nikic wrote: > Reason for using operand bundle -- The original discussion for this change is > captured here: https://reviews.llvm.org/D105909 - My takeaway from the > discussion is that the type id metadata was dropped by optimizations leading > to not so clean implementation using metadata.
To avoid dropping call metadata during optimizations you need to implement combineMetadata handling for it -- and design the metadata to be amenable to preservation during optimizations. In particular, rather than a single !type, you'd want a list of !types, such that when two indirect calls are merged (e.g. by SimplifyCFG) it is possible to preserve the callee type information. > Dropping the OB -- Yes. Dropping it will not affect correctness. If dropping it doesn't affect correctness, I don't think the use of operand bundles is appropriate. > "type" as a name -- Do you have recommendations for name? -- here's some that > comes to my mind - "type_hash", "fn_type_hash". Is the content of this OB the same as CFI `!type` metadata on function definitions? If so, I guess the use of "type" terminology is okay, though I'd at least make it the `callee_type`. https://github.com/llvm/llvm-project/pull/87573 _______________________________________________ llvm-branch-commits mailing list llvm-branch-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
