Re: [PATCH v2] rs6000: fmr gets used instead of faster xxlor [PR93571]

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

Hi!

On Sat, Feb 25, 2023 at 03:20:33PM +0530, Ajit Agarwal wrote:
> Here is the patch that uses xxlor instead of fmr where possible.
> Performance results shows that fmr is better in power9 and 
> power10 architectures whereas xxlor is better in power7 and
> power 8 architectures. fmr is the only option before p7.

>  ;; The ISA we implement.
> -(define_attr "isa" "any,p5,p6,p7,p7v,p8v,p9,p9v,p9kf,p9tf,p10"
> +(define_attr "isa" "any,p5,p6,p7,p7v,p8v,p7p8v,p9,p9v,p9kf,p9tf,p10"
>    (const_string "any"))

This isn't really about what insn we *can* use here.

> +     (and (eq_attr "isa" "p7p8v")
> +	  (match_test "TARGET_VSX && !TARGET_P9_VECTOR"))
> +     (const_int 1)

What is needed here is test the *tune* setting.  For example if someone
uses -mcpu=power8 -mtune=power9 (this is a setting that is really used,
or was a few years ago anyway) you *do* want fmr insns generated.

So don't do this via the isa attribute at all, just add some insn
condition (testing the tune setting)?


Segher