[Scummvm-devel] Code size

[Max Horn]

Hi folks

we recently talked about the size of a scummvm executable and how it  
grew with the merger of the GUI branch. I guess that is partially due  
to the fact that we now have an XML parser, too... ;) But I figured it  
might be helpful to generate some more data to figure out where we  
might be able to gain the most.

What I did was to compile a "release" binary for both trunk and the  
0.12.0 branch. So I run
   configure --enable-Werror --enable-release [ lots of --enable- 
ENGINE statements ]
and made fresh builds for both, on an Apple MacBook Pro, with Apple's  
modified GCC 4.0.1. Our "--enable-release" switch ensures that -O2 is  
passed to the compiler.

I then sorted all *.o files by size of their TEXT segment. Since  
engine code and infrastructure are kind of separate, I did this  
separately independently on both. (I took the liberty of re-inserting  
the column headers below for readability)

TRUNK rev 34812, non-engine code:

$ size */*.a | sort -nr | head
__TEXT	__DATA	__OBJC	others	dec	hex
87543	1180	0	0	88723	15a93	graphics/libgraphics.a(hq3x_i386.o)
65445	1204	0	0	66649	10459	graphics/libgraphics.a(hq2x_i386.o)
65016	1072	0	566	66654	1045e	gui/libgui.a(ThemeEngine.o)
54833	144	0	607	55584	d920	gui/libgui.a(ThemeParser.o)
46803	552	0	844	48199	bc47	gui/libgui.a(launcher.o)
46585	424	0	71	47080	b7e8	graphics/libgraphics.a(VectorRendererSpec.o)
39371	36	0	364	39771	9b5b	base/libbase.a(commandLine.o)
32745	308	0	405	33458	82b2	gui/libgui.a(options.o)
28307	1700	0	354	30361	7699	sound/libsound.a(audiostream.o)
22409	140	0	536	23085	5a2d	common/libcommon.a(advancedDetector.o)

$ ll -S */*.a
-rw-r--r-- 1 mhorn staff 736892 16. Okt 18:56 gui/libgui.a
-rw-r--r-- 1 mhorn staff 607148 16. Okt 18:56 graphics/libgraphics.a
-rw-r--r-- 1 mhorn staff 419884 16. Okt 18:56 sound/libsound.a
-rw-r--r-- 1 mhorn staff 365740 16. Okt 18:56 common/libcommon.a
-rw-r--r-- 1 mhorn staff 133636 16. Okt 18:56 backends/libbackends.a
-rw-r--r-- 1 mhorn staff 110956 16. Okt 18:56 base/libbase.a
-rw-r--r-- 1 mhorn staff  45876 16. Okt 18:56 engines/libengines.a


0.12.x BRANCH , non-engine code:

$ size */*.a | sort -nr | head
__TEXT	__DATA	__OBJC	others	dec	hex
87543	1180	0	0	88723	15a93	graphics/libgraphics.a(hq3x_i386.o)
65445	1204	0	0	66649	10459	graphics/libgraphics.a(hq2x_i386.o)
51949	160	0	370	52479	ccff	gui/libgui.a(ThemeModern.o)
38997	36	0	374	39407	99ef	base/libbase.a(commandLine.o)
32101	424	0	710	33235	81d3	gui/libgui.a(launcher.o)
29182	296	0	341	29819	747b	gui/libgui.a(options.o)
28307	1700	0	354	30361	7699	sound/libsound.a(audiostream.o)
27710	84	0	269	28063	6d9f	gui/libgui.a(theme-config.o)
22040	120	0	462	22622	585e	common/libcommon.a(advancedDetector.o)
20945	160	0	249	21354	536a	gui/libgui.a(ThemeClassic.o)

$ ll -S */*.a
-rw-r--r-- 1 mhorn staff 594716 16. Okt 19:16 gui/libgui.a
-rw-r--r-- 1 mhorn staff 524604 16. Okt 19:16 graphics/libgraphics.a
-rw-r--r-- 1 mhorn staff 416684 16. Okt 19:16 sound/libsound.a
-rw-r--r-- 1 mhorn staff 267020 16. Okt 19:16 common/libcommon.a
-rw-r--r-- 1 mhorn staff 146228 16. Okt 19:16 backends/libbackends.a
-rw-r--r-- 1 mhorn staff 118244 16. Okt 19:16 base/libbase.a
-rw-r--r-- 1 mhorn staff   8608 16. Okt 19:16 engines/libengines.a


So, the HQx scalers are big, but they are desktop-only, so no problem.  
But then, a very sizable chunk is taken by the new GUI code. The old  
code was not tiny either, but it's clear that libgui.a and  
libgraphis.a grew considerably (also libcommon.a grew, partially due  
to the XML parser, but other changes are in there as well). It sure  
would be nice to get those number back down a bit...

However, take all these numbers with a big grain of salt: We use lots  
of template code these days, and that causes the compiler to generate  
certain chunks of code multiple times, put into each *.o file  
separately. Only during linking time are these duplicates resolved and  
removed. So, the actual contribution of those *.o files to the  
resulting binary might be a lot less than what is listed above.

On the other hand, this immediately hints at a potential way to save  
memory: By carefully reconsidering how we use templates; esp. for the  
container classes, we might be able to save some KB of duplicate code  
(duplicate because the compiler generates it anew for each e.g.  
HashMap variant, even though the code each time is almost or  
completely identical). There are some ways to reduce that. If anybody  
is interested in looking into that, talk to me.


Here's what I got for engines (I enabled all engines):

TRUNK rev 34812, non-engine code:

$ size engines/*/*.a | sort -nr | head
__TEXT	__DATA	__OBJC	others	dec	hex
98802	2924	0	546	102272	18f80	engines/kyra/libkyra.a(staticres.o)
77058	64	0	553	77675	12f6b	engines/lure/liblure.a(hotspots.o)
55721	244	0	351	56316	dbfc	engines/touche/libtouche.a(touche.o)
54580	2112	0	333	57025	dec1	engines/kyra/libkyra.a(sound_towns.o)
53776	29904	0	613	84293	14945	engines/scumm/libscumm.a(detection.o)
51834	48	0	532	52414	ccbe	engines/kyra/libkyra.a(script_lok.o)
51821	44	0	346	52211	cbf3	engines/igor/libigor.a(igor.o)
51025	96	0	719	51840	ca80	engines/kyra/libkyra.a(script_hof.o)
50265	0	0	320	50585	c599	engines/drascula/libdrascula.a(animation.o)
48901	256	0	509	49666	c202	engines/drascula/libdrascula.a(rooms.o)

To put this into perspective, here's the size of the lib*.a files for  
each engine (to be taken with even more grains of salt, of course, but  
they give a good rough indication anyway):
$ ll -S engines/*/*.a
-rw-r--r-- 1 mhorn staff 2661204 16. Okt 18:51 engines/scumm/libscumm.a
-rw-r--r-- 1 mhorn staff 1960012 16. Okt 18:53 engines/kyra/libkyra.a
-rw-r--r-- 1 mhorn staff 1027996 16. Okt 18:52 engines/gob/libgob.a
-rw-r--r-- 1 mhorn staff  877820 16. Okt 18:54 engines/parallaction/ 
libparallaction.a
-rw-r--r-- 1 mhorn staff  823140 16. Okt 18:51 engines/agos/libagos.a
-rw-r--r-- 1 mhorn staff  740188 16. Okt 18:55 engines/saga/libsaga.a

[...]
-rw-r--r-- 1 mhorn staff  238412 16. Okt 18:54 engines/made/libmade.a

0.12.x BRANCH , non-engine code:

$ size engines/*/*.a | sort -nr | head
92384	3040	0	531	95955	176d3	engines/kyra/libkyra.a(staticres.o)
77074	64	0	553	77691	12f7b	engines/lure/liblure.a(hotspots.o)
57233	0	0	345	57578	e0ea	engines/drascula/libdrascula.a(animation.o)
55273	236	0	341	55850	da2a	engines/touche/libtouche.a(touche.o)
54956	2368	0	353	57677	e14d	engines/kyra/libkyra.a(sound_towns.o)
51882	48	0	532	52462	ccee	engines/kyra/libkyra.a(script_lok.o)
51854	36	0	346	52236	cc0c	engines/igor/libigor.a(igor.o)
51487	29832	0	558	81877	13fd5	engines/scumm/libscumm.a(detection.o)
50993	96	0	719	51808	ca60	engines/kyra/libkyra.a(script_hof.o)
49606	40	0	554	50200	c418	engines/drascula/libdrascula.a(rooms.o)

$ ll -S engines/*/*.a
-rw-r--r-- 1 mhorn staff 2690660 16. Okt 19:16 engines/scumm/libscumm.a
-rw-r--r-- 1 mhorn staff 1861188 16. Okt 19:16 engines/kyra/libkyra.a
-rw-r--r-- 1 mhorn staff  948060 16. Okt 19:16 engines/gob/libgob.a
-rw-r--r-- 1 mhorn staff  803404 16. Okt 19:16 engines/agos/libagos.a
-rw-r--r-- 1 mhorn staff  711500 16. Okt 19:16 engines/saga/libsaga.a
-rw-r--r-- 1 mhorn staff  696756 16. Okt 19:16 engines/parallaction/ 
libparallaction.a
[...]
-rw-r--r-- 1 mhorn staff  232492 16. Okt 19:16 engines/made/libmade.a

So it seems most engines got a bit bigger, somehow... This might be  
due to changes to our templated code in common, not sure yet.

Cheers,
Max