llvm dev - Apr 2009 - [LLVMdev] JIT relocations and stub memory issue

Hi All,

Can anyone explain why the MachineRelocation field "NeedsStub" is set
to
true for GlobalValues on x86_64, but false on plain x86? This is causing the
JITEmitter to allocate a stub for every GlobalValue relocation it encounters
on x86_64, quickly exhausting the 512K region for stubs in the default JIT
memory manager. Is this expected behavior? Is there anyway to override the
stub generation? In this case, why are stubs even being allocated, my
understanding was that they were merely placeholders for lazy JIT?

The same application running on plain x86 requires no such stubs and
therefore runs correctly. There is currently no stub deallocation
functionality in the default memory manager, so the x86_64 version can only
JIT a few thousand functions before the ExecutionEngine crashes with a JIT:
Ran out of space for function stubs error.

Thoughts?

Thanks,
Eric
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20090413/ad32a1c2/attachment.html>

On Apr 13, 2009, at 8:09 AM, Eric wrote:
> Hi All,
>
> Can anyone explain why the MachineRelocation field "NeedsStub" is
> set to true for GlobalValues on x86_64, but
That's true for large code model or Mac OS X. Right now, x86_64  
targets uses small code model by default. However, in JIT, the GV  
memory are allocated in heap and they are not guaranteed to be  
*close*. So the JIT is currently emitting stubs for all of them.

Evan
> false on plain x86? This is causing the JITEmitter to allocate a  
> stub for every GlobalValue relocation it encounters on x86_64,  
> quickly exhausting the 512K region for stubs in the default JIT  
> memory manager. Is this expected behavior? Is there anyway to  
> override the stub generation? In this case, why are stubs even being  
> allocated, my understanding was that they were merely placeholders  
> for lazy JIT?
>
> The same application running on plain x86 requires no such stubs and  
> therefore runs correctly. There is currently no stub deallocation  
> functionality in the default memory manager, so the x86_64 version  
> can only JIT a few thousand functions before the ExecutionEngine  
> crashes with a JIT: Ran out of space for function stubs error.
>
> Thoughts?
>
> Thanks,
> Eric
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev