Question about simplify_cond_and_lookup_avail_expr

[Andrew Pinski]

On Nov 24, 2004, at 9:03 PM, Devang Patel wrote:

>
> On Nov 24, 2004, at 4:14 PM, Jeffrey A Law wrote:
>
>> On Wed, 2004-11-24 at 11:57 -0800, Devang Patel wrote:
>>> Related question, Why do find_equivalent_equality_comparison() only 
>>> for
>>> EQ and NE ?
>> I can't think of a good reason offhand.  I might have been
>> overly paranoid about whether or not we can apply that
>> tranformation with relational tests.
>
> OK. I am playing with following patch. It allows me to remove few 
> unnecessary casts that prevents me to auto vectorize SPEC gzip loop. 
> With this patch, plus couple of other patches not yet in mainline, I 
> am able to vectorize candidate gzip loop.
>
> I have not polished it to be a real patch and full testing is not yet 
> complete, but am I on the right track?

No this is in the wrong place, this should be done by a tree combiner.
I did post an example of how to do a tree combiner a while back.
I also did post a pass which removed redundant casts before, it was
rejected for the same things as I have mentioned above.  If you want
I can post a new tree combiner but it needs some work.  Also fold does
all the work for you already.

Thanks,
Andrew Pinski