diff options
author | Jason Gross <jgross@mit.edu> | 2017-11-02 02:20:52 -0400 |
---|---|---|
committer | Jason Gross <jgross@mit.edu> | 2017-11-02 02:20:52 -0400 |
commit | 37a8b2a0c14082298d77f0a4858ab3402d9d1c6c (patch) | |
tree | bd1b4c8c34485332ff15d58272efb38468ea72e5 /src/Specific/solinas64_2e243m9/femul.c | |
parent | 11144d2e698c4a263e59acb02226d383865f74e0 (diff) |
Update display logs and c files
Diffstat (limited to 'src/Specific/solinas64_2e243m9/femul.c')
-rw-r--r-- | src/Specific/solinas64_2e243m9/femul.c | 106 |
1 files changed, 50 insertions, 56 deletions
diff --git a/src/Specific/solinas64_2e243m9/femul.c b/src/Specific/solinas64_2e243m9/femul.c index 2a5ad5f77..1f3dd625a 100644 --- a/src/Specific/solinas64_2e243m9/femul.c +++ b/src/Specific/solinas64_2e243m9/femul.c @@ -1,56 +1,50 @@ -#include <stdint.h> -#include <stdbool.h> -#include <x86intrin.h> -#include "liblow.h" - -#include "femul.h" - -typedef unsigned int uint128_t __attribute__((mode(TI))); - -#if (defined(__GNUC__) || defined(__GNUG__)) && !(defined(__clang__)||defined(__INTEL_COMPILER)) -// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81294 -#define _subborrow_u32 __builtin_ia32_sbb_u32 -#define _subborrow_u64 __builtin_ia32_sbb_u64 -#endif - -#undef force_inline -#define force_inline __attribute__((always_inline)) - -void force_inline femul(uint64_t* out, uint64_t x12, uint64_t x13, uint64_t x11, uint64_t x9, uint64_t x7, uint64_t x5, uint64_t x22, uint64_t x23, uint64_t x21, uint64_t x19, uint64_t x17, uint64_t x15) -{ uint128_t x24 = (((uint128_t)x5 * x22) + (((uint128_t)x7 * x23) + (((uint128_t)x9 * x21) + (((uint128_t)x11 * x19) + (((uint128_t)x13 * x17) + ((uint128_t)x12 * x15)))))); -{ uint128_t x25 = ((((uint128_t)x5 * x23) + ((0x2 * ((uint128_t)x7 * x21)) + (((uint128_t)x9 * x19) + ((0x2 * ((uint128_t)x11 * x17)) + ((uint128_t)x13 * x15))))) + (0x9 * (0x2 * ((uint128_t)x12 * x22)))); -{ uint128_t x26 = ((((uint128_t)x5 * x21) + (((uint128_t)x7 * x19) + (((uint128_t)x9 * x17) + ((uint128_t)x11 * x15)))) + (0x9 * (((uint128_t)x13 * x22) + ((uint128_t)x12 * x23)))); -{ uint128_t x27 = ((((uint128_t)x5 * x19) + ((0x2 * ((uint128_t)x7 * x17)) + ((uint128_t)x9 * x15))) + (0x9 * ((0x2 * ((uint128_t)x11 * x22)) + (((uint128_t)x13 * x23) + (0x2 * ((uint128_t)x12 * x21)))))); -{ uint128_t x28 = ((((uint128_t)x5 * x17) + ((uint128_t)x7 * x15)) + (0x9 * (((uint128_t)x9 * x22) + (((uint128_t)x11 * x23) + (((uint128_t)x13 * x21) + ((uint128_t)x12 * x19)))))); -{ uint128_t x29 = (((uint128_t)x5 * x15) + (0x9 * ((0x2 * ((uint128_t)x7 * x22)) + (((uint128_t)x9 * x23) + ((0x2 * ((uint128_t)x11 * x21)) + (((uint128_t)x13 * x19) + (0x2 * ((uint128_t)x12 * x17)))))))); -{ uint64_t x30 = (uint64_t) (x29 >> 0x29); -{ uint64_t x31 = ((uint64_t)x29 & 0x1ffffffffff); -{ uint128_t x32 = (x30 + x28); -{ uint64_t x33 = (uint64_t) (x32 >> 0x28); -{ uint64_t x34 = ((uint64_t)x32 & 0xffffffffff); -{ uint128_t x35 = (x33 + x27); -{ uint64_t x36 = (uint64_t) (x35 >> 0x29); -{ uint64_t x37 = ((uint64_t)x35 & 0x1ffffffffff); -{ uint128_t x38 = (x36 + x26); -{ uint64_t x39 = (uint64_t) (x38 >> 0x28); -{ uint64_t x40 = ((uint64_t)x38 & 0xffffffffff); -{ uint128_t x41 = (x39 + x25); -{ uint64_t x42 = (uint64_t) (x41 >> 0x29); -{ uint64_t x43 = ((uint64_t)x41 & 0x1ffffffffff); -{ uint128_t x44 = (x42 + x24); -{ uint64_t x45 = (uint64_t) (x44 >> 0x28); -{ uint64_t x46 = ((uint64_t)x44 & 0xffffffffff); -{ uint64_t x47 = (x31 + (0x9 * x45)); -{ uint64_t x48 = (x47 >> 0x29); -{ uint64_t x49 = (x47 & 0x1ffffffffff); -{ uint64_t x50 = (x48 + x34); -{ uint64_t x51 = (x50 >> 0x28); -{ uint64_t x52 = (x50 & 0xffffffffff); -out[0] = x46; -out[1] = x43; -out[2] = x40; -out[3] = x51 + x37; -out[4] = x52; -out[5] = x49; -}}}}}}}}}}}}}}}}}}}}}}}}}}}}} -// caller: uint64_t out[6]; +static void femul(uint64_t out[6], const uint64_t in1[6], const uint64_t in2[6]) { + { const uint64_t x12 = in1[5]; + { const uint64_t x13 = in1[4]; + { const uint64_t x11 = in1[3]; + { const uint64_t x9 = in1[2]; + { const uint64_t x7 = in1[1]; + { const uint64_t x5 = in1[0]; + { const uint64_t x22 = in2[5]; + { const uint64_t x23 = in2[4]; + { const uint64_t x21 = in2[3]; + { const uint64_t x19 = in2[2]; + { const uint64_t x17 = in2[1]; + { const uint64_t x15 = in2[0]; + { uint128_t x24 = (((uint128_t)x5 * x22) + (((uint128_t)x7 * x23) + (((uint128_t)x9 * x21) + (((uint128_t)x11 * x19) + (((uint128_t)x13 * x17) + ((uint128_t)x12 * x15)))))); + { uint128_t x25 = ((((uint128_t)x5 * x23) + ((0x2 * ((uint128_t)x7 * x21)) + (((uint128_t)x9 * x19) + ((0x2 * ((uint128_t)x11 * x17)) + ((uint128_t)x13 * x15))))) + (0x9 * (0x2 * ((uint128_t)x12 * x22)))); + { uint128_t x26 = ((((uint128_t)x5 * x21) + (((uint128_t)x7 * x19) + (((uint128_t)x9 * x17) + ((uint128_t)x11 * x15)))) + (0x9 * (((uint128_t)x13 * x22) + ((uint128_t)x12 * x23)))); + { uint128_t x27 = ((((uint128_t)x5 * x19) + ((0x2 * ((uint128_t)x7 * x17)) + ((uint128_t)x9 * x15))) + (0x9 * ((0x2 * ((uint128_t)x11 * x22)) + (((uint128_t)x13 * x23) + (0x2 * ((uint128_t)x12 * x21)))))); + { uint128_t x28 = ((((uint128_t)x5 * x17) + ((uint128_t)x7 * x15)) + (0x9 * (((uint128_t)x9 * x22) + (((uint128_t)x11 * x23) + (((uint128_t)x13 * x21) + ((uint128_t)x12 * x19)))))); + { uint128_t x29 = (((uint128_t)x5 * x15) + (0x9 * ((0x2 * ((uint128_t)x7 * x22)) + (((uint128_t)x9 * x23) + ((0x2 * ((uint128_t)x11 * x21)) + (((uint128_t)x13 * x19) + (0x2 * ((uint128_t)x12 * x17)))))))); + { uint64_t x30 = (uint64_t) (x29 >> 0x29); + { uint64_t x31 = ((uint64_t)x29 & 0x1ffffffffff); + { uint128_t x32 = (x30 + x28); + { uint64_t x33 = (uint64_t) (x32 >> 0x28); + { uint64_t x34 = ((uint64_t)x32 & 0xffffffffff); + { uint128_t x35 = (x33 + x27); + { uint64_t x36 = (uint64_t) (x35 >> 0x29); + { uint64_t x37 = ((uint64_t)x35 & 0x1ffffffffff); + { uint128_t x38 = (x36 + x26); + { uint64_t x39 = (uint64_t) (x38 >> 0x28); + { uint64_t x40 = ((uint64_t)x38 & 0xffffffffff); + { uint128_t x41 = (x39 + x25); + { uint64_t x42 = (uint64_t) (x41 >> 0x29); + { uint64_t x43 = ((uint64_t)x41 & 0x1ffffffffff); + { uint128_t x44 = (x42 + x24); + { uint64_t x45 = (uint64_t) (x44 >> 0x28); + { uint64_t x46 = ((uint64_t)x44 & 0xffffffffff); + { uint64_t x47 = (x31 + (0x9 * x45)); + { uint64_t x48 = (x47 >> 0x29); + { uint64_t x49 = (x47 & 0x1ffffffffff); + { uint64_t x50 = (x48 + x34); + { uint64_t x51 = (x50 >> 0x28); + { uint64_t x52 = (x50 & 0xffffffffff); + out[0] = x49; + out[1] = x52; + out[2] = (x51 + x37); + out[3] = x40; + out[4] = x43; + out[5] = x46; + }}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}} +} |