Represent 256-bit unaligned loads natively and remove the builtins. Similar change was made for 128-bit versions a while back.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@148919 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 37 insertions, 6 deletions

diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def
index 17a7cc63e1..2356b54d3c 100644
--- a/include/clang/Basic/BuiltinsX86.def
+++ b/include/clang/Basic/BuiltinsX86.def
@@ -453,11 +453,8 @@ BUILTIN(__builtin_ia32_vbroadcastsd256, "V4ddC*", "")
 BUILTIN(__builtin_ia32_vbroadcastss256, "V8ffC*", "")
 BUILTIN(__builtin_ia32_vbroadcastf128_pd256, "V4dV2dC*", "")
 BUILTIN(__builtin_ia32_vbroadcastf128_ps256, "V8fV4fC*", "")
-BUILTIN(__builtin_ia32_loadupd256, "V4ddC*", "")
-BUILTIN(__builtin_ia32_loadups256, "V8ffC*", "")
 BUILTIN(__builtin_ia32_storeupd256, "vd*V4d", "")
 BUILTIN(__builtin_ia32_storeups256, "vf*V8f", "")
-BUILTIN(__builtin_ia32_loaddqu256, "V32ccC*", "")
 BUILTIN(__builtin_ia32_storedqu256, "vc*V32c", "")
 BUILTIN(__builtin_ia32_lddqu256, "V32ccC*", "")
 BUILTIN(__builtin_ia32_movntdq256, "vV4LLi*V4LLi", "")
diff --git a/lib/Headers/avxintrin.h b/lib/Headers/avxintrin.h
index 620ee0fbb0..ce4b2264bf 100644
--- a/lib/Headers/avxintrin.h
+++ b/lib/Headers/avxintrin.h
@@ -751,13 +751,19 @@ _mm256_load_ps(float const *p)
 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
 _mm256_loadu_pd(double const *p)
 {
-  return (__m256d)__builtin_ia32_loadupd256(p);
+  struct __loadu_pd {
+    __m256d v;
+  } __attribute__((packed, may_alias));
+  return ((struct __loadu_pd*)p)->v;
 }
 
 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
 _mm256_loadu_ps(float const *p)
 {
-  return (__m256)__builtin_ia32_loadups256(p);
+  struct __loadu_ps {
+    __m256 v;
+  } __attribute__((packed, may_alias));
+  return ((struct __loadu_ps*)p)->v;
 }
 
 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
@@ -769,7 +775,10 @@ _mm256_load_si256(__m256i const *p)
 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
 _mm256_loadu_si256(__m256i const *p)
 {
-  return (__m256i)__builtin_ia32_loaddqu256((char const *)p);
+  struct __loadu_si256 {
+    __m256i v;
+  } __attribute__((packed, may_alias));
+  return ((struct __loadu_si256*)p)->v;
 }
 
 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
diff --git a/test/CodeGen/avx-builtins.c b/test/CodeGen/avx-builtins.c
new file mode 100644
index 0000000000..f03141225c
--- /dev/null
+++ b/test/CodeGen/avx-builtins.c
@@ -0,0 +1,25 @@
+// RUN: %clang_cc1 %s -O3 -triple=x86_64-apple-darwin -target-feature +avx -emit-llvm -o - | FileCheck %s
+
+// Don't include mm_malloc.h, it's system specific.
+#define __MM_MALLOC_H
+
+#include <immintrin.h>
+
+//
+// Test LLVM IR codegen of shuffle instructions
+//
+
+__m256 test__mm256_loadu_ps(void* p) {
+  // CHECK: load <8 x float>* %{.*}, align 1
+  return _mm256_loadu_ps(p);
+}
+
+__m256d test__mm256_loadu_pd(void* p) {
+  // CHECK: load <4 x double>* %{.*}, align 1
+  return _mm256_loadu_pd(p);
+}
+
+__m256i test__mm256_loadu_si256(void* p) {
+  // CHECK: load <4 x i64>* %0, align 1
+  return _mm256_loadu_si256(p);
+}