[llvm] [AMDGPU] Sink async DMA out of s_cbranch_execz then-blocks (PR #196374)

[Vigneshwar Jayakumar via llvm-commits]

================
@@ -654,6 +663,128 @@ void SILowerControlFlow::optimizeEndCf() {
   }
 }
 
+static MachineInstr *findExecRestoreAtBlockStart(MachineBasicBlock &JoinBB,
+                                                 const SIRegisterInfo *TRI) {
+  for (MachineInstr &MI : JoinBB) {
+    if (MI.isMetaInstruction() || MI.isDebugInstr())
+      continue;
+    return MI.modifiesRegister(AMDGPU::EXEC, TRI) ? &MI : nullptr;
+  }
+  return nullptr;
+}
+
+bool SILowerControlFlow::sinkAsyncDMAOutOfExeczBlocks(MachineFunction &MF) {
----------------
VigneshwarJ wrote:

Moved it to a new pass

https://github.com/llvm/llvm-project/pull/196374