RKSimon created this revision.
RKSimon added reviewers: qcolombet, craig.topper, ab, spatel, silvas.
RKSimon added a subscriber: cfe-commits.
RKSimon set the repository for this revision to rL LLVM.
Sister patch to D10555
As discussed on D10555, this patch replaces the use of the avx2.vbroadcast
(float) and avx2.pbroadcast (integer) broadcast intrinsics in avx2intrin.h with
generic __builtin_shufflevector calls.
At present all these changes still result in the expected
vbroadcast/vpbroadcast instructions in debug code. I can add a test that
ensures this if people wish but the conclusion on the discussion thread was
that this shouldn't be considered vital.
Repository:
rL LLVM
http://reviews.llvm.org/D12143
Files:
lib/Headers/avx2intrin.h
test/CodeGen/avx2-builtins.c
Index: test/CodeGen/avx2-builtins.c
===================================================================
--- test/CodeGen/avx2-builtins.c
+++ test/CodeGen/avx2-builtins.c
@@ -607,7 +607,9 @@
}
__m128 test_mm_broadcastss_ps(__m128 a) {
- // CHECK: @llvm.x86.avx2.vbroadcast.ss.ps
+ // CHECK-LABEL: test_mm_broadcastss_ps
+ // CHECK-NOT: @llvm.x86.avx2.vbroadcast.ss.ps
+ // CHECK: shufflevector <4 x float> %{{.*}}, <4 x float> %{{.*}}, <4 x i32> zeroinitializer
return _mm_broadcastss_ps(a);
}
@@ -617,12 +619,16 @@
}
__m256 test_mm256_broadcastss_ps(__m128 a) {
- // CHECK: @llvm.x86.avx2.vbroadcast.ss.ps.256
+ // CHECK-LABEL: test_mm256_broadcastss_ps
+ // CHECK-NOT: @llvm.x86.avx2.vbroadcast.ss.ps.256
+ // CHECK: shufflevector <4 x float> %{{.*}}, <4 x float> %{{.*}}, <8 x i32> zeroinitializer
return _mm256_broadcastss_ps(a);
}
__m256d test_mm256_broadcastsd_pd(__m128d a) {
- // check: @llvm.x86.avx2.vbroadcast.sd.pd.256
+ // CHECK-LABEL: test_mm256_broadcastsd_pd
+ // CHECK-NOT: @llvm.x86.avx2.vbroadcast.sd.pd.256
+ // CHECK: shufflevector <2 x double> %{{.*}}, <2 x double> %{{.*}}, <4 x i32> zeroinitializer
return _mm256_broadcastsd_pd(a);
}
@@ -646,42 +652,58 @@
}
__m256i test_mm256_broadcastb_epi8(__m128i a) {
- // CHECK: @llvm.x86.avx2.pbroadcastb.256
+ // CHECK-LABEL: test_mm256_broadcastb_epi8
+ // CHECK-NOT: @llvm.x86.avx2.pbroadcastb.256
+ // CHECK: shufflevector <16 x i8> %{{.*}}, <16 x i8> %{{.*}}, <32 x i32> zeroinitializer
return _mm256_broadcastb_epi8(a);
}
__m256i test_mm256_broadcastw_epi16(__m128i a) {
- // CHECK: @llvm.x86.avx2.pbroadcastw.256
+ // CHECK-LABEL: test_mm256_broadcastw_epi16
+ // CHECK-NOT: @llvm.x86.avx2.pbroadcastw.256
+ // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> %{{.*}}, <16 x i32> zeroinitializer
return _mm256_broadcastw_epi16(a);
}
__m256i test_mm256_broadcastd_epi32(__m128i a) {
- // CHECK: @llvm.x86.avx2.pbroadcastd.256
+ // CHECK-LABEL: test_mm256_broadcastd_epi32
+ // CHECK-NOT: @llvm.x86.avx2.pbroadcastd.256
+ // CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> %{{.*}}, <8 x i32> zeroinitializer
return _mm256_broadcastd_epi32(a);
}
__m256i test_mm256_broadcastq_epi64(__m128i a) {
- // CHECK: @llvm.x86.avx2.pbroadcastq.256
+ // CHECK-LABEL: test_mm256_broadcastq_epi64
+ // CHECK-NOT: @llvm.x86.avx2.pbroadcastq.256
+ // CHECK: shufflevector <2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <4 x i32> zeroinitializer
return _mm256_broadcastq_epi64(a);
}
__m128i test_mm_broadcastb_epi8(__m128i a) {
- // CHECK: @llvm.x86.avx2.pbroadcastb.128
+ // CHECK-LABEL: test_mm_broadcastb_epi8
+ // CHECK-NOT: @llvm.x86.avx2.pbroadcastb.128
+ // CHECK: shufflevector <16 x i8> %{{.*}}, <16 x i8> %{{.*}}, <16 x i32> zeroinitializer
return _mm_broadcastb_epi8(a);
}
__m128i test_mm_broadcastw_epi16(__m128i a) {
- // CHECK: @llvm.x86.avx2.pbroadcastw.128
+ // CHECK-LABEL: test_mm_broadcastw_epi16
+ // CHECK-NOT: @llvm.x86.avx2.pbroadcastw.128
+ // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> %{{.*}}, <8 x i32> zeroinitializer
return _mm_broadcastw_epi16(a);
}
__m128i test_mm_broadcastd_epi32(__m128i a) {
- // CHECK: @llvm.x86.avx2.pbroadcastd.128
+ // CHECK-LABEL: test_mm_broadcastd_epi32
+ // CHECK-NOT: @llvm.x86.avx2.pbroadcastd.128
+ // CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> %{{.*}}, <4 x i32> zeroinitializer
return _mm_broadcastd_epi32(a);
}
__m128i test_mm_broadcastq_epi64(__m128i a) {
- // CHECK: @llvm.x86.avx2.pbroadcastq.128
+ // CHECK-LABEL: test_mm_broadcastq_epi64
+ // CHECK-NOT: @llvm.x86.avx2.pbroadcastq.128
+ // CHECK: shufflevector <2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <2 x i32> zeroinitializer
return _mm_broadcastq_epi64(a);
}
Index: lib/Headers/avx2intrin.h
===================================================================
--- lib/Headers/avx2intrin.h
+++ lib/Headers/avx2intrin.h
@@ -760,7 +760,7 @@
static __inline__ __m128 __DEFAULT_FN_ATTRS
_mm_broadcastss_ps(__m128 __X)
{
- return (__m128)__builtin_ia32_vbroadcastss_ps((__v4sf)__X);
+ return (__m128)__builtin_shufflevector((__v4sf)__X, (__v4sf)__X, 0, 0, 0, 0);
}
static __inline__ __m128d __DEFAULT_FN_ATTRS
@@ -772,13 +772,13 @@
static __inline__ __m256 __DEFAULT_FN_ATTRS
_mm256_broadcastss_ps(__m128 __X)
{
- return (__m256)__builtin_ia32_vbroadcastss_ps256((__v4sf)__X);
+ return (__m256)__builtin_shufflevector((__v4sf)__X, (__v4sf)__X, 0, 0, 0, 0, 0, 0, 0, 0);
}
static __inline__ __m256d __DEFAULT_FN_ATTRS
_mm256_broadcastsd_pd(__m128d __X)
{
- return (__m256d)__builtin_ia32_vbroadcastsd_pd256((__v2df)__X);
+ return (__m256d)__builtin_shufflevector((__v2df)__X, (__v2df)__X, 0, 0, 0, 0);
}
static __inline__ __m256i __DEFAULT_FN_ATTRS
@@ -812,50 +812,50 @@
static __inline__ __m256i __DEFAULT_FN_ATTRS
_mm256_broadcastb_epi8(__m128i __X)
{
- return (__m256i)__builtin_ia32_pbroadcastb256((__v16qi)__X);
+ return (__m256i)__builtin_shufflevector((__v16qi)__X, (__v16qi)__X, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
}
static __inline__ __m256i __DEFAULT_FN_ATTRS
_mm256_broadcastw_epi16(__m128i __X)
{
- return (__m256i)__builtin_ia32_pbroadcastw256((__v8hi)__X);
+ return (__m256i)__builtin_shufflevector((__v8hi)__X, (__v8hi)__X, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
}
static __inline__ __m256i __DEFAULT_FN_ATTRS
_mm256_broadcastd_epi32(__m128i __X)
{
- return (__m256i)__builtin_ia32_pbroadcastd256((__v4si)__X);
+ return (__m256i)__builtin_shufflevector((__v4si)__X, (__v4si)__X, 0, 0, 0, 0, 0, 0, 0, 0);
}
static __inline__ __m256i __DEFAULT_FN_ATTRS
_mm256_broadcastq_epi64(__m128i __X)
{
- return (__m256i)__builtin_ia32_pbroadcastq256(__X);
+ return (__m256i)__builtin_shufflevector(__X, __X, 0, 0, 0, 0);
}
static __inline__ __m128i __DEFAULT_FN_ATTRS
_mm_broadcastb_epi8(__m128i __X)
{
- return (__m128i)__builtin_ia32_pbroadcastb128((__v16qi)__X);
+ return (__m128i)__builtin_shufflevector((__v16qi)__X, (__v16qi)__X, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
}
static __inline__ __m128i __DEFAULT_FN_ATTRS
_mm_broadcastw_epi16(__m128i __X)
{
- return (__m128i)__builtin_ia32_pbroadcastw128((__v8hi)__X);
+ return (__m128i)__builtin_shufflevector((__v8hi)__X, (__v8hi)__X, 0, 0, 0, 0, 0, 0, 0, 0);
}
static __inline__ __m128i __DEFAULT_FN_ATTRS
_mm_broadcastd_epi32(__m128i __X)
{
- return (__m128i)__builtin_ia32_pbroadcastd128((__v4si)__X);
+ return (__m128i)__builtin_shufflevector((__v4si)__X, (__v4si)__X, 0, 0, 0, 0);
}
static __inline__ __m128i __DEFAULT_FN_ATTRS
_mm_broadcastq_epi64(__m128i __X)
{
- return (__m128i)__builtin_ia32_pbroadcastq128(__X);
+ return (__m128i)__builtin_shufflevector(__X, __X, 0, 0);
}
static __inline__ __m256i __DEFAULT_FN_ATTRS
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
