[llvm] [AMDGPU] Allocate i1 argument to SGPRs (PR #72461)

[Matt Arsenault via llvm-commits]

================
@@ -296,6 +317,35 @@ bool AMDGPUCallLowering::canLowerReturn(MachineFunction &MF,
   return checkReturn(CCInfo, Outs, TLI.CCAssignFnForReturn(CallConv, IsVarArg));
 }
 
+/// Replace CallLowering::determineAndHandleAssignments() because we need to
+/// reserve ScratchRSrcReg when necessary.
+/// TODO: Investigate if reserving ScratchRSrcReg can be moved to calling conv
+/// functions. If so, then this function is not needed anymore -- we can just
+/// use CallLowering::determineAndHandleAssignments() as before.
+bool AMDGPUCallLowering::determineAndHandleAssignmentsLocal(
----------------
arsenm wrote:

Instead of creating determineAndHandleAssignmentsLocal, would be better to just add an overload of determineAndHandleAssignments in the base class that takes the CCInfo as an argument. IIRC the DAG has a similar split 

https://github.com/llvm/llvm-project/pull/72461