From cd28d129b6a5b84ac40b4a3f8060a6f764aa02b4 Mon Sep 17 00:00:00 2001 From: Matt Caswell Date: Mon, 19 Apr 2021 17:31:28 +0100 Subject: Avoid the need for Configure time 128-bit int detection We just detect this at compile time instead. This avoids cross-compilation problems where the host platform supports 128-bit ints, but the target platform does not (or vice versa). This was causing a problem on some platforms where, dependent on the CFLAGS, 128 bit ints were either supported or not. Fixes #14804 Reviewed-by: Richard Levitte (Merged from https://github.com/openssl/openssl/pull/14941) --- crypto/ec/curve448/arch_32/f_impl.c | 95 ---------------- crypto/ec/curve448/arch_32/f_impl32.c | 104 +++++++++++++++++ crypto/ec/curve448/arch_64/f_impl.c | 200 -------------------------------- crypto/ec/curve448/arch_64/f_impl64.c | 209 ++++++++++++++++++++++++++++++++++ 4 files changed, 313 insertions(+), 295 deletions(-) delete mode 100644 crypto/ec/curve448/arch_32/f_impl.c create mode 100644 crypto/ec/curve448/arch_32/f_impl32.c delete mode 100644 crypto/ec/curve448/arch_64/f_impl.c create mode 100644 crypto/ec/curve448/arch_64/f_impl64.c (limited to 'crypto/ec/curve448') diff --git a/crypto/ec/curve448/arch_32/f_impl.c b/crypto/ec/curve448/arch_32/f_impl.c deleted file mode 100644 index 2e9419b66d..0000000000 --- a/crypto/ec/curve448/arch_32/f_impl.c +++ /dev/null @@ -1,95 +0,0 @@ -/* - * Copyright 2017-2021 The OpenSSL Project Authors. All Rights Reserved. - * Copyright 2014 Cryptography Research, Inc. - * - * Licensed under the Apache License 2.0 (the "License"). You may not use - * this file except in compliance with the License. You can obtain a copy - * in the file LICENSE in the source distribution or at - * https://www.openssl.org/source/license.html - * - * Originally written by Mike Hamburg - */ - -#include "../field.h" - -void gf_mul(gf_s * RESTRICT cs, const gf as, const gf bs) -{ - const uint32_t *a = as->limb, *b = bs->limb; - uint32_t *c = cs->limb; - uint64_t accum0 = 0, accum1 = 0, accum2 = 0; - uint32_t mask = (1 << 28) - 1; - uint32_t aa[8], bb[8]; - int i, j; - - for (i = 0; i < 8; i++) { - aa[i] = a[i] + a[i + 8]; - bb[i] = b[i] + b[i + 8]; - } - - 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]; - accum1 += c[0]; - c[8] = ((uint32_t)(accum0)) & mask; - c[0] = ((uint32_t)(accum1)) & mask; - - accum0 >>= 28; - accum1 >>= 28; - c[9] += ((uint32_t)(accum0)); - c[1] += ((uint32_t)(accum1)); -} - -void gf_mulw_unsigned(gf_s * RESTRICT cs, const gf as, uint32_t b) -{ - const uint32_t *a = as->limb; - uint32_t *c = cs->limb; - uint64_t accum0 = 0, accum8 = 0; - uint32_t mask = (1 << 28) - 1; - int i; - - assert(b <= mask); - - 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; - c[9] += (uint32_t)(accum0 >> 28); - - accum8 += c[0]; - c[0] = ((uint32_t)accum8) & mask; - c[1] += (uint32_t)(accum8 >> 28); -} - -void gf_sqr(gf_s * RESTRICT cs, const gf as) -{ - gf_mul(cs, as, as); /* Performs better with a dedicated square */ -} diff --git a/crypto/ec/curve448/arch_32/f_impl32.c b/crypto/ec/curve448/arch_32/f_impl32.c new file mode 100644 index 0000000000..812c06d84a --- /dev/null +++ b/crypto/ec/curve448/arch_32/f_impl32.c @@ -0,0 +1,104 @@ +/* + * Copyright 2017-2021 The OpenSSL Project Authors. All Rights Reserved. + * Copyright 2014 Cryptography Research, Inc. + * + * Licensed under the Apache License 2.0 (the "License"). You may not use + * this file except in compliance with the License. You can obtain a copy + * in the file LICENSE in the source distribution or at + * https://www.openssl.org/source/license.html + * + * Originally written by Mike Hamburg + */ + +#include "openssl/macros.h" +#include "internal/numbers.h" + +#ifdef UINT128_MAX +/* We have support for 128 bit ints, so do nothing here */ +NON_EMPTY_TRANSLATION_UNIT +#else + +# include "../field.h" + +void gf_mul(gf_s * RESTRICT cs, const gf as, const gf bs) +{ + const uint32_t *a = as->limb, *b = bs->limb; + uint32_t *c = cs->limb; + uint64_t accum0 = 0, accum1 = 0, accum2 = 0; + uint32_t mask = (1 << 28) - 1; + uint32_t aa[8], bb[8]; + int i, j; + + for (i = 0; i < 8; i++) { + aa[i] = a[i] + a[i + 8]; + bb[i] = b[i] + b[i + 8]; + } + + 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]; + accum1 += c[0]; + c[8] = ((uint32_t)(accum0)) & mask; + c[0] = ((uint32_t)(accum1)) & mask; + + accum0 >>= 28; + accum1 >>= 28; + c[9] += ((uint32_t)(accum0)); + c[1] += ((uint32_t)(accum1)); +} + +void gf_mulw_unsigned(gf_s * RESTRICT cs, const gf as, uint32_t b) +{ + const uint32_t *a = as->limb; + uint32_t *c = cs->limb; + uint64_t accum0 = 0, accum8 = 0; + uint32_t mask = (1 << 28) - 1; + int i; + + assert(b <= mask); + + 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; + c[9] += (uint32_t)(accum0 >> 28); + + accum8 += c[0]; + c[0] = ((uint32_t)accum8) & mask; + c[1] += (uint32_t)(accum8 >> 28); +} + +void gf_sqr(gf_s * RESTRICT cs, const gf as) +{ + gf_mul(cs, as, as); /* Performs better with a dedicated square */ +} +#endif diff --git a/crypto/ec/curve448/arch_64/f_impl.c b/crypto/ec/curve448/arch_64/f_impl.c deleted file mode 100644 index 035355cf04..0000000000 --- a/crypto/ec/curve448/arch_64/f_impl.c +++ /dev/null @@ -1,200 +0,0 @@ -/* - * Copyright 2017-2021 The OpenSSL Project Authors. All Rights Reserved. - * Copyright 2014 Cryptography Research, Inc. - * - * Licensed under the OpenSSL license (the "License"). You may not use - * this file except in compliance with the License. You can obtain a copy - * in the file LICENSE in the source distribution or at - * https://www.openssl.org/source/license.html - * - * Originally written by Mike Hamburg - */ - -#include "../field.h" - -void gf_mul(gf_s * RESTRICT cs, const gf as, const gf bs) -{ - const uint64_t *a = as->limb, *b = bs->limb; - uint64_t *c = cs->limb; - uint128_t accum0 = 0, accum1 = 0, accum2; - uint64_t mask = (1ULL << 56) - 1; - uint64_t aa[4], bb[4], bbb[4]; - unsigned int i, j; - - for (i = 0; i < 4; i++) { - aa[i] = a[i] + a[i + 4]; - bb[i] = b[i] + b[i + 4]; - bbb[i] = bb[i] + b[i + 4]; - } - - for (i = 0; i < 4; i++) { - accum2 = 0; - - for (j = 0; j <= i; j++) { - accum2 += widemul(a[j], b[i - j]); - accum1 += widemul(aa[j], bb[i - j]); - accum0 += widemul(a[j + 4], b[i - j + 4]); - } - for (; j < 4; j++) { - accum2 += widemul(a[j], b[i - j + 8]); - accum1 += widemul(aa[j], bbb[i - j + 4]); - accum0 += widemul(a[j + 4], bb[i - j + 4]); - } - - accum1 -= accum2; - accum0 += accum2; - - c[i] = ((uint64_t)(accum0)) & mask; - c[i + 4] = ((uint64_t)(accum1)) & mask; - - accum0 >>= 56; - accum1 >>= 56; - } - - accum0 += accum1; - accum0 += c[4]; - accum1 += c[0]; - c[4] = ((uint64_t)(accum0)) & mask; - c[0] = ((uint64_t)(accum1)) & mask; - - accum0 >>= 56; - accum1 >>= 56; - - c[5] += ((uint64_t)(accum0)); - c[1] += ((uint64_t)(accum1)); -} - -void gf_mulw_unsigned(gf_s * RESTRICT cs, const gf as, uint32_t b) -{ - const uint64_t *a = as->limb; - uint64_t *c = cs->limb; - uint128_t accum0 = 0, accum4 = 0; - uint64_t mask = (1ULL << 56) - 1; - int i; - - for (i = 0; i < 4; i++) { - accum0 += widemul(b, a[i]); - accum4 += widemul(b, a[i + 4]); - c[i] = accum0 & mask; - accum0 >>= 56; - c[i + 4] = accum4 & mask; - accum4 >>= 56; - } - - accum0 += accum4 + c[4]; - c[4] = accum0 & mask; - c[5] += accum0 >> 56; - - accum4 += c[0]; - c[0] = accum4 & mask; - c[1] += accum4 >> 56; -} - -void gf_sqr(gf_s * RESTRICT cs, const gf as) -{ - const uint64_t *a = as->limb; - uint64_t *c = cs->limb; - uint128_t accum0 = 0, accum1 = 0, accum2; - uint64_t mask = (1ULL << 56) - 1; - uint64_t aa[4]; - unsigned int i; - - /* For some reason clang doesn't vectorize this without prompting? */ - for (i = 0; i < 4; i++) - aa[i] = a[i] + a[i + 4]; - - accum2 = widemul(a[0], a[3]); - accum0 = widemul(aa[0], aa[3]); - accum1 = widemul(a[4], a[7]); - - accum2 += widemul(a[1], a[2]); - accum0 += widemul(aa[1], aa[2]); - accum1 += widemul(a[5], a[6]); - - accum0 -= accum2; - accum1 += accum2; - - c[3] = ((uint64_t)(accum1)) << 1 & mask; - c[7] = ((uint64_t)(accum0)) << 1 & mask; - - accum0 >>= 55; - accum1 >>= 55; - - accum0 += widemul(2 * aa[1], aa[3]); - accum1 += widemul(2 * a[5], a[7]); - accum0 += widemul(aa[2], aa[2]); - accum1 += accum0; - - accum0 -= widemul(2 * a[1], a[3]); - accum1 += widemul(a[6], a[6]); - - accum2 = widemul(a[0], a[0]); - accum1 -= accum2; - accum0 += accum2; - - accum0 -= widemul(a[2], a[2]); - accum1 += widemul(aa[0], aa[0]); - accum0 += widemul(a[4], a[4]); - - c[0] = ((uint64_t)(accum0)) & mask; - c[4] = ((uint64_t)(accum1)) & mask; - - accum0 >>= 56; - accum1 >>= 56; - - accum2 = widemul(2 * aa[2], aa[3]); - accum0 -= widemul(2 * a[2], a[3]); - accum1 += widemul(2 * a[6], a[7]); - - accum1 += accum2; - accum0 += accum2; - - accum2 = widemul(2 * a[0], a[1]); - accum1 += widemul(2 * aa[0], aa[1]); - accum0 += widemul(2 * a[4], a[5]); - - accum1 -= accum2; - accum0 += accum2; - - c[1] = ((uint64_t)(accum0)) & mask; - c[5] = ((uint64_t)(accum1)) & mask; - - accum0 >>= 56; - accum1 >>= 56; - - accum2 = widemul(aa[3], aa[3]); - accum0 -= widemul(a[3], a[3]); - accum1 += widemul(a[7], a[7]); - - accum1 += accum2; - accum0 += accum2; - - accum2 = widemul(2 * a[0], a[2]); - accum1 += widemul(2 * aa[0], aa[2]); - accum0 += widemul(2 * a[4], a[6]); - - accum2 += widemul(a[1], a[1]); - accum1 += widemul(aa[1], aa[1]); - accum0 += widemul(a[5], a[5]); - - accum1 -= accum2; - accum0 += accum2; - - c[2] = ((uint64_t)(accum0)) & mask; - c[6] = ((uint64_t)(accum1)) & mask; - - accum0 >>= 56; - accum1 >>= 56; - - accum0 += c[3]; - accum1 += c[7]; - c[3] = ((uint64_t)(accum0)) & mask; - c[7] = ((uint64_t)(accum1)) & mask; - - /* we could almost stop here, but it wouldn't be stable, so... */ - - accum0 >>= 56; - accum1 >>= 56; - c[4] += ((uint64_t)(accum0)) + ((uint64_t)(accum1)); - c[0] += ((uint64_t)(accum1)); -} diff --git a/crypto/ec/curve448/arch_64/f_impl64.c b/crypto/ec/curve448/arch_64/f_impl64.c new file mode 100644 index 0000000000..bdafc0de92 --- /dev/null +++ b/crypto/ec/curve448/arch_64/f_impl64.c @@ -0,0 +1,209 @@ +/* + * Copyright 2017-2021 The OpenSSL Project Authors. All Rights Reserved. + * Copyright 2014 Cryptography Research, Inc. + * + * Licensed under the OpenSSL license (the "License"). You may not use + * this file except in compliance with the License. You can obtain a copy + * in the file LICENSE in the source distribution or at + * https://www.openssl.org/source/license.html + * + * Originally written by Mike Hamburg + */ + +#include "openssl/macros.h" +#include "internal/numbers.h" + +#ifndef UINT128_MAX +/* No support for 128 bit ints, so do nothing here */ +NON_EMPTY_TRANSLATION_UNIT +#else + +# include "../field.h" + +void gf_mul(gf_s * RESTRICT cs, const gf as, const gf bs) +{ + const uint64_t *a = as->limb, *b = bs->limb; + uint64_t *c = cs->limb; + uint128_t accum0 = 0, accum1 = 0, accum2; + uint64_t mask = (1ULL << 56) - 1; + uint64_t aa[4], bb[4], bbb[4]; + unsigned int i, j; + + for (i = 0; i < 4; i++) { + aa[i] = a[i] + a[i + 4]; + bb[i] = b[i] + b[i + 4]; + bbb[i] = bb[i] + b[i + 4]; + } + + for (i = 0; i < 4; i++) { + accum2 = 0; + + for (j = 0; j <= i; j++) { + accum2 += widemul(a[j], b[i - j]); + accum1 += widemul(aa[j], bb[i - j]); + accum0 += widemul(a[j + 4], b[i - j + 4]); + } + for (; j < 4; j++) { + accum2 += widemul(a[j], b[i - j + 8]); + accum1 += widemul(aa[j], bbb[i - j + 4]); + accum0 += widemul(a[j + 4], bb[i - j + 4]); + } + + accum1 -= accum2; + accum0 += accum2; + + c[i] = ((uint64_t)(accum0)) & mask; + c[i + 4] = ((uint64_t)(accum1)) & mask; + + accum0 >>= 56; + accum1 >>= 56; + } + + accum0 += accum1; + accum0 += c[4]; + accum1 += c[0]; + c[4] = ((uint64_t)(accum0)) & mask; + c[0] = ((uint64_t)(accum1)) & mask; + + accum0 >>= 56; + accum1 >>= 56; + + c[5] += ((uint64_t)(accum0)); + c[1] += ((uint64_t)(accum1)); +} + +void gf_mulw_unsigned(gf_s * RESTRICT cs, const gf as, uint32_t b) +{ + const uint64_t *a = as->limb; + uint64_t *c = cs->limb; + uint128_t accum0 = 0, accum4 = 0; + uint64_t mask = (1ULL << 56) - 1; + int i; + + for (i = 0; i < 4; i++) { + accum0 += widemul(b, a[i]); + accum4 += widemul(b, a[i + 4]); + c[i] = accum0 & mask; + accum0 >>= 56; + c[i + 4] = accum4 & mask; + accum4 >>= 56; + } + + accum0 += accum4 + c[4]; + c[4] = accum0 & mask; + c[5] += accum0 >> 56; + + accum4 += c[0]; + c[0] = accum4 & mask; + c[1] += accum4 >> 56; +} + +void gf_sqr(gf_s * RESTRICT cs, const gf as) +{ + const uint64_t *a = as->limb; + uint64_t *c = cs->limb; + uint128_t accum0 = 0, accum1 = 0, accum2; + uint64_t mask = (1ULL << 56) - 1; + uint64_t aa[4]; + unsigned int i; + + /* For some reason clang doesn't vectorize this without prompting? */ + for (i = 0; i < 4; i++) + aa[i] = a[i] + a[i + 4]; + + accum2 = widemul(a[0], a[3]); + accum0 = widemul(aa[0], aa[3]); + accum1 = widemul(a[4], a[7]); + + accum2 += widemul(a[1], a[2]); + accum0 += widemul(aa[1], aa[2]); + accum1 += widemul(a[5], a[6]); + + accum0 -= accum2; + accum1 += accum2; + + c[3] = ((uint64_t)(accum1)) << 1 & mask; + c[7] = ((uint64_t)(accum0)) << 1 & mask; + + accum0 >>= 55; + accum1 >>= 55; + + accum0 += widemul(2 * aa[1], aa[3]); + accum1 += widemul(2 * a[5], a[7]); + accum0 += widemul(aa[2], aa[2]); + accum1 += accum0; + + accum0 -= widemul(2 * a[1], a[3]); + accum1 += widemul(a[6], a[6]); + + accum2 = widemul(a[0], a[0]); + accum1 -= accum2; + accum0 += accum2; + + accum0 -= widemul(a[2], a[2]); + accum1 += widemul(aa[0], aa[0]); + accum0 += widemul(a[4], a[4]); + + c[0] = ((uint64_t)(accum0)) & mask; + c[4] = ((uint64_t)(accum1)) & mask; + + accum0 >>= 56; + accum1 >>= 56; + + accum2 = widemul(2 * aa[2], aa[3]); + accum0 -= widemul(2 * a[2], a[3]); + accum1 += widemul(2 * a[6], a[7]); + + accum1 += accum2; + accum0 += accum2; + + accum2 = widemul(2 * a[0], a[1]); + accum1 += widemul(2 * aa[0], aa[1]); + accum0 += widemul(2 * a[4], a[5]); + + accum1 -= accum2; + accum0 += accum2; + + c[1] = ((uint64_t)(accum0)) & mask; + c[5] = ((uint64_t)(accum1)) & mask; + + accum0 >>= 56; + accum1 >>= 56; + + accum2 = widemul(aa[3], aa[3]); + accum0 -= widemul(a[3], a[3]); + accum1 += widemul(a[7], a[7]); + + accum1 += accum2; + accum0 += accum2; + + accum2 = widemul(2 * a[0], a[2]); + accum1 += widemul(2 * aa[0], aa[2]); + accum0 += widemul(2 * a[4], a[6]); + + accum2 += widemul(a[1], a[1]); + accum1 += widemul(aa[1], aa[1]); + accum0 += widemul(a[5], a[5]); + + accum1 -= accum2; + accum0 += accum2; + + c[2] = ((uint64_t)(accum0)) & mask; + c[6] = ((uint64_t)(accum1)) & mask; + + accum0 >>= 56; + accum1 >>= 56; + + accum0 += c[3]; + accum1 += c[7]; + c[3] = ((uint64_t)(accum0)) & mask; + c[7] = ((uint64_t)(accum1)) & mask; + + /* we could almost stop here, but it wouldn't be stable, so... */ + + accum0 >>= 56; + accum1 >>= 56; + c[4] += ((uint64_t)(accum0)) + ((uint64_t)(accum1)); + c[0] += ((uint64_t)(accum1)); +} +#endif -- cgit v1.2.3