lkml.org 
[lkml]   [2011]   [Dec]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v3 3/5] clk: introduce the common clock framework
On Mon, 5 Dec 2011, Russell King - ARM Linux wrote:

> On Mon, Dec 05, 2011 at 02:15:56PM -0700, Paul Walmsley wrote:
>
> > But I'd propose that we instead increase the size of struct clk.rate to be
> > s64:
> >
> > s64 clk_round_rate(struct clk *clk, s64 desired_rate);
> > int clk_set_rate(struct clk *clk, s64 rate);
> > s64 clk_get_rate(struct clk *clk);
> >
> > struct clk {
> > ...
> > s64 rate;
> > ...
> > };
> >
> > That way the clock framework can accommodate current clock rates, as well
> > as any conceivable future clock rate. (Some production CPUs are already
> > running at clock rates greater than 4 GiHZ[1]. RF devices with 4 GiHz+
> > clock rates are also common, such as 802.11a devices running in the 5.8
> > GHz band, and drivers for those may eventually wish to use the clock
> > framework.)
>
> Yuck. You are aware that 64-bit math on 32-bit CPUs sucks? So burdening
> _everything_ with 64-bit rate quantities is absurd. As for making then
> 64-bit signed quantities, that's asking for horrid code from gcc for the
> majority of cases.

64-bit divides would be the only real issue in the clock framework
context. And those are easily avoided on clock hardware where the parent
rate can never exceed 32 bits.

For example, here's a trivial implementation for rate recalculation for a
integer divider clock node (that can't be handled with a right shift):

s64 div(struct clk *clk, u32 div) {
if (clk->flags & CLK_PARENT_RATE_MAX_U32)
return ((u32)(clk->parent->rate & 0xffffffff)) / div;

clk->rate = clk->parent->rate;
do_div(clk->rate, div);
return clk->rate;
}

gcc generates this for ARMv6:

00000010 <div>:
10: e92d4038 push {r3, r4, r5, lr}
14: e1a05000 mov r5, r0
18: e5d03010 ldrb r3, [r0, #16] @ clk->flags
1c: e1a04001 mov r4, r1
20: e3130001 tst r3, #1 @ 32-bit parent rate?
24: 1a000007 bne 48 <div+0x38> @ do 32-bit divide

/* 64-bit divide by 32-bit follows */

28: e5903000 ldr r3, [r0]
2c: e1c300d8 ldrd r0, [r3, #8]
30: ebfffffe bl 0 <__do_div64>
34: e1a00002 mov r0, r2
38: e1a01003 mov r1, r3
3c: e5852008 str r2, [r5, #8]
40: e585300c str r3, [r5, #12]
44: e8bd8038 pop {r3, r4, r5, pc}

/* 32-bit divide follows */

48: e5900000 ldr r0, [r0]
4c: e5900008 ldr r0, [r0, #8]
50: ebfffffe bl 0 <__aeabi_uidiv>
54: e3a01000 mov r1, #0
58: e8bd8038 pop {r3, r4, r5, pc}


Not bad at all, and this isn't even optimized. (The conditional could be
avoided completely with a little work during clock init.) What little
overhead there is seems quite trivial if it means that the clock framework
will be useful for present and future devices.


- Paul


\
 
 \ /
  Last update: 2011-12-06 02:39    [W:0.675 / U:0.228 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site