[SelectionDAG] Bail out of mergeTruncStores when not optimizing

[aeubanks]

Summary

With unoptimized code, we may see lots of stores and spend too much time in mergeTruncStores.

Fixes PR51827.

[aeubanks]

not really sure how to test this, pointers appreciated

[spatel]

I see similar checks sprinkled around DAGCombiner, so I guess this is ok too. Not sure if it is possible to do this at a higher level and remove those checks though.

But shouldn't this change the codegen for a reduced example of https://llvm.org/PR51827 ? Or is there some other optnone check that hides this one?

[aeubanks]

In D111596#3062068, @spatel wrote:

I see similar checks sprinkled around DAGCombiner, so I guess this is ok too. Not sure if it is possible to do this at a higher level and remove those checks though.

But shouldn't this change the codegen for a reduced example of https://llvm.org/PR51827 ? Or is there some other optnone check that hides this one?

This patch doesn't affect the codegen of the attached IR, but speeds it up a lot. It's probably bailing somewhere in mergeTruncStores() for most inputs.

b.ll11 MBDownload

[spatel]

LGTM - it would be nicer if we had a higher-level bailout for optnone / -O0 instead of individual checks scattered around the combiner, but that might be (practically) impossible?

[aeubanks]

A question I had while looking into this, why is this (and other store optimizations) being done in isel instead on the IR optimization side?

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

 /// =>
 ///  *((i32)p) = BSWAP(val);
 SDValue DAGCombiner::mergeTruncStores(StoreSDNode *N) {
   // The matching looks for "store (trunc x)" patterns that appear early but are
   // likely to be replaced by truncating store nodes during combining.
   // TODO: If there is evidence that running this later would help, this
   //       limitation could be removed. Legality checks may need to be added
   //       for the created store and optional bswap/rotate.
-  if (LegalOperations)
+  if (LegalOperations || OptLevel == CodeGenOpt::None)
     return SDValue();
 
   // We only handle merging simple stores of 1-4 bytes.
   // TODO: Allow unordered atomics when wider type is legal (see D66309)
   EVT MemVT = N->getMemoryVT();
   if (!(MemVT == MVT::i8 || MemVT == MVT::i16 || MemVT == MVT::i32) ||
       !N->isSimple() || N->isIndexed())
     return SDValue();