# What's changed?
* Support for new 'require by string' RFC with relative paths and
aliases in now enabled in Luau REPL application
### New Type Solver
* Fixed assertion failure on generic table keys (`[expr] = value`)
* Fixed an issue with type substitution traversing into the substituted
parts during type instantiation
* Fixed crash in union simplification when that union contained
uninhabited unions and other types inside
* Union types in binary type families like `add<a | b, c>` are expanded
into `add<a, c> | add<b, c>` to handle
* Added handling for type family solving creating new type families
* Fixed a bug with normalization operation caching types with unsolved
parts
* Tables with uninhabited properties are now simplified to `never`
* Fixed failures found by fuzzer
### Native Code Generation
* Added support for shared code generation between multiple Luau VM
instances
* Fixed issue in load-store propagation and new tagged LOAD_TVALUE
instructions
* Fixed issues with partial register dead store elimination causing
failures in GC assists
---
### 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: James McNellis <jmcnellis@roblox.com>
Co-authored-by: Vighnesh Vijay <vvijay@roblox.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
# What's Changed
* Fix a case where the stack wasn't completely cleaned up where
`debug.info` errored when passed `"f"` option and a thread.
* Fix a case of uninitialized field in `luaF_newproto`.
### New Type Solver
* When a local is captured in a function, don't add a new entry to the
`DfgScope::bindings` if the capture occurs within a loop.
* Fix a poor performance characteristic during unification by not trying
to simplify an intersection.
* Fix a case of multiple constraints mutating the same blocked type
causing incorrect inferences.
* Fix a case of assertion failure when overload resolution encounters a
return typepack mismatch.
* When refining a property of the top `table` type, we no longer signal
an unknown property error.
* Fix a misuse of free types when trying to infer the type of a
subscript expression.
* Fix a case of assertion failure when trying to resolve an overload
from `never`.
### Native Code Generation
* Fix dead store optimization issues caused by partial stores.
---
### Internal Contributors
Co-authored-by: Aaron Weiss <aaronweiss@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>
---------
Co-authored-by: Aaron Weiss <aaronweiss@roblox.com>
Co-authored-by: Andy Friesen <afriesen@roblox.com>
Co-authored-by: Vighnesh <vvijay@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>
When the input is a constant, we use a fairly inefficient sequence of
fmov+fcvt+dup or, when the double isn't encodable in fmov,
adr+ldr+fcvt+dup.
Instead, we can use the same lowering as X64 when the input is a
constant, and load the vector from memory. However, if the constant is
encodable via fmov, we can use a vector fmov instead (which is just one
instruction and doesn't need constant space).
Fortunately the bit encoding of fmov for 32-bit floating point numbers
matches that of 64-bit: the decoding algorithm is a little different
because it expands into a larger exponent, but the values are
compatible, so if a double can be encoded into a scalar fmov with a
given abcdefgh pattern, the same pattern should encode the same float;
due to the very limited number of mantissa and exponent bits, all values
that are encodable are also exact in both 32-bit and 64-bit floats.
This strategy is ~same as what gcc uses. For complex vectors, we
previously used 4 instructions and 8 bytes of constant storage, and now
we use 2 instructions and 16 bytes of constant storage, so the memory
footprint is the same; for simple vectors we just need 1 instruction (4
bytes).
clang lowers vector constants a little differently, opting to synthesize
a 64-bit integer using 4 instructions (mov/movk) and then move it to the
vector register - this requires 5 instructions and 20 bytes, vs ours/gcc
2 instructions and 8+16=24 bytes. I tried a simpler version of this that
would be more compact - synthesize a 32-bit integer constant with
mov+movk, and move it to vector register via dup.4s - but this was a
little slower on M2, so for now we prefer the slightly larger version as
it's not a regression vs current implementation.
On the vector approximation benchmark we get:
- Before this PR (flag=false): ~7.85 ns/op
- After this PR (flag=true): ~7.74 ns/op
- After this PR, with 0.125 instead of 0.123 in the benchmark code (to
use fmov): ~7.52 ns/op
- Not part of this PR, but the mov/dup strategy described above: ~8.00
ns/op
This change replaces scalar versions of vector opcodes for A64 with
actual vector instructions.
We take the approach similar to X64: patch last component with zero,
perform the math, patch last component with type tag. I'm hoping that in
the future the type tag will be placed separately (separate IR opcode?)
because right now chains of math operations result in excessive type tag
operations.
To patch the type tag without always keeping a mask in a register,
ins.4s instructions can be used; unfortunately it's only capable of
patching a register in-place, so we need an extra register copy in case
it's not last-use. Usually it's last-use so the patch is free; probably
with IR rework mentioned above all of this can be improved (e.g.
load-with-patch will never need to copy).
~It's not 100% clear if we *have* to patch type tag: Apple does preserve
denormals but we'd need to benchmark this to see if there's an actual
performance impact. But for now we're playing it safe.~
This was tested by running the conformance tests, and new opcode
implementations were checked by comparing the result with
https://armconverter.com/.
Performance testing is complicated by the fact that OSS Luau doesn't
support vector constructor out of the box, and other limitations of
codegen. I've hacked vector constructor/type into REPL and confirmed
that on a test that calls this function in a loop (not inlined):
```
function fma(a: vector, b: vector, c: vector)
return a * b + c
end
```
... this PR improves performance by ~6% (note that probably most of the
overhead here is the call dispatch; I didn't want to brave testing a
more complex expression). The assembly for an individual operation
changes as follows:
Before:
```
# %14 = MUL_VEC %12, %13 ; useCount: 2, lastUse: %22
dup s29,v31.s[0]
dup s28,v30.s[0]
fmul s29,s29,s28
ins v31.s[0],v29.s[0]
dup s29,v31.s[1]
dup s28,v30.s[1]
fmul s29,s29,s28
ins v31.s[1],v29.s[0]
dup s29,v31.s[2]
dup s28,v30.s[2]
fmul s29,s29,s28
ins v31.s[2],v29.s[0]
```
After:
```
# %14 = MUL_VEC %12, %13 ; useCount: 2, lastUse: %22
ins v31.s[3],w31
ins v30.s[3],w31
fmul v31.4s,v31.4s,v30.4s
movz w17,#4
ins v31.s[3],w17
```
**edit** final form (see comments):
```
# %14 = MUL_VEC %12, %13 ; useCount: 2, lastUse: %22
fmul v31.4s,v31.4s,v30.4s
movz w17,#4
ins v31.s[3],w17
```
# What's changed?
* Check interrupt handler inside the pattern match engine to eliminate
potential for programs to hang during string library function execution.
* Allow iteration over table properties to pass the old type solver.
### Native Code Generation
* Use in-place memory operands for math library operations on x64.
* Replace opaque bools with separate enum classes in IrDump to improve
code maintainability.
* Translate operations on inferred vectors to IR.
* Enable support for debugging native-compiled functions in Roblox
Studio.
### New Type Solver
* Rework type inference for boolean and string literals to introduce
bounded free types (bounded below by the singleton type, and above by
the primitive type) and reworked primitive type constraint to decide
which is the appropriate type for the literal.
* Introduce `FunctionCheckConstraint` to handle bidirectional
typechecking for function calls, pushing the expected parameter types
from the function onto the arguments.
* Introduce `union` and `intersect` type families to compute deferred
simplified unions and intersections to be employed by the constraint
generation logic in the new solver.
* Implement support for expanding the domain of local types in
`Unifier2`.
* Rework type inference for iteration variables bound by for in loops to
use local types.
* Change constraint blocking logic to use a set to prevent accidental
re-blocking.
* Add logic to detect missing return statements in functions.
### 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: Vyacheslav Egorov <vegorov@roblox.com>
---------
Co-authored-by: Alexander McCord <amccord@roblox.com>
Co-authored-by: Andy Friesen <afriesen@roblox.com>
Co-authored-by: Vighnesh <vvijay@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>
### What's Changed
- Improve readability of unions and intersections by limiting the number
of elements of those types that can be presented on a single line (gated
under `FFlag::LuauToStringSimpleCompositeTypesSingleLine`)
- Adds a new option to the compiler `--record-stats` to record and
output compilation statistics
- `if...then...else` expressions are now optimized into `AND/OR` form
when possible.
### VM
- Add a new `buffer` type to Luau based on the [buffer
RFC](https://github.com/Roblox/luau/pull/739) and additional C API
functions to work with it; this release does not include the library.
- Internal C API to work with string buffers has been updated to align
with Lua version more closely
### Native Codegen
- Added support for new X64 instruction (rev) and new A64 instruction
(bswap) in the assembler
- Simplified the way numerical loop condition is translated to IR
### New Type Solver
- Operator inference now handled by type families
- Created a new system called `Type Paths` to explain why subtyping
tests fail in order to improve the quality of error messages.
- Systematic changes to implement Data Flow analysis in the new solver
(`Breadcrumb` removed and replaced with `RefinementKey`)
---
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: Vighnesh Vijay <vvijay@roblox.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
---------
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: Lily Brown <lbrown@roblox.com>
Co-authored-by: Aaron Weiss <aaronweiss@roblox.com>
Co-authored-by: Alexander McCord <amccord@roblox.com>
Co-authored-by: Aviral Goel <agoel@roblox.com>
- Updated Roblox copyright to 2023
- Floor division operator `//` (implements #832)
- Autocomplete now shows `end` within `do` blocks
- Restore BraceType when using `Lexer::lookahead` (fixes#1019)
# New typechecker
- Subtyping tests between metatables and tables
- Subtyping tests between string singletons and tables
- Subtyping tests for class types
# Native codegen
- Fixed macOS test failure (wrong spill restore offset)
- Fixed clobbering of non-volatile xmm registers on Windows
- Fixed wrong storage location of SSA reg spills
- Implemented A64 support for add/sub extended
- Eliminated zextReg from A64 lowering
- Remove identical table slot lookups
- Propagate values from predecessor into the linear block
- Disabled reuse slot optimization
- Keep `LuaNode::val` check for nil when optimizing `CHECK_SLOT_MATCH`
- Implemented IR translation of `table.insert` builtin
- Fixed mmap error handling on macOS/Linux
# Tooling
- Used `|` as a column separator instead of `+` in `bench.py`
- Added a `table.sort` micro-benchmark
- Switched `libprotobuf-mutator` to a less problematic version
* Added luau-compile executable target to build/test compilation without
having full REPL included.
In our new typechecker:
* Fixed the order in which constraints are checked to get more
deterministic errors in different environments
* Fixed `isNumber`/`isString` checks to fix false positive errors in
binary comparisons
* CannotCallNonFunction error is reported when calling an intersection
type of non-functions
In our native code generation (jit):
* Outlined X64 return instruction code to improve code size
* Improved performance of return instruction on A64
* Added construction of a dominator tree for future optimizations
* Fixed gcc warning about uninitialized `std::optional`
* Fixed inlining of functions when they are used to compute their own
arguments
In the new type solver:
* Type families that are not part of a function signature cannot be
resolved at instantiation time and will now produce an error. This will
be relaxed in the future when we get constraint clauses on function
signatures (internally)
* `never` type is now comparable
* Improved typechecking of `for..in` statements
* Fixed checks for number type in `Add` type family
* Performance was improved, with particularly large gains on large
projects
And in native code generation (jit):
* We eliminated the call instruction overhead when native code support
is enabled in the VM
* Small optimizations to arm64 lowering
* Reworked LOP_GETIMPORT handling to reduce assembly code size
* Fixed non-deterministic binary output
* Fixed bad code generation caused by incorrect SSA to VM register links
invalidation
---------
Co-authored-by: Arseny Kapoulkine <arseny.kapoulkine@gmail.com>
Co-authored-by: Andy Friesen <afriesen@roblox.com>
Lots of things going on this week:
* Fix a crash that could occur in the presence of a cyclic union. We
shouldn't be creating cyclic unions, but we shouldn't be crashing when
they arise either.
* Minor cleanup of `luau_precall`
* Internal change to make L->top handling slightly more uniform
* Optimize SETGLOBAL & GETGLOBAL fallback C functions.
* https://github.com/Roblox/luau/pull/929
* The syntax to the `luau-reduce` commandline tool has changed. It now
accepts a script, a command to execute, and an error to search for. It
no longer automatically passes the script to the command which makes it
a lot more flexible. Also be warned that it edits the script it is
passed **in place**. Do not point it at something that is not in source
control!
New solver
* Switch to a greedier but more fallible algorithm for simplifying union
and intersection types that are created as part of refinement
calculation. This has much better and more predictable performance.
* Fix a constraint cycle in recursive function calls.
* Much improved inference of binary addition. Functions like `function
add(x, y) return x + y end` can now be inferred without annotations. We
also accurately typecheck calls to functions like this.
* Many small bugfixes surrounding things like table indexers
* Add support for indexers on class types. This was previously added to
the old solver; we now add it to the new one for feature parity.
JIT
* https://github.com/Roblox/luau/pull/931
* Fuse key.value and key.tt loads for CEHCK_SLOT_MATCH in A64
* Implement remaining aliases of BFM for A64
* Implement new callinfo flag for A64
* Add instruction simplification for int->num->int conversion chains
* Don't even load execdata for X64 calls
* Treat opcode fallbacks the same as manually written fallbacks
---------
Co-authored-by: Arseny Kapoulkine <arseny.kapoulkine@gmail.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
* `ClassType` can now have an indexer defined on it. This allows custom
types to be used in `t[x]` expressions.
* Fixed search for closest executable breakpoint line. Previously,
breakpoints might have been skipped in `else` blocks at the end of a
function
* Fixed how unification is performed for two optional types `a? <: b?`,
previously it might have unified either 'a' or 'b' with 'nil'. Note that
this fix is not enabled by default yet (see the list in
`ExperimentalFlags.h`)
In the new type solver, a concept of 'Type Families' has been
introduced.
Type families can be thought of as type aliases with custom type
inference/reduction logic included with them.
For example, we can have an `Add<T, U>` type family that will resolve
the type that is the result of adding two values together.
This will help type inference to figure out what 'T' and 'U' might be
when explicit type annotations are not provided.
In this update we don't define any type families, but they will be added
in the near future.
It is also possible for Luau embedders to define their own type families
in the global/environment scope.
Other changes include:
* Fixed scope used to find out which generic types should be included in
the function generic type list
* Fixed a crash after cyclic bound types were created during unification
And in native code generation (jit):
* Use of arm64 target on M1 now requires macOS 13
* Entry into native code has been optimized. This is especially
important for coroutine call/pcall performance as they involve going
through a C call frame
* LOP_LOADK(X) translation into IR has been improved to enable type
tag/constant propagation
* arm64 can use integer immediate values to synthesize floating-point
values
* x64 assembler removes duplicate 64bit numbers from the data section to
save space
* Linux `perf` can now be used to profile native Luau code (when running
with --codegen-perf CLI argument)
* Added a limit on how many instructions the Compiler can safely produce
(reported by @TheGreatSageEqualToHeaven)
C++ API Changes:
* With work started on read-only and write-only properties,
`Property::type` member variable has been replaced with `TypeId type()`
and `setType(TypeId)` functions.
* New `LazyType` unwrap callback now has a `void` return type, all
that's required from the callback is to write into `unwrapped` field.
In our work on the new type solver, the following issues were fixed:
* Work has started to support https://github.com/Roblox/luau/pull/77 and
https://github.com/Roblox/luau/pull/79
* Refinements are no longer applied on l-values, removing some
false-positive errors
* Improved overload resolution against expected result type
* `Frontend::prepareModuleScope` now works in the new solver
* Cofinite strings are now comparable
And these are the changes in native code generation (JIT):
* Fixed MIN_NUM and MAX_NUM constant fold when one of the arguments is
NaN
* Added constant folding for number conversions and bit operations
* Value spilling and rematerialization is now supported on arm64
* Improved FASTCALL2K IR generation to support second argument constant
* Added value numbering and load/store propagation optimizations
* Added STORE_VECTOR on arm64, completing the IR lowering on this target
* Work toward affording parallel type checking
* The interface to `LazyType` has changed:
* `LazyType` now takes a second callback that is passed the `LazyType&`
itself. This new callback is responsible for populating the field
`TypeId LazyType::unwrapped`. Multithreaded implementations should
acquire a lock in this callback.
* Modules now retain their `humanReadableNames`. This reduces the number
of cases where type checking has to call back to a `ModuleResolver`.
* https://github.com/Roblox/luau/pull/902
* Add timing info to the Luau REPL compilation output
We've also fixed some bugs and crashes in the new solver as we march
toward readiness.
* Thread ICEs (Internal Compiler Errors) back to the Frontend properly
* Refinements are no longer applied to lvalues
* More miscellaneous stability improvements
Lots of activity in the new JIT engine:
* Implement register spilling/restore for A64
* Correct Luau IR value restore location tracking
* Fixed use-after-free in x86 register allocator spill restore
* Use btz for bit tests
* Finish branch assembly support for A64
* Codesize and performance improvements for A64
* The bit32 library has been implemented for arm and x64
---------
Co-authored-by: Arseny Kapoulkine <arseny.kapoulkine@gmail.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
* Fixed exported types not being suggested in autocomplete
* `T...` is now convertible to `...any` (Fixes
https://github.com/Roblox/luau/issues/767)
* Fixed issue with `T?` not being convertible to `T | T` or `T?`
(sometimes when internal pointer identity is different)
* Fixed potential crash in missing table key error suggestion to use a
similar existing key
* `lua_topointer` now returns a pointer for strings
C++ API Changes:
* `prepareModuleScope` callback has moved from TypeChecker to Frontend
* For LSPs, AstQuery functions (and `isWithinComment`) can be used
without full Frontend data
A lot of changes in our two experimental components as well.
In our work on the new type-solver, the following issues were fixed:
* Fixed table union and intersection indexing
* Correct custom type environments are now used
* Fixed issue with values of `free & number` type not accepted in
numeric operations
And these are the changes in native code generation (JIT):
* arm64 lowering is almost complete with support for 99% of IR commands
and all fastcalls
* Fixed x64 assembly encoding for extended byte registers
* More external x64 calls are aware of register allocator
* `math.min`/`math.max` with more than 2 arguments are now lowered to IR
as well
* Fixed correctness issues with `math` library calls with multiple
results in variadic context and with x64 register conflicts
* x64 register allocator learnt to restore values from VM memory instead
of always using stack spills
* x64 exception unwind information now supports multiple functions and
fixes function start offset in Dwarf2 info
Once again, all of our changes this week are for new type solver and the
JIT.
In the new type solver, we fixed cyclic type alias handling and multiple
stability issues.
In the JIT, our main progress was for arm64, where, after lowering 36%
of instructions, we start seeing first Luau functions executing
natively.
For x64, we performed code cleanup and refactoring to allow for future
optimizations.
All of our changes this week have been focused on the new type solver
and the JIT.
As we march toward feature parity with the old solver, we've tightened
up a bunch of lingering issues with overload resolution, unsealed
tables, and type normalization. We've also fixed a bunch of crashes and
assertion failures in the new solver.
On the JIT front, we've started work on an A64 backend, improved the IR
analysis in a bunch of cases, and implemented assembly generation for
the builtin functions `type()` and `typeof()`.
---------
Co-authored-by: Arseny Kapoulkine <arseny.kapoulkine@gmail.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
* Fixed incorrect lexeme generated for string parts in the middle of an
interpolated string (Fixes https://github.com/Roblox/luau/issues/744)
* DeprecatedApi lint can report some issues without type inference
information
* Fixed performance of autocomplete requests when suggestions have large
intersection types (Solves
https://github.com/Roblox/luau/discussions/847)
* Marked `table.getn`/`foreach`/`foreachi` as deprecated ([RFC:
Deprecate
table.getn/foreach/foreachi](https://github.com/Roblox/luau/blob/master/rfcs/deprecate-table-getn-foreach.md))
* With -O2 optimization level, we now optimize builtin calls based on
known argument/return count.
Note that this change can be observable if `getfenv/setfenv` is used to
substitute a builtin, especially if arity is different.
Fastcall heavy tests show a 1-2% improvement.
* Luau can now be built with clang-cl (Fixes
https://github.com/Roblox/luau/issues/736)
We also made many improvements to our experimental components.
For our new type solver:
* Overhauled data flow analysis system, fixed issues with 'repeat'
loops, global variables and type annotations
* Type refinements now work on generic table indexing with a string
literal
* Type refinements will properly track potentially 'nil' values (like
t[x] for a missing key) and their further refinements
* Internal top table type is now isomorphic to `{}` which fixes issues
when `typeof(v) == 'table'` type refinement is handled
* References to non-existent types in type annotations no longer resolve
to 'error' type like in old solver
* Improved handling of class unions in property access expressions
* Fixed default type packs
* Unsealed tables can now have metatables
* Restored expected types for function arguments
And for native code generation:
* Added min and max IR instructions mapping to vminsd/vmaxsd on x64
* We now speculatively extract direct execution fast-paths based on
expected types of expressions which provides better optimization
opportunities inside a single basic block
* Translated existing math fastcalls to IR form to improve tag guard
removal and constant propagation
* Type inference of `a and b` and `a or b` has been improved (Fixes
https://github.com/Roblox/luau/issues/730)
* Improved error message when `for ... in x` loop iterates over a value
that could be 'nil'
* Return type of `next` not includes 'nil' (Fixes
https://github.com/Roblox/luau/issues/706)
* Improved type inference of `string` type
* Luau library table type names are now reported as `typeof(string)`/etc
instead of just `string` which was misleading
* Added parsing error when optional type alias type parameter wasn't
provided after `=` token
* Improved tagged union type refinement in conditional expressions, type
in `else` branch should no longer include previously handled union
options
* Reduce the stack utilization of type checking.
* Improve the error message that's reported when a delimiting comma is
missing from a table literal. eg
```lua
local t = {
first = 1
second = 2
}```
