Introduction to Dart VM


This document is work in progress and is currently being written. Please contact Vyacheslav Egorov (by mail or @mraleph) if you have any questions, suggestions, bug reports. Last update: January 6 2019

Purpose of this document

This document is intended as a reference for new members of the Dart VM team, potential external contributors or just anybody interested in VM internals. It starts with a high-level overview of the Dart VM and then proceeds to describe various components of the VM in more details.

Dart VM is a collection of components for executing Dart code natively. Notably it includes the following:

The name "Dart VM" is historical. Dart VM is a virtual machine in a sense that it provides an execution environment for a high-level programming language, however it does not imply that Dart is always interpreted or JIT-compiled, when executing on Dart VM. For example, Dart code can be compiled into machine code using Dart VM AOT pipeline and then executed within a stripped version of the Dart VM, called precompiled runtime, which does not contain any compiler components and is incapable of loading Dart source code dynamically.

1 How Dart VM runs your code?

Dart VM has multiple ways to execute the code, for example:

However the main difference between these lies in when and how VM converts Dart source code to executable code. The runtime environment that facilitates the execution remains the same.


Any Dart code within the VM is running within some isolate, which can be best described as an isolated Dart universe with its own memory (heap) and usually with its own thread of control (mutator thread). There can be many isolates executing Dart code concurrently, but they cannot share any state directly and can only communicate by message passing through ports (not to be confused with network ports!).

The relationship between OS threads and isolates is a bit blurry and highly dependent on how VM is embedded into an application. Only the following is guaranteed:

However the same OS thread can first enter one isolate, execute Dart code, then leave this isolate and enter another isolate. Alternatively many different OS threads can enter an isolate and execute Dart code inside it, just not simultaneously.

In addition to a single mutator thread an isolate can also be associated with multiple helper threads, for example:

Internally VM uses a thread pool (ThreadPool) to manage OS threads and the code is structured around ThreadPool::Task concept rather than around a concept of an OS thread. For example, instead of spawning a dedicated thread to perform background sweeping after a GC VM posts a SweeperTask to the global VM thread pool and thread pool implementation either selects an idling thread or spawns a new thread if no threads are available. Similarly the default implementation of an event loop for isolate message processing does not actually spawn a dedicated event loop thread, instead it posts a MessageHandlerTask to the thread pool whenever a new message arrives.

Source to read

Class Isolate represents an isolate, class Heap - isolate's heap. Class Thread describes the state associated with a thread attached to an isolate. Note that the name Thread is somewhat confusing because all OS threads attached to the same isolate as a mutator would reuse the same Thread instance. See Dart_RunLoop and MessageHandler for the default implementation of an isolate's message handling.

1.1 Running from source via JIT.

This section tries to cover what happens when you try to execute Dart from the command line:

// hello.dart
main() => print('Hello, World!');
$ dart hello.dart
Hello, World!

Since Dart 2 VM no longer has the ability to directly execute Dart from raw source, instead VM expects to be given Kernel binaries (also called dill files) which contain serialized Kernel ASTs. The task of translating Dart source into Kernel AST is handled by the common front-end (CFE) written in Dart and shared between different Dart tools (e.g. VM, dart2js, Dart Dev Compiler).

Dart to Kernel

To preserve convenience of executing Dart directly from source standalone dart executable hosts a helper isolate called kernel service, which handles compilation of Dart source into Kernel. VM then would run resulting Kernel binary.

Running from Source in Dart 2

However this setup is not the only way to arrange CFE and VM to run Dart code. For example Flutter completely separates compilation to Kernel and execution from Kernel by putting them onto different devices: compilation happens on the developer machine (host) and execution is handled on the target mobile device, which receives Kernel binaries send to it by flutter tool.

Dart to Kernel

Note that flutter tool does not handle parsing of Dart itself - instead it spawns another persistent process frontend_server, which is essentially a thin wrapper around CFE and a some Flutter specific Kernel-to-Kernel transformations. frontend_server compiles Dart source into Kernel files, which flutter tool then sends to the device. Persistence of the frontend_server process comes into play when developer requests hot reload: in this case frontend_server can reuse CFE state from the previous compilation and only recompile parts which actually changed.

Once Kernel binary is loaded into the VM it is parsed to create objects representing various program entities. However this is done lazily: at first only basic information about libraries and classes is loaded. Each entity originating from a Kernel binary keeps a pointer back to the binary, so that later more information can be loaded as needed. We use Raw... prefix whenever we talk about specific objects allocated internally by the VM. This follows VM own naming convention: layout of internal VM objects is defined using C++ classes with names starting with Raw in the header file raw_object.h. For example RawClass is a VM object describing Dart class, RawField is a VM object describing a Dart field within a Dart class and so on. We will return to this in a section covering runtime system and object model.

Kernel Loading. Stage 1

Information about the class is fully deserialized only when runtime later needs it (e.g. to lookup a class member, to allocate an instance, etc). At this stage class members are read from the Kernel binary. However full function bodies are not deserialized at this stage, only their signatures.

Kernel Loading. Stage 2

At this point enough information is loaded from Kernel binary for runtime to successfully resolve and invoke methods. For example it could resolve and invoke main function from a library.

Source to read

package:kernel/ast.dart defines classes describing the Kernel AST. package:front_end handles parsing Dart source and building Kernel AST from it. kernel::KernelLoader::LoadEntireProgram is an entry point for deserialization of Kernel AST into corresponding VM objects. pkg/vm/bin/kernel_service.dart implements the Kernel Service isolate, runtime/vm/ glues Dart implementation to the rest of the VM. package:vm hosts most of the Kernel based VM specific functionality, e.g various Kernel-to-Kernel transformations. However some VM specific transformations still live in package:kernel for historical reasons. A good example of a complicated transformation is package:kernel/transformations/continuation.dart, which desugars async,async* and sync* functions.

Trying it

If you are interested in Kernel format and its VM specific usage, then you can use pkg/vm/bin/gen_kernel.dart to produce a Kernel binary file from Dart source. Resulting binary can then be dumped using pkg/vm/bin/dump_kernel.dart.

# Take hello.dart and compile it to hello.dill Kernel binary using CFE.
$ dart pkg/vm/bin/gen_kernel.dart                        \
       --platform out/ReleaseX64/vm_platform_strong.dill \
       -o hello.dill                                     \

# Dump textual representation of Kernel AST.
$ dart pkg/vm/bin/dump_kernel.dart hello.dill hello.kernel.txt

When you try using gen_kernel.dart you will notice that it requires something called platform, a Kernel binary containing AST for all core libraries (dart:core, dart:async, etc). If you have Dart SDK build configured then you can just use platform file from the out directory, e.g. out/ReleaseX64/vm_platform_strong.dill. Alternatively you can use pkg/front_end/tool/_fasta/compile_platform.dart to generate the platform

# Produce outline and platform files using the given libraries list.
$ dart pkg/front_end/tool/_fasta/compile_platform.dart \
       dart:core                                       \
       sdk/lib/libraries.json                          \
       vm_outline.dill vm_platform.dill vm_outline.dill

Initially all functions have a placeholder instead of an actually executable code for their bodies: they point to LazyCompileStub, which simply asks runtime system to generate executable code for the current function and then tail-calls this newly generated code.

Lazy Compilation

When the function is compiled for the first time this is done by unoptimizing compiler.

Unoptimized Compilation

Unoptimizing compiler produces machine code in two passes:

  1. Serialized AST for the function's body is walked to generate an control flow graph (CFG) for the function body. CFG consists of basic blocks filled with intermediate language (IL) instructions. IL instructions used at this stage resemble instructions of a stack based virtual machine: they take operands from the stack, perform operations and then push results to the same stack. In reality not all functions have actual Dart / Kernel AST bodies, e.g. natives defined in C++ or artificial tear-off functions generated by Dart VM - in these cases IL is just created from thin air, instead of generating it from Kernel AST.

  2. resulting CFG is directly compiled to machine code using one-to-many lowering of IL instructions: each IL instruction expands to multiple machine language instructions.

There are no optimizations performed at this stage. The main goal of unoptimizing compiler is to produce executable code quickly.

This also means that unoptimizing compiler does not attempt to statically resolve any calls that were not resolved in Kernel binary, so calls (MethodInvocation or PropertyGet AST nodes) are compiled as if they were completely dynamic. VM currently does not use any form of virtual table or interface table based dispatch and instead implements dynamic calls using inline caching.

The core idea behind inline caching is to cache results of method resolution in a call site specific cache. Inline caching mechanism used by the VM consists of Original implementations of inline caching were actually patching the native code of the function - hence the name inline caching. The idea of inline caching dates far back to Smalltalk-80, see Efficient implementation of the Smalltalk-80 system :

The picture below illustrates the structure and the state of an inline cache associated with animal.toFace() call site, which was executed twice with an instance of Dog and once with an instance of a Cat.

Inline Caching

Unoptimizing compiler by itself is enough to execute any possible Dart code. However the code it produces is rather slow, that is why VM also implements adaptive optimizing compilation pipeline. The idea behind adaptive optimization is to use execution profile of a running program to drive optimization decisions.

As unoptimized code is running it collects the following information:

When an execution counter associated with a function reaches certain threshold, this function is submitted to a background optimizing compiler for optimization.

Optimizing compilations starts in the same way as unoptimizing compilation does: by walking serialized Kernel AST to build unoptimized IL for the function that is being optimized. However instead of directly lowering that IL into machine code, optimizing compiler proceeds to translate unoptimized IL into static single assignment (SSA) form based optimized IL. SSA based IL is then subjected to speculative specialization based on the collected type feedback and passed through a sequence of classical and Dart specific optimizations: e.g. inlining, range analysis, type propagation, representation selection, store-to-load and load-to-load forwarding, global value numbering, allocation sinking, etc. At the end optimized IL is lowered into machine code using linear scan register allocator and a simple one-to-many lowering of IL instructions.

Once compilation is complete background compiler requests mutator thread to enter a safepoint and attaches optimized code to the function. Next time the function is called - it will use optimized code. Some functions contain very long running loops, so it makes sense to switch execution from unoptimized to optimized code while the function is still running. This process is called on stack replacement (OSR) owing its name to the fact that a stack frame for one version of the function is transparently replaced with a stack frame for another version of the same function.

Optimizing Compilation

Source to read

Compiler sources are in the runtime/vm/compiler directory. Compilation pipeline entry point is CompileParsedFunctionHelper::Compile. IL is defined in runtime/vm/compiler/backend/il.h. Kernel-to-IL translation starts in kernel::StreamingFlowGraphBuilder::BuildGraph, and this function also handles construction of IL for various artificial functions. StubCode::GenerateNArgsCheckInlineCacheStub generates machine code for inline-cache stub, while InlineCacheMissHandler handles IC misses. runtime/vm/compiler/ defines optimizing compiler passes and their order. JitCallSpecializer does most of the type-feedback based specializations.

Trying it

VM also has flags which can be used to control JIT and to make it dump IL and generated machine code for the functions that are being compiled by the JIT.

Flag Description
--print-flow-graph[-optimized] Print IL for all (or only optimized) compilations
--disassemble[-optimized] Disassemble all (or only optimized) compiled functions
--print-flow-graph-filter=xyz,abc,... Restrict output triggered by previous flags only to the functions which contain one of the comma separated substrings in their names
--compiler-passes=... Fine control over compiler passes: force IL to be printed before/after a certain pass. Disable passes by name. Pass help for more information
--no-background-compilation Disable background compilation, and compile all hot functions on the main thread. Useful for experimentation, otherwise short running programs might finish before background compiler compiles hot function

For example

# Run test.dart and dump optimized IL and machine code for
# function(s) that contain(s) "myFunction" in its name.
# Disable background compilation for determinism.
$ dart --print-flow-graph-optimized         \
       --disassemble-optimized              \
       --print-flow-graph-filter=myFunction \
       --no-background-compilation          \

It is important to highlight that the code generated by optimizing compiler is specialized under speculative assumptions based on the execution profile of the application. For example, a dynamic call site that only observed instances of a single class C as a receiver will be converted into a direct call preceeded by a check verifying that receiver has an expected class C. However these assumptions might be violated later during execution of the program:

void printAnimal(obj) {
  print('Animal {');
  print('  ${obj.toString()}');

// Call printAnimal(...) a lot of times with an intance of Cat.
// As a result printAnimal(...) will be optimized under the
// assumption that obj is always a Cat.
for (var i = 0; i < 50000; i++)

// Now call printAnimal(...) with a Dog - optimized version
// can not handle such an object, because it was
// compiled under assumption that obj is always a Cat.
// This leads to deoptimization.

Whenever optimized code is making some assumptions which can't be derived from statically immutable information it needs to guard against violation of those assumptions and be able to recover if such violation occurs.

This process of recovery is known as deoptimization: whenever optimized version hits a case which it can't handle it simply transfers execution into the matching point of unoptimized function and continues execution there. Unoptimized version of a function does not make any assumptions and can handle all possible inputs. Entering unoptimized function at the right spot is absolutely crucial because code has side-effects (e.g. in the function above deoptimization happens after we already executed the first print). Matching instructions that deoptimize to positions in the unoptimized code in VM is done using deopt ids

VM usually discards optimized version of the function after deoptimization and then reoptimizes it again later - using updated type feedback.

There are two ways VM guards speculative assumptions made by the compiler:

Source to read

Deoptimizer machinery resides in runtime/vm/ It is essentially a mini-interpreter for deoptimization instructions which describe how to reconstruct needed state of the unoptimized code from the state of optimized code. Deoptimization instructions are generated by CompilerDeoptInfo::CreateDeoptInfo for every potential deoptimization location in optimized code during compilation.

Trying it

Flag --trace-deoptimization makes VM print information about the cause and location of every deoptimization that occurs. --trace-deoptimization-verbose makes VM print a line for every deoptimization instruction it executes during deoptimization.

1.2 Running from Snapshots

VM has ability to serialize isolate's heap or more precisely object graph residing in the heap into a binary snapshot. Snapshot then can be used to recreate the same state when starting VM isolates.


Snapshot's format is low level and optimized for fast startup - it is essentially a list of objects to create and instructions on how to connect them together. That was the original idea behind snapshots: instead of parsing Dart source and gradually creating internal VM data structures, VM can just spin an isolate up with all necessary data structures quickly unpacked from the snapshot.

Initially snapshots did not include machine code, however this capability was later added when AOT compiler was developed. Motivation for developing AOT compiler and snapshots-with-code was to allow VM to be used on the platforms where JITing is impossible due to platform level restrictions.

Snapshots-with-code work almost in the same way as normal snapshots with a minor difference: they include a code section which unlike the rest of the snapshot does not require deserialization. This code section laid in way that allows it to directly become part of the heap after it was mapped into memory.


Source to read

runtime/vm/ handles serialization and deserialization of snapshots. A family of API functions Dart_CreateXyzSnapshot[AsAssembly] are responsible for writing out snapshots of the heap (e.g. Dart_CreateAppJITSnapshotAsBlobs and Dart_CreateAppAOTSnapshotAsAssembly). On the other hand Dart_CreateIsolate optionally takes snapshot data to start an isolate from.

1.3 Running from AppJIT snapshots

AppJIT snapshots were introduced to reduce JIT warm up time for large Dart applications like dartanalyzer or dart2js. When these tools are used on small projects they spent as much time doing actual work as VM spends JIT compiling these apps.

AppJIT snapshots allow to address this problem: an application can be run on the VM using some mock training data and then all generated code and VM internal data structures are serialized into an AppJIT snapshot. This snapshot can then be distributed instead of distributing application in the source (or Kernel binary) form. VM starting from this snapshot can still JIT - if it turns out that execution profile on the real data does not match execution profile observed during training.


Trying it

dart binary will generate AppJIT snapshot after running the application if you pass --snapshot-kind=app-jit --snapshot=path-to-snapshot to it. Here is an example of generating and using an AppJIT snapshot for dart2js.

# Run from source in JIT mode.
$ dart pkg/compiler/lib/src/dart2js.dart -o hello.js hello.dart
Compiled 7,359,592 characters Dart to 10,620 characters JavaScript in 2.07 seconds
Dart file (hello.dart) compiled to JavaScript: hello.js

# Training run to generate app-jit snapshot
$ dart --snapshot-kind=app-jit --snapshot=dart2js.snapshot \
       pkg/compiler/lib/src/dart2js.dart -o hello.js hello.dart
Compiled 7,359,592 characters Dart to 10,620 characters JavaScript in 2.05 seconds
Dart file (hello.dart) compiled to JavaScript: hello.js

# Run from app-jit snapshot.
$ dart dart2js.snapshot -o hello.js hello.dart
Compiled 7,359,592 characters Dart to 10,620 characters JavaScript in 0.73 seconds
Dart file (hello.dart) compiled to JavaScript: hello.js

1.4 Running from AppAOT snapshots

AOT snapshots were originally introduced for platforms which make JIT compilation impossible, but they can also be used in situations where fast startup and consistent performance is worth potential performance penalty. There is usually a lot of confusion around how performance characteristics of JIT and AOT compare. JIT has access to precise local type information and execution profile of the running application, however it has to pay for it with warmup. AOT can infer and prove various properties globally (for which it has to pay with compile time), but has no information of how the program will actually be executing - on the other hand AOT compiled code reaches its peak performance almost immediately with virtual no warmup. Currently Dart VM JIT has best peak performance, while Dart VM AOT has best startup time.

Inability to JIT implies that:

  1. AOT snapshot must contain executable code for each and every function that could be invoked during application execution;
  2. the executable code must not rely on any speculative assumptions that could be violated during execution;

To satisfy these requirements the process of AOT compilation does global static analysis (type flow analysis or TFA) to determine which parts of the application are reachable from known set of entry points, instances of which classes are allocated and how types flow through the program. All of these analyses are conservative: meaning that they err on the side of correctness - which is in stark contrast with JIT which can err on the side of performance, because it can always deoptimize into unoptimized code to implement correct behavior.

All potentially reachable functions are then compiled to native code without any speculative optimizations. However type flow information is still used to specialize the code (e.g. devirtualize calls).

Once all functions are compiled a snapshot of the heap can be taken.

Resulting snapshot can then be run using precompiled runtime, a special variant of the Dart VM which excludes components like JIT and dynamic code loading facilities.

AOT pipeline

Source to read

package:vm/transformations/type_flow/transformer.dart is an entry point to the type flow analysis and transformation based on TFA results. Precompiler::DoCompileAll is an entry point to the AOT compilation loop in the VM.

Trying it

AOT compilation pipeline is currently not packaged into Dart SDK and projects that rely on it (like Flutter) must build it by hand out of pieces provided by the SDK. pkg/vm/tool/precompiler2 script is a good reference for how the pipeline is structured and which binary artifacts must be build to use it.

# Need to build normal dart executable and runtime for running AOT code.
$ tool/ -m release -a x64 runtime dart_precompiled_runtime

# Now compile an application using AOT compiler
$ pkg/vm/tool/precompiler2 hello.dart hello.aot

# Execute AOT snapshot using runtime for AOT code
$ out/ReleaseX64/dart_precompiled_runtime hello.aot
Hello, World!

Note that it is possible to pass options like --print-flow-graph-optimized and --disassemble-optimized to the precompiler2 script if you would like to inspect generated AOT code.

1.4.1 Switchable Calls

Even with global and local analyses AOT compiled code might still contain call sites which could not be devirtualized statically. To compensate for this AOT compiled code and runtime use an extension of inline caching technique utilized in JIT. This extended version is called switchable calls.

JIT section already described that each inline cache associated with a call site consists of two pieces: a cache object (represented by an instance of RawICData) and a chunk of native code to invoke (e.g. a InlineCacheStub). In JIT mode runtime would only update the cache itself. However in AOT runtime can choose to replace both the cache and the native code to invoke depending on the state of the inline cache.

AOT IC. Unlinked

Initially all dynamic calls start in the unlinked state. When such call-site is reached for the first time UnlinkedCallStub is invoked, which simply calls into runtime helper DRT_UnlinkedCall to link this call site.

If possible DRT_UnlinkedCall tries to transition the call site into a monomorphic state. In this state call site turns into a direct call, which enters method through a special entry point which verifies that receiver has expected class.

AOT IC: Monomorphic

In the example above we assume that when obj.method() was executed for the first time obj was an instance of C and obj.method resolved to C.method.

Next time we execute the same call-site it would invoke C.method directly, bypassing any sort of method lookup process. However it would enter C.method through a special entry point, which would verify that obj is still an instance of C. If that is not the case DRT_MonomorphicMiss would be invoked and will try to select the next call site state.

C.method might still be a valid target for an invocation, e.g obj is an instance of the class D which extends C but does not override C.method. In this case we check if call site could transition into a single target state, implemented by SingleTargetCallStub (see also RawSingleTargetCache).

AOT IC: Single Target

This stub is based on the fact that for AOT compilation most classes are assigned integer ids using depth-first traversal of the inheritance hierarchy. If C is a base class with subclasses D0, ..., Dn and none of those override C.method then C.:cid <= classId(obj) <= max(D0.:cid, ..., Dn.:cid) implies that obj.method resolves to C.method. In this circumstances instead of comparing to a single class (monomorphic state), we can use class id range check (single target state) which would work for all subclasses of C.

Otherwise call site would be switched to use linear search inline cache, similar to the one used in JIT mode (see ICCallThroughCodeStub, RawICData and DRT_MegamorphicCacheMissHandler).

AOT IC: linear IC call

Finally if the number of checks in the linear array grows past threshold the call site is switched to use a dictionary like structure (see MegamorphicCallStub, RawMegamorphicCache and DRT_MegamorphicCacheMissHandler).

AOT IC: dictionary

2 Runtime System


This section will be written next.

2.1 Object Model