Profiler modifications to Kaffe's JIT system

Initial work from Nathan M. Hanish and William E. Cohen.
Improved by Edouard G. Parmelan.


Instrumented Java Virtual Machine Overview:

What system does the instrumentation work with?

	* Kaffe Virtual Machine, Version 1.0.5 with JIT and JIT3 Translator
	Intel Pentium(I,II,III,MMX,Pro), Celeron, or Xeon processor based
	system that provides the IA-32 timestamp register (Intel instruction
	rdtsc).

	* Tested on Intel GNU/Linux.


What does the instrumentation record?

	* A count of the total number of invocations for each method,
	including native method.
	
	* The total amount of CPU time spend between the entry and exit of
	each method.
	
	* The total amount of CPU time spend between the entry and exit of
	each of the methods children (not applicable for native methods).
	
	* Time spent in just that method.
	
	* The total amount of CPU time spend between the performing JIT
	translation of executed methods.
	
	* A timing resolution of 1 clock cycle.
	
	* Low overhead, roughly 10 instructions per method invocation.


What doesn't the instrumentation provide?

	* A call graph.
	
	* Proper handling of exceptions and multi-threaded code.  Upon
	exception, only the self+children time is properly updated.    
	
	* Time spent in particular C functions called from native method.
	

How do to use the instrumentation?

	Install the latest version of the instrumented virtual machine.
	And execute Kaffe with -prof option.  The output is dumped to
	file prof.out in current directory.


What is the format of the output?

	Each line of prof.out contains the profiling information for one
	method.  The fields are printed as human readable ASCII text and are
	separated by spaces.  All times are in milliseconds.
	Each line has the following format:

		time		time	time		time
	count   self+child	self	children	jit	 method-name


How is the profiling implemented?

	* Four fields added to the structure that describes each method:
		- Count of invocations (callsCount)
		- Time spent in self and children (totalClicks)
		- Time spent in children (totalChildrenClick)
		- Time spent in JIT (jitClicks)

	* The JIT translator modified to produce additional inline code
	(config/i386/{jit,jit3}-i386.def):
		Prologue:
			Increment count of invocations
			Subtract current time from time in self and children
		Epilogue:
			Add current time from time in self and children
		Calls:
			Read current time and subtract from time in child
			Call function
			Read current time and add to time in child

	* Dumping data:
		Add function printProfilerStats() to dump profiling data
		in kaffe/kaffevm/{jit,jit3}/machine.c.

	* Headers:
		config/i386/common.h defines the profiler_click_t type and
		macro profiler_get_clicks().

		kaffe/kaffevm/classMethod.h has additional fields in the
		methods structure for storing the profiling data.


How to port to other arch ?
    
	* Create type profiler_click_t and macro profiler_get_clicks() in
	config/ARCH/common.h and define KAFFE_PROFILER.

	* Create macros profiler_start() and profiler_end() for your JIT
	engine in config/ARCH/jit-ARCH.def and/or config/ARCH/jit3-ARCH.def
	and use them in prologue, epilogue and all call functions.
	
	* Don't forget to enclose your code in #if defined(KAFFE_PROFILER)
	and #endif block.


Last edit: Sun Dec  5 02:01:40 CET 1999