[llvm-dev] [RFC] Framework for Finding and Using Similarity at the IR Level

[Andrew Litteken via llvm-dev]

Hi Yvan,

Glad to hear you’re interested! I’m not sure which ordering you’re looking at the patches, but a cost model was not added to the outliner until D87299. So, if that hasn’t been added I would expect some pretty big code increases since it just outlines whatever similarity it finds first otherwise.

The cost model has also been slightly reworked since the initial results to be more general, and try to use target hooks, but not all of the target hooks that we would really want exist. So, if it has been applied, I’ll definitely need to take a look.

Andrew

> On Sep 30, 2020, at 6:22 AM, Yvan Roux <yvan.roux at linaro.org> wrote:
> 
> Hi Andrew,
> 
> Thanks for this great work and presentation, I've also a huge interest in code
> size reduction.  I already looked at the first IR Outliner proposal,
> and I really
> like your more general approach.
> 
> I wanted to look at the impact on ARM targets, and managed to apply
> your patchset
> up to D87311 (the remaining ones which are related to call instructions and GEP
> need to be reworked to be applied or work correctly). I looked at CTMark code
> size other -Oz for both ARM and X86 and got a huge code size increase on both
> targets (around 30% or 40% in Geomean) when compiled with -mllvm -ir-outliner.
> 
> Does it rings some bells on your side, is it expected given that the full series
> is not applied, or did I just miss something ? From what I can observe, and as
> you mentioned in your first mail, a lot of code is introduced to handle
> parameters passing to outlined functions and the biggest code size increase was
> observed after applying D87296.
> 
> Thanks a lot for this work and any hint on how to reproduce your results, I'm
> eager to play with it and give some help if needed.
> 
> Cheers,
> Yvan
> 
> 
>> On Thu, 10 Sep 2020 at 00:00, Andrew Litteken via llvm-dev
>> <llvm-dev at lists.llvm.org> wrote:
>> 
>> Thanks for the positive feedback!
>> 
>> All the patches are up now:
>> 
>> Elevating different constants to arguments: [IRSim][IROutliner] Adding support for consolidating functions with different output arguments.
>> Extracting require outputs: [IRSim][IROutliner] Adding support for consolidating functions with different output arguments.
>> Merging Output Blocks: [IRSim][IROutliner] Merging output blocks for extracted functions with outputs
>> Cost Model: [IRSim][IROutliner] Adding a cost model, and debug option to turn the model off.
>> OptRemarks: [IRSim][IROutliner] Adding OptRemarks for the IROutliner.
>> Function Attribute Merging: [IRSim][IROutliner] Adding consistent function attribute merging
>> DebugInfo Handling: [IRSim][IROutliner] Adding DebugInfo handling for IR outlined functions.
>> Outlining from linkonceodr: [IRSim][IROutliner] Adding option to enable outlining from linkonceodr functions
>> Flexibility for Isomorphic Predicates: [IRSim] Adding support for isomorphic predicates
>> Flexibility for Commutative Instructions: [IRSim] Adding commutativity matching to structure checking
>> Matching call instructions for similarity identification: [IRSim] Letting call instructions be legal for similarity identification.
>> Outlining Calls: [IRSim][IROutliner] Allowing call instructions to be outlined.
>> Matching GEP instructions for similarity identification: [IRSim] Letting gep instructions be legal for similarity identification.
>> Outlining GEP instructions: [IRSim][IROutliner] Allowing GEP instructions to be outlined.
>> 
>> Feedback, comments and reviews are welcome!
>> 
>> Andrew
>> 
>> On Sep 2, 2020, at 5:04 AM, Renato Golin <rengolin at gmail.com> wrote:
>> 
>> Indeed, an awesome project and an excellent report!
>> 
>> Code size doesn't really get much attention, so the level of detail and the strong roadmap is refreshing.
>> 
>>> Hopefully, the project will provide execution times along with code-size reductions.
>> 
>> I doubt it.
>> 
>> Outlining will (almost?) always make for slower code due to a lot more calls being made. But that's ok for embedded targets, that don't have strong speculative machinery.
>> 
>> Also, if the program fits versus doesn't, then it will run infinitely faster. :D
>> 
>> Kristof, do you know who at Arm would be good to review those patches?
>> 
>> cheers,
>> --renato
>> 
>> 
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev