dexonsmith added a comment. In D134456#3827263 <https://reviews.llvm.org/D134456#3827263>, @aaron.ballman wrote:
> In D134456#3827185 <https://reviews.llvm.org/D134456#3827185>, @dexonsmith > wrote: > >> This safety scenario sounds like it could differ within a file. Is a flag >> really the right control? Maybe what you want is a `[[noprofile]]`, similar >> to `[[noopt]]`, which selectively disables the profile where the user wants >> fine-grained control to ignore the profile. > > My understanding is that it's pretty rare for safety critical code to mix > with non-safety critical code, so a flag sounds like the correct granularity > to me in terms of the use cases I'm aware of. In that case, maybe those files should just turn off PGO entirely. There's already a command-line for that: - `-fprofile-use=...`: compiler should use the profile to improve performance (user hints lose when there's disagreement) - default: no access to a profile; user hints by default Maybe (if it doesn't exist yet) `-fno-profile-use` would be useful for easy cancellation. In safety-critical code where you don't entirely trust the profile/coverage/compiler to do the right thing, why risk having the feature enabled at all? Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D134456/new/ https://reviews.llvm.org/D134456 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
