Patched clang to emit x86 blends as shufflevectors.

Summary:
Most of the clang header patch by Simon Pilgrim @ SCEE.
Also fixed (or added) clang tests for these intrinsics.

LLVM tests to make sure we get the blend instruction out of these
shufflevectors are at http://reviews.llvm.org/D3600

Reviewers: eli.friedman, craig.topper, rafael

Subscribers: cfe-commits

Differential Revision: http://reviews.llvm.org/D3601

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@208664 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Headers/avx2intrin.h b/lib/Headers/avx2intrin.h
index 95744693ac7768eca0a849bc79f56a9fa973db8d..394fdfee96522ca4b4cd5cfe44f7a97f1b82f006 100644 (file)
--- a/lib/Headers/avx2intrin.h
+++ b/lib/Headers/avx2intrin.h
@@ -160,7 +160,23 @@ _mm256_blendv_epi8(__m256i __V1, __m256i __V2, __m256i __M)
 #define _mm256_blend_epi16(V1, V2, M) __extension__ ({ \
   __m256i __V1 = (V1); \
   __m256i __V2 = (V2); \
-  (__m256i)__builtin_ia32_pblendw256((__v16hi)__V1, (__v16hi)__V2, (M)); })
+  (__m256d)__builtin_shufflevector((__v16hi)__V1, (__v16hi)__V2, \
+                                   (((M) & 0x01) ? 16 : 0), \
+                                   (((M) & 0x02) ? 17 : 1), \
+                                   (((M) & 0x04) ? 18 : 2), \
+                                   (((M) & 0x08) ? 19 : 3), \
+                                   (((M) & 0x10) ? 20 : 4), \
+                                   (((M) & 0x20) ? 21 : 5), \
+                                   (((M) & 0x40) ? 22 : 6), \
+                                   (((M) & 0x80) ? 23 : 7), \
+                                   (((M) & 0x01) ? 24 : 8), \
+                                   (((M) & 0x02) ? 25 : 9), \
+                                   (((M) & 0x04) ? 26 : 10), \
+                                   (((M) & 0x08) ? 27 : 11), \
+                                   (((M) & 0x10) ? 28 : 12), \
+                                   (((M) & 0x20) ? 29 : 13), \
+                                   (((M) & 0x40) ? 30 : 14), \
+                                   (((M) & 0x80) ? 31 : 15)); })
 
 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
 _mm256_cmpeq_epi8(__m256i __a, __m256i __b)
@@ -761,12 +777,24 @@ _mm256_broadcastsi128_si256(__m128i __X)
 #define _mm_blend_epi32(V1, V2, M) __extension__ ({ \
   __m128i __V1 = (V1); \
   __m128i __V2 = (V2); \
-  (__m128i)__builtin_ia32_pblendd128((__v4si)__V1, (__v4si)__V2, (M)); })
+  (__m128i)__builtin_shufflevector((__v4si)__V1, (__v4si)__V2, \
+                                   (((M) & 0x01) ? 4 : 0), \
+                                   (((M) & 0x02) ? 5 : 1), \
+                                   (((M) & 0x04) ? 6 : 2), \
+                                   (((M) & 0x08) ? 7 : 3)); })
 
 #define _mm256_blend_epi32(V1, V2, M) __extension__ ({ \
   __m256i __V1 = (V1); \
   __m256i __V2 = (V2); \
-  (__m256i)__builtin_ia32_pblendd256((__v8si)__V1, (__v8si)__V2, (M)); })
+  (__m256i)__builtin_shufflevector((__v8si)__V1, (__v8si)__V2, \
+                                   (((M) & 0x01) ?  8 : 0), \
+                                   (((M) & 0x02) ?  9 : 1), \
+                                   (((M) & 0x04) ? 10 : 2), \
+                                   (((M) & 0x08) ? 11 : 3), \
+                                   (((M) & 0x10) ? 12 : 4), \
+                                   (((M) & 0x20) ? 13 : 5), \
+                                   (((M) & 0x40) ? 14 : 6), \
+                                   (((M) & 0x80) ? 15 : 7)); })
 
 static __inline__ __m256i __attribute__((__always_inline__, __nodebug__))
 _mm256_broadcastb_epi8(__m128i __X)

diff --git a/lib/Headers/avxintrin.h b/lib/Headers/avxintrin.h
index 141c4d994bbcf97e35062f6c70382192fb48a72e..3d50439d36aa11c2c9dac0dc1435966d2cd07d41 100644 (file)
--- a/lib/Headers/avxintrin.h
+++ b/lib/Headers/avxintrin.h
@@ -308,12 +308,24 @@ _mm256_permutevar_ps(__m256 __a, __m256i __c)
 #define _mm256_blend_pd(V1, V2, M) __extension__ ({ \
   __m256d __V1 = (V1); \
   __m256d __V2 = (V2); \
-  (__m256d)__builtin_ia32_blendpd256((__v4df)__V1, (__v4df)__V2, (M)); })
+  (__m256d)__builtin_shufflevector((__v4df)__V1, (__v4df)__V2, \
+                                   (((M) & 0x01) ? 4 : 0), \
+                                   (((M) & 0x02) ? 5 : 1), \
+                                   (((M) & 0x04) ? 6 : 2), \
+                                   (((M) & 0x08) ? 7 : 3)); })
 
 #define _mm256_blend_ps(V1, V2, M) __extension__ ({ \
   __m256 __V1 = (V1); \
   __m256 __V2 = (V2); \
-  (__m256)__builtin_ia32_blendps256((__v8sf)__V1, (__v8sf)__V2, (M)); })
+  (__m256)__builtin_shufflevector((__v8sf)__V1, (__v8sf)__V2, \
+                                  (((M) & 0x01) ?  8 : 0), \
+                                  (((M) & 0x02) ?  9 : 1), \
+                                  (((M) & 0x04) ? 10 : 2), \
+                                  (((M) & 0x08) ? 11 : 3), \
+                                  (((M) & 0x10) ? 12 : 4), \
+                                  (((M) & 0x20) ? 13 : 5), \
+                                  (((M) & 0x40) ? 14 : 6), \
+                                  (((M) & 0x80) ? 15 : 7)); })
 
 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
 _mm256_blendv_pd(__m256d __a, __m256d __b, __m256d __c)

diff --git a/lib/Headers/smmintrin.h b/lib/Headers/smmintrin.h
index 53b3ccb4310c70bed2a52258d48a720ba62b2c32..6e35734f6e0d95402fc4b42a4097f1666043df3b 100644 (file)
--- a/lib/Headers/smmintrin.h
+++ b/lib/Headers/smmintrin.h
@@ -79,12 +79,18 @@
 #define _mm_blend_pd(V1, V2, M) __extension__ ({ \
   __m128d __V1 = (V1); \
   __m128d __V2 = (V2); \
-  (__m128d) __builtin_ia32_blendpd ((__v2df)__V1, (__v2df)__V2, (M)); })
+  (__m128d)__builtin_shufflevector((__v2df)__V1, (__v2df)__V2, \
+                                   (((M) & 0x01) ? 2 : 0), \
+                                   (((M) & 0x02) ? 3 : 1)); })
 
 #define _mm_blend_ps(V1, V2, M) __extension__ ({ \
   __m128 __V1 = (V1); \
   __m128 __V2 = (V2); \
-  (__m128) __builtin_ia32_blendps ((__v4sf)__V1, (__v4sf)__V2, (M)); })
+  (__m128)__builtin_shufflevector((__v4sf)__V1, (__v4sf)__V2, \
+                                  (((M) & 0x01) ? 4 : 0), \
+                                  (((M) & 0x02) ? 5 : 1), \
+                                  (((M) & 0x04) ? 6 : 2), \
+                                  (((M) & 0x08) ? 7 : 3)); })
 
 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_blendv_pd (__m128d __V1, __m128d __V2, __m128d __M)
@@ -110,7 +116,15 @@ _mm_blendv_epi8 (__m128i __V1, __m128i __V2, __m128i __M)
 #define _mm_blend_epi16(V1, V2, M) __extension__ ({ \
   __m128i __V1 = (V1); \
   __m128i __V2 = (V2); \
-  (__m128i) __builtin_ia32_pblendw128 ((__v8hi)__V1, (__v8hi)__V2, (M)); })
+  (__m128i)__builtin_shufflevector((__v8hi)__V1, (__v8hi)__V2, \
+                                   (((M) & 0x01) ?  8 : 0), \
+                                   (((M) & 0x02) ?  9 : 1), \
+                                   (((M) & 0x04) ? 10 : 2), \
+                                   (((M) & 0x08) ? 11 : 3), \
+                                   (((M) & 0x10) ? 12 : 4), \
+                                   (((M) & 0x20) ? 13 : 5), \
+                                   (((M) & 0x40) ? 14 : 6), \
+                                   (((M) & 0x80) ? 15 : 7)); })
 
 /* SSE4 Dword Multiply Instructions.  */
 static __inline__  __m128i __attribute__((__always_inline__, __nodebug__))

diff --git a/test/CodeGen/avx-builtins.c b/test/CodeGen/avx-builtins.c
index c88946fe856750a2b27731dd7df54d4351a8b6a5..5b5b6a56b4b6d054335903f2607b79e74f133e40 100644 (file)
--- a/test/CodeGen/avx-builtins.c
+++ b/test/CodeGen/avx-builtins.c
@@ -111,3 +111,15 @@ int test_extract_epi8(__m256i __a) {
   // CHECK: extractelement <32 x i8> %{{.*}}, i32 0
   return _mm256_extract_epi8(__a, 32);
 }
+
+__m256d test_256_blend_pd(__m256d __a, __m256d __b) {
+  // CHECK-LABEL: @test_256_blend_pd
+  // CHECK: shufflevector <4 x double> %{{.*}}, <4 x double> %{{.*}}, <4 x i32> <i32 4, i32 1, i32 6, i32 3>
+  return _mm256_blend_pd(__a, __b, 0x35);
+}
+
+__m256 test_256_blend_ps(__m256 __a, __m256 __b) {
+  // CHECK-LABEL: @test_256_blend_ps
+  // CHECK: shufflevector <8 x float> %{{.*}}, <8 x float> %{{.*}}, <8 x i32> <i32 8, i32 1, i32 10, i32 3, i32 12, i32 13, i32 6, i32 7>
+  return _mm256_blend_ps(__a, __b, 0x35);
+}

diff --git a/test/CodeGen/avx2-builtins.c b/test/CodeGen/avx2-builtins.c
index 948483db8d3d2e2b875ef236537025b8c847e4e8..03b4775d21b7001df875092d8b8c5c1ff355585d 100644 (file)
--- a/test/CodeGen/avx2-builtins.c
+++ b/test/CodeGen/avx2-builtins.c
@@ -5,6 +5,11 @@
 
 #include <immintrin.h>
 
+// FIXME: We should lower as many of these as possible to LLVM IR without
+// intrinsics which will allow us to remove redundant intrinsics in the
+// future. Taking care that we manage to optimize them afterwards and emit
+// the actual instruction (or better code).
+
 __m256i test_mm256_mpsadbw_epu8(__m256i x, __m256i y) {
   // CHECK: @llvm.x86.avx2.mpsadbw({{.*}}, {{.*}}, i32 3)
   return _mm256_mpsadbw_epu8(x, y, 3);
@@ -177,7 +182,8 @@ __m256i test_mm256_blendv_epi8(__m256i a, __m256i b, __m256i m) {
 }
 
 __m256i test_mm256_blend_epi16(__m256i a, __m256i b) {
-  // CHECK: @llvm.x86.avx2.pblendw(<16 x i16> %{{.*}}, <16 x i16> %{{.*}}, i32 2)
+  // CHECK-LABEL: test_mm256_blend_epi16
+  // CHECK: shufflevector <16 x i16> %{{.*}}, <16 x i16> %{{.*}}, <16 x i32> <i32 0, i32 17, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 25, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
   return _mm256_blend_epi16(a, b, 2);
 }
 
@@ -612,13 +618,13 @@ __m256i test_mm256_broadcastsi128_si256(__m128i a) {
 }
 
 __m128i test_mm_blend_epi32(__m128i a, __m128i b) {
-  // CHECK: @llvm.x86.avx2.pblendd.128
-  return _mm_blend_epi32(a, b, 57);
+  // 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: @llvm.x86.avx2.pblendd.256
-  return _mm256_blend_epi32(a, b, 57);
+  // 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) {

diff --git a/test/CodeGen/sse-builtins.c b/test/CodeGen/sse-builtins.c
index 1f5cb8e9e1074d2e09aa6c230be11ad9e7042680..874cf6dee05f497efca778a35932ee2fec0cc94a 100644 (file)
--- a/test/CodeGen/sse-builtins.c
+++ b/test/CodeGen/sse-builtins.c
@@ -237,3 +237,21 @@ void test_insert_epi32(__m128i __a, int b) {
   // CHECK: insertelement <4 x i32> %{{.*}}, i32 %{{.*}}, i32 0
    _mm_insert_epi32(__a, b, 4);
 }
+
+__m128d test_blend_pd(__m128d V1, __m128d V2) {
+  // CHECK-LABEL: @test_blend_pd
+  // CHECK: shufflevector <2 x double> %{{.*}}, <2 x double> %{{.*}}, <2 x i32> <i32 2, i32 1>
+  return _mm_blend_pd(V1, V2, 1);
+}
+
+__m128 test_blend_ps(__m128 V1, __m128 V2) {
+  // CHECK-LABEL: @test_blend_ps
+  // CHECK: shufflevector <4 x float> %{{.*}}, <4 x float> %{{.*}}, <4 x i32> <i32 4, i32 1, i32 6, i32 3>
+  return _mm_blend_ps(V1, V2, 5);
+}
+
+__m128i test_blend_epi16(__m128i V1, __m128i V2) {
+  // CHECK-LABEL: @test_blend_epi16
+  // CHECK: shufflevector <8 x i16> %{{.*}}, <8 x i16> %{{.*}}, <8 x i32> <i32 0, i32 9, i32 2, i32 11, i32 4, i32 13, i32 6, i32 7>
+  return _mm_blend_epi16(V1, V2, 42);
+}