[Bug tree-optimization/114151] [14 Regression] weird and inefficient codegen and addressing modes since r14-9193

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114151

--- Comment #22 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Andrew Macleod from comment #21)
> (In reply to Richard Biener from comment #19)
> 
> > 
> > While ranger has a range_on_exit API this doesn't work on GENERIC expressions
> > as far as I can see but only SSA names but I guess that could be "fixed"
> > given range_on_exit also looks at the last stmt and eventually defers to
> > range_of_expr (or range_on_entry), but possibly get_tree_range needs
> > variants for on_entry/on_exit (it doesn't seem to use it's 'stmt' context
> > very consistently, notably not for SSA_NAMEs ...).
> 
> That would appear to be an oversight. That API has not been used very much
> for arbitrary generic trees.  I think the original reason support for tree
> expressions was added was a "try this" for some other PR. It was simple to
> do so we lef tit in, but it never got any real traction.  At least as far as
> I can recall :-)
> 
> Currently, I think mosrt, if not all, uses of get_tree_range() are either
> !gimple_ssa_range_p() (commonly constants or unsupported types) or ssa_names
> on abnormal edges. 
> 
> For abnormal edges, we ought to be getting the global range directly these
> days instad of calling that routine.   Then in get_tree_range (), we ought
> to be calling range_of_expr for SSA_NAMES with the provided context.  I'll
> poke at that too. The support for general tree expressions changed the
> original intent of the function, and it should be adjusted. 
> 
> As for the on-exit/on-entry bits... we haven't had a need for entry/exit
> outside of ranger in the past.  I had toyed with exporting those routines
> and making them a part of the official API for value-query, but hadn't run
> across the need as yet.
> 
> Let me think about that for a minute. It can certainly be done. I guess we
> really only need an on-entry and on-exit version of range_of_expr to do
> everything.  So if we end up with something like:  
>   range_of_expr (r, expr, stmt)
>   range_of_expr_on_entry  (r, expr, bb)
>   range_of_expr_on_exit (r, expr, bb)
> 
> And have that all work with general trees expressions.. That would solve
> much of this for you?

Yes, I wouldn't mind if range_on_{entry,exit} handle general tree expressions,
there's enough APIs to be confused with already ;)

> 
> 
> 
> > 
> > Interestingly enough we somehow still need the
> > 
> 
> > 
> > hunk of Andrews patch to do it :/
> > 
> 
> That probably means there is another call somewhere in the chain with no
> context. However, I will say that functionality is more important than it
> seems. Should have been there from the start :-P.

Possibly yes.  It might be we fill rangers cache with VARYING and when
we re-do the query as a dependent one but with context we don't recompute
it?  I also only patched up a single place in SCEV with the context so
I possibly missed some others that end up with a range query, for example
through niter analysis that might be triggered.