[llvm] [AMDGPU] Add IR LiveReg type-based optimization (PR #66838)

[Matt Arsenault via llvm-commits]

================
@@ -341,6 +342,85 @@ class AMDGPUCodeGenPrepare : public FunctionPass {
   StringRef getPassName() const override { return "AMDGPU IR optimizations"; }
 };
 
+class LiveRegConversion {
+private:
+  // The instruction which defined the original virtual register used across
+  // blocks
+  Instruction *LiveRegDef;
+  // The original type
+  Type *OriginalType;
+  // The desired type
+  Type *NewType;
+  // The instruction sequence that converts the virtual register, to be used
+  // instead of the original
+  std::optional<Instruction *> Converted;
----------------
arsenm wrote:

I think std::optional with a pointer has negative value, it's not really any safer than the raw pointer and more syntactically heavy 

https://github.com/llvm/llvm-project/pull/66838