This change adds support for tagged lightuserdata and optional custom
typenames for lightuserdata.
Background: Lightuserdata is an efficient representation for many kinds
of unmanaged handles and resources in a game engine. However, currently
the VM only supports one kind of lightuserdata, which makes it
problematic in practice. For example, it's not possible to distinguish
between different kinds of lightuserdata in Lua bindings, which can lead
to unsafe practices and even crashes when a wrong kind of lightuserdata
is passed to a binding function. Tagged lightuserdata work similarly to
tagged userdata, i.e. they allow checking the tag quickly using
lua_tolightuserdatatagged (or lua_lightuserdatatag).
The tag is stored in the 'extra' field of TValue so it will add no cost
to the (untagged) lightuserdata type.
Alternatives would be to use full userdata values or use bitpacking to
embed type information into lightuserdata on application level.
Unfortunately these options are not that great in practice: full
userdata have major performance implications and bitpacking fails in
cases where full 64 bits are already used (e.g. pointers or 64-bit
hashes).
Lightuserdata names are not strictly necessary but they are rather
convenient when debugging Lua code. More precise error messages and
tostring returning more specific typename are useful to have in practice
(e.g. "resource" or "entity" instead of the more generic "userdata").
Impl note: I did not add support for renaming tags in
lua_setlightuserdataname as I'm not sure if it's possible to free fixed
strings. If it's simple enough, maybe we should allow renaming (although
I can't think of a specific need for it)?
---------
Co-authored-by: Petri Häkkinen <petrih@rmd.remedy.fi>
New Solver
* Improvements to data flow analysis
Native Code Generation
* Block limit is now per-function instead of per-module
Co-authored-by: Alexander McCord <amccord@roblox.com>
Co-authored-by: Andy Friesen <afriesen@roblox.com>
Co-authored-by: Aviral Goel <agoel@roblox.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
This PR resolves the RequireByString test suite failures in CI on
Ubuntu. The issue was that paths on linux are case sensitive and thus
MacOS/Windows machines simply behaved as if the paths to the test files
were case-insensitive, but the linux machine was unable to find the test
files. Those tests were `#ifdef'd` out, and have been un `#ifdef'd` to
demonstrate these changes work.
Old style require is now called with `pcall` to support comparing
against Lua.
New style require is now a third option.
Edit: this will be a temporary solution until the 'paths' support in
.luaurc is fixed.
`cmake_minimum_required` is a blocking check and raising required
minimal CMake version only for the fuzzer is not good for the users
(<3.19 was supported before, but not sure which exact one)
- Implemented [Require by String with Relative
Paths](https://github.com/luau-lang/rfcs/blob/master/docs/new-require-by-string-semantics.md)
RFC
- Implemented [Require by String with
Aliases](https://github.com/luau-lang/rfcs/blob/master/docs/require-by-string-aliases.md)
RFC with support for `paths` and `alias` arrays in .luarc
- Added SUBRK and DIVRK bytecode instructions to speed up
constant-number and constant/number operations
- Added `--vector-lib`, `--vector-ctor` and `--vector-type` options to
luau-compile to support code with vectors
New Solver
- Correctness fixes to subtyping
- Improvements to dataflow analysis
Native Code Generation
- Added bytecode analysis pass to predict type tags used in operations
- Fixed rare cases of numerical loops being generated without an
interrupt instruction
- Restored optimization data propagation into the linear block
- Duplicate buffer length checks are optimized away
Miscellaneous
- Small performance improvements to new non-strict mode
- Introduced more scripts for fuzzing Luau and processing the results,
including fuzzer build support for CMake
Co-authored-by: Alexander McCord <amccord@roblox.com>
Co-authored-by: Andy Friesen <afriesen@roblox.com>
Co-authored-by: Aviral Goel <agoel@roblox.com>
Co-authored-by: David Cope <dcope@roblox.com>
Co-authored-by: Lily Brown <lbrown@roblox.com>
Co-authored-by: Vighnesh Vijay <vvijay@roblox.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
---------
Co-authored-by: Aaron Weiss <aaronweiss@roblox.com>
Co-authored-by: Alexander McCord <amccord@roblox.com>
Co-authored-by: Andy Friesen <afriesen@roblox.com>
Co-authored-by: Aviral Goel <agoel@roblox.com>
Co-authored-by: David Cope <dcope@roblox.com>
Co-authored-by: Lily Brown <lbrown@roblox.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
- Add SUBRK and DIVRK bytecode instructions
- Enables future performance optimizations
Miscellaneous
- Small performance improvements to new non-strict mode
- Introduce more scripts for fuzzing
- Improcements to dataflow analysis
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>
We need ubuntu-20.04 for coverage analysis (clang after 10 doesn't seem
to properly interact with gcov version used for codecov) and for
releases (to produce binaries targeting earlier glibc).
However, we still should be verifying that Luau builds on latest,
because newer toolchains have stricter standard library headers and/or
warnings; without this we're at risk of constantly regressing the build
for packaging or external applications.
Note that release.yml can probably just be deleted but for now we simply
adjust it.
Closes#818
We set the `moduleName` of the source module to the provided
`packageName`.
This then populates the relevant `definitionModuleName`'s on
CTV/FTV/TTVs, so it allows us to lookup where they originated from.
The one place I can see this having an impact inside of Luau code is the
following:
1fa8311a18/Analysis/src/Error.cpp (L70-L90)
Which will potentially change they way error messages are formatted.
Should there be a way to exclude definition files when calling
`getDefinitionModuleName`, so we can preserve this behaviour?
New Solver
* New algorithm for inferring the types of locals that have no
annotations. This
algorithm is very conservative by default, but is augmented with some
control
flow awareness to handle most common scenarios.
* Fix bugs in type inference of tables
* Improve performance of by switching out standard C++ containers for
`DenseHashMap`
* Infrastructure to support clearer error messages in strict mode
Native Code Generation
* Fix a lowering issue with buffer.writeu8 and 0x80-0xff values: A
constant
argument wasn't truncated to the target type range and that causes an
assertion failure in `build.mov`.
* Store full lightuserdata value in loop iteration protocol lowering
* Add analysis to compute function bytecode distribution
* This includes a class to analyze the bytecode operator distribution
per
function and a CLI tool that produces a JSON report. See the new cmake
target `Luau.Bytecode.CLI`
---------
Co-authored-by: Aaron Weiss <aaronweiss@roblox.com>
Co-authored-by: Alexander McCord <amccord@roblox.com>
Co-authored-by: Andy Friesen <afriesen@roblox.com>
Co-authored-by: Aviral Goel <agoel@roblox.com>
Co-authored-by: Lily Brown <lbrown@roblox.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
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>
With gcc 14 some C++ Standard Library headers have been changed to no
longer include other headers that were used internally by the library.
In luau's case it is the `<algorithm>` header.
# What's changed?
- Record the location of properties for table types (closes#802)
- Implement stricter UTF-8 validations as per the RFC
(https://github.com/luau-lang/rfcs/pull/1)
- Implement `buffer` as a new type in both the old and new solvers.
- Changed errors produced by some `buffer` builtins to be a bit more
generic to avoid platform-dependent error messages.
- Fixed a bug where `Unifier` would copy some persistent types, tripping
some internal assertions.
- Type checking rules on relational operators is now a little bit more
lax.
- Improve dead code elimination for some `if` statements with complex
always-false conditions
## New type solver
- Dataflow analysis now generates phi nodes on exit of branches.
- Dataflow analysis avoids producing a new definition for locals or
properties that are not owned by that loop.
- If a function parameter has been constrained to `never`, report errors
at all uses of that parameter within that function.
- Switch to using the new `Luau::Set` to replace `std::unordered_set` to
alleviate some poor allocation characteristics which was negatively
affecting overall performance.
- Subtyping can now report many failing reasons instead of just the
first one that we happened to find during the test.
- Subtyping now also report reasons for type pack mismatches.
- When visiting `if` statements or expressions, the resulting context
are the common terms in both branches.
## Native codegen
- Implement support for `buffer` builtins to its IR for x64 and A64.
- Optimized `table.insert` by not inserting a table barrier if it is
fastcalled with a constant.
## Internal Contributors
Co-authored-by: Aaron Weiss <aaronweiss@roblox.com>
Co-authored-by: Alexander McCord <amccord@roblox.com>
Co-authored-by: Andy Friesen <afriesen@roblox.com>
Co-authored-by: Arseny Kapoulkine <arseny@roblox.com>
Co-authored-by: Aviral Goel <agoel@roblox.com>
Co-authored-by: Lily Brown <lbrown@roblox.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
# What's changed?
* Fixed a bug in type cloning by maintaining persistent types.
* We now parse imprecise integer literals to report the imprecision as a
warning to developers.
* Add a compiler flag to specify the name of the statistics output file.
### New type solver
* Renamed `ConstraintGraphBuilder` to `ConstraintGenerator`
* LValues now take into account the type being assigned during
constraint generation.
* Normalization performance has been improved by 33% by replacing the an
internal usage of `std::unordered_set` with `DenseHashMap`.
* Normalization now has a helper to identify types that are equivalent
to `unknown`, which is being used to fix some bugs in subtyping.
* Uses of the old unifier in the new type solver have been eliminated.
* Improved error explanations for subtyping errors in `TypeChecker2`.
### Native code generation
* Expanded some of the statistics recorded during compilation to include
the number of instructions and blocks.
* Introduce instruction and block count limiters for controlling what
bytecode is translated into native code.
* Implement code generation for byteswap instruction.
### Internal Contributors
Co-authored-by: Aaron Weiss <aaronweiss@roblox.com>
Co-authored-by: Alexander McCord <amccord@roblox.com>
Co-authored-by: Andy Friesen <afriesen@roblox.com>
Co-authored-by: Aviral Goel <agoel@roblox.com>
Co-authored-by: Lily Brown <lbrown@roblox.com>
For table type aliases, records the location of the property in the
alias declaration.
This is an alternate solution to the particular case noted in #802.
Instead of tracking the type alias definition for FTVs, it tracks it for
the encompassing property instead.
Note that we still don't have positions in the following case:
```
type Func = () -> ()
```
Closes#802 (Although not completely...)
## What's changed
- `bit32.byteswap` added
([RFC](4f543ec23b/docs/function-bit32-byteswap.md))
- Buffer library implementation
([RFC](4f543ec23b/docs/type-byte-buffer.md))
- Fixed a missing `stdint.h` include
- Fixed parser limiter for recursive type annotations being kind of
weird (fixes#645)
### Native Codegen
- Fixed a pair of issues when lowering `bit32.extract`
- Fixed a narrow edge case that could result in an infinite loop without
an interruption
- Fixed a negative array out-of-bounds access issue
- Temporarily reverted linear block predecessor value propagation
### New type solver
- We now type check assignments to annotated variables
- Fixed some test cases under local type inference
- Moved `isPending` checks for type families to improve performance
- Optimized our process for testing if a free type is sufficiently
solved
- Removed "none ptr" from lea instruction disassembly logging
### Build system & tooling
- CMake configuration now validates dependencies to maintain separation
between components
- Improvements to the fuzzer coverage
- Deduplicator for fuzzed callstacks
---------
Co-authored-by: Arseny Kapoulkine <arseny.kapoulkine@gmail.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
Co-authored-by: Andy Friesen <afriesen@roblox.com>
Co-authored-by: Aaron Weiss <aaronweiss@roblox.com>
Co-authored-by: Vighnesh Vijay <vvijay@roblox.com>
