This came up in shenandoah testing with XX:+ExtendedDTraceProbes.

make_runtime_call() is called through make_dtrace_method_exit() from Parse::return_current(). Memory state at this point is:

 137 Phi === 135 _ _ 91 [[ 74 141 145 150 152 162 166 168 179 182 187 193 202 211 216 225 228 237 242 258 266 274 282 ]] #memory Memory: @BotPTR *+bot, idx=Bot; !jvms: IdentityHashMap::put @ bci:24
 141 MergeMem === _ 1 137 1 1 279 1 275 282 [[ 142 ]] { - - N279:java/lang/Object+-8 * - N275:narrowoop: java/lang/Object *[int:>=0]+-8 * N282:narrowoop: java/lang/Object *[int:>=0]+any * [narrow] } Memory: @BotPTR *+bot, idx=Bot; !jvms: IdentityHashMap::put @ bci:24

The Phi is a loop phi so not all its inputs are set yet. The following code in GraphKit::make_runtime_call():
    assert(!wide_out, "narrow in => narrow out");
    Node* narrow_mem = memory(adr_type);
    prev_mem = reset_memory();
    map()->set_memory(narrow_mem);

set the entire memory state to the phi. Next in GraphKit::set_predefined_input_for_runtime_call():

  Node* memory = reset_memory();

causes the current memory state (the Phi) to be transformed which the GVN transforms to:
 91 Phi === 89 _ _ 73 [[ 137 100 103 105 113 116 118 126 129 131 ]] #memory Memory: @BotPTR *+bot, idx=Bot; !jvms: IdentityHashMap::put @ bci:24

the out of loop memory state and so the wrong state.