diff --git a/include/fast_io_hosted/platforms/linux/aarch64.h b/include/fast_io_hosted/platforms/linux/aarch64.h
index f2458f5c..44c3ddaa 100644
--- a/include/fast_io_hosted/platforms/linux/aarch64.h
+++ b/include/fast_io_hosted/platforms/linux/aarch64.h
@@ -1,5 +1,9 @@
 #pragma once
 
+#include <cstddef>
+#include <cstdint>
+#include <concepts>
+
 namespace fast_io
 {
 
@@ -24,6 +28,7 @@ inline return_value_type system_call(auto p1) noexcept
 }
 
 template <::std::uint_least64_t syscall_number>
+[[noreturn]]
 inline void system_call_no_return(auto p1) noexcept
 {
 	register ::std::uint_least64_t x8 __asm__("x8") = syscall_number;
@@ -43,6 +48,18 @@ inline return_value_type system_call(auto p1, auto p2) noexcept
 	return static_cast<return_value_type>(x0);
 }
 
+template <::std::uint_least64_t syscall_number, ::std::signed_integral return_value_type>
+	requires(1 < sizeof(return_value_type))
+[[__gnu__::__always_inline__]]
+inline return_value_type inline_syscall(auto p1, auto p2) noexcept
+{
+	register ::std::uint_least64_t x8 __asm__("x8") = syscall_number;
+	register ::std::uint_least64_t x0 __asm__("x0") = (::std::uint_least64_t)p1;
+	register ::std::uint_least64_t x1 __asm__("x1") = (::std::uint_least64_t)p2;
+	__asm__ __volatile__("svc 0" : "=r"(x0) : "r"(x8), "0"(x0), "r"(x1) : "memory", "cc");
+	return static_cast<return_value_type>(x0);
+}
+
 template <::std::uint_least64_t syscall_number, ::std::signed_integral return_value_type>
 	requires(1 < sizeof(return_value_type))
 inline return_value_type system_call(auto p1, auto p2, auto p3) noexcept
@@ -101,6 +118,7 @@ inline return_value_type system_call(auto p1, auto p2, auto p3, auto p4, auto p5
 }
 
 template <::std::integral I>
+[[noreturn]]
 inline void fast_exit(I ret) noexcept
 {
 	system_call_no_return<__NR_exit>(ret);
diff --git a/include/fast_io_hosted/platforms/linux/amd64.h b/include/fast_io_hosted/platforms/linux/amd64.h
index 5994ab9d..b6f097a5 100644
--- a/include/fast_io_hosted/platforms/linux/amd64.h
+++ b/include/fast_io_hosted/platforms/linux/amd64.h
@@ -1,6 +1,10 @@
 #pragma once
 // https://github.com/bminor/glibc/master/sysdeps/unix/sysv/linux/x86_64/sysdep.h
 
+#include <cstddef>
+#include <cstdint>
+#include <concepts>
+
 namespace fast_io
 {
 
@@ -31,6 +35,7 @@ inline return_value_type system_call(auto p1) noexcept
 }
 
 template <::std::uint_least64_t syscall_number>
+[[noreturn]]
 inline void system_call_no_return(auto p1) noexcept
 {
 	::std::uint_least64_t ret;
@@ -55,6 +60,20 @@ inline return_value_type system_call(auto p1, auto p2) noexcept
 	return ret;
 }
 
+template <::std::uint_least64_t syscall_number, ::std::signed_integral return_value_type>
+	requires(1 < sizeof(return_value_type))
+[[__gnu__::__always_inline__]]
+inline return_value_type inline_syscall(auto p1, auto p2) noexcept
+{
+	return_value_type ret;
+	__asm__ __volatile__("syscall"
+						 : "=a"(ret)
+						 // EDI      RSI       RDX
+						 : "0"(syscall_number), "D"(p1), "S"(p2)
+						 : "memory", "cc", "r11", "cx");
+	return ret;
+}
+
 template <::std::uint_least64_t syscall_number, ::std::signed_integral return_value_type>
 	requires(1 < sizeof(return_value_type))
 inline return_value_type system_call(auto p1, auto p2, auto p3) noexcept
@@ -114,6 +133,7 @@ inline return_value_type system_call(auto p1, auto p2, auto p3, auto p4, auto p5
 }
 
 template <::std::integral I>
+[[noreturn]]
 inline void fast_exit(I ret) noexcept
 {
 	system_call_no_return<__NR_exit>(ret);
diff --git a/include/fast_io_hosted/platforms/linux/generic.h b/include/fast_io_hosted/platforms/linux/generic.h
index 5cc440f8..794379e7 100644
--- a/include/fast_io_hosted/platforms/linux/generic.h
+++ b/include/fast_io_hosted/platforms/linux/generic.h
@@ -1,4 +1,8 @@
 #pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <concepts>
 #include <unistd.h>
 #include <sys/syscall.h>
 
@@ -13,7 +17,16 @@ inline return_value_type system_call(Args... args) noexcept
 	return static_cast<return_value_type>(ret);
 }
 
+template <::std::uint_least64_t syscall_number, ::std::signed_integral return_value_type>
+	requires(1 < sizeof(return_value_type))
+[[__gnu__::__always_inline__]]
+inline return_value_type inline_syscall(auto p1, auto p2) noexcept
+{
+	static_assert(false, "force inline for generic syscall is not supported");
+}
+
 template <::std::size_t syscall_number>
+[[noreturn]]
 inline void system_call_no_return(auto p1) noexcept
 {
 	::syscall(syscall_number, p1);
@@ -21,6 +34,7 @@ inline void system_call_no_return(auto p1) noexcept
 }
 
 template <::std::integral I>
+[[noreturn]]
 inline void fast_exit(I ret) noexcept
 {
 	system_call_no_return<__NR_exit>(ret);
diff --git a/include/fast_io_hosted/platforms/linux/loongarch64.h b/include/fast_io_hosted/platforms/linux/loongarch64.h
index d3dbe486..5a4426b5 100644
--- a/include/fast_io_hosted/platforms/linux/loongarch64.h
+++ b/include/fast_io_hosted/platforms/linux/loongarch64.h
@@ -5,6 +5,10 @@ glibc loongarch
 loongarch syscall is very similar to riscv. but it is using syscall instruction like x86_64
 */
 
+#include <cstddef>
+#include <cstdint>
+#include <concepts>
+
 namespace fast_io
 {
 
@@ -31,13 +35,14 @@ inline return_value_type system_call(auto p1) noexcept
 }
 
 template <::std::uint_least64_t syscall_number>
+[[noreturn]]
 inline void system_call_no_return(auto p1) noexcept
 {
 	register ::std::uint_least64_t a7 __asm__("$a7") = syscall_number;
 	register ::std::uint_least64_t a0 __asm__("$a0") = (::std::uint_least64_t)p1;
 	__asm__ __volatile__("syscall 0" : "+r"(a0) : "r"(a7) : "$t0", "$t1", "$t2", "$t3", "$t4",
 															"$t5", "$t6", "$t7", "$t8", "memory");
-	__builtin_unreachable(); 
+	__builtin_unreachable();
 }
 
 template <::std::uint_least64_t syscall_number, ::std::signed_integral return_value_type>
@@ -52,6 +57,19 @@ inline return_value_type system_call(auto p1, auto p2) noexcept
 	return static_cast<return_value_type>(a0);
 }
 
+template <::std::uint_least64_t syscall_number, ::std::signed_integral return_value_type>
+	requires(1 < sizeof(return_value_type))
+[[__gnu__::__always_inline__]]
+inline return_value_type inline_syscall(auto p1, auto p2) noexcept
+{
+	register ::std::uint_least64_t a7 __asm__("$a7") = syscall_number;
+	register ::std::uint_least64_t a0 __asm__("$a0") = (::std::uint_least64_t)p1;
+	register ::std::uint_least64_t a1 __asm__("$a1") = (::std::uint_least64_t)p2;
+	__asm__ __volatile__("syscall 0" : "+r"(a0) : "r"(a7), "r"(a1) : "$t0", "$t1", "$t2", "$t3", "$t4",
+																	 "$t5", "$t6", "$t7", "$t8", "memory");
+	return static_cast<return_value_type>(a0);
+}
+
 template <::std::uint_least64_t syscall_number, ::std::signed_integral return_value_type>
 	requires(1 < sizeof(return_value_type))
 inline return_value_type system_call(auto p1, auto p2, auto p3) noexcept
@@ -111,6 +129,7 @@ inline return_value_type system_call(auto p1, auto p2, auto p3, auto p4, auto p5
 }
 
 template <::std::integral I>
+[[noreturn]]
 inline void fast_exit(I ret) noexcept
 {
 	system_call_no_return<__NR_exit>(ret);
diff --git a/include/fast_io_hosted/platforms/linux/riscv64.h b/include/fast_io_hosted/platforms/linux/riscv64.h
index bdc1786d..5d547e45 100644
--- a/include/fast_io_hosted/platforms/linux/riscv64.h
+++ b/include/fast_io_hosted/platforms/linux/riscv64.h
@@ -5,6 +5,10 @@
 Do we need to deal with big endian with extra code???
 */
 
+#include <cstddef>
+#include <cstdint>
+#include <concepts>
+
 namespace fast_io
 {
 
@@ -29,6 +33,7 @@ inline return_value_type system_call(auto p1) noexcept
 }
 
 template <::std::uint_least64_t syscall_number>
+[[noreturn]]
 inline void system_call_no_return(auto p1) noexcept
 {
 	register ::std::uint_least64_t a7 __asm__("a7") = syscall_number;
@@ -48,6 +53,18 @@ inline return_value_type system_call(auto p1, auto p2) noexcept
 	return static_cast<return_value_type>(a0);
 }
 
+template <::std::uint_least64_t syscall_number, ::std::signed_integral return_value_type>
+	requires(1 < sizeof(return_value_type))
+[[__gnu__::__always_inline__]]
+inline return_value_type inline_syscall(auto p1, auto p2) noexcept
+{
+	register ::std::uint_least64_t a7 __asm__("a7") = syscall_number;
+	register ::std::uint_least64_t a0 __asm__("a0") = (::std::uint_least64_t)p1;
+	register ::std::uint_least64_t a1 __asm__("a1") = (::std::uint_least64_t)p2;
+	__asm__ __volatile__("ecall" : "+r"(a0) : "r"(a7), "r"(a1) : "memory");
+	return static_cast<return_value_type>(a0);
+}
+
 template <::std::uint_least64_t syscall_number, ::std::signed_integral return_value_type>
 	requires(1 < sizeof(return_value_type))
 inline return_value_type system_call(auto p1, auto p2, auto p3) noexcept
@@ -103,6 +120,7 @@ inline return_value_type system_call(auto p1, auto p2, auto p3, auto p4, auto p5
 }
 
 template <::std::integral I>
+[[noreturn]]
 inline void fast_exit(I ret) noexcept
 {
 	system_call_no_return<__NR_exit>(ret);