diff --git a/alt.lds b/alt.lds
index 571618e..ccd8937 100644
--- a/alt.lds
+++ b/alt.lds
@@ -1,7 +1,8 @@
 OUTPUT_FORMAT("elf32-shl", "elf32-shl", "elf32-shl")
 MEMORY
 {
-  p1ram : ORIGIN = 0x8c020000, LENGTH = 0xff0000
-  p2ram : ORIGIN = 0xac020000, LENGTH = 0xff0000
+  p1ram : ORIGIN = 0x8c020000, LENGTH = 0xfe0000
+  p2ram : ORIGIN = 0xac020000, LENGTH = 0xfe0000
 }
+__stack_reservation = 0x1000;
 INCLUDE "common.lds"
diff --git a/client.py b/client.py
index a26f832..07a4055 100644
--- a/client.py
+++ b/client.py
@@ -70,7 +70,11 @@ def do(ser, b):
         sync(ser, b'prime', wait=0)
         do(ser, b)
     print("\nDATA")
+    start = time.monotonic()
     ret = symmetric(ser, b)
+    end = time.monotonic()
+    duration = end - start
+    print("duration", duration)
     print(ret[-5:])
     if ret[:-5] != b:
         print("ret != b; dumped to asdf.bin")
@@ -96,8 +100,8 @@ def console(ser):
 with open(sys.argv[1], 'rb') as f:
     b = f.read()
 
-#with serial.Serial('/dev/ttyUSB0', 120192, timeout=1) as ser:
-with serial.Serial('/dev/ttyUSB0', 312500, timeout=1) as ser:
+with serial.Serial('/dev/ttyUSB0', 120192, timeout=1) as ser:
+#with serial.Serial('/dev/ttyUSB0', 312500, timeout=1) as ser:
     #console(ser)
     print("waiting: ", end=' ')
     sys.stdout.flush()
diff --git a/common.lds b/common.lds
index 0cdaac2..0f36d99 100644
--- a/common.lds
+++ b/common.lds
@@ -44,13 +44,5 @@ SECTIONS
   INCLUDE "debug.lds"
 }
 
-__p1ram_start = ORIGIN(p1ram);
-__p1ram_end = ORIGIN(p1ram) + LENGTH(p1ram);
-
-__bss_link_start = ADDR(.bss);
-__bss_link_end = ADDR(.bss) + SIZEOF(.bss);
-
-__ctors_link_start = ADDR(.ctors);
-__ctors_link_end = ADDR(.ctors) + SIZEOF(.ctors);
-
+INCLUDE "symbols.lds"
 INCLUDE "addresses.lds"
diff --git a/common.mk b/common.mk
index 6dba415..bcef840 100644
--- a/common.mk
+++ b/common.mk
@@ -2,7 +2,7 @@ MAKEFILE_PATH := $(abspath $(lastword $(MAKEFILE_LIST)))
 DIR := $(dir $(MAKEFILE_PATH))
 
 LIB ?= .
-OPT ?= -O3
+OPT ?= -Og
 GENERATED ?=
 
 AARCH = --isa=sh4 --little
@@ -32,6 +32,11 @@ IP_OBJ = \
 	sg/sg_ini.o \
 	sg/aip.o
 
+%.o: %.obj
+	$(OBJCOPY) -g \
+		--rename-section IP=.text.$* \
+		$< $@
+
 ip.elf: $(IP_OBJ)
 	$(LD) --orphan-handling=error --print-memory-usage -T $(LIB)/ip.lds $^ -o $@
 
@@ -42,11 +47,6 @@ START_OBJ = \
 
 include base.mk
 
-%.o: %.obj
-	$(OBJCOPY) -g \
-		--rename-section IP=.text.$* \
-		$< $@
-
 sine.pcm: common.mk
 	sox \
 		--rate 44100 \
diff --git a/example/example.mk b/example/example.mk
index e90640a..3dd3627 100644
--- a/example/example.mk
+++ b/example/example.mk
@@ -16,7 +16,9 @@ SPRITE_OBJ = \
 	holly/core.o \
 	holly/region_array.o \
 	holly/background.o \
-	holly/ta_fifo_polygon_converter.o
+	holly/ta_fifo_polygon_converter.o \
+	sh7091/serial.o \
+	$(LIBGCC)
 
 example/sprite.elf: LDSCRIPT = $(LIB)/alt.lds
 example/sprite.elf: $(START_OBJ) $(SPRITE_OBJ)
@@ -149,7 +151,7 @@ WIFFLE_ATTENUATION_OBJ = \
 	holly/background.o \
 	holly/ta_fifo_polygon_converter.o
 
-example/wiffle_attenuation.elf: LDSCRIPT = $(LIB)/alt.lds
+example/wiffle_attenuation.elf: LDSCRIPT = $(LIB)/main.lds
 example/wiffle_attenuation.elf: $(START_OBJ) $(WIFFLE_ATTENUATION_OBJ)
 
 MODIFIER_VOLUME_OBJ = \
@@ -322,7 +324,7 @@ SERIAL_TRANSFER_OBJ = \
 	sh7091/serial.o \
 	serial_load.o
 
-example/serial_transfer.elf: LDSCRIPT = $(LIB)/main.lds
+example/serial_transfer.elf: LDSCRIPT = $(LIB)/loader.lds
 example/serial_transfer.elf: $(START_OBJ) $(SERIAL_TRANSFER_OBJ)
 
 INTERRUPT_OBJ = \
diff --git a/example/macaw.cpp b/example/macaw.cpp
index e90eb53..42a4bed 100644
--- a/example/macaw.cpp
+++ b/example/macaw.cpp
@@ -185,5 +185,8 @@ void main()
 
     theta += half_degree;
     frame_ix += 1;
+
+    if (frame_ix > 10)
+      break;
   }
 }
diff --git a/example/serial_transfer.cpp b/example/serial_transfer.cpp
index c95b6c6..5c05424 100644
--- a/example/serial_transfer.cpp
+++ b/example/serial_transfer.cpp
@@ -6,24 +6,19 @@
 
 #include "serial_load.hpp"
 
-extern uint32_t __bss_link_start __asm("__bss_link_start");
-extern uint32_t __bss_link_end __asm("__bss_link_end");
+void main() __attribute__((section(".text.main")));
 
 void main()
 {
-  serial::init(4);
+  //serial::init(12);
   load_init();
 
   while (1) {
     using namespace scif;
 
-    while ((sh7091.SCIF.SCFSR2 & scfsr2::tdfe::bit_mask) == 0) {
-      // wait
-    }
     while ((scfdr2::receive_data_bytes(sh7091.SCIF.SCFDR2)) > 0) {
-      uint8_t c = sh7091.SCIF.SCFRDR2;
+      const uint8_t c = sh7091.SCIF.SCFRDR2;
       load_recv(c);
     }
-    sh7091.SCIF.SCFSR2 = sh7091.SCIF.SCFSR2 & (~scfsr2::rdf::bit_mask);
   }
 }
diff --git a/loader.lds b/loader.lds
new file mode 100644
index 0000000..601e03b
--- /dev/null
+++ b/loader.lds
@@ -0,0 +1,61 @@
+OUTPUT_FORMAT("elf32-shl", "elf32-shl", "elf32-shl")
+MEMORY
+{
+  p1ram : ORIGIN = 0x8c020000, LENGTH = 0xfe0000
+  p2ram : ORIGIN = 0xac020000, LENGTH = 0xfe0000
+  ldram : ORIGIN = 0x8cfff000, LENGTH = 0x1000
+}
+SECTIONS
+{
+  . = ORIGIN(p2ram);
+
+  .text.startup ALIGN(4) : SUBALIGN(4)
+  {
+    KEEP(*(.text.start))
+    *(.text.startup.*)
+    . = ALIGN(4);
+  } > p2ram AT> p2ram
+
+  .ctors ALIGN(4) : SUBALIGN(4)
+  {
+    KEEP(*(.ctors))
+    KEEP(*(.ctors.*))
+    . = ALIGN(4);
+  } > p2ram AT> p2ram
+
+  . = ORIGIN(ldram);
+
+  .text ALIGN(4) : SUBALIGN(4)
+  {
+    *(.text.*)
+    *(.text)
+    . = ALIGN(4);
+  } > ldram AT> p2ram
+
+  .data ALIGN(4) : SUBALIGN(4)
+  {
+    *(.data)
+    *(.data.*)
+    . = ALIGN(4);
+  } > ldram AT> p2ram
+
+  .rodata ALIGN(4) : SUBALIGN(4)
+  {
+    *(.rodata)
+    *(.rodata.*)
+    . = ALIGN(4);
+  } > ldram AT> p2ram
+
+  .bss ALIGN(4) (NOLOAD) : SUBALIGN(4)
+  {
+    *(.bss)
+    *(.bss.*)
+    *(COMMON)
+    . = ALIGN(4);
+  } > ldram
+
+  INCLUDE "debug.lds"
+}
+__stack_reservation = 0x0000;
+INCLUDE "symbols.lds"
+INCLUDE "addresses.lds"
diff --git a/main.lds b/main.lds
index 70cfe91..94bfbc3 100644
--- a/main.lds
+++ b/main.lds
@@ -4,4 +4,5 @@ MEMORY
   p1ram : ORIGIN = 0x8c010000, LENGTH = 0xff0000
   p2ram : ORIGIN = 0xac010000, LENGTH = 0xff0000
 }
+__stack_reservation = 0x1000;
 INCLUDE "common.lds"
diff --git a/runtime.cpp b/runtime.cpp
index 6460fe8..55a0fd7 100644
--- a/runtime.cpp
+++ b/runtime.cpp
@@ -1,27 +1,76 @@
 #include <cstdint>
+#include <type_traits>
 
 #include "sh7091/cache.hpp"
 
+#include "string.hpp"
+
+extern uint32_t __text_link_start __asm("__text_link_start");
+extern uint32_t __text_link_end __asm("__text_link_end");
+extern uint32_t __text_load_start __asm("__text_load_start");
+
+extern uint32_t __data_link_start __asm("__data_link_start");
+extern uint32_t __data_link_end __asm("__data_link_end");
+extern uint32_t __data_load_start __asm("__data_load_start");
+
+extern uint32_t __rodata_link_start __asm("__rodata_link_start");
+extern uint32_t __rodata_link_end __asm("__rodata_link_end");
+extern uint32_t __rodata_load_start __asm("__rodata_load_start");
+
+extern uint32_t __ctors_link_start __asm("__ctors_link_start");
+extern uint32_t __ctors_link_end __asm("__ctors_link_end");
+
 extern uint32_t __bss_link_start __asm("__bss_link_start");
 extern uint32_t __bss_link_end __asm("__bss_link_end");
 
 extern void main();
 
+typedef void(init_t)(void);
+
 extern "C"
 void runtime_init()
 __attribute__((section(".text.startup.runtime_init")));
 
+void copy(uint32_t * start, const uint32_t * end, uint32_t * load)
+__attribute__((section(".text.startup.copy")));
+
+void copy(uint32_t * start, const uint32_t * end, uint32_t * load)
+{
+  if (start != load) {
+    while (start < end) {
+      *start++ = *load++;
+    }
+  }
+}
+
 extern "C"
 void runtime_init()
 {
+  // relocate text (if necessary)
+  copy(&__text_link_start, &__text_link_end, &__text_load_start);
+
+  // relocate data (if necessary)
+  copy(&__data_link_start, &__data_link_end, &__data_load_start);
+
+  // relocate rodata (if necessary)
+  copy(&__rodata_link_start, &__rodata_link_end, &__rodata_load_start);
+
+  uint32_t * start;
+  uint32_t * end;
+
   // clear BSS
-  uint32_t * start = &__bss_link_start;
-  uint32_t * end = &__bss_link_end;
+  start = &__bss_link_start;
+  end = &__bss_link_end;
   while (start < end) {
     *start++ = 0;
   }
 
-  cache::init();
+  // call ctors
+  start = &__ctors_link_start;
+  end = &__ctors_link_end;
+  while (start < end) {
+    ((init_t*)(*start++))();
+  }
 
-  main();
+  cache::init();
 }
diff --git a/serial_load.cpp b/serial_load.cpp
index d76564c..f4be31b 100644
--- a/serial_load.cpp
+++ b/serial_load.cpp
@@ -2,6 +2,7 @@
 
 #include "sh7091/sh7091.hpp"
 #include "sh7091/sh7091_bits.hpp"
+#include "sh7091/serial.hpp"
 #include "holly/holly.hpp"
 
 enum load_command {
@@ -54,6 +55,24 @@ void debug(const char * s)
   }
 }
 
+void jump_to_func(const uint32_t addr)
+{
+  serial::string("jump to: ");
+  serial::integer<uint32_t>(addr);
+  // save our stack
+  asm volatile ("ldc r15, gbr; "
+		"mov #0, r15; "
+		"jsr @%0; "
+		"nop; "
+		"stc gbr, r15; "
+                :
+                : "r"(addr) /* input */
+	        /* clobbered register */
+                : "r0","r1","r2","r3","r4","r5","r6","r7","r8","r9","r10","r11","r12","macl","mach","gbr","pr"
+                );
+  // restore our stack
+}
+
 void load_recv(uint8_t c)
 {
   if (state.command == CMD_NONE)
@@ -119,8 +138,7 @@ void load_recv(uint8_t c)
       state.command = CMD_NONE;
       debug("prejump\n");
       holly.VO_BORDER_COL = (31 << 11);
-      void (*fptr)(void) = (void (*)(void))state.addr1;
-      fptr();
+      jump_to_func(state.addr1);
       holly.VO_BORDER_COL = (63 << 5) | (31 << 0);
       debug("postjump\n");
       return;
diff --git a/sh7091/serial.cpp b/sh7091/serial.cpp
index ca8361e..cc4073a 100644
--- a/sh7091/serial.cpp
+++ b/sh7091/serial.cpp
@@ -19,11 +19,20 @@ void init(uint8_t bit_rate)
 
   sh7091.SCIF.SCFCR2 = scfcr2::tfrst::reset_operation_enabled
                      | scfcr2::rfrst::reset_operation_enabled;
-  // tx/rx trigger on 1 byte
-  sh7091.SCIF.SCFCR2 = 0;
+
+  sh7091.SCIF.SCFCR2 = scfcr2::rtrg::trigger_on_1_byte
+		     | scfcr2::ttrg::trigger_on_8_bytes
+		     | scfcr2::mce::modem_signals_disabled;
 
   sh7091.SCIF.SCSPTR2 = 0;
-  sh7091.SCIF.SCLSR2 = 0;
+  sh7091.SCIF.SCFSR2 = (~scfsr2::er::bit_mask)
+		     & (~scfsr2::tend::bit_mask)
+		     & (~scfsr2::tdfe::bit_mask)
+		     & (~scfsr2::brk::bit_mask)
+		     & (~scfsr2::rdf::bit_mask)
+		     & (~scfsr2::dr::bit_mask)
+		     & 0xffff;
+  sh7091.SCIF.SCLSR2 = 0; // clear ORER
 
   sh7091.SCIF.SCSCR2 = scscr2::te::transmission_enabled
                      | scscr2::re::reception_enabled;
diff --git a/start.s b/start.s
index 2268c9b..fe70eb7 100644
--- a/start.s
+++ b/start.s
@@ -10,11 +10,22 @@ _start:
         or      r1,r0
         ldc     r0,sr
 
+	/* save pr */
+	sts.l pr,@-r15
+
         /* jump to runtime_init */
         mov.l runtime_init_ptr,r0
-        jmp @r0
+        jsr @r0
         nop
 
+	/* restore pr */
+	lds.l @r15+,pr
+
+	/* jump to main */
+	mov.l main_ptr,r0
+	jmp @r0
+	nop
+
         .align 4
 p1ram_end_ptr:
         .long __p1ram_end
@@ -22,3 +33,5 @@ imask_all:
         .long 0xf0
 runtime_init_ptr:
         .long _runtime_init
+main_ptr:
+	.long _main
diff --git a/symbols.lds b/symbols.lds
new file mode 100644
index 0000000..02f1ea8
--- /dev/null
+++ b/symbols.lds
@@ -0,0 +1,21 @@
+__p1ram_start = ORIGIN(p1ram);
+/* reserve bytes for serial program loader */
+__p1ram_end = ORIGIN(p1ram) + LENGTH(p1ram) - __stack_reservation;
+
+__text_link_start = ADDR(.text);
+__text_link_end = ADDR(.text) + SIZEOF(.text);
+__text_load_start = LOADADDR(.text);
+
+__data_link_start = ADDR(.data);
+__data_link_end = ADDR(.data) + SIZEOF(.data);
+__data_load_start = LOADADDR(.data);
+
+__rodata_link_start = ADDR(.rodata);
+__rodata_link_end = ADDR(.rodata) + SIZEOF(.rodata);
+__rodata_load_start = LOADADDR(.rodata);
+
+__ctors_link_start = ADDR(.ctors);
+__ctors_link_end = ADDR(.ctors) + SIZEOF(.ctors);
+
+__bss_link_start = ADDR(.bss);
+__bss_link_end = ADDR(.bss) + SIZEOF(.bss);