Re: Use compiler intrinsics for bit ops in hash

Hi David,
On Sun, Mar 8, 2020 at 11:34 AM David Fetter <david(at)fetter(dot)org> wrote:
>
> On Mon, Mar 02, 2020 at 12:45:21PM -0800, Jesse Zhang wrote:
> > Hi David,
>
> Per discussion on IRC with Andrew (RhodiumToad) Gierth:
>
> The runtime detection means there's always an indirect call overhead
> and no way to inline. This is counter to what using compiler
> intrinsics is supposed to do.
>
> It's better to rely on the compiler, because:
> (a) The compiler often knows whether the value can or can't be 0 and
> can therefore skip a conditional jump.

Yes, the compiler would know to eliminate the branch if the inlined
function is called with a literal argument, or it infers an invariant
from the context (like nesting inside a conditional block, or a previous
conditional "noreturn" path).

> (b) If you're targeting a recent microarchitecture, the compiler can
> just use the right instruction.

I might be more conservative than you are on (b). The thought of
building a binary that cannot run "somewhere" where the compiler
supports by default still mortifies me.

> (c) Even if the conditional branch is left in, it's not a big overhead.
>

I 100% agree with (c), see benchmarking results upthread.

Cheers,
Jesse