On 12/01/16 14:04, Richard Biener wrote:
> On Tue, 12 Jan 2016, Tom de Vries wrote:
>
>> On 12/01/16 12:22, Richard Biener wrote:
>>> Doesnt' the same issue apply to
>>>
>>>>> unsigned int *p;
>>>>>
>>>>> static void __attribute__((noinline, noclone))
>>>>> foo (void)
>>>>> {
>>>>>    unsigned int z;
>>>>>
>>>>>    for (z = 0; z < N; ++z)
>>>>>      ++(*p);
>>>>> }
>>> thus when we have a MEM_REF[p_1]?  SCEV will not analyze
>>> its evolution to a POLYNOMIAL_CHREC and thus access_fns will
>>> be NULL again.
>>>
>>
>> I didn't manage to trigger this scenario, though I could probably make it
>> happen by modifying ftree-loop-im to work in one case (the load of the value
>> of p) but not the other (the *p load and store).
>>
>>> I think avoiding a NULL access_fns is ok but it should be done
>>> unconditionally, not only for the DECL_P case.
>>
>> Ok, I'll retest and commit this patch.
>
> Please add a comment as well.

Patch updated with comment.

During testing however, I ran into two testsuite regressions:

1.

-PASS: gfortran.dg/graphite/pr39516.f   -O  (test for excess errors)
+FAIL: gfortran.dg/graphite/pr39516.f   -O  (internal compiler error)
+FAIL: gfortran.dg/graphite/pr39516.f   -O  (test for excess errors)

AFAIU, this is a duplicate of PR68976.

Should I wait with committing the patch until PR68976 is fixed?

2.

-XFAIL: gcc.dg/graphite/scop-pr66980.c scan-tree-dump-times graphite 
"number of SCoPs: 1" 1
+XPASS: gcc.dg/graphite/scop-pr66980.c scan-tree-dump-times graphite 
"number of SCoPs: 1" 1

AFAIU, this is not a real regression, but the testcase needs to be 
updated. I'm not sure how. Sebastian, perhaps you have an idea there?

Thanks,
- Tom