[X86][SSE] Replace PMULDQ GetDemandedBits combine with SimplifyMultipleUseDemandedBit...

Recommit rL367100 which was reverted at rL367141. Until PR42777 is fixed, we no longer get the benefits of peeking through bitcasts but it does still remove a GetDemandedBits user and gives us the equivalent combines.

llvm-svn: 367172

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 3086651..001862b 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -34381,6 +34381,18 @@ bool X86TargetLowering::SimplifyDemandedBitsForTargetNode(
     if (SimplifyDemandedBits(RHS, DemandedMask, OriginalDemandedElts, KnownOp,
                              TLO, Depth + 1))
       return true;
+
+    // Aggressively peek through ops to get at the demanded low bits.
+    SDValue DemandedLHS = SimplifyMultipleUseDemandedBits(
+        LHS, DemandedMask, OriginalDemandedElts, TLO.DAG, Depth + 1);
+    SDValue DemandedRHS = SimplifyMultipleUseDemandedBits(
+        RHS, DemandedMask, OriginalDemandedElts, TLO.DAG, Depth + 1);
+    if (DemandedLHS || DemandedRHS) {
+      DemandedLHS = DemandedLHS ? DemandedLHS : LHS;
+      DemandedRHS = DemandedRHS ? DemandedRHS : RHS;
+      return TLO.CombineTo(
+          Op, TLO.DAG.getNode(Opc, SDLoc(Op), VT, DemandedLHS, DemandedRHS));
+    }
     break;
   }
   case X86ISD::VSHLI: {
@@ -44220,15 +44232,6 @@ static SDValue combinePMULDQ(SDNode *N, SelectionDAG &DAG,
   if (ISD::isBuildVectorAllZeros(RHS.getNode()))
     return RHS;
 
-  // Aggressively peek through ops to get at the demanded low bits.
-  APInt DemandedMask = APInt::getLowBitsSet(64, 32);
-  SDValue DemandedLHS = DAG.GetDemandedBits(LHS, DemandedMask);
-  SDValue DemandedRHS = DAG.GetDemandedBits(RHS, DemandedMask);
-  if (DemandedLHS || DemandedRHS)
-    return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
-                       DemandedLHS ? DemandedLHS : LHS,
-                       DemandedRHS ? DemandedRHS : RHS);
-
   // PMULDQ/PMULUDQ only uses lower 32 bits from each vector element.
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   if (TLI.SimplifyDemandedBits(SDValue(N, 0), APInt::getAllOnesValue(64), DCI))