[Bug tree-optimization/79725] Sinking opportunity missed if blocked by dead stmt

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79725

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2017-02-27
          Component|c                           |tree-optimization
            Summary|Sinking opportunity missed  |Sinking opportunity missed
                   |if complex type is changed  |if blocked by dead stmt
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
We seem to be quite unlucky with initial IL:

  p = __complex__ (1.0e+0, 0.0);
  {
    int i;

    i = 0;
    goto <D.2520>;
    <D.2519>:
    _1 = (long unsigned int) i;
    _2 = _1 * 16;
    _3 = x + _2;
    D.2523 = *_3;
    _4 = REALPART_EXPR <D.2523>;
    _5 = (float) _4;
    _6 = IMAGPART_EXPR <D.2523>;
    _7 = (float) _6;
    p = COMPLEX_EXPR <_5, _7>;
    i = i + 1;
    <D.2520>:
    if (i <= 999999) goto <D.2519>; else goto <D.2521>;
    <D.2521>:
  }
  p.0 = p;
  _8 = REALPART_EXPR <p.0>;
  _9 = (double) _8;
  _10 = IMAGPART_EXPR <p.0>;
  _11 = (double) _10;
  D.2524 = COMPLEX_EXPR <_9, _11>;
  return D.2524;

this causes a PHI node for p inside the loop and the uses partly vanish
after threading.  But before sinking we still have a dead expression that
blocks the sinking because we do

static bool
statement_sink_location (gimple *stmt, basic_block frombb,
                         gimple_stmt_iterator *togsi)
{
...
  /* Return if there are no immediate uses of this stmt.  */
  if (has_zero_uses (DEF_FROM_PTR (def_p)))
    return false;

it's understandable we don't want to deal with dead code but in this case
it's "premature compile-time optimization" ;)  (OTOH we just crash if
we remove that check).

PRE figured out the redundancy in

  _5 = (float) _4;
  _7 = (float) _6;
  p_18 = COMPLEX_EXPR <_5, _7>;
  i_19 = i_24 + 1;
  if (i_19 != 1000000)
    goto <bb 5>; [98.99%]
  else
    goto <bb 4>; [1.01%]

  <bb 5> [98.00%]:
  goto <bb 3>; [100.00%]

  <bb 4> [1.00%]:
  _9 = (double) _5;
  _11 = (double) _7;
  _14 = COMPLEX_EXPR <_9, _11>;

and made p_18 unused but left the dead p_18 around (it's not its job to do the
DCE).

If you change complex float to complex double the initial IL doesn't contain
the decomposition and the situation is much easier.

I'll have a look to mitigate this in sinking for GCC 8.