Hi,
On 2019-02-14 16:45:38 -0500, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > On 2019-02-14 15:47:13 -0300, Alvaro Herrera wrote:
> >> Hah, I just realized you have to add -mlzcnt in order for these builtins
> >> to use the lzcnt instructions. It goes from something like
> >>
> >> bsrq %rax, %rax
> >> xorq $63, %rax
>
> > I'm confused how this is a general count leading zero operation? Did you
> > use constants or something that allowed ot infer a range in the test? If
> > so the compiler probably did some optimizations allowing it to do the
> > above.
>
> No. If you compile
>
> int myclz(unsigned long long x)
> {
> return __builtin_clzll(x);
> }
>
> at -O2, on just about any x86_64 gcc, you will get
>
> myclz:
> .LFB1:
> .cfi_startproc
> bsrq %rdi, %rax
> xorq $63, %rax
> ret
> .cfi_endproc
Yea, sorry for the noise. I misremembered the bsrq mnemonic.
bsr has a latency of three cycles, xor of one. lzcnt a latency of
three. So it's mildly faster to use lzcnt (it uses fewer ports, and has
a shorter latency). But I doubt we have code where that's noticable.
Greetings,
Andres Freund