This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
RE: [patch] Vectorizer cost model implementation
- From: "Jagasia, Harsha" <harsha dot jagasia at amd dot com>
- To: gcc-patches at gcc dot gnu dot org
- Cc: "Meissner, Michael" <michael dot meissner at amd dot com>
- Date: Thu, 5 Apr 2007 15:37:57 -0500
- Subject: RE: [patch] Vectorizer cost model implementation
Currently 6 of the existing vector tests do not get vectorized and hence
do not pass, because the cost model (correctly) estimates that it's not
profitable to vectorize them. So we have disabled the cost model on all
of the existing vect tests. I plan to start by adding appropriate copies
of those tests that fail, with the cost model enabled. I will also add
more tests to verify various decisions that the cost model makes.
Please continue to provide review on the patch.
Thanks,
Harsha Jagasia
>-----Original Message-----
>From: Jagasia, Harsha
>Sent: Thursday, April 05, 2007 3:21 PM
>To: Jagasia, Harsha
>Subject:
>
>> -----Original Message-----
>> From: Steven Bosscher [mailto:stevenb.gcc@gmail.com]
>>
>> Is it your plan to have this accepted for the trunk or for a branch?
>
>The plan is to have it accepted to the trunk, I believe.
>
>>
>> If so, where's the ChangeLog entry?
>
>Oops! I forgot. My apologies.
>
>>
>> How did you test this?
>>
>
>I used the vect tests ... a number of them "fail" with the cost model
>enabled, and the cost model is correct in its assessment. I disabled
>the cost model for the test suite for now so that those tests can
>continue to function as intended. I.e. test specific vectorization
>capabilities. Those tests will be made into cost model tests.
>Additionally, more tests should be added as well that target the cost
>model.
>
>I know that these issues should have been addressed before the patch
was
>submitted and at the very least I should have done a better job of
>explaining the situation in my submission. So, I must again apologize.
>I'm not going to be working on this project any longer, and it was
>desired that I submit the patch under my name prior to leaving. As I
>said, though, my colleague Harsha will be addressing any shortcomings
in
>this patch and will be supporting it going forward.
>
>Thanks for you comments.
>
>Tony
>
>
>Regards,
>Harsha
>
>--------------------------------------------------
>Harsha Jagasia GNU Tools Team (SEE)
>512-602-1775 Advanced Micro Devices
>harsha.jagasia@amd.com Austin, TX
>