[LLVMdev] An enhancement for MCJIT::getFunctionAddress

[Yaron Keren]

Hi Jim,

Clearly searching for name should not be linear in modules but with a map
of some kind.

After compilation to IR each module has a StringMap symbol table.
After compiling to MC and loading the object file, the dynamic linker has
a StringMap symbol table for all loaded modules.

In the usual use case you'll load module(s) into MCJIT and then compile /
link them and all is well, no linear search.

The linear-search use case happens when the MCJIT sort-of lazy compiles: it
gets a list of many, possibly 1000s of function-modules (one function per
module) but had not bothered to compile them to MC and load them since they
were not used yet. It will do so only when needed. In this case, if asked
to find a function it will (currently) linear search for it module by
module.

The solution would be to construct a StringMap for the later use-case,
however in the first use-case it's a wasted effort since all modules are
going to be loaded anyhow and the dynamic linker will build its own map.

MCJIT doesn't know the use-case and so one solution would be to lazy build
the map only when first asked to search, assuming more searches will
follow. This may be wrong assumption so maybe the better solution is to
have a flag or optional function to build this map at the programmer
decision who does know the use-case.

One possible optimization would be for the MCJIT not to build from scratch
but merge the modules stringmaps into one large virtual stringmap - I don't
know if it's possible with this data structure or cheaper than
reconstructing. This data structure would also have to erasing the modules
which are compiled.

Yaron



2013/10/22 Jim Grosbach <grosbach at apple.com>

> I don’t follow. Why are we looking at the module at all? That query should
> work even (especially) after the Module is deleted. We should be able to
> have a local symbol table that’s a DenseMap or something similar to resolve
> from names to target addresses. That map would be updated as part of the
> compilation when the object’s symbol table gets read.
>
> -Jim
>
>
> On Oct 21, 2013, at 4:55 PM, Kaylor, Andrew <andrew.kaylor at intel.com>
> wrote:
>
> There’s probably a lot that we could do, but I can’t think of anything
> easy.****
>
> Basically every time we need to look up a symbol by name we’re going to
> each module and saying “Do you have this symbol?”  It would likely be much
> better if we grabbed the function names from the module and did the search
> ourselves so that we could keep some information about the things that
> didn’t match and optimize the next search.****
>
> -Andy****
>
> *From:* Jim Grosbach [mailto:grosbach at apple.com <grosbach at apple.com>]
> *Sent:* Monday, October 21, 2013 4:39 PM
> *To:* Kaylor, Andrew; Yaron Keren
> *Cc:* <llvmdev at cs.uiuc.edu>
> *Subject:* Re: [LLVMdev] An enhancement for MCJIT::getFunctionAddress****
> ** **
> The search is linear? If that’s really true, we should fix that.****
> ** **
> On Oct 21, 2013, at 10:14 AM, Kaylor, Andrew <andrew.kaylor at intel.com>
> wrote:****
>
>
> ****
> I should have read this before sending my previous reply. :-)****
>  ****
> I’m not a big fan of default parameters, but some form of what you are
> suggesting may be useful.  See my other comments on this topic in the other
> reply.****
>  ****
> -Andy****
>   ****
>  ****
> *From:* Yaron Keren [mailto:yaron.keren at gmail.com <yaron.keren at gmail.com>]
>
> *Sent:* Saturday, October 19, 2013 3:17 PM
> *To:* Kaylor, Andrew; <llvmdev at cs.uiuc.edu>
> *Subject:* An enhancement for MCJIT::getFunctionAddress****
>  ****
> In MCJIT, the old JIT functions are deprecated in favor
> of getFunctionAddress. ****
> Code like:****
>  ****
>   llvm::Function *F = M->getFunction(FuncName);****
>   void *FN = EE->getPointerToFunction(F);****
>  ****
> should be rewritten as****
>  ****
>   uint64_t FN = EE->getFunctionAddress(FuncName);****
>  ****
> While functionally identical, in case the correct module is known the new
> version will be much slower, linear with the number of added (but not
> loaded) modules, since it has to (possibly) search for the correct module
> while old code directly searches the correct module.****
>  ****
> To solve the issue, getFunctionAddress could get an optional Module "hint"
> (=NULL by default) which - if provided - will make getSymbolAddress skip
> calling findModuleForSymbol (a very slow operation) and instead directly
> use the Module provided.****
>  ****
> Yaron****
>  ****
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20131022/196e00f8/attachment.html>