Hi Serguei,
Thanks a lot for taking a look at this. Yes, the primary motivation for changing this is to have this code "do the same" as the GC in respect of InterpretedVFrame (granted, we do care about value slots here, where GC doesn't). Sorry about making this more complex, I have tried to make some of it a bit easier to read based on your input, please take a look if you think this is easier to follow: New webrev suggestion: http://cr.openjdk.java.net/~mgronlun/8039905/webrev02/ And I agree with all the points about T_CONFLICT and dead locals that we need to consider as a next step. Thanks again for taking the time Many thanks Markus From: Serguei Spitsyn Sent: den 3 juli 2014 12:59 To: Markus Grönlund; hotspot-runtime-dev; Serviceability Dev Subject: Re: RFR(M): 8039905: heapdump/OnOOMToFile and heapdump/OnOOMToPath fail with "assert(fr().interpreter_frame_expression_stack_size() >= length) failed: error in expression stack!" Markus, I've done another pass through the webrev. It looks good in general. Thank you for fixing it! As for the ::locals() and ::expressions() refactoring, my guess is that your motivation was to make it closer to what the GC is doing. Is it true? You can keep it as you like because my comment is not a big or strong statement. It is just my personal preference. But, the refactoring made it harder to review the changes as the original mapping has gone. As we privately discussed, this code may need further improvements related to the dead locals and so, we may need to open new bugs for recognized issues. Also, It seems you are right on the meaning of the T_CONFLICT that it is set for the dead locals. However, in my bug/testcase context the T_CONFLICT value was not set for dead local. The observed value of the dead local was T_INT (as uninitialized data) which looks pretty strange. It is possible, we have a deal with another bug here. Thanks, Serguei On 7/2/14 10:04 PM, HYPERLINK "mailto:serguei.spit...@oracle.com"serguei.spit...@oracle.com wrote: Hi Markus, Sorry for the latency. I hope this review is still needed. src/share/vm/interpreter/oopMapCache.hpp A minor comment: The function blocks indentation was originally aligned, but this fix partially broke it: - uintptr_t entry_at(int offset) { int i = offset * bits_per_entry; return bit_mask()[i / BitsPerWord] >> (i % BitsPerWord); } + uintptr_t entry_at(int offset) const { int i = offset * bits_per_entry; return bit_mask()[i / BitsPerWord] >> (i % BitsPerWord); } void set_expression_stack_size(int sz) { _expression_stack_size = sz; } #ifdef ENABLE_ZAP_DEAD_LOCALS - bool is_dead(int offset) { return (entry_at(offset) & (1 << dead_bit_number)) != 0; } + bool is_dead(int offset) const { return (entry_at(offset) & (1 << dead_bit_number)) != 0; } #endif // Lookup - bool match(methodHandle method, int bci) { return _method == method() && _bci == bci; } - bool is_empty(); + bool match(methodHandle method, int bci) const { return _method == method() && _bci == bci; } + bool is_empty() const; src/share/vm/runtime/vframe.cpp The implementation of the ::stack_data () combined the logics from ::locals() and ::expressions(). As a result, the function became unreasonably more complex to have a deal with this combination. The original approach looks better even though some fragments are repeated twice. In other words, more simple and flat lines is better than less lines with more complexity as it adds another dimension to track down. I need a little bit more time to better understand the vframe.cpp part of the fix. Thanks, Serguei On 6/25/14 4:58 AM, Markus Grönlund wrote: Greetings, Kindly looking for reviews for the following change: Bug: http://bugs.openjdk.java.net/browse/JDK-8039905 Webrev: HYPERLINK "http://cr.openjdk.java.net/%7Emgronlun/8039905/webrev01/"http://cr.openjdk.java.net/~mgronlun/8039905/webrev01/ Description: JVMTI inspection code for following references makes use of a VM_HeapWalkOperation in order to follow-up root sets. In bug: https://bugs.openjdk.java.net/browse/JDK-8038624 - "interpretedVFrame::expressions() must respect InterpreterOopMap for liveness", it was found that the interpretedVFrame code had a discrepancy between basing length information from both asking the interpreter frame (which saw live expression slots for calls instructions) as well as the oop map (which did not). The liveness decisions for a particular BCI should be based on what the oopmap gives, and that was done as of that bug (8038624). In that process, I added an assert in an attempt to validate certain assumptions, and this assert has triggered during nightly testing in some cases. I have therefore inspected the GC code which follow-up roots from an interpreted frame (please see frame::oops_interpreted_do() and InterpreterFrameClosure::do_offset() (in frame.cpp) for reference), and reworked InterpretedVFrame so that inspections for the locals and expression slots are done in the same way as the GC code (especially in regards to taking decisions on the InterpreterOopMap). I needed to use InterpreterOopMap from a const context, and this is why I have "constified" this class where needed, as well as making "number_of_entries()" a const public accessor (to easily reach oop_mask length info). Testing completed: nsk/jdi* tests (especially the problematic ones reported for 8039905) compiler/6507107/HeapWalkingTest Thanks Markus
