[Mlir-commits] [clang] [flang] [llvm] [mlir] [IR] Generalize !amdgpu.ignore.denormal.mode into !atomic.ignore.denormal.mode (PR #217585)

[Nikita Popov]

nikic wrote:

> I can't help but wonder if (partly because I stumbled on the same issue) there's a way to do the metadata upgrades (we seem to keep collecting more of them) without O(transformations * IR size) passes. Perhaps step 1 is to have a map from metadata kinds to sets of instructions that are annotated with them?

We should do all the metadata upgrades in one pass over the IR, not one pass per metadata...

Also, textual IR generally should not get auto-upgraded, unless it's something very common that will introduce inconvenience when trying to check old reproducers. (This is not the case here, as the metadata will just get ignored, not produce a parse error.)

https://github.com/llvm/llvm-project/pull/217585