phoebewang wrote: > For the future, I wonder if we can avoid breaking changes like this and > instead make the deprecated flags and targets no-ops. It is not just a matter > of passing -mavx10.1-512 flags or not. > > For dynamic dispatch, we have had to specify "evex512" attributes in code, > and that has now transitioned from required to build-breaking. Because we are > not only dealing with discrete clang releases, such behavior changes are > quite painful. > > It would help to add some way to detect from code whether the breaking change > has occurred, for example a builtin `#define`.
SW has to match with HW spec, so it's unavoidable sometimes when HW changed. We have tried to reduce the disturbance. So "evex512" turns to a no-op. "avx10.1-512" and "avx10.1-512" are aliases of "avx10.1" and "avx10.2". However, "no-evex512", "avx10.1-256" and "avx10.2-256" have no substitutable solutions because we don't have HW support. https://github.com/llvm/llvm-project/pull/157034 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
