You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@harmony.apache.org by "Yu-Nan He (JIRA)" <ji...@apache.org> on 2007/06/13 03:43:26 UTC
[jira] Commented: (HARMONY-3558) [drlvm][gc] drlvm crashes with
large pages enabled
[ https://issues.apache.org/jira/browse/HARMONY-3558?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12504086 ]
Yu-Nan He commented on HARMONY-3558:
------------------------------------
Hi,
The exception is generated after encountering a wrong object address 0xffffffff8xxxxxxx in my machine. This is a JIT bug and it's caused by type cast.
Below is a piece of the code generation log of JET:
---------------------------------
;; 2) GETFIELD 32 {org/apache/harmony/kernel/vm/InternMap::elementData [Lorg/apache/harmony/kernel/vm/InternMap$Entry;}
;;>check.npe for GR1 - skipped
; mov(i32) gr3, [GR1+8]
; movP gr2, 0000000080000000
; add gr3, gr2
8B 53 08 48 B9 00 00 00 80 FF FF FF FF 48 01 CA
-----------------------------------
After code generation, it's expected that the value is 0x0000000080000000, but actually it's 0xffffffff80000000 in binary code.
The wrong function movp_impl() [jitrino/src/jet/enc_ia32.cpp]:
-----------------------
void Encoder::movp_impl(AR op0, const void *op1)
{
EncoderBase::Operands args;
args.add(devirt(op0));
#ifdef _EM64T_
args.add(EncoderBase::Operand(OpndSize_64, (::int_ptr)op1));
#else
args.add(EncoderBase::Operand(OpndSize_32, (::int_ptr)op1));
#endif
ip(EncoderBase::encode(ip(), Mnemonic_MOV, args));
}
------------------------
"int_ptr" is a signed type and in Operand() constructor the type of second parameter is "long long".
I modify the constructor and movp_impl() to an unsigned type. The bug disappears. (Both JET and server)
> [drlvm][gc] drlvm crashes with large pages enabled
> ---------------------------------------------------
>
> Key: HARMONY-3558
> URL: https://issues.apache.org/jira/browse/HARMONY-3558
> Project: Harmony
> Issue Type: Bug
> Components: DRLVM
> Environment: winx64
> Reporter: Rustem Rafikov
> Assignee: Xiao-Feng Li
>
> DRLVM 64bit build 20070330 crashes both with gcv4.1 and gcv5 if using large pages with large enough heap size.
> Details:
> 1. GC4.1:
> VM crashes if heap size requested > ~900M and works otherwise. It crashes with:
> win64-20070330/bin/java.exe -Xms1024m -Xmx1024m -Xem:server -XX:gc.lp=true -Xverbose:gc.lp -c
> p ./jbb.jar;./check.jar spec.jbb.JBBmain -propfile SPECjbb.props
> gc.lp: large pages are allocated
> java/lang/ExceptionInInitializerError : (null)
> HMYEXEL062E Internal VM error: Failed to create Java VM
> FAILED to invoke JVM.
> 2. GCv5:
> Due to a bug in gcv5 parsing parameters functtion yuo should specify -XX:gc.use_large_page=true -XX:gc.large_page=true
> Command line:
> win64-20070330/bin/java.exe -Xem:server -XX:vm.dlls=gc_gen.dll -XX:gc.use_large_page=true -XX:gc.large_page=true -X
> s1500m -Xmx1500m -cp ./jbb.jar;./check.jar spec.jbb.JBBmain -propfile SPECjbb.props.short
> it crashes with the message:
> java/lang/ExceptionInInitializerError : (null)
> HMYEXEL062E Internal VM error: Failed to create Java VM
> FAILED to invoke JVM.
> GC use large pages.
> So, in the both cases there is the messages that large pages are allocated.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.