-
Notifications
You must be signed in to change notification settings - Fork 183
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Remove the zero index from VarZeroVec #5601
Changes from 1 commit
e3e5dd1
9e85a34
6885722
047a49a
f2ce727
f8c4d38
585cd2a
1cc0814
27f2b23
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
- Loading branch information
There are no files selected for viewing
Original file line number | Diff line number | Diff line change |
---|---|---|
|
@@ -178,6 +178,7 @@ pub struct VarZeroVecComponents<'a, T: ?Sized, F> { | |
/// The number of elements | ||
len: u32, | ||
/// The list of indices into the `things` slice | ||
/// Since the first element is always at things[0], the first element of the indices array is for the *second* element | ||
indices: &'a [u8], | ||
/// The contiguous list of `T::VarULE`s | ||
things: &'a [u8], | ||
|
@@ -215,10 +216,12 @@ impl<'a, T: VarULE + ?Sized, F: VarZeroVecFormat> VarZeroVecComponents<'a, T, F> | |
/// Construct a new VarZeroVecComponents, checking invariants about the overall buffer size: | ||
/// | ||
/// - There must be either zero or at least four bytes (if four, this is the "length" parsed as a usize) | ||
/// - There must be at least `4*length + 4` bytes total, to form the array `indices` of indices | ||
/// - `indices[i]..indices[i+1]` must index into a valid section of | ||
/// - There must be at least `4*(length - 1) + 4` bytes total, to form the array `indices` of indices | ||
/// - `0..indices[0]` must index into a valid section of | ||
/// `things` (the data after `indices`), such that it parses to a `T::VarULE` | ||
/// - `indices[len - 1]..things.len()` must index into a valid section of | ||
/// - `indices[i - 1]..indices[i]` must index into a valid section of | ||
/// `things` (the data after `indices`), such that it parses to a `T::VarULE` | ||
/// - `indices[len - 2]..things.len()` must index into a valid section of | ||
/// `things`, such that it parses to a `T::VarULE` | ||
#[inline] | ||
pub fn parse_byte_slice(slice: &'a [u8]) -> Result<Self, VarZeroVecFormatError> { | ||
|
@@ -262,20 +265,23 @@ impl<'a, T: VarULE + ?Sized, F: VarZeroVecFormat> VarZeroVecComponents<'a, T, F> | |
len: u32, | ||
slice: &'a [u8], | ||
) -> Result<Self, VarZeroVecFormatError> { | ||
let len_minus_one = len.checked_sub(1); | ||
// The empty VZV is special-cased to the empty slice | ||
if len == 0 { | ||
let Some(len_minus_one) = len_minus_one else { | ||
return Ok(VarZeroVecComponents { | ||
len: 0, | ||
indices: &[], | ||
things: &[], | ||
marker: PhantomData, | ||
}); | ||
} | ||
}; | ||
// The indices array is one element shorter since the first index is always 0, | ||
// so we use len_minus_one | ||
let indices_bytes = slice | ||
.get(..F::Index::SIZE * (len as usize)) | ||
.get(..F::Index::SIZE * (len_minus_one as usize)) | ||
.ok_or(VarZeroVecFormatError::Metadata)?; | ||
let things = slice | ||
.get(F::Index::SIZE * (len as usize)..) | ||
.get(F::Index::SIZE * (len_minus_one as usize)..) | ||
.ok_or(VarZeroVecFormatError::Metadata)?; | ||
|
||
let borrowed = VarZeroVecComponents { | ||
|
@@ -331,17 +337,20 @@ impl<'a, T: VarULE + ?Sized, F: VarZeroVecFormat> VarZeroVecComponents<'a, T, F> | |
/// The len,bytes must have previously successfully run through | ||
/// [`VarZeroVecComponents::parse_byte_slice_with_length()`] | ||
pub unsafe fn from_bytes_unchecked_with_length(len: u32, slice: &'a [u8]) -> Self { | ||
let len_minus_one = len.checked_sub(1); | ||
// The empty VZV is special-cased to the empty slice | ||
if len == 0 { | ||
let Some(len_minus_one) = len_minus_one else { | ||
return VarZeroVecComponents { | ||
len: 0, | ||
indices: &[], | ||
things: &[], | ||
marker: PhantomData, | ||
}; | ||
} | ||
let indices_bytes = slice.get_unchecked(..F::Index::SIZE * (len as usize)); | ||
let things = slice.get_unchecked(F::Index::SIZE * (len as usize)..); | ||
}; | ||
// The indices array is one element shorter since the first index is always 0, | ||
// so we use len_minus_one | ||
let indices_bytes = slice.get_unchecked(..F::Index::SIZE * (len_minus_one as usize)); | ||
let things = slice.get_unchecked(F::Index::SIZE * (len_minus_one as usize)..); | ||
|
||
VarZeroVecComponents { | ||
len, | ||
|
@@ -360,7 +369,7 @@ impl<'a, T: VarULE + ?Sized, F: VarZeroVecFormat> VarZeroVecComponents<'a, T, F> | |
/// Returns `true` if the vector contains no elements. | ||
#[inline] | ||
pub fn is_empty(self) -> bool { | ||
self.indices.is_empty() | ||
self.len == 0 | ||
} | ||
|
||
/// Get the idx'th element out of this slice. Returns `None` if out of bounds. | ||
|
@@ -389,11 +398,17 @@ impl<'a, T: VarULE + ?Sized, F: VarZeroVecFormat> VarZeroVecComponents<'a, T, F> | |
/// - `idx` must be in bounds (`idx < self.len()`) | ||
#[inline] | ||
pub(crate) unsafe fn get_things_range(self, idx: usize) -> Range<usize> { | ||
let start = self.indices_slice().get_unchecked(idx).iule_to_usize(); | ||
let start = if let Some(idx_minus_one) = idx.checked_sub(1) { | ||
self.indices_slice() | ||
.get_unchecked(idx_minus_one) | ||
.iule_to_usize() | ||
} else { | ||
0 | ||
}; | ||
let end = if idx + 1 == self.len() { | ||
self.things.len() | ||
} else { | ||
self.indices_slice().get_unchecked(idx + 1).iule_to_usize() | ||
self.indices_slice().get_unchecked(idx).iule_to_usize() | ||
}; | ||
debug_assert!(start <= end); | ||
start..end | ||
|
@@ -417,26 +432,30 @@ impl<'a, T: VarULE + ?Sized, F: VarZeroVecFormat> VarZeroVecComponents<'a, T, F> | |
#[inline] | ||
#[allow(clippy::len_zero)] // more explicit to enforce safety invariants | ||
fn check_indices_and_things(self) -> Result<(), VarZeroVecFormatError> { | ||
assert_eq!(self.len(), self.indices_slice().len()); | ||
if self.len() == 0 { | ||
if self.things.len() > 0 { | ||
return Err(VarZeroVecFormatError::Metadata); | ||
} else { | ||
return Ok(()); | ||
} | ||
} | ||
let indices_slice = self.indices_slice(); | ||
assert_eq!(self.len(), indices_slice.len() + 1); | ||
// Safety: i is in bounds (assertion above) | ||
let mut start = (unsafe { *self.indices_slice().get_unchecked(0) }).iule_to_usize(); | ||
if start != 0 { | ||
return Err(VarZeroVecFormatError::Metadata); | ||
} | ||
let mut start = 0; | ||
for i in 0..self.len() { | ||
let end = if i == self.len() - 1 { | ||
self.things.len() | ||
// The indices array is offset by 1: indices[0] is the end of the first | ||
// element and the start of the next, since the start of the first element | ||
// is always things[0]. So to get the end we get element `i`. | ||
let end = if let Some(end) = indices_slice.get(i) { | ||
end.iule_to_usize() | ||
} else { | ||
// Safety: i+1 is in bounds (assertion above) | ||
(unsafe { *self.indices_slice().get_unchecked(i + 1) }).iule_to_usize() | ||
// This only happens at i = self.len() - 1 = indices_slice.len() + 1 - 1 | ||
// = indices_slice.len(). This is the last `end`, which is always the size of | ||
// `things` and thus never stored in the array | ||
self.things.len() | ||
}; | ||
|
||
if start > end { | ||
return Err(VarZeroVecFormatError::Metadata); | ||
} | ||
|
@@ -454,17 +473,27 @@ impl<'a, T: VarULE + ?Sized, F: VarZeroVecFormat> VarZeroVecComponents<'a, T, F> | |
/// Create an iterator over the Ts contained in VarZeroVecComponents | ||
#[inline] | ||
pub fn iter(self) -> impl Iterator<Item = &'a T> { | ||
self.indices_slice() | ||
.iter() | ||
.copied() | ||
.map(IntegerULE::iule_to_usize) | ||
// The indices array doesn't contain 0 or len, we need to graft it on | ||
// However we don't want to graft it on for an empty vector. | ||
let (begin, end) = if self.is_empty() { | ||
(None, None) | ||
} else { | ||
(Some(0), Some(self.things.len())) | ||
}; | ||
begin | ||
.into_iter() | ||
.chain( | ||
self.indices_slice() | ||
.iter() | ||
.copied() | ||
.map(IntegerULE::iule_to_usize), | ||
) | ||
.zip( | ||
self.indices_slice() | ||
.iter() | ||
.copied() | ||
.map(IntegerULE::iule_to_usize) | ||
.skip(1) | ||
.chain([self.things.len()]), | ||
.chain(end), | ||
Manishearth marked this conversation as resolved.
Show resolved
Hide resolved
|
||
) | ||
.map(move |(start, end)| unsafe { self.things.get_unchecked(start..end) }) | ||
.map(|bytes| unsafe { T::from_byte_slice_unchecked(bytes) }) | ||
|
@@ -502,16 +531,15 @@ where | |
/// Binary searches a sorted `VarZeroVecComponents<T>` for the given element. For more information, see | ||
/// the primitive function [`binary_search`](slice::binary_search). | ||
pub fn binary_search(&self, needle: &T) -> Result<usize, usize> { | ||
self.binary_search_impl(|probe| probe.cmp(needle), self.indices_slice()) | ||
self.binary_search_by(|probe| probe.cmp(needle)) | ||
} | ||
|
||
pub fn binary_search_in_range( | ||
&self, | ||
needle: &T, | ||
range: Range<usize>, | ||
) -> Option<Result<usize, usize>> { | ||
let indices_slice = self.indices_slice().get(range)?; | ||
Some(self.binary_search_impl(|probe| probe.cmp(needle), indices_slice)) | ||
self.binary_search_in_range_by(|probe| probe.cmp(needle), range) | ||
} | ||
} | ||
|
||
|
@@ -524,58 +552,77 @@ where | |
/// Binary searches a sorted `VarZeroVecComponents<T>` for the given predicate. For more information, see | ||
/// the primitive function [`binary_search_by`](slice::binary_search_by). | ||
pub fn binary_search_by(&self, predicate: impl FnMut(&T) -> Ordering) -> Result<usize, usize> { | ||
self.binary_search_impl(predicate, self.indices_slice()) | ||
// Safety: 0 and len are in range | ||
unsafe { self.binary_search_in_range_unchecked(predicate, 0..self.len()) } | ||
} | ||
|
||
// Binary search within a range. | ||
// Values returned are relative to the range start! | ||
pub fn binary_search_in_range_by( | ||
&self, | ||
predicate: impl FnMut(&T) -> Ordering, | ||
range: Range<usize>, | ||
) -> Option<Result<usize, usize>> { | ||
let indices_slice = self.indices_slice().get(range)?; | ||
Some(self.binary_search_impl(predicate, indices_slice)) | ||
} | ||
|
||
/// Binary searches a sorted `VarZeroVecComponents<T>` with the given predicate. For more information, see | ||
/// the primitive function [`binary_search`](slice::binary_search). | ||
fn binary_search_impl( | ||
if range.start > self.len() { | ||
return None; | ||
} | ||
if range.end > self.len() { | ||
return None; | ||
} | ||
if range.end < range.start { | ||
return None; | ||
} | ||
Manishearth marked this conversation as resolved.
Show resolved
Hide resolved
|
||
// Safety: We bounds checked above | ||
let range_absolute = | ||
unsafe { self.binary_search_in_range_unchecked(predicate, range.clone()) }; | ||
// The values returned are relative to the range start | ||
Some( | ||
range_absolute | ||
.map(|o| o - range.start) | ||
.map_err(|e| e - range.start), | ||
) | ||
} | ||
|
||
/// Safety: range must be in range for the slice (start <= len, end <= len, start <= end) | ||
unsafe fn binary_search_in_range_unchecked( | ||
&self, | ||
mut predicate: impl FnMut(&T) -> Ordering, | ||
indices_slice: &[F::Index], | ||
range: Range<usize>, | ||
) -> Result<usize, usize> { | ||
// This code is an absolute atrocity. This code is not a place of honor. This | ||
// code is known to the State of California to cause cancer. | ||
// | ||
// Unfortunately, the stdlib's `binary_search*` functions can only operate on slices. | ||
// We do not have a slice. We have something we can .get() and index on, but that is not | ||
// a slice. | ||
// | ||
// The `binary_search*` functions also do not have a variant where they give you the element's | ||
// index, which we could otherwise use to directly index `self`. | ||
// We do have `self.indices`, but these are indices into a byte buffer, which cannot in | ||
// isolation be used to recoup the logical index of the element they refer to. | ||
// | ||
// However, `binary_search_by()` provides references to the elements of the slice being iterated. | ||
// Since the layout of Rust slices is well-defined, we can do pointer arithmetic on these references | ||
// to obtain the index being used by the search. | ||
// | ||
// It's worth noting that the slice we choose to search is irrelevant, as long as it has the appropriate | ||
// length. `self.indices` is defined to have length `self.len()`, so it is convenient to use | ||
// here and does not require additional allocations. | ||
// | ||
// The alternative to doing this is to implement our own binary search. This is significantly less fun. | ||
|
||
// Note: We always use zero_index relative to the whole indices array, even if we are | ||
// only searching a subslice of it. | ||
let zero_index = self.indices.as_ptr() as *const _ as usize; | ||
indices_slice.binary_search_by(|probe: &_| { | ||
// `self.indices` is a vec of unaligned F::Index::SIZE values, so we divide by F::Index::SIZE | ||
// to get the actual index | ||
let index = (probe as *const _ as usize - zero_index) / F::Index::SIZE; | ||
// safety: we know this is in bounds | ||
let actual_probe = unsafe { self.get_unchecked(index) }; | ||
predicate(actual_probe) | ||
}) | ||
// Function invariant: size is always end - start | ||
let mut start = range.start; | ||
let mut end = range.end; | ||
let mut size; | ||
|
||
// Loop invariant: 0 <= start < end <= len | ||
// This invariant is initialized by the function safety invariants and the loop condition | ||
while start < end { | ||
size = end - start; | ||
// This establishes mid < end (which implies mid < len) | ||
// size is end - start. start + size is end (which is <= len). | ||
// mid = start + size/2 will be less than end | ||
let mid = start + size / 2; | ||
|
||
// Safety: mid is < end <= len, so in-range | ||
let cmp = predicate(self.get_unchecked(mid)); | ||
|
||
match cmp { | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Question: why did you expand the binary_search impl instead of using the std one? There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. The hacky impl was written assuming a slice to search on. Now I could still use a byte slice that I have cut out appropriately, but the indices slice that we currently use is no longer long enough for this to work, and it was going to get more complicated, so I decided to dispense with the weird hack. I can bring it back and make it work with the weird hack if desired. |
||
Ordering::Less => { | ||
// This retains the loop invariant since it | ||
// increments start, and we already have 0 <= start | ||
// start < end is enforced by the loop condition | ||
start = mid + 1; | ||
} | ||
Ordering::Greater => { | ||
// mid < end, so this decreases end. | ||
// This means end <= len is still true, and | ||
// end > start is enforced by the loop condition | ||
end = mid; | ||
} | ||
Ordering::Equal => return Ok(mid), | ||
} | ||
} | ||
Err(start) | ||
} | ||
} | ||
|
||
|
@@ -619,34 +666,37 @@ where | |
// idx_offset = offset from the start of the buffer for the next index | ||
let mut idx_offset: usize = 0; | ||
// first_dat_offset = offset from the start of the buffer of the first data block | ||
let first_dat_offset: usize = idx_offset + elements.len() * F::Index::SIZE; | ||
let first_dat_offset: usize = idx_offset + (elements.len() - 1) * F::Index::SIZE; | ||
// dat_offset = offset from the start of the buffer of the next data block | ||
let mut dat_offset: usize = first_dat_offset; | ||
|
||
for element in elements.iter() { | ||
for (i, element) in elements.iter().enumerate() { | ||
let element_len = element.encode_var_ule_len(); | ||
|
||
let idx_limit = idx_offset + F::Index::SIZE; | ||
#[allow(clippy::indexing_slicing)] // Function contract allows panicky behavior | ||
let idx_slice = &mut output[idx_offset..idx_limit]; | ||
// VZV expects data offsets to be stored relative to the first data block | ||
let idx = dat_offset - first_dat_offset; | ||
assert!(idx <= F::Index::MAX_VALUE as usize); | ||
#[allow(clippy::expect_used)] // this function is explicitly panicky | ||
let bytes_to_write = F::Index::iule_from_usize(idx).expect(F::Index::TOO_LARGE_ERROR); | ||
idx_slice.copy_from_slice(ULE::as_byte_slice(&[bytes_to_write])); | ||
// The first index is always 0. We don't write it, or update the idx offset. | ||
if i != 0 { | ||
let idx_limit = idx_offset + F::Index::SIZE; | ||
#[allow(clippy::indexing_slicing)] // Function contract allows panicky behavior | ||
let idx_slice = &mut output[idx_offset..idx_limit]; | ||
// VZV expects data offsets to be stored relative to the first data block | ||
let idx = dat_offset - first_dat_offset; | ||
assert!(idx <= F::Index::MAX_VALUE as usize); | ||
#[allow(clippy::expect_used)] // this function is explicitly panicky | ||
let bytes_to_write = F::Index::iule_from_usize(idx).expect(F::Index::TOO_LARGE_ERROR); | ||
idx_slice.copy_from_slice(ULE::as_byte_slice(&[bytes_to_write])); | ||
|
||
idx_offset = idx_limit; | ||
} | ||
|
||
let dat_limit = dat_offset + element_len; | ||
#[allow(clippy::indexing_slicing)] // Function contract allows panicky behavior | ||
let dat_slice = &mut output[dat_offset..dat_limit]; | ||
element.encode_var_ule_write(dat_slice); | ||
debug_assert_eq!(T::validate_byte_slice(dat_slice), Ok(())); | ||
|
||
idx_offset = idx_limit; | ||
dat_offset = dat_limit; | ||
} | ||
|
||
debug_assert_eq!(idx_offset, F::Index::SIZE * elements.len()); | ||
debug_assert_eq!(idx_offset, F::Index::SIZE * (elements.len() - 1)); | ||
assert_eq!(dat_offset, output.len()); | ||
} | ||
|
||
|
@@ -682,10 +732,12 @@ where | |
A: EncodeAsVarULE<T>, | ||
F: VarZeroVecFormat, | ||
{ | ||
if elements.is_empty() { | ||
let elements_len = elements.len(); | ||
let Some(elements_len_minus_one) = elements_len.checked_sub(1) else { | ||
// Empty vec is optimized to an empty byte representation | ||
return Some(0); | ||
} | ||
let idx_len: u32 = u32::try_from(elements.len()) | ||
}; | ||
let idx_len: u32 = u32::try_from(elements_len_minus_one) | ||
.ok()? | ||
.checked_mul(F::Index::SIZE as u32)?; | ||
let data_len: u32 = elements | ||
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thought: you should use
split_at_unchecked
here, but I think it isn't MSRV yet