Avoid memcpy in codegen for more types, notably Vec

scottmcm · scottmcm · commit 2627981b7eb4 · 2023-08-13T23:36:15.000-07:00
PR 111999 set up the framework to be able to do this; this PR expands it to more types than just arrays. Most interestingly, this allows it to work with `Vec<T>` and `String`, so swapping those no longer ends up going through stack like it does today (<https://rust.godbolt.org/z/cKG7o8aaW>). And since this is done in codegen, it's not special for `swap`, and thus will hopefully allow types like this to better optimize in lots of places, with easier SRoA.
diff --git a/compiler/rustc_codegen_llvm/src/type_of.rs b/compiler/rustc_codegen_llvm/src/type_of.rs
@@ -407,7 +407,7 @@ impl<'tcx> LayoutLlvmExt<'tcx> for TyAndLayout<'tcx> {
         // arrays but don't count as aggregate types
         if let FieldsShape::Array { count, .. } = self.layout.fields()
             && let element = self.field(cx, 0)
-            && element.ty.is_integral()
+            && element.ty.is_primitive()
         {
             // `cx.type_ix(bits)` is tempting here, but while that works great
             // for things that *stay* as memory-to-memory copies, it also ends
@@ -418,8 +418,36 @@ impl<'tcx> LayoutLlvmExt<'tcx> for TyAndLayout<'tcx> {
             return Some(cx.type_vector(ety, *count));
         }
 
-        // FIXME: The above only handled integer arrays; surely more things
-        // would also be possible. Be careful about provenance, though!
-        None
+        // Ensure the type isn't too complex nor otherwise ineligible
+        is_scalar_copy_reasonable(4, self.ty, cx)?;
+
+        // Otherwise we can load/store it via a long-enough integer type
+        Some(cx.type_ix(self.layout.size().bits()))
+    }
+}
+
+fn is_scalar_copy_reasonable<'a, 'tcx>(
+    max_fields: u32,
+    t: Ty<'tcx>,
+    cx: &CodegenCx<'a, 'tcx>,
+) -> Option<u32> {
+    if t.is_any_ptr() || t.is_primitive() {
+        return max_fields.checked_sub(1);
+    }
+
+    match t.kind() {
+        ty::Tuple(field_tys) => field_tys
+            .into_iter()
+            .try_fold(max_fields, |mf, tt| is_scalar_copy_reasonable(mf, tt, cx)),
+        // Unions are magic and can carry anything, regardless of their field
+        // types, so force them to always go through `memcpy`.
+        ty::Adt(adt_def, _) if adt_def.is_union() => None,
+        // If there could be multiple variants, just use `memcpy` for now.
+        ty::Adt(adt_def, _) if adt_def.variants().len() != 1 => None,
+        ty::Adt(adt_def, substs) => adt_def.all_fields().try_fold(max_fields, |mf, field_def| {
+            let field_ty = field_def.ty(cx.tcx, substs);
+            is_scalar_copy_reasonable(mf, field_ty, cx)
+        }),
+        _ => None,
     }
 }
diff --git a/tests/assembly/swap-strings.rs b/tests/assembly/swap-strings.rs
@@ -0,0 +1,23 @@
+// assembly-output: emit-asm
+// compile-flags: --crate-type=lib -O -C llvm-args=-x86-asm-syntax=intel
+// only-x86_64
+// ignore-sgx
+// ignore-debug
+
+// Ensure that the swap uses SIMD registers and does not go to stack.
+
+// CHECK-LABEL: swap_strings_xmm:
+#[no_mangle]
+pub fn swap_strings_xmm(a: &mut String, b: &mut String) {
+    // CHECK-DAG: movups  [[A1:xmm.+]], xmmword ptr [[AX:.+]]
+    // CHECK-DAG: mov     [[A2:r.+]], qword ptr [[AQ:.+]]
+    // CHECK-DAG: movups  [[B1:xmm.+]], xmmword ptr [[BX:.+]]
+    // CHECK-DAG: mov     [[B2:r.+]], qword ptr [[BQ:.+]]
+    // CHECK-NOT: mov
+    // CHECK-DAG: movups  xmmword ptr [[AX]], [[B1]]
+    // CHECK-DAG: mov     qword ptr [[AQ]], [[B2]]
+    // CHECK-DAG: movups  xmmword ptr [[BX]], [[A1]]
+    // CHECK-DAG: mov     qword ptr [[BQ]], [[A2]]
+    // CHECK: ret
+    std::mem::swap(a, b);
+}
diff --git a/tests/codegen/issues/issue-15953.rs b/tests/codegen/issues/issue-15953.rs
@@ -1,29 +1,32 @@
 // Test that llvm generates `memcpy` for moving a value
 // inside a function and moving an argument.
 
+#[derive(Default, Debug)]
+struct RatherLargeType(usize, isize, usize, isize, usize, isize);
+
 struct Foo {
-    x: Vec<i32>,
+    x: RatherLargeType,
 }
 
 #[inline(never)]
 #[no_mangle]
 // CHECK: memcpy
-fn interior(x: Vec<i32>) -> Vec<i32> {
+fn interior(x: RatherLargeType) -> RatherLargeType {
     let Foo { x } = Foo { x: x };
     x
 }
 
 #[inline(never)]
 #[no_mangle]
 // CHECK: memcpy
-fn exterior(x: Vec<i32>) -> Vec<i32> {
+fn exterior(x: RatherLargeType) -> RatherLargeType {
     x
 }
 
 fn main() {
-    let x = interior(Vec::new());
+    let x = interior(RatherLargeType::default());
     println!("{:?}", x);
 
-    let x = exterior(Vec::new());
+    let x = exterior(RatherLargeType::default());
     println!("{:?}", x);
 }
diff --git a/tests/codegen/issues/issue-86106.rs b/tests/codegen/issues/issue-86106.rs
@@ -2,14 +2,15 @@
 // compile-flags: -C opt-level=3 -Z merge-functions=disabled
 
 // The below two functions ensure that both `String::new()` and `"".to_string()`
-// produce the identical code.
+// generate their values directly, rather that creating a constant and copying
+// that constant (which takes more instructions because of PIC).
 
 #![crate_type = "lib"]
 
 // CHECK-LABEL: define {{(dso_local )?}}void @string_new
 #[no_mangle]
 pub fn string_new() -> String {
-    // CHECK: store ptr inttoptr
+    // CHECK: store {{i16|i32|i64}} 1, ptr %_0,
     // CHECK-NEXT: getelementptr
     // CHECK-NEXT: call void @llvm.memset
     // CHECK-NEXT: ret void
@@ -19,10 +20,8 @@ pub fn string_new() -> String {
 // CHECK-LABEL: define {{(dso_local )?}}void @empty_to_string
 #[no_mangle]
 pub fn empty_to_string() -> String {
-    // CHECK: store ptr inttoptr
-    // CHECK-NEXT: getelementptr
-    // CHECK-NEXT: call void @llvm.memset
-    // CHECK-NEXT: ret void
+    // CHECK: store {{i48|i96|i192}} 1, ptr %_0, align {{2|4|8}}
+    // CHECK-NEXT: ret
     "".to_string()
 }
 
@@ -32,7 +31,7 @@ pub fn empty_to_string() -> String {
 // CHECK-LABEL: @empty_vec
 #[no_mangle]
 pub fn empty_vec() -> Vec<u8> {
-    // CHECK: store ptr inttoptr
+    // CHECK: store ptr inttoptr ({{i16|i32|i64}} 1 to ptr), ptr %_0,
     // CHECK-NEXT: getelementptr
     // CHECK-NEXT: call void @llvm.memset
     // CHECK-NEXT: ret void
@@ -42,9 +41,19 @@ pub fn empty_vec() -> Vec<u8> {
 // CHECK-LABEL: @empty_vec_clone
 #[no_mangle]
 pub fn empty_vec_clone() -> Vec<u8> {
-    // CHECK: store ptr inttoptr
+    // CHECK: store {{i16|i32|i64}} 1, ptr %_0,
     // CHECK-NEXT: getelementptr
     // CHECK-NEXT: call void @llvm.memset
     // CHECK-NEXT: ret void
     vec![].clone()
 }
+
+// CHECK-LABEL: @empty_vec_from_array
+#[no_mangle]
+pub fn empty_vec_from_array() -> Vec<u8> {
+    // CHECK: store ptr inttoptr ({{i16|i32|i64}} 1 to ptr), ptr %_0,
+    // CHECK-NEXT: getelementptr
+    // CHECK-NEXT: call void @llvm.memset
+    // CHECK-NEXT: ret void
+    [].into()
+}
diff --git a/tests/codegen/loads.rs b/tests/codegen/loads.rs
@@ -136,12 +136,22 @@ pub fn small_array_alignment(x: [i8; 4]) -> [i8; 4] {
     x
 }
 
-// CHECK-LABEL: small_struct_alignment
+// CHECK-LABEL: i32 @small_struct_alignment(i32 %0)
 // The struct is loaded as i32, but its alignment is lower, go with 1 byte to avoid target
 // dependent alignment
 #[no_mangle]
 pub fn small_struct_alignment(x: Bytes) -> Bytes {
-    // CHECK: [[VAR:%[0-9]+]] = load i32, ptr %{{.*}}, align 1
+    // CHECK: [[RETP:%.+]] = alloca %Bytes, align 1
+    // CHECK: [[ALIGNED:%.+]] = alloca i32, align 4
+    // CHECK: %x = alloca %Bytes, align 1
+
+    // CHECK: store i32 %0, ptr [[ALIGNED]], align 4
+    // CHECK: call void @llvm.memcpy{{.+}}(ptr align 1 %x, ptr align 4 %1, i64 4, i1 false)
+
+    // CHECK: [[TEMP:%[0-9]+]] = load i32, ptr %x, align 1
+    // CHECK: store i32 [[TEMP]], ptr [[RETP]], align 1
+
+    // CHECK: [[VAR:%[0-9]+]] = load i32, ptr [[RETP]], align 1
     // CHECK: ret i32 [[VAR]]
     x
 }
diff --git a/tests/codegen/mem-replace-simple-type.rs b/tests/codegen/mem-replace-simple-type.rs
@@ -37,8 +37,19 @@ pub fn replace_ref_str<'a>(r: &mut &'a str, v: &'a str) -> &'a str {
 pub fn replace_short_array(r: &mut [u32; 3], v: [u32; 3]) -> [u32; 3] {
     // CHECK-NOT: alloca
     // CHECK: %[[R:.+]] = load <3 x i32>, ptr %r, align 4
-    // CHECK: store <3 x i32> %[[R]], ptr %result
+    // CHECK: store <3 x i32> %[[R]], ptr %result, align 4
     // CHECK: %[[V:.+]] = load <3 x i32>, ptr %v, align 4
-    // CHECK: store <3 x i32> %[[V]], ptr %r
+    // CHECK: store <3 x i32> %[[V]], ptr %r, align 4
+    std::mem::replace(r, v)
+}
+
+#[no_mangle]
+// CHECK-LABEL: @replace_string(
+pub fn replace_string(r: &mut String, v: String) -> String {
+    // CHECK-NOT: alloca
+    // CHECK: %[[R:.+]] = load i192, ptr %r, align 8
+    // CHECK: store i192 %[[R]], ptr %result, align 8
+    // CHECK: %[[V:.+]] = load i192, ptr %v, align 8
+    // CHECK: store i192 %[[V]], ptr %r, align 8
     std::mem::replace(r, v)
 }
diff --git a/tests/codegen/packed.rs b/tests/codegen/packed.rs
@@ -119,14 +119,22 @@ pub struct Packed2Pair(u8, u32);
 // CHECK-LABEL: @pkd1_pair
 #[no_mangle]
 pub fn pkd1_pair(pair1: &mut Packed1Pair, pair2: &mut Packed1Pair) {
-// CHECK: call void @llvm.memcpy.{{.*}}(ptr align 1 %{{.*}}, ptr align 1 %{{.*}}, i{{[0-9]+}} 5, i1 false)
+// CHECK: [[ALLOCA:%.+]] = alloca %Packed1Pair, align 1
+// CHECK: [[TEMP1:%.+]] = load i40, ptr %pair1, align 1
+// CHECK: store i40 [[TEMP1]], ptr [[ALLOCA]], align 1
+// CHECK: [[TEMP2:%.+]] = load i40, ptr [[ALLOCA]], align 1
+// CHECK: store i40 [[TEMP2]], ptr %pair2, align 1
     *pair2 = *pair1;
 }
 
 // CHECK-LABEL: @pkd2_pair
 #[no_mangle]
 pub fn pkd2_pair(pair1: &mut Packed2Pair, pair2: &mut Packed2Pair) {
-// CHECK: call void @llvm.memcpy.{{.*}}(ptr align 2 %{{.*}}, ptr align 2 %{{.*}}, i{{[0-9]+}} 6, i1 false)
+// CHECK: [[ALLOCA:%.+]] = alloca %Packed2Pair, align 2
+// CHECK: [[TEMP1:%.+]] = load i48, ptr %pair1, align 2
+// CHECK: store i48 [[TEMP1]], ptr [[ALLOCA]], align 2
+// CHECK: [[TEMP2:%.+]] = load i48, ptr [[ALLOCA]], align 2
+// CHECK: store i48 [[TEMP2]], ptr %pair2, align 2
     *pair2 = *pair1;
 }
 
@@ -141,13 +149,21 @@ pub struct Packed2NestedPair((u32, u32));
 // CHECK-LABEL: @pkd1_nested_pair
 #[no_mangle]
 pub fn pkd1_nested_pair(pair1: &mut Packed1NestedPair, pair2: &mut Packed1NestedPair) {
-// CHECK: call void @llvm.memcpy.{{.*}}(ptr align 1 %{{.*}}, ptr align 1 %{{.*}}, i{{[0-9]+}} 8, i1 false)
+// CHECK: [[ALLOCA:%.+]] = alloca %Packed1NestedPair, align 1
+// CHECK: [[TEMP1:%.+]] = load i64, ptr %pair1, align 1
+// CHECK: store i64 [[TEMP1]], ptr [[ALLOCA]], align 1
+// CHECK: [[TEMP2:%.+]] = load i64, ptr [[ALLOCA]], align 1
+// CHECK: store i64 [[TEMP2]], ptr %pair2, align 1
     *pair2 = *pair1;
 }
 
 // CHECK-LABEL: @pkd2_nested_pair
 #[no_mangle]
 pub fn pkd2_nested_pair(pair1: &mut Packed2NestedPair, pair2: &mut Packed2NestedPair) {
-// CHECK: call void @llvm.memcpy.{{.*}}(ptr align 2 %{{.*}}, ptr align 2 %{{.*}}, i{{[0-9]+}} 8, i1 false)
+// CHECK: [[ALLOCA:%.+]] = alloca %Packed2NestedPair, align 2
+// CHECK: [[TEMP1:%.+]] = load i64, ptr %pair1, align 2
+// CHECK: store i64 [[TEMP1]], ptr [[ALLOCA]], align 2
+// CHECK: [[TEMP2:%.+]] = load i64, ptr [[ALLOCA]], align 2
+// CHECK: store i64 [[TEMP2]], ptr %pair2, align 2
     *pair2 = *pair1;
 }
diff --git a/tests/codegen/simd-intrinsic/simd-intrinsic-transmute-array.rs b/tests/codegen/simd-intrinsic/simd-intrinsic-transmute-array.rs
@@ -35,7 +35,8 @@ pub fn vector_align() -> usize {
 // CHECK-LABEL: @build_array_s
 #[no_mangle]
 pub fn build_array_s(x: [f32; 4]) -> S<4> {
-    // CHECK: call void @llvm.memcpy.{{.+}}({{.*}} align [[VECTOR_ALIGN]] {{.*}} align [[ARRAY_ALIGN]] {{.*}}, [[USIZE]] 16, i1 false)
+    // CHECK: %[[VAL:.+]] = load <4 x float>, ptr %x, align [[ARRAY_ALIGN]]
+    // CHECK: store <4 x float> %[[VAL:.+]], ptr %_0, align [[VECTOR_ALIGN]]
     S::<4>(x)
 }
 
@@ -50,7 +51,8 @@ pub fn build_array_transmute_s(x: [f32; 4]) -> S<4> {
 // CHECK-LABEL: @build_array_t
 #[no_mangle]
 pub fn build_array_t(x: [f32; 4]) -> T {
-    // CHECK: call void @llvm.memcpy.{{.+}}({{.*}} align [[VECTOR_ALIGN]] {{.*}} align [[ARRAY_ALIGN]] {{.*}}, [[USIZE]] 16, i1 false)
+    // CHECK: %[[VAL:.+]] = load <4 x float>, ptr %x, align [[ARRAY_ALIGN]]
+    // CHECK: store <4 x float> %[[VAL:.+]], ptr %_0, align [[VECTOR_ALIGN]]
     T(x)
 }
 
diff --git a/tests/codegen/swap-small-types.rs b/tests/codegen/swap-small-types.rs
@@ -6,6 +6,21 @@
 
 use std::mem::swap;
 
+// CHECK-LABEL: @swap_fat_ptrs
+#[no_mangle]
+pub fn swap_fat_ptrs<'a>(x: &mut &'a str, y: &mut &'a str) {
+    // CHECK-NOT: alloca
+    // CHECK: %[[X0:.+]] = load ptr, ptr %x, align 8
+    // CHECK: %[[X1:.+]] = load i64, ptr %[[PX1:.+]], align 8
+    // CHECK: %[[Y0:.+]] = load ptr, ptr %y, align 8
+    // CHECK: %[[Y1:.+]] = load i64, ptr %[[PY1:.+]], align 8
+    // CHECK: store ptr %[[Y0]], ptr %x, align 8
+    // CHECK: store i64 %[[Y1]], ptr %[[PX1]], align 8
+    // CHECK: store ptr %[[X0]], ptr %y, align 8
+    // CHECK: store i64 %[[X1]], ptr %[[PY1]], align 8
+    swap(x, y)
+}
+
 type RGB48 = [u16; 3];
 
 // CHECK-LABEL: @swap_rgb48_manually(
@@ -40,9 +55,9 @@ type RGB24 = [u8; 3];
 // CHECK-LABEL: @swap_rgb24_slices
 #[no_mangle]
 pub fn swap_rgb24_slices(x: &mut [RGB24], y: &mut [RGB24]) {
-// CHECK-NOT: alloca
-// CHECK: load <{{[0-9]+}} x i8>
-// CHECK: store <{{[0-9]+}} x i8>
+    // CHECK-NOT: alloca
+    // CHECK: load <{{[0-9]+}} x i8>
+    // CHECK: store <{{[0-9]+}} x i8>
     if x.len() == y.len() {
         x.swap_with_slice(y);
     }
@@ -51,12 +66,23 @@ pub fn swap_rgb24_slices(x: &mut [RGB24], y: &mut [RGB24]) {
 // This one has a power-of-two size, so we iterate over it directly
 type RGBA32 = [u8; 4];
 
+// CHECK-LABEL: @swap_rgba32
+#[no_mangle]
+pub fn swap_rgba32(x: &mut RGBA32, y: &mut RGBA32) {
+    // CHECK-NOT: alloca
+    // CHECK: load <4 x i8>
+    // CHECK: load <4 x i8>
+    // CHECK: store <4 x i8>
+    // CHECK: store <4 x i8>
+    swap(x, y)
+}
+
 // CHECK-LABEL: @swap_rgba32_slices
 #[no_mangle]
 pub fn swap_rgba32_slices(x: &mut [RGBA32], y: &mut [RGBA32]) {
-// CHECK-NOT: alloca
-// CHECK: load <{{[0-9]+}} x i32>
-// CHECK: store <{{[0-9]+}} x i32>
+    // CHECK-NOT: alloca
+    // CHECK: load <{{[0-9]+}} x i32>
+    // CHECK: store <{{[0-9]+}} x i32>
     if x.len() == y.len() {
         x.swap_with_slice(y);
     }
@@ -69,10 +95,36 @@ const _: () = assert!(!std::mem::size_of::<String>().is_power_of_two());
 // CHECK-LABEL: @swap_string_slices
 #[no_mangle]
 pub fn swap_string_slices(x: &mut [String], y: &mut [String]) {
-// CHECK-NOT: alloca
-// CHECK: load <{{[0-9]+}} x i64>
-// CHECK: store <{{[0-9]+}} x i64>
+    // CHECK-NOT: alloca
+    // CHECK: load <{{[0-9]+}} x i64>
+    // CHECK: store <{{[0-9]+}} x i64>
     if x.len() == y.len() {
         x.swap_with_slice(y);
     }
 }
+
+// It's wasteful to do three `memcpy`s when a `String` is just three fields.
+
+// CHECK-LABEL: @swap_strings
+#[no_mangle]
+pub fn swap_strings(x: &mut String, y: &mut String) {
+    // CHECK-NOT: alloca
+    // CHECK: load i192
+    // CHECK: load i192
+    // CHECK: store i192
+    // CHECK: store i192
+    swap(x, y)
+}
+
+// CHECK-LABEL: @swap_tuple_with_padding
+#[no_mangle]
+pub fn swap_tuple_with_padding(x: &mut (u8, u32, u8), y: &mut (u8, u32, u8)) {
+    // CHECK-NOT: alloca
+    // CHECK: load i64
+    // CHECK-NOT: noundef
+    // CHECK: load i64
+    // CHECK-NOT: noundef
+    // CHECK: store i64
+    // CHECK: store i64
+    swap(x, y)
+}
diff --git a/tests/codegen/vec-in-place.rs b/tests/codegen/vec-in-place.rs

Original file line number	Diff line number	Diff line change
`@@ -1,29 +1,32 @@`
`1`	`1`	// Test that llvm generates `memcpy` for moving a value
`2`	`2`	`// inside a function and moving an argument.`
`3`	`3`
	`4`	`+#[derive(Default, Debug)]`
	`5`	`+struct RatherLargeType(usize, isize, usize, isize, usize, isize);`
	`6`	`+`
`4`	`7`	`struct Foo {`
`5`		`- x: Vec<i32>,`
	`8`	`+ x: RatherLargeType,`
`6`	`9`	`}`
`7`	`10`
`8`	`11`	`#[inline(never)]`
`9`	`12`	`#[no_mangle]`
`10`	`13`	`// CHECK: memcpy`
`11`		`-fn interior(x: Vec<i32>) -> Vec<i32> {`
	`14`	`+fn interior(x: RatherLargeType) -> RatherLargeType {`
`12`	`15`	`let Foo { x } = Foo { x: x };`
`13`	`16`	`x`
`14`	`17`	`}`
`15`	`18`
`16`	`19`	`#[inline(never)]`
`17`	`20`	`#[no_mangle]`
`18`	`21`	`// CHECK: memcpy`
`19`		`-fn exterior(x: Vec<i32>) -> Vec<i32> {`
	`22`	`+fn exterior(x: RatherLargeType) -> RatherLargeType {`
`20`	`23`	`x`
`21`	`24`	`}`
`22`	`25`
`23`	`26`	`fn main() {`
`24`		`- let x = interior(Vec::new());`
	`27`	`+ let x = interior(RatherLargeType::default());`
`25`	`28`	`println!("{:?}", x);`
`26`	`29`
`27`		`- let x = exterior(Vec::new());`
	`30`	`+ let x = exterior(RatherLargeType::default());`
`28`	`31`	`println!("{:?}", x);`
`29`	`32`	`}`
Original file line number	Diff line number	Diff line change
`@@ -136,12 +136,22 @@ pub fn small_array_alignment(x: [i8; 4]) -> [i8; 4] {`
`136`	`136`	`x`
`137`	`137`	`}`
`138`	`138`
`139`		`-// CHECK-LABEL: small_struct_alignment`
	`139`	`+// CHECK-LABEL: i32 @small_struct_alignment(i32 %0)`
`140`	`140`	`// The struct is loaded as i32, but its alignment is lower, go with 1 byte to avoid target`
`141`	`141`	`// dependent alignment`
`142`	`142`	`#[no_mangle]`
`143`	`143`	`pub fn small_struct_alignment(x: Bytes) -> Bytes {`
`144`		`- // CHECK: [[VAR:%[0-9]+]] = load i32, ptr %{{.*}}, align 1`
	`144`	`+ // CHECK: [[RETP:%.+]] = alloca %Bytes, align 1`
	`145`	`+ // CHECK: [[ALIGNED:%.+]] = alloca i32, align 4`
	`146`	`+ // CHECK: %x = alloca %Bytes, align 1`
	`147`	`+`
	`148`	`+ // CHECK: store i32 %0, ptr [[ALIGNED]], align 4`
	`149`	`+ // CHECK: call void @llvm.memcpy{{.+}}(ptr align 1 %x, ptr align 4 %1, i64 4, i1 false)`
	`150`	`+`
	`151`	`+ // CHECK: [[TEMP:%[0-9]+]] = load i32, ptr %x, align 1`
	`152`	`+ // CHECK: store i32 [[TEMP]], ptr [[RETP]], align 1`
	`153`	`+`
	`154`	`+ // CHECK: [[VAR:%[0-9]+]] = load i32, ptr [[RETP]], align 1`
`145`	`155`	`// CHECK: ret i32 [[VAR]]`
`146`	`156`	`x`
`147`	`157`	`}`