[PATCH] Reduce garbage produced by invert_truthvalue

[Richard Guenther]

On Thu, 15 Jun 2006, Roger Sayle wrote:

> 
> On Wed, 14 Jun 2006, Richard Guenther wrote:
> > 2006-06-14  Richard Guenther  <rguenther@suse.de>
> >
> > 	* fold-const.c (fold_truth_not_expr): Rename from
> > 	invert_truthvalue.  Give it fold_* semantics to avoid
> > 	generating garbage.
> > 	(invert_truthvalue): New function.  Wrapper around
> > 	fold_truth_not_expr.
> > 	(optimize_minmax_comparison): Avoid creating garbage.
> > 	(fold_unary): Use fold_truth_not_expr for folding
> > 	TRUTH_NOT_EXPR.
> > 	(fold_ternary): Replace uses of invert_truthvalue with
> > 	fold_truth_not_expr where applicable.
> > 	* tree.h (fold_truth_not_expr): Prototype.
> 
> This is OK for mainline.  invert_truthvalue is popular enough to
> show up in GCC's profiles, so reducing its overhead is a good
> thing even in stage3.
> 
> !    FIXME: one would think we would fold the result, but it causes
> !    problems with the dominator optimizer.  */
> 
> Did you check whether this prexisting comment is still true?  DOM
> has probably changed significantly since this comment was written,
> and not being able to canonicalize/fold trees may indicate a real
> latent bug somewhere.

No, I didn't check that.  Given that we are in stage3 I tried to
minimize observable difference for the patch.

Richard.

--
Richard Guenther <rguenther@suse.de>
Novell / SUSE Labs