[Mlir-commits] [mlir] [mlir][ArmSME] Use liveness information in the tile allocator (PR #90448)

[Andrzej Warzyński]

================
@@ -137,172 +129,508 @@ static ArrayRef<TileMask> getMasks(ArmSMETileType type) {
   }
 }
 
-/// Allocates and returns a tile ID. Returns an error if there are no tiles
-/// left.
-static FailureOr<unsigned> allocateTileId(ArmSMETileType tileType,
-                                          TileMask &tilesInUse) {
-  auto masks = getMasks(tileType);
-  for (auto [tileId, tileMask] : llvm::enumerate(masks)) {
-    if ((tilesInUse & tileMask) == TileMask::kNone) {
-      tilesInUse |= tileMask;
-      return tileId;
+class TileAllocator {
+public:
+  /// Allocates and returns a tile ID. Fails if there are no tiles left.
+  FailureOr<unsigned> allocateTileId(ArmSMETileType tileType) {
+    auto masks = getMasks(tileType);
+    for (auto [tileId, tileMask] : llvm::enumerate(masks)) {
+      if ((tilesInUse & tileMask) == TileMask::kNone) {
+        tilesInUse |= tileMask;
+        return tileId;
+      }
     }
+    return failure();
+  }
+
+  /// Releases a previously allocated tile ID.
+  void releaseTileId(ArmSMETileType tileType, unsigned tileId) {
+    TileMask tileMask = getMasks(tileType)[tileId];
+    assert((tilesInUse & tileMask) != TileMask::kNone &&
+           "cannot release unallocated tile!");
+    tilesInUse ^= tileMask;
+  }
+
+  /// Allocates an in-memory tile ID.
+  unsigned allocateInMemoryTileId() {
+    // Note: We never release in-memory tile IDs. We could, which may allow
+    // reusing an allocation, but as we _never_ want to spill an SME tile this
+    // is not optimized.
+    return nextInMemoryTileId++;
   }
-  return failure();
-}
 
-/// Collects transitive uses of a root value through control flow. This can
-/// handle basic SCF constructs, along with control flow (br and cond_br).
-/// Simple loops work at the SCF level, while more complex control flow can be
-/// dealt with after lowering to CF. This is used to implement basic tile
-/// allocation.
-static void findDependantOps(Value rootValue,
-                             SetVector<Operation *> &dependantOps) {
-  auto traverseCorrespondingValues = [&](auto inputValues, auto exitValues) {
-    for (auto [idx, value] : llvm::enumerate(inputValues)) {
-      if (value == rootValue)
-        findDependantOps(exitValues[idx], dependantOps);
+private:
+  TileMask tilesInUse = TileMask::kNone;
+  unsigned nextInMemoryTileId = kInMemoryTileIdBase;
+};
+
+// Add new intermediate blocks for the true and false destinations of a
+// `cf.cond_br`. This prevents spurious liveness overlaps due to copies at
+// branches.
+void splitCondBranches(IRRewriter &rewriter, FunctionOpInterface function) {
+  SmallVector<cf::CondBranchOp> worklist;
+  function.walk([&](cf::CondBranchOp condBranch) {
+    if (llvm::any_of(condBranch->getOperands(), [&](Value value) {
+          return isValidSMETileVectorType(value.getType());
+        })) {
+      worklist.push_back(condBranch);
     }
+  });
+
+  auto insertJump = [&](Location loc, Block *source, Block *dest, auto args) {
+    rewriter.setInsertionPointToEnd(source);
+    rewriter.create<cf::BranchOp>(loc, dest, args);
   };
-  for (Operation *user : rootValue.getUsers()) {
-    if (dependantOps.contains(user))
+
+  for (auto condBranch : worklist) {
+    auto loc = condBranch.getLoc();
+    Block *block = condBranch->getBlock();
+    auto newTrueBranch = rewriter.splitBlock(block, block->end());
+    auto newFalseBranch = rewriter.splitBlock(block, block->end());
+    insertJump(loc, newTrueBranch, condBranch.getTrueDest(),
+               condBranch.getTrueDestOperands());
+    insertJump(loc, newFalseBranch, condBranch.getFalseDest(),
+               condBranch.getFalseDestOperands());
+    condBranch.getFalseDestOperandsMutable().clear();
+    condBranch.getTrueDestOperandsMutable().clear();
+    condBranch.setSuccessor(newTrueBranch, 0);
+    condBranch.setSuccessor(newFalseBranch, 1);
+  }
+}
+
+/// Inserts tile copies at `cf.br` operations.
+void insertCopiesAtBranches(IRRewriter &rewriter,
+                            FunctionOpInterface function) {
+  splitCondBranches(rewriter, function);
+  for (Block &block : function.getBlocks()) {
+    Operation *terminator = block.getTerminator();
+    if (!isa<cf::BranchOp>(terminator))
       continue;
-    dependantOps.insert(user);
-    TypeSwitch<Operation *>(user)
-        .Case<cf::BranchOp>([&](auto branchOp) {
-          // (CF) Follow branch.
-          traverseCorrespondingValues(branchOp.getDestOperands(),
-                                      branchOp.getDest()->getArguments());
-        })
-        .Case<cf::CondBranchOp>([&](auto condBranchOp) {
-          // (CF) Follow true branch.
-          traverseCorrespondingValues(
-              condBranchOp.getTrueOperands(),
-              condBranchOp.getTrueDest()->getArguments());
-          // (CF) Follow false branch.
-          traverseCorrespondingValues(
-              condBranchOp.getFalseOperands(),
-              condBranchOp.getFalseDest()->getArguments());
-        })
-        .Case<LoopLikeOpInterface>([&](auto loopOp) {
-          // (SCF) Follow iter_args of (basic) loops (e.g. for loops).
-          traverseCorrespondingValues(loopOp.getInits(),
-                                      loopOp.getRegionIterArgs());
-        })
-        .Case<scf::YieldOp>([&](auto yieldOp) {
-          // (SCF) Follow yields of (basic) control flow (e.g. for loops).
-          auto parent = user->getParentOp();
-          traverseCorrespondingValues(user->getOperands(),
-                                      parent->getResults());
+    rewriter.setInsertionPoint(terminator);
----------------
banach-space wrote:

I think that you are referring to what's happening in _this_ method _today_. However, this method might get updated and also the same `rewriter` is share between other hooks. Would adding an insertion guard do any harm?

https://github.com/llvm/llvm-project/pull/90448