On 9/9/24 17:23, Andres Salomon wrote:
On 9/9/24 17:16, Sylvestre Ledru wrote:Le lundi 9 septembre 2024 à 23:12, Andres Salomon <dilin...@queued.net> a écrit :> >> We're currently using clang-16 in sid because it's also in bookworm, and> it's easier to use the same compiler for both distributions. > > It's likely* we'll need to backport clang-18 or 19 to stable - can you > give me a timeline for when you're planning to remove clang-16 from > stable?We aren't planning to remove it from stable. I guess you meant testing/sid, right ?Yes, sorry, I meant sid. :)> If it's in the near-term, then we'll probably start testing out > chromium in sid with clang-18 and backporting it to stable. If it's a > few months out, then (with the assumption that clang 19.1.0 final will > be released any day now and hopefully transition to trixie relatively > quickly,) we'll focus on clang-19. I think you should start soon if you can :)Okay thanks, I will do that!
*Sigh*. I was hoping to wait ~6 months before backporting clang-19, but it's looking like we might have to move that timeline up and go with clang-18 in bookworm.
Tim's been hitting some compiler bugs:"Just because I happened to notice the miscompile in one location in all of v8 on ppc64 doesn't mean that amd64/arm64 aren't affected in a less dramatic way -- the bug only triggers on a couple of sites for me, but what's making me wonder more about miscompilation on all architectures is the fact that upstream Chromium doesn't understand the NULL pointer bug I patched out in the GPU process. That one absolutely hit amd64; I was having the browser process crash every second time I tried a Google search on the AMD GPU systems. Since the miscompilation here basically changes a variable to NULL, it's starting to smell like the same compiler bug..."
and "I tried compiling the same problematic code for aarch64 on llc-16 and llc-17. There is a significant difference in generated assembly in the exact same location of the codebase as ppc64; the only difference is I think on aarch64 the differences will lead to incorrect operation vs. an outright crash. I really, *really* would not trust clang 16 on current Chromium releases at this point!"Any strong opinions (or suggestions) from the llvm maintainers or RMs at this point?
OpenPGP_signature.asc
Description: OpenPGP digital signature
