[llvm-commits] [PATCH] Add Pattern ID Information

[Rotem, Nadav]

Hi Bruno, 

You are right. I meant that the code which lowers generic shuffles (with cross-lane access) to AVX256 is missing. Due to the restrictive nature of pshufs on avx256, we need to manually lower these shuffles to pshufs + permute or even vinsertf128 + shuf + permute. The linux calling convention is implemented, not the win64 one. 

Thanks, 
Nadav



-----Original Message-----
From: Bruno Cardoso Lopes [mailto:bruno.cardoso at gmail.com] 
Sent: Thursday, August 04, 2011 10:41
To: Rotem, Nadav
Cc: Chris Lattner; David Greene; llvm-commits at cs.uiuc.edu
Subject: Re: [llvm-commits] [PATCH] Add Pattern ID Information

Nadav,

> None of the AVX shuffle code is in place, the broadcast instruction is not implemented, the avx calling conventions are not implemented, etc. I believe that there is still a ton of work.

That's not true. A important part of the shuffle code is in place,
which doesn't mean it's perfect and optimized, but it works. If by
calling convention you mean the x86_64 abi for 256-bit vectors, it's
already implemented in clang. The broadcast instruction should land
soon, as this is obviously a WIP. There isn't any objection to
contribution here, in the same way that there's no need for any
heavy-duty AVX merging.

-- 
Bruno Cardoso Lopes
http://www.brunocardoso.cc
---------------------------------------------------------------------
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.