[llvm] [AMDGPU][True16][CodeGen] Implement sgpr folding in true16 (PR #128929)

[Matt Arsenault via llvm-commits]

================
@@ -1040,6 +1053,14 @@ void SIFoldOperandsImpl::foldOperand(
       }
     }
 
+    // Allow immediates COPYd into sgpr_lo16 to be further folded while
+    // still being legal if not further folded
+    if (DestRC == &AMDGPU::SGPR_LO16RegClass) {
+      assert(ST->useRealTrue16Insts());
+      MRI->setRegClass(DestReg, &AMDGPU::SGPR_32RegClass);
----------------
arsenm wrote:

You should not modify the register class of the value until fully committing the rewrite of the instruction. This is speculatively modifying the class, so we may change the class and then not perform the fold based on the below conditions 

https://github.com/llvm/llvm-project/pull/128929