[LLVMdev] GSOC Adaptive Compilation Framework for LLVM JIT Compiler

[Xin Tong Utoronto]

On Tue, Mar 29, 2011 at 4:45 PM, Eric Christopher <echristo at apple.com>wrote:

> >
> > Project Outline:
> >
> >
> >
> > Currently, the LLVM JIT serves as a management layer for the executed
> LLVM IR, it manages the compiled code and calls the LLVM code generator to
> do the real work. There are levels of optimizations for the LLVM code
> generator, and depends on how much optimizations the code generator is asked
> to do, the time taken may vary significantly. The adaptive compilation
> mechanism should be able to detect when a method is getting hot, compiling
> or recompiling it using the appropriate optimization levels. Moreover, this
> should happen transparently to the running application. In order to keep
> track of how many times a JITed function is called. This involves inserting
> instrumentational code into the function's LLVM bitcode before it is sent to
> the code generator. This code will increment a counter when the function is
> called. And when the counter reaches a threshold, the function gives control
> back to the LLVM JIT. Then the JIT will look at the hotness of all the
> methods and find the one that triggered the recompilation threshold. The JIT
> can then choose to raise the level of optimization based on the algorithm
> below or some other algorithms developed later.
> >
> >
> > IF (getCompilationCount(method) > 50 in the last 100 samples) = >
> Recompile at Aggressive
> > ELSE Recompile at the next optimization level.
> >
> >
> > Even though the invocation counting introduces a few lines of binary, but
> the advantages of adaptive optimization should far overweigh the extra few
> lines of binary introduced. Note the adaptive compilation framework I
> propose here is orthogonal to the LLVM profile-guided optimizations. The
> profile-guided optimization is a technique used to optimize code with some
> profiling or external information. But the adaptive compilation framework is
> concerned with level of optimizations instead of how the optimizations are
> to be performed.
> >
>
> So, one way that current projects use the JIT is via getPointerToFunction()
> which returns an address that can then be casted and called with the
> appropriate arguments. The compile task itself is often done on a separate
> thread. How would you deal with the updating problem in the calling
> application? What sort of use cases for the JIT have you looked at so far?
>

I assume the updating problem means the problem when a method gets
recompiled. Here is an algorithm to deal with that. Say A calls B. when B
gets recompiled we patch B with *br helper* at the beginning of its code,
then when A calls B, B branches to the helper and the helper patches the *br
B* in A with *br newB*. as we don't know all the callers of B, we have to
wait until they call B to know who they are and patch them one-by-one. The
helper can get the address of the *br B* in A from the link register or some
specific registers or memory locations. For newly compiled code, the address
of the newB can be used. There is another problem with recompilation.
obsolete methods(methods that have recompiled copies) need to be recycled.
In order to do that, we will need to keep a *br helper* in place of the old
method and reclaim the old method body.

As for use case, the LLVM JIT is used as an execution engine for a few
number of ported languages, for example JIT compiler for PHP, in 2008
GSOC.  There
are also people using LLVM JIT for industry work,
https://llvm.org/svn/llvm-project/www-pubs/trunk/2010-01-Wennborg-Thesis.pdf
. As LLVM is growing more and more powerful, LLVM JIT will become more and
more attractive to language designer and implementer. And I think that is
one of the most important reasons we need to have an adaptive compilation
framework. This framework can also work together with the LLVM
profile-guided optimizations to make LLVM JIT a much faster execution
engine.
>
>
> >
> > This is a relatively small project and does not involve a lot of coding,
> but good portion of the time will be spent benchmarking, tuning and
> experimenting with different algorithms, i.e. what would be the algorithm to
> raise the compilation level when a method recompilation threshold is
> reached, can we make this algorithm adaptive too, etc. Therefore, my
> timeline for the project is as follow
> >
> >
> > Week 1
> > Benchmarking the current LLVM JIT compiler, measuring compilation speed
> differences for different levels of compilation. This piece of information
> is required to understand why a heuristic will outperform others
> >
> > Week 10 - 13
> > Benchmarking, tuning and experimenting with different recompilation
> algorithms. Typically benchmarking test cases would be
> >
>
> What do you have in mind for benchmarking? Which of the jitted problems
> were you looking at, or just running large programs through lli and that
> interface? (Which isn't threaded and therefore doesn't have the problems I
> mentioned above - it has other problems).
>

Widely known benchmarks, such as SPEC CPU, would be good candidates.  In
addition to these benchmarks, we may want to introduce some specific tests
for Just-In-Time compilers, like ones with a small portions of the methods
taking up 80%+ of the time and ones with all the methods spend about the
same amount of time and ones in the middle of the two.

>
> >
> > Week 14
> > Test and organize code. Documentation
> >
>
> As a general note all of these things would need to be done during the
> project along with incremental changes made to the repository (on a branch
> if possible).
>
> > Overall Goals:
> >
> >
> > My main goal at the end of the summer is to have an automated profiling
> and adaptive compilation framework for the LLVM. Even though the performance
> improvements are still unclear at this point, I believe that this adaptive
> compilation framework will definitely give noticeable performance benefits,
> as the current JIT compilation is either too simple to give a reasonably
> fast code or too expensive to apply to all functions.
>
> My comments above aside, I think this is a great idea for a project. It is
> aggressive so the amount of time you put in will likely be larger than a
> scaled back project.


>From the questions you asked, I now understand why this project might take
more time than I originally anticipated. Thank You.

- Xin

>


> -eric




-- 
Kind Regards

Xin Tong
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20110331/11bf1efc/attachment.html>