[gsharp-devel] redraw buglets

[Christophe Rhodes]

Christophe Rhodes <csr21 at cam.ac.uk> writes:

>    1    213   5.9    213   5.9 SB-BIGNUM:BIGNUM-GCD
>    2    171   4.8    384  10.7 SB-BIGNUM::SUBTRACT-BIGNUM-BUFFERS
>    3    139   3.9    523  14.5 "XEP for #'(SB-KERNEL:INSTANCE-LAMBDA (SB-PCL::.ARG0.) (LET NIL # ...))"
>    4    122   3.4    645  17.9 SB-BIGNUM::MAKE-GCD-BIGNUM-ODD
>    5    100   2.8    745  20.7 SB-BIGNUM::BIGNUM-GCD-ORDER-AND-SUBTRACT
>    6     96   2.7    841  23.4 "XEP for #'(SB-KERNEL:INSTANCE-LAMBDA (SB-PCL::.ARG0. SB-PCL::.ARG1. SB-PCL::.ARG2. ...) (LET NIL # ...))"
>    7     95   2.6    936  26.0 SB-KERNEL::%%TYPEP
>    8     89   2.5   1025  28.5 SB-BIGNUM::BIGNUM-BUFFER-ASHIFT-RIGHT
>    9     88   2.4   1113  31.0 SB-KERNEL:TWO-ARG-+
>   10     81   2.3   1194  33.2 SB-KERNEL:CLASSOID-TYPEP

Here are some more numbers, from profiling (with the regular profiler)
all functions in gsharp's own packages (but not CLIM or X-related
ones).

  seconds  |    consed   |  calls  |  sec/call  |  name
-------------------------------------------------------------
    39.229 | 550,292,048 |       1 |  39.229393 | GSHARP::RUN-GSHARP
    17.950 |  42,999,328 |  23,237 |   0.000772 | SCORE-PANE::MAKE-PIXMAP-RECORD
    16.849 |  36,449,032 |  21,135 |   0.000797 | SCORE-PANE::MAKE-RECTANGLE-RECORD
     7.168 |  32,446,368 |  21,135 |   0.000339 | SCORE-PANE::ADD-NEW-RECTANGLE-RECORD
     6.644 |   8,482,896 |  32,962 |   0.000202 | SCORE-PANE::MEDIUM-DRAW-NOTEHEAD
     6.269 |  11,886,152 |  10,366 |   0.000605 | SCORE-PANE::DRAW-SEGMENT
     5.188 |   8,387,400 |  25,310 |   0.000205 | SCORE-PANE::MEDIUM-DRAW-STEM
     4.645 |   6,632,224 |  23,237 |   0.000200 | SCORE-PANE::ADD-NEW-PIXMAP-RECORD

Some of the blame might lie in gsharp's MAKE-PIXMAP-RECORD and
MAKE-RECTANGLE-RECORD: the way they're written inhibits a CLOS
optimization: if make-instance sees a constant class name and constant
initargs, and there are no applicable methods on
(setf slot-value-using-class), then a certain amount of heavyweight
dispatch can be avoided.  However, given the statistical profiling
results indicating that bignums or ratios were the bottleneck, this is
likely only to be a side issue.

I think I'd investigate the medium transformation, and the various
coordinates associated with pixmaps and rectangles, and see if any of
those are overly precise...

Cheers,

Christophe
-- 
http://www-jcsu.jesus.cam.ac.uk/~csr21/       +44 1223 510 299/+44 7729 383 757
(set-pprint-dispatch 'number (lambda (s o) (declare (special b)) (format s b)))
(defvar b "~&Just another Lisp hacker~%")    (pprint #36rJesusCollegeCambridge)