Fix libm-test.inc ulps calculation for subnormals (bug 14064)

[Joseph S. Myers]

When fixing problems with exp10 I noted a problem, bug 14064, with how
libm-test.inc calculates ulps values for subnormal expected results,
so causing last-bit errors to be wrongly reported as huge numbers of
ulps.

I propose this patch to fix this libm-test.inc bug.  Tested x86_64,
x86 and (together with my exp10 patch
<http://sourceware.org/ml/libc-alpha/2012-05/msg00201.html>, where the
error is now reported as 1ulp not 35184372088832ulp) powerpc.

2012-05-06  Joseph Myers  <joseph@codesourcery.com>

	[BZ #14064]
	* math/libm-test.inc (check_float_internal): Correct ulp
	calculation for subnormal expected results.

diff --git a/math/libm-test.inc b/math/libm-test.inc
index 59192ed..86ec315 100644
--- a/math/libm-test.inc
+++ b/math/libm-test.inc
@@ -503,8 +503,7 @@ check_float_internal (const char *test_name, FLOAT computed, FLOAT expected,
 	  ulp = diff / FUNC(ldexp) (1.0, FUNC(ilogb) (expected) - MANT_DIG);
 	  break;
 	case FP_SUBNORMAL:
-	  ulp = (FUNC(ldexp) (diff, MANT_DIG)
-		 / FUNC(ldexp) (1.0, FUNC(ilogb) (expected)));
+	  ulp = (FUNC(ldexp) (diff, MANT_DIG) / min_value);
 	  break;
 	default:
 	  /* It should never happen. */

-- 
Joseph S. Myers
joseph@codesourcery.com