qiongsiwu wrote:

> > If there are concurrent modifications of the input headers while the 
> > modules are getting build, I think this error will be triggered.
> 
> > I think for this particular case, this diagnostic is good enough, as it 
> > spells out the exact header that is modified to debug the potential race.
> 
> Given our prior experience in this area, that has not been the case, though. 
> It's been logic bugs and saving state that went stale in a non-racy but still 
> in a way that breaks our assumptions (e.g., negative stat caching). This also 
> happens much more often when there's a pre-populated module cache. Or maybe 
> I'm misunderstanding what cases `pcm_not_dropped` was intended to capture? In 
> my opinion, when the compiler is in the position to implicitly build a module 
> (therefore, the build system & user are not in charge of creating PCMs), this 
> diagnostic is not straightforward because it should be able to rebuild the 
> module, unlike when this happens e.g. precompiled headers, which is driven by 
> the build system. We should still keep the explicit error about the 
> problematic input file, no disagreement there.

Ok! I added a note specifically for the input change case to indicate that 
there may be a compile bug. 

https://github.com/llvm/llvm-project/pull/209857
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to