[PATCH] clk: k210: Fix error calculation on 32bit

[Sean Anderson]

On 10/16/22 10:50, Heinrich Schuchardt wrote:
> 
> 
> On 10/16/22 16:48, Sean Anderson wrote:
>> On 10/16/22 10:40, Heinrich Schuchardt wrote:
>>>
>>>
>>> On 10/16/22 16:19, Sean Anderson wrote:
>>>> On 10/16/22 07:46, Heinrich Schuchardt wrote:
>>>>>
>>>>>
>>>>> On 10/16/22 09:51, Michal Suchánek wrote:
>>>>>> On Thu, Oct 13, 2022 at 10:34:29PM +0200, Michal Suchanek wrote:
>>>>>>> k210 is 64bit but the driver and tests are also built in sandbox, and
>>>>>>> that can be built 32bit.
>>>>>>>
>>>>>>> BIT(32) does not work on 32bit, shift before subtraction to fit into
>>>>>>> 32bit integer with BIT values.
>>>>>>
>>>>>>
>>>>>> Also see
>>>>>> https://patchwork.ozlabs.org/project/uboot/patch/20221016071035.461454-1-heinrich.schuchardt@canonical.com/
>>>>>>>
>>>>>>> Signed-off-by: Michal Suchanek <msuchanek at suse.de>
>>>>>>> ---
>>>>>>>
>>>>>>>   drivers/clk/clk_k210.c | 2 +-
>>>>>>>   test/dm/k210_pll.c     | 2 +-
>>>>>>>   2 files changed, 2 insertions(+), 2 deletions(-)
>>>>>>>
>>>>>>> diff --git a/drivers/clk/clk_k210.c b/drivers/clk/clk_k210.c
>>>>>>> index 1961efaa5e..e85f8ae14a 100644
>>>>>>> --- a/drivers/clk/clk_k210.c
>>>>>>> +++ b/drivers/clk/clk_k210.c
>>>>>>> @@ -846,7 +846,7 @@ again:
>>>>>>>           error = DIV_ROUND_CLOSEST_ULL(f * inv_ratio, r * od);
>>>>>>>           /* The lower 16 bits are spurious */
>>>>>>> -        error = abs((error - BIT(32))) >> 16;
>>>>>>> +        error = abs((error >> 16) - BIT(32 - 16));
>>>>>
>>>>>
>>>>> Your patch changes the results. Replacing BIT() by BIT_ULL() does not change the result:
>>>>>
>>>>> error = -1 = 0xffffffffffffffff
>>>>> before
>>>>> error - BIT(32) = -4294967297 = 0xfffffffeffffffff
>>>>> abs(error - BIT(32)) = 4294967297 = 0x100000001
>>>>> abs(error - BIT(32)) >> 16 = 65536 = 0x10000
>>>>> after
>>>>> error >> 16 = -1 = 0xffffffffffffffff
>>>>> (error >> 16) - BIT(32 - 16) = -65537 = 0xfffffffffffeffff
>>>>> abs((error >> 16) - BIT(32 - 16)) = 65537 = 0x10001
>>>>> using BIT_ULL
>>>>> abs(error - BIT_ULL(32)) >> 16 = 65536 = 0x10000
>>>>>
>>>>> error = 70000 = 0x11170
>>>>> before
>>>>> error - BIT(32) = -4294897296 = 0xffffffff00011170
>>>>> abs(error - BIT(32)) = 4294897296 = 0xfffeee90
>>>>> abs(error - BIT(32)) >> 16 = 65534 = 0xfffe
>>>>> after
>>>>> error >> 16 = 1 = 0x1
>>>>> (error >> 16) - BIT(32 - 16) = -65535 = 0xffffffffffff0001
>>>>> abs((error >> 16) - BIT(32 - 16)) = 65535 = 0xffff
>>>>> using BIT_ULL
>>>>> abs(error - BIT_ULL(32)) >> 16 = 65534 = 0xfffe
>>>>>
>>>>> I assume this change of results is unintended.
>>>>>
>>>>> We also replace abs() by abs64() to get correct results on a 32bit system:
>>>>>
>>>>> error = -1 = 0xffffffffffffffff
>>>>> abs(error - BIT_ULL(32)) >> 16 = 0 = 0x0
>>>>> abs64(error - BIT_ULL(32)) >> 16 = 65536 = 0x10000
>>>>> error = 70000 = 0x11170
>>>>> abs(error - BIT_ULL(32)) >> 16 = 1 = 0x1
>>>>> abs64(error - BIT_ULL(32)) >> 16 = 65534 = 0xfffe
>>>>>
>>>>> Best regards
>>>>>
>>>>> Heinrich
>>>>>
>>>>>>>           if (error < best_error) {
>>>>>>>               best->r = r;
>>>>>>> diff --git a/test/dm/k210_pll.c b/test/dm/k210_pll.c
>>>>>>> index a0cc84c396..622b1c9bef 100644
>>>>>>> --- a/test/dm/k210_pll.c
>>>>>>> +++ b/test/dm/k210_pll.c
>>>>>>> @@ -33,7 +33,7 @@ static int dm_test_k210_pll_calc_config(u32 rate, u32 rate_in,
>>>>>>>                   error = DIV_ROUND_CLOSEST_ULL(f * inv_ratio,
>>>>>>>                                     r * od);
>>>>>>>                   /* The lower 16 bits are spurious */
>>>>>>> -                error = abs((error - BIT(32))) >> 16;
>>>>>>> +                error = abs((error >> 16) - BIT(32 - 16));
>>>>>>>                   if (error < best_error) {
>>>>>>>                       best->r = r;
>>>>>>>                       best->f = f;
>>>>>>> -- 
>>>>>>> 2.37.3
>>>>>>>
>>>>
>>>> IMO the driver should just be changed to depend on 64-bit. The k210 is 64-bit,
>>>> and I didn't write anything with 32-bit in mind.
>>>
>>> Michal, Simon, and I are striving to get the sandbox working on ilp32 systems.
>>>
>>> Do you suggest to remove  the driver from sandbox_defconfig?
>>>
>>> This would imply that the unit test is not executed on Gitlab CI. You will still be able to execute it on the actual hardware.
>>
>> It's still enabled for sandbox64, so it should still be executed there.
> 
> No, the sandbox64 is also to be compiled on ilp32. It models physical address extension (PAE).

Do we have a config for word size? PHYS_64BIT seems to be for the address size.

--Sean