Fwd: [PATCH] Scheduling result adjustment to enable macro-fusion

[Wei Mi]

On Mon, Nov 25, 2013 at 2:08 AM, Alexander Monakov <amonakov@ispras.ru> wrote:
> On Sat, 23 Nov 2013, Wei Mi wrote:
>> For the failed testcase, it was compiled using -fmodulo-sched.
>> modulo-sched phase set SCHED_GROUP_P of a jump insn to be true, which
>> means the jump insn should be scheduled with prev insn as a group.
>
> SMS doesn't set SCHED_GROUP_P by itself; did you mean that SCHED_GROUP_P is
> set by dependency analysis code similar to sched2?
>

SCHED_GROUP_P is set in sched_analyze for "call + return value" group
and other groups, and in sched_init for macrofusion. Both
sched_analyze and sched_init are used by SMS (sched_analyze is used by
creating ddg). I think sched1 may have the same problem when it set
SCHED_GROUP_P and sched2 uses it.

>> When modulo scheduling is finished, the flag of SCHED_GROUP_P is not
>> cleaned up. After that, pass_jump2 phase split the bb and move the
>> prev insn to another bb. Then pass_sched2 see the flag and mistakenly
>> try to bind the jump insn with a code label.
>
> I think the analysis is incomplete.  Looking at the backtrace posted in the
> bug report, the failure path goes through chain_to_prev_insn, which protects
> against such failure:
>
>   prev_nonnote = prev_nonnote_nondebug_insn (insn);
>   if (BLOCK_FOR_INSN (insn) == BLOCK_FOR_INSN (prev_nonnote)
>       && ! sched_insns_conditions_mutex_p (insn, prev_nonnote))
>     add_dependence (insn, prev_nonnote, REG_DEP_ANTI);
>
> Why does it end up with a label at the assertion failure point?
>
> Alexander

Because code label is not a note or debug insn.

I think it is impossible to detect such inconsistency in
chain_to_prev_insn. If the prev_nonnote is not a code label, the bug
will not be exposed this time. Suppose some other optimizations insert
a real insn before the jump marked as SCHED_GROUP_P, following
scheduler pass will schedule them together silently. That is why I
think it is necessary to cleanup SCHED_GROUP_P when a scheduling pass
is finished.

Thanks,
Wei.