RFR: 8312980: C2: "malformed control flow" created during incremental inlining [v6]

[Emanuel Peter]

On Mon, 23 Oct 2023 11:22:54 GMT, Roland Westrelin <roland at openjdk.org> wrote:

>> When `testHelper()` is late inlined, `o` is recorded in the replaced
>> nodes because of the cast to `B` in that method. When late inlining
>> finishes, c2 tries to replace `o` with the resulting `CheckCastPP` to
>> `B` for uses dominated by the call. ``test` and `testHelper2` have 2
>> type checks for A: there are 2 `CheckCastPP` nodes pinned at 2
>> different controls (one is dominated by the `testHelper()` call, one
>> is not), a single `CmpP` that's shared by the 2 type checks (one of
>> the type check is dominated, the other is not). To check if a
>> replacement is legal, for each use of the node to be replaced, the
>> code in `ReplacedNodes::apply()` follow uses and uses of uses until it
>> finds a node that's a CFG node or is pinned. It then uses the IGVN
>> heuristics to figure out if the CFG is dominated by the call or
>> not. If it finds a CFG node that is not dominated, then that use is
>> skipped.
>> 
>> What happens here is that `ReplacedNodes::apply()` checks the control
>> input for each `CheckCastPP` and finds one to be dominated and the
>> other not. So it performs the replacement only for the one that's
>> dominated. For the shared `CmpP`, it follows uses until it reaches the
>> `If` nodes of the type checks. One is dominated. The other is not. So
>> it declares `CmpP` to not be dominated and skips it.
>> 
>> When IGVN runs next, the `CheckCastPP` that had its input replaced now
>> casts a `B` to a `A` which results to top but the check that `o` is of
>> type `A` that guards the type check still tests that `o` of type
>> `Object` is a `B`. That check doesn't constant fold. Replaced nodes
>> introduced an inconsistency. What we would have needed, is for both
>> the `CmpP` and `CheckCastPP` inputs to be changed. But that wasn't
>> possible because the `CmpP` is shared.
>> 
>> The fix I propose is a tweak to `ReplacedNodes::apply()` so it goes
>> depth first over the use of the node to be replaced (let's call it N)
>> and the use of its uses. When it hits a node that's pinned or a CFG,
>> it checks if it is dominated. If that is the case, the chain of use
>> that leads from N is cloned and the clone of the use of N gets the
>> improved input. That way if a node on that path is shared and used in
>> some CFG path that's not dominated, it is unaffected. We then let IGVN
>> cleans up extraneous clones.
>
> Roland Westrelin has updated the pull request incrementally with one additional commit since the last revision:
> 
>   run tests with release build

Tests are passing. Looks better now. Thanks for the updates!
@TobiHartmann should re-review this, or someone else.

src/hotspot/share/opto/replacednodes.cpp line 211:

> 209:     }
> 210:     // Map from current node to cloned/replaced node
> 211:     ResizeableResourceHashtable<Node*, Node*, AnyObj::RESOURCE_AREA, mtCompiler> clones(hash_table_size, hash_table_size);

What arena does this allocate from?

src/hotspot/share/opto/replacednodes.cpp line 232:

> 230:       Node* clone = n->clone();
> 231:       bool added = clones.put(n, clone);
> 232:       assert(added, "");

Suggestion:

      assert(added, "clone node must be added to mapping");

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

Marked as reviewed by epeter (Reviewer).

PR Review: https://git.openjdk.org/jdk/pull/15905#pullrequestreview-1694373379
PR Review Comment: https://git.openjdk.org/jdk/pull/15905#discussion_r1369872274
PR Review Comment: https://git.openjdk.org/jdk/pull/15905#discussion_r1369856258