phoebewang wrote: > Can you give me a better idea of the stages you intend to follow with this. > The patch title suggests removing all KNL/KNM handling but the patch itself > looks to be just about the KNL/KNM specific features. > > Removing the (incomplete) KNL/KNM specific features (ER/PF/etc) I don't think > will be any major cause for concern, ideally we'd keep support at the > assembly level only (which has been suggested for MMX as well at some point). > But even that might not be necessary.
It's a good idea to preserve the assembly level support. I think GCC's proposal just wants to deprecate GCC support rather than remove support from binutils too. > But what do you expect to happen if a KNL workstation attempts to use clang > 19 with `-march=native`? Do we still enable it as a AVX512F+AVX512CD target > or are you suggesting we degrade it to match x86-64-v3 arch spec? There're discussions in GCC mailing loop about `-mtune/arch=knl`. IIUC, the final decision is to remove the march support too. For `-march=native`, I'm not sure whether we detect it by CPU model or feature list. For the former, we didn't remove the KNL/KNM model, we can emit error for it; for the latter, we can return an approaching target has the same (remaining) features. > How can the test cases then can be simplified? Are you suggesting that we > merge more of the AVX512 attributes in X86.td (e.g. AVX512VL always enables > BW/DQ / VBMI2 enables VBMI etc?). I think we can't make AVX512VL being implied by default, but can gradually relax a lot of `Subtarget.hasVLX()` in CPP and TD files. We don't need to test the combinations that have AVX512 features but no AVX512VL and remove them from existing tests gradually. >From the developmental trend of HW design, I don't think we will fall back to >support 512-bit only features in the future. So from my point of view, >simplifying backend design and testing would be the major benefit we can >achieve. https://github.com/llvm/llvm-project/pull/75580 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
