The driver handled one file. It now loads the graph rooted at the entry file
and checks every unit in it.
The import root is derived rather than configured: a unit named `a.b` read
from `<root>/a/b.fe` fixes `<root>`, so a sibling `import c.d;` is looked for
at `<root>/c/d.fe`. That is enough for the fixtures and for any tree that
follows 8.1, and it means there is no path flag to get wrong yet.
Loading is depth-first with the chain of units currently open kept on a stack,
so meeting one again is a cycle rather than a revisit -- a unit reached twice
by different paths is loaded once. Cycles, imports with no source file, and two
imports claiming the same binding are all reported.
One thing this exposed: FeDiags held a single source buffer, so once a build
spanned several files every excerpt was drawn from whichever file was parsed
last. `cycle/b.fe:3` printed the text of a.fe. fe_diags_source switches it, and
loading and checking both set it per unit.
The cycle fixtures lost their line markers on purpose. Which import closes the
cycle depends on which file you enter from -- entering at a.fe reports b.fe,
entering at b.fe reports a.fe -- so pinning a line would pin an arbitrary half
of a symmetric pair. The message is pinned; the line is not.
units: missing, bindconf and cycle pass, on top of the identity cases.
146 -> 148 of 188. What is left in units/ needs cross-unit name resolution and
visibility: an importer still cannot see `util.answer`.
Units start here, with the part that needs no import graph: what a unit is
called and where it must live.
The parser only ever read a single identifier after `unit` and `import`, so
`unit game.main;` and `import std.io;` were syntax errors -- which is why the
dotted fixtures failed at the semicolon. It now reads a dotted path and stores
it canonically, dots included, since that spelling is the unit's identity
everywhere else. `import a.b as c;` parses too, with the alias on the node.
resolve.c is the new pass between parsing and checking, for the questions that
span files. It carries SPEC 8.1 so far: each path segment is ASCII lowercase,
starts with a letter, continues with letters, digits or underscore, and is at
most eight characters; and the dotted path must match the source path it was
read from, so game.world.map has to come from game/world/map.fe. The source
side is folded to lowercase before comparing, because a case-insensitive host
must not let two spellings become two units.
That rule then applied to the fixtures, which were not obeying it: 57 declared
a unit name unrelated to their file, left over from the milestone directories,
and eight had names too long to be legal. Both are now aligned -- the rule is
worth having only if the tree follows it.
units: badupper, badlong and unitbad pass. 138 -> 146 of 188. The rest of
units/ needs the import graph, which is the next piece: resolution, cycles,
bindings and visibility.
Checking the markers for the first time found five disagreements in areas that
are implemented. Four were the marker's fault:
- own/badarg pinned "self", but the rule being broken is that a returned
reference must derive from a parameter -- `self` has nothing to do with it.
- own/badbrmov pinned line 9, which is the closing brace; the second destroy is
on line 8.
- own/badloop pinned line 8, the destroy after the loop. The diagnostic is on
line 6, inside it, and line 6 is right: the second iteration moves the same
value again, so the loop body is where it is caught. Whoever wrote the marker
expected the error after the loop.
- optional/badcatch pinned line 14, the body of the catch block. The catch
expression on line 13 is what cannot fall through.
The fifth was the compiler's. own/badweak assigns a `&mut i32` to a `&i32` and
got "initializer type mismatch", which says nothing about why. Weakening an
exclusive borrow to a shared one is a specific rule and now says so, for
references and slices alike.
own/ is fully green: 50/50. Overall 133 -> 138 of 188. The six remaining marker
disagreements are all under units/ and generic/, where nothing is implemented
yet, so there is no diagnostic to compare against and no way to tell whether
the marker is right.
The milestone structure had stopped describing the compiler and started
shaping it: m7.c, check_m7.c, tests/m2..m9, and a checker and emitter that had
each grown past 2,500 lines because there was nowhere else to put anything.
Restart from the pipeline instead.
What is left is the front end -- lexer, parser, types, ownership, semantic
analysis -- and the fixtures that describe it. The C backend, the DOSBox-X
runner, the milestone registry and the batch build are removed. The driver now
stops after semantic analysis; a code generator attaches where emit_c did.
Fixtures move from milestone directories to what they check:
parse/ grammar own/ ownership and borrowing
types/ type rules optional/ optionals and error unions
format/ formatting, try units/ units and visibility
generic/ generics pending-backend/
pending-backend/ holds the three fixtures that can only be checked by running
a program -- that the bounds check traps, that --no-checks removes it, and that
drops and defers actually fire, verified through a fake allocator. Those are
not front-end tests and are not pretending to be; they come back first when
there is a code generator.
tests/run.py replaces the DOSBox-X harness. It builds the front end with the
pinned Watcom's Windows-hosted driver and runs every fixture in about two
seconds, and it does something the old runner structurally could not: it reads
the `// ERROR:line:text` marker each fixture carries and checks the diagnostic
against it. Those markers have been in the tree all along, unverified, because
DOS could not redirect the compiler's stderr and only the exit code was ever
compared.
133/188 pass. The 55 failures are not regressions -- they are what was already
true and invisible:
- units (27) and generic (23): `import`, `comptime` and generic declarations
parse and are then dropped on the floor. No pass looks at them. The old
registry did not list these fixtures at all, so nothing said so.
- five in own/, optional/ and generic/: a marker disagrees with the diagnostic
about the line or the wording. Each is either a wrong marker or a wrong
diagnostic and has to be read individually.
Everything removed is in git history.
The host gate accepted a WATCOM environment override and skipped when nothing
was found. Both are wrong for what it is: a system-wide Open Watcom is a
different version reporting different diagnostics, and a gate that skips is a
gate that is not running, which is the exact shape of the problem this file was
added to close. It now uses .dosboxx/watcom only and fails with the setup
command when that is absent, matching how dosboxx.py already behaves.
Nothing else in the project reaches for a system install: the DOS session sets
WATCOM=W: before calling BUILD.BAT, so the C:\DEVEL\WATCOMC fallback inside it
is unreachable.
fec/test-dos.bat was tracked but dead -- the runner generates RUN.BAT and only
copies build-dos.bat -- so it goes, along with the comment claiming it drives
the build and the three fixture READMEs that still pointed at it. The registry
decides what runs now.
(.qemu/ is untracked local debris from the same era and is left alone.)
m7_emit_call reimplements call emission and only carried over mem.destroy and
mem.replace, so every mem.create/mem.alloc_slice fell through to the generic
member path and emitted `fe_missing.create(0)`. wcc386 does not diagnose that
-- it terminates with exit 255, which wcl386 reports as "Unable to invoke
wcc386.exe" with no message at all.
The breakage covered all 17 allocation sites in m5/runtime.fe, and it was
invisible because m5-owned and m5-defer only emit C; m5-runtime is the one M5
fixture that compiles and links what was generated.
Also make runtime.fe legal: `run` used `try` while returning i32, which SPEC
allows only in a function returning an error union. It is `-> !i32` now, which
M7 accepts because contextual success construction lands with it, and
runtime.c takes the { error, value } struct the error union lowers to.
Found by calling wcc386 directly instead of through wcl386, which is the only
way to see a compiler crash here.
M1-M7: 183 passed.
SPEC.md allows `try` only inside a function returning an error union, but
check.c only tests it in the FE_N_EXPR_STMT case, so `var x = try e;` and
`x = try e;` walk straight past. m5/owned.fe and m5/runtime.fe both depend on
that gap.
Moving the check onto the try expression is a four-line change and it is
correct, but it cannot land yet. runtime.fe's `run` allocates and returns a
value, so closing the hole forces it to return an error union -- and master
rejects `return <value>;` in `-> !i32` ("return type mismatch") as well as a
bare `return;` in `-> !void` ("void expression returned from value function").
Both need contextual success construction, which is M7 work. `catch` and
`@trap()`, the two spellings SPEC offers as alternatives, are also M7-only, so
there is no way to express `run` legally on master today. All three paths were
tried in DOSBox-X, not assumed.
Fix owned.fe now, since `main() -> !void` is legal today and matches
m4/try-fpr.fe, and leave the checker alone until M7 lands with the rest.
Verified: 155 passed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BScg8CF1sAAM2zVHAu5zvW
The bounds-no-checks cases emitted and built but never ran, so they only
proved that --no-checks produces compilable C -- not that it removes the
check, which is the entire point of the flag.
A run case could not simply be appended. BOUNDS.FE returns the out-of-bounds
element directly, so with checks removed its exit code is whatever sits past
the array on the stack and there is no correct status to assert. Asserting on
the generated C instead does not work either: emit_c.c defines fe_trap_bounds
unconditionally and --no-checks only suppresses the call sites.
Add NOCHK.FE, which reads one element past a [2]i32 and returns x - x. That
is 0 for whatever garbage the unchecked read produced, so the same source has
a defined outcome both ways: compiled with checks it must trap, compiled with
--no-checks it must run to completion and exit 0. Register both halves and
drop the two BOUNDS-N cases they supersede.
Verified in DOSBox-X: 155 passed, including m3-nochk-trap failing as expected
and m3-nochk-off-run succeeding.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BScg8CF1sAAM2zVHAu5zvW
조건부 이동, 이중 destroy, 직접 drop 호출에 대한 실패 fixture를 추가하고
runtime harness와 test-dos.bat를 그에 맞춰 갱신한다.
M5는 아직 완료가 아니다. 모든 경로에서 정확히 1회 cleanup, defer/drop의 선언
역순 병합, try 전파 경로 cleanup, MaybeMoved 런타임 live flag, struct drop과
필드 역순 drop, 분기/루프 상태 합류, 누수/이중해제 카운터 harness가 남아 있다.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012PQm6oAvWX4Lp3iSN5AHGT
coolguy
