Power Programmierung 2

home *** CD-ROM | disk | FTP | other *** search

/ Power Programmierung 2 / Power-Programmierung CD 2 (Tewi)(1994).iso / tools / profiler / tcprof / tcprof.doc next >

Wrap

Text File | 1988-10-04 | 40.2 KB | 1,239 lines

Turbo C profiler (TCPROF version 1.5) Copyright (C) 1988 Andre Nancoz all rights reserved Contents INTRODUCTION 1 Overview . . . . . . . . . . . . . . . . . . . . . . . 1 History . . . . . . . . . . . . . . . . . . . . . . . . 1 Copyright information . . . . . . . . . . . . . . . . . 1 Trademark information . . . . . . . . . . . . . . . . . 1 System requirements . . . . . . . . . . . . . . . . . . 2 Compilation and execution requirements . . . . . . . . 2 Restrictions and caveats . . . . . . . . . . . . . . . 2 USING THE PROFILER 4 Executing the profiler . . . . . . . . . . . . . . . . 4 Profiler options . . . . . . . . . . . . . . . . . . . 4 INPUT/OUTPUT 6 File usage . . . . . . . . . . . . . . . . . . . . . . 6 Report content . . . . . . . . . . . . . . . . . . . . 6 MEMORY USAGE 9 Lines within functions . . . . . . . . . . . . . . . . 9 Functions only . . . . . . . . . . . . . . . . . . . . 9 ERROR MESSAGES 10 APPENDIX A 11 LISTING 1..............................................11 EXAMPLE 1..............................................12 EXAMPLE 2..............................................13 APPENDIX B 14 Registered user information . . . . . . . . . . . . . .14 Registered user features . . . . . . . . . . . . . . .14 APPENDIX C 17 PROBLEM REPORT/SUGGESTIONS . . . . . . . . . . . . . .17 REGISTERED USER ORDER FORM . . . . . . . . . . . . . .18 i INTRODUCTION Overview TCPROF will show you where the most time is being spent in a program written with Borlands' Turbo C. The output report indicates the percentage of time (numerically and graphically) spent in each function and line of your program. Several options allow you to vary the amount of information shown on the report. No changes to your source code are required in order to run the profiler. The hardware timer is read but not modified. This program is distributed at no charge and is designed for users who have an occasional need or curiosity about the efficiency of their programs. A registered user version of this program (TCPROFRU) is available which allows more extensive tailoring of the output and can reduce the amount of overhead and size of the report (See Appendix B for more information). History Version 1.0 First release. Version 1.5 Update for Turbo C release 2.0. Timing logic change to increase consistency. Copyright information Turbo C profiler program TCPROF and documentation Copyright (C) 1988 by Andre Nancoz all rights reserved The Turbo C profiler is not public domain software. TCPROF and it's accompanying documentation may be distributed freely provided no fee is charged and it is not part of a package for which a charge is made. Furthermore, the program and documentation must be distributed with no modifications. TCPROF is distributed as-is with no expressed or implied warranties. Trademark information Turbo C is a trademark of Borland International Inc. IBM-PC is a trademark of International Business Machines Corporation. MS-DOS is a trademark of Microsoft Corporation. October 4, 1988 INTRODUCTION - 1 System requirements IBM-PC and true compatibles. MS-DOS version 3.0 or above. Minimum memory required is 39K above that required for the program being profiled. Additional memory may be required depending on the program being profiled. See the section on memory usage for more information. Compilation and execution requirements In order to run a program under the profiler the program must be compiled with, at least, the following options: TCC command line; -y generate line numbers in .obj file. -M full link map. TLINK command line; /m or /s to generate a link map. /l output line numbers in link map. Integrated Development Environment; O/C/Code generation/Line numbers..On O/L/Map file..Publics or Detailed NOTE: Line numbers and map options are required even if the functions only option is used. When running the profiler the .EXE and the .MAP file must be available using the standard DOS search path, i.e. the current directory of the current drive is checked first. If not found, each directory in the environment PATH is checked. Restrictions and caveats Timer If your program steals the timer interrupt it must chain to the previous owner of the interrupt. The profiler does NOT modify the timer rate but does read the timer chip. If your program modifies the timer rate the profiler may still operate correctly. Single Step and Interrupt 3 If your program uses these, the profiler will not work correctly. It may be reboot time. longjmp Likely to mess up timing. Static functions All line numbers for a function defined as static will be included with the last non-static function. This will likely cause the warning message regarding inaccurate timing. The profiler keeps track of entrances to and exits from INTRODUCTION - 2 functions. Since static function addresses do not appear in the link map the profiler does not have an accurate indication of where the static function began or where the previous non-static function ended. Assembler Assembler modules should be assembled without the line number option(TASM /ZD). The profilers' timing resolution cannot handle interrupts at each machine instruction. Unlike previous versions of Turbo C, version 2.0 passes the line number option on to TASM when inline assembler is used. When a single machine instruction is generated for a line of C source it is likely that they will be reported as 0% due to the profilers timing resolution. Some functions such as enable(), disable(), etc. will cause the same problem. Use of the emit() function to output one instruction per line will also cause the problem. If you are using previous versions of Turbo C, use of the asm keyword within a C module will cause no line numbers to be in the map, TCC invokes MASM without the line number option. Therefore, functions which contain the asm keyword should be compiled in a separate module from functions which you wish to time. TSRs I have tested the profiler with several TSRs with no problems. The timing may be different for the same program depending on which TSRs are loaded. Therefore, in order to compare results of two profiler runs, the same environment should be used for each run. Of course activating a TSR during a profiler run would disturb the timing. Bugs This program is not a debugger and expects a reasonably behaved program. The profiler may be useful in detecting some problems such as loops or functions executing the wrong number of times or not executing at all, etc. INTRODUCTION - 3 USING THE PROFILER Executing the profiler TCPROF [prof options] program-name [program-name options] The brackets [] denote optional information, they are not entered. If the program-name is not entered, help information is displayed. Program-name options are passed to program-name in the same manner as if the program was executed stand-alone. The argv[0] parameter contains program-names' full path. The profiler report is written to stdout, any redirection applies to both program-name and the profiler report. The profiler uses stderr for error messages which cause it to abort the run, also for a few messages to the console when stdout is redirected. Profiler options Profiler options may be entered in upper or lower case. The leading character may be either a - or a /. All profiler options must precede the name of the program to be profiled. -F Profile by function. The default is a detailed report by line number within function. Using this option creates an output report with only one line per function and substantially reduces the profiler overhead when your program is running. -Cnn Print only lines whose execution count is equal to or greater than nn. The default value is 0. This option only influences the amount of report output. The option has no effect if the -F option is chosen. The nn is interpreted as an unsigned long integer. -Lnn Number of lines per page. The default is 25. If nn is omitted, 0 or an invalid number, the default is 58 (a good value for 8.5 inch printed output). This option does not change the "pause between each page" output action, which is controlled by whether or not stdout has been redirected. -Snn This option controls an internal memory allocation parameter. Entry to a function requires a word on an internal stack, the word is released at function exit. The value specified is the number of words to allocate. The maximum is 32751 minus the number of USING THE PROFILER - 4 functions profiled. It should only be required if your program uses recursion. The default value is 10. If the warning message "Timing may be inaccurate" occurs then this parameter should be used. It should be set to the maximum level of recursions that can occur at any one time. USING THE PROFILER - 5 INPUT/OUTPUT File usage When the program to be profiled is run only the standard Turbo C start up files are open. The spawn function passes these files to the program to be run. The profiler uses stdout and stderr before and after running your program. Any redirection applies to both the profiler and the program being profiled. Delimiter lines are written by the profiler to stdout before and after running your program. Besides the program to be profiled, the .MAP file must be available for read only access, this file is not open while your program is running. All file are located via the standard DOS path search. Report content The profile report is written to stdout with a default of 25 lines to the page, the L option can be used to change the default. If stdout has not been redirected the output will pause at the end of each page and you may either display the next page or cancel the remaining output, a clear screen is issued between each page. When output has been redirected, no pause or clear screen occurs between pages. Appendix A contains a listing and two reports of a small test program. The first report is by line number. This report exhibits very high profiler overhead since majority of time is spent within loops which generate a small number of machine instructions per line of C code. In more realistic programs overhead is in the 30% range. The second report is the same program but it was run with the functions only option. Only functions which are compiled with the line number option "on" are shown on the report. The time for functions compiled without line numbers are included in the lines of your program from where they are called. As a consequence, all the time for the system run time library functions (such as printf) will be included in the lines from which they are called. In the test program line 23 took 1 percent of the main functions' time, this includes the time it took to execute the printf function (with redirection to disk). On the other hand, the time for line 19 only includes the time to call and return from test_func, since test_func line numbers are included in the map. The first page of output contains general information about the program being profiled. In reality this may be many pages of output since it contains the stdout of the program being profiled. The options section on this page shows all the profiler options which were in effect during this execution. INPUT/OUTPUT - 6 Prior to running your program, the .MAP file is analyzed and a table is built in memory. After this is done the amount of memory available to run your program is printed. The "parameters" line contains the argv[0] to be passed to your program, subsequent line(s) contain argv[1] to argv[n] parameters each enclosed in quotes. A line is then output to delimit the stdout of your program from that of the profiler, it contains the full path and name of the program to be executed. If stdout has not been redirected the profiler will pause with a message allowing you to cancel execution or run the profile, optionally you may clear the screen prior to running your program. After returning from your program, a line is output giving the return code and the calculated time your program would have taken if the profiler had not been used. This time is the basis for the percentage calculations made in the body of the report. The next line contains the start date, time, end time and the percentage of the elapsed time spent within the profiler. The body of the report starts on the next page. In Appendix A, example 1 is lines within functions, example 2 is a functions only report. Whenever a new module is encountered in the .MAP file a line is output with the module name. When the beginning of a function is identified a line is output with the name, the percentage of time the function executed as compared to your programs total time from page 1 and a graphic representation of the percentage. In addition, for a functions only report the count column contains the number of times the function was executed, for a lines within function report this count appears on the first line within the function. The line number report has additional information. The first column contains the starting address of the line, the line number, the number of times the line was executed, the time taken by the line as a percentage of the programs' time, the time taken by the line as a percentage of the functions' time and a graphic representation of the percentages. Each column in the graph represents approximately 2.5%. Four characters are used within the graph, asterisks, periods, equal signs and dashes. The asterisks are used for function lines only, no other characters appear on these graph lines. Even though a report by lines is requested, if a function is not executed only the function line will appear on the report. The period only appears in the first column of the graph, it is used when a line was executed at least once but the time was less than 2.5%. The end of the equal signs represents the lines' time compared to the total time. The end of the dashes represents the lines' time compared to the functions' time. INPUT/OUTPUT - 7 For the lines within functions report, the format of the source file can affect the report. The listing in Appendix A contains some examples which will be referred to in the following discussion. The source for lines 33 and 34, and line 39 are essentially the same and generate identical machine instructions. On the other hand, the report is different for these two instruction sequences due to the map output. The compiler generates the initialization code of the "for loop" and outputs the line address for the .MAP file. Since the calculation is contained on the same line nothing is output to the .MAP file for it. Therefore, line 39 reports that the line was executed only once and all the time for the loop is included with line 39. In the case of lines 33 and 34, the compiler outputs the line 33 address for the initialization code and the line 34 address for the calculation which shows 19999 executions of the expression. In version 2 of Turbo C other loop constructs (while, do) are now generated similarly, so that lines 21 and 37 are now reported in the same manner as line 39. All times for the last line of a function are reported with the line which called the function. Therefore for best results, the closing bracket of a function should appear on a separate line. The count for the last line of a function is correct. The difference in times between lines 21 and 39 indicate the efficiencies of register variables. In the main function two dummy variables were allocated as registers so that the other variables, which are used in line 21, could not be automatically assigned by the compiler to registers. In test_func the variables for the loop were allocated to registers. In version 2.0 of Turbo C the compiler outputs line numbers to the .MAP file for floating point instructions. Therefore, these lines are now reported correctly. If you are using version 1.5 of Turbo C, lines 41 and 42 exhibit a problem when using floating point. If the first machine instruction to be generated on a line is for a floating point operation the compiler does not output a line number for the .MAP file. Therefore, line 42 does not appear in the report and all the time for line 42 is reported in line 41. If you need to see one or more floating point lines separately the only option would be to precede each statement with a non-floating point statement, such as an increment of a dummy integer variable (better yet, upgrade to version 2, it's worth it). Example: dummy++; r = a * b; /* on one line */ INPUT/OUTPUT - 8 MEMORY USAGE Lines within functions "Memory required" = 16 * "number of lines mapped" + 2 * ("number of functions within the line map" + "value of -S option") + 13 * "number of modules within the line map" + "your program requirements" + 39k Functions only "Memory required" = 16 * "number of functions within the line map" * 2 + 2 * ("number of functions within the line map" + "value of -S option") + 13 * "number of modules within the line map" + "your program requirements" + 39k MEMORY USAGE - 9 ERROR MESSAGES Return Code Message 35 DOS version must be 3.0 or later. 36 Parameter invalid in response file. 37 File not found (.exe and .map required). 38 Invalid map file, no line numbers. In the chapter "INTRODUCTION" see section "Compiler and execution requirements". 39 Insufficient memory. 40 Logic. This message is accompanied by other information. If it should occur please notify me, giving the full message displayed, details of your system and compile options used on the program being profiled. 41 Not a Turbo C program. 42 Invalid map file, no public values found. Refer to the section "Compiler and execution requirements". 43 NOTE: Timing inaccurate, see -S in documentation. Refer to the section "Profiler options". 44 NOTE: Timing inaccurate, see caveats in documentation. See the discussion on static functions under "Restrictions and caveats". 99 Parameter display or error. All return codes below 35 are identical to the errno values used by Turbo C. ERROR MESSAGES - 10 APPENDIX A --------------------------------- LISTING 1 --------------------------------- 1 #include <stdio.h> 2 #include <stdlib.h> 3 4 5 6 7 /* Prototype Definitions */ 8 void test_func(void); 9 /* End Prototypes */ 10 11 #define LOOP_CNT 20000 12 13 main() 14 { 15 register int reg_si, reg_di; /* dummy variables to use up regs */ 16 int a = 10; 17 int i = LOOP_CNT; 18 19 test_func(); 20 21 while (--i) a = a * a / a; 22 23 printf("Proftest executed"); 24 return (0); 25 } /* end of main */ 26 27 void test_func(void) 28 { 29 register unsigned int i; 30 register unsigned int a = 10; 31 float b, c; 32 33 for (i = LOOP_CNT; --i;) 34 a = a * a / a; 35 36 i = LOOP_CNT; 37 while (--i) a = a * a / a; 38 39 for (i = LOOP_CNT; --i;) a = a * a / a; 40 41 for (i = 50, b = 10.0; i > 0; i--) 42 c = b * b / b; 43 } /* end of test_func */ APPENDIX A - 11 ---------------------------------- EXAMPLE 1 ---------------------------------- Turbo C Profiler (TCPROF V1.5) 1 PROFILER OPTIONS: Minimum line count printed(-C) 0: Lines per page(-L) 58 Internal stack size(-S) 10 default Analyzing C:\C\PROFTEST.MAP 18 lines in 2 functions in 1 module will be profiled Building profile table Memory available 437432 bytes Parameters: C:\C\PROFTEST.EXE ***** STDOUT FOR C:\C\PROFTEST.EXE ***** Proftest executed *** End of Program Output Return code 0 Time mm:ss 0:03 *** Start 9-25-1988 16:41:26.67 End 16:41:42.22 Profiler overhead 84% -------- Example 1 Page 1: tcprof -L proftest >proftest.ex1 -------- PROFTEST.EXE Turbo C Profiler (TCPROF V1.5) 2 Func/Addr Line Count %Tot %Fun...1...2...3...4...5...6...7...8...9... MODULE PROFTEST main 36 *************** 0000:01FA 13 1 0 0 . 0000:0202 16 1 0 0 . 0000:0207 17 1 0 0 . 0000:020C 19 1 0 0 . 0000:020F 21 1 36 99 ==============-------------------------- 0000:0229 23 1 0 1 . 0000:0232 24 1 0 0 . 0000:0236 25 1 0 0 . test_func 64 ************************* 0000:023C 27 1 0 0 . 0000:0244 30 1 0 0 . 0000:0247 33 1 0 0 . 0000:024C 34 19999 21 33 ========----- 0000:025D 36 1 0 0 . 0000:0260 37 1 18 28 =======---- 0000:0273 39 1 18 28 =======---- 0000:0289 41 1 0 0 . 0000:0299 42 50 7 11 ===- 0000:02B6 44 1 0 0 . APPENDIX A - 12 *** End of profile report for C:\C\PROFTEST.EXE *** -------- Example 1 Page 2: tcprof -L proftest >proftest.ex1 -------- ---------------------------------- EXAMPLE 2 ---------------------------------- Turbo C Profiler (TCPROF V1.5) 1 PROFILER OPTIONS: FUNCTIONS ONLY(-F): Lines per page(-L) 58 Internal stack size(-S) 10 default Analyzing C:\C\PROFTEST.MAP 2 functions in 1 module will be profiled Building profile table Memory available 437656 bytes Parameters: C:\C\PROFTEST.EXE ***** STDOUT FOR C:\C\PROFTEST.EXE ***** Proftest executed *** End of Program Output Return code 0 Time mm:ss 0:03 *** Start 9-25-1988 16:42:58.56 End 16:43:02.08 Profiler overhead 12% ------- Example 2 Page 1: tcprof -L -F proftest >proftest.ex2 ------- PROFTEST.EXE Turbo C Profiler (TCPROF V1.5) 2 Func/Addr Line Count %Tot %Fun...1...2...3...4...5...6...7...8...9... MODULE PROFTEST main 1 38 *************** test_func 1 62 ************************* *** End of profile report for C:\C\PROFTEST.EXE *** ------- Example 2 Page 2: tcprof -L -F proftest >proftest.ex2 ------- APPENDIX A - 13 APPENDIX B Registered user information A registered users version of this program, TCPROFRU, is available for $20.00 ($25.00 for 3 1/2" diskettes). Please use the order form at the end of this document. The registered user version has additional options which allow you to tailor the amount of information gathered by the program. On larger programs this reduces the overhead of the profile runs and the size of the report. Registered user features -A[-]func-nm Alter the treatment of the named function. If the profiler is being run in the default mode of lines by function, each function named by this option will NOT have line numbers traced. A single line, as for a functions only report, is output for each named function. If the functions only option is "on" (-F) then the named functions will have line number tracing on. For either case above, if a minus sign precedes the named function it will not be traced and all times for the function will be included in the lines or functions that call it. Multiple A options may be used. The function name is case sensitive. Ex: tcprofru -F -Afn1 -A-fn2 tstprg Functions only plus fn1 lines and no trace on fn2. -M[-]mod-nm Alter the treatment of all functions in the named module. Depending on the status of the F switch the treatment of all functions within the named module will be modified. This is a shorthand method of altering(A[]) the treatment of all the functions in a module. Multiple M options may be used. The module name is not case sensitive. Ex: tcprofru -Mmod1 -M-mod2 testprog Profile functions only in mod1, no profile of functions in mod2, all other functions are profiled by line. APPENDIX B - 14 -F[-][+] Alter the treatment of all functions in the program. The -F- option causes all functions to be skipped except those specified by A or M options. The functions included by a -A or -M option are profiled by line unless the name is preceded by a -, in which case a single line per function is reported. This is convenient when only a few functions in a large program needs to be profiled. The -F+ option resets the default, profile by line. It would be used if a response file (-@ option) contains a -F option and you wish to override it. The last -F option encountered on the command line is the one in effect for all functions. The interaction between the -F, -M and -A options result in 27 possibilities. The table below is taken from the help screen of TCPROFRU. -F[-][+] | No or + | Yes | with - | -M[-]module | N | Y | - | N | Y | - | N | Y | - | -A[-]function|N Y -|N Y -|N Y -|N Y -|N Y -|N Y -|N Y -|N Y -|N Y -| resulting L F S F L S S L F F L S L F S S L F S L F L F S F L S profile: L = lines; F = function; S = skip, add time to caller The second, third and fourth columns are interpreted as follows: N means that the option not used for a particular module or function. Y means that the option was used for a particular module or function. The - means that a particular module or function was preceded by a minus. Ex: tcprofru -M-module1 -Amod1fn2 Refering to the preceding table: The "resulting profile" line gives the effect for a particular function. the effect on functions mod1fn1 and mod1fn2 in module1 is: No -F option therefore, select -F column 1; -M- specified for module1 therefore, select -M column 3; for mod1fn1 -A was not specified therefore, select -A column 7 which results in this function not appearing in the report and its time being added to the caller. On the other hand, mod1fn2 is specified by a -A option which selects -A column 8 and results in each line of the function being profiled. APPENDIX B - 15 Ex: tcprofru -F- -Afn1 -M-mod2 tstprg The -F- option turns off profiling of all functions in the program. The -A and -M options then modify this action as follows: Profile by lines function fn1, one line per function will be profiled for the functions in module mod2, all other function times are included with the caller's. -PnnF&|T Only print lines whose percentage exceeds nn. F = Line's percentage of function: T = Line's percentage of total. If both are specified the line is printed if either percentage exceeds the nn value. If both are to be the same percentage it may be specified as P20FT. Ex: tcprofru -P50F -p10t tstprg Only print lines, which are 50% or more of the function or are 10% or more of the total. -I[-] Include C source in the report. This option interleaves the C source code with the report. For each C module encountered in the .MAP file a standard DOS path search is made. If the source code is found it is printed with the report. The -I- option can be used to override a -I specified in a response file (-@ option). -@file-name Merge the named file with the command line options. When the @ option is encountered on the command line the file is opened and each line within the file is treated as one profiler command line parameter. The response file cannot be used to specify the program to be profiled or it's options. The options must be in the same sequence as they would appear on the command line. An @ option may not be contained within the file. Multiple @ options may be used on the command line. Ex: tcprofru -@op.dat -Aadd_fn tstprg APPENDIX B - 16 APPENDIX C PROBLEM REPORT/SUGGESTIONS TCPROF version 1.5 From: Name: ___________________________________ Company: ___________________________________ Address: ___________________________________ City, State, Zip: ___________________________________ Received program from: _____________________________________ Hardware: _________________________________________ Operating system: ________________________ version: ______ Resident programs: _________________________________________ __________________________________________ Error message (if problem report): _________________________________________________________ Description of error or suggestions: If appropriate and available, printed output of the profile report and a listing of the program or a fragment illustrating the problem would be helpful. Mail to: Mr. Andre Nancoz 349 Quincy Ave. Bronx, NY 10465 REGISTERED USER ORDER FORM Mail to: Mr. Andre Nancoz 349 Quincy Ave. Bronx, NY 10465 .......................................................... Order for TC profiler registered users version (TCPROFRU): Name: __________________________________________ Company: __________________________________________ Address: __________________________________________ City, State, Zip: __________________________________________ Received TCPROF program from: __________________________________________________________ Hardware: _________________________________________ Operating system: ________________________ version: ______ Send: ___ copies of 5 1/4" DSDD at $20.00 each ________ ___ copies of 3 1/2" DSDD at $25.00 each ________ Total enclosed. $________ Checks only. U. S. funds. NO COD.