[Bug lto/41578] New: 7% slower runtime with -flto than without

[Bug lto/41578] New: 7% slower runtime with -flto than without

[burnus at gcc dot gnu dot org]

AMD Athlon64 4800+ (dual core, 1MB L2 cache each, SSE, SSE2, SSE3), 4GB DDR400
with
[LTO]   gfortran -march=native -ffast-math -funroll-loops -flto -O3
[noLTO] gfortran -march=opteron -ffast-math -funroll-loops -ftree-loop-linear
-ftree-vectorize -msse3

When running the Polyhedron Benchmark suite, the LTO version beats the non-LTO
version in the geometric mean value: 23.91s [100%] vs 23.83s [99%]
 (http://www.polyhedron.co.uk/MFL6VW74649).

However, the capacita benchmark is significantly slower, others are also a bit
slower:
            [noLTO]       [LTO]
  capacita  81.54 [100]   87.41 [107]
  mdbx      19.49 [100]   20.07 [102]
  rnflow    34.91 [100]   36.04 [103]
  test_fpu  21.66 [100]   22.40 [103]

(Largest performance gain: aermod (35.14 [100], 31.46 [89]) followed by induct 
(36.51 [100], 35.18 [96]); the others are 2% to 0% faster with LTO.)


-- 
           Summary: 7% slower runtime with -flto than without
           Product: gcc
           Version: 4.5.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: lto
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: burnus at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41578

[Bug lto/41578] 7% slower runtime with -flto than without

[rguenth at gcc dot gnu dot org]

------- Comment #1 from rguenth at gcc dot gnu dot org  2009-10-05 13:10 -------
Try enabling -fwhole-program together with -flto.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41578

[Bug lto/41578] 7% slower runtime with -flto than without

[burnus at gcc dot gnu dot org]

------- Comment #2 from burnus at gcc dot gnu dot org  2009-10-06 12:14 -------
(In reply to comment #1)
> Try enabling -fwhole-program together with -flto.

I did now - and thanks to sanitizing the tree, there is no failure (contrary to
using only -fwhole-program w/o -flto).

capacita is now only <1% slower; but mdbx, rnflow, and test_fpu are still 2% to
3% slower. (The overall performance [geometric mean] is still 1% faster.)

(Comparison just for fun: Pathscale 3.2.99 is (geometric mean) 9%, ifort 11.1
is 8% faster.)


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41578

[Bug lto/41578] 7% slower runtime with -flto than without

[steven at gcc dot gnu dot org]

------- Comment #3 from steven at gcc dot gnu dot org  2009-10-23 19:38 -------
This one should just be monitored from week to week. There is so much stuff
going into the trunk right now that results vary from one day to another by +5%
to -5% in the last four days alone.


-- 

steven at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2009-10-23 19:38:35
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41578

[Bug lto/41578] 7% slower runtime with -flto than without

[rguenth at gcc dot gnu dot org]

------- Comment #4 from rguenth at gcc dot gnu dot org  2009-10-23 22:27 -------
The bugreport is not useful in its current form anyway (no testcase).  Let's
close it.


-- 

rguenth at gcc dot gnu dot org changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41578