Commit Graph
4195 Commits
Author SHA1 Message Date
Chris Tate dcd86a17da Merge main 2026-10-22 20:52:05 -05:00
Chris Tate 8c67f04062 Throw Node's TypeError for class constructors called through call/apply (#1892) 2026-10-22 20:36:01 -05:00
Chris Tate 5966f23bd3 Keep the class-value call fence when a subclass declares a static call
`C.call(...)` through a class value typed as `Base` may run a subclass's own
static `call`, which Node calls instead of the class constructor. The
constructor-call lowering now applies only when no class the receiver can
hold declares a static `call` or `apply`: the class and its bases (any
static member of that name, including accessors, fields without
initializers and computed names), and, for a receiver that may hold a
subclass, every subclass. Otherwise the call keeps its existing
diagnostic.
2026-10-22 20:33:33 -05:00
Chris Tate 210dee9eb0 Read the global object without rebuilding its name table (#1895) 2026-10-22 20:26:51 -05:00
Chris Tate a9ee38611e Register the global object's names again after a library re-init 2026-10-22 20:20:56 -05:00
Chris Tate 8807c24526 Compile zod 3 natively (#1835) 2026-10-22 20:13:04 -05:00
Chris Tate 0c6a35996d Fix the native compiler build on main (#1913) 2026-10-22 20:10:49 -05:00
Malte UblandClaude Opus 5.5 122c98cf4f Read the global object without rebuilding its name table
Every evaluation of `globalThis` as a value built an array literal of
every name the global type declares (hundreds of strings) and registered
each one in the known-name table again. A server that reads a singleton
through `globalThis` per request spent about a tenth of its time there.

Emit the name table as one immortal static array per program (the
tagged-template site array, keyed by its contents), and return the
global object at once when the runtime sees a table it has already
registered.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 18:10:05 -07:00
Chris Tate ae5c7d7780 Map interfaces that extend an any-typed base (zod z.infer) to checked-dynamic values (#1912) 2026-10-22 20:08:05 -05:00
Chris Tate 33b22c5bac Fix the native compiler build on main 2026-10-22 20:07:43 -05:00
Chris Tate 1660531c7b Support FinalizationRegistry with bookkeeping and no cleanup callbacks (#1854) 2026-10-22 20:03:31 -05:00
Chris Tate 62a0426193 Keep Linux poller timers in a heap behind one timerfd (#1896) 2026-10-22 20:02:08 -05:00
Chris Tate 636ffa21be Merge main 2026-10-22 19:59:51 -05:00
Malte UblandClaude Opus 5.5 d0334f93f5 Record the preflight/order baseline for the zod3-infer npm case
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 17:57:59 -07:00
Chris Tate b90a01fb80 Update the diagnostics fixtures now that Iterable values compile (#1910) 2026-10-22 19:57:19 -05:00
Chris Tate 56149fde04 Count each epoll test deadline from its own arm call
The poller test recorded every initial deadline as the same start time
plus the delay, while the poller reads the clock at each arm call; when
arming the 2,000 timers took over 1 ms, the recorded order could differ
from the delivered one by more than the assertion's slack. Each deadline
now counts from the time read just before its scrp_timer_arm call, as
the re-arm loop's do, and a 3 ms pause halfway through arming keeps the
test exercising that case.
2026-10-22 19:56:53 -05:00
Chris Tate fbaa71a66f Run the fromEntries aliasing program and drop compiled Iterable entries from the iterator fences
Since Iterable<T> values compile (#1875), two diagnostics fixtures changed:

- fromentries-aliased-entries.ts compiles and prints what Node prints:
  the entries' one array stays shared by both fields. It moves to the run
  corpus as 7440-fromentries-aliased-entries.ts.
- iterator-slot-fences.ts no longer reports its Iterable<number>
  parameter (SC2020) and the call through it (SC2004). That part leaves
  the fixture, which keeps the class-iterator fence; corpus 7380-7383
  cover Iterable values.
2026-10-22 19:54:17 -05:00
Chris Tate b19a589c12 Keep the global object behind a typeof globalThis intersection cast (#1894) 2026-10-22 19:53:35 -05:00
Chris Tate b7d355ed13 Stringify live typed records and arrays without a key list (#1897) 2026-10-22 19:52:19 -05:00
Chris Tate 7a96b0de57 Spread untyped sources into typed object literals (#1893) 2026-10-22 19:48:13 -05:00
Chris Tate 3cace3e785 Throw Node's TypeError for class constructors called through call/apply
`C.call(thisArg, ...)` and `C.apply(thisArg, args)` on a program class that
declares no static `call` or `apply` now lower as a call through the boxed
class value. It evaluates the arguments in order, applies apply's
argument-list checks, and throws `TypeError: Class constructor C cannot be
invoked without 'new'` with the runtime class name, as Node does. Before,
TypeScript reported SC2020 at compile time and JavaScript threw the
standard-library fence's Error at run time. This covers ES5 constructors that
call a class as their base (`function Sub() { Base.call(this); }`).

Runtime-provided classes (Error, EventEmitter, streams, EventTarget), classes
with runtime statics, spread arguments, and receivers other than a name,
`this` or a namespace member keep their existing lowering.
2026-10-22 19:36:04 -05:00
Chris Tate 2442872d62 Version charAt loops and compare characters as code units (#1889) 2026-10-22 19:28:24 -05:00
Chris Tate ba8889104a Free arrays whose elements need no release without a teardown (#1890) 2026-10-22 19:27:12 -05:00
Chris Tate c38792f5f3 Keep the unreached JavaScript coverage fixture failing to compile (#1891) 2026-10-22 19:25:07 -05:00
Chris Tate 36209f6d17 Keep the unreached JavaScript coverage fixture failing to compile
The fixture's unreached statement pushed values onto `{ items: [] }`, which
mapped to a `number[]` and could not hold them. Since #1881 a JavaScript
`[]` literal field is a checked-dynamic array, so the push compiles and the
test "a failed unreached JavaScript statement is always explained" found no
failed statement. The fixture now starts `items` as `[0]`, a `number[]`, so
the push of a boolean onto it still does not compile and reports SC1090 at
the same line.
2026-10-22 19:24:07 -05:00
Chris Tate d4a7b82dd8 Free arrays whose elements need no release without a teardown
When the last reference to an array dropped, the runtime ran the generic
teardown: scr_rc_destroy's depth bookkeeping, an indirect call to the array
destroy, and a loop over the element slots. Arrays of numbers, booleans and
compact ProgramLifetime references release nothing, and neither do arrays of
other ProgramLifetime instances, whose release does nothing.

- scr_arr_release frees an acyclic array without an ext directly when its
  elements are numbers, booleans or compact references, or its descriptor
  has SCR_ARR_OPS_RELEASE_NOTHING. Every other array keeps the full
  teardown.
- The backend sets SCR_ARR_OPS_RELEASE_NOTHING in the descriptor's flags
  word on 64-bit targets for ProgramLifetime element types that are not
  compact, plain or as the pointer arm of a nullable union, in programs
  without lifetime regions. Registration copies it into a table indexed by
  the descriptor's registry id before the id is published. Descriptors
  without the flag read 0 and keep the full teardown.

Corpus 7401 drops arrays of ProgramLifetime instances, numbers and booleans
through every array operation, with holes and far indices, next to arrays of
ordinary objects, and matches Node.
2026-10-22 19:22:55 -05:00
Chris Tate 93d2797bac Version charAt loops and compare characters as code units
`for (let i = 0; i < s.length; i++) if (s[i] === " ") n++` and the same
loop over `s.charAt(i)` got no loop version: `s[i]` lowers to charAt,
and the version covered only charCodeAt reads. Each character paid the
inline helper's ASCII and bounds tests and kept its opaque fallback.
`const c = s[i]; if (c >= "a" && c <= "z") n++` also built the
character string and compared strings twice.

- The typed-array loop version covers `s.charAt(index)` like
  `s.charCodeAt(index)`, and a counted loop that reads either at its
  counter takes an integer counter. In the fast copy, a charAt used as a
  string is the immortal one-character string for its byte, and the
  equality and switch rewrites to charCodeAt read the byte.
- A comparison of `s.charAt(i)` with a one-unit literal compares the
  code unit, as the equality does. An index out of range reads "",
  which is below every one-unit literal; its unit read is NaN, which the
  unordered predicates count as below.
- A const string local initialized from a one-argument charAt and used
  only in one-unit equalities, comparisons and switches holds the code
  unit instead (char-unit-locals.ts).
2026-10-22 19:22:20 -05:00
Chris Tate 42d7d9a0c2 Revert "Enumerate each record in the key order its object literal wrote" (#1888) 2026-10-22 19:04:00 -05:00
Chris Tate f5f0af3203 Inline string releases in release builds (#1887) 2026-10-22 19:03:12 -05:00
Chris Tate 3024e57eeb Revert "Enumerate each record in the key order its object literal wrote (#1802)" 2026-10-22 19:02:41 -05:00
Chris Tate bc6a824b55 Inline string releases in release builds
Release builds called scr_str_release for every string release, so a record update by object spread paid a call for each string field it dropped. A string release now uses the emitted fast-path helper, as string retains do; in release builds a function with more than 8 string release sites keeps the runtime call at each, so string-heavy builders and module initializers grow little. Speed builds are unchanged.
2026-10-22 18:59:44 -05:00
Chris Tate 22d136e6ac Read records by keyof keys when their fields differ in type (#1883) 2026-10-22 18:54:16 -05:00
Chris Tate 575bd9485b Copy unaligned AST word tables in one block (#1884) 2026-10-22 18:50:32 -05:00
Chris Tate e6317b3a88 Support construct, writev, decodeStrings and defaultEncoding on streams (#1842) 2026-10-22 18:48:26 -05:00
Chris Tate 86a94cf074 Copy unaligned AST word tables in one block 2026-10-22 18:47:46 -05:00
Chris Tate 3a23e98c7d Merge main 2026-10-22 18:45:27 -05:00
Chris Tate 53c43dd140 Merge main 2026-10-22 18:44:10 -05:00
Malte UblandClaude Opus 5.5 481f9a5e31 Stringify live typed records and arrays without a key list
JSON.stringify of a value reached through `unknown` walks native records
and arrays through their live capsules. The protocol walk built a
key-string array for every record and formatted every array index with
snprintf into a fresh string, only to look the member up again by that
key. A JSON API response of a few records spent most of its
serialization time there.

Without a replacer, walk a live record by field index over a snapshot of
its present fields (its keys are the immortal field names) and a live
array by element index, forming an index string only for a toJSON call
that receives it. Members are still read as scr_json_member reads them,
so callbacks that mutate or materialize the container see the same
results. Index keys elsewhere use the integer formatter.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 16:43:54 -07:00
Malte UblandClaude Opus 5.5 cbafabe610 Keep the global object behind a typeof globalThis intersection cast
`globalThis as typeof globalThis & { key?: T }` (the usual global
singleton spelling) matched the WebAssembly cast lowering, because the
target inherits the global's own WebAssembly member, and produced a fresh
object holding only WebAssembly on every evaluation. Writes through the
view were lost, so a `g[key] ??= new Repo()` singleton was rebuilt on
every call and its state never persisted.

Take the WebAssembly view only when the target's member is not the
global's own, and map an intersection that contains the intrinsic
`typeof globalThis` to the global object (dyn), as `typeof globalThis`
alone already maps.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 16:43:46 -07:00
Chris Tate f132cf5f39 Compile Iterable<T> values (#1875) 2026-10-22 18:30:45 -05:00
Chris Tate e3e37940f1 Iterate iterable unions without an array arm and adapt into optional Iterables
for-of over an iterable union read its array arm without checking that it
had one, so iterating a `Set<number> | Generator<number, void, undefined>`
crashed the compiler; a union without an array arm now runs the generator
loop only. A generator with a non-void return channel adapted into a
required Iterable<T> slot but not into an optional or nullable one, where
the conversion failed with SC2003; the slot conversion now accepts
iterable unions with undefined or null arms beside the iterable arms.
2026-10-22 18:27:44 -05:00
Chris Tate b40ab2b3a1 Fix JavaScript literal field mappings that broke zod 3 and the Effect 4 CLI (#1881) 2026-10-22 18:25:33 -05:00
Chris Tate de28aeca20 Map JavaScript [] literal fields dynamically and keep union-added fields unit-only
Two JavaScript object literal mappings from #1799 broke packages.

A property initialized to `[]` has the checker type `never[]`, and never's
element maps to f64. A literal with such a property builds as a
checked-dynamic object when it lowers on its own. A literal nested in one
that maps to a record (`{ common: { issues: [] }, path: params?.path || [] }`,
zod 3's parse context) lowers against its record type, so the property was a
`number[]`. `ctx.common.issues.push(issue)` then threw "TypeError: expected
number at $, got object", where Node pushes the record. Such a property now
maps to the checked-dynamic array, as a property initialized to `undefined`
or `null` maps to the checked-dynamic value.

That `undefined`/`null` rule also matched the properties the checker adds
when it normalizes a union of literals: `{ _tag, value } | { _tag, error }`
gains an optional `error?: undefined` and `value?: undefined`. Mapping those
to the checked-dynamic value let `{ _tag, value }` fit both arms, and Effect
4's CLI parser (`consumeFlagValue`) failed with SC2003 "union conversion has
no unambiguous layout mapping". The rule now skips optional properties, so
they keep their unit-only type and each literal converts to its own arm.
2026-10-22 18:24:10 -05:00
Chris Tate e0c70c0c01 Enumerate each record in the key order its object literal wrote (#1802) 2026-10-22 18:23:26 -05:00
Chris Tate 81891d55f7 Fix boxed record field hooks and moving key writes, and merge main 2026-10-22 18:19:14 -05:00
Chris Tate 3b4a3c3cdd Write and hash the LLVM module text in slices instead of whole copies (#1878) 2026-10-22 18:16:13 -05:00
Chris Tate 0ebd2c2e4c Compile conditional spreads of several properties and (#1820) 2026-10-22 18:15:22 -05:00
Chris Tate 067be15683 Write and hash the LLVM module text in slices instead of whole copies
The native compiler wrote the module's LLVM text, 150 MB on a large
program such as tsc-ts, by encoding the whole string at once; its frontend
cache entry by first escaping the whole text into a second string and then
encoding that; and it read the written file back whole to key the object
cache. Those copies were the build's peak.

Text parts are now encoded and written in slices of at most 1 MiB UTF-16
units that never split a surrogate pair. The frontend cache entry is
generated while it is written: the LLVM text is escaped slice by slice,
which is the same JSON as escaping it whole, the entry is hashed in the
same pass, and an entry past one native string abandons its write as it
was skipped before. The object cache key hashes the file in 1 MiB blocks.
The files and cache keys are unchanged.
2026-10-22 18:12:38 -05:00
Chris Tate a0c207cc25 Merge main 2026-10-22 18:12:20 -05:00
Chris Tate b955ea3817 Give arrays state bytes only once they hold a hole or undefined (#1827) 2026-10-22 18:09:47 -05:00