[LLVMdev] Code review for gather and scatter intrinsics

[Renato Golin]

I see. I probably got it wrong, then.

I was under the impression that we wouldn't do stride.load/store or
gather/scatter directly, but implement both of them via
index.load/store and differentiate gather/scatter with the argument as
a vector of pointers, not a pointer of vectors.

I also thought that we'd use the mask to complement the indexed load,
which would in turn help the vectorizer to create prologues/epilogues
and the back end to generate the best instruction.

cheers,
--renato

On 16 April 2015 at 20:17, Demikhovsky, Elena
<elena.demikhovsky at intel.com> wrote:
> Hi Renato,
>
> I fully agree with you, but indexed load and store is the next step.
> I'm asking to review gather and scatter code.
>
> Thanks.
>
> -  Elena
>
> -----Original Message-----
> From: Renato Golin [mailto:renato.golin at linaro.org]
> Sent: Thursday, April 16, 2015 17:17
> To: Demikhovsky, Elena
> Cc: llvmdev at cs.uiuc.edu; Chandler Carruth; James Molloy
> Subject: Re: [LLVMdev] Code review for gather and scatter intrinsics
>
> On 16 April 2015 at 14:44, Demikhovsky, Elena <elena.demikhovsky at intel.com> wrote:
>> http://reviews.llvm.org/D7665.
>> I presented this work on LLVM Euro and people are interested in this
>> feature.
>> Can anybody review this code, please?
>
> Hi Elena,
>
> Sorry for the delay, I'm still catching up with my emails after a long holiday. :)
>
> The only concern to this feature I remember was Chandler's comment that we should try to encode everything into loads and shuffles.
>
> Correct me if I'm wrong, but on the strided vectorizer thread we have reached a consensus that indexed intrinsics would be the least problematic (compared to strided access intrinsics or plain
> load+shuffle) because they're restricted to load/store of patterns
> that could later be lowered to shuffles, if the hardware doesn't support it, but they'd also keep the pattern intact even after other optimization passes have gone through the same code.
>
> Chandler,
>
> As I said at EuroLLVM, the fear is that we'd get the pattern destroyed and lose the ability of using masked / strided access at all. However, I'm haven't looked at great depth at this problem to know what are the cases and why they could break. Maybe Elena or James could help with that.
>
> cheers,
> --renato
> ---------------------------------------------------------------------
> Intel Israel (74) Limited
>
> This e-mail and any attachments may contain confidential material for
> the sole use of the intended recipient(s). Any review or distribution
> by others is strictly prohibited. If you are not the intended
> recipient, please contact the sender and delete all copies.