On 2026/01/26 18:14, Rafael Sadowski wrote:
> On Mon Jan 26, 2026 at 02:44:39PM +0000, Stuart Henderson wrote:
> > On 2026/01/26 15:23, Rafael Sadowski wrote:
> > > On Mon Jan 26, 2026 at 03:05:32PM +0100, Claudio Jeker wrote:
> > > > ports-gcc is modern enough to support -flto but our binutils and ld in
> > > > base are not. So compiles break whenever lto is enabled on archs like
> > > > sparc64.
> > > > 
> > > > This diff disables -flto and -fno-fat-lto-objects for OpenBSD which is
> > > > maybe a big hammer but the simplest fix I came up with.
> > > 
> > > We could hide it with MODCMAKE_PORT_BUILD like. MODCMAKE_PORT_BUILD
> > > only defined during ports build. This changes the behaviour only for
> > > ports, but not for end users, who receive it as expected by cmake
> > > behaviour.
> > 
> > Whether it's a ports build or not, LTO will not work with the old
> > ld.bfd that's in base, so this is a compatibility fix and needed for
> > any use whether that's in ports builds or not, this differs from
> > the usual things that get hidden behind MODCMAKE_PORT_BUILD.
> 
> You cannot know what someone is doing with cmake under OpenBSD and which
> linkers etc. they are using. You may want to use GCC15 and a wip-linker.
> Unlikely, but possible.
> 
> > 
> > The best fix for cmake would be to check the linker version as well
> > as the compiler version before enabling LTO, I think (and that is
> > probably valid to go upstream). 
> 
> Agree.
> 
> > But if that's too awkward then
> > Claudio's diff seems reasonably targetted (though will disable LTO
> > in some cases where it works - i.e. a ports-gcc build on an arch
> > using ld.lld).
> > 
> 
> Yes, see above
> 

But disabling LTO in some unlikely corner cases seems better than
breaking the build in some more-common caes..

Reply via email to