A Multi-Technique C Inliner

Chengyan Zhao

Supervisor : Dr. Owen Kaser

Graduate Academic Unit of

Computer Science

University of New BrunswickJuly 7, 1998

MCS thesis

by

2

Organization of presentation

1. Introduction

2. Parsing

3. Inlining Technique

4. Decision-making algorithm

5. Conclusion

3

1 Introduction What is Inlining?

…

t = strlen(“Hello”); /*callsite */

...

int strlen(char * s){

register int n;

n = 0;

while (*s++) n++;

return ;}

…

{ int AA0000 = 0;

char * s = “Hello”;

register int n;

n = 0;

while ( *s++) n++;

AA0000 = n;

t = AA0000;

}…

int strlen(char * s){

/* body of strlen */ }

Replacement of callsite

Parameter Passing Simulation

Return simulation

1. Introduction

4

Why Perform Inlining?

• Remove expensive call-return instructions• Remove parameter loads and stores• Increased opportunities for optimizations

t = A(15);

…

int A(int index){

return 3 * index + 1;

}

{ int temp_i = 15;

int AA0001 = 0;

{ AA0001 = 3 * temp_i + 1;

goto exit_01;}

exit_01: t = AA0001;

} ...

Before Inlining After inlining

1. Introduction

5

Inlining vs. Textual Replacement

if ( f(x) < 10)

…

int f(int index){

/* body of function f */

}

if( { /* duplicated body of function f(x) */}

< 10)

…

int f(int index){

/* body of function f */

}

Reason:ANSI C standard does not allow curly brackets and commands in expressions

Solution:Callsite standardization

Invalid Modification

1. Introduction

6

CSourceCode

CPreprocessing

ParsingC

preprocessed

code

Swapping andSplitting

CollectingProfileInformation

Parse tree

MakingInliningDecision

InliningSelected functioncallsite to beinlined

Adjusting andRegenerating

Incremental inlining

Creating theFinal Result

Parse tree

Standardizedparse tree

Parse tree

Done inlining

8 Major steps

Action of each step

Top View of Inlining

1. Introduction

7

2 Parsing• Build a parser• Parse-tree management -- message system

– Overview

Tree1 :Tree

node1 : BinaryNode

node2 ='j' :StringNode

Tree3 :Tree

node3 = '=' :StringNode

Tree4 :Tree

node4 = '2' :IntegerNode

Tree2 :Tree

1:

2:4:

6:

3: 5: 7:

Parse-tree for “j=2;”

Parse-tree introduction

Tree composition

2. Parsing

8

Message Connector

void Tree:::BroadCastMessage(CMessage &Msg){ node_->BroadCastMessage (Msg); }

Message Processorvoid ParseNode::BroadCastMessage(CMessage & Msg){ switch(Msg.Id){ case Type_of_message: ProcMsg(Msg); break; … } }

2. Parsing

Message broadcasting and receiving

Message BroadcasterTree * Tree::MsgBroadcaster(Parameter){ CMessage Msg; Compose Message; BroadCastMessage(Msg); Return Result; }

9

• Standardized format– f(e1,e2,…,en);– y = f(e1,e2,…,en);

• Swapping for statements– Overview of swapping – Swapping process

• Swap conditional controls• Swap return statement• Swap declaration

• Splitting for expressions– Comma operator removal – Split short-cut operators (and, or)– Split ? : expression– Split expression– Split nested callsite– Limitation on splitting

3 Inlining Technique3 Inlining Technique

10

Overview of Swapping General rule for if construct swapping

if(condition)

then_statement else_statement

Tree List

‘=’ if(temp1)

temp1 condition then else statement statement

Swapping

Parse-tree representation for swapping if construct

Parse-node message processing

void IfThenElseNode::BroadCastMessage(CMessage &Msg){ Switch(Msg.Id){ case ck_SwapIfCond: SwapIfCond(Msg); break; … } // end of switch … } // end of message filter

3. Inlining Technique

if(condition)

then_statement

else

else_statement;

{ int temp01 = condition;

/* progressing */

if(temp01) then_statement;

else

else_statement; }

11

Spliting

Rule for removing comma operator Comma operator removal

Split short-cut operator (&&)

temp01 = expr1, expr2, …,exprn;

expr1;expr2;…temp01 = exprn;

Comma operator removal

t = (f1(x) && (f2(b)+10));

{ int temp01 = 0; int temp02 = f1(x); if(temp02){ int temp03 = f2(b) + 10; if(temp03) temp01 = 1; } t = temp01; }

3. Inlining Technique

12

Split expression (may not always work)

… /*assume t = 10 */x = (t++)+f(t) +(t--);…

…type_of_f temp01;temp01 = f(t);x=(t++)+ temp01 +(t--);…

3. Inlining Technique

13

• Inline a callsite– Body duplication– Parameter passing simulation

• One-dimensional array passing

– Renaming– Return simulation– Specialization opportunity– Callsite removal

3 Inlining Technique

One-dimensional array parameter passing Simulation…f(s); /* callsite, s is an one-dimensional array type */…void f(int a[ ]){ … a[1] = 2; t = a[5]; … }

{int * temp01 = s; … temp01[1] = 2; t = temp01[5];…}…void f(int a[ ]){… a[1] = 2; t = a[5]; … }

14

4 Decision algorithm Eliminate uninlineable callsites Profile information collection

Profiler limitation Dummy functions

Algorithm Inliner’s features

Automatic, source-level Profile-guided Multi-technique (const-propagation, version issue, cache issue)

Testings

4 algorithm

15

Profile Information Collection- Profiler Build-in limitation

- Dummy function creation

f( ){ … f1( ); … f1( ); … }

… … f … 15 … f1 … …

Profile sourceprogram

Source program Profile information

4 algorithm

f(void){

f1( ); …

f2( ); … }

void Dummy01(void);

void Dummy01(void);

…

f(void){

{ f1( ); Dummy01( );}

...

{ f1( ); Dummy02( ); }

…

}

…

void Dummy01(void){}

void Dummy01(void) {}

… … f

… 15 … f1

… 10 … Dummy01

… 5 … Dummy02

… ...

16

Inlining Decision Algorithm

for each callsite in the program { Determine the benefit for each type of inliningAdjust benefits for constant propagation Adjust benefits for hazardous situations }

Decide best callsite and best type Do Inlining Update parse tree and decision-making data structures

4 algorithm

17

Experimental ResultsSpeed Improvement after 20% code

expansion

3.71

18.48

11.37 10.6317.28

05

101520

1

Programs

%

strcpynfibtaknqueensumfrom

Speed improvement after 50% code expansion

3.6922.25 20.88 19.64

40.58

0204060

1

Programs

%

strcpynfibtaknqueensumfrom

Speed improvement after 100% code expansion

7.9126.16 20.96

32.2942.53

0204060

1

Programs

%

strcpynfibtaknqueensumfrom

4 algorithm

Gcc 2.7.2.1

PentiumPro 266 machine

18

5. Conclusion and future work Conclusion

OO parser Inlining optimization (4% ~ 43%) Design patterns

Future work Multi-file support Pipeline interlock Database in inlining

5. Conclusion

19

References[1] Flex lexical analyzer generator v 2.91, Free Software Foundation, last updated at 09/10/96

[2] Berkeley Yacc Parser Generator, University of California at Berkeley, last update at:01/30/90

[3] Frank Buschmann, Regine Meunier, Hans Rohnert, Peter Sommerlad and Michael Stal,Pattern-Oriented Software Architecture, John Wiley & Sons, 1996

[4] American National Standard Institute C grammar, Comp.compilers Archie site located at:http:/ / iecc.com/cgi-bin/getarticle?91-09-030 ANSI 1988

[5] Jack W. Davidson and Anne M. Holler, A Study of a C Function Inliner. Software –Practice and Experience, 18:775-790, 1988

[6] Jack W. Davidson and Anne M. Holler. Subprogram Inlining: A Study of its Effects onProgram Execution Time. IEEE Transactions on Software Engineering, Vol. 18, No.2, Feb.1992.

[7] Pohua P. Chang, Scott A. Mahlke, William Y. Chen and Wen-mei W. Hwu. Profile-GuidedAutomatic Inline Expansion for C Programs. Software – Practice and Experience, 25:349-369,1992

[8] Keith D. Cooper, Mary W. Hall and Linda Torczon. Unexpected Side Effects of InlineSubstitution: A case study. ACM Letters on Programming Languages and Systems, 1:22-32,1992

[9] Jeffery Dean and Craig Chambers. Towards Better Inlining Decisions Using Inlining Trials.ACM Conference on LISP and Functional Programming, Orlando, Fl, June 1994

[10] Wen-mei W. Hwu and Pohua P. Chang. Inline Function Expansion for Compiling CPrograms. The SIGPLAN ’89 Conference on Programming Language Design andImplementation, pages 246-255, 1989

[11] Owen Kaser and C.R. Ramakrishnan. Evaluating Inlining Techniques. Technical ReportTR 96-001, Dept of Math, Stat and Computer Science, U. of New Brunswick, Saint John,Canada, Aug 1996. Scheduled to appear in Programming Languages, 1998

[12] Scott McFarling. Procedure Merging with Instruction Caches. Proceedings of theSIGPLAN’ 91 Conference on Programming Language Design and Implementation, ACM,1991

20

[13] Robert W. Scheifler. An Analysis of Inline Substitution for a Structured ProgrammingLanguage. Communications of the ACM, 20(9): 647-654, 1977.

[14] Graham, S.L., Kessler, P.B., McKusick, M.K., g̀prof: A Call Graph Execution Profiler,Proceedings of the SIGPLAN '82 Symposium on Compiler Construction, SIGPLANNotices, Vol. 17, No. 6, pp. 120-126, June 1982.http:/ /www.cs.ubc.ca/ local/ software/ GNU_info (accessible April 1998)

[15] Minda Zhang, Analysis of Object File Formats for Embedded Systems, Intel Corporation,June 1995.http:/ / developer.intel.com/design/ intarch/ PAPERS/ESC_FILE.HTM (accessible April1998)

[16] Executable Language Specification, Intel Corporationftp:/ / ftp.intel.com/ pub/ tis/ elf11g.zip (accessible March, 1998)

[17] Keith D. Cooper, Mary W. Hall and Linda Torczon, Unexpected Side Effects on InlineSubstitution: A Case Study, ACM Letters on Programming Languages and Systems. Vol. 1,No. 1, March 1992. Pages 22-32

[18] Rinus Plasmeijer and Marko van Eekelen, Functional Programming and Parallel GraphRewriting, Addison-Wesley, 1993

[19] S. Richardson and M. Ganapathi. Interprocedural analysis vs. procedure integration.Information Processing Letters, 32(3): 137-142, August 1989

[20] S. Richardson and M. Ganapathi. Interprocedural optimization: Experimental results.Software -–Practice and Experience, 19(2): 149-169, February 1989

[21] Alfred V. Aho, Ravi Sethi and Jeffrey D. Ullman, COMPILERS Principles, Techniquesand Tools, Addison Wesley Publishing Company, 1986.

[22] HP Compiler Optimization White Paper, Hewlett Packard Corporation. November 1996http:/ /www2.hp.com/wsg/ ssa/ fortran/optimiz.html

[23] Owen Kaser, C.R. Ramakrishnan, I.V. Ramakrishnan, R.C. Sekar, EQUALS – A FastParallel Implementation of a Lazy Language, Journal of Functional Programming, Vol. 7 No.2, 1997.

[24] GNU C++ language compiler package, version 2.7.2.1, Free Software Foundation. Ftpsite at http:/ /www.cs.ubc.ca/ local/ software/GNU_info gnu Canadian ftp mirror.

[25] Linux operating system package for PC version 5.0, kernel 2.0.29, RedHat SoftwareIncorporated. Ftp download site at http:/ /www.redhat.com/ (installed in May 1997)

21

[26] Intel C/C++ Compiler White Paper, Intel Corporation.Http site located at http:/ / 134.134.214.2/design/perftool/ icl24/ icl24wht.htm (accessibleApril 1998)

[27] Unified Modeling Language standard published by Rational Rose Software.Http site located at http:/ /www.rational.com/uml/html/ semantics/ semanta1.html (accessibleApril 1998)