Repository navigation
runtime: a JS Number has one encoding; delete the INT32 Number consumers (P3 B4b) - #14658
Conversation
…ers (P3 B4b) No producer of an INT32-tagged Number exists since B4a and codegen stopped reading one in B3. The runtime, stdlib and extension consumers go: equality and compare, ToString / ToNumber, JSON, console formatting, key and index decoding, argument validation. JSValue::is_int32 / as_int32 and is_int32_bits are deleted; the 0x7FFE tag is INTERNAL_WORD_TAG (runtime internal words only). The classifiers that read raw slot words keep the I32 slot word of an at-rest Array or I32 field lane (layout_slot).
…on-object instanceof operand
Walkthrough
Merge Risk:
 |
| Check name | Status | Explanation |
|---|---|---|
| Title check |  |
The title names the main change: the runtime moves to one encoding for JS Numbers and deletes the INT32 Number consumers. The diff matches this. The title is short and specific, though the phrase 'P3 … |
| Description check | |
The description follows the template's main intent. It has a clear summary of what changed and why, a detailed list of changes, a related-work reference, and a test plan with measured results and gate… |
Full details: Title check
-
Explanation
The title names the main change: the runtime moves to one encoding for JS Numbers and deletes the INT32 Number consumers. The diff matches this. The title is short and specific, though the phrase 'P3 B4b' refers to an internal plan stage a reader may not know.
Full details: Description check
-
Explanation
The description follows the template's main intent. It has a clear summary of what changed and why, a detailed list of changes, a related-work reference, and a test plan with measured results and gates. It does not follow the template's section headings exactly: it has no 'Related issue' section (it cites PR
#14372and#14395instead), and the test plan does not use the template checklist. The template's required checklist items about not bumping the version and not editing CHANGELOG.md are not shown as checked. The raw summary also lists a changelog entry, which the author may have added as a template-compliant item. The description is largely complete.
- Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.
Conflicts: typedarray/bigint.rs (main moved the ToBigInt label into bigint_coerce.rs; its INT32 arm is deleted there), check_file_size.sh (counts re-recorded). The sweep's own S0 hunk is dropped: main fixed the canonical-number arm.
What
The last phase of the canonical-Number plan (representation RFC §10 P3; after B1, B2, B4a #14372 and B3 #14395).
No code can make an INT32-tagged Number since B4a, and codegen stopped reading one in B3. This PR deletes the
runtime side: every place that still decoded a
0x7FFE_0000_xxxx_xxxxword as a Number.JSValue::is_int32/as_int32,value::is_int32_bitsandINT32_MASKare gone. About 330 consumer arms gowith them in perry-runtime (equality and relational compare, loose equality, ToNumber / ToString / radix,
truthiness, typeof, JSON writers, console and
util.format, key and index decoding for Arrays, typed arrays,Buffers, Maps and Sets, argument validation in fs / net / dgram / tls / zlib / child_process / bun:ffi / sqlite),
perry-stdlib (crypto, webcrypto, sqlite, streams, events, querystring) and the http, net, sharp and
better-sqlite3 extensions.
INTERNAL_WORD_TAGin perry-runtime: the0x7FFEband holds runtime-internal words onlyand never a JS value.
JsValue::is_number,to_numberandto_int32read a plain double only.JsValue::is_int32stays,deprecated and always false, so an out-of-tree extension that calls it keeps compiling.
canonicalize_typed_slot_store_bits(andslot_holds_raw_f64), andjs_closure_call_array's defensiveargument decode with its scratch buffer;
numify_arith_operandandint32_or_class_reprlikewise.One thing the design did not know: I32 at rest
DESIGN-p3 was written before I32 at rest (N1). Since then the
0x7FFEband also holds the I32 slot words of anat-rest Array and of an I32 field lane (
layout_slot.rs); they are decoded by the layout accessor before a valueleaves the slot. So the sweep separates two kinds of reader:
layout_slot::is_i32_word(the samebits the old predicate tested, so nothing changes for it):
array::value_bits_to_number, the check-mode kindtests (
wpfacts_check::kind_of,contracts::kind_of), the GC census, typed feedback's stable kind and thestructured-clone writer.
A generic reader that wpfacts admits never meets an at-rest slot (that is the at-rest admission rule), so no
value-side arm was reachable through one. The language parity corpus and check mode below are the evidence.
Not in this PR
INT32_TAGconstants (the at-rest word and the control cells) keep their names.wasm_runtime.js(not checkable here)._slot_indexparameter of
runtime_store_jsvalue_slot_layout_deferred, and codegen's calls of the now-identityjs_typed_f64_arg_to_raw/js_typed_i32_arg_to_raw. Listed in FOLLOWUPS.Switch, probe
None: a pure deletion with no fact and no lowering change, like B3. The off arm is main dd083f3. The chess facts dump is byte-identical to main's (
facts_shabelow).Chess (qb10,
instructions:u, each arm x2, same session; off = main dd083f3, on = this branch 9f48f8a = main dd083f3 merged in)Second runs: search 11 848.183 / 11 845.511, perft 258.931 / 258.911, evaluate 1 191.814 / 1 190.590, legal
354.793 / 354.774. Binary 9 320 576 -> 9 308 288 B (−12 288). Facts dump identical (
facts_sha69df1bea5e8af5b7 inboth). Chess check mode: all four modes rc 0, 0 misses, in both arms.
Gates run
All on the merged head 9f48f8a (qb10). Earlier runs on the pre-merge heads 75baabc and 5c46e43 (base
b44f944) gave the same picture and are superseded by these.
tests/parity/language/number/test_p3_b4b_one_number_encoding.ts: = node in default, DCE off, check mode,check mode with DCE off.
cargo test -p perry-runtime --lib: 6819 passed, 0 failed.-p perry-ffi --lib: 55 passed.cargo test -p perry-stdlib --lib: 342 passed, 1 failed:sqlite::bind::tests::typed_arrays_are_not_named_parameter_objects. It fails the same way on main (checked onb44f944: the test binary has no typed-array family hooks); not from this PR.
dd083f3; against the newest one (main 12cace2) no program that matched there fails here, and the
number-births branch on the same base gives the identical per-file result.
canonnumruntime built from this branch: 203 executable controls passed(
canonical-number-negative: 0 / 86 / 86 for the array-rep arm).scripts/check_file_size.sh: OK (four allowlist counts re-recorded, all smaller).The merge of main
Main moved by 46 commits in the same files while the gates ran; it is merged in (not rebased). Two conflicts:
typedarray/bigint.rs(main moved the ToBigInt label intobigint_coerce.rs; its INT32 arm is deleted there) andthe
check_file_size.shcounts. The S0 hunk this branch first carried (acceptarray repincanonical-number-negative) is dropped: main fixed that arm itself, and the branch has main's script unchanged.instanceofand internal wordsOn main
x instanceof Canswers false whenxis a0x7FFEword that names no class, through the INT32 arms ofinstanceof_lhs_is_primitiveand of the prototype walk. Deleting those arms made the walk throw for such a word.Both are restored as "an internal word is no object" (unit tests
ordinary_has_instance_walk_is_false_for_primitive_lhs,lhs_primitive_classification_is_tag_only). No fixtureline: a JS program cannot reach the case, since no code makes a
0x7FFEword as a value (B4a) and an I32 slot wordis decoded before it leaves its slot.
Not run
Full suites other than the three lib suites above, parity IR census, node-suite check mode, tssrc, Supabase, skin
size, hello / hello2 startup Ir, the ext crates' ABI tests.