[llvm] [BasicBlockSections] Apply path cloning with -basic-block-sections. (PR #68860)

[Sriraman Tallam via llvm-commits]

================
@@ -285,6 +314,130 @@ static bool hasInstrProfHashMismatch(MachineFunction &MF) {
   return false;
 }
 
+// Returns if we can legally apply all clonings specified in `ClonePaths`.
+static bool
+IsValidCloning(const MachineFunction &MF,
+               const DenseMap<unsigned, MachineBasicBlock *> &BBIDToBlock,
+               const SmallVector<unsigned> &ClonePath) {
+  const MachineBasicBlock *PrevBB = nullptr;
+  for (size_t I = 0; I < ClonePath.size(); ++I) {
+    unsigned BBID = ClonePath[I];
+    const MachineBasicBlock *PathBB = BBIDToBlock.lookup(BBID);
+    if (!PathBB) {
+      WithColor::warning() << "no block with id " << BBID << " in function "
+                           << MF.getName() << "\n";
+      return false;
+    }
+
+    if (PrevBB && !PrevBB->isSuccessor(PathBB)) {
+      WithColor::warning() << "block #" << BBID
+                           << " is not a successor of block #"
+                           << *PrevBB->getBBID() << " in function "
+                           << MF.getName() << "\n";
+      return false;
+    }
+
+    if (I != ClonePath.size() - 1 && !PathBB->empty() &&
+        PathBB->back().isIndirectBranch()) {
+      WithColor::warning()
+          << "block #" << BBID
+          << " has indirect branch and appears as the non-tail block of a "
+             "path in function "
+          << MF.getName() << "\n";
+      return false;
+    }
+    PrevBB = PathBB;
+  }
+  return true;
+}
+
+// Applies all clonings specified in `ClonePaths` to `MF` and returns a map
+// from `ProfileBBID`s of all clone blocks to their BBIDs (assigned by
+// `MachineFunction`).
+static DenseMap<ProfileBBID, unsigned>
+ApplyCloning(MachineFunction &MF,
+             const SmallVector<SmallVector<unsigned>> &ClonePaths) {
+  DenseMap<ProfileBBID, unsigned> CloneBBIDMap;
+  // Map from the final BB IDs to the `MachineBasicBlock`s.
+  DenseMap<unsigned, MachineBasicBlock *> BBIDToBlock;
+  for (auto &BB : MF)
+    BBIDToBlock.try_emplace(*BB.getBBID(), &BB);
+
+  DenseMap<unsigned, unsigned> NClonesForBBID;
+  auto TII = MF.getSubtarget().getInstrInfo();
+  for (const auto &ClonePath : ClonePaths) {
+    if (!IsValidCloning(MF, BBIDToBlock, ClonePath)) {
+      // We still need to increment the number of clones so we can map
+      // to the cluster info correctly.
+      for (unsigned BBID : ClonePath)
+        ++NClonesForBBID[BBID];
+      continue;
+    }
+    MachineBasicBlock *PrevBB = nullptr;
+    for (unsigned BBID : ClonePath) {
+      MachineBasicBlock *OrigBB = BBIDToBlock.at(BBID);
+      if (PrevBB == nullptr) {
+        if (auto FT = OrigBB->getFallThrough(/*JumpToFallThrough=*/false)) {
----------------
tmsri wrote:

Why is this done only for the first block in the clone path?

https://github.com/llvm/llvm-project/pull/68860