diff --git a/hblang/src/codegen.rs b/hblang/src/codegen.rs
index 9bf7462..b63b560 100644
--- a/hblang/src/codegen.rs
+++ b/hblang/src/codegen.rs
@@ -371,12 +371,18 @@ impl Func {
     }
 }
 
-#[derive(Default, PartialEq, Eq, Debug)]
+#[derive(Default, PartialEq, Eq)]
 pub struct RegAlloc {
     free:     Vec<Reg>,
     max_used: Reg,
 }
 
+impl std::fmt::Debug for RegAlloc {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("RegAlloc").finish()
+    }
+}
+
 impl RegAlloc {
     fn init_callee(&mut self) {
         self.free.clear();
@@ -564,18 +570,19 @@ impl<'a> Codegen<'a> {
                     }
 
                     let fn_label = self.labels[frame.label as usize].clone();
+                    self.gpa.borrow_mut().init_callee();
 
                     log::dbg!("fn-args");
                     let mut parama = 3..12;
                     for (arg, &ty) in args.iter().zip(fn_label.args.iter()) {
-                        let loc = self.load_arg(ty, &mut parama);
+                        let refed = arg.last.is_some_and(|l| l.get() & parser::REFERENCED != 0);
+                        let loc = self.load_arg(refed, ty, &mut parama);
                         self.vars.push(Variable {
                             id:    arg.id,
                             value: Value { ty, loc },
                         });
                     }
 
-                    self.gpa.borrow_mut().init_callee();
                     self.ret = fn_label.ret;
 
                     log::dbg!("fn-body");
@@ -1630,10 +1637,15 @@ impl<'a> Codegen<'a> {
         }
     }
 
-    fn load_arg(&mut self, ty: Type, parama: &mut Range<u8>) -> Loc {
+    fn load_arg(&mut self, referenced: bool, ty: Type, parama: &mut Range<u8>) -> Loc {
         let size = self.size_of(ty);
         match size {
             0 => Loc::Imm(0),
+            ..=8 if !referenced => {
+                let reg = self.alloc_reg();
+                self.code.encode(instrs::cp(reg.0, parama.next().unwrap()));
+                Loc::Reg(reg)
+            }
             ..=8 => {
                 let stack = self.alloc_stack(size as _);
                 self.store_stack(parama.next().unwrap(), stack.offset, size as _);
diff --git a/hblang/tests/codegen_tests_fb_driver.txt b/hblang/tests/codegen_tests_fb_driver.txt
index f79bef2..83862c7 100644
--- a/hblang/tests/codegen_tests_fb_driver.txt
+++ b/hblang/tests/codegen_tests_fb_driver.txt
@@ -1,3 +1,3 @@
-code size: 758
+code size: 678
 ret: 0
 status: Ok(())
diff --git a/hblang/tests/codegen_tests_functions.txt b/hblang/tests/codegen_tests_functions.txt
index 9f1d250..190c4f1 100644
--- a/hblang/tests/codegen_tests_functions.txt
+++ b/hblang/tests/codegen_tests_functions.txt
@@ -1,3 +1,3 @@
-code size: 388
+code size: 320
 ret: 33
 status: Ok(())
diff --git a/hblang/tests/codegen_tests_if_statements.txt b/hblang/tests/codegen_tests_if_statements.txt
index 5fadd38..cda3ca7 100644
--- a/hblang/tests/codegen_tests_if_statements.txt
+++ b/hblang/tests/codegen_tests_if_statements.txt
@@ -1,3 +1,3 @@
-code size: 347
+code size: 293
 ret: 55
 status: Ok(())
diff --git a/hblang/tests/codegen_tests_loops.txt b/hblang/tests/codegen_tests_loops.txt
index 5fadd38..23eb54c 100644
--- a/hblang/tests/codegen_tests_loops.txt
+++ b/hblang/tests/codegen_tests_loops.txt
@@ -1,3 +1,3 @@
-code size: 347
+code size: 296
 ret: 55
 status: Ok(())
diff --git a/hblang/tests/codegen_tests_pointers.txt b/hblang/tests/codegen_tests_pointers.txt
index 5216a64..9417ba6 100644
--- a/hblang/tests/codegen_tests_pointers.txt
+++ b/hblang/tests/codegen_tests_pointers.txt
@@ -1,3 +1,3 @@
-code size: 437
+code size: 382
 ret: 0
 status: Ok(())
diff --git a/hblang/tests/codegen_tests_structs.txt b/hblang/tests/codegen_tests_structs.txt
index db6996d..e81084a 100644
--- a/hblang/tests/codegen_tests_structs.txt
+++ b/hblang/tests/codegen_tests_structs.txt
@@ -1,3 +1,3 @@
-code size: 530
+code size: 483
 ret: 3
 status: Ok(())