[llvm] r191030 - llvm-c: Add LLVMGetPointerToFunction

Fri Sep 20 09:58:54 PDT 2013

On Sep 19, 2013, at 6:25 PM, Kaylor, Andrew <andrew.kaylor at intel.com> wrote:

> I’ve been working on the MCJIT interface today.  There are a lot of ugly things there as a result of sharing a base class with the old JIT engine.

Very much so. I’ve often wondered whether it would have been better to split off an entirely new class originally to avoid this sort of thing. Probably better the current way, but ugh.

> With regard to the getPointerToX functions, I was thing about adding new methods getXAddress that return uint64_t values and calling the old ones deprecated.  I’m not sure that helps with the C API, but I thought I’d mention it.

That’s a really good idea. For the C API, we just expose those instead. We should put a bit of thought into the inputs there, though. The current interfaces use pointers from the IR (Value* and such), but that’s not long term viable, as there’s nothing in the MCJIT which requires that the lifetime of the Module be >= the lifetime of the resultant compiled code. I’d much rather the interfaces be via name, but that has other complications (platform specific name mangling, etc..).

>  Is there any reason not to have this function return a uint64_t?  Even the void pointer would need to be cast to something else to be useful.

Just that there can be arches w/ > 64 bit pointers (fat pointers on x86, perhaps). I think that’s a reasonable limitation, though.

> BTW, the current implementation of MCJIT::runFunction() will definitely not work if the target is remote.

Yeah, it’s totally broken. I’m in favor of nuking it entirely.

>  
> -Andy
>  
> From: llvm-commits-bounces at cs.uiuc.edu [mailto:llvm-commits-bounces at cs.uiuc.edu] On Behalf Of Jim Grosbach
> Sent: Thursday, September 19, 2013 2:27 PM
> To: Filip Pizlo
> Cc: llvm-commits at cs.uiuc.edu
> Subject: Re: [llvm] r191030 - llvm-c: Add LLVMGetPointerToFunction
>  
>  
> On Sep 19, 2013, at 2:04 PM, Filip Pizlo <fpizlo at apple.com> wrote:
> 
> 
>  
> On Sep 19, 2013, at 1:49 PM, Jim Grosbach <grosbach at apple.com> wrote:
> 
> 
> 
> On Sep 19, 2013, at 1:46 PM, Filip Pizlo <fpizlo at apple.com> wrote:
> 
> 
> How is that different from LLVMGetPointerToGlobal()?
>  
> It’s not. That’s broken, too.
> 
> 
> And I agree that this should work.  It's great that the JIT could be used for remote execution, but some clients will never use the JIT in that fashion.  They will just want a void*.
>  
> Then they will use a memory manager that can make additional guarantees such as the opaque uint64_t value is actual just a bitcasted local pointer which can be safely cast back to void* or whatever. That’s a client-process usage, however, and not a guarantee of the JIT API.
>  
> Are you saying that if a LLVM C API client doesn't install a custom memory manager, then the MCJIT is semantically permitted to JIT the code into a remote process?
>  
> No, I’m saying the API has to be semantically rich enough to be agnostic about remote vs. local. That’s a function of the memory manager interface. The default memory manager does indeed to local currently. I have mixed feelings about making that an explicit guarantee.
> 
> 
> Currently the C API has no facility for asking for code to be JITed into something remote, and it seems like it would be a bug if the JIT did this without the client asking for it.  It makes far more sense to me that the C API continues to have a hard constraint that JITing is local, and that the end result is a function that lives in your address space.
>  
> I strongly disagree. The C API is about exposing functionality, not imposing additional constraints.
>  
>  
> -Jim
> 
> 
>  
>  
> On Sep 19, 2013, at 1:36 PM, Eric Christopher <echristo at gmail.com> wrote:
> 
> 
> Hrm. That's a good point and one I hadn't thought about... what's the
> best way forward here? There's definitely a useful need in getting a
> void * to a function out of the JIT from the C level.
> 
> -eric
> 
> On Thu, Sep 19, 2013 at 1:30 PM, Jim Grosbach <grosbach at apple.com> wrote:
> 
> This isn’t going to work when the execution target isn’t local. The pointer to the function is in the executing process address space and this returns a pointer in the compiler’s address space.
> 
> That’s a fundamental problem with the old JIT, and is semi-baked into the ExecutionEngine API, unfortunately, and cleaning that up is going to be painful. We shouldn’t, however, make the problem worse.
> 
> -Jim
> 
> On Sep 19, 2013, at 12:55 PM, Anders Waldenborg <anders at 0x63.nu> wrote:
> 
> 
> Author: andersg
> Date: Thu Sep 19 14:55:06 2013
> New Revision: 191030
> 
> URL: http://llvm.org/viewvc/llvm-project?rev=191030&view=rev
> Log:
> llvm-c: Add LLVMGetPointerToFunction
> 
> Differential Revision: http://llvm-reviews.chandlerc.com/D1715
> 
> 
> Modified:
>   llvm/trunk/include/llvm-c/ExecutionEngine.h
>   llvm/trunk/lib/ExecutionEngine/ExecutionEngineBindings.cpp
> 
> Modified: llvm/trunk/include/llvm-c/ExecutionEngine.h
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm-c/ExecutionEngine.h?rev=191030&r1=191029&r2=191030&view=diff
> ==============================================================================
> --- llvm/trunk/include/llvm-c/ExecutionEngine.h (original)
> +++ llvm/trunk/include/llvm-c/ExecutionEngine.h Thu Sep 19 14:55:06 2013
> @@ -141,6 +141,8 @@ LLVMGenericValueRef LLVMRunFunction(LLVM
>                                    unsigned NumArgs,
>                                    LLVMGenericValueRef *Args);
> 
> +void *LLVMGetPointerToFunction(LLVMExecutionEngineRef EE, LLVMValueRef F);
> +
> void LLVMFreeMachineCodeForFunction(LLVMExecutionEngineRef EE, LLVMValueRef F);
> 
> void LLVMAddModule(LLVMExecutionEngineRef EE, LLVMModuleRef M);
> 
> Modified: llvm/trunk/lib/ExecutionEngine/ExecutionEngineBindings.cpp
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/ExecutionEngine/ExecutionEngineBindings.cpp?rev=191030&r1=191029&r2=191030&view=diff
> ==============================================================================
> --- llvm/trunk/lib/ExecutionEngine/ExecutionEngineBindings.cpp (original)
> +++ llvm/trunk/lib/ExecutionEngine/ExecutionEngineBindings.cpp Thu Sep 19 14:55:06 2013
> @@ -276,6 +276,10 @@ LLVMGenericValueRef LLVMRunFunction(LLVM
>  return wrap(Result);
> }
> 
> +void *LLVMGetPointerToFunction(LLVMExecutionEngineRef EE, LLVMValueRef F) {
> +  return unwrap(EE)->getPointerToFunction(unwrap<Function>(F));
> +}
> +
> void LLVMFreeMachineCodeForFunction(LLVMExecutionEngineRef EE, LLVMValueRef F) {
>  unwrap(EE)->freeMachineCodeForFunction(unwrap<Function>(F));
> }
> 
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> 
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130920/ad89b564/attachment.html>