[Bug gold/20805] gcc's ThreadSanitizer broken by gold

[ccoutant at gmail dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=20805

--- Comment #12 from Cary Coutant <ccoutant at gmail dot com> ---
(In reply to Markus Trippelsdorf from comment #8)
> But Jakub wrote in the gcc bug:
> 
> »As for the ld.bfd optimization that makes linking with ld.bfd work, that is
> an optimization if there are any IE relocations then other GD/LD relocations
> are turned into them too, because once you have any IE relocations, all of
> the TLS has to be static anyway, so the TLS GD/LD calls are just a waste of
> time.«
> 
> So either this is a missed optimization for gold, or ld.bfd generates wrong
> code.

Yes, we could take the presence of a R_X86_64_GOTTPOFF relocation as a signal
that it's OK to optimize for the IE model. But to do it reliably, it would
require us to pre-scan the relocations before making any optimization
decisions. If we see a GD relocation that could have been optimized before we
see any GOTTPOFF relocations, we wouldn't be able to optimize that relocation.

I don't think that slowing the link down by adding another pass over the
relocations is worth it.

-- 
You are receiving this mail because:
You are on the CC list for the bug.