[X86][AVX2] Replace avx2.pbroadcast / avx2.vbroadcast intrinsics usage in avx2intrin.h with __builtin_shufflevector

[RKSimon]

Summary

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.

[ab]

Heh, this is D10556, no? :P

[RKSimon]

In D12143#228006, @ab wrote:

Heh, this is D10556, no? :P

Yes you're right (and you remembered to kill the builtin defs) - please can you add some more reviewers so that we can get it dealt with along with D10555?

lib/Headers/avx2intrin.h

 _mm256_stream_load_si256(__m256i *__V)
 {
   return (__m256i)__builtin_ia32_movntdqa256((__v4di *)__V);
 }
 
 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
 _mm_broadcastsd_pd(__m128d __a)
 {
   return __builtin_shufflevector(__a, __a, 0, 0);
 }
 
 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
 _mm256_broadcastsi128_si256(__m128i __X)
 {
   return (__m256i)__builtin_shufflevector(__X, __X, 0, 1, 0, 1);
 }
 
(__m256i)__builtin_shufflevector((__v8si)__V1, (__v8si)__V2, \
                                    (((M) & 0x10) ? 12 : 4), \
                                    (((M) & 0x20) ? 13 : 5), \
                                    (((M) & 0x40) ? 14 : 6), \
                                    (((M) & 0x80) ? 15 : 7)); })
 
 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
 _mm256_permutevar8x32_epi32(__m256i __a, __m256i __b)
 {
   return (__m256i)__builtin_ia32_permvarsi256((__v8si)__a, (__v8si)__b);
 }
 

test/CodeGen/avx2-builtins.c

 }
 
 __m256i test_mm256_stream_load_si256(__m256i *a) {
   // CHECK: @llvm.x86.avx2.movntdqa
   return _mm256_stream_load_si256(a);
 }
 
 __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);
 }
 
 __m128d test_mm_broadcastsd_pd(__m128d a) {
   // CHECK: shufflevector <2 x double> %{{.*}}, <2 x double> %{{.*}}, <2 x i32> zeroinitializer
   return _mm_broadcastsd_pd(a);
 }
 
 __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);
 }
 
 __m256i test_mm256_broadcastsi128_si256(__m128i a) {
   // CHECK: shufflevector <2 x i64> %{{.*}}, <2 x i64> %{{.*}}, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
   return _mm256_broadcastsi128_si256(a);
 }
 
 __m128i test_mm_blend_epi32(__m128i a, __m128i b) {
   // CHECK-LABEL: test_mm_blend_epi32
   // CHECK-NOT: @llvm.x86.avx2.pblendd.128
   // CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> %{{.*}}, <4 x i32> <i32 4, i32 1, i32 6, i32 3>
   return _mm_blend_epi32(a, b, 0x35);
 }
 
 __m256i test_mm256_blend_epi32(__m256i a, __m256i b) {
   // CHECK-LABEL: test_mm256_blend_epi32
   // CHECK-NOT: @llvm.x86.avx2.pblendd.256
   // CHECK: shufflevector <8 x i32> %{{.*}}, <8 x i32> %{{.*}}, <8 x i32> <i32 8, i32 1, i32 10, i32 3, i32 12, i32 13, i32 6, i32 7>
   return _mm256_blend_epi32(a, b, 0x35);
 }
 
 __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);
 }
 
 __m256i test_mm256_permutevar8x32_epi32(__m256i a, __m256i b) {
   // CHECK: @llvm.x86.avx2.permd
   return _mm256_permutevar8x32_epi32(a, b);
 }