Speed up in the presence of timeouts.

Currently regexp2 can be quite slow when a MatchTimeout is
supplied (a micro-benchmark shows 2300ns compared to 34ns
when no timeout is supplied). This slowdown is caused by
repeated timeout checks which call time.Now().

The new approach introduces a fast but approximate clock that
is just an atomic variable updated by a goroutine once very
10ms. The timeout check can now compare this variable.

Removed "timeout check skip" mechanism since a timeout
check is now very cheap.

Added a simple micro-benchmark that compares the speed
of searching 100 byte text with and without a timeout.

Performance impact:
1. A micro-benchmark that looks for an "easy" regexp in a 100
   byte string goes from ~2300ns to ~34ns.
2. Chroma (syntax highlighter) speeds up from ~500ms to ~50ms
   on a 24KB source file.

fastclock.go

@@ -0,0 +1,104 @@
+package regexp2
+
+import (
+	"sync"
+	"sync/atomic"
+	"time"
+)
+
+// fasttime holds a time value (ticks since clock initialization)
+type fasttime int64
+
+// fastclock provides a fast clock implementation.
+//
+// A background goroutine periodically stores the current time
+// into an atomic variable.
+//
+// A deadline can be quickly checked for expiration by comparing
+// its value to the clock stored in the atomic variable.
+//
+// The goroutine automatically stops once clockEnd is reached.
+// (clockEnd covers the largest deadline seen so far + some
+// extra time). This ensures that if regexp2 with timeouts
+// stops being used we will stop background work.
+type fastclock struct {
+	// current and clockEnd can be read via atomic loads.
+	// Reads and writes of other fields require mu to be held.
+	mu sync.Mutex
+
+	start    time.Time  // Time corresponding to fasttime(0)
+	current  atomicTime // Current time (approximate)
+	clockEnd atomicTime // When clock updater is supposed to stop (>= any existing deadline)
+	running  bool       // Is a clock updater running?
+}
+
+var fast fastclock
+
+// reached returns true if current time is at or past t.
+func (t fasttime) reached() bool {
+	return fast.current.read() >= t
+}
+
+// makeDeadline returns a time that is approximately time.Now().Add(d)
+func makeDeadline(d time.Duration) fasttime {
+	// Increase the deadline since the clock we are reading may be
+	// just about to tick forwards.
+	end := fast.current.read() + durationToTicks(d+clockPeriod)
+
+	// Start or extend clock if necessary.
+	if end > fast.clockEnd.read() {
+		extendClock(end)
+	}
+	return end
+}
+
+// extendClock ensures that clock is live and will run until at least end.
+func extendClock(end fasttime) {
+	fast.mu.Lock()
+	defer fast.mu.Unlock()
+
+	if fast.start.IsZero() {
+		fast.start = time.Now()
+	}
+
+	// Extend the running time to cover end as well as a bit of slop.
+	if shutdown := end + durationToTicks(time.Second); shutdown > fast.clockEnd.read() {
+		fast.clockEnd.write(shutdown)
+	}
+
+	// Start clock if necessary
+	if !fast.running {
+		fast.running = true
+		go runClock()
+	}
+}
+
+func durationToTicks(d time.Duration) fasttime {
+	// Downscale nanoseconds to approximately a millisecond so that we can avoid
+	// overflow even if the caller passes in math.MaxInt64.
+	return fasttime(d) >> 20
+}
+
+// clockPeriod is the approximate interval between updates of approximateClock.
+const clockPeriod = time.Millisecond * 10
+
+func runClock() {
+	fast.mu.Lock()
+	defer fast.mu.Unlock()
+
+	for fast.current.read() <= fast.clockEnd.read() {
+		// Unlock while sleeping.
+		fast.mu.Unlock()
+		time.Sleep(clockPeriod)
+		fast.mu.Lock()
+
+		newTime := durationToTicks(time.Since(fast.start))
+		fast.current.write(newTime)
+	}
+	fast.running = false
+}
+
+type atomicTime struct{ v int64 } // Should change to atomic.Int64 when we can use go 1.19
+
+func (t *atomicTime) read() fasttime   { return fasttime(atomic.LoadInt64(&t.v)) }
+func (t *atomicTime) write(v fasttime) { atomic.StoreInt64(&t.v, int64(v)) }

fastclock_test.go

@@ -0,0 +1,33 @@
+package regexp2
+
+import (
+	"fmt"
+	"testing"
+	"time"
+)
+
+func TestDeadline(t *testing.T) {
+	for _, delay := range []time.Duration{
+		clockPeriod / 10,
+		clockPeriod,
+		clockPeriod * 5,
+		clockPeriod * 10,
+	} {
+		t.Run(fmt.Sprint(delay), func(t *testing.T) {
+			t.Parallel()
+			start := time.Now()
+			d := makeDeadline(delay)
+			if d.reached() {
+				t.Fatalf("deadline (%v) unexpectedly expired immediately", delay)
+			}
+			time.Sleep(delay / 2)
+			if d.reached() {
+				t.Fatalf("deadline (%v) expired too soon (after %v)", delay, time.Since(start))
+			}
+			time.Sleep(delay/2 + 2*clockPeriod) // Give clock time to tick
+			if !d.reached() {
+				t.Fatalf("deadline (%v) did not expire within %v", delay, time.Since(start))
+			}
+		})
+	}
+}

regexp_performance_test.go

@@ -3,6 +3,7 @@ package regexp2
 import (
 	"strings"
 	"testing"
+	"time"
 )
 
 func BenchmarkLiteral(b *testing.B) {
@@ -305,3 +306,25 @@ func BenchmarkLeading(b *testing.B) {
 		}
 	}
 }
+
+func BenchmarkShortSearch(b *testing.B) {
+	for _, name := range []string{"no-timeout", "timeout"} {
+		b.Run(name, func(b *testing.B) {
+			b.StopTimer()
+			r := MustCompile(easy0, 0)
+			if name == "timeout" {
+				r.MatchTimeout = time.Second
+			}
+			t := makeText(100)
+			b.SetBytes(int64(len(t)))
+			b.StartTimer()
+			for i := 0; i < b.N; i++ {
+				if m, err := r.MatchRunes(t); m {
+					b.Fatal("match!")
+				} else if err != nil {
+					b.Fatalf("Err %v", err)
+				}
+			}
+		})
+	}
+}

runner.go

@@ -58,10 +58,9 @@ type runner struct {
 
 	runmatch *Match // result object
 
-	ignoreTimeout       bool
-	timeout             time.Duration // timeout in milliseconds (needed for actual)
-	timeoutChecksToSkip int
-	timeoutAt           time.Time
+	ignoreTimeout bool
+	timeout       time.Duration // timeout in milliseconds (needed for actual)
+	deadline      fasttime
 
 	operator        syntax.InstOp
 	codepos         int
@@ -1551,39 +1550,15 @@ func (r *runner) isECMABoundary(index, startpos, endpos int) bool {
 		(index < endpos && syntax.IsECMAWordChar(r.runtext[index]))
 }
 
-// this seems like a comment to justify randomly picking 1000 :-P
-// We have determined this value in a series of experiments where x86 retail
-// builds (ono-lab-optimized) were run on different pattern/input pairs. Larger values
-// of TimeoutCheckFrequency did not tend to increase performance; smaller values
-// of TimeoutCheckFrequency tended to slow down the execution.
-const timeoutCheckFrequency int = 1000
-
 func (r *runner) startTimeoutWatch() {
 	if r.ignoreTimeout {
 		return
 	}
-
-	r.timeoutChecksToSkip = timeoutCheckFrequency
-	r.timeoutAt = time.Now().Add(r.timeout)
+	r.deadline = makeDeadline(r.timeout)
 }
 
 func (r *runner) checkTimeout() error {
-	if r.ignoreTimeout {
-		return nil
-	}
-	r.timeoutChecksToSkip--
-	if r.timeoutChecksToSkip != 0 {
-		return nil
-	}
-
-	r.timeoutChecksToSkip = timeoutCheckFrequency
-	return r.doCheckTimeout()
-}
-
-func (r *runner) doCheckTimeout() error {
-	current := time.Now()
-
-	if current.Before(r.timeoutAt) {
+	if r.ignoreTimeout || !r.deadline.reached() {
 		return nil
 	}