[SelectionDAG] Use correct result type in visitEXTRACT_VECTOR_ELT

[dlee992]

In visitEXTRACT_VECTOR_ELT, we should use ScalarVT instead of LVT as the result VT for return and method call.

If the DAG has these two nodes (this is generated from llvm-project/llvm/test/CodeGen/Generic/extractelement-shuffle.ll):

        t22: v4i32,ch = load<(load (s128) from %stack.0)> t43, FrameIndex:i32<0>, undef:i32
      t24: i64 = extract_vector_elt t22, Constant:i32<3>

It could apply this combiner rule (i.e., (vextract (v4i32 load $addr), c) -> (i32 load $addr+c*size)) then hit a failure:

Combining: t24: i64 = extract_vector_elt t22, Constant:i32<3>
Creating constant: t47: i32 = Constant<63>
Creating constant: t48: i32 = Constant<4>
Creating constant: t49: i32 = Constant<12>
Creating new node: t50: i32 = add FrameIndex:i32<0>, Constant:i32<12>
Creating new node: t51: i32,ch = load<(load (s32) from %stack.0 + 12)> t43, t50, undef:i32
 ... into: t51: i32,ch = load<(load (s32) from %stack.0 + 12)> t43, t50, undef:i32
llc: ${HOME}/llvm-project/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp:1830: void (anonymous namespace)::DAGCombiner::Run(CombineLevel): Assertion `N->getValueType(0) == RV.getValueType() && N->getNumValues() == 1 && "Type mismatch"' failed.

This PR will fix this issue.

[github-actions]

Thank you for submitting a Pull Request (PR) to the LLVM Project!

This PR will be automatically labeled and the relevant teams will be notified.

If you wish to, you can add reviewers by using the "Reviewers" section on this page.

If this is not working for you, it is probably because you do not have write permissions for the repository. In which case you can instead tag reviewers by name in a comment by using @ followed by their GitHub username.

If you have received no comments on your PR for a week, you can request a review by "ping"ing the PR by adding a comment “Ping”. The common courtesy "ping" rate is once a week. Please remember that you are asking for valuable time from other developers.

If you have further questions, they may be answered by the LLVM GitHub User Guide.

You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums.

[llvmbot]

@llvm/pr-subscribers-llvm-selectiondag

Author: Da Li (李达) (dlee992)

Changes

In visitEXTRACT_VECTOR_ELT, we should use ScalarVT instead of LVT as the result VT for return and method call.

If the DAG has these two nodes (this is generated from llvm-project/llvm/test/CodeGen/Generic/extractelement-shuffle.ll with a private backend implementation):

        t22: v64i32,ch = load<(load (s2048) from %stack.0, align 64)> t43, FrameIndex:i64<0>, undef:i64
      t24: i64 = extract_vector_elt t22, Constant:i64<3>

It could apply this combiner rule (i.e., (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)) then hit a failure:

Combining: t24: i64 = extract_vector_elt t22, Constant:i64<3>
Creating constant: t47: i64 = Constant<63>
Creating constant: t48: i64 = Constant<4>
Creating constant: t49: i64 = Constant<12>
Creating new node: t50: i64 = add FrameIndex:i64<0>, Constant:i64<12>
Creating new node: t51: i32,ch = load<(load (s32) from %stack.0 + 12)> t43, t50, undef:i64
 ... into: t51: i32,ch = load<(load (s32) from %stack.0 + 12)> t43, t50, undef:i64
llc: ${HOME}/llvm-project/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp:1830: void (anonymous namespace)::DAGCombiner::Run(CombineLevel): Assertion `N->getValueType(0) == RV.getValueType() && N->getNumValues() == 1 && "Type mismatch"' failed.

This PR will fix this issue.

---

Full diff: https://github.com/llvm/llvm-project/pull/148707.diff

1 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+3-3)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 8136f1794775e..f26ee0ace6919 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -23440,10 +23440,10 @@ SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
 
   // If Idx was -1 above, Elt is going to be -1, so just return undef.
   if (Elt == -1)
-    return DAG.getUNDEF(LVT);
+    return DAG.getUNDEF(ScalarVT);
 
-  if (SDValue Scalarized =
-          TLI.scalarizeExtractedVectorLoad(LVT, DL, VecVT, Index, LN0, DAG)) {
+  if (SDValue Scalarized = TLI.scalarizeExtractedVectorLoad(ScalarVT, DL, VecVT,
+                                                            Index, LN0, DAG)) {
     ++OpsNarrowed;
     return Scalarized;
   }

[topperc]

It could apply this combiner rule (i.e., (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)) then hit a failure:

Was that supposed to be i32 instead of f32?

[topperc]

Do you have a test case?

[dlee992]

Do you have a test case?

I don't have one right now. I need to make up one with public targets, but I'm unfamiliar with public targets yet. (or anyone can help? Thanks!)

This PR is extremely similar with this regression fix, so also cc @arsenm .

[arsenm]

LGTM but should be tested. I think ARM makes the most use of the illegal-scalar-type vectors