| |||||||||||||
What's coming in $NEXT_KERNEL_VERSION, part 1What's coming in $NEXT_KERNEL_VERSION, part 1Posted May 26, 2011 14:55 UTC (Thu) by pr1268 (subscriber, #24648)Parent article: What's coming in $NEXT_KERNEL_VERSION, part 1
Just curious, does the kernel have to work independently of any standard library? The kstrtol_from_user function appears to be nearly identical to strtol(3) and even the bsearch kernel function has the exact same prototype signature as the CStdLib function of the same name. Also, how much improvement is -Os supposed to have over, e.g. -O2? My personal experience with code I write is that -Os binaries are not significantly smaller (or execute any faster). I have read the gcc man page, but perhaps there's more to it than what the man page says. Thanks! Respectfully,
(Log in to post comments)
What's coming in $NEXT_KERNEL_VERSION, part 1 Posted May 26, 2011 18:03 UTC (Thu) by nevets (subscriber, #11875) [Link]
For you first question, yes. The kernel does not link to any other library. Basically everything used by the kernel is defined by the kernel. Standard C libraries do too much that is not allowed in the kernel (like any floating point operations).
The kstrtol_from_user() does much more than strtol. That "from_user" part is critical. It is accessing data from an untrusted source. It must be prepared for that address to segfault, and make sure it does not belong to something that the user should not be touching.
For your second question, well, that's the point for switching back to -O2 :) The idea was that a smaller kernel would have less pressure on the instruction cache and be a bigger win in performance. Anytime you take a miss in the instruction cache, the CPU must stall to retrieve the memory that contains the instructions. As CPUs are much faster than memory access, the less you miss instruction cache, the faster you run. But if the -Os is not affecting this much at all, there's no point in using it.
What's coming in $NEXT_KERNEL_VERSION, part 1 Posted May 27, 2011 22:01 UTC (Fri) by oak (subscriber, #2786) [Link]
I think that at some point during v4.x development, gcc -Os option was changed from producing smaller code without much performance hit to something that mainly aims at smaller code. If I remember right, there was even a gcc bug filed by Mozilla devs about this as they thought it a regression...
What's coming in $NEXT_KERNEL_VERSION, part 1 Posted May 31, 2011 13:17 UTC (Tue) by pbonzini (subscriber, #60935) [Link]
> gcc -Os option was changed from producing smaller code without much
> performance hit to something that mainly aims at smaller code
That actually had always been the aim, but the tuning was not so good. For example, Mozilla was making some simple functions non-static that should have been static. These functions were very simple, but they were used only once and they still had to be included in the final binary. So GCC started being worried about including two copies of them and stopped inlining them. (This was fixed by tuning inliner heuristics to detect the optimization opportunities coming from the inlining of such simple functions).
|
|||||||||||||