[llvm-dev] [RFC] dereferenceable metadata

[Dinar Temirbulatov via llvm-dev]

>it's already implicit that it is dereferenceable for the size of the access (otherwise it would be undefined behavior).
The values we are loading after restoring those loads are not matter
for us and it is just that chain of loads for vectorization that
important for us. And those load would never get rematerialization if
vectorization is not profitable for the shuffle operation. Once we
remove the load in instcombine and there is no way to know that memory
is dereferenceable from the previous load in the chain of loads.
                                Thanks, Dinar.

On Thu, Jul 20, 2017 at 4:07 AM, Filipe Cabecinhas <me at filcab.net> wrote:
> Indeed. But the problem here is that Dinar is trying to keep information
> after a load/store is removed by instcombine
>
> For example:
>
> v4sf v = {p[0], p[1], p[2], p[3]};
> v4sf v2 = shuffle(v, 0, 0, 2, 2);
>
> Some pass comes in and removes the p[3] and p[1].
>
> Now you have smaller code, but lost the ability to use a vector load for all
> those values + shuffle. The code got scalarized because we lost the
> information that p[3] is valid.
>
> The attribute on a value/load/something might not be ideal (especially since
> we'd be changing other loads (ones we didn't remove)). But I think Dinar
> wanted to start a conversation about how we can keep this information around
> even after we delete some loads.
>
> Thank you,
>
>  Filipe
>
> On Wed, 19 Jul 2017 at 12:53, Nuno Lopes via llvm-dev
> <llvm-dev at lists.llvm.org> wrote:
>>
>> When a pointer is passed to load/store, it's already implicit that it is
>> dereferenceable for the size of the access (otherwise it would be
>> undefined
>> behavior).
>> Isn't that information sufficient for your use case?  (sorry, didn't read
>> the whole thread carefully).
>>
>> Nuno
>>
>> -----Original Message-----
>> From: Dinar Temirbulatov via llvm-dev
>> Sent: Tuesday, July 18, 2017 7:36 PM
>> Subject: [llvm-dev] [RFC] dereferenceable metadata
>>
>> Hi,
>> While working on PR21780, I used "dereferenceable_or_null" metadata
>> and I realized now that it is not correct for my solution to use this
>> metadata type since it might point to an address that it is not
>> dereferenceable but null.  I think that we need another new metadata
>> type, something  like  "dereferenceable" with that we could annotate
>> any load (not just pointer type like with dereferenceable_or_null).
>> For example, we could annotate this load : "%ld0 = load double,
>> double* %ptr, align 8" with dereferenceable<2> and that means that for
>> %ptr address memory with length 16-bytes  is known to be
>> accessible(dereferenceable). Originally in PR21780, InstCombine pass
>> removed some loads as dead(in a Scalar IR form) and later that
>> prevented us to vectorize the operation, but before removing such
>> loads we could annotate the remaining loads in the series with
>> "dereferenceable" and later restore IR in a way that is suitable for
>> vectorization(see https://reviews.llvm.org/D35139 ). Also, please note
>> that’s the information that is lost when InstCombine kills the last
>> load in the series and there is no way to restore this information
>> later in following passes.
>>                                          Thanks, Dinar.
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev