RKSimon created this revision.
RKSimon added reviewers: craig.topper, spatel, andreadb.
RKSimon added a subscriber: cfe-commits.
RKSimon set the repository for this revision to rL LLVM.
According to the gcc headers, intel intrinsics docs and msdn codegen the
_mm_store1_ps/_mm_store1_pd (and their _mm_store_ps1/_mm_store_pd1 analogues)
should require an aligned pointer - the clang headers are the only
implementation I can find that assume non-aligned stores (by storing with
_mm_storeu_ps/_mm_storeu_pd).
This patch raises the alignment requirements to match the other implementations
by calling _mm_store_ps/_mm_store_pd instead.
I've also added the missing _mm_store_pd1 intrinsic (which maps to
_mm_store1_pd like _mm_store_ps1 does to _mm_store1_ps).
As a followup I'll update the llvm fast-isel tests to match this codegen.
Repository:
rL LLVM
http://reviews.llvm.org/D20617
Files:
lib/Headers/emmintrin.h
lib/Headers/xmmintrin.h
test/CodeGen/sse-builtins.c
test/CodeGen/sse2-builtins.c
Index: test/CodeGen/sse2-builtins.c
===================================================================
--- test/CodeGen/sse2-builtins.c
+++ test/CodeGen/sse2-builtins.c
@@ -1205,6 +1205,13 @@
_mm_store_pd(A, B);
}
+void test_mm_store_pd1(double* x, __m128d y) {
+ // CHECK-LABEL: test_mm_store_pd1
+ // CHECK: shufflevector <2 x double> %{{.*}}, <2 x double> %{{.*}}, <2 x i32> zeroinitializer
+ // CHECK: store <2 x double> %{{.*}}, <2 x double>* {{.*}}, align 16
+ _mm_store_pd1(x, y);
+}
+
void test_mm_store_sd(double* A, __m128d B) {
// CHECK-LABEL: test_mm_store_sd
// CHECK: extractelement <2 x double> %{{.*}}, i32 0
@@ -1220,9 +1227,8 @@
void test_mm_store1_pd(double* x, __m128d y) {
// CHECK-LABEL: test_mm_store1_pd
- // CHECK: extractelement <2 x double> %{{.*}}, i32 0
- // CHECK: store {{.*}} double* {{.*}}, align 1{{$}}
- // CHECK: store {{.*}} double* {{.*}}, align 1{{$}}
+ // CHECK: shufflevector <2 x double> %{{.*}}, <2 x double> %{{.*}}, <2 x i32> zeroinitializer
+ // CHECK: store <2 x double> %{{.*}}, <2 x double>* {{.*}}, align 16
_mm_store1_pd(x, y);
}
Index: test/CodeGen/sse-builtins.c
===================================================================
--- test/CodeGen/sse-builtins.c
+++ test/CodeGen/sse-builtins.c
@@ -651,7 +651,7 @@
void test_mm_store_ps1(float* x, __m128 y) {
// CHECK-LABEL: test_mm_store_ps1
// CHECK: shufflevector <4 x float> %{{.*}}, <4 x float> %{{.*}}, <4 x i32> zeroinitializer
- // CHECK: call void @llvm.x86.sse.storeu.ps(i8* %{{.*}}, <4 x float> %{{.*}})
+ // CHECK: store <4 x float> %{{.*}}, <4 x float>* {{.*}}, align 16
_mm_store_ps1(x, y);
}
@@ -665,7 +665,7 @@
void test_mm_store1_ps(float* x, __m128 y) {
// CHECK-LABEL: test_mm_store1_ps
// CHECK: shufflevector <4 x float> %{{.*}}, <4 x float> %{{.*}}, <4 x i32> zeroinitializer
- // CHECK: call void @llvm.x86.sse.storeu.ps(i8* %{{.*}}, <4 x float> %{{.*}})
+ // CHECK: store <4 x float> %{{.*}}, <4 x float>* {{.*}}, align 16
_mm_store1_ps(x, y);
}
Index: lib/Headers/xmmintrin.h
===================================================================
--- lib/Headers/xmmintrin.h
+++ lib/Headers/xmmintrin.h
@@ -1590,22 +1590,22 @@
}
static __inline__ void __DEFAULT_FN_ATTRS
-_mm_store1_ps(float *__p, __m128 __a)
+_mm_store_ps(float *__p, __m128 __a)
{
- __a = __builtin_shufflevector((__v4sf)__a, (__v4sf)__a, 0, 0, 0, 0);
- _mm_storeu_ps(__p, __a);
+ *(__m128 *)__p = __a;
}
static __inline__ void __DEFAULT_FN_ATTRS
-_mm_store_ps1(float *__p, __m128 __a)
+_mm_store1_ps(float *__p, __m128 __a)
{
- return _mm_store1_ps(__p, __a);
+ __a = __builtin_shufflevector((__v4sf)__a, (__v4sf)__a, 0, 0, 0, 0);
+ _mm_store_ps(__p, __a);
}
static __inline__ void __DEFAULT_FN_ATTRS
-_mm_store_ps(float *__p, __m128 __a)
+_mm_store_ps1(float *__p, __m128 __a)
{
- *(__m128 *)__p = __a;
+ return _mm_store1_ps(__p, __a);
}
static __inline__ void __DEFAULT_FN_ATTRS
Index: lib/Headers/emmintrin.h
===================================================================
--- lib/Headers/emmintrin.h
+++ lib/Headers/emmintrin.h
@@ -582,19 +582,22 @@
}
static __inline__ void __DEFAULT_FN_ATTRS
+_mm_store_pd(double *__dp, __m128d __a)
+{
+ *(__m128d *)__dp = __a;
+}
+
+static __inline__ void __DEFAULT_FN_ATTRS
_mm_store1_pd(double *__dp, __m128d __a)
{
- struct __mm_store1_pd_struct {
- double __u[2];
- } __attribute__((__packed__, __may_alias__));
- ((struct __mm_store1_pd_struct*)__dp)->__u[0] = __a[0];
- ((struct __mm_store1_pd_struct*)__dp)->__u[1] = __a[0];
+ __a = __builtin_shufflevector((__v2df)__a, (__v2df)__a, 0, 0);
+ _mm_store_pd(__dp, __a);
}
static __inline__ void __DEFAULT_FN_ATTRS
-_mm_store_pd(double *__dp, __m128d __a)
+_mm_store_pd1(double *__dp, __m128d __a)
{
- *(__m128d *)__dp = __a;
+ return _mm_store1_pd(__dp, __a);
}
static __inline__ void __DEFAULT_FN_ATTRS
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
