llvm dev - Jan 2016 - LTO Module splitting and metadata

Ladies and Gentlemen,

  I am seeking some clarity in dealing with module splitting and metadata/debug
info.

I know the topic has been debated before, so I am trying to understate the
current state of things and potential development momentum.

What am I trying to do:
  - I am using llvm::SplitModule to break up a large LTO created module prior to
codegen.
    - The hope is to use parallel processing to address time and number of
sections per module.
    - I cannot afford globalizing any local symbols (see
http://reviews.llvm.org/D16124 for details).
    - Time/memory footprint is a big concern here.

What is the issue:
  - When I split an LTO module in presence of Dwarf data (full -g) I have
duplicate symbol mess during linking on metadata objects (llvm.dbg.cu)

Why it happens:
  - The llvm::SplitModule  uses CloneModule with conditional function to select
GlobalValue for each partition. But it does not do it for metadata - it seems to
simply copy _all_ metadata to all partitions.

What I know:
  - ThinLTO had a similar issue, and was forced(?) to use
FunctionImport/linkInModule for lazy picking of functions into a module (I am
sure I am oversimplifying here...)

What I do not know (and seek guidance for)
  - How I best achieve my objective given the above description?

Can I reuse the ThinLTO mechanism without duplicating functionality, or do I
need to augment CloneModule to discriminate on metadata copying?
Which would be more useful in the long run? Maybe there is already a better way
to achieve what I need? Did I get the whole picture wrong?

Any input is very much appreciated.

Thanks.

Sergei



---
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The
Linux Foundation

> -----Original Message-----
> From: Sergei Larin [mailto:slarin at codeaurora.org]
> Sent: Thursday, January 14, 2016 10:56 AM
> To: slarin at codeaurora.org; peter at pcc.me.uk
> Cc: mehdi.amini at apple.com; mehdi_amini at apple.com;
> tobias at codeaurora.org; llvm-commits at lists.llvm.org
> Subject: Re: [PATCH] D16124: Add to the split module utility an SCC based
> method which allows not to globalize any local variables.
> 
> slarin added a comment.
> 
> This would be even better if EquivalenceClasses would guarantee
> deterministic order... but anyhow - this should reflect Peter's
comments.
> 
> The re-globalization patch will likely appear as an independent change
later
> today.
> 
> 
> http://reviews.llvm.org/D16124
> 
>

Hi Sergei,

ThinLTO is a little different in how it behaves because we are doing
the metadata minimization for imported functions by delaying
materialization and linking in of the metadata. In your case, metadata
was materialized long before for the merged module.

There is a thread discussing metadata minimization for bugpoint that
seems like it might be related? See the thread starting at
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20160111/323906.html.

Teresa

On Fri, Jan 15, 2016 at 8:55 AM, Sergei Larin <slarin at codeaurora.org>
wrote:
>
> Ladies and Gentlemen,
>
>   I am seeking some clarity in dealing with module splitting and
metadata/debug info.
>
> I know the topic has been debated before, so I am trying to understate the
current state of things and potential development momentum.
>
> What am I trying to do:
>   - I am using llvm::SplitModule to break up a large LTO created module
prior to codegen.
>     - The hope is to use parallel processing to address time and number of
sections per module.
>     - I cannot afford globalizing any local symbols (see
http://reviews.llvm.org/D16124 for details).
>     - Time/memory footprint is a big concern here.
>
> What is the issue:
>   - When I split an LTO module in presence of Dwarf data (full -g) I have
duplicate symbol mess during linking on metadata objects (llvm.dbg.cu)
>
> Why it happens:
>   - The llvm::SplitModule  uses CloneModule with conditional function to
select GlobalValue for each partition. But it does not do it for metadata - it
seems to simply copy _all_ metadata to all partitions.
>
> What I know:
>   - ThinLTO had a similar issue, and was forced(?) to use
FunctionImport/linkInModule for lazy picking of functions into a module (I am
sure I am oversimplifying here...)
>
> What I do not know (and seek guidance for)
>   - How I best achieve my objective given the above description?
>
> Can I reuse the ThinLTO mechanism without duplicating functionality, or do
I need to augment CloneModule to discriminate on metadata copying?
> Which would be more useful in the long run? Maybe there is already a better
way to achieve what I need? Did I get the whole picture wrong?
>
> Any input is very much appreciated.
>
> Thanks.
>
> Sergei
>
>
>
> ---
> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted
by The Linux Foundation
>
>
>> -----Original Message-----
>> From: Sergei Larin [mailto:slarin at codeaurora.org]
>> Sent: Thursday, January 14, 2016 10:56 AM
>> To: slarin at codeaurora.org; peter at pcc.me.uk
>> Cc: mehdi.amini at apple.com; mehdi_amini at apple.com;
>> tobias at codeaurora.org; llvm-commits at lists.llvm.org
>> Subject: Re: [PATCH] D16124: Add to the split module utility an SCC
based
>> method which allows not to globalize any local variables.
>>
>> slarin added a comment.
>>
>> This would be even better if EquivalenceClasses would guarantee
>> deterministic order... but anyhow - this should reflect Peter's
comments.
>>
>> The re-globalization patch will likely appear as an independent change
later
>> today.
>>
>>
>> http://reviews.llvm.org/D16124
>>
>>
>
>


-- 
Teresa Johnson | Software Engineer | tejohnson at google.com | 408-460-2413