x805, x853);
   uint32_t x881;
   fiat_p384_uint1 x882;
-  fiat_p384_addcarryx_u32(&x881, &x882, x880, x855, x807);
+  fiat_p384_addcarryx_u32(&x881, &x882, x880, x807, x855);
   uint32_t x883;
   fiat_p384_uint1 x884;
-  fiat_p384_addcarryx_u32(&x883, &x884, x882, x857, x809);
+  fiat_p384_addcarryx_u32(&x883, &x884, x882, x809, x857);
   uint32_t x885;
   fiat_p384_uint1 x886;
-  fiat_p384_addcarryx_u32(&x885, &x886, x884, x859, x811);
+  fiat_p384_addcarryx_u32(&x885, &x886, x884, x811, x859);
   uint32_t x887;
   fiat_p384_uint1 x888;
-  fiat_p384_addcarryx_u32(&x887, &x888, x886, x861, x813);
+  fiat_p384_addcarryx_u32(&x887, &x888, x886, x813, x861);
   uint32_t x889;
   fiat_p384_uint1 x890;
-  fiat_p384_addcarryx_u32(&x889, &x890, x888, x863, x815);
+  fiat_p384_addcarryx_u32(&x889, &x890, x888, x815, x863);
   uint32_t x891;
   fiat_p384_uint1 x892;
-  fiat_p384_addcarryx_u32(&x891, &x892, x890, x865, x817);
+  fiat_p384_addcarryx_u32(&x891, &x892, x890, x817, x865);
   uint32_t x893;
   fiat_p384_uint1 x894;
-  fiat_p384_addcarryx_u32(&x893, &x894, x892, x867, x819);
+  fiat_p384_addcarryx_u32(&x893, &x894, x892, x819, x867);
   uint32_t x895;
   fiat_p384_uint1 x896;
-  fiat_p384_addcarryx_u32(&x895, &x896, x894, x869, x821);
+  fiat_p384_addcarryx_u32(&x895, &x896, x894, x821, x869);
   uint32_t x897;
   fiat_p384_uint1 x898;
-  fiat_p384_addcarryx_u32(&x897, &x898, x896, x871, x823);
+  fiat_p384_addcarryx_u32(&x897, &x898, x896, x823, x871);
   uint32_t x899;
   fiat_p384_uint1 x900;
-  fiat_p384_addcarryx_u32(&x899, &x900, x898, x873, x825);
+  fiat_p384_addcarryx_u32(&x899, &x900, x898, x825, x873);
   uint32_t x901;
   uint32_t x902;
   fiat_p384_mulx_u32(&x901, &x902, x875, UINT32_C(0xffffffff));
@@ -1492,73 +1492,73 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x919, &x920, x875, UINT32_C(0xffffffff));
   uint32_t x921;
   fiat_p384_uint1 x922;
-  fiat_p384_addcarryx_u32(&x921, &x922, 0x0, x915, x918);
+  fiat_p384_addcarryx_u32(&x921, &x922, 0x0, x918, x915);
   uint32_t x923;
   fiat_p384_uint1 x924;
-  fiat_p384_addcarryx_u32(&x923, &x924, x922, x913, x916);
+  fiat_p384_addcarryx_u32(&x923, &x924, x922, x916, x913);
   uint32_t x925;
   fiat_p384_uint1 x926;
-  fiat_p384_addcarryx_u32(&x925, &x926, x924, x911, x914);
+  fiat_p384_addcarryx_u32(&x925, &x926, x924, x914, x911);
   uint32_t x927;
   fiat_p384_uint1 x928;
-  fiat_p384_addcarryx_u32(&x927, &x928, x926, x909, x912);
+  fiat_p384_addcarryx_u32(&x927, &x928, x926, x912, x909);
   uint32_t x929;
   fiat_p384_uint1 x930;
-  fiat_p384_addcarryx_u32(&x929, &x930, x928, x907, x910);
+  fiat_p384_addcarryx_u32(&x929, &x930, x928, x910, x907);
   uint32_t x931;
   fiat_p384_uint1 x932;
-  fiat_p384_addcarryx_u32(&x931, &x932, x930, x905, x908);
+  fiat_p384_addcarryx_u32(&x931, &x932, x930, x908, x905);
   uint32_t x933;
   fiat_p384_uint1 x934;
-  fiat_p384_addcarryx_u32(&x933, &x934, x932, x903, x906);
+  fiat_p384_addcarryx_u32(&x933, &x934, x932, x906, x903);
   uint32_t x935;
   fiat_p384_uint1 x936;
-  fiat_p384_addcarryx_u32(&x935, &x936, x934, x901, x904);
+  fiat_p384_addcarryx_u32(&x935, &x936, x934, x904, x901);
   uint32_t x937;
   fiat_p384_uint1 x938;
-  fiat_p384_addcarryx_u32(&x937, &x938, x936, 0x0, x902);
+  fiat_p384_addcarryx_u32(&x937, &x938, x936, x902, 0x0);
   uint32_t x939;
   fiat_p384_uint1 x940;
-  fiat_p384_addcarryx_u32(&x939, &x940, 0x0, x919, x875);
+  fiat_p384_addcarryx_u32(&x939, &x940, 0x0, x875, x919);
   uint32_t x941;
   fiat_p384_uint1 x942;
-  fiat_p384_addcarryx_u32(&x941, &x942, x940, x920, x877);
+  fiat_p384_addcarryx_u32(&x941, &x942, x940, x877, x920);
   uint32_t x943;
   fiat_p384_uint1 x944;
-  fiat_p384_addcarryx_u32(&x943, &x944, x942, 0x0, x879);
+  fiat_p384_addcarryx_u32(&x943, &x944, x942, x879, 0x0);
   uint32_t x945;
   fiat_p384_uint1 x946;
-  fiat_p384_addcarryx_u32(&x945, &x946, x944, x917, x881);
+  fiat_p384_addcarryx_u32(&x945, &x946, x944, x881, x917);
   uint32_t x947;
   fiat_p384_uint1 x948;
-  fiat_p384_addcarryx_u32(&x947, &x948, x946, x921, x883);
+  fiat_p384_addcarryx_u32(&x947, &x948, x946, x883, x921);
   uint32_t x949;
   fiat_p384_uint1 x950;
-  fiat_p384_addcarryx_u32(&x949, &x950, x948, x923, x885);
+  fiat_p384_addcarryx_u32(&x949, &x950, x948, x885, x923);
   uint32_t x951;
   fiat_p384_uint1 x952;
-  fiat_p384_addcarryx_u32(&x951, &x952, x950, x925, x887);
+  fiat_p384_addcarryx_u32(&x951, &x952, x950, x887, x925);
   uint32_t x953;
   fiat_p384_uint1 x954;
-  fiat_p384_addcarryx_u32(&x953, &x954, x952, x927, x889);
+  fiat_p384_addcarryx_u32(&x953, &x954, x952, x889, x927);
   uint32_t x955;
   fiat_p384_uint1 x956;
-  fiat_p384_addcarryx_u32(&x955, &x956, x954, x929, x891);
+  fiat_p384_addcarryx_u32(&x955, &x956, x954, x891, x929);
   uint32_t x957;
   fiat_p384_uint1 x958;
-  fiat_p384_addcarryx_u32(&x957, &x958, x956, x931, x893);
+  fiat_p384_addcarryx_u32(&x957, &x958, x956, x893, x931);
   uint32_t x959;
   fiat_p384_uint1 x960;
-  fiat_p384_addcarryx_u32(&x959, &x960, x958, x933, x895);
+  fiat_p384_addcarryx_u32(&x959, &x960, x958, x895, x933);
   uint32_t x961;
   fiat_p384_uint1 x962;
-  fiat_p384_addcarryx_u32(&x961, &x962, x960, x935, x897);
+  fiat_p384_addcarryx_u32(&x961, &x962, x960, x897, x935);
   uint32_t x963;
   fiat_p384_uint1 x964;
-  fiat_p384_addcarryx_u32(&x963, &x964, x962, x937, x899);
+  fiat_p384_addcarryx_u32(&x963, &x964, x962, x899, x937);
   uint32_t x965;
   fiat_p384_uint1 x966;
-  fiat_p384_addcarryx_u32(&x965, &x966, x964, 0x0, x900);
+  fiat_p384_addcarryx_u32(&x965, &x966, x964, x900, 0x0);
   uint32_t x967;
   uint32_t x968;
   fiat_p384_mulx_u32(&x967, &x968, x7, (arg2[11]));
@@ -1597,79 +1597,79 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x989, &x990, x7, (arg2[0]));
   uint32_t x991;
   fiat_p384_uint1 x992;
-  fiat_p384_addcarryx_u32(&x991, &x992, 0x0, x987, x990);
+  fiat_p384_addcarryx_u32(&x991, &x992, 0x0, x990, x987);
   uint32_t x993;
   fiat_p384_uint1 x994;
-  fiat_p384_addcarryx_u32(&x993, &x994, x992, x985, x988);
+  fiat_p384_addcarryx_u32(&x993, &x994, x992, x988, x985);
   uint32_t x995;
   fiat_p384_uint1 x996;
-  fiat_p384_addcarryx_u32(&x995, &x996, x994, x983, x986);
+  fiat_p384_addcarryx_u32(&x995, &x996, x994, x986, x983);
   uint32_t x997;
   fiat_p384_uint1 x998;
-  fiat_p384_addcarryx_u32(&x997, &x998, x996, x981, x984);
+  fiat_p384_addcarryx_u32(&x997, &x998, x996, x984, x981);
   uint32_t x999;
   fiat_p384_uint1 x1000;
-  fiat_p384_addcarryx_u32(&x999, &x1000, x998, x979, x982);
+  fiat_p384_addcarryx_u32(&x999, &x1000, x998, x982, x979);
   uint32_t x1001;
   fiat_p384_uint1 x1002;
-  fiat_p384_addcarryx_u32(&x1001, &x1002, x1000, x977, x980);
+  fiat_p384_addcarryx_u32(&x1001, &x1002, x1000, x980, x977);
   uint32_t x1003;
   fiat_p384_uint1 x1004;
-  fiat_p384_addcarryx_u32(&x1003, &x1004, x1002, x975, x978);
+  fiat_p384_addcarryx_u32(&x1003, &x1004, x1002, x978, x975);
   uint32_t x1005;
   fiat_p384_uint1 x1006;
-  fiat_p384_addcarryx_u32(&x1005, &x1006, x1004, x973, x976);
+  fiat_p384_addcarryx_u32(&x1005, &x1006, x1004, x976, x973);
   uint32_t x1007;
   fiat_p384_uint1 x1008;
-  fiat_p384_addcarryx_u32(&x1007, &x1008, x1006, x971, x974);
+  fiat_p384_addcarryx_u32(&x1007, &x1008, x1006, x974, x971);
   uint32_t x1009;
   fiat_p384_uint1 x1010;
-  fiat_p384_addcarryx_u32(&x1009, &x1010, x1008, x969, x972);
+  fiat_p384_addcarryx_u32(&x1009, &x1010, x1008, x972, x969);
   uint32_t x1011;
   fiat_p384_uint1 x1012;
-  fiat_p384_addcarryx_u32(&x1011, &x1012, x1010, x967, x970);
+  fiat_p384_addcarryx_u32(&x1011, &x1012, x1010, x970, x967);
   uint32_t x1013;
   fiat_p384_uint1 x1014;
-  fiat_p384_addcarryx_u32(&x1013, &x1014, x1012, 0x0, x968);
+  fiat_p384_addcarryx_u32(&x1013, &x1014, x1012, x968, 0x0);
   uint32_t x1015;
   fiat_p384_uint1 x1016;
-  fiat_p384_addcarryx_u32(&x1015, &x1016, 0x0, x989, x941);
+  fiat_p384_addcarryx_u32(&x1015, &x1016, 0x0, x941, x989);
   uint32_t x1017;
   fiat_p384_uint1 x1018;
-  fiat_p384_addcarryx_u32(&x1017, &x1018, x1016, x991, x943);
+  fiat_p384_addcarryx_u32(&x1017, &x1018, x1016, x943, x991);
   uint32_t x1019;
   fiat_p384_uint1 x1020;
-  fiat_p384_addcarryx_u32(&x1019, &x1020, x1018, x993, x945);
+  fiat_p384_addcarryx_u32(&x1019, &x1020, x1018, x945, x993);
   uint32_t x1021;
   fiat_p384_uint1 x1022;
-  fiat_p384_addcarryx_u32(&x1021, &x1022, x1020, x995, x947);
+  fiat_p384_addcarryx_u32(&x1021, &x1022, x1020, x947, x995);
   uint32_t x1023;
   fiat_p384_uint1 x1024;
-  fiat_p384_addcarryx_u32(&x1023, &x1024, x1022, x997, x949);
+  fiat_p384_addcarryx_u32(&x1023, &x1024, x1022, x949, x997);
   uint32_t x1025;
   fiat_p384_uint1 x1026;
-  fiat_p384_addcarryx_u32(&x1025, &x1026, x1024, x999, x951);
+  fiat_p384_addcarryx_u32(&x1025, &x1026, x1024, x951, x999);
   uint32_t x1027;
   fiat_p384_uint1 x1028;
-  fiat_p384_addcarryx_u32(&x1027, &x1028, x1026, x1001, x953);
+  fiat_p384_addcarryx_u32(&x1027, &x1028, x1026, x953, x1001);
   uint32_t x1029;
   fiat_p384_uint1 x1030;
-  fiat_p384_addcarryx_u32(&x1029, &x1030, x1028, x1003, x955);
+  fiat_p384_addcarryx_u32(&x1029, &x1030, x1028, x955, x1003);
   uint32_t x1031;
   fiat_p384_uint1 x1032;
-  fiat_p384_addcarryx_u32(&x1031, &x1032, x1030, x1005, x957);
+  fiat_p384_addcarryx_u32(&x1031, &x1032, x1030, x957, x1005);
   uint32_t x1033;
   fiat_p384_uint1 x1034;
-  fiat_p384_addcarryx_u32(&x1033, &x1034, x1032, x1007, x959);
+  fiat_p384_addcarryx_u32(&x1033, &x1034, x1032, x959, x1007);
   uint32_t x1035;
   fiat_p384_uint1 x1036;
-  fiat_p384_addcarryx_u32(&x1035, &x1036, x1034, x1009, x961);
+  fiat_p384_addcarryx_u32(&x1035, &x1036, x1034, x961, x1009);
   uint32_t x1037;
   fiat_p384_uint1 x1038;
-  fiat_p384_addcarryx_u32(&x1037, &x1038, x1036, x1011, x963);
+  fiat_p384_addcarryx_u32(&x1037, &x1038, x1036, x963, x1011);
   uint32_t x1039;
   fiat_p384_uint1 x1040;
-  fiat_p384_addcarryx_u32(&x1039, &x1040, x1038, x1013, x965);
+  fiat_p384_addcarryx_u32(&x1039, &x1040, x1038, x965, x1013);
   uint32_t x1041;
   uint32_t x1042;
   fiat_p384_mulx_u32(&x1041, &x1042, x1015, UINT32_C(0xffffffff));
@@ -1702,73 +1702,73 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x1059, &x1060, x1015, UINT32_C(0xffffffff));
   uint32_t x1061;
   fiat_p384_uint1 x1062;
-  fiat_p384_addcarryx_u32(&x1061, &x1062, 0x0, x1055, x1058);
+  fiat_p384_addcarryx_u32(&x1061, &x1062, 0x0, x1058, x1055);
   uint32_t x1063;
   fiat_p384_uint1 x1064;
-  fiat_p384_addcarryx_u32(&x1063, &x1064, x1062, x1053, x1056);
+  fiat_p384_addcarryx_u32(&x1063, &x1064, x1062, x1056, x1053);
   uint32_t x1065;
   fiat_p384_uint1 x1066;
-  fiat_p384_addcarryx_u32(&x1065, &x1066, x1064, x1051, x1054);
+  fiat_p384_addcarryx_u32(&x1065, &x1066, x1064, x1054, x1051);
   uint32_t x1067;
   fiat_p384_uint1 x1068;
-  fiat_p384_addcarryx_u32(&x1067, &x1068, x1066, x1049, x1052);
+  fiat_p384_addcarryx_u32(&x1067, &x1068, x1066, x1052, x1049);
   uint32_t x1069;
   fiat_p384_uint1 x1070;
-  fiat_p384_addcarryx_u32(&x1069, &x1070, x1068, x1047, x1050);
+  fiat_p384_addcarryx_u32(&x1069, &x1070, x1068, x1050, x1047);
   uint32_t x1071;
   fiat_p384_uint1 x1072;
-  fiat_p384_addcarryx_u32(&x1071, &x1072, x1070, x1045, x1048);
+  fiat_p384_addcarryx_u32(&x1071, &x1072, x1070, x1048, x1045);
   uint32_t x1073;
   fiat_p384_uint1 x1074;
-  fiat_p384_addcarryx_u32(&x1073, &x1074, x1072, x1043, x1046);
+  fiat_p384_addcarryx_u32(&x1073, &x1074, x1072, x1046, x1043);
   uint32_t x1075;
   fiat_p384_uint1 x1076;
-  fiat_p384_addcarryx_u32(&x1075, &x1076, x1074, x1041, x1044);
+  fiat_p384_addcarryx_u32(&x1075, &x1076, x1074, x1044, x1041);
   uint32_t x1077;
   fiat_p384_uint1 x1078;
-  fiat_p384_addcarryx_u32(&x1077, &x1078, x1076, 0x0, x1042);
+  fiat_p384_addcarryx_u32(&x1077, &x1078, x1076, x1042, 0x0);
   uint32_t x1079;
   fiat_p384_uint1 x1080;
-  fiat_p384_addcarryx_u32(&x1079, &x1080, 0x0, x1059, x1015);
+  fiat_p384_addcarryx_u32(&x1079, &x1080, 0x0, x1015, x1059);
   uint32_t x1081;
   fiat_p384_uint1 x1082;
-  fiat_p384_addcarryx_u32(&x1081, &x1082, x1080, x1060, x1017);
+  fiat_p384_addcarryx_u32(&x1081, &x1082, x1080, x1017, x1060);
   uint32_t x1083;
   fiat_p384_uint1 x1084;
-  fiat_p384_addcarryx_u32(&x1083, &x1084, x1082, 0x0, x1019);
+  fiat_p384_addcarryx_u32(&x1083, &x1084, x1082, x1019, 0x0);
   uint32_t x1085;
   fiat_p384_uint1 x1086;
-  fiat_p384_addcarryx_u32(&x1085, &x1086, x1084, x1057, x1021);
+  fiat_p384_addcarryx_u32(&x1085, &x1086, x1084, x1021, x1057);
   uint32_t x1087;
   fiat_p384_uint1 x1088;
-  fiat_p384_addcarryx_u32(&x1087, &x1088, x1086, x1061, x1023);
+  fiat_p384_addcarryx_u32(&x1087, &x1088, x1086, x1023, x1061);
   uint32_t x1089;
   fiat_p384_uint1 x1090;
-  fiat_p384_addcarryx_u32(&x1089, &x1090, x1088, x1063, x1025);
+  fiat_p384_addcarryx_u32(&x1089, &x1090, x1088, x1025, x1063);
   uint32_t x1091;
   fiat_p384_uint1 x1092;
-  fiat_p384_addcarryx_u32(&x1091, &x1092, x1090, x1065, x1027);
+  fiat_p384_addcarryx_u32(&x1091, &x1092, x1090, x1027, x1065);
   uint32_t x1093;
   fiat_p384_uint1 x1094;
-  fiat_p384_addcarryx_u32(&x1093, &x1094, x1092, x1067, x1029);
+  fiat_p384_addcarryx_u32(&x1093, &x1094, x1092, x1029, x1067);
   uint32_t x1095;
   fiat_p384_uint1 x1096;
-  fiat_p384_addcarryx_u32(&x1095, &x1096, x1094, x1069, x1031);
+  fiat_p384_addcarryx_u32(&x1095, &x1096, x1094, x1031, x1069);
   uint32_t x1097;
   fiat_p384_uint1 x1098;
-  fiat_p384_addcarryx_u32(&x1097, &x1098, x1096, x1071, x1033);
+  fiat_p384_addcarryx_u32(&x1097, &x1098, x1096, x1033, x1071);
   uint32_t x1099;
   fiat_p384_uint1 x1100;
-  fiat_p384_addcarryx_u32(&x1099, &x1100, x1098, x1073, x1035);
+  fiat_p384_addcarryx_u32(&x1099, &x1100, x1098, x1035, x1073);
   uint32_t x1101;
   fiat_p384_uint1 x1102;
-  fiat_p384_addcarryx_u32(&x1101, &x1102, x1100, x1075, x1037);
+  fiat_p384_addcarryx_u32(&x1101, &x1102, x1100, x1037, x1075);
   uint32_t x1103;
   fiat_p384_uint1 x1104;
-  fiat_p384_addcarryx_u32(&x1103, &x1104, x1102, x1077, x1039);
+  fiat_p384_addcarryx_u32(&x1103, &x1104, x1102, x1039, x1077);
   uint32_t x1105;
   fiat_p384_uint1 x1106;
-  fiat_p384_addcarryx_u32(&x1105, &x1106, x1104, 0x0, x1040);
+  fiat_p384_addcarryx_u32(&x1105, &x1106, x1104, x1040, 0x0);
   uint32_t x1107;
   uint32_t x1108;
   fiat_p384_mulx_u32(&x1107, &x1108, x8, (arg2[11]));
@@ -1807,79 +1807,79 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x1129, &x1130, x8, (arg2[0]));
   uint32_t x1131;
   fiat_p384_uint1 x1132;
-  fiat_p384_addcarryx_u32(&x1131, &x1132, 0x0, x1127, x1130);
+  fiat_p384_addcarryx_u32(&x1131, &x1132, 0x0, x1130, x1127);
   uint32_t x1133;
   fiat_p384_uint1 x1134;
-  fiat_p384_addcarryx_u32(&x1133, &x1134, x1132, x1125, x1128);
+  fiat_p384_addcarryx_u32(&x1133, &x1134, x1132, x1128, x1125);
   uint32_t x1135;
   fiat_p384_uint1 x1136;
-  fiat_p384_addcarryx_u32(&x1135, &x1136, x1134, x1123, x1126);
+  fiat_p384_addcarryx_u32(&x1135, &x1136, x1134, x1126, x1123);
   uint32_t x1137;
   fiat_p384_uint1 x1138;
-  fiat_p384_addcarryx_u32(&x1137, &x1138, x1136, x1121, x1124);
+  fiat_p384_addcarryx_u32(&x1137, &x1138, x1136, x1124, x1121);
   uint32_t x1139;
   fiat_p384_uint1 x1140;
-  fiat_p384_addcarryx_u32(&x1139, &x1140, x1138, x1119, x1122);
+  fiat_p384_addcarryx_u32(&x1139, &x1140, x1138, x1122, x1119);
   uint32_t x1141;
   fiat_p384_uint1 x1142;
-  fiat_p384_addcarryx_u32(&x1141, &x1142, x1140, x1117, x1120);
+  fiat_p384_addcarryx_u32(&x1141, &x1142, x1140, x1120, x1117);
   uint32_t x1143;
   fiat_p384_uint1 x1144;
-  fiat_p384_addcarryx_u32(&x1143, &x1144, x1142, x1115, x1118);
+  fiat_p384_addcarryx_u32(&x1143, &x1144, x1142, x1118, x1115);
   uint32_t x1145;
   fiat_p384_uint1 x1146;
-  fiat_p384_addcarryx_u32(&x1145, &x1146, x1144, x1113, x1116);
+  fiat_p384_addcarryx_u32(&x1145, &x1146, x1144, x1116, x1113);
   uint32_t x1147;
   fiat_p384_uint1 x1148;
-  fiat_p384_addcarryx_u32(&x1147, &x1148, x1146, x1111, x1114);
+  fiat_p384_addcarryx_u32(&x1147, &x1148, x1146, x1114, x1111);
   uint32_t x1149;
   fiat_p384_uint1 x1150;
-  fiat_p384_addcarryx_u32(&x1149, &x1150, x1148, x1109, x1112);
+  fiat_p384_addcarryx_u32(&x1149, &x1150, x1148, x1112, x1109);
   uint32_t x1151;
   fiat_p384_uint1 x1152;
-  fiat_p384_addcarryx_u32(&x1151, &x1152, x1150, x1107, x1110);
+  fiat_p384_addcarryx_u32(&x1151, &x1152, x1150, x1110, x1107);
   uint32_t x1153;
   fiat_p384_uint1 x1154;
-  fiat_p384_addcarryx_u32(&x1153, &x1154, x1152, 0x0, x1108);
+  fiat_p384_addcarryx_u32(&x1153, &x1154, x1152, x1108, 0x0);
   uint32_t x1155;
   fiat_p384_uint1 x1156;
-  fiat_p384_addcarryx_u32(&x1155, &x1156, 0x0, x1129, x1081);
+  fiat_p384_addcarryx_u32(&x1155, &x1156, 0x0, x1081, x1129);
   uint32_t x1157;
   fiat_p384_uint1 x1158;
-  fiat_p384_addcarryx_u32(&x1157, &x1158, x1156, x1131, x1083);
+  fiat_p384_addcarryx_u32(&x1157, &x1158, x1156, x1083, x1131);
   uint32_t x1159;
   fiat_p384_uint1 x1160;
-  fiat_p384_addcarryx_u32(&x1159, &x1160, x1158, x1133, x1085);
+  fiat_p384_addcarryx_u32(&x1159, &x1160, x1158, x1085, x1133);
   uint32_t x1161;
   fiat_p384_uint1 x1162;
-  fiat_p384_addcarryx_u32(&x1161, &x1162, x1160, x1135, x1087);
+  fiat_p384_addcarryx_u32(&x1161, &x1162, x1160, x1087, x1135);
   uint32_t x1163;
   fiat_p384_uint1 x1164;
-  fiat_p384_addcarryx_u32(&x1163, &x1164, x1162, x1137, x1089);
+  fiat_p384_addcarryx_u32(&x1163, &x1164, x1162, x1089, x1137);
   uint32_t x1165;
   fiat_p384_uint1 x1166;
-  fiat_p384_addcarryx_u32(&x1165, &x1166, x1164, x1139, x1091);
+  fiat_p384_addcarryx_u32(&x1165, &x1166, x1164, x1091, x1139);
   uint32_t x1167;
   fiat_p384_uint1 x1168;
-  fiat_p384_addcarryx_u32(&x1167, &x1168, x1166, x1141, x1093);
+  fiat_p384_addcarryx_u32(&x1167, &x1168, x1166, x1093, x1141);
   uint32_t x1169;
   fiat_p384_uint1 x1170;
-  fiat_p384_addcarryx_u32(&x1169, &x1170, x1168, x1143, x1095);
+  fiat_p384_addcarryx_u32(&x1169, &x1170, x1168, x1095, x1143);
   uint32_t x1171;
   fiat_p384_uint1 x1172;
-  fiat_p384_addcarryx_u32(&x1171, &x1172, x1170, x1145, x1097);
+  fiat_p384_addcarryx_u32(&x1171, &x1172, x1170, x1097, x1145);
   uint32_t x1173;
   fiat_p384_uint1 x1174;
-  fiat_p384_addcarryx_u32(&x1173, &x1174, x1172, x1147, x1099);
+  fiat_p384_addcarryx_u32(&x1173, &x1174, x1172, x1099, x1147);
   uint32_t x1175;
   fiat_p384_uint1 x1176;
-  fiat_p384_addcarryx_u32(&x1175, &x1176, x1174, x1149, x1101);
+  fiat_p384_addcarryx_u32(&x1175, &x1176, x1174, x1101, x1149);
   uint32_t x1177;
   fiat_p384_uint1 x1178;
-  fiat_p384_addcarryx_u32(&x1177, &x1178, x1176, x1151, x1103);
+  fiat_p384_addcarryx_u32(&x1177, &x1178, x1176, x1103, x1151);
   uint32_t x1179;
   fiat_p384_uint1 x1180;
-  fiat_p384_addcarryx_u32(&x1179, &x1180, x1178, x1153, x1105);
+  fiat_p384_addcarryx_u32(&x1179, &x1180, x1178, x1105, x1153);
   uint32_t x1181;
   uint32_t x1182;
   fiat_p384_mulx_u32(&x1181, &x1182, x1155, UINT32_C(0xffffffff));
@@ -1912,73 +1912,73 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x1199, &x1200, x1155, UINT32_C(0xffffffff));
   uint32_t x1201;
   fiat_p384_uint1 x1202;
-  fiat_p384_addcarryx_u32(&x1201, &x1202, 0x0, x1195, x1198);
+  fiat_p384_addcarryx_u32(&x1201, &x1202, 0x0, x1198, x1195);
   uint32_t x1203;
   fiat_p384_uint1 x1204;
-  fiat_p384_addcarryx_u32(&x1203, &x1204, x1202, x1193, x1196);
+  fiat_p384_addcarryx_u32(&x1203, &x1204, x1202, x1196, x1193);
   uint32_t x1205;
   fiat_p384_uint1 x1206;
-  fiat_p384_addcarryx_u32(&x1205, &x1206, x1204, x1191, x1194);
+  fiat_p384_addcarryx_u32(&x1205, &x1206, x1204, x1194, x1191);
   uint32_t x1207;
   fiat_p384_uint1 x1208;
-  fiat_p384_addcarryx_u32(&x1207, &x1208, x1206, x1189, x1192);
+  fiat_p384_addcarryx_u32(&x1207, &x1208, x1206, x1192, x1189);
   uint32_t x1209;
   fiat_p384_uint1 x1210;
-  fiat_p384_addcarryx_u32(&x1209, &x1210, x1208, x1187, x1190);
+  fiat_p384_addcarryx_u32(&x1209, &x1210, x1208, x1190, x1187);
   uint32_t x1211;
   fiat_p384_uint1 x1212;
-  fiat_p384_addcarryx_u32(&x1211, &x1212, x1210, x1185, x1188);
+  fiat_p384_addcarryx_u32(&x1211, &x1212, x1210, x1188, x1185);
   uint32_t x1213;
   fiat_p384_uint1 x1214;
-  fiat_p384_addcarryx_u32(&x1213, &x1214, x1212, x1183, x1186);
+  fiat_p384_addcarryx_u32(&x1213, &x1214, x1212, x1186, x1183);
   uint32_t x1215;
   fiat_p384_uint1 x1216;
-  fiat_p384_addcarryx_u32(&x1215, &x1216, x1214, x1181, x1184);
+  fiat_p384_addcarryx_u32(&x1215, &x1216, x1214, x1184, x1181);
   uint32_t x1217;
   fiat_p384_uint1 x1218;
-  fiat_p384_addcarryx_u32(&x1217, &x1218, x1216, 0x0, x1182);
+  fiat_p384_addcarryx_u32(&x1217, &x1218, x1216, x1182, 0x0);
   uint32_t x1219;
   fiat_p384_uint1 x1220;
-  fiat_p384_addcarryx_u32(&x1219, &x1220, 0x0, x1199, x1155);
+  fiat_p384_addcarryx_u32(&x1219, &x1220, 0x0, x1155, x1199);
   uint32_t x1221;
   fiat_p384_uint1 x1222;
-  fiat_p384_addcarryx_u32(&x1221, &x1222, x1220, x1200, x1157);
+  fiat_p384_addcarryx_u32(&x1221, &x1222, x1220, x1157, x1200);
   uint32_t x1223;
   fiat_p384_uint1 x1224;
-  fiat_p384_addcarryx_u32(&x1223, &x1224, x1222, 0x0, x1159);
+  fiat_p384_addcarryx_u32(&x1223, &x1224, x1222, x1159, 0x0);
   uint32_t x1225;
   fiat_p384_uint1 x1226;
-  fiat_p384_addcarryx_u32(&x1225, &x1226, x1224, x1197, x1161);
+  fiat_p384_addcarryx_u32(&x1225, &x1226, x1224, x1161, x1197);
   uint32_t x1227;
   fiat_p384_uint1 x1228;
-  fiat_p384_addcarryx_u32(&x1227, &x1228, x1226, x1201, x1163);
+  fiat_p384_addcarryx_u32(&x1227, &x1228, x1226, x1163, x1201);
   uint32_t x1229;
   fiat_p384_uint1 x1230;
-  fiat_p384_addcarryx_u32(&x1229, &x1230, x1228, x1203, x1165);
+  fiat_p384_addcarryx_u32(&x1229, &x1230, x1228, x1165, x1203);
   uint32_t x1231;
   fiat_p384_uint1 x1232;
-  fiat_p384_addcarryx_u32(&x1231, &x1232, x1230, x1205, x1167);
+  fiat_p384_addcarryx_u32(&x1231, &x1232, x1230, x1167, x1205);
   uint32_t x1233;
   fiat_p384_uint1 x1234;
-  fiat_p384_addcarryx_u32(&x1233, &x1234, x1232, x1207, x1169);
+  fiat_p384_addcarryx_u32(&x1233, &x1234, x1232, x1169, x1207);
   uint32_t x1235;
   fiat_p384_uint1 x1236;
-  fiat_p384_addcarryx_u32(&x1235, &x1236, x1234, x1209, x1171);
+  fiat_p384_addcarryx_u32(&x1235, &x1236, x1234, x1171, x1209);
   uint32_t x1237;
   fiat_p384_uint1 x1238;
-  fiat_p384_addcarryx_u32(&x1237, &x1238, x1236, x1211, x1173);
+  fiat_p384_addcarryx_u32(&x1237, &x1238, x1236, x1173, x1211);
   uint32_t x1239;
   fiat_p384_uint1 x1240;
-  fiat_p384_addcarryx_u32(&x1239, &x1240, x1238, x1213, x1175);
+  fiat_p384_addcarryx_u32(&x1239, &x1240, x1238, x1175, x1213);
   uint32_t x1241;
   fiat_p384_uint1 x1242;
-  fiat_p384_addcarryx_u32(&x1241, &x1242, x1240, x1215, x1177);
+  fiat_p384_addcarryx_u32(&x1241, &x1242, x1240, x1177, x1215);
   uint32_t x1243;
   fiat_p384_uint1 x1244;
-  fiat_p384_addcarryx_u32(&x1243, &x1244, x1242, x1217, x1179);
+  fiat_p384_addcarryx_u32(&x1243, &x1244, x1242, x1179, x1217);
   uint32_t x1245;
   fiat_p384_uint1 x1246;
-  fiat_p384_addcarryx_u32(&x1245, &x1246, x1244, 0x0, x1180);
+  fiat_p384_addcarryx_u32(&x1245, &x1246, x1244, x1180, 0x0);
   uint32_t x1247;
   uint32_t x1248;
   fiat_p384_mulx_u32(&x1247, &x1248, x9, (arg2[11]));
@@ -2017,79 +2017,79 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x1269, &x1270, x9, (arg2[0]));
   uint32_t x1271;
   fiat_p384_uint1 x1272;
-  fiat_p384_addcarryx_u32(&x1271, &x1272, 0x0, x1267, x1270);
+  fiat_p384_addcarryx_u32(&x1271, &x1272, 0x0, x1270, x1267);
   uint32_t x1273;
   fiat_p384_uint1 x1274;
-  fiat_p384_addcarryx_u32(&x1273, &x1274, x1272, x1265, x1268);
+  fiat_p384_addcarryx_u32(&x1273, &x1274, x1272, x1268, x1265);
   uint32_t x1275;
   fiat_p384_uint1 x1276;
-  fiat_p384_addcarryx_u32(&x1275, &x1276, x1274, x1263, x1266);
+  fiat_p384_addcarryx_u32(&x1275, &x1276, x1274, x1266, x1263);
   uint32_t x1277;
   fiat_p384_uint1 x1278;
-  fiat_p384_addcarryx_u32(&x1277, &x1278, x1276, x1261, x1264);
+  fiat_p384_addcarryx_u32(&x1277, &x1278, x1276, x1264, x1261);
   uint32_t x1279;
   fiat_p384_uint1 x1280;
-  fiat_p384_addcarryx_u32(&x1279, &x1280, x1278, x1259, x1262);
+  fiat_p384_addcarryx_u32(&x1279, &x1280, x1278, x1262, x1259);
   uint32_t x1281;
   fiat_p384_uint1 x1282;
-  fiat_p384_addcarryx_u32(&x1281, &x1282, x1280, x1257, x1260);
+  fiat_p384_addcarryx_u32(&x1281, &x1282, x1280, x1260, x1257);
   uint32_t x1283;
   fiat_p384_uint1 x1284;
-  fiat_p384_addcarryx_u32(&x1283, &x1284, x1282, x1255, x1258);
+  fiat_p384_addcarryx_u32(&x1283, &x1284, x1282, x1258, x1255);
   uint32_t x1285;
   fiat_p384_uint1 x1286;
-  fiat_p384_addcarryx_u32(&x1285, &x1286, x1284, x1253, x1256);
+  fiat_p384_addcarryx_u32(&x1285, &x1286, x1284, x1256, x1253);
   uint32_t x1287;
   fiat_p384_uint1 x1288;
-  fiat_p384_addcarryx_u32(&x1287, &x1288, x1286, x1251, x1254);
+  fiat_p384_addcarryx_u32(&x1287, &x1288, x1286, x1254, x1251);
   uint32_t x1289;
   fiat_p384_uint1 x1290;
-  fiat_p384_addcarryx_u32(&x1289, &x1290, x1288, x1249, x1252);
+  fiat_p384_addcarryx_u32(&x1289, &x1290, x1288, x1252, x1249);
   uint32_t x1291;
   fiat_p384_uint1 x1292;
-  fiat_p384_addcarryx_u32(&x1291, &x1292, x1290, x1247, x1250);
+  fiat_p384_addcarryx_u32(&x1291, &x1292, x1290, x1250, x1247);
   uint32_t x1293;
   fiat_p384_uint1 x1294;
-  fiat_p384_addcarryx_u32(&x1293, &x1294, x1292, 0x0, x1248);
+  fiat_p384_addcarryx_u32(&x1293, &x1294, x1292, x1248, 0x0);
   uint32_t x1295;
   fiat_p384_uint1 x1296;
-  fiat_p384_addcarryx_u32(&x1295, &x1296, 0x0, x1269, x1221);
+  fiat_p384_addcarryx_u32(&x1295, &x1296, 0x0, x1221, x1269);
   uint32_t x1297;
   fiat_p384_uint1 x1298;
-  fiat_p384_addcarryx_u32(&x1297, &x1298, x1296, x1271, x1223);
+  fiat_p384_addcarryx_u32(&x1297, &x1298, x1296, x1223, x1271);
   uint32_t x1299;
   fiat_p384_uint1 x1300;
-  fiat_p384_addcarryx_u32(&x1299, &x1300, x1298, x1273, x1225);
+  fiat_p384_addcarryx_u32(&x1299, &x1300, x1298, x1225, x1273);
   uint32_t x1301;
   fiat_p384_uint1 x1302;
-  fiat_p384_addcarryx_u32(&x1301, &x1302, x1300, x1275, x1227);
+  fiat_p384_addcarryx_u32(&x1301, &x1302, x1300, x1227, x1275);
   uint32_t x1303;
   fiat_p384_uint1 x1304;
-  fiat_p384_addcarryx_u32(&x1303, &x1304, x1302, x1277, x1229);
+  fiat_p384_addcarryx_u32(&x1303, &x1304, x1302, x1229, x1277);
   uint32_t x1305;
   fiat_p384_uint1 x1306;
-  fiat_p384_addcarryx_u32(&x1305, &x1306, x1304, x1279, x1231);
+  fiat_p384_addcarryx_u32(&x1305, &x1306, x1304, x1231, x1279);
   uint32_t x1307;
   fiat_p384_uint1 x1308;
-  fiat_p384_addcarryx_u32(&x1307, &x1308, x1306, x1281, x1233);
+  fiat_p384_addcarryx_u32(&x1307, &x1308, x1306, x1233, x1281);
   uint32_t x1309;
   fiat_p384_uint1 x1310;
-  fiat_p384_addcarryx_u32(&x1309, &x1310, x1308, x1283, x1235);
+  fiat_p384_addcarryx_u32(&x1309, &x1310, x1308, x1235, x1283);
   uint32_t x1311;
   fiat_p384_uint1 x1312;
-  fiat_p384_addcarryx_u32(&x1311, &x1312, x1310, x1285, x1237);
+  fiat_p384_addcarryx_u32(&x1311, &x1312, x1310, x1237, x1285);
   uint32_t x1313;
   fiat_p384_uint1 x1314;
-  fiat_p384_addcarryx_u32(&x1313, &x1314, x1312, x1287, x1239);
+  fiat_p384_addcarryx_u32(&x1313, &x1314, x1312, x1239, x1287);
   uint32_t x1315;
   fiat_p384_uint1 x1316;
-  fiat_p384_addcarryx_u32(&x1315, &x1316, x1314, x1289, x1241);
+  fiat_p384_addcarryx_u32(&x1315, &x1316, x1314, x1241, x1289);
   uint32_t x1317;
   fiat_p384_uint1 x1318;
-  fiat_p384_addcarryx_u32(&x1317, &x1318, x1316, x1291, x1243);
+  fiat_p384_addcarryx_u32(&x1317, &x1318, x1316, x1243, x1291);
   uint32_t x1319;
   fiat_p384_uint1 x1320;
-  fiat_p384_addcarryx_u32(&x1319, &x1320, x1318, x1293, x1245);
+  fiat_p384_addcarryx_u32(&x1319, &x1320, x1318, x1245, x1293);
   uint32_t x1321;
   uint32_t x1322;
   fiat_p384_mulx_u32(&x1321, &x1322, x1295, UINT32_C(0xffffffff));
@@ -2122,73 +2122,73 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x1339, &x1340, x1295, UINT32_C(0xffffffff));
   uint32_t x1341;
   fiat_p384_uint1 x1342;
-  fiat_p384_addcarryx_u32(&x1341, &x1342, 0x0, x1335, x1338);
+  fiat_p384_addcarryx_u32(&x1341, &x1342, 0x0, x1338, x1335);
   uint32_t x1343;
   fiat_p384_uint1 x1344;
-  fiat_p384_addcarryx_u32(&x1343, &x1344, x1342, x1333, x1336);
+  fiat_p384_addcarryx_u32(&x1343, &x1344, x1342, x1336, x1333);
   uint32_t x1345;
   fiat_p384_uint1 x1346;
-  fiat_p384_addcarryx_u32(&x1345, &x1346, x1344, x1331, x1334);
+  fiat_p384_addcarryx_u32(&x1345, &x1346, x1344, x1334, x1331);
   uint32_t x1347;
   fiat_p384_uint1 x1348;
-  fiat_p384_addcarryx_u32(&x1347, &x1348, x1346, x1329, x1332);
+  fiat_p384_addcarryx_u32(&x1347, &x1348, x1346, x1332, x1329);
   uint32_t x1349;
   fiat_p384_uint1 x1350;
-  fiat_p384_addcarryx_u32(&x1349, &x1350, x1348, x1327, x1330);
+  fiat_p384_addcarryx_u32(&x1349, &x1350, x1348, x1330, x1327);
   uint32_t x1351;
   fiat_p384_uint1 x1352;
-  fiat_p384_addcarryx_u32(&x1351, &x1352, x1350, x1325, x1328);
+  fiat_p384_addcarryx_u32(&x1351, &x1352, x1350, x1328, x1325);
   uint32_t x1353;
   fiat_p384_uint1 x1354;
-  fiat_p384_addcarryx_u32(&x1353, &x1354, x1352, x1323, x1326);
+  fiat_p384_addcarryx_u32(&x1353, &x1354, x1352, x1326, x1323);
   uint32_t x1355;
   fiat_p384_uint1 x1356;
-  fiat_p384_addcarryx_u32(&x1355, &x1356, x1354, x1321, x1324);
+  fiat_p384_addcarryx_u32(&x1355, &x1356, x1354, x1324, x1321);
   uint32_t x1357;
   fiat_p384_uint1 x1358;
-  fiat_p384_addcarryx_u32(&x1357, &x1358, x1356, 0x0, x1322);
+  fiat_p384_addcarryx_u32(&x1357, &x1358, x1356, x1322, 0x0);
   uint32_t x1359;
   fiat_p384_uint1 x1360;
-  fiat_p384_addcarryx_u32(&x1359, &x1360, 0x0, x1339, x1295);
+  fiat_p384_addcarryx_u32(&x1359, &x1360, 0x0, x1295, x1339);
   uint32_t x1361;
   fiat_p384_uint1 x1362;
-  fiat_p384_addcarryx_u32(&x1361, &x1362, x1360, x1340, x1297);
+  fiat_p384_addcarryx_u32(&x1361, &x1362, x1360, x1297, x1340);
   uint32_t x1363;
   fiat_p384_uint1 x1364;
-  fiat_p384_addcarryx_u32(&x1363, &x1364, x1362, 0x0, x1299);
+  fiat_p384_addcarryx_u32(&x1363, &x1364, x1362, x1299, 0x0);
   uint32_t x1365;
   fiat_p384_uint1 x1366;
-  fiat_p384_addcarryx_u32(&x1365, &x1366, x1364, x1337, x1301);
+  fiat_p384_addcarryx_u32(&x1365, &x1366, x1364, x1301, x1337);
   uint32_t x1367;
   fiat_p384_uint1 x1368;
-  fiat_p384_addcarryx_u32(&x1367, &x1368, x1366, x1341, x1303);
+  fiat_p384_addcarryx_u32(&x1367, &x1368, x1366, x1303, x1341);
   uint32_t x1369;
   fiat_p384_uint1 x1370;
-  fiat_p384_addcarryx_u32(&x1369, &x1370, x1368, x1343, x1305);
+  fiat_p384_addcarryx_u32(&x1369, &x1370, x1368, x1305, x1343);
   uint32_t x1371;
   fiat_p384_uint1 x1372;
-  fiat_p384_addcarryx_u32(&x1371, &x1372, x1370, x1345, x1307);
+  fiat_p384_addcarryx_u32(&x1371, &x1372, x1370, x1307, x1345);
   uint32_t x1373;
   fiat_p384_uint1 x1374;
-  fiat_p384_addcarryx_u32(&x1373, &x1374, x1372, x1347, x1309);
+  fiat_p384_addcarryx_u32(&x1373, &x1374, x1372, x1309, x1347);
   uint32_t x1375;
   fiat_p384_uint1 x1376;
-  fiat_p384_addcarryx_u32(&x1375, &x1376, x1374, x1349, x1311);
+  fiat_p384_addcarryx_u32(&x1375, &x1376, x1374, x1311, x1349);
   uint32_t x1377;
   fiat_p384_uint1 x1378;
-  fiat_p384_addcarryx_u32(&x1377, &x1378, x1376, x1351, x1313);
+  fiat_p384_addcarryx_u32(&x1377, &x1378, x1376, x1313, x1351);
   uint32_t x1379;
   fiat_p384_uint1 x1380;
-  fiat_p384_addcarryx_u32(&x1379, &x1380, x1378, x1353, x1315);
+  fiat_p384_addcarryx_u32(&x1379, &x1380, x1378, x1315, x1353);
   uint32_t x1381;
   fiat_p384_uint1 x1382;
-  fiat_p384_addcarryx_u32(&x1381, &x1382, x1380, x1355, x1317);
+  fiat_p384_addcarryx_u32(&x1381, &x1382, x1380, x1317, x1355);
   uint32_t x1383;
   fiat_p384_uint1 x1384;
-  fiat_p384_addcarryx_u32(&x1383, &x1384, x1382, x1357, x1319);
+  fiat_p384_addcarryx_u32(&x1383, &x1384, x1382, x1319, x1357);
   uint32_t x1385;
   fiat_p384_uint1 x1386;
-  fiat_p384_addcarryx_u32(&x1385, &x1386, x1384, 0x0, x1320);
+  fiat_p384_addcarryx_u32(&x1385, &x1386, x1384, x1320, 0x0);
   uint32_t x1387;
   uint32_t x1388;
   fiat_p384_mulx_u32(&x1387, &x1388, x10, (arg2[11]));
@@ -2227,79 +2227,79 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x1409, &x1410, x10, (arg2[0]));
   uint32_t x1411;
   fiat_p384_uint1 x1412;
-  fiat_p384_addcarryx_u32(&x1411, &x1412, 0x0, x1407, x1410);
+  fiat_p384_addcarryx_u32(&x1411, &x1412, 0x0, x1410, x1407);
   uint32_t x1413;
   fiat_p384_uint1 x1414;
-  fiat_p384_addcarryx_u32(&x1413, &x1414, x1412, x1405, x1408);
+  fiat_p384_addcarryx_u32(&x1413, &x1414, x1412, x1408, x1405);
   uint32_t x1415;
   fiat_p384_uint1 x1416;
-  fiat_p384_addcarryx_u32(&x1415, &x1416, x1414, x1403, x1406);
+  fiat_p384_addcarryx_u32(&x1415, &x1416, x1414, x1406, x1403);
   uint32_t x1417;
   fiat_p384_uint1 x1418;
-  fiat_p384_addcarryx_u32(&x1417, &x1418, x1416, x1401, x1404);
+  fiat_p384_addcarryx_u32(&x1417, &x1418, x1416, x1404, x1401);
   uint32_t x1419;
   fiat_p384_uint1 x1420;
-  fiat_p384_addcarryx_u32(&x1419, &x1420, x1418, x1399, x1402);
+  fiat_p384_addcarryx_u32(&x1419, &x1420, x1418, x1402, x1399);
   uint32_t x1421;
   fiat_p384_uint1 x1422;
-  fiat_p384_addcarryx_u32(&x1421, &x1422, x1420, x1397, x1400);
+  fiat_p384_addcarryx_u32(&x1421, &x1422, x1420, x1400, x1397);
   uint32_t x1423;
   fiat_p384_uint1 x1424;
-  fiat_p384_addcarryx_u32(&x1423, &x1424, x1422, x1395, x1398);
+  fiat_p384_addcarryx_u32(&x1423, &x1424, x1422, x1398, x1395);
   uint32_t x1425;
   fiat_p384_uint1 x1426;
-  fiat_p384_addcarryx_u32(&x1425, &x1426, x1424, x1393, x1396);
+  fiat_p384_addcarryx_u32(&x1425, &x1426, x1424, x1396, x1393);
   uint32_t x1427;
   fiat_p384_uint1 x1428;
-  fiat_p384_addcarryx_u32(&x1427, &x1428, x1426, x1391, x1394);
+  fiat_p384_addcarryx_u32(&x1427, &x1428, x1426, x1394, x1391);
   uint32_t x1429;
   fiat_p384_uint1 x1430;
-  fiat_p384_addcarryx_u32(&x1429, &x1430, x1428, x1389, x1392);
+  fiat_p384_addcarryx_u32(&x1429, &x1430, x1428, x1392, x1389);
   uint32_t x1431;
   fiat_p384_uint1 x1432;
-  fiat_p384_addcarryx_u32(&x1431, &x1432, x1430, x1387, x1390);
+  fiat_p384_addcarryx_u32(&x1431, &x1432, x1430, x1390, x1387);
   uint32_t x1433;
   fiat_p384_uint1 x1434;
-  fiat_p384_addcarryx_u32(&x1433, &x1434, x1432, 0x0, x1388);
+  fiat_p384_addcarryx_u32(&x1433, &x1434, x1432, x1388, 0x0);
   uint32_t x1435;
   fiat_p384_uint1 x1436;
-  fiat_p384_addcarryx_u32(&x1435, &x1436, 0x0, x1409, x1361);
+  fiat_p384_addcarryx_u32(&x1435, &x1436, 0x0, x1361, x1409);
   uint32_t x1437;
   fiat_p384_uint1 x1438;
-  fiat_p384_addcarryx_u32(&x1437, &x1438, x1436, x1411, x1363);
+  fiat_p384_addcarryx_u32(&x1437, &x1438, x1436, x1363, x1411);
   uint32_t x1439;
   fiat_p384_uint1 x1440;
-  fiat_p384_addcarryx_u32(&x1439, &x1440, x1438, x1413, x1365);
+  fiat_p384_addcarryx_u32(&x1439, &x1440, x1438, x1365, x1413);
   uint32_t x1441;
   fiat_p384_uint1 x1442;
-  fiat_p384_addcarryx_u32(&x1441, &x1442, x1440, x1415, x1367);
+  fiat_p384_addcarryx_u32(&x1441, &x1442, x1440, x1367, x1415);
   uint32_t x1443;
   fiat_p384_uint1 x1444;
-  fiat_p384_addcarryx_u32(&x1443, &x1444, x1442, x1417, x1369);
+  fiat_p384_addcarryx_u32(&x1443, &x1444, x1442, x1369, x1417);
   uint32_t x1445;
   fiat_p384_uint1 x1446;
-  fiat_p384_addcarryx_u32(&x1445, &x1446, x1444, x1419, x1371);
+  fiat_p384_addcarryx_u32(&x1445, &x1446, x1444, x1371, x1419);
   uint32_t x1447;
   fiat_p384_uint1 x1448;
-  fiat_p384_addcarryx_u32(&x1447, &x1448, x1446, x1421, x1373);
+  fiat_p384_addcarryx_u32(&x1447, &x1448, x1446, x1373, x1421);
   uint32_t x1449;
   fiat_p384_uint1 x1450;
-  fiat_p384_addcarryx_u32(&x1449, &x1450, x1448, x1423, x1375);
+  fiat_p384_addcarryx_u32(&x1449, &x1450, x1448, x1375, x1423);
   uint32_t x1451;
   fiat_p384_uint1 x1452;
-  fiat_p384_addcarryx_u32(&x1451, &x1452, x1450, x1425, x1377);
+  fiat_p384_addcarryx_u32(&x1451, &x1452, x1450, x1377, x1425);
   uint32_t x1453;
   fiat_p384_uint1 x1454;
-  fiat_p384_addcarryx_u32(&x1453, &x1454, x1452, x1427, x1379);
+  fiat_p384_addcarryx_u32(&x1453, &x1454, x1452, x1379, x1427);
   uint32_t x1455;
   fiat_p384_uint1 x1456;
-  fiat_p384_addcarryx_u32(&x1455, &x1456, x1454, x1429, x1381);
+  fiat_p384_addcarryx_u32(&x1455, &x1456, x1454, x1381, x1429);
   uint32_t x1457;
   fiat_p384_uint1 x1458;
-  fiat_p384_addcarryx_u32(&x1457, &x1458, x1456, x1431, x1383);
+  fiat_p384_addcarryx_u32(&x1457, &x1458, x1456, x1383, x1431);
   uint32_t x1459;
   fiat_p384_uint1 x1460;
-  fiat_p384_addcarryx_u32(&x1459, &x1460, x1458, x1433, x1385);
+  fiat_p384_addcarryx_u32(&x1459, &x1460, x1458, x1385, x1433);
   uint32_t x1461;
   uint32_t x1462;
   fiat_p384_mulx_u32(&x1461, &x1462, x1435, UINT32_C(0xffffffff));
@@ -2332,73 +2332,73 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x1479, &x1480, x1435, UINT32_C(0xffffffff));
   uint32_t x1481;
   fiat_p384_uint1 x1482;
-  fiat_p384_addcarryx_u32(&x1481, &x1482, 0x0, x1475, x1478);
+  fiat_p384_addcarryx_u32(&x1481, &x1482, 0x0, x1478, x1475);
   uint32_t x1483;
   fiat_p384_uint1 x1484;
-  fiat_p384_addcarryx_u32(&x1483, &x1484, x1482, x1473, x1476);
+  fiat_p384_addcarryx_u32(&x1483, &x1484, x1482, x1476, x1473);
   uint32_t x1485;
   fiat_p384_uint1 x1486;
-  fiat_p384_addcarryx_u32(&x1485, &x1486, x1484, x1471, x1474);
+  fiat_p384_addcarryx_u32(&x1485, &x1486, x1484, x1474, x1471);
   uint32_t x1487;
   fiat_p384_uint1 x1488;
-  fiat_p384_addcarryx_u32(&x1487, &x1488, x1486, x1469, x1472);
+  fiat_p384_addcarryx_u32(&x1487, &x1488, x1486, x1472, x1469);
   uint32_t x1489;
   fiat_p384_uint1 x1490;
-  fiat_p384_addcarryx_u32(&x1489, &x1490, x1488, x1467, x1470);
+  fiat_p384_addcarryx_u32(&x1489, &x1490, x1488, x1470, x1467);
   uint32_t x1491;
   fiat_p384_uint1 x1492;
-  fiat_p384_addcarryx_u32(&x1491, &x1492, x1490, x1465, x1468);
+  fiat_p384_addcarryx_u32(&x1491, &x1492, x1490, x1468, x1465);
   uint32_t x1493;
   fiat_p384_uint1 x1494;
-  fiat_p384_addcarryx_u32(&x1493, &x1494, x1492, x1463, x1466);
+  fiat_p384_addcarryx_u32(&x1493, &x1494, x1492, x1466, x1463);
   uint32_t x1495;
   fiat_p384_uint1 x1496;
-  fiat_p384_addcarryx_u32(&x1495, &x1496, x1494, x1461, x1464);
+  fiat_p384_addcarryx_u32(&x1495, &x1496, x1494, x1464, x1461);
   uint32_t x1497;
   fiat_p384_uint1 x1498;
-  fiat_p384_addcarryx_u32(&x1497, &x1498, x1496, 0x0, x1462);
+  fiat_p384_addcarryx_u32(&x1497, &x1498, x1496, x1462, 0x0);
   uint32_t x1499;
   fiat_p384_uint1 x1500;
-  fiat_p384_addcarryx_u32(&x1499, &x1500, 0x0, x1479, x1435);
+  fiat_p384_addcarryx_u32(&x1499, &x1500, 0x0, x1435, x1479);
   uint32_t x1501;
   fiat_p384_uint1 x1502;
-  fiat_p384_addcarryx_u32(&x1501, &x1502, x1500, x1480, x1437);
+  fiat_p384_addcarryx_u32(&x1501, &x1502, x1500, x1437, x1480);
   uint32_t x1503;
   fiat_p384_uint1 x1504;
-  fiat_p384_addcarryx_u32(&x1503, &x1504, x1502, 0x0, x1439);
+  fiat_p384_addcarryx_u32(&x1503, &x1504, x1502, x1439, 0x0);
   uint32_t x1505;
   fiat_p384_uint1 x1506;
-  fiat_p384_addcarryx_u32(&x1505, &x1506, x1504, x1477, x1441);
+  fiat_p384_addcarryx_u32(&x1505, &x1506, x1504, x1441, x1477);
   uint32_t x1507;
   fiat_p384_uint1 x1508;
-  fiat_p384_addcarryx_u32(&x1507, &x1508, x1506, x1481, x1443);
+  fiat_p384_addcarryx_u32(&x1507, &x1508, x1506, x1443, x1481);
   uint32_t x1509;
   fiat_p384_uint1 x1510;
-  fiat_p384_addcarryx_u32(&x1509, &x1510, x1508, x1483, x1445);
+  fiat_p384_addcarryx_u32(&x1509, &x1510, x1508, x1445, x1483);
   uint32_t x1511;
   fiat_p384_uint1 x1512;
-  fiat_p384_addcarryx_u32(&x1511, &x1512, x1510, x1485, x1447);
+  fiat_p384_addcarryx_u32(&x1511, &x1512, x1510, x1447, x1485);
   uint32_t x1513;
   fiat_p384_uint1 x1514;
-  fiat_p384_addcarryx_u32(&x1513, &x1514, x1512, x1487, x1449);
+  fiat_p384_addcarryx_u32(&x1513, &x1514, x1512, x1449, x1487);
   uint32_t x1515;
   fiat_p384_uint1 x1516;
-  fiat_p384_addcarryx_u32(&x1515, &x1516, x1514, x1489, x1451);
+  fiat_p384_addcarryx_u32(&x1515, &x1516, x1514, x1451, x1489);
   uint32_t x1517;
   fiat_p384_uint1 x1518;
-  fiat_p384_addcarryx_u32(&x1517, &x1518, x1516, x1491, x1453);
+  fiat_p384_addcarryx_u32(&x1517, &x1518, x1516, x1453, x1491);
   uint32_t x1519;
   fiat_p384_uint1 x1520;
-  fiat_p384_addcarryx_u32(&x1519, &x1520, x1518, x1493, x1455);
+  fiat_p384_addcarryx_u32(&x1519, &x1520, x1518, x1455, x1493);
   uint32_t x1521;
   fiat_p384_uint1 x1522;
-  fiat_p384_addcarryx_u32(&x1521, &x1522, x1520, x1495, x1457);
+  fiat_p384_addcarryx_u32(&x1521, &x1522, x1520, x1457, x1495);
   uint32_t x1523;
   fiat_p384_uint1 x1524;
-  fiat_p384_addcarryx_u32(&x1523, &x1524, x1522, x1497, x1459);
+  fiat_p384_addcarryx_u32(&x1523, &x1524, x1522, x1459, x1497);
   uint32_t x1525;
   fiat_p384_uint1 x1526;
-  fiat_p384_addcarryx_u32(&x1525, &x1526, x1524, 0x0, x1460);
+  fiat_p384_addcarryx_u32(&x1525, &x1526, x1524, x1460, 0x0);
   uint32_t x1527;
   uint32_t x1528;
   fiat_p384_mulx_u32(&x1527, &x1528, x11, (arg2[11]));
@@ -2437,79 +2437,79 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x1549, &x1550, x11, (arg2[0]));
   uint32_t x1551;
   fiat_p384_uint1 x1552;
-  fiat_p384_addcarryx_u32(&x1551, &x1552, 0x0, x1547, x1550);
+  fiat_p384_addcarryx_u32(&x1551, &x1552, 0x0, x1550, x1547);
   uint32_t x1553;
   fiat_p384_uint1 x1554;
-  fiat_p384_addcarryx_u32(&x1553, &x1554, x1552, x1545, x1548);
+  fiat_p384_addcarryx_u32(&x1553, &x1554, x1552, x1548, x1545);
   uint32_t x1555;
   fiat_p384_uint1 x1556;
-  fiat_p384_addcarryx_u32(&x1555, &x1556, x1554, x1543, x1546);
+  fiat_p384_addcarryx_u32(&x1555, &x1556, x1554, x1546, x1543);
   uint32_t x1557;
   fiat_p384_uint1 x1558;
-  fiat_p384_addcarryx_u32(&x1557, &x1558, x1556, x1541, x1544);
+  fiat_p384_addcarryx_u32(&x1557, &x1558, x1556, x1544, x1541);
   uint32_t x1559;
   fiat_p384_uint1 x1560;
-  fiat_p384_addcarryx_u32(&x1559, &x1560, x1558, x1539, x1542);
+  fiat_p384_addcarryx_u32(&x1559, &x1560, x1558, x1542, x1539);
   uint32_t x1561;
   fiat_p384_uint1 x1562;
-  fiat_p384_addcarryx_u32(&x1561, &x1562, x1560, x1537, x1540);
+  fiat_p384_addcarryx_u32(&x1561, &x1562, x1560, x1540, x1537);
   uint32_t x1563;
   fiat_p384_uint1 x1564;
-  fiat_p384_addcarryx_u32(&x1563, &x1564, x1562, x1535, x1538);
+  fiat_p384_addcarryx_u32(&x1563, &x1564, x1562, x1538, x1535);
   uint32_t x1565;
   fiat_p384_uint1 x1566;
-  fiat_p384_addcarryx_u32(&x1565, &x1566, x1564, x1533, x1536);
+  fiat_p384_addcarryx_u32(&x1565, &x1566, x1564, x1536, x1533);
   uint32_t x1567;
   fiat_p384_uint1 x1568;
-  fiat_p384_addcarryx_u32(&x1567, &x1568, x1566, x1531, x1534);
+  fiat_p384_addcarryx_u32(&x1567, &x1568, x1566, x1534, x1531);
   uint32_t x1569;
   fiat_p384_uint1 x1570;
-  fiat_p384_addcarryx_u32(&x1569, &x1570, x1568, x1529, x1532);
+  fiat_p384_addcarryx_u32(&x1569, &x1570, x1568, x1532, x1529);
   uint32_t x1571;
   fiat_p384_uint1 x1572;
-  fiat_p384_addcarryx_u32(&x1571, &x1572, x1570, x1527, x1530);
+  fiat_p384_addcarryx_u32(&x1571, &x1572, x1570, x1530, x1527);
   uint32_t x1573;
   fiat_p384_uint1 x1574;
-  fiat_p384_addcarryx_u32(&x1573, &x1574, x1572, 0x0, x1528);
+  fiat_p384_addcarryx_u32(&x1573, &x1574, x1572, x1528, 0x0);
   uint32_t x1575;
   fiat_p384_uint1 x1576;
-  fiat_p384_addcarryx_u32(&x1575, &x1576, 0x0, x1549, x1501);
+  fiat_p384_addcarryx_u32(&x1575, &x1576, 0x0, x1501, x1549);
   uint32_t x1577;
   fiat_p384_uint1 x1578;
-  fiat_p384_addcarryx_u32(&x1577, &x1578, x1576, x1551, x1503);
+  fiat_p384_addcarryx_u32(&x1577, &x1578, x1576, x1503, x1551);
   uint32_t x1579;
   fiat_p384_uint1 x1580;
-  fiat_p384_addcarryx_u32(&x1579, &x1580, x1578, x1553, x1505);
+  fiat_p384_addcarryx_u32(&x1579, &x1580, x1578, x1505, x1553);
   uint32_t x1581;
   fiat_p384_uint1 x1582;
-  fiat_p384_addcarryx_u32(&x1581, &x1582, x1580, x1555, x1507);
+  fiat_p384_addcarryx_u32(&x1581, &x1582, x1580, x1507, x1555);
   uint32_t x1583;
   fiat_p384_uint1 x1584;
-  fiat_p384_addcarryx_u32(&x1583, &x1584, x1582, x1557, x1509);
+  fiat_p384_addcarryx_u32(&x1583, &x1584, x1582, x1509, x1557);
   uint32_t x1585;
   fiat_p384_uint1 x1586;
-  fiat_p384_addcarryx_u32(&x1585, &x1586, x1584, x1559, x1511);
+  fiat_p384_addcarryx_u32(&x1585, &x1586, x1584, x1511, x1559);
   uint32_t x1587;
   fiat_p384_uint1 x1588;
-  fiat_p384_addcarryx_u32(&x1587, &x1588, x1586, x1561, x1513);
+  fiat_p384_addcarryx_u32(&x1587, &x1588, x1586, x1513, x1561);
   uint32_t x1589;
   fiat_p384_uint1 x1590;
-  fiat_p384_addcarryx_u32(&x1589, &x1590, x1588, x1563, x1515);
+  fiat_p384_addcarryx_u32(&x1589, &x1590, x1588, x1515, x1563);
   uint32_t x1591;
   fiat_p384_uint1 x1592;
-  fiat_p384_addcarryx_u32(&x1591, &x1592, x1590, x1565, x1517);
+  fiat_p384_addcarryx_u32(&x1591, &x1592, x1590, x1517, x1565);
   uint32_t x1593;
   fiat_p384_uint1 x1594;
-  fiat_p384_addcarryx_u32(&x1593, &x1594, x1592, x1567, x1519);
+  fiat_p384_addcarryx_u32(&x1593, &x1594, x1592, x1519, x1567);
   uint32_t x1595;
   fiat_p384_uint1 x1596;
-  fiat_p384_addcarryx_u32(&x1595, &x1596, x1594, x1569, x1521);
+  fiat_p384_addcarryx_u32(&x1595, &x1596, x1594, x1521, x1569);
   uint32_t x1597;
   fiat_p384_uint1 x1598;
-  fiat_p384_addcarryx_u32(&x1597, &x1598, x1596, x1571, x1523);
+  fiat_p384_addcarryx_u32(&x1597, &x1598, x1596, x1523, x1571);
   uint32_t x1599;
   fiat_p384_uint1 x1600;
-  fiat_p384_addcarryx_u32(&x1599, &x1600, x1598, x1573, x1525);
+  fiat_p384_addcarryx_u32(&x1599, &x1600, x1598, x1525, x1573);
   uint32_t x1601;
   uint32_t x1602;
   fiat_p384_mulx_u32(&x1601, &x1602, x1575, UINT32_C(0xffffffff));
@@ -2542,73 +2542,73 @@ static void fiat_p384_mul(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_mulx_u32(&x1619, &x1620, x1575, UINT32_C(0xffffffff));
   uint32_t x1621;
   fiat_p384_uint1 x1622;
-  fiat_p384_addcarryx_u32(&x1621, &x1622, 0x0, x1615, x1618);
+  fiat_p384_addcarryx_u32(&x1621, &x1622, 0x0, x1618, x1615);
   uint32_t x1623;
   fiat_p384_uint1 x1624;
-  fiat_p384_addcarryx_u32(&x1623, &x1624, x1622, x1613, x1616);
+  fiat_p384_addcarryx_u32(&x1623, &x1624, x1622, x1616, x1613);
   uint32_t x1625;
   fiat_p384_uint1 x1626;
-  fiat_p384_addcarryx_u32(&x1625, &x1626, x1624, x1611, x1614);
+  fiat_p384_addcarryx_u32(&x1625, &x1626, x1624, x1614, x1611);
   uint32_t x1627;
   fiat_p384_uint1 x1628;
-  fiat_p384_addcarryx_u32(&x1627, &x1628, x1626, x1609, x1612);
+  fiat_p384_addcarryx_u32(&x1627, &x1628, x1626, x1612, x1609);
   uint32_t x1629;
   fiat_p384_uint1 x1630;
-  fiat_p384_addcarryx_u32(&x1629, &x1630, x1628, x1607, x1610);
+  fiat_p384_addcarryx_u32(&x1629, &x1630, x1628, x1610, x1607);
   uint32_t x1631;
   fiat_p384_uint1 x1632;
-  fiat_p384_addcarryx_u32(&x1631, &x1632, x1630, x1605, x1608);
+  fiat_p384_addcarryx_u32(&x1631, &x1632, x1630, x1608, x1605);
   uint32_t x1633;
   fiat_p384_uint1 x1634;
-  fiat_p384_addcarryx_u32(&x1633, &x1634, x1632, x1603, x1606);
+  fiat_p384_addcarryx_u32(&x1633, &x1634, x1632, x1606, x1603);
   uint32_t x1635;
   fiat_p384_uint1 x1636;
-  fiat_p384_addcarryx_u32(&x1635, &x1636, x1634, x1601, x1604);
+  fiat_p384_addcarryx_u32(&x1635, &x1636, x1634, x1604, x1601);
   uint32_t x1637;
   fiat_p384_uint1 x1638;
-  fiat_p384_addcarryx_u32(&x1637, &x1638, x1636, 0x0, x1602);
+  fiat_p384_addcarryx_u32(&x1637, &x1638, x1636, x1602, 0x0);
   uint32_t x1639;
   fiat_p384_uint1 x1640;
-  fiat_p384_addcarryx_u32(&x1639, &x1640, 0x0, x1619, x1575);
+  fiat_p384_addcarryx_u32(&x1639, &x1640, 0x0, x1575, x1619);
   uint32_t x1641;
   fiat_p384_uint1 x1642;
-  fiat_p384_addcarryx_u32(&x1641, &x1642, x1640, x1620, x1577);
+  fiat_p384_addcarryx_u32(&x1641, &x1642, x1640, x1577, x1620);
   uint32_t x1643;
   fiat_p384_uint1 x1644;
-  fiat_p384_addcarryx_u32(&x1643, &x1644, x1642, 0x0, x1579);
+  fiat_p384_addcarryx_u32(&x1643, &x1644, x1642, x1579, 0x0);
   uint32_t x1645;
   fiat_p384_uint1 x1646;
-  fiat_p384_addcarryx_u32(&x1645, &x1646, x1644, x1617, x1581);
+  fiat_p384_addcarryx_u32(&x1645, &x1646, x1644, x1581, x1617);
   uint32_t x1647;
   fiat_p384_uint1 x1648;
-  fiat_p384_addcarryx_u32(&x1647, &x1648, x1646, x1621, x1583);
+  fiat_p384_addcarryx_u32(&x1647, &x1648, x1646, x1583, x1621);
   uint32_t x1649;
   fiat_p384_uint1 x1650;
-  fiat_p384_addcarryx_u32(&x1649, &x1650, x1648, x1623, x1585);
+  fiat_p384_addcarryx_u32(&x1649, &x1650, x1648, x1585, x1623);
   uint32_t x1651;
   fiat_p384_uint1 x1652;
-  fiat_p384_addcarryx_u32(&x1651, &x1652, x1650, x1625, x1587);
+  fiat_p384_addcarryx_u32(&x1651, &x1652, x1650, x1587, x1625);
   uint32_t x1653;
   fiat_p384_uint1 x1654;
-  fiat_p384_addcarryx_u32(&x1653, &x1654, x1652, x1627, x1589);
+  fiat_p384_addcarryx_u32(&x1653, &x1654, x1652, x1589, x1627);
   uint32_t x1655;
   fiat_p384_uint1 x1656;
-  fiat_p384_addcarryx_u32(&x1655, &x1656, x1654, x1629, x1591);
+  fiat_p384_addcarryx_u32(&x1655, &x1656, x1654, x1591, x1629);
   uint32_t x1657;
   fiat_p384_uint1 x1658;
-  fiat_p384_addcarryx_u32(&x1657, &x1658, x1656, x1631, x1593);
+  fiat_p384_addcarryx_u32(&x1657, &x1658, x1656, x1593, x1631);
   uint32_t x1659;
   fiat_p384_uint1 x1660;
-  fiat_p384_addcarryx_u32(&x1659, &x1660, x1658, x1633, x1595);
+  fiat_p384_addcarryx_u32(&x1659, &x1660, x1658, x1595, x1633);
   uint32_t x1661;
   fiat_p384_uint1 x1662;
-  fiat_p384_addcarryx_u32(&x1661, &x1662, x1660, x1635, x1597);
+  fiat_p384_addcarryx_u32(&x1661, &x1662, x1660, x1597, x1635);
   uint32_t x1663;
   fiat_p384_uint1 x1664;
-  fiat_p384_addcarryx_u32(&x1663, &x1664, x1662, x1637, x1599);
+  fiat_p384_addcarryx_u32(&x1663, &x1664, x1662, x1599, x1637);
   uint32_t x1665;
   fiat_p384_uint1 x1666;
-  fiat_p384_addcarryx_u32(&x1665, &x1666, x1664, 0x0, x1600);
+  fiat_p384_addcarryx_u32(&x1665, &x1666, x1664, x1600, 0x0);
   uint32_t x1667;
   fiat_p384_uint1 x1668;
   fiat_p384_subborrowx_u32(&x1667, &x1668, 0x0, x1641, UINT32_C(0xffffffff));
@@ -2750,40 +2750,40 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x35, &x36, x12, (arg1[0]));
   uint32_t x37;
   fiat_p384_uint1 x38;
-  fiat_p384_addcarryx_u32(&x37, &x38, 0x0, x33, x36);
+  fiat_p384_addcarryx_u32(&x37, &x38, 0x0, x36, x33);
   uint32_t x39;
   fiat_p384_uint1 x40;
-  fiat_p384_addcarryx_u32(&x39, &x40, x38, x31, x34);
+  fiat_p384_addcarryx_u32(&x39, &x40, x38, x34, x31);
   uint32_t x41;
   fiat_p384_uint1 x42;
-  fiat_p384_addcarryx_u32(&x41, &x42, x40, x29, x32);
+  fiat_p384_addcarryx_u32(&x41, &x42, x40, x32, x29);
   uint32_t x43;
   fiat_p384_uint1 x44;
-  fiat_p384_addcarryx_u32(&x43, &x44, x42, x27, x30);
+  fiat_p384_addcarryx_u32(&x43, &x44, x42, x30, x27);
   uint32_t x45;
   fiat_p384_uint1 x46;
-  fiat_p384_addcarryx_u32(&x45, &x46, x44, x25, x28);
+  fiat_p384_addcarryx_u32(&x45, &x46, x44, x28, x25);
   uint32_t x47;
   fiat_p384_uint1 x48;
-  fiat_p384_addcarryx_u32(&x47, &x48, x46, x23, x26);
+  fiat_p384_addcarryx_u32(&x47, &x48, x46, x26, x23);
   uint32_t x49;
   fiat_p384_uint1 x50;
-  fiat_p384_addcarryx_u32(&x49, &x50, x48, x21, x24);
+  fiat_p384_addcarryx_u32(&x49, &x50, x48, x24, x21);
   uint32_t x51;
   fiat_p384_uint1 x52;
-  fiat_p384_addcarryx_u32(&x51, &x52, x50, x19, x22);
+  fiat_p384_addcarryx_u32(&x51, &x52, x50, x22, x19);
   uint32_t x53;
   fiat_p384_uint1 x54;
-  fiat_p384_addcarryx_u32(&x53, &x54, x52, x17, x20);
+  fiat_p384_addcarryx_u32(&x53, &x54, x52, x20, x17);
   uint32_t x55;
   fiat_p384_uint1 x56;
-  fiat_p384_addcarryx_u32(&x55, &x56, x54, x15, x18);
+  fiat_p384_addcarryx_u32(&x55, &x56, x54, x18, x15);
   uint32_t x57;
   fiat_p384_uint1 x58;
-  fiat_p384_addcarryx_u32(&x57, &x58, x56, x13, x16);
+  fiat_p384_addcarryx_u32(&x57, &x58, x56, x16, x13);
   uint32_t x59;
   fiat_p384_uint1 x60;
-  fiat_p384_addcarryx_u32(&x59, &x60, x58, 0x0, x14);
+  fiat_p384_addcarryx_u32(&x59, &x60, x58, x14, 0x0);
   uint32_t x61;
   uint32_t x62;
   fiat_p384_mulx_u32(&x61, &x62, x35, UINT32_C(0xffffffff));
@@ -2816,70 +2816,70 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x79, &x80, x35, UINT32_C(0xffffffff));
   uint32_t x81;
   fiat_p384_uint1 x82;
-  fiat_p384_addcarryx_u32(&x81, &x82, 0x0, x75, x78);
+  fiat_p384_addcarryx_u32(&x81, &x82, 0x0, x78, x75);
   uint32_t x83;
   fiat_p384_uint1 x84;
-  fiat_p384_addcarryx_u32(&x83, &x84, x82, x73, x76);
+  fiat_p384_addcarryx_u32(&x83, &x84, x82, x76, x73);
   uint32_t x85;
   fiat_p384_uint1 x86;
-  fiat_p384_addcarryx_u32(&x85, &x86, x84, x71, x74);
+  fiat_p384_addcarryx_u32(&x85, &x86, x84, x74, x71);
   uint32_t x87;
   fiat_p384_uint1 x88;
-  fiat_p384_addcarryx_u32(&x87, &x88, x86, x69, x72);
+  fiat_p384_addcarryx_u32(&x87, &x88, x86, x72, x69);
   uint32_t x89;
   fiat_p384_uint1 x90;
-  fiat_p384_addcarryx_u32(&x89, &x90, x88, x67, x70);
+  fiat_p384_addcarryx_u32(&x89, &x90, x88, x70, x67);
   uint32_t x91;
   fiat_p384_uint1 x92;
-  fiat_p384_addcarryx_u32(&x91, &x92, x90, x65, x68);
+  fiat_p384_addcarryx_u32(&x91, &x92, x90, x68, x65);
   uint32_t x93;
   fiat_p384_uint1 x94;
-  fiat_p384_addcarryx_u32(&x93, &x94, x92, x63, x66);
+  fiat_p384_addcarryx_u32(&x93, &x94, x92, x66, x63);
   uint32_t x95;
   fiat_p384_uint1 x96;
-  fiat_p384_addcarryx_u32(&x95, &x96, x94, x61, x64);
+  fiat_p384_addcarryx_u32(&x95, &x96, x94, x64, x61);
   uint32_t x97;
   fiat_p384_uint1 x98;
-  fiat_p384_addcarryx_u32(&x97, &x98, x96, 0x0, x62);
+  fiat_p384_addcarryx_u32(&x97, &x98, x96, x62, 0x0);
   uint32_t x99;
   fiat_p384_uint1 x100;
-  fiat_p384_addcarryx_u32(&x99, &x100, 0x0, x79, x35);
+  fiat_p384_addcarryx_u32(&x99, &x100, 0x0, x35, x79);
   uint32_t x101;
   fiat_p384_uint1 x102;
-  fiat_p384_addcarryx_u32(&x101, &x102, x100, x80, x37);
+  fiat_p384_addcarryx_u32(&x101, &x102, x100, x37, x80);
   uint32_t x103;
   fiat_p384_uint1 x104;
-  fiat_p384_addcarryx_u32(&x103, &x104, x102, 0x0, x39);
+  fiat_p384_addcarryx_u32(&x103, &x104, x102, x39, 0x0);
   uint32_t x105;
   fiat_p384_uint1 x106;
-  fiat_p384_addcarryx_u32(&x105, &x106, x104, x77, x41);
+  fiat_p384_addcarryx_u32(&x105, &x106, x104, x41, x77);
   uint32_t x107;
   fiat_p384_uint1 x108;
-  fiat_p384_addcarryx_u32(&x107, &x108, x106, x81, x43);
+  fiat_p384_addcarryx_u32(&x107, &x108, x106, x43, x81);
   uint32_t x109;
   fiat_p384_uint1 x110;
-  fiat_p384_addcarryx_u32(&x109, &x110, x108, x83, x45);
+  fiat_p384_addcarryx_u32(&x109, &x110, x108, x45, x83);
   uint32_t x111;
   fiat_p384_uint1 x112;
-  fiat_p384_addcarryx_u32(&x111, &x112, x110, x85, x47);
+  fiat_p384_addcarryx_u32(&x111, &x112, x110, x47, x85);
   uint32_t x113;
   fiat_p384_uint1 x114;
-  fiat_p384_addcarryx_u32(&x113, &x114, x112, x87, x49);
+  fiat_p384_addcarryx_u32(&x113, &x114, x112, x49, x87);
   uint32_t x115;
   fiat_p384_uint1 x116;
-  fiat_p384_addcarryx_u32(&x115, &x116, x114, x89, x51);
+  fiat_p384_addcarryx_u32(&x115, &x116, x114, x51, x89);
   uint32_t x117;
   fiat_p384_uint1 x118;
-  fiat_p384_addcarryx_u32(&x117, &x118, x116, x91, x53);
+  fiat_p384_addcarryx_u32(&x117, &x118, x116, x53, x91);
   uint32_t x119;
   fiat_p384_uint1 x120;
-  fiat_p384_addcarryx_u32(&x119, &x120, x118, x93, x55);
+  fiat_p384_addcarryx_u32(&x119, &x120, x118, x55, x93);
   uint32_t x121;
   fiat_p384_uint1 x122;
-  fiat_p384_addcarryx_u32(&x121, &x122, x120, x95, x57);
+  fiat_p384_addcarryx_u32(&x121, &x122, x120, x57, x95);
   uint32_t x123;
   fiat_p384_uint1 x124;
-  fiat_p384_addcarryx_u32(&x123, &x124, x122, x97, x59);
+  fiat_p384_addcarryx_u32(&x123, &x124, x122, x59, x97);
   uint32_t x125;
   fiat_p384_uint1 x126;
   fiat_p384_addcarryx_u32(&x125, &x126, x124, 0x0, 0x0);
@@ -2921,79 +2921,79 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x149, &x150, x1, (arg1[0]));
   uint32_t x151;
   fiat_p384_uint1 x152;
-  fiat_p384_addcarryx_u32(&x151, &x152, 0x0, x147, x150);
+  fiat_p384_addcarryx_u32(&x151, &x152, 0x0, x150, x147);
   uint32_t x153;
   fiat_p384_uint1 x154;
-  fiat_p384_addcarryx_u32(&x153, &x154, x152, x145, x148);
+  fiat_p384_addcarryx_u32(&x153, &x154, x152, x148, x145);
   uint32_t x155;
   fiat_p384_uint1 x156;
-  fiat_p384_addcarryx_u32(&x155, &x156, x154, x143, x146);
+  fiat_p384_addcarryx_u32(&x155, &x156, x154, x146, x143);
   uint32_t x157;
   fiat_p384_uint1 x158;
-  fiat_p384_addcarryx_u32(&x157, &x158, x156, x141, x144);
+  fiat_p384_addcarryx_u32(&x157, &x158, x156, x144, x141);
   uint32_t x159;
   fiat_p384_uint1 x160;
-  fiat_p384_addcarryx_u32(&x159, &x160, x158, x139, x142);
+  fiat_p384_addcarryx_u32(&x159, &x160, x158, x142, x139);
   uint32_t x161;
   fiat_p384_uint1 x162;
-  fiat_p384_addcarryx_u32(&x161, &x162, x160, x137, x140);
+  fiat_p384_addcarryx_u32(&x161, &x162, x160, x140, x137);
   uint32_t x163;
   fiat_p384_uint1 x164;
-  fiat_p384_addcarryx_u32(&x163, &x164, x162, x135, x138);
+  fiat_p384_addcarryx_u32(&x163, &x164, x162, x138, x135);
   uint32_t x165;
   fiat_p384_uint1 x166;
-  fiat_p384_addcarryx_u32(&x165, &x166, x164, x133, x136);
+  fiat_p384_addcarryx_u32(&x165, &x166, x164, x136, x133);
   uint32_t x167;
   fiat_p384_uint1 x168;
-  fiat_p384_addcarryx_u32(&x167, &x168, x166, x131, x134);
+  fiat_p384_addcarryx_u32(&x167, &x168, x166, x134, x131);
   uint32_t x169;
   fiat_p384_uint1 x170;
-  fiat_p384_addcarryx_u32(&x169, &x170, x168, x129, x132);
+  fiat_p384_addcarryx_u32(&x169, &x170, x168, x132, x129);
   uint32_t x171;
   fiat_p384_uint1 x172;
-  fiat_p384_addcarryx_u32(&x171, &x172, x170, x127, x130);
+  fiat_p384_addcarryx_u32(&x171, &x172, x170, x130, x127);
   uint32_t x173;
   fiat_p384_uint1 x174;
-  fiat_p384_addcarryx_u32(&x173, &x174, x172, 0x0, x128);
+  fiat_p384_addcarryx_u32(&x173, &x174, x172, x128, 0x0);
   uint32_t x175;
   fiat_p384_uint1 x176;
-  fiat_p384_addcarryx_u32(&x175, &x176, 0x0, x149, x101);
+  fiat_p384_addcarryx_u32(&x175, &x176, 0x0, x101, x149);
   uint32_t x177;
   fiat_p384_uint1 x178;
-  fiat_p384_addcarryx_u32(&x177, &x178, x176, x151, x103);
+  fiat_p384_addcarryx_u32(&x177, &x178, x176, x103, x151);
   uint32_t x179;
   fiat_p384_uint1 x180;
-  fiat_p384_addcarryx_u32(&x179, &x180, x178, x153, x105);
+  fiat_p384_addcarryx_u32(&x179, &x180, x178, x105, x153);
   uint32_t x181;
   fiat_p384_uint1 x182;
-  fiat_p384_addcarryx_u32(&x181, &x182, x180, x155, x107);
+  fiat_p384_addcarryx_u32(&x181, &x182, x180, x107, x155);
   uint32_t x183;
   fiat_p384_uint1 x184;
-  fiat_p384_addcarryx_u32(&x183, &x184, x182, x157, x109);
+  fiat_p384_addcarryx_u32(&x183, &x184, x182, x109, x157);
   uint32_t x185;
   fiat_p384_uint1 x186;
-  fiat_p384_addcarryx_u32(&x185, &x186, x184, x159, x111);
+  fiat_p384_addcarryx_u32(&x185, &x186, x184, x111, x159);
   uint32_t x187;
   fiat_p384_uint1 x188;
-  fiat_p384_addcarryx_u32(&x187, &x188, x186, x161, x113);
+  fiat_p384_addcarryx_u32(&x187, &x188, x186, x113, x161);
   uint32_t x189;
   fiat_p384_uint1 x190;
-  fiat_p384_addcarryx_u32(&x189, &x190, x188, x163, x115);
+  fiat_p384_addcarryx_u32(&x189, &x190, x188, x115, x163);
   uint32_t x191;
   fiat_p384_uint1 x192;
-  fiat_p384_addcarryx_u32(&x191, &x192, x190, x165, x117);
+  fiat_p384_addcarryx_u32(&x191, &x192, x190, x117, x165);
   uint32_t x193;
   fiat_p384_uint1 x194;
-  fiat_p384_addcarryx_u32(&x193, &x194, x192, x167, x119);
+  fiat_p384_addcarryx_u32(&x193, &x194, x192, x119, x167);
   uint32_t x195;
   fiat_p384_uint1 x196;
-  fiat_p384_addcarryx_u32(&x195, &x196, x194, x169, x121);
+  fiat_p384_addcarryx_u32(&x195, &x196, x194, x121, x169);
   uint32_t x197;
   fiat_p384_uint1 x198;
-  fiat_p384_addcarryx_u32(&x197, &x198, x196, x171, x123);
+  fiat_p384_addcarryx_u32(&x197, &x198, x196, x123, x171);
   uint32_t x199;
   fiat_p384_uint1 x200;
-  fiat_p384_addcarryx_u32(&x199, &x200, x198, x173, (fiat_p384_uint1)x125);
+  fiat_p384_addcarryx_u32(&x199, &x200, x198, (fiat_p384_uint1)x125, x173);
   uint32_t x201;
   uint32_t x202;
   fiat_p384_mulx_u32(&x201, &x202, x175, UINT32_C(0xffffffff));
@@ -3026,73 +3026,73 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x219, &x220, x175, UINT32_C(0xffffffff));
   uint32_t x221;
   fiat_p384_uint1 x222;
-  fiat_p384_addcarryx_u32(&x221, &x222, 0x0, x215, x218);
+  fiat_p384_addcarryx_u32(&x221, &x222, 0x0, x218, x215);
   uint32_t x223;
   fiat_p384_uint1 x224;
-  fiat_p384_addcarryx_u32(&x223, &x224, x222, x213, x216);
+  fiat_p384_addcarryx_u32(&x223, &x224, x222, x216, x213);
   uint32_t x225;
   fiat_p384_uint1 x226;
-  fiat_p384_addcarryx_u32(&x225, &x226, x224, x211, x214);
+  fiat_p384_addcarryx_u32(&x225, &x226, x224, x214, x211);
   uint32_t x227;
   fiat_p384_uint1 x228;
-  fiat_p384_addcarryx_u32(&x227, &x228, x226, x209, x212);
+  fiat_p384_addcarryx_u32(&x227, &x228, x226, x212, x209);
   uint32_t x229;
   fiat_p384_uint1 x230;
-  fiat_p384_addcarryx_u32(&x229, &x230, x228, x207, x210);
+  fiat_p384_addcarryx_u32(&x229, &x230, x228, x210, x207);
   uint32_t x231;
   fiat_p384_uint1 x232;
-  fiat_p384_addcarryx_u32(&x231, &x232, x230, x205, x208);
+  fiat_p384_addcarryx_u32(&x231, &x232, x230, x208, x205);
   uint32_t x233;
   fiat_p384_uint1 x234;
-  fiat_p384_addcarryx_u32(&x233, &x234, x232, x203, x206);
+  fiat_p384_addcarryx_u32(&x233, &x234, x232, x206, x203);
   uint32_t x235;
   fiat_p384_uint1 x236;
-  fiat_p384_addcarryx_u32(&x235, &x236, x234, x201, x204);
+  fiat_p384_addcarryx_u32(&x235, &x236, x234, x204, x201);
   uint32_t x237;
   fiat_p384_uint1 x238;
-  fiat_p384_addcarryx_u32(&x237, &x238, x236, 0x0, x202);
+  fiat_p384_addcarryx_u32(&x237, &x238, x236, x202, 0x0);
   uint32_t x239;
   fiat_p384_uint1 x240;
-  fiat_p384_addcarryx_u32(&x239, &x240, 0x0, x219, x175);
+  fiat_p384_addcarryx_u32(&x239, &x240, 0x0, x175, x219);
   uint32_t x241;
   fiat_p384_uint1 x242;
-  fiat_p384_addcarryx_u32(&x241, &x242, x240, x220, x177);
+  fiat_p384_addcarryx_u32(&x241, &x242, x240, x177, x220);
   uint32_t x243;
   fiat_p384_uint1 x244;
-  fiat_p384_addcarryx_u32(&x243, &x244, x242, 0x0, x179);
+  fiat_p384_addcarryx_u32(&x243, &x244, x242, x179, 0x0);
   uint32_t x245;
   fiat_p384_uint1 x246;
-  fiat_p384_addcarryx_u32(&x245, &x246, x244, x217, x181);
+  fiat_p384_addcarryx_u32(&x245, &x246, x244, x181, x217);
   uint32_t x247;
   fiat_p384_uint1 x248;
-  fiat_p384_addcarryx_u32(&x247, &x248, x246, x221, x183);
+  fiat_p384_addcarryx_u32(&x247, &x248, x246, x183, x221);
   uint32_t x249;
   fiat_p384_uint1 x250;
-  fiat_p384_addcarryx_u32(&x249, &x250, x248, x223, x185);
+  fiat_p384_addcarryx_u32(&x249, &x250, x248, x185, x223);
   uint32_t x251;
   fiat_p384_uint1 x252;
-  fiat_p384_addcarryx_u32(&x251, &x252, x250, x225, x187);
+  fiat_p384_addcarryx_u32(&x251, &x252, x250, x187, x225);
   uint32_t x253;
   fiat_p384_uint1 x254;
-  fiat_p384_addcarryx_u32(&x253, &x254, x252, x227, x189);
+  fiat_p384_addcarryx_u32(&x253, &x254, x252, x189, x227);
   uint32_t x255;
   fiat_p384_uint1 x256;
-  fiat_p384_addcarryx_u32(&x255, &x256, x254, x229, x191);
+  fiat_p384_addcarryx_u32(&x255, &x256, x254, x191, x229);
   uint32_t x257;
   fiat_p384_uint1 x258;
-  fiat_p384_addcarryx_u32(&x257, &x258, x256, x231, x193);
+  fiat_p384_addcarryx_u32(&x257, &x258, x256, x193, x231);
   uint32_t x259;
   fiat_p384_uint1 x260;
-  fiat_p384_addcarryx_u32(&x259, &x260, x258, x233, x195);
+  fiat_p384_addcarryx_u32(&x259, &x260, x258, x195, x233);
   uint32_t x261;
   fiat_p384_uint1 x262;
-  fiat_p384_addcarryx_u32(&x261, &x262, x260, x235, x197);
+  fiat_p384_addcarryx_u32(&x261, &x262, x260, x197, x235);
   uint32_t x263;
   fiat_p384_uint1 x264;
-  fiat_p384_addcarryx_u32(&x263, &x264, x262, x237, x199);
+  fiat_p384_addcarryx_u32(&x263, &x264, x262, x199, x237);
   uint32_t x265;
   fiat_p384_uint1 x266;
-  fiat_p384_addcarryx_u32(&x265, &x266, x264, 0x0, x200);
+  fiat_p384_addcarryx_u32(&x265, &x266, x264, x200, 0x0);
   uint32_t x267;
   uint32_t x268;
   fiat_p384_mulx_u32(&x267, &x268, x2, (arg1[11]));
@@ -3131,79 +3131,79 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x289, &x290, x2, (arg1[0]));
   uint32_t x291;
   fiat_p384_uint1 x292;
-  fiat_p384_addcarryx_u32(&x291, &x292, 0x0, x287, x290);
+  fiat_p384_addcarryx_u32(&x291, &x292, 0x0, x290, x287);
   uint32_t x293;
   fiat_p384_uint1 x294;
-  fiat_p384_addcarryx_u32(&x293, &x294, x292, x285, x288);
+  fiat_p384_addcarryx_u32(&x293, &x294, x292, x288, x285);
   uint32_t x295;
   fiat_p384_uint1 x296;
-  fiat_p384_addcarryx_u32(&x295, &x296, x294, x283, x286);
+  fiat_p384_addcarryx_u32(&x295, &x296, x294, x286, x283);
   uint32_t x297;
   fiat_p384_uint1 x298;
-  fiat_p384_addcarryx_u32(&x297, &x298, x296, x281, x284);
+  fiat_p384_addcarryx_u32(&x297, &x298, x296, x284, x281);
   uint32_t x299;
   fiat_p384_uint1 x300;
-  fiat_p384_addcarryx_u32(&x299, &x300, x298, x279, x282);
+  fiat_p384_addcarryx_u32(&x299, &x300, x298, x282, x279);
   uint32_t x301;
   fiat_p384_uint1 x302;
-  fiat_p384_addcarryx_u32(&x301, &x302, x300, x277, x280);
+  fiat_p384_addcarryx_u32(&x301, &x302, x300, x280, x277);
   uint32_t x303;
   fiat_p384_uint1 x304;
-  fiat_p384_addcarryx_u32(&x303, &x304, x302, x275, x278);
+  fiat_p384_addcarryx_u32(&x303, &x304, x302, x278, x275);
   uint32_t x305;
   fiat_p384_uint1 x306;
-  fiat_p384_addcarryx_u32(&x305, &x306, x304, x273, x276);
+  fiat_p384_addcarryx_u32(&x305, &x306, x304, x276, x273);
   uint32_t x307;
   fiat_p384_uint1 x308;
-  fiat_p384_addcarryx_u32(&x307, &x308, x306, x271, x274);
+  fiat_p384_addcarryx_u32(&x307, &x308, x306, x274, x271);
   uint32_t x309;
   fiat_p384_uint1 x310;
-  fiat_p384_addcarryx_u32(&x309, &x310, x308, x269, x272);
+  fiat_p384_addcarryx_u32(&x309, &x310, x308, x272, x269);
   uint32_t x311;
   fiat_p384_uint1 x312;
-  fiat_p384_addcarryx_u32(&x311, &x312, x310, x267, x270);
+  fiat_p384_addcarryx_u32(&x311, &x312, x310, x270, x267);
   uint32_t x313;
   fiat_p384_uint1 x314;
-  fiat_p384_addcarryx_u32(&x313, &x314, x312, 0x0, x268);
+  fiat_p384_addcarryx_u32(&x313, &x314, x312, x268, 0x0);
   uint32_t x315;
   fiat_p384_uint1 x316;
-  fiat_p384_addcarryx_u32(&x315, &x316, 0x0, x289, x241);
+  fiat_p384_addcarryx_u32(&x315, &x316, 0x0, x241, x289);
   uint32_t x317;
   fiat_p384_uint1 x318;
-  fiat_p384_addcarryx_u32(&x317, &x318, x316, x291, x243);
+  fiat_p384_addcarryx_u32(&x317, &x318, x316, x243, x291);
   uint32_t x319;
   fiat_p384_uint1 x320;
-  fiat_p384_addcarryx_u32(&x319, &x320, x318, x293, x245);
+  fiat_p384_addcarryx_u32(&x319, &x320, x318, x245, x293);
   uint32_t x321;
   fiat_p384_uint1 x322;
-  fiat_p384_addcarryx_u32(&x321, &x322, x320, x295, x247);
+  fiat_p384_addcarryx_u32(&x321, &x322, x320, x247, x295);
   uint32_t x323;
   fiat_p384_uint1 x324;
-  fiat_p384_addcarryx_u32(&x323, &x324, x322, x297, x249);
+  fiat_p384_addcarryx_u32(&x323, &x324, x322, x249, x297);
   uint32_t x325;
   fiat_p384_uint1 x326;
-  fiat_p384_addcarryx_u32(&x325, &x326, x324, x299, x251);
+  fiat_p384_addcarryx_u32(&x325, &x326, x324, x251, x299);
   uint32_t x327;
   fiat_p384_uint1 x328;
-  fiat_p384_addcarryx_u32(&x327, &x328, x326, x301, x253);
+  fiat_p384_addcarryx_u32(&x327, &x328, x326, x253, x301);
   uint32_t x329;
   fiat_p384_uint1 x330;
-  fiat_p384_addcarryx_u32(&x329, &x330, x328, x303, x255);
+  fiat_p384_addcarryx_u32(&x329, &x330, x328, x255, x303);
   uint32_t x331;
   fiat_p384_uint1 x332;
-  fiat_p384_addcarryx_u32(&x331, &x332, x330, x305, x257);
+  fiat_p384_addcarryx_u32(&x331, &x332, x330, x257, x305);
   uint32_t x333;
   fiat_p384_uint1 x334;
-  fiat_p384_addcarryx_u32(&x333, &x334, x332, x307, x259);
+  fiat_p384_addcarryx_u32(&x333, &x334, x332, x259, x307);
   uint32_t x335;
   fiat_p384_uint1 x336;
-  fiat_p384_addcarryx_u32(&x335, &x336, x334, x309, x261);
+  fiat_p384_addcarryx_u32(&x335, &x336, x334, x261, x309);
   uint32_t x337;
   fiat_p384_uint1 x338;
-  fiat_p384_addcarryx_u32(&x337, &x338, x336, x311, x263);
+  fiat_p384_addcarryx_u32(&x337, &x338, x336, x263, x311);
   uint32_t x339;
   fiat_p384_uint1 x340;
-  fiat_p384_addcarryx_u32(&x339, &x340, x338, x313, x265);
+  fiat_p384_addcarryx_u32(&x339, &x340, x338, x265, x313);
   uint32_t x341;
   uint32_t x342;
   fiat_p384_mulx_u32(&x341, &x342, x315, UINT32_C(0xffffffff));
@@ -3236,73 +3236,73 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x359, &x360, x315, UINT32_C(0xffffffff));
   uint32_t x361;
   fiat_p384_uint1 x362;
-  fiat_p384_addcarryx_u32(&x361, &x362, 0x0, x355, x358);
+  fiat_p384_addcarryx_u32(&x361, &x362, 0x0, x358, x355);
   uint32_t x363;
   fiat_p384_uint1 x364;
-  fiat_p384_addcarryx_u32(&x363, &x364, x362, x353, x356);
+  fiat_p384_addcarryx_u32(&x363, &x364, x362, x356, x353);
   uint32_t x365;
   fiat_p384_uint1 x366;
-  fiat_p384_addcarryx_u32(&x365, &x366, x364, x351, x354);
+  fiat_p384_addcarryx_u32(&x365, &x366, x364, x354, x351);
   uint32_t x367;
   fiat_p384_uint1 x368;
-  fiat_p384_addcarryx_u32(&x367, &x368, x366, x349, x352);
+  fiat_p384_addcarryx_u32(&x367, &x368, x366, x352, x349);
   uint32_t x369;
   fiat_p384_uint1 x370;
-  fiat_p384_addcarryx_u32(&x369, &x370, x368, x347, x350);
+  fiat_p384_addcarryx_u32(&x369, &x370, x368, x350, x347);
   uint32_t x371;
   fiat_p384_uint1 x372;
-  fiat_p384_addcarryx_u32(&x371, &x372, x370, x345, x348);
+  fiat_p384_addcarryx_u32(&x371, &x372, x370, x348, x345);
   uint32_t x373;
   fiat_p384_uint1 x374;
-  fiat_p384_addcarryx_u32(&x373, &x374, x372, x343, x346);
+  fiat_p384_addcarryx_u32(&x373, &x374, x372, x346, x343);
   uint32_t x375;
   fiat_p384_uint1 x376;
-  fiat_p384_addcarryx_u32(&x375, &x376, x374, x341, x344);
+  fiat_p384_addcarryx_u32(&x375, &x376, x374, x344, x341);
   uint32_t x377;
   fiat_p384_uint1 x378;
-  fiat_p384_addcarryx_u32(&x377, &x378, x376, 0x0, x342);
+  fiat_p384_addcarryx_u32(&x377, &x378, x376, x342, 0x0);
   uint32_t x379;
   fiat_p384_uint1 x380;
-  fiat_p384_addcarryx_u32(&x379, &x380, 0x0, x359, x315);
+  fiat_p384_addcarryx_u32(&x379, &x380, 0x0, x315, x359);
   uint32_t x381;
   fiat_p384_uint1 x382;
-  fiat_p384_addcarryx_u32(&x381, &x382, x380, x360, x317);
+  fiat_p384_addcarryx_u32(&x381, &x382, x380, x317, x360);
   uint32_t x383;
   fiat_p384_uint1 x384;
-  fiat_p384_addcarryx_u32(&x383, &x384, x382, 0x0, x319);
+  fiat_p384_addcarryx_u32(&x383, &x384, x382, x319, 0x0);
   uint32_t x385;
   fiat_p384_uint1 x386;
-  fiat_p384_addcarryx_u32(&x385, &x386, x384, x357, x321);
+  fiat_p384_addcarryx_u32(&x385, &x386, x384, x321, x357);
   uint32_t x387;
   fiat_p384_uint1 x388;
-  fiat_p384_addcarryx_u32(&x387, &x388, x386, x361, x323);
+  fiat_p384_addcarryx_u32(&x387, &x388, x386, x323, x361);
   uint32_t x389;
   fiat_p384_uint1 x390;
-  fiat_p384_addcarryx_u32(&x389, &x390, x388, x363, x325);
+  fiat_p384_addcarryx_u32(&x389, &x390, x388, x325, x363);
   uint32_t x391;
   fiat_p384_uint1 x392;
-  fiat_p384_addcarryx_u32(&x391, &x392, x390, x365, x327);
+  fiat_p384_addcarryx_u32(&x391, &x392, x390, x327, x365);
   uint32_t x393;
   fiat_p384_uint1 x394;
-  fiat_p384_addcarryx_u32(&x393, &x394, x392, x367, x329);
+  fiat_p384_addcarryx_u32(&x393, &x394, x392, x329, x367);
   uint32_t x395;
   fiat_p384_uint1 x396;
-  fiat_p384_addcarryx_u32(&x395, &x396, x394, x369, x331);
+  fiat_p384_addcarryx_u32(&x395, &x396, x394, x331, x369);
   uint32_t x397;
   fiat_p384_uint1 x398;
-  fiat_p384_addcarryx_u32(&x397, &x398, x396, x371, x333);
+  fiat_p384_addcarryx_u32(&x397, &x398, x396, x333, x371);
   uint32_t x399;
   fiat_p384_uint1 x400;
-  fiat_p384_addcarryx_u32(&x399, &x400, x398, x373, x335);
+  fiat_p384_addcarryx_u32(&x399, &x400, x398, x335, x373);
   uint32_t x401;
   fiat_p384_uint1 x402;
-  fiat_p384_addcarryx_u32(&x401, &x402, x400, x375, x337);
+  fiat_p384_addcarryx_u32(&x401, &x402, x400, x337, x375);
   uint32_t x403;
   fiat_p384_uint1 x404;
-  fiat_p384_addcarryx_u32(&x403, &x404, x402, x377, x339);
+  fiat_p384_addcarryx_u32(&x403, &x404, x402, x339, x377);
   uint32_t x405;
   fiat_p384_uint1 x406;
-  fiat_p384_addcarryx_u32(&x405, &x406, x404, 0x0, x340);
+  fiat_p384_addcarryx_u32(&x405, &x406, x404, x340, 0x0);
   uint32_t x407;
   uint32_t x408;
   fiat_p384_mulx_u32(&x407, &x408, x3, (arg1[11]));
@@ -3341,79 +3341,79 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x429, &x430, x3, (arg1[0]));
   uint32_t x431;
   fiat_p384_uint1 x432;
-  fiat_p384_addcarryx_u32(&x431, &x432, 0x0, x427, x430);
+  fiat_p384_addcarryx_u32(&x431, &x432, 0x0, x430, x427);
   uint32_t x433;
   fiat_p384_uint1 x434;
-  fiat_p384_addcarryx_u32(&x433, &x434, x432, x425, x428);
+  fiat_p384_addcarryx_u32(&x433, &x434, x432, x428, x425);
   uint32_t x435;
   fiat_p384_uint1 x436;
-  fiat_p384_addcarryx_u32(&x435, &x436, x434, x423, x426);
+  fiat_p384_addcarryx_u32(&x435, &x436, x434, x426, x423);
   uint32_t x437;
   fiat_p384_uint1 x438;
-  fiat_p384_addcarryx_u32(&x437, &x438, x436, x421, x424);
+  fiat_p384_addcarryx_u32(&x437, &x438, x436, x424, x421);
   uint32_t x439;
   fiat_p384_uint1 x440;
-  fiat_p384_addcarryx_u32(&x439, &x440, x438, x419, x422);
+  fiat_p384_addcarryx_u32(&x439, &x440, x438, x422, x419);
   uint32_t x441;
   fiat_p384_uint1 x442;
-  fiat_p384_addcarryx_u32(&x441, &x442, x440, x417, x420);
+  fiat_p384_addcarryx_u32(&x441, &x442, x440, x420, x417);
   uint32_t x443;
   fiat_p384_uint1 x444;
-  fiat_p384_addcarryx_u32(&x443, &x444, x442, x415, x418);
+  fiat_p384_addcarryx_u32(&x443, &x444, x442, x418, x415);
   uint32_t x445;
   fiat_p384_uint1 x446;
-  fiat_p384_addcarryx_u32(&x445, &x446, x444, x413, x416);
+  fiat_p384_addcarryx_u32(&x445, &x446, x444, x416, x413);
   uint32_t x447;
   fiat_p384_uint1 x448;
-  fiat_p384_addcarryx_u32(&x447, &x448, x446, x411, x414);
+  fiat_p384_addcarryx_u32(&x447, &x448, x446, x414, x411);
   uint32_t x449;
   fiat_p384_uint1 x450;
-  fiat_p384_addcarryx_u32(&x449, &x450, x448, x409, x412);
+  fiat_p384_addcarryx_u32(&x449, &x450, x448, x412, x409);
   uint32_t x451;
   fiat_p384_uint1 x452;
-  fiat_p384_addcarryx_u32(&x451, &x452, x450, x407, x410);
+  fiat_p384_addcarryx_u32(&x451, &x452, x450, x410, x407);
   uint32_t x453;
   fiat_p384_uint1 x454;
-  fiat_p384_addcarryx_u32(&x453, &x454, x452, 0x0, x408);
+  fiat_p384_addcarryx_u32(&x453, &x454, x452, x408, 0x0);
   uint32_t x455;
   fiat_p384_uint1 x456;
-  fiat_p384_addcarryx_u32(&x455, &x456, 0x0, x429, x381);
+  fiat_p384_addcarryx_u32(&x455, &x456, 0x0, x381, x429);
   uint32_t x457;
   fiat_p384_uint1 x458;
-  fiat_p384_addcarryx_u32(&x457, &x458, x456, x431, x383);
+  fiat_p384_addcarryx_u32(&x457, &x458, x456, x383, x431);
   uint32_t x459;
   fiat_p384_uint1 x460;
-  fiat_p384_addcarryx_u32(&x459, &x460, x458, x433, x385);
+  fiat_p384_addcarryx_u32(&x459, &x460, x458, x385, x433);
   uint32_t x461;
   fiat_p384_uint1 x462;
-  fiat_p384_addcarryx_u32(&x461, &x462, x460, x435, x387);
+  fiat_p384_addcarryx_u32(&x461, &x462, x460, x387, x435);
   uint32_t x463;
   fiat_p384_uint1 x464;
-  fiat_p384_addcarryx_u32(&x463, &x464, x462, x437, x389);
+  fiat_p384_addcarryx_u32(&x463, &x464, x462, x389, x437);
   uint32_t x465;
   fiat_p384_uint1 x466;
-  fiat_p384_addcarryx_u32(&x465, &x466, x464, x439, x391);
+  fiat_p384_addcarryx_u32(&x465, &x466, x464, x391, x439);
   uint32_t x467;
   fiat_p384_uint1 x468;
-  fiat_p384_addcarryx_u32(&x467, &x468, x466, x441, x393);
+  fiat_p384_addcarryx_u32(&x467, &x468, x466, x393, x441);
   uint32_t x469;
   fiat_p384_uint1 x470;
-  fiat_p384_addcarryx_u32(&x469, &x470, x468, x443, x395);
+  fiat_p384_addcarryx_u32(&x469, &x470, x468, x395, x443);
   uint32_t x471;
   fiat_p384_uint1 x472;
-  fiat_p384_addcarryx_u32(&x471, &x472, x470, x445, x397);
+  fiat_p384_addcarryx_u32(&x471, &x472, x470, x397, x445);
   uint32_t x473;
   fiat_p384_uint1 x474;
-  fiat_p384_addcarryx_u32(&x473, &x474, x472, x447, x399);
+  fiat_p384_addcarryx_u32(&x473, &x474, x472, x399, x447);
   uint32_t x475;
   fiat_p384_uint1 x476;
-  fiat_p384_addcarryx_u32(&x475, &x476, x474, x449, x401);
+  fiat_p384_addcarryx_u32(&x475, &x476, x474, x401, x449);
   uint32_t x477;
   fiat_p384_uint1 x478;
-  fiat_p384_addcarryx_u32(&x477, &x478, x476, x451, x403);
+  fiat_p384_addcarryx_u32(&x477, &x478, x476, x403, x451);
   uint32_t x479;
   fiat_p384_uint1 x480;
-  fiat_p384_addcarryx_u32(&x479, &x480, x478, x453, x405);
+  fiat_p384_addcarryx_u32(&x479, &x480, x478, x405, x453);
   uint32_t x481;
   uint32_t x482;
   fiat_p384_mulx_u32(&x481, &x482, x455, UINT32_C(0xffffffff));
@@ -3446,73 +3446,73 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x499, &x500, x455, UINT32_C(0xffffffff));
   uint32_t x501;
   fiat_p384_uint1 x502;
-  fiat_p384_addcarryx_u32(&x501, &x502, 0x0, x495, x498);
+  fiat_p384_addcarryx_u32(&x501, &x502, 0x0, x498, x495);
   uint32_t x503;
   fiat_p384_uint1 x504;
-  fiat_p384_addcarryx_u32(&x503, &x504, x502, x493, x496);
+  fiat_p384_addcarryx_u32(&x503, &x504, x502, x496, x493);
   uint32_t x505;
   fiat_p384_uint1 x506;
-  fiat_p384_addcarryx_u32(&x505, &x506, x504, x491, x494);
+  fiat_p384_addcarryx_u32(&x505, &x506, x504, x494, x491);
   uint32_t x507;
   fiat_p384_uint1 x508;
-  fiat_p384_addcarryx_u32(&x507, &x508, x506, x489, x492);
+  fiat_p384_addcarryx_u32(&x507, &x508, x506, x492, x489);
   uint32_t x509;
   fiat_p384_uint1 x510;
-  fiat_p384_addcarryx_u32(&x509, &x510, x508, x487, x490);
+  fiat_p384_addcarryx_u32(&x509, &x510, x508, x490, x487);
   uint32_t x511;
   fiat_p384_uint1 x512;
-  fiat_p384_addcarryx_u32(&x511, &x512, x510, x485, x488);
+  fiat_p384_addcarryx_u32(&x511, &x512, x510, x488, x485);
   uint32_t x513;
   fiat_p384_uint1 x514;
-  fiat_p384_addcarryx_u32(&x513, &x514, x512, x483, x486);
+  fiat_p384_addcarryx_u32(&x513, &x514, x512, x486, x483);
   uint32_t x515;
   fiat_p384_uint1 x516;
-  fiat_p384_addcarryx_u32(&x515, &x516, x514, x481, x484);
+  fiat_p384_addcarryx_u32(&x515, &x516, x514, x484, x481);
   uint32_t x517;
   fiat_p384_uint1 x518;
-  fiat_p384_addcarryx_u32(&x517, &x518, x516, 0x0, x482);
+  fiat_p384_addcarryx_u32(&x517, &x518, x516, x482, 0x0);
   uint32_t x519;
   fiat_p384_uint1 x520;
-  fiat_p384_addcarryx_u32(&x519, &x520, 0x0, x499, x455);
+  fiat_p384_addcarryx_u32(&x519, &x520, 0x0, x455, x499);
   uint32_t x521;
   fiat_p384_uint1 x522;
-  fiat_p384_addcarryx_u32(&x521, &x522, x520, x500, x457);
+  fiat_p384_addcarryx_u32(&x521, &x522, x520, x457, x500);
   uint32_t x523;
   fiat_p384_uint1 x524;
-  fiat_p384_addcarryx_u32(&x523, &x524, x522, 0x0, x459);
+  fiat_p384_addcarryx_u32(&x523, &x524, x522, x459, 0x0);
   uint32_t x525;
   fiat_p384_uint1 x526;
-  fiat_p384_addcarryx_u32(&x525, &x526, x524, x497, x461);
+  fiat_p384_addcarryx_u32(&x525, &x526, x524, x461, x497);
   uint32_t x527;
   fiat_p384_uint1 x528;
-  fiat_p384_addcarryx_u32(&x527, &x528, x526, x501, x463);
+  fiat_p384_addcarryx_u32(&x527, &x528, x526, x463, x501);
   uint32_t x529;
   fiat_p384_uint1 x530;
-  fiat_p384_addcarryx_u32(&x529, &x530, x528, x503, x465);
+  fiat_p384_addcarryx_u32(&x529, &x530, x528, x465, x503);
   uint32_t x531;
   fiat_p384_uint1 x532;
-  fiat_p384_addcarryx_u32(&x531, &x532, x530, x505, x467);
+  fiat_p384_addcarryx_u32(&x531, &x532, x530, x467, x505);
   uint32_t x533;
   fiat_p384_uint1 x534;
-  fiat_p384_addcarryx_u32(&x533, &x534, x532, x507, x469);
+  fiat_p384_addcarryx_u32(&x533, &x534, x532, x469, x507);
   uint32_t x535;
   fiat_p384_uint1 x536;
-  fiat_p384_addcarryx_u32(&x535, &x536, x534, x509, x471);
+  fiat_p384_addcarryx_u32(&x535, &x536, x534, x471, x509);
   uint32_t x537;
   fiat_p384_uint1 x538;
-  fiat_p384_addcarryx_u32(&x537, &x538, x536, x511, x473);
+  fiat_p384_addcarryx_u32(&x537, &x538, x536, x473, x511);
   uint32_t x539;
   fiat_p384_uint1 x540;
-  fiat_p384_addcarryx_u32(&x539, &x540, x538, x513, x475);
+  fiat_p384_addcarryx_u32(&x539, &x540, x538, x475, x513);
   uint32_t x541;
   fiat_p384_uint1 x542;
-  fiat_p384_addcarryx_u32(&x541, &x542, x540, x515, x477);
+  fiat_p384_addcarryx_u32(&x541, &x542, x540, x477, x515);
   uint32_t x543;
   fiat_p384_uint1 x544;
-  fiat_p384_addcarryx_u32(&x543, &x544, x542, x517, x479);
+  fiat_p384_addcarryx_u32(&x543, &x544, x542, x479, x517);
   uint32_t x545;
   fiat_p384_uint1 x546;
-  fiat_p384_addcarryx_u32(&x545, &x546, x544, 0x0, x480);
+  fiat_p384_addcarryx_u32(&x545, &x546, x544, x480, 0x0);
   uint32_t x547;
   uint32_t x548;
   fiat_p384_mulx_u32(&x547, &x548, x4, (arg1[11]));
@@ -3551,79 +3551,79 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x569, &x570, x4, (arg1[0]));
   uint32_t x571;
   fiat_p384_uint1 x572;
-  fiat_p384_addcarryx_u32(&x571, &x572, 0x0, x567, x570);
+  fiat_p384_addcarryx_u32(&x571, &x572, 0x0, x570, x567);
   uint32_t x573;
   fiat_p384_uint1 x574;
-  fiat_p384_addcarryx_u32(&x573, &x574, x572, x565, x568);
+  fiat_p384_addcarryx_u32(&x573, &x574, x572, x568, x565);
   uint32_t x575;
   fiat_p384_uint1 x576;
-  fiat_p384_addcarryx_u32(&x575, &x576, x574, x563, x566);
+  fiat_p384_addcarryx_u32(&x575, &x576, x574, x566, x563);
   uint32_t x577;
   fiat_p384_uint1 x578;
-  fiat_p384_addcarryx_u32(&x577, &x578, x576, x561, x564);
+  fiat_p384_addcarryx_u32(&x577, &x578, x576, x564, x561);
   uint32_t x579;
   fiat_p384_uint1 x580;
-  fiat_p384_addcarryx_u32(&x579, &x580, x578, x559, x562);
+  fiat_p384_addcarryx_u32(&x579, &x580, x578, x562, x559);
   uint32_t x581;
   fiat_p384_uint1 x582;
-  fiat_p384_addcarryx_u32(&x581, &x582, x580, x557, x560);
+  fiat_p384_addcarryx_u32(&x581, &x582, x580, x560, x557);
   uint32_t x583;
   fiat_p384_uint1 x584;
-  fiat_p384_addcarryx_u32(&x583, &x584, x582, x555, x558);
+  fiat_p384_addcarryx_u32(&x583, &x584, x582, x558, x555);
   uint32_t x585;
   fiat_p384_uint1 x586;
-  fiat_p384_addcarryx_u32(&x585, &x586, x584, x553, x556);
+  fiat_p384_addcarryx_u32(&x585, &x586, x584, x556, x553);
   uint32_t x587;
   fiat_p384_uint1 x588;
-  fiat_p384_addcarryx_u32(&x587, &x588, x586, x551, x554);
+  fiat_p384_addcarryx_u32(&x587, &x588, x586, x554, x551);
   uint32_t x589;
   fiat_p384_uint1 x590;
-  fiat_p384_addcarryx_u32(&x589, &x590, x588, x549, x552);
+  fiat_p384_addcarryx_u32(&x589, &x590, x588, x552, x549);
   uint32_t x591;
   fiat_p384_uint1 x592;
-  fiat_p384_addcarryx_u32(&x591, &x592, x590, x547, x550);
+  fiat_p384_addcarryx_u32(&x591, &x592, x590, x550, x547);
   uint32_t x593;
   fiat_p384_uint1 x594;
-  fiat_p384_addcarryx_u32(&x593, &x594, x592, 0x0, x548);
+  fiat_p384_addcarryx_u32(&x593, &x594, x592, x548, 0x0);
   uint32_t x595;
   fiat_p384_uint1 x596;
-  fiat_p384_addcarryx_u32(&x595, &x596, 0x0, x569, x521);
+  fiat_p384_addcarryx_u32(&x595, &x596, 0x0, x521, x569);
   uint32_t x597;
   fiat_p384_uint1 x598;
-  fiat_p384_addcarryx_u32(&x597, &x598, x596, x571, x523);
+  fiat_p384_addcarryx_u32(&x597, &x598, x596, x523, x571);
   uint32_t x599;
   fiat_p384_uint1 x600;
-  fiat_p384_addcarryx_u32(&x599, &x600, x598, x573, x525);
+  fiat_p384_addcarryx_u32(&x599, &x600, x598, x525, x573);
   uint32_t x601;
   fiat_p384_uint1 x602;
-  fiat_p384_addcarryx_u32(&x601, &x602, x600, x575, x527);
+  fiat_p384_addcarryx_u32(&x601, &x602, x600, x527, x575);
   uint32_t x603;
   fiat_p384_uint1 x604;
-  fiat_p384_addcarryx_u32(&x603, &x604, x602, x577, x529);
+  fiat_p384_addcarryx_u32(&x603, &x604, x602, x529, x577);
   uint32_t x605;
   fiat_p384_uint1 x606;
-  fiat_p384_addcarryx_u32(&x605, &x606, x604, x579, x531);
+  fiat_p384_addcarryx_u32(&x605, &x606, x604, x531, x579);
   uint32_t x607;
   fiat_p384_uint1 x608;
-  fiat_p384_addcarryx_u32(&x607, &x608, x606, x581, x533);
+  fiat_p384_addcarryx_u32(&x607, &x608, x606, x533, x581);
   uint32_t x609;
   fiat_p384_uint1 x610;
-  fiat_p384_addcarryx_u32(&x609, &x610, x608, x583, x535);
+  fiat_p384_addcarryx_u32(&x609, &x610, x608, x535, x583);
   uint32_t x611;
   fiat_p384_uint1 x612;
-  fiat_p384_addcarryx_u32(&x611, &x612, x610, x585, x537);
+  fiat_p384_addcarryx_u32(&x611, &x612, x610, x537, x585);
   uint32_t x613;
   fiat_p384_uint1 x614;
-  fiat_p384_addcarryx_u32(&x613, &x614, x612, x587, x539);
+  fiat_p384_addcarryx_u32(&x613, &x614, x612, x539, x587);
   uint32_t x615;
   fiat_p384_uint1 x616;
-  fiat_p384_addcarryx_u32(&x615, &x616, x614, x589, x541);
+  fiat_p384_addcarryx_u32(&x615, &x616, x614, x541, x589);
   uint32_t x617;
   fiat_p384_uint1 x618;
-  fiat_p384_addcarryx_u32(&x617, &x618, x616, x591, x543);
+  fiat_p384_addcarryx_u32(&x617, &x618, x616, x543, x591);
   uint32_t x619;
   fiat_p384_uint1 x620;
-  fiat_p384_addcarryx_u32(&x619, &x620, x618, x593, x545);
+  fiat_p384_addcarryx_u32(&x619, &x620, x618, x545, x593);
   uint32_t x621;
   uint32_t x622;
   fiat_p384_mulx_u32(&x621, &x622, x595, UINT32_C(0xffffffff));
@@ -3656,73 +3656,73 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x639, &x640, x595, UINT32_C(0xffffffff));
   uint32_t x641;
   fiat_p384_uint1 x642;
-  fiat_p384_addcarryx_u32(&x641, &x642, 0x0, x635, x638);
+  fiat_p384_addcarryx_u32(&x641, &x642, 0x0, x638, x635);
   uint32_t x643;
   fiat_p384_uint1 x644;
-  fiat_p384_addcarryx_u32(&x643, &x644, x642, x633, x636);
+  fiat_p384_addcarryx_u32(&x643, &x644, x642, x636, x633);
   uint32_t x645;
   fiat_p384_uint1 x646;
-  fiat_p384_addcarryx_u32(&x645, &x646, x644, x631, x634);
+  fiat_p384_addcarryx_u32(&x645, &x646, x644, x634, x631);
   uint32_t x647;
   fiat_p384_uint1 x648;
-  fiat_p384_addcarryx_u32(&x647, &x648, x646, x629, x632);
+  fiat_p384_addcarryx_u32(&x647, &x648, x646, x632, x629);
   uint32_t x649;
   fiat_p384_uint1 x650;
-  fiat_p384_addcarryx_u32(&x649, &x650, x648, x627, x630);
+  fiat_p384_addcarryx_u32(&x649, &x650, x648, x630, x627);
   uint32_t x651;
   fiat_p384_uint1 x652;
-  fiat_p384_addcarryx_u32(&x651, &x652, x650, x625, x628);
+  fiat_p384_addcarryx_u32(&x651, &x652, x650, x628, x625);
   uint32_t x653;
   fiat_p384_uint1 x654;
-  fiat_p384_addcarryx_u32(&x653, &x654, x652, x623, x626);
+  fiat_p384_addcarryx_u32(&x653, &x654, x652, x626, x623);
   uint32_t x655;
   fiat_p384_uint1 x656;
-  fiat_p384_addcarryx_u32(&x655, &x656, x654, x621, x624);
+  fiat_p384_addcarryx_u32(&x655, &x656, x654, x624, x621);
   uint32_t x657;
   fiat_p384_uint1 x658;
-  fiat_p384_addcarryx_u32(&x657, &x658, x656, 0x0, x622);
+  fiat_p384_addcarryx_u32(&x657, &x658, x656, x622, 0x0);
   uint32_t x659;
   fiat_p384_uint1 x660;
-  fiat_p384_addcarryx_u32(&x659, &x660, 0x0, x639, x595);
+  fiat_p384_addcarryx_u32(&x659, &x660, 0x0, x595, x639);
   uint32_t x661;
   fiat_p384_uint1 x662;
-  fiat_p384_addcarryx_u32(&x661, &x662, x660, x640, x597);
+  fiat_p384_addcarryx_u32(&x661, &x662, x660, x597, x640);
   uint32_t x663;
   fiat_p384_uint1 x664;
-  fiat_p384_addcarryx_u32(&x663, &x664, x662, 0x0, x599);
+  fiat_p384_addcarryx_u32(&x663, &x664, x662, x599, 0x0);
   uint32_t x665;
   fiat_p384_uint1 x666;
-  fiat_p384_addcarryx_u32(&x665, &x666, x664, x637, x601);
+  fiat_p384_addcarryx_u32(&x665, &x666, x664, x601, x637);
   uint32_t x667;
   fiat_p384_uint1 x668;
-  fiat_p384_addcarryx_u32(&x667, &x668, x666, x641, x603);
+  fiat_p384_addcarryx_u32(&x667, &x668, x666, x603, x641);
   uint32_t x669;
   fiat_p384_uint1 x670;
-  fiat_p384_addcarryx_u32(&x669, &x670, x668, x643, x605);
+  fiat_p384_addcarryx_u32(&x669, &x670, x668, x605, x643);
   uint32_t x671;
   fiat_p384_uint1 x672;
-  fiat_p384_addcarryx_u32(&x671, &x672, x670, x645, x607);
+  fiat_p384_addcarryx_u32(&x671, &x672, x670, x607, x645);
   uint32_t x673;
   fiat_p384_uint1 x674;
-  fiat_p384_addcarryx_u32(&x673, &x674, x672, x647, x609);
+  fiat_p384_addcarryx_u32(&x673, &x674, x672, x609, x647);
   uint32_t x675;
   fiat_p384_uint1 x676;
-  fiat_p384_addcarryx_u32(&x675, &x676, x674, x649, x611);
+  fiat_p384_addcarryx_u32(&x675, &x676, x674, x611, x649);
   uint32_t x677;
   fiat_p384_uint1 x678;
-  fiat_p384_addcarryx_u32(&x677, &x678, x676, x651, x613);
+  fiat_p384_addcarryx_u32(&x677, &x678, x676, x613, x651);
   uint32_t x679;
   fiat_p384_uint1 x680;
-  fiat_p384_addcarryx_u32(&x679, &x680, x678, x653, x615);
+  fiat_p384_addcarryx_u32(&x679, &x680, x678, x615, x653);
   uint32_t x681;
   fiat_p384_uint1 x682;
-  fiat_p384_addcarryx_u32(&x681, &x682, x680, x655, x617);
+  fiat_p384_addcarryx_u32(&x681, &x682, x680, x617, x655);
   uint32_t x683;
   fiat_p384_uint1 x684;
-  fiat_p384_addcarryx_u32(&x683, &x684, x682, x657, x619);
+  fiat_p384_addcarryx_u32(&x683, &x684, x682, x619, x657);
   uint32_t x685;
   fiat_p384_uint1 x686;
-  fiat_p384_addcarryx_u32(&x685, &x686, x684, 0x0, x620);
+  fiat_p384_addcarryx_u32(&x685, &x686, x684, x620, 0x0);
   uint32_t x687;
   uint32_t x688;
   fiat_p384_mulx_u32(&x687, &x688, x5, (arg1[11]));
@@ -3761,79 +3761,79 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x709, &x710, x5, (arg1[0]));
   uint32_t x711;
   fiat_p384_uint1 x712;
-  fiat_p384_addcarryx_u32(&x711, &x712, 0x0, x707, x710);
+  fiat_p384_addcarryx_u32(&x711, &x712, 0x0, x710, x707);
   uint32_t x713;
   fiat_p384_uint1 x714;
-  fiat_p384_addcarryx_u32(&x713, &x714, x712, x705, x708);
+  fiat_p384_addcarryx_u32(&x713, &x714, x712, x708, x705);
   uint32_t x715;
   fiat_p384_uint1 x716;
-  fiat_p384_addcarryx_u32(&x715, &x716, x714, x703, x706);
+  fiat_p384_addcarryx_u32(&x715, &x716, x714, x706, x703);
   uint32_t x717;
   fiat_p384_uint1 x718;
-  fiat_p384_addcarryx_u32(&x717, &x718, x716, x701, x704);
+  fiat_p384_addcarryx_u32(&x717, &x718, x716, x704, x701);
   uint32_t x719;
   fiat_p384_uint1 x720;
-  fiat_p384_addcarryx_u32(&x719, &x720, x718, x699, x702);
+  fiat_p384_addcarryx_u32(&x719, &x720, x718, x702, x699);
   uint32_t x721;
   fiat_p384_uint1 x722;
-  fiat_p384_addcarryx_u32(&x721, &x722, x720, x697, x700);
+  fiat_p384_addcarryx_u32(&x721, &x722, x720, x700, x697);
   uint32_t x723;
   fiat_p384_uint1 x724;
-  fiat_p384_addcarryx_u32(&x723, &x724, x722, x695, x698);
+  fiat_p384_addcarryx_u32(&x723, &x724, x722, x698, x695);
   uint32_t x725;
   fiat_p384_uint1 x726;
-  fiat_p384_addcarryx_u32(&x725, &x726, x724, x693, x696);
+  fiat_p384_addcarryx_u32(&x725, &x726, x724, x696, x693);
   uint32_t x727;
   fiat_p384_uint1 x728;
-  fiat_p384_addcarryx_u32(&x727, &x728, x726, x691, x694);
+  fiat_p384_addcarryx_u32(&x727, &x728, x726, x694, x691);
   uint32_t x729;
   fiat_p384_uint1 x730;
-  fiat_p384_addcarryx_u32(&x729, &x730, x728, x689, x692);
+  fiat_p384_addcarryx_u32(&x729, &x730, x728, x692, x689);
   uint32_t x731;
   fiat_p384_uint1 x732;
-  fiat_p384_addcarryx_u32(&x731, &x732, x730, x687, x690);
+  fiat_p384_addcarryx_u32(&x731, &x732, x730, x690, x687);
   uint32_t x733;
   fiat_p384_uint1 x734;
-  fiat_p384_addcarryx_u32(&x733, &x734, x732, 0x0, x688);
+  fiat_p384_addcarryx_u32(&x733, &x734, x732, x688, 0x0);
   uint32_t x735;
   fiat_p384_uint1 x736;
-  fiat_p384_addcarryx_u32(&x735, &x736, 0x0, x709, x661);
+  fiat_p384_addcarryx_u32(&x735, &x736, 0x0, x661, x709);
   uint32_t x737;
   fiat_p384_uint1 x738;
-  fiat_p384_addcarryx_u32(&x737, &x738, x736, x711, x663);
+  fiat_p384_addcarryx_u32(&x737, &x738, x736, x663, x711);
   uint32_t x739;
   fiat_p384_uint1 x740;
-  fiat_p384_addcarryx_u32(&x739, &x740, x738, x713, x665);
+  fiat_p384_addcarryx_u32(&x739, &x740, x738, x665, x713);
   uint32_t x741;
   fiat_p384_uint1 x742;
-  fiat_p384_addcarryx_u32(&x741, &x742, x740, x715, x667);
+  fiat_p384_addcarryx_u32(&x741, &x742, x740, x667, x715);
   uint32_t x743;
   fiat_p384_uint1 x744;
-  fiat_p384_addcarryx_u32(&x743, &x744, x742, x717, x669);
+  fiat_p384_addcarryx_u32(&x743, &x744, x742, x669, x717);
   uint32_t x745;
   fiat_p384_uint1 x746;
-  fiat_p384_addcarryx_u32(&x745, &x746, x744, x719, x671);
+  fiat_p384_addcarryx_u32(&x745, &x746, x744, x671, x719);
   uint32_t x747;
   fiat_p384_uint1 x748;
-  fiat_p384_addcarryx_u32(&x747, &x748, x746, x721, x673);
+  fiat_p384_addcarryx_u32(&x747, &x748, x746, x673, x721);
   uint32_t x749;
   fiat_p384_uint1 x750;
-  fiat_p384_addcarryx_u32(&x749, &x750, x748, x723, x675);
+  fiat_p384_addcarryx_u32(&x749, &x750, x748, x675, x723);
   uint32_t x751;
   fiat_p384_uint1 x752;
-  fiat_p384_addcarryx_u32(&x751, &x752, x750, x725, x677);
+  fiat_p384_addcarryx_u32(&x751, &x752, x750, x677, x725);
   uint32_t x753;
   fiat_p384_uint1 x754;
-  fiat_p384_addcarryx_u32(&x753, &x754, x752, x727, x679);
+  fiat_p384_addcarryx_u32(&x753, &x754, x752, x679, x727);
   uint32_t x755;
   fiat_p384_uint1 x756;
-  fiat_p384_addcarryx_u32(&x755, &x756, x754, x729, x681);
+  fiat_p384_addcarryx_u32(&x755, &x756, x754, x681, x729);
   uint32_t x757;
   fiat_p384_uint1 x758;
-  fiat_p384_addcarryx_u32(&x757, &x758, x756, x731, x683);
+  fiat_p384_addcarryx_u32(&x757, &x758, x756, x683, x731);
   uint32_t x759;
   fiat_p384_uint1 x760;
-  fiat_p384_addcarryx_u32(&x759, &x760, x758, x733, x685);
+  fiat_p384_addcarryx_u32(&x759, &x760, x758, x685, x733);
   uint32_t x761;
   uint32_t x762;
   fiat_p384_mulx_u32(&x761, &x762, x735, UINT32_C(0xffffffff));
@@ -3866,73 +3866,73 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x779, &x780, x735, UINT32_C(0xffffffff));
   uint32_t x781;
   fiat_p384_uint1 x782;
-  fiat_p384_addcarryx_u32(&x781, &x782, 0x0, x775, x778);
+  fiat_p384_addcarryx_u32(&x781, &x782, 0x0, x778, x775);
   uint32_t x783;
   fiat_p384_uint1 x784;
-  fiat_p384_addcarryx_u32(&x783, &x784, x782, x773, x776);
+  fiat_p384_addcarryx_u32(&x783, &x784, x782, x776, x773);
   uint32_t x785;
   fiat_p384_uint1 x786;
-  fiat_p384_addcarryx_u32(&x785, &x786, x784, x771, x774);
+  fiat_p384_addcarryx_u32(&x785, &x786, x784, x774, x771);
   uint32_t x787;
   fiat_p384_uint1 x788;
-  fiat_p384_addcarryx_u32(&x787, &x788, x786, x769, x772);
+  fiat_p384_addcarryx_u32(&x787, &x788, x786, x772, x769);
   uint32_t x789;
   fiat_p384_uint1 x790;
-  fiat_p384_addcarryx_u32(&x789, &x790, x788, x767, x770);
+  fiat_p384_addcarryx_u32(&x789, &x790, x788, x770, x767);
   uint32_t x791;
   fiat_p384_uint1 x792;
-  fiat_p384_addcarryx_u32(&x791, &x792, x790, x765, x768);
+  fiat_p384_addcarryx_u32(&x791, &x792, x790, x768, x765);
   uint32_t x793;
   fiat_p384_uint1 x794;
-  fiat_p384_addcarryx_u32(&x793, &x794, x792, x763, x766);
+  fiat_p384_addcarryx_u32(&x793, &x794, x792, x766, x763);
   uint32_t x795;
   fiat_p384_uint1 x796;
-  fiat_p384_addcarryx_u32(&x795, &x796, x794, x761, x764);
+  fiat_p384_addcarryx_u32(&x795, &x796, x794, x764, x761);
   uint32_t x797;
   fiat_p384_uint1 x798;
-  fiat_p384_addcarryx_u32(&x797, &x798, x796, 0x0, x762);
+  fiat_p384_addcarryx_u32(&x797, &x798, x796, x762, 0x0);
   uint32_t x799;
   fiat_p384_uint1 x800;
-  fiat_p384_addcarryx_u32(&x799, &x800, 0x0, x779, x735);
+  fiat_p384_addcarryx_u32(&x799, &x800, 0x0, x735, x779);
   uint32_t x801;
   fiat_p384_uint1 x802;
-  fiat_p384_addcarryx_u32(&x801, &x802, x800, x780, x737);
+  fiat_p384_addcarryx_u32(&x801, &x802, x800, x737, x780);
   uint32_t x803;
   fiat_p384_uint1 x804;
-  fiat_p384_addcarryx_u32(&x803, &x804, x802, 0x0, x739);
+  fiat_p384_addcarryx_u32(&x803, &x804, x802, x739, 0x0);
   uint32_t x805;
   fiat_p384_uint1 x806;
-  fiat_p384_addcarryx_u32(&x805, &x806, x804, x777, x741);
+  fiat_p384_addcarryx_u32(&x805, &x806, x804, x741, x777);
   uint32_t x807;
   fiat_p384_uint1 x808;
-  fiat_p384_addcarryx_u32(&x807, &x808, x806, x781, x743);
+  fiat_p384_addcarryx_u32(&x807, &x808, x806, x743, x781);
   uint32_t x809;
   fiat_p384_uint1 x810;
-  fiat_p384_addcarryx_u32(&x809, &x810, x808, x783, x745);
+  fiat_p384_addcarryx_u32(&x809, &x810, x808, x745, x783);
   uint32_t x811;
   fiat_p384_uint1 x812;
-  fiat_p384_addcarryx_u32(&x811, &x812, x810, x785, x747);
+  fiat_p384_addcarryx_u32(&x811, &x812, x810, x747, x785);
   uint32_t x813;
   fiat_p384_uint1 x814;
-  fiat_p384_addcarryx_u32(&x813, &x814, x812, x787, x749);
+  fiat_p384_addcarryx_u32(&x813, &x814, x812, x749, x787);
   uint32_t x815;
   fiat_p384_uint1 x816;
-  fiat_p384_addcarryx_u32(&x815, &x816, x814, x789, x751);
+  fiat_p384_addcarryx_u32(&x815, &x816, x814, x751, x789);
   uint32_t x817;
   fiat_p384_uint1 x818;
-  fiat_p384_addcarryx_u32(&x817, &x818, x816, x791, x753);
+  fiat_p384_addcarryx_u32(&x817, &x818, x816, x753, x791);
   uint32_t x819;
   fiat_p384_uint1 x820;
-  fiat_p384_addcarryx_u32(&x819, &x820, x818, x793, x755);
+  fiat_p384_addcarryx_u32(&x819, &x820, x818, x755, x793);
   uint32_t x821;
   fiat_p384_uint1 x822;
-  fiat_p384_addcarryx_u32(&x821, &x822, x820, x795, x757);
+  fiat_p384_addcarryx_u32(&x821, &x822, x820, x757, x795);
   uint32_t x823;
   fiat_p384_uint1 x824;
-  fiat_p384_addcarryx_u32(&x823, &x824, x822, x797, x759);
+  fiat_p384_addcarryx_u32(&x823, &x824, x822, x759, x797);
   uint32_t x825;
   fiat_p384_uint1 x826;
-  fiat_p384_addcarryx_u32(&x825, &x826, x824, 0x0, x760);
+  fiat_p384_addcarryx_u32(&x825, &x826, x824, x760, 0x0);
   uint32_t x827;
   uint32_t x828;
   fiat_p384_mulx_u32(&x827, &x828, x6, (arg1[11]));
@@ -3971,79 +3971,79 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x849, &x850, x6, (arg1[0]));
   uint32_t x851;
   fiat_p384_uint1 x852;
-  fiat_p384_addcarryx_u32(&x851, &x852, 0x0, x847, x850);
+  fiat_p384_addcarryx_u32(&x851, &x852, 0x0, x850, x847);
   uint32_t x853;
   fiat_p384_uint1 x854;
-  fiat_p384_addcarryx_u32(&x853, &x854, x852, x845, x848);
+  fiat_p384_addcarryx_u32(&x853, &x854, x852, x848, x845);
   uint32_t x855;
   fiat_p384_uint1 x856;
-  fiat_p384_addcarryx_u32(&x855, &x856, x854, x843, x846);
+  fiat_p384_addcarryx_u32(&x855, &x856, x854, x846, x843);
   uint32_t x857;
   fiat_p384_uint1 x858;
-  fiat_p384_addcarryx_u32(&x857, &x858, x856, x841, x844);
+  fiat_p384_addcarryx_u32(&x857, &x858, x856, x844, x841);
   uint32_t x859;
   fiat_p384_uint1 x860;
-  fiat_p384_addcarryx_u32(&x859, &x860, x858, x839, x842);
+  fiat_p384_addcarryx_u32(&x859, &x860, x858, x842, x839);
   uint32_t x861;
   fiat_p384_uint1 x862;
-  fiat_p384_addcarryx_u32(&x861, &x862, x860, x837, x840);
+  fiat_p384_addcarryx_u32(&x861, &x862, x860, x840, x837);
   uint32_t x863;
   fiat_p384_uint1 x864;
-  fiat_p384_addcarryx_u32(&x863, &x864, x862, x835, x838);
+  fiat_p384_addcarryx_u32(&x863, &x864, x862, x838, x835);
   uint32_t x865;
   fiat_p384_uint1 x866;
-  fiat_p384_addcarryx_u32(&x865, &x866, x864, x833, x836);
+  fiat_p384_addcarryx_u32(&x865, &x866, x864, x836, x833);
   uint32_t x867;
   fiat_p384_uint1 x868;
-  fiat_p384_addcarryx_u32(&x867, &x868, x866, x831, x834);
+  fiat_p384_addcarryx_u32(&x867, &x868, x866, x834, x831);
   uint32_t x869;
   fiat_p384_uint1 x870;
-  fiat_p384_addcarryx_u32(&x869, &x870, x868, x829, x832);
+  fiat_p384_addcarryx_u32(&x869, &x870, x868, x832, x829);
   uint32_t x871;
   fiat_p384_uint1 x872;
-  fiat_p384_addcarryx_u32(&x871, &x872, x870, x827, x830);
+  fiat_p384_addcarryx_u32(&x871, &x872, x870, x830, x827);
   uint32_t x873;
   fiat_p384_uint1 x874;
-  fiat_p384_addcarryx_u32(&x873, &x874, x872, 0x0, x828);
+  fiat_p384_addcarryx_u32(&x873, &x874, x872, x828, 0x0);
   uint32_t x875;
   fiat_p384_uint1 x876;
-  fiat_p384_addcarryx_u32(&x875, &x876, 0x0, x849, x801);
+  fiat_p384_addcarryx_u32(&x875, &x876, 0x0, x801, x849);
   uint32_t x877;
   fiat_p384_uint1 x878;
-  fiat_p384_addcarryx_u32(&x877, &x878, x876, x851, x803);
+  fiat_p384_addcarryx_u32(&x877, &x878, x876, x803, x851);
   uint32_t x879;
   fiat_p384_uint1 x880;
-  fiat_p384_addcarryx_u32(&x879, &x880, x878, x853, x805);
+  fiat_p384_addcarryx_u32(&x879, &x880, x878, x805, x853);
   uint32_t x881;
   fiat_p384_uint1 x882;
-  fiat_p384_addcarryx_u32(&x881, &x882, x880, x855, x807);
+  fiat_p384_addcarryx_u32(&x881, &x882, x880, x807, x855);
   uint32_t x883;
   fiat_p384_uint1 x884;
-  fiat_p384_addcarryx_u32(&x883, &x884, x882, x857, x809);
+  fiat_p384_addcarryx_u32(&x883, &x884, x882, x809, x857);
   uint32_t x885;
   fiat_p384_uint1 x886;
-  fiat_p384_addcarryx_u32(&x885, &x886, x884, x859, x811);
+  fiat_p384_addcarryx_u32(&x885, &x886, x884, x811, x859);
   uint32_t x887;
   fiat_p384_uint1 x888;
-  fiat_p384_addcarryx_u32(&x887, &x888, x886, x861, x813);
+  fiat_p384_addcarryx_u32(&x887, &x888, x886, x813, x861);
   uint32_t x889;
   fiat_p384_uint1 x890;
-  fiat_p384_addcarryx_u32(&x889, &x890, x888, x863, x815);
+  fiat_p384_addcarryx_u32(&x889, &x890, x888, x815, x863);
   uint32_t x891;
   fiat_p384_uint1 x892;
-  fiat_p384_addcarryx_u32(&x891, &x892, x890, x865, x817);
+  fiat_p384_addcarryx_u32(&x891, &x892, x890, x817, x865);
   uint32_t x893;
   fiat_p384_uint1 x894;
-  fiat_p384_addcarryx_u32(&x893, &x894, x892, x867, x819);
+  fiat_p384_addcarryx_u32(&x893, &x894, x892, x819, x867);
   uint32_t x895;
   fiat_p384_uint1 x896;
-  fiat_p384_addcarryx_u32(&x895, &x896, x894, x869, x821);
+  fiat_p384_addcarryx_u32(&x895, &x896, x894, x821, x869);
   uint32_t x897;
   fiat_p384_uint1 x898;
-  fiat_p384_addcarryx_u32(&x897, &x898, x896, x871, x823);
+  fiat_p384_addcarryx_u32(&x897, &x898, x896, x823, x871);
   uint32_t x899;
   fiat_p384_uint1 x900;
-  fiat_p384_addcarryx_u32(&x899, &x900, x898, x873, x825);
+  fiat_p384_addcarryx_u32(&x899, &x900, x898, x825, x873);
   uint32_t x901;
   uint32_t x902;
   fiat_p384_mulx_u32(&x901, &x902, x875, UINT32_C(0xffffffff));
@@ -4076,73 +4076,73 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x919, &x920, x875, UINT32_C(0xffffffff));
   uint32_t x921;
   fiat_p384_uint1 x922;
-  fiat_p384_addcarryx_u32(&x921, &x922, 0x0, x915, x918);
+  fiat_p384_addcarryx_u32(&x921, &x922, 0x0, x918, x915);
   uint32_t x923;
   fiat_p384_uint1 x924;
-  fiat_p384_addcarryx_u32(&x923, &x924, x922, x913, x916);
+  fiat_p384_addcarryx_u32(&x923, &x924, x922, x916, x913);
   uint32_t x925;
   fiat_p384_uint1 x926;
-  fiat_p384_addcarryx_u32(&x925, &x926, x924, x911, x914);
+  fiat_p384_addcarryx_u32(&x925, &x926, x924, x914, x911);
   uint32_t x927;
   fiat_p384_uint1 x928;
-  fiat_p384_addcarryx_u32(&x927, &x928, x926, x909, x912);
+  fiat_p384_addcarryx_u32(&x927, &x928, x926, x912, x909);
   uint32_t x929;
   fiat_p384_uint1 x930;
-  fiat_p384_addcarryx_u32(&x929, &x930, x928, x907, x910);
+  fiat_p384_addcarryx_u32(&x929, &x930, x928, x910, x907);
   uint32_t x931;
   fiat_p384_uint1 x932;
-  fiat_p384_addcarryx_u32(&x931, &x932, x930, x905, x908);
+  fiat_p384_addcarryx_u32(&x931, &x932, x930, x908, x905);
   uint32_t x933;
   fiat_p384_uint1 x934;
-  fiat_p384_addcarryx_u32(&x933, &x934, x932, x903, x906);
+  fiat_p384_addcarryx_u32(&x933, &x934, x932, x906, x903);
   uint32_t x935;
   fiat_p384_uint1 x936;
-  fiat_p384_addcarryx_u32(&x935, &x936, x934, x901, x904);
+  fiat_p384_addcarryx_u32(&x935, &x936, x934, x904, x901);
   uint32_t x937;
   fiat_p384_uint1 x938;
-  fiat_p384_addcarryx_u32(&x937, &x938, x936, 0x0, x902);
+  fiat_p384_addcarryx_u32(&x937, &x938, x936, x902, 0x0);
   uint32_t x939;
   fiat_p384_uint1 x940;
-  fiat_p384_addcarryx_u32(&x939, &x940, 0x0, x919, x875);
+  fiat_p384_addcarryx_u32(&x939, &x940, 0x0, x875, x919);
   uint32_t x941;
   fiat_p384_uint1 x942;
-  fiat_p384_addcarryx_u32(&x941, &x942, x940, x920, x877);
+  fiat_p384_addcarryx_u32(&x941, &x942, x940, x877, x920);
   uint32_t x943;
   fiat_p384_uint1 x944;
-  fiat_p384_addcarryx_u32(&x943, &x944, x942, 0x0, x879);
+  fiat_p384_addcarryx_u32(&x943, &x944, x942, x879, 0x0);
   uint32_t x945;
   fiat_p384_uint1 x946;
-  fiat_p384_addcarryx_u32(&x945, &x946, x944, x917, x881);
+  fiat_p384_addcarryx_u32(&x945, &x946, x944, x881, x917);
   uint32_t x947;
   fiat_p384_uint1 x948;
-  fiat_p384_addcarryx_u32(&x947, &x948, x946, x921, x883);
+  fiat_p384_addcarryx_u32(&x947, &x948, x946, x883, x921);
   uint32_t x949;
   fiat_p384_uint1 x950;
-  fiat_p384_addcarryx_u32(&x949, &x950, x948, x923, x885);
+  fiat_p384_addcarryx_u32(&x949, &x950, x948, x885, x923);
   uint32_t x951;
   fiat_p384_uint1 x952;
-  fiat_p384_addcarryx_u32(&x951, &x952, x950, x925, x887);
+  fiat_p384_addcarryx_u32(&x951, &x952, x950, x887, x925);
   uint32_t x953;
   fiat_p384_uint1 x954;
-  fiat_p384_addcarryx_u32(&x953, &x954, x952, x927, x889);
+  fiat_p384_addcarryx_u32(&x953, &x954, x952, x889, x927);
   uint32_t x955;
   fiat_p384_uint1 x956;
-  fiat_p384_addcarryx_u32(&x955, &x956, x954, x929, x891);
+  fiat_p384_addcarryx_u32(&x955, &x956, x954, x891, x929);
   uint32_t x957;
   fiat_p384_uint1 x958;
-  fiat_p384_addcarryx_u32(&x957, &x958, x956, x931, x893);
+  fiat_p384_addcarryx_u32(&x957, &x958, x956, x893, x931);
   uint32_t x959;
   fiat_p384_uint1 x960;
-  fiat_p384_addcarryx_u32(&x959, &x960, x958, x933, x895);
+  fiat_p384_addcarryx_u32(&x959, &x960, x958, x895, x933);
   uint32_t x961;
   fiat_p384_uint1 x962;
-  fiat_p384_addcarryx_u32(&x961, &x962, x960, x935, x897);
+  fiat_p384_addcarryx_u32(&x961, &x962, x960, x897, x935);
   uint32_t x963;
   fiat_p384_uint1 x964;
-  fiat_p384_addcarryx_u32(&x963, &x964, x962, x937, x899);
+  fiat_p384_addcarryx_u32(&x963, &x964, x962, x899, x937);
   uint32_t x965;
   fiat_p384_uint1 x966;
-  fiat_p384_addcarryx_u32(&x965, &x966, x964, 0x0, x900);
+  fiat_p384_addcarryx_u32(&x965, &x966, x964, x900, 0x0);
   uint32_t x967;
   uint32_t x968;
   fiat_p384_mulx_u32(&x967, &x968, x7, (arg1[11]));
@@ -4181,79 +4181,79 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x989, &x990, x7, (arg1[0]));
   uint32_t x991;
   fiat_p384_uint1 x992;
-  fiat_p384_addcarryx_u32(&x991, &x992, 0x0, x987, x990);
+  fiat_p384_addcarryx_u32(&x991, &x992, 0x0, x990, x987);
   uint32_t x993;
   fiat_p384_uint1 x994;
-  fiat_p384_addcarryx_u32(&x993, &x994, x992, x985, x988);
+  fiat_p384_addcarryx_u32(&x993, &x994, x992, x988, x985);
   uint32_t x995;
   fiat_p384_uint1 x996;
-  fiat_p384_addcarryx_u32(&x995, &x996, x994, x983, x986);
+  fiat_p384_addcarryx_u32(&x995, &x996, x994, x986, x983);
   uint32_t x997;
   fiat_p384_uint1 x998;
-  fiat_p384_addcarryx_u32(&x997, &x998, x996, x981, x984);
+  fiat_p384_addcarryx_u32(&x997, &x998, x996, x984, x981);
   uint32_t x999;
   fiat_p384_uint1 x1000;
-  fiat_p384_addcarryx_u32(&x999, &x1000, x998, x979, x982);
+  fiat_p384_addcarryx_u32(&x999, &x1000, x998, x982, x979);
   uint32_t x1001;
   fiat_p384_uint1 x1002;
-  fiat_p384_addcarryx_u32(&x1001, &x1002, x1000, x977, x980);
+  fiat_p384_addcarryx_u32(&x1001, &x1002, x1000, x980, x977);
   uint32_t x1003;
   fiat_p384_uint1 x1004;
-  fiat_p384_addcarryx_u32(&x1003, &x1004, x1002, x975, x978);
+  fiat_p384_addcarryx_u32(&x1003, &x1004, x1002, x978, x975);
   uint32_t x1005;
   fiat_p384_uint1 x1006;
-  fiat_p384_addcarryx_u32(&x1005, &x1006, x1004, x973, x976);
+  fiat_p384_addcarryx_u32(&x1005, &x1006, x1004, x976, x973);
   uint32_t x1007;
   fiat_p384_uint1 x1008;
-  fiat_p384_addcarryx_u32(&x1007, &x1008, x1006, x971, x974);
+  fiat_p384_addcarryx_u32(&x1007, &x1008, x1006, x974, x971);
   uint32_t x1009;
   fiat_p384_uint1 x1010;
-  fiat_p384_addcarryx_u32(&x1009, &x1010, x1008, x969, x972);
+  fiat_p384_addcarryx_u32(&x1009, &x1010, x1008, x972, x969);
   uint32_t x1011;
   fiat_p384_uint1 x1012;
-  fiat_p384_addcarryx_u32(&x1011, &x1012, x1010, x967, x970);
+  fiat_p384_addcarryx_u32(&x1011, &x1012, x1010, x970, x967);
   uint32_t x1013;
   fiat_p384_uint1 x1014;
-  fiat_p384_addcarryx_u32(&x1013, &x1014, x1012, 0x0, x968);
+  fiat_p384_addcarryx_u32(&x1013, &x1014, x1012, x968, 0x0);
   uint32_t x1015;
   fiat_p384_uint1 x1016;
-  fiat_p384_addcarryx_u32(&x1015, &x1016, 0x0, x989, x941);
+  fiat_p384_addcarryx_u32(&x1015, &x1016, 0x0, x941, x989);
   uint32_t x1017;
   fiat_p384_uint1 x1018;
-  fiat_p384_addcarryx_u32(&x1017, &x1018, x1016, x991, x943);
+  fiat_p384_addcarryx_u32(&x1017, &x1018, x1016, x943, x991);
   uint32_t x1019;
   fiat_p384_uint1 x1020;
-  fiat_p384_addcarryx_u32(&x1019, &x1020, x1018, x993, x945);
+  fiat_p384_addcarryx_u32(&x1019, &x1020, x1018, x945, x993);
   uint32_t x1021;
   fiat_p384_uint1 x1022;
-  fiat_p384_addcarryx_u32(&x1021, &x1022, x1020, x995, x947);
+  fiat_p384_addcarryx_u32(&x1021, &x1022, x1020, x947, x995);
   uint32_t x1023;
   fiat_p384_uint1 x1024;
-  fiat_p384_addcarryx_u32(&x1023, &x1024, x1022, x997, x949);
+  fiat_p384_addcarryx_u32(&x1023, &x1024, x1022, x949, x997);
   uint32_t x1025;
   fiat_p384_uint1 x1026;
-  fiat_p384_addcarryx_u32(&x1025, &x1026, x1024, x999, x951);
+  fiat_p384_addcarryx_u32(&x1025, &x1026, x1024, x951, x999);
   uint32_t x1027;
   fiat_p384_uint1 x1028;
-  fiat_p384_addcarryx_u32(&x1027, &x1028, x1026, x1001, x953);
+  fiat_p384_addcarryx_u32(&x1027, &x1028, x1026, x953, x1001);
   uint32_t x1029;
   fiat_p384_uint1 x1030;
-  fiat_p384_addcarryx_u32(&x1029, &x1030, x1028, x1003, x955);
+  fiat_p384_addcarryx_u32(&x1029, &x1030, x1028, x955, x1003);
   uint32_t x1031;
   fiat_p384_uint1 x1032;
-  fiat_p384_addcarryx_u32(&x1031, &x1032, x1030, x1005, x957);
+  fiat_p384_addcarryx_u32(&x1031, &x1032, x1030, x957, x1005);
   uint32_t x1033;
   fiat_p384_uint1 x1034;
-  fiat_p384_addcarryx_u32(&x1033, &x1034, x1032, x1007, x959);
+  fiat_p384_addcarryx_u32(&x1033, &x1034, x1032, x959, x1007);
   uint32_t x1035;
   fiat_p384_uint1 x1036;
-  fiat_p384_addcarryx_u32(&x1035, &x1036, x1034, x1009, x961);
+  fiat_p384_addcarryx_u32(&x1035, &x1036, x1034, x961, x1009);
   uint32_t x1037;
   fiat_p384_uint1 x1038;
-  fiat_p384_addcarryx_u32(&x1037, &x1038, x1036, x1011, x963);
+  fiat_p384_addcarryx_u32(&x1037, &x1038, x1036, x963, x1011);
   uint32_t x1039;
   fiat_p384_uint1 x1040;
-  fiat_p384_addcarryx_u32(&x1039, &x1040, x1038, x1013, x965);
+  fiat_p384_addcarryx_u32(&x1039, &x1040, x1038, x965, x1013);
   uint32_t x1041;
   uint32_t x1042;
   fiat_p384_mulx_u32(&x1041, &x1042, x1015, UINT32_C(0xffffffff));
@@ -4286,73 +4286,73 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x1059, &x1060, x1015, UINT32_C(0xffffffff));
   uint32_t x1061;
   fiat_p384_uint1 x1062;
-  fiat_p384_addcarryx_u32(&x1061, &x1062, 0x0, x1055, x1058);
+  fiat_p384_addcarryx_u32(&x1061, &x1062, 0x0, x1058, x1055);
   uint32_t x1063;
   fiat_p384_uint1 x1064;
-  fiat_p384_addcarryx_u32(&x1063, &x1064, x1062, x1053, x1056);
+  fiat_p384_addcarryx_u32(&x1063, &x1064, x1062, x1056, x1053);
   uint32_t x1065;
   fiat_p384_uint1 x1066;
-  fiat_p384_addcarryx_u32(&x1065, &x1066, x1064, x1051, x1054);
+  fiat_p384_addcarryx_u32(&x1065, &x1066, x1064, x1054, x1051);
   uint32_t x1067;
   fiat_p384_uint1 x1068;
-  fiat_p384_addcarryx_u32(&x1067, &x1068, x1066, x1049, x1052);
+  fiat_p384_addcarryx_u32(&x1067, &x1068, x1066, x1052, x1049);
   uint32_t x1069;
   fiat_p384_uint1 x1070;
-  fiat_p384_addcarryx_u32(&x1069, &x1070, x1068, x1047, x1050);
+  fiat_p384_addcarryx_u32(&x1069, &x1070, x1068, x1050, x1047);
   uint32_t x1071;
   fiat_p384_uint1 x1072;
-  fiat_p384_addcarryx_u32(&x1071, &x1072, x1070, x1045, x1048);
+  fiat_p384_addcarryx_u32(&x1071, &x1072, x1070, x1048, x1045);
   uint32_t x1073;
   fiat_p384_uint1 x1074;
-  fiat_p384_addcarryx_u32(&x1073, &x1074, x1072, x1043, x1046);
+  fiat_p384_addcarryx_u32(&x1073, &x1074, x1072, x1046, x1043);
   uint32_t x1075;
   fiat_p384_uint1 x1076;
-  fiat_p384_addcarryx_u32(&x1075, &x1076, x1074, x1041, x1044);
+  fiat_p384_addcarryx_u32(&x1075, &x1076, x1074, x1044, x1041);
   uint32_t x1077;
   fiat_p384_uint1 x1078;
-  fiat_p384_addcarryx_u32(&x1077, &x1078, x1076, 0x0, x1042);
+  fiat_p384_addcarryx_u32(&x1077, &x1078, x1076, x1042, 0x0);
   uint32_t x1079;
   fiat_p384_uint1 x1080;
-  fiat_p384_addcarryx_u32(&x1079, &x1080, 0x0, x1059, x1015);
+  fiat_p384_addcarryx_u32(&x1079, &x1080, 0x0, x1015, x1059);
   uint32_t x1081;
   fiat_p384_uint1 x1082;
-  fiat_p384_addcarryx_u32(&x1081, &x1082, x1080, x1060, x1017);
+  fiat_p384_addcarryx_u32(&x1081, &x1082, x1080, x1017, x1060);
   uint32_t x1083;
   fiat_p384_uint1 x1084;
-  fiat_p384_addcarryx_u32(&x1083, &x1084, x1082, 0x0, x1019);
+  fiat_p384_addcarryx_u32(&x1083, &x1084, x1082, x1019, 0x0);
   uint32_t x1085;
   fiat_p384_uint1 x1086;
-  fiat_p384_addcarryx_u32(&x1085, &x1086, x1084, x1057, x1021);
+  fiat_p384_addcarryx_u32(&x1085, &x1086, x1084, x1021, x1057);
   uint32_t x1087;
   fiat_p384_uint1 x1088;
-  fiat_p384_addcarryx_u32(&x1087, &x1088, x1086, x1061, x1023);
+  fiat_p384_addcarryx_u32(&x1087, &x1088, x1086, x1023, x1061);
   uint32_t x1089;
   fiat_p384_uint1 x1090;
-  fiat_p384_addcarryx_u32(&x1089, &x1090, x1088, x1063, x1025);
+  fiat_p384_addcarryx_u32(&x1089, &x1090, x1088, x1025, x1063);
   uint32_t x1091;
   fiat_p384_uint1 x1092;
-  fiat_p384_addcarryx_u32(&x1091, &x1092, x1090, x1065, x1027);
+  fiat_p384_addcarryx_u32(&x1091, &x1092, x1090, x1027, x1065);
   uint32_t x1093;
   fiat_p384_uint1 x1094;
-  fiat_p384_addcarryx_u32(&x1093, &x1094, x1092, x1067, x1029);
+  fiat_p384_addcarryx_u32(&x1093, &x1094, x1092, x1029, x1067);
   uint32_t x1095;
   fiat_p384_uint1 x1096;
-  fiat_p384_addcarryx_u32(&x1095, &x1096, x1094, x1069, x1031);
+  fiat_p384_addcarryx_u32(&x1095, &x1096, x1094, x1031, x1069);
   uint32_t x1097;
   fiat_p384_uint1 x1098;
-  fiat_p384_addcarryx_u32(&x1097, &x1098, x1096, x1071, x1033);
+  fiat_p384_addcarryx_u32(&x1097, &x1098, x1096, x1033, x1071);
   uint32_t x1099;
   fiat_p384_uint1 x1100;
-  fiat_p384_addcarryx_u32(&x1099, &x1100, x1098, x1073, x1035);
+  fiat_p384_addcarryx_u32(&x1099, &x1100, x1098, x1035, x1073);
   uint32_t x1101;
   fiat_p384_uint1 x1102;
-  fiat_p384_addcarryx_u32(&x1101, &x1102, x1100, x1075, x1037);
+  fiat_p384_addcarryx_u32(&x1101, &x1102, x1100, x1037, x1075);
   uint32_t x1103;
   fiat_p384_uint1 x1104;
-  fiat_p384_addcarryx_u32(&x1103, &x1104, x1102, x1077, x1039);
+  fiat_p384_addcarryx_u32(&x1103, &x1104, x1102, x1039, x1077);
   uint32_t x1105;
   fiat_p384_uint1 x1106;
-  fiat_p384_addcarryx_u32(&x1105, &x1106, x1104, 0x0, x1040);
+  fiat_p384_addcarryx_u32(&x1105, &x1106, x1104, x1040, 0x0);
   uint32_t x1107;
   uint32_t x1108;
   fiat_p384_mulx_u32(&x1107, &x1108, x8, (arg1[11]));
@@ -4391,79 +4391,79 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x1129, &x1130, x8, (arg1[0]));
   uint32_t x1131;
   fiat_p384_uint1 x1132;
-  fiat_p384_addcarryx_u32(&x1131, &x1132, 0x0, x1127, x1130);
+  fiat_p384_addcarryx_u32(&x1131, &x1132, 0x0, x1130, x1127);
   uint32_t x1133;
   fiat_p384_uint1 x1134;
-  fiat_p384_addcarryx_u32(&x1133, &x1134, x1132, x1125, x1128);
+  fiat_p384_addcarryx_u32(&x1133, &x1134, x1132, x1128, x1125);
   uint32_t x1135;
   fiat_p384_uint1 x1136;
-  fiat_p384_addcarryx_u32(&x1135, &x1136, x1134, x1123, x1126);
+  fiat_p384_addcarryx_u32(&x1135, &x1136, x1134, x1126, x1123);
   uint32_t x1137;
   fiat_p384_uint1 x1138;
-  fiat_p384_addcarryx_u32(&x1137, &x1138, x1136, x1121, x1124);
+  fiat_p384_addcarryx_u32(&x1137, &x1138, x1136, x1124, x1121);
   uint32_t x1139;
   fiat_p384_uint1 x1140;
-  fiat_p384_addcarryx_u32(&x1139, &x1140, x1138, x1119, x1122);
+  fiat_p384_addcarryx_u32(&x1139, &x1140, x1138, x1122, x1119);
   uint32_t x1141;
   fiat_p384_uint1 x1142;
-  fiat_p384_addcarryx_u32(&x1141, &x1142, x1140, x1117, x1120);
+  fiat_p384_addcarryx_u32(&x1141, &x1142, x1140, x1120, x1117);
   uint32_t x1143;
   fiat_p384_uint1 x1144;
-  fiat_p384_addcarryx_u32(&x1143, &x1144, x1142, x1115, x1118);
+  fiat_p384_addcarryx_u32(&x1143, &x1144, x1142, x1118, x1115);
   uint32_t x1145;
   fiat_p384_uint1 x1146;
-  fiat_p384_addcarryx_u32(&x1145, &x1146, x1144, x1113, x1116);
+  fiat_p384_addcarryx_u32(&x1145, &x1146, x1144, x1116, x1113);
   uint32_t x1147;
   fiat_p384_uint1 x1148;
-  fiat_p384_addcarryx_u32(&x1147, &x1148, x1146, x1111, x1114);
+  fiat_p384_addcarryx_u32(&x1147, &x1148, x1146, x1114, x1111);
   uint32_t x1149;
   fiat_p384_uint1 x1150;
-  fiat_p384_addcarryx_u32(&x1149, &x1150, x1148, x1109, x1112);
+  fiat_p384_addcarryx_u32(&x1149, &x1150, x1148, x1112, x1109);
   uint32_t x1151;
   fiat_p384_uint1 x1152;
-  fiat_p384_addcarryx_u32(&x1151, &x1152, x1150, x1107, x1110);
+  fiat_p384_addcarryx_u32(&x1151, &x1152, x1150, x1110, x1107);
   uint32_t x1153;
   fiat_p384_uint1 x1154;
-  fiat_p384_addcarryx_u32(&x1153, &x1154, x1152, 0x0, x1108);
+  fiat_p384_addcarryx_u32(&x1153, &x1154, x1152, x1108, 0x0);
   uint32_t x1155;
   fiat_p384_uint1 x1156;
-  fiat_p384_addcarryx_u32(&x1155, &x1156, 0x0, x1129, x1081);
+  fiat_p384_addcarryx_u32(&x1155, &x1156, 0x0, x1081, x1129);
   uint32_t x1157;
   fiat_p384_uint1 x1158;
-  fiat_p384_addcarryx_u32(&x1157, &x1158, x1156, x1131, x1083);
+  fiat_p384_addcarryx_u32(&x1157, &x1158, x1156, x1083, x1131);
   uint32_t x1159;
   fiat_p384_uint1 x1160;
-  fiat_p384_addcarryx_u32(&x1159, &x1160, x1158, x1133, x1085);
+  fiat_p384_addcarryx_u32(&x1159, &x1160, x1158, x1085, x1133);
   uint32_t x1161;
   fiat_p384_uint1 x1162;
-  fiat_p384_addcarryx_u32(&x1161, &x1162, x1160, x1135, x1087);
+  fiat_p384_addcarryx_u32(&x1161, &x1162, x1160, x1087, x1135);
   uint32_t x1163;
   fiat_p384_uint1 x1164;
-  fiat_p384_addcarryx_u32(&x1163, &x1164, x1162, x1137, x1089);
+  fiat_p384_addcarryx_u32(&x1163, &x1164, x1162, x1089, x1137);
   uint32_t x1165;
   fiat_p384_uint1 x1166;
-  fiat_p384_addcarryx_u32(&x1165, &x1166, x1164, x1139, x1091);
+  fiat_p384_addcarryx_u32(&x1165, &x1166, x1164, x1091, x1139);
   uint32_t x1167;
   fiat_p384_uint1 x1168;
-  fiat_p384_addcarryx_u32(&x1167, &x1168, x1166, x1141, x1093);
+  fiat_p384_addcarryx_u32(&x1167, &x1168, x1166, x1093, x1141);
   uint32_t x1169;
   fiat_p384_uint1 x1170;
-  fiat_p384_addcarryx_u32(&x1169, &x1170, x1168, x1143, x1095);
+  fiat_p384_addcarryx_u32(&x1169, &x1170, x1168, x1095, x1143);
   uint32_t x1171;
   fiat_p384_uint1 x1172;
-  fiat_p384_addcarryx_u32(&x1171, &x1172, x1170, x1145, x1097);
+  fiat_p384_addcarryx_u32(&x1171, &x1172, x1170, x1097, x1145);
   uint32_t x1173;
   fiat_p384_uint1 x1174;
-  fiat_p384_addcarryx_u32(&x1173, &x1174, x1172, x1147, x1099);
+  fiat_p384_addcarryx_u32(&x1173, &x1174, x1172, x1099, x1147);
   uint32_t x1175;
   fiat_p384_uint1 x1176;
-  fiat_p384_addcarryx_u32(&x1175, &x1176, x1174, x1149, x1101);
+  fiat_p384_addcarryx_u32(&x1175, &x1176, x1174, x1101, x1149);
   uint32_t x1177;
   fiat_p384_uint1 x1178;
-  fiat_p384_addcarryx_u32(&x1177, &x1178, x1176, x1151, x1103);
+  fiat_p384_addcarryx_u32(&x1177, &x1178, x1176, x1103, x1151);
   uint32_t x1179;
   fiat_p384_uint1 x1180;
-  fiat_p384_addcarryx_u32(&x1179, &x1180, x1178, x1153, x1105);
+  fiat_p384_addcarryx_u32(&x1179, &x1180, x1178, x1105, x1153);
   uint32_t x1181;
   uint32_t x1182;
   fiat_p384_mulx_u32(&x1181, &x1182, x1155, UINT32_C(0xffffffff));
@@ -4496,73 +4496,73 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x1199, &x1200, x1155, UINT32_C(0xffffffff));
   uint32_t x1201;
   fiat_p384_uint1 x1202;
-  fiat_p384_addcarryx_u32(&x1201, &x1202, 0x0, x1195, x1198);
+  fiat_p384_addcarryx_u32(&x1201, &x1202, 0x0, x1198, x1195);
   uint32_t x1203;
   fiat_p384_uint1 x1204;
-  fiat_p384_addcarryx_u32(&x1203, &x1204, x1202, x1193, x1196);
+  fiat_p384_addcarryx_u32(&x1203, &x1204, x1202, x1196, x1193);
   uint32_t x1205;
   fiat_p384_uint1 x1206;
-  fiat_p384_addcarryx_u32(&x1205, &x1206, x1204, x1191, x1194);
+  fiat_p384_addcarryx_u32(&x1205, &x1206, x1204, x1194, x1191);
   uint32_t x1207;
   fiat_p384_uint1 x1208;
-  fiat_p384_addcarryx_u32(&x1207, &x1208, x1206, x1189, x1192);
+  fiat_p384_addcarryx_u32(&x1207, &x1208, x1206, x1192, x1189);
   uint32_t x1209;
   fiat_p384_uint1 x1210;
-  fiat_p384_addcarryx_u32(&x1209, &x1210, x1208, x1187, x1190);
+  fiat_p384_addcarryx_u32(&x1209, &x1210, x1208, x1190, x1187);
   uint32_t x1211;
   fiat_p384_uint1 x1212;
-  fiat_p384_addcarryx_u32(&x1211, &x1212, x1210, x1185, x1188);
+  fiat_p384_addcarryx_u32(&x1211, &x1212, x1210, x1188, x1185);
   uint32_t x1213;
   fiat_p384_uint1 x1214;
-  fiat_p384_addcarryx_u32(&x1213, &x1214, x1212, x1183, x1186);
+  fiat_p384_addcarryx_u32(&x1213, &x1214, x1212, x1186, x1183);
   uint32_t x1215;
   fiat_p384_uint1 x1216;
-  fiat_p384_addcarryx_u32(&x1215, &x1216, x1214, x1181, x1184);
+  fiat_p384_addcarryx_u32(&x1215, &x1216, x1214, x1184, x1181);
   uint32_t x1217;
   fiat_p384_uint1 x1218;
-  fiat_p384_addcarryx_u32(&x1217, &x1218, x1216, 0x0, x1182);
+  fiat_p384_addcarryx_u32(&x1217, &x1218, x1216, x1182, 0x0);
   uint32_t x1219;
   fiat_p384_uint1 x1220;
-  fiat_p384_addcarryx_u32(&x1219, &x1220, 0x0, x1199, x1155);
+  fiat_p384_addcarryx_u32(&x1219, &x1220, 0x0, x1155, x1199);
   uint32_t x1221;
   fiat_p384_uint1 x1222;
-  fiat_p384_addcarryx_u32(&x1221, &x1222, x1220, x1200, x1157);
+  fiat_p384_addcarryx_u32(&x1221, &x1222, x1220, x1157, x1200);
   uint32_t x1223;
   fiat_p384_uint1 x1224;
-  fiat_p384_addcarryx_u32(&x1223, &x1224, x1222, 0x0, x1159);
+  fiat_p384_addcarryx_u32(&x1223, &x1224, x1222, x1159, 0x0);
   uint32_t x1225;
   fiat_p384_uint1 x1226;
-  fiat_p384_addcarryx_u32(&x1225, &x1226, x1224, x1197, x1161);
+  fiat_p384_addcarryx_u32(&x1225, &x1226, x1224, x1161, x1197);
   uint32_t x1227;
   fiat_p384_uint1 x1228;
-  fiat_p384_addcarryx_u32(&x1227, &x1228, x1226, x1201, x1163);
+  fiat_p384_addcarryx_u32(&x1227, &x1228, x1226, x1163, x1201);
   uint32_t x1229;
   fiat_p384_uint1 x1230;
-  fiat_p384_addcarryx_u32(&x1229, &x1230, x1228, x1203, x1165);
+  fiat_p384_addcarryx_u32(&x1229, &x1230, x1228, x1165, x1203);
   uint32_t x1231;
   fiat_p384_uint1 x1232;
-  fiat_p384_addcarryx_u32(&x1231, &x1232, x1230, x1205, x1167);
+  fiat_p384_addcarryx_u32(&x1231, &x1232, x1230, x1167, x1205);
   uint32_t x1233;
   fiat_p384_uint1 x1234;
-  fiat_p384_addcarryx_u32(&x1233, &x1234, x1232, x1207, x1169);
+  fiat_p384_addcarryx_u32(&x1233, &x1234, x1232, x1169, x1207);
   uint32_t x1235;
   fiat_p384_uint1 x1236;
-  fiat_p384_addcarryx_u32(&x1235, &x1236, x1234, x1209, x1171);
+  fiat_p384_addcarryx_u32(&x1235, &x1236, x1234, x1171, x1209);
   uint32_t x1237;
   fiat_p384_uint1 x1238;
-  fiat_p384_addcarryx_u32(&x1237, &x1238, x1236, x1211, x1173);
+  fiat_p384_addcarryx_u32(&x1237, &x1238, x1236, x1173, x1211);
   uint32_t x1239;
   fiat_p384_uint1 x1240;
-  fiat_p384_addcarryx_u32(&x1239, &x1240, x1238, x1213, x1175);
+  fiat_p384_addcarryx_u32(&x1239, &x1240, x1238, x1175, x1213);
   uint32_t x1241;
   fiat_p384_uint1 x1242;
-  fiat_p384_addcarryx_u32(&x1241, &x1242, x1240, x1215, x1177);
+  fiat_p384_addcarryx_u32(&x1241, &x1242, x1240, x1177, x1215);
   uint32_t x1243;
   fiat_p384_uint1 x1244;
-  fiat_p384_addcarryx_u32(&x1243, &x1244, x1242, x1217, x1179);
+  fiat_p384_addcarryx_u32(&x1243, &x1244, x1242, x1179, x1217);
   uint32_t x1245;
   fiat_p384_uint1 x1246;
-  fiat_p384_addcarryx_u32(&x1245, &x1246, x1244, 0x0, x1180);
+  fiat_p384_addcarryx_u32(&x1245, &x1246, x1244, x1180, 0x0);
   uint32_t x1247;
   uint32_t x1248;
   fiat_p384_mulx_u32(&x1247, &x1248, x9, (arg1[11]));
@@ -4601,79 +4601,79 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x1269, &x1270, x9, (arg1[0]));
   uint32_t x1271;
   fiat_p384_uint1 x1272;
-  fiat_p384_addcarryx_u32(&x1271, &x1272, 0x0, x1267, x1270);
+  fiat_p384_addcarryx_u32(&x1271, &x1272, 0x0, x1270, x1267);
   uint32_t x1273;
   fiat_p384_uint1 x1274;
-  fiat_p384_addcarryx_u32(&x1273, &x1274, x1272, x1265, x1268);
+  fiat_p384_addcarryx_u32(&x1273, &x1274, x1272, x1268, x1265);
   uint32_t x1275;
   fiat_p384_uint1 x1276;
-  fiat_p384_addcarryx_u32(&x1275, &x1276, x1274, x1263, x1266);
+  fiat_p384_addcarryx_u32(&x1275, &x1276, x1274, x1266, x1263);
   uint32_t x1277;
   fiat_p384_uint1 x1278;
-  fiat_p384_addcarryx_u32(&x1277, &x1278, x1276, x1261, x1264);
+  fiat_p384_addcarryx_u32(&x1277, &x1278, x1276, x1264, x1261);
   uint32_t x1279;
   fiat_p384_uint1 x1280;
-  fiat_p384_addcarryx_u32(&x1279, &x1280, x1278, x1259, x1262);
+  fiat_p384_addcarryx_u32(&x1279, &x1280, x1278, x1262, x1259);
   uint32_t x1281;
   fiat_p384_uint1 x1282;
-  fiat_p384_addcarryx_u32(&x1281, &x1282, x1280, x1257, x1260);
+  fiat_p384_addcarryx_u32(&x1281, &x1282, x1280, x1260, x1257);
   uint32_t x1283;
   fiat_p384_uint1 x1284;
-  fiat_p384_addcarryx_u32(&x1283, &x1284, x1282, x1255, x1258);
+  fiat_p384_addcarryx_u32(&x1283, &x1284, x1282, x1258, x1255);
   uint32_t x1285;
   fiat_p384_uint1 x1286;
-  fiat_p384_addcarryx_u32(&x1285, &x1286, x1284, x1253, x1256);
+  fiat_p384_addcarryx_u32(&x1285, &x1286, x1284, x1256, x1253);
   uint32_t x1287;
   fiat_p384_uint1 x1288;
-  fiat_p384_addcarryx_u32(&x1287, &x1288, x1286, x1251, x1254);
+  fiat_p384_addcarryx_u32(&x1287, &x1288, x1286, x1254, x1251);
   uint32_t x1289;
   fiat_p384_uint1 x1290;
-  fiat_p384_addcarryx_u32(&x1289, &x1290, x1288, x1249, x1252);
+  fiat_p384_addcarryx_u32(&x1289, &x1290, x1288, x1252, x1249);
   uint32_t x1291;
   fiat_p384_uint1 x1292;
-  fiat_p384_addcarryx_u32(&x1291, &x1292, x1290, x1247, x1250);
+  fiat_p384_addcarryx_u32(&x1291, &x1292, x1290, x1250, x1247);
   uint32_t x1293;
   fiat_p384_uint1 x1294;
-  fiat_p384_addcarryx_u32(&x1293, &x1294, x1292, 0x0, x1248);
+  fiat_p384_addcarryx_u32(&x1293, &x1294, x1292, x1248, 0x0);
   uint32_t x1295;
   fiat_p384_uint1 x1296;
-  fiat_p384_addcarryx_u32(&x1295, &x1296, 0x0, x1269, x1221);
+  fiat_p384_addcarryx_u32(&x1295, &x1296, 0x0, x1221, x1269);
   uint32_t x1297;
   fiat_p384_uint1 x1298;
-  fiat_p384_addcarryx_u32(&x1297, &x1298, x1296, x1271, x1223);
+  fiat_p384_addcarryx_u32(&x1297, &x1298, x1296, x1223, x1271);
   uint32_t x1299;
   fiat_p384_uint1 x1300;
-  fiat_p384_addcarryx_u32(&x1299, &x1300, x1298, x1273, x1225);
+  fiat_p384_addcarryx_u32(&x1299, &x1300, x1298, x1225, x1273);
   uint32_t x1301;
   fiat_p384_uint1 x1302;
-  fiat_p384_addcarryx_u32(&x1301, &x1302, x1300, x1275, x1227);
+  fiat_p384_addcarryx_u32(&x1301, &x1302, x1300, x1227, x1275);
   uint32_t x1303;
   fiat_p384_uint1 x1304;
-  fiat_p384_addcarryx_u32(&x1303, &x1304, x1302, x1277, x1229);
+  fiat_p384_addcarryx_u32(&x1303, &x1304, x1302, x1229, x1277);
   uint32_t x1305;
   fiat_p384_uint1 x1306;
-  fiat_p384_addcarryx_u32(&x1305, &x1306, x1304, x1279, x1231);
+  fiat_p384_addcarryx_u32(&x1305, &x1306, x1304, x1231, x1279);
   uint32_t x1307;
   fiat_p384_uint1 x1308;
-  fiat_p384_addcarryx_u32(&x1307, &x1308, x1306, x1281, x1233);
+  fiat_p384_addcarryx_u32(&x1307, &x1308, x1306, x1233, x1281);
   uint32_t x1309;
   fiat_p384_uint1 x1310;
-  fiat_p384_addcarryx_u32(&x1309, &x1310, x1308, x1283, x1235);
+  fiat_p384_addcarryx_u32(&x1309, &x1310, x1308, x1235, x1283);
   uint32_t x1311;
   fiat_p384_uint1 x1312;
-  fiat_p384_addcarryx_u32(&x1311, &x1312, x1310, x1285, x1237);
+  fiat_p384_addcarryx_u32(&x1311, &x1312, x1310, x1237, x1285);
   uint32_t x1313;
   fiat_p384_uint1 x1314;
-  fiat_p384_addcarryx_u32(&x1313, &x1314, x1312, x1287, x1239);
+  fiat_p384_addcarryx_u32(&x1313, &x1314, x1312, x1239, x1287);
   uint32_t x1315;
   fiat_p384_uint1 x1316;
-  fiat_p384_addcarryx_u32(&x1315, &x1316, x1314, x1289, x1241);
+  fiat_p384_addcarryx_u32(&x1315, &x1316, x1314, x1241, x1289);
   uint32_t x1317;
   fiat_p384_uint1 x1318;
-  fiat_p384_addcarryx_u32(&x1317, &x1318, x1316, x1291, x1243);
+  fiat_p384_addcarryx_u32(&x1317, &x1318, x1316, x1243, x1291);
   uint32_t x1319;
   fiat_p384_uint1 x1320;
-  fiat_p384_addcarryx_u32(&x1319, &x1320, x1318, x1293, x1245);
+  fiat_p384_addcarryx_u32(&x1319, &x1320, x1318, x1245, x1293);
   uint32_t x1321;
   uint32_t x1322;
   fiat_p384_mulx_u32(&x1321, &x1322, x1295, UINT32_C(0xffffffff));
@@ -4706,73 +4706,73 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x1339, &x1340, x1295, UINT32_C(0xffffffff));
   uint32_t x1341;
   fiat_p384_uint1 x1342;
-  fiat_p384_addcarryx_u32(&x1341, &x1342, 0x0, x1335, x1338);
+  fiat_p384_addcarryx_u32(&x1341, &x1342, 0x0, x1338, x1335);
   uint32_t x1343;
   fiat_p384_uint1 x1344;
-  fiat_p384_addcarryx_u32(&x1343, &x1344, x1342, x1333, x1336);
+  fiat_p384_addcarryx_u32(&x1343, &x1344, x1342, x1336, x1333);
   uint32_t x1345;
   fiat_p384_uint1 x1346;
-  fiat_p384_addcarryx_u32(&x1345, &x1346, x1344, x1331, x1334);
+  fiat_p384_addcarryx_u32(&x1345, &x1346, x1344, x1334, x1331);
   uint32_t x1347;
   fiat_p384_uint1 x1348;
-  fiat_p384_addcarryx_u32(&x1347, &x1348, x1346, x1329, x1332);
+  fiat_p384_addcarryx_u32(&x1347, &x1348, x1346, x1332, x1329);
   uint32_t x1349;
   fiat_p384_uint1 x1350;
-  fiat_p384_addcarryx_u32(&x1349, &x1350, x1348, x1327, x1330);
+  fiat_p384_addcarryx_u32(&x1349, &x1350, x1348, x1330, x1327);
   uint32_t x1351;
   fiat_p384_uint1 x1352;
-  fiat_p384_addcarryx_u32(&x1351, &x1352, x1350, x1325, x1328);
+  fiat_p384_addcarryx_u32(&x1351, &x1352, x1350, x1328, x1325);
   uint32_t x1353;
   fiat_p384_uint1 x1354;
-  fiat_p384_addcarryx_u32(&x1353, &x1354, x1352, x1323, x1326);
+  fiat_p384_addcarryx_u32(&x1353, &x1354, x1352, x1326, x1323);
   uint32_t x1355;
   fiat_p384_uint1 x1356;
-  fiat_p384_addcarryx_u32(&x1355, &x1356, x1354, x1321, x1324);
+  fiat_p384_addcarryx_u32(&x1355, &x1356, x1354, x1324, x1321);
   uint32_t x1357;
   fiat_p384_uint1 x1358;
-  fiat_p384_addcarryx_u32(&x1357, &x1358, x1356, 0x0, x1322);
+  fiat_p384_addcarryx_u32(&x1357, &x1358, x1356, x1322, 0x0);
   uint32_t x1359;
   fiat_p384_uint1 x1360;
-  fiat_p384_addcarryx_u32(&x1359, &x1360, 0x0, x1339, x1295);
+  fiat_p384_addcarryx_u32(&x1359, &x1360, 0x0, x1295, x1339);
   uint32_t x1361;
   fiat_p384_uint1 x1362;
-  fiat_p384_addcarryx_u32(&x1361, &x1362, x1360, x1340, x1297);
+  fiat_p384_addcarryx_u32(&x1361, &x1362, x1360, x1297, x1340);
   uint32_t x1363;
   fiat_p384_uint1 x1364;
-  fiat_p384_addcarryx_u32(&x1363, &x1364, x1362, 0x0, x1299);
+  fiat_p384_addcarryx_u32(&x1363, &x1364, x1362, x1299, 0x0);
   uint32_t x1365;
   fiat_p384_uint1 x1366;
-  fiat_p384_addcarryx_u32(&x1365, &x1366, x1364, x1337, x1301);
+  fiat_p384_addcarryx_u32(&x1365, &x1366, x1364, x1301, x1337);
   uint32_t x1367;
   fiat_p384_uint1 x1368;
-  fiat_p384_addcarryx_u32(&x1367, &x1368, x1366, x1341, x1303);
+  fiat_p384_addcarryx_u32(&x1367, &x1368, x1366, x1303, x1341);
   uint32_t x1369;
   fiat_p384_uint1 x1370;
-  fiat_p384_addcarryx_u32(&x1369, &x1370, x1368, x1343, x1305);
+  fiat_p384_addcarryx_u32(&x1369, &x1370, x1368, x1305, x1343);
   uint32_t x1371;
   fiat_p384_uint1 x1372;
-  fiat_p384_addcarryx_u32(&x1371, &x1372, x1370, x1345, x1307);
+  fiat_p384_addcarryx_u32(&x1371, &x1372, x1370, x1307, x1345);
   uint32_t x1373;
   fiat_p384_uint1 x1374;
-  fiat_p384_addcarryx_u32(&x1373, &x1374, x1372, x1347, x1309);
+  fiat_p384_addcarryx_u32(&x1373, &x1374, x1372, x1309, x1347);
   uint32_t x1375;
   fiat_p384_uint1 x1376;
-  fiat_p384_addcarryx_u32(&x1375, &x1376, x1374, x1349, x1311);
+  fiat_p384_addcarryx_u32(&x1375, &x1376, x1374, x1311, x1349);
   uint32_t x1377;
   fiat_p384_uint1 x1378;
-  fiat_p384_addcarryx_u32(&x1377, &x1378, x1376, x1351, x1313);
+  fiat_p384_addcarryx_u32(&x1377, &x1378, x1376, x1313, x1351);
   uint32_t x1379;
   fiat_p384_uint1 x1380;
-  fiat_p384_addcarryx_u32(&x1379, &x1380, x1378, x1353, x1315);
+  fiat_p384_addcarryx_u32(&x1379, &x1380, x1378, x1315, x1353);
   uint32_t x1381;
   fiat_p384_uint1 x1382;
-  fiat_p384_addcarryx_u32(&x1381, &x1382, x1380, x1355, x1317);
+  fiat_p384_addcarryx_u32(&x1381, &x1382, x1380, x1317, x1355);
   uint32_t x1383;
   fiat_p384_uint1 x1384;
-  fiat_p384_addcarryx_u32(&x1383, &x1384, x1382, x1357, x1319);
+  fiat_p384_addcarryx_u32(&x1383, &x1384, x1382, x1319, x1357);
   uint32_t x1385;
   fiat_p384_uint1 x1386;
-  fiat_p384_addcarryx_u32(&x1385, &x1386, x1384, 0x0, x1320);
+  fiat_p384_addcarryx_u32(&x1385, &x1386, x1384, x1320, 0x0);
   uint32_t x1387;
   uint32_t x1388;
   fiat_p384_mulx_u32(&x1387, &x1388, x10, (arg1[11]));
@@ -4811,79 +4811,79 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x1409, &x1410, x10, (arg1[0]));
   uint32_t x1411;
   fiat_p384_uint1 x1412;
-  fiat_p384_addcarryx_u32(&x1411, &x1412, 0x0, x1407, x1410);
+  fiat_p384_addcarryx_u32(&x1411, &x1412, 0x0, x1410, x1407);
   uint32_t x1413;
   fiat_p384_uint1 x1414;
-  fiat_p384_addcarryx_u32(&x1413, &x1414, x1412, x1405, x1408);
+  fiat_p384_addcarryx_u32(&x1413, &x1414, x1412, x1408, x1405);
   uint32_t x1415;
   fiat_p384_uint1 x1416;
-  fiat_p384_addcarryx_u32(&x1415, &x1416, x1414, x1403, x1406);
+  fiat_p384_addcarryx_u32(&x1415, &x1416, x1414, x1406, x1403);
   uint32_t x1417;
   fiat_p384_uint1 x1418;
-  fiat_p384_addcarryx_u32(&x1417, &x1418, x1416, x1401, x1404);
+  fiat_p384_addcarryx_u32(&x1417, &x1418, x1416, x1404, x1401);
   uint32_t x1419;
   fiat_p384_uint1 x1420;
-  fiat_p384_addcarryx_u32(&x1419, &x1420, x1418, x1399, x1402);
+  fiat_p384_addcarryx_u32(&x1419, &x1420, x1418, x1402, x1399);
   uint32_t x1421;
   fiat_p384_uint1 x1422;
-  fiat_p384_addcarryx_u32(&x1421, &x1422, x1420, x1397, x1400);
+  fiat_p384_addcarryx_u32(&x1421, &x1422, x1420, x1400, x1397);
   uint32_t x1423;
   fiat_p384_uint1 x1424;
-  fiat_p384_addcarryx_u32(&x1423, &x1424, x1422, x1395, x1398);
+  fiat_p384_addcarryx_u32(&x1423, &x1424, x1422, x1398, x1395);
   uint32_t x1425;
   fiat_p384_uint1 x1426;
-  fiat_p384_addcarryx_u32(&x1425, &x1426, x1424, x1393, x1396);
+  fiat_p384_addcarryx_u32(&x1425, &x1426, x1424, x1396, x1393);
   uint32_t x1427;
   fiat_p384_uint1 x1428;
-  fiat_p384_addcarryx_u32(&x1427, &x1428, x1426, x1391, x1394);
+  fiat_p384_addcarryx_u32(&x1427, &x1428, x1426, x1394, x1391);
   uint32_t x1429;
   fiat_p384_uint1 x1430;
-  fiat_p384_addcarryx_u32(&x1429, &x1430, x1428, x1389, x1392);
+  fiat_p384_addcarryx_u32(&x1429, &x1430, x1428, x1392, x1389);
   uint32_t x1431;
   fiat_p384_uint1 x1432;
-  fiat_p384_addcarryx_u32(&x1431, &x1432, x1430, x1387, x1390);
+  fiat_p384_addcarryx_u32(&x1431, &x1432, x1430, x1390, x1387);
   uint32_t x1433;
   fiat_p384_uint1 x1434;
-  fiat_p384_addcarryx_u32(&x1433, &x1434, x1432, 0x0, x1388);
+  fiat_p384_addcarryx_u32(&x1433, &x1434, x1432, x1388, 0x0);
   uint32_t x1435;
   fiat_p384_uint1 x1436;
-  fiat_p384_addcarryx_u32(&x1435, &x1436, 0x0, x1409, x1361);
+  fiat_p384_addcarryx_u32(&x1435, &x1436, 0x0, x1361, x1409);
   uint32_t x1437;
   fiat_p384_uint1 x1438;
-  fiat_p384_addcarryx_u32(&x1437, &x1438, x1436, x1411, x1363);
+  fiat_p384_addcarryx_u32(&x1437, &x1438, x1436, x1363, x1411);
   uint32_t x1439;
   fiat_p384_uint1 x1440;
-  fiat_p384_addcarryx_u32(&x1439, &x1440, x1438, x1413, x1365);
+  fiat_p384_addcarryx_u32(&x1439, &x1440, x1438, x1365, x1413);
   uint32_t x1441;
   fiat_p384_uint1 x1442;
-  fiat_p384_addcarryx_u32(&x1441, &x1442, x1440, x1415, x1367);
+  fiat_p384_addcarryx_u32(&x1441, &x1442, x1440, x1367, x1415);
   uint32_t x1443;
   fiat_p384_uint1 x1444;
-  fiat_p384_addcarryx_u32(&x1443, &x1444, x1442, x1417, x1369);
+  fiat_p384_addcarryx_u32(&x1443, &x1444, x1442, x1369, x1417);
   uint32_t x1445;
   fiat_p384_uint1 x1446;
-  fiat_p384_addcarryx_u32(&x1445, &x1446, x1444, x1419, x1371);
+  fiat_p384_addcarryx_u32(&x1445, &x1446, x1444, x1371, x1419);
   uint32_t x1447;
   fiat_p384_uint1 x1448;
-  fiat_p384_addcarryx_u32(&x1447, &x1448, x1446, x1421, x1373);
+  fiat_p384_addcarryx_u32(&x1447, &x1448, x1446, x1373, x1421);
   uint32_t x1449;
   fiat_p384_uint1 x1450;
-  fiat_p384_addcarryx_u32(&x1449, &x1450, x1448, x1423, x1375);
+  fiat_p384_addcarryx_u32(&x1449, &x1450, x1448, x1375, x1423);
   uint32_t x1451;
   fiat_p384_uint1 x1452;
-  fiat_p384_addcarryx_u32(&x1451, &x1452, x1450, x1425, x1377);
+  fiat_p384_addcarryx_u32(&x1451, &x1452, x1450, x1377, x1425);
   uint32_t x1453;
   fiat_p384_uint1 x1454;
-  fiat_p384_addcarryx_u32(&x1453, &x1454, x1452, x1427, x1379);
+  fiat_p384_addcarryx_u32(&x1453, &x1454, x1452, x1379, x1427);
   uint32_t x1455;
   fiat_p384_uint1 x1456;
-  fiat_p384_addcarryx_u32(&x1455, &x1456, x1454, x1429, x1381);
+  fiat_p384_addcarryx_u32(&x1455, &x1456, x1454, x1381, x1429);
   uint32_t x1457;
   fiat_p384_uint1 x1458;
-  fiat_p384_addcarryx_u32(&x1457, &x1458, x1456, x1431, x1383);
+  fiat_p384_addcarryx_u32(&x1457, &x1458, x1456, x1383, x1431);
   uint32_t x1459;
   fiat_p384_uint1 x1460;
-  fiat_p384_addcarryx_u32(&x1459, &x1460, x1458, x1433, x1385);
+  fiat_p384_addcarryx_u32(&x1459, &x1460, x1458, x1385, x1433);
   uint32_t x1461;
   uint32_t x1462;
   fiat_p384_mulx_u32(&x1461, &x1462, x1435, UINT32_C(0xffffffff));
@@ -4916,73 +4916,73 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x1479, &x1480, x1435, UINT32_C(0xffffffff));
   uint32_t x1481;
   fiat_p384_uint1 x1482;
-  fiat_p384_addcarryx_u32(&x1481, &x1482, 0x0, x1475, x1478);
+  fiat_p384_addcarryx_u32(&x1481, &x1482, 0x0, x1478, x1475);
   uint32_t x1483;
   fiat_p384_uint1 x1484;
-  fiat_p384_addcarryx_u32(&x1483, &x1484, x1482, x1473, x1476);
+  fiat_p384_addcarryx_u32(&x1483, &x1484, x1482, x1476, x1473);
   uint32_t x1485;
   fiat_p384_uint1 x1486;
-  fiat_p384_addcarryx_u32(&x1485, &x1486, x1484, x1471, x1474);
+  fiat_p384_addcarryx_u32(&x1485, &x1486, x1484, x1474, x1471);
   uint32_t x1487;
   fiat_p384_uint1 x1488;
-  fiat_p384_addcarryx_u32(&x1487, &x1488, x1486, x1469, x1472);
+  fiat_p384_addcarryx_u32(&x1487, &x1488, x1486, x1472, x1469);
   uint32_t x1489;
   fiat_p384_uint1 x1490;
-  fiat_p384_addcarryx_u32(&x1489, &x1490, x1488, x1467, x1470);
+  fiat_p384_addcarryx_u32(&x1489, &x1490, x1488, x1470, x1467);
   uint32_t x1491;
   fiat_p384_uint1 x1492;
-  fiat_p384_addcarryx_u32(&x1491, &x1492, x1490, x1465, x1468);
+  fiat_p384_addcarryx_u32(&x1491, &x1492, x1490, x1468, x1465);
   uint32_t x1493;
   fiat_p384_uint1 x1494;
-  fiat_p384_addcarryx_u32(&x1493, &x1494, x1492, x1463, x1466);
+  fiat_p384_addcarryx_u32(&x1493, &x1494, x1492, x1466, x1463);
   uint32_t x1495;
   fiat_p384_uint1 x1496;
-  fiat_p384_addcarryx_u32(&x1495, &x1496, x1494, x1461, x1464);
+  fiat_p384_addcarryx_u32(&x1495, &x1496, x1494, x1464, x1461);
   uint32_t x1497;
   fiat_p384_uint1 x1498;
-  fiat_p384_addcarryx_u32(&x1497, &x1498, x1496, 0x0, x1462);
+  fiat_p384_addcarryx_u32(&x1497, &x1498, x1496, x1462, 0x0);
   uint32_t x1499;
   fiat_p384_uint1 x1500;
-  fiat_p384_addcarryx_u32(&x1499, &x1500, 0x0, x1479, x1435);
+  fiat_p384_addcarryx_u32(&x1499, &x1500, 0x0, x1435, x1479);
   uint32_t x1501;
   fiat_p384_uint1 x1502;
-  fiat_p384_addcarryx_u32(&x1501, &x1502, x1500, x1480, x1437);
+  fiat_p384_addcarryx_u32(&x1501, &x1502, x1500, x1437, x1480);
   uint32_t x1503;
   fiat_p384_uint1 x1504;
-  fiat_p384_addcarryx_u32(&x1503, &x1504, x1502, 0x0, x1439);
+  fiat_p384_addcarryx_u32(&x1503, &x1504, x1502, x1439, 0x0);
   uint32_t x1505;
   fiat_p384_uint1 x1506;
-  fiat_p384_addcarryx_u32(&x1505, &x1506, x1504, x1477, x1441);
+  fiat_p384_addcarryx_u32(&x1505, &x1506, x1504, x1441, x1477);
   uint32_t x1507;
   fiat_p384_uint1 x1508;
-  fiat_p384_addcarryx_u32(&x1507, &x1508, x1506, x1481, x1443);
+  fiat_p384_addcarryx_u32(&x1507, &x1508, x1506, x1443, x1481);
   uint32_t x1509;
   fiat_p384_uint1 x1510;
-  fiat_p384_addcarryx_u32(&x1509, &x1510, x1508, x1483, x1445);
+  fiat_p384_addcarryx_u32(&x1509, &x1510, x1508, x1445, x1483);
   uint32_t x1511;
   fiat_p384_uint1 x1512;
-  fiat_p384_addcarryx_u32(&x1511, &x1512, x1510, x1485, x1447);
+  fiat_p384_addcarryx_u32(&x1511, &x1512, x1510, x1447, x1485);
   uint32_t x1513;
   fiat_p384_uint1 x1514;
-  fiat_p384_addcarryx_u32(&x1513, &x1514, x1512, x1487, x1449);
+  fiat_p384_addcarryx_u32(&x1513, &x1514, x1512, x1449, x1487);
   uint32_t x1515;
   fiat_p384_uint1 x1516;
-  fiat_p384_addcarryx_u32(&x1515, &x1516, x1514, x1489, x1451);
+  fiat_p384_addcarryx_u32(&x1515, &x1516, x1514, x1451, x1489);
   uint32_t x1517;
   fiat_p384_uint1 x1518;
-  fiat_p384_addcarryx_u32(&x1517, &x1518, x1516, x1491, x1453);
+  fiat_p384_addcarryx_u32(&x1517, &x1518, x1516, x1453, x1491);
   uint32_t x1519;
   fiat_p384_uint1 x1520;
-  fiat_p384_addcarryx_u32(&x1519, &x1520, x1518, x1493, x1455);
+  fiat_p384_addcarryx_u32(&x1519, &x1520, x1518, x1455, x1493);
   uint32_t x1521;
   fiat_p384_uint1 x1522;
-  fiat_p384_addcarryx_u32(&x1521, &x1522, x1520, x1495, x1457);
+  fiat_p384_addcarryx_u32(&x1521, &x1522, x1520, x1457, x1495);
   uint32_t x1523;
   fiat_p384_uint1 x1524;
-  fiat_p384_addcarryx_u32(&x1523, &x1524, x1522, x1497, x1459);
+  fiat_p384_addcarryx_u32(&x1523, &x1524, x1522, x1459, x1497);
   uint32_t x1525;
   fiat_p384_uint1 x1526;
-  fiat_p384_addcarryx_u32(&x1525, &x1526, x1524, 0x0, x1460);
+  fiat_p384_addcarryx_u32(&x1525, &x1526, x1524, x1460, 0x0);
   uint32_t x1527;
   uint32_t x1528;
   fiat_p384_mulx_u32(&x1527, &x1528, x11, (arg1[11]));
@@ -5021,79 +5021,79 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x1549, &x1550, x11, (arg1[0]));
   uint32_t x1551;
   fiat_p384_uint1 x1552;
-  fiat_p384_addcarryx_u32(&x1551, &x1552, 0x0, x1547, x1550);
+  fiat_p384_addcarryx_u32(&x1551, &x1552, 0x0, x1550, x1547);
   uint32_t x1553;
   fiat_p384_uint1 x1554;
-  fiat_p384_addcarryx_u32(&x1553, &x1554, x1552, x1545, x1548);
+  fiat_p384_addcarryx_u32(&x1553, &x1554, x1552, x1548, x1545);
   uint32_t x1555;
   fiat_p384_uint1 x1556;
-  fiat_p384_addcarryx_u32(&x1555, &x1556, x1554, x1543, x1546);
+  fiat_p384_addcarryx_u32(&x1555, &x1556, x1554, x1546, x1543);
   uint32_t x1557;
   fiat_p384_uint1 x1558;
-  fiat_p384_addcarryx_u32(&x1557, &x1558, x1556, x1541, x1544);
+  fiat_p384_addcarryx_u32(&x1557, &x1558, x1556, x1544, x1541);
   uint32_t x1559;
   fiat_p384_uint1 x1560;
-  fiat_p384_addcarryx_u32(&x1559, &x1560, x1558, x1539, x1542);
+  fiat_p384_addcarryx_u32(&x1559, &x1560, x1558, x1542, x1539);
   uint32_t x1561;
   fiat_p384_uint1 x1562;
-  fiat_p384_addcarryx_u32(&x1561, &x1562, x1560, x1537, x1540);
+  fiat_p384_addcarryx_u32(&x1561, &x1562, x1560, x1540, x1537);
   uint32_t x1563;
   fiat_p384_uint1 x1564;
-  fiat_p384_addcarryx_u32(&x1563, &x1564, x1562, x1535, x1538);
+  fiat_p384_addcarryx_u32(&x1563, &x1564, x1562, x1538, x1535);
   uint32_t x1565;
   fiat_p384_uint1 x1566;
-  fiat_p384_addcarryx_u32(&x1565, &x1566, x1564, x1533, x1536);
+  fiat_p384_addcarryx_u32(&x1565, &x1566, x1564, x1536, x1533);
   uint32_t x1567;
   fiat_p384_uint1 x1568;
-  fiat_p384_addcarryx_u32(&x1567, &x1568, x1566, x1531, x1534);
+  fiat_p384_addcarryx_u32(&x1567, &x1568, x1566, x1534, x1531);
   uint32_t x1569;
   fiat_p384_uint1 x1570;
-  fiat_p384_addcarryx_u32(&x1569, &x1570, x1568, x1529, x1532);
+  fiat_p384_addcarryx_u32(&x1569, &x1570, x1568, x1532, x1529);
   uint32_t x1571;
   fiat_p384_uint1 x1572;
-  fiat_p384_addcarryx_u32(&x1571, &x1572, x1570, x1527, x1530);
+  fiat_p384_addcarryx_u32(&x1571, &x1572, x1570, x1530, x1527);
   uint32_t x1573;
   fiat_p384_uint1 x1574;
-  fiat_p384_addcarryx_u32(&x1573, &x1574, x1572, 0x0, x1528);
+  fiat_p384_addcarryx_u32(&x1573, &x1574, x1572, x1528, 0x0);
   uint32_t x1575;
   fiat_p384_uint1 x1576;
-  fiat_p384_addcarryx_u32(&x1575, &x1576, 0x0, x1549, x1501);
+  fiat_p384_addcarryx_u32(&x1575, &x1576, 0x0, x1501, x1549);
   uint32_t x1577;
   fiat_p384_uint1 x1578;
-  fiat_p384_addcarryx_u32(&x1577, &x1578, x1576, x1551, x1503);
+  fiat_p384_addcarryx_u32(&x1577, &x1578, x1576, x1503, x1551);
   uint32_t x1579;
   fiat_p384_uint1 x1580;
-  fiat_p384_addcarryx_u32(&x1579, &x1580, x1578, x1553, x1505);
+  fiat_p384_addcarryx_u32(&x1579, &x1580, x1578, x1505, x1553);
   uint32_t x1581;
   fiat_p384_uint1 x1582;
-  fiat_p384_addcarryx_u32(&x1581, &x1582, x1580, x1555, x1507);
+  fiat_p384_addcarryx_u32(&x1581, &x1582, x1580, x1507, x1555);
   uint32_t x1583;
   fiat_p384_uint1 x1584;
-  fiat_p384_addcarryx_u32(&x1583, &x1584, x1582, x1557, x1509);
+  fiat_p384_addcarryx_u32(&x1583, &x1584, x1582, x1509, x1557);
   uint32_t x1585;
   fiat_p384_uint1 x1586;
-  fiat_p384_addcarryx_u32(&x1585, &x1586, x1584, x1559, x1511);
+  fiat_p384_addcarryx_u32(&x1585, &x1586, x1584, x1511, x1559);
   uint32_t x1587;
   fiat_p384_uint1 x1588;
-  fiat_p384_addcarryx_u32(&x1587, &x1588, x1586, x1561, x1513);
+  fiat_p384_addcarryx_u32(&x1587, &x1588, x1586, x1513, x1561);
   uint32_t x1589;
   fiat_p384_uint1 x1590;
-  fiat_p384_addcarryx_u32(&x1589, &x1590, x1588, x1563, x1515);
+  fiat_p384_addcarryx_u32(&x1589, &x1590, x1588, x1515, x1563);
   uint32_t x1591;
   fiat_p384_uint1 x1592;
-  fiat_p384_addcarryx_u32(&x1591, &x1592, x1590, x1565, x1517);
+  fiat_p384_addcarryx_u32(&x1591, &x1592, x1590, x1517, x1565);
   uint32_t x1593;
   fiat_p384_uint1 x1594;
-  fiat_p384_addcarryx_u32(&x1593, &x1594, x1592, x1567, x1519);
+  fiat_p384_addcarryx_u32(&x1593, &x1594, x1592, x1519, x1567);
   uint32_t x1595;
   fiat_p384_uint1 x1596;
-  fiat_p384_addcarryx_u32(&x1595, &x1596, x1594, x1569, x1521);
+  fiat_p384_addcarryx_u32(&x1595, &x1596, x1594, x1521, x1569);
   uint32_t x1597;
   fiat_p384_uint1 x1598;
-  fiat_p384_addcarryx_u32(&x1597, &x1598, x1596, x1571, x1523);
+  fiat_p384_addcarryx_u32(&x1597, &x1598, x1596, x1523, x1571);
   uint32_t x1599;
   fiat_p384_uint1 x1600;
-  fiat_p384_addcarryx_u32(&x1599, &x1600, x1598, x1573, x1525);
+  fiat_p384_addcarryx_u32(&x1599, &x1600, x1598, x1525, x1573);
   uint32_t x1601;
   uint32_t x1602;
   fiat_p384_mulx_u32(&x1601, &x1602, x1575, UINT32_C(0xffffffff));
@@ -5126,73 +5126,73 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_mulx_u32(&x1619, &x1620, x1575, UINT32_C(0xffffffff));
   uint32_t x1621;
   fiat_p384_uint1 x1622;
-  fiat_p384_addcarryx_u32(&x1621, &x1622, 0x0, x1615, x1618);
+  fiat_p384_addcarryx_u32(&x1621, &x1622, 0x0, x1618, x1615);
   uint32_t x1623;
   fiat_p384_uint1 x1624;
-  fiat_p384_addcarryx_u32(&x1623, &x1624, x1622, x1613, x1616);
+  fiat_p384_addcarryx_u32(&x1623, &x1624, x1622, x1616, x1613);
   uint32_t x1625;
   fiat_p384_uint1 x1626;
-  fiat_p384_addcarryx_u32(&x1625, &x1626, x1624, x1611, x1614);
+  fiat_p384_addcarryx_u32(&x1625, &x1626, x1624, x1614, x1611);
   uint32_t x1627;
   fiat_p384_uint1 x1628;
-  fiat_p384_addcarryx_u32(&x1627, &x1628, x1626, x1609, x1612);
+  fiat_p384_addcarryx_u32(&x1627, &x1628, x1626, x1612, x1609);
   uint32_t x1629;
   fiat_p384_uint1 x1630;
-  fiat_p384_addcarryx_u32(&x1629, &x1630, x1628, x1607, x1610);
+  fiat_p384_addcarryx_u32(&x1629, &x1630, x1628, x1610, x1607);
   uint32_t x1631;
   fiat_p384_uint1 x1632;
-  fiat_p384_addcarryx_u32(&x1631, &x1632, x1630, x1605, x1608);
+  fiat_p384_addcarryx_u32(&x1631, &x1632, x1630, x1608, x1605);
   uint32_t x1633;
   fiat_p384_uint1 x1634;
-  fiat_p384_addcarryx_u32(&x1633, &x1634, x1632, x1603, x1606);
+  fiat_p384_addcarryx_u32(&x1633, &x1634, x1632, x1606, x1603);
   uint32_t x1635;
   fiat_p384_uint1 x1636;
-  fiat_p384_addcarryx_u32(&x1635, &x1636, x1634, x1601, x1604);
+  fiat_p384_addcarryx_u32(&x1635, &x1636, x1634, x1604, x1601);
   uint32_t x1637;
   fiat_p384_uint1 x1638;
-  fiat_p384_addcarryx_u32(&x1637, &x1638, x1636, 0x0, x1602);
+  fiat_p384_addcarryx_u32(&x1637, &x1638, x1636, x1602, 0x0);
   uint32_t x1639;
   fiat_p384_uint1 x1640;
-  fiat_p384_addcarryx_u32(&x1639, &x1640, 0x0, x1619, x1575);
+  fiat_p384_addcarryx_u32(&x1639, &x1640, 0x0, x1575, x1619);
   uint32_t x1641;
   fiat_p384_uint1 x1642;
-  fiat_p384_addcarryx_u32(&x1641, &x1642, x1640, x1620, x1577);
+  fiat_p384_addcarryx_u32(&x1641, &x1642, x1640, x1577, x1620);
   uint32_t x1643;
   fiat_p384_uint1 x1644;
-  fiat_p384_addcarryx_u32(&x1643, &x1644, x1642, 0x0, x1579);
+  fiat_p384_addcarryx_u32(&x1643, &x1644, x1642, x1579, 0x0);
   uint32_t x1645;
   fiat_p384_uint1 x1646;
-  fiat_p384_addcarryx_u32(&x1645, &x1646, x1644, x1617, x1581);
+  fiat_p384_addcarryx_u32(&x1645, &x1646, x1644, x1581, x1617);
   uint32_t x1647;
   fiat_p384_uint1 x1648;
-  fiat_p384_addcarryx_u32(&x1647, &x1648, x1646, x1621, x1583);
+  fiat_p384_addcarryx_u32(&x1647, &x1648, x1646, x1583, x1621);
   uint32_t x1649;
   fiat_p384_uint1 x1650;
-  fiat_p384_addcarryx_u32(&x1649, &x1650, x1648, x1623, x1585);
+  fiat_p384_addcarryx_u32(&x1649, &x1650, x1648, x1585, x1623);
   uint32_t x1651;
   fiat_p384_uint1 x1652;
-  fiat_p384_addcarryx_u32(&x1651, &x1652, x1650, x1625, x1587);
+  fiat_p384_addcarryx_u32(&x1651, &x1652, x1650, x1587, x1625);
   uint32_t x1653;
   fiat_p384_uint1 x1654;
-  fiat_p384_addcarryx_u32(&x1653, &x1654, x1652, x1627, x1589);
+  fiat_p384_addcarryx_u32(&x1653, &x1654, x1652, x1589, x1627);
   uint32_t x1655;
   fiat_p384_uint1 x1656;
-  fiat_p384_addcarryx_u32(&x1655, &x1656, x1654, x1629, x1591);
+  fiat_p384_addcarryx_u32(&x1655, &x1656, x1654, x1591, x1629);
   uint32_t x1657;
   fiat_p384_uint1 x1658;
-  fiat_p384_addcarryx_u32(&x1657, &x1658, x1656, x1631, x1593);
+  fiat_p384_addcarryx_u32(&x1657, &x1658, x1656, x1593, x1631);
   uint32_t x1659;
   fiat_p384_uint1 x1660;
-  fiat_p384_addcarryx_u32(&x1659, &x1660, x1658, x1633, x1595);
+  fiat_p384_addcarryx_u32(&x1659, &x1660, x1658, x1595, x1633);
   uint32_t x1661;
   fiat_p384_uint1 x1662;
-  fiat_p384_addcarryx_u32(&x1661, &x1662, x1660, x1635, x1597);
+  fiat_p384_addcarryx_u32(&x1661, &x1662, x1660, x1597, x1635);
   uint32_t x1663;
   fiat_p384_uint1 x1664;
-  fiat_p384_addcarryx_u32(&x1663, &x1664, x1662, x1637, x1599);
+  fiat_p384_addcarryx_u32(&x1663, &x1664, x1662, x1599, x1637);
   uint32_t x1665;
   fiat_p384_uint1 x1666;
-  fiat_p384_addcarryx_u32(&x1665, &x1666, x1664, 0x0, x1600);
+  fiat_p384_addcarryx_u32(&x1665, &x1666, x1664, x1600, 0x0);
   uint32_t x1667;
   fiat_p384_uint1 x1668;
   fiat_p384_subborrowx_u32(&x1667, &x1668, 0x0, x1641, UINT32_C(0xffffffff));
@@ -5288,40 +5288,40 @@ static void fiat_p384_square(uint32_t out1[12], const uint32_t arg1[12]) {
 static void fiat_p384_add(uint32_t out1[12], const uint32_t arg1[12], const uint32_t arg2[12]) {
   uint32_t x1;
   fiat_p384_uint1 x2;
-  fiat_p384_addcarryx_u32(&x1, &x2, 0x0, (arg2[0]), (arg1[0]));
+  fiat_p384_addcarryx_u32(&x1, &x2, 0x0, (arg1[0]), (arg2[0]));
   uint32_t x3;
   fiat_p384_uint1 x4;
-  fiat_p384_addcarryx_u32(&x3, &x4, x2, (arg2[1]), (arg1[1]));
+  fiat_p384_addcarryx_u32(&x3, &x4, x2, (arg1[1]), (arg2[1]));
   uint32_t x5;
   fiat_p384_uint1 x6;
-  fiat_p384_addcarryx_u32(&x5, &x6, x4, (arg2[2]), (arg1[2]));
+  fiat_p384_addcarryx_u32(&x5, &x6, x4, (arg1[2]), (arg2[2]));
   uint32_t x7;
   fiat_p384_uint1 x8;
-  fiat_p384_addcarryx_u32(&x7, &x8, x6, (arg2[3]), (arg1[3]));
+  fiat_p384_addcarryx_u32(&x7, &x8, x6, (arg1[3]), (arg2[3]));
   uint32_t x9;
   fiat_p384_uint1 x10;
-  fiat_p384_addcarryx_u32(&x9, &x10, x8, (arg2[4]), (arg1[4]));
+  fiat_p384_addcarryx_u32(&x9, &x10, x8, (arg1[4]), (arg2[4]));
   uint32_t x11;
   fiat_p384_uint1 x12;
-  fiat_p384_addcarryx_u32(&x11, &x12, x10, (arg2[5]), (arg1[5]));
+  fiat_p384_addcarryx_u32(&x11, &x12, x10, (arg1[5]), (arg2[5]));
   uint32_t x13;
   fiat_p384_uint1 x14;
-  fiat_p384_addcarryx_u32(&x13, &x14, x12, (arg2[6]), (arg1[6]));
+  fiat_p384_addcarryx_u32(&x13, &x14, x12, (arg1[6]), (arg2[6]));
   uint32_t x15;
   fiat_p384_uint1 x16;
-  fiat_p384_addcarryx_u32(&x15, &x16, x14, (arg2[7]), (arg1[7]));
+  fiat_p384_addcarryx_u32(&x15, &x16, x14, (arg1[7]), (arg2[7]));
   uint32_t x17;
   fiat_p384_uint1 x18;
-  fiat_p384_addcarryx_u32(&x17, &x18, x16, (arg2[8]), (arg1[8]));
+  fiat_p384_addcarryx_u32(&x17, &x18, x16, (arg1[8]), (arg2[8]));
   uint32_t x19;
   fiat_p384_uint1 x20;
-  fiat_p384_addcarryx_u32(&x19, &x20, x18, (arg2[9]), (arg1[9]));
+  fiat_p384_addcarryx_u32(&x19, &x20, x18, (arg1[9]), (arg2[9]));
   uint32_t x21;
   fiat_p384_uint1 x22;
-  fiat_p384_addcarryx_u32(&x21, &x22, x20, (arg2[10]), (arg1[10]));
+  fiat_p384_addcarryx_u32(&x21, &x22, x20, (arg1[10]), (arg2[10]));
   uint32_t x23;
   fiat_p384_uint1 x24;
-  fiat_p384_addcarryx_u32(&x23, &x24, x22, (arg2[11]), (arg1[11]));
+  fiat_p384_addcarryx_u32(&x23, &x24, x22, (arg1[11]), (arg2[11]));
   uint32_t x25;
   fiat_p384_uint1 x26;
   fiat_p384_subborrowx_u32(&x25, &x26, 0x0, x1, UINT32_C(0xffffffff));
@@ -5455,40 +5455,40 @@ static void fiat_p384_sub(uint32_t out1[12], const uint32_t arg1[12], const uint
   fiat_p384_cmovznz_u32(&x25, x24, 0x0, UINT32_C(0xffffffff));
   uint32_t x26;
   fiat_p384_uint1 x27;
-  fiat_p384_addcarryx_u32(&x26, &x27, 0x0, (x25 & UINT32_C(0xffffffff)), x1);
+  fiat_p384_addcarryx_u32(&x26, &x27, 0x0, x1, (x25 & UINT32_C(0xffffffff)));
   uint32_t x28;
   fiat_p384_uint1 x29;
-  fiat_p384_addcarryx_u32(&x28, &x29, x27, 0x0, x3);
+  fiat_p384_addcarryx_u32(&x28, &x29, x27, x3, 0x0);
   uint32_t x30;
   fiat_p384_uint1 x31;
-  fiat_p384_addcarryx_u32(&x30, &x31, x29, 0x0, x5);
+  fiat_p384_addcarryx_u32(&x30, &x31, x29, x5, 0x0);
   uint32_t x32;
   fiat_p384_uint1 x33;
-  fiat_p384_addcarryx_u32(&x32, &x33, x31, (x25 & UINT32_C(0xffffffff)), x7);
+  fiat_p384_addcarryx_u32(&x32, &x33, x31, x7, (x25 & UINT32_C(0xffffffff)));
   uint32_t x34;
   fiat_p384_uint1 x35;
-  fiat_p384_addcarryx_u32(&x34, &x35, x33, (x25 & UINT32_C(0xfffffffe)), x9);
+  fiat_p384_addcarryx_u32(&x34, &x35, x33, x9, (x25 & UINT32_C(0xfffffffe)));
   uint32_t x36;
   fiat_p384_uint1 x37;
-  fiat_p384_addcarryx_u32(&x36, &x37, x35, (x25 & UINT32_C(0xffffffff)), x11);
+  fiat_p384_addcarryx_u32(&x36, &x37, x35, x11, (x25 & UINT32_C(0xffffffff)));
   uint32_t x38;
   fiat_p384_uint1 x39;
-  fiat_p384_addcarryx_u32(&x38, &x39, x37, (x25 & UINT32_C(0xffffffff)), x13);
+  fiat_p384_addcarryx_u32(&x38, &x39, x37, x13, (x25 & UINT32_C(0xffffffff)));
   uint32_t x40;
   fiat_p384_uint1 x41;
-  fiat_p384_addcarryx_u32(&x40, &x41, x39, (x25 & UINT32_C(0xffffffff)), x15);
+  fiat_p384_addcarryx_u32(&x40, &x41, x39, x15, (x25 & UINT32_C(0xffffffff)));
   uint32_t x42;
   fiat_p384_uint1 x43;
-  fiat_p384_addcarryx_u32(&x42, &x43, x41, (x25 & UINT32_C(0xffffffff)), x17);
+  fiat_p384_addcarryx_u32(&x42, &x43, x41, x17, (x25 & UINT32_C(0xffffffff)));
   uint32_t x44;
   fiat_p384_uint1 x45;
-  fiat_p384_addcarryx_u32(&x44, &x45, x43, (x25 & UINT32_C(0xffffffff)), x19);
+  fiat_p384_addcarryx_u32(&x44, &x45, x43, x19, (x25 & UINT32_C(0xffffffff)));
   uint32_t x46;
   fiat_p384_uint1 x47;
-  fiat_p384_addcarryx_u32(&x46, &x47, x45, (x25 & UINT32_C(0xffffffff)), x21);
+  fiat_p384_addcarryx_u32(&x46, &x47, x45, x21, (x25 & UINT32_C(0xffffffff)));
   uint32_t x48;
   fiat_p384_uint1 x49;
-  fiat_p384_addcarryx_u32(&x48, &x49, x47, (x25 & UINT32_C(0xffffffff)), x23);
+  fiat_p384_addcarryx_u32(&x48, &x49, x47, x23, (x25 & UINT32_C(0xffffffff)));
   out1[0] = x26;
   out1[1] = x28;
   out1[2] = x30;
@@ -5557,40 +5557,40 @@ static void fiat_p384_opp(uint32_t out1[12], const uint32_t arg1[12]) {
   fiat_p384_cmovznz_u32(&x25, x24, 0x0, UINT32_C(0xffffffff));
   uint32_t x26;
   fiat_p384_uint1 x27;
-  fiat_p384_addcarryx_u32(&x26, &x27, 0x0, (x25 & UINT32_C(0xffffffff)), x1);
+  fiat_p384_addcarryx_u32(&x26, &x27, 0x0, x1, (x25 & UINT32_C(0xffffffff)));
   uint32_t x28;
   fiat_p384_uint1 x29;
-  fiat_p384_addcarryx_u32(&x28, &x29, x27, 0x0, x3);
+  fiat_p384_addcarryx_u32(&x28, &x29, x27, x3, 0x0);
   uint32_t x30;
   fiat_p384_uint1 x31;
-  fiat_p384_addcarryx_u32(&x30, &x31, x29, 0x0, x5);
+  fiat_p384_addcarryx_u32(&x30, &x31, x29, x5, 0x0);
   uint32_t x32;
   fiat_p384_uint1 x33;
-  fiat_p384_addcarryx_u32(&x32, &x33, x31, (x25 & UINT32_C(0xffffffff)), x7);
+  fiat_p384_addcarryx_u32(&x32, &x33, x31, x7, (x25 & UINT32_C(0xffffffff)));
   uint32_t x34;
   fiat_p384_uint1 x35;
-  fiat_p384_addcarryx_u32(&x34, &x35, x33, (x25 & UINT32_C(0xfffffffe)), x9);
+  fiat_p384_addcarryx_u32(&x34, &x35, x33, x9, (x25 & UINT32_C(0xfffffffe)));
   uint32_t x36;
   fiat_p384_uint1 x37;
-  fiat_p384_addcarryx_u32(&x36, &x37, x35, (x25 & UINT32_C(0xffffffff)), x11);
+  fiat_p384_addcarryx_u32(&x36, &x37, x35, x11, (x25 & UINT32_C(0xffffffff)));
   uint32_t x38;
   fiat_p384_uint1 x39;
-  fiat_p384_addcarryx_u32(&x38, &x39, x37, (x25 & UINT32_C(0xffffffff)), x13);
+  fiat_p384_addcarryx_u32(&x38, &x39, x37, x13, (x25 & UINT32_C(0xffffffff)));
   uint32_t x40;
   fiat_p384_uint1 x41;
-  fiat_p384_addcarryx_u32(&x40, &x41, x39, (x25 & UINT32_C(0xffffffff)), x15);
+  fiat_p384_addcarryx_u32(&x40, &x41, x39, x15, (x25 & UINT32_C(0xffffffff)));
   uint32_t x42;
   fiat_p384_uint1 x43;
-  fiat_p384_addcarryx_u32(&x42, &x43, x41, (x25 & UINT32_C(0xffffffff)), x17);
+  fiat_p384_addcarryx_u32(&x42, &x43, x41, x17, (x25 & UINT32_C(0xffffffff)));
   uint32_t x44;
   fiat_p384_uint1 x45;
-  fiat_p384_addcarryx_u32(&x44, &x45, x43, (x25 & UINT32_C(0xffffffff)), x19);
+  fiat_p384_addcarryx_u32(&x44, &x45, x43, x19, (x25 & UINT32_C(0xffffffff)));
   uint32_t x46;
   fiat_p384_uint1 x47;
-  fiat_p384_addcarryx_u32(&x46, &x47, x45, (x25 & UINT32_C(0xffffffff)), x21);
+  fiat_p384_addcarryx_u32(&x46, &x47, x45, x21, (x25 & UINT32_C(0xffffffff)));
   uint32_t x48;
   fiat_p384_uint1 x49;
-  fiat_p384_addcarryx_u32(&x48, &x49, x47, (x25 & UINT32_C(0xffffffff)), x23);
+  fiat_p384_addcarryx_u32(&x48, &x49, x47, x23, (x25 & UINT32_C(0xffffffff)));
   out1[0] = x26;
   out1[1] = x28;
   out1[2] = x30;
@@ -5652,37 +5652,37 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_mulx_u32(&x20, &x21, x1, UINT32_C(0xffffffff));
   uint32_t x22;
   fiat_p384_uint1 x23;
-  fiat_p384_addcarryx_u32(&x22, &x23, 0x0, x16, x19);
+  fiat_p384_addcarryx_u32(&x22, &x23, 0x0, x19, x16);
   uint32_t x24;
   fiat_p384_uint1 x25;
-  fiat_p384_addcarryx_u32(&x24, &x25, x23, x14, x17);
+  fiat_p384_addcarryx_u32(&x24, &x25, x23, x17, x14);
   uint32_t x26;
   fiat_p384_uint1 x27;
-  fiat_p384_addcarryx_u32(&x26, &x27, x25, x12, x15);
+  fiat_p384_addcarryx_u32(&x26, &x27, x25, x15, x12);
   uint32_t x28;
   fiat_p384_uint1 x29;
-  fiat_p384_addcarryx_u32(&x28, &x29, x27, x10, x13);
+  fiat_p384_addcarryx_u32(&x28, &x29, x27, x13, x10);
   uint32_t x30;
   fiat_p384_uint1 x31;
-  fiat_p384_addcarryx_u32(&x30, &x31, x29, x8, x11);
+  fiat_p384_addcarryx_u32(&x30, &x31, x29, x11, x8);
   uint32_t x32;
   fiat_p384_uint1 x33;
-  fiat_p384_addcarryx_u32(&x32, &x33, x31, x6, x9);
+  fiat_p384_addcarryx_u32(&x32, &x33, x31, x9, x6);
   uint32_t x34;
   fiat_p384_uint1 x35;
-  fiat_p384_addcarryx_u32(&x34, &x35, x33, x4, x7);
+  fiat_p384_addcarryx_u32(&x34, &x35, x33, x7, x4);
   uint32_t x36;
   fiat_p384_uint1 x37;
-  fiat_p384_addcarryx_u32(&x36, &x37, x35, x2, x5);
+  fiat_p384_addcarryx_u32(&x36, &x37, x35, x5, x2);
   uint32_t x38;
   fiat_p384_uint1 x39;
-  fiat_p384_addcarryx_u32(&x38, &x39, 0x0, x20, x1);
+  fiat_p384_addcarryx_u32(&x38, &x39, 0x0, x1, x20);
   uint32_t x40;
   fiat_p384_uint1 x41;
-  fiat_p384_addcarryx_u32(&x40, &x41, x39, x21, 0x0);
+  fiat_p384_addcarryx_u32(&x40, &x41, x39, 0x0, x21);
   uint32_t x42;
   fiat_p384_uint1 x43;
-  fiat_p384_addcarryx_u32(&x42, &x43, 0x0, (arg1[1]), x40);
+  fiat_p384_addcarryx_u32(&x42, &x43, 0x0, x40, (arg1[1]));
   uint32_t x44;
   uint32_t x45;
   fiat_p384_mulx_u32(&x44, &x45, x42, UINT32_C(0xffffffff));
@@ -5715,112 +5715,112 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_mulx_u32(&x62, &x63, x42, UINT32_C(0xffffffff));
   uint32_t x64;
   fiat_p384_uint1 x65;
-  fiat_p384_addcarryx_u32(&x64, &x65, 0x0, x58, x61);
+  fiat_p384_addcarryx_u32(&x64, &x65, 0x0, x61, x58);
   uint32_t x66;
   fiat_p384_uint1 x67;
-  fiat_p384_addcarryx_u32(&x66, &x67, x65, x56, x59);
+  fiat_p384_addcarryx_u32(&x66, &x67, x65, x59, x56);
   uint32_t x68;
   fiat_p384_uint1 x69;
-  fiat_p384_addcarryx_u32(&x68, &x69, x67, x54, x57);
+  fiat_p384_addcarryx_u32(&x68, &x69, x67, x57, x54);
   uint32_t x70;
   fiat_p384_uint1 x71;
-  fiat_p384_addcarryx_u32(&x70, &x71, x69, x52, x55);
+  fiat_p384_addcarryx_u32(&x70, &x71, x69, x55, x52);
   uint32_t x72;
   fiat_p384_uint1 x73;
-  fiat_p384_addcarryx_u32(&x72, &x73, x71, x50, x53);
+  fiat_p384_addcarryx_u32(&x72, &x73, x71, x53, x50);
   uint32_t x74;
   fiat_p384_uint1 x75;
-  fiat_p384_addcarryx_u32(&x74, &x75, x73, x48, x51);
+  fiat_p384_addcarryx_u32(&x74, &x75, x73, x51, x48);
   uint32_t x76;
   fiat_p384_uint1 x77;
-  fiat_p384_addcarryx_u32(&x76, &x77, x75, x46, x49);
+  fiat_p384_addcarryx_u32(&x76, &x77, x75, x49, x46);
   uint32_t x78;
   fiat_p384_uint1 x79;
-  fiat_p384_addcarryx_u32(&x78, &x79, x77, x44, x47);
+  fiat_p384_addcarryx_u32(&x78, &x79, x77, x47, x44);
   uint32_t x80;
   fiat_p384_uint1 x81;
   fiat_p384_addcarryx_u32(&x80, &x81, x43, 0x0, 0x0);
   uint32_t x82;
   fiat_p384_uint1 x83;
-  fiat_p384_addcarryx_u32(&x82, &x83, 0x0, x62, x42);
+  fiat_p384_addcarryx_u32(&x82, &x83, 0x0, x42, x62);
   uint32_t x84;
   fiat_p384_uint1 x85;
-  fiat_p384_addcarryx_u32(&x84, &x85, x83, x63, (fiat_p384_uint1)x80);
+  fiat_p384_addcarryx_u32(&x84, &x85, x83, (fiat_p384_uint1)x80, x63);
   uint32_t x86;
   fiat_p384_uint1 x87;
-  fiat_p384_addcarryx_u32(&x86, &x87, x85, 0x0, x18);
+  fiat_p384_addcarryx_u32(&x86, &x87, x85, x18, 0x0);
   uint32_t x88;
   fiat_p384_uint1 x89;
-  fiat_p384_addcarryx_u32(&x88, &x89, x87, x60, x22);
+  fiat_p384_addcarryx_u32(&x88, &x89, x87, x22, x60);
   uint32_t x90;
   fiat_p384_uint1 x91;
-  fiat_p384_addcarryx_u32(&x90, &x91, x89, x64, x24);
+  fiat_p384_addcarryx_u32(&x90, &x91, x89, x24, x64);
   uint32_t x92;
   fiat_p384_uint1 x93;
-  fiat_p384_addcarryx_u32(&x92, &x93, x91, x66, x26);
+  fiat_p384_addcarryx_u32(&x92, &x93, x91, x26, x66);
   uint32_t x94;
   fiat_p384_uint1 x95;
-  fiat_p384_addcarryx_u32(&x94, &x95, x93, x68, x28);
+  fiat_p384_addcarryx_u32(&x94, &x95, x93, x28, x68);
   uint32_t x96;
   fiat_p384_uint1 x97;
-  fiat_p384_addcarryx_u32(&x96, &x97, x95, x70, x30);
+  fiat_p384_addcarryx_u32(&x96, &x97, x95, x30, x70);
   uint32_t x98;
   fiat_p384_uint1 x99;
-  fiat_p384_addcarryx_u32(&x98, &x99, x97, x72, x32);
+  fiat_p384_addcarryx_u32(&x98, &x99, x97, x32, x72);
   uint32_t x100;
   fiat_p384_uint1 x101;
-  fiat_p384_addcarryx_u32(&x100, &x101, x99, x74, x34);
+  fiat_p384_addcarryx_u32(&x100, &x101, x99, x34, x74);
   uint32_t x102;
   fiat_p384_uint1 x103;
-  fiat_p384_addcarryx_u32(&x102, &x103, x101, x76, x36);
+  fiat_p384_addcarryx_u32(&x102, &x103, x101, x36, x76);
   uint32_t x104;
   fiat_p384_uint1 x105;
-  fiat_p384_addcarryx_u32(&x104, &x105, x37, 0x0, x3);
+  fiat_p384_addcarryx_u32(&x104, &x105, x37, x3, 0x0);
   uint32_t x106;
   fiat_p384_uint1 x107;
-  fiat_p384_addcarryx_u32(&x106, &x107, x103, x78, x104);
+  fiat_p384_addcarryx_u32(&x106, &x107, x103, x104, x78);
   uint32_t x108;
   fiat_p384_uint1 x109;
-  fiat_p384_addcarryx_u32(&x108, &x109, x79, 0x0, x45);
+  fiat_p384_addcarryx_u32(&x108, &x109, x79, x45, 0x0);
   uint32_t x110;
   fiat_p384_uint1 x111;
-  fiat_p384_addcarryx_u32(&x110, &x111, x107, x108, 0x0);
+  fiat_p384_addcarryx_u32(&x110, &x111, x107, 0x0, x108);
   uint32_t x112;
   fiat_p384_uint1 x113;
-  fiat_p384_addcarryx_u32(&x112, &x113, 0x0, (arg1[2]), x84);
+  fiat_p384_addcarryx_u32(&x112, &x113, 0x0, x84, (arg1[2]));
   uint32_t x114;
   fiat_p384_uint1 x115;
-  fiat_p384_addcarryx_u32(&x114, &x115, x113, 0x0, x86);
+  fiat_p384_addcarryx_u32(&x114, &x115, x113, x86, 0x0);
   uint32_t x116;
   fiat_p384_uint1 x117;
-  fiat_p384_addcarryx_u32(&x116, &x117, x115, 0x0, x88);
+  fiat_p384_addcarryx_u32(&x116, &x117, x115, x88, 0x0);
   uint32_t x118;
   fiat_p384_uint1 x119;
-  fiat_p384_addcarryx_u32(&x118, &x119, x117, 0x0, x90);
+  fiat_p384_addcarryx_u32(&x118, &x119, x117, x90, 0x0);
   uint32_t x120;
   fiat_p384_uint1 x121;
-  fiat_p384_addcarryx_u32(&x120, &x121, x119, 0x0, x92);
+  fiat_p384_addcarryx_u32(&x120, &x121, x119, x92, 0x0);
   uint32_t x122;
   fiat_p384_uint1 x123;
-  fiat_p384_addcarryx_u32(&x122, &x123, x121, 0x0, x94);
+  fiat_p384_addcarryx_u32(&x122, &x123, x121, x94, 0x0);
   uint32_t x124;
   fiat_p384_uint1 x125;
-  fiat_p384_addcarryx_u32(&x124, &x125, x123, 0x0, x96);
+  fiat_p384_addcarryx_u32(&x124, &x125, x123, x96, 0x0);
   uint32_t x126;
   fiat_p384_uint1 x127;
-  fiat_p384_addcarryx_u32(&x126, &x127, x125, 0x0, x98);
+  fiat_p384_addcarryx_u32(&x126, &x127, x125, x98, 0x0);
   uint32_t x128;
   fiat_p384_uint1 x129;
-  fiat_p384_addcarryx_u32(&x128, &x129, x127, 0x0, x100);
+  fiat_p384_addcarryx_u32(&x128, &x129, x127, x100, 0x0);
   uint32_t x130;
   fiat_p384_uint1 x131;
-  fiat_p384_addcarryx_u32(&x130, &x131, x129, 0x0, x102);
+  fiat_p384_addcarryx_u32(&x130, &x131, x129, x102, 0x0);
   uint32_t x132;
   fiat_p384_uint1 x133;
-  fiat_p384_addcarryx_u32(&x132, &x133, x131, 0x0, x106);
+  fiat_p384_addcarryx_u32(&x132, &x133, x131, x106, 0x0);
   uint32_t x134;
   fiat_p384_uint1 x135;
-  fiat_p384_addcarryx_u32(&x134, &x135, x133, 0x0, x110);
+  fiat_p384_addcarryx_u32(&x134, &x135, x133, x110, 0x0);
   uint32_t x136;
   uint32_t x137;
   fiat_p384_mulx_u32(&x136, &x137, x112, UINT32_C(0xffffffff));
@@ -5853,112 +5853,112 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_mulx_u32(&x154, &x155, x112, UINT32_C(0xffffffff));
   uint32_t x156;
   fiat_p384_uint1 x157;
-  fiat_p384_addcarryx_u32(&x156, &x157, 0x0, x150, x153);
+  fiat_p384_addcarryx_u32(&x156, &x157, 0x0, x153, x150);
   uint32_t x158;
   fiat_p384_uint1 x159;
-  fiat_p384_addcarryx_u32(&x158, &x159, x157, x148, x151);
+  fiat_p384_addcarryx_u32(&x158, &x159, x157, x151, x148);
   uint32_t x160;
   fiat_p384_uint1 x161;
-  fiat_p384_addcarryx_u32(&x160, &x161, x159, x146, x149);
+  fiat_p384_addcarryx_u32(&x160, &x161, x159, x149, x146);
   uint32_t x162;
   fiat_p384_uint1 x163;
-  fiat_p384_addcarryx_u32(&x162, &x163, x161, x144, x147);
+  fiat_p384_addcarryx_u32(&x162, &x163, x161, x147, x144);
   uint32_t x164;
   fiat_p384_uint1 x165;
-  fiat_p384_addcarryx_u32(&x164, &x165, x163, x142, x145);
+  fiat_p384_addcarryx_u32(&x164, &x165, x163, x145, x142);
   uint32_t x166;
   fiat_p384_uint1 x167;
-  fiat_p384_addcarryx_u32(&x166, &x167, x165, x140, x143);
+  fiat_p384_addcarryx_u32(&x166, &x167, x165, x143, x140);
   uint32_t x168;
   fiat_p384_uint1 x169;
-  fiat_p384_addcarryx_u32(&x168, &x169, x167, x138, x141);
+  fiat_p384_addcarryx_u32(&x168, &x169, x167, x141, x138);
   uint32_t x170;
   fiat_p384_uint1 x171;
-  fiat_p384_addcarryx_u32(&x170, &x171, x169, x136, x139);
+  fiat_p384_addcarryx_u32(&x170, &x171, x169, x139, x136);
   uint32_t x172;
   fiat_p384_uint1 x173;
-  fiat_p384_addcarryx_u32(&x172, &x173, 0x0, x154, x112);
+  fiat_p384_addcarryx_u32(&x172, &x173, 0x0, x112, x154);
   uint32_t x174;
   fiat_p384_uint1 x175;
-  fiat_p384_addcarryx_u32(&x174, &x175, x173, x155, x114);
+  fiat_p384_addcarryx_u32(&x174, &x175, x173, x114, x155);
   uint32_t x176;
   fiat_p384_uint1 x177;
-  fiat_p384_addcarryx_u32(&x176, &x177, x175, 0x0, x116);
+  fiat_p384_addcarryx_u32(&x176, &x177, x175, x116, 0x0);
   uint32_t x178;
   fiat_p384_uint1 x179;
-  fiat_p384_addcarryx_u32(&x178, &x179, x177, x152, x118);
+  fiat_p384_addcarryx_u32(&x178, &x179, x177, x118, x152);
   uint32_t x180;
   fiat_p384_uint1 x181;
-  fiat_p384_addcarryx_u32(&x180, &x181, x179, x156, x120);
+  fiat_p384_addcarryx_u32(&x180, &x181, x179, x120, x156);
   uint32_t x182;
   fiat_p384_uint1 x183;
-  fiat_p384_addcarryx_u32(&x182, &x183, x181, x158, x122);
+  fiat_p384_addcarryx_u32(&x182, &x183, x181, x122, x158);
   uint32_t x184;
   fiat_p384_uint1 x185;
-  fiat_p384_addcarryx_u32(&x184, &x185, x183, x160, x124);
+  fiat_p384_addcarryx_u32(&x184, &x185, x183, x124, x160);
   uint32_t x186;
   fiat_p384_uint1 x187;
-  fiat_p384_addcarryx_u32(&x186, &x187, x185, x162, x126);
+  fiat_p384_addcarryx_u32(&x186, &x187, x185, x126, x162);
   uint32_t x188;
   fiat_p384_uint1 x189;
-  fiat_p384_addcarryx_u32(&x188, &x189, x187, x164, x128);
+  fiat_p384_addcarryx_u32(&x188, &x189, x187, x128, x164);
   uint32_t x190;
   fiat_p384_uint1 x191;
-  fiat_p384_addcarryx_u32(&x190, &x191, x189, x166, x130);
+  fiat_p384_addcarryx_u32(&x190, &x191, x189, x130, x166);
   uint32_t x192;
   fiat_p384_uint1 x193;
-  fiat_p384_addcarryx_u32(&x192, &x193, x191, x168, x132);
+  fiat_p384_addcarryx_u32(&x192, &x193, x191, x132, x168);
   uint32_t x194;
   fiat_p384_uint1 x195;
-  fiat_p384_addcarryx_u32(&x194, &x195, x193, x170, x134);
+  fiat_p384_addcarryx_u32(&x194, &x195, x193, x134, x170);
   uint32_t x196;
   fiat_p384_uint1 x197;
-  fiat_p384_addcarryx_u32(&x196, &x197, x111, 0x0, 0x0);
+  fiat_p384_addcarryx_u32(&x196, &x197, x171, x137, 0x0);
   uint32_t x198;
   fiat_p384_uint1 x199;
-  fiat_p384_addcarryx_u32(&x198, &x199, x135, 0x0, (fiat_p384_uint1)x196);
+  fiat_p384_addcarryx_u32(&x198, &x199, x111, 0x0, 0x0);
   uint32_t x200;
   fiat_p384_uint1 x201;
-  fiat_p384_addcarryx_u32(&x200, &x201, x171, 0x0, x137);
+  fiat_p384_addcarryx_u32(&x200, &x201, x135, (fiat_p384_uint1)x198, 0x0);
   uint32_t x202;
   fiat_p384_uint1 x203;
-  fiat_p384_addcarryx_u32(&x202, &x203, x195, x200, x198);
+  fiat_p384_addcarryx_u32(&x202, &x203, x195, x200, x196);
   uint32_t x204;
   fiat_p384_uint1 x205;
-  fiat_p384_addcarryx_u32(&x204, &x205, 0x0, (arg1[3]), x174);
+  fiat_p384_addcarryx_u32(&x204, &x205, 0x0, x174, (arg1[3]));
   uint32_t x206;
   fiat_p384_uint1 x207;
-  fiat_p384_addcarryx_u32(&x206, &x207, x205, 0x0, x176);
+  fiat_p384_addcarryx_u32(&x206, &x207, x205, x176, 0x0);
   uint32_t x208;
   fiat_p384_uint1 x209;
-  fiat_p384_addcarryx_u32(&x208, &x209, x207, 0x0, x178);
+  fiat_p384_addcarryx_u32(&x208, &x209, x207, x178, 0x0);
   uint32_t x210;
   fiat_p384_uint1 x211;
-  fiat_p384_addcarryx_u32(&x210, &x211, x209, 0x0, x180);
+  fiat_p384_addcarryx_u32(&x210, &x211, x209, x180, 0x0);
   uint32_t x212;
   fiat_p384_uint1 x213;
-  fiat_p384_addcarryx_u32(&x212, &x213, x211, 0x0, x182);
+  fiat_p384_addcarryx_u32(&x212, &x213, x211, x182, 0x0);
   uint32_t x214;
   fiat_p384_uint1 x215;
-  fiat_p384_addcarryx_u32(&x214, &x215, x213, 0x0, x184);
+  fiat_p384_addcarryx_u32(&x214, &x215, x213, x184, 0x0);
   uint32_t x216;
   fiat_p384_uint1 x217;
-  fiat_p384_addcarryx_u32(&x216, &x217, x215, 0x0, x186);
+  fiat_p384_addcarryx_u32(&x216, &x217, x215, x186, 0x0);
   uint32_t x218;
   fiat_p384_uint1 x219;
-  fiat_p384_addcarryx_u32(&x218, &x219, x217, 0x0, x188);
+  fiat_p384_addcarryx_u32(&x218, &x219, x217, x188, 0x0);
   uint32_t x220;
   fiat_p384_uint1 x221;
-  fiat_p384_addcarryx_u32(&x220, &x221, x219, 0x0, x190);
+  fiat_p384_addcarryx_u32(&x220, &x221, x219, x190, 0x0);
   uint32_t x222;
   fiat_p384_uint1 x223;
-  fiat_p384_addcarryx_u32(&x222, &x223, x221, 0x0, x192);
+  fiat_p384_addcarryx_u32(&x222, &x223, x221, x192, 0x0);
   uint32_t x224;
   fiat_p384_uint1 x225;
-  fiat_p384_addcarryx_u32(&x224, &x225, x223, 0x0, x194);
+  fiat_p384_addcarryx_u32(&x224, &x225, x223, x194, 0x0);
   uint32_t x226;
   fiat_p384_uint1 x227;
-  fiat_p384_addcarryx_u32(&x226, &x227, x225, 0x0, x202);
+  fiat_p384_addcarryx_u32(&x226, &x227, x225, x202, 0x0);
   uint32_t x228;
   uint32_t x229;
   fiat_p384_mulx_u32(&x228, &x229, x204, UINT32_C(0xffffffff));
@@ -5991,112 +5991,112 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_mulx_u32(&x246, &x247, x204, UINT32_C(0xffffffff));
   uint32_t x248;
   fiat_p384_uint1 x249;
-  fiat_p384_addcarryx_u32(&x248, &x249, 0x0, x242, x245);
+  fiat_p384_addcarryx_u32(&x248, &x249, 0x0, x245, x242);
   uint32_t x250;
   fiat_p384_uint1 x251;
-  fiat_p384_addcarryx_u32(&x250, &x251, x249, x240, x243);
+  fiat_p384_addcarryx_u32(&x250, &x251, x249, x243, x240);
   uint32_t x252;
   fiat_p384_uint1 x253;
-  fiat_p384_addcarryx_u32(&x252, &x253, x251, x238, x241);
+  fiat_p384_addcarryx_u32(&x252, &x253, x251, x241, x238);
   uint32_t x254;
   fiat_p384_uint1 x255;
-  fiat_p384_addcarryx_u32(&x254, &x255, x253, x236, x239);
+  fiat_p384_addcarryx_u32(&x254, &x255, x253, x239, x236);
   uint32_t x256;
   fiat_p384_uint1 x257;
-  fiat_p384_addcarryx_u32(&x256, &x257, x255, x234, x237);
+  fiat_p384_addcarryx_u32(&x256, &x257, x255, x237, x234);
   uint32_t x258;
   fiat_p384_uint1 x259;
-  fiat_p384_addcarryx_u32(&x258, &x259, x257, x232, x235);
+  fiat_p384_addcarryx_u32(&x258, &x259, x257, x235, x232);
   uint32_t x260;
   fiat_p384_uint1 x261;
-  fiat_p384_addcarryx_u32(&x260, &x261, x259, x230, x233);
+  fiat_p384_addcarryx_u32(&x260, &x261, x259, x233, x230);
   uint32_t x262;
   fiat_p384_uint1 x263;
-  fiat_p384_addcarryx_u32(&x262, &x263, x261, x228, x231);
+  fiat_p384_addcarryx_u32(&x262, &x263, x261, x231, x228);
   uint32_t x264;
   fiat_p384_uint1 x265;
-  fiat_p384_addcarryx_u32(&x264, &x265, 0x0, x246, x204);
+  fiat_p384_addcarryx_u32(&x264, &x265, 0x0, x204, x246);
   uint32_t x266;
   fiat_p384_uint1 x267;
-  fiat_p384_addcarryx_u32(&x266, &x267, x265, x247, x206);
+  fiat_p384_addcarryx_u32(&x266, &x267, x265, x206, x247);
   uint32_t x268;
   fiat_p384_uint1 x269;
-  fiat_p384_addcarryx_u32(&x268, &x269, x267, 0x0, x208);
+  fiat_p384_addcarryx_u32(&x268, &x269, x267, x208, 0x0);
   uint32_t x270;
   fiat_p384_uint1 x271;
-  fiat_p384_addcarryx_u32(&x270, &x271, x269, x244, x210);
+  fiat_p384_addcarryx_u32(&x270, &x271, x269, x210, x244);
   uint32_t x272;
   fiat_p384_uint1 x273;
-  fiat_p384_addcarryx_u32(&x272, &x273, x271, x248, x212);
+  fiat_p384_addcarryx_u32(&x272, &x273, x271, x212, x248);
   uint32_t x274;
   fiat_p384_uint1 x275;
-  fiat_p384_addcarryx_u32(&x274, &x275, x273, x250, x214);
+  fiat_p384_addcarryx_u32(&x274, &x275, x273, x214, x250);
   uint32_t x276;
   fiat_p384_uint1 x277;
-  fiat_p384_addcarryx_u32(&x276, &x277, x275, x252, x216);
+  fiat_p384_addcarryx_u32(&x276, &x277, x275, x216, x252);
   uint32_t x278;
   fiat_p384_uint1 x279;
-  fiat_p384_addcarryx_u32(&x278, &x279, x277, x254, x218);
+  fiat_p384_addcarryx_u32(&x278, &x279, x277, x218, x254);
   uint32_t x280;
   fiat_p384_uint1 x281;
-  fiat_p384_addcarryx_u32(&x280, &x281, x279, x256, x220);
+  fiat_p384_addcarryx_u32(&x280, &x281, x279, x220, x256);
   uint32_t x282;
   fiat_p384_uint1 x283;
-  fiat_p384_addcarryx_u32(&x282, &x283, x281, x258, x222);
+  fiat_p384_addcarryx_u32(&x282, &x283, x281, x222, x258);
   uint32_t x284;
   fiat_p384_uint1 x285;
-  fiat_p384_addcarryx_u32(&x284, &x285, x283, x260, x224);
+  fiat_p384_addcarryx_u32(&x284, &x285, x283, x224, x260);
   uint32_t x286;
   fiat_p384_uint1 x287;
-  fiat_p384_addcarryx_u32(&x286, &x287, x285, x262, x226);
+  fiat_p384_addcarryx_u32(&x286, &x287, x285, x226, x262);
   uint32_t x288;
   fiat_p384_uint1 x289;
-  fiat_p384_addcarryx_u32(&x288, &x289, x203, 0x0, 0x0);
+  fiat_p384_addcarryx_u32(&x288, &x289, x263, x229, 0x0);
   uint32_t x290;
   fiat_p384_uint1 x291;
-  fiat_p384_addcarryx_u32(&x290, &x291, x227, 0x0, (fiat_p384_uint1)x288);
+  fiat_p384_addcarryx_u32(&x290, &x291, x203, 0x0, 0x0);
   uint32_t x292;
   fiat_p384_uint1 x293;
-  fiat_p384_addcarryx_u32(&x292, &x293, x263, 0x0, x229);
+  fiat_p384_addcarryx_u32(&x292, &x293, x227, (fiat_p384_uint1)x290, 0x0);
   uint32_t x294;
   fiat_p384_uint1 x295;
-  fiat_p384_addcarryx_u32(&x294, &x295, x287, x292, x290);
+  fiat_p384_addcarryx_u32(&x294, &x295, x287, x292, x288);
   uint32_t x296;
   fiat_p384_uint1 x297;
-  fiat_p384_addcarryx_u32(&x296, &x297, 0x0, (arg1[4]), x266);
+  fiat_p384_addcarryx_u32(&x296, &x297, 0x0, x266, (arg1[4]));
   uint32_t x298;
   fiat_p384_uint1 x299;
-  fiat_p384_addcarryx_u32(&x298, &x299, x297, 0x0, x268);
+  fiat_p384_addcarryx_u32(&x298, &x299, x297, x268, 0x0);
   uint32_t x300;
   fiat_p384_uint1 x301;
-  fiat_p384_addcarryx_u32(&x300, &x301, x299, 0x0, x270);
+  fiat_p384_addcarryx_u32(&x300, &x301, x299, x270, 0x0);
   uint32_t x302;
   fiat_p384_uint1 x303;
-  fiat_p384_addcarryx_u32(&x302, &x303, x301, 0x0, x272);
+  fiat_p384_addcarryx_u32(&x302, &x303, x301, x272, 0x0);
   uint32_t x304;
   fiat_p384_uint1 x305;
-  fiat_p384_addcarryx_u32(&x304, &x305, x303, 0x0, x274);
+  fiat_p384_addcarryx_u32(&x304, &x305, x303, x274, 0x0);
   uint32_t x306;
   fiat_p384_uint1 x307;
-  fiat_p384_addcarryx_u32(&x306, &x307, x305, 0x0, x276);
+  fiat_p384_addcarryx_u32(&x306, &x307, x305, x276, 0x0);
   uint32_t x308;
   fiat_p384_uint1 x309;
-  fiat_p384_addcarryx_u32(&x308, &x309, x307, 0x0, x278);
+  fiat_p384_addcarryx_u32(&x308, &x309, x307, x278, 0x0);
   uint32_t x310;
   fiat_p384_uint1 x311;
-  fiat_p384_addcarryx_u32(&x310, &x311, x309, 0x0, x280);
+  fiat_p384_addcarryx_u32(&x310, &x311, x309, x280, 0x0);
   uint32_t x312;
   fiat_p384_uint1 x313;
-  fiat_p384_addcarryx_u32(&x312, &x313, x311, 0x0, x282);
+  fiat_p384_addcarryx_u32(&x312, &x313, x311, x282, 0x0);
   uint32_t x314;
   fiat_p384_uint1 x315;
-  fiat_p384_addcarryx_u32(&x314, &x315, x313, 0x0, x284);
+  fiat_p384_addcarryx_u32(&x314, &x315, x313, x284, 0x0);
   uint32_t x316;
   fiat_p384_uint1 x317;
-  fiat_p384_addcarryx_u32(&x316, &x317, x315, 0x0, x286);
+  fiat_p384_addcarryx_u32(&x316, &x317, x315, x286, 0x0);
   uint32_t x318;
   fiat_p384_uint1 x319;
-  fiat_p384_addcarryx_u32(&x318, &x319, x317, 0x0, x294);
+  fiat_p384_addcarryx_u32(&x318, &x319, x317, x294, 0x0);
   uint32_t x320;
   uint32_t x321;
   fiat_p384_mulx_u32(&x320, &x321, x296, UINT32_C(0xffffffff));
@@ -6129,112 +6129,112 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_mulx_u32(&x338, &x339, x296, UINT32_C(0xffffffff));
   uint32_t x340;
   fiat_p384_uint1 x341;
-  fiat_p384_addcarryx_u32(&x340, &x341, 0x0, x334, x337);
+  fiat_p384_addcarryx_u32(&x340, &x341, 0x0, x337, x334);
   uint32_t x342;
   fiat_p384_uint1 x343;
-  fiat_p384_addcarryx_u32(&x342, &x343, x341, x332, x335);
+  fiat_p384_addcarryx_u32(&x342, &x343, x341, x335, x332);
   uint32_t x344;
   fiat_p384_uint1 x345;
-  fiat_p384_addcarryx_u32(&x344, &x345, x343, x330, x333);
+  fiat_p384_addcarryx_u32(&x344, &x345, x343, x333, x330);
   uint32_t x346;
   fiat_p384_uint1 x347;
-  fiat_p384_addcarryx_u32(&x346, &x347, x345, x328, x331);
+  fiat_p384_addcarryx_u32(&x346, &x347, x345, x331, x328);
   uint32_t x348;
   fiat_p384_uint1 x349;
-  fiat_p384_addcarryx_u32(&x348, &x349, x347, x326, x329);
+  fiat_p384_addcarryx_u32(&x348, &x349, x347, x329, x326);
   uint32_t x350;
   fiat_p384_uint1 x351;
-  fiat_p384_addcarryx_u32(&x350, &x351, x349, x324, x327);
+  fiat_p384_addcarryx_u32(&x350, &x351, x349, x327, x324);
   uint32_t x352;
   fiat_p384_uint1 x353;
-  fiat_p384_addcarryx_u32(&x352, &x353, x351, x322, x325);
+  fiat_p384_addcarryx_u32(&x352, &x353, x351, x325, x322);
   uint32_t x354;
   fiat_p384_uint1 x355;
-  fiat_p384_addcarryx_u32(&x354, &x355, x353, x320, x323);
+  fiat_p384_addcarryx_u32(&x354, &x355, x353, x323, x320);
   uint32_t x356;
   fiat_p384_uint1 x357;
-  fiat_p384_addcarryx_u32(&x356, &x357, 0x0, x338, x296);
+  fiat_p384_addcarryx_u32(&x356, &x357, 0x0, x296, x338);
   uint32_t x358;
   fiat_p384_uint1 x359;
-  fiat_p384_addcarryx_u32(&x358, &x359, x357, x339, x298);
+  fiat_p384_addcarryx_u32(&x358, &x359, x357, x298, x339);
   uint32_t x360;
   fiat_p384_uint1 x361;
-  fiat_p384_addcarryx_u32(&x360, &x361, x359, 0x0, x300);
+  fiat_p384_addcarryx_u32(&x360, &x361, x359, x300, 0x0);
   uint32_t x362;
   fiat_p384_uint1 x363;
-  fiat_p384_addcarryx_u32(&x362, &x363, x361, x336, x302);
+  fiat_p384_addcarryx_u32(&x362, &x363, x361, x302, x336);
   uint32_t x364;
   fiat_p384_uint1 x365;
-  fiat_p384_addcarryx_u32(&x364, &x365, x363, x340, x304);
+  fiat_p384_addcarryx_u32(&x364, &x365, x363, x304, x340);
   uint32_t x366;
   fiat_p384_uint1 x367;
-  fiat_p384_addcarryx_u32(&x366, &x367, x365, x342, x306);
+  fiat_p384_addcarryx_u32(&x366, &x367, x365, x306, x342);
   uint32_t x368;
   fiat_p384_uint1 x369;
-  fiat_p384_addcarryx_u32(&x368, &x369, x367, x344, x308);
+  fiat_p384_addcarryx_u32(&x368, &x369, x367, x308, x344);
   uint32_t x370;
   fiat_p384_uint1 x371;
-  fiat_p384_addcarryx_u32(&x370, &x371, x369, x346, x310);
+  fiat_p384_addcarryx_u32(&x370, &x371, x369, x310, x346);
   uint32_t x372;
   fiat_p384_uint1 x373;
-  fiat_p384_addcarryx_u32(&x372, &x373, x371, x348, x312);
+  fiat_p384_addcarryx_u32(&x372, &x373, x371, x312, x348);
   uint32_t x374;
   fiat_p384_uint1 x375;
-  fiat_p384_addcarryx_u32(&x374, &x375, x373, x350, x314);
+  fiat_p384_addcarryx_u32(&x374, &x375, x373, x314, x350);
   uint32_t x376;
   fiat_p384_uint1 x377;
-  fiat_p384_addcarryx_u32(&x376, &x377, x375, x352, x316);
+  fiat_p384_addcarryx_u32(&x376, &x377, x375, x316, x352);
   uint32_t x378;
   fiat_p384_uint1 x379;
-  fiat_p384_addcarryx_u32(&x378, &x379, x377, x354, x318);
+  fiat_p384_addcarryx_u32(&x378, &x379, x377, x318, x354);
   uint32_t x380;
   fiat_p384_uint1 x381;
-  fiat_p384_addcarryx_u32(&x380, &x381, x295, 0x0, 0x0);
+  fiat_p384_addcarryx_u32(&x380, &x381, x355, x321, 0x0);
   uint32_t x382;
   fiat_p384_uint1 x383;
-  fiat_p384_addcarryx_u32(&x382, &x383, x319, 0x0, (fiat_p384_uint1)x380);
+  fiat_p384_addcarryx_u32(&x382, &x383, x295, 0x0, 0x0);
   uint32_t x384;
   fiat_p384_uint1 x385;
-  fiat_p384_addcarryx_u32(&x384, &x385, x355, 0x0, x321);
+  fiat_p384_addcarryx_u32(&x384, &x385, x319, (fiat_p384_uint1)x382, 0x0);
   uint32_t x386;
   fiat_p384_uint1 x387;
-  fiat_p384_addcarryx_u32(&x386, &x387, x379, x384, x382);
+  fiat_p384_addcarryx_u32(&x386, &x387, x379, x384, x380);
   uint32_t x388;
   fiat_p384_uint1 x389;
-  fiat_p384_addcarryx_u32(&x388, &x389, 0x0, (arg1[5]), x358);
+  fiat_p384_addcarryx_u32(&x388, &x389, 0x0, x358, (arg1[5]));
   uint32_t x390;
   fiat_p384_uint1 x391;
-  fiat_p384_addcarryx_u32(&x390, &x391, x389, 0x0, x360);
+  fiat_p384_addcarryx_u32(&x390, &x391, x389, x360, 0x0);
   uint32_t x392;
   fiat_p384_uint1 x393;
-  fiat_p384_addcarryx_u32(&x392, &x393, x391, 0x0, x362);
+  fiat_p384_addcarryx_u32(&x392, &x393, x391, x362, 0x0);
   uint32_t x394;
   fiat_p384_uint1 x395;
-  fiat_p384_addcarryx_u32(&x394, &x395, x393, 0x0, x364);
+  fiat_p384_addcarryx_u32(&x394, &x395, x393, x364, 0x0);
   uint32_t x396;
   fiat_p384_uint1 x397;
-  fiat_p384_addcarryx_u32(&x396, &x397, x395, 0x0, x366);
+  fiat_p384_addcarryx_u32(&x396, &x397, x395, x366, 0x0);
   uint32_t x398;
   fiat_p384_uint1 x399;
-  fiat_p384_addcarryx_u32(&x398, &x399, x397, 0x0, x368);
+  fiat_p384_addcarryx_u32(&x398, &x399, x397, x368, 0x0);
   uint32_t x400;
   fiat_p384_uint1 x401;
-  fiat_p384_addcarryx_u32(&x400, &x401, x399, 0x0, x370);
+  fiat_p384_addcarryx_u32(&x400, &x401, x399, x370, 0x0);
   uint32_t x402;
   fiat_p384_uint1 x403;
-  fiat_p384_addcarryx_u32(&x402, &x403, x401, 0x0, x372);
+  fiat_p384_addcarryx_u32(&x402, &x403, x401, x372, 0x0);
   uint32_t x404;
   fiat_p384_uint1 x405;
-  fiat_p384_addcarryx_u32(&x404, &x405, x403, 0x0, x374);
+  fiat_p384_addcarryx_u32(&x404, &x405, x403, x374, 0x0);
   uint32_t x406;
   fiat_p384_uint1 x407;
-  fiat_p384_addcarryx_u32(&x406, &x407, x405, 0x0, x376);
+  fiat_p384_addcarryx_u32(&x406, &x407, x405, x376, 0x0);
   uint32_t x408;
   fiat_p384_uint1 x409;
-  fiat_p384_addcarryx_u32(&x408, &x409, x407, 0x0, x378);
+  fiat_p384_addcarryx_u32(&x408, &x409, x407, x378, 0x0);
   uint32_t x410;
   fiat_p384_uint1 x411;
-  fiat_p384_addcarryx_u32(&x410, &x411, x409, 0x0, x386);
+  fiat_p384_addcarryx_u32(&x410, &x411, x409, x386, 0x0);
   uint32_t x412;
   uint32_t x413;
   fiat_p384_mulx_u32(&x412, &x413, x388, UINT32_C(0xffffffff));
@@ -6267,112 +6267,112 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_mulx_u32(&x430, &x431, x388, UINT32_C(0xffffffff));
   uint32_t x432;
   fiat_p384_uint1 x433;
-  fiat_p384_addcarryx_u32(&x432, &x433, 0x0, x426, x429);
+  fiat_p384_addcarryx_u32(&x432, &x433, 0x0, x429, x426);
   uint32_t x434;
   fiat_p384_uint1 x435;
-  fiat_p384_addcarryx_u32(&x434, &x435, x433, x424, x427);
+  fiat_p384_addcarryx_u32(&x434, &x435, x433, x427, x424);
   uint32_t x436;
   fiat_p384_uint1 x437;
-  fiat_p384_addcarryx_u32(&x436, &x437, x435, x422, x425);
+  fiat_p384_addcarryx_u32(&x436, &x437, x435, x425, x422);
   uint32_t x438;
   fiat_p384_uint1 x439;
-  fiat_p384_addcarryx_u32(&x438, &x439, x437, x420, x423);
+  fiat_p384_addcarryx_u32(&x438, &x439, x437, x423, x420);
   uint32_t x440;
   fiat_p384_uint1 x441;
-  fiat_p384_addcarryx_u32(&x440, &x441, x439, x418, x421);
+  fiat_p384_addcarryx_u32(&x440, &x441, x439, x421, x418);
   uint32_t x442;
   fiat_p384_uint1 x443;
-  fiat_p384_addcarryx_u32(&x442, &x443, x441, x416, x419);
+  fiat_p384_addcarryx_u32(&x442, &x443, x441, x419, x416);
   uint32_t x444;
   fiat_p384_uint1 x445;
-  fiat_p384_addcarryx_u32(&x444, &x445, x443, x414, x417);
+  fiat_p384_addcarryx_u32(&x444, &x445, x443, x417, x414);
   uint32_t x446;
   fiat_p384_uint1 x447;
-  fiat_p384_addcarryx_u32(&x446, &x447, x445, x412, x415);
+  fiat_p384_addcarryx_u32(&x446, &x447, x445, x415, x412);
   uint32_t x448;
   fiat_p384_uint1 x449;
-  fiat_p384_addcarryx_u32(&x448, &x449, 0x0, x430, x388);
+  fiat_p384_addcarryx_u32(&x448, &x449, 0x0, x388, x430);
   uint32_t x450;
   fiat_p384_uint1 x451;
-  fiat_p384_addcarryx_u32(&x450, &x451, x449, x431, x390);
+  fiat_p384_addcarryx_u32(&x450, &x451, x449, x390, x431);
   uint32_t x452;
   fiat_p384_uint1 x453;
-  fiat_p384_addcarryx_u32(&x452, &x453, x451, 0x0, x392);
+  fiat_p384_addcarryx_u32(&x452, &x453, x451, x392, 0x0);
   uint32_t x454;
   fiat_p384_uint1 x455;
-  fiat_p384_addcarryx_u32(&x454, &x455, x453, x428, x394);
+  fiat_p384_addcarryx_u32(&x454, &x455, x453, x394, x428);
   uint32_t x456;
   fiat_p384_uint1 x457;
-  fiat_p384_addcarryx_u32(&x456, &x457, x455, x432, x396);
+  fiat_p384_addcarryx_u32(&x456, &x457, x455, x396, x432);
   uint32_t x458;
   fiat_p384_uint1 x459;
-  fiat_p384_addcarryx_u32(&x458, &x459, x457, x434, x398);
+  fiat_p384_addcarryx_u32(&x458, &x459, x457, x398, x434);
   uint32_t x460;
   fiat_p384_uint1 x461;
-  fiat_p384_addcarryx_u32(&x460, &x461, x459, x436, x400);
+  fiat_p384_addcarryx_u32(&x460, &x461, x459, x400, x436);
   uint32_t x462;
   fiat_p384_uint1 x463;
-  fiat_p384_addcarryx_u32(&x462, &x463, x461, x438, x402);
+  fiat_p384_addcarryx_u32(&x462, &x463, x461, x402, x438);
   uint32_t x464;
   fiat_p384_uint1 x465;
-  fiat_p384_addcarryx_u32(&x464, &x465, x463, x440, x404);
+  fiat_p384_addcarryx_u32(&x464, &x465, x463, x404, x440);
   uint32_t x466;
   fiat_p384_uint1 x467;
-  fiat_p384_addcarryx_u32(&x466, &x467, x465, x442, x406);
+  fiat_p384_addcarryx_u32(&x466, &x467, x465, x406, x442);
   uint32_t x468;
   fiat_p384_uint1 x469;
-  fiat_p384_addcarryx_u32(&x468, &x469, x467, x444, x408);
+  fiat_p384_addcarryx_u32(&x468, &x469, x467, x408, x444);
   uint32_t x470;
   fiat_p384_uint1 x471;
-  fiat_p384_addcarryx_u32(&x470, &x471, x469, x446, x410);
+  fiat_p384_addcarryx_u32(&x470, &x471, x469, x410, x446);
   uint32_t x472;
   fiat_p384_uint1 x473;
-  fiat_p384_addcarryx_u32(&x472, &x473, x387, 0x0, 0x0);
+  fiat_p384_addcarryx_u32(&x472, &x473, x447, x413, 0x0);
   uint32_t x474;
   fiat_p384_uint1 x475;
-  fiat_p384_addcarryx_u32(&x474, &x475, x411, 0x0, (fiat_p384_uint1)x472);
+  fiat_p384_addcarryx_u32(&x474, &x475, x387, 0x0, 0x0);
   uint32_t x476;
   fiat_p384_uint1 x477;
-  fiat_p384_addcarryx_u32(&x476, &x477, x447, 0x0, x413);
+  fiat_p384_addcarryx_u32(&x476, &x477, x411, (fiat_p384_uint1)x474, 0x0);
   uint32_t x478;
   fiat_p384_uint1 x479;
-  fiat_p384_addcarryx_u32(&x478, &x479, x471, x476, x474);
+  fiat_p384_addcarryx_u32(&x478, &x479, x471, x476, x472);
   uint32_t x480;
   fiat_p384_uint1 x481;
-  fiat_p384_addcarryx_u32(&x480, &x481, 0x0, (arg1[6]), x450);
+  fiat_p384_addcarryx_u32(&x480, &x481, 0x0, x450, (arg1[6]));
   uint32_t x482;
   fiat_p384_uint1 x483;
-  fiat_p384_addcarryx_u32(&x482, &x483, x481, 0x0, x452);
+  fiat_p384_addcarryx_u32(&x482, &x483, x481, x452, 0x0);
   uint32_t x484;
   fiat_p384_uint1 x485;
-  fiat_p384_addcarryx_u32(&x484, &x485, x483, 0x0, x454);
+  fiat_p384_addcarryx_u32(&x484, &x485, x483, x454, 0x0);
   uint32_t x486;
   fiat_p384_uint1 x487;
-  fiat_p384_addcarryx_u32(&x486, &x487, x485, 0x0, x456);
+  fiat_p384_addcarryx_u32(&x486, &x487, x485, x456, 0x0);
   uint32_t x488;
   fiat_p384_uint1 x489;
-  fiat_p384_addcarryx_u32(&x488, &x489, x487, 0x0, x458);
+  fiat_p384_addcarryx_u32(&x488, &x489, x487, x458, 0x0);
   uint32_t x490;
   fiat_p384_uint1 x491;
-  fiat_p384_addcarryx_u32(&x490, &x491, x489, 0x0, x460);
+  fiat_p384_addcarryx_u32(&x490, &x491, x489, x460, 0x0);
   uint32_t x492;
   fiat_p384_uint1 x493;
-  fiat_p384_addcarryx_u32(&x492, &x493, x491, 0x0, x462);
+  fiat_p384_addcarryx_u32(&x492, &x493, x491, x462, 0x0);
   uint32_t x494;
   fiat_p384_uint1 x495;
-  fiat_p384_addcarryx_u32(&x494, &x495, x493, 0x0, x464);
+  fiat_p384_addcarryx_u32(&x494, &x495, x493, x464, 0x0);
   uint32_t x496;
   fiat_p384_uint1 x497;
-  fiat_p384_addcarryx_u32(&x496, &x497, x495, 0x0, x466);
+  fiat_p384_addcarryx_u32(&x496, &x497, x495, x466, 0x0);
   uint32_t x498;
   fiat_p384_uint1 x499;
-  fiat_p384_addcarryx_u32(&x498, &x499, x497, 0x0, x468);
+  fiat_p384_addcarryx_u32(&x498, &x499, x497, x468, 0x0);
   uint32_t x500;
   fiat_p384_uint1 x501;
-  fiat_p384_addcarryx_u32(&x500, &x501, x499, 0x0, x470);
+  fiat_p384_addcarryx_u32(&x500, &x501, x499, x470, 0x0);
   uint32_t x502;
   fiat_p384_uint1 x503;
-  fiat_p384_addcarryx_u32(&x502, &x503, x501, 0x0, x478);
+  fiat_p384_addcarryx_u32(&x502, &x503, x501, x478, 0x0);
   uint32_t x504;
   uint32_t x505;
   fiat_p384_mulx_u32(&x504, &x505, x480, UINT32_C(0xffffffff));
@@ -6405,112 +6405,112 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_mulx_u32(&x522, &x523, x480, UINT32_C(0xffffffff));
   uint32_t x524;
   fiat_p384_uint1 x525;
-  fiat_p384_addcarryx_u32(&x524, &x525, 0x0, x518, x521);
+  fiat_p384_addcarryx_u32(&x524, &x525, 0x0, x521, x518);
   uint32_t x526;
   fiat_p384_uint1 x527;
-  fiat_p384_addcarryx_u32(&x526, &x527, x525, x516, x519);
+  fiat_p384_addcarryx_u32(&x526, &x527, x525, x519, x516);
   uint32_t x528;
   fiat_p384_uint1 x529;
-  fiat_p384_addcarryx_u32(&x528, &x529, x527, x514, x517);
+  fiat_p384_addcarryx_u32(&x528, &x529, x527, x517, x514);
   uint32_t x530;
   fiat_p384_uint1 x531;
-  fiat_p384_addcarryx_u32(&x530, &x531, x529, x512, x515);
+  fiat_p384_addcarryx_u32(&x530, &x531, x529, x515, x512);
   uint32_t x532;
   fiat_p384_uint1 x533;
-  fiat_p384_addcarryx_u32(&x532, &x533, x531, x510, x513);
+  fiat_p384_addcarryx_u32(&x532, &x533, x531, x513, x510);
   uint32_t x534;
   fiat_p384_uint1 x535;
-  fiat_p384_addcarryx_u32(&x534, &x535, x533, x508, x511);
+  fiat_p384_addcarryx_u32(&x534, &x535, x533, x511, x508);
   uint32_t x536;
   fiat_p384_uint1 x537;
-  fiat_p384_addcarryx_u32(&x536, &x537, x535, x506, x509);
+  fiat_p384_addcarryx_u32(&x536, &x537, x535, x509, x506);
   uint32_t x538;
   fiat_p384_uint1 x539;
-  fiat_p384_addcarryx_u32(&x538, &x539, x537, x504, x507);
+  fiat_p384_addcarryx_u32(&x538, &x539, x537, x507, x504);
   uint32_t x540;
   fiat_p384_uint1 x541;
-  fiat_p384_addcarryx_u32(&x540, &x541, 0x0, x522, x480);
+  fiat_p384_addcarryx_u32(&x540, &x541, 0x0, x480, x522);
   uint32_t x542;
   fiat_p384_uint1 x543;
-  fiat_p384_addcarryx_u32(&x542, &x543, x541, x523, x482);
+  fiat_p384_addcarryx_u32(&x542, &x543, x541, x482, x523);
   uint32_t x544;
   fiat_p384_uint1 x545;
-  fiat_p384_addcarryx_u32(&x544, &x545, x543, 0x0, x484);
+  fiat_p384_addcarryx_u32(&x544, &x545, x543, x484, 0x0);
   uint32_t x546;
   fiat_p384_uint1 x547;
-  fiat_p384_addcarryx_u32(&x546, &x547, x545, x520, x486);
+  fiat_p384_addcarryx_u32(&x546, &x547, x545, x486, x520);
   uint32_t x548;
   fiat_p384_uint1 x549;
-  fiat_p384_addcarryx_u32(&x548, &x549, x547, x524, x488);
+  fiat_p384_addcarryx_u32(&x548, &x549, x547, x488, x524);
   uint32_t x550;
   fiat_p384_uint1 x551;
-  fiat_p384_addcarryx_u32(&x550, &x551, x549, x526, x490);
+  fiat_p384_addcarryx_u32(&x550, &x551, x549, x490, x526);
   uint32_t x552;
   fiat_p384_uint1 x553;
-  fiat_p384_addcarryx_u32(&x552, &x553, x551, x528, x492);
+  fiat_p384_addcarryx_u32(&x552, &x553, x551, x492, x528);
   uint32_t x554;
   fiat_p384_uint1 x555;
-  fiat_p384_addcarryx_u32(&x554, &x555, x553, x530, x494);
+  fiat_p384_addcarryx_u32(&x554, &x555, x553, x494, x530);
   uint32_t x556;
   fiat_p384_uint1 x557;
-  fiat_p384_addcarryx_u32(&x556, &x557, x555, x532, x496);
+  fiat_p384_addcarryx_u32(&x556, &x557, x555, x496, x532);
   uint32_t x558;
   fiat_p384_uint1 x559;
-  fiat_p384_addcarryx_u32(&x558, &x559, x557, x534, x498);
+  fiat_p384_addcarryx_u32(&x558, &x559, x557, x498, x534);
   uint32_t x560;
   fiat_p384_uint1 x561;
-  fiat_p384_addcarryx_u32(&x560, &x561, x559, x536, x500);
+  fiat_p384_addcarryx_u32(&x560, &x561, x559, x500, x536);
   uint32_t x562;
   fiat_p384_uint1 x563;
-  fiat_p384_addcarryx_u32(&x562, &x563, x561, x538, x502);
+  fiat_p384_addcarryx_u32(&x562, &x563, x561, x502, x538);
   uint32_t x564;
   fiat_p384_uint1 x565;
-  fiat_p384_addcarryx_u32(&x564, &x565, x479, 0x0, 0x0);
+  fiat_p384_addcarryx_u32(&x564, &x565, x539, x505, 0x0);
   uint32_t x566;
   fiat_p384_uint1 x567;
-  fiat_p384_addcarryx_u32(&x566, &x567, x503, 0x0, (fiat_p384_uint1)x564);
+  fiat_p384_addcarryx_u32(&x566, &x567, x479, 0x0, 0x0);
   uint32_t x568;
   fiat_p384_uint1 x569;
-  fiat_p384_addcarryx_u32(&x568, &x569, x539, 0x0, x505);
+  fiat_p384_addcarryx_u32(&x568, &x569, x503, (fiat_p384_uint1)x566, 0x0);
   uint32_t x570;
   fiat_p384_uint1 x571;
-  fiat_p384_addcarryx_u32(&x570, &x571, x563, x568, x566);
+  fiat_p384_addcarryx_u32(&x570, &x571, x563, x568, x564);
   uint32_t x572;
   fiat_p384_uint1 x573;
-  fiat_p384_addcarryx_u32(&x572, &x573, 0x0, (arg1[7]), x542);
+  fiat_p384_addcarryx_u32(&x572, &x573, 0x0, x542, (arg1[7]));
   uint32_t x574;
   fiat_p384_uint1 x575;
-  fiat_p384_addcarryx_u32(&x574, &x575, x573, 0x0, x544);
+  fiat_p384_addcarryx_u32(&x574, &x575, x573, x544, 0x0);
   uint32_t x576;
   fiat_p384_uint1 x577;
-  fiat_p384_addcarryx_u32(&x576, &x577, x575, 0x0, x546);
+  fiat_p384_addcarryx_u32(&x576, &x577, x575, x546, 0x0);
   uint32_t x578;
   fiat_p384_uint1 x579;
-  fiat_p384_addcarryx_u32(&x578, &x579, x577, 0x0, x548);
+  fiat_p384_addcarryx_u32(&x578, &x579, x577, x548, 0x0);
   uint32_t x580;
   fiat_p384_uint1 x581;
-  fiat_p384_addcarryx_u32(&x580, &x581, x579, 0x0, x550);
+  fiat_p384_addcarryx_u32(&x580, &x581, x579, x550, 0x0);
   uint32_t x582;
   fiat_p384_uint1 x583;
-  fiat_p384_addcarryx_u32(&x582, &x583, x581, 0x0, x552);
+  fiat_p384_addcarryx_u32(&x582, &x583, x581, x552, 0x0);
   uint32_t x584;
   fiat_p384_uint1 x585;
-  fiat_p384_addcarryx_u32(&x584, &x585, x583, 0x0, x554);
+  fiat_p384_addcarryx_u32(&x584, &x585, x583, x554, 0x0);
   uint32_t x586;
   fiat_p384_uint1 x587;
-  fiat_p384_addcarryx_u32(&x586, &x587, x585, 0x0, x556);
+  fiat_p384_addcarryx_u32(&x586, &x587, x585, x556, 0x0);
   uint32_t x588;
   fiat_p384_uint1 x589;
-  fiat_p384_addcarryx_u32(&x588, &x589, x587, 0x0, x558);
+  fiat_p384_addcarryx_u32(&x588, &x589, x587, x558, 0x0);
   uint32_t x590;
   fiat_p384_uint1 x591;
-  fiat_p384_addcarryx_u32(&x590, &x591, x589, 0x0, x560);
+  fiat_p384_addcarryx_u32(&x590, &x591, x589, x560, 0x0);
   uint32_t x592;
   fiat_p384_uint1 x593;
-  fiat_p384_addcarryx_u32(&x592, &x593, x591, 0x0, x562);
+  fiat_p384_addcarryx_u32(&x592, &x593, x591, x562, 0x0);
   uint32_t x594;
   fiat_p384_uint1 x595;
-  fiat_p384_addcarryx_u32(&x594, &x595, x593, 0x0, x570);
+  fiat_p384_addcarryx_u32(&x594, &x595, x593, x570, 0x0);
   uint32_t x596;
   uint32_t x597;
   fiat_p384_mulx_u32(&x596, &x597, x572, UINT32_C(0xffffffff));
@@ -6543,112 +6543,112 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_mulx_u32(&x614, &x615, x572, UINT32_C(0xffffffff));
   uint32_t x616;
   fiat_p384_uint1 x617;
-  fiat_p384_addcarryx_u32(&x616, &x617, 0x0, x610, x613);
+  fiat_p384_addcarryx_u32(&x616, &x617, 0x0, x613, x610);
   uint32_t x618;
   fiat_p384_uint1 x619;
-  fiat_p384_addcarryx_u32(&x618, &x619, x617, x608, x611);
+  fiat_p384_addcarryx_u32(&x618, &x619, x617, x611, x608);
   uint32_t x620;
   fiat_p384_uint1 x621;
-  fiat_p384_addcarryx_u32(&x620, &x621, x619, x606, x609);
+  fiat_p384_addcarryx_u32(&x620, &x621, x619, x609, x606);
   uint32_t x622;
   fiat_p384_uint1 x623;
-  fiat_p384_addcarryx_u32(&x622, &x623, x621, x604, x607);
+  fiat_p384_addcarryx_u32(&x622, &x623, x621, x607, x604);
   uint32_t x624;
   fiat_p384_uint1 x625;
-  fiat_p384_addcarryx_u32(&x624, &x625, x623, x602, x605);
+  fiat_p384_addcarryx_u32(&x624, &x625, x623, x605, x602);
   uint32_t x626;
   fiat_p384_uint1 x627;
-  fiat_p384_addcarryx_u32(&x626, &x627, x625, x600, x603);
+  fiat_p384_addcarryx_u32(&x626, &x627, x625, x603, x600);
   uint32_t x628;
   fiat_p384_uint1 x629;
-  fiat_p384_addcarryx_u32(&x628, &x629, x627, x598, x601);
+  fiat_p384_addcarryx_u32(&x628, &x629, x627, x601, x598);
   uint32_t x630;
   fiat_p384_uint1 x631;
-  fiat_p384_addcarryx_u32(&x630, &x631, x629, x596, x599);
+  fiat_p384_addcarryx_u32(&x630, &x631, x629, x599, x596);
   uint32_t x632;
   fiat_p384_uint1 x633;
-  fiat_p384_addcarryx_u32(&x632, &x633, 0x0, x614, x572);
+  fiat_p384_addcarryx_u32(&x632, &x633, 0x0, x572, x614);
   uint32_t x634;
   fiat_p384_uint1 x635;
-  fiat_p384_addcarryx_u32(&x634, &x635, x633, x615, x574);
+  fiat_p384_addcarryx_u32(&x634, &x635, x633, x574, x615);
   uint32_t x636;
   fiat_p384_uint1 x637;
-  fiat_p384_addcarryx_u32(&x636, &x637, x635, 0x0, x576);
+  fiat_p384_addcarryx_u32(&x636, &x637, x635, x576, 0x0);
   uint32_t x638;
   fiat_p384_uint1 x639;
-  fiat_p384_addcarryx_u32(&x638, &x639, x637, x612, x578);
+  fiat_p384_addcarryx_u32(&x638, &x639, x637, x578, x612);
   uint32_t x640;
   fiat_p384_uint1 x641;
-  fiat_p384_addcarryx_u32(&x640, &x641, x639, x616, x580);
+  fiat_p384_addcarryx_u32(&x640, &x641, x639, x580, x616);
   uint32_t x642;
   fiat_p384_uint1 x643;
-  fiat_p384_addcarryx_u32(&x642, &x643, x641, x618, x582);
+  fiat_p384_addcarryx_u32(&x642, &x643, x641, x582, x618);
   uint32_t x644;
   fiat_p384_uint1 x645;
-  fiat_p384_addcarryx_u32(&x644, &x645, x643, x620, x584);
+  fiat_p384_addcarryx_u32(&x644, &x645, x643, x584, x620);
   uint32_t x646;
   fiat_p384_uint1 x647;
-  fiat_p384_addcarryx_u32(&x646, &x647, x645, x622, x586);
+  fiat_p384_addcarryx_u32(&x646, &x647, x645, x586, x622);
   uint32_t x648;
   fiat_p384_uint1 x649;
-  fiat_p384_addcarryx_u32(&x648, &x649, x647, x624, x588);
+  fiat_p384_addcarryx_u32(&x648, &x649, x647, x588, x624);
   uint32_t x650;
   fiat_p384_uint1 x651;
-  fiat_p384_addcarryx_u32(&x650, &x651, x649, x626, x590);
+  fiat_p384_addcarryx_u32(&x650, &x651, x649, x590, x626);
   uint32_t x652;
   fiat_p384_uint1 x653;
-  fiat_p384_addcarryx_u32(&x652, &x653, x651, x628, x592);
+  fiat_p384_addcarryx_u32(&x652, &x653, x651, x592, x628);
   uint32_t x654;
   fiat_p384_uint1 x655;
-  fiat_p384_addcarryx_u32(&x654, &x655, x653, x630, x594);
+  fiat_p384_addcarryx_u32(&x654, &x655, x653, x594, x630);
   uint32_t x656;
   fiat_p384_uint1 x657;
-  fiat_p384_addcarryx_u32(&x656, &x657, x571, 0x0, 0x0);
+  fiat_p384_addcarryx_u32(&x656, &x657, x631, x597, 0x0);
   uint32_t x658;
   fiat_p384_uint1 x659;
-  fiat_p384_addcarryx_u32(&x658, &x659, x595, 0x0, (fiat_p384_uint1)x656);
+  fiat_p384_addcarryx_u32(&x658, &x659, x571, 0x0, 0x0);
   uint32_t x660;
   fiat_p384_uint1 x661;
-  fiat_p384_addcarryx_u32(&x660, &x661, x631, 0x0, x597);
+  fiat_p384_addcarryx_u32(&x660, &x661, x595, (fiat_p384_uint1)x658, 0x0);
   uint32_t x662;
   fiat_p384_uint1 x663;
-  fiat_p384_addcarryx_u32(&x662, &x663, x655, x660, x658);
+  fiat_p384_addcarryx_u32(&x662, &x663, x655, x660, x656);
   uint32_t x664;
   fiat_p384_uint1 x665;
-  fiat_p384_addcarryx_u32(&x664, &x665, 0x0, (arg1[8]), x634);
+  fiat_p384_addcarryx_u32(&x664, &x665, 0x0, x634, (arg1[8]));
   uint32_t x666;
   fiat_p384_uint1 x667;
-  fiat_p384_addcarryx_u32(&x666, &x667, x665, 0x0, x636);
+  fiat_p384_addcarryx_u32(&x666, &x667, x665, x636, 0x0);
   uint32_t x668;
   fiat_p384_uint1 x669;
-  fiat_p384_addcarryx_u32(&x668, &x669, x667, 0x0, x638);
+  fiat_p384_addcarryx_u32(&x668, &x669, x667, x638, 0x0);
   uint32_t x670;
   fiat_p384_uint1 x671;
-  fiat_p384_addcarryx_u32(&x670, &x671, x669, 0x0, x640);
+  fiat_p384_addcarryx_u32(&x670, &x671, x669, x640, 0x0);
   uint32_t x672;
   fiat_p384_uint1 x673;
-  fiat_p384_addcarryx_u32(&x672, &x673, x671, 0x0, x642);
+  fiat_p384_addcarryx_u32(&x672, &x673, x671, x642, 0x0);
   uint32_t x674;
   fiat_p384_uint1 x675;
-  fiat_p384_addcarryx_u32(&x674, &x675, x673, 0x0, x644);
+  fiat_p384_addcarryx_u32(&x674, &x675, x673, x644, 0x0);
   uint32_t x676;
   fiat_p384_uint1 x677;
-  fiat_p384_addcarryx_u32(&x676, &x677, x675, 0x0, x646);
+  fiat_p384_addcarryx_u32(&x676, &x677, x675, x646, 0x0);
   uint32_t x678;
   fiat_p384_uint1 x679;
-  fiat_p384_addcarryx_u32(&x678, &x679, x677, 0x0, x648);
+  fiat_p384_addcarryx_u32(&x678, &x679, x677, x648, 0x0);
   uint32_t x680;
   fiat_p384_uint1 x681;
-  fiat_p384_addcarryx_u32(&x680, &x681, x679, 0x0, x650);
+  fiat_p384_addcarryx_u32(&x680, &x681, x679, x650, 0x0);
   uint32_t x682;
   fiat_p384_uint1 x683;
-  fiat_p384_addcarryx_u32(&x682, &x683, x681, 0x0, x652);
+  fiat_p384_addcarryx_u32(&x682, &x683, x681, x652, 0x0);
   uint32_t x684;
   fiat_p384_uint1 x685;
-  fiat_p384_addcarryx_u32(&x684, &x685, x683, 0x0, x654);
+  fiat_p384_addcarryx_u32(&x684, &x685, x683, x654, 0x0);
   uint32_t x686;
   fiat_p384_uint1 x687;
-  fiat_p384_addcarryx_u32(&x686, &x687, x685, 0x0, x662);
+  fiat_p384_addcarryx_u32(&x686, &x687, x685, x662, 0x0);
   uint32_t x688;
   uint32_t x689;
   fiat_p384_mulx_u32(&x688, &x689, x664, UINT32_C(0xffffffff));
@@ -6681,112 +6681,112 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_mulx_u32(&x706, &x707, x664, UINT32_C(0xffffffff));
   uint32_t x708;
   fiat_p384_uint1 x709;
-  fiat_p384_addcarryx_u32(&x708, &x709, 0x0, x702, x705);
+  fiat_p384_addcarryx_u32(&x708, &x709, 0x0, x705, x702);
   uint32_t x710;
   fiat_p384_uint1 x711;
-  fiat_p384_addcarryx_u32(&x710, &x711, x709, x700, x703);
+  fiat_p384_addcarryx_u32(&x710, &x711, x709, x703, x700);
   uint32_t x712;
   fiat_p384_uint1 x713;
-  fiat_p384_addcarryx_u32(&x712, &x713, x711, x698, x701);
+  fiat_p384_addcarryx_u32(&x712, &x713, x711, x701, x698);
   uint32_t x714;
   fiat_p384_uint1 x715;
-  fiat_p384_addcarryx_u32(&x714, &x715, x713, x696, x699);
+  fiat_p384_addcarryx_u32(&x714, &x715, x713, x699, x696);
   uint32_t x716;
   fiat_p384_uint1 x717;
-  fiat_p384_addcarryx_u32(&x716, &x717, x715, x694, x697);
+  fiat_p384_addcarryx_u32(&x716, &x717, x715, x697, x694);
   uint32_t x718;
   fiat_p384_uint1 x719;
-  fiat_p384_addcarryx_u32(&x718, &x719, x717, x692, x695);
+  fiat_p384_addcarryx_u32(&x718, &x719, x717, x695, x692);
   uint32_t x720;
   fiat_p384_uint1 x721;
-  fiat_p384_addcarryx_u32(&x720, &x721, x719, x690, x693);
+  fiat_p384_addcarryx_u32(&x720, &x721, x719, x693, x690);
   uint32_t x722;
   fiat_p384_uint1 x723;
-  fiat_p384_addcarryx_u32(&x722, &x723, x721, x688, x691);
+  fiat_p384_addcarryx_u32(&x722, &x723, x721, x691, x688);
   uint32_t x724;
   fiat_p384_uint1 x725;
-  fiat_p384_addcarryx_u32(&x724, &x725, 0x0, x706, x664);
+  fiat_p384_addcarryx_u32(&x724, &x725, 0x0, x664, x706);
   uint32_t x726;
   fiat_p384_uint1 x727;
-  fiat_p384_addcarryx_u32(&x726, &x727, x725, x707, x666);
+  fiat_p384_addcarryx_u32(&x726, &x727, x725, x666, x707);
   uint32_t x728;
   fiat_p384_uint1 x729;
-  fiat_p384_addcarryx_u32(&x728, &x729, x727, 0x0, x668);
+  fiat_p384_addcarryx_u32(&x728, &x729, x727, x668, 0x0);
   uint32_t x730;
   fiat_p384_uint1 x731;
-  fiat_p384_addcarryx_u32(&x730, &x731, x729, x704, x670);
+  fiat_p384_addcarryx_u32(&x730, &x731, x729, x670, x704);
   uint32_t x732;
   fiat_p384_uint1 x733;
-  fiat_p384_addcarryx_u32(&x732, &x733, x731, x708, x672);
+  fiat_p384_addcarryx_u32(&x732, &x733, x731, x672, x708);
   uint32_t x734;
   fiat_p384_uint1 x735;
-  fiat_p384_addcarryx_u32(&x734, &x735, x733, x710, x674);
+  fiat_p384_addcarryx_u32(&x734, &x735, x733, x674, x710);
   uint32_t x736;
   fiat_p384_uint1 x737;
-  fiat_p384_addcarryx_u32(&x736, &x737, x735, x712, x676);
+  fiat_p384_addcarryx_u32(&x736, &x737, x735, x676, x712);
   uint32_t x738;
   fiat_p384_uint1 x739;
-  fiat_p384_addcarryx_u32(&x738, &x739, x737, x714, x678);
+  fiat_p384_addcarryx_u32(&x738, &x739, x737, x678, x714);
   uint32_t x740;
   fiat_p384_uint1 x741;
-  fiat_p384_addcarryx_u32(&x740, &x741, x739, x716, x680);
+  fiat_p384_addcarryx_u32(&x740, &x741, x739, x680, x716);
   uint32_t x742;
   fiat_p384_uint1 x743;
-  fiat_p384_addcarryx_u32(&x742, &x743, x741, x718, x682);
+  fiat_p384_addcarryx_u32(&x742, &x743, x741, x682, x718);
   uint32_t x744;
   fiat_p384_uint1 x745;
-  fiat_p384_addcarryx_u32(&x744, &x745, x743, x720, x684);
+  fiat_p384_addcarryx_u32(&x744, &x745, x743, x684, x720);
   uint32_t x746;
   fiat_p384_uint1 x747;
-  fiat_p384_addcarryx_u32(&x746, &x747, x745, x722, x686);
+  fiat_p384_addcarryx_u32(&x746, &x747, x745, x686, x722);
   uint32_t x748;
   fiat_p384_uint1 x749;
-  fiat_p384_addcarryx_u32(&x748, &x749, x663, 0x0, 0x0);
+  fiat_p384_addcarryx_u32(&x748, &x749, x723, x689, 0x0);
   uint32_t x750;
   fiat_p384_uint1 x751;
-  fiat_p384_addcarryx_u32(&x750, &x751, x687, 0x0, (fiat_p384_uint1)x748);
+  fiat_p384_addcarryx_u32(&x750, &x751, x663, 0x0, 0x0);
   uint32_t x752;
   fiat_p384_uint1 x753;
-  fiat_p384_addcarryx_u32(&x752, &x753, x723, 0x0, x689);
+  fiat_p384_addcarryx_u32(&x752, &x753, x687, (fiat_p384_uint1)x750, 0x0);
   uint32_t x754;
   fiat_p384_uint1 x755;
-  fiat_p384_addcarryx_u32(&x754, &x755, x747, x752, x750);
+  fiat_p384_addcarryx_u32(&x754, &x755, x747, x752, x748);
   uint32_t x756;
   fiat_p384_uint1 x757;
-  fiat_p384_addcarryx_u32(&x756, &x757, 0x0, (arg1[9]), x726);
+  fiat_p384_addcarryx_u32(&x756, &x757, 0x0, x726, (arg1[9]));
   uint32_t x758;
   fiat_p384_uint1 x759;
-  fiat_p384_addcarryx_u32(&x758, &x759, x757, 0x0, x728);
+  fiat_p384_addcarryx_u32(&x758, &x759, x757, x728, 0x0);
   uint32_t x760;
   fiat_p384_uint1 x761;
-  fiat_p384_addcarryx_u32(&x760, &x761, x759, 0x0, x730);
+  fiat_p384_addcarryx_u32(&x760, &x761, x759, x730, 0x0);
   uint32_t x762;
   fiat_p384_uint1 x763;
-  fiat_p384_addcarryx_u32(&x762, &x763, x761, 0x0, x732);
+  fiat_p384_addcarryx_u32(&x762, &x763, x761, x732, 0x0);
   uint32_t x764;
   fiat_p384_uint1 x765;
-  fiat_p384_addcarryx_u32(&x764, &x765, x763, 0x0, x734);
+  fiat_p384_addcarryx_u32(&x764, &x765, x763, x734, 0x0);
   uint32_t x766;
   fiat_p384_uint1 x767;
-  fiat_p384_addcarryx_u32(&x766, &x767, x765, 0x0, x736);
+  fiat_p384_addcarryx_u32(&x766, &x767, x765, x736, 0x0);
   uint32_t x768;
   fiat_p384_uint1 x769;
-  fiat_p384_addcarryx_u32(&x768, &x769, x767, 0x0, x738);
+  fiat_p384_addcarryx_u32(&x768, &x769, x767, x738, 0x0);
   uint32_t x770;
   fiat_p384_uint1 x771;
-  fiat_p384_addcarryx_u32(&x770, &x771, x769, 0x0, x740);
+  fiat_p384_addcarryx_u32(&x770, &x771, x769, x740, 0x0);
   uint32_t x772;
   fiat_p384_uint1 x773;
-  fiat_p384_addcarryx_u32(&x772, &x773, x771, 0x0, x742);
+  fiat_p384_addcarryx_u32(&x772, &x773, x771, x742, 0x0);
   uint32_t x774;
   fiat_p384_uint1 x775;
-  fiat_p384_addcarryx_u32(&x774, &x775, x773, 0x0, x744);
+  fiat_p384_addcarryx_u32(&x774, &x775, x773, x744, 0x0);
   uint32_t x776;
   fiat_p384_uint1 x777;
-  fiat_p384_addcarryx_u32(&x776, &x777, x775, 0x0, x746);
+  fiat_p384_addcarryx_u32(&x776, &x777, x775, x746, 0x0);
   uint32_t x778;
   fiat_p384_uint1 x779;
-  fiat_p384_addcarryx_u32(&x778, &x779, x777, 0x0, x754);
+  fiat_p384_addcarryx_u32(&x778, &x779, x777, x754, 0x0);
   uint32_t x780;
   uint32_t x781;
   fiat_p384_mulx_u32(&x780, &x781, x756, UINT32_C(0xffffffff));
@@ -6819,112 +6819,112 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_mulx_u32(&x798, &x799, x756, UINT32_C(0xffffffff));
   uint32_t x800;
   fiat_p384_uint1 x801;
-  fiat_p384_addcarryx_u32(&x800, &x801, 0x0, x794, x797);
+  fiat_p384_addcarryx_u32(&x800, &x801, 0x0, x797, x794);
   uint32_t x802;
   fiat_p384_uint1 x803;
-  fiat_p384_addcarryx_u32(&x802, &x803, x801, x792, x795);
+  fiat_p384_addcarryx_u32(&x802, &x803, x801, x795, x792);
   uint32_t x804;
   fiat_p384_uint1 x805;
-  fiat_p384_addcarryx_u32(&x804, &x805, x803, x790, x793);
+  fiat_p384_addcarryx_u32(&x804, &x805, x803, x793, x790);
   uint32_t x806;
   fiat_p384_uint1 x807;
-  fiat_p384_addcarryx_u32(&x806, &x807, x805, x788, x791);
+  fiat_p384_addcarryx_u32(&x806, &x807, x805, x791, x788);
   uint32_t x808;
   fiat_p384_uint1 x809;
-  fiat_p384_addcarryx_u32(&x808, &x809, x807, x786, x789);
+  fiat_p384_addcarryx_u32(&x808, &x809, x807, x789, x786);
   uint32_t x810;
   fiat_p384_uint1 x811;
-  fiat_p384_addcarryx_u32(&x810, &x811, x809, x784, x787);
+  fiat_p384_addcarryx_u32(&x810, &x811, x809, x787, x784);
   uint32_t x812;
   fiat_p384_uint1 x813;
-  fiat_p384_addcarryx_u32(&x812, &x813, x811, x782, x785);
+  fiat_p384_addcarryx_u32(&x812, &x813, x811, x785, x782);
   uint32_t x814;
   fiat_p384_uint1 x815;
-  fiat_p384_addcarryx_u32(&x814, &x815, x813, x780, x783);
+  fiat_p384_addcarryx_u32(&x814, &x815, x813, x783, x780);
   uint32_t x816;
   fiat_p384_uint1 x817;
-  fiat_p384_addcarryx_u32(&x816, &x817, 0x0, x798, x756);
+  fiat_p384_addcarryx_u32(&x816, &x817, 0x0, x756, x798);
   uint32_t x818;
   fiat_p384_uint1 x819;
-  fiat_p384_addcarryx_u32(&x818, &x819, x817, x799, x758);
+  fiat_p384_addcarryx_u32(&x818, &x819, x817, x758, x799);
   uint32_t x820;
   fiat_p384_uint1 x821;
-  fiat_p384_addcarryx_u32(&x820, &x821, x819, 0x0, x760);
+  fiat_p384_addcarryx_u32(&x820, &x821, x819, x760, 0x0);
   uint32_t x822;
   fiat_p384_uint1 x823;
-  fiat_p384_addcarryx_u32(&x822, &x823, x821, x796, x762);
+  fiat_p384_addcarryx_u32(&x822, &x823, x821, x762, x796);
   uint32_t x824;
   fiat_p384_uint1 x825;
-  fiat_p384_addcarryx_u32(&x824, &x825, x823, x800, x764);
+  fiat_p384_addcarryx_u32(&x824, &x825, x823, x764, x800);
   uint32_t x826;
   fiat_p384_uint1 x827;
-  fiat_p384_addcarryx_u32(&x826, &x827, x825, x802, x766);
+  fiat_p384_addcarryx_u32(&x826, &x827, x825, x766, x802);
   uint32_t x828;
   fiat_p384_uint1 x829;
-  fiat_p384_addcarryx_u32(&x828, &x829, x827, x804, x768);
+  fiat_p384_addcarryx_u32(&x828, &x829, x827, x768, x804);
   uint32_t x830;
   fiat_p384_uint1 x831;
-  fiat_p384_addcarryx_u32(&x830, &x831, x829, x806, x770);
+  fiat_p384_addcarryx_u32(&x830, &x831, x829, x770, x806);
   uint32_t x832;
   fiat_p384_uint1 x833;
-  fiat_p384_addcarryx_u32(&x832, &x833, x831, x808, x772);
+  fiat_p384_addcarryx_u32(&x832, &x833, x831, x772, x808);
   uint32_t x834;
   fiat_p384_uint1 x835;
-  fiat_p384_addcarryx_u32(&x834, &x835, x833, x810, x774);
+  fiat_p384_addcarryx_u32(&x834, &x835, x833, x774, x810);
   uint32_t x836;
   fiat_p384_uint1 x837;
-  fiat_p384_addcarryx_u32(&x836, &x837, x835, x812, x776);
+  fiat_p384_addcarryx_u32(&x836, &x837, x835, x776, x812);
   uint32_t x838;
   fiat_p384_uint1 x839;
-  fiat_p384_addcarryx_u32(&x838, &x839, x837, x814, x778);
+  fiat_p384_addcarryx_u32(&x838, &x839, x837, x778, x814);
   uint32_t x840;
   fiat_p384_uint1 x841;
-  fiat_p384_addcarryx_u32(&x840, &x841, x755, 0x0, 0x0);
+  fiat_p384_addcarryx_u32(&x840, &x841, x815, x781, 0x0);
   uint32_t x842;
   fiat_p384_uint1 x843;
-  fiat_p384_addcarryx_u32(&x842, &x843, x779, 0x0, (fiat_p384_uint1)x840);
+  fiat_p384_addcarryx_u32(&x842, &x843, x755, 0x0, 0x0);
   uint32_t x844;
   fiat_p384_uint1 x845;
-  fiat_p384_addcarryx_u32(&x844, &x845, x815, 0x0, x781);
+  fiat_p384_addcarryx_u32(&x844, &x845, x779, (fiat_p384_uint1)x842, 0x0);
   uint32_t x846;
   fiat_p384_uint1 x847;
-  fiat_p384_addcarryx_u32(&x846, &x847, x839, x844, x842);
+  fiat_p384_addcarryx_u32(&x846, &x847, x839, x844, x840);
   uint32_t x848;
   fiat_p384_uint1 x849;
-  fiat_p384_addcarryx_u32(&x848, &x849, 0x0, (arg1[10]), x818);
+  fiat_p384_addcarryx_u32(&x848, &x849, 0x0, x818, (arg1[10]));
   uint32_t x850;
   fiat_p384_uint1 x851;
-  fiat_p384_addcarryx_u32(&x850, &x851, x849, 0x0, x820);
+  fiat_p384_addcarryx_u32(&x850, &x851, x849, x820, 0x0);
   uint32_t x852;
   fiat_p384_uint1 x853;
-  fiat_p384_addcarryx_u32(&x852, &x853, x851, 0x0, x822);
+  fiat_p384_addcarryx_u32(&x852, &x853, x851, x822, 0x0);
   uint32_t x854;
   fiat_p384_uint1 x855;
-  fiat_p384_addcarryx_u32(&x854, &x855, x853, 0x0, x824);
+  fiat_p384_addcarryx_u32(&x854, &x855, x853, x824, 0x0);
   uint32_t x856;
   fiat_p384_uint1 x857;
-  fiat_p384_addcarryx_u32(&x856, &x857, x855, 0x0, x826);
+  fiat_p384_addcarryx_u32(&x856, &x857, x855, x826, 0x0);
   uint32_t x858;
   fiat_p384_uint1 x859;
-  fiat_p384_addcarryx_u32(&x858, &x859, x857, 0x0, x828);
+  fiat_p384_addcarryx_u32(&x858, &x859, x857, x828, 0x0);
   uint32_t x860;
   fiat_p384_uint1 x861;
-  fiat_p384_addcarryx_u32(&x860, &x861, x859, 0x0, x830);
+  fiat_p384_addcarryx_u32(&x860, &x861, x859, x830, 0x0);
   uint32_t x862;
   fiat_p384_uint1 x863;
-  fiat_p384_addcarryx_u32(&x862, &x863, x861, 0x0, x832);
+  fiat_p384_addcarryx_u32(&x862, &x863, x861, x832, 0x0);
   uint32_t x864;
   fiat_p384_uint1 x865;
-  fiat_p384_addcarryx_u32(&x864, &x865, x863, 0x0, x834);
+  fiat_p384_addcarryx_u32(&x864, &x865, x863, x834, 0x0);
   uint32_t x866;
   fiat_p384_uint1 x867;
-  fiat_p384_addcarryx_u32(&x866, &x867, x865, 0x0, x836);
+  fiat_p384_addcarryx_u32(&x866, &x867, x865, x836, 0x0);
   uint32_t x868;
   fiat_p384_uint1 x869;
-  fiat_p384_addcarryx_u32(&x868, &x869, x867, 0x0, x838);
+  fiat_p384_addcarryx_u32(&x868, &x869, x867, x838, 0x0);
   uint32_t x870;
   fiat_p384_uint1 x871;
-  fiat_p384_addcarryx_u32(&x870, &x871, x869, 0x0, x846);
+  fiat_p384_addcarryx_u32(&x870, &x871, x869, x846, 0x0);
   uint32_t x872;
   uint32_t x873;
   fiat_p384_mulx_u32(&x872, &x873, x848, UINT32_C(0xffffffff));
@@ -6957,112 +6957,112 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_mulx_u32(&x890, &x891, x848, UINT32_C(0xffffffff));
   uint32_t x892;
   fiat_p384_uint1 x893;
-  fiat_p384_addcarryx_u32(&x892, &x893, 0x0, x886, x889);
+  fiat_p384_addcarryx_u32(&x892, &x893, 0x0, x889, x886);
   uint32_t x894;
   fiat_p384_uint1 x895;
-  fiat_p384_addcarryx_u32(&x894, &x895, x893, x884, x887);
+  fiat_p384_addcarryx_u32(&x894, &x895, x893, x887, x884);
   uint32_t x896;
   fiat_p384_uint1 x897;
-  fiat_p384_addcarryx_u32(&x896, &x897, x895, x882, x885);
+  fiat_p384_addcarryx_u32(&x896, &x897, x895, x885, x882);
   uint32_t x898;
   fiat_p384_uint1 x899;
-  fiat_p384_addcarryx_u32(&x898, &x899, x897, x880, x883);
+  fiat_p384_addcarryx_u32(&x898, &x899, x897, x883, x880);
   uint32_t x900;
   fiat_p384_uint1 x901;
-  fiat_p384_addcarryx_u32(&x900, &x901, x899, x878, x881);
+  fiat_p384_addcarryx_u32(&x900, &x901, x899, x881, x878);
   uint32_t x902;
   fiat_p384_uint1 x903;
-  fiat_p384_addcarryx_u32(&x902, &x903, x901, x876, x879);
+  fiat_p384_addcarryx_u32(&x902, &x903, x901, x879, x876);
   uint32_t x904;
   fiat_p384_uint1 x905;
-  fiat_p384_addcarryx_u32(&x904, &x905, x903, x874, x877);
+  fiat_p384_addcarryx_u32(&x904, &x905, x903, x877, x874);
   uint32_t x906;
   fiat_p384_uint1 x907;
-  fiat_p384_addcarryx_u32(&x906, &x907, x905, x872, x875);
+  fiat_p384_addcarryx_u32(&x906, &x907, x905, x875, x872);
   uint32_t x908;
   fiat_p384_uint1 x909;
-  fiat_p384_addcarryx_u32(&x908, &x909, 0x0, x890, x848);
+  fiat_p384_addcarryx_u32(&x908, &x909, 0x0, x848, x890);
   uint32_t x910;
   fiat_p384_uint1 x911;
-  fiat_p384_addcarryx_u32(&x910, &x911, x909, x891, x850);
+  fiat_p384_addcarryx_u32(&x910, &x911, x909, x850, x891);
   uint32_t x912;
   fiat_p384_uint1 x913;
-  fiat_p384_addcarryx_u32(&x912, &x913, x911, 0x0, x852);
+  fiat_p384_addcarryx_u32(&x912, &x913, x911, x852, 0x0);
   uint32_t x914;
   fiat_p384_uint1 x915;
-  fiat_p384_addcarryx_u32(&x914, &x915, x913, x888, x854);
+  fiat_p384_addcarryx_u32(&x914, &x915, x913, x854, x888);
   uint32_t x916;
   fiat_p384_uint1 x917;
-  fiat_p384_addcarryx_u32(&x916, &x917, x915, x892, x856);
+  fiat_p384_addcarryx_u32(&x916, &x917, x915, x856, x892);
   uint32_t x918;
   fiat_p384_uint1 x919;
-  fiat_p384_addcarryx_u32(&x918, &x919, x917, x894, x858);
+  fiat_p384_addcarryx_u32(&x918, &x919, x917, x858, x894);
   uint32_t x920;
   fiat_p384_uint1 x921;
-  fiat_p384_addcarryx_u32(&x920, &x921, x919, x896, x860);
+  fiat_p384_addcarryx_u32(&x920, &x921, x919, x860, x896);
   uint32_t x922;
   fiat_p384_uint1 x923;
-  fiat_p384_addcarryx_u32(&x922, &x923, x921, x898, x862);
+  fiat_p384_addcarryx_u32(&x922, &x923, x921, x862, x898);
   uint32_t x924;
   fiat_p384_uint1 x925;
-  fiat_p384_addcarryx_u32(&x924, &x925, x923, x900, x864);
+  fiat_p384_addcarryx_u32(&x924, &x925, x923, x864, x900);
   uint32_t x926;
   fiat_p384_uint1 x927;
-  fiat_p384_addcarryx_u32(&x926, &x927, x925, x902, x866);
+  fiat_p384_addcarryx_u32(&x926, &x927, x925, x866, x902);
   uint32_t x928;
   fiat_p384_uint1 x929;
-  fiat_p384_addcarryx_u32(&x928, &x929, x927, x904, x868);
+  fiat_p384_addcarryx_u32(&x928, &x929, x927, x868, x904);
   uint32_t x930;
   fiat_p384_uint1 x931;
-  fiat_p384_addcarryx_u32(&x930, &x931, x929, x906, x870);
+  fiat_p384_addcarryx_u32(&x930, &x931, x929, x870, x906);
   uint32_t x932;
   fiat_p384_uint1 x933;
-  fiat_p384_addcarryx_u32(&x932, &x933, x847, 0x0, 0x0);
+  fiat_p384_addcarryx_u32(&x932, &x933, x907, x873, 0x0);
   uint32_t x934;
   fiat_p384_uint1 x935;
-  fiat_p384_addcarryx_u32(&x934, &x935, x871, 0x0, (fiat_p384_uint1)x932);
+  fiat_p384_addcarryx_u32(&x934, &x935, x847, 0x0, 0x0);
   uint32_t x936;
   fiat_p384_uint1 x937;
-  fiat_p384_addcarryx_u32(&x936, &x937, x907, 0x0, x873);
+  fiat_p384_addcarryx_u32(&x936, &x937, x871, (fiat_p384_uint1)x934, 0x0);
   uint32_t x938;
   fiat_p384_uint1 x939;
-  fiat_p384_addcarryx_u32(&x938, &x939, x931, x936, x934);
+  fiat_p384_addcarryx_u32(&x938, &x939, x931, x936, x932);
   uint32_t x940;
   fiat_p384_uint1 x941;
-  fiat_p384_addcarryx_u32(&x940, &x941, 0x0, (arg1[11]), x910);
+  fiat_p384_addcarryx_u32(&x940, &x941, 0x0, x910, (arg1[11]));
   uint32_t x942;
   fiat_p384_uint1 x943;
-  fiat_p384_addcarryx_u32(&x942, &x943, x941, 0x0, x912);
+  fiat_p384_addcarryx_u32(&x942, &x943, x941, x912, 0x0);
   uint32_t x944;
   fiat_p384_uint1 x945;
-  fiat_p384_addcarryx_u32(&x944, &x945, x943, 0x0, x914);
+  fiat_p384_addcarryx_u32(&x944, &x945, x943, x914, 0x0);
   uint32_t x946;
   fiat_p384_uint1 x947;
-  fiat_p384_addcarryx_u32(&x946, &x947, x945, 0x0, x916);
+  fiat_p384_addcarryx_u32(&x946, &x947, x945, x916, 0x0);
   uint32_t x948;
   fiat_p384_uint1 x949;
-  fiat_p384_addcarryx_u32(&x948, &x949, x947, 0x0, x918);
+  fiat_p384_addcarryx_u32(&x948, &x949, x947, x918, 0x0);
   uint32_t x950;
   fiat_p384_uint1 x951;
-  fiat_p384_addcarryx_u32(&x950, &x951, x949, 0x0, x920);
+  fiat_p384_addcarryx_u32(&x950, &x951, x949, x920, 0x0);
   uint32_t x952;
   fiat_p384_uint1 x953;
-  fiat_p384_addcarryx_u32(&x952, &x953, x951, 0x0, x922);
+  fiat_p384_addcarryx_u32(&x952, &x953, x951, x922, 0x0);
   uint32_t x954;
   fiat_p384_uint1 x955;
-  fiat_p384_addcarryx_u32(&x954, &x955, x953, 0x0, x924);
+  fiat_p384_addcarryx_u32(&x954, &x955, x953, x924, 0x0);
   uint32_t x956;
   fiat_p384_uint1 x957;
-  fiat_p384_addcarryx_u32(&x956, &x957, x955, 0x0, x926);
+  fiat_p384_addcarryx_u32(&x956, &x957, x955, x926, 0x0);
   uint32_t x958;
   fiat_p384_uint1 x959;
-  fiat_p384_addcarryx_u32(&x958, &x959, x957, 0x0, x928);
+  fiat_p384_addcarryx_u32(&x958, &x959, x957, x928, 0x0);
   uint32_t x960;
   fiat_p384_uint1 x961;
-  fiat_p384_addcarryx_u32(&x960, &x961, x959, 0x0, x930);
+  fiat_p384_addcarryx_u32(&x960, &x961, x959, x930, 0x0);
   uint32_t x962;
   fiat_p384_uint1 x963;
-  fiat_p384_addcarryx_u32(&x962, &x963, x961, 0x0, x938);
+  fiat_p384_addcarryx_u32(&x962, &x963, x961, x938, 0x0);
   uint32_t x964;
   uint32_t x965;
   fiat_p384_mulx_u32(&x964, &x965, x940, UINT32_C(0xffffffff));
@@ -7095,76 +7095,76 @@ static void fiat_p384_from_montgomery(uint32_t out1[12], const uint32_t arg1[12]
   fiat_p384_mulx_u32(&x982, &x983, x940, UINT32_C(0xffffffff));
   uint32_t x984;
   fiat_p384_uint1 x985;
-  fiat_p384_addcarryx_u32(&x984, &x985, 0x0, x978, x981);
+  fiat_p384_addcarryx_u32(&x984, &x985, 0x0, x981, x978);
   uint32_t x986;
   fiat_p384_uint1 x987;
-  fiat_p384_addcarryx_u32(&x986, &x987, x985, x976, x979);
+  fiat_p384_addcarryx_u32(&x986, &x987, x985, x979, x976);
   uint32_t x988;
   fiat_p384_uint1 x989;
-  fiat_p384_addcarryx_u32(&x988, &x989, x987, x974, x977);
+  fiat_p384_addcarryx_u32(&x988, &x989, x987, x977, x974);
   uint32_t x990;
   fiat_p384_uint1 x991;
-  fiat_p384_addcarryx_u32(&x990, &x991, x989, x972, x975);
+  fiat_p384_addcarryx_u32(&x990, &x991, x989, x975, x972);
   uint32_t x992;
   fiat_p384_uint1 x993;
-  fiat_p384_addcarryx_u32(&x992, &x993, x991, x970, x973);
+  fiat_p384_addcarryx_u32(&x992, &x993, x991, x973, x970);
   uint32_t x994;
   fiat_p384_uint1 x995;
-  fiat_p384_addcarryx_u32(&x994, &x995, x993, x968, x971);
+  fiat_p384_addcarryx_u32(&x994, &x995, x993, x971, x968);
   uint32_t x996;
   fiat_p384_uint1 x997;
-  fiat_p384_addcarryx_u32(&x996, &x997, x995, x966, x969);
+  fiat_p384_addcarryx_u32(&x996, &x997, x995, x969, x966);
   uint32_t x998;
   fiat_p384_uint1 x999;
-  fiat_p384_addcarryx_u32(&x998, &x999, x997, x964, x967);
+  fiat_p384_addcarryx_u32(&x998, &x999, x997, x967, x964);
   uint32_t x1000;
   fiat_p384_uint1 x1001;
-  fiat_p384_addcarryx_u32(&x1000, &x1001, 0x0, x982, x940);
+  fiat_p384_addcarryx_u32(&x1000, &x1001, 0x0, x940, x982);
   uint32_t x1002;
   fiat_p384_uint1 x1003;
-  fiat_p384_addcarryx_u32(&x1002, &x1003, x1001, x983, x942);
+  fiat_p384_addcarryx_u32(&x1002, &x1003, x1001, x942, x983);
   uint32_t x1004;
   fiat_p384_uint1 x1005;
-  fiat_p384_addcarryx_u32(&x1004, &x1005, x1003, 0x0, x944);
+  fiat_p384_addcarryx_u32(&x1004, &x1005, x1003, x944, 0x0);
   uint32_t x1006;
   fiat_p384_uint1 x1007;
-  fiat_p384_addcarryx_u32(&x1006, &x1007, x1005, x980, x946);
+  fiat_p384_addcarryx_u32(&x1006, &x1007, x1005, x946, x980);
   uint32_t x1008;
   fiat_p384_uint1 x1009;
-  fiat_p384_addcarryx_u32(&x1008, &x1009, x1007, x984, x948);
+  fiat_p384_addcarryx_u32(&x1008, &x1009, x1007, x948, x984);
   uint32_t x1010;
   fiat_p384_uint1 x1011;
-  fiat_p384_addcarryx_u32(&x1010, &x1011, x1009, x986, x950);
+  fiat_p384_addcarryx_u32(&x1010, &x1011, x1009, x950, x986);
   uint32_t x1012;
   fiat_p384_uint1 x1013;
-  fiat_p384_addcarryx_u32(&x1012, &x1013, x1011, x988, x952);
+  fiat_p384_addcarryx_u32(&x1012, &x1013, x1011, x952, x988);
   uint32_t x1014;
   fiat_p384_uint1 x1015;
-  fiat_p384_addcarryx_u32(&x1014, &x1015, x1013, x990, x954);
+  fiat_p384_addcarryx_u32(&x1014, &x1015, x1013, x954, x990);
   uint32_t x1016;
   fiat_p384_uint1 x1017;
-  fiat_p384_addcarryx_u32(&x1016, &x1017, x1015, x992, x956);
+  fiat_p384_addcarryx_u32(&x1016, &x1017, x1015, x956, x992);
   uint32_t x1018;
   fiat_p384_uint1 x1019;
-  fiat_p384_addcarryx_u32(&x1018, &x1019, x1017, x994, x958);
+  fiat_p384_addcarryx_u32(&x1018, &x1019, x1017, x958, x994);
   uint32_t x1020;
   fiat_p384_uint1 x1021;
-  fiat_p384_addcarryx_u32(&x1020, &x1021, x1019, x996, x960);
+  fiat_p384_addcarryx_u32(&x1020, &x1021, x1019, x960, x996);
   uint32_t x1022;
   fiat_p384_uint1 x1023;
-  fiat_p384_addcarryx_u32(&x1022, &x1023, x1021, x998, x962);
+  fiat_p384_addcarryx_u32(&x1022, &x1023, x1021, x962, x998);
   uint32_t x1024;
   fiat_p384_uint1 x1025;
-  fiat_p384_addcarryx_u32(&x1024, &x1025, x939, 0x0, 0x0);
+  fiat_p384_addcarryx_u32(&x1024, &x1025, x999, x965, 0x0);
   uint32_t x1026;
   fiat_p384_uint1 x1027;
-  fiat_p384_addcarryx_u32(&x1026, &x1027, x963, 0x0, (fiat_p384_uint1)x1024);
+  fiat_p384_addcarryx_u32(&x1026, &x1027, x939, 0x0, 0x0);
   uint32_t x1028;
   fiat_p384_uint1 x1029;
-  fiat_p384_addcarryx_u32(&x1028, &x1029, x999, 0x0, x965);
+  fiat_p384_addcarryx_u32(&x1028, &x1029, x963, (fiat_p384_uint1)x1026, 0x0);
   uint32_t x1030;
   fiat_p384_uint1 x1031;
-  fiat_p384_addcarryx_u32(&x1030, &x1031, x1023, x1028, x1026);
+  fiat_p384_addcarryx_u32(&x1030, &x1031, x1023, x1028, x1024);
   uint32_t x1032;
   fiat_p384_uint1 x1033;
   fiat_p384_subborrowx_u32(&x1032, &x1033, 0x0, x1002, UINT32_C(0xffffffff));