[PATCH] Fix write-back value propagation for pre-indexed addressing modes

[Hal Finkel]

----- Original Message -----
> From: "Silviu Baranga" <Silviu.Baranga at arm.com>
> To: "Hal Finkel" <hfinkel at anl.gov>
> Cc: llvm-commits at cs.uiuc.edu, "Dmitry Antipov" <antipov at dev.rtsoft.ru>
> Sent: Friday, April 26, 2013 5:21:35 AM
> Subject: RE: [PATCH] Fix write-back value propagation for pre-indexed addressing modes
> 
> > -----Original Message-----
> > From: Hal Finkel [mailto:hfinkel at anl.gov]
> > Sent: 25 April 2013 20:58
> > To: Silviu Baranga
> > Cc: llvm-commits at cs.uiuc.edu; Dmitry Antipov
> > Subject: Re: [PATCH] Fix write-back value propagation for
> > pre-indexed
> > addressing modes
> 
> 
> > I understand; we're kind of talking past each other at this point.
> > In
> > any case, looking at the ARM implementation, I disagree with
> > Dmitry's
> > interpretation of the problem. As I wrote the code in question,
> > please
> > feel free to commit your fix.
> 
> Sorry about that. The problem is reasonably complex and we both seem
> to
> have strong divergent views about it.

It is not your fault, and we're on the same page. I had been trying to construct an explanation based on Dimitry's diagnosis of the problem, which essentially included the possibility that PRE_DEC on ARM was essentially a PRE_INC with a negative offset. But looking at the code myself, I think that he was misinterpreting things.

> 
> > I would really like, however, if we had a test case for this. If
> > you
> > had code that the current implementation miscompiled, can you place
> > an
> > assert that triggers in the bad case (pre_dec + swap), and use
> > bugpoint
> > to reduce the test case?
> 
> I've only seen the pre_dec + swap failure in a user of the LLVM
> libraries,
> and I cannot reproduce the failure with llc (I've really tried).
> Also, the
> failure goes away with even small modifications of the test case.
> 
> I would actually like to include Dmitry's test case if he's OK with
> that
> so at least the patch does gets a test case.

Okay. FWIW, Dmitry did not respond to my last e-mail from last week, so if he does not respond promptly, I'm in favor of committing the fix and then the test case later.

 -Hal 

> 
> Thanks,
> Silviu
> 
> 
> -- IMPORTANT NOTICE: The contents of this email and any attachments
> are confidential and may also be privileged. If you are not the
> intended recipient, please notify the sender immediately and do not
> disclose the contents to any other person, use it for any purpose,
> or store or copy the information in any medium.  Thank you.
>