[cairo] Unreliable cairo-perf results; was Re: Mark cairo_fixed_t conversion functions with attribute(const)

[Chris Wilson]

Ross Burton (ross at burtonini.com) said: 
> You're producing some really interesting looking patches, but I don't
> see any proof that there are performance improvements.  I believe there
> are, but I like to see hard numbers!  Can you run cairo-perf-diff?

Hmm, would love to provide conclusive evidence showing that this change
improves gcc's output. However I've not found any machine where
cairo-perf gives reliable results,  e.g.

ickle at cube:~/src/cairo/perf$ git diff HEAD^ HEAD
diff --git a/src/cairo.h b/src/cairo.h
index a80efde..7701ec4 100644
--- a/src/cairo.h
+++ b/src/cairo.h
@@ -43,6 +43,8 @@
 
 CAIRO_BEGIN_DECLS
  
 +/* force a recompilation */
 +
  #define CAIRO_VERSION_ENCODE(major, minor, micro) (     \
			          ((major) * 10000)                             \
			        + ((minor) *   100)                             \
ickle at cube:~/src/cairo/perf$ ./cairo-perf-diff HEAD^ HEAD -- -i 100
Speedups
========
image-rgb     paint_image_rgba_source-256  174.08µs 0.25% -> 146.91µs 0.32%:  1.18x speedup
▏
image-rgb        paint_image_rgb_over-256  162.36µs 1.22% -> 140.51µs 0.84%:  1.16x speedup
▏
image-rgb   paint_similar_rgba_source-256  169.44µs 0.53% -> 147.17µs 0.39%:  1.15x speedup
▏
image-rgb      paint_image_rgb_source-256  160.88µs 0.79% -> 140.12µs 0.72%:  1.15x speedup
▏
...
For the record, that machine is an idle, single processor AMD64 with
1GiB of RAM.

Sorry for the whine - this has been driving me crazy.
--
Chris Wilson