[U-Boot] [PATCH 3/4 V2] OneNAND: Add simple OneNAND SPL

[Scott Wood <scottwood@freescale.com>]

On 11/03/2011 11:56 AM, Marek Vasut wrote:
>> On 11/02/2011 07:15 PM, Marek Vasut wrote:
>>>> On 11/01/2011 05:54 PM, Marek Vasut wrote:
>>>>> +static void spl_onenand_get_geometry(struct spl_onenand_data *data)
>>>>> +{
>>>
>>> [...]
>>>
>>>>> +	/* The page can be either 2k or 4k, avoid using DIV_ROUND_UP. */
>>>>> +	if (data.pagesize == 2048) {
>>>>> +		total_pages = len / 2048;
>>>>> +		page = offset / 2048;
>>>>> +		total_pages += !!(len & 2047);
>>>>> +	} else if (data.pagesize == 4096) {
>>>>> +		total_pages = len / 4096;
>>>>> +		page = offset / 4096;
>>>>> +		total_pages += !!(len & 4095);
>>>>> +	}
>>>>
>>>> What's wrong with DIV_ROUND_UP?  It should produce smaller code than
>>>> what you've done here...
>>>
>>> It pulls in aeabi_*div* functions, which won't fit into block 0 of
>>> Onenand.
>>
>> It shouldn't do that if the divisor is a constant power of 2.  The
>> compiler will turn it into a shift, just like with the other divides in
>> the above code fragment.
>>
>> You can't use DIV_ROUND_UP directly on data.pagesize, but you can use it
>> in each branch of the if statement instead of that awkward and slightly
>> more expensive !!(len & 4095) construct.
> 
> Expensive in what way?

Compare the resulting asm code.  You're replacing this:

	a = (b + 4095) >> 12;

with this:

	a = b >> 12;
	if (b & 4095)
		a++;

>  Either way, I don't think this matters that much.

This is code that has to fit in 1K -- why waste instructions by writing
code in a way that is *less* readable?

-Scott