[X86][AVX512] Converted the VPERMPD/VPERMQ intrinsics to generic IR

RKSimon · RKSimon · commit 4fa1f773ded5 · 2016-07-04T13:34:44.000Z
git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@274502 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Headers/avx512fintrin.h b/lib/Headers/avx512fintrin.h
@@ -8678,35 +8678,49 @@ _mm_mask3_fnmsub_sd (__m128d __W, __m128d __X, __m128d __Y, __mmask8 __U)
                                           -(__v2df)(__m128d)(Y), \
                                           (__mmask8)(U), (int)(R)); })
 
-#define _mm512_permutex_pd(X, M) __extension__ ({ \
-  (__m512d)__builtin_ia32_permdf512_mask((__v8df)(__m512d)(X), (int)(M), \
-                                         (__v8df)_mm512_undefined_pd(), \
-                                         (__mmask8)-1); })
-
-#define _mm512_mask_permutex_pd(W, U, X, M) __extension__ ({ \
-  (__m512d)__builtin_ia32_permdf512_mask((__v8df)(__m512d)(X), (int)(M), \
-                                         (__v8df)(__m512d)(W), \
-                                         (__mmask8)(U)); })
-
-#define _mm512_maskz_permutex_pd(U, X, M) __extension__ ({ \
-  (__m512d)__builtin_ia32_permdf512_mask((__v8df)(__m512d)(X), (int)(M), \
-                                         (__v8df)_mm512_setzero_pd(), \
-                                         (__mmask8)(U)); })
-
-#define _mm512_permutex_epi64(X, I) __extension__ ({ \
-  (__m512i)__builtin_ia32_permdi512_mask((__v8di)(__m512i)(X), (int)(I), \
-                                         (__v8di)_mm512_undefined_epi32(), \
-                                         (__mmask8)-1); })
+#define _mm512_permutex_pd(X, C) __extension__ ({ \
+  (__m512d)__builtin_shufflevector((__v8df)(__m512d)(X), \
+                                   (__v8df)_mm512_undefined_pd(), \
+                                   0 + (((C) & 0x03) >> 0), \
+                                   0 + (((C) & 0x0c) >> 2), \
+                                   0 + (((C) & 0x30) >> 4), \
+                                   0 + (((C) & 0xc0) >> 6), \
+                                   4 + (((C) & 0x03) >> 0), \
+                                   4 + (((C) & 0x0c) >> 2), \
+                                   4 + (((C) & 0x30) >> 4), \
+                                   4 + (((C) & 0xc0) >> 6)); })
+
+#define _mm512_mask_permutex_pd(W, U, X, C) __extension__ ({ \
+  (__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \
+                                       (__v8df)_mm512_permutex_pd((X), (C)), \
+                                       (__v8df)(__m512d)(W)); })
 
-#define _mm512_mask_permutex_epi64(W, M, X, I) __extension__ ({ \
-  (__m512i)__builtin_ia32_permdi512_mask((__v8di)(__m512i)(X), (int)(I), \
-                                         (__v8di)(__m512i)(W), \
-                                         (__mmask8)(M)); })
+#define _mm512_maskz_permutex_pd(U, X, C) __extension__ ({ \
+  (__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \
+                                       (__v8df)_mm512_permutex_pd((X), (C)), \
+                                       (__v8df)_mm512_setzero_pd()); })
 
-#define _mm512_maskz_permutex_epi64(M, X, I) __extension__ ({ \
-  (__m512i)__builtin_ia32_permdi512_mask((__v8di)(__m512i)(X), (int)(I), \
-                                         (__v8di)_mm512_setzero_si512(), \
-                                         (__mmask8)(M)); })
+#define _mm512_permutex_epi64(X, C) __extension__ ({ \
+  (__m512i)__builtin_shufflevector((__v8di)(__m512i)(X), \
+                                   (__v8di)_mm512_undefined_epi32(), \
+                                   0 + (((C) & 0x03) >> 0), \
+                                   0 + (((C) & 0x0c) >> 2), \
+                                   0 + (((C) & 0x30) >> 4), \
+                                   0 + (((C) & 0xc0) >> 6), \
+                                   4 + (((C) & 0x03) >> 0), \
+                                   4 + (((C) & 0x0c) >> 2), \
+                                   4 + (((C) & 0x30) >> 4), \
+                                   4 + (((C) & 0xc0) >> 6)); })
+
+#define _mm512_mask_permutex_epi64(W, U, X, C) __extension__ ({ \
+  (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \
+                                      (__v8di)_mm512_permutex_epi64((X), (C)), \
+                                      (__v8di)(__m512i)(W)); })
+
+#define _mm512_maskz_permutex_epi64(U, X, C) __extension__ ({ \
+  (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \
+                                      (__v8di)_mm512_permutex_epi64((X), (C)), \
+                                      (__v8di)_mm512_setzero_si512()); })
 
 static __inline__ __m512d __DEFAULT_FN_ATTRS
 _mm512_permutexvar_pd (__m512i __X, __m512d __Y)
diff --git a/lib/Headers/avx512vlintrin.h b/lib/Headers/avx512vlintrin.h
@@ -8806,35 +8806,37 @@ _mm256_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
                                         (__v8si)(__m256i)(index), \
                                         (__mmask8)(mask), (int)(scale)); })
 
-#define _mm256_mask_permutex_pd(W, U, X, imm) __extension__ ({ \
-  (__m256d)__builtin_ia32_permdf256_mask((__v4df)(__m256d)(X), (int)(imm), \
-                                         (__v4df)(__m256d)(W), \
-                                         (__mmask8)(U)); })
-
-#define _mm256_maskz_permutex_pd(U, X, imm) __extension__ ({ \
-  (__m256d)__builtin_ia32_permdf256_mask((__v4df)(__m256d)(X), (int)(imm), \
-                                         (__v4df)_mm256_setzero_pd(), \
-                                         (__mmask8)(U)); })
-
-#define _mm256_permutex_pd(X, M) __extension__ ({ \
-  (__m256d)__builtin_ia32_permdf256_mask((__v4df)(__m256d)(X), (int)(M), \
-                                         (__v4df)_mm256_undefined_pd(), \
-                                         (__mmask8)-1); })
+#define _mm256_permutex_pd(X, C) __extension__ ({ \
+  (__m256d)__builtin_shufflevector((__v4df)(__m256d)(X), \
+                                   (__v4df)_mm256_undefined_pd(), \
+                                   (C) & 0x3, ((C) & 0xc) >> 2, \
+                                   ((C) & 0x30) >> 4, ((C) & 0xc0) >> 6); })
 
-#define _mm256_mask_permutex_epi64(W, M, X, I) __extension__ ({ \
-  (__m256i)__builtin_ia32_permdi256_mask((__v4di)(__m256i)(X), (int)(I), \
-                                         (__v4di)(__m256i)(W), \
-                                         (__mmask8)(M)); })
+#define _mm256_mask_permutex_pd(W, U, X, C) __extension__ ({ \
+  (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \
+                                       (__v4df)_mm256_permutex_pd((X), (C)), \
+                                       (__v4df)(__m256d)(W)); })
 
-#define _mm256_maskz_permutex_epi64(M, X, I) __extension__ ({ \
-  (__m256i)__builtin_ia32_permdi256_mask((__v4di)(__m256i)(X), (int)(I), \
-                                         (__v4di)_mm256_setzero_si256(), \
-                                         (__mmask8)(M)); })
+#define _mm256_maskz_permutex_pd(U, X, C) __extension__ ({ \
+  (__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \
+                                       (__v4df)_mm256_permutex_pd((X), (C)), \
+                                       (__v4df)_mm256_setzero_pd()); })
 
-#define _mm256_permutex_epi64(X, I) __extension__ ({ \
-  (__m256i)__builtin_ia32_permdi256_mask((__v4di)(__m256i)(X), (int)(I), \
-                                         (__v4di)_mm256_undefined_si256(), \
-                                         (__mmask8)-1); })
+#define _mm256_permutex_epi64(X, C) __extension__ ({ \
+  (__m256i)__builtin_shufflevector((__v4di)(__m256i)(X), \
+                                   (__v4di)_mm256_undefined_si256(), \
+                                   (C) & 0x3, ((C) & 0xc) >> 2, \
+                                   ((C) & 0x30) >> 4, ((C) & 0xc0) >> 6); })
+
+#define _mm256_mask_permutex_epi64(W, U, X, C) __extension__ ({ \
+  (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \
+                                      (__v4di)_mm256_permutex_epi64((X), (C)), \
+                                      (__v4di)(__m256i)(W)); })
+
+#define _mm256_maskz_permutex_epi64(U, X, C) __extension__ ({ \
+  (__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \
+                                      (__v4di)_mm256_permutex_epi64((X), (C)), \
+                                      (__v4di)_mm256_setzero_si256()); })
 
 static __inline__ __m256d __DEFAULT_FN_ATTRS
 _mm256_permutexvar_pd (__m256i __X, __m256d __Y)
diff --git a/test/CodeGen/avx512f-builtins.c b/test/CodeGen/avx512f-builtins.c
@@ -5704,38 +5704,42 @@ __m128d test_mm_mask3_fnmsub_round_sd(__m128d __W, __m128d __X, __m128d __Y, __m
 
 __m512d test_mm512_permutex_pd(__m512d __X) {
   // CHECK-LABEL: @test_mm512_permutex_pd
-  // CHECK: @llvm.x86.avx512.mask.perm.df.512
-  return _mm512_permutex_pd(__X, 0); 
+  // 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);
 }
 
 __m512d test_mm512_mask_permutex_pd(__m512d __W, __mmask8 __U, __m512d __X) {
   // CHECK-LABEL: @test_mm512_mask_permutex_pd
-  // CHECK: @llvm.x86.avx512.mask.perm.df.512
-  return _mm512_mask_permutex_pd(__W, __U, __X, 0); 
+  // 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>
+  // CHECK: select <8 x i1> %{{.*}}, <8 x double> %{{.*}}, <8 x double> %{{.*}}
+  return _mm512_mask_permutex_pd(__W, __U, __X, 0);
 }
 
 __m512d test_mm512_maskz_permutex_pd(__mmask8 __U, __m512d __X) {
   // CHECK-LABEL: @test_mm512_maskz_permutex_pd
-  // CHECK: @llvm.x86.avx512.mask.perm.df.512
-  return _mm512_maskz_permutex_pd(__U, __X, 0); 
+  // 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>
+  // CHECK: select <8 x i1> %{{.*}}, <8 x double> %{{.*}}, <8 x double> %{{.*}}
+  return _mm512_maskz_permutex_pd(__U, __X, 0);
 }
 
 __m512i test_mm512_permutex_epi64(__m512i __X) {
   // CHECK-LABEL: @test_mm512_permutex_epi64
-  // CHECK: @llvm.x86.avx512.mask.perm.di.512
-  return _mm512_permutex_epi64(__X, 0); 
+  // CHECK: shufflevector <8 x i64> %{{.*}}, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 4, i32 4, i32 4, i32 4>
+  return _mm512_permutex_epi64(__X, 0);
 }
 
 __m512i test_mm512_mask_permutex_epi64(__m512i __W, __mmask8 __M, __m512i __X) {
   // CHECK-LABEL: @test_mm512_mask_permutex_epi64
-  // CHECK: @llvm.x86.avx512.mask.perm.di.512
-  return _mm512_mask_permutex_epi64(__W, __M, __X, 0); 
+  // CHECK: shufflevector <8 x i64> %{{.*}}, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 4, i32 4, i32 4, i32 4>
+  // CHECK: select <8 x i1> %{{.*}}, <8 x i64> %{{.*}}, <8 x i64> %{{.*}}
+  return _mm512_mask_permutex_epi64(__W, __M, __X, 0);
 }
 
 __m512i test_mm512_maskz_permutex_epi64(__mmask8 __M, __m512i __X) {
   // CHECK-LABEL: @test_mm512_maskz_permutex_epi64
-  // CHECK: @llvm.x86.avx512.mask.perm.di.512
-  return _mm512_maskz_permutex_epi64(__M, __X, 0); 
+  // CHECK: shufflevector <8 x i64> %{{.*}}, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 4, i32 4, i32 4, i32 4>
+  // CHECK: select <8 x i1> %{{.*}}, <8 x i64> %{{.*}}, <8 x i64> %{{.*}}
+  return _mm512_maskz_permutex_epi64(__M, __X, 0);
 }
 
 __m512d test_mm512_permutexvar_pd(__m512i __X, __m512d __Y) {
diff --git a/test/CodeGen/avx512vl-builtins.c b/test/CodeGen/avx512vl-builtins.c
