FAST COSINE DOUBLE PRECISION FP

@Madis731: When I showed that moving through memory was faster than through a register on Core 2 Duo, the reason wasn't because of latency but because it forced the operation to go through ports than were otheriwse unused. The latency is I think 4 clocks more for the memory move than for the register move.

I prefer movddup for duplicating inputs into a register as it's a pretty old processor that doesn't have SSE3 by now. Are you aware that when Agner Fog leaves the latency field blank in his instruction tables that it means that his tests weren't able to determine their latency, not that the latency is zero? Operations from memory normally have the same latency for the destination (register) operand and something like 2 or 3 clocks more latency for the source (memory) operand. There are some operations such as movsd that are different semantically for a register source and a memory source where this may not be true. It's difficult to measure the latency of a load because it can't be disentangled from that of the store that meant you had an event you had to wait on in the first place. Core 2 Duo also has an additional clock latency between a load and a floating point operation on the modified register and a store operation after a floating point operation, IIRC. Latencies through memory are obviously much worse if store forwarding rules aren't followed. I timed some of this explicitly at some point but I'm too lazy to dig out the code just now.

The fastest way to compute SIN and COS is to store up a few first quadrant angles then compute their trig functions simultaneously. Argument reduction is the longest part of the algorithm if it can't be avoided and current processors need lots of parallel work even for single-threaded code to be efficient. Apologies if these facts have already been pointed out in this thread.