The bad code happens with the new inline assembly code for semaphores (it
_probably_ happens with other inline assembly code too, though - the
semaphore code is in no way special among linux inline asm usage). The
"down()" operation is defined as:
extern inline void down(struct semaphore * sem)
{
__asm__ __volatile__(
"# atomic down operation\n"
"1:\n\t"
"movl $1b,%%eax\n\t"
#ifdef __SMP__
"lock ; "
#endif
"decl %0\n\t"
"js " SYMBOL_NAME_STR(down_failed)
:/* no outputs */
:"m" (sem->count), "c" (sem)
:"ax","dx","memory");
}
Note especially the fact that the asm statement marks registers ax and dx as
being destroyed by the statement (%eax is used as a "return" address and is
loaded here, while %edx is potentially destroyed by the functions calls that
are called from "down_failed").
The problem is that this bad gcc doesn't take this into account, the code
produced in sd_init_onedisk looks like this:
...
call scsi_do_cmd
addl $0x1c,%esp
movl 16(%esp),%ecx
movl 16(%esp),%eax
/* down operation starts here */
movl $0x18d9dc,%eax
decl (%eax)
js 0x19bfa0 <down_failed>
/* end of down operations */
movl 0xf0(%edi),%esi
incl %ebx
testl %esi,%esi
je sd_init_onedisk+633
movb 0x8a(%edi),%dl
...
Note the use of "decl (%eax)" - gcc is allocating %eax for the input of the
down operation, even though the asm has marked %eax clobbered. And I quote
from the gcc info pages:
"The input operands are guaranteed not to use any of the clobbered
registers, and neither will the output operands' addresses, so you can read
and write the clobbered registers as many times as you like"
so this is definitely a buggy gcc that has produced the code above. But from
all reports it seems that gcc-2.7.2 and gcc-2.7.2.1 are ok, and there are
also version of the no-strength-reduce bugfix that are ok. Werner Fink is on
vacation right now, so I still don't know exactly what version of gcc he is
using (he did mention gnat, are there patches to the gcc back-end for ada
too?).
Linus