[Bug tree-optimization/34244] [4.3 Regression] VRP/SCEV miscompiles Firefox

[rakdver at gcc dot gnu dot org]

------- Comment #5 from rakdver at gcc dot gnu dot org  2007-11-27 17:00 -------
> # of iteration analysis records an assumption that offset_46 >= 0.  However,
> this is simplified to true, as the value range of offset_46 is set to [0,0] by
> vrp (which seems to be wrong); so the problem is probably somewhere else in
> vrp.

So the problem is the following:  we have a code like

if (something)
  offset = 100;
else
  offset = -100;

while (offset > 0)
  offset--;

if (offset == 0)
  launch_nuclear_rockets ();

VRP starts simulating the code, first executing the true branch of the if
(something) condition, getting offset = 100.  It then proceeds with the loop,
determining that number of iterations is offset (since we just now believe that
offset==100, this is correct, without any assumptions), thus the final value of
offset is 0 and the nuclear war always starts.

Later, VRP evaluates the false branch of the if (something) condition, setting
the value range of offset to [-100,100], and proceeds to re-evaluate the
effects of the loop.  However, scev caches the number of iterations of the
loop, so it is not re-evaluated, and we keep believing that the number of
iterations is always offset.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34244