[llvm] [AMDGPU] Fix waterfall inreg call args on gfx90a: hoist defs + avoid AGPR sources (PR #194890)

[Amit Tiwari via llvm-commits]

================
@@ -7109,6 +7109,29 @@ static void emitLoadScalarOpsFromVGPRLoop(
     unsigned NumSubRegs = RegSize / 32;
     Register VScalarOp = ScalarOp->getReg();
 
+    auto LaneClassOf = [&](const TargetRegisterClass *RC) {
+      if (TRI->getRegSizeInBits(*RC) <= 32)
+        return RC;
+      const TargetRegisterClass *Sub =
+          TRI->getSubRegisterClass(RC, AMDGPU::sub0);
+      return Sub ? Sub : RC;
+    };
----------------
amitamd7 wrote:

Was added it as if RC is wider than 32 bits, ask for the `AMDGPU::sub0` subreg class.

https://github.com/llvm/llvm-project/pull/194890