lenary wrote: > > Gentle Ping. I'm looking for answers to two questions: > > > > * [Should I] prepare a fixup commit that [uses `MVT::Untyped`] (and removes > > the `riscv_*_pair` MVTs), if we think that's a better target-independent > > approach? > > I guess so. I didn't know about the SystemZ change before.
I'll try it, and push a separate commit, so we can see what else is needed. > > I do find it a little weird having 2 register classes with the same spill > size and registers. I believe the tablegen generated code considers them > subclasses of each other. X86 has some register classes like this too so I > guess its not a big deal. I'm not very happy about the duplication either, but it's because of a limitation where fairly early on, selectiondag uses the first legal type in a class, rather than the one you might want. > > * Any advice on whether I should be digging deeply into changing > > SelectionDAGBuilder.cpp (affecting all targets) so we don't need the > > `getNumRegisters` override? The options here IMO are a) live with the > > override, b) change SelectionDAGBuilder now, c) change SelectionDAGBuilder > > in a follow-up and remove the override once it is not needed. > > I think we can live with the override since SystemZ is doing it too. https://github.com/llvm/llvm-project/pull/112983 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
