[llvm] [AMDGPU] Mark AGPR tuple implicit in the first instr of AGPR spills. (PR #115285)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> > > > Looking at copyPhysReg, I think the same bug exists there for the V_PK_MOV_B32 case
> 
> > > 
> 
> > > 
> 
> > > Are you going to look into this one?
> 
> > 
> 
> > Yes, I am.
> 
> 
> 
> I found https://github.com/llvm/llvm-project/blob/e821f642fdc75922b1a020447485acccf3e7fa92/llvm/test/CodeGen/AMDGPU/v_mov_b64_expansion.mir i.e expansion of V_MOV_B64_PSEUDO which uses V_PK_MOV_B32 where implicit is not used. Are you referring to the same case? or something else?


This test is unrelated. I mean the handling of COPY for wide tuples when v_pk_mov_b32 is used 

https://github.com/llvm/llvm-project/pull/115285