aboutsummaryrefslogtreecommitdiff
path: root/src
diff options
context:
space:
mode:
authorWladimir J. van der Laan <laanwj@protonmail.com>2020-04-22 15:07:16 +0200
committerWladimir J. van der Laan <laanwj@protonmail.com>2020-04-22 15:09:19 +0200
commitce4e1f0282c6a622e8252f06028f1da416a79e23 (patch)
tree79522d5117216cfd8182b595de6366b0af5d4ce9 /src
parent9e8e813df5f19ff154d9314f2d039eb6153f06c4 (diff)
parent850847309458f43fc7ce6c13fa08c86e1cae042a (diff)
downloadbitcoin-ce4e1f0282c6a622e8252f06028f1da416a79e23.tar.xz
Merge #18553: Avoid non-trivial global constants in SHA-NI code
850847309458f43fc7ce6c13fa08c86e1cae042a Avoid non-trivial global constants in SHA-NI code (Pieter Wuille) Pull request description: This is a potential solution for #18456. It seems that the compiler cannot turn `_mm_set_epi64x(<constant>,<constnant>)` into a constant itself, and thus emits a global initializer for the `MASK`, `INIT0`, and `INIT1` global constants in the sha-ni SHA256 implementation. Change this by turning them into dumb byte arrays, loading them into an SSE variable whenever needed. Tested on a SHA-NI capable machine. I do not observe any obvious performance impact (but this is hard to measure, it's already very fast...). ACKs for top commit: laanwj: Code review ACK 850847309458f43fc7ce6c13fa08c86e1cae042a elichai: ACK 850847309458f43fc7ce6c13fa08c86e1cae042a Tree-SHA512: 07049cf1a33624c22df2be48b814d5636c037b368861eb13ee073bdce2b7c902a56e96518218961f55a2a1631a40825ded6dbbc28d7fe0e7fec267d704e39112
Diffstat (limited to 'src')
-rw-r--r--src/crypto/sha256_shani.cpp36
1 files changed, 17 insertions, 19 deletions
diff --git a/src/crypto/sha256_shani.cpp b/src/crypto/sha256_shani.cpp
index 92f67710fb..3473f6e39f 100644
--- a/src/crypto/sha256_shani.cpp
+++ b/src/crypto/sha256_shani.cpp
@@ -11,13 +11,11 @@
#include <stdint.h>
#include <immintrin.h>
-
-
namespace {
-const __m128i MASK = _mm_set_epi64x(0x0c0d0e0f08090a0bULL, 0x0405060700010203ULL);
-const __m128i INIT0 = _mm_set_epi64x(0x6a09e667bb67ae85ull, 0x510e527f9b05688cull);
-const __m128i INIT1 = _mm_set_epi64x(0x3c6ef372a54ff53aull, 0x1f83d9ab5be0cd19ull);
+alignas(__m128i) const uint8_t MASK[16] = {0x03, 0x02, 0x01, 0x00, 0x07, 0x06, 0x05, 0x04, 0x0b, 0x0a, 0x09, 0x08, 0x0f, 0x0e, 0x0d, 0x0c};
+alignas(__m128i) const uint8_t INIT0[16] = {0x8c, 0x68, 0x05, 0x9b, 0x7f, 0x52, 0x0e, 0x51, 0x85, 0xae, 0x67, 0xbb, 0x67, 0xe6, 0x09, 0x6a};
+alignas(__m128i) const uint8_t INIT1[16] = {0x19, 0xcd, 0xe0, 0x5b, 0xab, 0xd9, 0x83, 0x1f, 0x3a, 0xf5, 0x4f, 0xa5, 0x72, 0xf3, 0x6e, 0x3c};
void inline __attribute__((always_inline)) QuadRound(__m128i& state0, __m128i& state1, uint64_t k1, uint64_t k0)
{
@@ -67,12 +65,12 @@ void inline __attribute__((always_inline)) Unshuffle(__m128i& s0, __m128i& s1)
__m128i inline __attribute__((always_inline)) Load(const unsigned char* in)
{
- return _mm_shuffle_epi8(_mm_loadu_si128((const __m128i*)in), MASK);
+ return _mm_shuffle_epi8(_mm_loadu_si128((const __m128i*)in), _mm_load_si128((const __m128i*)MASK));
}
void inline __attribute__((always_inline)) Save(unsigned char* out, __m128i s)
{
- _mm_storeu_si128((__m128i*)out, _mm_shuffle_epi8(s, MASK));
+ _mm_storeu_si128((__m128i*)out, _mm_shuffle_epi8(s, _mm_load_si128((const __m128i*)MASK)));
}
}
@@ -149,8 +147,8 @@ void Transform_2way(unsigned char* out, const unsigned char* in)
__m128i bm0, bm1, bm2, bm3, bs0, bs1, bso0, bso1;
/* Transform 1 */
- bs0 = as0 = INIT0;
- bs1 = as1 = INIT1;
+ bs0 = as0 = _mm_load_si128((const __m128i*)INIT0);
+ bs1 = as1 = _mm_load_si128((const __m128i*)INIT1);
am0 = Load(in);
bm0 = Load(in + 64);
QuadRound(as0, as1, am0, 0xe9b5dba5b5c0fbcfull, 0x71374491428a2f98ull);
@@ -219,10 +217,10 @@ void Transform_2way(unsigned char* out, const unsigned char* in)
ShiftMessageC(bm1, bm2, bm3);
QuadRound(as0, as1, am3, 0xc67178f2bef9A3f7ull, 0xa4506ceb90befffaull);
QuadRound(bs0, bs1, bm3, 0xc67178f2bef9A3f7ull, 0xa4506ceb90befffaull);
- as0 = _mm_add_epi32(as0, INIT0);
- bs0 = _mm_add_epi32(bs0, INIT0);
- as1 = _mm_add_epi32(as1, INIT1);
- bs1 = _mm_add_epi32(bs1, INIT1);
+ as0 = _mm_add_epi32(as0, _mm_load_si128((const __m128i*)INIT0));
+ bs0 = _mm_add_epi32(bs0, _mm_load_si128((const __m128i*)INIT0));
+ as1 = _mm_add_epi32(as1, _mm_load_si128((const __m128i*)INIT1));
+ bs1 = _mm_add_epi32(bs1, _mm_load_si128((const __m128i*)INIT1));
/* Transform 2 */
aso0 = as0;
@@ -275,8 +273,8 @@ void Transform_2way(unsigned char* out, const unsigned char* in)
bm1 = bs1;
/* Transform 3 */
- bs0 = as0 = INIT0;
- bs1 = as1 = INIT1;
+ bs0 = as0 = _mm_load_si128((const __m128i*)INIT0);
+ bs1 = as1 = _mm_load_si128((const __m128i*)INIT1);
QuadRound(as0, as1, am0, 0xe9b5dba5B5c0fbcfull, 0x71374491428a2f98ull);
QuadRound(bs0, bs1, bm0, 0xe9b5dba5B5c0fbcfull, 0x71374491428a2f98ull);
QuadRound(as0, as1, am1, 0xab1c5ed5923f82a4ull, 0x59f111f13956c25bull);
@@ -339,10 +337,10 @@ void Transform_2way(unsigned char* out, const unsigned char* in)
ShiftMessageC(bm1, bm2, bm3);
QuadRound(as0, as1, am3, 0xc67178f2bef9a3f7ull, 0xa4506ceb90befffaull);
QuadRound(bs0, bs1, bm3, 0xc67178f2bef9a3f7ull, 0xa4506ceb90befffaull);
- as0 = _mm_add_epi32(as0, INIT0);
- bs0 = _mm_add_epi32(bs0, INIT0);
- as1 = _mm_add_epi32(as1, INIT1);
- bs1 = _mm_add_epi32(bs1, INIT1);
+ as0 = _mm_add_epi32(as0, _mm_load_si128((const __m128i*)INIT0));
+ bs0 = _mm_add_epi32(bs0, _mm_load_si128((const __m128i*)INIT0));
+ as1 = _mm_add_epi32(as1, _mm_load_si128((const __m128i*)INIT1));
+ bs1 = _mm_add_epi32(bs1, _mm_load_si128((const __m128i*)INIT1));
/* Extract hash into out */
Unshuffle(as0, as1);