Testing optimize-for-size suitability?
Is there a benchmark or set of benchmarks that would allow me to test the suitability of the CONFIG_CC_OPTIMIZE_FOR_SIZE kernel config option?
It seems to me that the benefit of this option is very dependant on the amount of CPU cache installed, with the compiler code generation being a secondary factor. The use, or not, of CONFIG_CC_OPTIMIZE_FOR_SIZE is basically an act of faith without knowing how it impacts my particular environment.
I've got a Pentium4 CPU with 512KB of L2 cache, and I'm using GCC v3.3.3. How can I determine whether or not CONFIG_CC_OPTIMIZE_FOR_SIZE should be used for my system?
Thanks.
Quoted 9 lines · Steve Snyder
Is there a benchmark or set of benchmarks that would allow me to test the suitability of the
CONFIG_CC_OPTIMIZE_FOR_SIZEkernel config option?It seems to me that the benefit of this option is very dependant on the amount of CPU cache installed, with the compiler code generation being a secondary factor. The use, or not, of
CONFIG_CC_OPTIMIZE_FOR_SIZEis basically an act of faith without knowing how it impacts my particular environment.
Quoted 3 lines
I've got a Pentium4 CPU with 512KB of L2 cache, and I'm using GCC v3.3.3. How can I determine whether or not
CONFIG_CC_OPTIMIZE_FOR_SIZEshould be used for my system?
it also saves 400 kb of memory that can be used by the pagecache now... that doesn't show in a microbenchmark but it's a quite sizable gain if you remember that a disk seek is 10msec..that is a LOT of cpu cycles..
Quoted 8 lines · Arjan van de Ven
Quoted 3 lines · Steve Snyder
Is there a benchmark or set of benchmarks that would allow me to test the suitability of the
CONFIG_CC_OPTIMIZE_FOR_SIZEkernel config option?it also saves 400 kb of memory that can be used by the pagecache now... that doesn't show in a microbenchmark but it's a quite sizable gain if you remember that a disk seek is 10msec..that is a LOT of cpu cycles..
Since I'm using a Duron 600MHz (the lowest model, AFAIK), I decided to compile a new kernel with the -Os option. Despite the scary warning on the help, it seems to be working fine for the first few moments with Debian's testing gcc (3.3.5).
I haven't noticed any other improvement, but I guess that more memory available to applications would never hurt (as long as stability is preserved).
I will also try to compile other applications (like Firefox), since my main memory is so limited and the small cache of my processor would surely appreciate it.
Perhaps it is time to give the tiny-linux project a try to see what other optimizations I can achieve.
Just another data point, Rogério.
Signature · 3 lines
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= Rogério Brito - [email protected] - http://www.ime.usp.br/~rbrito =-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Quoted 9 lines · Steve Snyder
Is there a benchmark or set of benchmarks that would allow me to test the suitability of the
CONFIG_CC_OPTIMIZE_FOR_SIZEkernel config option?It seems to me that the benefit of this option is very dependant on the amount of CPU cache installed, with the compiler code generation being a secondary factor. The use, or not, of
CONFIG_CC_OPTIMIZE_FOR_SIZEis basically an act of faith without knowing how it impacts my particular environment.
Quoted 5 lines
I've got a Pentium4 CPU with 512KB of L2 cache, and I'm using GCC v3.3.3. How can I determine whether or not
CONFIG_CC_OPTIMIZE_FOR_SIZEshould be used for my system?Thanks.
In theory, -O2 should produce faster code.
In practice, I don't know about any recent benchmarks comparing -Os/-O2 kernels.
In practice, I doubt it would make any noticable difference if the kernel might be faster by let's say 1% with one option compared to the other one.
The main disadvantage of -Os is that it's much less tested for kernel compilations, and therefore miscompilations are slightly more likely.
cu
Adrian
Signature · 4 lines
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon SeedHi Adrian,
Quoted 11 lines · Adrian Bunk
In theory, -O2 should produce faster code.
In practice, I don't know about any recent benchmarks comparing -Os/-O2 kernels.
In practice, I doubt it would make any noticable difference if the kernel might be faster by let's say 1% with one option compared to the other one.
The main disadvantage of -Os is that it's much less tested for kernel compilations, and therefore miscompilations are slightly more likely.
In fact, I've been compiling all my kernels with -Os for at least 2 years with gcc-2.95.3. -Os and -O2 produce nearly the same code on this compiler, it's even difficult to find algorithms which produce fairly different code with it. But things get different with gcc-3.3. -Os produces *really* smaller code (sometimes up to 20% smaller than -Os on gcc-2.95.3), but this code also becomes fairly slower, and disassembling it sometimes shows what can be called stupid code, because speed optimization completely disappears eventhough sometimes both size and speed could be optimized (eg: by switching two instructions to prevent pipelines stalls). On various code, I found gcc-3.3 -Os to deliver about 30% less performance than -O2. On the other hand, gcc-3.3 -O2 produces bigger and sometimes faster code than gcc-2.95 -O2. So it's difficult to say which one is better, it really depends on what you do with it.
I have not benchmarked any gcc-3.3 -Os kernel yet, though I've already been running some of them accidentely because of an old .config which gets rebuild on a machine which only has gcc-3.3. I've not noticed miscompilations nor really perceptible slowdowns, but I've not measured this last point.
What I often found efficient to both reduce code size and improve speed is to play with code alignment and stack boundary. Using "-mpreferred-stack-boundary=2" keeps the stack 32-bit aligned, which removes some entry and leave code in all functions (and esp, ...). Code alignment with -malign-loops=0, -malign-jumps=0 and -malign-functions=0 reduces the code size while not really affecting its speed (or just slightly increase some of these params if you find speed problem).
Regards,
Willy