On 2019-Feb-14, Tom Lane wrote:
> I'd bet a fair amount of money that we'd be better off *not* using
> lzcnt, even if available, because then we could just expose things
> along this line:
>
> static inline int
> pg_clz(...)
> {
> #ifdef HAVE__BUILTIN_CLZ
> return __builtin_clz(x);
> #else
> handwritten implementation;
> #endif
> }
>
> Avoiding a function call (that has to indirect through a pointer) probably
> saves much more than the difference between lzcnt and the other way.
I see ... makes sense.
That leads me to the attached patch. It creates a new file
pg_popcount.c which is the only one compiled with -mpopcnt (if
available); if there's no compiler switch to enable POPCNT, we just
don't compile the file. I'm not sure that's kosher -- in particular I'm
not sure if it can fail when POPCNT is enabled by other flags and
-mpopcnt is not needed at all. I think our c-compiler.m4 stuff is a bit
too simplistic there: it just assumes that -mpopcnt is always required.
But what if the user passes it in CFLAGS?
I left CPUID alone for the CLZ/CTZ builtins. So we either use the
table, or the builtins. We never try the instructions.
--
Álvaro Herrera https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services