With Valhalla, we see the following Phi/Mergemem pattern that is not folded and blocks optimizations:

287 MergeMem === _ 1 78 1 1 1 1 1 1 1 1 1 1 1 7 [[ 250 ]] { - - - - - - - - - - - N7:ConditionalTest$BoxInterface:NotNull:exact+12 *,iid=181 [narrow] } Memory: @BotPTR *+bot, idx=Bot; !jvms: ConditionalTest$BoxInterface::<init> @ bci:6 ConditionalTest$BoxInterface::from @ bci:10 ConditionalTest::test3 @ bci:18
 215 MergeMem === _ 1 78 1 1 1 1 1 1 1 1 1 1 1 7 [[ 250 ]] { - - - - - - - - - - - N7:ConditionalTest$BoxInterface:NotNull:exact+12 *,iid=181 [narrow] } Memory: @BotPTR *+bot, idx=Bot; !jvms: ConditionalTest$BoxInterface::<init> @ bci:1 ConditionalTest$BoxInterface::from @ bci:10 ConditionalTest::test3 @ bci:18
 250 Phi === 247 215 287 [[ 206 516 ]] #memory Memory: @BotPTR *+bot, idx=Bot; !orig=[428],[433] !jvms: ConditionalTest$BoxInterface::<init> @ bci:6 ConditionalTest$BoxInterface::from @ bci:10 ConditionalTest::test3 @ bci:18

With JDK-8246603, I've enabled splitting of Phi's through mergemems if all inputs are mergemems. We need to revisit if this is correct.

TestLWorld::test109 should trigger this and IR verification fails without splitting phis through mergemems.