[PATCH][RFA][PR tree-optimization/79095] Improve overflow test optimization and avoid invalid warnings

[Richard Biener]

On Fri, Jan 27, 2017 at 11:21 PM, Jeff Law <law@redhat.com> wrote:
> On 01/27/2017 02:35 PM, Richard Biener wrote:
>>
>> On January 27, 2017 7:30:07 PM GMT+01:00, Jeff Law <law@redhat.com> wrote:
>>>
>>> On 01/27/2017 05:08 AM, Richard Biener wrote:
>>>>
>>>> On Fri, Jan 27, 2017 at 10:02 AM, Marc Glisse <marc.glisse@inria.fr>
>>>
>>> wrote:
>>>>>
>>>>> On Thu, 26 Jan 2017, Jeff Law wrote:
>>>>>
>>>>>>> I assume this causes a regression for code like
>>>>>>>
>>>>>>> unsigned f(unsigned a){
>>>>>>>   unsigned b=a+1;
>>>>>>>   if(b<a)return 42;
>>>>>>>   return b;
>>>>>>> }
>>>>>>
>>>>>>
>>>>>> Yes.  The transformation ruins the conversion into ADD_OVERFLOW for
>>>
>>> the +-
>>>>>>
>>>>>> 1 case.  However, ISTM that we could potentially recover the
>>>
>>> ADD_OVERFLOW in
>>>>>>
>>>>>> phi-opt.  It's a very simple pattern that would be presented to
>>>
>>> phi-opt, so
>>>>>>
>>>>>> it might not be terrible to recover -- which has the advantage that
>>>
>>> if a
>>>>>>
>>>>>> user wrote an optimized overflow test we'd be able to recover
>>>
>>> ADD_OVERFLOW
>>>>>>
>>>>>> for it.
>>>>>
>>>>>
>>>>>
>>>>> phi-opt is a bit surprising at first glance because there can be
>>>
>>> overflow
>>>>>
>>>>> checking without condition/PHI, but if it is convenient to catch
>>>
>>> many
>>>>>
>>>>> cases...
>>>>
>>>>
>>>> Yeah, and it's still on my TODO to add some helpers exercising
>>>> match.pd COND_EXPR
>>>> patterns from PHI nodes and their controlling condition.
>>>
>>> It turns out to be better to fix the existing machinery to detect
>>> ADD_OVERFLOW in the transformed case than to add new detection to
>>> phi-opt.
>>>
>>> The problem with improving the detection of ADD_OVERFLOW is that the
>>> transformed test may allow the ADD/SUB to be sunk.  So by the time we
>>> run the pass to detect ADD_OVERFLOW, the test and arithmetic may be in
>>> different blocks -- ugh.
>>>
>>> The more I keep thinking about this the more I wonder if transforming
>>> the conditional is just more of a headache than its worth -- the main
>>> need here is to drive propagation of known constants into the THEN/ELSE
>>>
>>> clauses.  Transforming the conditional makes that easy for VRP & DOM to
>>>
>>> discover those constant and the transform is easy to write in match.pd.
>>>
>>> But we could just go back to discovering the case in VRP or DOM via
>>> open-coding detection, then propagating the known constants without
>>> transforming the conditional.
>>
>>
>> Indeed we can do that.  And in fact with named patterns in match.pd you
>> could even avoid the open-coding.
>
> negate_expr_p being the example?  That does look mighty interesting... After
> recognition we'd still have to extract the operands, but it does look like
> it might handle the detection part.

Yes, the (match ..) stuff is actually exported in gimple-match.c (just
no declarations
are emitted to headers yet).  logical_inverted_value might be a better example
given it has an "output":

(match (logical_inverted_value @0)
 (truth_not @0))

bool
gimple_logical_inverted_value (tree t, tree *res_ops, tree
(*valueize)(tree) ATTRIBUTE_UNUSED)
{

you get @0 in res_ops[0] if that returns true.  I've at some point
written some of tree-vect-patterns.c
as match.pd (match...) but never really completed it.  In the end it
would be nice to write the patterns
in-inline at use points, say,

  /* (match (foo @0) (....)) */
  if (gimple_foo (...))
   {
   }

and have some gen-program extract those patterns from source files
(plus inserting the necessary
prototypes locally?).

Richard.



> jeff