2021-10-30 04:25:12 +08:00
// This file is part of the Luau programming language and is licensed under MIT License; see LICENSE.txt for details
# pragma once
// clang-format off
// This header contains the bytecode definition for Luau interpreter
// Creating the bytecode is outside the scope of this file and is handled by bytecode builder (BytecodeBuilder.h) and bytecode compiler (Compiler.h)
// Note that ALL enums declared in this file are order-sensitive since the values are baked into bytecode that needs to be processed by legacy clients.
2022-06-24 09:56:00 +08:00
// # Bytecode definitions
2021-10-30 04:25:12 +08:00
// Bytecode instructions are using "word code" - each instruction is one or many 32-bit words.
// The first word in the instruction is always the instruction header, and *must* contain the opcode (enum below) in the least significant byte.
//
// Instruction word can be encoded using one of the following encodings:
// ABC - least-significant byte for the opcode, followed by three bytes, A, B and C; each byte declares a register index, small index into some other table or an unsigned integral value
// AD - least-significant byte for the opcode, followed by A byte, followed by D half-word (16-bit integer). D is a signed integer that commonly specifies constant table index or jump offset
// E - least-significant byte for the opcode, followed by E (24-bit integer). E is a signed integer that commonly specifies a jump offset
//
// Instruction word is sometimes followed by one extra word, indicated as AUX - this is just a 32-bit word and is decoded according to the specification for each opcode.
// For each opcode the encoding is *static* - that is, based on the opcode you know a-priory how large the instruction is, with the exception of NEWCLOSURE
2022-06-24 09:56:00 +08:00
// # Bytecode indices
2021-10-30 04:25:12 +08:00
// Bytecode instructions commonly refer to integer values that define offsets or indices for various entities. For each type, there's a maximum encodable value.
// Note that in some cases, the compiler will set a lower limit than the maximum encodable value is to prevent fragile code into bumping against the limits whenever we change the compilation details.
// Additionally, in some specific instructions such as ANDK, the limit on the encoded value is smaller; this means that if a value is larger, a different instruction must be selected.
//
// Registers: 0-254. Registers refer to the values on the function's stack frame, including arguments.
2023-03-18 03:20:37 +08:00
// Upvalues: 0-199. Upvalues refer to the values stored in the closure object.
2021-10-30 04:25:12 +08:00
// Constants: 0-2^23-1. Constants are stored in a table allocated with each proto; to allow for future bytecode tweaks the encodable value is limited to 23 bits.
// Closures: 0-2^15-1. Closures are created from child protos via a child index; the limit is for the number of closures immediately referenced in each function.
2023-02-09 22:12:58 +08:00
// Jumps: -2^23..2^23. Jump offsets are specified in word increments, so jumping over an instruction may sometimes require an offset of 2 or more. Note that for jump instructions with AUX, the AUX word is included as part of the jump offset.
2022-06-24 09:56:00 +08:00
// # Bytecode versions
// Bytecode serialized format embeds a version number, that dictates both the serialized form as well as the allowed instructions. As long as the bytecode version falls into supported
// range (indicated by LBC_BYTECODE_MIN / LBC_BYTECODE_MAX) and was produced by Luau compiler, it should load and execute correctly.
//
// Note that Luau runtime doesn't provide indefinite bytecode compatibility: support for older versions gets removed over time. As such, bytecode isn't a durable storage format and it's expected
// that Luau users can recompile bytecode from source on Luau version upgrades if necessary.
2022-08-05 06:35:33 +08:00
// # Bytecode version history
//
// Note: due to limitations of the versioning scheme, some bytecode blobs that carry version 2 are using features from version 3. Starting from version 3, version should be sufficient to indicate bytecode compatibility.
//
// Version 1: Baseline version for the open-source release. Supported until 0.521.
2023-02-25 05:49:38 +08:00
// Version 2: Adds Proto::linedefined. Supported until 0.544.
2022-08-05 06:35:33 +08:00
// Version 3: Adds FORGPREP/JUMPXEQK* and enhances AUX encoding for FORGLOOP. Removes FORGLOOP_NEXT/INEXT and JUMPIFEQK/JUMPIFNOTEQK. Currently supported.
2023-09-02 01:58:27 +08:00
// Version 4: Adds Proto::flags, typeinfo, and floor division opcodes IDIV/IDIVK. Currently supported.
Add SUBRK and DIVRK bytecode instructions to bytecode v5 (#1115)
Right now, we can compile R\*K for all arithmetic instructions, but K\*R
gets compiled into two instructions (LOADN/LOADK + arithmetic opcode).
This is problematic since it leads to reduced performance for some code.
However, we'd like to avoid adding reverse variants of ADDK et al for
all opcodes to avoid the increase in I$ footprint for interpreter.
Looking at the arithmetic instructions, % and // don't have interesting
use cases for K\*V; ^ is sometimes used with constant on the left hand
side but this would need to call pow() by necessity in all cases so it
would be slow regardless of the dispatch overhead. This leaves the four
basic arithmetic operations.
For + and \*, we can implement a compiler-side optimization in the
future that transforms K\*R to R\*K automatically. This could either be
done unconditionally at -O2, or conditionally based on the type of the
value (driven by type annotations / inference) -- this technically
changes behavior in presence of metamethods, although it might be
sensible to just always do this because non-commutative +/* are evil.
However, for - and / it is impossible for the compiler to optimize this
in the future, so we need dedicated opcodes. This only increases the
interpreter size by ~300 bytes (~1.5%) on X64.
This makes spectral-norm and math-partial-sums 6% faster; maybe more
importantly, voxelgen gets 1.5% faster (so this change does have
real-world impact).
To avoid the proliferation of bytecode versions this change piggybacks
on the bytecode version bump that was just made in 604 for vector
constants; we would still be able to enable these independently but
we'll consider v5 complete when both are enabled.
Related: #626
---------
Co-authored-by: vegorov-rbx <75688451+vegorov-rbx@users.noreply.github.com>
2023-11-28 23:35:01 +08:00
// Version 5: Adds SUBRK/DIVRK and vector constants. Currently supported.
2022-08-05 06:35:33 +08:00
2024-04-26 06:26:09 +08:00
// # Bytecode type information history
// Version 1: (from bytecode version 4) Type information for function signature. Currently supported.
// Version 2: (from bytecode version 4) Type information for arguments, upvalues, locals and some temporaries. Currently supported.
2022-06-24 09:56:00 +08:00
// Bytecode opcode, part of the instruction header
2021-10-30 04:25:12 +08:00
enum LuauOpcode
{
// NOP: noop
LOP_NOP ,
// BREAK: debugger break
LOP_BREAK ,
// LOADNIL: sets register to nil
// A: target register
LOP_LOADNIL ,
// LOADB: sets register to boolean and jumps to a given short offset (used to compile comparison results into a boolean)
// A: target register
// B: value (0/1)
// C: jump offset
LOP_LOADB ,
// LOADN: sets register to a number literal
// A: target register
// D: value (-32768..32767)
LOP_LOADN ,
Optimize vector literals by storing them in the constant table (#1096)
With this optimization, built-in vector constructor calls with 3/4 arguments are detected by the compiler and turned into vector constants when the arguments are constant numbers.
Requires optimization level 2 because built-ins are not folded otherwise by the compiler.
Bytecode version is bumped because of the new constant type, but old bytecode versions can still be loaded.
The following synthetic benchmark shows ~6.6x improvement.
```
local v
for i = 1, 10000000 do
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
end
```
Also tried a more real world scenario and could see a few percent improvement.
Added a new fast flag LuauVectorLiterals for enabling the feature.
---------
Co-authored-by: Petri Häkkinen <petrih@rmd.remedy.fi>
Co-authored-by: vegorov-rbx <75688451+vegorov-rbx@users.noreply.github.com>
Co-authored-by: Arseny Kapoulkine <arseny.kapoulkine@gmail.com>
2023-11-17 20:54:32 +08:00
// LOADK: sets register to an entry from the constant table from the proto (number/vector/string)
2021-10-30 04:25:12 +08:00
// A: target register
// D: constant table index (0..32767)
LOP_LOADK ,
// MOVE: move (copy) value from one register to another
// A: target register
// B: source register
LOP_MOVE ,
// GETGLOBAL: load value from global table using constant string as a key
// A: target register
// C: predicted slot index (based on hash)
// AUX: constant table index
LOP_GETGLOBAL ,
// SETGLOBAL: set value in global table using constant string as a key
// A: source register
// C: predicted slot index (based on hash)
// AUX: constant table index
LOP_SETGLOBAL ,
// GETUPVAL: load upvalue from the upvalue table for the current function
// A: target register
2023-03-18 03:20:37 +08:00
// B: upvalue index
2021-10-30 04:25:12 +08:00
LOP_GETUPVAL ,
// SETUPVAL: store value into the upvalue table for the current function
// A: target register
2023-03-18 03:20:37 +08:00
// B: upvalue index
2021-10-30 04:25:12 +08:00
LOP_SETUPVAL ,
// CLOSEUPVALS: close (migrate to heap) all upvalues that were captured for registers >= target
// A: target register
LOP_CLOSEUPVALS ,
// GETIMPORT: load imported global table global from the constant table
// A: target register
// D: constant table index (0..32767); we assume that imports are loaded into the constant table
// AUX: 3 10-bit indices of constant strings that, combined, constitute an import path; length of the path is set by the top 2 bits (1,2,3)
LOP_GETIMPORT ,
// GETTABLE: load value from table into target register using key from register
// A: target register
// B: table register
// C: index register
LOP_GETTABLE ,
// SETTABLE: store source register into table using key from register
// A: source register
// B: table register
// C: index register
LOP_SETTABLE ,
// GETTABLEKS: load value from table into target register using constant string as a key
// A: target register
// B: table register
// C: predicted slot index (based on hash)
// AUX: constant table index
LOP_GETTABLEKS ,
// SETTABLEKS: store source register into table using constant string as a key
// A: source register
// B: table register
// C: predicted slot index (based on hash)
// AUX: constant table index
LOP_SETTABLEKS ,
// GETTABLEN: load value from table into target register using small integer index as a key
// A: target register
// B: table register
// C: index-1 (index is 1..256)
LOP_GETTABLEN ,
// SETTABLEN: store source register into table using small integer index as a key
// A: source register
// B: table register
// C: index-1 (index is 1..256)
LOP_SETTABLEN ,
// NEWCLOSURE: create closure from a child proto; followed by a CAPTURE instruction for each upvalue
// A: target register
// D: child proto index (0..32767)
LOP_NEWCLOSURE ,
// NAMECALL: prepare to call specified method by name by loading function from source register using constant index into target register and copying source register into target register + 1
// A: target register
// B: source register
// C: predicted slot index (based on hash)
// AUX: constant table index
// Note that this instruction must be followed directly by CALL; it prepares the arguments
// This instruction is roughly equivalent to GETTABLEKS + MOVE pair, but we need a special instruction to support custom __namecall metamethod
LOP_NAMECALL ,
// CALL: call specified function
// A: register where the function object lives, followed by arguments; results are placed starting from the same register
// B: argument count + 1, or 0 to preserve all arguments up to top (MULTRET)
// C: result count + 1, or 0 to preserve all values and adjust top (MULTRET)
LOP_CALL ,
// RETURN: returns specified values from the function
// A: register where the returned values start
// B: number of returned values + 1, or 0 to return all values up to top (MULTRET)
LOP_RETURN ,
// JUMP: jumps to target offset
// D: jump offset (-32768..32767; 0 means "next instruction" aka "don't jump")
LOP_JUMP ,
// JUMPBACK: jumps to target offset; this is equivalent to JUMP but is used as a safepoint to be able to interrupt while/repeat loops
// D: jump offset (-32768..32767; 0 means "next instruction" aka "don't jump")
LOP_JUMPBACK ,
// JUMPIF: jumps to target offset if register is not nil/false
// A: source register
// D: jump offset (-32768..32767; 0 means "next instruction" aka "don't jump")
LOP_JUMPIF ,
// JUMPIFNOT: jumps to target offset if register is nil/false
// A: source register
// D: jump offset (-32768..32767; 0 means "next instruction" aka "don't jump")
LOP_JUMPIFNOT ,
// JUMPIFEQ, JUMPIFLE, JUMPIFLT, JUMPIFNOTEQ, JUMPIFNOTLE, JUMPIFNOTLT: jumps to target offset if the comparison is true (or false, for NOT variants)
// A: source register 1
2023-02-09 22:12:58 +08:00
// D: jump offset (-32768..32767; 1 means "next instruction" aka "don't jump")
2021-10-30 04:25:12 +08:00
// AUX: source register 2
LOP_JUMPIFEQ ,
LOP_JUMPIFLE ,
LOP_JUMPIFLT ,
LOP_JUMPIFNOTEQ ,
LOP_JUMPIFNOTLE ,
LOP_JUMPIFNOTLT ,
// ADD, SUB, MUL, DIV, MOD, POW: compute arithmetic operation between two source registers and put the result into target register
// A: target register
// B: source register 1
// C: source register 2
LOP_ADD ,
LOP_SUB ,
LOP_MUL ,
LOP_DIV ,
LOP_MOD ,
LOP_POW ,
// ADDK, SUBK, MULK, DIVK, MODK, POWK: compute arithmetic operation between the source register and a constant and put the result into target register
// A: target register
// B: source register
Add SUBRK and DIVRK bytecode instructions to bytecode v5 (#1115)
Right now, we can compile R\*K for all arithmetic instructions, but K\*R
gets compiled into two instructions (LOADN/LOADK + arithmetic opcode).
This is problematic since it leads to reduced performance for some code.
However, we'd like to avoid adding reverse variants of ADDK et al for
all opcodes to avoid the increase in I$ footprint for interpreter.
Looking at the arithmetic instructions, % and // don't have interesting
use cases for K\*V; ^ is sometimes used with constant on the left hand
side but this would need to call pow() by necessity in all cases so it
would be slow regardless of the dispatch overhead. This leaves the four
basic arithmetic operations.
For + and \*, we can implement a compiler-side optimization in the
future that transforms K\*R to R\*K automatically. This could either be
done unconditionally at -O2, or conditionally based on the type of the
value (driven by type annotations / inference) -- this technically
changes behavior in presence of metamethods, although it might be
sensible to just always do this because non-commutative +/* are evil.
However, for - and / it is impossible for the compiler to optimize this
in the future, so we need dedicated opcodes. This only increases the
interpreter size by ~300 bytes (~1.5%) on X64.
This makes spectral-norm and math-partial-sums 6% faster; maybe more
importantly, voxelgen gets 1.5% faster (so this change does have
real-world impact).
To avoid the proliferation of bytecode versions this change piggybacks
on the bytecode version bump that was just made in 604 for vector
constants; we would still be able to enable these independently but
we'll consider v5 complete when both are enabled.
Related: #626
---------
Co-authored-by: vegorov-rbx <75688451+vegorov-rbx@users.noreply.github.com>
2023-11-28 23:35:01 +08:00
// C: constant table index (0..255); must refer to a number
2021-10-30 04:25:12 +08:00
LOP_ADDK ,
LOP_SUBK ,
LOP_MULK ,
LOP_DIVK ,
LOP_MODK ,
LOP_POWK ,
2021-11-04 22:50:46 +08:00
// AND, OR: perform `and` or `or` operation (selecting first or second register based on whether the first one is truthy) and put the result into target register
2021-10-30 04:25:12 +08:00
// A: target register
// B: source register 1
// C: source register 2
LOP_AND ,
LOP_OR ,
2021-11-04 22:50:46 +08:00
// ANDK, ORK: perform `and` or `or` operation (selecting source register or constant based on whether the source register is truthy) and put the result into target register
2021-10-30 04:25:12 +08:00
// A: target register
// B: source register
// C: constant table index (0..255)
LOP_ANDK ,
LOP_ORK ,
// CONCAT: concatenate all strings between B and C (inclusive) and put the result into A
// A: target register
// B: source register start
// C: source register end
LOP_CONCAT ,
// NOT, MINUS, LENGTH: compute unary operation for source register and put the result into target register
// A: target register
// B: source register
LOP_NOT ,
LOP_MINUS ,
LOP_LENGTH ,
// NEWTABLE: create table in target register
// A: target register
// B: table size, stored as 0 for v=0 and ceil(log2(v))+1 for v!=0
// AUX: array size
LOP_NEWTABLE ,
// DUPTABLE: duplicate table using the constant table template to target register
// A: target register
// D: constant table index (0..32767)
LOP_DUPTABLE ,
// SETLIST: set a list of values to table in target register
// A: target register
// B: source register start
// C: value count + 1, or 0 to use all values up to top (MULTRET)
// AUX: table index to start from
LOP_SETLIST ,
// FORNPREP: prepare a numeric for loop, jump over the loop if first iteration doesn't need to run
// A: target register; numeric for loops assume a register layout [limit, step, index, variable]
// D: jump offset (-32768..32767)
// limit/step are immutable, index isn't visible to user code since it's copied into variable
LOP_FORNPREP ,
// FORNLOOP: adjust loop variables for one iteration, jump back to the loop header if loop needs to continue
// A: target register; see FORNPREP for register layout
// D: jump offset (-32768..32767)
LOP_FORNLOOP ,
// FORGLOOP: adjust loop variables for one iteration of a generic for loop, jump back to the loop header if loop needs to continue
// A: target register; generic for loops assume a register layout [generator, state, index, variables...]
// D: jump offset (-32768..32767)
2022-07-01 07:52:43 +08:00
// AUX: variable count (1..255) in the low 8 bits, high bit indicates whether to use ipairs-style traversal in the fast path
2021-10-30 04:25:12 +08:00
// loop variables are adjusted by calling generator(state, index) and expecting it to return a tuple that's copied to the user variables
// the first variable is then copied into index; generator/state are immutable, index isn't visible to user code
LOP_FORGLOOP ,
2022-09-09 06:14:25 +08:00
// FORGPREP_INEXT: prepare FORGLOOP with 2 output variables (no AUX encoding), assuming generator is luaB_inext, and jump to FORGLOOP
// A: target register (see FORGLOOP for register layout)
2021-10-30 04:25:12 +08:00
LOP_FORGPREP_INEXT ,
2022-09-09 06:14:25 +08:00
// removed in v3
LOP_DEP_FORGLOOP_INEXT ,
// FORGPREP_NEXT: prepare FORGLOOP with 2 output variables (no AUX encoding), assuming generator is luaB_next, and jump to FORGLOOP
// A: target register (see FORGLOOP for register layout)
2021-10-30 04:25:12 +08:00
LOP_FORGPREP_NEXT ,
2022-09-09 06:14:25 +08:00
2023-05-26 05:36:34 +08:00
// NATIVECALL: start executing new function in native code
// this is a pseudo-instruction that is never emitted by bytecode compiler, but can be constructed at runtime to accelerate native code dispatch
LOP_NATIVECALL ,
2021-10-30 04:25:12 +08:00
// GETVARARGS: copy variables into the target register from vararg storage for current function
// A: target register
// B: variable count + 1, or 0 to copy all variables and adjust top (MULTRET)
LOP_GETVARARGS ,
// DUPCLOSURE: create closure from a pre-created function object (reusing it unless environments diverge)
// A: target register
// D: constant table index (0..32767)
LOP_DUPCLOSURE ,
// PREPVARARGS: prepare stack for variadic functions so that GETVARARGS works correctly
// A: number of fixed arguments
LOP_PREPVARARGS ,
// LOADKX: sets register to an entry from the constant table from the proto (number/string)
// A: target register
// AUX: constant table index
LOP_LOADKX ,
// JUMPX: jumps to the target offset; like JUMPBACK, supports interruption
// E: jump offset (-2^23..2^23; 0 means "next instruction" aka "don't jump")
LOP_JUMPX ,
// FASTCALL: perform a fast call of a built-in function
// A: builtin function id (see LuauBuiltinFunction)
// C: jump offset to get to following CALL
// FASTCALL is followed by one of (GETIMPORT, MOVE, GETUPVAL) instructions and by CALL instruction
// This is necessary so that if FASTCALL can't perform the call inline, it can continue normal execution
// If FASTCALL *can* perform the call, it jumps over the instructions *and* over the next CALL
// Note that FASTCALL will read the actual call arguments, such as argument/result registers and counts, from the CALL instruction
LOP_FASTCALL ,
// COVERAGE: update coverage information stored in the instruction
// E: hit count for the instruction (0..2^23-1)
// The hit count is incremented by VM every time the instruction is executed, and saturates at 2^23-1
LOP_COVERAGE ,
// CAPTURE: capture a local or an upvalue as an upvalue into a newly created closure; only valid after NEWCLOSURE
// A: capture type, see LuauCaptureType
// B: source register (for VAL/REF) or upvalue index (for UPVAL/UPREF)
LOP_CAPTURE ,
Add SUBRK and DIVRK bytecode instructions to bytecode v5 (#1115)
Right now, we can compile R\*K for all arithmetic instructions, but K\*R
gets compiled into two instructions (LOADN/LOADK + arithmetic opcode).
This is problematic since it leads to reduced performance for some code.
However, we'd like to avoid adding reverse variants of ADDK et al for
all opcodes to avoid the increase in I$ footprint for interpreter.
Looking at the arithmetic instructions, % and // don't have interesting
use cases for K\*V; ^ is sometimes used with constant on the left hand
side but this would need to call pow() by necessity in all cases so it
would be slow regardless of the dispatch overhead. This leaves the four
basic arithmetic operations.
For + and \*, we can implement a compiler-side optimization in the
future that transforms K\*R to R\*K automatically. This could either be
done unconditionally at -O2, or conditionally based on the type of the
value (driven by type annotations / inference) -- this technically
changes behavior in presence of metamethods, although it might be
sensible to just always do this because non-commutative +/* are evil.
However, for - and / it is impossible for the compiler to optimize this
in the future, so we need dedicated opcodes. This only increases the
interpreter size by ~300 bytes (~1.5%) on X64.
This makes spectral-norm and math-partial-sums 6% faster; maybe more
importantly, voxelgen gets 1.5% faster (so this change does have
real-world impact).
To avoid the proliferation of bytecode versions this change piggybacks
on the bytecode version bump that was just made in 604 for vector
constants; we would still be able to enable these independently but
we'll consider v5 complete when both are enabled.
Related: #626
---------
Co-authored-by: vegorov-rbx <75688451+vegorov-rbx@users.noreply.github.com>
2023-11-28 23:35:01 +08:00
// SUBRK, DIVRK: compute arithmetic operation between the constant and a source register and put the result into target register
// A: target register
// B: source register
// C: constant table index (0..255); must refer to a number
LOP_SUBRK ,
LOP_DIVRK ,
2021-10-30 04:25:12 +08:00
// FASTCALL1: perform a fast call of a built-in function using 1 register argument
// A: builtin function id (see LuauBuiltinFunction)
// B: source argument register
// C: jump offset to get to following CALL
LOP_FASTCALL1 ,
// FASTCALL2: perform a fast call of a built-in function using 2 register arguments
// A: builtin function id (see LuauBuiltinFunction)
// B: source argument register
// C: jump offset to get to following CALL
// AUX: source register 2 in least-significant byte
LOP_FASTCALL2 ,
// FASTCALL2K: perform a fast call of a built-in function using 1 register argument and 1 constant argument
// A: builtin function id (see LuauBuiltinFunction)
// B: source argument register
// C: jump offset to get to following CALL
// AUX: constant index
LOP_FASTCALL2K ,
2022-05-06 08:03:43 +08:00
// FORGPREP: prepare loop variables for a generic for loop, jump to the loop backedge unconditionally
// A: target register; generic for loops assume a register layout [generator, state, index, variables...]
// D: jump offset (-32768..32767)
LOP_FORGPREP ,
2022-08-05 06:35:33 +08:00
// JUMPXEQKNIL, JUMPXEQKB: jumps to target offset if the comparison with constant is true (or false, see AUX)
// A: source register 1
2023-02-09 22:12:58 +08:00
// D: jump offset (-32768..32767; 1 means "next instruction" aka "don't jump")
2022-08-05 06:35:33 +08:00
// AUX: constant value (for boolean) in low bit, NOT flag (that flips comparison result) in high bit
LOP_JUMPXEQKNIL ,
LOP_JUMPXEQKB ,
// JUMPXEQKN, JUMPXEQKS: jumps to target offset if the comparison with constant is true (or false, see AUX)
// A: source register 1
2023-02-09 22:12:58 +08:00
// D: jump offset (-32768..32767; 1 means "next instruction" aka "don't jump")
2022-08-05 06:35:33 +08:00
// AUX: constant table index in low 24 bits, NOT flag (that flips comparison result) in high bit
LOP_JUMPXEQKN ,
LOP_JUMPXEQKS ,
2023-09-02 01:58:27 +08:00
// IDIV: compute floor division between two source registers and put the result into target register
// A: target register
// B: source register 1
// C: source register 2
LOP_IDIV ,
// IDIVK compute floor division between the source register and a constant and put the result into target register
// A: target register
// B: source register
// C: constant table index (0..255)
LOP_IDIVK ,
2021-10-30 04:25:12 +08:00
// Enum entry for number of opcodes, not a valid opcode by itself!
LOP__COUNT
} ;
// Bytecode instruction header: it's always a 32-bit integer, with low byte (first byte in little endian) containing the opcode
// Some instruction types require more data and have more 32-bit integers following the header
# define LUAU_INSN_OP(insn) ((insn) & 0xff)
// ABC encoding: three 8-bit values, containing registers or small numbers
# define LUAU_INSN_A(insn) (((insn) >> 8) & 0xff)
# define LUAU_INSN_B(insn) (((insn) >> 16) & 0xff)
# define LUAU_INSN_C(insn) (((insn) >> 24) & 0xff)
// AD encoding: one 8-bit value, one signed 16-bit value
# define LUAU_INSN_D(insn) (int32_t(insn) >> 16)
// E encoding: one signed 24-bit value
# define LUAU_INSN_E(insn) (int32_t(insn) >> 8)
// Bytecode tags, used internally for bytecode encoded as a string
enum LuauBytecodeTag
{
2022-06-24 09:56:00 +08:00
// Bytecode version; runtime supports [MIN, MAX], compiler emits TARGET by default but may emit a higher version when flags are enabled
2022-09-16 06:38:17 +08:00
LBC_VERSION_MIN = 3 ,
Optimize vector literals by storing them in the constant table (#1096)
With this optimization, built-in vector constructor calls with 3/4 arguments are detected by the compiler and turned into vector constants when the arguments are constant numbers.
Requires optimization level 2 because built-ins are not folded otherwise by the compiler.
Bytecode version is bumped because of the new constant type, but old bytecode versions can still be loaded.
The following synthetic benchmark shows ~6.6x improvement.
```
local v
for i = 1, 10000000 do
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
end
```
Also tried a more real world scenario and could see a few percent improvement.
Added a new fast flag LuauVectorLiterals for enabling the feature.
---------
Co-authored-by: Petri Häkkinen <petrih@rmd.remedy.fi>
Co-authored-by: vegorov-rbx <75688451+vegorov-rbx@users.noreply.github.com>
Co-authored-by: Arseny Kapoulkine <arseny.kapoulkine@gmail.com>
2023-11-17 20:54:32 +08:00
LBC_VERSION_MAX = 5 ,
2024-02-16 10:04:39 +08:00
LBC_VERSION_TARGET = 5 ,
2023-06-17 01:35:18 +08:00
// Type encoding version
2024-04-26 06:26:09 +08:00
LBC_TYPE_VERSION_DEPRECATED = 1 ,
2024-06-01 03:18:18 +08:00
LBC_TYPE_VERSION_MIN = 1 ,
LBC_TYPE_VERSION_MAX = 3 ,
LBC_TYPE_VERSION_TARGET = 3 ,
2021-10-30 04:25:12 +08:00
// Types of constant table entries
LBC_CONSTANT_NIL = 0 ,
LBC_CONSTANT_BOOLEAN ,
LBC_CONSTANT_NUMBER ,
LBC_CONSTANT_STRING ,
LBC_CONSTANT_IMPORT ,
LBC_CONSTANT_TABLE ,
LBC_CONSTANT_CLOSURE ,
Optimize vector literals by storing them in the constant table (#1096)
With this optimization, built-in vector constructor calls with 3/4 arguments are detected by the compiler and turned into vector constants when the arguments are constant numbers.
Requires optimization level 2 because built-ins are not folded otherwise by the compiler.
Bytecode version is bumped because of the new constant type, but old bytecode versions can still be loaded.
The following synthetic benchmark shows ~6.6x improvement.
```
local v
for i = 1, 10000000 do
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
v = vector(1, 2, 3)
end
```
Also tried a more real world scenario and could see a few percent improvement.
Added a new fast flag LuauVectorLiterals for enabling the feature.
---------
Co-authored-by: Petri Häkkinen <petrih@rmd.remedy.fi>
Co-authored-by: vegorov-rbx <75688451+vegorov-rbx@users.noreply.github.com>
Co-authored-by: Arseny Kapoulkine <arseny.kapoulkine@gmail.com>
2023-11-17 20:54:32 +08:00
LBC_CONSTANT_VECTOR ,
2021-10-30 04:25:12 +08:00
} ;
2023-06-17 01:35:18 +08:00
// Type table tags
2023-07-15 02:08:53 +08:00
enum LuauBytecodeType
2023-06-17 01:35:18 +08:00
{
LBC_TYPE_NIL = 0 ,
LBC_TYPE_BOOLEAN ,
LBC_TYPE_NUMBER ,
LBC_TYPE_STRING ,
LBC_TYPE_TABLE ,
LBC_TYPE_FUNCTION ,
LBC_TYPE_THREAD ,
LBC_TYPE_USERDATA ,
LBC_TYPE_VECTOR ,
2023-10-28 05:18:41 +08:00
LBC_TYPE_BUFFER ,
2023-06-17 01:35:18 +08:00
LBC_TYPE_ANY = 15 ,
2024-06-01 03:18:18 +08:00
LBC_TYPE_TAGGED_USERDATA_BASE = 64 ,
LBC_TYPE_TAGGED_USERDATA_END = 64 + 32 ,
2023-06-17 01:35:18 +08:00
LBC_TYPE_OPTIONAL_BIT = 1 < < 7 ,
LBC_TYPE_INVALID = 256 ,
} ;
2021-10-30 04:25:12 +08:00
// Builtin function ids, used in LOP_FASTCALL
enum LuauBuiltinFunction
{
LBF_NONE = 0 ,
// assert()
LBF_ASSERT ,
// math.
LBF_MATH_ABS ,
LBF_MATH_ACOS ,
LBF_MATH_ASIN ,
LBF_MATH_ATAN2 ,
LBF_MATH_ATAN ,
LBF_MATH_CEIL ,
LBF_MATH_COSH ,
LBF_MATH_COS ,
LBF_MATH_DEG ,
LBF_MATH_EXP ,
LBF_MATH_FLOOR ,
LBF_MATH_FMOD ,
LBF_MATH_FREXP ,
LBF_MATH_LDEXP ,
LBF_MATH_LOG10 ,
LBF_MATH_LOG ,
LBF_MATH_MAX ,
LBF_MATH_MIN ,
LBF_MATH_MODF ,
LBF_MATH_POW ,
LBF_MATH_RAD ,
LBF_MATH_SINH ,
LBF_MATH_SIN ,
LBF_MATH_SQRT ,
LBF_MATH_TANH ,
LBF_MATH_TAN ,
// bit32.
LBF_BIT32_ARSHIFT ,
LBF_BIT32_BAND ,
LBF_BIT32_BNOT ,
LBF_BIT32_BOR ,
LBF_BIT32_BXOR ,
LBF_BIT32_BTEST ,
LBF_BIT32_EXTRACT ,
LBF_BIT32_LROTATE ,
LBF_BIT32_LSHIFT ,
LBF_BIT32_REPLACE ,
LBF_BIT32_RROTATE ,
LBF_BIT32_RSHIFT ,
// type()
LBF_TYPE ,
// string.
LBF_STRING_BYTE ,
LBF_STRING_CHAR ,
LBF_STRING_LEN ,
// typeof()
LBF_TYPEOF ,
// string.
LBF_STRING_SUB ,
// math.
LBF_MATH_CLAMP ,
LBF_MATH_SIGN ,
LBF_MATH_ROUND ,
// raw*
LBF_RAWSET ,
LBF_RAWGET ,
LBF_RAWEQUAL ,
// table.
LBF_TABLE_INSERT ,
LBF_TABLE_UNPACK ,
// vector ctor
LBF_VECTOR ,
2021-11-12 22:27:34 +08:00
// bit32.count
LBF_BIT32_COUNTLZ ,
LBF_BIT32_COUNTRZ ,
2022-01-28 07:46:05 +08:00
// select(_, ...)
LBF_SELECT_VARARG ,
2022-07-01 07:52:43 +08:00
// rawlen
LBF_RAWLEN ,
2022-08-19 05:32:08 +08:00
// bit32.extract(_, k, k)
LBF_BIT32_EXTRACTK ,
2022-09-30 06:23:10 +08:00
// get/setmetatable
LBF_GETMETATABLE ,
LBF_SETMETATABLE ,
2023-07-28 23:13:53 +08:00
// tonumber/tostring
LBF_TONUMBER ,
LBF_TOSTRING ,
2023-10-23 23:00:48 +08:00
// bit32.byteswap(n)
LBF_BIT32_BYTESWAP ,
2023-10-28 05:18:41 +08:00
// buffer.
LBF_BUFFER_READI8 ,
LBF_BUFFER_READU8 ,
LBF_BUFFER_WRITEU8 ,
LBF_BUFFER_READI16 ,
LBF_BUFFER_READU16 ,
LBF_BUFFER_WRITEU16 ,
LBF_BUFFER_READI32 ,
LBF_BUFFER_READU32 ,
LBF_BUFFER_WRITEU32 ,
LBF_BUFFER_READF32 ,
LBF_BUFFER_WRITEF32 ,
LBF_BUFFER_READF64 ,
LBF_BUFFER_WRITEF64 ,
2021-10-30 04:25:12 +08:00
} ;
// Capture type, used in LOP_CAPTURE
enum LuauCaptureType
{
LCT_VAL = 0 ,
LCT_REF ,
LCT_UPVAL ,
} ;
2023-07-28 23:13:53 +08:00
// Proto flag bitmask, stored in Proto::flags
enum LuauProtoFlag
{
// used to tag main proto for modules with --!native
LPF_NATIVE_MODULE = 1 < < 0 ,
2023-10-07 03:02:32 +08:00
// used to tag individual protos as not profitable to compile natively
LPF_NATIVE_COLD = 1 < < 1 ,
2023-07-28 23:13:53 +08:00
} ;