RE: Fix the inaccuracy of j0f/y0f/j1f/y1f

[Joseph Myers <joseph@codesourcery.com>]

On Tue, 23 Feb 2021, Michael Morrell wrote:

> Paul,
> 
> You state that the IEEE 754 standard doesn't require correct rounding, 
> but I don't see where it says that.  I see Section 9.2 ("Additional 
> mathematical operations") and the list in Table 9.1 which seems to 
> include most of the math functions in glibc (I notice that the Bessel 
> functions like y0 aren't included) and these all must be correctly 
> rounded.  I agree that a language may choose not to define all of the 
> operations in 9.1, but that doesn't seem to give it the right to 
> implement the function inaccurately.

If you look at the current C2x draft N2596, 5.2.4.2.2 paragraph 8 says 
that the accuracy of the library functions is implementation-defined.  
Annex F says how C operations relate to operations defined in IEEE 754; 
the table in F.3 paragraph 1 includes an entry for sqrt, for example, so 
that corresponds to the IEEE squareRoot operation, but doesn't include 
most of the math.h functions; paragraph 20 gives the mapping for other 
functions and says that correct rounding is not required.

-- 
Joseph S. Myers
joseph@codesourcery.com