386897 - compile cairo/pixman with full speed optimization

Status: RESOLVED FIXED

(Core :: Graphics, defect)

Description

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Attachment: optimize cairo/pixman for speed, not space

I saw some significant codegen differences when cairo was compiled with -O2 vs. -O1 (optimize for speed instead of our default optimize for space).  In particular, the _cairo_lround function which is used in some critical paths (win32 glyph drawing for one) is very expensive with -O1 but is inconsequential with -O2 (like 2500ms -> 30ms for a simple benchmark of the function).

This patch enables speed optimizations for cairo and pixman.  I don't know what the codesize hit would be, mind you, but I'm pretty sure that we should take it.

Comment 1

[Eli Friedman]

Do we trust GCC with -O3?

Comment 2

[dolphinling]

Gentoo's compilation optimization guide says "In 3.x, -O3 has been shown to lead to marginally faster execution times over -O2, but this is no longer the case with gcc 4.x."

Feel free to draw you own conclusions, but mine is that if -O2 gives the 2500ms -> 30ms jump, then stick with it, -O3 probably isn't needed.

(also, perhaps -Os should be benchmarked... it includes many of the optimizations in -O2 over -O1, and possibly the one that matters here is one of those)

Comment 3

[Stuart Parmenter]

Comment on attachment 270959 [details] [diff] [review]
optimize cairo/pixman for speed, not space

would it be better to optimize all of gfx this way?

Comment 4

[Jesse Ruderman]

2500ms -> 30ms by changing optimization options?  Wow.  I wonder if the gcc team would consider that a bug.

Comment 5

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

-O2 should be fine for gcc -- I was more speaking in terms of -O2 for msvc.  The optimization block is what we use in a few other places where we want to compile with optimizations: sqlite and some other bits, though -O3 for gcc is probably a leftover from gcc 3.x.

The 2500ms -> 30ms was under win32 msvc; it might or might not be a bug.  That is, there was a significant performance gain from no optimizations to optimize-size, and another from optimize-size to optimize-speed; the function probably is a little bigger.

Comment 6

[Ryan VanderMeulen [:RyanVM]]

FWIW, MS recommends you use -O2 over -Ox
http://msdn2.microsoft.com/en-us/library/59a3b321(VS.80).aspx

Comment 7

[Vladimir Vukicevic [:vlad] [:vladv] (needinfo me, slow to respond)]

Checked in with -O2 for both gcc and msvc.