Re: [RFC PATCH] tree-ssa-sink: do not sink to in front of setjmp

[Richard Biener <richard.guenther@gmail.com>]

On Mon, Jan 3, 2022 at 5:35 PM Alexander Monakov <amonakov@ispras.ru> wrote:
>
> On Mon, 3 Jan 2022, Richard Biener wrote:
>
> > > @@ -5674,6 +5675,14 @@ gimple_verify_flow_info (void)
> > >                err = 1;
> > >              }
> > >
> > > +         if (prev_stmt && stmt_starts_bb_p (stmt, prev_stmt))
> >
> > stmt_starts_bb_p is really a helper used during CFG build, I'd rather
> > test explicitely for a GIMPLE call with ECF_RETURNS_TWICE, or maybe,
> > verify that iff a block has abnormal predecessors it starts with such
> > a call (because IIRC we in some cases elide abnormal edges and then
> > it's OK to move "down" the calls).  So yes, if a block has abnormal preds
> > it should start with a ECF_RETURNS_TWICE call, I think we cannot
> > verify the reverse reliably - abnormal edges cannot easily be re-built
> > in late stages (it's a bug that we do so during RTL expansion).
>
> Thanks, I could rewrite the patch along those lines, but I'm not sure where
> this is going: the ~100 extra FAILs will still be there. What would the next
> steps be for this patch and the initial tree-ssa-sink patch?

I approved the initial sink patch (maybe not clearly enough).  Can you open
a bugreport about the missing CFG verification and list the set of FAILs
(all errors in some passes similar to the one you fixed in sinking I guess)?
It indeed sounds like something to tackle during next stage1 (unless you
already narrowed down the culprit to a single offender...)

Richard.

>
> Alexander