[llvm-commits] [llvm] r58825 - /llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp
Evan Cheng
evan.cheng at apple.com
Thu Nov 6 17:50:16 PST 2008
On Nov 6, 2008, at 5:28 PM, Dale Johannesen wrote:
> Author: johannes
> Date: Thu Nov 6 19:28:02 2008
> New Revision: 58825
>
> URL: http://llvm.org/viewvc/llvm-project?rev=58825&view=rev
> Log:
> When we're doing a compare of load-AND-constant to 0
> (e.g. a bitfield test) narrow the load as much as possible.
> The has the potential to avoid unnecessary partial-word
> load-after-store conflicts, which cause stalls on several targets.
> Also a size win on x86 (testb vs testl).
This can a loss if the load is narrowed to i16 though. Assuming we
have the checks in place, this is probably ok. But I still wonder if
this is the approach we ought to take. Perhaps we need to attack all
of the load-after-store issues?
Evan
>
>
>
> Modified:
> llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp
>
> Modified: llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp?rev=58825&r1=58824&r2=58825&view=diff
>
> =
> =
> =
> =
> =
> =
> =
> =
> ======================================================================
> --- llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp (original)
> +++ llvm/trunk/lib/CodeGen/SelectionDAG/TargetLowering.cpp Thu Nov
> 6 19:28:02 2008
> @@ -1352,6 +1352,52 @@
> Zero, Cond);
> }
> }
> +
> + // If the LHS is '(and load, const)', the RHS is 0,
> + // the test is for equality or unsigned, and all 1 bits of
> the const are
> + // in the same partial word, see if we can shorten the load.
> + if (DCI.isBeforeLegalize() &&
> + N0.getOpcode() == ISD::AND && C1 == 0 &&
> + isa<LoadSDNode>(N0.getOperand(0)) &&
> + N0.getOperand(0).getNode()->hasOneUse() &&
> + isa<ConstantSDNode>(N0.getOperand(1))) {
> + LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0));
> + uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))-
> >getZExtValue();
> + unsigned bestWidth = 0, bestOffset = 0;
> + if (!Lod->isVolatile()) {
> + unsigned origWidth = N0.getValueType().getSizeInBits();
> + for (unsigned width = origWidth / 2; width>=8; width /=
> 2) {
> + uint64_t newMask = (1ULL << width) - 1;
> + for (unsigned offset=0; offset<origWidth/width; offset+
> +) {
> + if ((newMask & Mask)==Mask) {
> + bestOffset = (uint64_t)offset * (width/8);
> + bestWidth = width;
> + break;
> + }
> + newMask = newMask << width;
> + }
> + }
> + }
> + if (bestWidth) {
> + MVT newVT = MVT::getIntegerVT(bestWidth);
> + if (newVT.isRound()) {
> + uint64_t bestMask = Mask >> (bestOffset * 8);
> + MVT PtrType = Lod->getOperand(1).getValueType();
> + SDValue Ptr = Lod->getBasePtr();
> + if (bestOffset != 0)
> + Ptr = DAG.getNode(ISD::ADD, PtrType, Lod->getBasePtr(),
> + DAG.getConstant(bestOffset,
> PtrType));
> + unsigned NewAlign = MinAlign(Lod->getAlignment(),
> bestOffset);
> + SDValue NewLoad = DAG.getLoad(newVT, Lod->getChain(),
> Ptr,
> + Lod->getSrcValue(),
> + Lod->getSrcValueOffset()
> + bestOffset,
> + false, NewAlign);
> + return DAG.getSetCC(VT, DAG.getNode(ISD::AND, newVT,
> NewLoad,
> +
> DAG.getConstant(bestMask, newVT)),
> + DAG.getConstant(0LL, newVT),
> Cond);
> + }
> + }
> + }
>
> // If the LHS is a ZERO_EXTEND, perform the comparison on the
> input.
> if (N0.getOpcode() == ISD::ZERO_EXTEND) {
>
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
More information about the llvm-commits
mailing list