Tweak the threshold for chunked swapping

Thanks to 98892 for the tests I brought in here, as it demonstrated that 3×usize is currently suboptimal.

library/core/src/mem/mod.rs

@@ -736,7 +736,7 @@ pub const fn swap<T>(x: &mut T, y: &mut T) {
         // tends to copy the whole thing to stack rather than doing it one part
         // at a time, so instead treat them as one-element slices and piggy-back
         // the slice optimizations that will split up the swaps.
-        if size_of::<T>() / align_of::<T>() > 4 {
+        if const { size_of::<T>() / align_of::<T>() > 2 } {
             // SAFETY: exclusive references always point to one non-overlapping
             // element and are non-null and properly aligned.
             return unsafe { ptr::swap_nonoverlapping(x, y, 1) };

tests/codegen/swap-small-types.rs

@@ -48,16 +48,32 @@ pub fn swap_rgba64(x: &mut RGBA64, y: &mut RGBA64) {
     swap(x, y)
 }
 
+// CHECK-LABEL: @swap_vecs
+#[no_mangle]
+pub fn swap_vecs(x: &mut Vec<u32>, y: &mut Vec<u32>) {
+    // CHECK-NOT: alloca
+    // CHECK: ret void
+    swap(x, y)
+}
+
+// CHECK-LABEL: @swap_slices
+#[no_mangle]
+pub fn swap_slices<'a>(x: &mut &'a [u32], y: &mut &'a [u32]) {
+    // CHECK-NOT: alloca
+    // CHECK: ret void
+    swap(x, y)
+}
+
 // LLVM doesn't vectorize a loop over 3-byte elements,
 // so we chunk it down to bytes and loop over those instead.
 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);
     }
@@ -69,9 +85,9 @@ type RGBA32 = [u8; 4];
 // 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);
     }
@@ -84,10 +100,24 @@ 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);
     }
 }
+
+#[repr(C, packed)]
+pub struct Packed {
+    pub first: bool,
+    pub second: usize,
+}
+
+// CHECK-LABEL: @swap_packed_structs
+#[no_mangle]
+pub fn swap_packed_structs(x: &mut Packed, y: &mut Packed) {
+    // CHECK-NOT: alloca
+    // CHECK: ret void
+    swap(x, y)
+}