Transcript slides

Specialization
Run-time Code
Generation in the
context of Dynamic
Compilation
presented by Mujtaba Ali
Earlier in the Semester

“VCODE: A Retargetable, Extensible, Very Fast
Dynamic Code Generation System”
–
–

Generate “assembly” at run-time
Run-time Code Generation, but not in the realm of
Dynamic Compilation
“Fast, Effective Dynamic Compilation”
–
–
–
Our only foray into Run-time Specialization
Same authors as the first paper we will discuss
Precursor to DyC – we’ll call it the “DyC paper”
Example Compiler Optimizations
Constant Propagation
(everyone’s favorite)
Copy Propagation
Constant Folding
Loop Unrolling
Branch Elimination
Dead Store Elimination
(Dead-assignment Elimination)
Order Matters
foo (b) {
a = 3;
print(a);
a = b;
print(a);
}
foo (b) {
a = 3;
print(3);
a = b;
print(a);
}
Constant Propagation
foo (b) {
a = 3;
print(3);
a = b;
print(b);
}
Copy Propagation
DA Elimination
foo (b) {
print(3);
print(b);
}
Partial Evaluation
Complete Set of
Compiler Optimizations
Typically:


Partial Evaluation


Constant Propagation
Constant Folding
Loop Unrolling
Inlining
DyC Paper Refresher
Towards Automatic Construction
of Staged Compilers
Matthai Philipose,
Craig Chambers, and
Susan J. Eggers
Contributions

Eases the burden on the compiler writer
–

Compiler writer simply writes optimizations as usual
Supporting contributions:
–
–
Enhancements to Partial Evaluation
Enhancements to Dead-assignment Elimination
Compiler
Writer
Staged Compilation Framework

–
–

Compiler
User
SCF-ML
Language in which compiler optimizations
are written
First-order side-effect free subset of ML
Stager
–
–
Specializes optimizations written in SCF-ML
to separate compile-time and run-time
optimizations
Requires “approximate” information about
program input
High-level Overview of SCF
Concrete Example

Function to be optimized:
–
Say this function is the only contents of “mul_add.c”
Concrete Example (con’t)

At compile-time, compiler user does the following:
O1:
const prop
(SCF-ML)
I1
O2:
copy prop
(SCF-ML)
stager
stager
I2
AST of
mul_add.c
Info:
a is const
at run-time
O3:
dae
(SCF-ML)
O1’:
staged
const prop
(SCF-ML)
stub code
stager
I3
O2’:
staged
copy prop
(SCF-ML)
O3’:
staged
dae
(SCF-ML)
Concrete Example (con’t)

In ML:
Concrete Example (con’t)

What does the stub function do?
P1
AST of
mul_add.c
O1’:
staged
const prop
(SCF-ML)
P2
O2’:
staged
copy prop
(SCF-ML)
Run-time
const value
of a
–
All this is happening at run-time
P3
O3’:
staged
dae
(SCF-ML)
Popt
Concrete Example (con’t)

In ML:
Stager

Stager specializes the optimizations
–
–

Partial Evaluator optimizes writer’s optimization and
generates run-time/staged part of optimization
Performance is crucial for run-time part of
optimizations hence Dead-assignment Elimination
It appears as if there are two level of
specialization
–
One of the “alleged” levels is just the automatic
separation between compile-time and run-time
Alternate View of Specialization
Standard compiler:
P
O1
M
Specializing compiler:
P
O1’
O1’/2
run it
M/2
I


M’ is specialized to I
Note that O1’ will execute faster than O1
M’
Compare and Contrast with DyC

Compiler writer’s job is much easier
–
–
–

Stager automatically does all the work
Not specific to C
Not limited to optimizations that come with the
staged compiler
Compiler user still has the same API:
–
–
Input (run-time constraints)
User program
Benchmarking


Definitely easier than DyC
But is automatic staging as effective wrt:
–
–


Unstaged (standard) compiler
Hand-staged optimizations (as in DyC)
How fast are the combined static-time and runtime stages of the staged compiler?
Also, what about the size of the staged
compiler?
Benchmarking (con’t)

vs. Unstaged (standard) compiler
–

vs. Hand-staged optimizations
–
–

Hand-staged optimizations perform much better
Don’t know why, though
Unstaged compiler vs. staged compiler (both optimized)
–

Noticeably faster (it better be)
Who cares since the unstaged compiler does every thing at
static-time
Size of the automatically staged compiler
–
In practice, size grows linearly in the size of the input program
Drawbacks

Compiler user must understand specialization
–


Not clear if code generation during stub
function execution is language-specific
Implemented in ML
–
–

Doesn’t seem to be a way around this
Language interoperability (ML  C)
ML is garbage collected and bounds checked
Not robust enough for public consumption
Supporting Contributions

Partial Evaluators have been developed for
many domains
–
–
However, partially evaluating optimization programs
presents its own challenges
SCF’s Partial Evaluator:


Knows about map and set types
Uses identity tags for better equality analysis
Supporting Contributions(con’t)

Also, SCF employs a smarter representation of
abstract values
–
–
More amenable to specialization of code
implementing optimizations
Abstract values are used to loosely interpret code
during partial evaluation
Towards Automatic Specialization
of Java Programs
Ulrik Pagh Schultz,
Julia L. Lawall,
Charles Consel, and
Gilles Muller
Contributions

Specialization is useful in an OO setting

Supporting contributions:
–
–
–
OO languages lend themselves to specialization
“Quick and dirty” proof of concept
Performance benchmarks
General Idea

The more generic you get, the more your
performance suffers
Genericity
Sluggish
–
–
–
Specificity
Efficient
OO is generic by nature
Specialization is our savior
Specialization will convert a generic program to a
specific one
Automatic Specialization?

What is meant by “automatic”?
–

Specialized code is not generated by hand
What is different about the “specialization”
here?
–
–
–
Let’s try and mimic run-time specialization at
compile-time
Specialize for each member of the most likely inputs
If actual input at run-time is not in this set, fall back
on the generic, non-specialized code
Concrete Example

Consider this Java class for the power function:
Concrete Example (con’t)

Assume the exponent is always 3
–
We can specialize the calculate() method
–
JSCC would allow us to automatically specialize calculate()
Concrete Example (con’t)

With JSCC, we would define a specialization
class as such:
–
JSCC will generate the specialized code and…
Concrete Example (con’t)

… JSCC adds a method for switching between
specialized implementations:
Specialization in an OO Context

How does OO naturally lend itself to
specialization?
–
–
–
Data Encapsulation
Virtual Method Calls
In the case of Java, specializing the “VM”
Specializing Data Encapsulation

Traditional program specialization, but…
–
… save on sequence of pointer dereferences (due to OO)
Specializing Object Types

Information fed to specializer can help eliminate
dynamic dispatch overhead
Specializing the Virtual Machine

Such low-level optimizations are possible because
Java is eventually translated to C in this scheme
Two Levels of Specialization (Again)

Specialization of original Java program
–
JSCC


Example earlier
Specialization of translated C program
–
Tempo

Only the compile-time specialization capabilities of Tempo
are exploited
“Quick and Dirty” Prototype
Benchmarks
Image Processing Application
Benchmarks (con’t)

What are “Prospective” Benchmarks?
Sparc
–
Pentium
Benchmarks on manual back-translation to Java
Future Work

Run-time specialization
–

Automatic back translation to Java
–

Tempo already supports run-time specialization
Why? Portability.
Apply to situations involving extreme genericity
–
–
Software components
For example: JavaBeans
Drawbacks

Lack of run-time specialization
–
Forces tedious compile-time specialization

How will exploiting Tempo’s run-time
specialization affect backporting to Java?

Looks and feels like an ugly hack
The Present

JSPEC is the new incarnation
–
–

Run-time specialization
Java-to-Java translation
JSPEC paper is unpublished but available
Compare and Constrast with SCF

Unlike SCF, adding additional optimizations to
the specialization is non-trivial
–
–

You’re stuck with the provided partial evaluator
But partial evaluation covers a lot of ground
No run-time specialization
–
Most significant drawback
That’s All Folks!

Use Cyclone because it’s good for you!
–
Cyclone is the bestest language.