Re: [PATCH bpf-next 7/7] nfp: bpf: migrate to advanced reciprocal divide in reciprocal_div.h

[Jiong Wang <jiong.wang@netronome.com>]

On 26/06/2018 21:59, Jakub Kicinski wrote:
> On Sun, 24 Jun 2018 20:54:21 -0700, Jakub Kicinski wrote:
>> +	 * NOTE: because we are using "reciprocal_value_adv" which doesn't
>> +	 * support dividend with MSB set, so we need to JIT separate NFP
>> +	 * sequence to handle such case. It could be a simple sequence if there
>> +	 * is conditional move, however there isn't for NFP. So, we don't bother
>> +	 * generating compare-if-set-branch sequence by rejecting the program
>> +	 * straight away when the u32 dividend has MSB set. Divide by such a
>> +	 * large constant would be rare in practice. Also, the programmer could
>> +	 * simply rewrite it as "result = divisor >= the_const".
> Thinking about this again, can we just use carry bit?

Good catch, yes we can.

> The code may end
> up shorter than the explanation why we don't support that case :P
>
> immed[c, 0]
> alu[--, a, -, b]
> alu[c, c, +carry, 0]

eBPF input will be "a = a / b", given "immed" doesn't affect carry bit,
I'd reorder the sequence so we only need one tmp register for holding
"b" who is constant.

   alu[--, a, -, b]
   immed[b, 0]
   alu[a, b, +carry, 0]
  
Thanks.
Regards,
Jiong

>
> Should be equivalent to:
>
> c = a >= b
>
> (Thanks to Edwin for double-checking the carry semantics.)