| 1 |
/*===---- pmmintrin.h - Implementation of SSE3 intrinsics on PowerPC -------=== |
| 2 |
* |
| 3 |
* Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 |
* See https://llvm.org/LICENSE.txt for license information. |
| 5 |
* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 |
* |
| 7 |
*===-----------------------------------------------------------------------=== |
| 8 |
*/ |
| 9 |
|
| 10 |
/* Implemented from the specification included in the Intel C++ Compiler |
| 11 |
User Guide and Reference, version 9.0. */ |
| 12 |
|
| 13 |
#ifndef NO_WARN_X86_INTRINSICS |
| 14 |
/* This header is distributed to simplify porting x86_64 code that |
| 15 |
makes explicit use of Intel intrinsics to powerpc64le. |
| 16 |
It is the user's responsibility to determine if the results are |
| 17 |
acceptable and make additional changes as necessary. |
| 18 |
Note that much code that uses Intel intrinsics can be rewritten in |
| 19 |
standard C or GNU C extensions, which are more portable and better |
| 20 |
optimized across multiple targets. |
| 21 |
|
| 22 |
In the specific case of X86 SSE3 intrinsics, the PowerPC VMX/VSX ISA |
| 23 |
is a good match for most SIMD operations. However the Horizontal |
| 24 |
add/sub requires the data pairs be permuted into a separate |
| 25 |
registers with vertical even/odd alignment for the operation. |
| 26 |
And the addsub operation requires the sign of only the even numbered |
| 27 |
elements be flipped (xored with -0.0). |
| 28 |
For larger blocks of code using these intrinsic implementations, |
| 29 |
the compiler be should be able to schedule instructions to avoid |
| 30 |
additional latency. |
| 31 |
|
| 32 |
In the specific case of the monitor and mwait instructions there are |
| 33 |
no direct equivalent in the PowerISA at this time. So those |
| 34 |
intrinsics are not implemented. */ |
| 35 |
#error \ |
| 36 |
"Please read comment above. Use -DNO_WARN_X86_INTRINSICS to disable this warning." |
| 37 |
#endif |
| 38 |
|
| 39 |
#ifndef PMMINTRIN_H_ |
| 40 |
#define PMMINTRIN_H_ |
| 41 |
|
| 42 |
#if defined(__powerpc64__) && \ |
| 43 |
(defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)) |
| 44 |
|
| 45 |
/* We need definitions from the SSE2 and SSE header files*/ |
| 46 |
#include <emmintrin.h> |
| 47 |
|
| 48 |
extern __inline __m128 |
| 49 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 50 |
_mm_addsub_ps(__m128 __X, __m128 __Y) { |
| 51 |
const __v4sf __even_n0 = {-0.0, 0.0, -0.0, 0.0}; |
| 52 |
__v4sf __even_neg_Y = vec_xor(__Y, __even_n0); |
| 53 |
return (__m128)vec_add(__X, __even_neg_Y); |
| 54 |
} |
| 55 |
|
| 56 |
extern __inline __m128d |
| 57 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 58 |
_mm_addsub_pd(__m128d __X, __m128d __Y) { |
| 59 |
const __v2df __even_n0 = {-0.0, 0.0}; |
| 60 |
__v2df __even_neg_Y = vec_xor(__Y, __even_n0); |
| 61 |
return (__m128d)vec_add(__X, __even_neg_Y); |
| 62 |
} |
| 63 |
|
| 64 |
extern __inline __m128 |
| 65 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 66 |
_mm_hadd_ps(__m128 __X, __m128 __Y) { |
| 67 |
__vector unsigned char __xform2 = {0x00, 0x01, 0x02, 0x03, 0x08, 0x09, |
| 68 |
0x0A, 0x0B, 0x10, 0x11, 0x12, 0x13, |
| 69 |
0x18, 0x19, 0x1A, 0x1B}; |
| 70 |
__vector unsigned char __xform1 = {0x04, 0x05, 0x06, 0x07, 0x0C, 0x0D, |
| 71 |
0x0E, 0x0F, 0x14, 0x15, 0x16, 0x17, |
| 72 |
0x1C, 0x1D, 0x1E, 0x1F}; |
| 73 |
return (__m128)vec_add(vec_perm((__v4sf)__X, (__v4sf)__Y, __xform2), |
| 74 |
vec_perm((__v4sf)__X, (__v4sf)__Y, __xform1)); |
| 75 |
} |
| 76 |
|
| 77 |
extern __inline __m128 |
| 78 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 79 |
_mm_hsub_ps(__m128 __X, __m128 __Y) { |
| 80 |
__vector unsigned char __xform2 = {0x00, 0x01, 0x02, 0x03, 0x08, 0x09, |
| 81 |
0x0A, 0x0B, 0x10, 0x11, 0x12, 0x13, |
| 82 |
0x18, 0x19, 0x1A, 0x1B}; |
| 83 |
__vector unsigned char __xform1 = {0x04, 0x05, 0x06, 0x07, 0x0C, 0x0D, |
| 84 |
0x0E, 0x0F, 0x14, 0x15, 0x16, 0x17, |
| 85 |
0x1C, 0x1D, 0x1E, 0x1F}; |
| 86 |
return (__m128)vec_sub(vec_perm((__v4sf)__X, (__v4sf)__Y, __xform2), |
| 87 |
vec_perm((__v4sf)__X, (__v4sf)__Y, __xform1)); |
| 88 |
} |
| 89 |
|
| 90 |
extern __inline __m128d |
| 91 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 92 |
_mm_hadd_pd(__m128d __X, __m128d __Y) { |
| 93 |
return (__m128d)vec_add(vec_mergeh((__v2df)__X, (__v2df)__Y), |
| 94 |
vec_mergel((__v2df)__X, (__v2df)__Y)); |
| 95 |
} |
| 96 |
|
| 97 |
extern __inline __m128d |
| 98 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 99 |
_mm_hsub_pd(__m128d __X, __m128d __Y) { |
| 100 |
return (__m128d)vec_sub(vec_mergeh((__v2df)__X, (__v2df)__Y), |
| 101 |
vec_mergel((__v2df)__X, (__v2df)__Y)); |
| 102 |
} |
| 103 |
|
| 104 |
#ifdef _ARCH_PWR8 |
| 105 |
extern __inline __m128 |
| 106 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 107 |
_mm_movehdup_ps(__m128 __X) { |
| 108 |
return (__m128)vec_mergeo((__v4su)__X, (__v4su)__X); |
| 109 |
} |
| 110 |
#endif |
| 111 |
|
| 112 |
#ifdef _ARCH_PWR8 |
| 113 |
extern __inline __m128 |
| 114 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 115 |
_mm_moveldup_ps(__m128 __X) { |
| 116 |
return (__m128)vec_mergee((__v4su)__X, (__v4su)__X); |
| 117 |
} |
| 118 |
#endif |
| 119 |
|
| 120 |
extern __inline __m128d |
| 121 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 122 |
_mm_loaddup_pd(double const *__P) { |
| 123 |
return (__m128d)vec_splats(*__P); |
| 124 |
} |
| 125 |
|
| 126 |
extern __inline __m128d |
| 127 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 128 |
_mm_movedup_pd(__m128d __X) { |
| 129 |
return _mm_shuffle_pd(__X, __X, _MM_SHUFFLE2(0, 0)); |
| 130 |
} |
| 131 |
|
| 132 |
extern __inline __m128i |
| 133 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 134 |
_mm_lddqu_si128(__m128i const *__P) { |
| 135 |
return (__m128i)(vec_vsx_ld(0, (signed int const *)__P)); |
| 136 |
} |
| 137 |
|
| 138 |
/* POWER8 / POWER9 have no equivalent for _mm_monitor nor _mm_wait. */ |
| 139 |
|
| 140 |
#else |
| 141 |
#include_next <pmmintrin.h> |
| 142 |
#endif /* defined(__powerpc64__) && \ |
| 143 |
* (defined(__linux__) || defined(__FreeBSD__) || defined(_AIX)) */ |
| 144 |
|
| 145 |
#endif /* PMMINTRIN_H_ */ |