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

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@91269 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/clang/Basic/BuiltinsX86.def b/include/clang/Basic/BuiltinsX86.def
index adb1b7643f8ea50bb2d8059672d230614d39d70c..6315c16dd80a85b263606329546a44c02233eb0a 100644 (file)
--- a/include/clang/Basic/BuiltinsX86.def
+++ b/include/clang/Basic/BuiltinsX86.def
@@ -251,7 +251,7 @@ BUILTIN(__builtin_ia32_monitor, "vv*UiUi", "")
 BUILTIN(__builtin_ia32_mwait, "vUiUi", "")
 BUILTIN(__builtin_ia32_lddqu, "V16ccC*", "")
 BUILTIN(__builtin_ia32_palignr128, "V16cV16cV16cc", "")
-BUILTIN(__builtin_ia32_palignr, "V8cV8cV8cc", "")
+BUILTIN(__builtin_ia32_palignr, "V1LLiV1LLiV1LLic", "")
 BUILTIN(__builtin_ia32_insertps128, "V4fV4fV4fi", "")
 
 BUILTIN(__builtin_ia32_storelv4si, "vV2i*V2LLi", "")

diff --git a/lib/CodeGen/CGBuiltin.cpp b/lib/CodeGen/CGBuiltin.cpp
index 5d5caa2e9a7dd30e91421918226aa5ef7ee3b7e4..c70443245c76d99756807934fa0695639c8c0323 100644 (file)
--- a/lib/CodeGen/CGBuiltin.cpp
+++ b/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/lib/Headers/tmmintrin.h b/lib/Headers/tmmintrin.h
index 374a27ecd77bf08038bd4ec118fa3f35d61d3684..7adb776fef769624fa71a07b5decdae69585a991 100644 (file)
--- a/lib/Headers/tmmintrin.h
+++ b/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)

diff --git a/test/CodeGen/palignr.c b/test/CodeGen/palignr.c
index c0c7e773841eabe943725a5bfe2d549554704829..41e48bd2854d9b1fc620895d03b17bf48ddefa1e 100644 (file)
--- a/test/CodeGen/palignr.c
+++ b/test/CodeGen/palignr.c
@@ -1,8 +1,12 @@
 // RUN: clang-cc %s -triple=i686-apple-darwin -target-feature +ssse3 -O1 -S -o - | FileCheck %s
 
 #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*8)))
+typedef __attribute__((vector_size(8))) int int2;
 typedef __attribute__((vector_size(16))) int int4;
 
+// CHECK: palignr
+int2 mmx_align1(int2 a, int2 b) { return _mm_alignr_pi8(a, b, 7); }
 // CHECK: palignr
 int4 align1(int4 a, int4 b) { return _mm_alignr_epi8(a, b, 15); }
 // CHECK: ret