[AVX-512] Use scalar vfmsub/vfnmsub mask3 intrinsics instead of inverting the mask argument of a vfmadd intrinsic.

[craig.topper]

Summary

Inverting the mask argument currently fails isel and even if that was fixed it does not reflect the intended semantics of the intrinsic.

[craig.topper]

Ping

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

 TARGET_BUILTIN(__builtin_ia32_getexppd512_mask, "V8dV8dV8dUcIi","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_getexpps512_mask, "V16fV16fV16fUsIi","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_vfmaddss3_mask,  "V4fV4fV4fV4fUcIi", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_vfmaddss3_maskz, "V4fV4fV4fV4fUcIi", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_vfmaddss3_mask3, "V4fV4fV4fV4fUcIi", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_vfmaddsd3_mask,  "V2dV2dV2dV2dUcIi", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_vfmaddsd3_maskz, "V2dV2dV2dV2dUcIi", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_vfmaddsd3_mask3, "V2dV2dV2dV2dUcIi", "", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmsubsd3_mask3, "V2dV2dV2dV2dUcIi", "", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfmsubss3_mask3, "V4fV4fV4fV4fUcIi", "", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubsd3_mask3, "V2dV2dV2dV2dUcIi", "", "avx512f")
+TARGET_BUILTIN(__builtin_ia32_vfnmsubss3_mask3, "V4fV4fV4fV4fUcIi", "", "avx512f")
 TARGET_BUILTIN(__builtin_ia32_permvarhi512_mask, "V32sV32sV32sV32sUi","","avx512bw")
 TARGET_BUILTIN(__builtin_ia32_permvardf512_mask, "V8dV8dV8LLiV8dUc","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_permvardi512_mask, "V8LLiV8LLiV8LLiV8LLiUc","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_permvarsf512_mask, "V16fV16fV16iV16fUs","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_permvarsi512_mask, "V16iV16iV16iV16iUs","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_permvarqi512_mask, "V64cV64cV64cV64cULLi","","avx512vbmi")
 TARGET_BUILTIN(__builtin_ia32_permvarqi128_mask, "V16cV16cV16cV16cUs","","avx512vbmi,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_permvarqi256_mask, "V32cV32cV32cV32cUi","","avx512vbmi,avx512vl")

cfe/trunk/lib/Headers/avx512fintrin.h

#define _mm_maskz_fmsub_round_ss(U, A, B, C, R) __extension__ ({\
   (__m128)__builtin_ia32_vfmaddss3_maskz((__v4sf)(__m128)(A), \
                                          (__v4sf)(__m128)(B), \
                                          -(__v4sf)(__m128)(C), (__mmask8)(U), \
                                          (int)(R)); })
 
 static __inline__ __m128 __DEFAULT_FN_ATTRS
 _mm_mask3_fmsub_ss (__m128 __W, __m128 __X, __m128 __Y, __mmask8 __U)
 {
- return (__m128) __builtin_ia32_vfmaddss3_mask3 ((__v4sf) __W,
+ return (__m128) __builtin_ia32_vfmsubss3_mask3 ((__v4sf) __W,
           (__v4sf) __X,
-          -(__v4sf) __Y,
+          (__v4sf) __Y,
           (__mmask8) __U,
           _MM_FROUND_CUR_DIRECTION);
 }
 
 #define _mm_mask3_fmsub_round_ss(W, X, Y, U, R) __extension__ ({\
-  (__m128)__builtin_ia32_vfmaddss3_mask3((__v4sf)(__m128)(W), \
+  (__m128)__builtin_ia32_vfmsubss3_mask3((__v4sf)(__m128)(W), \
                                          (__v4sf)(__m128)(X), \
-                                         -(__v4sf)(__m128)(Y), (__mmask8)(U), \
+                                         (__v4sf)(__m128)(Y), (__mmask8)(U), \
                                          (int)(R)); })
 
 static __inline__ __m128 __DEFAULT_FN_ATTRS
 _mm_mask_fnmadd_ss (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
 {
  return (__m128) __builtin_ia32_vfmaddss3_mask ((__v4sf) __W,
           -(__v4sf) __A,
           (__v4sf) __B,
#define _mm_maskz_fnmsub_round_ss(U, A, B, C, R) __extension__ ({\
   (__m128)__builtin_ia32_vfmaddss3_maskz(-(__v4sf)(__m128)(A), \
                                          (__v4sf)(__m128)(B), \
                                          -(__v4sf)(__m128)(C), (__mmask8)(U), \
                                          _MM_FROUND_CUR_DIRECTION); })
 
 static __inline__ __m128 __DEFAULT_FN_ATTRS
 _mm_mask3_fnmsub_ss (__m128 __W, __m128 __X, __m128 __Y, __mmask8 __U)
 {
- return (__m128) __builtin_ia32_vfmaddss3_mask3 (-(__v4sf) __W,
+ return (__m128) __builtin_ia32_vfnmsubss3_mask3 ((__v4sf) __W,
           (__v4sf) __X,
-          -(__v4sf) __Y,
+          (__v4sf) __Y,
           (__mmask8) __U,
           _MM_FROUND_CUR_DIRECTION);
 }
 
 #define _mm_mask3_fnmsub_round_ss(W, X, Y, U, R) __extension__({\
-  (__m128)__builtin_ia32_vfmaddss3_mask3(-(__v4sf)(__m128)(W), \
+  (__m128)__builtin_ia32_vfnmsubss3_mask3((__v4sf)(__m128)(W), \
                                          (__v4sf)(__m128)(X), \
-                                         -(__v4sf)(__m128)(Y), (__mmask8)(U), \
+                                         (__v4sf)(__m128)(Y), (__mmask8)(U), \
                                          (int)(R)); })
 
 static __inline__ __m128d __DEFAULT_FN_ATTRS
 _mm_mask_fmadd_sd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
 {
  return (__m128d) __builtin_ia32_vfmaddsd3_mask ( (__v2df) __W,
           (__v2df) __A,
           (__v2df) __B,
#define _mm_maskz_fmsub_round_sd(U, A, B, C, R) __extension__ ({\
   (__m128d)__builtin_ia32_vfmaddsd3_maskz((__v2df)(__m128d)(A), \
                                           (__v2df)(__m128d)(B), \
                                           -(__v2df)(__m128d)(C), \
                                           (__mmask8)(U), (int)(R)); })
 
 static __inline__ __m128d __DEFAULT_FN_ATTRS
 _mm_mask3_fmsub_sd (__m128d __W, __m128d __X, __m128d __Y, __mmask8 __U)
 {
- return (__m128d) __builtin_ia32_vfmaddsd3_mask3 ((__v2df) __W,
+ return (__m128d) __builtin_ia32_vfmsubsd3_mask3 ((__v2df) __W,
           (__v2df) __X,
-          -(__v2df) __Y,
+          (__v2df) __Y,
           (__mmask8) __U,
           _MM_FROUND_CUR_DIRECTION);
 }
 
 #define _mm_mask3_fmsub_round_sd(W, X, Y, U, R) __extension__ ({\
-  (__m128d)__builtin_ia32_vfmaddsd3_mask3((__v2df)(__m128d)(W), \
+  (__m128d)__builtin_ia32_vfmsubsd3_mask3((__v2df)(__m128d)(W), \
                                           (__v2df)(__m128d)(X), \
-                                          -(__v2df)(__m128d)(Y), \
+                                          (__v2df)(__m128d)(Y), \
                                           (__mmask8)(U), (int)(R)); })
 
 static __inline__ __m128d __DEFAULT_FN_ATTRS
 _mm_mask_fnmadd_sd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
 {
  return (__m128d) __builtin_ia32_vfmaddsd3_mask ( (__v2df) __W,
           -(__v2df) __A,
           (__v2df) __B,
(__m128d)__builtin_ia32_vfmaddsd3_maskz(-(__v2df)(__m128d)(A), \
                                           (__v2df)(__m128d)(B), \
                                           -(__v2df)(__m128d)(C), \
                                           (__mmask8)(U), \
                                           _MM_FROUND_CUR_DIRECTION); })
 
 static __inline__ __m128d __DEFAULT_FN_ATTRS
 _mm_mask3_fnmsub_sd (__m128d __W, __m128d __X, __m128d __Y, __mmask8 __U)
 {
- return (__m128d) __builtin_ia32_vfmaddsd3_mask3 (-(__v2df) (__W),
+ return (__m128d) __builtin_ia32_vfnmsubsd3_mask3 ((__v2df) (__W),
           (__v2df) __X,
-          -(__v2df) (__Y),
+          (__v2df) (__Y),
           (__mmask8) __U,
           _MM_FROUND_CUR_DIRECTION);
 }
 
 #define _mm_mask3_fnmsub_round_sd(W, X, Y, U, R) __extension__({\
-  (__m128d)__builtin_ia32_vfmaddsd3_mask3(-(__v2df)(__m128d)(W), \
+  (__m128d)__builtin_ia32_vfnmsubsd3_mask3((__v2df)(__m128d)(W), \
                                           (__v2df)(__m128d)(X), \
-                                          -(__v2df)(__m128d)(Y), \
+                                          (__v2df)(__m128d)(Y), \
                                           (__mmask8)(U), (int)(R)); })
 
 #define _mm512_permutex_pd(X, C) __extension__ ({ \
   (__m512d)__builtin_shufflevector((__v8df)(__m512d)(X), \
                                    (__v8df)_mm512_undefined_pd(), \
                                    0 + (((C) >> 0) & 0x3), \
                                    0 + (((C) >> 2) & 0x3), \
                                    0 + (((C) >> 4) & 0x3), \

cfe/trunk/test/CodeGen/avx512f-builtins.c

 __m128 test_mm_maskz_fmsub_round_ss(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C){
   // CHECK-LABEL: @test_mm_maskz_fmsub_round_ss
   // CHECK: @llvm.x86.avx512.maskz.vfmadd.ss
   return _mm_maskz_fmsub_round_ss(__U, __A, __B, __C, _MM_FROUND_CUR_DIRECTION);
 }
 
 __m128 test_mm_mask3_fmsub_ss(__m128 __W, __m128 __X, __m128 __Y, __mmask8 __U){
   // CHECK-LABEL: @test_mm_mask3_fmsub_ss
-  // CHECK: @llvm.x86.avx512.mask3.vfmadd.ss
+  // CHECK: @llvm.x86.avx512.mask3.vfmsub.ss
   return _mm_mask3_fmsub_ss(__W, __X, __Y, __U);
 }
 
 __m128 test_mm_mask3_fmsub_round_ss(__m128 __W, __m128 __X, __m128 __Y, __mmask8 __U){
   // CHECK-LABEL: @test_mm_mask3_fmsub_round_ss
-  // CHECK: @llvm.x86.avx512.mask3.vfmadd.ss
+  // CHECK: @llvm.x86.avx512.mask3.vfmsub.ss
   return _mm_mask3_fmsub_round_ss(__W, __X, __Y, __U, _MM_FROUND_CUR_DIRECTION);
 }
 
 __m128 test_mm_mask_fnmadd_ss(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B){
   // CHECK-LABEL: @test_mm_mask_fnmadd_ss
   // CHECK: @llvm.x86.avx512.mask.vfmadd.ss
   return _mm_mask_fnmadd_ss(__W, __U, __A, __B);
 }

 __m128 test_mm_maskz_fnmsub_round_ss(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C){
   // CHECK-LABEL: @test_mm_maskz_fnmsub_round_ss
   // CHECK: @llvm.x86.avx512.maskz.vfmadd.ss
   return _mm_maskz_fnmsub_round_ss(__U, __A, __B, __C, _MM_FROUND_CUR_DIRECTION);
 }
 
 __m128 test_mm_mask3_fnmsub_ss(__m128 __W, __m128 __X, __m128 __Y, __mmask8 __U){
   // CHECK-LABEL: @test_mm_mask3_fnmsub_ss
-  // CHECK: @llvm.x86.avx512.mask3.vfmadd.ss
+  // CHECK: @llvm.x86.avx512.mask3.vfnmsub.ss
   return _mm_mask3_fnmsub_ss(__W, __X, __Y, __U);
 }
 
 __m128 test_mm_mask3_fnmsub_round_ss(__m128 __W, __m128 __X, __m128 __Y, __mmask8 __U){
   // CHECK-LABEL: @test_mm_mask3_fnmsub_round_ss
-  // CHECK: @llvm.x86.avx512.mask3.vfmadd.ss
+  // CHECK: @llvm.x86.avx512.mask3.vfnmsub.ss
   return _mm_mask3_fnmsub_round_ss(__W, __X, __Y, __U, _MM_FROUND_CUR_DIRECTION);
 }
 
 __m128d test_mm_mask_fmadd_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B){
   // CHECK-LABEL: @test_mm_mask_fmadd_sd
   // CHECK: @llvm.x86.avx512.mask.vfmadd.sd
   return _mm_mask_fmadd_sd(__W, __U, __A, __B);
 }

 __m128d test_mm_maskz_fmsub_round_sd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C){
   // CHECK-LABEL: @test_mm_maskz_fmsub_round_sd
   // CHECK: @llvm.x86.avx512.maskz.vfmadd.sd
   return _mm_maskz_fmsub_round_sd(__U, __A, __B, __C, _MM_FROUND_CUR_DIRECTION);
 }
 
 __m128d test_mm_mask3_fmsub_sd(__m128d __W, __m128d __X, __m128d __Y, __mmask8 __U){
   // CHECK-LABEL: @test_mm_mask3_fmsub_sd
-  // CHECK: @llvm.x86.avx512.mask3.vfmadd.sd
+  // CHECK: @llvm.x86.avx512.mask3.vfmsub.sd
   return _mm_mask3_fmsub_sd(__W, __X, __Y, __U);
 }
 
 __m128d test_mm_mask3_fmsub_round_sd(__m128d __W, __m128d __X, __m128d __Y, __mmask8 __U){
   // CHECK-LABEL: @test_mm_mask3_fmsub_round_sd
-  // CHECK: @llvm.x86.avx512.mask3.vfmadd.sd
+  // CHECK: @llvm.x86.avx512.mask3.vfmsub.sd
   return _mm_mask3_fmsub_round_sd(__W, __X, __Y, __U, _MM_FROUND_CUR_DIRECTION);
 }
 
 __m128d test_mm_mask_fnmadd_sd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B){
   // CHECK-LABEL: @test_mm_mask_fnmadd_sd
   // CHECK: @llvm.x86.avx512.mask.vfmadd.sd
   return _mm_mask_fnmadd_sd(__W, __U, __A, __B);
 }

 __m128d test_mm_maskz_fnmsub_round_sd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C){
   // CHECK-LABEL: @test_mm_maskz_fnmsub_round_sd
   // CHECK: @llvm.x86.avx512.maskz.vfmadd.sd
   return _mm_maskz_fnmsub_round_sd(__U, __A, __B, __C, _MM_FROUND_CUR_DIRECTION);
 }
 
 __m128d test_mm_mask3_fnmsub_sd(__m128d __W, __m128d __X, __m128d __Y, __mmask8 __U){
   // CHECK-LABEL: @test_mm_mask3_fnmsub_sd
-  // CHECK: @llvm.x86.avx512.mask3.vfmadd.sd
+  // CHECK: @llvm.x86.avx512.mask3.vfnmsub.sd
   return _mm_mask3_fnmsub_sd(__W, __X, __Y, __U);
 }
 
 __m128d test_mm_mask3_fnmsub_round_sd(__m128d __W, __m128d __X, __m128d __Y, __mmask8 __U){
   // CHECK-LABEL: @test_mm_mask3_fnmsub_round_sd
-  // CHECK: @llvm.x86.avx512.mask3.vfmadd.sd
+  // CHECK: @llvm.x86.avx512.mask3.vfnmsub.sd
   return _mm_mask3_fnmsub_round_sd(__W, __X, __Y, __U, _MM_FROUND_CUR_DIRECTION);
 }
 
 __m512d test_mm512_permutex_pd(__m512d __X) {
   // CHECK-LABEL: @test_mm512_permutex_pd
   // CHECK: shufflevector <8 x double> %{{.*}}, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 4, i32 4, i32 4, i32 4>
   return _mm512_permutex_pd(__X, 0);
 }