patch with larger errors for "float"

[Joseph Myers]

On Mon, 20 Jan 2020, Carlos O'Donell wrote:

> (1) Does it make sense to add more input data points for the tests?
> 
> What you are suggestion with your patch makes sense, and now your patch
> includes the input, output, and ULPs adjustment which makes it a complete
> patch that should pass the regression testing on x86_64.

FWIW, I previously added some test inputs based on random testing (but 
random testing with many fewer inputs).

> (2) Are we changing all the things we need to change?
> 
> Yes, you add inputs to 10 functions, and then add outputs to 10 function
> files for testing. That looks correct.

However, there are libm-test-ulps updates for only 6 functions (not for 
atan2, cosh, erf or sin).  Presuming the tests do indeed pass with those 
updates for only 6 functions, maybe this indicates some difference in the 
details of how glibc measures the error in a function versus how MPCHECK 
does?  (I think adding the extra tests for the other 4 functions is fine 
even if they don't involve increased errors as measured by glibc.)

-- 
Joseph S. Myers
joseph@codesourcery.com