Revision control
Copy as Markdown
Other Tools
// Copyright Mozilla Foundation. See the COPYRIGHT
// file at the top-level directory of this distribution.
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// option. This file may not be copied, modified, or distributed
// except according to those terms.
// It's assumed that in due course Rust will have explicit SIMD but will not
// be good at run-time selection of SIMD vs. no-SIMD. In such a future,
// x86_64 will always use SSE2 and 32-bit x86 will use SSE2 when compiled with
// a Mozilla-shipped rustc. SIMD support and especially detection on ARM is a
// mess. Under the circumstances, it seems to make sense to optimize the ALU
// case for ARMv7 rather than x86. Annoyingly, I was unable to get useful
// numbers of the actual ARMv7 CPU I have access to, because (thermal?)
// throttling kept interfering. Since Raspberry Pi 3 (ARMv8 core but running
// ARMv7 code) produced reproducible performance numbers, that's the ARM
// computer that this code ended up being optimized for in the ALU case.
// Less popular CPU architectures simply get the approach that was chosen based
// on Raspberry Pi 3 measurements. The UTF-16 and UTF-8 ALU cases take
// different approaches based on benchmarking on Raspberry Pi 3.
#[cfg(all(
feature = "simd-accel",
any(
target_feature = "sse2",
all(target_endian = "little", target_arch = "aarch64"),
all(target_endian = "little", target_feature = "neon")
)
))]
use crate::simd_funcs::*;
cfg_if! {
if #[cfg(feature = "simd-accel")] {
#[allow(unused_imports)]
use ::core::intrinsics::unlikely;
#[allow(unused_imports)]
use ::core::intrinsics::likely;
} else {
#[allow(dead_code)]
#[inline(always)]
fn unlikely(b: bool) -> bool {
b
}
#[allow(dead_code)]
#[inline(always)]
fn likely(b: bool) -> bool {
b
}
}
}
// Safety invariants for masks: data & mask = 0 for valid ASCII or basic latin utf-16
// `as` truncates, so works on 32-bit, too.
#[allow(dead_code)]
pub const ASCII_MASK: usize = 0x8080_8080_8080_8080u64 as usize;
// `as` truncates, so works on 32-bit, too.
#[allow(dead_code)]
pub const BASIC_LATIN_MASK: usize = 0xFF80_FF80_FF80_FF80u64 as usize;
#[allow(unused_macros)]
macro_rules! ascii_naive {
($name:ident, $src_unit:ty, $dst_unit:ty) => {
/// Safety: src and dst must have len_unit elements and be aligned
/// Safety-usable invariant: will return Some() when it fails
/// to convert. The first value will be a u8 that is > 127.
#[inline(always)]
pub unsafe fn $name(
src: *const $src_unit,
dst: *mut $dst_unit,
len: usize,
) -> Option<($src_unit, usize)> {
// Yes, manually omitting the bound check here matters
// a lot for perf.
for i in 0..len {
// Safety: len invariant used here
let code_unit = *(src.add(i));
// Safety: Upholds safety-usable invariant here
if code_unit > 127 {
return Some((code_unit, i));
}
// Safety: len invariant used here
*(dst.add(i)) = code_unit as $dst_unit;
}
return None;
}
};
}
#[allow(unused_macros)]
macro_rules! ascii_alu {
($name:ident,
// safety invariant: src/dst MUST be u8
$src_unit:ty,
$dst_unit:ty,
// Safety invariant: stride_fn must consume and produce two usizes, and return the index of the first non-ascii when it fails
$stride_fn:ident) => {
/// Safety: src and dst must have len elements, src is valid for read, dst is valid for
/// write
/// Safety-usable invariant: will return Some() when it fails
/// to convert. The first value will be a u8 that is > 127.
#[cfg_attr(feature = "cargo-clippy", allow(never_loop, cast_ptr_alignment))]
#[inline(always)]
pub unsafe fn $name(
src: *const $src_unit,
dst: *mut $dst_unit,
len: usize,
) -> Option<($src_unit, usize)> {
let mut offset = 0usize;
// This loop is only broken out of as a `goto` forward
loop {
// Safety: until_alignment becomes the number of bytes we need to munch until we are aligned to usize
let mut until_alignment = {
// Check if the other unit aligns if we move the narrower unit
// to alignment.
// if ::core::mem::size_of::<$src_unit>() == ::core::mem::size_of::<$dst_unit>() {
// ascii_to_ascii
let src_alignment = (src as usize) & ALU_ALIGNMENT_MASK;
let dst_alignment = (dst as usize) & ALU_ALIGNMENT_MASK;
if src_alignment != dst_alignment {
// Safety: bails early and ends up in the naïve branch where usize-alignment doesn't matter
break;
}
(ALU_ALIGNMENT - src_alignment) & ALU_ALIGNMENT_MASK
// } else if ::core::mem::size_of::<$src_unit>() < ::core::mem::size_of::<$dst_unit>() {
// ascii_to_basic_latin
// let src_until_alignment = (ALIGNMENT - ((src as usize) & ALIGNMENT_MASK)) & ALIGNMENT_MASK;
// if (dst.add(src_until_alignment) as usize) & ALIGNMENT_MASK != 0 {
// break;
// }
// src_until_alignment
// } else {
// basic_latin_to_ascii
// let dst_until_alignment = (ALIGNMENT - ((dst as usize) & ALIGNMENT_MASK)) & ALIGNMENT_MASK;
// if (src.add(dst_until_alignment) as usize) & ALIGNMENT_MASK != 0 {
// break;
// }
// dst_until_alignment
// }
};
if until_alignment + ALU_STRIDE_SIZE <= len {
// Moving pointers to alignment seems to be a pessimization on
// x86_64 for operations that have UTF-16 as the internal
// Unicode representation. However, since it seems to be a win
// on ARM (tested ARMv7 code running on ARMv8 [rpi3]), except
// mixed results when encoding from UTF-16 and since x86 and
// x86_64 should be using SSE2 in due course, keeping the move
// to alignment here. It would be good to test on more ARM CPUs
// and on real MIPS and POWER hardware.
//
// Safety: This is the naïve code once again, for `until_alignment` bytes
while until_alignment != 0 {
let code_unit = *(src.add(offset));
if code_unit > 127 {
// Safety: Upholds safety-usable invariant here
return Some((code_unit, offset));
}
*(dst.add(offset)) = code_unit as $dst_unit;
// Safety: offset is the number of bytes copied so far
offset += 1;
until_alignment -= 1;
}
let len_minus_stride = len - ALU_STRIDE_SIZE;
loop {
// Safety: num_ascii is known to be a byte index of a non-ascii byte due to stride_fn's invariant
if let Some(num_ascii) = $stride_fn(
// Safety: These are known to be valid and aligned since we have at
// least ALU_STRIDE_SIZE data in these buffers, and offset is the
// number of elements copied so far, which according to the
// until_alignment calculation above will cause both src and dst to be
// aligned to usize after this add
src.add(offset) as *const usize,
dst.add(offset) as *mut usize,
) {
offset += num_ascii;
// Safety: Upholds safety-usable invariant here by indexing into non-ascii byte
return Some((*(src.add(offset)), offset));
}
// Safety: offset continues to be the number of bytes copied so far, and
// maintains usize alignment for the next loop iteration
offset += ALU_STRIDE_SIZE;
// Safety: This is `offset > len - stride. This loop will continue as long as
// `offset <= len - stride`, which means there are `stride` bytes to still be read.
if offset > len_minus_stride {
break;
}
}
}
break;
}
// Safety: This is the naïve code, same as ascii_naive, and has no requirements
// other than src/dst being valid for the the right lens
while offset < len {
// Safety: len invariant used here
let code_unit = *(src.add(offset));
if code_unit > 127 {
// Safety: Upholds safety-usable invariant here
return Some((code_unit, offset));
}
// Safety: len invariant used here
*(dst.add(offset)) = code_unit as $dst_unit;
offset += 1;
}
None
}
};
}
#[allow(unused_macros)]
macro_rules! basic_latin_alu {
($name:ident,
// safety invariant: use u8 for src/dest for ascii, and u16 for basic_latin
$src_unit:ty,
$dst_unit:ty,
// safety invariant: stride function must munch ALU_STRIDE_SIZE*size(src_unit) bytes off of src and
// write ALU_STRIDE_SIZE*size(dst_unit) bytes to dst
$stride_fn:ident) => {
/// Safety: src and dst must have len elements, src is valid for read, dst is valid for
/// write
/// Safety-usable invariant: will return Some() when it fails
/// to convert. The first value will be a u8 that is > 127.
#[cfg_attr(
feature = "cargo-clippy",
allow(never_loop, cast_ptr_alignment, cast_lossless)
)]
#[inline(always)]
pub unsafe fn $name(
src: *const $src_unit,
dst: *mut $dst_unit,
len: usize,
) -> Option<($src_unit, usize)> {
let mut offset = 0usize;
// This loop is only broken out of as a `goto` forward
loop {
// Safety: until_alignment becomes the number of bytes we need to munch from src/dest until we are aligned to usize
// We ensure basic-latin has the same alignment as ascii, starting with ascii since it is smaller.
let mut until_alignment = {
// Check if the other unit aligns if we move the narrower unit
// to alignment.
// if ::core::mem::size_of::<$src_unit>() == ::core::mem::size_of::<$dst_unit>() {
// ascii_to_ascii
// let src_alignment = (src as usize) & ALIGNMENT_MASK;
// let dst_alignment = (dst as usize) & ALIGNMENT_MASK;
// if src_alignment != dst_alignment {
// break;
// }
// (ALIGNMENT - src_alignment) & ALIGNMENT_MASK
// } else
if ::core::mem::size_of::<$src_unit>() < ::core::mem::size_of::<$dst_unit>() {
// ascii_to_basic_latin
let src_until_alignment = (ALU_ALIGNMENT
- ((src as usize) & ALU_ALIGNMENT_MASK))
& ALU_ALIGNMENT_MASK;
if (dst.wrapping_add(src_until_alignment) as usize) & ALU_ALIGNMENT_MASK
!= 0
{
break;
}
src_until_alignment
} else {
// basic_latin_to_ascii
let dst_until_alignment = (ALU_ALIGNMENT
- ((dst as usize) & ALU_ALIGNMENT_MASK))
& ALU_ALIGNMENT_MASK;
if (src.wrapping_add(dst_until_alignment) as usize) & ALU_ALIGNMENT_MASK
!= 0
{
break;
}
dst_until_alignment
}
};
if until_alignment + ALU_STRIDE_SIZE <= len {
// Moving pointers to alignment seems to be a pessimization on
// x86_64 for operations that have UTF-16 as the internal
// Unicode representation. However, since it seems to be a win
// on ARM (tested ARMv7 code running on ARMv8 [rpi3]), except
// mixed results when encoding from UTF-16 and since x86 and
// x86_64 should be using SSE2 in due course, keeping the move
// to alignment here. It would be good to test on more ARM CPUs
// and on real MIPS and POWER hardware.
//
// Safety: This is the naïve code once again, for `until_alignment` bytes
while until_alignment != 0 {
let code_unit = *(src.add(offset));
if code_unit > 127 {
// Safety: Upholds safety-usable invariant here
return Some((code_unit, offset));
}
*(dst.add(offset)) = code_unit as $dst_unit;
// Safety: offset is the number of bytes copied so far
offset += 1;
until_alignment -= 1;
}
let len_minus_stride = len - ALU_STRIDE_SIZE;
loop {
if !$stride_fn(
// Safety: These are known to be valid and aligned since we have at
// least ALU_STRIDE_SIZE data in these buffers, and offset is the
// number of elements copied so far, which according to the
// until_alignment calculation above will cause both src and dst to be
// aligned to usize after this add
src.add(offset) as *const usize,
dst.add(offset) as *mut usize,
) {
break;
}
// Safety: offset continues to be the number of bytes copied so far, and
// maintains usize alignment for the next loop iteration
offset += ALU_STRIDE_SIZE;
// Safety: This is `offset > len - stride. This loop will continue as long as
// `offset <= len - stride`, which means there are `stride` bytes to still be read.
if offset > len_minus_stride {
break;
}
}
}
break;
}
// Safety: This is the naïve code once again, for leftover bytes
while offset < len {
// Safety: len invariant used here
let code_unit = *(src.add(offset));
if code_unit > 127 {
// Safety: Upholds safety-usable invariant here
return Some((code_unit, offset));
}
// Safety: len invariant used here
*(dst.add(offset)) = code_unit as $dst_unit;
offset += 1;
}
None
}
};
}
#[allow(unused_macros)]
macro_rules! latin1_alu {
// safety invariant: stride function must munch ALU_STRIDE_SIZE*size(src_unit) bytes off of src and
// write ALU_STRIDE_SIZE*size(dst_unit) bytes to dst
($name:ident, $src_unit:ty, $dst_unit:ty, $stride_fn:ident) => {
/// Safety: src and dst must have len elements, src is valid for read, dst is valid for
/// write
#[cfg_attr(
feature = "cargo-clippy",
allow(never_loop, cast_ptr_alignment, cast_lossless)
)]
#[inline(always)]
pub unsafe fn $name(src: *const $src_unit, dst: *mut $dst_unit, len: usize) {
let mut offset = 0usize;
// This loop is only broken out of as a `goto` forward
loop {
// Safety: until_alignment becomes the number of bytes we need to munch from src/dest until we are aligned to usize
// We ensure the UTF-16 side has the same alignment as the Latin-1 side, starting with Latin-1 since it is smaller.
let mut until_alignment = {
if ::core::mem::size_of::<$src_unit>() < ::core::mem::size_of::<$dst_unit>() {
// unpack
let src_until_alignment = (ALU_ALIGNMENT
- ((src as usize) & ALU_ALIGNMENT_MASK))
& ALU_ALIGNMENT_MASK;
if (dst.wrapping_add(src_until_alignment) as usize) & ALU_ALIGNMENT_MASK
!= 0
{
break;
}
src_until_alignment
} else {
// pack
let dst_until_alignment = (ALU_ALIGNMENT
- ((dst as usize) & ALU_ALIGNMENT_MASK))
& ALU_ALIGNMENT_MASK;
if (src.wrapping_add(dst_until_alignment) as usize) & ALU_ALIGNMENT_MASK
!= 0
{
break;
}
dst_until_alignment
}
};
if until_alignment + ALU_STRIDE_SIZE <= len {
// Safety: This is the naïve code once again, for `until_alignment` bytes
while until_alignment != 0 {
let code_unit = *(src.add(offset));
*(dst.add(offset)) = code_unit as $dst_unit;
// Safety: offset is the number of bytes copied so far
offset += 1;
until_alignment -= 1;
}
let len_minus_stride = len - ALU_STRIDE_SIZE;
loop {
$stride_fn(
// Safety: These are known to be valid and aligned since we have at
// least ALU_STRIDE_SIZE data in these buffers, and offset is the
// number of elements copied so far, which according to the
// until_alignment calculation above will cause both src and dst to be
// aligned to usize after this add
src.add(offset) as *const usize,
dst.add(offset) as *mut usize,
);
// Safety: offset continues to be the number of bytes copied so far, and
// maintains usize alignment for the next loop iteration
offset += ALU_STRIDE_SIZE;
// Safety: This is `offset > len - stride. This loop will continue as long as
// `offset <= len - stride`, which means there are `stride` bytes to still be read.
if offset > len_minus_stride {
break;
}
}
}
break;
}
// Safety: This is the naïve code once again, for leftover bytes
while offset < len {
// Safety: len invariant used here
let code_unit = *(src.add(offset));
*(dst.add(offset)) = code_unit as $dst_unit;
offset += 1;
}
}
};
}
#[allow(unused_macros)]
macro_rules! ascii_simd_check_align {
(
$name:ident,
$src_unit:ty,
$dst_unit:ty,
// Safety: This function must require aligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_both_aligned:ident,
// Safety: This function must require aligned/unaligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_src_aligned:ident,
// Safety: This function must require unaligned/aligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_dst_aligned:ident,
// Safety: This function must require unaligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_neither_aligned:ident
) => {
/// Safety: src/dst must be valid for reads/writes of `len` elements of their units.
///
/// Safety-usable invariant: will return Some() when it encounters non-ASCII, with the first element in the Some being
/// guaranteed to be non-ASCII (> 127), and the second being the offset where it is found
#[inline(always)]
pub unsafe fn $name(
src: *const $src_unit,
dst: *mut $dst_unit,
len: usize,
) -> Option<($src_unit, usize)> {
let mut offset = 0usize;
// Safety: if this check succeeds we're valid for reading/writing at least `SIMD_STRIDE_SIZE` elements.
if SIMD_STRIDE_SIZE <= len {
let len_minus_stride = len - SIMD_STRIDE_SIZE;
// XXX Should we first process one stride unconditionally as unaligned to
// avoid the cost of the branchiness below if the first stride fails anyway?
// XXX Should we just use unaligned SSE2 access unconditionally? It seems that
// on Haswell, it would make sense to just use unaligned and not bother
// checking. Need to benchmark older architectures before deciding.
let dst_masked = (dst as usize) & SIMD_ALIGNMENT_MASK;
// Safety: checking whether src is aligned
if ((src as usize) & SIMD_ALIGNMENT_MASK) == 0 {
// Safety: Checking whether dst is aligned
if dst_masked == 0 {
loop {
// Safety: We're valid to read/write SIMD_STRIDE_SIZE elements and have the appropriate alignments
if !$stride_both_aligned(src.add(offset), dst.add(offset)) {
break;
}
offset += SIMD_STRIDE_SIZE;
// Safety: This is `offset > len - SIMD_STRIDE_SIZE` which means we always have at least `SIMD_STRIDE_SIZE` elements to munch next time.
if offset > len_minus_stride {
break;
}
}
} else {
loop {
// Safety: We're valid to read/write SIMD_STRIDE_SIZE elements and have the appropriate alignments
if !$stride_src_aligned(src.add(offset), dst.add(offset)) {
break;
}
offset += SIMD_STRIDE_SIZE;
// Safety: This is `offset > len - SIMD_STRIDE_SIZE` which means we always have at least `SIMD_STRIDE_SIZE` elements to munch next time.
if offset > len_minus_stride {
break;
}
}
}
} else {
if dst_masked == 0 {
loop {
// Safety: We're valid to read/write SIMD_STRIDE_SIZE elements and have the appropriate alignments
if !$stride_dst_aligned(src.add(offset), dst.add(offset)) {
break;
}
offset += SIMD_STRIDE_SIZE;
// Safety: This is `offset > len - SIMD_STRIDE_SIZE` which means we always have at least `SIMD_STRIDE_SIZE` elements to munch next time.
if offset > len_minus_stride {
break;
}
}
} else {
loop {
// Safety: We're valid to read/write SIMD_STRIDE_SIZE elements and have the appropriate alignments
if !$stride_neither_aligned(src.add(offset), dst.add(offset)) {
break;
}
offset += SIMD_STRIDE_SIZE;
// Safety: This is `offset > len - SIMD_STRIDE_SIZE` which means we always have at least `SIMD_STRIDE_SIZE` elements to munch next time.
if offset > len_minus_stride {
break;
}
}
}
}
}
while offset < len {
// Safety: uses len invariant here and below
let code_unit = *(src.add(offset));
if code_unit > 127 {
// Safety: upholds safety-usable invariant
return Some((code_unit, offset));
}
*(dst.add(offset)) = code_unit as $dst_unit;
offset += 1;
}
None
}
};
}
#[allow(unused_macros)]
macro_rules! ascii_simd_check_align_unrolled {
(
$name:ident,
$src_unit:ty,
$dst_unit:ty,
// Safety: This function must require aligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_both_aligned:ident,
// Safety: This function must require aligned/unaligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_src_aligned:ident,
// Safety: This function must require unaligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_neither_aligned:ident,
// Safety: This function must require aligned src/dest that are valid for reading/writing 2*SIMD_STRIDE_SIZE src_unit/dst_unit
$double_stride_both_aligned:ident,
// Safety: This function must require aligned/unaligned src/dest that are valid for reading/writing 2*SIMD_STRIDE_SIZE src_unit/dst_unit
$double_stride_src_aligned:ident
) => {
/// Safety: src/dst must be valid for reads/writes of `len` elements of their units.
///
/// Safety-usable invariant: will return Some() when it encounters non-ASCII, with the first element in the Some being
/// guaranteed to be non-ASCII (> 127), and the second being the offset where it is found #[inline(always)]
pub unsafe fn $name(
src: *const $src_unit,
dst: *mut $dst_unit,
len: usize,
) -> Option<($src_unit, usize)> {
let unit_size = ::core::mem::size_of::<$src_unit>();
let mut offset = 0usize;
// This loop is only broken out of as a goto forward without
// actually looping
'outer: loop {
// Safety: if this check succeeds we're valid for reading/writing at least `SIMD_STRIDE_SIZE` elements.
if SIMD_STRIDE_SIZE <= len {
// First, process one unaligned
// Safety: this is safe to call since we're valid for this read/write
if !$stride_neither_aligned(src, dst) {
break 'outer;
}
offset = SIMD_STRIDE_SIZE;
// We have now seen 16 ASCII bytes. Let's guess that
// there will be enough more to justify more expense
// in the case of non-ASCII.
// Use aligned reads for the sake of old microachitectures.
//
// Safety: this correctly calculates the number of src_units that need to be read before the remaining list is aligned.
// This is less that SIMD_ALIGNMENT, which is also SIMD_STRIDE_SIZE (as documented)
let until_alignment = ((SIMD_ALIGNMENT
- ((src.add(offset) as usize) & SIMD_ALIGNMENT_MASK))
& SIMD_ALIGNMENT_MASK)
/ unit_size;
// Safety: This addition won't overflow, because even in the 32-bit PAE case the
// address space holds enough code that the slice length can't be that
// close to address space size.
// offset now equals SIMD_STRIDE_SIZE, hence times 3 below.
//
// Safety: if this check succeeds we're valid for reading/writing at least `2 * SIMD_STRIDE_SIZE` elements plus `until_alignment`.
// The extra SIMD_STRIDE_SIZE in the condition is because `offset` is already `SIMD_STRIDE_SIZE`.
if until_alignment + (SIMD_STRIDE_SIZE * 3) <= len {
if until_alignment != 0 {
// Safety: this is safe to call since we're valid for this read/write (and more), and don't care about alignment
// This will copy over bytes that get decoded twice since it's not incrementing `offset` by SIMD_STRIDE_SIZE. This is fine.
if !$stride_neither_aligned(src.add(offset), dst.add(offset)) {
break;
}
offset += until_alignment;
}
// Safety: At this point we're valid for reading/writing 2*SIMD_STRIDE_SIZE elements
// Safety: Now `offset` is aligned for `src`
let len_minus_stride_times_two = len - (SIMD_STRIDE_SIZE * 2);
// Safety: This is whether dst is aligned
let dst_masked = (dst.add(offset) as usize) & SIMD_ALIGNMENT_MASK;
if dst_masked == 0 {
loop {
// Safety: both are aligned, we can call the aligned function. We're valid for reading/writing double stride from the initial condition
// and the loop break condition below
if let Some(advance) =
$double_stride_both_aligned(src.add(offset), dst.add(offset))
{
offset += advance;
let code_unit = *(src.add(offset));
// Safety: uses safety-usable invariant on ascii_to_ascii_simd_double_stride to return
// guaranteed non-ascii
return Some((code_unit, offset));
}
offset += SIMD_STRIDE_SIZE * 2;
// Safety: This is `offset > len - 2 * SIMD_STRIDE_SIZE` which means we always have at least `2 * SIMD_STRIDE_SIZE` elements to munch next time.
if offset > len_minus_stride_times_two {
break;
}
}
// Safety: We're valid for reading/writing one more, and can still assume alignment
if offset + SIMD_STRIDE_SIZE <= len {
if !$stride_both_aligned(src.add(offset), dst.add(offset)) {
break 'outer;
}
offset += SIMD_STRIDE_SIZE;
}
} else {
loop {
// Safety: only src is aligned here. We're valid for reading/writing double stride from the initial condition
// and the loop break condition below
if let Some(advance) =
$double_stride_src_aligned(src.add(offset), dst.add(offset))
{
offset += advance;
let code_unit = *(src.add(offset));
// Safety: uses safety-usable invariant on ascii_to_ascii_simd_double_stride to return
// guaranteed non-ascii
return Some((code_unit, offset));
}
offset += SIMD_STRIDE_SIZE * 2;
// Safety: This is `offset > len - 2 * SIMD_STRIDE_SIZE` which means we always have at least `2 * SIMD_STRIDE_SIZE` elements to munch next time.
if offset > len_minus_stride_times_two {
break;
}
}
// Safety: We're valid for reading/writing one more, and can still assume alignment
if offset + SIMD_STRIDE_SIZE <= len {
if !$stride_src_aligned(src.add(offset), dst.add(offset)) {
break 'outer;
}
offset += SIMD_STRIDE_SIZE;
}
}
} else {
// At most two iterations, so unroll
if offset + SIMD_STRIDE_SIZE <= len {
// Safety: The check above ensures we're allowed to read/write this, and we don't use alignment
if !$stride_neither_aligned(src.add(offset), dst.add(offset)) {
break;
}
offset += SIMD_STRIDE_SIZE;
if offset + SIMD_STRIDE_SIZE <= len {
// Safety: The check above ensures we're allowed to read/write this, and we don't use alignment
if !$stride_neither_aligned(src.add(offset), dst.add(offset)) {
break;
}
offset += SIMD_STRIDE_SIZE;
}
}
}
}
break 'outer;
}
while offset < len {
// Safety: relies straightforwardly on the `len` invariant
let code_unit = *(src.add(offset));
if code_unit > 127 {
// Safety-usable invariant upheld here
return Some((code_unit, offset));
}
*(dst.add(offset)) = code_unit as $dst_unit;
offset += 1;
}
None
}
};
}
#[allow(unused_macros)]
macro_rules! latin1_simd_check_align {
(
$name:ident,
$src_unit:ty,
$dst_unit:ty,
// Safety: This function must require aligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_both_aligned:ident,
// Safety: This function must require aligned/unaligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_src_aligned:ident,
// Safety: This function must require unaligned/aligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_dst_aligned:ident,
// Safety: This function must require unaligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_neither_aligned:ident
) => {
/// Safety: src/dst must be valid for reads/writes of `len` elements of their units.
#[inline(always)]
pub unsafe fn $name(src: *const $src_unit, dst: *mut $dst_unit, len: usize) {
let mut offset = 0usize;
// Safety: if this check succeeds we're valid for reading/writing at least `SIMD_STRIDE_SIZE` elements.
if SIMD_STRIDE_SIZE <= len {
let len_minus_stride = len - SIMD_STRIDE_SIZE;
// Whether dst is aligned
let dst_masked = (dst as usize) & SIMD_ALIGNMENT_MASK;
// Whether src is aligned
if ((src as usize) & SIMD_ALIGNMENT_MASK) == 0 {
if dst_masked == 0 {
loop {
// Safety: Both were aligned, we can use the aligned function
$stride_both_aligned(src.add(offset), dst.add(offset));
offset += SIMD_STRIDE_SIZE;
// Safety: This is `offset > len - SIMD_STRIDE_SIZE`, which means in the next iteration we're valid for
// reading/writing at least SIMD_STRIDE_SIZE elements.
if offset > len_minus_stride {
break;
}
}
} else {
loop {
// Safety: src was aligned, dst was not
$stride_src_aligned(src.add(offset), dst.add(offset));
offset += SIMD_STRIDE_SIZE;
// Safety: This is `offset > len - SIMD_STRIDE_SIZE`, which means in the next iteration we're valid for
// reading/writing at least SIMD_STRIDE_SIZE elements.
if offset > len_minus_stride {
break;
}
}
}
} else {
if dst_masked == 0 {
loop {
// Safety: src was aligned, dst was not
$stride_dst_aligned(src.add(offset), dst.add(offset));
offset += SIMD_STRIDE_SIZE;
// Safety: This is `offset > len - SIMD_STRIDE_SIZE`, which means in the next iteration we're valid for
// reading/writing at least SIMD_STRIDE_SIZE elements.
if offset > len_minus_stride {
break;
}
}
} else {
loop {
// Safety: Neither were aligned
$stride_neither_aligned(src.add(offset), dst.add(offset));
offset += SIMD_STRIDE_SIZE;
// Safety: This is `offset > len - SIMD_STRIDE_SIZE`, which means in the next iteration we're valid for
// reading/writing at least SIMD_STRIDE_SIZE elements.
if offset > len_minus_stride {
break;
}
}
}
}
}
while offset < len {
// Safety: relies straightforwardly on the `len` invariant
let code_unit = *(src.add(offset));
*(dst.add(offset)) = code_unit as $dst_unit;
offset += 1;
}
}
};
}
#[allow(unused_macros)]
macro_rules! latin1_simd_check_align_unrolled {
(
$name:ident,
$src_unit:ty,
$dst_unit:ty,
// Safety: This function must require aligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_both_aligned:ident,
// Safety: This function must require aligned/unaligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_src_aligned:ident,
// Safety: This function must require unaligned/aligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_dst_aligned:ident,
// Safety: This function must require unaligned src/dest that are valid for reading/writing SIMD_STRIDE_SIZE src_unit/dst_unit
$stride_neither_aligned:ident
) => {
/// Safety: src/dst must be valid for reads/writes of `len` elements of their units.
#[inline(always)]
pub unsafe fn $name(src: *const $src_unit, dst: *mut $dst_unit, len: usize) {
let unit_size = ::core::mem::size_of::<$src_unit>();
let mut offset = 0usize;
// Safety: if this check succeeds we're valid for reading/writing at least `SIMD_STRIDE_SIZE` elements.
if SIMD_STRIDE_SIZE <= len {
// Safety: this correctly calculates the number of src_units that need to be read before the remaining list is aligned.
// This is by definition less than SIMD_STRIDE_SIZE.
let mut until_alignment = ((SIMD_STRIDE_SIZE
- ((src as usize) & SIMD_ALIGNMENT_MASK))
& SIMD_ALIGNMENT_MASK)
/ unit_size;
while until_alignment != 0 {
// Safety: This is a straightforward copy, since until_alignment is < SIMD_STRIDE_SIZE < len, this is in-bounds
*(dst.add(offset)) = *(src.add(offset)) as $dst_unit;
offset += 1;
until_alignment -= 1;
}
// Safety: here offset will be `until_alignment`, i.e. enough to align `src`.
let len_minus_stride = len - SIMD_STRIDE_SIZE;
// Safety: if this check succeeds we're valid for reading/writing at least `2 * SIMD_STRIDE_SIZE` elements.
if offset + SIMD_STRIDE_SIZE * 2 <= len {
let len_minus_stride_times_two = len_minus_stride - SIMD_STRIDE_SIZE;
// Safety: at this point src is known to be aligned at offset, dst is not.
if (dst.add(offset) as usize) & SIMD_ALIGNMENT_MASK == 0 {
loop {
// Safety: We checked alignment of dst above, we can use the alignment functions. We're allowed to read/write 2*SIMD_STRIDE_SIZE elements, which we do.
$stride_both_aligned(src.add(offset), dst.add(offset));
offset += SIMD_STRIDE_SIZE;
$stride_both_aligned(src.add(offset), dst.add(offset));
offset += SIMD_STRIDE_SIZE;
// Safety: This is `offset > len - 2 * SIMD_STRIDE_SIZE` which means we always have at least `2 * SIMD_STRIDE_SIZE` elements to munch next time.
if offset > len_minus_stride_times_two {
break;
}
}
} else {
loop {
// Safety: we ensured alignment of src already.
$stride_src_aligned(src.add(offset), dst.add(offset));
offset += SIMD_STRIDE_SIZE;
$stride_src_aligned(src.add(offset), dst.add(offset));
offset += SIMD_STRIDE_SIZE;
// Safety: This is `offset > len - 2 * SIMD_STRIDE_SIZE` which means we always have at least `2 * SIMD_STRIDE_SIZE` elements to munch next time.
if offset > len_minus_stride_times_two {
break;
}
}
}
}
// Safety: This is `offset > len - SIMD_STRIDE_SIZE` which means we are valid to munch SIMD_STRIDE_SIZE more elements, which we do
if offset < len_minus_stride {
$stride_src_aligned(src.add(offset), dst.add(offset));
offset += SIMD_STRIDE_SIZE;
}
}
while offset < len {
// Safety: uses len invariant here and below
let code_unit = *(src.add(offset));
// On x86_64, this loop autovectorizes but in the pack
// case there are instructions whose purpose is to make sure
// each u16 in the vector is truncated before packing. However,
// since we don't care about saturating behavior of SSE2 packing
// when the input isn't Latin1, those instructions are useless.
// Unfortunately, using the `assume` intrinsic to lie to the
// optimizer doesn't make LLVM omit the trunctation that we
// don't need. Possibly this loop could be manually optimized
// to do the sort of thing that LLVM does but without the
// ANDing the read vectors of u16 with a constant that discards
// the high half of each u16. As far as I can tell, the
// optimization assumes that doing a SIMD read past the end of
// the array is OK.
*(dst.add(offset)) = code_unit as $dst_unit;
offset += 1;
}
}
};
}
#[allow(unused_macros)]
macro_rules! ascii_simd_unalign {
// Safety: stride_neither_aligned must be a function that requires src/dest be valid for unaligned reads/writes for SIMD_STRIDE_SIZE elements of type src_unit/dest_unit
($name:ident, $src_unit:ty, $dst_unit:ty, $stride_neither_aligned:ident) => {
/// Safety: src and dst must be valid for reads/writes of len elements of type src_unit/dst_unit
///
/// Safety-usable invariant: will return Some() when it encounters non-ASCII, with the first element in the Some being
/// guaranteed to be non-ASCII (> 127), and the second being the offset where it is found
#[inline(always)]
pub unsafe fn $name(
src: *const $src_unit,
dst: *mut $dst_unit,
len: usize,
) -> Option<($src_unit, usize)> {
let mut offset = 0usize;
// Safety: if this check succeeds we're valid for reading/writing at least `stride` elements.
if SIMD_STRIDE_SIZE <= len {
let len_minus_stride = len - SIMD_STRIDE_SIZE;
loop {
// Safety: We know we're valid for `stride` reads/writes, so we can call this function. We don't need alignment.
if !$stride_neither_aligned(src.add(offset), dst.add(offset)) {
break;
}
offset += SIMD_STRIDE_SIZE;
// This is `offset > len - stride` which means we always have at least `stride` elements to munch next time.
if offset > len_minus_stride {
break;
}
}
}
while offset < len {
// Safety: Uses len invariant here and below
let code_unit = *(src.add(offset));
if code_unit > 127 {
// Safety-usable invariant upheld here
return Some((code_unit, offset));
}
*(dst.add(offset)) = code_unit as $dst_unit;
offset += 1;
}
None
}
};
}
#[allow(unused_macros)]
macro_rules! latin1_simd_unalign {
// Safety: stride_neither_aligned must be a function that requires src/dest be valid for unaligned reads/writes for SIMD_STRIDE_SIZE elements of type src_unit/dest_unit
($name:ident, $src_unit:ty, $dst_unit:ty, $stride_neither_aligned:ident) => {
/// Safety: src and dst must be valid for unaligned reads/writes of len elements of type src_unit/dst_unit
#[inline(always)]
pub unsafe fn $name(src: *const $src_unit, dst: *mut $dst_unit, len: usize) {
let mut offset = 0usize;
// Safety: if this check succeeds we're valid for reading/writing at least `stride` elements.
if SIMD_STRIDE_SIZE <= len {
let len_minus_stride = len - SIMD_STRIDE_SIZE;
loop {
// Safety: We know we're valid for `stride` reads/writes, so we can call this function. We don't need alignment.
$stride_neither_aligned(src.add(offset), dst.add(offset));
offset += SIMD_STRIDE_SIZE;
// This is `offset > len - stride` which means we always have at least `stride` elements to munch next time.
if offset > len_minus_stride {
break;
}
}
}
while offset < len {
// Safety: Uses len invariant here
let code_unit = *(src.add(offset));
*(dst.add(offset)) = code_unit as $dst_unit;
offset += 1;
}
}
};
}
#[allow(unused_macros)]
macro_rules! ascii_to_ascii_simd_stride {
// Safety: load/store must be valid for 16 bytes of read/write, which may be unaligned. (candidates: `(load|store)(16|8)_(unaligned|aligned)` functions)
($name:ident, $load:ident, $store:ident) => {
/// Safety: src and dst must be valid for 16 bytes of read/write according to
/// the $load/$store fn, which may allow for unaligned reads/writes or require
/// alignment to either 16x8 or u8x16.
#[inline(always)]
pub unsafe fn $name(src: *const u8, dst: *mut u8) -> bool {
let simd = $load(src);
if !simd_is_ascii(simd) {
return false;
}
$store(dst, simd);
true
}
};
}
#[allow(unused_macros)]
macro_rules! ascii_to_ascii_simd_double_stride {
// Safety: store must be valid for 32 bytes of write, which may be unaligned (candidates: `store(8|16)_(aligned|unaligned)`)
($name:ident, $store:ident) => {
/// Safety: src must be valid for 32 bytes of aligned u8x16 read
/// dst must be valid for 32 bytes of unaligned write according to
/// the $store fn, which may allow for unaligned writes or require
/// alignment to either 16x8 or u8x16.
///
/// Safety-usable invariant: Returns Some(index) if the element at `index` is invalid ASCII
#[inline(always)]
pub unsafe fn $name(src: *const u8, dst: *mut u8) -> Option<usize> {
let first = load16_aligned(src);
let second = load16_aligned(src.add(SIMD_STRIDE_SIZE));
$store(dst, first);
if unlikely(!simd_is_ascii(first | second)) {
// Safety: mask_ascii produces a mask of all the high bits.
let mask_first = mask_ascii(first);
if mask_first != 0 {
// Safety: on little endian systems this will be the number of ascii bytes
// before the first non-ascii, i.e. valid for indexing src
// TODO SAFETY: What about big-endian systems?
return Some(mask_first.trailing_zeros() as usize);
}
$store(dst.add(SIMD_STRIDE_SIZE), second);
let mask_second = mask_ascii(second);
// Safety: on little endian systems this will be the number of ascii bytes
// before the first non-ascii, i.e. valid for indexing src
return Some(SIMD_STRIDE_SIZE + mask_second.trailing_zeros() as usize);
}
$store(dst.add(SIMD_STRIDE_SIZE), second);
None
}
};
}
#[allow(unused_macros)]
macro_rules! ascii_to_basic_latin_simd_stride {
// Safety: load/store must be valid for 16 bytes of read/write, which may be unaligned. (candidates: `(load|store)(16|8)_(unaligned|aligned)` functions)
($name:ident, $load:ident, $store:ident) => {
/// Safety: src and dst must be valid for 16/32 bytes of read/write according to
/// the $load/$store fn, which may allow for unaligned reads/writes or require
/// alignment to either 16x8 or u8x16.
#[inline(always)]
pub unsafe fn $name(src: *const u8, dst: *mut u16) -> bool {
let simd = $load(src);
if !simd_is_ascii(simd) {
return false;
}
let (first, second) = simd_unpack(simd);
$store(dst, first);
$store(dst.add(8), second);
true
}
};
}
#[allow(unused_macros)]
macro_rules! ascii_to_basic_latin_simd_double_stride {
// Safety: store must be valid for 16 bytes of write, which may be unaligned
($name:ident, $store:ident) => {
/// Safety: src must be valid for 2*SIMD_STRIDE_SIZE bytes of aligned reads,
/// aligned to either 16x8 or u8x16.
/// dst must be valid for 2*SIMD_STRIDE_SIZE bytes of aligned or unaligned reads
#[inline(always)]
pub unsafe fn $name(src: *const u8, dst: *mut u16) -> Option<usize> {
let first = load16_aligned(src);
let second = load16_aligned(src.add(SIMD_STRIDE_SIZE));
let (a, b) = simd_unpack(first);
$store(dst, a);
// Safety: divide by 2 since it's a u16 pointer
$store(dst.add(SIMD_STRIDE_SIZE / 2), b);
if unlikely(!simd_is_ascii(first | second)) {
let mask_first = mask_ascii(first);
if mask_first != 0 {
return Some(mask_first.trailing_zeros() as usize);
}
let (c, d) = simd_unpack(second);
$store(dst.add(SIMD_STRIDE_SIZE), c);
$store(dst.add(SIMD_STRIDE_SIZE + (SIMD_STRIDE_SIZE / 2)), d);
let mask_second = mask_ascii(second);
return Some(SIMD_STRIDE_SIZE + mask_second.trailing_zeros() as usize);
}
let (c, d) = simd_unpack(second);
$store(dst.add(SIMD_STRIDE_SIZE), c);
$store(dst.add(SIMD_STRIDE_SIZE + (SIMD_STRIDE_SIZE / 2)), d);
None
}
};
}
#[allow(unused_macros)]
macro_rules! unpack_simd_stride {
// Safety: load/store must be valid for 16 bytes of read/write, which may be unaligned. (candidates: `(load|store)(16|8)_(unaligned|aligned)` functions)
($name:ident, $load:ident, $store:ident) => {
/// Safety: src and dst must be valid for 16 bytes of read/write according to
/// the $load/$store fn, which may allow for unaligned reads/writes or require
/// alignment to either 16x8 or u8x16.
#[inline(always)]
pub unsafe fn $name(src: *const u8, dst: *mut u16) {
let simd = $load(src);
let (first, second) = simd_unpack(simd);
$store(dst, first);
$store(dst.add(8), second);
}
};
}
#[allow(unused_macros)]
macro_rules! basic_latin_to_ascii_simd_stride {
// Safety: load/store must be valid for 16 bytes of read/write, which may be unaligned. (candidates: `(load|store)(16|8)_(unaligned|aligned)` functions)
($name:ident, $load:ident, $store:ident) => {
/// Safety: src and dst must be valid for 32/16 bytes of read/write according to
/// the $load/$store fn, which may allow for unaligned reads/writes or require
/// alignment to either 16x8 or u8x16.
#[inline(always)]
pub unsafe fn $name(src: *const u16, dst: *mut u8) -> bool {
let first = $load(src);
let second = $load(src.add(8));
if simd_is_basic_latin(first | second) {
$store(dst, simd_pack(first, second));
true
} else {
false
}
}
};
}
#[allow(unused_macros)]
macro_rules! pack_simd_stride {
// Safety: load/store must be valid for 16 bytes of read/write, which may be unaligned. (candidates: `(load|store)(16|8)_(unaligned|aligned)` functions)
($name:ident, $load:ident, $store:ident) => {
/// Safety: src and dst must be valid for 32/16 bytes of read/write according to
/// the $load/$store fn, which may allow for unaligned reads/writes or require
/// alignment to either 16x8 or u8x16.
#[inline(always)]
pub unsafe fn $name(src: *const u16, dst: *mut u8) {
let first = $load(src);
let second = $load(src.add(8));
$store(dst, simd_pack(first, second));
}
};
}
cfg_if! {
if #[cfg(all(feature = "simd-accel", target_endian = "little", target_arch = "aarch64"))] {
// SIMD with the same instructions for aligned and unaligned loads and stores
pub const SIMD_STRIDE_SIZE: usize = 16;
pub const MAX_STRIDE_SIZE: usize = 16;
// pub const ALIGNMENT: usize = 8;
pub const ALU_STRIDE_SIZE: usize = 16;
pub const ALU_ALIGNMENT: usize = 8;
pub const ALU_ALIGNMENT_MASK: usize = 7;
// Safety for stride macros: We stick to the load8_aligned/etc family of functions. We consistently produce
// neither_unaligned variants using only unaligned inputs.
ascii_to_ascii_simd_stride!(ascii_to_ascii_stride_neither_aligned, load16_unaligned, store16_unaligned);
ascii_to_basic_latin_simd_stride!(ascii_to_basic_latin_stride_neither_aligned, load16_unaligned, store8_unaligned);
unpack_simd_stride!(unpack_stride_neither_aligned, load16_unaligned, store8_unaligned);
basic_latin_to_ascii_simd_stride!(basic_latin_to_ascii_stride_neither_aligned, load8_unaligned, store16_unaligned);
pack_simd_stride!(pack_stride_neither_aligned, load8_unaligned, store16_unaligned);
// Safety for conversion macros: We use the unalign macro with unalign functions above. All stride functions were produced
// by stride macros that universally munch a single SIMD_STRIDE_SIZE worth of elements.
ascii_simd_unalign!(ascii_to_ascii, u8, u8, ascii_to_ascii_stride_neither_aligned);
ascii_simd_unalign!(ascii_to_basic_latin, u8, u16, ascii_to_basic_latin_stride_neither_aligned);
ascii_simd_unalign!(basic_latin_to_ascii, u16, u8, basic_latin_to_ascii_stride_neither_aligned);
latin1_simd_unalign!(unpack_latin1, u8, u16, unpack_stride_neither_aligned);
latin1_simd_unalign!(pack_latin1, u16, u8, pack_stride_neither_aligned);
} else if #[cfg(all(feature = "simd-accel", target_endian = "little", target_feature = "neon"))] {
// SIMD with different instructions for aligned and unaligned loads and stores.
//
// Newer microarchitectures are not supposed to have a performance difference between
// aligned and unaligned SSE2 loads and stores when the address is actually aligned,
// but the benchmark results I see don't agree.
pub const SIMD_STRIDE_SIZE: usize = 16;
pub const MAX_STRIDE_SIZE: usize = 16;
pub const SIMD_ALIGNMENT_MASK: usize = 15;
// Safety for stride macros: We stick to the load8_aligned/etc family of functions. We consistently name
// aligned/unaligned functions according to src/dst being aligned/unaligned
ascii_to_ascii_simd_stride!(ascii_to_ascii_stride_both_aligned, load16_aligned, store16_aligned);
ascii_to_ascii_simd_stride!(ascii_to_ascii_stride_src_aligned, load16_aligned, store16_unaligned);
ascii_to_ascii_simd_stride!(ascii_to_ascii_stride_dst_aligned, load16_unaligned, store16_aligned);
ascii_to_ascii_simd_stride!(ascii_to_ascii_stride_neither_aligned, load16_unaligned, store16_unaligned);
ascii_to_basic_latin_simd_stride!(ascii_to_basic_latin_stride_both_aligned, load16_aligned, store8_aligned);
ascii_to_basic_latin_simd_stride!(ascii_to_basic_latin_stride_src_aligned, load16_aligned, store8_unaligned);
ascii_to_basic_latin_simd_stride!(ascii_to_basic_latin_stride_dst_aligned, load16_unaligned, store8_aligned);
ascii_to_basic_latin_simd_stride!(ascii_to_basic_latin_stride_neither_aligned, load16_unaligned, store8_unaligned);
unpack_simd_stride!(unpack_stride_both_aligned, load16_aligned, store8_aligned);
unpack_simd_stride!(unpack_stride_src_aligned, load16_aligned, store8_unaligned);
unpack_simd_stride!(unpack_stride_dst_aligned, load16_unaligned, store8_aligned);
unpack_simd_stride!(unpack_stride_neither_aligned, load16_unaligned, store8_unaligned);
basic_latin_to_ascii_simd_stride!(basic_latin_to_ascii_stride_both_aligned, load8_aligned, store16_aligned);
basic_latin_to_ascii_simd_stride!(basic_latin_to_ascii_stride_src_aligned, load8_aligned, store16_unaligned);
basic_latin_to_ascii_simd_stride!(basic_latin_to_ascii_stride_dst_aligned, load8_unaligned, store16_aligned);
basic_latin_to_ascii_simd_stride!(basic_latin_to_ascii_stride_neither_aligned, load8_unaligned, store16_unaligned);
pack_simd_stride!(pack_stride_both_aligned, load8_aligned, store16_aligned);
pack_simd_stride!(pack_stride_src_aligned, load8_aligned, store16_unaligned);
pack_simd_stride!(pack_stride_dst_aligned, load8_unaligned, store16_aligned);
pack_simd_stride!(pack_stride_neither_aligned, load8_unaligned, store16_unaligned);
// Safety for conversion macros: We use the correct pattern of both/src/dst/neither here. All stride functions were produced
// by stride macros that universally munch a single SIMD_STRIDE_SIZE worth of elements.
ascii_simd_check_align!(ascii_to_ascii, u8, u8, ascii_to_ascii_stride_both_aligned, ascii_to_ascii_stride_src_aligned, ascii_to_ascii_stride_dst_aligned, ascii_to_ascii_stride_neither_aligned);
ascii_simd_check_align!(ascii_to_basic_latin, u8, u16, ascii_to_basic_latin_stride_both_aligned, ascii_to_basic_latin_stride_src_aligned, ascii_to_basic_latin_stride_dst_aligned, ascii_to_basic_latin_stride_neither_aligned);
ascii_simd_check_align!(basic_latin_to_ascii, u16, u8, basic_latin_to_ascii_stride_both_aligned, basic_latin_to_ascii_stride_src_aligned, basic_latin_to_ascii_stride_dst_aligned, basic_latin_to_ascii_stride_neither_aligned);
latin1_simd_check_align!(unpack_latin1, u8, u16, unpack_stride_both_aligned, unpack_stride_src_aligned, unpack_stride_dst_aligned, unpack_stride_neither_aligned);
latin1_simd_check_align!(pack_latin1, u16, u8, pack_stride_both_aligned, pack_stride_src_aligned, pack_stride_dst_aligned, pack_stride_neither_aligned);
} else if #[cfg(all(feature = "simd-accel", target_feature = "sse2"))] {
// SIMD with different instructions for aligned and unaligned loads and stores.
//
// Newer microarchitectures are not supposed to have a performance difference between
// aligned and unaligned SSE2 loads and stores when the address is actually aligned,
// but the benchmark results I see don't agree.
pub const SIMD_STRIDE_SIZE: usize = 16;
/// Safety-usable invariant: This should be identical to SIMD_STRIDE_SIZE (used by ascii_simd_check_align_unrolled)
pub const SIMD_ALIGNMENT: usize = 16;
pub const MAX_STRIDE_SIZE: usize = 16;
pub const SIMD_ALIGNMENT_MASK: usize = 15;
// Safety for stride macros: We stick to the load8_aligned/etc family of functions. We consistently name
// aligned/unaligned functions according to src/dst being aligned/unaligned
ascii_to_ascii_simd_double_stride!(ascii_to_ascii_simd_double_stride_both_aligned, store16_aligned);
ascii_to_ascii_simd_double_stride!(ascii_to_ascii_simd_double_stride_src_aligned, store16_unaligned);
ascii_to_basic_latin_simd_double_stride!(ascii_to_basic_latin_simd_double_stride_both_aligned, store8_aligned);
ascii_to_basic_latin_simd_double_stride!(ascii_to_basic_latin_simd_double_stride_src_aligned, store8_unaligned);
ascii_to_ascii_simd_stride!(ascii_to_ascii_stride_both_aligned, load16_aligned, store16_aligned);
ascii_to_ascii_simd_stride!(ascii_to_ascii_stride_src_aligned, load16_aligned, store16_unaligned);
ascii_to_ascii_simd_stride!(ascii_to_ascii_stride_neither_aligned, load16_unaligned, store16_unaligned);
ascii_to_basic_latin_simd_stride!(ascii_to_basic_latin_stride_both_aligned, load16_aligned, store8_aligned);
ascii_to_basic_latin_simd_stride!(ascii_to_basic_latin_stride_src_aligned, load16_aligned, store8_unaligned);
ascii_to_basic_latin_simd_stride!(ascii_to_basic_latin_stride_neither_aligned, load16_unaligned, store8_unaligned);
unpack_simd_stride!(unpack_stride_both_aligned, load16_aligned, store8_aligned);
unpack_simd_stride!(unpack_stride_src_aligned, load16_aligned, store8_unaligned);
basic_latin_to_ascii_simd_stride!(basic_latin_to_ascii_stride_both_aligned, load8_aligned, store16_aligned);
basic_latin_to_ascii_simd_stride!(basic_latin_to_ascii_stride_src_aligned, load8_aligned, store16_unaligned);
basic_latin_to_ascii_simd_stride!(basic_latin_to_ascii_stride_dst_aligned, load8_unaligned, store16_aligned);
basic_latin_to_ascii_simd_stride!(basic_latin_to_ascii_stride_neither_aligned, load8_unaligned, store16_unaligned);
pack_simd_stride!(pack_stride_both_aligned, load8_aligned, store16_aligned);
pack_simd_stride!(pack_stride_src_aligned, load8_aligned, store16_unaligned);
// Safety for conversion macros: We use the correct pattern of both/src/dst/neither/double_both/double_src here. All stride functions were produced
// by stride macros that universally munch a single SIMD_STRIDE_SIZE worth of elements.
ascii_simd_check_align_unrolled!(ascii_to_ascii, u8, u8, ascii_to_ascii_stride_both_aligned, ascii_to_ascii_stride_src_aligned, ascii_to_ascii_stride_neither_aligned, ascii_to_ascii_simd_double_stride_both_aligned, ascii_to_ascii_simd_double_stride_src_aligned);
ascii_simd_check_align_unrolled!(ascii_to_basic_latin, u8, u16, ascii_to_basic_latin_stride_both_aligned, ascii_to_basic_latin_stride_src_aligned, ascii_to_basic_latin_stride_neither_aligned, ascii_to_basic_latin_simd_double_stride_both_aligned, ascii_to_basic_latin_simd_double_stride_src_aligned);
ascii_simd_check_align!(basic_latin_to_ascii, u16, u8, basic_latin_to_ascii_stride_both_aligned, basic_latin_to_ascii_stride_src_aligned, basic_latin_to_ascii_stride_dst_aligned, basic_latin_to_ascii_stride_neither_aligned);
latin1_simd_check_align_unrolled!(unpack_latin1, u8, u16, unpack_stride_both_aligned, unpack_stride_src_aligned, unpack_stride_dst_aligned, unpack_stride_neither_aligned);
latin1_simd_check_align_unrolled!(pack_latin1, u16, u8, pack_stride_both_aligned, pack_stride_src_aligned, pack_stride_dst_aligned, pack_stride_neither_aligned);
} else if #[cfg(all(target_endian = "little", target_pointer_width = "64"))] {
// Aligned ALU word, little-endian, 64-bit
/// Safety invariant: this is the amount of bytes consumed by
/// unpack_alu. This will be twice the pointer width, as it consumes two usizes.
/// This is also the number of bytes produced by pack_alu.
/// This is also the number of u16 code units produced/consumed by unpack_alu/pack_alu respectively.
pub const ALU_STRIDE_SIZE: usize = 16;
pub const MAX_STRIDE_SIZE: usize = 16;
// Safety invariant: this is the pointer width in bytes
pub const ALU_ALIGNMENT: usize = 8;
// Safety invariant: this is a mask for getting the bits of a pointer not aligned to ALU_ALIGNMENT
pub const ALU_ALIGNMENT_MASK: usize = 7;
/// Safety: dst must point to valid space for writing four `usize`s
#[inline(always)]
unsafe fn unpack_alu(word: usize, second_word: usize, dst: *mut usize) {
let first = ((0x0000_0000_FF00_0000usize & word) << 24) |
((0x0000_0000_00FF_0000usize & word) << 16) |
((0x0000_0000_0000_FF00usize & word) << 8) |
(0x0000_0000_0000_00FFusize & word);
let second = ((0xFF00_0000_0000_0000usize & word) >> 8) |
((0x00FF_0000_0000_0000usize & word) >> 16) |
((0x0000_FF00_0000_0000usize & word) >> 24) |
((0x0000_00FF_0000_0000usize & word) >> 32);
let third = ((0x0000_0000_FF00_0000usize & second_word) << 24) |
((0x0000_0000_00FF_0000usize & second_word) << 16) |
((0x0000_0000_0000_FF00usize & second_word) << 8) |
(0x0000_0000_0000_00FFusize & second_word);
let fourth = ((0xFF00_0000_0000_0000usize & second_word) >> 8) |
((0x00FF_0000_0000_0000usize & second_word) >> 16) |
((0x0000_FF00_0000_0000usize & second_word) >> 24) |
((0x0000_00FF_0000_0000usize & second_word) >> 32);
// Safety: fn invariant used here
*dst = first;
*(dst.add(1)) = second;
*(dst.add(2)) = third;
*(dst.add(3)) = fourth;
}
/// Safety: dst must point to valid space for writing two `usize`s
#[inline(always)]
unsafe fn pack_alu(first: usize, second: usize, third: usize, fourth: usize, dst: *mut usize) {
let word = ((0x00FF_0000_0000_0000usize & second) << 8) |
((0x0000_00FF_0000_0000usize & second) << 16) |
((0x0000_0000_00FF_0000usize & second) << 24) |
((0x0000_0000_0000_00FFusize & second) << 32) |
((0x00FF_0000_0000_0000usize & first) >> 24) |
((0x0000_00FF_0000_0000usize & first) >> 16) |
((0x0000_0000_00FF_0000usize & first) >> 8) |
(0x0000_0000_0000_00FFusize & first);
let second_word = ((0x00FF_0000_0000_0000usize & fourth) << 8) |
((0x0000_00FF_0000_0000usize & fourth) << 16) |
((0x0000_0000_00FF_0000usize & fourth) << 24) |
((0x0000_0000_0000_00FFusize & fourth) << 32) |
((0x00FF_0000_0000_0000usize & third) >> 24) |
((0x0000_00FF_0000_0000usize & third) >> 16) |
((0x0000_0000_00FF_0000usize & third) >> 8) |
(0x0000_0000_0000_00FFusize & third);
// Safety: fn invariant used here
*dst = word;
*(dst.add(1)) = second_word;
}
} else if #[cfg(all(target_endian = "little", target_pointer_width = "32"))] {
// Aligned ALU word, little-endian, 32-bit
/// Safety invariant: this is the amount of bytes consumed by
/// unpack_alu. This will be twice the pointer width, as it consumes two usizes.
/// This is also the number of bytes produced by pack_alu.
/// This is also the number of u16 code units produced/consumed by unpack_alu/pack_alu respectively.
pub const ALU_STRIDE_SIZE: usize = 8;
pub const MAX_STRIDE_SIZE: usize = 8;
// Safety invariant: this is the pointer width in bytes
pub const ALU_ALIGNMENT: usize = 4;
// Safety invariant: this is a mask for getting the bits of a pointer not aligned to ALU_ALIGNMENT
pub const ALU_ALIGNMENT_MASK: usize = 3;
/// Safety: dst must point to valid space for writing four `usize`s
#[inline(always)]
unsafe fn unpack_alu(word: usize, second_word: usize, dst: *mut usize) {
let first = ((0x0000_FF00usize & word) << 8) |
(0x0000_00FFusize & word);
let second = ((0xFF00_0000usize & word) >> 8) |
((0x00FF_0000usize & word) >> 16);
let third = ((0x0000_FF00usize & second_word) << 8) |
(0x0000_00FFusize & second_word);
let fourth = ((0xFF00_0000usize & second_word) >> 8) |
((0x00FF_0000usize & second_word) >> 16);
// Safety: fn invariant used here
*dst = first;
*(dst.add(1)) = second;
*(dst.add(2)) = third;
*(dst.add(3)) = fourth;
}
/// Safety: dst must point to valid space for writing two `usize`s
#[inline(always)]
unsafe fn pack_alu(first: usize, second: usize, third: usize, fourth: usize, dst: *mut usize) {
let word = ((0x00FF_0000usize & second) << 8) |
((0x0000_00FFusize & second) << 16) |
((0x00FF_0000usize & first) >> 8) |
(0x0000_00FFusize & first);
let second_word = ((0x00FF_0000usize & fourth) << 8) |
((0x0000_00FFusize & fourth) << 16) |
((0x00FF_0000usize & third) >> 8) |
(0x0000_00FFusize & third);
// Safety: fn invariant used here
*dst = word;
*(dst.add(1)) = second_word;
}
} else if #[cfg(all(target_endian = "big", target_pointer_width = "64"))] {
// Aligned ALU word, big-endian, 64-bit
/// Safety invariant: this is the amount of bytes consumed by
/// unpack_alu. This will be twice the pointer width, as it consumes two usizes.
/// This is also the number of bytes produced by pack_alu.
/// This is also the number of u16 code units produced/consumed by unpack_alu/pack_alu respectively.
pub const ALU_STRIDE_SIZE: usize = 16;
pub const MAX_STRIDE_SIZE: usize = 16;
// Safety invariant: this is the pointer width in bytes
pub const ALU_ALIGNMENT: usize = 8;
// Safety invariant: this is a mask for getting the bits of a pointer not aligned to ALU_ALIGNMENT
pub const ALU_ALIGNMENT_MASK: usize = 7;
/// Safety: dst must point to valid space for writing four `usize`s
#[inline(always)]
unsafe fn unpack_alu(word: usize, second_word: usize, dst: *mut usize) {
let first = ((0xFF00_0000_0000_0000usize & word) >> 8) |
((0x00FF_0000_0000_0000usize & word) >> 16) |
((0x0000_FF00_0000_0000usize & word) >> 24) |
((0x0000_00FF_0000_0000usize & word) >> 32);
let second = ((0x0000_0000_FF00_0000usize & word) << 24) |
((0x0000_0000_00FF_0000usize & word) << 16) |
((0x0000_0000_0000_FF00usize & word) << 8) |
(0x0000_0000_0000_00FFusize & word);
let third = ((0xFF00_0000_0000_0000usize & second_word) >> 8) |
((0x00FF_0000_0000_0000usize & second_word) >> 16) |
((0x0000_FF00_0000_0000usize & second_word) >> 24) |
((0x0000_00FF_0000_0000usize & second_word) >> 32);
let fourth = ((0x0000_0000_FF00_0000usize & second_word) << 24) |
((0x0000_0000_00FF_0000usize & second_word) << 16) |
((0x0000_0000_0000_FF00usize & second_word) << 8) |
(0x0000_0000_0000_00FFusize & second_word);
// Safety: fn invariant used here
*dst = first;
*(dst.add(1)) = second;
*(dst.add(2)) = third;
*(dst.add(3)) = fourth;
}
/// Safety: dst must point to valid space for writing two `usize`s
#[inline(always)]
unsafe fn pack_alu(first: usize, second: usize, third: usize, fourth: usize, dst: *mut usize) {
let word = ((0x00FF0000_00000000usize & first) << 8) |
((0x000000FF_00000000usize & first) << 16) |
((0x00000000_00FF0000usize & first) << 24) |
((0x00000000_000000FFusize & first) << 32) |
((0x00FF0000_00000000usize & second) >> 24) |
((0x000000FF_00000000usize & second) >> 16) |
((0x00000000_00FF0000usize & second) >> 8) |
(0x00000000_000000FFusize & second);
let second_word = ((0x00FF0000_00000000usize & third) << 8) |
((0x000000FF_00000000usize & third) << 16) |
((0x00000000_00FF0000usize & third) << 24) |
((0x00000000_000000FFusize & third) << 32) |
((0x00FF0000_00000000usize & fourth) >> 24) |
((0x000000FF_00000000usize & fourth) >> 16) |
((0x00000000_00FF0000usize & fourth) >> 8) |
(0x00000000_000000FFusize & fourth);
// Safety: fn invariant used here
*dst = word;
*(dst.add(1)) = second_word;
}
} else if #[cfg(all(target_endian = "big", target_pointer_width = "32"))] {
// Aligned ALU word, big-endian, 32-bit
/// Safety invariant: this is the amount of bytes consumed by
/// unpack_alu. This will be twice the pointer width, as it consumes two usizes.
/// This is also the number of bytes produced by pack_alu.
/// This is also the number of u16 code units produced/consumed by unpack_alu/pack_alu respectively.
pub const ALU_STRIDE_SIZE: usize = 8;
pub const MAX_STRIDE_SIZE: usize = 8;
// Safety invariant: this is the pointer width in bytes
pub const ALU_ALIGNMENT: usize = 4;
// Safety invariant: this is a mask for getting the bits of a pointer not aligned to ALU_ALIGNMENT
pub const ALU_ALIGNMENT_MASK: usize = 3;
/// Safety: dst must point to valid space for writing four `usize`s
#[inline(always)]
unsafe fn unpack_alu(word: usize, second_word: usize, dst: *mut usize) {
let first = ((0xFF00_0000usize & word) >> 8) |
((0x00FF_0000usize & word) >> 16);
let second = ((0x0000_FF00usize & word) << 8) |
(0x0000_00FFusize & word);
let third = ((0xFF00_0000usize & second_word) >> 8) |
((0x00FF_0000usize & second_word) >> 16);
let fourth = ((0x0000_FF00usize & second_word) << 8) |
(0x0000_00FFusize & second_word);
// Safety: fn invariant used here
*dst = first;
*(dst.add(1)) = second;
*(dst.add(2)) = third;
*(dst.add(3)) = fourth;
}
/// Safety: dst must point to valid space for writing two `usize`s
#[inline(always)]
unsafe fn pack_alu(first: usize, second: usize, third: usize, fourth: usize, dst: *mut usize) {
let word = ((0x00FF_0000usize & first) << 8) |
((0x0000_00FFusize & first) << 16) |
((0x00FF_0000usize & second) >> 8) |
(0x0000_00FFusize & second);
let second_word = ((0x00FF_0000usize & third) << 8) |
((0x0000_00FFusize & third) << 16) |
((0x00FF_0000usize & fourth) >> 8) |
(0x0000_00FFusize & fourth);
// Safety: fn invariant used here
*dst = word;
*(dst.add(1)) = second_word;
}
} else {
ascii_naive!(ascii_to_ascii, u8, u8);
ascii_naive!(ascii_to_basic_latin, u8, u16);
ascii_naive!(basic_latin_to_ascii, u16, u8);
}
}
cfg_if! {
// Safety-usable invariant: this counts the zeroes from the "first byte" of utf-8 data packed into a usize
// with the target endianness
if #[cfg(target_endian = "little")] {
#[allow(dead_code)]
#[inline(always)]
fn count_zeros(word: usize) -> u32 {
word.trailing_zeros()
}
} else {
#[allow(dead_code)]
#[inline(always)]
fn count_zeros(word: usize) -> u32 {
word.leading_zeros()
}
}
}
cfg_if! {
if #[cfg(all(feature = "simd-accel", target_endian = "little", target_arch = "disabled"))] {
/// Safety-usable invariant: Will return the value and position of the first non-ASCII byte in the slice in a Some if found.
/// In other words, the first element of the Some is always `> 127`
#[inline(always)]
pub fn validate_ascii(slice: &[u8]) -> Option<(u8, usize)> {
let src = slice.as_ptr();
let len = slice.len();
let mut offset = 0usize;
// Safety: if this check succeeds we're valid for reading/writing at least `stride` elements.
if SIMD_STRIDE_SIZE <= len {
let len_minus_stride = len - SIMD_STRIDE_SIZE;
loop {
// Safety: src at offset is valid for a `SIMD_STRIDE_SIZE` read
let simd = unsafe { load16_unaligned(src.add(offset)) };
if !simd_is_ascii(simd) {
break;
}
offset += SIMD_STRIDE_SIZE;
// This is `offset > len - SIMD_STRIDE_SIZE` which means we always have at least `SIMD_STRIDE_SIZE` elements to munch next time.
if offset > len_minus_stride {
break;
}
}
}
while offset < len {
let code_unit = slice[offset];
if code_unit > 127 {
// Safety: Safety-usable invariant upheld here
return Some((code_unit, offset));
}
offset += 1;
}
None
}
} else if #[cfg(all(feature = "simd-accel", target_feature = "sse2"))] {
/// Safety-usable invariant: will return Some() when it encounters non-ASCII, with the first element in the Some being
/// guaranteed to be non-ASCII (> 127), and the second being the offset where it is found
#[inline(always)]
pub fn validate_ascii(slice: &[u8]) -> Option<(u8, usize)> {
let src = slice.as_ptr();
let len = slice.len();
let mut offset = 0usize;
// Safety: if this check succeeds we're valid for reading at least `stride` elements.
if SIMD_STRIDE_SIZE <= len {
// First, process one unaligned vector
// Safety: src is valid for a `SIMD_STRIDE_SIZE` read
let simd = unsafe { load16_unaligned(src) };
let mask = mask_ascii(simd);
if mask != 0 {
offset = mask.trailing_zeros() as usize;
let non_ascii = unsafe { *src.add(offset) };
return Some((non_ascii, offset));
}
offset = SIMD_STRIDE_SIZE;
// Safety: Now that offset has changed we don't yet know how much it is valid for
// We have now seen 16 ASCII bytes. Let's guess that
// there will be enough more to justify more expense
// in the case of non-ASCII.
// Use aligned reads for the sake of old microachitectures.
// Safety: this correctly calculates the number of src_units that need to be read before the remaining list is aligned.
// This is by definition less than SIMD_ALIGNMENT, which is defined to be equal to SIMD_STRIDE_SIZE.
let until_alignment = unsafe { (SIMD_ALIGNMENT - ((src.add(offset) as usize) & SIMD_ALIGNMENT_MASK)) & SIMD_ALIGNMENT_MASK };
// This addition won't overflow, because even in the 32-bit PAE case the
// address space holds enough code that the slice length can't be that
// close to address space size.
// offset now equals SIMD_STRIDE_SIZE, hence times 3 below.
//
// Safety: if this check succeeds we're valid for reading at least `2 * SIMD_STRIDE_SIZE` elements plus `until_alignment`.
// The extra SIMD_STRIDE_SIZE in the condition is because `offset` is already `SIMD_STRIDE_SIZE`.
if until_alignment + (SIMD_STRIDE_SIZE * 3) <= len {
if until_alignment != 0 {
// Safety: this is safe to call since we're valid for this read (and more), and don't care about alignment
// This will copy over bytes that get decoded twice since it's not incrementing `offset` by SIMD_STRIDE_SIZE. This is fine.
let simd = unsafe { load16_unaligned(src.add(offset)) };
let mask = mask_ascii(simd);
if mask != 0 {
offset += mask.trailing_zeros() as usize;
let non_ascii = unsafe { *src.add(offset) };
return Some((non_ascii, offset));
}
offset += until_alignment;
}
// Safety: At this point we're valid for reading 2*SIMD_STRIDE_SIZE elements
// Safety: Now `offset` is aligned for `src`
let len_minus_stride_times_two = len - (SIMD_STRIDE_SIZE * 2);
loop {
// Safety: We were valid for this read, and were aligned.
let first = unsafe { load16_aligned(src.add(offset)) };
let second = unsafe { load16_aligned(src.add(offset + SIMD_STRIDE_SIZE)) };
if !simd_is_ascii(first | second) {
// Safety: mask_ascii produces a mask of all the high bits.
let mask_first = mask_ascii(first);
if mask_first != 0 {
// Safety: on little endian systems this will be the number of ascii bytes
// before the first non-ascii, i.e. valid for indexing src
// TODO SAFETY: What about big-endian systems?
offset += mask_first.trailing_zeros() as usize;
} else {
let mask_second = mask_ascii(second);
// Safety: on little endian systems this will be the number of ascii bytes
// before the first non-ascii, i.e. valid for indexing src
offset += SIMD_STRIDE_SIZE + mask_second.trailing_zeros() as usize;
}
// Safety: We know this is non-ASCII, and can uphold the safety-usable invariant here
let non_ascii = unsafe { *src.add(offset) };
return Some((non_ascii, offset));
}
offset += SIMD_STRIDE_SIZE * 2;
// Safety: This is `offset > len - 2 * SIMD_STRIDE_SIZE` which means we always have at least `2 * SIMD_STRIDE_SIZE` elements to munch next time.
if offset > len_minus_stride_times_two {
break;
}
}
// Safety: if this check succeeds we're valid for reading at least `SIMD_STRIDE_SIZE`
if offset + SIMD_STRIDE_SIZE <= len {
// Safety: We were valid for this read, and were aligned.
let simd = unsafe { load16_aligned(src.add(offset)) };
// Safety: mask_ascii produces a mask of all the high bits.
let mask = mask_ascii(simd);
if mask != 0 {
// Safety: on little endian systems this will be the number of ascii bytes
// before the first non-ascii, i.e. valid for indexing src
offset += mask.trailing_zeros() as usize;
let non_ascii = unsafe { *src.add(offset) };
// Safety: We know this is non-ASCII, and can uphold the safety-usable invariant here
return Some((non_ascii, offset));
}
offset += SIMD_STRIDE_SIZE;
}
} else {
// Safety: this is the unaligned branch
// At most two iterations, so unroll
// Safety: if this check succeeds we're valid for reading at least `SIMD_STRIDE_SIZE`
if offset + SIMD_STRIDE_SIZE <= len {
// Safety: We're valid for this read but must use an unaligned read
let simd = unsafe { load16_unaligned(src.add(offset)) };
let mask = mask_ascii(simd);
if mask != 0 {
offset += mask.trailing_zeros() as usize;
let non_ascii = unsafe { *src.add(offset) };
// Safety-usable invariant upheld here (same as above)
return Some((non_ascii, offset));
}
offset += SIMD_STRIDE_SIZE;
// Safety: if this check succeeds we're valid for reading at least `SIMD_STRIDE_SIZE`
if offset + SIMD_STRIDE_SIZE <= len {
// Safety: We're valid for this read but must use an unaligned read
let simd = unsafe { load16_unaligned(src.add(offset)) };
let mask = mask_ascii(simd);
if mask != 0 {
offset += mask.trailing_zeros() as usize;
let non_ascii = unsafe { *src.add(offset) };
// Safety-usable invariant upheld here (same as above)
return Some((non_ascii, offset));
}
offset += SIMD_STRIDE_SIZE;
}
}
}
}
while offset < len {
// Safety: relies straightforwardly on the `len` invariant
let code_unit = unsafe { *(src.add(offset)) };
if code_unit > 127 {
// Safety-usable invariant upheld here
return Some((code_unit, offset));
}
offset += 1;
}
None
}
} else {
// Safety-usable invariant: returns byte index of first non-ascii byte
#[inline(always)]
fn find_non_ascii(word: usize, second_word: usize) -> Option<usize> {
let word_masked = word & ASCII_MASK;
let second_masked = second_word & ASCII_MASK;
if (word_masked | second_masked) == 0 {
// Both are ascii, invariant upheld
return None;
}
if word_masked != 0 {
let zeros = count_zeros(word_masked);
// `zeros` now contains 0 to 7 (for the seven bits of masked ASCII in little endian,
// or up to 7 bits of non-ASCII in big endian if the first byte is non-ASCII)
// plus 8 times the number of ASCII in text order before the
// non-ASCII byte in the little-endian case or 8 times the number of ASCII in
// text order before the non-ASCII byte in the big-endian case.
let num_ascii = (zeros >> 3) as usize;
// Safety-usable invariant upheld here
return Some(num_ascii);
}
let zeros = count_zeros(second_masked);
// `zeros` now contains 0 to 7 (for the seven bits of masked ASCII in little endian,
// or up to 7 bits of non-ASCII in big endian if the first byte is non-ASCII)
// plus 8 times the number of ASCII in text order before the
// non-ASCII byte in the little-endian case or 8 times the number of ASCII in
// text order before the non-ASCII byte in the big-endian case.
let num_ascii = (zeros >> 3) as usize;
// Safety-usable invariant upheld here
Some(ALU_ALIGNMENT + num_ascii)
}
/// Safety: `src` must be valid for the reads of two `usize`s
///
/// Safety-usable invariant: will return byte index of first non-ascii byte
#[inline(always)]
unsafe fn validate_ascii_stride(src: *const usize) -> Option<usize> {
let word = *src;
let second_word = *(src.add(1));
find_non_ascii(word, second_word)
}
/// Safety-usable invariant: will return Some() when it encounters non-ASCII, with the first element in the Some being
/// guaranteed to be non-ASCII (> 127), and the second being the offset where it is found
#[cfg_attr(feature = "cargo-clippy", allow(cast_ptr_alignment))]
#[inline(always)]
pub fn validate_ascii(slice: &[u8]) -> Option<(u8, usize)> {
let src = slice.as_ptr();
let len = slice.len();
let mut offset = 0usize;
let mut until_alignment = (ALU_ALIGNMENT - ((src as usize) & ALU_ALIGNMENT_MASK)) & ALU_ALIGNMENT_MASK;
// Safety: If this check fails we're valid to read `until_alignment + ALU_STRIDE_SIZE` elements
if until_alignment + ALU_STRIDE_SIZE <= len {
while until_alignment != 0 {
let code_unit = slice[offset];
if code_unit > 127 {
// Safety-usable invairant upheld here
return Some((code_unit, offset));
}
offset += 1;
until_alignment -= 1;
}
// Safety: At this point we have read until_alignment elements and
// are valid for `ALU_STRIDE_SIZE` more.
let len_minus_stride = len - ALU_STRIDE_SIZE;
loop {
// Safety: we were valid for this read
let ptr = unsafe { src.add(offset) as *const usize };
if let Some(num_ascii) = unsafe { validate_ascii_stride(ptr) } {
offset += num_ascii;
// Safety-usable invairant upheld here using the invariant from validate_ascii_stride()
return Some((unsafe { *(src.add(offset)) }, offset));
}
offset += ALU_STRIDE_SIZE;
// Safety: This is `offset > ALU_STRIDE_SIZE` which means we always have at least `2 * ALU_STRIDE_SIZE` elements to munch next time.
if offset > len_minus_stride {
break;
}
}
}
while offset < len {
let code_unit = slice[offset];
if code_unit > 127 {
// Safety-usable invairant upheld here
return Some((code_unit, offset));
}
offset += 1;
}
None
}
}
}
cfg_if! {
if #[cfg(all(feature = "simd-accel", any(target_feature = "sse2", all(target_endian = "little", target_arch = "aarch64"))))] {
} else if #[cfg(all(feature = "simd-accel", target_endian = "little", target_feature = "neon"))] {
// Even with NEON enabled, we use the ALU path for ASCII validation, because testing
// on Exynos 5 indicated that using NEON isn't worthwhile where there are only
// vector reads without vector writes.
pub const ALU_STRIDE_SIZE: usize = 8;
pub const ALU_ALIGNMENT: usize = 4;
pub const ALU_ALIGNMENT_MASK: usize = 3;
} else {
// Safety: src points to two valid `usize`s, dst points to four valid `usize`s
#[inline(always)]
unsafe fn unpack_latin1_stride_alu(src: *const usize, dst: *mut usize) {
// Safety: src safety invariant used here
let word = *src;
let second_word = *(src.add(1));
// Safety: dst safety invariant passed down
unpack_alu(word, second_word, dst);
}
// Safety: src points to four valid `usize`s, dst points to two valid `usize`s
#[inline(always)]
unsafe fn pack_latin1_stride_alu(src: *const usize, dst: *mut usize) {
// Safety: src safety invariant used here
let first = *src;
let second = *(src.add(1));
let third = *(src.add(2));
let fourth = *(src.add(3));
// Safety: dst safety invariant passed down
pack_alu(first, second, third, fourth, dst);
}
// Safety: src points to two valid `usize`s, dst points to four valid `usize`s
#[inline(always)]
unsafe fn ascii_to_basic_latin_stride_alu(src: *const usize, dst: *mut usize) -> bool {
// Safety: src safety invariant used here
let word = *src;
let second_word = *(src.add(1));
// Check if the words contains non-ASCII
if (word & ASCII_MASK) | (second_word & ASCII_MASK) != 0 {
return false;
}
// Safety: dst safety invariant passed down
unpack_alu(word, second_word, dst);
true
}
// Safety: src points four valid `usize`s, dst points to two valid `usize`s
#[inline(always)]
unsafe fn basic_latin_to_ascii_stride_alu(src: *const usize, dst: *mut usize) -> bool {
// Safety: src safety invariant used here
let first = *src;
let second = *(src.add(1));
let third = *(src.add(2));
let fourth = *(src.add(3));
if (first & BASIC_LATIN_MASK) | (second & BASIC_LATIN_MASK) | (third & BASIC_LATIN_MASK) | (fourth & BASIC_LATIN_MASK) != 0 {
return false;
}
// Safety: dst safety invariant passed down
pack_alu(first, second, third, fourth, dst);
true
}
// Safety: src, dst both point to two valid `usize`s each
// Safety-usable invariant: Will return byte index of first non-ascii byte.
#[inline(always)]
unsafe fn ascii_to_ascii_stride(src: *const usize, dst: *mut usize) -> Option<usize> {
// Safety: src safety invariant used here
let word = *src;
let second_word = *(src.add(1));
// Safety: src safety invariant used here
*dst = word;
*(dst.add(1)) = second_word;
// Relies on safety-usable invariant here
find_non_ascii(word, second_word)
}
basic_latin_alu!(ascii_to_basic_latin, u8, u16, ascii_to_basic_latin_stride_alu);
basic_latin_alu!(basic_latin_to_ascii, u16, u8, basic_latin_to_ascii_stride_alu);
latin1_alu!(unpack_latin1, u8, u16, unpack_latin1_stride_alu);
latin1_alu!(pack_latin1, u16, u8, pack_latin1_stride_alu);
// Safety invariant upheld: ascii_to_ascii_stride will return byte index of first non-ascii if found
ascii_alu!(ascii_to_ascii, u8, u8, ascii_to_ascii_stride);
}
}
pub fn ascii_valid_up_to(bytes: &[u8]) -> usize {
match validate_ascii(bytes) {
None => bytes.len(),
Some((_, num_valid)) => num_valid,
}
}
pub fn iso_2022_jp_ascii_valid_up_to(bytes: &[u8]) -> usize {
for (i, b_ref) in bytes.iter().enumerate() {
let b = *b_ref;
if b >= 0x80 || b == 0x1B || b == 0x0E || b == 0x0F {
return i;
}
}
bytes.len()
}
// Any copyright to the test code below this comment is dedicated to the
#[cfg(all(test, feature = "alloc"))]
mod tests {
use super::*;
use alloc::vec::Vec;
macro_rules! test_ascii {
($test_name:ident, $fn_tested:ident, $src_unit:ty, $dst_unit:ty) => {
#[test]
fn $test_name() {
let mut src: Vec<$src_unit> = Vec::with_capacity(32);
let mut dst: Vec<$dst_unit> = Vec::with_capacity(32);
for i in 0..32 {
src.clear();
dst.clear();
dst.resize(32, 0);
for j in 0..32 {
let c = if i == j { 0xAA } else { j + 0x40 };
src.push(c as $src_unit);
}
match unsafe { $fn_tested(src.as_ptr(), dst.as_mut_ptr(), 32) } {
None => unreachable!("Should always find non-ASCII"),
Some((non_ascii, num_ascii)) => {
assert_eq!(non_ascii, 0xAA);
assert_eq!(num_ascii, i);
for j in 0..i {
assert_eq!(dst[j], (j + 0x40) as $dst_unit);
}
}
}
}
}
};
}
test_ascii!(test_ascii_to_ascii, ascii_to_ascii, u8, u8);
test_ascii!(test_ascii_to_basic_latin, ascii_to_basic_latin, u8, u16);
test_ascii!(test_basic_latin_to_ascii, basic_latin_to_ascii, u16, u8);
}