wbrawner/openssl
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Remove unrolled loops

Browse source 
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
(Merged from https://github.com/openssl/openssl/pull/5105)
This commit is contained in:
Matt Caswell 2018-01-31 13:37:42 +00:00
parent 909c68ae72
commit cb5ed32665
3 changed files with 36 additions and 53 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

67
crypto/ec/curve448/arch_32/f_impl.c
View file

@ -12,14 +12,6 @@
#include "field.h"
#if (defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) && !defined(I_HATE_UNROLLED_LOOPS)) \
|| defined(C448_FORCE_UNROLL)
# define REPEAT8(_x) _x _x _x _x _x _x _x _x
# define FOR_LIMB(_i,_start,_end,_x) do { _i=_start; REPEAT8( if (_i<_end) { _x; } _i++;) } while (0)
#else
# define FOR_LIMB(_i,_start,_end,_x) do { for (_i=_start; _i<_end; _i++) _x; } while (0)
#endif
void gf_mul(gf_s * RESTRICT cs, const gf as, const gf bs)
{
const uint32_t *a = as->limb, *b = bs->limb;
@ -34,30 +26,28 @@ void gf_mul(gf_s * RESTRICT cs, const gf as, const gf bs)
bb[i] = b[i] + b[i + 8];
}
FOR_LIMB(j, 0, 8, {
accum2 = 0;
FOR_LIMB(i, 0, j + 1, {
accum2 += widemul(a[j - i], b[i]);
accum1 += widemul(aa[j - i], bb[i]);
accum0 += widemul(a[8 + j - i], b[8 + i]);
}
); accum1 -= accum2; accum0 += accum2;
accum2 = 0;
FOR_LIMB(i, j + 1, 8, {
accum0 -=
widemul(a[8 + j - i], b[i]);
accum2 +=
widemul(aa[8 + j - i],
bb[i]);
accum1 += widemul(a[16 + j - i], b[8 + i]);
}
);
accum1 += accum2;
accum0 += accum2;
c[j] = ((uint32_t)(accum0)) & mask;
c[j + 8] = ((uint32_t)(accum1)) & mask;
accum0 >>= 28; accum1 >>= 28;
});
for (j = 0; j < 8; j++) {
accum2 = 0;
for (i = 0; i < j + 1; i++) {
accum2 += widemul(a[j - i], b[i]);
accum1 += widemul(aa[j - i], bb[i]);
accum0 += widemul(a[8 + j - i], b[8 + i]);
}
accum1 -= accum2;
accum0 += accum2;
accum2 = 0;
for (i = j + 1; i < 8; i++) {
accum0 -= widemul(a[8 + j - i], b[i]);
accum2 += widemul(aa[8 + j - i], bb[i]);
accum1 += widemul(a[16 + j - i], b[8 + i]);
}
accum1 += accum2;
accum0 += accum2;
c[j] = ((uint32_t)(accum0)) & mask;
c[j + 8] = ((uint32_t)(accum1)) & mask;
accum0 >>= 28;
accum1 >>= 28;
}
accum0 += accum1;
accum0 += c[8];
@ -81,11 +71,14 @@ void gf_mulw_unsigned(gf_s * RESTRICT cs, const gf as, uint32_t b)
assert(b < 1 << 28);
FOR_LIMB(i, 0, 8, {
accum0 += widemul(b, a[i]); accum8 += widemul(b, a[i + 8]);
c[i] = accum0 & mask; accum0 >>= 28;
c[i + 8] = accum8 & mask; accum8 >>= 28;
});
for (i = 0; i < 8; i++) {
accum0 += widemul(b, a[i]);
accum8 += widemul(b, a[i + 8]);
c[i] = accum0 & mask;
accum0 >>= 28;
c[i + 8] = accum8 & mask;
accum8 >>= 28;
}
accum0 += accum8 + c[8];
c[8] = ((uint32_t)accum0) & mask;

10
crypto/ec/curve448/f_generic.c
View file

@ -30,7 +30,7 @@ void gf_serialize(uint8_t serial[SER_BYTES], const gf x, int with_hibit)
if (!with_hibit)
assert(gf_hibit(red) == 0);
UNROLL for (i = 0; i < (with_hibit ? X_SER_BYTES : SER_BYTES); i++) {
for (i = 0; i < (with_hibit ? X_SER_BYTES : SER_BYTES); i++) {
if (fill < 8 && j < NLIMBS) {
buffer |= ((dword_t) red->limb[LIMBPERM(j)]) << fill;
fill += LIMB_PLACE_VALUE(LIMBPERM(j));
@ -73,9 +73,11 @@ mask_t gf_deserialize(gf x, const uint8_t serial[SER_BYTES], int with_hibit,
unsigned int i;
mask_t succ;
UNROLL for (i = 0; i < NLIMBS; i++) {
UNROLL while (fill < LIMB_PLACE_VALUE(LIMBPERM(i)) && j < nbytes) {
uint8_t sj = serial[j];
for (i = 0; i < NLIMBS; i++) {
while (fill < LIMB_PLACE_VALUE(LIMBPERM(i)) && j < nbytes) {
uint8_t sj;
sj = serial[j];
if (j == nbytes - 1)
sj &= ~hi_nmask;
buffer |= ((dword_t) sj) << fill;

12
crypto/ec/curve448/word.h
View file

@ -59,18 +59,6 @@ typedef int64_t dsword_t;
# error "For now we only support 32- and 64-bit architectures."
# endif
/* PERF: vectorize vs unroll */
# ifdef __clang__
# if 100*__clang_major__ + __clang_minor__ > 305
# define UNROLL _Pragma("clang loop unroll(full)")
# endif
# endif
# ifndef UNROLL
# define UNROLL
# endif
/*
* The plan on booleans: The external interface uses c448_bool_t, but this
* might be a different size than our particular arch's word_t (and thus