Hi,
Also, is this the only test that's failing?
Nope, there was another one later on IIRC. I'll disable the PosQNan2 and
NegQNan* checks and investigate.
It so turns out that NaN handling by the testsuite was all that fails on
my machine, due to compiler and binutils assumptions about what's counting
as NaN on m68k. Basically, cpp does substitute anything that it knows to
not represent a valid float by a const nan of positive sign.
The following patch to the HalfTest suite makes the testsuite failure go
away. Please apply. In the alternative, please indicate whether a binNMU
would be in order.
Michael
--- ilmbase-1.0.1/HalfTest/testClassification.cpp.org 2008-04-28
01:21:13.000000000 +0200
+++ ilmbase-1.0.1/HalfTest/testClassification.cpp 2008-04-28
00:19:22.000000000 +0200
@@ -156,9 +156,10 @@
testClass (-0.3, 1, 1, 0, 0, 0, 0, 1);
testClass (-HALF_MAX, 1, 1, 0, 0, 0, 0, 1);
testClass (floatNegInfinity(), 0, 0, 0, 0, 0, 1, 1);
+#ifndef __mc68000__
testClass (floatNegQNan1(), 0, 0, 0, 0, 1, 0,
1);
testClass (floatNegQNan2(), 0, 0, 0, 0, 1, 0,
1);
-
+#endif
cout << "\n";
testClass (half::posInf(), 0, 0, 0, 0, 0, 1, 0);
--- ilmbase-1.0.1/HalfTest/testBitPatterns.cpp.org 2008-04-28
01:21:51.000000000 +0200
+++ ilmbase-1.0.1/HalfTest/testBitPatterns.cpp 2008-04-27 23:57:37.000000000
+0200
@@ -305,9 +305,11 @@
testBits (floatPosQNan1(),
"0 11111 1111111111", // nan
"0 11111111 11111111110000000000000"); // nan
+#ifndef __mc68000__
testBits (floatPosQNan2(),
"0 11111 1010101010", // nan
"0 11111111 10101010100000000000000"); // nan
+#endif
//
// Numbers close to -1.0
@@ -473,12 +475,14 @@
testBits (floatNegInfinity(),
"1 11111 0000000000", // +infinity
"1 11111111 00000000000000000000000"); // +infinity
+#ifndef __mc68000__
testBits (floatNegQNan1(),
"1 11111 1111111111", // nan
"1 11111111 11111111110000000000000"); // nan
testBits (floatNegQNan2(),
"1 11111 1010101010", // nan
"1 11111111 10101010100000000000000"); // nan
+#endif
cout << "ok\n\n" << flush;
}
--
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
