This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [patch, IRA] Fix PR46002
On 2011/2/22 äå 12:59, Jeff Law wrote:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
> On 02/21/11 08:59, Vladimir Makarov wrote:
>> On 02/19/2011 08:06 AM, Chung-Lin Tang wrote:
>>> Hi Vladimir,
>>> Thanks for reviewing my other patch for PR46178. Although you mentioned
>>> the priority algorithm is bound for removal, I still have another
>>> related patch.
>>>
>>> This fix is for PR46002: during update_copy_costs(), when we obtain the
>>> cover class for a copy-related allocno, it is not sufficient to test if
>>> the two reg-classes intersect; the assigned register in the current
>>> allocno must be *in* in the intersection. Otherwise the classes may
>>> intersect, but the currently assigned register is not valid for
>>> assigning to the other allocno (then ICE in the index>= 0 assertion).
>>>
>>> This patch changes the copy skipping test, from testing cover class
>>> intersection, into a test of if 'hard_regno' is within the cover class
>>> of 'another_allocno'.
>>>
>>> Bootstrapped and tested on x86_64 with no regressions. Comparing test
>>> results under -fira-algorithm=priority also shows many fewer FAILs. Okay
>>> for trunk?
>>>
>> Yes, it is ok for me. Thanks for working priority coloring PRs and FAILs.
> Likewise. Though we also need to be sure tests are being added to the
> testsuite for these failures.
>
> Also note that even though these PRs aren't marked as regressions, they
> clearly are regressions.
Committed patch at r170388.
Jeff, thanks for adding the testcases; I should have done that myself.
Chung-Lin