[LLVMdev] MCJIT and Lazy Compilation

Mon Mar 11 14:55:12 PDT 2013

Hi Andrew, the prototype looks great! Andy Kaylor is on vacation, so I'm not sure if he'll see this thread until next week.

You definitely have MCJIT moving in the right direction, but the API you're using (ExecutionEngine) has to be stretched a little bit to accomplish your goals.  In your implementation, do clients call recompileAndRelinkFunction() for every module that is needed to satisfy inter-module dependencies (if so, can the same thing be accomplished with the existing ExecutionEngine::addModule()?) or do clients call it only to update ("hotswap") a module that has been previously compiled but has since changed?

Off hand, it sounds like ExecutionEngine might need another function like:

updateModule(Module* M);

which does what your modified recompileAndRelinkFunction() does. It should be easy to add any new functions you need to the API in ExecutionEngine.h as virtuals (with default, empty, implementations in order to keep the old JIT happy) and overrides in MCJIT that do the right thing.

Regarding tests, I'm thinking that the existing lli-based MCJIT integration tests (tests/ExecutionEngine/MCJIT) are going to be harder to extend to support multiple modules (with function dependencies, globals, etc) than the unit tests (unittests/ExecutionEngine/MCJIT). Anyways, it sounds like you have already been testing lots of cases, but I'm not sure if that's been done manually or in some automatic fashion. There is a (commented-out) test in MCJITTests.cpp that intends to test the case of dependencies between modules, but there's probably lots of other cases that can be added too. In any case, I find that writing unit tests is a big help to figure out what an API should look like.

Also, I believe there's an assumption (in at least one place) that ExecutionEngines take ownership of Modules that are passed in. It sounds like you removed it (from MCJIT at least) by getting rid of the field M and the corresponding assertions, but I'm wondering if there's any other common code in ExecutionEngine that makes the same assumption about a single module whose ownership is transferred.

Can you elaborate on what you mean by "1) The current single Module case may need to be retained??". I don't see why the single-module case cannot be handled in exactly the same way as the multiple module cases.

Ashok (cc'd) has in the past looked at MCJIT memory managers, and specifically at applying permissions, so he might have some more thoughts, and I'll leave it to Andy K to review at the relocation stuff as that's his domain of expertise!

At this point though, I think lots people are excited about MCJIT (specifically, multiple modules with dependencies and hot-swapping) and would love to see a patch when you have a moment to rebase it on top of current trunk. Don't worry if the patch is incomplete; just seeing its design might help elicit some more targeted comments!

Thanks,
Dan

From: Andrew Sorensen <digegoo at gmail.com<mailto:digegoo at gmail.com>>
Date: Saturday, 9 March, 2013 4:26 AM
To: Andrew Kaylor <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>>
Cc: LLVM List <llvmdev at cs.uiuc.edu<mailto:llvmdev at cs.uiuc.edu>>
Subject: Re: [LLVMdev] MCJIT and Lazy Compilation

Hi Andy/Albert,

Sorry for the slow reply, my day job caught up with me.

Two bits of progress. (a) MCJIT is working nicely for non-trivial
examples in Extempore on x86 and ARM, and (b) the page
permissions are now RO again. For your amusement a *very*
cheesy video of Extempore running on-the-fly with MCJIT on an
ARM Pandaboard. Viewer discretion is advised!
https://vimeo.com/60407237

Here is the overview of changes I promised a couple of weeks back.
These comments are based on the 3.3 trunk of about 3 weeks ago.

-- RuntimeDyld --

Relocations needs to be cleared between
each emitObject.

New method required: clearRelocations

void RuntimeDyldImpl::clearRelocations() {
  ExternalSymbolRelocations.clear();
  Relocations.clear();
}

void RuntimeDyld::clearRelocations() {
  Dyld->clearRelocations();
}

-- RuntimeDyldImpl --

  void clearRelocations();

-- SectionMemoryManager --

Removed option in allocateSection to use existing 'free' memory
regions - i.e. now always allocate new memory. This means that once a
module is emitted and its memory permissions are applied we don't have
to touch it again. This means we don't have to set exec sections
writable before a new emit. I'm sure there is a nicer way to achieve
this.

-- MCJIT --

MCJIT currently takes a single Module and only supports a single call
to emitObject. A number of assertion and conditional checks enforce
this. These checks can all be removed (assertions and conditionals),
making M obsolete.

At the moment there are a number of call sites for emitObject. I
removed all of them and replaced them with a single emitObject call
site accessible through the extant, but unused,
recompileAndRelinkFunction method. This then becomes the clients
access point to compile each individual module. This is obviously a
hack to maintain the integrity of the existing API (i.e. bad name,
evil type munging F->M etc..).

void *MCJIT::recompileAndRelinkFunction(Function *F) {
  emitObject((Module*) F);
  finalizeObject();
  return NULL;
}

FinalizeObject should not call resolveRelocations. Instead it should
call the *new* clearRelocations.

void MCJIT::finalizeObject() {
  // New Dyld call clearRelocations
  Dyld.clearRelocations();

  // Set page permissions.
  MemMgr->applyPermissions();
}

At the moment I am just leaking allocated sections in SMM. i.e.

void MCJIT::freeMachineCodeForFunction(Function *F) {
  dbgs() << "free machine code not yet supported in MCJIT\n";
  return;
}

But it should be relatively straight forward to maintain some kind of
Module -> SMM Section map. Hotswapping currently works fine because
the relocations all update as expected. So this is *just* a leakage
problem.

-- AsmParser/LLParser --

Some forward refs in the LLParser need fixing. I'm working directly
from IR though, so C++ API people will need a different fix.

-------------------------

I might have missed a few things as it has been a couple of weeks
since I looked at the code but that is the general drift. As
previously mentioned there really is not much required to make all of
this function well for the individual module case.

Of course this does not address the issues that clients may face
moving from JIT to MCJIT as a result of these changes. However, I
expect that many have similar solutions to Extempore, and should not
have any major problems moving to a multi-module approach. Albert do
you see this being much of a problem for Pure?

There are obviously some things that will need fixing.

1) The current single Module case may need to be retained??
2) We need a more appropriate call than recompileAndRelinkFunction
3) We need proper section reclamation/erasure
4) Better solution for permissions re. freemem in section allocation.

Andy, aside from these 4 points does everything here sound OK to you,
and if so how would you like to proceed. I can put *some* time in but
you'll need to give me some direction about what you would like done.

Cheers,
Andrew.

On Sat, Feb 16, 2013 at 5:26 PM, Andrew Sorensen <digegoo at gmail.com<mailto:digegoo at gmail.com>> wrote:
Hey Andy,

Yep I've tested some non-trivial examples with loads of dependencies,
both code and data, global, local and external symbol resolution etc..

Actually this was truly a piece of cake, nothing to do, the memory manager
is working really nicely so far as I can tell.  Relocations to sections are all working
as expected (aside from previously mentioned ARM issue which is probably just
something that I'm doing wrong) with all global symbol relocs managed persistently
by the MM between object injections.  All in all it just works ;)  I had to make a few
minor adjustments to things like the LLParser for forward dependencies but overall
really simple stuff.

There certainly are some section management issues that will need to be addressed,
but I don't see any major hurdles there.  I was going to take a look into this next week?

The biggest issue for multi-module is probably going to be client side not LLVM side,
although this has not been a huge problem for me as most of this bookkeeping is
already managed "client side" in extempore.

I'm happy to send you code although it might be more useful for me to write a
followup email outlining exactly what changes were made and then let the experts
decide how best to proceed ;)  Tomorrows a little hectic but I'll try to send a note
through on Monday.

Cheers,
Andrew.

On Sat, Feb 16, 2013 at 8:13 AM, Kaylor, Andrew <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>> wrote:
This is great news.

Do you have any dependencies between your modules?  For instance, one calling a function in another?  If so, how did you handle that?

Any chance you could share some code snippets or the relevant portions?

-Andy

From: Andrew Sorensen [mailto:digegoo at gmail.com<mailto:digegoo at gmail.com>]
Sent: Thursday, February 14, 2013 11:48 PM

To: Kaylor, Andrew
Cc: llvmdev at cs.uiuc.edu<mailto:llvmdev at cs.uiuc.edu>
Subject: Re: [LLVMdev] MCJIT and Lazy Compilation

OK, so I have some *preliminary* results, which are on the whole quite encouraging!

I haven't had a great deal of time, but I have managed to get Extempore up and
running with function (actually lexical closures so composed of quite a bit of additional
guff) level compilation using Andy's multi module suggestion. I also have on-the-fly
recompilation of existing closures working (caveats below) so from an end-user
perspective this means that Extempore appears functionally equivalent with MCJIT
and the old legacy JIT - hot-swapping audio signal processing code on-the-fly using
MCJIT for example.

Firstly multi-module definitely proved to be considerably easier than attempting to hack
solutions for incremental *monolithic* module builds - which I also investigated.

So the only major obstacle that I have run into so far are page permissions in relation
to code relocations.  I have a *safe* hack which is to toggle section permissions between
rw and exec/ro in-between new object injections - however this is obviously problematic
for code that is executing concurrently (i.e. secondary threads).  I also have an *unsafe*
hack, (purely for experimentation :-) whereby exec sections are left rw, and although
very evil it works for test purposes (i.e. the audio example mentioned above).  These
solutions are obviously both inappropriate and I will investigate a *real* solution when
I find some time.

Also I didn't bother to implement section erasure, at the moment I'm just allocating
new sections for each compile regardless of whether the new code replaces existing
functionality. Having said that I don't see this as much of an issue, I was just to
lazy to bother implementing it.  I'll check this when I have some further free time.

FYI this is all under x86.  I did try to run under ARM but bombed out on an assertion error
in the ARM ELF relocation code - specifically   assert((*TargetPtr & 0x000F0FFF) == 0);
I assume this is a result of something evil that I have done but I haven't yet had time to
investigate any further.  Again I'll let you know when I have some more time.

Just a quick heads up but In general my initial thoughts are that MCJIT is really not
that far off.

Cheers,
Andrew.

On Fri, Feb 8, 2013 at 7:36 AM, Kaylor, Andrew <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>> wrote:
That’s awesome!

I think at this point having people try out various approaches and seeing what works and what doesn’t is our biggest need in this area.  Please do keep me informed about what you find out.

-Andy

From: Andrew Sorensen [mailto:digegoo at gmail.com<mailto:digegoo at gmail.com>]
Sent: Wednesday, February 06, 2013 4:33 PM
To: Kaylor, Andrew
Cc: llvmdev at cs.uiuc.edu<mailto:llvmdev at cs.uiuc.edu>
Subject: Re: [LLVMdev] MCJIT and Lazy Compilation

Thanks for the update Andy.

I'm very happy to be involved in anyway that is helpful.  If you would like me to test ideas, or contribute to further discussions, then please let me know.

I currently have extempore running nicely with MCJIT for the "monolithic" case and am working on various LLVM hacks to better understand the issues involved with non-monolithic approaches - in particular I'm starting with your multi-module approach.  I will report back when (and if) I have something useful to contribute.

Cheers,
Andrew.

On Wed, Feb 6, 2013 at 4:08 AM, Kaylor, Andrew <andrew.kaylor at intel.com<mailto:andrew.kaylor at intel.com>> wrote:
Hi Andrew,

I was about to write a belated reply to this message (sorry for the delay), but then I realized that pretty much everything useful that I have to say on the subject is contained in this message (which is in a thread Albert Graef already linked to):

https://groups.google.com/d/msg/llvm-dev/Rk9cWdRX0Wg/Fa1Mn6cyS9UJ

Generally, I do hope that MCJIT will be capable of replacing the old JIT someday soon, though obviously it cannot do so until it provides equivalent functionality.  I doubt it will ever be a “drop-in” replacement, but I hope that minimal rework will be needed.  Most significantly, as can be seen in earlier discussions, things will need to be made Module-centric rather than Function-centric.  It ought to be possible to write a utility class that takes a monolithic Module and breaks it up into sub-Modules for individual functions, but I think that would need to happen outside of the MCJIT engine because not all clients would want that kind of granularity.

There’s definitely a lot of work to be done here to get this right, and hopefully we’ll get active participation in any design discussions to make sure the solution meets everyone’s needs.  I don’t have a time table for this right now.  I will file a Bugzilla report as soon as the LLVM server is ready.

-Andy

From:llvmdev-bounces at cs.uiuc.edu<mailto:llvmdev-bounces at cs.uiuc.edu> [mailto:llvmdev-bounces at cs.uiuc.edu<mailto:llvmdev-bounces at cs.uiuc.edu>] On Behalf Of Andrew Sorensen
Sent: Thursday, January 31, 2013 7:56 PM
To: llvmdev at cs.uiuc.edu<mailto:llvmdev at cs.uiuc.edu>
Subject: [LLVMdev] MCJIT and Lazy Compilation

Does anyone have a roadmap for MCJIT with what I think people are
calling lazy compilation.

Is this even on the cards?

I spent the last few hours moving my project (extempore.moso.com.au<http://extempore.moso.com.au>)
over to MCJIT (particularly for ARM), and am a little horrified to discover
no ability to compile, and just as importantly to recompile, at a function level.
This is absolutely mandatory for my project.

I have been looking enviously at MCJIT's ARM+DWARF support for a
couple of years and was under the misapprehension that MCJIT was
attempting to be a *drop-in* replacement for JIT.  So I wasn't overly
concerned about the primary JIT being largely neglected. This is obviously
my fault, I wasn't paying close enough attention.

I am now wondering what the LLVM project, in the large, plans regarding
just-in-time compilation moving forward.  Is MCJIT the future, and
if so what kind of roadmap is there to replicate current JIT functionality.
In my case in relation to function level (re)compilation.

I appreciate everyones efforts, and that we all have our own agendas.
I'm just trying to put my own roadmap in place.

Cheers,
Andrew.