This way they get lowered through the ARMISD::BUILD_VECTOR, which can produce more efficient D register moves. Also helps D115653 not get stuck in a loop.
FoldTwoEntryPHINode()