[PATCH] D69778: Make -fmodules-codegen and -fmodules-debuginfo work also with precompiled headers

[David Blaikie via Phabricator via cfe-commits]

dblaikie added a comment.

In D69778#1928417 <https://reviews.llvm.org/D69778#1928417>, @llunak wrote:

> In D69778#1928111 <https://reviews.llvm.org/D69778#1928111>, @dblaikie wrote:
>
> > In D69778#1927761 <https://reviews.llvm.org/D69778#1927761>, @llunak wrote:
> >
> > > It comes from 08c5a7b8fda1b97df9d027d1ac096f93edeb6c2f . AFAICT -fmodules-codegen is a superset of -building-pch-with-obj, the important difference is that -building-pch-with-obj decides which code will be shared while building TUs, while -fmodules-codegen decides while building the PCH.
> >
> >
> > I'm not sure I follow - how would that be possible? building the object from the PCH is done independently of any usage, so it can't depend on the way the PCH is used by any TU that includes it, I think?
>
>
> If a TU uses a PCH, it'll result in some code emitted in every TU that uses the same PCH (e.g. template instantiations from the PCH). And -building-pch-with-obj detects these and discards them in all TUs except the PCH's object file. That means that the code shared in the PCH's object file is only code that would be duplicated in all TUs using the PCH.


I'm not sure how that could be possible - there's no data transferred between the compilation of the TU's object files and the PCH's object file, right?

Looks to me like the original patch claims the PCH object file contains all the dllexported inline functions from the header. Which, yes, is different from -fmodules-codegen, but doesn't sound like it's based on usage & I'm not sure how it could be based on usage.

> This is fundamentally different from -fmodules-codegen, which decides what will be emitted and shared already when creating the PCH, which has the consequence that it selects also a lot of code that will be eventually unused, which causes problems like the -Wl,--gc-sections part or not being necessarily worth it (IIRC even your -fmodules-codegen commit says so). So while -fmodules-codegen should be generally superior, -building-pch-with-obj is trivial to use. That's my understanding of both the approaches (and I spent quite a lot of time looking at both when working on D69778 <https://reviews.llvm.org/D69778> and D69585 <https://reviews.llvm.org/D69585>).




Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D69778/new/

https://reviews.llvm.org/D69778