On Sat, Feb 9, 2019 at 12:26 PM Ard Biesheuvel <ard.biesheu...@linaro.org> wrote: > > On Sat, 9 Feb 2019 at 12:19, Miguel Ojeda > <miguel.ojeda.sando...@gmail.com> wrote: > > > > It also affects the optimizer in two different ways AFAIK: > > > > * For the function itself, it gets optimized for size instead of speed. > > * For callers, the paths that lead to the calls are treated as unlikely. > > > > That seems reasonable, but that still does not mean it is necessarily > a problem if you apply 'cold' to one but not the other.
Indeed. As I said, it is likely that you missed the attribute, not a sure thing (i.e. that you didn't do it explicitly). > > So GCC reports it because you would be (likely) missing the > > optimizations you expected if you are using the alias instead of the > > target. > > > > I see how that could be reasonable for extern declarations that do not > match the definition, since in that case, it is assumed that there is > only one instance of the function. For function pointers, I don't > think this assumption is valid. It sounds reasonable to have another warning for declarations-definition attribute mismatches too. However, I don't see why you would warn differently. Even if you have one instance of the function, you may also be using the declaration to explicitly avoid the unlikely treatment. Now, whether the warning is worth or not or at which "level", it depends. I guess the rationale behind having it under -Wall is that they expect people to have missed copying the attributes, rather than they are using aliases specifically to avoid a cold/... attribute. > > In our case in patch 3, we do not want the optimization for callers, > > which is why we don't mark the extern declaration as __cold (see the > > commit message). > > > > You did not cc me on the whole set, so I don't have the patch. But in > any case, GCC 9 has not been released so we should still have time to > talk sense into the GCC guys. I only CC'd people on the relevant patches according to get_maintainers (but yeah, 2 & 3 are related, I could have merged those lists). Anyway, using the lore.kernel.org server makes it easy to see an entire series when you have already one: https://lore.kernel.org/lkml/20190209000840.11018-1-miguel.ojeda.sando...@gmail.com/ As for GCC, Martin (the author of the features) is CC'd, so he can chime in (and I am sure he appreciates the feedback :-) Cheers, Miguel
