diff --git a/mlir/lib/Dialect/Shape/IR/Shape.cpp b/mlir/lib/Dialect/Shape/IR/Shape.cpp
--- a/mlir/lib/Dialect/Shape/IR/Shape.cpp
+++ b/mlir/lib/Dialect/Shape/IR/Shape.cpp
@@ -606,12 +606,45 @@
     return success();
   }
 };
+
+template <typename OpTy>
+struct CanonicalizeCastedExtentTensorOperandsPattern
+    : public OpRewritePattern<OpTy> {
+  using OpRewritePattern<OpTy>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(OpTy op,
+                                PatternRewriter &rewriter) const override {
+    // Canonicalize operands.
+    bool anyChange = false;
+    auto canonicalizeOperand = [&](Value operand) {
+      if (auto castOp = operand.getDefiningOp<tensor::CastOp>()) {
+        // Only eliminate the cast if it holds no shape information.
+        bool isInformationLoosingCast =
+            castOp.getType().cast<RankedTensorType>().isDynamicDim(0);
+        if (isInformationLoosingCast) {
+          anyChange = true;
+          return castOp.source();
+        }
+      }
+      return operand;
+    };
+    auto newOperands = llvm::to_vector<8>(
+        llvm::map_range(op.getOperands(), canonicalizeOperand));
+
+    // Rewrite op if any change required.
+    if (!anyChange)
+      return failure();
+    rewriter.replaceOpWithNewOp<OpTy>(op, op->getResultTypes(), newOperands);
+    return success();
+  }
+};
 } // namespace
 
 void BroadcastOp::getCanonicalizationPatterns(RewritePatternSet &patterns,
                                               MLIRContext *context) {
   patterns.add<BroadcastFoldConstantOperandsPattern,
                BroadcastForwardSingleOperandPattern,
+               CanonicalizeCastedExtentTensorOperandsPattern<BroadcastOp>,
                RemoveDuplicateOperandsPattern<BroadcastOp>,
                RemoveEmptyShapeOperandsPattern<BroadcastOp>>(context);
 }
@@ -694,9 +727,11 @@
   // Canonicalization patterns have overlap with the considerations during
   // folding in case additional shape information is inferred at some point that
   // does not result in folding.
-  patterns.add<CstrBroadcastableEqOps,
-               RemoveDuplicateOperandsPattern<CstrBroadcastableOp>,
-               RemoveEmptyShapeOperandsPattern<CstrBroadcastableOp>>(context);
+  patterns
+      .add<CanonicalizeCastedExtentTensorOperandsPattern<CstrBroadcastableOp>,
+           CstrBroadcastableEqOps,
+           RemoveDuplicateOperandsPattern<CstrBroadcastableOp>,
+           RemoveEmptyShapeOperandsPattern<CstrBroadcastableOp>>(context);
 }
 
 // Return true if there is exactly one attribute not representing a scalar
diff --git a/mlir/test/Dialect/Shape/canonicalize.mlir b/mlir/test/Dialect/Shape/canonicalize.mlir
--- a/mlir/test/Dialect/Shape/canonicalize.mlir
+++ b/mlir/test/Dialect/Shape/canonicalize.mlir
@@ -1325,3 +1325,21 @@
   %2 = shape.cstr_broadcastable %0, %1: tensor<2xindex>, tensor<1xindex>
   "use"(%2) : (!shape.witness) -> ()
 }
+
+// -----
+
+// CHECK-LABEL: @casted_extent_tensor_operands
+// CHECK-SAME: (%[[ARG0:.*]]: tensor<?xindex>, %[[ARG1:.*]]: tensor<3xindex>)
+func @casted_extent_tensor_operands(%arg0 : tensor<?xindex>,
+    %arg1 : tensor<3xindex>) -> (!shape.witness, tensor<?xindex>) {
+  // CHECK: %[[CASTED_ARG0:.*]] = tensor.cast %[[ARG0]] : tensor<?xindex> to tensor<3xindex>
+  // CHECK: %[[WIT:.*]] = shape.cstr_broadcastable %[[CASTED_ARG0]], %[[ARG1]] : tensor<3xindex>, tensor<3xindex>
+  // CHECK: %[[RES:.*]] = shape.broadcast %[[CASTED_ARG0]], %[[ARG1]] : tensor<3xindex>, tensor<3xindex>
+  // CHECK: return %[[WIT]], %[[RES]]
+  %0 = tensor.cast %arg0 : tensor<?xindex> to tensor<3xindex>
+  %1 = tensor.cast %arg1 : tensor<3xindex> to tensor<?xindex>
+  %2 = shape.cstr_broadcastable %0, %1 : tensor<3xindex>, tensor<?xindex>
+  %3 = shape.broadcast %0, %1 :tensor<3xindex>, tensor<?xindex>
+      -> tensor<?xindex>
+  return %2, %3 : !shape.witness, tensor<?xindex>
+}