Merge branch 'aarch64' into 'develop'

Merge Aarch64 into develop

See merge request !11

Cargo.toml

@@ -21,6 +21,7 @@ ast = {path = "src/ast"}
 utils = {path = "src/utils"}
 gc = {path = "src/gc"}
 
+
 field-offset = "0.1.1"
 libloading = "0.3"
 lazy_static = "0.1.15"
@@ -34,4 +35,5 @@ time = "0.1.34"
 maplit = "0.1.4"
 docopt = "0.6"
 petgraph = "0.4.1"
-extprim = "*"
\ No newline at end of file
+extprim = "*"
+num-traits = "*"

ci.sh

+set -e
+if [ -z "$MU_ZEBU" ]
+then
+	export MU_ZEBU=$PWD
+fi
+export MU_LOG_LEVEL=none
+export RUST_TEST_THREADS=1
+export RUST_BACKTRACE=0
+export PYTHONPATH="$MU_ZEBU/tests/test_jit/mu-client-pypy/:$MU_ZEBU/tests/test_jit/RPySOM/src"
+export LD_LIBRARY_PATH="$MU_ZEBU/tests/test_jit/:$MU_ZEBU/tests/test_jit"
+export ZEBU_BUILD=release
+
+rm -rf $MU_ZEBU/emit
+rm -rf $MU_ZEBU/tests/test_jit/emit
+
+#cargo clean
+cargo test --release --no-run --color=always 2>&1 | tee build_out.txt
+
+/usr/bin/time -f "finished in %e secs" -a -o cargo_test_out.txt ./test-release --color=always 2>/dev/null | tee cargo_test_out.txt
+
+cd $MU_ZEBU/tests/test_jit/mu-client-pypy
+git pull
+
+cd $MU_ZEBU/tests/test_jit/RPySOM
+git pull
+
+cd $MU_ZEBU/tests/test_jit/
+ZEBU_BUILD=release LD_LIBRARY_PATH=. PYTHONPATH=mu-client-pypy:RPySOM/src pytest test_*.py -v --color=yes 2>&1 | tee $MU_ZEBU/pytest_out.txt
+

src/ast/src/ir.rs

@@ -735,6 +735,14 @@ impl Value {
         })
     }
 
+    pub fn is_int_ex_const(&self) -> bool {
+        match self.v {
+            Value_::Constant(Constant::IntEx(_)) => true,
+            _ => false
+        }
+    }
+
+
     pub fn is_int_const(&self) -> bool {
         match self.v {
             Value_::Constant(Constant::Int(_)) => true,
@@ -742,7 +750,13 @@ impl Value {
             _ => false
         }
     }
-    
+    pub fn is_fp_const(&self) -> bool {
+        match self.v {
+            Value_::Constant(Constant::Float(_)) => true,
+            Value_::Constant(Constant::Double(_)) => true,
+            _ => false
+        }
+    }
     pub fn extract_int_const(&self) -> u64 {
         match self.v {
             Value_::Constant(Constant::Int(val)) => val,
@@ -751,6 +765,13 @@ impl Value {
         }
     }
 
+    pub fn extract_int_ex_const(&self) -> Vec<u64> {
+        match self.v {
+            Value_::Constant(Constant::IntEx(ref val)) => val.clone(),
+            _ => panic!("expect int ex const")
+        }
+    }
+
     pub fn extract_ssa_id(&self) -> Option<MuID> {
         match self.v {
             Value_::SSAVar(id) => Some(id),

src/ast/src/op.rs

@@ -211,7 +211,7 @@ impl CmpOp {
             FUGT => FULT,
             FULT => FUGT,
 
-            _ => self, // all other comparisons are reflexive
+            _ => self, // all other comparisons are symmetric
         }
     }
     pub fn invert(self) -> CmpOp {
@@ -257,6 +257,18 @@ impl CmpOp {
             FTRUE => FFALSE,
         }
     }
+
+    // gets the unsigned version of the comparison
+    pub fn get_unsigned(self) -> CmpOp {
+        use op::CmpOp::*;
+        match self {
+            SGE => UGE,
+            SLT => ULT,
+            SGT => UGT,
+            SLE => ULE,
+            _   => self,
+        }
+    }
     pub fn is_signed(self) -> bool {
         use op::CmpOp::*;
         match self {
@@ -264,6 +276,14 @@ impl CmpOp {
             _ => false
         }
     }
+
+    pub fn is_symmetric(self) -> bool {
+        use op::CmpOp::*;
+        match self {
+            EQ | NE | FORD| FUNO| FUNE | FUEQ | FONE | FOEQ => true,
+            _ => false
+        }
+    }
 }
 
 #[derive(Copy, Clone, Debug, PartialEq, RustcEncodable, RustcDecodable)]

src/compiler/backend/arch/aarch64/codegen.rs

@@ -23,8 +23,10 @@ pub trait CodeGenerator {
     fn end_block(&mut self, block_name: MuName);
 
     // add CFI info
+    fn add_cfi_sections(&mut self, arg: &str);
     fn add_cfi_startproc(&mut self);
     fn add_cfi_endproc(&mut self);
+    fn add_cfi_def_cfa(&mut self, reg: Reg, offset: i32);
     fn add_cfi_def_cfa_register(&mut self, reg: Reg);
     fn add_cfi_def_cfa_offset(&mut self, offset: i32);
     fn add_cfi_offset(&mut self, reg: Reg, offset: i32);
@@ -35,6 +37,10 @@ pub trait CodeGenerator {
     fn emit_frame_grow(&mut self); // Emits a SUB
     fn emit_frame_shrink(&mut self); // Emits an ADD
 
+    // Used to pass a string that the assembler will interpret as an immediate argument
+    // (This is neccesary to support the use of ELF relocations like ':tprel_hi12:foo')
+    fn emit_add_str(&mut self, dest: Reg, src1: Reg, src2: &str);
+
     // stack minimpulation
     fn emit_push_pair(&mut self, src1: Reg, src2: Reg, stack: Reg); // Emits a STP
     fn emit_pop_pair(&mut self, dest1: Reg, dest2: Reg, stack: Reg); // Emits a LDP

src/compiler/frame.rs

@@ -186,7 +186,7 @@ impl FrameSlot {
                     base: aarch64::FP.clone(),
                     offset: Some(Value::make_int_const(vm.next_id(), self.offset as u64)),
                     scale: 1,
-                    signed: false
+                    signed: true
                 }
             )
         })

src/runtime/entrypoints.rs

@@ -144,8 +144,8 @@ lazy_static! {
     pub static ref UDIV_U128 : RuntimeEntrypoint = RuntimeEntrypoint {
         sig: P(MuFuncSig {
             hdr: MuEntityHeader::unnamed(ir::new_internal_id()),
-            ret_tys: vec![UINT64_TYPE.clone(); 2],
-            arg_tys: vec![UINT64_TYPE.clone(); 4]
+            ret_tys: vec![UINT128_TYPE.clone(); 1],
+            arg_tys: vec![UINT128_TYPE.clone(); 2]
         }),
         aot: ValueLocation::Relocatable(RegGroup::GPR, String::from("muentry_udiv_u128")),
         jit: RwLock::new(None)
@@ -154,8 +154,8 @@ lazy_static! {
     pub static ref SDIV_I128 : RuntimeEntrypoint = RuntimeEntrypoint {
         sig: P(MuFuncSig {
             hdr: MuEntityHeader::unnamed(ir::new_internal_id()),
-            ret_tys: vec![UINT64_TYPE.clone(); 2],
-            arg_tys: vec![UINT64_TYPE.clone(); 4]
+            ret_tys: vec![UINT128_TYPE.clone(); 1],
+            arg_tys: vec![UINT128_TYPE.clone(); 2]
         }),
         aot: ValueLocation::Relocatable(RegGroup::GPR, String::from("muentry_sdiv_i128")),
         jit: RwLock::new(None)
@@ -164,8 +164,8 @@ lazy_static! {
     pub static ref UREM_U128 : RuntimeEntrypoint = RuntimeEntrypoint {
         sig: P(MuFuncSig {
             hdr: MuEntityHeader::unnamed(ir::new_internal_id()),
-            ret_tys: vec![UINT64_TYPE.clone(); 2],
-            arg_tys: vec![UINT64_TYPE.clone(); 4]
+            ret_tys: vec![UINT128_TYPE.clone(); 1],
+            arg_tys: vec![UINT128_TYPE.clone(); 2]
         }),
         aot: ValueLocation::Relocatable(RegGroup::GPR, String::from("muentry_urem_u128")),
         jit: RwLock::new(None)
@@ -174,13 +174,92 @@ lazy_static! {
     pub static ref SREM_I128 : RuntimeEntrypoint = RuntimeEntrypoint {
         sig: P(MuFuncSig {
             hdr: MuEntityHeader::unnamed(ir::new_internal_id()),
-            ret_tys: vec![UINT64_TYPE.clone(); 2],
-            arg_tys: vec![UINT64_TYPE.clone(); 4]
+            ret_tys: vec![UINT128_TYPE.clone(); 1],
+            arg_tys: vec![UINT128_TYPE.clone(); 2]
         }),
         aot: ValueLocation::Relocatable(RegGroup::GPR, String::from("muentry_srem_i128")),
         jit: RwLock::new(None)
     };
 
+    pub static ref FPTOUI_DOUBLE_U128 : RuntimeEntrypoint = RuntimeEntrypoint {
+        sig: P(MuFuncSig {
+            hdr: MuEntityHeader::unnamed(ir::new_internal_id()),
+            ret_tys: vec![UINT128_TYPE.clone(); 1],
+            arg_tys: vec![DOUBLE_TYPE.clone(); 1]
+        }),
+        aot: ValueLocation::Relocatable(RegGroup::GPR, String::from("muentry_fptoui_double_u128")),
+        jit: RwLock::new(None)
+    };
+
+    pub static ref FPTOSI_DOUBLE_I128 : RuntimeEntrypoint = RuntimeEntrypoint {
+        sig: P(MuFuncSig {
+            hdr: MuEntityHeader::unnamed(ir::new_internal_id()),
+            ret_tys: vec![UINT128_TYPE.clone(); 1],
+            arg_tys: vec![DOUBLE_TYPE.clone(); 1]
+        }),
+        aot: ValueLocation::Relocatable(RegGroup::GPR, String::from("muentry_fptosi_double_i128")),
+        jit: RwLock::new(None)
+    };
+    pub static ref UITOFP_U128_DOUBLE : RuntimeEntrypoint = RuntimeEntrypoint {
+        sig: P(MuFuncSig {
+            hdr: MuEntityHeader::unnamed(ir::new_internal_id()),
+            ret_tys: vec![DOUBLE_TYPE.clone(); 1],
+            arg_tys: vec![UINT128_TYPE.clone(); 1]
+        }),
+        aot: ValueLocation::Relocatable(RegGroup::GPR, String::from("muentry_uitofp_u128_double")),
+        jit: RwLock::new(None)
+    };
+
+    pub static ref SITOFP_I128_DOUBLE : RuntimeEntrypoint = RuntimeEntrypoint {
+        sig: P(MuFuncSig {
+            hdr: MuEntityHeader::unnamed(ir::new_internal_id()),
+            ret_tys: vec![DOUBLE_TYPE.clone(); 1],
+            arg_tys: vec![UINT128_TYPE.clone(); 1]
+        }),
+        aot: ValueLocation::Relocatable(RegGroup::GPR, String::from("muentry_sitofp_i128_double")),
+        jit: RwLock::new(None)
+    };
+
+    // Conversion to/from int<128> from/to float
+        pub static ref FPTOUI_FLOAT_U128 : RuntimeEntrypoint = RuntimeEntrypoint {
+        sig: P(MuFuncSig {
+            hdr: MuEntityHeader::unnamed(ir::new_internal_id()),
+            ret_tys: vec![UINT128_TYPE.clone(); 1],
+            arg_tys: vec![FLOAT_TYPE.clone(); 1]
+        }),
+        aot: ValueLocation::Relocatable(RegGroup::GPR, String::from("muentry_fptoui_float_u128")),
+        jit: RwLock::new(None)
+    };
+
+    pub static ref FPTOSI_FLOAT_I128 : RuntimeEntrypoint = RuntimeEntrypoint {
+        sig: P(MuFuncSig {
+            hdr: MuEntityHeader::unnamed(ir::new_internal_id()),
+            ret_tys: vec![UINT128_TYPE.clone(); 1],
+            arg_tys: vec![FLOAT_TYPE.clone(); 1]
+        }),
+        aot: ValueLocation::Relocatable(RegGroup::GPR, String::from("muentry_fptosi_float_i128")),
+        jit: RwLock::new(None)
+    };
+    pub static ref UITOFP_U128_FLOAT : RuntimeEntrypoint = RuntimeEntrypoint {
+        sig: P(MuFuncSig {
+            hdr: MuEntityHeader::unnamed(ir::new_internal_id()),
+            ret_tys: vec![FLOAT_TYPE.clone(); 1],
+            arg_tys: vec![UINT128_TYPE.clone(); 1]
+        }),
+        aot: ValueLocation::Relocatable(RegGroup::GPR, String::from("muentry_uitofp_u128_float")),
+        jit: RwLock::new(None)
+    };
+
+    pub static ref SITOFP_I128_FLOAT : RuntimeEntrypoint = RuntimeEntrypoint {
+        sig: P(MuFuncSig {
+            hdr: MuEntityHeader::unnamed(ir::new_internal_id()),
+            ret_tys: vec![FLOAT_TYPE.clone(); 1],
+            arg_tys: vec![UINT128_TYPE.clone(); 1]
+        }),
+        aot: ValueLocation::Relocatable(RegGroup::GPR, String::from("muentry_sitofp_i128_float")),
+        jit: RwLock::new(None)
+    };
+
     // impl/decl: mod.rs
     pub static ref PRINT_HEX : RuntimeEntrypoint = RuntimeEntrypoint {
         sig: P(MuFuncSig {

src/runtime/exception_aarch64.rs

@@ -5,7 +5,6 @@ use utils::Address;
 use utils::Word;
 use utils::POINTER_SIZE;
 use runtime::thread;
-use runtime::PRINT_BACKTRACE;
 
 use std::sync::RwLock;
 use std::sync::RwLockReadGuard;
@@ -45,7 +44,9 @@ pub extern fn throw_exception_internal(exception_obj: Address, last_frame_callee
     trace!("throwing exception: {}", exception_obj);
 
     trace!("callee saved registers of last frame is saved at {}", last_frame_callee_saved);
-    inspect_higher_address(last_frame_callee_saved, 20);
+    if cfg!(debug_assertions) {
+        inspect_higher_address(last_frame_callee_saved, 20);
+    }
 
     let mut cur_thread = thread::MuThread::current_mut();
     // set exception object
@@ -100,7 +101,7 @@ pub extern fn throw_exception_internal(exception_obj: Address, last_frame_callee
         }
     };
 
-    if PRINT_BACKTRACE {
+    if cfg!(debug_assertions) {
         print_backtrace(throw_frame_callsite, cursor.clone());
     }
 

src/runtime/exception_x64.rs

@@ -45,8 +45,10 @@ pub extern fn throw_exception_internal(exception_obj: Address, last_frame_callee
     trace!("throwing exception: {}", exception_obj);
 
     trace!("callee saved registers of last frame is saved at {}", last_frame_callee_saved);
-    inspect_nearby_address(last_frame_callee_saved, 8);
-    
+    if cfg!(debug_assertions) {
+        inspect_nearby_address(last_frame_callee_saved, 8);
+    }
+
     let mut cur_thread = thread::MuThread::current_mut();
     // set exception object
     cur_thread.exception_obj = exception_obj;
@@ -84,7 +86,7 @@ pub extern fn throw_exception_internal(exception_obj: Address, last_frame_callee
         }
     };
 
-    if PRINT_BACKTRACE {
+    if cfg!(debug_assertions) {
         print_backtrace(throw_frame_callsite, cursor.clone());
     }
     

src/runtime/math.rs

@@ -14,29 +14,46 @@ pub extern fn muentry_frem64(a: f64, b: f64) -> f64 {
     a.rem(b)
 }
 
+extern crate num_traits;
 use extprim::u128::u128;
 use extprim::i128::i128;
+use runtime::math::num_traits::ToPrimitive;
+use runtime::math::num_traits::FromPrimitive;
 
 #[no_mangle]
-#[allow(unreachable_code)]
 pub extern fn muentry_udiv_u128(a: u128, b: u128) -> u128 {
     a.wrapping_div(b)
 }
 
 #[no_mangle]
-#[allow(unreachable_code)]
 pub extern fn muentry_sdiv_i128(a: i128, b: i128) -> i128 {
     a.wrapping_div(b)
 }
 
 #[no_mangle]
-#[allow(unreachable_code)]
 pub extern fn muentry_urem_u128(a: u128, b: u128) -> u128 {
     a.wrapping_rem(b)
 }
 
 #[no_mangle]
-#[allow(unreachable_code)]
 pub extern fn muentry_srem_i128(a: i128, b: i128) -> i128 {
     a.wrapping_rem(b)
-}
\ No newline at end of file
+}
+
+#[no_mangle]
+pub extern fn muentry_fptoui_double_u128(a: f64) -> u128 { u128::from_f64(a).unwrap() }
+#[no_mangle]
+pub extern fn muentry_fptosi_double_i128(a: f64) -> i128 { i128::from_f64(a).unwrap() }
+#[no_mangle]
+pub extern fn muentry_uitofp_u128_double(a: u128) -> f64 { a.to_f64().unwrap() }
+#[no_mangle]
+pub extern fn muentry_sitofp_i128_double(a: i128) -> f64 { a.to_f64().unwrap() }
+
+#[no_mangle]
+pub extern fn muentry_fptoui_float_u128(a: f32) -> u128 { u128::from_f32(a).unwrap() }
+#[no_mangle]
+pub extern fn muentry_fptosi_float_i128(a: f32) -> i128 { i128::from_f32(a).unwrap() }
+#[no_mangle]
+pub extern fn muentry_uitofp_u128_float(a: u128) -> f32 { a.to_f32().unwrap() }
+#[no_mangle]
+pub extern fn muentry_sitofp_i128_float(a: i128) -> f32 { a.to_f32().unwrap() }
\ No newline at end of file

src/runtime/swap_stack_aarch64_sysv.S

@@ -24,23 +24,41 @@ begin_func swap_to_mu_stack
           # -- on new stack --
           # arguments (reverse order of thread.rs - runtime_load_args)
           pop_arguments
-          //enter_frame // Unncesary Note saves the return address of the function that called this function
 
-          # If entry tries to return (IT SHOULDN'T) it will return to entry_returned
+          # If entry tries to return (IT SHOULDN'T) it will call entry_returned (which will panic!)
           ADR LR, entry_returned
 
           # branch to entry
           RET X9
 end_func swap_to_mu_stack
 
-# For debuging only, executes
+# The error message to print when entry returns
+	.type .Lentry_returned_message, @object
+	.section .rodata.str1.1,"aMS",@progbits,1
+.Lentry_returned_message:
+        .asciz  "ERROR: The entry function returned\n"
+        .size   .Lentry_returned_message, 36
+
+# Prints an error mesage and calls exit(1)
 begin_func entry_returned
-          HLT #0
-end_func entry_returned
+    ADRP x0, .Lentry_returned_message
+    ADD x0, x0, :lo12:.Lentry_returned_message // Loads x0 with the adress of .Lentry_returned_message
+
+    MOV X1, #35 // Length of string
 
+	MOV X2, #1 // Number of elements
+
+    ADRP X3, :got:stderr
+    LDR X3, [X3, :got_lo12:stderr]
+    LDR X3, [X3] // Load X3 with the address of stderr
+
+    BL fwrite // calls fwrite(x0, x1, x2, x3)
+    MOV W0, 1
+    BL exit // call exit(1)
+end_func entry_returned
 
-# _swap_back_to_native_stack(sp_loc: Address)
-#                            X0
+// _swap_back_to_native_stack(sp_loc: Address)
+//                            X0
 begin_func muentry_swap_back_to_native_stack
           # SP = *sp_loc
           LDR X9, [X0]

src/runtime/thread.rs

@@ -127,6 +127,7 @@ impl MuStack {
             match reg_group {
                 RegGroup::GPR => gpr_used.push(word),
                 RegGroup::FPR => fpr_used.push(word),
+                RegGroup::GPREX => unimplemented!(),
             }
         }
 

tests/test_compiler/test_call.rs

@@ -1291,4 +1291,96 @@ fn store_funcref() -> VM {
     });
 
     vm
+}
+
+use test_compiler::test_int128::add_u128;
+
+#[test]
+fn test_call_int128_arg() {
+    VM::start_logging_trace();
+
+    let vm = Arc::new(add_u128());
+    call_add_u128(&vm);
+
+    let func_add  = vm.id_of("add_u128");
+    let func_call = vm.id_of("call_add_u128");
+
+    let compiler = Compiler::new(CompilerPolicy::default(), &vm);
+    {
+        let funcs = vm.funcs().read().unwrap();
+        let func_vers = vm.func_vers().read().unwrap();
+
+        {
+            let func = funcs.get(&func_add).unwrap().read().unwrap();
+            let mut func_ver = func_vers.get(&func.cur_ver.unwrap()).unwrap().write().unwrap();
+
+            compiler.compile(&mut func_ver);
+        }
+        {
+            let func = funcs.get(&func_call).unwrap().read().unwrap();
+            let mut func_ver = func_vers.get(&func.cur_ver.unwrap()).unwrap().write().unwrap();
+
+            compiler.compile(&mut func_ver);
+        }
+    }
+
+    vm.make_primordial_thread(func_call, true, vec![]);
+    backend::emit_context(&vm);
+
+    let executable = aot::link_primordial(vec![Mu("add_u128"), Mu("call_add_u128")], "test_call_int128_arg", &vm);
+    let output = aot::execute_nocheck(executable);
+
+    // exit with (84)
+    assert!(output.status.code().is_some());
+    assert_eq!(output.status.code().unwrap(), 84);
+}
+
+fn call_add_u128(vm: &VM) {
+    let add_u128_sig = vm.get_func_sig(vm.id_of("sig"));
+    let add_u128_id  = vm.id_of("add_u128");
+
+    typedef!    ((vm) int64  = mu_int(64));
+    typedef!    ((vm) int128 = mu_int(128));
+
+    constdef!   ((vm) <int128> int128_42 = Constant::IntEx(vec![42, 0]));
+
+    typedef!    ((vm) funcref_add_u128 = mu_funcref(add_u128_sig));
+    constdef!   ((vm) <funcref_add_u128> const_funcref_add_u128 = Constant::FuncRef(add_u128_id));
+
+    funcsig!    ((vm) call_add_u128_sig = () -> ());
+    funcdecl!   ((vm) <call_add_u128_sig> call_add_u128);
+    funcdef!    ((vm) <call_add_u128_sig> call_add_u128 VERSION call_add_u128_v1);
+
+    // blk_entry
+    block!      ((vm, call_add_u128_v1) blk_entry);
+
+    // EXPRCALL add_u128 (42, 42)
+    ssa!        ((vm, call_add_u128_v1) <int128> res);
+    consta!     ((vm, call_add_u128_v1) int128_42_local = int128_42);
+    consta!     ((vm, call_add_u128_v1) const_funcref_add_u128_local = const_funcref_add_u128);
+    inst!       ((vm, call_add_u128_v1) blk_entry_call:
+        res = EXPRCALL (CallConvention::Mu, is_abort: false) const_funcref_add_u128_local (int128_42_local, int128_42_local)
+    );
+
+    ssa!        ((vm, call_add_u128_v1) <int64> trunc_res);
+    inst!       ((vm, call_add_u128_v1) blk_entry_trunc:
+        trunc_res = CONVOP (ConvOp::TRUNC) <int128 int64> res
+    );
+
+    let blk_entry_exit = gen_ccall_exit(trunc_res.clone(), &mut call_add_u128_v1, &vm);
+
+    inst!       ((vm, call_add_u128_v1) blk_entry_ret:
+        RET
+    );
+
+    define_block!((vm, call_add_u128_v1) blk_entry() {
+        blk_entry_call,
+        blk_entry_trunc,
+        blk_entry_exit,
+        blk_entry_ret
+    });
+
+    define_func_ver!((vm) call_add_u128_v1 (entry: blk_entry) {
+        blk_entry
+    });
 }
\ No newline at end of file

tests/test_compiler/test_int128.rs

@@ -29,8 +29,8 @@ fn test_add_u128() {
     }
 }
 
-fn add_u128() -> VM {
-    let vm = VM::new();
+pub fn add_u128() -> VM {
+    let vm = VM::new_with_opts("init_mu --disable-inline");
 
     typedef!    ((vm) u128 = mu_int(128));
 
@@ -61,6 +61,56 @@ fn add_u128() -> VM {
     vm
 }
 
+#[test]
+fn test_sub_u128() {
+    let lib = testutil::compile_fnc("sub_u128", &sub_u128);
+
+    unsafe {
+        use std::u64;
+
+        let sub_u128 : libloading::Symbol<unsafe extern fn(u64, u64, u64, u64) -> (u64, u64)> = lib.get(b"sub_u128").unwrap();
+
+        let res = sub_u128(1, 0, 1, 0);
+        println!("sub_u128(1, 1) = {:?}", res);
+        assert!(res == (0, 0));
+
+        let res = sub_u128(u64::MAX, 0, u64::MAX, u64::MAX);
+        println!("sub_u128(u64::MAX, -1) = {:?}", res);
+        assert!(res == (0, 1));
+    }
+}
+
+fn sub_u128() -> VM {
+    let vm = VM::new();
+
+    typedef!    ((vm) u128 = mu_int(128));
+
+    funcsig!    ((vm) sig = (u128, u128) -> (u128));
+    funcdecl!   ((vm) <sig> sub_u128);
+    funcdef!    ((vm) <sig> sub_u128 VERSION sub_u128_v1);
+
+    block!      ((vm, sub_u128_v1) blk_entry);
+    ssa!        ((vm, sub_u128_v1) <u128> a);
+    ssa!        ((vm, sub_u128_v1) <u128> b);
+
+    // sum = sub %a %b
+    ssa!        ((vm, sub_u128_v1) <u128> sum);
+    inst!       ((vm, sub_u128_v1) blk_entry_sub_u128:
+        sum = BINOP (BinOp::Sub) a b
+    );
+
+    inst!       ((vm, sub_u128_v1) blk_entry_ret:
+        RET (sum)
+    );
+
+    define_block!   ((vm, sub_u128_v1) blk_entry(a, b) {
+        blk_entry_sub_u128, blk_entry_ret
+    });
+
+    define_func_ver!((vm) sub_u128_v1 (entry: blk_entry) {blk_entry});
+
+    vm
+}
 #[test]
 fn test_add_const_u128() {
     let lib = testutil::compile_fnc("add_const_u128", &add_const_u128);