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_2e512m569 | |
parent | 11144d2e698c4a263e59acb02226d383865f74e0 (diff) |
Update display logs and c files
Diffstat (limited to 'src/Specific/solinas64_2e512m569')
-rw-r--r-- | src/Specific/solinas64_2e512m569/femul.c | 154 | ||||
-rw-r--r-- | src/Specific/solinas64_2e512m569/fesquare.c | 144 | ||||
-rw-r--r-- | src/Specific/solinas64_2e512m569/freeze.c | 77 |
3 files changed, 199 insertions, 176 deletions
diff --git a/src/Specific/solinas64_2e512m569/femul.c b/src/Specific/solinas64_2e512m569/femul.c index e4b8904f7..46e0f261b 100644 --- a/src/Specific/solinas64_2e512m569/femul.c +++ b/src/Specific/solinas64_2e512m569/femul.c @@ -1,76 +1,78 @@ -#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 x20, uint64_t x21, uint64_t x19, uint64_t x17, uint64_t x15, uint64_t x13, uint64_t x11, uint64_t x9, uint64_t x7, uint64_t x5, uint64_t x38, uint64_t x39, uint64_t x37, uint64_t x35, uint64_t x33, uint64_t x31, uint64_t x29, uint64_t x27, uint64_t x25, uint64_t x23) -{ uint128_t x40 = (((uint128_t)x5 * x38) + ((0x2 * ((uint128_t)x7 * x39)) + ((0x2 * ((uint128_t)x9 * x37)) + ((0x2 * ((uint128_t)x11 * x35)) + (((uint128_t)x13 * x33) + (((uint128_t)x15 * x31) + ((0x2 * ((uint128_t)x17 * x29)) + ((0x2 * ((uint128_t)x19 * x27)) + ((0x2 * ((uint128_t)x21 * x25)) + ((uint128_t)x20 * x23)))))))))); -{ uint128_t x41 = ((((uint128_t)x5 * x39) + ((0x2 * ((uint128_t)x7 * x37)) + ((0x2 * ((uint128_t)x9 * x35)) + (((uint128_t)x11 * x33) + (((uint128_t)x13 * x31) + (((uint128_t)x15 * x29) + ((0x2 * ((uint128_t)x17 * x27)) + ((0x2 * ((uint128_t)x19 * x25)) + ((uint128_t)x21 * x23))))))))) + (0x239 * ((uint128_t)x20 * x38))); -{ uint128_t x42 = ((((uint128_t)x5 * x37) + ((0x2 * ((uint128_t)x7 * x35)) + (((uint128_t)x9 * x33) + (((uint128_t)x11 * x31) + (((uint128_t)x13 * x29) + (((uint128_t)x15 * x27) + ((0x2 * ((uint128_t)x17 * x25)) + ((uint128_t)x19 * x23)))))))) + (0x239 * (((uint128_t)x21 * x38) + ((uint128_t)x20 * x39)))); -{ uint128_t x43 = ((((uint128_t)x5 * x35) + (((uint128_t)x7 * x33) + (((uint128_t)x9 * x31) + (((uint128_t)x11 * x29) + (((uint128_t)x13 * x27) + (((uint128_t)x15 * x25) + ((uint128_t)x17 * x23))))))) + (0x239 * (((uint128_t)x19 * x38) + (((uint128_t)x21 * x39) + ((uint128_t)x20 * x37))))); -{ uint128_t x44 = ((((uint128_t)x5 * x33) + ((0x2 * ((uint128_t)x7 * x31)) + ((0x2 * ((uint128_t)x9 * x29)) + ((0x2 * ((uint128_t)x11 * x27)) + ((0x2 * ((uint128_t)x13 * x25)) + ((uint128_t)x15 * x23)))))) + (0x239 * ((0x2 * ((uint128_t)x17 * x38)) + ((0x2 * ((uint128_t)x19 * x39)) + ((0x2 * ((uint128_t)x21 * x37)) + (0x2 * ((uint128_t)x20 * x35))))))); -{ uint128_t x45 = ((((uint128_t)x5 * x31) + ((0x2 * ((uint128_t)x7 * x29)) + ((0x2 * ((uint128_t)x9 * x27)) + ((0x2 * ((uint128_t)x11 * x25)) + ((uint128_t)x13 * x23))))) + (0x239 * (((uint128_t)x15 * x38) + ((0x2 * ((uint128_t)x17 * x39)) + ((0x2 * ((uint128_t)x19 * x37)) + ((0x2 * ((uint128_t)x21 * x35)) + ((uint128_t)x20 * x33))))))); -{ uint128_t x46 = ((((uint128_t)x5 * x29) + ((0x2 * ((uint128_t)x7 * x27)) + ((0x2 * ((uint128_t)x9 * x25)) + ((uint128_t)x11 * x23)))) + (0x239 * (((uint128_t)x13 * x38) + (((uint128_t)x15 * x39) + ((0x2 * ((uint128_t)x17 * x37)) + ((0x2 * ((uint128_t)x19 * x35)) + (((uint128_t)x21 * x33) + ((uint128_t)x20 * x31)))))))); -{ uint128_t x47 = ((((uint128_t)x5 * x27) + ((0x2 * ((uint128_t)x7 * x25)) + ((uint128_t)x9 * x23))) + (0x239 * (((uint128_t)x11 * x38) + (((uint128_t)x13 * x39) + (((uint128_t)x15 * x37) + ((0x2 * ((uint128_t)x17 * x35)) + (((uint128_t)x19 * x33) + (((uint128_t)x21 * x31) + ((uint128_t)x20 * x29))))))))); -{ uint128_t x48 = ((((uint128_t)x5 * x25) + ((uint128_t)x7 * x23)) + (0x239 * (((uint128_t)x9 * x38) + (((uint128_t)x11 * x39) + (((uint128_t)x13 * x37) + (((uint128_t)x15 * x35) + (((uint128_t)x17 * x33) + (((uint128_t)x19 * x31) + (((uint128_t)x21 * x29) + ((uint128_t)x20 * x27)))))))))); -{ uint128_t x49 = (((uint128_t)x5 * x23) + (0x239 * ((0x2 * ((uint128_t)x7 * x38)) + ((0x2 * ((uint128_t)x9 * x39)) + ((0x2 * ((uint128_t)x11 * x37)) + ((0x2 * ((uint128_t)x13 * x35)) + (((uint128_t)x15 * x33) + ((0x2 * ((uint128_t)x17 * x31)) + ((0x2 * ((uint128_t)x19 * x29)) + ((0x2 * ((uint128_t)x21 * x27)) + (0x2 * ((uint128_t)x20 * x25)))))))))))); -{ uint128_t x50 = (x49 >> 0x34); -{ uint64_t x51 = ((uint64_t)x49 & 0xfffffffffffff); -{ uint128_t x52 = (x50 + x48); -{ uint128_t x53 = (x52 >> 0x33); -{ uint64_t x54 = ((uint64_t)x52 & 0x7ffffffffffff); -{ uint128_t x55 = (x53 + x47); -{ uint128_t x56 = (x55 >> 0x33); -{ uint64_t x57 = ((uint64_t)x55 & 0x7ffffffffffff); -{ uint128_t x58 = (x56 + x46); -{ uint128_t x59 = (x58 >> 0x33); -{ uint64_t x60 = ((uint64_t)x58 & 0x7ffffffffffff); -{ uint128_t x61 = (x59 + x45); -{ uint128_t x62 = (x61 >> 0x33); -{ uint64_t x63 = ((uint64_t)x61 & 0x7ffffffffffff); -{ uint128_t x64 = (x62 + x44); -{ uint128_t x65 = (x64 >> 0x34); -{ uint64_t x66 = ((uint64_t)x64 & 0xfffffffffffff); -{ uint128_t x67 = (x65 + x43); -{ uint128_t x68 = (x67 >> 0x33); -{ uint64_t x69 = ((uint64_t)x67 & 0x7ffffffffffff); -{ uint128_t x70 = (x68 + x42); -{ uint128_t x71 = (x70 >> 0x33); -{ uint64_t x72 = ((uint64_t)x70 & 0x7ffffffffffff); -{ uint128_t x73 = (x71 + x41); -{ uint64_t x74 = (uint64_t) (x73 >> 0x33); -{ uint64_t x75 = ((uint64_t)x73 & 0x7ffffffffffff); -{ uint128_t x76 = (x74 + x40); -{ uint64_t x77 = (uint64_t) (x76 >> 0x33); -{ uint64_t x78 = ((uint64_t)x76 & 0x7ffffffffffff); -{ uint128_t x79 = (x51 + ((uint128_t)0x239 * x77)); -{ uint64_t x80 = (uint64_t) (x79 >> 0x34); -{ uint64_t x81 = ((uint64_t)x79 & 0xfffffffffffff); -{ uint64_t x82 = (x80 + x54); -{ uint64_t x83 = (x82 >> 0x33); -{ uint64_t x84 = (x82 & 0x7ffffffffffff); -out[0] = x78; -out[1] = x75; -out[2] = x72; -out[3] = x69; -out[4] = x66; -out[5] = x63; -out[6] = x60; -out[7] = x83 + x57; -out[8] = x84; -out[9] = x81; -}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}} -// caller: uint64_t out[10]; +static void femul(uint64_t out[10], const uint64_t in1[10], const uint64_t in2[10]) { + { const uint64_t x20 = in1[9]; + { const uint64_t x21 = in1[8]; + { const uint64_t x19 = in1[7]; + { const uint64_t x17 = in1[6]; + { const uint64_t x15 = 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 x38 = in2[9]; + { const uint64_t x39 = in2[8]; + { const uint64_t x37 = in2[7]; + { const uint64_t x35 = in2[6]; + { const uint64_t x33 = in2[5]; + { const uint64_t x31 = in2[4]; + { const uint64_t x29 = in2[3]; + { const uint64_t x27 = in2[2]; + { const uint64_t x25 = in2[1]; + { const uint64_t x23 = in2[0]; + { uint128_t x40 = (((uint128_t)x5 * x38) + ((0x2 * ((uint128_t)x7 * x39)) + ((0x2 * ((uint128_t)x9 * x37)) + ((0x2 * ((uint128_t)x11 * x35)) + (((uint128_t)x13 * x33) + (((uint128_t)x15 * x31) + ((0x2 * ((uint128_t)x17 * x29)) + ((0x2 * ((uint128_t)x19 * x27)) + ((0x2 * ((uint128_t)x21 * x25)) + ((uint128_t)x20 * x23)))))))))); + { uint128_t x41 = ((((uint128_t)x5 * x39) + ((0x2 * ((uint128_t)x7 * x37)) + ((0x2 * ((uint128_t)x9 * x35)) + (((uint128_t)x11 * x33) + (((uint128_t)x13 * x31) + (((uint128_t)x15 * x29) + ((0x2 * ((uint128_t)x17 * x27)) + ((0x2 * ((uint128_t)x19 * x25)) + ((uint128_t)x21 * x23))))))))) + (0x239 * ((uint128_t)x20 * x38))); + { uint128_t x42 = ((((uint128_t)x5 * x37) + ((0x2 * ((uint128_t)x7 * x35)) + (((uint128_t)x9 * x33) + (((uint128_t)x11 * x31) + (((uint128_t)x13 * x29) + (((uint128_t)x15 * x27) + ((0x2 * ((uint128_t)x17 * x25)) + ((uint128_t)x19 * x23)))))))) + (0x239 * (((uint128_t)x21 * x38) + ((uint128_t)x20 * x39)))); + { uint128_t x43 = ((((uint128_t)x5 * x35) + (((uint128_t)x7 * x33) + (((uint128_t)x9 * x31) + (((uint128_t)x11 * x29) + (((uint128_t)x13 * x27) + (((uint128_t)x15 * x25) + ((uint128_t)x17 * x23))))))) + (0x239 * (((uint128_t)x19 * x38) + (((uint128_t)x21 * x39) + ((uint128_t)x20 * x37))))); + { uint128_t x44 = ((((uint128_t)x5 * x33) + ((0x2 * ((uint128_t)x7 * x31)) + ((0x2 * ((uint128_t)x9 * x29)) + ((0x2 * ((uint128_t)x11 * x27)) + ((0x2 * ((uint128_t)x13 * x25)) + ((uint128_t)x15 * x23)))))) + (0x239 * ((0x2 * ((uint128_t)x17 * x38)) + ((0x2 * ((uint128_t)x19 * x39)) + ((0x2 * ((uint128_t)x21 * x37)) + (0x2 * ((uint128_t)x20 * x35))))))); + { uint128_t x45 = ((((uint128_t)x5 * x31) + ((0x2 * ((uint128_t)x7 * x29)) + ((0x2 * ((uint128_t)x9 * x27)) + ((0x2 * ((uint128_t)x11 * x25)) + ((uint128_t)x13 * x23))))) + (0x239 * (((uint128_t)x15 * x38) + ((0x2 * ((uint128_t)x17 * x39)) + ((0x2 * ((uint128_t)x19 * x37)) + ((0x2 * ((uint128_t)x21 * x35)) + ((uint128_t)x20 * x33))))))); + { uint128_t x46 = ((((uint128_t)x5 * x29) + ((0x2 * ((uint128_t)x7 * x27)) + ((0x2 * ((uint128_t)x9 * x25)) + ((uint128_t)x11 * x23)))) + (0x239 * (((uint128_t)x13 * x38) + (((uint128_t)x15 * x39) + ((0x2 * ((uint128_t)x17 * x37)) + ((0x2 * ((uint128_t)x19 * x35)) + (((uint128_t)x21 * x33) + ((uint128_t)x20 * x31)))))))); + { uint128_t x47 = ((((uint128_t)x5 * x27) + ((0x2 * ((uint128_t)x7 * x25)) + ((uint128_t)x9 * x23))) + (0x239 * (((uint128_t)x11 * x38) + (((uint128_t)x13 * x39) + (((uint128_t)x15 * x37) + ((0x2 * ((uint128_t)x17 * x35)) + (((uint128_t)x19 * x33) + (((uint128_t)x21 * x31) + ((uint128_t)x20 * x29))))))))); + { uint128_t x48 = ((((uint128_t)x5 * x25) + ((uint128_t)x7 * x23)) + (0x239 * (((uint128_t)x9 * x38) + (((uint128_t)x11 * x39) + (((uint128_t)x13 * x37) + (((uint128_t)x15 * x35) + (((uint128_t)x17 * x33) + (((uint128_t)x19 * x31) + (((uint128_t)x21 * x29) + ((uint128_t)x20 * x27)))))))))); + { uint128_t x49 = (((uint128_t)x5 * x23) + (0x239 * ((0x2 * ((uint128_t)x7 * x38)) + ((0x2 * ((uint128_t)x9 * x39)) + ((0x2 * ((uint128_t)x11 * x37)) + ((0x2 * ((uint128_t)x13 * x35)) + (((uint128_t)x15 * x33) + ((0x2 * ((uint128_t)x17 * x31)) + ((0x2 * ((uint128_t)x19 * x29)) + ((0x2 * ((uint128_t)x21 * x27)) + (0x2 * ((uint128_t)x20 * x25)))))))))))); + { uint128_t x50 = (x49 >> 0x34); + { uint64_t x51 = ((uint64_t)x49 & 0xfffffffffffff); + { uint128_t x52 = (x50 + x48); + { uint128_t x53 = (x52 >> 0x33); + { uint64_t x54 = ((uint64_t)x52 & 0x7ffffffffffff); + { uint128_t x55 = (x53 + x47); + { uint128_t x56 = (x55 >> 0x33); + { uint64_t x57 = ((uint64_t)x55 & 0x7ffffffffffff); + { uint128_t x58 = (x56 + x46); + { uint128_t x59 = (x58 >> 0x33); + { uint64_t x60 = ((uint64_t)x58 & 0x7ffffffffffff); + { uint128_t x61 = (x59 + x45); + { uint128_t x62 = (x61 >> 0x33); + { uint64_t x63 = ((uint64_t)x61 & 0x7ffffffffffff); + { uint128_t x64 = (x62 + x44); + { uint128_t x65 = (x64 >> 0x34); + { uint64_t x66 = ((uint64_t)x64 & 0xfffffffffffff); + { uint128_t x67 = (x65 + x43); + { uint128_t x68 = (x67 >> 0x33); + { uint64_t x69 = ((uint64_t)x67 & 0x7ffffffffffff); + { uint128_t x70 = (x68 + x42); + { uint128_t x71 = (x70 >> 0x33); + { uint64_t x72 = ((uint64_t)x70 & 0x7ffffffffffff); + { uint128_t x73 = (x71 + x41); + { uint64_t x74 = (uint64_t) (x73 >> 0x33); + { uint64_t x75 = ((uint64_t)x73 & 0x7ffffffffffff); + { uint128_t x76 = (x74 + x40); + { uint64_t x77 = (uint64_t) (x76 >> 0x33); + { uint64_t x78 = ((uint64_t)x76 & 0x7ffffffffffff); + { uint128_t x79 = (x51 + ((uint128_t)0x239 * x77)); + { uint64_t x80 = (uint64_t) (x79 >> 0x34); + { uint64_t x81 = ((uint64_t)x79 & 0xfffffffffffff); + { uint64_t x82 = (x80 + x54); + { uint64_t x83 = (x82 >> 0x33); + { uint64_t x84 = (x82 & 0x7ffffffffffff); + out[0] = x81; + out[1] = x84; + out[2] = (x83 + x57); + out[3] = x60; + out[4] = x63; + out[5] = x66; + out[6] = x69; + out[7] = x72; + out[8] = x75; + out[9] = x78; + }}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}} +} diff --git a/src/Specific/solinas64_2e512m569/fesquare.c b/src/Specific/solinas64_2e512m569/fesquare.c index 61f6a5c8c..c8ecbdfda 100644 --- a/src/Specific/solinas64_2e512m569/fesquare.c +++ b/src/Specific/solinas64_2e512m569/fesquare.c @@ -1,76 +1,68 @@ -#include <stdint.h> -#include <stdbool.h> -#include <x86intrin.h> -#include "liblow.h" - -#include "fesquare.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 fesquare(uint64_t* out, uint64_t x17, uint64_t x18, uint64_t x16, uint64_t x14, uint64_t x12, uint64_t x10, uint64_t x8, uint64_t x6, uint64_t x4, uint64_t x2) -{ uint128_t x19 = (((uint128_t)x2 * x17) + ((0x2 * ((uint128_t)x4 * x18)) + ((0x2 * ((uint128_t)x6 * x16)) + ((0x2 * ((uint128_t)x8 * x14)) + (((uint128_t)x10 * x12) + (((uint128_t)x12 * x10) + ((0x2 * ((uint128_t)x14 * x8)) + ((0x2 * ((uint128_t)x16 * x6)) + ((0x2 * ((uint128_t)x18 * x4)) + ((uint128_t)x17 * x2)))))))))); -{ uint128_t x20 = ((((uint128_t)x2 * x18) + ((0x2 * ((uint128_t)x4 * x16)) + ((0x2 * ((uint128_t)x6 * x14)) + (((uint128_t)x8 * x12) + (((uint128_t)x10 * x10) + (((uint128_t)x12 * x8) + ((0x2 * ((uint128_t)x14 * x6)) + ((0x2 * ((uint128_t)x16 * x4)) + ((uint128_t)x18 * x2))))))))) + (0x239 * ((uint128_t)x17 * x17))); -{ uint128_t x21 = ((((uint128_t)x2 * x16) + ((0x2 * ((uint128_t)x4 * x14)) + (((uint128_t)x6 * x12) + (((uint128_t)x8 * x10) + (((uint128_t)x10 * x8) + (((uint128_t)x12 * x6) + ((0x2 * ((uint128_t)x14 * x4)) + ((uint128_t)x16 * x2)))))))) + (0x239 * (((uint128_t)x18 * x17) + ((uint128_t)x17 * x18)))); -{ uint128_t x22 = ((((uint128_t)x2 * x14) + (((uint128_t)x4 * x12) + (((uint128_t)x6 * x10) + (((uint128_t)x8 * x8) + (((uint128_t)x10 * x6) + (((uint128_t)x12 * x4) + ((uint128_t)x14 * x2))))))) + (0x239 * (((uint128_t)x16 * x17) + (((uint128_t)x18 * x18) + ((uint128_t)x17 * x16))))); -{ uint128_t x23 = ((((uint128_t)x2 * x12) + ((0x2 * ((uint128_t)x4 * x10)) + ((0x2 * ((uint128_t)x6 * x8)) + ((0x2 * ((uint128_t)x8 * x6)) + ((0x2 * ((uint128_t)x10 * x4)) + ((uint128_t)x12 * x2)))))) + (0x239 * ((0x2 * ((uint128_t)x14 * x17)) + ((0x2 * ((uint128_t)x16 * x18)) + ((0x2 * ((uint128_t)x18 * x16)) + (0x2 * ((uint128_t)x17 * x14))))))); -{ uint128_t x24 = ((((uint128_t)x2 * x10) + ((0x2 * ((uint128_t)x4 * x8)) + ((0x2 * ((uint128_t)x6 * x6)) + ((0x2 * ((uint128_t)x8 * x4)) + ((uint128_t)x10 * x2))))) + (0x239 * (((uint128_t)x12 * x17) + ((0x2 * ((uint128_t)x14 * x18)) + ((0x2 * ((uint128_t)x16 * x16)) + ((0x2 * ((uint128_t)x18 * x14)) + ((uint128_t)x17 * x12))))))); -{ uint128_t x25 = ((((uint128_t)x2 * x8) + ((0x2 * ((uint128_t)x4 * x6)) + ((0x2 * ((uint128_t)x6 * x4)) + ((uint128_t)x8 * x2)))) + (0x239 * (((uint128_t)x10 * x17) + (((uint128_t)x12 * x18) + ((0x2 * ((uint128_t)x14 * x16)) + ((0x2 * ((uint128_t)x16 * x14)) + (((uint128_t)x18 * x12) + ((uint128_t)x17 * x10)))))))); -{ uint128_t x26 = ((((uint128_t)x2 * x6) + ((0x2 * ((uint128_t)x4 * x4)) + ((uint128_t)x6 * x2))) + (0x239 * (((uint128_t)x8 * x17) + (((uint128_t)x10 * x18) + (((uint128_t)x12 * x16) + ((0x2 * ((uint128_t)x14 * x14)) + (((uint128_t)x16 * x12) + (((uint128_t)x18 * x10) + ((uint128_t)x17 * x8))))))))); -{ uint128_t x27 = ((((uint128_t)x2 * x4) + ((uint128_t)x4 * x2)) + (0x239 * (((uint128_t)x6 * x17) + (((uint128_t)x8 * x18) + (((uint128_t)x10 * x16) + (((uint128_t)x12 * x14) + (((uint128_t)x14 * x12) + (((uint128_t)x16 * x10) + (((uint128_t)x18 * x8) + ((uint128_t)x17 * x6)))))))))); -{ uint128_t x28 = (((uint128_t)x2 * x2) + (0x239 * ((0x2 * ((uint128_t)x4 * x17)) + ((0x2 * ((uint128_t)x6 * x18)) + ((0x2 * ((uint128_t)x8 * x16)) + ((0x2 * ((uint128_t)x10 * x14)) + (((uint128_t)x12 * x12) + ((0x2 * ((uint128_t)x14 * x10)) + ((0x2 * ((uint128_t)x16 * x8)) + ((0x2 * ((uint128_t)x18 * x6)) + (0x2 * ((uint128_t)x17 * x4)))))))))))); -{ uint128_t x29 = (x28 >> 0x34); -{ uint64_t x30 = ((uint64_t)x28 & 0xfffffffffffff); -{ uint128_t x31 = (x29 + x27); -{ uint128_t x32 = (x31 >> 0x33); -{ uint64_t x33 = ((uint64_t)x31 & 0x7ffffffffffff); -{ uint128_t x34 = (x32 + x26); -{ uint128_t x35 = (x34 >> 0x33); -{ uint64_t x36 = ((uint64_t)x34 & 0x7ffffffffffff); -{ uint128_t x37 = (x35 + x25); -{ uint128_t x38 = (x37 >> 0x33); -{ uint64_t x39 = ((uint64_t)x37 & 0x7ffffffffffff); -{ uint128_t x40 = (x38 + x24); -{ uint128_t x41 = (x40 >> 0x33); -{ uint64_t x42 = ((uint64_t)x40 & 0x7ffffffffffff); -{ uint128_t x43 = (x41 + x23); -{ uint128_t x44 = (x43 >> 0x34); -{ uint64_t x45 = ((uint64_t)x43 & 0xfffffffffffff); -{ uint128_t x46 = (x44 + x22); -{ uint128_t x47 = (x46 >> 0x33); -{ uint64_t x48 = ((uint64_t)x46 & 0x7ffffffffffff); -{ uint128_t x49 = (x47 + x21); -{ uint128_t x50 = (x49 >> 0x33); -{ uint64_t x51 = ((uint64_t)x49 & 0x7ffffffffffff); -{ uint128_t x52 = (x50 + x20); -{ uint64_t x53 = (uint64_t) (x52 >> 0x33); -{ uint64_t x54 = ((uint64_t)x52 & 0x7ffffffffffff); -{ uint128_t x55 = (x53 + x19); -{ uint64_t x56 = (uint64_t) (x55 >> 0x33); -{ uint64_t x57 = ((uint64_t)x55 & 0x7ffffffffffff); -{ uint128_t x58 = (x30 + ((uint128_t)0x239 * x56)); -{ uint64_t x59 = (uint64_t) (x58 >> 0x34); -{ uint64_t x60 = ((uint64_t)x58 & 0xfffffffffffff); -{ uint64_t x61 = (x59 + x33); -{ uint64_t x62 = (x61 >> 0x33); -{ uint64_t x63 = (x61 & 0x7ffffffffffff); -out[0] = x57; -out[1] = x54; -out[2] = x51; -out[3] = x48; -out[4] = x45; -out[5] = x42; -out[6] = x39; -out[7] = x62 + x36; -out[8] = x63; -out[9] = x60; -}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}} -// caller: uint64_t out[10]; +static void fesquare(uint64_t out[10], const uint64_t in1[10]) { + { const uint64_t x17 = in1[9]; + { const uint64_t x18 = in1[8]; + { const uint64_t x16 = in1[7]; + { const uint64_t x14 = in1[6]; + { const uint64_t x12 = in1[5]; + { const uint64_t x10 = in1[4]; + { const uint64_t x8 = in1[3]; + { const uint64_t x6 = in1[2]; + { const uint64_t x4 = in1[1]; + { const uint64_t x2 = in1[0]; + { uint128_t x19 = (((uint128_t)x2 * x17) + ((0x2 * ((uint128_t)x4 * x18)) + ((0x2 * ((uint128_t)x6 * x16)) + ((0x2 * ((uint128_t)x8 * x14)) + (((uint128_t)x10 * x12) + (((uint128_t)x12 * x10) + ((0x2 * ((uint128_t)x14 * x8)) + ((0x2 * ((uint128_t)x16 * x6)) + ((0x2 * ((uint128_t)x18 * x4)) + ((uint128_t)x17 * x2)))))))))); + { uint128_t x20 = ((((uint128_t)x2 * x18) + ((0x2 * ((uint128_t)x4 * x16)) + ((0x2 * ((uint128_t)x6 * x14)) + (((uint128_t)x8 * x12) + (((uint128_t)x10 * x10) + (((uint128_t)x12 * x8) + ((0x2 * ((uint128_t)x14 * x6)) + ((0x2 * ((uint128_t)x16 * x4)) + ((uint128_t)x18 * x2))))))))) + (0x239 * ((uint128_t)x17 * x17))); + { uint128_t x21 = ((((uint128_t)x2 * x16) + ((0x2 * ((uint128_t)x4 * x14)) + (((uint128_t)x6 * x12) + (((uint128_t)x8 * x10) + (((uint128_t)x10 * x8) + (((uint128_t)x12 * x6) + ((0x2 * ((uint128_t)x14 * x4)) + ((uint128_t)x16 * x2)))))))) + (0x239 * (((uint128_t)x18 * x17) + ((uint128_t)x17 * x18)))); + { uint128_t x22 = ((((uint128_t)x2 * x14) + (((uint128_t)x4 * x12) + (((uint128_t)x6 * x10) + (((uint128_t)x8 * x8) + (((uint128_t)x10 * x6) + (((uint128_t)x12 * x4) + ((uint128_t)x14 * x2))))))) + (0x239 * (((uint128_t)x16 * x17) + (((uint128_t)x18 * x18) + ((uint128_t)x17 * x16))))); + { uint128_t x23 = ((((uint128_t)x2 * x12) + ((0x2 * ((uint128_t)x4 * x10)) + ((0x2 * ((uint128_t)x6 * x8)) + ((0x2 * ((uint128_t)x8 * x6)) + ((0x2 * ((uint128_t)x10 * x4)) + ((uint128_t)x12 * x2)))))) + (0x239 * ((0x2 * ((uint128_t)x14 * x17)) + ((0x2 * ((uint128_t)x16 * x18)) + ((0x2 * ((uint128_t)x18 * x16)) + (0x2 * ((uint128_t)x17 * x14))))))); + { uint128_t x24 = ((((uint128_t)x2 * x10) + ((0x2 * ((uint128_t)x4 * x8)) + ((0x2 * ((uint128_t)x6 * x6)) + ((0x2 * ((uint128_t)x8 * x4)) + ((uint128_t)x10 * x2))))) + (0x239 * (((uint128_t)x12 * x17) + ((0x2 * ((uint128_t)x14 * x18)) + ((0x2 * ((uint128_t)x16 * x16)) + ((0x2 * ((uint128_t)x18 * x14)) + ((uint128_t)x17 * x12))))))); + { uint128_t x25 = ((((uint128_t)x2 * x8) + ((0x2 * ((uint128_t)x4 * x6)) + ((0x2 * ((uint128_t)x6 * x4)) + ((uint128_t)x8 * x2)))) + (0x239 * (((uint128_t)x10 * x17) + (((uint128_t)x12 * x18) + ((0x2 * ((uint128_t)x14 * x16)) + ((0x2 * ((uint128_t)x16 * x14)) + (((uint128_t)x18 * x12) + ((uint128_t)x17 * x10)))))))); + { uint128_t x26 = ((((uint128_t)x2 * x6) + ((0x2 * ((uint128_t)x4 * x4)) + ((uint128_t)x6 * x2))) + (0x239 * (((uint128_t)x8 * x17) + (((uint128_t)x10 * x18) + (((uint128_t)x12 * x16) + ((0x2 * ((uint128_t)x14 * x14)) + (((uint128_t)x16 * x12) + (((uint128_t)x18 * x10) + ((uint128_t)x17 * x8))))))))); + { uint128_t x27 = ((((uint128_t)x2 * x4) + ((uint128_t)x4 * x2)) + (0x239 * (((uint128_t)x6 * x17) + (((uint128_t)x8 * x18) + (((uint128_t)x10 * x16) + (((uint128_t)x12 * x14) + (((uint128_t)x14 * x12) + (((uint128_t)x16 * x10) + (((uint128_t)x18 * x8) + ((uint128_t)x17 * x6)))))))))); + { uint128_t x28 = (((uint128_t)x2 * x2) + (0x239 * ((0x2 * ((uint128_t)x4 * x17)) + ((0x2 * ((uint128_t)x6 * x18)) + ((0x2 * ((uint128_t)x8 * x16)) + ((0x2 * ((uint128_t)x10 * x14)) + (((uint128_t)x12 * x12) + ((0x2 * ((uint128_t)x14 * x10)) + ((0x2 * ((uint128_t)x16 * x8)) + ((0x2 * ((uint128_t)x18 * x6)) + (0x2 * ((uint128_t)x17 * x4)))))))))))); + { uint128_t x29 = (x28 >> 0x34); + { uint64_t x30 = ((uint64_t)x28 & 0xfffffffffffff); + { uint128_t x31 = (x29 + x27); + { uint128_t x32 = (x31 >> 0x33); + { uint64_t x33 = ((uint64_t)x31 & 0x7ffffffffffff); + { uint128_t x34 = (x32 + x26); + { uint128_t x35 = (x34 >> 0x33); + { uint64_t x36 = ((uint64_t)x34 & 0x7ffffffffffff); + { uint128_t x37 = (x35 + x25); + { uint128_t x38 = (x37 >> 0x33); + { uint64_t x39 = ((uint64_t)x37 & 0x7ffffffffffff); + { uint128_t x40 = (x38 + x24); + { uint128_t x41 = (x40 >> 0x33); + { uint64_t x42 = ((uint64_t)x40 & 0x7ffffffffffff); + { uint128_t x43 = (x41 + x23); + { uint128_t x44 = (x43 >> 0x34); + { uint64_t x45 = ((uint64_t)x43 & 0xfffffffffffff); + { uint128_t x46 = (x44 + x22); + { uint128_t x47 = (x46 >> 0x33); + { uint64_t x48 = ((uint64_t)x46 & 0x7ffffffffffff); + { uint128_t x49 = (x47 + x21); + { uint128_t x50 = (x49 >> 0x33); + { uint64_t x51 = ((uint64_t)x49 & 0x7ffffffffffff); + { uint128_t x52 = (x50 + x20); + { uint64_t x53 = (uint64_t) (x52 >> 0x33); + { uint64_t x54 = ((uint64_t)x52 & 0x7ffffffffffff); + { uint128_t x55 = (x53 + x19); + { uint64_t x56 = (uint64_t) (x55 >> 0x33); + { uint64_t x57 = ((uint64_t)x55 & 0x7ffffffffffff); + { uint128_t x58 = (x30 + ((uint128_t)0x239 * x56)); + { uint64_t x59 = (uint64_t) (x58 >> 0x34); + { uint64_t x60 = ((uint64_t)x58 & 0xfffffffffffff); + { uint64_t x61 = (x59 + x33); + { uint64_t x62 = (x61 >> 0x33); + { uint64_t x63 = (x61 & 0x7ffffffffffff); + out[0] = x60; + out[1] = x63; + out[2] = (x62 + x36); + out[3] = x39; + out[4] = x42; + out[5] = x45; + out[6] = x48; + out[7] = x51; + out[8] = x54; + out[9] = x57; + }}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}} +} diff --git a/src/Specific/solinas64_2e512m569/freeze.c b/src/Specific/solinas64_2e512m569/freeze.c index ab6b2036f..af890dae2 100644 --- a/src/Specific/solinas64_2e512m569/freeze.c +++ b/src/Specific/solinas64_2e512m569/freeze.c @@ -1,25 +1,54 @@ -#include <stdint.h> -#include <stdbool.h> -#include <x86intrin.h> -#include "liblow.h" - -#include "freeze.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 freeze(uint64_t* out, uint64_t x17, uint64_t x18, uint64_t x16, uint64_t x14, uint64_t x12, uint64_t x10, uint64_t x8, uint64_t x6, uint64_t x4, uint64_t x2) -out[0] = uint64_t x20; -out[1] = uint8_t x21 = Op Syntax.SubWithGetBorrow 52 Syntax.TWord 3 Syntax.TWord 6 Syntax.TWord 6 Syntax.TWord 6 Syntax.TWord 3 0x0; -out[2] = x2; -out[3] = 0xffffffffffdc7;; +static void freeze(uint64_t out[10], const uint64_t in1[10]) { + { const uint64_t x17 = in1[9]; + { const uint64_t x18 = in1[8]; + { const uint64_t x16 = in1[7]; + { const uint64_t x14 = in1[6]; + { const uint64_t x12 = in1[5]; + { const uint64_t x10 = in1[4]; + { const uint64_t x8 = in1[3]; + { const uint64_t x6 = in1[2]; + { const uint64_t x4 = in1[1]; + { const uint64_t x2 = in1[0]; + { uint64_t x20, uint8_t x21 = Op (Syntax.SubWithGetBorrow 52 (Syntax.TWord 3) (Syntax.TWord 6) (Syntax.TWord 6) (Syntax.TWord 6) (Syntax.TWord 3)) (0x0, Return x2, 0xffffffffffdc7); + { uint64_t x23; uint8_t x24 = _subborrow_u51(x21, x4, 0x7ffffffffffff, &x23); + { uint64_t x26; uint8_t x27 = _subborrow_u51(x24, x6, 0x7ffffffffffff, &x26); + { uint64_t x29; uint8_t x30 = _subborrow_u51(x27, x8, 0x7ffffffffffff, &x29); + { uint64_t x32; uint8_t x33 = _subborrow_u51(x30, x10, 0x7ffffffffffff, &x32); + { uint64_t x35, uint8_t x36 = Op (Syntax.SubWithGetBorrow 52 (Syntax.TWord 3) (Syntax.TWord 6) (Syntax.TWord 6) (Syntax.TWord 6) (Syntax.TWord 3)) (Return x33, Return x12, 0xfffffffffffff); + { uint64_t x38; uint8_t x39 = _subborrow_u51(x36, x14, 0x7ffffffffffff, &x38); + { uint64_t x41; uint8_t x42 = _subborrow_u51(x39, x16, 0x7ffffffffffff, &x41); + { uint64_t x44; uint8_t x45 = _subborrow_u51(x42, x18, 0x7ffffffffffff, &x44); + { uint64_t x47; uint8_t x48 = _subborrow_u51(x45, x17, 0x7ffffffffffff, &x47); + { uint64_t x49 = (uint64_t)cmovznz(x48, 0x0, 0xffffffffffffffffL); + { uint64_t x50 = (x49 & 0xffffffffffdc7); + { uint64_t x52, uint8_t x53 = Op (Syntax.AddWithGetCarry 52 (Syntax.TWord 3) (Syntax.TWord 6) (Syntax.TWord 6) (Syntax.TWord 6) (Syntax.TWord 3)) (0x0, Return x20, Return x50); + { uint64_t x54 = (x49 & 0x7ffffffffffff); + { uint64_t x56; uint8_t x57 = _addcarryx_u51(x53, x23, x54, &x56); + { uint64_t x58 = (x49 & 0x7ffffffffffff); + { uint64_t x60; uint8_t x61 = _addcarryx_u51(x57, x26, x58, &x60); + { uint64_t x62 = (x49 & 0x7ffffffffffff); + { uint64_t x64; uint8_t x65 = _addcarryx_u51(x61, x29, x62, &x64); + { uint64_t x66 = (x49 & 0x7ffffffffffff); + { uint64_t x68; uint8_t x69 = _addcarryx_u51(x65, x32, x66, &x68); + { uint64_t x70 = (x49 & 0xfffffffffffff); + { uint64_t x72, uint8_t x73 = Op (Syntax.AddWithGetCarry 52 (Syntax.TWord 3) (Syntax.TWord 6) (Syntax.TWord 6) (Syntax.TWord 6) (Syntax.TWord 3)) (Return x69, Return x35, Return x70); + { uint64_t x74 = (x49 & 0x7ffffffffffff); + { uint64_t x76; uint8_t x77 = _addcarryx_u51(x73, x38, x74, &x76); + { uint64_t x78 = (x49 & 0x7ffffffffffff); + { uint64_t x80; uint8_t x81 = _addcarryx_u51(x77, x41, x78, &x80); + { uint64_t x82 = (x49 & 0x7ffffffffffff); + { uint64_t x84; uint8_t x85 = _addcarryx_u51(x81, x44, x82, &x84); + { uint64_t x86 = (x49 & 0x7ffffffffffff); + { uint64_t x88; uint8_t _ = _addcarryx_u51(x85, x47, x86, &x88); + out[0] = x52; + out[1] = x56; + out[2] = x60; + out[3] = x64; + out[4] = x68; + out[5] = x72; + out[6] = x76; + out[7] = x80; + out[8] = x84; + out[9] = x88; + }}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}} } -// caller: uint64_t out[4]; |