Re: x86_64-gnu 14.2.0 cross-compiler -O2 removes THREAD_SETMEM in glibc sigreturn.c
Andreas Schwab, le dim. 24 nov. 2024 15:15:43 +0100, a ecrit: > On Nov 24 2024, Sergey Bugaev wrote: > > So are you saying that we always must mark any asm statement that > > might transfer control somewhere else w/o returning as 'asm goto', > > even if we don't actually need to jump to any of the C-level labels? > > An ordinary asm is not allowed to change flow control. I was suggesting on IRC to rather move the asm bit outside any function and just call it as a function (with attribute noreturn). Samuel
Re: x86_64-gnu 14.2.0 cross-compiler -O2 removes THREAD_SETMEM in glibc sigreturn.c
Samuel Thibault, le dim. 24 nov. 2024 12:44:00 +0100, a ecrit:
> Sergey Bugaev, le dim. 24 nov. 2024 14:35:33 +0300, a ecrit:
> > Reduced further:
> >
> > --8<--
> > struct hurd_sigstate;
> >
> > typedef struct
> > {
>
> [... the content doesn't actually matter]
>
> > unsigned int reply_port;
> > } tcbhead_t;
> >
> > void
> > __sigreturn2 (struct hurd_sigstate *ss, unsigned long *usp,
> > unsigned int sc_reply_port)
> > {
> > (* (unsigned int __seg_fs *) __builtin_offsetof (tcbhead_t,
> > reply_port) = sc_reply_port);
> >
> > #ifdef ADD_NOP
> > asm ("nop");
> > #endif
> >
> >
> > asm volatile ("movq %0, %%rsp\n"
> > "retq $128" :
> > : "rm" (usp));
> >
> > __builtin_unreachable ();
> > }
> > >8---
>
> Could it be simply because __builtin_unreachable tells gcc that the
> function is not supposed to actually execute, just because it doesn't
> know that the retq asm snippet is indeed a noreturn? Can we tell gcc
> that?
Sergey points me at
https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html#Goto-Labels
“
GCC assumes that asm execution falls through to the next statement (if
this is not the case, consider using the __builtin_unreachable intrinsic
after the asm statement)
”
so it's the documented way to make it so, but apparently it does more
than this and affects the fs-segmented store.
Samuel
Re: x86_64-gnu 14.2.0 cross-compiler -O2 removes THREAD_SETMEM in glibc sigreturn.c
Sergey Bugaev, le dim. 24 nov. 2024 14:35:33 +0300, a ecrit:
> Reduced further:
>
> --8<--
> struct hurd_sigstate;
>
> typedef struct
> {
[... the content doesn't actually matter]
> unsigned int reply_port;
> } tcbhead_t;
>
> void
> __sigreturn2 (struct hurd_sigstate *ss, unsigned long *usp,
> unsigned int sc_reply_port)
> {
> (* (unsigned int __seg_fs *) __builtin_offsetof (tcbhead_t,
> reply_port) = sc_reply_port);
>
> #ifdef ADD_NOP
> asm ("nop");
> #endif
>
>
> asm volatile ("movq %0, %%rsp\n"
> "retq $128" :
> : "rm" (usp));
>
> __builtin_unreachable ();
> }
> >8---
Could it be simply because __builtin_unreachable tells gcc that the
function is not supposed to actually execute, just because it doesn't
know that the retq asm snippet is indeed a noreturn? Can we tell gcc
that?
Samuel
Re: x86_64-gnu 14.2.0 cross-compiler -O2 removes THREAD_SETMEM in glibc sigreturn.c
Hello, Samuel Thibault, le dim. 24 nov. 2024 15:18:31 +0100, a ecrit: > Andreas Schwab, le dim. 24 nov. 2024 15:15:43 +0100, a ecrit: > > On Nov 24 2024, Sergey Bugaev wrote: > > > So are you saying that we always must mark any asm statement that > > > might transfer control somewhere else w/o returning as 'asm goto', > > > even if we don't actually need to jump to any of the C-level labels? > > > > An ordinary asm is not allowed to change flow control. > > I was suggesting on IRC to rather move the asm bit outside any function > and just call it as a function (with attribute noreturn). That fixes things on my box, I pushed it to the glibc tree. Samuel
