Re: [PATCH] Be careful about combined chain with length == 0 (PR, tree-optimization/70754).

[Richard Biener <richard dot guenther at gmail dot com>]

On Wed, Jan 18, 2017 at 4:32 PM, Bin.Cheng <amker.cheng@gmail.com> wrote:
> On Wed, Jan 18, 2017 at 2:54 PM, Richard Biener
> <richard.guenther@gmail.com> wrote:
>> On Wed, Jan 18, 2017 at 11:10 AM, Martin Liška <mliska@suse.cz> wrote:
>>> Hello.
>>>
>>>
>>> Patch can bootstrap on ppc64le-redhat-linux and survives regression tests.
>>>
>>> Ready to be installed?
>>
>> I'm not sure.  If we have such zero distance refs in the IL at the
>> time pcom runs then not handling
>> them will pessimize code-gen for cases where they are part of a larger
>> chain.  Esp. I don't like
> Do you mean different chains distributed because of MAX_DISTANCE by
> "larger chain"?  With the patch, such chain of refs would still be
> pred-commoned, just the arithmetic operation not combined, which could
> be handled by later DOM?
>> another stmt_dominates_stmt_p call and thus rather not handle length
>> == 0 at all...
> Not handle length == 0 chains at all may be sub-optimal.  As you said,
> such chain of refs at the point may simply because previous dom/cse
> fail to analyze the references.
>>
>> We already seem to go great length in associating stuff when combining
>> stuff thus isn't this
>> maybe an artifact of this association?  Maybe we simply need to sort
>> the new chain after
>> combining it so the root stmt comes last?
>>
>> Note that there seems to be only a single length per chain but not all
>> refs in a chain need to
>> have the same distance.  This means your fix is likely incomplete?
>> What prevents the situation
>> to arise for distance != 0?
> Yes, it's possible for two refs have the same distance in a chain with
> length > 0.  But that should not be a problem, because existing uses
> are replaced by newly generated PHI variables which always dominate
> the uses, right?

I must admit I don't know predcom in such detail but then can we handle
distance == 0 by simply inserting a PHI for those as well (a degenerate
one of course)?  Or can for distance == 0 the ref be not loop invariant?

Note that for length == 0 all refs in the chain will have a dependence distance
of zero.  So my first argument probably doesn't hold and we could simply
remove handling of length == 0 chains and rely on CSE?

Richard.

> Thanks,
> bin
>>
>> Richard.
>>
>>> Martin
>>>