[Mlir-commits] [mlir] [mlir][affine] remove divide zero check when simplifer affineMap (#64622) (PR #68519)

[Jakub Kuderski]

================
@@ -700,6 +708,94 @@ static std::optional<int64_t> getUpperBound(Value iv) {
   return forOp.getConstantUpperBound() - 1;
 }
 
+/// Get a lower or upper (depending on `isUpper`) bound for `expr` while using
+/// the constant lower and upper bounds for its inputs provided in
+/// `constLowerBounds` and `constUpperBounds`. Return std::nullopt if such a
+/// bound can't be computed. This method only handles simple sum of product
+/// expressions (w.r.t constant coefficients) so as to not depend on anything
+/// heavyweight in `Analysis`. Expressions of the form: c0*d0 + c1*d1 + c2*s0 +
+/// ... + c_n are handled. Expressions involving floordiv, ceildiv, mod or
+/// semi-affine ones will lead std::nullopt being returned.
+static std::optional<int64_t>
+getBoundForExpr(AffineExpr expr, unsigned numDims, unsigned numSymbols,
+                ArrayRef<std::optional<int64_t>> constLowerBounds,
+                ArrayRef<std::optional<int64_t>> constUpperBounds,
+                bool isUpper) {
+  // Handle divs and mods.
+  if (auto binOpExpr = expr.dyn_cast<AffineBinaryOpExpr>()) {
+    // If the LHS of a floor or ceil is bounded and the RHS is a constant, we
+    // can compute an upper bound.
+    if (binOpExpr.getKind() == AffineExprKind::FloorDiv) {
+      auto rhsConst = binOpExpr.getRHS().dyn_cast<AffineConstantExpr>();
+      if (!rhsConst || rhsConst.getValue() < 1)
+        return std::nullopt;
+      auto bound = getBoundForExpr(binOpExpr.getLHS(), numDims, numSymbols,
----------------
kuhar wrote:

Could you write the full type here? It's not obvious just based on the RHS.

https://github.com/llvm/llvm-project/pull/68519