On Wed, May 28, 2025 at 06:35:58PM +0200, Peter Zijlstra wrote:
> On Wed, May 28, 2025 at 06:30:35PM +0200, Peter Zijlstra wrote:
> > On Wed, May 28, 2025 at 09:44:52AM +0200, Peter Zijlstra wrote:
> > > On Tue, May 06, 2025 at 12:18:49PM -0700, Josh Poimboeuf wrote:
> > >
> > > > Weird, I'm not seeing that.
> > >
> > > I Ate'nt Crazeh...
> > >
> > > https://lore.kernel.org/all/[email protected]/T/#u
> > >
> > > I'll go poke at it, see if today is the day I can figure out WTF
> > > happens.
> >
> > It manages to trip the CFI_UNDEFINED case in op->dest.reg == cfa->base
> > in update_cfi_state().
> >
> > I figured it ought to tickle the regular 'mov %rbp, %rsp' case above
> > there, but it doesn't, for some reason it has cfa.base == SP at this
> > point.
> >
> > This happens... /me looks in scrollback ... at POP_REGS 'pop
> > %rbp'. ARGH!!
> >
More fun!
> > So the sequence of fail is:
> >
> > push %rbp
> > mov %rsp, %rbp # cfa.base = BP
> >
> > SAVE
sub $0x40,%rsp
and $0xffffffffffffffc0,%rsp
This hits the 'older GCC, drap with frame pointer' case in OP_SRC_AND.
Which means we then hard rely on the frame pointer to get things right.
However, per all the PUSH/POP_REGS nonsense, BP can get clobbered.
Specifically the code between the CALL and POP %rbp below are up in the
air. I don't think it can currently unwind properly there.
> > ...
> > push %rbp
> > ...
> > pop %rbp # cfa.base = SP
>
> This is the POP !drap and dest==base case.
>
> > ...
> > mov %rbp, %rsp # UNDEF
> > nop # FAIL
> > RESTORE
> >
> > Note that the MOV+NOP is the 4 bytes ERETS needs.
