Comp 512 Spring 2011 Welcome to COMP 512

  • Slides: 23
Download presentation
Comp 512 Spring 2011 Welcome to COMP 512 Copyright 2011, Keith D. Cooper &

Comp 512 Spring 2011 Welcome to COMP 512 Copyright 2011, Keith D. Cooper & Linda Torczon, all rights reserved. Students enrolled in Comp 512 at Rice University have explicit permission to make copies of these materials for their personal use. Faculty from other educational institutions may use these materials for nonprofit educational purposes, provided this copyright notice is preserved. COMP 512, Rice University 1

COMP 512 This is COMP 512 — “Advanced Compiler Construction” • Subject Matter >

COMP 512 This is COMP 512 — “Advanced Compiler Construction” • Subject Matter > Compiler-based code improvement techniques ® Sometimes called “optimization” Analysis required to support them > No vector or multiprocessor parallelism > ® See COMP 515, taught by Vivek Sarkar • Required Work Mid-term (25%), Final (25%), & Project (50%) > Details of project will depend on class size > Notice: Any student with a disability requiring accommodations in this class is encouraged to contact me after class or during office hours. Students should also contact Rice’s Coordinator for Disabled Student Services COMP 512, Rice University 2

What About Reading Materials? • We will use many different sources Chapters 8, 9,

What About Reading Materials? • We will use many different sources Chapters 8, 9, & 10 of “Engineering a Compiler” > The original papers > You will learn more if you actually read the papers > Part of your education is learning to read technical papers and think critically about their contents • Slides from lecture will be available on the web site http: //www. cs. rice. edu/~keith/512 > I will try to post them before class > Your part is to read the material before coming to class COMP 512, Rice University 3

COMP 512 My goals • Convey a fundamental understanding of the current state-of-the-art in

COMP 512 My goals • Convey a fundamental understanding of the current state-of-the-art in code optimization and code generation • Develop a mental framework for approaching these techniques • Differentiate between the past & the present • Motivate current research areas (and expose dead problems) Explicit non-goals • Cover every transformation in the “catalog” • Teach every data-flow analysis algorithm • Cover issues related to multiprocessor parallelism COMP 512, Rice University ||’ism 4

COMP 512 Rough syllabus • Introduction to optimization Motivation & history > An example

COMP 512 Rough syllabus • Introduction to optimization Motivation & history > An example compiler > Some simple examples > • Data-flow analysis (Fortran H) (new Chapter 8, Ea. C) (Chapter 9, Ea. C) Iterative algorithm > SSA construction > • Classical scalar optimization (Chapter 10) Taxonomy for transformations > Populate the taxonomy > • Combining optimizations • Analyzing and improving whole programs COMP 512, Rice University (papers) 5

COMP 512 For next class read R. G. Scarborough and H. G. Kolsky, “Improved

COMP 512 For next class read R. G. Scarborough and H. G. Kolsky, “Improved Optimization of FORTRAN Object Programs”, IBM Journal of Research and Development, November, 1980, pages 660 -676. The point of this particular reading is to show you that a legendary optimizing compiler does not need to do every known transformation. Instead, it can do a handful of things and do them well. COMP 512, Rice University 6

COMP 512 Many educated computer scientists have serious misperceptions about compiler-based code optimization •

COMP 512 Many educated computer scientists have serious misperceptions about compiler-based code optimization • For example, see the Wikipedia entry “Compiler Optimization” Look in the history section for the page as of roughly January 2009; it may change & should only get more accurate > You may feel compelled to provide edits to it > ® Please register if you do; anonymous edits lack authority • By the end of COMP 512, you will be literate in the field of scalar code optimization, to the point where you will be capable of writing a much better entry than the current one (1/1/09) COMP 512, Rice University Keith – Switch to Firefox. 7

COMP 512 How does optimization change the program? Source Program Compiler Target Program Optimizer

COMP 512 How does optimization change the program? Source Program Compiler Target Program Optimizer tries to 1. Eliminate overhead from language abstractions 2. Map source program onto hardware efficiently > Hide hardware weaknesses, utilize hardware strengths 3. Equal the efficiency of a good assembly programmer COMP 512, Rice University 8

COMP 512 What does optimization do? Output 1 Input 1 Compiler Output 2 Output

COMP 512 What does optimization do? Output 1 Input 1 Compiler Output 2 Output 3 Output 4 • The compiler can produce many outputs for a given input The user might want the fastest code > The user might want the smallest code > The user might want the code that pages least > The user might want the code that … > • Optimization tries to reshape the code to better fit the user’s goal COMP 512, Rice University 9

COMP 512 • Some inputs have always produced good code First Fortran compiler focused

COMP 512 • Some inputs have always produced good code First Fortran compiler focused on loops > PCC did well on assembly-like programs > Input 1 Output 1 Input 2 Output 2 Input 3 Input 4 Compiler Output 3 Output 4 • The compiler should provide robust optimization Small changes in the input should not produce wild changes in the output > Create (& fulfill) an expectation of excellent code quality > Broaden the set of inputs that produce good code > • Routinely attain large fraction of peak performance COMP 512, Rice University (not 5%) 10

COMP 512 Good optimizing compilers are crafted, not assembled • • • Consistent philosophy

COMP 512 Good optimizing compilers are crafted, not assembled • • • Consistent philosophy Careful selection of transformations Thorough application of those transformations Careful use of algorithms and data structures Attention to the output code Compilers are engineered objects • • Try to minimize running time of compiled code Try to minimize compile time Try to limit use of compile-time space With all these constraints, results are sometimes unexpected COMP 512, Rice University 11

COMP 512 One strategy may not work for all applications • Compiler may need

COMP 512 One strategy may not work for all applications • Compiler may need to adapt its strategies to fit specific programs Choice and order of optimizations > Parameters that control decisions & transformations > • Emerging field of “autotuning” or “adaptive compilation” Compiler writer cannot predict a single answer for all possible programs > Use learning, models, or search to find good strategies > • Currently, lots of people are working in this area Many submissions to CGO & PLDI > We’ll talk about some of the problems & solutions > COMP 512, Rice University 12

A quick look at real compilers Consider inline substitution • Replace procedure call with

A quick look at real compilers Consider inline substitution • Replace procedure call with body of called procedure Rename to handle naming issues > Widely used in optimizing OOPs > • How well do compilers handle inlined code? We will talk about inlining later. For now, focus on how compilers handle inlined code. Characteristics of inline substitution • Safety: almost always safe • Profitability: expect improvement from avoiding the overhead of a procedure call and from specialization of the code • Opportunity: inline leaf procedures, procedures called once, others where specialization seems likely COMP 512, Rice University 13

A quick look at real compilers The study Five good compilers! • Eight programs,

A quick look at real compilers The study Five good compilers! • Eight programs, five compilers, five processors • Eliminated over 99% of dynamic calls in 5 of programs • Measured speed of original versus transformed code Inliner Source Program Compiler Execute & time Experimental Setup • We expected uniform speed up, at least from call overhead • What really happened? COMP 512, Rice University 14

A quick look at real compilers Do you see a pattern in this data?

A quick look at real compilers Do you see a pattern in this data? COMP 512, Rice University 15

A quick look at real compilers And this happened with good compilers! What happened?

A quick look at real compilers And this happened with good compilers! What happened? • Input code violated assumptions made by compiler writers Longer procedures > More names > Different code shapes > • Exacerbated problems that are unimportant on “normal” code Imprecise analysis > Algorithms that scale poorly > Tradeoffs between global and local speed > Limitations in the implementations > The compiler writers were surprised COMP 512, Rice University ( most of them) 16

A quick look at real compilers One standout story • MIPS M 120/5, 16

A quick look at real compilers One standout story • MIPS M 120/5, 16 MB of memory • Running standalone, wanal 1 took > 95 hours to compile Original code, not the transformed code > 1, 252 lines of Fortran (not a large program) > COMP 512 met twice during the compilation > • Running standalone with 48 MB of memory, it took < 9 minutes • The compiler swapped for over 95 hours !? ! • For several years, wanal 1 was a popular benchmark > Compiler writers included it to show their compile times! COMP 512, Rice University 17

COMP 512 Intent of this class • Theory & practice of scalar optimization The

COMP 512 Intent of this class • Theory & practice of scalar optimization The underpinning for all modern compilers > Influences the practice of computer architecture > • • Learn not only “what” but also “how” and “why” Provide a framework for thinking about compilation Class will emphasize transformations Remember register windows? Analysis should be driven by needs of transformations Role of the lab • Critically important to provide hands-on experience • Little time pressure COMP 512, Rice University 18

Disclaimer: The following slides contain a rough history of code optimization from 1955 to

Disclaimer: The following slides contain a rough history of code optimization from 1955 to 2000. They are intended to convey to you my own impressions of what was happening in the field. They are quite subjective. They are quite incomplete. (Hundreds of papers were published during each five year period. I cannot, and did not, try to be comprehensive. ) They are based on perusing conference proceedings for the various periods. Events are listed when (in my perception) the subject came to the fore. In some cases, this is different than when the idea first appeared. For example, software pipelining was clearly invented by Glaser & Rau in 1981. That notwithstanding, the technique became widely known and understood in the latter half of the 1980’s, which is why I cited the two PLDI 88 papers. Again, this history is neither definitive or objective. - Keith COMP 512, Rice University 19

A Sense of History 1955 -1959 Commercial compilers generated good code Fortran Separation of

A Sense of History 1955 -1959 Commercial compilers generated good code Fortran Separation of concerns (Backus, 1956) Cobol Control-flow graph, register allocation (Haibt, 1957) 1960– 1964 Academics try to catch up with industrial trade secrets Algol 60 Early algorithms for “code generation” (1960, 1961) Relating theory to practice (Lavrov, 1962) Alpha project at Novosibirsk (Ershov, 1963 & 1965) 1965 -1969 Technology begins to spread PL/I Fortran H (Medlock & Lowry, 1967) Algol 68 Value numbering (Balke, 1967 ? ) Simula 67 Literature begins to emerge (Allen, 1969) COMP 512, Rice University 20

A Sense of History 1970 -1974 The literature explodes and optimization grows up SETL

A Sense of History 1970 -1974 The literature explodes and optimization grows up SETL Cocke & Schwartz, Allen-Cocke Catalog, 1971 Smalltalk Theory of analysis (Kildall, 1971, Allen & Cocke, 1972) Lisp Interprocedural analysis (Spillman, 1972) APL Strength reduction, dead code elimination, Live (SETL) Expression tree algorithms (Sethi, Aho & Ullman) 1975 -1979 Global optimization comes of age Pascal Full literature of data-flow analysis CLU Strength reduction (Cocke & Kennedy, 1977) Alphard Partial redundancy elimination (Morel & Renvoise, 1979) Com. Lisp Inline substitution studies (Scheiffler, 1977, Ball, 1979) Tail recursion elimination (Steele, 1978) Data dependence analysis (Bannerjee, 1979) COMP 512, Rice University 21

A Sense of History 1980 -1984 Programming environments and new architectures Smalltalk 80 Incremental

A Sense of History 1980 -1984 Programming environments and new architectures Smalltalk 80 Incremental analysis (Reps, 1982; Ryder, Zadeck, 1983) ADA Incremental compilation (Schwartz et al. , 1984) Scheme Interprocedural analysis (Myers, 1981; Cooper, 1984) RISC compilers (PL. 8, 1980; MIPS, 1983) Graph coloring allocation (Chaitin, 1981; Chow, 1983) Vectorization (Wolfe, 1982; R. Allen, 1983) 1985– 1989 Resurgence of interest in classical optimization C++ Constant propagation (Wegman & Zadeck, Torczon, 1985) ML Code motion of control structures (Cytron et al. , 1986) Modula-3 Value numbering (Alpern et al. , Rosen et al. , 1988) Software pipelining (Lam, Aiken & Nicolau, 1988) Pointer analysis (Ruggeri, 1988) SSA-form (Cytron et al. , 1989) COMP 512, Rice University 22

A Sense of History 1990 -1994 Architects (and memory speed) drive the process Fortran

A Sense of History 1990 -1994 Architects (and memory speed) drive the process Fortran 90 Hierarchical allocation (Koblenz & Callahan, 1991) Scalar replacement (Carr 1991) Cache blocking (Wolf, 1991) Prefetch placement (Mowry, 1992) Commercial interprocedural compilers (Convex, 1992) 1995– 1999 The internet & SSA both come of age Java JIT compilers (Everyone, from 1996 to present) Perl (? ) Code compression (Franz, 1995; Frasier et al. , 1997; . . . ) SSA-based formulations of old methods (lots of papers) Compile to VM (Java, 1995; Bernstein, 1998; … ) Memory layout optimizations (Smith, 19? ? ; others …) Widespread use of analysis (pointers, omega test, …) It’s still too early for an epitaph ! COMP 512, Rice University * 23