The sinuous path to Valhalla

Presentation of the OpenJDK project Valhalla at the meetup "dev de l'est parisien"

Transcript

1. The sinuous path to Valhalla Rémi Forax
2. Me, Myself and i Assistant Prof at Paris East University Open Source dev OpenJDK, ASM, Tatoo, Pro, etc Java Spec expert One of the Father of invokedynamic (Java 7) lambda (Java 8), module (Java 9), constant dynamic (Java 11) record (Java 14/15) and valhalla (Java ??)
3. We need you ! (or your money)
4. I’m not an Oracle Employee ! Java is open source since a long time now
5. Don’t believe what I’m saying !
6. The sinuous path to Valhalla Rémi Forax
7. Intro
8. Java at its inception Hardware in 1993 – Intel 486 SX 25 Mhz / 8 MB of RAM – Pointer dereference ~~ one addition Java Model – Every objects should be hold by references
9. Java Nowadays Hardware – Intel Kaby Lake 2.4 Ghz / 16 GB of RAM – Pointer dereference ~~ 500 additions Java Model – Every objects are still hold by references • Unless the JIT can prove that the identity is not used Escape Analysis (if … )
10. On heap representation An array of class Complex header length header re im header re im header re im header re im ...
11. Inline object (first draft) An inline object – no identity (no address) Runtime model – inlined into another object/an array cells on heap – scalarization into registers on stack
12. On heap representation An array of an inline class Complex no pointer no header at runtime header length re im re im ... re im re im
13. C struct ? typedef struct _complex { double re; double im; } complex; in C – An array of struct is not an array of pointers – Stack allocation can be scalarize • depends on inlining
14. Aliasing issue in C/C++ Same address in memory ?? void foo(complex& c1, complex& c2) { c1.re = 5; // is c2.re changed ? } Solution: make it immutable !
15. Inline object An inline object – no identity (no address) – not mutable – not nullable (no pointer) Runtime model – scalarization into registers on stack – inlined into another object/an array cells on heap
16. Inlineable What if ? – Inline class or array too big ? – Concurrent read/write of an inline object Solution: Inline objects are inlineable – VM decides if it is inlined or not • volatile/static fields are never inlined
17. Examples
18. Inline class Code like a class, works like an int public inline class Complex { private final double re; private final double im; public Complex(double re, double im) { this.re = re; this.im = im; } public Complex add(Complex c) { return new Complex(re + c.re, im + c.im); } }
19. Inline class a class with – keyword “inline” – no superclass but interfaces – all fields are final • Can not create a cycle ! – any methods
20. Mandelbrot Complex provides a useful abstraction for (var col = -32; col < +32; col++) { for (var row = -32; row < +32; row++) { var c = new Complex(row, col); var zn = new Complex(0.0, 0.0); var iteration = 0; while (zn.modulus() < 4 && iteration < max) { zn = zn.square().add(c); iteration++; } if (iteration < max) { image.setRGB(col, row, colors[iteration]); } else { image.setRGB(col, row, black); } } }
21. Benchmark with JMH Time to render the Mandelbrot 64x64 with 1000 iterations? With java -XX:+EnableValhalla score error primitive 1540 µs/op 22 µs/op indirect 2816 µs/op 48 µs/op inline 1541 µs/op 12 µs/op
22. Example: IntBox java.lang.Integer written as an inline class public inline class IntBox { private final int value; public IntBox(int value) { this.value = value; } public int intValue() { return value; } public static IntBox zero() { return new IntBox(0); } }
23. Array of inline objects Sum of an array of IntBox private static final IntBox[] ARRAY = new IntBox[100_000]; static { range(0, ARRAY.length).forEach(i -> ARRAY[i] = new IntBox(i)); } @Benchmark public int sum_IntBox() { var sum = 0; for(var value : ARRAY) { sum += value.intValue(); } return sum; }
24. Benchmark with JMH Time to sum an array of 100 000 values With java -XX:+EnableValhalla score error primitive (int) 27.1 µs/op 0.2 µs/op indirect (j.l.Integer) 60.7 µs/op 0.6 µs/op inline (IntBox) 27.0 µs/op 0.1 µs/op
25. Array of inline objects What if we shuffle the array ? private static final IntBox[] ARRAY = new IntBox[100_000]; static { range(0, ARRAY.length).forEach(i -> ARRAY[i] = new IntBox(i)); Collections.shuffle(Arrays.asList(ARRAY)); } @Benchmark public int sum_IntBox() { var sum = 0; for(var value : ARRAY) { sum += value.intValue(); } return sum; }
26. Benchmark with JMH Time to sum an array of 100 000 shuffled values With java -XX:+EnableValhalla score error primitive (int) 27.1 µs/op 0.2 µs/op indirect (j.l.Integer) 121.9 µs/op 5.1 µs/op inline (IntBox) 27.0 µs/op 0.1 µs/op
27. Inline class & concurrency
28. Problem: Value tearing inline class Value { final int x, y; public Value(int x, int y) { … } } public static void main(String[] args) { var box = new Object() { Value shared; }; var zero = new Value(0, 0); var one = new Value(1, 1); new Thread(() -> { for(;;) { box.shared = zero; } }).start(); new Thread(() -> { for(;;) { box.shared = one; } }).start(); for(;;) { var val = box.shared; if (val.x != val.y) { throw new AssertionError("oops"); } }
29. Problem: Value tearing inline class Value { final int x, y; public Value(int x, int y) { … } } public static void main(String[] args) { var box = new Object() { Value shared; }; var zero = new Value(0, 0); var one = new Value(1, 1); new Thread(() -> { for(;;) { box.shared = zero; } }).start(); new Thread(() -> { for(;;) { box.shared = one; } }).start(); for(;;) { var val = box.shared; if (val.x != val.y) { throw new AssertionError("oops"); } } Two stores Two reads Unprotected concurrent access
30. Value tearing Accessing a value type concurrently => several reads/stores Not a new problem long/double on 32bits VMs has the same issue No real solution => don’t do that !
31. Bytecode representation
32. Problem: class late loading In Java, a class is loaded as late as possible class Holder { … } class AnotherClass { public static Holder getMeAHolder() { return null; } } Even when getMeAHolder() is executed, Holder is not loaded
33. Inline class loading but inline classes need to be loaded early – memory layout – method calling sequence • Scalarization even if method not inlined ! – verifier checks null pollution solution: prefix class by ‘Q’ instead of ‘L’ ‘Q’ => loaded early
34. IntBox (again) java.lang.Integer written as an inline class public inline class IntBox { private final int value; public IntBox(int value) { this.value = value; } public int intValue() { return value; } – public static IntBox zero() { return new IntBox(0); } }
35. IntBox as classfile An inline class descriptor is prefixed by ‘Q’ public final inline class IntBox { private final int value; descriptor: I public static IntBox zero(); descriptor: ()QintBox; ... }
36. Problem: class initialization In a constructor, you can see the fields being initialized public inline class IntBox { private final int value; public IntBox(int value) { // here I can see the mutation !! this.value = value; } ... }
37. True immutability ? In a constructor, you can see the fields being initialized => not truly immutable Solution: only initialize through a static factory – Compiler transforms constructors to static methods “this” should not escape ! – Add 2 new opcodes: defaultvalue and withfield
38. Bytecode for IntBox constructor public final inline class IntBox { private final int value; public static <init>(I)QIntBox; 0: defaultvalue #1 // class IntBox 3: astore_1 4: iload_0 5: aload_1 6: swap 7: withfield #3 // Field value:I 10: astore_1 11: aload_1 12: areturn
39. Exposing defaultvalue An inline object has a default values All fields initialized to null, 0, 0.0, false, etc Java syntax .default var empty = Complex.default;
40. IntBox The opcode defaultvalue is accessible in Java public inline class IntBox { private final int value; public IntBox(int value) { this.value = value; } public static IntBox zero() { return IntBox.default; } }
41. OpenJDK projects dependency a.k.a more use cases
42. OpenJDK projects dependency Panama – Vector API (SIMD, AVX support) • Long128, Long256, etc – Native Access API (safe/managed malloc) • Pointer, Slice, etc Amber – Pattern matching with de-construction • Carrier object
43. Roadmap
44. Current Roadmap Early access prototypes (http://jdk.java.net/) – MVT – Lworld 1, 2 => We are here now ! – Lword 10 - preview release in Java • Support erased generics – Lworld 50 • Retrofit primitives to be inline • Support Value Based Classes – Lworld 100 – finale release • Reified generics
45. Integration with existing types
46. Minimum Value Type (first prototype)
47. j.l.Value vs j.l.Object java.lang.Object indirect objects java.lang.Integer java.util.ArrayList java.util.AbstractList
48. j.l.Value vs j.l.Object java.lang.Object indirect objects inline objects Complex IntBox java.lang.Integer java.util.ArrayList java.util.AbstractList java.lang.Value Complex.box IntBox.box boxing boxing
49. MVT Good perf with inline classes only code ! No perf regression in existing code ! Doesn’t work well with real applications Requires too much boxing • println(Object) • generics: collections, streams
50. L-world (second prototype)
51. Object = root of all indirect objects java.lang.Object indirect objects java.lang.String java.util.Optional java.lang.CharSequence java.util.ArrayList java.util.AbstractList
52. Object = root of every objects java.lang.Object indirect objects inline objects java.lang.String Complex IntBox java.lang.Integer java.lang.CharSequence java.util.ArrayList CompactString java.util.AbstractList
53. L-world conversion Subtype relationship Complex c = ... Object o = c; Downcast (checkcast + nullcheck) Object o2 = … Complex c2 = (Complex) o2;
54. Toward the preview release
55. Problem: array covariance An array of Object • an array of indirect type (array of pointer) • an array of inline type (array of structs) => size of an array cell not constant ! Iterating over an array can be megamorphic ! Solution: Aggressive loop hoisting + loop duplication
56. Problem: subtyping relationship All operations available on Object should be available on an inline class – No header • synchronized, wait/notify ? • System.identityHashCode ? – No identity • Meaning of o == o2 if inline objects ?
57. No header synchronized, wait/notify ? Throw an IllegalMonitorStateException • Avoid perf regression ? System.identityHashCode ? Redirected to hashCode()
58. No identity, == (aka acmp) 3 possible semantics – always return false (no pointer) – do a classcheck + components checks • perf regression in existing code ? – may stackoverflow / very slow if recursive – redirect to call equals() • Perf regression ? Still in flux – I prefer the former one (demote ==)
59. Problem: erased generics Inside a generics class, T is nullable but inline classes are not nullable Map<String,Complex>.get() – Can return null (API contract) – Can not return null (Complex not nullable)
60. Eclair interface
61. Eclair interface Write an inline class – derives automatically an empty sealed interface – Add unboxing between eclair and inline Non nullable type → inline class Nullable type → eclair interface
62. Example public inline class Complex { private final double re; private final double im; public Complex add(Complex c) { return new Complex(re + c.re, im + c.im); } }
63. Example public inline class Complex implements Complex.box { private final double re; private final double im; public Complex add(Complex c) { return new Complex(re + c.re, im + c.im); } // generated automatically public sealed interface box permits Complex { // empty } }
64. How to use it Erased generics don’t allow inline class List<Complex> l = … // doesn’t compile But using the eclair box is ok List<Complex.box> l = List.of(new Complex(…)); // subtyping Complex c = l.get(0); // unboxing, may throw NPE
65. Summary
66. Inline class - preview – no identity, not mutable, not nullable, tearable – runtime model • scalarization into registers on stack • inlined into another object/array cells on heap – Object as root of everything • don’t use == – erased generic => lightweight boxing (eclair)
67. Current Roadmap Early access prototypes (http://jdk.java.net/) – MVT – Lworld 1, 2 => We are here now ! – Lword 10 - preview release in Java • Support erased generics – Lworld 50 • Retrofit primitive to be inline • Support Value based classes – Lworld 100 – finale release • Reified generics