[x86][avx512] more changes in intrinsics to be align with gcc format

[delena]

LGTM

[echristo]

Why?

-eric

[llvm-commits]

To be compatible with GCC headers.

• Elena

[echristo]

Why again? We're already not compatible because we don't handle, nor do we
want to, other builtins.

-eric

[delena]

Hi Eric,

We did not complete the SKX intrinsics yet. We are adding them to the backend and to the clang. And we still have a lot of work to do.
The names and parameters should be the same in GCC and in Clang headers.
The current patch is just a minor fix.
Could you, please, explain your point?

• Elena

From: Eric Christopher [mailto:[e-mail]]
Sent: Thursday, December 10, 2015 20:23
To: [e-mail]; Demikhovsky, Elena <[e-mail]>
Subject: Re: [PATCH] D15328: [x86][avx512] more changes in intrinsics to be align with gcc format

Why again? We're already not compatible because we don't handle, nor do we want to, other builtins.

-eric

[llvm-commits]

The stance I had always taken was that if clang and gcc both have a builtin
for something we match them. If clang doesn't need a builtin and gcc does
then we just don't match.

• msg-32599-28.txt162 BDownload

[echristo]

Seems reasonable to me. It just seemed odd that we were changing. Header
compatibility isn't necessarily a good reason, but being able to keep
source compatibility if possible is nice.

Thanks!

[delena]

We’ll keep source compatibility. This part is under construction yet. Amount is huge, we are adding intrinsics and running a lot of tests for GCC compatibility.
Within a time this code will be stable.

• Elena

From: Eric Christopher [mailto:[e-mail]]
Sent: Thursday, December 10, 2015 23:43
To: Craig Topper <[e-mail]>; [e-mail]; Demikhovsky, Elena <[e-mail]>
Cc: [e-mail]
Subject: Re: [PATCH] D15328: [x86][avx512] more changes in intrinsics to be align with gcc format

Seems reasonable to me. It just seemed odd that we were changing. Header compatibility isn't necessarily a good reason, but being able to keep source compatibility if possible is nice.

Thanks!

[echristo]

Sounds good.

Thanks!

cfe/trunk/include/clang/Basic/BuiltinsX86.def

 
 BUILTIN(__builtin_ia32_rdpmc, "ULLii", "")
 BUILTIN(__builtin_ia32_rdtsc, "ULLi", "")
 BUILTIN(__builtin_ia32_rdtscp, "ULLiUi*", "")
 
 // AVX-512
 TARGET_BUILTIN(__builtin_ia32_sqrtpd512_mask, "V8dV8dV8dUcIi", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_sqrtps512_mask, "V16fV16fV16fUsIi", "", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rsqrt14sd_mask, "V2dV2dV2dV2dUc", "", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rsqrt14ss_mask, "V4fV4fV4fV4fUc", "", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14sd, "V2dV2dV2dV2dUc", "", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rsqrt14ss, "V4fV4fV4fV4fUc", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_rsqrt14pd512_mask, "V8dV8dV8dUc", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_rsqrt14ps512_mask, "V16fV16fV16fUs", "", "avx512f")
 
-TARGET_BUILTIN(__builtin_ia32_rsqrt28sd_mask, "V2dV2dV2dV2dUcIi", "", "avx512er")
-TARGET_BUILTIN(__builtin_ia32_rsqrt28ss_mask, "V4fV4fV4fV4fUcIi", "", "avx512er")
+TARGET_BUILTIN(__builtin_ia32_rsqrt28sd_round, "V2dV2dV2dV2dUcIi", "", "avx512er")
+TARGET_BUILTIN(__builtin_ia32_rsqrt28ss_round, "V4fV4fV4fV4fUcIi", "", "avx512er")
 TARGET_BUILTIN(__builtin_ia32_rsqrt28pd_mask, "V8dV8dV8dUcIi", "", "avx512er")
 TARGET_BUILTIN(__builtin_ia32_rsqrt28ps_mask, "V16fV16fV16fUsIi", "", "avx512er")
 
-TARGET_BUILTIN(__builtin_ia32_rcp14sd_mask, "V2dV2dV2dV2dUc", "", "avx512f")
-TARGET_BUILTIN(__builtin_ia32_rcp14ss_mask, "V4fV4fV4fV4fUc", "", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rcp14sd, "V2dV2dV2dV2dUc", "", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_rcp14ss, "V4fV4fV4fV4fUc", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_rcp14pd512_mask, "V8dV8dV8dUc", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_rcp14ps512_mask, "V16fV16fV16fUs", "", "avx512f")
 
-TARGET_BUILTIN(__builtin_ia32_rcp28sd_mask, "V2dV2dV2dV2dUcIi", "", "avx512er")
-TARGET_BUILTIN(__builtin_ia32_rcp28ss_mask, "V4fV4fV4fV4fUcIi", "", "avx512er")
+TARGET_BUILTIN(__builtin_ia32_rcp28sd_round, "V2dV2dV2dV2dUcIi", "", "avx512er")
+TARGET_BUILTIN(__builtin_ia32_rcp28ss_round, "V4fV4fV4fV4fUcIi", "", "avx512er")
 TARGET_BUILTIN(__builtin_ia32_rcp28pd_mask, "V8dV8dV8dUcIi", "", "avx512er")
 TARGET_BUILTIN(__builtin_ia32_rcp28ps_mask, "V16fV16fV16fUsIi", "", "avx512er")
 TARGET_BUILTIN(__builtin_ia32_exp2pd_mask, "V8dV8dV8dUcIi", "", "avx512er")
 TARGET_BUILTIN(__builtin_ia32_exp2ps_mask, "V16fV16fV16fUsIi", "", "avx512er")
 
 TARGET_BUILTIN(__builtin_ia32_cvttps2dq512_mask, "V16iV16fV16iUsIi", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_cvttps2udq512_mask, "V16iV16fV16iUsIi", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_cvttpd2dq512_mask, "V8iV8dV8iUcIi", "", "avx512f")

cfe/trunk/lib/Headers/avx512erintrin.h

 
 #define _mm512_mask_rsqrt28_ps(S, M, A) \
   _mm512_mask_rsqrt28_round_ps((S), (M), A, _MM_FROUND_CUR_DIRECTION)
 
 #define _mm512_maskz_rsqrt28_ps(M, A) \
   _mm512_maskz_rsqrt28_round_ps((M), (A), _MM_FROUND_CUR_DIRECTION)
 
 #define _mm_rsqrt28_round_ss(A, B, R) __extension__ ({ \
-  (__m128)__builtin_ia32_rsqrt28ss_mask((__v4sf)(__m128)(A), \
+  (__m128)__builtin_ia32_rsqrt28ss_round((__v4sf)(__m128)(A), \
                                         (__v4sf)(__m128)(B), \
                                         (__v4sf)_mm_setzero_ps(), \
                                         (__mmask8)-1, (R)); })
 
 #define _mm_mask_rsqrt28_round_ss(S, M, A, B, R) __extension__ ({ \
-  (__m128)__builtin_ia32_rsqrt28ss_mask((__v4sf)(__m128)(A), \
+  (__m128)__builtin_ia32_rsqrt28ss_round((__v4sf)(__m128)(A), \
                                         (__v4sf)(__m128)(B), \
                                         (__v4sf)(__m128)(S), \
                                         (__mmask8)(M), (R)); })
 
 #define _mm_maskz_rsqrt28_round_ss(M, A, B, R) __extension__ ({ \
-  (__m128)__builtin_ia32_rsqrt28ss_mask((__v4sf)(__m128)(A), \
+  (__m128)__builtin_ia32_rsqrt28ss_round((__v4sf)(__m128)(A), \
                                         (__v4sf)(__m128)(B), \
                                         (__v4sf)_mm_setzero_ps(), \
                                         (__mmask8)(M), (R)); })
 
 #define _mm_rsqrt28_ss(A, B) \
   _mm_rsqrt28_round_ss((A), (B), _MM_FROUND_CUR_DIRECTION)
 
 #define _mm_mask_rsqrt28_ss(S, M, A, B) \
   _mm_mask_rsqrt28_round_ss((S), (M), (A), (B), _MM_FROUND_CUR_DIRECTION)
 
 #define _mm_maskz_rsqrt28_ss(M, A, B) \
   _mm_maskz_rsqrt28_round_ss((M), (A), (B), _MM_FROUND_CUR_DIRECTION)
 
 #define _mm_rsqrt28_round_sd(A, B, R) __extension__ ({ \
-  (__m128d)__builtin_ia32_rsqrt28sd_mask((__v2df)(__m128d)(A), \
+  (__m128d)__builtin_ia32_rsqrt28sd_round((__v2df)(__m128d)(A), \
                                          (__v2df)(__m128d)(B), \
                                          (__v2df)_mm_setzero_pd(), \
                                          (__mmask8)-1, (R)); })
 
 #define _mm_mask_rsqrt28_round_sd(S, M, A, B, R) __extension__ ({ \
-  (__m128d)__builtin_ia32_rsqrt28sd_mask((__v2df)(__m128d)(A), \
+  (__m128d)__builtin_ia32_rsqrt28sd_round((__v2df)(__m128d)(A), \
                                          (__v2df)(__m128d)(B), \
                                          (__v2df)(__m128d)(S), \
                                          (__mmask8)(M), (R)); })
 
 #define _mm_maskz_rsqrt28_round_sd(M, A, B, R) __extension__ ({ \
-  (__m128d)__builtin_ia32_rsqrt28sd_mask((__v2df)(__m128d)(A), \
+  (__m128d)__builtin_ia32_rsqrt28sd_round((__v2df)(__m128d)(A), \
                                          (__v2df)(__m128d)(B), \
                                          (__v2df)_mm_setzero_pd(), \
                                          (__mmask8)(M), (R)); })
 
 #define _mm_rsqrt28_sd(A, B) \
   _mm_rsqrt28_round_sd((A), (B), _MM_FROUND_CUR_DIRECTION)
 
 #define _mm_mask_rsqrt28_sd(S, M, A, B) \

 
 #define _mm512_mask_rcp28_ps(S, M, A) \
   _mm512_mask_rcp28_round_ps((S), (M), (A), _MM_FROUND_CUR_DIRECTION)
 
 #define _mm512_maskz_rcp28_ps(M, A) \
   _mm512_maskz_rcp28_round_ps((M), (A), _MM_FROUND_CUR_DIRECTION)
 
 #define _mm_rcp28_round_ss(A, B, R) __extension__ ({ \
-  (__m128)__builtin_ia32_rcp28ss_mask((__v4sf)(__m128)(A), \
+  (__m128)__builtin_ia32_rcp28ss_round((__v4sf)(__m128)(A), \
                                       (__v4sf)(__m128)(B), \
                                       (__v4sf)_mm_setzero_ps(), \
                                       (__mmask8)-1, (R)); })
 
 #define _mm_mask_rcp28_round_ss(S, M, A, B, R) __extension__ ({ \
-  (__m128)__builtin_ia32_rcp28ss_mask((__v4sf)(__m128)(A), \
+  (__m128)__builtin_ia32_rcp28ss_round((__v4sf)(__m128)(A), \
                                       (__v4sf)(__m128)(B), \
                                       (__v4sf)(__m128)(S), \
                                       (__mmask8)(M), (R)); })
 
 #define _mm_maskz_rcp28_round_ss(M, A, B, R) __extension__ ({ \
-  (__m128)__builtin_ia32_rcp28ss_mask((__v4sf)(__m128)(A), \
+  (__m128)__builtin_ia32_rcp28ss_round((__v4sf)(__m128)(A), \
                                       (__v4sf)(__m128)(B), \
                                       (__v4sf)_mm_setzero_ps(), \
                                       (__mmask8)(M), (R)); })
 
 #define _mm_rcp28_ss(A, B) \
   _mm_rcp28_round_ss((A), (B), _MM_FROUND_CUR_DIRECTION)
 
 #define _mm_mask_rcp28_ss(S, M, A, B) \
   _mm_mask_rcp28_round_ss((S), (M), (A), (B), _MM_FROUND_CUR_DIRECTION)
 
 #define _mm_maskz_rcp28_ss(M, A, B) \
   _mm_maskz_rcp28_round_ss((M), (A), (B), _MM_FROUND_CUR_DIRECTION)
 
 #define _mm_rcp28_round_sd(A, B, R) __extension__ ({ \
-  (__m128d)__builtin_ia32_rcp28sd_mask((__v2df)(__m128d)(A), \
+  (__m128d)__builtin_ia32_rcp28sd_round((__v2df)(__m128d)(A), \
                                        (__v2df)(__m128d)(B), \
                                        (__v2df)_mm_setzero_pd(), \
                                        (__mmask8)-1, (R)); })
 
 #define _mm_mask_rcp28_round_sd(S, M, A, B, R) __extension__ ({ \
-  (__m128d)__builtin_ia32_rcp28sd_mask((__v2df)(__m128d)(A), \
+  (__m128d)__builtin_ia32_rcp28sd_round((__v2df)(__m128d)(A), \
                                        (__v2df)(__m128d)(B), \
                                        (__v2df)(__m128d)(S), \
                                        (__mmask8)(M), (R)); })
 
 #define _mm_maskz_rcp28_round_sd(M, A, B, R) __extension__ ({ \
-  (__m128d)__builtin_ia32_rcp28sd_mask((__v2df)(__m128d)(A), \
+  (__m128d)__builtin_ia32_rcp28sd_round((__v2df)(__m128d)(A), \
                                        (__v2df)(__m128d)(B), \
                                        (__v2df)_mm_setzero_pd(), \
                                        (__mmask8)(M), (R)); })
 
 #define _mm_rcp28_sd(A, B) \
   _mm_rcp28_round_sd((A), (B), _MM_FROUND_CUR_DIRECTION)
 
 #define _mm_mask_rcp28_sd(S, M, A, B) \
   _mm_mask_rcp28_round_sd((S), (M), (A), (B), _MM_FROUND_CUR_DIRECTION)
 
 #define _mm_maskz_rcp28_sd(M, A, B) \
   _mm_maskz_rcp28_round_sd((M), (A), (B), _MM_FROUND_CUR_DIRECTION)
 
 #endif // __AVX512ERINTRIN_H

cfe/trunk/lib/Headers/avx512fintrin.h

return (__m512) __builtin_ia32_rsqrt14ps512_mask ((__v16sf) __A,
                 (__v16sf)
                 _mm512_setzero_ps (),
                 (__mmask16) -1);
 }
 
 static  __inline__ __m128 __DEFAULT_FN_ATTRS
 _mm_rsqrt14_ss(__m128 __A, __m128 __B)
 {
-  return (__m128) __builtin_ia32_rsqrt14ss_mask ((__v4sf) __A,
+  return (__m128) __builtin_ia32_rsqrt14ss ((__v4sf) __A,
              (__v4sf) __B,
              (__v4sf)
              _mm_setzero_ps (),
              (__mmask8) -1);
 }
 
 static  __inline__ __m128d __DEFAULT_FN_ATTRS
 _mm_rsqrt14_sd(__m128d __A, __m128d __B)
 {
-  return (__m128d) __builtin_ia32_rsqrt14sd_mask ((__v2df) __A,
+  return (__m128d) __builtin_ia32_rsqrt14sd ((__v2df) __A,
               (__v2df) __B,
               (__v2df)
               _mm_setzero_pd (),
               (__mmask8) -1);
 }
 
 static  __inline__ __m512d __DEFAULT_FN_ATTRS
 _mm512_rcp14_pd(__m512d __A)
_mm512_rcp14_ps(__m512 __A)
   return (__m512) __builtin_ia32_rcp14ps512_mask ((__v16sf) __A,
               (__v16sf)
               _mm512_setzero_ps (),
               (__mmask16) -1);
 }
 static  __inline__ __m128 __DEFAULT_FN_ATTRS
 _mm_rcp14_ss(__m128 __A, __m128 __B)
 {
-  return (__m128) __builtin_ia32_rcp14ss_mask ((__v4sf) __A,
+  return (__m128) __builtin_ia32_rcp14ss ((__v4sf) __A,
                  (__v4sf) __B,
                  (__v4sf)
                  _mm_setzero_ps (),
                  (__mmask8) -1);
 }
 
 static  __inline__ __m128d __DEFAULT_FN_ATTRS
 _mm_rcp14_sd(__m128d __A, __m128d __B)
 {
-  return (__m128d) __builtin_ia32_rcp14sd_mask ((__v2df) __A,
+  return (__m128d) __builtin_ia32_rcp14sd ((__v2df) __A,
             (__v2df) __B,
             (__v2df)
             _mm_setzero_pd (),
             (__mmask8) -1);
 }
 
 static __inline __m512 __DEFAULT_FN_ATTRS
 _mm512_floor_ps(__m512 __A)