Old loop optimizer in 4.0?

[Zdenek Dvorak]

Hello,

I would like to know what is the public opinion on what to do
with the old loop optimizer in 4.0.  Basically there are three
possibilities:

1) Leave it enabled.
2) Leave it in, but disable it by default.
3) Remove it completely.

Some facts:

The possibility 3) has the nice property of getting rid of the last
pass in the middleend that does not work with CFG, but makes comparing
the old and new loop optimizer more difficult.  Also some optimization
are missing in new loop optimizer:

1) Array prefetching.  We have speculative prefecthing that may replace
   it partially, but can only be used with profile feedback.  Array
   prefetching is also mostly redundant on modern processors with
   hardware prefetching (in fact spoils the code sometimes, for example
   did it on x86_64 SPECs the last time I checked).  lno-branch loop
   prefetching is not quite ready yet (while it works nicely on
   micro-benchmarks, it spoils the code for SPECs quite a lot).
2) Loop reversal.  Should be trivial to implement with the tree-linear
   code, I guess.  Redundant with doloop optimization on some platforms.

Below are some benchmark comparisons from x86_64.  Base is -O2, peak
is -O2 -fno-loop-optimize -fmove-loop-invariants (i.e. old loop
optimizer disabled, loop invariant motion in new rtl loop optimizer
enabled).

(The wupwise failure is probably some misscompilation due to
-fmove-loop-invariants, I haven't checked yet).

Zdenek

   164.gzip          1400     193         726*     1400     193         726*
   175.vpr           1400     187         749*     1400     188         745*
   176.gcc           1100     116         952*     1100     116         952*
   181.mcf           1800     440         409*     1800     445         405*
   186.crafty        1000      74.3      1346*     1000      72.7      1376*
   197.parser        1800     326         551*     1800     323         556*
   252.eon           1300     103        1267*     1300     102        1279*
   253.perlbmk       1800     178        1009*     1800     182         989*
   254.gap           1100     144         762*     1100     142         772*
   255.vortex        1900     169        1126*     1900     165        1149*
   256.bzip2         1500     194         773*     1500     194         773*
   300.twolf         3000     350         858*     3000     367         818*
   Est. SPECint_base2000                  835
   Est. SPECint2000                                                     834

   168.wupwise       1600   200       799    *                             
   171.swim          3100   451       688    *     3100   451       687    *
   172.mgrid         1800   552       326    *     1800   486       370    *
   173.applu         2100   349       602    *     2100   348       603    *
   177.mesa          1400   135      1036    *     1400   137      1024    *
   178.galgel                                X                             X
   179.art           2600   278       937    *     2600   274       951    *
   183.equake        1300   156       831    *     1300   157       827    *
   187.facerec       1900   290       654    *     1900   307       619    *
   188.ammp          2200   248       888    *     2200   251       878    *
   189.lucas         2000   235       852    *     2000   232       863    *
   191.fma3d                                 X                             X
   200.sixtrack                              X                             X
   301.apsi          2600   341       762    *     2600   346       750    *
   Est. SPECfp_base2000               733    
   Est. SPECfp2000                                                  731