[Patch, i386] Turn on -fschedule-insns and -fsched-pressure by default for x86-64

[Fang, Changpeng]

________________________________________
From: Jack Howarth [howarth@bromo.med.uc.edu]
Sent: Tuesday, December 07, 2010 12:10 AM
To: Fang, Changpeng
Cc: gcc-patches@gcc.gnu.org; sebpop@gmail.com; hjl.tools@gmail.com; Yazdani, Reza

>According to the Polyhedron 2005 benchmarks, -fschedule-insns and -fsched-pressure
>result in a net performance loss (at least for -mtune=core2) at both -m32 and -m64
>on x86_64-apple-darwin10. The induct benchmark in particular suffers a severe
>regression with those options.

>x86_64-apple-darwin10
>gfortran -m64 -mtune=core2 -ffast-math -funroll-loops -O3 %n.f90 -o %n

>benchmark           stock  -fschedule-insns  %change
                           -fsched-pressure
>ac                   8.81    8.75             -0.7
>aermod              17.21   17.21              0.0
>air                  5.53    5.53              0.0
>capacita            32.39   32.89              1.5
>channel              1.84    1.84              0.0
>doduc               26.40   26.55              0.6
>fatigue              8.14    8.22              1.0
>gas_dyn              4.31    4.29             -0.5
>induct              12.85   15.89             19.1
>linpk               15.47   15.47              0.0
>mdbx                11.21   11.19             -0.2
>nf                  29.98   29.96             -0.1
>protein             34.03   33.44             -1.8
>rnflow              23.13   22.91             -1.0
>test_fpu             8.03    8.11              1.0
>tfft                 1.87    1.87              0.0

>Geometric Mean      10.84   10.99              1.4

For -m64, it seems only induct is a concern. The performance degradation depends on optimization level.
Here is my measurement on amdfam10 for induct:

     -O2            -O2 -fschedule-insns -fsched-pressure        improvement
     33.77                   34.11                                                      -1.01%

    -O3           -O3 -fschedule-insns -fsched-pressure        improvement
    28.44                    27.47                                                      +2.80%

    -Ofast       -Ofast -fschedule-insns -fsched-pressure        improvement
   16.94                      18.95                                                     -11.87%

    -Ofast -funroll-loops      -Ofast -funroll-loops -fschedule-insns -fsched-pressure        improvement
    17.14                      18.97                                                     -10.68%


I guess, with -Ofast (-O3 -ffast-math), register pressure is higher, which makes the scheduling optimization less
effective in general.

Thanks,

Changpeng