This change folds:
a * 1 => a
a / 1 => a
a * -1 => -a
a / -1 => -a
a * 2 => a + a
a / 2^k => a * 2^-k
a - 0 => a
a + (-0) => a
Note that the following folds are all invalid:
a + 0 => a (breaks for negative zero)
a - (-0) => a (breaks for negative zero)
a - a => 0 (breaks for Inf/NaN)
0 - a => -a (breaks for negative zero)
Various cases of UNM_NUM could be optimized (eg (-a) * (-b) = a * b),
but that doesn't happen in benchmarks.
While it would be possible to also fold inverse multiplications (k * v),
these do not happen in benchmarks and rarely happen in bytecode due
to type based optimizations. Maybe this can be improved with some sort
of
IR canonicalization in the future if necessary.
I've considered moving some of these, like division strength reduction,
to IR translation (as this is where POW is lowered presently) but it
didn't
seem better one way or the other.
This change improves performance on some benchmarks, e.g. trig and
voxelgen,
and should be a strict uplift as it never generates more instructions or
longer
latency chains. On Apple M2, without division->multiplication
optimization, both
benchmarks see 0.1-0.2% uplift. Division optimization makes trig 3%
faster; I expect
the gains on X64 will be more muted, but on Apple this seems to allow
loop iterations
to overlap better by removing the division bottleneck.
These were written before compiler optimizations and array invariant. It
is now impossible for t[1] to be stored in the hash part, as this would
violate the array invariant that says that elements 1..#t are stored in
the array.
For ipairs, it doesn't traverse the hash part anymore now, so we adjust
the code to make sure no elements outside of the 1..#t slice are
covered. For table.find, we can use find-with-offset to still access the
hash part.
Fixes#1283.
* Fixed `Frontend::markDirty` not working on modules that were not
typechecked yet
* Fixed generic variadic function unification succeeding when it should
have reported an error
New Type Solver:
* Implemented semantic subtyping check for function types
Native Code Generation:
* Improved performance of numerical loops with a constant step
* Simplified IR for `bit32.extract` calls extracting first/last bits
* Improved performance of NaN checks
- 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
Type checker/autocomplete:
* `Luau::autocomplete` no longer performs typechecking internally, make
sure to run `Frontend::check` before performing autocomplete requests
* Autocomplete string suggestions without "" are now only suggested
inside the ""
* Autocomplete suggestions now include `function (anonymous autofilled)`
key with a full suggestion for the function expression (with arguments
included) stored in `AutocompleteEntry::insertText`
* `AutocompleteEntry::indexedWithSelf` is provided for function call
suggestions made with `:`
* Cyclic modules now see each other type exports as `any` to prevent
memory use-after-free (similar to module return type)
Runtime:
* Updated inline/loop unroll cost model to better handle assignments
(Fixes https://github.com/Roblox/luau/issues/978)
* `math.noise` speed was improved by ~30%
* `table.concat` speed was improved by ~5-7%
* `tonumber` and `tostring` now have fastcall paths that execute ~1.5x
and ~2.5x faster respectively (fixes#777)
* Fixed crash in `luaL_typename` when index refers to a non-existing
value
* Fixed potential out of memory scenario when using `string.sub` or
`string.char` in a loop
* Fixed behavior of some fastcall builtins when called without arguments
under -O2 to match original functions
* Support for native code execution in VM is now enabled by default
(note: native code still has to be generated explicitly)
* `Codegen::compile` now accepts `CodeGen_OnlyNativeModules` flag. When
set, only modules that have a `--!native` hot-comment at the top will be
compiled to native code
In our new typechecker:
* Generic type packs are no longer considered to be variadic during
unification
* Timeout and cancellation now works in new solver
* Fixed false positive errors around 'table' and 'function' type
refinements
* Table literals now use covariant unification rules. This is sound
since literal has no type specified and has no aliases
* Fixed issues with blocked types escaping the constraint solver
* Fixed more places where error messages that should've been suppressed
were still reported
* Fixed errors when iterating over a top table type
In our native code generation (jit):
* 'DebugLuauAbortingChecks' flag is now supported on A64
* LOP_NEWCLOSURE has been translated to IR
* Fix#817
* Fix#850
* Optimize math.floor/ceil/round with SSE4.1
* Results in a ~7-9% speedup on the math-cordic benchmark.
* Optimized table.sort.
* table.sort is now ~4.1x faster (when not using a predicate) and ~2.1x
faster when using a simple predicate. Performance may improve further in
the future.
* Reorganize the memory ownership of builtin type definitions.
* This is a small initial step toward affording parallel typechecking.
The new type solver is coming along nicely. We are working on fixing
crashes and bugs.
A few major changes to native codegen landed this week:
* Fixed lowering of Luau IR mod instruction when first argument is a
constant
* Added VM register data-flow/capture analysis
* Fixed issues with optimizations in unreachable blocks
---------
Co-authored-by: Arseny Kapoulkine <arseny.kapoulkine@gmail.com>
Co-authored-by: Vyacheslav Egorov <vegorov@roblox.com>
* For autocomplete, additional information is included in Scope for type
alias name locations and names of imported modules
* Improved autocomplete suggestions in 'for' and 'while' loop headers
* String match functions return types are now optional strings and
numbers because match is not guaranteed at runtime
* Fixed build issue on gcc 11 and up (Fixes
https://github.com/Roblox/luau/issues/806)
* Support `["prop"]` syntax on class definitions in definition files.
(#704)
* Improve type checking performance for complex overloaded functions
* Fix rare cases of incorrect stack traces for out of memory errors at
runtime
* Reoptimized math.min/max/bit32 builtins assuming at least 2 arguments are used (1-2% lift on some benchmarks)
* Type errors that mention function types no longer have redundant parenthesis around return type
* Luau REPL now supports --compile=remarks which displays the source code with optimization remarks embedded as comments
* Builtin calls are slightly faster when called with 1-2 arguments (~1% improvement in some benchmarks)
- Fix rare type checking bugs with invalid generic types escaping the
module scope
- Fix type checking of variadic type packs in certain cases
- Implement type normalization, which resolves a large set of various
issues with unions/intersections in type checker
- Improve parse errors for trailing commas in function calls and type
lists
- Reduce profiling skew when using --profile with very high frequencies
- Improve performance of `lua_getinfo` (`debug.info`, `debug.traceback`
and profiling overhead are now 20% faster/smaller)
- Improve performance of polymorphic comparisons (1-2% lift on some
benchmarks)
- Improve performance of closure creation (1-2% lift on some benchmarks)
- Improve string comparison performance (4% lift on string sorting)
- Improve ComparisonPrecedence lint suggestions for three-way comparisons (X < Y < Z)
- Improve type checking stability
- Improve location information for errors when parsing invalid type annotations
- Compiler now generates bytecode version 3 in all configurations
- Improve performance of comparisons against numeric constants on AArch64
- Fix some cases where type checking would overflow the native stack
- Improve autocomplete behavior when assigning a partially written function call (not currently exposed through command line tools)
- Improve autocomplete type inference feedback for some expressions where previously the type would not be known
- Improve quantification performance during type checking for large types
- Improve type checking for table literals when the expected type of the table is known because of a type annotation
- Fix type checking errors in cases where required module has errors in the resulting type
- Fix debug line information for multi-line chained call sequences (Add function name information for "attempt to call a nil value" #255)
- lua_newuserdata now takes 2 arguments to match Lua/LuaJIT APIs better; lua_newuserdatatagged should be used if the third argument was non-0.
- lua_ref can no longer be used with LUA_REGISTRYINDEX to prevent mistakes when migrating Lua FFI (Inconsistency with lua_ref #247)
- Fix assertions and possible crashes when executing script code indirectly via metatable dispatch from lua_equal/lua_lessthan/lua_getfield/etc. (Hitting a crash in an assert after lua_equal is called. #259)
- Fix flamegraph scripts to run under Python 2
