[Bug fortran/48636] Enable more inlining with -O2 and higher

["dominiq at lps dot ens.fr" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48636

--- Comment #24 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2012-10-17 13:13:24 UTC ---
Summary for the polyhedron tests (pb05):

(a) revision 192449 unpatched
(b) revision 192516 with patch in comment #20
options: -fprotect-parens -Ofast -funroll-loops -ftree-loop-linear \
         -fomit-frame-pointer -fwhole-program -flto -fno-tree-loop-if-convert

Benchmark         Compile            Executable              Ave Run   
     Name          (secs)               (bytes)               (secs)  
              (a)       (b)        (a)         (b)        (a)       (b)     
---------   -------   -------  ----------  ----------   -------   -------
       ac      6.30      9.66       54904       54856      9.12      8.49
   aermod    129.56    178.13     1158904     1527680     17.69     17.28
      air     25.64     29.95      102752      106712      7.15      7.16
 capacita      6.34     18.99       69096      130536     42.00     40.57
  channel      2.99      3.51       34736       34736      3.03      3.02
    doduc     19.51     23.70      163528      196016     27.54     27.65
  fatigue      7.03      7.54       69312       69248      4.92      2.94
  gas_dyn     14.65     14.91       99320       99280      3.72      3.88
   induct     19.50     21.02      149160      161552     13.18     13.14
    linpk      1.84      2.91       22008       17832     21.65     21.69
     mdbx      6.59     10.14       68800       80968     12.85     12.78
       nf      6.34     24.25       59544      104544     18.95     18.88
  protein     22.27     46.34      119056      156048     35.65     35.41
   rnflow     17.75     31.23      114600      147280     23.56     23.54
 test_fpu     10.69     20.66       76640      117512     10.89     10.95
     tfft      1.62      2.93       22424       22384      3.34      3.33

Geometric Mean Execution Time =  (a) 11.88 seconds (b) 11.43 seconds

I also see many failures for the gcc.dg/tree-ssa/slsr-* tests: slsr-2.c to
slsr-11.c, slsr-14.c to slsr-20.c, slsr-24.c, and slsr-25.c, and for
gfortran.dg/vect/vect-8.f90: the loop

do while(ii >  1)
ipnt= ipntp
ipntp= ipntp+ii
ii= ishft(ii,-1)
i= ipntp+1
!dir$ vector always
       x(ipntp+2:ipntp+ii+1)=x(ipnt+2:ipntp:2)-v(ipnt+2:ipntp:2) &
     &*x(ipnt+1:ipntp-1:2)-v(ipnt+3:ipntp+1:2)*x(ipnt+3:ipntp+1:2)
END DO

is not vectorized because

154: not vectorized: complicated access pattern.