On 4/22/15 5:25 PM, David Miller wrote:
From: David Ahern <david.ahern@xxxxxxxxxx>
Date: Wed, 22 Apr 2015 17:19:23 -0600
Only thing left in my queue is optimized versions of the ffs / fls
families, but that patch is v9b specific, not M7.
Something faster than the popc thing in arch/sparc/lib/ffs.S?
hmmm... i saw that, but wasn't clear 1) how it got inserted and 2) the
overhead of a function call versus inline. Anyways, what I have is the
same 3 instructions as an inline. But really the __ffs was just along
for the ride; the focus was on __fls.
Are you thinking of using "lzcnt"? I wasn't impressed with the
performance of that instruction last time I played around with it.
A comparison of what I hacked together is attached (columns too wide for
inline). Data is from a T4-2. It shows lzcnt to be better for __fls, fls
and fl64.
I'd like to put some attention on precise mode for perf counters; it
just has not bubbled to the top.
That plus the backtrace deadlock thing we're discussing in another
thread, that bug is irritating because your pagefault_disable() change
should "just work".
oh, yes. forgot about that one. I spent too many hours trying to figure
out why processes get killed with a sigbus. I added an option to perf
tool to skip userspace chains until I can get back to it.
- "slow" means version from asm-generic.
- Times are in nsec.
- 'bit' column shown to ensure correct answer between current and lzcnt
- average of 10 back-to-back calls
| __fls | fls | fls64
word | lzcnt slow | lzcnt slow | lzcnt slow
| bit dt bit dt | bit dt bit dt | bit dt bit dt
0 | 0 15 0 67 | 0 19 0 21 | 0 14 0 14
1 | 0 13 0 50 | 1 32 1 61 | 1 20 1 51
80000000 | 31 13 31 39 | 32 30 32 49 | 64 25 64 37
8000000000000000 | 63 13 63 34 | 0 17 0 16 | 0 12 0 14
