Allow AVX2 and SSE42 to both be used if available (#45)

* don't try sse42 if avx2 found a match

* adjust longer request bench show benefit of sse42 combined with avx2

* allow runtime detection to use sse42 if not long enough for avx2

benches/parse.rs

@@ -20,7 +20,7 @@ Accept-Encoding: gzip,deflate\r\n\
 Accept-Charset: Shift_JIS,utf-8;q=0.7,*;q=0.7\r\n\
 Keep-Alive: 115\r\n\
 Connection: keep-alive\r\n\
-Cookie: wp_ozh_wsa_visits=2; wp_ozh_wsa_visit_lasttime=xxxxxxxxxx; __utma=xxxxxxxxx.xxxxxxxxxx.xxxxxxxxxx.xxxxxxxxxx.xxxxxxxxxx.x; __utmz=xxxxxxxxx.xxxxxxxxxx.x.x.utmccn=(referral)|utmcsr=reader.livedoor.com|utmcct=/reader/|utmcmd=referral\r\n\r\n";
+Cookie: wp_ozh_wsa_visits=2; wp_ozh_wsa_visit_lasttime=xxxxxxxxxx; __utma=xxxxxxxxx.xxxxxxxxxx.xxxxxxxxxx.xxxxxxxxxx.xxxxxxxxxx.x; __utmz=xxxxxxxxx.xxxxxxxxxx.x.x.utmccn=(referral)|utmcsr=reader.livedoor.com|utmcct=/reader/|utmcmd=referral|padding=under256\r\n\r\n";
 
 
 #[bench]

src/simd/avx2.rs

@@ -1,14 +1,23 @@
 use ::iter::Bytes;
 
-pub unsafe fn parse_uri_batch_32<'a>(bytes: &mut Bytes<'a>) {
+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,
+}
+
+
+pub unsafe fn parse_uri_batch_32<'a>(bytes: &mut Bytes<'a>) -> Scan {
     while bytes.as_ref().len() >= 32 {
         let advance = match_url_char_32_avx(bytes.as_ref());
         bytes.advance(advance);
 
         if advance != 32 {
-            break;
+            return Scan::Found;
         }
     }
+    Scan::TooShort
 }
 
 #[cfg(target_arch = "x86_64")]
@@ -57,15 +66,16 @@ unsafe fn match_url_char_32_avx(_: &[u8]) -> usize {
     unreachable!("AVX2 detection should be disabled for x86");
 }
 
-pub unsafe fn match_header_value_batch_32(bytes: &mut Bytes) {
+pub unsafe fn match_header_value_batch_32(bytes: &mut Bytes) -> Scan {
     while bytes.as_ref().len() >= 32 {
         let advance = match_header_value_char_32_avx(bytes.as_ref());
         bytes.advance(advance);
 
         if advance != 32 {
-            break;
+            return Scan::Found;
         }
     }
+    Scan::TooShort
 }
 
 #[cfg(target_arch = "x86_64")]

src/simd/mod.rs

@@ -64,14 +64,20 @@ mod runtime {
     const INIT: usize = 0;
     const SSE_42: usize = 1;
     const AVX_2: usize = 2;
+    const AVX_2_AND_SSE_42: usize = 3;
     const NONE: usize = ::core::usize::MAX;
 
     fn detect() -> usize {
         let feat = FEATURE.load(Ordering::Relaxed);
         if feat == INIT {
             if cfg!(target_arch = "x86_64") && is_x86_feature_detected!("avx2") {
-                FEATURE.store(AVX_2, Ordering::Relaxed);
-                return AVX_2;
+                if is_x86_feature_detected!("sse4.2") {
+                    FEATURE.store(AVX_2_AND_SSE_42, Ordering::Relaxed);
+                    return AVX_2_AND_SSE_42;
+                } else {
+                    FEATURE.store(AVX_2, Ordering::Relaxed);
+                    return AVX_2;
+                }
             } else if is_x86_feature_detected!("sse4.2") {
                 FEATURE.store(SSE_42, Ordering::Relaxed);
                 return SSE_42;
@@ -86,7 +92,13 @@ mod runtime {
         unsafe {
             match detect() {
                 SSE_42 => super::sse42::parse_uri_batch_16(bytes),
-                AVX_2 => super::avx2::parse_uri_batch_32(bytes),
+                AVX_2 => { super::avx2::parse_uri_batch_32(bytes); },
+                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)
+                },
                 _ => ()
             }
         }
@@ -98,7 +110,13 @@ mod runtime {
         unsafe {
             match detect() {
                 SSE_42 => super::sse42::match_header_value_batch_16(bytes),
-                AVX_2 => super::avx2::match_header_value_batch_32(bytes),
+                AVX_2 => { super::avx2::match_header_value_batch_32(bytes); },
+                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)
+                },
                 _ => ()
             }
         }
@@ -191,8 +209,12 @@ mod avx2_compile_time {
     pub fn match_header_value_vectored(bytes: &mut ::Bytes) {
         // do both, since avx2 only works when bytes.len() >= 32
         if cfg!(target_arch = "x86_64") && is_x86_feature_detected!("avx2") {
-            unsafe {
-                super::avx2::match_header_value_batch_32(bytes);
+            let scanned = unsafe {
+                super::avx2::match_header_value_batch_32(bytes)
+            };
+
+            if let super::avx2::Scan::Found = scanned {
+                return;
             }
         }
         if is_x86_feature_detected!("sse4.2") {