Re: [PATCH][GCC][AARCH64] Add even-pair register classes

[Matthew Malcomson <matthew dot malcomson at arm dot com>]

Hi Wilko,

On 28/09/18 13:33, Wilco Dijkstra wrote:
> Matthew wrote:
>
> > The canonical way to require even-odd pairs of registers to implement a TImode
> > pseudo register as mentioned in the documentation is to limit *all* TImode
> > registers to being even-odd by using the TARGET_HARD_REGNO_MODE_OK hook.
> And that is the best approach for cases like this. Adding new register classes is rarely
> a good idea since they complicate matters and slow down the register allocator. We've
> had several cases where a new register class caused unexpected code quality issues.
> How do we know that won't happen this time?
I didn't know that a new register class could often cause code quality 
issues
and slower compilation, I guess this is only when the corresponding 
constraint
is being used?
Or can it affect other patterns?
> > We don't want to use this method as it would apply to all register pairs.
> Why not? From our Arm/Thumb-2 experiments on DImode splitting and allocation we
> know even/odd pairs give a better allocation, and there are fewer cases where the
> allocation completely fails under high register pressure.
Do the Arm/Thumb-2 experiments you mention imply that it would be better to
force TImode to be put in even/odd pairs anyway, or would the ira allocation
preference already provide enough incentive?
I guess this would be influenced by DImode being used quite often in
Arm/Thumb-2?
> > (It is noteworthy that the current cost model in ira prefers to put values that
> > span two registers in an even-odd pair anyway, so the use of
> > TARGET_HARD_REGNO_MODE_OK would likely not cause much change in low register
> > pressure situations)
> Exactly. TImode is not used that much (unlike DImode on Arm/Thumb-2), so it would
> be hard to find any difference.
>
> So I would strongly suggest to keep things simple and just use even/odd pairs. Adding a
> new register class is generally not worth it unless you can show it will significantly improve
> code quality.
>
> Wilco
Can I clarify whether you're saying "limiting all register pairs to even/odd
might overall be more beneficial than just limiting specific patterns" 
or "there
probably won't be much difference between the two, and introducing a new
register class isn't worth the complexity"?
I initially thought you meant the second statement (because that was 
something
that I wondered about already), but am not certain.

Thanks,
M