Re: [PATCH] Optimize in VRP loads from constant arrays (take 2)

[Jakub Jelinek <jakub at redhat dot com>]

On Tue, May 02, 2017 at 02:50:16PM +0200, Richard Biener wrote:
> > If array_at_struct_end_p is wrong, it should be fixed ;)
> 
> Indeed.  It was originally meant to say false if you can trust
> TYPE_DOMAIN of the array but now it says false if there's some means
> to constrain the array size (the DECL_P path and now your STRING_CST
> one).  But all callers afterwards just look at TYPE_DOMAIN ...

So shall we verify that TYPE_DOMAIN is consistent with the object size
in that case inside of array_at_struct_end_p?

> > > I'd restructure the patch quite different, using for_each_index on the
> > > ref gather an array of index pointers (bail out on sth unhandled).
> > > Then I'd see if I have interesting ranges for them, if not, bail out.
> > > Also compute the size product of all ranges and test that against
> > > PARAM_MAX_VRP_CONSTANT_ARRAY_LOADS.  Then store the minimum range
> > > value in the index places (temporarily) and use get_base_ref_and_extent to
> > > get at the constant "starting" offset.  From there iterate using
> > > the remembered indices (remember the ref tree as well via for_each_index),
> > > directly adjusting the constant offset so you can feed
> > > fold_ctor_reference the constant offsets of all elements that need to
> > > be considered.  As optimization fold_ctor_reference would know how
> > > to start from the "last" offset (much refactoring would need to be
> > > done here given nested ctors and multiple indices I guess).
> > 
> > But for this, don't you want to take it over?
> 
> I can try.  Is there a PR for this?

Ok, filed PR80603, it is now all yours.

> > I agree that the current implementation is not very efficient and that is
> > why it is also limited to that small number of iterations.
> > As many cases just aren't able to use the valueize callback, handling
> > arbitrary numbers of non-constant indexes would be harder.
> 
> Sure.  I'd have expected you simply handle ARRAY_REF of a VAR_DECL
> and nothing else ;)

That would be too simple and boring ;)

	Jakub