sx/current/CHECKPOINT-ASM.md

# sx Inline Assembly — Checkpoint (ASM stream)

Companion to `current/PLAN-ASM.md`; design in
[docs/inline-asm-design.md](../docs/inline-asm-design.md). Update after every
commit, one step at a time per the cadence rule (no commit may both add a test
and make it pass).

## Last completed step
**G (read-write `+` place outputs)** — a `+r` / `+{reg}` `-> @place` output is now
implemented (the last substantive feature). LLVM has no `+` constraint, so a
read-write place lowers to: an output **`=`** constraint (return slot, stored back
through the place after the call; the leading `+` rewritten to `=` in
`appendAsmConstraints`), **plus** a **tied input** (the decimal index of that
output) appended **after** the regular inputs, seeded with the place's loaded
value passed as a call arg. Tied inputs come **last** so existing operand indices
(`%[name]`→`${N}`) are undisturbed — `asmOperandIndex` unchanged. Lowering
(`lowerAsmExpr`) no longer rejects `+` (indirect `*` still rejected loudly).
`emitInlineAsm` (`src/backend/llvm/ops.zig`): grows arg/param arrays by the rw
count (`n_args = n_inputs + n_rw`), loads each seed (`asm.rw.seed`), emits the
tied constraint, and the existing store-back path writes the modified output back.
New `asmIsReadWrite(e, op)` helper. Verified by **running**: increment-in-place
(41→42, IR `"=r,0"`) and a mixed case (rw place + regular input + value output) →
textbook `"=r,=r,r,0"` with correct `${N}` indices and args `(input, seed)`. Two
commits per cadence: (1) `examples/1650-platform-asm-rw-place.sx` locked the
rejection; (2) implemented + flipped 1650 to a runnable aarch64-pinned example
(`{ "target": "macos" }`, ir-only elsewhere). `zig build test` green (658 corpus,
446 unit). Files: `src/ir/lower/expr.zig`, `src/backend/llvm/ops.zig`,
`examples/1650-*`.

Prior: **2** — `-> @place` write-through outputs. An asm result can be **stored through
a place** (local / struct field) instead of returned; the place output does NOT
join the result tuple. Parser: `-> @place` parses the `@place` as an ordinary
address-of expression → an `out_place` operand (`src/parser.zig`). Lowering
(`lowerAsmExpr`): out_place operand = the lowered `@place` address, `out_ty` =
the pointee; read-write (`+`) and indirect-memory (`*`) constraints rejected
loudly (not yet implemented). Added `out_ty: TypeId` to the IR `AsmOperand`
(`src/ir/inst.zig`) so emit builds the **combined** return struct (ALL outputs).
`emitInlineAsm` rewrite (`src/backend/llvm/ops.zig`): the LLVM return type is now
built from every output's `out_ty`; after the call, out_place slots are
`store`d through their address and out_value slots rebuild the sx result — with a
**fast path** (no place outputs → the asm's struct return IS the result, so
pure-value asm IR is unchanged). Verified: write-to-local (`get42`→42), struct
field (`@p.b`), mixed value+place (`v=10 b=20`), `+` rejected. Locked with
`examples/1649-platform-asm-place-output.sx` (mixed, runs on aarch64). `zig build
test` green (657 corpus, 446 unit). Files: `src/parser.zig`, `src/ir/inst.zig`,
`src/ir/lower/expr.zig`, `src/backend/llvm/ops.zig`, `examples/1649-*`.

Prior: **F** — global (module-scope) asm. A top-level `asm { "tmpl", };` block (template
only) lowers to LLVM `module asm`, and a lib-less `extern` calls into the symbols
it defines. New `asm_global` AST node (`src/ast.zig`) + `parseAsmGlobal`
(`src/parser.zig`, dispatched from `parseTopLevel` on `kw_asm`) — rejects
`volatile` and any operands/clobbers. The node forced (and got) arms in the same
three `Node.Data` switches as `asm_expr` (`sema.zig` ×2, `semantic_diagnostics.zig`).
`Module` gains a `global_asm: ArrayList([]const u8)` (`src/ir/module.zig`);
`lowerMainAndComptime` captures each template (the dead `lowerDecls` is NOT the
top-level pass — `lowerRoot` Pass 2 uses `lowerMainAndComptime`); `emit_llvm.zig`'s
`emit()` appends each via `LLVMAppendModuleInlineAsm` (source order). Verified
end-to-end: an aarch64 `_my_add` global routine called via `extern` returns 42 —
**AOT only** (the ORC JIT doesn't link module-asm symbols, so `sx run` is wrong;
the design ties global-asm symbols to the final linked binary). Locked with
`examples/1648-platform-asm-global.sx` (`.build { "aot": true, "target": "macos" }`
→ AOT build+run on aarch64, ir-only elsewhere). `zig build test` green (656
corpus, 446 unit). Files: `src/ast.zig`, `src/parser.zig`, `src/sema.zig`,
`src/ir/semantic_diagnostics.zig`, `src/ir/module.zig`, `src/ir/lower/decl.zig`,
`src/ir/emit_llvm.zig`, `examples/1648-*`.

Prior: **E** — multi-output tuples. **Inline asm now returns tuples.** Replaced the
N>1 bail with a shared `asmResultType` helper (`src/ir/lower/expr.zig`, mixed
into `Lowering`) that derives the result type from the `out_value` operands
(0→void, 1→T, N→named tuple, named via the §II.5 effective-name rule). The key
realization: `toLLVMType(tuple)` already produces a literal struct `{T1,…,Tn}` —
exactly LLVM's multi-output asm return — so **emit needed NO change**; building
the op with a tuple result type makes the asm call return the struct, which IS
sx's tuple value (destructured by the normal `tuple_get` path). `inferType`'s
`.asm_expr` arm now also delegates to `asmResultType` (single owner), so
`return asm`, `x := asm`, and `q, r := asm` all agree on the type. Verified
end-to-end on aarch64: `split(0x1234)`→`(lo=52, hi=18)`, a udiv/msub divmod→
`(3, 2)`. IR is textbook: `call { i64, i64 } asm "divq ${4}",
"={rax},={rdx},{rax},{rdx},r,~{cc}"(…)` → extractvalue → tuple. Converted 1640 to
the x86_64 multi-output IR lock (ir-only) + added `1647-platform-asm-aarch64-multi`
(runs on aarch64). `zig build test` green (655 corpus, 446 unit). Files:
`src/ir/lower/expr.zig`, `src/ir/lower.zig`, `src/ir/expr_typer.zig`,
`examples/164{0,7}-*`.

Prior: **C.1 + D** — inline asm CODEGEN (lowering builds the op + LLVM emit). **Inline
assembly now runs end-to-end.** `lowerAsmExpr` (`src/ir/lower/expr.zig`) stops
bailing: it resolves each operand's effective name (§II.5 auto-naming), interns
template/constraints/clobbers, lowers input `Ref`s, derives the result `TypeId`
(0→void, 1→T), and builds the `inline_asm` op. Added a `%[name]`-references-a-
real-operand check (the last deferred validation). Multi-output (N>1) still bails
loudly ("Phase E"). `emitInlineAsm` (`src/backend/llvm/ops.zig`, port of Zig's
`airAssembly`): assembles the LLVM constraint string (outputs→inputs→`~{clobber}`,
`,`→`|`), rewrites the template (`%[name]`→`${N}`, `%%`→`%`, `$`→`$$`, `%=`→
`${:uid}`), then `LLVMGetInlineAsm` + `LLVMBuildCall2` (AT&T). Dispatch wired
(`emit_llvm.zig`, replacing the C.0 `@panic`). **`llvm_shim.c`**: added
`LLVMInitializeNativeAsmParser()` — the JIT must assemble inline asm at run time.
Verified end-to-end: aarch64 `add`/`mov` run on the host (exit 42), `nop volatile`
runs (1642 now exit 0), IR is textbook (`call i64 asm "add ${0},${1},${2}",
"=r,r,r"(…)`). Locked with `examples/1645-platform-asm-aarch64-add.sx` (runs on
aarch64, ir-only elsewhere via `.build` + `.ir`). Also added the `inferType`
`.asm_expr` arm (`src/ir/expr_typer.zig`, 0→void / 1→T) — without it a bare
`x := asm {…-> T}` binding inferred `.unresolved` and silently produced 0;
regression-locked with `examples/1646-platform-asm-value-binding.sx`. Updated
1640 (now Phase-E bail) + 1642 (now runs). `zig build test` green (654 corpus,
446 unit). Files: `src/ir/lower/expr.zig`, `src/backend/llvm/ops.zig`,
`src/ir/emit_llvm.zig`, `src/ir/expr_typer.zig`, `llvm_shim.c`,
`examples/164{0,2,5,6}-*`.

Prior: **C.0** — IR op `inline_asm` (lock; no behavior change). Added `inline_asm:
InlineAsm` to the IR `Op` union + the `InlineAsm` struct (`template: StringId`,
`operands: []const AsmOperand` {role/name/constraint/operand}, `clobbers:
[]const StringId`, `has_side_effects`) in `src/ir/inst.zig` — all strings
interned, operands in source order, result on `Inst.ty`. The new variant forced
(and got) arms in two exhaustive `Op` switches: `src/ir/interp.zig` (loud
`bailDetail` — inline asm is never comptime-evaluable) and `src/ir/print.zig`
(IR dump). `src/ir/emit_llvm.zig` gets a `@panic` **tripwire** — emit lands in
Phase D, and until then `lowerAsmExpr` still bails so no `inline_asm` op is ever
created (reaching emit would be a lowering-switched-over-too-early bug). Unit
test `inline_asm op shape` in `src/ir/inst.test.zig`. `zig build test` green
(652 corpus, 446 unit). Files: `src/ir/inst.zig`, `src/ir/interp.zig`,
`src/ir/print.zig`, `src/ir/emit_llvm.zig`, `src/ir/inst.test.zig`.

Prior: **B.1** — operand-name validation (design §II.5 auto-naming rule). Extended
`lowerAsmExpr` with a `pinnedRegister(constraint)` helper (`"={eax}"`→`eax`,
`"+{rax}"`→`rax`, `"=r"`→null) and two checks: (1) **reject the echo form**
`[eax] "={eax}"` — a label identical to its own pinned register is redundant
(the operand is already auto-named after the register); (2) **reject duplicate
operand names** (ambiguous `%[name]` / result field). Locked with
`examples/1643-platform-asm-echo-name.sx` + `1644-platform-asm-duplicate-name.sx`.
`zig build test` green (652 corpus, 0 failed; 445 unit). Files:
`src/ir/lower/expr.zig`.

Prior: **B.0** — asm shape validation (compile-path diagnostics). Restructured the
`.asm_expr` lowering arm into `lowerAsmExpr` (`src/ir/lower/expr.zig`, mixed into
`Lowering` in `src/ir/lower.zig`): it validates BEFORE the not-yet-implemented
codegen bail, so the user sees the real problem first. Two checklist items now
enforced with named diagnostics: (1) **template must be a compile-time-known
string** (`"..."` / `#string`); (2) **no value outputs ⇒ must be `volatile`**
(mirrors Zig — a result-less asm could be deleted). Valid shapes still bail with
the "codegen not yet implemented" message. Result-type derivation + auto-naming
stay deferred to a later step (observable only once Phase C produces a real IR
op). Locked with `examples/1641-platform-asm-missing-volatile.sx` (volatile
error) + `1642-platform-asm-nop-volatile.sx` (volatile no-output accepted →
codegen bail). `zig build test` green (650 corpus, 0 failed; 445 unit). Files:
`src/ir/lower/expr.zig`, `src/ir/lower.zig`, `examples/164{1,2}-*`.

Prior: **A.1** — parse `asm { … }` + loud lowering bail (folded A.1+A.2 into one honest
lock commit, since the loud bail IS current correct behavior — cadence option
(a)). Added `AsmExpr`/`AsmOperand` to `src/ast.zig` + the `asm_expr` `Node.Data`
arm; `parseAsmExpr` in `src/parser.zig` (`parsePrimary` `.kw_asm` dispatch) —
parses the template, flat operand list (`[name]? "constraint" -> Type` value
output / `= expr` input), and `clobbers(.…)`; `volatile`/`clobbers` recognized
contextually via `isContextualWord`. The new `asm_expr` tag forced (and got)
arms in three exhaustive `Node.Data` switches: `src/sema.zig` `analyzeNode` +
`findNodeAtOffset`, `src/ir/semantic_diagnostics.zig` `checkBindingNames` (all
recurse into template + operand payloads). Lowering bails LOUD + named in
`src/ir/lower/expr.zig` ("inline assembly codegen is not yet implemented…") via
an explicit `.asm_expr` arm (not the generic `unknown_expr` else) returning
`emitPlaceholder`. `-> @place` write-through is rejected with a clear "Phase 2"
parse error. Locked with `examples/1640-platform-asm-parse.sx` (multi-output
`divmod`, named operands, register pins, clobbers — parses then bails; called
from `main`). `zig build test` green (648 corpus, 0 failed; 445 unit). Files:
`src/ast.zig`, `src/parser.zig`, `src/sema.zig`, `src/ir/semantic_diagnostics.zig`,
`src/ir/lower/expr.zig`, `examples/1640-*`.

Prior: **A.0** — `kw_asm` keyword (first compiler code). Added the `kw_asm` `Token.Tag`
variant + `.{ "asm", .kw_asm }` keyword-map entry in `src/token.zig`; `volatile` /
`clobbers` deliberately stay OUT of the global table (contextual). New exhaustive
`Tag` switch in `src/lsp/server.zig` `classifyToken` flagged the missing arm (the
intended coverage tripwire) — added `.kw_asm` to the keyword group. Lock test in
new `src/lexer.test.zig` (`asm`→`kw_asm`, `volatile`/`clobbers`→`identifier`),
wired into the `src/root.zig` barrel as `lexer_tests`. `zig build test` green (648
corpus, 0 failed; 445 unit, 0 failed — +1). Files: `src/token.zig`,
`src/lexer.test.zig`, `src/root.zig`, `src/lsp/server.zig`.

Prior: **0.2** — CLAUDE.md docs for `<name>.build`; **Phase 0 COMPLETE**.
**0.1** — corpus runner **ir-only branch** for cross-target examples. Replaced
0.0's loud placeholder bail: when `cfg.target` doesn't match the host (`ir_only`),
`sweepRoot` skips run/build/exec and verifies via `sx ir --target` only —
asserting `.exit` (ir cmd) + `.ir` (normalized stdout) + `.stderr`, never
`.stdout` (write skipped in update mode, assertion skipped in verify mode). An
`.ir` snapshot is **required** in ir-only mode — its absence is a loud failure
("needs an .ir snapshot for ir-only mode"). Locked with
`examples/1639-platform-target-cross.sx` (asm-free `main :: () -> i64 { return 0;
}`), `.build` `{ "target": "x86_64-linux" }`, + checked-in `.ir`. Verified both
guards fire: corrupting the `.ir` → IR mismatch; deleting it → the require-failure.
`zig build test` green (647 corpus, 0 failed; 444 unit). Files:
`src/corpus_run.test.zig`, `examples/1639-*`.

## Current state
**Inline assembly works end-to-end: 0, 1, and N value outputs (tuples).** Full
pipeline: lex (A.0) → parse (A.1) → validate (B.0/B.1 + `%[name]` check) → IR op
(C.0) → lower-builds-op + LLVM emit + JIT asm-parser init (C.1/D) → multi-output
tuples (E). Register-class + register-pinned operands, inputs, clobbers, `#string`
multi-instruction templates, `%[name]`/`%%` rewriting, and the §II.5 auto-naming
rule all work and execute on the host JIT. Global `asm { … }` (Phase F) works AOT (call-into-asm
via lib-less `extern`). `-> @place` **write-through** outputs work (Phase 2) and
**read-write (`+`)** place outputs work (Phase G — tied-input lowering, runs on
aarch64). Indirect-memory (`*`) place outputs are still rejected loudly as
not-yet-implemented — the only remaining substantive feature. Smaller
follow-ups: the comptime-call guard for global asm (`#run` into a module-asm
symbol should fail loud via dlsym-miss — pin a test), a JIT-vs-global-asm note
(`sx run` silently mishandles module-asm symbols; AOT is correct), and the x86_64
syscall ir-only example. `readme.md` now has an "Inline Assembly" section.

Known orthogonal bug: **issue 0137** — `sx run` on a program with no `main`
segfaults (`src/target.zig:256-273`, unguarded JIT entry lookup). Pre-existing,
asm-independent; does NOT block the ASM stream (every example has a `main`).

Phase E–F feasibility already confirmed against the live tree
(`LLVMGetInlineAsm` / `LLVMBuildCall2` / `LLVMAppendModuleInlineAsm` in LLVM@19
`Core.h`; ERR-stream `extractvalue`→tuple in `emit_llvm.zig:726-927`; lib-less
`extern`, 60 sites; `--target` a global CLI flag).

## Next step
Inline assembly is **feature-complete for the common surface** plus read-write
(`+`) place outputs. Remaining work, all optional / additive (pick any):
- **Indirect-memory (`"=*m"`) outputs**: pass the place address as an arg, asm
  writes through it (no return slot). Currently rejected.
- **Output-to-`const` rejection** for `-> @place` (the place must be mutable).
- **Polish**: comptime-call guard test for global asm; make `sx run` error (not
  silently mishandle) a module-asm symbol; x86_64 syscall-write ir-only example.

Orthogonal: **issue 0137** (no-`main` segfault).

## Log
- (init) Plan + design doc written; ASM stream opened.
- (0.0) Corpus runner target-gating: `<name>.build` JSON config (replaces `.aot`
  marker), `--target` threading, `hostMatchesTarget` execute-gate, loud
  cross-target placeholder bail. Migrated 1226/1227 `.aot`→`.build`; locked with
  1638 fixture + unit tests. `zig build test` green.
- (0.1) ir-only branch: cross-target examples verify via `sx ir --target` only
  (exit+ir+stderr, no stdout; `.ir` required). Locked with 1639 fixture; verified
  corrupt-.ir → mismatch and missing-.ir → loud failure. `zig build test` green.
- (0.2) docs: CLAUDE.md documents `<name>.build` JSON sidecar (aot + target +
  ir-only gating), replacing stale `.aot` marker prose. **Phase 0 COMPLETE.**
- (A.0) `kw_asm` keyword in token.zig (+ map entry); LSP `classifyToken` switch
  coverage; lock test in new `lexer.test.zig` (wired via root.zig). `volatile` /
  `clobbers` stay contextual identifiers. `zig build test` green (445 unit, +1).
- (A.1) parse `asm { … }` → `AsmExpr` + loud lowering bail; `asm_expr` arms in 3
  exhaustive `Node.Data` switches; `-> @place` rejected (Phase 2). Adopted operand
  auto-naming rule (design §II.5). Locked with 1640 fixture. Filed orthogonal
  issue 0137 (no-`main` JIT segfault). `zig build test` green (648 corpus, 445 unit).
- (B.0) asm shape validation in `lowerAsmExpr`: comptime-string template +
  no-output⇒volatile, with named diagnostics before the codegen bail. Locked with
  1641 (volatile error) + 1642 (volatile accepted). `zig build test` green (650
  corpus, 445 unit).
- (B.1) operand-name validation: `pinnedRegister` helper + reject echo form
  (`[eax] "={eax}"`) and duplicate names. Locked with 1643 + 1644. `zig build
  test` green (652 corpus, 445 unit).
- (C.0) IR op `inline_asm: InlineAsm` + interp `bailDetail` + print arm + emit
  `@panic` tripwire (Phase D). No behavior change (lowering still bails). Unit
  test `inline_asm op shape`. `zig build test` green (652 corpus, 446 unit).
- (C.1+D) CODEGEN — `lowerAsmExpr` builds the op (effective names, interned
  strings, input Refs, 0/1 result type) + `%[name]` validation; `emitInlineAsm`
  (constraint string + template rewrite + `LLVMGetInlineAsm`/`BuildCall2`, AT&T);
  `inferType` arm; `LLVMInitializeNativeAsmParser` for the JIT. **Inline asm runs
  end-to-end.** N>1 bails (Phase E). Locked with 1645 (aarch64 add, runs) + 1646
  (`:=` binding); updated 1640/1642. `zig build test` green (654 corpus, 446 unit).
- (E) multi-output tuples — `asmResultType` helper (0→void/1→T/N→named tuple),
  shared by lowering + `inferType`. `toLLVMType(tuple)` == LLVM multi-output
  struct, so emit unchanged; the asm struct return IS the sx tuple. Runs on
  aarch64 (1647: `split`→`(lo,hi)`); 1640 → x86 multi-output IR lock (ir-only).
  `zig build test` green (655 corpus, 446 unit).
- (F) global asm — `asm_global` AST node + `parseAsmGlobal` (top-level, rejects
  volatile/operands); `Module.global_asm` captured in `lowerMainAndComptime`;
  `emit()` appends via `LLVMAppendModuleInlineAsm`; call-into via lib-less
  `extern`. AOT-verified (1648, `_my_add`→42). `zig build test` green (656 corpus).
- (docs) readme.md "Inline Assembly" section (b8800a2).
- (2) `-> @place` write-through — `out_place` operand; `out_ty` on the IR
  AsmOperand; `emitInlineAsm` builds the combined output struct + splits
  (out_place → store-through, out_value → result), fast-path when no places.
  `+`/`*` rejected. Locked with 1649 (mixed, runs). `zig build test` green (657
  corpus, 446 unit).
- (G) read-write `+` place outputs — `+` lowers to an output `=` + a tied input
  (output-index constraint) seeded with the place's loaded value, tied inputs
  appended last (operand indices undisturbed). `appendAsmConstraints` rewrites
  `+`→`=`; `emitInlineAsm` grows args by the rw count + loads seeds;
  `asmIsReadWrite` helper. Lowering stops rejecting `+` (`*` still rejected). Two
  commits (cadence): 1650 locked the rejection, then flipped to a runnable
  aarch64 example (`"=r,0"` IR). `zig build test` green (658 corpus, 446 unit).

## Known issues
- **0137** — `sx run` on a program with no `main` segfaults (unguarded JIT entry
  lookup, `src/target.zig:256-273`). Pre-existing, asm-independent. Filed
  `issues/0137-jit-run-no-main-segfault.md`. Does not block A.1.