[llvm-dev] RFC: a practical mechanism for applying Machine Learning for optimization policies in LLVM

[Mircea Trofin via llvm-dev]

On Thu, Apr 9, 2020 at 10:47 AM Adrian Prantl <aprantl at apple.com> wrote:

>
>
> > On Apr 8, 2020, at 2:04 PM, Mircea Trofin via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
> >
> > TL;DR; We can improve compiler optimizations driven by heuristics by
> replacing those heuristics with machine-learned policies (ML models).
> Policies are trained offline and ship as part of the compiler. Determinism
> is maintained because models are fixed when the compiler is operating in
> production. Fine-tuning or regressions may be handled by incorporating the
> interesting cases in the ML training set, retraining the compiler, and
> redeploying it.
> > For a first milestone, we chose inlining for size (-Oz) on X86-64. We
> were able to train an ML model to produce binaries 1.5-6% smaller than -Oz
> of tip-of-tree. The trained model appears to generalize well over a diverse
> set of binaries. Compile time is increased marginally (under 5%). The model
> also happens to produce slightly better - performing code under SPEC2006 -
> total score improvement by 1.75%. As we only wanted to verify there is no
> significant regression in SPEC, and given the milestone goals, we haven’t
> dug any deeper into the speed results.
> >
>
> How generally applicable are the results? I find it easy to believe that
> if, for example, you train a model on SPEC2006, you will get fantastic
> results when compiling SPEC2006, but does that translate well to other
> inputs?


The current model was trained on an internal corpus of ~25K modules,
obtained from an internal "search" binary. Only ~20 or so are shared with
llvm, for example. In particular, no SPEC modules were used for training.
So the results appear to generalize reasonably well.

Or is the idea to train the compiler on the exact input program to be
> compiled, (perhaps in a CI of sorts), so you end up with a specialized
> compiler for each input program?
>

That's the anti-goal :) The hope is to have reference policies, trained
from a representative corpus, which are 'good enough' for general use (just
like manual heuristics today), while having a systematic methodology for
interested parties to go the extra step and fine-tune (retrain) to their
specifics.


>
> -- adrian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200409/d0488e91/attachment.html>