Support 128-bit integers on platforms without native support (#103)

* Use sized integer types

* Add support for integer types not supported on some platforms

* Add feature to test non-native integers in CI

17 files changed, 1215 insertions, 350 deletions

diff --git a/.github/workflows/main.yml b/.github/workflows/ci.yml
index 98bed8ef387..337837c40bf 100644
--- a/.github/workflows/main.yml
+++ b/.github/workflows/ci.yml
@@ -10,10 +10,17 @@ jobs:
 
     strategy:
       fail-fast: false
+      matrix:
+        libgccjit_version: ["libgccjit.so", "libgccjit_without_int128.so"]
 
     steps:
     - uses: actions/checkout@v2
 
+    - uses: actions/checkout@v2
+      with:
+        repository: llvm/llvm-project
+        path: llvm
+
     - name: Install packages
       run: sudo apt-get install ninja-build ripgrep
 
@@ -21,19 +28,25 @@ jobs:
       uses: dawidd6/action-download-artifact@v2
       with:
           workflow: main.yml
-          name: libgccjit.so
+          name: ${{ matrix.libgccjit_version }}
           path: gcc-build
           repo: antoyo/gcc
+          search_artifacts: true # Because, instead, the action only check the last job ran and that won't work since we want multiple artifacts.
 
     - name: Setup path to libgccjit
       run: |
           echo $(readlink -f gcc-build) > gcc_path
+          # NOTE: the filename is still libgccjit.so even when the artifact name is different.
           ln gcc-build/libgccjit.so gcc-build/libgccjit.so.0
 
-    - name: Set LIBRARY_PATH
+    - name: Set env
       run: |
         echo "LIBRARY_PATH=$(cat gcc_path)" >> $GITHUB_ENV
         echo "LD_LIBRARY_PATH=$(cat gcc_path)" >> $GITHUB_ENV
+        echo "workspace="$GITHUB_WORKSPACE >> $GITHUB_ENV
+
+    - name: Set RUST_COMPILER_RT_ROOT
+      run: echo "RUST_COMPILER_RT_ROOT="${{ env.workspace }}/llvm/compiler-rt >> $GITHUB_ENV
 
     # https://github.com/actions/cache/issues/133
     - name: Fixup owner of ~/.cargo/
diff --git a/.gitignore b/.gitignore
index 1e2f9e3aebb..efda74b2633 100644
--- a/.gitignore
+++ b/.gitignore
@@ -18,3 +18,4 @@ gimple*
 res
 test-backend
 gcc_path
+benchmarks
diff --git a/Cargo.lock b/Cargo.lock
index 2688ea4a4e1..d4c407b0974 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -41,7 +41,7 @@ dependencies = [
 [[package]]
 name = "gccjit"
 version = "1.0.0"
-source = "git+https://github.com/antoyo/gccjit.rs#e68fce53af18dce4d40e6b7090f881ff86a2e892"
+source = "git+https://github.com/antoyo/gccjit.rs#cbb07c6601ba4246fc2967c4d770403c57192ca2"
 dependencies = [
  "gccjit_sys",
 ]
@@ -49,7 +49,7 @@ dependencies = [
 [[package]]
 name = "gccjit_sys"
 version = "0.0.1"
-source = "git+https://github.com/antoyo/gccjit.rs#e68fce53af18dce4d40e6b7090f881ff86a2e892"
+source = "git+https://github.com/antoyo/gccjit.rs#cbb07c6601ba4246fc2967c4d770403c57192ca2"
 dependencies = [
  "libc 0.1.12",
 ]
diff --git a/Readme.md b/Readme.md
index 1fcfb5f6e20..6e333f1b641 100644
--- a/Readme.md
+++ b/Readme.md
@@ -109,6 +109,13 @@ Or add a breakpoint to `add_error` in gdb and print the line number using:
 
 ```
 p loc->m_line
+p loc->m_filename->m_buffer
+```
+
+To print a debug representation of a tree:
+
+```c
+debug_tree(expr);
 ```
 
 To get the `rustc` command to run in `gdb`, add the `--verbose` flag to `cargo build`.
@@ -134,4 +141,5 @@ To get the `rustc` command to run in `gdb`, add the `--verbose` flag to `cargo b
  * Set `linker='-Clinker=m68k-linux-gcc'`.
  * Set the path to the cross-compiling libgccjit in `gcc_path`.
  * Disable the 128-bit integer types if the target doesn't support them by using `let i128_type = context.new_type::<i64>();` in `context.rs` (same for u128_type).
+ * Comment the line: `context.add_command_line_option("-masm=intel");` in src/base.rs.
  * (might not be necessary) Disable the compilation of libstd.so (and possibly libcore.so?).
diff --git a/build.sh b/build.sh
index 17a0d2ab3f0..8a621e12b04 100755
--- a/build.sh
+++ b/build.sh
@@ -3,7 +3,7 @@
 #set -x
 set -e
 
-if [ -f ./gcc_path ]; then 
+if [ -f ./gcc_path ]; then
     export GCC_PATH=$(cat gcc_path)
 else
     echo 'Please put the path to your custom build of libgccjit in the file `gcc_path`, see Readme.md for details'
@@ -13,13 +13,21 @@ fi
 export LD_LIBRARY_PATH="$GCC_PATH"
 export LIBRARY_PATH="$GCC_PATH"
 
+features=
+
+if [[ "$1" == "--features" ]]; then
+    shift
+    features="--features $1"
+    shift
+fi
+
 if [[ "$1" == "--release" ]]; then
     export CHANNEL='release'
-    CARGO_INCREMENTAL=1 cargo rustc --release
+    CARGO_INCREMENTAL=1 cargo rustc --release $features
 else
     echo $LD_LIBRARY_PATH
     export CHANNEL='debug'
-    cargo rustc
+    cargo rustc $features
 fi
 
 source config.sh
diff --git a/build_sysroot/build_sysroot.sh b/build_sysroot/build_sysroot.sh
index d1dcf495db8..a965ca971a0 100755
--- a/build_sysroot/build_sysroot.sh
+++ b/build_sysroot/build_sysroot.sh
@@ -22,7 +22,7 @@ if [[ "$1" == "--release" ]]; then
     RUSTFLAGS="$RUSTFLAGS -Zmir-opt-level=3" cargo build --target $TARGET_TRIPLE --release
 else
     sysroot_channel='debug'
-    cargo build --target $TARGET_TRIPLE
+    cargo build --target $TARGET_TRIPLE --features compiler_builtins/c
 fi
 
 # Copy files to sysroot
diff --git a/src/back/write.rs b/src/back/write.rs
index 334ef32f1d1..b503bd020f6 100644
--- a/src/back/write.rs
+++ b/src/back/write.rs
@@ -45,7 +45,7 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_han
                 if env::var("CG_GCCJIT_DUMP_MODULE_NAMES").as_deref() == Ok("1") {
                     println!("Module {}", module.name);
                 }
-                if env::var("CG_GCCJIT_DUMP_MODULE").as_deref() == Ok(&module.name) {
+                if env::var("CG_GCCJIT_DUMP_ALL_MODULES").as_deref() == Ok("1") || env::var("CG_GCCJIT_DUMP_MODULE").as_deref() == Ok(&module.name) {
                     println!("Dumping reproducer {}", module.name);
                     let _ = fs::create_dir("/tmp/reproducers");
                     // FIXME(antoyo): segfault in dump_reproducer_to_file() might be caused by
@@ -54,6 +54,11 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_han
                     context.dump_reproducer_to_file(&format!("/tmp/reproducers/{}.c", module.name));
                     println!("Dumped reproducer {}", module.name);
                 }
+                if env::var("CG_GCCJIT_DUMP_TO_FILE").as_deref() == Ok("1") {
+                    let _ = fs::create_dir("/tmp/gccjit_dumps");
+                    let path = &format!("/tmp/gccjit_dumps/{}.c", module.name);
+                    context.dump_to_file(path, true);
+                }
                 context.compile_to_file(OutputKind::ObjectFile, obj_out.to_str().expect("path to str"));
             }
 
diff --git a/src/base.rs b/src/base.rs
index 45e791a99d6..6808993182a 100644
--- a/src/base.rs
+++ b/src/base.rs
@@ -52,7 +52,7 @@ pub fn linkage_to_gcc(linkage: Linkage) -> FunctionType {
     }
 }
 
-pub fn compile_codegen_unit<'tcx>(tcx: TyCtxt<'tcx>, cgu_name: Symbol) -> (ModuleCodegen<GccContext>, u64) {
+pub fn compile_codegen_unit<'tcx>(tcx: TyCtxt<'tcx>, cgu_name: Symbol, supports_128bit_integers: bool) -> (ModuleCodegen<GccContext>, u64) {
     let prof_timer = tcx.prof.generic_activity("codegen_module");
     let start_time = Instant::now();
 
@@ -60,7 +60,7 @@ pub fn compile_codegen_unit<'tcx>(tcx: TyCtxt<'tcx>, cgu_name: Symbol) -> (Modul
     let (module, _) = tcx.dep_graph.with_task(
         dep_node,
         tcx,
-        cgu_name,
+        (cgu_name, supports_128bit_integers),
         module_codegen,
         Some(dep_graph::hash_result),
     );
@@ -71,7 +71,7 @@ pub fn compile_codegen_unit<'tcx>(tcx: TyCtxt<'tcx>, cgu_name: Symbol) -> (Modul
     // the time we needed for codegenning it.
     let cost = time_to_codegen.as_secs() * 1_000_000_000 + time_to_codegen.subsec_nanos() as u64;
 
-    fn module_codegen(tcx: TyCtxt<'_>, cgu_name: Symbol) -> ModuleCodegen<GccContext> {
+    fn module_codegen(tcx: TyCtxt<'_>, (cgu_name, supports_128bit_integers): (Symbol, bool)) -> ModuleCodegen<GccContext> {
         let cgu = tcx.codegen_unit(cgu_name);
         // Instantiate monomorphizations without filling out definitions yet...
         //let llvm_module = ModuleLlvm::new(tcx, &cgu_name.as_str());
@@ -106,7 +106,7 @@ pub fn compile_codegen_unit<'tcx>(tcx: TyCtxt<'tcx>, cgu_name: Symbol) -> (Modul
         }
 
         {
-            let cx = CodegenCx::new(&context, cgu, tcx);
+            let cx = CodegenCx::new(&context, cgu, tcx, supports_128bit_integers);
 
             let mono_items = cgu.items_in_deterministic_order(tcx);
             for &(mono_item, (linkage, visibility)) in &mono_items {
diff --git a/src/builder.rs b/src/builder.rs
index ccf8123000c..2969eda2d53 100644
--- a/src/builder.rs
+++ b/src/builder.rs
@@ -94,7 +94,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
     }
 
     fn atomic_extremum(&mut self, operation: ExtremumOperation, dst: RValue<'gcc>, src: RValue<'gcc>, order: AtomicOrdering) -> RValue<'gcc> {
-        let size = self.cx.int_width(src.get_type()) / 8;
+        let size = src.get_type().get_size();
 
         let func = self.current_func();
 
@@ -141,8 +141,8 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
     }
 
     fn compare_exchange(&self, dst: RValue<'gcc>, cmp: LValue<'gcc>, src: RValue<'gcc>, order: AtomicOrdering, failure_order: AtomicOrdering, weak: bool) -> RValue<'gcc> {
-        let size = self.cx.int_width(src.get_type());
-        let compare_exchange = self.context.get_builtin_function(&format!("__atomic_compare_exchange_{}", size / 8));
+        let size = src.get_type().get_size();
+        let compare_exchange = self.context.get_builtin_function(&format!("__atomic_compare_exchange_{}", size));
         let order = self.context.new_rvalue_from_int(self.i32_type, order.to_gcc());
         let failure_order = self.context.new_rvalue_from_int(self.i32_type, failure_order.to_gcc());
         let weak = self.context.new_rvalue_from_int(self.bool_type, weak as i32);
@@ -290,7 +290,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
 
         if return_type != void_type {
             unsafe { RETURN_VALUE_COUNT += 1 };
-            let result = current_func.new_local(None, return_type, &format!("returnValue{}", unsafe { RETURN_VALUE_COUNT }));
+            let result = current_func.new_local(None, return_type, &format!("ptrReturnValue{}", unsafe { RETURN_VALUE_COUNT }));
             current_block.add_assignment(None, result, self.cx.context.new_call_through_ptr(None, func_ptr, &args));
             result.to_rvalue()
         }
@@ -309,7 +309,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         }
     }
 
-    pub fn overflow_call(&mut self, func: Function<'gcc>, args: &[RValue<'gcc>], _funclet: Option<&Funclet>) -> RValue<'gcc> {
+    pub fn overflow_call(&self, func: Function<'gcc>, args: &[RValue<'gcc>], _funclet: Option<&Funclet>) -> RValue<'gcc> {
         // gccjit requires to use the result of functions, even when it's not used.
         // That's why we assign the result to a local.
         let return_type = self.context.new_type::<bool>();
@@ -317,7 +317,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         let current_func = current_block.get_function();
         // TODO(antoyo): return the new_call() directly? Since the overflow function has no side-effects.
         unsafe { RETURN_VALUE_COUNT += 1 };
-        let result = current_func.new_local(None, return_type, &format!("returnValue{}", unsafe { RETURN_VALUE_COUNT }));
+        let result = current_func.new_local(None, return_type, &format!("overflowReturnValue{}", unsafe { RETURN_VALUE_COUNT }));
         current_block.add_assignment(None, result, self.cx.context.new_call(None, func, &args));
         result.to_rvalue()
     }
@@ -468,23 +468,16 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         }
     }
 
-    fn add(&mut self, a: RValue<'gcc>, mut b: RValue<'gcc>) -> RValue<'gcc> {
-        // FIXME(antoyo): this should not be required.
-        if format!("{:?}", a.get_type()) != format!("{:?}", b.get_type()) {
-            b = self.context.new_cast(None, b, a.get_type());
-        }
-        a + b
+    fn add(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        self.gcc_add(a, b)
     }
 
     fn fadd(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
         a + b
     }
 
-    fn sub(&mut self, a: RValue<'gcc>, mut b: RValue<'gcc>) -> RValue<'gcc> {
-        if a.get_type() != b.get_type() {
-            b = self.context.new_cast(None, b, a.get_type());
-        }
-        a - b
+    fn sub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        self.gcc_sub(a, b)
     }
 
     fn fsub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
@@ -492,7 +485,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     fn mul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
-        a * b
+        self.gcc_mul(a, b)
     }
 
     fn fmul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
@@ -500,8 +493,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     fn udiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
-        // TODO(antoyo): convert the arguments to unsigned?
-        a / b
+        self.gcc_udiv(a, b)
     }
 
     fn exactudiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
@@ -511,8 +503,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     fn sdiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
-        // TODO(antoyo): convert the arguments to signed?
-        a / b
+        self.gcc_sdiv(a, b)
     }
 
     fn exactsdiv(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
@@ -529,11 +520,11 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     fn urem(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
-        a % b
+        self.gcc_urem(a, b)
     }
 
     fn srem(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
-        a % b
+        self.gcc_srem(a, b)
     }
 
     fn frem(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
@@ -549,81 +540,33 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     fn shl(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
-        // FIXME(antoyo): remove the casts when libgccjit can shift an unsigned number by an unsigned number.
-        let a_type = a.get_type();
-        let b_type = b.get_type();
-        if a_type.is_unsigned(self) && b_type.is_signed(self) {
-            let a = self.context.new_cast(None, a, b_type);
-            let result = a << b;
-            self.context.new_cast(None, result, a_type)
-        }
-        else if a_type.is_signed(self) && b_type.is_unsigned(self) {
-            let b = self.context.new_cast(None, b, a_type);
-            a << b
-        }
-        else {
-            a << b
-        }
+        self.gcc_shl(a, b)
     }
 
     fn lshr(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
-        // FIXME(antoyo): remove the casts when libgccjit can shift an unsigned number by an unsigned number.
-        // TODO(antoyo): cast to unsigned to do a logical shift if that does not work.
-        let a_type = a.get_type();
-        let b_type = b.get_type();
-        if a_type.is_unsigned(self) && b_type.is_signed(self) {
-            let a = self.context.new_cast(None, a, b_type);
-            let result = a >> b;
-            self.context.new_cast(None, result, a_type)
-        }
-        else if a_type.is_signed(self) && b_type.is_unsigned(self) {
-            let b = self.context.new_cast(None, b, a_type);
-            a >> b
-        }
-        else {
-            a >> b
-        }
+        self.gcc_lshr(a, b)
     }
 
     fn ashr(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
         // TODO(antoyo): check whether behavior is an arithmetic shift for >> .
-        // FIXME(antoyo): remove the casts when libgccjit can shift an unsigned number by an unsigned number.
-        let a_type = a.get_type();
-        let b_type = b.get_type();
-        if a_type.is_unsigned(self) && b_type.is_signed(self) {
-            let a = self.context.new_cast(None, a, b_type);
-            let result = a >> b;
-            self.context.new_cast(None, result, a_type)
-        }
-        else if a_type.is_signed(self) && b_type.is_unsigned(self) {
-            let b = self.context.new_cast(None, b, a_type);
-            a >> b
-        }
-        else {
-            a >> b
-        }
+        // It seems to be if the value is signed.
+        self.gcc_lshr(a, b)
     }
 
-    fn and(&mut self, a: RValue<'gcc>, mut b: RValue<'gcc>) -> RValue<'gcc> {
-        if a.get_type() != b.get_type() {
-            b = self.context.new_cast(None, b, a.get_type());
-        }
-        a & b
+    fn and(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        self.gcc_and(a, b)
     }
 
-    fn or(&mut self, a: RValue<'gcc>, mut b: RValue<'gcc>) -> RValue<'gcc> {
-        if a.get_type() != b.get_type() {
-            b = self.context.new_cast(None, b, a.get_type());
-        }
-        a | b
+    fn or(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        self.cx.gcc_or(a, b)
     }
 
     fn xor(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
-        a ^ b
+        self.gcc_xor(a, b)
     }
 
     fn neg(&mut self, a: RValue<'gcc>) -> RValue<'gcc> {
-        self.cx.context.new_unary_op(None, UnaryOp::Minus, a.get_type(), a)
+        self.gcc_neg(a)
     }
 
     fn fneg(&mut self, a: RValue<'gcc>) -> RValue<'gcc> {
@@ -631,14 +574,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     fn not(&mut self, a: RValue<'gcc>) -> RValue<'gcc> {
-        let operation =
-            if a.get_type().is_bool() {
-                UnaryOp::LogicalNegate
-            }
-            else {
-                UnaryOp::BitwiseNegate
-            };
-        self.cx.context.new_unary_op(None, operation, a.get_type(), a)
+        self.gcc_not(a)
     }
 
     fn unchecked_sadd(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
@@ -646,7 +582,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     fn unchecked_uadd(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
-        a + b
+        self.gcc_add(a, b)
     }
 
     fn unchecked_ssub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
@@ -655,7 +591,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
 
     fn unchecked_usub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
         // TODO(antoyo): should generate poison value?
-        a - b
+        self.gcc_sub(a, b)
     }
 
     fn unchecked_smul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
@@ -687,76 +623,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     fn checked_binop(&mut self, oop: OverflowOp, typ: Ty<'_>, lhs: Self::Value, rhs: Self::Value) -> (Self::Value, Self::Value) {
-        use rustc_middle::ty::{Int, IntTy::*, Uint, UintTy::*};
-
-        let new_kind =
-            match typ.kind() {
-                Int(t @ Isize) => Int(t.normalize(self.tcx.sess.target.pointer_width)),
-                Uint(t @ Usize) => Uint(t.normalize(self.tcx.sess.target.pointer_width)),
-                t @ (Uint(_) | Int(_)) => t.clone(),
-                _ => panic!("tried to get overflow intrinsic for op applied to non-int type"),
-            };
-
-        // TODO(antoyo): remove duplication with intrinsic?
-        let name =
-            match oop {
-                OverflowOp::Add =>
-                    match new_kind {
-                        Int(I8) => "__builtin_add_overflow",
-                        Int(I16) => "__builtin_add_overflow",
-                        Int(I32) => "__builtin_sadd_overflow",
-                        Int(I64) => "__builtin_saddll_overflow",
-                        Int(I128) => "__builtin_add_overflow",
-
-                        Uint(U8) => "__builtin_add_overflow",
-                        Uint(U16) => "__builtin_add_overflow",
-                        Uint(U32) => "__builtin_uadd_overflow",
-                        Uint(U64) => "__builtin_uaddll_overflow",
-                        Uint(U128) => "__builtin_add_overflow",
-
-                        _ => unreachable!(),
-                    },
-                OverflowOp::Sub =>
-                    match new_kind {
-                        Int(I8) => "__builtin_sub_overflow",
-                        Int(I16) => "__builtin_sub_overflow",
-                        Int(I32) => "__builtin_ssub_overflow",
-                        Int(I64) => "__builtin_ssubll_overflow",
-                        Int(I128) => "__builtin_sub_overflow",
-
-                        Uint(U8) => "__builtin_sub_overflow",
-                        Uint(U16) => "__builtin_sub_overflow",
-                        Uint(U32) => "__builtin_usub_overflow",
-                        Uint(U64) => "__builtin_usubll_overflow",
-                        Uint(U128) => "__builtin_sub_overflow",
-
-                        _ => unreachable!(),
-                    },
-                OverflowOp::Mul =>
-                    match new_kind {
-                        Int(I8) => "__builtin_mul_overflow",
-                        Int(I16) => "__builtin_mul_overflow",
-                        Int(I32) => "__builtin_smul_overflow",
-                        Int(I64) => "__builtin_smulll_overflow",
-                        Int(I128) => "__builtin_mul_overflow",
-
-                        Uint(U8) => "__builtin_mul_overflow",
-                        Uint(U16) => "__builtin_mul_overflow",
-                        Uint(U32) => "__builtin_umul_overflow",
-                        Uint(U64) => "__builtin_umulll_overflow",
-                        Uint(U128) => "__builtin_mul_overflow",
-
-                        _ => unreachable!(),
-                    },
-            };
-
-        let intrinsic = self.context.get_builtin_function(&name);
-        let res = self.current_func()
-            // TODO(antoyo): is it correct to use rhs type instead of the parameter typ?
-            .new_local(None, rhs.get_type(), "binopResult")
-            .get_address(None);
-        let overflow = self.overflow_call(intrinsic, &[lhs, rhs, res], None);
-        (res.dereference(None).to_rvalue(), overflow)
+        self.gcc_checked_binop(oop, typ, lhs, rhs)
     }
 
     fn alloca(&mut self, ty: Type<'gcc>, align: Align) -> RValue<'gcc> {
@@ -1003,7 +870,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     /* Casts */
     fn trunc(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
         // TODO(antoyo): check that it indeed truncate the value.
-        self.context.new_cast(None, value, dest_ty)
+        self.gcc_int_cast(value, dest_ty)
     }
 
     fn sext(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
@@ -1016,19 +883,19 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     fn fptoui(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
-        self.context.new_cast(None, value, dest_ty)
+        self.gcc_float_to_uint_cast(value, dest_ty)
     }
 
     fn fptosi(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
-        self.context.new_cast(None, value, dest_ty)
+        self.gcc_float_to_int_cast(value, dest_ty)
     }
 
     fn uitofp(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
-        self.context.new_cast(None, value, dest_ty)
+        self.gcc_uint_to_float_cast(value, dest_ty)
     }
 
     fn sitofp(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
-        self.context.new_cast(None, value, dest_ty)
+        self.gcc_int_to_float_cast(value, dest_ty)
     }
 
     fn fptrunc(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
@@ -1054,7 +921,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
 
     fn intcast(&mut self, value: RValue<'gcc>, dest_typ: Type<'gcc>, _is_signed: bool) -> RValue<'gcc> {
         // NOTE: is_signed is for value, not dest_typ.
-        self.cx.context.new_cast(None, value, dest_typ)
+        self.gcc_int_cast(value, dest_typ)
     }
 
     fn pointercast(&mut self, value: RValue<'gcc>, dest_ty: Type<'gcc>) -> RValue<'gcc> {
@@ -1075,21 +942,8 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     /* Comparisons */
-    fn icmp(&mut self, op: IntPredicate, mut lhs: RValue<'gcc>, mut rhs: RValue<'gcc>) -> RValue<'gcc> {
-        let left_type = lhs.get_type();
-        let right_type = rhs.get_type();
-        if left_type != right_type {
-            // NOTE: because libgccjit cannot compare function pointers.
-            if left_type.dyncast_function_ptr_type().is_some() && right_type.dyncast_function_ptr_type().is_some() {
-                lhs = self.context.new_cast(None, lhs, self.usize_type.make_pointer());
-                rhs = self.context.new_cast(None, rhs, self.usize_type.make_pointer());
-            }
-            // NOTE: hack because we try to cast a vector type to the same vector type.
-            else if format!("{:?}", left_type) != format!("{:?}", right_type) {
-                rhs = self.context.new_cast(None, rhs, left_type);
-            }
-        }
-        self.context.new_comparison(None, op.to_gcc_comparison(), lhs, rhs)
+    fn icmp(&mut self, op: IntPredicate, lhs: RValue<'gcc>, rhs: RValue<'gcc>) -> RValue<'gcc> {
+        self.gcc_icmp(op, lhs, rhs)
     }
 
     fn fcmp(&mut self, op: RealPredicate, lhs: RValue<'gcc>, rhs: RValue<'gcc>) -> RValue<'gcc> {
@@ -1156,7 +1010,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
         then_block.add_assignment(None, variable, then_val);
         then_block.end_with_jump(None, after_block);
 
-        if then_val.get_type() != else_val.get_type() {
+        if !then_val.get_type().is_compatible_with(else_val.get_type()) {
             else_val = self.context.new_cast(None, else_val, then_val.get_type());
         }
         else_block.add_assignment(None, variable, else_val);
@@ -1322,7 +1176,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
     }
 
     fn atomic_rmw(&mut self, op: AtomicRmwBinOp, dst: RValue<'gcc>, src: RValue<'gcc>, order: AtomicOrdering) -> RValue<'gcc> {
-        let size = self.cx.int_width(src.get_type()) / 8;
+        let size = src.get_type().get_size();
         let name =
             match op {
                 AtomicRmwBinOp::AtomicXchg => format!("__atomic_exchange_{}", size),
@@ -1396,7 +1250,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> {
             // Fix the code in codegen_ssa::base::from_immediate.
             return value;
         }
-        self.context.new_cast(None, value, dest_typ)
+        self.gcc_int_cast(value, dest_typ)
     }
 
     fn cx(&self) -> &CodegenCx<'gcc, 'tcx> {
@@ -1470,7 +1324,7 @@ impl<'tcx> HasTargetSpec for Builder<'_, '_, 'tcx> {
     }
 }
 
-trait ToGccComp {
+pub trait ToGccComp {
     fn to_gcc_comparison(&self) -> ComparisonOp;
 }
 
diff --git a/src/common.rs b/src/common.rs
index 5851826147d..89a3dc052d8 100644
--- a/src/common.rs
+++ b/src/common.rs
@@ -1,7 +1,5 @@
-use std::convert::TryFrom;
-
 use gccjit::LValue;
-use gccjit::{Block, CType, RValue, Type, ToRValue};
+use gccjit::{Block, RValue, Type, ToRValue};
 use rustc_codegen_ssa::mir::place::PlaceRef;
 use rustc_codegen_ssa::traits::{
     BaseTypeMethods,
@@ -111,29 +109,15 @@ impl<'gcc, 'tcx> ConstMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
     }
 
     fn const_int(&self, typ: Type<'gcc>, int: i64) -> RValue<'gcc> {
-        self.context.new_rvalue_from_long(typ, i64::try_from(int).expect("i64::try_from"))
+        self.gcc_int(typ, int)
     }
 
     fn const_uint(&self, typ: Type<'gcc>, int: u64) -> RValue<'gcc> {
-        self.context.new_rvalue_from_long(typ, u64::try_from(int).expect("u64::try_from") as i64)
+        self.gcc_uint(typ, int)
     }
 
     fn const_uint_big(&self, typ: Type<'gcc>, num: u128) -> RValue<'gcc> {
-        if num >> 64 != 0 {
-            // FIXME(antoyo): use a new function new_rvalue_from_unsigned_long()?
-            let low = self.context.new_rvalue_from_long(self.u64_type, num as u64 as i64);
-            let high = self.context.new_rvalue_from_long(typ, (num >> 64) as u64 as i64);
-
-            let sixty_four = self.context.new_rvalue_from_long(typ, 64);
-            (high << sixty_four) | self.context.new_cast(None, low, typ)
-        }
-        else if typ.is_i128(self) {
-            let num = self.context.new_rvalue_from_long(self.u64_type, num as u64 as i64);
-            self.context.new_cast(None, num, typ)
-        }
-        else {
-            self.context.new_rvalue_from_long(typ, num as u64 as i64)
-        }
+        self.gcc_uint_big(typ, num)
     }
 
     fn const_bool(&self, val: bool) -> RValue<'gcc> {
@@ -425,11 +409,11 @@ impl<'gcc, 'tcx> TypeReflection<'gcc, 'tcx> for Type<'gcc> {
     }
 
     fn is_i128(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
-        self.unqualified() == cx.context.new_c_type(CType::Int128t)
+        self.unqualified() == cx.i128_type.unqualified()
     }
 
     fn is_u128(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
-        self.unqualified() == cx.context.new_c_type(CType::UInt128t)
+        self.unqualified() == cx.u128_type.unqualified()
     }
 
     fn is_f32(&self, cx: &CodegenCx<'gcc, 'tcx>) -> bool {
diff --git a/src/context.rs b/src/context.rs
index d260a983347..795966d8183 100644
--- a/src/context.rs
+++ b/src/context.rs
@@ -62,6 +62,8 @@ pub struct CodegenCx<'gcc, 'tcx> {
     pub ulonglong_type: Type<'gcc>,
     pub sizet_type: Type<'gcc>,
 
+    pub supports_128bit_integers: bool,
+
     pub float_type: Type<'gcc>,
     pub double_type: Type<'gcc>,
 
@@ -110,22 +112,29 @@ pub struct CodegenCx<'gcc, 'tcx> {
 }
 
 impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
-    pub fn new(context: &'gcc Context<'gcc>, codegen_unit: &'tcx CodegenUnit<'tcx>, tcx: TyCtxt<'tcx>) -> Self {
+    pub fn new(context: &'gcc Context<'gcc>, codegen_unit: &'tcx CodegenUnit<'tcx>, tcx: TyCtxt<'tcx>, supports_128bit_integers: bool) -> Self {
         let check_overflow = tcx.sess.overflow_checks();
-        // TODO(antoyo): fix this mess. libgccjit seems to return random type when using new_int_type().
-        let isize_type = context.new_c_type(CType::LongLong);
-        let usize_type = context.new_c_type(CType::ULongLong);
-        let bool_type = context.new_type::<bool>();
-        let i8_type = context.new_type::<i8>();
-        let i16_type = context.new_type::<i16>();
-        let i32_type = context.new_type::<i32>();
-        let i64_type = context.new_c_type(CType::LongLong);
-        let i128_type = context.new_c_type(CType::Int128t).get_aligned(8); // TODO(antoyo): should the alignment be hard-coded?
-        let u8_type = context.new_type::<u8>();
-        let u16_type = context.new_type::<u16>();
-        let u32_type = context.new_type::<u32>();
-        let u64_type = context.new_c_type(CType::ULongLong);
-        let u128_type = context.new_c_type(CType::UInt128t).get_aligned(8); // TODO(antoyo): should the alignment be hard-coded?
+
+        let i8_type = context.new_c_type(CType::Int8t);
+        let i16_type = context.new_c_type(CType::Int16t);
+        let i32_type = context.new_c_type(CType::Int32t);
+        let i64_type = context.new_c_type(CType::Int64t);
+        let u8_type = context.new_c_type(CType::UInt8t);
+        let u16_type = context.new_c_type(CType::UInt16t);
+        let u32_type = context.new_c_type(CType::UInt32t);
+        let u64_type = context.new_c_type(CType::UInt64t);
+
+        let (i128_type, u128_type) =
+            if supports_128bit_integers {
+                let i128_type = context.new_c_type(CType::Int128t).get_aligned(8); // TODO(antoyo): should the alignment be hard-coded?;
+                let u128_type = context.new_c_type(CType::UInt128t).get_aligned(8); // TODO(antoyo): should the alignment be hard-coded?;
+                (i128_type, u128_type)
+            }
+            else {
+                let i128_type = context.new_array_type(None, i64_type, 2);
+                let u128_type = context.new_array_type(None, u64_type, 2);
+                (i128_type, u128_type)
+            };
 
         let tls_model = to_gcc_tls_mode(tcx.sess.tls_model());
 
@@ -139,8 +148,13 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
         let ulonglong_type = context.new_c_type(CType::ULongLong);
         let sizet_type = context.new_c_type(CType::SizeT);
 
-        assert_eq!(isize_type, i64_type);
-        assert_eq!(usize_type, u64_type);
+        let isize_type = context.new_c_type(CType::LongLong);
+        let usize_type = context.new_c_type(CType::ULongLong);
+        let bool_type = context.new_type::<bool>();
+
+        // TODO(antoyo): only have those assertions on x86_64.
+        assert_eq!(isize_type.get_size(), i64_type.get_size());
+        assert_eq!(usize_type.get_size(), u64_type.get_size());
 
         let mut functions = FxHashMap::default();
         let builtins = [
@@ -190,6 +204,8 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
             ulonglong_type,
             sizet_type,
 
+            supports_128bit_integers,
+
             float_type,
             double_type,
 
@@ -221,6 +237,41 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
         function
     }
 
+    pub fn is_native_int_type(&self, typ: Type<'gcc>) -> bool {
+        let types = [
+            self.u8_type,
+            self.u16_type,
+            self.u32_type,
+            self.u64_type,
+            self.i8_type,
+            self.i16_type,
+            self.i32_type,
+            self.i64_type,
+        ];
+
+        for native_type in types {
+            if native_type.is_compatible_with(typ) {
+                return true;
+            }
+        }
+
+        self.supports_128bit_integers &&
+            (self.u128_type.is_compatible_with(typ) || self.i128_type.is_compatible_with(typ))
+    }
+
+    pub fn is_non_native_int_type(&self, typ: Type<'gcc>) -> bool {
+        !self.supports_128bit_integers &&
+            (self.u128_type.is_compatible_with(typ) || self.i128_type.is_compatible_with(typ))
+    }
+
+    pub fn is_native_int_type_or_bool(&self, typ: Type<'gcc>) -> bool {
+        self.is_native_int_type(typ) || typ == self.bool_type
+    }
+
+    pub fn is_int_type_or_bool(&self, typ: Type<'gcc>) -> bool {
+        self.is_native_int_type(typ) || self.is_non_native_int_type(typ) || typ == self.bool_type
+    }
+
     pub fn sess(&self) -> &Session {
         &self.tcx.sess
     }
diff --git a/src/int.rs b/src/int.rs
new file mode 100644
index 00000000000..a1f28f3f881
--- /dev/null
+++ b/src/int.rs
@@ -0,0 +1,737 @@
+//! Module to handle integer operations.
+//! This module exists because some integer types are not supported on some gcc platforms, e.g.
+//! 128-bit integers on 32-bit platforms and thus require to be handled manually.
+
+use std::convert::TryFrom;
+
+use gccjit::{ComparisonOp, FunctionType, RValue, ToRValue, Type, UnaryOp, BinaryOp};
+use rustc_codegen_ssa::common::{IntPredicate, TypeKind};
+use rustc_codegen_ssa::traits::{BackendTypes, BaseTypeMethods, BuilderMethods, OverflowOp};
+use rustc_middle::ty::Ty;
+
+use crate::builder::ToGccComp;
+use crate::{builder::Builder, common::{SignType, TypeReflection}, context::CodegenCx};
+
+impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
+    pub fn gcc_urem(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        // 128-bit unsigned %: __umodti3
+        self.multiplicative_operation(BinaryOp::Modulo, "mod", false, a, b)
+    }
+
+    pub fn gcc_srem(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        // 128-bit signed %:   __modti3
+        self.multiplicative_operation(BinaryOp::Modulo, "mod", true, a, b)
+    }
+
+    pub fn gcc_not(&self, a: RValue<'gcc>) -> RValue<'gcc> {
+        let typ = a.get_type();
+        if self.is_native_int_type_or_bool(typ) {
+            let operation =
+                if typ.is_bool() {
+                    UnaryOp::LogicalNegate
+                }
+                else {
+                    UnaryOp::BitwiseNegate
+                };
+            self.cx.context.new_unary_op(None, operation, typ, a)
+        }
+        else {
+            // TODO(antoyo): use __negdi2 and __negti2 instead?
+            let element_type = typ.dyncast_array().expect("element type");
+            let values = [
+                self.cx.context.new_unary_op(None, UnaryOp::BitwiseNegate, element_type, self.low(a)),
+                self.cx.context.new_unary_op(None, UnaryOp::BitwiseNegate, element_type, self.high(a)),
+            ];
+            self.cx.context.new_array_constructor(None, typ, &values)
+        }
+    }
+
+    pub fn gcc_neg(&self, a: RValue<'gcc>) -> RValue<'gcc> {
+        let a_type = a.get_type();
+        if self.is_native_int_type(a_type) {
+            self.cx.context.new_unary_op(None, UnaryOp::Minus, a.get_type(), a)
+        }
+        else {
+            let param_a = self.context.new_parameter(None, a_type, "a");
+            let func = self.context.new_function(None, FunctionType::Extern, a_type, &[param_a], "__negti2", false);
+            self.context.new_call(None, func, &[a])
+        }
+    }
+
+    pub fn gcc_and(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        self.cx.bitwise_operation(BinaryOp::BitwiseAnd, a, b)
+    }
+
+    pub fn gcc_lshr(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        let a_type = a.get_type();
+        let b_type = b.get_type();
+        let a_native = self.is_native_int_type(a_type);
+        let b_native = self.is_native_int_type(b_type);
+        if a_native && b_native {
+            // FIXME(antoyo): remove the casts when libgccjit can shift an unsigned number by an unsigned number.
+            // TODO(antoyo): cast to unsigned to do a logical shift if that does not work.
+            if a_type.is_unsigned(self) && b_type.is_signed(self) {
+                let a = self.context.new_cast(None, a, b_type);
+                let result = a >> b;
+                self.context.new_cast(None, result, a_type)
+            }
+            else if a_type.is_signed(self) && b_type.is_unsigned(self) {
+                let b = self.context.new_cast(None, b, a_type);
+                a >> b
+            }
+            else {
+                a >> b
+            }
+        }
+        else if a_native && !b_native {
+            self.gcc_lshr(a, self.gcc_int_cast(b, a_type))
+        }
+        else {
+            // NOTE: we cannot use the lshr builtin because it's calling hi() (to get the most
+            // significant half of the number) which uses lshr.
+
+            let native_int_type = a_type.dyncast_array().expect("get element type");
+
+            let func = self.current_func();
+            let then_block = func.new_block("then");
+            let else_block = func.new_block("else");
+            let after_block = func.new_block("after");
+            let b0_block = func.new_block("b0");
+            let actual_else_block = func.new_block("actual_else");
+
+            let result = func.new_local(None, a_type, "shiftResult");
+
+            let sixty_four = self.gcc_int(native_int_type, 64);
+            let sixty_three = self.gcc_int(native_int_type, 63);
+            let zero = self.gcc_zero(native_int_type);
+            let b = self.gcc_int_cast(b, native_int_type);
+            let condition = self.gcc_icmp(IntPredicate::IntNE, self.gcc_and(b, sixty_four), zero);
+            self.llbb().end_with_conditional(None, condition, then_block, else_block);
+
+            // TODO(antoyo): take endianness into account.
+            let shift_value = self.gcc_sub(b, sixty_four);
+            let high = self.high(a);
+            let sign =
+                if a_type.is_signed(self) {
+                    high >> sixty_three
+                }
+                else {
+                    zero
+                };
+            let values = [
+                high >> shift_value,
+                sign,
+            ];
+            let array_value = self.context.new_array_constructor(None, a_type, &values);
+            then_block.add_assignment(None, result, array_value);
+            then_block.end_with_jump(None, after_block);
+
+            let condition = self.gcc_icmp(IntPredicate::IntEQ, b, zero);
+            else_block.end_with_conditional(None, condition, b0_block, actual_else_block);
+
+            b0_block.add_assignment(None, result, a);
+            b0_block.end_with_jump(None, after_block);
+
+            let shift_value = self.gcc_sub(sixty_four, b);
+            // NOTE: cast low to its unsigned type in order to perform a logical right shift.
+            let unsigned_type = native_int_type.to_unsigned(&self.cx);
+            let casted_low = self.context.new_cast(None, self.low(a), unsigned_type);
+            let shifted_low = casted_low >> self.context.new_cast(None, b, unsigned_type);
+            let shifted_low = self.context.new_cast(None, shifted_low, native_int_type);
+            let values = [
+                (high << shift_value) | shifted_low,
+                high >> b,
+            ];
+            let array_value = self.context.new_array_constructor(None, a_type, &values);
+            actual_else_block.add_assignment(None, result, array_value);
+            actual_else_block.end_with_jump(None, after_block);
+
+            // NOTE: since jumps were added in a place rustc does not expect, the current block in the
+            // state need to be updated.
+            self.block = Some(after_block);
+            *self.cx.current_block.borrow_mut() = Some(after_block);
+
+            result.to_rvalue()
+        }
+    }
+
+    fn additive_operation(&self, operation: BinaryOp, a: RValue<'gcc>, mut b: RValue<'gcc>) -> RValue<'gcc> {
+        let a_type = a.get_type();
+        let b_type = b.get_type();
+        if self.is_native_int_type_or_bool(a_type) && self.is_native_int_type_or_bool(b_type) {
+            if a.get_type() != b.get_type() {
+                b = self.context.new_cast(None, b, a.get_type());
+            }
+            self.context.new_binary_op(None, operation, a_type, a, b)
+        }
+        else {
+            let signed = a_type.is_compatible_with(self.i128_type);
+            let func_name =
+                match (operation, signed) {
+                    (BinaryOp::Plus, true) => "__rust_i128_add",
+                    (BinaryOp::Plus, false) => "__rust_u128_add",
+                    (BinaryOp::Minus, true) => "__rust_i128_sub",
+                    (BinaryOp::Minus, false) => "__rust_u128_sub",
+                    _ => unreachable!("unexpected additive operation {:?}", operation),
+                };
+            let param_a = self.context.new_parameter(None, a_type, "a");
+            let param_b = self.context.new_parameter(None, b_type, "b");
+            let func = self.context.new_function(None, FunctionType::Extern, a_type, &[param_a, param_b], func_name, false);
+            self.context.new_call(None, func, &[a, b])
+        }
+    }
+
+    pub fn gcc_add(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        self.additive_operation(BinaryOp::Plus, a, b)
+    }
+
+    pub fn gcc_mul(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        self.multiplicative_operation(BinaryOp::Mult, "mul", true, a, b)
+    }
+
+    pub fn gcc_sub(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        self.additive_operation(BinaryOp::Minus, a, b)
+    }
+
+    fn multiplicative_operation(&self, operation: BinaryOp, operation_name: &str, signed: bool, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        let a_type = a.get_type();
+        let b_type = b.get_type();
+        if self.is_native_int_type_or_bool(a_type) && self.is_native_int_type_or_bool(b_type) {
+            self.context.new_binary_op(None, operation, a_type, a, b)
+        }
+        else {
+            let sign =
+                if signed {
+                    ""
+                }
+                else {
+                    "u"
+                };
+            let func_name = format!("__{}{}ti3", sign, operation_name);
+            let param_a = self.context.new_parameter(None, a_type, "a");
+            let param_b = self.context.new_parameter(None, b_type, "b");
+            let func = self.context.new_function(None, FunctionType::Extern, a_type, &[param_a, param_b], func_name, false);
+            self.context.new_call(None, func, &[a, b])
+        }
+    }
+
+    pub fn gcc_sdiv(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        // TODO(antoyo): check if the types are signed?
+        // 128-bit, signed: __divti3
+        // TODO(antoyo): convert the arguments to signed?
+        self.multiplicative_operation(BinaryOp::Divide, "div", true, a, b)
+    }
+
+    pub fn gcc_udiv(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        // 128-bit, unsigned: __udivti3
+        self.multiplicative_operation(BinaryOp::Divide, "div", false, a, b)
+    }
+
+    pub fn gcc_checked_binop(&self, oop: OverflowOp, typ: Ty<'_>, lhs: <Self as BackendTypes>::Value, rhs: <Self as BackendTypes>::Value) -> (<Self as BackendTypes>::Value, <Self as BackendTypes>::Value) {
+        use rustc_middle::ty::{Int, IntTy::*, Uint, UintTy::*};
+
+        let new_kind =
+            match typ.kind() {
+                Int(t @ Isize) => Int(t.normalize(self.tcx.sess.target.pointer_width)),
+                Uint(t @ Usize) => Uint(t.normalize(self.tcx.sess.target.pointer_width)),
+                t @ (Uint(_) | Int(_)) => t.clone(),
+                _ => panic!("tried to get overflow intrinsic for op applied to non-int type"),
+            };
+
+        // TODO(antoyo): remove duplication with intrinsic?
+        let name =
+            if self.is_native_int_type(lhs.get_type()) {
+                match oop {
+                    OverflowOp::Add =>
+                        match new_kind {
+                            Int(I8) => "__builtin_add_overflow",
+                            Int(I16) => "__builtin_add_overflow",
+                            Int(I32) => "__builtin_sadd_overflow",
+                            Int(I64) => "__builtin_saddll_overflow",
+                            Int(I128) => "__builtin_add_overflow",
+
+                            Uint(U8) => "__builtin_add_overflow",
+                            Uint(U16) => "__builtin_add_overflow",
+                            Uint(U32) => "__builtin_uadd_overflow",
+                            Uint(U64) => "__builtin_uaddll_overflow",
+                            Uint(U128) => "__builtin_add_overflow",
+
+                            _ => unreachable!(),
+                        },
+                    OverflowOp::Sub =>
+                        match new_kind {
+                            Int(I8) => "__builtin_sub_overflow",
+                            Int(I16) => "__builtin_sub_overflow",
+                            Int(I32) => "__builtin_ssub_overflow",
+                            Int(I64) => "__builtin_ssubll_overflow",
+                            Int(I128) => "__builtin_sub_overflow",
+
+                            Uint(U8) => "__builtin_sub_overflow",
+                            Uint(U16) => "__builtin_sub_overflow",
+                            Uint(U32) => "__builtin_usub_overflow",
+                            Uint(U64) => "__builtin_usubll_overflow",
+                            Uint(U128) => "__builtin_sub_overflow",
+
+                            _ => unreachable!(),
+                        },
+                    OverflowOp::Mul =>
+                        match new_kind {
+                            Int(I8) => "__builtin_mul_overflow",
+                            Int(I16) => "__builtin_mul_overflow",
+                            Int(I32) => "__builtin_smul_overflow",
+                            Int(I64) => "__builtin_smulll_overflow",
+                            Int(I128) => "__builtin_mul_overflow",
+
+                            Uint(U8) => "__builtin_mul_overflow",
+                            Uint(U16) => "__builtin_mul_overflow",
+                            Uint(U32) => "__builtin_umul_overflow",
+                            Uint(U64) => "__builtin_umulll_overflow",
+                            Uint(U128) => "__builtin_mul_overflow",
+
+                            _ => unreachable!(),
+                        },
+                }
+            }
+            else {
+                match new_kind {
+                    Int(I128) | Uint(U128) => {
+                        let func_name =
+                            match oop {
+                                OverflowOp::Add =>
+                                    match new_kind {
+                                        Int(I128) => "__rust_i128_addo",
+                                        Uint(U128) => "__rust_u128_addo",
+                                        _ => unreachable!(),
+                                    },
+                                OverflowOp::Sub =>
+                                    match new_kind {
+                                        Int(I128) => "__rust_i128_subo",
+                                        Uint(U128) => "__rust_u128_subo",
+                                        _ => unreachable!(),
+                                    },
+                                OverflowOp::Mul =>
+                                    match new_kind {
+                                        Int(I128) => "__rust_i128_mulo", // TODO(antoyo): use __muloti4d instead?
+                                        Uint(U128) => "__rust_u128_mulo",
+                                        _ => unreachable!(),
+                                    },
+                            };
+                        let a_type = lhs.get_type();
+                        let b_type = rhs.get_type();
+                        let param_a = self.context.new_parameter(None, a_type, "a");
+                        let param_b = self.context.new_parameter(None, b_type, "b");
+                        let result_field = self.context.new_field(None, a_type, "result");
+                        let overflow_field = self.context.new_field(None, self.bool_type, "overflow");
+                        let return_type = self.context.new_struct_type(None, "result_overflow", &[result_field, overflow_field]);
+                        let func = self.context.new_function(None, FunctionType::Extern, return_type.as_type(), &[param_a, param_b], func_name, false);
+                        let result = self.context.new_call(None, func, &[lhs, rhs]);
+                        let overflow = result.access_field(None, overflow_field);
+                        let int_result = result.access_field(None, result_field);
+                        return (int_result, overflow);
+                    },
+                    _ => {
+                        match oop {
+                            OverflowOp::Mul =>
+                                match new_kind {
+                                    Int(I32) => "__mulosi4",
+                                    Int(I64) => "__mulodi4",
+                                    _ => unreachable!(),
+                                },
+                            _ => unimplemented!("overflow operation for {:?}", new_kind),
+                        }
+                    }
+                }
+            };
+
+        let intrinsic = self.context.get_builtin_function(&name);
+        let res = self.current_func()
+            // TODO(antoyo): is it correct to use rhs type instead of the parameter typ?
+            .new_local(None, rhs.get_type(), "binopResult")
+            .get_address(None);
+        let overflow = self.overflow_call(intrinsic, &[lhs, rhs, res], None);
+        (res.dereference(None).to_rvalue(), overflow)
+    }
+
+    pub fn gcc_icmp(&self, op: IntPredicate, mut lhs: RValue<'gcc>, mut rhs: RValue<'gcc>) -> RValue<'gcc> {
+        let a_type = lhs.get_type();
+        let b_type = rhs.get_type();
+        if self.is_non_native_int_type(a_type) || self.is_non_native_int_type(b_type) {
+            let signed = a_type.is_compatible_with(self.i128_type);
+            let sign =
+                if signed {
+                    ""
+                }
+                else {
+                    "u"
+                };
+            let func_name = format!("__{}cmpti2", sign);
+            let param_a = self.context.new_parameter(None, a_type, "a");
+            let param_b = self.context.new_parameter(None, b_type, "b");
+            let func = self.context.new_function(None, FunctionType::Extern, self.int_type, &[param_a, param_b], func_name, false);
+            let cmp = self.context.new_call(None, func, &[lhs, rhs]);
+            let (op, limit) =
+                match op {
+                    IntPredicate::IntEQ => {
+                        return self.context.new_comparison(None, ComparisonOp::Equals, cmp, self.context.new_rvalue_one(self.int_type));
+                    },
+                    IntPredicate::IntNE => {
+                        return self.context.new_comparison(None, ComparisonOp::NotEquals, cmp, self.context.new_rvalue_one(self.int_type));
+                    },
+                    IntPredicate::IntUGT => (ComparisonOp::Equals, 2),
+                    IntPredicate::IntUGE => (ComparisonOp::GreaterThanEquals, 1),
+                    IntPredicate::IntULT => (ComparisonOp::Equals, 0),
+                    IntPredicate::IntULE => (ComparisonOp::LessThanEquals, 1),
+                    IntPredicate::IntSGT => (ComparisonOp::Equals, 2),
+                    IntPredicate::IntSGE => (ComparisonOp::GreaterThanEquals, 1),
+                    IntPredicate::IntSLT => (ComparisonOp::Equals, 0),
+                    IntPredicate::IntSLE => (ComparisonOp::LessThanEquals, 1),
+                };
+            self.context.new_comparison(None, op, cmp, self.context.new_rvalue_from_int(self.int_type, limit))
+        }
+        else {
+            let left_type = lhs.get_type();
+            let right_type = rhs.get_type();
+            if left_type != right_type {
+                // NOTE: because libgccjit cannot compare function pointers.
+                if left_type.dyncast_function_ptr_type().is_some() && right_type.dyncast_function_ptr_type().is_some() {
+                    lhs = self.context.new_cast(None, lhs, self.usize_type.make_pointer());
+                    rhs = self.context.new_cast(None, rhs, self.usize_type.make_pointer());
+                }
+                // NOTE: hack because we try to cast a vector type to the same vector type.
+                else if format!("{:?}", left_type) != format!("{:?}", right_type) {
+                    rhs = self.context.new_cast(None, rhs, left_type);
+                }
+            }
+            self.context.new_comparison(None, op.to_gcc_comparison(), lhs, rhs)
+        }
+    }
+
+    pub fn gcc_xor(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        let a_type = a.get_type();
+        let b_type = b.get_type();
+        if self.is_native_int_type_or_bool(a_type) && self.is_native_int_type_or_bool(b_type) {
+            a ^ b
+        }
+        else {
+            let values = [
+                self.low(a) ^ self.low(b),
+                self.high(a) ^ self.high(b),
+            ];
+            self.context.new_array_constructor(None, a_type, &values)
+        }
+    }
+
+    pub fn gcc_shl(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        let a_type = a.get_type();
+        let b_type = b.get_type();
+        let a_native = self.is_native_int_type(a_type);
+        let b_native = self.is_native_int_type(b_type);
+        if a_native && b_native {
+            // FIXME(antoyo): remove the casts when libgccjit can shift an unsigned number by an unsigned number.
+            if a_type.is_unsigned(self) && b_type.is_signed(self) {
+                let a = self.context.new_cast(None, a, b_type);
+                let result = a << b;
+                self.context.new_cast(None, result, a_type)
+            }
+            else if a_type.is_signed(self) && b_type.is_unsigned(self) {
+                let b = self.context.new_cast(None, b, a_type);
+                a << b
+            }
+            else {
+                a << b
+            }
+        }
+        else if a_native && !b_native {
+            self.gcc_shl(a, self.gcc_int_cast(b, a_type))
+        }
+        else {
+            // NOTE: we cannot use the ashl builtin because it's calling widen_hi() which uses ashl.
+            let native_int_type = a_type.dyncast_array().expect("get element type");
+
+            let func = self.current_func();
+            let then_block = func.new_block("then");
+            let else_block = func.new_block("else");
+            let after_block = func.new_block("after");
+            let b0_block = func.new_block("b0");
+            let actual_else_block = func.new_block("actual_else");
+
+            let result = func.new_local(None, a_type, "shiftResult");
+
+            let b = self.gcc_int_cast(b, native_int_type);
+            let sixty_four = self.gcc_int(native_int_type, 64);
+            let zero = self.gcc_zero(native_int_type);
+            let condition = self.gcc_icmp(IntPredicate::IntNE, self.gcc_and(b, sixty_four), zero);
+            self.llbb().end_with_conditional(None, condition, then_block, else_block);
+
+            // TODO(antoyo): take endianness into account.
+            let values = [
+                zero,
+                self.low(a) << (b - sixty_four),
+            ];
+            let array_value = self.context.new_array_constructor(None, a_type, &values);
+            then_block.add_assignment(None, result, array_value);
+            then_block.end_with_jump(None, after_block);
+
+            let condition = self.gcc_icmp(IntPredicate::IntEQ, b, zero);
+            else_block.end_with_conditional(None, condition, b0_block, actual_else_block);
+
+            b0_block.add_assignment(None, result, a);
+            b0_block.end_with_jump(None, after_block);
+
+            // NOTE: cast low to its unsigned type in order to perform a logical right shift.
+            let unsigned_type = native_int_type.to_unsigned(&self.cx);
+            let casted_low = self.context.new_cast(None, self.low(a), unsigned_type);
+            let shift_value = self.context.new_cast(None, sixty_four - b, unsigned_type);
+            let high_low = self.context.new_cast(None, casted_low >> shift_value, native_int_type);
+            let values = [
+                self.low(a) << b,
+                (self.high(a) << b) | high_low,
+            ];
+
+            let array_value = self.context.new_array_constructor(None, a_type, &values);
+            actual_else_block.add_assignment(None, result, array_value);
+            actual_else_block.end_with_jump(None, after_block);
+
+            // NOTE: since jumps were added in a place rustc does not expect, the current block in the
+            // state need to be updated.
+            self.block = Some(after_block);
+            *self.cx.current_block.borrow_mut() = Some(after_block);
+
+            result.to_rvalue()
+        }
+    }
+
+    pub fn gcc_bswap(&mut self, mut arg: RValue<'gcc>, width: u64) -> RValue<'gcc> {
+        let arg_type = arg.get_type();
+        if !self.is_native_int_type(arg_type) {
+            let native_int_type = arg_type.dyncast_array().expect("get element type");
+            let lsb = self.context.new_array_access(None, arg, self.context.new_rvalue_from_int(self.int_type, 0)).to_rvalue();
+            let swapped_lsb = self.gcc_bswap(lsb, width / 2);
+            let swapped_lsb = self.context.new_cast(None, swapped_lsb, native_int_type);
+            let msb = self.context.new_array_access(None, arg, self.context.new_rvalue_from_int(self.int_type, 1)).to_rvalue();
+            let swapped_msb = self.gcc_bswap(msb, width / 2);
+            let swapped_msb = self.context.new_cast(None, swapped_msb, native_int_type);
+
+            // NOTE: we also need to swap the two elements here, in addition to swapping inside
+            // the elements themselves like done above.
+            return self.context.new_array_constructor(None, arg_type, &[swapped_msb, swapped_lsb]);
+        }
+
+        // TODO(antoyo): check if it's faster to use string literals and a
+        // match instead of format!.
+        let bswap = self.cx.context.get_builtin_function(&format!("__builtin_bswap{}", width));
+        // FIXME(antoyo): this cast should not be necessary. Remove
+        // when having proper sized integer types.
+        let param_type = bswap.get_param(0).to_rvalue().get_type();
+        if param_type != arg_type {
+            arg = self.bitcast(arg, param_type);
+        }
+        self.cx.context.new_call(None, bswap, &[arg])
+    }
+}
+
+impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> {
+    pub fn gcc_int(&self, typ: Type<'gcc>, int: i64) -> RValue<'gcc> {
+        if self.is_native_int_type_or_bool(typ) {
+            self.context.new_rvalue_from_long(typ, i64::try_from(int).expect("i64::try_from"))
+        }
+        else {
+            // NOTE: set the sign in high.
+            self.from_low_high(typ, int, -(int.is_negative() as i64))
+        }
+    }
+
+    pub fn gcc_uint(&self, typ: Type<'gcc>, int: u64) -> RValue<'gcc> {
+        if self.is_native_int_type_or_bool(typ) {
+            self.context.new_rvalue_from_long(typ, u64::try_from(int).expect("u64::try_from") as i64)
+        }
+        else {
+            self.from_low_high(typ, int as i64, 0)
+        }
+    }
+
+    pub fn gcc_uint_big(&self, typ: Type<'gcc>, num: u128) -> RValue<'gcc> {
+        let low = num as u64;
+        let high = (num >> 64) as u64;
+        if num >> 64 != 0 {
+            // FIXME(antoyo): use a new function new_rvalue_from_unsigned_long()?
+            if self.is_native_int_type(typ) {
+                let low = self.context.new_rvalue_from_long(self.u64_type, low as i64);
+                let high = self.context.new_rvalue_from_long(typ, high as i64);
+
+                let sixty_four = self.context.new_rvalue_from_long(typ, 64);
+                let shift = high << sixty_four;
+                shift | self.context.new_cast(None, low, typ)
+            }
+            else {
+                self.from_low_high(typ, low as i64, high as i64)
+            }
+        }
+        else if typ.is_i128(self) {
+            let num = self.context.new_rvalue_from_long(self.u64_type, num as u64 as i64);
+            self.gcc_int_cast(num, typ)
+        }
+        else {
+            self.gcc_uint(typ, num as u64)
+        }
+    }
+
+    pub fn gcc_zero(&self, typ: Type<'gcc>) -> RValue<'gcc> {
+        if self.is_native_int_type_or_bool(typ) {
+            self.context.new_rvalue_zero(typ)
+        }
+        else {
+            self.from_low_high(typ, 0, 0)
+        }
+    }
+
+    pub fn gcc_int_width(&self, typ: Type<'gcc>) -> u64 {
+        if self.is_native_int_type_or_bool(typ) {
+            typ.get_size() as u64 * 8
+        }
+        else {
+            // NOTE: the only unsupported types are u128 and i128.
+            128
+        }
+    }
+
+    fn bitwise_operation(&self, operation: BinaryOp, a: RValue<'gcc>, mut b: RValue<'gcc>) -> RValue<'gcc> {
+        let a_type = a.get_type();
+        let b_type = b.get_type();
+        let a_native = self.is_native_int_type_or_bool(a_type);
+        let b_native = self.is_native_int_type_or_bool(b_type);
+        if a_native && b_native {
+            if a_type != b_type {
+                b = self.context.new_cast(None, b, a_type);
+            }
+            self.context.new_binary_op(None, operation, a_type, a, b)
+        }
+        else {
+            assert!(!a_native && !b_native, "both types should either be native or non-native for or operation");
+            let native_int_type = a_type.dyncast_array().expect("get element type");
+            let values = [
+                self.context.new_binary_op(None, operation, native_int_type, self.low(a), self.low(b)),
+                self.context.new_binary_op(None, operation, native_int_type, self.high(a), self.high(b)),
+            ];
+            self.context.new_array_constructor(None, a_type, &values)
+        }
+    }
+
+    pub fn gcc_or(&self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> {
+        self.bitwise_operation(BinaryOp::BitwiseOr, a, b)
+    }
+
+    // TODO(antoyo): can we use https://github.com/rust-lang/compiler-builtins/blob/master/src/int/mod.rs#L379 instead?
+    pub fn gcc_int_cast(&self, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> {
+        let value_type = value.get_type();
+        if self.is_native_int_type_or_bool(dest_typ) && self.is_native_int_type_or_bool(value_type) {
+            self.context.new_cast(None, value, dest_typ)
+        }
+        else if self.is_native_int_type_or_bool(dest_typ) {
+            self.context.new_cast(None, self.low(value), dest_typ)
+        }
+        else if self.is_native_int_type_or_bool(value_type) {
+            let dest_element_type = dest_typ.dyncast_array().expect("get element type");
+
+            // NOTE: set the sign of the value.
+            let zero = self.context.new_rvalue_zero(value_type);
+            let is_negative = self.context.new_comparison(None, ComparisonOp::LessThan, value, zero);
+            let is_negative = self.gcc_int_cast(is_negative, dest_element_type);
+            let values = [
+                self.context.new_cast(None, value, dest_element_type),
+                self.context.new_unary_op(None, UnaryOp::Minus, dest_element_type, is_negative),
+            ];
+            self.context.new_array_constructor(None, dest_typ, &values)
+        }
+        else {
+            // Since u128 and i128 are the only types that can be unsupported, we know the type of
+            // value and the destination type have the same size, so a bitcast is fine.
+            self.context.new_bitcast(None, value, dest_typ)
+        }
+    }
+
+    fn int_to_float_cast(&self, signed: bool, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> {
+        let value_type = value.get_type();
+        if self.is_native_int_type_or_bool(value_type) {
+            return self.context.new_cast(None, value, dest_typ);
+        }
+
+        let name_suffix =
+            match self.type_kind(dest_typ) {
+                TypeKind::Float => "tisf",
+                TypeKind::Double => "tidf",
+                kind => panic!("cannot cast a non-native integer to type {:?}", kind),
+            };
+        let sign =
+            if signed {
+                ""
+            }
+            else {
+                "un"
+            };
+        let func_name = format!("__float{}{}", sign, name_suffix);
+        let param = self.context.new_parameter(None, value_type, "n");
+        let func = self.context.new_function(None, FunctionType::Extern, dest_typ, &[param], func_name, false);
+        self.context.new_call(None, func, &[value])
+    }
+
+    pub fn gcc_int_to_float_cast(&self, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> {
+        self.int_to_float_cast(true, value, dest_typ)
+    }
+
+    pub fn gcc_uint_to_float_cast(&self, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> {
+        self.int_to_float_cast(false, value, dest_typ)
+    }
+
+    fn float_to_int_cast(&self, signed: bool, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> {
+        let value_type = value.get_type();
+        if self.is_native_int_type_or_bool(dest_typ) {
+            return self.context.new_cast(None, value, dest_typ);
+        }
+
+        let name_suffix =
+            match self.type_kind(value_type) {
+                TypeKind::Float => "sfti",
+                TypeKind::Double => "dfti",
+                kind => panic!("cannot cast a {:?} to non-native integer", kind),
+            };
+        let sign =
+            if signed {
+                ""
+            }
+            else {
+                "uns"
+            };
+        let func_name = format!("__fix{}{}", sign, name_suffix);
+        let param = self.context.new_parameter(None, value_type, "n");
+        let func = self.context.new_function(None, FunctionType::Extern, dest_typ, &[param], func_name, false);
+        self.context.new_call(None, func, &[value])
+    }
+
+    pub fn gcc_float_to_int_cast(&self, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> {
+        self.float_to_int_cast(true, value, dest_typ)
+    }
+
+    pub fn gcc_float_to_uint_cast(&self, value: RValue<'gcc>, dest_typ: Type<'gcc>) -> RValue<'gcc> {
+        self.float_to_int_cast(false, value, dest_typ)
+    }
+
+    fn high(&self, value: RValue<'gcc>) -> RValue<'gcc> {
+        self.context.new_array_access(None, value, self.context.new_rvalue_from_int(self.int_type, 1))
+            .to_rvalue()
+    }
+
+    fn low(&self, value: RValue<'gcc>) -> RValue<'gcc> {
+        self.context.new_array_access(None, value, self.context.new_rvalue_from_int(self.int_type, 0))
+            .to_rvalue()
+    }
+
+    fn from_low_high(&self, typ: Type<'gcc>, low: i64, high: i64) -> RValue<'gcc> {
+        let native_int_type = typ.dyncast_array().expect("get element type");
+        let values = [
+            self.context.new_rvalue_from_long(native_int_type, low),
+            self.context.new_rvalue_from_long(native_int_type, high),
+        ];
+        self.context.new_array_constructor(None, typ, &values)
+    }
+}
diff --git a/src/intrinsic/mod.rs b/src/intrinsic/mod.rs
index 572ac559d09..7cd0f944f2f 100644
--- a/src/intrinsic/mod.rs
+++ b/src/intrinsic/mod.rs
@@ -1,7 +1,7 @@
 pub mod llvm;
 mod simd;
 
-use gccjit::{ComparisonOp, Function, RValue, ToRValue, Type, UnaryOp};
+use gccjit::{ComparisonOp, Function, RValue, ToRValue, Type, UnaryOp, FunctionType};
 use rustc_codegen_ssa::MemFlags;
 use rustc_codegen_ssa::base::wants_msvc_seh;
 use rustc_codegen_ssa::common::{IntPredicate, span_invalid_monomorphization_error};
@@ -175,11 +175,11 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> {
 
                                     let arg = args[0].immediate();
                                     let result = func.new_local(None, arg.get_type(), "zeros");
-                                    let zero = self.cx.context.new_rvalue_zero(arg.get_type());
-                                    let cond = self.cx.context.new_comparison(None, ComparisonOp::Equals, arg, zero);
+                                    let zero = self.cx.gcc_zero(arg.get_type());
+                                    let cond = self.gcc_icmp(IntPredicate::IntEQ, arg, zero);
                                     self.llbb().end_with_conditional(None, cond, then_block, else_block);
 
-                                    let zero_result = self.cx.context.new_rvalue_from_long(arg.get_type(), width as i64);
+                                    let zero_result = self.cx.gcc_uint(arg.get_type(), width);
                                     then_block.add_assignment(None, result, zero_result);
                                     then_block.end_with_jump(None, after_block);
 
@@ -195,8 +195,8 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> {
                                             sym::cttz => self.count_trailing_zeroes(width, arg),
                                             _ => unreachable!(),
                                         };
-                                    else_block.add_assignment(None, result, zeros);
-                                    else_block.end_with_jump(None, after_block);
+                                    self.llbb().add_assignment(None, result, zeros);
+                                    self.llbb().end_with_jump(None, after_block);
 
                                     // NOTE: since jumps were added in a place rustc does not
                                     // expect, the current blocks in the state need to be updated.
@@ -217,17 +217,7 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> {
                                         args[0].immediate() // byte swap a u8/i8 is just a no-op
                                     }
                                     else {
-                                        // TODO(antoyo): check if it's faster to use string literals and a
-                                        // match instead of format!.
-                                        let bswap = self.cx.context.get_builtin_function(&format!("__builtin_bswap{}", width));
-                                        let mut arg = args[0].immediate();
-                                        // FIXME(antoyo): this cast should not be necessary. Remove
-                                        // when having proper sized integer types.
-                                        let param_type = bswap.get_param(0).to_rvalue().get_type();
-                                        if param_type != arg.get_type() {
-                                            arg = self.bitcast(arg, param_type);
-                                        }
-                                        self.cx.context.new_call(None, bswap, &[arg])
+                                        self.gcc_bswap(args[0].immediate(), width)
                                     }
                                 },
                                 sym::bitreverse => self.bit_reverse(width, args[0].immediate()),
@@ -526,7 +516,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
 
         let value =
             if result_type.is_signed(self.cx) {
-                self.context.new_cast(None, value, typ)
+                self.gcc_int_cast(value, typ)
             }
             else {
                 value
@@ -673,30 +663,33 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
                 },
                 128 => {
                     // TODO(antoyo): find a more efficient implementation?
-                    let sixty_four = self.context.new_rvalue_from_long(typ, 64);
-                    let high = self.context.new_cast(None, value >> sixty_four, self.u64_type);
-                    let low = self.context.new_cast(None, value, self.u64_type);
+                    let sixty_four = self.gcc_int(typ, 64);
+                    let right_shift = self.gcc_lshr(value, sixty_four);
+                    let high = self.gcc_int_cast(right_shift, self.u64_type);
+                    let low = self.gcc_int_cast(value, self.u64_type);
 
                     let reversed_high = self.bit_reverse(64, high);
                     let reversed_low = self.bit_reverse(64, low);
 
-                    let new_low = self.context.new_cast(None, reversed_high, typ);
-                    let new_high = self.context.new_cast(None, reversed_low, typ) << sixty_four;
+                    let new_low = self.gcc_int_cast(reversed_high, typ);
+                    let new_high = self.shl(self.gcc_int_cast(reversed_low, typ), sixty_four);
 
-                    new_low | new_high
+                    self.gcc_or(new_low, new_high)
                 },
                 _ => {
                     panic!("cannot bit reverse with width = {}", width);
                 },
             };
 
-        self.context.new_cast(None, result, result_type)
+        self.gcc_int_cast(result, result_type)
     }
 
-    fn count_leading_zeroes(&self, width: u64, arg: RValue<'gcc>) -> RValue<'gcc> {
+    fn count_leading_zeroes(&mut self, width: u64, arg: RValue<'gcc>) -> RValue<'gcc> {
         // TODO(antoyo): use width?
         let arg_type = arg.get_type();
         let count_leading_zeroes =
+            // TODO(antoyo): write a new function Type::is_compatible_with(&Type) and use it here
+            // instead of using is_uint().
             if arg_type.is_uint(&self.cx) {
                 "__builtin_clz"
             }
@@ -712,9 +705,10 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
                 let result = self.current_func()
                     .new_local(None, array_type, "count_loading_zeroes_results");
 
-                let sixty_four = self.context.new_rvalue_from_long(arg_type, 64);
-                let high = self.context.new_cast(None, arg >> sixty_four, self.u64_type);
-                let low = self.context.new_cast(None, arg, self.u64_type);
+                let sixty_four = self.const_uint(arg_type, 64);
+                let shift = self.lshr(arg, sixty_four);
+                let high = self.gcc_int_cast(shift, self.u64_type);
+                let low = self.gcc_int_cast(arg, self.u64_type);
 
                 let zero = self.context.new_rvalue_zero(self.usize_type);
                 let one = self.context.new_rvalue_one(self.usize_type);
@@ -723,17 +717,18 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
                 let clzll = self.context.get_builtin_function("__builtin_clzll");
 
                 let first_elem = self.context.new_array_access(None, result, zero);
-                let first_value = self.context.new_cast(None, self.context.new_call(None, clzll, &[high]), arg_type);
+                let first_value = self.gcc_int_cast(self.context.new_call(None, clzll, &[high]), arg_type);
                 self.llbb()
                     .add_assignment(None, first_elem, first_value);
 
                 let second_elem = self.context.new_array_access(None, result, one);
-                let second_value = self.context.new_cast(None, self.context.new_call(None, clzll, &[low]), arg_type) + sixty_four;
+                let cast = self.gcc_int_cast(self.context.new_call(None, clzll, &[low]), arg_type);
+                let second_value = self.add(cast, sixty_four);
                 self.llbb()
                     .add_assignment(None, second_elem, second_value);
 
                 let third_elem = self.context.new_array_access(None, result, two);
-                let third_value = self.context.new_rvalue_from_long(arg_type, 128);
+                let third_value = self.const_uint(arg_type, 128);
                 self.llbb()
                     .add_assignment(None, third_elem, third_value);
 
@@ -749,13 +744,13 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
 
                 let res = self.context.new_array_access(None, result, index);
 
-                return self.context.new_cast(None, res, arg_type);
+                return self.gcc_int_cast(res.to_rvalue(), arg_type);
             }
             else {
-                let count_leading_zeroes = self.context.get_builtin_function("__builtin_clz");
-                let arg = self.context.new_cast(None, arg, self.uint_type);
-                let diff = self.int_width(self.uint_type) - self.int_width(arg_type);
-                let diff = self.context.new_rvalue_from_long(self.int_type, diff);
+                let count_leading_zeroes = self.context.get_builtin_function("__builtin_clzll");
+                let arg = self.context.new_cast(None, arg, self.ulonglong_type);
+                let diff = self.ulonglong_type.get_size() as i64 - arg_type.get_size() as i64;
+                let diff = self.context.new_rvalue_from_long(self.int_type, diff * 8);
                 let res = self.context.new_call(None, count_leading_zeroes, &[arg]) - diff;
                 return self.context.new_cast(None, res, arg_type);
             };
@@ -764,18 +759,20 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         self.context.new_cast(None, res, arg_type)
     }
 
-    fn count_trailing_zeroes(&self, _width: u64, arg: RValue<'gcc>) -> RValue<'gcc> {
+    fn count_trailing_zeroes(&mut self, _width: u64, arg: RValue<'gcc>) -> RValue<'gcc> {
         let result_type = arg.get_type();
         let arg =
             if result_type.is_signed(self.cx) {
                 let new_type = result_type.to_unsigned(self.cx);
-                self.context.new_cast(None, arg, new_type)
+                self.gcc_int_cast(arg, new_type)
             }
             else {
                 arg
             };
         let arg_type = arg.get_type();
         let (count_trailing_zeroes, expected_type) =
+            // TODO(antoyo): write a new function Type::is_compatible_with(&Type) and use it here
+            // instead of using is_uint().
             if arg_type.is_uchar(&self.cx) || arg_type.is_ushort(&self.cx) || arg_type.is_uint(&self.cx) {
                 // NOTE: we don't need to & 0xFF for uchar because the result is undefined on zero.
                 ("__builtin_ctz", self.cx.uint_type)
@@ -792,9 +789,10 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
                 let result = self.current_func()
                     .new_local(None, array_type, "count_loading_zeroes_results");
 
-                let sixty_four = self.context.new_rvalue_from_long(arg_type, 64);
-                let high = self.context.new_cast(None, arg >> sixty_four, self.u64_type);
-                let low = self.context.new_cast(None, arg, self.u64_type);
+                let sixty_four = self.gcc_int(arg_type, 64);
+                let shift = self.gcc_lshr(arg, sixty_four);
+                let high = self.gcc_int_cast(shift, self.u64_type);
+                let low = self.gcc_int_cast(arg, self.u64_type);
 
                 let zero = self.context.new_rvalue_zero(self.usize_type);
                 let one = self.context.new_rvalue_one(self.usize_type);
@@ -803,17 +801,17 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
                 let ctzll = self.context.get_builtin_function("__builtin_ctzll");
 
                 let first_elem = self.context.new_array_access(None, result, zero);
-                let first_value = self.context.new_cast(None, self.context.new_call(None, ctzll, &[low]), arg_type);
+                let first_value = self.gcc_int_cast(self.context.new_call(None, ctzll, &[low]), arg_type);
                 self.llbb()
                     .add_assignment(None, first_elem, first_value);
 
                 let second_elem = self.context.new_array_access(None, result, one);
-                let second_value = self.context.new_cast(None, self.context.new_call(None, ctzll, &[high]), arg_type) + sixty_four;
+                let second_value = self.gcc_add(self.gcc_int_cast(self.context.new_call(None, ctzll, &[high]), arg_type), sixty_four);
                 self.llbb()
                     .add_assignment(None, second_elem, second_value);
 
                 let third_elem = self.context.new_array_access(None, result, two);
-                let third_value = self.context.new_rvalue_from_long(arg_type, 128);
+                let third_value = self.gcc_int(arg_type, 128);
                 self.llbb()
                     .add_assignment(None, third_elem, third_value);
 
@@ -829,10 +827,20 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
 
                 let res = self.context.new_array_access(None, result, index);
 
-                return self.context.new_cast(None, res, result_type);
+                return self.gcc_int_cast(res.to_rvalue(), result_type);
             }
             else {
-                unimplemented!("count_trailing_zeroes for {:?}", arg_type);
+                let count_trailing_zeroes = self.context.get_builtin_function("__builtin_ctzll");
+                let arg_size = arg_type.get_size();
+                let casted_arg = self.context.new_cast(None, arg, self.ulonglong_type);
+                let byte_diff = self.ulonglong_type.get_size() as i64 - arg_size as i64;
+                let diff = self.context.new_rvalue_from_long(self.int_type, byte_diff * 8);
+                let mask = self.context.new_rvalue_from_long(arg_type, -1); // To get the value with all bits set.
+                let masked = mask & self.context.new_unary_op(None, UnaryOp::BitwiseNegate, arg_type, arg);
+                let cond = self.context.new_comparison(None, ComparisonOp::Equals, masked, mask);
+                let diff = diff * self.context.new_cast(None, cond, self.int_type);
+                let res = self.context.new_call(None, count_trailing_zeroes, &[casted_arg]) - diff;
+                return self.context.new_cast(None, res, result_type);
             };
         let count_trailing_zeroes = self.context.get_builtin_function(count_trailing_zeroes);
         let arg =
@@ -846,18 +854,14 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
         self.context.new_cast(None, res, result_type)
     }
 
-    fn int_width(&self, typ: Type<'gcc>) -> i64 {
-        self.cx.int_width(typ) as i64
-    }
-
-    fn pop_count(&self, value: RValue<'gcc>) -> RValue<'gcc> {
+    fn pop_count(&mut self, value: RValue<'gcc>) -> RValue<'gcc> {
         // TODO(antoyo): use the optimized version with fewer operations.
         let result_type = value.get_type();
         let value_type = result_type.to_unsigned(self.cx);
 
         let value =
             if result_type.is_signed(self.cx) {
-                self.context.new_cast(None, value, value_type)
+                self.gcc_int_cast(value, value_type)
             }
             else {
                 value
@@ -867,13 +871,14 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
             // TODO(antoyo): implement in the normal algorithm below to have a more efficient
             // implementation (that does not require a call to __popcountdi2).
             let popcount = self.context.get_builtin_function("__builtin_popcountll");
-            let sixty_four = self.context.new_rvalue_from_long(value_type, 64);
-            let high = self.context.new_cast(None, value >> sixty_four, self.cx.ulonglong_type);
+            let sixty_four = self.gcc_int(value_type, 64);
+            let right_shift = self.gcc_lshr(value, sixty_four);
+            let high = self.gcc_int_cast(right_shift, self.cx.ulonglong_type);
             let high = self.context.new_call(None, popcount, &[high]);
-            let low = self.context.new_cast(None, value, self.cx.ulonglong_type);
+            let low = self.gcc_int_cast(value, self.cx.ulonglong_type);
             let low = self.context.new_call(None, popcount, &[low]);
             let res = high + low;
-            return self.context.new_cast(None, res, result_type);
+            return self.gcc_int_cast(res, result_type);
         }
 
         // First step.
@@ -935,13 +940,14 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
 
     // Algorithm from: https://blog.regehr.org/archives/1063
     fn rotate_left(&mut self, value: RValue<'gcc>, shift: RValue<'gcc>, width: u64) -> RValue<'gcc> {
-        let max = self.context.new_rvalue_from_long(shift.get_type(), width as i64);
-        let shift = shift % max;
+        let max = self.const_uint(shift.get_type(), width);
+        let shift = self.urem(shift, max);
         let lhs = self.shl(value, shift);
+        let result_neg = self.neg(shift);
         let result_and =
             self.and(
-                self.context.new_unary_op(None, UnaryOp::Minus, shift.get_type(), shift),
-                self.context.new_rvalue_from_long(shift.get_type(), width as i64 - 1),
+                result_neg,
+                self.const_uint(shift.get_type(), width - 1),
             );
         let rhs = self.lshr(value, result_and);
         self.or(lhs, rhs)
@@ -949,13 +955,14 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
 
     // Algorithm from: https://blog.regehr.org/archives/1063
     fn rotate_right(&mut self, value: RValue<'gcc>, shift: RValue<'gcc>, width: u64) -> RValue<'gcc> {
-        let max = self.context.new_rvalue_from_long(shift.get_type(), width as i64);
-        let shift = shift % max;
+        let max = self.const_uint(shift.get_type(), width);
+        let shift = self.urem(shift, max);
         let lhs = self.lshr(value, shift);
+        let result_neg = self.neg(shift);
         let result_and =
             self.and(
-                self.context.new_unary_op(None, UnaryOp::Minus, shift.get_type(), shift),
-                self.context.new_rvalue_from_long(shift.get_type(), width as i64 - 1),
+                result_neg,
+                self.const_uint(shift.get_type(), width - 1),
             );
         let rhs = self.shl(value, result_and);
         self.or(lhs, rhs)
@@ -1015,31 +1022,52 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> {
     fn saturating_sub(&mut self, lhs: RValue<'gcc>, rhs: RValue<'gcc>, signed: bool, width: u64) -> RValue<'gcc> {
         if signed {
             // Also based on algorithm from: https://stackoverflow.com/a/56531252/389119
-            let func_name =
-                match width {
-                    8 => "__builtin_sub_overflow",
-                    16 => "__builtin_sub_overflow",
-                    32 => "__builtin_ssub_overflow",
-                    64 => "__builtin_ssubll_overflow",
-                    128 => "__builtin_sub_overflow",
-                    _ => unreachable!(),
-                };
-            let overflow_func = self.context.get_builtin_function(func_name);
             let result_type = lhs.get_type();
             let func = self.current_func.borrow().expect("func");
             let res = func.new_local(None, result_type, "saturating_diff");
-            let overflow = self.overflow_call(overflow_func, &[lhs, rhs, res.get_address(None)], None);
+            let supports_native_type = self.is_native_int_type(result_type);
+            let overflow =
+                if supports_native_type {
+                    let func_name =
+                        match width {
+                            8 => "__builtin_sub_overflow",
+                            16 => "__builtin_sub_overflow",
+                            32 => "__builtin_ssub_overflow",
+                            64 => "__builtin_ssubll_overflow",
+                            128 => "__builtin_sub_overflow",
+                            _ => unreachable!(),
+                        };
+                    let overflow_func = self.context.get_builtin_function(func_name);
+                    self.overflow_call(overflow_func, &[lhs, rhs, res.get_address(None)], None)
+                }
+                else {
+                    let func_name =
+                        match width {
+                            128 => "__rust_i128_subo",
+                            _ => unreachable!(),
+                        };
+                    let param_a = self.context.new_parameter(None, result_type, "a");
+                    let param_b = self.context.new_parameter(None, result_type, "b");
+                    let result_field = self.context.new_field(None, result_type, "result");
+                    let overflow_field = self.context.new_field(None, self.bool_type, "overflow");
+                    let return_type = self.context.new_struct_type(None, "result_overflow", &[result_field, overflow_field]);
+                    let func = self.context.new_function(None, FunctionType::Extern, return_type.as_type(), &[param_a, param_b], func_name, false);
+                    let result = self.context.new_call(None, func, &[lhs, rhs]);
+                    let overflow = result.access_field(None, overflow_field);
+                    let int_result = result.access_field(None, result_field);
+                    self.llbb().add_assignment(None, res, int_result);
+                    overflow
+                };
 
             let then_block = func.new_block("then");
             let after_block = func.new_block("after");
 
-            let unsigned_type = self.context.new_int_type(width as i32 / 8, false);
-            let shifted = self.context.new_cast(None, lhs, unsigned_type) >> self.context.new_rvalue_from_int(unsigned_type, width as i32 - 1);
-            let uint_max = self.context.new_unary_op(None, UnaryOp::BitwiseNegate, unsigned_type,
-                self.context.new_rvalue_from_int(unsigned_type, 0)
-            );
-            let int_max = uint_max >> self.context.new_rvalue_one(unsigned_type);
-            then_block.add_assignment(None, res, self.context.new_cast(None, shifted + int_max, result_type));
+            // NOTE: convert the type to unsigned to have an unsigned shift.
+            let unsigned_type = result_type.to_unsigned(&self.cx);
+            let shifted = self.gcc_lshr(self.gcc_int_cast(lhs, unsigned_type), self.gcc_int(unsigned_type, width as i64 - 1));
+            let uint_max = self.gcc_not(self.gcc_int(unsigned_type, 0));
+            let int_max = self.gcc_lshr(uint_max, self.gcc_int(unsigned_type, 1));
+            then_block.add_assignment(None, res, self.gcc_int_cast(self.gcc_add(shifted, int_max), result_type));
             then_block.end_with_jump(None, after_block);
 
             self.llbb().end_with_conditional(None, overflow, then_block, after_block);
diff --git a/src/lib.rs b/src/lib.rs
index 9b62bf41837..ca283e1380c 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -1,4 +1,5 @@
 /*
+ * TODO(antoyo): implement equality in libgccjit based on https://zpz.github.io/blog/overloading-equality-operator-in-cpp-class-hierarchy/ (for type equality?)
  * TODO(antoyo): support #[inline] attributes.
  * TODO(antoyo): support LTO (gcc's equivalent to Thin LTO is enabled by -fwhopr: https://stackoverflow.com/questions/64954525/does-gcc-have-thin-lto).
  *
@@ -21,6 +22,7 @@ extern crate rustc_middle;
 extern crate rustc_session;
 extern crate rustc_span;
 extern crate rustc_target;
+extern crate tempfile;
 
 // This prevents duplicating functions and statics that are already part of the host rustc process.
 #[allow(unused_extern_crates)]
@@ -40,15 +42,16 @@ mod context;
 mod coverageinfo;
 mod debuginfo;
 mod declare;
+mod int;
 mod intrinsic;
 mod mono_item;
 mod type_;
 mod type_of;
 
 use std::any::Any;
-use std::sync::Arc;
+use std::sync::{Arc, Mutex};
 
-use gccjit::{Context, OptimizationLevel};
+use gccjit::{Context, OptimizationLevel, CType};
 use rustc_ast::expand::allocator::AllocatorKind;
 use rustc_codegen_ssa::{CodegenResults, CompiledModule, ModuleCodegen};
 use rustc_codegen_ssa::base::codegen_crate;
@@ -65,6 +68,7 @@ use rustc_session::config::{Lto, OptLevel, OutputFilenames};
 use rustc_session::Session;
 use rustc_span::Symbol;
 use rustc_span::fatal_error::FatalError;
+use tempfile::TempDir;
 
 pub struct PrintOnPanic<F: Fn() -> String>(pub F);
 
@@ -77,13 +81,24 @@ impl<F: Fn() -> String> Drop for PrintOnPanic<F> {
 }
 
 #[derive(Clone)]
-pub struct GccCodegenBackend;
+pub struct GccCodegenBackend {
+    supports_128bit_integers: Arc<Mutex<bool>>,
+}
 
 impl CodegenBackend for GccCodegenBackend {
     fn init(&self, sess: &Session) {
         if sess.lto() != Lto::No {
             sess.warn("LTO is not supported. You may get a linker error.");
         }
+
+        let temp_dir = TempDir::new().expect("cannot create temporary directory");
+        let temp_file = temp_dir.into_path().join("result.asm");
+        let check_context = Context::default();
+        check_context.set_print_errors_to_stderr(false);
+        let _int128_ty = check_context.new_c_type(CType::UInt128t);
+        // NOTE: we cannot just call compile() as this would require other files than libgccjit.so.
+        check_context.compile_to_file(gccjit::OutputKind::Assembler, temp_file.to_str().expect("path to str"));
+        *self.supports_128bit_integers.lock().expect("lock") = check_context.get_last_error() == Ok(None);
     }
 
     fn codegen_crate<'tcx>(&self, tcx: TyCtxt<'tcx>, metadata: EncodedMetadata, need_metadata_module: bool) -> Box<dyn Any> {
@@ -129,7 +144,7 @@ impl ExtraBackendMethods for GccCodegenBackend {
     }
 
     fn compile_codegen_unit<'tcx>(&self, tcx: TyCtxt<'tcx>, cgu_name: Symbol) -> (ModuleCodegen<Self::Module>, u64) {
-        base::compile_codegen_unit(tcx, cgu_name)
+        base::compile_codegen_unit(tcx, cgu_name, *self.supports_128bit_integers.lock().expect("lock"))
     }
 
     fn target_machine_factory(&self, _sess: &Session, _opt_level: OptLevel) -> TargetMachineFactoryFn<Self> {
@@ -237,7 +252,9 @@ impl WriteBackendMethods for GccCodegenBackend {
 /// This is the entrypoint for a hot plugged rustc_codegen_gccjit
 #[no_mangle]
 pub fn __rustc_codegen_backend() -> Box<dyn CodegenBackend> {
-    Box::new(GccCodegenBackend)
+    Box::new(GccCodegenBackend {
+        supports_128bit_integers: Arc::new(Mutex::new(false)),
+    })
 }
 
 fn to_gcc_opt_level(optlevel: Option<OptLevel>) -> OptimizationLevel {
diff --git a/src/type_.rs b/src/type_.rs
index 28e2adc492b..3c96cd6afc2 100644
--- a/src/type_.rs
+++ b/src/type_.rs
@@ -7,7 +7,6 @@ use rustc_middle::bug;
 use rustc_middle::ty::layout::TyAndLayout;
 use rustc_target::abi::{AddressSpace, Align, Integer, Size};
 
-use crate::common::TypeReflection;
 use crate::context::CodegenCx;
 use crate::type_of::LayoutGccExt;
 
@@ -119,9 +118,15 @@ impl<'gcc, 'tcx> BaseTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
     }
 
     fn type_kind(&self, typ: Type<'gcc>) -> TypeKind {
-        if typ.is_integral() {
+        if self.is_int_type_or_bool(typ) {
             TypeKind::Integer
         }
+        else if typ == self.float_type {
+            TypeKind::Float
+        }
+        else if typ == self.double_type {
+            TypeKind::Double
+        }
         else if typ.dyncast_vector().is_some() {
             TypeKind::Vector
         }
@@ -175,24 +180,7 @@ impl<'gcc, 'tcx> BaseTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> {
     }
 
     fn int_width(&self, typ: Type<'gcc>) -> u64 {
-        if typ.is_i8(self) || typ.is_u8(self) {
-            8
-        }
-        else if typ.is_i16(self) || typ.is_u16(self) {
-            16
-        }
-        else if typ.is_i32(self) || typ.is_u32(self) {
-            32
-        }
-        else if typ.is_i64(self) || typ.is_u64(self) {
-            64
-        }
-        else if typ.is_i128(self) || typ.is_u128(self) {
-            128
-        }
-        else {
-            panic!("Cannot get width of int type {:?}", typ);
-        }
+        self.gcc_int_width(typ)
     }
 
     fn val_ty(&self, value: RValue<'gcc>) -> Type<'gcc> {
diff --git a/test.sh b/test.sh
index 70bd86edcbe..4f05013440b 100755
--- a/test.sh
+++ b/test.sh
@@ -4,7 +4,7 @@
 
 set -e
 
-if [ -f ./gcc_path ]; then 
+if [ -f ./gcc_path ]; then
     export GCC_PATH=$(cat gcc_path)
 else
     echo 'Please put the path to your custom build of libgccjit in the file `gcc_path`, see Readme.md for details'
@@ -14,14 +14,26 @@ fi
 export LD_LIBRARY_PATH="$GCC_PATH"
 export LIBRARY_PATH="$GCC_PATH"
 
+features=
+
+if [[ "$1" == "--features" ]]; then
+    shift
+    features="--features $1"
+    shift
+fi
+
 if [[ "$1" == "--release" ]]; then
     export CHANNEL='release'
-    CARGO_INCREMENTAL=1 cargo rustc --release
+    CARGO_INCREMENTAL=1 cargo rustc --release $features
     shift
 else
     echo $LD_LIBRARY_PATH
     export CHANNEL='debug'
-    cargo rustc
+    cargo rustc $features
+fi
+
+if [[ "$1" == "--build" ]]; then
+    exit
 fi
 
 source config.sh
@@ -206,6 +218,14 @@ case $1 in
         clean_ui_tests
         ;;
 
+    "--std-tests")
+        std_tests
+        ;;
+
+    "--build-sysroot")
+        build_sysroot
+        ;;
+
     *)
         clean
         mini_tests
diff --git a/tests/run/int.rs b/tests/run/int.rs
new file mode 100644
index 00000000000..7a62fc7d1f7
--- /dev/null
+++ b/tests/run/int.rs
@@ -0,0 +1,151 @@
+// Compiler:
+//
+// Run-time:
+//   status: 0
+
+#![feature(arbitrary_self_types, auto_traits, core_intrinsics, lang_items, start, intrinsics)]
+
+#![no_std]
+
+mod intrinsics {
+    extern "rust-intrinsic" {
+        pub fn abort() -> !;
+    }
+}
+
+/*
+ * Core
+ */
+
+mod libc {
+    #[link(name = "c")]
+    extern "C" {
+        pub fn puts(s: *const u8) -> i32;
+    }
+}
+
+#[panic_handler]
+fn panic_handler(_: &core::panic::PanicInfo) -> ! {
+    unsafe {
+        core::intrinsics::abort();
+    }
+}
+
+/*
+ * Code
+ */
+
+#[start]
+fn main(argc: isize, _argv: *const *const u8) -> isize {
+    let var = 134217856_u128;
+    let var2 = 10475372733397991552_u128;
+    let var3 = 193236519889708027473620326106273939584_u128;
+    let var4 = 123236519889708027473620326106273939584_u128;
+    let var5 = 153236519889708027473620326106273939584_u128;
+    let var6 = 18446744073709551616_i128;
+    let var7 = 170141183460469231731687303715884105728_u128;
+
+    // Shifts.
+    assert_eq!(var << (argc as u128 - 1), var);
+    assert_eq!(var << argc as u128, 268435712);
+    assert_eq!(var << (argc + 32) as u128, 1152922604118474752);
+    assert_eq!(var << (argc + 48) as u128, 75557935783508361347072);
+    assert_eq!(var << (argc + 60) as u128, 309485304969250248077606912);
+    assert_eq!(var << (argc + 62) as u128, 1237941219877000992310427648);
+    assert_eq!(var << (argc + 63) as u128, 2475882439754001984620855296);
+    assert_eq!(var << (argc + 80) as u128, 324518863143436548128224745357312);
+
+    assert_eq!(var2 << argc as u128, 20950745466795983104);
+    assert_eq!(var2 << (argc as u128 - 1), var2);
+    assert_eq!(var2 << (argc + 32) as u128, 89982766606709001335848566784);
+    assert_eq!(var2 << (argc + 48) as u128, 5897110592337281111546171672756224);
+    assert_eq!(var2 << (argc + 60) as u128, 24154564986213503432893119171609493504);
+    assert_eq!(var2 << (argc + 62) as u128, 96618259944854013731572476686437974016);
+    assert_eq!(var2 << (argc + 63) as u128, 193236519889708027463144953372875948032);
+
+    assert_eq!(var3 << argc as u128, 46190672858477591483866044780779667712);
+    assert_eq!(var3 << (argc as u128 - 1), var3);
+    assert_eq!(var3 << (argc + 32) as u128, 21267668304951024224840338247585366016);
+    assert_eq!(var3 << (argc + 48) as u128, 1335125106377253154015353231953100800);
+    assert_eq!(var3 << (argc + 60) as u128, 24154564986213503432893119171609493504);
+    assert_eq!(var3 << (argc + 62) as u128, 96618259944854013731572476686437974016);
+    assert_eq!(var3 << (argc + 63) as u128, 193236519889708027463144953372875948032);
+
+    assert_eq!(var >> (argc as u128 - 1), var);
+    assert_eq!(var >> argc as u128, 67108928);
+    assert_eq!(var >> (argc + 32) as u128, 0);
+    assert_eq!(var >> (argc + 48) as u128, 0);
+    assert_eq!(var >> (argc + 60) as u128, 0);
+    assert_eq!(var >> (argc + 62) as u128, 0);
+    assert_eq!(var >> (argc + 63) as u128, 0);
+
+    assert_eq!(var2 >> argc as u128, 5237686366698995776);
+    assert_eq!(var2 >> (argc as u128 - 1), var2);
+    assert_eq!(var2 >> (argc + 32) as u128, 1219493888);
+    assert_eq!(var2 >> (argc + 48) as u128, 18608);
+    assert_eq!(var2 >> (argc + 60) as u128, 4);
+    assert_eq!(var2 >> (argc + 62) as u128, 1);
+    assert_eq!(var2 >> (argc + 63) as u128, 0);
+
+    assert_eq!(var3 >> (argc as u128 - 1), var3);
+    assert_eq!(var3 >> argc as u128, 96618259944854013736810163053136969792);
+    assert_eq!(var3 >> (argc + 32) as u128, 22495691651677250335181635584);
+    assert_eq!(var3 >> (argc + 48) as u128, 343257013727985387194544);
+    assert_eq!(var3 >> (argc + 60) as u128, 83802981867183932420);
+    assert_eq!(var3 >> (argc + 62) as u128, 20950745466795983105);
+    assert_eq!(var3 >> (argc + 63) as u128, 10475372733397991552);
+    assert_eq!(var3 >> (argc + 80) as u128, 79920751444992);
+
+    assert_eq!(var6 >> argc as u128, 9223372036854775808);
+    assert_eq!((var6 - 1) >> argc as u128, 9223372036854775807);
+    assert_eq!(var7 >> argc as u128, 85070591730234615865843651857942052864);
+
+    // Casts
+    assert_eq!((var >> (argc + 32) as u128) as u64, 0);
+    assert_eq!((var >> argc as u128) as u64, 67108928);
+
+    // Addition.
+    assert_eq!(var + argc as u128, 134217857);
+
+    assert_eq!(var2 + argc as u128, 10475372733397991553);
+    assert_eq!(var2 + (var2 + argc as u128) as u128, 20950745466795983105);
+
+    assert_eq!(var3 + argc as u128, 193236519889708027473620326106273939585);
+
+    // Subtraction
+    assert_eq!(var - argc as u128, 134217855);
+
+    assert_eq!(var2 - argc as u128, 10475372733397991551);
+
+    assert_eq!(var3 - argc as u128, 193236519889708027473620326106273939583);
+
+    // Multiplication
+    assert_eq!(var * (argc + 1) as u128, 268435712);
+    assert_eq!(var * (argc as u128 + var2), 1405982069077538020949770368);
+
+    assert_eq!(var2 * (argc + 1) as u128, 20950745466795983104);
+    assert_eq!(var2 * (argc as u128 + var2), 109733433903618109003204073240861360256);
+
+    assert_eq!(var3 * argc as u128, 193236519889708027473620326106273939584);
+
+    assert_eq!(var4 * (argc + 1) as u128, 246473039779416054947240652212547879168);
+
+    assert_eq!(var5 * (argc + 1) as u128, 306473039779416054947240652212547879168);
+
+    // Division.
+    assert_eq!(var / (argc + 1) as u128, 67108928);
+    assert_eq!(var / (argc + 2) as u128, 44739285);
+
+    assert_eq!(var2 / (argc + 1) as u128, 5237686366698995776);
+    assert_eq!(var2 / (argc + 2) as u128, 3491790911132663850);
+
+    assert_eq!(var3 / (argc + 1) as u128, 96618259944854013736810163053136969792);
+    assert_eq!(var3 / (argc + 2) as u128, 64412173296569342491206775368757979861);
+    assert_eq!(var3 / (argc as u128 + var4), 1);
+    assert_eq!(var3 / (argc as u128 + var2), 18446744073709551615);
+
+    assert_eq!(var4 / (argc + 1) as u128, 61618259944854013736810163053136969792);
+    assert_eq!(var4 / (argc + 2) as u128, 41078839963236009157873442035424646528);
+
+    0
+}