[Bug tree-optimization/48822] [4.5/4.6/4.7 Regression] G++ gets stucks and never finishes compilation when enabling -O2/3 optimization options.

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48822

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
          Component|c++                         |tree-optimization
         AssignedTo|unassigned at gcc dot       |rguenth at gcc dot gnu.org
                   |gnu.org                     |
            Summary|[4.5 Regression] G++ gets   |[4.5/4.6/4.7 Regression]
                   |stucks and never finishes   |G++ gets stucks and never
                   |compilation when enabling   |finishes compilation when
                   |-O2/3 optimization options. |enabling -O2/3 optimization
                   |                            |options.

--- Comment #4 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-04-29 
14:06:24 UTC ---
  D.23284.num = 0;
  D.23284.den = 0;
...
<bb 11>:
  D.23599_34 = D.23284.den;
  if (D.23599_34 == 0)
    goto <bb 21> (<L10>);
  else
    goto <bb 12>;

<bb 12>:
  g_35 = D.23284.den;
  g_36 = D.23284.num;
  goto <bb 14>;

<bb 13>:
  g_39 = g_37 % g_38;

<bb 14>:
  # g_37 = PHI <g_36(12), g_38(13)>
  # g_38 = PHI <g_35(12), g_39(13)>
  if (g_38 != 0)
    goto <bb 13>;

and D.23599_34, g_35 and g_36 are correctly value-numbered to zero.  We then
start

Value numbering g_37 stmt = g_37 = PHI <g_36(12), g_38(13)>
Setting value number of g_37 to 0 (changed)
Value numbering g_38 stmt = g_38 = PHI <g_35(12), g_39(13)>
Setting value number of g_38 to 0 (changed)
Value numbering g_39 stmt = g_39 = g_37 % g_38;
Setting value number of g_39 to g_39 (changed)

g_39 has undefined result, thus

Value numbering g_37 stmt = g_37 = PHI <g_36(12), g_38(13)>
Setting value number of g_37 to 0
Value numbering g_38 stmt = g_38 = PHI <g_35(12), g_39(13)>
Setting value number of g_38 to g_38 (changed)
Value numbering g_39 stmt = g_39 = g_37 % g_38;
RHS g_37 % g_38 simplified to 0 has constants 1

where here is the bug that we go from VARYING to CONSTANT for g_39.

ISTR changing VN to never do that but that had lots of weird fallout.

I'm sure the issue is latent.