[PATCH] [OPENMP] llvm.loop.vectorize.enable metadata are lost after critical edge splitting

[Hal Finkel]

----- Original Message -----
> From: "Andrew Trick" <atrick at apple.com>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: reviews+D5539+public+7680b8f2f557d7c9 at reviews.llvm.org, "llvm commits" <llvm-commits at cs.uiuc.edu>, "zinovy nis"
> <zinovy.nis at gmail.com>, chandlerc at gmail.com
> Sent: Tuesday, September 30, 2014 7:54:06 PM
> Subject: Re: [PATCH] [OPENMP] llvm.loop.vectorize.enable metadata are lost after critical edge splitting
> 
> On Sep 30, 2014, at 2:57 PM, Hal Finkel < hfinkel at anl.gov > wrote:
> 
> I believe this was a known issue when we discussed the design; if we
> can come up with a better design, that'd (by definition) be great.
> The core issue is just how to associate metadata with a loop.
> 
> 
> This is clearly out-of-scope, but for the record I don’t think it’s
> hard to find the branch that the loop latch is control dependent on,
> and the loop metadata could just always reside there. That way, the
> metadata would only be invalidated if the loop test itself changed.
> I would do it by looking for a loop exit block that dominates the
> (unconditional) latch and doesn’t dominate any other loop exits.

Can you start a thread on llvm-dev about this? I'd like to discuss this (but obviously not in this thread) with more people listening.

 -Hal

> 
> 
> -Andy

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