more on ghc filename encodings

My last post missed an important thing about GHC 7.4's handling of encodings for FileName. It can in fact be safe to use FilePath to write a command like rm. This is because GHC internally uses a special encoding for FilePath data, that is documented to allow "arbitrary undecodable bytes to be round-tripped through it". (It seems to do this by encoding the undecodable bytes as very high unicode code points.) So, when presented with a filename that cannot be decoded using utf-8 (or whatever the system encoding is), it still handles it, and using the resulting FilePath will in fact operate on the right file. Whew!

Moral of the story is that if you're going to be using GHC 7.4 to read or write filenames from a pipe, or a file, you need to arrange for the Handle you're reading or writing to use this special encoding too. I use this to set up my Handles:

  import System.IO
import GHC.IO.Encoding
import GHC.IO.Handle

fileEncoding :: Handle -> IO ()
fileEncoding h = hSetEncoding h =<< getFileSystemEncoding

Even if you're only going to write a FilePath to stdout, you need to do this. Otherwise, your program will crash on some filenames! This doesn't seem quite right to me, but I hesitate to file a bug report. (And this is not a new problem in GHC anyway.) If I did, it would have this testcase:

  # touch "me¡"
# LANG=C ghc
Prelude> :m System.Directory
Prelude System.Directory> mapM_ putStrLn =<< getDirectoryContents "."
me*** Exception: <stdout>: hPutChar: invalid argument (invalid character)

Since git-annex reads lots of filenames from git commands and other places, I had to deal with this extensively. Unfortunatly I have not found a way to read Text from a Handle using the fileSystemEncoding. So I'm stuck with slow Strings. But, it does seem to work now.

PS: I found a bug in GHC 7.4 today where one of those famous Haskell immutable values seems to get well, mutated. Specifically a [FilePath] that is non-empty at the top of a function ends up empty at the bottom. Unless IO is done involving it at the top. Really. Hope to develop a test case soon. Happily, the code that triggered it did so while working around a bug in GHC that is fixed in 7.4. Language bugs.. gotta love em.

Syndicated 2012-02-03 20:11:32 from see shy jo

unicode ate my homework

I've just spent several days trying to adapt git-annex to changes in ghc 4.7's handling of unicode in filenames. And by spent, I mean, time withdrawn from the bank, and frittered away.

In kindergarten, the top of the classrom wall was encircled by the aA bB cC of the alphabet. I'll bet they still put that up on the walls. And all the kids who grow up to become involved with computers learn that was a lie. The alphabet doesn't stop at zZ. It wouldn't all fit on a wall anymore.

So we're in a transition period, where we've all learnt deeply the alphabet, but the reality is much more complicated. And the collision between that intuitive sense of the world and the real world makes things more complicated still. And so, until we get much farther along in this transition period, you have to be very lucky indeed to not have wasted time dealing with that complexity, or at least having encountered Mojibake.

Most of the pain centers around programming languages, and libraries, which are all at different stages of the transition from ascii and other legacy encodings to unicode.

• If you're using C, you likely deal with all characters as raw bytes, and rely on the backwards compatability built into UTF-8, or you go to long lengths to manually deal with wide characters, so you can intelligently manipulate strings. The transition has barely begin, and will, apparently, never end.
• If you're using perl (at least like I do in ikiwiki), everything is (probably) unicode internally, but every time you call a library or do IO you have to manually deal with conversions, that are generally not even documented. You constantly find new encoding bugs. (If you're lucky, you don't find outright language bugs... I have.) You're at a very uncomfortable midpoint of the transition.
• If you're using haskell, or probably lots of other languages like python and ruby, everything is unicode all the time.. except for when it's not.
• If you're using javascript, the transition is basically complete.

My most recent pain is because the haskell GHC compiler is moving along in the transition, getting closer to the end. Or at least finishing the second 80% and moving into the third 80%. (This is not a quick transition..)

The change involves filename encodings, a situation that, at least on unix systems, is a vast mess of its own. Any filename, anywhere, can be in any encoding, and there's no way to know what's the right one, if you dislike guessing.

Haskell folk like strongly typed stuff, so this ambiguity about what type of data is contained in a FilePath type was surely anathama. So GHC is changing to always use UTF-8 for operations on FilePath. (Or whatever the system encoding is set to, but let's just assume it's UTF-8.)

Which is great and all, unless you need to write a Haskell program that can deal with arbitrary files. Let's say you want to delete a file. Just a simple rm. Now there are two problems:

1. The input filename is assumed to be in the system encoding aka unicode. What if it cannot be validly interpreted in that encoding? Probably your rm throws an exception.
2. Once the FilePath is loaded, it's been decoded to unicode characters. In order to call unlink, these have to be re-encoded to get a filename. Will that be the same bytes as the input filename and the filename on disk? Possibly not, and then the rm will delete the wrong thing, or fail.

But haskell people are smart, so they thought of this problem, and provided a separate type that can deal with it. RawFilePath hearks back to kindergarten; the filename is simply a series of bytes with no encoding. Which means it cannot be converted to a FilePath without encountering the above problems. But does let you write a safe rm in ghc 4.7.

So I set out to make something more complicated than a rm, that still needs to deal with arbitrary filename encodings. And I soon saw it would be problimatic. Because the things ghc can do with RawFilePaths are limited. It can't even split the directory from the filename. We often do need to manipulate filenames in such ways, even if we don't know their encoding, when we're doing something more complicated than rm.

If you use a library that does anything useful with FilePath, it's not available for RawFilePath. If you used standard haskell stuff like readFile and writeFile, it's not available for RawFilePath either. Enjoy your low-level POSIX interface!

So, I went lowlevel, and wrote my own RawFilePath versions of pretty much all of System.FilePath, and System.Directory, and parts of MissingH and other libraries. (And noticed that I can understand all this Haskell code.. yay!) And I got it close enough to working that, I'm sure, if I wanted to chase type errors for a week, I could get git-annex, with ghc 4.7, to fully work on any encoding of filenames.

But, now I'm left wondering what to do, because all this work is regressive; it's swimming against the tide of the transition. GHC's change is certainly the right change to make for most programs, that are not like rm. And so most programs and libraries won't use RawFilePath. This risks leaving a program that does a fish out of water.

At this point, I'm inclined to make git-annex support only unicode (or the system encoding). That's easy. And maybe have a branch that uses RawFilePath, in a hackish and type-unsafe way, with no guarantees of correctness, for those who really need it.

Previously: unicode eye chart wanted on a bumper sticker abc boxes unpacking boxes

Syndicated 2012-02-02 22:12:02 from see shy jo

announcing github-backup

Partly as a followup to a Github survey, and partly because I had a free evening and the need to write more haskell code, any haskell code, I present to you, github-backup.

github-backup is a simple tool you run in a git repository you cloned from Github. It backs up everything Github knows about the repository, including other forks, issues, comments, milestones, pull requests, and watchers.

This is all stored in the repository, as regular files, on a "github" branch.

Available in Cabal now, in Debian maybe if someone packages haskell-github.

Syndicated 2012-01-26 04:44:10 from see shy jo

olduse.net 1982

Hard to believe I've consumed all of 1981's Usenet posts now on olduse.net, and it's been running for 7 months already.

Last night, there was a "very long" post, describing nearly every node on usenet in 1982. There had been a warning about this post the day before, since it would take many sites half an hour to download at 300 baud. It was handily formatted as a shell script, which created per-node files.

So, I ran this code nobody has run since 1982. It worked. I got files. I tossed them on the olduse.net wiki, and used some ikiwiki code TOVA contracted me to write just a few months ago, to make clickable links on my usenet map.

The map data was contributed in another post a while back. By 1982, usenet is getting nearly impossible to map with 1982 technology of ascii art. I enjoyed throwing graphviz, git, wikis, and the web at it.

So, we have a collaboration across time, me and "Mark" and a lot of people who described their usenet nodes and piles of technology that make creating a mashup easy. Awesome!

I blog about stuff I find on the olduse.net blog. It's an open blog; Koldfront also blogs there, and we welcome other bloggers.

Some of the highlights for me have included:

As the space shuttle program is winding down, reading the excitement about the first shuttle flights, and the play-by-play coverage of a launch, posted to net.columbia by a high school student borrowing his dad's account. (A usegroup name that's hard to read without remembering its fate).

The announcements of the Motorola M68k, the IBM PC, and the CD-ROM.

Reading the TCP-IP digest, and Postel's plans for launching IPv4 soon, while the world IPv6 launch is being planned now. (The nay-sayers are especially fun to read. Including the guy who was concerned about the address space size, in 1981!)

Learning that nethack ascention tales have a history streching back 30 years, to rogue, and that the stories back then had much the same flavor as they do today.

Various celebrity sightings. Dennis Ritchie teaching C and Unix. Bill Joy talking vi. RMS talking .. nuclear politics?

The general development of usenet. B-news being rolled out, groups proliferating, many first inklings of what will be major problems and developments in 5 or 10 years. A shift in tone is already apparent, by now usenet is not only about announcements, there are already some flames.

Still 9 years to go!

Syndicated 2012-01-21 20:58:05 from see shy jo

version numbers

Today I released two entirely different pieces of software with the identical version number 3.20120115. Debian developers also will be soon noticing a piece of software I released with the version number 9.20120115.

I expect to move more of my software to this version number scheme over time, unless I find something badly wrong with it. It reflects how I think about versions for my software; there's a kind of continual "now" that development progresses through, in which individual releases have little discrete meaning and at the same time, there can also be significant discontinuities, that require the user to do something to deal with (such as a new debhelper compat version, or a new git-annex repository format).

Those two things are really all that I need a version number for my software to communicate. I can do without the rest of the things that version numbers are used for:

• The marketing of version 1.0 and 2.0.
• The comparative nuances such as whether 1.0 to 1.1 is a relatively big change, and 1.0 to 1.0.1 is a relatively small change
• The implication that 0.99 is almost 1.0 ready, and 1.1a is some kind of alpha release.

There is so much software, with so many version numbers that any signal encoded in such version numbers is swamped in the noise. Even on projects that I develop a version number like 2.88 is meaningless to me. All I care about is, how long ago was that version? Has there been a major change breaking compatibility since that version? "2.88" doesn't answer these questions well; "3.20111111" does.

It is a little wordy to have the full year in there, and it can be annoying to remember to set the version to the right date on release day (TODO: automate). This is balanced with the version not being so wordy as to include the time of day, which means I might have to do a 3.20120115.1 if I goof up. These minor problems are worth it to instantly know how old a version is when a user pastes it into a bug report.

And that is probably all I will ever have to say about version numbers. :)

Syndicated 2012-01-16 02:21:07 from see shy jo