[X86] combineX86ShufflesRecursively - split the getTargetShuffleInputs call from...

Exposes an issue in getFauxShuffleMask where the OR(SHUFFLE,SHUFFLE) decode should always resolve zero/undef elements.

Part of the fix for PR43024 where ideally we shouldn't call resolveTargetShuffleAndZeroables for Depth == 0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@374928 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index df08ebfe07692abe9fafdb86e36a5d01776590b1..bdc05ad65f032ff9fbad79cc14a9cef47ec65adf 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -6932,9 +6932,9 @@ static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts,
     SmallVector<int, 64> SrcMask0, SrcMask1;
     SmallVector<SDValue, 2> SrcInputs0, SrcInputs1;
     if (!getTargetShuffleInputs(N0, SrcInputs0, SrcMask0, DAG, Depth + 1,
-                                ResolveKnownElts) ||
+                                true) ||
         !getTargetShuffleInputs(N1, SrcInputs1, SrcMask1, DAG, Depth + 1,
-                                ResolveKnownElts))
+                                true))
       return false;
     size_t MaskSize = std::max(SrcMask0.size(), SrcMask1.size());
     SmallVector<int, 64> Mask0, Mask1;
@@ -33012,11 +33012,17 @@ static SDValue combineX86ShufflesRecursively(
          "Can only combine shuffles of the same vector register size.");
 
   // Extract target shuffle mask and resolve sentinels and inputs.
+  // TODO - determine Op's demanded elts from RootMask.
   SmallVector<int, 64> OpMask;
   SmallVector<SDValue, 2> OpInputs;
-  if (!getTargetShuffleInputs(Op, OpInputs, OpMask, DAG, Depth))
+  APInt OpUndef, OpZero;
+  APInt OpDemandedElts = APInt::getAllOnesValue(VT.getVectorNumElements());
+  if (!getTargetShuffleInputs(Op, OpDemandedElts, OpInputs, OpMask, OpUndef,
+                              OpZero, DAG, Depth, false))
     return SDValue();
 
+  resolveTargetShuffleAndZeroables(OpMask, OpUndef, OpZero);
+
   // Add the inputs to the Ops list, avoiding duplicates.
   SmallVector<SDValue, 16> Ops(SrcOps.begin(), SrcOps.end());