[Headers][X86] Ensure all AVX broadcast scalar load intrinsics are un…

…aligned Similar to the existing _mm_load1_pd/_mm_loaddup_pd and broadcast vector loads, these intrinsic should ensure the loads are unaligned and not assume type alignment Fixes #62325
llvm · Jul 3, 2023 · d963420 · d963420
1 parent f292ca1
commit d963420
Show file tree

Hide file tree

Showing 2 changed files with 16 additions and 7 deletions.
diff --git a/clang/lib/Headers/avxintrin.h b/clang/lib/Headers/avxintrin.h
@@ -3017,8 +3017,11 @@ _mm256_zeroupper(void)
 static __inline __m128 __DEFAULT_FN_ATTRS128
 _mm_broadcast_ss(float const *__a)
 {
-  float __f = *__a;
-  return __extension__ (__m128)(__v4sf){ __f, __f, __f, __f };
+  struct __mm_broadcast_ss_struct {
+    float __f;
+  } __attribute__((__packed__, __may_alias__));
+  float __f = ((const struct __mm_broadcast_ss_struct*)__a)->__f;
+  return __extension__ (__m128){ __f, __f, __f, __f };
 }
 
 /// Loads a scalar double-precision floating point value from the
@@ -3036,7 +3039,10 @@ _mm_broadcast_ss(float const *__a)
 static __inline __m256d __DEFAULT_FN_ATTRS
 _mm256_broadcast_sd(double const *__a)
 {
-  double __d = *__a;
+  struct __mm256_broadcast_sd_struct {
+    double __d;
+  } __attribute__((__packed__, __may_alias__));
+  double __d = ((const struct __mm256_broadcast_sd_struct*)__a)->__d;
   return __extension__ (__m256d)(__v4df){ __d, __d, __d, __d };
 }
 
@@ -3055,7 +3061,10 @@ _mm256_broadcast_sd(double const *__a)
 static __inline __m256 __DEFAULT_FN_ATTRS
 _mm256_broadcast_ss(float const *__a)
 {
-  float __f = *__a;
+  struct __mm256_broadcast_ss_struct {
+    float __f;
+  } __attribute__((__packed__, __may_alias__));
+  float __f = ((const struct __mm256_broadcast_ss_struct*)__a)->__f;
   return __extension__ (__m256)(__v8sf){ __f, __f, __f, __f, __f, __f, __f, __f };
 }
 

diff --git a/clang/test/CodeGen/X86/avx-builtins.c b/clang/test/CodeGen/X86/avx-builtins.c
@@ -99,7 +99,7 @@ __m256 test_mm256_broadcast_ps(__m128* A) {
 
 __m256d test_mm256_broadcast_sd(double* A) {
   // CHECK-LABEL: test_mm256_broadcast_sd
-  // CHECK: load double, ptr %{{.*}}
+  // CHECK: load double, ptr %{{.*}}, align 1{{$}}
   // CHECK: insertelement <4 x double> undef, double %{{.*}}, i32 0
   // CHECK: insertelement <4 x double> %{{.*}}, double %{{.*}}, i32 1
   // CHECK: insertelement <4 x double> %{{.*}}, double %{{.*}}, i32 2
@@ -109,7 +109,7 @@ __m256d test_mm256_broadcast_sd(double* A) {
 
 __m128 test_mm_broadcast_ss(float* A) {
   // CHECK-LABEL: test_mm_broadcast_ss
-  // CHECK: load float, ptr %{{.*}}
+  // CHECK: load float, ptr %{{.*}}, align 1{{$}}
   // CHECK: insertelement <4 x float> undef, float %{{.*}}, i32 0
   // CHECK: insertelement <4 x float> %{{.*}}, float %{{.*}}, i32 1
   // CHECK: insertelement <4 x float> %{{.*}}, float %{{.*}}, i32 2
@@ -119,7 +119,7 @@ __m128 test_mm_broadcast_ss(float* A) {
 
 __m256 test_mm256_broadcast_ss(float* A) {
   // CHECK-LABEL: test_mm256_broadcast_ss
-  // CHECK: load float, ptr %{{.*}}
+  // CHECK: load float, ptr %{{.*}}, align 1{{$}}
   // CHECK: insertelement <8 x float> undef, float %{{.*}}, i32 0
   // CHECK: insertelement <8 x float> %{{.*}}, float %{{.*}}, i32 1
   // CHECK: insertelement <8 x float> %{{.*}}, float %{{.*}}, i32 2