[PATCH][LAA] Fix estimation of number of memchecks

[Silviu Baranga]

Hi Adam,
 
Yes, actually counting seems like a much better solution, so let's go with
that.
I'll make the interface change, get the benchmark results and upload the
changes to Phab.
 
Thanks,
Silviu
 
 
From: Adam Nemet [mailto:anemet at apple.com] 
Sent: 02 June 2015 20:45
To: Silviu Baranga
Cc: llvm-commits at cs.uiuc.edu
Subject: Re: [PATCH][LAA] Fix estimation of number of memchecks
 
 
On Jun 2, 2015, at 12:02 PM, Adam Nemet <anemet at apple.com> wrote:
 
Hi,
 
The patch looks good but I'd like to see some benchmarking data before fully
signing off on this.  Most importantly, do we vectorize more now?  My
concern is that we may have tuned the threshold with the incorrect formula.
 
Also since we're benchmarking in this area it would be great to completely
move away from using an estimate for the number of memchecks.
 
There is a new API in LAA (RuntimePointerCheck::needsAnyChecking) that uses
the actual number of memchecks needed.  Most notably it excludes checks that
the dependence analysis deems unnecessary (i.e. the DependencySetId check in
needsChecking).
 
We could add a new API returning the actual number instead of estimating
this.
 
Adam
 
 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150603/e1b007b9/attachment.html>