[PATCH] [AArch64] Improve and enable the SeparateConstOffsetFromGEP for AArch64 backend.

[Hao Liu]

Hi James,

 

From: mankeyrabbit at gmail.com [mailto:mankeyrabbit at gmail.com] On Behalf Of James Molloy
Sent: Monday, October 27, 2014 11:38 PM
To: reviews+D5864+public+2de5fd757ea4c578 at reviews.llvm.org
Cc: Hao Liu; Tim Northover; ssijaric at codeaurora.org; Chad Rosier; Jingyue Wu; LLVM Commits
Subject: Re: [PATCH] [AArch64] Improve and enable the SeparateConstOffsetFromGEP for AArch64 backend.

 

Hi Hao,

 

I'm not sure about this. Firstly I have concerns about the statistical significance of the difference in results - we've seen swings of +/- 1% on A57, and here we're talking about a 0.5% difference. Are you satisfied that the difference is genuine?

 

[Hao Liu] I’m also curious about the result. I suspect both versions should have not big difference. But I think the difference is genuine. The benchmark astar is quite stable and should never have variance about +/- 1%.

Secondly, I don't see a good reason why useAA should be worse than !useAA. To my mind, if there is a regression when useAA is enabled, it should be one of two things:

  1) A legitimate regression, in which case let's fix useAA!

  2) Random code permutation outside of the compiler's model. If this is the case we shouldn't be adjusting the compiler to compensate.

 

useAA has the potential to improve codegen. We're not really using it at the moment, but it's good that it's available to use in the future. If we rip it out, we should have a good reason for it, I think!

 

Can I suggest that you make the initial commit, which improves performance, then have a look at why !useAA is better and fix that incrementally?

 

[Hao Liu] OK, I’d like to do that. Thanks.

Cheers,

 

James
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20141028/6208e4f0/attachment.html>