[Bug middle-end/57503] [4.7/4.8 Regression] Expand uses wrong multiply routine

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57503

--- Comment #6 from Georg-Johann Lay <gjl at gcc dot gnu.org> ---
Created attachment 30252
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=30252&action=edit
.165r.expand: dump file with -Os -mmcu=atmega168 -fno-expensive-optimizations

The tree-ssa part for widening_mul can be disabled by
-fno-expensive-optimizations, but the code is still wrong.  Cf. the attached
.expand dump of

long
func1 (unsigned char a, unsigned char b, int c)
{
    unsigned ab = a * b;
    return (long) ab * c;
}

ab is computed in insn 10 with zero-extended inputs:

(insn 8 5 9 2 (set (reg:HI 53 [ D.1447 ])
        (zero_extend:HI (reg/v:QI 49 [ a ]))) cecky.c:4 -1
     (nil))
(insn 9 8 10 2 (set (reg:HI 54 [ D.1447 ])
        (zero_extend:HI (reg/v:QI 50 [ b ]))) cecky.c:4 -1
     (nil))
(insn 10 9 11 2 (set (reg:HI 55 [ D.1447 ])
        (mult:HI (reg:HI 53 [ D.1447 ])
            (reg:HI 54 [ D.1447 ]))) cecky.c:4 -1

This is correct, but then in insn 11 ab gets sign-extended even though it is
unsigned:

(insn 11 10 12 2 (set (reg:SI 56 [ D.1448 ])
        (sign_extend:SI (reg:HI 55 [ D.1447 ]))) cecky.c:5 -1
     (nil))

Insn combine combines this wrong extension into a widening multiplication.