odd instruction reordering (was Re: [Sbcl-devel] cannot build with sb-thread enabled)

[Daniel B. <da...@te...>]

=2D----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Istvan Marko <mi...@im...> writes:

> Dump of assembler code for function arch_os_get_current_thread:
> 0x80559b8 <arch_os_get_current_thread>: mov    %fs:0x38,%eax
> 0x80559be <arch_os_get_current_thread+6>:       push   %ebp
> 0x80559bf <arch_os_get_current_thread+7>:       mov    %esp,%ebp
> 0x80559c1 <arch_os_get_current_thread+9>:       xor    %edx,%edx
> 0x80559c3 <arch_os_get_current_thread+11>:      cmpl   $0x0,0x8061a30
> 0x80559ca <arch_os_get_current_thread+18>:      cmovne %eax,%edx
> 0x80559cd <arch_os_get_current_thread+21>:      mov    %edx,%eax
> 0x80559cf <arch_os_get_current_thread+23>:      mov    %ebp,%esp
> 0x80559d1 <arch_os_get_current_thread+25>:      pop    %ebp
> 0x80559d2 <arch_os_get_current_thread+26>:      ret=20=20=20=20
> End of assembler dump.

> (gdb) info registers=20
> fs             0x0      0

A load from %fs:0x38 is probably not a really good thing to do before
it's been set up to point to a valid ldt entry (which happens later,
in thread init).

But that's why we test if(all_threads) before doing it.  The version
of gcc you're using seems to have reordered stuff so that the thread
dereference happens before the test, which is obviously undesirable.
I don't know if this is a gcc bug or (at least as likely) there's some
constraint missing on the asm instruction that should be sitting there
to prevent it being reordered.

I don't suppose we have any more competent gcc users on the list, do we?

=2D -dan

=2D --=20

   http://www.cliki.net/ - Link farm for free CL-on-Unix resources=20
=2D----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.1 (GNU/Linux)

iD8DBQE+17hdHDK5ZnWQiRMRAqBAAJ9maJYNxi6ackAfNkz0Xi7bsjlORgCcCbda
OkyFK1DUJ4FRCR1BH67V7MY=3D
=3Dx1nI
=2D----END PGP SIGNATURE-----