[Bug debug/108656] [12/13 Regression] '-fcompare-debug' failure (length) w/ -O2 -fno-ipa-pure-const -fno-tree-dce --param early-inlining-insns=0 since r12-5236

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108656

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #7)
> (In reply to Richard Biener from comment #6)
> > (In reply to Jakub Jelinek from comment #5)
> > > Created attachment 54412 [details]
> > > gcc13-pr108656.patch
> > > 
> > > So shall we fix it like this then?
> > 
> > But isn't this the wrong "side"?  returns_twice means it is the abnormal
> > control _receiver_, it doesn't perform an abnormal goto itself.
> > 
> > stmt_starts_bb_p is correct here, so where does it go wrong?
> 
> I think it needs both.  The thing is, when it returns the second time, it
> does that again by returning from itself, not through returning from some
> unrelated function.
> Say, if we have pure + returns_twice call and no other call in a function,
> the abnormal edge to the pure + returns_twice call would be optimized away,
> even when the call remains, because there would be no edge from some call to
> the .ABNORMAL_DISPATCHER block.
> __attribute__((pure, returns_twice)) int foo (void);
> 
> int
> bar (void)
> {
>   for (int i = 0; i < 64; ++i)
>     {
>       int x = foo ();
>       if (x == 26)
>         return -42;
>     }
>   return 42;
> }
> doesn't even have because of this any abnormal edges created.
> Or, if there is some other non-pure call somewhere else, we model through
> the abnormal
> edges that that other call can pass control back to the start of the
> returns_twice call to make it return again.

Ah, but then the issue is that we assume that 'foo' doesn't longjmp,
independent on whether it is returns_twice or not?  Can a setjmp ()
function perform a longjmp () to its own context?  Would it even appear
as returning twice then?  Would calling setjmp in a loop like above
and then jumping to another iteration via longjmp even be valid?

That said, for your example I would think not having any abnormal edges
is correct - there's no frame that could transfer control back to the
returns_twice receiver, no?