Handle (shl (anyext (shr ...))) in SimpilfyDemandedBits
This is really an extension of the current (shl (shr ...)) -> shl optimization.
The main difference is that certain upper bits must also not be demanded.
The motivating examples are the first two in the testcase, which occur
in llvmpipe output.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@192783 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index 17631c0..9d6a3b4 100644
--- a/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -635,6 +635,31 @@
TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(),
NarrowShl));
}
+ // Repeat the SHL optimization above in cases where an extension
+ // intervenes: (shl (anyext (shr x, c1)), c2) to
+ // (shl (anyext x), c2-c1). This requires that the bottom c1 bits
+ // aren't demanded (as above) and that the shifted upper c1 bits of
+ // x aren't demanded.
+ if (InOp.hasOneUse() &&
+ InnerOp.getOpcode() == ISD::SRL &&
+ InnerOp.hasOneUse() &&
+ isa<ConstantSDNode>(InnerOp.getOperand(1))) {
+ uint64_t InnerShAmt = cast<ConstantSDNode>(InnerOp.getOperand(1))
+ ->getZExtValue();
+ if (InnerShAmt < ShAmt &&
+ InnerShAmt < InnerBits &&
+ NewMask.lshr(InnerBits - InnerShAmt + ShAmt) == 0 &&
+ NewMask.trunc(ShAmt) == 0) {
+ SDValue NewSA =
+ TLO.DAG.getConstant(ShAmt - InnerShAmt,
+ Op.getOperand(1).getValueType());
+ EVT VT = Op.getValueType();
+ SDValue NewExt = TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT,
+ InnerOp.getOperand(0));
+ return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, VT,
+ NewExt, NewSA));
+ }
+ }
}
KnownZero <<= SA->getZExtValue();