Skip to content
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.

Sign up for GitHub

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

Jump to bottom

cleanup: SIMD runtime detection #132

Merged
merged 1 commit into from
Apr 20, 2023
Merged

cleanup: SIMD runtime detection #132

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
50 changes: 22 additions & 28 deletions src/simd/avx2.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,29 +1,24 @@
use crate::iter::Bytes;

pub enum Scan {
/// Returned when an implementation finds a noteworthy token.
Found,
/// Returned when an implementation couldn't keep running because the input was too short.
TooShort,
}

#[cfg(target_arch = "x86")]
AaronO marked this conversation as resolved.
Show resolved Hide resolved
unsafe fn parse_uri_batch_32(_: &[u8]) -> usize {
pub unsafe fn match_uri_vectored(_: &mut Bytes) {
unreachable!("AVX2 detection should be disabled for x86");
}

#[inline]
#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2")]
pub unsafe fn parse_uri_batch_32(bytes: &mut Bytes) -> Scan {
#[target_feature(enable = "avx2", enable = "sse4.2")]
pub unsafe fn match_uri_vectored(bytes: &mut Bytes) {
while bytes.as_ref().len() >= 32 {
let advance = match_url_char_32_avx(bytes.as_ref());
bytes.advance(advance);

if advance != 32 {
return Scan::Found;
return;
}
}
Scan::TooShort
// do both, since avx2 only works when bytes.len() >= 32
super::sse42::match_uri_vectored(bytes)
}

#[inline(always)]
Expand Down Expand Up @@ -69,22 +64,23 @@ unsafe fn match_url_char_32_avx(buf: &[u8]) -> usize {
}

#[cfg(target_arch = "x86")]
unsafe fn match_header_value_batch_32(_: &[u8]) -> usize {
pub unsafe fn match_header_value_vectored(_: &mut Bytes) {
unreachable!("AVX2 detection should be disabled for x86");
}

#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2")]
pub unsafe fn match_header_value_batch_32(bytes: &mut Bytes) -> Scan {
#[target_feature(enable = "avx2", enable = "sse4.2")]
pub unsafe fn match_header_value_vectored(bytes: &mut Bytes) {
while bytes.as_ref().len() >= 32 {
let advance = match_header_value_char_32_avx(bytes.as_ref());
bytes.advance(advance);

if advance != 32 {
return Scan::Found;
return;
}
}
Scan::TooShort
// do both, since avx2 only works when bytes.len() >= 32
super::sse42::match_header_value_vectored(bytes)
}

#[inline(always)]
Expand Down Expand Up @@ -120,17 +116,16 @@ unsafe fn match_header_value_char_32_avx(buf: &[u8]) -> usize {

#[test]
fn avx2_code_matches_uri_chars_table() {
match super::detect() {
super::AVX_2 | super::AVX_2_AND_SSE_42 => {},
_ => return,
if !is_x86_feature_detected!("avx2") {
return;
}

unsafe {
assert!(byte_is_allowed(b'_', parse_uri_batch_32));
assert!(byte_is_allowed(b'_', match_uri_vectored));

for (b, allowed) in crate::URI_MAP.iter().cloned().enumerate() {
assert_eq!(
byte_is_allowed(b as u8, parse_uri_batch_32), allowed,
byte_is_allowed(b as u8, match_uri_vectored), allowed,
"byte_is_allowed({:?}) should be {:?}", b, allowed,
);
}
Expand All @@ -139,25 +134,24 @@ fn avx2_code_matches_uri_chars_table() {

#[test]
fn avx2_code_matches_header_value_chars_table() {
match super::detect() {
super::AVX_2 | super::AVX_2_AND_SSE_42 => {},
_ => return,
if !is_x86_feature_detected!("avx2") {
return;
}

unsafe {
assert!(byte_is_allowed(b'_', match_header_value_batch_32));
assert!(byte_is_allowed(b'_', match_header_value_vectored));

for (b, allowed) in crate::HEADER_VALUE_MAP.iter().cloned().enumerate() {
assert_eq!(
byte_is_allowed(b as u8, match_header_value_batch_32), allowed,
byte_is_allowed(b as u8, match_header_value_vectored), allowed,
"byte_is_allowed({:?}) should be {:?}", b, allowed,
);
}
}
}

#[cfg(test)]
unsafe fn byte_is_allowed(byte: u8, f: unsafe fn(bytes: &mut Bytes<'_>) -> Scan) -> bool {
unsafe fn byte_is_allowed(byte: u8, f: unsafe fn(bytes: &mut Bytes<'_>)) -> bool {
let slice = [
b'_', b'_', b'_', b'_',
b'_', b'_', b'_', b'_',
Expand Down
198 changes: 19 additions & 179 deletions src/simd/mod.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,45 +38,6 @@ mod sse42;
))]
mod avx2;

#[cfg(all(
httparse_simd,
any(
target_arch = "x86",
target_arch = "x86_64",
),
))]
pub const SSE_42: usize = 1;
#[cfg(all(
httparse_simd,
any(not(httparse_simd_target_feature_sse42), httparse_simd_target_feature_avx2),
any(
target_arch = "x86",
target_arch = "x86_64",
),
))]
pub const AVX_2: usize = 2;
#[cfg(all(
httparse_simd,
any(
not(httparse_simd_target_feature_sse42),
httparse_simd_target_feature_avx2,
test,
),
any(
target_arch = "x86",
target_arch = "x86_64",
),
))]
pub const AVX_2_AND_SSE_42: usize = 3;

#[cfg(all(
httparse_simd,
any(
target_arch = "x86",
target_arch = "x86_64",
),
))]
const NONE: usize = std::usize::MAX;
#[cfg(all(
httparse_simd,
not(any(
Expand All @@ -88,77 +49,7 @@ const NONE: usize = std::usize::MAX;
target_arch = "x86_64",
),
))]
mod runtime {
//! Runtime detection of simd features. Used when the build script
//! doesn't notice any target features at build time.
//!
//! While `is_x86_feature_detected!` has it's own caching built-in,
//! at least in 1.27.0, the functions don't inline, leaving using it
//! actually *slower* than just using the scalar fallback.

use core::sync::atomic::{AtomicUsize, Ordering};

static FEATURE: AtomicUsize = AtomicUsize::new(0);

const INIT: usize = 0;

pub fn detect() -> usize {
let feat = FEATURE.load(Ordering::Relaxed);
if feat == INIT {
if cfg!(target_arch = "x86_64") && is_x86_feature_detected!("avx2") {
if is_x86_feature_detected!("sse4.2") {
FEATURE.store(super::AVX_2_AND_SSE_42, Ordering::Relaxed);
return super::AVX_2_AND_SSE_42;
} else {
FEATURE.store(super::AVX_2, Ordering::Relaxed);
return super::AVX_2;
}
} else if is_x86_feature_detected!("sse4.2") {
FEATURE.store(super::SSE_42, Ordering::Relaxed);
return super::SSE_42;
} else {
FEATURE.store(super::NONE, Ordering::Relaxed);
}
}
feat
}

pub fn match_uri_vectored(bytes: &mut crate::iter::Bytes) {
unsafe {
match detect() {
super::SSE_42 => super::sse42::parse_uri_batch_16(bytes),
super::AVX_2 => { super::avx2::parse_uri_batch_32(bytes); },
super::AVX_2_AND_SSE_42 => {
if let super::avx2::Scan::Found = super::avx2::parse_uri_batch_32(bytes) {
return;
}
super::sse42::parse_uri_batch_16(bytes)
},
_ => ()
}
}

// else do nothing
}

pub fn match_header_value_vectored(bytes: &mut crate::iter::Bytes) {
unsafe {
match detect() {
super::SSE_42 => super::sse42::match_header_value_batch_16(bytes),
super::AVX_2 => { super::avx2::match_header_value_batch_32(bytes); },
super::AVX_2_AND_SSE_42 => {
if let super::avx2::Scan::Found = super::avx2::match_header_value_batch_32(bytes) {
return;
}
super::sse42::match_header_value_batch_16(bytes)
},
_ => ()
}
}

// else do nothing
}
}
mod runtime;

#[cfg(all(
httparse_simd,
Expand All @@ -183,32 +74,16 @@ pub use self::runtime::*;
),
))]
mod sse42_compile_time {
pub fn match_uri_vectored(bytes: &mut crate::iter::Bytes) {
if detect() == super::SSE_42 {
unsafe {
super::sse42::parse_uri_batch_16(bytes);
}
}

// else do nothing
#[inline(always)]
pub fn match_uri_vectored(b: &mut crate::iter::Bytes<'_>) {
// SAFETY: calls are guarded by a compile time feature check
unsafe { crate::simd::sse42::match_uri_vectored(b) }
}

pub fn match_header_value_vectored(bytes: &mut crate::iter::Bytes) {
if detect() == super::SSE_42 {
unsafe {
super::sse42::match_header_value_batch_16(bytes);
}
}

// else do nothing
}

pub fn detect() -> usize {
if is_x86_feature_detected!("sse4.2") {
super::SSE_42
} else {
super::NONE
}

#[inline(always)]
pub fn match_header_value_vectored(b: &mut crate::iter::Bytes<'_>) {
// SAFETY: calls are guarded by a compile time feature check
unsafe { crate::simd::sse42::match_header_value_vectored(b) }
}
}

Expand All @@ -232,51 +107,16 @@ pub use self::sse42_compile_time::*;
),
))]
mod avx2_compile_time {
pub fn match_uri_vectored(bytes: &mut crate::iter::Bytes) {
// do both, since avx2 only works when bytes.len() >= 32
if detect() == super::AVX_2_AND_SSE_42 {
unsafe {
super::avx2::parse_uri_batch_32(bytes);
}

}
if detect() == super::SSE_42 {
unsafe {
super::sse42::parse_uri_batch_16(bytes);
}
}

// else do nothing
#[inline(always)]
pub fn match_uri_vectored(b: &mut crate::iter::Bytes<'_>) {
// SAFETY: calls are guarded by a compile time feature check
unsafe { crate::simd::avx2::match_uri_vectored(b) }
}

pub fn match_header_value_vectored(bytes: &mut crate::iter::Bytes) {
// do both, since avx2 only works when bytes.len() >= 32
if detect() == super::AVX_2_AND_SSE_42 {
let scanned = unsafe {
super::avx2::match_header_value_batch_32(bytes)
};

if let super::avx2::Scan::Found = scanned {
return;
}
}
if detect() == super::SSE_42 {
unsafe {
super::sse42::match_header_value_batch_16(bytes);
}
}

// else do nothing
}

pub fn detect() -> usize {
if cfg!(target_arch = "x86_64") && is_x86_feature_detected!("avx2") {
super::AVX_2_AND_SSE_42
} else if is_x86_feature_detected!("sse4.2") {
super::SSE_42
} else {
super::NONE
}

#[inline(always)]
pub fn match_header_value_vectored(b: &mut crate::iter::Bytes<'_>) {
// SAFETY: calls are guarded by a compile time feature check
unsafe { crate::simd::avx2::match_header_value_vectored(b) }
}
}

Expand Down
Loading