Optimizing ML with run-time code generation

Lee, Peter; Leone, Mark P.

doi:10.1145/231379.231407

Cited by 116 publications

(27 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…More generally, run-time code generation has been investigated e.g. in [14] and [7]. Some of the results presented in these papers is likely to be of interest to future low-level optimizations by Psyco, although all frameworks considered so far (VCODE [7], GNU Lightning 22 , LLVM [13]) focus on whole-function code generation and thus cannot be reused directly.…”

Section: Related Workmentioning

confidence: 99%

Representation-based just-in-time specialization and the psyco prototype for python

Rigo

2004

Proceedings of the 2004 ACM SIGPLAN Symposium on Partial Evaluation and Semantics-Based Program Manipulation

View full text Add to dashboard Cite

A powerful application of specialization is to remove interpretative overhead: a language can be implemented with an interpreter, whose performance is then improved by specializing it for a given program source. This approach is only moderately successful with very high level languages, where the operation of each single step can be highly dependent on run-time data and context. In the present paper, the Psyco prototype for the Python language is presented. It introduces two novel techniques. The first is just-in-time specialization, or specialization by need, which introduces the "unlifting" ability for a value to be promoted from runtime to compile-time during specialization -the inverse of the lift operator of partial evaluation. Its presence gives an unusual and powerful perspective on the specialization process. The second technique is representations, a theory of data-oriented specialization generalizing the traditional specialization domains (i.e. the compile-time/run-time dichotomy).

show abstract

Section: Related Workmentioning

confidence: 99%

Representation-based just-in-time specialization and the psyco prototype for python

Rigo

2004

Proceedings of the 2004 ACM SIGPLAN Symposium on Partial Evaluation and Semantics-Based Program Manipulation

View full text Add to dashboard Cite

show abstract

“…For instance, there are many examples (kernel implementation (Massalin, 1992), graphics (Pike et al, 1985), interactive media (Draves, 1998), method dispatch in object-oriented languages (Deutsch & Schiffman, 1984;Chambers & Ungar, 1989), etc.) where run-time code generation can be employed to reap significant gain in run-time performance (Leone & Lee, 1996). To illustrate this point, we define a function evalPoly as follows in Scheme for evaluating a polynomial p at a given point x.…”

Section: Introductionmentioning

confidence: 99%

Meta-programming through typeful code representation

Chen

2005

J. Funct. Prog.

View full text Add to dashboard Cite

Link to this article: http://journals.cambridge.org/abstract_S0956796805005617How to cite this article: CHIYAN CHEN and HONGWEI XI (2005). Meta-programming through typeful code representation. JFP 15 (6): 797-835, 2005. AbstractBy allowing the programmer to write code that can generate code at run-time, metaprogramming offers a powerful approach to program construction. For instance, metaprogramming can often be employed to enhance program efficiency and facilitate the construction of generic programs. However, meta-programming, especially in an untyped setting, is notoriously error-prone. In this paper, we aim at making meta-programming less error-prone by providing a type system to facilitate the construction of correct meta-programs. We first introduce some code constructors for constructing typeful code representation in which program variables are represented in terms of deBruijn indexes, and then formally demonstrate how such typeful code representation can be used to support meta-programming. With our approach, a particular interesting feature is that code becomes first-class values, which can be inspected as well as executed at run-time. The main contribution of the paper lies in the recognition and then the formalization of a novel approach to typed meta-programming that is practical, general and flexible.

show abstract

“…For example, profile data can be used to measure how many times a procedure is called versus how many times it is created, and this ratio could be used to guide lambda lifting [12]. Combined with an estimate of the cost of generating code for the procedure, this ratio could also help determine when run-time code generation [22] would be profitable. Our system can also be extended to recompile (specialize) a closure based on the run-time values of its free variables.…”

Section: Discussionmentioning

confidence: 99%

“…Several recent research projects have focused on lightweight run-time code generation [8,15,22,20,25], with code generation costs on the order of five to 100 instructions executed for each instruction generated. Although the overhead inherent in our model is greater, the potential benefits are greater as well.…”

Section: Related Workmentioning

confidence: 99%

“…On the other hand, it prevents the compiler from exploiting properties of the program, its input, and its execution environment that can be determined only at run time. Recent research has shown that dynamic compilation can dramatically improve the performance of a wide range of applications including network packet demultiplexing, sparse matrix computations, pattern matching, and mobile code [10,8,13,16,22,23]. Dynamic optimization works when a program is staged in such a way that the cost of dynamic recompilation can be amortized over many runs of the optimized code [21].…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

An infrastructure for profile-driven dynamic recompilation

Burger

Dybvig

Proceedings of the 1998 International Conference on Computer Languages (Cat. No.98CB36225)

View full text Add to dashboard Cite

Dynamic optimization of computer programs can dramatically improve their performance on a variety of applications. This paper presents an efficient infrastructure for dynamic recompilation that can support a wide range of dynamic optimizations including profile-driven optimizations. The infrastructure allows any section of code to be optimized and regenerated on-the-fly, even code for currently active procedures. The infrastructure incorporates a low-overhead edge-count profiling strategy that supports first-class continuations and reinstrumentation of active procedures. Profiling instrumentation can be added and removed dynamically, and the data can be displayed graphically in terms of the original source to provide useful feedback to the programmer.

show abstract

Optimizing ML with run-time code generation

Cited by 116 publications

References 22 publications

Representation-based just-in-time specialization and the psyco prototype for python

Representation-based just-in-time specialization and the psyco prototype for python

Meta-programming through typeful code representation

An infrastructure for profile-driven dynamic recompilation

Contact Info

Product

Resources

About