fix(vm): make findAll linear by stopping matchAt when no threads remain

tests/regression.zigmodified+88-0

Changes to tests/regression.zig

		@@ -1,4 +1,5 @@
1	1	const std = ;
	2	+const builtin = ;
2	3	const Regex = .Regex;
3	4	const RegexError = .RegexError;
4	5
		⋯174 linesShow the 174 lines above this hunk@@ -179,3 +180,90 @@ test "regression: double pattern compile and deinit" {
179	180	try std.testing.expect(try regex.isMatch("def"));
180	181	}
181	182	}
	183	+
	184	+// --- findAll quadratic blowup (issue: "findAll is O(n^2), not linear") ---
	185	+//
	186	+// Before the fix, matchAt kept iterating to input.len after all threads died,
	187	+// and findAll restarted matchAt per match, so a scan with m matches over n
	188	+// bytes cost ~O(n*m). The guard below doubles the input and checks the time
	189	+// ratio: a linear scan stays near ~2x, the pre-fix quadratic path was ~4x and
	190	+// growing. A ratio (not an absolute bound) keeps this independent of build
	191	+// mode (Debug vs ReleaseFast) and machine speed.
	192	+
	193	+fn monotonicNs() u64 {
	194	+ const clk: std.c.clockid_t = switch (builtin.os.tag) {
	195	+ .macos, .ios, .tvos, .watchos, .visionos => .UPTIME_RAW,
	196	+ else => .MONOTONIC,
	197	+ };
	198	+ var ts: std.c.timespec = undefined;
	199	+ _ = std.c.clock_gettime(clk, &ts);
	200	+ return @as(u64, @intCast(ts.sec)) * 1_000_000_000 + @as(u64, @intCast(ts.nsec));
	201	+}
	202	+
	203	+const FindAllScan = struct { ns: u64, count: usize, first_start: usize, last_end: usize };
	204	+
	205	+fn timeFindAll(allocator: std.mem.Allocator, regex: *const Regex, n: usize) !FindAllScan {
	206	+ const buf = try allocator.alloc(u8, n);
	207	+ defer allocator.free(buf);
	208	+ @memset(buf, '.');
	209	+ var i: usize = 0;
	210	+ while (i + 8 <= n) : (i += 64) @memcpy(buf[i .. i + 8], "Sherlock");
	211	+
	212	+ // Warm up (allocator caches, code paths), then take the min of a few runs
	213	+ // to suppress scheduler noise without depending on absolute timing.
	214	+ {
	215	+ const warm = try regex.findAll(allocator, buf);
	216	+ for (warm) |*m| {
	217	+ var mm = m.*;
	218	+ mm.deinit(allocator);
	219	+ }
	220	+ allocator.free(warm);
	221	+ }
	222	+
	223	+ var best: u64 = std.math.maxInt(u64);
	224	+ var count: usize = 0;
	225	+ var first_start: usize = 0;
	226	+ var last_end: usize = 0;
	227	+ var rep: usize = 0;
	228	+ while (rep < 3) : (rep += 1) {
	229	+ const t0 = monotonicNs();
	230	+ const matches = try regex.findAll(allocator, buf);
	231	+ const dt = monotonicNs() - t0;
	232	+ count = matches.len;
	233	+ if (matches.len > 0) {
	234	+ first_start = matches[0].start;
	235	+ last_end = matches[matches.len - 1].end;
	236	+ }
	237	+ for (matches) |*m| {
	238	+ var mm = m.*;
	239	+ mm.deinit(allocator);
	240	+ }
	241	+ allocator.free(matches);
	242	+ if (dt < best) best = dt;
	243	+ }
	244	+ return .{ .ns = best, .count = count, .first_start = first_start, .last_end = last_end };
	245	+}
	246	+
	247	+test "regression: findAll scales linearly (no O(n^2) blowup)" {
	248	+ const allocator = std.testing.allocator;
	249	+
	250	+ var regex = try Regex.compile(allocator, "Sherlock");
	251	+ defer regex.deinit();
	252	+
	253	+ const n1: usize = 16 * 1024;
	254	+ const n2: usize = 32 * 1024; // exactly 2x
	255	+
	256	+ const r1 = try timeFindAll(allocator, &regex, n1);
	257	+ const r2 = try timeFindAll(allocator, &regex, n2);
	258	+
	259	+ // Correctness: a "Sherlock" every 64 bytes, found at the right places.
	260	+ try std.testing.expectEqual(n1 / 64, r1.count);
	261	+ try std.testing.expectEqual(n2 / 64, r2.count);
	262	+ try std.testing.expectEqual(@as(usize, 0), r2.first_start);
	263	+ try std.testing.expectEqual(((n2 - 8) / 64) * 64 + 8, r2.last_end);
	264	+
	265	+ // Linearity guard: doubling input must not more-than-triple the time.
	266	+ // Linear ≈ 2x; the pre-fix quadratic path was ≈ 4x (and worsening).
	267	+ try std.testing.expect(r1.ns > 0);
	268	+ try std.testing.expect(r2.ns * 10 < r1.ns * 30); // r2/r1 < 3.0
	269	+}