Re: [PATCH][GCC][ARM] Adjust costs so udiv is preferred over sdiv when both are valid. [Patch (2/2)]

[Kyrill Tkachov]

On 02/05/17 16:37, 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.

Similar to aarch64 this patch along with my other two related mid-end changes
makes a big difference in division by constants.

Given:

int f2(int x)
{
  return ((x * x) % 300) + ((x * x) / 300);
}

we now generate

f2:
        mul     r3, r0, r0
        mov     r0, r3
        ldr     r1, .L3
        umull   r2, r3, r0, r1
        lsr     r2, r3, #5
        add     r3, r2, r2, lsl #2
        rsb     r3, r3, r3, lsl #4
        sub     r0, r0, r3, lsl #2
        add     r0, r0, r2
        bx      lr

as opposed to

f2:
        mul     r3, r0, r0
        mov     r0, r3
        ldr     r3, .L4
        push    {r4, r5}
        smull   r4, r5, r0, r3
        asr     r3, r0, #31
        rsb     r3, r3, r5, asr #5
        add     r2, r3, r3, lsl #2
        rsb     r2, r2, r2, lsl #4
        sub     r0, r0, r2, lsl #2
        add     r0, r0, r3
        pop     {r4, r5}
        bx      lr

Bootstrapped and reg tested on arm-none-eabi
with no regressions.

OK for trunk?

Thanks,
Tamar


gcc/
2017-05-02  Tamar Christina  <tamar.christ...@arm.com>

        * config/arm/arm.c (arm_rtx_costs_internal): Make sdiv more expensive 
than udiv.


gcc/testsuite/
2017-05-02  Tamar Christina  <tamar.christ...@arm.com>

        * gcc.target/arm/sdiv_costs_1.c: New.

diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index 
b24143e32e2f100000f3b150f7ed0df4fabb3cc8..ecc7688b1db6309a4dd694a8e254e64abe14d7e3
 100644
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -9258,6 +9258,8 @@ arm_rtx_costs_internal (rtx x, enum rtx_code code, enum 
rtx_code outer_code,
        *cost += COSTS_N_INSNS (speed_p ? extra_cost->mult[0].idiv : 0);
       else
        *cost = LIBCALL_COST (2);
+
+      *cost += (code == DIV ? 1 : 0);
       return false;    /* All arguments must be in registers.  */
 

We usually try to avoid adjusting the costs in units other than COSTS_N_INSNS.
Would adding COSTS_N_INSNS (1) here work?
If so, could you also add a comment here to describe why we're adjusting the 
cost.

     case MOD:
@@ -9280,7 +9282,7 @@ arm_rtx_costs_internal (rtx x, enum rtx_code code, enum 
rtx_code outer_code,
 
     /* Fall-through.  */
     case UMOD:
-      *cost = LIBCALL_COST (2);
+      *cost = LIBCALL_COST (2) + (code == MOD ? 1 : 0);

Same here.

Thanks,
Kyrill