use super::{MMB16Dbl, MMB32Dbl};
use crate::mem as basic;
use crate::utils::*;
#[cfg(target_arch = "x86")]
use core::arch::x86 as mmx;
#[cfg(target_arch = "x86_64")]
use core::arch::x86_64 as mmx;
use mmx::__m128i;
use mmx::_mm_cmpeq_epi8;
use mmx::_mm_load_si128;
use mmx::_mm_loadu_si128;
use mmx::_mm_movemask_epi8;
use mmx::_mm_or_si128;
use mmx::__m256i;
use mmx::_mm256_cmpeq_epi8;
use mmx::_mm256_load_si256;
use mmx::_mm256_loadu_si256;
use mmx::_mm256_movemask_epi8;
use mmx::_mm256_or_si256;
#[cfg(any(target_arch = "x86_64", target_arch = "x86"))]
use super::cpuid;
use core::sync::atomic::AtomicPtr;
use core::sync::atomic::Ordering;
type FuncType = fn(&[u8], B1Dbl) -> Option<usize>;
const FUNC: FuncType = fnptr_setup_func;
static FUNC_PTR_ATOM: AtomicPtr<FuncType> = AtomicPtr::new(FUNC as *mut FuncType);
#[inline(never)]
fn fnptr_setup_func(buf: &[u8], needle: B1Dbl) -> Option<usize> {
#[cfg(target_arch = "x86_64")]
let func = if cpuid::has_avx2() {
_memnechr_dbl_avx2
} else {
_memnechr_dbl_sse2
};
#[cfg(target_arch = "x86")]
let func = if cpuid::has_avx2() {
_memnechr_dbl_avx2
} else if cpuid::has_sse2() {
_memnechr_dbl_sse2
} else {
_memnechr_dbl_basic
};
FUNC_PTR_ATOM.store(func as *mut FuncType, Ordering::Relaxed);
unsafe { func(buf, needle) }
}
#[inline(always)]
pub(crate) fn _memnechr_dbl_impl(buf: &[u8], needle: B1Dbl) -> Option<usize> {
let func_u = FUNC_PTR_ATOM.load(Ordering::Relaxed);
#[allow(clippy::crosspointer_transmute)]
let func: FuncType = unsafe { core::mem::transmute(func_u) };
func(buf, needle)
}
unsafe fn _memnechr_dbl_basic(buf: &[u8], needle: B1Dbl) -> Option<usize> {
basic::_memnechr_dbl_impl(buf, needle)
}
#[cfg(any(target_arch = "x86_64", target_arch = "x86"))]
#[target_feature(enable = "sse2")]
#[allow(clippy::missing_safety_doc)]
pub unsafe fn _memnechr_dbl_sse2(buf: &[u8], needle: B1Dbl) -> Option<usize> {
_memnechr_dbl_sse2_impl(buf, needle)
}
#[cfg(any(target_arch = "x86_64", target_arch = "x86"))]
#[target_feature(enable = "avx2")]
#[allow(clippy::missing_safety_doc)]
pub unsafe fn _memnechr_dbl_avx2(buf: &[u8], needle: B1Dbl) -> Option<usize> {
_memnechr_dbl_avx2_impl(buf, needle)
}
#[inline(always)]
fn _memnechr_dbl_sse2_impl(buf: &[u8], needle: B1Dbl) -> Option<usize> {
let buf_len = buf.len();
let mut buf_ptr = buf.as_ptr();
let start_ptr = buf_ptr;
let end_ptr = unsafe { buf_ptr.add(buf_len) };
buf_ptr.prefetch_read_data();
if buf_len >= 16 {
let cc: MMB16Dbl = needle.into();
{
if !buf_ptr.is_aligned_u128() {
#[cfg(not(feature = "test_alignment_check"))]
{
let r = unsafe { _nechr_c16_uu_x1(buf_ptr, cc, start_ptr) };
if r.is_some() {
return r;
}
let remaining_align = 0x10_usize - ((buf_ptr as usize) & 0x0F_usize);
buf_ptr = unsafe { buf_ptr.add(remaining_align) };
}
#[cfg(feature = "test_alignment_check")]
{
let r = basic::_nechr_dbl_to_aligned_u128(buf_ptr, needle, start_ptr);
if let Some(p) = r.0 {
buf_ptr = p;
} else if let Some(v) = r.1 {
return Some(v);
}
}
}
}
{
let (r, p) = _unroll_loop_with_prefetch::<4, 16, _>(buf_ptr, end_ptr, |p| unsafe {
_nechr_c16_aa_x1(p, cc, start_ptr)
});
if r.is_some() {
return r;
}
buf_ptr = p;
}
{
let (r, p) = _unroll_loop::<1, 16, _>(buf_ptr, end_ptr, |p| unsafe {
_nechr_c16_aa_x1(p, cc, start_ptr)
});
if r.is_some() {
return r;
}
buf_ptr = p;
}
}
let cc: B8Dbl = needle.into();
basic::_memnechr_dbl_remaining_15_bytes_impl(buf_ptr, cc, start_ptr, end_ptr)
}
#[inline(always)]
fn _memnechr_dbl_avx2_impl(buf: &[u8], needle: B1Dbl) -> Option<usize> {
let buf_len = buf.len();
let mut buf_ptr = buf.as_ptr();
let start_ptr = buf_ptr;
let end_ptr = unsafe { buf_ptr.add(buf_len) };
buf_ptr.prefetch_read_data();
if buf_len >= 32 {
let cc: MMB32Dbl = needle.into();
{
if !buf_ptr.is_aligned_u256() {
#[cfg(not(feature = "test_alignment_check"))]
{
let r = unsafe { _nechr_c32_uu_x1(buf_ptr, cc, start_ptr) };
if r.is_some() {
return r;
}
let remaining_align = 0x20_usize - ((buf_ptr as usize) & 0x1F_usize);
buf_ptr = unsafe { buf_ptr.add(remaining_align) };
}
#[cfg(feature = "test_alignment_check")]
{
let r = basic::_nechr_dbl_to_aligned_u256(buf_ptr, needle, start_ptr);
if let Some(p) = r.0 {
buf_ptr = p;
} else if let Some(v) = r.1 {
return Some(v);
}
}
}
}
{
let (r, p) = _unroll_loop::<1, 32, _>(buf_ptr, end_ptr, |p| unsafe {
_nechr_c32_aa_x1(p, cc, start_ptr)
});
if r.is_some() {
return r;
}
buf_ptr = p;
}
{
let cc: MMB16Dbl = needle.into();
let (r, p) = _unroll_loop::<1, 16, _>(buf_ptr, end_ptr, |p| unsafe {
_nechr_c16_aa_x1(p, cc, start_ptr)
});
if r.is_some() {
return r;
}
buf_ptr = p;
}
} else if buf_len >= 16 {
{
let cc: MMB16Dbl = needle.into();
if buf_ptr.is_aligned_u128() {
let (r, p) = _unroll_loop::<1, 16, _>(buf_ptr, end_ptr, |p| unsafe {
_nechr_c16_aa_x1(p, cc, start_ptr)
});
if r.is_some() {
return r;
}
buf_ptr = p;
} else {
#[cfg(not(feature = "test_alignment_check"))]
{
let (r, p) = _unroll_loop::<1, 16, _>(buf_ptr, end_ptr, |p| unsafe {
_nechr_c16_uu_x1(p, cc, start_ptr)
});
if r.is_some() {
return r;
}
buf_ptr = p;
}
#[cfg(feature = "test_alignment_check")]
{
let r = basic::_nechr_dbl_to_aligned_u128(buf_ptr, needle, start_ptr);
if let Some(p) = r.0 {
buf_ptr = p;
} else if let Some(v) = r.1 {
return Some(v);
}
let (r, p) = _unroll_loop::<1, 16, _>(buf_ptr, end_ptr, |p| unsafe {
_nechr_c16_aa_x1(p, cc, start_ptr)
});
if r.is_some() {
return r;
}
buf_ptr = p;
}
}
}
}
let cc: B8Dbl = needle.into();
basic::_memnechr_dbl_remaining_15_bytes_impl(buf_ptr, cc, start_ptr, end_ptr)
}
#[inline(always)]
fn _return_nechr_dbl<T, PU>(base: T, mask_a: PU) -> Option<usize>
where
T: core::ops::Add<usize, Output = usize>,
PU: BitOrt,
{
if !mask_a.is_highs() {
let idx1 = mask_a.trailing_ones() as usize;
Some(base + idx1)
} else {
None
}
}
#[inline(always)]
unsafe fn _nechr_c16_uu_x1(
buf_ptr: *const u8,
mm_c16: MMB16Dbl,
st_ptr: *const u8,
) -> Option<usize> {
let mask_0_ab = unsafe {
let mm_0 = _mm_loadu_si128(buf_ptr as *const __m128i);
let mm_0_eq_a = _mm_cmpeq_epi8(mm_0, mm_c16.v1);
let mm_0_eq_b = _mm_cmpeq_epi8(mm_0, mm_c16.v2);
let mm_0_eq_ab = _mm_or_si128(mm_0_eq_a, mm_0_eq_b);
_mm_movemask_epi8(mm_0_eq_ab) as u16
};
let base = buf_ptr.usz_offset_from(st_ptr);
_return_nechr_dbl(base, mask_0_ab)
}
#[inline(always)]
unsafe fn _nechr_c16_aa_x1(
buf_ptr: *const u8,
mm_c16: MMB16Dbl,
st_ptr: *const u8,
) -> Option<usize> {
let mask_0_ab = unsafe {
let mm_0 = _mm_load_si128(buf_ptr as *const __m128i);
let mm_0_eq_a = _mm_cmpeq_epi8(mm_0, mm_c16.v1);
let mm_0_eq_b = _mm_cmpeq_epi8(mm_0, mm_c16.v2);
let mm_0_eq_ab = _mm_or_si128(mm_0_eq_a, mm_0_eq_b);
_mm_movemask_epi8(mm_0_eq_ab) as u16
};
let base = buf_ptr.usz_offset_from(st_ptr);
_return_nechr_dbl(base, mask_0_ab)
}
#[inline(always)]
unsafe fn _nechr_c32_uu_x1(
buf_ptr: *const u8,
mm_c32: MMB32Dbl,
st_ptr: *const u8,
) -> Option<usize> {
let mask_0_ab = unsafe {
let mm_0 = _mm256_loadu_si256(buf_ptr as *const __m256i);
let mm_0_eq_a = _mm256_cmpeq_epi8(mm_0, mm_c32.v1);
let mm_0_eq_b = _mm256_cmpeq_epi8(mm_0, mm_c32.v2);
let mm_0_eq_ab = _mm256_or_si256(mm_0_eq_a, mm_0_eq_b);
_mm256_movemask_epi8(mm_0_eq_ab) as u32
};
let base = buf_ptr.usz_offset_from(st_ptr);
_return_nechr_dbl(base, mask_0_ab)
}
#[inline(always)]
unsafe fn _nechr_c32_aa_x1(
buf_ptr: *const u8,
mm_c32: MMB32Dbl,
st_ptr: *const u8,
) -> Option<usize> {
let mask_0_ab = unsafe {
let mm_0 = _mm256_load_si256(buf_ptr as *const __m256i);
let mm_0_eq_a = _mm256_cmpeq_epi8(mm_0, mm_c32.v1);
let mm_0_eq_b = _mm256_cmpeq_epi8(mm_0, mm_c32.v2);
let mm_0_eq_ab = _mm256_or_si256(mm_0_eq_a, mm_0_eq_b);
_mm256_movemask_epi8(mm_0_eq_ab) as u32
};
let base = buf_ptr.usz_offset_from(st_ptr);
_return_nechr_dbl(base, mask_0_ab)
}