[Bug c++/13351] New: 10x slower execution speed GCC 3.3 (compared with GCC 3.2)

[Bug c++/13351] New: 10x slower execution speed GCC 3.3 (compared with GCC 3.2)

[cline at parashift dot com]

The execution speed of the following simple program is 10x slower on GCC 3.3 
than on GCC 3.2 (using cygwin; I haven't tested Linux yet).  To reproduce, 
save it and compile with:
  g++ -Wall -pedantic -O -DNDEBUG speedtest2.cpp

========================================================================
#include <iostream>
#include <string>
#include <vector>
#include <map>
#include <sys/types.h>
#include <sys/timeb.h>

inline
double timeToMillisec()
{
  timeb x;    // Change to _timeb with MSVC++
  ftime(&x);  // Change to _ftime with MSVC++
  return double(x.time) + double(x.millitm)/1000.0;
}

typedef std::string           String;
typedef std::vector<String>   List;
typedef std::map<String,List> Map;

List foo(unsigned numEntries)
{
  List v;
  for (unsigned i = 0u; i < numEntries; ++i)
    v.push_back(String("foo") + char(i % 256));
  return v;
}

Map bar(unsigned numEntries, List v)
{
  Map m;
  for (unsigned i = 0u; i < numEntries; ++i)
    m[String("foo") + char(i % 256)] = v;
  return m;
}

int main()
{
  const unsigned perIteration = 1000u;

  std::cout << "Testing: " << std::flush;
  const double begin = timeToMillisec();
  double elapsed;
  unsigned total = 0u;

  do {
    bar(perIteration, foo(perIteration));
    total += perIteration * perIteration;
    elapsed = timeToMillisec() - begin;
  } while (elapsed < 1.0);

  std::cout << total << " in " << elapsed << " sec = "
            << (elapsed * 1000.0 * 1000.0 / total) << " ns/per\n";

  return 0;
}
========================================================================

-- 
           Summary: 10x slower execution speed GCC 3.3 (compared with GCC
                    3.2)
           Product: gcc
           Version: 3.3
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: c++
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: cline at parashift dot com
                CC: gcc-bugs at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug c++/13351] [3.3 Regression] 10x slower execution speed GCC 3.3 (compared with GCC 3.2)

[pinskia at gcc dot gnu dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |compile-time-hog
            Summary|10x slower execution speed  |[3.3 Regression] 10x slower
                   |GCC 3.3 (compared with GCC  |execution speed GCC 3.3
                   |3.2)                        |(compared with GCC 3.2)
   Target Milestone|---                         |3.3.3


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug c++/13351] [3.3 Regression] 10x slower generated code (GCC 3.3 compared with GCC 3.2)

[cline at parashift dot com]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[3.3 Regression] 10x slower |[3.3 Regression] 10x slower
                   |execution speed GCC 3.3     |generated code (GCC 3.3
                   |(compared with GCC 3.2)     |compared with GCC 3.2)


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug c++/13351] [3.3 Regression] 10x slower generated code (GCC 3.3 compared with GCC 3.2)

[cline at parashift dot com]

------- Additional Comments From cline at parashift dot com  2003-12-08 10:42 -------
Subject: RE:  [3.3 Regression] 10x slower execution speed GCC 3.3 (compared with GCC 3.2)

Sorry to bother, but the keyword you've added is incorrect: the issue is the
speed of the *generated* code, not the compilation itself.

10x slower is a bit extreme, so I hope you consider this to be a high
priority issue assuming you can confirm that its existence.

Marshall


-----Original Message-----
From: pinskia at gcc dot gnu dot org [mailto:gcc-bugzilla@gcc.gnu.org] 
Sent: Monday, December 08, 2003 1:43 AM
To: cline@parashift.com
Subject: [Bug c++/13351] [3.3 Regression] 10x slower execution speed GCC 3.3
(compared with GCC 3.2)




-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |compile-time-hog
            Summary|10x slower execution speed  |[3.3 Regression] 10x slower
                   |GCC 3.3 (compared with GCC  |execution speed GCC 3.3
                   |3.2)                        |(compared with GCC 3.2)
   Target Milestone|---                         |3.3.3


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

------- You are receiving this mail because: -------
You reported the bug, or are watching the reporter.




-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug libstdc++/13351] [3.3 Regression] 10x slower generated code (GCC 3.3 compared with GCC 3.2)

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2003-12-08 11:16 -------
tin:~/src/gnu/gcctest>~/ia32_linux_gcc3_3/bin/g++ pr13351.cc -O -static
tin:~/src/gnu/gcctest>./a.out
Testing: 100000000 in 1.371 sec = 1.371 ns/per
tin:~/src/gnu/gcctest>~/ia32_linux_gcc3_2/bin/g++ pr13351.cc -O -static
tin:~/src/gnu/gcctest>./a.out
Testing: 100000000 in 1.381 sec = 1.381 ns/per
tin:~/src/gnu/gcctest>g++ pr13351.cc -O -static <-- 3.4 (20031125)
Testing: 100000000 in 1.447 sec = 1.447 ns/per

To me it looks faster (3.4 has regressed some what though).

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|c++                         |libstdc++
           Keywords|compile-time-hog            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug libstdc++/13351] [3.3 Regression] Generated code is noticeable slower (compared to 3.2)

[giovannibajo at libero dot it]

------- Additional Comments From giovannibajo at libero dot it  2003-12-08 12:10 -------
Under cygwin, I get a slowdown, but it's nowhere as significant as 10x:

$ g++-3.2.2 -O pr13351.cc && time ./a.exe
Testing: 9000000 in 1.032 sec = 0.114667 ns/per

real    0m1.120s
user    0m0.010s
sys     0m0.040s

$ g++-3.3.1 -O pr13351.cc && time ./a.exe
Testing: 2000000 in 1.509 sec = 0.7545 ns/per

real    0m1.776s
user    0m1.502s
sys     0m0.050s


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
   Last reconfirmed|0000-00-00 00:00:00         |2003-12-08 12:10:42
               date|                            |
            Summary|[3.3 Regression] 10x slower |[3.3 Regression] Generated
                   |generated code (GCC 3.3     |code is noticeable slower
                   |compared with GCC 3.2)      |(compared to 3.2)


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug libstdc++/13351] [3.3 Regression] Generated code is noticeable slower (compared to 3.2)

[cline at parashift dot com]

------- Additional Comments From cline at parashift dot com  2003-12-08 13:00 -------
Subject: RE:  [3.3 Regression] Generated code is noticeable slower (compared to 3.2)

Actually you got pretty close to what I got.  Your numbers show a slow-down
of roughly 6x or 7x (the ns/per number is the significant one; the others
are merely used to compute that number).

Marshall


-----Original Message-----
From: giovannibajo at libero dot it [mailto:gcc-bugzilla@gcc.gnu.org] 
Sent: Monday, December 08, 2003 6:11 AM
To: cline@parashift.com
Subject: [Bug libstdc++/13351] [3.3 Regression] Generated code is noticeable
slower (compared to 3.2)



------- Additional Comments From giovannibajo at libero dot it  2003-12-08
12:10 -------
Under cygwin, I get a slowdown, but it's nowhere as significant as 10x:

$ g++-3.2.2 -O pr13351.cc && time ./a.exe
Testing: 9000000 in 1.032 sec = 0.114667 ns/per

real    0m1.120s
user    0m0.010s
sys     0m0.040s

$ g++-3.3.1 -O pr13351.cc && time ./a.exe
Testing: 2000000 in 1.509 sec = 0.7545 ns/per

real    0m1.776s
user    0m1.502s
sys     0m0.050s


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
   Last reconfirmed|0000-00-00 00:00:00         |2003-12-08 12:10:42
               date|                            |
            Summary|[3.3 Regression] 10x slower |[3.3 Regression] Generated
                   |generated code (GCC 3.3     |code is noticeable slower
                   |compared with GCC 3.2)      |(compared to 3.2)


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

------- You are receiving this mail because: -------
You reported the bug, or are watching the reporter.




-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug libstdc++/13351] [3.3 Regression] BIG slowdown in generated code (compared to 3.2)

[giovannibajo at libero dot it]

------- Additional Comments From giovannibajo at libero dot it  2003-12-08 13:09 -------
Ok, the example confused me. Much easier to not play games with timers and just 
do this in the main:

int main()
{
  const unsigned perIteration = 1000u;

  for (unsigned i=0;i<30;i++)
    bar(perIteration, foo(perIteration));

  return 0;
}

I'll attach my simplified testcase. Now I get this:

$ g++-3.2.2 -O pr13351.cc && time ./a.exe

real    0m3.525s
user    0m0.010s
sys     0m0.020s


$ g++-3.3.1 -O pr13351.cc && time ./a.exe

real    0m22.814s
user    0m21.510s
sys     0m0.050s


Which is indeed a BIG slowdown.


-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[3.3 Regression] Generated  |[3.3 Regression] BIG
                   |code is noticeable slower   |slowdown in generated code
                   |(compared to 3.2)           |(compared to 3.2)


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug libstdc++/13351] [3.3 Regression] BIG slowdown in generated code (compared to 3.2)

[falk at debian dot org]

------- Additional Comments From falk at debian dot org  2003-12-08 13:31 -------
I cannot reproduce this on Alpha. How's the speed at -O2? How about compilation
time and binary size? If compilation time or binary size are better, I would
find the slowdown quite acceptable at -O1 (but not at -O2).

falk@juist:/tmp% time g++-3.2 -O test.cc && ls -l a.out && time ./a.out
g++-3.2 -O test.cc  3.70s user 0.10s system 88% cpu 4.320 total
-rwxr-xr-x    1 falk     falk        36629 Dec  8 14:26 a.out*
./a.out  2.54s user 0.05s system 96% cpu 2.667 total
falk@juist:/tmp% time g++-3.3 -O test.cc && ls -l a.out && time ./a.out
g++-3.3 -O test.cc  2.85s user 0.20s system 86% cpu 3.535 total
-rwxr-xr-x    1 falk     falk        32434 Dec  8 14:26 a.out*
./a.out  2.54s user 0.03s system 91% cpu 2.795 total

At least on Alpha, 3.3 seems like a noticeable improvement :)


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug libstdc++/13351] [3.3 Regression] BIG slowdown in generated code (compared to 3.2)

[carlo at alinoe dot com]

------- Additional Comments From carlo at alinoe dot com  2003-12-08 15:33 -------
Subject: Re:  [3.3 Regression] 10x slower generated code (GCC 3.3 compared with GCC 3.2)

Not the case on i686-pc-linux-gnu (athlon 900 Mhz)

g++ 3.2   : 0.2116 ns/per
g++ 3.2.1 : 0.32025 ns/per
g++ 3.2.2 : 0.31925 ns/per
g++ 3.2.3 : 0.3185 ns/per

g++ 3.3   : doesn't compile
g++ 3.3.1 : 0.102 ns/per
g++ 3.3.2 : 0.1034 ns/per

mainline: 0.0818462 ns/per


So I'd say - it became 3 times FASTER, instead.



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug libstdc++/13351] [3.3 Regression] BIG slowdown in generated code (compared to 3.2)

[gdr at integrable-solutions dot net]

------- Additional Comments From gdr at integrable-solutions dot net  2003-12-08 17:38 -------
Subject: Re:  [3.3 Regression] 10x slower generated code (GCC 3.3 compared with GCC 3.2)

"cline at parashift dot com" <gcc-bugzilla@gcc.gnu.org> writes:

| 10x slower is a bit extreme, so I hope you consider this to be a high
| priority issue assuming you can confirm that its existence.

Unless, the patch is relatively trivial to solve that problem, I don't
think it is going to be fixed in 3.3.3.

-- Gaby


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug libstdc++/13351] [3.3 Regression] BIG slowdown in generated code (compared to 3.2)

[bkoz at gcc dot gnu dot org]

------- Additional Comments From bkoz at gcc dot gnu dot org  2003-12-09 21:30 -------

Another datapoint, on linux:


%time 3.3.2.out
1.790u 0.040s 0:01.84 99.4%     0+0k 0+0io 209pf+0w

%time 3.4-cur.out
1.650u 0.040s 0:01.70 99.4%     0+0k 0+0io 201pf+0w

Please try to get something that the linux people can reproduce. Right now, I'm
not even convinced this is a problem.

-benjamin


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug libstdc++/13351] [3.3 Regression] BIG slowdown in generated code (compared to 3.2)

[cline at parashift dot com]

------- Additional Comments From cline at parashift dot com  2003-12-13 22:02 -------
Subject: RE:  [3.3 Regression] BIG slowdown in generated code (compared to 3.2)

On Tue, Dec 09, 2003, bkoz wrote:
>Please try to get something that the linux people can reproduce. Right now,
I'm
>not even convinced this is a problem.
>-benjamin

Admirable request, but apparently not one that can be fulfilled: others have
reproduced the problem on Cygwin, but apparently the problem does not show
up on Linux.

Marshall




-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351

[Bug libstdc++/13351] [3.3 Regression] BIG slowdown in generated code (compared to 3.2)

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2003-12-23 09:47 -------
Will not fix for 3.3.3, will be invasive.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |WONTFIX


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13351