Share
Embed
Metaprogramming, Metaobject Protocols, Gradual Type Checks: Optimizing the "Unoptimizable" Using Old Ideas
Metaobject Protocols and Type Checks, do they have much in common? Perhaps not from a language perspective. However, under the hood of a modern virtual machine, they turn out to show very similar behavior and can be optimized very similarly.
This talk will go back to the days of Terminator 2, The Naked Gun 2 1/2, and Star Trek VI. We will revisit the early days of just-in-time compilation, the basic insights that are still true, and see how to apply them to metaprogramming techniques of different shapes and forms.
Stefan Marr
October 20, 2019
Transcript
62 slides
-
Metaprogramming, Metaobject Protocols, Gradual Type Checks Optimizing the “Unoptimizable” Using
Old Ideas Stefan Marr Athens, October 2019 Creative Commons Attribution-ShareAlike 4.0 License -
The Knights on the Quest for Excellent Performance 4 Excellent
Performance Reflection Metaprogramming Gradual Typing Metaobject Protocols -
Metaprogramming 6 meth.invoke() 1.7x slower Dynamic Proxies 7.5x slower Reflection
Metaprogramming Excellent Performance Don’t Use It Inn -
Metaobject Protocols LoggingClass extends Metaclass { writeToField(obj, fieldName, value) {
console.log(`${fieldName}: ${value}`) obj.setField(fieldName, value) } } 7 Metaobject Protocols Excellent Performance -
Metaobject Protocols LoggingClass extends Metaclass { writeToField(obj, fieldName, value) {
console.log(`${fieldName}: ${value}`) obj.setField(fieldName, value) } } 8 Metaobject Protocols Excellent Performance AOP -
Gradual Typing async addMessage(user: User, message) { const msg =
`<img src="/image/${user.profilePicture}"> ${message}</span>`; this.outputElem .insertAdjacentHTML( 'beforeend', msg); 9 Excellent Performance Gradual Typing -
Gradual Typing async addMessage(user: User, message) { const msg =
`<img src="/image/${user.profilePicture}"> ${message}</span>`; this.outputElem .insertAdjacentHTML( 'beforeend', msg); 10 Excellent Performance Gradual Typing Somewhat True… The whole Truth is a little more complex -
The Knights Found New Homes 11 Excellent Performance Reflection Metaprogramming
Gradual Typing Metaobject Protocols AOP Land of Engineering Short Cuts -
The (Movie) Heroes of ‘91 15 Polymorphic Inline Caches Just-in-time
Compilation Maps (Hidden Classes) Terminator 2 The Naked Gun 2 1/2 Star Trek VI -
A Class Hierarchy of Widgets 18 class Widget { fitsInto(width)
{ return this.width <= width; } } class Button extends Widget {} class RadioButton extends Button {} fn findAllThatFit(arr, width) { const result = []; for (const w of arr) if (w.fitsInto(width)) result.append(w) return result; } -
Lookups can be frequent and costly 19 class Widget {
fitsInto(width) { return this.width <= width; } } class Button extends Widget {} class RadioButton extends Button {} fn findAllThatFit(arr, width) { const result = []; for (const w of arr) if (w.fitsInto(width)) result.append(w) return result; } RadioButton Button fitsInto Widget superclass superclass For each fitsInto call hasMethod: 3x getSuperclass: 2x -
Solution: Lookup Caching 20 w.fitsInto(width) could be various functions, but
we don’t need to do the same lookup repeatedly method method (in case we see widget of different class) PIC: check for receiver and jump to method directly in machine code Useful because: • Most sends are monomorphic • Few are polymorphic • And just a couple are megamorphic -
Just-in-time Compilation • Produces native code, optimized, avoiding the overhead
of interpretation • At run time, can utilize knowledge about program execution • Ahead-of-time compilation, i.e., classic static compilation can only guess how a program is used 24 Missing Slide added after the talk -
With PICs, we can know 25 class Widget { fitsInto(width)
{ return this.width <= width; } } class Button extends Widget {} class RadioButton extends Button {} fn findAllThatFit(arr, width) { const result = []; for (const w of arr) if (w.fitsInto(width)) result.append(w) return result; } 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 RadioButton Widget.fitsInto Array RadioButton Integer -
And Generate Efficient Code 26 fn findAllThatFit(arr, width) { const
result = []; for (const w of arr) if (w.width(field) <=(int) width) result.append(w) return result; } 1 2 3 4 5 6 7 inlined fitsInto Specialized to types Resulting in Efficient Machine Code -
Polymorphic Inline Caches • Give us type information • Enable
method inlining 27 -
Our Hero has a lot of Luck and Achieves his
Goal 28 Relies on Developers not utilizing flexibility And, understanding 90’s humor -
The Power of Dynamic Languages 30 o = {foo: 33}
Object with 1 field o.bar = new Object() Object with 2 fields o.float = 4.2 Object with 3 fields o.float = "string" And you can store anything -
Data Representation for Objects 31 o = {foo: 33} o.bar
= new Object() o.baz = "string" o.float = 4.2 Obj 1 2 3 … 8 foo 33 "string" 4.2 bar baz float Full Power of Dynamic Languages: Rarely Used -
Maps, Hidden Classes, Object Shapes 32 o = {foo: 33}
o.bar = new Object() o.baz = "string" o.float = 4.2 Obj 33 4.2 "string" Shape 1: foo(int) 2: baz(ptr) 3: float(float) 4: bar(ptr) Combined with inline caches, a field access is a simple access at memory offset -
METAPROGRAMMING Is there hope for our Knight from the Land
of Engineering Shortcuts to find the true treasure? 34 Reflection Metaprogramming Excellent Performance -
Optimizing Reflective Method Invocation 37 cnt.invoke("+", [1]) + string Int.+
method Nesting of Lookup Caches Eliminates Potential Variability * string Int.* method -
Zero-Overhead Metaprogramming: Reflection and Metaobject Protocols Fast and without Compromises.
Marr, S., Seaton, C. & Ducasse, S. (2015). PLDI’15 Simple Metaprogramming: Zero Overhead 38 http://stefan-marr.de/papers/pldi-marr-et-al-zero-overhead-metaprogramming-artifacts/ Reflective & Direct: Identical Machine Code! -
Ruby Image Processing Code using Metaprogramming 39 • • •
• • 10.0 12.5 15.0 17.5 20.0 Compose Color Burn Compose Color Dodge Compose Darken Compose Difference Compose Exclusion Compose Hard Light Compose Hard Mix Compose Lighten Compose Linear Burn Compose Linear Dodge Compose Linear Light Compose Multiply Compose Normal Compose Overlay Compose Pin Light Compose Screen Compose Soft Light Compose Vivid Light Speedup over unoptimized (higher is better) Speedup over unoptimized (higher is better) -
METAOBJECT PROTOCOLS Is there hope for our Knight from the
Land of Engineering Shortcuts to find the true treasure? 40 Excellent Performance Metaobject Protocols -
Metaobject Protocols WriteLogging extends Metaclass { writeToField(obj, fieldName, value) {
console.log(`${fieldName}: ${value}`) obj.setField(fieldName, value) } } 41 Redefine the Language from within the Language -
Problem obj.field = 12; writeToField(obj, "field", 12) fn writeToField(obj, fieldName,
value) { console.log(`${fieldName}: ${value}`) obj.setField(fieldName, value) } } 42 turns into AOP Looks very Hard! -
Ownership-based Metaobject Protocol Building a Safe Actor Framework class ActorDomain
extends Domain { fn writeToField(obj, fieldIdx, value) { if (Domain.current() == this) { obj.setField(fieldIdx, value); } else { throw new IsolationError(obj); } } /* ... */ } 43 http://stefan-marr.de/research/omop/ -
An Actor Example 44 actor.fieldA := 1 semantic depends on
metaobject AD.writeToField Cache Desired Language Semantics Eliminates Potential Variability Std write -
OMOP Overhead 45 meta-tracing partial evaluation Overhead: 4% (min. -1%,
max. 19%) Overhead: 9% (min. -7%, max. 38%) • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • 1.00 1.05 1.10 1.15 1.20 Bounce BubbleSort Dispatch Fannkuch Fibonacci FieldLoop IntegerLoop List Loop Permute QuickSort Recurse Storage Sum Towers TreeSort WhileLoop DeltaBlue Mandelbrot NBody Richards Runtime Ratio to run without OMOP SOMMT • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • 0.8 1.0 1.2 1.4 Bounce BubbleSort Dispatch Fannkuch Fibonacci FieldLoop IntegerLoop List Loop Permute QuickSort Recurse Storage Sum Towers TreeSort WhileLoop DeltaBlue Mandelbrot NBody Richards Runtime Ratio to run without OMOP SOMPE -
GRADUAL TYPING Is there hope for our Knight from the
Land of Engineering Shortcuts to find the true treasure? 47 Excellent Performance Gradual Typing -
Gradual Typing without Run-Time Semantics async addMessage(user: User, message) {
const msg = `<img src="/image/${user.profilePicture}"> ${message}</span>`; this.outputElem .insertAdjacentHTML( 'beforeend', msg); 48 Very Useful in Practice. And rather popular. Gradual Typing -
Transient Gradual Typing type Vehicle = interface { registration registerTo(_)
} type Department = interface { code } var companyCar: Vehicle := object { method registerTo(d: Department) { print "{d.code}" } } companyCar.registerTo( object { var name := "R&D" }) 49 Types are shallow. Method names matter, but arguments don’t. Object only has name, no code method Assignment to registerTo(d: Department) should error -
Transient Gradual Typing tmp := object { method registerTo(d) {
typeCheck d is Department print "{d.code}" } } typeCheck tmp is Vehicle var companyCar = tmp companyCar.registerTo( object { var name := "R&D" }) 50 Very simple semantics. Other Gradual systems have blame, and are more complex Possibly many many checks. Looks very Hard! Gradual Typing -
How to get rid of these checks without losing run-time
semantics ? tmp := object { method registerTo(d) { typeCheck d is Department print "{d.code}" } } typeCheck tmp is Vehicle var companyCar = tmp companyCar.registerTo( object { var name := "R&D" }) 51 -
Shapes to the Rescue 52 Shape 1: foo(int) 2: baz(ptr)
3: float(float) 4: bar(ptr) Shape 1: foo(int) 2: baz(ptr) 3: float(float) 4: bar(ptr) Implicitly Compatible to: - Type 1 - Type 2 1. Check object is compatible 2. Shape implies compatibility -
Final optimized code tmp := object { method registerTo(d) {
check d hasShape s1 print "{d.code}" } } check tmp hasShape s2 var companyCar = tmp companyCar.registerTo( object { var name := "R&D" }) 53 need to do type check only once per lexical location s1.code s2.registerTo JIT Compiler can remove redundant checks -
Our Heroes 56 Eliminates Potential Variability Provides a Supporting Structure
Makes it Fast With a little Luck Polymorphic Inline Caches Just-in-time Compilation Maps, Hidden Classes, Shapes -
Things I didn’t talk about Failure cases: Deoptimization An Efficient
Implementation of SELF a Dynamically-Typed Object- Oriented Language Based on Prototypes. Chambers, C., Ungar, D. & Lee, E. (1989). OOPSLA’89 Object shapes are useful for other things Efficient and Thread-Safe Objects for Dynamically- Typed Languages. B. Daloze, S. Marr, D. Bonetta, and H. Mössenböck. OOPSLA'16 58 And many other modern optimizations -
Our Knights Made it With Some Help of our 90’s
Heroes 59 Excellent Performance Reflection Metaprogramming Gradual Typing Metaobject Protocols -
Research and Literature • Efficient Implementation of the Smalltalk-80 System.
Deutsch, L. P. & Schiffman, A. M. (1984). POPL’84 • Optimizing Dynamically-Typed Object-Oriented Languages With Polymorphic Inline Caches. Hölzle, U., Chambers, C. & Ungar, D. (1991). ECOOP’91 • Zero-Overhead Metaprogramming: Reflection and Metaobject Protocols Fast and without Compromises. Marr, S., Seaton, C. & Ducasse, S. (2015). PLDI’15 • Optimizing prototypes in V8 https://mathiasbynens.be/notes/ prototypes • https://mathiasbynens.be/notes/ shapes-ics • https://mrale.ph/blog/2012/06/0 3/explaining-js-vms-in-js-inline- caches.html 61 -
Research and Literature • An Efficient Implementation of SELF a
Dynamically-Typed Object-Oriented Language Based on Prototypes. Chambers, C., Ungar, D. & Lee, E. (1989). OOPSLA’89 • An Object Storage Model for the Truffle Language Implementation Framework. A. Wöß, C. Wirth, D. Bonetta, C. Seaton, C. Humer, and H. Mössenböck. PPPJ’14. • Storage Strategies for Collections in Dynamically Typed Languages. C. F. Bolz, L. Diekmann, and L. Tratt. OOPSLA’13. • Memento Mori: Dynamic Allocation-site-based Optimizations. Clifford, D., Payer, H., Stanton, M. & Titzer, B. L. (2015). ISMM’15 62 -
Research and Literature • Virtual Machine Warmup Blows Hot and
Cold. Barrett, E., Bolz-Tereick, C. F., Killick, R., Mount, S. & Tratt, L. (2017). OOPSLA’17 • Quantifying Performance Changes with Effect Size Confidence Intervals. Kalibera, T. & Jones, R. (2012). Technical Report, University of Kent. • Rigorous Benchmarking in Reasonable Time. Kalibera, T. & Jones, R. (2013). ISMM’13 • How Not to Lie With Statistics: The Correct Way to Summarize Benchmark Results. Fleming, P. J. & Wallace, J. J. (1986). Commun. ACM • SIGPLAN Empirical Evaluation Guidelines https://www.sigplan.org/Resources/E mpiricalEvaluation/ • Systems Benchmarking Crimes, Gernot Heiser https://www.cse.unsw.edu.au/~gern ot/benchmarking-crimes.html • Benchmarking Crimes: An Emerging Threat in Systems Security. van der Kouwe, E., Andriesse, D., Bos, H., Giuffrida, C. & Heiser, G. (2018). arxiv:1801.02381 • http://btorpey.github.io/blog/2014/0 2/18/clock-sources-in-linux/ • Generating an Artefact From a Benchmarking Setup as Part of CI https://stefan- marr.de/2019/05/artifacts-from-ci/ 63