[Bug target/109855] [14 Regression] ICE: in curr_insn_transform, at lra-constraints.cc:4231 unable to generate reloads for {aarch64_mlav4hi_vec_concatz_le} at -O1

["ktkachov at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109855

--- Comment #6 from ktkachov at gcc dot gnu.org ---
(In reply to ktkachov from comment #5)
> (In reply to rsandifo@gcc.gnu.org from comment #4)
> > I guess the problem is that the define_subst output template has:
> > 
> >   (match_operand:<VDBL> 0)
> > 
> > which creates a new operand 0 with an empty predicate and constraint,
> > as opposed to a (match_dup 0), which would be substituted with the
> > original operand 0.  Unfortunately
> > 
> >   (match_dup:<VDBL> 0)
> > 
> > doesn't work as a way of inserting the original destination with
> > a different mode, since the :<VDBL> is ignored.  Perhaps we should
> > “fix” that.  Alternatively:
> > 
> >   (match_operand:<VDBL> 0 "register_operand" "=w")
> > 
> > should work, but probably locks us into using patterns that have one
> > alternative only.
> 
> I think this approach is the most promising and probably okay for the vast
> majority of cases we want to handle with these substs.

Interestingly, it does seem to do the right thing for multi-alternative
patterns too. For example:
(define_insn ("aarch64_cmltv4hf_vec_concatz_le")
     [
        (set (match_operand:V8HI 0 ("register_operand") ("=w,w"))
            (vec_concat:V8HI (neg:V4HI (lt:V4HI (match_operand:V4HF 1
("register_operand") ("w,w"))
                        (match_operand:V4HF 2 ("aarch64_simd_reg_or_zero")
("w,YDz"))))
                (match_operand:V4HI 3 ("aarch64_simd_or_scalar_imm_zero")
(""))))
    ] ("(!BYTES_BIG_ENDIAN) && ((TARGET_SIMD) && (TARGET_SIMD_F16INST))") ("@
  fcmgt\t%0.4h, %2.4h, %1.4h
  fcmlt\t%0.4h, %1.4h, 0")
     [
        (set_attr ("type") ("neon_fp_compare_s"))
        (set_attr ("add_vec_concat_subst_le") ("no"))
    ])