This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [patch] More improvements for the code generation of the if-conversion
On Wed, Jun 30, 2010 at 03:22, Richard Guenther <rguenther@suse.de> wrote:
>> A substitution of depth 1 doesn't seem to be enough: if I am limiting
>> the level to 1, then with my current patch-set, I get all these failures:
>
> How is it that an improvement patch creates new FAILs? ?W/o your
> patch we also don't need to "unfold".
>
The patch-set contains the if-conversion of memory writes, so these
fails correspond to vect not able to handle the codes generated by the
if-conversion, see http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44710
> What kind of expressions do you need to fold? ?I suppose they
> are all comparisons and logical/bitwise and/or or nots?
>
Yes, see the gimple code in the PR, although I guess that this case
would be handled by level=1.
> Please have a look at the infrastructure Sandra added
> for the patch for PRs 39874 and 28685 then. ?This is how it should
> be done - not using fold and building large trees for any
> expression kinds you are running into. ?I bet you can simply
> re-use what Sandra added.
>
Thanks for the pointer, I will try to reuse this.
Sebastian