Revert mmx palignr to use an intrinsic, since mmx shuffle patterns are missing.

llvm-svn: 91269

2 files changed, 5 insertions, 2 deletions

diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 5d5caa2e9a7..c70443245c7 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -805,8 +805,11 @@ Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
     Ops[0] = Builder.CreateBitCast(Ops[0], PtrTy);
     return Builder.CreateStore(Ops[1], Ops[0]);
   }
-  case X86::BI__builtin_ia32_palignr128:
   case X86::BI__builtin_ia32_palignr: {
+    Function *F = CGM.getIntrinsic(Intrinsic::x86_ssse3_palign_r);
+    return Builder.CreateCall(F, &Ops[0], &Ops[0] + Ops.size());
+  }
+  case X86::BI__builtin_ia32_palignr128: {
     unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
     
     // If palignr is shifting the pair of input vectors less than 17 bytes,
diff --git a/clang/lib/Headers/tmmintrin.h b/clang/lib/Headers/tmmintrin.h
index 374a27ecd77..7adb776fef7 100644
--- a/clang/lib/Headers/tmmintrin.h
+++ b/clang/lib/Headers/tmmintrin.h
@@ -67,7 +67,7 @@ _mm_abs_epi32(__m128i a)
 }
 
 #define _mm_alignr_epi8(a, b, n) (__builtin_ia32_palignr128((a), (b), (n)))
-#define _mm_alignr_pi8(a, b, n) (__builtin_ia32_palignr((a), (b), (n)))
+#define _mm_alignr_pi8(a, b, n) (__builtin_ia32_palignr((a), (b), (n*8)))
 
 static inline __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_hadd_epi16(__m128i a, __m128i b)