[PATCH v2] x32: Use THREAD_SELF for RSEQ_* accessors

[H.J. Lu]

On Sat, Aug 23, 2025 at 1:34 AM Uros Bizjak <ubizjak@gmail.com> wrote:
>
> On Fri, Aug 22, 2025 at 10:19 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> >
> > Use THREAD_SELF to get the %fs register to work around the GCC bug:
> >
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121613
> >
> > with the negative offset from the %fs register.  Comparing against the
> > asm statements, for RSEQ_GETMEM_ONCE in sched_getcpu:
> >
> >   asm statements                       C with THREAD_SELF
> >     endbr64                              endbr64
> >     subl    $40, %esp                    subl    $40, %esp
> >                                   >      movl    __rseq_offset(%rip),
> >     movl    %fs:24, %eax                 movl    %fs:24, %eax
> >     movl    %eax, 28(%rsp)               movl    %eax, 28(%rsp)
> >     movslq  __rseq_offset(%rip),  |      movl    %fs:8, %eax
> >     movl %fs:4(%rax),%eax         |      leal    4(%rax,%rdx), %eax
> >                                   >      movl    (%eax), %eax
>
> So, instead of one memory access:
>
> movl %fs:4(%rax),%eax
>
> we now have two:
>
> movl    %fs:8, %eax
> movl    (%eax), %eax
>
> for *every* access to the RSEQ area. This nullifies all optimization
> efforts to use direct %fs references to TCB, and is IMO too high a
> price to pay for a more readable source. For this workaround, assembly
> is hands down a better choice, and the eventual GCC fix will also be
> similar: sign-extend the non-const value to a temporary DImode
> register
>
> IMO the approach that creates direct %fs reference with assembly
> outweighs the slightly more complex source. Also, please note that the
> current assembler implementation lacks "memory" clobber, for which I
> am preparing a patch.

I dropped my patch.  I will wait for your updated patch.

> An alternative approach would be to figure why is __rseq_offset
> negative. One would expect that since %fs points to the beginning of
> the TCB block, then all offsets from %fs should be positive. Perhaps
> the RSEQ area could be placed elsewhere in the TCB block to guarantee
> positive offset?

I think it is done on purpose to keep the TCB size unchanged.

-- 
H.J.