[Bug tree-optimization/90838] Detect table-based ctz implementation

["wilco at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90838

--- Comment #17 from Wilco <wilco at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #16)
> (In reply to Wilco from comment #15)
> > It would make more sense to move x86 backends to CTZ_DEFINED_VALUE_AT_ZERO
> > == 2 so that you always get the same result even when you don't have tzcnt.
> > A conditional move would be possible, so it adds an extra 2 instructions at
> > worst (ie. still significantly faster than doing the table lookup, multiply
> > etc). And it could be optimized when you know CLZ/CTZ input is non-zero.
> 
> Conditional moves are a lottery on x86, in many cases very bad idea.  And
> when people actually use __builtin_clz*, they state that they don't care
> about the 0 value, so emitting terribly performing code for it just in case
> would be wrong.
> If forwprop emits the conditional in separate blocks for the CTZ_DVAZ!=2
> case, on targets where conditional moves are beneficial for it it can also
> emit them, or emit the jump which say on x86 will be most likely faster than
> cmov.

Well GCC emits a cmov for this (-O2 -march=x86-64-v2):

int ctz(long a)
{
  return (a == 0) ? 64 : __builtin_ctzl (a);
}

ctz:
        xor     edx, edx
        mov     eax, 64
        rep bsf rdx, rdi
        test    rdi, rdi
        cmovne  eax, edx
        ret

Note the extra 'test' seems redundant since IIRC bsf sets Z=1 if the input is
zero.

On Zen 2 this has identical performance as the plain builtin when you loop it
as res = ctz (res) + 1; (ie. measuring latency of non-zero case). So I find it
hard to believe cmov is expensive on modern cores.