[x86][AVX512] Lowering kunpack intrinsics to LLVM IR

This patch, together with a matching llvm patch (https://reviews.llvm.org/D39720), implements the lowering of X86 kunpack intrinsics to IR.

Differential Revision: https://reviews.llvm.org/D39719

Change-Id: Id5d3cb394ad33b98be79a6783d1d15569e2b798d

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@319777 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Headers/avx512bwintrin.h b/lib/Headers/avx512bwintrin.h
index 064300a48798e5a55eb626df8385d7b29100fe0b..3ff0e3aafdcc294e67d70056e65d2656b1906f86 100644 (file)
--- a/lib/Headers/avx512bwintrin.h
+++ b/lib/Headers/avx512bwintrin.h
@@ -1854,15 +1854,13 @@ _mm512_maskz_set1_epi8 (__mmask64 __M, char __A)
 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
 _mm512_kunpackd (__mmask64 __A, __mmask64 __B)
 {
-  return (__mmask64) __builtin_ia32_kunpckdi ((__mmask64) __A,
-                (__mmask64) __B);
+  return (__mmask64)  (( __A  & 0xFFFFFFFF) | ( __B << 32));
 }
 
 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
 _mm512_kunpackw (__mmask32 __A, __mmask32 __B)
 {
-  return (__mmask32) __builtin_ia32_kunpcksi ((__mmask32) __A,
-                (__mmask32) __B);
+return (__mmask32)  (( __A  & 0xFFFF) | ( __B << 16));
 }
 
 static __inline__ __m512i __DEFAULT_FN_ATTRS

diff --git a/lib/Headers/avx512fintrin.h b/lib/Headers/avx512fintrin.h
index f5137428ba3f7d7613bb73158d86da4a91becb9b..d34f0b1327ae7035194c73715ee3c9069658cb00 100644 (file)
--- a/lib/Headers/avx512fintrin.h
+++ b/lib/Headers/avx512fintrin.h
@@ -8787,7 +8787,7 @@ _mm512_kortestz (__mmask16 __A, __mmask16 __B)
 static __inline__ __mmask16 __DEFAULT_FN_ATTRS
 _mm512_kunpackb (__mmask16 __A, __mmask16 __B)
 {
-  return (__mmask16) __builtin_ia32_kunpckhi ((__mmask16) __A, (__mmask16) __B);
+  return (__mmask16)  (( __A  & 0xFF) | ( __B << 8));
 }
 
 static __inline__ __mmask16 __DEFAULT_FN_ATTRS

diff --git a/test/CodeGen/avx512bw-builtins.c b/test/CodeGen/avx512bw-builtins.c
index 190387b2d97a6191a074d9889b07d8571cffd0ca..3160a6667c0003d393ff0ee8ccd145c59d302457 100644 (file)
--- a/test/CodeGen/avx512bw-builtins.c
+++ b/test/CodeGen/avx512bw-builtins.c
@@ -1626,16 +1626,26 @@ __m512i test_mm512_maskz_set1_epi8(__mmask64 __M, char __A) {
   return _mm512_maskz_set1_epi8(__M, __A); 
 }
 
-__mmask64 test_mm512_kunpackd(__mmask64 __A, __mmask64 __B) {
+__mmask64 test_mm512_kunpackd(__m512i __A, __m512i __B, __m512i __C, __m512i __D, __m512i __E, __m512i __F) {
   // CHECK-LABEL: @test_mm512_kunpackd
-  // CHECK: @llvm.x86.avx512.kunpck.dq
-  return _mm512_kunpackd(__A, __B); 
+  // CHECK: bitcast <64 x i1> %{{.*}} to i64
+  // CHECK: bitcast <64 x i1> %{{.*}} to i64
+  // CHECK: and i64 %{{.*}}, 4294967295
+  // CHECK: shl i64 %{{.*}}, 32
+  // CHECK: or i64 %{{.*}}, %{{.*}}
+  // CHECK: bitcast i64 %{{.*}} to <64 x i1>
+  return _mm512_mask_cmpneq_epu8_mask(_mm512_kunpackd(_mm512_cmpneq_epu8_mask(__B, __A),_mm512_cmpneq_epu8_mask(__C, __D)), __E, __F); 
 }
 
-__mmask32 test_mm512_kunpackw(__mmask32 __A, __mmask32 __B) {
+__mmask32 test_mm512_kunpackw(__m512i __A, __m512i __B, __m512i __C, __m512i __D, __m512i __E, __m512i __F) {
   // CHECK-LABEL: @test_mm512_kunpackw
-  // CHECK: @llvm.x86.avx512.kunpck.wd
-  return _mm512_kunpackw(__A, __B); 
+  // CHECK: bitcast <32 x i1> %{{.*}} to i32
+  // CHECK: bitcast <32 x i1> %{{.*}} to i32
+  // CHECK: and i32 %{{.*}}, 65535
+  // CHECK: shl i32 %{{.*}}, 16
+  // CHECK: or i32 %{{.*}}, %{{.*}}
+  // CHECK: bitcast i32 %{{.*}} to <32 x i1>
+  return _mm512_mask_cmpneq_epu16_mask(_mm512_kunpackw(_mm512_cmpneq_epu16_mask(__B, __A),_mm512_cmpneq_epu16_mask(__C, __D)), __E, __F); 
 }
 
 __m512i test_mm512_mask_loadu_epi16(__m512i __W, __mmask32 __U, void const *__P) {

diff --git a/test/CodeGen/avx512f-builtins.c b/test/CodeGen/avx512f-builtins.c
index 4c97b2e3f904e9d3c45d296e33b75ec9f93c2dcc..a4e9e3bc9df00e23c796cd56a1aac0f115512083 100644 (file)
--- a/test/CodeGen/avx512f-builtins.c
+++ b/test/CodeGen/avx512f-builtins.c
@@ -6241,10 +6241,17 @@ int test_mm512_kortestz(__mmask16 __A, __mmask16 __B) {
   return _mm512_kortestz(__A, __B); 
 }
 
-__mmask16 test_mm512_kunpackb(__mmask16 __A, __mmask16 __B) {
+__mmask16 test_mm512_kunpackb(__m512i __A, __m512i __B, __m512i __C, __m512i __D, __m512i __E, __m512i __F) {
   // CHECK-LABEL: @test_mm512_kunpackb
-  // CHECK: @llvm.x86.avx512.kunpck.bw
-  return _mm512_kunpackb(__A, __B); 
+  // CHECK: bitcast <16 x i1> %{{.*}} to i16
+  // CHECK: bitcast <16 x i1> %{{.*}} to i16
+  // CHECK: and i32 %{{.*}}, 255
+  // CHECK: shl i32 %{{.*}}, 8
+  // CHECK: or i32 %{{.*}}, %{{.*}}
+  // CHECK: bitcast i16 %{{.*}} to <16 x i1>
+  return _mm512_mask_cmpneq_epu32_mask(_mm512_kunpackb(_mm512_cmpneq_epu32_mask(__A, __B),
+                                                       _mm512_cmpneq_epu32_mask(__C, __D)),
+                                                       __E, __F);
 }
 
 __mmask16 test_mm512_kxnor(__mmask16 __A, __mmask16 __B) {