Fix wordsize-64 cosh regression (bug 14273)

[Carlos O'Donell]

On 6/21/2012 12:27 PM, Joseph S. Myers wrote:
> This patch fixes bug 14273, spurious cosh overflows that are a
> regression in 2.15 and 2.16 introduced by the dbl-64/wordsize-64
> version of cosh.
> 
> Tested x86_64 and i386; no ulps updates needed, but it's certainly
> possible that some other architecture will need an ulps update, so do
> we wish to defer the testcase until after 2.16 branches and apply only
> the code change to master before then (and apply only the code change
> to 2.15 branch)?
> 
> (The input used in the tests is the result of rounding the input given
> in the bug report to 53-bit precision.)
> 
> 2012-06-21  Joseph Myers  <joseph@codesourcery.com>
> 
> 	[BZ #14273]
> 	* sysdeps/ieee754/dbl-64/wordsize-64/e_cosh.c: Clear sign bit of
> 	64-bit integer value before comparing against overflow value.
> 	* math/libm-test.inc (cosh_test): Add more tests.

I agree.

Fix the bug in 2.16.
Fix the bug in 2.15.
Apply the testcase in 2.17 to prevent regressions.

It's too late in the freeze to introduce a new testsuite regression 
that requires everyone to check for ULPs updates.

Cheers,
Carlos.