Marmot

1
Marmot

• Bob Fitzgerald
• Todd Knoblock
• Erik Ruf
• Bjarne Steensgaard
• David Tarditi

2
Why Study Java?

• Good features static typing, type safe, OO,
  exceptions, gc
• Normal enough that MS programmers might use
  (versus ML, Scheme, ...)
• Real implementations available for
  bootstrap/comparison
• Benchmark code available

3
Motivation

• The quality of C/C software is not as good as
  it could be.
• Features such as clean design, type safety,
  memory safety, and garbage collection can
  substantially improve software quality.
• Butpeople are very concerned about the
  efficiency of Java

4
Counter Examples

• Office suite in Java
• Excel in Java

5
Inherent in Java, or artifact of implementations?

• Applets
• interpret/Jit
• Small memory
• Small code
• Mostly UI
• Less than 1/10 the speed of C.
• Applications
• ?

6
Goals

• Well-balanced performance
• Research infrastructure

7
The Marmot System
8
Buzzwords

• static
• type-directed
• SSA-based
• native code generation
• research platform

9
Language subset

• Correct thread, exception, memory semantics
• Most libraries, including awt, net
• No dynamic loading
• Minimal reflection

10
Where did the time go?
11
Architecture
12
Standard Scalar Optimizations

• constant/copy prop with conditional branch elim
• dead variable/code elim
• CSE
• redundant load elim (fields and array elts)
• invariant code motion
• strength reduction

• control flow simplifications
• inlining
• operator lowering, reanalysis
• coloring register alloc
• instruction scheduling
• missing PRE, loop unrolling

13
Standard OO Optimizations

• interprocedural allocation/invocation
• treeshaking
• intraprocedural type propagation
• static method binding
• check elimination(array store, cast, instanceof,
  null)

14
Bytecode-related optimizations

• type elaboration
• idiom recognition
• ints representing booleans
• control/value issues
• subroutines
• statically initialzed arrays

15
Intermediate Representation

• traditional CFG plus exception arcs

16
Block splitting enables SSA
17
Array Bound Check Optimization

• more difficult in Java
• precise exceptions limit code motion
• array use not limited to intraprocedural loop
  nests
• often performed incorrectly
• array indices are signed quantities
• i, i1, ..., ik doesnt increase monotonically

18
Interprocedural Flow Analysis

• non flow based
• RTA, TBAA, field analysis
• type inference based
• possibly-null,
• type propagation,
• escape frame/thread,
• thread specific synchronization/allocation

19
Flow analysis efficiency

• interprocedural
• must be near-linear time in practice
• intraprocedural
• O(n3) typically ok
• are procedures the proper units for this
  distinction?

20
Runtime

• Multiple GC implementations
• conservative, copying, generational, sliding,
  thread-specfic, stack allocation
• card marking, SSB write barriers
• compact tables
• Efficient primitives
• type tests, synchronization, exception handling,
  interface dispatch (but heavyweight threads)
• exposure in IR lowering, redundancy elimination

21
Backend

• coloring register allocation with lifetime
  splitting
• effective use of x86 instruction set
• basic instruction scheduling

22
Libraries

• custom implementation
• written Java wherever possible

23
Performance
24
Where does the time go now?
25
Performance Of Other Systems Relative to Marmot
Higher is better
26
(No Transcript)
27
Assessment

• performance is improving 18 in 1999, 24 more
  in 2000
• 10x faster than JIT
• 85 of C on average for cognates
• whats an acceptable price?

28
Some Open Questions

• (Things we know less about now than when we
  started)

29
Lowerings

• How many?
• How different?
• Atomizing operators
• Null/sync/finally, etc.

30
Is SSA a good idea?

• Expense and proliferation of locals
• Imputed definitions (Range, null, type, etc.)
• Arrays and fields
• Transforming the dominance relation
• Complications for exceptions

31
Space/Time Tradeoffs
32
What is Local, What is global?

• Basic block
• Method
• Class
• Application
• Processor
• System
• LAN
• World
• Universe

• O(n3) local
• O(n) global
• Are the source programs divisions meaningful?

33
Current Work
34
Interprocedural Flow Analysis

• goal model polymorphic containers
• e.g. Vector, Hashtable, Object
• primary sources of downcasts, virtual calls
• presence limits quality of interprocedural opts
• example VectorltTgt.elementAt(i) returns a T
• uses call rebinding, synchronization
  elimination, stackframe/thread allocation
• expensive substitute for source-level generics

35
Profile Based Optimization

• improve existing phases
• inlining
• field layout
• gc selection/tuning
• register allocation
• instruction scheduling
• issue relating object-code results to IR at
  various optimization levels

36
Backend

• register allocation
• allocation quality limits inlining,
  specialization
• especially important on 6-register CPU
• instruction scheduling
• CISCification, rematerialization

37
Storage Management

• improve tables, metadata mechanisms
• object layout, splitting, inlining
• compile time optimization, staging of runtime
  operations
• compile time gc (reuse, regions)

38
Language Issues

• annoyances scoping, co/contravariance
• parametric polymorphism and abstract datatypes
• invariants/design by contract
• alternative typing disciplines
• module systems

39
Q A