This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Re: Scheduling fix (fixes gcc.c-torture/execute/980506-1.c on ppc)

  In message <>you write:
  > For the test case, the branch is actually 0% probability, and haifa
  > then mis-schedules the procedure.
  > If the fix below is not acceptable, then you must go through
  > haifa-sched.c and fix every place where it examines the flow graph
  > (that is, everywhere in haifa-sched.c) to allow for blocks like block
  > 2 in the above.  The immediate problem is in compute_dom_prob_ps.
As I mentioned separately, the real fix for this bug was elsewhere.

However, that does not mean we should not try to fix this particular

If you look at the debugging dumps, you'll find that flow does know
that the last block reaches the exit.  This implies that haifa is
the problem, not the cfg provided by flow:

Breakpoint 4, find_rgns (s_preds=0xbffff598, s_succs=0xbffff58c, 
    num_preds=0xbffff580, num_succs=0xbffff574, dom=0x82add90)
    at /puke/law/egcs-1.1//gcc/haifa-sched.c:1475
1475      char no_loops = 1;
(gdb) p dump_bb_data (asm_out_file, s_preds, s_succs)
BB data

BB 0, start 4, end 9
  preds: entry
  succs: 2
BB 1, start 11, end 14
  preds: 2
  succs: 2
BB 2, start 15, end 18
  preds: 1 0
  succs: exit 1

All the necessary information is in the cfg computed by flow; haifa
just isn't using it properly.  So if we're going to fix the branch
probabilities, we need to fix haifa, not flow.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]