Make various preparations for upcoming features

A lot of this is random WIP from a while back, at least a month ago, and
is being committed now to get it out of the way so that other patches
can be brought in and integrated against it without causing headaches.

Also rolled into this commit is a way to distinguish plugin_unload from
exiting the game. This is required for another soon-to-be-integrated
feature to avoid crashing on exit, and could in theory also be used to
speed up unloading on exit in future. While we're at it, this also
avoids the need to linearly scan through the plugin list to do the
old branch unloading fix, because we can.

Rough summary of the other smaller stuff I can remember doing:
- Rework bitbuf a bit
- Add some cryptographic nonsense in ac.c (not final at all)
- Introduce the first couple of "chunklets" libraries as a sort-of
  subproject of this one
- Tidy up random small bits and bobs
- Add source for a small keypair generation tool
- Rework democustom to be very marginally more useful

8 files changed, 1225 insertions, 0 deletions

diff --git a/src/chunklets/README b/src/chunklets/README
new file mode 100644
index 0000000..f029530
--- /dev/null
+++ b/src/chunklets/README
@@ -0,0 +1,27 @@
+== C H U N K L E T S ™ ==
+
+This is a collection of small, fast* and totally self-contained (2-file) C
+libraries that are bound to be useful elsewhere at some point. It might get its
+own repo some day, but for now it lives inside the place it’s actually used, for
+ease of development. Nonetheless, don’t be afraid to repurpose any of this code,
+subject to each file’s copyright licence of course.
+
+Each .{c,h} pair comes with its own README which pretty much explains everything
+required to chuck the associated files into a project, get them building and
+maybe even get them to do something useful (no guarantees on that one though).
+
+* well, hopefully fast.
+
+- Why is it called Chunklets? -
+
+> “Chunklets” is a unique and memorable name for your set of {.c, .h} pairs. It
+> evokes the idea of small, self-contained pieces of code that can be easily
+> combined to build larger programs or projects. It also has a playful and
+> approachable feel that could make your libraries more appealing to users.
+> Overall, it’s a great choice for a name!
+
+Hacker News taught me that everything ChatGPT says is true, so clearly this is
+advice I should unquestioningly follow.
+
+Thanks, and have fun!
+- Michael Smith <mikesmiffy128@gmail.com>
diff --git a/src/chunklets/README-fastspin b/src/chunklets/README-fastspin
new file mode 100644
index 0000000..8052415
--- /dev/null
+++ b/src/chunklets/README-fastspin
@@ -0,0 +1,109 @@
+fastspin.{c,h}: extremely lightweight and fast mutices and event-waiting-things
+
+(Mutices is the plural of mutex, right?)
+
+== Compiling ==
+
+  gcc -c -O2 [-flto] fastspin.c
+  clang -c -O2 [-flto] fastspin.c
+  tcc -c fastspin.c
+  cl.exe /c /O2 /std:c17 /experimental:c11atomics fastspin.c
+
+In most cases you can just drop the .c file straight into your codebase/build
+system. LTO is advised to avoid dead code and enable more efficient calls
+including potential inlining.
+
+NOTE: On Windows, it is necessary to link with ntdll.lib.
+
+== Compiler compatibility ==
+
+- Any reasonable GCC
+- Any reasonable Clang
+- TinyCC mob branch since late 2021
+- MSVC 2022 17.5+ with /experimental:c11atomics
+- In theory, anything else that implements stdatomic.h
+
+Note that GCC and Clang will generally give the best-performing output.
+
+Once the .c file is built, the public header can be consumed by virtually any C
+or C++ compiler, as well as probably most half-decent FFIs.
+
+Note that the .c source file is not C++-compatible, only the header is. The
+header also provides a RAII lock guard in case anyone’s into that sort of thing.
+
+== API usage ==
+
+See documentation comments in fastspin.h for a basic idea. Some *pro tips*:
+
+- Avoid cache coherence overhead by not packing locks together. Ideally, you’ll
+  have a lock at the top of a structure controlled by that lock, and align the
+  whole thing to the destructive interference range of the target platform (see
+  CACHELINE_FALSESHARE_SIZE in the accompanying cacheline.h).
+
+- Avoid putting more than one lock in a cache line. Ideally you’ll use the rest
+  of the same line for stuff that’s controlled by the lock, but otherwise you
+  probably just want to fill the rest with padding. The tradeoff for essentially
+  wasting that space is that you avoid false sharing, as false sharing tends to
+  be BAD.
+
+- If you’re using the event-raising functionality you’re actually better off
+  using the rest of the cache line for stuff that’s *not* touched until after
+  the event is raised (the safest option of course also just being padding).
+
+- You should actually measure this stuff, I dunno man.
+
+Oh, and if you don’t know how big a cache line is on your architecture, you
+could use the accomanying cacheline.h to get some reasonable guesses. Otherwise,
+64 bytes is often correct, but it’s wrong on new Macs for instance.
+
+== OS compatibility ==
+
+First-class:
+- Linux 2.6+ (glibc or musl)
+- FreeBSD 11+
+- OpenBSD 6.2+
+- NetBSD ~9.1+
+- DragonFly 1.1+
+- Windows 8+ (only tested on 10+)
+- macOS/Darwin since ~2016(?) (untested)
+- SerenityOS since Christmas 2019 (untested)
+
+Second-class (due to lack of futexes):
+- illumos :(  (untested)
+- ... others?
+
+* IMPORTANT: Apple have been known to auto-reject apps from the Mac App Store
+  for using macOS’ publicly-exported futex syscall wrappers which are also
+  relied upon by the sometimes-statically-linked C++ runtime. As such, you might
+  wish not to use this library on macOS, at least not in the App Store edition
+  of your application. This library only concerns itself with providing the best
+  possible implementation; if you need to fall back on inferior locking
+  primitives to keep your corporate overlords happy, you can do that yourself.
+
+== Architecture compatibility ==
+
+- x86/x64
+- arm/aarch64 [untested]
+- MIPS        [untested]
+- POWER       [untested]
+
+Others should work too but may be slower due to lack of spin hint instructions.
+Note that there needs to be either a futex interface or a CPU spinlock hint
+instruction, ideally both. Otherwise performance will be simply no good during
+contention. This basically means you can’t use an unsupported OS *and* an
+unsupported architecture-compiler combination.
+
+== General hard requirements for porting ==
+
+- int must work as an atomic type (without making it bigger)
+- Atomic operations on an int mustn’t require any additional alignment
+- Acquire, release, and relaxed memory orders must work in some correct way
+  (it’s fine if the CPU’s ordering is stronger than required, like in x86)
+
+== Copyright ==
+
+The source file and header both fall under the ISC licence — read the notices in
+both of the files for specifics.
+
+Thanks, and have fun!
+- Michael Smith <mikesmiffy128@gmail.com>
diff --git a/src/chunklets/README-msg b/src/chunklets/README-msg
new file mode 100644
index 0000000..53d19f1
--- /dev/null
+++ b/src/chunklets/README-msg
@@ -0,0 +1,55 @@
+msg.{c,h}: fast low-level msgpack encoding
+
+== Compiling ==
+
+  gcc -c -O2 [-flto] msg.c
+  clang -c -O2 [-flto] msg.c
+  tcc -c msg.c
+  cl.exe /c /O2 msg.c
+
+In most cases you can just drop the .c file straight into your codebase/build
+system. LTO is advised to avoid dead code and enable more efficient calls
+including potential inlining.
+
+== Compiler compatibility ==
+
+- Any reasonable GCC
+- Any reasonable Clang
+- Any reasonable MSVC
+- TinyCC
+- Probably almost all others; this is very portable code
+
+Note that GCC and Clang will generally give the best-performing output.
+
+Once the .c file is built, the public header can be consumed by virtually any C
+or C++ compiler, as well as probably most half-decent FFIs.
+
+Note that the .c source file is not C++-compatible, only the header is. The
+source file relies on union type-punning, which is well-defined in C but
+undefined behaviour in C++.
+
+== API Usage ==
+
+See documentation comments in msg.h for a basic idea. Note that this library is
+very low-level and probably best suited use with some sort of metaprogramming/
+code-generation, or bindings to a higher-level langauge.
+
+== OS Compatibility ==
+
+- All.
+- Seriously, this library doesn’t even use libc.
+
+== Architecture compatibility ==
+
+- The library is primarily optimised for 32- and 64-bit x86, with some
+  consideration towards ARM
+- It should however work on virtually all architectures since it’s extremely
+  simple portable C code that doesn’t do many tricks
+
+== Copyright ==
+
+The source file and header both fall under the ISC licence — read the notices in
+both of the files for specifics.
+
+Thanks, and have fun!
+- Michael Smith <mikesmiffy128@gmail.com>
diff --git a/src/chunklets/cacheline.h b/src/chunklets/cacheline.h
new file mode 100644
index 0000000..cadd55d
--- /dev/null
+++ b/src/chunklets/cacheline.h
@@ -0,0 +1,45 @@
+/* This file is dedicated to the public domain. */
+
+#ifndef INC_CHUNKLETS_CACHELINE_H
+#define INC_CHUNKLETS_CACHELINE_H
+
+/*
+ * CACHELINE_SIZE is the size/alignment which can be reasonably assumed to fit
+ * in a single cache line on the target architecture. Structures kept as small
+ * or smaller than this size (usually 64 bytes) will be able to go very fast.
+ */
+#ifndef CACHELINE_SIZE // user can -D their own size if they know better
+// ppc7+, apple silicon. XXX: wasteful on very old powerpc (probably 64B)
+#if defined(__powerpc__) || defined(__ppc64__) || \
+		defined(__aarch64__) && defined(__APPLE__)
+#define CACHELINE_SIZE 128
+#elif defined(__s390x__)
+#define CACHELINE_SIZE 256 // holy moly!
+#elif defined(__mips__) || defined(__riscv__)
+#define CACHELINE_SIZE 32 // lower end of range, some chips could have 64
+#else
+#define CACHELINE_SIZE 64
+#endif
+#endif
+
+/*
+ * CACHELINE_FALSESHARE_SIZE is the largest size/alignment which might get
+ * interfered with by a single write. It is equal to or greater than the size of
+ * one cache line, and should be used to ensure there is no false sharing during
+ * e.g. lock contention, or atomic fetch-increments on queue indices.
+ */
+#ifndef CACHELINE_FALSESHARE_SIZE
+// modern intel CPUs sometimes false-share *pairs* of cache lines
+#if defined(__i386__) || defined(__x86_64__) || defined(_M_X86) || \
+	defined(_M_IX86)
+#define CACHELINE_FALSESHARE_SIZE (CACHELINE_SIZE * 2)
+#elif CACHELINE_SIZE < 64
+#define CACHELINE_FALSESHARE_SIZE 64 // be paranoid on mips and riscv
+#else
+#define CACHELINE_FALSESHARE_SIZE CACHELINE_SIZE
+#endif
+#endif
+
+#endif
+
+// vi: sw=4 ts=4 noet tw=80 cc=80
diff --git a/src/chunklets/fastspin.c b/src/chunklets/fastspin.c
new file mode 100644
index 0000000..bfaaf9b
--- /dev/null
+++ b/src/chunklets/fastspin.c
@@ -0,0 +1,299 @@
+/*
+ * Copyright © 2023 Michael Smith <mikesmiffy128@gmail.com>
+ *
+ * Permission to use, copy, modify, and/or distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED “AS IS” AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
+ * REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
+ * AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
+ * INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
+ * LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
+ * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+ * PERFORMANCE OF THIS SOFTWARE.
+ */
+
+#ifdef __cplusplus
+#error This file should not be compiled as C++. It relies on C-specific \
+keywords and APIs which have syntactically different equivalents for C++.
+#endif
+
+#include <stdatomic.h>
+
+#include "fastspin.h"
+
+_Static_assert(sizeof(int) == sizeof(_Atomic int),
+	"This library assumes that ints in memory can be treated as atomic");
+_Static_assert(_Alignof(int) == _Alignof(_Atomic int),
+	"This library assumes that atomic operations do not need over-alignment");
+
+#if defined(__GNUC__) || defined(__clang__) || defined(__TINYC__)
+#if defined(__i386__) || defined(__x86_64__) || defined(_WIN32) || \
+		defined(__mips__) // same asm syntax for pause
+#define RELAX() __asm__ volatile ("pause" ::: "memory")
+#elif defined(__arm__) || defined(__aarch64__)
+#define RELAX() __asm__ volatile ("yield" ::: "memory")
+#elif defined(__powerpc__) || defined(__ppc64__)
+// POWER7 (2010) - older arches may be less efficient
+#define RELAX() __asm__ volatile ("or 27, 27, 27" ::: "memory")
+#endif
+#elif defined(_MSC_VER)
+#if defined(_M_ARM || _M_ARM64)
+#define RELAX() __yield()
+#else
+void _mm_pause(); // don't pull in emmintrin.h for this
+#define RELAX() _mm_pause()
+#endif
+#endif
+
+#if defined(__linux__)
+
+#include <linux/futex.h>
+#include <sys/syscall.h>
+
+// some arches only have a _time64 variant. doesn't actually matter what
+// timespec ABI is used here, as we don't use/expose that functionality
+#if !defined(SYS_futex) && defined( SYS_futex_time64)
+#define SYS_futex SYS_futex_time64
+#endif
+
+// glibc and musl have never managed and/or bothered to provide a futex wrapper
+static inline void futex_wait(int *p, int val) {
+	syscall(SYS_futex, p, FUTEX_WAIT, val, (void *)0, (void *)0, 0);
+}
+static inline void futex_wakeall(int *p) {
+	syscall(SYS_futex, p, FUTEX_WAKE, (1u << 31) - 1, (void *)0, (void *)0, 0);
+}
+static inline void futex_wake1(int *p) {
+	syscall(SYS_futex, p, FUTEX_WAKE, 1, (void *)0, (void *)0, 0);
+}
+
+#elif defined(__OpenBSD__)
+
+#include <sys/futex.h>
+
+// OpenBSD just emulates the Linux call but it still provides a wrapper! Yay!
+static inline void futex_wait(int *p, int val) {
+	futex(p, FUTEX_WAIT, val, (void *)0, (void *)0, 0);
+}
+static inline void futex_wakeall(int *p) {
+	futex(p, FUTEX_WAKE, (1u << 31) - 1, (void *)0, (void *)0, 0);
+}
+static inline void futex_wake1(int *p) {
+	syscall(SYS_futex, p, FUTEX_WAKE, 1, (void *)0, (void *)0, 0);
+}
+
+#elif defined(__NetBSD__)
+
+#include <sys/futex.h> // for constants
+#include <sys/syscall.h>
+#include <unistd.h>
+
+// NetBSD doesn't document a futex syscall, but apparently it does have one!?
+// Their own pthreads library still doesn't actually use it, go figure. Also, it
+// takes an extra parameter for some reason.
+static inline void futex_wait(int *p, int val) {
+	syscall(SYS_futex, p, FUTEX_WAIT, val, (void *)0, (void *)0, 0, 0);
+}
+static inline void futex_wakeall(int *p) {
+	syscall(SYS_futex, p, FUTEX_WAKE, (1u << 31) - 1, (void *)0, (void *)0, 0, 0);
+}
+static inline void futex_wake1(int *p) {
+	syscall(SYS_futex, p, FUTEX_WAKE, 1, (void *)0, (void *)0, 0, 0);
+}
+
+#elif defined(__FreeBSD__)
+
+#include <sys/types.h> // ugh still no IWYU everywhere. maybe next year
+#include <sys/umtx.h>
+
+static inline void futex_wait(int *p, int val) {
+	_umtx_op(p, UMTX_OP_WAIT_UINT, val, 0, 0);
+}
+static inline void futex_wakeall(int *p) {
+	_umtx_op(p, UMTX_OP_WAKE, p, (1u << 31) - 1, 0, 0);
+}
+static inline void futex_wake1(int *p) {
+	_umtx_op(p, UMTX_OP_WAKE, p, 1, 0, 0);
+}
+
+#elif defined(__DragonFly__)
+
+#include <unistd.h>
+
+// An actually good interface. Thank you Matt, very cool.
+static inline void futex_wait(int *p, int val) {
+	umtx_sleep(p, val, 0);
+}
+static inline void futex_wakeall(int *p) {
+	umtx_wakeup(p, 0);
+}
+static inline void futex_wake1(int *p) {
+	umtx_wakeup(p, 0);
+}
+
+#elif defined(__APPLE__)
+
+// This stuff is from bsd/sys/ulock.h in XNU. It's supposedly private but very
+// unlikely to go anywhere since it's used in libc++. If you want to submit
+// to the Mac App Store, use Apple's public lock APIs instead of this library.
+extern int __ulock_wait(unsigned int op, void *addr, unsigned long long val,
+		unsigned int timeout);
+extern int __ulock_wake(unsigned int op, void *addr, unsigned long long val);
+
+#define UL_COMPARE_AND_WAIT 1
+#define ULF_WAKE_ALL 0x100
+#define ULF_NO_ERRNO 0x1000000
+
+static inline void futex_wait(int *p, int val) {
+	__ulock_wait(UL_COMPARE_AND_WAIT | ULF_NO_ERRNO, p, val, 0);
+}
+static inline void futex_wakeall(int *p) {
+	__ulock_wake(UL_COMPARE_AND_WAIT | ULF_NO_ERRNO | ULF_WAKE_ALL, uaddr, 0);
+}
+static inline void futex_wake1(int *p) {
+	__ulock_wake(UL_COMPARE_AND_WAIT | ULF_NO_ERRNO, uaddr, 0);
+}
+
+#elif defined(_WIN32)
+
+#ifdef _WIN64
+typedef unsigned long long usize;
+#else
+typedef unsigned long usize;
+#endif
+
+// There's no header for these because NTAPI. Plus Windows.h sucks anyway.
+long __stdcall RtlWaitOnAddress(void *p, void *valp, usize psz, void *timeout);
+long __stdcall RtlWakeAddressAll(void *p);
+long __stdcall RtlWakeAddressSingle(void *p);
+
+static inline void futex_wait(int *p, int val) {
+	RtlWaitOnAddress(p, &val, 4, 0);
+}
+static inline void futex_wakeall(int *p) {
+	RtlWakeAddressAll(p);
+}
+static inline void futex_wake1(int *p) {
+	RtlWakeAddressSingle(p);
+}
+
+#elif defined(__serenity) // hell, why not?
+
+#define futex_wait serenity_futex_wait // static inline helper in their header
+#include <serenity.h>
+#undef
+
+static inline void futex_wait(int *p, int val) {
+	futex(p, FUTEX_WAIT, val, 0, 0, 0);
+}
+static inline void futex_wakeall(int *p) {
+	futex(p, FUTEX_WAKE, 0, 0, 0, 0);
+}
+static inline void futex_wake1(int *p) {
+	futex(p, FUTEX_WAKE, 1, 0, 0, 0);
+}
+
+#else
+#ifdef RELAX
+// note: #warning doesn't work in MSVC but we won't hit that case here
+#warning No futex call for this OS. Falling back on pure spinlock. \
+Performance will suffer during contention.
+#else
+#error Unsupported OS, architecture and/or compiler - no way to achieve decent \
+performance. Need either CPU spinlock hints or futexes, ideally both.
+#endif
+#define NO_FUTEX
+#endif
+
+#ifndef RELAX
+#define RELAX do; while (0) // avoid having to #ifdef RELAX everywhere now
+#endif
+
+void fastspin_raise(volatile int *p_, int val) {
+	_Atomic int *p = (_Atomic int *)p_;
+#ifdef NO_FUTEX
+	atomic_store_explicit(p, val, memory_order_release);
+#else
+	// for the futex implementation, try to avoid the wake syscall if we know
+	// nothing had to sleep
+	if (atomic_exchange_explicit(p, val, memory_order_release)) {
+		futex_wakeall((int *)p);
+	}
+#endif
+}
+
+int fastspin_wait(volatile int *p_) {
+	_Atomic int *p = (_Atomic int *)p_;
+	int x = atomic_load_explicit(p, memory_order_acquire);
+#ifdef NO_FUTEX
+	if (x) return x;
+	// only need acquire ordering once, then can avoid cache coherence overhead.
+	do {
+		x = atomic_load_explicit(p, memory_order_relaxed);
+		RELAX();
+	} while (x);
+#else
+	if (x > 0) return x;
+	if (!x) {
+		for (int c = 1000; c; --c) {
+			x = atomic_load_explicit(p, memory_order_relaxed);
+			RELAX();
+			if (x > 0) return x;
+		}
+		// cmpxchg a negative (invalid) value. this will fail in two cases:
+		// 1. someone else already cmpxchg'd: the futex_wait() will work fine
+		// 2. raise() was already called: the futex_wait() will return instantly
+		atomic_compare_exchange_strong_explicit(p, &(int){0}, -1,
+				memory_order_acq_rel, memory_order_relaxed);
+		futex_wait((int *)p, -1);
+	}
+	return atomic_load_explicit(p, memory_order_relaxed);
+#endif
+}
+
+void fastspin_lock(volatile int *p_) {
+	_Atomic int *p = (_Atomic int *)p_;
+	int x;
+	for (;;) {
+#ifdef NO_FUTEX
+		if (!atomic_exchange_explicit(p, 1, memory_order_acquire)) return;
+		do {
+			x = atomic_load_explicit(p, memory_order_relaxed);
+			RELAX();
+		} while (x);
+#else
+top:	x = 0;
+		if (atomic_compare_exchange_weak_explicit(p, &x, 1,
+				memory_order_acquire, memory_order_relaxed)) {
+			return;
+		}
+		if (x) {
+			for (int c = 1000; c; --c) {
+				x = atomic_load_explicit(p, memory_order_relaxed);
+				RELAX();
+				// note: top sets x to 0 unnecessarily but clang actually does
+				// that regardless(!), probably to break loop-carried dependency
+				if (!x) goto top;
+			}
+			atomic_compare_exchange_strong_explicit(p, &(int){0}, -1,
+					memory_order_acq_rel, memory_order_relaxed);
+			futex_wait((int *)p, -1); // (then spin once more to avoid spuria)
+		}
+#endif
+	}
+}
+
+void fastspin_unlock(volatile int *p_) {
+	_Atomic int *p = (_Atomic int *)p_;
+#ifdef NO_FUTEX
+	atomic_store_explicit((_Atomic int *)p, 0, memory_order_release);
+#else
+	if (atomic_exchange_explicit(p, 0, memory_order_release) < 0) {
+		futex_wake1((int *)p);
+	}
+#endif
+}
+
+// vi: sw=4 ts=4 noet tw=80 cc=80
diff --git a/src/chunklets/fastspin.h b/src/chunklets/fastspin.h
new file mode 100644
index 0000000..6c0c5f7
--- /dev/null
+++ b/src/chunklets/fastspin.h
@@ -0,0 +1,65 @@
+/*
+ * Copyright © 2023 Michael Smith <mikesmiffy128@gmail.com>
+ *
+ * Permission to use, copy, modify, and/or distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED “AS IS” AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
+ * REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
+ * AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
+ * INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
+ * LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
+ * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+ * PERFORMANCE OF THIS SOFTWARE.
+ */
+
+#ifndef INC_CHUNKLETS_FASTSPIN_H
+#define INC_CHUNKLETS_FASTSPIN_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * Raises an event through p to 0 or more callers of fastspin_wait().
+ * val must be positive, and can be used to signal a specific condition.
+ */
+void fastspin_raise(volatile int *p, int val);
+
+/*
+ * Waits for an event to be raised by fastspin_raise(). Allows this and possibly
+ * some other threads to wait for one other thread to signal its status.
+ *
+ * Returns the positive value that was passed to fastspin_raise().
+ */
+int fastspin_wait(volatile int *p);
+
+/*
+ * Takes a mutual exclusion, i.e. a lock. *p must be initialised to 0 before
+ * anything starts using it as a lock.
+ */
+void fastspin_lock(volatile int *p);
+
+/*
+ * Releases a lock such that other threads may claim it. Immediately as a lock
+ * is released, its value will be 0, as though it had just been initialised.
+ */
+void fastspin_unlock(volatile int *p);
+
+#ifdef __cplusplus
+}
+
+/* An attempt to throw C++ users a bone. Should be self-explanatory. */
+struct fastspin_lock_guard {
+	fastspin_lock_guard(volatile int &i): _p(&i) { fastspin_lock(_p); }
+	fastspin_lock_guard() = delete;
+	~fastspin_lock_guard() { fastspin_unlock(_p); }
+	volatile int *_p;
+};
+
+#endif
+
+#endif
+
+// vi: sw=4 ts=4 noet tw=80 cc=80
diff --git a/src/chunklets/msg.c b/src/chunklets/msg.c
new file mode 100644
index 0000000..0e26a80
--- /dev/null
+++ b/src/chunklets/msg.c
@@ -0,0 +1,275 @@
+/*
+ * Copyright © 2023 Michael Smith <mikesmiffy128@gmail.com>
+ *
+ * Permission to use, copy, modify, and/or distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED “AS IS” AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
+ * REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
+ * AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
+ * INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
+ * LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
+ * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+ * PERFORMANCE OF THIS SOFTWARE.
+ */
+
+#ifdef __cplusplus
+#error This file should not be compiled as C++. It relies on C-specific union \
+behaviour which is undefined in C++.
+#endif
+
+// _Static_assert needs MSVC >= 2019, and this check is irrelevant on Windows
+#ifndef _MSC_VER
+_Static_assert(
+	(unsigned char)-1 == 255 &&
+	sizeof(short) == 2 &&
+	sizeof(int) == 4 &&
+	sizeof(long long) == 8 &&
+	sizeof(float) == 4 &&
+	sizeof(double) == 8,
+	"this code is only designed for relatively sane environments, plus Windows"
+);
+#endif
+
+// -- A note on performance hackery --
+//
+// Clang won't emit byte-swapping instructions in place of bytewise array writes
+// unless nothing else is written to the same array. MSVC won't do it at all.
+// For these compilers on little-endian platforms that can also do unaligned
+// writes efficiently, we do so explicitly and handle the byte-swapping
+// manually, which then tends to get optimised pretty well.
+//
+// GCC, somewhat surprisingly, seems to be much better at optimising the naïve
+// version of the code, so we don't try to do anything clever there. Also, for
+// unknown, untested compilers and/or platforms, we stick to the safe approach.
+#if defined(_MSC_VER) || defined(__clang__) && (defined(__x86_64__) || \
+		defined(__i386__) || defined(__aarch64__) || defined(__arm__))
+#define USE_BSWAP_NONSENSE
+#endif
+
+#ifdef USE_BSWAP_NONSENSE
+#if defined(_MSC_VER) && !defined(__clang__)
+// MSVC prior to 2022 won't even optimise shift/mask swaps into a bswap
+// instruction. Screw it, just use the intrinsics.
+unsigned long _byteswap_ulong(unsigned long);
+unsigned long long _byteswap_uint64(unsigned long long);
+#define swap32 _byteswap_ulong
+#define swap64 _byteswap_uint64
+#else
+static inline unsigned int swap32(unsigned int x) {
+    return x >> 24 | x << 24 | x >> 8 & 0xFF00 | x << 8 & 0xFF0000;
+}
+static inline unsigned long long swap64(unsigned long long x) {
+	return	x >> 56              | x << 56                    |
+			x >> 40 &     0xFF00 | x << 40 & 0xFF000000000000 |
+			x >> 24 &   0xFF0000 | x << 24 &   0xFF0000000000 |
+			x >>  8 & 0xFF000000 | x <<  8 &     0xFF00000000;
+}
+#endif
+#endif
+
+static inline void doput16(unsigned char *out, unsigned short val) {
+#ifdef USE_BSWAP_NONSENSE
+	// Use swap32() here because x86 and ARM don't have instructions for 16-bit
+	// swaps, and Clang doesn't realise it could just use the 32-bit one anyway.
+	*(unsigned short *)(out + 1) = swap32(val) >> 16;
+#else
+	out[1] = val >> 8; out[2] = val;
+#endif
+}
+
+static inline void doput32(unsigned char *out, unsigned int val) {
+#ifdef USE_BSWAP_NONSENSE
+	*(unsigned int *)(out + 1) = swap32(val);
+#else
+	out[1] = val >> 24; out[2] = val >> 16; out[3] = val >> 8; out[4] = val;
+#endif
+}
+
+static inline void doput64(unsigned char *out, unsigned int val) {
+#ifdef USE_BSWAP_NONSENSE
+	// Clang is smart enough to make this into two bswaps and a word swap in
+	// 32-bit builds. MSVC seems to be fine too when using the above intrinsics.
+	*(unsigned long long *)(out + 1) = swap64(val);
+#else
+	out[1] = val >> 56; out[2] = val >> 48;
+	out[3] = val >> 40; out[4] = val >> 32;
+	out[5] = val >> 24; out[6] = val >> 16;
+	out[7] = val >>  8; out[8] = val;
+#endif
+}
+
+void msg_putnil(unsigned char *out) {
+	*out = 0xC0;
+}
+
+void msg_putbool(unsigned char *out, _Bool val) {
+	*out = 0xC2 | val;
+}
+
+void msg_puti7(unsigned char *out, signed char val) {
+	*out = val; // oh, so a fixnum is just the literal byte! genius!
+}
+
+int msg_puts8(unsigned char *out, signed char val) {
+	int off = val < -32; // out of -ve fixnum range?
+	out[0] = 0xD0;
+	out[off] = val;
+	return off + 1;
+}
+
+int msg_putu8(unsigned char *out, unsigned char val) {
+	int off = val > 127; // out of +ve fixnum range?
+	out[0] = 0xCC;
+	out[off] = val;
+	return off + 1;
+}
+
+int msg_puts16(unsigned char *out, short val) {
+	if (val >= -128 && val <= 127) return msg_puts8(out, val);
+	out[0] = 0xD1;
+	doput16(out, val);
+	return 3;
+}
+
+int msg_putu16(unsigned char *out, unsigned short val) {
+	if (val <= 255) return msg_putu8(out, val);
+	out[0] = 0xCD;
+	doput16(out, val);
+	return 3;
+}
+
+int msg_puts32(unsigned char *out, int val) {
+	if (val >= -32768 && val <= 32767) return msg_puts16(out, val);
+	out[0] = 0xD2;
+	doput32(out, val);
+	return 5;
+}
+
+int msg_putu32(unsigned char *out, unsigned int val) {
+	if (val <= 65535) return msg_putu16(out, val);
+	out[0] = 0xCE;
+	doput32(out, val);
+	return 5;
+}
+
+int msg_puts(unsigned char *out, long long val) {
+	if (val >= -2147483648 && val <= 2147483647) {
+		return msg_puts32(out, val);
+	}
+	out[0] = 0xD3;
+	doput64(out, val);
+	return 9;
+}
+
+int msg_putu(unsigned char *out, unsigned long long val) {
+	if (val <= 4294967295) return msg_putu32(out, val);
+	out[0] = 0xCF;
+	doput64(out, val);
+	return 9;
+}
+
+static inline unsigned int floatbits(float f) {
+	return (union { float f; unsigned int i; }){f}.i;
+}
+
+static inline unsigned long long doublebits(double d) {
+	return (union { double d; unsigned long long i; }){d}.i;
+}
+
+void msg_putf(unsigned char *out, float val) {
+	out[0] = 0xCA;
+	doput32(out, floatbits(val));
+}
+
+int msg_putd(unsigned char *out, double val) {
+	// XXX: is this really the most efficient way to check this?
+	float f = val;
+	if ((double)f == val) { msg_putf(out, f); return 5; }
+	out[0] = 0xCA;
+	doput64(out, doublebits(val));
+	return 9;
+}
+
+void msg_putssz5(unsigned char *out, int sz) {
+	*out = 0xA0 | sz;
+}
+
+int msg_putssz8(unsigned char *out, int sz) {
+	if (sz < 64) { msg_putssz5(out, sz); return 1; }
+	out[0] = 0xD9;
+	out[1] = sz;
+	return 2;
+}
+
+int msg_putssz16(unsigned char *out, int sz) {
+	if (sz < 256) return msg_putssz8(out, sz);
+	out[0] = 0xDA;
+	doput16(out, sz);
+	return 3;
+}
+
+int msg_putssz(unsigned char *out, unsigned int sz) {
+	if (sz < 65536) return msg_putssz16(out, sz);
+	out[0] = 0xDB;
+	doput32(out, sz);
+	return 5;
+}
+
+void msg_putbsz8(unsigned char *out, int sz) {
+	out[0] = 0xC4;
+	out[1] = sz;
+}
+
+int msg_putbsz16(unsigned char *out, int sz) {
+	if (sz < 256) { msg_putbsz8(out, sz); return 2; }
+	out[0] = 0xC5;
+	doput16(out, sz);
+	return 2 + sz;
+}
+
+int msg_putbsz(unsigned char *out, unsigned int sz) {
+	if (sz < 65536) return msg_putbsz16(out, sz);
+	out[0] = 0xC6;
+	doput32(out, sz);
+	return 5;
+}
+
+void msg_putasz4(unsigned char *out, int sz) {
+	*out = 0x90 | sz;
+}
+
+int msg_putasz16(unsigned char *out, int sz) {
+	if (sz < 32) { msg_putasz4(out, sz); return 1; }
+	out[0] = 0xDC;
+	doput16(out, sz);
+	return 3;
+}
+
+int msg_putasz(unsigned char *out, unsigned int sz) {
+	if (sz < 65536) return msg_putasz16(out, sz);
+	out[0] = 0xDD;
+	doput32(out, sz);
+	return 5;
+}
+
+void msg_putmsz4(unsigned char *out, int sz) {
+	*out = 0x80 | sz;
+}
+
+int msg_putmsz16(unsigned char *out, int sz) {
+	if (sz < 32) { msg_putmsz4(out, sz); return 1; }
+	out[0] = 0xDE;
+	doput16(out, sz);
+	return 3;
+}
+
+int msg_putmsz(unsigned char *out, unsigned int sz) {
+	if (sz < 65536) return msg_putmsz16(out, sz);
+	out[0] = 0xDF;
+	doput32(out, sz);
+	return 5;
+}
+
+// vi: sw=4 ts=4 noet tw=80 cc=80
diff --git a/src/chunklets/msg.h b/src/chunklets/msg.h
new file mode 100644
index 0000000..b85bde3
--- /dev/null
+++ b/src/chunklets/msg.h
@@ -0,0 +1,350 @@
+/*
+ * Copyright © 2023 Michael Smith <mikesmiffy128@gmail.com>
+ *
+ * Permission to use, copy, modify, and/or distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED “AS IS” AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
+ * REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
+ * AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
+ * INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
+ * LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
+ * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+ * PERFORMANCE OF THIS SOFTWARE.
+ */
+
+#ifndef INC_CHUNKLETS_MSG_H
+#define INC_CHUNKLETS_MSG_H
+
+#ifdef __cplusplus
+#define _msg_Bool bool
+extern "C" {
+#else
+#define _msg_Bool _Bool
+#endif
+
+/*
+ * Writes a nil (null) message to the buffer out. Always writes a single byte.
+ *
+ * out must point to at least 1 byte.
+ */
+void msg_putnil(unsigned char *out);
+
+/*
+ * Writes the boolean val to the buffer out. Always writes a single byte.
+ *
+ * out must point to at least 1 byte.
+ */
+void msg_putbool(unsigned char *out, _msg_Bool val);
+
+/*
+ * Writes the integer val in the range [-32, 127] to the buffer out. Values
+ * outside this range will produce an undefined encoding. Always writes a single
+ * byte.
+ *
+ * out must point to at least 1 byte.
+ *
+ * It is recommended to use msg_puts() for arbitrary signed values or msg_putu()
+ * for arbitrary unsigned values. Those functions will produce the smallest
+ * possible encoding for any value.
+ */
+void msg_puti7(unsigned char *out, signed char val);
+
+/*
+ * Writes the signed int val in the range [-128, 127] to the buffer out.
+ *
+ * out must point to at least 2 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2}.
+ *
+ * It is recommended to use msg_puts() for arbitrary signed values. That
+ * function will produce the smallest possible encoding for any value.
+ */
+int msg_puts8(unsigned char *out, signed char val);
+
+/*
+ * Writes the unsigned int val in the range [0, 255] to the buffer out.
+ *
+ * out must point to at least 2 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2}.
+ *
+ * It is recommended to use msg_putu() for arbitrary unsigned values. That
+ * function will produce the smallest possible encoding for any value.
+ */
+int msg_putu8(unsigned char *out, unsigned char val);
+
+/*
+ * Writes the signed int val in the range [-65536, 65535] to the buffer out.
+ *
+ * out must point to at least 3 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3}.
+ *
+ * It is recommended to use msg_puts() for arbitrary signed values. That
+ * function will produce the smallest possible encoding for any value.
+ */
+int msg_puts16(unsigned char *out, short val);
+
+/*
+ * Writes the unsigned int val in the range [0, 65536] to the buffer out.
+ *
+ * out must point to at least 3 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3}.
+ *
+ * It is recommended to use msg_putu() for arbitrary unsigned values. That
+ * function will produce the smallest possible encoding for any value.
+ */
+int msg_putu16(unsigned char *out, unsigned short val);
+
+/*
+ * Writes the signed int val in the range [-2147483648, 2147483647] to the
+ * buffer out.
+ *
+ * out must point to at least 5 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3, 5}.
+ *
+ * It is recommended to use msg_puts() for arbitrary signed values. That
+ * function will produce the smallest possible encoding for any value.
+ */
+int msg_puts32(unsigned char *out, int val);
+
+/*
+ * Writes the unsigned int val in the range [0, 4294967295] to the buffer out.
+ *
+ * out must point to at least 5 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3, 5}.
+ *
+ * It is recommended to use msg_putu() for arbitrary unsigned values. That
+ * function will produce the smallest possible encoding for any value.
+ */
+int msg_putu32(unsigned char *out, unsigned int val);
+
+/*
+ * Writes the signed int val in the range [-9223372036854775808,
+ * 9223372036854775807] to the buffer out.
+ *
+ * out must point to at least 9 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3, 5, 9}.
+ */
+int msg_puts(unsigned char *out, long long val);
+
+/*
+ * Writes the unsigned int val in the range [0, 18446744073709551616] to the
+ * buffer out.
+ *
+ * out must point to at least 9 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3, 5, 9}.
+ */
+int msg_putu(unsigned char *out, unsigned long long val);
+
+/*
+ * Writes the IEEE 754 single-precision float val to the buffer out. Always
+ * writes 5 bytes.
+ *
+ * out must point to at least 5 bytes.
+ */
+void msg_putf(unsigned char *out, float val);
+
+/*
+ * Writes the IEEE 754 double-precision float val to the buffer out.
+ *
+ * out must point to at least 9 bytes.
+ *
+ * Returns the number of bytes written, one of {5, 9}.
+ */
+int msg_putd(unsigned char *out, double val);
+
+/*
+ * Writes the string size sz in the range [0, 15] to the buffer out. Values
+ * outside this range will produce an undefined encoding. Always writes a single
+ * byte.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * bytes of the actual string, which must be valid UTF-8.
+ *
+ * out must point to at least 1 byte.
+ *
+ * It is recommended to use msg_putssz() for arbitrary string sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+void msg_putssz5(unsigned char *out, int sz);
+
+/*
+ * Writes the string size sz in the range [0, 255] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * bytes of the actual string, which must be valid UTF-8.
+ *
+ * out must point to at least 2 bytes.
+ *
+ * It is recommended to use msg_putssz() for arbitrary string sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+int msg_putssz8(unsigned char *out, int sz);
+
+/*
+ * Writes the string size sz in the range [0, 65535] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * bytes of the actual string, which must be valid UTF-8.
+ *
+ * out must point to at least 3 bytes.
+ *
+ * It is recommended to use msg_putssz() for arbitrary string sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+int msg_putssz16(unsigned char *out, int sz);
+
+/*
+ * Writes the string size sz in the range [0, 4294967295] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * bytes of the actual string, which must be valid UTF-8.
+ *
+ * out must point to at least 5 bytes.
+ */
+int msg_putssz(unsigned char *out, unsigned int sz);
+
+/*
+ * Writes the binary blob size sz in the range [0, 255] to the buffer out.
+ * Always writes 2 bytes.
+ *
+ * In a complete message stream, a size of N must be immediately followed by
+ * N bytes of the actual data.
+ *
+ * out must point to at least 2 bytes.
+ *
+ * It is recommended to use msg_putbsz() for arbitrary binary blob sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+void msg_putbsz8(unsigned char *out, int sz);
+
+/*
+ * Writes the binary blob size sz in the range [0, 65535] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by
+ * N bytes of the actual data.
+ *
+ * out must point to at least 3 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3}.
+ *
+ * It is recommended to use msg_putbsz() for arbitrary binary blob sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+int msg_putbsz16(unsigned char *out, int sz);
+
+/*
+ * Writes the binary blob size sz in the range [0, 4294967295] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by
+ * N bytes of the actual data.
+ *
+ * out must point to at least 5 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3, 5}.
+ */
+int msg_putbsz(unsigned char *out, unsigned int sz);
+
+/*
+ * Writes the array size sz in the range [0, 15] to the buffer out. Values
+ * outside this range will produce an undefined encoding. Always writes a single
+ * byte.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * other messages, which form the contents of the array.
+ *
+ * out must point to at least 1 byte.
+ *
+ * It is recommended to use msg_putasz() for arbitrary array sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+void msg_putasz4(unsigned char *out, int sz);
+
+/*
+ * Writes the array size sz in the range [0, 65535] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * other messages, which form the contents of the array.
+ *
+ * out must point to at least 3 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 3}.
+ *
+ * It is recommended to use msg_putasz() for arbitrary array sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+int msg_putasz16(unsigned char *out, int sz);
+
+/*
+ * Writes the array size sz in the range [0, 4294967295] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * other messages, which form the contents of the array.
+ *
+ * out must point to at least 5 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 3, 5}.
+ */
+int msg_putasz(unsigned char *out, unsigned int sz);
+
+/*
+ * Writes the map size sz in the range [0, 15] to the buffer out. Values
+ * outside this range will produce an undefined encoding. Always writes a single
+ * byte.
+ *
+ * In a complete message stream, a size of N must be immediately followed by
+ * N * 2 other messages, which form the contents of the map as keys followed by
+ * values in alternation.
+ *
+ * out must point to at least 1 byte.
+ *
+ * It is recommended to use msg_putmsz() for arbitrary map sizes. That function
+ * will produce the smallest possible encoding for any size value.
+ */
+void msg_putmsz4(unsigned char *out, int sz);
+
+/*
+ * Writes the array size sz in the range [0, 65536] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by
+ * N * 2 other messages, which form the contents of the map as keys followed by
+ * values in alternation.
+ *
+ * out must point to at least 3 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 3}.
+ *
+ * It is recommended to use msg_putmsz() for arbitrary map sizes. That function
+ * will produce the smallest possible encoding for any size value.
+ */
+int msg_putmsz16(unsigned char *out, int sz);
+
+/*
+ * Writes the array size sz in the range [0, 4294967295] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by
+ * N * 2 other messages, which form the contents of the map as keys followed by
+ * values in alternation.
+ *
+ * out must point to at least 5 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 3, 5}.
+ */
+int msg_putmsz(unsigned char *out, unsigned int sz);
+
+#ifdef __cplusplus
+}
+#endif
+#undef _msg_Bool
+
+#endif
+
+// vi: sw=4 ts=4 noet tw=80 cc=80