https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89474
--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> --- I admit I don't know much about df infrastructure, but it looks completely wrong to me to first modify the various instructions to refer to at that point created, but uninitialized v4sf_const0 pseudo, including df_insn_rescan etc., then df_analyze including DF_DU_CHAIN | DF_UD_CHAIN etc. and finally add an initializer for that pseudo. I'd say df_analyze shouldn't be called if the IL is in so inconsistent state.
