7 May 2004 graydon   » (Master)


met some really nice, slightly older and significantly wiser folks for dinner yesterday; it's always neat to talk with people who have more long-term perspective on computing trends. also spent a few days in north carolina with the red hat gang soaking up the corporate cheer. it was a bit rushed, but I had a lot of fun and met many interesting and friendly hackers. next up: gcc summit!


I wanted to lend a moment's support to the D language, which is sensibly crafted and sticks to pragmatics rather than dogma. pragmatics are important. I'm especially happy to see it promoting DBC and formalized support for unit testing.

languages in general

I've been tinkering with some ideas for programming languages recently. not necessarily "new languages", but aspects of programming which I feel ought to be expressable in language. there are two issues in particular I'd very much like to see formalized into language constructs:

  1. "design by accounting" (in a sense analogous to DBC): the notion that the signature of a function includes resource use accounting; typically in terms of memory and time, though conceivably also in terms of user-defined costs. the notion being that a function has a certain cost which is the sum of the costs of its sub-expressions, with loops costing some bounded-above multiple of the cost of their body. costs would be calculated as much as possible by the compiler, and statically verified as much as possible. analogous to other forms of mixed static/dynamic checking, those cost checks which could not be statically verified (say at module boundaries, i/o operations, or hard-to-analyze loop heads) would be left in the program, making "withdrawls" from a runtime-managed cost center, with underflow causing an exception. the technology for this is just a redeployment of conventional value range propagation and bounds checking.

  2. an "ownership tree". this came up last night too. every time I get to discussing modern languages, recently, an uncomfortable fact comes up: I don't like garbage collection. I don't like it because it encourages the programmer to not think about ownership, so they build systems which slowly leak space, by accidental accumulation of dangling references. I don't claim that C's "explicit free" model is any better mind you; it makes you think about ownership but gives you no vocabulary with which to express your thoughts, so you wind up acting on mistaken assumptions (double free, forgotten free).

    I think the solution to this is to differentiate, at a low level, between owning and non-owning (weak) pointers. you should get an owning pointer from an allocation, and each allocation should (until it dies) have exactly 1 owning pointer in the program. you enforce this by prohibiting direct copying of owning pointers, but make 3 variants of the concept of "copying an owning pointer": one which clones the pointee into a new owning pointer, one which forms a weak pointer, and one which forms a non-storable reference (for use as a function argument). if you want an owning pointer you can move around (and possibly set to null), you make a discriminated union of "nothing" and an owning pointer (like haskell's Maybe 'a or ML's 'a option), and add a language construct which takes two of these values and transfers an owning pointer between them atomically. passing an owning pointer as an argument -- another form of "making a copy" -- is similarly prohibited.

    in other words, the rules of the language should make cyclical ownership, null pointers, and double-ownership inexpressible.

    this strikes some people as draconian, but I claim that it is nothing more than the extension of stack discipline from control into data. we already enforce, in many languages, that all control structures must be expressed as a tree of function calls, rather than an arbitrary graph of gotos. at any time, there is a single, obvious stack of activation frames from main() to your current context. this simplifies reasoning about control; imagine debugging without stack traces! all I'm saying is that we should force all data structures to express their ownership as a proper tree, rather than an arbitrary graph of pointers.


spent some weeks digging through the repaint, double buffer management, clipping, and paint coalescing sections of swing. I think I have it mostly right now, but then cairo went and changed underfoot. alas.


I released a version of monotone with support for human readable version names and win32. two more boxes ticked off the list.


a year ago we bought an extra athlon as a workhorse for compute-heavy jobs. frances needed it mostly, but I use it sometimes for big builds or tests. it turns out the athlon was much faster than the one we meant to buy (accidentally got the wrong part) and came with a very bad fan. this gave it the unfortunate habit of critically overheating and shutting down.

three things were done to fix this:

  1. a new heat sink made of copper.
  2. a program called athcool which twiddled the necessary registers to have the chip actually enter a low power state when the OS runs HLT. this is not the default. if you have an athlon, you probably want this program to run at boot.
  3. a homebrew thermostat consisting of the lm_sensors module, a gkrellm alarm, and a shell script. the sensors monitor temperature and the alarm trips after a certain temperature is reached. this runs the shell script, which pulls out the top process id on the system, sends it SIGSTOP, and schedules an 'at' job for 5 minutes in the future to send SIGCONT.

Latest blog entries     Older blog entries

New Advogato Features

New HTML Parser: The long-awaited libxml2 based HTML parser code is live. It needs further work but already handles most markup better than the original parser.

Keep up with the latest Advogato features by reading the Advogato status blog.

If you're a C programmer with some spare time, take a look at the mod_virgule project page and help us with one of the tasks on the ToDo list!