From: Craig Topper <craig.topper@intel.com>
Date: Fri, 22 Feb 2019 01:49:53 +0000 (+0000)
Subject: [X86] Add a DAG combine for (aext_vector_inreg (aext_vector_inreg X)) -> (aext_vector... 
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=c12778d01415b5b489712e936100460ae3ac4948;p=llvm

[X86] Add a DAG combine for (aext_vector_inreg (aext_vector_inreg X)) -> (aext_vector_inreg X) to fix a regression from my previous commit.

Type legalization is causing two nodes to be created here, but we can use a single node to extend from v8i16 to v2i64.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@354648 91177308-0d34-0410-b5e6-96231b3b80d8
---

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index dfc7d8ba200..62b044caae2 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1873,6 +1873,7 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
   setTargetDAGCombine(ISD::ANY_EXTEND);
   setTargetDAGCombine(ISD::SIGN_EXTEND);
   setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
+  setTargetDAGCombine(ISD::ANY_EXTEND_VECTOR_INREG);
   setTargetDAGCombine(ISD::SINT_TO_FP);
   setTargetDAGCombine(ISD::UINT_TO_FP);
   setTargetDAGCombine(ISD::SETCC);
@@ -42108,6 +42109,25 @@ static SDValue combinePMULDQ(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+static SDValue combineExtInVec(SDNode *N, SelectionDAG &DAG) {
+  // Disabling for widening legalization for now. We can enable if we find a
+  // case that needs it. Otherwise it can be deleted when we switch to
+  // widening legalization.
+  if (ExperimentalVectorWideningLegalization)
+    return SDValue();
+
+  EVT VT = N->getValueType(0);
+  SDValue In = N->getOperand(0);
+
+  // Combine (ext_invec (ext_invec X)) -> (ext_invec X)
+  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+  if (In.getOpcode() == N->getOpcode() &&
+      TLI.isTypeLegal(VT) && TLI.isTypeLegal(In.getOperand(0).getValueType()))
+    return DAG.getNode(N->getOpcode(), SDLoc(N), VT, In.getOperand(0));
+
+  return SDValue();
+}
+
 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
                                              DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
@@ -42169,6 +42189,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::ZERO_EXTEND:    return combineZext(N, DAG, DCI, Subtarget);
   case ISD::SIGN_EXTEND:    return combineSext(N, DAG, DCI, Subtarget);
   case ISD::SIGN_EXTEND_INREG: return combineSignExtendInReg(N, DAG, Subtarget);
+  case ISD::ANY_EXTEND_VECTOR_INREG: return combineExtInVec(N, DAG);
   case ISD::SETCC:          return combineSetCC(N, DAG, Subtarget);
   case X86ISD::SETCC:       return combineX86SetCC(N, DAG, Subtarget);
   case X86ISD::BRCOND:      return combineBrCond(N, DAG, Subtarget);
diff --git a/test/CodeGen/X86/vector-narrow-binop.ll b/test/CodeGen/X86/vector-narrow-binop.ll
index a452ca44476..2ef7fa7f941 100644
--- a/test/CodeGen/X86/vector-narrow-binop.ll
+++ b/test/CodeGen/X86/vector-narrow-binop.ll
@@ -147,16 +147,15 @@ define <2 x i8> @PR39893_2(<2 x float> %x) {
 ; SSE-NEXT:    xorps %xmm1, %xmm1
 ; SSE-NEXT:    subps %xmm0, %xmm1
 ; SSE-NEXT:    punpcklbw {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3],xmm1[4],xmm0[4],xmm1[5],xmm0[5],xmm1[6],xmm0[6],xmm1[7],xmm0[7]
-; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,1,0,3]
-; SSE-NEXT:    pshufhw {{.*#+}} xmm0 = xmm0[0,1,2,3,5,5,6,7]
+; SSE-NEXT:    punpcklwd {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1],xmm1[2],xmm0[2],xmm1[3],xmm0[3]
+; SSE-NEXT:    pshufd {{.*#+}} xmm0 = xmm1[0,1,1,3]
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: PR39893_2:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
 ; AVX-NEXT:    vsubps %xmm0, %xmm1, %xmm0
-; AVX-NEXT:    vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
-; AVX-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
+; AVX-NEXT:    vpmovzxbq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,zero,zero,zero,zero,xmm0[1],zero,zero,zero,zero,zero,zero,zero
 ; AVX-NEXT:    retq
   %fsub = fsub <2 x float> zeroinitializer, %x
   %bc = bitcast <2 x float> %fsub to <8 x i8>