[Bug tree-optimization/71261] [7 Regression] Trunk GCC hangs on knl and broadwell targets

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71261

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
I have a simple patch improving it to

 tree loop if-conversion :   3.16 (52%) usr   0.82 (59%) sys   3.98 (53%) wall 
315828 kB (79%) ggc
 rest of compilation     :   2.49 (41%) usr   0.67 (49%) sys   2.97 (40%) wall 
242780 kB (61%) ggc
 unaccounted todo        :   0.02 ( 0%) usr   0.23 (17%) sys   0.25 ( 3%) wall 
 49155 kB (12%) ggc
 TOTAL                 :   6.11             1.38             7.50            
400732 kB

the most time is spent in immediate use walking in ifcvt_split_def_stmt while
it looks only for uses on a single stmt.  That can be sped up by walking
uses on that stmt only.

It's still slow, so a bigger reproducer is still appreciated.