[llvm-dev] [RFC] Profile guided section layout

[Xinliang David Li via llvm-dev]

On Thu, Jun 15, 2017 at 2:39 PM, Sean Silva <chisophugis at gmail.com> wrote:

>
>
> On Thu, Jun 15, 2017 at 2:33 PM, Xinliang David Li <xinliangli at gmail.com>
> wrote:
>
>>
>>
>> On Thu, Jun 15, 2017 at 2:30 PM, Sean Silva <chisophugis at gmail.com>
>> wrote:
>>
>>>
>>>
>>> On Thu, Jun 15, 2017 at 11:09 AM, Xinliang David Li via llvm-dev <
>>> llvm-dev at lists.llvm.org> wrote:
>>>
>>>>
>>>>
>>>> On Thu, Jun 15, 2017 at 10:55 AM, Michael Spencer via llvm-dev <
>>>> llvm-dev at lists.llvm.org> wrote:
>>>>
>>>>> On Thu, Jun 15, 2017 at 10:08 AM, Tobias Edler von Koch <
>>>>> tobias at codeaurora.org> wrote:
>>>>>
>>>>>> Hi Michael,
>>>>>>
>>>>>> This is cool stuff, thanks for sharing!
>>>>>>
>>>>>> On 06/15/2017 11:51 AM, Michael Spencer via llvm-dev wrote:
>>>>>>
>>>>>>> The first is a new llvm pass which uses branch frequency info to get
>>>>>>> counts for each call instruction and then adds a module flags metatdata
>>>>>>> table of function -> function edges along with their counts.
>>>>>>>
>>>>>>> The second takes the module flags metadata and writes it into a
>>>>>>> .note.llvm.callgraph section in the object file. This currently just dumps
>>>>>>> it as text, but could save space by reusing the string table.
>>>>>>>
>>>>>> Have you considered reading the profile in the linker and extracting
>>>>>> that information directly from the profile? The profile should contain call
>>>>>> sites and their sample counts and you could match these up with relocations
>>>>>> (calls) in the section?
>>>>>
>>>>>
>>>> The main reason is that IPO transformations such as inlining and
>>>> clonining will change the hotness of functions, so the original profile can
>>>> not be directly for the purpose of function layout.   There is a similar
>>>> support in Gold plugin for Google GCC.
>>>>
>>>
>>> Will this cause issues with ThinLTO? E.g. the thinlto backends are doing
>>> inlining of imported functions. Do we have a mechanism for those decisions
>>> to be reflected in a global profile for the linker to look at?
>>>
>>> In theory the thinlto backends can keep a history of their IPO decisions
>>> in metadata or something and emit a section for the linker to aggregate and
>>> reconstruct an accurate global profile, but that seems relatively invasive.
>>>
>>
>>
>> Yes, it will cause problems which is also known to GCC's LIPO. We have an
>> intern working on that problem :)
>>
>
> Nice! What approach is being used to solve it?
>
>

The method and results will be shared at some point. One approach is what
you mentioned which uses meta data and pass them to linker plugin. Another
way to be explored is to let thin link phase to make a very quick cross
module inlinining decisions globally and teach the backend inliner to honor
the results .  The global inline decisions are not intended to replace the
backend inline analysis. The global analysis can for instance, 1) expose
linker GC opportunities; 2) can focus on hot functions with few callsites
or very few really hot callsites; 3) identify a single module that has the
most # of calls to a function and assign that module to be be final info
updater; 4) decide the post-link hotness for those functions whose
decisions are made at thin-link time.

David



> -- Sean Silva
>
>
>>
>> David
>>
>>
>>>
>>> -- Sean Silva
>>>
>>>
>>>>
>>>> David
>>>>
>>>>
>>>>
>>>>
>>>>>
>>>>> I did this using IR PGO instead of sample PGO so the profile data can
>>>>> only be applied in the same place in the pipeline it is generated. Even for
>>>>> sample based this would be complicated as the linker would actually need to
>>>>> generate machine basic blocks from sections to be able to accurately match
>>>>> sample counts to relocations, as there may be cold calls in hot functions.
>>>>>
>>>>> It may be useful however for the linker to directly accept an
>>>>> externally generated call graph profile. The current approach can actually
>>>>> do this by embedding it into an extra object file.
>>>>>
>>>>>
>>>>>>
>>>>>>
>>>>>> It doesn't currently work for LTO as the llvm pass needs to be run
>>>>>>> after all inlining decisions have been made and LTO codegen has to be done
>>>>>>> with -ffunction-sections.
>>>>>>>
>>>>>> So this is just an implementation issue, right? You can make LTO run
>>>>>> with -ffunction-sections (by setting TargetOptions.FunctionSections=true)
>>>>>> and insert your pass in the appropriate place in the pipeline.
>>>>>>
>>>>>
>>>>> Yeah, just an implementation issue. Just need to build the pass
>>>>> pipeline differently for LTO and add a way to do -ffunction-sections in lld.
>>>>>
>>>>> - Michael Spencer
>>>>>
>>>>>
>>>>>>
>>>>>> Thanks,
>>>>>> Tobias
>>>>>>
>>>>>> --
>>>>>> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
>>>>>> a Linux Foundation Collaborative Project.
>>>>>>
>>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> LLVM Developers mailing list
>>>>> llvm-dev at lists.llvm.org
>>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>>>
>>>>>
>>>>
>>>> _______________________________________________
>>>> LLVM Developers mailing list
>>>> llvm-dev at lists.llvm.org
>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170615/f62cc25a/attachment-0001.html>