[PATCH] Global Structure Vectorization

[Hal Finkel]

----- Original Message -----
> From: "Arnold Schwaighofer" <aschwaighofer at apple.com>
> To: "Nadav Rotem" <nrotem at apple.com>
> Cc: "Renato Golin" <renato.golin at linaro.org>, "LLVM Commits" <llvm-commits at cs.uiuc.edu>, "Hal Finkel"
> <hfinkel at anl.gov>
> Sent: Saturday, February 23, 2013 11:55:52 AM
> Subject: Re: [PATCH] Global Structure Vectorization
> 
> Hi all,
> 
> On Feb 20, 2013, at 1:18 PM, Nadav Rotem <nrotem at apple.com> wrote:
> 
> > Hi Renato,
> > 
> > Thanks for working on this. The code looks correct. I agree with
> > Arnold that this function is becoming even more complicated and it
> > would be nice to refactor it somehow, but we can do it later. I
> > don't mind if the tests are in one file or multiple files.  I
> > assume that you ran the test-suite and found no failures. Did you
> > see any newly vectorized loops ?
> 
> I just did a quick analysis of how many more loops we handle due to
> this patch. This is running over test-suite + externals (specs,
> povray) on x86-64/darwin. I also have a modified copy where we
> always check alias analysis (i.e. also when the identified object is
> not a global variable). Here are the numbers:
> 
> no aa vectorization: 4249
> current trunk: 4277 (delta 28)
> always aa vec: 4320 (delta 43)

Is there any reason not to make the "always aa vec" default? (or at least add an option to enable it)?

 -Hal

> 
> What is missing of course is a qualitative analysis: do any of those
> matter?
> 
> 
> Here are some benchmarks that had (more) vectorized loops due to
> variant "always aa vec":
> 
> Povray
> bzip2
> gcc
> gobmk
> h264ref
> ClamAV
> JM/ldecod
> JM/lencod
> bzip2
> 7zip
> gsm/toast
> mpeg2/mpeg2dec
> MiBench/consumer-lame
> MiBench/telecomm-gsm
> ReedSolomon
> 
> 
> 
> Best,
> Arnold
>