Java Virtual Machine and Jikes RVM

1
Java Virtual Machine and Jikes RVM

• Topics
• Virtual Machine
• Performance
• Jikes RVM

2
Virtual machine (Abstract computer)

• JVM (Java Virtual Machine)
• Collection of libraries that runs java code
• CLR (Common Language Runtime)
• Similar to JVM but more target languages
• VMM (Virtual Machine Monitor)
• PVM (Parallel Virtual Machine)
• Network computers-like

3
Background

• JVM in Fig1 (layered abstraction)
• Bytecode in Fig 2 (machine language of JVM)
• Stream of bytecode is a sequence of inst, each
  has a 1-byte opcode, zero or more operand
• Mnemonic assembly language in JVM

Fig 1
Fig 2
4
Bytecode example
Method count() 0 aload_0 1 invokespecial 1
4 return Method
void main(java.lang.String) 0 iconst_2 1
istore_1 2 iconst_0 3 istore_2 4
iconst_0 5 istore_3 6 goto 16 9 iload_1
10 iload_3 11 iadd 12 istore_2 13 iinc 3
1 16 iload_3 17 iconst_5 18 if_icmplt 9
21 getstatic 2
24 iload_2 25 invokevirtual 3 println(int) 28 return

• class count
• public static void main(String s)
• int x2, y0, i
• for (i0 i
• y x i
• System.out.println( y )

5
JVM Facts

• Stack machine
• Object oriented
• Method invocation (static, virtual, interface)
• Exceptions
• Type checking
• Multithreading
• Memory management
• Class loading

6
Virtual Hardware

• Four basic parts
• register, stack, garbage-collected heap, method
  area
• 32-bit address, 4GB addressable memory, integer
  is 32-bit long
• Stack machine (vs. register machine)
• Registers PC, optop, frame, and vars

7
JVM performance

• Do not expect Java compiler/Jikes to perform
  many clever optimization due to Javas rather
  strict sequencing and thread semantics, there is
  a little the compiler can safely do, in contrast
  to compilers for less strictly defined languages
  such C/Fortran.

8
(No Transcript)
9
(No Transcript)
10
(No Transcript)
11
JVM Runtime state

• Method area
• like .text, store bytecode (pc)
• Stack
• stack frame storing params and result
• 3 sections local var, execution env, operand
  stack
• Heap
• created dynamically by new
• operator, runtime env tracks it

12
(No Transcript)
13
JVM Instruction-set family

• Load and Store
• load, store, push
• Arithmetic and Logic
• add, sub, mul, div, rem, neg, shl, shr, or, xor
• Conversions
• i2l, i2f, i2b, i2c, i2s
• Objects
• new, newarray, getfield, putfield, aload,
  astore, instanceof
• Stack management
• pop, dup, dup_x, swap

14

• Control transfer
• if_icmpeq, ifreq, iflt, ifnull, tableswitch,
  cmp, goto, jsr, ret, athrow
• Method invocation
• invokevirtual, invokeinterface, invokespecial,
  invokestatic
• Multiple encodings
• iconst_m1, iconst_0, bipush, sipush, ldc
• Constant pool
• Utf8, Inteter, Float, Class, NameAndType,
  Fieldref, MethodRef, InterfaceMethodref

15
Jikes RVM Introduction

• Formerly called Jalapeho project at IBM
• Based on JIT (Just-in-time) compilation
• Not a full VM, meant for research
• Nothing to do with source to bytecode compiler
• Implemented in Java (203k Java, Feb 02) with
  GNUs CLASSPATH
• Runs on AIX/PowerPC and Linux/IA32

16
Jikes RVM subsystem

• Two compilers (baseline and optimizing compiler)
  and Adaptive Optimization System (AOS).
• Major java features
• Bootstrapping, Method invocation, Method
  dispatch, interface method invocation,
• Exception, Dynamic type checking, Memory memory
  management, Garbage collection,
• Threads, Class loading etc

17
Baseline Compiler

• Based on Just-In-Time compiler
• Two main components
• Generate GC map for safe point
• Generate executable (209 switch statement)
• Generate machine code
• Emulate stack machine
• Faster in compilation but much slower (5) than
  optimizing compiler

18
Optimizing compiler

• Translating code into IR (Intermediate
  Representation) and perform optimization.
• Three levels of IRs, multiple optimization levels
  (many classical, some novel), 100K.
• Template driven, instruction selection, IA32
  assembler
• Three groups local (basic block), global (method
  level) and inter-procedural (across methods).
• Interface to adaptive optimization system.

19
Transformation

• Four transformations
• - HIR (High IR)
• - LIR (Low IR)
• - MIR (Machine IR)
• - Final assembly to machine code
• Transformation
• - Preserve original meaning
• - Measurable speedup
• - Worth effort

transform
20
Bytecode to IRs
21

• High level IR (HIR)
• Recognize arch independent
• bytecode and resemble them
• - Use register transfer language
• (JVM stack abstraction)
• - Array bound check
• - Inline of bytecode subroutine
• - Devirtualization
• Low level IR (LIR)
• - Arch independent, Complex HIR operators (new,
  call virtual) are expanded
• Machine Level IR (MIR)
• Arch specific, defining MIR file (29 format)
• Inst selection, register allocation
• One-to-one mapping between operators and target
  ISA

22

• Common subexpression
• for (i0) xai aiaj
• t64i t6 4i
• xat6 x at6
• t74i at6 t9
• at7t9
• Dead-code
• if (debug) print
• Loop optimization (induction variables)
• while (i
• while (i

23
HIR (BC2IR)

• Transform to HIR
• Abstract interpretation of bytecode
• Bytecode subroutines inlined
• Translate stack inst to 3-address inst
• On-the-fly dataflow optimizations
• (const folding, branch optimization, unreachable
  code elimination, const and type propagation, )
• Build FCFG
• Ref ACM Java Grande 99

24
HIR Example
25
FCFG example

• Reduced of nodes
• Improve basic block analysis
• Adapt forward/backward analysis

26
MIR (HIR2LIR)

• Convert to LIR
• Introduce Jikes RVM detail into IR
• VM services
• (allocation, locks, type, checks, write
  barriers)
• Object model
• (field, array, static layout, method invocation)
• Other low-level expansion
• (switch, compare/branch, exception checks)

27
LIR Example
t1 new java.lang.Object return t1
public static Object quickNewScalar(int size,
Object tib, boolean hasFinalizer) throws
OutOfMemoryError Object rrt
VM_Allocator.allocateScalar(size, tib) if (
hasFinalizer ) VM_Finalizer.addElement(ret) retu
rn ret
public static Object allocateScalar(int size,
Object tib) throws OutOfMemoryError
VM_Address regiongetHeapSpaceFast(size) Obje
ct newObj VM_ObjectMode.initializeScalar( re
gion, tob, size) return newObj
28
MIR (LIR2MIR)

• Convert to MIR
• Instruction selection rules for each architecture
• Each rule in LIR2MIR.rules is defined by a
  four-line record (PRODUCTION, COST, FLAGS,
  TEMPLATE)
• Ref Bottom-up, rewrite system (Fraser 92,
• Sarkar 01)

29
Code optimization techniques

• Three address code
• Keep heavily used var in registers
• Control-flow improvement
• Common subexpression
• Dead-code elimination
• Constant folding
• Loop optimization
• Basic block optimization (flow graph)
• Data Flow analysis and flow graph

30
More

• Flow insensitive optimizations (register list)
• Inlining any method with on-the-fly optimization
  (guarded inlining)
• Load Elimination
• Loop Normalization
• Loop Unrolling
• Global code placement
• Commoning (separate exeception from computation)
• Heap Array SSA
• FCFG forward analysis
• Simple Escape analysis (use register list to
  identify object)

31
Final Assembly

• Machine code generation
• GC maps
• Exception tables
• Machine code info for
• Online profiling, stackframe inspection,
  debugging, dynamic linking, lazy compilation

32
Other interesting work

• GUN Kissme JVM
• Porting the Jikes RVM to Linux/IA32
• Cache performance in JVM
• What missed Java Memory Management, Jikes
  specific optimization.