1==================== 2Constant Interpreter 3==================== 4 5.. contents:: 6 :local: 7 8Introduction 9============ 10 11The constexpr interpreter aims to replace the existing tree evaluator in clang, improving performance on constructs which are executed inefficiently by the evaluator. The interpreter is activated using the following flags: 12 13* ``-fexperimental-new-constant-interpreter`` enables the interpreter, emitting an error if an unsupported feature is encountered 14 15Bytecode Compilation 16==================== 17 18Bytecode compilation is handled in ``ByteCodeStmtGen.h`` for statements and ``ByteCodeExprGen.h`` for expressions. The compiler has two different backends: one to generate bytecode for functions (``ByteCodeEmitter``) and one to directly evaluate expressions as they are compiled, without generating bytecode (``EvalEmitter``). All functions are compiled to bytecode, while toplevel expressions used in constant contexts are directly evaluated since the bytecode would never be reused. This mechanism aims to pave the way towards replacing the evaluator, improving its performance on functions and loops, while being just as fast on single-use toplevel expressions. 19 20The interpreter relies on stack-based, strongly-typed opcodes. The glue logic between the code generator, along with the enumeration and description of opcodes, can be found in ``Opcodes.td``. The opcodes are implemented as generic template methods in ``Interp.h`` and instantiated with the relevant primitive types by the interpreter loop or by the evaluating emitter. 21 22Primitive Types 23--------------- 24 25* ``PT_{U|S}int{8|16|32|64}`` 26 27 Signed or unsigned integers of a specific bit width, implemented using the ```Integral``` type. 28 29* ``PT_{U|S}intFP`` 30 31 Signed or unsigned integers of an arbitrary, but fixed width used to implement 32 integral types which are required by the target, but are not supported by the host. 33 Under the hood, they rely on APValue. The ``Integral`` specialisation for these 34 types is required by opcodes to share an implementation with fixed integrals. 35 36* ``PT_Bool`` 37 38 Representation for boolean types, essentially a 1-bit unsigned ``Integral``. 39 40* ``PT_RealFP`` 41 42 Arbitrary, but fixed precision floating point numbers. Could be specialised in 43 the future similarly to integers in order to improve floating point performance. 44 45* ``PT_Ptr`` 46 47 Pointer type, defined in ``"Pointer.h"``. 48 49* ``PT_FnPtr`` 50 51 Function pointer type, can also be a null function pointer. Defined in ``"Pointer.h"``. 52 53* ``PT_MemPtr`` 54 55 Member pointer type, can also be a null member pointer. Defined in ``"Pointer.h"`` 56 57Composite types 58--------------- 59 60The interpreter distinguishes two kinds of composite types: arrays and records. Unions are represented as records, except a single field can be marked as active. The contents of inactive fields are kept until they 61are reactivated and overwritten. 62 63 64Bytecode Execution 65================== 66 67Bytecode is executed using a stack-based interpreter. The execution context consists of an ``InterpStack``, along with a chain of ``InterpFrame`` objects storing the call frames. Frames are built by call instructions and destroyed by return instructions. They perform one allocation to reserve space for all locals in a single block. These objects store all the required information to emit stack traces whenever evaluation fails. 68 69Memory Organisation 70=================== 71 72Memory management in the interpreter relies on 3 data structures: ``Block`` 73object which store the data and associated inline metadata, ``Pointer`` objects 74which refer to or into blocks, and ``Descriptor`` structures which describe 75blocks and subobjects nested inside blocks. 76 77Blocks 78------ 79 80Blocks contain data interleaved with metadata. They are allocated either statically 81in the code generator (globals, static members, dummy parameter values etc.) or 82dynamically in the interpreter, when creating the frame containing the local variables 83of a function. Blocks are associated with a descriptor that characterises the entire 84allocation, along with a few additional attributes: 85 86* ``IsStatic`` indicates whether the block has static duration in the interpreter, i.e. it is not a local in a frame. 87 88* ``IsExtern`` indicates that the block was created for an extern and the storage cannot be read or written. 89 90* ``DeclID`` identifies each global declaration (it is set to an invalid and irrelevant value for locals) in order to prevent illegal writes and reads involving globals and temporaries with static storage duration. 91 92Static blocks are never deallocated, but local ones might be deallocated even when there are live pointers to them. Pointers are only valid as long as the blocks they point to are valid, so a block with pointers to it whose lifetime ends is kept alive until all pointers to it go out of scope. Since the frame is destroyed on function exit, such blocks are turned into a ``DeadBlock`` and copied to storage managed by the interpreter itself, not the frame. Reads and writes to these blocks are illegal and cause an appropriate diagnostic to be emitted. When the last pointer goes out of scope, dead blocks are also deallocated. 93 94The lifetime of blocks is managed through 3 methods stored in the descriptor of the block: 95 96* **CtorFn**: initializes the metadata which is store in the block, alongside actual data. Invokes the default constructors of objects which are not trivial (``Pointer``, ``RealFP``, etc.) 97* **DtorFn**: invokes the destructors of non-trivial objects. 98* **MoveFn**: moves a block to dead storage. 99 100Non-static blocks track all the pointers into them through an intrusive doubly-linked list, this is required in order to adjust all pointers when transforming a block into a dead block. 101 102Descriptors 103----------- 104 105Descriptor are generated at bytecode compilation time and contain information required to determine if a particular memory access is allowed in constexpr. Even though there is a single descriptor object, it encodes information for several kinds of objects: 106 107* **Primitives** 108 109 A block containing a primitive reserved storage only for the primitive. 110 111* **Arrays of primitives** 112 113 An array of primitives contains a pointer to an ``InitMap`` storage as its first field: the initialisation map is a bit map indicating all elements of the array which were initialised. If the pointer is null, no elements were initialised, while a value of ``(InitMap)-1`` indicates that the object was fully initialised. when all fields are initialised, the map is deallocated and replaced with that token. 114 115 Array elements are stored sequentially, without padding, after the pointer to the map. 116 117* **Arrays of composites and records** 118 119 Each element in an array of composites is preceded by an ``InlineDescriptor``. Descriptors and elements are stored sequentially in the block. Records are laid out identically to arrays of composites: each field and base class is preceded by an inline descriptor. The ``InlineDescriptor`` has the following field: 120 121 * **Offset**: byte offset into the array or record, used to step back to the parent array or record. 122 * **IsConst**: flag indicating if the field is const-qualified. 123 * **IsInitialized**: flag indicating whether the field or element was initialized. For non-primitive fields, this is only relevant for base classes. 124 * **IsBase**: flag indicating whether the record is a base class. In that case, the offset can be used to identify the derived class. 125 * **IsActive**: indicates if the field is the active field of a union. 126 * **IsMutable**: indicates if the field is marked as mutable. 127 128Inline descriptors are filled in by the `CtorFn` of blocks, which leaves storage in an uninitialised, but valid state. 129 130Pointers 131-------- 132 133Pointers track a ``Pointee``, the block to which they point or ``nullptr`` for null pointers, along with a ``Base`` and an ``Offset``. The base identifies the innermost field, while the offset points to an array element relative to the base (including one-past-end pointers). Most subobject the pointer points to in block, while the offset identifies the array element the pointer points to. These two fields allow all pointers to be uniquely identified and disambiguated. 134 135As an example, consider the following structure: 136 137.. code-block:: c 138 139 struct A { 140 struct B { 141 int x; 142 int y; 143 } b; 144 struct C { 145 int a; 146 int b; 147 } c[2]; 148 int z; 149 }; 150 constexpr A a; 151 152On the target, ``&a`` and ``&a.b.x`` are equal. So are ``&a.c[0]`` and ``&a.c[0].a``. In the interpreter, all these pointers must be distinguished since the are all allowed to address distinct range of memory. 153 154In the interpreter, the object would require 240 bytes of storage and would have its field interleaved with metadata. The pointers which can be derived to the object are illustrated in the following diagram: 155 156:: 157 158 0 16 32 40 56 64 80 96 112 120 136 144 160 176 184 200 208 224 240 159 +---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+ 160 + B | D | D | x | D | y | D | D | D | a | D | b | D | D | a | D | b | D | z | 161 +---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+ 162 ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ 163 | | | | | | | &a.c[0].b | | &a.c[1].b | 164 a |&a.b.x &a.y &a.c |&a.c[0].a |&a.c[1].a | 165 &a.b &a.c[0] &a.c[1] &a.z 166 167The ``Base`` offset of all pointers points to the start of a field or an array and is preceded by an inline descriptor (unless ``Base == 0``, pointing to the root). All the relevant attributes can be read from either the inline descriptor or the descriptor of the block. 168 169Array elements are identified by the ``Offset`` field of pointers, pointing to past the inline descriptors for composites and before the actual data in the case of primitive arrays. The ``Offset`` points to the offset where primitives can be read from. As an example, ``a.c + 1`` would have the same base as ``a.c`` since it is an element of ``a.c``, but its offset would point to ``&a.c[1]``. The ``*`` operation narrows the scope of the pointer, adjusting the base to ``&a.c[1]``. The reverse operator, ``&``, expands the scope of ``&a.c[1]``, turning it into ``a.c + 1``. When a one-past-end pointer is narrowed, its offset is set to ``-1`` to indicate that it is an invalid value (expanding returns the past-the-end pointer). As a special case, narrowing ``&a.c`` results in ``&a.c[0]``. The `narrow` and `expand` methods can be used to follow the chain of equivalent pointers. 170 171TODO 172==== 173 174Missing Language Features 175------------------------- 176 177* Definition of externs must override previous declaration 178* Changing the active field of unions 179* Union copy constructors 180* ``typeid`` 181* ``volatile`` 182* ``__builtin_constant_p`` 183* ``std::initializer_list`` 184* lambdas 185* range-based for loops 186* ``vector_size`` 187* ``dynamic_cast`` 188 189Known Bugs 190---------- 191 192* Pointer comparison for equality needs to narrow/expand pointers 193* If execution fails, memory storing APInts and APFloats is leaked when the stack is cleared 194