diff --git a/clients/drcachesim/tracer/output.cpp b/clients/drcachesim/tracer/output.cpp
index 99472c41f02..6f28e14e0f8 100644
--- a/clients/drcachesim/tracer/output.cpp
+++ b/clients/drcachesim/tracer/output.cpp
@@ -728,9 +728,9 @@ output_buffer(void *drcontext, per_thread_t *data, byte *buf_base, byte *buf_ptr
               size_t header_size)
 {
     byte *pipe_start = buf_base;
-    byte *pipe_end = pipe_start;
     if (!op_offline.get_value()) {
         byte *post_header = buf_base + header_size;
+        byte *last_ok_to_split_ref = nullptr;
         // Pipe split headers are just the tid.
         header_size = instru->sizeof_entry();
         for (byte *mem_ref = post_header; mem_ref < buf_ptr;
@@ -743,7 +743,6 @@ output_buffer(void *drcontext, per_thread_t *data, byte *buf_base, byte *buf_ptr
             // it or one instr after.
             if (is_ok_to_split_before(instru->get_entry_type(mem_ref),
                                       instru->get_entry_size(mem_ref))) {
-                pipe_end = mem_ref;
                 // We check the end of this entry + the max # of delay entries to
                 // avoid splitting an instr from its subsequent bundle entry.
                 // An alternative is to have the reader use per-thread state.
@@ -752,8 +751,23 @@ output_buffer(void *drcontext, per_thread_t *data, byte *buf_base, byte *buf_ptr
                     DR_ASSERT(is_ok_to_split_before(
                         instru->get_entry_type(pipe_start + header_size),
                         instru->get_entry_size(pipe_start + header_size)));
-                    pipe_start = atomic_pipe_write(drcontext, pipe_start, pipe_end,
-                                                   get_local_window(data));
+                    // Check if we went over the edge waiting for enough entries to
+                    // write. If we did, we simply write till the last ok-to-split ref.
+                    if (mem_ref - pipe_start > ipc_pipe.get_atomic_write_size()) {
+                        DR_ASSERT_MSG(
+                            last_ok_to_split_ref != nullptr,
+                            "Found too many entries without an ok-to-split point");
+                        pipe_start =
+                            atomic_pipe_write(drcontext, pipe_start, last_ok_to_split_ref,
+                                              get_local_window(data));
+                        last_ok_to_split_ref = mem_ref;
+                    } else {
+                        pipe_start = atomic_pipe_write(drcontext, pipe_start, mem_ref,
+                                                       get_local_window(data));
+                        last_ok_to_split_ref = nullptr;
+                    }
+                } else {
+                    last_ok_to_split_ref = mem_ref;
                 }
             }
         }
@@ -767,7 +781,9 @@ output_buffer(void *drcontext, per_thread_t *data, byte *buf_base, byte *buf_ptr
             DR_ASSERT(
                 is_ok_to_split_before(instru->get_entry_type(pipe_start + header_size),
                                       instru->get_entry_size(pipe_start + header_size)));
-            pipe_start = atomic_pipe_write(drcontext, pipe_start, pipe_end,
+            DR_ASSERT_MSG(last_ok_to_split_ref != nullptr,
+                          "Found too many entries without an ok-to-split point");
+            pipe_start = atomic_pipe_write(drcontext, pipe_start, last_ok_to_split_ref,
                                            get_local_window(data));
         }
         if ((buf_ptr - pipe_start) > (ssize_t)buf_hdr_slots_size) {