DOCSLIB.ORG

Automatic and Interactive Parallelization

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Automatic and Interactive Parallelization RICE UNIVERSITY Automatic and Interactive Parallelization by c Kathryn S M Kinley A Thesis Submitted in Partial Fulfillment of the Requirements for the Degree Do ctor of Philosophy Approved Thesis Committee Ken Kennedy Noah Harding Professor chair Computer Science Keith D Co op er Asso ciate Professor Computer Science Don H Johnson Professor Electrical and Computer Engineering Danny C Sorensen Professor Mathematical Sciences Houston Texas March Automatic and Interactive Parallelization c Kathryn S M Kinley Abstract The goal of this dissertation is to give programmers the ability to achieve high p er formance by fo cusing on developing parallel algorithms rather than on architecture sp ecic details The advantages of this approach also include program p ortability and legibility To achieve high p erformance we provide automatic compilation techniques that tailor parallel algorithms to sharedmemory multipro cessors with lo cal caches and a common bus In particular the compiler maps complete applications onto the sp ecics of a machine exploiting b oth parallelism and memory To optimize complete applications we develop novel general algorithms to trans form lo ops that contain arbitrary conditional control ow In addition we provide new interpro cedural transformations which enable optimization across pro cedure b ound aries These techniques provide the basis for a robust automatic parallelizing algo rithm that is applicable to complete programs The algorithm for automatic parallel co de generation takes into consideration the interaction of parallelism and data lo cality as well as the overhead of parallelism The algorithm is based on a simple cost mo del that accurately predicts cache line reuse from multiple accesses to the same memory lo cation and from consecutive accesses The optimizer uses this mo del to improve data lo cality It also uses the mo del to discover and intro duce eective parallelism that complements the b enets of data lo cality The optimizer further improves the eectiveness of parallelism by seeking to increase its granularity Parallelism is intro duced only when granularity is sucient to overcome its asso ciated costs The algorithm for parallel co de generation is shown to b e ecient and several of its comp onent algorithms are proven optimal The ecacy of the optimizer is illustrated with exp erimental results In most cases it is very eective and either achieves or improves the p erformance of handcrafted parallel programs When p erformance is not satisfactory we provide an interactive parallel programming to ol which combines compiler analysis and algorithms with human exp ertise Acknowledgments Ken Kennedy provided me with the three most imp ortant elements of supp ort in graduate scho ol intellectual p olitical and nancial In addition Ken Keith Co op er and Linda Torczon fostered a research atmosphere and working environment whose b enets are untold I would also like to recognize the other memb ers of my committee Keith Co op er Don Johnson and Danny Sorensen Keith has b een an endless source of encouragement and wisdom throughout my graduate career Don Johnson gave me my rst taste of research and ho oked me for life I am fortunate that many of my fellow graduate students and friends supp orted my research intellectually emotionally and with implementations I would esp ecially like to thank ChauWen Tseng Marina Kalem Mary Hall Paul Havlak Nat McIntosh Preston Briggs Ben Chase and the entire compiler group I am extremely gratefully to my entire family As always my parents were a constant source of love and encouragement To my husband Scotty Strahan I hop e I am the ro ck for you that you have b een for me A little Madness in the Spring Is wholesome even for the King Emily Dickinson Contents Abstract ii Acknowledgments iii List of Illustrations ix Intro duction Automatic parallelization Interactive parallelization Overview Technical Background Dep endence Analysis Interpro cedural dep endence analysis Augmented call graph Interactive Parallel Programming Intro duction Work Mo del Transformations Reordering transformations Dep endence breaking transformations Memory hierarchy transformations Miscellaneous transformations Transformation algorithms Lo op interchange Lo op skewing Lo op distribution Unroll and jam Incremental analysis after edits User and compiler interaction v Related work Discussion Lo op Transformations with Arbitrary Control Flow Motivation Lo op distribution Mechanics Restructuring Co de generation Other transformations Lo op skewing Lo op reversal Lo op p ermutation Strip mining Privatization Scalar expansion Lo op fusion Lo op p eeling Related work Discussion Interpro cedural Transformations Intro duction Technical background Augmented call graph Interpro cedural section analysis Supp ort for interpro cedural optimization The ParaScop e compilation system Recompilation analysis Interpro cedural transformation Lo op extraction Lo op emb edding Intrapro cedural transformations Lo op fusion Lo op p ermutation vi Exp erimental results Sp ec Ocean Related work Discussion Optimizing for Parallelism and Data Lo cality Intro duction Memory and language mo del Tradeos in optimization Optimizing data lo cality Sources of data reuse Simplifying assumptions Lo op cost Reference groups Lo op cost algorithm Imp erfectly nested lo ops Lo op p ermutation Memory order Permuting to achieve memory order Data lo cality exp erimental results Matrix multiply Stencil computations Jacobi and SOR Erlebacher Parallelism Performance estimation Intro ducing parallelism Strip mining Parallelization algorithm Optimization algorithm Exp erimental results Matrix multiply Dmxpy Related work Discussion vii An Automatic Parallel Co de Generator Intro duction Parallel co de generation Driving co de generation Pro cedure cloning Lo opbased optimization Partitioning for lo op distribution and lo op fusion Simple partition algorithm Merging the solutions Discussion Lo op fusion Lo op distribution Integrating interpro cedural transformations Selecting the appropriate interpro cedural transformation Extensions to pro cedure cloning .
Load more

Recommended publications
  • Redundancy Elimination Common Subexpression Elimination
    Redundancy Elimination Aim: Eliminate redundant operations in dynamic execution Why occur? Loop-invariant code: Ex: constant assignment in loop Same expression computed Ex: addressing Value numbering is an example Requires dataflow analysis Other optimizations: Constant subexpression elimination Loop-invariant code motion Partial redundancy elimination Common Subexpression Elimination Replace recomputation of expression by use of temp which holds value Ex. (s1) y := a + b Ex. (s1) temp := a + b (s1') y := temp (s2) z := a + b (s2) z := temp Illegal? How different from value numbering? Ex. (s1) read(i) (s2) j := i + 1 (s3) k := i i + 1, k+1 (s4) l := k + 1 no cse, same value number Why need temp? Local and Global ¡ Local CSE (BB) Ex. (s1) c := a + b (s1) t1 := a + b (s2) d := m&n (s1') c := t1 (s3) e := a + b (s2) d := m&n (s4) m := 5 (s5) if( m&n) ... (s3) e := t1 (s4) m := 5 (s5) if( m&n) ... 5 instr, 4 ops, 7 vars 6 instr, 3 ops, 8 vars Always better? Method: keep track of expressions computed in block whose operands have not changed value CSE Hash Table (+, a, b) (&,m, n) Global CSE example i := j i := j a := 4*i t := 4*i i := i + 1 i := i + 1 b := 4*i t := 4*i b := t c := 4*i c := t Assumes b is used later ¡ Global CSE An expression e is available at entry to B if on every path p from Entry to B, there is an evaluation of e at B' on p whose values are not redefined between B' and B.
    [Show full text]
  • Polyhedral Compilation As a Design Pattern for Compiler Construction
    Polyhedral Compilation as a Design Pattern for Compiler Construction PLISS, May 19-24, 2019 [email protected] Polyhedra? Example: Tiles Polyhedra? Example: Tiles How many of you read “Design Pattern”? → Tiles Everywhere 1. Hardware Example: Google Cloud TPU Architectural Scalability With Tiling Tiles Everywhere 1. Hardware Google Edge TPU Edge computing zoo Tiles Everywhere 1. Hardware 2. Data Layout Example: XLA compiler, Tiled data layout Repeated/Hierarchical Tiling e.g., BF16 (bfloat16) on Cloud TPU (should be 8x128 then 2x1) Tiles Everywhere Tiling in Halide 1. Hardware 2. Data Layout Tiled schedule: strip-mine (a.k.a. split) 3. Control Flow permute (a.k.a. reorder) 4. Data Flow 5. Data Parallelism Vectorized schedule: strip-mine Example: Halide for image processing pipelines vectorize inner loop https://halide-lang.org Meta-programming API and domain-specific language (DSL) for loop transformations, numerical computing kernels Non-divisible bounds/extent: strip-mine shift left/up redundant computation (also forward substitute/inline operand) Tiles Everywhere TVM example: scan cell (RNN) m = tvm.var("m") n = tvm.var("n") 1. Hardware X = tvm.placeholder((m,n), name="X") s_state = tvm.placeholder((m,n)) 2. Data Layout s_init = tvm.compute((1,n), lambda _,i: X[0,i]) s_do = tvm.compute((m,n), lambda t,i: s_state[t-1,i] + X[t,i]) 3. Control Flow s_scan = tvm.scan(s_init, s_do, s_state, inputs=[X]) s = tvm.create_schedule(s_scan.op) 4. Data Flow // Schedule to run the scan cell on a CUDA device block_x = tvm.thread_axis("blockIdx.x") 5. Data Parallelism thread_x = tvm.thread_axis("threadIdx.x") xo,xi = s[s_init].split(s_init.op.axis[1], factor=num_thread) s[s_init].bind(xo, block_x) Example: Halide for image processing pipelines s[s_init].bind(xi, thread_x) xo,xi = s[s_do].split(s_do.op.axis[1], factor=num_thread) https://halide-lang.org s[s_do].bind(xo, block_x) s[s_do].bind(xi, thread_x) print(tvm.lower(s, [X, s_scan], simple_mode=True)) And also TVM for neural networks https://tvm.ai Tiling and Beyond 1.
    [Show full text]
  • A Tiling Perspective for Register Optimization Fabrice Rastello, Sadayappan Ponnuswany, Duco Van Amstel
    A Tiling Perspective for Register Optimization Fabrice Rastello, Sadayappan Ponnuswany, Duco van Amstel To cite this version: Fabrice Rastello, Sadayappan Ponnuswany, Duco van Amstel. A Tiling Perspective for Register Op- timization. [Research Report] RR-8541, Inria. 2014, pp.24. hal-00998915 HAL Id: hal-00998915 https://hal.inria.fr/hal-00998915 Submitted on 3 Jun 2014 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. A Tiling Perspective for Register Optimization Łukasz Domagała, Fabrice Rastello, Sadayappan Ponnuswany, Duco van Amstel RESEARCH REPORT N° 8541 May 2014 Project-Teams GCG ISSN 0249-6399 ISRN INRIA/RR--8541--FR+ENG A Tiling Perspective for Register Optimization Lukasz Domaga la∗, Fabrice Rastello†, Sadayappan Ponnuswany‡, Duco van Amstel§ Project-Teams GCG Research Report n° 8541 — May 2014 — 21 pages Abstract: Register allocation is a much studied problem. A particularly important context for optimizing register allocation is within loops, since a significant fraction of the execution time of programs is often inside loop code. A variety of algorithms have been proposed in the past for register allocation, but the complexity of the problem has resulted in a decoupling of several important aspects, including loop unrolling, register promotion, and instruction reordering.
    [Show full text]
  • Loop Transformations and Parallelization
    Loop transformations and parallelization Claude Tadonki LAL/CNRS/IN2P3 University of Paris-Sud [email protected] December 2010 Claude Tadonki Loop transformations and parallelization C. Tadonki – Loop transformations Introduction Most of the time, the most time consuming part of a program is on loops. Thus, loops optimization is critical in high performance computing. Depending on the target architecture, the goal of loops transformations are: improve data reuse and data locality efficient use of memory hierarchy reducing overheads associated with executing loops instructions pipeline maximize parallelism Loop transformations can be performed at different levels by the programmer, the compiler, or specialized tools. At high level, some well known transformations are commonly considered: loop interchange loop (node) splitting loop unswitching loop reversal loop fusion loop inversion loop skewing loop fission loop vectorization loop blocking loop unrolling loop parallelization Claude Tadonki Loop transformations and parallelization C. Tadonki – Loop transformations Dependence analysis Extract and analyze the dependencies of a computation from its polyhedral model is a fundamental step toward loop optimization or scheduling. Definition For a given variable V and given indexes I1, I2, if the computation of X(I1) requires the value of X(I2), then I1 ¡ I2 is called a dependence vector for variable V . Drawing all the dependence vectors within the computation polytope yields the so-called dependencies diagram. Example The dependence vectors are (1; 0); (0; 1); (¡1; 1). Claude Tadonki Loop transformations and parallelization C. Tadonki – Loop transformations Scheduling Definition The computation on the entire domain of a given loop can be performed following any valid schedule.A timing function tV for variable V yields a valid schedule if and only if t(x) > t(x ¡ d); 8d 2 DV ; (1) where DV is the set of all dependence vectors for variable V .
    [Show full text]
  • Incremental Computation of Static Single Assignment Form
    Incremental Computation of Static Single Assignment Form Jong-Deok Choi 1 and Vivek Sarkar 2 and Edith Schonberg 3 IBM T.J. Watson Research Center, P. O. Box 704, Yorktown Heights, NY 10598 ([email protected]) 2 IBM Software Solutions Division, 555 Bailey Avenue, San Jose, CA 95141 ([email protected]) IBM T.J. Watson Research Center, P. O. Box 704, Yorktown Heights, NY 10598 ([email protected]) Abstract. Static single assignment (SSA) form is an intermediate rep- resentation that is well suited for solving many data flow optimization problems. However, since the standard algorithm for building SSA form is exhaustive, maintaining correct SSA form throughout a multi-pass com- pilation process can be expensive. In this paper, we present incremental algorithms for restoring correct SSA form after program transformations. First, we specify incremental SSA algorithms for insertion and deletion of a use/definition of a variable, and for a large class of updates on intervals. We characterize several cases for which the cost of these algorithms will be proportional to the size of the transformed region and hence poten- tially much smaller than the cost of the exhaustive algorithm. Secondly, we specify customized SSA-update algorithms for a set of common loop transformations. These algorithms are highly efficient: the cost depends at worst on the size of the transformed code, and in many cases the cost is independent of the loop body size and depends only on the number of loops. 1 Introduction Static single assignment (SSA) [8, 6] form is a compact intermediate program representation that is well suited to a large number of compiler optimization algorithms, including constant propagation [19], global value numbering [3], and program equivalence detection [21].
    [Show full text]
  • Compiler-Based Code-Improvement Techniques
    Compiler-Based Code-Improvement Techniques KEITH D. COOPER, KATHRYN S. MCKINLEY, and LINDA TORCZON Since the earliest days of compilation, code quality has been recognized as an important problem [18]. A rich literature has developed around the issue of improving code quality. This paper surveys one part of that literature: code transformations intended to improve the running time of programs on uniprocessor machines. This paper emphasizes transformations intended to improve code quality rather than analysis methods. We describe analytical techniques and specific data-flow problems to the extent that they are necessary to understand the transformations. Other papers provide excellent summaries of the various sub-fields of program analysis. The paper is structured around a simple taxonomy that classifies transformations based on how they change the code. The taxonomy is populated with example transformations drawn from the literature. Each transformation is described at a depth that facilitates broad understanding; detailed references are provided for deeper study of individual transformations. The taxonomy provides the reader with a framework for thinking about code-improving transformations. It also serves as an organizing principle for the paper. Copyright 1998, all rights reserved. You may copy this article for your personal use in Comp 512. Further reproduction or distribution requires written permission from the authors. 1INTRODUCTION This paper presents an overview of compiler-based methods for improving the run-time behavior of programs — often mislabeled code optimization. These techniques have a long history in the literature. For example, Backus makes it quite clear that code quality was a major concern to the implementors of the first Fortran compilers [18].
    [Show full text]
  • Introduction to Machine-Independent Optimizations - 1
    Introduction to Machine-Independent Optimizations - 1 Y.N. Srikant Department of Computer Science and Automation Indian Institute of Science Bangalore 560 012 NPTEL Course on Principles of Compiler Design Y.N. Srikant Introduction to Optimizations Outline of the Lecture What is code optimization? Illustrations of code optimizations Examples of data-flow analysis Fundamentals of control-flow analysis Algorithms for two machine-independent optimizations SSA form and optimizations Y.N. Srikant Introduction to Optimizations Machine-independent Code Optimization Intermediate code generation process introduces many inefficiencies Extra copies of variables, using variables instead of constants, repeated evaluation of expressions, etc. Code optimization removes such inefficiencies and improves code Improvement may be time, space, or power consumption It changes the structure of programs, sometimes of beyond recognition Inlines functions, unrolls loops, eliminates some programmer-defined variables, etc. Code optimization consists of a bunch of heuristics and percentage of improvement depends on programs (may be zero also) Y.N. Srikant Introduction to Optimizations Examples of Machine-Independant Optimizations Global common sub-expression elimination Copy propagation Constant propagation and constant folding Loop invariant code motion Induction variable elimination and strength reduction Partial redundancy elimination Loop unrolling Function inlining Tail recursion removal Vectorization and Concurrentization Loop interchange, and loop blocking Y.N. Srikant Introduction to Optimizations Bubble Sort Running Example Y.N. Srikant Introduction to Optimizations Control Flow Graph of Bubble Sort Y.N. Srikant Introduction to Optimizations GCSE Conceptual Example Y.N. Srikant Introduction to Optimizations GCSE on Running Example - 1 Y.N. Srikant Introduction to Optimizations GCSE on Running Example - 2 Y.N. Srikant Introduction to Optimizations Copy Propagation on Running Example Y.N.
    [Show full text]
  • Code Improvement:Thephasesof Compilation Devoted to Generating Good Code
    Code17 Improvement In Chapter 15 we discussed the generation, assembly, and linking of target code in the middle and back end of a compiler. The techniques we presented led to correct but highly suboptimal code: there were many redundant computations, and inefficient use of the registers, multiple functional units, and cache of a mod- ern microprocessor. This chapter takes a look at code improvement:thephasesof compilation devoted to generating good code. For the most part we will interpret “good” to mean fast. In a few cases we will also consider program transforma- tions that decrease memory requirements. On occasion a real compiler may try to minimize power consumption, dollar cost of execution under a particular ac- counting system, or demand for some other resource; we will not consider these issues here. There are several possible levels of “aggressiveness” in code improvement. In a very simple compiler, or in a “nonoptimizing” run of a more sophisticated com- piler, we can use a peephole optimizer to peruse already-generated target code for obviously suboptimal sequences of adjacent instructions. At a slightly higher level, typical of the baseline behavior of production-quality compilers, we can generate near-optimal code for basic blocks. As described in Chapter 15, a basic block is a maximal-length sequence of instructions that will always execute in its entirety (assuming it executes at all). In the absence of delayed branches, each ba- sic block in assembly language or machine code begins with the target of a branch or with the instruction after a conditional branch, and ends with a branch or with the instruction before the target of a branch.
    [Show full text]
  • A Performance-Based Approach to Automatic Parallelization
    A Performance-based Approach to Automatic Parallelization Lecture Notes for ECE 663 Advanced Optimizing Compilers Draft Rudolf Eigenmann School of Electrical and Computer Engineering Purdue University c Rudolf Eigenmann 2 Chapter 1 Motivation and Introduction Slide 2: Optimizing Compilers are in the Center of the (Software) Universe Compilers are the translators between “human and machine”. This interface is one of the more challenging and important issues in all of computer science. Through programming languages, software engineers tell machines what do to. Today’s programming languages are still rather cryptic, borrowing a few words from the English language vocabulary. Perhaps, tomorrow’s languages will be more like natural languages. They may offer higher-level ex- pressions, allowing problems to be specified, rather than coding the detailed problem solution algorithms. Translating these languages onto modern architectures with their low-level ma- chine code languages is already a challenge. As every generation of computer architectures tends to grow in complexity, translating future programming languages onto future machines will be an even grander challenge. In performing such translation, optimization is important. Basic translation is usually inefficient. Performance almost always matters. The are other optimization criteria, as well. Energy efficiency is of increasing importance and is critical for devices that are battery operated. Small code and memory size can be of great value for embedded system. In this course, we will focus mainly on performance. Optimizations for parallel machines are the focus of this course. Today’s CPUs contain multiple cores. They are one of the most important classes of current parallel machines. This was not always so.
    [Show full text]
  • Manual Micro-Optimizations in C++ an Investigation of Four Micro-Optimizations and Their Usefulness
    Technology and Society Computer Science and Media Technology Bachelor thesis 15 credits, first cycle Manual micro-optimizations in C++ An investigation of four micro-optimizations and their usefulness Manuella mikrooptimeringar i C++ En undersökning av fyra mikrooptimeringar och deras nytta Viktor Ekström Degree: Bachelor 180hp Supervisor: Olle Lindeberg Main field: Computer Science Examiner: Farid Naisan Programme: Application Developer Final seminar: 4/6 2019 Abstract Optimization is essential for utilizing the full potential of the computer. There are several different approaches to optimization, including so-called micro-optimizations. Micro-optimizations are local adjustments that do not change an algorithm. This study investigates four micro-optimizations: loop interchange, loop unrolling, cache loop end value, and iterator incrementation, to see when they provide performance benefit in C++. This is investigated through an experiment, where the running time of test cases with and without micro-optimization is measured and then compared between cases. Measurements are made on two compilers. Results show several circumstances where micro-optimizations provide benefit. However, value can vary greatly depending on the compiler even when using the same code. A micro-optimization providing benefit with one compiler may be detrimental to performance with another compiler. This shows that understanding the compiler, and measuring performance, remains important. 1 2 Sammanfattning Optimering är oumbärligt för att utnyttja datorns fulla potential. Det finns flera olika infallsvinklar till optimering, däribland så kallade mikrooptimeringar. Mikrooptimeringar är lokala ändringar som inte ändrar på någon algoritm. Denna studie undersöker fyra mikrooptimeringar: loop interchange, loop unrolling, cache loop end value, och iterator-inkrementering, för att se när de bidrar med prestandaförbättring i C++.
    [Show full text]
  • CSE 501: Course Outline Implementation of Programming Languages
    CSE 501: Course outline Implementation of Programming Languages Models of compilation/analysis Main focus: program analysis and transformation ¥ how to represent programs? Standard optimizing transformations ¥ how to analyze programs? what to analyze? ¥ how to transform programs? what transformations to apply? Basic representations and analyses Study imperative, functional, and object-oriented languages Fancier representations and analyses Official prerequisites: Interprocedural representations, analyses, and transformations ¥ CSE 401 or equivalent ¥ for imperative, functional, and OO languages ¥ CSE 505 or equivalent Run-time system issues ¥ garbage collection Reading: Appel’s “Modern Compiler Implementation” ¥ compiling dynamic dispatch, first-class functions, ... + ~20 papers from literature ¥ “Compilers: Principles, Techniques, & Tools”, Dynamic (JIT) compilation a.k.a. the Dragon Book, as a reference Other program analysis frameworks and tools Coursework: ¥ model checking, constraints, best-effort "bug finders" ¥ periodic homework assignments ¥ major course project ¥ midterm, final Craig Chambers 1 CSE 501 Craig Chambers 2 CSE 501 Why study compilers? Goals for language implementation Meeting area of programming languages, architectures Correctness ¥ capabilities of compilers greatly influence design of these others Efficiency ¥ of: time, data space, code space Program representation, analysis, and transformation ¥ at: compile-time, run-time is widely useful beyond pure compilation ¥ software engineering tools Support expressive,
    [Show full text]
  • Cray Fortran 90 Optimization
    Cray Fortran 90 Optimization Sylvia Crain, Cray Research, A Silicon Graphics Company, 655F Lone Oak Drive, Eagan, Minnesota 55121 ABSTRACT: This paper provides a discussion of the optimization technology used in the Cray CF90 compiler. It also outlines some specific optimizations important to the compiler and provides numbers on Cray CF90 performance relative to the Cray CF77 compiler. Lastly, it suggests strategies for transitioning from Fortran 77 to Fortran 90. 1 The Optimizer Technology these intrinsics allows for other optimizations to take place on the loop bodies and also eliminates the overhead of the subrou- The optimizer used by Cray’s Fortran 90 compiler is called tine calls. Even with no other optimization, removal of the PDGCS which stands for Program Dependence Graph subroutine call overhead can provide a good win. Compiling System. This optimizer provides state of the art tech- nology to support the Fortran 90 language. The premise of the Loop fusion is an important optimization to aid array syntax. optimizer is hierarchical memory optimization which structures After an array intrinsics is expand into a series of loops, the optimizations to reduce memory traffic, take advantage of loops can then be fused or combined into a single larger loop cache, and reduce the number of loads and stores required where other potential optimizations can then be performed. within a program. The program dependence graph (PDG) is at Loop Interchange is swap inner and outer loops. This allows the heart of the optimizer and is integral to the design. Modifi- for moving dependencies out of the loop thus providing greater cations to the PDG build on top of one another.
    [Show full text]