ExecutionEngine::finalizeModule()

[Kaylor, Andrew]

> So the loaded state purpose is having modules in "standby" should they be required. OK.

Not quite.  The purpose of the "added" state is having modules in standby. (added=known but not compiled)  The purpose of the "loaded" state is to have an intermediate state wherein clients can remap the addresses of sections, using MCJIT::mapSectionAddress, (in preparation for later copying to a new location or an external process) before relocations and page permissions are applied.  (loaded=compiled and required sections copied into local memory but not yet ready for execution)


> Maybe it's possible to split a module into function-modules automatically in a addModuleFunctions() function.

That's certainly possible.  I'm on the fence as to whether or not it's desirable.  I could be convinced by a good argument either way.  It seems to me like it belongs more in a "helper" area.


> 1) Is the purpose of having loaded and finalized states to enable an external action on the module between loading and finalizing?

Strictly speaking it is to enable externally initiated action on the object that was generated from the module between loading and finalizing.

It might not be obvious at first glance, but the "remote-mcjit" case in the lli tool does this.  In that case, the intermediate action is taken by the RemoteMemoryManager in response to the notifyObjectLoaded function being called.


> 2) Is it OK to finalize a set of modules, add more modules and finalize again?

Yes!


-Andy


From: Yaron Keren [mailto:yaron.keren at gmail.com]
Sent: Friday, October 18, 2013 3:04 PM
To: Kaylor, Andrew
Cc: llvm-commits at cs.uiuc.edu
Subject: Re: ExecutionEngine::finalizeModule()

This data structure sounds good .

I actually didn't consider lazy compilation with regard about module states because I knew that "MCJIT does not have lazy compilation".  So the loaded state purpose is having modules in "standby" should they be required. OK.

If you put one function into one module exactly (call this a function-module) you almost have the lazy compilation capability. Maybe it's possible to split a module into function-modules automatically in a addModuleFunctions() function.

Some questions -

1) Is the purpose of having loaded and finalized states to enable an external action on the module between loading and finalizing?

2) Is it OK to finalize a set of modules, add more modules and finalize again?

Yaron


2013/10/19 Kaylor, Andrew <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>>
I've actually got an idea for this and am working on a patch today.  I'll send it to you for input when I have something working.

Basically my idea is to replace the existing Modules vector that MCJIT inherits from ExecutionEngine with a smarter container class that will internally maintain separate lists/sets for added, loaded and finalized modules.   This class would provide iterator access that MCJIT could use to iterate through the entire combined set of modules (I think I can make that work reasonably) but also iterators to go through just the modules in a particular state.  This would let us get rid of the clumsy ModuleState thing completely and provide faster access to the subsets we need.

The "loaded" state is needed because when relocations are processed (during the finalize handling) they are applied to all modules that have been loaded and the finalize operation on the memory manager (which sets page protection, invalidates code caches, etc.) is applied to the sections of all loaded modules.  It would require a significant amount of extra infrastructure to make this behave otherwise and I couldn't think of a compelling reason why it would be needed.

But maybe there's some confusion here.  The emitted state is entirely useless because it is always immediately followed by object loading.  The state I'm currently calling "loaded" is probably what you're referring to as "compiled".  We definitely do not want to immediately compile modules when addModule is called because that would eliminate the only facsimile of lazy compilation MCJIT currently has.  The idea is that you can add a module to provide function definitions but if nothing in that module is referenced by a module in which a function is executed (note the wording here because it's not exactly the ideal case) then the module never gets compiled.

Eventually, we'll want to add another level of lazy compilation in which we generate stubs to resolve external function during linking/finalization and only compile their dependent modules if the stub is called.  But that's a job for another day.

-Andy

From: Yaron Keren [mailto:yaron.keren at gmail.com<mailto:yaron.keren at gmail.com>]
Sent: Friday, October 18, 2013 1:41 PM

To: Kaylor, Andrew
Cc: llvm-commits at cs.uiuc.edu<mailto:llvm-commits at cs.uiuc.edu>
Subject: Re: ExecutionEngine::finalizeModule()

Ok, it's simpler than I thought since only three states are relevant.

But first - do we need really all three states? what good is loaded state? is just means being on the list, and any operation that is possible on a loaded module should be possible before adding the module to MCJIT, right?

If we merge loading and emitted by having the addModule immediately emit the loaded module we'll need only two states:  "Compiled" and "Linked" (or "Finalized").

addModule emits module and adds the module to an "Compiled" vector of modules.
Finalize loops over "Compiled" vector, each module is finalized, added to "Linked" list and then the "Compiled" list is cleared.

If we still need three lists then just the same, three vectors would be required with add inserting the loaded vector, emit moving modules from loaded to compiled and finalize moving them from compiled to linked.

Yaron




2013/10/18 Kaylor, Andrew <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>>
Your suggestion to introduce an optimization within MCJIT to avoid inefficiency in finalizeLoadedModules is the correct direction.  I definitely prefer to improve the MCJIT implementation rather than increasing the surface of the interface and requiring clients to optimize their own use of MCJIT.

We've talked about removing runFunction from the MCJIT interface.  It really doesn't belong there and was just implemented as a convenience to ease the transition of legacy code to MCJIT.  As such, I think of it in the same way I think of finalizeObject.  That is, I'm hoping it will go away.  It might make sense to provide an independent class or perhaps a static function that provides the signature decoding that's contained in MCJIT::runFunction, which would ease the pain of deprecating it.  That method isn't entirely complete as it stands, BTW.

The Module state handling in MCJIT definitely isn't mature.  The only states that really have any significance are "Added" (which needs a better name), "Loaded" and "Finalized".  The current state tracking is just something I put in there to get the multiple module support up and running.  I would be happy to see it replaced with something better.

-Andy


From: Yaron Keren [mailto:yaron.keren at gmail.com<mailto:yaron.keren at gmail.com>]
Sent: Thursday, October 17, 2013 4:02 PM
To: Kaylor, Andrew
Cc: llvm-commits at cs.uiuc.edu<mailto:llvm-commits at cs.uiuc.edu>
Subject: Re: ExecutionEngine::finalizeModule()

The changes you are making make sense. Return valid ready-to-run function address rather than depend upon the user calling finalize is right.

The issue is with scalability of the current  implementation. Let's take your Kaliedoscope example (which was very helpful) as the use-case but every REPL will do the same.

After adding  a new module it and only it needs to be finalized to run the new code. But finalizeLoadedModules loops over all modules to see which one should be finalized. It's not efficient.

I initially thought using finalizeModule would be the solution, calling it after creating a module, saving the loop.

However this goes deeper and finalizeModule isn't the solution.  As you say getFunctionAddress calls finalizeLoadedModules() if a symbol found. finalizeLoadedModules always loops on all modules so it's wasteful. In an extreme case, we may know that all modules are already finalized but can't prevent the needless loop on module.

The solution would be other data structures.
Simplified case, instead of looping on all modules and checking module states  MCJIT could have a vector of un-finalized modules and a vector of finalized modules. New modules enter the un-finalized vector and finalizeLoadedModules() loops over the un-finalized vector, finalize every module, appends them to finalized vector and clears the unfinalized vector.

This is simplified since there are more ModuleStates so maybe other data structures should be used. I have not looked in depth the ModuleStates.

Regarding getFunctionAddress, runFunction() does not use it but calls the older getPointerToFunction, so either runFunction() should be changed to the newer API or we still need to manually finalize modules.

Yaron

2013/10/18 Kaylor, Andrew <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>>
Hi Yaron,

What do you see as the use case for adding this?  I was thinking about removing finalizeObject() from the interface -- it's mostly there to support code that was written when it was necessary.  It was probably a misstep to introduce finalizeObject in the EE interface in the first place.  It's a case of leaking an implementation detail into the interface.

With the current trunk implementation of MCJIT, calling getFunctionAddress will implicitly finalize the module that contains the function (as well as any other modules that have been loaded).  The idea is that the address you get back is immediately useful (which was not true of the older getPointerToFunction method).

Also note that within MCJIT calling finalizeModule has the effect of finalizing all loaded modules.  In fact, I plan to update the implementation of finalizeModule to call finalizeLoadedModules soon.

-Andy


From: Yaron Keren [mailto:yaron.keren at gmail.com<mailto:yaron.keren at gmail.com>]
Sent: Thursday, October 17, 2013 2:32 PM
To: llvm-commits at cs.uiuc.edu<mailto:llvm-commits at cs.uiuc.edu>; Kaylor, Andrew
Subject: ExecutionEngine::finalizeModule()

Hi,

Make finalizeModule() accessible for MCJIT created through ExecutionEngine.

Maybe it would make sense to expose finalizeLoadedModules() as well.

Yaron





-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20131018/4fac38cb/attachment.html>