Download Demo Summary - Ptolemy Project

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
The Ptolemy C Code Generator
Ankush Varma
Shuvra Bhattacharyya
University of Maryland, College Park
5th Biennial Ptolemy Miniconference
Berkeley, CA, May 9, 2003
Why C?
• Fast
– fewer runtime checks.
– no JVM overhead.
• Small
– full library not needed.
• Retargetability
– ANSI C source code is portable.
– No JVM needed.
– Can run on embedded systems with no JVM and no OS.
• More optimized
– C compilers have highly configurable, well-studied optimizations.
C Code Generation 2
Cracking the Code
• Problem: Java Bytecode is stack-based. C uses local variables.
– Use Soot to unroll the stack into local variables while transforming it into
intermediate representation.
• Problem: Java is Object-Oriented. C is not.
– Use hashing to give each method/field a different name.
– Use structures for Objects/Classes.
– Dare to use indirect function pointers (scary, but it works).
• Problem: Java relies heavily on exceptions. C has no support for
exceptions.
– Build support for exceptions in C using setjmp-longjmp.
• Problem: Java has automatic garbage collection, C does not
– Use a C-based GC as a plugin.
• Problem: Java relies on JVM or OS for some functionality
– Build a runtime library to provide the required functionality.
C Code Generation 3
The Dark Art of Java-to-C Compilation
Java Code
Java Compiler
Portable Bytecode (.class file)
Soot
Jimple Intermediate Representation
Optimizations
Optimized Jimple
Code Generator
Libraries, C code and headers
Automatically generated makefile
C Compiler
Executable Application
C Code Generation 4
Trimming the Tree
• Set up MethodCallGraph
• Start with main class and all its methods as required.
• Start worklist-based algorithm
– If you see a class:
• look at its initialization methods.
• look at its superclasses.
– If you see a field:
• look at the class declaring it.
• look at the class of its type.
– If you see a method:
• look at all fields it references.
• look at all methods it calls.
• look at its class.
• look at the classes of exceptions it throws/catches.
C Code Generation 5
Strategies for Speeding Up Compilation
• Create single untrimmed shared static library
– too simplistic.
– generating library takes too long.
– code bloat.
• Generate separate library trimmed for each application
– still takes a long time.
– mostly spending time figuring out polymorphic method calls
within library classes.
C Code Generation 6
Strategies for Speeding Up Compilation
(continued)
• Divide code into application and library classes:
– cache library method targets in a disk file.
– compile-time analysis is simplified.
– extensible to any library classes (Ptolemy?)
User Methods
Library Methods
C Code Generation 7
Interpreted Java
java JIT
gcj
Method
Float
String
Logic
Loop
80000
70000
60000
50000
40000
30000
20000
10000
0
Sieve
Faster …
Java-to-C
C Code Generation 8
… and much, much smaller.
2000
1500
gcj
Java-to-C
1000
500
Te
st
Pr
in
t
et
s
hS
H
as
ist
s
ed
L
e
Li
nk
af
fe
in
C
ac
k
0
Li
np
Code Size (KB)
2500
C Code Generation 9
Roadmap
•
•
•
•
•
•
•
•
•
•
•
Automatic makefile generation.
Generated fully ANSI-compliant C.
Inheritance.
Exceptions.
Multidimensional arrays.
Interfaces.
Automatic Garbage Collection.
Plugging in to ptolemy.
Further generic optimizations.
Further ptolemy-specific optimizations.
Run java code on a DSP.
C Code Generation 10
C Code Generation 11