[Bug ipa/84658] [7/8 Regression] -O3 -fmerge-all-constants causes incorrect for-each loop generation.

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84658

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |alias
             Status|NEW                         |ASSIGNED

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
This is a points-to issue.  After inlining we have

  static const intD.9 aD.2281[16] = {0, 1, 2, 3, 4, 5, 8, 15, 16, 17, 512,
1020, 1021, 1022, 1023, 1024};
  intD.9 bD.2313;
  const intD.9 * __for_beginD.2314;
  static const intD.9 aD.2291ptD.2281[16];
...
  <bb 3> [100.00%]:
  # PT = { D.2281 } (nonlocal)
  # ALIGN = 4, MISALIGN = 0
  # __for_begin_8 = PHI <&aD.2281(2), __for_begin_10(4)>
  if (__for_begin_8 == &MEM[(voidD.43 *)&aD.2281 + 64B])
    goto <bb 5>; [5.88%]

good copy!

  <bb 6> [100.00%]:
  # PT = { D.2291 } (nonlocal)
  # ALIGN = 4, MISALIGN = 0
  # __for_begin_5 = PHI <&aD.2291ptD.2281(5), __for_begin_7(7)>
  if (__for_begin_5 == &MEM[(voidD.43 *)&aD.2291ptD.2281 + 64B])
    goto <bb 8>; [5.88%]
  else
    goto <bb 7>; [94.12%]

bad copy!  See how __for_begin_5 only points to D.2291 -- remember the
points-to sets are just bits.  But the DECL we check against has
been adjusted to the DECL_PT_UID of 2281...

I guess this is finally a case where we wondered if we get things correct... :/
I remember saying you need to adjust all existing points-to sets....

Note for the function bodies after inlining I see foo () has points-to
retained but bar () has it seemingly cleared?  But maybe this just dump
before applying the IPA ICF transform.

The ICF dump says:

Unified; Variable alias has been created.
  Setting points-to UID of [a.2291] as 2281

but then existing points-to sets containing 2291 are not adjusted.