X86: Prefer using VPSHUFD over VPERMIL because it has better throughput.

llvm-svn: 169624

lib/Target/X86/X86ISelLowering.cpp

@@ -6781,12 +6781,13 @@ X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
 
     unsigned TargetMask = getShuffleSHUFImmediate(SVOp);
 
-    if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
-      return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1, TargetMask, DAG);
-
     if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
       return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
 
+    if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
+      return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1, TargetMask,
+                                  DAG);
+
     return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1,
                                 TargetMask, DAG);
   }

test/CodeGen/X86/avx-intrinsics-x86.ll

@@ -2348,7 +2348,7 @@ declare <4 x double> @llvm.x86.avx.vpermil.pd.256(<4 x double>, i8) nounwind rea
 
 
 define <4 x float> @test_x86_avx_vpermil_ps(<4 x float> %a0) {
-  ; CHECK: vpermilps
+  ; CHECK: vpshufd
   %res = call <4 x float> @llvm.x86.avx.vpermil.ps(<4 x float> %a0, i8 7) ; <<4 x float>> [#uses=1]
   ret <4 x float> %res
 }

test/CodeGen/X86/avx-shuffle.ll

@@ -6,7 +6,7 @@ define <4 x float> @test1(<4 x float> %a) nounwind {
   ret <4 x float> %b
 ; CHECK: test1:
 ; CHECK: vshufps
-; CHECK: vpermilps
+; CHECK: vpshufd
 }
 
 ; rdar://10538417
@@ -106,7 +106,7 @@ define <4 x float> @test11(<4 x float> %a) nounwind  {
 
 define <4 x float> @test12(<4 x float>* %a) nounwind  {
 ; CHECK: test12
-; CHECK: vpermilps $27, (
+; CHECK: vpshufd
   %tmp0 = load <4 x float>* %a
   %tmp1 = shufflevector <4 x float> %tmp0, <4 x float> undef, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
   ret <4 x float> %tmp1

test/CodeGen/X86/avx-splat.ll

@@ -84,7 +84,7 @@ define <8 x float> @funcF(i32 %val) nounwind {
   ret <8 x float> %tmp
 }
 
-; CHECK: vpermilps  $0
+; CHECK: vpshufd  $0
 ; CHECK-NEXT: vinsertf128  $1
 define <8 x float> @funcG(<8 x float> %a) nounwind uwtable readnone ssp {
 entry:
@@ -93,7 +93,7 @@ entry:
 }
 
 ; CHECK: vextractf128  $1
-; CHECK-NEXT: vpermilps  $85
+; CHECK-NEXT: vpshufd
 ; CHECK-NEXT: vinsertf128  $1
 define <8 x float> @funcH(<8 x float> %a) nounwind uwtable readnone ssp {
 entry: