diff --git a/6502.S b/6502.S
index 762086f..33ff11a 100644
--- a/6502.S
+++ b/6502.S
@@ -12,9 +12,8 @@
 .extern prn
 .macro prn reg
   push {r0-r3}
-  mov r0, \reg
-  bl prn
-  pop {r0-r3}
+  bl   prn
+  pop  {r0-r3}
 .endm
 
 #define T0 r4
@@ -36,7 +35,18 @@
 #define FLAG_C (1 << 0)
 
 .macro _ldb
+.ifdef USE_FRAMEPOINTER
+  // If -fomit-frame-pointer is not specified, save and restore r12 here.
+  push {r1}
+  mov  r1, r12
+  push {r1}
+.endif
   bl   cpu6502_load
+.ifdef USE_FRAMEPOINTER
+  pop  {r1}
+  mov  r12, r1
+  pop  {r1}
+.endif
 .endm
 
 .macro _ldw
@@ -50,7 +60,17 @@
 .endm
 
 .macro _stb
+.ifdef USE_FRAMEPOINTER
+  push {r1}
+  mov  r1, r12
+  push {r1}
+.endif
   bl   cpu6502_store
+.ifdef USE_FRAMEPOINTER
+  pop  {r1}
+  mov  r12, r1
+  pop  {r1}
+.endif
 .endm
 
 .macro _stw
@@ -1244,7 +1264,10 @@ op99:  // STA - Absolute, Y
   _decode
 
 op9a:  // TXS
-  _t   RX, SR
+  _t   RX, SP
+  movs r0, #1
+  lsls r0, r0, #8
+  add  SP, SP, r0
   _decode
 
 op9d:  // STA - Absolute, X
diff --git a/Makefile b/Makefile
index 6a40d07..b8b2d42 100644
--- a/Makefile
+++ b/Makefile
@@ -13,6 +13,8 @@ CLOCK	= 12000
 ROM	= applebasic
 ROMOBJ	= $(ROM).o
 ROMFLAG	= -I binary -O elf32-littlearm -B arm --rename-section .data=.rodata
+FRAMEPTR= -fomit-frame-pointer
+#FRAMEPTR= -DUSE_FRAMEPOINTER
 OBJS	= vectors.o reset.o 6502.o apple2.o uart.o $(ROMOBJ)
 
 $(APP).bin: $(APP)
@@ -37,8 +39,8 @@ $(ROM).rom:
 		--strip-symbol _binary_$(ROM)_rom_size
 
 # Test binary that runs on qemu user mode emulation for testing
-test: 6502.o test.c
-	$(CC) -mthumb -static test.c 6502.o -o test && qemu-arm test
+test: 6502.S test.c
+	$(CC) -mthumb -static $(FRAMEPTR) test.c 6502.S -o test && qemu-arm test
 
 # Assume a CQ Mary comaptible board.
 run: $(APP).bin
diff --git a/test.c b/test.c
index aa8af8d..c291ca1 100644
--- a/test.c
+++ b/test.c
@@ -5,27 +5,39 @@
 uint8_t mem[0x10000];
 
 void prn(int c) {
+  uint8_t fp;
+  __asm__("movs %0, r12": "=r"(fp));
   printf("### $%02x ###\n", c);
+  __asm__("movs r12, %0":: "r"(fp));
 }
 
 void cpu6502_dump(
     uint16_t pc, uint8_t a, uint8_t x, uint8_t y, uint8_t sp, uint8_t sr) {
+  uint8_t fp;
+  __asm__("movs %0, r12": "=r"(fp));
   printf("*** dump *** PC=$%04x A=$%02x X=$%02x Y=$%02x SP=$%02x "
          "NV-B_DIZC=%d%d-%d_%d%d%d%d\n",
          pc, a, x, y, sp, (sr >> 7) & 1, (sr >> 6) & 1, (sr >> 4) & 1,
          (sr >> 3) & 1, (sr >> 2) & 1, (sr >> 1) & 1, sr & 1);
   fflush(stdout);
+  __asm__("movs r12, %0":: "r"(fp));
 }
 
 uint8_t cpu6502_load(uint16_t addr) {
+  uint8_t fp;
+  __asm__("movs %0, r12": "=r"(fp));
   printf("load  $%04x => $%02x\n", addr, mem[addr]);
   fflush(stdout);
+  __asm__("movs r12, %0":: "r"(fp));
   return mem[addr];
 }
 
 void cpu6502_store(uint16_t addr, uint8_t data) {
+  uint8_t fp;
+  __asm__("movs %0, r12": "=r"(fp));
   printf("store $%04x <= $%02x\n", addr, data);
   fflush(stdout);
+  __asm__("movs r12, %0":: "r"(fp));
   mem[addr] = data;
 }