[llvm] [AMDGPU] Hints for prefering allocating AGPR registers for buffer loads (PR #215697)

[Matt Arsenault via llvm-commits]

================
@@ -96,6 +119,16 @@ bool AMDGPUPrepareAGPRAllocImpl::run(MachineFunction &MF) {
   bool Changed = false;
   for (MachineBasicBlock &MBB : MF) {
     for (MachineInstr &MI : MBB) {
+      if (BufferLoadPreferAGPR && TII.isMUBUF(MI) && MI.mayLoad() &&
+          !MI.mayStore() && MI.getOperand(0).isReg() &&
----------------
arsenm wrote:

Preferring only loads and not stores doesn't make much sense 

https://github.com/llvm/llvm-project/pull/215697