This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Fix PR34706, PR34683 -- use more precise types for fortran arrays
On Jan 8, 2008 9:53 PM, Janus Weil <jaydub66@googlemail.com> wrote:
> Hi Richard,
> thanks for taking care of this. Already the other patches you
> committed in connection with PR34683 have helped a bit. With rev.
> 131401 the test case compiles in about 60sec using 570MB of memory
> (with -O1 -fstrict aliasing). This is already significantly better
> than the 800MB from before.
> After applying your trans-types.c patch, this further improves to
> 33sec and 440MB. Which is nice, but still way above last years'
> October builds. Most of the memory is still used for "tree operand
> scan" and "tree FRE". So I guess there still must be something going
> wrong.
Yes, as I analyzed in the PR the real algorithmic problem left is the
loop unroller, which does work in the complexity O (N^2). Unfortunately
I think we will not be able to fix this for 4.3.
Your observation with -O2 / -O3 differences is because the Fortran FE patch
effectiveness depends on memory partitioning, which heuristics change
between the optimization levels.
Richard.