Fix bug in visitDivExpr, visitMulExpr and visitModExpr

Whenever the result of a div or mod affine expression is a
constant expression, place the value in the constant index of the
flattened expression instead of adding it as a local expression.

Author: arnab-polymage

mlir/include/mlir/IR/AffineExprVisitor.h

@@ -418,6 +418,14 @@ class SimpleAffineExprFlattener
                                              AffineExpr localExpr);
 
 private:
+  /// Flatten binary expression `expr` and add it to `result`. If `expr` is a
+  /// dimension, symbol or constant, we add it to appropriate index in `result`.
+  /// Otherwise we add it in the local variable section. `lhs` and `rhs` are the
+  /// operands of `expr`.
+  LogicalResult addExprToFlattenedList(AffineExpr expr, ArrayRef<int64_t> lhs,
+                                       ArrayRef<int64_t> rhs,
+                                       SmallVectorImpl<int64_t> &result);
+
   /// Adds `localExpr`, which may be mod, ceildiv, floordiv or mod expression
   /// representing the affine expression corresponding to the quantifier
   /// introduced as the local variable corresponding to `localExpr`. If the

mlir/lib/IR/AffineExpr.cpp

@@ -1177,10 +1177,9 @@ static AffineExpr getSemiAffineExprFromFlatForm(ArrayRef<int64_t> flatExprs,
     if (flatExprs[numDims + numSymbols + it.index()] == 0)
       continue;
     AffineExpr expr = it.value();
-    auto binaryExpr = dyn_cast<AffineBinaryOpExpr>(expr);
-    if (!binaryExpr)
-      continue;
-
+    // A local expression cannot be a dimension, symbol or a constant -- it
+    // should be a binary op expression.
+    auto binaryExpr = cast<AffineBinaryOpExpr>(expr);
     AffineExpr lhs = binaryExpr.getLHS();
     AffineExpr rhs = binaryExpr.getRHS();
     if (!((isa<AffineDimExpr>(lhs) || isa<AffineSymbolExpr>(lhs)) &&
@@ -1274,6 +1273,27 @@ SimpleAffineExprFlattener::SimpleAffineExprFlattener(unsigned numDims,
   operandExprStack.reserve(8);
 }
 
+LogicalResult SimpleAffineExprFlattener::addExprToFlattenedList(
+    AffineExpr expr, ArrayRef<int64_t> lhs, ArrayRef<int64_t> rhs,
+    SmallVectorImpl<int64_t> &result) {
+  if (auto constExpr = dyn_cast<AffineConstantExpr>(expr)) {
+    std::fill(result.begin(), result.end(), 0);
+    result[getConstantIndex()] = constExpr.getValue();
+    return success();
+  }
+  if (auto dimExpr = dyn_cast<AffineDimExpr>(expr)) {
+    std::fill(result.begin(), result.end(), 0);
+    result[getDimStartIndex() + dimExpr.getPosition()] = 1;
+    return success();
+  }
+  if (auto symExpr = dyn_cast<AffineSymbolExpr>(expr)) {
+    std::fill(result.begin(), result.end(), 0);
+    result[getSymbolStartIndex() + symExpr.getPosition()] = 1;
+    return success();
+  }
+  return addLocalVariableSemiAffine(lhs, rhs, expr, result, result.size());
+}
+
 // In pure affine t = expr * c, we multiply each coefficient of lhs with c.
 //
 // In case of semi affine multiplication expressions, t = expr * symbolic_expr,
@@ -1295,7 +1315,7 @@ LogicalResult SimpleAffineExprFlattener::visitMulExpr(AffineBinaryOpExpr expr) {
                                              localExprs, context);
     AffineExpr b = getAffineExprFromFlatForm(rhs, numDims, numSymbols,
                                              localExprs, context);
-    return addLocalVariableSemiAffine(mulLhs, rhs, a * b, lhs, lhs.size());
+    return addExprToFlattenedList(a * b, mulLhs, rhs, lhs);
   }
 
   // Get the RHS constant.
@@ -1347,8 +1367,7 @@ LogicalResult SimpleAffineExprFlattener::visitModExpr(AffineBinaryOpExpr expr) {
         lhs, numDims, numSymbols, localExprs, context);
     AffineExpr divisorExpr = getAffineExprFromFlatForm(rhs, numDims, numSymbols,
                                                        localExprs, context);
-    AffineExpr modExpr = dividendExpr % divisorExpr;
-    return addLocalVariableSemiAffine(modLhs, rhs, modExpr, lhs, lhs.size());
+    return addExprToFlattenedList(dividendExpr % divisorExpr, modLhs, rhs, lhs);
   }
 
   int64_t rhsConst = rhs[getConstantIndex()];
@@ -1482,7 +1501,7 @@ LogicalResult SimpleAffineExprFlattener::visitDivExpr(AffineBinaryOpExpr expr,
     AffineExpr b = getAffineExprFromFlatForm(rhs, numDims, numSymbols,
                                              localExprs, context);
     AffineExpr divExpr = isCeil ? a.ceilDiv(b) : a.floorDiv(b);
-    return addLocalVariableSemiAffine(divLhs, rhs, divExpr, lhs, lhs.size());
+    return addExprToFlattenedList(divExpr, divLhs, rhs, lhs);
   }
 
   // This is a pure affine expr; the RHS is a positive constant.

mlir/test/Dialect/Affine/simplify-structures.mlir

@@ -608,3 +608,32 @@ func.func @semiaffine_simplification_floordiv_and_ceildiv_const(%arg0: tensor<?x
   // CHECK-NEXT: return %[[C6]], %[[C7]]
   return %a, %b : index, index
 }
+
+// -----
+
+// CHECK-DAG: #[[$MAP:.*]] = affine_map<()[s0] -> (13 mod s0)>
+// CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0) -> (d0 * 2)>
+// CHECK-LABEL: semiaffine_simplification_local_expr_folded_into_non_binary_expr
+func.func @semiaffine_simplification_local_expr_folded_into_non_binary_expr(%arg0: memref<?x?xf32>) -> (index, index) {
+  %c0 = arith.constant 0 : index
+  %c1 = arith.constant 1 : index
+  %c4 = arith.constant 4 : index
+  %c13 = arith.constant 13 : index
+  // CHECK: %[[DIM:.*]] = memref.dim
+  %dim = memref.dim %arg0, %c0 : memref<?x?xf32>
+  // CHECK: %[[VAL0:.*]] = affine.apply #[[$MAP]]()[%[[DIM]]]
+  %c = affine.apply affine_map<()[s0, s1, s2, s3] -> (s0 mod (s1 + (-s1 + s3) * (-s1 + s1 * s2 + 1)))>()[%c13, %dim, %c1, %dim]
+  %alloc = memref.alloc() : memref<1xindex>
+  affine.for %iv = 0 to 1 {
+    %d = affine.apply affine_map<(d0)[s1, s2] -> ((d0 - s1 + s1 * s2) * (s1 + (-s1 + 2) * (-s1 + s1 * s2 + 1)))>(%iv)[%dim, %c1]
+    affine.store %d, %alloc[0] : memref<1xindex>
+  }
+  // CHECK:      affine.for %[[IV:.*]] = 0 to 1 {
+  // CHECK-NEXT:   %[[VAL:.*]] = affine.apply #[[$MAP1]](%[[IV]])
+  // CHECK-NEXT:   affine.store %[[VAL]], %{{.*}}[0] : memref<1xindex>
+  // CHECK-NEXT: }
+  // CHECK: %[[VAL1:.*]] = affine.load %{{.*}}[0]
+  %d = affine.load %alloc[0] : memref<1xindex>
+  // CHECK: return %[[VAL0]], %[[VAL1]]
+  return %c, %d : index, index
+}