Integrated: 8319764: C2 compilation asserts during incremental inlining because Phi input is out of bounds

[Roland Westrelin]

On Tue, 14 Nov 2023 08:56:09 GMT, Roland Westrelin <roland at openjdk.org> wrote:

> In the test case:
> 
> `lateInlined1()` and `lateInlined2()` are inlined incrementally
> because they are invoked wih a method handle which is known constant
> only after igvn. For the failure to reproduce, `lateInlined2()` must
> be inlined after `lateInlined1()` which depends on igvn processing
> order (that's why the test needs `StressIGVN`). `lateInlined1()` has 2
> virtual calls that are inlined with bimorphic inlining. At each call
> site, exception state is merged from `A.m()` and `B.m()`, the 2
> methods that are inlined. `c` is casted to non null in `A.m()`. At the
> first call site, `c` is live and merging the exception states causes
> the creation of a `Phi` for `c` that merges `c` and the cast to non
> null of `c`. At the second call site, `c` is not live so it's ignored
> when merging exception states. When the exception states for the 2
> call sites are combined, the region that merges the states at the
> first call site is extended to have 4 inputs but the `Phi` for `c` at
> the first call site is left with 2 inputs. When `lateInlined2()` is
> inlined, it records a replaced node for a cast of `c` to non null. On
> completion of incremental inlining, that replaced nodes is "applied".
> In the process, the dead `Phi` for `c` created at the previous
> incremental inline is found and because it has less inputs that its
> region, the crash occurs.
> 
> The fix I propose is to simply ignore phi nodes that don't have the
> same number of inputs as their region because they can only be a dead
> Phi.

This pull request has now been integrated.

Changeset: f33c874b
Author:    Roland Westrelin <roland at openjdk.org>
URL:       https://git.openjdk.org/jdk/commit/f33c874b6e624ad81572a2f806e198dd692a31a6
Stats:     141 lines in 2 files changed: 135 ins; 0 del; 6 mod

8319764: C2 compilation asserts during incremental inlining because Phi input is out of bounds

Reviewed-by: thartmann, chagedorn

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

PR: https://git.openjdk.org/jdk/pull/16648