This week I learned...

• A shibboleth is language that sends cultural signals beyond its plain meaning. The word comes from a fairly amazing Bible story.

  Now I want a word for language invented to annoy, like “Democrat Party”.

• Aristotle believed slavery to be “expedient and right”. All the best arguments by learned apologists for slavery in the U.S. South were from his writings, particularly in the Politics.

• Incidentally, Aristotle thought democracy was a crummy form of government. And, in his ideal society, homeschooling would be banned.

• There are expressions of the Golden Rule in the scriptures of many religions.

• I don't understand what this means, exactly, but the x86 instruction set contains an FNOP instruction. The manual describes it as a floating-point no-operation instruction. How this might be different from a regular NOP I don't know.

• So there's a widely known interview question: write a function to count the number of bits of a 32-bit int that are set to 1. This function is called the population count—that link describes a surprising use for it.

  The last time I tested this, summing 4 queries into a 256-value lookup table is fastest for 32-bit integers, faster than the awesomely clever bit-twiddling solution. I shouldn't have been surprised. The lookup table fits easily into cache. The bit-twiddling solution has a lot of dependencies; the CPU can't find any instruction-level parallelism there.

  Anyway, what I learned a week or two ago is that future Intel chips will have an SSE instruction, POPCNT, that does this, in parallel, for several words at a time. (Someone I mentioned this to commented that he doesn't want to be fired for pronouncing that.)

• Often when I write these blog entries, I'm still unsure of the significance of some of the things I've just learned. For example, I learned something about Java monitors (or pthreads condition variables, which are the same thing). When thread 1 notifies, waking thread 2 on a separate CPU, the lock associated with the monitor ensures that CPU 1's writes are flushed to main memory and CPU 2 sees them before thread 2 starts running. There's no need for write-barrier magic in Object.notify itself.

• Apple's Shark profiler has a feature that lets you compare two profiles. But the result is calculated by comparing percentages, not comparing the absolute number of samples. So, for my purposes, useless.

• The .mshark files produced by Shark are gzipped binary property lists, but the actual samples are stored in there as raw binary data which I haven't tried to decipher.