llvm patch

[Hal Finkel]

----- Original Message -----
> From: "Jan Hranac" <hranac at codasip.com>
> To: "Matt Arsenault" <arsenm2 at gmail.com>
> Cc: llvm-commits at cs.uiuc.edu
> Sent: Monday, July 27, 2015 4:54:26 AM
> Subject: Re: llvm patch
> 
> Thank you for your reply.
> The problem is that the original testcase (which includes an indirect
> specification of an architecture) is an intelectual property of
> Codasip
> (codasip.com) and of our customers. Hence, I'm rather bound by
> various
> NDAs.
> However, I can tell you that it occurred when there was a legal type
> v2i16 (with legal load/store operations) in an architecture. I'm
> including an edited fragment of a test application (please do not
> publish this). The loop needs to be vectorized in middle-end. The
> arrays
> will keep their aligment=2 which will lead to a call of
> ExpandUnalignedLoad on a v2i16 load.
> I apologize for the complication and thank you for your patience.

Understood; a useful test case would need to involve IR for an in-tree target regardless. I agree, however, that this looks like an obviously-correct fix. r243956.

 -Hal

> 
> On 25/07/15 18:35, Matt Arsenault wrote:
> >> On Jul 23, 2015, at 10:51 PM, Jan Hranac <hranac at codasip.com>
> >> wrote:
> >>
> >> The attached patch fixes loss of chaining in ExpandUnalignedLoad
> >> for legal (but unaligned) vectors and floats. The function was
> >> returning an input chain, not an output one in that case.
> >> <patch.txt>_______________________________________________
> >> llvm-commits mailing list
> >> llvm-commits at cs.uiuc.edu
> >> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> > LGTM. Do you have a testcase you can add which breaks from this?
> 
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> 

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory