Prechádzať zdrojové kódy

Implement wasm mini loader and refine footprint of loader and runtime (#276)

wenyongh 5 rokov pred
rodič
commit
7a287fd1a9

+ 8 - 2
build-scripts/config_common.cmake

@@ -133,10 +133,16 @@ endif ()
 if (WAMR_BUILD_SPEC_TEST EQUAL 1)
   add_definitions (-DWASM_ENABLE_SPEC_TEST=1)
   message ("     spec test compatible mode is on")
-endif()
+endif ()
 if (WAMR_BUILD_BULK_MEMORY EQUAL 1)
   add_definitions (-DWASM_ENABLE_BULK_MEMORY=1)
   message ("     Bulk memory feature enabled")
 else ()
   add_definitions (-DWASM_ENABLE_BULK_MEMORY=0)
-endif()
+endif ()
+if (WAMR_BUILD_MINI_LOADER EQUAL 1)
+  add_definitions (-DWASM_ENABLE_MINI_LOADER=1)
+  message ("     WASM mini loader enabled")
+else ()
+  add_definitions (-DWASM_ENABLE_MINI_LOADER=0)
+endif ()

+ 5 - 0
core/config.h

@@ -130,6 +130,11 @@ enum {
 #define WASM_ENABLE_MULTI_MODULE 0
 #endif
 
+/* Enable wasm mini loader or not */
+#ifndef WASM_ENABLE_MINI_LOADER
+#define WASM_ENABLE_MINI_LOADER 0
+#endif
+
 /* Heap and stack profiling */
 #define BH_ENABLE_MEMORY_PROFILING 0
 

+ 69 - 115
core/iwasm/aot/aot_loader.c

@@ -59,13 +59,24 @@ static union {
 
 #define is_little_endian() (__ue.b == 1)
 
-#define CHECK_BUF(buf, buf_end, length) do {                \
-    if (buf + length > buf_end) {                           \
-      set_error_buf(error_buf, error_buf_size,              \
-                    "Read data failed: unexpected end.");   \
-      goto fail;                                            \
-    }                                                       \
-  } while (0)
+static bool
+check_buf(const uint8 *buf, const uint8 *buf_end, uint32 length,
+          char *error_buf, uint32 error_buf_size)
+{
+    if (buf + length > buf_end) {
+        set_error_buf(error_buf, error_buf_size,
+                      "AOT module load failed: unexpect end.");
+        return false;
+    }
+    return true;
+}
+
+#define CHECK_BUF(buf, buf_end, length) do {    \
+  if (!check_buf(buf, buf_end, length,          \
+                 error_buf, error_buf_size)) {  \
+      goto fail;                                \
+  }                                             \
+} while (0)
 
 static uint8*
 align_ptr(const uint8 *p, uint32 b)
@@ -150,17 +161,32 @@ GET_U64_FROM_ADDR(uint32 *addr)
 /* Legal values for e_version */
 #define E_VERSION_CURRENT  1        /* Current version */
 
+static void *
+loader_malloc(uint64 size, char *error_buf, uint32 error_buf_size)
+{
+    void *mem;
+
+    if (size >= UINT32_MAX
+        || !(mem = wasm_runtime_malloc((uint32)size))) {
+        set_error_buf(error_buf, error_buf_size,
+                      "AOT module load failed: "
+                      "allocate memory failed.");
+        return NULL;
+    }
+
+    memset(mem, 0, (uint32)size);
+    return mem;
+}
+
 static char*
 const_str_set_insert(const uint8 *str, int32 len, AOTModule *module,
                      char* error_buf, uint32 error_buf_size)
 {
     HashMap *set = module->const_str_set;
-    char *c_str = wasm_runtime_malloc((uint32)len + 1), *value;
+    char *c_str, *value;
 
-    if (!c_str) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module load failed: "
-                      "allocate memory failed.");
+    if (!(c_str = loader_malloc((uint32)len + 1,
+                                error_buf, error_buf_size))) {
         return NULL;
     }
 
@@ -348,17 +374,11 @@ load_mem_init_data_list(const uint8 **p_buf, const uint8 *buf_end,
 
     /* Allocate memory */
     size = sizeof(AOTMemInitData *) * (uint64)module->mem_init_data_count;
-    if (size >= UINT32_MAX
-        || !(module->mem_init_data_list =
-             data_list = wasm_runtime_malloc((uint32)size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module load failed: "
-                      "allocate memory failed.");
+    if (!(module->mem_init_data_list = data_list =
+                loader_malloc(size, error_buf, error_buf_size))) {
         return false;
     }
 
-    memset(data_list, 0, size);
-
     /* Create each memory data segment */
     for (i = 0; i < module->mem_init_data_count; i++) {
         uint32 init_expr_type, byte_count;
@@ -372,11 +392,8 @@ load_mem_init_data_list(const uint8 **p_buf, const uint8 *buf_end,
         read_uint64(buf, buf_end, init_expr_value);
         read_uint32(buf, buf_end, byte_count);
         size = offsetof(AOTMemInitData, bytes) + (uint64)byte_count;
-        if (size >= UINT32_MAX
-            || !(data_list[i] = wasm_runtime_malloc((uint32)size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "AOT module load failed: "
-                          "allocate memory failed.");
+        if (!(data_list[i] = loader_malloc
+                    (size, error_buf, error_buf_size))) {
             return false;
         }
 
@@ -447,17 +464,11 @@ load_table_init_data_list(const uint8 **p_buf, const uint8 *buf_end,
 
     /* Allocate memory */
     size = sizeof(AOTTableInitData *) * (uint64)module->table_init_data_count;
-    if (size >= UINT32_MAX
-        || !(module->table_init_data_list =
-             data_list = wasm_runtime_malloc((uint32)size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module load failed: "
-                      "allocate memory failed.");
+    if (!(module->table_init_data_list = data_list =
+                loader_malloc(size, error_buf, error_buf_size))) {
         return false;
     }
 
-    memset(data_list, 0, size);
-
     /* Create each table data segment */
     for (i = 0; i < module->table_init_data_count; i++) {
         uint32 init_expr_type, func_index_count;
@@ -469,11 +480,8 @@ load_table_init_data_list(const uint8 **p_buf, const uint8 *buf_end,
 
         size1 = sizeof(uint32) * (uint64)func_index_count;
         size = offsetof(AOTTableInitData, func_indexes) + size1;
-        if (size >= UINT32_MAX
-            || !(data_list[i] = wasm_runtime_malloc((uint32)size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "AOT module load failed: "
-                          "allocate memory failed.");
+        if (!(data_list[i] = loader_malloc
+                    (size, error_buf, error_buf_size))) {
             return false;
         }
 
@@ -535,16 +543,11 @@ load_func_types(const uint8 **p_buf, const uint8 *buf_end,
 
     /* Allocate memory */
     size = sizeof(AOTFuncType *) * (uint64)module->func_type_count;
-    if (size >= UINT32_MAX
-        || !(module->func_types = func_types = wasm_runtime_malloc((uint32)size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module load failed: "
-                      "allocate memory failed.");
+    if (!(module->func_types = func_types = loader_malloc
+                (size, error_buf, error_buf_size))) {
         return false;
     }
 
-    memset(func_types, 0, size);
-
     /* Create each function type */
     for (i = 0; i < module->func_type_count; i++) {
         uint32 param_count, result_count;
@@ -555,11 +558,8 @@ load_func_types(const uint8 **p_buf, const uint8 *buf_end,
 
         size1 = (uint64)param_count + (uint64)result_count;
         size = offsetof(AOTFuncType, types) + size1;
-        if (size >= UINT32_MAX
-            || !(func_types[i] = wasm_runtime_malloc((uint32)size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "AOT module load failed: "
-                          "allocate memory failed.");
+        if (!(func_types[i] = loader_malloc
+                    (size, error_buf, error_buf_size))) {
             return false;
         }
 
@@ -613,17 +613,11 @@ load_import_globals(const uint8 **p_buf, const uint8 *buf_end,
 
     /* Allocate memory */
     size = sizeof(AOTImportGlobal) * (uint64)module->import_global_count;
-    if (size >= UINT32_MAX
-        || !(module->import_globals =
-             import_globals = wasm_runtime_malloc((uint32)size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module load failed: "
-                      "allocate memory failed.");
+    if (!(module->import_globals = import_globals =
+                loader_malloc(size, error_buf, error_buf_size))) {
         return false;
     }
 
-    memset(import_globals, 0, size);
-
     /* Create each import global */
     for (i = 0; i < module->import_global_count; i++) {
         buf = (uint8*)align_ptr(buf, 2);
@@ -685,16 +679,11 @@ load_globals(const uint8 **p_buf, const uint8 *buf_end,
 
     /* Allocate memory */
     size = sizeof(AOTGlobal) * (uint64)module->global_count;
-    if (size >= UINT32_MAX
-        || !(module->globals = globals = wasm_runtime_malloc((uint32)size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module load failed: "
-                      "allocate memory failed.");
+    if (!(module->globals = globals = loader_malloc
+                (size, error_buf, error_buf_size))) {
         return false;
     }
 
-    memset(globals, 0, size);
-
     if (module->import_global_count > 0) {
         last_import_global =
             &module->import_globals[module->import_global_count - 1];
@@ -767,17 +756,11 @@ load_import_funcs(const uint8 **p_buf, const uint8 *buf_end,
 
     /* Allocate memory */
     size = sizeof(AOTImportFunc) * (uint64)module->import_func_count;
-    if (size >= UINT32_MAX
-        || !(module->import_funcs =
-             import_funcs = wasm_runtime_malloc((uint32)size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module load failed: "
-                      "allocate memory failed.");
+    if (!(module->import_funcs = import_funcs =
+                loader_malloc(size, error_buf, error_buf_size))) {
         return false;
     }
 
-    memset(import_funcs, 0, size);
-
     /* Create each import func */
     for (i = 0; i < module->import_func_count; i++) {
         read_uint16(buf, buf_end, import_funcs[i].func_type_index);
@@ -860,17 +843,11 @@ load_object_data_sections(const uint8 **p_buf, const uint8 *buf_end,
 
     /* Allocate memory */
     size = sizeof(AOTObjectDataSection) * (uint64)module->data_section_count;
-    if (size >= UINT32_MAX
-        || !(module->data_sections =
-             data_sections = wasm_runtime_malloc((uint32)size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module load failed: "
-                      "allocate memory failed.");
+    if (!(module->data_sections = data_sections =
+                loader_malloc(size, error_buf, error_buf_size))) {
         return false;
     }
 
-    memset(data_sections, 0, size);
-
     /* Create each data section */
     for (i = 0; i < module->data_section_count; i++) {
         int map_prot = MMAP_PROT_READ | MMAP_PROT_WRITE;
@@ -1021,10 +998,8 @@ load_function_section(const uint8 *buf, const uint8 *buf_end,
     uint64 size, text_offset;
 
     size = sizeof(void*) * (uint64)module->func_count;
-    if (size >= UINT32_MAX
-        || !(module->func_ptrs = wasm_runtime_malloc((uint32)size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module load failed: allocate memory failed.");
+    if (!(module->func_ptrs = loader_malloc
+                (size, error_buf, error_buf_size))) {
         return false;
     }
 
@@ -1065,10 +1040,8 @@ load_function_section(const uint8 *buf, const uint8 *buf_end,
     }
 
     size = sizeof(uint32) * (uint64)module->func_count;
-    if (size >= UINT32_MAX
-        || !(module->func_type_indexes = wasm_runtime_malloc((uint32)size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module load failed: allocate memory failed.");
+    if (!(module->func_type_indexes = loader_malloc
+                (size, error_buf, error_buf_size))) {
         return false;
     }
 
@@ -1112,17 +1085,11 @@ load_export_funcs(const uint8 **p_buf, const uint8 *buf_end,
 
     /* Allocate memory */
     size = sizeof(AOTExportFunc) * (uint64)module->export_func_count;
-    if (size >= UINT32_MAX
-        || !(module->export_funcs =
-             export_funcs = wasm_runtime_malloc((uint32)size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module load failed: "
-                      "allocate memory failed.");
+    if (!(module->export_funcs = export_funcs =
+                loader_malloc(size, error_buf, error_buf_size))) {
         return false;
     }
 
-    memset(export_funcs, 0, size);
-
     /* Create each export func */
     for (i = 0; i < module->export_func_count; i++) {
         read_uint32(buf, buf_end, export_funcs[i].func_index);
@@ -1234,10 +1201,8 @@ do_text_relocation(AOTModule *module,
         if (symbol_len + 1 <= sizeof(symbol_buf))
             symbol = symbol_buf;
         else {
-            if (!(symbol = wasm_runtime_malloc(symbol_len + 1))) {
-                set_error_buf(error_buf, error_buf_size,
-                              "AOT module load failed: "
-                              "allocate memory failed.");
+            if (!(symbol = loader_malloc(symbol_len + 1,
+                                         error_buf, error_buf_size))) {
                 return false;
             }
         }
@@ -1432,15 +1397,10 @@ load_relocation_section(const uint8 *buf, const uint8 *buf_end,
 
     /* Allocate memory for relocation groups */
     size = sizeof(AOTRelocationGroup) * (uint64)group_count;
-    if (size >= UINT32_MAX || !(groups = wasm_runtime_malloc((uint32)size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module load failed: "
-                      "allocate memory failed.");
+    if (!(groups = loader_malloc(size, error_buf, error_buf_size))) {
         goto fail;
     }
 
-    memset(groups, 0, size);
-
     /* Load each relocation group */
     for (i = 0, group = groups; i < group_count; i++, group++) {
         AOTRelocation *relocation;
@@ -1473,18 +1433,12 @@ load_relocation_section(const uint8 *buf, const uint8 *buf_end,
 
         /* Allocate memory for relocations */
         size = sizeof(AOTRelocation) * (uint64)group->relocation_count;
-        if (size >= UINT32_MAX
-            || !(group->relocations = relocation =
-                 wasm_runtime_malloc((uint32)size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "AOT module load failed: "
-                          "allocate memory failed.");
+        if (!(group->relocations = relocation =
+                    loader_malloc(size, error_buf, error_buf_size))) {
             ret = false;
             goto fail;
         }
 
-        memset(group->relocations, 0, size);
-
         /* Load each relocation */
         for (j = 0; j < group->relocation_count; j++, relocation++) {
             uint32 symbol_index;

+ 25 - 25
core/iwasm/aot/aot_runtime.c

@@ -14,6 +14,23 @@ set_error_buf(char *error_buf, uint32 error_buf_size, const char *string)
         snprintf(error_buf, error_buf_size, "%s", string);
 }
 
+static void *
+runtime_malloc(uint64 size, char *error_buf, uint32 error_buf_size)
+{
+    void *mem;
+
+    if (size >= UINT32_MAX
+        || !(mem = wasm_runtime_malloc((uint32)size))) {
+        set_error_buf(error_buf, error_buf_size,
+                      "AOT module instantiate failed: "
+                      "allocate memory failed.");
+        return NULL;
+    }
+
+    memset(mem, 0, (uint32)size);
+    return mem;
+}
+
 static bool
 global_instantiate(AOTModuleInstance *module_inst, AOTModule *module,
                    char *error_buf, uint32 error_buf_size)
@@ -142,15 +159,10 @@ memory_instantiate(AOTModuleInstance *module_inst, AOTModule *module,
     uint8 *p;
 
     /* Allocate memory */
-    if (total_size >= UINT32_MAX
-        || !(p = wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module instantiate failed: allocate memory failed.");
+    if (!(p = runtime_malloc(total_size, error_buf, error_buf_size))) {
         return false;
     }
 
-    memset(p, 0, (uint32)total_size);
-
     /* Initialize heap info */
     module_inst->heap_data.ptr = p;
     p += heap_size;
@@ -270,15 +282,11 @@ init_func_ptrs(AOTModuleInstance *module_inst, AOTModule *module,
         ((uint64)module->import_func_count + module->func_count) * sizeof(void*);
 
     /* Allocate memory */
-    if (total_size >= UINT32_MAX
-        || !(module_inst->func_ptrs.ptr = wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module instantiate failed: allocate memory failed.");
+    if (!(module_inst->func_ptrs.ptr = runtime_malloc
+                (total_size, error_buf, error_buf_size))) {
         return false;
     }
 
-    memset(module_inst->func_ptrs.ptr, 0, (uint32)total_size);
-
     /* Set import function pointers */
     func_ptrs = (void**)module_inst->func_ptrs.ptr;
     for (i = 0; i < module->import_func_count; i++, func_ptrs++)
@@ -299,16 +307,11 @@ init_func_type_indexes(AOTModuleInstance *module_inst, AOTModule *module,
         ((uint64)module->import_func_count + module->func_count) * sizeof(uint32);
 
     /* Allocate memory */
-    if (total_size >= UINT32_MAX
-        || !(module_inst->func_type_indexes.ptr =
-                                wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module instantiate failed: allocate memory failed.");
+    if (!(module_inst->func_type_indexes.ptr =
+                runtime_malloc(total_size, error_buf, error_buf_size))) {
         return false;
     }
 
-    memset(module_inst->func_type_indexes.ptr, 0, (uint32)total_size);
-
     /* Set import function type indexes */
     func_type_index = (uint32*)module_inst->func_type_indexes.ptr;
     for (i = 0; i < module->import_func_count; i++, func_type_index++)
@@ -381,14 +384,11 @@ aot_instantiate(AOTModule *module,
         heap_size = APP_HEAP_SIZE_MAX;
 
     /* Allocate module instance, global data, table data and heap data */
-    if (total_size >= UINT32_MAX
-        || !(module_inst = wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "AOT module instantiate failed: allocate memory failed.");
+    if (!(module_inst = runtime_malloc(total_size,
+                                       error_buf, error_buf_size))) {
         return NULL;
     }
 
-    memset(module_inst, 0, total_size);
     module_inst->module_type = Wasm_Module_AoT;
     module_inst->aot_module.ptr = module;
 
@@ -801,7 +801,7 @@ aot_enlarge_memory(AOTModuleInstance *module_inst, uint32 inc_page_count)
     /* Destroy heap's lock firstly, if its memory is re-allocated,
        we cannot access its lock again. */
     mem_allocator_destroy_lock(module_inst->heap_handle.ptr);
-    if (!(heap_data = wasm_runtime_realloc(heap_handle_old, (uint32)total_size))) {
+    if (!(heap_data = wasm_runtime_realloc(heap_data_old, (uint32)total_size))) {
         if (!(heap_data = wasm_runtime_malloc((uint32)total_size))) {
             /* Restore heap's lock if memory re-alloc failed */
             mem_allocator_reinit_lock(module_inst->heap_handle.ptr);

+ 47 - 41
core/iwasm/common/wasm_runtime_common.c

@@ -60,6 +60,29 @@ set_error_buf(char *error_buf, uint32 error_buf_size, const char *string)
         snprintf(error_buf, error_buf_size, "%s", string);
 }
 
+static void *
+runtime_malloc(uint64 size, WASMModuleInstanceCommon *module_inst,
+               char *error_buf, uint32 error_buf_size)
+{
+    void *mem;
+
+    if (size >= UINT32_MAX
+        || !(mem = wasm_runtime_malloc((uint32)size))) {
+        if (module_inst != NULL) {
+            wasm_runtime_set_exception(module_inst,
+                                       "allocate memory failed.");
+        }
+        else if (error_buf != NULL) {
+            set_error_buf(error_buf, error_buf_size,
+                          "allocate memory failed.");
+        }
+        return NULL;
+    }
+
+    memset(mem, 0, (uint32)size);
+    return mem;
+}
+
 static bool
 wasm_runtime_env_init()
 {
@@ -91,12 +114,12 @@ wasm_runtime_env_init()
 static bool
 wasm_runtime_exec_env_check(WASMExecEnv *exec_env)
 {
-    return !(!exec_env
-        || !exec_env->module_inst
-        || exec_env->wasm_stack_size == 0
-        || exec_env->wasm_stack.s.top_boundary !=
+    return exec_env
+           && exec_env->module_inst
+           && exec_env->wasm_stack_size > 0
+           && exec_env->wasm_stack.s.top_boundary ==
                 exec_env->wasm_stack.s.bottom + exec_env->wasm_stack_size
-        || exec_env->wasm_stack.s.top > exec_env->wasm_stack.s.top_boundary);
+           && exec_env->wasm_stack.s.top <= exec_env->wasm_stack.s.top_boundary;
 }
 
 bool
@@ -240,13 +263,10 @@ wasm_runtime_register_module_internal(const char *module_name,
     }
 
     /* module hasn't been registered */
-    node = wasm_runtime_malloc(sizeof(WASMRegisteredModule));
+    node = runtime_malloc(sizeof(WASMRegisteredModule), NULL, NULL, 0);
     if (!node) {
         LOG_DEBUG("malloc WASMRegisteredModule failed. SZ=%d",
                   sizeof(WASMRegisteredModule));
-        set_error_buf_v(error_buf, error_buf_size,
-                        "malloc WASMRegisteredModule failed. SZ=%d",
-                        sizeof(WASMRegisteredModule));
         return false;
     }
 
@@ -377,16 +397,15 @@ wasm_runtime_destroy_registered_module_list()
 }
 
 bool
-wasm_runtime_add_loading_module(const char *module_name, char *error_buf,
-                                uint32 error_buf_size)
+wasm_runtime_add_loading_module(const char *module_name,
+                                char *error_buf, uint32 error_buf_size)
 {
     LOG_DEBUG("add %s into a loading list", module_name);
-    LoadingModule *loadingModule = wasm_runtime_malloc(sizeof(LoadingModule));
+    LoadingModule *loadingModule =
+            runtime_malloc(sizeof(LoadingModule), NULL,
+                           error_buf, error_buf_size);
 
     if (!loadingModule) {
-        set_error_buf_v(error_buf, error_buf_size,
-                        "malloc LoadingModule failed. SZ=%d",
-                        sizeof(LoadingModule));
         return false;
     }
 
@@ -1145,13 +1164,11 @@ wasm_runtime_init_wasi(WASMModuleInstanceCommon *module_inst,
     uint64 total_size;
     uint32 i;
 
-    if (!(wasi_ctx = wasm_runtime_malloc(sizeof(WASIContext)))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "Init wasi environment failed: allocate memory failed.");
+    if (!(wasi_ctx = runtime_malloc(sizeof(WASIContext), NULL,
+                                    error_buf, error_buf_size))) {
         return false;
     }
 
-    memset(wasi_ctx, 0, sizeof(WASIContext));
     wasm_runtime_set_wasi_ctx(module_inst, wasi_ctx);
 
 #if WASM_ENABLE_INTERP != 0
@@ -1645,12 +1662,11 @@ resolve_function(const WASMModuleInstanceCommon *module_inst,
     char *function_name = NULL;
     uint32 length = strlen(name) + 1;
 
-    orig_name = wasm_runtime_malloc(sizeof(char) * length);
+    orig_name = runtime_malloc(sizeof(char) * length, NULL, NULL, 0);
     if (!orig_name) {
         return NULL;
     }
 
-    memset(orig_name, 0, sizeof(char) * length);
     strncpy(orig_name, name, length);
 
     if (!parse_function_name(orig_name, &sub_module_name, &function_name)) {
@@ -1808,9 +1824,8 @@ wasm_application_execute_func(WASMModuleInstanceCommon *module_inst,
     }
 
     total_size = sizeof(uint32) * (uint64)(argc1 > 2 ? argc1 : 2);
-    if (total_size >= UINT32_MAX
-        || (!(argv1 = wasm_runtime_malloc((uint32)total_size)))) {
-        wasm_runtime_set_exception(module_inst, "allocate memory failed.");
+    if ((!(argv1 = runtime_malloc((uint32)total_size, module_inst,
+                                  NULL, 0)))) {
         goto fail;
     }
 
@@ -2008,13 +2023,10 @@ wasm_runtime_invoke_native_raw(WASMExecEnv *exec_env, void *func_ptr,
     argc1 = func_type->param_count;
     if (argc1 > sizeof(argv_buf) / sizeof(uint64)) {
         size = sizeof(uint64) * (uint64)argc1;
-        if (size >= UINT32_MAX
-            || !(argv1 = wasm_runtime_malloc((uint32)size))) {
-            wasm_runtime_set_exception(exec_env->module_inst,
-                                       "allocate memory failed.");
+        if (!(argv1 = runtime_malloc((uint32)size, exec_env->module_inst,
+                                     NULL, 0))) {
             return false;
         }
-        memset(argv1, 0, (uint32)size);
     }
 
     argv_dst = argv1;
@@ -2208,10 +2220,8 @@ wasm_runtime_invoke_native(WASMExecEnv *exec_env, void *func_ptr,
     argc1 = MAX_REG_INTS + MAX_REG_FLOATS + n_stacks;
     if (argc1 > sizeof(argv_buf) / sizeof(uint32)) {
         size = sizeof(uint32) * (uint32)argc1;
-        if (size >= UINT32_MAX
-                || !(argv1 = wasm_runtime_malloc((uint32)size))) {
-            wasm_runtime_set_exception(exec_env->module_inst,
-                                       "allocate memory failed.");
+        if (!(argv1 = runtime_malloc((uint32)size, exec_env->module_inst,
+                                     NULL, 0))) {
             return false;
         }
     }
@@ -2386,10 +2396,8 @@ wasm_runtime_invoke_native(WASMExecEnv *exec_env, void *func_ptr,
 
     if (argc1 > sizeof(argv_buf) / sizeof(uint32)) {
         size = sizeof(uint32) * (uint64)argc1;
-        if (size >= UINT_MAX
-            || !(argv1 = wasm_runtime_malloc((uint32)size))) {
-            wasm_runtime_set_exception(exec_env->module_inst,
-                                       "allocate memory failed.");
+        if (!(argv1 = runtime_malloc((uint32)size, exec_env->module_inst,
+                                     NULL, 0))) {
             return false;
         }
     }
@@ -2543,10 +2551,8 @@ wasm_runtime_invoke_native(WASMExecEnv *exec_env, void *func_ptr,
     argc1 = 1 + MAX_REG_FLOATS + func_type->param_count + 2;
     if (argc1 > sizeof(argv_buf) / sizeof(uint64)) {
         size = sizeof(uint64) * (uint64)argc1;
-        if (size >= UINT32_MAX
-            || !(argv1 = wasm_runtime_malloc((uint32)size))) {
-            wasm_runtime_set_exception(exec_env->module_inst,
-                                       "allocate memory failed.");
+        if (!(argv1 = runtime_malloc((uint32)size, exec_env->module_inst,
+                                     NULL, 0))) {
             return false;
         }
     }

+ 7 - 1
core/iwasm/interpreter/iwasm_interp.cmake

@@ -13,8 +13,14 @@ else ()
     set (INTERPRETER     "wasm_interp_classic.c")
 endif ()
 
+if (WAMR_BUILD_MINI_LOADER EQUAL 1)
+    set (LOADER          "wasm_mini_loader.c")
+else ()
+    set (LOADER          "wasm_loader.c")
+endif ()
+
 file (GLOB_RECURSE source_all
-    ${IWASM_INTERP_DIR}/wasm_loader.c
+    ${IWASM_INTERP_DIR}/${LOADER}
     ${IWASM_INTERP_DIR}/wasm_runtime.c
     ${IWASM_INTERP_DIR}/${INTERPRETER}
 )

+ 107 - 155
core/iwasm/interpreter/wasm_loader.c

@@ -24,21 +24,43 @@ set_error_buf(char *error_buf, uint32 error_buf_size, const char *string)
         snprintf(error_buf, error_buf_size, "%s", string);
 }
 
-#define CHECK_BUF(buf, buf_end, length) do {                \
-  if (buf + length > buf_end) {                             \
-    set_error_buf(error_buf, error_buf_size,                \
-                  "WASM module load failed: "               \
-                  "unexpected end of section or function"); \
-    return false;                                           \
-  }                                                         \
+static bool
+check_buf(const uint8 *buf, const uint8 *buf_end, uint32 length,
+          char *error_buf, uint32 error_buf_size)
+{
+    if (buf + length > buf_end) {
+        set_error_buf(error_buf, error_buf_size,
+                      "WASM module load failed: "
+                      "unexpected end of section or function");
+        return false;
+    }
+    return true;
+}
+
+static bool
+check_buf1(const uint8 *buf, const uint8 *buf_end, uint32 length,
+           char *error_buf, uint32 error_buf_size)
+{
+    if (buf + length > buf_end) {
+        set_error_buf(error_buf, error_buf_size,
+                      "WASM module load failed: unexpected end");
+        return false;
+    }
+    return true;
+}
+
+#define CHECK_BUF(buf, buf_end, length) do {    \
+  if (!check_buf(buf, buf_end, length,          \
+                 error_buf, error_buf_size)) {  \
+      return false;                             \
+  }                                             \
 } while (0)
 
-#define CHECK_BUF1(buf, buf_end, length) do {               \
-  if (buf + length > buf_end) {                             \
-    set_error_buf(error_buf, error_buf_size,                \
-                  "WASM module load failed: unexpected end");\
-    return false;                                           \
-  }                                                         \
+#define CHECK_BUF1(buf, buf_end, length) do {   \
+  if (!check_buf1(buf, buf_end, length,         \
+                  error_buf, error_buf_size)) { \
+      return false;                             \
+  }                                             \
 } while (0)
 
 static bool
@@ -193,6 +215,23 @@ fail_integer_too_large:
   res = (int32)res64;                               \
 } while (0)
 
+static void *
+loader_malloc(uint64 size, char *error_buf, uint32 error_buf_size)
+{
+    void *mem;
+
+    if (size >= UINT32_MAX
+        || !(mem = wasm_runtime_malloc((uint32)size))) {
+        set_error_buf(error_buf, error_buf_size,
+                      "WASM module load failed: "
+                      "allocate memory failed.");
+        return NULL;
+    }
+
+    memset(mem, 0, (uint32)size);
+    return mem;
+}
+
 static bool
 check_utf8_str(const uint8* str, uint32 len)
 {
@@ -256,10 +295,8 @@ const_str_list_insert(const uint8 *str, uint32 len, WASMModule *module,
         return node->str;
     }
 
-    if (!(node = wasm_runtime_malloc(sizeof(StringNode) + len + 1))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "WASM module load failed: "
-                      "allocate memory failed.");
+    if (!(node = loader_malloc(sizeof(StringNode) + len + 1,
+                               error_buf, error_buf_size))) {
         return NULL;
     }
 
@@ -361,15 +398,11 @@ load_type_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
     if (type_count) {
         module->type_count = type_count;
         total_size = sizeof(WASMType*) * (uint64)type_count;
-        if (total_size >= UINT32_MAX
-            || !(module->types = wasm_runtime_malloc((uint32)total_size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "Load type section failed: allocate memory failed.");
+        if (!(module->types = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
             return false;
         }
 
-        memset(module->types, 0, (uint32)total_size);
-
         for (i = 0; i < type_count; i++) {
             CHECK_BUF(p, p_end, 1);
             flag = read_uint8(p);
@@ -396,11 +429,8 @@ load_type_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
 
             total_size = offsetof(WASMType, types) +
                          sizeof(uint8) * (uint64)(param_count + result_count);
-            if (total_size >= UINT32_MAX
-                || !(type = module->types[i] =
-                                wasm_runtime_malloc((uint32)total_size))) {
-                set_error_buf(error_buf, error_buf_size,
-                              "Load type section failed: allocate memory failed.");
+            if (!(type = module->types[i] =
+                        loader_malloc(total_size, error_buf, error_buf_size))) {
                 return false;
             }
 
@@ -808,12 +838,6 @@ load_table_import(WASMModule *sub_module, const char *sub_module_name,
     }
     *p_buf = p;
 
-    if ((declare_max_size_flag & 1) && declare_init_size > declare_max_size) {
-        set_error_buf(error_buf, error_buf_size,
-                      "size minimum must not be greater than maximum");
-        return false;
-    }
-
 #if WASM_ENABLE_MULTI_MODULE != 0
     if (!wasm_runtime_is_built_in_module(sub_module_name)) {
         linked_table = wasm_loader_resolve_table(
@@ -1093,12 +1117,6 @@ load_table(const uint8 **p_buf, const uint8 *buf_end, WASMTable *table,
     else
         table->max_size = 0x10000;
 
-    if ((table->flags & 1) && table->init_size > table->max_size) {
-        set_error_buf(error_buf, error_buf_size,
-                      "size minimum must not be greater than maximum");
-        return false;
-    }
-
     *p_buf = p;
     return true;
 }
@@ -1327,15 +1345,11 @@ load_import_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
     if (import_count) {
         module->import_count = import_count;
         total_size = sizeof(WASMImport) * (uint64)import_count;
-        if (total_size >= UINT32_MAX
-            || !(module->imports = wasm_runtime_malloc((uint32)total_size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "Load import section failed: allocate memory failed.");
+        if (!(module->imports = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
             return false;
         }
 
-        memset(module->imports, 0, (uint32)total_size);
-
         p_old = p;
 
         /* Scan firstly to get import count of each type */
@@ -1576,10 +1590,8 @@ init_function_local_offsets(WASMFunction *func,
     uint32 i, local_offset = 0;
     uint64 total_size = sizeof(uint16) * ((uint64)param_count + local_count);
 
-    if (total_size >= UINT32_MAX
-        || !(func->local_offsets = wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "Load function section failed: allocate memory failed.");
+    if (!(func->local_offsets =
+                loader_malloc(total_size, error_buf, error_buf_size))) {
         return false;
     }
 
@@ -1627,15 +1639,11 @@ load_function_section(const uint8 *buf, const uint8 *buf_end,
     if (func_count) {
         module->function_count = func_count;
         total_size = sizeof(WASMFunction*) * (uint64)func_count;
-        if (total_size >= UINT32_MAX
-            || !(module->functions = wasm_runtime_malloc((uint32)total_size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "Load function section failed: allocate memory failed.");
+        if (!(module->functions =
+                    loader_malloc(total_size, error_buf, error_buf_size))) {
             return false;
         }
 
-        memset(module->functions, 0, (uint32)total_size);
-
         for (i = 0; i < func_count; i++) {
             /* Resolve function type */
             read_leb_uint32(p, p_end, type_index);
@@ -1680,17 +1688,12 @@ load_function_section(const uint8 *buf, const uint8 *buf_end,
             code_size = (uint32)(p_code_end - p_code);
 
             total_size = sizeof(WASMFunction) + (uint64)local_count;
-            if (total_size >= UINT32_MAX
-                || !(func = module->functions[i] =
-                            wasm_runtime_malloc((uint32)total_size))) {
-                set_error_buf(error_buf, error_buf_size,
-                              "Load function section failed: "
-                              "allocate memory failed.");
+            if (!(func = module->functions[i] =
+                        loader_malloc(total_size, error_buf, error_buf_size))) {
                 return false;
             }
 
             /* Set function type, local count, code size and code body */
-            memset(func, 0, (uint32)total_size);
             func->func_type = module->types[type_index];
             func->local_count = local_count;
             if (local_count > 0)
@@ -1775,15 +1778,11 @@ load_table_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
     if (table_count) {
         module->table_count = table_count;
         total_size = sizeof(WASMTable) * (uint64)table_count;
-        if (total_size >= UINT32_MAX
-            || !(module->tables = wasm_runtime_malloc((uint32)total_size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "Load table section failed: allocate memory failed.");
+        if (!(module->tables = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
             return false;
         }
 
-        memset(module->tables, 0, (uint32)total_size);
-
         /* load each table */
         table = module->tables;
         for (i = 0; i < table_count; i++, table++)
@@ -1820,15 +1819,11 @@ load_memory_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
     if (memory_count) {
         module->memory_count = memory_count;
         total_size = sizeof(WASMMemory) * (uint64)memory_count;
-        if (total_size >= UINT32_MAX
-            || !(module->memories = wasm_runtime_malloc((uint32)total_size))) {
-            set_error_buf(error_buf, error_buf_size,
-                         "Load memory section failed: allocate memory failed.");
+        if (!(module->memories = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
             return false;
         }
 
-        memset(module->memories, 0, (uint32)total_size);
-
         /* load each memory */
         memory = module->memories;
         for (i = 0; i < memory_count; i++, memory++)
@@ -1861,16 +1856,11 @@ load_global_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
     if (global_count) {
         module->global_count = global_count;
         total_size = sizeof(WASMGlobal) * (uint64)global_count;
-        if (total_size >= UINT32_MAX
-            || !(module->globals = wasm_runtime_malloc((uint32)total_size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "Load global section failed: "
-                          "allocate memory failed.");
+        if (!(module->globals = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
             return false;
         }
 
-        memset(module->globals, 0, (uint32)total_size);
-
         global = module->globals;
 
         for(i = 0; i < global_count; i++, global++) {
@@ -1932,16 +1922,11 @@ load_export_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
     if (export_count) {
         module->export_count = export_count;
         total_size = sizeof(WASMExport) * (uint64)export_count;
-        if (total_size >= UINT32_MAX
-            || !(module->exports = wasm_runtime_malloc((uint32)total_size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "Load export section failed: "
-                          "allocate memory failed.");
+        if (!(module->exports = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
             return false;
         }
 
-        memset(module->exports, 0, (uint32)total_size);
-
         export = module->exports;
         for (i = 0; i < export_count; i++, export++) {
             read_leb_uint32(p, p_end, str_len);
@@ -2038,16 +2023,11 @@ load_table_segment_section(const uint8 *buf, const uint8 *buf_end, WASMModule *m
     if (table_segment_count) {
         module->table_seg_count = table_segment_count;
         total_size = sizeof(WASMTableSeg) * (uint64)table_segment_count;
-        if (total_size >= UINT32_MAX
-            || !(module->table_segments = wasm_runtime_malloc((uint32)total_size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "Load table segment section failed: "
-                          "allocate memory failed.");
+        if (!(module->table_segments = loader_malloc
+                (total_size, error_buf, error_buf_size))) {
             return false;
         }
 
-        memset(module->table_segments, 0, (uint32)total_size);
-
         table_segment = module->table_segments;
         for (i = 0; i < table_segment_count; i++, table_segment++) {
             if (p >= p_end) {
@@ -2074,12 +2054,8 @@ load_table_segment_section(const uint8 *buf, const uint8 *buf_end, WASMModule *m
             read_leb_uint32(p, p_end, function_count);
             table_segment->function_count = function_count;
             total_size = sizeof(uint32) * (uint64)function_count;
-            if (total_size >= UINT32_MAX
-                || !(table_segment->func_indexes = (uint32 *)
-                        wasm_runtime_malloc((uint32)total_size))) {
-                set_error_buf(error_buf, error_buf_size,
-                              "Load table segment section failed: "
-                              "allocate memory failed.");
+            if (!(table_segment->func_indexes = (uint32 *)
+                    loader_malloc(total_size, error_buf, error_buf_size))) {
                 return false;
             }
             for (j = 0; j < function_count; j++) {
@@ -2134,16 +2110,11 @@ load_data_segment_section(const uint8 *buf, const uint8 *buf_end,
     if (data_seg_count) {
         module->data_seg_count = data_seg_count;
         total_size = sizeof(WASMDataSeg*) * (uint64)data_seg_count;
-        if (total_size >= UINT32_MAX
-            || !(module->data_segments = wasm_runtime_malloc((uint32)total_size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "Load data segment section failed: "
-                          "allocate memory failed.");
+        if (!(module->data_segments = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
             return false;
         }
 
-        memset(module->data_segments, 0, (uint32)total_size);
-
         for (i = 0; i < data_seg_count; i++) {
             read_leb_uint32(p, p_end, mem_index);
 #if WASM_ENABLE_BULK_MEMORY != 0
@@ -2157,7 +2128,6 @@ load_data_segment_section(const uint8 *buf, const uint8 *buf_end,
                     /* no memory index, treat index as 0 */
                     mem_index = 0;
                     goto check_mem_index;
-                    break;
                 case 0x02:
                     /* read following memory index */
                     read_leb_uint32(p, p_end, mem_index);
@@ -2193,11 +2163,8 @@ check_mem_index:
 
             read_leb_uint32(p, p_end, data_seg_len);
 
-            if (!(dataseg = module->data_segments[i] =
-                        wasm_runtime_malloc((uint32)sizeof(WASMDataSeg)))) {
-                set_error_buf(error_buf, error_buf_size,
-                              "Load data segment section failed: "
-                              "allocate memory failed.");
+            if (!(dataseg = module->data_segments[i] = loader_malloc
+                        (sizeof(WASMDataSeg), error_buf, error_buf_size))) {
                 return false;
             }
 
@@ -2477,18 +2444,19 @@ load_from_sections(WASMModule *module, WASMSection *sections,
 #endif
 
     total_size = sizeof(BlockAddr) * (uint64)BLOCK_ADDR_CACHE_SIZE * BLOCK_ADDR_CONFLICT_SIZE;
-    if (total_size >= UINT32_MAX
-        || !(block_addr_cache = wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "WASM module load failed: allocate memory failed");
+    if (!(block_addr_cache = loader_malloc
+                (total_size, error_buf, error_buf_size))) {
         return false;
     }
 
     for (i = 0; i < module->function_count; i++) {
         WASMFunction *func = module->functions[i];
         memset(block_addr_cache, 0, (uint32)total_size);
-        if (!wasm_loader_prepare_bytecode(module, func, block_addr_cache, error_buf, error_buf_size))
+        if (!wasm_loader_prepare_bytecode(module, func, block_addr_cache,
+                                          error_buf, error_buf_size)) {
+            wasm_runtime_free(block_addr_cache);
             return false;
+        }
     }
     wasm_runtime_free(block_addr_cache);
 
@@ -2616,17 +2584,13 @@ static void wasm_loader_free(void *ptr)
 static WASMModule*
 create_module(char *error_buf, uint32 error_buf_size)
 {
-    WASMModule *module = wasm_runtime_malloc(sizeof(WASMModule));
+    WASMModule *module = loader_malloc(sizeof(WASMModule),
+                                       error_buf, error_buf_size);
 
     if (!module) {
-        set_error_buf(error_buf, error_buf_size,
-                      "WASM module load failed: "
-                      "allocate memory failed.");
         return NULL;
     }
 
-    memset(module, 0, sizeof(WASMModule));
-
     module->module_type = Wasm_Module_Bytecode;
 
     /* Set start_function to -1, means no start function */
@@ -2732,14 +2696,11 @@ create_sections(const uint8 *buf, uint32 size,
             read_leb_uint32(p, p_end, section_size);
             CHECK_BUF1(p, p_end, section_size);
 
-            if (!(section = wasm_runtime_malloc(sizeof(WASMSection)))) {
-                set_error_buf(error_buf, error_buf_size,
-                              "WASM module load failed: "
-                              "allocate memory failed.");
+            if (!(section = loader_malloc(sizeof(WASMSection),
+                                          error_buf, error_buf_size))) {
                 return false;
             }
 
-            memset(section, 0, sizeof(WASMSection));
             section->section_type = section_type;
             section->section_body = (uint8*)p;
             section->section_body_size = section_size;
@@ -3416,16 +3377,12 @@ memory_realloc(void *mem_old, uint32 size_old, uint32 size_new,
 {
     uint8 *mem_new;
     bh_assert(size_new > size_old);
-    if ((mem_new = wasm_runtime_malloc(size_new))) {
+    if ((mem_new = loader_malloc
+                (size_new, error_buf, error_buf_size))) {
         bh_memcpy_s(mem_new, size_new, mem_old, size_old);
         memset(mem_new + size_old, 0, size_new - size_old);
         wasm_runtime_free(mem_old);
     }
-    else {
-        set_error_buf(error_buf, error_buf_size,
-                    "WASM loader prepare bytecode failed: "
-                    "allocate memory failed.");
-    }
     return mem_new;
 }
 
@@ -3485,7 +3442,8 @@ check_offset_pop(WASMLoaderContext *ctx, uint32 cells)
     return true;
 }
 
-static void free_label_patch_list(BranchBlock *frame_csp)
+static void
+free_label_patch_list(BranchBlock *frame_csp)
 {
     BranchBlockPatch *label_patch = frame_csp->patch_list;
     BranchBlockPatch *next;
@@ -3497,7 +3455,8 @@ static void free_label_patch_list(BranchBlock *frame_csp)
     frame_csp->patch_list = NULL;
 }
 
-static void free_all_label_patch_lists(BranchBlock *frame_csp, uint32 csp_num)
+static void
+free_all_label_patch_lists(BranchBlock *frame_csp, uint32 csp_num)
 {
     BranchBlock *tmp_csp = frame_csp;
 
@@ -3525,7 +3484,6 @@ fail:
     return false;
 }
 
-
 static bool
 check_stack_top_values(uint8 *frame_ref, int32 stack_cell_num, uint8 type,
                        char *error_buf, uint32 error_buf_size)
@@ -3580,7 +3538,8 @@ check_stack_pop(WASMLoaderContext *ctx, uint8 type,
     return true;
 }
 
-static void wasm_loader_ctx_destroy(WASMLoaderContext *ctx)
+static void
+wasm_loader_ctx_destroy(WASMLoaderContext *ctx)
 {
     if (ctx) {
         if (ctx->frame_ref_bottom)
@@ -4055,11 +4014,9 @@ add_label_patch_to_list(BranchBlock *frame_csp,
                         uint8 patch_type, uint8 *p_code_compiled,
                         char *error_buf, uint32 error_buf_size)
 {
-    BranchBlockPatch *patch = wasm_runtime_malloc(sizeof(BranchBlockPatch));
+    BranchBlockPatch *patch = loader_malloc
+        (sizeof(BranchBlockPatch), error_buf, error_buf_size);
     if (!patch) {
-        set_error_buf(error_buf, error_buf_size,
-                      "WASM loader prepare bytecode failed: "
-                      "allocate memory failed");
         return false;
     }
     patch->patch_type = patch_type;
@@ -5505,7 +5462,7 @@ re_scan:
 
             case WASM_OP_SET_GLOBAL:
             {
-                bool is_multable = false;
+                bool is_mutable = false;
                 read_leb_uint32(p, p_end, global_idx);
                 if (global_idx >= global_count) {
                     set_error_buf(error_buf, error_buf_size,
@@ -5514,12 +5471,12 @@ re_scan:
                     goto fail;
                 }
 
-                is_multable =
+                is_mutable =
                   global_idx < module->import_global_count
                     ? module->import_globals[global_idx].u.global.is_mutable
                     : module->globals[global_idx - module->import_global_count]
                         .is_mutable;
-                if (!is_multable) {
+                if (!is_mutable) {
                     set_error_buf(error_buf,
                                   error_buf_size,
                                   "global is immutable");
@@ -5772,8 +5729,6 @@ re_scan:
                 POP2_AND_PUSH(VALUE_TYPE_F64, VALUE_TYPE_I32);
                 break;
 
-                break;
-
             case WASM_OP_I32_CLZ:
             case WASM_OP_I32_CTZ:
             case WASM_OP_I32_POPCNT:
@@ -6092,13 +6047,10 @@ fail_data_cnt_sec_require:
 
     func->const_cell_num = loader_ctx->const_cell_num;
     if (!(func->consts = func_const =
-                wasm_runtime_malloc(func->const_cell_num * 4))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "WASM loader prepare bytecode failed: "
-                      "allocate memory failed");
+                loader_malloc(func->const_cell_num * 4,
+                              error_buf, error_buf_size))) {
         goto fail;
     }
-    memset(func->consts, 0, func->const_cell_num * 4);
     func_const_end = func->consts + func->const_cell_num * 4;
     // reverse the const buf
     for (int i = loader_ctx->num_const - 1; i >= 0; i--) {

+ 4947 - 0
core/iwasm/interpreter/wasm_mini_loader.c

@@ -0,0 +1,4947 @@
+/*
+ * Copyright (C) 2019 Intel Corporation.  All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+ */
+
+#include "wasm_loader.h"
+#include "bh_common.h"
+#include "bh_log.h"
+#include "wasm.h"
+#include "wasm_opcode.h"
+#include "wasm_runtime.h"
+#include "../common/wasm_native.h"
+
+/* Read a value of given type from the address pointed to by the given
+   pointer and increase the pointer to the position just after the
+   value being read.  */
+#define TEMPLATE_READ_VALUE(Type, p)                    \
+    (p += sizeof(Type), *(Type *)(p - sizeof(Type)))
+
+static void
+set_error_buf(char *error_buf, uint32 error_buf_size, const char *string)
+{
+    if (error_buf != NULL)
+        snprintf(error_buf, error_buf_size, "%s", string);
+}
+
+#define CHECK_BUF(buf, buf_end, length) do {                \
+    bh_assert(buf + length <= buf_end);                     \
+  } while (0)
+
+#define CHECK_BUF1(buf, buf_end, length) do {               \
+    bh_assert(buf + length <= buf_end);                     \
+  } while (0)
+
+static void
+skip_leb(const uint8 **p_buf, const uint8 *buf_end, uint32 maxbits,
+         char* error_buf, uint32 error_buf_size)
+{
+    const uint8 *buf = *p_buf;
+    uint32 offset = 0, bcnt = 0;
+    uint64 byte;
+
+    while (true) {
+        bh_assert(bcnt + 1 <= (maxbits + 6) / 7);
+        CHECK_BUF(buf, buf_end, offset + 1);
+        byte = buf[offset];
+        offset += 1;
+        bcnt += 1;
+        if ((byte & 0x80) == 0) {
+            break;
+        }
+    }
+
+    *p_buf += offset;
+}
+
+#define skip_leb_int64(p, p_end) do {               \
+    skip_leb(&p, p_end, 64,                         \
+             error_buf, error_buf_size);            \
+  } while (0)
+
+#define skip_leb_uint32(p, p_end) do {              \
+    skip_leb(&p, p_end, 32,                         \
+             error_buf, error_buf_size);            \
+  } while (0)
+
+#define skip_leb_int32(p, p_end) do {               \
+    skip_leb(&p, p_end, 32,                         \
+             error_buf, error_buf_size);            \
+  } while (0)
+
+static void
+read_leb(uint8 **p_buf, const uint8 *buf_end,
+         uint32 maxbits, bool sign, uint64 *p_result,
+         char* error_buf, uint32 error_buf_size)
+{
+    const uint8 *buf = *p_buf;
+    uint64 result = 0;
+    uint32 shift = 0;
+    uint32 offset = 0, bcnt = 0;
+    uint64 byte;
+
+    while (true) {
+        bh_assert(bcnt + 1 <= (maxbits + 6) / 7);
+        CHECK_BUF(buf, buf_end, offset + 1);
+        byte = buf[offset];
+        offset += 1;
+        result |= ((byte & 0x7f) << shift);
+        shift += 7;
+        bcnt += 1;
+        if ((byte & 0x80) == 0) {
+            break;
+        }
+    }
+
+    if (!sign && maxbits == 32 && shift >= maxbits) {
+        /* The top bits set represent values > 32 bits */
+        bh_assert(!(((uint8)byte) & 0xf0));
+    }
+    else if (sign && maxbits == 32) {
+        if (shift < maxbits) {
+            /* Sign extend */
+            result = (((int32)result) << (maxbits - shift))
+                     >> (maxbits - shift);
+        }
+        else {
+            /* The top bits should be a sign-extension of the sign bit */
+            bool sign_bit_set = ((uint8)byte) & 0x8;
+            int top_bits = ((uint8)byte) & 0xf0;
+            bh_assert(!((sign_bit_set && top_bits != 0x70)
+                        || (!sign_bit_set && top_bits != 0)));
+            (void)top_bits;
+            (void)sign_bit_set;
+        }
+    }
+    else if (sign && maxbits == 64) {
+        if (shift < maxbits) {
+            /* Sign extend */
+            result = (((int64)result) << (maxbits - shift))
+                     >> (maxbits - shift);
+        }
+        else {
+            /* The top bits should be a sign-extension of the sign bit */
+            bool sign_bit_set = ((uint8)byte) & 0x1;
+            int top_bits = ((uint8)byte) & 0xfe;
+
+            bh_assert(!((sign_bit_set && top_bits != 0x7e)
+                        || (!sign_bit_set && top_bits != 0)));
+            (void)top_bits;
+            (void)sign_bit_set;
+        }
+    }
+
+    *p_buf += offset;
+    *p_result = result;
+}
+
+#define read_uint8(p)  TEMPLATE_READ_VALUE(uint8, p)
+#define read_uint32(p) TEMPLATE_READ_VALUE(uint32, p)
+#define read_bool(p)   TEMPLATE_READ_VALUE(bool, p)
+
+#define read_leb_int64(p, p_end, res) do {          \
+  uint64 res64;                                     \
+  read_leb((uint8**)&p, p_end, 64, true, &res64,    \
+                error_buf, error_buf_size);         \
+  res = (int64)res64;                               \
+} while (0)
+
+#define read_leb_uint32(p, p_end, res) do {         \
+  uint64 res64;                                     \
+  read_leb((uint8**)&p, p_end, 32, false, &res64,   \
+                error_buf, error_buf_size);         \
+  res = (uint32)res64;                              \
+} while (0)
+
+#define read_leb_int32(p, p_end, res) do {          \
+  uint64 res64;                                     \
+  read_leb((uint8**)&p, p_end, 32, true, &res64,    \
+                error_buf, error_buf_size);         \
+  res = (int32)res64;                               \
+} while (0)
+
+static void *
+loader_malloc(uint64 size, char *error_buf, uint32 error_buf_size)
+{
+    void *mem;
+
+    if (size >= UINT32_MAX
+        || !(mem = wasm_runtime_malloc((uint32)size))) {
+        set_error_buf(error_buf, error_buf_size,
+                      "WASM module load failed: "
+                      "allocate memory failed.");
+        return NULL;
+    }
+
+    memset(mem, 0, (uint32)size);
+    return mem;
+}
+
+static char *
+const_str_list_insert(const uint8 *str, uint32 len, WASMModule *module,
+                     char *error_buf, uint32 error_buf_size)
+{
+    StringNode *node, *node_next;
+
+    /* Search const str list */
+    node = module->const_str_list;
+    while (node) {
+        node_next = node->next;
+        if (strlen(node->str) == len
+            && !memcmp(node->str, str, len))
+            break;
+        node = node_next;
+    }
+
+    if (node) {
+        LOG_DEBUG("reuse %s", node->str);
+        return node->str;
+    }
+
+    if (!(node = loader_malloc(sizeof(StringNode) + len + 1,
+                               error_buf, error_buf_size))) {
+        return NULL;
+    }
+
+    node->str = ((char*)node) + sizeof(StringNode);
+    bh_memcpy_s(node->str, len + 1, str, len);
+    node->str[len] = '\0';
+
+    if (!module->const_str_list) {
+        /* set as head */
+        module->const_str_list = node;
+        node->next = NULL;
+    }
+    else {
+        /* insert it */
+        node->next = module->const_str_list;
+        module->const_str_list = node;
+    }
+
+    return node->str;
+}
+
+static bool
+load_init_expr(const uint8 **p_buf, const uint8 *buf_end,
+               InitializerExpression *init_expr, uint8 type,
+               char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = *p_buf, *p_end = buf_end;
+    uint8 flag, end_byte, *p_float;
+    uint32 i;
+
+    CHECK_BUF(p, p_end, 1);
+    init_expr->init_expr_type = read_uint8(p);
+    flag = init_expr->init_expr_type;
+
+    switch (flag) {
+        /* i32.const */
+        case INIT_EXPR_TYPE_I32_CONST:
+            bh_assert(type == VALUE_TYPE_I32);
+            read_leb_int32(p, p_end, init_expr->u.i32);
+            break;
+        /* i64.const */
+        case INIT_EXPR_TYPE_I64_CONST:
+            bh_assert(type == VALUE_TYPE_I64);
+            read_leb_int64(p, p_end, init_expr->u.i64);
+            break;
+        /* f32.const */
+        case INIT_EXPR_TYPE_F32_CONST:
+            bh_assert(type == VALUE_TYPE_F32);
+            CHECK_BUF(p, p_end, 4);
+            p_float = (uint8*)&init_expr->u.f32;
+            for (i = 0; i < sizeof(float32); i++)
+                *p_float++ = *p++;
+            break;
+        /* f64.const */
+        case INIT_EXPR_TYPE_F64_CONST:
+            bh_assert(type == VALUE_TYPE_F64);
+            CHECK_BUF(p, p_end, 8);
+            p_float = (uint8*)&init_expr->u.f64;
+            for (i = 0; i < sizeof(float64); i++)
+                *p_float++ = *p++;
+            break;
+        /* get_global */
+        case INIT_EXPR_TYPE_GET_GLOBAL:
+            read_leb_uint32(p, p_end, init_expr->u.global_index);
+            break;
+        default:
+            bh_assert(0);
+            break;
+    }
+    CHECK_BUF(p, p_end, 1);
+    end_byte = read_uint8(p);
+    bh_assert(end_byte == 0x0b);
+    *p_buf = p;
+
+    (void)end_byte;
+    return true;
+}
+
+static bool
+load_type_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
+                  char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end, *p_org;
+    uint32 type_count, param_count, result_count, i, j;
+    uint64 total_size;
+    uint8 flag;
+    WASMType *type;
+
+    read_leb_uint32(p, p_end, type_count);
+
+    if (type_count) {
+        module->type_count = type_count;
+        total_size = sizeof(WASMType*) * (uint64)type_count;
+        if (!(module->types = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
+            return false;
+        }
+
+        for (i = 0; i < type_count; i++) {
+            CHECK_BUF(p, p_end, 1);
+            flag = read_uint8(p);
+            bh_assert(flag == 0x60);
+
+            read_leb_uint32(p, p_end, param_count);
+
+            /* Resolve param count and result count firstly */
+            p_org = p;
+            CHECK_BUF(p, p_end, param_count);
+            p += param_count;
+            read_leb_uint32(p, p_end, result_count);
+            bh_assert(result_count <= 1);
+            CHECK_BUF(p, p_end, result_count);
+            p = p_org;
+
+            total_size = offsetof(WASMType, types) +
+                         sizeof(uint8) * (uint64)(param_count + result_count);
+            if (!(type = module->types[i] =
+                        loader_malloc(total_size, error_buf, error_buf_size))) {
+                return false;
+            }
+
+            /* Resolve param types and result types */
+            type->param_count = param_count;
+            type->result_count = result_count;
+            for (j = 0; j < param_count; j++) {
+                CHECK_BUF(p, p_end, 1);
+                type->types[j] = read_uint8(p);
+            }
+            read_leb_uint32(p, p_end, result_count);
+            for (j = 0; j < result_count; j++) {
+                CHECK_BUF(p, p_end, 1);
+                type->types[param_count + j] = read_uint8(p);
+            }
+        }
+    }
+
+    bh_assert(p == p_end);
+    LOG_VERBOSE("Load type section success.\n");
+    (void)flag;
+    return true;
+}
+
+static bool
+load_function_import(const WASMModule *parent_module, WASMModule *sub_module,
+                     char *sub_module_name, char *function_name,
+                     const uint8 **p_buf, const uint8 *buf_end,
+                     WASMFunctionImport *function,
+                     char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = *p_buf, *p_end = buf_end;
+    uint32 declare_type_index = 0;
+    WASMType *declare_func_type = NULL;
+    WASMFunction *linked_func = NULL;
+    const char *linked_signature = NULL;
+    void *linked_attachment = NULL;
+    bool linked_call_conv_raw = false;
+    bool is_built_in_module = false;
+
+    CHECK_BUF(p, p_end, 1);
+    read_leb_uint32(p, p_end, declare_type_index);
+    *p_buf = p;
+
+    bh_assert(declare_type_index < parent_module->type_count);
+
+    declare_func_type = parent_module->types[declare_type_index];
+
+    is_built_in_module = wasm_runtime_is_built_in_module(sub_module_name);
+    if (is_built_in_module) {
+        LOG_DEBUG("%s is a function of a built-in module %s",
+                  function_name,
+                  sub_module_name);
+        /* check built-in modules */
+        linked_func = wasm_native_resolve_symbol(sub_module_name,
+                                                 function_name,
+                                                 declare_func_type,
+                                                 &linked_signature,
+                                                 &linked_attachment,
+                                                 &linked_call_conv_raw);
+    }
+
+    if (!linked_func) {
+#if WASM_ENABLE_SPEC_TEST != 0
+        set_error_buf(error_buf,
+                      error_buf_size,
+                      "unknown import or incompatible import type");
+        return false;
+#else
+#if WASM_ENABLE_WAMR_COMPILER == 0
+        LOG_WARNING(
+          "warning: fail to link import function (%s, %s)",
+          sub_module_name, function_name);
+#endif
+#endif
+    }
+
+    function->module_name = sub_module_name;
+    function->field_name = function_name;
+    function->func_type = declare_func_type;
+    /* func_ptr_linked is for built-in functions */
+    function->func_ptr_linked = is_built_in_module ? linked_func : NULL;
+    function->signature = linked_signature;
+    function->attachment = linked_attachment;
+    function->call_conv_raw = linked_call_conv_raw;
+    return true;
+}
+
+static bool
+load_table_import(WASMModule *sub_module, const char *sub_module_name,
+                  const char *table_name, const uint8 **p_buf,
+                  const uint8 *buf_end, WASMTableImport *table,
+                  char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = *p_buf, *p_end = buf_end;
+    uint32 declare_elem_type = 0;
+    uint32 declare_max_size_flag = 0;
+    uint32 declare_init_size = 0;
+    uint32 declare_max_size = 0;
+
+    CHECK_BUF(p, p_end, 1);
+    /* 0x70 */
+    declare_elem_type = read_uint8(p);
+    bh_assert(TABLE_ELEM_TYPE_ANY_FUNC == declare_elem_type);
+
+    read_leb_uint32(p, p_end, declare_max_size_flag);
+    read_leb_uint32(p, p_end, declare_init_size);
+    if (declare_max_size_flag & 1) {
+        read_leb_uint32(p, p_end, declare_max_size);
+        bh_assert(table->init_size <= table->max_size);
+    } else {
+        declare_max_size = 0x10000;
+    }
+    *p_buf = p;
+
+    bh_assert(!((declare_max_size_flag & 1)
+                && declare_init_size > declare_max_size));
+
+    /* now we believe all declaration are ok */
+    table->elem_type = declare_elem_type;
+    table->init_size = declare_init_size;
+    table->flags = declare_max_size_flag;
+    table->max_size = declare_max_size;
+    return true;
+}
+
+unsigned
+wasm_runtime_memory_pool_size();
+
+static bool
+load_memory_import(WASMModule *sub_module, const char *sub_module_name,
+                   const char *memory_name, const uint8 **p_buf,
+                   const uint8 *buf_end, WASMMemoryImport *memory,
+                   char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = *p_buf, *p_end = buf_end;
+    uint32 pool_size = wasm_runtime_memory_pool_size();
+#if WASM_ENABLE_APP_FRAMEWORK != 0
+    uint32 max_page_count = pool_size * APP_MEMORY_MAX_GLOBAL_HEAP_PERCENT
+                            / DEFAULT_NUM_BYTES_PER_PAGE;
+#else
+    uint32 max_page_count = pool_size / DEFAULT_NUM_BYTES_PER_PAGE;
+#endif /* WASM_ENABLE_APP_FRAMEWORK */
+    uint32 declare_max_page_count_flag = 0;
+    uint32 declare_init_page_count = 0;
+    uint32 declare_max_page_count = 0;
+
+    read_leb_uint32(p, p_end, declare_max_page_count_flag);
+    read_leb_uint32(p, p_end, declare_init_page_count);
+    bh_assert(declare_init_page_count <= 65536);
+
+    if (declare_max_page_count_flag & 1) {
+        read_leb_uint32(p, p_end, declare_max_page_count);
+        bh_assert(declare_init_page_count <= declare_max_page_count);
+        bh_assert(declare_max_page_count <= 65536);
+        if (declare_max_page_count > max_page_count) {
+            declare_max_page_count = max_page_count;
+        }
+    }
+    else {
+        /* Limit the maximum memory size to max_page_count */
+        declare_max_page_count = max_page_count;
+    }
+
+    /* now we believe all declaration are ok */
+    memory->flags = declare_max_page_count_flag;
+    memory->init_page_count = declare_init_page_count;
+    memory->max_page_count = declare_max_page_count;
+    memory->num_bytes_per_page = DEFAULT_NUM_BYTES_PER_PAGE;
+
+    *p_buf = p;
+    return true;
+}
+
+static bool
+load_global_import(const WASMModule *parent_module,
+                   WASMModule *sub_module,
+                   char *sub_module_name, char *global_name,
+                   const uint8 **p_buf, const uint8 *buf_end,
+                   WASMGlobalImport *global,
+                   char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = *p_buf, *p_end = buf_end;
+    uint8 declare_type = 0;
+    uint8 declare_mutable = 0;
+    bool is_mutable = false;
+    bool ret = false;
+
+    CHECK_BUF(p, p_end, 2);
+    declare_type = read_uint8(p);
+    declare_mutable = read_uint8(p);
+    *p_buf = p;
+
+    bh_assert(declare_mutable < 2);
+
+    is_mutable = declare_mutable & 1 ? true : false;
+
+#if WASM_ENABLE_LIBC_BUILTIN != 0
+    ret = wasm_runtime_is_built_in_module(sub_module_name);
+    if (ret) {
+        /* check built-in modules */
+        ret = wasm_native_lookup_libc_builtin_global(sub_module_name,
+                                                     global_name, global);
+        if (ret) {
+            LOG_DEBUG("(%s, %s) is a global of a built-in module",
+                      sub_module_name, global_name);
+        }
+    }
+#endif /* WASM_ENABLE_LIBC_BUILTIN */
+
+    if (!ret) {
+        set_error_buf_v(error_buf, error_buf_size,
+                        "unknown import or incompatible import type");
+        return false;
+    }
+
+    global->module_name = sub_module_name;
+    global->field_name = global_name;
+    global->type = declare_type;
+    global->is_mutable = is_mutable;
+    (void)p_end;
+    return true;
+}
+
+static bool
+load_table(const uint8 **p_buf, const uint8 *buf_end, WASMTable *table,
+           char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = *p_buf, *p_end = buf_end;
+
+    CHECK_BUF(p, p_end, 1);
+    /* 0x70 */
+    table->elem_type = read_uint8(p);
+    bh_assert(TABLE_ELEM_TYPE_ANY_FUNC == table->elem_type);
+
+    read_leb_uint32(p, p_end, table->flags);
+    read_leb_uint32(p, p_end, table->init_size);
+    if (table->flags & 1) {
+        read_leb_uint32(p, p_end, table->max_size);
+        bh_assert(table->init_size <= table->max_size);
+    }
+    else
+        table->max_size = 0x10000;
+
+    bh_assert(!((table->flags & 1)
+                && table->init_size > table->max_size));
+
+    *p_buf = p;
+    return true;
+}
+
+static bool
+load_memory(const uint8 **p_buf, const uint8 *buf_end, WASMMemory *memory,
+            char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = *p_buf, *p_end = buf_end;
+    uint32 pool_size = wasm_runtime_memory_pool_size();
+#if WASM_ENABLE_APP_FRAMEWORK != 0
+    uint32 max_page_count = pool_size * APP_MEMORY_MAX_GLOBAL_HEAP_PERCENT
+                            / DEFAULT_NUM_BYTES_PER_PAGE;
+#else
+    uint32 max_page_count = pool_size / DEFAULT_NUM_BYTES_PER_PAGE;
+#endif
+
+    read_leb_uint32(p, p_end, memory->flags);
+    read_leb_uint32(p, p_end, memory->init_page_count);
+    bh_assert(memory->init_page_count <= 65536);
+
+    if (memory->flags & 1) {
+        read_leb_uint32(p, p_end, memory->max_page_count);
+        bh_assert(memory->init_page_count <= memory->max_page_count);
+        bh_assert(memory->max_page_count <= 65536);
+        if (memory->max_page_count > max_page_count)
+            memory->max_page_count = max_page_count;
+    }
+    else
+        /* Limit the maximum memory size to max_page_count */
+        memory->max_page_count = max_page_count;
+
+    memory->num_bytes_per_page = DEFAULT_NUM_BYTES_PER_PAGE;
+
+    *p_buf = p;
+    return true;
+}
+
+static bool
+load_import_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
+                    char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end, *p_old;
+    uint32 import_count, name_len, type_index, i, u32, flags;
+    uint64 total_size;
+    WASMImport *import;
+    WASMImport *import_functions = NULL, *import_tables = NULL;
+    WASMImport *import_memories = NULL, *import_globals = NULL;
+    char *sub_module_name, *field_name;
+    uint8 u8, kind;
+
+    read_leb_uint32(p, p_end, import_count);
+
+    if (import_count) {
+        module->import_count = import_count;
+        total_size = sizeof(WASMImport) * (uint64)import_count;
+        if (!(module->imports = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
+            return false;
+        }
+
+        p_old = p;
+
+        /* Scan firstly to get import count of each type */
+        for (i = 0; i < import_count; i++) {
+            /* module name */
+            read_leb_uint32(p, p_end, name_len);
+            CHECK_BUF(p, p_end, name_len);
+            p += name_len;
+
+            /* field name */
+            read_leb_uint32(p, p_end, name_len);
+            CHECK_BUF(p, p_end, name_len);
+            p += name_len;
+
+            CHECK_BUF(p, p_end, 1);
+            /* 0x00/0x01/0x02/0x03 */
+            kind = read_uint8(p);
+
+            switch (kind) {
+                case IMPORT_KIND_FUNC: /* import function */
+                    read_leb_uint32(p, p_end, type_index);
+                    module->import_function_count++;
+                    break;
+
+                case IMPORT_KIND_TABLE: /* import table */
+                    CHECK_BUF(p, p_end, 1);
+                    /* 0x70 */
+                    u8 = read_uint8(p);
+                    read_leb_uint32(p, p_end, flags);
+                    read_leb_uint32(p, p_end, u32);
+                    if (flags & 1)
+                        read_leb_uint32(p, p_end, u32);
+                    module->import_table_count++;
+                    bh_assert(module->import_table_count <= 1);
+                    break;
+
+                case IMPORT_KIND_MEMORY: /* import memory */
+                    read_leb_uint32(p, p_end, flags);
+                    read_leb_uint32(p, p_end, u32);
+                    if (flags & 1)
+                        read_leb_uint32(p, p_end, u32);
+                    module->import_memory_count++;
+                    bh_assert(module->import_memory_count <= 1);
+                    break;
+
+                case IMPORT_KIND_GLOBAL: /* import global */
+                    CHECK_BUF(p, p_end, 2);
+                    p += 2;
+                    module->import_global_count++;
+                    break;
+
+                default:
+                    bh_assert(0);
+                    break;
+            }
+        }
+
+        if (module->import_function_count)
+            import_functions = module->import_functions = module->imports;
+        if (module->import_table_count)
+            import_tables = module->import_tables =
+                module->imports + module->import_function_count;
+        if (module->import_memory_count)
+            import_memories = module->import_memories =
+                module->imports + module->import_function_count + module->import_table_count;
+        if (module->import_global_count)
+            import_globals = module->import_globals =
+                module->imports + module->import_function_count + module->import_table_count
+                + module->import_memory_count;
+
+        p = p_old;
+
+        // TODO: move it out of the loop
+        /* insert "env", "wasi_unstable" and "wasi_snapshot_preview1" to const str list */
+        if (!const_str_list_insert((uint8*)"env", 3, module, error_buf, error_buf_size)
+            || !const_str_list_insert((uint8*)"wasi_unstable", 13, module,
+                                     error_buf, error_buf_size)
+            || !const_str_list_insert((uint8*)"wasi_snapshot_preview1", 22, module,
+                                     error_buf, error_buf_size)) {
+            return false;
+        }
+
+        /* Scan again to read the data */
+        for (i = 0; i < import_count; i++) {
+            WASMModule *sub_module = NULL;
+
+            /* load module name */
+            read_leb_uint32(p, p_end, name_len);
+            CHECK_BUF(p, p_end, name_len);
+            if (!(sub_module_name = const_str_list_insert(
+                    p, name_len, module, error_buf, error_buf_size))) {
+                return false;
+            }
+            p += name_len;
+
+            /* load field name */
+            read_leb_uint32(p, p_end, name_len);
+            CHECK_BUF(p, p_end, name_len);
+            if (!(field_name = const_str_list_insert(
+                    p, name_len, module, error_buf, error_buf_size))) {
+                return false;
+            }
+            p += name_len;
+
+            LOG_DEBUG("import #%d: (%s, %s)", i, sub_module_name, field_name);
+
+            CHECK_BUF(p, p_end, 1);
+            /* 0x00/0x01/0x02/0x03 */
+            kind = read_uint8(p);
+            switch (kind) {
+                case IMPORT_KIND_FUNC: /* import function */
+                    bh_assert(import_functions);
+                    import = import_functions++;
+                    if (!load_function_import(module, sub_module,
+                                              sub_module_name, field_name, &p,
+                                              p_end, &import->u.function,
+                                              error_buf, error_buf_size)) {
+                        return false;
+                    }
+                    break;
+
+                case IMPORT_KIND_TABLE: /* import table */
+                    bh_assert(import_tables);
+                    import = import_tables++;
+                    if (!load_table_import(sub_module,
+                                           sub_module_name,
+                                           field_name,
+                                           &p,
+                                           p_end,
+                                           &import->u.table,
+                                           error_buf,
+                                           error_buf_size)) {
+                        LOG_DEBUG("can not import such a table (%s,%s)",
+                                  sub_module_name, field_name);
+                        return false;
+                    }
+                    break;
+
+                case IMPORT_KIND_MEMORY: /* import memory */
+                    bh_assert(import_memories);
+                    import = import_memories++;
+                    if (!load_memory_import(sub_module,
+                                            sub_module_name,
+                                            field_name,
+                                            &p,
+                                            p_end,
+                                            &import->u.memory,
+                                            error_buf,
+                                            error_buf_size)) {
+                        return false;
+                    }
+                    break;
+
+                case IMPORT_KIND_GLOBAL: /* import global */
+                    bh_assert(import_globals);
+                    import = import_globals++;
+                    if (!load_global_import(module,
+                                            sub_module,
+                                            sub_module_name, field_name, &p,
+                                            p_end, &import->u.global,
+                                            error_buf, error_buf_size)) {
+                        return false;
+                    }
+                    break;
+
+                default:
+                    bh_assert(0);
+                    import = NULL;
+                    break;
+            }
+            import->kind = kind;
+            import->u.names.module_name = sub_module_name;
+            import->u.names.field_name = field_name;
+            (void)sub_module;
+        }
+
+#if WASM_ENABLE_LIBC_WASI != 0
+        import = module->import_functions;
+        for (i = 0; i < module->import_function_count; i++, import++) {
+            if (!strcmp(import->u.names.module_name, "wasi_unstable")
+                || !strcmp(import->u.names.module_name, "wasi_snapshot_preview1")) {
+                module->is_wasi_module = true;
+                break;
+            }
+        }
+#endif
+    }
+
+    bh_assert(p == p_end);
+
+    LOG_VERBOSE("Load import section success.\n");
+    (void)u8;
+    (void)u32;
+    (void)type_index;
+    return true;
+}
+
+static bool
+init_function_local_offsets(WASMFunction *func,
+                            char *error_buf, uint32 error_buf_size)
+{
+    WASMType *param_type = func->func_type;
+    uint32 param_count = param_type->param_count;
+    uint8 *param_types = param_type->types;
+    uint32 local_count = func->local_count;
+    uint8 *local_types = func->local_types;
+    uint32 i, local_offset = 0;
+    uint64 total_size = sizeof(uint16) * ((uint64)param_count + local_count);
+
+    if (!(func->local_offsets =
+                loader_malloc(total_size, error_buf, error_buf_size))) {
+        return false;
+    }
+
+    for (i = 0; i < param_count; i++) {
+        func->local_offsets[i] = (uint16)local_offset;
+        local_offset += wasm_value_type_cell_num(param_types[i]);
+    }
+
+    for (i = 0; i < local_count; i++) {
+        func->local_offsets[param_count + i] = (uint16)local_offset;
+        local_offset += wasm_value_type_cell_num(local_types[i]);
+    }
+
+    bh_assert(local_offset == func->param_cell_num + func->local_cell_num);
+    return true;
+}
+
+static bool
+load_function_section(const uint8 *buf, const uint8 *buf_end,
+                      const uint8 *buf_code, const uint8 *buf_code_end,
+                      WASMModule *module,
+                      char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end;
+    const uint8 *p_code = buf_code, *p_code_end, *p_code_save;
+    uint32 func_count;
+    uint64 total_size;
+    uint32 code_count = 0, code_size, type_index, i, j, k, local_type_index;
+    uint32 local_count, local_set_count, sub_local_count;
+    uint8 type;
+    WASMFunction *func;
+
+    read_leb_uint32(p, p_end, func_count);
+
+    if (buf_code)
+        read_leb_uint32(p_code, buf_code_end, code_count);
+
+    bh_assert(func_count == code_count);
+
+    if (func_count) {
+        module->function_count = func_count;
+        total_size = sizeof(WASMFunction*) * (uint64)func_count;
+        if (!(module->functions =
+                    loader_malloc(total_size, error_buf, error_buf_size))) {
+            return false;
+        }
+
+        for (i = 0; i < func_count; i++) {
+            /* Resolve function type */
+            read_leb_uint32(p, p_end, type_index);
+            bh_assert(type_index < module->type_count);
+
+            read_leb_uint32(p_code, buf_code_end, code_size);
+            bh_assert(code_size > 0
+                      && p_code + code_size <= buf_code_end);
+
+            /* Resolve local set count */
+            p_code_end = p_code + code_size;
+            local_count = 0;
+            read_leb_uint32(p_code, buf_code_end, local_set_count);
+            p_code_save = p_code;
+
+            /* Calculate total local count */
+            for (j = 0; j < local_set_count; j++) {
+                read_leb_uint32(p_code, buf_code_end, sub_local_count);
+                bh_assert(sub_local_count <= UINT32_MAX - local_count);
+
+                CHECK_BUF(p_code, buf_code_end, 1);
+                /* 0x7F/0x7E/0x7D/0x7C */
+                type = read_uint8(p_code);
+                local_count += sub_local_count;
+            }
+
+            /* Alloc memory, layout: function structure + local types */
+            code_size = (uint32)(p_code_end - p_code);
+
+            total_size = sizeof(WASMFunction) + (uint64)local_count;
+            if (!(func = module->functions[i] =
+                        loader_malloc(total_size, error_buf, error_buf_size))) {
+                return false;
+            }
+
+            /* Set function type, local count, code size and code body */
+            func->func_type = module->types[type_index];
+            func->local_count = local_count;
+            if (local_count > 0)
+                func->local_types = (uint8*)func + sizeof(WASMFunction);
+            func->code_size = code_size;
+            /*
+             * we shall make a copy of code body [p_code, p_code + code_size]
+             * when we are worrying about inappropriate releasing behaviour.
+             * all code bodies are actually in a buffer which user allocates in
+             * his embedding environment and we don't have power on them.
+             * it will be like:
+             * code_body_cp = malloc(code_size);
+             * memcpy(code_body_cp, p_code, code_size);
+             * func->code = code_body_cp;
+             */
+            func->code = (uint8*)p_code;
+
+            /* Load each local type */
+            p_code = p_code_save;
+            local_type_index = 0;
+            for (j = 0; j < local_set_count; j++) {
+                read_leb_uint32(p_code, buf_code_end, sub_local_count);
+                bh_assert(!(local_type_index + sub_local_count <= local_type_index
+                            || local_type_index + sub_local_count > local_count));
+
+                CHECK_BUF(p_code, buf_code_end, 1);
+                /* 0x7F/0x7E/0x7D/0x7C */
+                type = read_uint8(p_code);
+                bh_assert(type >= VALUE_TYPE_F64 && type <= VALUE_TYPE_I32);
+                for (k = 0; k < sub_local_count; k++) {
+                    func->local_types[local_type_index++] = type;
+                }
+            }
+
+            func->param_cell_num = wasm_type_param_cell_num(func->func_type);
+            func->ret_cell_num = wasm_type_return_cell_num(func->func_type);
+            func->local_cell_num =
+                wasm_get_cell_num(func->local_types, func->local_count);
+
+            if (!init_function_local_offsets(func, error_buf, error_buf_size))
+                return false;
+
+            p_code = p_code_end;
+        }
+    }
+
+    bh_assert(p == p_end);
+    LOG_VERBOSE("Load function section success.\n");
+    (void)code_count;
+    return true;
+}
+
+static bool
+load_table_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
+                   char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end;
+    uint32 table_count, i;
+    uint64 total_size;
+    WASMTable *table;
+
+    read_leb_uint32(p, p_end, table_count);
+    bh_assert(module->import_table_count + table_count <= 1);
+
+    if (table_count) {
+        module->table_count = table_count;
+        total_size = sizeof(WASMTable) * (uint64)table_count;
+        if (!(module->tables = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
+            return false;
+        }
+
+        /* load each table */
+        table = module->tables;
+        for (i = 0; i < table_count; i++, table++)
+            if (!load_table(&p, p_end, table, error_buf, error_buf_size))
+                return false;
+    }
+
+    bh_assert(p == p_end);
+    LOG_VERBOSE("Load table section success.\n");
+    return true;
+}
+
+static bool
+load_memory_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
+                    char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end;
+    uint32 memory_count, i;
+    uint64 total_size;
+    WASMMemory *memory;
+
+    read_leb_uint32(p, p_end, memory_count);
+    bh_assert(module->import_memory_count + memory_count <= 1);
+
+    if (memory_count) {
+        module->memory_count = memory_count;
+        total_size = sizeof(WASMMemory) * (uint64)memory_count;
+        if (!(module->memories = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
+            return false;
+        }
+
+        /* load each memory */
+        memory = module->memories;
+        for (i = 0; i < memory_count; i++, memory++)
+            if (!load_memory(&p, p_end, memory, error_buf, error_buf_size))
+                return false;
+    }
+
+    bh_assert(p == p_end);
+    LOG_VERBOSE("Load memory section success.\n");
+    return true;
+}
+
+static bool
+load_global_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
+                    char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end;
+    uint32 global_count, i;
+    uint64 total_size;
+    WASMGlobal *global;
+    uint8 mutable;
+
+    read_leb_uint32(p, p_end, global_count);
+
+    if (global_count) {
+        module->global_count = global_count;
+        total_size = sizeof(WASMGlobal) * (uint64)global_count;
+        if (!(module->globals = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
+            return false;
+        }
+
+        global = module->globals;
+
+        for(i = 0; i < global_count; i++, global++) {
+            CHECK_BUF(p, p_end, 2);
+            global->type = read_uint8(p);
+            mutable = read_uint8(p);
+            bh_assert(mutable < 2);
+            global->is_mutable = mutable ? true : false;
+
+            /* initialize expression */
+            if (!load_init_expr(&p, p_end, &(global->init_expr),
+                                global->type, error_buf, error_buf_size))
+                return false;
+
+            if (INIT_EXPR_TYPE_GET_GLOBAL == global->init_expr.init_expr_type) {
+                /**
+                 * Currently, constant expressions occurring as initializers
+                 * of globals are further constrained in that contained
+                 * global.get instructions are
+                 * only allowed to refer to imported globals.
+                 */
+                uint32 target_global_index = global->init_expr.u.global_index;
+                bh_assert(target_global_index < module->import_global_count);
+                (void)target_global_index;
+            }
+        }
+    }
+
+    bh_assert(p == p_end);
+    LOG_VERBOSE("Load global section success.\n");
+    return true;
+}
+
+static bool
+load_export_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
+                    char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end;
+    uint32 export_count, i, j, index;
+    uint64 total_size;
+    uint32 str_len;
+    WASMExport *export;
+    const char *name;
+
+    read_leb_uint32(p, p_end, export_count);
+
+    if (export_count) {
+        module->export_count = export_count;
+        total_size = sizeof(WASMExport) * (uint64)export_count;
+        if (!(module->exports = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
+            return false;
+        }
+
+        export = module->exports;
+        for (i = 0; i < export_count; i++, export++) {
+            read_leb_uint32(p, p_end, str_len);
+            CHECK_BUF(p, p_end, str_len);
+
+            for (j = 0; j < i; j++) {
+                name = module->exports[j].name;
+                bh_assert(!(strlen(name) == str_len
+                            && memcmp(name, p, str_len) == 0));
+            }
+
+            if (!(export->name = const_str_list_insert(p, str_len, module,
+                            error_buf, error_buf_size))) {
+                return false;
+            }
+
+            p += str_len;
+            CHECK_BUF(p, p_end, 1);
+            export->kind = read_uint8(p);
+            read_leb_uint32(p, p_end, index);
+            export->index = index;
+
+            switch(export->kind) {
+                /*function index*/
+                case EXPORT_KIND_FUNC:
+                    bh_assert(index < module->function_count
+                                      + module->import_function_count);
+                    break;
+                /*table index*/
+                case EXPORT_KIND_TABLE:
+                    bh_assert(index < module->table_count
+                                      + module->import_table_count);
+                    break;
+                /*memory index*/
+                case EXPORT_KIND_MEMORY:
+                    bh_assert(index < module->memory_count
+                                      + module->import_memory_count);
+                    break;
+                /*global index*/
+                case EXPORT_KIND_GLOBAL:
+                    bh_assert(index < module->global_count
+                                      + module->import_global_count);
+                    break;
+                default:
+                    bh_assert(0);
+                    break;
+            }
+        }
+    }
+
+    bh_assert(p == p_end);
+    LOG_VERBOSE("Load export section success.\n");
+    (void)name;
+    return true;
+}
+
+static bool
+load_table_segment_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
+                           char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end;
+    uint32 table_segment_count, i, j, table_index, function_count, function_index;
+    uint64 total_size;
+    WASMTableSeg *table_segment;
+
+    read_leb_uint32(p, p_end, table_segment_count);
+
+    if (table_segment_count) {
+        module->table_seg_count = table_segment_count;
+        total_size = sizeof(WASMTableSeg) * (uint64)table_segment_count;
+        if (!(module->table_segments = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
+            return false;
+        }
+
+        table_segment = module->table_segments;
+        for (i = 0; i < table_segment_count; i++, table_segment++) {
+            bh_assert(p < p_end);
+            read_leb_uint32(p, p_end, table_index);
+            bh_assert(table_index < module->import_table_count
+                                    + module->table_count);
+
+            table_segment->table_index = table_index;
+
+            /* initialize expression */
+            if (!load_init_expr(&p, p_end, &(table_segment->base_offset),
+                                VALUE_TYPE_I32, error_buf, error_buf_size))
+                return false;
+
+            read_leb_uint32(p, p_end, function_count);
+            table_segment->function_count = function_count;
+            total_size = sizeof(uint32) * (uint64)function_count;
+            if (!(table_segment->func_indexes = (uint32 *)
+                    loader_malloc(total_size, error_buf, error_buf_size))) {
+                return false;
+            }
+
+            for (j = 0; j < function_count; j++) {
+                read_leb_uint32(p, p_end, function_index);
+                bh_assert(function_index < module->function_count
+                                           + module->function_count);
+                table_segment->func_indexes[j] = function_index;
+            }
+        }
+    }
+
+    bh_assert(p == p_end);
+    LOG_VERBOSE("Load table segment section success.\n");
+    return true;
+}
+
+static bool
+load_data_segment_section(const uint8 *buf, const uint8 *buf_end,
+                          WASMModule *module,
+                          char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end;
+    uint32 data_seg_count, i, mem_index, data_seg_len;
+    uint64 total_size;
+    WASMDataSeg *dataseg;
+    InitializerExpression init_expr;
+#if WASM_ENABLE_BULK_MEMORY != 0
+    bool is_passive = false;
+    uint32 mem_flag;
+#endif
+
+    read_leb_uint32(p, p_end, data_seg_count);
+
+#if WASM_ENABLE_BULK_MEMORY != 0
+    bh_assert(module->data_seg_count1 == 0
+              || data_seg_count == module->data_seg_count1);
+#endif
+
+    if (data_seg_count) {
+        module->data_seg_count = data_seg_count;
+        total_size = sizeof(WASMDataSeg*) * (uint64)data_seg_count;
+        if (!(module->data_segments = loader_malloc
+                    (total_size, error_buf, error_buf_size))) {
+            return false;
+        }
+
+        for (i = 0; i < data_seg_count; i++) {
+            read_leb_uint32(p, p_end, mem_index);
+#if WASM_ENABLE_BULK_MEMORY != 0
+            is_passive = false;
+            mem_flag = mem_index & 0x03;
+            switch (mem_flag) {
+                case 0x01:
+                    is_passive = true;
+                    break;
+                case 0x00:
+                    /* no memory index, treat index as 0 */
+                    mem_index = 0;
+                    goto check_mem_index;
+                case 0x02:
+                    /* read following memory index */
+                    read_leb_uint32(p, p_end, mem_index);
+check_mem_index:
+                    bh_assert(mem_index < module->import_memory_count
+                                          + module->memory_count);
+                    break;
+                case 0x03:
+                default:
+                    bh_assert(0);
+                    break;
+            }
+#else
+            bh_assert(mem_index < module->import_memory_count
+                                  + module->memory_count);
+#endif /* WASM_ENABLE_BULK_MEMORY */
+
+#if WASM_ENABLE_BULK_MEMORY != 0
+            if (!is_passive)
+#endif
+                if (!load_init_expr(&p, p_end, &init_expr, VALUE_TYPE_I32,
+                                    error_buf, error_buf_size))
+                    return false;
+
+            read_leb_uint32(p, p_end, data_seg_len);
+
+            if (!(dataseg = module->data_segments[i] = loader_malloc
+                        (sizeof(WASMDataSeg), error_buf, error_buf_size))) {
+                return false;
+            }
+
+#if WASM_ENABLE_BULK_MEMORY != 0
+            dataseg->is_passive = is_passive;
+            if (!is_passive)
+#endif
+            {
+                bh_memcpy_s(&dataseg->base_offset, sizeof(InitializerExpression),
+                            &init_expr, sizeof(InitializerExpression));
+
+                dataseg->memory_index = mem_index;
+            }
+
+            dataseg->data_length = data_seg_len;
+            CHECK_BUF(p, p_end, data_seg_len);
+            dataseg->data = (uint8*)p;
+            p += data_seg_len;
+        }
+    }
+
+    bh_assert(p == p_end);
+    LOG_VERBOSE("Load data segment section success.\n");
+    return true;
+}
+
+#if WASM_ENABLE_BULK_MEMORY != 0
+static bool
+load_datacount_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
+                       char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end;
+    uint32 data_seg_count1 = 0;
+
+    read_leb_uint32(p, p_end, data_seg_count1);
+    module->data_seg_count1 = data_seg_count1;
+
+    bh_assert(p == p_end);
+    LOG_VERBOSE("Load datacount section success.\n");
+    return true;
+}
+#endif
+
+static bool
+load_code_section(const uint8 *buf, const uint8 *buf_end,
+                  const uint8 *buf_func,
+                  const uint8 *buf_func_end,
+                  WASMModule *module,
+                  char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end;
+    const uint8 *p_func = buf_func;
+    uint32 func_count = 0, code_count;
+
+    /* code has been loaded in function section, so pass it here, just check
+     * whether function and code section have inconsistent lengths */
+    read_leb_uint32(p, p_end, code_count);
+
+    if (buf_func)
+        read_leb_uint32(p_func, buf_func_end, func_count);
+
+    bh_assert(func_count == code_count);
+    LOG_VERBOSE("Load code segment section success.\n");
+    (void)code_count;
+    (void)func_count;
+    return true;
+}
+
+static bool
+load_start_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
+                   char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end;
+    WASMType *type;
+    uint32 start_function;
+
+    read_leb_uint32(p, p_end, start_function);
+
+    bh_assert(start_function < module->function_count
+                               + module->import_function_count);
+
+    if (start_function < module->import_function_count)
+        type = module->import_functions[start_function].u.function.func_type;
+    else
+        type = module->functions[start_function - module->import_function_count]
+                ->func_type;
+
+    bh_assert(type->param_count == 0 && type->result_count == 0);
+
+    module->start_function = start_function;
+
+    bh_assert(p == p_end);
+    LOG_VERBOSE("Load start section success.\n");
+    (void)type;
+    return true;
+}
+
+static bool
+load_user_section(const uint8 *buf, const uint8 *buf_end, WASMModule *module,
+                  char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *p = buf, *p_end = buf_end;
+    uint32 name_len;
+
+    bh_assert(p < p_end);
+
+    read_leb_uint32(p, p_end, name_len);
+
+    bh_assert(name_len > 0
+              && p + name_len <= p_end);
+    LOG_VERBOSE("Load custom section success.\n");
+    (void)name_len;
+    return true;
+}
+
+
+static bool
+wasm_loader_prepare_bytecode(WASMModule *module, WASMFunction *func,
+                             BlockAddr *block_addr_cache,
+                             char *error_buf, uint32 error_buf_size);
+
+#if WASM_ENABLE_FAST_INTERP != 0
+void **
+wasm_interp_get_handle_table();
+
+static void **handle_table;
+#endif
+
+static bool
+load_from_sections(WASMModule *module, WASMSection *sections,
+                   char *error_buf, uint32 error_buf_size)
+{
+    WASMExport *export;
+    WASMSection *section = sections;
+    const uint8 *buf, *buf_end, *buf_code = NULL, *buf_code_end = NULL,
+                *buf_func = NULL, *buf_func_end = NULL;
+    WASMGlobal *llvm_data_end_global = NULL, *llvm_heap_base_global = NULL;
+    WASMGlobal *llvm_stack_top_global = NULL, *global;
+    uint32 llvm_data_end = UINT32_MAX, llvm_heap_base = UINT32_MAX;
+    uint32 llvm_stack_top = UINT32_MAX, global_index, i;
+    uint32 data_end_global_index = UINT32_MAX;
+    uint32 heap_base_global_index = UINT32_MAX;
+    uint32 stack_top_global_index = UINT32_MAX;
+    BlockAddr *block_addr_cache;
+    uint64 total_size;
+
+    /* Find code and function sections if have */
+    while (section) {
+        if (section->section_type == SECTION_TYPE_CODE) {
+            buf_code = section->section_body;
+            buf_code_end = buf_code + section->section_body_size;
+        }
+        else if (section->section_type == SECTION_TYPE_FUNC) {
+            buf_func = section->section_body;
+            buf_func_end = buf_func + section->section_body_size;
+        }
+        section = section->next;
+    }
+
+    section = sections;
+    while (section) {
+        buf = section->section_body;
+        buf_end = buf + section->section_body_size;
+        LOG_DEBUG("to section %d", section->section_type);
+        switch (section->section_type) {
+            case SECTION_TYPE_USER:
+                /* unsupported user section, ignore it. */
+                if (!load_user_section(buf, buf_end, module, error_buf, error_buf_size))
+                    return false;
+                break;
+            case SECTION_TYPE_TYPE:
+                if (!load_type_section(buf, buf_end, module, error_buf, error_buf_size))
+                    return false;
+                break;
+            case SECTION_TYPE_IMPORT:
+                if (!load_import_section(buf, buf_end, module, error_buf, error_buf_size))
+                    return false;
+                break;
+            case SECTION_TYPE_FUNC:
+                if (!load_function_section(buf, buf_end, buf_code, buf_code_end,
+                            module, error_buf, error_buf_size))
+                    return false;
+                break;
+            case SECTION_TYPE_TABLE:
+                if (!load_table_section(buf, buf_end, module, error_buf, error_buf_size))
+                    return false;
+                break;
+            case SECTION_TYPE_MEMORY:
+                if (!load_memory_section(buf, buf_end, module, error_buf, error_buf_size))
+                    return false;
+                break;
+            case SECTION_TYPE_GLOBAL:
+                if (!load_global_section(buf, buf_end, module, error_buf, error_buf_size))
+                    return false;
+                break;
+            case SECTION_TYPE_EXPORT:
+                if (!load_export_section(buf, buf_end, module, error_buf, error_buf_size))
+                    return false;
+                break;
+            case SECTION_TYPE_START:
+                if (!load_start_section(buf, buf_end, module, error_buf, error_buf_size))
+                    return false;
+                break;
+            case SECTION_TYPE_ELEM:
+                if (!load_table_segment_section(buf, buf_end, module, error_buf, error_buf_size))
+                    return false;
+                break;
+            case SECTION_TYPE_CODE:
+                if (!load_code_section(buf, buf_end, buf_func, buf_func_end,
+                                       module, error_buf, error_buf_size))
+                    return false;
+                break;
+            case SECTION_TYPE_DATA:
+                if (!load_data_segment_section(buf, buf_end, module, error_buf, error_buf_size))
+                    return false;
+                break;
+#if WASM_ENABLE_BULK_MEMORY != 0
+            case SECTION_TYPE_DATACOUNT:
+                if (!load_datacount_section(buf, buf_end, module, error_buf, error_buf_size))
+                    return false;
+                break;
+#endif
+            default:
+                set_error_buf(error_buf, error_buf_size,
+                              "WASM module load failed: invalid section id");
+                return false;
+        }
+
+        section = section->next;
+    }
+
+#if WASM_ENABLE_FAST_INTERP != 0
+    handle_table = wasm_interp_get_handle_table();
+#endif
+
+    total_size = sizeof(BlockAddr) * (uint64)BLOCK_ADDR_CACHE_SIZE * BLOCK_ADDR_CONFLICT_SIZE;
+    if (!(block_addr_cache = loader_malloc
+                (total_size, error_buf, error_buf_size))) {
+        return false;
+    }
+
+    for (i = 0; i < module->function_count; i++) {
+        WASMFunction *func = module->functions[i];
+        memset(block_addr_cache, 0, (uint32)total_size);
+        if (!wasm_loader_prepare_bytecode(module, func, block_addr_cache,
+                                          error_buf, error_buf_size)) {
+            wasm_runtime_free(block_addr_cache);
+            return false;
+        }
+    }
+    wasm_runtime_free(block_addr_cache);
+
+    /* Resolve llvm auxiliary data/stack/heap info and reset memory info */
+    if (!module->possible_memory_grow) {
+        export = module->exports;
+        for (i = 0; i < module->export_count; i++, export++) {
+            if (export->kind == EXPORT_KIND_GLOBAL) {
+                if (!strcmp(export->name, "__heap_base")) {
+                    global_index = export->index - module->import_global_count;
+                    global = module->globals + global_index;
+                    if (global->type == VALUE_TYPE_I32
+                        && !global->is_mutable
+                        && global->init_expr.init_expr_type ==
+                                INIT_EXPR_TYPE_I32_CONST) {
+                        heap_base_global_index = global_index;
+                        llvm_heap_base_global = global;
+                        llvm_heap_base = global->init_expr.u.i32;
+                        LOG_VERBOSE("found llvm __heap_base global, value: %d\n",
+                                    llvm_heap_base);
+                    }
+                }
+                else if (!strcmp(export->name, "__data_end")) {
+                    global_index = export->index - module->import_global_count;
+                    global = module->globals + global_index;
+                    if (global->type == VALUE_TYPE_I32
+                        && !global->is_mutable
+                        && global->init_expr.init_expr_type ==
+                                INIT_EXPR_TYPE_I32_CONST) {
+                        data_end_global_index = global_index;
+                        llvm_data_end_global = global;
+                        llvm_data_end = global->init_expr.u.i32;
+                        LOG_VERBOSE("found llvm __data_end global, value: %d\n",
+                                    llvm_data_end);
+
+                        llvm_data_end = align_uint(llvm_data_end, 16);
+                    }
+                }
+
+                if (llvm_data_end_global && llvm_heap_base_global) {
+                    if ((data_end_global_index == heap_base_global_index + 1
+                         && (int32)data_end_global_index > 1)
+                        || (heap_base_global_index == data_end_global_index + 1
+                            && (int32)heap_base_global_index > 1)) {
+                        global_index =
+                            data_end_global_index < heap_base_global_index
+                            ? data_end_global_index - 1 : heap_base_global_index - 1;
+                        global = module->globals + global_index;
+                        if (global->type == VALUE_TYPE_I32
+                            && global->is_mutable
+                            && global->init_expr.init_expr_type ==
+                                        INIT_EXPR_TYPE_I32_CONST) {
+                            llvm_stack_top_global = global;
+                            llvm_stack_top = global->init_expr.u.i32;
+                            stack_top_global_index = global_index;
+                            LOG_VERBOSE("found llvm stack top global, "
+                                        "value: %d, global index: %d\n",
+                                        llvm_stack_top, global_index);
+                        }
+                    }
+                    break;
+                }
+            }
+        }
+
+        if (llvm_data_end_global
+            && llvm_heap_base_global
+            && llvm_stack_top_global
+            && llvm_stack_top <= llvm_heap_base) {
+            WASMMemoryImport *memory_import;
+            WASMMemory *memory;
+            uint64 init_memory_size;
+            uint32 shrunk_memory_size = llvm_heap_base > llvm_data_end
+                                        ? llvm_heap_base : llvm_data_end;
+            if (module->import_memory_count) {
+                memory_import = &module->import_memories[0].u.memory;
+                init_memory_size = (uint64)memory_import->num_bytes_per_page *
+                                   memory_import->init_page_count;
+                if (llvm_heap_base <= init_memory_size
+                    && llvm_data_end <= init_memory_size) {
+                    /* Reset memory info to decrease memory usage */
+                    memory_import->num_bytes_per_page = shrunk_memory_size;
+                    memory_import->init_page_count = 1;
+                    LOG_VERBOSE("reset import memory size to %d\n",
+                                shrunk_memory_size);
+                }
+            }
+            if (module->memory_count) {
+                memory = &module->memories[0];
+                init_memory_size = (uint64)memory->num_bytes_per_page *
+                             memory->init_page_count;
+                if (llvm_heap_base <= init_memory_size
+                    && llvm_data_end <= init_memory_size) {
+                    /* Reset memory info to decrease memory usage */
+                    memory->num_bytes_per_page = shrunk_memory_size;
+                    memory->init_page_count = 1;
+                    LOG_VERBOSE("reset memory size to %d\n", shrunk_memory_size);
+                }
+            }
+
+            module->llvm_aux_data_end = llvm_data_end;
+            module->llvm_aux_stack_bottom = llvm_stack_top;
+            module->llvm_aux_stack_size = llvm_stack_top > llvm_data_end
+                                          ? llvm_stack_top - llvm_data_end
+                                          : llvm_stack_top;
+            module->llvm_aux_stack_global_index = stack_top_global_index;
+            LOG_VERBOSE("aux stack bottom: %d, size: %d\n",
+                        module->llvm_aux_stack_bottom,
+                        module->llvm_aux_stack_size);
+        }
+    }
+
+    return true;
+}
+
+#if BH_ENABLE_MEMORY_PROFILING != 0
+static void wasm_loader_free(void *ptr)
+{
+    wasm_runtime_free(ptr);
+}
+#else
+#define wasm_loader_free wasm_free
+#endif
+
+static WASMModule*
+create_module(char *error_buf, uint32 error_buf_size)
+{
+    WASMModule *module = loader_malloc(sizeof(WASMModule),
+                                       error_buf, error_buf_size);
+
+    if (!module) {
+        return NULL;
+    }
+
+    module->module_type = Wasm_Module_Bytecode;
+
+    /* Set start_function to -1, means no start function */
+    module->start_function = (uint32)-1;
+
+#if WASM_ENABLE_MULTI_MODULE != 0
+    module->import_module_list = &module->import_module_list_head;
+#endif
+    return module;
+}
+
+WASMModule *
+wasm_loader_load_from_sections(WASMSection *section_list,
+                               char *error_buf, uint32 error_buf_size)
+{
+    WASMModule *module = create_module(error_buf, error_buf_size);
+    if (!module)
+        return NULL;
+
+    if (!load_from_sections(module, section_list, error_buf, error_buf_size)) {
+        wasm_loader_unload(module);
+        return NULL;
+    }
+
+    LOG_VERBOSE("Load module from sections success.\n");
+    return module;
+}
+
+static void
+destroy_sections(WASMSection *section_list)
+{
+    WASMSection *section = section_list, *next;
+    while (section) {
+        next = section->next;
+        wasm_runtime_free(section);
+        section = next;
+    }
+}
+
+static uint8 section_ids[] = {
+    SECTION_TYPE_USER,
+    SECTION_TYPE_TYPE,
+    SECTION_TYPE_IMPORT,
+    SECTION_TYPE_FUNC,
+    SECTION_TYPE_TABLE,
+    SECTION_TYPE_MEMORY,
+    SECTION_TYPE_GLOBAL,
+    SECTION_TYPE_EXPORT,
+    SECTION_TYPE_START,
+    SECTION_TYPE_ELEM,
+#if WASM_ENABLE_BULK_MEMORY != 0
+    SECTION_TYPE_DATACOUNT,
+#endif
+    SECTION_TYPE_CODE,
+    SECTION_TYPE_DATA
+};
+
+static uint8
+get_section_index(uint8 section_type)
+{
+    uint8 max_id = sizeof(section_ids) / sizeof(uint8);
+
+    for (uint8 i = 0; i < max_id; i++) {
+        if (section_type == section_ids[i])
+            return i;
+    }
+
+    return (uint8)-1;
+}
+
+static bool
+create_sections(const uint8 *buf, uint32 size,
+                WASMSection **p_section_list,
+                char *error_buf, uint32 error_buf_size)
+{
+    WASMSection *section_list_end = NULL, *section;
+    const uint8 *p = buf, *p_end = buf + size/*, *section_body*/;
+    uint8 section_type, section_index, last_section_index = (uint8)-1;
+    uint32 section_size;
+
+    bh_assert(!*p_section_list);
+
+    p += 8;
+    while (p < p_end) {
+        CHECK_BUF(p, p_end, 1);
+        section_type = read_uint8(p);
+        section_index = get_section_index(section_type);
+        if (section_index != (uint8)-1) {
+            if (section_type != SECTION_TYPE_USER) {
+                /* Custom sections may be inserted at any place,
+                   while other sections must occur at most once
+                   and in prescribed order. */
+                bh_assert(last_section_index == (uint8)-1
+                          || last_section_index < section_index);
+                last_section_index = section_index;
+            }
+            CHECK_BUF1(p, p_end, 1);
+            read_leb_uint32(p, p_end, section_size);
+            CHECK_BUF1(p, p_end, section_size);
+
+            if (!(section = loader_malloc(sizeof(WASMSection),
+                                          error_buf, error_buf_size))) {
+                return false;
+            }
+
+            section->section_type = section_type;
+            section->section_body = (uint8*)p;
+            section->section_body_size = section_size;
+
+            if (!*p_section_list)
+                *p_section_list = section_list_end = section;
+            else {
+                section_list_end->next = section;
+                section_list_end = section;
+            }
+
+            p += section_size;
+        }
+        else {
+            bh_assert(0);
+        }
+    }
+
+    (void)last_section_index;
+    return true;
+}
+
+static void
+exchange32(uint8* p_data)
+{
+    uint8 value = *p_data;
+    *p_data = *(p_data + 3);
+    *(p_data + 3) = value;
+
+    value = *(p_data + 1);
+    *(p_data + 1) = *(p_data + 2);
+    *(p_data + 2) = value;
+}
+
+static union {
+    int a;
+    char b;
+} __ue = { .a = 1 };
+
+#define is_little_endian() (__ue.b == 1)
+
+static bool
+load(const uint8 *buf, uint32 size, WASMModule *module,
+     char *error_buf, uint32 error_buf_size)
+{
+    const uint8 *buf_end = buf + size;
+    const uint8 *p = buf, *p_end = buf_end;
+    uint32 magic_number, version;
+    WASMSection *section_list = NULL;
+
+    CHECK_BUF1(p, p_end, sizeof(uint32));
+    magic_number = read_uint32(p);
+    if (!is_little_endian())
+        exchange32((uint8*)&magic_number);
+
+    bh_assert(magic_number == WASM_MAGIC_NUMBER);
+
+    CHECK_BUF1(p, p_end, sizeof(uint32));
+    version = read_uint32(p);
+    if (!is_little_endian())
+        exchange32((uint8*)&version);
+
+    if (version != WASM_CURRENT_VERSION) {
+        set_error_buf(error_buf, error_buf_size,
+                      "WASM module load failed: unknown binary version");
+        return false;
+    }
+
+    if (!create_sections(buf, size, &section_list, error_buf, error_buf_size)
+        || !load_from_sections(module, section_list, error_buf, error_buf_size)) {
+        destroy_sections(section_list);
+        return false;
+    }
+
+    destroy_sections(section_list);
+    (void)p_end;
+    return true;
+}
+
+WASMModule*
+wasm_loader_load(const uint8 *buf, uint32 size, char *error_buf, uint32 error_buf_size)
+{
+    WASMModule *module = create_module(error_buf, error_buf_size);
+    if (!module) {
+        return NULL;
+    }
+
+    if (!load(buf, size, module, error_buf, error_buf_size)) {
+        LOG_VERBOSE("Load module failed, %s", error_buf);
+        goto fail;
+    }
+
+    LOG_VERBOSE("Load module success");
+    return module;
+
+fail:
+    wasm_loader_unload(module);
+    return NULL;
+}
+
+void
+wasm_loader_unload(WASMModule *module)
+{
+    uint32 i;
+
+    if (!module)
+        return;
+
+    if (module->types) {
+        for (i = 0; i < module->type_count; i++) {
+            if (module->types[i])
+                wasm_runtime_free(module->types[i]);
+        }
+        wasm_runtime_free(module->types);
+    }
+
+    if (module->imports)
+        wasm_runtime_free(module->imports);
+
+    if (module->functions) {
+        for (i = 0; i < module->function_count; i++) {
+            if (module->functions[i]) {
+                if (module->functions[i]->local_offsets)
+                    wasm_runtime_free(module->functions[i]->local_offsets);
+#if WASM_ENABLE_FAST_INTERP != 0
+                if (module->functions[i]->code_compiled)
+                    wasm_runtime_free(module->functions[i]->code_compiled);
+                if (module->functions[i]->consts)
+                    wasm_runtime_free(module->functions[i]->consts);
+#endif
+                wasm_runtime_free(module->functions[i]);
+            }
+        }
+        wasm_runtime_free(module->functions);
+    }
+
+    if (module->tables)
+        wasm_runtime_free(module->tables);
+
+    if (module->memories)
+        wasm_runtime_free(module->memories);
+
+    if (module->globals)
+        wasm_runtime_free(module->globals);
+
+    if (module->exports)
+        wasm_runtime_free(module->exports);
+
+    if (module->table_segments) {
+        for (i = 0; i < module->table_seg_count; i++) {
+            if (module->table_segments[i].func_indexes)
+                wasm_runtime_free(module->table_segments[i].func_indexes);
+        }
+        wasm_runtime_free(module->table_segments);
+    }
+
+    if (module->data_segments) {
+        for (i = 0; i < module->data_seg_count; i++) {
+            if (module->data_segments[i])
+                wasm_runtime_free(module->data_segments[i]);
+        }
+        wasm_runtime_free(module->data_segments);
+    }
+
+    if (module->const_str_list) {
+        StringNode *node = module->const_str_list, *node_next;
+        while (node) {
+            node_next = node->next;
+            wasm_runtime_free(node);
+            node = node_next;
+        }
+    }
+
+    wasm_runtime_free(module);
+}
+
+bool
+wasm_loader_find_block_addr(BlockAddr *block_addr_cache,
+                            const uint8 *start_addr,
+                            const uint8 *code_end_addr,
+                            uint8 block_type,
+                            uint8 **p_else_addr,
+                            uint8 **p_end_addr,
+                            char *error_buf,
+                            uint32 error_buf_size)
+{
+    const uint8 *p = start_addr, *p_end = code_end_addr;
+    uint8 *else_addr = NULL;
+    uint32 block_nested_depth = 1, count, i, j, t;
+    uint8 opcode, u8;
+    BlockAddr block_stack[16] = { 0 }, *block;
+
+    i = ((uintptr_t)start_addr) % BLOCK_ADDR_CACHE_SIZE;
+    block = block_addr_cache + BLOCK_ADDR_CONFLICT_SIZE * i;
+
+    for (j = 0; j < BLOCK_ADDR_CONFLICT_SIZE; j++) {
+        if (block[j].start_addr == start_addr) {
+            /* Cache hit */
+            *p_else_addr = block[j].else_addr;
+            *p_end_addr = block[j].end_addr;
+            return true;
+        }
+    }
+
+    /* Cache unhit */
+    block_stack[0].start_addr = start_addr;
+
+    while (p < code_end_addr) {
+        opcode = *p++;
+
+        switch (opcode) {
+            case WASM_OP_UNREACHABLE:
+            case WASM_OP_NOP:
+                break;
+
+            case WASM_OP_BLOCK:
+            case WASM_OP_LOOP:
+            case WASM_OP_IF:
+                CHECK_BUF(p, p_end, 1);
+                /* block result type: 0x40/0x7F/0x7E/0x7D/0x7C */
+                u8 = read_uint8(p);
+                if (block_nested_depth < sizeof(block_stack)/sizeof(BlockAddr)) {
+                    block_stack[block_nested_depth].start_addr = p;
+                    block_stack[block_nested_depth].else_addr = NULL;
+                }
+                block_nested_depth++;
+                break;
+
+            case WASM_OP_ELSE:
+                if (block_type == BLOCK_TYPE_IF && block_nested_depth == 1)
+                    else_addr = (uint8*)(p - 1);
+                if (block_nested_depth - 1 < sizeof(block_stack)/sizeof(BlockAddr))
+                    block_stack[block_nested_depth - 1].else_addr = (uint8*)(p - 1);
+                break;
+
+            case WASM_OP_END:
+                if (block_nested_depth == 1) {
+                    if (block_type == BLOCK_TYPE_IF)
+                        *p_else_addr = else_addr;
+                    *p_end_addr = (uint8*)(p - 1);
+
+                    block_stack[0].end_addr = (uint8*)(p - 1);
+                    for (t = 0; t < sizeof(block_stack)/sizeof(BlockAddr); t++) {
+                        start_addr = block_stack[t].start_addr;
+                        if (start_addr) {
+                            i = ((uintptr_t)start_addr) % BLOCK_ADDR_CACHE_SIZE;
+                            block = block_addr_cache + BLOCK_ADDR_CONFLICT_SIZE * i;
+                            for (j = 0; j < BLOCK_ADDR_CONFLICT_SIZE; j++)
+                                if (!block[j].start_addr)
+                                    break;
+
+                            if (j == BLOCK_ADDR_CONFLICT_SIZE) {
+                                memmove(block + 1, block, (BLOCK_ADDR_CONFLICT_SIZE - 1) *
+                                                          sizeof(BlockAddr));
+                                j = 0;
+
+                            }
+                            block[j].start_addr = block_stack[t].start_addr;
+                            block[j].else_addr = block_stack[t].else_addr;
+                            block[j].end_addr = block_stack[t].end_addr;
+                        }
+                        else
+                            break;
+                    }
+                    return true;
+                }
+                else {
+                    block_nested_depth--;
+                    if (block_nested_depth < sizeof(block_stack)/sizeof(BlockAddr))
+                        block_stack[block_nested_depth].end_addr = (uint8*)(p - 1);
+                }
+                break;
+
+            case WASM_OP_BR:
+            case WASM_OP_BR_IF:
+                skip_leb_uint32(p, p_end); /* labelidx */
+                break;
+
+            case WASM_OP_BR_TABLE:
+                read_leb_uint32(p, p_end, count); /* lable num */
+                for (i = 0; i <= count; i++) /* lableidxs */
+                    skip_leb_uint32(p, p_end);
+                break;
+
+            case WASM_OP_RETURN:
+                break;
+
+            case WASM_OP_CALL:
+                skip_leb_uint32(p, p_end); /* funcidx */
+                break;
+
+            case WASM_OP_CALL_INDIRECT:
+                skip_leb_uint32(p, p_end); /* typeidx */
+                CHECK_BUF(p, p_end, 1);
+                u8 = read_uint8(p); /* 0x00 */
+                break;
+
+            case WASM_OP_DROP:
+            case WASM_OP_SELECT:
+            case WASM_OP_DROP_64:
+            case WASM_OP_SELECT_64:
+                break;
+
+            case WASM_OP_GET_LOCAL:
+            case WASM_OP_SET_LOCAL:
+            case WASM_OP_TEE_LOCAL:
+            case WASM_OP_GET_GLOBAL:
+            case WASM_OP_SET_GLOBAL:
+                skip_leb_uint32(p, p_end); /* localidx */
+                break;
+
+            case EXT_OP_GET_LOCAL_FAST:
+            case EXT_OP_SET_LOCAL_FAST:
+            case EXT_OP_TEE_LOCAL_FAST:
+                CHECK_BUF(p, p_end, 1);
+                p++;
+                break;
+
+            case WASM_OP_I32_LOAD:
+            case WASM_OP_I64_LOAD:
+            case WASM_OP_F32_LOAD:
+            case WASM_OP_F64_LOAD:
+            case WASM_OP_I32_LOAD8_S:
+            case WASM_OP_I32_LOAD8_U:
+            case WASM_OP_I32_LOAD16_S:
+            case WASM_OP_I32_LOAD16_U:
+            case WASM_OP_I64_LOAD8_S:
+            case WASM_OP_I64_LOAD8_U:
+            case WASM_OP_I64_LOAD16_S:
+            case WASM_OP_I64_LOAD16_U:
+            case WASM_OP_I64_LOAD32_S:
+            case WASM_OP_I64_LOAD32_U:
+            case WASM_OP_I32_STORE:
+            case WASM_OP_I64_STORE:
+            case WASM_OP_F32_STORE:
+            case WASM_OP_F64_STORE:
+            case WASM_OP_I32_STORE8:
+            case WASM_OP_I32_STORE16:
+            case WASM_OP_I64_STORE8:
+            case WASM_OP_I64_STORE16:
+            case WASM_OP_I64_STORE32:
+                skip_leb_uint32(p, p_end); /* align */
+                skip_leb_uint32(p, p_end); /* offset */
+                break;
+
+            case WASM_OP_MEMORY_SIZE:
+            case WASM_OP_MEMORY_GROW:
+                skip_leb_uint32(p, p_end); /* 0x00 */
+                break;
+
+            case WASM_OP_I32_CONST:
+                skip_leb_int32(p, p_end);
+                break;
+            case WASM_OP_I64_CONST:
+                skip_leb_int64(p, p_end);
+                break;
+            case WASM_OP_F32_CONST:
+                p += sizeof(float32);
+                break;
+            case WASM_OP_F64_CONST:
+                p += sizeof(float64);
+                break;
+
+            case WASM_OP_I32_EQZ:
+            case WASM_OP_I32_EQ:
+            case WASM_OP_I32_NE:
+            case WASM_OP_I32_LT_S:
+            case WASM_OP_I32_LT_U:
+            case WASM_OP_I32_GT_S:
+            case WASM_OP_I32_GT_U:
+            case WASM_OP_I32_LE_S:
+            case WASM_OP_I32_LE_U:
+            case WASM_OP_I32_GE_S:
+            case WASM_OP_I32_GE_U:
+            case WASM_OP_I64_EQZ:
+            case WASM_OP_I64_EQ:
+            case WASM_OP_I64_NE:
+            case WASM_OP_I64_LT_S:
+            case WASM_OP_I64_LT_U:
+            case WASM_OP_I64_GT_S:
+            case WASM_OP_I64_GT_U:
+            case WASM_OP_I64_LE_S:
+            case WASM_OP_I64_LE_U:
+            case WASM_OP_I64_GE_S:
+            case WASM_OP_I64_GE_U:
+            case WASM_OP_F32_EQ:
+            case WASM_OP_F32_NE:
+            case WASM_OP_F32_LT:
+            case WASM_OP_F32_GT:
+            case WASM_OP_F32_LE:
+            case WASM_OP_F32_GE:
+            case WASM_OP_F64_EQ:
+            case WASM_OP_F64_NE:
+            case WASM_OP_F64_LT:
+            case WASM_OP_F64_GT:
+            case WASM_OP_F64_LE:
+            case WASM_OP_F64_GE:
+            case WASM_OP_I32_CLZ:
+            case WASM_OP_I32_CTZ:
+            case WASM_OP_I32_POPCNT:
+            case WASM_OP_I32_ADD:
+            case WASM_OP_I32_SUB:
+            case WASM_OP_I32_MUL:
+            case WASM_OP_I32_DIV_S:
+            case WASM_OP_I32_DIV_U:
+            case WASM_OP_I32_REM_S:
+            case WASM_OP_I32_REM_U:
+            case WASM_OP_I32_AND:
+            case WASM_OP_I32_OR:
+            case WASM_OP_I32_XOR:
+            case WASM_OP_I32_SHL:
+            case WASM_OP_I32_SHR_S:
+            case WASM_OP_I32_SHR_U:
+            case WASM_OP_I32_ROTL:
+            case WASM_OP_I32_ROTR:
+            case WASM_OP_I64_CLZ:
+            case WASM_OP_I64_CTZ:
+            case WASM_OP_I64_POPCNT:
+            case WASM_OP_I64_ADD:
+            case WASM_OP_I64_SUB:
+            case WASM_OP_I64_MUL:
+            case WASM_OP_I64_DIV_S:
+            case WASM_OP_I64_DIV_U:
+            case WASM_OP_I64_REM_S:
+            case WASM_OP_I64_REM_U:
+            case WASM_OP_I64_AND:
+            case WASM_OP_I64_OR:
+            case WASM_OP_I64_XOR:
+            case WASM_OP_I64_SHL:
+            case WASM_OP_I64_SHR_S:
+            case WASM_OP_I64_SHR_U:
+            case WASM_OP_I64_ROTL:
+            case WASM_OP_I64_ROTR:
+            case WASM_OP_F32_ABS:
+            case WASM_OP_F32_NEG:
+            case WASM_OP_F32_CEIL:
+            case WASM_OP_F32_FLOOR:
+            case WASM_OP_F32_TRUNC:
+            case WASM_OP_F32_NEAREST:
+            case WASM_OP_F32_SQRT:
+            case WASM_OP_F32_ADD:
+            case WASM_OP_F32_SUB:
+            case WASM_OP_F32_MUL:
+            case WASM_OP_F32_DIV:
+            case WASM_OP_F32_MIN:
+            case WASM_OP_F32_MAX:
+            case WASM_OP_F32_COPYSIGN:
+            case WASM_OP_F64_ABS:
+            case WASM_OP_F64_NEG:
+            case WASM_OP_F64_CEIL:
+            case WASM_OP_F64_FLOOR:
+            case WASM_OP_F64_TRUNC:
+            case WASM_OP_F64_NEAREST:
+            case WASM_OP_F64_SQRT:
+            case WASM_OP_F64_ADD:
+            case WASM_OP_F64_SUB:
+            case WASM_OP_F64_MUL:
+            case WASM_OP_F64_DIV:
+            case WASM_OP_F64_MIN:
+            case WASM_OP_F64_MAX:
+            case WASM_OP_F64_COPYSIGN:
+            case WASM_OP_I32_WRAP_I64:
+            case WASM_OP_I32_TRUNC_S_F32:
+            case WASM_OP_I32_TRUNC_U_F32:
+            case WASM_OP_I32_TRUNC_S_F64:
+            case WASM_OP_I32_TRUNC_U_F64:
+            case WASM_OP_I64_EXTEND_S_I32:
+            case WASM_OP_I64_EXTEND_U_I32:
+            case WASM_OP_I64_TRUNC_S_F32:
+            case WASM_OP_I64_TRUNC_U_F32:
+            case WASM_OP_I64_TRUNC_S_F64:
+            case WASM_OP_I64_TRUNC_U_F64:
+            case WASM_OP_F32_CONVERT_S_I32:
+            case WASM_OP_F32_CONVERT_U_I32:
+            case WASM_OP_F32_CONVERT_S_I64:
+            case WASM_OP_F32_CONVERT_U_I64:
+            case WASM_OP_F32_DEMOTE_F64:
+            case WASM_OP_F64_CONVERT_S_I32:
+            case WASM_OP_F64_CONVERT_U_I32:
+            case WASM_OP_F64_CONVERT_S_I64:
+            case WASM_OP_F64_CONVERT_U_I64:
+            case WASM_OP_F64_PROMOTE_F32:
+            case WASM_OP_I32_REINTERPRET_F32:
+            case WASM_OP_I64_REINTERPRET_F64:
+            case WASM_OP_F32_REINTERPRET_I32:
+            case WASM_OP_F64_REINTERPRET_I64:
+            case WASM_OP_I32_EXTEND8_S:
+            case WASM_OP_I32_EXTEND16_S:
+            case WASM_OP_I64_EXTEND8_S:
+            case WASM_OP_I64_EXTEND16_S:
+            case WASM_OP_I64_EXTEND32_S:
+                break;
+            case WASM_OP_MISC_PREFIX:
+            {
+                opcode = read_uint8(p);
+                switch (opcode) {
+                    case WASM_OP_I32_TRUNC_SAT_S_F32:
+                    case WASM_OP_I32_TRUNC_SAT_U_F32:
+                    case WASM_OP_I32_TRUNC_SAT_S_F64:
+                    case WASM_OP_I32_TRUNC_SAT_U_F64:
+                    case WASM_OP_I64_TRUNC_SAT_S_F32:
+                    case WASM_OP_I64_TRUNC_SAT_U_F32:
+                    case WASM_OP_I64_TRUNC_SAT_S_F64:
+                    case WASM_OP_I64_TRUNC_SAT_U_F64:
+                        break;
+#if WASM_ENABLE_BULK_MEMORY != 0
+                    case WASM_OP_MEMORY_INIT:
+                        skip_leb_uint32(p, p_end);
+                        /* skip memory idx */
+                        p++;
+                        break;
+                    case WASM_OP_DATA_DROP:
+                        skip_leb_uint32(p, p_end);
+                        break;
+                    case WASM_OP_MEMORY_COPY:
+                        /* skip two memory idx */
+                        p += 2;
+                        break;
+                    case WASM_OP_MEMORY_FILL:
+                        /* skip memory idx */
+                        p++;
+                        break;
+#endif
+                    default:
+                        bh_assert(0);
+                        break;
+                }
+                break;
+            }
+
+            default:
+                bh_assert(0);
+                break;
+        }
+    }
+
+    (void)u8;
+    return false;
+}
+
+#define REF_I32   VALUE_TYPE_I32
+#define REF_F32   VALUE_TYPE_F32
+#define REF_I64_1 VALUE_TYPE_I64
+#define REF_I64_2 VALUE_TYPE_I64
+#define REF_F64_1 VALUE_TYPE_F64
+#define REF_F64_2 VALUE_TYPE_F64
+#define REF_ANY   VALUE_TYPE_ANY
+
+#if WASM_ENABLE_FAST_INTERP != 0
+
+#if WASM_DEBUG_PREPROCESSOR != 0
+#define LOG_OP(...)       os_printf(__VA_ARGS__)
+#else
+#define LOG_OP(...)
+#endif
+
+#define PATCH_ELSE 0
+#define PATCH_END  1
+typedef struct BranchBlockPatch {
+    struct BranchBlockPatch *next;
+    uint8 patch_type;
+    uint8 *code_compiled;
+} BranchBlockPatch;
+#endif
+
+typedef struct BranchBlock {
+    uint8 block_type;
+    uint8 return_type;
+    uint8 *start_addr;
+    uint8 *else_addr;
+    uint8 *end_addr;
+    uint32 stack_cell_num;
+#if WASM_ENABLE_FAST_INTERP != 0
+    uint16 dynamic_offset;
+    uint8 *code_compiled;
+    BranchBlockPatch *patch_list;
+#endif
+
+    /* Indicate the operand stack is in polymorphic state.
+     * If the opcode is one of unreachable/br/br_table/return, stack is marked
+     * to polymorphic state until the block's 'end' opcode is processed.
+     * If stack is in polymorphic state and stack is empty, instruction can
+     * pop any type of value directly without decreasing stack top pointer
+     * and stack cell num. */
+    bool is_stack_polymorphic;
+} BranchBlock;
+
+typedef struct WASMLoaderContext {
+    /* frame ref stack */
+    uint8 *frame_ref;
+    uint8 *frame_ref_bottom;
+    uint8 *frame_ref_boundary;
+    uint32 frame_ref_size;
+    uint32 stack_cell_num;
+    uint32 max_stack_cell_num;
+
+    /* frame csp stack */
+    BranchBlock *frame_csp;
+    BranchBlock *frame_csp_bottom;
+    BranchBlock *frame_csp_boundary;
+    uint32 frame_csp_size;
+    uint32 csp_num;
+    uint32 max_csp_num;
+
+#if WASM_ENABLE_FAST_INTERP != 0
+    /* frame offset stack */
+    int16 *frame_offset;
+    int16 *frame_offset_bottom;
+    int16 *frame_offset_boundary;
+    uint32 frame_offset_size;
+    int16 dynamic_offset;
+    int16 start_dynamic_offset;
+    int16 max_dynamic_offset;
+
+    /* preserved local offset */
+    int16 preserved_local_offset;
+
+    /* const buffer */
+    uint8 *const_buf;
+    uint16 num_const;
+    uint16 const_buf_size;
+    uint16 const_cell_num;
+
+    /* processed code */
+    uint8 *p_code_compiled;
+    uint8 *p_code_compiled_end;
+    uint32 code_compiled_size;
+#endif
+} WASMLoaderContext;
+
+typedef struct Const {
+    WASMValue value;
+    uint16 slot_index;
+    uint8 value_type;
+} Const;
+
+static void*
+memory_realloc(void *mem_old, uint32 size_old, uint32 size_new,
+               char *error_buf, uint32 error_buf_size)
+{
+    uint8 *mem_new;
+    bh_assert(size_new > size_old);
+    if ((mem_new = loader_malloc
+                (size_new, error_buf, error_buf_size))) {
+        bh_memcpy_s(mem_new, size_new, mem_old, size_old);
+        memset(mem_new + size_old, 0, size_new - size_old);
+        wasm_runtime_free(mem_old);
+    }
+    return mem_new;
+}
+
+#define MEM_REALLOC(mem, size_old, size_new) do {                \
+    void *mem_new = memory_realloc(mem, size_old, size_new,      \
+                                   error_buf, error_buf_size);   \
+    if (!mem_new)                                                \
+        goto fail;                                               \
+    mem = mem_new;                                               \
+  } while (0)
+
+#define CHECK_CSP_PUSH() do {                                    \
+    if (ctx->frame_csp >= ctx->frame_csp_boundary) {             \
+      MEM_REALLOC(ctx->frame_csp_bottom, ctx->frame_csp_size,    \
+                  (uint32)(ctx->frame_csp_size                   \
+                           + 8 * sizeof(BranchBlock)));          \
+      ctx->frame_csp_size += (uint32)(8 * sizeof(BranchBlock));  \
+      ctx->frame_csp_boundary = ctx->frame_csp_bottom +          \
+                    ctx->frame_csp_size / sizeof(BranchBlock);   \
+      ctx->frame_csp = ctx->frame_csp_bottom + ctx->csp_num;     \
+    }                                                            \
+  } while (0)
+
+#define CHECK_CSP_POP() do {                                     \
+    bh_assert(ctx->csp_num >= 1);                                \
+  } while (0)
+
+#if WASM_ENABLE_FAST_INTERP != 0
+static bool
+check_offset_push(WASMLoaderContext *ctx,
+                  char *error_buf, uint32 error_buf_size)
+{
+    uint32 cell_num = (ctx->frame_offset - ctx->frame_offset_bottom);
+    if (ctx->frame_offset >= ctx->frame_offset_boundary) {
+        MEM_REALLOC(ctx->frame_offset_bottom, ctx->frame_offset_size,
+                    ctx->frame_offset_size + 16);
+        ctx->frame_offset_size += 16;
+        ctx->frame_offset_boundary = ctx->frame_offset_bottom +
+                    ctx->frame_offset_size / sizeof(int16);
+        ctx->frame_offset = ctx->frame_offset_bottom + cell_num;
+    }
+    return true;
+fail:
+    return false;
+}
+
+static bool
+check_offset_pop(WASMLoaderContext *ctx, uint32 cells)
+{
+    if (ctx->frame_offset - cells < ctx->frame_offset_bottom)
+        return false;
+    return true;
+}
+
+static void
+free_label_patch_list(BranchBlock *frame_csp)
+{
+    BranchBlockPatch *label_patch = frame_csp->patch_list;
+    BranchBlockPatch *next;
+    while (label_patch != NULL) {
+        next = label_patch->next;
+        wasm_runtime_free(label_patch);
+        label_patch = next;
+    }
+    frame_csp->patch_list = NULL;
+}
+
+static void
+free_all_label_patch_lists(BranchBlock *frame_csp, uint32 csp_num)
+{
+    BranchBlock *tmp_csp = frame_csp;
+
+    for (uint32 i = 0; i < csp_num; i++) {
+        free_label_patch_list(tmp_csp);
+        tmp_csp ++;
+    }
+}
+
+#endif
+
+static bool
+check_stack_push(WASMLoaderContext *ctx,
+                 char *error_buf, uint32 error_buf_size)
+{
+    if (ctx->frame_ref >= ctx->frame_ref_boundary) {
+        MEM_REALLOC(ctx->frame_ref_bottom, ctx->frame_ref_size,
+                    ctx->frame_ref_size + 16);
+        ctx->frame_ref_size += 16;
+        ctx->frame_ref_boundary = ctx->frame_ref_bottom + ctx->frame_ref_size;
+        ctx->frame_ref = ctx->frame_ref_bottom + ctx->stack_cell_num;
+    }
+    return true;
+fail:
+    return false;
+}
+
+
+static bool
+check_stack_top_values(uint8 *frame_ref, int32 stack_cell_num, uint8 type,
+                       char *error_buf, uint32 error_buf_size)
+{
+    bh_assert(!(((type == VALUE_TYPE_I32 || type == VALUE_TYPE_F32)
+                 && stack_cell_num < 1)
+                || ((type == VALUE_TYPE_I64 || type == VALUE_TYPE_F64)
+                    && stack_cell_num < 2)));
+
+    bh_assert(!((type == VALUE_TYPE_I32 && *(frame_ref - 1) != REF_I32)
+                || (type == VALUE_TYPE_F32 && *(frame_ref - 1) != REF_F32)
+                || (type == VALUE_TYPE_I64
+                    && (*(frame_ref - 2) != REF_I64_1
+                        || *(frame_ref - 1) != REF_I64_2))
+                || (type == VALUE_TYPE_F64
+                    && (*(frame_ref - 2) != REF_F64_1
+                        || *(frame_ref - 1) != REF_F64_2))));
+    return true;
+}
+
+static bool
+check_stack_pop(WASMLoaderContext *ctx, uint8 type,
+                char *error_buf, uint32 error_buf_size)
+{
+    int32 block_stack_cell_num = (int32)
+        (ctx->stack_cell_num - (ctx->frame_csp - 1)->stack_cell_num);
+
+    if (block_stack_cell_num > 0
+        && *(ctx->frame_ref - 1) == VALUE_TYPE_ANY) {
+        /* the stack top is a value of any type, return success */
+        return true;
+    }
+
+    if (!check_stack_top_values(ctx->frame_ref, block_stack_cell_num,
+                                type, error_buf, error_buf_size))
+        return false;
+
+    return true;
+}
+
+static void
+wasm_loader_ctx_destroy(WASMLoaderContext *ctx)
+{
+    if (ctx) {
+        if (ctx->frame_ref_bottom)
+            wasm_runtime_free(ctx->frame_ref_bottom);
+        if (ctx->frame_csp_bottom) {
+#if WASM_ENABLE_FAST_INTERP != 0
+            free_all_label_patch_lists(ctx->frame_csp_bottom, ctx->csp_num);
+#endif
+            wasm_runtime_free(ctx->frame_csp_bottom);
+        }
+#if WASM_ENABLE_FAST_INTERP != 0
+        if (ctx->frame_offset_bottom)
+            wasm_runtime_free(ctx->frame_offset_bottom);
+        if (ctx->const_buf)
+            wasm_runtime_free(ctx->const_buf);
+#endif
+        wasm_runtime_free(ctx);
+    }
+}
+
+static WASMLoaderContext*
+wasm_loader_ctx_init(WASMFunction *func)
+{
+    WASMLoaderContext *loader_ctx =
+        wasm_runtime_malloc(sizeof(WASMLoaderContext));
+    if (!loader_ctx)
+        return false;
+    memset(loader_ctx, 0, sizeof(WASMLoaderContext));
+
+    loader_ctx->frame_ref_size = 32;
+    if (!(loader_ctx->frame_ref_bottom = loader_ctx->frame_ref =
+            wasm_runtime_malloc(loader_ctx->frame_ref_size)))
+        goto fail;
+    memset(loader_ctx->frame_ref_bottom, 0, loader_ctx->frame_ref_size);
+    loader_ctx->frame_ref_boundary = loader_ctx->frame_ref_bottom +
+                                        loader_ctx->frame_ref_size;
+
+    loader_ctx->frame_csp_size = sizeof(BranchBlock) * 8;
+    if (!(loader_ctx->frame_csp_bottom = loader_ctx->frame_csp =
+            wasm_runtime_malloc(loader_ctx->frame_csp_size)))
+        goto fail;
+    memset(loader_ctx->frame_csp_bottom, 0, loader_ctx->frame_csp_size);
+    loader_ctx->frame_csp_boundary = loader_ctx->frame_csp_bottom + 8;
+
+#if WASM_ENABLE_FAST_INTERP != 0
+    loader_ctx->frame_offset_size = sizeof(int16) * 32;
+    if (!(loader_ctx->frame_offset_bottom = loader_ctx->frame_offset =
+            wasm_runtime_malloc(loader_ctx->frame_offset_size)))
+        goto fail;
+    memset(loader_ctx->frame_offset_bottom, 0,
+           loader_ctx->frame_offset_size);
+    loader_ctx->frame_offset_boundary = loader_ctx->frame_offset_bottom + 32;
+
+    loader_ctx->num_const = 0;
+    loader_ctx->const_buf_size = sizeof(Const) * 8;
+    if (!(loader_ctx->const_buf = wasm_runtime_malloc(loader_ctx->const_buf_size)))
+        goto fail;
+    memset(loader_ctx->const_buf, 0, loader_ctx->const_buf_size);
+
+    loader_ctx->start_dynamic_offset = loader_ctx->dynamic_offset =
+        loader_ctx->max_dynamic_offset = func->param_cell_num +
+                                            func->local_cell_num;
+#endif
+    return loader_ctx;
+
+fail:
+    wasm_loader_ctx_destroy(loader_ctx);
+    return NULL;
+}
+
+static bool
+wasm_loader_push_frame_ref(WASMLoaderContext *ctx, uint8 type,
+                           char *error_buf, uint32 error_buf_size)
+{
+    if (type == VALUE_TYPE_VOID)
+        return true;
+
+    if (!check_stack_push(ctx, error_buf, error_buf_size))
+        return false;
+
+    *ctx->frame_ref++ = type;
+    ctx->stack_cell_num++;
+    if (ctx->stack_cell_num > ctx->max_stack_cell_num)
+        ctx->max_stack_cell_num = ctx->stack_cell_num;
+
+    if (type == VALUE_TYPE_I32
+        || type == VALUE_TYPE_F32
+        || type == VALUE_TYPE_ANY)
+        return true;
+
+    if (!check_stack_push(ctx, error_buf, error_buf_size))
+        return false;
+    *ctx->frame_ref++ = type;
+    ctx->stack_cell_num++;
+    if (ctx->stack_cell_num > ctx->max_stack_cell_num)
+        ctx->max_stack_cell_num = ctx->stack_cell_num;
+    return true;
+}
+
+static bool
+wasm_loader_pop_frame_ref(WASMLoaderContext *ctx, uint8 type,
+                          char *error_buf, uint32 error_buf_size)
+{
+    BranchBlock *cur_block = ctx->frame_csp - 1;
+    int32 available_stack_cell = (int32)
+        (ctx->stack_cell_num - cur_block->stack_cell_num);
+
+    /* Directly return success if current block is in stack
+     * polymorphic state while stack is empty. */
+    if (available_stack_cell <= 0 && cur_block->is_stack_polymorphic)
+        return true;
+
+    if (type == VALUE_TYPE_VOID)
+        return true;
+
+    if (!check_stack_pop(ctx, type, error_buf, error_buf_size))
+        return false;
+
+    ctx->frame_ref--;
+    ctx->stack_cell_num--;
+
+    if (type == VALUE_TYPE_I32
+        || type == VALUE_TYPE_F32
+        || *ctx->frame_ref == VALUE_TYPE_ANY)
+        return true;
+
+    ctx->frame_ref--;
+    ctx->stack_cell_num--;
+    return true;
+}
+
+static bool
+wasm_loader_push_pop_frame_ref(WASMLoaderContext *ctx, uint8 pop_cnt,
+                               uint8 type_push, uint8 type_pop,
+                               char *error_buf, uint32 error_buf_size)
+{
+    for (int i = 0; i < pop_cnt; i++) {
+        if (!wasm_loader_pop_frame_ref(ctx, type_pop, error_buf, error_buf_size))
+            return false;
+    }
+    if (!wasm_loader_push_frame_ref(ctx, type_push, error_buf, error_buf_size))
+        return false;
+    return true;
+}
+
+static bool
+wasm_loader_push_frame_csp(WASMLoaderContext *ctx, uint8 type,
+                           uint8 ret_type, uint8* start_addr,
+                           char *error_buf, uint32 error_buf_size)
+{
+    CHECK_CSP_PUSH();
+    memset(ctx->frame_csp, 0, sizeof(BranchBlock));
+    ctx->frame_csp->block_type = type;
+    ctx->frame_csp->return_type = ret_type;
+    ctx->frame_csp->start_addr = start_addr;
+    ctx->frame_csp->stack_cell_num = ctx->stack_cell_num;
+#if WASM_ENABLE_FAST_INTERP != 0
+    ctx->frame_csp->dynamic_offset = ctx->dynamic_offset;
+    ctx->frame_csp->patch_list = NULL;
+#endif
+    ctx->frame_csp++;
+    ctx->csp_num++;
+    if (ctx->csp_num > ctx->max_csp_num)
+        ctx->max_csp_num = ctx->csp_num;
+    return true;
+fail:
+    return false;
+}
+
+static bool
+wasm_loader_pop_frame_csp(WASMLoaderContext *ctx,
+                          char *error_buf, uint32 error_buf_size)
+{
+    CHECK_CSP_POP();
+    ctx->frame_csp--;
+    ctx->csp_num--;
+    return true;
+}
+
+static bool
+wasm_loader_check_br(WASMLoaderContext *ctx, uint32 depth,
+                     char *error_buf, uint32 error_buf_size)
+{
+    BranchBlock *target_block, *cur_block;
+    int32 available_stack_cell;
+
+    bh_assert(ctx->csp_num >= depth + 1);
+
+    target_block = ctx->frame_csp - (depth + 1);
+    cur_block = ctx->frame_csp - 1;
+
+    available_stack_cell = (int32)
+        (ctx->stack_cell_num - cur_block->stack_cell_num);
+
+    if (available_stack_cell <= 0 && target_block->is_stack_polymorphic)
+        return true;
+
+    if (target_block->block_type != BLOCK_TYPE_LOOP) {
+        uint8 type = target_block->return_type;
+        if (!check_stack_top_values(ctx->frame_ref, available_stack_cell,
+                                    type, error_buf, error_buf_size))
+            return false;
+    }
+    return true;
+}
+
+#if WASM_ENABLE_FAST_INTERP != 0
+
+#if WASM_ENABLE_ABS_LABEL_ADDR != 0
+
+#define emit_label(opcode) do {                                     \
+    wasm_loader_emit_ptr(loader_ctx, handle_table[opcode]);         \
+    LOG_OP("\nemit_op [%02x]\t", opcode);                           \
+  } while (0)
+
+#define skip_label() do {                                           \
+    wasm_loader_emit_backspace(loader_ctx, sizeof(void *));         \
+    LOG_OP("\ndelete last op\n");                                   \
+  } while (0)
+
+#else
+
+#define emit_label(opcode) do {                                     \
+    int32 offset = (int32)(handle_table[opcode] - handle_table[0]); \
+    bh_assert(offset >= INT16_MIN && offset < INT16_MAX);           \
+    wasm_loader_emit_int16(loader_ctx, offset);                     \
+    LOG_OP("\nemit_op [%02x]\t", opcode);                           \
+  } while (0)
+
+// drop local.get / const / block / loop / end
+#define skip_label() do {                                           \
+    wasm_loader_emit_backspace(loader_ctx, sizeof(int16));          \
+    LOG_OP("\ndelete last op\n");                                   \
+  } while (0)
+
+#endif /* WASM_ENABLE_ABS_LABEL_ADDR */
+
+#define emit_empty_label_addr_and_frame_ip(type) do {               \
+    if (!add_label_patch_to_list(loader_ctx->frame_csp - 1, type,   \
+                                 loader_ctx->p_code_compiled,       \
+                                 error_buf, error_buf_size))        \
+        goto fail;                                                  \
+    /* label address, to be patched */                              \
+    wasm_loader_emit_ptr(loader_ctx, NULL);                         \
+  } while (0)
+
+#define emit_br_info(frame_csp) do {                                \
+    if (!wasm_loader_emit_br_info(loader_ctx, frame_csp,            \
+                                  error_buf, error_buf_size))       \
+        goto fail;                                                  \
+  } while (0)
+
+#define LAST_OP_OUTPUT_I32() (last_op >= WASM_OP_I32_EQZ                \
+                                && last_op <= WASM_OP_I32_ROTR)         \
+                            || (last_op == WASM_OP_I32_LOAD             \
+                                || last_op == WASM_OP_F32_LOAD)         \
+                            || (last_op >= WASM_OP_I32_LOAD8_S          \
+                                && last_op <= WASM_OP_I32_LOAD16_U)     \
+                            || (last_op >= WASM_OP_F32_ABS              \
+                                && last_op <= WASM_OP_F32_COPYSIGN)     \
+                            || (last_op >= WASM_OP_I32_WRAP_I64         \
+                                && last_op <= WASM_OP_I32_TRUNC_U_F64)  \
+                            || (last_op >= WASM_OP_F32_CONVERT_S_I32    \
+                                && last_op <= WASM_OP_F32_DEMOTE_F64)   \
+                            || (last_op == WASM_OP_I32_REINTERPRET_F32) \
+                            || (last_op == WASM_OP_F32_REINTERPRET_I32) \
+                            || (last_op == EXT_OP_COPY_STACK_TOP)
+
+#define LAST_OP_OUTPUT_I64() (last_op >= WASM_OP_I64_CLZ                \
+                                && last_op <= WASM_OP_I64_ROTR)         \
+                            || (last_op >= WASM_OP_F64_ABS              \
+                                && last_op <= WASM_OP_F64_COPYSIGN)     \
+                            || (last_op == WASM_OP_I64_LOAD             \
+                                || last_op == WASM_OP_F64_LOAD)         \
+                            || (last_op >= WASM_OP_I64_LOAD8_S          \
+                                && last_op <= WASM_OP_I64_LOAD32_U)     \
+                            || (last_op >= WASM_OP_I64_EXTEND_S_I32     \
+                                && last_op <= WASM_OP_I64_TRUNC_U_F64)  \
+                            || (last_op >= WASM_OP_F64_CONVERT_S_I32    \
+                                && last_op <= WASM_OP_F64_PROMOTE_F32)  \
+                            || (last_op == WASM_OP_I64_REINTERPRET_F64) \
+                            || (last_op == WASM_OP_F64_REINTERPRET_I64) \
+                            || (last_op == EXT_OP_COPY_STACK_TOP_I64)
+
+#define GET_CONST_OFFSET(type, val) do {                                \
+    if (!(wasm_loader_get_const_offset(loader_ctx, type,                \
+                                       &val, &operand_offset,           \
+                                       error_buf, error_buf_size)))     \
+        goto fail;                                                      \
+  } while (0)
+
+#define GET_CONST_F32_OFFSET(type, fval) do {                           \
+    if (!(wasm_loader_get_const_offset(loader_ctx, type,                \
+                                       &fval, &operand_offset,          \
+                                       error_buf, error_buf_size)))     \
+        goto fail;                                                      \
+  } while (0)
+
+#define GET_CONST_F64_OFFSET(type, fval) do {                           \
+    if (!(wasm_loader_get_const_offset(loader_ctx, type,                \
+                                       &fval, &operand_offset,          \
+                                       error_buf, error_buf_size)))     \
+        goto fail;                                                      \
+  } while (0)
+
+#define emit_operand(ctx, offset) do {                              \
+    wasm_loader_emit_int16(ctx, offset);                            \
+    LOG_OP("%d\t", offset);                                         \
+  } while (0)
+
+#define emit_byte(ctx, byte) do {                                   \
+    wasm_loader_emit_uint8(ctx, byte);                              \
+    LOG_OP("%d\t", byte);                                           \
+  } while (0)
+
+#define emit_leb() do {                                             \
+    wasm_loader_emit_leb(loader_ctx, p_org, p);                     \
+  } while (0)
+
+#define emit_const(value) do {                                      \
+    GET_CONST_OFFSET(VALUE_TYPE_I32, value);                        \
+    emit_operand(loader_ctx, operand_offset);                       \
+  } while (0)
+
+static bool
+wasm_loader_ctx_reinit(WASMLoaderContext *ctx)
+{
+    if (!(ctx->p_code_compiled = wasm_runtime_malloc(ctx->code_compiled_size)))
+        return false;
+    memset(ctx->p_code_compiled, 0, ctx->code_compiled_size);
+    ctx->p_code_compiled_end = ctx->p_code_compiled +
+                                    ctx->code_compiled_size;
+
+    /* clean up frame ref */
+    memset(ctx->frame_ref_bottom, 0, ctx->frame_ref_size);
+    ctx->frame_ref = ctx->frame_ref_bottom;
+    ctx->stack_cell_num = 0;
+
+    /* clean up frame csp */
+    memset(ctx->frame_csp_bottom, 0, ctx->frame_csp_size);
+    ctx->frame_csp = ctx->frame_csp_bottom;
+    ctx->csp_num = 0;
+    ctx->max_csp_num = 0;
+
+    /* clean up frame offset */
+    memset(ctx->frame_offset_bottom, 0, ctx->frame_offset_size);
+    ctx->frame_offset = ctx->frame_offset_bottom;
+    ctx->dynamic_offset = ctx->start_dynamic_offset;
+
+    /* init preserved local offsets */
+    ctx->preserved_local_offset = ctx->max_dynamic_offset;
+
+    /* const buf is reserved */
+    return true;
+}
+
+static void
+wasm_loader_emit_int16(WASMLoaderContext *ctx, int16 value)
+{
+    if (ctx->p_code_compiled) {
+        *(int16*)(ctx->p_code_compiled) = value;
+        ctx->p_code_compiled += sizeof(int16);
+    }
+    else
+        ctx->code_compiled_size += sizeof(int16);
+}
+
+static void
+wasm_loader_emit_uint8(WASMLoaderContext *ctx, uint8 value)
+{
+    if (ctx->p_code_compiled) {
+        *(ctx->p_code_compiled) = value;
+        ctx->p_code_compiled += sizeof(uint8);
+    }
+    else
+        ctx->code_compiled_size += sizeof(uint8);
+}
+
+static void
+wasm_loader_emit_ptr(WASMLoaderContext *ctx, void *value)
+{
+    if (ctx->p_code_compiled) {
+        *(uint8**)(ctx->p_code_compiled) = value;
+        ctx->p_code_compiled += sizeof(void *);
+    }
+    else
+        ctx->code_compiled_size += sizeof(void *);
+}
+
+static void
+wasm_loader_emit_backspace(WASMLoaderContext *ctx, uint32 size)
+{
+    if (ctx->p_code_compiled) {
+        ctx->p_code_compiled -= size;
+    }
+    else
+        ctx->code_compiled_size -= size;
+}
+
+static void
+wasm_loader_emit_leb(WASMLoaderContext *ctx, uint8* start, uint8* end)
+{
+    if (ctx->p_code_compiled) {
+        bh_memcpy_s(ctx->p_code_compiled,
+                    ctx->p_code_compiled_end - ctx->p_code_compiled,
+                    start, end - start);
+        ctx->p_code_compiled += (end - start);
+    }
+    else {
+        ctx->code_compiled_size += (end - start);
+    }
+
+}
+
+static bool
+preserve_referenced_local(WASMLoaderContext *loader_ctx, uint8 opcode,
+                          uint32 local_index, uint32 local_type, bool *preserved,
+                          char *error_buf, uint32 error_buf_size)
+{
+    int16 preserved_offset = (int16)local_index;
+    *preserved = false;
+    for (uint32 i = 0; i < loader_ctx->stack_cell_num; i++) {
+        /* move previous local into dynamic space before a set/tee_local opcode */
+        if (loader_ctx->frame_offset_bottom[i] == (int16)local_index) {
+            if (preserved_offset == (int16)local_index) {
+                *preserved = true;
+                skip_label();
+                if (local_type == VALUE_TYPE_I32
+                    || local_type == VALUE_TYPE_F32) {
+                    preserved_offset = loader_ctx->preserved_local_offset;
+                    /* Only increase preserve offset in the second traversal */
+                    if (loader_ctx->p_code_compiled)
+                        loader_ctx->preserved_local_offset++;
+                    emit_label(EXT_OP_COPY_STACK_TOP);
+                }
+                else {
+                    preserved_offset = loader_ctx->preserved_local_offset;
+                    if (loader_ctx->p_code_compiled)
+                        loader_ctx->preserved_local_offset += 2;
+                    emit_label(EXT_OP_COPY_STACK_TOP_I64);
+                }
+                emit_operand(loader_ctx, local_index);
+                emit_operand(loader_ctx, preserved_offset);
+                emit_label(opcode);
+            }
+            loader_ctx->frame_offset_bottom[i] = preserved_offset;
+        }
+    }
+
+    return true;
+
+#if WASM_ENABLE_ABS_LABEL_ADDR == 0
+fail:
+    return false;
+#endif
+}
+
+static bool
+add_label_patch_to_list(BranchBlock *frame_csp,
+                        uint8 patch_type, uint8 *p_code_compiled,
+                        char *error_buf, uint32 error_buf_size)
+{
+    BranchBlockPatch *patch = loader_malloc
+        (sizeof(BranchBlockPatch), error_buf, error_buf_size);
+    if (!patch) {
+        return false;
+    }
+    patch->patch_type = patch_type;
+    patch->code_compiled = p_code_compiled;
+    if (!frame_csp->patch_list) {
+        frame_csp->patch_list = patch;
+        patch->next = NULL;
+    }
+    else {
+        patch->next = frame_csp->patch_list;
+        frame_csp->patch_list = patch;
+    }
+    return true;
+}
+
+static void
+apply_label_patch(WASMLoaderContext *ctx, uint8 depth,
+                  uint8 patch_type)
+{
+    BranchBlock *frame_csp = ctx->frame_csp - depth;
+    BranchBlockPatch *node = frame_csp->patch_list;
+    BranchBlockPatch *node_prev = NULL, *node_next;
+
+    if (!ctx->p_code_compiled)
+        return;
+
+    while (node) {
+        node_next = node->next;
+        if (node->patch_type == patch_type) {
+            *((uint8**)node->code_compiled) = ctx->p_code_compiled;
+            if (node_prev == NULL) {
+                frame_csp->patch_list = node_next;
+            }
+            else {
+                node_prev->next = node_next;
+            }
+            wasm_runtime_free(node);
+        }
+        else {
+            node_prev = node;
+        }
+        node = node_next;
+    }
+}
+
+static bool
+wasm_loader_emit_br_info(WASMLoaderContext *ctx, BranchBlock *frame_csp,
+                         char *error_buf, uint32 error_buf_size)
+{
+    emit_operand(ctx, frame_csp->dynamic_offset);
+    if (frame_csp->block_type == BLOCK_TYPE_LOOP ||
+        frame_csp->return_type == VALUE_TYPE_VOID) {
+        emit_byte(ctx, 0);
+        emit_operand(ctx, 0);
+    }
+    else if (frame_csp->return_type == VALUE_TYPE_I32
+        || frame_csp->return_type == VALUE_TYPE_F32) {
+        emit_byte(ctx, 1);
+        emit_operand(ctx, *(int16*)(ctx->frame_offset - 1));
+    }
+    else if (frame_csp->return_type == VALUE_TYPE_I64
+             || frame_csp->return_type == VALUE_TYPE_F64) {
+        emit_byte(ctx, 2);
+        emit_operand(ctx, *(int16*)(ctx->frame_offset - 2));
+    }
+
+    if (frame_csp->block_type == BLOCK_TYPE_LOOP) {
+        wasm_loader_emit_ptr(ctx, frame_csp->code_compiled);
+    }
+    else {
+        if (!add_label_patch_to_list(frame_csp, PATCH_END,
+                                     ctx->p_code_compiled,
+                                     error_buf, error_buf_size))
+            return false;
+        /* label address, to be patched */
+        wasm_loader_emit_ptr(ctx, NULL);
+    }
+    return true;
+}
+
+static bool
+wasm_loader_push_frame_offset(WASMLoaderContext *ctx, uint8 type,
+                              bool disable_emit, int16 operand_offset,
+                              char *error_buf, uint32 error_buf_size)
+{
+    if (type == VALUE_TYPE_VOID)
+        return true;
+
+    // only check memory overflow in first traverse
+    if (ctx->p_code_compiled == NULL) {
+        if (!check_offset_push(ctx, error_buf, error_buf_size))
+            return false;
+    }
+
+    if (disable_emit)
+        *(ctx->frame_offset)++ = operand_offset;
+    else {
+        emit_operand(ctx, ctx->dynamic_offset);
+        *(ctx->frame_offset)++ = ctx->dynamic_offset;
+        ctx->dynamic_offset++;
+        if (ctx->dynamic_offset > ctx->max_dynamic_offset)
+            ctx->max_dynamic_offset = ctx->dynamic_offset;
+    }
+
+    if (type == VALUE_TYPE_I32 || type == VALUE_TYPE_F32)
+        return true;
+
+    if (ctx->p_code_compiled == NULL) {
+        if (!check_offset_push(ctx, error_buf, error_buf_size))
+            return false;
+    }
+
+    ctx->frame_offset++;
+    if (!disable_emit) {
+        ctx->dynamic_offset++;
+        if (ctx->dynamic_offset > ctx->max_dynamic_offset)
+            ctx->max_dynamic_offset = ctx->dynamic_offset;
+    }
+    return true;
+}
+
+/* This function should be in front of wasm_loader_pop_frame_ref
+    as they both use ctx->stack_cell_num, and ctx->stack_cell_num
+    will be modified by wasm_loader_pop_frame_ref */
+static bool
+wasm_loader_pop_frame_offset(WASMLoaderContext *ctx, uint8 type,
+                             char *error_buf, uint32 error_buf_size)
+{
+    /* if ctx->frame_csp equals ctx->frame_csp_bottom,
+        then current block is the function block */
+    uint32 depth = ctx->frame_csp > ctx->frame_csp_bottom ? 1 : 0;
+    BranchBlock *cur_block = ctx->frame_csp - depth;
+    int32 available_stack_cell = (int32)
+        (ctx->stack_cell_num - cur_block->stack_cell_num);
+
+    /* Directly return success if current block is in stack
+     * polymorphic state while stack is empty. */
+    if (available_stack_cell <= 0 && cur_block->is_stack_polymorphic)
+        return true;
+
+    if (type == VALUE_TYPE_VOID)
+        return true;
+
+    if (type == VALUE_TYPE_I32 || type == VALUE_TYPE_F32) {
+        /* Check the offset stack bottom to ensure the frame offset
+            stack will not go underflow. But we don't thrown error
+            and return true here, because the error msg should be
+            given in wasm_loader_pop_frame_ref */
+        if (!check_offset_pop(ctx, 1))
+            return true;
+
+        ctx->frame_offset -= 1;
+        if ((*(ctx->frame_offset) > ctx->start_dynamic_offset)
+            && (*(ctx->frame_offset) < ctx->max_dynamic_offset))
+            ctx->dynamic_offset -= 1;
+    }
+    else {
+        if (!check_offset_pop(ctx, 2))
+            return true;
+
+        ctx->frame_offset -= 2;
+        if ((*(ctx->frame_offset) > ctx->start_dynamic_offset)
+            && (*(ctx->frame_offset) < ctx->max_dynamic_offset))
+            ctx->dynamic_offset -= 2;
+    }
+    emit_operand(ctx, *(ctx->frame_offset));
+    return true;
+}
+
+static bool
+wasm_loader_push_pop_frame_offset(WASMLoaderContext *ctx, uint8 pop_cnt,
+                                  uint8 type_push, uint8 type_pop,
+                                  bool disable_emit, int16 operand_offset,
+                                  char *error_buf, uint32 error_buf_size)
+{
+    for (int i = 0; i < pop_cnt; i++) {
+        if (!wasm_loader_pop_frame_offset(ctx, type_pop, error_buf, error_buf_size))
+            return false;
+    }
+    if (!wasm_loader_push_frame_offset(ctx, type_push,
+                                       disable_emit, operand_offset,
+                                       error_buf, error_buf_size))
+        return false;
+
+    return true;
+}
+
+static bool
+wasm_loader_push_frame_ref_offset(WASMLoaderContext *ctx, uint8 type,
+                                  bool disable_emit, int16 operand_offset,
+                                  char *error_buf, uint32 error_buf_size)
+{
+    if (!(wasm_loader_push_frame_offset(ctx, type, disable_emit, operand_offset,
+                                        error_buf, error_buf_size)))
+        return false;
+    if (!(wasm_loader_push_frame_ref(ctx, type, error_buf, error_buf_size)))
+        return false;
+
+    return true;
+}
+
+static bool
+wasm_loader_pop_frame_ref_offset(WASMLoaderContext *ctx, uint8 type,
+                                 char *error_buf, uint32 error_buf_size)
+{
+    /* put wasm_loader_pop_frame_offset in front of wasm_loader_pop_frame_ref */
+    if (!wasm_loader_pop_frame_offset(ctx, type, error_buf, error_buf_size))
+        return false;
+    if (!wasm_loader_pop_frame_ref(ctx, type, error_buf, error_buf_size))
+        return false;
+
+    return true;
+}
+
+static bool
+wasm_loader_push_pop_frame_ref_offset(WASMLoaderContext *ctx, uint8 pop_cnt,
+                                      uint8 type_push, uint8 type_pop,
+                                      bool disable_emit, int16 operand_offset,
+                                      char *error_buf, uint32 error_buf_size)
+{
+    if (!wasm_loader_push_pop_frame_offset(ctx, pop_cnt, type_push, type_pop,
+                                           disable_emit, operand_offset,
+                                           error_buf, error_buf_size))
+        return false;
+    if (!wasm_loader_push_pop_frame_ref(ctx, pop_cnt, type_push, type_pop,
+                                        error_buf, error_buf_size))
+        return false;
+
+    return true;
+}
+
+static bool
+wasm_loader_get_const_offset(WASMLoaderContext *ctx, uint8 type,
+                             void *value, int16 *offset,
+                             char *error_buf, uint32 error_buf_size)
+{
+    int16 operand_offset = 0;
+    Const *c;
+    for (c = (Const *)ctx->const_buf;
+         (uint8*)c < ctx->const_buf + ctx->num_const * sizeof(Const); c ++) {
+        if ((type == c->value_type)
+            && ((type == VALUE_TYPE_I64 && *(int64*)value == c->value.i64)
+            || (type == VALUE_TYPE_I32 && *(int32*)value == c->value.i32)
+            || (type == VALUE_TYPE_F64
+                && (0 == memcmp(value, &(c->value.f64), sizeof(float64))))
+            || (type == VALUE_TYPE_F32
+                && (0 == memcmp(value, &(c->value.f32), sizeof(float32)))))) {
+            operand_offset = c->slot_index;
+            break;
+        }
+        if (c->value_type == VALUE_TYPE_I64
+            || c->value_type == VALUE_TYPE_F64)
+            operand_offset += 2;
+        else
+            operand_offset += 1;
+    }
+    if ((uint8 *)c == ctx->const_buf + ctx->num_const * sizeof(Const)) {
+        if ((uint8 *)c == ctx->const_buf + ctx->const_buf_size) {
+            MEM_REALLOC(ctx->const_buf,
+                        ctx->const_buf_size,
+                        ctx->const_buf_size + 4 * sizeof(Const));
+            ctx->const_buf_size += 4 * sizeof(Const);
+            c = (Const *)(ctx->const_buf + ctx->num_const * sizeof(Const));
+        }
+        c->value_type = type;
+        switch (type) {
+        case VALUE_TYPE_F64:
+            bh_memcpy_s(&(c->value.f64), sizeof(WASMValue), value, sizeof(float64));
+            ctx->const_cell_num += 2;
+            /* The const buf will be reversed, we use the second cell */
+            /* of the i64/f64 const so the finnal offset is corrent */
+            operand_offset ++;
+            break;
+        case VALUE_TYPE_I64:
+            c->value.i64 = *(int64*)value;
+            ctx->const_cell_num += 2;
+            operand_offset ++;
+            break;
+        case VALUE_TYPE_F32:
+            bh_memcpy_s(&(c->value.f32), sizeof(WASMValue), value, sizeof(float32));
+            ctx->const_cell_num ++;
+            break;
+        case VALUE_TYPE_I32:
+            c->value.i32 = *(int32*)value;
+            ctx->const_cell_num ++;
+            break;
+        default:
+            break;
+        }
+        c->slot_index = operand_offset;
+        ctx->num_const ++;
+        LOG_OP("#### new const [%d]: %ld\n",
+            ctx->num_const, (int64)c->value.i64);
+    }
+    /* use negetive index for const */
+    operand_offset = -(operand_offset + 1);
+    *offset = operand_offset;
+    return true;
+fail:
+    return false;
+}
+
+/*
+    PUSH(POP)_XXX = push(pop) frame_ref + push(pop) frame_offset
+    -- Mostly used for the binary / compare operation
+    PUSH(POP)_OFFSET_TYPE only push(pop) the frame_offset stack
+    -- Mostly used in block / control instructions
+
+    The POP will always emit the offset on the top of the frame_offset stack
+    PUSH can be used in two ways:
+    1. directly PUSH:
+            PUSH_XXX();
+        will allocate a dynamic space and emit
+    2. silent PUSH:
+            operand_offset = xxx; disable_emit = true;
+            PUSH_XXX();
+        only push the frame_offset stack, no emit
+*/
+#define PUSH_I32() do {                                                 \
+    if (!wasm_loader_push_frame_ref_offset(loader_ctx, VALUE_TYPE_I32,  \
+                                           disable_emit, operand_offset,\
+                                           error_buf, error_buf_size))  \
+        goto fail;                                                      \
+  } while (0)
+
+#define PUSH_F32() do {                                                 \
+    if (!wasm_loader_push_frame_ref_offset(loader_ctx, VALUE_TYPE_F32,  \
+                                           disable_emit, operand_offset,\
+                                           error_buf, error_buf_size))  \
+        goto fail;                                                      \
+  } while (0)
+
+#define PUSH_I64() do {                                                 \
+    if (!wasm_loader_push_frame_ref_offset(loader_ctx, VALUE_TYPE_I64,  \
+                                           disable_emit, operand_offset,\
+                                           error_buf, error_buf_size))  \
+        goto fail;                                                      \
+  } while (0)
+
+#define PUSH_F64() do {                                                 \
+    if (!wasm_loader_push_frame_ref_offset(loader_ctx, VALUE_TYPE_F64,  \
+                                           disable_emit, operand_offset,\
+                                           error_buf, error_buf_size))  \
+        goto fail;                                                      \
+  } while (0)
+
+#define POP_I32() do {                                                  \
+    if (!wasm_loader_pop_frame_ref_offset(loader_ctx, VALUE_TYPE_I32,   \
+                                          error_buf, error_buf_size))   \
+        goto fail;                                                      \
+  } while (0)
+
+#define POP_F32() do {                                                  \
+    if (!wasm_loader_pop_frame_ref_offset(loader_ctx, VALUE_TYPE_F32,   \
+                                          error_buf, error_buf_size))   \
+        goto fail;                                                      \
+  } while (0)
+
+#define POP_I64() do {                                                  \
+    if (!wasm_loader_pop_frame_ref_offset(loader_ctx, VALUE_TYPE_I64,   \
+                                          error_buf, error_buf_size))   \
+        goto fail;                                                      \
+  } while (0)
+
+#define POP_F64() do {                                                  \
+    if (!wasm_loader_pop_frame_ref_offset(loader_ctx, VALUE_TYPE_F64,   \
+                                          error_buf, error_buf_size))   \
+        goto fail;                                                      \
+  } while (0)
+
+#define PUSH_OFFSET_TYPE(type) do {                                     \
+    if (!(wasm_loader_push_frame_offset(loader_ctx, type,               \
+                                        disable_emit, operand_offset,   \
+                                        error_buf, error_buf_size)))    \
+        goto fail;                                                      \
+  } while (0)
+
+#define POP_OFFSET_TYPE(type) do {                                      \
+    if (!(wasm_loader_pop_frame_offset(loader_ctx, type,                \
+                                       error_buf, error_buf_size)))     \
+        goto fail;                                                      \
+  } while (0)
+
+#define POP_AND_PUSH(type_pop, type_push) do {                                \
+    if (!(wasm_loader_push_pop_frame_ref_offset(loader_ctx, 1,                \
+                                                type_push, type_pop,          \
+                                                disable_emit, operand_offset, \
+                                                error_buf, error_buf_size)))  \
+        goto fail;                                                            \
+  } while (0)
+
+/* type of POPs should be the same */
+#define POP2_AND_PUSH(type_pop, type_push) do {                               \
+    if (!(wasm_loader_push_pop_frame_ref_offset(loader_ctx, 2,                \
+                                                type_push, type_pop,          \
+                                                disable_emit, operand_offset, \
+                                                error_buf, error_buf_size)))  \
+        goto fail;                                                            \
+  } while (0)
+
+#else /* WASM_ENABLE_FAST_INTERP */
+
+#define PUSH_I32() do {                                             \
+    if (!(wasm_loader_push_frame_ref(loader_ctx, VALUE_TYPE_I32,    \
+                                     error_buf, error_buf_size)))   \
+        goto fail;                                                  \
+  } while (0)
+
+#define PUSH_F32() do {                                             \
+    if (!(wasm_loader_push_frame_ref(loader_ctx, VALUE_TYPE_F32,    \
+                                     error_buf, error_buf_size)))   \
+        goto fail;                                                  \
+  } while (0)
+
+#define PUSH_I64() do {                                             \
+    if (!(wasm_loader_push_frame_ref(loader_ctx, VALUE_TYPE_I64,    \
+                                     error_buf, error_buf_size)))   \
+        goto fail;                                                  \
+  } while (0)
+
+#define PUSH_F64() do {                                             \
+    if (!(wasm_loader_push_frame_ref(loader_ctx, VALUE_TYPE_F64,    \
+                                     error_buf, error_buf_size)))   \
+        goto fail;                                                  \
+  } while (0)
+
+#define POP_I32() do {                                              \
+    if (!(wasm_loader_pop_frame_ref(loader_ctx, VALUE_TYPE_I32,     \
+                                    error_buf, error_buf_size)))    \
+        goto fail;                                                  \
+  } while (0)
+
+#define POP_F32() do {                                              \
+    if (!(wasm_loader_pop_frame_ref(loader_ctx, VALUE_TYPE_F32,     \
+                                    error_buf, error_buf_size)))    \
+        goto fail;                                                  \
+  } while (0)
+
+#define POP_I64() do {                                              \
+    if (!(wasm_loader_pop_frame_ref(loader_ctx, VALUE_TYPE_I64,     \
+                                    error_buf, error_buf_size)))    \
+        goto fail;                                                  \
+  } while (0)
+
+#define POP_F64() do {                                              \
+    if (!(wasm_loader_pop_frame_ref(loader_ctx, VALUE_TYPE_F64,     \
+                                    error_buf, error_buf_size)))    \
+        goto fail;                                                  \
+  } while (0)
+
+#define POP_AND_PUSH(type_pop, type_push) do {                           \
+    if (!(wasm_loader_push_pop_frame_ref(loader_ctx, 1,                  \
+                                         type_push, type_pop,            \
+                                         error_buf, error_buf_size)))    \
+        goto fail;                                                       \
+  } while (0)
+
+/* type of POPs should be the same */
+#define POP2_AND_PUSH(type_pop, type_push) do {                          \
+    if (!(wasm_loader_push_pop_frame_ref(loader_ctx, 2,                  \
+                                         type_push, type_pop,            \
+                                         error_buf, error_buf_size)))    \
+        goto fail;                                                       \
+  } while (0)
+#endif /* WASM_ENABLE_FAST_INTERP */
+
+#if WASM_ENABLE_FAST_INTERP != 0
+
+static bool
+reserve_block_ret(WASMLoaderContext *loader_ctx, uint8 opcode, bool disable_emit,
+                  char *error_buf, uint32 error_buf_size)
+{
+    int16 operand_offset = 0;
+    uint8 block_depth = 0;
+    if (opcode == WASM_OP_ELSE)
+        block_depth = 1;
+    else
+        block_depth = 0;
+
+    if ((loader_ctx->frame_csp - block_depth)->return_type != VALUE_TYPE_VOID) {
+        uint8 return_cells;
+        if ((loader_ctx->frame_csp - block_depth)->return_type == VALUE_TYPE_I32
+            || (loader_ctx->frame_csp - block_depth)->return_type == VALUE_TYPE_F32)
+            return_cells = 1;
+        else
+            return_cells = 2;
+        if ((loader_ctx->frame_csp - block_depth)->dynamic_offset !=
+                *(loader_ctx->frame_offset - return_cells)) {
+
+            /* insert op_copy before else opcode */
+            if (opcode == WASM_OP_ELSE)
+                skip_label();
+
+            if (return_cells == 1)
+                emit_label(EXT_OP_COPY_STACK_TOP);
+            else
+                emit_label(EXT_OP_COPY_STACK_TOP_I64);
+            emit_operand(loader_ctx, *(loader_ctx->frame_offset - return_cells));
+            emit_operand(loader_ctx, (loader_ctx->frame_csp - block_depth)->dynamic_offset);
+
+            if (opcode == WASM_OP_ELSE) {
+                *(loader_ctx->frame_offset - return_cells) =
+                    (loader_ctx->frame_csp - block_depth)->dynamic_offset;
+            }
+            else {
+                loader_ctx->frame_offset -= return_cells;
+                loader_ctx->dynamic_offset = loader_ctx->frame_csp->dynamic_offset;
+                PUSH_OFFSET_TYPE((loader_ctx->frame_csp - block_depth)->return_type);
+                wasm_loader_emit_backspace(loader_ctx, sizeof(int16));
+            }
+            if (opcode == WASM_OP_ELSE)
+                emit_label(opcode);
+        }
+    }
+
+    return true;
+
+fail:
+    return false;
+}
+
+#endif /* WASM_ENABLE_FAST_INTERP */
+
+#define RESERVE_BLOCK_RET() do {                                    \
+     if (!reserve_block_ret(loader_ctx, opcode, disable_emit,       \
+                            error_buf, error_buf_size))             \
+        goto fail;                                                  \
+  } while (0)
+
+#define PUSH_TYPE(type) do {                                        \
+    if (!(wasm_loader_push_frame_ref(loader_ctx, type,              \
+                                     error_buf, error_buf_size)))   \
+        goto fail;                                                  \
+  } while (0)
+
+#define POP_TYPE(type) do {                                         \
+    if (!(wasm_loader_pop_frame_ref(loader_ctx, type,               \
+                                    error_buf, error_buf_size)))    \
+        goto fail;                                                  \
+  } while (0)
+
+#define PUSH_CSP(type, ret_type, _start_addr) do {              \
+    if (!wasm_loader_push_frame_csp(loader_ctx, type, ret_type, \
+                                    _start_addr, error_buf,     \
+                                    error_buf_size))            \
+        goto fail;                                              \
+  } while (0)
+
+#define POP_CSP() do {                                          \
+    if (!wasm_loader_pop_frame_csp(loader_ctx,                  \
+                                   error_buf, error_buf_size))  \
+        goto fail;                                              \
+  } while (0)
+
+#define GET_LOCAL_INDEX_TYPE_AND_OFFSET() do {      \
+    read_leb_uint32(p, p_end, local_idx);           \
+    bh_assert(local_idx < param_count + local_count);\
+    local_type = local_idx < param_count            \
+        ? param_types[local_idx]                    \
+        : local_types[local_idx - param_count];     \
+    local_offset = local_offsets[local_idx];        \
+  } while (0)
+
+#define CHECK_BR(depth) do {                                        \
+    if (!wasm_loader_check_br(loader_ctx, depth,                    \
+                              error_buf, error_buf_size))           \
+        goto fail;                                                  \
+  } while (0)
+
+#define CHECK_MEMORY() do {                 \
+    bh_assert(module->import_memory_count   \
+              + module->memory_count > 0);  \
+  } while (0)
+
+#define CHECK_BLOCK_TYPE(type) do {         \
+    bh_assert(type == VALUE_TYPE_I32        \
+              || type == VALUE_TYPE_I64     \
+              || type == VALUE_TYPE_F32     \
+              || type == VALUE_TYPE_F64     \
+              || type == VALUE_TYPE_VOID);  \
+  } while (0)
+
+static BranchBlock *
+check_branch_block(WASMLoaderContext *loader_ctx,
+                   uint8 **p_buf, uint8 *buf_end,
+                   char *error_buf, uint32 error_buf_size)
+{
+    uint8 *p = *p_buf, *p_end = buf_end;
+    BranchBlock *frame_csp_tmp;
+    uint32 depth;
+
+    read_leb_uint32(p, p_end, depth);
+    CHECK_BR(depth);
+    frame_csp_tmp = loader_ctx->frame_csp - depth - 1;
+#if WASM_ENABLE_FAST_INTERP != 0
+    emit_br_info(frame_csp_tmp);
+#endif
+
+    *p_buf = p;
+    return frame_csp_tmp;
+fail:
+    return NULL;
+}
+
+static bool
+check_branch_block_ret(WASMLoaderContext *loader_ctx,
+                       BranchBlock *frame_csp_tmp,
+                       char *error_buf, uint32 error_buf_size)
+{
+#if WASM_ENABLE_FAST_INTERP != 0
+    BranchBlock *cur_block = loader_ctx->frame_csp - 1;
+    bool disable_emit = true;
+    int16 operand_offset = 0;
+#endif
+    if (frame_csp_tmp->block_type != BLOCK_TYPE_LOOP) {
+        uint8 block_return_type = frame_csp_tmp->return_type;
+#if WASM_ENABLE_FAST_INTERP != 0
+        /* If the stack is in polymorphic state, do fake pop and push on
+            offset stack to keep the depth of offset stack to be the same
+            with ref stack */
+        if (cur_block->is_stack_polymorphic) {
+            POP_OFFSET_TYPE(block_return_type);
+            PUSH_OFFSET_TYPE(block_return_type);
+        }
+#endif
+        POP_TYPE(block_return_type);
+        PUSH_TYPE(block_return_type);
+    }
+    return true;
+fail:
+    return false;
+}
+
+static bool
+check_block_stack(WASMLoaderContext *ctx, BranchBlock *block,
+                  char *error_buf, uint32 error_buf_size)
+{
+    uint8 type = block->return_type;
+    int32 available_stack_cell = (int32)
+        (ctx->stack_cell_num - block->stack_cell_num);
+
+    if (type != VALUE_TYPE_VOID
+        && available_stack_cell <= 0
+        && block->is_stack_polymorphic) {
+        if (!(wasm_loader_push_frame_ref(ctx, type, error_buf, error_buf_size))
+#if WASM_ENABLE_FAST_INTERP != 0
+            || !(wasm_loader_push_frame_offset(ctx, type, true, 0, error_buf, error_buf_size))
+#endif
+            )
+            return false;
+        return true;
+    }
+
+    if (type != VALUE_TYPE_VOID
+        && available_stack_cell == 1
+        && *(ctx->frame_ref - 1) == VALUE_TYPE_ANY) {
+        if (type == VALUE_TYPE_I32 || type == VALUE_TYPE_F32) {
+            /* If the stack top is a value of any type, change its type to the
+             * same as block return type and return success */
+            *(ctx->frame_ref - 1) = type;
+        }
+        else {
+            if (!(wasm_loader_push_frame_ref(ctx, VALUE_TYPE_I32,
+                                             error_buf, error_buf_size))
+#if WASM_ENABLE_FAST_INTERP != 0
+                || !(wasm_loader_push_frame_offset(ctx, VALUE_TYPE_I32,
+                                                   true, 0,
+                                                   error_buf, error_buf_size))
+#endif
+                                                   )
+                return false;
+            *(ctx->frame_ref - 1) = *(ctx->frame_ref - 2) = type;
+        }
+        return true;
+    }
+
+    bh_assert(!(((type == VALUE_TYPE_I32 || type == VALUE_TYPE_F32)
+                 && available_stack_cell != 1)
+                || ((type == VALUE_TYPE_I64 || type == VALUE_TYPE_F64)
+                    && available_stack_cell != 2)
+                || (type == VALUE_TYPE_VOID
+                    && available_stack_cell > 0)));
+
+    if (!check_stack_top_values(ctx->frame_ref, available_stack_cell,
+                                type, error_buf, error_buf_size))
+        return false;
+
+    return true;
+}
+
+/* reset the stack to the state of before entering the last block */
+#if WASM_ENABLE_FAST_INTERP != 0
+#define RESET_STACK() do {                                                   \
+    loader_ctx->stack_cell_num =                                             \
+               (loader_ctx->frame_csp - 1)->stack_cell_num;                  \
+    loader_ctx->frame_ref =                                                  \
+               loader_ctx->frame_ref_bottom + loader_ctx->stack_cell_num;    \
+    loader_ctx->frame_offset =                                               \
+               loader_ctx->frame_offset_bottom + loader_ctx->stack_cell_num; \
+} while (0)
+#else
+#define RESET_STACK() do {                                                \
+    loader_ctx->stack_cell_num =                                          \
+               (loader_ctx->frame_csp - 1)->stack_cell_num;               \
+    loader_ctx->frame_ref =                                               \
+               loader_ctx->frame_ref_bottom + loader_ctx->stack_cell_num; \
+} while (0)
+#endif
+
+/* set current block's stack polymorphic state */
+#define SET_CUR_BLOCK_STACK_POLYMORPHIC_STATE(flag) do {                  \
+    BranchBlock *cur_block = loader_ctx->frame_csp - 1;                   \
+    cur_block->is_stack_polymorphic = flag;                               \
+} while (0)
+
+static bool
+wasm_loader_prepare_bytecode(WASMModule *module, WASMFunction *func,
+                             BlockAddr *block_addr_cache,
+                             char *error_buf, uint32 error_buf_size)
+{
+    uint8 *p = func->code, *p_end = func->code + func->code_size, *p_org;
+    uint32 param_count, local_count, global_count;
+    uint8 *param_types, ret_type, *local_types, local_type, global_type;
+    uint16 *local_offsets, local_offset;
+    uint32 count, i, local_idx, global_idx, u32, align, mem_offset;
+    int32 i32, i32_const = 0;
+    int64 i64;
+    uint8 opcode, u8, block_return_type;
+    bool return_value = false;
+    WASMLoaderContext *loader_ctx;
+    BranchBlock *frame_csp_tmp;
+#if WASM_ENABLE_BULK_MEMORY != 0
+    uint32 segment_index;
+#endif
+#if WASM_ENABLE_FAST_INTERP != 0
+    uint8 *func_const_end, *func_const;
+    int16 operand_offset;
+    uint8 last_op = 0;
+    bool disable_emit, preserve_local = false;
+    float32 f32;
+    float64 f64;
+
+    LOG_OP("\nProcessing func | [%d] params | [%d] locals | [%d] return\n",
+        func->param_cell_num,
+        func->local_cell_num,
+        func->ret_cell_num);
+#endif
+
+    global_count = module->import_global_count + module->global_count;
+
+    param_count = func->func_type->param_count;
+    param_types = func->func_type->types;
+    ret_type = func->func_type->result_count
+               ? param_types[param_count] : VALUE_TYPE_VOID;
+
+    local_count = func->local_count;
+    local_types = func->local_types;
+    local_offsets = func->local_offsets;
+
+    if (!(loader_ctx = wasm_loader_ctx_init(func))) {
+        set_error_buf(error_buf, error_buf_size,
+                      "WASM loader prepare bytecode failed: "
+                      "allocate memory failed");
+        goto fail;
+    }
+
+#if WASM_ENABLE_FAST_INTERP != 0
+re_scan:
+    if (loader_ctx->code_compiled_size > 0) {
+        if (!wasm_loader_ctx_reinit(loader_ctx)) {
+            set_error_buf(error_buf, error_buf_size,
+                          "WASM loader prepare bytecode failed: "
+                          "allocate memory failed");
+            goto fail;
+        }
+        p = func->code;
+        func->code_compiled = loader_ctx->p_code_compiled;
+    }
+#endif
+
+    PUSH_CSP(BLOCK_TYPE_FUNCTION, ret_type, p);
+
+    while (p < p_end) {
+        opcode = *p++;
+#if WASM_ENABLE_FAST_INTERP != 0
+        p_org = p;
+        disable_emit = false;
+        emit_label(opcode);
+#endif
+
+        switch (opcode) {
+            case WASM_OP_UNREACHABLE:
+                RESET_STACK();
+                SET_CUR_BLOCK_STACK_POLYMORPHIC_STATE(true);
+                break;
+
+            case WASM_OP_NOP:
+#if WASM_ENABLE_FAST_INTERP != 0
+                skip_label();
+#endif
+                break;
+
+            case WASM_OP_BLOCK:
+                /* 0x40/0x7F/0x7E/0x7D/0x7C */
+                block_return_type = read_uint8(p);
+                CHECK_BLOCK_TYPE(block_return_type);
+                PUSH_CSP(BLOCK_TYPE_BLOCK, block_return_type, p);
+#if WASM_ENABLE_FAST_INTERP != 0
+                skip_label();
+#endif
+                break;
+
+            case WASM_OP_LOOP:
+                /* 0x40/0x7F/0x7E/0x7D/0x7C */
+                block_return_type = read_uint8(p);
+                CHECK_BLOCK_TYPE(block_return_type);
+                PUSH_CSP(BLOCK_TYPE_LOOP, block_return_type, p);
+#if WASM_ENABLE_FAST_INTERP != 0
+                skip_label();
+                (loader_ctx->frame_csp - 1)->code_compiled =
+                    loader_ctx->p_code_compiled;
+#endif
+                break;
+
+            case WASM_OP_IF:
+                POP_I32();
+                /* 0x40/0x7F/0x7E/0x7D/0x7C */
+                block_return_type = read_uint8(p);
+                CHECK_BLOCK_TYPE(block_return_type);
+                PUSH_CSP(BLOCK_TYPE_IF, block_return_type, p);
+#if WASM_ENABLE_FAST_INTERP != 0
+                emit_empty_label_addr_and_frame_ip(PATCH_ELSE);
+                emit_empty_label_addr_and_frame_ip(PATCH_END);
+#endif
+                break;
+
+            case WASM_OP_ELSE:
+                bh_assert(loader_ctx->csp_num >= 2
+                          && (loader_ctx->frame_csp - 1)->block_type
+                                == BLOCK_TYPE_IF);
+
+                /* check whether if branch's stack matches its result type */
+                if (!check_block_stack(loader_ctx, loader_ctx->frame_csp - 1,
+                                       error_buf, error_buf_size))
+                    goto fail;
+
+                (loader_ctx->frame_csp - 1)->else_addr = p - 1;
+
+#if WASM_ENABLE_FAST_INTERP != 0
+                /* if the result of if branch is in local or const area, add a copy op */
+                RESERVE_BLOCK_RET();
+
+                emit_empty_label_addr_and_frame_ip(PATCH_END);
+                apply_label_patch(loader_ctx, 1, PATCH_ELSE);
+#endif
+                RESET_STACK();
+                SET_CUR_BLOCK_STACK_POLYMORPHIC_STATE(false);
+                break;
+
+            case WASM_OP_END:
+            {
+
+                /* check whether block stack matches its result type */
+                if (!check_block_stack(loader_ctx, loader_ctx->frame_csp - 1,
+                                       error_buf, error_buf_size))
+                    goto fail;
+
+                /* if has return value, but no else branch, fail */
+                bh_assert(!((loader_ctx->frame_csp - 1)->block_type == BLOCK_TYPE_IF
+                            && (loader_ctx->frame_csp - 1)->return_type != VALUE_TYPE_VOID
+                            && !(loader_ctx->frame_csp - 1)->else_addr));
+
+                POP_CSP();
+
+#if WASM_ENABLE_FAST_INTERP != 0
+                skip_label();
+                /* copy the result to the block return address */
+                RESERVE_BLOCK_RET();
+
+                apply_label_patch(loader_ctx, 0, PATCH_END);
+                free_label_patch_list(loader_ctx->frame_csp);
+                if (loader_ctx->frame_csp->block_type == BLOCK_TYPE_FUNCTION) {
+                    emit_label(WASM_OP_RETURN);
+                    POP_OFFSET_TYPE(loader_ctx->frame_csp->return_type);
+                }
+#endif
+                if (loader_ctx->csp_num > 0) {
+                    loader_ctx->frame_csp->end_addr = p - 1;
+                }
+                else {
+                    /* end of function block, function will return,
+                       ignore the following bytecodes */
+                    p = p_end;
+
+                    continue;
+                }
+
+                SET_CUR_BLOCK_STACK_POLYMORPHIC_STATE(false);
+                break;
+            }
+
+            case WASM_OP_BR:
+            {
+                if (!(frame_csp_tmp = check_branch_block(loader_ctx, &p, p_end,
+                                                         error_buf, error_buf_size)))
+                    goto fail;
+
+                if (!check_branch_block_ret(loader_ctx, frame_csp_tmp,
+                                            error_buf, error_buf_size))
+                    goto fail;
+
+                RESET_STACK();
+                SET_CUR_BLOCK_STACK_POLYMORPHIC_STATE(true);
+                break;
+            }
+
+            case WASM_OP_BR_IF:
+            {
+                POP_I32();
+
+                if (!(frame_csp_tmp = check_branch_block(loader_ctx, &p, p_end,
+                                                         error_buf, error_buf_size)))
+                    goto fail;
+
+                if (!check_branch_block_ret(loader_ctx, frame_csp_tmp,
+                                            error_buf, error_buf_size))
+                    goto fail;
+
+                break;
+            }
+
+            case WASM_OP_BR_TABLE:
+            {
+                uint8 ret_type;
+
+                read_leb_uint32(p, p_end, count);
+#if WASM_ENABLE_FAST_INTERP != 0
+                emit_const(count);
+#endif
+                POP_I32();
+
+                /* TODO: check the const */
+                for (i = 0; i <= count; i++) {
+                    if (!(frame_csp_tmp = check_branch_block(loader_ctx, &p, p_end,
+                                                             error_buf, error_buf_size)))
+                        goto fail;
+
+                    if (!check_branch_block_ret(loader_ctx, frame_csp_tmp,
+                                                error_buf, error_buf_size))
+                        goto fail;
+
+                    if (i == 0) {
+                        ret_type = frame_csp_tmp->block_type == BLOCK_TYPE_LOOP ?
+                                   VALUE_TYPE_VOID : frame_csp_tmp->return_type;
+                    }
+                    else {
+                        /* Check whether all table items have the same return type */
+                        uint8 tmp_ret_type = frame_csp_tmp->block_type == BLOCK_TYPE_LOOP ?
+                                             VALUE_TYPE_VOID : frame_csp_tmp->return_type;
+                        bh_assert(ret_type == tmp_ret_type);
+                        (void)tmp_ret_type;
+                    }
+                }
+
+                RESET_STACK();
+                SET_CUR_BLOCK_STACK_POLYMORPHIC_STATE(true);
+                (void)ret_type;
+                break;
+            }
+
+            case WASM_OP_RETURN:
+            {
+                POP_TYPE(ret_type);
+                PUSH_TYPE(ret_type);
+
+#if WASM_ENABLE_FAST_INTERP != 0
+                // emit the offset after return opcode
+                POP_OFFSET_TYPE(ret_type);
+#endif
+
+                RESET_STACK();
+                SET_CUR_BLOCK_STACK_POLYMORPHIC_STATE(true);
+                break;
+            }
+
+            case WASM_OP_CALL:
+            {
+                WASMType *func_type;
+                uint32 func_idx;
+                int32 idx;
+
+                read_leb_uint32(p, p_end, func_idx);
+#if WASM_ENABLE_FAST_INTERP != 0
+                // we need to emit func_idx before arguments
+                emit_const(func_idx);
+#endif
+
+                bh_assert(func_idx < module->import_function_count
+                                     + module->function_count);
+
+                if (func_idx < module->import_function_count)
+                    func_type = module->import_functions[func_idx].u.function.func_type;
+                else
+                    func_type =
+                        module->functions[func_idx - module->import_function_count]->func_type;
+
+                if (func_type->param_count > 0) {
+                    for (idx = (int32)(func_type->param_count - 1); idx >= 0; idx--) {
+                        POP_TYPE(func_type->types[idx]);
+#if WASM_ENABLE_FAST_INTERP != 0
+                        POP_OFFSET_TYPE(func_type->types[idx]);
+#endif
+                    }
+                }
+
+                if (func_type->result_count) {
+                    PUSH_TYPE(func_type->types[func_type->param_count]);
+#if WASM_ENABLE_FAST_INTERP != 0
+                    PUSH_OFFSET_TYPE(func_type->types[func_type->param_count]);
+#endif
+                }
+
+                func->has_op_func_call = true;
+                break;
+            }
+
+            case WASM_OP_CALL_INDIRECT:
+            {
+                int32 idx;
+                WASMType *func_type;
+                uint32 type_idx;
+
+                bh_assert(module->import_table_count
+                          + module->table_count > 0);
+
+                read_leb_uint32(p, p_end, type_idx);
+#if WASM_ENABLE_FAST_INTERP != 0
+                // we need to emit func_idx before arguments
+                emit_const(type_idx);
+#endif
+
+                /* reserved byte 0x00 */
+                bh_assert(*p == 0x00);
+                p++;
+
+                POP_I32();
+
+                bh_assert(type_idx < module->type_count);
+
+                func_type = module->types[type_idx];
+
+                if (func_type->param_count > 0) {
+                    for (idx = (int32)(func_type->param_count - 1); idx >= 0; idx--) {
+                        POP_TYPE(func_type->types[idx]);
+#if WASM_ENABLE_FAST_INTERP != 0
+                        POP_OFFSET_TYPE(func_type->types[idx]);
+#endif
+                    }
+                }
+
+                if (func_type->result_count > 0) {
+                    PUSH_TYPE(func_type->types[func_type->param_count]);
+#if WASM_ENABLE_FAST_INTERP != 0
+                    PUSH_OFFSET_TYPE(func_type->types[func_type->param_count]);
+#endif
+                }
+
+                func->has_op_func_call = true;
+                break;
+            }
+
+            case WASM_OP_DROP:
+            case WASM_OP_DROP_64:
+            {
+                BranchBlock *cur_block = loader_ctx->frame_csp - 1;
+                int32 available_stack_cell = (int32)
+                    (loader_ctx->stack_cell_num - cur_block->stack_cell_num);
+
+                bh_assert(!(available_stack_cell <= 0
+                            && !cur_block->is_stack_polymorphic));
+
+                if (available_stack_cell > 0) {
+                    if (*(loader_ctx->frame_ref - 1) == REF_I32
+                        || *(loader_ctx->frame_ref - 1) == REF_F32) {
+                        loader_ctx->frame_ref--;
+                        loader_ctx->stack_cell_num--;
+#if WASM_ENABLE_FAST_INTERP != 0
+                        skip_label();
+                        loader_ctx->frame_offset--;
+                        if (*(loader_ctx->frame_offset) >
+                                loader_ctx->start_dynamic_offset)
+                            loader_ctx->dynamic_offset --;
+#endif
+                    }
+                    else {
+                        loader_ctx->frame_ref -= 2;
+                        loader_ctx->stack_cell_num -= 2;
+#if (WASM_ENABLE_FAST_INTERP == 0) || (WASM_ENABLE_JIT != 0)
+                        *(p - 1) = WASM_OP_DROP_64;
+#endif
+#if WASM_ENABLE_FAST_INTERP != 0
+                        skip_label();
+                        loader_ctx->frame_offset -= 2;
+                        if (*(loader_ctx->frame_offset) >
+                                loader_ctx->start_dynamic_offset)
+                            loader_ctx->dynamic_offset -= 2;
+#endif
+                    }
+                }
+                else {
+#if WASM_ENABLE_FAST_INTERP != 0
+                    skip_label();
+#endif
+                }
+                break;
+            }
+
+            case WASM_OP_SELECT:
+            case WASM_OP_SELECT_64:
+            {
+                uint8 ref_type;
+                BranchBlock *cur_block = loader_ctx->frame_csp - 1;
+                int32 available_stack_cell;
+
+                POP_I32();
+
+                available_stack_cell = (int32)
+                    (loader_ctx->stack_cell_num - cur_block->stack_cell_num);
+
+                bh_assert(!(available_stack_cell <= 0
+                            && !cur_block->is_stack_polymorphic));
+
+                if (available_stack_cell > 0) {
+                    switch (*(loader_ctx->frame_ref - 1)) {
+                        case REF_I32:
+                        case REF_F32:
+                            break;
+                        case REF_I64_2:
+                        case REF_F64_2:
+#if (WASM_ENABLE_FAST_INTERP == 0) || (WASM_ENABLE_JIT != 0)
+                            *(p - 1) = WASM_OP_SELECT_64;
+#endif
+#if WASM_ENABLE_FAST_INTERP != 0
+                            if (loader_ctx->p_code_compiled) {
+#if WASM_ENABLE_ABS_LABEL_ADDR != 0
+                                *(void**)(loader_ctx->p_code_compiled - 2 - sizeof(void*)) =
+                                    handle_table[WASM_OP_SELECT_64];
+#else
+                                *((int16*)loader_ctx->p_code_compiled - 2) = (int16)
+                                    (handle_table[WASM_OP_SELECT_64] - handle_table[0]);
+#endif
+                            }
+#endif
+                            break;
+                    }
+
+                    ref_type = *(loader_ctx->frame_ref - 1);
+#if WASM_ENABLE_FAST_INTERP != 0
+                    POP_OFFSET_TYPE(ref_type);
+#endif
+                    POP_TYPE(ref_type);
+#if WASM_ENABLE_FAST_INTERP != 0
+                    POP_OFFSET_TYPE(ref_type);
+#endif
+                    POP_TYPE(ref_type);
+#if WASM_ENABLE_FAST_INTERP != 0
+                    PUSH_OFFSET_TYPE(ref_type);
+#endif
+                    PUSH_TYPE(ref_type);
+                }
+                else {
+#if WASM_ENABLE_FAST_INTERP != 0
+                    PUSH_OFFSET_TYPE(VALUE_TYPE_ANY);
+#endif
+                    PUSH_TYPE(VALUE_TYPE_ANY);
+                }
+                break;
+            }
+
+            case WASM_OP_GET_LOCAL:
+            {
+                p_org = p - 1;
+                GET_LOCAL_INDEX_TYPE_AND_OFFSET();
+                PUSH_TYPE(local_type);
+
+#if WASM_ENABLE_FAST_INTERP != 0
+                /* Get Local is optimized out */
+                skip_label();
+                disable_emit = true;
+                operand_offset = local_offset;
+                PUSH_OFFSET_TYPE(local_type);
+#else
+#if (WASM_ENABLE_WAMR_COMPILER == 0) && (WASM_ENABLE_JIT == 0)
+                if (local_offset < 0x80) {
+                    *p_org++ = EXT_OP_GET_LOCAL_FAST;
+                    if (local_type == VALUE_TYPE_I32
+                        || local_type == VALUE_TYPE_F32)
+                        *p_org++ = (uint8)local_offset;
+                    else
+                        *p_org++ = (uint8)(local_offset | 0x80);
+                    while (p_org < p)
+                        *p_org++ = WASM_OP_NOP;
+                }
+#endif
+#endif
+                break;
+            }
+
+            case WASM_OP_SET_LOCAL:
+            {
+                p_org = p - 1;
+                GET_LOCAL_INDEX_TYPE_AND_OFFSET();
+                POP_TYPE(local_type);
+
+#if WASM_ENABLE_FAST_INTERP != 0
+                if (!(preserve_referenced_local(loader_ctx, opcode, local_offset,
+                                                local_type, &preserve_local,
+                                                error_buf, error_buf_size)))
+                    goto fail;
+
+                if (local_offset < 256) {
+                    skip_label();
+                    if ((!preserve_local) && (LAST_OP_OUTPUT_I32())) {
+                        if (loader_ctx->p_code_compiled)
+                            *(int16*)(loader_ctx->p_code_compiled - 2) = local_offset;
+                        loader_ctx->frame_offset --;
+                        loader_ctx->dynamic_offset --;
+                    }
+                    else if ((!preserve_local) && (LAST_OP_OUTPUT_I64())) {
+                        if (loader_ctx->p_code_compiled)
+                            *(int16*)(loader_ctx->p_code_compiled - 2) = local_offset;
+                        loader_ctx->frame_offset -= 2;
+                        loader_ctx->dynamic_offset -= 2;
+                    }
+                    else {
+                        if (local_type == VALUE_TYPE_I32
+                            || local_type == VALUE_TYPE_F32) {
+                            emit_label(EXT_OP_SET_LOCAL_FAST);
+                            emit_byte(loader_ctx, local_offset);
+                        }
+                        else {
+                            emit_label(EXT_OP_SET_LOCAL_FAST_I64);
+                            emit_byte(loader_ctx, local_offset);
+                        }
+                        POP_OFFSET_TYPE(local_type);
+                    }
+                }
+                else {   /* local index larger than 255, reserve leb */
+                    p_org ++;
+                    emit_leb();
+                    POP_OFFSET_TYPE(local_type);
+                }
+#else
+#if (WASM_ENABLE_WAMR_COMPILER == 0) && (WASM_ENABLE_JIT == 0)
+                if (local_offset < 0x80) {
+                    *p_org++ = EXT_OP_SET_LOCAL_FAST;
+                    if (local_type == VALUE_TYPE_I32
+                        || local_type == VALUE_TYPE_F32)
+                        *p_org++ = (uint8)local_offset;
+                    else
+                        *p_org++ = (uint8)(local_offset | 0x80);
+                    while (p_org < p)
+                        *p_org++ = WASM_OP_NOP;
+                }
+#endif
+#endif
+                break;
+            }
+
+            case WASM_OP_TEE_LOCAL:
+            {
+                p_org = p - 1;
+                GET_LOCAL_INDEX_TYPE_AND_OFFSET();
+#if WASM_ENABLE_FAST_INTERP != 0
+                /* If the stack is in polymorphic state, do fake pop and push on
+                    offset stack to keep the depth of offset stack to be the same
+                    with ref stack */
+                BranchBlock *cur_block = loader_ctx->frame_csp - 1;
+                if (cur_block->is_stack_polymorphic) {
+                    POP_OFFSET_TYPE(local_type);
+                    PUSH_OFFSET_TYPE(local_type);
+                }
+#endif
+                POP_TYPE(local_type);
+                PUSH_TYPE(local_type);
+
+#if WASM_ENABLE_FAST_INTERP != 0
+                if (!(preserve_referenced_local(loader_ctx, opcode, local_offset,
+                                                local_type, &preserve_local,
+                                                error_buf, error_buf_size)))
+                    goto fail;
+
+                if (local_offset < 256) {
+                    skip_label();
+                    if (local_type == VALUE_TYPE_I32
+                        || local_type == VALUE_TYPE_F32) {
+                        emit_label(EXT_OP_TEE_LOCAL_FAST);
+                        emit_byte(loader_ctx, local_offset);
+                    }
+                    else {
+                        emit_label(EXT_OP_TEE_LOCAL_FAST_I64);
+                        emit_byte(loader_ctx, local_offset);
+                    }
+                }
+                else {  /* local index larger than 255, reserve leb */
+                    p_org ++;
+                    emit_leb();
+                }
+                emit_operand(loader_ctx, *(loader_ctx->frame_offset -
+                        wasm_value_type_cell_num(local_type)));
+#else
+#if (WASM_ENABLE_WAMR_COMPILER == 0) && (WASM_ENABLE_JIT == 0)
+                if (local_offset < 0x80) {
+                    *p_org++ = EXT_OP_TEE_LOCAL_FAST;
+                    if (local_type == VALUE_TYPE_I32
+                        || local_type == VALUE_TYPE_F32)
+                        *p_org++ = (uint8)local_offset;
+                    else
+                        *p_org++ = (uint8)(local_offset | 0x80);
+                    while (p_org < p)
+                        *p_org++ = WASM_OP_NOP;
+                }
+#endif
+#endif
+                break;
+            }
+
+            case WASM_OP_GET_GLOBAL:
+            {
+                read_leb_uint32(p, p_end, global_idx);
+                bh_assert(global_idx < global_count);
+
+                global_type = global_idx < module->import_global_count
+                              ? module->import_globals[global_idx].u.global.type
+                              :module->globals[global_idx - module->import_global_count].type;
+
+                PUSH_TYPE(global_type);
+#if WASM_ENABLE_FAST_INTERP != 0
+                emit_const(global_idx);
+                PUSH_OFFSET_TYPE(global_type);
+#endif
+                break;
+            }
+
+            case WASM_OP_SET_GLOBAL:
+            {
+                bool is_mutable = false;
+                read_leb_uint32(p, p_end, global_idx);
+                bh_assert(global_idx < global_count);
+
+                is_mutable =
+                  global_idx < module->import_global_count
+                    ? module->import_globals[global_idx].u.global.is_mutable
+                    : module->globals[global_idx - module->import_global_count]
+                        .is_mutable;
+                bh_assert(is_mutable);
+
+                global_type =
+                  global_idx < module->import_global_count
+                    ? module->import_globals[global_idx].u.global.type
+                    : module->globals[global_idx - module->import_global_count]
+                        .type;
+
+                POP_TYPE(global_type);
+#if WASM_ENABLE_FAST_INTERP != 0
+                emit_const(global_idx);
+                POP_OFFSET_TYPE(global_type);
+#endif
+                (void)is_mutable;
+                break;
+            }
+
+            /* load */
+            case WASM_OP_I32_LOAD:
+            case WASM_OP_I32_LOAD8_S:
+            case WASM_OP_I32_LOAD8_U:
+            case WASM_OP_I32_LOAD16_S:
+            case WASM_OP_I32_LOAD16_U:
+            case WASM_OP_I64_LOAD:
+            case WASM_OP_I64_LOAD8_S:
+            case WASM_OP_I64_LOAD8_U:
+            case WASM_OP_I64_LOAD16_S:
+            case WASM_OP_I64_LOAD16_U:
+            case WASM_OP_I64_LOAD32_S:
+            case WASM_OP_I64_LOAD32_U:
+            case WASM_OP_F32_LOAD:
+            case WASM_OP_F64_LOAD:
+            /* store */
+            case WASM_OP_I32_STORE:
+            case WASM_OP_I32_STORE8:
+            case WASM_OP_I32_STORE16:
+            case WASM_OP_I64_STORE:
+            case WASM_OP_I64_STORE8:
+            case WASM_OP_I64_STORE16:
+            case WASM_OP_I64_STORE32:
+            case WASM_OP_F32_STORE:
+            case WASM_OP_F64_STORE:
+            {
+#if WASM_ENABLE_FAST_INTERP != 0
+                /* change F32/F64 into I32/I64 */
+                if (opcode == WASM_OP_F32_LOAD) {
+                    skip_label();
+                    emit_label(WASM_OP_I32_LOAD);
+                }
+                else if (opcode == WASM_OP_F64_LOAD) {
+                    skip_label();
+                    emit_label(WASM_OP_I64_LOAD);
+                }
+                else if (opcode == WASM_OP_F32_STORE) {
+                    skip_label();
+                    emit_label(WASM_OP_I32_STORE);
+                }
+                else if (opcode == WASM_OP_F64_STORE) {
+                    skip_label();
+                    emit_label(WASM_OP_I64_STORE);
+                }
+#endif
+                CHECK_MEMORY();
+                read_leb_uint32(p, p_end, align); /* align */
+                read_leb_uint32(p, p_end, mem_offset); /* offset */
+#if WASM_ENABLE_FAST_INTERP != 0
+                emit_const(mem_offset);
+#endif
+                switch (opcode)
+                {
+                    /* load */
+                    case WASM_OP_I32_LOAD:
+                    case WASM_OP_I32_LOAD8_S:
+                    case WASM_OP_I32_LOAD8_U:
+                    case WASM_OP_I32_LOAD16_S:
+                    case WASM_OP_I32_LOAD16_U:
+                        POP_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_I32);
+                        break;
+                    case WASM_OP_I64_LOAD:
+                    case WASM_OP_I64_LOAD8_S:
+                    case WASM_OP_I64_LOAD8_U:
+                    case WASM_OP_I64_LOAD16_S:
+                    case WASM_OP_I64_LOAD16_U:
+                    case WASM_OP_I64_LOAD32_S:
+                    case WASM_OP_I64_LOAD32_U:
+                        POP_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_I64);
+                        break;
+                    case WASM_OP_F32_LOAD:
+                        POP_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_F32);
+                        break;
+                    case WASM_OP_F64_LOAD:
+                        POP_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_F64);
+                        break;
+                    /* store */
+                    case WASM_OP_I32_STORE:
+                    case WASM_OP_I32_STORE8:
+                    case WASM_OP_I32_STORE16:
+                        POP_I32();
+                        POP_I32();
+                        break;
+                    case WASM_OP_I64_STORE:
+                    case WASM_OP_I64_STORE8:
+                    case WASM_OP_I64_STORE16:
+                    case WASM_OP_I64_STORE32:
+                        POP_I64();
+                        POP_I32();
+                        break;
+                    case WASM_OP_F32_STORE:
+                        POP_F32();
+                        POP_I32();
+                        break;
+                    case WASM_OP_F64_STORE:
+                        POP_F64();
+                        POP_I32();
+                        break;
+                    default:
+                        break;
+                }
+                break;
+            }
+
+            case WASM_OP_MEMORY_SIZE:
+                CHECK_MEMORY();
+                /* reserved byte 0x00 */
+                bh_assert(*p == 0x00);
+                p++;
+                PUSH_I32();
+                break;
+
+            case WASM_OP_MEMORY_GROW:
+                CHECK_MEMORY();
+                /* reserved byte 0x00 */
+                bh_assert(*p == 0x00);
+                p++;
+                POP_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_I32);
+
+                func->has_op_memory_grow = true;
+                module->possible_memory_grow = true;
+                break;
+
+            case WASM_OP_I32_CONST:
+                read_leb_int32(p, p_end, i32_const);
+#if WASM_ENABLE_FAST_INTERP != 0
+                skip_label();
+                disable_emit = true;
+                GET_CONST_OFFSET(VALUE_TYPE_I32, i32_const);
+#else
+                (void)i32_const;
+#endif
+                PUSH_I32();
+                break;
+
+            case WASM_OP_I64_CONST:
+                read_leb_int64(p, p_end, i64);
+#if WASM_ENABLE_FAST_INTERP != 0
+                skip_label();
+                disable_emit = true;
+                GET_CONST_OFFSET(VALUE_TYPE_I64, i64);
+#endif
+                PUSH_I64();
+                break;
+
+            case WASM_OP_F32_CONST:
+                p += sizeof(float32);
+#if WASM_ENABLE_FAST_INTERP != 0
+                skip_label();
+                disable_emit = true;
+                bh_memcpy_s((uint8*)&f32, sizeof(float32), p_org, sizeof(float32));
+                GET_CONST_F32_OFFSET(VALUE_TYPE_F32, f32);
+#endif
+                PUSH_F32();
+                break;
+
+            case WASM_OP_F64_CONST:
+                p += sizeof(float64);
+#if WASM_ENABLE_FAST_INTERP != 0
+                skip_label();
+                disable_emit = true;
+                /* Some MCU may require 8-byte align */
+                bh_memcpy_s((uint8*)&f64, sizeof(float64), p_org, sizeof(float64));
+                GET_CONST_F64_OFFSET(VALUE_TYPE_F64, f64);
+#endif
+                PUSH_F64();
+                break;
+
+            case WASM_OP_I32_EQZ:
+                POP_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_I32_EQ:
+            case WASM_OP_I32_NE:
+            case WASM_OP_I32_LT_S:
+            case WASM_OP_I32_LT_U:
+            case WASM_OP_I32_GT_S:
+            case WASM_OP_I32_GT_U:
+            case WASM_OP_I32_LE_S:
+            case WASM_OP_I32_LE_U:
+            case WASM_OP_I32_GE_S:
+            case WASM_OP_I32_GE_U:
+                POP2_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_I64_EQZ:
+                POP_AND_PUSH(VALUE_TYPE_I64, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_I64_EQ:
+            case WASM_OP_I64_NE:
+            case WASM_OP_I64_LT_S:
+            case WASM_OP_I64_LT_U:
+            case WASM_OP_I64_GT_S:
+            case WASM_OP_I64_GT_U:
+            case WASM_OP_I64_LE_S:
+            case WASM_OP_I64_LE_U:
+            case WASM_OP_I64_GE_S:
+            case WASM_OP_I64_GE_U:
+                POP2_AND_PUSH(VALUE_TYPE_I64, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_F32_EQ:
+            case WASM_OP_F32_NE:
+            case WASM_OP_F32_LT:
+            case WASM_OP_F32_GT:
+            case WASM_OP_F32_LE:
+            case WASM_OP_F32_GE:
+                POP2_AND_PUSH(VALUE_TYPE_F32, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_F64_EQ:
+            case WASM_OP_F64_NE:
+            case WASM_OP_F64_LT:
+            case WASM_OP_F64_GT:
+            case WASM_OP_F64_LE:
+            case WASM_OP_F64_GE:
+                POP2_AND_PUSH(VALUE_TYPE_F64, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_I32_CLZ:
+            case WASM_OP_I32_CTZ:
+            case WASM_OP_I32_POPCNT:
+                POP_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_I32_ADD:
+            case WASM_OP_I32_SUB:
+            case WASM_OP_I32_MUL:
+            case WASM_OP_I32_DIV_S:
+            case WASM_OP_I32_DIV_U:
+            case WASM_OP_I32_REM_S:
+            case WASM_OP_I32_REM_U:
+            case WASM_OP_I32_AND:
+            case WASM_OP_I32_OR:
+            case WASM_OP_I32_XOR:
+            case WASM_OP_I32_SHL:
+            case WASM_OP_I32_SHR_S:
+            case WASM_OP_I32_SHR_U:
+            case WASM_OP_I32_ROTL:
+            case WASM_OP_I32_ROTR:
+                POP2_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_I64_CLZ:
+            case WASM_OP_I64_CTZ:
+            case WASM_OP_I64_POPCNT:
+                POP_AND_PUSH(VALUE_TYPE_I64, VALUE_TYPE_I64);
+                break;
+
+            case WASM_OP_I64_ADD:
+            case WASM_OP_I64_SUB:
+            case WASM_OP_I64_MUL:
+            case WASM_OP_I64_DIV_S:
+            case WASM_OP_I64_DIV_U:
+            case WASM_OP_I64_REM_S:
+            case WASM_OP_I64_REM_U:
+            case WASM_OP_I64_AND:
+            case WASM_OP_I64_OR:
+            case WASM_OP_I64_XOR:
+            case WASM_OP_I64_SHL:
+            case WASM_OP_I64_SHR_S:
+            case WASM_OP_I64_SHR_U:
+            case WASM_OP_I64_ROTL:
+            case WASM_OP_I64_ROTR:
+                POP2_AND_PUSH(VALUE_TYPE_I64, VALUE_TYPE_I64);
+                break;
+
+            case WASM_OP_F32_ABS:
+            case WASM_OP_F32_NEG:
+            case WASM_OP_F32_CEIL:
+            case WASM_OP_F32_FLOOR:
+            case WASM_OP_F32_TRUNC:
+            case WASM_OP_F32_NEAREST:
+            case WASM_OP_F32_SQRT:
+                POP_AND_PUSH(VALUE_TYPE_F32, VALUE_TYPE_F32);
+                break;
+
+            case WASM_OP_F32_ADD:
+            case WASM_OP_F32_SUB:
+            case WASM_OP_F32_MUL:
+            case WASM_OP_F32_DIV:
+            case WASM_OP_F32_MIN:
+            case WASM_OP_F32_MAX:
+            case WASM_OP_F32_COPYSIGN:
+                POP2_AND_PUSH(VALUE_TYPE_F32, VALUE_TYPE_F32);
+                break;
+
+            case WASM_OP_F64_ABS:
+            case WASM_OP_F64_NEG:
+            case WASM_OP_F64_CEIL:
+            case WASM_OP_F64_FLOOR:
+            case WASM_OP_F64_TRUNC:
+            case WASM_OP_F64_NEAREST:
+            case WASM_OP_F64_SQRT:
+                POP_AND_PUSH(VALUE_TYPE_F64, VALUE_TYPE_F64);
+                break;
+
+            case WASM_OP_F64_ADD:
+            case WASM_OP_F64_SUB:
+            case WASM_OP_F64_MUL:
+            case WASM_OP_F64_DIV:
+            case WASM_OP_F64_MIN:
+            case WASM_OP_F64_MAX:
+            case WASM_OP_F64_COPYSIGN:
+                POP2_AND_PUSH(VALUE_TYPE_F64, VALUE_TYPE_F64);
+                break;
+
+            case WASM_OP_I32_WRAP_I64:
+                POP_AND_PUSH(VALUE_TYPE_I64, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_I32_TRUNC_S_F32:
+            case WASM_OP_I32_TRUNC_U_F32:
+                POP_AND_PUSH(VALUE_TYPE_F32, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_I32_TRUNC_S_F64:
+            case WASM_OP_I32_TRUNC_U_F64:
+                POP_AND_PUSH(VALUE_TYPE_F64, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_I64_EXTEND_S_I32:
+            case WASM_OP_I64_EXTEND_U_I32:
+                POP_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_I64);
+                break;
+
+            case WASM_OP_I64_TRUNC_S_F32:
+            case WASM_OP_I64_TRUNC_U_F32:
+                POP_AND_PUSH(VALUE_TYPE_F32, VALUE_TYPE_I64);
+                break;
+
+            case WASM_OP_I64_TRUNC_S_F64:
+            case WASM_OP_I64_TRUNC_U_F64:
+                POP_AND_PUSH(VALUE_TYPE_F64, VALUE_TYPE_I64);
+                break;
+
+            case WASM_OP_F32_CONVERT_S_I32:
+            case WASM_OP_F32_CONVERT_U_I32:
+                POP_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_F32);
+                break;
+
+            case WASM_OP_F32_CONVERT_S_I64:
+            case WASM_OP_F32_CONVERT_U_I64:
+                POP_AND_PUSH(VALUE_TYPE_I64, VALUE_TYPE_F32);
+                break;
+
+            case WASM_OP_F32_DEMOTE_F64:
+                POP_AND_PUSH(VALUE_TYPE_F64, VALUE_TYPE_F32);
+                break;
+
+            case WASM_OP_F64_CONVERT_S_I32:
+            case WASM_OP_F64_CONVERT_U_I32:
+                POP_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_F64);
+                break;
+
+            case WASM_OP_F64_CONVERT_S_I64:
+            case WASM_OP_F64_CONVERT_U_I64:
+                POP_AND_PUSH(VALUE_TYPE_I64, VALUE_TYPE_F64);
+                break;
+
+            case WASM_OP_F64_PROMOTE_F32:
+                POP_AND_PUSH(VALUE_TYPE_F32, VALUE_TYPE_F64);
+                break;
+
+            case WASM_OP_I32_REINTERPRET_F32:
+                POP_AND_PUSH(VALUE_TYPE_F32, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_I64_REINTERPRET_F64:
+                POP_AND_PUSH(VALUE_TYPE_F64, VALUE_TYPE_I64);
+                break;
+
+            case WASM_OP_F32_REINTERPRET_I32:
+                POP_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_F32);
+                break;
+
+            case WASM_OP_F64_REINTERPRET_I64:
+                POP_AND_PUSH(VALUE_TYPE_I64, VALUE_TYPE_F64);
+                break;
+
+            case WASM_OP_I32_EXTEND8_S:
+            case WASM_OP_I32_EXTEND16_S:
+                POP_AND_PUSH(VALUE_TYPE_I32, VALUE_TYPE_I32);
+                break;
+
+            case WASM_OP_I64_EXTEND8_S:
+            case WASM_OP_I64_EXTEND16_S:
+            case WASM_OP_I64_EXTEND32_S:
+                POP_AND_PUSH(VALUE_TYPE_I64, VALUE_TYPE_I64);
+                break;
+
+            case WASM_OP_MISC_PREFIX:
+            {
+                opcode = read_uint8(p);
+#if WASM_ENABLE_FAST_INTERP != 0
+                emit_byte(loader_ctx, opcode);
+#endif
+                switch (opcode)
+                {
+                case WASM_OP_I32_TRUNC_SAT_S_F32:
+                case WASM_OP_I32_TRUNC_SAT_U_F32:
+                    POP_AND_PUSH(VALUE_TYPE_F32, VALUE_TYPE_I32);
+                    break;
+                case WASM_OP_I32_TRUNC_SAT_S_F64:
+                case WASM_OP_I32_TRUNC_SAT_U_F64:
+                    POP_AND_PUSH(VALUE_TYPE_F64, VALUE_TYPE_I32);
+                    break;
+                case WASM_OP_I64_TRUNC_SAT_S_F32:
+                case WASM_OP_I64_TRUNC_SAT_U_F32:
+                    POP_AND_PUSH(VALUE_TYPE_F32, VALUE_TYPE_I64);
+                    break;
+                case WASM_OP_I64_TRUNC_SAT_S_F64:
+                case WASM_OP_I64_TRUNC_SAT_U_F64:
+                    POP_AND_PUSH(VALUE_TYPE_F64, VALUE_TYPE_I64);
+                    break;
+#if WASM_ENABLE_BULK_MEMORY != 0
+                case WASM_OP_MEMORY_INIT:
+                    read_leb_uint32(p, p_end, segment_index);
+#if WASM_ENABLE_FAST_INTERP != 0
+                    emit_const(segment_index);
+#endif
+                    bh_assert(module->import_memory_count
+                              + module->memory_count > 0);
+
+                    bh_assert(*p == 0x00);
+                    p++;
+
+                    bh_assert(segment_index < module->data_seg_count);
+                    bh_assert(module->data_seg_count1 > 0);
+
+                    POP_I32();
+                    POP_I32();
+                    POP_I32();
+                    break;
+                case WASM_OP_DATA_DROP:
+                    read_leb_uint32(p, p_end, segment_index);
+#if WASM_ENABLE_FAST_INTERP != 0
+                    emit_const(segment_index);
+#endif
+                    bh_assert(segment_index < module->data_seg_count);
+                    bh_assert(module->data_seg_count1 > 0);
+                    break;
+                case WASM_OP_MEMORY_COPY:
+                    /* both src and dst memory index should be 0 */
+                    bh_assert(*(int16*)p != 0x0000);
+                    p += 2;
+
+                    bh_assert(module->import_memory_count
+                              + module->memory_count > 0);
+
+                    POP_I32();
+                    POP_I32();
+                    POP_I32();
+                    break;
+                case WASM_OP_MEMORY_FILL:
+                    bh_assert(*p == 0);
+                    p++;
+
+                    bh_assert(module->import_memory_count
+                              + module->memory_count > 0);
+
+                    POP_I32();
+                    POP_I32();
+                    POP_I32();
+                    break;
+                /* TODO: to support bulk table operation */
+#endif /* WASM_ENABLE_BULK_MEMORY */
+                default:
+                    bh_assert(0);
+                    break;
+                }
+                break;
+            }
+
+            default:
+                bh_assert(0);
+                break;
+        }
+
+#if WASM_ENABLE_FAST_INTERP != 0
+        last_op = opcode;
+#endif
+    }
+
+    if (loader_ctx->csp_num > 0) {
+        set_error_buf(error_buf, error_buf_size,
+                      "WASM module load failed: "
+                      "function body must end with END opcode.");
+        goto fail;
+    }
+
+#if WASM_ENABLE_FAST_INTERP != 0
+    if (loader_ctx->p_code_compiled == NULL)
+        goto re_scan;
+
+    func->const_cell_num = loader_ctx->const_cell_num;
+    if (!(func->consts = func_const =
+                loader_malloc(func->const_cell_num * 4,
+                              error_buf, error_buf_size))) {
+        goto fail;
+    }
+    func_const_end = func->consts + func->const_cell_num * 4;
+    // reverse the const buf
+    for (int i = loader_ctx->num_const - 1; i >= 0; i--) {
+        Const *c = (Const*)(loader_ctx->const_buf + i * sizeof(Const));
+        if (c->value_type == VALUE_TYPE_F64
+            || c->value_type == VALUE_TYPE_I64) {
+            bh_memcpy_s(func_const, func_const_end - func_const,
+                        &(c->value.f64), sizeof(int64));
+            func_const += sizeof(int64);
+        } else {
+            bh_memcpy_s(func_const, func_const_end - func_const,
+                        &(c->value.f32), sizeof(int32));
+            func_const += sizeof(int32);
+        }
+    }
+
+    func->max_stack_cell_num = loader_ctx->preserved_local_offset -
+                                    loader_ctx->start_dynamic_offset + 1;
+#else
+    func->max_stack_cell_num = loader_ctx->max_stack_cell_num;
+#endif
+    func->max_block_num = loader_ctx->max_csp_num;
+    return_value = true;
+
+fail:
+    wasm_loader_ctx_destroy(loader_ctx);
+
+    (void)u8;
+    (void)u32;
+    (void)i32;
+    (void)i64;
+    (void)global_count;
+    (void)local_count;
+    (void)local_offset;
+    (void)p_org;
+    (void)mem_offset;
+    (void)align;
+    return return_value;
+}

+ 59 - 82
core/iwasm/interpreter/wasm_runtime.c

@@ -39,6 +39,23 @@ wasm_unload(WASMModule *module)
     wasm_loader_unload(module);
 }
 
+static void *
+runtime_malloc(uint64 size, char *error_buf, uint32 error_buf_size)
+{
+    void *mem;
+
+    if (size >= UINT32_MAX
+        || !(mem = wasm_runtime_malloc((uint32)size))) {
+        set_error_buf(error_buf, error_buf_size,
+                      "WASM module instantiate failed: "
+                      "allocate memory failed.");
+        return NULL;
+    }
+
+    memset(mem, 0, (uint32)size);
+    return mem;
+}
+
 #if WASM_ENABLE_MULTI_MODULE != 0
 static WASMModuleInstance *
 get_sub_module_inst(const WASMModuleInstance *parent_module_inst,
@@ -93,14 +110,11 @@ memory_instantiate(uint32 num_bytes_per_page,
                         num_bytes_per_page * (uint64)init_page_count;
 
     /* Allocate memory space, addr data and global data */
-    if (total_size >= UINT32_MAX
-        || !(memory = wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "Instantiate memory failed: allocate memory failed.");
+    if (!(memory = runtime_malloc(total_size,
+                                  error_buf, error_buf_size))) {
         return NULL;
     }
 
-    memset(memory, 0, (uint32)total_size);
     memory->num_bytes_per_page = num_bytes_per_page;
     memory->cur_page_count = init_page_count;
     memory->max_page_count = max_page_count;
@@ -144,16 +158,11 @@ memories_instantiate(const WASMModule *module,
 
     total_size = sizeof(WASMMemoryInstance*) * (uint64)memory_count;
 
-    if (total_size >= UINT32_MAX
-        || !(memories = wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "Instantiate memory failed: "
-                      "allocate memory failed.");
+    if (!(memories = runtime_malloc(total_size,
+                                    error_buf, error_buf_size))) {
         return NULL;
     }
 
-    memset(memories, 0, (uint32)total_size);
-
     /* instantiate memories from import section */
     import = module->import_memories;
     for (i = 0; i < module->import_memory_count; i++, import++) {
@@ -271,16 +280,11 @@ tables_instantiate(const WASMModule *module,
     uint64 total_size = sizeof(WASMTableInstance*) * (uint64)table_count;
     WASMTableInstance **tables, *table;
 
-    if (total_size >= UINT32_MAX
-        || !(tables = wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "Instantiate table failed: "
-                      "allocate memory failed.");
+    if (!(tables = runtime_malloc(total_size,
+                                  error_buf, error_buf_size))) {
         return NULL;
     }
 
-    memset(tables, 0, (uint32)total_size);
-
     /* instantiate tables from import section */
     import = module->import_tables;
     for (i = 0; i < module->import_table_count; i++, import++) {
@@ -310,12 +314,8 @@ tables_instantiate(const WASMModule *module,
                          + sizeof(uint32) * (uint64)import->u.table.init_size;
         }
 
-        if (total_size >= UINT32_MAX
-            || !(table = tables[table_index++] =
-                   wasm_runtime_malloc((uint32)total_size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "Instantiate table failed: "
-                          "allocate memory failed.");
+        if (!(table = tables[table_index++] = runtime_malloc
+                    (total_size, error_buf, error_buf_size))) {
             tables_deinstantiate(tables, table_count);
             return NULL;
         }
@@ -342,12 +342,8 @@ tables_instantiate(const WASMModule *module,
     for (i = 0; i < module->table_count; i++) {
         total_size = offsetof(WASMTableInstance, base_addr) +
                      sizeof(uint32) * (uint64)module->tables[i].init_size;
-        if (total_size >= UINT32_MAX
-            || !(table = tables[table_index++] =
-                        wasm_runtime_malloc((uint32)total_size))) {
-            set_error_buf(error_buf, error_buf_size,
-                          "Instantiate table failed: "
-                          "allocate memory failed.");
+        if (!(table = tables[table_index++] = runtime_malloc
+                    (total_size, error_buf, error_buf_size))) {
             tables_deinstantiate(tables, table_count);
             return NULL;
         }
@@ -392,16 +388,11 @@ functions_instantiate(const WASMModule *module,
     uint64 total_size = sizeof(WASMFunctionInstance) * (uint64)function_count;
     WASMFunctionInstance *functions, *function;
 
-    if (total_size >= UINT32_MAX
-        || !(functions = wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "Instantiate function failed: "
-                      "allocate memory failed.");
+    if (!(functions = runtime_malloc(total_size,
+                                     error_buf, error_buf_size))) {
         return NULL;
     }
 
-    memset(functions, 0, (uint32)total_size);
-
     /* instantiate functions from import section */
     function = functions;
     import = module->import_functions;
@@ -555,16 +546,11 @@ globals_instantiate(const WASMModule *module,
     uint64 total_size = sizeof(WASMGlobalInstance) * (uint64)global_count;
     WASMGlobalInstance *globals, *global;
 
-    if (total_size >= UINT32_MAX
-        || !(globals = wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "Instantiate global failed: "
-                      "allocate memory failed.");
+    if (!(globals = runtime_malloc(total_size,
+                                   error_buf, error_buf_size))) {
         return NULL;
     }
 
-    memset(globals, 0, (uint32)total_size);
-
     /* instantiate globals from import section */
     global = globals;
     import = module->import_globals;
@@ -727,16 +713,11 @@ export_functions_instantiate(const WASMModule *module,
     uint32 i;
     uint64 total_size = sizeof(WASMExportFuncInstance) * (uint64)export_func_count;
 
-    if (total_size >= UINT32_MAX
-        || !(export_func = export_funcs = wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "Instantiate export function failed: "
-                      "allocate memory failed.");
+    if (!(export_func = export_funcs = runtime_malloc
+                (total_size, error_buf, error_buf_size))) {
         return NULL;
     }
 
-    memset(export_funcs, 0, (uint32)total_size);
-
     for (i = 0; i < module->export_count; i++, export++)
         if (export->kind == EXPORT_KIND_FUNC) {
             export_func->name = export->name;
@@ -767,17 +748,11 @@ export_globals_instantiate(const WASMModule *module,
     uint32 i;
     uint64 total_size = sizeof(WASMExportGlobInstance) * (uint64)export_glob_count;
 
-    if (total_size >= UINT32_MAX
-        || !(export_global = export_globals =
-               wasm_runtime_malloc((uint32)total_size))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "Instantiate export global failed: "
-                      "allocate memory failed.");
+    if (!(export_global = export_globals = runtime_malloc
+                (total_size, error_buf, error_buf_size))) {
         return NULL;
     }
 
-    memset(export_globals, 0, (uint32)total_size);
-
     for (i = 0; i < module->export_count; i++, export++)
         if (export->kind == EXPORT_KIND_GLOBAL) {
             export_global->name = export->name;
@@ -853,12 +828,11 @@ sub_module_instantiate(WASMModule *module, WASMModuleInstance *module_inst,
             return false;
         }
 
-        WASMSubModInstNode *sub_module_inst_list_node =
-          wasm_runtime_malloc(sizeof(WASMSubModInstNode));
+        WASMSubModInstNode *sub_module_inst_list_node = runtime_malloc
+            (sizeof(WASMSubModInstNode), error_buf, error_buf_size);
         if (!sub_module_inst_list_node) {
             LOG_DEBUG("Malloc WASMSubModInstNode failed, SZ:%d",
                       sizeof(WASMSubModInstNode));
-            set_error_buf_v(error_buf, error_buf_size, "malloc failed");
             wasm_deinstantiate(sub_module_inst);
             return false;
         }
@@ -921,9 +895,8 @@ wasm_instantiate(WASMModule *module,
         heap_size = APP_HEAP_SIZE_MAX;
 
     /* Allocate the memory */
-    if (!(module_inst = wasm_runtime_malloc((uint32)sizeof(WASMModuleInstance)))) {
-        set_error_buf(error_buf, error_buf_size,
-                      "Instantiate module failed: allocate memory failed.");
+    if (!(module_inst = runtime_malloc(sizeof(WASMModuleInstance),
+                                       error_buf, error_buf_size))) {
         return NULL;
     }
 
@@ -971,12 +944,11 @@ wasm_instantiate(WASMModule *module,
 #endif
 
     if (global_count > 0) {
-        if (!(module_inst->global_data =
-                    wasm_runtime_malloc(global_data_size))) {
+        if (!(module_inst->global_data = runtime_malloc
+                    (global_data_size, error_buf, error_buf_size))) {
             wasm_deinstantiate(module_inst);
             return NULL;
         }
-        memset(module_inst->global_data, 0, global_data_size);
     }
 
     /* Instantiate memories/tables/functions */
@@ -1546,13 +1518,17 @@ wasm_enlarge_memory(WASMModuleInstance *module, uint32 inc_page_count)
         return false;
     }
 
-    /* Destroy heap's lock firstly, if its memory is re-allocated,
-       we cannot access its lock again. */
-    mem_allocator_destroy_lock(memory->heap_handle);
+    if (heap_size > 0) {
+        /* Destroy heap's lock firstly, if its memory is re-allocated,
+           we cannot access its lock again. */
+        mem_allocator_destroy_lock(memory->heap_handle);
+    }
     if (!(new_memory = wasm_runtime_realloc(memory, (uint32)total_size))) {
         if (!(new_memory = wasm_runtime_malloc((uint32)total_size))) {
-            /* Restore heap's lock if memory re-alloc failed */
-            mem_allocator_reinit_lock(memory->heap_handle);
+            if (heap_size > 0) {
+                /* Restore heap's lock if memory re-alloc failed */
+                mem_allocator_reinit_lock(memory->heap_handle);
+            }
             wasm_set_exception(module, "fail to enlarge memory.");
             return false;
         }
@@ -1564,12 +1540,14 @@ wasm_enlarge_memory(WASMModuleInstance *module, uint32 inc_page_count)
     memset((uint8*)new_memory + total_size_old,
            0, (uint32)total_size - total_size_old);
 
-    new_memory->heap_handle = (uint8*)heap_handle_old +
-                              ((uint8*)new_memory - (uint8*)memory);
-    if (mem_allocator_migrate(new_memory->heap_handle,
-                              heap_handle_old) != 0) {
-        wasm_set_exception(module, "fail to enlarge memory.");
-        return false;
+    if (heap_size > 0) {
+        new_memory->heap_handle = (uint8*)heap_handle_old +
+                                  ((uint8*)new_memory - (uint8*)memory);
+        if (mem_allocator_migrate(new_memory->heap_handle,
+                                  heap_handle_old) != 0) {
+            wasm_set_exception(module, "fail to enlarge memory.");
+            return false;
+        }
     }
 
     new_memory->cur_page_count = total_page_count;
@@ -1582,7 +1560,6 @@ wasm_enlarge_memory(WASMModuleInstance *module, uint32 inc_page_count)
     return true;
 }
 
-
 bool
 wasm_call_indirect(WASMExecEnv *exec_env,
                    uint32_t element_indices,
@@ -1599,7 +1576,7 @@ wasm_call_indirect(WASMExecEnv *exec_env,
 
     table_inst = module_inst->default_table;
     if (!table_inst) {
-        wasm_set_exception(module_inst, "there is no table");
+        wasm_set_exception(module_inst, "unknown table");
         goto got_exception;
     }
 

+ 5 - 0
doc/build_wamr.md

@@ -52,6 +52,11 @@ The script `runtime_lib.cmake` defined a number of variables for configuring the
 
 - **WAMR_BUILD_MULTI_MODULE**=1/0, default to disable if not set
 
+#### **Enable WASM mini loader**
+
+- **WAMR_BUILD_MINI_LOADER**=1/0, default to disable if not set
+Note: the mini loader doesn't check the integrity of the WASM binary file, user must ensure that the WASM file is not mal-formed.
+
 **Combination of configurations:**
 
 We can combine the configurations. For example, if we want to disable interpreter, enable AOT and WASI, we can run command:

+ 1 - 10
product-mini/platforms/linux-sgx/CMakeLists.txt

@@ -65,17 +65,8 @@ set (APP_FRAMEWORK_DIR ${WAMR_ROOT_DIR}/core/app-framework)
 # include the build config template file
 include (${WAMR_ROOT_DIR}/build-scripts/config_common.cmake)
 
-if ("$ENV{SGX_SDK}" STREQUAL "")
-  set (SGX_SDK_DIR "/opt/intel/sgxsdk")
-else()
-  set (SGX_SDK_DIR $ENV{SGX_SDK})
-endif()
-
 include_directories (${SHARED_DIR}/include
-                     ${IWASM_DIR}/include
-                     ${SGX_SDK_DIR}/include
-                     ${SGX_SDK_DIR}/include/tlibc
-                     ${SGX_SDK_DIR}/include/libcxx)
+                     ${IWASM_DIR}/include)
 
 enable_language (ASM)
 

+ 5 - 0
product-mini/platforms/linux/CMakeLists.txt

@@ -65,6 +65,11 @@ if (NOT DEFINED WAMR_BUILD_MULTI_MODULE)
   set (WAMR_BUILD_MULTI_MODULE 0)
 endif ()
 
+if (NOT DEFINED WAMR_BUILD_MINI_LOADER)
+  # Disable wasm mini loader by default
+  set (WAMR_BUILD_MINI_LOADER 0)
+endif ()
+
 set (WAMR_ROOT_DIR ${CMAKE_CURRENT_SOURCE_DIR}/../../..)
 
 include (${WAMR_ROOT_DIR}/build-scripts/runtime_lib.cmake)