[Bug c/54896] [4.7/4.8 Regression] Some optimization slowness with GCC 4.7.2

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54896

Steven Bosscher <steven at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |alias, compile-time-hog
             Status|WAITING                     |NEW
                 CC|                            |rguenth at gcc dot gnu.org
      Known to work|                            |4.3.2, 4.4.0
            Summary|Some optimization slowness  |[4.7/4.8 Regression] Some
                   |with GCC 4.7.2              |optimization slowness with
                   |                            |GCC 4.7.2
      Known to fail|                            |4.7.2, 4.8.0

--- Comment #3 from Steven Bosscher <steven at gcc dot gnu.org> 2012-10-10 22:51:12 UTC ---
Thanks for the test case!

Bug is confirmed with GCC 4.8 (trunk revision 192219).

Problem areas at -O1:
 alias stmt walking    :  31.68 (36%) usr
 tree DSE              :  10.83 (12%) usr
 CSE                   :   9.02 (10%) usr
 reload CSE regs       :  23.17 (26%) usr
 TOTAL                 :  87.99

Problem areas at -O2 are pretty much the same:
 alias stmt walking    :  47.29 (28%) usr
 tree DSE              :  10.84 ( 7%) usr
 CSE                   :   9.12 ( 5%) usr
 CSE 2                 :  34.66 (21%) usr
 reload CSE regs       :  45.37 (27%) usr
 TOTAL                 : 166.12


GCC 4.3.2 has the same problems at the RTL level:
 tree DSE              :  11.60 (17%) usr
 CSE                   :  19.74 (28%) usr
 CSE 2                 :  12.63 (18%) usr
 reload CSE regs       :   8.49 (12%) usr
 TOTAL                 :  69.68


tree-DSE, CSE and reload-CSE are both quadratic in the number of
instructions per basic block, and all basic blocks in this test
case contain O(10^3) instructions.

The alias stmt walking slowness is a regression, but I suspect
there a dup that Rick Biener already knows about...