[Bug tree-optimization/90510] [10 Regression] Unnecessary permutation

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90510

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
So we have

(insn 10 9 11 2 (set (reg:V2DF 92)
        (vec_duplicate:V2DF (reg/v:DF 84 [ tem ]))) "t.c":8:42 2984
{vec_dupv2df}
     (expr_list:REG_DEAD (reg/v:DF 84 [ tem ])
        (nil)))
(insn 11 10 16 2 (set (reg:V2DF 91)
        (vec_merge:V2DF (reg:V2DF 92)
            (reg/v:V2DF 87 [ x ])
            (const_int 1 [0x1]))) "t.c":8:10 2983 {sse2_movsd}
     (expr_list:REG_DEAD (reg:V2DF 92)
        (expr_list:REG_DEAD (reg/v:V2DF 87 [ x ])
            (nil))))

and nothing figures out insn 10 is unnecessary because the upper half of 92
is dead.  I somehow expected combine/simplify-rtx to merge these but I
guess without target support it is hard to see that sse2_movsd can handle
sth like a paradoxical vector subreg of the scalar:

(insn 11 10 16 2 (set (reg:V2DF 91)
        (vec_merge:V2DF (subreg:V2DF (reg:DF 84) 0)
            (reg/v:V2DF 87 [ x ])
            (const_int 1 [0x1]))) "t.c":8:10 2983 {sse2_movsd}
     (expr_list:REG_DEAD (reg:V2DF 92)
        (expr_list:REG_DEAD (reg/v:V2DF 87 [ x ])
            (nil))))


Now we can pattern-match both

  _4 = BIT_FIELD_REF <x_7(D), 64, 64>;
  z_6 = {_3, _4};

and

  _7 = {_3, _3};
  z_6 = VEC_PERM_EXPR <x_4(D), _7, { 2, 1 }>;

to

  z_6 = BIT_INSERT_EXPR <x_4(D), _3, 0>

which we can hope to be expanded better.