RFR: 8271341: Opcode() != Op_If && Opcode() != Op_RangeCheck) || outcnt() == 2 assert failure with Test7179138_1.java

[Vladimir Kozlov]

On Fri, 30 Jul 2021 08:27:47 GMT, Roland Westrelin <roland at openjdk.org> wrote:

> The root cause of this bug is in PhaseStringOpts. In the middle of the
> chain of calls that are optimized out, there's a diamond Region/If. On
> most executions this diamond is optimized out by IGVN because once
> PhaseStringOpts is over, all the Region's Phis are removed. But
> because one input of the If/Bool/Cmp is set to top by PhaseStringOpts
> when calls are removed, it sometimes happen that top propagates to the
> If before the Region is optimized out. That causes control flow below
> the If to become dead while it should still be reachable.
> 
> The fix I propose is to have PhaseStringOpts removed the Region/If in
> that case.

src/hotspot/share/opto/stringopts.cpp line 282:

> 280:     } else if (n->is_Region()) {
> 281:       Node* iff = n->in(1)->in(0);
> 282:       assert(iff->is_If(), "no if for the diamond");

I assume that the only Region node listed in _control list is diamond Region.
May be add assert to check that it is diamond region. Up to you.
I would like to have separate RFE to have `is_diamond_region()` - we seems check it in few places.

-------------

PR: https://git.openjdk.java.net/jdk/pull/4944