1
0
Fork 0
mirror of https://github.com/azalea-rs/simdnbt.git synced 2025-08-02 07:26:04 +00:00

move fastvec into its own module

This commit is contained in:
mat 2025-03-15 08:32:49 +00:00
parent a44d709489
commit ffe5b2575d
4 changed files with 181 additions and 111 deletions

View file

@ -157,14 +157,16 @@ fn read_with_stack<'a>(
Ok(())
}
#[derive(Default)]
pub(crate) struct Tapes<'a> {
main: MainTape,
extra: ExtraTapes<'a>,
}
impl Tapes<'_> {
fn new() -> Self {
Self::default()
Self {
main: MainTape::with_capacity(1024),
extra: ExtraTapes::default(),
}
}
}
impl Debug for Tapes<'_> {

View file

@ -1,121 +1,19 @@
use std::{
alloc::{self, Allocator, Layout},
fmt::{self, Debug},
mem,
ptr::NonNull,
};
use crate::common::{
BYTE_ARRAY_ID, BYTE_ID, COMPOUND_ID, DOUBLE_ID, FLOAT_ID, INT_ARRAY_ID, INT_ID, LONG_ARRAY_ID,
LONG_ID, SHORT_ID, STRING_ID,
use crate::{
common::{
BYTE_ARRAY_ID, BYTE_ID, COMPOUND_ID, DOUBLE_ID, FLOAT_ID, INT_ARRAY_ID, INT_ID,
LONG_ARRAY_ID, LONG_ID, SHORT_ID, STRING_ID,
},
fastvec::FastVec,
};
const DEFAULT_CAPACITY: usize = 1024;
// this is faster than a Vec mainly because we store a `cur` pointer which
// allows us to avoid having to add the start+length when pushing
#[derive(Debug)]
pub struct MainTape<A: Allocator = alloc::Global> {
/// The next address we'll be writing to.
cur: NonNull<TapeElement>,
/// The last (probably uninitialized) element in the tape.
end: NonNull<TapeElement>,
/// The start of the tape.
ptr: NonNull<u8>,
alloc: A,
}
impl<A: Allocator> MainTape<A> {
#[inline(always)]
pub fn push(&mut self, element: TapeElement) {
if self.cur == self.end {
self.grow()
}
unsafe { self.push_unchecked(element) };
}
#[inline]
pub unsafe fn push_unchecked(&mut self, element: TapeElement) {
unsafe { self.cur.write(element) };
self.cur = unsafe { self.cur.add(1) };
}
fn grow(&mut self) {
let old_cap = self.capacity();
let extending_by = old_cap;
let new_cap = old_cap + extending_by;
let element_size = mem::size_of::<TapeElement>();
let old_cap_bytes = old_cap * element_size;
let new_cap_bytes = new_cap * element_size;
let new_ptr = unsafe {
self.alloc.grow(
self.ptr,
Layout::from_size_align_unchecked(old_cap_bytes, element_size),
Layout::from_size_align_unchecked(new_cap_bytes, element_size),
)
};
let new_ptr = new_ptr.expect("allocation failed");
self.ptr = new_ptr.cast();
// update cur in case the ptr changed
self.cur = NonNull::new(self.ptr.as_ptr() as *mut TapeElement).unwrap();
self.cur = unsafe { self.cur.add(old_cap) };
// and end has to be updated anyways since we're updating the capacity
self.end = unsafe { self.cur.add(extending_by) };
}
#[inline]
pub fn len(&self) -> usize {
unsafe { self.cur.offset_from(self.ptr.cast()) as usize }
}
#[inline]
fn capacity(&self) -> usize {
unsafe { self.end.offset_from(self.ptr.cast()) as usize }
}
#[inline]
pub unsafe fn get_unchecked_mut(&mut self, index: usize) -> &mut TapeElement {
debug_assert!(index < self.len());
self.ptr.cast::<TapeElement>().add(index).as_mut()
}
#[inline]
pub fn as_ptr(&self) -> *const TapeElement {
self.ptr.cast().as_ptr()
}
}
impl Default for MainTape {
fn default() -> Self {
let element_size = mem::size_of::<TapeElement>();
let ptr = unsafe {
// this is faster than Layout::array
alloc::alloc(Layout::from_size_align_unchecked(
DEFAULT_CAPACITY * element_size,
element_size,
))
};
let ptr = NonNull::new(ptr as *mut TapeElement).expect("allocation failed");
let end = unsafe { ptr.add(DEFAULT_CAPACITY) };
Self {
cur: ptr,
end,
ptr: ptr.cast(),
alloc: alloc::Global,
}
}
}
impl<A: Allocator> Drop for MainTape<A> {
fn drop(&mut self) {
unsafe {
self.alloc.deallocate(
self.ptr.cast(),
Layout::array::<TapeElement>(self.capacity()).unwrap(),
)
};
}
}
pub type MainTape = FastVec<TapeElement>;
#[derive(Clone, Copy)]
pub struct TapeElement(u64);

169
simdnbt/src/fastvec.rs Normal file
View file

@ -0,0 +1,169 @@
use std::{
alloc::{self, Allocator, Layout},
mem::{self},
ptr::NonNull,
};
/// A slightly faster (but more limited) append-only Vec.
#[derive(Debug)]
pub struct FastVec<T, A: Allocator = alloc::Global> {
/// The next address we'll be writing to.
cur: NonNull<T>,
/// The last (probably uninitialized) element.
end: NonNull<T>,
/// The start of the vec.
ptr: NonNull<u8>,
alloc: A,
}
impl<T> FastVec<T, alloc::Global> {
/// Create a new FastVec with the given capacity.
///
/// Note that a capacity of 0 is not supported (as it would require an extra
/// instruction that's usually unnecessary), and will be treated as 1.
pub fn with_capacity(capacity: usize) -> Self {
let capacity = capacity.min(1);
let element_size = mem::size_of::<T>();
let ptr = unsafe {
// this is faster than Layout::array
alloc::alloc(Layout::from_size_align_unchecked(
capacity * element_size,
element_size,
))
};
let ptr = NonNull::new(ptr as *mut T).expect("allocation failed");
let end = unsafe { ptr.add(DEFAULT_CAPACITY) };
Self {
cur: ptr,
end,
ptr: ptr.cast(),
alloc: alloc::Global,
}
}
}
impl<T, A: Allocator> FastVec<T, A> {
#[inline(always)]
pub fn push(&mut self, element: T) {
if self.cur == self.end {
self.grow()
}
unsafe { self.push_unchecked(element) };
}
#[inline]
pub unsafe fn push_unchecked(&mut self, element: T) {
unsafe { self.cur.write(element) };
self.cur = unsafe { self.cur.add(1) };
}
#[inline]
fn grow(&mut self) {
let old_cap = self.capacity();
let extending_by = old_cap;
let new_cap = old_cap + extending_by;
let element_size = mem::size_of::<T>();
let old_cap_bytes = old_cap * element_size;
let new_cap_bytes = new_cap * element_size;
let new_ptr = unsafe {
self.alloc.grow(
self.ptr,
Layout::from_size_align_unchecked(old_cap_bytes, element_size),
Layout::from_size_align_unchecked(new_cap_bytes, element_size),
)
};
let new_ptr = new_ptr.expect("allocation failed");
self.ptr = new_ptr.cast();
// update cur in case the ptr changed
self.cur = NonNull::new(self.ptr.as_ptr() as *mut T).unwrap();
self.cur = unsafe { self.cur.add(old_cap) };
// and end has to be updated anyways since we're updating the capacity
self.end = unsafe { self.cur.add(extending_by) };
}
#[inline]
pub fn len(&self) -> usize {
unsafe { self.cur.offset_from(self.ptr.cast()) as usize }
}
#[inline]
fn capacity(&self) -> usize {
unsafe { self.end.offset_from(self.ptr.cast()) as usize }
}
#[inline]
pub unsafe fn get_unchecked_mut(&mut self, index: usize) -> &mut T {
debug_assert!(index < self.len());
self.ptr.cast::<T>().add(index).as_mut()
}
#[inline]
pub fn as_ptr(&self) -> *const T {
self.ptr.cast().as_ptr()
}
pub fn to_vec(self) -> Vec<T> {
let vec = unsafe {
Vec::from_raw_parts(
self.ptr.cast().as_ptr() as *mut T,
self.len(),
self.capacity(),
)
};
// the allocation was moved so don't drop it
mem::forget(self);
vec
}
}
const DEFAULT_CAPACITY: usize = 1024;
impl<T> Default for FastVec<T> {
fn default() -> Self {
let element_size = mem::size_of::<T>();
let ptr = unsafe {
// this is faster than Layout::array
alloc::alloc(Layout::from_size_align_unchecked(
DEFAULT_CAPACITY * element_size,
element_size,
))
};
let ptr = NonNull::new(ptr as *mut T).expect("allocation failed");
let end = unsafe { ptr.add(DEFAULT_CAPACITY) };
Self {
cur: ptr,
end,
ptr: ptr.cast(),
alloc: alloc::Global,
}
}
}
impl<T, A: Allocator> Drop for FastVec<T, A> {
fn drop(&mut self) {
unsafe {
self.alloc.deallocate(
self.ptr.cast(),
Layout::array::<T>(self.capacity()).unwrap(),
)
};
}
}
impl<T> From<Vec<T>> for FastVec<T> {
fn from(vec: Vec<T>) -> Self {
let mut fastvec = FastVec::with_capacity(vec.len());
for element in vec {
fastvec.push(element);
}
fastvec
}
}
impl<T> From<FastVec<T>> for Vec<T> {
fn from(fastvec: FastVec<T>) -> Self {
fastvec.to_vec()
}
}

View file

@ -11,6 +11,7 @@ compile_error!("simdnbt only supports 64-bit platforms");
pub mod borrow;
mod common;
mod error;
mod fastvec;
mod mutf8;
pub mod owned;
pub mod raw_list;