Logs the 2a.C/2a.D pair (6b7a66b + e6d6903) — the if-return
regression in issue-0045's original fix, now resolved via the
SSA "return-done block" pattern.
Step 2 is functionally sufficient for the impl-matching payoff
(step 5 of the plan); per-call-shape monomorphisation deferred
until a real workload demands it.
Step 3 ("Type-reflection intrinsics") flagged as the natural
next slice: `$args[$i]` in type positions unlocks the block-
trampoline body in stdlib's generic Into(Block) impl, which is
the visible payoff of the whole pack feature.
Fixes the regression locked in by 2a.C (commit 6b7a66b).
issue-0045's original fix set `block_terminated = true` after
each inline `return X;` to skip dead code in the inlined body.
But the flag leaked past structured control flow — an `if cond
{ return X; }` whose merge block continued to subsequent
statements would short-circuit the trailing code at the
`lowerBlockValue` loop's `if (self.block_terminated) return
null;` check.
Switched to the classical SSA "return-done block" shape:
- `InlineReturnInfo` carries a third field `done_bb: BlockId`
— a fresh basic block allocated by `lowerComptimeCall` per
comptime-call instance.
- `lowerReturn`'s inline path stores into the slot, drains
defers, and emits `br done_bb`. The basic block's terminator
is what carries the "no fall-through" signal; the
`block_terminated` flag is no longer touched.
- `lowerComptimeCall` allocates the slot + done_bb, lowers the
body, then switches to done_bb and loads the slot. Tail-
expression bodies that fall through (rare when has_return is
true) get a synthetic store + br so the CFG is well-formed.
For `if cond { return 42; }; return -1;`:
- cond=true: then's `return 42` stores 42, br done_bb. Merge
block has only the false predecessor, doesn't run the
trailing return. Load done_bb → 42.
- cond=false: condBr skips to merge. Merge runs `return -1;`
→ store -1, br done_bb. Load → -1.
`examples/157-pack-if-return.sx` flips from `8354116000` (the
uninitialised slot load on the false path) to `-1`. A
three-way `classify(..$args)` smoke confirms multi-path
inline-return works for any of the three branches.
Dead-code-after-return inside the inlined body still trips the
LLVM verifier (same shape as a regular `return X; print("dead");`
which also crashes today). Acceptable consistency — user code
shouldn't write unreachable code in either context.
196/196 example tests + `zig build test` green.
Follow-up to issue-0045's fix (commit 9e78790). The fix routes
inline-comptime-body `return X;` into a result slot but sets
`block_terminated = true` after the inline return — and that
flag leaks past the enclosing `if`'s merge block.
Body shape:
maybe :: (..$args) -> s64 {
if args.len > 0 { return 42; }
return -1;
}
For `maybe()` (zero call-args), the false-condition path skips
the then-branch's `return 42;` and should fall through to
`return -1;`. Today's flow:
- Then-branch's `return 42;` stores 42 to slot and sets
block_terminated = true.
- if lowering switches to merge_bb. block_terminated stays
true (never reset across the if/merge boundary).
- lowerBlockValue's loop sees block_terminated and returns
null without processing the trailing `return -1;`.
- lowerComptimeCall loads slot — slot was never written on
the false-condition path → garbage (8354116000 on this
machine; stable across runs).
`maybe(99)` works because the cond is true; the then-branch's
store wins.
Next commit reshapes the inline-return mechanism to use a
dedicated "return-done" basic block: each inline `return X;`
stores to slot and branches to ret_done; after the body
lowers, lowerComptimeCall switches to ret_done and loads. The
basic block CFG carries the control-flow termination — no
need for the leaking `block_terminated` flag.
196/196 example tests + `zig build test` green (the new test
captures the wrong value as the snapshot to flip).
Comptime fn body containing BOTH a nested comptime call
(`print(...)`) AND a `return X;` fails in one of two shapes
depending on the comptime-param flavour: a `storeAtRawPtr`
panic in the interp (plain `$x: s32` comptime) or "unresolved
'result'" at compile time (pack-fn `..$args`).
Same root: my issue-0045 fix's `inline_return_target` slot
setup interacts badly with the recursive comptime-call path
that invokes `#insert build_format(fmt)` → interpreter →
parse-and-lower of `result := ...` statements.
Pre-issue-0045-fix the pattern crashed at the LLVM verifier
("Terminator found in the middle of a basic block") so the
recursive path never ran. The fix exposed the deeper bug; it
didn't create it.
Not blocking the next pack-feature slices:
- Step 2a tests use arrow-form bodies with no nested print.
- Steps 2b/3 don't inherently require nested comptime calls —
builders run inside `#insert` contexts, not inside public
pack-fn bodies.
- Will bite when step 5 refactors stdlib's `print`/`format` to
`..$args` or when user code writes a pack-fn with both
`print` debug output and an early `return`.
Investigation prompt in the issue file points at
`createComptimeFunction`'s saved/restored state list (missing
`inline_return_target`, `pack_arg_nodes`,
`comptime_param_nodes`) as the most likely angle.
Logs the two step-2a commits (223ec3d lock-in, cd36784 fix):
pack-fn bodies' `args[<int_literal>]` now substitutes the i-th
call-site argument's lowered value directly, propagating the
concrete type through field access and typed coercion.
Out-of-scope follow-ups noted: non-literal comptime indices,
type-position `$args[$i]` (step 3), per-mono mangling, and the
pre-existing nested-comptime-call scope bug (a pack-fn body
calling `print(...)` AND containing `return X;` trips
"unresolved 'result'" — same shape as the issue-0045 family,
worth filing separately if a later slice needs it).
Pack-fn bodies that index the pack via `args[<int_literal>]`
now resolve to the i-th call-site argument's lowered value
directly, propagating the call arg's concrete type instead
of the boxed `Any` that the `[]Any` slice path returns.
New plumbing in `src/ir/lower.zig`:
- `pack_arg_nodes: ?std.StringHashMap([]const *const Node)` on
Lowering. Maps a pack param name (e.g. "args") to the slice
of call-site arg AST nodes.
- `lowerComptimeCall` populates the map when the variadic
param is heterogeneous (`is_variadic AND is_comptime`, i.e.
the `..$args` form). Plain `args: ..Any` keeps the existing
`[]Any` slice path so stdlib's `format`/`print` continue
unchanged. The map is saved/restored across nested calls
mirroring `comptime_param_nodes`.
- `packArgNodeAt(ie)` returns the call-arg node when an
index_expr matches `<pack_name>[<comptime_int_literal>]`
with the index in range; null otherwise (fall through to
standard slice indexing for runtime indices or non-pack
bases).
- `lowerIndexExpr` checks `packArgNodeAt` first; on a hit it
lowers the call arg node directly. `inferExprType`'s
`index_expr` arm does the parallel check so AST-level type
inference (e.g., for field-access type checking) sees the
concrete call-arg type.
`examples/156-pack-typed-index.sx` flips from
"field 'x' not found on type 'Any'" to `7` — `args[0].x` now
resolves through the concrete `Point` type instead of Any.
Out of scope (deferred): non-literal comptime indices
(`args[$i]` where `$i` is an arbitrary comptime expression);
`$args[$i]` in type positions (step 3); per-mono mangling
(monomorphisation stays inline-only).
195/195 example tests + `zig build test` green.
Step 2 of the variadic heterogeneous type packs feature: typed
runtime indexing (`args[$i]` at comptime-known `$i`). Today's
pack-fn body lowers `args[i]` through the `[]Any` slice path —
the static type returned is `Any`, so any downstream field
access / typed-coercion / further indexing fails the moment it
needs more than primitive auto-unboxing.
`examples/156-pack-typed-index.sx` pins the simplest visible
failure: `args[0].x` on a struct-typed call arg trips
"field 'x' not found on type 'Any'" at the field-access site
because AST-level type inference for `args[0]` returns Any.
Next commit teaches `lowerIndexExpr` (and `inferExprType` for
the same shape) to detect an index_expr whose base is a
pack-name binding from the enclosing comptime call AND whose
index is a comptime int literal — substitutes the i-th
call-site arg's lowered value directly, propagating the call
arg's concrete type through field access, typed assignments,
and further indexing. The `[]Any` slice path stays as the
runtime-indexed fallback for `args[i]` where `i` is not a
comptime constant.
195/195 example tests + `zig build test` green.
Logs the issue-0045 fix (inline-return slot for comptime-call
bodies, commits 3d32ab0 + 9e78790) — the LLVM verifier crash
that surfaced when probing step 2 of the pack feature. Test
count now 194/194 with the new regression test.
`lowerComptimeCall` now scans the body for `return` statements
via `fnBodyHasReturn`. When found, it allocates a stack slot
typed to the fn's return type and installs it as
`self.inline_return_target` before lowering the body.
`lowerReturn` checks `inline_return_target` first:
- If set, it stores the coerced return value into the slot,
drains pending defers, sets `block_terminated = true`, and
returns without emitting a `ret` into the caller's basic
block.
- Otherwise it emits the standard `ret` as before.
After the body lowers, the inliner either returns the
tail-expression value (existing fast path — bodies with no
`return` skip the slot entirely) or loads the slot when
`block_terminated` is set.
Why the bug was invisible until now: `format`/`print` and
every other stdlib comptime fn use arrow form (`=> expr`) or
`#insert`-only bodies — no `return` statement, no path through
`lowerReturn`. Step 1.b of the pack feature made `..$args`
parseable; the natural smoke test
`foo :: (..$args) -> s64 { return 42; }` was the first
comptime-fn body to take the `return`-with-trailing-statements
path, surfacing the LLVM verifier crash.
`examples/issue-0045.sx` flips from the lock-in failure to
`42`. 194/194 example tests + `zig build test` green.
Filed `issues/0045-pack-fn-call-llvm-verifier-failure.md`.
Surfaced by probing step 2 territory of the variadic
heterogeneous type packs feature: any `..$args` fn whose body
is a block containing `return X;` (or any comptime fn with a
non-void return, comptime params, and explicit `return` in a
block body) trips LLVM's "Terminator found in the middle of a
basic block" verifier.
`lowerComptimeCall` inlines the body's statements directly into
the caller's LLVM function. `lowerReturn` then emits a `ret`
into the caller's basic block — but the caller still has
trailing instructions, hence the verifier failure.
`examples/issue-0045.sx` reproduces the crash with the minimum
pack-fn shape (`foo :: (..$args) -> s64 { return 42; }`). Same
shape with a plain comptime param (`($x: s32) -> s64 { return
42; }`) reproduces identically, so the bug is broader than
packs. Arrow-form bodies (`=> 42`) work today because they have
no `return` statement.
Next commit teaches `lowerComptimeCall` to allocate a result
slot when the body contains a `return`, and reroutes
`lowerReturn` to store into that slot + flag the block as
terminated so the inliner picks up the value.
Logs the four commits that closed out step 1 of the variadic
heterogeneous type packs feature:
- 1c.A: parser-rejection lock-in for `..$args` inside `Closure(...)`.
- 1c.B: parser + AST + types.zig `pack_start` representation.
- 1d.A: impl-matching concrete-only miss lock-in.
- 1d.B: pack-aware impl matching with $args + $R binding through
`param_impl_pack_map` and `pack_bindings`.
Next step is plan step 2 — runtime `args[$i]` indexing + per-mono
mangling — opening the door to body-side pack reflection that
step 5 needs to retire the hand-rolled `Into(Block)` impls.
Also catches up the 1b entry which the prior session left
uncommitted in the working tree.
Pack-shaped impls (`impl P(...) for Closure(..$args) -> $R`) now
match concrete closure sources at xx resolution time. Concrete
impls keep their priority — pack matching only fires on a
concrete-key miss in `param_impl_map`.
New plumbing in src/ir/lower.zig:
- `PackParamImplEntry` carries the pack-shaped source TypeId plus
the pack-var and ret-var names extracted from the impl AST's
`target_type_expr`. `registerParamImpl` detects pack-shaped
sources via `pack_start != null` on the resolved closure type
and additionally registers in a new `param_impl_pack_map`
keyed by `"Proto\x00<arg_mangled>"` (no source suffix).
- `tryUserConversion` re-shapes the concrete lookup so the pack
path runs on miss. `tryPackImplMatch` walks the pack entries,
verifies the source's fixed prefix matches the impl's prefix,
binds the pack-var to the source's tail param TypeIds, binds
the ret-var (when the impl's return is generic) to the source
return, and monomorphises the convert method. Mangled name
stays keyed on the concrete source so distinct call shapes
monomorphise separately.
- `pack_bindings: ?StringHashMap([]const TypeId)` is saved/
restored around monomorphisation, mirroring `type_bindings`.
- `resolveClosureTypeWithBindings` handles the closure_type_expr
node during type resolution: when the closure carries a
`pack_name` AND `pack_bindings` has a binding for it, the
bound TypeIds are appended after the fixed prefix and the
result is a concrete (non-pack) closure type — so the impl
body's `self: Closure(..$args) -> $R` substitutes to the
concrete source closure during monomorphisation. Without an
active binding, the pack shape is preserved.
`examples/155-pack-impl-match.sx` flips from the
"no Into(Block) for cl_s32_bool__bool" lock-in diagnostic to
"pack impl match ok": one user-declared
`impl Into(Block) for Closure(..$args) -> $R` covers a
`Closure(s32, bool) -> bool` source that stdlib has no
hand-rolled impl for. Constructed Block isn't invoked
(invoke=null) — the test exercises only the matching +
monomorphisation, not the trampoline (step 5 of the plan).
Existing concrete-impl paths unchanged: 95-objc-block-noop,
96-objc-block-multi-arg, and stdlib's hand-rolled
`Into(Block) for Closure(bool) -> void` continue to pass through
the concrete map first. Same-file duplicate pack impls
diagnose at registration; cross-module visibility and
multi-pack-impl specificity stay TODOs (matching the deferred
Phase 5 work on the concrete path).
193/193 example tests + `zig build test` green.
Step 1d lock-in test pinning today's matching behaviour.
`registerParamImpl` records every impl in `param_impl_map` keyed
by `"Proto\x00<arg_mangled>\x00<src_mangled>"`. For a pack impl
`Into(Block) for Closure(..$args) -> $R` the key contains the
pack-shaped closure's mangle (interns with `pack_start = Some(0)`
after 1c.B). At the `xx cl : *Block` site the lookup mangles the
concrete `Closure(s32, bool) -> bool` source and finds nothing —
the existing focused diagnostic fires:
no `Into(Block) for cl_s32_bool__bool` impl — add a per-signature
`__block_invoke_<sig>` trampoline + Into impl alongside the
existing ones in modules/std/objc_block.sx, or declare it in
your own code
The pack impl is reachable in the file but never considered.
Next commit (1d.B):
- New `param_impl_pack_map` keyed by `"Proto\x00<arg_mangled>"`
(no src) — populated by `registerParamImpl` when the source
is pack-shaped.
- `tryUserConversion` walks the pack map on concrete-key miss.
Pack shape matches when the impl's fixed prefix equals the
source's matching prefix; the remainder binds to `$args` and
the source's return type binds to `$R`. Concrete impls win
over pack impls (specificity).
- `resolveTypeWithBindings` learns the closure_type_expr path
so the impl body's `self: Closure(..$args) -> $R` substitutes
to the concrete source closure during monomorphisation.
The `Closure(s32, bool) -> bool` shape is not covered by stdlib
or 96-block-multi-arg's hand-rolled impls, so the pack impl is
the only candidate post-1d.B.
193/193 example tests + `zig build test` green.
`parseTypeExpr`'s `Closure(...)` arm now accepts a trailing
`..$name` (sigil optional) as a variadic-pack marker. Pack must
be terminal — `)` is the only token accepted after the name.
`ClosureTypeExpr` AST gains `pack_name: ?[]const u8` carrying the
identifier so later slices can name the binding.
`FunctionInfo` / `ClosureInfo` in src/ir/types.zig grow a
`pack_start: ?u32 = null` field. `Closure(..$args) -> R` interns
as `params = []`, `pack_start = Some(0)` — distinct from any
concrete `Closure(...) -> R` shape thanks to updated hash/eql
arms. New constructor pair `closureTypePack` /
`functionTypePack` keeps the existing single-shape constructors
unchanged.
`type_bridge.resolveClosureType` calls `closureTypePack` when
`pack_name != null`. The pack starts after the fixed prefix,
so `Closure(Prefix, ..$args)` resolves with `params = [Prefix]`,
`pack_start = Some(1)`.
No semantic effect yet — the signature exists in the type table
but no matching code reads `pack_start`. Step 1d wires impl
matching: `Closure(..$args) -> $R` binds against any concrete
closure source type in `tryUserConversion` / `registerParamImpl`.
`examples/154-pack-type-rep.sx` flips from rejecting-with-error
to positive parse smoke (prints "pack type rep ok").
192/192 example tests + `zig build test` green.
Next slice of the variadic heterogeneous type packs (`..$args`)
feature: type-system representation. Per the FFI cadence rule, this
commit locks in the parser-rejection behavior so the next commit's
type-rep extension surfaces as a behavior shift.
examples/154-pack-type-rep.sx uses `..$args` inside a `Closure(...)`
type expression — the pack-shape spelling used by impl headers like
`impl Into(Block) for Closure(..$args) -> $R`. Today's parser
recognizes `..$args` only at the parameter-list site (1b);
`parseTypeExpr`'s `Closure(...)` arm calls `parseTypeExpr` per
position and hits "expected type name" at the `..` token. Snapshot
captures the rejection at line 18, column 26.
Next commit (1c.B):
- Parser: `parseTypeExpr` Closure arm accepts `..$args` as the
trailing pack marker. AST gets a `pack_name: ?[]const u8` (or
equivalent) field on `ClosureTypeExpr`.
- types.zig: `FunctionInfo` / `ClosureInfo` gain `pack_start: ?u32`
so the pack shape is distinct from any concrete arity in the
type table. Hash/eql updated.
- type_bridge: `resolveClosureType` threads pack_start through.
- 154 flips green.
192/192 example tests + `zig build test` green.
Extends parseParams in src/parser.zig:1558 to recognize a leading
`..` before the optional `$` sigil and the parameter name. The
old `args: ..T` form (variadic marker after the colon) still
works — both paths set the same `is_variadic` flag.
A pack declaration `..$args` parses as:
- `is_variadic = true` (from the leading `..`)
- `is_comptime = true` (from the `$` sigil)
- `type_expr = inferred_type` (no `:` annotation)
The no-colon branch now propagates `is_variadic` and `is_comptime`
onto the Param struct so later slices (type rep, impl matching,
monomorphisation) can read both flags from the parsed AST without
re-deriving from token sequence.
`examples/150-pack-parse.sx` flips from rejecting-with-error to
positive parse smoke. No semantic effect yet — `foo` is declared
but never instantiated.
191/191 example tests + `zig build test` green.
First slice of the `..$args` (variadic heterogeneous type pack)
feature. Locks in the current parser-rejection behavior so the
next commit's parser extension shows up as a behavior shift.
`examples/150-pack-parse.sx` declares `foo :: (..$args) -> s64`.
Today's parser hits `..` where it expects a parameter name
(parseParams in src/parser.zig:1558 only handles `..` inside the
type position after a colon) and emits "expected parameter name".
Expected output captures this rejection.
Per FFI cadence rule, this is the "test fails today, passes after
next commit's parser change" pair.
Pack feature plan saved at
~/.claude/plans/lets-see-options-for-merry-dijkstra.md ("Variadic
heterogeneous type packs" section). Motivates replacing the
hand-rolled per-signature `Into(Block)` impls with one generic
`impl Into(Block) for Closure(..$args) -> $R`; also unlocks
compile-time arity/type errors for `print`/`format`.
191/191 example tests + `zig build test` green.
Reconsidered the M5.A.2 cleanup. The compiler-synthesised trampoline
path was hidden behaviour — a user reading their code couldn't tell
how `xx my_closure : Block` worked without reading lower.zig. That's
exactly the kind of magic sx's design has been pushing against.
New design (strict mode):
1. Stdlib's modules/std/objc_block.sx hand-rolls
`__block_invoke_void` + `Into(Block) for Closure() -> void` and
the same pair for `Closure(bool) -> void` (restored from M5.A.2).
These are readable reference implementations of the bridge ABI.
2. The compiler intercept fires NO synthesis — instead, when
`tryUserConversion` can't find a reachable `Into(Block)` impl for
the closure's signature, it emits a focused diagnostic:
"no `Into(Block) for <Closure-sig>` impl — add a per-signature
`__block_invoke_<sig>` trampoline + Into impl alongside the
existing ones in modules/std/objc_block.sx, or declare it in
your own code"
3. Per-signature declarations live in stdlib (for common signatures)
or in user code (for app-specific ones). 96-objc-block-multi-arg
now demonstrates the user-side pattern in-file — it declares its
own `__block_invoke_void_s32_p` + `Into(Block) for Closure(s32,
*void) -> void` impl alongside its main().
Net effect:
- Every block bridge is source-visible. No hidden compiler magic.
- Users see exactly how the Apple ABI shape is constructed in sx
source — stdlib serves as the reference implementation.
- Compiler enforces the discipline: missing impl → clear diagnostic
pointing at the template.
- Coverage for arbitrary signatures requires conscious user opt-in,
not silent fallthrough.
Removed from lower.zig: `tryClosureToBlockConversion`,
`emitBlockInvokeTrampoline`, `mangleClosureSigForBlock`,
`mangleTypeForBlock`, and the `block_invoke_trampolines` dedup
state field. Net: the synthesis machinery is gone; only the
detection helper `isClosureToBlockCast` remains, used by the
diagnostic.
190/190 example tests pass; chess on iOS-sim green.
A signature the hand-rolled stdlib never covered: `Closure(s32, *void) -> void`.
Pre-M5.A this code wouldn't compile (no `Into(Block) for Closure(s32, *void) -> void`
declaration); post-M5.A the compiler emits `__block_invoke_v_i_p` on
demand and the call site goes through it.
The test uses two-arg side-effect capture (globals `g_sum`, `g_tag`)
to verify both args reached the closure body. Confirms the
trampoline's calling convention forwards
`(__sx_default_context, sx_env, arg0, arg1)` correctly through to
the closure's underlying fn.
Note: return-value signatures (e.g. `Closure(s32) -> s32`) are
recognised by the trampoline emitter — `cinfo.ret` flows through
to `beginFunction`'s return slot — but exercising them requires
closure-return-type inference that the test runner stumbled on
during authoring (`(n: s32) => { return n+1; }` infers void). The
void-returning shape is the more common Cocoa pattern (animation
bodies, dispatch_async, completion handlers); return-value
signatures land properly once the closure inference catches up
(orthogonal to M5.A).
190/190 example tests pass.
The compiler-synthesised trampoline path (previous commit) covers
every closure signature on demand; the hand-rolled stdlib impls
were only for two specific shapes (`Closure() -> void`,
`Closure(bool) -> void`) and are now strictly redundant.
Kept: the `Block` struct, `BlockDescriptor`, the
`_NSConcreteStackBlock` extern decl, and the shared
`__sx_block_descriptor` global. The compiler-emitted code
references all four; users still need to `#import
"modules/std/objc_block.sx";` to bring them into the module.
Removed: `__block_invoke_void`, `__block_invoke_bool`, and both
`impl Into(Block) for Closure(...) -> void` blocks. Replaced with
a comment block explaining how the compiler now handles the cast.
After this commit, `xx my_closure : Block` works for ANY closure
signature with no per-signature stdlib boilerplate. 189/189
example tests pass; chess on iOS-sim green.
`xx closure : Block` casts now bypass the user-space Into(Block)
protocol path entirely. The compiler intercepts in
`tryUserConversion` BEFORE the Into lookup, detects when src is
`Closure(...)` and dst is `Block`, and emits:
1. A C-ABI trampoline `__block_invoke_<sig>` (deduped per closure
signature via `block_invoke_trampolines` map). Body matches the
existing hand-rolled `__block_invoke_void` exactly: load
block_self struct, extract sx_env (field 5) + sx_fn (field 6),
call sx_fn(__sx_default_context, sx_env, ...user_args), return.
2. Inline Block-struct construction at the cast site:
`Block { isa = &_NSConcreteStackBlock, flags=0, reserved=0,
invoke = &__block_invoke_<sig>,
descriptor = &__sx_block_descriptor,
sx_env = closure.env, sx_fn = closure.fn_ptr }`
Signature mangling: compact codes — `v` void, `b` bool, `i` s32,
`q` s64, `f` f32, `d` f64, `c/C/s/S/I/Q` for other ints, `p` for
pointers/aggregates that lower to a machine word. Return first,
then params underscore-joined. `Closure() -> void` mangles to `v`;
`Closure(bool) -> void` mangles to `v_b`.
Loud failures at the cast site:
- `Block` struct missing → "requires #import \"modules/std/objc_block.sx\";"
- `_NSConcreteStackBlock` extern missing → same diagnostic.
- `__sx_block_descriptor` global missing → same.
- `__sx_default_context` missing inside the trampoline emitter →
compiler-bug diagnostic (the scan pass should always register it).
The existing hand-rolled stdlib impls (`__block_invoke_void`,
`__block_invoke_bool`, the two `Into(Block) for Closure(...)`
impls) are now redundant — the compiler-synthesised trampoline
takes over via the intercept. Next commit (M5.A.2) removes them.
95-objc-block-noop continues to pass; IR shows `__block_invoke_v`
(the synthesised name) replacing the hand-rolled
`__block_invoke_void` at the cast site. 189/189 example tests
pass; chess on iOS-sim green.
12 commits this session shipped the entire M4 milestone. sx-defined
Obj-C classes now honor `context.allocator` end-to-end, route through
NSObject's retain/release at the source-language level, and emit
correct ARC ops in property setters/getters/dealloc per the Apple
ABI contract.
189/189 example tests pass; chess on iOS-sim green throughout.
Ready for M5 (closure↔block bridge) or M1.1.b (Class(T) phantom
typing) next session.
emitObjcDefinedClassDeallocImp now walks the class's #property fields
BEFORE freeing the state struct. For each:
- assign → no-op (primitives, no ARC traffic).
- strong → val = load field; objc_release(val).
- copy → same as strong (the stored value is a +1 retained copy
produced by the setter's [val copy]; we release it here).
- weak → objc_destroyWeak(&field) — unregisters the slot from
libobjc's side-table so the runtime stops tracking it.
Order matters: property releases happen BEFORE freeing the state
struct (which would invalidate the pointers we need to read), which
happens BEFORE [super dealloc] (which eventually frees the Obj-C
instance's own memory). The full sequence is now:
%state = object_getIvar(self, __sx_state_ivar)
// M4.B (this commit):
for each strong/copy property P:
val = load struct_gep(state, P.idx); objc_release(val)
for each weak property P:
objc_destroyWeak(struct_gep(state, P.idx))
// M4.0c (already shipped):
allocator = load struct_gep(state, 0)
allocator.dealloc(state)
object_setIvar(self, ivar, null)
// M1.2 A.6:
[super dealloc] // → objc_msgSendSuper2
ffi-objc-arc-02-strong-property now passes: child held by parent's
strong property gets released when parent deallocates, refcount → 0,
child deallocates, both states freed via tracker. Balanced 2/2.
189/189 example tests pass; chess on iOS-sim green. M4 complete.
emitObjcDefinedPropertyGetter dispatches on objcPropertyKind. The
strong/copy/assign paths keep their bare load. The weak path:
retained = objc_loadWeakRetained(field_addr)
autoreleased = objc_autorelease(retained)
return autoreleased
`objc_loadWeakRetained` does the race-safe upgrade via libobjc's
side-table: if the target has deinitialized (or is mid-dealloc on
another thread), returns null; otherwise returns the target with
refcount bumped (+1 retained, transferred to caller).
`objc_autorelease` drops the +1 into the current pool so the
caller doesn't need to manually balance — matches Apple's auto-nil
weak-getter contract.
The bare-load weak path (still in place pre-M4.B-getter) worked
for the single-threaded test scenario because the runtime nils the
slot before the load happens. The load-retained version covers the
multi-threaded "between load and use, target deinit's" race that
silent bare-load can't.
189/189 example tests pass; chess on iOS-sim green.
emitObjcDefinedPropertySetter now dispatches on objcPropertyKind to
emit the right runtime ops per Apple's ARC contract:
- assign → bare store (primitives, explicitly opted-out object slots).
- strong → load old; objc_retain(new); store new; objc_release(old).
Apple's runtime treats release(NULL) as a safe no-op, so
no explicit null-check on the old value.
- weak → objc_storeWeak(field_addr, val) — handles first-store
(init) and re-store (destroy + init) atomically. Registers
the slot with libobjc's side-table; the runtime auto-nils
it when the target deallocates.
- copy → [val copy] (sends `copy` selector — returns retained per
the NSCopying contract); load old; store the copied
instance; release old.
Side-effect on the weak path: even with the bare-load getter still in
place (loaded directly from the slot), weak reads work because Apple's
runtime side-table-nils the slot at target dealloc. The getter
improvement via objc_loadWeakRetained is the next commit and is
needed for race-safe reads (between load and use, the target could
deinit on another thread); for the single-threaded test scenarios
the bare load is sufficient.
ffi-objc-arc-02-strong-property advances from "child dealloc'd at
midpoint" to "unbalanced; alloc=2 dealloc=1" — strong setter now
retains, but the M4.B-dealloc cleanup hasn't landed so the child
held by the property isn't released when the parent deallocates.
Final commit (M4.B dealloc) closes the loop.
ffi-objc-arc-03-weak-property turns fully green: storeWeak +
auto-nil side-table do the work.
189/189 example tests pass; chess on iOS-sim green.
Three pieces, no behavior change yet:
1. `ObjcPropertyKind` enum (strong/weak/copy/assign) + `objcPropertyKind`
helper in lower.zig. Reads `field.property_modifiers`, applies the
default rule (`*<ObjC-class>` → strong; primitives → assign), and
emits loud diagnostics for the silent-error budget:
- unknown modifier name (typo) → "expected one of: strong, weak, copy, ..."
- conflicting modifiers (e.g. `strong,weak`) → "mutually exclusive"
- `weak` on non-object slot → "requires a pointer-to-Obj-C-class type"
- `copy` on non-object slot → same
- `strong` (default or explicit) on `*void` → "ambiguous: specify
#property(strong|weak|copy|assign) explicitly"
Called from `emitObjcDefinedClassPropertyImps` for validation; the
returned kind isn't wired into setter/getter/dealloc yet — that's
the next three commits.
2. `ensureArcRuntimeDecls` lazily declares libobjc's ARC helpers:
objc_retain, objc_release, objc_storeWeak, objc_loadWeakRetained,
objc_initWeak, objc_destroyWeak. Uses the existing
`ensureCRuntimeDecl` pattern; idempotent.
3. Fix existing NSObject method names in std/objc.sx — `isEqual_`,
`isKindOfClass_`, `respondsToSelector_` had trailing underscores
that the selector mangling turned into double-colon selectors
(`isEqual::`). Removed the trailing underscore so the selectors
come out as `isEqual:`, `isKindOfClass:`, `respondsToSelector:`
as Apple's runtime expects.
4. Two xfail regression tests:
- ffi-objc-arc-02-strong-property: assigns child to parent's strong
property, releases the original child reference. Midpoint check:
child's dealloc should NOT have fired (strong setter retained).
Pre-M4.B-setter: child dealloc fires immediately → "FAIL: child
dealloc'd at midpoint" snapshot. Exit code 1.
- ffi-objc-arc-03-weak-property: assigns target to holder's weak
property, releases target. Reads holder.target → should be null
(auto-niled). Pre-M4.B-getter/setter: reads stale pointer →
"FAIL: weak property didn't auto-nil" snapshot.
These will turn green as M4.B setter (commit 2), getter (commit 3),
and dealloc-cleanup (commit 4) land. Each subsequent commit updates
the snapshot to reflect the now-passing output.
189/189 example tests pass; chess on iOS-sim green.
Two regression tests pinning down the silent-error surface in M4.0:
ffi-objc-arc-00 — single sx-defined-class instance round-trips
through a TrackingAllocator-wrapped GPA. Captures alloc/dealloc
deltas around the lifecycle, verifies (+1, +1). Pre-M4.0 the +alloc
IMP used libc malloc and -dealloc used libc free; tracker would
have observed (+0, +0) and missed the leak silently.
ffi-objc-arc-00b — three instances alloc'd and released. Catches
bugs where:
- the captured allocator becomes shared (one global slot vs
per-instance);
- alloc captures the wrong allocator on the 2nd+ instance;
- dealloc reads garbage if state[0] is overwritten between
instances.
Both tests are macos-only (libobjc + NSObject must be present at
runtime). Both wrap the lifecycle in `push Context.{ allocator =
xx tracker }` so the threading path is exercised.
Important authoring note: `print` inside the push-block also routes
through tracker (string formatting allocs), polluting the leak
delta. Tests capture before/after counts WITHOUT any prints between
alloc and release, then verify the BALANCE — every alloc paired
with a dealloc — rather than absolute counts. Discovered while
writing 00: an initial naive "leak_count() == 0" assertion failed
not because M4.0 was broken but because print's string allocs
weren't freed at scope exit.
187/187 example tests pass.
Snapshot of FFI progress mid-M4. Allocator-aware sx-defined-class
lifecycle is done end-to-end (state struct + +alloc + -dealloc).
Stdlib NSObject + autoreleasepool helper landed; defer-release
pattern works at user code. Property ARC ops (M4.B) is the next
slice.
185/185 example tests pass; chess on iOS-sim regression-verified
at every M4 sub-commit.
Declare `NSObject` in std/objc.sx as `#foreign #objc_class("NSObject")`
with the canonical instance + class-method surface every Obj-C class
inherits: `retain`/`release`/`autorelease`/`new`/`alloc`/`init`/
`description`/`hash`/`isEqual_`/`isKindOfClass_`/`respondsToSelector_`/
`class`. Root the foreign-class hierarchy in uikit.sx at NSObject by
adding `#extends NSObject;` to every previously-unrooted declaration
(NSValue, NSNumber, NSDictionary, NSSet, NSNotification, NSBundle,
NSNotificationCenter, NSRunLoop, CADisplayLink, CALayer, EAGLContext,
UIScreen, UIResponder) plus deeper chain fixes (NSMutableDictionary
extends NSDictionary; UIWindow extends UIView; UIViewController
extends UIResponder). After this, M2.3's extends-chain walk finds
`retain`/`release` on any UIKit-typed value:
view := UIView.alloc().init();
defer view.release(); // canonical sx idiom — no language magic
Plus `autoreleasepool(body: Closure())` stdlib helper that wraps
`body` in `objc_autoreleasePoolPush` / `defer objc_autoreleasePoolPop`.
Required for Foundation factory returns; closure-call frame is real
cost so hot loops should inline the push/defer-pop pattern manually.
Smoke test `ffi-objc-arc-01-autoreleasepool.sx` exercises both
patterns; refresh of two IR snapshots picks up the new stdlib decls
appearing in test outputs that include `modules/std/objc.sx`.
185/185 example tests pass; chess on iOS-sim green.
The synthesized -dealloc IMP now loads `state->__sx_allocator` (the
slot captured at +alloc time by M4.0a + M4.0b) and dispatches
`allocator.dealloc(state)` through the inline-protocol fn-ptr at
slot 2. Old behaviour was `free(state)` — went straight to libc,
ignoring whatever allocator the instance was constructed with.
After this commit, the per-instance allocator design from M1.2 A.5
is finally end-to-end correct:
push Context.{ allocator = arena } {
f := SxFoo.alloc(); ← arena.alloc(STATE_SIZE) + capture
// ... use f ...
}
// refcount → 0 ⇒ -dealloc:
// load state->__sx_allocator = arena
// arena.dealloc(state) ← same allocator round-trips
TrackingAllocator now sees the alloc/dealloc pair; the deferred M1.2
A.5 work is done. Closes the loop on M4.0.
The dealloc IMP passes `__sx_default_context` as the implicit __sx_ctx
when invoking the dealloc fn-ptr — the IMP itself has no caller-side
ctx (it's called by Apple's runtime at refcount-zero), and the
default GPA is the right baseline for any nested allocations the
dealloc body might perform.
Each compiler-internal lookup that "can't fail" (Context type,
__sx_default_context global) emits a loud diagnostic instead of
silent fall-through, per the silent-error budget.
184/184 example tests pass; chess on iOS-sim green.
Two converging paths now allocate the state struct via the protocol's
allocator instead of raw malloc:
(1) sx-side `Cls.alloc()`: compiler intercepts in `lowerObjcStaticCall`
when the receiver is a sx-defined `#objc_class` and the method is
the niladic `alloc`. Emits the inline alloc-and-init sequence
using the caller's `current_ctx_ref` as the context — so
`push Context.{ allocator = my_arena } { let f := SxFoo.alloc(); }`
honors `my_arena` end-to-end. The msgSend dispatch is bypassed
entirely for this case.
(2) Obj-C-runtime `[Cls alloc]` (Info.plist principal class, NSCoder,
UIKit reflection): the synthesized `+alloc` IMP shim reads
`__sx_default_context.allocator` and calls into the same shared
helper. The IMP has `has_implicit_ctx = false` and runs with no
caller-side context — the default GPA is the right policy choice
for "everything Apple's runtime instantiates".
Shared helper `emitObjcDefinedAllocAndInit(fcd, cls_ref, ctx_addr)`
does the work: `class_createInstance` → `ctx.allocator.alloc(STATE_SIZE)`
via the inline-protocol fn-ptr → memset 0 → store allocator at
state[0] (the M4.0a slot, captured for -dealloc's later use) →
`object_setIvar(instance, __sx_state_ivar, state)`. Loud failures
on missing globals via the diagnostics system.
The sx-side interception must explicitly bitcast the
`class_createInstance` result from `*void` to the method's declared
return type (`*<Cls>` or `?*<Cls>`). lowerVarDecl reads the Ref's IR
type when no type annotation is present, and coerceToType is a
no-op for ptr→ptr — without the bitcast, `let f := SxFoo.alloc();`
binds `f` at `*void` and downstream `f.class` / `f.method()` fails
to find anything.
-dealloc still uses `free(state)` (M4.0c rewrites it). 184/184 tests
pass; chess on iOS-sim green.
State struct for an sx-defined `#objc_class` now leads with an
Allocator field at index 0 — captured at +alloc time, read by
-dealloc to free the state through the same allocator. User fields
shift to index 1+; the existing by-name lookups in
emitObjcDefinedClassPropertyImps + lookupObjcDefinedStateFieldOnPointer
naturally resolve them at the new indices.
This step is the layout change only; the +alloc IMP still mallocs
(M4.0b will rewrite it to thread context.allocator through), and
-dealloc still uses free() (M4.0c). The field is allocated but
uninitialised; nobody reads it yet.
Storage type comes from `Context.fields[0].ty` via the new
`objcStateAllocatorType` helper — same Allocator value-shape the
implicit context machinery has used all along. If Context isn't
registered (early-init paths), the helper falls back to omitting
the field rather than synthesising a half-broken layout.
IR snapshot for 142-objc-class-method-lowering updated to reflect
the new struct shape and the +24-byte state allocation. Chess on
iOS-sim green; 184/184 example tests pass.
Three threads, one commit because they're entangled:
1. Helper free functions on `*UIKitPlatform` (refresh_safe_insets,
read_screen_scale, create_gl_context, setup_renderbuffer,
present_renderbuffer, compute_layer_pixel_size) → methods on the
`impl Platform for UIKitPlatform` block. IMP-shape trampolines
(`uikit_keyboard_will_change_frame`, `uikit_scene_will_connect[_ios]`,
`uikit_gl_view_tick/layout/touches_*`, `uikit_subscribe_keyboard_notifications`)
also collapse into methods on UIKitPlatform — the
`(self: *void, _cmd: *void, ...)` form is no longer needed since
M3 made the #objc_class trampolines compiler-synthesized. Class
method bodies in SxAppDelegate / SxSceneDelegate / SxGLView /
SxMetalView now read `if g_uikit_plat == null { return; }
g_uikit_plat.x();` — no more `xx self, xx 0` casts at every IMP
call site.
2. Declarative `layerClass` form. SxGLView and SxMetalView promote
from the M2.1(a) constant-with-runtime-string-lookup workaround
(`layerClass :: *void = objc_getClass("CAEAGLLayer".ptr);`) to
the class-method expression-body form
(`layerClass :: () => CAEAGLLayer.class();`). Type stays `*void`
until M1.1.b lands `Class(T)` parameterisation; the value side
already matches the plan. Backing this: foreign-class declarations
for CAEAGLLayer (extended with `class :: () -> *void;`) and a new
CAMetalLayer foreign-class declaration alongside it. Both
`#extends CALayer` so the dispatch chain knows about the parent.
3. Optional-shape idiom pass on uikit.sx. `xx`-as-optional-wrap on
field assignments (`plat.gl_ctx = xx ctx`, `plat.text_field = xx tf`,
`plat.display_link = xx link`) dropped — implicit `T → ?T` does
the right thing. `!` force-unwraps replaced with `if val := opt
{ ... }` safe-narrowing (the keyboard handler, the GL-context
read in setup/present renderbuffer, the gl_view read in scene
bootstrap). `orelse` (Zig keyword) that briefly snuck into the
keyboard handler removed in favour of the `if win := plat.window`
narrowing pattern. Result: no `xx` casts left on the implicit
T→?T path; all optional access goes through `if val :=`.
IR snapshots `ffi-objc-call-06-sret-return.ir` and
`ffi-objc-dsl-07-mangling-table.ir` refresh to pick up the new
`object_getIvar` / `object_setIvar` runtime-helper declarations
introduced when M1.2 A.3 made instance-method bodies route field
access through the state ivar.
Chess on iOS-sim green throughout. 184/184 example tests pass.
For UFCS dispatch on foreign-class receivers (`#foreign #objc_class`
aliases), `resolveCallParamTypes` was returning an empty slice — both
`resolveFuncByName(qualified)` and `fn_ast_map.get(qualified)` miss
for `#foreign` methods (they live in `foreign_class_map`, not the
regular fn maps). With `param_types` empty, the per-arg `target_type`
assignment in `lowerCall` was skipped, leaving `self.target_type` as
whatever it held on entry — usually the enclosing function's return
type. Inside a `-> BOOL` method, `xx ptr` then lowered with target
type `i8`: `ptrtoint ptr to i64` → `trunc i64 to i8`, sending the low
byte of the pointer through.
Symptom: chess on iOS-sim crashed in
`-[NSNotificationCenter addObserver:selector:name:object:]` with
`observer = 0xC0` (low byte of the SxAppDelegate receiver) when the
AppDelegate method's first param was renamed to anything other than
`self`. The original session diagnosed it as a `self`-vs-`this`
hardcoding in `lower.zig`, but those hardcoded `"self"` strings are
all on compiler-synthesized parameters (init scopes, JNI stubs,
property IMPs, dealloc IMPs) — not the user-facing #objc_class body
params. The bug was in arg-type resolution.
Fix walks `foreign_class_map` + `findForeignMethodInChain` to recover
the declared param types (skipping the implicit `*Self` for instance
methods). Regression test `examples/issue-0044.sx` exercises the
BOOL-return + foreign-class arg shape; pre-fix the receiver round-trip
prints WRONG, post-fix it prints ok.
The UIKitPlatform struct had a string of '*void = null; // UIWindow*'
fields — the type lived in a comment, every callsite had to 'xx'-cast
back to the real type. Migrated to the real foreign-class pointer
types now that M3 declared all the relevant '#objc_class' aliases:
window: ?*UIWindow
root_vc: ?*UIViewController
gl_view: ?*UIView (SxGLView OR SxMetalView — both extend UIView)
gl_layer: ?*CALayer (CAEAGLLayer OR CAMetalLayer)
gl_ctx: ?*EAGLContext
display_link: ?*CADisplayLink
Each field is wrapped in '?' since the platform may not have set
it yet (gl_ctx is null in metal mode, display_link is null before
the first frame, etc.).
SxSceneDelegate's window getter/setter now take/return '?*UIWindow'
instead of '*void' so calling code doesn't need an xx-cast.
Required fix in objcTypeEncodingFromSignature: '?T' (optional) was
bailing with 'type kind not yet supported'. Apple's runtime treats
nullability as 'pointer may be null' — the wire encoding is the
same as T. Recursive unwrap handles ?*UIView → '@', ?*CADisplayLink
→ '@', etc.
Chess on iOS-sim: board renders, full pipeline intact. 183 tests
+ zig build test green.
Three holdover free functions from the pre-M3 era were each
just two or three lines that forwarded to a global. With M3
finished, every call site is one #objc_class method body, so
the wrapper indirection earns nothing — inline them.
Deleted:
uikit_window_getter → body of SxSceneDelegate.window
uikit_window_setter → body of SxSceneDelegate.setWindow
uikit_did_finish_launching → body of
SxAppDelegate.application_didFinishLaunchingWithOptions
The bigger helpers (uikit_keyboard_will_change_frame,
uikit_scene_will_connect, uikit_gl_view_tick/_layout, the four
uikit_gl_view_touches_*) stay — their bodies are 30-80 lines
each, so wrapping them in a small forwarding method body inside
#objc_class is the cleaner factoring.
Chess on iOS-sim: board renders, full game state intact. 183
example tests + zig build test green.
Three slices in one commit since they're tightly coupled (the
M3.5 deletion only makes sense after M3.3 and M3.4):
M3.3 — SxGLView migrated to declarative '#objc_class("SxGLView")':
- '#extends UIView' for the view-hierarchy + responder chain.
- 'layerClass :: *void = objc_getClass("CAEAGLLayer".ptr);' uses
the M2.1(a) class-level constant form. Registered on the
metaclass; UIView's +layerClass override dispatches here so
EAGL gets the right backing layer.
- Six instance methods (sxTick, layoutSubviews, four touch
selectors) forward to existing legacy IMP free functions.
M3.4 — SxMetalView migrated, same shape as SxGLView; differs only
in the 'layerClass' constant returning CAMetalLayer instead of
CAEAGLLayer. The five shared IMPs (sxTick/layoutSubviews/4 touch
handlers) reach the same free functions — they already branch on
plat.gpu_mode for GL-specific renderbuffer code.
M3.5 — uikit_register_classes() and the two helper registration
functions are deleted outright. Every sx-defined Obj-C class in
this module now goes through the compiler's M1.2 / M2.1(a)
synthesis path at module init. The call site inside
UIKitPlatform.init is gone too — just a comment marking the
migration point.
Chess on iOS-sim: board renders, scene-delegate connection still
fires, GL/Metal layer setup intact, touch dispatch routes through
the synthesized IMP trampolines. 183 example tests + zig build
test green.
End of M3. The platform layer's Obj-C-runtime wiring is fully
declarative.
Remaining: M4 (autoreleasepool + ARC ops), M5 (closure↔block),
M6 (auto-import + production hardening). M1.1.b (Class(T)
parameterization + instancetype) is still deferred — none of
the migrated uikit code needed it.
Migrates SxSceneDelegate from the hand-rolled
objc_allocateClassPair + class_addMethod + class_addProtocol
sequence to the declarative form:
SxSceneDelegate :: #objc_class("SxSceneDelegate") {
#extends UIResponder;
#implements UISceneDelegate;
#implements UIWindowSceneDelegate;
scene_willConnectToSession_options :: (self, scene, session, options) { ... }
window :: (self) -> *void { ... }
setWindow :: (self, w) { ... }
}
emit_llvm now honors '#implements' in the class-pair init
constructor — for each #implements ProtocolAlias on the cache
entry's AST, emit before objc_registerClassPair:
proto = objc_getProtocol("ProtocolName")
class_addProtocol(cls, proto)
iOS checks 'class_conformsToProtocol' when instantiating scene
delegates; without the conformance the runtime silently rejects
the class and a default scene with no delegate gets created
instead. The protocol-getter returns null on dead-strip /
runtime mismatch (rare but possible) — the runtime treats
class_addProtocol(cls, null) as a no-op, so no explicit null
check needed.
Method bodies forward to the existing legacy free IMP functions
(uikit_scene_will_connect, uikit_window_getter,
uikit_window_setter) so we don't have to inline the scene-
connect setup logic (~80 lines).
uikit_register_classes is now tiny — just the two remaining
view-class helpers (M3.3 SxGLView + M3.4 SxMetalView). M3.5
deletes the function entirely once those land.
Chess on iOS-sim: board renders, scene delegate fires, touch
events route correctly. 183 example tests + zig build test
green.
Two coupled changes that unblock the uikit_register_classes
migration:
1) M1.2 A.3 — body's 'self' is the Obj-C id (opaque), NOT the
state struct. Matches Apple's ObjC semantics where 'self' IS
the object. Cocoa idiom 'xx self → id' works at runtime calls
(addObserver:, etc.); previously the trampoline replaced
'self' with the state-struct pointer, breaking any runtime
call that expected an id.
'*Self' substitution in resolveTypeWithBindings now points at
foreignClassStructType(fcd) — the opaque class stub — instead
of objcDefinedStateStructType(fcd).
'self.field' access on a sx-defined class instance field is
rewritten by lowerFieldAccess to go through the __sx_state
ivar:
state = object_getIvar(self, load(__<Cls>_state_ivar))
val = struct_gep(state, field_idx) → load
Both read (lowerFieldAccess) and write (lowerAssignment) take
this path. Compound ops (+=, -=, etc.) are supported via
storeOrCompound. The lookup is filtered: skip property fields
(those still go through the M2.2 msgSend getter/setter
dispatch) and foreign classes (no state).
New helpers in lower.zig:
- lookupObjcDefinedStateFieldOnPointer — match check.
- lowerObjcDefinedStateForObj — emit the object_getIvar +
ivar-global-load idiom (shared between read + write paths).
- lowerObjcDefinedStateFieldRead — the load path.
Also moved the @llvm.global_ctors registration out of the
sx-defined class-pair init constructor — global_ctors fires
DURING dyld's framework load, before UIKit registers its Obj-C
classes. objc_getClass("UIResponder") returned null, super
was null, objc_registerClassPair crashed. main's entry block
is post-framework-load but pre-user-code — exactly the right
window. New helper injectCtorIntoMain.
2) M3.1 — SxAppDelegate migrated to declarative #objc_class.
uikit_register_classes' hand-rolled objc_allocateClassPair +
class_addMethod for SxAppDelegate is gone; the compiler
synthesises the class at module init. The method bodies
forward to the existing legacy IMP free functions
(uikit_did_finish_launching, uikit_keyboard_will_change_frame)
so we don't have to inline 70+ lines of keyboard-frame logic
right now.
Also adds UIResponder foreign-class declaration and chains
UIView / UITextField to it via #extends UIResponder so the
methods that previously lived on UITextField directly
(becomeFirstResponder etc.) move to their proper home.
Chess on iOS-sim: board renders, full state intact. 183 example
tests + zig build test green.
When 'obj.method()' is called on a foreign-class pointer and the
method isn't declared on the receiver's class, the compiler walks
the '#extends' chain to find an ancestor that declared it.
Property lookup (M2.2) flows through the same chain walker.
ParentX :: #foreign #objc_class("...") { foo :: ... }
ChildX :: #foreign #objc_class("...") { #extends ParentX; }
child.foo() // now resolves — was 'no method foo on ChildX'
Two new helpers in lower.zig:
- findForeignMethodInChain(fcd, name) walks the cache via
fcd.members[i].extends → foreign_class_map[parent] → ...
Depth-capped at 16 to break accidental cycles.
- findForeignPropertyInChain(fcd, name) — same shape for fields.
ALSO fixes a latent class-hierarchy bug uncovered while testing
M2.3: emit_llvm was passing the sx alias name to
objc_allocateClassPair(super, ...) rather than the actual Obj-C
runtime class name. For 'SxThing :: #objc_class(...) { #extends
NSObjectBase; }' where 'NSObjectBase' is aliased to "NSObject",
emit_llvm produced 'objc_getClass("NSObjectBase")' → NULL →
'objc_allocateClassPair(NULL, ...)' → SxThing's super-class link
was broken → '[sx_thing hash]' bypassed NSObject and crashed in
the forwarding machinery.
Fix: ObjcDefinedClassEntry gains a 'parent_objc_name' field
pre-resolved by lower.zig's 'resolveObjcParentName' through
foreign_class_map (which has the alias → foreign_path mapping).
emit_llvm just reads the resolved name from the entry.
153-objc-extends-chain.sx exercises both fixes:
1-level: SxThing → NSObject — t.hash() walks one #extends.
2-level: SxLeaf → SxMiddle → NSObject — chained #extends.
Both return real NSObject.hash values from libobjc.
183 example tests pass (+1). zig build test green.
Properties on sx-defined #objc_class declarations now synthesize
getter (always) and setter (unless 'readonly') IMPs that GEP into
the hidden state struct and load / store the corresponding field.
The state struct already holds every user-declared field
(objcDefinedStateStructType), so no new layout work — the IMPs
just dispatch a struct_gep + load/store through the __sx_state
ivar.
For each '#property' field on a sx-defined class:
Getter '__<Cls>_<field>_imp(self, _cmd) -> T':
state = object_getIvar(self, load(__<Cls>_state_ivar))
return state.<field>
Setter '__<Cls>_set<Field>_imp(self, _cmd, val) -> void':
state = object_getIvar(self, load(__<Cls>_state_ivar))
state.<field> = val
Both IMPs land in the cache's methods slice (mirroring the
method-IMP wiring from M1.2 A.4b.iii) so emit_llvm's
class_addMethod loop registers them on the class without
special-casing. Selector mangling:
getter: <field> (e.g. 'width')
setter: set<Field>: (e.g. 'setWidth:')
Type encoding derived from the field's resolved IR TypeId.
'readonly' (the only modifier honored in this slice) skips the
setter emission AND the corresponding method entry — so the
runtime reports the selector as absent. Other modifiers
(strong, weak, copy, assign) parse fine but stay no-ops until
M4.2 wires up ARC ops in the setter body.
152-objc-property-sx-defined.sx round-trips on macOS:
b.width = 10; b.height = 7;
read back through getter IMPs.
area is readonly — class_getInstanceMethod(SxBox, sel(setArea:))
returns NULL, confirming the setter is absent.
182 example tests pass (+1). zig build test green.
Inside a '#objc_class { ... }' block, 'name :: Type = expr;' is
accepted alongside the existing method form. Parsed as sugar for
'name :: () -> Type => expr;' — a niladic class method with an
expression body. The synthesized class method flows through the
M2.1(b) class-method pipeline: a C-ABI IMP is emitted and
registered on the metaclass.
Apple's runtime sees zero distinction — '[Cls foo]' dispatches to
our IMP regardless of source spelling. The constant form is
purely syntactic sugar; it reads better for static metadata
returns:
SxGLView :: #objc_class("SxGLView") {
layerClass :: Class = CAEAGLLayer.class();
}
vs. the equivalent method form:
layerClass :: () -> Class => CAEAGLLayer.class();
Parser change: after 'name ::' if the next token isn't '(' we
take the constant branch — parse a type expr, expect '=', parse
the value expr, expect ';'. The result is a ForeignMethodDecl
with is_static=true, empty params, return_type=Type, body=block
wrapping the expr. Pure parser-level transformation; no new AST
nodes, no new lowering passes.
150-objc-class-level-constant.sx exercises both shapes on macOS:
a primitive (s32 answer) and a pointer ('*NSObject seedClass'
— the canonical '+layerClass'-style factory return).
180 example tests pass (+1). zig build test green.
M2.1 complete: both (a) the constant form and (b) the
expression-bodied class method shape land.
Next: M2.2 — 'field: T #property(modifiers...)' synthesizes
getter/setter pairs.
Bodied methods without a '*Self' first param (parser marks
is_static=true) are now registered as Obj-C CLASS methods on
the metaclass.
Each such method gets:
- A synthesized FnDecl + body lowering through the existing
M1.2 A.2 path.
- A C-ABI trampoline 'emitObjcDefinedClassStaticImp' — same
shape as the instance trampoline but skips the __sx_state
ivar read (no instance state) and passes only
'__sx_default_context' (plus user args) to the sx body.
- An entry in ObjcDefinedMethodEntry with 'is_class=true'.
emit_llvm's class-pair init constructor now computes the
metaclass once up-front (via object_getClass(cls)) and shares
it between the +alloc IMP registration (M1.2 A.5) and the
M2.1(b) class-method registrations. The per-method registration
loop picks the target via 'method.is_class ? metaclass : cls'.
149-objc-class-method-static-imp.sx end-to-end on macOS:
SxFoo :: #objc_class("SxFoo") {
answer :: () -> s32 { return 42; }
}
// [SxFoo answer] via objc_msgSend → 42
// class_getClassMethod(SxFoo, sel_answer) → non-null
Still TODO for M2.1: the (a) class-LEVEL constant form
'layerClass :: Class = CAEAGLLayer.class();' — needs parser
extension to recognize 'name :: Type = expr;' inside #objc_class
blocks, plus lazy-init-slot synthesis.
179 example tests pass (+1). zig build test green.
Adds a special case to lowerFieldAccess: when the field is
literally 'class' and the receiver is a pointer to an Obj-C
(or Obj-C protocol) foreign-class struct, emit
'object_getClass(obj)' instead of falling through to struct GEP.
Returns 'Class' (the M1.1 first-pass alias for *void;
parameterized Class(T) covariance is deferred to M1.1.b).
f := SxFoo.alloc();
cls := f.class; // → object_getClass(f)
cls == objc_getClass("SxFoo".ptr); // ok
New helper isObjcClassPointer(ty) detects 'ptr -> struct in
foreign_class_map under .objc_class / .objc_protocol'. The
check fires BEFORE the auto-deref so the runtime call sees the
opaque Obj-C pointer rather than the load'd struct stub.
148-objc-self-class-accessor.sx exercises both shapes end-to-end
against the macOS runtime: sx-defined class (SxFoo) and foreign
class (NSObject). Round-trips against objc_getClass(name).
178 example tests pass. zig build test green.
This effectively closes Month 1 — M1.0, M1.1 (first pass), M1.2,
M1.3 all done. Remaining: M1.1.b (Class(T) covariance +
instancetype), then Month 2 (declarative sugar).
Delete the bail at lower.zig:4407 that diagnosed sx-defined Obj-C
class dispatch as 'not yet supported'. Both foreign and
sx-defined '#objc_class' decls now flow through the same
'lowerObjcMethodCall' path — instance methods on sx-defined
classes dispatch via objc_msgSend, and the registered IMP
trampolines (M1.2 A.4b.iii) route to the sx bodies.
The runtime non-Obj-C branch (.swift_class / .swift_struct /
.swift_protocol) keeps its 'not yet supported' diagnostic;
M1.2 only addresses the Obj-C runtimes.
Constructor reorder in emit_llvm: emitObjcDefinedClassInit
runs BEFORE emitObjcClassInit. Otherwise the Phase 3.1
class-cache populator calls objc_getClass("SxFoo") before our
constructor registers the class — cache slot stored null and
'SxFoo.method()' dispatched against a null class pointer.
ffi-objc-defined-class-01-instance.sx (the integration test
from the plan) now runs the full lifecycle on macOS:
f := SxFoo.alloc() // synthesized +alloc IMP fires
f.bump() // dispatch → IMP trampoline → sx body
f.bump() // state persists across calls
f.bump()
f.get() // → 3
release_fn(f, sel_release) // synthesized -dealloc fires
The user declares 'alloc :: () -> *SxFoo;' bodyless to give the
synthesized +alloc IMP a typed contract at sx call sites —
same convention as foreign classes today.
M1.2 complete: A.0 A.1 A.2 A.3 A.4 A.4b.i A.4b.ii A.4b.iii
A.5 A.6 A.7. End-to-end class-synthesis foundation works.
177 example tests pass (+1 from the integration test). zig
build test green.
For every sx-defined #objc_class, emit a C-callconv -dealloc IMP
that runs at refcount-zero. Frees the sx state struct, nils the
ivar, then chains to [super dealloc] so NSObject's runtime
cleanup (object_dispose, associated-object teardown, KVO, etc.)
runs as usual.
-dealloc IMP (self: id, _cmd: SEL) -> void
state = object_getIvar(self, load @__<Cls>_state_ivar)
free(state) // free(NULL) is safe
object_setIvar(self, ivar, NULL)
sup = alloca { receiver: *void, super_class: *void }
sup.receiver = self
sup.super_class = load @__<Cls>_class
sel_dealloc = sel_registerName("dealloc")
objc_msgSendSuper2(&sup, sel_dealloc)
return
Two new per-class globals:
- '__<Cls>_class' : *void — populated by emit_llvm's
class-pair init constructor with the freshly-allocated Class
pointer (after objc_registerClassPair).
- The existing '__<Cls>_state_ivar' is also consulted to find
the state struct.
The -dealloc IMP is registered on the class itself (instance
method) via class_addMethod with encoding 'v@:'. emit_llvm
ALSO stores cls_val into '__<Cls>_class' so the trampoline
can build the objc_super struct.
internStringConstantGlobal helper added to lower.zig — interns
C strings as [N:0]u8 globals with byte-level aggregate inits.
Used here for the 'dealloc' selector string.
147-objc-class-dealloc-roundtrip.sx verifies end-to-end on
macOS: alloc + release fires the IMP, and a second alloc/release
cycle proves runtime state isn't corrupted. class_getMethod-
Implementation confirms the IMP is registered.
176 example tests pass (+1). zig build test green.
Still gated: sx-side 'obj.method()' calls bail at lower.zig:4407
with the existing diagnostic. A.7 opens the gate — last sub-step
of M1.2.
For every sx-defined #objc_class, emit a C-callconv +alloc IMP
that the Obj-C runtime calls when '[Cls alloc]' fires (from sx
code, UIKit instantiation, Info.plist principal class, etc.):
+alloc IMP (cls: Class, _cmd: SEL) -> id
instance = class_createInstance(cls, 0)
state = malloc(STATE_SIZE)
memset(state, 0, STATE_SIZE)
object_setIvar(instance, load(@__<Cls>_state_ivar), state)
return instance
STATE_SIZE = max(typeSizeBytes(state struct), 1) — always at
least one byte so the ivar is never null after +alloc returns.
The IMP is registered on the METACLASS (class methods live there
— every Class object's isa points to the metaclass) in emit_llvm's
class-pair init constructor:
metaclass = object_getClass(cls)
sel_alloc = sel_registerName("alloc")
class_addMethod(metaclass, sel_alloc, alloc_imp, "@@:")
That override wins over NSObject's default +alloc; runtime
instantiations get the __sx_state ivar bound automatically.
Per-instance allocator binding (the plan's full design — store
the Allocator value in the state struct so -dealloc frees through
the same one) is deferred. libc malloc/free is fine for v1; we'll
upgrade once Month 4's autoreleasepool + ARC ops shake out.
REFACTOR: collapsed five duplicate 'get<Name>Fid' helpers and
their cache fields (object_getIvar, object_setIvar,
class_createInstance, malloc, memset) into a single
'ensureCRuntimeDecl(name, params, ret) -> FuncId'. The helper
checks for an existing decl by name first (avoids the
'class_createInstance.1' duplicate-symbol crash when stdlib's
'#foreign' decl is already in the module). One helper instead
of one-per-function = ~150 lines deleted.
object_getIvar / object_setIvar added to stdlib std/objc.sx
so user code can use them too (146 exercises object_getIvar
to verify __sx_state was bound to a non-null state pointer
after +alloc).
146-objc-class-alloc-roundtrip.sx end-to-end against macOS:
'[SxFoo alloc]' returns non-null AND object_getIvar(instance,
__sx_state) returns the state ptr. Real Obj-C runtime, no
mocks.
175 example tests pass (+1). zig build test green.
For each instance method on a sx-defined '#objc_class', the
class-pair init constructor now:
sel = sel_registerName("selector_string")
imp = @__<Cls>_<method>_imp (M1.2 A.4b.ii)
class_addMethod(cls, sel, imp, "<encoding>")
before objc_registerClassPair. The IMP trampoline (A.4b.ii)
already bridges C-ABI -> sx body. With registration in place,
'objc_msgSend(obj, sel_bump)' now routes to the trampoline,
which reads __sx_state ivar and forwards to '@<Cls>.<method>'.
To get selector + type-encoding strings out of lower.zig and
into emit_llvm, ObjcDefinedClassEntry gains a 'methods' slice:
pub const ObjcDefinedMethodEntry = struct {
sel: []const u8, // mangled selector (M1.2 A.1's deriveObjcSelector)
encoding: []const u8, // type encoding (M1.2 A.1's objcTypeEncodingFromSignature)
imp_name: []const u8, // C-callconv trampoline symbol
};
registerObjcDefinedClassMethods populates this when it declares
each method's body function; Module.setObjcDefinedClassMethods
attaches the slice to the cache entry by name. Static (class-
side) methods are skipped — A.4b only covers instance methods;
class-method hooks like '+layerClass' land in M2.1.
emit_llvm reads entry.methods and emits class_addMethod inside
the per-class init block, before objc_registerClassPair (the
runtime locks the method list at register time on some SDK
versions).
145-objc-class-method-dispatch.sx verifies end-to-end:
class_getMethodImplementation(SxFoo, sel_registerName("bump"))
returns non-null after main starts. Both niladic ('bump') and
single-arg ('add:') selectors checked.
Still gated (A.7): sx-side 'obj.bump()' calls. The dispatch
gate at lower.zig:4407 hasn't opened — A.5 (+alloc) and A.6
(-dealloc) need to land first so the integration test
ffi-objc-defined-class-01-instance.sx (full state round-trip)
can exercise the full lifecycle.
174 example tests pass (+1 from 145). zig build test green.
For each bodied instance method on a sx-defined #objc_class,
emit a C-callconv trampoline function '__<Cls>_<method>_imp':
void __SxFoo_bump_imp(ptr obj, ptr _cmd, ...user_args) {
ivar = load @__SxFoo_state_ivar
state = object_getIvar(obj, ivar)
call @SxFoo.bump(__sx_default_context, state, ...user_args)
ret
}
The trampoline bridges the Obj-C runtime's IMP calling convention
('id self, SEL _cmd, ...args' as C ABI) to the sx body's
default-callconv shape ('__sx_ctx ptr, state ptr, ...user_args').
Implicit context comes from '&__sx_default_context'; the body
keeps its sx-side personality intact and can use 'self.field'
through the substituted state-struct pointer (M1.2 A.2b + A.3).
New helpers in lower.zig:
- 'getObjcObjectGetIvarFid' lazily declares object_getIvar.
- 'emitObjcDefinedClassImps' + 'emitObjcDefinedClassImp' walk the
cache and synthesise each trampoline.
- 'lookupGlobalIdByName' for finding the per-class ivar handle
global. Linear scan — same N-is-small rationale as the other
Obj-C caches.
Dead code at this commit: the trampolines exist in the module
but no class_addMethod call registers them with the runtime.
'objc_msgSend(obj, sel_bump)' would still fall through to the
parent class (NSObject 'doesNotRecognizeSelector:') today.
A.4b.iii wires up class_addMethod in emit_llvm's class-pair-init
constructor — that's when the trampolines come alive.
142's IR snapshot refreshed to show the trampoline.
173 example tests pass. zig build test green.
