cannam@154: /* Copyright (c) 2014, Cisco Systems, INC
cannam@154:    Written by XiangMingZhu WeiZhou MinPeng YanWang
cannam@154: 
cannam@154:    Redistribution and use in source and binary forms, with or without
cannam@154:    modification, are permitted provided that the following conditions
cannam@154:    are met:
cannam@154: 
cannam@154:    - Redistributions of source code must retain the above copyright
cannam@154:    notice, this list of conditions and the following disclaimer.
cannam@154: 
cannam@154:    - Redistributions in binary form must reproduce the above copyright
cannam@154:    notice, this list of conditions and the following disclaimer in the
cannam@154:    documentation and/or other materials provided with the distribution.
cannam@154: 
cannam@154:    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
cannam@154:    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
cannam@154:    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
cannam@154:    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
cannam@154:    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
cannam@154:    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
cannam@154:    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
cannam@154:    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
cannam@154:    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
cannam@154:    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
cannam@154:    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
cannam@154: */
cannam@154: 
cannam@154: #if !defined(X86CPU_H)
cannam@154: # define X86CPU_H
cannam@154: 
cannam@154: # if defined(OPUS_X86_MAY_HAVE_SSE)
cannam@154: #  define MAY_HAVE_SSE(name) name ## _sse
cannam@154: # else
cannam@154: #  define MAY_HAVE_SSE(name) name ## _c
cannam@154: # endif
cannam@154: 
cannam@154: # if defined(OPUS_X86_MAY_HAVE_SSE2)
cannam@154: #  define MAY_HAVE_SSE2(name) name ## _sse2
cannam@154: # else
cannam@154: #  define MAY_HAVE_SSE2(name) name ## _c
cannam@154: # endif
cannam@154: 
cannam@154: # if defined(OPUS_X86_MAY_HAVE_SSE4_1)
cannam@154: #  define MAY_HAVE_SSE4_1(name) name ## _sse4_1
cannam@154: # else
cannam@154: #  define MAY_HAVE_SSE4_1(name) name ## _c
cannam@154: # endif
cannam@154: 
cannam@154: # if defined(OPUS_X86_MAY_HAVE_AVX)
cannam@154: #  define MAY_HAVE_AVX(name) name ## _avx
cannam@154: # else
cannam@154: #  define MAY_HAVE_AVX(name) name ## _c
cannam@154: # endif
cannam@154: 
cannam@154: # if defined(OPUS_HAVE_RTCD)
cannam@154: int opus_select_arch(void);
cannam@154: # endif
cannam@154: 
cannam@154: /*gcc appears to emit MOVDQA's to load the argument of an _mm_cvtepi8_epi32()
cannam@154:   or _mm_cvtepi16_epi32() when optimizations are disabled, even though the
cannam@154:   actual PMOVSXWD instruction takes an m32 or m64. Unlike a normal memory
cannam@154:   reference, these require 16-byte alignment and load a full 16 bytes (instead
cannam@154:   of 4 or 8), possibly reading out of bounds.
cannam@154: 
cannam@154:   We can insert an explicit MOVD or MOVQ using _mm_cvtsi32_si128() or
cannam@154:   _mm_loadl_epi64(), which should have the same semantics as an m32 or m64
cannam@154:   reference in the PMOVSXWD instruction itself, but gcc is not smart enough to
cannam@154:   optimize this out when optimizations ARE enabled.
cannam@154: 
cannam@154:   Clang, in contrast, requires us to do this always for _mm_cvtepi8_epi32
cannam@154:   (which is fair, since technically the compiler is always allowed to do the
cannam@154:   dereference before invoking the function implementing the intrinsic).
cannam@154:   However, it is smart enough to eliminate the extra MOVD instruction.
cannam@154:   For _mm_cvtepi16_epi32, it does the right thing, though does *not* optimize out
cannam@154:   the extra MOVQ if it's specified explicitly */
cannam@154: 
cannam@154: # if defined(__clang__) || !defined(__OPTIMIZE__)
cannam@154: #  define OP_CVTEPI8_EPI32_M32(x) \
cannam@154:  (_mm_cvtepi8_epi32(_mm_cvtsi32_si128(*(int *)(x))))
cannam@154: # else
cannam@154: #  define OP_CVTEPI8_EPI32_M32(x) \
cannam@154:  (_mm_cvtepi8_epi32(*(__m128i *)(x)))
cannam@154: #endif
cannam@154: 
cannam@154: /* similar reasoning about the instruction sequence as in the 32-bit macro above,
cannam@154:  */
cannam@154: # if defined(__clang__) || !defined(__OPTIMIZE__)
cannam@154: #  define OP_CVTEPI16_EPI32_M64(x) \
cannam@154:  (_mm_cvtepi16_epi32(_mm_loadl_epi64((__m128i *)(x))))
cannam@154: # else
cannam@154: #  define OP_CVTEPI16_EPI32_M64(x) \
cannam@154:  (_mm_cvtepi16_epi32(*(__m128i *)(x)))
cannam@154: # endif
cannam@154: 
cannam@154: #endif