llvm dev - May 2014 - [LLVMdev] LLVM JIT and PGO in JIT

Hi,

I haven't touched LLVM codes for many years. I have a few questions related
to LLVM JIT and PGO in JIT:
1) Is LLVM JIT method-based or trace-based JIT?
2) How does LLVM do gear shift from the interpreter mode to the JIT mode?
3) Any profiling mechanism implemented in LLVM JIT?
4) Any profile-guided optimization implemented in LLVM JIT?

thanks
Qiong
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140527/d6eb7e39/attachment.html>

On 05/27/2014 02:36 PM, Qiong Cai wrote:
> Hi,
>
> I haven't touched LLVM codes for many years. I have a few questions 
> related to LLVM JIT and PGO in JIT:
> 1) Is LLVM JIT method-based or trace-based JIT?
It could be used to implement either.  Method-based is the most natural 
usage though.
> 2) How does LLVM do gear shift from the interpreter mode to the JIT mode?
It doesn't.  LLVM provides the framework for building the JIT; you have 
to provide the runtime, interpreter, and compilation
policy.
> 3) Any profiling mechanism implemented in LLVM JIT?
There's some limited support for instrumentation passes, but I don't 
believe it's particularly solid.  I haven't looked at this area in any 
detail though.
> 4) Any profile-guided optimization implemented in LLVM JIT?
Yes.  At the moment, they're fairly limited, but there's a fair amount 
of work happening in this area over the last few months. Contributions 
welcome.

Philip