[llvm] [MachineLICM] Allow hoisting loads from invariant address (PR #70796)

[Florian Hahn via llvm-commits]

================
@@ -1333,6 +1346,46 @@ void MachineLICMBase::InitCSEMap(MachineBasicBlock *BB) {
     CSEMap[BB][MI.getOpcode()].push_back(&MI);
 }
 
+/// Initialize AllowedToHoistLoads with information about whether invariant
+/// loads can be moved outside a given loop
+void MachineLICMBase::InitializeLoadsHoistableLoops() {
+  SmallVector<MachineLoop *, 8> Worklist(MLI->begin(), MLI->end());
+  SmallVector<MachineLoop *, 8> LoopsInPreOrder;
+
+  // Mark all loops as hoistable initially and prepare a list of loops in
+  // pre-order DFS.
+  while (!Worklist.empty()) {
+    auto *L = Worklist.pop_back_val();
+    AllowedToHoistLoads[L] = true;
+    LoopsInPreOrder.push_back(L);
+    Worklist.insert(Worklist.end(), L->getSubLoops().begin(),
+                    L->getSubLoops().end());
+  }
+
+  // Going from the innermost to outermost loops, check if a loop has
+  // instructions preventing invariant load hoisting. If such instruction is
+  // found, mark this loop and its parent as non-hoistable and continue
+  // investigating the next loop.
+  // Visiting in a reversed pre-ordered DFS manner
+  // allows us to not process all the instructions of the outer loop if the
+  // inner loop is proved to be non-load-hoistable.
+  for (auto *Loop : reverse(LoopsInPreOrder)) {
+    for (auto *MBB : Loop->blocks()) {
+      // If this loop has already been marked as non-hoistable, skip it.
+      if (!AllowedToHoistLoads[Loop])
+        continue;
+      for (auto &MI : *MBB) {
+        if (!MI.mayStore() && !MI.isCall() &&
----------------
fhahn wrote:

It looks like this doesn't consider instructions with  unmodelled sideeffects, like memory barriers and we now (incorrectly) hoist loads across memory barriers. I put up https://github.com/llvm/llvm-project/pull/116987

https://github.com/llvm/llvm-project/pull/70796