[llvm-dev] [RFC] Improving Speculative Compilation in Concurrent ORC JIT
preejackie via llvm-dev
Dear All,
The new concurrent orc jit support
speculative compilation with a naive implementation and
laziness. Compiling everything up-front can consume more CPU and
memory than needed and also impose significant effect on
resource constraint systems due to context switching of JIT
threads (coarse level parallelism). The main idea here is to
develop heuristics to find which functions are worth compiling
into executable bits, speculatively using jit threads.
Heuristics can be extracted by analyzing higher level program representations such as ASTs, CFGs and from LLVM IR. Profile guided optimization can also used to determine heuristics with greater confidence level based on previous execution steps.
Goals:
1) Create & Improve APIs to support heuristic guided speculative compilation in concurrent orc jit.
2) Finding key heuristics to decide what to compile next.
3) Bench marking jit with eagerly speculating full module vs heuristic based approach for different jit threads.
Possible caveats:
1) More time spent in analyzing what to compile next than actually compiling. These can addressed once we have basic working prototype, so that we can find out the bottlenecks and remove it. Or may be parallelizing analysis process.
Currently, I'm trying to find heuristics
that can be used for this project and familiarizing with
concurrent jit design and APIs. Also,
please note that this idea has been proposed by Lang Hames and
I would like to propose this idea & implementation
for GSoC 2019 :) and I'm improvising this idea. I would like to
know the community responses!
This can serve as a good documentation for
me to write a proposal and to keep things in track.
Any thoughts on finding useful heuristics are highly appreciated :) I have also started a thread in Numba JIT project to see what heuristics they are using.
Also anyone interested in mentoring this
project please reply!
References:
Have a great day! PreeJackie
-- Have a great day! PreeJackie
preejackie via llvm-dev
Hi all,
I have come up with
some idea for the speculation task:
1) Construct the
local(per-module) function ordering list based on the sequence
of their call in control flow graphs and create stub for each
symbol from other module that is referenced by the current
function, put the module in waiting list.
2) Replace the stub with actual native code address and notify the modules in the waiting list.
The creation of
function ordering list involves analysing the control flow
graphs and branch probability for conditional function calls.
I'm also trying to figure out whether the function atttributes
will fit in this picture + more like using module summaries from
thinlto builds.
These speculation
actions should be buried into internals of concurrent compiler
and should not be visible to clients right?
How I can proceed with plan, I'm open to comments!
Anyone interested in
mentoring this project for GSoC 2019. Please reply!