Test-first scaffolding for LLVM backend modularization (Phase A7.1) before the
type/ABI helpers move into src/backend/llvm/{types,abi}.zig. Visibility-only
change to the targets — no behavior change. Closes the ARCH-SAFETY "no generic
ABI snapshot" gap.
- 2 new emit_llvm.test.zig tests:
- abiCoerceParamType across every C-ABI size bucket: <=8 -> i64, 9-16 ->
[2 x i64], >16 -> ptr, HFA (all-float/all-double, <=4 fields) -> unchanged,
string -> ptr, slice -> ptr, scalar -> unchanged. Built via a local
internStruct helper (field slice in the module arena -> no testing-allocator
leak); asserts against emitter.cached_* + LLVMArrayType2.
- needsByval: true only for >16-byte non-HFA struct; false for <=16 / HFA /
string / slice / non-struct.
- 1 new .ir snapshot: 1202-ffi-cc-c-large-aggregate (the canonical callconv(.c)
>16-byte byval example that directly documents abiCoerceParamType) — pins the
byval param path end-to-end (5 byval + entry reload + 2 sret from Arena.init).
Path-free + idempotent (verified across two captures). Suite count unchanged
(snapshot added to an existing example).
- Widened abiCoerceParamType + needsByval to pub (visibility only;
abiCoerceParamTypeEx/materializeByvalArg/verifySizes stay private — move with
callers in sub-step 2). No logic touched.
- Recorded the A7.1 coverage inventory + residual gaps (wasm32 usize->i32 branch,
fn-ptr large-aggregate 1203/1204) in ARCH-SAFETY.md.
Gate: zig build, zig build test, bash tests/run_examples.sh -> 361/0 (no churn
beyond the new 1202 .ir).
A source path with no directory component (`sx build main.sx` from the
project dir — what the chess app does) made `diFileFor` emit a `DIFile`
with an empty `directory:`, so the compile unit's `DW_AT_comp_dir` was
"". Apple's ld then silently drops the *entire* object's debug map (0
N_OSO) and the binary is undebuggable — lldb resolves no sx source.
Builds whose path had any directory (`.sx-tmp/x.sx`, `examples/x.sx`)
were unaffected, which is why small repros + the stepping smoke passed
and only the bundled chess app hit it.
Fix: diFileFor falls back to "." (and "/" for a root-level file) when
the path has no directory component, so comp_dir is never empty.
Verified: chess (`sx build --target macos --emit-obj main.sx`) now
links with OSO=1 and lldb resolves `frame at main.sx:82:8`. Regression
guard added to the DWARF unit test (asserts `DIFile(... directory: ".")`
for a bare filename). Gates: zig build, zig build test, run_examples.sh
-> 291 passed, debug-stepping smoke ok.
Attach LLVM debug metadata so a captured return-address PC resolves to
file:line:col (the runtime half E3.3 needs) and sx binaries become
debuggable in lldb/gdb.
- llvm_api.zig: bind llvm-c/DebugInfo.h (DIBuilder C API was unbound).
- emit_llvm.zig: DIBuilder + one DICompileUnit/DIFile on the main file,
a DISubprogram per function (LLVMSetSubprogram), and a DILocation per
instruction from Inst.span (errors.SourceLoc.compute, scoped to the
subprogram). Plus the "Debug Info Version"/"Dwarf Version" module
flags and LLVMDIBuilderFinalize.
- Gated on opt none/less + a wired source map (setDebugContext from
core.zig), mirroring lower.zig's tracesEnabled; release strips it.
Verified: sx ir/sx asm --opt none show correct DILocations + .loc
directives; the 290-example JIT suite (-O0 -> debug on) verifies and
runs unchanged. +2 DWARF unit tests.
root.zig had no `test` block, so the test binary discovered zero tests and
trivially "passed" — every src test had silently rotted. Add
`refAllDecls(@This())` to root.zig so all 185 tests run, then fix the rot it
surfaced:
- emit_llvm.test: operands were constants, so LLVM folded the very
instructions being asserted (fadd/sub/icmp/insertvalue/extractvalue/sext).
Rewrite to use function-parameter operands; `main` now returns i32 (entry
convention); tagged-union enum_init lowers via memory, not insertvalue.
- interp.test: switch the per-test allocator to an arena (the interpreter is
arena-style and intentionally frees little) — clears the transient-Value
leaks without an ownership-ambiguous source change.
- lower.test: pass `is_imported` to lowerFunction; mark two helpers `pub`; the
if/else block test now uses a runtime (param) condition since lowering folds
`if true`.
- print.test: SSA numbering — params occupy %0/%1, so consts start at %2.
- jni_java_emit.test: nested-class refs render in Java source form
(`SurfaceHolder.Callback`), not the JNI `$` form.
Leaks fixed at the source where ownership was clear: Module gains an arena for
the operand slices the Builder dupes (struct/call/branch/switch args, block
params, lowerFunction params); objcDefinedStateStructType builds its field
slice in that arena and frees its temp name string.
allocViaContext used to fall back to a direct `.heap_alloc` (libc
malloc) when `Context` wasn't registered — i.e. when the program
didn't import std.sx. That was a silent escape hatch: a program could
appear to allocate fine without a `Context`, sidestepping protocol
dispatch entirely. Same shape as the matchContextAllocCall trap we
removed, just in a different code path.
Now: every site that needs `Context` emits a clear diagnostic when
the type isn't in scope, pointing the user at the required import.
- `allocViaContext`: the three fallback branches (no implicit_ctx, no
Context type, malformed Context struct) all call the new
`diagnoseMissingContext("heap allocation")` and return a
placeholder. Codegen no longer emits libc malloc as the silent
no-import path.
- `lowerPush`: the no-Context branches used to silently drop the
push and just lower the body. Now diagnose first, then lower
(keeping the body's other diagnostics flowing).
- `lowerIdentifier` for "context": used to silently fall through to
`global_names.get("context")` (which would emit an unresolved
identifier with no actionable hint). Now diagnose with the
required-import message.
With every consumer gone, the `.heap_alloc` and `.heap_free` IR ops
are deleted entirely:
- `inst.zig`: drop the Op variants.
- `interp.zig`: drop the execInst arms.
- `emit_llvm.zig`: drop the arms (the `getOrDeclareMalloc/Free`
helpers stay — they're still used by the foreign-decl path for
user-level `malloc`/`free` foreign bindings).
- `print.zig`: drop the printers + the isVoidOp arm.
- `emit_llvm.test.zig`: drop the unit test (op no longer exists).
155/155 example tests pass. Unit tests green. Chess green on macOS /
iOS sim / Android. A program that doesn't import std.sx and tries to
use `context.allocator.alloc` or `push Context.{}` or the `context`
identifier now gets a real error:
error: heap allocation requires the Context type — add
`#import "modules/std.sx";` (or a module that imports it)
Closes the last silent allocation-protocol escape.
