The induction variable canonicalization pass inserts new canonical induction
variables if a loop is lacking one, but it always chooses to insert a 32-bit
induction variable.  This can turn legal loops into infinite loops, such as this
one:

---
long long V;
int main() {
  long long X;
  for (X = 1; X < (1ULL << 33); ++X)
    V = X;
  printf("%lld\n", V);    // erroneously not reached
}
---

Though always introducing 64-bit induction variables would be acceptable, the
right approach is to insert a variable which is the largest of the existing
auxillary induction variables.

-Chris

Ugh, this gets worse.  The indvars pass is attempting to take an existing
canonical induction variable if it exists, instead of introducing a new one (a
nobel goal).  However, since it's ignoring the size of the existing variable, it
can cause neat miscompilations is cases like this:

void test() {
  int I; unsigned char C;

  for (C = 0, I = 1; I != 12300; ++C, ++I)
    foo(I);
}

Luckily, it appears extremely uncommon for programs to use loops with mixed
sized induction variables.

-Chris

Fixed.  Testcase here:
http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20031215/010403.html

Patch here:
http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20031215/010409.html

-Chris