Hello, On 2/7/20 9:29 AM, Martin KaFai Lau wrote: > On Fri, Feb 07, 2020 at 07:34:57AM +0100, Uwe Kleine-König wrote: >> Hello, >> >> On 2/7/20 12:47 AM, Linus Torvalds wrote: >>> Instead, it could just calculate the nbuckets first, and then do the >>> "log2()" on that: >>> >>> /* Use at least 2 buckets, select_bucket() is undefined >>> behavior with 1 bucket */ >>> nbuckets = max_t(u32, 2, roundup_pow_of_two(num_possible_cpus())); >>> smap->bucket_log = ilog2(buckets); >> >> >> Isn't it kind of ineffective to first round to a power of two and then >> take the ilog2 of it? >> >> At a first glance I'd say that >> >> ilog2(roundup_pow_of_two(x)) == ilog(x - 1) + 1 >> >> for x > 1. (Maybe even for x == 1? Didn't care to check, I think it >> doesn't matter for the case at hand.) >> >> This RHS might be easier to optimize for the compiler?! > I believe x == 1 needs an extra case since ilog2(0) won't work. > > Since this function (map creation) is not on the fast path, > I currently opt for Linus's suggestion which its code is more > self-explanatory. If you put my suggestion in a macro that is called for example ilog2_roundup() you might have both: easier code and self-explanation. Best regards Uwe
Attachment:
signature.asc
Description: OpenPGP digital signature
