speed up String::push and String::insert

Author: lincot

library/alloc/src/lib.rs

@@ -105,6 +105,7 @@
 #![feature(async_fn_traits)]
 #![feature(async_iterator)]
 #![feature(box_uninit_write)]
+#![feature(char_internals)]
 #![feature(clone_to_uninit)]
 #![feature(coerce_unsized)]
 #![feature(const_align_of_val)]

library/alloc/src/string.rs

@@ -1417,9 +1417,14 @@ impl String {
     #[inline]
     #[stable(feature = "rust1", since = "1.0.0")]
     pub fn push(&mut self, ch: char) {
-        match ch.len_utf8() {
-            1 => self.vec.push(ch as u8),
-            _ => self.vec.extend_from_slice(ch.encode_utf8(&mut [0; 4]).as_bytes()),
+        let len = self.len();
+        let ch_len = ch.len_utf8();
+        self.reserve(ch_len);
+
+        // SAFETY: Just reserved capacity for at least the length needed to encode `ch`.
+        unsafe {
+            core::char::encode_utf8_raw_unchecked(ch as u32, self.as_mut_ptr().add(self.len()));
+            self.vec.set_len(len + ch_len);
         }
     }
 
@@ -1716,24 +1721,30 @@ impl String {
     #[rustc_confusables("set")]
     pub fn insert(&mut self, idx: usize, ch: char) {
         assert!(self.is_char_boundary(idx));
-        let mut bits = [0; 4];
-        let bits = ch.encode_utf8(&mut bits).as_bytes();
 
+        let len = self.len();
+        let ch_len = ch.len_utf8();
+        self.reserve(ch_len);
+
+        // SAFETY: Shift data `ch_len` bytes to the right,
+        // capacity was just reserved for at least that many bytes.
         unsafe {
-            self.insert_bytes(idx, bits);
+            ptr::copy(
+                self.vec.as_ptr().add(idx),
+                self.vec.as_mut_ptr().add(idx + ch_len),
+                len - idx,
+            );
         }
-    }
 
-    #[cfg(not(no_global_oom_handling))]
-    unsafe fn insert_bytes(&mut self, idx: usize, bytes: &[u8]) {
-        let len = self.len();
-        let amt = bytes.len();
-        self.vec.reserve(amt);
+        // SAFETY: Encode the character into the space left after the shift
+        // if `idx != len`, or into the uninitialized spare capacity otherwise.
+        unsafe {
+            core::char::encode_utf8_raw_unchecked(ch as u32, self.vec.as_mut_ptr().add(idx));
+        }
 
+        // SAFETY: `ch_len` initialized bytes have been added.
         unsafe {
-            ptr::copy(self.vec.as_ptr().add(idx), self.vec.as_mut_ptr().add(idx + amt), len - idx);
-            ptr::copy_nonoverlapping(bytes.as_ptr(), self.vec.as_mut_ptr().add(idx), amt);
-            self.vec.set_len(len + amt);
+            self.vec.set_len(len + ch_len);
         }
     }
 
@@ -1763,8 +1774,14 @@ impl String {
     pub fn insert_str(&mut self, idx: usize, string: &str) {
         assert!(self.is_char_boundary(idx));
 
+        let len = self.len();
+        let amt = string.len();
+        self.reserve(amt);
+
         unsafe {
-            self.insert_bytes(idx, string.as_bytes());
+            ptr::copy(self.vec.as_ptr().add(idx), self.vec.as_mut_ptr().add(idx + amt), len - idx);
+            ptr::copy_nonoverlapping(string.as_ptr(), self.vec.as_mut_ptr().add(idx), amt);
+            self.vec.set_len(len + amt);
         }
     }
 

library/core/src/char/methods.rs

@@ -1789,32 +1789,66 @@ pub const fn encode_utf8_raw(code: u32, dst: &mut [u8]) -> &mut [u8] {
         );
     }
     let len = len_utf8(code);
-    match (len, &mut *dst) {
-        (1, [a, ..]) => {
-            *a = code as u8;
-        }
-        (2, [a, b, ..]) => {
-            *a = (code >> 6 & 0x1F) as u8 | TAG_TWO_B;
-            *b = (code & 0x3F) as u8 | TAG_CONT;
-        }
-        (3, [a, b, c, ..]) => {
-            *a = (code >> 12 & 0x0F) as u8 | TAG_THREE_B;
-            *b = (code >> 6 & 0x3F) as u8 | TAG_CONT;
-            *c = (code & 0x3F) as u8 | TAG_CONT;
-        }
-        (4, [a, b, c, d, ..]) => {
-            *a = (code >> 18 & 0x07) as u8 | TAG_FOUR_B;
-            *b = (code >> 12 & 0x3F) as u8 | TAG_CONT;
-            *c = (code >> 6 & 0x3F) as u8 | TAG_CONT;
-            *d = (code & 0x3F) as u8 | TAG_CONT;
-        }
+    if dst.len() < len {
         // FIXME(const-hack): We would prefer to have streamlined panics when formatters become const-friendly.
-        _ => const_eval_select((code, len, dst.len()), panic_at_const, panic_at_rt),
-    };
+        const_eval_select((code, len, dst.len()), panic_at_const, panic_at_rt);
+    }
+
+    // SAFETY: `dst` is checked to be at least the length needed to encode the codepoint.
+    unsafe { encode_utf8_raw_unchecked(code, dst.as_mut_ptr()) };
+
     // SAFETY: `<&mut [u8]>::as_mut_ptr` is guaranteed to return a valid pointer and `len` has been tested to be within bounds.
     unsafe { slice::from_raw_parts_mut(dst.as_mut_ptr(), len) }
 }
 
+/// Encodes a raw u32 value as UTF-8 to the provided destination buffer.
+///
+/// Unlike `char::encode_utf8`, this method also handles codepoints in the surrogate range.
+/// (Creating a `char` in the surrogate range is UB.)
+/// The result is valid [generalized UTF-8] but not valid UTF-8.
+///
+/// [generalized UTF-8]: https://simonsapin.github.io/wtf-8/#generalized-utf8
+///
+/// # Safety
+///
+/// The behavior is undefined if the buffer pointed to by `dst` is not
+/// large enough to hold the encoded codepoint. A buffer of length four
+/// is large enough to encode any `char`.
+///
+/// For a safe version of this function, see the [`encode_utf8_raw`] function.
+#[unstable(feature = "char_internals", reason = "exposed only for libstd", issue = "none")]
+#[doc(hidden)]
+#[inline]
+pub const unsafe fn encode_utf8_raw_unchecked(code: u32, dst: *mut u8) {
+    let len = len_utf8(code);
+    // SAFETY: The caller must guarantee that the buffer pointed to by `dst`
+    // is at least `len` bytes long.
+    unsafe {
+        match len {
+            1 => {
+                *dst = code as u8;
+            }
+            2 => {
+                *dst = (code >> 6 & 0x1F) as u8 | TAG_TWO_B;
+                *dst.add(1) = (code & 0x3F) as u8 | TAG_CONT;
+            }
+            3 => {
+                *dst = (code >> 12 & 0x0F) as u8 | TAG_THREE_B;
+                *dst.add(1) = (code >> 6 & 0x3F) as u8 | TAG_CONT;
+                *dst.add(2) = (code & 0x3F) as u8 | TAG_CONT;
+            }
+            4 => {
+                *dst = (code >> 18 & 0x07) as u8 | TAG_FOUR_B;
+                *dst.add(1) = (code >> 12 & 0x3F) as u8 | TAG_CONT;
+                *dst.add(2) = (code >> 6 & 0x3F) as u8 | TAG_CONT;
+                *dst.add(3) = (code & 0x3F) as u8 | TAG_CONT;
+            }
+            // SAFETY: `char` always takes between 1 and 4 bytes to encode in UTF-8.
+            _ => crate::hint::unreachable_unchecked(),
+        }
+    }
+}
+
 /// Encodes a raw `u32` value as UTF-16 into the provided `u16` buffer,
 /// and then returns the subslice of the buffer that contains the encoded character.
 ///

library/core/src/char/mod.rs

@@ -38,7 +38,7 @@ pub use self::decode::{DecodeUtf16, DecodeUtf16Error};
 #[unstable(feature = "char_internals", reason = "exposed only for libstd", issue = "none")]
 pub use self::methods::encode_utf16_raw; // perma-unstable
 #[unstable(feature = "char_internals", reason = "exposed only for libstd", issue = "none")]
-pub use self::methods::encode_utf8_raw; // perma-unstable
+pub use self::methods::{encode_utf8_raw, encode_utf8_raw_unchecked}; // perma-unstable
 
 #[rustfmt::skip]
 use crate::ascii;