diff --git a/src/record_ts.cpp b/src/record_ts.cpp
index a536cb1b..ef7533bf 100644
--- a/src/record_ts.cpp
+++ b/src/record_ts.cpp
@@ -769,11 +769,8 @@ int Recording::replay_launch(Operation &op) {
     }
 
     // Change kernel size in `kernel_params`
-    if (backend == JitBackend::CUDA) {
-        uintptr_t size = 0;
-        std::memcpy(&size, &launch_size, sizeof(uint32_t));
-        kernel_params[0] = (void *) size;
-    }
+    if (backend == JitBackend::CUDA)
+        kernel_params[0] = (void *) (uintptr_t) launch_size;
 
     if (!dry_run) {
 #ifndef NDEBUG
@@ -1530,7 +1527,7 @@ void RecordThreadState::record_aggregate(void *dst, AggregationEntry *agg,
              size);
 
     uint32_t start = m_recording.dependencies.size();
-    
+
     add_out_param(dst, VarType::UInt8);
 
     for (uint32_t i = 0; i < size; ++i) {
diff --git a/src/record_ts.h b/src/record_ts.h
index c91e0af6..fe3b13b1 100644
--- a/src/record_ts.h
+++ b/src/record_ts.h
@@ -195,7 +195,7 @@ struct Operation {
         struct {
             /// The reduce type of a prefix reduction operation
             ReduceOp rtype;
-            
+
             /// Whether a prefix sum operation is exclusive
             bool exclusive;
             bool reverse;
@@ -450,7 +450,7 @@ struct RecordThreadState : ThreadState {
 protected:
     /// Indicates that recording has been paused.
     bool m_paused = false;
-    
+
     /// Mapping from data pointer of a variable to a index into the slot of the
     /// recording.
     PtrToSlot ptr_to_slot;
@@ -556,7 +556,7 @@ struct RecordThreadState : ThreadState {
      * function will create a new variable and mapping from the pointer to it.
      */
     void notify_free(const void *ptr) override;
-    
+
     // =============================================================
 
     ~RecordThreadState() {}
diff --git a/tests/test.h b/tests/test.h
index 4e282874..22f1fd7a 100644
--- a/tests/test.h
+++ b/tests/test.h
@@ -337,7 +337,7 @@ template <typename Func> class FrozenFunction {
             return record(input_vector, args...);
         } else {
             std::vector<uint32_t> output_vector(m_outputs, 0);
-            
+
             jit_log(LogLevel::Debug, "replay:");
             // replay adds borrowing references to the \c output_vector
             jit_freeze_replay(m_recording, input_vector.data(),