[Bug fortran/30388] New: gfortran42 is slower than g77 3.4 about 10%

[Bug fortran/30388] New: gfortran42 is slower than g77 3.4 about 10%

[chat95 at mac dot com]

1. download http://accc.riken.jp/HPC/HimenoBMT/Load_module/himenoBMTxp_m.lzh
% wget http://accc.riken.jp/HPC/HimenoBMT/Load_module/himenoBMTxp_m.lzh
2. extract with lha
% lha x himenoBMTxp_m.lzh 
himenobmtxp_m.f - Melted   :  o
3. apply following patch:
--- himenobmtxp_m.f~    Mon Jan  7 13:54:48 2002
+++ himenobmtxp_m.f     Sat Jan  6 09:02:22 2007
@@ -39,7 +39,6 @@
 C "use portlib" statement on the next line is for Visual fortran
 C to use UNIX libraries. Please remove it if your system is UNIX.
 C -------------------
-      use portlib
       IMPLICIT REAL*4(a-h,o-z)
 C
 C      PARAMETER (mimax=513,mjmax=257,mkmax=257)

4. compile with gfortran42 and g77-34.
% g77-34 -O -o himenobmtxp_m himenobmtxp_m.f
% ./himenobmtxp_m
  mimax= 257 mjmax= 129 mkmax= 129
  imax= 256 jmax= 128 kmax= 128
  Start rehearsal measurement process.
  Measure the performance in 3 times.
   MFLOPS:  421.095337  time(s):  0.976774991  0.00169381942
 Now, start the actual measurement process.
 The loop will be excuted in 184 times.
 This will take about one minute.
 Wait for a while.
  Loop executed for  184 times
  Gosa :  0.00126111682
  MFLOPS:  431.486053  time(s):  58.4661865
  Score based on Pentium III 600MHz :  5.20866823
PAUSE  statement executed
To resume execution, type go.  Other input will terminate the job.

STOP
% gfortran42 -O -o himenobmtxp_m himenobmtxp_m.f
himenobmtxp_m.f:111.72:

      pause                                                             
                                                                       1
Warning: Obsolete: PAUSE statement at (1)
% ./himenobmtxp_m
  mimax=         257  mjmax=         129  mkmax=         129
  imax=         256  jmax=         128  kmax=         128
  Start rehearsal measurement process.
  Measure the performance in 3 times.
   MFLOPS:   190.3127       time(s):   2.161261      1.6937795E-03
 Now, start the actual measurement process.
 The loop will be excuted in          83  times.
 This will take about one minute.
 Wait for a while.
  Loop executed for           83  times
  Gosa :  1.4101295E-03
  MFLOPS:   372.9143       time(s):   30.51566    
  Score based on Pentium III 600MHz :   4.501622    
PAUSE
To resume execution, type go.  Other input will terminate the job.

5 result:
g77-34 MFLOPS: 431.486053
gfortran MFLOPS: 372.9143

% g77-34 -v
Reading specs from
/usr/local/lib/gcc/i386-portbld-freebsd6.1/3.4.6/gcc/i386-portbld-freebsd6.1/3.4.6/specs
Configured with: ./..//gcc-3.4.6/configure --disable-nls --with-system-zlib
--with-libiconv-prefix=/usr/local --program-suffix=34
--libdir=/usr/local/lib/gcc/i386-portbld-freebsd6.1/3.4.6
--with-gxx-include-dir=/usr/local/lib/gcc/i386-portbld-freebsd6.1/3.4.6/include/c++/
--infodir=/usr/local/info/gcc34 --prefix=/usr/local i386-portbld-freebsd6.1
Thread model: posix
gcc version 3.4.6 [FreeBSD]

% gfortran42 -v
Using built-in specs.
Target: i386-portbld-freebsd6.1
Configured with: ./..//gcc-4.2-20070102/configure --disable-nls
--with-system-zlib --with-libiconv-prefix=/usr/local --with-gmp=/usr/local
--program-suffix=42 --libdir=/usr/local/lib/gcc-4.2.0
--with-gxx-include-dir=/usr/local/lib/gcc-4.2.0/include/c++/
--infodir=/usr/local/info/gcc42 --disable-rpath --prefix=/usr/local
i386-portbld-freebsd6.1
Thread model: posix
gcc version 4.2.0 20070102 (prerelease)


-- 
           Summary: gfortran42 is slower than g77 3.4 about 10%
           Product: gcc
           Version: 4.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: fortran
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: chat95 at mac dot com
 GCC build triplet: i386-portbld-freebsd6.1
  GCC host triplet: i386-portbld-freebsd6.1


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[chat95 at mac dot com]

------- Comment #1 from chat95 at mac dot com  2007-01-06 00:32 -------
My CPU is : 
Jan  6 06:59:54 debussy kernel: ACPI APIC Table: <INTEL  D845PESV>
Jan  6 06:59:54 debussy kernel: Timecounter "i8254" frequency 1193182 Hz
quality 0
Jan  6 06:59:54 debussy kernel: CPU: Intel(R) Pentium(R) 4 CPU 2.40GHz
(2400.10-MHz 686-
class CPU)
Jan  6 06:59:54 debussy kernel: Origin = "GenuineIntel"  Id = 0xf27  Stepping =
7
Jan  6 06:59:54 debussy kernel:
Features=0xbfebfbff<FPU,VME,DE,PSE,TSC,MSR,PAE,MCE,CX8,A
PIC,SEP,MTRR,PGE,MCA,CMOV,PAT,PSE36,CLFLUSH,DTS,ACPI,MMX,FXSR,SSE,SSE2,SS,HTT,TM,PBE>
Jan  6 06:59:54 debussy kernel: Features2=0x4400<CNTX-ID,<b14>>
Jan  6 06:59:54 debussy kernel: real memory  = 1072955392 (1023 MB)
Jan  6 06:59:54 debussy kernel: avail memory = 1040994304 (992 MB)


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[pinskia at gcc dot gnu dot org]

------- Comment #2 from pinskia at gcc dot gnu dot org  2007-01-06 00:33 -------
What happens if you use -O2 instead of -O ?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[chat95 at mac dot com]

------- Comment #3 from chat95 at mac dot com  2007-01-06 00:38 -------
% gfortran42 -o himenobmtxp_m.gfortran42 -O2 -ffast-math -finline-functions
-fomit-frame-pointer -funroll-loops -fexpensive-optimizations himenobmtxp_m.f
% ./himenobmtxp_m.gfortran42 
  mimax=         257  mjmax=         129  mkmax=         129
  imax=         256  jmax=         128  kmax=         128
  Start rehearsal measurement process.
  Measure the performance in 3 times.
   MFLOPS:   194.9637       time(s):   2.109703      1.6937795E-03
 Now, start the actual measurement process.
 The loop will be excuted in          85  times.
 This will take about one minute.
 Wait for a while.
  Loop executed for           85  times
  Gosa :  1.4063698E-03
  MFLOPS:   391.9939       time(s):   29.72989    
  Score based on Pentium III 600MHz :   4.731940    
PAUSE
To resume execution, type go.  Other input will terminate the job.
% g77-34 -o himenobmtxp_m.g77 -O2 -ffast-math -finline-functions
-fomit-frame-pointer -funroll-loops -fexpensive-optimizations himenobmtxp_m.f
% ./himenobmtxp_m.g77 
  mimax= 257 mjmax= 129 mkmax= 129
  imax= 256 jmax= 128 kmax= 128
  Start rehearsal measurement process.
  Measure the performance in 3 times.
   MFLOPS:  242.952499  time(s):  1.69298697  0.00169377949
 Now, start the actual measurement process.
 The loop will be excuted in 106 times.
 This will take about one minute.
 Wait for a while.
  Loop executed for  106 times
  Gosa :  0.00136957248
  MFLOPS:  246.231003  time(s):  59.0223999
  Score based on Pentium III 600MHz :  2.97236848
PAUSE  statement executed
To resume execution, type go.  Other input will terminate the job.

Summary:

g77-34 with -O  : 431.486053 MFLOPS
g77-34 with -O2 -ffast-math -finline-functions -fomit-frame-pointer
-funroll-loops -fexpensive-optimizations
                : 246.231003 MFLOPS
gfortran42 with -O : 372.9143 MFLOPS
gfortran42 with -O2 -ffast-math -finline-functions -fomit-frame-pointer
-funroll-loops -fexpensive-optimizations
                   : 391.9939 MFLOPS

gfortran42 with -O2 ... produced faster code than gfortran42 -O, but
slower than g77-34 with -O, but g77-34 with -O2 ... produces very slow
code.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[jvdelisle at gcc dot gnu dot org]

------- Comment #4 from jvdelisle at gcc dot gnu dot org  2007-01-06 02:13 -------
Could you please post a gzipped version of the test case?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[brooks at gcc dot gnu dot org]

------- Comment #5 from brooks at gcc dot gnu dot org  2007-01-06 02:36 -------
Created an attachment (id=12863)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=12863&action=view)
Original source file (extracted from archive)


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[jvdelisle at gcc dot gnu dot org]

------- Comment #6 from jvdelisle at gcc dot gnu dot org  2007-01-06 04:44 -------
With my system:  x86-64 Linux

Using -O3 -ffast-math in both cases:

g77:       MFLOPS:  893.198364

gfortran:  MFLOPS:   895.7278


Using -O3 -ffast-math in both cases:

g77:       MFLOPS:  930.39801

gfortran:   MFLOPS:   891.1194


Using -O1 -ffast-math in both cases:

g77:        MFLOPS:  962.237854

gfortran:   MFLOPS:   890.9428


Using -O1 -ffast-math in both cases:

g77:        MFLOPS:  259.945648

gfortran:   MFLOPS:   199.7145 


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[jvdelisle at gcc dot gnu dot org]

------- Comment #7 from jvdelisle at gcc dot gnu dot org  2007-01-06 05:05 -------
According to gprof, not unexpected, 97% of time is in the jacoobi routine. 
This might warrant further study by the optimizer experts.  Certainly not by
me. :)


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[chat95 at mac dot com]

------- Comment #8 from chat95 at mac dot com  2007-01-06 05:08 -------

Using -O3 -ffast-math in both cases:

Using -O3 -ffast-math in both cases:

Using -O1 -ffast-math in both cases:

Using -O1 -ffast-math in both cases:

what's the difference between theset two?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[jvdelisle at gcc dot gnu dot org]

------- Comment #9 from jvdelisle at gcc dot gnu dot org  2007-01-06 05:26 -------
Typo on my part:

They are:

-O3
-O2
-O1
-O0

In that order.  I apologize


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[rguenth at gcc dot gnu dot org]

------- Comment #10 from rguenth at gcc dot gnu dot org  2007-01-06 10:17 -------
Well, it all boils down to differences in induction variable choices.  The
choice made by 3.4 (for i686) doesn't look too good to me compared to 4.1.2
here.

4.1: MFLOPS:   990.4130       time(s):   28.10176
3.4: MFLOPS:  449.016235  time(s):  59.8477402

(-O2 -ffast-math -funroll-loops -fomit-frame-pointer)

The predcomm branch should be able to do a lot better here.

With SSE math I get

4.1: MFLOPS:   954.2382       time(s):   31.46597
3.4: MFLOPS:  467.92215  time(s):  59.1876984


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rakdver at gcc dot gnu dot
                   |                            |org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[fxcoudert at gcc dot gnu dot org]

------- Comment #11 from fxcoudert at gcc dot gnu dot org  2007-01-23 15:23 -------
Well, I can confirm that one the reporter's code, on my i686-linux (Intel(R)
Pentium(R) III CPU family      1266MHz), whatever optimisation flags I give to
gfortran-4.3, "g77-3.4.6 -O" does a better job.


-- 

fxcoudert at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |fxcoudert at gcc dot gnu dot
                   |                            |org
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2007-01-23 15:23:57
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[jvdelisle at gcc dot gnu dot org]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 795 bytes --]



------- Comment #12 from jvdelisle at gcc dot gnu dot org  2007-12-17 06:10 -------
Here is a twist on the test case

 gfc -O2 -m64 -ffast-math -funroll-all-loops -fipa-pta -fivopts poisson.f 
poisson.f:111.72:

      pause                                                             
                                                                       1
Warning: Deleted feature: PAUSE statement at (1)
poisson.f: In function ‘jacobi’:
poisson.f:233: error: stmt (0x2aaaaeb4f300) marked modified after optimization
pass: 
D.1063_9 = *nn_8(D);
poisson.f:233: internal compiler error: verify_ssa failed
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[steven at gcc dot gnu dot org]

------- Comment #13 from steven at gcc dot gnu dot org  2007-12-17 06:33 -------
-fipa-pta is broken iirc.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[jvdelisle at gcc dot gnu dot org]

------- Comment #14 from jvdelisle at gcc dot gnu dot org  2008-02-17 04:11 -------
Two questions on this PR.

Is there really anything on the gfortran side we can do to make this better or
is it really a middle-end / back-end issue?

Can we close this pr or change the component to other than fortran?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388

[Bug fortran/30388] gfortran42 is slower than g77 3.4 about 10%

[jvdelisle at gcc dot gnu dot org]

------- Comment #15 from jvdelisle at gcc dot gnu dot org  2008-11-26 04:05 -------
Not a gfortran frontend issue, so closing.


-- 

jvdelisle at gcc dot gnu dot org changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30388