Cracking libxul

All timestamps are based on your local time of:

Posted by: stak
Tags: mozilla
Posted on: 2014-05-22 09:02:20

For a while now I've been wanting to take a look inside libxul to see why it's so big. In particular I wanted to know what the impact of using templates so heavily in our code was - things like nsTArray and nsRefPtr are probably used on hundreds of different types throughout our codebase. Last night I was have trouble sleeping so I decided to crack open libxul and see if I could figure it out. I didn't persist enough to get the exact answers I wanted, but I got close enough. It was also kind of fun and I figured I'd post about it partly as an educational thing and partly to inspire others to dig deeper into this.

First step: build libxul. I had a debug build on my Linux machine with recent gecko, so I just used the from that.

Second step: disassemble libxul.

objdump -d > libxul.disasm

Although I've looked at disassemblies before I had to look at the file in vim a little bit to figure the best way to parse it to get what I wanted, which was the size of every function defined in the library. This turned out to be a fairly simple awk script.

Third step: get function sizes. (snippet below is reformatted for easier reading)

awk 'BEGIN { addr=0; label="";}
     /:$/ && !/Disassembly of section/ { naddr = sprintf("%d", "0x" $1);
                                         print (naddr-addr), label;
                                         label=$2 }'
    libxul.disasm > libxul.sizes

For those of you unfamiliar with awk, this identifies every line that ends in a colon, but doesn't have the text "Disassembly of section" (I determined this would be sufficient to match the line that starts off every function disassembly). It then takes the address (which is in hex in the dump), converts it to decimal, and subtracts it from the address of the previous matching line. Finally it dumps out the size/name pairs. I inspected the file to make sure it looked ok, and removed a bad line at the top of the file (easier to fix it manually than fix the awk script).

Now that I had the size of each function, I did a quick sanity check to make sure it added up to a reasonable number:

awk '{ total += $1 } END { print total }' libxul.sizes

The value spit out is around 40 megs. This seemed to be in the right order of magnitude for code in libxul so I proceeded further.

Fourth step: see what's biggest!

sort -rn libxul.sizes | head -n 20
57984 <_ZL9InterpretP9JSContextRN2js8RunStateE>:
43798 <_ZN20nsHtml5AttributeName17initializeStaticsEv>:
41614 <_ZN22nsWindowMemoryReporter14CollectReportsEP25nsIMemoryReporterCallbackP11nsISupports>:
39792 <_Z7JS_Initv>:
32722 <vp9_fdct32x32_sse2>:
28674 <encode_mcu_huff>:
24365 <_Z7yyparseP13TParseContext>:
21800 <_ZN18nsHtml5ElementName17initializeStaticsEv>:
20558 <_ZN7mozilla3dom14PContentParent17OnMessageReceivedERKN3IPC7MessageE.part.1247>:
20302 <_ZN16nsHtml5Tokenizer9stateLoopI23nsHtml5ViewSourcePolicyEEiiDsiPDsbii>:
18367 <sctp_setopt>:
17900 <vp9_find_best_sub_pixel_comp_tree>:
16952 <_ZN7mozilla3dom13PBrowserChild17OnMessageReceivedERKN3IPC7MessageE>:
16096 <vp9_sad64x64x4d_sse2>:
15996 <_ZN7mozilla12_GLOBAL__N_119WebGLImageConverter3runILNS_16WebGLTexelFormatE17EEEvS3_NS_29WebGLTexelPremultiplicationOpE>:
15594 <_ZN7mozilla12_GLOBAL__N_119WebGLImageConverter3runILNS_16WebGLTexelFormatE16EEEvS3_NS_29WebGLTexelPremultiplicationOpE>:
14963 <vp9_idct32x32_1024_add_sse2>:
14838 <_ZN7mozilla12_GLOBAL__N_119WebGLImageConverter3runILNS_16WebGLTexelFormatE4EEEvS3_NS_29WebGLTexelPremultiplicationOpE>:
14792 <_ZN7mozilla12_GLOBAL__N_119WebGLImageConverter3runILNS_16WebGLTexelFormatE21EEEvS3_NS_29WebGLTexelPremultiplicationOpE>:
14740 <_ZN16nsHtml5Tokenizer9stateLoopI19nsHtml5SilentPolicyEEiiDsiPDsbii>:

That output looks reasonable. Top of the list is something to do with interpreting JS, followed by some HTML name static initializer thing. Guessing from the symbol names it seems like everything there would be pretty big. So far so good.

Fifth step: see how much space nsTArray takes up. As you can see above, the function names in the disassembly are mangled, and while I could spend some time trying to figure out how to demangle them it didn't seem particularly worth the time. Instead I just looked for symbols that started with nsTArray_Impl which by visual inspection seemed to match what I was looking for, and would at least give me a ballpark figure.

grep "<_ZN13nsTArray_Impl" libxul.sizes | awk '{ total += $1 } END { print total }'

That's around 377k of stuff just to deal with nsTArray_Impl functions. You can compare that to the total libxul number and the largest functions listed above to get a sense of how much that is. I did the same for nsRefPtr and got 92k. Looking for ZNSt6vector, which I presume is the std::vector class, returned 101k.

That more or less answered the questions I had and gave me an idea of how much space was being used by a particular template class. I tried a few more things like grouping by the first 20 characters of the function name and summing up the sizes, but it didn't give particularly useful results. I had hoped it would approximate the total size taken up by each class but because of the variability in name lengths I would really need a demangler before being able to get that.

Posted by glandium at 2014-05-22 10:14:12
To get symbol sizes, instead of starting from the disassembly, which is huge, you could just start from the output of:
objdump -T
readelf -s
[ Reply to this ]
Posted by stak at 2014-05-22 10:41:37
Ah, cool. The disassembly was actually around the same size as the debug-build (~680 MB). awk processes it pretty quickly on my machine.

But yeah, readelf -s seems nice, and seems to give around the same results for nsTArray (~376k).
[ Reply to this ]
Posted by pao at 2014-05-22 15:02:47
For demangling, you should be able to just run your text through c++filt.
[ Reply to this ]
Posted by glandium at 2014-05-22 18:16:29
Or ask objdump to do it for you :)
[ Reply to this ]
Posted by njn at 2014-05-22 20:27:14
Nathan Froyd has looked at this sort of thing in the recent past. You two could have some scintillating conversations :P
[ Reply to this ]
Posted by Neil Rashbrook at 2014-05-23 06:23:39
nsHTML5AttributeName::initializeStatics makes me a little sad. It's code translated to C++ from Java, so it's being hurt by a level of abstraction that wouldn't normally exist in C++.
Allowed expansions in comments/replies: [i]italic[/i], [u]underline[/u], [b]bold[/b], [code]code[/code], [sub]subscript[/sub], [sup]superscript[/sup], [url=http://some.url]linked text[/url]
Human verification: Sum of thirty-three and thirty-one =

[ Add a new comment ]

(c) Kartikaya Gupta, 2004-2024. User comments owned by their respective posters. All rights reserved.
You are accessing this website via IPv4. Consider upgrading to IPv6!