diff --git a/klippy/chelper/__init__.py b/klippy/chelper/__init__.py
index f7ba372c..ed5e156c 100644
--- a/klippy/chelper/__init__.py
+++ b/klippy/chelper/__init__.py
@@ -11,7 +11,9 @@ import cffi
 # c_helper.so compiling
 ######################################################################
 
-COMPILE_CMD = "gcc -Wall -g -O2 -shared -fPIC -o %s %s"
+COMPILE_CMD = ("gcc -Wall -g -O2 -shared -fPIC"
+               " -flto -fwhole-program -fno-use-linker-plugin"
+               " -o %s %s")
 SOURCE_FILES = ['stepcompress.c', 'serialqueue.c', 'pyhelper.c']
 DEST_LIB = "c_helper.so"
 OTHER_FILES = ['list.h', 'serialqueue.h', 'pyhelper.h']
@@ -56,9 +58,6 @@ defs_serialqueue = """
     void serialqueue_free_commandqueue(struct command_queue *cq);
     void serialqueue_send(struct serialqueue *sq, struct command_queue *cq
         , uint8_t *msg, int len, uint64_t min_clock, uint64_t req_clock);
-    void serialqueue_encode_and_send(struct serialqueue *sq
-        , struct command_queue *cq, uint32_t *data, int len
-        , uint64_t min_clock, uint64_t req_clock);
     void serialqueue_pull(struct serialqueue *sq
         , struct pull_queue_message *pqm);
     void serialqueue_set_baud_adjust(struct serialqueue *sq, double baud_adjust);
diff --git a/klippy/chelper/pyhelper.c b/klippy/chelper/pyhelper.c
index 6fa4817f..947e5ee1 100644
--- a/klippy/chelper/pyhelper.c
+++ b/klippy/chelper/pyhelper.c
@@ -1,6 +1,6 @@
 // Helper functions for C / Python interface
 //
-// Copyright (C) 2016  Kevin O'Connor <kevin@koconnor.net>
+// Copyright (C) 2016-2018  Kevin O'Connor <kevin@koconnor.net>
 //
 // This file may be distributed under the terms of the GNU GPLv3 license.
 
@@ -10,10 +10,11 @@
 #include <stdio.h> // fprintf
 #include <string.h> // strerror
 #include <time.h> // struct timespec
+#include "compiler.h" // __visible
 #include "pyhelper.h" // get_monotonic
 
 // Return the monotonic system time as a double
-double
+double __visible
 get_monotonic(void)
 {
     struct timespec ts;
@@ -41,7 +42,7 @@ default_logger(const char *msg)
 
 static void (*python_logging_callback)(const char *msg) = default_logger;
 
-void
+void __visible
 set_python_logging_callback(void (*func)(const char *))
 {
     python_logging_callback = func;
diff --git a/klippy/chelper/serialqueue.c b/klippy/chelper/serialqueue.c
index 4a18e845..8da2a8d5 100644
--- a/klippy/chelper/serialqueue.c
+++ b/klippy/chelper/serialqueue.c
@@ -1,6 +1,6 @@
 // Serial port command queuing
 //
-// Copyright (C) 2016  Kevin O'Connor <kevin@koconnor.net>
+// Copyright (C) 2016-2018  Kevin O'Connor <kevin@koconnor.net>
 //
 // This file may be distributed under the terms of the GNU GPLv3 license.
 //
@@ -23,6 +23,7 @@
 #include <string.h> // memset
 #include <termios.h> // tcflush
 #include <unistd.h> // pipe
+#include "compiler.h" // __visible
 #include "list.h" // list_add_tail
 #include "pyhelper.h" // get_monotonic
 #include "serialqueue.h" // struct queue_message
@@ -797,7 +798,7 @@ background_thread(void *data)
 }
 
 // Create a new 'struct serialqueue' object
-struct serialqueue *
+struct serialqueue * __visible
 serialqueue_alloc(int serial_fd, int write_only)
 {
     struct serialqueue *sq = malloc(sizeof(*sq));
@@ -859,7 +860,7 @@ fail:
 }
 
 // Request that the background thread exit
-void
+void __visible
 serialqueue_exit(struct serialqueue *sq)
 {
     pollreactor_do_exit(&sq->pr);
@@ -870,7 +871,7 @@ serialqueue_exit(struct serialqueue *sq)
 }
 
 // Free all resources associated with a serialqueue
-void
+void __visible
 serialqueue_free(struct serialqueue *sq)
 {
     if (!sq)
@@ -895,7 +896,7 @@ serialqueue_free(struct serialqueue *sq)
 }
 
 // Allocate a 'struct command_queue'
-struct command_queue *
+struct command_queue * __visible
 serialqueue_alloc_commandqueue(void)
 {
     struct command_queue *cq = malloc(sizeof(*cq));
@@ -906,7 +907,7 @@ serialqueue_alloc_commandqueue(void)
 }
 
 // Free a 'struct command_queue'
-void
+void __visible
 serialqueue_free_commandqueue(struct command_queue *cq)
 {
     if (!cq)
@@ -956,7 +957,7 @@ serialqueue_send_batch(struct serialqueue *sq, struct command_queue *cq
 
 // Schedule the transmission of a message on the serial port at a
 // given time and priority.
-void
+void __visible
 serialqueue_send(struct serialqueue *sq, struct command_queue *cq
                  , uint8_t *msg, int len, uint64_t min_clock, uint64_t req_clock)
 {
@@ -988,7 +989,7 @@ serialqueue_encode_and_send(struct serialqueue *sq, struct command_queue *cq
 
 // Return a message read from the serial port (or wait for one if none
 // available)
-void
+void __visible
 serialqueue_pull(struct serialqueue *sq, struct pull_queue_message *pqm)
 {
     pthread_mutex_lock(&sq->lock);
@@ -1022,7 +1023,7 @@ exit:
     pthread_mutex_unlock(&sq->lock);
 }
 
-void
+void __visible
 serialqueue_set_baud_adjust(struct serialqueue *sq, double baud_adjust)
 {
     pthread_mutex_lock(&sq->lock);
@@ -1030,7 +1031,7 @@ serialqueue_set_baud_adjust(struct serialqueue *sq, double baud_adjust)
     pthread_mutex_unlock(&sq->lock);
 }
 
-void
+void __visible
 serialqueue_set_receive_window(struct serialqueue *sq, int receive_window)
 {
     pthread_mutex_lock(&sq->lock);
@@ -1040,7 +1041,7 @@ serialqueue_set_receive_window(struct serialqueue *sq, int receive_window)
 
 // Set the estimated clock rate of the mcu on the other end of the
 // serial port
-void
+void __visible
 serialqueue_set_clock_est(struct serialqueue *sq, double est_freq
                           , double last_clock_time, uint64_t last_clock)
 {
@@ -1052,7 +1053,7 @@ serialqueue_set_clock_est(struct serialqueue *sq, double est_freq
 }
 
 // Return a string buffer containing statistics for the serial port
-void
+void __visible
 serialqueue_get_stats(struct serialqueue *sq, char *buf, int len)
 {
     struct serialqueue stats;
@@ -1074,7 +1075,7 @@ serialqueue_get_stats(struct serialqueue *sq, char *buf, int len)
 }
 
 // Extract old messages stored in the debug queues
-int
+int __visible
 serialqueue_extract_old(struct serialqueue *sq, int sentq
                         , struct pull_queue_message *q, int max)
 {
diff --git a/klippy/chelper/stepcompress.c b/klippy/chelper/stepcompress.c
index c65bb431..6e7a80d4 100644
--- a/klippy/chelper/stepcompress.c
+++ b/klippy/chelper/stepcompress.c
@@ -1,6 +1,6 @@
 // Stepper pulse schedule compression
 //
-// Copyright (C) 2016,2017  Kevin O'Connor <kevin@koconnor.net>
+// Copyright (C) 2016-2018  Kevin O'Connor <kevin@koconnor.net>
 //
 // This file may be distributed under the terms of the GNU GPLv3 license.
 //
@@ -229,7 +229,7 @@ check_line(struct stepcompress *sc, struct step_move move)
  ****************************************************************/
 
 // Allocate a new 'stepcompress' object
-struct stepcompress *
+struct stepcompress * __visible
 stepcompress_alloc(uint32_t max_error, uint32_t queue_step_msgid
                    , uint32_t set_next_step_dir_msgid, uint32_t invert_sdir
                    , uint32_t oid)
@@ -247,7 +247,7 @@ stepcompress_alloc(uint32_t max_error, uint32_t queue_step_msgid
 }
 
 // Free memory associated with a 'stepcompress' object
-void
+void __visible
 stepcompress_free(struct stepcompress *sc)
 {
     if (!sc)
@@ -328,7 +328,7 @@ set_next_step_dir(struct stepcompress *sc, int sdir)
 }
 
 // Reset the internal state of the stepcompress object
-int
+int __visible
 stepcompress_reset(struct stepcompress *sc, uint64_t last_step_clock)
 {
     int ret = stepcompress_flush(sc, UINT64_MAX);
@@ -340,7 +340,7 @@ stepcompress_reset(struct stepcompress *sc, uint64_t last_step_clock)
 }
 
 // Indicate the stepper is in homing mode (or done homing if zero)
-int
+int __visible
 stepcompress_set_homing(struct stepcompress *sc, uint64_t homing_clock)
 {
     int ret = stepcompress_flush(sc, UINT64_MAX);
@@ -351,7 +351,7 @@ stepcompress_set_homing(struct stepcompress *sc, uint64_t homing_clock)
 }
 
 // Queue an mcu command to go out in order with stepper commands
-int
+int __visible
 stepcompress_queue_msg(struct stepcompress *sc, uint32_t *data, int len)
 {
     int ret = stepcompress_flush(sc, UINT64_MAX);
@@ -502,7 +502,7 @@ static inline double safe_sqrt(double v) {
 }
 
 // Schedule a step event at the specified step_clock time
-int32_t
+int32_t __visible
 stepcompress_push(struct stepcompress *sc, double print_time, int32_t sdir)
 {
     int ret = set_next_step_dir(sc, !!sdir);
@@ -522,7 +522,7 @@ stepcompress_push(struct stepcompress *sc, double print_time, int32_t sdir)
 // Otherwise it uses the formula:
 //  step_time = (print_time + sqrt(2*step_num/accel + (start_sv/accel)**2)
 //               - start_sv/accel)
-int32_t
+int32_t __visible
 stepcompress_push_const(
     struct stepcompress *sc, double print_time
     , double step_offset, double steps, double start_sv, double accel)
@@ -672,7 +672,7 @@ _stepcompress_push_delta(
     return res;
 }
 
-int32_t
+int32_t __visible
 stepcompress_push_delta(
     struct stepcompress *sc, double print_time, double move_sd
     , double start_sv, double accel
@@ -729,7 +729,7 @@ struct steppersync {
 };
 
 // Allocate a new 'steppersync' object
-struct steppersync *
+struct steppersync * __visible
 steppersync_alloc(struct serialqueue *sq, struct stepcompress **sc_list
                   , int sc_num, int move_num)
 {
@@ -750,7 +750,7 @@ steppersync_alloc(struct serialqueue *sq, struct stepcompress **sc_list
 }
 
 // Free memory associated with a 'steppersync' object
-void
+void __visible
 steppersync_free(struct steppersync *ss)
 {
     if (!ss)
@@ -762,7 +762,7 @@ steppersync_free(struct steppersync *ss)
 }
 
 // Set the conversion rate of 'print_time' to mcu clock
-void
+void __visible
 steppersync_set_time(struct steppersync *ss, double time_offset, double mcu_freq)
 {
     int i;
@@ -798,7 +798,7 @@ heap_replace(struct steppersync *ss, uint64_t req_clock)
 }
 
 // Find and transmit any scheduled steps prior to the given 'move_clock'
-int
+int __visible
 steppersync_flush(struct steppersync *ss, uint64_t move_clock)
 {
     // Flush each stepcompress to the specified move_clock