It looks like we are miscompiling indexes through C99 VLA's in some cases.  File
this into the "wow, I didn't know you could do that" catagory.  :)

Testcase:

---
int foo(int len, char arr[][len], int X) {
  return arr[X][0];
}
---

The correct addressing arithmetic to generate is "arr + X*len", which we don't
do in this case.

-Chris

Here's another horrible testcase with the same issue:


int sub1 (int i, int j) {
  typedef struct {
   int  c[i+2];
  }c;
  int x[10], y[10];

  if (j == 2) {
      memcpy (x, y, 10 * sizeof (int));
      return sizeof (c);
  } else
    return sizeof (c) * 3;
}

int main() {
  typedef struct {
   int  c[22];
  }c;
  if (sub1 (20, 3) != sizeof (c)*3)
    abort ();
  return 0;
}

This patch:
http://mail.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20040503/014167.html

Fixes this testcase (test/Regression/CFrontend/2004-05-07-VarArrays.c):

int foo(int len, char arr[][len], int X) {
  return arr[X][0];
}

The second testcase attached to this bug is actually a bug in the non-LLVM
specific portion of GCC, so it will magically be fixed when new bits are merged in.

-Chris