[AVX512] Add intrinsics for 512-bit insertf32x8/insertf32x4/inserti32x4.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@269617 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def
index 71423df9099a2828ba4c836148fe85abacafb01c..3878f7fe8126d3677fb941a45693913e9df307cd 100644 (file)
--- a/include/clang/Basic/BuiltinsX86.def
+++ b/include/clang/Basic/BuiltinsX86.def
@@ -2150,6 +2150,7 @@ TARGET_BUILTIN(__builtin_ia32_extractf64x2_256_mask, "V2dV4dIiV2dUc","","avx512d
 TARGET_BUILTIN(__builtin_ia32_extracti64x2_256_mask, "V2LLiV4LLiIiV2LLiUc","","avx512dq,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_extractf32x4_256_mask, "V4fV8fIiV4fUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_extracti32x4_256_mask, "V4iV8iIiV4iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_insertf32x8_mask, "V16fV16fV8fIiV16fUs","","avx512dq")
 TARGET_BUILTIN(__builtin_ia32_insertf64x2_512_mask, "V8dV8dV2dIiV8dUc","","avx512dq")
 TARGET_BUILTIN(__builtin_ia32_inserti32x8_mask, "V16iV16iV8iIiV16iUs","","avx512dq")
 TARGET_BUILTIN(__builtin_ia32_inserti64x2_512_mask, "V8LLiV8LLiV2LLiIiV8LLiUc","","avx512dq")
@@ -2159,6 +2160,8 @@ TARGET_BUILTIN(__builtin_ia32_insertf64x2_256_mask, "V4dV4dV2dIiV4dUc","","avx51
 TARGET_BUILTIN(__builtin_ia32_inserti64x2_256_mask, "V4LLiV4LLiV2LLiIiV4LLiUc","","avx512dq,avx512vl")
 TARGET_BUILTIN(__builtin_ia32_insertf32x4_256_mask, "V8fV8fV4fIiV8fUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_inserti32x4_256_mask, "V8iV8iV4iIiV8iUc","","avx512vl")
+TARGET_BUILTIN(__builtin_ia32_insertf32x4_mask, "V16fV16fV4fIiV16fUs","","avx512f")
+TARGET_BUILTIN(__builtin_ia32_inserti32x4_mask, "V16iV16iV4iIiV16iUs","","avx512f")
 TARGET_BUILTIN(__builtin_ia32_getmantpd128_mask, "V2dV2diV2dUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_getmantpd256_mask, "V4dV4diV4dUc","","avx512vl")
 TARGET_BUILTIN(__builtin_ia32_getmantps128_mask, "V4fV4fiV4fUc","","avx512vl")

diff --git a/lib/Headers/avx512dqintrin.h b/lib/Headers/avx512dqintrin.h
index ab970f149f445474d33d7420708de435a4b2bc6b..74e920fdae46575cbbf4e42e641fba751491ee6a 100644 (file)
--- a/lib/Headers/avx512dqintrin.h
+++ b/lib/Headers/avx512dqintrin.h
@@ -1027,6 +1027,30 @@ __builtin_ia32_extracti64x2_512_mask ((__v8di)( __A),\
                 (__mmask8) ( __U));\
 })
 
+#define _mm512_insertf32x8( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_insertf32x8_mask ((__v16sf)( __A),\
+                (__v8sf)( __B),\
+                ( __imm),\
+                (__v16sf) _mm512_setzero_ps (),\
+                (__mmask16) -1);\
+})
+
+#define _mm512_mask_insertf32x8( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_insertf32x8_mask ((__v16sf)( __A),\
+                (__v8sf)( __B),\
+                ( __imm),\
+                (__v16sf)( __W),\
+                (__mmask16)( __U));\
+})
+
+#define _mm512_maskz_insertf32x8( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_insertf32x8_mask ((__v16sf)( __A),\
+                (__v8sf)( __B),\
+                ( __imm),\
+                (__v16sf) _mm512_setzero_ps (),\
+                (__mmask16)( __U));\
+})
+
 #define _mm512_insertf64x2( __A, __B, __imm) __extension__ ({ \
 __builtin_ia32_insertf64x2_512_mask ((__v8df)( __A),\
                 (__v2df)( __B),\

diff --git a/lib/Headers/avx512fintrin.h b/lib/Headers/avx512fintrin.h
index d70709f0933c26c06df58dffd5e3cc6d78149115..f967f6ec8233ad0350957be3ee08f8004aac12a7 100644 (file)
--- a/lib/Headers/avx512fintrin.h
+++ b/lib/Headers/avx512fintrin.h
@@ -7427,6 +7427,54 @@ __builtin_ia32_inserti64x4_mask ((__v8di)( __A),\
                 (__mmask8)( __U));\
 })
 
+#define _mm512_insertf32x4( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_insertf32x4_mask ((__v16sf)( __A),\
+                (__v4sf)( __B),\
+                ( __imm),\
+                (__v16sf) _mm512_undefined_ps (),\
+                (__mmask16) -1);\
+})
+
+#define _mm512_mask_insertf32x4( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_insertf32x4_mask ((__v16sf)( __A),\
+                (__v4sf)( __B),\
+                ( __imm),\
+                (__v16sf)( __W),\
+                (__mmask16)( __U));\
+})
+
+#define _mm512_maskz_insertf32x4( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_insertf32x4_mask ((__v16sf)( __A),\
+                (__v4sf)( __B),\
+                ( __imm),\
+                (__v16sf) _mm512_setzero_ps (),\
+                (__mmask16)( __U));\
+})
+
+#define _mm512_inserti32x4( __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_inserti32x4_mask ((__v16si)( __A),\
+                (__v4si)( __B),\
+                ( __imm),\
+                (__v16si) _mm512_setzero_si512 (),\
+                (__mmask16) -1);\
+})
+
+#define _mm512_mask_inserti32x4( __W, __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_inserti32x4_mask ((__v16si)( __A),\
+                (__v4si)( __B),\
+                ( __imm),\
+                (__v16si)( __W),\
+                (__mmask16)( __U));\
+})
+
+#define _mm512_maskz_inserti32x4( __U, __A, __B, __imm) __extension__ ({ \
+__builtin_ia32_inserti32x4_mask ((__v16si)( __A),\
+                (__v4si)( __B),\
+                ( __imm),\
+                (__v16si) _mm512_setzero_si512 (),\
+                (__mmask16)( __U));\
+})
+
 #define _mm512_getmant_round_pd( __A, __B, __C, __R) __extension__ ({ \
 __builtin_ia32_getmantpd512_mask ((__v8df)( __A),\
                  (__C << 2) |( __B),\

diff --git a/test/CodeGen/avx512dq-builtins.c b/test/CodeGen/avx512dq-builtins.c
index f9cc917a9c2740fa09f638a6a57c26af689b96dc..f9bc504a5e06fdc29ad816dddb6042f969193dbd 100644 (file)
--- a/test/CodeGen/avx512dq-builtins.c
+++ b/test/CodeGen/avx512dq-builtins.c
@@ -946,6 +946,24 @@ __m128i test_mm512_maskz_extracti64x2_epi64(__mmask8 __U, __m512i __A) {
   return _mm512_maskz_extracti64x2_epi64(__U, __A, 3); 
 }
 
+__m512 test_mm512_insertf32x8(__m512 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm512_insertf32x8
+  // CHECK: @llvm.x86.avx512.mask.insertf32x8
+  return _mm512_insertf32x8(__A, __B, 1); 
+}
+
+__m512 test_mm512_mask_insertf32x8(__m512 __W, __mmask16 __U, __m512 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm512_mask_insertf32x8
+  // CHECK: @llvm.x86.avx512.mask.insertf32x8
+  return _mm512_mask_insertf32x8(__W, __U, __A, __B, 1); 
+}
+
+__m512 test_mm512_maskz_insertf32x8(__mmask16 __U, __m512 __A, __m256 __B) {
+  // CHECK-LABEL: @test_mm512_maskz_insertf32x8
+  // CHECK: @llvm.x86.avx512.mask.insertf32x8
+  return _mm512_maskz_insertf32x8(__U, __A, __B, 1); 
+}
+
 __m512d test_mm512_insertf64x2(__m512d __A, __m128d __B) {
   // CHECK-LABEL: @test_mm512_insertf64x2
   // CHECK: @llvm.x86.avx512.mask.insertf64x2

diff --git a/test/CodeGen/avx512f-builtins.c b/test/CodeGen/avx512f-builtins.c
index 178b0fb0c10c7f83c4f533803abceb7db6e6ff32..acc04f6d912983c59a262a723d21a9bc6efb7591 100644 (file)
--- a/test/CodeGen/avx512f-builtins.c
+++ b/test/CodeGen/avx512f-builtins.c
@@ -4483,6 +4483,42 @@ __m512i test_mm512_maskz_inserti64x4(__mmask8 __U, __m512i __A, __m256i __B) {
   return _mm512_maskz_inserti64x4(__U, __A, __B, 1); 
 }
 
+__m512 test_mm512_insertf32x4(__m512 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm512_insertf32x4
+  // CHECK: @llvm.x86.avx512.mask.insertf32x4
+  return _mm512_insertf32x4(__A, __B, 1);
+}
+
+__m512 test_mm512_mask_insertf32x4(__m512 __W, __mmask16 __U, __m512 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm512_mask_insertf32x4
+  // CHECK: @llvm.x86.avx512.mask.insertf32x4
+  return _mm512_mask_insertf32x4(__W, __U, __A, __B, 1); 
+}
+
+__m512 test_mm512_maskz_insertf32x4(__mmask16 __U, __m512 __A, __m128 __B) {
+  // CHECK-LABEL: @test_mm512_maskz_insertf32x4
+  // CHECK: @llvm.x86.avx512.mask.insertf32x4
+  return _mm512_maskz_insertf32x4(__U, __A, __B, 1); 
+}
+
+__m512i test_mm512_inserti32x4(__m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_inserti32x4
+  // CHECK: @llvm.x86.avx512.mask.inserti32x4
+  return _mm512_inserti32x4(__A, __B, 1); 
+}
+
+__m512i test_mm512_mask_inserti32x4(__m512i __W, __mmask16 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_mask_inserti32x4
+  // CHECK: @llvm.x86.avx512.mask.inserti32x4
+  return _mm512_mask_inserti32x4(__W, __U, __A, __B, 1); 
+}
+
+__m512i test_mm512_maskz_inserti32x4(__mmask16 __U, __m512i __A, __m128i __B) {
+  // CHECK-LABEL: @test_mm512_maskz_inserti32x4
+  // CHECK: @llvm.x86.avx512.mask.inserti32x4
+  return _mm512_maskz_inserti32x4(__U, __A, __B, 1); 
+}
+
 __m512d test_mm512_getmant_round_pd(__m512d __A) {
   // CHECK-LABEL: @test_mm512_getmant_round_pd
   // CHECK: @llvm.x86.avx512.mask.getmant.pd.512