[Bug middle-end/30364] [4.1 Regression] Wrong variable ranges due to constant folding

[jakub at gcc dot gnu dot org]

------- Comment #17 from jakub at gcc dot gnu dot org  2007-08-24 16:33 -------
Does this make sense to do this for POINTER_TYPE_P in 4.1/4.2, which
reassociate
in many other places?
I belive in 4.3 that changed with POINTER_PLUS_EXPR stuff.

struct A { int i; int v[10000]; };
int *foo (struct A *a, int x)
{
  return &a->v[x - 2147483646];
}

in 4.2/4.1 gives:

foo (a, x)
{
<bb 2>:
  return &a->v[0] - 4294967288B + (int *) ((unsigned int) x * 4);

}

so it will overflow (well, underflow).

E.g. on Linux kernel code this commit causes some stack size usage regressions.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30364