[Bug tree-optimization/112612] Holding on the loop variable rather than a derived value which can replace it

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112612

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2023-11-20

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.

Candidate 1:
  Var befor: ivtmp.5
  Var after: ivtmp.5
  Incr POS: before exit test
  IV struct:
    Type:       unsigned long
    Base:       0
    Step:       1
    Biv:        N
    Overflowness wrto loop niter:       No-overflow
...
Candidate 8:
  Var befor: ivtmp.10
  Var after: ivtmp.10
  Incr POS: before exit test
  IV struct:
    Type:       unsigned int
    Base:       0
    Step:       2
    Biv:        N
    Overflowness wrto loop niter:       No-overflow

Candidate 9:
  Var befor: ivtmp.11
  Var after: ivtmp.11
  Incr POS: before exit test
  IV struct:
    Type:       unsigned long
    Base:       (unsigned long) a_8(D)
    Step:       4
    Object:     (void *) a_8(D)
    Biv:        N
    Overflowness wrto loop niter:       Overflow

so we do have this candidate.

Improved to:
  cost: 16 (complexity 0)
  reg_cost: 4
  cand_cost: 10
  cand_group_cost: 2 (complexity 0)
  candidates: 8, 9
   group:0 --> iv_cand:8, cost=(0,0)
   group:1 --> iv_cand:9, cost=(2,0)
   group:2 --> iv_cand:8, cost=(0,0)
  invariant variables:
  invariant expressions:

Initial set of candidates:
  cost: 15 (complexity 2)
  reg_cost: 3
  cand_cost: 5
  cand_group_cost: 7 (complexity 2)
  candidates: 1
   group:0 --> iv_cand:1, cost=(4,0)
   group:1 --> iv_cand:1, cost=(3,2)
   group:2 --> iv_cand:1, cost=(0,0)
  invariant variables: 1
  invariant expressions:

but somehow we fail to express(?) some of the uses with just candidate 8?

It "works" with -m32 added:

  <bb 3> [local count: 1063004408]:
  # ivtmp.10_12 = PHI <ivtmp.10_11(5), 0(2)>
  val_7 = (int) ivtmp.10_12;
  MEM[(int *)a_8(D) + ivtmp.10_12 * 2] = val_7;
  ivtmp.10_11 = ivtmp.10_12 + 2;
  if (ivtmp.10_11 != 200)
    goto <bb 5>; [98.99%]

so it might be the 32->64bit promotion is what gets us off.  We might
possibly want to consider a 'unsigned long' candidate with step 2.

.L2:
        movl    %eax, (%edx,%eax,2)
        addl    $2, %eax
        cmpl    $200, %eax
        jne     .L2