cheng/libsodium
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Force alignment for _mm_loadl_epi64() in DEBUG mode

Browse Source 
Required to work around gcc sanitizer
This commit is contained in:
Frank Denis 2015-12-07 22:44:24 +01:00
parent ea43d1b18b
commit 731f2e1c12
1 changed files with 15 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
src/libsodium/crypto_onetimeauth/poly1305/sse2/poly1305_sse2.c
View File

@ -53,6 +53,21 @@ typedef struct poly1305_state_internal_t {
unsigned char buffer[poly1305_block_size]; /* 32 bytes */ unsigned char buffer[poly1305_block_size]; /* 32 bytes */
} poly1305_state_internal_t; /* 164 bytes total */ } poly1305_state_internal_t; /* 164 bytes total */
/*
* _mm_loadl_epi64() is turned into a simple MOVQ. So, unaligned accesses are totally fine, even though this intrinsic requires a __m128i* input.
* This confuses dynamic analysis, so force alignment, only in debug mode.
*/
#ifdef DEBUG
static xmmi
_fakealign_mm_loadl_epi64(const void *m)
{
xmmi tmp;
memcpy(&tmp, m, 8);
return _mm_loadl_epi64(&tmp);
}
# define _mm_loadl_epi64(X) _fakealign_mm_loadl_epi64(X)
#endif
/* copy 0-31 bytes */ /* copy 0-31 bytes */
static inline void static inline void
poly1305_block_copy31(unsigned char *dst, const unsigned char *src, unsigned long long bytes) poly1305_block_copy31(unsigned char *dst, const unsigned char *src, unsigned long long bytes)
@ -166,7 +181,6 @@ poly1305_blocks(poly1305_state_internal_t *st, const unsigned char *m,
if (!(st->flags & poly1305_started)) { if (!(st->flags & poly1305_started)) {
/* H = [Mx,My] */ /* H = [Mx,My] */
/* Note that _mm_loadl_epi64() is turned into a simple MOVQ. So, unaligned accesses are totally fine, even though this intrinsic requires a __m128i* input */
T5 = _mm_unpacklo_epi64(_mm_loadl_epi64((const xmmi *)(const void *)(m + 0)), _mm_loadl_epi64((const xmmi *)(const void *)(m + 16))); T5 = _mm_unpacklo_epi64(_mm_loadl_epi64((const xmmi *)(const void *)(m + 0)), _mm_loadl_epi64((const xmmi *)(const void *)(m + 16)));
T6 = _mm_unpacklo_epi64(_mm_loadl_epi64((const xmmi *)(const void *)(m + 8)), _mm_loadl_epi64((const xmmi *)(const void *)(m + 24))); T6 = _mm_unpacklo_epi64(_mm_loadl_epi64((const xmmi *)(const void *)(m + 8)), _mm_loadl_epi64((const xmmi *)(const void *)(m + 24)));
H0 = _mm_and_si128(MMASK, T5); H0 = _mm_and_si128(MMASK, T5);