svn commit: r300949 - head/sys/netpfil/ipfw

[Don Lewis]

On 29 May, Bruce Evans wrote:
> On Sun, 29 May 2016, Don Lewis wrote:
> 
>> Log:
>>  Cast some expressions that multiply a long long constant by a
>>  floating point constant to int64_t.  This avoids the runtime
>>  conversion of the the other operand in a set of comparisons from
>>  int64_t to floating point and doing the comparisions in floating
>>  point.
>>
>>  Suggested by:	lidl
>>  Submitted by:	Rasool Al-Saadi <ralsaadi at swin.edu.au>
>>  MFC after:	2 weeks (with r300779)
> 
> Compilers are still permitted to (and perhaps even required to)
> evaluate FP constant expressions at runtime (to get rounding and/or
> exception flags right).  They probably don't in practice, but it is
> unclear what happens for -O0 and the rules for rounding are too hard
> to understand.
> 
>> Modified: head/sys/netpfil/ipfw/dn_aqm_pie.c
>> ==============================================================================
>> --- head/sys/netpfil/ipfw/dn_aqm_pie.c	Sun May 29 07:14:51 2016	(r300948)
>> +++ head/sys/netpfil/ipfw/dn_aqm_pie.c	Sun May 29 07:23:56 2016	(r300949)
>> @@ -244,17 +244,17 @@ calculate_drop_prob(void *x)
>> 	p *= (PIE_MAX_PROB << 12) / AQM_TIME_1S;
>>
>> 	/* auto-tune drop probability */
>> -	if (prob< PIE_MAX_PROB * 0.000001)
>> +	if (prob < (int64_t)(PIE_MAX_PROB * 0.000001))
>> 		p >>= 11 + PIE_FIX_POINT_BITS+12;
>> +	else if (prob < (int64_t)(PIE_MAX_PROB * 0.00001))
>> 		p >>= 9 + PIE_FIX_POINT_BITS+12;
> 
> Why not just divide by integer powers of 10?
> 
> This might not give a suitably monotonic/continuous scaling at the
> endpoints, but it is unclear if the FP gives that either even if we
> are more careful with the rounding mode.
> 
> A table of endpoints could be used to get precise control.  Then FP
> can be used more safely, since it is clear that constants in tables
> must be evaluated at compile time.
> 
>> ...
> 
> Similarly for all cases.

Thanks.  I've passed this upstream.