[Bug rtl-optimization/40101] [4.5 Regression] 200.sixtrack ICEs in get_seqno_by_preds, at sel-sched-ir.c:3752

[abel at gcc dot gnu dot org]

------- Comment #2 from abel at gcc dot gnu dot org  2009-05-15 13:20 -------
The bug happens when we compute a seqno for the newly created bookkeping insn. 
Seqnos exist in the algorithm (roughly) to guide the movement of fences so
that, first, all unscheduled insns will be found and scheduled, and second,
"dynamic" back-edges during pipelining will be properly marked.  

The assert we hit just means that the existing code was unable to compute the
proper seqno, because it looks only on direct succs/preds of the place we want
to insert bookkeeping, yet in the test case the fences are processed in such an
order that all of those succs/preds are already scheduled (we're pipelining). 
The fix would be to extend this code to walk further to find yet unscheduled
insns and to derive the seqno from them.  

However, thinking about it further, I believe that we can get rid of seqnos
altogether.  We already have topological sorting for basic blocks in the
scheduler, and it should be enough just to check that we don't hit other fence
or already visited insn (or already scheduled insn if no pipelining) when
computing our availability sets, and it should be enough just to move fences to
unscheduled code.  The original approach also used seqnos to mark parallel
groups, but this is not needed for gcc.  So, as we are in Stage 1, and there is
no rush with this bug, I will try to get rid of seqnos first.  In case it turns
out I've overlooked anything, I will resort to the former, more conservative
fix.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40101