Bug#914770: llvm-toolchain-7: baseline violation on i386

[Sylvestre Ledru]

Le 27/11/2018 à 23:54, Fanael Linithien a écrit :
> I fail to see how #914770 and #914838 are related in any way at all,
> the former blockers of #894840 and #632472 are much more relevant in
> my opinion, even if still not ideal.
yeah, my bad
>
> Anyway, it's still a baseline violation that affects *binary packages
> already in sid*, in my case breaking seemingly unrelated packages like
> mesa, and due to the recent change of how LLVM is built (i.e.
> bootstrapped using GCC for stage 1, then Clang for stage 2), it
> affects *all packages* built from src:llvm-toolchain-7 and thus should
> be reported against that.
>
> Furthermore, I don't think changing the baseline would be easy,
> because of Firefox: stable uses the latest ESR, which in turn depends
> on recent Rust, which in turn depends on recent LLVM, so in one or two
> Firefox ESR cycles we (Debian) may end up requiring LLVM 7 in *jessie
> and stretch*, and while changing the baseline of buster may be an
> option[1], changing it retroactively in *oldstable and stable* doesn't
> seem like a very good idea.
>
> At any rate, the fix to this particular issue seems quite easy: just
> pass -march=i686 to Clang when building stage 2 of LLVM and Clang.
>
> [1] And even then, I'd be willing to contribute to keep the baseline
> as is for as long as possible: I have the old hardware to test stuff
> on, I have the knowledge to fix stuff when needed, and I have the
> willingness to make Debian better.

Sure, don't hesitate to propose a patch. Should be similar to:

https://salsa.debian.org/pkg-llvm-team/llvm-toolchain/blob/7/debian/patches/clang-arm-default-vfp3-on-armv7a.patch

S