https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113114
--- Comment #6 from Alex Coplan <acoplan at gcc dot gnu.org> ---
Hmm, it's worth noting that the ILP32 case is a bit different, though, in that
we have:
(rr) call debug (insn->rtl ())
(insn 16 21 19 3 (parallel [
(set (reg:DF 62 v30)
(unspec:DF [
(mem:V2x8QI (reg/v/f:DI 0 x0 [orig:123 a ] [123]) [0 +0
S16 A64])
] UNSPEC_LDP_FST))
(set (reg:DF 63 v31)
(unspec:DF [
(mem:V2x8QI (reg/v/f:DI 0 x0 [orig:123 a ] [123]) [0 +0
S16 A64])
] UNSPEC_LDP_SND))
]) 88 {*load_pair_8}
(nil))
(rr) call debug (trailing_add->rtl ())
(insn 20 18 41 3 (set (reg:SI 0 x0 [orig:118 ivtmp.22 ] [118])
(plus:SI (reg:SI 0 x0 [orig:123 a ] [123])
(const_int 8 [0x8]))) 119 {*addsi3_aarch64}
(nil))
i.e. x0 appears as DImode in the load pair addresses but the trailing update is
done in SImode, which means we end up not matching when forming the final
pattern.
I don't think either case is particularly interesting, so I'm leaning towards
just bailing out if recog fails in the pass (in which case both of these just
become missed-optimizations).