On Jan 21, 2008, at 3:15 PM, Theodore Tso wrote:
On Mon, Jan 21, 2008 at 03:01:43PM -0500, Kevin Ballard wrote:You seem to be under the impression that I'm advocating that git treat all filenames as unicode strings, and thus change its hashing algorithm as described. I am not. I am saying that, if git only had to deal with HFS+, then it could treat all filenames as strings, etc. However, since git does not only have to deal with HFS+, this will not work. What I am describingis an ideal, not a practicality.Well, why are you arguing on the git list about precisely that (when you reponsed to Linus), then?
Because of the way in which an argument evolves. This started out as "HFS+ is stupid because it normalizes", and I was arguing that said normalization wasn't stupid. This turned into an argument as to why HFS + wasn't stupid for normalization, which is basically this argument of the ideal. Yes, I realize that it's not producing any practical results, but I'm stubborn (as, apparently, are most of you), and I believe that if the official stance of the git project is "HFS+ is stupid" then there's a lower chance of a patch being accepted then if people accept that "HFS+ is different in an incompatible fashion".
In other words, what I'm saying is that treating filenames as strings works perfectly fine, *provided you can do that 100% of the time*. git cannot do that 100% of the time, therefore it's not appropriate here. The purpose of this argument is to illustrate that treating filenames as strings isn't wrong, it's simply incompatible with treating filenames as byte sequences.No, it's still broken, because of the Unicode-is-not-static problem. What happens when you start adding more composable characters, which some future version of HFS+ will start breaking apart?
If you need a static representation, you normalize to a specific form. And in fact, adding new composable characters doesn't matter, since if they didn't exist before, you couldn't have possibly used them. Unless you mean adding new composed forms of existing simpler characters, at which point you seem to be arguing for NFD instead of NFC.
Presumably the whole *reason* why HFS+ was corrupting strings was so that "stupid applications" that only did byte comparisons would work correctly. But when you upgrade from Mac OS 10.5 to 10.6, and it adds support for new composable characters, and you now take a USB hard drive that was hooked up to a MacBook Air, running one version of MacOS, and hook it up to another Macintosh, running another version of MacOS, the normalization algorithm will be different, so the byte comparisons won't work.
I doubt that HFS+ normalized so that "stupid applications" could do byte comparisons. But even if that were the case, see previous paragraph.
So all of this extra work which MacOS put in to corrupt filenames behind our back doesn't actually do any good; applications still need to be smart, or there will be rare, hard to reproduce bugs nevertheless. So if MacOS wants to supply Unicode libraries that compare strings keeping in mind Unicode "equivalences" it can be our guest (although how they deal with different versions of Unicode with different equivalence classes will be their cross to bear). BUT MacOS X SHOULD NOT BE CORRUPTING FILENAMES. TO DO SO IS BROKEN.
Your entire argument is based on the assumption that HFS+ "corrupts" filenames in order to allow dumb clients to do byte comparisons, and I don't believe that to be the case. In fact, it's only considered a corruption if you care about the byte sequence of filenames, and my argument is that, on HFS+, you aren't supposed to care about the byte sequence.
-Kevin Ballard -- Kevin Ballard http://kevin.sb.org kevin@xxxxxx http://www.tildesoft.com
<<attachment: smime.p7s>>
