Hacker News Reader

Comment by mrcode007 | Hacker News Reader

It uses SWAR technique for counting bits because target platform wasn’t specified in compiler flags properly. popcnt and family are specific family extensions and without specifying the target flags the compiler generates lowest common denominator among available instructions

Edit: add -msse4 and you will see popcnt appear because popcnt is the SSE4 extension

View on HN

I think the issue that the article is not explaining very well, is that what it's actually trying to do is check if a number is a power of two, and one part of the compiler says "use popcnt!" and then later another part says "oh sorry, we'll have to emulate that with a ton of instructions". Whereas it could have just used a shift and and.

The power of two test is simply

x & (x-1) == 0

which could be special cased through templates too

> It was a Clang compiler, ARM-based platform

> I do think it is a bug in this case, native popcnt could maybe be defended, but should not be used if compiling for architecture that does not support it.

> If you would like to experiment yourself, Compiler Explorer snippet can be found here

https://gcc.godbolt.org/z/qGzWo39b6

Which is an x86-64 sample

It's a little impossible to show the difference between using popcnt and the expanded emulation code on a platform that doesn't support popcnt.

It doesn't change the fact that the actual program itself was being complied on a platform that doesn't support popcnt, so no amount of compiler arguments would fix that.

NEON extension on ARM comes with CNT instruction that counts set bits.

The article says all of this?

Reply on news.ycombinator.com