Status fields:
| creation_ts: | 2008-08-07 13:53 |
|---|---|
| component: | jit |
| version: | default branch |
| rep_platform: | arm |
| op_sys: | All |
| bug_status: | RESOLVED |
| resolution: | FIXED |
| reporter: | twisti@complang.tuwien.ac.at |
cacaojvm.org
I long thought this bug is somewhere down in OpenJDK's native code, but I'm not sure
anymore. Here is a stacktrace:
$ java test
Exception in thread "main" java.lang.IllegalArgumentException: disparate values
at sun.misc.FDBigInt.quoRemIteration(FloatingDecimal.java:2795)
at sun.misc.FloatingDecimal.dtoa(FloatingDecimal.java:833)
at sun.misc.FloatingDecimal.<init>(FloatingDecimal.java:460)
at sun.misc.FloatingDecimal.parseHexString(FloatingDecimal.java:2297)
at sun.misc.FloatingDecimal.readJavaFormatString(FloatingDecimal.java:1064)
at java.lang.Double.valueOf(Double.java:492)
at test.main(test.java:43)
Somewhere in Double.valueOf() we call this method (this is the first method where arguments differ): x86_64: sun.misc.FpUtils.rawCopySign(DD)D PUBLIC STATIC(1.79769e+308 (0x7fefffffffffffff), 1 (0x3ff0000000000000)) arm: called: sun.misc.FpUtils.rawCopySign(DD)D PUBLIC STATIC(2.64731e-315 (0x1ff0000000000000), 1 (0x000000003ff00000)) Looks like another byte-swap problem.
On an --enable-softfloat ARM build the test is OK: $ java test $ The call trace tells us that the values are properly passed: sun.misc.FpUtils.rawCopySign(DD)D PUBLIC STATIC(1.79769e+308 (0x7fefffffffffffff), 1 (0x3ff0000000000000)) I start to hate FPA on LE ARMs.
Btw. sun.misc.FpUtils.rawCopySign() is called from sun.misc.FloatingDecimal.parseHexString().
Created an attachment (id=45) testcase
Xerxes sent me a call trace and it seems we have different problem, although I still don't know why because he also has a soft-float build. x86_64: called: sun.misc.FloatingDecimal.dtoa(IJI)V PRIVATE(0x00000000005f4400 (Class = "sun.misc.FloatingDecimal"), 1023 (0x000003ff), 9007199254740991 (0x001fffffffffffff), 53 (0x00000035)) called: java.lang.StrictMath.floor(D)D PUBLIC STATIC NATIVE(308.275 (0x4073446485c91398)) finished: java.lang.StrictMath.floor(D)D->308 (0x4073400000000000) Xerxes' ARM: called: sun.misc.FloatingDecimal.dtoa(IJI)V PRIVATE(0x00076270 (Class = "sun.misc.FloatingDecimal"), 1023 (0x000003ff), 9007199254740991 (0x001fffffffffffff), 1144114972 (0x4431cf1c)) called: java.lang.StrictMath.floor(D)D PUBLIC STATIC NATIVE(308.275 (0x4073446485c91398)) finished: java.lang.StrictMath.floor(D)D->1.59098e-314 (0x00000000bff00000) That looks very much like a native library problem. Iyonix ARM (hard-float): called: sun.misc.FloatingDecimal.dtoa(IJI)V PRIVATE(0x0008c2d0 (Class = "sun.misc.FloatingDecimal"), -1042 (0xfffffbee), 4504059029487616 (0x0010006af6800000), 63104 (0x0000f680)) called: java.lang.StrictMath.floor(D)D PUBLIC STATIC NATIVE(2.71828 (0x800000004005bf0a)) finished: java.lang.StrictMath.floor(D)D->0 (0x0000000000000000)
Xerxes' ARM: Recompiled all native OpenJDK classes and cacao copiled with softfloat passes the testcase. xerxes@armel:~/test$ /home/xerxes/cacao-clean/icedtea6/bootstrap/icedtea/bin/java -version java version "1.6.0_0" OpenJDK Runtime Environment (build 1.6.0_0-b11) CACAO (build 0.99.2, JIT mode) xerxes@armel:~/test$ /home/xerxes/cacao-clean/icedtea6/bootstrap/icedtea/bin/javac test.java xerxes@armel:~/test$ /home/xerxes/cacao-clean/icedtea6/bootstrap/icedtea/bin/java test xerxes@armel:~/test$
I get called: sun.misc.FloatingDecimal.dtoa(IJI)V PRIVATE(0x00a524b0 (Class = "sun.misc.FloatingDecimal"), 1023 (0x000003ff), 9007199254740991 (0x001fffffffffffff), 0 (0x00000000)) called: java.lang.Math.floor(D)D PUBLIC STATIC(308.275 (0x4073446485c91398)) finished: java.lang.Math.floor(D)D->308 (0x4073400000000000) The last argument of dtoa is wrong, but that seems to be more a problem with +TraceJavaCalls than with the FP handling itself.
| date: | 2008-08-07 14:37 |
|---|---|
| desc: | testcase |
| type: | text/x-java |
| download: | test.java |