aboutsummaryrefslogtreecommitdiff
path: root/src/Specific/solinas32_2e336m3/femul.c
diff options
context:
space:
mode:
authorGravatar Jason Gross <jgross@mit.edu>2017-11-02 02:20:52 -0400
committerGravatar Jason Gross <jgross@mit.edu>2017-11-02 02:20:52 -0400
commit37a8b2a0c14082298d77f0a4858ab3402d9d1c6c (patch)
treebd1b4c8c34485332ff15d58272efb38468ea72e5 /src/Specific/solinas32_2e336m3/femul.c
parent11144d2e698c4a263e59acb02226d383865f74e0 (diff)
Update display logs and c files
Diffstat (limited to 'src/Specific/solinas32_2e336m3/femul.c')
-rw-r--r--src/Specific/solinas32_2e336m3/femul.c178
1 files changed, 92 insertions, 86 deletions
diff --git a/src/Specific/solinas32_2e336m3/femul.c b/src/Specific/solinas32_2e336m3/femul.c
index e1eb4f061..4d7bf5a52 100644
--- a/src/Specific/solinas32_2e336m3/femul.c
+++ b/src/Specific/solinas32_2e336m3/femul.c
@@ -1,86 +1,92 @@
-#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 x24, uint64_t x25, uint64_t x23, 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 x46, uint64_t x47, uint64_t x45, uint64_t x43, uint64_t x41, uint64_t x39, uint64_t x37, uint64_t x35, uint64_t x33, uint64_t x31, uint64_t x29, uint64_t x27)
-{ uint64_t x48 = (((uint64_t)x5 * x46) + (((uint64_t)x7 * x47) + (((uint64_t)x9 * x45) + (((uint64_t)x11 * x43) + (((uint64_t)x13 * x41) + (((uint64_t)x15 * x39) + (((uint64_t)x17 * x37) + (((uint64_t)x19 * x35) + (((uint64_t)x21 * x33) + (((uint64_t)x23 * x31) + (((uint64_t)x25 * x29) + ((uint64_t)x24 * x27))))))))))));
-{ uint64_t x49 = ((((uint64_t)x5 * x47) + (((uint64_t)x7 * x45) + (((uint64_t)x9 * x43) + (((uint64_t)x11 * x41) + (((uint64_t)x13 * x39) + (((uint64_t)x15 * x37) + (((uint64_t)x17 * x35) + (((uint64_t)x19 * x33) + (((uint64_t)x21 * x31) + (((uint64_t)x23 * x29) + ((uint64_t)x25 * x27))))))))))) + (0x3 * ((uint64_t)x24 * x46)));
-{ uint64_t x50 = ((((uint64_t)x5 * x45) + (((uint64_t)x7 * x43) + (((uint64_t)x9 * x41) + (((uint64_t)x11 * x39) + (((uint64_t)x13 * x37) + (((uint64_t)x15 * x35) + (((uint64_t)x17 * x33) + (((uint64_t)x19 * x31) + (((uint64_t)x21 * x29) + ((uint64_t)x23 * x27)))))))))) + (0x3 * (((uint64_t)x25 * x46) + ((uint64_t)x24 * x47))));
-{ uint64_t x51 = ((((uint64_t)x5 * x43) + (((uint64_t)x7 * x41) + (((uint64_t)x9 * x39) + (((uint64_t)x11 * x37) + (((uint64_t)x13 * x35) + (((uint64_t)x15 * x33) + (((uint64_t)x17 * x31) + (((uint64_t)x19 * x29) + ((uint64_t)x21 * x27))))))))) + (0x3 * (((uint64_t)x23 * x46) + (((uint64_t)x25 * x47) + ((uint64_t)x24 * x45)))));
-{ uint64_t x52 = ((((uint64_t)x5 * x41) + (((uint64_t)x7 * x39) + (((uint64_t)x9 * x37) + (((uint64_t)x11 * x35) + (((uint64_t)x13 * x33) + (((uint64_t)x15 * x31) + (((uint64_t)x17 * x29) + ((uint64_t)x19 * x27)))))))) + (0x3 * (((uint64_t)x21 * x46) + (((uint64_t)x23 * x47) + (((uint64_t)x25 * x45) + ((uint64_t)x24 * x43))))));
-{ uint64_t x53 = ((((uint64_t)x5 * x39) + (((uint64_t)x7 * x37) + (((uint64_t)x9 * x35) + (((uint64_t)x11 * x33) + (((uint64_t)x13 * x31) + (((uint64_t)x15 * x29) + ((uint64_t)x17 * x27))))))) + (0x3 * (((uint64_t)x19 * x46) + (((uint64_t)x21 * x47) + (((uint64_t)x23 * x45) + (((uint64_t)x25 * x43) + ((uint64_t)x24 * x41)))))));
-{ ℤ x54 = ((((uint64_t)x5 * x37) + (((uint64_t)x7 * x35) + (((uint64_t)x9 * x33) + (((uint64_t)x11 * x31) + (((uint64_t)x13 * x29) + ((uint64_t)x15 * x27)))))) +ℤ (0x3 * (((uint64_t)x17 * x46) + (((uint64_t)x19 * x47) + (((uint64_t)x21 * x45) + (((uint64_t)x23 * x43) + (((uint64_t)x25 * x41) + ((uint64_t)x24 * x39))))))));
-{ ℤ x55 = ((((uint64_t)x5 * x35) + (((uint64_t)x7 * x33) + (((uint64_t)x9 * x31) + (((uint64_t)x11 * x29) + ((uint64_t)x13 * x27))))) +ℤ (0x3 * (((uint64_t)x15 * x46) + (((uint64_t)x17 * x47) + (((uint64_t)x19 * x45) + (((uint64_t)x21 * x43) + (((uint64_t)x23 * x41) + (((uint64_t)x25 * x39) + ((uint64_t)x24 * x37)))))))));
-{ ℤ x56 = ((((uint64_t)x5 * x33) + (((uint64_t)x7 * x31) + (((uint64_t)x9 * x29) + ((uint64_t)x11 * x27)))) +ℤ (0x3 *ℤ (((uint64_t)x13 * x46) + (((uint64_t)x15 * x47) + (((uint64_t)x17 * x45) + (((uint64_t)x19 * x43) + (((uint64_t)x21 * x41) + (((uint64_t)x23 * x39) + (((uint64_t)x25 * x37) + ((uint64_t)x24 * x35))))))))));
-{ ℤ x57 = ((((uint64_t)x5 * x31) + (((uint64_t)x7 * x29) + ((uint64_t)x9 * x27))) +ℤ (0x3 *ℤ (((uint64_t)x11 * x46) + (((uint64_t)x13 * x47) + (((uint64_t)x15 * x45) + (((uint64_t)x17 * x43) + (((uint64_t)x19 * x41) + (((uint64_t)x21 * x39) + (((uint64_t)x23 * x37) + (((uint64_t)x25 * x35) + ((uint64_t)x24 * x33)))))))))));
-{ ℤ x58 = ((((uint64_t)x5 * x29) + ((uint64_t)x7 * x27)) +ℤ (0x3 *ℤ (((uint64_t)x9 * x46) + (((uint64_t)x11 * x47) + (((uint64_t)x13 * x45) + (((uint64_t)x15 * x43) + (((uint64_t)x17 * x41) + (((uint64_t)x19 * x39) + (((uint64_t)x21 * x37) + (((uint64_t)x23 * x35) + (((uint64_t)x25 * x33) + ((uint64_t)x24 * x31))))))))))));
-{ ℤ x59 = (((uint64_t)x5 * x27) +ℤ (0x3 *ℤ (((uint64_t)x7 * x46) + (((uint64_t)x9 * x47) + (((uint64_t)x11 * x45) + (((uint64_t)x13 * x43) + (((uint64_t)x15 * x41) + (((uint64_t)x17 * x39) + (((uint64_t)x19 * x37) + (((uint64_t)x21 * x35) + (((uint64_t)x23 * x33) + (((uint64_t)x25 * x31) + ((uint64_t)x24 * x29)))))))))))));
-{ uint64_t x60 = (x59 >> 0x1c);
-{ uint32_t x61 = (x59 & 0xfffffff);
-{ ℤ x62 = (x60 +ℤ x58);
-{ uint64_t x63 = (x62 >> 0x1c);
-{ uint32_t x64 = (x62 & 0xfffffff);
-{ ℤ x65 = (x63 +ℤ x57);
-{ uint64_t x66 = (x65 >> 0x1c);
-{ uint32_t x67 = (x65 & 0xfffffff);
-{ ℤ x68 = (x66 +ℤ x56);
-{ uint64_t x69 = (x68 >> 0x1c);
-{ uint32_t x70 = (x68 & 0xfffffff);
-{ ℤ x71 = (x69 +ℤ x55);
-{ uint64_t x72 = (x71 >> 0x1c);
-{ uint32_t x73 = (x71 & 0xfffffff);
-{ ℤ x74 = (x72 +ℤ x54);
-{ uint64_t x75 = (x74 >> 0x1c);
-{ uint32_t x76 = (x74 & 0xfffffff);
-{ uint64_t x77 = (x75 + x53);
-{ uint64_t x78 = (x77 >> 0x1c);
-{ uint32_t x79 = ((uint32_t)x77 & 0xfffffff);
-{ uint64_t x80 = (x78 + x52);
-{ uint64_t x81 = (x80 >> 0x1c);
-{ uint32_t x82 = ((uint32_t)x80 & 0xfffffff);
-{ uint64_t x83 = (x81 + x51);
-{ uint64_t x84 = (x83 >> 0x1c);
-{ uint32_t x85 = ((uint32_t)x83 & 0xfffffff);
-{ uint64_t x86 = (x84 + x50);
-{ uint64_t x87 = (x86 >> 0x1c);
-{ uint32_t x88 = ((uint32_t)x86 & 0xfffffff);
-{ uint64_t x89 = (x87 + x49);
-{ uint64_t x90 = (x89 >> 0x1c);
-{ uint32_t x91 = ((uint32_t)x89 & 0xfffffff);
-{ uint64_t x92 = (x90 + x48);
-{ uint64_t x93 = (x92 >> 0x1c);
-{ uint32_t x94 = ((uint32_t)x92 & 0xfffffff);
-{ uint64_t x95 = (x61 + (0x3 * x93));
-{ uint32_t x96 = (uint32_t) (x95 >> 0x1c);
-{ uint32_t x97 = ((uint32_t)x95 & 0xfffffff);
-{ uint32_t x98 = (x96 + x64);
-{ uint32_t x99 = (x98 >> 0x1c);
-{ uint32_t x100 = (x98 & 0xfffffff);
-out[0] = x94;
-out[1] = x91;
-out[2] = x88;
-out[3] = x85;
-out[4] = x82;
-out[5] = x79;
-out[6] = x76;
-out[7] = x73;
-out[8] = x70;
-out[9] = x99 + x67;
-out[10] = x100;
-out[11] = x97;
-}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}
-// caller: uint64_t out[12];
+static void femul(uint32_t out[12], const uint32_t in1[12], const uint32_t in2[12]) {
+ { const uint32_t x24 = in1[11];
+ { const uint32_t x25 = in1[10];
+ { const uint32_t x23 = in1[9];
+ { const uint32_t x21 = in1[8];
+ { const uint32_t x19 = in1[7];
+ { const uint32_t x17 = in1[6];
+ { const uint32_t x15 = in1[5];
+ { const uint32_t x13 = in1[4];
+ { const uint32_t x11 = in1[3];
+ { const uint32_t x9 = in1[2];
+ { const uint32_t x7 = in1[1];
+ { const uint32_t x5 = in1[0];
+ { const uint32_t x46 = in2[11];
+ { const uint32_t x47 = in2[10];
+ { const uint32_t x45 = in2[9];
+ { const uint32_t x43 = in2[8];
+ { const uint32_t x41 = in2[7];
+ { const uint32_t x39 = in2[6];
+ { const uint32_t x37 = in2[5];
+ { const uint32_t x35 = in2[4];
+ { const uint32_t x33 = in2[3];
+ { const uint32_t x31 = in2[2];
+ { const uint32_t x29 = in2[1];
+ { const uint32_t x27 = in2[0];
+ { uint64_t x48 = (((uint64_t)x5 * x46) + (((uint64_t)x7 * x47) + (((uint64_t)x9 * x45) + (((uint64_t)x11 * x43) + (((uint64_t)x13 * x41) + (((uint64_t)x15 * x39) + (((uint64_t)x17 * x37) + (((uint64_t)x19 * x35) + (((uint64_t)x21 * x33) + (((uint64_t)x23 * x31) + (((uint64_t)x25 * x29) + ((uint64_t)x24 * x27))))))))))));
+ { uint64_t x49 = ((((uint64_t)x5 * x47) + (((uint64_t)x7 * x45) + (((uint64_t)x9 * x43) + (((uint64_t)x11 * x41) + (((uint64_t)x13 * x39) + (((uint64_t)x15 * x37) + (((uint64_t)x17 * x35) + (((uint64_t)x19 * x33) + (((uint64_t)x21 * x31) + (((uint64_t)x23 * x29) + ((uint64_t)x25 * x27))))))))))) + (0x3 * ((uint64_t)x24 * x46)));
+ { uint64_t x50 = ((((uint64_t)x5 * x45) + (((uint64_t)x7 * x43) + (((uint64_t)x9 * x41) + (((uint64_t)x11 * x39) + (((uint64_t)x13 * x37) + (((uint64_t)x15 * x35) + (((uint64_t)x17 * x33) + (((uint64_t)x19 * x31) + (((uint64_t)x21 * x29) + ((uint64_t)x23 * x27)))))))))) + (0x3 * (((uint64_t)x25 * x46) + ((uint64_t)x24 * x47))));
+ { uint64_t x51 = ((((uint64_t)x5 * x43) + (((uint64_t)x7 * x41) + (((uint64_t)x9 * x39) + (((uint64_t)x11 * x37) + (((uint64_t)x13 * x35) + (((uint64_t)x15 * x33) + (((uint64_t)x17 * x31) + (((uint64_t)x19 * x29) + ((uint64_t)x21 * x27))))))))) + (0x3 * (((uint64_t)x23 * x46) + (((uint64_t)x25 * x47) + ((uint64_t)x24 * x45)))));
+ { uint64_t x52 = ((((uint64_t)x5 * x41) + (((uint64_t)x7 * x39) + (((uint64_t)x9 * x37) + (((uint64_t)x11 * x35) + (((uint64_t)x13 * x33) + (((uint64_t)x15 * x31) + (((uint64_t)x17 * x29) + ((uint64_t)x19 * x27)))))))) + (0x3 * (((uint64_t)x21 * x46) + (((uint64_t)x23 * x47) + (((uint64_t)x25 * x45) + ((uint64_t)x24 * x43))))));
+ { uint64_t x53 = ((((uint64_t)x5 * x39) + (((uint64_t)x7 * x37) + (((uint64_t)x9 * x35) + (((uint64_t)x11 * x33) + (((uint64_t)x13 * x31) + (((uint64_t)x15 * x29) + ((uint64_t)x17 * x27))))))) + (0x3 * (((uint64_t)x19 * x46) + (((uint64_t)x21 * x47) + (((uint64_t)x23 * x45) + (((uint64_t)x25 * x43) + ((uint64_t)x24 * x41)))))));
+ { ℤ x54 = ((((uint64_t)x5 * x37) + (((uint64_t)x7 * x35) + (((uint64_t)x9 * x33) + (((uint64_t)x11 * x31) + (((uint64_t)x13 * x29) + ((uint64_t)x15 * x27)))))) +ℤ (0x3 * (((uint64_t)x17 * x46) + (((uint64_t)x19 * x47) + (((uint64_t)x21 * x45) + (((uint64_t)x23 * x43) + (((uint64_t)x25 * x41) + ((uint64_t)x24 * x39))))))));
+ { ℤ x55 = ((((uint64_t)x5 * x35) + (((uint64_t)x7 * x33) + (((uint64_t)x9 * x31) + (((uint64_t)x11 * x29) + ((uint64_t)x13 * x27))))) +ℤ (0x3 * (((uint64_t)x15 * x46) + (((uint64_t)x17 * x47) + (((uint64_t)x19 * x45) + (((uint64_t)x21 * x43) + (((uint64_t)x23 * x41) + (((uint64_t)x25 * x39) + ((uint64_t)x24 * x37)))))))));
+ { ℤ x56 = ((((uint64_t)x5 * x33) + (((uint64_t)x7 * x31) + (((uint64_t)x9 * x29) + ((uint64_t)x11 * x27)))) +ℤ (0x3 *ℤ (((uint64_t)x13 * x46) + (((uint64_t)x15 * x47) + (((uint64_t)x17 * x45) + (((uint64_t)x19 * x43) + (((uint64_t)x21 * x41) + (((uint64_t)x23 * x39) + (((uint64_t)x25 * x37) + ((uint64_t)x24 * x35))))))))));
+ { ℤ x57 = ((((uint64_t)x5 * x31) + (((uint64_t)x7 * x29) + ((uint64_t)x9 * x27))) +ℤ (0x3 *ℤ (((uint64_t)x11 * x46) + (((uint64_t)x13 * x47) + (((uint64_t)x15 * x45) + (((uint64_t)x17 * x43) + (((uint64_t)x19 * x41) + (((uint64_t)x21 * x39) + (((uint64_t)x23 * x37) + (((uint64_t)x25 * x35) + ((uint64_t)x24 * x33)))))))))));
+ { ℤ x58 = ((((uint64_t)x5 * x29) + ((uint64_t)x7 * x27)) +ℤ (0x3 *ℤ (((uint64_t)x9 * x46) + (((uint64_t)x11 * x47) + (((uint64_t)x13 * x45) + (((uint64_t)x15 * x43) + (((uint64_t)x17 * x41) + (((uint64_t)x19 * x39) + (((uint64_t)x21 * x37) + (((uint64_t)x23 * x35) + (((uint64_t)x25 * x33) + ((uint64_t)x24 * x31))))))))))));
+ { ℤ x59 = (((uint64_t)x5 * x27) +ℤ (0x3 *ℤ (((uint64_t)x7 * x46) + (((uint64_t)x9 * x47) + (((uint64_t)x11 * x45) + (((uint64_t)x13 * x43) + (((uint64_t)x15 * x41) + (((uint64_t)x17 * x39) + (((uint64_t)x19 * x37) + (((uint64_t)x21 * x35) + (((uint64_t)x23 * x33) + (((uint64_t)x25 * x31) + ((uint64_t)x24 * x29)))))))))))));
+ { uint64_t x60 = (x59 >> 0x1c);
+ { uint32_t x61 = (x59 & 0xfffffff);
+ { ℤ x62 = (x60 +ℤ x58);
+ { uint64_t x63 = (x62 >> 0x1c);
+ { uint32_t x64 = (x62 & 0xfffffff);
+ { ℤ x65 = (x63 +ℤ x57);
+ { uint64_t x66 = (x65 >> 0x1c);
+ { uint32_t x67 = (x65 & 0xfffffff);
+ { ℤ x68 = (x66 +ℤ x56);
+ { uint64_t x69 = (x68 >> 0x1c);
+ { uint32_t x70 = (x68 & 0xfffffff);
+ { ℤ x71 = (x69 +ℤ x55);
+ { uint64_t x72 = (x71 >> 0x1c);
+ { uint32_t x73 = (x71 & 0xfffffff);
+ { ℤ x74 = (x72 +ℤ x54);
+ { uint64_t x75 = (x74 >> 0x1c);
+ { uint32_t x76 = (x74 & 0xfffffff);
+ { uint64_t x77 = (x75 + x53);
+ { uint64_t x78 = (x77 >> 0x1c);
+ { uint32_t x79 = ((uint32_t)x77 & 0xfffffff);
+ { uint64_t x80 = (x78 + x52);
+ { uint64_t x81 = (x80 >> 0x1c);
+ { uint32_t x82 = ((uint32_t)x80 & 0xfffffff);
+ { uint64_t x83 = (x81 + x51);
+ { uint64_t x84 = (x83 >> 0x1c);
+ { uint32_t x85 = ((uint32_t)x83 & 0xfffffff);
+ { uint64_t x86 = (x84 + x50);
+ { uint64_t x87 = (x86 >> 0x1c);
+ { uint32_t x88 = ((uint32_t)x86 & 0xfffffff);
+ { uint64_t x89 = (x87 + x49);
+ { uint64_t x90 = (x89 >> 0x1c);
+ { uint32_t x91 = ((uint32_t)x89 & 0xfffffff);
+ { uint64_t x92 = (x90 + x48);
+ { uint64_t x93 = (x92 >> 0x1c);
+ { uint32_t x94 = ((uint32_t)x92 & 0xfffffff);
+ { uint64_t x95 = (x61 + (0x3 * x93));
+ { uint32_t x96 = (uint32_t) (x95 >> 0x1c);
+ { uint32_t x97 = ((uint32_t)x95 & 0xfffffff);
+ { uint32_t x98 = (x96 + x64);
+ { uint32_t x99 = (x98 >> 0x1c);
+ { uint32_t x100 = (x98 & 0xfffffff);
+ out[0] = x97;
+ out[1] = x100;
+ out[2] = (x99 + x67);
+ out[3] = x70;
+ out[4] = x73;
+ out[5] = x76;
+ out[6] = x79;
+ out[7] = x82;
+ out[8] = x85;
+ out[9] = x88;
+ out[10] = x91;
+ out[11] = x94;
+ }}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}
+}