Add SIMD accelerated multiple pattern search.

This uses the "Teddy" algorithm, as learned from the Hyperscan regular
expression library: https://01.org/hyperscan

This support optional, subject to the following:

1. A nightly compiler.
2. Enabling the `simd-accel` feature.
3. Adding `RUSTFLAGS="-C target-feature=+ssse3"` when compiling.

Author: BurntSushi

.travis.yml

@@ -9,7 +9,7 @@ script:
   - cargo build --verbose
   - cargo build --verbose --manifest-path=regex-debug/Cargo.toml
   - if [ "$TRAVIS_RUST_VERSION" = "nightly" ]; then
-      travis_wait cargo test --verbose --features pattern;
+      RUSTFLAGS="-C target-feature=+ssse3" cargo test --verbose --features 'simd-accel pattern';
     else
       travis_wait cargo test --verbose;
     fi

Cargo.toml

@@ -21,6 +21,8 @@ memchr = "0.1.9"
 thread_local = "0.2.4"
 # For parsing regular expressions.
 regex-syntax = { path = "regex-syntax", version = "0.3.1" }
+# For accelerating text search.
+simd = { version = "0.1.0", optional = true }
 # For compiling UTF-8 decoding into automata.
 utf8-ranges = "0.1.3"
 
@@ -35,6 +37,8 @@ rand = "0.3"
 [features]
 # Enable to use the unstable pattern traits defined in std.
 pattern = []
+# Enable to use simd acceleration.
+simd-accel = ["simd"]
 
 [lib]
 # There are no benchmarks in the library code itself

bench/Cargo.toml

@@ -17,7 +17,7 @@ libc = "0.2"
 onig = { version = "0.4", optional = true }
 libpcre-sys = { version = "0.2", optional = true }
 memmap = "0.2"
-regex = { version = "0.1", path = ".." }
+regex = { version = "0.1", path = "..", features = ["simd-accel"] }
 regex_macros = { version = "0.1", path = "../regex_macros", optional = true }
 regex-syntax = { version = "0.3", path = "../regex-syntax" }
 rustc-serialize = "0.3"

bench/compile

@@ -1,5 +1,8 @@
 #!/bin/sh
 
+# Enable SIMD.
+export RUSTFLAGS="-C target-feature=+ssse3"
+
 exec cargo build \
   --release \
   --features 're-onig re-pcre1 re-pcre2 re-re2 re-rust re-rust-bytes re-tcl' \

bench/run

@@ -9,6 +9,9 @@ if [ $# = 0 ] || [ $1 = '-h' ] || [ $1 = '--help' ]; then
   usage
 fi
 
+# Enable SIMD.
+export RUSTFLAGS="-C target-feature=+ssse3"
+
 which="$1"
 shift
 case $which in

src/lib.rs

@@ -449,6 +449,7 @@
 #![deny(missing_docs)]
 #![cfg_attr(test, deny(warnings))]
 #![cfg_attr(feature = "pattern", feature(pattern))]
+#![cfg_attr(feature = "simd-accel", feature(cfg_target_feature))]
 #![doc(html_logo_url = "https://www.rust-lang.org/logos/rust-logo-128x128-blk-v2.png",
        html_favicon_url = "https://www.rust-lang.org/favicon.ico",
        html_root_url = "https://doc.rust-lang.org/regex/")]
@@ -458,6 +459,7 @@ extern crate memchr;
 extern crate thread_local;
 #[cfg(test)] extern crate quickcheck;
 extern crate regex_syntax as syntax;
+#[cfg(feature = "simd-accel")] extern crate simd;
 extern crate utf8_ranges;
 
 pub use error::Error;
@@ -582,6 +584,11 @@ mod re_plugin;
 mod re_set;
 mod re_trait;
 mod re_unicode;
+#[cfg(feature = "simd-accel")]
+mod simd_accel;
+#[cfg(not(feature = "simd-accel"))]
+#[path = "simd_fallback/mod.rs"]
+mod simd_accel;
 mod sparse;
 
 /// The `internal` module exists to support the `regex!` macro and other

src/literals.rs

@@ -15,6 +15,7 @@ use memchr::{memchr, memchr2, memchr3};
 use syntax;
 
 use freqs::BYTE_FREQUENCIES;
+use simd_accel::teddy128::Teddy;
 
 /// A prefix extracted from a compiled regular expression.
 ///
@@ -51,6 +52,8 @@ enum Matcher {
     Single(SingleSearch),
     /// An Aho-Corasick automaton.
     AC(FullAcAutomaton<syntax::Lit>),
+    /// A simd accelerated multiple string matcher.
+    Teddy128(Teddy),
 }
 
 impl LiteralSearcher {
@@ -100,6 +103,7 @@ impl LiteralSearcher {
             Bytes(ref sset) => sset.find(haystack).map(|i| (i, i + 1)),
             Single(ref s) => s.find(haystack).map(|i| (i, i + s.len())),
             AC(ref aut) => aut.find(haystack).next().map(|m| (m.start, m.end)),
+            Teddy128(ref ted) => ted.find(haystack).map(|m| (m.start, m.end)),
         }
     }
 
@@ -136,6 +140,9 @@ impl LiteralSearcher {
             Matcher::Bytes(ref sset) => LiteralIter::Bytes(&sset.dense),
             Matcher::Single(ref s) => LiteralIter::Single(&s.pat),
             Matcher::AC(ref ac) => LiteralIter::AC(ac.patterns()),
+            Matcher::Teddy128(ref ted) => {
+                LiteralIter::Teddy128(ted.patterns())
+            }
         }
     }
 
@@ -162,6 +169,7 @@ impl LiteralSearcher {
             Bytes(ref sset) => sset.dense.len(),
             Single(_) => 1,
             AC(ref aut) => aut.len(),
+            Teddy128(ref ted) => ted.len(),
         }
     }
 
@@ -173,6 +181,7 @@ impl LiteralSearcher {
             Bytes(ref sset) => sset.approximate_size(),
             Single(ref single) => single.approximate_size(),
             AC(ref aut) => aut.heap_bytes(),
+            Teddy128(ref ted) => ted.approximate_size(),
         }
     }
 }
@@ -190,23 +199,34 @@ impl Matcher {
 
     fn new(lits: &syntax::Literals, sset: SingleByteSet) -> Self {
         if lits.literals().is_empty() {
-            Matcher::Empty
-        } else if sset.dense.len() >= 26 {
+            return Matcher::Empty;
+        }
+        if sset.dense.len() >= 26 {
             // Avoid trying to match a large number of single bytes.
             // This is *very* sensitive to a frequency analysis comparison
             // between the bytes in sset and the composition of the haystack.
             // No matter the size of sset, if its members all are rare in the
             // haystack, then it'd be worth using it. How to tune this... IDK.
             // ---AG
-            Matcher::Empty
-        } else if sset.complete {
-            Matcher::Bytes(sset)
-        } else if lits.literals().len() == 1 {
-            Matcher::Single(SingleSearch::new(lits.literals()[0].to_vec()))
-        } else {
-            let pats = lits.literals().to_owned();
-            Matcher::AC(AcAutomaton::new(pats).into_full())
+            return Matcher::Empty;
+        }
+        if sset.complete {
+            return Matcher::Bytes(sset);
+        }
+        if lits.literals().len() == 1 {
+            let lit = lits.literals()[0].to_vec();
+            return Matcher::Single(SingleSearch::new(lit));
         }
+        // Only try Teddy if Aho-Corasick can't use memchr.
+        // Also, in its current form, Teddy doesn't scale well to lots of
+        // literals.
+        if sset.dense.len() > 1 && lits.literals().len() <= 32 {
+            if let Some(ted) = Teddy::new(lits) {
+                return Matcher::Teddy128(ted);
+            }
+        }
+        let pats = lits.literals().to_owned();
+        Matcher::AC(AcAutomaton::new(pats).into_full())
     }
 }
 
@@ -215,6 +235,7 @@ pub enum LiteralIter<'a> {
     Bytes(&'a [u8]),
     Single(&'a [u8]),
     AC(&'a [syntax::Lit]),
+    Teddy128(&'a [Vec<u8>]),
 }
 
 impl<'a> Iterator for LiteralIter<'a> {
@@ -250,6 +271,15 @@ impl<'a> Iterator for LiteralIter<'a> {
                     Some(&**next)
                 }
             }
+            LiteralIter::Teddy128(ref mut lits) => {
+                if lits.is_empty() {
+                    None
+                } else {
+                    let next = &lits[0];
+                    *lits = &lits[1..];
+                    Some(&**next)
+                }
+            }
         }
     }
 }

src/simd_accel/mod.rs

@@ -0,0 +1,5 @@
+#[cfg(target_feature = "ssse3")]
+pub mod teddy128;
+#[cfg(not(target_feature = "ssse3"))]
+#[path = "../simd_fallback/teddy128.rs"]
+pub mod teddy128;