Tag der mündlichen Prüfung: 03. Juni 2004 Dekan / Dekanin: Prof. Dr. Bernhard Steffen Gutachter / Gutachterinnen: Prof. Dr. Francky Catthoor, Prof. Dr

Transcription

1 Source Code Optimization Techniques for Data Flow Dominated Embedded Software Dissertation zur Erlangung des Grades eines Doktors der Naturwissenschaften der Universität Dortmund am Fachbereich Informatik von Heiko Falk Dortmund 2004

2 Tag der mündlichen Prüfung: 03. Juni 2004 Dekan / Dekanin: Prof. Dr. Bernhard Steffen Gutachter / Gutachterinnen: Prof. Dr. Francky Catthoor, Prof. Dr. Peter Marwedel, Prof. Dr. Peter Padawitz

3

4

5 Contents v Contents List of Figures List of Tables ix xiii 1 Introduction Why Source Code Optimization? Abstraction Levels of Code Optimization Survey of the traditional Code Optimization Process Scopes for Code Optimization Target Application Domain Goals and Contributions Outline of the Thesis Author's Contribution to the Thesis Existing Code Optimization Techniques Specification Optimization Algorithm Selection Memory Hierarchy Exploitation Processor-independent Source Code Optimizations Processor-specific Source Code Optimizations Compiler Optimizations Loop Optimizations for High-Performance Computing Code Generation for embedded Processors Fundamental Concepts for Optimization and Evaluation Polyhedral Modeling Optimization using Genetic Algorithms Benchmarking Methodology Profiling of Pipeline and Cache Performance Compilation for Runtime and Code Size Measurement Estimation of Energy Dissipation

6 vi Contents 3.4 Summary Intermediate Representations Low-level Intermediate Representations GNU RTL Trimaran ELCOR IR Medium-level Intermediate Representations Sun IR IR-C / LANCE High-level Intermediate Representations SUIF IMPACT Selection of an IR for Source Code Optimization Summary Loop Nest Splitting Introduction Control Flow Overhead in Data Flow dominated Software Control Flow Overhead caused by Data Partitioning Splitting of Loop Nests for Control Flow Optimization Related Work Analysis and Optimization Techniques for Loop Nest Splitting Preliminaries Condition Satisfiability Condition Optimization Chromosomal Representation Fitness Function Polytope Generation Global Search Space Construction Global Search Space Exploration Chromosomal Representation Fitness Function Source Code Transformation Generation of the splitting If-Statement Loop Nest Duplication Extensions for Loops with non-constant Bounds Experimental Results Stand-alone Loop Nest Splitting Pipeline and Cache Performance Execution Times and Code Sizes Energy Consumption

7 Contents vii Combined Data Partitioning and Loop Nest Splitting for energy-efficient Scratchpad Utilization Execution Times and Code Sizes Energy Consumption Summary Advanced Code Hoisting A motivating Example Related Work Analysis Techniques for Advanced Code Hoisting Common Subexpression Identification Collection of equivalent Expressions Computation of Live Ranges of Expressions Determination of the outermost Loop for a CSE Computation of Execution Frequencies using Polytope Models Experimental Results Pipeline and Cache Performance Execution Times and Code Sizes Energy Consumption Summary Ring Buffer Replacement Motivation Optimization Steps Ring Buffer Scalarization Loop Unrolling for Ring Buffers Experimental Results Pipeline and Cache Performance Execution Times and Code Sizes Energy Consumption Summary Summary and Conclusions Summary and Contribution to Research Future Work A Experimental Comparison of SUIF and IR-C / LANCE 199

8 viii Contents B Benchmarking Data for Loop Nest Splitting 201 B.1 Values of performance-monitoring Counters B.1.1 Intel Pentium III B.1.2 Sun UltraSPARC III B.1.3 MIPS R B.2 Execution Times and Code Sizes B.3 Energy Consumption of an ARM7TDMI Core B.4 Combined Data Partitioning and Loop Nest Splitting B.4.1 Execution Times and Code Sizes B.4.2 Energy Consumption C Benchmarking Data for Advanced Code Hoisting 213 C.1 Values of performance-monitoring Counters C.1.1 Intel Pentium III C.1.2 Sun UltraSPARC III C.1.3 MIPS R C.2 Execution Times and Code Sizes C.3 Energy Consumption of an ARM7TDMI Core D Benchmarking Data for Ring Buffer Replacement 219 D.1 Values of performance-monitoring Counters D.1.1 Intel Pentium III D.1.2 Sun UltraSPARC III D.1.3 MIPS R D.2 Execution Times and Code Sizes D.3 Energy Consumption of an ARM7TDMI Core References 223 Index 235

9 List of Figures ix List of Figures 1.1 Examples of typical Embedded Systems Abstraction Levels and achievable Improvements of Code Optimizations Structure of Optimizing Compilers Implementation of a saturating Addition Abstraction Levels of Code Optimizations Overview of the DTSE Methodology Replacement of complex Code by atic6x Intrinsic A convex Polytope A non-convex Set of Points Schematic Overview of Genetic Algorithms Genetic Algorithm Encoding Crossover Mutation Overview of the Workflow employed in the encc Compiler Experimental Comparison of Runtimes based on SUIF and IR-C / LANCE Atypical Loop Nest Atypical Code Fragment before and after Data Partitioning Loop Nest after Splitting Conventional Loop Splitting Conventional Loop Unswitching Design Flow of Loop Nest Splitting Conditions as Polytopes Structure of the Fitness Function for Global Search Space Exploration Atypical Loop Nest with variable Loop Bounds Structure of the Fitness Function for Condition Optimization. 103

10 x List of Figures 5.11 A Loop Nest with variable Loop Bounds after Splitting Relative Pipeline and Cache Behavior for Intel Pentium III Relative Pipeline and Cache Behavior for Sun UltraSPARC III Relative Pipeline and Cache Behavior for MIPS R Relative Runtimes after Loop Nest Splitting Relative CodeSizes after Loop Nest Splitting Relative Energy Consumption after Loop Nest Splitting Relative Runtimes after Data Partitioning and Loop Nest Splitting Relative Code Sizes after Data Partitioning and Loop Nest Splitting Relative Energy Consumption after Data Partitioning Relative Energy Consumption after Data Partitioning and Loop Nest Splitting ACodeSequence of GSM Codebook Search The GSM Codebook Search after Code Hoisting without Control Flow Analysis The GSM Codebook Search after Advanced Code Hoisting Local Common Subexpression Elimination Global Common Subexpression Elimination Loop-invariant CodeMotion Parts of DTSE causing Control Flow and Addressing Overhead AFragment of the CAVITY Benchmark before and after DTSE Algorithm for Advanced Code Hoisting Algorithm to identify Common Subexpressions The SUIF Abstract Syntax Tree Bottom-up Syntax Tree Traversal Algorithm to determine the Parent of expr i;c and expr i;c Basic Algorithm for Syntax Tree Traversal Algorithm to compute Live Ranges Algorithm to determine the outermost Loop for a CSE General Structure of Control Flow surrounding an Expression expr Relative Pipeline and Cache Behavior for Intel Pentium III Relative Pipeline and Cache Behavior for Sun UltraSPARC III Relative Pipeline and Cache Behavior for MIPS R Relative Runtimes after Advanced Code Hoisting Runtime Comparison of Advanced Code Hoisting and Common Subexpression Elimination Relative CodeSizes after Advanced Code Hoisting

11 List of Figures xi 6.24 Relative Energy Consumption after Advanced Code Hoisting Structure of a Ring Buffer Realization of a Ring Buffer using an Array Original Structure of the CAVITY Application Structure of the CAVITY Application after Data Reuse Transformation Access Analysis of Line Buffers during In-Place Mapping Scalar Replacement of Array Elements Principal operating Mode of Ring Buffer Replacement Ring Buffer Scalarization Loop Unrolling for Ring Buffers Relative Pipeline and Cache Behavior for Intel Pentium III Relative Pipeline and Cache Behavior for Sun UltraSPARC III Relative Pipeline and Cache Behavior for MIPS R Relative Runtimes after Ring Buffer Replacement Relative CodeSizes after Ring Buffer Replacement Relative Energy Consumption after Ring Buffer Replacement Relative Runtimes after entire Optimization Sequence Possible Speed / Size Trade-Offs for Loop Nest Splitting B.1 Structure of the Intel Pentium III Architecture B.2 Structure of the Sun UltraSPARC III Architecture

12 xii

13 List of Tables xiii List of Tables 5.1 Characteristic Properties of Benchmarks A.1 Runtimes of original Benchmark Versions A.2 Runtimes of SUIF Benchmark Versions A.3 Runtimes of IR-C Benchmark Versions B.1 PMC Values for original Benchmark Versions B.2 PMC Values for Benchmarks after Loop Nest Splitting B.3 PIC Values for original Benchmark Versions B.4 PIC Values for Benchmarks after Loop Nest Splitting B.5 PC Values for original Benchmark Versions B.6 PC Values for Benchmarks after Loop Nest Splitting B.7 Compilers and Optimization Levels for Runtime and Code Size Measurements B.8 Runtimes of original Benchmark Versions B.9 Runtimes of Benchmarks after Loop Nest Splitting B.10 Code Sizes of original Benchmark Versions B.11 Code Sizes of Benchmarks after Loop Nest Splitting B.12 Energy Consumption of original Benchmark Versions B.13 Energy Consumption of Benchmarks after Loop Nest Splitting 209 B.14 Runtimes of Benchmarks after Data Partitioning and Loop Nest Splitting B.15 Code Sizes of Benchmarks after Data Partitioning and Loop Nest Splitting B.16 Energy Consumption of original Benchmark Versions B.17 Energy Consumption of Benchmarks after Data Partitioning. 211 B.18 Energy Consumption of Benchmarks after Loop Nest Splitting 211 C.1 PMC Values for original Benchmark Versions C.2 PMC Values for Benchmarks after Advanced Code Hoisting. 214 C.3 PIC Values for original Benchmark Versions

14 xiv List of Tables C.4 PIC Values for Benchmarks after Advanced Code Hoisting C.5 PC Values for original Benchmark Versions C.6 PC Values for Benchmarks after Advanced Code Hoisting C.7 Runtimes of original Benchmark Versions C.8 Runtimes of Benchmarks after Advanced Code Hoisting C.9 Code Sizes of original Benchmark Versions C.10 Code Sizes of Benchmarks after Advanced Code Hoisting C.11 Energy Consumption of original Benchmark Versions C.12 Energy Consumption of Benchmarks after Advanced Code Hoisting D.1 PMC Values of Ring Buffer Replacement for CAVITY D.2 PIC Values of Ring Buffer Replacement for CAVITY D.3 PC Values of Ring Buffer Replacement for CAVITY D.4 Runtimes of Ring Buffer Replacement for CAVITY D.5 Code Sizes of Ring Buffer Replacement for CAVITY D.6 Energy Consumption of Ring Buffer Replacement for CAVITY 222