On Thu, Oct 17, 2019 at 9:16 AM Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote: > > On Wed, Oct 16, 2019 at 04:29:54PM -0700, Linus Torvalds wrote: > > On Wed, Oct 16, 2019 at 4:11 PM Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote: > > > > > > Looks like it was fixed soon after the complain: > > > > > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63567 > > > > Ahh, so there are gcc versions which essentially do this wrong, and > > I'm not seeing it because it was fixed. > > > > Ho humm. Considering that this was fixed in gcc five years ago, and we > > already require gc-4.6, and did that two years ago, maybe we can just > > raise the requirement a bit further. > > > > BUT. > > > > It's not clear which versions are ok with this. In your next email you said: > > > > > It would mean bumping GCC version requirements to 4.7. > > > > which I think would be reasonable, but is it actually ok in 4.7? > > I think, not. I don't have 4.7 around, but 4.9.3 has the issue if > -std=gnu99 is used. > > > The bugzilla entry says "Target Milestone: 5.0", and I'm not sure how > > to check what that "revision=216440" ends up actually meaning. > > > > I have a git tree of gcc, and in that one 216440 is commit > > d303aeafa9b, but that seems to imply it only made it into 5.1: > > > > [torvalds@i7 gcc]$ git name-rev --tags > > d303aeafa9b46e06cd853696acb6345dff51a6b9 > > d303aeafa9b46e06cd853696acb6345dff51a6b9 tags/gcc-5_1_0-release~3943 > > > > so we'd have to jump forward a _lot_. > > > > That's a bit sad and annoying. I'd be ok with jumping to 4.7, but I'm > > not sure we can jump to 5.1. > > > > Although maybe we should be a _lot_ more aggressive about gcc > > versions, I'm on gcc-9.2.1 right now, and gcc-5.1 is from April 22, > > 2015. > > 5.4.1 builds kernel fine for me with allmodconfig (minus retpoline which > requires compiler support). Both -std=gnu99 and -std=gnu11. > > Note that GCC has changed their version scheme. 5.4.1 is bug-fix release > of GCC-5. I tested -std=gnu99 for ARM with pre-built Linaro toolchains. GCC 4.9.4 was NG, GCC 5.3.1 was OK. If we increase the minimal GCC version, we might end up with dropping more architecture. I can no longer get the toolchains for hexagon, unicore32. https://mirrors.edge.kernel.org/pub/tools/crosstool/ provides hexagon compilers, but only for GCC 4.6.1 -- Best Regards Masahiro Yamada
