[llvm-commits] [RFC] Patch 1/3 for the native code compiler of Erlang

[Jakob Stoklund Olesen]

On May 4, 2012, at 11:35 AM, Yiannis Tsiouris wrote:

> On 05/03/2012 05:52 PM, Jakob Stoklund Olesen wrote:
>> Yes, except the MI part was a thinko. It should just be an MF reference:
>> 
>>  virtual const TargetRegisterClass *getPointerRegClass(const MachineFunction &MF, unsigned Kind=0) const;
> Using an MI reference might be more clear as MI is almost everywhere
> available. That's why i preferred it.

It should be an MF because the result should not be allowed to depend on which instruction you are passing in.

>>> Furthermore, to do this properly I should change the interface of
>>> getPointerRegClass target hook in all Targets and the usage of it by
>>> changing TRI->getPointerRegClass() wherever needed, right?
>> Yes, and it looks like you will have to add the same MF argument to TargetInstrInfo::getRegClass().
>> 
>> If you find any places that call TII->getRegClass() without having an MF reference available, we may have to find a different solution.
> This *almost* thing is what makes me worried. I tried to implement this
> as follows:
> 
> 1) Changed getPointerRegClass definition to match:
> 
>  virtual const TargetRegisterClass *getPointerRegClass(const MachineFunction &MF, unsigned Kind=0) const;
> 
> 2) I also had to change getRegClass from TargetRegisterInfo as you
> suspected.

You mean TargetInstrInfo::getRegClass(), right?

> But there are some references to it that didn't have an MI
> reference available.

These callers all have an MF reference available:

> Such examples are:
> * lib/CodeGen/MachineInstr.cpp: getRegClassConstraint

getParent()->getParent() (You can assert if either is NULL).

> * lib/CodeGen/MachineRegisterInfo.cpp: recomputeRegClass

Doesn't call TII::getRegClass() directly.

> * lib/CodeGen/SelectionDAG/InstrEmitter.cpp: EmitCopyFromReg

*MF

> * lib/CodeGen/SelectionDAG/ScheduleDAGRRList.cpp: GetCostForDef

Just add an MF argument.

> * lib/Target/Hexagon/HexagonISelDAGToDAG.cpp: SelectIntrinsicWOChain

*MF

The last function also demonstrates why using an MI pointer won't work: There aren't any available in SelectionDAGISel.

/jakob