[Bug tree-optimization/71347] [7 regression] Performance drop after r235513 on x86-64 in 32-bit mode.

[amker at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71347

--- Comment #2 from amker at gcc dot gnu.org ---
Thanks for reporting this.

The dump after IVOPT now is:


  <bb 3>:
  # prephitmp_21 = PHI <pretmp_20(4), powmult_2(2)>
  # prephitmp_23 = PHI <pretmp_22(4), powmult_2(2)>
  # ivtmp.17_16 = PHI <ivtmp.17_9(4), ivtmp.17_24(2)>
  _6 = prephitmp_21 * prephitmp_23;
  _4 = (void *) ivtmp.17_16;
  MEM[base: _4, offset: 0B] = _6;
  ivtmp.17_9 = ivtmp.17_16 + 8;
  if (ivtmp.17_9 != _26)
    goto <bb 4>;
  else
    goto <bb 5>;

  <bb 4>:
  _5 = (void *) ivtmp.17_9;
  pretmp_20 = MEM[base: _5, offset: 4294967288B];
  pretmp_22 = X[1];
  goto <bb 3>;

That patch skips computing cost for sub iv_uses in a group:
Group 0:
  Type: ADDRESS
  Use 0.0:
    At stmt:    X[i_18] = _6;
    At pos:     X[i_18]
    IV struct:
      Type:     double *
      Base:     (double *) (&X + 16)
      Step:     8
      Object:   (void *) &X
      Biv:      N
      Overflowness wrto loop niter:     Overflow
  Use 0.1:
    At stmt:    pretmp_20 = X[_15];
    At pos:     X[_15]
    IV struct:
      Type:     double *
      Base:     (double *) (&X + 16)
      Step:     8
      Object:   (void *) &X
      Biv:      N
      Overflowness wrto loop niter:     Overflow

Though use 0.0/0.1 have same {base, step}, but there are in different program
point, so if iv_cand is increased before use 0.0, the first use is transformed
into: MEM[var_before, 0], and use 0.1 will be transformed into: MEM[var_after,
-8].  Now the two memory reference have different expressions, though access to
the same object.  Afterwards, DOM failed to CSE them.

It is hard to decide which sub iv_use we should compute the cost, and which one
we can skip computing the cost.  Maybe I should revert this part of code which
was introduced to save a small amount of compilation time.