Re: floating point accuracy, was Compiler bugs

(Christian Bau) wrote thusly:
> Or let my ask this question: The fdlibm functions for sine/cosine
> contain some huge tables. If one bit in these tables were incorrect,
> producing completely wrong results for x > 1e100 (but still in the range
> -1 to +1), would anyone notice?
>


The answer, sadly, is probably not.


In support of that, I'll point out that I once worked for a herein
unnamed (but at the time well-established and respected) "scientific"
computer vendor that released a version of their math library with the
sine and cosine routines reversed. Nobody, including a large
scientific customer base, noticed for quite some time.


Turns out, both sin(x) and cos(x) were tested over large ranges of x via the
"(((sin(x)**2)*(cos(x)**2)).eq.1)" identity... and both routines were quite
accurate, just completely wrong. No individual values of either sin(x) or
cos(x) were ever examined for correctness by the QA people.


Hey, Bob! The sine of 60 degrees is 0.5, right?'dily,


-S

Return to the comp.compilers page.
Search the comp.compilers archives again.