diff options
author | Craig Topper <craig.topper@gmail.com> | 2017-03-13 18:17:46 +0000 |
---|---|---|
committer | Craig Topper <craig.topper@gmail.com> | 2017-03-13 18:17:46 +0000 |
commit | eb7ea28bdd71c97bb7ddd22cd671c606d5d9956e (patch) | |
tree | 8b55f92b8fe60c1706f9f2e83ce6509a89a1063f /llvm/lib/Target/X86/X86ISelLowering.cpp | |
parent | d81f557fe202572609a5c7a020482e3767488f88 (diff) | |
download | bcm5719-llvm-eb7ea28bdd71c97bb7ddd22cd671c606d5d9956e.tar.gz bcm5719-llvm-eb7ea28bdd71c97bb7ddd22cd671c606d5d9956e.zip |
[AVX-512] If gather mask is all ones, force the input to a zero vector.
We were already forcing undef inputs to become a zero vector, this now catches an all ones mask too.
Ideally we'd use undef and let execution dep fix handle picking the best register/clearance for the undef, but I don't think it can handle the early clobber today.
llvm-svn: 297651
Diffstat (limited to 'llvm/lib/Target/X86/X86ISelLowering.cpp')
-rw-r--r-- | llvm/lib/Target/X86/X86ISelLowering.cpp | 5 |
1 files changed, 4 insertions, 1 deletions
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index cf8304e565a..56f62dce002 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -19868,7 +19868,10 @@ static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG, SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other); SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32); SDValue Segment = DAG.getRegister(0, MVT::i32); - if (Src.isUndef()) + // If source is undef or we know it won't be used, use a zero vector + // to break register dependency. + // TODO: use undef instead and let ExeDepsFix deal with it? + if (Src.isUndef() || ISD::isBuildVectorAllOnes(VMask.getNode())) Src = getZeroVector(Op.getSimpleValueType(), Subtarget, DAG, dl); SDValue Ops[] = {Src, VMask, Base, Scale, Index, Disp, Segment, Chain}; SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops); |