Last modified 7 months ago Last modified on 09/12/13 11:33:05


JavaScriptCore is the built-in JavaScript engine for WebKit. It currently implements ECMAScript as in ECMA-262 specification.

JavaScriptCore is often referred with different names, such as SquirrelFish and SquirrelFish Extreme. Within the context of Safari, Nitro and Nitro Extreme (the marketing terms from Apple) are also commonly used. However, the name of the project and the library is always JavaScriptCore.

JavaScriptCore source code resides in WebKit source tree, it's under Source/JavaScriptCore directory.

Core Engine

JavaScriptCore is an optimizing virtual machine. JavaScriptCore consists of the following building blocks: lexer, parser, start-up interpreter (LLInt), baseline JIT, a low-latency optimizing JIT (DFG), and a high-throughput optimizing JIT (FTL).

Lexer is responsible for the lexical analysis, i.e. breaking down the script source into a series of tokens. JavaScriptCore lexer is hand-written, it is mostly in parser/Lexer.h and parser/Lexer.cpp.

Parser carries out the syntactic analysis, i.e. consuming the tokens from the lexer and building the corresponding syntax tree. JavaScriptCore uses a hand-written recursive descent parser, the code is in parser/JSParser.h and parser/JSParser.cpp.

LLInt, short for Low Level Interpreter, executes the bytecodes produced by the parser. The bulk of the Low Level Interpreter is in llint/. The LLInt is written in a portable assembly called offlineasm (see offlineasm/, which can compile to x86, ARMv7, and C. The LLInt is intended to have zero start-up cost besides lexing and parsing, while obeying the calling, stack, and register conventions used by the just-in-time compilers. For example, calling a LLInt function works "as if" the function was compiled to native code, except that the machine code entrypoint is actually a shared LLInt prologue. The LLInt includes optimizations such as inline caching to ensure fast property access.

Baseline JIT kicks in for functions that are invoked at least 6 times, or take a loop at least 100 times (or some combination - like 3 invocations with 50 loop iterations total). Note, these numbers are approximate; the actual heuristics depend on function size and current memory pressure. The LLInt will on-stack-replace (OSR) to the JIT if it is stuck in a loop; as well all callers of the function are relinked to point to the compiled code as opposed to the LLInt prologue. The Baseline JIT also acts as a fall-back for functions that are compiled by the optimizing JIT: if the optimized code encounters a case it cannot handle, it bails (via an OSR exit) to the Baseline JIT. The Baseline JIT is in jit/. The Baseline JIT also performs sophisticated polymorphic inline caching for almost all heap accesses.

Both the LLInt and Baseline JIT collect light-weight profiling information to enable speculative execution by the next tier of execution (the DFG). Information collected includes recent values loaded into arguments, loaded from the heap, or loaded from a call return. Additionally, all inline caching in the LLInt and Baseline JIT is engineered to enable the DFG to scrape type information easily: for example the DFG can detect that a heap access sometimes, often, or always sees a particular type just by looking at the current state of an inline cache; this can be used to determine the most profitable level of speculation. A more thorough overview of type inference in JavaScriptCore is provided in the next section.

DFG JIT kicks in for functions that are invoked at least 60 times, or that took a loop at least 1000 times. Again, these numbers are approximate and are subject to additional heuristics. The DFG performs aggressive type speculation based on profiling information collected by the lower tiers. This allows it to forward-propagate type information, eliding many type checks. Sometimes the DFG goes further and speculates on values themselves - for example it may speculate that a value loaded from the heap is always some known function in order to enable inlining. The DFG uses deoptimization (we call it "OSR exit") to handle cases where speculation fails. Deoptimization may be synchronous (for example, a branch that checks that the type of a value is that which was expected) or asynchronous (for example, the runtime may observe that the shape or value of some object or variable has changed in a way that contravenes assumptions made by the DFG). The latter is referred to as "watchpointing" in the DFG codebase. Altogether, the Baseline JIT and the DFG JIT share a two-way OSR relationship: the Baseline JIT may OSR into the DFG when a function gets hot, and the DFG may OSR to the Baseline JIT in the case of deoptimization. Repeated OSR exits from the DFG serve as an additional profiling hint: the DFG OSR exit machinery records the reason of the exit (including potentially the values that failed speculation) as well as the frequency with which it occurred; if an exit is taken often enough then reoptimization kicks in: callers are relinked to the Baseline JIT for the affected function, more profiling is gathered, and then the DFG may be later reinvoked. Reoptimization uses exponential back-off to defend against pathological code. The DFG is in dfg/.

FTL JIT kicks in for functions that are invoked thousands of times, or loop tens of thousands of times. It is still a work in progress and is disabled by default. See FTLJIT for more information.

At any time, functions, eval blocks, and global code in JavaScriptCore may be executing in a mix of the LLInt, Baseline JIT, and DFG. In the extreme case of a recursive function, there may even be multiple stack frames where one frame is in the LLInt, another is in the Baseline JIT, while another still is in the DFG; even more extreme are cases where one stack frame is executing an old DFG compilation and another is executing a new DFG compilation because recompilation kicked in but execution did not yet return to the old DFG code. These three engines are designed to maintain identical execution semantics, and so even if multiple functions in a JavaScript program are executing in a mix of these engines the only perceptible effect ought to be execution performance.

Type Inference

Type inference is achieved by profiling values, predicting what types operations will produce based on those profiles, inserting type checks based on the type predictions, and then attempting to construct type proofs about the types of values based on the type checks.

Consider this example to motivate type inference, and optimization, of JavaScript:

o.x * o.x + o.y * o.y

Say that in the context where this code is used, 'o' is an object, and it indeed has properties 'x' and 'y', and those properties are nothing special - in particular, they are not accessors. Let's also say that 'o.x' and 'o.y' usually return doubles, but may sometimes return integers - JavaScript does not have a built-in notion of an integer value, but for efficiency JavaScriptCore will represent most integers as int32 rather than as double. To understand both the problem of type inference, and its solution in JavaScriptCore, it's useful to first consider what a JavaScript engine would have to do if it had no information about 'o', 'o.x', or 'o.y':

  • The expression 'o.x' has to first check if 'o' has any special handling of property access. It may be a DOM object, and DOM objects may intercept accesses to their properties in non-obvious ways. If it doesn't have special handling, the engine must look up the property named "x" (where "x" is literally a string) in the object. Objects are just tables mapping strings to either values or accessors. If it maps to an accessor, the accessor must be called. If it isn't an accessor, then the value is returned. If "x" is not found in 'o', then the process repeats for o's prototype. The inference required for optimizing the object access is not covered in this section.
  • The binary multiply operation in the expression 'o.x * o.x' has to first check the types of its operands. Either operand may be an object, in which case its 'valueOf' method must be called. Either operand may be a string, in which case it must be converted to a number. Once both operands are appropriately converted to numbers (or if they were numbers already), the engine must check if they are both integers; if so, then an integer multiply is performed. This may overflow, in which case a double multiply is performed instead. If either operand is a double, then both are converted to doubles, and a double multiply is performed. Thus 'o.x * o.x' may return either an integer or a double. There is no way of proving, for the generic JavaScript multiply, what kind of number it will return and how that number will be represented.
  • The binary addition operation in the expression 'o.x * o.x + o.y * o.y' has to proceed roughly as the multiply did, except it has to consider the possibility that its operands are strings, in which case a string concatenation is performed. In this case, we could statically prove that this isn't the case - multiply must have returned a number. But still, addition must perform checks for integers versus doubles on both operands, since we do not know which of those types was returned by the multiplication expressions. As a result, the addition may also return either an integer, or a double.

The intuition behind JavaScriptCore's type inference is that we can say with good probability what type a numerical operation (such as addition or multiplication) will return, and which path it will take, if we could guess what types flowed into that operation. This forms a kind of induction step that applies to operations that don't generally operate over the heap: if we can predict their inputs, then we can predict their outputs. But induction requires a base case. In the case of JavaScript operations, the base cases are operations that get non-local values: for example, loading a value from the heap (as in 'o.x'), accessing an argument to a function, or using a value returned from a function call. For simplicity, we refer to all such non-local values as heap values, and all operations that place heap values into local variables as heap operations. For arguments, we treat the function prologue as a "heap operation" of sorts, which "loads" the arguments into the argument variables. We bootstrap our inductive reasoning about type predictions by using value profiling: both the LLInt and Baseline JIT will record the most recent value seen at any heap operation. Each heap operation has exactly one value profile bucket associated with it, and each value profile bucket will store exactly one recent value.

A straw-man view of JavaScriptCore's type inference is that we convert each value profile's most recent value into a type, and then apply the induction step to propagate this type through all operations in the function. This gives us type predictions at each value-producing operation in the function. All variables become predictively typed as well.

In reality, JavaScriptCore includes in each value profile a second field, which is a type that bounds a random subset of the values seen. This type uses the SpeculatedType (or SpecType for short) type system, which is implemented in SpeculatedType.h. The type of each value profile starts out as SpecNone - i.e. the type corresponding to no values. You can also think of this as bottom (see Lattice), or the "contradiction". When the Baseline JIT's execution counter exceeds a threshold (see JIT::emitOptimizationCheck in JIT.cpp), it will produce a new type to bound both the last type, and the most recent value. It may also then choose to invoke the DFG, or it may choose to let the baseline code run more. Our heuristics favor the latter, meaning that when DFG compilation kicks in, each value profile will typically have a type that bounds multiple different values.

The propagation of SpecTypes derived at value profiles to all operations and variables in a function is performed using a standard forward data flow formulation, implemented as a flow-insensitive fixpoint. This is one of the first phases of DFG compilation, and is only activated once the Baseline JIT decides, based on its execution count, that a function would be better off executing optimized code. See DFGPredictionPropagationPhase.cpp.

After each value in a function is labeled with a predicted type, we insert speculative type checks based on those predictions. For example, in a numeric operation (like 'o.x * o.y'), we insert speculate-double checks on the operands of the multiplication. If a speculation check fails, execution is diverted from optimized DFG code back to the Baseline JIT. This has the effect of proving the type for subsequent code within the DFG. Consider how the simple addition operation 'a + b' will be decomposed, if 'a' and 'b' both had SpecInt32 as their predicted type:

check if a is Int32 -> else OSR exit to Baseline JIT
check if b is Int32 -> else OSR exit to Baseline JIT
result = a + b // integer addition
check if overflow -> else OSR exit to Baseline JIT

After this operation completes, we know that:

  • 'a' is an integer.
  • 'b' is an integer.
  • the result is an integer.

Any subsequent operations on either 'a' or 'b' do not need to check their types. Likewise for operations on the result. The elimination of subsequent checks is achieved by a second data flow analysis, called simply the DFG CFA. Unlike the prediction propagation phase, which is concerned with constructing type predictions, the CFA is concerned with constructing type proofs. The CFA, found in DFGCFAPhase.cpp and DFGAbstractState.cpp, follows a flow-sensitive forward data flow formulation. It also implements sparse conditional constant propagation, which gives it the ability to sometimes prove that values are constants, as well as proving their types.

Putting this together, the expression 'o.x * o.x + o.y * o.y' will only require type checks on the value loaded from 'o.x' and the value loaded from 'o.y'. After that, we know that the values are doubles, and we know that we only have to emit a double multiply path followed by a double addition. When combined with type check hoisting, DFG code will usually execute a type check at most once per heap load.

Object Access Optimization

(to be written)

Attached are some slides from a recent talk about JavaScriptCore.