Re: [Sbcl-devel] [Sbcl-commits] CVS: sbcl/src/code linux-os.lisp, 1.12, 1.13 time. lisp, 1.31, 1.32 unix.lisp, 1.66, 1.67 win32-os.lisp, 1 .2, 1.3 win32.lisp , 1.5, 1.6

[Yaroslav K. <kav...@je...>]

Nikodemus Siivola wrote:
> I assume you mean this as a jest, surely you're not seriously suggesting
> we pun FILETIME into an integer?

For what use structure if not use its fields?
In win32.lisp FILETIME used only for get-process-times.

To do that:
(flet ((filetime-to-100-ns (time)
                    (+ (ash (slot time 'dw-high-datetime) 32)
                       (slot time 'dw-low-datetime))))

instead of simply get (unsigned 64)?

> My quick benchmarks gave no idication of this being a bottleneck --
> indeed, you need to do quite a few calls to GET-INTERNAL-RUN-TIME in
> order to see it move to the next tick. (Actually, we might see how far
> we could push the resolution of the value we return, both on Win32 and
> elsewhere.)

My quick benchmarks:

get-process-times1 use FILETIME as structure
get-process-times2 use FILETIME as unsigned-long-long

CL-USER> (time
	  (dotimes (x 5000000)
	    (multiple-value-bind (res c-time e-time k-time u-time)
		(get-process-times1)
	      (declare (ignore res c-time e-time k-time u-time)))))
Evaluation took:
   9.0 seconds of real time
   7.203125 seconds of user run time
   1.890625 seconds of system run time
   [Run times include 0.469 seconds GC run time.]
   0 page faults and
   760,006,936 bytes consed.
NIL

CL-USER> (time
	  (dotimes (x 5000000)
	    (multiple-value-bind (res c-time e-time k-time u-time)
		(get-process-times2)
	      (declare (ignore res c-time e-time k-time u-time)))))
Evaluation took:
   6.0 seconds of real time
   3.59375 seconds of user run time
   1.765625 seconds of system run time
   [Run times include 0.156 seconds GC run time.]
   0 page faults and
   280,002,008 bytes consed.
NIL

Tics superfluous are not.

-- 
WBR, Yaroslav Kavenchuk.