Re: [PATCH PR71734] Add missed check that reference defined inside loop.

[Yuri Rumyantsev <ysrumyan at gmail dot com>]

Richard,

The patch proposed by you does not work properly for
g++.dg/vect/pr70729-nest.cc test since the reference for this->S_n has
been cached as dependent for outer loop and loop is not vectorized:

 g++ -Ofast -fopenmp -mavx2 pr70729-nest.cc -c
-fdump-tree-vect-details
grep 'LOOP VECTORIZED' pr70729-nest.cc.149t.vect
<not found>

You missed additional check I added before check on cached dependence.

2016-08-09 13:00 GMT+03:00 Richard Biener <richard.guenther@gmail.com>:
> On Tue, Aug 9, 2016 at 11:20 AM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>> Yes it is impossible since all basic blocks are handled from outer
>> loops to innermost so we can not have the sequence with wrong
>> dependence, i.e. we cached that reference is independent (due to
>> safelen) but the same reference in outer loop must be evaluated as
>> dependent. So we must re-evaluate only dependent references in loops
>> having non-zero safelen attribute.
>
> Hmm.  I don't like depending on this implementation detail.  Does the
> attached patch work
> which simply avoids any positive/negative caching on safelen affected
> refs?  It also makes
> the query cheaper by avoiding the dive into child loops.
>
> Richard.
>
>> 2016-08-09 11:44 GMT+03:00 Richard Biener <richard.guenther@gmail.com>:
>>> On Fri, Aug 5, 2016 at 4:57 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>>>> Richard,
>>>>
>>>> I added additional check before caching dependencies since (1) all
>>>> statements in loop are handled in loop postorder order, i.e. form
>>>> outer to inner; (2) we can change dependency for reference in subloops
>>>> which have non-zero safelen attribute. So I propose to re-evaluate it
>>>> in such cases. I don't see why we need to avoid dependence caching for
>>>> all loop nests since pragma omp simd is used very rarely.
>>>
>>> You think it is impossible to construct a testcase which hits the
>>> correctness issue?
>>> "very rarely" is not a good argument to generate wrong code.
>>>
>>> Richard.
>>>
>>>> 2016-08-05 16:50 GMT+03:00 Richard Biener <richard.guenther@gmail.com>:
>>>>> On Fri, Aug 5, 2016 at 3:28 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>>>>>> Richard,
>>>>>>
>>>>>> Here is updated patch which implements your proposal - I pass loop
>>>>>> instead of stmt to determine either REF is defined inside LOOP nest or
>>>>>> not. I checked that for pr70729-nest.cc the reference this->S_n  for
>>>>>> statements which are out of innermost loop are  not considered as
>>>>>> independent as you pointed out.
>>>>>>
>>>>>> Regression testing did not show any new failures and both failed tests
>>>>>> from libgomp.fortran suite now passed.
>>>>>>
>>>>>> Is it OK for trunk?
>>>>>
>>>>> I don't quite understand
>>>>>
>>>>> +  /* Ignore dependence for loops having greater safelen.  */
>>>>> +  if (new_safelen == safelen
>>>>> +      && bitmap_bit_p (&ref->dep_loop, LOOP_DEP_BIT (loop->num, stored_p)))
>>>>>      return false;
>>>>>
>>>>> this seems to suggest (correctly I think) that we cannot rely on the caching
>>>>> for safelen, neither for optimal results (you seem to address that) but also
>>>>> not for correctness (we cache the no-dep result from a safelen run and
>>>>> then happily re-use that info for a ref that is not safe for safelen).
>>>>>
>>>>> It seems to me we need to avoid any caching if we made things independent
>>>>> because of safelen and simply not do the dep test afterwards.  this means
>>>>> inlining ref_indep_loop_p_1 partly into _2 (not sure if there's a great way
>>>>> to do this w/o confusing the control flow).
>>>>>
>>>>> Richard.
>>>>>
>>>>>> ChangeLog:
>>>>>> 2016-08-05  Yuri Rumyantsev  <ysrumyan@gmail.com>
>>>>>>
>>>>>> PR tree-optimization/71734
>>>>>> * tree-ssa-loop-im.c (ref_indep_loop_p): Add new argument REF_LOOP.
>>>>>> (outermost_indep_loop): Pass LOOP argumnet where REF was defined to
>>>>>> ref_indep_loop_p.
>>>>>> (ref_indep_loop_p_1): Fix commentary.
>>>>>> (ref_indep_loop_p_2): Add additional argument REF_LOOP, introduce new
>>>>>> variable NEW_SAFELEN which may have new value for SAFELEN, ignore
>>>>>> dependencde for loop having greater safelen value, pass REF_LOOP in
>>>>>> recursive call.
>>>>>> (can_sm_ref_p): Pass LOOP as additional argument to ref_indep_loop_p.
>>>>>>
>>>>>> 2016-08-03 16:44 GMT+03:00 Richard Biener <richard.guenther@gmail.com>:
>>>>>>> On Fri, Jul 29, 2016 at 4:00 PM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>>>>>>>> Hi Richard.
>>>>>>>>
>>>>>>>> It turned out that the fix proposed by you does not work for liggomp
>>>>>>>> tests simd3 and simd4.
>>>>>>>> The reason is that we can't change safelen value for references not
>>>>>>>> defined inside loop. So I add missed check on it to patch.
>>>>>>>> Is it OK for trunk?
>>>>>>>
>>>>>>> Hmm, I don't like the walk of all subloops in ref_defined_in_loop_p as
>>>>>>> that operation can end up being quadratic in the loop depth/width.
>>>>>>>
>>>>>>> But I also wonder about correctness given that LIM "commons"
>>>>>>> references.  So we can have
>>>>>>>
>>>>>>>   for (;;)
>>>>>>>     .. = ref;  (1)
>>>>>>>     for (;;) // safelen == 2  (2)
>>>>>>>       ... = ref;
>>>>>>>
>>>>>>> and when looking at the ref at (1) which according to you should not
>>>>>>> have safelen applied your function will happily return that ref is defined
>>>>>>> in the inner loop.
>>>>>>>
>>>>>>> So it looks like to be able to apply safelen the caller of ref_indep_loop_p
>>>>>>> needs to pass down a ref plus a location (a stmt).  In which case your
>>>>>>> function can simply use flow_loop_nested_p (loop, gimple_bb
>>>>>>> (stmt)->loop_father);
>>>>>>>
>>>>>>> Richard.
>>>>>>>
>>>>>>>> ChangeLog:
>>>>>>>> 2016-07-29  Yuri Rumyantsev  <ysrumyan@gmail.com>
>>>>>>>>
>>>>>>>> PR tree-optimization/71734
>>>>>>>> * tree-ssa-loop-im.c (ref_defined_in_loop_p): New function.
>>>>>>>> (ref_indep_loop_p_2): Change SAFELEN value for REF defined inside LOOP.
>>>>>>>>
>>>>>>>> 2016-07-29 13:08 GMT+03:00 Yuri Rumyantsev <ysrumyan@gmail.com>:
>>>>>>>>> Sorry H.J.
>>>>>>>>>
>>>>>>>>> I checked both these tests manually but forgot to pass "-fopenmp" option.
>>>>>>>>> I will fix the issue asap.
>>>>>>>>>
>>>>>>>>> 2016-07-29 0:33 GMT+03:00 H.J. Lu <hjl.tools@gmail.com>:
>>>>>>>>>> On Thu, Jul 28, 2016 at 6:49 AM, Yuri Rumyantsev <ysrumyan@gmail.com> wrote:
>>>>>>>>>>> Richard,
>>>>>>>>>>>
>>>>>>>>>>> I prepare a patch which is based on yours. New test is also included.
>>>>>>>>>>> Bootstrapping and regression testing did not show any new failures.
>>>>>>>>>>> Is it OK for trunk?
>>>>>>>>>>>
>>>>>>>>>>> Thanks.
>>>>>>>>>>> ChangeLog:
>>>>>>>>>>> 2016-07-28  Yuri Rumyantsev  <ysrumyan@gmail.com>
>>>>>>>>>>>
>>>>>>>>>>> PR tree-optimization/71734
>>>>>>>>>>> * tree-ssa-loop-im.c (ref_indep_loop_p_1): Pass value of safelen
>>>>>>>>>>> attribute instead of REF_LOOP and use it.
>>>>>>>>>>> (ref_indep_loop_p_2): Use SAFELEN argument instead of REF_LOOP and
>>>>>>>>>>> set it for Loops having non-zero safelen attribute.
>>>>>>>>>>> (ref_indep_loop_p): Pass zero as initial value for safelen.
>>>>>>>>>>> gcc/testsuite/ChangeLog:
>>>>>>>>>>>         * g++.dg/vect/pr70729-nest.cc: New test.
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Does this cause
>>>>>>>>>>
>>>>>>>>>> FAIL: libgomp.fortran/pr71734-1.f90   -O3 -fomit-frame-pointer
>>>>>>>>>> -funroll-loops -fpeel-loops -ftracer -finline-functions  execution
>>>>>>>>>> test
>>>>>>>>>> FAIL: libgomp.fortran/pr71734-1.f90   -O3 -g  execution test
>>>>>>>>>> FAIL: libgomp.fortran/pr71734-2.f90   -O3 -fomit-frame-pointer
>>>>>>>>>> -funroll-loops -fpeel-loops -ftracer -finline-functions  execution
>>>>>>>>>> test
>>>>>>>>>> FAIL: libgomp.fortran/pr71734-2.f90   -O3 -g  execution test
>>>>>>>>>>
>>>>>>>>>> on AVX machines and
>>>>>>>>>>
>>>>>>>>>> FAIL: libgomp.fortran/simd3.f90   -O3 -fomit-frame-pointer
>>>>>>>>>> -funroll-loops -fpeel-loops -ftracer -finline-functions  execution
>>>>>>>>>> test
>>>>>>>>>> FAIL: libgomp.fortran/simd3.f90   -O3 -g  execution test
>>>>>>>>>> FAIL: libgomp.fortran/simd4.f90   -O3 -fomit-frame-pointer
>>>>>>>>>> -funroll-loops -fpeel-loops -ftracer -finline-functions  execution
>>>>>>>>>> test
>>>>>>>>>> FAIL: libgomp.fortran/simd4.f90   -O3 -g  execution test
>>>>>>>>>>
>>>>>>>>>> on non-AVX machines?
>>>>>>>>>>
>>>>>>>>>> --
>>>>>>>>>> H.J.