diff --git a/thrust/testing/cuda/mismatch.cu b/thrust/testing/cuda/mismatch.cu
index 540dccdbb96..1fcf097953e 100644
--- a/thrust/testing/cuda/mismatch.cu
+++ b/thrust/testing/cuda/mismatch.cu
@@ -93,3 +93,36 @@ void TestMismatchCudaStreams()
   cudaStreamDestroy(s);
 }
 DECLARE_UNITTEST(TestMismatchCudaStreams);
+
+// see https://github.com/NVIDIA/cccl/issues/3591
+template <typename T>
+class Wrapper
+{
+public:
+  Wrapper()
+  {
+    ++my_count;
+  }
+
+  _CCCL_HOST_DEVICE bool operator==(const Wrapper& input) const
+  {
+    return true;
+  }
+
+  ~Wrapper()
+  {
+    --my_count;
+  }
+
+private:
+  static std::atomic<size_t> my_count;
+  T dummy;
+};
+
+void TestMismatchBug3591()
+{
+  using T = Wrapper<int32_t>;
+  T* p    = nullptr;
+  thrust::mismatch(thrust::device, p, p, p, cuda::std::equal_to<T>());
+}
+DECLARE_UNITTEST(TestMismatchBug3591);
diff --git a/thrust/thrust/system/cuda/detail/mismatch.h b/thrust/thrust/system/cuda/detail/mismatch.h
index 9987799faca..63e64b722ec 100644
--- a/thrust/thrust/system/cuda/detail/mismatch.h
+++ b/thrust/thrust/system/cuda/detail/mismatch.h
@@ -63,17 +63,145 @@ THRUST_NAMESPACE_END
 THRUST_NAMESPACE_BEGIN
 namespace cuda_cub
 {
+namespace detail
+{
+template <class ValueType, class InputIt1, class InputIt2, class BinaryOp>
+struct transform_pair_of_input_iterators_t
+{
+  using self_t            = transform_pair_of_input_iterators_t;
+  using difference_type   = typename iterator_traits<InputIt1>::difference_type;
+  using value_type        = ValueType;
+  using pointer           = void;
+  using reference         = value_type;
+  using iterator_category = std::random_access_iterator_tag;
+
+  InputIt1 input1;
+  InputIt2 input2;
+  mutable BinaryOp op;
+
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE
+  transform_pair_of_input_iterators_t(InputIt1 input1_, InputIt2 input2_, BinaryOp op_)
+      : input1(input1_)
+      , input2(input2_)
+      , op(op_)
+  {}
+
+  transform_pair_of_input_iterators_t(const self_t&) = default;
+
+  // BinaryOp might not be copy assignable, such as when it is a lambda.
+  // Define an explicit copy assignment operator that doesn't try to assign it.
+  _CCCL_HOST_DEVICE self_t& operator=(const self_t& o)
+  {
+    input1 = o.input1;
+    input2 = o.input2;
+    return *this;
+  }
+
+  /// Postfix increment
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE self_t operator++(int)
+  {
+    self_t retval = *this;
+    ++input1;
+    ++input2;
+    return retval;
+  }
+
+  /// Prefix increment
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE self_t operator++()
+  {
+    ++input1;
+    ++input2;
+    return *this;
+  }
+
+  /// Indirection
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE reference operator*() const
+  {
+    return op(*input1, *input2);
+  }
+  /// Indirection
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE reference operator*()
+  {
+    return op(*input1, *input2);
+  }
+
+  /// Addition
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE self_t operator+(difference_type n) const
+  {
+    return self_t(input1 + n, input2 + n, op);
+  }
+
+  /// Addition assignment
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE self_t& operator+=(difference_type n)
+  {
+    input1 += n;
+    input2 += n;
+    return *this;
+  }
+
+  /// Subtraction
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE self_t operator-(difference_type n) const
+  {
+    return self_t(input1 - n, input2 - n, op);
+  }
+
+  /// Subtraction assignment
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE self_t& operator-=(difference_type n)
+  {
+    input1 -= n;
+    input2 -= n;
+    return *this;
+  }
+
+  /// Distance
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE difference_type operator-(self_t other) const
+  {
+    return input1 - other.input1;
+  }
+
+  /// Array subscript
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE reference operator[](difference_type n) const
+  {
+    return op(input1[n], input2[n]);
+  }
+
+  /// Equal to
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE bool operator==(const self_t& rhs) const
+  {
+    return (input1 == rhs.input1) && (input2 == rhs.input2);
+  }
+
+  /// Not equal to
+  _CCCL_HOST_DEVICE _CCCL_FORCEINLINE bool operator!=(const self_t& rhs) const
+  {
+    return (input1 != rhs.input1) || (input2 != rhs.input2);
+  }
+
+}; // struct transform_pair_of_input_iterators_t
+} // namespace detail
 
 template <class Derived, class InputIt1, class InputIt2, class BinaryPred>
 pair<InputIt1, InputIt2> _CCCL_HOST_DEVICE
 mismatch(execution_policy<Derived>& policy, InputIt1 first1, InputIt1 last1, InputIt2 first2, BinaryPred binary_pred)
 {
+  const auto transform_first =
+    detail::transform_pair_of_input_iterators_t<bool, InputIt1, InputIt2, BinaryPred>(first1, first2, binary_pred);
+  const auto result = cuda_cub::find_if_not(
+    policy, transform_first, transform_first + thrust::distance(first1, last1), ::cuda::std::__identity{});
+  return thrust::make_pair(first1 + thrust::distance(transform_first, result),
+                           first2 + thrust::distance(transform_first, result));
+
+  // FIXME(bgruber): the following code should be equivalent and not require a dedicated iterator. However, it
+  // additionally requires the value_type to constructible/destructible on the device, which should be fixed at some
+  // point. See also: https://github.com/NVIDIA/cccl/issues/3591
+#  if 0
   const auto n            = thrust::distance(first1, last1);
   const auto first        = make_zip_iterator(first1, first2);
   const auto last         = make_zip_iterator(last1, first2 + n);
   const auto mismatch_pos = cuda_cub::find_if_not(policy, first, last, make_zip_function(binary_pred));
   const auto dist         = thrust::distance(first, mismatch_pos);
   return thrust::make_pair(first1 + dist, first2 + dist);
+#  endif
 }
 
 template <class Derived, class InputIt1, class InputIt2>