[Bug rtl-optimization/47477] [4.7/4.8/4.9 regression] Sub-optimal mov at end of method

[ktietz at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47477

--- Comment #17 from Kai Tietz <ktietz at gcc dot gnu.org> ---
What optimization you expect here?  I see by the new type-demotion pass some
changes in optimized tree-output:

foo ()
{
  int i;
  short int _4;
  char _5;
  unsigned short _6;
  unsigned short _8;
  short int _9;
  unsigned short _10;
  unsigned short _11;
  short int _12;
  sizetype _25;

  <bb 2>:
  goto <bb 4>;

  <bb 3>:

  <bb 4>:
  # i_17 = PHI <i_14(3), 0(2)>
  _25 = (sizetype) i_17;
  _4 = MEM[symbol: a, index: _25, step: 2, offset: 0B];
  _5 = (char) _4;
  _6 = (unsigned short) _5;
  _9 = MEM[symbol: b, index: _25, step: 2, offset: 0B];
  _8 = (unsigned short) _9;
  _10 = _8 + 17;
  _11 = _10 + _6;
  _12 = (short int) _11;
  MEM[symbol: a, index: _25, step: 2, offset: 0B] = _12;
  i_14 = i_17 + 1;
  if (i_14 != 1024)
    goto <bb 3>;
  else
    goto <bb 5>;

  <bb 5>:
  return;
}

what then gets simplified to the following assembler on IA32:
_foo:
        xorl    %eax, %eax
        .p2align 4,,10
L2:
        movsbw  _a(%eax,%eax), %dx
        movzwl  _b(%eax,%eax), %ecx
        leal    17(%ecx,%edx), %edx
        movw    %dx, _a(%eax,%eax)
        addl    $1, %eax
        cmpl    $1024, %eax
        jne     L2
        rep ret

The same assembler gets produced for my with all compilers back to 4.6.0, just
tree-optimization output differs.