[llvm] [WIP][AMDGPU] Improve the handling of `inreg` arguments (PR #133614)

[Matt Arsenault via llvm-commits]

================
@@ -3130,8 +3221,17 @@ SDValue SITargetLowering::LowerFormalArguments(
       llvm_unreachable("Unexpected register class in LowerFormalArguments!");
     EVT ValVT = VA.getValVT();
 
-    Reg = MF.addLiveIn(Reg, RC);
-    SDValue Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
+    SDValue Val;
+    // If an argument is marked inreg but gets pushed to a VGPR, it indicates
+    // we've run out of SGPRs for argument passing. In such cases, we'd prefer
+    // to start packing inreg arguments into individual lanes of VGPRs, rather
+    // than placing them directly into VGPRs.
+    if (RC == &AMDGPU::VGPR_32RegClass && Arg.Flags.isInReg()) {
+      Val = Spiller.readLane(Chain, DL, Reg, VT);
+    } else {
+      Reg = MF.addLiveIn(Reg, RC);
+      Val = DAG.getCopyFromReg(Chain, DL, Reg, VT);
----------------
arsenm wrote:

will probably also need to be a WWM copy 

https://github.com/llvm/llvm-project/pull/133614