[X86] Combine zext(packus(x),packus(y)) -> concat(x,y) (PR39637)

Its proving tricky to combine shuffles across multiple vector sizes, so for now I'm adding this more specific combine - the pattern is common enough to be worth it as a first step.

llvm-svn: 354757

1 files changed, 14 insertions, 0 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index ae7fd6c93b3..dc50eca4258 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -40328,6 +40328,20 @@ static SDValue combineZext(SDNode *N, SelectionDAG &DAG,
   if (SDValue R = combineOrCmpEqZeroToCtlzSrl(N, DAG, DCI, Subtarget))
     return R;
 
+  // TODO: Combine with any target/faux shuffle.
+  if (N0.getOpcode() == X86ISD::PACKUS && N0.getValueSizeInBits() == 128 &&
+      VT.getScalarSizeInBits() == N0.getOperand(0).getScalarValueSizeInBits()) {
+    SDValue N00 = N0.getOperand(0);
+    SDValue N01 = N0.getOperand(1);
+    unsigned NumSrcElts = N00.getValueType().getVectorNumElements();
+    unsigned NumSrcEltBits = N00.getScalarValueSizeInBits();
+    APInt ZeroMask = APInt::getHighBitsSet(NumSrcEltBits, NumSrcEltBits / 2);
+    if ((N00.isUndef() || DAG.MaskedValueIsZero(N00, ZeroMask)) &&
+        (N01.isUndef() || DAG.MaskedValueIsZero(N01, ZeroMask))) {
+      return concatSubVectors(N00, N01, VT, NumSrcElts * 2, DAG, dl, 128);
+    }
+  }
+
   return SDValue();
 }