[PATCH] Add two new calling conventions for runtime calls

[Philip Reames]

Your revised naming scheme would be fine by me.

We actually have several different calling conventions in use. Our 
compiled-code-to-assembly-stub conventions is quite close to your 
proposed CPreservesMost, except that we occasionally use a dual-register 
return variant for some calls.  Our compiled-to-compiled convention is a 
slightly rotated form of the C calling convention.  When we make calls 
from compiled code into C code (via a hand written assembly stub), we 
want to be able to easily insert a leading argument without having to 
shift all other arguments. (i.e. for JNI)

Things to consider for your documentation:
- Is this defined to match the objc runtime?  If so, which?  Or is 
defined in terms of a variant of the C calling convention which ObjC 
happens to use?  Or is it defined in terms of purpose and only valid 
against code compiled by the same version of LLVM with the same 
convention name?  (i.e. compatibility guarantees)
- What is the difference between these and existing calling 
conventions?  In particular, "cold"?  When should each be used?
- How should target neutral front ends use the conventions?

You don't need to answer any of those directly to me.  Just provide some 
documentation which addresses them to some degree. (The most important 
is the first.)

Philip

On 1/16/14 11:23 AM, Juergen Ributzka wrote:
> Hi Philip,
>
> thanks for looking at this. I am definitely open for better names for 
> the calling conventions. Although I would prefer not to use GPR or XMM 
> in them, because they are not only meant for X86. Also on a different 
> architecture with different standard calling conventions a different 
> mix of registers might make more sense. ObjC is currently the initial 
> use for this, but we don’t want to limit it to just this one 
> particular runtime. The more people can use this the better.
>
> What about CPreservesMost and CPreservesAll?
>
> How different is your special calling convention?
>
> Thanks
>
> Cheers,
> Juergen
>
>
> On Jan 16, 2014, at 10:45 AM, Philip Reames <listmail at philipreames.com 
> <mailto:listmail at philipreames.com>> wrote:
>
>> On 1/16/14 10:39 AM, Philip Reames wrote:
>>> My only objection to this patch is the naming.  We also have runtime 
>>> calls with special calling conventions, but ours aren't the same as 
>>> the ones specified here. Could we rename these?
>>>
>>> A few random ideas for names:
>>> CPreserveGPR
>>> CPreserveGPRXMM
>>> WebKitRuntimeCC
>> Correction: ObjCRuntime.  It helps if I read the original email more 
>> closely.
>>>
>>> Also, should these be prefixed with X86_?
>>>
>>> Philip
>>>
>>> On 1/15/14 10:41 AM, Lang Hames wrote:
>>>> Hi Juergen,
>>>>
>>>> If these calling conventions are intended to be generic, should a
>>>> description be added to the LangRef?
>>>>
>>>> Otherwise LGTM.
>>>>
>>>> - Lang.
>>>>
>>>> On Tue, Jan 14, 2014 at 3:11 PM, Juergen Ributzka 
>>>> <juergen at apple.com <mailto:juergen at apple.com>> wrote:
>>>>> ping
>>>>>
>>>>> On Jan 10, 2014, at 7:45 PM, Juergen Ributzka <juergen at apple.com 
>>>>> <mailto:juergen at apple.com>> wrote:
>>>>>
>>>>>> New rebased patch and a small fix to the label check in the unit 
>>>>>> test.
>>>>>>
>>>>>> Cheers,
>>>>>> Juergen
>>>>>>
>>>>>> <0002-Add-two-new-calling-conventions-for-runtime-calls.patch>
>>>>>> On Jan 10, 2014, at 3:47 PM, Juergen Ributzka <juergen at apple.com 
>>>>>> <mailto:juergen at apple.com>> wrote:
>>>>>>
>>>>>>> Hi @ll,
>>>>>>>
>>>>>>> this patch adds two new target-independent calling conventions 
>>>>>>> for runtime calls - Runtime and FastRuntime.
>>>>>>>
>>>>>>> The target-specific implementation for X86-64 is defined as 
>>>>>>> following:
>>>>>>>      - Arguments are passed as for the default C calling convention
>>>>>>>      - The same applies for the return value(s)
>>>>>>>      - for runtimecc the callee preserves all GPRs - except R11
>>>>>>>      - for fastruntimecc the callee preserves all GPRs and all 
>>>>>>> XMMs - except R11
>>>>>>>
>>>>>>> The idea is to provide calling conventions for calls to very hot 
>>>>>>> runtime functions that are normally just a few lines of assembly 
>>>>>>> code and don’t require a lot of registers. This could be used by 
>>>>>>> the ObjectiveC runtime, or any other runtime that provides 
>>>>>>> performance critical functions.
>>>>>>>
>>>>>>> The FastRuntime CC is intended to be used for small codes that 
>>>>>>> don’t call any other functions at all.
>>>>>>> The Runtime CC is also intended for small codes that usually 
>>>>>>> don’t call other functions on the fast path, but might have to 
>>>>>>> on the slow path.
>>>>>>>
>>>>>>> Cheers,
>>>>>>> Juergen
>>>>>>>
>>>>>>> <0001-Add-two-new-calling-conventions-for-runtime-calls.patch>
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> llvm-commits mailing list
>>>>>>> llvm-commits at cs.uiuc.edu <mailto: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 <mailto: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 <mailto: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 <mailto: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/20140116/3922a376/attachment.html>