Messages in this thread | ![/](/images/icornerl.gif) | | Date | Wed, 24 Sep 2003 22:22:18 -0400 | Subject | Re: Minimizing the Kernel | From | rwhron@earthlin ... |
| |
> Hmm, has anyone tried -Os with gcc3+ ? > Maybe that'd be good for size optimization?
2.6.0-test3 compiled with gcc-3.3.1 and redhat 7.2's gcc-2.96-112. gcc-3.3.1 saves about 275k text (10%). -Os is more effective on gcc-3.3.1 than 2.96. These were all built with the same .config.
size vmlinux-* text data bss dec hex filename 2120419 449928 131748 2702095 293b0f vmlinux-3.3.1-Os 2124890 449928 131748 2706566 294c86 vmlinux-3.3.1-Os-falign=2 2334482 457304 125952 2917738 2c856a vmlinux-2.96-112-Os 2405382 449960 131748 2987090 2d9452 vmlinux-3.3.1 2408343 457332 125952 2991627 2da60b vmlinux-2.96-112
Most frequently saved instruction with gcc-3.3.1 -Os is nop.
This was on x86. The -falign=2 version had -falign-functions=2 -falign-jumps=2 -falign-labels=2 -falign-loops=2. I believe the default with -Os is no alignment (i.e. -falign-*=0).
I benchmarked those compilers/options on a K6/2. You could wade through http://home.earthlink.net/~rwhron/kernel/bigbox.html to see all the results.
Quick summary: gcc-3.3.1 -Os -falign=2 was best for most LMbench tests. For other benchmarks that wasn't always true. K6/2 has small L1 cache (32+32K) L2 cache is 1M, but not much faster than RAM on my box.
These generalizations can be made: 1) gcc-3.3.1 -Os kernel code is about 10% smaller than gcc-2.96 -Os. 2) Actual memory savings is not 10%, because dynamic structure sizes don't change. 3) gcc-3.3.1 takes significantly longer to compile source.
YMMV
-- Randy Hron http://home.earthlink.net/~rwhron/kernel/bigbox.html
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| ![\](/images/icornerr.gif) |