shithub: opus

Info  •  Files  •  Log  •  Branches

Disable dangerous SSE 4.1 intrinsic optimizations

These could result in 16-byte-aligned loads on unaligned data, causing
a segfault.

--- a/celt/x86/x86cpu.h

+++ b/celt/x86/x86cpu.h

@@ -56,40 +56,10 @@

 int opus_select_arch(void);

 # endif

-/*gcc appears to emit MOVDQA's to load the argument of an _mm_cvtepi8_epi32()

-  or _mm_cvtepi16_epi32() when optimizations are disabled, even though the

-  actual PMOVSXWD instruction takes an m32 or m64. Unlike a normal memory

-  reference, these require 16-byte alignment and load a full 16 bytes (instead

-  of 4 or 8), possibly reading out of bounds.

-

-  We can insert an explicit MOVD or MOVQ using _mm_cvtsi32_si128() or

-  _mm_loadl_epi64(), which should have the same semantics as an m32 or m64

-  reference in the PMOVSXWD instruction itself, but gcc is not smart enough to

-  optimize this out when optimizations ARE enabled.

-

-  Clang, in contrast, requires us to do this always for _mm_cvtepi8_epi32

-  (which is fair, since technically the compiler is always allowed to do the

-  dereference before invoking the function implementing the intrinsic).

-  However, it is smart enough to eliminate the extra MOVD instruction.

-  For _mm_cvtepi16_epi32, it does the right thing, though does *not* optimize out

-  the extra MOVQ if it's specified explicitly */

-

-# if defined(__clang__) || !defined(__OPTIMIZE__)

-#  define OP_CVTEPI8_EPI32_M32(x) \

+#define OP_CVTEPI8_EPI32_M32(x) \

  (_mm_cvtepi8_epi32(_mm_cvtsi32_si128(*(int *)(x))))

-# else

-#  define OP_CVTEPI8_EPI32_M32(x) \

- (_mm_cvtepi8_epi32(*(__m128i *)(x)))

-#endif

-/* similar reasoning about the instruction sequence as in the 32-bit macro above,

- */

-# if defined(__clang__) || !defined(__OPTIMIZE__)

-#  define OP_CVTEPI16_EPI32_M64(x) \

+#define OP_CVTEPI16_EPI32_M64(x) \

  (_mm_cvtepi16_epi32(_mm_loadl_epi64((__m128i *)(x))))

-# else

-#  define OP_CVTEPI16_EPI32_M64(x) \

- (_mm_cvtepi16_epi32(*(__m128i *)(x)))

-# endif

 #endif