[Bug optimization/12419] [3.3/3.4 Regression] Performace regression: poor optimization of const memory

["janis187 at us dot ibm dot com" <gcc-bugzilla at gcc dot gnu dot org>]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=12419



------- Additional Comments From janis187 at us dot ibm dot com  2003-09-29 22:42 -------
The change reported in PR 12419 is somewhat erratic, although a
regression hunt on the 3.3 branch identified this patch:

--- gcc/gcc/ChangeLog ---

2003-04-07  Glen Nakamura  <glen@imodulo.com>

          PR opt/8634
          * explow.c (maybe_set_unchanging): Don't flag non-static const
          aggregate type initializers with RTX_UNCHANGING_P.

which includes the following new comment text:

       We cannot do this for non-static aggregates, because of the double
       writes that can be generated by store_constructor, depending on the
       contents of the initializer.  Yes, this does eliminate a good fraction
       of the number of uses of RTX_UNCHANGING_P for a language like Ada.
       It also eliminates a good quantity of bugs.  Let this be incentive to
       eliminate RTX_UNCHANGING_P entirely in favour of a more reliable
       solution, perhaps based on alias sets.

The regression hunt on the 3.3-branch took place on i686-pc-linux-gnu
using the submitter's test case compiled with -O2.

A similar hunt on mainline identified the merge of the 3.4 improvements
branch in December; very odd.  Even more odd is that I get different
results with recent mainline compilers depending on whether I use gcc or
cc1 directly.  Sounds like something is uninitialized.