[llvm] [CodeGen] [EarlyIfConversion] Prioritize conversion of hard to predict branches (PR #174457)

[Florian Hahn via llvm-commits]

================
@@ -896,6 +909,133 @@ void EarlyIfConverter::invalidateTraces() {
   Traces->verifyAnalysis();
 }
 
+/// Check if a register's value comes from a memory load by walking the
+/// def-use chain. We want to prioritize converting branches which
+/// depend on values loaded from memory (unless they are loop invariant,
+/// or come from a constant pool).
+/// Results are cached for virtual registers only.
+bool EarlyIfConverter::doOperandsComeFromMemory(Register Reg) {
+  if (!Reg.isVirtual())
+    return false;
+
+  // Check cache first.
+  auto It = OperandMemoryCache.find(Reg);
+  if (It != OperandMemoryCache.end())
+    return It->second;
+
+  // Walk the def-use chain.
+  SmallPtrSet<const MachineInstr *, 8> Visited;
+  SmallVector<const MachineInstr *> Worklist;
+  SmallVector<Register, 16> VisitedRegs;
+
+  MachineInstr *DefMI = MRI->getVRegDef(Reg);
+  // The operand is defined outside of the function - it does not
+  // come from memory access.
+  if (!DefMI) {
+    OperandMemoryCache[Reg] = false;
+    return false;
+  }
+
+  Worklist.push_back(DefMI);
+  VisitedRegs.push_back(Reg);
+
+  while (!Worklist.empty()) {
+    const MachineInstr *MI = Worklist.pop_back_val();
+    if (!Visited.insert(MI).second)
----------------
fhahn wrote:

We should probably limit walking to a number of defs, to avoid excessive compile time on huge functions

https://github.com/llvm/llvm-project/pull/174457