erichkeane wrote: > > CC @AaronBallman for an opinion about the AST represenation > > This is a bit awkward in that it's only used for the `annotate` attributes' > argument lists, which should pass all of their arguments down to LLVM IR. > However, we don't typically generate LLVM IR for discarded value statements > in other circumstances. So I'm not certain _what_ we should pass down to LLVM > IR in this case. The arguments could be used positionally, so in that case > we'd want to pass _something_ down so the arguments line up to the source. > The arguments could be used semantically, so in that case, there's nothing to > pass down because there are no semantics. > > (CC @erichkeane for more opinions)
My immediate read of this before seeing the discussion was that the positional arguments are important, so making one 'disappear' could be problematic. I think a value of 'None', plus a reasonable IR for that would be our best way forward. As far as what a 'reasonable IR' for a 'None' value would be, I don't know if we have any viable experts here we can discuss this with. The nature of 'annotate' is that we HAVE no expert on the feature (since it is intended to be used by just about anyone in the backend). I think we'd need to see a codegen test for 'none' however. So in short: 1- Make it a None ConstantExpr 2- Give that a reasonable output for its attribute in IR 3- Write a codegen test that shows it off 4- Add a release note. https://github.com/llvm/llvm-project/pull/105789 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
