https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98209
--- Comment #10 from rguenther at suse dot de <rguenther at suse dot de> --- On Thu, 28 Jan 2021, jakub at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98209 > > Jakub Jelinek <jakub at gcc dot gnu.org> changed: > > What |Removed |Added > ---------------------------------------------------------------------------- > CC| |jakub at gcc dot gnu.org > > --- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> --- > But e.g. for the *intrin.h intrinsics, ignoring the ISA requirements that it > got from the target pragmas is highly undesirable, we do need the diagnostics > and refuse to inline them, otherwise they will ICE all over. > And it is unclear how to differentiate between ISA requirements that were > added > because of the command line option flags vs. ISA requirements that got added > through target pragmas or target attributes. For select cases we have sth like cfun->uses_math and chose to ignore some (target specific) settings based on things we use in the IL. We could go much further here (but of course _not_ scan the IL in the respective function - we'd have to pre-compute things somehow, like adding target specific fields in the inline summary / merging and when computing that, allow the target to amend). I think the *intrin.h example is not really relevant as their direct caller is very unlikely to differ in setting.
