In JDK mainline, AOTConstantPoolResolver::preresolve_class_cp_entries() always tries to resolve the entries specified in preresolve_list(). We rely on AOTConstantPoolResolver::is_resolution_deterministic() to decide whether some of those preresolved entries need to be reverted back to the unresolved state.

However, in the premain branch, this check is added for some unspecified reasons:

https://github.com/openjdk/leyden/blob/478f86f9cd6df6b92c037c83d0540b9c5fe97e5c/src/hotspot/share/cds/aotConstantPoolResolver.cpp#L239-L246

  if (!CDSConfig::is_dumping_aot_linked_classes()) {
    // TODO: Why is this check needed in Leyden?
    // The following 3 tests fails when this "if" check is removed (when -XX:-AOTClassLinking is NOT enabled)
    // - runtime/cds/appcds/methodHandles/MethodHandlesAsCollectorTest.java
    // - runtime/cds/appcds/methodHandles/MethodHandlesGeneralTest.java
    // - runtime/cds/appcds/methodHandles/MethodHandlesSpreadArgumentsTest.java
    return;
  }


This causes fewer CP entries to be resolved in CDS archive when AOTClassLinking is false.

=========

The reason for the failure is probably because AOTConstantPoolResolver::is_resolution_deterministic is incorrect when AOTClassLinking is false.