[LLVMdev] Compile function with limited set of registers? Jump to another function?

Tue Feb 1 23:32:53 PST 2011

Hi James,

Joshua is right, what you're trying to accomplish is quite known in the Java
VM world (
http://domino.research.ibm.com/comm/research_people.nsf/pages/dgrove.oopsla01.html
).

In order to express the "thunk" code in LLVM you need a full control of how
registers are used (because otherwise they would mess up with the
arguments). I haven't investigated enough to know if that's possible today
in LLVM (I think it wasn't a few years ago when I added the optimization).
So what I ended up with was a generic IR that walks the interface table of
an object and detects collisions. The IR was inline in the caller to make
sure that arguments in registers are not overriden.

This may not be the best approach today, but I believe it was the easiest
way to have something efficient at that point.

Cheers,
Nicolas

On Mon, Jan 31, 2011 at 11:01 PM, James Williams <junk at giantblob.com> wrote:

> Thanks, that's a good idea - I'll have a look through the VMKit source.
>
> -- James
>
>
> On 31 January 2011 21:39, Joshua Warner <joshuawarner32 at gmail.com> wrote:
>
>> Hi James,
>>
>> I see the problem now.  You might look at VMKit (a Java VM build with the
>> LLVM JIT compiler)  - I would expect it uses a similar method for resolving
>> interface calls (the method, if I understand it correctly, is well-known in
>> the Java world).
>>
>> I've CC'd the main dev behind VMKit - he might be able to lend some
>> insight.
>>
>> --Joshua
>>
>> On Mon, Jan 31, 2011 at 2:24 PM, James Williams <junk at giantblob.com>wrote:
>>
>>> Hi Joshua,
>>>
>>> Thanks - I was hoping that would be the case.
>>>
>>> However, I've had a think about this since I posted to the list and I
>>> believe the only way to handle these issues safely in LLVM IR would be to
>>> define the thunk as varargs. I'm not sure how well LLVM handles varargs but
>>> ideally it would all compile down to nothing since the parameters to the the
>>> thunk would be in the same registers/stack locations as required by the
>>> target method.
>>>
>>> Unfortunately, varargs has some downsides: there's the additional
>>> overhead for the extra hidden parameter to every interface method call for
>>> the parameter count plus it doesn't (I don't think) support tail calls.
>>>
>>> -- James
>>>
>>>
>>> On 27 January 2011 17:37, Joshua Warner <joshuawarner32 at gmail.com>wrote:
>>>
>>>> Hi James,
>>>>
>>>> I'm no expert in LLVM IR, but I think that just encoding each *actual*
>>>> method invocation in the thunk as a tail call would work.  This would
>>>> require trusting that LLVM passes / code generators will translate down to a
>>>> jump, as is normal.  If the passes / code generators are smart, I see no
>>>> reason that LLVM wouldn't emit code that fits your requirements.  Either
>>>> way, you know that your thunk will be correct - it just might not be as
>>>> efficient as you want.
>>>>
>>>> I would suggest experimenting with generating a thunk this way, and look
>>>> at the resultant target assembly to make sure it's doing what you want.
>>>>
>>>> -Joshua
>>>>
>>>>  On Tue, Jan 25, 2011 at 2:04 AM, James Williams <junk at giantblob.com>wrote:
>>>>
>>>>>  Hi,
>>>>>
>>>>> Can anyone tell me, is it possible to express in LLVM IR:
>>>>>   - that, for a specific function, register allocator can use only
>>>>> limited set of registers? (specifically, cannot touch any registers that
>>>>> might contain parameters)
>>>>>   - that stack can't be touched? (or at least must balance on exit from
>>>>> thunk)
>>>>>   - jump, not call, to another function leaving any received parameters
>>>>> unchanged in registers and on stack?
>>>>>
>>>>> Thanks,
>>>>> -- James Williams
>>>>>
>>>>> Background:
>>>>>
>>>>> I'm looking for some advice on implementing thunks required by my
>>>>> language's interface call mechanism. This is a fairly conventional
>>>>> arrangement where method selectors in interfaces are hashed to determine
>>>>> their index within a vtable and hash collisions are disambiguated at runtime
>>>>> by a thunk, which determines which method to call from a selector id passed
>>>>> as the first method parameter.
>>>>>
>>>>> I'm currently using a single thunk (written in assembly) for all
>>>>> collisions that walks a table to determine what method to call. This works
>>>>> but it's inefficient and requires the a hand written thunk for each
>>>>> supported target.
>>>>>
>>>>> I'd like to instead generate IR for a specific thunk for each vtable
>>>>> collisoin that does a binary search of possible selectors because this will
>>>>> avoid some pointer dereferences and an additional indirect call.
>>>>>
>>>>> The problem is that a thunk may need to decide between methods with
>>>>> different signatures without disturbing parameters in registers and on the
>>>>> stack and then jump to, rather than call, another function:
>>>>>
>>>>> interface X:
>>>>>   method A(a, b)
>>>>>
>>>>> interface Y:
>>>>>   method B(c, d, e)
>>>>>
>>>>> class Z implements X, y:
>>>>>   method A(a, b) ...
>>>>>   method B(c, d, e) ...
>>>>>
>>>>> X.A + Y.B happen to hash to same vtable index, say -3
>>>>>
>>>>> This would require a thunk something like:
>>>>>
>>>>> vtable[-3] =
>>>>>   thunk_Z_AorB(selector_id, ...)
>>>>>     // binary search for matching selector id:
>>>>>     if selector_id <= selector_Z_A then
>>>>>       Z.A(selector_id, ...)
>>>>>     else
>>>>>       Z.B(selector_id, ...)
>>>>>     fi
>>>>>
>>>>> which would ideally would compile on x64 to something like:
>>>>>
>>>>> thunk_Z_AorB:
>>>>>   cmp $selector_Z_A, %rdi
>>>>>   jle Z.A
>>>>>   jmp Z.B
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> 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/20110202/daca0361/attachment.html>