HiPE version 1.0 Kostis Sagonas Uppsala University.

HiPEHiPEversion 1.0version 1.0

Kostis SagonasUppsala University

Structure of this talk

1. Recent past

2. Present

3. Near future

Part 1: Recent Past

Historical introduction

During the last few years, the following people:

Erik “Happi” JohanssonMikael PetterssonRichard CarlssonKostis Sagonas

have been having fun developing and maintaining the HiPE system...

HiPE: High Performance Erlang

A “just-in-time” native code compiler for Erlang

– Allows flexible, user-controlled compilation of Erlang programs to native machine code

– Fine-grained: Compilation unit is a single function

• HiPE 0.92 released as open-source in March 2000 (see also www.csd.uu.se/projects/hipe)– Released version was JAM-based (Erlang R4)– Available only for SPARC machines

HiPE: Technical Details

• Both virtual machine code and native code can happily co-exist in the runtime system

• HiPE optimizes calls to functions which execute in the same mode

• Preserves tail-calls• The generated native code is quite efficient

– HiPE significantly outperforms all other Erlang implementations

– has similar performance to e.g. Concurrent SML/NJ

Erlang Run-Time System HiPE Compiler

JAMEmulator

Code area

JAMDissassembler

HiPELoader

JAM

Bytecode

OtherData

NativeCode

SymbolicJAM

Icode

RTL

SPARC

The HiPE system

Old HiPE Architecture

Speedup of HiPE (0.92) over other Erlang implementations

0

5

10

15

20

25

fib huff length nrev qsort smith tak decode

Sp

ee

du

p o

ve

r J

AM

HiPE

Etos

JAM

BEAM

Success of HiPE: Let’s take a quick poll...

Who has used HiPE ?

My educated guess: A few…

Who is using HiPE instead of using Erlang/OTP ?

Most probable answer: None!

Part 2: Present

HiPE version 1.0: The current HiPE team

At Uppsala University:Erik “Happi” Johansson

Mikael PetterssonRichard CarlssonKostis Sagonas

+Jesper Wilhelmsson

Recent addition from the Erlang/OTP team:Bjorn Gustavsson

HiPE version 1.0: Main Features of Interest

• HiPE is fully and tightly integrated within Open Source Erlang/OTP Release 8

• Handles the full Erlang language

• Back-ends for:– SPARC – x86-based machines running Linux or Solaris

• Usually very reasonable compilation times

• Acceptable sizes of object code

Erlang Run-Time System HiPE Compiler

BEAMEmulator

Code area

BEAMDissassembler

HiPELoader

BEAM

Bytecode

OtherData

NativeCode

SymbolicBEAM

Icode

RTL

SPARC X86

A HiPE-enabled Erlang/OTP system

New HiPE Architecture

HiPE version 1.0: Installation Guide

1. Get Open Source Erlang/OTP R8

2. If on SPARC or x86, instead of typing: ./configure ./make type: ./configure --enable-hipe ./make

HiPE: Invoking the compiler (novice user)

• Instead of typing: 1> c(Module, Options). types: 1> c(Module, [native|Options]).

• Alternatively, instead of typing: erlc … File types: erlc +native … File

HiPE: Invoking the compiler (seasoned user)

• Instead of typing: 1> c(Module, Options). types: 1> c(Module, [native,{hipe,

[’O3’,...]}|Options]).

• Learns about the full set of HiPE compiler options by typing:

1> hipe:help().

HiPE: Invoking the compiler (expert user)

• Selectively and dynamically compiles individual Erlang functions using:

1> c(M). … 42> hipe:c({M,F,A},

HiPE_Opts).

• Reports bugs to: hipe-bugs@csd.uu.se

HiPE: Invoking the compiler (HiPE hacker)

Sends bug fixes and compiler improvements to: hipe@csd.uu.se

HiPE version 1.0: Currently known limitations

• Native code will not be unloaded once loaded into the system (slight memory leak)

• Tracing and debugging support is non-existentHint: Use HiPE once your application is debugged!

• Floating point arithmetic slower than in BEAM• Exception information often slightly differs between

BEAM and HiPE • in particular, no tracelist is provided

• Fixed size (i.e., non-resizable) constant pool

HiPE version 1.0: Performance Expectations

0

2

4

6

8BEAM

HiPE/SPARC

HiPE/x86

HiPE version 1.0: Useful Hints

To get the most out of HiPE:

• Compile all time-critical parts of your Erlang application into native code– Separate hot from cold code

(perhaps by placing them into different modules)

• Avoid easily avoidable mode-switches

Part 3: Near Future

HiPE: Current Work

• Optimization of inter-process communication and efficient memory management for concurrent processes

• Formal specification of the Core Erlang language and its full integration within HiPE and Erlang/OTP

• New inliner for the BEAM compiler

• Experimental extensions of the Erlang language

Alternative Memory Architectures for Erlang

• Erlang/OTP has a process-centric view of memory management:

each process allocates each own memory area

Process 1

heap

stack

Process 2

heap

stack

Process n

heap

stack

...

...Global

ETS

Table

Interprocess communication is quite expensive

Alternative Memory Architectures for Erlang

• We (mainly Jesper Wilhelmsson) have designed and implemented an Erlang/OTP system where:– the heap is shared between all processes– but each process allocates its own stack

Process 1

stack

Process 2

stack

Process n

stack... Global

ETS

Table

Interprocess communication does not involve copying of messages

Global Heap

Unified Heap Architecture: Installation Guide

1. Get Erlang/OTP R82. Install by typing: ./configure --enable-unified-heap ./makeWarnings:

– Highly experimental at this point!– Does not work with HiPE

Request:Send us your favourite concurrent Erlang

application

Core Erlang:Formal Specification and Use in OSE

• A small and relatively clean subset of Erlang– Syntactic sugar is removed– Resembles other “core” FP languages

• An appropriate medium to:– base the compiler’s front-end (already part of R8)– apply high-level transformations such as:

• inlining (currently under development)• deforestation (prototype; results so far inconclusive)

– base work of static analysis or verification

A formal definition of Core Erlang is currently available as a tech report (Richard Carlsson et al)

Core Erlang Inliner: User’s Manual

• Instead of typing: 1> c(Module, Options). type: 1> c(Module, [inline|Options]).

Extensions of the Erlang language

• Parameterized Erlang modules– Design laid out; issue is efficient

implementation– Current work by Richard Carlsson & Mikael

Pettersson

• User-defined parametric datatypes (a la ML)

• Structured module system for Erlang