diff --git a/am/arch/mips32-npc/include/npc.h b/am/arch/mips32-npc/include/npc.h
index b98a23c6..6192aba8 100644
--- a/am/arch/mips32-npc/include/npc.h
+++ b/am/arch/mips32-npc/include/npc.h
@@ -90,10 +90,10 @@ typedef struct {
 #define EXC_TRAP    13
 
 #define MFC0(dst, src, sel) \
-asm volatile("mfc0 %0, $"_STR(src)", %1\n\t":"=r"(dst):"i"(sel))
+asm volatile("nop; nop; nop; mfc0 %0, $"_STR(src)", %1; nop; nop; nop\n\t":"=r"(dst):"i"(sel))
 
 #define MTC0(dst, src, sel) \
-asm volatile("mtc0 %0, $"_STR(dst)", %1\n\t"::"g"(src),"i"(sel))
+asm volatile("nop; nop; mtc0 %0, $"_STR(dst)", %1; nop; nop\n\t"::"g"(src),"i"(sel))
 
 #define _STR(x) _VAL(x)
 #define _VAL(x) #x
diff --git a/am/arch/mips32-npc/src/asye.c b/am/arch/mips32-npc/src/asye.c
index dd3e2f35..d3d6a1df 100644
--- a/am/arch/mips32-npc/src/asye.c
+++ b/am/arch/mips32-npc/src/asye.c
@@ -8,13 +8,11 @@ static _RegSet* (*H) (_Event, _RegSet*) = NULL;
 void print_timer() {
   int compare = 0;
   MFC0(compare, CP0_COMPARE, 0);
-  int count = 0;
-  MFC0(count, CP0_COUNT, 0);
-  printk("\e[33m[AM]\e[0m:%d, %d\n", compare, count);
-
-  static int counter = 0;
-  counter ++;
-  if(counter > 40) _halt(0);
+  int count0 = 0;
+  MFC0(count0, CP0_COUNT, 0);
+  int count1 = 0;
+  MFC0(count1, CP0_COUNT, 1);
+  printk("\e[33m[AM]\e[0m: compare:%d, count0:%d, count1:%d\n", compare, count0, count1);
 }
 
 void update_timer(uint32_t step) {
@@ -53,7 +51,7 @@ _RegSet *_make(_Area kstack, void (*entry)(void *), void *args){
 
 void _yield(){
   init_timer(INTERVAL);
-  asm volatile("addiu $a0, $0, -1; syscall; nop");
+  asm volatile("nop; li $a0, -1; syscall; nop");
 }
 
 int _intr_read() {
@@ -73,6 +71,9 @@ void irq_handle(struct _RegSet *regs){
   cp0_cause_t *cause = (void*)&(regs->cause);
   uint32_t exccode = cause->ExcCode;
   uint32_t ipcode = cause->IP;
+  
+  // print_timer();
+  // printk("[AM] cause:%x, status:%x, code:%x, ip:%x\n", regs->cause, regs->status, exccode, ipcode);
 
   _Event ev;
   ev.event = _EVENT_NULL;
@@ -107,7 +108,8 @@ void irq_handle(struct _RegSet *regs){
     case EXC_RI:
     case EXC_OV:
     default:
-	  printk("unhandled exccode = %x, epc:%08x\n", exccode, regs->epc);
+	  printk("unhandled exccode = %x, epc:%08x, badvaddr:%08x\n", exccode, regs->epc, regs->badvaddr);
+	  printk("cp0: base:%08x, cause:%08x, status:%08x\n", regs->base, regs->cause, regs->status);
 	  _halt(-1);
   }
 
@@ -117,7 +119,16 @@ void irq_handle(struct _RegSet *regs){
 	  _RegSet *next = H(ev, regs);
 	  if(next != NULL) ret = next;
   }
-  // printf("rets: t0:%x, t1:%x, t2:%x, t3:%x\n", ret->t0, ret->t1, ret->t2, ret->t3);
+  /*
+  printk("======================\n");
+  int count0 = 0;
+  MFC0(count0, CP0_COUNT, 0);
+  while(count0 > 0) {
+	_putc(count0 % 10 + '0');
+	count0 /= 10;
+  }
+  _putc('\n');
+  */
 
   // restore common registers
   asm volatile(
diff --git a/am/arch/mips32-npc/src/ioe.c b/am/arch/mips32-npc/src/ioe.c
index b4d4f3c2..db2e574f 100644
--- a/am/arch/mips32-npc/src/ioe.c
+++ b/am/arch/mips32-npc/src/ioe.c
@@ -171,7 +171,7 @@ int e0_scan_code[256] = {
   [0x7D] = _KEY_PAGEUP,
 };
 
-size_t input_read(uintptr_t reg, void *buf, size_t size) {
+static size_t keyboard_read(uintptr_t reg, void *buf, size_t size) {
   int code = in_scancode();
   int *table = normal_scancode;
 
@@ -192,7 +192,7 @@ size_t input_read(uintptr_t reg, void *buf, size_t size) {
 
 static _Device mips32_npc_dev[] = {
   {_DEV_SERIAL,  "NOOP Serial Controller", uartlite_read, uartlite_write},
-  {_DEV_INPUT,   "NOOP Keyboard Controller", input_read, NULL},
+  {_DEV_INPUT,   "NOOP Keyboard Controller", keyboard_read, NULL},
   {_DEV_TIMER,   "NOOP Fake Timer",   timer_read, NULL},
   {_DEV_VIDEO,   "NoneStandard VGA Controller",  video_read, video_write},
 };
diff --git a/am/arch/mips32-npc/src/trap.S b/am/arch/mips32-npc/src/trap.S
index 6de76c08..c14e2529 100644
--- a/am/arch/mips32-npc/src/trap.S
+++ b/am/arch/mips32-npc/src/trap.S
@@ -45,15 +45,35 @@ _trap_entry:
 
 	addu	$sp, $k0, $0
 
+	nop
+	nop
 	mfc0	$k0, $14
+	nop
+	nop
 	sw		$k0, 0x80($sp)
+	nop
+	nop
 	mfc0	$k0, $13
+	nop
+	nop
 	sw		$k0, 0x84($sp)
+	nop
+	nop
 	mfc0	$k0, $12
+	nop
+	nop
 	sw		$k0, 0x88($sp)
+	nop
+	nop
 	mfc0	$k0, $8
+	nop
+	nop
 	sw		$k0, 0x8c($sp)
+	nop
+	nop
 	mfc0	$k0, $7
+	nop
+	nop
 	sw		$k0, 0x90($sp)
 # give trapframe pointer to a0
 	addiu   $a0, $sp, 0x4
diff --git a/tests/asyetest/main.c b/tests/asyetest/main.c
index 7a3bb6ae..c49979b4 100644
--- a/tests/asyetest/main.c
+++ b/tests/asyetest/main.c
@@ -35,9 +35,8 @@ _RegSet* handler(_Event ev, _RegSet *regs) {
 #define _STR(x) _VAL(x)
 #define _VAL(x) #x
 
-/*
 #define MFC0(dst, src, sel) \
-asm volatile("mfc0 %0, $"_STR(src)", %1\n\t":"=r"(dst):"i"(sel))
+asm volatile("nop; nop; mfc0 %0, $"_STR(src)", %1; nop; nop\n\t":"=r"(dst):"i"(sel))
 
 void read_write_cp0() {
   int reg;
@@ -45,29 +44,29 @@ void read_write_cp0() {
   reg = 0x12345678; \
   MFC0(reg, src, sel); \
   printf("CP0:{" #src ", " #sel "}=%x\n", reg);
-  CP0_OUT(8, 0)
-  CP0_OUT(9, 0)
-  CP0_OUT(9, 1)
-  CP0_OUT(11, 0)
-  CP0_OUT(12, 0)
-  CP0_OUT(13, 0)
-  CP0_OUT(14, 0)
+  CP0_OUT(8, 0) // badvaddr
+  CP0_OUT(9, 0) // count0
+  CP0_OUT(9, 1) // count1
+  CP0_OUT(11, 0) // compare
+  CP0_OUT(12, 0) // status
+  CP0_OUT(13, 0) // cause
+  CP0_OUT(14, 0) // epc
 }
-*/
 
 int main(){
   _ioe_init();
   _asye_init(handler);
-  // read_write_cp0();
+  read_write_cp0();
   assert(!_intr_read());
   _intr_write(1);
   //_make(_heap, void*)main, 0);
   printf("intr:%x\n", _intr_read());
+  _yield();
   while (1) {
 	for(int i = 0; i < 10; i++)
 	  printf("-");
 	printf(":yield, intr:%x\n", _intr_read());
-    _yield();
+	while(1); // no yield
   }
   return 0;
 }
diff --git a/tests/cputest/tests/clz.c b/tests/cputest/tests/clz.c
new file mode 100644
index 00000000..7b5e1c51
--- /dev/null
+++ b/tests/cputest/tests/clz.c
@@ -0,0 +1,35 @@
+#include "trap.h"
+
+int clz(uint32_t x) {
+  if(!x) return 32;
+  static const char debruijn32[32] = {
+	0, 31, 9, 30, 3, 8, 13, 29, 2, 5, 7, 21, 12, 24, 28, 19,
+	1, 10, 4, 14, 6, 22, 25, 20, 11, 15, 23, 26, 16, 27, 17, 18
+  };
+  x |= x>>1;
+  x |= x>>2;
+  x |= x>>4;
+  x |= x>>8;
+  x |= x>>16;
+  x++;
+  return debruijn32[x*0x076be629>>27];
+}
+
+int main() {
+#ifdef __ISA_MIPS32__
+  {
+	// special case, in == 0
+	int in = 0, out = 0;
+	asm volatile("clz %0, %1": "=r"(out) : "r"(in));
+	nemu_assert(out == clz(in));
+  }
+
+  for(int i = 0; i < 32; i++) {
+	int in = 1 << i;
+	int out = 0;
+	asm volatile("clz %0, %1": "=r"(out) : "r"(in));
+	nemu_assert(out == clz(in));
+  }
+#endif
+  return 0;
+}
diff --git a/tests/segmenttest/Makefile b/tests/segmenttest/Makefile
new file mode 100644
index 00000000..efcaa34f
--- /dev/null
+++ b/tests/segmenttest/Makefile
@@ -0,0 +1,19 @@
+NAME = segmenttest
+SRCS = main.c initrd.S
+LIBS += klib compiler-rt
+ARCH ?= mips32-npc
+
+TESTCASE = load-store
+INITRD_S = initrd.S
+
+TESTBIN = $(AM_HOME)/tests/cputest/build/$(TESTCASE)-$(ARCH).bin
+
+$(TESTBIN): $(AM_HOME)/tests/cputest/tests/$(TESTCASE).c
+	make -s -C $(@D)/.. ARCH=$(ARCH)
+
+$(INITRD_S): $(TESTBIN)
+	@cp $< build/fsimg
+	@touch $@
+	echo + $@
+
+include $(AM_HOME)/Makefile.app
diff --git a/tests/segmenttest/initrd.S b/tests/segmenttest/initrd.S
new file mode 100644
index 00000000..28574af4
--- /dev/null
+++ b/tests/segmenttest/initrd.S
@@ -0,0 +1,5 @@
+.section .data
+.global __fsimg_start, __fsimg_end
+__fsimg_start:
+.incbin "build/fsimg"
+__fsimg_end:
diff --git a/tests/segmenttest/main.c b/tests/segmenttest/main.c
new file mode 100644
index 00000000..a5730e4a
--- /dev/null
+++ b/tests/segmenttest/main.c
@@ -0,0 +1,49 @@
+#include <am.h>
+#include <amdev.h>
+#include <klib.h>
+
+_RegSet testcase_regs;
+
+int ntraps = 0;
+_RegSet* handler(_Event ev, _RegSet *regs) {
+  switch (ev.event) {
+    case _EVENT_IRQ_TIMER:
+      printf(".");
+      break;
+    case _EVENT_IRQ_IODEV: 
+      while (1) {
+        int key = read_key();
+        if (key == _KEY_NONE) break;
+        if (key & 0x8000) {
+          printf("[D:%d]", key ^ 0x8000);
+        } else {
+          printf("[U:%d]/[TRAPS:%d]", key, ntraps);
+          ntraps = 0;
+        }
+      }
+      break;
+    case _EVENT_YIELD:
+      ntraps++;
+      break;
+  }
+  return &testcase_regs;
+}
+
+void load_testcase() {
+  extern char __fsimg_start;
+  extern char __fsimg_end;
+}
+
+int main(){
+  _ioe_init();
+  _asye_init(handler);
+
+  assert(!_intr_read());
+  _intr_write(1);
+
+  load_testcase();
+
+  _yield();
+  while(1) { }
+  return 0;
+}