Compiler Optimisation

Transcription

1 Compiler Optimisation Michael O Boyle mob@inf.e.ac.uk Room 1.06 January, 2014

2 1 Two recommene books for the course Recommene texts Engineering a Compiler Engineering a Compiler by K. D. Cooper an L. Torczon. Publishe by Morgan Kaufmann 2003 Optimizing Compilers for Moern Architectures: A Depenence-base Approach by R. Allen an K. Kenney. Publishe Morgan Kaufmann 2001 Avance Compiler Design an Implementation by Steven S. Muchnick, publishe by Morgan Kaufmann. (extra reaing - not require) Aitional papers especially for the later part of course - beyon books Note Slies o not replace books. Provie motivation, concepts an examples not etails.

3 2 How to get the most of the course Rea ahea incluing exam questions an use lectures to ask questions L1 is a recap an sets the stage. Check you are comfortable Take notes. Do the course work an write well. Straightforwar - scheule smartly. Exam results ten to be highly bi-moal If you are struggling, ask earlier rather than later If you on t unerstan - it s probably my fault - so ask!

4 3 Course Structure L1 Introuction an Recap L2 Course Work - again upate from last year 4-5 lectures on classical optimisation (Base on Engineering a Compiler) 5-6 lectures on high level/parallel (Base on Kenney s book + papers) 4-5 lectures on aaptive compilation (Base on papers) Aitional lectures on course work/ revision/ external talks/ research irections

5 4 Overview - Recap Compilation as translation an optimisation Compiler structure Phase orer lexical, syntactic, semantic analysis Naive coe generation an optimisation Next lecture looks at coursework an then focus on scalar optimisation -mile en

6 5 Compilation Compilers : map user programs to harware. Translation - must be correct Hie unerlying complexity. Machines are not Von Neumann Current focus : Optimisation go faster, smaller, cooler. 40+ years. In general uneciable, sub-problems at least NP-complete Try to solve uneciable problem in less time than execution! Tackling a universal systems problem: Java to x86, VHDL to netlists etc. Gap between potential performance an actual wiening - compilers help?

7 6 Compilation as translation vs optimisation Moern focus is on exploiting architecture features Exploiting parallelism: instruction, threa, multi-core, accelerators Effective management of memory hierarchy registers,li,l2,l3,mem,disk Small architectural changes have big impact Compilers have to be architecture aware -coesign e.g. RISC Optimisation at many levels source, internal formats, assembler

8 7 Compiler structure source Front En HL AST Restruct HL AST Mile En Low IR Back En assembler Front en translates strings of characters into a structure abstract syntax tree Mile en attempt machine inepenent optimisation. Can also inclue source to source transformations - restructurer - outputs a lower level intermeiate format Many choices for IRs. Affect form an strength of later analysis or optimisation Backen: coe generation, instruction scheuling an register allocation

9 8 Phase Orer Lexical Analysis: Fins an verifies basic syntactic items - lexemes, tokens using finite state automata Syntax Analysis: Checks tokens follow a grammar base on a context free grammar an buils an Abstract Syntax Tree (AST) Semantic Analysis: Checks all names are consistently use. Various type checking schemes employe. Attribute grammar to Milner type inference. Buils a symbol table Optimisation + Coe generation - later lectures

10 Lexical Analysis Tokens inclue keywors int, ientifiers main upate an constants 10E6 Tokens efine using regular expression (RE), alphabet Σ,,, ǫ Input to scanner generators translate to NFA an simplifie to DFA Number of states = size of table. No impact on scan time complexity Moern languages use white space as separators. DO i = 1. 16! l = (a b... z A B... Z) = ( ) integer = real =. exp =. E 9

11 Lexical Analysis as eterministic finite automata exit Int 10. exit Real. E exit Exp How are the following classifie? 0, 01, 2.6, 2., 2.6E2 an 2E20

12 Syntax Analysis Tokens form the wors or terminals for the grammar. RE not powerful enough. Use context free grammar (CFG) base on BNF variants Next strip out syntax sugar an buils AST Form of CFG etermines type of language an parser family. Top own vs Bottom up. Automation, error hanling. Grammar rewriting expr = term (op expr) term = number i op = + Example: parse x 2 y 11

13 12 Syntax Analysis x 2 y expr term op expr i term op expr x number 2 * term i y Impact on bining of operators. x 2 y is parse as x (2 y). What about x 2 y?

14 13 term expr op i term op expr x number 2 The Abstract Syntax Tree expr * term i y x * 2 y The straightforwar parse tree has many intermeiate steps that can be eliminate This cutown tree is known as the abstract syntax tree an is a central ata structure use by compilers

15 14 Semantic Analysis One name can be use for ifferent vars epening on scope. Symbol table Type checking. Attribute grammars augment BNF rules with type rules expr = term (op expr) term = num i op = + expr.type = term.type (F op expr.type) term.type = num.type i.type F op = F F + F x 2 y int:x, real:y, int <real Difficult to a non-local knowlege : A-hoc syntax approaches, yacc Higher orer functional languages an ynamic typing make things interesting

16 15 Semantic Analysis x 2 y int:x, real:y, int <real real F int real int x * real int int real int 2 y real real real real Can be use for type inconsistencies/errors F int int real ouble int real ouble real real real ouble ouble ouble

17 16 Basic Coe Generation Translate AST in to assembler. Walk through the tree an emit coe base on noe type Hanle proceure calls an storage layouts. Assume activation recor pointer in register r 0 Loaing value x into register r 3 - ILOC instruction set (EaC) > r > r 1 (Not a mem op) Loa aress offset loaa0 r 0, r 1 > r 3 Mem[r 0 + r 1 ] > r 3

18 Coe Generation Typical top own generator - left to right x 2 y case op gen(left(noe), right(noe), op(noe)) case ientifier reg = nextreg() gen( loai, offset(noe),reg) gen( loaa0, r0,reg,reg) case num gen(loai, val(noe),nextreg()) Optimisations inclue elimination of reunancy. Unnecessary loas This scheme assumes unboune registers - nextreg() 17

19 18 5 Coe Generation 1 3 registers use x * 4 2 y 2 3 -> r1 1 loaa0 r0,r1 -> r1 1 loai 2 -> r2 2 -> r3 3 loaa0 r0,r3 ->r3 3 mult r2,r3 -> r3 4 sub r1,r3->r3 5

20 19 Optimisation Generate more efficient coe -eliminate reunancy Different traversal - less registers a = b*c + t = b*c e = 2-b*c a = t + e = 2- t 4 x * 2 5 -> r1 3 loaa0 r0,r1 -> r1 loai 2 -> r2 mult r2,r1 -> r1 -> r2 loaa0 r0,r2->r2 sub r2,r1->r2 2 y 1

21 20 In first part of course Machine moels/ Optimisation goals Assume uni-processor with instruction level parallelism, registers an memory Generate assembler shoul not perform any reunant computation Shoul utilise all available functional units an minimise impact of latency Register access is fast compare to memory but limite in number. Use wisely Two flavours consiere superscalar out-of-orer vs VLIW: Dynamic vs static scheuling Later consier multi-core architecture

22 21 Summary Compilation as translation an optimisation Compiler structure Phase orer lexical, syntactic, semantic analysis Naive coe generation an optimisation Next lecture course work Monay next week Jan 20 lecture postpone Then scalar optimisation - mile en