wwstage: expand spread variants in match check + size them off flattened members (#209)

The wwstage checker walked a match's raw AST variant list and never expanded a ...inner spread variant, so it rejected fmt's match over field = (...formattable | *mods) ('not a variant of scrutinee'). cstage's resolve_type flattens the spread at type-build. Mirror that in the AST exhaustiveness walk (casevariantin + a recursive checkvariantcovered): when a variant resolves to N_TTAGGED via a spread, recurse into its members. Additive + spread-gated -- typeeqast / casevariantpairmatch (#13) / casecovers untouched, so non-spread matches and 990-997 byte-id are unaffected. Also size a spread N_TTAGGED off each flattened member (mirror cstage check.c), dropping the inner union tag word (field 40B to 32B). Closes the #209 CHECKER reject; full fmt-byte-id still awaits cgen cluster #226 (io.read nominal-remap) + #227 (spread-widen ABI), so fmt tests stay cstage-only with retargeted comments. Adds test 792; regenerates w6c/wwdump combined.ww.
2026-05-31 20:31:40 +09:00
parent 497f1fa0d3
commit 3d44f742a0
10 changed files with 471 additions and 113 deletions
--- a/test/wcc/777_fmt_handle_run.c
+++ b/test/wcc/777_fmt_handle_run.c
@@ -16,18 +16,19 @@
 * (the io.stream arm) and reads back via memio.string — the two-arm
 * coverage #5 graduated.
 *
- * Cstage-only per row (no STAGE_WW, no byte_id) — pre-existing wwstage
- * bug #209 (sibling of #190): the wwstage checker bails `case: not a
- * variant of scrutinee (X)` + `match: variant not handled
- * (formattable)` on fmt's match-on-formattable arms whenever a
- * downstream probe `import fmt;`s the package. The bug bites identically
- * — `fmt.errorln("hi")` from a probe trips the same trace — and is
- * pre-existing (reproduces on master). Every fmt test is cstage-only
- * (970 fmttest runs the @test fixture under cstage `ww run` only); 995
- * self-rebuild dodges it because no selfhost main pulls err.ww
- * transitively, so wwstage never compiles fmt in the bootstrap. Fix =
- * one-class checker repair, out of #5 commit-2 scope; byte-id graduates
- * with #209 close.
+ * Cstage-only per row (no STAGE_WW, no byte_id). #209 (the wwstage
+ * checker bail `case: not a variant of scrutinee` + `match: variant not
+ * handled (formattable)` on fmt's spread-union match-arms) is now CLOSED
+ * — wwstage compiles fmt. The residual blocker is a SEPARATE pre-existing
+ * cgen cluster surfaced once fmt actually codegens under wwstage:
+ *   - #226: io.read's error widening (cgwidentagremap) loses the
+ *     io.eof/io.error NAMED tinfo identity under fmt-presence
+ *     (flatvariantidxt → -1 → tag collapses to 0); cs!=ww whole-file, so
+ *     byte-id fails even though io.read is unused here (it's emitted
+ *     regardless). The #10/#218/tinfo-lossy-nominal family.
+ *   - #227: spread-union widen + return-ABI, broken in BOTH stages.
+ * 995 self-rebuild dodges all of this because no selfhost main pulls fmt
+ * transitively. Byte-id graduates when #226 (+#227) land.
 *
 *   row                  | what it pins
 *   ---------------------+--------------------------------------
@@ -293,8 +294,8 @@ run_row(const char *driver, const char *cwd, const struct row *r, int seq)
 /* asm_byte_identical — diff cstage vs wwstage .s. Parallel trees so
 * ww_ww writing intermediates next to the source doesn't clobber the
 * cstage .s (CLAUDE.md rule 14 phase split). Mirror of 776's. Unused
- * while every row is cstage-only (#209), kept for the byte-id
- * graduation once #209 closes. */
+ * while every row is cstage-only (#226/#227; #209 is closed), kept for
+ * the byte-id graduation once those cgen folds land. */
 static int
 asm_byte_identical(const char *cdrv, const char *wdrv, const char *cwd,
    const struct row *r, int seq)
@@ -356,7 +357,7 @@ main(void)
 	int total = 0, fail = 0;
 	int wwpresent = (access(wdrv, X_OK) == 0);
 	int seq = 0;
-	(void)asm_byte_identical;	/* cstage-only until #209 closes */
+	(void)asm_byte_identical;	/* cstage-only until #226/#227 land */

 	for (int i = 0; i < n; i++) {
 		if (rows[i].stage_mask & STAGE_CS) {
--- a/test/wcc/778_bufio_vstream_run.c
+++ b/test/wcc/778_bufio_vstream_run.c
@@ -15,9 +15,12 @@
 *     scanline / finish.
 *
 * Each row imports os + bufio + memio + io (NOT fmt — bufio doesn't
- * transitively pull fmt, so #209's formattable-match bail doesn't
- * propagate; both stages run on every row, cs.s==ww.s — bufio is NOT
- * compiler-embedded, so these byte-id rows are its ONLY byte-id cover).
+ * transitively pull fmt, so it dodges the fmt-presence cgen cluster
+ * #226 (io.read error-remap nominal-identity) + #227 (spread-union
+ * widen/return-ABI) that blocks the fmt-importing tests' byte-id; both
+ * stages run on every row, cs.s==ww.s — bufio is NOT compiler-embedded,
+ * so these byte-id rows are its ONLY byte-id cover). (#209, the wwstage
+ * formattable match-arm CHECKER bail, is now closed.)
 *
 *   row                       | what it pins
 *   --------------------------+--------------------------------------
--- a/test/wcc/779_log_vstream_run.c
+++ b/test/wcc/779_log_vstream_run.c
@@ -12,18 +12,16 @@
 * Each row imports os + log + memio + io. log itself imports fmt,
 * so the formattable / field types flow through transitively.
 *
- * Cstage-only per row (no STAGE_WW, no byte_id) — pre-existing
- * wwstage bug #209 (sibling of #190): the wwstage checker bails on
- * match-arm-over-formattable when fmt is imported transitively, which
- * lib/log does for the logger vtable signatures + the fmt.fprint /
- * fprintf dispatch in stdprintln / stdprintfln. The bug bites the
- * OLD log surface identically — `log.println("hi")` from a probe
- * trips the same trace. Existing 970 logtest runs the @test fixture
- * under cstage `ww run` only, so the OLD surface is fine via cstage.
- * 995 self-rebuild dodges it because no selfhost cmd imports log
- * transitively. Fix-of-the-bug = one-class checker repair, out-of-
- * scope for the additive fold-e5; closes the wwstage half here when
- * it lands. Byte-id graduates with #209 close.
+ * Cstage-only per row (no STAGE_WW, no byte_id). #209 (the wwstage
+ * checker bail on match-arm-over-formattable when fmt is imported
+ * transitively — which lib/log does for the logger vtable signatures +
+ * the fmt.fprint / fprintf dispatch in stdprintln / stdprintfln) is now
+ * CLOSED; wwstage compiles fmt/log. The residual blocker is the
+ * pre-existing cgen cluster #226 (io.read error-remap nominal-identity)
+ * + #227 (spread-union widen/return-ABI), surfaced once fmt codegens
+ * under wwstage. Existing 971 logtest runs the @test fixture under
+ * cstage `ww run` only. 995 self-rebuild dodges it because no selfhost
+ * cmd imports log transitively. Byte-id graduates when #226 (+#227) land.
 *
 *   row                       | what it pins
 *   --------------------------+--------------------------------------
@@ -70,8 +68,9 @@
 *     using `os.trywrite(...)?`; same shape memio.ww + fmt.ww +
 *     bufio.ww adopt.
 *
- *   - #209 (wwstage formattable match-arm bail): every row is
- *     STAGE_CS-only; byte-id deferred until #209 lands.
+ *   - #226 (io.read error-remap nominal-identity) + #227 (spread-union
+ *     widen/return-ABI): every row is STAGE_CS-only; byte-id deferred
+ *     until they land. #209 (the checker bail) is already closed.
 *
 * BOOTSTRAP-EMBED CHECK: log is NOT embedded in any selfhost
 * combined.ww (grep `package log\|import log` returns empty), so the
@@ -251,8 +250,9 @@ run_row(const char *driver, const char *cwd, const struct row *r, int seq)
 /* asm_byte_identical — diff cstage vs wwstage .s. Parallel trees so
 * ww_ww writing intermediates next to the source doesn't clobber the
 * cstage .s (CLAUDE.md rule 14 phase split). Mirror of 778's. Unused
- * for fold-e5 — every row is STAGE_CS-only per #209 — but kept
- * scaffolded for when #209 closes and byte-id graduates. */
+ * for fold-e5 — every row is STAGE_CS-only per #226/#227 (#209, the
+ * checker bail, is closed) — but kept scaffolded for when those cgen
+ * folds land and byte-id graduates. */
 static int
 asm_byte_identical(const char *cdrv, const char *wdrv, const char *cwd,
    const struct row *r, int seq)
--- a/test/wcc/780_fmt_mods_run.c
+++ b/test/wcc/780_fmt_mods_run.c
@@ -11,11 +11,12 @@
 * calls fprintf with a {:mods} format string + one arg, and asserts the
 * exact byte content read back.
 *
- * Cstage-only per row — pre-existing wwstage bug #209 (sibling of
- * #190): the wwstage checker bails on fmt.formattable match-arms
- * whenever a downstream probe `import fmt;`s the package. Identical
- * carve-out to 777_fmt_handle_run (cited there). Byte-id graduates with
- * #209 close.
+ * Cstage-only per row. #209 (the wwstage checker bail on fmt's
+ * spread-union match-arms) is now CLOSED; the residual blocker is the
+ * pre-existing cgen cluster #226 (io.read error-remap nominal-identity)
+ * + #227 (spread-union widen/return-ABI). Identical carve-out to
+ * 777_fmt_handle_run (cited there). Byte-id graduates when #226 (+#227)
+ * land.
 *
 *   row              | format     | arg          | expect    | exit
 *   -----------------+------------+--------------+-----------+-----
@@ -44,8 +45,10 @@
 *
 * DEFERRALS / RELATED (NOT addressed by these rows):
 *
- *   - #209 (wwstage formattable match-arm bail): blocks STAGE_WW here,
- *     same as 777_fmt_handle_run. Cstage-only until close.
+ *   - #226 (io.read error-remap nominal-identity) + #227 (spread-union
+ *     widen/return-ABI): block STAGE_WW here, same as 777_fmt_handle_run.
+ *     #209 (the checker bail) is closed; these cgen folds are the
+ *     residual. Cstage-only until they land.
 *
 * BOOTSTRAP-EMBED CHECK: fmt is NOT embedded in any selfhost
 * combined.ww (grep confirmed), so the #5 fmt graduation does NOT
--- a/test/wcc/781_fmt_compositions_run.c
+++ b/test/wcc/781_fmt_compositions_run.c
@@ -33,13 +33,13 @@
 *                       | io.close. Asserts the owned str bytes, then
 *                       | frees via os.free.
 *
- * Cstage-only per row (no STAGE_WW, no byte_id) — pre-existing wwstage
- * bug #209 (sibling of #190): the wwstage checker bails `case: not a
- * variant of scrutinee (X)` + `match: variant not handled
- * (formattable)` on fmt.formattable match-arms whenever a downstream
- * probe `import fmt;`s the package. Identical carve-out to
- * 777_fmt_handle_run + 780_fmt_mods_run. Byte-id graduates with #209
- * close.
+ * Cstage-only per row (no STAGE_WW, no byte_id). #209 (the wwstage
+ * checker bail `case: not a variant of scrutinee` + `match: variant not
+ * handled (formattable)` on fmt's spread-union match-arms) is now CLOSED;
+ * wwstage compiles fmt. The residual blocker is the pre-existing cgen
+ * cluster #226 (io.read error-remap nominal-identity) + #227 (spread-union
+ * widen/return-ABI). Identical carve-out to 777_fmt_handle_run +
+ * 780_fmt_mods_run. Byte-id graduates when #226 (+#227) land.
 *
 * IMPORT-ORDER WORKAROUND (sibling task #208, pre-existing): every row
 * places `import os;` FIRST for the same reason 776/777/780 document
@@ -48,8 +48,10 @@
 *
 * DEFERRALS / RELATED (NOT addressed by these rows):
 *
- *   - #209 (wwstage formattable match-arm bail): blocks STAGE_WW here,
- *     same as 777/780. Cstage-only until close.
+ *   - #226 (io.read error-remap nominal-identity) + #227 (spread-union
+ *     widen/return-ABI): block STAGE_WW here, same as 777/780. #209 (the
+ *     checker bail) is closed; these cgen folds are the residual.
+ *     Cstage-only until they land.
 *   - #173 (TRY-on-tagged-return both-stages broken): bsprintf widens
 *     nomem into io.error inline rather than `memio.fixed(buf)?` for
 *     the same reason memio.ww does.
--- a/test/wcc/792_spread_variant_match.c
+++ b/test/wcc/792_spread_variant_match.c
@@ -0,0 +1,163 @@
+/*
+ * 792_spread_variant_match — project #209 close. Pins that the wwstage
+ * checker accepts a `match` over a tagged union with a `...inner` SPREAD
+ * variant, byte-identically with cstage (rule-10).
+ *
+ * THE BUG (wwstage-CHECKER-only, cs!=ww): a spread variant
+ * `type field = (...inner | str)` flattens its inner union's members
+ * into `field` (i64|bool|rune|str). cstage's resolve_type performs the
+ * flatten at type-build (cmd/wcc/check.c:651-674), so its match-arm
+ * validity + exhaustiveness see the flattened members. wwstage's checker
+ * (selfhost/cmd/wcc/check.ww checkmatchexhaust) walked the RAW AST
+ * `u.list`, which keeps the spread unexpanded — every member arm tripped
+ * "case: not a variant of scrutinee" and the phantom spread node tripped
+ * "match: variant not handled". So any module that matches over a spread
+ * union (e.g. fmt's `field = (...formattable | *mods)`) was
+ * wwstage-uncompilable. cstage compiled it fine. Hare flattens spreads;
+ * align UP to cstage (a too-strict checker, NOT a down-align).
+ *
+ * THE FIX (#209): casevariantin + a checkvariantcovered coverage helper
+ * recurse into a spread variant's inner-union members (resolvealias →
+ * N_TTAGGED → walk members), mirroring cstage's flatten at the AST
+ * layer; the #61a tinfo build additionally sizes the union off the
+ * flattened MEMBERS (not the whole inner union) so the slot matches
+ * cstage's. cgen already dispatches off the flattened tinfo.params
+ * (flatvariantidxt), so no cgen change is needed — checker-only.
+ *
+ * The PRE-FIX failure mode is a wwstage CHECKER REJECT, so the
+ * discriminator is `w6c_ww` on the driver-produced combined.ww
+ * succeeding AT ALL — pre-fix it errored out; post-fix it succeeds AND
+ * is byte-id with cstage's w6c. The probe stays on the param+match shape
+ * (a fn taking the spread union and matching it, exhaustive, no
+ * default): that is exactly what #209's checker fix covers and is
+ * byte-id. (Spread-union CONSTRUCTION/return-ABI is a separate,
+ * pre-existing both-stage cgen path — out of #209 scope — so it is
+ * deliberately NOT exercised here.) Self-contained single-file probe.
+ */
+#define _GNU_SOURCE
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <sys/wait.h>
+
+static int
+runwait(const char *cmd)
+{
+	int rc = system(cmd);
+	if (rc == -1) return -1;
+	if (WIFEXITED(rc)) return WEXITSTATUS(rc);
+	return -1;
+}
+
+static int
+slurp_eq(const char *a, const char *b)
+{
+	FILE *fa = fopen(a, "rb");
+	FILE *fb = fopen(b, "rb");
+	if (!fa || !fb) { if (fa) fclose(fa); if (fb) fclose(fb); return -1; }
+	int rc = 0;
+	for (;;) {
+		int ca = fgetc(fa);
+		int cb = fgetc(fb);
+		if (ca != cb) { rc = -1; break; }
+		if (ca == EOF) break;
+	}
+	fclose(fa); fclose(fb);
+	return rc;
+}
+
+static const char SRC[] =
+	"package main;\n"
+	"\n"
+	"type inner = (i64 | bool | rune);\n"
+	"type field = (...inner | str);\n"
+	"\n"
+	"fn classify(f: field) i64 = {\n"
+	"	match (f) {\n"
+	"	case let n: i64 => return n;\n"
+	"	case let b: bool => return 1;\n"
+	"	case let r: rune => return 2;\n"
+	"	case let s: str => return (s.len: i64);\n"
+	"	};\n"
+	"};\n"
+	"\n"
+	"export fn main() i32 = { return 0; };\n";
+
+int
+main(void)
+{
+	const char *bin = getenv("BIN");
+	if (!bin) bin = "out/bin";
+	char absbin[2048];
+	if (bin[0] != '/') {
+		char cwd[1024];
+		if (getcwd(cwd, sizeof cwd) == NULL) return 1;
+		snprintf(absbin, sizeof absbin, "%s/%s", cwd, bin);
+		bin = absbin;
+	}
+
+	char w6c[2100], w6c_ww[2100];
+	snprintf(w6c, sizeof w6c, "%s/w6c", bin);
+	snprintf(w6c_ww, sizeof w6c_ww, "%s/w6c_ww", bin);
+	if (access(w6c_ww, X_OK) != 0) {
+		fprintf(stderr, "792: w6c_ww missing — cannot run the cs==ww "
+		    "byte-id gate (the whole point of this test)\n");
+		return 1;
+	}
+
+	char dir[] = "/tmp/ww792_XXXXXX";
+	if (mkdtemp(dir) == NULL) { fprintf(stderr, "792: mkdtemp\n"); return 1; }
+
+	char path[1024], cmd[4096], comb[1024], cs_s[1024], ws_s[1024];
+	int rc = 0;
+
+	snprintf(path, sizeof path, "%s/main.ww", dir);
+	FILE *f = fopen(path, "wb");
+	if (!f) { fprintf(stderr, "792: write main.ww\n"); rc = 1; goto done; }
+	fputs(SRC, f);
+	fclose(f);
+
+	snprintf(comb, sizeof comb, "%s/main.combined.ww", dir);
+
+	/* cstage driver build: produces the combined.ww + must accept. */
+	snprintf(cmd, sizeof cmd, "cd %s && %s/ww build -I %s %s/main.ww",
+	    dir, bin, dir, dir);
+	if (runwait(cmd) != 0) {
+		fprintf(stderr, "792: cstage build failed\n");
+		rc = 1; goto done;
+	}
+	if (access(comb, 0) != 0) {
+		fprintf(stderr, "792: no combined.ww produced\n");
+		rc = 1; goto done;
+	}
+
+	/* The #209 discriminator: raw w6c_ww on the combined. Pre-fix the
+	 * wwstage checker REJECTED the spread-union match arms (non-zero
+	 * exit). Post-fix it accepts AND is byte-id with cstage's w6c. */
+	snprintf(cs_s, sizeof cs_s, "%s/cs.s", dir);
+	snprintf(ws_s, sizeof ws_s, "%s/ww.s", dir);
+	snprintf(cmd, sizeof cmd, "%s -o %s %s 2>/dev/null", w6c, cs_s, comb);
+	if (runwait(cmd) != 0) {
+		fprintf(stderr, "792: w6c on combined failed\n");
+		rc = 1; goto done;
+	}
+	snprintf(cmd, sizeof cmd, "%s -o %s %s 2>/dev/null", w6c_ww, ws_s, comb);
+	if (runwait(cmd) != 0) {
+		fprintf(stderr, "792: w6c_ww on combined failed "
+		    "(#209 spread-union match reject?)\n");
+		rc = 1; goto done;
+	}
+	if (slurp_eq(cs_s, ws_s) != 0) {
+		fprintf(stderr, "792: cs.s/ww.s DIFFER (rule-10 byte-id "
+		    "violation)\n");
+		rc = 1; goto done;
+	}
+
+	printf("spread_variant_match: ok (w6c_ww accepts spread match + "
+	    "cs==ww byte-id)\n");
+
+done:
+	snprintf(cmd, sizeof cmd, "rm -rf %s", dir);
+	(void)runwait(cmd);
+	return rc;
+}