https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120447
ktkachov at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed|2025-05-27 00:00:00 |2025-5-30
CC| |ktkachov at gcc dot gnu.org,
| |rsandifo at gcc dot gnu.org
--- Comment #4 from ktkachov at gcc dot gnu.org ---
The crash through aarch64_emit_load_store_through_mode happens when to do a
move from:
(mem/u/c:VNx4QI (reg/f:DI 119) [0 S[4, 4] A32])
to
(reg:VNx4QI 115)
through a QImode subreg.
In the code aarch64_emit_load_store_through_mode
if (MEM_P (src))
{
rtx tmp = force_reg (new_mode, adjust_address (src, new_mode, 0));
tmp = force_lowpart_subreg (int_mode, tmp, new_mode);
emit_move_insn (dest, force_lowpart_subreg (mode, tmp, int_mode));
}
The line:
tmp = force_lowpart_subreg (int_mode, tmp, new_mode);
returns NULL_RTX which crashes the emit_move_insn.
>From what I can tell the validate_subreg check inside the eventual callee
inside the force_lowpart_subreg chain returns false when validating a VNx4QI
subreg of reg:QI. This is because the [4, 4] size of VNx4QI is not ordered_p
wrt REGMODE_NATURAL_SIZE (QImode), which is 8.
I guess we could do such checks in aarch64_expand_maskloadstore before calling
aarch64_emit_load_store_through_mode but we should be able to make this
optimisation for QImode so is there a way perhaps to allow this?
