# Issue A few test failed intermittently with -Xcomp on Mac due to an unschedulable graph in C2: - valhalla/valuetypes/WeakReferenceTest.java - valhalla/valuetypes/ProxyTest.java - valhalla/valuetypes/ObjectNewInstance.java - valhalla/valuetypes/ObjectMethodsViaCondy.java
# Causes The origin of the issue seems to be coming from strength reduction (`process_late_inline_calls_no_inline`) where we replace virtual and MH calls with direct calls. https://github.com/dafedafe/valhalla/blob/75e2dd95df5d847d7d6e35a23016d22705681cf4/src/hotspot/share/opto/compile.cpp#L3072 If the return type of the methods are not loaded, we add a call to runtime's `store_inline_type_fields_to_buf` right after the actual method call, to save the scalarized return into a oop. This happens first for the caller at parse time and then for the callee when strength-reducing the virtual call to a direct one. The return projections of the inline fields of the call are added to `store_inline_type_fields_to_buf` and its oop projection is then added as input to the other `store_inline_type_fields_to_buf` which fundamentally leaves the graph around it in an awkward state. If this happens in combination with loop unswitching it can lead to a graph not being schedulable, which is what happens in the failure of this issue, where we have: * a virtual call with a following `store_inline_type_fields_to_buf` (1) in a loop. * the loop undergoes unswitching, which creates 2 copies of the body (including a copy of the virtual call). All outputs of the new virtual call are phi-d with the one of the other path as input of the `store_inline_type_fields_to_buf`. * the new copy of the virtual call is later replaced with a direct call: the creation of the new direct call adds a `store_inline_type_fields_to_buf` (2) right after the direct call. All the inline type return values of the call being inlined are removed, so the phis now only have one input and are removed by a later GVN pass. <img width="600" alt="Screenshot 2025-11-26 at 10 33 51" src="https://github.com/user-attachments/assets/35e947c2-350e-414f-9d44-72559d480a88"; /> * this creates an issue later during GCM since the `store_inline_type_fields_to_buf` (1) call is logically not dominated by any of the 2 sides of the unswitched loop and lands in a separate dominance path of the arguments whose phis have been removed (which are still dominated by the original virtual call). # Solution The issue happens only when strength-reducing to a direct call and the return type is not loaded. So, there seems to be 2 conceivable fixes: either we avoid strength-reduction if the type is not loaded or we avoid creating the second `store_inline_type_fields_to_buf` and rewire the output projections of the strength-reduced call to the output of the projections of the original virtual call. By choosing the first solution we potentially skip a useful optimization (in particular if the return value is always null, which won't trigger the deoptimization). So, the rewiring solution seems to be the best option in this case. # Testing * repeated testing with failing tests (same conditions) * replay of failing compilations * Tier 1-3 Unfortunately creating a regression test that would consistently trigger the specific issue with unloaded return type and strength-reduction to direct call has proven to be unviable. ------------- Commit messages: - JDK-8369045: remove new line - Revert "JDK-8369045: add bool for rt_loaded" - JDK-8369045: add comment - JDK-8369045: add bool for rt_loaded - JDK-8369045: replace result projections instead of no strength reduction - JDK-8369045: [lworld] valhalla/valuetypes/WeakReferenceTest.java has an unschedulable graph Changes: https://git.openjdk.org/valhalla/pull/1768/files Webrev: https://webrevs.openjdk.org/?repo=valhalla&pr=1768&range=00 Issue: https://bugs.openjdk.org/browse/JDK-8369045 Stats: 49 lines in 3 files changed: 28 ins; 3 del; 18 mod Patch: https://git.openjdk.org/valhalla/pull/1768.diff Fetch: git fetch https://git.openjdk.org/valhalla.git pull/1768/head:pull/1768 PR: https://git.openjdk.org/valhalla/pull/1768
