diff options
Diffstat (limited to 'js/src/doc/JITOptimizations')
| -rw-r--r-- | js/src/doc/JITOptimizations/Outcomes.md | 581 | ||||
| -rw-r--r-- | js/src/doc/JITOptimizations/Strategies.md | 582 | ||||
| -rw-r--r-- | js/src/doc/JITOptimizations/config.sh | 15 |
3 files changed, 1178 insertions, 0 deletions
diff --git a/js/src/doc/JITOptimizations/Outcomes.md b/js/src/doc/JITOptimizations/Outcomes.md new file mode 100644 index 000000000..b0eb9c43a --- /dev/null +++ b/js/src/doc/JITOptimizations/Outcomes.md @@ -0,0 +1,581 @@ +# JIT Optimization Outcomes + +SpiderMonkey's optimizing JIT, IonMonkey, uses a number of different +optimization strategies to speed up various operations. The most commonplace +operations that are relevant for fast program execution are property accesses +and function calls. + +This page documents the meaning of different optimization outcomes. + +## General Outcomes + +General outcomes shared between various optimization strategies. + +### GenericFailure + +The optimization attempt failed, and the reason was not recorded. + +### GenericSuccess + +Optimization succeeded. + +### Disabled + +The optimization has been explicitly disallowed. + +### NoTypeInfo + +Optimization failed because there was no type information associated with +object containing the property. This failure mode is unlikely, and occurs +if the target object is obtained in some roundabout way. + +### NoAnalysisInfo + +TODO + +### NoShapeInfo + +The baseline compiler recorded no usable shape information for this operation. + +### UnknownObject + +The type of the object is not known. This can happen if the operation sees many different types of objects, and so the type of the input to the operation cannot be resolved to a single type. + +### UnknownProperties + +Optimization failed because the object containing the property was marked +as having unknown properties. This can happen if too many properties are +defined on the object, or if `delete` is used to remove one of the object's +properties. + +### Singleton + +One of the types present in the typeset was a singleton type, preventing the optimization from being enabled. + +### NotSingleton + +Optimization failed because the object containing the property did not +have a 'singleton' type. Singleton types are assigned to objects that are +"one of a kind", such as global objects, literal objects declared in the +global scope, and top-level function objects. + +### NotFixedSlot + +The property being accessed is not stored at a known location in the object. This can occur if one of the expected types of objects to be used in this operation has unknown properties, or if different instances of the object store the property at different locations (for example, some instances have the property assigned in a different order than others). + +### InconsistentFixedSlot + +The property being accessed is not stored at a known location in the object. This can occur if the operation is polymorphic on different object types and one or more of the object types contain the property at a different slot than the others. + +### NotObject + +Optimization failed because the stored in the property could potentially +be a non-object value. Since only objects can be uniquely typed, the +optimization strategy fails in this case. + +### NotStruct + +The object holding the property is not a typed struct object. + +### NotUnboxed + +The object whose property is being accessed is not formatted as an +unboxed object. + +### UnboxedConvertedToNative + +The object whose property is being accessed was previously unboxed, +but was deoptimized and converted to a native object. + +### StructNoField + +The unboxed property being accessed does not correspond to a field on typed +object. + +### InconsistentFieldType + +The type of an unboxed field is not consistent across all the different types of objects it could be accessed from. + +### InconsistentFieldOffset + +The offset of an unboxed field is not consistent across all the different types of objects it could be accessed from. + +### NeedsTypeBarrier + +Optimization failed because somehow the property was accessed in a way +that returned a different type than the expected constant. This is an +unlikely failure mode, and should not occur. + +### InDictionaryMode + +The object whose property is being accessed is in dictionary mode. Objects which are used in ways that suggest they are hashtables, are turned into dictionary objects and their types marked as such. + +### NoProtoFound + +A prototype object was not found for all the object used by this operation. + +### MultiProtoPaths + +Objects used in this operation had differing prototypes. + +### NonWritableProperty + +The property being assigned to is not writable for some types of objects which are used in this operation. + +### ProtoIndexedProps + +The object being accessed has indexed properties that are exotic (for example, defined as a property on a prototype object and left as a hole in the underlying object). + +### ArrayBadFlags + +The array being accessed may have flags that the optimization strategy cannot handle. For example, if the array has sparse indexes, or has indexes that overflow the array's length, the optimization strategy may fail. + +### ArrayDoubleConversion + +The type-system indicates that some arrays at this site should be converted to packed arrays of doubles, while others should not. The optimization strategy fails for this condition. + +### ArrayRange + +Could not accurately calculate the range attributes of an inline array creation. + +### ArraySeenNegativeIndex + +Arrays at this element access location have seen negative indexes. + +### TypedObjectHasDetachedBuffer + +The storage for the typed object being accessed at this location might be a detached ArrayBuffer. (This can happen if the typed object, or its underlying buffer as accessed using `TypedObject.storage(typedObject).buffer`, is transferred using the structured clone algorithm.) + +### TypedObjectArrayRange + +Failed to do range check of element access on a typed object. + +### AccessNotDense + +### AccessNotSimdObject + +The observed type of the target of the property access doesn't guarantee +that it is a SIMD object. + +### AccessNotTypedObject + +The observed type of the target of the property access doesn't guarantee +that it is a TypedObject. + +### AccessNotTypedArray + +The observed type of the target of the property access doesn't guarantee +that it is a TypedArray. + +### AccessNotString + +The observed type of the target of the property access doesn't guarantee +that it is a String. + +### OperandNotString + +Optimization failed because of failing to speculate the operand is a string. + +### OperandNotNumber + +Optimization failed because of failing to speculate the operand is a number. + +### OperandNotStringOrNumber + +Optimization failed because of failing to speculate the operand is a string or a number. + +### OperandNotSimpleArith + +Optimization failed because of failing to speculate the operand is a simple arithmetic type. I.e. definitely not an object, string, symbol or internal magic type. + +### StaticTypedArrayUint32 + +Typed Arrays of uint32 values are not yet fully optimized. + +### StaticTypedArrayCantComputeMask +### OutOfBounds + +The element access has been observed to be out of the length bounds of +the string being accessed. + +### GetElemStringNotCached + +IonMonkey does not generate inline caches for element accesses on +target values which may be strings. + +### NonNativeReceiver + +IonMonkey does not generate inline caches for indexed element accesses on +target values which may be non-native objects (e.g. DOM elements). + +### IndexType + +IonMonkey does not generate inline caches for element reads in which +the keys have never been observed to be a String, Symbol, or Int32. + +### SetElemNonDenseNonTANotCached + +IonMonkey only generates inline caches for element accesses which are +either on dense objects (e.g. dense Arrays), or Typed Arrays. + +### NoSimdJitSupport + +Optimization failed because SIMD JIT support was not enabled. + +### SimdTypeNotOptimized + +The type observed as being retrieved from this property access did not +match an optimizable type. + +### HasCommonInliningPath + +Inlining was abandoned because the inlining call path was repeated. A +repeated call path is indicative of a potentially mutually recursive +function call chain. + +### Inlined + +Method has been successfully inlined. + +### DOM + +Successfully optimized a call to a DOM getter or setter function. + +### Monomorphic + +Successfully optimized a guarded monomorphic property access. This +property access is about twice as expensive as a definite property access. +A guarded monomorphic property access is basically a direct memory load +from an object, guarded by a type or shape check. In pseucodcode: + + if HiddenType(obj) === CONSTANT_TYPE: + return obj[CONSTANT_SLOT_OFFSET] + else + BAILOUT() + +### Polymorphic + +Successfully optimized a guarded polymorphic property access. This +property access is similar to a monomorphic property access, except +that multiple different hidden types have been observed, and each of +them are being checked for. In pseudocode: + + if HiddenType(obj) === CONSTANT_TYPE_1: + return obj[CONSTANT_SLOT_OFFSET_1] + elif HiddenType(obj) === CONSTANT_TYPE_2: + return obj[CONSTANT_SLOT_OFFSET_2] + ... + elif HiddenType(obj) === CONSTANT_TYPE_K: + return obj[CONSTANT_SLOT_OFFSET_K] + else + BAILOUT() + +## Inline Cache Outcomes + +Outcomes describing inline cache stubs that were generated. + +### ICOptStub_GenericSuccess + +Generic success condition for generating an optimized inline cache stub. + +### ICGetPropStub_ReadSlot + +An inline cache property read which loads a value from a known slot +location within an object. + +### ICGetPropStub_CallGetter + +An inline cache property read which calls a getter function. + +### ICGetPropStub_ArrayLength + +An inline cache property read which retrieves the length of an array +object. + +### ICGetPropStub_UnboxedRead + +An inline cache property read which retrieves an value from an unboxed +object. + +### ICGetPropStub_UnboxedReadExpando + +An inline cache property read which retrieves an value from the expando +component of an unboxed object. + +### ICGetPropStub_UnboxedArrayLength + +An inline cache property read which retrieves the length of an unboxed +array object. + +### ICGetPropStub_TypedArrayLength + +An inline cache property read which retrieves the length of an typed array +object. + +### ICGetPropStub_DOMProxyShadowed + +An inline cache property read which retrieves a shadowed property from a +DOM object. A shadowed property is an inbuilt DOM property that has been +overwritten by JS code. + +### ICGetPropStub_DOMProxyUnshadowed + +An inline cache property read which retrieves an unshadowed property from +a DOM object. + +### ICGetPropStub_GenericProxy + +An inline cache property read which retrieves the property by calling into +a generic proxy trap. + +### ICGetPropStub_ArgumentsLength + +An inline cache property read which reads the length value from an +arguments object. + +### ICSetPropStub_Slot + +An inline cache property write which sets a value to a known slot location +in an object. + +### ICSetPropStub_GenericProxy + +An inline cache property write which sets the value by calling into a +generic proxy trap. + +### ICSetPropStub_DOMProxyShadowed + +An inline cache property write which sets a shadowed property from a DOM +object. A shadowed property is an inbuilt DOM property that has been +overwritten by JS code. + +### ICSetPropStub_DOMProxyUnshadowed + +An inline cache property write which sets an unshadowed property on an +DOM object. + +### ICSetPropStub_CallSetter + +An inline cache property write which calls a setter function on an object. + +### ICSetPropStub_AddSlot + +An inline cache property write which adds a new slot to an object, +because the object is known to not have such a property already. + +### ICSetPropStub_SetUnboxed + +An inline cache property write which sets an unboxed value on an unboxed +object. + +### ICGetElemStub_ReadSlot + +An inline cache element read which loads a value from a known property +slot in an object. + +### ICGetElemStub_CallGetter + +An inline cache element read which calls a getter function on the object. + +### ICGetElemStub_ReadUnboxed + +An inline cache element read which loads an unboxed value from a known +property slot on an unboxed object. + +### ICGetElemStub_Dense + +An inline cache element read which loads an element from a dense array. + +### ICGetElemStub_DenseHole + +An inline cache element read which loads an element from a dense array +which might have a hole. + +### ICGetElemStub_TypedArray + +An inline cache element read which loads an element from a typed array. + +### ICGetElemStub_ArgsElement + +An inline cache element read which loads an element from an `arguments` +object. + +### ICGetElemStub_ArgsElementStrict + +An inline cache element read which loads an element from an `arguments` +object in strict mode. + +### ICSetElemStub_Dense + +An inline cache element write which sets an element on a dense array. + +### ICSetElemStub_TypedArray + +An inline cache element write which sets an element on a typed array. + +### ICNameStub_ReadSlot + +An inline cache element which loads a bare variable name from a slot on +a scope chain object. + +### ICNameStub_CallGetter + +An inline cache element which loads a bare variable name by calling a +getter function on the global object. + +### ICNameStub_TypeOfNoProperty + +An inline cache element which loads undefined for the type +of a missing property. + +## Call Inlining Outcomes + +Optimization outcomes of attempts to inline function calls. + +### CantInlineGeneric + +Generic failure-to-inline outcome. + +### CantInlineClassConstructor + +Cannot inline a class constructor function. + +### CantInlineExceededDepth + +Inlining stopped because inlining depth was exceeded. + +### CantInlineExceededTotalBytecodeLength + +Inlining stopped because the total bytecode length of all inlined +functions exceeded a threshold. + +### CantInlineBigCaller + +Inlining was aborted because the caller function is very large. + +### CantInlineBigCallee + +Inlining was aborted because the callee function is very large. + +### CantInlineBigCalleeInlinedBytecodeLength + +Inlining was aborted because the callee is known to inline a lot of code. + +### CantInlineNotHot + +Inlining was aborted because the callee is not hot enough. + +### CantInlineNotInDispatch + +This failure mode relates to inlining a method call of the form: + + obj.method() + +where a hidden type check on the target object reveals the method to be +inlined. If the property access for `method` does not identify +the method to be inlined, the inlining fails with this code. + +### CantInlineNativeBadType + +Inlining attempt of native function was aborted because the hidden +type for the function, or the result of the function, was not +compatible with one of the known inline-able native functions. + +### CantInlineNoTarget + +Unable to inline function call. The callee (target) function, f in f(x), +is not known at JIT time. + +### CantInlineNotInterpreted + +Unable to inline function call. The callee function is not an interpreted +function. For example, it could be a native function for which Ion has no +built-in specialization. + +### CantInlineNoBaseline + +Unable to inline function call. The interpreted callee function could not +be compiled by the Baseline compiler. + +### CantInlineLazy + +Unable to inline function call. The interpreted callee function has a +lazy, compiled on-demand script instead of an already compiled script. + +### CantInlineNotConstructor + +Unable to inline function call. Rare. The interpreted callee function is +invoked with new but cannot be called as a constructor. Property +accessors, Function.prototype, and arrow (=>) functions cannot be called +as constructors. + +### CantInlineDisableIon + +Unable to inline function call. The interpreted callee function has been +explicitly blacklisted against Ion compilation. + +### CantInlineTooManyArgs + +Unable to inline function call. The interpreted callee function either has +too many parameters or is called with too many arguments. These thresholds +are subject to change. + +### CantInlineRecursive + +Unable to inline function call. The interpreted callee function recurs for +more than one level. The first level of recursion is inlineable. + +### CantInlineHeavyweight + +Unable to inline function call. The interpreted callee function contains +variables bindings that are closed over. For example, +`function f() { var x; return function () { x++; } }` +closes over x in f. + +### CantInlineNeedsArgsObj + +Unable to inline function call. The interpreted callee function requires +an arguments object to be created. + +### CantInlineDebuggee + +Unable to inline function call. The interpreted callee function is being +debugged by the Debugger API. + +### CantInlineUnknownProps + +Unable to inline function call. The engine knows nothing definite about +the type of the callee function object. + +### CantInlineUnreachable + +Unable to inline function call. The call site has not been observed to +have ever been executed. It lacks observed type information for its arguments, +its return value, or both. + +### CantInlineBound + +Unable to inline function call. The interpreted callee is a bound function +generated from `Function.prototype.bind` that failed some sub-checks. +(expand) + +### CantInlineNativeNoSpecialization + +Unable to inline function call. Ion does not have built-in specialization for +the native (implemented in C++) callee function. + +### CantInlineNativeBadForm + +Unable to inline function call. The native callee function was called with an +unsupported number of arguments, or calling non-constructing functions with new. + +### CantInlineBadType + +Unable to inline function call. The native callee function was called with +arguments with types that the built-in specialization does not support. For +example, calling Math functions on objects. + +### CantInlineNativeNoTemplateObj + +Unable to inline function call. Cannot inline a native constructor (e.g., new +Array) because no template object was cached by the Baseline compiler. (expand) diff --git a/js/src/doc/JITOptimizations/Strategies.md b/js/src/doc/JITOptimizations/Strategies.md new file mode 100644 index 000000000..d4feef310 --- /dev/null +++ b/js/src/doc/JITOptimizations/Strategies.md @@ -0,0 +1,582 @@ +# JIT Optimization Strategies + +SpiderMonkey's optimizing JIT, IonMonkey, uses a number of different +optimization strategies to speed up various operations. The most commonplace +operations that are relevant for fast program execution are property accesses +and function calls. + +Optimization information is currently collected for the following operations: + +- [BinaryArith](#binaryarith) (`+-/*%`) +- [GetProperty](#getprop) (`obj.prop`) +- [SetProperty](#setprop) (`obj.prop = val`) +- [GetElement](#getelem) (`obj[elemName]`) +- [SetElement](#setelem) (`obj[elemName] = val`) +- [Call](#call) (`func(...)`) + +At each operation site, IonMonkey tries a battery of <i>strategies</i>, from +the most optimized but most restrictive to the least optimized but least +restrictive. For each strategy attempted, its <i>outcome</i> is tracked. An +outcome is either success or a reason why the strategy failed to apply. + +This page documents the various optimization strategies and their outcomes. It +provides information on what they attempt to do, what general level of +speed-up they provide, what kind of program characteristics can prevent them +from being used, and common ways to enable the engine to utilize that +optimization. + +## <a name="binaryarith"></a>BinaryArith + +### BinaryArith_Concat + +Attempts to optimize an addition to string concatenation. The types of the operands are checked to contain a hint of being a concatenation. Both have to be string or one has to be a string and the other a type that easily can get converted to string (like numbers). + +### BinaryArith_SpecializedTypes + +Attempts to do a numeric arithmetic based on operand types. One of the inputs need to be a numeric type and the other a simple arithmetic type. + +### BinaryArith_SpecializedOnBaselineTypes + +Just like BinrayArith_SpecializedTypes tries to do a numeric arithmetic, but based on the observed types in the Baseline Compiler. If it succeeds it will roughly give same optimization as BinaryArith_SpecializedTypes, but the deduction is more fragile. In the best case one should try to get this optimization based on the input types. + +### BinaryArith_SharedCache + +Attempts to optimize a binary arithmetic using inline cache. + +### BinaryArith_Call + +Last resort which cannot get specialized at all and takes the slow path to do the arithmetic. + + +## <a name="getprop"></a>GetProperty + +### GetProp_ArgumentsLength + +Attempts to optimize an `arguments.length` property access. This optimization +only works if the arguments object is used in well-understood ways within the +function. The function containing the arguments.length is allowed to use the +arguments object in the following ways without disabling this optimization: + +- Access `arguments.length` +- Access `arguments.callee` +- Access individual args using `arguments[i]` +- Save arguments into variables, as long as those variables cannot be accessed + by any nested function, and as long as there exists no eval nywhere within + the function or nested function definitions. +- Call a function using `f.apply(obj, arguments)` + +If the function contains any use of the arguments object that falls out of the +cases defined above, this optimization will be suppressed. In particular, +`arguments` cannot be returned from the function, or passed as an argument into +calls (except for the `apply` case above). + +### GetProp_ArgumentsCallee + +Attempts to optimize an `arguments.callee` property access. This optimization +only works if the arguments object is used in well-understood ways within the +function. The function containing the `arguments.callee` is allowed to use the +arguments object in the following ways without disabling this optimization: + +- Access arguments.length +- Access arguments.callee +- Access individual args using arguments[i] +- Save arguments into variables, as long as those variables cannot be accessed + by any nested function, and as long as there exists no eval nywhere within + the function or nested function definitions. +- Call a function using `f.apply(obj, arguments)` + +If the function contains any use of the `arguments` object that falls out of +the cases defined above, this optimization will be suppressed. In particular, +arguments cannot be returned from the function, or passed as an argument into +calls (except for the `apply` example listed above). + +### GetProp_InferredConstant + +Attempts to optimize an access to a property that seems to be a constant. It +applies to property accesses on objects which are global-like in that there is +only one instance of them per program. This includes global objects, object +literals defined at the top-level of a script, and top-level function objects. + +This optimization makes the assumption that a property that has not changed +after it was first assigned, is likely a constant property. It then directly +inlines the value of the property into hot code that accesses it. For +example, in the following code: + + var Constants = {}; + Constants.N = 100; + + function testArray(array) { + for (var i = 0; i < array.length; i++) { + if (array[i] > Constants.N) + return true; + } + return false; + } + +Will have the loop compiled into the following when `testArray` gets hot. + + for (var i = 0; i < array.length; i++) { + if (array[i] > 100) + return true; + } + +When this optimization is successful, property access is eliminated entirely +and replaced with an inline constant. + +### GetProp_Constant + +Attempts to optimize reading a property that contains a uniquely-typed (or +"singleton") object. With uniquely-typed objects, it is guaranteed that +no other object has that same type. Unique (or "singleton") types are +assigned to certain kinds of objects, like global objects, top-level +functions, and object literals declared at the top-level of a script. If +a property has always contained the same uniquely-typed object, then the +engine can use the unique type to map back to a specific object, and +eliminate the property access, replacing it with a constant reference to +the object. + +When this optimization is successful, property access is eliminated entirely +and replaced with an inline constant. The different success and failure +conditions are documented below: + +### GetProp_StaticName + +Attempts to optimize a property access on `window` which refers to +a property on the global object. + +### GetProp_TypedObject + +Optimizes accesses to properties on TypedObjects. + +### GetProp_DefiniteSlot + +Optimizes access to a well-known regular property on an object. For this +optimization to succeed, the property needs to be well-defined on the object. +For objects constructed by constructor functions, this means that the property +needs to be defined in the constructor, before any complex logic occurs within +the constructor. + +This is the best case for a regular "field" type property that is not +turned into a constant. It compiles down to a single CPU-level load +instruction. + + function SomeConstructor() { + this.x = 10; // x is a definite slot property + this.y = 10; // y is a definite slot property + someComplicatedFunctionCall(); + this.z = 20; // z is not a definite slot property. + } + +In the above example, the properties `x` and `y` can be determined to always +exist on any instance of `SomeConstructor` at definite locations, allowing +the engine to deterministically infer the position of `x` without a shape +guard. + +This optimization can fail for a number of reasons. If the types observed +at the property access are polymorphic (more than one type), this optimization +cannot succeed. Furthermore, even if the object type is monomorphic, the +optimization will fail if the property being accessed is not a definite +slot as described above. + +### GetProp_Unboxed + +Similar to `GetProp_DefiniteSlot`. Unboxed property reads are possible on +properties which satisfy all the characteristics of a definite slot, and +additionally have been observed to only store values of one kind of value. + +Consider the following constructor: + + function Point(x, y) { + this.x = x; + this.y = y; + } + +If only integers are ever stored in the `x` and `y` properties, +then the instances of `Point` will be represented in an "unboxed" mode - +with the property values stored as raw 4-byte values within the object. + +Objects which have the unboxed optimization are more compact. + +### GetProp_CommonGetter + +Optimizes access to properties which are implemented by a getter function, +where the getter is shared between multiple types. + +This optimization applies most often when the property access site is +polymorphic, but all the object types are derived variants of a single +base class, where the property access refers to a getter on the base +class. + +Consider the following example: + function Base() {} + Base.prototype = { + get x() { return 3; } + }; + + function Derived1() {} + Derived1.prototype = Object.create(Base.prototype); + + function Derived2() {} + Derived1.prototype = Object.create(Base.prototype); + +If a property access for `d.x` sees only instances of both `Derived1` and +`Derived2` for `d`, it can optimize the access to `x` to a call to the +getter function defined on `Base`. + +This optimization applies to shared getters on both pure JS objects as well +as DOM objects. + +### GetProp_InlineAccess + +Optimizes a polymorphic property access where there are only a few different +types of objects seen, and the property on all of the different types is +determinable through a shape-check. + +If a property access is monomorphic and the property's location is determinable +from the object's shape, but the property is not definite (see: +GetProp_DefiniteProperty), then this optimization may be used. + +Alternatively, if the property access is polymorphic, but only has a few +different shapes observed at the access site, this optimization may be used. + +This optimization compiles down to one-or more shape-guarded direct loads +from the object. The following pseudocode describes the kind of machine +code generated by this optimization: + + if obj.shape == Shape1 then + obj.slots[0] + elif obj.shape == Shape2 then + obj.slots[5] + elif obj.shape == Shape3 then + obj.slots[2] + ... + end + +### GetProp_Innerize + +Attempts to optimize a situation where a property access of the form +`window.PROP` can be directly translated into a property access on +the inner global object. + +This optimization will always fail on property accesses which are +not on the window object. + +It is useful because accessing global names via the 'window' object +is a common idiom in web programming. + +### GetProp_InlineCache + +This is the worst-case scenario for a property access optimization. This +strategy is used when all the others fail. The engine simply inserts +an inline cache at the property access site. + +Inline caches start off as a jump to a separate piece of code called +a "fallback". The fallback calls into the interpreter VM (which is +very slow) to perform the operation, and then decides if the operation +can be optimized in that particular case. If so, it generates a new +"stub" (or freestanding piece of jitcode) and changes the inline cache +to jump to the stub. The stub attempts to optimize further occurrences +of that same kind of operation. + +Inline caches are an order of magnitude slower than the other optimization +strategies, and are an indication that the type inference engine has +failed to collect enough information to guide the optimization process. + +## <a name="setprop"></a>SetProperty + +### SetProp_CommonSetter + +Optimizes access to properties which are implemented by a setter function, +where the setter is shared between multiple types. + +This optimization applies most often when the property access site is +polymorphic, but all the object types are derived variants of a single +base class, where the property access refers to a setter on the base +class. + +Consider the following example: + function Base() {} + Base.prototype = { + set x(val) { ... } + }; + + function Derived1() {} + Derived1.prototype = Object.create(Base.prototype); + + function Derived2() {} + Derived1.prototype = Object.create(Base.prototype); + +If a property write for `d.x = val` sees only instances of both `Derived1` and +`Derived2` for `d`, it can optimize the write to `x` to a call to the +setter function defined on `Base`. + +This optimization applies to shared setters on both pure JS objects as well +as DOM objects. + +### SetProp_TypedObject + +Optimizes accesses to properties on TypedObjects. + +### SetProp_DefiniteSlot + +Optimizes a write to a well-known regular property on an object. For this +optimization to succeed, the property needs to be well-defined on the object. +For objects constructed by constructor functions, this means that the property +needs to be defined in the constructor, before any complex logic occurs within +the constructor. + +This is the best case for a regular "field" type property that is not +turned into a constant. It compiles down to a single CPU-level load +instruction. + + function SomeConstructor() { + this.x = 10; // x is a definite slot property + this.y = 10; // y is a definite slot property + someComplicatedFunctionCall(); + this.z = 20; // z is not a definite slot property. + } + +In the above example, the properties `x` and `y` can be determined to always +exist on any instance of `SomeConstructor` at definite locations, allowing +the engine to deterministically infer the position of `x` without a shape +guard. + +This optimization can fail for a number of reasons. If the types observed +at the property access are polymorphic (more than one type), this optimization +cannot succeed. Furthermore, even if the object type is monomorphic, the +optimization will fail if the property being written is not a definite +slot as described above. + +### SetProp_Unboxed + +Similar to `SetProp_DefiniteSlot`. Unboxed property writes are possible on +properties which satisfy all the characteristics of a definite slot, and +additionally have been observed to only store values of one kind of value. + +Consider the following constructor: + + function Point(x, y) { + this.x = x; + this.y = y; + } + +If only integers are ever stored in the `x` and `y` properties, +then the instances of `Point` will be represented in an "unboxed" mode - +with the property values stored as raw 4-byte values within the object. + +Objects which have the unboxed optimization are more compact. + +### SetProp_InlineAccess + +Optimizes a polymorphic property write where there are only a few different +types of objects seen, and the property on all of the different types is +determinable through a shape-check. + +If a property write is monomorphic and the property's location is determinable +from the object's shape, but the property is not definite (see: +GetProp_DefiniteProperty), then this optimization may be used. + +Alternatively, if the property write is polymorphic, but only has a few +different shapes observed at the access site, this optimization may be used. + +This optimization compiles down to one-or more shape-guarded direct stores +to the object. The following pseudocode describes the kind of machine +code generated by this optimization: + + if obj.shape == Shape1 then + obj.slots[0] = val + elif obj.shape == Shape2 then + obj.slots[5] = val + elif obj.shape == Shape3 then + obj.slots[2] = val + ... + end + +### SetProp_InlineCache + +This is the worst-case scenario for a property access optimization. This +strategy is used when all the others fail. The engine simply inserts +an inline cache at the property write site. + +Inline caches start off as a jump to a separate piece of code called +a "fallback". The fallback calls into the interpreter VM (which is +very slow) to perform the operation, and then decides if the operation +can be optimized in that particular case. If so, it generates a new +"stub" (or freestanding piece of jitcode) and changes the inline cache +to jump to the stub. The stub attempts to optimize further occurrences +of that same kind of operation. + +Inline caches are an order of magnitude slower than the other optimization +strategies, and are an indication that the type inference engine has +failed to collect enough information to guide the optimization process. + +## <a name="getelem"></a>GetElement + +### GetElem_TypedObject + +Attempts to optimized element accesses on array Typed Objects. + +### GetElem_Dense + +Attempts to optimize element accesses on densely packed array objects. Dense +arrays are arrays which do not have any 'holes'. This means that the array +has valid values for all indexes from `0` to `length-1`. + +### GetElem_TypedStatic + +Attempts to optimize element accesses on a typed array that can be determined +to always refer to the same array object. If this optimization succeeds, the +'array' object is treated as a constant, and is not looked up or retrieved from +a variable. + +### GetElem_TypedArray + +Attempts to optimize element accesses on a typed array. + +### GetElem_String + +Attempts to optimize element accesses on a string. + +### GetElem_Arguments + +Attempts to optimize element accesses on the `arguments` special object +available in functions. This optimization only works if the arguments +object is used in well-understood ways within the function. The function +containing the arguments.length is allowed to use the arguments object in +the following ways without disabling this optimization: + +- Access `arguments.length` +- Access `arguments.callee` +- Access individual args using `arguments[i]` +- Save `arguments` into variables, as long as those variables cannot be + accessed by any nested function, and as long as there exists no `eval` + anywhere within the function or nested function definitions. +- Call a function using `f.apply(obj, arguments)` + +If the function contains any use of the arguments object that falls out of the +cases defined above, this optimization will be suppressed. In particular, +`arguments` cannot be returned from the function, or passed as an argument into +calls (except for the `apply` case above). + +### GetElem_ArgumentsInlined + +Similar to GetEelem_Arguments, but optimizes cases where the access on +`arguments` is happening within an inlined function. In these cases, an +access of the form `arguments[i]` can be directly translated into a +direct reference to the corresponding argument value in the inlined call. + +Consider the following: + function foo(arg) { + return bar(arg, 3); + } + function bar() { + return arguments[0] + arguments[1]; + } + +In the above case, if `foo` is compiled with Ion, and the call to `bar` +is inlined, then this optimization can transform the entire procedure to +the following pseudo-code: + + compiled foo(arg): + // inlined call to bar(arg, 3) { + return arg + 3; + // } + +### GetElem_InlineCache + +This is the worst-case scenario for a element access optimization. This +strategy is used when all the others fail. The engine simply inserts +an inline cache at the property write site. + +Inline caches start off as a jump to a separate piece of code called +a "fallback". The fallback calls into the interpreter VM (which is +very slow) to perform the operation, and then decides if the operation +can be optimized in that particular case. If so, it generates a new +"stub" (or freestanding piece of jitcode) and changes the inline cache +to jump to the stub. The stub attempts to optimize further occurrences +of that same kind of operation. + +Inline caches are an order of magnitude slower than the other optimization +strategies, and are an indication that the type inference engine has +failed to collect enough information to guide the optimization process. + +## <a name="setelem"></a>SetElement + +### SetElem_TypedObject + +Attempts to optimized element writes on array Typed Objects. + +### SetElem_TypedStatic + +Attempts to optimize element writes on a typed array that can be determined +to always refer to the same array object. If this optimization succeeds, the +'array' object is treated as a constant, and is not looked up or retrieved from +a variable. + +### SetElem_TypedArray + +Attempts to optimize element writes on a typed array. + +### SetElem_Dense + +Attempts to optimize element writes on densely packed array objects. Dense +arrays are arrays which do not have any 'holes'. This means that the array +has valid values for all indexes from `0` to `length-1`. + +### SetElem_Arguments + +Attempts to optimize element writes to the `arguments` special object +available in functions. This optimization only works if the arguments +object is used in well-understood ways within the function. The function +containing the arguments.length is allowed to use the arguments object in +the following ways without disabling this optimization: + +- Access `arguments.length` +- Access `arguments.callee` +- Access individual args using `arguments[i]` +- Save `arguments` into variables, as long as those variables cannot be + accessed by any nested function, and as long as there exists no `eval` + anywhere within the function or nested function definitions. +- Call a function using `f.apply(obj, arguments)` + +If the function contains any use of the arguments object that falls out of the +cases defined above, this optimization will be suppressed. In particular, +`arguments` cannot be returned from the function, or passed as an argument into +calls (except for the `apply` case above). + +### SetElem_InlineCache + +This is the worst-case scenario for a element write optimization. This +strategy is used when all the others fail. The engine simply inserts +an inline cache at the property write site. + +Inline caches start off as a jump to a separate piece of code called +a "fallback". The fallback calls into the interpreter VM (which is +very slow) to perform the operation, and then decides if the operation +can be optimized in that particular case. If so, it generates a new +"stub" (or freestanding piece of jitcode) and changes the inline cache +to jump to the stub. The stub attempts to optimize further occurrences +of that same kind of operation. + +Inline caches are an order of magnitude slower than the other optimization +strategies, and are an indication that the type inference engine has +failed to collect enough information to guide the optimization process. + +## <a name="call"></a>Call + +### Call_Inline + +A function call `f(x)` usually pushes a frame onto the call stack. Inlining a +call site conceptually copies the body of the callee function and pastes it +in place of the call site and avoids pushing a new execution frame. Usually, +hot functions do well to be inlined. This is one of the most important +optimizations the JIT performs. + +Ion inlines both interpreted (i.e., written in JavaScript) functions and +native (i.e., built-ins such as `Math.sin` implemented in C++). + +A successfully inlined call site has the outcome Inlined. + +Failure to inline comes in two flavors: unable (e.g., unable to determine +exact callee) and unwilling (e.g., heuristics concluded that the time-space +tradeoff will not pay off). diff --git a/js/src/doc/JITOptimizations/config.sh b/js/src/doc/JITOptimizations/config.sh new file mode 100644 index 000000000..087384ec6 --- /dev/null +++ b/js/src/doc/JITOptimizations/config.sh @@ -0,0 +1,15 @@ +base-url https://developer.mozilla.org/en-US/docs/Tools/ + +markdown Strategies.md JIT-Optimization/Strategies + label 'jitstrategies' 'JIT Optimization Strategies' + label 'getprop' '#getproperty' 'JIT Optimization Strategies : GetProperty' + label 'setprop' '#setproperty' 'JIT Optimization Strategies : SetProperty' + label 'getelem' '#getelement' 'JIT Optimization Strategies : GetElement' + label 'setelem' '#setelement' 'JIT Optimization Strategies : SetElement' + label 'call' '#call' 'JIT Optimization Strategies : Call' + +markdown Strategies.md JIT-Optimization/Outcomes + label 'jitoutcomes' 'JIT Optimization Outcomes' + label 'general' '#general' 'JIT Optimization Outcomes : General Outcomes' + label 'ic' '#inlinecache' 'JIT Optimization Outcomes : Inline Cache Outcomes' + label 'callinline' '#callinline' 'JIT Optimization Outcomes : Call Inlining Outcomes' |