Why is this code 6.5x slower with optimizations enabled?Unit Testing C CodeWith arrays, why is it the case that a[5] == 5[a]?Why doesn't GCC optimize a*a*a*a*a*a to (a*a*a)*(a*a*a)?Why are elementwise additions much faster in separate loops than in a combined loop?What is “:-!!” in C code?Why is my program slow when looping over exactly 8192 elements?Obfuscated C Code Contest 2006. Please explain sykes2.cWhy does the C preprocessor interpret the word “linux” as the constant “1”?Why does GCC generate 15-20% faster code if I optimize for size instead of speed?How is the linking done for string functions in C?

My colleague's body is amazing

What is GPS' 19 year rollover and does it present a cybersecurity issue?

Example of a relative pronoun

How does one intimidate enemies without having the capacity for violence?

Is it possible to make sharp wind that can cut stuff from afar?

Can I make popcorn with any corn?

Can you lasso down a wizard who is using the Levitate spell?

N.B. ligature in Latex

New order #4: World

Concept of linear mappings are confusing me

Do airline pilots ever risk not hearing communication directed to them specifically, from traffic controllers?

Is there a minimum number of transactions in a block?

Does the radius of the Spirit Guardians spell depend on the size of the caster?

least quadratic residue under GRH: an EXPLICIT bound

Infinite past with a beginning?

What makes Graph invariants so useful/important?

Finding files for which a command fails

What is the offset in a seaplane's hull?

If Manufacturer spice model and Datasheet give different values which should I use?

When blogging recipes, how can I support both readers who want the narrative/journey and ones who want the printer-friendly recipe?

Shell script can be run only with sh command

Are tax years 2016 & 2017 back taxes deductible for tax year 2018?

How old can references or sources in a thesis be?

Why has Russell's definition of numbers using equivalence classes been finally abandoned? ( If it has actually been abandoned).



Why is this code 6.5x slower with optimizations enabled?


Unit Testing C CodeWith arrays, why is it the case that a[5] == 5[a]?Why doesn't GCC optimize a*a*a*a*a*a to (a*a*a)*(a*a*a)?Why are elementwise additions much faster in separate loops than in a combined loop?What is “:-!!” in C code?Why is my program slow when looping over exactly 8192 elements?Obfuscated C Code Contest 2006. Please explain sykes2.cWhy does the C preprocessor interpret the word “linux” as the constant “1”?Why does GCC generate 15-20% faster code if I optimize for size instead of speed?How is the linking done for string functions in C?






.everyoneloves__top-leaderboard:empty,.everyoneloves__mid-leaderboard:empty,.everyoneloves__bot-mid-leaderboard:empty height:90px;width:728px;box-sizing:border-box;








23















I wanted to benchmark glibc's strlen function for some reason and found out it apparently performs much slower with optimizations enabled in GCC and I have no idea why.



Here's my code:



#include <time.h>
#include <string.h>
#include <stdlib.h>
#include <stdio.h>

int main()
char *s = calloc(1 << 20, 1);
memset(s, 65, 1000000);
clock_t start = clock();
for (int i = 0; i < 128; ++i)
s[strlen(s)] = 'A';

clock_t end = clock();
printf("%lldn", (long long)(end-start));
return 0;



On my machine it outputs:



$ gcc test.c && ./a.out
13336
$ gcc -O1 test.c && ./a.out
199004
$ gcc -O2 test.c && ./a.out
83415
$ gcc -O3 test.c && ./a.out
83415


Somehow, enabling optimizations causes it to execute longer.









share
























  • With gcc-8.2 debug version takes 51334, release 8246. Release compiler options -O3 -march=native -DNDEBUG

    – Maxim Egorushkin
    9 hours ago







  • 1





    Please report it to gcc's bugzilla.

    – Marc Glisse
    8 hours ago






  • 1





    Using -fno-builtin makes the problem go away. So presumably the issue is that in this particular instance, GCC's builtin strlen is slower than the library's.

    – David Schwartz
    8 hours ago











  • It is generating repnz scasb for strlen at -O1.

    – Marc Glisse
    8 hours ago







  • 1





    @MarcGlisse It's already been filed: gcc.gnu.org/bugzilla/show_bug.cgi?id=88809

    – Justin
    3 hours ago

















23















I wanted to benchmark glibc's strlen function for some reason and found out it apparently performs much slower with optimizations enabled in GCC and I have no idea why.



Here's my code:



#include <time.h>
#include <string.h>
#include <stdlib.h>
#include <stdio.h>

int main()
char *s = calloc(1 << 20, 1);
memset(s, 65, 1000000);
clock_t start = clock();
for (int i = 0; i < 128; ++i)
s[strlen(s)] = 'A';

clock_t end = clock();
printf("%lldn", (long long)(end-start));
return 0;



On my machine it outputs:



$ gcc test.c && ./a.out
13336
$ gcc -O1 test.c && ./a.out
199004
$ gcc -O2 test.c && ./a.out
83415
$ gcc -O3 test.c && ./a.out
83415


Somehow, enabling optimizations causes it to execute longer.









share
























  • With gcc-8.2 debug version takes 51334, release 8246. Release compiler options -O3 -march=native -DNDEBUG

    – Maxim Egorushkin
    9 hours ago







  • 1





    Please report it to gcc's bugzilla.

    – Marc Glisse
    8 hours ago






  • 1





    Using -fno-builtin makes the problem go away. So presumably the issue is that in this particular instance, GCC's builtin strlen is slower than the library's.

    – David Schwartz
    8 hours ago











  • It is generating repnz scasb for strlen at -O1.

    – Marc Glisse
    8 hours ago







  • 1





    @MarcGlisse It's already been filed: gcc.gnu.org/bugzilla/show_bug.cgi?id=88809

    – Justin
    3 hours ago













23












23








23


1






I wanted to benchmark glibc's strlen function for some reason and found out it apparently performs much slower with optimizations enabled in GCC and I have no idea why.



Here's my code:



#include <time.h>
#include <string.h>
#include <stdlib.h>
#include <stdio.h>

int main()
char *s = calloc(1 << 20, 1);
memset(s, 65, 1000000);
clock_t start = clock();
for (int i = 0; i < 128; ++i)
s[strlen(s)] = 'A';

clock_t end = clock();
printf("%lldn", (long long)(end-start));
return 0;



On my machine it outputs:



$ gcc test.c && ./a.out
13336
$ gcc -O1 test.c && ./a.out
199004
$ gcc -O2 test.c && ./a.out
83415
$ gcc -O3 test.c && ./a.out
83415


Somehow, enabling optimizations causes it to execute longer.









share
















I wanted to benchmark glibc's strlen function for some reason and found out it apparently performs much slower with optimizations enabled in GCC and I have no idea why.



Here's my code:



#include <time.h>
#include <string.h>
#include <stdlib.h>
#include <stdio.h>

int main()
char *s = calloc(1 << 20, 1);
memset(s, 65, 1000000);
clock_t start = clock();
for (int i = 0; i < 128; ++i)
s[strlen(s)] = 'A';

clock_t end = clock();
printf("%lldn", (long long)(end-start));
return 0;



On my machine it outputs:



$ gcc test.c && ./a.out
13336
$ gcc -O1 test.c && ./a.out
199004
$ gcc -O2 test.c && ./a.out
83415
$ gcc -O3 test.c && ./a.out
83415


Somehow, enabling optimizations causes it to execute longer.







c gcc glibc





share














share












share



share








edited 9 hours ago









Fei Xiang

2,1634822




2,1634822










asked 9 hours ago









TsarNTsarN

11816




11816












  • With gcc-8.2 debug version takes 51334, release 8246. Release compiler options -O3 -march=native -DNDEBUG

    – Maxim Egorushkin
    9 hours ago







  • 1





    Please report it to gcc's bugzilla.

    – Marc Glisse
    8 hours ago






  • 1





    Using -fno-builtin makes the problem go away. So presumably the issue is that in this particular instance, GCC's builtin strlen is slower than the library's.

    – David Schwartz
    8 hours ago











  • It is generating repnz scasb for strlen at -O1.

    – Marc Glisse
    8 hours ago







  • 1





    @MarcGlisse It's already been filed: gcc.gnu.org/bugzilla/show_bug.cgi?id=88809

    – Justin
    3 hours ago

















  • With gcc-8.2 debug version takes 51334, release 8246. Release compiler options -O3 -march=native -DNDEBUG

    – Maxim Egorushkin
    9 hours ago







  • 1





    Please report it to gcc's bugzilla.

    – Marc Glisse
    8 hours ago






  • 1





    Using -fno-builtin makes the problem go away. So presumably the issue is that in this particular instance, GCC's builtin strlen is slower than the library's.

    – David Schwartz
    8 hours ago











  • It is generating repnz scasb for strlen at -O1.

    – Marc Glisse
    8 hours ago







  • 1





    @MarcGlisse It's already been filed: gcc.gnu.org/bugzilla/show_bug.cgi?id=88809

    – Justin
    3 hours ago
















With gcc-8.2 debug version takes 51334, release 8246. Release compiler options -O3 -march=native -DNDEBUG

– Maxim Egorushkin
9 hours ago






With gcc-8.2 debug version takes 51334, release 8246. Release compiler options -O3 -march=native -DNDEBUG

– Maxim Egorushkin
9 hours ago





1




1





Please report it to gcc's bugzilla.

– Marc Glisse
8 hours ago





Please report it to gcc's bugzilla.

– Marc Glisse
8 hours ago




1




1





Using -fno-builtin makes the problem go away. So presumably the issue is that in this particular instance, GCC's builtin strlen is slower than the library's.

– David Schwartz
8 hours ago





Using -fno-builtin makes the problem go away. So presumably the issue is that in this particular instance, GCC's builtin strlen is slower than the library's.

– David Schwartz
8 hours ago













It is generating repnz scasb for strlen at -O1.

– Marc Glisse
8 hours ago






It is generating repnz scasb for strlen at -O1.

– Marc Glisse
8 hours ago





1




1





@MarcGlisse It's already been filed: gcc.gnu.org/bugzilla/show_bug.cgi?id=88809

– Justin
3 hours ago





@MarcGlisse It's already been filed: gcc.gnu.org/bugzilla/show_bug.cgi?id=88809

– Justin
3 hours ago












1 Answer
1






active

oldest

votes


















21














Testing your code on Godbolt's Compiler Explorer provides this explanation:



  • at -O0 or without optimisations, the generated code call the C library function strlen

  • at -O1 the generated code uses a simple inline expansion using a rep scasb instruction.

  • at -O2 and above, the generated code uses a more elaborate inline expansion.

Benchmarking your code repeatedly shows a substantial variation from one run to another, but increasing the number of iterations shows that:



  • the -O1 code is much slower than the C library implementation: 32240 vs 3090

  • the -O2 code is faster than the -O1 but still substantially slower than the C ibrary code: 8570 vs 3090.

This behavior is specific to gcc and the glibc. The same test on OS/X with clang and Apple's Libc does not show a significant difference, which is not a surprise as Godbolt shows that clang generates a call to the C library strlen at all optimisation levels.



This could be considered a bug in gcc/glibc but more extensive benchmarking might show that the overhead of calling strlen has a more important impact than the lack of performance of the inline code for small strings. The strings on which you benchmark are uncommonly large, so focusing the benchmark on ultra-long strings might not give meaningful results.



I updated the benchmark for smaller strings and it shows similar performance for string lengths varying from 0 to 100 at -O0 and -O2 but still a much worse performance at -O1, 3 times slower.



Here is the updated code:



#include <stdlib.h>
#include <string.h>
#include <time.h>

void benchmark(int repeat, int minlen, int maxlen)
char *s = malloc(maxlen + 1);
memset(s, 'A', minlen);
long long bytes = 0, calls = 0;
clock_t clk = clock();
for (int n = 0; n < repeat; n++)
for (int i = minlen; i < maxlen; ++i)
bytes += i + 1;
calls += 1;
s[i] = '';
s[strlen(s)] = 'A';


clk = clock() - clk;
free(s);
double avglen = (minlen + maxlen - 1) / 2.0;
double ns = (double)clk * 1e9 / CLOCKS_PER_SEC;
printf("average length %7.0f -> avg time: %7.3f ns/byte, %7.3f ns/calln",
avglen, ns / bytes, ns / calls);


int main()
benchmark(10000000, 0, 1);
benchmark(1000000, 0, 10);
benchmark(1000000, 5, 15);
benchmark(100000, 0, 100);
benchmark(100000, 50, 150);
benchmark(10000, 0, 1000);
benchmark(10000, 500, 1500);
benchmark(1000, 0, 10000);
benchmark(1000, 5000, 15000);
benchmark(100, 1000000 - 50, 1000000 + 50);
return 0;



Here is the output:




chqrlie> gcc -std=c99 -O0 benchstrlen.c && ./a.out
average length 0 -> avg time: 14.000 ns/byte, 14.000 ns/call
average length 4 -> avg time: 2.364 ns/byte, 13.000 ns/call
average length 10 -> avg time: 1.238 ns/byte, 13.000 ns/call
average length 50 -> avg time: 0.317 ns/byte, 16.000 ns/call
average length 100 -> avg time: 0.169 ns/byte, 17.000 ns/call
average length 500 -> avg time: 0.074 ns/byte, 37.000 ns/call
average length 1000 -> avg time: 0.068 ns/byte, 68.000 ns/call
average length 5000 -> avg time: 0.064 ns/byte, 318.000 ns/call
average length 10000 -> avg time: 0.062 ns/byte, 622.000 ns/call
average length 1000000 -> avg time: 0.062 ns/byte, 62000.000 ns/call
chqrlie> gcc -std=c99 -O1 benchstrlen.c && ./a.out
average length 0 -> avg time: 20.000 ns/byte, 20.000 ns/call
average length 4 -> avg time: 3.818 ns/byte, 21.000 ns/call
average length 10 -> avg time: 2.190 ns/byte, 23.000 ns/call
average length 50 -> avg time: 0.990 ns/byte, 50.000 ns/call
average length 100 -> avg time: 0.816 ns/byte, 82.000 ns/call
average length 500 -> avg time: 0.679 ns/byte, 340.000 ns/call
average length 1000 -> avg time: 0.664 ns/byte, 664.000 ns/call
average length 5000 -> avg time: 0.651 ns/byte, 3254.000 ns/call
average length 10000 -> avg time: 0.649 ns/byte, 6491.000 ns/call
average length 1000000 -> avg time: 0.648 ns/byte, 648000.000 ns/call
chqrlie> gcc -std=c99 -O2 benchstrlen.c && ./a.out
average length 0 -> avg time: 10.000 ns/byte, 10.000 ns/call
average length 4 -> avg time: 2.000 ns/byte, 11.000 ns/call
average length 10 -> avg time: 1.048 ns/byte, 11.000 ns/call
average length 50 -> avg time: 0.337 ns/byte, 17.000 ns/call
average length 100 -> avg time: 0.299 ns/byte, 30.000 ns/call
average length 500 -> avg time: 0.202 ns/byte, 101.000 ns/call
average length 1000 -> avg time: 0.188 ns/byte, 188.000 ns/call
average length 5000 -> avg time: 0.174 ns/byte, 868.000 ns/call
average length 10000 -> avg time: 0.172 ns/byte, 1716.000 ns/call
average length 1000000 -> avg time: 0.172 ns/byte, 172000.000 ns/call





share|improve this answer

























  • Wouldn't it still be better for the inlined version to use the same optimizations as the library strlen, giving the best of both worlds?

    – Daniel H
    8 hours ago






  • 1





    It would, but the hand optimized version in the C library might be larger and more complicated to inline. I have not looked into this recently, but there used to be a mix of complex platform specific macros in <string.h> and hard coded optimisations in the gcc code generator. Definitely still room for improvement on intel targets.

    – chqrlie
    8 hours ago











  • Does it change if you use -march=native -mtune=native?

    – Deduplicator
    7 hours ago











  • Note that the GNU C library function for strlen() is likely optimised for extremely large strings (that no sane programmer will care about) at the expense of performance for small strings (that are extremely common); and the optimisations done by the library version should never be done. The problem here is that the OP's code doesn't keep track of the string's length itself (e.g. with an int len; variable) and should not have used strlen() at all, making the code so bad for performance that "optimised for something no sane programmer would care about" actually helped.

    – Brendan
    7 hours ago







  • 1





    @chqrlie: I'd also say that this is partly a symptom of a larger problem - code in libraries can't be optimised for any specific case and therefore must be "un-optimal" for some cases. To work around that it would've been nice if there was a strlen_small() and a separate strlen_large(), but there isn't.

    – Brendan
    5 hours ago











Your Answer






StackExchange.ifUsing("editor", function ()
StackExchange.using("externalEditor", function ()
StackExchange.using("snippets", function ()
StackExchange.snippets.init();
);
);
, "code-snippets");

StackExchange.ready(function()
var channelOptions =
tags: "".split(" "),
id: "1"
;
initTagRenderer("".split(" "), "".split(" "), channelOptions);

StackExchange.using("externalEditor", function()
// Have to fire editor after snippets, if snippets enabled
if (StackExchange.settings.snippets.snippetsEnabled)
StackExchange.using("snippets", function()
createEditor();
);

else
createEditor();

);

function createEditor()
StackExchange.prepareEditor(
heartbeatType: 'answer',
autoActivateHeartbeat: false,
convertImagesToLinks: true,
noModals: true,
showLowRepImageUploadWarning: true,
reputationToPostImages: 10,
bindNavPrevention: true,
postfix: "",
imageUploader:
brandingHtml: "Powered by u003ca class="icon-imgur-white" href="https://imgur.com/"u003eu003c/au003e",
contentPolicyHtml: "User contributions licensed under u003ca href="https://creativecommons.org/licenses/by-sa/3.0/"u003ecc by-sa 3.0 with attribution requiredu003c/au003e u003ca href="https://stackoverflow.com/legal/content-policy"u003e(content policy)u003c/au003e",
allowUrls: true
,
onDemand: true,
discardSelector: ".discard-answer"
,immediatelyShowMarkdownHelp:true
);



);













draft saved

draft discarded


















StackExchange.ready(
function ()
StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fstackoverflow.com%2fquestions%2f55563598%2fwhy-is-this-code-6-5x-slower-with-optimizations-enabled%23new-answer', 'question_page');

);

Post as a guest















Required, but never shown

























1 Answer
1






active

oldest

votes








1 Answer
1






active

oldest

votes









active

oldest

votes






active

oldest

votes









21














Testing your code on Godbolt's Compiler Explorer provides this explanation:



  • at -O0 or without optimisations, the generated code call the C library function strlen

  • at -O1 the generated code uses a simple inline expansion using a rep scasb instruction.

  • at -O2 and above, the generated code uses a more elaborate inline expansion.

Benchmarking your code repeatedly shows a substantial variation from one run to another, but increasing the number of iterations shows that:



  • the -O1 code is much slower than the C library implementation: 32240 vs 3090

  • the -O2 code is faster than the -O1 but still substantially slower than the C ibrary code: 8570 vs 3090.

This behavior is specific to gcc and the glibc. The same test on OS/X with clang and Apple's Libc does not show a significant difference, which is not a surprise as Godbolt shows that clang generates a call to the C library strlen at all optimisation levels.



This could be considered a bug in gcc/glibc but more extensive benchmarking might show that the overhead of calling strlen has a more important impact than the lack of performance of the inline code for small strings. The strings on which you benchmark are uncommonly large, so focusing the benchmark on ultra-long strings might not give meaningful results.



I updated the benchmark for smaller strings and it shows similar performance for string lengths varying from 0 to 100 at -O0 and -O2 but still a much worse performance at -O1, 3 times slower.



Here is the updated code:



#include <stdlib.h>
#include <string.h>
#include <time.h>

void benchmark(int repeat, int minlen, int maxlen)
char *s = malloc(maxlen + 1);
memset(s, 'A', minlen);
long long bytes = 0, calls = 0;
clock_t clk = clock();
for (int n = 0; n < repeat; n++)
for (int i = minlen; i < maxlen; ++i)
bytes += i + 1;
calls += 1;
s[i] = '';
s[strlen(s)] = 'A';


clk = clock() - clk;
free(s);
double avglen = (minlen + maxlen - 1) / 2.0;
double ns = (double)clk * 1e9 / CLOCKS_PER_SEC;
printf("average length %7.0f -> avg time: %7.3f ns/byte, %7.3f ns/calln",
avglen, ns / bytes, ns / calls);


int main()
benchmark(10000000, 0, 1);
benchmark(1000000, 0, 10);
benchmark(1000000, 5, 15);
benchmark(100000, 0, 100);
benchmark(100000, 50, 150);
benchmark(10000, 0, 1000);
benchmark(10000, 500, 1500);
benchmark(1000, 0, 10000);
benchmark(1000, 5000, 15000);
benchmark(100, 1000000 - 50, 1000000 + 50);
return 0;



Here is the output:




chqrlie> gcc -std=c99 -O0 benchstrlen.c && ./a.out
average length 0 -> avg time: 14.000 ns/byte, 14.000 ns/call
average length 4 -> avg time: 2.364 ns/byte, 13.000 ns/call
average length 10 -> avg time: 1.238 ns/byte, 13.000 ns/call
average length 50 -> avg time: 0.317 ns/byte, 16.000 ns/call
average length 100 -> avg time: 0.169 ns/byte, 17.000 ns/call
average length 500 -> avg time: 0.074 ns/byte, 37.000 ns/call
average length 1000 -> avg time: 0.068 ns/byte, 68.000 ns/call
average length 5000 -> avg time: 0.064 ns/byte, 318.000 ns/call
average length 10000 -> avg time: 0.062 ns/byte, 622.000 ns/call
average length 1000000 -> avg time: 0.062 ns/byte, 62000.000 ns/call
chqrlie> gcc -std=c99 -O1 benchstrlen.c && ./a.out
average length 0 -> avg time: 20.000 ns/byte, 20.000 ns/call
average length 4 -> avg time: 3.818 ns/byte, 21.000 ns/call
average length 10 -> avg time: 2.190 ns/byte, 23.000 ns/call
average length 50 -> avg time: 0.990 ns/byte, 50.000 ns/call
average length 100 -> avg time: 0.816 ns/byte, 82.000 ns/call
average length 500 -> avg time: 0.679 ns/byte, 340.000 ns/call
average length 1000 -> avg time: 0.664 ns/byte, 664.000 ns/call
average length 5000 -> avg time: 0.651 ns/byte, 3254.000 ns/call
average length 10000 -> avg time: 0.649 ns/byte, 6491.000 ns/call
average length 1000000 -> avg time: 0.648 ns/byte, 648000.000 ns/call
chqrlie> gcc -std=c99 -O2 benchstrlen.c && ./a.out
average length 0 -> avg time: 10.000 ns/byte, 10.000 ns/call
average length 4 -> avg time: 2.000 ns/byte, 11.000 ns/call
average length 10 -> avg time: 1.048 ns/byte, 11.000 ns/call
average length 50 -> avg time: 0.337 ns/byte, 17.000 ns/call
average length 100 -> avg time: 0.299 ns/byte, 30.000 ns/call
average length 500 -> avg time: 0.202 ns/byte, 101.000 ns/call
average length 1000 -> avg time: 0.188 ns/byte, 188.000 ns/call
average length 5000 -> avg time: 0.174 ns/byte, 868.000 ns/call
average length 10000 -> avg time: 0.172 ns/byte, 1716.000 ns/call
average length 1000000 -> avg time: 0.172 ns/byte, 172000.000 ns/call





share|improve this answer

























  • Wouldn't it still be better for the inlined version to use the same optimizations as the library strlen, giving the best of both worlds?

    – Daniel H
    8 hours ago






  • 1





    It would, but the hand optimized version in the C library might be larger and more complicated to inline. I have not looked into this recently, but there used to be a mix of complex platform specific macros in <string.h> and hard coded optimisations in the gcc code generator. Definitely still room for improvement on intel targets.

    – chqrlie
    8 hours ago











  • Does it change if you use -march=native -mtune=native?

    – Deduplicator
    7 hours ago











  • Note that the GNU C library function for strlen() is likely optimised for extremely large strings (that no sane programmer will care about) at the expense of performance for small strings (that are extremely common); and the optimisations done by the library version should never be done. The problem here is that the OP's code doesn't keep track of the string's length itself (e.g. with an int len; variable) and should not have used strlen() at all, making the code so bad for performance that "optimised for something no sane programmer would care about" actually helped.

    – Brendan
    7 hours ago







  • 1





    @chqrlie: I'd also say that this is partly a symptom of a larger problem - code in libraries can't be optimised for any specific case and therefore must be "un-optimal" for some cases. To work around that it would've been nice if there was a strlen_small() and a separate strlen_large(), but there isn't.

    – Brendan
    5 hours ago















21














Testing your code on Godbolt's Compiler Explorer provides this explanation:



  • at -O0 or without optimisations, the generated code call the C library function strlen

  • at -O1 the generated code uses a simple inline expansion using a rep scasb instruction.

  • at -O2 and above, the generated code uses a more elaborate inline expansion.

Benchmarking your code repeatedly shows a substantial variation from one run to another, but increasing the number of iterations shows that:



  • the -O1 code is much slower than the C library implementation: 32240 vs 3090

  • the -O2 code is faster than the -O1 but still substantially slower than the C ibrary code: 8570 vs 3090.

This behavior is specific to gcc and the glibc. The same test on OS/X with clang and Apple's Libc does not show a significant difference, which is not a surprise as Godbolt shows that clang generates a call to the C library strlen at all optimisation levels.



This could be considered a bug in gcc/glibc but more extensive benchmarking might show that the overhead of calling strlen has a more important impact than the lack of performance of the inline code for small strings. The strings on which you benchmark are uncommonly large, so focusing the benchmark on ultra-long strings might not give meaningful results.



I updated the benchmark for smaller strings and it shows similar performance for string lengths varying from 0 to 100 at -O0 and -O2 but still a much worse performance at -O1, 3 times slower.



Here is the updated code:



#include <stdlib.h>
#include <string.h>
#include <time.h>

void benchmark(int repeat, int minlen, int maxlen)
char *s = malloc(maxlen + 1);
memset(s, 'A', minlen);
long long bytes = 0, calls = 0;
clock_t clk = clock();
for (int n = 0; n < repeat; n++)
for (int i = minlen; i < maxlen; ++i)
bytes += i + 1;
calls += 1;
s[i] = '';
s[strlen(s)] = 'A';


clk = clock() - clk;
free(s);
double avglen = (minlen + maxlen - 1) / 2.0;
double ns = (double)clk * 1e9 / CLOCKS_PER_SEC;
printf("average length %7.0f -> avg time: %7.3f ns/byte, %7.3f ns/calln",
avglen, ns / bytes, ns / calls);


int main()
benchmark(10000000, 0, 1);
benchmark(1000000, 0, 10);
benchmark(1000000, 5, 15);
benchmark(100000, 0, 100);
benchmark(100000, 50, 150);
benchmark(10000, 0, 1000);
benchmark(10000, 500, 1500);
benchmark(1000, 0, 10000);
benchmark(1000, 5000, 15000);
benchmark(100, 1000000 - 50, 1000000 + 50);
return 0;



Here is the output:




chqrlie> gcc -std=c99 -O0 benchstrlen.c && ./a.out
average length 0 -> avg time: 14.000 ns/byte, 14.000 ns/call
average length 4 -> avg time: 2.364 ns/byte, 13.000 ns/call
average length 10 -> avg time: 1.238 ns/byte, 13.000 ns/call
average length 50 -> avg time: 0.317 ns/byte, 16.000 ns/call
average length 100 -> avg time: 0.169 ns/byte, 17.000 ns/call
average length 500 -> avg time: 0.074 ns/byte, 37.000 ns/call
average length 1000 -> avg time: 0.068 ns/byte, 68.000 ns/call
average length 5000 -> avg time: 0.064 ns/byte, 318.000 ns/call
average length 10000 -> avg time: 0.062 ns/byte, 622.000 ns/call
average length 1000000 -> avg time: 0.062 ns/byte, 62000.000 ns/call
chqrlie> gcc -std=c99 -O1 benchstrlen.c && ./a.out
average length 0 -> avg time: 20.000 ns/byte, 20.000 ns/call
average length 4 -> avg time: 3.818 ns/byte, 21.000 ns/call
average length 10 -> avg time: 2.190 ns/byte, 23.000 ns/call
average length 50 -> avg time: 0.990 ns/byte, 50.000 ns/call
average length 100 -> avg time: 0.816 ns/byte, 82.000 ns/call
average length 500 -> avg time: 0.679 ns/byte, 340.000 ns/call
average length 1000 -> avg time: 0.664 ns/byte, 664.000 ns/call
average length 5000 -> avg time: 0.651 ns/byte, 3254.000 ns/call
average length 10000 -> avg time: 0.649 ns/byte, 6491.000 ns/call
average length 1000000 -> avg time: 0.648 ns/byte, 648000.000 ns/call
chqrlie> gcc -std=c99 -O2 benchstrlen.c && ./a.out
average length 0 -> avg time: 10.000 ns/byte, 10.000 ns/call
average length 4 -> avg time: 2.000 ns/byte, 11.000 ns/call
average length 10 -> avg time: 1.048 ns/byte, 11.000 ns/call
average length 50 -> avg time: 0.337 ns/byte, 17.000 ns/call
average length 100 -> avg time: 0.299 ns/byte, 30.000 ns/call
average length 500 -> avg time: 0.202 ns/byte, 101.000 ns/call
average length 1000 -> avg time: 0.188 ns/byte, 188.000 ns/call
average length 5000 -> avg time: 0.174 ns/byte, 868.000 ns/call
average length 10000 -> avg time: 0.172 ns/byte, 1716.000 ns/call
average length 1000000 -> avg time: 0.172 ns/byte, 172000.000 ns/call





share|improve this answer

























  • Wouldn't it still be better for the inlined version to use the same optimizations as the library strlen, giving the best of both worlds?

    – Daniel H
    8 hours ago






  • 1





    It would, but the hand optimized version in the C library might be larger and more complicated to inline. I have not looked into this recently, but there used to be a mix of complex platform specific macros in <string.h> and hard coded optimisations in the gcc code generator. Definitely still room for improvement on intel targets.

    – chqrlie
    8 hours ago











  • Does it change if you use -march=native -mtune=native?

    – Deduplicator
    7 hours ago











  • Note that the GNU C library function for strlen() is likely optimised for extremely large strings (that no sane programmer will care about) at the expense of performance for small strings (that are extremely common); and the optimisations done by the library version should never be done. The problem here is that the OP's code doesn't keep track of the string's length itself (e.g. with an int len; variable) and should not have used strlen() at all, making the code so bad for performance that "optimised for something no sane programmer would care about" actually helped.

    – Brendan
    7 hours ago







  • 1





    @chqrlie: I'd also say that this is partly a symptom of a larger problem - code in libraries can't be optimised for any specific case and therefore must be "un-optimal" for some cases. To work around that it would've been nice if there was a strlen_small() and a separate strlen_large(), but there isn't.

    – Brendan
    5 hours ago













21












21








21







Testing your code on Godbolt's Compiler Explorer provides this explanation:



  • at -O0 or without optimisations, the generated code call the C library function strlen

  • at -O1 the generated code uses a simple inline expansion using a rep scasb instruction.

  • at -O2 and above, the generated code uses a more elaborate inline expansion.

Benchmarking your code repeatedly shows a substantial variation from one run to another, but increasing the number of iterations shows that:



  • the -O1 code is much slower than the C library implementation: 32240 vs 3090

  • the -O2 code is faster than the -O1 but still substantially slower than the C ibrary code: 8570 vs 3090.

This behavior is specific to gcc and the glibc. The same test on OS/X with clang and Apple's Libc does not show a significant difference, which is not a surprise as Godbolt shows that clang generates a call to the C library strlen at all optimisation levels.



This could be considered a bug in gcc/glibc but more extensive benchmarking might show that the overhead of calling strlen has a more important impact than the lack of performance of the inline code for small strings. The strings on which you benchmark are uncommonly large, so focusing the benchmark on ultra-long strings might not give meaningful results.



I updated the benchmark for smaller strings and it shows similar performance for string lengths varying from 0 to 100 at -O0 and -O2 but still a much worse performance at -O1, 3 times slower.



Here is the updated code:



#include <stdlib.h>
#include <string.h>
#include <time.h>

void benchmark(int repeat, int minlen, int maxlen)
char *s = malloc(maxlen + 1);
memset(s, 'A', minlen);
long long bytes = 0, calls = 0;
clock_t clk = clock();
for (int n = 0; n < repeat; n++)
for (int i = minlen; i < maxlen; ++i)
bytes += i + 1;
calls += 1;
s[i] = '';
s[strlen(s)] = 'A';


clk = clock() - clk;
free(s);
double avglen = (minlen + maxlen - 1) / 2.0;
double ns = (double)clk * 1e9 / CLOCKS_PER_SEC;
printf("average length %7.0f -> avg time: %7.3f ns/byte, %7.3f ns/calln",
avglen, ns / bytes, ns / calls);


int main()
benchmark(10000000, 0, 1);
benchmark(1000000, 0, 10);
benchmark(1000000, 5, 15);
benchmark(100000, 0, 100);
benchmark(100000, 50, 150);
benchmark(10000, 0, 1000);
benchmark(10000, 500, 1500);
benchmark(1000, 0, 10000);
benchmark(1000, 5000, 15000);
benchmark(100, 1000000 - 50, 1000000 + 50);
return 0;



Here is the output:




chqrlie> gcc -std=c99 -O0 benchstrlen.c && ./a.out
average length 0 -> avg time: 14.000 ns/byte, 14.000 ns/call
average length 4 -> avg time: 2.364 ns/byte, 13.000 ns/call
average length 10 -> avg time: 1.238 ns/byte, 13.000 ns/call
average length 50 -> avg time: 0.317 ns/byte, 16.000 ns/call
average length 100 -> avg time: 0.169 ns/byte, 17.000 ns/call
average length 500 -> avg time: 0.074 ns/byte, 37.000 ns/call
average length 1000 -> avg time: 0.068 ns/byte, 68.000 ns/call
average length 5000 -> avg time: 0.064 ns/byte, 318.000 ns/call
average length 10000 -> avg time: 0.062 ns/byte, 622.000 ns/call
average length 1000000 -> avg time: 0.062 ns/byte, 62000.000 ns/call
chqrlie> gcc -std=c99 -O1 benchstrlen.c && ./a.out
average length 0 -> avg time: 20.000 ns/byte, 20.000 ns/call
average length 4 -> avg time: 3.818 ns/byte, 21.000 ns/call
average length 10 -> avg time: 2.190 ns/byte, 23.000 ns/call
average length 50 -> avg time: 0.990 ns/byte, 50.000 ns/call
average length 100 -> avg time: 0.816 ns/byte, 82.000 ns/call
average length 500 -> avg time: 0.679 ns/byte, 340.000 ns/call
average length 1000 -> avg time: 0.664 ns/byte, 664.000 ns/call
average length 5000 -> avg time: 0.651 ns/byte, 3254.000 ns/call
average length 10000 -> avg time: 0.649 ns/byte, 6491.000 ns/call
average length 1000000 -> avg time: 0.648 ns/byte, 648000.000 ns/call
chqrlie> gcc -std=c99 -O2 benchstrlen.c && ./a.out
average length 0 -> avg time: 10.000 ns/byte, 10.000 ns/call
average length 4 -> avg time: 2.000 ns/byte, 11.000 ns/call
average length 10 -> avg time: 1.048 ns/byte, 11.000 ns/call
average length 50 -> avg time: 0.337 ns/byte, 17.000 ns/call
average length 100 -> avg time: 0.299 ns/byte, 30.000 ns/call
average length 500 -> avg time: 0.202 ns/byte, 101.000 ns/call
average length 1000 -> avg time: 0.188 ns/byte, 188.000 ns/call
average length 5000 -> avg time: 0.174 ns/byte, 868.000 ns/call
average length 10000 -> avg time: 0.172 ns/byte, 1716.000 ns/call
average length 1000000 -> avg time: 0.172 ns/byte, 172000.000 ns/call





share|improve this answer















Testing your code on Godbolt's Compiler Explorer provides this explanation:



  • at -O0 or without optimisations, the generated code call the C library function strlen

  • at -O1 the generated code uses a simple inline expansion using a rep scasb instruction.

  • at -O2 and above, the generated code uses a more elaborate inline expansion.

Benchmarking your code repeatedly shows a substantial variation from one run to another, but increasing the number of iterations shows that:



  • the -O1 code is much slower than the C library implementation: 32240 vs 3090

  • the -O2 code is faster than the -O1 but still substantially slower than the C ibrary code: 8570 vs 3090.

This behavior is specific to gcc and the glibc. The same test on OS/X with clang and Apple's Libc does not show a significant difference, which is not a surprise as Godbolt shows that clang generates a call to the C library strlen at all optimisation levels.



This could be considered a bug in gcc/glibc but more extensive benchmarking might show that the overhead of calling strlen has a more important impact than the lack of performance of the inline code for small strings. The strings on which you benchmark are uncommonly large, so focusing the benchmark on ultra-long strings might not give meaningful results.



I updated the benchmark for smaller strings and it shows similar performance for string lengths varying from 0 to 100 at -O0 and -O2 but still a much worse performance at -O1, 3 times slower.



Here is the updated code:



#include <stdlib.h>
#include <string.h>
#include <time.h>

void benchmark(int repeat, int minlen, int maxlen)
char *s = malloc(maxlen + 1);
memset(s, 'A', minlen);
long long bytes = 0, calls = 0;
clock_t clk = clock();
for (int n = 0; n < repeat; n++)
for (int i = minlen; i < maxlen; ++i)
bytes += i + 1;
calls += 1;
s[i] = '';
s[strlen(s)] = 'A';


clk = clock() - clk;
free(s);
double avglen = (minlen + maxlen - 1) / 2.0;
double ns = (double)clk * 1e9 / CLOCKS_PER_SEC;
printf("average length %7.0f -> avg time: %7.3f ns/byte, %7.3f ns/calln",
avglen, ns / bytes, ns / calls);


int main()
benchmark(10000000, 0, 1);
benchmark(1000000, 0, 10);
benchmark(1000000, 5, 15);
benchmark(100000, 0, 100);
benchmark(100000, 50, 150);
benchmark(10000, 0, 1000);
benchmark(10000, 500, 1500);
benchmark(1000, 0, 10000);
benchmark(1000, 5000, 15000);
benchmark(100, 1000000 - 50, 1000000 + 50);
return 0;



Here is the output:




chqrlie> gcc -std=c99 -O0 benchstrlen.c && ./a.out
average length 0 -> avg time: 14.000 ns/byte, 14.000 ns/call
average length 4 -> avg time: 2.364 ns/byte, 13.000 ns/call
average length 10 -> avg time: 1.238 ns/byte, 13.000 ns/call
average length 50 -> avg time: 0.317 ns/byte, 16.000 ns/call
average length 100 -> avg time: 0.169 ns/byte, 17.000 ns/call
average length 500 -> avg time: 0.074 ns/byte, 37.000 ns/call
average length 1000 -> avg time: 0.068 ns/byte, 68.000 ns/call
average length 5000 -> avg time: 0.064 ns/byte, 318.000 ns/call
average length 10000 -> avg time: 0.062 ns/byte, 622.000 ns/call
average length 1000000 -> avg time: 0.062 ns/byte, 62000.000 ns/call
chqrlie> gcc -std=c99 -O1 benchstrlen.c && ./a.out
average length 0 -> avg time: 20.000 ns/byte, 20.000 ns/call
average length 4 -> avg time: 3.818 ns/byte, 21.000 ns/call
average length 10 -> avg time: 2.190 ns/byte, 23.000 ns/call
average length 50 -> avg time: 0.990 ns/byte, 50.000 ns/call
average length 100 -> avg time: 0.816 ns/byte, 82.000 ns/call
average length 500 -> avg time: 0.679 ns/byte, 340.000 ns/call
average length 1000 -> avg time: 0.664 ns/byte, 664.000 ns/call
average length 5000 -> avg time: 0.651 ns/byte, 3254.000 ns/call
average length 10000 -> avg time: 0.649 ns/byte, 6491.000 ns/call
average length 1000000 -> avg time: 0.648 ns/byte, 648000.000 ns/call
chqrlie> gcc -std=c99 -O2 benchstrlen.c && ./a.out
average length 0 -> avg time: 10.000 ns/byte, 10.000 ns/call
average length 4 -> avg time: 2.000 ns/byte, 11.000 ns/call
average length 10 -> avg time: 1.048 ns/byte, 11.000 ns/call
average length 50 -> avg time: 0.337 ns/byte, 17.000 ns/call
average length 100 -> avg time: 0.299 ns/byte, 30.000 ns/call
average length 500 -> avg time: 0.202 ns/byte, 101.000 ns/call
average length 1000 -> avg time: 0.188 ns/byte, 188.000 ns/call
average length 5000 -> avg time: 0.174 ns/byte, 868.000 ns/call
average length 10000 -> avg time: 0.172 ns/byte, 1716.000 ns/call
average length 1000000 -> avg time: 0.172 ns/byte, 172000.000 ns/call






share|improve this answer














share|improve this answer



share|improve this answer








edited 7 hours ago

























answered 8 hours ago









chqrliechqrlie

63.1k849108




63.1k849108












  • Wouldn't it still be better for the inlined version to use the same optimizations as the library strlen, giving the best of both worlds?

    – Daniel H
    8 hours ago






  • 1





    It would, but the hand optimized version in the C library might be larger and more complicated to inline. I have not looked into this recently, but there used to be a mix of complex platform specific macros in <string.h> and hard coded optimisations in the gcc code generator. Definitely still room for improvement on intel targets.

    – chqrlie
    8 hours ago











  • Does it change if you use -march=native -mtune=native?

    – Deduplicator
    7 hours ago











  • Note that the GNU C library function for strlen() is likely optimised for extremely large strings (that no sane programmer will care about) at the expense of performance for small strings (that are extremely common); and the optimisations done by the library version should never be done. The problem here is that the OP's code doesn't keep track of the string's length itself (e.g. with an int len; variable) and should not have used strlen() at all, making the code so bad for performance that "optimised for something no sane programmer would care about" actually helped.

    – Brendan
    7 hours ago







  • 1





    @chqrlie: I'd also say that this is partly a symptom of a larger problem - code in libraries can't be optimised for any specific case and therefore must be "un-optimal" for some cases. To work around that it would've been nice if there was a strlen_small() and a separate strlen_large(), but there isn't.

    – Brendan
    5 hours ago

















  • Wouldn't it still be better for the inlined version to use the same optimizations as the library strlen, giving the best of both worlds?

    – Daniel H
    8 hours ago






  • 1





    It would, but the hand optimized version in the C library might be larger and more complicated to inline. I have not looked into this recently, but there used to be a mix of complex platform specific macros in <string.h> and hard coded optimisations in the gcc code generator. Definitely still room for improvement on intel targets.

    – chqrlie
    8 hours ago











  • Does it change if you use -march=native -mtune=native?

    – Deduplicator
    7 hours ago











  • Note that the GNU C library function for strlen() is likely optimised for extremely large strings (that no sane programmer will care about) at the expense of performance for small strings (that are extremely common); and the optimisations done by the library version should never be done. The problem here is that the OP's code doesn't keep track of the string's length itself (e.g. with an int len; variable) and should not have used strlen() at all, making the code so bad for performance that "optimised for something no sane programmer would care about" actually helped.

    – Brendan
    7 hours ago







  • 1





    @chqrlie: I'd also say that this is partly a symptom of a larger problem - code in libraries can't be optimised for any specific case and therefore must be "un-optimal" for some cases. To work around that it would've been nice if there was a strlen_small() and a separate strlen_large(), but there isn't.

    – Brendan
    5 hours ago
















Wouldn't it still be better for the inlined version to use the same optimizations as the library strlen, giving the best of both worlds?

– Daniel H
8 hours ago





Wouldn't it still be better for the inlined version to use the same optimizations as the library strlen, giving the best of both worlds?

– Daniel H
8 hours ago




1




1





It would, but the hand optimized version in the C library might be larger and more complicated to inline. I have not looked into this recently, but there used to be a mix of complex platform specific macros in <string.h> and hard coded optimisations in the gcc code generator. Definitely still room for improvement on intel targets.

– chqrlie
8 hours ago





It would, but the hand optimized version in the C library might be larger and more complicated to inline. I have not looked into this recently, but there used to be a mix of complex platform specific macros in <string.h> and hard coded optimisations in the gcc code generator. Definitely still room for improvement on intel targets.

– chqrlie
8 hours ago













Does it change if you use -march=native -mtune=native?

– Deduplicator
7 hours ago





Does it change if you use -march=native -mtune=native?

– Deduplicator
7 hours ago













Note that the GNU C library function for strlen() is likely optimised for extremely large strings (that no sane programmer will care about) at the expense of performance for small strings (that are extremely common); and the optimisations done by the library version should never be done. The problem here is that the OP's code doesn't keep track of the string's length itself (e.g. with an int len; variable) and should not have used strlen() at all, making the code so bad for performance that "optimised for something no sane programmer would care about" actually helped.

– Brendan
7 hours ago






Note that the GNU C library function for strlen() is likely optimised for extremely large strings (that no sane programmer will care about) at the expense of performance for small strings (that are extremely common); and the optimisations done by the library version should never be done. The problem here is that the OP's code doesn't keep track of the string's length itself (e.g. with an int len; variable) and should not have used strlen() at all, making the code so bad for performance that "optimised for something no sane programmer would care about" actually helped.

– Brendan
7 hours ago





1




1





@chqrlie: I'd also say that this is partly a symptom of a larger problem - code in libraries can't be optimised for any specific case and therefore must be "un-optimal" for some cases. To work around that it would've been nice if there was a strlen_small() and a separate strlen_large(), but there isn't.

– Brendan
5 hours ago





@chqrlie: I'd also say that this is partly a symptom of a larger problem - code in libraries can't be optimised for any specific case and therefore must be "un-optimal" for some cases. To work around that it would've been nice if there was a strlen_small() and a separate strlen_large(), but there isn't.

– Brendan
5 hours ago



















draft saved

draft discarded
















































Thanks for contributing an answer to Stack Overflow!


  • Please be sure to answer the question. Provide details and share your research!

But avoid


  • Asking for help, clarification, or responding to other answers.

  • Making statements based on opinion; back them up with references or personal experience.

To learn more, see our tips on writing great answers.




draft saved


draft discarded














StackExchange.ready(
function ()
StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fstackoverflow.com%2fquestions%2f55563598%2fwhy-is-this-code-6-5x-slower-with-optimizations-enabled%23new-answer', 'question_page');

);

Post as a guest















Required, but never shown





















































Required, but never shown














Required, but never shown












Required, but never shown







Required, but never shown

































Required, but never shown














Required, but never shown












Required, but never shown







Required, but never shown







Popular posts from this blog

Can not update quote_id field of “quote_item” table magento 2Magento 2.1 - We can't remove the item. (Shopping Cart doesnt allow us to remove items before becomes empty)Add value for custom quote item attribute using REST apiREST API endpoint v1/carts/cartId/items always returns error messageCorrect way to save entries to databaseHow to remove all associated quote objects of a customer completelyMagento 2 - Save value from custom input field to quote_itemGet quote_item data using quote id and product id filter in Magento 2How to set additional data to quote_item table from controller in Magento 2?What is the purpose of additional_data column in quote_item table in magento2Set Custom Price to Quote item magento2 from controller

Magento 2 disable Secret Key on URL's from terminal The Next CEO of Stack OverflowMagento 2 Shortcut/GUI tool to perform commandline tasks for windowsIn menu add configuration linkMagento oAuth : Generating access token and access secretMagento 2 security key issue in Third-Party API redirect URIPublic actions in admin controllersHow to Disable Cache in Custom WidgetURL Key not changing in Magento 2Product URL Key gets deleted when importing custom options - Magento 2Problem with reindex terminalMagento 2 - bin/magento Commands not working in Cpanel Terminal

Aasi (pallopeli) Navigointivalikko