Out-of-sight compile times for calculate_loop_depth

Michael Hayes m.hayes@elec.canterbury.ac.nz
Mon Feb 7 15:11:00 GMT 2000


Brad Lucier writes:

 > http://www.math.purdue.edu/~lucier/_meroon.i.gz

Is this some machine generated program?  I guess it has lots of loops
with a large number of edges in the CFG from looking at the profiling
info.  I'll grab it later on when I've got a spare minute.

 > All routines with more than 5% runtime are:
 > 
 > Each sample counts as 0.000976562 seconds.
 >   %   cumulative   self              self     total           
 >  time   seconds   seconds    calls  ms/call  ms/call  name    
 >  36.66    104.63   104.63    59615     1.76     1.76  sbitmap_intersection_of_preds
 >  21.87    167.07    62.44 21551141     0.00     0.00  sbitmap_a_subset_b_p
 >   6.41    185.37    18.30     6596     2.77     2.77  flow_loop_exits_find
 >   5.15    200.08    14.71     6596     2.23     2.23  flow_loop_pre_header_find


This testcase exposes problems with the dominator calculation
algorithm and with flow_loop_node_add.  The latter I never
particularly liked since it can exhibit quadratic behaviour.  As an
interim solution, adding the loops to the loop tree in reverse order
should reduce the insertion time.

Michael.




More information about the Gcc mailing list