Re: PR81136: ICE from inconsistent DR_MISALIGNMENTs

[Richard Sandiford <richard dot sandiford at linaro dot org>]

Richard Biener <richard.guenther@gmail.com> writes:
> On Mon, Jun 26, 2017 at 1:14 PM, Richard Sandiford
> <richard.sandiford@linaro.org> wrote:
>> Richard Biener <richard.guenther@gmail.com> writes:
>>> On Fri, Jun 23, 2017 at 2:05 PM, Richard Sandiford
>>> <richard.sandiford@linaro.org> wrote:
>>>> Richard Biener <richard.guenther@gmail.com> writes:
>>>>> On Thu, Jun 22, 2017 at 1:30 PM, Richard Sandiford
>>>>> <richard.sandiford@linaro.org> wrote:
>>>>>> The test case triggered this assert in vect_update_misalignment_for_peel:
>>>>>>
>>>>>>           gcc_assert (DR_MISALIGNMENT (dr) / dr_size ==
>>>>>>                       DR_MISALIGNMENT (dr_peel) / dr_peel_size);
>>>>>>
>>>>>> We knew that the two DRs had the same misalignment at runtime, but when
>>>>>> considered in isolation, one data reference guaranteed a higher compile-time
>>>>>> base alignment than the other.
>>>>>>
>>>>>> In the test case this looks like a missed opportunity.  Both references
>>>>>> are unconditional, so it should be possible to use the highest of the
>>>>>> available base alignment guarantees when analyzing each reference.
>>>>>> The patch does this.
>>>>>>
>>>>>> However, as the comment in the patch says, the base alignment guarantees
>>>>>> provided by a conditional reference only apply if the reference occurs
>>>>>> at least once.  In this case it would be legitimate for two references
>>>>>> to have the same runtime misalignment and for one reference to provide a
>>>>>> stronger compile-time guarantee than the other about what the misalignment
>>>>>> actually is.  The patch therefore relaxes the assert to handle that case.
>>>>>
>>>>> Hmm, but you don't actually check whether a reference occurs only
>>> conditional,
>>>>> do you?  You just seem to say that for masked loads/stores the reference
>>>>> is conditional (I believe that's not true).  But for a loop like
>>>>>
>>>>>  for (;;)
>>>>>    if (a[i])
>>>>>      sum += b[j];
>>>>>
>>>>> you still assume b[j] executes unconditionally?
>>>>
>>>> Maybe the documentation isn't clear enough, but DR_IS_CONDITIONAL
>>>> was supposed to mean "even if the containing statement executes
>>>> and runs to completion, the reference might not actually occur".
>>>> The example above isn't conditional in that sense because the
>>>> reference to b[j] does occur if the store is reached and completes.
>>>>
>>>> Masked loads and stores are conditional in that sense though.
>>>> The reference only occurs if the mask is nonzero; the memory
>>>> isn't touched otherwise.  The functions are used to if-convert
>>>> things like:
>>>>
>>>>    for (...)
>>>>      a[i] = b[i] ? c[i] : d[i];
>>>>
>>>> where there's no guarantee that it's safe to access c[i] when !b[i]
>>>> (or d[i] when b[i]).  No reference occurs for an all-false mask.
>>>
>>> But as you touch generic data-ref code here you should apply more
>>> sensible semantics to DR_IS_CONDITIONAL than just marking
>>> masked loads/stores but not DRs occuring inside BBs only executed
>>> conditionally ...
>>
>> I don't see why that's more sensible though.  If a statement is only
>> conditionally executed in a loop, it's up to the consumer to decide
>> what to do about that.  The conditions under which the statement
>> is reached are a control-flow issue and tree-data-ref.c doesn't
>> have any special information about it.
>>
>> Masked loads and stores are special because the DR_REFs created by
>> tree-data-ref.c are artificial: they didn't exist as MEM_REFs in the
>> original DR_STMT.  And AIUI they didn't exist as MEM_REFs precisely
>> because they're not guaranteed to happen, even if the load or store
>> statement itself is executed.  So in this case the DR_IS_CONDITIONAL
>> is reflecting something that tree-data-ref.c itself has done.
>>
>> How about calling it DR_IS_CONDITIONAL_IN_STMT to avoid the
>> general-sounding name?
>
> That sounds better and avoids the ambiguity.

OK.

>>>>> The vectorizer of course only sees unconditionally executed stmts.
>>>>>
>>>>> So - I'd simply not add this DR_IS_CONDITIONAL.  Did you run into
>>>>> any real-world (testsuite) issues without this?
>>>>
>>>> Dropping DR_IS_CONDITIONAL would cause us to make invalid alignment
>>>> assumptions in silly corner cases.  I could add a scan test for it,
>>>> for targets with masked loads and stores.  It wouldn't trigger
>>>> an execution failure though because we assume that targets with
>>>> masked loads and stores allow unaligned accesses:
>>>>
>>>>   /* For now assume all conditional loads/stores support unaligned
>>>>      access without any special code.  */
>>>>   if (is_gimple_call (stmt)
>>>>       && gimple_call_internal_p (stmt)
>>>>       && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD
>>>>           || gimple_call_internal_fn (stmt) == IFN_MASK_STORE))
>>>>     return dr_unaligned_supported;
>>>>
>>>> So the worst that would happen is that we'd supposedly peel for
>>>> alignment, but actually misalign everything instead, and so make
>>>> things slower rather than quicker.
>>>>
>>>>> Note that the assert is to prevent bogus information.  Iff we aligned
>>>>> DR with base alignment 8 and misalign 3 then if another same-align
>>>>> DR has base alignment 16 we can't simply zero its DR_MISALIGNMENT
>>>>> as it still can be 8 after aligning DR.
>>>>>
>>>>> So I think it's wrong to put DRs with differing base-alignment into
>>>>> the same-align-refs chain, those should get their DR_MISALIGNMENT
>>>>> updated independenlty after peeling.
>>>>
>>>> DR_MISALIGNMENT is relative to the vector alignment rather than
>>>> the base alignment though.  So:
>>>
>>> We seem to use it that way, yes (looking at set_ptr_info_alignment
>>> uses).  So why not fix the assert then by capping the alignment/misalignment
>>> we compute at this value as well?  (and document this in the header
>>> around DR_MISALIGNMENT)
>>>
>>> Ideally we'd do alignment analysis independent of the vector size
>>> though (for those stupid targets with multiple vector sizes to consider...).
>>>
>>>> a) when looking for references *A1 and *A2 with the same alignment,
>>>>    we simply have to prove that A1 % vecalign == A2 % vecalign.
>>>>    This doesn't require any knowledge about the base alignment.
>>>>    If we break the addresses down as:
>>>>
>>>>       A1 = BASE1 + REST1,  REST1 = INIT1 + OFFSET1 + X * STEP1
>>>>       A2 = BASE2 + REST2,  REST2 = INIT2 + OFFSET2 + X * STEP2
>>>>
>>>>    and can prove that BASE1 == BASE2, the alignment of that base
>>>>    isn't important.  We simply need to prove that REST1 % vecalign
>>>>    == REST2 % vecalign for all X.
>>>>
>>>> b) In the assert, we've peeled the loop so that DR_PEEL is guaranteed
>>>>    to be vector-aligned.  If DR_PEEL is A1 in the example above, we have
>>>>    A1 % vecalign == 0, so A2 % vecalign must be 0 too.  This again doesn't
>>>>    rely on the base alignment being known.
>>>>
>>>> What a high base alignment for DR_PEEL gives us is the ability to know
>>>> at compile how many iterations need to be peeled to make DR_PEEL aligned.
>>>> But the points above apply regardless of whether we know that value at
>>>> compile time or not.
>>>>
>>>> In examples like the test case, we would have known at compile time that
>>>> VF-1 iterations would need to be peeled if we'd picked the store as the
>>>> DR_PEEL, but would have treated the number of peels as variable if we'd
>>>> picked the load.  The value calculated at runtime would still have been
>>>> VF-1, it's just that the code wouldn't have been as efficient.
>>>>
>>>> One of the benefits of pooling the alignments for unconditional references
>>>> is that it no longer matters which DR we pick: the number of peels will
>>>> be a compile-time constant both ways.
>>>>
>>>> Thanks,
>>>> Richard
>>>>
>>>>> I'd rather not mix fixing this with the improvement to eventuall use a
>>>>> larger align for the other DR if possible.
>>>
>>> ^^^
>>>
>>> So can you fix the ICE with capping base alignment / DR_MISALIGNMENT?
>>
>> I don't think the problem is the lack of a cap.  In the test case we
>> see that:
>>
>> 1. B is known at compile time to be X * vecsize + Y when considered in
>>    isolation, because the base alignment derived from its DR_REF >= vecsize.
>>    So DR_MISALIGNMENT (B) == Y.
>>
>> 2. A's misalignment wrt vecsize is not known at compile time when
>>    considered in isolation, because no useful base alignment can be
>>    derived from its DR_REF.  (The DR_REF is to a plain int rather than
>>    to a structure with a high alignment.)  So DR_MISALIGNMENT (A) == -1.
>>
>> 3. A and B when considered as a pair trivially have the same misalignment
>>    wrt vecsize, for the reasons above.
>>
>> Each of these results is individually correct.  The problem is that the
>> assert is conflating two things: it's saying that if we know two datarefs
>> have the same misaligment, we must either be able to calculate a
>> compile-time misalignment for both datarefs in isolation, or we must
>> fail to calculate a compile-time misalignment for both datarefs in
>> isolation.  That isn't true: it's valid to have situations in which the
>> compile-time misalignment is known for one dataref in isolation but not
>> for the other.
>
> True.  So the assert should then become
>
>       gcc_assert (! known_alignment_for_access_p (dr)
>                   || DR_MISALIGNMENT (dr) / dr_size ==
>                     DR_MISALIGNMENT (dr_peel) / dr_peel_size);
>
> ?

I think it would need to be:

      gcc_assert (!known_alignment_for_access_p (dr)
		  || !known_alignment_for_access_p (dr_peel)
		  || (DR_MISALIGNMENT (dr) / dr_size
		      == DR_MISALIGNMENT (dr_peel) / dr_peel_size));

But yeah, that would work too.  The idea with the assert in the patch was
that for unconditional references we probably *do* want to try to compute
the same compile-time misalignment, but for optimisation reasons rather
than correctness.  Maybe that's more properly a gcc_checking_assert
though, since nothing goes wrong if it fails.  So perhaps:

     gcc_checking_assert (DR_IS_CONDITIONAL_IN_STMT (dr)
			  || DR_IS_CONDITIONAL_IN_STMT (dr_peel)
			  || (known_alignment_for_access_p (dr)
			      == known_alignment_for_access_p (dr_peel)));

as a follow-on assert.

Should I split it into two patches, one to change the gcc_assert and
another to add the optimisation?

Thanks,
Richard