Files
ww/test/wcc/726_alias_leaf_collision.c
Hojun-Cho 85af051cd1 selfhost+test: prefer same-module aliaslookup match (#27)
wwstage UNDER — aliaslookup's leaf-only first-match walk let a
cross-module leaf collision (`type invalid = !i32;` ahead of
`type invalid = !void;` in c.aliases) shadow module M's own
alias. Silent-correct-by-zero-init: the let-decl prologue zeroed
the slot 8B-wide (typeis8byteprimitive's void-aliased path,
post-#22), so MOVSXD on the misresolved !i32 produced the right
value while diverging from cstage's MOVQ — bootstrap byte-id
held until any caller bumped the alias-chain ordering. Mirrors
cstage scope_lookup_prefer (cmd/wcc/check.c:65); module-
qualified pkg.alias path unchanged.

Polarity catalog: wwstage UNDER — aliaslookup missing module-
preferring scope discipline. Convergence wwstage → cstage's
resolver pattern (rule 10; cstage already correct via
scope_lookup_prefer). Two-pass walk: same-module first, then
existing first-match fallback. Sea-of-stars shape preserved.

Surfaced by lib/strings landing: utf8's `type invalid = !void;`
and strconv's `type invalid = !i32;` registered in the same flat
c.aliases under one combined.ww, with strconv's later-registered
entry sitting at the head of the chain. utf8.next/decode's
`return e;` (e: invalid) packed via MOVSXD instead of MOVQ. Four
sites in main.s, contributing to 993/995 byte-id divergence in
the wwstage rebuild path.

Tests:
  - 726_alias_leaf_collision row 1 pins MOVQ post-zero-init on
    both stages and cstage↔wwstage cmp -s byte-id for the
    `(invalid:!void via beta)` shape with `alpha.invalid = !i32`
    seeded ahead in c.aliases. Sentinel-flip-verified: revert →
    wwstage emits MOVSXD post-zero-init + cmp diverges.
  - Row 2 (i32_local_read_keeps_movsxd) gates against future
    symptom-fix attempts: legitimate `let i: i32; return (i:i64);`
    must still emit MOVSXD on both stages (>=2 narrow signed loads
    in the promote fn TEXT). Independent of the aliaslookup fix.

98/98 ok. 995_self_rebuild stays green (ww2==ww3==ww4 byte-id).
2026-05-18 09:03:25 +09:00

278 lines
8.3 KiB
C

/*
* 726_alias_leaf_collision — sentinel for #27. Pins wwstage's
* aliaslookup to a same-module-first leaf walk so a bare `invalid`
* inside module M (with `type invalid = !void;`) resolves to M's
* underlying void, not to some other module's `type invalid = !i32;`
* registered later in c.aliases.
*
* Pre-fix wwstage's `aliaslookup` (selfhost/cmd/wcc/cgen.ww) walked
* c.aliases head-first by leaf name, returning the FIRST match. With
* combined-source build orderings where a `type invalid = !i32;` from
* lib/strconv sits ahead of utf8's `type invalid = !void;` in the
* chain, `localloadop` resolved invalid → !i32 (size 4, signed) and
* emitted MOVSXD on a slot the let-decl prologue zero-inited 8B-wide
* (typeis8byteprimitive walks to void → MOVQ $0). Silent-correct-by-
* zero-init: the upper 4B happened to be zero from the zero-init, so
* the 4B sign-extend produced the right value, but the asm diverged
* from cstage's MOVQ and 995_self_rebuild broke once any caller of
* lib/strings/strings.byteindex (the (str|rune) match-arm shape that
* dragged utf8 into the bootstrap closure) bumped the alias-chain
* ordering.
*
* Cstage's `resolve_typename` (cmd/wcc/check.c:65) calls
* `scope_lookup_prefer(cur, cur_mod, nm)` — same-module-first then
* fallback. The fix mirrors that discipline in wwstage's aliaslookup.
* Asserts byte-id between cstage and wwstage on a probe shape that
* loses the asm divergence under the fix.
*/
#include <stdio.h>
#include <stdlib.h>
#include <string.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;
}
struct row {
const char *label;
const char *src;
/* `want` selects which assertion family runs on the emitted .s:
* "movq_invalid" — post-zero-init load on beta.yield_more's
* invalid-slot must be MOVQ (the collision case).
* "movsxd_i32" — local i32 read in promote must be MOVSXD on
* BOTH stages. Guards future "fix" attempts
* that pessimize localloadop and would break
* legitimate signed-narrow loads. */
const char *want;
};
/* Module ordering picked so the colliding `type invalid = !i32;` lands
* later in c.aliases (head-prepend → at the head), masking module M's
* own `type invalid = !void;`. `gamma` brings both modules into scope
* via use directives so decl_mod tags both as imports — that gates
* scope_define_in_module's per-mod dedup path on cstage's side. */
static const struct row rows[] = {
{ "void_invalid_under_i32_collision",
"// MODULE: gamma\n"
"use alpha;\n"
"use beta;\n"
"export fn main() i32 = { return 0; };\n"
"// MODULE: beta\n"
"type more = void;\n"
"type invalid = !void;\n"
"fn yield_more() (rune | more | invalid) = {\n"
" let e: invalid;\n"
" return e;\n"
"};\n"
"// MODULE: alpha\n"
"export type invalid = !i32;\n",
"movq_invalid" },
/* Regression guard: a real signed-narrow `let i: i32 = ...;` read
* MUST stay on MOVSXD. Any future "fix" that pessimizes
* localloadop to blanket-MOVQ in the name of #27 would silently
* regress this — losing sign-extension on i32 → i64 promotion
* (also drops the high half of a deref-store-narrowed slot). */
{ "i32_local_read_keeps_movsxd",
"fn promote(x: i32) i64 = {\n"
" let i: i32 = x;\n"
" return (i: i64);\n"
"};\n",
"movsxd_i32" },
};
static int
emit_s(const char *w6c, const struct row *r, int i, char *out_s, size_t cap)
{
char src[64], cmd[1024];
snprintf(src, sizeof src, "/tmp/alc_%d_%d.ww", getpid(), i);
snprintf(out_s, cap, "/tmp/alc_%d_%d_%s.s",
getpid(), i, w6c[strlen(w6c) - 1] == 'w' ? "ww" : "c");
FILE *f = fopen(src, "wb");
if (!f) return -1;
fputs(r->src, f);
fclose(f);
snprintf(cmd, sizeof cmd, "%s -o %s %s 2>/dev/null", w6c, out_s, src);
int rc = runwait(cmd);
unlink(src);
return rc;
}
static int
slurp(const char *path, char *buf, size_t cap)
{
FILE *f = fopen(path, "rb");
if (!f) return -1;
size_t n = fread(buf, 1, cap - 1, f);
fclose(f);
buf[n] = '\0';
return (int)n;
}
/* The yield_more body must read its tagged-return payload through an
* 8B MOVQ. A MOVSXD on the same slot is the pre-fix silent-correct-
* by-zero-init shape. */
static int
check_movq_after_zeroinit(const char *spath, const struct row *r,
const char *stage)
{
char buf[1 << 14];
if (slurp(spath, buf, sizeof buf) < 0) {
fprintf(stderr, "row[%s][%s]: cannot read %s\n",
r->label, stage, spath);
return -1;
}
const char *fn = strstr(buf, "TEXT beta.yield_more");
if (!fn) {
fprintf(stderr, "row[%s][%s]: no TEXT beta.yield_more in %s\n",
r->label, stage, spath);
return -1;
}
const char *zi = strstr(fn, "MOVQ\t$0, ");
if (!zi) {
fprintf(stderr, "row[%s][%s]: no zero-init store\n",
r->label, stage);
return -1;
}
const char *bad = strstr(zi, "MOVSXD\t");
const char *good = strstr(zi, "\n\tMOVQ\t-");
const char *ret = strstr(zi, "RET");
if (bad && (ret == NULL || bad < ret)) {
fprintf(stderr,
"row[%s][%s]: post-zero-init slot read uses MOVSXD\n",
r->label, stage);
return -1;
}
if (!good || (ret && good > ret)) {
fprintf(stderr,
"row[%s][%s]: post-zero-init slot read missing MOVQ\n",
r->label, stage);
return -1;
}
return 0;
}
/* Real i32 local read MUST stay MOVSXD. Any MOVQ load of `i` (the
* narrow signed local) inside promote means a later "fix" pessimized
* localloadop and silently dropped sign-extension. */
static int
check_movsxd_kept(const char *spath, const struct row *r, const char *stage)
{
char buf[1 << 14];
if (slurp(spath, buf, sizeof buf) < 0) {
fprintf(stderr, "row[%s][%s]: cannot read %s\n",
r->label, stage, spath);
return -1;
}
const char *fn = strstr(buf, "TEXT promote");
if (!fn) {
fprintf(stderr, "row[%s][%s]: no TEXT promote in %s\n",
r->label, stage, spath);
return -1;
}
const char *ret = strstr(fn, "RET");
const char *bp = fn;
int sxd = 0;
for (;;) {
const char *p = strstr(bp, "MOVSXD\t-");
if (!p || (ret && p > ret)) break;
sxd++;
bp = p + 1;
}
if (sxd < 2) {
fprintf(stderr,
"row[%s][%s]: expected MOVSXD on i32 local reads, found %d\n",
r->label, stage, sxd);
return -1;
}
return 0;
}
int
main(void)
{
const char *bin = getenv("BIN");
if (!bin) bin = "out/bin";
char absbin[512];
if (bin[0] != '/') {
char cwd[256];
if (getcwd(cwd, sizeof cwd) == NULL) return 1;
snprintf(absbin, sizeof absbin, "%s/%s", cwd, bin);
bin = absbin;
}
char w6c[640], w6c_ww[640];
snprintf(w6c, sizeof w6c, "%s/w6c", bin);
snprintf(w6c_ww, sizeof w6c_ww, "%s/w6c_ww", bin);
int have_ww = (access(w6c_ww, X_OK) == 0);
int n = (int)(sizeof rows / sizeof rows[0]);
int total = 0, fail = 0;
for (int i = 0; i < n; i++) {
char cs_path[128], ws_path[128];
if (emit_s(w6c, &rows[i], i, cs_path, sizeof cs_path) != 0) {
fprintf(stderr,
"alias_leaf_collision[cstage][%s]: w6c failed\n",
rows[i].label);
fail++; total++; continue;
}
total++;
int rc;
if (strcmp(rows[i].want, "movsxd_i32") == 0) {
rc = check_movsxd_kept(cs_path, &rows[i], "cstage");
} else {
rc = check_movq_after_zeroinit(cs_path, &rows[i], "cstage");
}
if (rc != 0) fail++;
if (!have_ww) { unlink(cs_path); continue; }
if (emit_s(w6c_ww, &rows[i], i, ws_path, sizeof ws_path) != 0) {
fprintf(stderr,
"alias_leaf_collision[wwstage][%s]: w6c_ww failed\n",
rows[i].label);
fail++; total++;
unlink(cs_path); continue;
}
total++;
if (strcmp(rows[i].want, "movsxd_i32") == 0) {
rc = check_movsxd_kept(ws_path, &rows[i], "wwstage");
} else {
rc = check_movq_after_zeroinit(ws_path, &rows[i], "wwstage");
}
if (rc != 0) fail++;
/* byte-id between stages on the same input — the principled
* sentinel for the polarity catalog. */
total++;
char cmd[512];
snprintf(cmd, sizeof cmd, "cmp -s %s %s", cs_path, ws_path);
if (runwait(cmd) != 0) {
fprintf(stderr,
"alias_leaf_collision[%s]: cstage vs wwstage asm differs\n",
rows[i].label);
fail++;
}
unlink(cs_path); unlink(ws_path);
}
if (fail) {
fprintf(stderr,
"alias_leaf_collision: %d/%d fixtures failed\n", fail, total);
return 1;
}
printf("alias_leaf_collision: %d/%d ok\n", total, total);
return 0;
}