RFR: 8315920: C2: "control input must dominate current control" assert failure

[Christian Hagedorn]

On Wed, 13 Sep 2023 14:35:55 GMT, Roland Westrelin <roland at openjdk.org> wrote:

> The assert fires during range check eliminationin code that was added
> recently. It catches a bug that's been going unnoticed. The `array[0]`
> load of the test case ends up on the exit of the loop above it. After
> pre/main/post loops are created for this loop (the inner loop is
> optimized out), the load is at a region merging control from the just
> created pre/main/post loops but its control (as returned by
> `PhaseIdealLoop::get_ctrl()`) is not updated by pre/main/post loops
> creation code so it's still the exit projection of one of the
> loops. This causes the assert to fire (the load has bad control) when
> RC runs next in the same loop opts pass.
> 
> The current logic for pre/main/post loops creation does update the
> control of nodes on an exit projection but only if the node is
> referenced from some other node in the loop body.
> 
> I think a similar bug with a node assigned control at the exit but not
> pinned is likely to exist. So I doubt simply going over nodes pinned
> at exits during pre/main/post and updating their control is good
> enough. What I propose instead is to `lazy_replace()` the projection
> by the region so that, for all nodes that are assigned the exit as
> control, `get_ctrl()` will then return the region. It's not entirely
> straightforward to implement because the `lazy_replace()` mechanism
> assumes the control being replaced is dead which is not the case
> here. That's why in the patch the exit projection is cloned so the
> existing one is dead and can be used in `lazy_replace()`. This also
> makes the calls to `sink_use()` redundant.

Looks good to me, too.

src/hotspot/share/opto/loopopts.cpp line 2133:

> 2131:       Node *prev = use->is_CFG() ? use : get_ctrl(use);
> 2132:       assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
> 2133:       Node *cfg = (prev->_idx >= new_counter && prev->is_Region())

Suggestion:

      Node* cfg = (prev->_idx >= new_counter && prev->is_Region())

src/hotspot/share/opto/loopopts.cpp line 2617:

> 2615: 
> 2616:         assert(use->is_Proj(), "loop exit should be projections");
> 2617:         Node* use_clone = use->clone();

Maybe also add a comment here why you need to clone the node for `lazy_replace()`.

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

Marked as reviewed by chagedorn (Reviewer).

PR Review: https://git.openjdk.org/jdk/pull/15720#pullrequestreview-1635712463
PR Review Comment: https://git.openjdk.org/jdk/pull/15720#discussion_r1331677014
PR Review Comment: https://git.openjdk.org/jdk/pull/15720#discussion_r1331675100