Use `load`-`store` instead of `memcpy` for short integer arrays

4 files changed, 65 insertions, 15 deletions

diff --git a/tests/codegen/array-codegen.rs b/tests/codegen/array-codegen.rs
new file mode 100644
index 00000000000..98488eb92ee
--- /dev/null
+++ b/tests/codegen/array-codegen.rs
@@ -0,0 +1,35 @@
+// compile-flags: -O -C no-prepopulate-passes
+// min-llvm-version: 15.0 (for opaque pointers)
+
+#![crate_type = "lib"]
+
+// CHECK-LABEL: @array_load
+#[no_mangle]
+pub fn array_load(a: &[u8; 4]) -> [u8; 4] {
+    // CHECK: %0 = alloca [4 x i8], align 1
+    // CHECK: %[[TEMP1:.+]] = load <4 x i8>, ptr %a, align 1
+    // CHECK: store <4 x i8> %[[TEMP1]], ptr %0, align 1
+    // CHECK: %[[TEMP2:.+]] = load i32, ptr %0, align 1
+    // CHECK: ret i32 %[[TEMP2]]
+    *a
+}
+
+// CHECK-LABEL: @array_store
+#[no_mangle]
+pub fn array_store(a: [u8; 4], p: &mut [u8; 4]) {
+    // CHECK: %a = alloca [4 x i8]
+    // CHECK: %[[TEMP:.+]] = load <4 x i8>, ptr %a, align 1
+    // CHECK-NEXT: store <4 x i8> %[[TEMP]], ptr %p, align 1
+    *p = a;
+}
+
+// CHECK-LABEL: @array_copy
+#[no_mangle]
+pub fn array_copy(a: &[u8; 4], p: &mut [u8; 4]) {
+    // CHECK: %[[LOCAL:.+]] = alloca [4 x i8], align 1
+    // CHECK: %[[TEMP1:.+]] = load <4 x i8>, ptr %a, align 1
+    // CHECK: store <4 x i8> %[[TEMP1]], ptr %[[LOCAL]], align 1
+    // CHECK: %[[TEMP2:.+]] = load <4 x i8>, ptr %[[LOCAL]], align 1
+    // CHECK: store <4 x i8> %[[TEMP2]], ptr %p, align 1
+    *p = *a;
+}
diff --git a/tests/codegen/mem-replace-simple-type.rs b/tests/codegen/mem-replace-simple-type.rs
index 4253ef13666..6151177de15 100644
--- a/tests/codegen/mem-replace-simple-type.rs
+++ b/tests/codegen/mem-replace-simple-type.rs
@@ -32,3 +32,14 @@ pub fn replace_ref_str<'a>(r: &mut &'a str, v: &'a str) -> &'a str {
     // CHECK: ret { ptr, i64 } %[[P2]]
     std::mem::replace(r, v)
 }
+
+#[no_mangle]
+// CHECK-LABEL: @replace_short_array(
+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 %0
+    // CHECK: %[[V:.+]] = load <3 x i32>, ptr %v, align 4
+    // CHECK: store <3 x i32> %[[V]], ptr %r
+    std::mem::replace(r, v)
+}
diff --git a/tests/codegen/swap-simd-types.rs b/tests/codegen/swap-simd-types.rs
index c90b277eb44..3472a42b0e6 100644
--- a/tests/codegen/swap-simd-types.rs
+++ b/tests/codegen/swap-simd-types.rs
@@ -30,3 +30,12 @@ pub fn swap_m256_slice(x: &mut [__m256], y: &mut [__m256]) {
         x.swap_with_slice(y);
     }
 }
+
+// CHECK-LABEL: @swap_bytes32
+#[no_mangle]
+pub fn swap_bytes32(x: &mut [u8; 32], y: &mut [u8; 32]) {
+// CHECK-NOT: alloca
+// CHECK: load <32 x i8>{{.+}}align 1
+// CHECK: store <32 x i8>{{.+}}align 1
+    swap(x, y)
+}
diff --git a/tests/codegen/swap-small-types.rs b/tests/codegen/swap-small-types.rs
index 6289d7af3a0..419645a3fc6 100644
--- a/tests/codegen/swap-small-types.rs
+++ b/tests/codegen/swap-small-types.rs
@@ -1,4 +1,4 @@
-// compile-flags: -O
+// compile-flags: -O -Z merge-functions=disabled
 // only-x86_64
 // ignore-debug: the debug assertions get in the way
 
@@ -12,13 +12,10 @@ type RGB48 = [u16; 3];
 #[no_mangle]
 pub fn swap_rgb48_manually(x: &mut RGB48, y: &mut RGB48) {
     // CHECK-NOT: alloca
-    // CHECK: %temp = alloca [3 x i16]
-    // CHECK-NOT: alloca
-    // CHECK-NOT: call void @llvm.memcpy
-    // CHECK: call void @llvm.memcpy.{{.+}}({{.+}} %temp, {{.+}} %x, {{.+}} 6, {{.+}})
-    // CHECK: call void @llvm.memcpy.{{.+}}({{.+}} %x, {{.+}} %y, {{.+}} 6, {{.+}})
-    // CHECK: call void @llvm.memcpy.{{.+}}({{.+}} %y, {{.+}} %temp, {{.+}} 6, {{.+}})
-    // CHECK-NOT: call void @llvm.memcpy
+    // CHECK: %[[TEMP0:.+]] = load <3 x i16>, ptr %x, align 2
+    // CHECK: %[[TEMP1:.+]] = load <3 x i16>, ptr %y, align 2
+    // CHECK: store <3 x i16> %[[TEMP1]], ptr %x, align 2
+    // CHECK: store <3 x i16> %[[TEMP0]], ptr %y, align 2
 
     let temp = *x;
     *x = *y;
@@ -28,13 +25,11 @@ pub fn swap_rgb48_manually(x: &mut RGB48, y: &mut RGB48) {
 // CHECK-LABEL: @swap_rgb48
 #[no_mangle]
 pub fn swap_rgb48(x: &mut RGB48, y: &mut RGB48) {
-    // FIXME MIR inlining messes up LLVM optimizations.
-    // If these checks start failing, please update this test.
-    // CHECK: alloca [3 x i16]
-    // CHECK: call void @llvm.memcpy
-// WOULD-CHECK-NOT: alloca
-// WOULD-CHECK: load i48
-// WOULD-CHECK: store i48
+    // CHECK-NOT: alloca
+    // CHECK: load <3 x i16>
+    // CHECK: load <3 x i16>
+    // CHECK: store <3 x i16>
+    // CHECK: store <3 x i16>
     swap(x, y)
 }