[llvm] [AMDGPU] Insert readfirstlane for uniform VGPR arguments (PR #178198)

[Matt Arsenault via llvm-commits]

================
@@ -3627,6 +3627,14 @@ SDValue SITargetLowering::LowerFormalArguments(
 
     Reg = MF.addLiveIn(Reg, RC);
     SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
+    if (Arg.Flags.isInReg() && RC == &AMDGPU::VGPR_32RegClass) {
+      // FIXME: Need to forward the chains created by `CopyFromReg`s, make sure
----------------
arsenm wrote:

I think so? It's more of a pre-existing issue for all of the argument lowering cases

https://github.com/llvm/llvm-project/pull/178198