Issue 241480
Summary [LoopVectorize] `different types inferred for different operands` in truncateToMinimalBitwidths
Labels new issue
Assignees
Reporter arie2026
    In IR the two `or`s share no operand, so MinBWs gives them different widths. VPlan then folds `sext i8 1 to i32` into
the live-in `ir<1>`, which the first `or` also uses. `truncateToMinimalBitwidths` caches truncates per operand only
(`ProcessedTruncs`), so the second `or` reuses a trunc of `ir<1>` made for the other width, and cloning the recipe asserts.

Reproducer (`opt -mtriple=x86_64-unknown-linux-gnu -mattr=+avx2 -passes=loop-vectorize`; also reproduces on 5e457cda8d5a and f2808fc79):
```llvm
define void @f(i32 %a, ptr %dst) {
entry:
  br label %loop

loop:
  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
  %o1 = or i32 %a, 1
  %c0 = sext i8 1 to i32
  %cmp = icmp sle i32 %o1, 0
  %z = zext i1 %cmp to i32
  %o2 = or i32 %c0, %z
  %t = trunc i32 %o2 to i16
 %found = icmp eq i16 %t, 0
  %gep = getelementptr i1, ptr %dst, i64 %iv
 store i1 %found, ptr %gep
  %iv.next = add i64 %iv, 1
  %ec = icmp eq i64 %iv.next, 1024
  br i1 %ec, label %exit, label %loop

exit:
  ret void
}
```

```
$ opt -mtriple=x86_64-unknown-linux-gnu -mattr=+avx2 -passes=loop-vectorize -disable-output test.ll
opt: llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp:498: auto llvm::computeScalarTypeForInstruction(unsigned int, ArrayRef<VPValue *>)::(anonymous class)::operator()(unsigned int, Type *) const: Assertion `(!OpTy || OpTy == ExpectedTy) && "different types inferred for different operands"' failed.
#11 llvm::computeScalarTypeForInstruction(...)
#12 llvm::VPWidenRecipe::VPWidenRecipe(...)
#13 llvm::VPWidenRecipe::cloneWithOperands(...)
#14 llvm::VPlanTransforms::truncateToMinimalBitwidths(...)
#15 llvm::VPlanTransforms::runPass<...>(...)
#16 llvm::LoopVectorizationPlanner::buildVPlans(...)
#17 llvm::LoopVectorizationPlanner::plan(...)
#18 llvm::LoopVectorizePass::processLoop(...)
#19 llvm::LoopVectorizePass::runImpl(...)
```

Expected: no crash. Each narrowed recipe gets operands truncated to its own width.

Notes: if the second `or` uses a literal `1` instead of `%c0`, the sharing exists in IR, MinBWs merges the two `or`s,
and there is no crash. So the problem is sharing that VPlan creates after MinBWs was computed. Possible fix: key
`ProcessedTruncs` by `(Op, NewResTy)`, or don't reuse a cached trunc whose type differs.

Found with a mutation-based IR fuzzer. Triage, test-case minimization and parts of this write-up were
assisted by an LLM; verified the reproducer on 9dce18f0e.

Assisted-by: Claude (Anthropic)
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs

Reply via email to