Re: [PATCH] tree-optimization/104263 - avoid retaining abnormal edges for non-call/goto stmts

[Richard Biener <rguenther@suse.de>]

On Fri, 28 Jan 2022, Jakub Jelinek wrote:

> On Fri, Jan 28, 2022 at 11:29:38AM +0100, Richard Biener wrote:
> > This removes a premature optimization from
> > gimple_purge_dead_abnormal_call_edges which, after eliding the
> > last setjmp (or computed goto) statement from a function and
> > thus clearing cfun->calls_setjmp, leaves us with the abnormal
> > edges from other calls that are elided for example via inlining
> > or DCE.  That's a CFG / IL combination that should be impossible
> > (not addressing the fact that with cfun->calls_setjmp and
> > cfun->has_nonlocal_label cleared we should not have any abnormal
> > edge at all).
> > 
> > For the testcase in the PR this means that IPA inlining will
> > remove the abormal edges from the block after inlining the call
> > the edge was coming from.
> 
> Couldn't DCE when it clears calls_setjmp and doesn't set it again
> (I think we never clear has_nonlocal_label) temporarily set
> calls_setjmp and gimple_purge_all_dead_abnormal_call_edges
> with it?
> Or have next to calls_setjmp a maybe_calls_setjmp flag that
> would be sticky like has_nonlocal_labels and would be never cleared?

I suppose we could do things like this.  Note in CFG cleanup
we call gimple_purge_dead_eh_edges on each block but not
gimple_purge_dead_abnormal_call_edges.  DCE, when resetting the
flag could also manually axe all abnormal edges in the function.

Still I think assuming there are no abnormal edges when neither
of the flag is set is premature (as can be seen here).  I also
don't think what we do in the function is very timing critical,
but sure, we walk all successor edges.

uninit has interesting code checking ->calls_setjmp conditionally
ignoring abnormal SSA names but only then ... (it should be
able to ignore them when _none_ of the flags is set instead).

That said, gimple_purge_dead_abnormal_call_edges wants to check
"are there possibly any abnormal edges in the function" and clearly
testing just the flags doesn't do it but resetting the flag was
important enough to cut out (sometimes bogus?) checks before
optimizations.

Richard.