[ScalarEvolution] code on nsw inconsistent with comment

+llvmdev

The code looks incorrect, AFAICT. Can you write a test case where
this leads to a miscompile, just to confirm? I suspect this will show
up in something like

for (...) {
  if (cond)
    t = inbounds_gep(A, idx)
  m = gep(A, idx)
  v = m s< A
}

SCEV should incorrectly fold v to false.