[llvm] [SCEV] Require that addrec operands dominate the loop (PR #67030)

[Philip Reames via llvm-commits]

================
@@ -3670,9 +3670,12 @@ ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands,
            "SCEVAddRecExpr operand types don't match!");
     assert(!Operands[i]->getType()->isPointerTy() && "Step must be integer");
   }
-  for (unsigned i = 0, e = Operands.size(); i != e; ++i)
+  for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
     assert(isLoopInvariant(Operands[i], L) &&
            "SCEVAddRecExpr operand is not loop-invariant!");
+    assert(properlyDominates(Operands[i], L->getHeader()) &&
----------------
preames wrote:

I find it quite surprising that a value defined after L (i.e. whose live range doesn't include L) is consider loop invariant in L.  I wonder if we don't have a whole class of bugs based on that confusion - i.e. beyond this single change.

Do you have any example of a case where we create a SCEV with operands defined after the defining scope of the SCEV?  

https://github.com/llvm/llvm-project/pull/67030