<div dir="ltr">Isn't what's stored in the PointerInfo of a MMO usually a Value * from IR? Except when its a stack reference or a constant pool or one of the other PseudoSources?<div><br clear="all"><div><div dir="ltr" class="gmail_signature" data-smartmail="gmail_signature">~Craig</div></div><br></div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Wed, Mar 18, 2020 at 5:32 PM Philip Reames <<a href="mailto:listmail@philipreames.com">listmail@philipreames.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
<div>
<p>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}, ....</p>
<p>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.</p>
<p>Philip<br>
</p>
<div>On 3/18/20 4:28 PM, Craig Topper wrote:<br>
</div>
<blockquote type="cite">
<div dir="ltr">
<div>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?</div>
<div><br>
</div>
<div>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.</div>
<br clear="all">
<div>
<div dir="ltr">~Craig</div>
</div>
<br>
</div>
<br>
<div class="gmail_quote">
<div dir="ltr" class="gmail_attr">On Wed, Mar 18, 2020 at 4:16
PM Philip Reames via llvm-commits <<a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a>>
wrote:<br>
</div>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">Just
curious, we have the ability to describe multiple memory
locations <br>
for a single instruction. Why not simply track each lane
independently <br>
and keep the right locations on each half?<br>
<br>
Philip<br>
<br>
On 3/18/20 4:12 PM, Craig Topper via llvm-commits wrote:<br>
> Author: Craig Topper<br>
> Date: 2020-03-18T16:07:15-07:00<br>
> New Revision: c69a4d6bef0a1350f509f35beb450dccc2a6c5e2<br>
><br>
> URL: <a href="https://github.com/llvm/llvm-project/commit/c69a4d6bef0a1350f509f35beb450dccc2a6c5e2" rel="noreferrer" target="_blank">https://github.com/llvm/llvm-project/commit/c69a4d6bef0a1350f509f35beb450dccc2a6c5e2</a><br>
> DIFF: <a href="https://github.com/llvm/llvm-project/commit/c69a4d6bef0a1350f509f35beb450dccc2a6c5e2.diff" rel="noreferrer" target="_blank">https://github.com/llvm/llvm-project/commit/c69a4d6bef0a1350f509f35beb450dccc2a6c5e2.diff</a><br>
><br>
> LOG: [SelectionDAG] When splitting gathers/scatters in
type legalization, set MMO size to UnknownSize<br>
><br>
> 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.<br>
><br>
> 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?<br>
><br>
> Differential Revision: <a href="https://reviews.llvm.org/D76388" rel="noreferrer" target="_blank">https://reviews.llvm.org/D76388</a><br>
><br>
> Added:<br>
> <br>
><br>
> Modified:<br>
>
llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp<br>
><br>
> Removed:<br>
> <br>
><br>
><br>
>
################################################################################<br>
> diff --git
a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp<br>
> index 5a4b4c615bc0..09934bbb29fe 100644<br>
> ---
a/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp<br>
> +++
b/llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp<br>
> @@ -20,10 +20,11 @@<br>
>
//===----------------------------------------------------------------------===//<br>
> <br>
> #include "LegalizeTypes.h"<br>
> +#include "llvm/Analysis/MemoryLocation.h"<br>
> #include "llvm/IR/DataLayout.h"<br>
> #include "llvm/Support/ErrorHandling.h"<br>
> -#include "llvm/Support/raw_ostream.h"<br>
> #include "llvm/Support/TypeSize.h"<br>
> +#include "llvm/Support/raw_ostream.h"<br>
> using namespace llvm;<br>
> <br>
> #define DEBUG_TYPE "legalize-types"<br>
> @@ -1627,11 +1628,6 @@ void
DAGTypeLegalizer::SplitVecRes_MGATHER(MaskedGatherSDNode *MGT,<br>
> std::tie(MaskLo, MaskHi) = DAG.SplitVector(Mask,
dl);<br>
> }<br>
> <br>
> - EVT MemoryVT = MGT->getMemoryVT();<br>
> - EVT LoMemVT, HiMemVT;<br>
> - // Split MemoryVT<br>
> - std::tie(LoMemVT, HiMemVT) =
DAG.GetSplitDestVTs(MemoryVT);<br>
> -<br>
> SDValue PassThruLo, PassThruHi;<br>
> if (getTypeAction(PassThru.getValueType()) ==
TargetLowering::TypeSplitVector)<br>
> GetSplitVector(PassThru, PassThruLo, PassThruHi);<br>
> @@ -1644,10 +1640,10 @@ void
DAGTypeLegalizer::SplitVecRes_MGATHER(MaskedGatherSDNode *MGT,<br>
> else<br>
> std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index,
dl);<br>
> <br>
> - MachineMemOperand *MMO = DAG.getMachineFunction().<br>
> - getMachineMemOperand(MGT->getPointerInfo(),<br>
> - MachineMemOperand::MOLoad,
LoMemVT.getStoreSize(),<br>
> - Alignment, MGT->getAAInfo(),
MGT->getRanges());<br>
> + MachineMemOperand *MMO =
DAG.getMachineFunction().getMachineMemOperand(<br>
> + MGT->getPointerInfo(),
MachineMemOperand::MOLoad,<br>
> + MemoryLocation::UnknownSize, Alignment,
MGT->getAAInfo(),<br>
> + MGT->getRanges());<br>
> <br>
> SDValue OpsLo[] = {Ch, PassThruLo, MaskLo, Ptr,
IndexLo, Scale};<br>
> Lo = DAG.getMaskedGather(DAG.getVTList(LoVT,
MVT::Other), LoVT, dl, OpsLo,<br>
> @@ -2376,13 +2372,10 @@ SDValue
DAGTypeLegalizer::SplitVecOp_MSCATTER(MaskedScatterSDNode *N,<br>
> SDValue Index = N->getIndex();<br>
> SDValue Scale = N->getScale();<br>
> SDValue Data = N->getValue();<br>
> - EVT MemoryVT = N->getMemoryVT();<br>
> unsigned Alignment = N->getOriginalAlignment();<br>
> SDLoc DL(N);<br>
> <br>
> // Split all operands<br>
> - EVT LoMemVT, HiMemVT;<br>
> - std::tie(LoMemVT, HiMemVT) =
DAG.GetSplitDestVTs(MemoryVT);<br>
> <br>
> SDValue DataLo, DataHi;<br>
> if (getTypeAction(Data.getValueType()) ==
TargetLowering::TypeSplitVector)<br>
> @@ -2409,20 +2402,14 @@ SDValue
DAGTypeLegalizer::SplitVecOp_MSCATTER(MaskedScatterSDNode *N,<br>
> std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index,
DL);<br>
> <br>
> SDValue Lo;<br>
> - MachineMemOperand *MMO = DAG.getMachineFunction().<br>
> - getMachineMemOperand(N->getPointerInfo(),<br>
> - MachineMemOperand::MOStore,
LoMemVT.getStoreSize(),<br>
> - Alignment, N->getAAInfo(),
N->getRanges());<br>
> + MachineMemOperand *MMO =
DAG.getMachineFunction().getMachineMemOperand(<br>
> + N->getPointerInfo(),
MachineMemOperand::MOStore,<br>
> + MemoryLocation::UnknownSize, Alignment,
N->getAAInfo(), N->getRanges());<br>
> <br>
> SDValue OpsLo[] = {Ch, DataLo, MaskLo, Ptr, IndexLo,
Scale};<br>
> Lo = DAG.getMaskedScatter(DAG.getVTList(MVT::Other),
DataLo.getValueType(),<br>
> DL, OpsLo, MMO,
N->getIndexType());<br>
> <br>
> - MMO = DAG.getMachineFunction().<br>
> - getMachineMemOperand(N->getPointerInfo(),<br>
> - MachineMemOperand::MOStore,
HiMemVT.getStoreSize(),<br>
> - Alignment, N->getAAInfo(),
N->getRanges());<br>
> -<br>
> // The order of the Scatter operation after split is
well defined. The "Hi"<br>
> // part comes after the "Lo". So these two operations
should be chained one<br>
> // after another.<br>
><br>
><br>
> <br>
> _______________________________________________<br>
> llvm-commits mailing list<br>
> <a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a><br>
> <a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits" rel="noreferrer" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits</a><br>
_______________________________________________<br>
llvm-commits mailing list<br>
<a href="mailto:llvm-commits@lists.llvm.org" target="_blank">llvm-commits@lists.llvm.org</a><br>
<a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits" rel="noreferrer" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits</a><br>
</blockquote>
</div>
</blockquote>
</div>
</blockquote></div>