[main]Notes on TeXmacs

Scheming

We discuss some aspects of the choice of the Scheme language implementation used in TeXmacs.

[6.1.2021]

In TeXmacs we use GNU Guile version 1.8 as embedded Scheme interpreter. Lately, this has been a source of problems, mostly related to packaging. On one hand Guile 1.8 is no more supported in standard Linux distributions. This implies that we have either to ship Guile in our codebase or maintain a private repository for Guile 1.8. On the other hand more recent versions of Guile compile the sources to bytecode and introduce a separated expansion and evaluation phases which break part of our codebase. Moreover while Guile 2 has a fine compiler, it has been for some time quite slow in interpreting code and only recently with the introduction of a JIT compiler in Guile 3 the interpreted code runs as fast as in Guile 1.8. Finally we have some difficulties in compiling Guile 2 on Windows and Guile 3 seems not yet be supported there. The GNU Lilypond project seems to have run into similar issues with Guile and they also remained with Guile 1.8.8 (see here).

This situation is not very nice and we would like to move forward and remove these problems. Therefore we are currently evaluating various possible strategies to evolve the way we run Scheme code within TeXmacs. The following are the criteria we want to take into account.

We use Scheme as an extension language to run on top of the GUI and the typesetting code in order to implement the dynamical and programmable UI. Additionally, if the implementation speed allows we can shift some of the conversion code in Scheme. We have C++ at our disposal anyway so we can always drop there for speed-critical tasks, but is nicer to be able to write Scheme code instead of C++.
TeXmacs already needs a huge library like Qt in order to be cross-platform, so in this respect we would like not to carry over a huge all-purpose Scheme implementation which would come with all its own libraries dependencies (e.g. Guile needs GMP).
We need to be cross-platform so we would like to minimize the problem related to porting the Scheme implementation on the various OSs and architectures.
Guile 1.8 speed is OK so far. However we cannot loose performance (especially at boot time and for standard operations like moving the cursor around). We could even hope to gain some speed by using an alternative implementation.
TeXmacs do not currently use Unicode for its internal string representations but a custom formate (TeXmacs universal representation), so marshaling data to Unicode-aware Schemes is a mildly issue for us.

As for the available alternatives, the situation so far is the following.

An easy choice would be to just integrate in our codebase the sources of Guile 1.8 (which is not more maintained) in TeXmacs. This would not require changes to our scheme code. However the packaging problems remains, moreover there are large portions of the library we do not use.
We can go with Guile 3 and get the nice speed improvement wrt. 1.8. See here for some updates on the development of Guile. This will require more or less extensive changes to our Scheme code. These changes are been implemented in a git development branch and so far the situation seems promising. One drawback of this choice is the fact that support for platforms like Windows is not the primary goal of the Guile developers (as far as we understand) and also that Guile is become a large standalone scheme with many library dependencies, far from its origin as extension language. So the benefits of the compiler should be weighted against the difficulty of packaging and also the stability of the cross-platform support. Also Guile 2/3 has encoding-aware strings which poses a (small) problem for us and would require some hackery to get around.
An attractive option is to switch to another small, embeddable Scheme interpreter like Chibi Scheme or S7 Scheme. For performance reasons it seems that the clear winner for us is S7 which is a very compact interpreter written in C with no external dependencies. We are currently experimenting in integrating S7 with TeXmacs and we have a working development branch (with still some issue).
A radical alternative is to consider the embedding of a full-fledged Scheme compiler. By looking at the multitude of Scheme implementations around we singled out Chez Scheme as a compiled system which seems flexible enough to be able to be embedded in our setting. Its main features are: compactness, speed, industrial quality, open-source (Apache 2.0), cross-platform and with no external dependencies. (For an interesting history of this implementation refer to this paper.)

The development branches of TeXmacs for Guile 3 and S7 can be found here:

We performed some test of these alternative implementations to help us make a preliminary picture. We used some standard R7RS benchmarks found here and we got the results shown below on a MacBook Air (2019). They show that S7 is consistently the fastest interpreter wrt. Chibi or Guile 1.8. We also see that, as expected, Chez is among the fastest implementations available with consistent timings all across the benchmarks.

We currently have working implementations of TeXmacs with Guile 1.8, S7 and Guile 3.0.4 and what we see from preliminary tests is that boot time of S7 is comparable to that of Guile 3.0.4 (0.4 sec) and half of that of Guile 1.8 (0.8 sec). Also compile the full manual require the roughly the same time to all the three implementations (~15 sec). We would need to consider more intensive tests to discriminate the tradeoffs in performances of the various choices.

Here are the results of the R7RS benchmarks. The symbol ∅ denotes that the program exceeded a conventional time limit and was interrupted.

test	s7	chibi 0.9.1	chez 9.5.1	guile 1.8.8	guile 3.0.4
browse:2000	24.27		0.84	76.63	12.06
deriv:10000000	25.19	97.11	1.15	67.27	18.58
destruc:600:50:4000	52.08	94.90	2.21		7.14
diviter:1000:1000000	9.69	55.88	1.81	82.75	15.45
divrec:1000:1000000	11.80	50.91	2.19	82.04	17.41
puzzle:1000	27.72	270.98	1.66	221.49	18.09
triangl:22:1:50	33.93	110.86	1.91	107.15	8.52
tak:40:20:11:1	12.93	55.19	1.66	134.21	4.76
takl:40:20:12:1	20.97		3.71		9.46
ntakl:40:20:12:1	17.07	95.98	3.65		9.52
cpstak:40:20:11:1	103.36	222.13	4.21	258.62	59.44
ctak:32:16:8:1	44.14		0.96
fib:40:5	10.22	96.25	3.63	236.65	12.09
fibc:30:10	25.80		0.71
fibfp:35.0:10	1.89	42.79	3.18	56.26	22.00
sum:10000:200000	6.64	99.79	3.59		6.87
sumfp:1000000.0:500	2.50	85.36	4.25	111.78	42.06
fft:65536:100	32.20	69.76	3.32		7.69
mbrot:75:1000	24.40	209.51	5.39		50.09
mbrotZ:75:1000	18.56		9.26		67.01
nucleic:50	19.95	79.05	2.59	69.32	15.35
pi	–		0.60		0.56
pnpoly:1000000	17.98	253.84	4.73		24.89
ray:50	20.46	119.63	2.83		18.51
simplex:1000000	46.34	182.76	2.34		13.90
ack:3:12:2	10.57	74.51	3.12		8.41
array1:1000000:500	11.48	64.18	8.16	138.45	9.24
string:500000:100	1.71	6.34	6.54	1.81	1.87
sum1:25	0.47	121.78	2.02	1.71	4.43
cat:50	1.19	70.95	2.69		28.40
tail:50	1.19	11.21	3.57		9.82
wc:inputs/bib:50	8.27		1.76	73.34	16.96
read1:2500	0.41	281.23	1.28	2.69	5.80
compiler:2000	41.16	115.83	2.99		5.15
conform:500	51.03	199.42	2.10		10.51
dynamic:500	22.74	288.25	4.01	71.60	7.37
earley			5.03		9.49
graphs:7:3	127.61		2.27		23.03
lattice:44:10	139.28		3.91		15.94
matrix:5:5:2500	72.07	214.16	1.63		9.88
maze:20:7:10000	23.26	84.15	1.66		4.70
mazefun:11:11:10000	19.51	122.02	2.86	128.66	9.66
nqueens:13:10	55.11	165.25	4.99		19.37
paraffins:23:10	31.42		5.97		4.25
parsing:2500	39.44		3.35		10.69
peval:2000	29.68	177.10	2.05	107.05	15.64
primes:1000:10000	7.73	21.08	2.64	43.73	7.52
quicksort:10000:2500	94.00		3.90		13.25
scheme:100000	71.46	154.36	2.55		15.14
slatex:500	32.07	173.60	3.73	43.82	45.05
chudnovsky	–		0.32		0.31
nboyer:5:1	39.27	41.54	3.95	142.86	5.10
sboyer:5:1	31.54	39.14	1.30	155.49	4.76
gcbench:20:1	20.54		1.87		3.51
mperm:20:10:2:1	173.33	659.14	13.33		10.65
equal:100:8:1000:2000:5000	0.78	54.11	0.99
bv2string:1000:1000:100	10.78	11.17	2.47		4.49

[8.1.2021]

Some more benchmarks. Populating the autocompletion tree (i.e. pressing Tab in a Scheme session within TeXmacs)

	symbols	time (msec)
guile 1.8.8	8639	686
guile 3.0.5	10300	217
S7	7360	255

[12.4.2021]

Another interesting system is femtolisp which is currenlty in use within Julia for the initial stage of parsing and syntactic elaboration. It is a small lisp with lexical scoping and bytecode interpreter. The benchmarks have been updated to include it. S7 seems still faster on average.

test	femtolisp	s7	chibi 0.9.1	chez 9.5.1	guile 1.8.8	guile 3.0.4
browse:2000		24.27		0.84	76.63	12.06
deriv:10000000	11.56	25.19	97.11	1.15	67.27	18.58
destruc:600:50:4000	27.05	52.08	94.90	2.21		7.14
diviter:1000:1000000	20.02	9.69	55.88	1.81	82.75	15.45
divrec:1000:1000000	20.64	11.80	50.91	2.19	82.04	17.41
puzzle:1000	74.62	27.72	270.98	1.66	221.49	18.09
triangl:22:1:50	34.29	33.93	110.86	1.91	107.15	8.52
tak:40:20:11:1	24.78	12.93	55.19	1.66	134.21	4.76
takl:40:20:12:1	74.92	20.97		3.71		9.46
ntakl:40:20:12:1	60.71	17.07	95.98	3.65		9.52
cpstak:40:20:11:1	49.13	103.36	222.13	4.21	258.62	59.44
ctak:32:16:8:1	101.5	44.14		0.96
fib:40:5	46.50	10.22	96.25	3.63	236.65	12.09
fibc:30:10	64.58	25.80		0.71
fibfp:35.0:10	45.34	1.89	42.79	3.18	56.26	22.00
sum:10000:200000	55.43	6.64	99.79	3.59		6.87
sumfp:1000000.0:500	32.92	2.50	85.36	4.25	111.78	42.06
fft:65536:100	36.87	32.20	69.76	3.32		7.69
mbrot:75:1000		24.40	209.51	5.39		50.09
mbrotZ:75:1000		18.56		9.26		67.01
nucleic:50	50.83	19.95	79.05	2.59	69.32	15.35
pi		–		0.60		0.56
pnpoly:1000000	107.00	17.98	253.84	4.73		24.89
ray:50	27.11	20.46	119.63	2.83		18.51
simplex:1000000	85.77	46.34	182.76	2.34		13.90
ack:3:12:2	44.28	10.57	74.51	3.12		8.41
array1:1000000:500	59.10	11.48	64.18	8.16	138.45	9.24
string:500000:100	6.30	1.71	6.34	6.54	1.81	1.87
sum1:25	0.73	0.47	121.78	2.02	1.71	4.43
cat:50	28.36	1.19	70.95	2.69		28.40
tail:50	1.08	1.19	11.21	3.57		9.82
wc:inputs/bib:50	39.55	8.27		1.76	73.34	16.96
read1:2500	1.53	0.41	281.23	1.28	2.69	5.80
compiler:2000		41.16	115.83	2.99		5.15
conform:500	47.04	51.03	199.42	2.10		10.51
dynamic:500		22.74	288.25	4.01	71.60	7.37
earley	35.82			5.03		9.49
graphs:7:3	99.22	127.61		2.27		23.03
lattice:44:10	135.43	139.28		3.91		15.94
matrix:5:5:2500	45.10	72.07	214.16	1.63		9.88
maze:20:7:10000		23.26	84.15	1.66		4.70
mazefun:11:11:10000	28.50	19.51	122.02	2.86	128.66	9.66
nqueens:13:10	63.33	55.11	165.25	4.99		19.37
paraffins:23:10	6.12	31.42		5.97		4.25
parsing:2500		39.44		3.35		10.69
peval:2000	32.29	29.68	177.10	2.05	107.05	15.64
primes:1000:10000	13.52	7.73	21.08	2.64	43.73	7.52
quicksort:10000:2500	119.74	94.00		3.90		13.25
scheme:100000		71.46	154.36	2.55		15.14
slatex:500		32.07	173.60	3.73	43.82	45.05
chudnovsky		–		0.32		0.31
nboyer:5:1	12.48	39.27	41.54	3.95	142.86	5.10
sboyer:5:1	15.42	31.54	39.14	1.30	155.49	4.76
gcbench:20:1		20.54		1.87		3.51
mperm:20:10:2:1		173.33	659.14	13.33		10.65
equal:100:8:1000:2000:5000	3.61	0.78	54.11	0.99
bv2string:1000:1000:100	10.21	10.78	11.17	2.47		4.49