[LLVMdev] [RFC] MCJIT usage models

[Larry Gritz]

On Dec 9, 2013, at 3:59 PM, Kevin Modzelewski <kmod at dropbox.com> wrote:

> About lazy compilation, I'm still of the opinion that that's better handled
> outside of MCJIT.  For the people asking for it, would it be enough to have
> a wrapper around MCJIT that automatically splits modules and adds stubs to
> do lazy compilation?


I think that would be sufficient for me.

The case I care about is that there is large body of functions that the dynamic code *could* call, but any particular JITed program will only call a tiny fraction of that.  I pre-compile this library to bitcode at build time, and seed the module with it.

If I call llvm::ParseBitcodeFile to pull it in, I incur about a 0.25s overhead.  If I use llvm::getLazyBitcodeModule, it's less than 0.01s for small programs, which is fast enough to compile/recompile these functions frequently in an interactive application.  I'd very much like to see this kind of gain for MCJIT.

I don't care if it's literally part of MCJIT or provided as a wrapper. Either is fine, as long as I don't have to write it myself (mostly because I don't trust myself to not make it subtly buggy, compared to somebody who deeply knows MCJIT internals).


--
Larry Gritz
lg at larrygritz.com