Graal and Truffle: One VM to Rule Them All

1. Graal and Truffle: One VM to Rule Them All Thomas Wuerthinger Oracle Labs @thomaswue 12-December-2013, at ETH Zurich

2. Disclaimer The following is intended to provide some insight into a line of research in Oracle Labs. It is intended for information purposes only, and may not be incorporated into any contract. It is not a commitment to deliver any material, code, or functionality, and should not be relied upon in making purchasing decisions. The development, release, and timing of any features or functionality described in connection with any Oracle product or service remains at the sole discretion of Oracle. Any views expressed in this presentation are my own and do not necessarily reflect the views of Oracle. 2 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

6. Relative Speed of Programming Languages (as measured by the Computer Language Benchmarks Game, ~1y ago) One VM to for all languages means interoperability and being able to choose the best language for the task! 3 Goal: 6 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

8. Static versus Dynamic Compilation (1) §  Static (or ahead-of-time) Compilation –  Compilation happens before program is run. –  Can include profiling feedback from sample application runs. §  Dynamic (or just-in-time) Compilation –  Compilation happens while the program is running. –  Base line execution (interpreter or simple compiler) gathers profiling feeback. –  Optimization => Deoptimization => Reoptimization cycles. –  On-stack-replacement (OSR) to switch between the tiers (two or more execution modes. 8 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

9. Static versus Dynamic Compilation (2) §  Static (or ahead-of-time) Compilation –  Fast start-up, because compilation and profiling is not part of application execution time. –  Predictable performance as only the source program affects the generated machine code. §  Dynamic (or just-in-time) Compilation –  Can exploit exact target platform properties when generating machine code. –  Profiling feedback captures part of the application behavior and increases code quality. –  The deoptimization capabilities allow the optimized code to be incomplete and/or use aggressive speculation. –  Can use assumptions about the current state of the system (e.g., loaded classes) in the generated code. 9 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

10. Profiling Feedback for Java §  Branch probabilities –  Never taken branches can be omitted. –  Exact probabilities allows if-cascade reordering. §  Loop frequencies –  Guide loop unrolling and loop invariant motion. §  Type profile –  Optimize instanceof, checkcast type checks (i.e., speculate that only a specific set of types occurs) –  Optimize virtual calls or interface calls. Profiling feedback only helps when the program behavior during the observed period matches the overall program behavior. 10 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

11. Static Single Assignment (SSA) Form §  Every variable is assigned only once. §  Phis capture values coming from different control flow branches. §  Commonly used in compilers as it simplifies optimizations and traversal along the def-use and use-def chain. ... if (condition) { x = value1 + value2; } else { x = value2; } return x; 11 Copyright © 2013, Oracle and/or its affiliates. All rights reserved. ... if (condition) { x1 = value1 + value2; } else { x2 = value2; } x3 = phi(x1, x2); return x3;

13. Graal is an … ... extensible, dynamic compiler using object-oriented Java programming, a graph intermediate representation, and Java snippets. 13 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

14. HotSpotVM versus GraalVM 30k LOC 120k LOC 60k LOC Client Server Graal Compiler Interface Compilation Queue Compilation Queue Compiler Interface HotSpot HotSpot C++ 14 Copyright © 2013, Oracle and/or its affiliates. All rights reserved. Java

15. Why Java? Robustness: Runtime exceptions not fatal. Reflection: Annotations instead of macros. Meta-Evaluation: IR subgraph expressible in Java code. Extensibility: No language barrier to the application. Tooling: Java IDEs speed up the development process. 15 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

16. Snippets for Graph Construction Manual construction: Node max(ValueNode a, ValueNode b) { IfNode ifNode = new IfNode(new IntegerLessThanNode(a, b)); ifNode.trueSuccessor().setNext(new ReturnNode(a)); ifNode.falseSuccessor().setNext(new ReturnNode(b)); return ifNode; } Expression as snippet: int max(int a, int b) { if (a > b) return a; else return b; } 16 Copyright © 2013, Oracle and/or its affiliates. All rights reserved. Data Code

17. Lowering §  Replace one node with multiple other nodes. –  New nodes provide more detailed description of semantics. –  New nodes can be optimized and moved separately. §  General Java lowerings –  Example: Replace an array store with null check, bounds check, store check, write operation. if (array != null && index >= 0 && index < array.length && canAssign(array.getClass().getComponentType(), value)) { *(array + 16 + index*8) = value; } else { deoptimize; } §  VM specific lowerings –  Examples: Replace a monitorenter with the code dependent on the locking schemes used by the VM 17 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

18. Gradual Lowering 3 Nodes per bytecode 2.5 2 Graal 1.5 Client Server 1 0.5 0 After parsing After optimizations After lowering Before code emission Numbers obtained while running the DaCapo benchmark suite. 18 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

19. Extensibility •  Multiple Target Platforms (AMD64, SPARC, PTX, HSAIL) •  Multiple Runtimes (HotSpot and Maxine) •  Adding new types of Nodes •  Adding new compiler Phases abstract class Phase { abstract void run(Graph g); } for (IfNode n : graph.getNodes(IfNode.class)) { ... } Compiler has about 100 different individual modules. 19 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

20. Graph IR •  Static single assignment (SSA) form with def-use and use-def edges. •  Program dependence graph (sea of nodes), but with explicit distinction between control flow and data flow edges. •  Graph visualization tools: IdealGraphVisualizer and c1visualizer. ... condition If ... if (condition) { result = value1 + value2; } else { result = value2; } return result; Begin Begin End End Merge Copyright © 2013, Oracle and/or its affiliates. All rights reserved. Add Phi Return 20 value1 value2

23. Eliding Exception Edges Catch Operation Operation Operation Actual Potential Invoke 1296646 14454 1.11% BoundsCheck 166770 498 0.30% NullCheck 1525061 686 0.04% OutOfMemory 110078 0 0.00% CheckCast 99192 0 0.00% DivRem 6082 0 0.00% MonitorNullCheck 33631 0 0.00% TOTAL 3237460 15638 0.48% Numbers obtained while running the DaCapo benchmark suite. 23 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

25. Java Peak Performance §  SPECjvm2008 114 120 100 100 80 76 60 40 20 0 Client Graal Server Conﬁgura*on: Intel Core i7-‐3770 @ 3,4 Ghz, 4 Cores 8 Threads, 16 GB RAM Comparison against HotSpot changeset tag hs25-‐b37 from June 13, 2013 25 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

26. Scala Peak Performance §  Scala-Dacapo Benchmark Suite 120 100 100 106 80 61 60 40 20 0 Client Graal Server Conﬁgura*on: Intel Core i7-‐3770 @ 3,4 Ghz, 4 Cores 8 Threads, 16 GB RAM Comparison against HotSpot changeset tag hs25-‐b37 from June 13, 2013 26 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

27. Your Compiler Extension? http://openjdk.java.net/projects/graal/ graal-dev@openjdk.java.net $ hg clone http://hg.openjdk.java.net/graal/graal $ cd graal $ ./mx.sh --vm graal build $ ./mx.sh ideinit $ ./mx.sh --vm graal vm §  Graal Resources https://wiki.openjdk.java.net/display/Graal/Main §  Graal License: GPLv2 27 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

29. “Write Your Own Language” Current situation Prototype a new language Parser and language work to build syntax tree (AST), AST Interpreter Write a “real” VM In C/C++, still using AST interpreter, spend a lot of time implementing runtime system, GC, … People start using it People complain about performance Define a bytecode format and write bytecode interpreter Performance is still bad Write a JIT compiler Improve the garbage collector 29 Copyright © 2013, Oracle and/or its affiliates. All rights reserved. How it should be Prototype a new language in Java Parser and language work to build syntax tree (AST) Execute using AST interpreter People start using it And it is already fast

30. Truffle: System Structure Written by: Application Developer Written in: Guest Language Application Guest Language Language Developer Guest Language Implementation Managed Host Language VM Expert Host Services Managed Host Language or Unmanaged Language OS Expert OS 30 Copyright © 2013, Oracle and/or its affiliates. All rights reserved. Unmanaged Language (typically C or C++)

31. Speculate and Optimize … Node Rewriting for Profiling Feedback U Compilation using Partial Evaluation G G U U Node Transitions U U I Uninitialized S AST Interpreter Uninitialized Nodes I G I I D String Double G Generic 31 I Integer I U I G Copyright © 2013, Oracle and/or its affiliates. All rights reserved. AST Interpreter Rewritten Nodes Compiled Code

32. Partial Evaluation §  Example function: –  f(x, y) = x + y + 1 §  Partial evaluation of example function: –  g(y) = f(1, y) = 1 + y + 1 = y + 2 §  Interpreter function: –  f(program, arguments) = calculations to interpret the program §  Partial evaluation of interpreter function (first Futamura projection): –  g(arguments) = f(#specificProgram, arguments) = compiled version of #specificProgram that takes arguments as parameters 32 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

33. … and Deoptimize and Reoptimize! Deoptimization to AST Interpreter Node Rewriting to Update Profiling Feedback G Recompilation using Partial Evaluation G G I I G I G D G I I I 33 D G Copyright © 2013, Oracle and/or its affiliates. All rights reserved. I I D G D

34. Object add(Object a, Object b) { if(a instanceof Integer && b instanceof Integer) { return (int)a + (int)b; } else if (a instanceof String && b instanceof String) { return (String)a + (String)b; } else { return genericAdd(a, b); } } int add(int a, String add(String a, int b) { Copyright © 2013, Oracle and/or its affiliates. All rights reserved. return genericAdd(a, b); return a + b; } 34 Object b) { String b) { return a + b; } Object add(Object a, }

35. Node Implementation class IAddNode extends BinaryNode { int executeInt(Frame f) throws UnexpectedResult { int a; try { a = left.executeInt(f); } catch (UnexpectedResult ex) { throw rewrite(f, ex.result, right.execute(f)); } int b; try { b = right.executeInt(f); } catch (UnexpectedResult ex) { throw rewrite(f, a, ex.result); } try { return Math.addExact(a, b); } catch (ArithmeticException ex) { throw rewrite(f, a, b); } } 35 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

37. Truffle DSL @Specialization(rewriteOn=ArithmeticException.class) int addInt(int a, int b) { return Math.addExact(a, b); } @Specialization double addDouble(double a, double b) { return a + b; } @Generic Object addGeneric(Frame f, Object a, Object b) { // Handling of String omitted for simplicity. Number aNum = Runtime.toNumber(f, a); Number bNum = Runtime.toNumber(f, b); return Double.valueOf(aNum.doubleValue() + bNum.doubleValue()); } 37 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

41. Truffle API Compiler Directives §  Guards if(condition) { // some code that is only valid if condition is true } else { CompilerDirectives.transferToInterpreter(); } §  Assumptions Assumption assumption = Truffle.getRuntime().createAssumption(); assumption.check(); // some code that is only valid if assumption is true assumption.invalidate(); 41 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

42. Performance Number Disclaimers §  All Truffle numbers reflect the current development snapshot. –  Subject to change at any time (hopefully improve) –  You have to know a benchmark to understand why it is slow or fast §  We are not claiming to have complete language implementations. –  JavaScript: quite complete, passing 99.8% of ECMAScript262 tests –  Ruby: passing >45% of RubySpec language tests –  R: early prototype §  We measure against latest versions of competitors. §  We measure peak performance (i.e., giving each benchmark enough iterations to warmup before starting measurement). §  Benchmarks that are not shown –  may not run at all, or –  may not run fast 42 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

43. Peak Performance: JavaScript Speedup relative to V8 2.6 3.0 Truffle SpiderMonkey 2.5 0.8 1.0 0.9 1.2 1.1 0.9 1.1 0.5 0.6 0.7 0.7 1.0 0.6 1.0 0.8 1.0 0.7 1.5 1.4 1.5 1.6 2.0 0.5 te u C om po si em gb x2 bo bo yrle d r ye y la ea na vi er -s to k sp es e ra y tra c to cr yp bl lta de ric ha rd s ue 0.0 Selection of benchmarks from Google‘s Octane benchmark suite v1.0 43 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

44. Peak Performance: Ruby Speedup relative to JRuby 1.7.5 14 14 14 16 MRI 2.0.0 Topaz 12 Truf f le 10 0 44 Copyright © 2013, Oracle and/or its affiliates. All rights reserved. 0.6 1.0 1.7 1.8 0.8 1.7 2.7 1.1 0.5 0.2 0.4 0.3 0.7 2 1.7 2.7 4 0.7 4.7 4.5 6 4.9 8

45. Peak Performance: R 94 Speedup relative to GNUR 100.0 90.0 80.0 70.0 60.0 22 30.0 0.0 45 Copyright © 2013, Oracle and/or its affiliates. All rights reserved. 0.8 2.7 2.1 10.0 2.0 14 20.0 23 40.0 24 38 39 50.0

47. Your Language? http://openjdk.java.net/projects/graal/ graal-dev@openjdk.java.net $ hg clone http://hg.openjdk.java.net/graal/graal $ cd graal $ ./mx.sh --vm server build $ ./mx.sh ideinit $ ./mx.sh --vm server unittest SumTest §  Truffle API Resources https://wiki.openjdk.java.net/display/Graal/Truffle+FAQ+and+Guidelines §  Truffle API License: GPLv2 with Classpath Exception 47 Copyright © 2013, Oracle and/or its affiliates. All rights reserved.

48. Acknowledgements Oracle Labs Laurent Daynès Erik Eckstein Michael Haupt Peter Kessler Christos Kotselidis David Leibs Roland Schatz Chris Seaton Doug Simon Michael Van De Vanter Christian Wimmer Christian Wirth Mario Wolczko Thomas Würthinger Laura Hill (Manager) Interns Danilo Ansaloni Daniele Bonetta Shams Imam Stephen Kell Gregor Richards Rifat Shariyar 48 Copyright © 2013, Oracle and/or its affiliates. All rights reserved. JKU Linz Prof. Hanspeter Mössenböck Gilles Duboscq Matthias Grimmer Christian Häubl Josef Haider Christian Humer Christian Huber Manuel Rigger Lukas Stadler Bernhard Urban Andreas Wöß University of Edinburgh Christophe Dubach Juan José Fumero Alfonso Ranjeet Singh Toomas Remmelg LaBRI Floréal Morandat University of California, Irvine Prof. Michael Franz Codrut Stancu Gulfem Savrun Yeniceri Wei Zhang Purdue University Prof. Jan Vitek Tomas Kalibera Petr Maj  Lei Zhao T. U. Dortmund Prof. Peter Marwedel Helena Kotthaus Ingo Korb University of California, Davis Prof. Duncan Temple Lang Nicholas Ulle