237041 - compile options for improved minimo speed on arm.

Status: RESOLVED INVALID

(Firefox Build System :: General, defect)

Description

[chris hofmann]

Here are the options to try with GCC 3.3 when
compiling for ARM target:

-Os -finline-limit=1
-mno-apcs-frame -fno-schedule-insns
-mno-thumb-iterwork -fno-optimize-sibling-calls
-fomit-frame-pointer -fno-if-conversion
-ffunction-sections -fno-thread-jumps
-fno-force-mem -fno-hosted
-fno-foce-addr -fno-expensive-optimizations
-fno-inline-functions -fargument-noalias-global
-fbranch-probabilities
-fno-gcse -fno-rerun-cse-after-loop

These options should be optimal for COMPILATION SPEED.

( IF you want to optimize for size, you should
 replace "-fno-gcse -fno-rerun-cse-after-loop"
 with "-fno-data-sections -fnew-ra". BUT the
 optimized set of options for compilation speed
 produces SMALLER code that "-Os" alone... )

I guess somebody should check that these options
are safe. For example, -fargument-noalias-global
seems suspect...

let's optimize for speed and generate new performance and memory graphs to compare.

Comment 1

[chris hofmann]

adding cls in case he know anything about any of these options.

Comment 2

[cls]

Sorry, I haven't played with any of those besides -Os.  Maybe bryner has?

Comment 3

[chris hofmann]

 ------- Additional Comment #4 From chris hofmann  2004-03-20 07:55 PST  [reply]
-------

Created an attachment (id=144405)
mem use graph with these options


------- Additional Comment #5 From chris hofmann 2004-03-20 07:59 PST [reply]
-------

comparison mem us graph from release 0.1.
http://www.mozilla.org/projects/minimo/releases/arm/0.1/minimo-mem.gif

some improvememt, but it is not quite apples to apples comparison since we might
have made ground with other fixes between minimo 0.1 (mozilla 1.7 alpha) and
1.7b where this minimo snapshot came from.

Comment 4

[chris hofmann]

Attachment: data details

Comment 5

[Doug Turner (:dougt)]

In the build that the data was collected against, I did not use the follow
compiler options since they did _not_ produce a launchable build.

-fno-foce-addr
-mno-apcs-frame                                                                 
-mno-thumb-iterwork

Here is the build that was used to collect this data:

http://www.meer.net/~dougt/minimo_0.2-a1_arm.ipk

Comment 6

[Doug Turner (:dougt)]

this would be a great thing to have someone test again.

Comment 7

[Guillaume Letellier]

Do we really want to use all those "black magic" options? In particular -fargument-noalias-global will cause GCC to make assumptions which AFAIK violate the C/C++ language standards and -fnew-ra is considered broken by the GCC developers.

Most of the time on ARM platforms we only use:   -g -O2 -march=armv5t  (generic for most of the platforms (armv5, V6...)
or   -g -O2 -march=armv6 for armv6 platforms  like the N800
or more specifically   -g -O2 -mcpu=arm1136jf-s -mfpu=vfp -mfloat-abi=softfp   
for arm11 platforms like the N800

Comment 8

[Doug Turner (:dougt)]

i do not think we want to use any black magic unless it pays off.  these options were given to us from an embedder a long time ago, and we created this bug to track them.  Since we didn't have any arm/gcc experts, we didn't really know what to make of the options.

Comment 9

[Stuart Parmenter]

agreed, lets not do any black magic