[Bug middle-end/56552] conditional move can generate unnecessary conversion code

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56552

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-03-07 13:32:30 UTC ---
I had to use -O3 -march=loongson2f to trigger it.

During combine, it seems the combiner figures everything out, but mips.md lacks
needed conditional move patterns:
(insn 36 35 37 2 (set (reg:SI 298)
        (if_then_else:SI (eq:SI (reg:SI 299)
                (const_int 0 [0]))
            (reg/v:SI 194 [ a2+-2 ])
            (reg/v:SI 213 [ a2+-2 ]))) 602 {*movsi_on_si}
     (expr_list:REG_DEAD (reg:SI 299)
        (expr_list:REG_DEAD (reg/v:SI 213 [ a2+-2 ])
            (expr_list:REG_DEAD (reg/v:SI 194 [ a2+-2 ])
                (nil)))))
(insn 37 36 40 2 (set (reg/v:SI 214 [ a2+-2 ])
        (zero_extend:SI (subreg:HI (reg:SI 298) 2))) 188 {*zero_extendhisi2}
     (expr_list:REG_DEAD (reg:SI 298)
        (nil)))

and combiner has:
Trying 36 -> 37:
Failed to match this instruction:
(set (reg/v:SI 214 [ a2+-2 ])
    (if_then_else:SI (reg:SI 299)
        (reg/v:SI 213 [ a2+-2 ])
        (reg/v:SI 194 [ a2+-2 ])))

I.e. it figures that the masking isn't needed, but during
simplification/canonicalization transforms that (ne:SI (reg:SI 299) (const_int
0)) in first argument of IF_THEN_ELSE into (reg:SI 299).
If mips.md had patterns that accepted for movcc the first argument of
IF_THEN_ELSE being a register_operand of GPR mode as if it was (ne:GPR (that
register) (const_int 0)), the testcase would be fixed.