// time — clocks, instants, durations. Mirrors Hare's lib/time
// (ref/hare/time/duration.ha, instant.ha, arithm.ha,
// +linux/functions.ha). Calendar / date / strftime / timezone /
// sleep live in separate Hare modules and graduate when callers /
// supporting stdlib arrive.
//
// `duration` is a NAMED alias of i64 (lib/math/random precedent
// at lib/math/random/random.ww:8); ww treats NAMED as a newtype,
// so cross-i64 arithmetic inside this module needs explicit casts.
// Hare's structural alias semantics let those casts vanish, but
// our type checker is strict.

package time;

@symbol("rt_syscall") fn syscall2(num: i64, a: i64, b: i64) i64;
@symbol("rt_abort")   fn abort(msg: str) void;

def SYS_CLOCK_GETTIME: i64 = 228;

// ref/hare/time/duration.ha:6. 290y representable range.
export type duration = i64;

// ref/hare/time/duration.ha:9-18. Plan-9 naming (lowercase)
// diverges from Hare's uppercase per project rule 4.
export def nanosecond:  duration = 1i64;
export def microsecond: duration = 1000i64;
export def millisecond: duration = 1000000i64;
export def second:      duration = 1000000000i64;

// ref/hare/time/instant.ha:9. (sec, nsec) pair — NOT POSIX struct
// timespec (which uses u32 nsec). Layout matches Linux's struct
// timespec on 64-bit (i64+i64) so we can pass &instant directly
// to clock_gettime.
export type instant = struct {
	sec:  i64,
	nsec: i64,
};

// ref/hare/time/+linux/functions.ha:84. First cut exposes only
// realtime and monotonic; Hare's process_cpu / thread_cpu / boot /
// realtime_alarm / boot_alarm / tai graduate when a caller needs
// them (CLAUDE.md rule 9 — Hare-fidelity, no premature surface).
export type clock = enum i32 {
	realtime  = 0,
	monotonic = 1,
};

// ref/hare/time/+linux/functions.ha:138. Hare's now() also aborts
// on impossible errnos. (instant | oserror) is deliberately not
// the return shape — EINVAL / EFAULT are programmer errors (bad
// clock id, bad ptr), and a 1-word-payload sum return walks into
// task #9's cgen-divergence trap.
export fn now(c: clock) instant = {
	let i: instant;
	let rc = syscall2(SYS_CLOCK_GETTIME, (c as i32): i64, (&i): i64);
	if (rc != 0i64) { abort("time.now: clock_gettime failed"); };
	return i;
};

// ref/hare/time/arithm.ha:9. Adds duration to instant. The
// negative-duration branch normalises nsec into [0, second).
export fn add(i: instant, x: duration) instant = {
	let r: instant;
	let xi: i64 = x: i64;
	let sec: i64 = second: i64;
	let nsec: i64 = nanosecond: i64;
	if (xi == 0i64) {
		r.sec  = i.sec;
		r.nsec = i.nsec;
		return r;
	};
	if (xi > 0i64) {
		r.sec  = i.sec + (i.nsec + xi) / sec;
		r.nsec = (i.nsec + xi) % sec;
		return r;
	};
	r.sec  = i.sec + (i.nsec + xi - sec + nsec) / sec;
	r.nsec = (i.nsec + (xi % sec) + sec) % sec;
	return r;
};

// ref/hare/time/arithm.ha:26. Returns duration from a to b.
// Sign convention: b - a.
export fn diff(a: instant, b: instant) duration = {
	let sec: i64 = second: i64;
	let v: i64 = ((b.sec - a.sec) * sec) + (b.nsec - a.nsec);
	return v: duration;
};

// ref/hare/time/arithm.ha:32. -1 if a < b, 0 if equal, +1 if a > b.
export fn compare(a: instant, b: instant) i8 = {
	if (a.sec  < b.sec)  { return -1i8; };
	if (a.sec  > b.sec)  { return  1i8; };
	if (a.nsec < b.nsec) { return -1i8; };
	if (a.nsec > b.nsec) { return  1i8; };
	return 0i8;
};

// os — process and filesystem facade. The body of each call lands
// either in libwwrt.a (rt_syscall trampoline) or libc bindings,
// depending on how the program was linked.

package os;

import time;

@symbol("rt_syscall") fn syscall0(num: nr) i64;
@symbol("rt_syscall") fn syscall1(num: nr, a: i64) i64;
@symbol("rt_syscall") fn syscall2(num: nr, a: i64, b: i64) i64;
@symbol("rt_syscall") fn syscall3(num: nr, a: i64, b: i64, c: i64) i64;
@symbol("rt_syscall") fn syscall4(num: nr, a: i64, b: i64, c: i64, d: i64) i64;

@symbol("rt_free")  export fn free(p: *void, n: u64) void;
@symbol("rt_abort") fn abort(msg: str) void;

// Hare-style runtime check. Caller passes a message that's printed
// to stderr before exit(1).
export fn assert(cond: bool, msg: str) void = {
	if (!cond) { abort(msg); };
};

// Linux amd64 syscall numbers. Internal to this module — passed as
// the first arg of syscall0..4 via libwwrt's rt_syscall trampoline.
// `nr` is the type so the call sites can't accidentally pass an
// arbitrary i64 (`syscall1(0i64, ...)` no longer typechecks).
type nr = enum i64 {
	READ        = 0,
	WRITE       = 1,
	OPEN        = 2,
	CLOSE       = 3,
	LSEEK       = 8,
	ACCESS      = 21,
	DUP2        = 33,
	GETPID      = 39,
	FORK        = 57,
	EXECVE      = 59,
	EXIT        = 60,
	WAIT4       = 61,
	MKDIR       = 83,
	RMDIR       = 84,
	UNLINK      = 87,
	GETCWD      = 79,
	GETDENTS64  = 217,
	NEWFSTATAT  = 262,
};

// open(2) flags. Linux values, matching <fcntl.h>. Hare names them
// `fs::flag::RDONLY` etc; we use the same leaf names so callers say
// `os.flag.RDONLY` and `os.flag.WRONLY | os.flag.CREATE`.
export type flag = enum i32 {
	RDONLY  = 0,
	WRONLY  = 1,
	RDWR    = 2,
	CREATE  = 64,	// 0x40
	EXCL    = 128,	// 0x80 — pair with CREATE to fail on existing path
	TRUNC   = 512,	// 0x200
};

// lseek(2) whence. Hare names it `io::whence`.
export type whence = enum i32 {
	SET = 0,
	CUR = 1,
	END = 2,
};

export fn exit(code: i32) void = {
	syscall1(nr.EXIT, code: i64);
};

// PATH_MAX / pathbuf / kpath — port of Hare's ref/hare/sys/+linux/
// syscalls.ha:25,27,29-55. Hare's `path` accepts a sum `(str |
// []u8 | *const u8)`; ww's lib/os public surface narrows to `str`
// (the Hare-faithful surface at ref/hare/os/os.ha:37,47,50 etc).
// Internally, [[kpath]] copies the `str` bytes into a single
// module-level [[pathbuf]] scratch slot and NUL-terminates so the
// raw Linux syscalls (which require C strings) see a valid
// terminator. Same precedent as Hare's static `pathbuf`.
//
// Non-reentrant: one buffer, every [[stat]] / [[open]] / etc.
// rewrites it. Same caveat as strconv's `*tos` family (overwritten
// on next call). Caller must NOT hold a kpath-returned pointer
// across another lib/os path call. Graduates when ww grows a
// thread story.
//
// `nil`-as-overflow over `(*u8 | oserror)`: wwstage over-allocates
// 1-word-payload tagged returns to 24B (cstage emits 16B).
// Task #9; revert at task #10 when fixed. Repro at
// .ai/probe_tagged_return_pointer_payload.ww.
export def PATH_MAX: i32 = 4096;
let pathbuf: [4096]u8;

// ref/hare/sys/+linux/types.ha:886-888. ww folds `sys` into `os`, so the
// std fd NUMBERS live here (the sys role). Typed i32, NOT io.file as in
// Hare's os::stdout_file (ref/hare/os/+linux/stdfd.ha:28): Hare's `os`
// imports `io`, but ww's `os` is the import floor and must never import
// io (lib/CLAUDE.md) — so the io.file/io.handle binding can't live here.
// Consumers (lib/fmt's stdio wrappers) cast i32→io.file at the use site,
// where the handle layer is already in scope.
export def STDIN_FILENO:  i32 = 0;
export def STDOUT_FILENO: i32 = 1;
export def STDERR_FILENO: i32 = 2;

fn kpath(p: str) *u8 = {
	if (p.len + 1 >= PATH_MAX) { return nil: *u8; };	// ENAMETOOLONG
	let i: i32 = 0;
	for (i < p.len) { pathbuf[i] = p[i]; i += 1; };
	pathbuf[p.len] = 0u8;
	return &pathbuf[0];
};

// Raw, non-fallible primitives. These return Linux's int conventions
// (negative = -errno, non-negative = bytes/fd/etc). Callers wanting a
// Hare-style fallible API use the wrappers below.
export fn write(fd: i32, buf: *u8, n: u64) i64 = {
	return syscall3(nr.WRITE, fd: i64, buf: i64, n: i64);
};

export fn read(fd: i32, buf: *u8, n: u64) i64 = {
	return syscall3(nr.READ, fd: i64, buf: i64, n: i64);
};

export fn close(fd: i32) i32 = {
	return syscall1(nr.CLOSE, fd: i64): i32;
};

// dup2(2): make `newfd` refer to the same description as `oldfd`,
// closing `newfd` first if open. Returns `newfd` on success or a
// negative errno. Used by w6c_ww to redirect stdout into an output
// file without changing the cgen emit path.
export fn dup2(oldfd: i32, newfd: i32) i32 = {
	return syscall2(nr.DUP2, oldfd: i64, newfd: i64): i32;
};

// Fallible wrappers. The error variant is `oserror` (an i64 carrying
// -errno). The sum type makes success/failure explicit and lets
// callers `?` the result up the stack.
export fn tryread(fd: i32, buf: *u8, n: u64) (i64 | oserror) = {
	let r: i64 = read(fd, buf, n);
	if (r < 0) { return r: oserror; };
	return r;
};

export fn trywrite(fd: i32, buf: *u8, n: u64) (i64 | oserror) = {
	let r: i64 = write(fd, buf, n);
	if (r < 0) { return r: oserror; };
	return r;
};

// open — Linux open(2). Returns -errno on failure, fd otherwise.
// Higher-level callers prefer `tryopen`. Mirrors Hare's os::open
// (ref/hare/os/os.ha:117); kpath lands the bytes in pathbuf.
// Returns -ENAMETOOLONG (-36) if the path overflows PATH_MAX.
export fn open(path: str, flags: flag, mode: i32) i32 = {
	let p: *u8 = kpath(path);
	if (p == nil: *u8) { return -36i32; };	// ENAMETOOLONG
	return syscall3(nr.OPEN, p: i64, (flags as i32): i64, mode: i64): i32;
};

export fn tryopen(path: str, flags: flag, mode: i32) (i32 | oserror) = {
	let fd: i32 = open(path, flags, mode);
	if (fd < 0) { return fd: i64: oserror; };
	return fd;
};

// lseek — set/inspect the fd's position. Returns the new offset or
// a negative errno. We use this for fstat-free file-size discovery
// (open ⇒ lseek to end ⇒ lseek back).
export fn lseek(fd: i32, off: i64, w: whence) i64 = {
	return syscall3(nr.LSEEK, fd: i64, off, (w as i32): i64);
};

// oserror — the underlying errno from a failed syscall, as a
// negative i64 (Linux's int convention; e.g. -2 = ENOENT). The
// `!`-flagged alias makes ?-propagation pick this variant as the
// error half of any (T | oserror) shape. Hare's analogue is
// errors::errno carried inside io::error.
export type oserror = !i64;

// errno — the raw Linux errno as a positive code (ref/hare/sys/+linux/
// errno.ha:5, `errno = !int`). ww folds Hare's `sys` role into os
// (lib/CLAUDE.md), so the sys::errno machinery lands here. Spelled i32
// rather than int: Linux errnos are kernel ints (32-bit), keeping os's
// kernel-facing surface uniformly i32. Distinct from [[oserror]] (!i64,
// the syscall's *negative* raw return) — the two model different
// things, so they are not unified; the negative→positive normalization
// lives at the oserror→errors.error boundary in those callers.
export type errno = !i32;

// Mapped errno values, ref/hare/sys/+linux/errno.ha:559-682. Positive,
// matching Hare's defs (the kernel returns -N; the wrap-to-positive is
// the caller's concern). Subset: exactly the errnos [[errors.errno]]
// maps to a named condition; grow as callers surface more.
export def ENOENT:       errno = 2;
export def EINTR:        errno = 4;
export def EAGAIN:       errno = 11;
export def EACCES:       errno = 13;
export def EBUSY:        errno = 16;
export def EEXIST:       errno = 17;
export def EINVAL:       errno = 22;
export def EOVERFLOW:    errno = 75;
export def ENETUNREACH:  errno = 101;
export def ETIMEDOUT:    errno = 110;
export def ECONNREFUSED: errno = 111;
export def ECANCELED:    errno = 125;

// strerror — human-readable text for an [[errno]] (Hare's
// sys::strerror, ref/hare/sys/+linux/errno.ha:18). FAITHFUL MINIMAL
// SUBSET: the mapped errnos above plus a generic fallback; grow the
// switch as callers surface more (lib/CLAUDE.md documented-subset, not
// a workaround). Messages verbatim from the reference. Hare's
// unknown_errno formats the numeric value; that is deferred.
export fn strerror(err: errno) str = {
	switch (err) {
	case ENOENT:       return "No such file or directory";
	case EINTR:        return "Interrupted system call";
	case EAGAIN:       return "Resource temporarily unavailable";
	case EACCES:       return "Permission denied";
	case EBUSY:        return "Device or resource busy";
	case EEXIST:       return "File exists";
	case EINVAL:       return "Invalid argument";
	case EOVERFLOW:    return "Value too large for defined data type";
	case ENETUNREACH:  return "Network is unreachable";
	case ETIMEDOUT:    return "Connection timed out";
	case ECONNREFUSED: return "Connection refused";
	case ECANCELED:    return "Operation canceled";
	};
	return "Unknown error";
};

// filesize — byte length of an open fd via lseek-to-end-and-back.
export fn filesize(fd: i32) (i64 | oserror) = {
	let end: i64 = lseek(fd, 0i64, whence.END);
	if (end < 0) { return end: oserror; };
	let r: i64 = lseek(fd, 0i64, whence.SET);
	if (r < 0) { return r: oserror; };
	return end;
};

// readall — keep reading until `n` bytes have arrived or the fd
// closes early. Hare name (io::readall); the buffer is caller-
// supplied, matching the Plan 9 subset convention.
export fn readall(fd: i32, buf: *u8, n: u64) (i64 | oserror) = {
	let got: u64 = 0u64;
	for (got < n) {
		let r: i64 = read(fd, buf + got, n - got);
		if (r < 0) { return r: oserror; };
		if (r == 0) { return got: i64; };	// short read: caller decides
		got += r: u64;
	};
	return got: i64;
};

// writeall — keep writing until `n` bytes have been accepted or the
// fd refuses progress. Hare name (io::writeall).
export fn writeall(fd: i32, buf: *u8, n: u64) (i64 | oserror) = {
	let sent: u64 = 0u64;
	for (sent < n) {
		let r: i64 = write(fd, buf + sent, n - sent);
		if (r < 0) { return r: oserror; };
		if (r == 0) { return sent: i64; };
		sent += r: u64;
	};
	return sent: i64;
};

// ---- process and filesystem helpers used by the `ww` driver ----------

// access(2): returns 0 if the file is reachable, negative errno
// otherwise. mode is the bitset described in <unistd.h> (F_OK=0).
// Mirrors Hare's os::access (ref/hare/os/+linux/fs.ha:access).
// Returns -ENAMETOOLONG (-36) if the path overflows PATH_MAX.
export fn access(path: str, mode: i32) i32 = {
	let p: *u8 = kpath(path);
	if (p == nil: *u8) { return -36i32; };
	return syscall2(nr.ACCESS, p: i64, mode: i64): i32;
};

// remove — unlink(2). Mirrors Hare's os::remove
// (ref/hare/os/os.ha:12).
export fn remove(path: str) i32 = {
	let p: *u8 = kpath(path);
	if (p == nil: *u8) { return -36i32; };
	return syscall1(nr.UNLINK, p: i64): i32;
};

// mkdir — mkdir(2). Mode is the unix permission bitset (e.g. 0o700).
// Returns 0 on success, negative errno otherwise. Mirrors Hare's
// os::mkdir (ref/hare/os/os.ha:50).
export fn mkdir(path: str, mode: i32) i32 = {
	let p: *u8 = kpath(path);
	if (p == nil: *u8) { return -36i32; };
	return syscall2(nr.MKDIR, p: i64, mode: i64): i32;
};

// rmdir — rmdir(2). Mirrors Hare's os::rmdir
// (ref/hare/os/os.ha:58).
export fn rmdir(path: str) i32 = {
	let p: *u8 = kpath(path);
	if (p == nil: *u8) { return -36i32; };
	return syscall1(nr.RMDIR, p: i64): i32;
};

// mkdirs — recursive mkdir. Creates `path` and any non-existent
// parent directories with the given mode. EEXIST is silently
// accepted (matches Hare's `errors::exists` skip in os::mkdirs);
// any other syscall failure surfaces as `oserror`.
//
// Mirrors Hare's os::mkdirs (ref/hare/os/os.ha:54). The in-place
// '/' → NUL splice walks the kpath-loaded [[pathbuf]] directly
// instead of recursing through [[mkdir]] — re-entering kpath would
// clobber the buffer mid-walk (single static slot, see kpath's
// non-reentrancy note above).
export fn mkdirs(path: str, mode: i32) (void | oserror) = {
	let cp: *u8 = kpath(path);
	if (cp == nil: *u8) { return -36i64: oserror; };
	let n: i32 = path.len;
	if (n == 0) { return; };

	// Walk forward; at each '/' boundary, NUL-terminate the prefix,
	// raw MKDIR syscall on pathbuf, restore the slash, continue.
	// Skip index 0 so a leading '/' on absolute paths doesn't
	// trigger an empty mkdir.
	let i: i32 = 1;
	for (i < n) {
		if (pathbuf[i] == 47u8) {	// '/'
			pathbuf[i] = 0u8;
			let r: i32 = syscall2(nr.MKDIR,
			    (&pathbuf[0]): i64, mode: i64): i32;
			pathbuf[i] = 47u8;
			if (r < 0) {
				if (r != -17) { return r: i64: oserror; };
			};
		};
		i += 1;
	};

	let r: i32 = syscall2(nr.MKDIR,
	    (&pathbuf[0]): i64, mode: i64): i32;
	if (r < 0) {
		if (r != -17) { return r: i64: oserror; };
	};
	return;
};

// getpid(2). Used by the driver to mint unique scratch paths.
export fn getpid() i32 = {
	return syscall0(nr.GETPID): i32;
};

// fork(2): 0 in the child, child pid in the parent, negative errno
// on failure.
export fn fork() i32 = {
	return syscall0(nr.FORK): i32;
};

// execve(2): on success, does not return. Mirrors Hare's
// os::exec::exec path arg (str). argv/envp stay `**u8` — the
// kernel takes a NUL-pointer-terminated table of NUL-terminated
// C strings, a different shape from a path.
export fn execve(path: str, argv: **u8, envp: **u8) i32 = {
	let p: *u8 = kpath(path);
	if (p == nil: *u8) { return -36i32; };
	return syscall3(nr.EXECVE, p: i64, argv: i64, envp: i64): i32;
};

// wait4(2): wait for `pid` (or any child if -1), store status in
// `*status`, return the pid that ended (or negative errno).
export fn wait4(pid: i32, status: *i32, options: i32, rusage: *void) i32 = {
	return syscall4(nr.WAIT4, pid: i64, status: i64,
	    options: i64, rusage: i64): i32;
};

// getcwd(2) — Linux flavour. Writes the NUL-terminated cwd into `buf`
// and returns the number of bytes written (including the NUL), or a
// negative errno. The driver uses it to expand `.` to the cwd's
// basename for `ww build` / `ww test`.
export fn getcwd(buf: *u8, n: u64) i64 = {
	return syscall2(nr.GETCWD, buf: i64, n: i64);
};

// getdents64(2) — Linux directory enumeration. The fd must be opened
// with O_RDONLY on a directory. `buf` receives a packed sequence of
// linux_dirent64 records:
//
//   struct linux_dirent64 {
//       u64  d_ino;       // 0..7
//       i64  d_off;       // 8..15
//       u16  d_reclen;    // 16..17  — total bytes for this record
//       u8   d_type;      // 18      — DT_REG/DT_DIR/...
//       u8   d_name[];    // 19..    — NUL-terminated name + padding
//   };
//
// Returns bytes written into `buf` (advance by d_reclen to walk),
// 0 at end-of-directory, or a negative errno.
export fn getdents64(fd: i32, buf: *u8, n: u64) i64 = {
	return syscall3(nr.GETDENTS64, fd: i64, buf: i64, n: i64);
};

// ---- environment ------------------------------------------------------

// rt_envp — runtime-side getter. rt/start.s captures envp into a DATAW
// slot before calling main; this binding lifts the captured pointer
// into ww. Same FFI shape as rt_syscall / rt_malloc / rt_abort: a TEXT
// symbol the linker resolves. The returned `**u8` is a NUL-terminated
// table of `*u8` entries, each pointing at a NUL-terminated
// "NAME=VALUE" byte sequence.
//
// We don't expose `rtenvp` directly; [[getenv]] is the only consumer.
@symbol("rt_envp") fn rtenvp() **u8;

// getenv — POSIX getenv. Returns a borrowed `str` view over the value
// bytes of the named environment variable, or void if the name is not
// present. The view is valid for the process lifetime — the bytes
// live in the kernel-supplied envp table at process entry. A future
// `setenv` (separate task) that grows the table behind the scenes
// would invalidate prior views; v1 has no setenv, so callers can
// hold the view indefinitely.
//
// Mirrors Hare's os::tryenv shape (returns void rather than panicking
// on missing). Hare also ships os::getenv (`(str | void)`) and
// os::mustenv (panic-on-missing); ww collapses to the single
// `(str | void)` form for now — consumers wanting "must" semantics
// abort at the call site.
//
// Algorithm: walk the NUL-pointer-terminated `environ` table doing a
// "name=" prefix match against each entry, byte-wise. NUL inside
// `name` would never match a real env var (env var names cannot
// contain '\0'), so we don't filter — POSIX puts that responsibility
// on the caller.
export fn getenv(name: str) (str | void) = {
	let envp: **u8 = rtenvp();
	let i: i32 = 0;
	for (true) {
		let entry: *u8 = envp[i];
		if (entry == nil: *u8) { return; };
		let j: i32 = 0;
		let matched: bool = true;
		for (j < name.len) {
			if (entry[j] == 0u8) { matched = false; break; };
			if (entry[j] != name[j]) { matched = false; break; };
			j += 1;
		};
		if (matched) {
			if (entry[name.len] == 61u8) {	// '='
				let val: *u8 = entry + ((name.len + 1): u64);
				let n: i32 = 0;
				for (val[n] != 0u8) { n += 1; };
				let r: str;
				r.ptr = val;
				r.len = n;
				return r;
			};
		};
		i += 1;
	};
	return;
};

// ---- stat / lstat / fstat / exists -----------------------------------
//
// Ports of Hare's stat family (ref/hare/fs/fs.ha:172,196 +
// ref/hare/sys/+linux/stat.ha:24-58). The Hare surface returns
// `filestat` by value; ww's cgreturn ABI tops out at 24B today (see
// STATUS task #21) and filestat is 80B, so [[stat]] / [[lstat]] /
// [[fstat]] take an out-parameter and return `(void | oserror)`.
// Re-evaluate the by-value shape when full sret lands.
//
// `filestat`, `mode`, and `stat_mask` live in lib/os because ww has
// no lib/fs yet; Hare puts them in `fs::`. These types graduate to
// lib/fs when that module ships — callers should expect a future
// re-export.
//
// Underlying syscall is SYS_newfstatat (262), which unifies
// stat/lstat/fstat through the `dirfd + flags` triple:
//   stat  = newfstatat(AT_FDCWD, path,  0)
//   lstat = newfstatat(AT_FDCWD, path,  AT_SYMLINK_NOFOLLOW)
//   fstat = newfstatat(fd,       "",    AT_EMPTY_PATH)
// Avoiding SYS_statx — its 256B variable layout would buy btime,
// but Hare's filestat doesn't expose btime either, so we stay on
// the simpler 144B kernel struct.

// fstatat(2) flag values. Linux constants from <linux/fcntl.h>.
// Names mirror Hare's ref/hare/sys/+linux/types.ha:45-51 (capital-
// AT_ prefix, top-level `def`s).
export def AT_FDCWD:            i32 = -100;
export def AT_SYMLINK_NOFOLLOW: i32 = 256;	// 0x100
export def AT_EMPTY_PATH:       i32 = 4096;	// 0x1000

// mode — file-mode bits. Mirrors Hare's fs::mode (ref/hare/fs/
// types.ha:63). Permission bits are the standard Unix octal subset;
// type bits live in the S_IFMT = 0o170000 region. Type-bit test:
//
//   let t: u32 = (fi.mode as u32) & 61440u32;	// 0o170000 mask
//   if (t == os.mode.DIR as u32) { /* directory */ };
//
// Numeric values are octal in Hare's source; ww has no octal
// literals so they're written as decimal with the octal in a
// trailing comment.
export type mode = enum u32 {
	// permission bits
	USER_RWX  = 448u32,	// 0o700
	USER_RW   = 384u32,	// 0o600
	USER_RX   = 320u32,	// 0o500
	USER_R    = 256u32,	// 0o400
	USER_W    = 128u32,	// 0o200
	USER_X    = 64u32,	// 0o100
	GROUP_RWX = 56u32,	// 0o070
	GROUP_RW  = 48u32,	// 0o060
	GROUP_RX  = 40u32,	// 0o050
	GROUP_R   = 32u32,	// 0o040
	GROUP_W   = 16u32,	// 0o020
	GROUP_X   = 8u32,	// 0o010
	OTHER_RWX = 7u32,	// 0o007
	OTHER_RW  = 6u32,	// 0o006
	OTHER_RX  = 5u32,	// 0o005
	OTHER_R   = 4u32,	// 0o004
	OTHER_W   = 2u32,	// 0o002
	OTHER_X   = 1u32,	// 0o001
	SETUID    = 2048u32,	// 0o4000
	SETGID    = 1024u32,	// 0o2000
	STICKY    = 512u32,	// 0o1000
	// file-type bits (S_IFMT mask = 0o170000 = 61440)
	UNKNOWN   = 0u32,
	FIFO      = 4096u32,	// 0o010000
	CHR       = 8192u32,	// 0o020000
	DIR       = 16384u32,	// 0o040000
	BLK       = 24576u32,	// 0o060000
	REG       = 32768u32,	// 0o100000
	LINK      = 40960u32,	// 0o120000
	SOCK      = 49152u32,	// 0o140000
};

// stat_mask — which filestat fields the call populated. Mirrors
// Hare's fs::stat_mask (ref/hare/fs/types.ha:129). newfstatat fills
// every field, so [[stat]] / [[lstat]] / [[fstat]] always set all
// seven bits OR-folded (see [[fillfilestat]]); per-bit testing is
// the documented sparse-backend pattern (cf. Hare's fs::fs network
// backends that only populate mtime+size).
export type stat_mask = enum u32 {
	UID   = 1u32,
	GID   = 2u32,
	SIZE  = 4u32,
	INODE = 8u32,
	ATIME = 16u32,
	MTIME = 32u32,
	CTIME = 64u32,
};

// filestat — Hare's fs::filestat (ref/hare/fs/types.ha:141). 80
// bytes. Times are time.instant (ref/hare/time/instant.ha:9) — the
// canonical Hare shape. See module-header note re: graduation to
// lib/fs.
export type filestat = struct {
	mask:  stat_mask,	// 0  (4)
	mode:  mode,		// 4  (4)
	uid:   u32,		// 8  (4)
	gid:   u32,		// 12 (4)
	sz:    u64,		// 16 (8)
	inode: u64,		// 24 (8)
	atime: time.instant,	// 32 (16)
	mtime: time.instant,	// 48 (16)
	ctime: time.instant,	// 64 (16) — ends at 80
};

// kstat — x86_64 kernel `struct stat` layout. Mirrors
// arch/x86/include/uapi/asm/stat.h (`__kernel_ulong_t`-keyed
// fields). 144 bytes. Module-internal; SYS_newfstatat writes into
// this buffer and the public stat fns then copy the bits into the
// Hare-shaped [[filestat]].
type kstat = struct {
	dev:        u64,	// 0
	ino:        u64,	// 8
	nlink:      u64,	// 16
	mode:       u32,	// 24
	uid:        u32,	// 28
	gid:        u32,	// 32
	pad0:       u32,	// 36
	rdev:       u64,	// 40
	sz:         i64,	// 48
	blksize:    i64,	// 56
	blocks:     i64,	// 64
	atime_sec:  i64,	// 72
	atime_nsec: i64,	// 80
	mtime_sec:  i64,	// 88
	mtime_nsec: i64,	// 96
	ctime_sec:  i64,	// 104
	ctime_nsec: i64,	// 112
	unused0:    i64,	// 120
	unused1:    i64,	// 128
	unused2:    i64,	// 136 — ends at 144
};

// emptypath — single-NUL byte used as the `pathname` arg to
// newfstatat with AT_EMPTY_PATH. The kernel requires a non-NULL
// pointer to a zero-length C string, NOT a null pointer. Bytes are
// read-only from the kernel's view; ww has no module-level const so
// this is a writable `let`.
let emptypath: [1]u8 = [0u8];

// fillfilestat — copy a 144B kstat into the 80B Hare-shaped
// filestat. Internal helper used by all three public entry points.
// Mirrors Hare's st_to_filestat (ref/hare/os/+linux/dirfdfs.ha:259):
// newfstatat populates every field, so the mask is the OR-fold of
// all seven Hare stat_mask bits.
fn fillfilestat(out: *filestat, k: *kstat) void = {
	out.mask        = stat_mask.UID | stat_mask.GID | stat_mask.SIZE
	    | stat_mask.INODE | stat_mask.ATIME | stat_mask.MTIME
	    | stat_mask.CTIME;
	out.mode        = k.mode: mode;
	out.uid         = k.uid;
	out.gid         = k.gid;
	out.sz          = k.sz: u64;
	out.inode       = k.ino;
	out.atime.sec   = k.atime_sec;
	out.atime.nsec  = k.atime_nsec;
	out.mtime.sec   = k.mtime_sec;
	out.mtime.nsec  = k.mtime_nsec;
	out.ctime.sec   = k.ctime_sec;
	out.ctime.nsec  = k.ctime_nsec;
};

// stat — fill *out with metadata for `path`. Follows symlinks.
// Returns ENAMETOOLONG (-36) as `oserror` if the path overflows
// PATH_MAX.
//
// Mirrors Hare's sys::stat (ref/hare/sys/+linux/stat.ha:51) modulo
// the out-param shape forced by the cgreturn 24B cap. Note: Hare's
// higher-level fs::stat (ref/hare/fs/fs.ha:172) instead has lstat
// semantics — we follow sys::stat's POSIX-stat behavior here.
export fn stat(out: *filestat, path: str) (void | oserror) = {
	let cp: *u8 = kpath(path);
	if (cp == nil: *u8) { return -36i64: oserror; };
	let k: kstat;
	let r: i64 = syscall4(nr.NEWFSTATAT,
	    AT_FDCWD: i64, cp: i64, (&k): i64, 0i64);
	if (r < 0) { return r: oserror; };
	fillfilestat(out, &k);
};

// lstat — like [[stat]] but does NOT follow a terminal symlink.
// Mirrors Hare's sys::lstat (ref/hare/sys/+linux/stat.ha:57).
export fn lstat(out: *filestat, path: str) (void | oserror) = {
	let cp: *u8 = kpath(path);
	if (cp == nil: *u8) { return -36i64: oserror; };
	let k: kstat;
	let r: i64 = syscall4(nr.NEWFSTATAT,
	    AT_FDCWD: i64, cp: i64, (&k): i64,
	    AT_SYMLINK_NOFOLLOW: i64);
	if (r < 0) { return r: oserror; };
	fillfilestat(out, &k);
};

// fstat — like [[stat]] but addresses the file by fd. Uses
// newfstatat(fd, "", AT_EMPTY_PATH); the kernel resolves the fd
// directly. Mirrors Hare's sys::fstat (ref/hare/sys/+linux/stat.ha:54).
export fn fstat(out: *filestat, fd: i32) (void | oserror) = {
	let k: kstat;
	let r: i64 = syscall4(nr.NEWFSTATAT,
	    fd: i64, (&emptypath[0]): i64, (&k): i64,
	    AT_EMPTY_PATH: i64);
	if (r < 0) { return r: oserror; };
	fillfilestat(out, &k);
};

// exists — true if `path` resolves to anything (regular file,
// directory, symlink, ...). Stat-shaped (Hare's `fs::exists`,
// ref/hare/fs/fs.ha:196) — no separate syscall. Symlinks are
// followed; a dangling symlink is `false`. ENAMETOOLONG is
// swallowed as `false` — Hare's os::exists doc says "true if a
// node exists at the given path, or false if not."
//
// Race warning: prefer "open and handle the error" over "exists
// then open" in real code (Hare's docstring carries the same
// note). The race is unavoidable in this shape.
//
// Goes through SYS_newfstatat directly rather than match'ing on
// [[stat]]'s `(void | oserror)` return. Functionally identical;
// the direct shape sidesteps a cstage/wwstage cgen disagreement
// on the slot size of `(void | oserror)` (cstage 16B, wwstage 24B
// — same class as STATUS #22, surfaced first time a match on this
// shape combined with an 80B local-struct local frame). Use the
// match shape once #22 lands.
export fn exists(path: str) bool = {
	let cp: *u8 = kpath(path);
	if (cp == nil: *u8) { return false; };
	let k: kstat;
	let r: i64 = syscall4(nr.NEWFSTATAT,
	    AT_FDCWD: i64, cp: i64, (&k): i64, 0i64);
	return r >= 0i64;
};

// rt — runtime primitives exposed to ww programs.
// Mirrors Hare's rt:: module placement (ref/hare/rt/).

package rt;

// malloc — mmap-backed page allocator. Untyped: `malloc(n)` returns a
// `*void`; callers cast to the target type. Diverges from Hare: Hare
// exposes `alloc` / `free` as typed language builtins that the
// compiler lowers to rt::malloc/rt::free; ww has no such builtins,
// so the rt-symbol surface is exposed directly. Stdlib callers that
// need a typed allocation pattern wrap this with a cast plus a stored
// capacity (see [[strings.dup]], [[memio.dynamic]]).
//
// OOM: rt_malloc is a bare mmap(MAP_ANON|MAP_PRIVATE) wrapper with no
// error path. The raw Linux mmap syscall returns a negative errno cast
// to `*void` on failure (e.g. `(void*)-12` for ENOMEM); the
// `MAP_FAILED` (`(void*)-1`) value is a libc-wrapper convention that
// rt_malloc doesn't apply. Neither `== nil` nor `== (void*)-1` catches
// it; any deref of such a return faults. Today the stdlib does not
// check; OOM faults on first dereference. A typed fallible variant is
// a future task (task #39). ref/hare/rt/malloc.ha:27.
@symbol("rt_malloc") export fn malloc(n: u64) *void;

// selfhost/cmd/w6l/sym.ww — port of cmd/w6l/sym.c.
//
// Linker symbol table. Singly-linked list, usually a few hundred
// entries; hashing isn't worth it yet.

package w6l;

type lsym = struct {
	name:        str,
	val:         u64,	// offset within combined .text (or .data when
				// indata=1) once linked
	defined:     i32,	// 1 if some lobj defines this symbol
	indata:      i32,	// 1 if defined in .data (writable globals)
	owner:       *lobj,
	idxinowner:  i32,
	// Dynamic-linking fields. Set by resolve when an undefined sym
	// is provided by some loaded lso. pltidx and dynsymidx default
	// to -1 (set explicitly by resolve; alloc-zeroing gives 0, not -1).
	isdyn:       i32,
	dynlib:      *lso,
	dynversion:  str,	// matched export's version; len 0 if none
	pltidx:      i32,
	dynsymidx:   i32,
	snext:       *lsym,
};

type lrel = struct {
	off:    u64,	// offset within the relocation's section
	section: i32,	// 0 = .text, 1 = .data
	kind:   i32,	// R_X86_64_*
	sym:    *lsym,
	addend: i64,
	rnext:  *lrel,
};

type lobj = struct {
	path:     str,
	buf:      *u8,	// object bytes
	len:      u64,
	textoff:  u64,	// offset of .text in combined output
	textsize: u64,
	dataoff:  u64,	// offset of .data in combined output
	datasize: u64,	// bytes contributed to combined .data (0 if none)
	onext:    *lobj,
};

// lexport — one entry per GLOBAL/WEAK symbol exported by a loaded .so.
// Stored as a chain in the order the .so's dynsym presents them, so
// soprovides_v's first-match semantics agree with the C version.
type lexport = struct {
	name:    str,
	version: str,	// len 0 for unversioned globals
	enext:   *lexport,
};

type lso = struct {
	path:    str,	// full filesystem path used to load
	soname:  str,	// DT_SONAME, or basename if missing
	exports: *lexport,	// dynsym-order chain of exported names
	sonext:  *lso,
};

type lnk = struct {
	objs:    *lobj,
	sos:     *lso,
	syms:    *lsym,
	rels:    *lrel,
	text:    *u8,	// combined .text
	textcap: u64,
	textlen: u64,
	// Combined .data (writable). Empty unless any input .o has a
	// .data PROGBITS section.
	data:    *u8,
	datacap: u64,
	datalen: u64,
	errs:    i32,
	dynn:    i32,	// number of syms routed through PLT
};

fn streq(a: str, b: str) bool = {
	if (a.len != b.len) { return false; };
	let i: i32 = 0;
	for (i < a.len) {
		if (a[i] != b[i]) { return false; };
		i += 1;
	};
	return true;
};

export fn intern(l: *lnk, name: str) *lsym = {
	let s: *lsym = l.syms;
	for (s != nil) {
		if (streq(s.name, name)) { return s; };
		s = s.snext;
	};
	let n: *lsym = alloc(lsym { name = name, snext = l.syms })!;
	l.syms = n;
	return n;
};

export fn lookup(l: *lnk, name: str) *lsym = {
	let s: *lsym = l.syms;
	for (s != nil) {
		if (streq(s.name, name)) { return s; };
		s = s.snext;
	};
	return nil;
};

// types — integer limits. Mirrors Hare's types::limits (I8_MAX, …)
// platform-fixed for amd64. Numeric helpers live in lib/math, matching
// Hare's split between types::limits and math::.

package types;

def I8_MAX:  i8  = 127;
def I16_MAX: i16 = 32767;
def I32_MAX: i32 = 2147483647;
def I64_MAX: i64 = 9223372036854775807;

def I8_MIN:  i8  = -128;
def I16_MIN: i16 = -32768;
def I32_MIN: i32 = -2147483648;
def I64_MIN: i64 = -9223372036854775808;

def U8_MAX:  u8  = 255;
def U16_MAX: u16 = 65535;
def U32_MAX: u32 = 4294967295;
def U64_MAX: u64 = 18446744073709551615;

def U8_MIN:  u8  = 0;
def U16_MIN: u16 = 0;
def U32_MIN: u32 = 0;
def U64_MIN: u64 = 0;

// int/uint are machine-word (Go-style, type.c:58); limits derived from
// size(int) per #114 + user ruling; cf Go math.MaxInt; diverges from
// Hare's per-arch literal (arch+x86_64.ha) because ww's int is 64-bit.
def INT_MAX:  int  = (1 << (size(int)*8 - 1)) - 1;
def INT_MIN:  int  = -1 << (size(int)*8 - 1);
def UINT_MIN: uint = 0;
def UINT_MAX: uint = ~(0: uint);

// size is 8B on amd64; no cast needed (size ∈ unsigned class per #113).
def SIZE_MIN: size = U64_MIN;
def SIZE_MAX: size = U64_MAX;

// uintptr not in the unsigned class, so the cast is required (Hare's form).
def UINTPTR_MIN: uintptr = U64_MIN: uintptr;
def UINTPTR_MAX: uintptr = U64_MAX: uintptr;

def RUNE_MIN: rune = '\0';

// bytes — slice operations over []u8. Mirrors Hare's bytes module
// (ref/hare/bytes/) for the in-tree subset: search/equality/prefix
// helpers used by lib/encoding, lib/bufio, lib/memio.
//
// Documented divergences from Hare:
//   - index_slice / rindex_slice use naive O(n·m); Hare specialises
//     2/3/4-byte needles and falls back to two_way (Crochemore-Perrin)
//     for longer (ref/hare/bytes/index.ha:61, ref/hare/bytes/two_way.ha).
//     Correctness equivalent.
//   - peek_token dispatches index/rindex by branching on `reverse`
//     rather than a function-pointer `ifunc` (ref/hare/bytes/tokenize.ha:97).
//     ww has no fn pointers in scope yet — same pattern as lib/strings
//     `move`. Outwardly identical.
//   - tokenize / rtokenize zero the `delim` field on the constructed
//     tokenizer when `in` is empty, rather than mutating the variadic
//     param before the struct write (ref/hare/bytes/tokenize.ha:26-28).
//     Semantically identical; the variadic param is borrowed and
//     captured-by-value into the struct, so mutating either side
//     yields the same observable state.

package bytes;

import os;
import types;

// done — iteration sentinel returned by next_token / peek_token at
// end-of-input. ref/hare/bytes/tokenize.ha uses the built-in `done`
// token; ww spells it per-package the same way lib/encoding/utf8 does
// (utf8.ww:36). Plain `void` (not `!void`): continuation signal.
export type done = void;

// tokenizer — cursor over an input slice. Layout mirrors
// ref/hare/bytes/tokenize.ha:6-10. `p` is the cached peek-position;
// I64_MAX (forward) / I64_MIN (reverse) are the unprimed sentinels.
// p < 0 also identifies a reverse-direction iterator.
export type tokenizer = struct {
	in:    []u8,
	delim: []u8,
	p:     i64,
};

// equal — true iff `a` and `b` have the same length and contents.
// ref/hare/bytes/equal.ha:9.
export fn equal(a: []u8, b: []u8) bool = {
	if (a.len != b.len) { return false; };
	let i: i32 = 0;
	for (i < a.len) {
		if (a[i] != b[i]) { return false; };
		i += 1;
	};
	return true;
};

// index — first offset of `needle` in `s`. u8 needle scans for the
// byte; []u8 needle scans for the substring. void if absent.
// ref/hare/bytes/index.ha:6.
export fn index(s: []u8, needle: (u8 | []u8)) (i32 | void) = {
	match (needle) {
	case let c: u8 => {
		let i: i32 = 0;
		for (i < s.len) {
			if (s[i] == c) { return i; };
			i += 1;
		};
		return;
	};
	case let sub: []u8 => {
		if (sub.len == 0) { return 0; };
		if (sub.len > s.len) { return; };
		let last: i32 = s.len - sub.len;
		let i: i32 = 0;
		for (i <= last) {
			let j: i32 = 0;
			let ok: bool = true;
			for (j < sub.len) {
				if (s[i + j] != sub[j]) { ok = false; j = sub.len; }
				else { j += 1; };
			};
			if (ok) { return i; };
			i += 1;
		};
		return;
	};
	};
	return;
};

// rindex — last offset of `needle` in `s`. Empty []u8 needle returns
// s.len (ref/hare/bytes/index.ha:103 — Hare's loop yields r-0 at i=0).
// ref/hare/bytes/index.ha:86.
export fn rindex(s: []u8, needle: (u8 | []u8)) (i32 | void) = {
	match (needle) {
	case let c: u8 => {
		let i: i32 = s.len - 1;
		for (i >= 0) {
			if (s[i] == c) { return i; };
			i -= 1;
		};
		return;
	};
	case let sub: []u8 => {
		if (sub.len == 0) { return s.len; };
		if (sub.len > s.len) { return; };
		let i: i32 = s.len - sub.len;
		for (i >= 0) {
			let j: i32 = 0;
			let ok: bool = true;
			for (j < sub.len) {
				if (s[i + j] != sub[j]) { ok = false; j = sub.len; }
				else { j += 1; };
			};
			if (ok) { return i; };
			i -= 1;
		};
		return;
	};
	};
	return;
};

// contains — true iff any of `needles` (byte or sub-slice) appears in `s`.
// ref/hare/bytes/contains.ha:6.
export fn contains(s: []u8, needles: (u8 | []u8)...) bool = {
	let i: i32 = 0;
	for (i < needles.len) {
		match (needles[i]) {
		case let b: u8   => {
			match (index(s, b)) {
			case let bo: i32 => return true;
			case void        => void;
			};
		};
		case let n: []u8 => {
			match (index(s, n)) {
			case let bo: i32 => return true;
			case void        => void;
			};
		};
		};
		i += 1;
	};
	return false;
};

// ltrim — borrowed view of `in` with leading bytes in `trim` stripped.
// `trim` must be non-empty. ref/hare/bytes/trim.ha:7.
export fn ltrim(in: []u8, trim: u8...) []u8 = {
	os.assert(trim.len > 0, "bytes.ltrim called with empty trim set");
	let i: i32 = 0;
	for (i < in.len && contains(trim, in[i])) { i += 1; };
	let r: []u8;
	r.ptr = in.ptr + (i: u64);
	r.len = in.len - i;
	r.cap = r.len;
	return r;
};

// rtrim — borrowed view of `in` with trailing bytes in `trim` stripped.
// `trim` must be non-empty. ref/hare/bytes/trim.ha:17. Hare's loop uses
// `size` underflow at i==0 to terminate; ww indices are signed i32, so
// the equivalent termination is spelled `i >= 0` explicitly.
export fn rtrim(in: []u8, trim: u8...) []u8 = {
	os.assert(trim.len > 0, "bytes.rtrim called with empty trim set");
	let i: i32 = in.len - 1;
	for (i >= 0 && contains(trim, in[i])) { i -= 1; };
	let r: []u8;
	r.ptr = in.ptr;
	r.len = i + 1;
	r.cap = r.len;
	return r;
};

// trim — borrowed view of `in` with both ends in `trim` stripped.
// ref/hare/bytes/trim.ha:27.
export fn trim(in: []u8, trim: u8...) []u8 = {
	return ltrim(rtrim(in, trim...), trim...);
};

// hasprefix — true iff `s` starts with `pre`.
// ref/hare/bytes/contains.ha:21.
export fn hasprefix(s: []u8, pre: []u8) bool = {
	if (pre.len > s.len) { return false; };
	let i: i32 = 0;
	for (i < pre.len) {
		if (s[i] != pre[i]) { return false; };
		i += 1;
	};
	return true;
};

// hassuffix — true iff `s` ends with `suf`.
// ref/hare/bytes/contains.ha:35.
export fn hassuffix(s: []u8, suf: []u8) bool = {
	if (suf.len > s.len) { return false; };
	let off: i32 = s.len - suf.len;
	let i: i32 = 0;
	for (i < suf.len) {
		if (s[off + i] != suf[i]) { return false; };
		i += 1;
	};
	return true;
};

// reverse — in-place reverse of `s`. ref/hare/bytes/reverse.ha:5.
export fn reverse(s: []u8) void = {
	let i: i32 = 0;
	let j: i32 = s.len - 1;
	for (i < j) {
		let t: u8 = s[i];
		s[i] = s[j];
		s[j] = t;
		i += 1;
		j -= 1;
	};
};

// zero — set every byte of `s` to 0. ref/hare/bytes/zero.ha:5.
export fn zero(s: []u8) void = {
	let i: i32 = 0;
	for (i < s.len) {
		s[i] = 0u8;
		i += 1;
	};
};

// tokenize — iterator yielding tokens from `in` separated by any byte
// in `delim`. Leading / trailing / adjacent delims yield empty tokens.
// `delim` is borrowed; caller keeps it valid for the tokenizer's
// lifetime. ref/hare/bytes/tokenize.ha:22.
export fn tokenize(in: []u8, delim: u8...) tokenizer = {
	os.assert(delim.len > 0, "bytes.tokenize called with empty slice");
	os.assert((in.len: i64) < types.I64_MAX,
		"bytes.tokenize: input length exceeds I64_MAX");
	let t: tokenizer;
	t.in = in;
	t.delim = delim;
	if (in.len == 0) {
		t.delim.len = 0;
		t.delim.cap = 0;
	};
	t.p = types.I64_MAX;
	return t;
};

// rtokenize — reverse-direction tokenize. First next_token yields the
// last token, last next_token yields the first. ref/hare/bytes/tokenize.ha:40.
export fn rtokenize(in: []u8, delim: u8...) tokenizer = {
	os.assert(delim.len > 0, "bytes.rtokenize called with empty slice");
	os.assert((in.len: i64) < types.I64_MAX,
		"bytes.rtokenize: input length exceeds I64_MAX");
	let t: tokenizer;
	t.in = in;
	t.delim = delim;
	if (in.len == 0) {
		t.delim.len = 0;
		t.delim.cap = 0;
	};
	t.p = types.I64_MIN;
	return t;
};

// peek_token — next token without advancing the cursor. Returns done
// once `s.delim` has been zeroed by a prior past-end next_token.
// ref/hare/bytes/tokenize.ha:91.
export fn peek_token(s: *tokenizer) ([]u8 | done) = {
	if (s.delim.len == 0) {
		let d: done; return d;
	};

	let reverse: bool = s.p < 0i64;
	let known: bool = false;
	if (reverse) {
		if (s.p != types.I64_MIN) { known = true; };
	} else {
		if (s.p != types.I64_MAX) { known = true; };
	};
	if (!known) {
		let i: i64 = types.I64_MAX;
		if (reverse) { i = types.I64_MIN; };
		let dlen: i64 = 0i64;
		let slen: i64 = s.in.len: i64;

		let k: i32 = 0;
		for (k < s.delim.len) {
			let d: u8 = s.delim[k];
			let ix_found: bool = false;
			let ix_val: i32 = 0;
			if (reverse) {
				match (rindex(s.in, d)) {
				case let v: i32 => { ix_found = true; ix_val = v; };
				case void       => void;
				};
			} else {
				match (index(s.in, d)) {
				case let v: i32 => { ix_found = true; ix_val = v; };
				case void       => void;
				};
			};
			if (ix_found) {
				if (!reverse) {
					if ((ix_val: i64) < i) { i = ix_val: i64; dlen = 1i64; };
				} else {
					if ((ix_val: i64) > i) { i = ix_val: i64; dlen = 1i64; };
				};
			} else {
				if (!reverse) {
					if (slen < i) { i = slen; };
				} else {
					if (0i64 > i) { i = 0i64; };
				};
			};
			k += 1;
		};

		if (reverse) {
			if (i == slen) {
				s.p = -(slen + 1i64);
			} else {
				s.p = i + dlen - slen - 1i64;
			};
		} else {
			s.p = i;
		};
	};

	let r: []u8;
	if (reverse) {
		let start: i32 = (s.in.len: i64 + s.p + 1i64): i32;
		r.ptr = s.in.ptr + (start: u64);
		r.len = s.in.len - start;
		r.cap = r.len;
	} else {
		let end: i32 = s.p: i32;
		r.ptr = s.in.ptr;
		r.len = end;
		r.cap = end;
	};
	return r;
};

// next_token — current token, then advance past it and the delim.
// Once the input is exhausted, returns done and zeros `s.delim` so
// subsequent peeks short-circuit. ref/hare/bytes/tokenize.ha:59.
export fn next_token(s: *tokenizer) ([]u8 | done) = {
	let b: []u8;
	match (peek_token(s)) {
	case let v: []u8 => { b = v; };
	case done        => { let d: done; return d; };
	};

	let slen: i64 = s.in.len: i64;
	let reverse: bool = s.p < 0i64;
	if (reverse) {
		if (slen + s.p + 1i64 == 0i64) {
			s.delim.len = 0;
			s.delim.cap = 0;
			s.in.len = 0;
			s.in.cap = 0;
		} else {
			let end: i32 = (slen + s.p + 1i64 - 1i64): i32;
			s.in.len = end;
			s.in.cap = end;
		};
		s.p = types.I64_MIN;
	} else {
		if (s.p == slen) {
			s.delim.len = 0;
			s.delim.cap = 0;
			s.in.len = 0;
			s.in.cap = 0;
		} else {
			let adv: u64 = (s.p: u64) + 1u64;
			let adv_i32: i32 = (s.p: i32) + 1;
			s.in.ptr = s.in.ptr + adv;
			s.in.len = s.in.len - adv_i32;
			s.in.cap = s.in.cap - adv_i32;
		};
		s.p = types.I64_MAX;
	};
	return b;
};

// remaining_tokens — the unconsumed portion of `s.in`. Read-only view.
// ref/hare/bytes/tokenize.ha:145.
export fn remaining_tokens(s: *tokenizer) []u8 = {
	return s.in;
};

// rt_ensure is the runtime slice-growth helper invoked by the
// `append(s, v)` builtin. We bind it directly because the builtin's
// expansion stores only 8 bytes of the new element (cgen emits a
// single MOVQ), losing the .len/.cap fields of a []u8 element (24B).
// Mirrors the same workaround in lib/shlex.shlex (appendstr, 16B) and
// lib/getopt.getopt (appendoption, 24B); collapses in one go when the
// append builtin learns to store the full element width.
@symbol("rt_ensure") fn rtensure(s: *void, membsz: u64) void;

// appendslice — grow `*slice` by one and store `item` (24B). Mirror
// of [[shlex.appendstr]] / [[getopt.appendoption]]. Bypasses the
// `append` builtin's first-8B-only-store gap for a slice-element.
fn appendslice(slice: *[][]u8, item: []u8) void = {
	let newlen: i32 = slice.len + 1;
	slice.len = newlen;
	rtensure(slice: *void, 24u64);
	let dst: *[]u8 = &slice.ptr[newlen - 1];
	dst.ptr = item.ptr;
	dst.len = item.len;
	dst.cap = item.cap;
};

// splitn — split `in` on any byte in `delim`, returning up to `n`
// tokens via forward iteration. The trailing slot (when more than
// `n - 1` tokens exist) holds the unconsumed remainder.
//
// The caller frees the returned slice via
// `os.free(r.ptr: *void, (r.cap: u64) * 24u64)`. Element bytes are
// borrowed from `in`.
//
// Hare's `([][]u8 | nomem)` collapses to `[][]u8` here: ww os.alloc
// has no recoverable failure path. Same precedent as
// shlex.split / getopt.tryparse.
//
// ref/hare/bytes/tokenize.ha:156.
export fn splitn(in: []u8, delim: []u8, n: i32) [][]u8 = {
	os.assert(delim.len > 0,
		"bytes.splitn must not be called with an empty delimiter");
	let toks: [][]u8;
	toks.ptr = nil: *[]u8;
	toks.len = 0;
	toks.cap = 0;
	let tok: tokenizer = tokenize(in, delim...);
	let i: i32 = 0;
	for (i < n - 1) {
		match (next_token(&tok)) {
		case let s: []u8 => { appendslice(&toks, s); };
		case done        => { return toks; };
		};
		i += 1;
	};
	match (peek_token(&tok)) {
	case done         => void;
	case let pk: []u8 => {
		let r: []u8 = remaining_tokens(&tok);
		appendslice(&toks, r);
	};
	};
	return toks;
};

// rsplitn — reverse-direction counterpart to [[splitn]]: tokens are
// collected from the end of `in`. The trailing slot holds the
// unconsumed prefix (everything before the n-th-from-last delim hit).
//
// When the input has fewer than n tokens, the `done` short-circuit
// returns toks UN-reversed (in last-token-first order). Mirrors Hare
// at ref/hare/bytes/tokenize.ha:196-199 where the in-place reverse
// step is gated behind the n-1 loop running to completion. Only the
// "loop ran to completion AND peek saw a remainder" path applies the
// reverse; both early-exit paths skip it.
//
// ref/hare/bytes/tokenize.ha:186.
export fn rsplitn(in: []u8, delim: []u8, n: i32) [][]u8 = {
	os.assert(delim.len > 0,
		"bytes.rsplitn called with empty delimiter");
	let toks: [][]u8;
	toks.ptr = nil: *[]u8;
	toks.len = 0;
	toks.cap = 0;
	let tok: tokenizer = rtokenize(in, delim...);
	let i: i32 = 0;
	for (i < n - 1) {
		match (next_token(&tok)) {
		case let s: []u8 => { appendslice(&toks, s); };
		case done        => { return toks; };
		};
		i += 1;
	};
	match (peek_token(&tok)) {
	case done         => void;
	case let pk: []u8 => {
		let r: []u8 = remaining_tokens(&tok);
		appendslice(&toks, r);
	};
	};

	// In-place reverse so callers see argv-order, matching Hare
	// (ref/hare/bytes/tokenize.ha:207). Element copy is field-wise
	// through `*[]u8` because `toks[i] = toks[j]` (full 24B slice
	// store) lands in the multi-word-store gap noted at
	// cmd/w6c/cgen.c:6515-6523.
	let a: i32 = 0;
	let b: i32 = toks.len - 1;
	for (a < b) {
		let pa: *[]u8 = &toks.ptr[a];
		let pb: *[]u8 = &toks.ptr[b];
		let tp: *u8 = pa.ptr;
		let tl: i32 = pa.len;
		let tc: i32 = pa.cap;
		pa.ptr = pb.ptr;
		pa.len = pb.len;
		pa.cap = pb.cap;
		pb.ptr = tp;
		pb.len = tl;
		pb.cap = tc;
		a += 1;
		b -= 1;
	};
	return toks;
};

// split — full split of `in` on `delim` (no token cap). Mirrors
// `splitn(in, delim, types::SIZE_MAX)`. ww uses `types.I32_MAX`
// because the index type is i32 (lib/CLAUDE.md).
//
// ref/hare/bytes/tokenize.ha:225.
export fn split(in: []u8, delim: []u8) [][]u8 = {
	return splitn(in, delim, types.I32_MAX);
};

// cut — split `in` along the first instance of `delim`, returning the
// portion before and the portion after the delimiter as a borrowed
// tuple. When `delim` is absent, the whole input is the first half and
// the second is empty. ref/hare/bytes/tokenize.ha:392.
//
// Delim is spelled (u8 | []u8) to match index/rindex (bytes.ww:57/91);
// the tagged union is an unordered set, so this is the same type as
// Hare's ([]u8 | u8), not a divergence.
export fn cut(in: []u8, delim: (u8 | []u8)) ([]u8, []u8) = {
	let ln: i32 = match (delim) {
	case let c: u8     => yield 1i32;
	case let sub: []u8 => {
		os.assert(sub.len > 0,
			"bytes.cut called with empty delimiter");
		yield sub.len;
	};
	};
	match (index(in, delim)) {
	case let i: i32 => {
		let lo: i32 = i + ln;
		return (in[0:i], in[lo:in.len]);
	};
	case void => {
		let empty: []u8;
		empty.ptr = nil; empty.len = 0; empty.cap = 0;
		return (in, empty);
	};
	};
};

// rcut — like [[cut]] but splits along the last instance of `delim`.
// ref/hare/bytes/tokenize.ha:413.
export fn rcut(in: []u8, delim: (u8 | []u8)) ([]u8, []u8) = {
	let ln: i32 = match (delim) {
	case let c: u8     => yield 1i32;
	case let sub: []u8 => {
		os.assert(sub.len > 0,
			"bytes.rcut called with empty delimiter");
		yield sub.len;
	};
	};
	match (rindex(in, delim)) {
	case let i: i32 => {
		let lo: i32 = i + ln;
		return (in[0:i], in[lo:in.len]);
	};
	case void => {
		let empty: []u8;
		empty.ptr = nil; empty.len = 0; empty.cap = 0;
		return (in, empty);
	};
	};
};

// encoding/utf8 — UTF-8 encode/decode. Hare port; see
// ref/hare/encoding/utf8/{types,rune,encode,decode,decodetable}.ha.
//
// The decoder is Hoehrmann's branchless DFA, originally published
// at <https://bjoern.hoehrmann.de/utf-8/decoder/dfa/>. Hare's
// ref/hare/encoding/utf8/decodetable.ha:4 restructures Hoehrmann's
// flat table to 2D `[8][256]i8`; we flatten back to 1D `[2048]i8`
// because ww cgen does not yet ship 2D arrays (task #20).
//
// Surface deviation from ref/hare/encoding/utf8:
//
//   - `encoderune` takes a caller-supplied `out: []u8` and returns
//     the byte count. Hare returns a slice into a `static let buf`;
//     the caller-buffer form skips the static-buffer/slice-return pair.
//
// Deferred (no in-tree caller, follow-up tasks): `appendrune`,
// `strencode`, `strdecode`. Hare's string-iteration surface
// (`strings::iterator`/`strings::next` — ref/hare/strings/iter.ha)
// lives under lib/strings, not here.

// ref/hare/encoding/utf8/types.ha:6 — incomplete trailing sequence.
// Plain `void` (not `!void`): a truncated tail is a control-flow
// signal, not an error caller can ignore.
package utf8;

export type more = void;

// ref/hare/encoding/utf8/types.ha:9 — invalid UTF-8 sequence.
export type invalid = !void;

// ref/hare/encoding/utf8/types.ha:12 — fixed message; `invalid` carries
// no payload, so the rendering is constant.
export fn strerror(err: invalid) str = {
	return "Invalid UTF-8";
};

// `done` is not a built-in singleton in ww (Hare ships it as part of
// the type system). Plain `void` (not `!void`): end-of-input is a
// continuation signal, not an error. lib/io spells its EOF the same
// way (lib/io/io.ww:8-11).
export type done = void;

// ref/hare/encoding/utf8/decodetable.ha:4 — Hoehrmann's UTF-8 DFA,
// flat 1D `[2048]i8`. Layout: dfa[state*256 + byte] gives the next
// state (>0), the accept transition (0 — emit rune), or invalid (-1).
// Values match ref/hare/encoding/utf8/decodetable.ha verbatim.
let dfa: [2048]i8 = [
	// state 0 — initial byte: ASCII accepts (0), continuation/illegal
	// byte rejects (-1), legal multibyte start emits a state.
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	3i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 4i8, 2i8, 2i8,
	5i8, 6i8, 6i8, 6i8, 7i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 1 — expecting one continuation byte (0x80..0xBF).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 2 — expecting one continuation byte (full 0x80..0xBF range).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 3 — first byte was 0xE0; continuation byte must be 0xA0..0xBF
	// (rejects overlong 3-byte encodings).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 4 — first byte was 0xED; continuation byte must be 0x80..0x9F
	// (rejects UTF-16 surrogate codepoints U+D800..U+DFFF).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 5 — first byte was 0xF0; continuation byte must be 0x90..0xBF
	// (rejects overlong 4-byte encodings).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 6 — middle continuation byte of a 4-byte sequence (0x80..0xBF).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 7 — first byte was 0xF4; continuation byte must be 0x80..0x8F
	// (rejects codepoints above U+10FFFF).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
];

// ref/hare/encoding/utf8/decode.ha:17 — payload-bit masks. Hare's
// [2][8]u8 flattened to 1D [16]u8; row 0 (offsets 0..7) is the
// continuation-byte mask (always 0x3F), row 1 (offsets 8..15) is the
// initial-byte payload mask indexed by the transition class.
let masks: [16]u8 = [
	0x3fu8, 0x3fu8, 0x3fu8, 0x3fu8, 0x3fu8, 0x3fu8, 0x3fu8, 0x3fu8,
	0x7fu8, 0x1fu8, 0x0fu8, 0x0fu8, 0x0fu8, 0x07u8, 0x07u8, 0x07u8,
];

// ref/hare/encoding/utf8/decode.ha:6 — incremental decoder state.
export type decoder = struct {
	offs: i32,
	src:  []u8,
};

// ref/hare/encoding/utf8/decode.ha:12.
export fn decode(src: []u8) decoder = {
	let d: decoder;
	d.src = src;
	d.offs = 0;
	return d;
};

// ref/hare/encoding/utf8/decode.ha:27. Returns the next rune from a
// decoder, `done` at end-of-input, `more` on truncated trailing
// sequence, `invalid` on malformed input (overlong, surrogate,
// out-of-range, bad continuation).
//
// Algorithm is verbatim Hoehrmann (see file header). One structural
// rewrite: Hare encodes the "initial vs continuation byte" decision
// as the branchless `(state - 1): uint >> 31`, which assumes a 32-bit
// uint. ww's uint is 64-bit (cmd/wcc/type.c:58), so the shift answer
// would be 0x1_ffff_ffff rather than 1. We spell the same predicate
// with an explicit conditional.
export fn next(d: *decoder) (rune | done | more | invalid) = {
	if (d.offs == d.src.len) {
		let dn: done; return dn;
	};
	let nx: i32 = 0;
	let state: i32 = 0;
	let r: u32 = 0u32;
	for (d.offs < d.src.len) {
		let b: u8 = d.src[d.offs];
		let bi: i32 = b: i32;
		let row: i32 = state * 256 + bi;
		let cell: i8 = dfa[row];
		nx = cell: i32;
		let mi: i32 = 0;
		if (state == 0) { mi = 1; };
		let m: u8 = masks[mi * 8 + (nx & 7)];
		r = (r << 6u32) | ((b & m): u32);
		if (nx <= 0) {
			d.offs += 1;
			if (nx == 0) { return r: rune; };
			let e: invalid; return e;
		};
		state = nx;
		d.offs += 1;
	};
	let mr: more; return mr;
};

// ref/hare/encoding/utf8/decode.ha:207. Strict whole-input check.
// The hot path: tight DFA loop, no rune assembly. Bails the moment
// the table returns -1 so malformed inputs don't pay for the rest
// of the buffer.
export fn validate(src: []u8) (void | invalid) = {
	let state: i32 = 0;
	let i: i32 = 0;
	for (i < src.len) {
		if (state < 0) { break; };
		let bi: i32 = src[i]: i32;
		let cell: i8 = dfa[state * 256 + bi];
		state = cell: i32;
		i += 1;
	};
	if (state == 0) { return; };
	let e: invalid; return e;
};

// ref/hare/encoding/utf8/rune.ha:5. Encoded byte length of `r` as
// UTF-8. Callers in ww use this to size the buffer they hand to
// [[encoderune]]; values >0x10FFFF or negative are not legal Unicode
// codepoints and Hare aborts on them in `encoderune` itself, so we
// keep `runesz` infallible (matches Hare).
export fn runesz(r: rune) i32 = {
	let ch: u32 = r: u32;
	if (ch < 128u32)    { return 1; };
	if (ch < 2048u32)   { return 2; };
	if (ch < 65536u32)  { return 3; };
	return 4;
};

// ref/hare/encoding/utf8/rune.ha:15. Expected byte length of the
// codepoint that starts with `c`, or `invalid` if `c` cannot start
// a legal UTF-8 sequence. Constants written in decimal because ww
// doesn't accept Hare's `0b1000_0000` binary syntax: 0x80=128,
// 0xC2=194, 0xE0=224, 0xF0=240, 0xF8=248.
export fn utf8sz(c: u8) (i32 | invalid) = {
	if (c < 128u8) { return 1; };
	if (c < 194u8) { let e: invalid; return e; };
	if (c >= 248u8) { let e: invalid; return e; };
	if (c < 224u8) { return 2; };
	if (c < 240u8) { return 3; };
	return 4;
};

// ref/hare/encoding/utf8/encode.ha:7. Encode `r` into `out` (caller-
// supplied; must hold at least [[runesz]](r) bytes) and return the
// byte count. ABORT if `r` is a UTF-16 surrogate or above U+10FFFF —
// same precondition Hare asserts at ref/hare/encoding/utf8/encode.ha:9.
//
// Surface deviation: Hare returns `[]u8` (slice into a static buf).
// ww uses the caller-buffer form; caller can reuse a [4]u8 stack
// scratch across encodes.
export fn encoderune(out: []u8, r: rune) i32 = {
	let ch: u32 = r: u32;
	if (ch >= 0xD800u32) {
		if (ch <= 0xDFFFu32) {
			abort("utf8.encoderune: surrogate codepoint");
		};
	};
	if (ch > 0x10FFFFu32) {
		abort("utf8.encoderune: codepoint > U+10FFFF");
	};

	let n: i32 = 0;
	let first: u8 = 0u8;
	if (ch < 0x80u32) {
		first = 0u8; n = 1;
	} else if (ch < 0x800u32) {
		first = 0xC0u8; n = 2;
	} else if (ch < 0x10000u32) {
		first = 0xE0u8; n = 3;
	} else {
		first = 0xF0u8; n = 4;
	};

	let v: u32 = ch;
	let i: i32 = n - 1;
	for (i > 0) {
		out[i] = ((v: u8) & 0x3Fu8) | 0x80u8;
		v = v >> 6u32;
		i -= 1;
	};
	out[0] = (v: u8) | first;
	return n;
};

// ref/hare/encoding/utf8/decode.ha:52. Walks back from `d.offs` to a
// byte that could start a codepoint (state-0 dfa cell != -1), re-decodes
// forward from there, and confirms the forward decode lands back at the
// original offset. Returns `done` at start-of-input; `invalid` if no
// initial byte appears within 4 steps (no legal UTF-8 codepoint exceeds
// 4 bytes), if the forward decode returns `more`/`invalid`, or if it
// lands at a different offset than expected. Returns `more` when the
// walk reaches byte 0 without finding any initial byte.
//
// Hare's `for (d.offs < len(d.src); d.offs -= 1)` relies on size_t
// wrap-around to exit when offs underflows past 0; ww's offs is i32,
// so we spell the same exit as `d.offs >= 0`. Hare's `defer d.offs = t`
// is inlined in each match arm — ww has no defer.
export fn prev(d: *decoder) (rune | done | more | invalid) = {
	if (d.offs == 0) {
		let dn: done; return dn;
	};
	let n: i32 = d.offs;
	d.offs -= 1;
	for (d.offs >= 0) {
		let b: u8 = d.src[d.offs];
		let bi: i32 = b: i32;
		let cell: i8 = dfa[bi];
		if (cell: i32 != -1) {
			let t: i32 = d.offs;
			match (next(d)) {
			case let r: rune => {
				let landed: i32 = d.offs;
				d.offs = t;
				if (landed != n) {
					let e: invalid; return e;
				};
				return r;
			};
			case let dn: done => {
				d.offs = t;
				let e: invalid; return e;
			};
			case let m: more => {
				d.offs = t;
				let e: invalid; return e;
			};
			case let e: invalid => {
				d.offs = t;
				let e2: invalid; return e2;
			};
			};
		};
		if (n - d.offs == 4) {
			let e: invalid; return e;
		};
		d.offs -= 1;
	};
	let mr: more; return mr;
};

// ref/hare/encoding/utf8/decode.ha:74. Borrowed view of the bytes from
// the decoder's current position to the end of its source.
export fn remaining(d: *decoder) []u8 = {
	let r: []u8;
	r.ptr = d.src.ptr + (d.offs: u64);
	r.len = d.src.len - d.offs;
	r.cap = d.src.len - d.offs;
	return r;
};

// ref/hare/encoding/utf8/decode.ha:80. Borrowed view of the bytes
// between two decoders' positions. Precondition (Hare asserts both):
// the decoders share the same source, and `begin.offs <= end.offs`.
export fn slice(begin: *decoder, end: *decoder) []u8 = {
	if (begin.src.ptr != end.src.ptr) {
		abort("utf8.slice: decoders from different sources");
	};
	if (begin.offs > end.offs) {
		abort("utf8.slice: begin past end");
	};
	let r: []u8;
	r.ptr = begin.src.ptr + (begin.offs: u64);
	r.len = end.offs - begin.offs;
	r.cap = end.offs - begin.offs;
	return r;
};

// ref/hare/encoding/utf8/decode.ha:203. Byte position of the decoder
// in its source.
export fn position(d: *decoder) i32 = {
	return d.offs;
};


// strings — operations over str ({ptr,len}). Hare port; see
// ref/hare/strings/.
//
// Documented divergences from Hare:
//
//   - `byteindex` / `rbyteindex` rune arms encode via
//     `utf8.encoderune`; the legacy impls scanned for `r: u8` (an
//     undocumented ASCII-only restriction that silently dropped
//     to the wrong byte for U+80..U+7FF and higher).
//   - `dup(s: str) str` — Hare returns `(str | nomem)`. ww's
//     `os.alloc` aborts on OOM (no `nomem` type), so we return plain
//     `str`. Empty input returns `{nil, 0}`; Hare returns the static
//     empty string — same observable result.
//   - `iterator` is flattened (`offs`, `src`, `reverse` fields).
//     Hare uses anonymous-embedded `utf8::decoder`
//     (ref/hare/strings/iter.ha:6-9); ww has no anonymous-embed
//     syntax, so `next`/`prev`/`slice` copy `offs`/`src` into a
//     local `utf8.decoder` for the call (and `next`/`prev` write
//     `offs` back).
//   - Hare's private `move()` helper dispatches on a `forward: bool`
//     using a function-pointer `let fun = if (forward) &utf8::next
//     else &utf8::prev`. ww has no fn-pointers in scope yet, so the
//     dispatch is a branch on `forward` selecting the call site.

package strings;

import bytes;
import encoding.utf8;
import os;
import rt;
import types;

// toutf8 — borrowed []u8 view of `s`. ref/hare/strings/utf8.ha:29.
// `cap` equals `len`; the slice does not own a separate allocation.
export fn toutf8(s: str) []u8 = {
	let r: []u8;
	r.ptr = s.ptr;
	r.len = s.len;
	r.cap = s.len;
	return r;
};

// frombytes — borrowed str view of `in`. Pure reinterpret per
// CLAUDE.md rule 9 carve-out; ref/hare/strings/utf8.ha:10.
export fn frombytes(in: []u8) str = {
	let r: str;
	r.ptr = in.ptr;
	r.len = in.len;
	return r;
};

// compare — three-way bytewise codepoint-order comparison. Return is
// a sign (neg/zero/pos), not an index, so it tracks Hare's `int`
// rather than the str-index i32 (#8). ref/hare/strings/compare.ha:12.
export fn compare(a: str, b: str) int = {
	let n: i32 = a.len;
	if (b.len < n) { n = b.len; };
	let i: i32 = 0;
	for (i < n) {
		if (a[i] != b[i]) { return (a[i]: int) - (b[i]: int); };
		i += 1;
	};
	return (a.len: int) - (b.len: int);
};

// dup — allocate a fresh copy of `s`. Caller releases with
// `os.free(r.ptr, r.len: u64)`. ref/hare/strings/dup.ha:7.
export fn dup(s: str) str = {
	let r: str;
	r.ptr = nil;
	r.len = 0;
	if (s.len == 0) { return r; };
	let buf: []u8 = alloc([], s.len: u64)!;
	let i: i32 = 0;
	for (i < s.len) { buf[i] = s[i]; i += 1; };
	buf.len = s.len;
	return frombytes(buf);
};

// dupall — fresh `[]str` whose elements are independent copies of
// `s`'s elements. Caller releases via [[freeall]].
// ref/hare/strings/dup.ha:26 (#6).
//
// Hare gates the per-element dup behind `?` and rolls back via
// `defer if (!ok) freeall(newsl)`. ww has no `defer if`; more
// importantly, ww's [[dup]] is still unchecked (returns plain `str`,
// aborts via os.alloc on OOM — see top-of-file divergence note),
// so the only nomem propagation point is the initial slice alloc.
// With no inner failure path, the rollback is structurally a no-op
// and is omitted; it returns once dup graduates to `(str | nomem)`
// (#46). The pre-allocated slice has `cap == s.len`, so appendstr's
// rt_ensure call never reaches the grow branch.
//
// Empty input bypasses the alloc: rt_malloc(0) is an mmap of 0 bytes
// which returns -EINVAL, and the alloc-slice `?` shortcut routes
// that through nomem — Hare's heap allocator hands back a sentinel
// instead (#47). Return `{nil, 0, 0}` directly so callers get the
// Hare-observable shape (len==0, freeall is a no-op via cap==0).
export fn dupall(s: []str) ([]str | nomem) = {
	if (s.len == 0) {
		let r: []str;
		r.ptr = nil: *str;
		r.len = 0;
		r.cap = 0;
		return r;
	};
	let newsl: []str = alloc([], s.len)?;
	let i: i32 = 0;
	for (i < s.len) {
		appendstr(&newsl, dup(s[i]));
		i += 1;
	};
	return newsl;
};

// freeall — release each element + the slice header. The natural
// disposer for any `[]str` of dup'd elements (e.g. shlex.split).
// ref/hare/strings/dup.ha:38.
//
// Empty elements (`{nil, 0}` from a zero-length dup) are skipped:
// os.free on a nil pointer at len 0 tickles the rt_free guard. The
// slice header itself is freed at `cap * size(str)` — the literal
// would drift under #1's str-layout bump, so route through the
// typ.ww SSoT. A never-grown slice (cap == 0) skips the header free.
export fn freeall(s: []str) void = {
	let i: i32 = 0;
	for (i < s.len) {
		if (s[i].len > 0) {
			os.free(s[i].ptr: *void, s[i].len: u64);
		};
		i += 1;
	};
	if (s.cap > 0) {
		os.free(s.ptr: *void, (s.cap: u64) * size(str): u64);
	};
};

// concat — fresh allocation containing each element of `strs` in
// order. Caller releases with `os.free(r.ptr, r.len: u64)`.
// ref/hare/strings/concat.ha:5. Hare's `nomem` return is dropped:
// `os.alloc` aborts on OOM.
export fn concat(strs: str...) str = {
	let total: i32 = 0;
	let i: i32 = 0;
	for (i < strs.len) { total += strs[i].len; i += 1; };
	let r: str;
	r.ptr = nil;
	r.len = 0;
	if (total == 0) { return r; };
	let buf: []u8 = alloc([], total: u64)!;
	let off: i32 = 0;
	i = 0;
	for (i < strs.len) {
		let j: i32 = 0;
		for (j < strs[i].len) {
			buf[off + j] = strs[i][j];
			j += 1;
		};
		off += strs[i].len;
		i += 1;
	};
	buf.len = total;
	return frombytes(buf);
};

// join — fresh allocation with `delim` placed between each element of
// `strs`. Caller releases with `os.free(r.ptr, r.len: u64)`.
// ref/hare/strings/concat.ha:46. Hare's `nomem` return is dropped:
// `os.alloc` aborts on OOM.
export fn join(delim: str, strs: str...) str = {
	let total: i32 = 0;
	let i: i32 = 0;
	for (i < strs.len) {
		total += strs[i].len;
		if (i + 1 < strs.len) { total += delim.len; };
		i += 1;
	};
	let r: str;
	r.ptr = nil;
	r.len = 0;
	if (total == 0) { return r; };
	let buf: []u8 = alloc([], total: u64)!;
	let off: i32 = 0;
	i = 0;
	for (i < strs.len) {
		let j: i32 = 0;
		for (j < strs[i].len) {
			buf[off + j] = strs[i][j];
			j += 1;
		};
		off += strs[i].len;
		if (i + 1 < strs.len) {
			j = 0;
			for (j < delim.len) {
				buf[off + j] = delim[j];
				j += 1;
			};
			off += delim.len;
		};
		i += 1;
	};
	buf.len = total;
	return frombytes(buf);
};

// utf8bytelenbounded — walk `it` forward `end` runes and return the
// resulting byte offset. ref/hare/strings/sub.ha:10. Aborts on
// short input per Hare's contract for the rune-wise [[sub]].
fn utf8bytelenbounded(it: *iterator, end: i32) i32 = {
	let i: i32 = 0;
	for (i < end) {
		match (next(it)) {
		case let r: rune => void;
		case utf8.done   => abort("strings.sub: index exceeds string length");
		};
		i += 1;
	};
	return it.offs;
};

// sub — borrowed substring [start, end) where start/end are rune
// indices. ref/hare/strings/sub.ha:30. Hare's 2-arg `sub(s, start)`
// defaulting end=END is omitted: ww has no default-parameter syntax
// (filed as #37). Byte-indexed counterpart: [[bytesub]].
export fn sub(s: str, start: i32, end: i32) str = {
	os.assert(start <= end, "strings.sub: start is higher than end");
	let it: iterator = iter(s);
	let starti: i32 = utf8bytelenbounded(&it, start);
	let endi: i32 = utf8bytelenbounded(&it, end - start);
	let r: str;
	r.ptr = s.ptr + (starti: u64);
	r.len = endi - starti;
	return r;
};

// bytesub — borrowed substring [start, end) where start/end are byte
// offsets. ref/hare/strings/sub.ha:59 (#7). Returns `utf8.invalid` if
// either endpoint lands on a continuation byte (would split a
// codepoint); the equivalent Hare predicate is `s[i] & 0xc0 == 0x80`
// at ref/hare/strings/sub.ha:72-73.
export fn bytesub(s: str, start: i32, end: i32) (str | utf8.invalid) = {
	os.assert(start <= end, "strings.bytesub: start is higher than end");
	os.assert(end <= s.len, "strings.bytesub: end exceeds string length");
	if (start < s.len) {
		if ((s[start] & 0xC0u8) == 0x80u8) {
			let e: utf8.invalid; return e;
		};
	};
	if (end < s.len) {
		if ((s[end] & 0xC0u8) == 0x80u8) {
			let e: utf8.invalid; return e;
		};
	};
	let r: str;
	r.ptr = s.ptr + (start: u64);
	r.len = end - start;
	return r;
};

// runebytes — encode `r` into caller's `scratch` (must hold 4 bytes)
// and return the borrowed slice trimmed to the encoded length. Hare
// inlines the same shape at ref/hare/strings/index.ha:132.
fn runebytes(scratch: []u8, r: rune) []u8 = {
	let n: i32 = utf8.encoderune(scratch, r);
	let s: []u8;
	s.ptr = scratch.ptr;
	s.len = n;
	s.cap = n;
	return s;
};

// hasprefix — true iff `in` begins with `prefix`.
// ref/hare/strings/suffix.ha:8.
export fn hasprefix(in: str, prefix: (str | rune)) bool = {
	let scratch: [4]u8;
	let p: []u8 = match (prefix) {
	case let s: str  => yield toutf8(s);
	case let r: rune => yield runebytes(scratch[0:4], r);
	};
	return bytes.hasprefix(toutf8(in), p);
};

// hassuffix — true iff `in` ends with `suff`.
// ref/hare/strings/suffix.ha:26.
export fn hassuffix(in: str, suff: (str | rune)) bool = {
	let scratch: [4]u8;
	let s: []u8 = match (suff) {
	case let v: str  => yield toutf8(v);
	case let r: rune => yield runebytes(scratch[0:4], r);
	};
	return bytes.hassuffix(toutf8(in), s);
};

// byteindex — byte-wise offset of `needle` in `haystack`, or void if
// absent. ref/hare/strings/index.ha:127. Rune arm encodes via
// utf8.encoderune (Hare passes the encoded slice straight to
// bytes::index).
export fn byteindex(haystack: str, needle: (str | rune)) (i32 | void) = {
	let scratch: [4]u8;
	let n: []u8 = match (needle) {
	case let s: str  => yield toutf8(s);
	case let r: rune => yield runebytes(scratch[0:4], r);
	};
	return bytes.index(toutf8(haystack), n);
};

// rbyteindex — byte-wise offset of the last `needle` in `haystack`.
// ref/hare/strings/index.ha:138.
export fn rbyteindex(haystack: str, needle: (str | rune)) (i32 | void) = {
	let scratch: [4]u8;
	let n: []u8 = match (needle) {
	case let s: str  => yield toutf8(s);
	case let r: rune => yield runebytes(scratch[0:4], r);
	};
	return bytes.rindex(toutf8(haystack), n);
};

// indexstring — str-arm of [[index]]. Dual-rune-iterator walk: at each
// candidate rune index `i`, compare `haystack` from that position
// against `needle` rune-by-rune until needle is exhausted (match) or
// a mismatch / haystack-exhaustion breaks the inner loop. Mirrors
// ref/hare/strings/index.ha:59 (#10). Hare copies `rest_iter = s_iter`
// directly via struct assignment; ww re-seats `rest_iter` field-wise
// because the let-init struct-copy form diverges between cstage and
// wwstage on this iterator type (993_ww_ww + 995_self_rebuild fail,
// filed as #41) and rule #10 (CLAUDE.md) forbids stage asymmetry.
fn indexstring(haystack: str, needle: str) (i32 | void) = {
	let s_iter: iterator = iter(haystack);
	let i: i32 = 0;
	for (true) {
		let rest_iter: iterator;
		rest_iter.src = s_iter.src;
		rest_iter.offs = s_iter.offs;
		rest_iter.reverse = s_iter.reverse;
		let needle_iter: iterator = iter(needle);
		let matched: bool = false;
		for (true) {
			let rest_done:   bool = false;
			let rest_r:      rune;
			match (next(&rest_iter)) {
			case let r: rune => rest_r = r;
			case utf8.done   => rest_done = true;
			};
			let needle_done: bool = false;
			let needle_r:    rune;
			match (next(&needle_iter)) {
			case let r: rune => needle_r = r;
			case utf8.done   => needle_done = true;
			};
			if (rest_done && !needle_done) { break; };
			if (needle_done) { matched = true; break; };
			if (rest_r != needle_r) { break; };
		};
		if (matched) { return i; };
		match (next(&s_iter)) {
		case let r: rune => i += 1;
		case utf8.done   => return;
		};
	};
	return;
};

// index — rune-wise offset of `needle`'s first occurrence in
// `haystack`, or void if absent. ref/hare/strings/index.ha:10. The
// str-arm delegates to [[indexstring]] (dual-iterator rune-by-rune
// walk per Hare's `index_string`, #10); the rune-arm mirrors Hare's
// `index_rune` (ref/hare/strings/index.ha:31).
export fn index(haystack: str, needle: (str | rune)) (i32 | void) = {
	match (needle) {
	case let s: str  => return indexstring(haystack, s);
	case let r: rune => {
		let it: iterator = iter(haystack);
		let i: i32 = 0;
		for (true) {
			match (next(&it)) {
			case let n: rune => {
				if (n == r) { return i; };
				i += 1;
			};
			case utf8.done   => return;
			};
		};
	};
	};
	return;
};

// rindex — rune-wise offset of `needle`'s last occurrence in
// `haystack`, or void if absent. ref/hare/strings/index.ha:22. The
// str-arm reuses `rbyteindex`; the rune-arm walks forward tracking
// the most recent matching rune index (Hare's `rindex_rune` with
// `riter` returns a byte-offset value for multibyte strings, which
// disagrees with the rune-wise docstring; we keep the docstring's
// contract).
export fn rindex(haystack: str, needle: (str | rune)) (i32 | void) = {
	match (needle) {
	case let s: str  => {
		match (rbyteindex(haystack, s)) {
		case void        => return;
		case let bo: i32 => {
			let it: iterator = iter(haystack);
			let i: i32 = 0;
			for (position(&it) < bo) {
				match (next(&it)) {
				case let r: rune => i += 1;
				case utf8.done   => break;
				};
			};
			return i;
		};
		};
	};
	case let r: rune => {
		let it: iterator = iter(haystack);
		let i: i32 = 0;
		let last: i32 = -1;
		for (true) {
			match (next(&it)) {
			case let n: rune => {
				if (n == r) { last = i; };
				i += 1;
			};
			case utf8.done   => break;
			};
		};
		if (last < 0) { return; };
		return last;
	};
	};
	return;
};

// contains — true iff any of `needles` occurs in `haystack`.
// ref/hare/strings/contains.ha:9.
export fn contains(haystack: str, needles: (str | rune)...) bool = {
	let i: i32 = 0;
	for (i < needles.len) {
		match (needles[i]) {
		case let s: str  => {
			match (byteindex(haystack, s)) {
			case let bo: i32 => return true;
			case void        => void;
			};
		};
		case let r: rune => {
			match (byteindex(haystack, r)) {
			case let bo: i32 => return true;
			case void        => void;
			};
		};
		};
		i += 1;
	};
	return false;
};

// trimprefix — `s` with `prefix` stripped from the front, or `s`
// unchanged if it doesn't start with `prefix`. Borrowed view.
// ref/hare/strings/trim.ha:60.
export fn trimprefix(input: str, prefix: str) str = {
	if (!hasprefix(input, prefix)) { return input; };
	let r: str;
	r.ptr = input.ptr + (prefix.len: u64);
	r.len = input.len - prefix.len;
	return r;
};

// trimsuffix — symmetric. ref/hare/strings/trim.ha:69.
export fn trimsuffix(input: str, suffix: str) str = {
	if (!hassuffix(input, suffix)) { return input; };
	let r: str;
	r.ptr = input.ptr;
	r.len = input.len - suffix.len;
	return r;
};

// whitespace — ASCII whitespace set used by the 0-arg ltrim/rtrim/trim
// branches (#9). ref/hare/strings/trim.ha:6.
let whitespace: [4]u8 = [0x20u8, 0x0Au8, 0x09u8, 0x0Du8];

// ltrim — strip leading runes that occur in `trim`. Borrowed view.
// 0-arg strips ASCII whitespace via [[bytes.ltrim]] (#9).
// ref/hare/strings/trim.ha:11. The spread expression is inlined
// because `let ws: []u8 = whitespace[0:4]` produces a slice whose
// ptr doesn't track the module-level array storage (filed as #40);
// `b.flush = flushdefault[0:1]` in lib/bufio is the same shape via
// the working field-assign path.
export fn ltrim(input: str, trim: rune...) str = {
	if (trim.len == 0) {
		return frombytes(bytes.ltrim(toutf8(input), whitespace[0:4]...));
	};
	let it: iterator = iter(input);
	for (true) {
		match (next(&it)) {
		case let r: rune => {
			let j: i32 = 0;
			let found: bool = false;
			for (j < trim.len) {
				if (r == trim[j]) { found = true; j = trim.len; }
				else { j += 1; };
			};
			if (!found) {
				match (prev(&it)) {
				case let r2: rune => void;
				case utf8.done    => void;
				};
				break;
			};
		};
		case utf8.done => break;
		};
	};
	return iterstr(&it);
};

// rtrim — strip trailing runes that occur in `trim`. Borrowed view.
// 0-arg strips ASCII whitespace via [[bytes.rtrim]] (#9). Spread is
// inlined to dodge #40 — see [[ltrim]].
// ref/hare/strings/trim.ha:32.
export fn rtrim(input: str, trim: rune...) str = {
	if (trim.len == 0) {
		return frombytes(bytes.rtrim(toutf8(input), whitespace[0:4]...));
	};
	let it: iterator = riter(input);
	for (true) {
		match (next(&it)) {
		case let r: rune => {
			let j: i32 = 0;
			let found: bool = false;
			for (j < trim.len) {
				if (r == trim[j]) { found = true; j = trim.len; }
				else { j += 1; };
			};
			if (!found) {
				match (prev(&it)) {
				case let r2: rune => void;
				case utf8.done    => void;
				};
				break;
			};
		};
		case utf8.done => break;
		};
	};
	return iterstr(&it);
};

// trim — strip from both ends. ref/hare/strings/trim.ha:54.
export fn trim(input: str, trim: rune...) str = {
	return ltrim(rtrim(input, trim...), trim...);
};

// iterator — UTF-8 rune cursor over a `str`. Layout flattens Hare's
// anonymous-embedded `utf8::decoder` (ref/hare/strings/iter.ha:6-9) to
// explicit fields. `reverse` selects walk direction: forward iterators
// (`iter`) advance through utf8.next; reverse iterators (`riter`) advance
// through utf8.prev. May be copied to save state.
export type iterator = struct {
	offs:    i32,
	src:     []u8,
	reverse: bool,
};

// iter — initialize a forward iterator at the start of `src`.
// ref/hare/strings/iter.ha:24.
export fn iter(src: str) iterator = {
	let r: iterator;
	r.src = toutf8(src);
	r.offs = 0;
	r.reverse = false;
	return r;
};

// riter — initialize a reverse iterator at the end of `src`. `next`
// on a reverse iterator walks back through the string.
// ref/hare/strings/iter.ha:32.
export fn riter(src: str) iterator = {
	let r: iterator;
	r.src = toutf8(src);
	r.offs = src.len;
	r.reverse = true;
	return r;
};

// move — private dispatch shared by next/prev. `forward` selects
// utf8.next vs utf8.prev. Aborts on more/invalid per Hare's
// ref/hare/strings/iter.ha:51-58 ("Invalid UTF-8 string (this should
// not happen)"). Hare picks the utf8 function via a fn-pointer; ww
// branches on `forward` at each call site instead.
fn move(forward: bool, it: *iterator) (rune | utf8.done) = {
	let d: utf8.decoder;
	d.src = it.src;
	d.offs = it.offs;
	if (forward) {
		match (utf8.next(&d)) {
		case let r: rune          => { it.offs = d.offs; return r; };
		case let dn: utf8.done    => return dn;
		case let m:  utf8.more    => abort("strings.move: invalid UTF-8");
		case let e:  utf8.invalid => abort("strings.move: invalid UTF-8");
		};
	} else {
		match (utf8.prev(&d)) {
		case let r: rune          => { it.offs = d.offs; return r; };
		case let dn: utf8.done    => return dn;
		case let m:  utf8.more    => abort("strings.move: invalid UTF-8");
		case let e:  utf8.invalid => abort("strings.move: invalid UTF-8");
		};
	};
};

// next — advance the iterator one rune. Forward iterators step
// through utf8.next; reverse iterators (riter) step backward through
// utf8.prev. Returns utf8.done at end-of-walk. ref/hare/strings/iter.ha:45.
export fn next(it: *iterator) (rune | utf8.done) = {
	return move(!it.reverse, it);
};

// prev — step back one rune. Dual to next: on a forward iterator
// this walks utf8.prev; on a reverse iterator (riter) it walks
// utf8.next. ref/hare/strings/iter.ha:49.
export fn prev(it: *iterator) (rune | utf8.done) = {
	return move(it.reverse, it);
};

// iterstr — borrowed view of the bytes remaining in the iterator's
// walk direction. Forward iter: bytes from offs to end; reverse iter:
// bytes from start to offs. ref/hare/strings/iter.ha:63.
export fn iterstr(it: *iterator) str = {
	let r: []u8;
	if (it.reverse) {
		r = it.src[0:it.offs];
	} else {
		r = it.src[it.offs:it.src.len];
	};
	return frombytes(r);
};

// slice — borrowed substring between two iterator positions.
// ref/hare/strings/iter.ha:75. Hare passes `*iterator` directly where
// `*utf8::decoder` is expected via anonymous-embed coercion; ww has
// no anonymous embed, so we reconstruct a local utf8.decoder for each
// endpoint and forward — same pattern as `move` above.
export fn slice(begin: *iterator, end: *iterator) str = {
	let b: utf8.decoder;
	b.src = begin.src;
	b.offs = begin.offs;
	let e: utf8.decoder;
	e.src = end.src;
	e.offs = end.offs;
	return frombytes(utf8.slice(&b, &e));
};

// position — byte-wise offset of the iterator in its source.
// ref/hare/strings/iter.ha:82.
export fn position(it: *iterator) i32 = {
	return it.offs;
};

// tokenizer — re-export of bytes.tokenizer. ref/hare/strings/tokenize.ha:7.
// First cross-module type alias in tree; needs #22's transitive
// alias-chain unwrap (cstage type_chase_named + wwstage
// structlookupchain) to walk struct fields through the chain.
export type tokenizer = bytes.tokenizer;

// tokenize — yield substrings of `s` split on any byte in `delim`.
// Leading / trailing / adjacent delims yield empty tokens. `s` and
// `delim` are borrowed; caller keeps them live for the tokenizer's
// lifetime. ref/hare/strings/tokenize.ha:32. ASCII-only delim
// asserted per Hare lines 35-37: a multibyte rune in delim would
// split on a single continuation byte and yield invalid UTF-8.
export fn tokenize(s: str, delim: str) tokenizer = {
	let d: []u8 = toutf8(delim);
	let i: i32 = 0;
	for (i < d.len) {
		os.assert((d[i] & 0x80u8) == 0u8,
			"strings.tokenize cannot tokenize on non-ASCII delimiters");
		i += 1;
	};
	return bytes.tokenize(toutf8(s), d...);
};

// rtokenize — reverse-direction counterpart to [[tokenize]]. First
// next_token yields the last token, last yields the first.
// ref/hare/strings/tokenize.ha:44.
export fn rtokenize(s: str, delim: str) tokenizer = {
	let d: []u8 = toutf8(delim);
	let i: i32 = 0;
	for (i < d.len) {
		os.assert((d[i] & 0x80u8) == 0u8,
			"strings.rtokenize cannot tokenize on non-ASCII delimiters");
		i += 1;
	};
	return bytes.rtokenize(toutf8(s), d...);
};

// next_token — current token, advancing the cursor.
// ref/hare/strings/tokenize.ha:62.
export fn next_token(s: *tokenizer) (str | bytes.done) = {
	let b: *bytes.tokenizer = s: *bytes.tokenizer;
	match (bytes.next_token(b)) {
	case let v: []u8 => return frombytes(v);
	case bytes.done  => { let d: bytes.done; return d; };
	};
};

// peek_token — current token without advancing.
// ref/hare/strings/tokenize.ha:71.
export fn peek_token(s: *tokenizer) (str | bytes.done) = {
	let b: *bytes.tokenizer = s: *bytes.tokenizer;
	match (bytes.peek_token(b)) {
	case let v: []u8 => return frombytes(v);
	case bytes.done  => { let d: bytes.done; return d; };
	};
};

// remaining_tokens — unconsumed portion of the input ahead of the
// cursor. ref/hare/strings/tokenize.ha:79.
export fn remaining_tokens(s: *tokenizer) str = {
	let b: *bytes.tokenizer = s: *bytes.tokenizer;
	return frombytes(bytes.remaining_tokens(b));
};

// cut — split `in` along the first instance of `delim`, returning the
// portions before and after it. When `delim` is absent the whole input
// is the first half and the second is empty. Both halves are borrowed
// from `in`; caller ensures `delim` is non-empty.
// ref/hare/strings/tokenize.ha:288.
export fn cut(in: str, delim: str) (str, str) = {
	let (a, b) = bytes.cut(toutf8(in), toutf8(delim));
	return (frombytes(a), frombytes(b));
};

// rcut — like [[cut]] but split along the LAST instance of `delim`.
// ref/hare/strings/tokenize.ha:302.
export fn rcut(in: str, delim: str) (str, str) = {
	let (a, b) = bytes.rcut(toutf8(in), toutf8(delim));
	return (frombytes(a), frombytes(b));
};

// rt_ensure is the runtime slice-growth helper invoked by the
// `append(s, v)` builtin. Direct bind for the same reason as
// lib/shlex.shlex (appendstr, 16B): the builtin's expansion stores
// only 8B of the new element, losing the `.len` half of a `str`.
@symbol("rt_ensure") fn rtensure(s: *void, membsz: u64) void;

// appendstr — grow `*slice` by one and store `item` (16B). Mirror of
// lib/shlex.shlex appendstr. Collapses when the append builtin learns
// to store the full element width.
fn appendstr(slice: *[]str, item: str) void = {
	let newlen: i32 = slice.len + 1;
	slice.len = newlen;
	rtensure(slice: *void, size(str): u64);
	let dst: *str = &slice.ptr[newlen - 1];
	dst.ptr = item.ptr;
	dst.len = item.len;
};

// splitn — split `in` on any byte in `delim`, returning up to `n`
// tokens via forward iteration. The trailing slot (when more than
// `n - 1` tokens exist) holds the unconsumed remainder. Strings
// within the result are borrowed from `in`.
//
// The caller frees the returned slice via
// `os.free(r.ptr: *void, (r.cap: u64) * size(str): u64)`.
//
// Hare's `([]str | nomem)` collapses to `[]str` here: ww os.alloc
// has no recoverable failure path. Same precedent as
// shlex.split / bytes.splitn.
//
// ref/hare/strings/tokenize.ha:172.
export fn splitn(in: str, delim: str, n: i32) []str = {
	let toks: []str;
	toks.ptr = nil: *str;
	toks.len = 0;
	toks.cap = 0;
	let tok: tokenizer = tokenize(in, delim);
	let i: i32 = 0;
	for (i < n - 1) {
		match (next_token(&tok)) {
		case let s: str => { appendstr(&toks, s); };
		case bytes.done => { return toks; };
		};
		i += 1;
	};
	match (peek_token(&tok)) {
	case bytes.done  => void;
	case let pk: str => {
		let r: str = remaining_tokens(&tok);
		appendstr(&toks, r);
	};
	};
	return toks;
};

// rsplitn — reverse-direction counterpart to [[splitn]]: tokens are
// collected from the end of `in`. The trailing slot holds the
// unconsumed prefix (everything before the n-th-from-last delim hit).
//
// When the input has fewer than n tokens, the `done` short-circuit
// returns toks UN-reversed (in last-token-first order). Mirrors Hare
// at ref/hare/strings/tokenize.ha:219-224 where the in-place reverse
// step is gated behind the n-1 loop running to completion.
//
// ref/hare/strings/tokenize.ha:200.
export fn rsplitn(in: str, delim: str, n: i32) []str = {
	let toks: []str;
	toks.ptr = nil: *str;
	toks.len = 0;
	toks.cap = 0;
	let tok: tokenizer = rtokenize(in, delim);
	let i: i32 = 0;
	for (i < n - 1) {
		match (next_token(&tok)) {
		case let s: str => { appendstr(&toks, s); };
		case bytes.done => { return toks; };
		};
		i += 1;
	};
	match (peek_token(&tok)) {
	case bytes.done  => void;
	case let pk: str => {
		let r: str = remaining_tokens(&tok);
		appendstr(&toks, r);
	};
	};

	// In-place reverse so callers see argv-order, matching Hare
	// (ref/hare/strings/tokenize.ha:220). Element copy is field-wise
	// through `*str` because `toks[i] = toks[j]` (full 16B str store)
	// lands in the multi-word-store gap noted at cmd/w6c/cgen.c:6515.
	let a: i32 = 0;
	let b: i32 = toks.len - 1;
	for (a < b) {
		let pa: *str = &toks.ptr[a];
		let pb: *str = &toks.ptr[b];
		let tp: *u8 = pa.ptr;
		let tl: i32 = pa.len;
		pa.ptr = pb.ptr;
		pa.len = pb.len;
		pb.ptr = tp;
		pb.len = tl;
		a += 1;
		b -= 1;
	};
	return toks;
};

// split — full split of `in` on `delim` (no token cap). Mirrors
// `splitn(in, delim, types::SIZE_MAX)`. ww uses `types.I32_MAX`
// because the index type is i32 (lib/CLAUDE.md).
//
// ref/hare/strings/tokenize.ha:242.
export fn split(in: str, delim: str) []str = {
	return splitn(in, delim, types.I32_MAX);
};

// lpad — left-pad `s` with `p` rune until the result reaches `maxlen`
// bytes. Length comparison is BYTES, mirroring Hare's `len(s) >= maxlen`
// at ref/hare/strings/pad.ha:9. A multibyte `p` whose encoded width
// doesn't divide `maxlen - s.len` evenly leaves a trailing pad byte
// pair sliced mid-codepoint at byte `maxlen-1`, exactly as Hare's
// `res[..maxlen]` does (ref/hare/strings/pad.ha:20). When
// `(maxlen - s.len) * pad.len >= maxlen` (multibyte pad overflows the
// budget), `s` is entirely sliced off — same as Hare. Caller releases
// with `os.free(r.ptr, r.len: u64)`. Hare's `nomem` return is dropped:
// `os.alloc` aborts on OOM. Buf size == r.len keeps the free-contract
// shape of [[dup]] / [[concat]] / [[join]]; Hare's `alloc([], maxlen)!`
// over-allocs via append then slices, but Hare's slice-free recovers
// the true capacity from the heap allocator (rt/ensure.ha:24), which
// ww's munmap-based `os.free` cannot do.
export fn lpad(s: str, p: rune, maxlen: i32) str = {
	if (s.len >= maxlen) { return dup(s); };
	let scratch: [4]u8;
	let pad: []u8 = runebytes(scratch[0:4], p);
	let buf: []u8 = alloc([], maxlen: u64)!;
	let padwrite: i32 = (maxlen - s.len) * pad.len;
	if (padwrite > maxlen) { padwrite = maxlen; };
	let off: i32 = 0;
	for (off < padwrite) {
		buf[off] = pad.ptr[off % pad.len];
		off += 1;
	};
	let k: i32 = 0;
	let srem: i32 = maxlen - off;
	if (srem > s.len) { srem = s.len; };
	for (k < srem) {
		buf[off + k] = s[k];
		k += 1;
	};
	buf.len = maxlen;
	return frombytes(buf);
};

// replace — fresh allocation of `s` with every non-overlapping
// occurrence of `needle` replaced by `target`. Caller releases with
// `os.free(r.ptr, r.len: u64)`. ref/hare/strings/replace.ha:8 (#4).
//
// Hare delegates to [[multireplace]] with a single pair; ww has no
// `(str, str)` variadic shape today (#39), so this is a standalone
// two-pass implementation: pass 1 counts matches to size the result,
// pass 2 copies chunks and `target` into a single fresh buffer.
// Single nomem path (the `alloc([], total)?`) preserves Hare's
// signature without a per-write `append(...)?` (ww's append builtin
// aborts on OOM, #11). Empty `needle` would hasprefix-match every
// position with a zero stride — same infinite loop Hare exhibits at
// ref/hare/strings/replace.ha:31; not gated.
export fn replace(s: str, needle: str, target: str) (str | nomem) = {
	let sb: []u8 = toutf8(s);
	let nb: []u8 = toutf8(needle);
	let tb: []u8 = toutf8(target);
	let count: i32 = 0;
	let i: i32 = 0;
	for (i < sb.len) {
		if (bytes.hasprefix(sb[i:sb.len], nb)) {
			count += 1;
			i += nb.len;
		} else {
			i += 1;
		};
	};
	let total: i32 = sb.len + count * (tb.len - nb.len);
	if (total == 0) {
		let r: str;
		r.ptr = nil;
		r.len = 0;
		return r;
	};
	let res: []u8 = alloc([], total)?;
	let off: i32 = 0;
	i = 0;
	for (i < sb.len) {
		if (bytes.hasprefix(sb[i:sb.len], nb)) {
			let j: i32 = 0;
			for (j < tb.len) {
				res.ptr[off + j] = tb.ptr[j];
				j += 1;
			};
			off += tb.len;
			i += nb.len;
		} else {
			res.ptr[off] = sb.ptr[i];
			off += 1;
			i += 1;
		};
	};
	res.len = total;
	return frombytes(res);
};

// rpad — right-pad `s` with `p` rune until the result reaches `maxlen`
// bytes. Symmetric with [[lpad]]. ref/hare/strings/pad.ha:39.
export fn rpad(s: str, p: rune, maxlen: i32) str = {
	if (s.len >= maxlen) { return dup(s); };
	let scratch: [4]u8;
	let pad: []u8 = runebytes(scratch[0:4], p);
	let buf: []u8 = alloc([], maxlen: u64)!;
	let k: i32 = 0;
	for (k < s.len) {
		buf[k] = s[k];
		k += 1;
	};
	let padwrite: i32 = maxlen - s.len;
	let i: i32 = 0;
	for (i < padwrite) {
		buf[s.len + i] = pad.ptr[i % pad.len];
		i += 1;
	};
	buf.len = maxlen;
	return frombytes(buf);
};

// selfhost/cmd/w6l/obj.ww — port of cmd/w6l/obj.c.
//
// Loads relocatable ELF64 .o files emitted by w6a, appends .text to
// the combined image, and pulls in symbols + relocations with
// offsets adjusted to the combined section.
//
// Also handles SysV `ar` archives (libwwrt.a). The two-pass loader
// indexes members on the first pass and iteratively pulls members
// that define currently-undefined symbols on subsequent passes.

package w6l;

import os;
import rt;
import strings;
import sym;

def ET_REL:        i32 = 1;
def EM_X86_64:     i32 = 62;
def SHT_PROGBITS:  i32 = 1;
def SHT_SYMTAB:    i32 = 2;
def SHT_STRTAB:    i32 = 3;
def SHT_RELA:      i32 = 4;

// ---- little-endian byte readers ----------------------------------------
// w6a/w6l use straight LE on amd64. Reading via byte offsets keeps us off
// the cgen's u16 field-load story for now (MOVZBQ exists; MOVZWQ doesn't).

fn rdu16(p: *u8, off: u64) u16 = {
	let b0: u16 = p[off]: u16;
	let b1: u16 = p[off + 1u64]: u16;
	return b0 | (b1 << 8u16);
};

fn rdu32(p: *u8, off: u64) u32 = {
	let b0: u32 = p[off]: u32;
	let b1: u32 = p[off + 1u64]: u32;
	let b2: u32 = p[off + 2u64]: u32;
	let b3: u32 = p[off + 3u64]: u32;
	return b0 | (b1 << 8u32) | (b2 << 16u32) | (b3 << 24u32);
};

fn rdu64(p: *u8, off: u64) u64 = {
	let lo: u64 = rdu32(p, off): u64;
	let hi: u64 = rdu32(p, off + 4u64): u64;
	return lo | (hi << 32u64);
};

// ---- ELF64 section header offsets (40 bytes total) --------------------
def SHDR_SIZE:     u64 = 64u64;	// sizeof(Shdr) per ELF64 spec
def SHDR_NAME:     u64 = 0u64;
def SHDR_TYPE:     u64 = 4u64;
def SHDR_OFFSET:   u64 = 24u64;
def SHDR_SIZE_F:   u64 = 32u64;
def SHDR_LINK:     u64 = 40u64;

// ELF64 ehdr field offsets
def EHDR_SIZE:     u64 = 64u64;
def EHDR_TYPE:     u64 = 16u64;
def EHDR_MACHINE:  u64 = 18u64;
def EHDR_SHOFF:    u64 = 40u64;
def EHDR_SHENTSIZE: u64 = 58u64;
def EHDR_SHNUM:    u64 = 60u64;
def EHDR_SHSTRNDX: u64 = 62u64;

// ELF64 sym entry: 24 bytes
def SYM_SIZE:      u64 = 24u64;
def SYM_NAME:      u64 = 0u64;
def SYM_INFO:      u64 = 4u64;
def SYM_SHNDX:     u64 = 6u64;
def SYM_VALUE:     u64 = 8u64;

// ELF64 RELA entry: 24 bytes
def RELA_SIZE:     u64 = 24u64;
def RELA_OFFSET:   u64 = 0u64;
def RELA_INFO:     u64 = 8u64;
def RELA_ADDEND:   u64 = 16u64;

// ---- file slurp --------------------------------------------------------

fn slurp(path: *u8) (*u8, u64) = {
	let fd: i32 = os.open(pathstr(path), os.flag.RDONLY, 0i32);
	if (fd < 0) { return nil, 0u64; };
	let szr: (i64 | os.oserror) = os.filesize(fd);
	let n: i64 = 0i64;
	match (szr) {
	case let v: i64 => n = v;
	case let e: os.oserror => { os.close(fd); return nil, 0u64; };
	};
	let buf: []u8 = alloc([], n: u64)!;
	buf.len = n: i32;
	let rr: (i64 | os.oserror) = os.readall(fd, buf.ptr, n: u64);
	os.close(fd);
	let got: i64 = 0i64;
	match (rr) {
	case let v: i64 => got = v;
	case let e: os.oserror => return nil, 0u64;
	};
	if (got != n) { return nil, 0u64; };
	return buf.ptr, n: u64;
};

// ---- text buffer growth ------------------------------------------------

fn emittext(l: *lnk, src: *u8, n: u64) void = {
	if (l.textlen + n > l.textcap) {
		let nc: u64 = l.textcap;
		if (nc == 0u64) { nc = 4096u64; };
		for (nc < l.textlen + n) { nc = nc * 2u64; };
		// Grow by mmap'ing a fresh region and copying. The old buffer
		// is leaked into the page allocator; for a linker run this is
		// trivial waste.
		let nb: []u8 = alloc([], nc)!;
		let i: u64 = 0u64;
		for (i < l.textlen) {
			nb[i] = l.text[i];
			i += 1u64;
		};
		l.text = nb.ptr;
		l.textcap = nc;
	};
	let i: u64 = 0u64;
	for (i < n) {
		l.text[l.textlen + i] = src[i];
		i += 1u64;
	};
	l.textlen += n;
};

fn emitdata(l: *lnk, src: *u8, n: u64) void = {
	if (l.datalen + n > l.datacap) {
		let nc: u64 = l.datacap;
		if (nc == 0u64) { nc = 256u64; };
		for (nc < l.datalen + n) { nc = nc * 2u64; };
		let nb: []u8 = alloc([], nc)!;
		let i: u64 = 0u64;
		for (i < l.datalen) {
			nb[i] = l.data[i];
			i += 1u64;
		};
		l.data = nb.ptr;
		l.datacap = nc;
	};
	let i: u64 = 0u64;
	for (i < n) {
		l.data[l.datalen + i] = src[i];
		i += 1u64;
	};
	l.datalen += n;
};

// ---- C-string helpers --------------------------------------------------

fn cstrlen(p: *u8) u64 = {
	let n: u64 = 0u64;
	for (p[n] != 0u8) { n += 1u64; };
	return n;
};

// pathstr — view a NUL-terminated *u8 as a str. Bridges argv-style
// callers to lib/os entrypoints (str post-task-#23). Shared with
// main.ww and dyn.ww via the w6l bundle.
fn pathstr(p: *u8) str = {
	let r: str;
	r.ptr = p;
	r.len = cstrlen(p): i32;
	return r;
};

fn cstreq(p: *u8, lit: str) bool = {
	let n: u64 = lit.len: u64;
	let i: u64 = 0u64;
	for (i < n) {
		let li: i32 = i: i32;
		if (p[i] != lit[li]) { return false; };
		i += 1u64;
	};
	if (p[i] != 0u8) { return false; };
	return true;
};

// Build a ww str from a NUL-terminated *u8 (for passing to intern).
fn cstrtostr(p: *u8) str = {
	let n: u64 = cstrlen(p);
	let view: str;
	view.ptr = p;
	view.len = n: i32;
	return strings.dup(view);
};

// ---- archive (SysV ar) types and helpers -------------------------------
//
// Each archive member starts with a 60-byte ar_hdr. The fields we care
// about are the first byte (member type) and the size at offset 48 (a
// 10-byte, space-padded decimal). Member bodies are 2-byte aligned.

type defent = struct {
	name:  str,
	dnext: *defent,
};

type armember = struct {
	data:   *u8,		// owned heap copy of the member's ELF bytes
	size:   u64,
	defs:   *defent,	// linked list of defined globals
	loaded: i32,
	mnext:  *armember,
};

fn isarchive(p: *u8, len: u64) bool = {
	if (len < 8u64) { return false; };
	if (p[0u64] != 33u8)  { return false; };	// '!'
	if (p[1u64] != 60u8)  { return false; };	// '<'
	if (p[2u64] != 97u8)  { return false; };	// 'a'
	if (p[3u64] != 114u8) { return false; };	// 'r'
	if (p[4u64] != 99u8)  { return false; };	// 'c'
	if (p[5u64] != 104u8) { return false; };	// 'h'
	if (p[6u64] != 62u8)  { return false; };	// '>'
	if (p[7u64] != 10u8)  { return false; };	// '\n'
	return true;
};

// arfield — parse a space-padded decimal integer of width n.
fn arfield(p: *u8, n: u64) u64 = {
	let v: u64 = 0u64;
	let i: u64 = 0u64;
	for (i < n) {
		let c: u8 = p[i];
		if (c < 48u8) { return v; };	// space, NUL, etc.
		if (c > 57u8) { return v; };
		v = v * 10u64 + ((c - 48u8): u64);
		i += 1u64;
	};
	return v;
};

// elfglobals — return a linked list of names of globally-defined
// (STB_GLOBAL) symbols whose section is `.text`. Names are owned
// heap copies, so the source ELF buffer can be freed afterward.
fn elfglobals(buf: *u8, len: u64) *defent = {
	if (len < EHDR_SIZE) { return nil; };
	if (buf[0u64] != 127u8) { return nil; };
	if (buf[1u64] != 69u8)  { return nil; };
	if (buf[2u64] != 76u8)  { return nil; };
	if (buf[3u64] != 70u8)  { return nil; };

	let shoff:    u64 = rdu64(buf, EHDR_SHOFF);
	let shnum:    u32 = rdu16(buf, EHDR_SHNUM): u32;
	let shstrndx: u32 = rdu16(buf, EHDR_SHSTRNDX): u32;

	let shstrshoff: u64 = rdu64(buf, shoff + (shstrndx: u64) * SHDR_SIZE + SHDR_OFFSET);
	let shstr: *u8 = buf + shstrshoff;

	let idxtext:   i32 = -1;
	let idxdata:   i32 = -1;
	let idxsymtab: i32 = -1;
	let i: u32 = 0u32;
	for (i < shnum) {
		let secoff: u64 = shoff + (i: u64) * SHDR_SIZE;
		let shtype: u32 = rdu32(buf, secoff + SHDR_TYPE);
		let shname: u32 = rdu32(buf, secoff + SHDR_NAME);
		let nm: *u8 = shstr + (shname: u64);
		if (shtype == SHT_PROGBITS: u32) {
			if (cstreq(nm, ".text")) { idxtext = i: i32; };
			if (cstreq(nm, ".data")) { idxdata = i: i32; };
		};
		if (shtype == SHT_SYMTAB: u32)  { idxsymtab = i: i32; };
		i += 1u32;
	};
	if (idxtext < 0) { return nil; };
	if (idxsymtab < 0) { return nil; };

	let symsh:   u64 = shoff + (idxsymtab: u64) * SHDR_SIZE;
	let symoff:  u64 = rdu64(buf, symsh + SHDR_OFFSET);
	let symsize: u64 = rdu64(buf, symsh + SHDR_SIZE_F);
	let symlink: u32 = rdu32(buf, symsh + SHDR_LINK);
	let nsyms:    u64 = symsize / SYM_SIZE;

	let strsh:  u64 = shoff + (symlink: u64) * SHDR_SIZE;
	let stroff: u64 = rdu64(buf, strsh + SHDR_OFFSET);
	let strtab:  *u8 = buf + stroff;

	let head: *defent = nil;
	let si: u64 = 1u64;
	for (si < nsyms) {
		let symp: u64 = symoff + si * SYM_SIZE;
		let stname:  u32 = rdu32(buf, symp + SYM_NAME);
		let stinfo:  u8  = buf[symp + SYM_INFO];
		let stshndx: u16 = rdu16(buf, symp + SYM_SHNDX);
		let bind:     u32 = (stinfo: u32) >> 4u32;
		// STB_GLOBAL = 1; defined in .text or .data. Both are
		// included so an archive member that owns a data global
		// gets pulled in when something references it.
		if (bind == 1u32) {
			if (stshndx != 0u16) {
				let intext: bool = (stshndx: i32) == idxtext;
				let indt:   bool = false;
				if (idxdata >= 0) {
					indt = (stshndx: i32) == idxdata;
				};
				if (intext || indt) {
					let nmp: *u8 = strtab + (stname: u64);
					if (nmp[0u64] != 0u8) {
						let nm: str = cstrtostr(nmp);
						let de: *defent = alloc(defent { name = nm, dnext = head })!;
						head = de;
					};
				};
			};
		};
		si += 1u64;
	};
	return head;
};

// memberdefinesundef — true if any of m's defined globals matches a
// currently-undefined symbol in the linker's symbol table. Names not
// already interned are uninteresting (the link doesn't need them yet).
fn memberdefinesundef(l: *lnk, m: *armember) bool = {
	let de: *defent = m.defs;
	for (de != nil) {
		let s: *lsym = lookup(l, de.name);
		if (s != nil) {
			if (s.defined == 0) { return true; };
		};
		de = de.dnext;
	};
	return false;
};

// loadarchive — port of cmd/w6l/obj.c:load_archive.
//
// Pass 1 indexes every regular member. Pass 2 iteratively pulls in any
// member that supplies a currently-undefined symbol; each pull may
// introduce fresh undefs, so we loop until quiescent.
fn loadarchive(l: *lnk, path: *u8, buf: *u8, len: u64) i32 = {
	let head: *armember = nil;
	let tail: *armember = nil;
	let pos:  u64 = 8u64;	// past "!<arch>\n"
	for (pos + 60u64 <= len) {
		let hdrsize: u64 = arfield(buf + pos + 48u64, 10u64);
		let hdrend:  u64 = pos + 60u64;
		if (hdrend + hdrsize > len) { break; };
		let first: u8 = buf[pos];
		// Skip the symbol table ('/'), long-name table ('//'), and
		// any padding entries (NUL or space leading byte).
		if (first != 47u8) { if (first != 0u8) { if (first != 32u8) {
			let m: *armember = alloc(armember { size = hdrsize })!;
			let mbs: []u8 = alloc([], hdrsize)!;
			let mb: *u8 = mbs.ptr;
			let i: u64 = 0u64;
			for (i < hdrsize) {
				mb[i] = buf[hdrend + i];
				i += 1u64;
			};
			m.data = mb;
			m.defs = elfglobals(mb, hdrsize);
			if (head == nil) { head = m; }
			else { tail.mnext = m; };
			tail = m;
		}; }; };
		pos = hdrend + hdrsize;
		if ((hdrsize & 1u64) != 0u64) { pos = pos + 1u64; };
	};

	let changed: i32 = 1;
	for (changed != 0) {
		changed = 0;
		let m: *armember = head;
		for (m != nil) {
			if (m.loaded == 0) {
				if (memberdefinesundef(l, m)) {
					if (loadimage(l, path, m.data, m.size) == 0) {
						m.loaded = 1;
						changed = 1;
					};
				};
			};
			m = m.mnext;
		};
	};
	return 0;
};

// ---- main loader -------------------------------------------------------

export fn load(l: *lnk, path: *u8) i32 = {
	let bufp: *u8;
	let buflen: u64;
	bufp, buflen = slurp(path);
	if (bufp == nil) {
		os.write(2, "w6l: cannot read object\n".ptr, 23u64);
		return -1;
	};
	if (isarchive(bufp, buflen)) {
		return loadarchive(l, path, bufp, buflen);
	};
	return loadimage(l, path, bufp, buflen);
};

fn loadimage(l: *lnk, path: *u8, buf: *u8, len: u64) i32 = {
	if (len < EHDR_SIZE) { return -1; };
	// magic: 0x7f, 'E', 'L', 'F'
	if (buf[0u64] != 127u8) { return -1; };
	if (buf[1u64] != 69u8)  { return -1; };
	if (buf[2u64] != 76u8)  { return -1; };
	if (buf[3u64] != 70u8)  { return -1; };
	if (buf[4u64] != 2u8)   { return -1; };	// ELFCLASS64
	if (rdu16(buf, EHDR_TYPE)    != ET_REL: u16)    { return -1; };
	if (rdu16(buf, EHDR_MACHINE) != EM_X86_64: u16) { return -1; };

	let shoff:    u64 = rdu64(buf, EHDR_SHOFF);
	let shnum:    u32 = rdu16(buf, EHDR_SHNUM): u32;
	let shstrndx: u32 = rdu16(buf, EHDR_SHSTRNDX): u32;

	let shstrshoff: u64 = rdu64(buf, shoff + (shstrndx: u64) * SHDR_SIZE + SHDR_OFFSET);
	let shstr: *u8 = buf + shstrshoff;

	// find .text, .data, .symtab, .rela.text, .rela.data
	let idxtext:   i32 = -1;
	let idxdata:   i32 = -1;
	let idxsymtab: i32 = -1;
	let idxrela:   i32 = -1;
	let idxrelad:  i32 = -1;
	let i: u32 = 0u32;
	for (i < shnum) {
		let secoff: u64 = shoff + (i: u64) * SHDR_SIZE;
		let shtype: u32 = rdu32(buf, secoff + SHDR_TYPE);
		let shname: u32 = rdu32(buf, secoff + SHDR_NAME);
		let nm: *u8 = shstr + (shname: u64);
		if (shtype == SHT_PROGBITS: u32) {
			if (cstreq(nm, ".text")) { idxtext = i: i32; };
			if (cstreq(nm, ".data")) { idxdata = i: i32; };
		};
		if (shtype == SHT_SYMTAB: u32)  { idxsymtab = i: i32; };
		if (shtype == SHT_RELA: u32) {
			if (cstreq(nm, ".rela.text")) { idxrela = i: i32; };
			if (cstreq(nm, ".rela.data")) { idxrelad = i: i32; };
		};
		i += 1u32;
	};
	if (idxtext < 0) {
		os.write(2, "w6l: missing .text\n".ptr, 18u64);
		return -1;
	};
	if (idxsymtab < 0) {
		os.write(2, "w6l: missing .symtab\n".ptr, 20u64);
		return -1;
	};

	let textsh:   u64 = shoff + (idxtext: u64) * SHDR_SIZE;
	let textoff:  u64 = rdu64(buf, textsh + SHDR_OFFSET);
	let textsize: u64 = rdu64(buf, textsh + SHDR_SIZE_F);

	let symsh:   u64 = shoff + (idxsymtab: u64) * SHDR_SIZE;
	let symoff:  u64 = rdu64(buf, symsh + SHDR_OFFSET);
	let symsize: u64 = rdu64(buf, symsh + SHDR_SIZE_F);
	let symlink: u32 = rdu32(buf, symsh + SHDR_LINK);
	let nsyms:    u64 = symsize / SYM_SIZE;

	let strsh:  u64 = shoff + (symlink: u64) * SHDR_SIZE;
	let stroff: u64 = rdu64(buf, strsh + SHDR_OFFSET);
	let strtab:  *u8 = buf + stroff;

	let datasize: u64 = 0u64;
	let dataoff:   u64 = 0u64;
	if (idxdata >= 0) {
		let datash: u64 = shoff + (idxdata: u64) * SHDR_SIZE;
		dataoff  = rdu64(buf, datash + SHDR_OFFSET);
		datasize = rdu64(buf, datash + SHDR_SIZE_F);
	};

	// Track this object.
	let ob: *lobj = alloc(lobj {
		path     = cstrtostr(path),
		buf      = buf,
		len      = len,
		textoff  = l.textlen,
		textsize = textsize,
		dataoff  = l.datalen,
		datasize = datasize,
		onext    = l.objs,
	})!;
	l.objs = ob;

	// Append .text bytes to the combined image.
	emittext(l, buf + textoff, textsize);
	// Append .data bytes (if present) to the combined .data buffer.
	if (idxdata >= 0) {
		if (datasize > 0u64) {
			emitdata(l, buf + dataoff, datasize);
		};
	};

	// Walk symbols. We don't keep a per-object map[] of *lsym. Instead
	// the reloc loop re-walks symtab and re-interns by name. Simpler
	// than dancing around the cgen's u64-shift gaps.
	let si: u64 = 1u64;	// skip index 0 (always undef sentinel)
	for (si < nsyms) {
		let symp: u64 = symoff + si * SYM_SIZE;
		let stname:  u32 = rdu32(buf, symp + SYM_NAME);
		let stshndx: u16 = rdu16(buf, symp + SYM_SHNDX);
		let stvalue: u64 = rdu64(buf, symp + SYM_VALUE);
		let nmp: *u8 = strtab + (stname: u64);
		if (nmp[0u64] != 0u8) {
			let nm: str = cstrtostr(nmp);
			let gs: *lsym = intern(l, nm);
			if (stshndx != 0u16) {
				let intext: bool = (stshndx: i32) == idxtext;
				let indt:   bool = false;
				if (idxdata >= 0) {
					indt = (stshndx: i32) == idxdata;
				};
				if (intext) { if (indt) { indt = false; }; };
				if (intext) {
					if (gs.defined != 0) {
						os.write(2, "w6l: duplicate symbol\n".ptr, 21u64);
						l.errs += 1;
					} else {
						gs.defined = 1;
						gs.owner = ob;
						gs.idxinowner = si: i32;
						gs.val = ob.textoff + stvalue;
					};
				};
				if (indt) {
					if (gs.defined != 0) {
						os.write(2, "w6l: duplicate symbol\n".ptr, 21u64);
						l.errs += 1;
					} else {
						gs.defined = 1;
						gs.indata = 1;
						gs.owner = ob;
						gs.idxinowner = si: i32;
						gs.val = ob.dataoff + stvalue;
					};
				};
			};
		};
		si += 1u64;
	};

	// Per-object relocation collection.
	if (idxrela >= 0) {
		let relash:   u64 = shoff + (idxrela: u64) * SHDR_SIZE;
		let relaoff:  u64 = rdu64(buf, relash + SHDR_OFFSET);
		let relasize: u64 = rdu64(buf, relash + SHDR_SIZE_F);
		let nrel:      u64 = relasize / RELA_SIZE;
		let ri: u64 = 0u64;
		for (ri < nrel) {
			let rp: u64 = relaoff + ri * RELA_SIZE;
			let roff:    u64 = rdu64(buf, rp + RELA_OFFSET);
			let rinfo:   u64 = rdu64(buf, rp + RELA_INFO);
			let raddend: u64 = rdu64(buf, rp + RELA_ADDEND);
			let rsymidx: u32 = (rinfo >> 32u64): u32;
			let rkind:    i32 = ((rinfo & 4294967295u64): u32): i32;
			let nr: *lrel = alloc(lrel {
				off     = ob.textoff + roff,
				section = 0,
				kind    = rkind,
				addend  = raddend: i64,
				rnext   = l.rels,
			})!;
			// Look up the referenced sym by name (re-walk symtab).
			if ((rsymidx: u64) < nsyms) {
				let sp: u64 = symoff + (rsymidx: u64) * SYM_SIZE;
				let sname: u32 = rdu32(buf, sp + SYM_NAME);
				let snm: *u8 = strtab + (sname: u64);
				if (snm[0u64] != 0u8) {
					let nm: str = cstrtostr(snm);
					nr.sym = intern(l, nm);
				};
			};
			l.rels = nr;
			ri += 1u64;
		};
	};

	// Data-reloc collection. Offsets land in .data, shifted by
	// this object's data_off so they index the combined buffer.
	if (idxrelad >= 0) {
		let relash:   u64 = shoff + (idxrelad: u64) * SHDR_SIZE;
		let relaoff:  u64 = rdu64(buf, relash + SHDR_OFFSET);
		let relasize: u64 = rdu64(buf, relash + SHDR_SIZE_F);
		let nrel:      u64 = relasize / RELA_SIZE;
		let ri: u64 = 0u64;
		for (ri < nrel) {
			let rp: u64 = relaoff + ri * RELA_SIZE;
			let roff:    u64 = rdu64(buf, rp + RELA_OFFSET);
			let rinfo:   u64 = rdu64(buf, rp + RELA_INFO);
			let raddend: u64 = rdu64(buf, rp + RELA_ADDEND);
			let rsymidx: u32 = (rinfo >> 32u64): u32;
			let rkind:    i32 = ((rinfo & 4294967295u64): u32): i32;
			let nr: *lrel = alloc(lrel {
				off     = ob.dataoff + roff,
				section = 1,
				kind    = rkind,
				addend  = raddend: i64,
				rnext   = l.rels,
			})!;
			if ((rsymidx: u64) < nsyms) {
				let sp: u64 = symoff + (rsymidx: u64) * SYM_SIZE;
				let sname: u32 = rdu32(buf, sp + SYM_NAME);
				let snm: *u8 = strtab + (sname: u64);
				if (snm[0u64] != 0u8) {
					let nm: str = cstrtostr(snm);
					nr.sym = intern(l, nm);
				};
			};
			l.rels = nr;
			ri += 1u64;
		};
	};

	return 0;
};

// selfhost/cmd/w6l/dyn.ww — port of cmd/w6l/dyn.c.
//
// Load a shared object (ET_DYN) so the linker knows which symbols it
// exports and which DT_NEEDED entry to record. We do not pull bytes
// from the .so; the dynamic loader maps it at runtime.
//
// Each call appends one lso to lnk->sos. l_so_provides_v answers
// "does this .so export the named symbol, and at which version?" —
// l_resolve uses that to promote unresolved references to dynamic.

package w6l;

import os;
import rt;
import strings;
import sym;

def ET_DYN_SO:     u16 = 3u16;
def EM_X86_64_SO:  u16 = 62u16;

def SHT_DYNAMIC:        u32 = 6u32;
def SHT_DYNSYM:         u32 = 11u32;
// GNU extensions, sh_type values.
def SHT_GNU_VERDEF:     u32 = 1879048189u32;	// 0x6ffffffd
def SHT_GNU_VERNEED:    u32 = 1879048190u32;	// 0x6ffffffe
def SHT_GNU_VERSYM:     u32 = 1879048191u32;	// 0x6fffffff

def DT_NULL_TAG:        i64 = 0i64;
def DT_SONAME_TAG:      i64 = 14i64;

// Versym special values.
def VER_NDX_LOCAL_C:    u16 = 0u16;
def VER_NDX_GLOBAL_C:   u16 = 1u16;
def VERSYM_HIDDEN_C:    u16 = 32768u16;	// 0x8000
def VERSYM_VERSION_C:   u16 = 32767u16;	// 0x7fff

// ELF64 ehdr field offsets (subset)
def EH_SHOFF:      u64 = 40u64;
def EH_ETYPE:      u64 = 16u64;
def EH_EMACHINE:   u64 = 18u64;
def EH_SHENTSIZE:  u64 = 58u64;
def EH_SHNUM:      u64 = 60u64;
def EH_SHSTRNDX:   u64 = 62u64;

// ELF64 Shdr (64 bytes)
def SH_SIZE:       u64 = 64u64;
def SH_TYPE:       u64 = 4u64;
def SH_OFFSET:     u64 = 24u64;
def SH_SIZE_F:     u64 = 32u64;
def SH_LINK:       u64 = 40u64;
def SH_ENTSIZE:    u64 = 56u64;

// ELF64 Sym (24 bytes)
def SY_SIZE:       u64 = 24u64;
def SY_NAME:       u64 = 0u64;
def SY_INFO:       u64 = 4u64;
def SY_SHNDX:      u64 = 6u64;

// ELF64 Dyn (16 bytes)
def DY_SIZE:       u64 = 16u64;
def DY_TAG:        u64 = 0u64;
def DY_VAL:        u64 = 8u64;

// Verdef (20 bytes)
def VD_SIZE:       u64 = 20u64;
def VD_NDX:        u64 = 4u64;
def VD_CNT:        u64 = 6u64;
def VD_AUX:        u64 = 12u64;
def VD_NEXT:       u64 = 16u64;

// Verdaux (8 bytes)
def VA_NAME:       u64 = 0u64;
def VA_NEXT:       u64 = 4u64;

// ---- little-endian byte readers ---------------------------------------

fn du16(p: *u8, off: u64) u16 = {
	let b0: u16 = p[off]: u16;
	let b1: u16 = p[off + 1u64]: u16;
	return b0 | (b1 << 8u16);
};

fn du32(p: *u8, off: u64) u32 = {
	let b0: u32 = p[off]: u32;
	let b1: u32 = p[off + 1u64]: u32;
	let b2: u32 = p[off + 2u64]: u32;
	let b3: u32 = p[off + 3u64]: u32;
	return b0 | (b1 << 8u32) | (b2 << 16u32) | (b3 << 24u32);
};

fn du64(p: *u8, off: u64) u64 = {
	let lo: u64 = du32(p, off): u64;
	let hi: u64 = du32(p, off + 4u64): u64;
	return lo | (hi << 32u64);
};

fn di64(p: *u8, off: u64) i64 = {
	return du64(p, off): i64;
};

// ---- C-string helpers --------------------------------------------------

fn dcstrlen(p: *u8) u64 = {
	let n: u64 = 0u64;
	for (p[n] != 0u8) { n += 1u64; };
	return n;
};

fn dcstrtostr(p: *u8) str = {
	let n: u64 = dcstrlen(p);
	let view: str;
	view.ptr = p;
	view.len = n: i32;
	return strings.dup(view);
};

// basename: scan for last '/' and return pointer past it.
fn dbasename(p: *u8) *u8 = {
	let n: u64 = dcstrlen(p);
	let i: u64 = n;
	for (i > 0u64) {
		i -= 1u64;
		if (p[i] == 47u8) {	// '/'
			return p + i + 1u64;
		};
	};
	return p;
};

// ---- file slurp --------------------------------------------------------

fn slurpso(path: *u8) (*u8, u64) = {
	let fd: i32 = os.open(pathstr(path), os.flag.RDONLY, 0i32);
	if (fd < 0) { return nil, 0u64; };
	let szr: (i64 | os.oserror) = os.filesize(fd);
	let n: i64 = 0i64;
	match (szr) {
	case let v: i64 => n = v;
	case let e: os.oserror => { os.close(fd); return nil, 0u64; };
	};
	let buf: []u8 = alloc([], n: u64)!;
	buf.len = n: i32;
	let rr: (i64 | os.oserror) = os.readall(fd, buf.ptr, n: u64);
	os.close(fd);
	let got: i64 = 0i64;
	match (rr) {
	case let v: i64 => got = v;
	case let e: os.oserror => return nil, 0u64;
	};
	if (got != n) { return nil, 0u64; };
	return buf.ptr, n: u64;
};

// ---- verdef helpers ----------------------------------------------------

// vdnameat — walk verdef records and return the name (as *u8 into
// the .so's verstr buffer) for the entry whose vd_ndx == ndx. The name
// is the first Verdaux's vda_name (subsequent auxes are predecessor
// names). Returns nil if no entry matches.
fn vdnameat(buf: *u8, verdefoff: u64, verdefsize: u64,
		verstr: *u8, ndx: u16) *u8 = {
	let off: u64 = 0u64;
	for (off < verdefsize) {
		let vdp: u64 = verdefoff + off;
		let vdndx: u16 = du16(buf, vdp + VD_NDX);
		let vdaux: u32 = du32(buf, vdp + VD_AUX);
		let vdnext: u32 = du32(buf, vdp + VD_NEXT);
		if (vdndx == ndx) {
			let auxp: u64 = vdp + (vdaux: u64);
			let vdaname: u32 = du32(buf, auxp + VA_NAME);
			return verstr + (vdaname: u64);
		};
		if (vdnext == 0u32) { return nil; };
		off += vdnext: u64;
	};
	return nil;
};

// ---- entry points ------------------------------------------------------

export fn loadso(l: *lnk, path: *u8) i32 = {
	let buf: *u8;
	let blen: u64;
	buf, blen = slurpso(path);
	if (buf == nil) {
		os.write(2, "w6l: cannot read .so\n".ptr, 20u64);
		return -1;
	};
	if (blen < 64u64) {
		os.write(2, "w6l: short ELF\n".ptr, 14u64);
		return -1;
	};
	if (buf[0u64] != 127u8) { return soerr("not ELF"); };
	if (buf[1u64] != 69u8)  { return soerr("not ELF"); };
	if (buf[2u64] != 76u8)  { return soerr("not ELF"); };
	if (buf[3u64] != 70u8)  { return soerr("not ELF"); };
	if (buf[4u64] != 2u8)   { return soerr("not ELFCLASS64"); };
	if (du16(buf, EH_EMACHINE) != EM_X86_64_SO) {
		return soerr("not amd64");
	};
	if (du16(buf, EH_ETYPE) != ET_DYN_SO) {
		return soerr("not ET_DYN");
	};

	let shoff: u64 = du64(buf, EH_SHOFF);
	let shnum: u32 = du16(buf, EH_SHNUM): u32;
	if (shoff == 0u64) { return soerr("stripped .so unsupported"); };
	if (shnum == 0u32) { return soerr("stripped .so unsupported"); };

	// Locate the four sections we care about.
	let idxdynsym:  i32 = -1;
	let idxdynamic: i32 = -1;
	let idxversym:  i32 = -1;
	let idxverdef:  i32 = -1;
	let i: u32 = 0u32;
	for (i < shnum) {
		let shp: u64 = shoff + (i: u64) * SH_SIZE;
		let shtype: u32 = du32(buf, shp + SH_TYPE);
		if (shtype == SHT_DYNSYM)       { idxdynsym = i: i32; };
		if (shtype == SHT_DYNAMIC)      { idxdynamic = i: i32; };
		if (shtype == SHT_GNU_VERSYM)   { idxversym = i: i32; };
		if (shtype == SHT_GNU_VERDEF)   { idxverdef = i: i32; };
		i += 1u32;
	};
	if (idxdynsym < 0) {
		return soerr("no .dynsym");
	};

	let dynsymsh: u64 = shoff + (idxdynsym: u64) * SH_SIZE;
	let dynsymoff: u64 = du64(buf, dynsymsh + SH_OFFSET);
	let dynsymsize: u64 = du64(buf, dynsymsh + SH_SIZE_F);
	let dynsymlink: u32 = du32(buf, dynsymsh + SH_LINK);
	let nsyms: u64 = dynsymsize / SY_SIZE;

	let dynstrsh: u64 = shoff + (dynsymlink: u64) * SH_SIZE;
	let dynstroff: u64 = du64(buf, dynstrsh + SH_OFFSET);
	let dynstr: *u8 = buf + dynstroff;

	// SONAME: .dynamic strings live in the section pointed at by its
	// sh_link (almost always .dynstr).
	let sonamecs: *u8 = nil;
	if (idxdynamic >= 0) {
		let dynsh: u64 = shoff + (idxdynamic: u64) * SH_SIZE;
		let dynoff: u64 = du64(buf, dynsh + SH_OFFSET);
		let dynsize: u64 = du64(buf, dynsh + SH_SIZE_F);
		let dynlink: u32 = du32(buf, dynsh + SH_LINK);
		let dstrsh: u64 = shoff + (dynlink: u64) * SH_SIZE;
		let dstroff: u64 = du64(buf, dstrsh + SH_OFFSET);
		let dstr: *u8 = buf + dstroff;
		let nd: u64 = dynsize / DY_SIZE;
		let di: u64 = 0u64;
		for (di < nd) {
			let dp: u64 = dynoff + di * DY_SIZE;
			let dtag: i64 = di64(buf, dp + DY_TAG);
			if (dtag == DT_NULL_TAG) {
				di = nd;	// break
			} else {
				if (dtag == DT_SONAME_TAG) {
					let dval: u64 = du64(buf, dp + DY_VAL);
					sonamecs = dstr + dval;
					di = nd;	// break
				} else {
					di += 1u64;
				};
			};
		};
	};
	if (sonamecs == nil) {
		sonamecs = dbasename(path);
	};

	// Versym is one u16 per dynsym entry.
	let versymoff: u64 = 0u64;
	let hasversym: i32 = 0;
	if (idxversym >= 0) {
		let vssh: u64 = shoff + (idxversym: u64) * SH_SIZE;
		versymoff = du64(buf, vssh + SH_OFFSET);
		hasversym = 1;
	};

	// Verdef section bounds + the .dynstr-like string section it uses.
	let verdefoff: u64 = 0u64;
	let verdefsize: u64 = 0u64;
	let verstr: *u8 = nil;
	if (idxverdef >= 0) {
		let vdsh: u64 = shoff + (idxverdef: u64) * SH_SIZE;
		verdefoff = du64(buf, vdsh + SH_OFFSET);
		verdefsize = du64(buf, vdsh + SH_SIZE_F);
		let vdlink: u32 = du32(buf, vdsh + SH_LINK);
		let vstrsh: u64 = shoff + (vdlink: u64) * SH_SIZE;
		let vstroff: u64 = du64(buf, vstrsh + SH_OFFSET);
		verstr = buf + vstroff;
	};

	// Build the lso. Exports are appended in dynsym order so
	// soprovides_v's first-match semantics match the C version.
	let so: *lso = alloc(lso { path = dcstrtostr(path), soname = dcstrtostr(sonamecs) })!;
	let tail: *lexport = nil;

	let si: u64 = 1u64;
	for (si < nsyms) {
		let sp: u64 = dynsymoff + si * SY_SIZE;
		let stshndx: u16 = du16(buf, sp + SY_SHNDX);
		if (stshndx == 0u16) { si += 1u64; } else {
			let stinfo: u8 = buf[sp + SY_INFO];
			let bind: u32 = (stinfo: u32) >> 4u32;
			if (bind != 1u32) { if (bind != 2u32) {
				// not GLOBAL/WEAK
				si += 1u64;
				continue;
			}; };
			let stname: u32 = du32(buf, sp + SY_NAME);
			let nmp: *u8 = dynstr + (stname: u64);
			if (nmp[0u64] == 0u8) {
				si += 1u64;
				continue;
			};

			// Determine version. Skip non-default (hidden) and
			// local entries.
			let vernamecs: *u8 = nil;
			let keep: i32 = 1;
			if (hasversym != 0) {
				let v: u16 = du16(buf, versymoff + si * 2u64);
				if ((v & VERSYM_HIDDEN_C) != 0u16) {
					keep = 0;	// non-default
				} else {
					let vidx: u16 = v & VERSYM_VERSION_C;
					if (vidx == VER_NDX_LOCAL_C) {
						keep = 0;	// not exported
					} else { if (vidx == VER_NDX_GLOBAL_C) {
						vernamecs = nil;
					} else { if (vidx == 1u16) {
						// glibc's BASE entry: treat as
						// unversioned. (The C version
						// notes that vidx==1 in Verdef
						// maps to the SONAME BASE.)
						vernamecs = nil;
					} else {
						if (verstr != nil) {
							let nm: *u8 = vdnameat(buf, verdefoff, verdefsize, verstr, vidx);
							vernamecs = nm;
						};
					}; }; };
				};
			};

			if (keep != 0) {
				let e: *lexport = alloc(lexport { name = dcstrtostr(nmp) })!;
				if (vernamecs != nil) {
					e.version = dcstrtostr(vernamecs);
				};
				if (tail == nil) {
					so.exports = e;
				} else {
					tail.enext = e;
				};
				tail = e;
			};
			si += 1u64;
		};
	};

	so.sonext = l.sos;
	l.sos = so;
	return 0;
};

fn soerr(msg: str) i32 = {
	os.write(2, "w6l: ".ptr, 4u64);
	os.write(2, msg.ptr, msg.len: u64);
	os.write(2, "\n".ptr, 1u64);
	return -1;
};

// soprovides — 1 if so exports name, 0 otherwise.
export fn soprovides(so: *lso, name: str) i32 = {
	if (so == nil) { return 0; };
	let e: *lexport = so.exports;
	for (e != nil) {
		if (streq(e.name, name)) { return 1; };
		e = e.enext;
	};
	return 0;
};

// soversion — the version of so's export named `name`, or an empty
// str (ptr=nil, len=0) if the export is unversioned or not present.
export fn soversion(so: *lso, name: str) str = {
	let result: str;
	result.ptr = nil;
	result.len = 0i32;
	if (so == nil) { return result; };
	let e: *lexport = so.exports;
	for (e != nil) {
		if (streq(e.name, name)) {
			result.ptr = e.version.ptr;
			result.len = e.version.len;
			return result;
		};
		e = e.enext;
	};
	return result;
};

// `streq` lives in sym.ww — same bundle, single definition.

// selfhost/cmd/w6l/pass.ww — port of cmd/w6l/pass.c.
//
// Resolution + relocation. l_resolve flags every undefined symbol
// referenced by a relocation, and promotes those provided by some
// loaded .so to "dynamic" with a freshly-assigned PLT slot.
// l_relocate walks the rel list and patches the .text bytes in place
// once the final virtual base is known. Dynamic refs are deferred:
// their site is patched later in dynout, once the PLT vaddr is known.
//
// Supported relocation kinds: PC32 (=2), PLT32 (=4); both are 32-bit
// PC-relative displacements (PLT32 == PC32 for static).

package w6l;

import os;
import sym;
import dyn;

def R_X86_64_64:    i32 = 1;
def R_X86_64_PC32:  i32 = 2;
def R_X86_64_PLT32: i32 = 4;

export fn resolve(l: *lnk) i32 = {
	// Initialise dynamic-linking sentinels. alloc(T{})! zeroes, so
	// isdyn/dynlib start clean — but pltidx and dynsymidx
	// must be -1, not 0.
	let si: *lsym = l.syms;
	for (si != nil) {
		si.pltidx = -1;
		si.dynsymidx = -1;
		si = si.snext;
	};

	// Promote each undefined sym that some lso exports to dynamic
	// and hand it a PLT slot. Iteration order over the relocation
	// list determines slot numbering and is stable across runs.
	let r: *lrel = l.rels;
	for (r != nil) {
		if (r.sym != nil) {
			if (r.sym.defined == 0) {
				if (r.sym.isdyn == 0) {
					let so: *lso = l.sos;
					for (so != nil) {
						if (soprovides(so, r.sym.name) != 0) {
							r.sym.isdyn = 1;
							r.sym.dynlib = so;
							r.sym.pltidx = l.dynn;
							l.dynn += 1;
							so = nil;	// break
						} else {
							so = so.sonext;
						};
					};
				};
			};
		};
		r = r.rnext;
	};

	// What remains undefined truly is undefined.
	let r2: *lrel = l.rels;
	for (r2 != nil) {
		if (r2.sym != nil) {
			if (r2.sym.defined == 0) {
				if (r2.sym.isdyn == 0) {
					os.write(2, "w6l: undefined reference to '".ptr, 28u64);
					let nm: str = r2.sym.name;
					os.write(2, nm.ptr, nm.len: u64);
					os.write(2, "'\n".ptr, 2u64);
					l.errs += 1;
				};
			};
		};
		r2 = r2.rnext;
	};
	return l.errs;
};

fn patchu32(p: *u8, v: u32) void = {
	p[0] = (v & 255u32): u8;
	p[1] = ((v >> 8u32) & 255u32): u8;
	p[2] = ((v >> 16u32) & 255u32): u8;
	p[3] = ((v >> 24u32) & 255u32): u8;
};

fn patchu64(p: *u8, v: u64) void = {
	let i: i32 = 0;
	for (i < 8) {
		p[i] = ((v >> (i: u64 * 8u64)) & 255u64): u8;
		i += 1;
	};
};

export fn relocate(l: *lnk, textva: u64, datava: u64) i32 = {
	let r: *lrel = l.rels;
	for (r != nil) {
		if (r.sym != nil) {
			// Dynamic refs are patched later in dynout once the
			// PLT vaddr is known.
			if (r.sym.isdyn != 0) {
				r = r.rnext;
				continue;
			};
			if (r.sym.defined != 0) {
				let symva: u64 = textva + r.sym.val;
				if (r.sym.indata != 0) { symva = datava + r.sym.val; };
				let k: i32 = r.kind;
				if (k == R_X86_64_PC32) {
					let site: u64 = textva + r.off;
					let rel: i64 = (symva: i64 - site: i64) + r.addend;
					patchu32(l.text + r.off, rel: u32);
				} else { if (k == R_X86_64_PLT32) {
					let site: u64 = textva + r.off;
					let rel: i64 = (symva: i64 - site: i64) + r.addend;
					patchu32(l.text + r.off, rel: u32);
				} else { if (k == R_X86_64_64) {
					// Absolute 64-bit. Currently used only
					// for DATAR slots in .data.
					let v: u64 = (symva: i64 + r.addend): u64;
					if (r.section == 1) {
						patchu64(l.data + r.off, v);
					} else {
						patchu64(l.text + r.off, v);
					};
				} else {
					os.write(2, "w6l: unsupported reloc kind\n".ptr, 27u64);
					l.errs += 1;
				};};};
			};
		};
		r = r.rnext;
	};
	return l.errs;
};

// selfhost/cmd/w6l/dynout.ww — port of cmd/w6l/dynout.c.
//
// Emit a dynamic-linked ELF executable. The shape is the simplest
// valid one: PT_INTERP + PT_DYNAMIC + DT_BIND_NOW so the loader
// resolves every PLT slot at startup (no lazy binding, no PLT0
// trampoline). SysV .hash, not .gnu.hash. Non-PIE, fixed base.
//
// Layout:
//   [0]            Ehdr
//   [64]           Phdrs (PT_LOAD R+X, PT_LOAD R+W, PT_INTERP, PT_DYNAMIC)
//   [interp_off]   "/lib64/ld-linux-x86-64.so.2\0"
//   [dynstr_off]   .dynstr
//   [dynsym_off]   .dynsym
//   [hash_off]     .hash
//   [versym_off]   .gnu.version
//   [verneed_off]  .gnu.version_r
//   [relaplt_off]  .rela.plt
//   [pad to 0x1000]
//   [text_off]     .text
//   [plt_off]      .plt
//   [pad to next page]
//   [gotplt_off]   .got.plt   (writable; mapped by PT_LOAD #2)
//   [dynamic_off]  .dynamic   (writable; covered by PT_DYNAMIC)

package w6l;

import os;
import rt;
import sym;

// ELF constants
def ET_EXEC_D:     u16 = 2u16;
def EM_X86_64_D:   u16 = 62u16;
def EV_CURRENT_D:  u32 = 1u32;
def ELFCLASS64_D:  u8  = 2u8;
def ELFDATA2LSB_D: u8  = 1u8;

def PT_LOAD_D:     u32 = 1u32;
def PT_DYNAMIC_D:  u32 = 2u32;
def PT_INTERP_D:   u32 = 3u32;
def PF_X_D:        u32 = 1u32;
def PF_W_D:        u32 = 2u32;
def PF_R_D:        u32 = 4u32;

def DT_NULL:       i64 = 0i64;
def DT_NEEDED:     i64 = 1i64;
def DT_PLTRELSZ:   i64 = 2i64;
def DT_PLTGOT:     i64 = 3i64;
def DT_HASH:       i64 = 4i64;
def DT_STRTAB:     i64 = 5i64;
def DT_SYMTAB:     i64 = 6i64;
def DT_STRSZ:      i64 = 10i64;
def DT_SYMENT:     i64 = 11i64;
def DT_PLTREL:     i64 = 20i64;
def DT_RELA:       i64 = 7i64;
def DT_JMPREL:     i64 = 23i64;
def DT_BIND_NOW:   i64 = 24i64;
def DT_VERSYM:     i64 = 1879048176i64;	// 0x6ffffff0
def DT_VERNEED:    i64 = 1879048190i64;	// 0x6ffffffe
def DT_VERNEEDNUM: i64 = 1879048191i64;	// 0x6fffffff

def VER_NDX_LOCAL_D:  u16 = 0u16;
def VER_NDX_GLOBAL_D: u16 = 1u16;

def R_X86_64_PC32_D:        i32 = 2;
def R_X86_64_PLT32_D:       i32 = 4;
def R_X86_64_JUMP_SLOT_D:   u32 = 7u32;

def STB_GLOBAL_D: u8 = 1u8;
def STT_FUNC_D:   u8 = 2u8;

def PLT_STUB_BYTES_D: u64 = 8u64;
def PAGE: u64 = 4096u64;

def INTERP: str = "/lib64/ld-linux-x86-64.so.2";

// ---- byte writers ------------------------------------------------------

fn dwr8(buf: *u8, off: u64, v: u8) void = {
	buf[off] = v;
};

fn dwr16(buf: *u8, off: u64, v: u16) void = {
	buf[off]         = (v & 255u16): u8;
	buf[off + 1u64]  = ((v >> 8u16) & 255u16): u8;
};

fn dwr32(buf: *u8, off: u64, v: u32) void = {
	buf[off]         = (v & 255u32): u8;
	buf[off + 1u64]  = ((v >> 8u32)  & 255u32): u8;
	buf[off + 2u64]  = ((v >> 16u32) & 255u32): u8;
	buf[off + 3u64]  = ((v >> 24u32) & 255u32): u8;
};

fn dwr64(buf: *u8, off: u64, v: u64) void = {
	dwr32(buf, off,        (v & 4294967295u64): u32);
	dwr32(buf, off + 4u64, ((v >> 32u64) & 4294967295u64): u32);
};

fn dwri64(buf: *u8, off: u64, v: i64) void = {
	dwr64(buf, off, v: u64);
};

fn dwri32(buf: *u8, off: u64, v: i32) void = {
	dwr32(buf, off, v: u32);
};

// ---- byte readers ------------------------------------------------------

fn drdu16(p: *u8, off: u64) u16 = {
	let b0: u16 = p[off]: u16;
	let b1: u16 = p[off + 1u64]: u16;
	return b0 | (b1 << 8u16);
};

fn drdu32(p: *u8, off: u64) u32 = {
	let b0: u32 = p[off]: u32;
	let b1: u32 = p[off + 1u64]: u32;
	let b2: u32 = p[off + 2u64]: u32;
	let b3: u32 = p[off + 3u64]: u32;
	return b0 | (b1 << 8u32) | (b2 << 16u32) | (b3 << 24u32);
};

fn drdi32(p: *u8, off: u64) i32 = {
	return drdu32(p, off): i32;
};

fn dbcopy(dst: *u8, off: u64, src: *u8, n: u64) void = {
	let i: u64 = 0u64;
	for (i < n) {
		dst[off + i] = src[i];
		i += 1u64;
	};
};

// elfhash — SysV ELF hash. Used for .gnu.version_r's vna_hash.
fn elfhash(name: str) u32 = {
	let h: u32 = 0u32;
	let i: i32 = 0;
	for (i < name.len) {
		let c: u32 = (name[i]: u8): u32;
		h = (h << 4u32) + c;
		let g: u32 = h & 4026531840u32;	// 0xf0000000
		if (g != 0u32) { h = h ^ (g >> 24u32); };
		h = h & ~g;
		i += 1;
	};
	return h;
};

fn alignup(off: u64, a: u64) u64 = {
	return (off + a - 1u64) & ~(a - 1u64);
};

fn streqd(a: str, b: str) bool = {
	if (a.len != b.len) { return false; };
	let i: i32 = 0;
	for (i < a.len) {
		if (a[i] != b[i]) { return false; };
		i += 1;
	};
	return true;
};

// ---- main entry --------------------------------------------------------

export fn emitdynelf(l: *lnk, fd: i32, base: u64, entry: u64) i32 = {
	// .data shares the R+W PT_LOAD with .got.plt and .dynamic.
	// Placed after .dynamic so the segment is one contiguous run;
	// relocate runs from here so the dyn layout's datava lands in
	// patched offsets.
	let n: i32 = l.dynn;
	let nu: u64 = n: u64;

	// ---- collect dyn syms into a plt_idx-indexed array ----
	let dynsyms: []*lsym = alloc([], nu)!;
	let s: *lsym = l.syms;
	for (s != nil) {
		if (s.isdyn != 0) {
			if (s.pltidx >= 0) {
				if (s.pltidx < n) {
					dynsyms[s.pltidx] = s;
				};
			};
		};
		s = s.snext;
	};
	let i: i32 = 0;
	for (i < n) {
		if (dynsyms[i] == nil) {
			os.write(2, "w6l: dynout: no sym for plt_idx\n".ptr, 31u64);
			return 1;
		};
		i += 1;
	};

	// ---- collect used .so's (in l.sos order) ----
	let maxsos: i32 = 0;
	let so: *lso = l.sos;
	for (so != nil) { maxsos += 1; so = so.sonext; };
	let sosused: []*lso = alloc([], maxsos: u64)!;
	let nsos: i32 = 0;
	so = l.sos;
	for (so != nil) {
		let used: i32 = 0;
		let j: i32 = 0;
		for (j < n) {
			let dsm: *lsym = dynsyms[j];
			let dl: *lso = dsm.dynlib;
			if (dl == so) { used = 1; j = n; }
			else { j += 1; };
		};
		if (used != 0) {
			sosused[nsos] = so;
			nsos += 1;
		};
		so = so.sonext;
	};

	// ---- build flat version table grouped by vlib ----
	// vlib_sos_idx[k] = sos_used index for vlib k.
	// vlib_first[k]   = ver index of first version under vlib k.
	// vlib_count[k]   = number of versions under vlib k.
	// ver_lib_idx[v]  = vlib index that version v belongs to.
	// ver_name_ptr_arr[v] = name's *u8 (interned in the .so's verdef strings).
	// ver_name_len_buf[v] = name length (i32).
	// ver_dynstr_off[v]   = offset within .dynstr (assigned after layout).
	// ver_vna_other[v]    = versym index (starting at 2).

	let vlibsosidxbuf: []u8 = alloc([], (maxsos: u64) * 4u64)!;
	let vlibfirstbuf:   []u8 = alloc([], (maxsos: u64) * 4u64)!;
	let vlibcountbuf:   []u8 = alloc([], (maxsos: u64) * 4u64)!;
	let nvlibs: i32 = 0;

	let verlibidxbuf:  []u8 = alloc([], nu * 4u64)!;
	let vernameptrarr: []*u8 = alloc([], nu)!;
	let vernamelenbuf: []u8 = alloc([], nu * 4u64)!;
	let verdynstroff:   []u8 = alloc([], nu * 4u64)!;
	let vervnaother:    []u8 = alloc([], nu * 2u64)!;
	let nvers: i32 = 0;

	let si: i32 = 0;
	for (si < nsos) {
		let curso: *lso = sosused[si];
		let has: i32 = 0;
		let j: i32 = 0;
		for (j < n) {
			let dsm: *lsym = dynsyms[j];
			let dl: *lso = dsm.dynlib;
			if (dl == curso) {
				let nm0: str = dsm.name;
				let dv: str = soversion(curso, nm0);
				if (dv.len > 0) {
					has = 1; j = n;
				} else { j += 1; };
			} else { j += 1; };
		};
		if (has != 0) {
			dwr32(vlibsosidxbuf.ptr, (nvlibs: u64) * 4u64, si: u32);
			dwr32(vlibfirstbuf.ptr,   (nvlibs: u64) * 4u64, nvers: u32);
			let added: i32 = 0;
			let jj: i32 = 0;
			for (jj < n) {
				let dsm2: *lsym = dynsyms[jj];
				let dl2: *lso = dsm2.dynlib;
				if (dl2 == curso) {
					let nm2: str = dsm2.name;
					let vname: str = soversion(curso, nm2);
					if (vname.len > 0) {
						let seen: i32 = 0;
						let k: i32 = 0;
						for (k < added) {
							let kk: i32 = nvers - added + k;
							let existing: str;
							existing.ptr = vernameptrarr[kk];
							existing.len = drdi32(vernamelenbuf.ptr, (kk: u64) * 4u64);
							if (streqd(existing, vname)) {
								seen = 1; k = added;
							} else { k += 1; };
						};
						if (seen == 0) {
							dwr32(verlibidxbuf.ptr, (nvers: u64) * 4u64, nvlibs: u32);
							vernameptrarr[nvers] = vname.ptr;
							dwri32(vernamelenbuf.ptr, (nvers: u64) * 4u64, vname.len);
							nvers += 1;
							added += 1;
						};
					};
				};
				jj += 1;
			};
			dwr32(vlibcountbuf.ptr, (nvlibs: u64) * 4u64, added: u32);
			nvlibs += 1;
		};
		si += 1;
	};

	// Assign vna_other indices starting at 2, walking vlib then per-version.
	let nextvna: u16 = 2u16;
	let vi: i32 = 0;
	for (vi < nvlibs) {
		let first: i32 = drdi32(vlibfirstbuf.ptr, (vi: u64) * 4u64);
		let cnt:   i32 = drdi32(vlibcountbuf.ptr, (vi: u64) * 4u64);
		let k: i32 = 0;
		for (k < cnt) {
			dwr16(vervnaother.ptr, ((first + k): u64) * 2u64, nextvna);
			nextvna += 1u16;
			k += 1;
		};
		vi += 1;
	};

	// ---- compute dynstr size ----
	let dynstrsz: u64 = 1u64;	// leading NUL
	let pi: i32 = 0;
	for (pi < nsos) {
		let so4: *lso = sosused[pi];
		dynstrsz += so4.soname.len: u64;
		dynstrsz += 1u64;
		pi += 1;
	};
	pi = 0;
	for (pi < n) {
		let dsm4: *lsym = dynsyms[pi];
		dynstrsz += dsm4.name.len: u64;
		dynstrsz += 1u64;
		pi += 1;
	};
	pi = 0;
	for (pi < nvers) {
		let nmlen: i32 = drdi32(vernamelenbuf.ptr, (pi: u64) * 4u64);
		dynstrsz += nmlen: u64;
		dynstrsz += 1u64;
		pi += 1;
	};

	// ---- fill dynstr ----
	let dynstr: []u8 = alloc([], dynstrsz)!;
	let dynstrpos: u64 = 1u64;	// past leading NUL

	let sonamestr: []u8 = alloc([], (nsos: u64) * 4u64)!;
	pi = 0;
	for (pi < nsos) {
		dwr32(sonamestr.ptr, (pi: u64) * 4u64, dynstrpos: u32);
		let so2: *lso = sosused[pi];
		let snm: str = so2.soname;
		dbcopy(dynstr.ptr, dynstrpos, snm.ptr, snm.len: u64);
		dynstrpos += snm.len: u64;
		dynstr[dynstrpos] = 0u8;
		dynstrpos += 1u64;
		pi += 1;
	};
	let symnamestr: []u8 = alloc([], nu * 4u64)!;
	pi = 0;
	for (pi < n) {
		dwr32(symnamestr.ptr, (pi: u64) * 4u64, dynstrpos: u32);
		let dsm: *lsym = dynsyms[pi];
		let snm: str = dsm.name;
		dbcopy(dynstr.ptr, dynstrpos, snm.ptr, snm.len: u64);
		dynstrpos += snm.len: u64;
		dynstr[dynstrpos] = 0u8;
		dynstrpos += 1u64;
		pi += 1;
	};
	pi = 0;
	for (pi < nvers) {
		dwr32(verdynstroff.ptr, (pi: u64) * 4u64, dynstrpos: u32);
		let nmp: *u8 = vernameptrarr[pi];
		let nmlen: i32 = drdi32(vernamelenbuf.ptr, (pi: u64) * 4u64);
		dbcopy(dynstr.ptr, dynstrpos, nmp, nmlen: u64);
		dynstrpos += nmlen: u64;
		dynstr[dynstrpos] = 0u8;
		dynstrpos += 1u64;
		pi += 1;
	};

	// ---- per-dyn-sym versym index ----
	let versymfor: []u8 = alloc([], nu * 2u64)!;
	pi = 0;
	for (pi < n) {
		let dsm3: *lsym = dynsyms[pi];
		let dl3: *lso = dsm3.dynlib;
		let nm3: str = dsm3.name;
		let vname: str = soversion(dl3, nm3);
		if (vname.len == 0) {
			dwr16(versymfor.ptr, (pi: u64) * 2u64, VER_NDX_GLOBAL_D);
		} else {
			let matched: i32 = 0;
			let vk: i32 = 0;
			for (vk < nvers) {
				let vlx:  i32 = drdi32(verlibidxbuf.ptr, (vk: u64) * 4u64);
				let sosx: i32 = drdi32(vlibsosidxbuf.ptr, (vlx: u64) * 4u64);
				if (sosused[sosx] == dl3) {
					let exi: str;
					exi.ptr = vernameptrarr[vk];
					exi.len = drdi32(vernamelenbuf.ptr, (vk: u64) * 4u64);
					if (streqd(exi, vname)) {
						let other: u16 = drdu16(vervnaother.ptr, (vk: u64) * 2u64);
						dwr16(versymfor.ptr, (pi: u64) * 2u64, other);
						matched = 1;
						vk = nvers;
					} else { vk += 1; };
				} else { vk += 1; };
			};
			if (matched == 0) {
				dwr16(versymfor.ptr, (pi: u64) * 2u64, VER_NDX_GLOBAL_D);
			};
		};
		pi += 1;
	};

	// ---- compute byte sizes ----
	let ehdrsz: u64 = 64u64;
	let nphdrs: u64 = 4u64;
	let phdrsz: u64 = nphdrs * 56u64;
	let interpsz: u64 = (INTERP.len: u64) + 1u64;

	let nsymstotal: u64 = 1u64 + nu;
	let dynsymsz: u64 = nsymstotal * 24u64;

	let nbuckets: u32 = 1u32;
	let nchain: u32 = nsymstotal: u32;
	let hashsz: u64 = (2u64 + (nbuckets: u64) + (nchain: u64)) * 4u64;

	let relapltsz: u64 = nu * 24u64;
	let pltsz: u64 = nu * PLT_STUB_BYTES_D;
	let gotpltsz: u64 = (3u64 + nu) * 8u64;
	let versymsz: u64 = nsymstotal * 2u64;

	let verneedsz: u64 = 0u64;
	let vli: i32 = 0;
	for (vli < nvlibs) {
		let cnt: i32 = drdi32(vlibcountbuf.ptr, (vli: u64) * 4u64);
		verneedsz += 16u64 + 16u64 * (cnt: u64);
		vli += 1;
	};

	let withver: i32 = 0;
	if (nvlibs > 0) { withver = 1; };
	let extra: u64 = 0u64;
	if (withver != 0) { extra = 3u64; };
	let ndyn: u64 = (nsos: u64) + 11u64 + extra;
	let dynamicsz: u64 = ndyn * 16u64;

	// ---- compute file offsets ----
	let off: u64 = ehdrsz + phdrsz;
	let interpoff: u64 = off;       off += interpsz;
	off = alignup(off, 8u64);
	let dynstroff: u64 = off;       off += dynstrsz;
	off = alignup(off, 8u64);
	let dynsymoff: u64 = off;       off += dynsymsz;
	let hashoff: u64 = off;         off += hashsz;
	off = alignup(off, 2u64);
	let versymoff: u64 = off;       off += versymsz;
	off = alignup(off, 4u64);
	let verneedoff: u64 = off;      off += verneedsz;
	off = alignup(off, 8u64);
	let relapltoff: u64 = off;      off += relapltsz;

	let textoff: u64 = alignup(off, PAGE);
	let pltoff: u64 = textoff + l.textlen;
	let rxend: u64 = pltoff + pltsz;

	let gotpltoff: u64 = alignup(rxend, PAGE);
	let dynamicoff: u64 = gotpltoff + gotpltsz;
	let dataoff: u64 = dynamicoff + dynamicsz;
	let fileend: u64 = dataoff + l.datalen;

	let interpva: u64 = base + interpoff;
	let dynstrva: u64 = base + dynstroff;
	let dynsymva: u64 = base + dynsymoff;
	let hashva: u64 = base + hashoff;
	let versymva: u64 = base + versymoff;
	let verneedva: u64 = base + verneedoff;
	let relapltva: u64 = base + relapltoff;
	let textva: u64 = base + textoff;
	let pltva: u64 = base + pltoff;
	let gotpltva: u64 = base + gotpltoff;
	let dynamicva: u64 = base + dynamicoff;
	let datava: u64 = base + dataoff;

	// Apply relocations now that the dyn layout's textva/datava are
	// pinned. main.ww defers this so each path uses its own VAs.
	if (relocate(l, textva, datava) != 0) { return 1; };

	// BSS optimisation — same trailing-zero scan as out.ww.
	let bsslen: u64 = 0u64;
	if (l.datalen > 0u64) {
		for (bsslen < l.datalen) {
			let b: u8 = l.data[l.datalen - 1u64 - bsslen];
			if (b != 0u8) { break; };
			bsslen += 1u64;
		};
	};
	let datafilelen: u64 = l.datalen - bsslen;
	let filedataend: u64 = dataoff + datafilelen;

	// ---- build .dynsym ----
	let dynsymbuf: []u8 = alloc([], dynsymsz)!;
	pi = 0;
	for (pi < n) {
		let eoff: u64 = (1u64 + (pi: u64)) * 24u64;
		dwr32(dynsymbuf.ptr, eoff + 0u64, drdu32(symnamestr.ptr, (pi: u64) * 4u64));
		dwr8(dynsymbuf.ptr, eoff + 4u64, (STB_GLOBAL_D << 4u8) | (STT_FUNC_D & 15u8));
		dwr8(dynsymbuf.ptr, eoff + 5u64, 0u8);
		dwr16(dynsymbuf.ptr, eoff + 6u64, 0u16);
		dwr64(dynsymbuf.ptr, eoff + 8u64, 0u64);
		dwr64(dynsymbuf.ptr, eoff + 16u64, 0u64);
		pi += 1;
	};

	// ---- build .hash (SysV, 1 bucket) ----
	let hashbuf: []u8 = alloc([], hashsz)!;
	dwr32(hashbuf.ptr, 0u64, nbuckets);
	dwr32(hashbuf.ptr, 4u64, nchain);
	let bucket0: u32 = 0u32;
	if (nsymstotal > 1u64) { bucket0 = 1u32; };
	dwr32(hashbuf.ptr, 8u64, bucket0);
	let ci: u64 = 1u64;
	for (ci < nsymstotal) {
		let nxt: u32 = 0u32;
		if (ci + 1u64 < nsymstotal) { nxt = (ci + 1u64): u32; };
		dwr32(hashbuf.ptr, 8u64 + (nbuckets: u64) * 4u64 + ci * 4u64, nxt);
		ci += 1u64;
	};

	// ---- build .rela.plt ----
	let relapltbuf: []u8 = alloc([], relapltsz)!;
	pi = 0;
	for (pi < n) {
		let roff: u64 = (pi: u64) * 24u64;
		dwr64(relapltbuf.ptr, roff + 0u64, gotpltva + (3u64 + (pi: u64)) * 8u64);
		let info: u64 = ((1u64 + (pi: u64)) << 32u64) | (R_X86_64_JUMP_SLOT_D: u64);
		dwr64(relapltbuf.ptr, roff + 8u64, info);
		dwri64(relapltbuf.ptr, roff + 16u64, 0i64);
		pi += 1;
	};

	// ---- build .gnu.version (u16 per dynsym entry) ----
	let versymbuf: []u8 = alloc([], versymsz)!;
	dwr16(versymbuf.ptr, 0u64, VER_NDX_LOCAL_D);
	pi = 0;
	for (pi < n) {
		dwr16(versymbuf.ptr, 2u64 + (pi: u64) * 2u64, drdu16(versymfor.ptr, (pi: u64) * 2u64));
		pi += 1;
	};

	// ---- build .gnu.version_r ----
	let verneedbuf: []u8 = alloc([], verneedsz)!;
	if (verneedsz > 0u64) {
		let vnoff: u64 = 0u64;
		vli = 0;
		for (vli < nvlibs) {
			let sosidx: i32 = drdi32(vlibsosidxbuf.ptr, (vli: u64) * 4u64);
			let first:   i32 = drdi32(vlibfirstbuf.ptr,   (vli: u64) * 4u64);
			let cnt:     i32 = drdi32(vlibcountbuf.ptr,   (vli: u64) * 4u64);
			let vnstart: u64 = vnoff;
			dwr16(verneedbuf.ptr, vnoff + 0u64, 1u16);
			dwr16(verneedbuf.ptr, vnoff + 2u64, cnt: u16);
			dwr32(verneedbuf.ptr, vnoff + 4u64, drdu32(sonamestr.ptr, (sosidx: u64) * 4u64));
			dwr32(verneedbuf.ptr, vnoff + 8u64, 16u32);
			vnoff += 16u64;
			let k: i32 = 0;
			for (k < cnt) {
				let vk: i32 = first + k;
				let nm: str;
				nm.ptr = vernameptrarr[vk];
				nm.len = drdi32(vernamelenbuf.ptr, (vk: u64) * 4u64);
				let h: u32 = elfhash(nm);
				dwr32(verneedbuf.ptr, vnoff + 0u64, h);
				dwr16(verneedbuf.ptr, vnoff + 4u64, 0u16);
				dwr16(verneedbuf.ptr, vnoff + 6u64, drdu16(vervnaother.ptr, (vk: u64) * 2u64));
				dwr32(verneedbuf.ptr, vnoff + 8u64, drdu32(verdynstroff.ptr, (vk: u64) * 4u64));
				let nxt: u32 = 0u32;
				if (k + 1 < cnt) { nxt = 16u32; };
				dwr32(verneedbuf.ptr, vnoff + 12u64, nxt);
				vnoff += 16u64;
				k += 1;
			};
			let vnnxt: u32 = 0u32;
			if (vli + 1 < nvlibs) { vnnxt = (vnoff - vnstart): u32; };
			dwr32(verneedbuf.ptr, vnstart + 12u64, vnnxt);
			vli += 1;
		};
	};

	// ---- build .plt ----
	let pltbuf: []u8 = alloc([], pltsz)!;
	pi = 0;
	for (pi < n) {
		let poff: u64 = (pi: u64) * PLT_STUB_BYTES_D;
		let stubva: u64 = pltva + poff;
		let nextip: u64 = stubva + 6u64;
		let slotva: u64 = gotpltva + (3u64 + (pi: u64)) * 8u64;
		let disp: i64 = (slotva: i64) - (nextip: i64);
		dwr8(pltbuf.ptr, poff + 0u64, 255u8);
		dwr8(pltbuf.ptr, poff + 1u64, 37u8);
		dwr32(pltbuf.ptr, poff + 2u64, (disp: i32): u32);
		pi += 1;
	};

	// ---- build .got.plt ----
	let gotpltbuf: []u8 = alloc([], gotpltsz)!;
	dwr64(gotpltbuf.ptr, 0u64, dynamicva);

	// ---- build .dynamic ----
	let dynamicbuf: []u8 = alloc([], dynamicsz)!;
	let dk: u64 = 0u64;
	pi = 0;
	for (pi < nsos) {
		dwri64(dynamicbuf.ptr, dk * 16u64 + 0u64, DT_NEEDED);
		dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, drdu32(sonamestr.ptr, (pi: u64) * 4u64): u64);
		dk += 1u64;
		pi += 1;
	};
	dwri64(dynamicbuf.ptr, dk * 16u64, DT_HASH);     dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, hashva);     dk += 1u64;
	dwri64(dynamicbuf.ptr, dk * 16u64, DT_STRTAB);   dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, dynstrva);   dk += 1u64;
	dwri64(dynamicbuf.ptr, dk * 16u64, DT_SYMTAB);   dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, dynsymva);   dk += 1u64;
	dwri64(dynamicbuf.ptr, dk * 16u64, DT_STRSZ);    dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, dynstrsz);   dk += 1u64;
	dwri64(dynamicbuf.ptr, dk * 16u64, DT_SYMENT);   dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, 24u64);       dk += 1u64;
	dwri64(dynamicbuf.ptr, dk * 16u64, DT_PLTGOT);   dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, gotpltva);   dk += 1u64;
	dwri64(dynamicbuf.ptr, dk * 16u64, DT_PLTRELSZ); dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, relapltsz);  dk += 1u64;
	dwri64(dynamicbuf.ptr, dk * 16u64, DT_PLTREL);   dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, DT_RELA: u64);dk += 1u64;
	dwri64(dynamicbuf.ptr, dk * 16u64, DT_JMPREL);   dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, relapltva);  dk += 1u64;
	dwri64(dynamicbuf.ptr, dk * 16u64, DT_BIND_NOW); dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, 0u64);        dk += 1u64;
	if (withver != 0) {
		dwri64(dynamicbuf.ptr, dk * 16u64, DT_VERSYM);     dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, versymva);   dk += 1u64;
		dwri64(dynamicbuf.ptr, dk * 16u64, DT_VERNEED);    dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, verneedva);  dk += 1u64;
		dwri64(dynamicbuf.ptr, dk * 16u64, DT_VERNEEDNUM); dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, nvlibs: u64);dk += 1u64;
	};
	dwri64(dynamicbuf.ptr, dk * 16u64, DT_NULL); dwr64(dynamicbuf.ptr, dk * 16u64 + 8u64, 0u64); dk += 1u64;
	if (dk != ndyn) {
		os.write(2, "w6l: dynamic entry count mismatch\n".ptr, 33u64);
		return 1;
	};

	// ---- patch .text relocs targeting dynamic syms ----
	let r: *lrel = l.rels;
	for (r != nil) {
		if (r.sym != nil) {
			let rsym: *lsym = r.sym;
			if (rsym.isdyn != 0) {
				if (r.kind != R_X86_64_PC32_D) {
					if (r.kind != R_X86_64_PLT32_D) {
						os.write(2, "w6l: dynamic reloc kind unsupported\n".ptr, 35u64);
						return 1;
					};
				};
				let site: u64 = textva + r.off;
				let stub: u64 = pltva + (rsym.pltidx: u64) * PLT_STUB_BYTES_D;
				let disp: i64 = (stub: i64) - (site: i64) + r.addend;
				dwr32(l.text, r.off, (disp: i32): u32);
			};
		};
		r = r.rnext;
	};

	// ---- assemble file buffer ----
	let filebuf: []u8 = alloc([], fileend)!;
	filebuf.len = fileend: i32;

	// Ehdr
	dwr8(filebuf.ptr, 0u64, 127u8);
	dwr8(filebuf.ptr, 1u64, 69u8);
	dwr8(filebuf.ptr, 2u64, 76u8);
	dwr8(filebuf.ptr, 3u64, 70u8);
	dwr8(filebuf.ptr, 4u64, ELFCLASS64_D);
	dwr8(filebuf.ptr, 5u64, ELFDATA2LSB_D);
	dwr8(filebuf.ptr, 6u64, EV_CURRENT_D: u8);
	dwr16(filebuf.ptr, 16u64, ET_EXEC_D);
	dwr16(filebuf.ptr, 18u64, EM_X86_64_D);
	dwr32(filebuf.ptr, 20u64, EV_CURRENT_D);
	dwr64(filebuf.ptr, 24u64, entry);
	dwr64(filebuf.ptr, 32u64, ehdrsz);
	dwr64(filebuf.ptr, 40u64, 0u64);
	dwr32(filebuf.ptr, 48u64, 0u32);
	dwr16(filebuf.ptr, 52u64, ehdrsz: u16);
	dwr16(filebuf.ptr, 54u64, 56u16);
	dwr16(filebuf.ptr, 56u64, nphdrs: u16);
	dwr16(filebuf.ptr, 58u64, 0u16);
	dwr16(filebuf.ptr, 60u64, 0u16);
	dwr16(filebuf.ptr, 62u64, 0u16);

	// Phdrs at offset 64.
	let p0: u64 = 64u64;
	dwr32(filebuf.ptr, p0 + 0u64,  PT_LOAD_D);
	dwr32(filebuf.ptr, p0 + 4u64,  PF_R_D | PF_X_D);
	dwr64(filebuf.ptr, p0 + 8u64,  0u64);
	dwr64(filebuf.ptr, p0 + 16u64, base);
	dwr64(filebuf.ptr, p0 + 24u64, base);
	dwr64(filebuf.ptr, p0 + 32u64, rxend);
	dwr64(filebuf.ptr, p0 + 40u64, rxend);
	dwr64(filebuf.ptr, p0 + 48u64, PAGE);

	let p1: u64 = 64u64 + 56u64;
	dwr32(filebuf.ptr, p1 + 0u64,  PT_LOAD_D);
	dwr32(filebuf.ptr, p1 + 4u64,  PF_R_D | PF_W_D);
	dwr64(filebuf.ptr, p1 + 8u64,  gotpltoff);
	dwr64(filebuf.ptr, p1 + 16u64, gotpltva);
	dwr64(filebuf.ptr, p1 + 24u64, gotpltva);
	// filesz trims the .data trailing zeros (BSS); memsz covers
	// .got.plt + .dynamic + the full .data so the loader zero-fills.
	dwr64(filebuf.ptr, p1 + 32u64, filedataend - gotpltoff);
	dwr64(filebuf.ptr, p1 + 40u64, fileend - gotpltoff);
	dwr64(filebuf.ptr, p1 + 48u64, PAGE);

	let p2: u64 = 64u64 + 112u64;
	dwr32(filebuf.ptr, p2 + 0u64,  PT_INTERP_D);
	dwr32(filebuf.ptr, p2 + 4u64,  PF_R_D);
	dwr64(filebuf.ptr, p2 + 8u64,  interpoff);
	dwr64(filebuf.ptr, p2 + 16u64, interpva);
	dwr64(filebuf.ptr, p2 + 24u64, interpva);
	dwr64(filebuf.ptr, p2 + 32u64, interpsz);
	dwr64(filebuf.ptr, p2 + 40u64, interpsz);
	dwr64(filebuf.ptr, p2 + 48u64, 1u64);

	let p3: u64 = 64u64 + 168u64;
	dwr32(filebuf.ptr, p3 + 0u64,  PT_DYNAMIC_D);
	dwr32(filebuf.ptr, p3 + 4u64,  PF_R_D | PF_W_D);
	dwr64(filebuf.ptr, p3 + 8u64,  dynamicoff);
	dwr64(filebuf.ptr, p3 + 16u64, dynamicva);
	dwr64(filebuf.ptr, p3 + 24u64, dynamicva);
	dwr64(filebuf.ptr, p3 + 32u64, dynamicsz);
	dwr64(filebuf.ptr, p3 + 40u64, dynamicsz);
	dwr64(filebuf.ptr, p3 + 48u64, 8u64);

	// Sections.
	dbcopy(filebuf.ptr, interpoff, INTERP.ptr, INTERP.len: u64);
	dwr8(filebuf.ptr, interpoff + (INTERP.len: u64), 0u8);
	dbcopy(filebuf.ptr, dynstroff, dynstr.ptr, dynstrsz);
	dbcopy(filebuf.ptr, dynsymoff, dynsymbuf.ptr, dynsymsz);
	dbcopy(filebuf.ptr, hashoff, hashbuf.ptr, hashsz);
	dbcopy(filebuf.ptr, versymoff, versymbuf.ptr, versymsz);
	if (verneedsz > 0u64) {
		dbcopy(filebuf.ptr, verneedoff, verneedbuf.ptr, verneedsz);
	};
	dbcopy(filebuf.ptr, relapltoff, relapltbuf.ptr, relapltsz);
	if (l.textlen > 0u64) {
		dbcopy(filebuf.ptr, textoff, l.text, l.textlen);
	};
	dbcopy(filebuf.ptr, pltoff, pltbuf.ptr, pltsz);
	dbcopy(filebuf.ptr, gotpltoff, gotpltbuf.ptr, gotpltsz);
	dbcopy(filebuf.ptr, dynamicoff, dynamicbuf.ptr, dynamicsz);
	if (datafilelen > 0u64) {
		dbcopy(filebuf.ptr, dataoff, l.data, datafilelen);
	};

	let wr: (i64 | os.oserror) = os.writeall(fd, filebuf.ptr, filedataend);
	match (wr) {
	case let v: i64 => { if (v != filedataend: i64) { return 1; }; };
	case let e: os.oserror => return 1;
	};
	return 0;
};

// selfhost/cmd/w6l/out.ww — port of cmd/w6l/out.c.
//
// Emit a static ELF64 executable. File layout (per the C original):
//   [0..64)       Ehdr
//   [64..120)     Phdr (one PT_LOAD)
//   [120..0x1000) zero pad
//   [0x1000..)    .text bytes
// Single PT_LOAD covers the whole file, R+X. No interpreter, no .bss.

package w6l;

import os;
import rt;
import sym;
import dynout;

def ET_EXEC:    u16 = 2u16;
def EM_X86_64_W: u16 = 62u16;
def EV_CURRENT: u32 = 1u32;
def ELFCLASS64: u8 = 2u8;
def ELFDATA2LSB: u8 = 1u8;
def PT_LOAD:    u32 = 1u32;
def PF_X:       u32 = 1u32;
def PF_W:       u32 = 2u32;
def PF_R:       u32 = 4u32;

def TEXT_OFF: u64 = 4096u64;	// 0x1000
def PAGE_SZ:  u64 = 4096u64;

// ---- little-endian byte writers ----------------------------------------

fn wru16(buf: *u8, off: u64, v: u16) void = {
	buf[off]         = (v & 255u16): u8;
	buf[off + 1u64]  = ((v >> 8u16) & 255u16): u8;
};

fn wru32(buf: *u8, off: u64, v: u32) void = {
	buf[off]         = (v & 255u32): u8;
	buf[off + 1u64]  = ((v >> 8u32)  & 255u32): u8;
	buf[off + 2u64]  = ((v >> 16u32) & 255u32): u8;
	buf[off + 3u64]  = ((v >> 24u32) & 255u32): u8;
};

fn wru64(buf: *u8, off: u64, v: u64) void = {
	wru32(buf, off,        (v & 4294967295u64): u32);
	wru32(buf, off + 4u64, ((v >> 32u64) & 4294967295u64): u32);
};

// ---- emit ---------------------------------------------------------------

export fn emitelf(l: *lnk, fd: i32, base: u64, entry: u64) i32 = {
	// Dispatch: any loaded shared object plus any dynamic ref means
	// we owe the loader a real PT_INTERP/PT_DYNAMIC binary.
	if (l.sos != nil) {
		if (l.dynn > 0) {
			return emitdynelf(l, fd, base, entry);
		};
	};

	let hasdata: bool = l.datalen > 0u64;
	let rxend:  u64 = TEXT_OFF + l.textlen;
	// .data lands at the next page boundary so the loader can give
	// it fresh R+W permissions without overlapping the R+X mapping.
	let dataoff: u64 = 0u64;
	let datava: u64 = 0u64;
	if (hasdata) {
		dataoff = (rxend + PAGE_SZ - 1u64) & ~(PAGE_SZ - 1u64);
		datava = base + dataoff;
	};

	// Apply relocations now that the layout's textva/datava are
	// known. Deferred from main.ww so the dyn path uses its own
	// datava.
	if (relocate(l, base + TEXT_OFF, datava) != 0) { return -1; };

	// BSS optimisation: trailing zero bytes in .data can be left
	// out of the file. The loader zero-fills the gap between
	// p_filesz and p_memsz. Scan after l_relocate has applied any
	// DATAR patches — anything still zero at the tail genuinely is
	// zero-init. Matches cmd/w6l/out.c byte-for-byte.
	let bsslen: u64 = 0u64;
	if (hasdata) {
		for (bsslen < l.datalen) {
			let b: u8 = l.data[l.datalen - 1u64 - bsslen];
			if (b != 0u8) { break; };
			bsslen += 1u64;
		};
	};
	let datafilelen: u64 = l.datalen - bsslen;

	// One contiguous header buffer covering [0..0x1000), then .text.
	let hdr: []u8 = alloc([], TEXT_OFF)!;
	hdr.len = TEXT_OFF: i32;

	// --- Ehdr (64 bytes) ---
	hdr[0u64] = 127u8;	// 0x7f
	hdr[1u64] = 69u8;	// 'E'
	hdr[2u64] = 76u8;	// 'L'
	hdr[3u64] = 70u8;	// 'F'
	hdr[4u64] = ELFCLASS64;
	hdr[5u64] = ELFDATA2LSB;
	hdr[6u64] = EV_CURRENT: u8;
	wru16(hdr.ptr, 16u64, ET_EXEC);		// e_type
	wru16(hdr.ptr, 18u64, EM_X86_64_W);	// e_machine
	wru32(hdr.ptr, 20u64, EV_CURRENT);		// e_version
	wru64(hdr.ptr, 24u64, entry);		// e_entry
	wru64(hdr.ptr, 32u64, 64u64);		// e_phoff = sizeof(Ehdr)
	wru64(hdr.ptr, 40u64, 0u64);		// e_shoff
	wru32(hdr.ptr, 48u64, 0u32);		// e_flags
	wru16(hdr.ptr, 52u64, 64u16);		// e_ehsize
	wru16(hdr.ptr, 54u64, 56u16);		// e_phentsize
	if (hasdata) { wru16(hdr.ptr, 56u64, 2u16); }
	else { wru16(hdr.ptr, 56u64, 1u16); };
	wru16(hdr.ptr, 58u64, 0u16);		// e_shentsize
	wru16(hdr.ptr, 60u64, 0u16);		// e_shnum
	wru16(hdr.ptr, 62u64, 0u16);		// e_shstrndx

	// --- Phdr #1 (R+X) at offset 64 ---
	wru32(hdr.ptr, 64u64, PT_LOAD);
	wru32(hdr.ptr, 68u64, PF_R | PF_X);
	wru64(hdr.ptr, 72u64, 0u64);		// p_offset
	wru64(hdr.ptr, 80u64, base);		// p_vaddr
	wru64(hdr.ptr, 88u64, base);		// p_paddr
	wru64(hdr.ptr, 96u64, rxend);		// p_filesz
	wru64(hdr.ptr, 104u64, rxend);		// p_memsz
	wru64(hdr.ptr, 112u64, TEXT_OFF);		// p_align

	if (hasdata) {
		// --- Phdr #2 (R+W) at offset 64+56=120 ---
		wru32(hdr.ptr, 120u64, PT_LOAD);
		wru32(hdr.ptr, 124u64, PF_R | PF_W);
		wru64(hdr.ptr, 128u64, dataoff);		// p_offset
		wru64(hdr.ptr, 136u64, base + dataoff);	// p_vaddr
		wru64(hdr.ptr, 144u64, base + dataoff);	// p_paddr
		wru64(hdr.ptr, 152u64, datafilelen);	// p_filesz
		wru64(hdr.ptr, 160u64, l.datalen);		// p_memsz
		wru64(hdr.ptr, 168u64, PAGE_SZ);		// p_align
	};

	// Write [0..0x1000) then .text.
	let r1: (i64 | os.oserror) = os.writeall(fd, hdr.ptr, TEXT_OFF);
	let n1: i64 = 0i64;
	match (r1) {
	case let v: i64 => n1 = v;
	case let e: os.oserror => return -1;
	};
	if (n1 != TEXT_OFF: i64) { return -1; };
	if (l.textlen > 0u64) {
		let r2: (i64 | os.oserror) = os.writeall(fd, l.text, l.textlen);
		let n2: i64 = 0i64;
		match (r2) {
		case let v: i64 => n2 = v;
		case let e: os.oserror => return -1;
		};
		if (n2 != l.textlen: i64) { return -1; };
	};
	if (hasdata && datafilelen > 0u64) {
		// Pad to the page-aligned data offset, then write only
		// the non-zero prefix of .data. The rest is BSS — the
		// loader zero-fills from p_filesz to p_memsz.
		let here: u64 = TEXT_OFF + l.textlen;
		let zero: u8 = 0u8;
		for (here < dataoff) {
			let r3: (i64 | os.oserror) = os.writeall(fd, &zero, 1u64);
			match (r3) {
			case let v: i64 => { };
			case let e: os.oserror => return -1;
			};
			here += 1u64;
		};
		let r4: (i64 | os.oserror) = os.writeall(fd, l.data, datafilelen);
		let n4: i64 = 0i64;
		match (r4) {
		case let v: i64 => n4 = v;
		case let e: os.oserror => return -1;
		};
		if (n4 != datafilelen: i64) { return -1; };
	};
	return 0;
};

// selfhost/cmd/w6l/main.ww — port of cmd/w6l/main.c.
//
// w6l = amd64 linker. Reads relocatable ELF .o files, SysV `ar`
// archives, and shared objects (ET_DYN). Resolves symbols, applies
// relocations, writes a static or dynamic-linked ELF executable.
//
//   w6l_ww -o out [-L<dir>...] [-l<name>...] file1.o file2.o ...

package main;

import os;
import rt;
import sym;
import obj;
import dyn;
import pass;
import out;

def BASE: u64 = 4194304u64;	// 0x400000
def CODE_VA_OFF: u64 = 4096u64;	// .text starts at base + 0x1000

fn mklnk() *lnk = {
	let l: *lnk = alloc(lnk { })!;
	return l;
};

// `cstreq` lives in obj.ww — same bundle, single definition.

// `cstrlen` lives in obj.ww — same bundle, single definition.

// Build "<dir>/lib<name>.<ext>" into dst (NUL-terminated). Returns total
// length excluding NUL. dst must be large enough.
fn buildpath(dst: *u8, dir: *u8, name: *u8, ext: str) u64 = {
	let i: u64 = 0u64;
	let dn: u64 = cstrlen(dir);
	let nn: u64 = cstrlen(name);
	let k: u64 = 0u64;
	for (k < dn) { dst[i] = dir[k]; i += 1u64; k += 1u64; };
	dst[i] = 47u8;	// '/'
	i += 1u64;
	dst[i] = 108u8;	// 'l'
	i += 1u64;
	dst[i] = 105u8;	// 'i'
	i += 1u64;
	dst[i] = 98u8;	// 'b'
	i += 1u64;
	k = 0u64;
	for (k < nn) { dst[i] = name[k]; i += 1u64; k += 1u64; };
	k = 0u64;
	for (k < ext.len: u64) {
		let li: i32 = k: i32;
		dst[i] = ext[li];
		i += 1u64;
		k += 1u64;
	};
	dst[i] = 0u8;
	return i;
};

// Append decimal n to dst at offset i. Returns new offset.
fn appenddec(dst: *u8, i: u64, n: u64) u64 = {
	if (n == 0u64) {
		dst[i] = 48u8;
		return i + 1u64;
	};
	let buf: [16]u8;
	let k: u64 = 0u64;
	let v: u64 = n;
	for (v > 0u64) {
		buf[k] = (v % 10u64): u8 + 48u8;
		v = v / 10u64;
		k += 1u64;
	};
	let oi: u64 = i;
	for (k > 0u64) {
		k -= 1u64;
		dst[oi] = buf[k];
		oi += 1u64;
	};
	return oi;
};

fn buildpathv(dst: *u8, dir: *u8, name: *u8, v: u64) u64 = {
	let i: u64 = 0u64;
	let dn: u64 = cstrlen(dir);
	let nn: u64 = cstrlen(name);
	let k: u64 = 0u64;
	for (k < dn) { dst[i] = dir[k]; i += 1u64; k += 1u64; };
	dst[i] = 47u8; i += 1u64;
	dst[i] = 108u8; i += 1u64;
	dst[i] = 105u8; i += 1u64;
	dst[i] = 98u8; i += 1u64;
	k = 0u64;
	for (k < nn) { dst[i] = name[k]; i += 1u64; k += 1u64; };
	dst[i] = 46u8; i += 1u64; dst[i] = 115u8; i += 1u64; dst[i] = 111u8; i += 1u64; dst[i] = 46u8; i += 1u64;
	i = appenddec(dst, i, v);
	dst[i] = 0u8;
	return i;
};

// islinkable: read first 8 bytes; require !<arch>\n or \x7fELF.
fn islinkable(path: *u8) bool = {
	let fd: i32 = os.open(pathstr(path), os.flag.RDONLY, 0i32);
	if (fd < 0) { return false; };
	let mp: [8]u8;
	let n: i64 = os.read(fd, &mp[0], 8u64);
	os.close(fd);
	if (n < 4i64) { return false; };
	// archive: "!<arch>\n"
	if (n >= 8i64) {
		if (mp[0u64] == 33u8) { if (mp[1u64] == 60u8) {
		if (mp[2u64] == 97u8) { if (mp[3u64] == 114u8) {
		if (mp[4u64] == 99u8) { if (mp[5u64] == 104u8) {
		if (mp[6u64] == 62u8) { if (mp[7u64] == 10u8) {
			return true;
		}; }; }; }; }; }; }; };
	};
	// ELF: "\x7fELF"
	if (mp[0u64] == 127u8) {
		if (mp[1u64] == 69u8) {
			if (mp[2u64] == 76u8) {
				if (mp[3u64] == 70u8) { return true; };
			};
		};
	};
	return false;
};

// Walk libdirs[0..n) trying lib<name>.so, then lib<name>.so.{0..8},
// then lib<name>.a. Return a heap-allocated NUL-terminated path on
// success, nil on miss.
fn resolvelib(name: *u8, libdirs: **u8, nlibdirs: i32) *u8 = {
	let bufp: []u8 = alloc([], 1024u64)!;
	bufp.len = 1024;
	let i: i32 = 0;
	for (i < nlibdirs) {
		let dir: *u8 = libdirs[i];
		let _l1: u64 = buildpath(bufp.ptr, dir, name, ".so");
		if (islinkable(bufp.ptr)) {
			let pl: u64 = cstrlen(bufp.ptr);
			let p: []u8 = alloc([], pl + 1u64)!;
			let k: u64 = 0u64;
			for (k <= pl) { p[k] = bufp[k]; k += 1u64; };
			return p.ptr;
		};
		let v: u64 = 0u64;
		for (v <= 8u64) {
			let _l2: u64 = buildpathv(bufp.ptr, dir, name, v);
			if (islinkable(bufp.ptr)) {
				let pl2: u64 = cstrlen(bufp.ptr);
				let p2: []u8 = alloc([], pl2 + 1u64)!;
				let k2: u64 = 0u64;
				for (k2 <= pl2) { p2[k2] = bufp[k2]; k2 += 1u64; };
				return p2.ptr;
			};
			v += 1u64;
		};
		let _l3: u64 = buildpath(bufp.ptr, dir, name, ".a");
		if (islinkable(bufp.ptr)) {
			let pl3: u64 = cstrlen(bufp.ptr);
			let p3: []u8 = alloc([], pl3 + 1u64)!;
			let k3: u64 = 0u64;
			for (k3 <= pl3) { p3[k3] = bufp[k3]; k3 += 1u64; };
			return p3.ptr;
		};
		i += 1;
	};
	return nil;
};

// Read first 20 bytes; return 1 for ET_DYN .so, 0 for ar/.o.
fn isso(path: *u8) i32 = {
	let fd: i32 = os.open(pathstr(path), os.flag.RDONLY, 0i32);
	if (fd < 0) { return 0; };
	let mp: [20]u8;
	let n: i64 = os.read(fd, &mp[0], 20u64);
	os.close(fd);
	if (n < 20i64) { return 0; };
	if (mp[0u64] != 127u8) { return 0; };
	if (mp[1u64] != 69u8) { return 0; };
	if (mp[2u64] != 76u8) { return 0; };
	if (mp[3u64] != 70u8) { return 0; };
	// e_type at offset 16, u16 little-endian
	let t: u16 = (mp[16u64]: u16) | ((mp[17u64]: u16) << 8u16);
	if (t == 3u16) { return 1; };
	return 0;
};

export fn main(argc: i32, argv: **u8) i32 = {
	let outpath: *u8 = nil;
	let maxinputs: i32 = 64;
	let inputs:  []*u8 = alloc([], maxinputs: u64)!;
	inputs.len = maxinputs;
	let ninputs: i32 = 0;
	let libdirs: []*u8 = alloc([], maxinputs: u64)!;
	libdirs.len = maxinputs;
	let nlibdirs: i32 = 0;
	let lflags:  []*u8 = alloc([], maxinputs: u64)!;
	lflags.len = maxinputs;
	let nlflags: i32 = 0;

	let i: i32 = 1;
	for (i < argc) {
		let a: *u8 = argv[i];
		if (cstreq(a, "-o")) {
			i += 1;
			if (i >= argc) {
				os.write(2, "w6l: -o requires argument\n".ptr, 25u64);
				return 2;
			};
			outpath = argv[i];
		} else { if (cstreq(a, "-L")) {
			i += 1;
			if (i >= argc) {
				os.write(2, "w6l: -L requires argument\n".ptr, 25u64);
				return 2;
			};
			libdirs[nlibdirs] = argv[i];
			nlibdirs += 1;
		} else { if (cstreq(a, "-l")) {
			i += 1;
			if (i >= argc) {
				os.write(2, "w6l: -l requires argument\n".ptr, 25u64);
				return 2;
			};
			lflags[nlflags] = argv[i];
			nlflags += 1;
		} else { if (a[0u64] == 45u8) {
			// -L<dir> joined form.
			if (a[1u64] == 76u8) {
				if (a[2u64] != 0u8) {
					libdirs[nlibdirs] = a + 2u64;
					nlibdirs += 1;
				} else {
					os.write(2, "w6l: bare -L\n".ptr, 12u64);
					return 2;
				};
			} else { if (a[1u64] == 108u8) {
				if (a[2u64] != 0u8) {
					lflags[nlflags] = a + 2u64;
					nlflags += 1;
				} else {
					os.write(2, "w6l: bare -l\n".ptr, 12u64);
					return 2;
				};
			} else {
				os.write(2, "w6l: unknown flag\n".ptr, 17u64);
				return 2;
			};};
		} else {
			if (ninputs >= maxinputs) {
				os.write(2, "w6l: too many inputs\n".ptr, 20u64);
				return 2;
			};
			inputs[ninputs] = a;
			ninputs += 1;
		};};};};
		i += 1;
	};

	if (outpath == nil) {
		os.write(2, "usage: w6l_ww -o exe [-L<dir>...] [-l<name>...] file1.o [file2.o...]\n".ptr, 68u64);
		return 2;
	};
	if (ninputs == 0) {
		os.write(2, "w6l: no inputs\n".ptr, 14u64);
		return 2;
	};

	let l: *lnk = mklnk();

	// Seed _start so libwwrt-style start.o is recognised as wanted.
	intern(l, "_start");

	// Load positional inputs first (preserving order).
	let k: i32 = 0;
	for (k < ninputs) {
		if (load(l, inputs[k]) != 0) {
			return 1;
		};
		k += 1;
	};

	// Then resolve -l flags and load each. Archives append; shared
	// objects register their exports.
	let lf: i32 = 0;
	for (lf < nlflags) {
		let p: *u8 = resolvelib(lflags[lf], libdirs.ptr, nlibdirs);
		if (p == nil) {
			os.write(2, "w6l: cannot find -l".ptr, 18u64);
			let nm: *u8 = lflags[lf];
			os.write(2, nm, cstrlen(nm));
			os.write(2, "\n".ptr, 1u64);
			return 1;
		};
		if (isso(p) != 0) {
			if (loadso(l, p) != 0) { return 1; };
		} else {
			if (load(l, p) != 0) { return 1; };
		};
		lf += 1;
	};

	if (resolve(l) != 0) { return 1; };
	// Relocation is deferred to the emit functions — each path
	// knows its own layout (textva, datava); the static and dyn
	// paths place .data at different VAs.

	let entrysym: *lsym = lookup(l, "_start");
	if (entrysym == nil) { entrysym = lookup(l, "main"); }
	else { if (entrysym.defined == 0) { entrysym = lookup(l, "main"); }; };
	if (entrysym == nil) {
		os.write(2, "w6l: no _start or main symbol\n".ptr, 29u64);
		return 1;
	};
	if (entrysym.defined == 0) {
		os.write(2, "w6l: no _start or main symbol\n".ptr, 29u64);
		return 1;
	};

	let flags: os.flag = os.flag.WRONLY | os.flag.CREATE | os.flag.TRUNC;
	let fd: i32 = os.open(pathstr(outpath), flags, 493i32);	// 0o755
	if (fd < 0) {
		os.write(2, "w6l: cannot open output\n".ptr, 23u64);
		return 1;
	};

	let entryva: u64 = BASE + CODE_VA_OFF + entrysym.val;
	let rc: i32 = emitelf(l, fd, BASE, entryva);
	os.close(fd);
	return rc;
};