[mlir] Automatic reference counting for Async values + runtime support for ref counted objects

[ezhulenev]

Summary

Depends On D89963

Automatic reference counting algorithm outline:

1. ReturnLike operations forward the reference counted values without modifying the reference count.
2. Use liveness analysis to find blocks in the CFG where the lifetime of reference counted values ends, and insert drop_ref operations after the last use of the value.
3. Insert add_ref before the async.execute operation capturing the value, and pairing drop_ref before the async body region terminator, to release the captured reference counted value when execution completes.
4. If the reference counted value is passed only to some of the block successors, insert drop_ref operations in the beginning of the blocks that do not have reference coutned value uses.

[ftynse]

mlir/lib/Conversion/AsyncToLLVM/AsyncToLLVM.cpp

669

You probably want to take the operand from operands rather than from the op directly in case it was modified by another pattern.

AddRefOpAdaptor is an autogenerated class that is constructible from ArrayRef<Value> and provides an API similar to the Op it models, i.e. you can call adaptor.operand().

681

Could we do something like

template <typename OpTy>
class RefToCallLoweringPattern : public OpConversionPattern<OpTy> {
  RefLoweringPatter(MLIRContext *ctx, StringRef funcName) : OpConversionPattern<OpTy>(ctx), funcName(funcName) {}

  matchAndRewrite(...) {
    ...
    rewruter.replaceOpWithNewOp<CallOp>(op, Type(), funcName, ValueRange(args));
  }
};

and remove duplicate code?

910

I would recommend to make ConstantOp legal, not the whole StandardDialect, which has lots of different things.

mlir/lib/Dialect/Async/IR/Async.cpp

349–350 ↗

(On Diff #302681)

Just declare it as IntNonNegative in ODS.

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

38

MLIR uses /// for top-level comments.

65

Out of scope: I am interested in seing this as a generic OpInterface, just yesterday the need for this popped up in another discussion.

144

Any particular reason for using 32bit integers for refcount? In this struct, it may not even save space because the compiler will insert padding.

276

19 looks very unconventional. We usually try to estimate what would be the common "small" number of entries and round it up to a power of two.

mlir/lib/ExecutionEngine/AsyncRuntime.cpp

94

please fix

[ezhulenev]

Remove code duplication in op lowering + fix style guide violations

mlir/lib/Conversion/AsyncToLLVM/AsyncToLLVM.cpp

669

Wouldn't the changes be also visible through the op? From the auto generated code is seems that they are identical:

::mlir::Value AddRefOpAdaptor::operand() {
  return *getODSOperands(0).begin();
}

vs

::mlir::Operation::operand_range AddRefOp::getODSOperands(unsigned index) {
  auto valueRange = getODSOperandIndexAndLength(index);
  return {std::next(getOperation()->operand_begin(), valueRange.first),
           std::next(getOperation()->operand_begin(), valueRange.first + valueRange.second)};
}

::mlir::Value AddRefOp::operand() {
  return *getODSOperands(0).begin();
}

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

65

Yeah, seems like a useful property in many contexts. Will leave it for the followup.

144

Not really, just to match the type of the count arg in add_ref/drop_ref ops, but that choice is also arbitrary.

276

That was a typo, it was supposed to be 10 :) Changed to 8 here and below, because that seems like a reasonable upper bound for number of uses for an async value,

[ezhulenev]

Use IntPositive trait for ref count attr

[ftynse]

mlir/lib/Conversion/AsyncToLLVM/AsyncToLLVM.cpp

669

No they will not be visible. Conversion almost never changes operations in-place. replaceOpWithNewOp and the likes inject a new op, and keep the old op until the conversion completes in case one needs to examine the original op or its operand. The list of the operands to the op being rewritten is formed by combining the results of the new ops if they were rewritten and existing ops if they were not. This is why we pass operands into matchAndRewrite, otherwise it would have been a useless copy of op->getOperands().

[rriddle]

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

40

Missing static on all of these?

[ezhulenev]

Add static to functions in AsyncRefCounting.cpp

[herhut]

mlir/lib/Conversion/AsyncToLLVM/AsyncToLLVM.cpp

675

Why not produce the ValueRange in place from the two arguments?

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

38

Nit: are.

170

I would argue for not having the users consume reference counts, as this makes it impossible to optimize the decrement operations in IR (they are tied to the ops). For instance, if you had inc_rc and dec_rc explicit, and both were in a loop, you could hoist the increments and sink the decrements, removing the overhead from the loop.

That might be a better way to optimize this in general. First insert all increments and decrements trivially where needed (the buffer deallocation pass could do this for you, see my comment on other CL) and then have a pass that pushes increments and decrements up/down, combining them where possible.

Seems less fragile and would work with existing interfaces for region control flow. It would also allow to pass async values to operations that do not implement the reference counting consumer interface.

[ezhulenev]

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

170

FWIW Swift SIL has all reference counting explicit (https://github.com/apple/swift/blob/main/docs/ARCOptimization.rst).

There are two types of ref-counted value users:

1. "forwarding": std.return, function call arg - they do not change the ref count
2. "consumers" - everything else.

Async automatic ref counting will need to either have a closed set of supported users, or rely in op interfaces to distinguish between user types.

[ezhulenev]

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

170

And there is also operation like mlirAsyncRuntimeAddTokenToGroup that consumes reference at some indeterminate point in the future, so if IR has drop_ref, then the operation will need to have add_ref to compensate for that or marked as "forwarding" (reference counting responsibility forwarded to the runtime)

[silvas]

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

170

It is unclear what "dynamic operation" means in this context and why scf.for is the "innermost". Can you adjust the comment?

I also don't understand "Inside this operation statically known number of uses is 1" - if %cond is false it will be 0.

181

nit: looks like line wrapping here forgot to insert //.Same on the async.drop_ref below.

273

nit: you might want to clarify somwhere that when you say "instances" here, it is "per instance of result's owner".

[ezhulenev]

Use liveness analysis for reference counting

[ezhulenev]

Construct ValueRange directly as an argument to create call

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

170

I've pushed a new revision based on liveness analysis and explicit drop_ref instead of implicit "ref consumer".

[silvas]

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

193

Why only ExecuteOp? Why not use NumberOfExecutions?

[ezhulenev]

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

193

Because operations after the async.execute can be executed before the operations nested under the async.execute, this is currently the only operation that has this property.

Example:

%token = ...
async.execute {
  async.await %token : !async.token // await #1
  async.yield
}
async.await %token : !async.token // await #2

It is impossible to determine which of the async.await operations will be the "last use" at runtime. Ref counting will pick second await as the last user and will create drop_ref after it, however if first await will be executed later it needs to keep the token alive.

[silvas]

mlir/include/mlir/Dialect/Async/IR/AsyncOps.td

234

nit: "All values are semantically created"

235

unclear what "owner" means in this context. Is this referring to a runtime construct or IR construct?

mlir/lib/Conversion/AsyncToLLVM/AsyncToLLVM.cpp

54

should it start with "create" to match the others?

670

rewriter has some helpers to avoid these raw get calls.

674

This should use operands[0] for the converted operands since this is doing a type conversion.

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

38

Discuss runtime refcounting ABI conventions for runtime functions in this comment. And conventions for IR functions that accept/return refcounted objects.

49

Add the explanation from your other review comment here justifying the special treatment of async.execute.

56

nit: typo coutned

63

typo: dialect types are

93

explain why not nested blocks (or leave TODO; also, we should probably signalPassFailure if we encounter uses in nested region)

108

typo: in in

122

findAncestorOpInBlock is tricky. Can you do this? (or leave a comment explaining the tricky case):

for (Operation *user : value.getUsers()) {
  if (user->getParent() == block) {
    userInTheBlock = user;
    break;
  }
}

Also, recommend putting this in a static helper, per https://llvm.org/docs/CodingStandards.html#use-early-exits-and-continue-to-simplify-code

212

I think you can avoid findAncestorBlockInRegion/findAncestorOpInBlock by just doing while (user->getRegion() != definingRegion). That would make this code simpler as well.

244

I would prefer to keep such optimizations in a separate pass. Advantages:

1. Easy to show and test tricky cases of this optimization (the current code requires a level of indirection -- one has to imagine which ops are inserted, and then removed)
2. When debugging a miscompile, it is easier to bisect by removing an optimization pass which should not affect correctness.
3. Can do this more efficiently. The current algorithm is O(BlockSize^3); many ML programs are single blocks of >1000 ops. I think this algorithm can be replaced with with a single walk of each block, applying the optimization to all refcounted Value's in that block at the same time.
4. Makes test cases for this pass clearer because users can see all the ops inserted and follow along with the code.

(if you want to omit this optimization from the initial patch, that is fine too).

[ezhulenev]

Add a separate AsyncRefCountingOptimization pass + address PR comments

[ezhulenev]

mlir/include/mlir/Dialect/Async/IR/AsyncOps.td

235

Changed the documentation to reflect the new implementation of automatic reference counting.

mlir/lib/Conversion/AsyncToLLVM/AsyncToLLVM.cpp

54

createTokenFunctionType == function type for createToken function. Renamed to addOrDropRefFunctionType to make it clear that it is for add_ref and drop_ref ops.

674

Yes, also fixed a similar bug below.

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

93

Added few lines to explain why ignoring nested regions is ok.

122

findAncestorOpInBlock required to find the last use in the block even if the "real" use is deep inside nested region.

%token = ...
scf.for %i = ... {      <<<----- `scf.for` will be the last user
  async.await %token : !async.token
}
asyn.drop_ref %token.   <<<---- will be added after the last use in the CFG

Cleaned up code a little bit.

244

I moved it to a separate async-ref-counting-optimization pass. It is still not as efficient as it could be, but I added a small preprocessing step + iterate only the blocks that have uses of value.

[ezhulenev]

Fix a bug in ref counting optimization

[ezhulenev]

Break the loop early in user is after dropRef

[ezhulenev]

ValueUser->UserInfo

[ezhulenev]

Mark symbol declaration private

[silvas]

Thanks! This looks great!

mlir/include/mlir/Dialect/Async/IR/AsyncOps.td

246

nit: could -> can

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

47

nit: "it is the responsibility of the async value user" seems to imply that it is not this pass's responsibility. Suggest "To implement automatic reference counting, we must insert a +1 reference before each Operation using the value".

76

typo: yied

mlir/lib/Dialect/Async/Transforms/AsyncRefCountingOptimization.cpp

40 ↗

(On Diff #306239)

suggest putting this helper in include/Dialect/Async/IR/Async.h; it is used in the other file too.

mlir/test/Dialect/Async/async-ref-counting-optimization.mlir

1 ↗

(On Diff #306239)

Is it interesting to test async.execute[%token]?

55 ↗

(On Diff #306239)

is scf.if essential to this test case? If not, remove it. if so, describe it in the comment.

58 ↗

(On Diff #306239)

The input IR here seems strange to me. Will it create a leak if %arg1 == false?

I don't see a test case that produces IR that looks like this in async-ref-counting.mlir. Perhaps it would be good to add.

64 ↗

(On Diff #306239)

nit: inconsistency of CHECK: drop_ref vs CHECK: async.drop_ref

mlir/test/Dialect/Async/async-ref-counting.mlir

146

Is there a missing CHECK: async.add_ref %[[TOKEN]] on the line before %token_0 = async.execute and a missing CHECK: async.drop_ref %[[TOKEN_0]] before the return?

(best to show all add_ref/drop_ref, or use CHECK-NOT to show that they are not produced there)

mlir/test/Dialect/Async/verify.mlir

25

generally we don't test propreties verified by traits/interfaces.

[ezhulenev]

Address PR comments

[ezhulenev]

Thanks for the review!

mlir/test/Dialect/Async/async-ref-counting-optimization.mlir

1 ↗

(On Diff #306239)

Added a test, it is indeed quite common pattern with nested async execute operations.

58 ↗

(On Diff #306239)

I was not really thinking about ref counting correctness when writing this tests :) Added an explicit note to the test where this property is violated.

mlir/test/Dialect/Async/async-ref-counting.mlir

146

Yes, forgot to update some tests after decoupling it from ref counting optimization. Added back missing checks to few other tests.

mlir/include/mlir/Dialect/Async/IR/AsyncBase.td

 
 def Async_AnyValueType : DialectType<AsyncDialect,
                            CPred<"$_self.isa<::mlir::async::ValueType>()">,
                                  "async value type">;
 
 def Async_AnyValueOrTokenType : AnyTypeOf<[Async_AnyValueType,
                                            Async_TokenType]>;
 
+def Async_AnyAsyncType : AnyTypeOf<[Async_AnyValueType,
+                                    Async_TokenType,
+                                    Async_GroupType]>;
+
 #endif // ASYNC_BASE_TD

mlir/include/mlir/Dialect/Async/IR/AsyncOps.td

def Async_AwaitAllOp : Async_Op<"await_all", []> {
   }];
 
   let arguments = (ins Async_GroupType:$operand);
   let results = (outs);
 
   let assemblyFormat = "$operand attr-dict";
 }
 
+//===----------------------------------------------------------------------===//
+// Async Dialect Automatic Reference Counting Operations.
+//===----------------------------------------------------------------------===//
+
+// All async values (values, tokens, groups) are reference counted at runtime
+// and automatically destructed when reference count drops to 0.
+//
+// All values semantically created with a reference count of +1 and it is
+// the responsibility of the async value owner to add/drop reference count
+// based on the number of uses.
+//
+// See `AsyncRefCountingPass` for the automatic reference counting
+// implementation details.
+
+def Async_AddRefOp : Async_Op<"add_ref"> {
+  let summary = "adds a reference to async value";
+  let description = [{
+    The `async.add_ref` operation adds a reference(s) to async value (token,
+    value or group).
+  }];
+
+  let arguments = (ins Async_AnyAsyncType:$operand,
+                       Confined<I32Attr, [IntNonNegative]>:$count);
+  let results = (outs );
+
+  let assemblyFormat = [{
+    $operand attr-dict `:` type($operand)
+  }];
+}
+
+def Async_DropRefOp : Async_Op<"drop_ref"> {
+  let summary = "drops a reference to async value";
+  let description = [{
+    The `async.drop_ref` operation drops a reference(s) to async value (token,
+    value or group).
+  }];
+
+  let arguments = (ins Async_AnyAsyncType:$operand,
+                       Confined<I32Attr, [IntNonNegative]>:$count);
+  let results = (outs );
+
+  let assemblyFormat = [{
+    $operand attr-dict `:` type($operand)
+  }];
+}
+
 #endif // ASYNC_OPS

mlir/include/mlir/Dialect/Async/Passes.h

 #define MLIR_DIALECT_ASYNC_PASSES_H_
 
 #include "mlir/Pass/Pass.h"
 
 namespace mlir {
 
 std::unique_ptr<OperationPass<FuncOp>> createAsyncParallelForPass();
 
+std::unique_ptr<OperationPass<FuncOp>> createAsyncRefCountingPass();
+
 //===----------------------------------------------------------------------===//
 // Registration
 //===----------------------------------------------------------------------===//
 
 /// Generate the code for registering passes.
 #define GEN_PASS_REGISTRATION
 #include "mlir/Dialect/Async/Passes.h.inc"
 
 } // namespace mlir
 
 #endif // MLIR_DIALECT_ASYNC_PASSES_H_

mlir/include/mlir/Dialect/Async/Passes.td

let options = [
     Option<"numConcurrentAsyncExecute", "num-concurrent-async-execute",
       "int32_t", /*default=*/"4",
       "The number of async.execute operations that will be used for concurrent "
       "loop execution.">
   ];
   let dependentDialects = ["async::AsyncDialect", "scf::SCFDialect"];
 }
 
+def AsyncRefCounting : FunctionPass<"async-ref-counting"> {
+  let summary = "Automatic reference counting for Async dialect data types";
+  let constructor = "mlir::createAsyncRefCountingPass()";
+  let dependentDialects = ["async::AsyncDialect"];
+}
+
 #endif // MLIR_DIALECT_ASYNC_PASSES

mlir/include/mlir/ExecutionEngine/AsyncRuntime.h

 typedef struct AsyncGroup MLIR_AsyncGroup;
 
 // Async runtime uses LLVM coroutines to represent asynchronous tasks. Task
 // function is a coroutine handle and a resume function that continue coroutine
 // execution from a suspension point.
 using CoroHandle = void *;           // coroutine handle
 using CoroResume = void (*)(void *); // coroutine resume function
 
+// Async runtime uses reference counting to manage the lifetime of async values
+// (values of async types like tokens, values and groups).
+using RefCountedObjPtr = void *;
+
+// Adds references to reference counted runtime object.
+extern "C" MLIR_ASYNCRUNTIME_EXPORT void
+    mlirAsyncRuntimeAddRef(RefCountedObjPtr, int32_t);
+
+// Drops references from reference counted runtime object.
+extern "C" MLIR_ASYNCRUNTIME_EXPORT void
+    mlirAsyncRuntimeDropRef(RefCountedObjPtr, int32_t);
+
 // Create a new `async.token` in not-ready state.
 extern "C" MLIR_ASYNCRUNTIME_EXPORT AsyncToken *mlirAsyncRuntimeCreateToken();
 
 // Create a new `async.group` in empty state.
 extern "C" MLIR_ASYNCRUNTIME_EXPORT AsyncGroup *mlirAsyncRuntimeCreateGroup();
 
 extern "C" MLIR_ASYNCRUNTIME_EXPORT int64_t
 mlirAsyncRuntimeAddTokenToGroup(AsyncToken *, AsyncGroup *);

mlir/integration_test/Dialect/Async/CPU/test-async-parallel-for-1d.mlir

 // RUN:   mlir-opt %s -async-parallel-for                                      \
+// RUN:               -async-ref-counting                                      \
 // RUN:               -convert-async-to-llvm                                   \
 // RUN:               -convert-scf-to-std                                      \
 // RUN:               -convert-std-to-llvm                                     \
 // RUN: | mlir-cpu-runner                                                      \
 // RUN:  -e entry -entry-point-result=void -O0                                 \
 // RUN:  -shared-libs=%mlir_integration_test_dir/libmlir_runner_utils%shlibext \
 // RUN:  -shared-libs=%mlir_integration_test_dir/libmlir_async_runtime%shlibext\
 // RUN: | FileCheck %s --dump-input=always

mlir/integration_test/Dialect/Async/CPU/test-async-parallel-for-2d.mlir

 // RUN:   mlir-opt %s -async-parallel-for                                      \
+// RUN:               -async-ref-counting                                      \
 // RUN:               -convert-async-to-llvm                                   \
 // RUN:               -convert-scf-to-std                                      \
 // RUN:               -convert-std-to-llvm                                     \
 // RUN: | mlir-cpu-runner                                                      \
 // RUN:  -e entry -entry-point-result=void -O0                                 \
 // RUN:  -shared-libs=%mlir_integration_test_dir/libmlir_runner_utils%shlibext \
 // RUN:  -shared-libs=%mlir_integration_test_dir/libmlir_async_runtime%shlibext\
 // RUN: | FileCheck %s --dump-input=always

mlir/lib/Conversion/AsyncToLLVM/AsyncToLLVM.cpp

 
 // Prefix for functions outlined from `async.execute` op regions.
 static constexpr const char kAsyncFnPrefix[] = "async_execute_fn";
 
 //===----------------------------------------------------------------------===//
 // Async Runtime C API declaration.
 //===----------------------------------------------------------------------===//
 
+static constexpr const char *kAddRef = "mlirAsyncRuntimeAddRef";
+static constexpr const char *kDropRef = "mlirAsyncRuntimeDropRef";
 static constexpr const char *kCreateToken = "mlirAsyncRuntimeCreateToken";
 static constexpr const char *kCreateGroup = "mlirAsyncRuntimeCreateGroup";
 static constexpr const char *kEmplaceToken = "mlirAsyncRuntimeEmplaceToken";
 static constexpr const char *kAwaitToken = "mlirAsyncRuntimeAwaitToken";
 static constexpr const char *kAwaitGroup = "mlirAsyncRuntimeAwaitAllInGroup";
 static constexpr const char *kExecute = "mlirAsyncRuntimeExecute";
 static constexpr const char *kAddTokenToGroup =
     "mlirAsyncRuntimeAddTokenToGroup";
 static constexpr const char *kAwaitAndExecute =
     "mlirAsyncRuntimeAwaitTokenAndExecute";
 static constexpr const char *kAwaitAllAndExecute =
     "mlirAsyncRuntimeAwaitAllInGroupAndExecute";
 
 namespace {
 // Async Runtime API function types.
 struct AsyncAPI {
+  static FunctionType refCountingFunctionType(MLIRContext *ctx) {
+    auto ref = LLVM::LLVMType::getInt8PtrTy(ctx);
+    auto count = IntegerType::get(32, ctx);
+    return FunctionType::get({ref, count}, {}, ctx);
+  }
+
   static FunctionType createTokenFunctionType(MLIRContext *ctx) {
     return FunctionType::get({}, {TokenType::get(ctx)}, ctx);
   }
 
   static FunctionType createGroupFunctionType(MLIRContext *ctx) {
     return FunctionType::get({}, {GroupType::get(ctx)}, ctx);
   }
 

 
 // Adds Async Runtime C API declarations to the module.
 static void addAsyncRuntimeApiDeclarations(ModuleOp module) {
   auto builder = OpBuilder::atBlockTerminator(module.getBody());
 
   MLIRContext *ctx = module.getContext();
   Location loc = module.getLoc();
 
+  if (!module.lookupSymbol(kAddRef))
+    builder.create<FuncOp>(loc, kAddRef,
+                           AsyncAPI::refCountingFunctionType(ctx));
+
+  if (!module.lookupSymbol(kDropRef))
+    builder.create<FuncOp>(loc, kDropRef,
+                           AsyncAPI::refCountingFunctionType(ctx));
+
   if (!module.lookupSymbol(kCreateToken))
     builder.create<FuncOp>(loc, kCreateToken,
                            AsyncAPI::createTokenFunctionType(ctx));
 
   if (!module.lookupSymbol(kCreateGroup))
     builder.create<FuncOp>(loc, kCreateGroup,
                            AsyncAPI::createGroupFunctionType(ctx));
 
rewriter.replaceOpWithNewOp<CallOp>(op, resultTypes, call.callee(),
                                         call.getOperands());
 
     return success();
   }
 };
 } // namespace
 
 //===----------------------------------------------------------------------===//
+// Async reference counting ops lowering (`async.add_ref` and `async.drop_ref`
+// to the corresponding API calls).
+//===----------------------------------------------------------------------===//
+
+namespace {
+
+template <typename RefCountingOp>
+class RefCountingOpLowering : public ConversionPattern {
+public:
+  explicit RefCountingOpLowering(MLIRContext *ctx, StringRef apiFunctionName)
+      : ConversionPattern(RefCountingOp::getOperationName(), 1, ctx),
+        apiFunctionName(apiFunctionName) {}
+
+  LogicalResult
+  matchAndRewrite(Operation *op, ArrayRef<Value> operands,
+                  ConversionPatternRewriter &rewriter) const override {
+    RefCountingOp refCountingOp = cast<RefCountingOp>(op);
+
+    auto i32 = IntegerType::get(32, op->getContext());
+    auto count = IntegerAttr::get(i32, refCountingOp.count());
+    auto countCst = rewriter.create<ConstantOp>(op->getLoc(), i32, count);
+
+    SmallVector<Value, 2> args = {refCountingOp.operand(), countCst};
+    rewriter.replaceOpWithNewOp<CallOp>(op, Type(), apiFunctionName,
+                                        ValueRange(args));
+    return success();
+  }
+
+private:
+  StringRef apiFunctionName;
+};
+
+// async.drop_ref op lowering to mlirAsyncRuntimeDropRef function call.
+class AddRefOpLowering : public RefCountingOpLowering<AddRefOp> {
+public:
+  explicit AddRefOpLowering(MLIRContext *ctx)
+      : RefCountingOpLowering(ctx, kAddRef) {}
+};
+
+// async.create_group op lowering to mlirAsyncRuntimeCreateGroup function call.
+class DropRefOpLowering : public RefCountingOpLowering<DropRefOp> {
+public:
+  explicit DropRefOpLowering(MLIRContext *ctx)
+      : RefCountingOpLowering(ctx, kDropRef) {}
+};
+
+} // namespace
+
+//===----------------------------------------------------------------------===//
 // async.create_group op lowering to mlirAsyncRuntimeCreateGroup function call.
 //===----------------------------------------------------------------------===//
 
 namespace {
 class CreateGroupOpLowering : public ConversionPattern {
 public:
   explicit CreateGroupOpLowering(MLIRContext *ctx)
       : ConversionPattern(CreateGroupOp::getOperationName(), 1, ctx) {}
void ConvertAsyncToLLVMPass::runOnOperation() {
   MLIRContext *ctx = &getContext();
 
   // Convert async dialect types and operations to LLVM dialect.
   AsyncRuntimeTypeConverter converter;
   OwningRewritePatternList patterns;
 
   populateFuncOpTypeConversionPattern(patterns, ctx, converter);
   patterns.insert<CallOpOpConversion>(ctx);
+  patterns.insert<AddRefOpLowering, DropRefOpLowering>(ctx);
   patterns.insert<CreateGroupOpLowering, AddToGroupOpLowering>(ctx);
   patterns.insert<AwaitOpLowering, AwaitAllOpLowering>(ctx, outlinedFunctions);
 
   ConversionTarget target(*ctx);
+  target.addLegalOp<ConstantOp>();
   target.addLegalDialect<LLVM::LLVMDialect>();
   target.addIllegalDialect<AsyncDialect>();
   target.addDynamicallyLegalOp<FuncOp>(
       [&](FuncOp op) { return converter.isSignatureLegal(op.getType()); });
   target.addDynamicallyLegalOp<CallOp>(
       [&](CallOp op) { return converter.isLegal(op.getResultTypes()); });
 
   if (failed(applyPartialConversion(module, target, std::move(patterns))))
     signalPassFailure();
 }
 } // namespace
 
 std::unique_ptr<OperationPass<ModuleOp>> mlir::createConvertAsyncToLLVMPass() {
   return std::make_unique<ConvertAsyncToLLVMPass>();
 }

mlir/lib/Dialect/Async/Transforms/AsyncRefCounting.cpp

+//===- AsyncRefCounting.cpp - Implementation of Async Ref Counting --------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements automatic reference counting for Async dialect data
+// types.
+//
+//===----------------------------------------------------------------------===//
+
+#include "PassDetail.h"
+#include "mlir/Dialect/Async/IR/Async.h"
+#include "mlir/Dialect/Async/Passes.h"
+#include "mlir/Dialect/StandardOps/IR/Ops.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+#include "llvm/ADT/SmallSet.h"
+
+using namespace mlir;
+using namespace mlir::async;
+
+#define DEBUG_TYPE "async-ref-counting"
+
+namespace {
+
+struct AsyncRefCountingPass
+    : public AsyncRefCountingBase<AsyncRefCountingPass> {
+  AsyncRefCountingPass() = default;
+  void runOnFunction() override;
+};
+
+} // namespace
+
+/// Returns true if the type is reference counted. All async dialect types a
+/// reference counted at runtime.
+bool isRefCounted(Type type) {
+  return type.isa<TokenType, ValueType, GroupType>();
+}
+
+/// Returns true if the operation `op` supports async reference counting.
+///
+/// It is the async value consumer responsibility to drop the reference count
+/// when the value is no longer needed. If the async value passed to the
+/// consumer that is not aware of reference counting, this async value will leak
+/// at runtime.
+bool isSupportedConsumer(Operation *op) {
+  // Return operation transfers ownership to the caller.
+  if (isa<ReturnOp>(op))
+    return true;
+
+  // Async dialect operations correctly handle reference counted values.
+  if (isa<ExecuteOp, AwaitOp, AwaitAllOp, AddToGroupOp>(op))
+    return true;
+
+  return false;
+}
+
+/// Returns the statically know number of instances for all operations in the
+/// attached region (the number of times each operation inside the attached
+/// region will be executed).
+Optional<int32_t> getStaticNumberOfInstances(Operation *op) {
+  assert(!op->getRegions().empty() && "operation must have attached regions");
+
+  // `async.execute` will execute all operations exactly once.
+  if (isa<ExecuteOp>(op))
+    return 1;
+
+  // TODO: Loops with statically known bounds have statically know number of
+  // operation instances in the loop body.
+  return None;
+}
+
+/// Returns the statically known number of instances of the `user` operation
+/// that consumes async values produced by the `owner` operation. Returns empty
+/// optional if the number of instances is dynamic.
+///
+/// Examples:
+///
+/// 1. `owner` and `user` are in the same region.
+///
+///    %token = ...
+///    "use"(%token): (!async.token) -> ()
+///
+///    Number of instances: 1
+///
+/// 2. `user` is inside the region with statically known execution.
+///
+///    %token = ...
+///    async.execute {
+///      "use"(%token): (!async.token) -> ()
+///    }
+///
+///    Number of instances: 1 (async.execute will execute all operations in the
+///                            attached body region)
+///
+/// 3. `user` is inside the dynamic control flow operation (e.g. `scf.if`,
+///    `scf.for` or `scf.parallel`).
+///
+///    %token = ...
+///    scf.if %condition {
+///      "use"(%token): (!async.token) -> ()
+///    } else {
+///      "some_other_operation"(): () -> ()
+///    }
+///
+///    Number of instances: <unknown> (it is not statically known if the
+///                                    execution will go into the first region).
+///
+/// If we know the number of `user` instances statically, we can increment the
+/// reference count for the async value produced by the `owner`:
+///
+///   %token = ...
+///   async.add_ref %token {count = <static-number-of-instances - 1>}
+///
+/// For dynamic instances we can safely add a reference only in the same region
+/// as the `user` parent region. See details below.
+Optional<int32_t> getStaticNumberOfInstances(Operation *owner,
+                                             Operation *user) {
+  int32_t result = 1;
+
+  Operation *ownerParent = owner->getParentOp();
+  Operation *userParent = user->getParentOp();
+
+  while (ownerParent != userParent) {
+    if (auto num = getStaticNumberOfInstances(userParent))
+      result *= *num;
+    else
+      return None;
+
+    userParent = userParent->getParentOp();
+  }
+
+  return result;
+}
+
+namespace {
+struct DynamicInstanceProperties {
+  /// The static number instances of the `user` operation inside the dynamic
+  /// operation.
+  int32_t staticNumberOfInstances;
+
+  /// The block owned by the region attached to the dynamic operation.
+  Block *dynamicBlock;
+
+  /// Operation in the same region as async value `owner` that contains the
+  /// dynamic operation (can be the dynamic operation itself). We'll use this as
+  /// an anchor to add explicit `async.drop_ref` operation after it.
+  Operation *refCountAnchor;
+};
+} // namespace
+
+/// Returns the dynamic instance properties of the `user` operation that
+/// consumes async value produced by the `owner` operation.
+///
+/// Example:
+///
+///  %token = ...
+///  scf.for %i = %c0 to %c2 step %c1 {
+///    scf.if %cond {
+///      async.execute {
+///        async.await %token : !async.token
+///      }
+///    }
+///  }
+///
+/// Innermost dynamic operation that contains the async value user `async.await`
+/// is `scf.for`. Inside this operation statically known number of uses is 1.
+///
+/// Dynamic reference counting must be added to the block that owns the
+/// operation that has statically known number of instances of the async uses.
+///
+/// With automatic reference counting this should become:
+///
+///  %token = ...
+///
+///  // Add a reference count statically because we know that we have one
+///  dynamic
+///  // use inside the `scf.for` operation.
+///  async.add_ref %token {count = 1 : i32} : !async.token
+///
+///  scf.for %i = %c0 to %c2 step %c1 {
+///    scf.if %cond {
+///      // Add a reference count dynamically.
+///      async.add_ref %token {count = 1 : i32} : !async.token
+///      async.execute {
+///        async.await %token : !async.token
+///      }
+///    }
+///  }
+///
+///  // Explicitly drop the static reference that we added on behalf of
+///  `scf.for` operation.
+///  async.drop_ref %token {count = 1 : i32} : !async.token
+DynamicInstanceProperties getDynamicInstanceProperties(Operation *owner,
+                                                       Operation *user) {
+  assert(!getStaticNumberOfInstances(owner, user).hasValue() &&
+         "user must have dynamic number of instance");
+
+  // Compute the number of static instances before we reach the first dynamic
+  // parent.
+  int32_t numberOfStaticInstances = 1;
+
+  // Operation with statically known number of instances of the `user`
+  // operation (can be `user` operation itself).
+  Operation *staticUser = user;
+
+  Operation *ownerParent = owner->getParentOp();
+  Operation *userParent = user->getParentOp();
+
+  // Find the parent with statically known number of instances.
+  while (ownerParent != userParent) {
+    if (auto n = getStaticNumberOfInstances(userParent))
+      numberOfStaticInstances *= *n;
+    else
+      break;
+
+    staticUser = userParent;
+    userParent = userParent->getParentOp();
+  }
+
+  assert(ownerParent != userParent && "did not find dynamic operation");
+
+  // Block that owns operation with statically known number of user instances,
+  // but the parent has dynamic nubmer of instances.
+  Block *dynamicBlock = staticUser->getBlock();
+  Operation *dynamicOperation = userParent;
+
+  assert(dynamicBlock->getParentOp() == dynamicOperation);
+
+  // Find the operation that owns the operation with dynamic nubmer of
+  // instances and has the same parent as the `owner`.
+  Operation *dynamicParent = dynamicOperation->getParentOp();
+  while (ownerParent != dynamicParent) {
+    dynamicOperation = dynamicParent;
+    dynamicParent = dynamicOperation->getParentOp();
+  }
+
+  return {numberOfStaticInstances, dynamicBlock, dynamicOperation};
+}
+
+LogicalResult addAutomaticRefCounting(OpResult result) {
+  Operation *op = result.getOwner();
+  MLIRContext *ctx = op->getContext();
+
+  Location loc = result.getLoc();
+
+  OpBuilder builder(op);
+  builder.setInsertionPointAfter(op);
+
+  auto i32 = IntegerType::get(32, ctx);
+
+  // Drop ref count -1 if the result has no users.
+  if (result.getUsers().empty()) {
+    builder.create<DropRefOp>(loc, result, IntegerAttr::get(i32, 1));
+    return success();
+  }
+
+  // Verify that all users support automatic reference counting.
+  for (Operation *user : result.getUsers()) {
+    if (!isSupportedConsumer(user)) {
+      op->emitError() << "result #" << result.getResultNumber()
+                      << " passed to the operation that does not support "
+                         "automatic async reference counting: "
+                      << user->getName();
+      return failure();
+    }
+  }
+
+  // The number of statically known uses of the `result`.
+  int32_t staticInstances = 0;
+
+  // Collect properties of the dynamic uses of the `result`.
+  SmallVector<DynamicInstanceProperties, 8> dynamicInstances;
+
+  for (const OpOperand &use : result.getUses()) {
+    Operation *user = use.getOwner();
+
+    // Check if we know the number of user instances statically.
+    if (auto knownStatically = getStaticNumberOfInstances(op, user)) {
+      staticInstances += *knownStatically;
+      continue;
+    }
+
+    // Collect dynamic instance properties otherwise.
+    DynamicInstanceProperties props = getDynamicInstanceProperties(op, user);
+    dynamicInstances.push_back(props);
+  }
+
+  // Remove redundant reference counting from the same anchors ...
+  llvm::SmallSet<Operation *, 8> anchors;
+  // ... and aggregate the number of static instances per dynamic block.
+  llvm::DenseMap<Block *, int32_t> blockStaticCounts;
+
+  for (DynamicInstanceProperties &props : dynamicInstances) {
+    blockStaticCounts[props.dynamicBlock] += props.staticNumberOfInstances;
+    anchors.insert(props.refCountAnchor);
+  }
+
+  // We'll add +1 reference for each static instance of the user operation, and
+  // also +1 for every dynamic instance anchor operation. Adding references for
+  // dynamic instance anchors is required to keep reference counted objects
+  // alive until the control flow reaches `async.add_ref` operation inside the
+  // dynamic region.
+  int32_t useCount = staticInstances + anchors.size();
+
+  // Add +1 reference for each result use to eventually drop the reference count
+  // to zero.
+  if (useCount > 1)
+    builder.create<AddRefOp>(loc, result, IntegerAttr::get(i32, useCount - 1));
+
+  // Drop reference count immediately after the anchor operation.
+  for (Operation *anchor : anchors) {
+    builder.setInsertionPointAfter(anchor);
+    builder.create<DropRefOp>(loc, result, IntegerAttr::get(i32, 1));
+  }
+
+  // Add statically know references at the beginning of the dynamic block.
+  for (auto &kv : blockStaticCounts) {
+    Block *block = kv.first;
+    int32_t count = kv.second;
+    builder.setInsertionPointToStart(block);
+    builder.create<AddRefOp>(loc, result, IntegerAttr::get(i32, count));
+  }
+
+  return success();
+}
+
+LogicalResult addAutomaticRefCounting(Operation *op) {
+  for (unsigned i = 0; i < op->getNumResults(); ++i) {
+    Type resultType = op->getResultTypes()[i];
+    if (!isRefCounted(resultType))
+      continue;
+
+    if (failed(addAutomaticRefCounting(op->getResult(i))))
+      return failure();
+  }
+  return success();
+}
+
+void AsyncRefCountingPass::runOnFunction() {
+  FuncOp func = getFunction();
+
+  // Add `async.add_ref` operations to match the number of uses for each async
+  // value.
+  WalkResult walkResult = func.walk([](Operation *op) -> WalkResult {
+    if (failed(addAutomaticRefCounting(op)))
+      return WalkResult::interrupt();
+    return WalkResult::advance();
+  });
+
+  if (walkResult.wasInterrupted())
+    signalPassFailure();
+}
+
+std::unique_ptr<OperationPass<FuncOp>> mlir::createAsyncRefCountingPass() {
+  return std::make_unique<AsyncRefCountingPass>();
+}

mlir/lib/Dialect/Async/Transforms/CMakeLists.txt

 add_mlir_dialect_library(MLIRAsyncTransforms
   AsyncParallelFor.cpp
+  AsyncRefCounting.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/Async
 
   DEPENDS
   MLIRAsyncPassIncGen
 
   LINK_LIBS PUBLIC
   MLIRIR
   MLIRAsync
   MLIRSCF
   MLIRPass
   MLIRTransforms
   MLIRTransformUtils
 )

mlir/lib/ExecutionEngine/AsyncRuntime.cpp

 //
 //===----------------------------------------------------------------------===//
 
 #include "mlir/ExecutionEngine/AsyncRuntime.h"
 
 #ifdef MLIR_ASYNCRUNTIME_DEFINE_FUNCTIONS
 
 #include <atomic>
+#include <cassert>
 #include <condition_variable>
 #include <functional>
 #include <iostream>
 #include <mutex>
 #include <thread>
 #include <vector>
 
 //===----------------------------------------------------------------------===//
 // Async runtime API.
 //===----------------------------------------------------------------------===//
 
-struct AsyncToken {
-  bool ready = false;
+namespace {
+
+// Forward declare class defined below.
+class RefCounted;
+
+// -------------------------------------------------------------------------- //
+// AsyncRuntime orchestrates all async operations and Async runtime API is built
+// on top of the default runtime instance.
+// -------------------------------------------------------------------------- //
+
+class AsyncRuntime {
+public:
+  AsyncRuntime() : numRefCountedObjects(0) {}
+
+  ~AsyncRuntime() {
+    assert(getNumRefCountedObjects() == 0 &&
+           "all ref counted objects must be destroyed");
+  }
+
+  int32_t getNumRefCountedObjects() {
+    return numRefCountedObjects.load(std::memory_order_relaxed);
+  }
+
+private:
+  friend class RefCounted;
+
+  // Count the total number of reference counted objects in this instance
+  // of an AsyncRuntime. For debugging purposes only.
+  void addNumRefCountedObjects() {
+    numRefCountedObjects.fetch_add(1, std::memory_order_relaxed);
+  }
+  void dropNumRefCountedObjects() {
+    numRefCountedObjects.fetch_sub(1, std::memory_order_relaxed);
+  }
+
+  std::atomic<int32_t> numRefCountedObjects;
+};
+
+// Returns the default per-process instance of an async runtime.
+AsyncRuntime *getDefaultAsyncRuntimeInstance() {
+  static auto runtime = std::make_unique<AsyncRuntime>();
+  return runtime.get();
+}
+
+// -------------------------------------------------------------------------- //
+// A base class for all reference counted objects created by the async runtime.
+// -------------------------------------------------------------------------- //
+
+class RefCounted {
+public:
+  RefCounted(AsyncRuntime *runtime, int32_t refCount = 1)
+      : runtime(runtime), refCount(refCount) {
+    runtime->addNumRefCountedObjects();
+  }
+
+  virtual ~RefCounted() {
+    assert(refCount.load() == 0 && "reference count must be zero");
+    runtime->dropNumRefCountedObjects();
+  }
+
+  RefCounted(const RefCounted &) = delete;
+  RefCounted &operator=(const RefCounted &) = delete;
+
+  void addRef(int32_t count = 1) { refCount.fetch_add(count); }
+
+  void dropRef(int32_t count = 1) {
+    int32_t previous = refCount.fetch_sub(count);
+    assert(previous >= count && "reference count should not go below zero");
+    if (previous == count)
+      destroy();
+  }
+
+protected:
+  virtual void destroy() { delete this; }
+
+private:
+  AsyncRuntime *runtime;
+  std::atomic<int32_t> refCount;
+};
+
+} // namespace
+
+struct AsyncToken : public RefCounted {
+  // AsyncToken created with a reference count of 2 because it will be returned
+  // to the `async.execute` caller and also will be later on emplaced by the
+  // asynchronously executed task. If the caller immediately will drop its
+  // reference we must ensure that the token will be alive until the
+  // asynchronous operation is completed.
+  AsyncToken(AsyncRuntime *runtime) : RefCounted(runtime, /*count=*/2) {}
+
+  // Internal state below guarded by a mutex.
   std::mutex mu;
   std::condition_variable cv;
+
+  bool ready = false;
   std::vector<std::function<void()>> awaiters;
 };
 
-struct AsyncGroup {
-  std::atomic<int> pendingTokens{0};
-  std::atomic<int> rank{0};
+struct AsyncGroup : public RefCounted {
+  AsyncGroup(AsyncRuntime *runtime)
+      : RefCounted(runtime), pendingTokens(0), rank(0) {}
+
+  std::atomic<int> pendingTokens;
+  std::atomic<int> rank;
+
+  // Internal state below guarded by a mutex.
   std::mutex mu;
   std::condition_variable cv;
+
   std::vector<std::function<void()>> awaiters;
 };
 
+// Adds references to reference counted runtime object.
+extern "C" MLIR_ASYNCRUNTIME_EXPORT void
+mlirAsyncRuntimeAddRef(RefCountedObjPtr ptr, int32_t count) {
+  RefCounted *refCounted = static_cast<RefCounted *>(ptr);
+  refCounted->addRef(count);
+}
+
+// Drops references from reference counted runtime object.
+extern "C" MLIR_ASYNCRUNTIME_EXPORT void
+mlirAsyncRuntimeDropRef(RefCountedObjPtr ptr, int32_t count) {
+  RefCounted *refCounted = static_cast<RefCounted *>(ptr);
+  refCounted->dropRef(count);
+}
+
 // Create a new `async.token` in not-ready state.
 extern "C" MLIR_ASYNCRUNTIME_EXPORT AsyncToken *mlirAsyncRuntimeCreateToken() {
-  AsyncToken *token = new AsyncToken;
+  AsyncToken *token = new AsyncToken(getDefaultAsyncRuntimeInstance());
   return token;
 }
 
 // Create a new `async.group` in empty state.
 extern "C" MLIR_ASYNCRUNTIME_EXPORT AsyncGroup *mlirAsyncRuntimeCreateGroup() {
-  AsyncGroup *group = new AsyncGroup;
+  AsyncGroup *group = new AsyncGroup(getDefaultAsyncRuntimeInstance());
   return group;
 }
 
 extern "C" MLIR_ASYNCRUNTIME_EXPORT int64_t
 mlirAsyncRuntimeAddTokenToGroup(AsyncToken *token, AsyncGroup *group) {
   std::unique_lock<std::mutex> lockToken(token->mu);
   std::unique_lock<std::mutex> lockGroup(group->mu);
 
+  // Get the rank of the token inside the group before we drop the reference.
+  int rank = group->rank.fetch_add(1);
   group->pendingTokens.fetch_add(1);
 
-  auto onTokenReady = [group]() {
+  auto onTokenReady = [group, token]() {
     // Run all group awaiters if it was the last token in the group.
     if (group->pendingTokens.fetch_sub(1) == 1) {
       group->cv.notify_all();
       for (auto &awaiter : group->awaiters)
         awaiter();
     }
+
+    // We no longer need the token or the group, drop references on them.
+    group->dropRef();
+    token->dropRef();
   };
 
   if (token->ready)
     onTokenReady();
   else
     token->awaiters.push_back([onTokenReady]() { onTokenReady(); });
 
-  return group->rank.fetch_add(1);
+  return rank;
 }
 
 // Switches `async.token` to ready state and runs all awaiters.
 extern "C" MLIR_ASYNCRUNTIME_EXPORT void
 mlirAsyncRuntimeEmplaceToken(AsyncToken *token) {
   std::unique_lock<std::mutex> lock(token->mu);
   token->ready = true;
   token->cv.notify_all();
   for (auto &awaiter : token->awaiters)
     awaiter();
+
+  token->dropRef();
 }
 
 extern "C" MLIR_ASYNCRUNTIME_EXPORT void
 mlirAsyncRuntimeAwaitToken(AsyncToken *token) {
   std::unique_lock<std::mutex> lock(token->mu);
   if (!token->ready)
     token->cv.wait(lock, [token] { return token->ready; });
+
+  token->dropRef();
 }
 
 extern "C" MLIR_ASYNCRUNTIME_EXPORT void
 mlirAsyncRuntimeAwaitAllInGroup(AsyncGroup *group) {
   std::unique_lock<std::mutex> lock(group->mu);
   if (group->pendingTokens != 0)
     group->cv.wait(lock, [group] { return group->pendingTokens == 0; });
+
+  group->dropRef();
 }
 
 extern "C" MLIR_ASYNCRUNTIME_EXPORT void
 mlirAsyncRuntimeExecute(CoroHandle handle, CoroResume resume) {
 #if LLVM_ENABLE_THREADS
   std::thread thread([handle, resume]() { (*resume)(handle); });
   thread.detach();
 #else
   (*resume)(handle);
 #endif
 }
 
 extern "C" MLIR_ASYNCRUNTIME_EXPORT void
 mlirAsyncRuntimeAwaitTokenAndExecute(AsyncToken *token, CoroHandle handle,
                                      CoroResume resume) {
   std::unique_lock<std::mutex> lock(token->mu);
 
-  auto execute = [handle, resume]() {
+  auto execute = [handle, resume, token]() {
+    token->dropRef();
     mlirAsyncRuntimeExecute(handle, resume);
   };
 
   if (token->ready)
     execute();
   else
     token->awaiters.push_back([execute]() { execute(); });
 }
 
 extern "C" MLIR_ASYNCRUNTIME_EXPORT void
 mlirAsyncRuntimeAwaitAllInGroupAndExecute(AsyncGroup *group, CoroHandle handle,
                                           CoroResume resume) {
   std::unique_lock<std::mutex> lock(group->mu);
 
-  auto execute = [handle, resume]() {
+  auto execute = [handle, resume, group]() {
+    group->dropRef();
     mlirAsyncRuntimeExecute(handle, resume);
   };
 
   if (group->pendingTokens == 0)
     execute();
   else
     group->awaiters.push_back([execute]() { execute(); });
 }

mlir/test/Conversion/AsyncToLLVM/convert-to-llvm.mlir

 // RUN: mlir-opt %s -split-input-file -convert-async-to-llvm | FileCheck %s
 
+// CHECK-LABEL: reference_counting
+func @reference_counting(%arg0: !async.token) {
+  // CHECK: %[[C2:.*]] = constant 2 : i32
+  // CHECK: call @mlirAsyncRuntimeAddRef(%arg0, %[[C2]])
+  async.add_ref %arg0 {count = 2 : i32} : !async.token
+
+  // CHECK: %[[C1:.*]] = constant 1 : i32
+  // CHECK: call @mlirAsyncRuntimeDropRef(%arg0, %[[C1]])
+  async.drop_ref %arg0 {count = 1 : i32} : !async.token
+
+  return
+}
+
+// -----
+
 // CHECK-LABEL: execute_no_async_args
 func @execute_no_async_args(%arg0: f32, %arg1: memref<1xf32>) {
   // CHECK: %[[TOKEN:.*]] = call @async_execute_fn(%arg0, %arg1)
   %token = async.execute {
     %c0 = constant 0 : index
     store %arg0, %arg1[%c0] : memref<1xf32>
     async.yield
   }

mlir/test/Dialect/Async/async-ref-counting.mlir

+// RUN: mlir-opt %s -async-ref-counting | FileCheck %s
+
+// CHECK-LABEL: @token_no_uses
+func @token_no_uses() {
+  // CHECK: %[[TOKEN:.*]] = async.execute
+  // CHECK: async.drop_ref %[[TOKEN]] {count = 1 : i32}
+  %token = async.execute {
+    async.yield
+  }
+  return
+}
+
+// CHECK-LABEL: @token_return
+func @token_return() -> !async.token {
+  // CHECK: %[[TOKEN:.*]] = async.execute
+  %token = async.execute {
+    async.yield
+  }
+  // CHECK: return %token
+  return %token : !async.token
+}
+
+// CHECK-LABEL: @token_with_await
+func @token_with_await() -> !async.token {
+  // CHECK: %[[TOKEN:.*]] = async.execute
+  // CHECK: async.add_ref %[[TOKEN]] {count = 1 : i32}
+  %token = async.execute {
+    async.yield
+  }
+  async.await %token : !async.token
+  // CHECK: return %token
+  return %token : !async.token
+}
+
+// CHECK-LABEL: @token_capture
+func @token_capture() {
+  // CHECK: %[[TOKEN:.*]] = async.execute
+  // CHECK: async.add_ref %[[TOKEN]] {count = 1 : i32}
+  %token = async.execute {
+    async.yield
+  }
+
+  // CHECK: %[[TOKEN_0:.*]] = async.execute
+  %token_0 = async.execute {
+    async.await %token : !async.token
+    async.yield
+  }
+
+  // CHECK: async.await %[[TOKEN]]
+  // CHECK: async.await %[[TOKEN_0]]
+  async.await %token : !async.token
+  async.await %token_0 : !async.token
+
+  // CHECK: return
+  return
+}
+
+// CHECK-LABEL: @token_dependency
+func @token_dependency() {
+  // CHECK: %[[TOKEN:.*]] = async.execute
+  // CHECK: async.add_ref %[[TOKEN]] {count = 1 : i32}
+  %token = async.execute {
+    async.yield
+  }
+
+  // CHECK: %[[TOKEN_0:.*]] = async.execute
+  %token_0 = async.execute[%token] {
+    async.yield
+  }
+
+  // CHECK: async.await %[[TOKEN]]
+  // CHECK: async.await %[[TOKEN_0]]
+  async.await %token : !async.token
+  async.await %token_0 : !async.token
+
+  // CHECK: return
+  return
+}
+
+// CHECK-LABEL: @value_operand
+func @value_operand() -> f32 {
+  // CHECK: %[[TOKEN:.*]], %[[RESULTS:.*]] = async.execute
+  // CHECK: async.add_ref %[[RESULTS]] {count = 1 : i32}
+  // CHECK: async.add_ref %[[TOKEN]] {count = 1 : i32}
+  %token, %results = async.execute -> !async.value<f32> {
+    %0 = constant 0.0 : f32
+    async.yield %0 : f32
+  }
+
+  // CHECK: %[[TOKEN_0:.*]] = async.execute
+  %token_0 = async.execute[%token](%results as %arg0 : !async.value<f32>)  {
+    async.yield
+  }
+
+  // CHECK: async.await %[[TOKEN]]
+  // CHECK: async.await %[[TOKEN_0]]
+  async.await %token : !async.token
+  async.await %token_0 : !async.token
+
+  // CHECK: async.await %[[RESULTS]]
+  %0 = async.await %results : !async.value<f32>
+
+  // CHECK: return
+  return %0 : f32
+}
+
+// CHECK-LABEL: @async_group
+func @async_group() {
+  // CHECK: %[[GROUP:.*]] = async.create_group
+  // CHECK: async.add_ref %[[GROUP]] {count = 1 : i32} : !async.group
+  %0 = async.create_group
+
+  // CHECK: %[[TOKEN:.*]] = async.execute
+  // CHECK: %[[TOKEN_0:.*]] = async.execute
+  %token = async.execute { async.yield }
+  %token_0 = async.execute { async.yield }
+
+  // CHECK: async.add_to_group %[[TOKEN]], %[[GROUP]]
+  // CHECK: async.add_to_group %[[TOKEN_0]], %[[GROUP]]
+  %1 = async.add_to_group %token, %0 : !async.token
+  %2 = async.add_to_group %token_0, %0 : !async.token
+
+  // CHECK: return
+  return
+}
+
+// CHECK-LABEL: @capture_by_scf_if
+func @capture_by_scf_if(%arg0 : i1) {
+  %token = async.execute { async.yield }
+
+  scf.if %arg0 {
+    // CHECK: async.add_ref %token {count = 2 : i32}
+    async.await %token : !async.token
+    async.await %token : !async.token
+  } else {
+    // CHECK: async.add_ref %token {count = 1 : i32}
+    async.await %token : !async.token
+  }
+  // CHECK: async.drop_ref %token {count = 1 : i32}
+
+  return
+}
+
+// CHECK-LABEL: @capture_by_scf_if_with_async_execute
+func @capture_by_scf_if_with_async_execute(%arg0 : i1) {
+  %token = async.execute { async.yield }
+
+  // `async.await` from the `async.execute` rolled up to the first
+  // operation with dynamic number of instances.
+  scf.if %arg0 {
+    // CHECK: async.add_ref %token {count = 2 : i32}
+    async.execute {
+      async.await %token : !async.token
+      async.await %token : !async.token
+      async.yield
+    }
+  } else {
+    // CHECK: async.add_ref %token {count = 1 : i32}
+    async.await %token : !async.token
+  }
+  // CHECK: async.drop_ref %token {count = 1 : i32}
+
+  return
+}
+
+// CHECK-LABEL: @capture_by_scf_for
+func @capture_by_scf_for() {
+  %token = async.execute { async.yield }
+
+  %c0 = constant 0 : index
+  %c1 = constant 1 : index
+  %c2 = constant 0 : index
+
+  scf.for %i = %c0 to %c2 step %c1 {
+    // CHECK: async.add_ref %token {count = 1 : i32}
+    async.await %token : !async.token
+  }
+  // CHECK: async.drop_ref %token {count = 1 : i32}
+
+  return
+}

mlir/test/Dialect/Async/ops.mlir

func @create_group_and_await_all(%arg0: !async.token, %arg1: !async.value<f32>) -> index {
   // CHECK: async.add_to_group %arg1
   %1 = async.add_to_group %arg0, %0 : !async.token
   %2 = async.add_to_group %arg1, %0 : !async.value<f32>
   async.await_all %0
 
   %3 = addi %1, %2 : index
   return %3 : index
 }
+
+// CHECK-LABEL: @add_ref
+func @add_ref(%arg0: !async.token) {
+  // CHECK: async.add_ref %arg0 {count = 1 : i32}
+  async.add_ref %arg0 {count = 1 : i32} : !async.token
+  return
+}
+
+// CHECK-LABEL: @drop_ref
+func @drop_ref(%arg0: !async.token) {
+  // CHECK: async.drop_ref %arg0 {count = 1 : i32}
+  async.drop_ref %arg0 {count = 1 : i32} : !async.token
+  return
+}

mlir/test/Dialect/Async/verify.mlir

 }
 
 // -----
 
 func @wrong_async_await_result_type(%arg0: !async.value<f32>) {
   // expected-error @+1 {{'async.await' op result type 'f64' does not match async value type 'f32'}}
   %0 = "async.await"(%arg0): (!async.value<f32>) -> f64
 }
+
+// -----
+
+func @wrong_add_ref_count(%arg0: !async.token) {
+  // expected-error @+1 {{'async.add_ref' op reference count must be greater than 0}}
+  async.add_ref %arg0 {count = 0 : i32} : !async.token
+}
+
+// -----
+
+func @wrong_drop_ref_count(%arg0: !async.token) {
+  // expected-error @+1 {{'async.drop_ref' op reference count must be greater than 0}}
+  async.drop_ref %arg0 {count = 0 : i32} : !async.token
+}

mlir/test/mlir-cpu-runner/async-group.mlir

-// RUN:   mlir-opt %s -convert-async-to-llvm                                   \
+// RUN:   mlir-opt %s -async-ref-counting                                      \
+// RUN:               -convert-async-to-llvm                                   \
 // RUN:               -convert-std-to-llvm                                     \
 // RUN: | mlir-cpu-runner                                                      \
 // RUN:     -e main -entry-point-result=void -O0                               \
 // RUN:     -shared-libs=%linalg_test_lib_dir/libmlir_c_runner_utils%shlibext  \
 // RUN:     -shared-libs=%linalg_test_lib_dir/libmlir_runner_utils%shlibext    \
 // RUN:     -shared-libs=%linalg_test_lib_dir/libmlir_async_runtime%shlibext   \
 // RUN: | FileCheck %s