COMPILER TECHNIQUES FOR MATLAB PROGRAMS ... - CiteSeerX

More documents

Recommendations

Info

Chapter 2 RELATED WORK There are many examples of typeless programming languages, including APL, MATLAB, ML [GMW79], Haskell [HWA + 88], Lisp [MAE + 65], and Smalltalk [GR83]. These languages can be divided in two classes: the statically typed languages (e.g., ML and Haskell) and the dynamically typed languages (e.g., Lisp, Smalltalk, APL, and MATLAB). Statically typed languages include type constraints as part of their language denition. Hence, type inference is necessary to ensure that these type constraints are satised. These type constraints allow compilers to deduce most of the type information, thereby making run-time type-checking unnecessary. For cases in which the type-checker cannot deduce the type of a valid expression, the user must provide the type information. Normally in these languages, the type inference must be accurate enough to satisfy the language's type constraints. Thus, the type-checker may sometimes nd a more generic type assignment than expected. A classical example is the language ML, a meta-language for theorem proving, where the main motivation for strict type-checking is to ensure that every computed value of type theorem is indeed a theorem [GMW79]. To facilitate type inferencing, several restrictions are imposed to the language. For example, ML does not allow expressions with mixed types. Thus, a function denition fun successor(n) = n+1 will be considered a function that takes an integer as argument (n) and returns another 7
integer. The reason for this is that the operator + can only be applied to two operands of the same type. Therefore, n must have intrinsic type integer because it is being added to the integer 1. Also, to allow thetype to be inferred statically, ML sometimes requires explicit type information to resolve overloading. in these cases, the lack ofthistype information will result in run-time type errors. Dynamically typed languages do not havetype constraints hence, run-time type-checking is a necessary part of the system. Due to the lack oftype constraints, compilers cannot deduce all type information. Therefore, for the cases where the type-checker cannot deduce the type of a valid expression, run-time type checks are required. For dynamically typed languages, type inference is primarily used for optimization purposes. Hence, the inferred type information should be as precise as possible in order to avoid the overhead of run-time type-checking. We now rst describe relevant type inference techniques, followed by some approaches for the compilation of APL, MATLAB, and MATLAB-like languages. These approaches range from research projects to commercial products. 2.1 Relevant Type Inference Approaches Type inference algorithms have been presented in the literature for both classes of languages described above. However, most of the work concentrates on dening a type system and inference rules for a particular language. One of our main objectives in this thesis is to evaluate the eectiveness of the inference mechanisms for detecting intrinsic type, rank, and shape on MATLAB programs. We make use in this work of data-ow analysis as described by Aho, Sethi, and Ullman in [ASU85], and type inference techniques developed for SETL [Sch75] and APL [Bud88, Chi86]. We extend these techniques where necessary to deal with peculiarities of the MAT- LAB language and to improve accuracy and performance. Examples of these techniques are: a structural inference mechanism, a symbolic dimension propagation analysis, and a value propagation analysis. 8
Page 1 and 2: COMPILER TECHNIQUES FOR MATLAB PROG
Page 3 and 4: COMPILER TECHNIQUES FOR MATLAB PROG
Page 5 and 6: Acknowledgments Iamindebtedwiththem
Page 7 and 8: Contents Chapter 1 INTRODUCTION :::
Page 9 and 10: 7 EXPERIMENTAL RESULTS ::::::::::::
Page 11 and 12: List of Figures 3.1 Phases of the M
Page 13 and 14: 7.7 CG speedups on the SGI Power Ch
Page 15 and 16: 1.1 High-Level Approach for Softwar
Page 17 and 18: executing their compiled versions.
Page 19: structure: (e.g., upper triangular,
Page 23 and 24: the high-level semantics of the APL
Page 25 and 26: Chapter 3 OVERALL STRATEGY 3.1 Rest
Page 27 and 28: Matlab Program lexical analyzer syn
Page 29 and 30: according to the grammar rules, pro
Page 31 and 32: S1: load %(V) S1: load %(V 1 ) S2:
Page 33 and 34: epresenting variable V 1 will conta
Page 35 and 36: Structural Inference One of the gre
Page 37 and 38: eciprocal estimator, which requires
Page 39 and 40: if (A T .eq. T COMPLEX) then if (B
Page 41 and 42: S1: temp = 1 S2: X = function(temp)
Page 43 and 44: if (a D1 .eq. 1) then if (a D2 .eq.
Page 45 and 46: Chapter 4 INTERNAL REPRESENTATION 4
Page 47 and 48: Algorithm: Dierentiation Between Va
Page 49 and 50: entry list variable’s instances V
Page 51 and 52: S0: load %(n) S0: load %(n 1 ) P1:
Page 53 and 54: Chapter 5 THE STATIC INFERENCE MECH
Page 55 and 56: Type of rst Type of second paramete
Page 57 and 58: Unknown Real Complex Integer Logica
Page 59 and 60: A = zeros(5,1) for ... ... A(1:5) =
Page 61 and 62: Assignment to a constant: Attribute
Page 63 and 64: A B B A scalar vector(p,1) vector(
Page 65 and 66: Unknown . . . κ i κ j κ l κ m N
Page 67 and 68: S1: B = rand(2) B 1 = rand(2) S2: A
Page 69 and 70: [L, U, P] = lu(A) y = Ln (P b) x =
Page 71 and 72:
MATLAB, and x would be computed in
Page 73 and 74:
if (a D1 .eq. 1) then if (c D1 .ne.
Page 75 and 76:
S1: A = ones(5) S2: B = function(A)
Page 77 and 78:
6.2 Dynamic Shape Inference The use
Page 79 and 80:
if (k .gt. P D1 .or. k .gt. P D2) t
Page 81 and 82:
A pointer (P s) to an AST node corr
Page 83 and 84:
x 1 will be already pointing to m 1
Page 85 and 86:
P nr also set to m 1 . Therefore, P
Page 87 and 88:
ALLOCATE statement can be placed ou
Page 89 and 90:
Test Programs: Problem size Lines S
Page 91 and 92:
FD: isanumeric approximation method
Page 93 and 94:
for j = 1:n ... r = sqrt(L(j,j) - s
Page 95 and 96:
1000 SGI Power Challenge 400 Speedu
Page 97 and 98:
7.2.2 Comparison of Compiler Genera
Page 99 and 100:
10 83 91 17 9 8 SPARCstation 10 Spe
Page 101 and 102:
were selected: AQ, due to its real
Page 103 and 104:
inference phases AQ CG 3D Di FD no
Page 105 and 106:
7 6 SGI Power Challenge CG 5 Second
Page 107 and 108:
program. Notice that although the M
Page 109 and 110:
uns, there were no preallocations o
Page 111 and 112:
Chapter 8 CONCLUSIONS AND FUTURE DI
Page 113 and 114:
of techniques for the generation of
Page 115 and 116:
[BE94] William Blume and Rudolf Eig
Page 117 and 118:
[GHR94] E. Gallopoulos, E. Houstis,
Page 119 and 120:
[Pac93] Pacic-Sierra Research Corpo
show all

COMPILER TECHNIQUES FOR MATLAB PROGRAMS ... - CiteSeerX

Create successful ePaper yourself

Delete template?

Save as template?