Re: Optimize three-valued comparison between integers

[Bruno Haible]

Florian Weimer wrote:
> On s390x, all three variants use conditional
> branches, but the first one is the shortest.

Indeed. Surprising.

> There is also this:
> 
> int sign4 (long n1, long n2)
> {
>   return (char) ((n1 > n2) - (n1 < n2));
> }

The case should be towards (signed char). Otherwise, on arm-linux-gnueabi
and powerpc-linux-gnu the 3 values are 0, 1, 255, not 0, 1, -1.

> It's one instruction shorter on x86-64 than sign3, but it's worse on
> other architectures.  (One of the x86-64 quirks is that conditional sets
> do not affect the full register, only a byte.)

Indeed. But IMO, the conditional jumps matter more than an immediate
instruction.

Bruno