s-barannikov wrote: > That might work, but that also is more limiting.
I think it is a good thing. All target-independent pseudo instructions should have the same semantics across all targets, otherwise they become target-dependent. If a target can override a standard pseudo instruction in any way it wishes, then there is nothing in common between these instructions across targets except for the opcode. This makes them unsuitable for use in generic code. The instructions from the list have few uses in generic code. It looks like the point in introducing them was to guarantee they have certain properties. If a target can override the properties or get them wrong accidentally, it creates opportunities for bugs to creep in. > In particular that doesn't allow you to set implicit uses / defs like you > would want for adjcallstack pseudos Those are no different from other target-specific instructions IMO. They just happen to have similar names on most targets. To make the argument more obvious/convincing, consider CALL instruction, which goes in the same row. All targets have call instructions, but does that fact alone makes CALL plausible for promotion? I don't think so. Call instructions are too different across targets and most targets have multiple variants of them. https://github.com/llvm/llvm-project/pull/159880 _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
