Reduce various x86/x86_64 libm ulps expectations

[Andreas Jaeger]

On 02/29/2012 05:49 PM, Joseph S. Myers wrote:
> Some libm tests for x86 and x86_64 long double had huge error
> expectations - hundreds of ulps, where otherwise the maximum expected
> error was 25ulp.
>
> I did some testing with those expectations removed, and actually I
> don't see those huge errors.  So I propose this patch that reduces
> various expectations based on my testing.  Variations between
> different x86 processors may require some of these to be increased a
> bit - but I hope that people will look carefully before increasing
> anything above 25 (or even above 9, which is now the biggest error
> listed for any x86 or x86_64 test).
>
> I did a spot check on the orphaned acos error - an expectation of
> 622ulps error on acos on i386, but without any particular test input
> expected to produce such an error.  It was originally an error for a
> test of acos (0.0625) (commit
> 1af437d6368694a0aea5f9ad1daec66bee2dd673, 2002-06-20).  That test
> failed to include the "L" suffix on the expected result, so that error
> was almost entirely the result of a testcase bug.  That was fixed in
> commit 96d10bdf030b75d28ddef3b342981ef810489ed8 (2002-07-02), and the
> expectation for that individual test removed - but not the overall
> expectation for acos.  My expectation is that the other large errors
> have similar stories: they were at one point actually seen in
> testsuite runs, but either testcases or function implementations have
> been changed since then so that those large errors will no longer be
> seen.

Thanks for digging into this. I suggest to commit this - and if any of 
us that runs the testsuite encounters problems which then might be due 
to compiler or hardware issues, we can increase the ULPs again.

Maintainers of other architectures might want to reset their ULPs files 
as well...

Andreas