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

[Matt Arsenault via llvm-commits]

================
@@ -369,9 +450,269 @@ bool AMDGPUCodeGenPrepareImpl::run(Function &F) {
       }
     }
   }
+
+  // GlobalISel should directly use the values, and do not need to emit
+  // CopyTo/CopyFrom Regs across blocks
+  if (UsesGlobalISel)
+    return MadeChange;
+
+  // "Optimize" the virtual regs that cross basic block boundaries. In such
+  // cases, vectors of illegal types will be scalarized and widened, with each
+  // scalar living in its own physical register. The optimization converts the
+  // vectors to equivalent vectors of legal type (which are convereted back
+  // before uses in subsequenmt blocks), to pack the bits into fewer physical
+  // registers (used in CopyToReg/CopyFromReg pairs).
+  LiveRegOptimizer LRO(Mod);
+  for (auto &BB : F) {
----------------
arsenm wrote:

Can you just stick with doing this in the visitor instead of introducing another loop over the function?

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