[llvm-dev] Issue with ThinLTO and gcc-9 [was: Re: [Release-testers] [9.0.0 Release] Release Candidate 3 is here]

[Hans Wennborg via llvm-dev]

On Mon, Sep 9, 2019 at 5:37 PM Sylvestre Ledru <sylvestre at debian.org> wrote:
>
> Hello,
>
> Le 30/08/2019 à 18:38, Hans Wennborg via Release-testers a écrit :
> > Hello everyone,
> >
> > 9.0.0-rc3 was tagged today from the release_90 branch at r370450. In
> > the Git monorepo, it's tagged as llvmorg-9.0.0-rc3.
> I experienced an issue with lto. Initially not clang related but might be.
>
> For a few months, I have been building packages using ThinLTO (
> -DLLVM_ENABLE_LTO="Thin" ) and stage2. gcc for stage1.
>
> Debian unstable and the next Ubuntu updated to gcc-9. This version is stricter
> on -flto (only accept understood values ([1]).
> It caused the build to fail on polly with "No ffs implementation found" (which is just
> cmake failing with -flto=thin) [2]
> This clearly makes sense.
> So, I tried to use -DBOOTSTRAP_LLVM_ENABLE_LTO="Thin" to have only -flto=Thin used in the stage2.
>
> However, this caused two issues:
> * Some packages became much bigger (21M => 141M)
> * Linking failed for some packages with:
> /usr/bin/ld: /usr/lib/llvm-9/lib/libclangCodeGen.a: error adding symbols: archive has no index; run ranlib to
> one [3]
>
> I tried to play with clang cmake file [4] but I haven't been able to fix it.
>
> For now, I am forcing the usage of gcc-8 but this won't scale.
>
> Anyway, probably not a blocker for the 9 release but I prefer to report this in case other are getting this issue.

Thanks for raising it.

I've never used our cmake files to bootstrap directly. Does it work if
you do the steps manually: first invoke cmake and do a regular build
with gcc 9, then invoke cmake separately for the second build, setting
the flags to use the clang and lld from the first step, with thinlto?