Hello back,
On Sun, Apr 10, 2011 at 12:11:56AM +0200, Lluís Batlle i Rossell wrote:
> Further investigations showed that there is some place where 'signaling nan'
> and
> 'quiet nan' are MIXED.
>
> For example, the C++ numeric_limits<double> in mips encodes the quiet nan this
> way: 0x7ff7ffffffffffff
>
> and the signaling nan this way: 0x7fffffffffffffff
The difference above is the most significant bit of the fraction.
I checked the manual (loongson2f user guide), and it says:
"For all floating-point formats, if v is NaN, the most-significant bit of f
determines whether the value is a signaling or quiet NaN: v is a signaling NaN
if the most-significant bit of f is set, otherwise, v is a quiet NaN."
So it means that the values in numeric_limits are correct according to the
manual.
But further checking with the C++ program below (which you have to modify to
make it use madd with a quiet nan passing the 'f') showed that all 'extra fp
instructions' look affected by this: madd.d, msub.d, nmadd.d and nmsub.d. Only
when having a qnan in any of the registers. The a-number cases, and signaling
nan, work without signal.
Gcc rendered *and called* those instructions only with "-O1". Trying other -O
values, that got into different instructions, and *none of them* got a signal,
either it was a signaling or a non-signalin nan.
Regards,
Lluís
> Here is the C++ code I used. Change it at will to test:
> ----------------
> #include <limits>
> #include <math.h>
> #include <iostream>
>
> double madd(double a, double b, double c)
> {
> return a*b+c;
> }
>
> int main()
> {
> double z = 0;
> double f = std::numeric_limits<double>::quiet_NaN();
>
> double g = madd(z, 1, 1);
>
> std::cout << "primer: " << g << std::endl;
>
> f = std::numeric_limits<double>::signaling_NaN();
>
> g = madd(z, f, f);
>
> std::cout << "segon: " << g << std::endl;
>
> return 0;
> }
> ----------------
--
You received this message because you are subscribed to the Google Groups
"loongson-dev" group.
To post to this group, send email to loongson-dev@googlegroups.com.
To unsubscribe from this group, send email to
loongson-dev+unsubscr...@googlegroups.com.
For more options, visit this group at
http://groups.google.com/group/loongson-dev?hl=en.
