Auto merge of #120050 - scottmcm:vec-resize-memset, r=<try>

`Vec::resize` for bytes should be a single `memset`

Really I just started by trying to see if specializing `iter::repeat_n` would help the perf issue that kept me from removing `Vec::extend_with` last time I tried, but I noticed in the process that a resize for bytes doesn't set all the new space with a single `memset`: <https://play.rust-lang.org/?version=nightly&mode=release&edition=2021&gist=35175ec844b46fcd95e2d0aad526859e>

So using `repeat_n` to implement it -- like `VecDeque` uses, with the specialization for `next` to avoid a branch -- means that the optimizer notices the resize can set all the values with a single memset.

Author: bors

Diff for: library/alloc/src/vec/mod.rs

@@ -145,6 +145,12 @@ use self::spec_extend::SpecExtend;
 #[cfg(not(no_global_oom_handling))]
 mod spec_extend;
 
+#[cfg(not(no_global_oom_handling))]
+use self::spec_extend_elem::SpecExtendElem;
+
+#[cfg(not(no_global_oom_handling))]
+mod spec_extend_elem;
+
 /// A contiguous growable array type, written as `Vec<T>`, short for 'vector'.
 ///
 /// # Examples
@@ -2871,7 +2877,7 @@ impl<T: Clone, A: Allocator> Vec<T, A> {
         let len = self.len();
 
         if new_len > len {
-            self.extend_with(new_len - len, value)
+            self.extend_elem(new_len - len, value)
         } else {
             self.truncate(new_len);
         }
@@ -2985,32 +2991,38 @@ impl<T, A: Allocator, const N: usize> Vec<[T; N], A> {
 
 impl<T: Clone, A: Allocator> Vec<T, A> {
     #[cfg(not(no_global_oom_handling))]
-    /// Extend the vector by `n` clones of value.
-    fn extend_with(&mut self, n: usize, value: T) {
+    #[inline]
+    /// Extend the vector by `n` clones of `value`.
+    ///
+    /// Uses specialization to dispatch to `extend_elem_copy` if possible.
+    fn extend_elem(&mut self, n: usize, value: T) {
+        self.spec_extend_elem(n, value);
+    }
+
+    #[cfg(not(no_global_oom_handling))]
+    #[inline]
+    /// Extend the vector by `n` *copies* of `value`.
+    fn extend_elem_copy(&mut self, n: usize, value: T)
+    where
+        T: Copy,
+    {
         self.reserve(n);
 
+        // SAFETY: We now have space for all the elements, so the pointer math
+        // is all in-bounds. Because `T: Copy`, there's no user code being run
+        // here (no `clone` nor any `drop` in the assignment).
         unsafe {
-            let mut ptr = self.as_mut_ptr().add(self.len());
-            // Use SetLenOnDrop to work around bug where compiler
-            // might not realize the store through `ptr` through self.set_len()
-            // don't alias.
-            let mut local_len = SetLenOnDrop::new(&mut self.len);
-
-            // Write all elements except the last one
-            for _ in 1..n {
-                ptr::write(ptr, value.clone());
-                ptr = ptr.add(1);
-                // Increment the length in every step in case clone() panics
-                local_len.increment_len(1);
-            }
-
-            if n > 0 {
-                // We can write the last element directly without cloning needlessly
-                ptr::write(ptr, value);
-                local_len.increment_len(1);
+            // Because there's no user code being run here, we can skip it for ZSTs.
+            // That helps tests in debug mode that do things like `vec![(); HUGE]`.
+            // See <https://github.com/rust-lang/rust/pull/118094>
+            if !T::IS_ZST {
+                let ptr = self.as_mut_ptr().add(self.len);
+                for i in 0..n {
+                    *ptr.add(i) = value;
+                }
             }
 
-            // len set by scope guard
+            self.len += n;
         }
     }
 }

Diff for: library/alloc/src/vec/spec_extend_elem.rs

@@ -0,0 +1,27 @@
+use core::iter::repeat_n;
+
+use super::Vec;
+use crate::alloc::Allocator;
+
+// Specialization trait used for Vec::extend_elem
+pub(super) trait SpecExtendElem<T> {
+    fn spec_extend_elem(&mut self, n: usize, value: T);
+}
+
+impl<T, A: Allocator> SpecExtendElem<T> for Vec<T, A>
+where
+    T: Clone,
+{
+    default fn spec_extend_elem(&mut self, n: usize, value: T) {
+        self.extend_trusted(repeat_n(value, n))
+    }
+}
+
+impl<T, A: Allocator> SpecExtendElem<T> for Vec<T, A>
+where
+    T: Copy,
+{
+    fn spec_extend_elem(&mut self, n: usize, value: T) {
+        self.extend_elem_copy(n, value)
+    }
+}

Diff for: library/alloc/src/vec/spec_from_elem.rs

@@ -12,7 +12,7 @@ pub(super) trait SpecFromElem: Sized {
 impl<T: Clone> SpecFromElem for T {
     default fn from_elem<A: Allocator>(elem: Self, n: usize, alloc: A) -> Vec<Self, A> {
         let mut v = Vec::with_capacity_in(n, alloc);
-        v.extend_with(n, elem);
+        v.extend_elem(n, elem);
         v
     }
 }
@@ -24,7 +24,7 @@ impl<T: Clone + IsZero> SpecFromElem for T {
             return Vec { buf: RawVec::with_capacity_zeroed_in(n, alloc), len: n };
         }
         let mut v = Vec::with_capacity_in(n, alloc);
-        v.extend_with(n, elem);
+        v.extend_elem(n, elem);
         v
     }
 }
@@ -58,18 +58,3 @@ impl SpecFromElem for u8 {
         v
     }
 }
-
-// A better way would be to implement this for all ZSTs which are `Copy` and have trivial `Clone`
-// but the latter cannot be detected currently
-impl SpecFromElem for () {
-    #[inline]
-    fn from_elem<A: Allocator>(_elem: (), n: usize, alloc: A) -> Vec<(), A> {
-        let mut v = Vec::with_capacity_in(n, alloc);
-        // SAFETY: the capacity has just been set to `n`
-        // and `()` is a ZST with trivial `Clone` implementation
-        unsafe {
-            v.set_len(n);
-        }
-        v
-    }
-}

Diff for: tests/codegen/vec-of-bytes-memset.rs

@@ -0,0 +1,23 @@
+//@ compile-flags: -O
+//@ only-64bit
+
+#![crate_type = "lib"]
+
+// CHECK-LABEL: @resize_bytes_is_one_memset
+#[no_mangle]
+pub fn resize_bytes_is_one_memset(x: &mut Vec<u8>) {
+    // CHECK: call void @llvm.memset.p0.i64({{.+}}, i8 123, i64 456789, i1 false)
+    let new_len = x.len() + 456789;
+    x.resize(new_len, 123);
+}
+
+#[derive(Copy, Clone)]
+struct ByteNewtype(i8);
+
+// CHECK-LABEL: @from_elem_is_one_memset
+#[no_mangle]
+pub fn from_elem_is_one_memset() -> Vec<ByteNewtype> {
+    // CHECK: %[[P:.+]] = tail call{{.+}}@__rust_alloc(i64 noundef 123456, i64 noundef 1)
+    // CHECK: call void @llvm.memset.p0.i64({{.+}} %[[P]], i8 42, i64 123456, i1 false)
+    vec![ByteNewtype(42); 123456]
+}