diff --git a/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp b/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp
--- a/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp
+++ b/mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp
@@ -67,6 +67,15 @@
     for (unsigned i = 0, e = t.getRank(); i < e; ++i)
       res.push_back(b.create<DimOp>(loc, v, i));
   }
+  if (getNumInitTensors() == 0 && getOperation()->getNumResults() != 0) {
+    OpBuilder::InsertionGuard guard(b);
+    b.setInsertionPointAfter(getOperation());
+    for (Value v : getOperation()->getResults()) {
+      ShapedType t = v.getType().template cast<ShapedType>();
+      for (unsigned i = 0, e = t.getRank(); i < e; ++i)
+        res.push_back(b.create<DimOp>(loc, v, i));
+    }
+  }
   return res;
 }
 
@@ -165,14 +174,6 @@
   HasAffineDimExprVisitor checkDimExpr(outputDims);
   if (checkDimExpr.visit(operandShapesToResultDimMap.getResult(0)))
     return llvm::None;
-
-  // Since map.compose(createFlatListOfOperands()) only works for the case where
-  // init tensors exist, drop the dims corresponding to output shapes in the
-  // map.
-  if (getNumInitTensors() == 0 && getOperation()->getNumResults() != 0) {
-    operandShapesToResultDimMap = getProjectedMap(
-        operandShapesToResultDimMap, llvm::to_vector<4>(outputDims));
-  }
   return applyMapToValues(b, loc, operandShapesToResultDimMap,
                           createFlatListOfOperandDims(b, loc))[0];
 }