On Thu, 6 Feb 2014, Jakub Jelinek wrote:
> On Thu, Feb 06, 2014 at 10:33:54AM +0100, Richard Biener wrote:
> > > Not even in call_can_make_abnormal_goto? I mean, as we split
> > > __builtin_setjmp into __builtin_setjmp{_setup,_receiver} and the former
> > > has abnormal edge out of it, there is no need to have another abnormal
> > > edge
> > > out of the receiver.
> >
> > longjmp has outgoing abnormal edges, setjmp has
> > incoming abnormal edges. So how's the dispatcher working exactly?
> > It receives all abnormal edges from all possibly longjmp calling
> > stmts and dispatches them to all setjmp calls?
>
> I've actually removed the special __builtin_setjmp_dispatcher and replaced
> it with the same ABNORMAL_DISPATCHER used for normal setjmp/returns_twice
> calls. __builtin_setjmp_receiver actually is not really a returns_twice
> function, it only has the abnormal predecessor edge (from
> ABNORMAL_DISPATCHER) and no other predecessor edge, so it really works as a
> normal call with just the requirement that it is at the start of the bb.
> We just need to treat it as receiver of the abnormal edges (we do already).
> It is __builtin_setjmp_setup that has both normal fallthru edge and abnormal
> edge to the dispatcher. The dispatcher is expanded as nothing at all (it's
> basic block is actually removed), the receiver as loading of frame pointer
> etc. from somewhere and setup as saving frame pointer somewhere.
>
> > So why would there be extra edges from the receiver?
>
> That is what I'd like to change. The reason for the extra edge from the
> receiver is that it is a non-leaf call. Supposedly if we set ECF_LEAF on
> __builtin_setjmp_receiver, it might work, but that looks too dangerous to
> me, at least right now.
Ah, so __builtin_setjmp_receiver is like setjmp in this regard
and setjmp is LEAF (it's a stmt that doesn't direct control-flow
anywhere else). So __builtin_setjmp_receiver should be LEAF as well
(and so should __builtin_setjmp_setup).
Doesn't sound dangerous at all to me ...
(same applies to trampoline setup and stack save/restore, _not_ to
__builtin_nonlocal_goto though)
> > > > stmts we add abnormal edges to. Btw, why isn't BUILT_IN_SETJMP_RECEIVER
> > > > ECF_RETURNS_TWICE?
> > >
> > > See PR60003, #c5 in particular which really didn't work at all, and Eric's
> > > preference in #c8.
> >
> > Hmm, but it really is ECF_RETURNS_TWICE ...
>
> It is not.
Then __builtin_setjmp_setup is?
Looks like I have to build a cross to a sjlj EH target ...
Richard.
