On Tue, May 02, 2017 at 04:37:16PM +0100, Tamar Christina wrote:
> Hi All,
>
> This patch adjusts the cost model so that when both sdiv and udiv are possible
> it prefers udiv over sdiv. This was done by making sdiv slightly more
> expensive
> instead of making udiv cheaper to keep the baseline costs of a division the
> same
> as before.
>
> For aarch64 this patch along with my other two related mid-end changes
> makes a big difference in division by constants.
This patch seems to have an unrelated change to the MOD/UMOD costs to delete
the handling of floating-point values. That change makes sense, but would
have been better in a separate patch.
> Given:
>
> int f2(int x)
> {
> return ((x * x) % 300) + ((x * x) / 300);
> }
>
> we now generate
>
> f2:
> mul w0, w0, w0
> mov w1, 33205
> movk w1, 0x1b4e, lsl 16
> mov w2, 300
> umull x1, w0, w1
> lsr x1, x1, 37
> msub w0, w1, w2, w0
> add w0, w0, w1
> ret
>
> as opposed to
>
> f2:
> mul w0, w0, w0
> mov w2, 33205
> movk w2, 0x1b4e, lsl 16
> mov w3, 300
> smull x1, w0, w2
> umull x2, w0, w2
> asr x1, x1, 37
> sub w1, w1, w0, asr 31
> lsr x2, x2, 37
> msub w0, w1, w3, w0
> add w0, w0, w2
> ret
>
> Bootstrapped and reg tested on aarch64-none-linux-gnu with no regressions.
>
> OK for trunk?
OK.
Thanks,
James
> gcc/
> 2017-05-02 Tamar Christina <tamar.christ...@arm.com>
>
> * config/aarch64/aarch64.c (aarch64_rtx_costs): Make sdiv more
> expensive than udiv.
> Remove floating point cases from mod.
>
> gcc/testsuite/
> 2017-05-02 Tamar Christina <tamar.christ...@arm.com>
>
> * gcc.target/aarch64/sdiv_costs_1.c: New.
