Re: [PATCH V4] rs6000: Optimize cmp on rotated 16bits constant

[Segher Boessenkool <segher@kernel.crashing.org>]

Hi!

On Thu, Aug 25, 2022 at 08:11:31PM +0800, Jiufu Guo wrote:
> Segher Boessenkool <segher@kernel.crashing.org> writes:
> > You usually can split fine if you cannot create new pseudos, by reusing
> > existing registers.
> >
> > FAIL will cause an ICE: the RTL instruction does match, but will fail
> > when trying to generate machine code for it.
> >
> Previous patch is using "gen_reg_rtx (DImode)" to generate a pseudo for
> the rotated result to prevent orignal one being changed accidently.
> So, an 'assert (can_create_pseudo_p ())' would catch it in after RA.

It sounds like you want a define_split, not a define_insn_and_split.
That is much more stomachable anyway.

Anything that creates conditional branches together with compars insns
belongs before RA, before sched1 even.

> To enable this splitter works after RA, we may need to reserve one
> register (clobber would be ok).  Such as below:
> 
>   [(set (pc)
> 	(if_then_else (eqne (match_operand:DI 1 "gpc_reg_operand" "r")
> 			    (match_operand:DI 2 "const_int_operand" "n"))
> 		      (label_ref (match_operand 0 ""))
> 		      (pc)))
>   (clobber (match_scratch:DI 3 "=r"))
>   (clobber (match_scratch:CCUNS 4 "=y"))]

Yes, that is one way to do it.  Another way is to reuse operand 1.  A
clobber is probably better in this case though :-)

If this is only so combine can match things, you certainly want just a
define_split, and the compare+branch in one pattern is not as bad then.


Segher