testing: only compute b.N once when passed -count > 1

When running a benchmark multiple times, instead of re-computing the
value of b.N each time, use the value found by the first run.

For

  go test -bench=. -benchtime 3s -count 2 p_test.go

on the benchmark in the linked issue; before:

  BenchmarkBenchmark-4   	     500	  10180593 ns/op
  --- BENCH: BenchmarkBenchmark-4
  	  p_test.go:13: single call took 10.111079ms
  	  p_test.go:13: single call took 1.017298685s
  	  p_test.go:13: single call took 5.090096124s
  BenchmarkBenchmark-4   	     500	  10182164 ns/op
  --- BENCH: BenchmarkBenchmark-4
  	  p_test.go:13: single call took 10.098169ms
  	  p_test.go:13: single call took 1.017712905s
  	  p_test.go:13: single call took 5.090898517s
  PASS
  ok  	command-line-arguments	12.244s

and after:

  BenchmarkBenchmark-4   	     500	  10177076 ns/op
  --- BENCH: BenchmarkBenchmark-4
  	  p_test.go:13: single call took 10.091301ms
  	  p_test.go:13: single call took 1.016943125s
  	  p_test.go:13: single call took 5.088376028s
  BenchmarkBenchmark-4   	     500	  10171497 ns/op
  --- BENCH: BenchmarkBenchmark-4
  	  p_test.go:13: single call took 10.140245ms
  	  p_test.go:13: single call took 5.085605921s
  PASS
  ok  	command-line-arguments	11.218s

Fixes #23423

Change-Id: Ie66a8c5ac43881eb8741e14105db28745b4d56d3
Reviewed-on: https://go-review.googlesource.com/110775
Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

1 files changed, 32 insertions, 15 deletions

diff --git a/src/testing/benchmark.go b/src/testing/benchmark.go
index 4d569b7971..ac9ca58397 100644
--- a/src/testing/benchmark.go
+++ b/src/testing/benchmark.go
@@ -251,27 +251,20 @@ func (b *B) run() {
 		b.context.processBench(b) // Must call doBench.
 	} else {
 		// Running func Benchmark.
-		b.doBench()
+		b.doBench(0)
 	}
 }
 
-func (b *B) doBench() BenchmarkResult {
-	go b.launch()
+func (b *B) doBench(hint int) BenchmarkResult {
+	go b.launch(hint)
 	<-b.signal
 	return b.result
 }
 
-// launch launches the benchmark function. It gradually increases the number
-// of benchmark iterations until the benchmark runs for the requested benchtime.
-// launch is run by the doBench function as a separate goroutine.
-// run1 must have been called on b.
-func (b *B) launch() {
-	// Signal that we're done whether we return normally
-	// or by FailNow's runtime.Goexit.
-	defer func() {
-		b.signal <- true
-	}()
-
+// autodetectN runs the benchmark function, gradually increasing the
+// number of iterations until the benchmark runs for the requested
+// benchtime.
+func (b *B) autodetectN() {
 	// Run the benchmark for at least the specified amount of time.
 	d := b.benchTime
 	for n := 1; !b.failed && b.duration < d && n < 1e9; {
@@ -289,6 +282,26 @@ func (b *B) launch() {
 		n = roundUp(n)
 		b.runN(n)
 	}
+}
+
+// launch launches the benchmark function for hintN iterations. If
+// hintN == 0, it autodetects the number of benchmark iterations based
+// on the requested benchtime.
+// launch is run by the doBench function as a separate goroutine.
+// run1 must have been called on b.
+func (b *B) launch(hintN int) {
+	// Signal that we're done whether we return normally
+	// or by FailNow's runtime.Goexit.
+	defer func() {
+		b.signal <- true
+	}()
+
+	if hintN == 0 {
+		b.autodetectN()
+	} else {
+		b.runN(hintN)
+	}
+
 	b.result = BenchmarkResult{b.N, b.duration, b.bytes, b.netAllocs, b.netBytes}
 }
 
@@ -426,6 +439,7 @@ func runBenchmarks(importPath string, matchString func(pat, str string) (bool, e
 // processBench runs bench b for the configured CPU counts and prints the results.
 func (ctx *benchContext) processBench(b *B) {
 	for i, procs := range cpuList {
+		var nHint int
 		for j := uint(0); j < *count; j++ {
 			runtime.GOMAXPROCS(procs)
 			benchName := benchmarkName(b.name, procs)
@@ -444,7 +458,10 @@ func (ctx *benchContext) processBench(b *B) {
 				}
 				b.run1()
 			}
-			r := b.doBench()
+			r := b.doBench(nHint)
+			if j == 0 {
+				nHint = b.N
+			}
 			if b.failed {
 				// The output could be very long here, but probably isn't.
 				// We print it all, regardless, because we don't want to trim the reason