Design and Implementation of a 64-bit PowerPC Port of Jikes RVM 2.0.3

1
Design and Implementation of a 64-bit PowerPC
Port of Jikes RVM 2.0.3

• Sergiy Kyrylkov, University of New Mexico
• Darko Stefanovic, University of New Mexico
• Eliot Moss, University of Massachusetts Amherst

2
Motivation for a 64-bit VM

• New set of research opportunities
• Particularly interesting for memory management
  research
• No 64-bit open-source VM with enough
  functionality available

3
Jikes RVM 2.0.3for 64-bit PowerPC

• 32/64-bit clean runtime
• 32/64-bit clean baseline compiler
• Most of GCTk collectors (SS, gen, Appel, OF,
  Beltway)
• No optimizing compiler
• No adaptive subsystem
• No watson collectors
• No JNI

4
Jikes RVM 2.0.3for 64-bit PowerPC

• Addresses are ADDRESS
• Words are WORD
• Offsets are int
• JTOC is int
• Operand stack is WORD
• Default object header is two words
• Method and interface IDs are int

5
Design DecisionsJTOC Layout
6
Design DecisionsObject Layout
7
Design DecisionsStack Frame Layout
8
Design DecisionsOperand Stack Layout
9
Porting Baseline Compiler

• New 64-bit instructions in VM_Assembler
• static final int LDtemplate 58ltlt26
• static final INSTRUCTION LD (int RT, int DS, int
  RA)
• return 58ltlt26 RTltlt21 RAltlt16 (DS0xFFFC)
• final void emitLD (int RT, int DS, int RA)
• if (VM.VerifyAssertions)
• VM.assert(fits(DS, 16))
• VM.assert(correctds(DS))
• INSTRUCTION mi LDtemplate RTltlt21 RAltlt16
  (DS0xFFFC)
• if (VM.TraceAssembler) asm(mIP, mi, "ld", RT,
  signedHex(DS), RA)
• mIP
• mc.addInstruction(mi)

10
Porting Baseline Compiler

• New methods in VM_Assembler
• final void emitLint (int RT, int D, int RA)
• //-if RVM_FOR_POWERPC32
• emitLWZ(RT, D, RA)
• //-endif
• //-if RVM_FOR_POWERPC64
• emitLWA(RT, D, RA)
• //-endif

11
Porting Baseline Compiler

• 32/64-bit clean VM_Compiler
• case 0x60 / --- iadd --- /
• if (VM.TraceAssembler) asm.noteBytecode("iadd")
• /
• asm.emitLWZ (T0, 0, SP)
• asm.emitLWZ (T1, 4, SP)
• asm.emitADDC (T2, T1, T0)
• asm.emitSTWU (T2, 4, SP)
• /
• asm.emitLint (T0, 0, SP)
• asm.emitLint (T1, BYTES_IN_WORD, SP)
• asm.emitADDC (T2, T1, T0)
• asm.emitSTWU (T2, BYTES_IN_WORD, SP)
• break

12
Porting Core Runtime

• Integer/address disambiguation
• int beg VM.objectAsAddress(instructions)
• private static int obsoleteMethodCount
• vs
• ADDRESS beg VM.objectAsAddress(instructions)
• private static int obsoleteMethodCount
• New system-wide constants
• int tibIndex method.getOffsetltlt2
• vs
• int tibIndex method.getOffsetltltBYTES_IN_ADDRESS_
  LOG

13
Porting Core Runtime

• New naming conventions
• sysCall1
• vs
• sysCall_X_rI, sysCall_I_rI, sysCall_A_rI
• sysCall_AI_rX, syscal_AIA_A
• Method/code splitting
• int sysCall1(int p1)
• vs
• int sysCall_I_rI(int p1)
• ADDRESS sysCall_A_rA(ADDRESS p1)

14
Porting Core Runtime

• Pure 64-bit implementation of magic in
  baseline-compiler-like style
• if (methodName VM_MagicNames.sysCall_I_rI)
• int size SIZE_INTEGER
• int valueOffset
• valueOffset generateSysCall1(asm, size,
  false)
• valueOffset SIZE_INTEGER
• asm.emitLint(3, valueOffset, SP) //
  load value
• generateSysCall2(asm, size)
• generateSysCallRet_I(asm, size)
• return

15
Lessons Learned

• Use appropriate data types (int vs int,
  ADDRESS, WORD)
• Use system-wide constants and variables(4 vs
  BYTES_IN_INT, BYTES_IN_ADDRESS)
• Choose appropriate naming conventions(getMemoryWo
  rd vs getIntAtAddress, getAddressAtAddress)

16
jess32 vs jess64
17
javac32 vs javac64
18
jack32 vs jack64
19
pseudojbb32 vs pseudojbb64
20
Conclusions

• Open-source 64-bit VM with most complete
  functionality
• Provides testbed for a number of new research
  opportunities
• Useful insights for future VM developers to
  create robust and portable software

21
Future Work

• Add optimizing compiler and adaptive subsystem
  for complete functionality
• Implement new GC algorithms (within JMTk)
  benefiting from 64-bit address space to evaluate
  their performance on a 64-bit PowerPC architecture