[DAG] Stop replace masked compressstore with normal masked store

[xiangzhangllvm]

Summary

The masked compressstore can not be replaced with normal masked store.
Because masked compressstore required adjacently store the masked elements
which normal masked store can not make sure.

[pengfei]

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

11033–11034

Do you mean cannot be a compress store?

[pengfei]

llvm/test/CodeGen/X86/masked_compressstore_isel.ll

15–17

Can we optimizate it to truncated compressstore?

[xiangzhangllvm]

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

11033–11034

Yes, right

llvm/test/CodeGen/X86/masked_compressstore_isel.ll

15–17

Do we have truncated compressstore?
May be can implement by targets.
But for target independent code we can not directly transfer to normal masted store.

[pengfei]

llvm/test/CodeGen/X86/masked_compressstore_isel.ll

15–17

I think the truncated xxx is just a code concept, it actually means turning the trunction from value to the mask.
So I think we should combine these 2 lines into VPCOMPRESSWZ256mrk ... just like mask store.

[xiangzhangllvm]

llvm/test/CodeGen/X86/masked_compressstore_isel.ll

15–17

Can we first let it in and then try do the optimization ?
I wish to close a must fix bug first. thanks!

[pengfei]

LGTM.

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

11033–11034

Don't forget update the comments.

llvm/test/CodeGen/X86/masked_compressstore_isel.ll

15–17

Sure. Better add a TODO in comments.

[xiangzhangllvm]

In D139682#3987319, @pengfei wrote:

LGTM.

Thank you! sure, let me add a todo here when commit.

[pengfei]

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

11033–11034

I meant the comment We can do this even if this is already a masked truncstore or a compress store is wrong. I'd expect it would be:

// If this is a TRUNC followed by a masked store, fold this into a masked
// truncating store.  We can do this even if this is already a masked
// truncstore.
// If this is a TRUNC followed by a compress store, we don't do the combine.
// TODO: Try combine to masked compress store if possiable.

[xiangzhangllvm]

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

11033–11034

Good catch! Done in 7557d94bd8d92e486084074e1017eb33149f9156 , thanks again!

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if (SimplifyDemandedBits(Value, TruncDemandedBits)) {
       if (N->getOpcode() != ISD::DELETED_NODE)
         AddToWorklist(N);
       return SDValue(N, 0);
     }
   }
 
   // If this is a TRUNC followed by a masked store, fold this into a masked
   // truncating store.  We can do this even if this is already a masked
-  // truncstore.
+  // truncstore or a compress store.
+  // TODO: Try combine to masked compress store if possiable.
   if ((Value.getOpcode() == ISD::TRUNCATE) && Value->hasOneUse() &&
-      MST->isUnindexed() &&
+      MST->isUnindexed() && !MST->isCompressingStore() &&
       TLI.canCombineTruncStore(Value.getOperand(0).getValueType(),
                                MST->getMemoryVT(), LegalOperations)) {
     auto Mask = TLI.promoteTargetBoolean(DAG, MST->getMask(),
                                          Value.getOperand(0).getValueType());
     return DAG.getMaskedStore(Chain, SDLoc(N), Value.getOperand(0), Ptr,
                               MST->getOffset(), Mask, MST->getMemoryVT(),
                               MST->getMemOperand(), MST->getAddressingMode(),
                               /*IsTruncating=*/true);

llvm/test/CodeGen/X86/masked_compressstore_isel.ll

 ; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -start-after=codegenprepare -stop-before finalize-isel | FileCheck %s
 
 define void @_Z3fooiPiPs(<8 x i32> %gepload, <8 x i1> %0) #0 {
 entry:
   %1 = trunc <8 x i32> %gepload to <8 x i16>
   tail call void @llvm.masked.compressstore.v8i16(<8 x i16> %1, i16* null, <8 x i1> %0)
   ret void
 }
 
 ; CHECK-LABEL: bb.0.entry:
 ; CHECK:         %1:vr128x = COPY $xmm1
 ; CHECK-NEXT:    %0:vr256x = COPY $ymm0
 ; CHECK-NEXT:    %2:vr128x = VPSLLWZ128ri %1, 15
 ; CHECK-NEXT:    %3:vk16wm = VPMOVW2MZ128rr killed %2
-; CHECK-NEXT:    VPMOVDWZ256mrk $noreg, 1, $noreg, 0, $noreg, killed %3, %0 :: (store unknown-size into `i16* null`, align 16)
+; CHECK-NEXT:    %4:vr128x = VPMOVDWZ256rr %0
+; CHECK-NEXT:    VPCOMPRESSWZ128mrk $noreg, 1, $noreg, 0, $noreg, killed %3, killed %4 :: (store unknown-size into `i16* null`, align 16)
 ; CHECK-NEXT:    RET 0
 
 ; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: write)
 declare void @llvm.masked.compressstore.v8i16(<8 x i16>, i16* nocapture, <8 x i1>) #1
 
 attributes #0 = { "target-cpu"="icelake-server" }
 attributes #1 = { nocallback nofree nosync nounwind willreturn memory(argmem: write) }