[llvm-commits] llvm/lib/Target/PowerPC/PPCRegisterInfo.cpp

[Chris Lattner]

On Mar 3, 2007, at 6:30 AM, Nicolas Geoffray wrote:

> Chris Lattner wrote:
>>
>>> Because R31 is not used as a callee saved register when  
>>> llvm::NoFramePointerElim is set to 1.
>>
>> NoFramePointerElim doesn't change the ABI, only the codegen model.
>
> Yes it doesn't change the ABI, but R31's offset in Linux/ELF is in  
> the callee saved area. In Macho it is not.

Ok, so it's not related to NoFramePointerElim?  If that's the case,  
you should just have Macho and ELF return different sets of callee  
saved regs.

>>> The algorithm
>>> that calculates the callee saved registers' addresses does not  
>>> save R31's offset when used as a frame pointer.
>>> And with the ELF Abi, R31's offset is in the callee saved  
>>> registers area. Therefore I have to force its allocation
>>> (at the -4 offset).
>>
>> Please see how the LR register forces spilling to a specific stack  
>> slot.
>>
>
> You're talking about this code right?
>  MachineLocation LRDst(MachineLocation::VirtualFP, LROffset);
>  MachineLocation LRSrc(IsPPC64 ? PPC::LR8 : PPC::LR);
>  Moves.push_back(MachineMove(FrameLabelId, LRDst, LRSrc));
>
> However this is not enough for R31 because R31's offset is in the  
> callee saved area. And if I don't add the code of this patch,
> the algorithm that calculates callee saved addresses does not take  
> into account the fact that R31 is in the callee saved area.

Hrm, the PPC backend changed.  It should override the  
'getCalleeSavedSpillSlots' hook, to say that LR (and R31) must be  
spilled as specific offsets on the stack.  I don't know how it works  
now.

> Btw, I'm just reproducing the code of  LowerDYNAMIC_STACKALLOC  
> (PPCISelLowering.cpp). As the compilation may not go
> through this method, I have to do the saving of R31's offset before  
> the callee saved scan.

LowerDYNAMIC_STACKALLOC is a bit different, on darwin, it has to  
ensure that the stack pointer/frame pointer are kept in sync as the  
stack is adjusted.  This is a very strange ABI rule, and I'd be  
surprised if ELF did the same thing.

-Chris