Crash on current head + switchpoint patch
Christian Thalinger
christian.thalinger at oracle.com
Tue Sep 6 01:40:12 PDT 2011
On Sep 6, 2011, at 7:50 AM, Tom Rodriguez wrote:
> The fix for 7071307 seems to have made it disappear, though I'm not sure why. It may just perturb the inlining such that it doesn't occur in this case. I think it's still a potential problem or at least it needs to be confirmed that it isn't a problem.
I think it's not a problem. The caller is a method handle adapter and so is the callee:
(dbx) p caller_method->is_method_handle_adapter()
caller_method->is_method_handle_adapter() = true
(dbx) p callee_method->is_method_handle_adapter()
callee_method->is_method_handle_adapter() = true
The logic in InlineTree::should_inline tests for caller_method->is_method_handle_adapter() and since the MDO data returns zero as invoke_count we hit the assert. But 7071307 adds this check:
+ if (invoke_count == 0) {
+ return "method handle not reached";
+ }
resulting in:
@ 200 java.lang.invoke.MethodHandle::invokeExact (41 bytes) inline (hot)
@ 8 java.lang.invoke.MethodHandle::invokeExact (20 bytes) method handle not reached
@ 17 java.lang.invoke.MethodHandle::invokeExact (12 bytes) method handle not reached
I think we are good and can close 7086187 as a duplicate of 7071307.
-- Christian
>
> tom
>
> On Sep 5, 2011, at 4:02 PM, Charles Oliver Nutter wrote:
>
>> The same case that blew up before doesn't blow up now with
>> hotspot-comp tip. Not sure if it's a heisenbug or if other changes
>> made the >0 check unnecessary.
>>
>> - Charlie
>>
>> On Mon, Sep 5, 2011 at 8:31 AM, Christian Thalinger
>> <christian.thalinger at oracle.com> wrote:
>>> I'm not able to reproduce it. What JRuby GIT revision fails?
>>>
>>> -- Christian
>>>
>>> On Sep 2, 2011, at 2:13 AM, Tom Rodriguez wrote:
>>>
>>>> I reproduced it. Basically we've chosen to inline a method handle because that's what we do but the profile for the call site says it's never been invoked, so the invoke_count is 0. When we try to use that zero as a scale for inlining tests of callees we get a div by zero. The problem has always been there I think but it was masked by the inline size cutoffs. We may want to emit uncommon traps for invokedynamic call sites that haven't been reached according to profiles. I filed 7086187 for this. If you need a workaround, this should work.
>>>>
>>>> diff -r a32de5085326 src/share/vm/opto/bytecodeInfo.cpp
>>>> --- a/src/share/vm/opto/bytecodeInfo.cpp
>>>> +++ b/src/share/vm/opto/bytecodeInfo.cpp
>>>> @@ -145,7 +145,7 @@
>>>> }
>>>>
>>>> assert(invoke_count != 0, "require invocation count greater than zero");
>>>> - int freq = call_site_count / invoke_count;
>>>> + int freq = invoke_count > 0 ? (call_site_count / invoke_count) : 0;
>>>>
>>>> // bump the max size if the call is frequent
>>>> if ((freq >= InlineFrequencyRatio) ||
>>>>
>>>> You'd need to disable the assert too if you wanted to run fastdebug.
>>>>
>>>> By the way, after working around it, I get this:
>>>>
>>>> dbx) c
>>>> MethodHandleStatics.java:102:in `newIllegalArgumentException': java.lang.IllegalArgumentException: target and fallback types must match: (ThreadContext,IRubyObject,IRubyObject,String)IRubyObject != (JRubyCallSite,ThreadContext,IRubyObject,IRubyObject,String)IRubyObject
>>>> from MethodHandles.java:2166:in `misMatchedTypes'
>>>> from MethodHandles.java:2150:in `guardWithTest'
>>>> from SwitchPoint.java:173:in `guardWithTest'
>>>> from InvokeDynamicSupport.java:660:in `updateInvocationTarget'
>>>> from InvokeDynamicSupport.java:462:in `invocationFallback'
>>>> from ./red_black_tree.rb:68:in `method__11$RUBY$initialize'
>>>> from $_dot_$red_black_tree$method__11$RUBY$initialize:65535:in `call'
>>>> from CachingCallSite.java:142:in `callBlock'
>>>> from CachingCallSite.java:148:in `call'
>>>> from RubyClass.java:798:in `newInstance'
>>>> from RubyClass$i$newInstance.gen:65535:in `call'
>>>> from JavaMethod.java:256:in `call'
>>>> from MethodHandle.java:566:in `invokeWithArguments'
>>>> from InvokeDynamicSupport.java:464:in `invocationFallback'
>>>> from bm1.rb:16:in `method__0$RUBY$rbt_bm'
>>>> from bm1$method__0$RUBY$rbt_bm:65535:in `call'
>>>> from bm1$method__0$RUBY$rbt_bm:65535:in `call'
>>>> from MethodHandle.java:566:in `invokeWithArguments'
>>>> from InvokeDynamicSupport.java:464:in `invocationFallback'
>>>> from bm1.rb:30:in `block_10$RUBY$__file__'
>>>> from bm1$block_10$RUBY$__file__:65535:in `call'
>>>> from CompiledBlock.java:112:in `yield'
>>>> from CompiledBlock.java:95:in `yield'
>>>> from Block.java:130:in `yield'
>>>> from RubyFixnum.java:252:in `times'
>>>> from MethodHandleImpl.java:1154:in `invoke_L5'
>>>> from MethodHandleImpl.java:1154:in `invoke_L5'
>>>> from MethodHandle.java:566:in `invokeWithArguments'
>>>> from InvokeDynamicSupport.java:538:in `invocationFallback'
>>>> from bm1.rb:29:in `__file__'
>>>> from bm1.rb:-1:in `load'
>>>> from Ruby.java:715:in `runScript'
>>>> from Ruby.java:708:in `runScript'
>>>> from Ruby.java:615:in `runNormally'
>>>> from Ruby.java:464:in `runFromMain'
>>>> from Main.java:317:in `doRunFromMain'
>>>> from Main.java:237:in `internalRun'
>>>> from Main.java:203:in `run'
>>>> from Main.java:187:in `run'
>>>> from Main.java:167:in `main'
>>>>
>>>> Anyway, thanks for the report.
>>>>
>>>> tom
>>>>
>>>> On Sep 1, 2011, at 3:49 PM, Charles Oliver Nutter wrote:
>>>>
>>>>> I'm running a build of hotspot-comp from today plus Christian's
>>>>> "switchpoint push invalidation" patch. Getting the following crash
>>>>> running the "redblack" benchmark (http://github.com/headius/redblack):
>>>>>
>>>>> headius at headius-vbox-ubuntu:~/projects/redblack$
>>>>> JAVA_HOME=../hotspot/build/linux/jdk-linux-amd64/ ../jruby/bin/jruby
>>>>> -Xinvokedynamic.all=true -X+C bm1.rb 20
>>>>> #
>>>>> # A fatal error has been detected by the Java Runtime Environment:
>>>>> #
>>>>> # SIGFPE (0x8) at pc=0x00007fdbad4439c0, pid=14498, tid=140581309708032
>>>>> #
>>>>> # JRE version: 7.0_02-b05
>>>>> # Java VM: OpenJDK 64-Bit Server VM (22.0-b02-internal mixed mode
>>>>> linux-amd64 compressed oops)
>>>>> # Problematic frame:
>>>>> # V [libjvm.so+0x2719c0] InlineTree::should_inline(ciMethod*,
>>>>> ciMethod*, int, ciCallProfile&, WarmCallInfo*) const+0x100
>>>>> #
>>>>> # Failed to write core dump. Core dumps have been disabled. To enable
>>>>> core dumping, try "ulimit -c unlimited" before starting Java again
>>>>> #
>>>>> # An error report file with more information is saved as:
>>>>> # /home/headius/projects/redblack/hs_err_pid14498.log
>>>>> #
>>>>> # If you would like to submit a bug report, please visit:
>>>>> # http://bugreport.sun.com/bugreport/crash.jsp
>>>>> #
>>>>> Aborted
>>>>>
>>>>> Should be easy to produce, but I can provide any info you need. JRuby
>>>>> master, amd64 product build. -Xinvokedynamic.all is the same as
>>>>> -Djruby.invokedynamic.all and turns on all current uses of
>>>>> invokedynamic (some are off because they're slow in Java 7 GA).
>>>>>
>>>>> - Charlie
>>>>
>>>
>>>
>
More information about the hotspot-compiler-dev
mailing list