@William-Mou @msimberg: does this fix the nvhpc issue for you? This should be more general than #36961 as it does not disable -Werror for all compilers all the time.

@tgamblin I can confirm this works for NVHPC, thanks! For what it's worth, it seems like NVHPC actually learned the -Wno-error flag in 23.5 (https://godbolt.org/z/zdrveGxfo), so perhaps the check can be

if self.spec.satisfies("%nvhpc@:23.3") or self.spec.satisfies("%pgi"):

?

Hi @tgamblin and @adamjstewart

Thank you very much, I think this is very helpful! As @msimberg mentioned, nvhpc 23.5 solved this issue. Maybe could change the condition to support it.

It looks like it learned -Wno-error, but not the specific form, e.g. -Wno-error=implicit-function-declaration. Lemme modify the condition for that.

Ok I updated this to handle what I think are all the ways nvhpc works. Specifically, it will disable general -Werror args but not specific -Werror=... args for newer nvhpc's.

I wonder if we should instead name this according to "when to replace", not "when not to replace".

We had a big debate about this when we did it. By default we filter out -Werror, and nobody knows about it. The simple way to say "don't do that" is "keep -Werror" instead of filtering it out. That's why the option is keep_werror -- it's designed to let specific packages opt out easily with keep_werror = True.

I don't think it's worth inverting the sense of this, especially now that it's been in Spack a while -- while it might make the logic for this one function simpler, it will make packages more complicated.