Re: [PATCH] [Aarch64] Optimize subtract in shift counts

[Andrew Pinski <pinskia@gmail.com>]

On Mon, Aug 7, 2017 at 1:36 PM, Michael Collison
<Michael.Collison@arm.com> wrote:
> This patch improves code generation for shifts with subtract instructions where the first operand to the subtract is equal to the bit-size of the operation.
>
>
> long f1(long x, int i)
> {
>   return x >> (64 - i);
> }
>
> int f2(int x, int i)
> {
>   return x << (32 - i);
> }
>
>
> With trunk at -O2 we generate:
>
> f1:
>         mov     w2, 64
>         sub     w1, w2, w1
>         asr     x0, x0, x1
>         ret
>
> f2:
>         mov     w2, 32
>         sub     w1, w2, w1
>         lsl     w0, w0, w1
>         ret
>
> with the patch we generate:
>
> f1:
>         neg     w2, w1
>         asr     x0, x0, x2
>         ret
>         .size   f1, .-f1
>         .align  2
>         .p2align 3,,7
>         .global f2
>         .type   f2, %function
> f2:
>         neg     w2, w1
>         lsl     w0, w0, w2
>         ret
>
> Okay for trunk?


Shouldn't this be handled in simplify-rtx instead of an aarch64
specific pattern?

That is simplify:
(SHIFT A (32 - B)) -> (SHIFT A (AND (NEG B) 31))
etc.

or maybe not.  I don't mind either way after thinking about it more.

Thanks,
Andrew

>
> 2017-08-07  Michael Collison <michael.collison@arm.com>
>
>         * config/aarch64/aarch64.md (*aarch64_reg_<optab>_minus<mode>3):
>         New pattern.
>
> 2016-08-07  Michael Collison <michael.collison@arm.com>
>
>         * gcc.target/aarch64/var_shift_mask_2.c: New test.