diff --git a/src/arch/x86_64/encode.rs b/src/arch/x86_64/encode.rs
index 456abf2..9499c62 100644
--- a/src/arch/x86_64/encode.rs
+++ b/src/arch/x86_64/encode.rs
@@ -172,9 +172,35 @@ impl Code {
         self.bytes.push(0xc3);
     }
 
-    pub fn sub(&mut self) {
-        // sub esp 40 iirc
-        self.bytes.extend([0x48, 0x83, 0xec, 0x28]);
+    pub fn sub(&mut self, mut dst: Reg, src: impl Into<Imm>) -> ERes {
+        let mut src = src.into();
+        let mut width = src.width_signed()?;
+        let dst_width = dst.width().min(Width::B32);
+
+        self.prefix16(dst_width);
+        self.rex(dst, 0, 0, dst);
+
+        if width > dst_width {
+            width = src.width_unsigned()?;
+            if dst.width() == Width::B64 || width > dst_width {
+                return Err("immediate overflow".into());
+            }
+            src = src.reinterpret(dst_width);
+            width = src.width_signed()?;
+        }
+        if dst.width() == Width::B8 {
+            self.bytes.push(0x80);
+        } else {
+            if width == Width::B8 {
+                self.bytes.push(0x83);
+            } else {
+                self.bytes.push(0x81);
+                width = dst_width;
+            }
+        }
+        self.bytes.push(modrm(0b11, 0b101, dst.base()));
+        self.imm(src, width);
+        Ok(())
     }
 
     fn prefix16(&mut self, width: impl Into<Width>) {
diff --git a/src/arch/x86_64/test/asm/mod.rs b/src/arch/x86_64/test/asm/mod.rs
new file mode 100644
index 0000000..e84f419
--- /dev/null
+++ b/src/arch/x86_64/test/asm/mod.rs
@@ -0,0 +1,48 @@
+mod setup;
+use setup::*;
+
+#[test]
+fn mov() {
+    let c = &mut TestCtx::new("mov");
+
+    for dst in regs() {
+        for src in regs() {
+            eq(c, format!("mov {dst}, {src}"), |c| c.mov(dst, src));
+        }
+    }
+
+    for dst in regs() {
+        for src in mems() {
+            eq(c, format!("mov {dst}, {src}"), |c| c.mov(dst, src));
+        }
+    }
+
+    for dst in regs() {
+        for src in imms() {
+            eq(c, format!("mov {dst}, {src}"), |c| c.mov(dst, src));
+        }
+    }
+
+    for dst in mems() {
+        for src in regs() {
+            eq(c, format!("mov {dst}, {src}"), |c| c.mov(dst, src));
+        }
+    }
+
+    for dst in mems() {
+        for src in imms() {
+            eq(c, format!("mov {dst}, {src}"), |c| c.mov(dst, src));
+        }
+    }
+}
+
+#[test]
+fn sub() {
+    let c = &mut TestCtx::new("mov");
+
+    for dst in regs() {
+        for src in imms() {
+            eq(c, format!("sub {dst}, {src}"), |c| c.sub(dst, src))
+        }
+    }
+}
diff --git a/src/arch/x86_64/test/nasm.rs b/src/arch/x86_64/test/asm/setup.rs
similarity index 81%
rename from src/arch/x86_64/test/nasm.rs
rename to src/arch/x86_64/test/asm/setup.rs
index 2b533c8..3050a40 100644
--- a/src/arch/x86_64/test/nasm.rs
+++ b/src/arch/x86_64/test/asm/setup.rs
@@ -32,50 +32,15 @@ const IMMS: &[i128] = &[
 
 const WIDTHS: &[Width] = &[Width::B8, Width::B16, Width::B32, Width::B64];
 
-#[test]
-fn mov() {
-    let c = &mut TestCtx::new("mov");
-
-    for dst in regs() {
-        for src in regs() {
-            eq(c, format!("mov {dst}, {src}"), |c| c.mov(dst, src));
-        }
-    }
-
-    for dst in regs() {
-        for src in mems() {
-            eq(c, format!("mov {dst}, {src}"), |c| c.mov(dst, src));
-        }
-    }
-
-    for dst in regs() {
-        for src in imms() {
-            eq(c, format!("mov {dst}, {src}"), |c| c.mov(dst, src));
-        }
-    }
-
-    for dst in mems() {
-        for src in regs() {
-            eq(c, format!("mov {dst}, {src}"), |c| c.mov(dst, src));
-        }
-    }
-
-    for dst in mems() {
-        for src in imms() {
-            eq(c, format!("mov {dst}, {src}"), |c| c.mov(dst, src));
-        }
-    }
-}
-
-fn imms() -> impl Iterator<Item = i128> {
+pub fn imms() -> impl Iterator<Item = i128> {
     IMMS.iter().cloned()
 }
 
-fn regs() -> impl Iterator<Item = Reg> {
+pub fn regs() -> impl Iterator<Item = Reg> {
     Reg::IMPORTANT.iter().cloned()
 }
 
-fn mems() -> impl Iterator<Item = Mem> {
+pub fn mems() -> impl Iterator<Item = Mem> {
     gen move {
         for &reg in Reg::IMPORTANT {
             for &disp in DISPS {
@@ -87,14 +52,14 @@ fn mems() -> impl Iterator<Item = Mem> {
     }
 }
 
-struct TestCtx {
+pub struct TestCtx {
     path: String,
     code: Code,
     cache: HashMap<String, Result<Vec<u8>, String>>,
     changed: bool,
 }
 
-fn eq(
+pub fn eq(
     ctx: &mut TestCtx,
     asm: impl AsRef<str>,
     instr: impl FnOnce(&mut Code) -> Result<(), CompilerMsg>,
@@ -176,7 +141,7 @@ fn write(path: &str, binary: &[u8]) {
 const CACHE_PATH: &str = "test/nasm_cache";
 
 impl TestCtx {
-    fn new(name: &str) -> Self {
+    pub fn new(name: &str) -> Self {
         let path = CACHE_PATH.to_string() + "/" + name;
         let cache = match std::fs::read(&path) {
             Ok(bytes) => bitcode::decode(&bytes).unwrap_or_default(),
diff --git a/src/arch/x86_64/test/full/windows.rs b/src/arch/x86_64/test/full/windows.rs
index 7c9b44c..478292e 100644
--- a/src/arch/x86_64/test/full/windows.rs
+++ b/src/arch/x86_64/test/full/windows.rs
@@ -11,7 +11,7 @@ fn hello() -> Result<(), CompilerMsg> {
     let entry = program.func(
         "main",
         [Instr::Asm(encode(|c| {
-            c.sub();
+            c.sub(esp, 0x28)?;
             // stdout
             c.mov(ecx, -11)?;
             c.call_mem(get_std_handle);
diff --git a/src/arch/x86_64/test/mod.rs b/src/arch/x86_64/test/mod.rs
index 3a6a91a..b47d229 100644
--- a/src/arch/x86_64/test/mod.rs
+++ b/src/arch/x86_64/test/mod.rs
@@ -1,2 +1,2 @@
 mod full;
-mod nasm;
+mod asm;
diff --git a/src/arch/x86_64/types.rs b/src/arch/x86_64/types.rs
index 664cdb6..3f592b6 100644
--- a/src/arch/x86_64/types.rs
+++ b/src/arch/x86_64/types.rs
@@ -53,6 +53,15 @@ impl Imm {
             _ => return Err(Self::overflow_msg()),
         })
     }
+
+    pub fn reinterpret(&self, width: Width) -> Self {
+        Self(match width {
+            Width::B8 => self.0 as i8 as i128,
+            Width::B16 => self.0 as i16 as i128,
+            Width::B32 => self.0 as i32 as i128,
+            Width::B64 => self.0 as i64 as i128,
+        })
+    }
 }
 
 impl TryFrom<Imm> for u8 {
diff --git a/test/bin/x86_64_test.exe b/test/bin/x86_64_test.exe
index 6e01c6a..ea88623 100755
Binary files a/test/bin/x86_64_test.exe and b/test/bin/x86_64_test.exe differ
diff --git a/test/nasm_cache/mov b/test/nasm_cache/mov
index e53395c..b65b4cb 100644
Binary files a/test/nasm_cache/mov and b/test/nasm_cache/mov differ