Allow spurious underflow / inexact for ldbl-128ibm [committed]

[Joseph Myers <joseph at codesourcery dot com>]

A large number of the test-ldouble failures seen for ldbl-128ibm are
spurious "underflow" and "inexact" exceptions.  These arise from such
exceptions in the underlying arithmetic; unlike other spurious
exceptions from that arithmetic, they do not in general relate to
cases where the returned result is also substantially inaccurate, are
not so readily avoidable by appropriately conditional libgcc patches,
and are widespread enough to be hard to handle through individual
XFAILing of the affected tests.

Thus, this patch documents relaxed accuracy goals for libm functions
for IBM long double and makes libm-test.inc reflect these spurious
exceptions in ldbl-128ibm arithmetic and always allow them in
ldbl-128ibm testing (while still not allowing these exceptions to be
missing where required to be present).  Tested for powerpc.  Committed.

2016-03-10  Joseph Myers  <joseph@codesourcery.com>

	* manual/math.texi (Errors in Math Functions): Document relaxed
	accuracy goals for IBM long double.
	* math/libm-test.inc (test_exceptions): Always allow spurious
	"underflow" and "inexact" exceptions for IBM long double.

diff --git a/manual/math.texi b/manual/math.texi
index 72f3fda..5c9f7b9 100644
--- a/manual/math.texi
+++ b/manual/math.texi
@@ -1327,6 +1327,16 @@ necessarily the exact value actually passed with discontiguous
 mantissa bits.
 
 @item
+For the IBM @code{long double} format, functions whose results are
+fully specified by reference to corresponding IEEE 754 floating-point
+operations have the same accuracy goals as other functions, but with
+the error bound being the same as that for division (3ulp).
+Furthermore, ``inexact'' and ``underflow'' exceptions may be raised
+for all functions for any inputs, even where such exceptions are
+inconsistent with the returned value, since the underlying
+floating-point arithmetic has that property.
+
+@item
 Functions behave as if the infinite-precision result computed is zero,
 infinity or NaN if and only if that is the mathematically correct
 infinite-precision result.  They behave as if the infinite-precision
diff --git a/math/libm-test.inc b/math/libm-test.inc
index c9925fd..f1ba7dd 100644
--- a/math/libm-test.inc
+++ b/math/libm-test.inc
@@ -672,13 +672,20 @@ test_exceptions (const char *test_name, int exception)
 	test_single_exception (test_name, exception, OVERFLOW_EXCEPTION,
 			       FE_OVERFLOW, "Overflow");
 #endif
+      /* Spurious "underflow" and "inexact" exceptions are always
+	 allowed for IBM long double, in line with the underlying
+	 arithmetic.  */
 #ifdef FE_UNDERFLOW
-      if ((exception & UNDERFLOW_EXCEPTION_OK) == 0)
+      if ((exception & UNDERFLOW_EXCEPTION_OK) == 0
+	  && !(TEST_COND_ldbl_128ibm
+	       && (exception & UNDERFLOW_EXCEPTION) == 0))
 	test_single_exception (test_name, exception, UNDERFLOW_EXCEPTION,
 			       FE_UNDERFLOW, "Underflow");
 #endif
 #ifdef FE_INEXACT
-      if ((exception & (INEXACT_EXCEPTION | NO_INEXACT_EXCEPTION)) != 0)
+      if ((exception & (INEXACT_EXCEPTION | NO_INEXACT_EXCEPTION)) != 0
+	  && !(TEST_COND_ldbl_128ibm
+	       && (exception & NO_INEXACT_EXCEPTION) != 0))
 	test_single_exception (test_name, exception, INEXACT_EXCEPTION,
 			       FE_INEXACT, "Inexact");
 #endif

-- 
Joseph S. Myers
joseph@codesourcery.com