diff --git a/Makefile b/Makefile
index a864783..93afc67 100644
--- a/Makefile
+++ b/Makefile
@@ -52,10 +52,18 @@ $(OUT)/tests/%.elf: tests/%.c $(OUT)/$(STAGE0)
 	chmod +x $@ ; $(PRINTF) "Running $@ ...\n"
 	$(Q)$(TARGET_EXEC) $@ && $(call pass)
 
-check: $(TESTBINS) tests/driver.sh
-	tests/driver.sh
+check: check-stage0 check-stage2
+
+check-stage0: $(OUT)/$(STAGE0) $(TESTBINS) tests/driver.sh
+	$(VECHO) "  TEST STAGE 0\n"
+	tests/driver.sh 0
+
+check-stage2: $(OUT)/$(STAGE2) $(TESTBINS) tests/driver.sh
+	$(VECHO) "  TEST STAGE 2\n"
+	tests/driver.sh 2
 
 check-snapshots: $(OUT)/$(STAGE0) $(SNAPSHOTS) tests/check-snapshots.sh
+	$(VECHO) "  TEST SNAPSHOTS\n"
 	tests/check-snapshots.sh
 
 $(OUT)/%.o: %.c
diff --git a/README.md b/README.md
index f22e636..fced1a8 100644
--- a/README.md
+++ b/README.md
@@ -112,17 +112,27 @@ Verify that the emitted IRs are identical to the snapshots by specifying `check-
 $ make check-snapshots
 ```
 
-`shecc` comes with unit tests. To run the tests, give `check` as an argument:
+`shecc` comes with unit tests consist of stage 0, stage 2. To run these tests, give `check` as an argument:
 ```shell
 $ make check
 ```
 
 Reference output:
 ```
+  TEST STAGE 0
 ...
 int main(int argc, int argv) { exit(sizeof(char)); } => 1
 int main(int argc, int argv) { int a; a = 0; switch (3) { case 0: return 2; case 3: a = 10; break; case 1: return 0; } exit(a); } => 10
 int main(int argc, int argv) { int a; a = 0; switch (3) { case 0: return 2; default: a = 10; break; } exit(a); } => 10
+OK
+  TEST STAGE 2
+...
+int main(int argc, int argv) { exit(sizeof(char*)); }
+exit code => 4
+output => 
+int main(int argc, int argv) { exit(sizeof(int*)); }
+exit code => 4
+output => 
 OK
 ```
 
diff --git a/lib/c.c b/lib/c.c
index 722d35b..8627086 100644
--- a/lib/c.c
+++ b/lib/c.c
@@ -14,6 +14,7 @@
 #define false 0
 
 #if defined(__arm__)
+#define __SIZEOF_POINTER__ 4
 #define __syscall_exit 1
 #define __syscall_read 3
 #define __syscall_write 4
@@ -23,6 +24,7 @@
 #define __syscall_munmap 91
 
 #elif defined(__riscv)
+#define __SIZEOF_POINTER__ 4
 #define __syscall_exit 93
 #define __syscall_read 63
 #define __syscall_write 64
diff --git a/src/arm-codegen.c b/src/arm-codegen.c
index 5835961..4658a4b 100644
--- a/src/arm-codegen.c
+++ b/src/arm-codegen.c
@@ -336,8 +336,11 @@ void emit_ph2_ir(ph2_ir_t *ph2_ir)
         }
         interm = __r8;
         /* div/mod emulation */
-        /* Preserve the values of the dividend and divisor */
-        emit(__stmdb(__AL, 1, __sp, (1 << rn) | (1 << rm)));
+        /* Preserve the values of the dividend and divisor. This also keeps
+         * stack pointer to be aligned.
+         */
+        emit(__stmdb(__AL, 1, __sp,
+                     (1 << __r11) | (1 << __r12) | (1 << rn) | (1 << rm)));
         /* Obtain absolute values of the dividend and divisor */
         emit(__srl_amt(__AL, 0, arith_rs, __r8, rn, 31));
         emit(__add_r(__AL, rn, rn, __r8));
@@ -382,7 +385,8 @@ void emit_ph2_ir(ph2_ir_t *ph2_ir)
          * in rd.
          */
         emit(__mov_r(__AL, __r9, rn));
-        emit(__ldm(__AL, 1, __sp, (1 << rn) | (1 << rm)));
+        emit(__ldm(__AL, 1, __sp,
+                   (1 << __r11) | (1 << __r12) | (1 << rn) | (1 << rm)));
         emit(__mov_r(__AL, rd, interm));
         /* Handle the correct sign for the quotient or remainder */
         emit(__cmp_i(__AL, __r10, 0));
diff --git a/src/parser.c b/src/parser.c
index 484de1b..4171725 100644
--- a/src/parser.c
+++ b/src/parser.c
@@ -844,7 +844,11 @@ void read_expr_operand(block_t *parent, basic_block_t **bb)
         read_ternary_operation(parent, bb);
         lex_expect(T_close_bracket);
     } else if (lex_accept(T_sizeof)) {
+        /* TODO: Use more generalized type grammar parsing function to handle
+         * type reading
+         */
         char token[MAX_TYPE_LEN];
+        int ptr_cnt = 0;
 
         lex_expect(T_open_bracket);
         int find_type_flag = lex_accept(T_struct) ? 2 : 1;
@@ -853,9 +857,12 @@ void read_expr_operand(block_t *parent, basic_block_t **bb)
         if (!type)
             error("Unable to find type");
 
+        while (lex_accept(T_asterisk))
+            ptr_cnt++;
+
         ph1_ir = add_ph1_ir(OP_load_constant);
         vd = require_var(parent);
-        vd->init_val = type->size;
+        vd->init_val = ptr_cnt ? PTR_SIZE : type->size;
         strcpy(vd->var_name, gen_name());
         ph1_ir->dest = vd;
         opstack_push(vd);
diff --git a/src/reg-alloc.c b/src/reg-alloc.c
index fe0dc5d..f2cbfb5 100644
--- a/src/reg-alloc.c
+++ b/src/reg-alloc.c
@@ -12,6 +12,18 @@
  * dead variable and does NOT wrtie it back to the stack.
  */
 
+/* Aligns size to nearest multiple of 4, this meets
+ * ARMv7's alignment requirement.
+ *
+ * This function should
+ * be called whenever handling with user-defined type's
+ * size.
+ */
+int align_size(int i)
+{
+    return i <= 4 ? 4 : (i + 3) & ~3;
+}
+
 bool check_live_out(basic_block_t *bb, var_t *var)
 {
     for (int i = 0; i < bb->live_out_idx; i++) {
@@ -239,11 +251,11 @@ void reg_alloc()
                 src0 = GLOBAL_FUNC.stack_size;
                 if (global_insn->rd->is_ptr)
                     GLOBAL_FUNC.stack_size +=
-                        (PTR_SIZE * global_insn->rd->array_size);
+                        align_size(PTR_SIZE * global_insn->rd->array_size);
                 else {
                     type_t *type = find_type(global_insn->rd->type_name, 0);
                     GLOBAL_FUNC.stack_size +=
-                        (global_insn->rd->array_size * type->size);
+                        align_size(global_insn->rd->array_size * type->size);
                 }
 
                 dest =
@@ -260,7 +272,7 @@ void reg_alloc()
                          strcmp(global_insn->rd->type_name, "char") &&
                          strcmp(global_insn->rd->type_name, "_Bool")) {
                     type_t *type = find_type(global_insn->rd->type_name, 0);
-                    GLOBAL_FUNC.stack_size += type->size;
+                    GLOBAL_FUNC.stack_size += align_size(type->size);
                 } else
                     /* 'char' is aligned to one byte for the convenience */
                     GLOBAL_FUNC.stack_size += 4;
@@ -365,9 +377,10 @@ void reg_alloc()
                     }
 
                     if (insn->rd->array_size)
-                        fn->func->stack_size += (insn->rd->array_size * sz);
+                        fn->func->stack_size +=
+                            align_size(insn->rd->array_size * sz);
                     else
-                        fn->func->stack_size += sz;
+                        fn->func->stack_size += align_size(sz);
 
                     dest = prepare_dest(bb, insn->rd, -1, -1);
                     ir = bb_add_ph2_ir(bb, OP_address_of);
diff --git a/src/riscv-codegen.c b/src/riscv-codegen.c
index 92329df..0858796 100644
--- a/src/riscv-codegen.c
+++ b/src/riscv-codegen.c
@@ -72,7 +72,7 @@ void update_elf_offset(ph2_ir_t *ph2_ir)
         if (hard_mul_div)
             elf_offset += 4;
         else
-            elf_offset += 104;
+            elf_offset += 108;
         return;
     case OP_load_data_address:
     case OP_neq:
@@ -331,12 +331,13 @@ void emit_ph2_ir(ph2_ir_t *ph2_ir)
         /* Unsigned integer division */
         emit(__addi(__t0, __zero, 0));
         emit(__addi(__t1, __zero, 1));
-        emit(__beq(__t3, __zero, 48));
-        emit(__beq(__t2, __zero, 44));
+        emit(__beq(__t3, __zero, 52));
+        emit(__beq(__t2, __zero, 48));
+        emit(__beq(__t2, __t3, 20));
         emit(__bltu(__t2, __t3, 16));
         emit(__slli(__t3, __t3, 1));
         emit(__slli(__t1, __t1, 1));
-        emit(__jal(__zero, -12));
+        emit(__jal(__zero, -16));
         emit(__bltu(__t2, __t3, 12));
         emit(__sub(__t2, __t2, __t3));
         emit(__add(__t0, __t0, __t1));
diff --git a/tests/driver.sh b/tests/driver.sh
index 2f318a6..3c2df09 100755
--- a/tests/driver.sh
+++ b/tests/driver.sh
@@ -2,7 +2,22 @@
 
 set -u
 
-readonly SHECC="$PWD/out/shecc"
+if [ "$#" != 1 ]; then
+    echo "Usage: $0 <stage>"
+    exit 1
+fi
+
+case "$1" in
+    "0")
+        readonly SHECC="$PWD/out/shecc" ;;
+    "1")
+        readonly SHECC="$PWD/out/shecc-stage1.elf" ;;
+    "2")
+        readonly SHECC="$PWD/out/shecc-stage2.elf" ;;
+    *)
+        echo "$1 is not a valid stage"
+        exit 1 ;;
+esac
 
 # try - test shecc with given code
 # Usage:
@@ -405,8 +420,36 @@ items 5 "int a; a = 10; a -= 5; return a;"
 items 20 "int *p; int a[3]; a[0] = 10; a[1] = 20; a[2] = 30; p = a; p+=1; return p[0];"
 
 # sizeof
-expr 4 "sizeof(int)";
+expr 0 "sizeof(void)";
+expr 1 "sizeof(_Bool)";
 expr 1 "sizeof(char)";
+expr 4 "sizeof(int)";
+# sizeof pointers
+expr 4 "sizeof(void*)";
+expr 4 "sizeof(_Bool*)";
+expr 4 "sizeof(char*)";
+expr 4 "sizeof(int*)";
+# sizeof multi-level pointer
+expr 4 "sizeof(void**)";
+expr 4 "sizeof(_Bool**)";
+expr 4 "sizeof(char**)";
+expr 4 "sizeof(int**)";
+# sizeof struct
+try_ 4 << EOF
+typedef struct {
+    int a;
+    int b;
+} struct_t;
+int main() { return sizeof(struct_t*); }
+EOF
+# sizeof enum
+try_ 4 << EOF
+typedef enum {
+    A,
+    B
+} enum_t;
+int main() { return sizeof(enum_t*); }
+EOF
 
 # switch-case
 items 10 "int a; a = 0; switch (3) { case 0: return 2; case 3: a = 10; break; case 1: return 0; } return a;"