[PATCH 13/25] Create TARGET_DISABLE_CURRENT_VECTOR_SIZE

[Richard Biener]

On Tue, Sep 18, 2018 at 10:22 PM Andrew Stubbs <ams@codesourcery.com> wrote:
>
> On 18/09/18 12:21, Richard Sandiford wrote:
> > Would the same be useful for GCN, or do you basically always
> > want a VF of 64?
>
> Always 64; the vector size varies between 512-bit and 4096-bit, as needed.
>
> > None of this is a fundamental restriction in theory.  It's just
> > something that needs to be fixed.
> >
> > One approach would be to get the loop vectoriser to iterate over the
> > number of lanes the target supports insteaad of all possible vector
> > sizes.  The problem is that on its own this would mean trying 4
> > lane counts even on targets with a single supported vector size.
> > So we'd need to do something a bit smarter...
>
> Yeah, that sounds like an interesting project, but way more than I think
> I need. Basically, we don't need to iterate over anything; there's only
> one option for each mode.
>
> For the purposes of this patch, might it be enough to track down all the
> places that use the current_vector_size and fix them up, somehow?
>
> Obviously, I'm not sure what that means just yet ...

I think the only part that wants a "fixed" size is the code iterating over
vector sizes.  All the rest of the code simply wants to commit to
a specific vector type for each DEF - to match the ISAs we've faced
sofar the approach is simply to choose the vector type of current_vector_size
size and proper element type.

I've long wanted to fix that part in a way to actually commit to vector types
later and compute the DEF vector type of a stmt by looking at the vector
type of the USEs and the operation.

So I guess the current_vector_size thing isn't too hard to get rid of, what
you'd end up with would be using that size when you decide for vector
types for loads (where there are no USEs with vector types, so for example
this would not apply to gathers).

So I'd say you want to refactor get_same_sized_vectype uses and
make the size argument to get_vectype_for_scalar_type_and_size
a hint only.

Richard.

>
> Andrew