Re:There Are Only Four Billion Floats, So Test Them All (2014)
by brucedawson on Thursday September 14, 2017 @10:47PM
(#15254186)
A few hundred random test cases? Double-precision numbers can have 2,048 different exponents. I've seen multiple bugs that only happen with denormals so you need thousands of tests to have any hope of hitting those.
If you're going to do random testing then you need to have a huge set of special numbers (denormals, NaNs, infinity, zero, exact powers of two, one more or less than exact powers of two) and you should do millions or billions of random numbers. Then you can start to feel confident.
Re:There Are Only Four Billion Floats, So Test Them All (2014)
by im3w1l on Friday September 15, 2017 @06:38AM
(#15256003)
>It's worth tuning your random input generator to produce inputs that commonly cause problems like -0.
This would definitely include denormals. As well as +-inf, nan, odd int, even int, negative odd and even int. +-1. Some .5s.
It's worth tuning your random input generator to produce inputs that commonly cause problems like -0.
I can still think of a few situations where exhaustiveness might be worth the effort:
- verifying code passed in from the outside (security considerations?)
- verifying code generated automatically (ie as part of program synthesis)
- verifying very general code you expect to be used in hard-to-predict contexts (ie library routines or compiler optimizations)