Greetings.
[EMAIL PROTECTED] wrote:
Hi,
Christian Thalinger (twisti) of the CACAO project has recently been looking at
using the OpenJDK libraries with CACAO as the VM (via a suitable libjvm.so)
to allow it to be built on non-HotSpot supported platforms (e.g. ppc, ppc64,
alpha). I've just been testing this on ppc64 this morning.
The trigonometric functions (sin, cos, tan) seem to run into difficulties on
these new platforms. On x86, x86_64 and sparc, sin(0.7) returns the correct
answer of 0.644217687237691. On alpha, the same test
gives -0.9589410493575494 while it returns the input value of 0.7 on both ppc
and ppc64. Ironically, gij on ppc64 gives the correct answer.
To give some background, in HotSpot the Math.{sin, cos, tan} methods are
often intrinsified, that is, they have JVM-internal implementation
(which meet the spec) rather than using the method compiled from the
source of java.lang.Math. The StrictMath class is specified to use the
fdlibm algorithms and currently it usually calls down to the C versions
of those functions compiled as part of the JDK.
The C code in fdlibm require a memory overlay between the bits of a
64-bit integer and a double. The source techniques used for this often
break when the C compiler optimization level is too high. We've had to
tweak the C compiler settings for fdlibm over the years on different
platforms and often compile fdlibm with different options than other
native code in the JDK so I would suspect some sort of build problem.
The Java language and vm require support for the full set of IEEE 754
floating-point values, including NaN and infinity. At least older Alpha
chops did not have full hardware support for these values and compiler
options were needed to enable their use. More recent Alphas do have
more complete hardware support but compiler options may be needed.
(Also note that the Alpha hardware uses a weaker memory model than
required by Java so some work would be needed in a vm to insert fences,
etc. as appropriate.)
-Joe Darcy
This was spotted when running SPECs check:
5$ cacao -cp ./check.jar spec/jbb/validity/PepTest
LOG: [0x30021cd0] JVM_GetStackAccessControlContext: IMPLEMENT ME!
LOG: [0x30021cd0] JVM_GetStackAccessControlContext: IMPLEMENT ME!
LOG: [0x30021cd0] JVM_FindSignal: name=HUP
LOG: [0x30021cd0] JVM_RegisterSignal: sig=0, handler=0x2, IMPLEMENT ME!
LOG: [0x30021cd0] JVM_FindSignal: name=INT
LOG: [0x30021cd0] JVM_RegisterSignal: sig=0, handler=0x2, IMPLEMENT ME!
LOG: [0x30021cd0] JVM_FindSignal: name=TERM
LOG: [0x30021cd0] JVM_RegisterSignal: sig=0, handler=0x2, IMPLEMENT ME!
LOG: [0x30021cd0] JVM_GetStackAccessControlContext: IMPLEMENT ME!
LOG: [0x30021cd0] JVM_GetStackAccessControlContext: IMPLEMENT ME!
LOG: [0x30021cd0] JVM_GetSystemPackage: IMPLEMENT ME!
LOG: [0x30021cd0] JVM_GetSystemPackage: IMPLEMENT ME!
testIf: OK
testArray: OK
testBitOps: OK
testFor: OK
testDiv: OK
testTableSwitch: OK
testLookupSwitch: OK
testHiddenField: OK
checkRemainders: long double OK
checkMathFcts: sin(0.7) evaluated to: 0.7, expected: 0.644218
cos(0.7) evaluated to: 1.0, expected: 0.764842
acos(0.7) evaluated to: 1.5707963267948966, expected: 0.795399
atan(0.7) evaluated to: 0.7, expected: 0.610726
******************************************
Some math function failed
******************************************
Primes less than 50: 2 3 5 7 11 13 17 19 23 29 31 37 41 43 47
testExc1(simple throw/catch): OK
testExc2(skip catch clauses): OK
testExc3(catch in inner): OK
testExc4(catch in outer): OK
testExc5(rethrow): OK
testExc6(throw accross call): OK
testExc7(throw accr. 2 calls): OK
loopExitContinueInExceptionHandler: OK
testStringHash: OK
testClone: OK
testObjectArray: OK
checkInstanceOf: OK
checkInterfaceInstanceOf: OK
testWaitNull: OK
testVarAndMethodNameClash: OK
testMisk: OK
testGC: OK
testRandom : OK
PepTest: error
Any idea what the problem may be and when this may be rectified?
Thanks,
