Documentation lies

For various reasons, I want to get the extent of a particular expression in clang. Most AST objects in clang have a method along the lines of getSourceRange(), which, on inspection, returns start and end locations for the object in question. Naturally, it's the perfect fit (more or less), so I start using it. However, it turns out that the documentation is a bit of a liar. I would expect the end location to return the location (either file offset or file/line) of the end of the expression, give or take a character depending on whether people prefer half-open or fully closed ranges. Instead, it returns the location of the last token in the expression. Getting the true last token requires this mess: sm.getFileOffset(Lexer::getLocForEndOfToken(end, 0, sm, features)). Oh yeah, in case you couldn't have guessed, that causes it to relex the file starting at that point. For an idea of my current mood, pick a random image from Jennifer's recent post and you'll likely capture it.

DXR updates

It's been almost a week since I last discussed DXR, but, if you look at the commit log, you can see that I've not exactly been laying back and doing nothing. No, the main reason is because most of the changes I've been doing so far aren't exactly ground-breaking.

In terms of UI, the only thing that's changed is that I know actually link the variable declarations correctly, a result of discovering the typo that was causing it not to work properly in the midst of the other changes. From a user's perspective, I've also radically altered the invocation of DXR. Whereas before it involved a long series of steps consisting of "build here, export these variables, run this script, build there, run that script, then run this file, and I hope you did everything correctly otherwise you'll wait half an hour to find you didn't" (most of which takes place in different directories), it's now down to . dxr/setup-env.sh (if you invoke that incorrectly, you get a loud error message telling you how to do it correctly. Unfortunately, the nature of shells prohibits me from just doing the right thing in the first place), build your code, and then python dxr/dxr-index.py in the directory with the proper dxr.config. Vastly improved, then. But most of my changes are merely in cleaning up the code.

The first major change I made was replacing the libclang API with a real clang plugin. There's only one thing I've found harder (figuring out inheritance, not that I'm sure I've had it right in the first place), and there are a lot of things I've found easier. Given how crazy people get with build systems, latching onto a compiler makes things go a lot smoother. The biggest downside I've seen with clang is that it's documentation is lacking. RecursiveASTVisitor is the main visitor API I'm using, but the documentation for that class is complete and utter crap, a result of doxygen failing to handle the hurdle of comprehending macros. I ended up using the libindex-based dxr to dump a database of all of the functions in the class, of which there are something like 1289, or around 60 times the functions listed in the documentation. Another example is Decl, where the inheritance picture is actually helpful. It, however, manages to have failed to document DeclCXX.h, which is a rather glaring omission if what you are working with is C++ source.

The last set of changes I did was rearchitecting the code to make it hackable by other people. I have started on a basic pluggable architecture for actually implementing new language modules, although most of the information is still hardcoded to just use cxx-clang. In addition, I've begun work on ripping out SQL as the exchange medium of choice: the sidebar list is now directly generated using the post-processing ball of information, and linkification is now set up so that it can be generated independent of tokenization. In the process, I've greatly sped up HTML generation times only to regress a lot of it by tokenizing the file twice (the latter part will unfortunately remain until I get around to changing tokenization API). It shouldn't take that much longer for me to rip SQL fully out of the HTML builder and shove SQL generation into a parallel process for improved end-to-end time.

Some things I've discovered about python along the way. First, python closures don't let you mutate closed-over-variables. That breaks things. Second, if you have a very large python object, don't pass it around as an argument to multiprocessing: just make it a global and let the kernel's copy-on-write code make cloning cheap. Third, apparently the fastest way to concatenate strings in python is to use join with a list comprehension. Finally, python dynamic module importing sucks big time.

Where am I going from here? I'll have HTMLification (i.e., remove SQL queries and replace with ball lookups) fixed up tomorrow; after that, I'll make the cgi scripts work again. The next major change after that is getting inheritance and references in good shape, and then making sure that I can do namespaces correctly. At that point in time, I'll think I'll be able to make a prerelease of DXR.

Fakeserver and future changes

This is perhaps a bit belated, but I thought I'd give a brief overview of where I think fakeserver should be going over the next several months (i.e., when I find time to work on it). For those who don't know, Fakeserver is a generalized testing framework for the 4 major mail protocols: IMAP, NNTP, SMTP, and POP. For various reasons, I'm actually getting around to fixing problems with it right now. The following are the features I am going to be adding:

Multiple connections

This is a long-standing bug in fakeserver, which I first discovered about 3 years ago, when I started testing the IMAP implementation. Fakeserver uses the same connection handler for every connection, which means the state of the connection is shared between all connections, obviously very problematic for a protocol as stateful as IMAP. In my defense, I cribbed from the stateless (more or less) HTTP server, and used the barely-stateful NNTP as my first testbed. This bug is probably the most API-breaking of any change to fakeserver code: every invocation of the server must replace new nsMailServer(new handler(daemon)) with new nsMailServer(function (d) { return new handler(d); }, daemon). Subsequent to that, handlers are no longer easily exposed by the server, which means all manipulation must be on the daemon end. This causes major changes to SMTP code and NNTP code for different reasons; pretty much everyone else hides between helper functions that act as a firewall to minimizing code changes.

Multi-process fakeserver

This change (which has no patch yet) will move the fakeserver running into another process. I have some working prototypes for manipulating objects across two different processes, but I don't yet have a good IPC mechanism (which Mozilla strangely lacks) for actually communicating this information. As for API changes, I haven't gotten far enough yet to know the impact of multiprocess fakeserver, but the answer will probably that handler information goes from "hard to get" to "impossible to get". On the plus side, this should essentially enable mozmill tests to be able to use fakeserver.

SSL fakeserver

I figure fakeserver needs SSL support--it's common in the mail world to use SSL instead of plaintext, so we need to support it. In terms of external API, the only change will be some added methods to start up an SSL server (and methods to get STARTTLS working to, provided my knowledge of how SSL works is correct). The API pseudo-internal to maild.js gets more convoluted, though: essentially, I need to support multiple pumping routines for communication. On the plus side, though, if the API I'm planning on implementing turns out to be feasible, we should also be able to get generic charset decoding for nearly free, for i18n tests.

LDAP fakeserver

LDAP address book code is almost completely untested. I'd like to see that fixed. However, LDAP is a binary protocol (BER, to be precise). If I can implement the SSL support in the API framework I want to, then it shouldn't be that much more to get a BER-to-text-ish layer thrown in on top of it. The downside is that I'm pretty much looking at using our own LDAP library to do BER encoding/decoding since I don't want to write that myself.

DXR thinking aloud

A primary goal of DXR is to be easy to use. Like any simple statement, translating that into design decisions is inordinately difficult, and I best approach such issues by thinking out loud. Normally, I do this in IRC channels, but today I think it would be best to think in a louder medium, since the problem is harder.

There are three distinct things that need to be made "easy to use". The first is the generation of the database and the subsequent creation of the web interface. The second is extension of DXR to new languages, while the last is the customization of DXR to provide more information. All of them have significant issues.

Building with DXR

Starting with the first one, build systems are complicated. For a simple GNU-ish utility, ./configure && make is a functioning build system. But where DXR is most useful is on the large, hydra-like projects where figuring out how to build the program is itself a nightmare: Mozilla, OpenJDK, Eclipse, etc. There is also a substantial number of non-autoconf based systems which throw great wrenches in everything. At the very least, I know this much about DXR: I need to set environment variables before configuring and building (i.e., replace the compiler), I need to "watch" the build process (i.e., follow warning spew), and I need to do things after the build finishes (post-processing). Potential options:

1. Tell the user to run this command before the build and after the build. On the plus side, this means that DXR needs to know absolutely nothing about how the build system works. On the down side, this requires confusing instructions: in particular, since I'm setting environment variables, the user has to specifically type ". " in the shell to get them set up properly. There are a lot of people who do not have significant shell exposure to actually understand why that is necessary, and general usage is different enough from the commands that people are liable to make mistakes doing so.
2. Guess what the build system looks like and try to do it all by ourselves. This is pretty much the opposite extreme, in that it foists all the work on DXR. If your program is "normal", this won't be a problem. If your program isn't... it will be a world of pain. Take also into consideration that any automated approach is likely to fail hard on Mozilla code to begin with, which effectively makes this a non-starter.
3. Have the user input their build system to a configuration file and go from there. A step down from the previous item, but it increases the need for configuration files.
4. Have DXR spawn a shell for the build system. Intriguing, solves some problems but causes others.

Conclusion: well, I don't like any of those options. While the goal of essentially being able to "click" DXR and have it Just Work™ is nice, I have reservations about such an approach being able to work in practice. I think I'll go for a "#1 and punt on this issue to someone with more experience."

Multiple language

I could devote an entire week's worth of blog posts to this topic, I think, and I would wager that this is more complicated and nebulous than even build systems are. In the end, all we really need to worry about with build systems is replacing compilers with our versions and getting to the end; with languages, we actually need to be very introspective and invasive to do our job.

Probably the best place to start is actually laying out what needs to be done. If the end goal is to produce the source viewer, then we need to at least be able to do syntax highlighting. That by itself is difficult, but people have done it before: I think my gut preference at this point is to basically ask authors of DXR plugins to expose something akin to vim's syntax highlighting instead of asking them to write a full lexer for their language.

On the other end of the spectrum is generating the database. The idea is to use an instrumenting compiler, but while that works for C++ or Java, someone whose primary code is a website in Perl or a large Python utility has a hard time writing a compiler. Perhaps the best option here is just parsing the source code when we walk the tree. There is also the question about what to do with the build system: surely people might want help understanding what it is their Makefile.in is really doing.

So what does the database look like? For standard programming languages, we appear to have a wide-ranging and clear notion of types/classes, functions, and variables, with slightly vaguer notions of inheritance, macros (in both the lexical preprocessing sense and the type-based sense of C++'s templates), and visibility. Dynamic languages like JavaScript or Python might lack some reliable information (e.g., variables don't have types, although people often still act as if they have implicit type information), but they still uphold this general contract. If you consider instead things like CSS and HTML or Makefiles in the build system, this general scheme completely fails to hold, but you can still desire information in the database: for example, it would help to be able to pinpoint which CSS rules apply to a given HTML element.

This begs the question, how does one handle multiple languages in the database? As I ponder this, I realize that there are multiple domains of knowledge: what is available in one language is not necessarily available in another. Of the languages Mozilla uses, C, assembly, C++, and Objective-C[++] all share the same ability to access any information written in the other languages; contrast this to JS code, which can only interact with native code via the use of a subset of IDL or interactions with native functions. IDL is a third space, which is a middle ground between native and JS code, but is insufficiently compatible with either to be lumped in with one. Options:

1. Dump each language into the same tables with no distinction. This has problems in so far as some languages can't be shoehorned into the same models, but I think that in such cases, one is probably looking for different enough information anyways that it doesn't matter. The advantage of this is that searching for an identifier will bring it up everywhere. The disadvantage... is that it gets brought up everywhere.
2. Similar to #1, but make an extra column for language, and let people filter by language.
3. Going a step further, take the extra language information and build up the notion of different bindings: this foo.bar method on a python object may be implemented by this Python_foo_bar C binding. In other words, add another table which lists this cross-pollination and takes it into account when searching or providing detailed information<.
4. Instead of the language column in tables, make different tables for every language.
5. Instead of tables, use databases?

Hmm. I think the binding cross-reference is important. On closer thought, it's not really languages themselves that are the issue here, it's essentially the target bindings: if we have a system that is doing non-trivial build system work that involves cross-compiling, it matters if what we are doing is being done for the host or being done for the target. Apart from that, I think right now that the best approach is to have different tables.

Extraneous information

The previous discussion bleeds into this final one, since they both ultimately concern themselves with one thing: the database. This time, the question is how to handle generation of information beyond the "standard" set of information. Information, as I see it, comes in a few forms. There is additional information at the granularity of identifiers (this function consumes 234 bytes of space or this is the documentation for the function), lines (this line was not executed in the test suite), files (this file gets compiled to this binary library), and arguably directories or other concepts not totally mappable to constructs in the source tree (e.g., output libraries).

The main question here is not on the design of the database: it's only a question of extra tables or extra columns (or both!). Instead, the real question is in the design of the programmatic mechanisms. In dxr+dehydra, the simple answer is to load multiple scripts. For dxr+clang, however, the question becomes a lot more difficult since the code is written in C++ and isn't dynamically loading modules like dehydra does. It also begins to beg the question of the exposed API. On the other hand, I'm not sure I know enough of the problem space to be able to actually come up with solutions. I think I'll leave this one for later

Who's responsible for libxul

As you might have gathered from my last post, I spend a fair amount of time thinking up ways of visualizing software metrics. Today's visualization is inspired by a lunch conversation I had yesterday. Someone asked "how much space does SVG take up in libxul?" This isn't a terribly hard problem to solve: we have a tool that can tell us the size of symbols (i.e., codesighs), and I have a tool that gives me a database of where every symbol is defined in source code.

Actually, it's not so simple. codesighs, as I might have predicted, doesn't appear to work that well, so I regressed back to nm. And when I actually poked the generated DXR database for mozilla-central using dehydra (I haven't had the guts to try my dxr-clang integration on m-c yet), I discovered that it appears to only lack the information about functions. Both static functions and member C++ functions—my suspicions of breakage are now confirmed. So I resorted to one of my "one-liner" scripts, as follows:

cat <<HEREDOC
<!DOCTYPE html>
<html><head>
<title>Code size by file</title>
<script type="application/javascript" src="https://www.google.com/jsapi"></script>
<script type="application/javascript">
google.load("visualization", "1", {packages: ["treemap"]});
google.setOnLoadCallback(drawChart);
function drawChart() {
  var data = new google.visualization.DataTable();
  data.addColumn('string', 'File');
  data.addColumn('string', 'Path');
  data.addColumn('number', 'Code size (B)');
HEREDOC
nm --format=bsd --size-sort /src/build/trunk/browser/dist/bin/libxul.so | cut -d' ' -f 3 > /tmp/libxul.symbols.txt
find -name "*.o"| xargs nm --format=bsd --size-sort --print-file-name | sed -e 's/:/ /g' | python -c '
import sys
symbols = set(open("/tmp/libxul.symbols.txt").readlines())
files = {}
for line in sys.stdin:
  args = line.split(" ")
  if args[-2] in "tTvVwW" and args[-1] in symbols:
    files[args[0]] = files.get(args[0], 0) + int(args[1], 16)
out_lines = set()
for f in files:
  f2 = f.replace("./", "")
  path = f2.split("/")
  str = "src/"
  for p in path[:-1]:
    out_lines.add((str + p + "/", str, 0))
    str += p + "/"
  out_lines.add((f2, str, files[f]))
print "data.addRows([" + ",\n".join([repr(list(x)) for x in out_lines]) + "]);"
'

cat <<HEREDOC
  data.addRow(["src/", null, 0]);
  var tree = new google.visualization.TreeMap(document.getElementById('tmap'));
  tree.draw(data, {maxDepth: 3, headerHeight: 0, noColor: "#ddd", maxColor: "#ddd",
    minColor: "#ddd", midColor: "#ddd"});
}
</script></head>
<body>
  <div id="tmap" style="width: 100%; height: 1000px;"></div>
</body>
</html>
HEREDOC

It's really 4 commands, which just happens to include two HEREDOC cats and a multi-line python script. The output is a littlenot-so-little HTML file that loads a treemap visualization using the google API. The end result can be seen here. To answer the question posed yesterday: SVG accounts for about 5-6% of libxul, in terms of content, and about another 1% for the layout component; altogether, about as much as our xpconnect code accounts for libxul. For Thunderbird folks, I've taken the time to also similarly divvy up libxul, and that can be found here.

P.S., if you're complaining about the lack of hard numbers, blame Google for giving me a crappy API that doesn't allow me to make custom tooltips.

Visualization toolkits suck

This is a rant in prelude to a blog post I expect to make tomorrow morning (or later today, depending on your current timezone). As someone who is mildly interested in data, I spend a fair amount of time thinking up things that would be interesting to just see dumped out visually. And I'm not particularly interested in small datasets: one of my smaller datasets I've been playing with is "every symbol in libxul.so"; one of my larger datasets is "every news message posted to Usenet this year, excluding binaries" (note: this is 13 GiB worth of data, and that's not 100% complete).

So while I have a fair amount of data, what I need is a simple way to visualize it. If what I have is a simple bar chart or scatter plot, it's possible to put it in Excel or LibreOffice... only to watch those programs choke on a mere few thousand data points (inputting a 120K file caused LibreOffice to hang for about 5 minutes to produce the scatter plot). But spreadsheet programs clearly lack the power for serious visualization information; the most glaring chart they lack is the box-and-whiskers plot. Of course, if the data I'm looking at isn't simple 1D or 2D data, then what I really need won't be satiable with them anyways.

I also need to visualize large tree data, typically in a tree map. Since the code for making squarified tree maps is more than I care to do for a simple vis project, I'd rather just use a simple toolkit. But which to use? Java-based toolkits (i.e., prefuse) require me to sit down and make a full-fledged application for what should be a quick data visualization, and I don't know any Flash to be able to use Flash toolkits (i.e., flare). For JavaScript, I've tried the JavaScript InfoVis Toolkit, Google's toolkit, and protovis, all without much luck. JIT and protovis both require too much baggage for a simple "what does this data look like", and Google's API is too inflexible to do anything more than "ooh, pretty treemap". Hence why my previous foray into an application for viewing code coverage used a Java applet: it was the only thing I could get working.

What I want is a toolkit that gracefully supports large datasets, allows me to easily drop data in and play with views (preferably one that doesn't try to dynamically change the display based on partial options reconfiguration like most office suites attempt), supports a variety of datasets, and has a fine-grained level of customizability. Kind of like SpotFire or Tableau, just a tad bit more in my price range. Ideally, it would also be easy to put on the web, too, although supporting crappy old IE versions isn't a major feature I need. Is that really too much to ask for?

Building with dxr-clang

As I'm building yet another program from scratch (I've now run dxr on four separate trees in the past week, and I'm still looking for a good one to use for tests), let me take the time to explain how to set up DXR on your own computer. Instead of stuffing it in this blogpost, most of the information is instead on wiki.mozilla.org. One thing not on that page is that I currently have the sql files in the clang dumper dump to the source directory instead of the object directory, so the DXR config needs to have the source and object directory be the same setting, even if you don't actually build the source in-place!