[Clang][AVX512][Builtin] Adding supporting to intrinsics of cvt{b|d|q}2mask{128|256...

Differential Revision: http://reviews.llvm.org/D19009

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@266188 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def
index bfd2806152bc820ab90df631af95254d0767d643..e3cad39ed5aea0ce06f4d16dae882313ec928a7f 100644 (file)
--- a/include/clang/Basic/BuiltinsX86.def
+++ b/include/clang/Basic/BuiltinsX86.def
@@ -1962,6 +1962,27 @@ TARGET_BUILTIN(__builtin_ia32_rsqrt14pd128_mask, "V2dV2dV2dUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_rsqrt14pd256_mask, "V4dV4dV4dUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_rsqrt14ps128_mask, "V4fV4fV4fUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_rsqrt14ps256_mask, "V8fV8fV8fUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtb2mask512, "ULLiV64c","","avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2b512, "V64cULLi","","avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2w512, "V32sUi","","avx512bw")
+TARGET_BUILTIN(__builtin_ia32_cvtd2mask512, "UsV16i","","avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2d512, "V16iUs","","avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2q512, "V8LLiUc","","avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtq2mask512, "UcV8LLi","","avx512dq")
+TARGET_BUILTIN(__builtin_ia32_cvtb2mask128, "UsV16c","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtb2mask256, "UiV32c","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2b128, "V16cUs","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2b256, "V32cUi","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2w128, "V8sUc","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2w256, "V16sUs","","avx512bw,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtd2mask128, "UcV4i","","avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtd2mask256, "UcV8i","","avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2d128, "V4iUc","","avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2d256, "V8iUc","","avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2q128, "V2LLiUc","","avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtmask2q256, "V4LLiUc","","avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtq2mask128, "UcV2LLi","","avx512dq,avx512vl")
+TARGET_BUILTIN(__builtin_ia32_cvtq2mask256, "UcV4LLi","","avx512dq,avx512vl")
 
 #undef BUILTIN
 #undef TARGET_BUILTIN

diff --git a/lib/Headers/avx512bwintrin.h b/lib/Headers/avx512bwintrin.h
index 90f685d5c3ecc0348f37e5723735fa0ad03e1505..688bf30aafa82ba7fed71807a5b5bc5a594fc16d 100644 (file)
--- a/lib/Headers/avx512bwintrin.h
+++ b/lib/Headers/avx512bwintrin.h
@@ -2057,6 +2057,24 @@ _mm512_mask_testn_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B)
              (__v32hi) __B, __U);
 }
 
+static __inline__ __mmask64 __DEFAULT_FN_ATTRS
+_mm512_movepi8_mask (__m512i __A)
+{
+  return (__mmask64) __builtin_ia32_cvtb2mask512 ((__v64qi) __A);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_movm_epi8 (__mmask64 __A)
+{
+  return (__m512i) __builtin_ia32_cvtmask2b512 (__A);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_movm_epi16 (__mmask32 __A)
+{
+  return (__m512i) __builtin_ia32_cvtmask2w512 (__A);
+}
+
 
 #undef __DEFAULT_FN_ATTRS
 

diff --git a/lib/Headers/avx512dqintrin.h b/lib/Headers/avx512dqintrin.h
index afee4903ba77e97f29db20871472c5950fcbec0c..a3c3d86ac03af8dd161e67df5a092d2c0cf0ec24 100644 (file)
--- a/lib/Headers/avx512dqintrin.h
+++ b/lib/Headers/avx512dqintrin.h
@@ -773,6 +773,31 @@ _mm512_maskz_cvtepu64_ps (__mmask8 __U, __m512i __A) {
   (__m512) __builtin_ia32_reduceps512_mask ((__v16sf) __A, __B,      \
                (__v16sf) _mm512_setzero_ps(), (__mmask16) __U, __R);})
 
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS
+_mm512_movepi32_mask (__m512i __A)
+{
+  return (__mmask16) __builtin_ia32_cvtd2mask512 ((__v16si) __A);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_movm_epi32 (__mmask16 __A)
+{
+  return (__m512i) __builtin_ia32_cvtmask2d512 (__A);
+}
+
+static __inline__ __m512i __DEFAULT_FN_ATTRS
+_mm512_movm_epi64 (__mmask8 __A)
+{
+  return (__m512i) __builtin_ia32_cvtmask2q512 (__A);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm512_movepi64_mask (__m512i __A)
+{
+  return (__mmask8) __builtin_ia32_cvtq2mask512 ((__v8di) __A);
+}
+
+
 #undef __DEFAULT_FN_ATTRS
 
 #endif

diff --git a/lib/Headers/avx512vlbwintrin.h b/lib/Headers/avx512vlbwintrin.h
index 57511762499a4de3f3da402edba747bc090b2174..511e05e68fd56e74ce4a6f891f8a8cf6a0c26b3c 100644 (file)
--- a/lib/Headers/avx512vlbwintrin.h
+++ b/lib/Headers/avx512vlbwintrin.h
@@ -3172,6 +3172,42 @@ _mm256_mask_testn_epi16_mask (__mmask16 __U, __m256i __A, __m256i __B)
              (__v16hi) __B, __U);
 }
 
+static __inline__ __mmask16 __DEFAULT_FN_ATTRS
+_mm_movepi8_mask (__m128i __A)
+{
+  return (__mmask16) __builtin_ia32_cvtb2mask128 ((__v16qi) __A);
+}
+
+static __inline__ __mmask32 __DEFAULT_FN_ATTRS
+_mm256_movepi8_mask (__m256i __A)
+{
+  return (__mmask32) __builtin_ia32_cvtb2mask256 ((__v32qi) __A);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_movm_epi8 (__mmask16 __A)
+{
+  return (__m128i) __builtin_ia32_cvtmask2b128 (__A);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_movm_epi8 (__mmask32 __A)
+{
+  return (__m256i) __builtin_ia32_cvtmask2b256 (__A);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_movm_epi16 (__mmask8 __A)
+{
+  return (__m128i) __builtin_ia32_cvtmask2w128 (__A);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_movm_epi16 (__mmask16 __A)
+{
+  return (__m256i) __builtin_ia32_cvtmask2w256 (__A);
+}
+
 
 
 #undef __DEFAULT_FN_ATTRS

diff --git a/lib/Headers/avx512vldqintrin.h b/lib/Headers/avx512vldqintrin.h
index dfd858e013da3fd49b536006073bb7ec3fe792de..1497526e54ea178b31264af8a8096a63d1e6e8ba 100644 (file)
--- a/lib/Headers/avx512vldqintrin.h
+++ b/lib/Headers/avx512vldqintrin.h
@@ -948,6 +948,54 @@ _mm256_maskz_cvtepu64_ps (__mmask8 __U, __m256i __A) {
   (__m256) __builtin_ia32_reduceps256_mask ((__v8sf) __A, __B,     \
                 (__v8sf) _mm256_setzero_ps(), (__mmask8) __U); })
 
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_movepi32_mask (__m128i __A)
+{
+  return (__mmask8) __builtin_ia32_cvtd2mask128 ((__v4si) __A);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_movepi32_mask (__m256i __A)
+{
+  return (__mmask8) __builtin_ia32_cvtd2mask256 ((__v8si) __A);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_movm_epi32 (__mmask8 __A)
+{
+  return (__m128i) __builtin_ia32_cvtmask2d128 (__A);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_movm_epi32 (__mmask8 __A)
+{
+  return (__m256i) __builtin_ia32_cvtmask2d256 (__A);
+}
+
+static __inline__ __m128i __DEFAULT_FN_ATTRS
+_mm_movm_epi64 (__mmask8 __A)
+{
+  return (__m128i) __builtin_ia32_cvtmask2q128 (__A);
+}
+
+static __inline__ __m256i __DEFAULT_FN_ATTRS
+_mm256_movm_epi64 (__mmask8 __A)
+{
+  return (__m256i) __builtin_ia32_cvtmask2q256 (__A);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm_movepi64_mask (__m128i __A)
+{
+  return (__mmask8) __builtin_ia32_cvtq2mask128 ((__v2di) __A);
+}
+
+static __inline__ __mmask8 __DEFAULT_FN_ATTRS
+_mm256_movepi64_mask (__m256i __A)
+{
+  return (__mmask8) __builtin_ia32_cvtq2mask256 ((__v4di) __A);
+}
+
 #undef __DEFAULT_FN_ATTRS
 
 #endif

diff --git a/test/CodeGen/avx512bw-builtins.c b/test/CodeGen/avx512bw-builtins.c
index 06f2e360bf622bb8b988500098cf3453bb981ebf..246bccc5e332f267b461f29b25d1d971f7f7af6f 100644 (file)
--- a/test/CodeGen/avx512bw-builtins.c
+++ b/test/CodeGen/avx512bw-builtins.c
@@ -1404,3 +1404,22 @@ __mmask32 test_mm512_mask_testn_epi16_mask(__mmask32 __U, __m512i __A, __m512i _
   return _mm512_mask_testn_epi16_mask(__U, __A, __B); 
 }
 
+__mmask64 test_mm512_movepi8_mask(__m512i __A) {
+  // CHECK-LABEL: @test_mm512_movepi8_mask
+  // CHECK: @llvm.x86.avx512.cvtb2mask.512
+  return _mm512_movepi8_mask(__A); 
+}
+
+__m512i test_mm512_movm_epi8(__mmask64 __A) {
+  // CHECK-LABEL: @test_mm512_movm_epi8
+  // CHECK: @llvm.x86.avx512.cvtmask2b.512
+  return _mm512_movm_epi8(__A); 
+}
+
+__m512i test_mm512_movm_epi16(__mmask32 __A) {
+  // CHECK-LABEL: @test_mm512_movm_epi16
+  // CHECK: @llvm.x86.avx512.cvtmask2w.512
+  return _mm512_movm_epi16(__A); 
+}
+
+

diff --git a/test/CodeGen/avx512dq-builtins.c b/test/CodeGen/avx512dq-builtins.c
index fc09a28ac11c6c014a8e64f59a225cee6873bca8..0be0e54488af01f1204824c6e97dd140f7b0e6df 100644 (file)
--- a/test/CodeGen/avx512dq-builtins.c
+++ b/test/CodeGen/avx512dq-builtins.c
@@ -743,3 +743,27 @@ __m512 test_mm512_maskz_reduce_round_ps(__mmask16 __U, __m512 __A) {
   return _mm512_maskz_reduce_round_ps(__U, __A, 4, 8); 
 }
 
+__mmask16 test_mm512_movepi32_mask(__m512i __A) {
+  // CHECK-LABEL: @test_mm512_movepi32_mask
+  // CHECK: @llvm.x86.avx512.cvtd2mask.512
+  return _mm512_movepi32_mask(__A); 
+}
+
+__m512i test_mm512_movm_epi32(__mmask16 __A) {
+  // CHECK-LABEL: @test_mm512_movm_epi32
+  // CHECK: @llvm.x86.avx512.cvtmask2d.512
+  return _mm512_movm_epi32(__A); 
+}
+
+__m512i test_mm512_movm_epi64(__mmask8 __A) {
+  // CHECK-LABEL: @test_mm512_movm_epi64
+  // CHECK: @llvm.x86.avx512.cvtmask2q.512
+  return _mm512_movm_epi64(__A); 
+}
+
+__mmask8 test_mm512_movepi64_mask(__m512i __A) {
+  // CHECK-LABEL: @test_mm512_movepi64_mask
+  // CHECK: @llvm.x86.avx512.cvtq2mask.512
+  return _mm512_movepi64_mask(__A); 
+}
+

diff --git a/test/CodeGen/avx512vlbw-builtins.c b/test/CodeGen/avx512vlbw-builtins.c
index b4836a71874587df8ba0129019dddd9db1462dc1..399ad0b4af95d82eae8d80d5a2c03325fc75ec36 100644 (file)
--- a/test/CodeGen/avx512vlbw-builtins.c
+++ b/test/CodeGen/avx512vlbw-builtins.c
@@ -2172,3 +2172,41 @@ __mmask16 test_mm256_mask_testn_epi16_mask(__mmask16 __U, __m256i __A, __m256i _
   return _mm256_mask_testn_epi16_mask(__U, __A, __B); 
 }
 
+__mmask16 test_mm_movepi8_mask(__m128i __A) {
+  // CHECK-LABEL: @test_mm_movepi8_mask
+  // CHECK: @llvm.x86.avx512.cvtb2mask.128
+  return _mm_movepi8_mask(__A); 
+}
+
+__mmask32 test_mm256_movepi8_mask(__m256i __A) {
+  // CHECK-LABEL: @test_mm256_movepi8_mask
+  // CHECK: @llvm.x86.avx512.cvtb2mask.256
+  return _mm256_movepi8_mask(__A); 
+}
+
+__m128i test_mm_movm_epi8(__mmask16 __A) {
+  // CHECK-LABEL: @test_mm_movm_epi8
+  // CHECK: @llvm.x86.avx512.cvtmask2b.128
+  return _mm_movm_epi8(__A); 
+}
+
+__m256i test_mm256_movm_epi8(__mmask32 __A) {
+  // CHECK-LABEL: @test_mm256_movm_epi8
+  // CHECK: @llvm.x86.avx512.cvtmask2b.256
+  return _mm256_movm_epi8(__A); 
+}
+
+__m128i test_mm_movm_epi16(__mmask8 __A) {
+  // CHECK-LABEL: @test_mm_movm_epi16
+  // CHECK: @llvm.x86.avx512.cvtmask2w.128
+  return _mm_movm_epi16(__A); 
+}
+
+__m256i test_mm256_movm_epi16(__mmask16 __A) {
+  // CHECK-LABEL: @test_mm256_movm_epi16
+  // CHECK: @llvm.x86.avx512.cvtmask2w.256
+  return _mm256_movm_epi16(__A); 
+}
+
+
+

diff --git a/test/CodeGen/avx512vldq-builtins.c b/test/CodeGen/avx512vldq-builtins.c
index 69bdc7a11045537606d3dd02022e4146fc5992d4..5e765f93774d69c4ad98d1edc4d67760f53b9829 100644 (file)
--- a/test/CodeGen/avx512vldq-builtins.c
+++ b/test/CodeGen/avx512vldq-builtins.c
@@ -808,3 +808,52 @@ __m256 test_mm256_maskz_reduce_ps(__mmask8 __U, __m256 __A) {
   // CHECK: @llvm.x86.avx512.mask.reduce.ps.256
   return _mm256_maskz_reduce_ps(__U, __A, 4); 
 }
+
+__mmask8 test_mm_movepi32_mask(__m128i __A) {
+  // CHECK-LABEL: @test_mm_movepi32_mask
+  // CHECK: @llvm.x86.avx512.cvtd2mask.128
+  return _mm_movepi32_mask(__A); 
+}
+
+__mmask8 test_mm256_movepi32_mask(__m256i __A) {
+  // CHECK-LABEL: @test_mm256_movepi32_mask
+  // CHECK: @llvm.x86.avx512.cvtd2mask.256
+  return _mm256_movepi32_mask(__A); 
+}
+
+__m128i test_mm_movm_epi32(__mmask8 __A) {
+  // CHECK-LABEL: @test_mm_movm_epi32
+  // CHECK: @llvm.x86.avx512.cvtmask2d.128
+  return _mm_movm_epi32(__A); 
+}
+
+__m256i test_mm256_movm_epi32(__mmask8 __A) {
+  // CHECK-LABEL: @test_mm256_movm_epi32
+  // CHECK: @llvm.x86.avx512.cvtmask2d.256
+  return _mm256_movm_epi32(__A); 
+}
+
+__m128i test_mm_movm_epi64(__mmask8 __A) {
+  // CHECK-LABEL: @test_mm_movm_epi64
+  // CHECK: @llvm.x86.avx512.cvtmask2q.128
+  return _mm_movm_epi64(__A); 
+}
+
+__m256i test_mm256_movm_epi64(__mmask8 __A) {
+  // CHECK-LABEL: @test_mm256_movm_epi64
+  // CHECK: @llvm.x86.avx512.cvtmask2q.256
+  return _mm256_movm_epi64(__A); 
+}
+
+__mmask8 test_mm_movepi64_mask(__m128i __A) {
+  // CHECK-LABEL: @test_mm_movepi64_mask
+  // CHECK: @llvm.x86.avx512.cvtq2mask.128
+  return _mm_movepi64_mask(__A); 
+}
+
+__mmask8 test_mm256_movepi64_mask(__m256i __A) {
+  // CHECK-LABEL: @test_mm256_movepi64_mask
+  // CHECK: @llvm.x86.avx512.cvtq2mask.256
+  return _mm256_movepi64_mask(__A); 
+}
+