[Bug tree-optimization/80237] float to double conversion is not optimized away

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80237

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2017-03-29
          Component|c                           |tree-optimization
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
We have (after inlining g into foo):

foo (float x)
{
  float _3;
  float _4;
  double iftmp.0_5;
  double _6;
  double iftmp.0_7;
  double iftmp.0_8;

  <bb 2> [100.00%]:
  if (x_2(D) > 0.0)
    goto <bb 3>; [60.23%]
  else
    goto <bb 4>; [39.77%]

  <bb 3> [60.23%]:
  _3 = f (x_2(D));
  iftmp.0_5 = (double) _3;
  goto <bb 5>; [100.00%]

  <bb 4> [39.77%]:
  _6 = (double) x_2(D);
  iftmp.0_7 = _6 + 1.0e+0;

  <bb 5> [100.00%]:
  # iftmp.0_8 = PHI <iftmp.0_5(3), iftmp.0_7(4)>
  _4 = (float) iftmp.0_8;
  return _4;

thus the tailcall is not exposed at GIMPLE level.

For the above we fail to optimize the partial redundancy of the conversion
in PRE.  During phi translation we do see that _4 is equal to _3 on the
3->5 edge but phi translation fails because _3 is not antic-in in bb 5.
Not really relevant, I think the check is somewhat bogus.  With that fixed
we do detect the tailcall:

  <bb 2> [100.00%]:
  if (x_2(D) > 0.0)
    goto <bb 3>; [60.23%]
  else
    goto <bb 4>; [39.77%]

  <bb 3> [60.23%]:
  _3 = f (x_2(D)); [tail call]
  goto <bb 5>; [100.00%]

  <bb 4> [39.77%]:
  _6 = (double) x_2(D);
  iftmp.0_7 = _6 + 1.0e+0;
  _9 = (float) iftmp.0_7;

  <bb 5> [100.00%]:
  # prephitmp_10 = PHI <_3(3), _9(4)>
  return prephitmp_10;

foo:
.LFB2:
        .cfi_startproc
        ucomiss .LC0(%rip), %xmm0
        jbe     .L7
        jmp     f
        .p2align 4,,10
        .p2align 3
.L7:
        cvtss2sd        %xmm0, %xmm0
        addsd   .LC1(%rip), %xmm0
        cvtsd2ss        %xmm0, %xmm0
        ret

Mine.