How not to Design a Scripting Language

How not to Design a Scripting Language

Paul Biggar

Department of Computer Science and StatisticsTrinity College Dublin

StackOverflow London, 28th October, 2009

How not to Design a Scripting Language Paul Biggar

How not to Design a Scripting Language

Paul Biggar

Department of Computer Science and StatisticsTrinity College Dublin

StackOverflow London, 28th October, 2009

About me

• PhD candidate, Trinity College Dublin• Topic: Compilers, optimizations, scripting languages.

PhD DissertationDesign and Implementation of an Ahead-of-time PHP Compiler

phc (http://phpcompiler.org)


1. means I justsubmitted

2. So that’s what thistalk is about

About me




About me





1. means I justsubmitted

2. So that’s what thistalk is about

3. Not much PHP

About me




How not to design a scripting language

• Compilers• Scripting Languages

• Speed


1. really, this is a talkabout compilersand scriptinglanguages



• Speed



• Speed


1. Thats whats in theback of my head allthe time



• Speed

What is a scripting language?• Javascript• Lua• Perl• PHP• Python• Ruby

Common Features:• Dynamic typing• Duck typing• Interpreted by default• FFI via C API


1. well, no-oneactually has agood definition ofscripting language!

2. I’m not talkingabout Bash orPowershell or VB,or some littlelanguage youwrote last week






1. well, no-oneactually has agood definition ofscripting language!

2. I’m not talkingabout Bash orPowershell or VB,or some littlelanguage youwrote last week

3. exposing theirinternals



Language implementation

• Interpreters: Easy, portable

• Compilers: Not too hard, sometimes portable,optimizations

NOT THE DRAGON BOOKEngineering a Compiler by Cooper/Torczon

Modern Compiler Implementation in X by Appel

• Just-in-time compilers: Very difficult, unportable, fastinterpreter.


1. Reads one line ata time (kinda)

2. hence used inmany scriptinglangs


• Interpreters: Easy, portable

• Compilers: Not too hard, sometimes portable,optimizations





• Interpreters: Easy, portable• Compilers: Not too hard, sometimes portable,

optimizations





1. Converts sourcecode into machinecode programs(kinda)

2. Lots of time tooptimize



optimizations






optimizations





1. 2 months, Joel,Dragon

2. For intermediateand advanced,there are alsomuch better books



optimizations






optimizations





1. Amazingoptimizations;Hotspot; dispatch,exceptions andarithmetic

2. Shockingly difficultto write



optimizations




What’s right with scripting languages?

1 Elegant and well designed,2 High level of abstraction,3 Dynamic typing (and duck typing).


1. dont dwell here -I’m not here toconvince you

2. the bad things donttouch these




1 Elegant and well designed,

2 High level of abstraction,3 Dynamic typing (and duck typing).




3. Python, Ruby; JStoo in a certain way


1 Elegant and well designed,

2 High level of abstraction,3 Dynamic typing (and duck typing).


1 Elegant and well designed,2 High level of abstraction,

3 Dynamic typing (and duck typing).




3. get things done, asJoel would say


1 Elegant and well designed,2 High level of abstraction,

3 Dynamic typing (and duck typing).






3. avoids manyproblems inherentin Java, C# andC++: verbosity,type systems



What’s wrong with scripting languages?

Symptoms: Speed, Portability

Problem: Language designed for interpretersProblem: Language designed for one specific interpreter

• Run-time source code execution• Foreign Function Interface


1. portable to .Net,JVM, etc?

2. not fast means youcant use it as much



Problem: Language designed for interpretersProblem: Language designed for one specific interpreter




Problem: Language designed for interpreters

Problem: Language designed for one specific interpreter

• Run-time source code execution

• Foreign Function Interface








• Run-time source code execution

• Foreign Function Interface














FFI

FFI

Foreign Function Interface based on CPython interpreter

• Access to C libraries• Script C applications using Python scripts• Rewrite hot code in C


1. define FFI2. glue code in C;

wrap things inpython datastructures; exposeinterpreterinternals

3. Works for allexcept JS - I’musing Python asexample

FFI

Foreign Function Interface based on CPython interpreter

• Access to C libraries• Script C applications using Python scripts• Rewrite hot code in C

FFI

FFI (good) implications

• Libraries not that slow• Can break out of Python for slow code.


FFI (good) implications

• Libraries not that slow• Can break out of Python for slow code.

FFI

FFI (bad) implications

• Language is allowed to be slow• Must break out of Python for speed.


1. nice things aboutpython are lost(high level, elegant

FFI (bad) implications

• Language is allowed to be slow• Must break out of Python for speed.

FFI

FFI (worse) implications

• Legacy issues

• Reimplementations


1. you cant changethe interpretermuch - it cant getfaster


• Legacy issues

• Reimplementations

FFI


• Legacy issues• Reimplementations


1. Jython,IronPython, PyPy,cant use the samecode

2. cant evenreimplement ownlanguage

3. by constrast, lookat JS


• Legacy issues• Reimplementations

FFI

FFI solution

Don’t expose yourself!

• Importing functions into Python with a Domain SpecificLanguage is good

• Only one way of FFI is better• Declarative is best

• Any reimplementation can reuse the same libraries withoutany modifications

• CPython itself can change without hassle


1. rule for life, really2. like pyrex, ctypes,

etc3. no implementation

specific code at all4. import functions

directly, andaccess them fromwithin Pythonwithout a line of C

5. Ruby libFFI

FFI solution






FFI

FFI solution



• Only one way of FFI is better

• Declarative is best





etc3. for

reimplementation4. no implementation



6. Ruby libFFI

FFI solution



• Only one way of FFI is better

• Declarative is best



FFI

FFI solution











5. Ruby libFFI

FFI solution






FFI

FFI solution











5. Ruby libFFI

FFI solution






Compiled and interpreted models

Dynamic source code generation

• eval and dynamic include/import

• meta-programming

eval (mysql_read (...)[0]);

• .rc files

username = "myname"password = "mypass"server = "srv.domain.com"

• localization

$lang = ....;include ("localisation/locale.$lang.php");


1. Should be used forsource inclusion.But becausethey’re interpreted,they can be usedfor:


• eval and dynamic include/import

• meta-programming


• .rc files


• localization




• eval and dynamic include/import• meta-programming


• .rc files


• localization







• .rc files


• localization






• .rc files


• localization







• .rc files


• localization






• .rc files


• localization







• .rc files


• localization




We don’t even know the fullprogram source!!


1. go to next slidestraight away


We don’t even know the fullprogram source!!


So they can’t be compiled (ahead-of-time)Downsides:

• Must use FFI for speed• Static analysis• Cool optimizations can’t happen

t = ...;for (i = 0; i < strlen(t); i++){

s[i] = t[i];}

t = ...;_temp = strlen(t);for (i = 0; i < _temp; i++){

s[i] = t[i];}

alert ($(’li’).get(0).nodeName);alert ($(’li’)[0].nodeName);


1. self-perpetuatingcycle

2. shame, scriptinglanguages couldreally use staticanalysis

3. high-level meanshigh leveloptimizations? no

4. redundencyelimination - somecant be done byhand



t = ...;for (i = 0; i < strlen(t); i++){s[i] = t[i];

}

t = ...;_temp = strlen(t);for (i = 0; i < _temp; i++){s[i] = t[i];

}





t = ...;for (i = 0; i < strlen(t); i++){

s[i] = t[i];}


s[i] = t[i];}










}


}





t = ...;for (i = 0; i < strlen(t); i++){

s[i] = t[i];}


s[i] = t[i];}










}


}





t = ...;for (i = 0; i < strlen(t); i++){

s[i] = t[i];}


s[i] = t[i];}

alert ($(’li’).get(0).nodeName);

alert ($(’li’)[0].nodeName);









}


}






t = ...;for (i = 0; i < strlen(t); i++){

s[i] = t[i];}


s[i] = t[i];}











}


}




JIT compiled

Tracemonkey

http://hacks.mozilla.org/2009/07/tracemonkey-overview/

Type Analysis for Javascript

Simon Holm Jensen, Anders Møller and Peter ThiemannSAS ’09http://www.brics.dk/TAJS/


1. instead, they mustbe JIT compiled.We’ll still never getcool optimizations

2. worse, theresearch just getsus, at run-time,expensively,information wecould get atcompile-time.

JIT compiled

Tracemonkey





JIT compiled

Tracemonkey





1. instead, they mustbe JIT compiled.We’ll still never getcool optimizations

2. worse, theresearch just getsus, at run-time,expensively,information wecould get atcompile-time.

JIT compiled

Tracemonkey





Fix at language design time

• No dynamic include; no eval.• Compile-time meta-programming

• .rc files• localization


1. do it like Perl, orC++


• No dynamic include; no eval.• Compile-time meta-programming

• .rc files• localization



• No dynamic include; no eval.• Compile-time meta-programming• .rc files

• localization


1. sandbox them


• No dynamic include; no eval.• Compile-time meta-programming• .rc files

• localization



• No dynamic include; no eval.• Compile-time meta-programming• .rc files• localization


1. With a compiledmodel, we know allthe files


• No dynamic include; no eval.• Compile-time meta-programming• .rc files• localization


Doing it right

• Factor• compiled model• compile-time meta-programming• declarative FFI


1. unfortunatelychose a paradigmthat nobody knows:stack-based;whoops!

Doing it right

• Factor• compiled model• compile-time meta-programming• declarative FFI


Open research problems

• Optimizing boxing• High-level optimizations• Combining ahead-of-time and JIT compilation


1. this is the shit weshould beresearching,instead of findingways aroundintracaibleproblems

2. javascript 10px3. table based4. should be trivial

Open research problems

• Optimizing boxing• High-level optimizations• Combining ahead-of-time and JIT compilation


Conclusion

Design the next scriptinglanguage right


Conclusion

Design the next scriptinglanguage right

Documents

How not to Design a Scripting Language