Steven Bosscher wrote:
Notice how the conditional sets of r14 and r17 in insns 9 and 10 have
been moved past insn 14, which uses these registers. Shouldn't there
be true dependencies on insns 9 and 10 for insn 14?
In theory, yes. But the scheduler (it is ebb scheduler as I understand)
frequently removes dependencies creating a barrier and all dependencies
to the insns before the barrier are changed by dependencies to the
barrier (in this case the barrier is the jump insn and dependency 14->9
is changed by 14->the jump). The code should work ok if there is no
movement down through jump insn (it is how the scheduler, ebb or
interblock one, should work). But the scheduler did it for insn #9.
I'll look at this PR today.
Vlad
