[Haskell-cafe] Optimizing Eq instances with reallyUnsafePtrEquality#

[Joachim Durchholz]

Am 08.06.21 um 16:25 schrieb Simon Jakobi via Haskell-Cafe:
> Hi everyone!
> 
> In https://github.com/haskell-unordered-containers/unordered-containers/issues/77
> we're wondering whether certain Eq instances, for example record types
> or strings, could be optimized by including a pointer equality check
> that detects when an object is compared with itself.

You have to be aware that the pointer comparison itself does not come 
for free. Execution prediction means that the "happy path" may be so 
rare it's not loaded into the CPU cache and you might end with a slower 
system - the case is somewhat pathological but not so rare that you can 
just assume that it will not happen to you.
A lot also depends on whether the data to be compared needs to be loaded 
anyway. If yes, the pointer comparison won't give you any gains, it will 
be just one instruction more to execute, creating more execution unit 
contention inside the CPU. If no, then obviously the pointer comparison 
will help.

In the end, you need to benchmark to see if it helps you.
And you cannot usefully benchmark unless you have also nailed down all 
performance-relevant compiler and runtime options, which you do only if 
you have exhausted all other optimization possibilities.

IOW if it complicates your code, don't do it - unless you are already 
long past the point where code reusability has taken a back seat to raw 
optimization.

Regards,
Jo