Merge pull request #6 from CleverCloud/arm_support

Move simd under a feature flag, use fast CharLookup if not enabled
CleverCloud · Aug 16, 2023 · 3f3e116 · 3f3e116
2 parents 68c22f4 + b74e75c
commit 3f3e116
Show file tree

Hide file tree

Showing 3 changed files with 86 additions and 15 deletions.
diff --git a/Cargo.toml b/Cargo.toml
@@ -14,9 +14,10 @@ version = "0.6.2"
 nom = "7.1.3"
 
 [features]
-# default = ["custom-vecdeque"]
+default = ["simd"]
 rc-alloc = []
 custom-vecdeque = []
+simd = []
 
 [profile.release]
 lto = true

diff --git a/src/protocol/h1/parser/primitives.rs b/src/protocol/h1/parser/primitives.rs
@@ -205,9 +205,9 @@ fn http_status(i: &[u8]) -> IResult<&[u8], (&[u8], u16)> {
 #[inline]
 #[allow(clippy::type_complexity)]
 pub fn parse_request_line(i: &[u8]) -> IResult<&[u8], (&[u8], &[u8], Version)> {
-    let (i, method) = tchar::take_while_simd(i)?;
+    let (i, method) = tchar::take_while_fast(i)?;
     let (i, _) = space(i)?;
-    let (i, uri) = vchar::take_while_simd(i)?;
+    let (i, uri) = vchar::take_while_fast(i)?;
     let (i, _) = space(i)?;
     let (i, version) = http_version(i)?;
     let (i, _) = crlf(i)?;
@@ -224,7 +224,7 @@ pub fn parse_response_line(i: &[u8]) -> IResult<&[u8], (Version, &[u8], u16, &[u
     let (i, _) = space(i)?;
     let (i, (status, code)) = http_status(i)?;
     let (i, _) = space(i)?;
-    let (i, reason) = achar::take_while_simd(i)?;
+    let (i, reason) = achar::take_while_fast(i)?;
     let (i, _) = crlf(i)?;
     Ok((i, (version, status, code, reason)))
 }
@@ -236,13 +236,13 @@ pub fn parse_response_line(i: &[u8]) -> IResult<&[u8], (Version, &[u8], u16, &[u
 #[inline]
 #[allow(clippy::type_complexity)]
 pub fn parse_header_or_cookie(i: &[u8]) -> IResult<&[u8], Option<(&[u8], &[u8])>> {
-    let (i, key) = tchar::take_while_simd(i)?;
+    let (i, key) = tchar::take_while_fast(i)?;
     let (i, _) = tag(b":")(i)?;
     let (i, _) = take_while(is_space)(i)?;
     if compare_no_case(key, b"cookie") {
         return Ok((i, None));
     }
-    let (i, val) = achar::take_while_simd(i)?;
+    let (i, val) = achar::take_while_fast(i)?;
     let (i, _) = crlf(i)?;
     Ok((i, Some((key, val))))
 }
@@ -253,10 +253,10 @@ pub fn parse_header_or_cookie(i: &[u8]) -> IResult<&[u8], Option<(&[u8], &[u8])>
 /// example: `Content-Length: 42\r\n`
 #[inline]
 pub fn parse_header(i: &[u8]) -> IResult<&[u8], (&[u8], &[u8])> {
-    let (i, key) = tchar::take_while_simd(i)?;
+    let (i, key) = tchar::take_while_fast(i)?;
     let (i, _) = tag(b":")(i)?;
     let (i, _) = take_while(is_space)(i)?;
-    let (i, val) = achar::take_while_simd(i)?;
+    let (i, val) = achar::take_while_fast(i)?;
     let (i, _) = crlf(i)?;
     Ok((i, (key, val)))
 }
@@ -277,8 +277,8 @@ pub fn parse_single_crumb(i: &[u8], first: bool) -> IResult<&[u8], (&[u8], &[u8]
     } else {
         i
     };
-    let (i, key) = ck_char::take_while_simd(i)?;
-    let (i, val) = opt(tuple((tag(b"="), cv_char::take_while_simd)))(i)?;
+    let (i, key) = ck_char::take_while_fast(i)?;
+    let (i, val) = opt(tuple((tag(b"="), cv_char::take_while_fast)))(i)?;
 
     match val {
         Some((_, val)) => Ok((i, (key, val))),

diff --git a/src/protocol/utils.rs b/src/protocol/utils.rs
@@ -103,8 +103,10 @@ macro_rules! compile_lookup {
         mod $name {
             use $crate::h1::parser::primitives::{CharLookup, CharRanges, CharTable};
             pub const LOOKUP: CharLookup = $crate::make_char_lookup!($($t)*);
-            pub const RANGES: CharRanges = LOOKUP.ranges;
             pub const TABLE: CharTable = LOOKUP.table;
+            #[allow(dead_code)]
+            pub const RANGES: CharRanges = LOOKUP.ranges;
+            #[allow(dead_code)]
             pub const LENGTH: i32 = LOOKUP.len;
 
             #[inline]
@@ -115,11 +117,11 @@ macro_rules! compile_lookup {
             }
 
             #[inline]
-            #[allow(dead_code)]
+            #[cfg(feature="simd")]
             /// Returns the longest string that fits the rule (simd optimized)
             ///
             /// *Streaming version* will return a Err::Incomplete(Needed::Unknown) if the pattern reaches the end of the input.
-            pub fn take_while_simd(input: &[u8]) -> nom::IResult<&[u8], &[u8]> {
+            fn take_while_simd(input: &[u8]) -> nom::IResult<&[u8], &[u8]> {
                 use std::arch::x86_64::{
                     _mm_cmpestri, _mm_lddqu_si128, _mm_loadu_si128, _SIDD_CMP_RANGES,
                     _SIDD_LEAST_SIGNIFICANT, _SIDD_UBYTE_OPS,
@@ -173,9 +175,9 @@ macro_rules! compile_lookup {
             }
 
             #[inline]
-            #[allow(dead_code)]
+            #[cfg(feature="simd")]
             /// Returns the longest string that fits the rule (simd optimized)
-            pub fn take_while_complete_simd(input: &[u8]) -> nom::IResult<&[u8], &[u8]> {
+            fn take_while_complete_simd(input: &[u8]) -> nom::IResult<&[u8], &[u8]> {
                 use std::arch::x86_64::{
                     _mm_cmpestri, _mm_lddqu_si128, _mm_loadu_si128, _SIDD_CMP_RANGES,
                     _SIDD_LEAST_SIGNIFICANT, _SIDD_UBYTE_OPS,
@@ -223,6 +225,74 @@ macro_rules! compile_lookup {
                     ))
                 }
             }
+
+            #[inline]
+            #[allow(dead_code)]
+            /// Returns the longest string that fits the rule (not simd optimized)
+            ///
+            /// *Streaming version* will return a Err::Incomplete(Needed::Unknown) if the pattern reaches the end of the input.
+            pub fn take_while(input: &[u8]) -> nom::IResult<&[u8], &[u8]> {
+                let mut i = 0;
+                while i < input.len() {
+                    if unsafe { !TABLE.get_unchecked(*input.get_unchecked(i) as usize) } {
+                        break;
+                    }
+                    i += 1;
+                }
+                if i == input.len() {
+                    return Err(nom::Err::Incomplete(nom::Needed::Unknown));
+                } else {
+                    unsafe {
+                        Ok((
+                            input.get_unchecked(i..),
+                            input.get_unchecked(..i),
+                        ))
+                    }
+                }
+            }
+
+            #[inline]
+            #[allow(dead_code)]
+            /// Returns the longest string that fits the rule (not simd optimized)
+            pub fn take_while_complete(input: &[u8]) -> nom::IResult<&[u8], &[u8]> {
+                let mut i = 0;
+                while i < input.len() {
+                    if unsafe { !TABLE.get_unchecked(*input.get_unchecked(i) as usize) } {
+                        break;
+                    }
+                    i += 1;
+                }
+                unsafe {
+                    Ok((
+                        input.get_unchecked(i..),
+                        input.get_unchecked(..i),
+                    ))
+                }
+            }
+
+            #[inline]
+            #[allow(dead_code)]
+            /// Returns the longest string that fits the rule (using simd if enabled)
+            ///
+            /// *Streaming version* will return a Err::Incomplete(Needed::Unknown) if the pattern reaches the end of the input.
+            pub fn take_while_fast(input: &[u8]) -> nom::IResult<&[u8], &[u8]> {
+                #[cfg(feature="simd")]
+                let result = take_while_simd(input);
+                #[cfg(not(feature="simd"))]
+                let result = take_while(input);
+                result
+            }
+
+            #[inline]
+            #[allow(dead_code)]
+            /// Returns the longest string that fits the rule (using simd if enabled)
+            pub fn take_while_complete_fast(input: &[u8]) -> nom::IResult<&[u8], &[u8]> {
+                #[cfg(feature="simd")]
+                let result = take_while_complete_simd(input);
+                #[cfg(not(feature="simd"))]
+                let result = take_while_complete(input);
+                result
+            }
         }
     }
 }