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

[Cullen Rhodes]

================
@@ -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) {
----------------
c-rhodes wrote:

this is untested

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