arne
/
cryptopp
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Fix unaligned load for _mm_loaddup_pd with GCC and UBsan 

This function was missed earlier. Unfortunately, it does not squash all of the unaligned load findings. I'm pretty sure it is a GCC problem
pull/552/head
Jeffrey Walton 2017-12-28 01:16:17 -05:00
parent 24883ca79f
commit 7b14ead0f3
No known key found for this signature in database
GPG Key ID: B36AB348921B1838
1 changed files with 28 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

40
adv-simd.h
View File

@ -36,6 +36,14 @@
# include <tmmintrin.h>
#endif
// https://www.spinics.net/lists/gcchelp/msg47735.html and
// https://www.spinics.net/lists/gcchelp/msg47749.html
#if (CRYPTOPP_GCC_VERSION >= 40900)
# define GCC_NO_UBSAN __attribute__ ((no_sanitize_undefined))
#else
# define GCC_NO_UBSAN
#endif
// ************************ All block ciphers *********************** //
ANONYMOUS_NAMESPACE_BEGIN
@ -631,7 +639,7 @@ size_t AdvancedProcessBlocks128_NEON2x6(F2 func2, F6 func6,
return length;
}
NAMESPACE_END
NAMESPACE_END // CryptoPP
#endif // CRYPTOPP_ARM_NEON_AVAILABLE
@ -656,6 +664,14 @@ NAMESPACE_END
# define CONST_M128_CAST(x) ((const __m128i *)(const void *)(x))
#endif
// GCC double casts, https://www.spinics.net/lists/gcchelp/msg47735.html
#ifndef DOUBLE_CAST
# define DOUBLE_CAST(x) ((double *)(void *)(x))
#endif
#ifndef CONST_DOUBLE_CAST
# define CONST_DOUBLE_CAST(x) ((const double *)(const void *)(x))
#endif
ANONYMOUS_NAMESPACE_BEGIN
using CryptoPP::word32;
@ -674,7 +690,7 @@ ANONYMOUS_NAMESPACE_END
NAMESPACE_BEGIN(CryptoPP)
template <typename F2, typename F6>
inline size_t AdvancedProcessBlocks64_SSE2x6(F2 func2, F6 func6,
inline size_t GCC_NO_UBSAN AdvancedProcessBlocks64_SSE2x6(F2 func2, F6 func6,
const word32 *subKeys, size_t rounds, const byte *inBlocks,
const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags)
{
@ -711,7 +727,7 @@ inline size_t AdvancedProcessBlocks64_SSE2x6(F2 func2, F6 func6,
// After the dup load we have two counters in the XMM word. Then we need
// to increment the low ctr by 0 and the high ctr by 1.
block0 = _mm_add_epi32(*CONST_M128_CAST(s_one32x4_1b), _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(inBlocks))));
_mm_loaddup_pd(CONST_DOUBLE_CAST(inBlocks))));
// After initial increment of {0,1} remaining counters increment by {2,2}.
const __m128i be2 = *CONST_M128_CAST(s_one32x4_2b);
@ -722,7 +738,7 @@ inline size_t AdvancedProcessBlocks64_SSE2x6(F2 func2, F6 func6,
block5 = _mm_add_epi32(be2, block4);
// Store the next counter. UBsan false positive; mem_addr can be unaligned.
_mm_store_sd(reinterpret_cast<double*>(const_cast<byte*>(inBlocks)),
_mm_store_sd(DOUBLE_CAST(inBlocks),
_mm_castsi128_pd(_mm_add_epi32(be2, block5)));
}
else
@ -802,14 +818,14 @@ inline size_t AdvancedProcessBlocks64_SSE2x6(F2 func2, F6 func6,
// After the dup load we have two counters in the XMM word. Then we need
// to increment the low ctr by 0 and the high ctr by 1.
block0 = _mm_add_epi32(*CONST_M128_CAST(s_one32x4_1b), _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(inBlocks))));
_mm_loaddup_pd(CONST_DOUBLE_CAST(inBlocks))));
// After initial increment of {0,1} remaining counters increment by {2,2}.
const __m128i be2 = *CONST_M128_CAST(s_one32x4_2b);
block1 = _mm_add_epi32(be2, block0);
// Store the next counter. UBsan false positive; mem_addr can be unaligned.
_mm_store_sd(reinterpret_cast<double*>(const_cast<byte*>(inBlocks)),
_mm_store_sd(DOUBLE_CAST(inBlocks),
_mm_castsi128_pd(_mm_add_epi64(be2, block1)));
}
else
@ -873,13 +889,13 @@ inline size_t AdvancedProcessBlocks64_SSE2x6(F2 func2, F6 func6,
__m128i block, zero = _mm_setzero_si128();
block = _mm_castpd_si128(
// UBsan false positive; mem_addr can be unaligned.
_mm_load_sd(reinterpret_cast<const double*>(inBlocks)));
_mm_load_sd(CONST_DOUBLE_CAST(inBlocks)));
if (flags & BT_XorInput)
{
block = _mm_xor_si128(block, _mm_castpd_si128(
// UBsan false positive; mem_addr can be unaligned.
_mm_load_sd(reinterpret_cast<const double*>(xorBlocks))));
_mm_load_sd(CONST_DOUBLE_CAST(xorBlocks))));
}
if (flags & BT_InBlockIsCounter)
@ -891,11 +907,11 @@ inline size_t AdvancedProcessBlocks64_SSE2x6(F2 func2, F6 func6,
{
block = _mm_xor_si128(block, _mm_castpd_si128(
// UBsan false positive; mem_addr can be unaligned.
_mm_load_sd(reinterpret_cast<const double*>(xorBlocks))));
_mm_load_sd(CONST_DOUBLE_CAST(xorBlocks))));
}
// UBsan false positive; mem_addr can be unaligned.
_mm_store_sd(reinterpret_cast<double*>(outBlocks), _mm_castsi128_pd(block));
_mm_store_sd(DOUBLE_CAST(outBlocks), _mm_castsi128_pd(block));
inBlocks += inIncrement;
outBlocks += outIncrement;
@ -1103,7 +1119,7 @@ inline size_t AdvancedProcessBlocks128_SSE1x4(F1 func1, F4 func4,
CRYPTOPP_ASSERT(length >= 16);
const ptrdiff_t blockSize = 16;
// const ptrdiff_t xmmBlockSize = 16;
// const ptrdiff_t xmmBlockSize = 16;
ptrdiff_t inIncrement = (flags & (BT_InBlockIsCounter|BT_DontIncrementInOutPointers)) ? 0 : blockSize;
ptrdiff_t xorIncrement = xorBlocks ? blockSize : 0;
@ -1212,7 +1228,7 @@ inline size_t AdvancedProcessBlocks128_SSE1x4(F1 func1, F4 func4,
return length;
}
NAMESPACE_END
NAMESPACE_END // CryptoPP
#endif // CRYPTOPP_SSSE3_AVAILABLE