LLVM doesn't always optimize away the four loads from this:

     (__m128){ p[0], p[1], p[2], p[3] }

which produces really bad code. This could be done in instcombine, but it's
probably better to do it in the front-end instead.
<rdar://problem/9424836>

llvm-svn: 131237

3 files changed, 4 insertions, 2 deletions

diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 46546177ae5..494dfaeff77 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -2143,6 +2143,8 @@ Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
     // If palignr is shifting the pair of vectors more than 32 bytes, emit zero.
     return llvm::Constant::getNullValue(ConvertType(E->getType()));
   }
+  case X86::BI__builtin_ia32_loadups:
+  case X86::BI__builtin_ia32_loadupd:
   case X86::BI__builtin_ia32_loaddqu: {
     const llvm::Type *VecTy = ConvertType(E->getType());
     const llvm::Type *IntTy = llvm::IntegerType::get(getLLVMContext(), 128);
diff --git a/clang/lib/Headers/emmintrin.h b/clang/lib/Headers/emmintrin.h
index 62c10b5134e..746e717a309 100644
--- a/clang/lib/Headers/emmintrin.h
+++ b/clang/lib/Headers/emmintrin.h
@@ -466,7 +466,7 @@ _mm_loadr_pd(double const *dp)
 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 _mm_loadu_pd(double const *dp)
 {
-  return (__m128d){ dp[0], dp[1] };
+  return __builtin_ia32_loadupd(dp);
 }
 
 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
diff --git a/clang/lib/Headers/xmmintrin.h b/clang/lib/Headers/xmmintrin.h
index 00760ed6d1e..42dd3e8d3b8 100644
--- a/clang/lib/Headers/xmmintrin.h
+++ b/clang/lib/Headers/xmmintrin.h
@@ -539,7 +539,7 @@ _mm_load_ps(const float *p)
 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
 _mm_loadu_ps(const float *p)
 {
-  return (__m128){ p[0], p[1], p[2], p[3] };
+  return __builtin_ia32_loadups(p);
 }
 
 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))