Make aliasing_component_refs_p to work harder when same_type_for_tbaa returns -1

[Richard Sandiford]

Richard Biener <rguenther@suse.de> writes:
> On Thu, 6 Jun 2019, Jan Hubicka wrote:
>
>> > > This is more of an accident and there are cases where we do not trip across
>> > > this -1 and we disambiguate array accesses that seems unsafe to me.
>> > > 
>> > > With my change aliasing_component_refs_p finds the match of the
>> > > array types (type_same_for_tbaa_p returns 1 with non-LTO becuase they
>> > > have same canonical type) and disambiguates based on disjoint access ranges.
>> > > 
>> > > I have thus went ahead and updated all uses of type_same_for_tbaa_p to special
>> > > case arrays and reffer to this testcase (which seems odd and is only one in
>> > > testsuite): We can still do useful disambiguation if the array is not toplevel
>> > > reference or we know that the memory object is not bigger. This is tested by a
>> > > testcase I added and is quite frequent in real world code.
>> > > 
>> > > I also added check to give up on VLAs since I can not convicne myself that
>> > > this is safe: I think early inlining VLAs and streaming them may lead to
>> > > same VLA type have two different sizes at a time enabling it to partially
>> > > overlap.
>> > 
>> > OK - sorry for the delay.  The array stuff gets a bit ugly so
>> > we eventually want to do sth about that ...
>> 
>> Thanks, no prob with the delay - I apprechiate we could discuss these
>> things carefully as they are by no means obvious :)
>> 
>> I agree that the way overlapping arrays support is done current is ugly
>> and seems incomplete.  I hope to clean it up now and craft more
>> testcases.  It seems bit odd decision to suport the partial overlaps as
>> done by alias-2.c. I wonder how important it is in practice.
>
> Probably not very, likely appears in the context of VLAs or
> multi-dimensional arrays.  IIRC at some point I tried to
> understand what the C standard says here but I don't remember
> the outcome ;)  Then there's also ARRAY_RANGE_REF ...

Dropping support for that would also allow cheaper runtime alias checks
in the vectoriser for some cases, via the DDR_COULD_BE_INDEPENDENT_P stuff.

Richard