[llvm] c69a4d6 - [SelectionDAG] When splitting gathers/scatters in type legalization, set MMO size to UnknownSize
Philip Reames via llvm-commits
llvm-commits at lists.llvm.org
Wed Mar 18 17:32:44 PDT 2020
I wasn't talking about the IR representation at all. At MC, we can
provide multiple MMO on a single instruction. Given the semantics of a
gather, we could encode each implied memory access separate with it's
own {base, offset} information. So a single gather instruction would
have MMOs: {base, offset1}, {base, offset2}, ....
I hadn't considered the SDAG representation. We handle that for
STATEPOINT (one user of multiple MMOs I happen to be familiar with) by
placing them directly on a target specific STATEPOINT node. So, yes,
we'd probably want to either allow MemSDNode to have multiple MMOs or
split the usage.
Philip
On 3/18/20 4:28 PM, Craig Topper wrote:
> There's probably something we can do there. We'd need a scalar pointer
> in IR to refer to though right? We likely have a vector GEP and none
> of the underlying APIs know what to do with that. Or we need to change
> those APIs to refer to a specific element of the vector?
>
> We'd also need a new base class for the gather and scatter nodes in
> SelectionDAG other than MemSDNode or we'd need to change MemSDNode to
> have more than just a single MemOperand pointer.
>
> ~Craig
>
>
> On Wed, Mar 18, 2020 at 4:16 PM Philip Reames via llvm-commits
> <llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>> wrote:
>
> Just curious, we have the ability to describe multiple memory
> locations
> for a single instruction. Why not simply track each lane
> independently
> and keep the right locations on each half?
>
> Philip
>
> On 3/18/20 4:12 PM, Craig Topper via llvm-commits wrote:
> > Author: Craig Topper
> > Date: 2020-03-18T16:07:15-07:00
> > New Revision: c69a4d6bef0a1350f509f35beb450dccc2a6c5e2
> >
> > URL:
> https://github.com/llvm/llvm-project/commit/c69a4d6bef0a1350f509f35beb450dccc2a6c5e2
> > DIFF:
> https://github.com/llvm/llvm-project/commit/c69a4d6bef0a1350f509f35beb450dccc2a6c5e2.diff
> >
> > LOG: [SelectionDAG] When splitting gathers/scatters in type
> legalization, set MMO size to UnknownSize
> >
> > Gather/scatter don't access one memory location, they access
> multiple disjoint locations. So using a fixed size isn't accurate.
> But we don't have a way to represent the true behavior so just use
> UnknownSize.
> >
> > Previously we "split" the memory VT and use that size for the
> MMO of each half. But the memory VT is scalar so splitting usually
> just returned the original scalar VT, but on 32-bit X86 if the
> scalar VT was i64 it probably returned i32?
> >
> > Differential Revision: https://reviews.llvm.org/D76388
> >
> > Added:
> >
> >
> > Modified:
> > llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
> >
> > Removed:
> >
> >
> >
> >
> ################################################################################
> > diff --git
> a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
> b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
> > index 5a4b4c615bc0..09934bbb29fe 100644
> > --- a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
> > +++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
> > @@ -20,10 +20,11 @@
> >
> //===----------------------------------------------------------------------===//
> >
> > #include "LegalizeTypes.h"
> > +#include "llvm/Analysis/MemoryLocation.h"
> > #include "llvm/IR/DataLayout.h"
> > #include "llvm/Support/ErrorHandling.h"
> > -#include "llvm/Support/raw_ostream.h"
> > #include "llvm/Support/TypeSize.h"
> > +#include "llvm/Support/raw_ostream.h"
> > using namespace llvm;
> >
> > #define DEBUG_TYPE "legalize-types"
> > @@ -1627,11 +1628,6 @@ void
> DAGTypeLegalizer::SplitVecRes_MGATHER(MaskedGatherSDNode *MGT,
> > std::tie(MaskLo, MaskHi) = DAG.SplitVector(Mask, dl);
> > }
> >
> > - EVT MemoryVT = MGT->getMemoryVT();
> > - EVT LoMemVT, HiMemVT;
> > - // Split MemoryVT
> > - std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
> > -
> > SDValue PassThruLo, PassThruHi;
> > if (getTypeAction(PassThru.getValueType()) ==
> TargetLowering::TypeSplitVector)
> > GetSplitVector(PassThru, PassThruLo, PassThruHi);
> > @@ -1644,10 +1640,10 @@ void
> DAGTypeLegalizer::SplitVecRes_MGATHER(MaskedGatherSDNode *MGT,
> > else
> > std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index, dl);
> >
> > - MachineMemOperand *MMO = DAG.getMachineFunction().
> > - getMachineMemOperand(MGT->getPointerInfo(),
> > - MachineMemOperand::MOLoad,
> LoMemVT.getStoreSize(),
> > - Alignment, MGT->getAAInfo(),
> MGT->getRanges());
> > + MachineMemOperand *MMO =
> DAG.getMachineFunction().getMachineMemOperand(
> > + MGT->getPointerInfo(), MachineMemOperand::MOLoad,
> > + MemoryLocation::UnknownSize, Alignment, MGT->getAAInfo(),
> > + MGT->getRanges());
> >
> > SDValue OpsLo[] = {Ch, PassThruLo, MaskLo, Ptr, IndexLo, Scale};
> > Lo = DAG.getMaskedGather(DAG.getVTList(LoVT, MVT::Other),
> LoVT, dl, OpsLo,
> > @@ -2376,13 +2372,10 @@ SDValue
> DAGTypeLegalizer::SplitVecOp_MSCATTER(MaskedScatterSDNode *N,
> > SDValue Index = N->getIndex();
> > SDValue Scale = N->getScale();
> > SDValue Data = N->getValue();
> > - EVT MemoryVT = N->getMemoryVT();
> > unsigned Alignment = N->getOriginalAlignment();
> > SDLoc DL(N);
> >
> > // Split all operands
> > - EVT LoMemVT, HiMemVT;
> > - std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
> >
> > SDValue DataLo, DataHi;
> > if (getTypeAction(Data.getValueType()) ==
> TargetLowering::TypeSplitVector)
> > @@ -2409,20 +2402,14 @@ SDValue
> DAGTypeLegalizer::SplitVecOp_MSCATTER(MaskedScatterSDNode *N,
> > std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index, DL);
> >
> > SDValue Lo;
> > - MachineMemOperand *MMO = DAG.getMachineFunction().
> > - getMachineMemOperand(N->getPointerInfo(),
> > - MachineMemOperand::MOStore,
> LoMemVT.getStoreSize(),
> > - Alignment, N->getAAInfo(),
> N->getRanges());
> > + MachineMemOperand *MMO =
> DAG.getMachineFunction().getMachineMemOperand(
> > + N->getPointerInfo(), MachineMemOperand::MOStore,
> > + MemoryLocation::UnknownSize, Alignment, N->getAAInfo(),
> N->getRanges());
> >
> > SDValue OpsLo[] = {Ch, DataLo, MaskLo, Ptr, IndexLo, Scale};
> > Lo = DAG.getMaskedScatter(DAG.getVTList(MVT::Other),
> DataLo.getValueType(),
> > DL, OpsLo, MMO, N->getIndexType());
> >
> > - MMO = DAG.getMachineFunction().
> > - getMachineMemOperand(N->getPointerInfo(),
> > - MachineMemOperand::MOStore,
> HiMemVT.getStoreSize(),
> > - Alignment, N->getAAInfo(),
> N->getRanges());
> > -
> > // The order of the Scatter operation after split is well
> defined. The "Hi"
> > // part comes after the "Lo". So these two operations should
> be chained one
> > // after another.
> >
> >
> >
> > _______________________________________________
> > llvm-commits mailing list
> > llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>
> > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200318/f9004ecb/attachment.html>
More information about the llvm-commits
mailing list