[LLVMdev] MCJit interface question
Lang Hames
lhames at gmail.com
Tue Jun 9 13:51:50 PDT 2015
Hi Joseph,
> LLILC is using ORC now, and it was a remarkably smooth transition / small
change
That's great news.
> This tiny amount of boilerplate that I ended up with for creating the
TargetMachine seems entirely reasonable, but going through the exercise and
looking at EngineBuilder code (which of course has several more paths than
reflected above) made me wonder how much of its logic is for building the
TargetMachine and whether pulling it out into a TargetMachineBuilder would
be useful to give ORC clients the support that EngineBuilder gives MCJit
clients.
The use of EngineBuilder().selectTarget() was a short-cut while I was
getting Orc up and running. I think a utility function for building the
TargetMachine sounds like a good idea, especially if the boilerplate gets
any bigger.
> So I simply defined a NullResolver...
That seems like a reasonable thing to include in-tree.
> Perhaps a LambdaObjectLayer that takes a function<void(const
ObjectFile*)> would be appropriate? Let me know if you'd like to see a
patch along those lines; I'm happy to contribute back, but don't have much
context on what would be useful for other clients.
I think an ObjectTransformLayer (analogous to the IRTransformLayer) would
be a useful utility.
Apologies for the lack of documentation for the "layer" concept. I'm doing
some work on Orc over the next couple of days and I'll try to add
documentation for this.
- Lang.
On Thu, Jun 4, 2015 at 1:15 PM, Joseph Tremoulet <jotrem at microsoft.com>
wrote:
> > Sounds good. Please let me know how your experiments go. I'm keen to
> improve the Orc APIs further, so your feedback would be very welcome.
>
>
>
> LLILC is using ORC now, and it was a remarkably smooth transition / small
> change (
> https://github.com/dotnet/llilc/commit/47513add13980e7a32f9b0ec3d2da3db0911bca2).
> The CoreCLR execution engine handles laziness (and symbol resolution) for
> us, so a straightforward application of IRCompileLayer using SimpleCompiler
> and ObjectLinkingLayer along the lines of the "initial" orc kaleidoscope
> example covered our current functionality neatly.
>
>
>
> There were just two spots that I found myself writing some boilerplate:
>
>
>
> One was in constructing the TargetMachine (to pass to the SimpleCompiler
> constructor). With MCJit, the EngineBuilder took care of creating the
> TargetMachine, and the code to set up the EngineBuilder looked like this:
>
>
>
> std::unique_ptr<Module> M = ..;
>
> M->setTargetTriple(LLILC_TARGET_TRIPLE);
>
>
>
> EngineBuilder Builder(std::move(M)); // Builder takes ownership of module
>
> std::string ErrStr;
>
> Builder.setErrorStr(&ErrStr);
>
> Builder.setOptLevel(/* value depending on what the CLR EE requests */);
>
> Builder.setTargetOptions(/* value depending on what the CLR EE requests
> */);
>
>
>
> ExecutionEngine *NewEngine = Builder.create();
>
>
>
> I noticed you used "EngineBuilder().selectTarget()" in the kaleidoscope
> sample, but I couldn't do exactly that for LLILC since we're threading the
> triple through the module that the builder wants to own, and also it seemed
> to make sense that we shouldn't need an EngineBuilder if we're not building
> an Engine, so I looked at what our Builder had been doing and wound up with
> this code to instead create the TargetMachine directly:
>
>
>
> std::string ErrStr;
>
> const llvm::Target *TheTarget =
>
> TargetRegistry::lookupTarget(LLILC_TARGET_TRIPLE, ErrStr);
>
> TargetOptions Options = /* value depending on what the CLR EE requests
> */;
>
> CodeGenOpt::Level OptLevel = /* value depending on what the CLR EE
> requests */;
>
> TargetMachine *TM = TheTarget->createTargetMachine(
>
> LLILC_TARGET_TRIPLE, "", "", Options, Reloc::Default,
> CodeModel::Default,
>
> OptLevel);
>
>
>
> This tiny amount of boilerplate that I ended up with for creating the
> TargetMachine seems entirely reasonable, but going through the exercise and
> looking at EngineBuilder code (which of course has several more paths than
> reflected above) made me wonder how much of its logic is for building the
> TargetMachine and whether pulling it out into a TargetMachineBuilder would
> be useful to give ORC clients the support that EngineBuilder gives MCJit
> clients.
>
>
>
>
>
> The second spot where I needed some code was getting a SymbolResolver to
> pass to the ObjectLinkingLayer. In LLILC we actually don't need
> cross-module symbols resolved because the CoreCLR execution engine resolves
> them and gives the Jit raw addresses. So I simply defined a NullResolver
> like so (I felt it was more readable this way than building an equivalent
> LambdaResolver):
>
>
>
> /// \brief Stub \p SymbolResolver expecting no resolution requests
>
> ///
>
> /// The ObjectLinkingLayer takes a SymbolResolver ctor parameter.
>
> /// The CLR EE resolves tokens to addresses for the Jit during IL reading,
>
> /// so no symbol resolution is actually needed at ObjLinking time.
>
> class NullResolver : public llvm::RuntimeDyld::SymbolResolver {
>
> public:
>
> llvm::RuntimeDyld::SymbolInfo findSymbol(const std::string &Name) final {
>
> llvm_unreachable("Reader resolves tokens directly to addresses");
>
> }
>
>
>
> llvm::RuntimeDyld::SymbolInfo
>
> findSymbolInLogicalDylib(const std::string &Name) final {
>
> llvm_unreachable("Reader resolves tokens directly to addresses");
>
> }
>
> };
>
>
>
> It occurs to me that if any other ORC clients are in the same boat,
> they'll need a resolver that look a lot like this, but I think perhaps
> we're in an unusual boat.
>
>
>
>
>
> One general comment I have is that I wish it were more easily discoverable
> what the constraints are on the various template parameters in ORC code,
> both from the client side (e.g. realizing that SymbolResolverPtrT should be
> a naked/unique/etc pointer to RuntimeDyld::SymbolResolver or a derived
> class) and the consumer side (e.g. realizing that an ObjSetHandleT needs to
> be moved rather than assigned). The comment on IRCompileLayer mentions
> that BaseLayerT "must implement the object layer concept", for example,
> but I wasn't sure what exactly constitutes the object layer concept. I had
> to dig into that one, as I went on to implement an object layer to satisfy
> the CoreCLR's requirements on EH frame reporting that started this thread.
> I figured that implementing the methods on ObjectLinkingLayer would be a
> good bet, but approached it by starting with just addObjectSet and adding
> method implementations as the compiler complained they were missing. With
> that approach, I never did need to implement findSymbolIn,
> mapSectionAddress, or emitAndFinalize, for reasons I haven't dug in to
> understand. The object layer class I produced in this way was itself
> mostly boilerplate:
>
>
>
> class ReserveSpaceLayerT {
>
> public:
>
> typedef LLILCJit::LoadLayerT::ObjSetHandleT ObjSetHandleT;
>
>
>
> ReserveSpaceLayerT(LLILCJit::LoadLayerT *Loader, llilc::EEMemoryManager
> *MM) {
>
> this->Loader = Loader;
>
> this->MM = MM;
>
> }
>
>
>
> template <typename ObjSetT, typename MemoryManagerPtrT, typename
> SymbolResolverPtrT>
>
> ObjSetHandleT addObjectSet(const ObjSetT &Objects,
>
> MemoryManagerPtrT MemMgr,
>
> SymbolResolverPtrT Resolver) {
>
> for (const auto& Obj : Objects) {
>
> MM->reserveUnwindSpace(*Obj);
>
> }
>
> return Loader->addObjectSet(Objects, MemMgr, Resolver);
>
> }
>
>
>
> void removeObjectSet(ObjSetHandleT H) {
>
> Loader->removeObjectSet(std::move(H));
>
> }
>
>
>
> orc::JITSymbol findSymbol(StringRef Name, bool ExportedSymbolsOnly) {
>
> return Loader->findSymbol(Name, ExportedSymbolsOnly);
>
> }
>
>
>
> template <typename OwningMBSet>
>
> void takeOwnershipOfBuffers(ObjSetHandleT H, OwningMBSet MBs) {
>
> Loader->takeOwnershipOfBuffers(std::move(H), std::move(MBs));
>
> }
>
>
>
> private:
>
> LLILCJit::LoadLayerT *Loader;
>
> llilc::EEMemoryManager *MM;
>
> };
>
>
>
> I could imagine that abstracting this might make a useful utility -- what
> I wanted was to call MM->reserveUnwindSpace on each ObjectFile as it passes
> through, and otherwise forward everything on to the base layer. Perhaps a
> LambdaObjectLayer that takes a function<void(const ObjectFile*)> would be
> appropriate? Let me know if you'd like to see a patch along those lines;
> I'm happy to contribute back, but don't have much context on what would be
> useful for other clients.
>
>
>
> Thanks
>
> -Joseph
>
>
>
>
>
> *From:* Lang Hames [mailto:lhames at gmail.com]
> *Sent:* Monday, June 1, 2015 1:41 PM
> *To:* Joseph Tremoulet
> *Cc:* Russell Hadley; llvmdev at cs.uiuc.edu
>
> *Subject:* Re: [LLVMdev] MCJit interface question
>
>
>
> Hi Russell, Joseph,
>
>
>
> > I'll look into moving LLILC to ORC.
>
>
>
> Sounds good. Please let me know how your experiments go. I'm keen to
> improve the Orc APIs further, so your feedback would be very welcome.
>
>
>
> Cheers,
>
> Lang.
>
>
>
> On Sat, May 30, 2015 at 11:14 AM, Joseph Tremoulet <jotrem at microsoft.com>
> wrote:
>
> Agreed, that sounds like the best plan. I'll look into moving LLILC to
> ORC.
>
>
>
> Thanks
>
> -Joseph
>
>
>
> *From:* Russell Hadley
> *Sent:* Friday, May 29, 2015 8:13 PM
> *To:* Lang Hames; Joseph Tremoulet
> *Cc:* llvmdev at cs.uiuc.edu
> *Subject:* RE: [LLVMdev] MCJit interface question
>
>
>
> Hey Joseph,
>
>
>
> What Lang said made me wonder. Is it the right time for us (LLILC) to
> move to ORC? The long term plan was to go there but this could be our
> forcing function.
>
>
>
> -R
>
>
>
> *From:* llvmdev-bounces at cs.uiuc.edu [mailto:llvmdev-bounces at cs.uiuc.edu
> <llvmdev-bounces at cs.uiuc.edu>] *On Behalf Of *Lang Hames
> *Sent:* Friday, May 29, 2015 2:23 PM
> *To:* Joseph Tremoulet
> *Cc:* llvmdev at cs.uiuc.edu
> *Subject:* Re: [LLVMdev] MCJit interface question
>
>
>
> Hi Joseph,
>
>
>
> There are several reasons that a client might want to access the object
> before it's loaded, so a general API like #2 seems like the way to go.
>
>
>
> To support this in MCJIT you can add this to the event listener API. Orc
> clients can already do this by adding a custom object-file layer.
>
>
>
> - Lang.
>
>
>
>
>
>
>
> On Fri, May 29, 2015 at 9:05 AM, Joseph Tremoulet <jotrem at microsoft.com>
> wrote:
>
> Hi,
>
>
>
> I think I need to make a small change to the MCJit interface, and would
> like some feedback on what the most appropriate option would be.
>
>
>
> I'm working on LLILC (a jit for the CoreCLR built on MCJit, which creates
> one module for each MSIL method, containing the main function and zero or
> more EH handler functions extracted from the MSIL method). The CoreCLR
> requires the jit to notify it of the size of the unwind info descriptors
> for each function in the module *before* reserving the memory it will be
> loaded into. So we need a call (or calls) from the Jit/loader into the
> MemoryManager that runs at-or-before reserveAllocationSpace, is
> conceptually similar to registerEHFrames in that it's reserving EH frames,
> but that additionally needs to separate out the per-function information.
>
>
>
> A few options come to mind:
>
> 1. Add a needsToReserveEHFrames callback on MemoryManager (to
> parallel needsToReserveAllocationSpace), and a reserveEHFrames callback
> (parallel to registerEHFrames) that the loader would use to notify the
> memory manager if needsToReserveEHFrames() is true. This seems at a
> high-level the most straightforward fit for the LLILC requirement, but I
> don't know if for other targets it would even be possible to move the
> identification of EH frames (which depends on the 'LocalSections' computed
> in loadObjectImpl) to before calling reserveAllocationSpace. I also don't
> know if that would be an undesirable "tight coupling" of RuntimeDyld with
> CoreCLR's particular interface. (and note that this is really two options,
> in that the code to separate out the per-function EH frame contribution
> could go in either the client memory manager or in the loader.)
>
> 2. Add a notification similar to NotifyObjectEmitted, but which
> runs just *before* the call to Dyld.loadObject. Something like
> NotifyObjectPreLoaded. The LLILC-specific MemoryManager could use this
> hook to scan the object::ObjectFile and pass whatever it needs to the
> CoreCLR. This seems like a decent option to me, but I don't know if it
> would be considered a bad loss of encapsulation to passs out the
> object::ObjectFile in the state where it's been 'created' but not yet
> 'loaded'.
>
> 3. Similar to #2, the existing reserveAllocationSpace callback on
> MemoryManager could simply take an additional parameter which is the
> object::ObjectFile. This would be a more minimal change than #2 in terms
> of how much surface area it adds to the MCJit interface, but also a more
> invasive change in that all MemoryManager implementations would need to be
> updated with the reserveAllocationSpace signature change (though they could
> just ignore the new parameter).
>
> 4. We could avoid needing to crack the object file for this
> information altogether; MCJit could add a hook where a client could insert
> passes into the PassManager used in emitObject; LLILC could attach a pass
> that would consult the MachineModuleInfo, where this information could be
> stored (it's similar to what's stored in the WinEHFuncInfos hanging off the
> MMI today). But adding hooks for client passes might be opening a can of
> worms…
>
>
>
> My inclination would be #2 or #3, but I would love some feedback on which
> of the tradeoffs seem most agreeable (and/or what options I've overlooked).
>
>
>
> Thanks
>
> -Joseph
>
>
>
>
> _______________________________________________
> 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/20150609/52fbaa23/attachment.html>
More information about the llvm-dev
mailing list