04:38 <whitequark> hey sb0

04:38 <whitequark> I was reading the output of ARTIQ and I thought up a great optimization

04:39 <whitequark> we can eliminate like seven memory accesses from a loop calling ttl.pulse

04:50 <sb0> yeah? how?

04:54 <whitequark> changing the calling convention to pin now to a register

04:58 <whitequark> right now `now` is RMWed all around, which is not optimized because now can be aliased

05:00 <sb0> ah, yes, I had seen that

05:01 <sb0> how much of a mess does it make to implement that?

05:01 <sb0> and what is the performance gain?

05:06 <whitequark> not especially messy. the easiest way is to put it as the first argument, which pushes the sixth normal argument onto the stack

05:06 <whitequark> I can't say what the gain will be until I have tried it

05:07 <whitequark> most of the advantage of hoisting stuff comes from the soft-FP, so this is unlikely to gain more than maybe dozens of nanoseconds of slack