On POWER4 PowerPC32 for logb(1) and POWER[4|5|6] PowerPC32 for logbl(1)
in FE_DOWNWARD rounding mode the result is -0.0, which is contrary to
IEEE 754-2008 (section 5.3.3).
This is bugzilla bug #887:
http://sourceware.org/bugzilla/show_bug.cgi?id=887
This is due to the nature of the Power Architecture fsub instruction
under hardware rounding mode FE_DOWNWARD.
GCC recently introduced the ability to use the fcfid instruction in
32-bit mode on power4, which fixes the issue for logb(1) and expanded
the use of fcfid for power6 which corrected the issue for logbl(1).
r186387 | meissner | 2012-04-12 13:10:27 -0400 (Thu, 12 Apr 2012) | 16 lines
[gcc]
2012-04-11 Michael Meissner<meissner@linux.vnet.ibm.com>
PR target/52775
* config/rs6000/rs6000.h (TARGET_FCFID): Add TARGET_PPC_GPOPT to
the list of options to enable the FCFID instruction.
(TARGET_EXTRA_BUILTINS): Adjust comment.
The included patch adds a test case to the libm test suite to detect
this issue if the compiler ever regresses.
I couldn't reproduce a similar nextafter() failure per the initial bug
report so I assume it's been fixed.
Ryan S. Arnold
IBM Linux Technology Center
2012-04-26 Ryan S. Arnold<rsa@linux.vnet.ibm.com>
[BZ #887]
* math/libm-test.inc (logb_test_downward): New test to expose
erroneous negative sign on -0.0 result of logb[l](1) in FE_DOWNWARD
rounding mode.
diff --git a/math/libm-test.inc b/math/libm-test.inc
index e0ac613..81de85b 100644
--- a/math/libm-test.inc
+++ b/math/libm-test.inc
@@ -5101,6 +5101,40 @@ logb_test (void)
END (logb);
}
+static void
+logb_test_downward (void)
+{
+ int save_round_mode;
+ errno = 0;
+
+ FUNC(logb) (0);
+ if (errno == ENOSYS)
+ /* Function not implemented. */
+ return;
+
+ START (logb_downward);
+
+ save_round_mode = fegetround ();
+
+ if (!fesetround (FE_DOWNWARD))
+ {
+
+ /* IEEE 754-2008 says (section 5.3.3) that "logB(1) is +0.". Libm
+ * should not return -0 from logb in any rounding mode. PowerPC32 has
+ * failed with this test for power4 logb (and logbl on all PowerPC
+ * platforms) in the past due to instruction selection. GCC PR 52775
+ * provides the availability of the fcfid insn in 32-bit mode which
+ * eliminates the use of fsub in this instance and prevents the negative
+ * signed 0.0. */
+ TEST_f_f (logb, 1.000e+0, plus_zero);
+ }
+
+ fesetround (save_round_mode);
+
+ END (logb_downward);
+}
static void
lround_test (void)
@@ -8210,6 +8244,7 @@ main (int argc, char **argv)
log1p_test ();
log2_test ();
logb_test ();
+ logb_test_downward ();
modf_test ();
ilogb_test ();
scalb_test ();