Re: [patch] for PR 18529

[Roger Sayle <roger at eyesopen dot com>]

On Fri, 19 Nov 2004, Zdenek Dvorak wrote:
> here is the new version of the patch, with subst renamed and the
> a == b && something transformation moved out of fold.

Many thanks, this is *much* better.  tree_simplify_using_condition
is clearly the correct place for this, and allows even stronger
optimization than your original changes to "fold".

My only remaining comments are
(1) that for the time being, "simplify_replace_tree" can be kept local
to tree-ssa-loop-niter.c, and made static.  This also avoids adding
the prototype to tree.h.  If at some point in the future, someone
else requires this functionality, it can be moved to fold-const.c
and prototyped in tree.h, so that it can be shared.

(2) The comments in my previous mail's about in the inefficient use
of operand_equal_p in fold_to_nonsharp_ineq_using_bound.  You should
either call "fold" for strong equivalence, or alternatively continue
to use "weak" equivalence, but with the more efficient implememtation
suggested previously.


Thanks again for your patience, but don't forget to list the target
triples that you've bootstrapped and regression tested on when you
repost your updated patch.

Roger
--