0xzre wrote: > This can't just break existing uses. Needs to implement bitcode autoupgrade, > and some compatibility on the builtin signature. > > Also, can we just infer clamping as an optimization fold based on the use > context?
Good catch. Will add a bitcode autupgrade for the old 4 op WMMA IU8 intrinsic to append clamp=0, plus a lit bitcode test. I'll keep source compatibility by keeping the existing builtin spelling mapped to clamp=0 and adding the 5 op form. Clamp inference can be a follow-up combine I think... this patch stays on correctness/compat. Would that be okay @arsenm ? https://github.com/llvm/llvm-project/pull/171069 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
