>
> > BTW: If abs is a define_insn_and_split, pr117722.c still fails
> > because combine can no longer fold abs(a - b) into the old form.
> > So I keep it as an expand.
>
> There shouldn't be any difference WRT combine. Or do you mean
> a
> (define_insn_and_split (
> [(set (match_operand:V_VLSI 0 "register_operand")
> (abs:V_VLSI
> (match_operand:V_VLSI 1 "register_operand")))]?
>
Yeah, i tried
```
(define_insn_and_split (
[(set (match_operand:V_VLSI 0 "register_operand")
(abs:V_VLSI
(match_operand:V_VLSI 1 "register_operand")))]?
```
the dump shows:
/tmp/pr117722-split/pr117722.c.309r.combine
513 (insn 21 20 22 4 (set (reg:RVVMF2HI 159 [ vect_patt_18.18_125 ])
514 (minus:RVVMF2HI (zero_extend:RVVMF2HI (reg:RVVMF4QI 155))
515 (zero_extend:RVVMF2HI (reg:RVVMF4QI 157))))
"gcc/testsuite/gcc.target/riscv/rvv/autovec/pr117722.c":16:27 34245
{*dual_widen_suburvvmf2hi}
516 (expr_list:REG_DEAD (reg:RVVMF4QI 157)
517 (expr_list:REG_DEAD (reg:RVVMF4QI 155)
518 (nil))))
519 (insn 22 21 23 4 (set (reg:RVVMF2HI 160 [ vect_patt_45.20_127 ])
520 (abs:RVVMF2HI (reg:RVVMF2HI 159 [ vect_patt_18.18_125
]))) "gcc/testsuite/gcc.target/riscv/rvv/autovec/pr117722.c":16:15
30864 {absrvvmf2hi2}
521 (expr_list:REG_DEAD (reg:RVVMF2HI 159 [ vect_patt_18.18_125 ])
522 (nil)))
combine still sees `abs`.
but with the `define_expand` version, it sees `neg + smax` sequence,
then rewrite from `neg (a - b)` into `(b - a)`.
