[llvm] r369780 - [SelectionDAG][X86] Enable iX SimplifyDemandedBits to vXi1 SimplifyDemandedVectorElts simplification. Add a hack to X86 to avoid a regression
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Fri Aug 23 10:14:58 PDT 2019
Author: ctopper
Date: Fri Aug 23 10:14:58 2019
New Revision: 369780
URL: http://llvm.org/viewvc/llvm-project?rev=369780&view=rev
Log:
[SelectionDAG][X86] Enable iX SimplifyDemandedBits to vXi1 SimplifyDemandedVectorElts simplification. Add a hack to X86 to avoid a regression
Patch showing the effect of enabling bool vector oversimplification.
Non-VLX builds can simplify a kshift shuffle, but VLX builds simplify:
insert_subvector v8i zeroinitializer, v2i --> insert_subvector v8i undef, v2i
Preventing the removal of the AND to clear the upper bits of result
Differential Revision: https://reviews.llvm.org/D53022
Modified:
llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp
llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
Modified: llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp?rev=369780&r1=369779&r2=369780&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp (original)
+++ llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp Fri Aug 23 10:14:58 2019
@@ -1819,9 +1819,7 @@ bool TargetLowering::SimplifyDemandedBit
// Bitcast from a vector using SimplifyDemanded Bits/VectorElts.
// Demand the elt/bit if any of the original elts/bits are demanded.
// TODO - bigendian once we have test coverage.
- // TODO - bool vectors once SimplifyDemandedVectorElts has SETCC support.
- if (SrcVT.isVector() && NumSrcEltBits > 1 &&
- (BitWidth % NumSrcEltBits) == 0 &&
+ if (SrcVT.isVector() && (BitWidth % NumSrcEltBits) == 0 &&
TLO.DAG.getDataLayout().isLittleEndian()) {
unsigned Scale = BitWidth / NumSrcEltBits;
unsigned NumSrcElts = SrcVT.getVectorNumElements();
Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=369780&r1=369779&r2=369780&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Fri Aug 23 10:14:58 2019
@@ -35380,6 +35380,26 @@ static SDValue combineBitcast(SDNode *N,
// type, widen both sides to avoid a trip through memory.
if ((SrcVT == MVT::v4i1 || SrcVT == MVT::v2i1) && VT.isScalarInteger() &&
Subtarget.hasAVX512()) {
+ // Use zeros for the widening if we already have some zeroes. This can
+ // allow SimplifyDemandedBits to remove scalar ANDs that may be down
+ // stream of this.
+ // FIXME: It might make sense to detect a concat_vectors with a mix of
+ // zeroes and undef and turn it into insert_subvector for i1 vectors as
+ // a separate combine. What we can't do is canonicalize the operands of
+ // such a concat or we'll get into a loop with SimplifyDemandedBits.
+ if (N0.getOpcode() == ISD::CONCAT_VECTORS) {
+ SDValue LastOp = N0.getOperand(N0.getNumOperands() - 1);
+ if (ISD::isBuildVectorAllZeros(LastOp.getNode())) {
+ SrcVT = LastOp.getValueType();
+ unsigned NumConcats = 8 / SrcVT.getVectorNumElements();
+ SmallVector<SDValue, 4> Ops(N0->op_begin(), N0->op_end());
+ Ops.resize(NumConcats, DAG.getConstant(0, dl, SrcVT));
+ N0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v8i1, Ops);
+ N0 = DAG.getBitcast(MVT::i8, N0);
+ return DAG.getNode(ISD::TRUNCATE, dl, VT, N0);
+ }
+ }
+
unsigned NumConcats = 8 / SrcVT.getVectorNumElements();
SmallVector<SDValue, 4> Ops(NumConcats, DAG.getUNDEF(SrcVT));
Ops[0] = N0;
More information about the llvm-commits
mailing list