Re: [PING^1][PATCH v4 3/3] PR80791 Consider doloop cmp use in ivopts

["Kewen.Lin" <linkw@linux.ibm.com>]

Hi Richard,

on 2019/7/12 下午8:11, Richard Biener wrote:
> On Wed, 10 Jul 2019, Kewen.Lin wrote:
> 
>> Hi all,
>>
>> I'd like to gentle ping the below patch:
>> https://gcc.gnu.org/ml/gcc-patches/2019-06/msg01225.html
>>
>> The previous version for more context/background:
>> https://gcc.gnu.org/ml/gcc-patches/2019-06/msg01126.html
>>
>> Thanks a lot in advance!
> 
> Again I would have hoped Bin to chime in here.
> 
> Am I correct that doloop HW implementations are constrainted
> by a decrement of one?  I see no code in the patch to constrain
> things this way.  

If my understanding is correct, under have_count_reg_decr_p
I don't think we should check the decrement one pattern, doloop
can transform the loop closing to decrement by 1 since it knows
the iteration total count.  Since it uses special hardware register
like Power count register, we don't expect it to be shared with
other uses.  Btw, it also doesn't require the compare should be the 
comp/decrement pattern, so this patch more focuses on this comp
is needed or not (should be considered in selection or not).

> I'm not familiar with the group code at all
> but I would have expected the patch to only affect costing
> of IVs of the appropriate form (decrement one and possibly
> no uses besides the one in the compare/decrement).  

But since we select IV cand for every IV uses, we never knows
this IV cand will have the only use till the whole selection
done.

> Since
> ivcanon already adds a canonical counter IV it's not
> necessary to generate an artificial candidate IV of the
> wanted style (that's something I might have expected as well).

This patch is only for the case guarded in have_count_reg_decr_p.
It doesn't requires to have the artificial candidate IV as well
as decrement-compare-jump code sequence.  The code on power looks
like:  mtctr Rx   // move Rx (which holding total_counter) 
                  // to ctr register
      L:                   
       loop body...
       bnze L     // decrease ctr register and jump to L if 
                  // ctr nonzero    

> 
> The rest should be just magic from the IVOPTs side?
> 
> There might be the need to only consider at most one counter IV
> in the costing code.

The current patch doesn't introduce any IV cands but focus on 
zeroing the cost of comp IV use since we know it will be eliminated.
Still to leverage the existing candidate selection algorithm to decide
the final optimal IV set.  Bring back the canonical counter IV only
if it's not selected by any IV uses to keep the doloop comp use
rewriting correct, but it shouldn't affect anything since the use will
be eliminated and is the only use, the IV and its related will be
removed as well.


Thanks,
Kewen