wangpc-pp wrote: > I think what @wangpc-pp advocated here (please correct me if I'm wrong) was > that user should be responsible annotating these registers as clobbered so > that we have more freedom on scheduling when the inline assembly is not using > any vector instructions. While other approaches -- regardless of marking them > as clobber or implicit-def these two registers in RISCVInsertVSETVLI -- are > taking a more conservative path to ensure correctness.
Yes this is exactly what I meant! > > It's a shame that we're not able to analyze the instructions within inline > assembly to make a better decision. Personally I think we should prioritize > correctness, therefore I'm more incline to the latter approaches. I do agree > that we probably should only do it when vector instructions are present, as > pointed out by one of the review comments by @wangpc-pp. I am not opposed to this. But I have to say, the use case is not so common. We don't mix assemblies with RVV instructions and RVV intrinsics in the same code. In this scenarios, we should write all RVV instructions in asm statements or just use RVV intrinsics. Even for these rare cases, the user can achieve the same effect if they put `vl/vtype` into clobbered registers explicitly. And let me quote @kito-cheng's comments: https://github.com/llvm/llvm-project/pull/97794#issuecomment-2228482748, there are some other approaches. https://github.com/llvm/llvm-project/pull/128636 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
