This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PING] Move costs, allocno costs, and per-register costs

Vladimir Makarov <> writes:
> Bernd Schmidt wrote:
>> Vladimir Makarov wrote:
>>> Richard Sandiford wrote:
>>> In
>>> you wrote
>>>  In summary, I think ira-costs.c:record_reg_classes should not add
>>>  ira_may_move_out_cost and ira_memory_cost for unused output operands.
>>>  I'd like to test that at some point, but I've got a lot on my plate
>>>  and probably won't find time.  (An agreement in principle might
>>>  add a bit of incentive though.)
>>> I think it would be interesting to see the results.  Another would be
>>> interesting code is to check dealing with equiv memory.  The code
>>> quality is very sensitive to changes in ira-costs.c.
>> Do you have something like the patch below in mind?
> Yes, thanks.
>>   I tested this
>> against Richard's patches for a Blackfin target, with inconclusive
>> results.  It did seem to help with the testcase Richard found that had a
>> regression; other than that there weren't too many differences in code
>> quality that I could find.
> I've just finished benchmarking it on x86 too.  Two programs from 
> SPECInt2000 has the same code.  Other SPECInt2000 program has a bit 
> better code (like 0.2% but it is in measurement error range even for my 
> most stable machine which has a very small measurement range).
> As for SPECFp2000 about half programs have the same code, the code 
> quality of the rest are simply the same.
> So I also see no cons or pros for the patch.

Oh well.  Thanks to both of you for trying.  TBH, I never really
expected this to make a vast difference to things like SPEC scores.
I just thought it might do more good than harm in the cases that
it did affect.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]