[Bug target/49089] New: Regression on CFP2006 on Bulldozer From Splitting AVX 32-byte Unaligned Loads

[Bug target/49089] New: Regression on CFP2006 on Bulldozer From Splitting AVX 32-byte Unaligned Loads

[changpeng.fang at amd dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49089

           Summary: Regression on CFP2006 on Bulldozer From Splitting AVX
                    32-byte Unaligned Loads
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: changpeng.fang@amd.com


The regression is caused by the following patch that splits AVX 32-byte
unaligned load and store:
http://gcc.gnu.org/ml/gcc-patches/2011-03/msg01839.html

Here is the performance impact on a Bulldozer System:

              store-split     load-split
410.bwaves    0.48          -0.48
416.gamess    0.55           0.00
433.milc    1.76          -3.96
434.zeusmp    3.48          -3.48
435.gromacs    0.51           1.54
436.cactusADM    -0.72          -0.72
437.leslie3d    10.33          -0.94
444.namd    1.03           0.00
447.dealII    0.70          -1.41
450.soplex    0.79           0.40
453.povray    -0.50           -0.50
454.calculix    5.07           -1.84
459.GemsFDTD    4.33           -6.25
465.tonto    1.27            0.00
470.lbm        -0.86            1.44
481.wrf        1.35            -3.59
482.sphinx3    0.00            -2.11
geomean        1.71            -1.31

While splitting store is good, Bulldozer seems not like unaligned
load splitting.

[Bug target/49089] Regression on CFP2006 on Bulldozer From Splitting AVX 32-byte Unaligned Loads

[changpeng.fang at amd dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49089

--- Comment #1 from Changpeng Fang <changpeng.fang at amd dot com> 2011-05-20 18:01:29 UTC ---
Apparently, this default option setting should only apply to systems that
splitting loads is bebeficial:

config/i386/i386.c(ix86_option_override_internal):

if (!(target_flags_explicit & MASK_AVX256_SPLIT_UNALIGNED_LOAD))
  target_flags |= MASK_AVX256_SPLIT_UNALIGNED_LOAD;

[Bug target/49089] Regression on CFP2006 on Bulldozer From Splitting AVX 32-byte Unaligned Loads

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49089

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2011.05.21 09:38:03
     Ever Confirmed|0                           |1

--- Comment #2 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-05-21 09:38:03 UTC ---
Confirmed.  The default setting should be part of the tuning flags.