diff options
author | Amitay Isaacs <amitay@ozlabs.org> | 2021-01-19 13:50:18 +1100 |
---|---|---|
committer | Matt Caswell <matt@openssl.org> | 2021-04-08 12:18:09 +0100 |
commit | 927e704e8c2ec47ca58a86f57dc1900d48ebb985 (patch) | |
tree | 161776f4e9ba250df252f64536ce4c5d38a78371 /crypto | |
parent | 0f4286c78a657958bcd5cee38a8d131b4bb2582a (diff) |
Partially Revert "Remove curve448 architecture specific files"
This reverts commit 7e492f3372ed83af074a63d5920f13de7e3455b6.
This brings back the 64-bit reference implementation for curve448.
Reviewed-by: Tomas Mraz <tomas@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/14784)
Diffstat (limited to 'crypto')
-rw-r--r-- | crypto/ec/curve448/arch_ref64/arch_intrinsics.h | 31 | ||||
-rw-r--r-- | crypto/ec/curve448/arch_ref64/f_impl.c | 308 | ||||
-rw-r--r-- | crypto/ec/curve448/arch_ref64/f_impl.h | 49 |
3 files changed, 388 insertions, 0 deletions
diff --git a/crypto/ec/curve448/arch_ref64/arch_intrinsics.h b/crypto/ec/curve448/arch_ref64/arch_intrinsics.h new file mode 100644 index 0000000000..650b63897d --- /dev/null +++ b/crypto/ec/curve448/arch_ref64/arch_intrinsics.h @@ -0,0 +1,31 @@ +/* + * Copyright 2017-2018 The OpenSSL Project Authors. All Rights Reserved. + * Copyright 2016 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 + */ + +#ifndef __ARCH_REF64_ARCH_INTRINSICS_H__ +# define __ARCH_REF64_ARCH_INTRINSICS_H__ + +# define ARCH_WORD_BITS 64 + +static __inline__ __attribute((always_inline, unused)) +uint64_t word_is_zero(uint64_t a) +{ + /* let's hope the compiler isn't clever enough to optimize this. */ + return (((__uint128_t) a) - 1) >> 64; +} + +static __inline__ __attribute((always_inline, unused)) +__uint128_t widemul(uint64_t a, uint64_t b) +{ + return ((__uint128_t) a) * b; +} + +#endif /* ARCH_REF64_ARCH_INTRINSICS_H__ */ diff --git a/crypto/ec/curve448/arch_ref64/f_impl.c b/crypto/ec/curve448/arch_ref64/f_impl.c new file mode 100644 index 0000000000..7cb5749cda --- /dev/null +++ b/crypto/ec/curve448/arch_ref64/f_impl.c @@ -0,0 +1,308 @@ +/* + * Copyright 2017-2018 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; + + 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]; + } + + int I_HATE_UNROLLED_LOOPS = 0; + + if (I_HATE_UNROLLED_LOOPS) { + /* + * The compiler probably won't unroll this, so it's like 80% slower. + */ + for (i = 0; i < 4; i++) { + accum2 = 0; + + unsigned int j; + 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; + } + } else { + accum2 = widemul(a[0], b[0]); + accum1 += widemul(aa[0], bb[0]); + accum0 += widemul(a[4], b[4]); + + accum2 += widemul(a[1], b[7]); + accum1 += widemul(aa[1], bbb[3]); + accum0 += widemul(a[5], bb[3]); + + accum2 += widemul(a[2], b[6]); + accum1 += widemul(aa[2], bbb[2]); + accum0 += widemul(a[6], bb[2]); + + accum2 += widemul(a[3], b[5]); + accum1 += widemul(aa[3], bbb[1]); + accum0 += widemul(a[7], bb[1]); + + accum1 -= accum2; + accum0 += accum2; + + c[0] = ((uint64_t)(accum0)) & mask; + c[4] = ((uint64_t)(accum1)) & mask; + + accum0 >>= 56; + accum1 >>= 56; + + accum2 = widemul(a[0], b[1]); + accum1 += widemul(aa[0], bb[1]); + accum0 += widemul(a[4], b[5]); + + accum2 += widemul(a[1], b[0]); + accum1 += widemul(aa[1], bb[0]); + accum0 += widemul(a[5], b[4]); + + accum2 += widemul(a[2], b[7]); + accum1 += widemul(aa[2], bbb[3]); + accum0 += widemul(a[6], bb[3]); + + accum2 += widemul(a[3], b[6]); + accum1 += widemul(aa[3], bbb[2]); + accum0 += widemul(a[7], bb[2]); + + accum1 -= accum2; + accum0 += accum2; + + c[1] = ((uint64_t)(accum0)) & mask; + c[5] = ((uint64_t)(accum1)) & mask; + + accum0 >>= 56; + accum1 >>= 56; + + accum2 = widemul(a[0], b[2]); + accum1 += widemul(aa[0], bb[2]); + accum0 += widemul(a[4], b[6]); + + accum2 += widemul(a[1], b[1]); + accum1 += widemul(aa[1], bb[1]); + accum0 += widemul(a[5], b[5]); + + accum2 += widemul(a[2], b[0]); + accum1 += widemul(aa[2], bb[0]); + accum0 += widemul(a[6], b[4]); + + accum2 += widemul(a[3], b[7]); + accum1 += widemul(aa[3], bbb[3]); + accum0 += widemul(a[7], bb[3]); + + accum1 -= accum2; + accum0 += accum2; + + c[2] = ((uint64_t)(accum0)) & mask; + c[6] = ((uint64_t)(accum1)) & mask; + + accum0 >>= 56; + accum1 >>= 56; + + accum2 = widemul(a[0], b[3]); + accum1 += widemul(aa[0], bb[3]); + accum0 += widemul(a[4], b[7]); + + accum2 += widemul(a[1], b[2]); + accum1 += widemul(aa[1], bb[2]); + accum0 += widemul(a[5], b[6]); + + accum2 += widemul(a[2], b[1]); + accum1 += widemul(aa[2], bb[1]); + accum0 += widemul(a[6], b[5]); + + accum2 += widemul(a[3], b[0]); + accum1 += widemul(aa[3], bb[0]); + accum0 += widemul(a[7], b[4]); + + accum1 -= accum2; + accum0 += accum2; + + c[3] = ((uint64_t)(accum0)) & mask; + c[7] = ((uint64_t)(accum1)) & mask; + + accum0 >>= 56; + accum1 >>= 56; + } /* !I_HATE_UNROLLED_LOOPS */ + + 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]; + + /* For some reason clang doesn't vectorize this without prompting? */ + unsigned int i; + 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_ref64/f_impl.h b/crypto/ec/curve448/arch_ref64/f_impl.h new file mode 100644 index 0000000000..8751ceecd9 --- /dev/null +++ b/crypto/ec/curve448/arch_ref64/f_impl.h @@ -0,0 +1,49 @@ +/* + * Copyright 2017-2018 The OpenSSL Project Authors. All Rights Reserved. + * Copyright 2014-2016 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 + */ + +#define GF_HEADROOM 9999 /* Everything is reduced anyway */ +#define FIELD_LITERAL(a,b,c,d,e,f,g,h) {{a,b,c,d,e,f,g,h}} + +#define LIMB_PLACE_VALUE(i) 56 + +void gf_add_RAW(gf out, const gf a, const gf b) +{ + for (unsigned int i = 0; i < 8; i++) + out->limb[i] = a->limb[i] + b->limb[i]; + gf_weak_reduce(out); +} + +void gf_sub_RAW(gf out, const gf a, const gf b) +{ + uint64_t co1 = ((1ull << 56) - 1) * 2, co2 = co1 - 2; + + for (unsigned int i = 0; i < 8; i++) + out->limb[i] = a->limb[i] - b->limb[i] + ((i == 4) ? co2 : co1); + gf_weak_reduce(out); +} + +void gf_bias(gf a, int amt) +{ + (void)a; + (void)amt; +} + +void gf_weak_reduce(gf a) +{ + uint64_t mask = (1ull << 56) - 1; + uint64_t tmp = a->limb[7] >> 56; + + a->limb[4] += tmp; + for (unsigned int i = 7; i > 0; i--) + a->limb[i] = (a->limb[i] & mask) + (a->limb[i - 1] >> 56); + a->limb[0] = (a->limb[0] & mask) + tmp; +} |