Request for review: 7034585 Adjust fillinStackTrace filtering to assist 6998871

[David Holmes]

Tom Rodriguez said the following on 04/09/11 09:22:
> Actually why are you checking for fillInStackTrace* instead of just fillInStackTrace?
> 
> if (method->name() == vmSymbols::fillIinStackTrace_name() &&
>     throwable->is_a(method->method_holder())) {

Added flexibility. There is still some debate as to the final naming 
scheme in the JDK changes. At present an overload of fillInStackTrace is 
used with a dummy int parameter. The normal naming scheme would have the 
native method be called fillInStackTrace0. That would have required a VM 
change, hence the overload. But we ended up needing a VM change anyway, 
hence it is now possible to change the method name. I was trying to 
accommodate this either way. But perhaps it is simpler to just add the 
symbol for fillInStackTrace0 and check for either?

David

> tom
> 
> On Apr 8, 2011, at 4:12 PM, Tom Rodriguez wrote:
> 
>> On Apr 8, 2011, at 3:57 PM, David Holmes wrote:
>>
>>> Hi Tom,
>>>
>>> Tom Rodriguez said the following on 04/09/11 06:53:
>>>> Performance of fillInStackTrace is somewhat important so I think you should avoid using as_C_string since that creates a copy of both fillInStackTrace and the method being checked.  Maybe something like this:
>>>> const char* fillInStackTrace = "fillInStackTrace";
>>>> int len = 16;
>>>> assert(strlen(fillInStackTrace) == len, "must agree");
>>>> if (method->name()->starts_with(fillInStackTrace, 16) &&
>>>>   throwable->is_a(method->method_holder())) {
>>> I hear you, but we shouldn't be hard-coding names like that - which is why we use the vmSymbols entry in the first place.
>> The reason we use the vmSymbols is so that we can do pointer compares.  An assert like this:
>>
>> assert(strcmp(vmSymbols::fillInstackTrace()->as_C_string(), fillInStackTrace) == 0, "must be the same");
>>
>> would mitigate your concern, though I think it's unlikely we will be changing the name fillInStackTrace any time soon.
>>
>>> It's a pity symbols don't have a set of operators for comparing against other symbols etc. I'll poke a little deeper to see what impact as_C_string will have and whether there is a way to mitigate it - perhaps use the variant where I supply the buffer eg:
>>>
>>> char[] buf = ...
>>> if (method->name()->starts_with(vmSymbols::fillInstackTrace()->as_C_string(buf), 16) { ...
>> It still seems like pointless work.  The name of the symbol is the value of the symbol so dynamically converting it to char* is just a waste.
>>
>> tom
>>
>>> Thanks,
>>> David
>>>
>>>> tom
>>>> On Apr 8, 2011, at 4:33 AM, David Holmes wrote:
>>>>> webrev:
>>>>>
>>>>> http://cr.openjdk.java.net/~dholmes/7034585/webrev/
>>>>>
>>>>> When an exception is created, fillInStacktrace is called to populate the backtrace information. This is done in java_lang_Throwable::fill_in_stack_trace in the VM. Because the interesting part of the stacktrace is from the location where the exception was created, and upwards, filtering is applied in fill_in_stack_trace to remove the entry for fillInStackTrace() itself, and the exception constructors.
>>>>>
>>>>> The current filtering code only expects to find a single frame for the fillInStackTrace method, so if an exception class overrides fillInStackTrace (and invokes super.fillInStackTrace) this actually disables the filtering of the constructors. Eg we see:
>>>>>
>>>>> Exception in thread "main" MyException
>>>>>      at MyException.fillInStackTrace(MyException.java:3)
>>>>>      at java.lang.Throwable.<init>(Throwable.java:260)
>>>>>      at java.lang.Exception.<init>(Exception.java:54)
>>>>>      at java.lang.RuntimeException.<init>(RuntimeException.java:51)
>>>>>      at MyException.<init>(MyException.java:1)
>>>>>      at MyException.main(MyException.java:7)
>>>>>
>>>>> instead of:
>>>>>
>>>>> Exception in thread "main" MyException
>>>>>      at MyException.main(MyException.java:7)
>>>>>
>>>>> The changes to Throwable.java for 6998871 effectively introduce an additional fillInStackTrace() frame and so this too disables the desired filtering of the backtrace.
>>>>>
>>>>> The proposal is quite simple: to modify fill_in_stack_trace so that it expects one or more fillInStackTrace frames, followed by zero or more constructor frames. This addresses the needs of 6998871 as well as fixing any user-defined overriding of fillInStackTrace.
>>>>>
>>>>> Thanks,
>>>>> David
>