[LLVMdev] getMinimalPhysRegClass

[Jakob Stoklund Olesen]

On Jan 26, 2012, at 3:19 PM, Reed Kotler wrote:

> Does anyone understand the purpose of this target inpdendent function?

It's actually a leftover from way back when a register would belong to ONE register class. This hasn't been true for a long time, and we have mostly eliminated code making such assumptions.

The function has since been changed to mean "give me the most precise register class containing PhysReg".

> By adding a new register class that is for MIPS16 but not even 
> referencing it, the compiler breaks because of this code.
> 
> def CPU16Regs : RegisterClass<"Mips", [i32], 32, (add
>   // Return Values and Arguments
>   V0, V1, A0, A1, A2, A3,
>   // Callee save
>   S0, S1
>  )>;
> 
> I'm trying to understand how the Arm compiler avoids this problem.
> 
> But still, what is the logic here?

PrologEpilogInsertion wants to spill S0 as a callee-saved register, but storeRegToStackSlot insists that you give it a register class. PEI uses getMinimalPhysRegClass() because it has no other knowledge of the register class.

MipsInstrInfo::storeRegToStackSlot() needs to be aware that register classes can have sub-classes.

Compare:

  if (RC == Mips::CPURegsRegisterClass)
    Opc = IsN64 ? Mips::SW_P8 : Mips::SW;
  else if (RC == Mips::CPU64RegsRegisterClass)
    Opc = IsN64 ? Mips::SD_P8 : Mips::SD;
 
To ARM:

      if (ARM::GPRRegClass.hasSubClassEq(RC)) {

/jakob