102/102 regression tests pass (+ffi-objc-call-04-primitive-returns
with xfail snapshot capturing today's diagnostic).
Pinned scenario: `[NSObject class]` — `#objc_call(*void)(null, "class")`.
Should return a non-null Class pointer once the lowering supports
non-void returns. Today the Phase 1.3 restriction trips with:
#objc_call: only `void` return + (recv, selector) is lowered today;
non-void / arg-bearing arities land in later phase-1 steps
The next commit (1.6b) introduces an `objc_msg_send` IR opcode that
bundles (recv, sel, args, ret_ty) and emit_llvm builds a per-call-
site LLVM function type, sharing one declared `@objc_msgSend`
symbol across return-type variants. Five primitive returns
(*void / bool / s32 / s64 / f64) get folded in across 1.6b–c.
101/101 regression tests pass; the IR snapshot for the selector-
sharing test diff flips from four per-call `sel_registerName` calls
to two (one per unique selector) routed through a module-init
constructor — matching what clang emits for `@selector(...)`.
Hot-path cost collapses from a libobjc hashtable lookup per call to
a single load of a static `SEL*` slot:
Before (Phase 1.3):
%sel = call ptr @sel_registerName(<"init">)
call ptr @objc_msgSend(<recv>, %sel)
After (Phase 1.5):
%sel = load ptr, ptr @OBJC_SELECTOR_REFERENCES_init
call ptr @objc_msgSend(<recv>, %sel)
+ @OBJC_SELECTOR_REFERENCES_init = internal global ptr null
+ @OBJC_SELECTOR_REFERENCES_release = internal global ptr null
+ define internal void @__sx_objc_selector_init() {
+ %sel = call ptr @sel_registerName(ptr @OBJC_METH_VAR_NAME_)
+ store ptr %sel, ptr @OBJC_SELECTOR_REFERENCES_init
+ %sel1 = call ptr @sel_registerName(ptr @OBJC_METH_VAR_NAME_.2)
+ store ptr %sel1, ptr @OBJC_SELECTOR_REFERENCES_release
+ ret void
+ }
+ @llvm.global_ctors = appending global [1 x { i32, ptr, ptr }]
+ [{ ..., ptr @__sx_objc_selector_init, ptr null }]
Implementation:
module.zig | new `objc_selector_cache: ArrayList(ObjcSelectorEntry)`
with `lookupObjcSelector` / `appendObjcSelector`. List
(not hashmap) keeps emit order stable across builds so
the IR snapshot doesn't flicker on rehash.
lower.zig | `internObjcSelector(sel)` creates the slot on first
use, returns the same `GlobalId` on every subsequent
call to the same selector. lowerFfiIntrinsicCall now
emits `global_addr + load` for literal selectors.
Non-literal selectors keep the `sel_registerName`
fallback. Declaring `sel_registerName` lazily on
first intern so emit_llvm finds it for the
constructor body.
emit_llvm.zig | new `emitObjcSelectorInit` pass synthesizes a void
constructor that loops over the cache, calls
`sel_registerName` for each unique selector string,
stores the result in the slot. Constructor is
registered in `@llvm.global_ctors` with default
priority (65535) so dyld runs it before main.
The `@OBJC_METH_VAR_NAME_` private string globals and unnamed-addr
flag match clang's exact emission shape — picked up by the system
linker into the right Mach-O sections on macOS / iOS. Chess
Android + iOS-sim still build clean (no `#objc_call` in chess yet —
phase-3 migration will start exercising this).
run_examples.sh now supports an optional `tests/expected/<name>.ir`
sibling to `.txt`/`.exit`. When present, the runner also captures
`sx ir <file>` output, normalizes target-/host-specific noise
(module ID, target triple/datalayout, attribute groups, LLVM's
auto-suffixed %temp numbering), and diffs against the snapshot.
`--update` regenerates it alongside the runtime output.
Catches lowering changes that don't affect what the program prints
— exactly the shape Phase 1.5's selector interning will produce
(same runtime output, very different IR).
First snapshot: `ffi-objc-call-03-selector-sharing.ir`. Today the
test emits four `call ptr @sel_registerName(ptr @str.N)` lines for
its four call sites; after 1.5 we expect two static
`@OBJC_SELECTOR_REFERENCES_<sel>` globals + loads at each call
site. The diff between the two snapshots will be the visible
artifact of the optimization.
101/101 regression tests pass (+ffi-objc-call-03-selector-sharing).
Test exercises four call sites — three sharing "init" and one
"release" — to pin the multi-site / multi-selector lowering before
1.5 changes how SEL lookups are cached.
Runtime behavior: identical before and after 1.5 (all call sites
hit nil receivers; libobjc returns 0 for void). The improvement is
visible only in the emitted IR — today:
$ ./zig-out/bin/sx ir examples/ffi-objc-call-03-selector-sharing.sx \\
| grep -c "call ptr @sel_registerName"
4
After 1.5 (planned): 2 — one `sel_registerName` per unique selector
string, materialized into a static `OBJC_SELECTOR_REFERENCES_<sel>`
global at module init, then loaded at each call site. Matches the
shape clang produces for `@selector(...)`. Worth re-running the
above grep after 1.5 lands as a manual sanity check.
The IR-shape snapshot harness (auto-diff of `sx ir` output) is
deferred; for now we verify by eye.
100/100 regression tests pass; ffi-objc-call-02-void-return flips
from xfail (codegen rejection) to passing ("ok").
Lowering for `#objc_call(void)(recv, "selector:")` lands in
lower.zig as `lowerFfiIntrinsicCall`:
%sel = call ptr @sel_registerName(<"selector:">)
%call = call ptr @objc_msgSend(<recv>, %sel)
Two extern decls (`sel_registerName(*u8) -> *void` and
`objc_msgSend(*void, *void) -> *void`) are declared lazily and
cached on the Lowering instance via `objc_msg_send_fid` /
`sel_register_name_fid`, so multiple call sites share one
declaration each.
Phase 1.3 deliberately keeps scope tight: only `void` return + just
(recv, selector) arity is wired. Non-void returns + variadic arity
fall through with a diagnostic and are owned by subsequent phase-1
steps (1.6 primitive returns; 1.7..1.9 struct shapes; 1.10 multi-
keyword selectors).
Selector resolution is still per-call-site `sel_registerName` —
the planned 1.5 interning turns the per-call hashtable lookup into
a single static-global load. Chess Android + iOS-sim builds clean
— no regression on the existing typed-`objc_msgSend`-cast pattern.
100/100 regression tests pass (+ffi-objc-call-02-void-return xfail
snapshot).
The intrinsic with no `inline if false` guard reaches sema/codegen
and trips an "unresolved: 'unknown_expr'" — the FfiIntrinsicCall
AST node from Phase 1.1 has no lowering rules in lower.zig /
emit_llvm.zig yet.
nil receiver was chosen so the test doesn't need a real Obj-C
object graph: the runtime guarantees `[nil msg]` is a no-op with
zero result for void returns. macOS-gated via `inline if OS == .macos`
so the runner stays portable.
Next commit: emit_llvm.zig produces the per-call-site
%sel = call ptr @sel_registerName(ptr "init.0")
call void @objc_msgSend(ptr null, ptr %sel)
lowering. Snapshot flips to "ok". Selector interning (one shared
global per unique selector string) lands as a separate step (1.5).
99/99 regression tests pass (+ffi-jni-call-01-parse).
Locks in the same parse-surface contract for the JNI intrinsics
that ffi-objc-call-01-parse pins for the Obj-C side:
#jni_call(*void)(null, null, "getWindow", "()Landroid/view/Window;");
#jni_static_call(s32)(null, null, "max", "(II)I", 3, 7);
#jni_call(bool)(null, null, "isShown", "()Z");
All three lower through the shared `FfiIntrinsicCall` AST node
added in 1.1; only the kind tag distinguishes them. `inline if false`
keeps sema/codegen out of the picture until later phase-1 steps
wire those in.
98/98 regression tests pass; ffi-objc-call-01-parse flips from
parse-error xfail to passing.
Shape: `#<intrinsic>(ReturnT)(args...)`. The return-type generic
sits in the first parens, the actual call args in the second. All
three intrinsics share the same parse rule; only the kind tag and
the downstream lowering differ.
token.zig | three new hash_* tags
lexer.zig | matches the directive keywords with the same
isIdentContinue boundary check as the rest
ast.zig | FfiIntrinsicCall node with `kind`, `return_type`,
and `args` fields; FfiIntrinsicKind enum
parser.zig | parseFfiIntrinsicCall — same call-arg loop shape
as Call, with the leading return-type slot
sema.zig | analyzeNode + findNodeAtOffset arms walk the args
+ return-type child nodes
lsp/server.zig | classify the new tokens as ST.keyword
Codegen for the new intrinsic isn't wired yet — examples that
reach the body of a non-suppressed call would fail at lowering.
The current parse test uses `inline if false { ... }` to suppress
the dead branch, so sema/codegen don't see the node. Phase 1.3+
adds the lowering and the gate comes off.
Chess Android + iOS-sim builds clean — no regression on the
existing `objc_msgSend` cast pattern or the JNI helper.
98/98 regression tests pass (+ffi-objc-call-01-parse with xfail
snapshot capturing today's parse error).
Phase 1 of PLAN-FFI.md introduces three compiler intrinsics
(`#objc_call`, `#jni_call`, `#jni_static_call`) that lift the
ceremony off the existing typed-`objc_msgSend` and JNI dispatch
patterns. This is the first step of the cadence:
1.0 (this commit): test-add. Locks the current parse rejection.
1.1 (next): make-green. Parser accepts the new syntax;
this snapshot updates to whatever the next
pipeline stage produces (sema/codegen still
can't lower the intrinsic — that's later
phase-1 steps).
1.3+: codegen lands; the test eventually runs
cleanly against Foundation.
`inline if false` wraps the call site so the AST carries the node
but no codegen runs for it. Lets Phase 1.1's parse-only test pass
without dragging in the sema/codegen plumbing prematurely.
97/97 regression tests pass (94 expected updated; +issue-0037 from
the prior commit).
The companion `94-foreign-global-helper.sx` ALSO declares
`__stdinp : *void #foreign;`. Two sx files referencing the same
extern symbol must link cleanly — LLVM dedupes the named global at
the module level, and the C linker resolves both refs to the one
libSystem definition.
The full ergonomic story (helper computes the *same* address as the
main file's direct read) is blocked on issue-0037: lower.zig's
`address_of(global)` branch produces `undef` when the body is a
non-main function, even single-file. Once issue-0037 closes, fold
the helper's address back into an equality check here.
The cross-file link itself works today and is the lemma we're
locking in. This is also the closest thing today to the cross-file
extern-global ergonomic issue-0030 wants — `#foreign` already works
across files; the missing piece is sx-side `extern` decls for
sx-defined globals.
Repro found while writing PLAN-FFI step 0.10.
In a single file:
__stdinp : *void #foreign;
stdinp_addr :: () -> u64 { xx @__stdinp; }
main :: () -> s32 {
a : u64 = xx @__stdinp; // a = real symbol address
b := stdinp_addr(); // b = 0
...
}
The emitted IR for the helper is `ret i64 undef`, suggesting the
`address_of(identifier=__stdinp)` branch in lower.zig (~line 1719)
doesn't see `__stdinp` in `global_names` at the moment the helper's
body is being lowered — even though the same lookup succeeds inside
main's body in the same compilation unit.
Likely cause: lazy-body lowering ordering vs. the pass that
registers extern global decls into `global_names`. Worth verifying
which before fixing — could also be per-function scoping of the
map. Phase 1 of the FFI plan doesn't depend on this, so it stays
filed as an open issue and gets addressed when convenient (or when
sx-side `extern` cross-file globals from issue-0030 land and need
the same lookup to work everywhere).
`inline if OS == { case .macos: ... case .ios: ... else: ... }` is
already supported (see library/modules/platform/sdl3.sx:42 and
examples/38-build-config.sx:30). Cleaner than the chained
`inline if OS == .a; inline if OS == .b; ...` form the prior
commit used.
Same expected output — only the macOS arm survives codegen on the
host. Snapshot unchanged.
96/96 regression tests pass (+ffi-09-foreign-result-chain).
Opaque C-handle pattern that mirrors how real sx code threads
MTLBuffer*, AAssetManager*, file pointers, etc. through composite
sx values. C side has a trivial heap-int handle (`ffi_chain_make`
returning `void*`, `ffi_chain_bump` / `_peek` / `_dispose`). The sx
side exercises:
1. Chained calls — make -> bump -> bump -> peek; one handle
threaded through four FFI sites in sequence.
2. Struct field — `Counter { handle: *void; label: string; }`
hosts the handle; methods/accesses go through
`.handle` to feed back into C.
3. List(*void) — push N handles, iterate, peek each, iterate
again to bump each, iterate again to read
back. Catches any aliasing / lifetime breakage
when handles round-trip through the slice
backing of List.
95/95 regression tests pass (+ffi-08-foreign-in-method).
One trivial C helper (`ffi_method_helper`) called from each of the
major sx surface constructs that can host an FFI site:
1. struct method body Counter.next
2. protocol impl method body impl Doubler for Counter
3. closure value body make_adder's `closure(...)`
4. comptime-gated branch `inline if OS == .macos { ... }`
No new ABI shapes — the lowering route a `#foreign` call site takes
shouldn't depend on its enclosing construct, and the test pins that
lemma. A future lowering refactor that, say, breaks protocol-dispatch
fast-paths for FFI-calling impl methods will fail here directly
instead of being caught only by the chess Android regression.
The `inline if` branches for ios/linux compile down to nothing on
macOS, so only the macOS arm fires at runtime — useful smoke test
that the comptime gate works around FFI sites too.
vendors/ is a third-party namespace (stb_image, kb_text_shape, etc.);
test fixtures don't belong there. The .c/.h companion files for the
Phase-0 FFI baselines now sit alongside the .sx that drives them in
examples/, with matching basenames:
examples/ffi-01-primitives.{sx,c,h} <- was vendors/ffi_primitives/
examples/ffi-02-small-struct.{sx,c,h} <- was vendors/ffi_structs/
examples/ffi-03-large-struct.{sx,c,h} <- was vendors/ffi_large_struct/
examples/ffi-04-fp-struct.{sx,c,h} <- was vendors/ffi_fp_struct/
examples/ffi-05-string-args.{sx,c,h} <- was vendors/ffi_strings/
examples/ffi-06-callback.{sx,c,h} <- was vendors/ffi_callback/
examples/101-ffi-medium-struct.{sx,c} <- was vendors/ffi_medium_struct/
`#source` / `#include` paths in the .sx files become bare filenames
(no prefix) since imports.zig's base_dir resolution finds them
relative to the importing .sx file's directory.
`library/vendors/sx_ffi_resolve_test/` stays put — that one's the
whole point: regression coverage for the stdlib-search branch of
the resolution chain, so it must live where ONLY that branch can
find it.
94/94 regression tests pass.
94/94 regression tests pass (+ffi-07-c-import-block).
Companion C helper lives only at
`library/vendors/sx_ffi_resolve_test/`. Critically NOT in
`sx/vendors/` (the sx repo root) and NOT in the importing
example's directory — so the `vendors/...` paths in this
example are findable solely via the stdlib search branch
(`<exe>/../../library`, `<exe>/../library`, `<exe>/library`).
That branch is the one the JNI insets bridge needs to reach
`library/vendors/sx_android_jni/sx_android_jni.c` without
forcing chess (or any consumer) to vendor an identically-named
copy. The test pins the resolution end-to-end:
- #include resolves; clang parses the .h; c_import.zig
synthesizes #foreign fn decls for `sx_ffi_resolve_test_add` /
`_mul`.
- #source resolves; the .c is compiled into the build's
object list.
- sx calls the synthesized decls and prints results.
Latent bug from the stdlib-path resolution introduced in 4849cfb.
The earlier shape captured `const ci = decl.data.c_import_decl;`
BEFORE mutating `decl.data.c_import_decl.{sources,includes}` with
the resolved paths, then passed the stale `ci.includes` to
`c_import.processCImport`. Result: `#include "vendors/..."` paths
that resolved via the stdlib branch (i.e. only existed under
sx/library/vendors/) reached clang as the original unresolved
string and failed to parse — silently producing no synthesized
`#foreign` decls.
`#source` survived because the source list is re-read from
decl.data later (collectCImportSources walks the AST), so it
picked up the mutated value. Only `#include`'s synthesis path was
broken.
Fix: do the resolution first inside its own scope, then re-bind
`ci` from `decl.data.c_import_decl` so the include list passed to
processCImport sees the resolved paths.
Caught by ffi-07 baseline (next commit) — the test deliberately
puts its C helper only under library/vendors/ so the path is
findable solely via the stdlib chain.
93/93 regression tests pass (+ffi-06-callback).
Mirrors the `app->onInputEvent` install pattern from
library/modules/platform/android.sx:
1. (s32) -> s32 — single primitive arg/return
2. (*void, s32) -> s32 — opaque ctx pointer + value
(the onInputEvent shape)
Side effects via two file-level globals so the test observes both
the return value AND state mutation across multiple calls:
- g_callback_hits = N proves the callback fired N times.
- g_callback_sum = sum of args proves each individual call landed
with the correct value.
The ctx-pointer variant casts `*void` back to `*s32` inside the
callback and reads through it (`p.*`), proving the pointer survives
the round-trip with no aliasing weirdness.
92/92 regression tests pass (+ffi-05-string-args).
Covers the four shapes that actually appear at the sx ↔ C boundary
today:
1. [:0]u8 string literal -> const char* (ffi_strlen, ffi_first_byte)
2. sx `string` value via .ptr (slice-decay branch in
coerceArg pulls the pointer)
3. [*]u8 raw buffer + length (ffi_sum_bytes, mutated via
ffi_write_byte and read back)
4. C-returned const char* (round-trips back as [*]u8)
The mutate-via-C path catches any pointer-aliasing regression — sx
allocates the fixed array `bytes : [4]u8`, passes `.ptr` to C which
writes index 1, and the sx side reads `bytes[1]` to confirm the
mutation took effect through the same memory.
91/91 regression tests pass (+ffi-04-fp-struct).
Single-file regression net for the all-float / all-double aggregate
ABI path:
FQuad — 16 B, 4×f32 (same slot as ffi-02's Vec4f)
DQuad — 32 B, 4×f64 (UIEdgeInsets-shape — the f32-vs-f64 landmine)
Already nominally covered by ffi-02's Vec4f, but pinning it as a
focused single-file test means a future ABI rule change that breaks
the HFA path fails *this* test directly without a noisy drag-in from
the multi-shape baseline.
DQuad at 32 B straddles the AAPCS64 HFA limit (≤4 floats of same
type, total ≤64 B); it stays as a struct value passed through
v0..v3 rather than going indirect. The snapshot confirms the values
arrive intact.
90/90 regression tests pass (+ffi-03-large-struct).
vendors/ffi_large_struct/{.h,.c} defines:
Big24 — 24 B, three s64 (byval params + sret return)
Big48 — 48 B, six s64 (same path, larger)
`make / rotate-or-reverse / sum` helpers per shape. The sx-side
example imports via `#source` only and declares matching structs +
hand-written #foreign decls.
Snapshot pins today's >16-byte aggregate ABI now that the
emit_llvm.zig sret-return transform is in place (previous commit).
That gives us a regression net for all four C-ABI aggregate slots
in one place:
≤8 B int — i64 coercion (ffi-01 vec-likes)
9..16 B int — [2 x i64] coercion (ffi-02 Pair64/Quad32, 101)
16 B HFA — struct, no coercion (ffi-02 Vec2/Vec4f)
>16 B — byval params + sret (this commit)
Foreign functions that return a >16-byte non-HFA aggregate (e.g.
Big24 / UIEdgeInsets on iOS / clang-shaped struct returns) need the
indirect-return ABI: caller allocates space, passes its pointer as a
hidden first arg with `sret(<T>)`, callee writes through it and
returns void. AAPCS64 puts the pointer in x8; SysV AMD64 puts it in
the first int register and treats the named return as void.
The existing >16-byte branch in `abiCoerceParamType` was returning
`ptr` for BOTH params and returns. That works for byval params (the
established pattern — caller alloca + store + pass ptr, callee loads
in prologue), but is wrong for returns: it caused the function decl
to look like `ptr @fn(...)` rather than `void @fn(ptr sret(<T>), ...)`,
and the call site read whatever happened to be in x0 as a struct
pointer — segfault on dereference (caught while writing the ffi-03
baseline).
Fix layered into the same `abiCoerceParamType` / call-site code path:
emitFunctionDecl:
- Compute `uses_sret = needs_c_abi && needsByval(ret_ty, raw_ret_ty)`.
- Ret type collapses to void.
- Prepend a `ptr` param at slot 0.
- Add `sret(<RetType>)` type attribute on param-index 1
(LLVMAttributeIndex 1 = first parameter; 0 = return value).
.call lowering:
- Detect callee_uses_sret via the same predicate.
- Allocate the result on the caller's stack (`sret.slot`).
- Prepend it as args[0] (with sret_off index alignment so the
original sx args land at args[1..]).
- After LLVMBuildCall2, set the same `sret(<T>)` attribute on
the call site's arg 1 (mirrors the fn-decl attribute — both
land in the AArch64 backend's lowering pass).
- Load the result from the slot to produce the IR value.
`call_indirect` (function-pointer dispatch — uikit.sx's typed
`objc_msgSend` casts) keeps its existing behavior for now; the iOS
path already round-trips UIEdgeInsets via that route. Folding the
same sret transform into call_indirect is a follow-up.
89/89 regression tests still pass. Chess Android + iOS-sim both
build clean.
Now that emit_llvm.zig bridges the struct<->[2 x i64] ABI mismatch
(previous commit), the 9..16-byte integer-only shapes round-trip
cleanly. Extended `examples/ffi-02-small-struct.sx` to cover all
four aggregate ABI slots in one place:
Vec2 — 8 B, two f32 (register pair, float)
Vec4f — 16 B, four f32 (HFA — homogeneous float aggregate)
Pair64 — 16 B, two s64 (9..16 B int — [2 x i64] coercion slot)
Quad32 — 16 B, four s32 (same slot as Pair64)
Vendor helpers (`vendors/ffi_structs/{ffi_structs.h,ffi_structs.c}`)
grow `ffi_pair64_*` + `ffi_quad32_*` companions. Snapshot updated
to capture the full output. 89/89 regression tests pass.
`examples/101-ffi-medium-struct.sx` keeps a minimal focused repro
of the Pair64 case so the issue's emergence-and-fix history stays
greppable.
Resolves issue-0036 (LLVM verifier failure on 16-byte integer-only
struct by value through #foreign). The mismatch:
Call parameter type does not match function signature!
%load = load { i64, i64 }, ptr %alloca, align 8
[2 x i64] %call = call [2 x i64] @fn({ i64, i64 } %load)
`abiCoerceParamType` had already chosen `[2 x i64]` for 9..16-byte
non-HFA structs (the AAPCS64 / SysV AMD64 register-pair ABI slot for
that size class) on the foreign-decl side, but `coerceArg` only knew
how to bridge struct<->integer (the ≤8 B case) — not struct<->array.
LLVM's verifier rejects type-mismatched call args, so the call site
never landed.
Added the symmetric branches in coerceArg:
- Struct -> Array : alloca <array>; store <struct>; load <array>
- Array -> Struct : alloca <array>; store <array>; load <struct>
Both use the LLVM opaque-pointer memory-bitcast pattern already in
place for the integer case. They're paired with the existing
i64 <-> small-struct bridge so all four (≤8 B int, 9..16 B int,
16 B HFA, >16 B byval) ABI slots round-trip cleanly through
emit_llvm now.
File mechanics: promotes the issue-0036 repro to a focused feature
example per CLAUDE.md's issue-resolution workflow:
examples/issue-0036.sx -> examples/101-ffi-medium-struct.sx
tests/expected/issue-0036.{txt,exit} -> tests/expected/101-ffi-medium-struct.{txt,exit}
vendors/issue_0036/issue_0036.c -> vendors/ffi_medium_struct/ffi_medium_struct.c
Snapshot updated to the passing output. 89/89 regression tests pass;
chess Android build still clean.
Surfaced while writing the ffi-02-small-struct.sx baseline. The sx
#foreign decl lowers `{ s64, s64 }` (and other 16-byte integer-only
shapes like `{ s32, s32, s32, s32 }`) to `[2 x i64]` for the small-
struct register-pair ABI on AAPCS64 / SysV AMD64, but the call site
loads the struct as `{ i64, i64 }`. The two types must agree for the
LLVM verifier to accept the call:
Call parameter type does not match function signature!
%load = load { i64, i64 }, ptr %alloca, align 8
[2 x i64] %call = call [2 x i64] @issue0036_swap({ i64, i64 } %load)
Float-only 16-byte aggregates (e.g. Vec4f) work because they route
through the HFA path which keeps the struct representation. See
examples/ffi-02-small-struct.sx for the working cases.
Phase 1's #foreign lowering rework is the natural place to unify
these representations; check there before fixing inline.
88/88 regression tests pass (+ffi-02-small-struct).
vendors/ffi_structs/ defines:
Vec2 — 8 B, two f32 — register-pair (float) ABI
Vec4f — 16 B, four f32 — homogeneous float aggregate (HFA) on AAPCS64
Both pass cleanly today: the sx-side struct declarations match the C
ABI for these float-only shapes, and the call-site / foreign-decl
type representations agree.
`#source` only (no `#include`) — c_import's type mapping rewrites
struct-typed params/returns to *void, which would link but pass
through the wrong ABI silently. The hand-written #foreign decls keep
sx's struct types end to end.
16-byte integer-only shapes (`{s64, s64}`, `{s32, s32, s32, s32}`)
discovered to trip the LLVM verifier (`[2 x i64]` vs `{ i64, i64 }`
mismatch between foreign decl and call site). Excluded from this
baseline; filed separately in the next commit as issue-0036.
87/87 regression tests pass (was 86; +ffi-01-primitives).
vendors/ffi_primitives/{.h,.c} exposes a trivial identity roundtrip
per primitive C type — int/uint/short/ushort/long long/unsigned long
long/signed char/unsigned char/float/double/void* — plus two-arg
add helpers (int + double) for multi-arg ABI exercise. The sx-side
example imports the .h via `#import c { #include / #source }` and
prints each result; the snapshot in tests/expected pins today's
parameter + return ABI so Phase 1's #objc_call / #jni_call lowering
work can't silently regress primitive marshalling.
Two findings logged in current/CHECKPOINT-FFI.md's Known issues
section (current behavior, not new bugs): (1) c_import.zig maps
`signed char` -> `u8` not `s8`, and (2) sx integer-literal parser
rejects values >= 2^63 as overflow even when the receiver is u64.
Both worked around in this test without blocking the baseline.
First step of the FFI ceremony reduction plan (current/PLAN-FFI.md).
Iterates a (target, example) tuple list, runs `sx build --target <t>
<example>`, asserts exit 0 + output file produced. Cross-compile
correctness only — these examples can't run on the host.
Initial tuple list is empty, so the script exits 0 on a clean tree
and contributors without the iOS SDK / Android NDK aren't blocked.
`toolchain_available` short-circuits with a SKIP line when the
requested toolchain isn't installed. Phase 1/2/3 cross-only examples
populate TUPLES as they land.
Four Android UX wins landing together; all verified end-to-end on a
Pixel 7 Pro (board fills width, info-panel text renders, status bar
inset honored, tap-to-select + tap-to-move plays 1. e4).
- AndroidPlatform.init reads density via AConfiguration_getDensity
(app->config at offset 32) and sets dpi_scale = density / 160. The
hardcoded 1.0 had been making every logical unit equal one physical
pixel; ChessBoardView's 520-default size_that_fits fallback then
rendered at ~half the framebuffer width on the device, and glyphs
rasterized at literal 11-13 physical pixels were essentially invisible
on a 2340-tall display.
- gles3.sx set_scissor un-stubbed; with dpi_scale right the renderer
feeds in valid pixel bounds and the Y-flip math lands inside the
framebuffer.
- New library/vendors/sx_android_jni/sx_android_jni.c walks
activity -> window -> decorView -> rootWindowInsets via JNI and
publishes the system-bar insets. safe_insets() lazy-queries the
first call after EGL is up (decor view isn't attached at bootstrap).
- sx_android_install_input_handler sets app->onInputEvent; sx-side
sx_android_input_event translates AMotionEvent DOWN/MOVE/UP/CANCEL
into existing mouse_down/mouse_moved/mouse_up Events so the chess
board's tap-to-select + ScrollView drag path Just Works. Coordinates
divided by dpi_scale so layout-side hit tests match. poll_events
drains its slice after returning (mirrors the SDL pattern).
- src/imports.zig now routes #import c { #source / #include } paths
through the same chain as #import (importing dir -> CWD -> stdlib
search paths). Lets library-owned C helpers like the JNI bridge
live in sx/library/vendors/ without forcing consumers to vendor a
copy. Existing CWD-relative consumer layouts (chess's vendors/...)
still resolve first, so no regression.
86/86 regression tests pass.
platform/android.sx: `sx_android_bootstrap(app)` now also reads the
ANativeActivity's `assetManager` (offset 64) and `internalDataPath`
(offset 32) into module globals so consumers can route file I/O
through the APK's bundled `assets/` tree.
target.zig (`createApk`): also zips the project's `./assets/`
directory into the APK alongside `lib/<arch>/`. Resolves relative
to the user's CWD at invoke time — matches the convention chess
uses (assets/ next to main.sx).
gles3.sx: scissor is currently a no-op on Android. The renderer's
ScrollView clip_push path feeds bounds that land outside the
framebuffer (clipping everything off-screen). With scissor disabled
the chess board + pieces render correctly. TODO recorded in the
file to fix the bounds path properly.
Mirror of metal.sx, talks to GLES3 via opengl.sx's runtime-loaded
fn-pointer variables. EGL bootstrap is owned by AndroidPlatform; this
module just calls `load_gl(@eglGetProcAddress)` once during `init` to
populate the pointers, then drives raw draw/state from there.
The renderer's vertex layout (12 floats: pos2/uv2/color4/params4 = 48
bytes, attribute locations 0-3) is hardcoded in a single shared VAO
the Gles3Gpu owns — `set_vertex_buffer` rebinds the active VBO against
it. `set_vertex_constants(slot=1, data, 64)` is treated as the 4x4
projection matrix; `set_texture(slot=0, ...)` binds texture unit 0 and
sets `uniform sampler2D uTex` — both match renderer.sx's shader
contract.
A subtle gotcha caught + recorded in the file header: declaring the
same GL name as a `#foreign` function while opengl.sx also declares it
as an fn-pointer global silently lets the global win, and calling
through the uninitialized variable jumps to PC=0. Solution: don't
re-declare; use opengl.sx's pointers and `load_gl` them.
renderer.sx: the GPU-protocol shader-source branch now passes
(UI_VERT_SRC_ES, UI_FRAG_SRC_ES) on Android (separate vert+frag) vs.
the combined MSL library on iOS. Both gated with `inline if OS == X`.
C imports (stb_image, stb_truetype) compiled via the embedded LLVM
clang library now resolve bionic headers on Android. main.zig auto-
fills target_config.sysroot with the NDK root (mirroring the iOS path
that auto-fills the iOS SDK path); c_import.zig derives the bionic
sysroot inside it and passes `--sysroot <ndk>/toolchains/llvm/prebuilt/<host>/sysroot`
to the embedded clang.
Android link branch in target.zig stops auto-appending entries from the
collected `#library` / `#framework` lists. Most `#library` directives
in the stdlib (`objc.sx`'s `objc :: #library "objc";`, frameworks set
by uikit.sx, etc.) describe Apple-specific intent that's nonsensical on
Android. Users opt into Android-side libs via `opts.add_link_flag(...)`
in build.sx — same shape as how the iOS branch already lists frameworks
in chess's configure_build.
Verified end-to-end: chess game compiles for Android, packages into a
debug-signed APK, installs and launches on Pixel 7 Pro. It crashes in
UIRenderer.init because raw GL calls run before EGL is up — same
architectural gap iOS bridges via the Metal GPU protocol. Next step
is a GLES3 GPU impl or a lazy-init in UIRenderer. 86/86 regression
tests + iOS-sim chess build clean.
User writes BOTH `main` and a 3-line `android_main(app)` trampoline.
The library provides `sx_android_bootstrap(app)` (stashes the NDK app
pointer into a platform-owned global) and `AndroidPlatform` impl of
the Platform protocol. The library NEVER references `main` — the OS-
shape entry symbol lives in user code where the other entry symbols
already live. iOS / SDL3 keep their existing shape; only Android adds
the trampoline.
Cross-cutting bits this commit ships:
library/modules/compiler.sx
Add `android` variant to `OperatingSystem`.
src/ir/lower.zig
- injectComptimeConstants: map TargetConfig.isAndroid() → .android.
- New Pass 4 `checkRequiredEntryPoints`: emit a clean diagnostic
when `--target android` is requested but `android_main` isn't
defined, instead of letting the user crash on a dlopen-time
missing-symbol error.
library/modules/platform/android.sx
AndroidPlatform impl of the Platform protocol — EGL bringup on
`APP_CMD_INIT_WINDOW`, ALooper(0) polling, dispatches the user's
frame closure each ~16 ms tick. `sx_android_bootstrap(app)` is the
only function exposed for the entry trampoline.
examples/99-android-egl-clear.sx
Rewritten to use the new pattern: minimum `main` + `android_main`
pair, AndroidPlatform-driven render loop. Doubles as the usage
reference users hand off to the compiler diagnostic.
Verified on Pixel 7 Pro: purple clear-color frame, periodic
`rendered 60 frames` logcat lines. iOS-sim chess + 86/86 regression
tests pass.
When lowering `self.cb()` from inside a method whose receiver is *Self,
the field-access path passed the receiver pointer (not the aggregate)
to `structGet`, which then produced `call void undef(ptr undef)` at
the LLVM level — undefined at runtime, corrupted adjacent globals when
it transferred control to a garbage pointer. Auto-load through the
pointer first so structGet receives a real aggregate.
Discovered while building the new AndroidPlatform's `run_frame_loop` —
calling the stored frame closure as `self.frame_closure()` zeroed
out adjacent globals because the undef call jumped into random memory.
Added examples/100-closure-field-call-via-self-ptr.sx as the locked-in
regression: both direct (`self.cb()`) and hoisted (`fn := self.cb; fn();`)
forms must yield identical IR + behavior. 86/86 regression tests pass.
Verified on Pixel 7 Pro: solid purple frame, 'rendered 60 frames'
logcat line every second. End-to-end exercise of the new
sx-build → libsxhello.so → APK toolchain shipped today: NDK clang
link + native_app_glue bundling + aapt2/zipalign/apksigner pipeline +
isExportedEntryName so android_main lands in .dynsym.
Notes the source captures so future Android work doesn't repeat the
debugging:
- android_app field offsets for arm64 NDK 29 (window @ 72,
destroyRequested @ 100, source process fn-ptr @ 16).
- ALooper_pollOnce(-1, ...) blows the stack inside Looper::pollOnce
on this device/OS combo; ALooper_pollOnce(0, ...) is fine. We
drive the event loop non-blocking and sleep 16ms.
Outside the regression set on purpose (no tests/expected/99-*.txt) —
same convention as 63-metal-clear.sx. Build instructions live in the
file's leading comment.
Android (toolchain):
--target android / --target android-arm64 → aarch64-linux-android21.
target.zig discovers $ANDROID_NDK_HOME (or scans
~/Library/Android/sdk/ndk/* for the newest), invokes the NDK clang
with -shared -fPIC and links libsxhello.so against -llog -landroid
-lEGL -lGLESv3 -lm -ldl. native_app_glue.c from the NDK is compiled
and linked alongside the sx .o so apps can use the conventional
android_main(struct android_app*) shape; -u ANativeActivity_onCreate
keeps glue's symbol live.
Android (APK):
--apk <out> wraps the .so into a debug-signed installable APK.
target.zig discovers the SDK at $ANDROID_HOME (or
~/Library/Android/sdk), picks the newest build-tools + platforms,
generates a NativeActivity AndroidManifest.xml from --bundle-id,
packages via aapt2 link, appends the lib/ tree, zipalign, then
apksigner against ~/.android/debug.keystore (auto-generated via
keytool on first use). One command end-to-end:
sx build --target android --apk out.apk \\
--bundle-id co.swipelab.foo main.sx
Verified on Pixel 7 Pro: install + launch reaches android_main.
Compiler (entry-point linkage):
Top-level fn defs default to LLVM internal linkage and are lazily
lowered (only `main` was eagerly lowered before). Added
isExportedEntryName() — a small allowlist for names the OS loader
calls: `main`, `android_main`, `ANativeActivity_onCreate`,
`JNI_OnLoad`. These get eagerly lowered AND keep external linkage,
so they actually land in .dynsym.
LSP (imports):
DocumentStore now takes the install-discovered stdlib_paths and
forwards them into resolveImportPath, mirroring the compiler. Before
this, every `#import "modules/..."` resolved through the stdlib path
failed silently inside the LSP and identifiers from those modules
showed as `undefined variable`. Repro on label.sx: 1 false positive
before, 0 after.
Three new method signatures on the GPU protocol. Metal backend sends
`release` to the MTLTexture/Buffer/RenderPipelineState and nulls the
slot in its backing List so the handle becomes inert; handles are not
re-used. glyph_cache.grow() now destroys the old atlas before
allocating its replacement, eliminating the per-grow leak the file's
comment had been flagging since Session 62.
Protocol structs registered via registerProtocolDecl carry a new
is_protocol flag; the ?T paths in sizeOf/typeSizeBytes/toLLVMType
recognise it and lay out ?Protocol as the protocol struct itself
(ctx == null IS the "none" state), matching how ?Closure / ?*T are
sentinel-shaped — no extra storage.
Method dispatch on ?Protocol auto-unwraps in lowerCall's field-access
path; the unwrap is structurally a no-op so we just rebind obj_ty to
the payload type. resolveCallParamTypes extended for optional-protocol
receivers so enum-literal args (gpu.create_texture(.r8, ...)) get the
right target_type and don't silently collapse to tag=0 : s32 — same
issue-0031-class bug closed in Session 66, one type-system layer
deeper.
Library: UIRenderer / UIPipeline / GlyphCache migrated from the verbose
gpu: GPU = ---; has_gpu: bool pattern to gpu: ?GPU = null. set_gpu no
longer maintains a parallel bool flag.
Bundled: dock.sx threads delta_time as a struct field rather than via
a global pointer (cleanup unrelated to issue-0028, committed alongside).
Verified: 85/85 regression tests pass; iOS-sim chess + macOS chess
both render correctly post-migration.
Two small cleanups in the Metal text path on top of the buffer-offset
fix from cc71d95:
- Drop the SDF-style `smoothstep(0.5 ± ew, alpha)` from the text mode
branch in UI_MSL_SRC. The glyph atlas stores alpha coverage straight
from stbtt_MakeGlyphBitmap, not signed distance, so the smoothstep
was thinning anti-aliased strokes by mapping mid-coverage values
(0.3–0.7) toward 0/1. Use the sampled value directly as alpha.
- Drop the 16-byte alignment pad on `mtl_buf_offset` in `flush()`. Each
batch's upload_size is already a multiple of UI_VERTEX_BYTES (48), so
the running offset stays vertex-aligned without the extra rounding.
- After `font.shape_text` + `font.flush` in `render_text`, re-bind
`font.texture_id`. If the atlas grew during shaping, the GPU texture
handle changed; without this rebind the next flush samples the old
(smaller) atlas which doesn't have the newly-rasterized glyphs.
- Use explicit s64-pointer arithmetic in `metal_update_buffer_at_ios`
so a future regression in `[*]u8` indexing can't quietly miscompile
the per-flush write offset.
Text at small sizes still renders dim on dark backgrounds — most glyph
pixels sit in 0.1–0.5 coverage and the linear blend doesn't push them
to bright values — tracked separately as the faint-text follow-up.
UIRenderer.flush wrote to mtl_vbuf at byte offset 0 on every flush.
Metal records draw commands but reads the buffer at GPU execution time,
so a frame with multiple flushes ended up rendering whatever the LAST
writer left in the buffer for every draw. Chess UI hit this hard:
each of the 32 pieces in the initial position triggers two bind_texture
flushes (atlas -> pieces -> atlas), so ~64 mid-frame flushes silently
rendered the final info-panel batch over the board and the sprites.
New GPU protocol method update_buffer_at(buf, data, size, byte_offset);
Metal impl writes at offset via [*]u8 arithmetic on [buf contents].
UIRenderer tracks mtl_buf_offset (reset in begin, advanced per flush,
aligned to 16B, wraps on overflow) and draws each batch with
vertex_off = byte_off / UI_VERTEX_BYTES. Metal buffer over-allocated
4x the per-flush max (~3 MB) for headroom. GL path untouched —
glBufferData already orphans the storage.
71/71 regression tests pass. Metal-clear example, macOS GL chess, and
WASM chess all still build.
Four root causes for "chess UI shows white screen" — all fixed:
1. Hybrid legacy-app + scene-API path on iOS 26. Without
UIApplicationSceneManifest in the Info.plist, iOS 26 booted us in
[rb-legacy] mode and -[UIApplication connectedScenes] returned an
empty set. didFinishLaunching's window-setup code bailed at "no scene"
and the UIWindow never appeared on screen. Fix: emit the manifest in
buildInfoPlist (src/target.zig) AND split the window/view/layer setup
from didFinishLaunching into a new SxSceneDelegate's
scene:willConnectToSession:options: IMP. didFinishLaunching now just
subscribes the keyboard observer and returns YES.
2. UISceneDelegate formal protocol conformance. iOS 26 checks
-[cls conformsToProtocol:@protocol(UISceneDelegate)] before
instantiating the scene delegate; without it the runtime logs
"SxSceneDelegate does not conform to the UISceneDelegate protocol"
and silently uses a default delegate that does nothing. Fix:
look up UISceneDelegate + UIWindowSceneDelegate via objc_getProtocol
and class_addProtocol BEFORE objc_registerClassPair. The protocol
metadata is present at link time (unlike UIApplicationDelegate per
the long-standing legacy note in CHECKPOINT).
3. Protocol method return types via type aliases lowered as void.
The GPU protocol declares `create_shader(...) -> ShaderHandle` where
`ShaderHandle :: u32`. The protocol-decl lowering at lower.zig:7547
passed the return AST node through type_bridge.resolveAstType which
doesn't know about the type_alias_map. resolveTypeName fell through
to its "assume named struct" branch and registered ShaderHandle as
an empty struct ({ }). LLVM IR for the protocol call_indirect then
read `call {} %fn_ptr(...)` — return value discarded; the
subsequent abi.coerce load from a zero-init'd alloca yielded 0.
Symptom: UIRenderer.mtl_shader = 0, set_shader sees state == null,
the render-encoder fires draw with no pipeline state bound, GPU
rejects the command buffer with MTLCommandBufferErrorInternal.
Fix: at the protocol-decl method-type resolution sites in
lower.zig, check type_alias_map BEFORE falling through to
type_bridge.resolveAstType for both params and return type. A
chess-side companion fix in /Users/agra/projects/game/main.sx
(separate commit) memsets the MetalGPU struct after alloc so the
List(*void) fields' len/cap/items aren't garbage.
After all four (this commit + memset companion in chess repo):
- 71/71 regression tests pass.
- Chess game now boots, scene-connects, ticks CADisplayLink, renders
dark-gray clear + UI text + panel dividers every frame on iOS sim.
- Metal-clear example still renders.
Chess board + pieces visual contrast and faint-text-color are remaining
visual-polish items, not compiler/platform-setup issues.
Three stacked compiler bugs were causing iOS-sim chess to crash inside
[MTLTexture replaceRegion:...]. Fixing them lets every replaceRegion call
site succeed (1×1 RGBA8, 1MB R8 atlas, 440×440 chess pieces).
Path B for callconv(.c) fn-pointer casts:
- FunctionInfo now carries call_conv: CallConv (TypeInfo.CallConv) so
function-type interning distinguishes sx-CC from C-CC. Inst.zig's
Function.CallingConvention aliases the same enum.
- Parser accepts an optional `callconv(.c)` suffix on fn-pointer type
spellings (factored into parseOptionalCallConv() shared with parseFnDecl
and parseLambda).
- resolveFunctionType passes the parsed CC through functionTypeCC().
- .call_indirect reads fp.call_conv == .c and applies the C-ABI
alloca+materialize for >16B aggregate args (Path A's behaviour at .call).
Apple ARM64 ABI (drop LLVM byval):
- Side-by-side asm diff vs clang's emission for the equivalent C call site
showed LLVM's `byval` attribute lowers Apple-arm64 byval on the stack,
while clang passes the struct via a pointer in the next int register
(x2 for replaceRegion:). The runtime objc_msgSend dispatch path expects
clang's convention.
- Dropped the byval attribute from the function-signature emission and
from both call sites (.call and .call_indirect). The materialize-into-
alloca + pass-plain-ptr pattern stays — the call site now matches
clang's `mov x2, sp` exactly.
- Path A's sx-to-sx case continues to work since both ends use plain ptr
(caller does alloca+store+pass, callee loads from the ptr in prologue).
Protocol dispatch (emitProtocolDispatch):
- Untargeted `null` lowers as const_null with type .void (per
target_type orelse .void). The "wrap-value-in-alloca-pass-pointer"
branch alloca'd a void slot, which LLVM's IRBuilder asserts on —
EXC_BREAKPOINT in getTypeSizeInBits, manifesting as exit 133 / SIGTRAP
when building the chess game. Fixed by re-emitting as
constNull(void_ptr) when arg_ty == .void && expected_ty == void_ptr.
- is_pointer_ty only recognized .pointer, so [*]T (many_pointer) was
alloca-wrapped — the heap pixels pointer from stbi_load was stored
into a stack slot and the slot's address was passed as the *void arg.
Fixed by extending the check to `.pointer or .many_pointer`.
metal.sx call sites + lifecycle guards:
- msg_replace (replaceRegion:, MTLRegion = 48B) and the two setScissorRect:
sites (MTLScissorRect = 32B) now spell their fn-pointer types with
by-value params + callconv(.c) — the *MTLRegion/@local workaround is
gone.
- metal_begin_frame_ios bails before nextDrawable when pixel_w/h are 0
(drawableSize 0×0 makes nextDrawable abort via XPC).
- metal_init_ios only sets drawableSize when dims are positive.
- begin_frame's encoder/cmd_buffer failure paths now clear self.drawable
so a partial failure doesn't leak a drawable back into the pool.
Examples + tests:
- examples/86-callconv-c-fnptr-large-aggregate.sx — new, covers Path B
with C-CC fn-ptr cast.
- examples/87-fnptr-cast-large-aggregate.sx — renamed from issue-0025.sx,
covers Path B with default sx-CC (the negative case).
- examples/85-cc-c-large-aggregate.sx — from Session 60, covers Path A.
- examples/issue-0014.sx, issue-0024.sx, issue-0025.sx — removed
(resolved earlier this work).
71 regression tests pass, 0 failed. Chess game builds clean for iOS sim
and reaches its frame loop without aborting. Runtime: chess UI still
doesn't render — remaining issue is in the UIKit lifecycle / CAMetalLayer
setup (legacy-app vs scene-API hybrid), not a compiler bug. See
current/CHECKPOINT.md "Next step" for the diagnosis + options.
Step 3b code is wired across UIRenderer + GlyphCache + UIPipeline +
chess game (gpu_mode = .metal on iOS, MetalGPU bound via the GPU
protocol). macOS GL chess, iOS-sim GLES chess, and iOS-sim Metal
triangle (63-metal-clear.sx) all still render.
iOS-sim Metal chess crashes inside replaceRegion uploading the 1MB
font atlas. Bisecting that crash exposed several sx-language issues
where mid-bisect tracers (NSLog inside if/else branch bodies) didn't
produce output, blocking further investigation.
Filing each finding as examples/issue-NNNN.sx rather than working
around piecemeal:
Bugs:
- 0024 NSLog/foreign-call inside if/else body not producing output
- 0025 C-ABI param coercion incomplete for composites >16B
(combined direct-call abiCoerceParamType TODO + call_indirect
path that doesn't apply C-ABI coercion at all)
- 0026 replaceRegion 1MB upload crash (likely downstream of 0025)
Features needed for step 4 + cleanup:
- 0027 Obj-C block bridge (^{...}) for animateWithDuration:
- 0028 Optional protocol box (?GPU = null) replaces T = ---; has_T: bool
- 0029 destroy_texture/buffer/shader on GPU protocol
- 0030 extern cross-file globals
Library-side: renderer.sx + glyph_cache.sx + pipeline.sx gain a
`gpu: GPU = ---; has_gpu: bool` field pair + branches that route every
GL touchpoint through the protocol when has_gpu. glyph_cache.init
saves/restores those fields around its memset. pipeline.set_gpu()
propagates to renderer + font. Renderer's MSL shader source added as
UI_MSL_SRC using packed_float2/packed_float4 to keep the 12-float
interleaved vertex layout tight (48 bytes).
metal.sx: dual-phase init (init(null, 0, 0) for eager device+queue,
re-init with the layer once UIKit installs the SxMetalView).
setStorageMode:.shared on every texture descriptor to ensure CPU-
writable atlas pixels on Apple Silicon iOS-sim.
Regression suite: 68 passing, 0 failed. WASM chess build currently
broken under step 3b state (silent compiler crash); documented in
CHECKPOINT.md, likely fallout from one of the filed issues (probably
0028 — the verbose protocol-box pattern). Step 3b resumes after
0024-0030 land.
Phase 8 step 3a of the Metal renderer port:
- New library/modules/gpu/ with types.sx (handles + ClearColor +
TextureFormat enum), api.sx (GPU :: protocol { ... } covering the
lifecycle / per-frame / resource / per-draw surface), and metal.sx
(MetalGPU backend implementing the protocol against CAMetalLayer).
Resource handles are 1-based indices into backend List(*void) tables.
MTL aggregates >16 bytes (MTLRegion, MTLScissorRect) pass via *T to
match arm64 Apple's indirect-by-reference ABI; MTLClearColor + CGSize
go through the HFA path as direct fn-pointer casts on objc_msgSend.
- UIKitPlatform got a gpu_mode: GpuMode toggle + sibling SxMetalView
class registration. In metal mode init skips EAGL context, the
did_finish_launching IMP skips the EAGL drawable-properties dict,
layoutSubviews reads the layer's bounds * dpi_scale into pixel_w/h
instead of allocating a GL renderbuffer, and end_frame is a no-op
(the MetalGPU owns its own present).
- examples/63-metal-clear.sx verifies the pipeline end-to-end on iOS
sim — compiles a pass-through MSL shader (packed_float2/packed_float4
to avoid alignment padding), uploads 3 vertices, draws a colored
triangle on a dark-blue clear.
Compiler fixes (filed-and-fixed in this branch):
- inline if X { return E; } followed by a fall-through final expression
no longer emits two terminators into the same basic block. Verified
by examples/83-inline-if-return-fallthrough.sx.
- Top-level type alias Name :: u32; now resolves correctly as the type
annotation on a global variable (was treated as ptr {}, breaking
comparisons + initializers). Verified by examples/84-global-type-alias.sx.
Issue->feature promotion:
- 16 historical examples/issue-NNNN.sx repros now confirmed-fixed and
renamed to focused feature names (67-82). Each gains a
tests/expected/*.txt + .exit pair so the regression suite covers them.
- 5 stale issue repros deleted (subsumed by broader tests).
Regression suite: 68 passing, 0 failed. macOS chess builds + runs; wasm
chess builds; iOS sim GLES chess still renders the full board; iOS sim
Metal demo renders the triangle.
Walked back the manual-interpolation + CABasicAnimation+presentationLayer
attempts at lockstep keyboard inset. Both leave a visible frame of lag
because the lockstep problem is structural, not implementation-detail:
- GL renderbuffer content is baked at presentRenderbuffer() time.
- The CoreAnimation compositor can interpolate the *position* of a
CALayer per-vsync but cannot reach into our renderbuffer's pixels.
- The GPU pipeline (CADisplayLink → command build → present →
compositor → display) is 2-3 frames deep on iOS GLES, so even
`targetTimestamp`-based prediction is one to two frames short.
The architectural escape that doesn't move the GL view (rejected for
edge cases) is to give CoreAnimation a renderable handle it can sync
on. That means **Metal**:
- CAMetalLayer + MTLDrawable.presentAtTime(_:) caps the pipeline at
exactly one frame.
- With targetTimestamp prediction + curve-accurate keyboard math,
our drawable lands at the same vsync as UIKit's keyboard.
- Renderer modernization (Metal/Vulkan/WebGPU per platform) was on
the roadmap anyway; lockstep is the forcing function.
This commit keeps the keyboard observer + show/hide_keyboard wiring
intact and SNAPS keyboard_height when the observer fires. Behavior:
the chess board doesn't shift during the keyboard animation; it shifts
in one step when the observer fires. Less smooth than the broken
attempt but honest.
Plan for the Metal port (next):
- library/modules/gpu/{metal,vulkan,webgpu}.sx + a `GPU` protocol
analogous to Platform.
- Port modules/ui/renderer.sx shaders from GLSL to MSL.
- SxGLView becomes SxMetalView; CAEAGLLayer becomes CAMetalLayer.
- Lockstep falls out of MTLDrawable.presentAtTime(targetTimestamp).
UIKitPlatform now reads `[UIView safeAreaInsets]` (UIEdgeInsets = 32-byte
struct: top, left, bottom, right CGFloats) in begin_frame, and subscribes
to UIKeyboardWillChangeFrameNotification on NSNotificationCenter. The
chess game's build_ui pads its root by `g_safe_insets`, so the Dynamic
Island no longer overlaps the board on iPhone 17 Pro — all 8 ranks and
files are visible.
Struct returns >16 bytes (UIEdgeInsets, CGRect) go through the arm64
x8 indirect-result-pointer convention; expressing the return type on a
typed `objc_msgSend` fn-pointer cast generates the right call sequence.
Same pattern used to unwrap the keyboard's CGRect from NSValue
(UIKeyboardFrameEndUserInfoKey).
show_keyboard / hide_keyboard now drive a hidden UITextField subview as
the firstResponder source. resignFirstResponder dismisses; observer
fires with height=0 → safe_insets bottom collapses.
Deferred (next iteration): wrap the inset update in
[UIView animateWithDuration: animations:^{ ... }] to land in the same
CoreAnimation transaction as the keyboard. sx doesn't have block
syntax yet — we'd need a C shim that takes an fn-ptr and builds the
block. Today the inset snaps while the keyboard slides; the lag is
visible but the rest of the wiring is in place.
examples/66-uikit-platform.sx updated: each tap toggles the keyboard
+ advances the clear color (red→green→blue), so the observer can be
observed firing via the visible keyboard slide.
What works on iOS sim now:
- pure-UIKit boot via UIApplicationMain (no SDL3 on iOS)
- SxGLView (CAEAGLLayer) + EAGLContext(GLES3) + CADisplayLink
- GLES3 shader path in modules/ui/renderer.sx (was wasm-only; now
wasm-OR-ios)
- UITouch -> ui.Event translation (mouse_down/moved/up) on touchesBegan/
Moved/Ended/Cancelled. Verified by tapping the chess board: the
expected pawn highlights and its legal moves show as green dots.
- chdir to NSBundle.mainBundle.resourcePath inside UIKitPlatform.init so
the game's relative fopen("assets/...") calls resolve.
Required restructuring to fix four problems discovered along the way:
1. GL context + load_gl must happen BEFORE UIApplicationMain so the
game's pipeline.init (which compiles shaders) doesn't crash on null
function pointers. Pulled EAGLContext creation + load_gl out of
didFinishLaunching: into UIKitPlatform.init via uikit_create_gl_context.
2. UIScreen.nativeScale returns CGFloat (=double on 64-bit Apple).
Reading it through a `(*void, *void) -> f32` msgSend signature
clobbers the value to 0 — the upper 32 bits of d0 land where the f32
reads from. Replaced msg_f with msg_d returning f64 (and added
msg_odbl for setContentScaleFactor: which takes CGFloat).
3. `xx <f64-call-result>` directly assigned to an f32 field through a
sema path lowers as `sitofp` (integer→float) on the double — LLVM
verification rejects it. Workaround: hoist into an `f64` local first.
4. The renderer was selecting the GLSL 330 core shader on every non-wasm
target, including iOS GLES3 where it silently fails to compile and
no quads render. Added OS == .ios to the GLES branch.
Game changes:
- main.sx: g_plat is now a boxed `Platform` (not concrete *SdlPlatform).
Backend chosen per-target via `inline if OS == .ios { ... }`. The
ESC-to-stop handling is OS-guarded (mobile apps don't quit on key
press, and SDL_Keycode references would force-link SDL on iOS).
- build.sx: iOS no longer adds SDL3; it adds UIKit + OpenGLES +
QuartzCore instead.
- delta_time and viewport dims are now mirrored to free globals so the
dock subsystem (`g_dock_delta_time = @g_delta_time`) and build_ui
layout decisions don't need a pointer through the boxed protocol.
Other:
- Added `stop()` to the Platform protocol (no-op on UIKitPlatform).
- examples/66-uikit-platform.sx updated: taps advance the clear color
(red → green → blue) — smoke test for the touch IMP wiring.
- shutdown() on UIKitPlatform is a no-op (mobile apps don't tear down).
Outstanding for next session:
- The Dynamic Island notch overlaps the top of the board because we
haven't read UIView.safeAreaInsets yet (CGRect/UIEdgeInsets struct
returns require a different msgSend ABI than we currently express).
- Keyboard observer (UIKeyboardWillChangeFrameNotification + animation
duration) — the load-bearing iOS feature.
- Real-device codesigning workflow for the new build.
Two more sx compiler bugs to file out of this work:
- xx(f64 call result) → f32 emits sitofp (problem #3 above).
- Inline `#import` inside `inline if` fails to parse (we worked around
by importing both backends unconditionally; the unused-backend's
Obj-C calls are gated by `inline if OS == .ios`).
End-to-end on iOS sim: UIKitPlatform boots an SxAppDelegate, installs
an SxGLView (UIView subclass overriding +layerClass to return
CAEAGLLayer) as the root view controller's view, sets the drawable
properties (EAGLColorFormatRGBA8, non-retained backing — looked up by
dlsym so pointer-identity-checked constants match), creates an
EAGLContext (GLES3), and registers a CADisplayLink that invokes the
user's frame closure on every vsync. end_frame presents the
renderbuffer via [EAGLContext presentRenderbuffer:].
The renderbuffer is allocated lazily in -[SxGLView layoutSubviews] once
the layer has its real on-screen bounds — allocating earlier (e.g. in
didFinishLaunching) failed with INCOMPLETE_ATTACHMENT because the
SxGLView's frame was still zero at that point. Setting the SxGLView
as the VC's `view` (via setView:) lets the standard VC layout pipeline
size it to the window without us having to read CGRect struct returns
from objc_msgSend.
EAGL drawableProperties dict keys/values are dlsym'd from OpenGLES —
the framework checks them by pointer identity, so synthesized NSString
literals with the same contents don't work.
examples/66-uikit-platform.sx — runnable smoke test that cycles the
screen color (red → green → blue every 30 frames) so you can confirm
the display-link tick and present pipeline.
modules/opengl.sx gains glGenFramebuffers, glGenRenderbuffers,
glBindFramebuffer, glBindRenderbuffer, glFramebufferRenderbuffer,
glGetRenderbufferParameteriv, glCheckFramebufferStatus — needed for
the iOS GLES FBO-to-renderbuffer setup. They're wired into load_gl
so SDL and the iOS dlsym loader both pick them up.
Compiles cleanly on macOS / WASM / iOS-sim. Non-iOS targets never
reference the unresolved UIKit/QuartzCore/OpenGLES symbols because
every Obj-C touch lives inside `inline if OS == .ios`.
Game's iOS path still goes through SDL3 for now. Touch events + game
wire-up + keyboard observer = next steps.
- library/modules/platform/sdl3.sx: SdlPlatform impl wrapping SDL3 init,
GL context, event pump, swap. run_frame_loop owns the loop: while loop
on desktop, emscripten_set_main_loop on WASM. Registers an event-watch
that re-invokes the frame closure during macOS modal resize-drag so
content keeps rendering at the new size. safe_insets / keyboard /
show_keyboard / hide_keyboard are no-ops (these targets have no soft
keyboard).
Two compiler bug repros uncovered during the refactor:
- examples/issue-0020.sx: global `Foo = .{}` zero-initializes, ignoring
struct field defaults. Local `Foo = .{}` correctly applies defaults.
Workaround: set fields explicitly in an init method or heap-allocate
the value.
- examples/issue-0021.sx: an enclosing function's return type bleeds
into `xx`'s target type inside an `if-then-else` expression on the
RHS of a struct-field assignment. The same expression in a `-> void`
function produces the right value; in a `-> bool` function it
silently produces 0. Bit the SX Chess game's dpi_scale calc inside
`SdlPlatform.init` (returns bool), making all text labels render
invisibly on retina. Workaround: hoist each `xx` cast into its own
f32 local.
Regression gate: 50/50 examples pass, macOS chess game runs at ~2700fps
(close to the pre-refactor 2900 baseline), WASM build still emits a
working .html/.js/.wasm/.data quad.
