[Bug middle-end/55481] [4.8 regression] -O2 generates a wrong-code infinite loop in C++Benchmark's simple_types_constant_folding int8 xor test

[rakdver at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55481

--- Comment #15 from Zdenek Dvorak <rakdver at gcc dot gnu.org> 2012-12-11 18:19:40 UTC ---
Created attachment 28928
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=28928
A proposed patch

This patch fixes the error (and also makes the code more clearly match what is
described in the comments).  Long story: we try to avoid rewriting computations
that define bivs, as that is usually unnecessary.  However, there is a special
case where rewriting is still needed; e.g.,

loop
{
  tmp = biv + 1;
  use (tmp);
  biv = tmp + 1;   (*)
}

Ivopts may decide to eliminate the temporary variable tmp by expressing its use
in some other way (based on a different biv, etc.).  So, if we did not rewrite
(*), it would refer to a variable that is no longer defined.

And in this case, we used a wrong way of expressing the biv computation that
may cause signed overflows.  The patch makes us instead to fall back to the
general rewriting case, which does not have this problem.  Untested (beyond the
single testcase + C-only bootstrap).