Author: rksimon Date: Sat Jul 29 08:33:34 2017 New Revision: 309488 URL: http://llvm.org/viewvc/llvm-project?rev=309488&view=rev Log: [X86][AVX] Ensure vector non-temporal load/store intrinsics force pointer alignment (PR33830)
Clang specifies a max type alignment of 16 bytes on darwin targets (annoyingly in the driver not via cc1), meaning that the builtin nontemporal stores don't correctly align the loads/stores to 32 or 64 bytes when required, resulting in lowering to temporal unaligned loads/stores. This patch casts the vectors to explicitly aligned types prior to the load/store to ensure that the require alignment is respected. Differential Revision: https://reviews.llvm.org/D35996 Modified: cfe/trunk/lib/Headers/avx2intrin.h cfe/trunk/lib/Headers/avx512fintrin.h cfe/trunk/lib/Headers/avxintrin.h cfe/trunk/test/CodeGen/x86-nontemporal.c Modified: cfe/trunk/lib/Headers/avx2intrin.h URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/Headers/avx2intrin.h?rev=309488&r1=309487&r2=309488&view=diff ============================================================================== --- cfe/trunk/lib/Headers/avx2intrin.h (original) +++ cfe/trunk/lib/Headers/avx2intrin.h Sat Jul 29 08:33:34 2017 @@ -832,7 +832,8 @@ _mm256_xor_si256(__m256i __a, __m256i __ static __inline__ __m256i __DEFAULT_FN_ATTRS _mm256_stream_load_si256(__m256i const *__V) { - return (__m256i)__builtin_nontemporal_load((const __v4di *)__V); + typedef __v4di __v4di_aligned __attribute__((aligned(32))); + return (__m256i)__builtin_nontemporal_load((const __v4di_aligned *)__V); } static __inline__ __m128 __DEFAULT_FN_ATTRS Modified: cfe/trunk/lib/Headers/avx512fintrin.h URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/Headers/avx512fintrin.h?rev=309488&r1=309487&r2=309488&view=diff ============================================================================== --- cfe/trunk/lib/Headers/avx512fintrin.h (original) +++ cfe/trunk/lib/Headers/avx512fintrin.h Sat Jul 29 08:33:34 2017 @@ -9035,25 +9035,29 @@ _mm512_kxor (__mmask16 __A, __mmask16 __ static __inline__ void __DEFAULT_FN_ATTRS _mm512_stream_si512 (__m512i * __P, __m512i __A) { - __builtin_nontemporal_store((__v8di)__A, (__v8di*)__P); + typedef __v8di __v8di_aligned __attribute__((aligned(64))); + __builtin_nontemporal_store((__v8di_aligned)__A, (__v8di_aligned*)__P); } static __inline__ __m512i __DEFAULT_FN_ATTRS _mm512_stream_load_si512 (void *__P) { - return (__m512i) __builtin_nontemporal_load((const __v8di *)__P); + typedef __v8di __v8di_aligned __attribute__((aligned(64))); + return (__m512i) __builtin_nontemporal_load((const __v8di_aligned *)__P); } static __inline__ void __DEFAULT_FN_ATTRS _mm512_stream_pd (double *__P, __m512d __A) { - __builtin_nontemporal_store((__v8df)__A, (__v8df*)__P); + typedef __v8df __v8df_aligned __attribute__((aligned(64))); + __builtin_nontemporal_store((__v8df_aligned)__A, (__v8df_aligned*)__P); } static __inline__ void __DEFAULT_FN_ATTRS _mm512_stream_ps (float *__P, __m512 __A) { - __builtin_nontemporal_store((__v16sf)__A, (__v16sf*)__P); + typedef __v16sf __v16sf_aligned __attribute__((aligned(64))); + __builtin_nontemporal_store((__v16sf_aligned)__A, (__v16sf_aligned*)__P); } static __inline__ __m512d __DEFAULT_FN_ATTRS Modified: cfe/trunk/lib/Headers/avxintrin.h URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/Headers/avxintrin.h?rev=309488&r1=309487&r2=309488&view=diff ============================================================================== --- cfe/trunk/lib/Headers/avxintrin.h (original) +++ cfe/trunk/lib/Headers/avxintrin.h Sat Jul 29 08:33:34 2017 @@ -3590,7 +3590,8 @@ _mm_maskstore_ps(float *__p, __m128i __m static __inline void __DEFAULT_FN_ATTRS _mm256_stream_si256(__m256i *__a, __m256i __b) { - __builtin_nontemporal_store((__v4di)__b, (__v4di*)__a); + typedef __v4di __v4di_aligned __attribute__((aligned(32))); + __builtin_nontemporal_store((__v4di_aligned)__b, (__v4di_aligned*)__a); } /// \brief Moves double-precision values from a 256-bit vector of [4 x double] @@ -3609,7 +3610,8 @@ _mm256_stream_si256(__m256i *__a, __m256 static __inline void __DEFAULT_FN_ATTRS _mm256_stream_pd(double *__a, __m256d __b) { - __builtin_nontemporal_store((__v4df)__b, (__v4df*)__a); + typedef __v4df __v4df_aligned __attribute__((aligned(32))); + __builtin_nontemporal_store((__v4df_aligned)__b, (__v4df_aligned*)__a); } /// \brief Moves single-precision floating point values from a 256-bit vector @@ -3629,7 +3631,8 @@ _mm256_stream_pd(double *__a, __m256d __ static __inline void __DEFAULT_FN_ATTRS _mm256_stream_ps(float *__p, __m256 __a) { - __builtin_nontemporal_store((__v8sf)__a, (__v8sf*)__p); + typedef __v8sf __v8sf_aligned __attribute__((aligned(32))); + __builtin_nontemporal_store((__v8sf_aligned)__a, (__v8sf_aligned*)__p); } /* Create vectors */ Modified: cfe/trunk/test/CodeGen/x86-nontemporal.c URL: http://llvm.org/viewvc/llvm-project/cfe/trunk/test/CodeGen/x86-nontemporal.c?rev=309488&r1=309487&r2=309488&view=diff ============================================================================== --- cfe/trunk/test/CodeGen/x86-nontemporal.c (original) +++ cfe/trunk/test/CodeGen/x86-nontemporal.c Sat Jul 29 08:33:34 2017 @@ -1,12 +1,12 @@ // RUN: %clang_cc1 -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +sse4.1 -target-feature +avx -target-feature +avx2 -target-feature +avx512f -emit-llvm -o - -Wall -Werror | FileCheck %s --check-prefix=CHECK // RUN: %clang_cc1 -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +sse4.1 -target-feature +avx -target-feature +avx2 -target-feature +avx512f -fno-signed-char -emit-llvm -o - -Wall -Werror | FileCheck %s --check-prefix=CHECK -// RUN: %clang_cc1 -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +sse4.1 -target-feature +avx -target-feature +avx2 -target-feature +avx512f -emit-llvm -o - -Wall -Werror -fmax-type-align=16 | FileCheck %s --check-prefix=CHECK16 -// RUN: %clang_cc1 -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +sse4.1 -target-feature +avx -target-feature +avx2 -target-feature +avx512f -fno-signed-char -emit-llvm -o - -Wall -Werror -fmax-type-align=16 | FileCheck %s --check-prefix=CHECK16 +// RUN: %clang_cc1 -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +sse4.1 -target-feature +avx -target-feature +avx2 -target-feature +avx512f -emit-llvm -o - -Wall -Werror -fmax-type-align=16 | FileCheck %s --check-prefix=CHECK +// RUN: %clang_cc1 -ffreestanding %s -triple=x86_64-apple-darwin -target-feature +sse4.1 -target-feature +avx -target-feature +avx2 -target-feature +avx512f -fno-signed-char -emit-llvm -o - -Wall -Werror -fmax-type-align=16 | FileCheck %s --check-prefix=CHECK #include <x86intrin.h> -// FIXME: (PR33830) Tests showing failure to correctly align non-temporal load/stores on darwin targets where fmax-type-align is set to 16. +// (PR33830) Tests ensure the correct alignment of non-temporal load/stores on darwin targets where fmax-type-align is set to 16. // // 128-bit vectors @@ -43,36 +43,24 @@ __m128i test_mm_stream_load_si128(__m128 void test_mm256_stream_pd(double* A, __m256d B) { // CHECK-LABEL: test_mm256_stream_pd // CHECK: store <4 x double> %{{.*}}, <4 x double>* %{{.*}}, align 32, !nontemporal - - // CHECK16-LABEL: test_mm256_stream_pd - // CHECK16: store <4 x double> %{{.*}}, <4 x double>* %{{.*}}, align 16, !nontemporal _mm256_stream_pd(A, B); } void test_mm256_stream_ps(float* A, __m256 B) { // CHECK-LABEL: test_mm256_stream_ps // CHECK: store <8 x float> %{{.*}}, <8 x float>* %{{.*}}, align 32, !nontemporal - - // CHECK16-LABEL: test_mm256_stream_ps - // CHECK16: store <8 x float> %{{.*}}, <8 x float>* %{{.*}}, align 16, !nontemporal _mm256_stream_ps(A, B); } void test_mm256_stream_si256(__m256i* A, __m256i B) { // CHECK-LABEL: test_mm256_stream_si256 // CHECK: store <4 x i64> %{{.*}}, <4 x i64>* %{{.*}}, align 32, !nontemporal - - // CHECK16-LABEL: test_mm256_stream_si256 - // CHECK16: store <4 x i64> %{{.*}}, <4 x i64>* %{{.*}}, align 16, !nontemporal _mm256_stream_si256(A, B); } __m256i test_mm256_stream_load_si256(__m256i const *A) { // CHECK-LABEL: test_mm256_stream_load_si256 // CHECK: load <4 x i64>, <4 x i64>* %{{.*}}, align 32, !nontemporal - - // CHECK16-LABEL: test_mm256_stream_load_si256 - // CHECK16: load <4 x i64>, <4 x i64>* %{{.*}}, align 16, !nontemporal return _mm256_stream_load_si256(A); } @@ -83,35 +71,23 @@ __m256i test_mm256_stream_load_si256(__m void test_mm512_stream_pd(double* A, __m512d B) { // CHECK-LABEL: test_mm512_stream_pd // CHECK: store <8 x double> %{{.*}}, <8 x double>* %{{.*}}, align 64, !nontemporal - - // CHECK16-LABEL: test_mm512_stream_pd - // CHECK16: store <8 x double> %{{.*}}, <8 x double>* %{{.*}}, align 16, !nontemporal _mm512_stream_pd(A, B); } void test_mm512_stream_ps(float* A, __m512 B) { // CHECK-LABEL: test_mm512_stream_ps // CHECK: store <16 x float> %{{.*}}, <16 x float>* %{{.*}}, align 64, !nontemporal - - // CHECK16-LABEL: test_mm512_stream_ps - // CHECK16: store <16 x float> %{{.*}}, <16 x float>* %{{.*}}, align 16, !nontemporal _mm512_stream_ps(A, B); } void test_mm512_stream_si512(__m512i* A, __m512i B) { // CHECK-LABEL: test_mm512_stream_si512 // CHECK: store <8 x i64> %{{.*}}, <8 x i64>* %{{.*}}, align 64, !nontemporal - - // CHECK16-LABEL: test_mm512_stream_si512 - // CHECK16: store <8 x i64> %{{.*}}, <8 x i64>* %{{.*}}, align 16, !nontemporal _mm512_stream_si512(A, B); } __m512i test_mm512_stream_load_si512(void *A) { // CHECK-LABEL: test_mm512_stream_load_si512 // CHECK: load <8 x i64>, <8 x i64>* %{{.*}}, align 64, !nontemporal - - // CHECK16-LABEL: test_mm512_stream_load_si512 - // CHECK16: load <8 x i64>, <8 x i64>* %{{.*}}, align 16, !nontemporal return _mm512_stream_load_si512(A); } _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
