[Polly] Generating code for changed memory accesses

[Tobias Grosser]

On 29/07/2014 10:03, 'Johannes Doerfert' wrote:
>> I don't see why removing the tuple name should be needed. The tuple names
>> should match. If they don't either I misunderstood something or you use them
>> in the wrong way.
> That's because I did something else (and had still the same idea in mind).
> With the reverse relation and pullback we don't need that anymore.
>
>> Sorry, I forgot to reverse the schedule map: Here the full flow for your
> That helps me, and the fact that I now understand the *_pullback_* functions.
>
> My first approach used the Stmt scattering which I applied with
> isl_map_apply_domain, but thats different from what _pullback_ does.
>
> I will come up with new patch (hopefully) tomorrow and integrate also your other
> comments. However, I would like to enable this kind of codegen for all access
> relations (not only imported ones) as long as delinearization is off and the
> memory access is not "non-affine/overaproximated".
> Is that OK with you or do you disagree?

Let's first get it in and tested for the easy ones. We can then discuss 
the switch.

This may in fact be a good idea, but I would like to understand a little 
bit better the motivation.

Cheers,
Tobias