------- Comment #12 from ro at CeBiTec dot Uni-Bielefeld dot DE 2010-03-09
19:14 -------
Subject: Re: amd64 TLS IE code sequence on Solaris 2/x86 violates spec
> ------- Comment #11 from hjl dot tools at gmail dot com 2010-03-09 18:45
> -------
> Sun linker changes
>
> 4: 64 48 8b 14 25 00 00 00 00 mov %fs:0x0,%rdx
> d: 48 8b 05 00 00 00 00 mov 0x0(%rip),%rax # 14 <main+0x14>
>
> 10: R_X86_64_GOTTPOFF cnt-0x4
>
> to
>
> 400e0c: 64 48 8b 04 25 00 00 00 00 mov %fs:0x0,%rax
> 400e15: 48 8d 80 f0 ff ff ff lea -0x10(%rax),%rax
>
> Nowhere in TLS spec allows the linker to change
>
> d: 48 8b 05 00 00 00 00 mov 0x0(%rip),%rax # 14 <main+0x14>
>
> 10: R_X86_64_GOTTPOFF cnt-0x4
>
> to
>
> 400e15: 48 8d 80 f0 ff ff ff lea -0x10(%rax),%rax
>
> It is Sun linker bug. Please report it to them.
True, this is a bug, but the input sequence isn't valid according to the
spec:
4: 64 48 8b 14 25 00 00 00 00 mov %fs:0x0,%rdx
d: 48 8b 05 00 00 00 00 mov 0x0(%rip),%rax # 14 <main+0x14>
This should be
mov %fs:0x0,%rax
instead. As I said, garbage in, garbage out. If you disagree, point me
at where the spec allows this. Even if so, we should either fix (if
allowing other registers is a GNU extension to the base spec) or work
around (if it is a misunderstanding on Sun's part) the problem if
targetting Solaris: there are linkers in the field that behave as
observed, and generating code that causes binaries to crash isn't a good
option.
Rainer
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43309
