freenode/#clasp - IRC Chatlog

14:06:53 Bike sub byte arrays are kind of difficult because the gcbitarray constructor works with words, not elements... and that kind of ties into how all the gcarray constructors are... but that's deep stuff to change

14:21:45 Bike using clz instead of checking each bit individually, for (position 1 simple-bit-vector), is like two hundred times faster, so that's kinda cool

14:36:43 beach Yes, sub-byte arrays are problematic, but tricks like that are worth the trouble.

14:40:39 Bike If a bit vector happens to be a multiple of 64/whatever long, it might be easier to reverse, too. apparently some architectures have bit reverse instructions

14:40:54 Bike not x86, but llvm can still put in a clever series of rotations and stuff for you

14:48:47 beach Interesting. I hadn't considered the reversal thing. In the past I would have done a 256-element table to apply for each byte, but nowadays I am not so sure.

14:54:40 Bike yeah, gotta worry about the memory latency

14:55:33 beach Exactly.

14:59:12 beach I wonder how well a decision tree would work.

14:59:39 beach (if (< x 128) (if (< x 64) ...))

14:59:51 beach For a single byte I mean.

15:29:25 Bike The usual on x86 seems to be a loop of shifts. Shifting left puts the former MSB in the carry flag, and then you can use "add x, x" to both shift the result left and put the carry flag in the LSB.

15:29:27 Bike Kind of cool.

15:36:17 Bike er, you shift right

15:36:25 Bike which still puts it in the carry i guess

15:38:56 beach Oh, nice.

15:40:24 beach Hmm, but that only works once, it seems.

15:41:22 Bike What do you mean?

15:41:57 beach If you do it one more time, the LSB will be shifted, but you don't want that.

15:42:35 Bike Er, well I mean you have a source register and a destination register.

15:42:47 Bike You shift out of the source with shr and into the destination with add.

15:43:13 Bike adc.

15:43:20 beach Ah, OK. Let me contemplate that some more.

15:43:51 beach But not today. Time to fix dinner for my (admittedly small) family.

15:43:57 beach ACTION is off.

18:51:47 Bike my psuedo copy constructor compiles without complaint but at runtime it makes a bit vector full of junk data, instead of a copy

18:51:47 Bike weird