gscm/doc/CHANGES [gscm

gscm/doc/CHANGES

1	Testing & debugging after completion of rewrite:
2
3	1. Read/write ordering bug in pop() -- returned top of stack after having
4	already removed it. Needed to store to a temporary.
5
6	2. "return SC_NULL" instead of "RETURN(SC_NULL)" in reader. Considered renaming
7	the macro, but by any other name it would still refer to the same concept and
8	still be subject to confusion. Seems like a fundamental pun between C and the
9	embedded stack machine language. Wouldn't happen in Scheme since there return
10	is just something that happens to the last expresssion in a sequence, not a
11	command.
12
13	3. Toplevel error handler seems to be getting run without an actual error.
14	Replacing toplevel with more basic evaluation tests for now.
15
16	1 -> ERROR: apply: not a procedure: -945832504593831248
17
18	The error is correct (1 is not a procedure) but the display is wrong.
19	(Incorrect untagging?)
20
21	Problem was in parse_integer / parse_decimal. Before, I used a C heap based
22	dynamic buffer mechanism, which tracked buffer pointer, allocation size and
23	filled size in a struct. In the rewrite I converted this to a string on the
24	Scheme heap, but since that doesn't distinguish fill from allocation, I
25	tracked fill in the lexeme_length global variable. (Ugly, but this one
26	internal use didn't seem to justify a whole new Scheme type.) But in
27	porting the integer parsers I used the Scheme string's length (allocation)
28	rather than the global. The chief hazard seems to be the two notions of
29	length, and passing around this un-truncated buffer in the reader internals
30	for efficiency. Reviewed other use of R_LEXEME for similar problems,
31	finding none. Changed the number parsers to take length as an argument to
32	at least reduce use of the global.
33
34	#(1 2 3) -> ERROR: apply: not a procedure: #()
35
36	Again, a correct error but bad read or print of the vector.
37
38	In the reader, the length header of the vector was being set after
39	the "len" variable had already been used to count down to zero.
40	Rather than building the vector "by hand" I replaced this with the
41	recently added make_vector_uninit function.
42
43	But this also revealed that the local variable "len" was being unsafely
44	used across a recursive call. Rather than push/popping it every time I took
45	it out of the loop to be computed later by list_length(). Except here we
46	already know the list is really a list, so I wrote a new unsafe_list_length
47	function to omit the type and cycle checks of list_length.
48
49	(letrec () 1) -> ERROR: eval: empty combination
50
51	LETREC was passing the body to EVAL_BODY through R_EXPR instead of R_BODY.
52	(Missed during manual register optimization.) Verified the other call to
53	EVAL_BODY. This fixed #3.
54
55	4. A register conflict in the error handler, where r_error_handler was assigned
56	to R_PROC, then R_ERR_DETAIL was referenced, but the latter two had been
57	optimized to share a register. Solved by reassigning R_ERR_DETAIL to a free
58	register. I did go to some pains to verify the register assignments but perhaps
59	gave short shrift to the error handler. It would be a good exercise to do a
60	full audit for register conflicts, once the code is a bit more settled. The
61	rules for this are detailed in a comment.
62
63	5. Two bugs relating to list_length, which is supposed to return -1 for safely
64	detecting non-lists. First, builtin_list_length stored its result in a "value"
65	instead of a "long", which being an unsigned type, prevented recognizing the
66	negative. Second, for odd length improper lists (e.g. (1 2 3 . 4)), the is_pair
67	check after the first of the two "fast = cdr(fast)" pointer advancements in the
68	loop returned directly, bypassing the "fast != SC_NULL" improper list check
69	after the loop. This irregularity resulted from being overly preoccupied with
70	(premature) hand-optimization (doing a single length += 2 after the two
71	advances in each loop instead of a length++ after each one), a tendency I've
72	had to fight in other places as well. (Chiefly, using the friendly type
73	constructor and accessor functions, which tag/untag values as needed, rather
74	than duplicating their functionality to save a few instructions on redundant
75	tag operations. In this, it helped that I convinced myself, by examining the
76	generated assembly, that the compiler optimizations really are smart enough to
77	factor out the tag operations, at least in some cases, and at least with static
78	functions.)
79
80	6. In SET, checked for cdr(R_EXPR) == SC_NULL after having overwritten the
81	intended R_EXPR with car(R_EXPR). Swapped the order. (Discovered by running the
82	"bad" semantic test set, but woulda come up soon enough anyway as it completely
83	broke "set!".)
84
85	(Later... somewhere around Jan 2017)
86
87	7. yes '(' \| gscm -> inf loop between gc/error handler for out of memory.
88	sc_error1 does a string() allocation to set R_ERR_MSG, but the stack doesn't
89	get dropped until after the longjmp. Fixed by dropping stack in the error
90	handler. Also wrote a toplevel error handler callable from Scheme implementing
91	the same behavior using a captured continuation.
92
93	8. The "apply" builtin did not copy its final argument (a list). The standard
94	doesn't seem to specify whether it should or not, but "list" is required to
95	return a newly allocated list, and my no-op implementation was assuming its
96	argument list was already a fresh copy. Wrote a test case and fixed by making
97	builtin_apply copy. Also broken in tinyscheme. Affects what is probably the
98	simplest way to copy a list, (apply list some-list).
99
100	Major refactoring: compiler, variable references...
101
102	October 2017:
103
104	9. Register effect bugs in assoc and member builtins (see tests-misc.scm).
105
106	10. Miscompilation of "delay" (see tests-misc.scm).
107
108	11. The compiler was not immune to redefinition of builtins. (REPL too?)
109	Implemented immutable environments with a separate copy for the interaction
110	environment. Immutable string/list/vector constants still need doing.
111
112	Implemented piped subprocess extension
113
114	12. Realized input_port()/output_port() could leak the file object and
115	descriptor if they hit an out of memory error.
116
117	November 2017: fuzz testing
118
119	13. Infinite loop in the lexer for EOF in a comment.
120
121	Implemented CLI options, memory stats, floor/ceiling/truncate/round/port?
122
123	January 2018:
124
125	14. Heap discipline broken in arithmetic builtins on flonums by way of "fold"
126	helper.
127
128	April 2018:
129
130	Partial implementation of the long-missing number->string and string->number
131	builtins, enabling removal of a bunch of printfs.
132
133	15. Discovered by inspection that the open_file helper was unsafely storing a
134	port in a local variable across a call to string_append_null. Made it pass the
135	port by register instead and refactored a bit.
136
137	16. Realized that the changes implemented for #12 can't actually solve the problem;
138	file objects and descriptors still leak if a subsequent unrelated call gets the
139	memory error. Backed them out in order to simplify port construction.
140
141	17. builtin_open_subprocess was a mess all around. Changed the interface,
142	improved comments, simplified the temporary argv construction (resolving a
143	strict aliasing warning), and tightened syscall error handling. As above,
144	realized that leaking resources due to allocation failures is not a problem
145	that can be solved within the scope of this function (so e.g. it's fine to
146	allocate the return value list after the child is spawned).
147
148	18. REPL now accepts multi-valued returns, printing them one per line.
149
150	19. Extensions are now available in an immutable (gales-scheme-environment),
151	analogous to (scheme-report-environment 5), in addition to the interaction
152	environment. This allows for hygienic packages making use of the extensions
153	without dependency injection. Probably useless flush-input-port removed. Some
154	toplevel error handler setup inconsistencies fixed, but the error plumbing
155	needs more scrutiny.
156
157	20. Implemented double-width multiply functions, signed and unsigned, in
158	portable C and x86_64 assembly, plus range checking and float conversion
159	helpers. Used these to fix long-standing overflows in * and expt builtins.
160	Implemented extensions for wrapped, carried and bitwise arithmetic on fixnums.
161	Factored tagging macros out of machine-specific #if-blocks and tightened their
162	types by casting. Need to finish correctness proof for signed wide multiply.
163
164	May 2018:
165
166	21. Fixed self-initialization bug in macro expansion in fx-/wrap builtin (and
167	added -Winit-self to gcc flags). Added fx-/carry-unsigned and fxmaj builtins.
168	Made some internal functions static. Added startup assertion to enforce the
169	assumption that unsigned long matches the value / pointer size. Minor tweaks to
170	build in C99 pedantic mode; would like to move toward C89.
171
172	22. Implemented if-exists extension for open-output-file. Switched to "let"
173	forms in toplevel. Made static symbol variable setup/usage more consistent.
174
175	23. Core:
176	- Catch I/O errors in flush_port, read_char and peek_char, simplifying spec for
177	flush-output-port builtin.
178	- Light comment editing. Remove specification comments for nonstandard
179	builtins, which now have their own extensions document.
180	- Make eqv? and equal? do the simpler checks first, short-circuiting deep
181	inspection where possible.
182	- Rename quit builtin to exit.
183	- Add a GC thrash factor (knob not exposed yet) to hasten OOM detection when GC
184	reclaims less than heap_size/factor space.
185	- Wipe registers on OOM detection (previously just the stack was dropped in
186	sc_error1, with reason unclear).
187	- Rename fallback error handlers to fatal/fatal1 for clarity.
188
189	Core/compiler/toplevel:
190	- Implement "error" as a normal builtin, parallel to sc_error1, making
191	set-error-handler! work for Scheme-signalled errors too and removing the
192	compiler's special error procedure with its null-terminator wrangling.
193	- Remove syn-env from the special compiler environment as it can be handled
194	purely from Scheme.
195	- Merge the compiler environment (now just one binding) into the privileged
196	toplevel environment.
197
198	Toplevel:
199	- Extend call-with-output-file and with-output-to-file to pass through the
200	if-exists option.
201	- Make with-input-from-file and with-output-to-file restore the current ports
202	on escape or reentry, not required by spec but clearly the right thing.
203	(Using a stub dynamic-wind, so won't actually work yet.)
204	- Fix all four of those to work with multi-valued returns.
205	- Make REPL error handler support multiple detail arguments (though the other
206	components don't yet).
207
208	24. Added sync and data-sync options to flush-output-port extension.
209	Refactoring of port argument wranglers.
210
211	25. Big move of builtins to library. Startup time and performance of the
212	builtins in question is degraded, but much hazardous and hard-to-verify
213	microcode is eliminated.
214
215	Removed compiler dependence on builtin forms:
216	and
217	or
218
219	Removed compiler dependence on builtin procedures:
220	append
221	apply
222	assq
223	for-each
224	list
225	list-tail
226	map
227	memq
228	vector->list
229	zero?
230
231	Compiler now depends on some extensions and uses eq? for fixnum comparison, in
232	hopes of reducing performance impact on replaced builtins.
233
234	Merged the syntax library into the toplevel, removing some startup complexity
235	and allowing the toplevel to use library syntax.
236
237	Fixed some edge cases of the "case" macro.
238
239	The "load" builtin, by way of exec-from-port, now reads the whole file before
240	evaluating the forms.
241
242	Forms moved from core to syntax library:
243	and
244	or
245
246	Builtins moved to library:
247	list
248	append
249	reverse
250	list-tail
251	list-ref
252	memq
253	memv
254	member
255	assq
256	assv
257	assoc
258	vector->list
259	apply
260	map
261	for-each
262
263	New privileged builtins:
264	apply/unchecked
265	car/unchecked
266	cdr/unchecked
267	set-cdr/unchecked!
268	vector-ref/unchecked
269	fx+/unchecked
270
271	Wrote some tests for the new library procedures. Accelerated negative tests by
272	replacing bash/sed gunk with a custom error handler to run all in one process.
273	Moved the currently failing immutability test to the negatives and expanded.
274
275	June 2018:
276
277	26. Make "string-append" run in linear rather than quadratic time, fixing a
278	FIXME. Make toplevel "apply" and "append" list copying helpers work tail
279	recursively, by mutation in one pass as in the original C, reducing memory
280	overhead.
281
282	27. Implemented dynamic-wind by adding a "spool" register and a new field in
283	continuation objects to capture it, the necessary additions to the APPLY
284	subroutine (now separated into APPLY_CONTINUATION), and a lowest common
285	ancestor helper function.
286
287	Substantially refactored error handling in both core and toplevel to:
288	* work with dynamic winding while avoiding infinite loops;
289	* avoid some unnecessary C to Scheme string conversions;
290	* remove string_append helper, a poor interface (per change 26);
291	* support multiple detail arguments to ERROR, per SRFI-23;
292	* be generally clearer, I think.
293
294	Similarly, the EXIT extension is now implemented using a continuation that
295	returns the given code from the toplevel (and sc_toplevel now returns the code
296	to main). This both makes it support unwinding and removes the need for a
297	builtin exposing exit(). (At some point it might be nice to eliminate remaining
298	uses of exit(), but so far it's still needed as an error handling fallback.)
299
300	Builtins moved to library:
301	string
302	string->list
303	vector
304	set-error-handler!
305	exit
306
307	New privileged builtins:
308	push-winding!
309	string-ref/unchecked
310	set-error-continuation!
311
312	Extension added to library:
313	error
314
315	builtin_error still exists, but only for reporting startup errors before it's
316	replaced by the toplevel.
317
318	Fixed a few cases where pointers were being followed into the heap without
319	untagging. This had gone unnoticed because the three tag bits happen to
320	overflow in the shift for array indexing on 64-bit.
321
322	Some minor clarity and error context improvements to library procedures.
323
324	Renamed list_length/unsafe_list_length, to avoid giving the impression that
325	safety can be assumed for these internal helper functions.
326
327	Some C89ist shuffling of declarations.
328
329	Optimized builtin_cons: reuse pair from arg list as it's freshly allocated.
330
331	sc_toplevel now returns the integer exit code,
332
333	TODO: think thorough / test what happens on error or other escape in
334	unwinding/rewinding, explicitly unspecified by R5RS.
335
336	28. Minor optimization: compile quoted empty lists and quoted vectors to the
337	values themselves, as they're not valid expressions and thus can be made
338	self-evaluating simply by removing the check. Deprioritize "quote" in the
339	builtin syntax dispatch (based on pure guesswork as to frequency). 1% speedup
340	observed in some existing tests, 5% in a contrived one.
341
342	Condense some redundant register aliases (R_OPERATOR->R_EXPR;
343	R_BODY->R_OPERANDS), avoiding some pedantic self-assignment.
344
345	29. Fixed a subtle semantics bug in FORCE, and wrote test: a promise's code may
346	be entered more than once if it forces itself recursively; if results differ,
347	the first must not be clobbered. Found by re-reading rationale in R5RS.
348
349	In the same spirit of "mutability is not your friend", reverted the
350	builtin_cons optimization from #27. While safe, arglist mutation is premature
351	at this point; for example it may hinder debugging, which ought to be a
352	strength of an interpreter. If I do decide to allow it, there's a bunch of
353	other builtins that could do it too. But a better optimization might be to
354	avoid building arglists in the first place for fixed-arity builtins.
355
356	30. New builtins, with tests:
357	immutable?
358	cons/immutable
359	string-copy/immutable
360	vector-copy/immutable
361
362	Slightly simplified some related code.
363
364	In previous immutability work, rather than expose new primitives I had the
365	reader construct strings and vectors as immutable. This was neither correct nor
366	useful. Literal constants are now properly petrified by the compiler, and with
367	that, all my tests are finally passing.
368
369	31. Mega-IO-refactor, replacing stdio with Scheme-managed buffering, and
370	implementing sockets extension.
371
372	Port-based write functions now all use R_PORT rather than an argument and
373	signal errors rather than returning status. This tidies up the printer quite a
374	bit.
375
376	Checking for open port on each read/write no longer needed - the kernel's FD
377	validity check is enough.
378	- Added bonus: (close-output-port (current-output-port)) at REPL was a segfault
379	as the check had been missed in the fallback error handler. Now "FATAL: Bad
380	file descriptor".
381	- Should probably prevent closing stdin/out though. (close-input-port
382	(current-input-port)) at REPL is an infinite loop.
383
384	Replaced posix_spawn* with vfork, simplifying builtin_open_subprocess, possibly
385	making it fast on more platforms, and removing another likely malloc user.
386
387	New privileged builtins operating at the file descriptor level:
388	inet-stream-socket
389	inet-dgram-socket
390	unix-stream-socket
391	unix-dgram-socket
392	socket-ports
393	getsockname
394	getpeername
395	connect-inet
396	connect-unix
397	listen
398	accept
399	close
400
401	New library procedures:
402	open-tcp-connection
403	open-tcp-listener
404	open-udp-socket
405	open-unix-connection
406	open-unix-listener
407	open-unix-datagram-socket
408	call-with-tcp-connection
409	call-with-unix-connection
410	sequential-tcp-server
411	sequential-unix-server
412
413	32. Fix assertion violation (buffer overflow) introduced in #31, when writing
414	to a closed port, catching the error, then writing again.
415
416	Relatedly, any other write errors on the underlying FD will now close the port.
417	(Rationale: if recovery were possible, the underlying system should already be
418	doing it, so write errors are generally permanent; even if not, they can be
419	asynchronous so no way to know how much was successfully written.)
420
421	REPL error handler changed back to restarting using a captured continuation. (I
422	think I had removed this deliberately in #27 in the spirit of simplifying, but
423	the result was the REPL exiting with error status on a normal ^D termination if
424	an error had ever been caught, because error handlers aren't really supposed to
425	return.)
426
427	Ugly (define-r5rs 'foo (lambda ...)) forms changed to internal definitions
428	followed by bulk registration, in the same manner as the extensions in #31.
429
430	33. New builtin: char-ready?, made possible by direct access to input buffers.
431	Stream sockets are made nonblocking to work around poll/select defects, but
432	undetected blocking is still possible, e.g. reading from disk or if standard
433	input is an externally provided socket.
434
435	Builtin moved to library and tests added: equal?
436
437	34. Libc in-sourcing:
438	- Removed stdio from main.c and indirect use via assert
439	- Removed last direct uses of malloc by replacing GC root linked list with an
440	array and heap creation with mmap, also enabling the hugepages option
441	- Removed last indirect use of malloc via atexit
442	- Replaced stdlib.h/string.h with individual declarations
443
444	Made write_err handle EINTR/EAGAIN/EWOULDBLOCK, sharing with flush_output_port.
445
446	July 2018:
447
448	35. Start of library-level bignum support, including addition, subtraction,
449	Comba multiplication, and automatic fixnum promotion and demotion. Using
450	non-opaque values until the proper privileged interface is settled.
451
452	It looks like the reader and printer will have to be done in Scheme now.
453	(They'd have to make non-tail calls back into Scheme to do base conversion, and
454	such C-level recursion is verboten.) So much the better for sanity and future
455	rewrites of the core though. A pared-down bootstrap reader will be preserved.
456
457	Core:
458	- More assertions
459	- Flush ports on abort (with re-entry check, as flushing itself has assertions)
460	- Bignum awareness in number predicates
461	- Start replacing car(cdr(x)) pattern with cadr macro
462
463	Builtins moved to library: + - *
464
465	New builtin: fx</unsigned
466
467	New privileged builtins:
468	fx-/unchecked
469	fx=/unchecked
470	fxneg/unchecked
471	fxnegative/unchecked?
472	bignum?
473	flonum?
474	flonum/unchecked
475	floneg/unchecked
476	flo+/unchecked
477	flo-/unchecked
478	flo*/unchecked
479
480	Renamed builtin: fx-/carry-unsigned to fx-/borrow-unsigned. (Makes more sense
481	to me as the overflow output is 1, not -1.)
482
483	Syntax library:
484	- SRFI-8 RECEIVE syntax to simplify call-with-values. (LET-VALUES is nicer but
485	ugly to implement.) Ideally this wouldn't be visible in the R5RS environment.
486	- Fix quirk where COND with no successful tests returned #f rather than the ()
487	I'm generally using for "unspecified". This also worked around the compiler
488	bug that caused it to reject a literal #f as the final test.
489
490	36. Serious bugfix: missed making the spool register a GC root in #27.
491
492	Minor bugfixes:
493	- Proper PEEK-CHAR semantics for devices with transient end-of-file e.g. tty.
494	- Detect excess arguments in the helper macro for unary character functions.
495
496	New builtin: read-token
497
498	New privileged builtin: reverse-list->vector/unchecked
499
500	Builtins moved to library: read string->number
501
502	The full range of numeric syntax is now supported except for rational and
503	complex: alternate radix (2/8/10/16), radix and exactness prefixes, exponents
504	and sharped-out digits (though significant figures aren't actually tracked).
505	Bignums come "for free", though base conversion is the naive quadratic
506	algorithm.
507
508	Simplified the (now) bootstrap number decoder by removing float syntax and
509	promotions.
510
511	Internal tweaks to the lexer and (now) bootstrap reader to support read-token.
512	Also loosened the overly anal interpretation of allowed whitespace.
513
514	Started downcasing labels (all-caps doesn't really serve a purpose and risks
515	interference from libc extensions).
516
517	37. Bignum division, conversion to flonum and output.
518
519	Bugfix: bignum negation failed to demote in the case of LEAST-FIXNUM.
520
521	Minor optimization: avoid some unnecessary untagging in fixnum ops.
522
523	Builtins moved to library:
524	zero?
525	positive?
526	negative?
527	abs
528	quotient
529	remainder
530	modulo
531	exact->inexact
532	number->string
533	write
534	display
535	newline
536
537	In addition to bignum support, quotient/remainder/modulo now properly maintain
538	inexactness for flonum inputs, though can produce inf/nan.
539
540	In the Scheme rewrite of the printer, I realized the special save/restore of
541	current ports for error recovery is unnecessary now that dynamic-wind is taking
542	care of it (they can't be set directly by user code).
543
544	New library procedures: gcd lcm
545
546	New builtins: fx= fx< fx<= fx<=/unsigned fxlength/unsigned
547
548	Renamed builtin: fxshift-unsigned to fxshift/unsigned
549
550	New privileged builtins:
551	fxdiv/unchecked
552	fxdiv/ext/unchecked
553	fixnum->dec/unchecked
554	fixnum->hex/unchecked
555	fixnum->oct/unchecked
556	fixnum->bin/unchecked
557	flonum->dec/unchecked
558	flonum/unsigned/unchecked
559	flodiv/unchecked
560	floquotient/unchecked
561	floremainder/unchecked
562	builtin?
563	builtin-name
564	promise?
565	continuation?
566
567	38. Full bignum integration.
568
569	Serious bugfix: (QUOTIENT LEAST-FIXNUM -1) => LEAST-FIXNUM
570	(On the first pass for #37 it didn't occur to me that signed fixnum quotient
571	had an edge case besides zero divisor. The equivalent problem existed in the
572	prior C-only implementation, or possibly worse as C leaves negative division
573	implementation-defined.)
574
575	Builtins moved to library:
576	eqv?
577	= < > <= >=
578	odd? even?
579	max min
580	/
581	floor ceiling truncate round
582	exp log
583	sin cos tan
584	asin acos atan
585	sqrt
586	expt
587	inexact->exact
588
589	Bignums now opaque and supported by all generic operators. Unlike the fixnum
590	and flonum ops, which could also exist as unprivileged extensions, the bignum
591	primitives are implicitly unchecked. In addition to bignum support, the
592	irrational functions now handle some special cases for exactness e.g. zero (as
593	encouraged but not required).
594
595	New privileged builtins:
596	toplevel-environment
597	set-car/unchecked!
598	fx</unchecked
599	fx<=/unchecked
600	fixnum->bin/unsigned/unchecked
601	flo=/unchecked
602	flo</unchecked
603	flo<=/unchecked
604	fraction/exponent/unchecked
605	load-exponent/unchecked
606	inf/unchecked?
607	flonum->fixnum/unchecked
608	floor/unchecked
609	ceiling/unchecked
610	truncate/unchecked
611	round/unchecked
612	exp/unchecked
613	log/unchecked
614	sin/unchecked
615	cos/unchecked
616	tan/unchecked
617	asin/unchecked
618	acos/unchecked
619	atan/unchecked
620	atan2/unchecked
621	sqrt/unchecked
622	make-bignum
623	bignum?
624	bignum-negative?
625	bignum-set-negative!
626	bignum-ref
627	bignum-set!
628	bignum-length
629	bignum
630	bignum/unsigned
631	bignum2
632	bignum-truncate!
633
634	Renamed privileged builtins:
635	fxdiv/unchecked -> fxdiv/unsigned/unchecked
636	fxdiv/ext/unchecked -> fxdiv/ext/unsigned/unchecked
637
638	(They were already unsigned; I think I'd been trying to avoid the crazy long
639	names, but it was a hazard.)
640
641	Core:
642	- Grabbed the last extended type slot to fit the bignum sign bit
643	- Type assertions added to fixnum_val, new counterpart unsigned_fixnum_val, and
644	flonum_val, expanding coverage while reducing boilerplate
645	- Bounds check assertion added to vector_ref and vector_set
646	- With flonum coercion headaches out of core, finally replaced safe_fixnum_val
647	and exact_fixnum_val with something simple and correct
648	- is_number optimized based on extended number type tag values being contiguous
649	- is_integer made possibly more standard conformant by using ROUND (nearbyint)
650	instead of FLOOR for flonums (don't see how this could make a difference but
651	neither am I sure it can't)
652	- Made shallow_print report particular environment specifiers, the internal
653	specials, immutability, and vector lengths
654	- make_str and make_vector size checks simplified based on the unsigned trick
655
656	Compiler:
657	- Removed fairly useless vector support from macro language
658	- Fixed old bug: improper rejection of macros expanding to #f
659	- Removed trivial dependency on generic <
660	- Switched to storing error context as symbol, avoiding SYMBOL->STRING calls in
661	the normal case
662
663	Toplevel:
664	- Pretty "evaluates to" symbol prefixing REPL results
665	- Tracing for library procedure calls, allowing removal of the growing sprawl
666	of ad-hoc error context reporting
667	- Filled in stubs for bignum to hex, octal and binary conversion
668	- Handled bignum status in EXIT by remaindering there rather than in core
669
670	asm-generic:
671	- Replaced "seems right but can't prove" implementation of signed double-width
672	multiply with ugly but more obvious method
673
674	August 2018:
675
676	39.
677	Serious bugfix: plug leaks of uncompiled subtrees in named LET and DO, while
678	simplifying, by feeding the full expansions back through compilation as is done
679	for ordinary macros. (Less seriously, the same for lambda annotation in plain
680	LET, at worst a performance issue.)
681
682	Core:
683	- Add is_port assertions to R_PORT-based I/O routines
684	- Remove unnecessary variable pre-check in SET, allowing resolve_variable_ref
685	to operate on unresolved refs only, saving a few branches
686	- Remove fairly pointless and expensive is_list assertion in EVAL
687	- Remove redundant is_fixnum assertion from builtin_make_bignum
688
689	Compiler:
690	- More consistently unquote self-evaluating objects
691	- Reject non-readable objects (possible through EVAL)
692
693	Syntax library:
694	- OR: optimize by avoiding unnecessary temporary binding (multiple evaluation
695	is safe for non-list expressions as they're all side-effect free and O(1))
696	- Explain how the hardcoded temporary names are safe
697	- COND: simplify by handling the temporary binding case with OR
698	- CASE: optimize by using EQV? instead of MEMV for the single-datum case
699
700	40.
701	Core/compiler:
702	- Prevent integer overflows by limiting parameter list length (number of
703	bindings per lexical frame)
704	- Save a word in the procedure object and a pair in the annotated lambda form
705	by encoding variadic status in the sign of the arity
706
707	Core:
708	- Avoid possibly undefined overflow behavior in untag_signed by doing the left
709	shift as unsigned
710	- Note a hazard in add_tag/ext_add_tag and audit
711
712	Compiler:
713	- Slight simplifications/optimizations: filterize CHECK-IDENT; avoid repeated
714	SYMBOL->STRING in duplicate search
715
716	40.maint1.
717
718	Starting a maintenance branch as the crucial variable lookup optimization has
719	not yet made it back after the compiler overhaul of revision 41.
720
721	Core:
722	- Remove assertions in stack ops, redundant since adding type assertions to
723	car/cdr long ago
724	- Remove invalid assertion added in r39: input_port_ready doesn't use R_PORT
725	- Remove mostly redundant is_list assertions in the full environment lookups
726
727	March 2020:
728
729	40.maint2.
730
731	Overall:
732	- Restructure tree for slashpackage
733	- Trim Makefile
734	- Write install + check scripts and README
735
736	Main (backports):
737	- Make default heap size build-time tunable (isolating magic numbers)
738	- Report build configuration in help text
739	- Remove "strtol" stdlib dependency
740
741	40.maint3.
742
743	Bulk reindent & reflow.
744	- Tabs: good
745	- Non-meaningful newlines in human text: bad
746	- "if" as a special "lisp word" producing less indentation: bad
747	- Code width: mostly sticking to 80 columns
748
749	September 2020:
750
751	40.maint4.
752
753	Core:
754	- Consistently allow HT, FF, and CR as whitespace in the lexer. (Early versions took a strict view of the standard and only allowed space or newline; at some point I loosened this to allow tab-indented files, but missed the delimiter lookahead cases.)
755
756	Main:
757	- Fix -m argument parsing from bad backport in 40.maint2.
758
759	Packaging:
760	- Fix version symlink update steps. (When a previous version existed, /package/gscm was being followed rather than replaced. djb got this right but apparently I tried to 'improve' it.)
761	- Placeholder file doesn't need to be hidden.
762
763	October 2020:
764
765	40.maint5.
766
767	Core:
768	- Bugfix: the string-fill! and vector-fill! builtins didn't enforce immutability. (I hadn't originally implemented immutable subtypes, which are hinted at but not specified in R5RS; when adding them I put checks in the obvious "set!" family of mutators, but missed the bulk fill operations. This might allow subversion of downstream checks that depend on immutability, though the only example coming to mind is the structural checking of expression trees passed to EVAL, and the parts of those that matter are all pairs which aren't affected here.)
769
770	November 2021:
771
772	40.maint6.
773
774	Packaging:
775	- Command symlink now installs to /usr/bin instead of /command, based on the view that the latter only complicates installation without actually solving any problem. The install script won't delete the old one if present, but it also won't create it e.g. for backward compatibility, as I figure it'll be easy enough to update any downstream references and I'd rather not perpetuate top-level filesystem clutter. At any rate I don't expect to change it again.
776
777	Documentation:
778	- README moved from 'package' subdir to top level and massaged a bit.