fix(ir): diagnose non-constant global initializers loudly (issue 0072)

globalInitValue's issue-0071 .identifier arm closed the bare-identifier hole,
but .field_access (and every other non-literal expression shape) still fell
through to `else => null`, so a global like `g : s32 = K.x;` was emitted with
no payload and silently zero-initialized (g=0).

Make the `else` emit a diagnostic — "global '<name>' must be initialized by a
compile-time constant" — instead of a null payload, so no unsupported shape can
silently zero. Two arms added alongside:

- `.null_literal => .null_val`: a `*void = null` global was previously a
  no-payload zero-init; this preserves the exact LLVMConstNull emission (fixes
  3 ffi examples that regressed on the first cut).
- explicit `.enum_literal => null` carve-out: the stdlib's
  `OS : OperatingSystem = .unknown;` zero-init is load-bearing for compile-time
  `inline if OS == .X`; documented, not folded into a silent fallthrough.

Field-access constant *evaluation* (materializing K.x -> 9) is intentionally
not implemented: a typed struct const like K is not registered in
module_const_map, so it would require new plumbing whose writes are read at
runtime — out of scope. The diagnostic is the issue-sanctioned outcome.

Regression: examples/1118-diagnostics-global-non-const-initializer-rejected.sx
(exit 1). Gate: zig build, zig build test, run_examples.sh -> 356/0.
This commit is contained in:
agra
2026-06-02 17:57:17 +03:00
parent ad7200c196
commit b72d49073e
6 changed files with 153 additions and 4 deletions

View File

@@ -0,0 +1,113 @@
# 0072 — global field-access constant initializer silently zero-initializes
> **RESOLVED.** Root cause: `Lowering.globalInitValue`'s issue-0071 `.identifier`
> arm closed the bare-identifier hole, but `.field_access` (and every other
> non-literal expression shape) still fell through to `else => null`, so the
> global was emitted with no payload and silently zero-initialized (`g=0`).
> Fix: the `else` now emits a diagnostic — "global '<name>' must be initialized
> by a compile-time constant" — instead of returning a null payload, so an
> unsupported initializer shape can never silently zero. Two arms were added
> alongside it: `.null_literal => .null_val` (a `*void = null` global was
> previously a no-payload zero-init; this preserves that exact emission —
> `LLVMConstNull`), and an explicit `.enum_literal => null` carve-out (the
> stdlib's `OS : OperatingSystem = .unknown;` zero-init is load-bearing for
> compile-time `inline if OS == .X`; documented, not folded into a silent
> fallthrough). Field-access constant *evaluation* (materializing `K.x` → 9) was
> intentionally not implemented: a typed struct const like `K` is not registered
> in `module_const_map`, so it would require new plumbing whose `module_const_map`
> writes are read at runtime — out of scope; the diagnostic is the chosen,
> issue-sanctioned outcome. Regression
> `examples/1118-diagnostics-global-non-const-initializer-rejected.sx` (exit 1).
## Symptom
A top-level global initialized from a field access on a module constant compiles
but is zero-initialized.
Observed: `g=0`
Expected: `g` should be initialized to `9`, or the compiler should reject the
initializer loudly if field-access global constants are not supported yet.
## Reproduction
```sx
#import "modules/std.sx";
Point :: struct {
x: s32;
y: s32;
}
K : Point : Point.{ x = 9, y = 4 };
g : s32 = K.x;
main :: () -> s32 {
print("g={}\n", g);
return g;
}
```
Run:
```sh
./zig-out/bin/sx run .sx-tmp/review-0071-field-access-const.sx
```
The repro is standalone; the inline source above is sufficient to recreate the
scratch file under `.sx-tmp/`.
## Investigation prompt
Fix issue 0072: a top-level global initialized from a field access on a module
constant must not silently become zero.
Context:
- This surfaced during Codex re-review of commit `ad7200c`, the issue-0071 fix.
- `ad7200c` correctly handles identifier initializers that name module constants
(`K : A : 42; g : A = K;`) and diagnoses identifiers that are not usable
constants.
- A remaining non-identifier expression shape still falls through silently:
`K : Point : Point.{ x = 9, y = 4 }; g : s32 = K.x;` emits a null global
initializer payload and runs as `g=0`.
Suspected area:
- `src/ir/lower.zig`, `Lowering.globalInitValue`.
- The `.identifier` arm now diagnoses unusable identifier initializers, but the
generic `else => null` still covers `.field_access` and other expression
forms.
- `constExprValue` currently handles literals, negative numeric literals, and
array literals; it does not evaluate field access or struct-literal module
constants. `constStructLiteral` can serialize direct struct literals when
called with the destination type, but that machinery is not used for `K.x`.
Likely fix:
- Add a loud diagnostic for unsupported top-level global initializer expression
shapes, or implement field-access constant evaluation in the same step.
- If implementing it, resolve the base identifier through
`ProgramIndex.module_const_map`, materialize the base constant with its
recorded type, then extract the named field using the IR struct layout.
- Do not replace the current bug with a real-type sentinel such as `.void`; an
unsupported initializer should either produce a concrete `ConstantValue` or a
user-visible diagnostic.
- Preserve the issue-0071 behavior:
`K : A : 42; g : A = K;` must still emit `g=42`.
- Preserve the issue-0070/0068 regressions:
`examples/0133-types-forward-alias-global.sx` and
`examples/1117-diagnostics-value-const-as-type-rejected.sx` must stay green.
- Scope-check enum-literal globals separately. `OS : OperatingSystem = .unknown`
currently relies on compile-time constant injection and pre-existing zero-init
behavior; decide explicitly whether that stays carved out or gets its own
separate issue.
Verification:
- Repro above should either print `g=9` and exit `9`, or fail with a clear
"global initializer must be a compile-time constant" diagnostic. It must not
print `g=0`.
- Run:
```sh
zig build
zig build test
bash tests/run_examples.sh
```