On 04/01/2020 00:49, Tom Lane wrote:
> Vik Fearing <vik.fearing@2ndquadrant.com> writes:
>> On 03/01/2020 20:14, Fabien COELHO wrote:
>>> The point of swapping is to a void possibly expensive modulo, but this
>>> should be done on absolute values, otherwise it may not achieve its
>>> purpose as stated by the comment?
>> Ah, true. How widespread are these architectures that need this special
>> treatment? Is it really worth handling?
> On some older RISC architectures, integer division is really slow, like
> slower than floating-point. I'm not sure if that's true on any platform
> people still care about though. In recent years, CPU architects have been
> able to throw all the transistors they needed at such problems. On a
> machine with single-cycle divide, it's likely that the extra
> compare-and-branch is a net loss.
OK.
> Might be worth checking it on ARM in particular, as being a RISC
> architecture that's still popular.
I don't know how I would check this.
> Also, if we end up having a "numeric" implementation, it absolutely is
> worth it for that, because there is nothing cheap about numeric_div.
The patch includes a numeric version, and I take care to short-circuit
everything I can.
> I'd be sort of inclined to have the swap in the other implementations
> just to keep the algorithms as much alike as possible.
They can't quite be the same behavior because numeric doesn't have the
unrepresentable -INT_MIN problem, and integers don't have NaN.
--
Vik Fearing