Mirrors/openssl
0
0
Fork 0
mirror of https://github.com/openssl/openssl.git synced 2026-01-18 17:11:31 +01:00
Code Issues Packages Projects Releases Wiki Activity

[crypto/ec] Remove unreachable AVX2 code in NISTZ256 implementation

Browse Source 
`crypto/ec/ecp_nistz256.c` contained code sections guarded by a
`ECP_NISTZ256_AVX2` define.

The relevant comment read:

> /*
>  * Note that by default ECP_NISTZ256_AVX2 is undefined. While it's great
>  * code processing 4 points in parallel, corresponding serial operation
>  * is several times slower, because it uses 29x29=58-bit multiplication
>  * as opposite to 64x64=128-bit in integer-only scalar case. As result
>  * it doesn't provide *significant* performance improvement. Note that
>  * just defining ECP_NISTZ256_AVX2 is not sufficient to make it work,
>  * you'd need to compile even asm/ecp_nistz256-avx.pl module.
>  */

Without diminishing the quality of the original submission, it's evident
that this code has been basically unreachable without modifications to
the library source code and is under-tested.

This commit removes these sections from the codebase.

Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
(Merged from https://github.com/openssl/openssl/pull/12019)
This commit is contained in:
Nicola Tuveri
2020-06-02 21:06:48 +03:00
parent c1fd005bfc
commit 00da0f6989
2 changed files with 47 additions and 2337 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2082
crypto/ec/asm/ecp_nistz256-avx2.pl
View File

File diff suppressed because it is too large Load Diff

302
crypto/ec/ecp_nistz256.c
View File

@@ -936,207 +936,6 @@ __owur static int ecp_nistz256_mult_precompute(EC_GROUP *group, BN_CTX *ctx)
return ret;
}
/*
* Note that by default ECP_NISTZ256_AVX2 is undefined. While it's great
* code processing 4 points in parallel, corresponding serial operation
* is several times slower, because it uses 29x29=58-bit multiplication
* as opposite to 64x64=128-bit in integer-only scalar case. As result
* it doesn't provide *significant* performance improvement. Note that
* just defining ECP_NISTZ256_AVX2 is not sufficient to make it work,
* you'd need to compile even asm/ecp_nistz256-avx.pl module.
*/
#if defined(ECP_NISTZ256_AVX2)
# if !(defined(__x86_64) || defined(__x86_64__) || \
defined(_M_AMD64) || defined(_M_X64)) || \
!(defined(__GNUC__) || defined(_MSC_VER)) /* this is for ALIGN32 */
# undef ECP_NISTZ256_AVX2
# else
/* Constant time access, loading four values, from four consecutive tables */
void ecp_nistz256_avx2_multi_gather_w7(void *result, const void *in,
int index0, int index1, int index2,
int index3);
void ecp_nistz256_avx2_transpose_convert(void *RESULTx4, const void *in);
void ecp_nistz256_avx2_convert_transpose_back(void *result, const void *Ax4);
void ecp_nistz256_avx2_point_add_affine_x4(void *RESULTx4, const void *Ax4,
const void *Bx4);
void ecp_nistz256_avx2_point_add_affines_x4(void *RESULTx4, const void *Ax4,
const void *Bx4);
void ecp_nistz256_avx2_to_mont(void *RESULTx4, const void *Ax4);
void ecp_nistz256_avx2_from_mont(void *RESULTx4, const void *Ax4);
void ecp_nistz256_avx2_set1(void *RESULTx4);
int ecp_nistz_avx2_eligible(void);
static void booth_recode_w7(unsigned char *sign,
unsigned char *digit, unsigned char in)
{
unsigned char s, d;
s = ~((in >> 7) - 1);
d = (1 << 8) - in - 1;
d = (d & s) | (in & ~s);
d = (d >> 1) + (d & 1);
*sign = s & 1;
*digit = d;
}
/*
* ecp_nistz256_avx2_mul_g performs multiplication by G, using only the
* precomputed table. It does 4 affine point additions in parallel,
* significantly speeding up point multiplication for a fixed value.
*/
static void ecp_nistz256_avx2_mul_g(P256_POINT *r,
unsigned char p_str[33],
const P256_POINT_AFFINE(*preComputedTable)[64])
{
const unsigned int window_size = 7;
const unsigned int mask = (1 << (window_size + 1)) - 1;
unsigned int wvalue;
/* Using 4 windows at a time */
unsigned char sign0, digit0;
unsigned char sign1, digit1;
unsigned char sign2, digit2;
unsigned char sign3, digit3;
unsigned int idx = 0;
BN_ULONG tmp[P256_LIMBS];
int i;
ALIGN32 BN_ULONG aX4[4 * 9 * 3] = { 0 };
ALIGN32 BN_ULONG bX4[4 * 9 * 2] = { 0 };
ALIGN32 P256_POINT_AFFINE point_arr[4];
ALIGN32 P256_POINT res_point_arr[4];
/* Initial four windows */
wvalue = *((u16 *) & p_str[0]);
wvalue = (wvalue << 1) & mask;
idx += window_size;
booth_recode_w7(&sign0, &digit0, wvalue);
wvalue = *((u16 *) & p_str[(idx - 1) / 8]);
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
booth_recode_w7(&sign1, &digit1, wvalue);
wvalue = *((u16 *) & p_str[(idx - 1) / 8]);
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
booth_recode_w7(&sign2, &digit2, wvalue);
wvalue = *((u16 *) & p_str[(idx - 1) / 8]);
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
booth_recode_w7(&sign3, &digit3, wvalue);
ecp_nistz256_avx2_multi_gather_w7(point_arr, preComputedTable[0],
digit0, digit1, digit2, digit3);
ecp_nistz256_neg(tmp, point_arr[0].Y);
copy_conditional(point_arr[0].Y, tmp, sign0);
ecp_nistz256_neg(tmp, point_arr[1].Y);
copy_conditional(point_arr[1].Y, tmp, sign1);
ecp_nistz256_neg(tmp, point_arr[2].Y);
copy_conditional(point_arr[2].Y, tmp, sign2);
ecp_nistz256_neg(tmp, point_arr[3].Y);
copy_conditional(point_arr[3].Y, tmp, sign3);
ecp_nistz256_avx2_transpose_convert(aX4, point_arr);
ecp_nistz256_avx2_to_mont(aX4, aX4);
ecp_nistz256_avx2_to_mont(&aX4[4 * 9], &aX4[4 * 9]);
ecp_nistz256_avx2_set1(&aX4[4 * 9 * 2]);
wvalue = *((u16 *) & p_str[(idx - 1) / 8]);
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
booth_recode_w7(&sign0, &digit0, wvalue);
wvalue = *((u16 *) & p_str[(idx - 1) / 8]);
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
booth_recode_w7(&sign1, &digit1, wvalue);
wvalue = *((u16 *) & p_str[(idx - 1) / 8]);
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
booth_recode_w7(&sign2, &digit2, wvalue);
wvalue = *((u16 *) & p_str[(idx - 1) / 8]);
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
booth_recode_w7(&sign3, &digit3, wvalue);
ecp_nistz256_avx2_multi_gather_w7(point_arr, preComputedTable[4 * 1],
digit0, digit1, digit2, digit3);
ecp_nistz256_neg(tmp, point_arr[0].Y);
copy_conditional(point_arr[0].Y, tmp, sign0);
ecp_nistz256_neg(tmp, point_arr[1].Y);
copy_conditional(point_arr[1].Y, tmp, sign1);
ecp_nistz256_neg(tmp, point_arr[2].Y);
copy_conditional(point_arr[2].Y, tmp, sign2);
ecp_nistz256_neg(tmp, point_arr[3].Y);
copy_conditional(point_arr[3].Y, tmp, sign3);
ecp_nistz256_avx2_transpose_convert(bX4, point_arr);
ecp_nistz256_avx2_to_mont(bX4, bX4);
ecp_nistz256_avx2_to_mont(&bX4[4 * 9], &bX4[4 * 9]);
/* Optimized when both inputs are affine */
ecp_nistz256_avx2_point_add_affines_x4(aX4, aX4, bX4);
for (i = 2; i < 9; i++) {
wvalue = *((u16 *) & p_str[(idx - 1) / 8]);
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
booth_recode_w7(&sign0, &digit0, wvalue);
wvalue = *((u16 *) & p_str[(idx - 1) / 8]);
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
booth_recode_w7(&sign1, &digit1, wvalue);
wvalue = *((u16 *) & p_str[(idx - 1) / 8]);
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
booth_recode_w7(&sign2, &digit2, wvalue);
wvalue = *((u16 *) & p_str[(idx - 1) / 8]);
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
booth_recode_w7(&sign3, &digit3, wvalue);
ecp_nistz256_avx2_multi_gather_w7(point_arr,
preComputedTable[4 * i],
digit0, digit1, digit2, digit3);
ecp_nistz256_neg(tmp, point_arr[0].Y);
copy_conditional(point_arr[0].Y, tmp, sign0);
ecp_nistz256_neg(tmp, point_arr[1].Y);
copy_conditional(point_arr[1].Y, tmp, sign1);
ecp_nistz256_neg(tmp, point_arr[2].Y);
copy_conditional(point_arr[2].Y, tmp, sign2);
ecp_nistz256_neg(tmp, point_arr[3].Y);
copy_conditional(point_arr[3].Y, tmp, sign3);
ecp_nistz256_avx2_transpose_convert(bX4, point_arr);
ecp_nistz256_avx2_to_mont(bX4, bX4);
ecp_nistz256_avx2_to_mont(&bX4[4 * 9], &bX4[4 * 9]);
ecp_nistz256_avx2_point_add_affine_x4(aX4, aX4, bX4);
}
ecp_nistz256_avx2_from_mont(&aX4[4 * 9 * 0], &aX4[4 * 9 * 0]);
ecp_nistz256_avx2_from_mont(&aX4[4 * 9 * 1], &aX4[4 * 9 * 1]);
ecp_nistz256_avx2_from_mont(&aX4[4 * 9 * 2], &aX4[4 * 9 * 2]);
ecp_nistz256_avx2_convert_transpose_back(res_point_arr, aX4);
/* Last window is performed serially */
wvalue = *((u16 *) & p_str[(idx - 1) / 8]);
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
booth_recode_w7(&sign0, &digit0, wvalue);
ecp_nistz256_gather_w7((P256_POINT_AFFINE *)r,
preComputedTable[36], digit0);
ecp_nistz256_neg(tmp, r->Y);
copy_conditional(r->Y, tmp, sign0);
memcpy(r->Z, ONE, sizeof(ONE));
/* Sum the four windows */
ecp_nistz256_point_add(r, r, &res_point_arr[0]);
ecp_nistz256_point_add(r, r, &res_point_arr[1]);
ecp_nistz256_point_add(r, r, &res_point_arr[2]);
ecp_nistz256_point_add(r, r, &res_point_arr[3]);
}
# endif
#endif
__owur static int ecp_nistz256_set_from_affine(EC_POINT *out, const EC_GROUP *group,
const P256_POINT_AFFINE *in,
BN_CTX *ctx)
@@ -1226,6 +1025,8 @@ __owur static int ecp_nistz256_points_mul(const EC_GROUP *group,
}
if (preComputedTable) {
BN_ULONG infty;
if ((BN_num_bits(scalar) > 256)
|| BN_is_negative(scalar)) {
if ((tmp_scalar = BN_CTX_get(ctx)) == NULL)
@@ -1257,67 +1058,58 @@ __owur static int ecp_nistz256_points_mul(const EC_GROUP *group,
for (; i < 33; i++)
p_str[i] = 0;
#if defined(ECP_NISTZ256_AVX2)
if (ecp_nistz_avx2_eligible()) {
ecp_nistz256_avx2_mul_g(&p.p, p_str, preComputedTable);
} else
#endif
{
BN_ULONG infty;
/* First window */
wvalue = (p_str[0] << 1) & mask;
idx += window_size;
/* First window */
wvalue = (p_str[0] << 1) & mask;
wvalue = _booth_recode_w7(wvalue);
ecp_nistz256_gather_w7(&p.a, preComputedTable[0],
wvalue >> 1);
ecp_nistz256_neg(p.p.Z, p.p.Y);
copy_conditional(p.p.Y, p.p.Z, wvalue & 1);
/*
* Since affine infinity is encoded as (0,0) and
* Jacobian is (,,0), we need to harmonize them
* by assigning "one" or zero to Z.
*/
infty = (p.p.X[0] | p.p.X[1] | p.p.X[2] | p.p.X[3] |
p.p.Y[0] | p.p.Y[1] | p.p.Y[2] | p.p.Y[3]);
if (P256_LIMBS == 8)
infty |= (p.p.X[4] | p.p.X[5] | p.p.X[6] | p.p.X[7] |
p.p.Y[4] | p.p.Y[5] | p.p.Y[6] | p.p.Y[7]);
infty = 0 - is_zero(infty);
infty = ~infty;
p.p.Z[0] = ONE[0] & infty;
p.p.Z[1] = ONE[1] & infty;
p.p.Z[2] = ONE[2] & infty;
p.p.Z[3] = ONE[3] & infty;
if (P256_LIMBS == 8) {
p.p.Z[4] = ONE[4] & infty;
p.p.Z[5] = ONE[5] & infty;
p.p.Z[6] = ONE[6] & infty;
p.p.Z[7] = ONE[7] & infty;
}
for (i = 1; i < 37; i++) {
unsigned int off = (idx - 1) / 8;
wvalue = p_str[off] | p_str[off + 1] << 8;
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
wvalue = _booth_recode_w7(wvalue);
ecp_nistz256_gather_w7(&p.a, preComputedTable[0],
wvalue >> 1);
ecp_nistz256_gather_w7(&t.a,
preComputedTable[i], wvalue >> 1);
ecp_nistz256_neg(p.p.Z, p.p.Y);
copy_conditional(p.p.Y, p.p.Z, wvalue & 1);
ecp_nistz256_neg(t.p.Z, t.a.Y);
copy_conditional(t.a.Y, t.p.Z, wvalue & 1);
/*
* Since affine infinity is encoded as (0,0) and
* Jacobian ias (,,0), we need to harmonize them
* by assigning "one" or zero to Z.
*/
infty = (p.p.X[0] | p.p.X[1] | p.p.X[2] | p.p.X[3] |
p.p.Y[0] | p.p.Y[1] | p.p.Y[2] | p.p.Y[3]);
if (P256_LIMBS == 8)
infty |= (p.p.X[4] | p.p.X[5] | p.p.X[6] | p.p.X[7] |
p.p.Y[4] | p.p.Y[5] | p.p.Y[6] | p.p.Y[7]);
infty = 0 - is_zero(infty);
infty = ~infty;
p.p.Z[0] = ONE[0] & infty;
p.p.Z[1] = ONE[1] & infty;
p.p.Z[2] = ONE[2] & infty;
p.p.Z[3] = ONE[3] & infty;
if (P256_LIMBS == 8) {
p.p.Z[4] = ONE[4] & infty;
p.p.Z[5] = ONE[5] & infty;
p.p.Z[6] = ONE[6] & infty;
p.p.Z[7] = ONE[7] & infty;
}
for (i = 1; i < 37; i++) {
unsigned int off = (idx - 1) / 8;
wvalue = p_str[off] | p_str[off + 1] << 8;
wvalue = (wvalue >> ((idx - 1) % 8)) & mask;
idx += window_size;
wvalue = _booth_recode_w7(wvalue);
ecp_nistz256_gather_w7(&t.a,
preComputedTable[i], wvalue >> 1);
ecp_nistz256_neg(t.p.Z, t.a.Y);
copy_conditional(t.a.Y, t.p.Z, wvalue & 1);
ecp_nistz256_point_add_affine(&p.p, &p.p, &t.a);
}
ecp_nistz256_point_add_affine(&p.p, &p.p, &t.a);
}
} else {
p_is_infinity = 1;