[AMDGPU] expand-fp: always report modifications #163153
Conversation
The last change to the pass lost the assignment to the "Modified" variable for one of the pass optimizations. Add it back.
|
This fixes the test failure in |
| continue; | ||
|
|
||
| addToWorklist(I, Worklist); | ||
| Modified = true; |
There was a problem hiding this comment.
How about removing all the assignments on add to worklist, and do it once when processing the worklist
There was a problem hiding this comment.
This is possible provided scalarize always pushes to the Worklist, i.e. the instructions created in this function cannot be folded. This seems to hold true for the instructions handled in this pass. I have implemented the change accordingly in this commit. Does that make sense or should I perhaps pass NoFolder to the IRBuilder ?
There was a problem hiding this comment.
I'll revert the additional changes for now to fix the build bot failures. We can still add them in a follow-up PR.
This reverts commit 0ea427f.
The last change to the pass in PR llvm#158588 lost the assignment to the "Modified" variable for one of the pass optimizations. Add it back. This fixes the test failure in `CodeGen/AMDGPU/itofp.i128.bf.ll` (in a `LLVM_ENABLE_EXPENSIVE_CHECKS=ON` build).
4 participants
The last change to the pass in PR #158588 lost the assignment to the "Modified" variable for one of the pass optimizations.
Add it back. This fixes the test failure in
CodeGen/AMDGPU/itofp.i128.bf.ll(in aLLVM_ENABLE_EXPENSIVE_CHECKS=ONbuild).