diff --git a/mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td b/mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td
--- a/mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td
+++ b/mlir/include/mlir/Dialect/Linalg/IR/LinalgOps.td
@@ -291,6 +291,8 @@
       "ArrayRef<OpFoldResult>":$low, "ArrayRef<OpFoldResult>":$high,
       CArg<"ArrayRef<NamedAttribute>", "{}">:$attrs)>
   ];
+
+  let hasCanonicalizer = 1;
 }
 
 def Linalg_RangeOp :
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
@@ -1119,6 +1119,28 @@
   return success();
 }
 
+namespace {
+// Folds linalg.pad_tensor when padding is static zeros.
+struct FoldStaticZeroPadding : public OpRewritePattern<PadTensorOp> {
+  using OpRewritePattern<PadTensorOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(PadTensorOp padTensorOp,
+                                PatternRewriter &rewriter) const override {
+    if (!padTensorOp.hasZeroLowPad() || !padTensorOp.hasZeroHighPad())
+      return failure();
+    rewriter.replaceOpWithNewOp<tensor::CastOp>(
+        padTensorOp, padTensorOp.result().getType(), padTensorOp.source());
+    return success();
+  }
+};
+
+} // namespace
+
+void PadTensorOp::getCanonicalizationPatterns(RewritePatternSet &results,
+                                              MLIRContext *context) {
+  results.add<FoldStaticZeroPadding>(context);
+}
+
 //===----------------------------------------------------------------------===//
 // ReshapeOp
 //===----------------------------------------------------------------------===//
diff --git a/mlir/test/Dialect/Linalg/canonicalize.mlir b/mlir/test/Dialect/Linalg/canonicalize.mlir
--- a/mlir/test/Dialect/Linalg/canonicalize.mlir
+++ b/mlir/test/Dialect/Linalg/canonicalize.mlir
@@ -1132,3 +1132,19 @@
 // CHECK-NEXT:     %[[SUM1:.+]] = addi %[[SUM0]], %[[ARG2]] : index
 // CHECK-NEXT:     %[[SUM2:.+]] = addi %[[SUM1]], %[[ARG3]] : index
 // CHECK-NEXT:     linalg.yield %[[SUM2]] : index
+
+// -----
+
+func @tensor_pad_cast_fold(%arg0: tensor<4x4xf32>) -> tensor<4x4xf32> {
+  %c0 = constant 0 : index
+  %cst = constant 0.0 : f32
+  %0 = tensor.cast %arg0 : tensor<4x4xf32> to tensor<?x?xf32>
+  %1 = linalg.pad_tensor %0 low[%c0, %c0] high[%c0, %c0]  {
+    ^bb0(%arg1: index, %arg2: index):  // no predecessors
+      linalg.yield %cst : f32
+  } : tensor<?x?xf32> to tensor<4x4xf32>
+  return %1 : tensor<4x4xf32>
+}
+// CHECK-LABEL: @tensor_pad_cast
+// CHECK-SAME: %[[ARG0:.+]]: tensor<4x4xf32>
+// CHECK: return %[[ARG0]]