On 8/29/2022 9:13 AM, Toon Moene wrote:
On 8/29/22 17:08, Jeff Law via Gcc-patches wrote:
However, I'm hoping to forget as many floating point details, as fast
as possible, as soon as I can ;-).
Actually FP isn't that bad -- I'd largely avoided it for decades, but
didn't have a choice earlier this year. And there's a lot more
headroom for improvements in the FP space than the integer space IMHO.
One of the more interesting ones is to try to limit the range of the
input to the trigonometric functions - that way you could use ones
without any argument reduction phase ...
The difficult part is that most of the trig stuff is in libraries, so we
don't have visibility into the full range.
What we do sometimes have is knowledge that the special values are
already handled which allows us to do things like automatically
transform a division into estimation + NR correction steps (atan2).
I guess we could do specialization based on the input range. So rather
than calling "sin" we could call a special one that didn't have the
reduction step when we know the input value is in a sensible range.
Jeff
