Rewrite handling of 64-bit palignr intrinsics to be vector shuffles.

Stop multiplying constant by 8 accordingly in the header and change
intrinsic definition for what types we expect.

Add to existing palignr test to check that we're emitting the correct things.

llvm-svn: 101332

4 files changed, 49 insertions, 8 deletions

diff --git a/clang/include/clang/Basic/BuiltinsX86.def b/clang/include/clang/Basic/BuiltinsX86.def
index 5c75d3799bc..a878dd1bd1a 100644
--- a/clang/include/clang/Basic/BuiltinsX86.def
+++ b/clang/include/clang/Basic/BuiltinsX86.def
@@ -245,7 +245,7 @@ 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, "V1LLiV1LLiV1LLic", "")
+BUILTIN(__builtin_ia32_palignr, "V8cV8cV8cc", "")
 BUILTIN(__builtin_ia32_insertps128, "V4fV4fV4fi", "")
 
 BUILTIN(__builtin_ia32_storelv4si, "vV2i*V2LLi", "")
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 85ab1dceba9..95c41db86e0 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -982,8 +982,38 @@ Value *CodeGenFunction::EmitX86BuiltinExpr(unsigned BuiltinID,
     return Builder.CreateStore(Ops[1], Ops[0]);
   }
   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());
+    unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
+    
+    // If palignr is shifting the pair of input vectors less than 9 bytes,
+    // emit a shuffle instruction.
+    if (shiftVal <= 8) {
+      const llvm::Type *IntTy = llvm::Type::getInt32Ty(VMContext);
+
+      llvm::SmallVector<llvm::Constant*, 8> Indices;
+      for (unsigned i = 0; i != 8; ++i)
+        Indices.push_back(llvm::ConstantInt::get(IntTy, shiftVal + i));
+      
+      Value* SV = llvm::ConstantVector::get(Indices.begin(), Indices.size());
+      return Builder.CreateShuffleVector(Ops[1], Ops[0], SV, "palignr");
+    }
+    
+    // If palignr is shifting the pair of input vectors more than 8 but less
+    // than 16 bytes, emit a logical right shift of the destination.
+    if (shiftVal < 16) {
+      // MMX has these as 1 x i64 vectors for some odd optimization reasons.
+      const llvm::Type *EltTy = llvm::Type::getInt64Ty(VMContext);
+      const llvm::Type *VecTy = llvm::VectorType::get(EltTy, 1);
+      
+      Ops[0] = Builder.CreateBitCast(Ops[0], VecTy, "cast");
+      Ops[1] = llvm::ConstantInt::get(VecTy, (shiftVal-8) * 8);
+      
+      // create i32 constant
+      llvm::Function *F = CGM.getIntrinsic(Intrinsic::x86_mmx_psrl_q);
+      return Builder.CreateCall(F, &Ops[0], &Ops[0] + 2, "palignr");
+    }
+    
+    // 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_palignr128: {
     unsigned shiftVal = cast<llvm::ConstantInt>(Ops[2])->getZExtValue();
diff --git a/clang/lib/Headers/tmmintrin.h b/clang/lib/Headers/tmmintrin.h
index 09ebc237801..07fea1c98bf 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*8)))
+#define _mm_alignr_pi8(a, b, n) (__builtin_ia32_palignr((a), (b), (n)))
 
 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 _mm_hadd_epi16(__m128i a, __m128i b)
diff --git a/clang/test/CodeGen/palignr.c b/clang/test/CodeGen/palignr.c
index 627e309bd94..6297b2e990f 100644
--- a/clang/test/CodeGen/palignr.c
+++ b/clang/test/CodeGen/palignr.c
@@ -1,13 +1,9 @@
 // RUN: %clang_cc1 %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
 // CHECK: ret
@@ -17,3 +13,18 @@ int4 align2(int4 a, int4 b) { return _mm_alignr_epi8(a, b, 16); }
 int4 align3(int4 a, int4 b) { return _mm_alignr_epi8(a, b, 17); }
 // CHECK: xor
 int4 align4(int4 a, int4 b) { return _mm_alignr_epi8(a, b, 32); }
+
+#define _mm_alignr_pi8(a, b, n) (__builtin_ia32_palignr((a), (b), (n)))
+typedef __attribute__((vector_size(8))) int int2;
+
+// CHECK-NOT: palignr
+int2 align5(int2 a, int2 b) { return _mm_alignr_pi8(a, b, 8); }
+
+// CHECK: psrlq
+int2 align6(int2 a, int2 b) { return _mm_alignr_pi8(a, b, 9); }
+
+// CHECK: xor
+int2 align7(int2 a, int2 b) { return _mm_alignr_pi8(a, b, 16); }
+
+// CHECK: palignr
+int2 align8(int2 a, int2 b) { return _mm_alignr_pi8(a, b, 7); }
\ No newline at end of file