[ASDAlexander77]

Summary

Phase 4a of 32-bit support: a tagged union now reads back every member's fields exactly, on x64 and on i686. This fixes a miscompile that has been in x64 all along. The plan is docs/superpowers/plans/2026-09-22-32-bit-phase-4a-union-byte-storage.md.

The defect

A tagged union lowered to a packed {ptr tag, S}, where S was the LLVM struct of its largest member. Members go in and out through memory, but any bytes that fall in S's padding are not part of the value once it is loaded as S. So a smaller member's field stored there was lost the next time the union was copied.

• x64: in {flag; a: s32; b: s32} | {flag; v: number}, a read back as 0 (gc) or garbage (rc, none). 00bigint_struct_layout could not assert it.
• i686: a pointer followed by an 8-aligned f64 hits this in common shapes. 02union_type printed Error 9.11037e-305 downloading instead of Error 1 downloading.

The change

• Storage without padding. The value field is now packed <{[N/P x iP], [N%P x i8]}>: pointer-sized words plus a byte tail, where P is the pointer size in bytes. N is the old alloc size, so no union changes size. Words rather than bytes, because LLVM flattens a first-class [N x i8] argument into N separate byte arguments. LLVMTypeConverterHelper::getUnionStorageSize/getUnionStorageType.
• Every reader of the union value was audited. Three changed:
  • the union→union cast, and castToAny of a union: both used the largest member's type as the value type;
  • castLLVMTypes: aggregates of 8 bytes or less went through a LoadSaveOp that moves exactly one pointer, which is 4 bytes at i686. That shortcut is removed; the constant-size memcpy is folded anyway.
• Unused tail bytes are zeroed when a smaller member goes into a union. Before, they were undefined: at -O0, any === any of two identical values boxed from a union could be false, and the conservative GC saw garbage words.
• Module-level unions initialized with a constant use a global constructor. Building a union now goes through memory, which is not allowed in a global initializer. Before this change, a non-largest member already failed with ops with side effects not allowed in global initializers; now let g: number | string = "x" compiles as well. Only tagged-union globals move to a constructor. Non-tagged unions (string | undefined, literal unions) are untouched.
• The types have different sizes warning no longer fires for member-into-union conversions, which differ in size by design.

x64

x64 output changes for every tagged union, by design. x64 .text over the corpus (gc, --opt_level=3) changes by −0.03% (340240 → 340129 bytes). Exported functions that take or return unions change ABI: default-lib or DLL binaries built by the previous compiler must be rebuilt.

New: tslang/test/probe-corpus.sh

Compiles and runs every corpus file as a single-file exe with the suite's flags, for x64 and i686 × gc/rc/none, and writes a TSV. --compare before.tsv after.tsv lists every row that passed before and does not pass after. It is hand-run, like the other check-*.sh scripts. About 100 files per model fail at x64 too; those are multi-file tests that need the test runner.

pass counts	x64 before → after	i686 before → after
gc	423 → 425	410 → 417
rc	418 → 420	409 → 418
none	418 → 420	411 → 418

The file count goes from 520 to 522, because of the two new tests. --compare finds 0 regressions and 29 newly passing rows. At i686 the newly passing files are 02union_type, 03union_type, 03union_type_case_order, 00typeof_static_fold_conditions, parser, and arrayLiterals under rc.

Still failing at i686 only:

• the six async files, for phase 4b;
• internals, whose inline_asm<i64> with =r fits no single i686 register;
• 02funcs_vararg, which is out of the suite.

Found, not fixed

• An importer's module-level global initialized from a DLL-exported global reads garbage. For example, const copiedN: number = cN prints a denormal, and a union copy crashes silently. The old compiler does the same, and it happens at x64 too. It is recorded in the spec's open issues.
• 00array8_tuple_spread and arrayLiterals2ES5 at i686 under rc flipped between two probe runs of the same compiler before this change.

Test plan

• Windows Release suite: ctest -j 16 -C Release passes 2769/2769. That includes the new 00union_member_padding and 00union_global_init (compile and jit), and 00bigint_struct_layout with a asserted again.
• The new and the existing union tests pass at x64 and i686 under gc, rc and none. 00union_global_init also passes with --di --opt_level=0 and with --emit=jit.
• probe-corpus.sh --compare: 0 regressions against the base-commit baseline at either arch.
• --verify-ownership -mm=rc gives output identical to the base compiler on 62 union corpus files at both arches. The one exception is 00union_global_init, which the base compiler cannot compile. The verifier runs before LowerToLLVM.
• Two independent reviews. The first found the global-initializer regression and the per-byte argument cost; both are fixed here. The second approved the fixes.
• Linux CI. Not run locally.

🤖 Generated with Claude Code