Hi Paul,
On 8/9/21 3:23 PM, Paul A. Clarke via Gcc-patches wrote:
Some compatibility implementations of x86 intrinsics include
Power intrinsics which require POWER8. Guard them.
emmintrin.h:
- _mm_cmpord_pd: Remove code which was ostensibly for pre-POWER8,
but which indeed depended on POWER8 (vec_cmpgt(v2du)/vcmpgtud).
The "POWER8" version works fine on pre-POWER8.
- _mm_mul_epu32: vec_mule(v4su) uses vmuleuw.
pmmintrin.h:
- _mm_movehdup_ps: vec_mergeo(v4su) uses vmrgow.
- _mm_moveldup_ps: vec_mergee(v4su) uses vmrgew.
smmintrin.h:
- _mm_cmpeq_epi64: vec_cmpeq(v2di) uses vcmpequd.
- _mm_mul_epi32: vec_mule(v4si) uses vmuluwm.
- _mm_cmpgt_epi64: vec_cmpgt(v2di) uses vcmpgtsd.
tmmintrin.h:
- _mm_sign_epi8: vec_neg(v4si) uses vsububm.
- _mm_sign_epi16: vec_neg(v4si) uses vsubuhm.
- _mm_sign_epi32: vec_neg(v4si) uses vsubuwm.
Note that the above three could actually be supported pre-POWER8,
but current GCC does not support them before POWER8.
- _mm_sign_pi8: depends on _mm_sign_epi8.
- _mm_sign_pi16: depends on _mm_sign_epi16.
- _mm_sign_pi32: depends on _mm_sign_epi32.
Testing, backports, etc.
2021-08-09 Paul A. Clarke <p...@us.ibm.com>
gcc
I think you mentioned there's a PR open for this.
* config/rs6000/emmintrin.h: Guard POWER8 intrinsics.
* config/rs6000/pmmintrin.h: Same.
* config/rs6000/smmintrin.h: Same.
* config/rs6000/tmmintrin.h: Same.
---
gcc/config/rs6000/emmintrin.h | 12 ++----------
gcc/config/rs6000/pmmintrin.h | 2 ++
gcc/config/rs6000/smmintrin.h | 4 ++++
gcc/config/rs6000/tmmintrin.h | 2 ++
4 files changed, 10 insertions(+), 10 deletions(-)
diff --git a/gcc/config/rs6000/emmintrin.h b/gcc/config/rs6000/emmintrin.h
index ce1287edf782..32ad72b4cc35 100644
--- a/gcc/config/rs6000/emmintrin.h
+++ b/gcc/config/rs6000/emmintrin.h
@@ -430,20 +430,10 @@ _mm_cmpnge_pd (__m128d __A, __m128d __B)
extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
_mm_cmpord_pd (__m128d __A, __m128d __B)
{
-#if _ARCH_PWR8
__v2du c, d;
/* Compare against self will return false (0's) if NAN. */
c = (__v2du)vec_cmpeq (__A, __A);
d = (__v2du)vec_cmpeq (__B, __B);
-#else
- __v2du a, b;
- __v2du c, d;
- const __v2du double_exp_mask = {0x7ff0000000000000, 0x7ff0000000000000};
- a = (__v2du)vec_abs ((__v2df)__A);
- b = (__v2du)vec_abs ((__v2df)__B);
- c = (__v2du)vec_cmpgt (double_exp_mask, a);
- d = (__v2du)vec_cmpgt (double_exp_mask, b);
-#endif
/* A != NAN and B != NAN. */
return ((__m128d)vec_and(c, d));
}
@@ -1472,6 +1462,7 @@ _mm_mul_su32 (__m64 __A, __m64 __B)
return ((__m64)a * (__m64)b);
}
+#ifdef _ARCH_PWR8
extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
_mm_mul_epu32 (__m128i __A, __m128i __B)
{
@@ -1498,6 +1489,7 @@ _mm_mul_epu32 (__m128i __A, __m128i __B)
return (__m128i) vec_mule ((__v4su)__A, (__v4su)__B);
#endif
}
+#endif
extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
_mm_slli_epi16 (__m128i __A, int __B)
diff --git a/gcc/config/rs6000/pmmintrin.h b/gcc/config/rs6000/pmmintrin.h
index eab712fdfa66..d5da1a6daa34 100644
--- a/gcc/config/rs6000/pmmintrin.h
+++ b/gcc/config/rs6000/pmmintrin.h
@@ -123,6 +123,7 @@ _mm_hsub_pd (__m128d __X, __m128d __Y)
vec_mergel ((__v2df) __X, (__v2df)__Y));
}
+#ifdef _ARCH_PWR8
extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
_mm_movehdup_ps (__m128 __X)
{
@@ -134,6 +135,7 @@ _mm_moveldup_ps (__m128 __X)
{
return (__m128)vec_mergee ((__v4su)__X, (__v4su)__X);
}
+#endif
Up to you and the maintainers, but I would have a slight preference for
putting the #ifdef/#endif around each individual function, rather than
groups of functions. It's more maintainable that way.
Otherwise LGTM.
Thanks!
Bill
extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
_mm_loaddup_pd (double const *__P)
diff --git a/gcc/config/rs6000/smmintrin.h b/gcc/config/rs6000/smmintrin.h
index 8d6ae98c7ce3..d2ba5f11de2e 100644
--- a/gcc/config/rs6000/smmintrin.h
+++ b/gcc/config/rs6000/smmintrin.h
@@ -278,6 +278,7 @@ _mm_extract_ps (__m128 __X, const int __N)
return ((__v4si)__X)[__N & 3];
}
+#ifdef _ARCH_PWR8
extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
_mm_blend_epi16 (__m128i __A, __m128i __B, const int __imm8)
{
@@ -289,6 +290,7 @@ _mm_blend_epi16 (__m128i __A, __m128i __B, const int __imm8)
#endif
return (__m128i) vec_sel ((__v8hu) __A, (__v8hu) __B, __shortmask);
}
+#endif
extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__,
__artificial__))
_mm_blendv_epi8 (__m128i __A, __m128i __B, __m128i __mask)
@@ -349,6 +351,7 @@ _mm_blend_pd (__m128d __A, __m128d __B, const int __imm8)
return (__m128d) __r;
}
+#ifdef _ARCH_PWR8
__inline __m128d
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
_mm_blendv_pd (__m128d __A, __m128d __B, __m128d __mask)
@@ -357,6 +360,7 @@ _mm_blendv_pd (__m128d __A, __m128d __B, __m128d __mask)
const __vector __bool long long __boolmask = vec_cmplt ((__v2di) __mask,
__zero);
return (__m128d) vec_sel ((__v2du) __A, (__v2du) __B, (__v2du) __boolmask);
}
+#endif
__inline int
__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
diff --git a/gcc/config/rs6000/tmmintrin.h b/gcc/config/rs6000/tmmintrin.h
index 971511260b78..4120cc1ea551 100644
--- a/gcc/config/rs6000/tmmintrin.h
+++ b/gcc/config/rs6000/tmmintrin.h
@@ -350,6 +350,7 @@ _mm_shuffle_pi8 (__m64 __A, __m64 __B)
return (__m64) ((__v2du) (__C))[0];
}
+#ifdef _ARCH_PWR8
extern __inline __m128i
__attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm_sign_epi8 (__m128i __A, __m128i __B)
@@ -418,6 +419,7 @@ _mm_sign_pi32 (__m64 __A, __m64 __B)
__C = (__v4si) _mm_sign_epi32 ((__m128i) __C, (__m128i) __D);
return (__m64) ((__v2du) (__C))[0];
}
+#endif
extern __inline __m128i
__attribute__((__gnu_inline__, __always_inline__, __artificial__))
