989e18b760
Replace the bare-paren tuple grammar with explicit, position-unambiguous
forms, mirroring how structs work:
type `(A, B)` -> `Tuple(A, B)` (named keeps `:`)
value `(a, b)` -> `.(a, b)` (named uses `=`)
typed (new) -> `Tuple(A, B).(a, b)` (like `Point.{...}`)
failable `-> (T, !)` -> `-> T !`
`-> (T1, T2, !)`-> `-> Tuple(T1, T2) !` (channel outside Tuple)
Bare `(...)` is now grouping only, everywhere; a comma in bare parens is a
hard error with a migration hint. Grouping, function types `(A, B) -> R`,
param lists, lambdas, and match bindings are unaffected.
`Tuple(...)` is strictly a TYPE in every position (including `size_of` /
`type_info` args); a tuple VALUE comes only from `.(...)` (anonymous) or
`Tuple(...).(...)` (explicitly typed). A bare `Tuple(1, 2)` is a tuple
type with non-type elements -> rejected.
The ~110 tuple-bearing corpus files were migrated with a one-shot
AST-aware migrator (the `sx migrate` tool from the prior commit, removed
here). New examples: 0130 (new syntax), 0131 (typed construction), 1060
(named-tuple failable return). 1116 golden updated for the new hint text.
15 lines
590 B
Plaintext
15 lines
590 B
Plaintext
// `??` default whose type can't match a 1-tuple optional payload is a clean
|
|
// diagnostic, not an LLVM PHI-type crash.
|
|
//
|
|
// Regression (issue 0180): `?(i32,) ?? 5` built a merge PHI of `{i32}` (the
|
|
// unwrapped 1-tuple payload) vs `i32` (the scalar default) and aborted the
|
|
// LLVM verifier. There is no implicit scalar->1-tuple coercion (a 1-tuple
|
|
// value is written `(5,)`), so the mismatch is now reported at the default.
|
|
#import "modules/std.sx";
|
|
|
|
main :: () {
|
|
o : ?Tuple(i32) = null;
|
|
x := o ?? 5; // default 'i64' vs payload '(i32,)' -> diagnostic
|
|
print("{}\n", x.0);
|
|
}
|