diff --git a/lite/api/paddle_place.cc b/lite/api/paddle_place.cc
index cb70fd43cdd80713a6a47afb61bbde24e131a5fa..9bc63e78aae92556a312eb36c3415f9d57c2239a 100644
--- a/lite/api/paddle_place.cc
+++ b/lite/api/paddle_place.cc
@@ -24,9 +24,9 @@ namespace lite_api {
 size_t Place::hash() const {
   std::hash<int> h;
   size_t hash = h(static_cast<int>(target));
-  hash = lite::hash_combine(hash, static_cast<int>(precision));
-  hash = lite::hash_combine(hash, static_cast<int>(layout));
-  hash = lite::hash_combine(hash, static_cast<int>(device));
+  lite::CombineHash(static_cast<int64_t>(precision), &hash);
+  lite::CombineHash(static_cast<int64_t>(layout), &hash);
+  lite::CombineHash(static_cast<int64_t>(device), &hash);
   return hash;
 }
 
diff --git a/lite/core/op_lite.cc b/lite/core/op_lite.cc
index de76f404f8a129eb94e645dc731a0d09c1ee3c77..7428a16d0afab9dba126189c19d0eb99f0d8e31d 100644
--- a/lite/core/op_lite.cc
+++ b/lite/core/op_lite.cc
@@ -42,16 +42,14 @@ bool OpLite::InferShapeWithCache() {
     // combined dims value into new_hash value.
     auto &element_dims = (*iter)->dims();
     for (size_t i = 0; i < element_dims.size(); i++) {
-      new_hash =
-          lite::hash_combine(new_hash, static_cast<int>(element_dims[i]));
+      lite::CombineHash(static_cast<int64_t>(element_dims[i]), &new_hash);
     }
     // combine lod value into new_hash valud.
     auto &emement_lods = (*iter)->lod();
     for (auto lod_iter = emement_lods.begin(); lod_iter != emement_lods.end();
          lod_iter++) {
       for (size_t i = 0; i < lod_iter->size(); i++) {
-        new_hash =
-            lite::hash_combine(new_hash, static_cast<int>(lod_iter->at(i)));
+        lite::CombineHash(static_cast<int64_t>(lod_iter->at(i)), &new_hash);
       }
     }
   }
diff --git a/lite/core/type_system.cc b/lite/core/type_system.cc
index 276d0c4a349794bed0ece755c924cf789a7cf54e..aaafd29841f44e671460a4c45babc7a8f663dacf 100644
--- a/lite/core/type_system.cc
+++ b/lite/core/type_system.cc
@@ -21,9 +21,9 @@ namespace lite {
 size_t ParamTypeRegistry::KernelIdTy::hash() const {
   std::hash<std::string> h;
   size_t hash = h(kernel_type);
-  hash = hash_combine(hash, place.hash());
-  hash = hash_combine(hash, std::hash<int>()(static_cast<int>(io)));
-  hash = hash_combine(hash, std::hash<std::string>()(arg_name));
+  lite::CombineHash(place.hash(), &hash);
+  lite::CombineHash(std::hash<int>()(static_cast<int>(io)), &hash);
+  lite::CombineHash(std::hash<std::string>()(arg_name), &hash);
   return hash;
 }
 
@@ -48,8 +48,7 @@ const Type *Type::GetTensorTy(TargetType target,
   // NOTE quite naive implementation here, but not performance sensitive.
   DataType::ID type_id = DataType::ID::Tensor;
 
-#define HASH_ONE(x) v = hash_combine(v, hasher(static_cast<int>(x)))
-
+#define HASH_ONE(x) CombineHash(hasher(static_cast<int>(x)), &v);
   std::hash<int> hasher;
   size_t v = hasher(static_cast<int>(type_id));
   HASH_ONE(target);
@@ -80,8 +79,7 @@ const Type *Type::GetTensorListTy(TargetType target,
   static std::map<size_t, const Type *> type_repo;
   DataType::ID type_id = DataType::ID::TensorList;
 
-#define HASH_ONE(x) v = hash_combine(v, hasher(static_cast<int>(x)))
-
+#define HASH_ONE(x) CombineHash(hasher(static_cast<int>(x)), &v);
   std::hash<int> hasher;
   size_t v = hasher(static_cast<int>(type_id));
   HASH_ONE(target);
diff --git a/lite/utils/hash.h b/lite/utils/hash.h
index a1fa3be02e58f0908b108a65431ca1993512c821..0135b53a8609a2a8168a25727738afbda4398dc7 100644
--- a/lite/utils/hash.h
+++ b/lite/utils/hash.h
@@ -18,10 +18,11 @@
 namespace paddle {
 namespace lite {
 
+// A simplified implementation of boost::hash_combine.
 template <typename T>
-inline size_t hash_combine(size_t s, const T& v) {
+inline void CombineHash(const T& from, size_t* to) {
   std::hash<T> h;
-  return (s ^ h(v)) + 0x9e3779b9 + (s << 6) + (s >> 2);
+  *to ^= h(from) + 0x9e3779b9 + (*to << 6) + (*to >> 2);
 }
 
 }  // namespace lite