[Bug tree-optimization/54717] [4.8 Regression] Runtime regression: polyhedron test "rnflow" degraded

[dominiq at lps dot ens.fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54717

--- Comment #8 from Dominique d'Humieres <dominiq at lps dot ens.fr> 2012-10-02 20:23:42 UTC ---
Created attachment 28333
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28333
bzipped tar archive of a reduced test

The tar archive contains the files
cptrf2_inl_1.f90  rnflow.in  rnflow_red.f90  rnfprm.h
and can be used as in

[macbook] dbg_rnflow/pr54717% gfc -c -Ofast -funroll-loops rnflow_red.f90
[macbook] dbg_rnflow/pr54717% gfc -c -O2 cptrf2_inl_1.f90
[macbook] dbg_rnflow/pr54717% gfc rnflow_red.o cptrf2_inl_1.o
[macbook] dbg_rnflow/pr54717% time a.out > /dev/null
21.036u 0.051s 0:21.09 99.9%    0+0k 0+0io 0pf+0w
[macbook] dbg_rnflow/pr54717% gfc -c -O2 -ftree-loop-if-convert
cptrf2_inl_1.f90
[macbook] dbg_rnflow/pr54717% gfc rnflow_red.o cptrf2_inl_1.o
[macbook] dbg_rnflow/pr54717% time a.out > /dev/null
27.150u 0.051s 0:27.20 100.0%    0+0k 0+0io 0pf+0w

This shows that the file cptrf2_inl_1.f90 compiled with -ftree-loop-if-convert
gives a slow executable without involving inlining or vectorization.