[AVX-512] Replace 512-bit masked packss/packus builtins and replace with new unmasked...

These new unmasked builtins will enable us to easily support optimizing these builtins in InstCombine in the backend.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@295291 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def
index cb73faca282543713bfde204c2e174737f6c0e12..e147c8fb86b5461c38fe6e5d1a7e628e42a98530 100644 (file)
--- a/include/clang/Basic/BuiltinsX86.def
+++ b/include/clang/Basic/BuiltinsX86.def
@@ -1068,10 +1068,10 @@ TARGET_BUILTIN(__builtin_ia32_ucmpw512_mask, "UiV32sV32sIiUi", "", "avx512bw")
 
 TARGET_BUILTIN(__builtin_ia32_pabsb512_mask, "V64cV64cV64cULLi", "", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_pabsw512_mask, "V32sV32sV32sUi", "", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_packssdw512_mask, "V32sV16iV16iV32sUi", "", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_packsswb512_mask, "V64cV32sV32sV64cULLi", "", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_packusdw512_mask, "V32sV16iV16iV32sUi", "", "avx512bw")
-TARGET_BUILTIN(__builtin_ia32_packuswb512_mask, "V64cV32sV32sV64cULLi", "", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_packssdw512, "V32sV16iV16i", "", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_packsswb512, "V64cV32sV32s", "", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_packusdw512, "V32sV16iV16i", "", "avx512bw")
+TARGET_BUILTIN(__builtin_ia32_packuswb512, "V64cV32sV32s", "", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_paddsb512_mask, "V64cV64cV64cV64cULLi", "", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_paddsw512_mask, "V32sV32sV32sV32sUi", "", "avx512bw")
 TARGET_BUILTIN(__builtin_ia32_paddusb512_mask, "V64cV64cV64cV64cULLi", "", "avx512bw")

diff --git a/lib/Headers/avx512bwintrin.h b/lib/Headers/avx512bwintrin.h
index 629dc8611a7f207e90327a7455a9e4c8e4e448fd..41958b7214e2f97b8ab4d7d16e3ba02e51982199 100644 (file)
--- a/lib/Headers/avx512bwintrin.h
+++ b/lib/Headers/avx512bwintrin.h
@@ -504,115 +504,91 @@ _mm512_maskz_abs_epi16 (__mmask32 __U, __m512i __A)
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_packs_epi32 (__m512i __A, __m512i __B)
+_mm512_packs_epi32(__m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A,
-              (__v16si) __B,
-              (__v32hi) _mm512_setzero_hi(),
-              (__mmask32) -1);
+  return (__m512i)__builtin_ia32_packssdw512((__v16si)__A, (__v16si)__B);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_maskz_packs_epi32 (__mmask32 __M, __m512i __A, __m512i __B)
+_mm512_maskz_packs_epi32(__mmask32 __M, __m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A,
-              (__v16si) __B,
-              (__v32hi) _mm512_setzero_hi(),
-              __M);
+  return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M,
+                                       (__v32hi)_mm512_packs_epi32(__A, __B),
+                                       (__v32hi)_mm512_setzero_hi());
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask_packs_epi32 (__m512i __W, __mmask32 __M, __m512i __A,
-       __m512i __B)
+_mm512_mask_packs_epi32(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A,
-              (__v16si) __B,
-              (__v32hi) __W,
-              __M);
+  return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M,
+                                       (__v32hi)_mm512_packs_epi32(__A, __B),
+                                       (__v32hi)__W);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_packs_epi16 (__m512i __A, __m512i __B)
+_mm512_packs_epi16(__m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A,
-              (__v32hi) __B,
-              (__v64qi) _mm512_setzero_qi(),
-              (__mmask64) -1);
+  return (__m512i)__builtin_ia32_packsswb512((__v32hi)__A, (__v32hi) __B);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask_packs_epi16 (__m512i __W, __mmask64 __M, __m512i __A,
-       __m512i __B)
+_mm512_mask_packs_epi16(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A,
-              (__v32hi) __B,
-              (__v64qi) __W,
-              (__mmask64) __M);
+  return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M,
+                                        (__v64qi)_mm512_packs_epi16(__A, __B),
+                                        (__v64qi)__W);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_maskz_packs_epi16 (__mmask64 __M, __m512i __A, __m512i __B)
+_mm512_maskz_packs_epi16(__mmask64 __M, __m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A,
-              (__v32hi) __B,
-              (__v64qi) _mm512_setzero_qi(),
-              __M);
+  return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M,
+                                        (__v64qi)_mm512_packs_epi16(__A, __B),
+                                        (__v64qi)_mm512_setzero_qi());
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_packus_epi32 (__m512i __A, __m512i __B)
+_mm512_packus_epi32(__m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A,
-              (__v16si) __B,
-              (__v32hi) _mm512_setzero_hi(),
-              (__mmask32) -1);
+  return (__m512i)__builtin_ia32_packusdw512((__v16si) __A, (__v16si) __B);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_maskz_packus_epi32 (__mmask32 __M, __m512i __A, __m512i __B)
+_mm512_maskz_packus_epi32(__mmask32 __M, __m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A,
-              (__v16si) __B,
-              (__v32hi) _mm512_setzero_hi(),
-              __M);
+  return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M,
+                                       (__v32hi)_mm512_packus_epi32(__A, __B),
+                                       (__v32hi)_mm512_setzero_hi());
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask_packus_epi32 (__m512i __W, __mmask32 __M, __m512i __A,
-        __m512i __B)
+_mm512_mask_packus_epi32(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A,
-              (__v16si) __B,
-              (__v32hi) __W,
-              __M);
+  return (__m512i)__builtin_ia32_selectw_512((__mmask32)__M,
+                                       (__v32hi)_mm512_packus_epi32(__A, __B),
+                                       (__v32hi)__W);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_packus_epi16 (__m512i __A, __m512i __B)
+_mm512_packus_epi16(__m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A,
-              (__v32hi) __B,
-              (__v64qi) _mm512_setzero_qi(),
-              (__mmask64) -1);
+  return (__m512i)__builtin_ia32_packuswb512((__v32hi) __A, (__v32hi) __B);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_mask_packus_epi16 (__m512i __W, __mmask64 __M, __m512i __A,
-        __m512i __B)
+_mm512_mask_packus_epi16(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A,
-              (__v32hi) __B,
-              (__v64qi) __W,
-              (__mmask64) __M);
+  return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M,
+                                        (__v64qi)_mm512_packus_epi16(__A, __B),
+                                        (__v64qi)__W);
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS
-_mm512_maskz_packus_epi16 (__mmask64 __M, __m512i __A, __m512i __B)
+_mm512_maskz_packus_epi16(__mmask64 __M, __m512i __A, __m512i __B)
 {
-  return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A,
-              (__v32hi) __B,
-              (__v64qi) _mm512_setzero_qi(),
-              (__mmask64) __M);
+  return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M,
+                                        (__v64qi)_mm512_packus_epi16(__A, __B),
+                                        (__v64qi)_mm512_setzero_qi());
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS

diff --git a/test/CodeGen/avx512bw-builtins.c b/test/CodeGen/avx512bw-builtins.c
index b4dfb5ccb9ece777ca61dc8f8f95888dc49dc419..8283365acc7fb0eba5635de8fb6e3353d14d3901 100644 (file)
--- a/test/CodeGen/avx512bw-builtins.c
+++ b/test/CodeGen/avx512bw-builtins.c
@@ -510,62 +510,70 @@ __m512i test_mm512_maskz_abs_epi16(__mmask32 __U, __m512i __A) {
 }
 __m512i test_mm512_packs_epi32(__m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_packs_epi32
-  // CHECK: @llvm.x86.avx512.mask.packssdw.512
+  // CHECK: @llvm.x86.avx512.packssdw.512
   return _mm512_packs_epi32(__A,__B); 
 }
 __m512i test_mm512_maskz_packs_epi32(__mmask32 __M, __m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_maskz_packs_epi32
-  // CHECK: @llvm.x86.avx512.mask.packssdw.512
+  // CHECK: @llvm.x86.avx512.packssdw.512
+  // CHECK: select <32 x i1> %{{.*}}, <32 x i16> %{{.*}}, <32 x i16> %{{.*}}
   return _mm512_maskz_packs_epi32(__M,__A,__B); 
 }
 __m512i test_mm512_mask_packs_epi32(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_mask_packs_epi32
-  // CHECK: @llvm.x86.avx512.mask.packssdw.512
+  // CHECK: @llvm.x86.avx512.packssdw.512
+  // CHECK: select <32 x i1> %{{.*}}, <32 x i16> %{{.*}}, <32 x i16> %{{.*}}
   return _mm512_mask_packs_epi32(__W,__M,__A,__B); 
 }
 __m512i test_mm512_packs_epi16(__m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_packs_epi16
-  // CHECK: @llvm.x86.avx512.mask.packsswb.512
+  // CHECK: @llvm.x86.avx512.packsswb.512
   return _mm512_packs_epi16(__A,__B); 
 }
 __m512i test_mm512_mask_packs_epi16(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_mask_packs_epi16
-  // CHECK: @llvm.x86.avx512.mask.packsswb.512
+  // CHECK: @llvm.x86.avx512.packsswb.512
+  // CHECK: select <64 x i1> %{{.*}}, <64 x i8> %{{.*}}, <64 x i8> %{{.*}}
   return _mm512_mask_packs_epi16(__W,__M,__A,__B); 
 }
 __m512i test_mm512_maskz_packs_epi16(__mmask64 __M, __m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_maskz_packs_epi16
-  // CHECK: @llvm.x86.avx512.mask.packsswb.512
+  // CHECK: @llvm.x86.avx512.packsswb.512
+  // CHECK: select <64 x i1> %{{.*}}, <64 x i8> %{{.*}}, <64 x i8> %{{.*}}
   return _mm512_maskz_packs_epi16(__M,__A,__B); 
 }
 __m512i test_mm512_packus_epi32(__m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_packus_epi32
-  // CHECK: @llvm.x86.avx512.mask.packusdw.512
+  // CHECK: @llvm.x86.avx512.packusdw.512
   return _mm512_packus_epi32(__A,__B); 
 }
 __m512i test_mm512_maskz_packus_epi32(__mmask32 __M, __m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_maskz_packus_epi32
-  // CHECK: @llvm.x86.avx512.mask.packusdw.512
+  // CHECK: @llvm.x86.avx512.packusdw.512
+  // CHECK: select <32 x i1> %{{.*}}, <32 x i16> %{{.*}}, <32 x i16> %{{.*}}
   return _mm512_maskz_packus_epi32(__M,__A,__B); 
 }
 __m512i test_mm512_mask_packus_epi32(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_mask_packus_epi32
-  // CHECK: @llvm.x86.avx512.mask.packusdw.512
+  // CHECK: @llvm.x86.avx512.packusdw.512
+  // CHECK: select <32 x i1> %{{.*}}, <32 x i16> %{{.*}}, <32 x i16> %{{.*}}
   return _mm512_mask_packus_epi32(__W,__M,__A,__B); 
 }
 __m512i test_mm512_packus_epi16(__m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_packus_epi16
-  // CHECK: @llvm.x86.avx512.mask.packuswb.512
+  // CHECK: @llvm.x86.avx512.packuswb.512
   return _mm512_packus_epi16(__A,__B); 
 }
 __m512i test_mm512_mask_packus_epi16(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_mask_packus_epi16
-  // CHECK: @llvm.x86.avx512.mask.packuswb.512
+  // CHECK: @llvm.x86.avx512.packuswb.512
+  // CHECK: select <64 x i1> %{{.*}}, <64 x i8> %{{.*}}, <64 x i8> %{{.*}}
   return _mm512_mask_packus_epi16(__W,__M,__A,__B); 
 }
 __m512i test_mm512_maskz_packus_epi16(__mmask64 __M, __m512i __A, __m512i __B) {
   // CHECK-LABEL: @test_mm512_maskz_packus_epi16
-  // CHECK: @llvm.x86.avx512.mask.packuswb.512
+  // CHECK: @llvm.x86.avx512.packuswb.512
+  // CHECK: select <64 x i1> %{{.*}}, <64 x i8> %{{.*}}, <64 x i8> %{{.*}}
   return _mm512_maskz_packus_epi16(__M,__A,__B); 
 }
 __m512i test_mm512_adds_epi8(__m512i __A, __m512i __B) {