codekingpro/portable-devtools

1// Copyright 2009 The Go Authors. All rights reserved.2// Use of this source code is governed by a BSD-style3// license that can be found in the LICENSE file.4 5package testing6 7import (8	"context"9	"flag"10	"fmt"11	"internal/sysinfo"12	"io"13	"math"14	"os"15	"runtime"16	"slices"17	"strconv"18	"strings"19	"sync"20	"sync/atomic"21	"time"22	"unicode"23)24 25func initBenchmarkFlags() {26	matchBenchmarks = flag.String("test.bench", "", "run only benchmarks matching `regexp`")27	benchmarkMemory = flag.Bool("test.benchmem", false, "print memory allocations for benchmarks")28	flag.Var(&benchTime, "test.benchtime", "run each benchmark for duration `d` or N times if `d` is of the form Nx")29}30 31var (32	matchBenchmarks *string33	benchmarkMemory *bool34 35	benchTime = durationOrCountFlag{d: 1 * time.Second} // changed during test of testing package36)37 38type durationOrCountFlag struct {39	d         time.Duration40	n         int41	allowZero bool42}43 44func (f *durationOrCountFlag) String() string {45	if f.n > 0 {46		return fmt.Sprintf("%dx", f.n)47	}48	return f.d.String()49}50 51func (f *durationOrCountFlag) Set(s string) error {52	if strings.HasSuffix(s, "x") {53		n, err := strconv.ParseInt(s[:len(s)-1], 10, 0)54		if err != nil || n < 0 || (!f.allowZero && n == 0) {55			return fmt.Errorf("invalid count")56		}57		*f = durationOrCountFlag{n: int(n)}58		return nil59	}60	d, err := time.ParseDuration(s)61	if err != nil || d < 0 || (!f.allowZero && d == 0) {62		return fmt.Errorf("invalid duration")63	}64	*f = durationOrCountFlag{d: d}65	return nil66}67 68// Global lock to ensure only one benchmark runs at a time.69var benchmarkLock sync.Mutex70 71// Used for every benchmark for measuring memory.72var memStats runtime.MemStats73 74// InternalBenchmark is an internal type but exported because it is cross-package;75// it is part of the implementation of the "go test" command.76type InternalBenchmark struct {77	Name string78	F    func(b *B)79}80 81// B is a type passed to [Benchmark] functions to manage benchmark82// timing and control the number of iterations.83//84// A benchmark ends when its Benchmark function returns or calls any of the methods85// [B.FailNow], [B.Fatal], [B.Fatalf], [B.SkipNow], [B.Skip], or [B.Skipf].86// Those methods must be called only from the goroutine running the Benchmark function.87// The other reporting methods, such as the variations of [B.Log] and [B.Error],88// may be called simultaneously from multiple goroutines.89//90// Like in tests, benchmark logs are accumulated during execution91// and dumped to standard output when done. Unlike in tests, benchmark logs92// are always printed, so as not to hide output whose existence may be93// affecting benchmark results.94type B struct {95	common96	importPath       string // import path of the package containing the benchmark97	bstate           *benchState98	N                int99	previousN        int           // number of iterations in the previous run100	previousDuration time.Duration // total duration of the previous run101	benchFunc        func(b *B)102	benchTime        durationOrCountFlag103	bytes            int64104	missingBytes     bool // one of the subbenchmarks does not have bytes set.105	timerOn          bool106	showAllocResult  bool107	result           BenchmarkResult108	parallelism      int // RunParallel creates parallelism*GOMAXPROCS goroutines109	// The initial states of memStats.Mallocs and memStats.TotalAlloc.110	startAllocs uint64111	startBytes  uint64112	// The net total of this test after being run.113	netAllocs uint64114	netBytes  uint64115	// Extra metrics collected by ReportMetric.116	extra map[string]float64117 118	// loop tracks the state of B.Loop119	loop struct {120		// n is the target number of iterations. It gets bumped up as we go.121		// When the benchmark loop is done, we commit this to b.N so users can122		// do reporting based on it, but we avoid exposing it until then.123		n uint64124		// i is the current Loop iteration. It's strictly monotonically125		// increasing toward n.126		//127		// The high bit is used to poison the Loop fast path and fall back to128		// the slow path.129		i uint64130 131		done bool // set when B.Loop return false132	}133}134 135// StartTimer starts timing a test. This function is called automatically136// before a benchmark starts, but it can also be used to resume timing after137// a call to [B.StopTimer].138func (b *B) StartTimer() {139	if !b.timerOn {140		runtime.ReadMemStats(&memStats)141		b.startAllocs = memStats.Mallocs142		b.startBytes = memStats.TotalAlloc143		b.start = highPrecisionTimeNow()144		b.timerOn = true145		b.loop.i &^= loopPoisonTimer146	}147}148 149// StopTimer stops timing a test. This can be used to pause the timer150// while performing steps that you don't want to measure.151func (b *B) StopTimer() {152	if b.timerOn {153		b.duration += highPrecisionTimeSince(b.start)154		runtime.ReadMemStats(&memStats)155		b.netAllocs += memStats.Mallocs - b.startAllocs156		b.netBytes += memStats.TotalAlloc - b.startBytes157		b.timerOn = false158		// If we hit B.Loop with the timer stopped, fail.159		b.loop.i |= loopPoisonTimer160	}161}162 163// ResetTimer zeroes the elapsed benchmark time and memory allocation counters164// and deletes user-reported metrics.165// It does not affect whether the timer is running.166func (b *B) ResetTimer() {167	if b.extra == nil {168		// Allocate the extra map before reading memory stats.169		// Pre-size it to make more allocation unlikely.170		b.extra = make(map[string]float64, 16)171	} else {172		clear(b.extra)173	}174	if b.timerOn {175		runtime.ReadMemStats(&memStats)176		b.startAllocs = memStats.Mallocs177		b.startBytes = memStats.TotalAlloc178		b.start = highPrecisionTimeNow()179	}180	b.duration = 0181	b.netAllocs = 0182	b.netBytes = 0183}184 185// SetBytes records the number of bytes processed in a single operation.186// If this is called, the benchmark will report ns/op and MB/s.187func (b *B) SetBytes(n int64) { b.bytes = n }188 189// ReportAllocs enables malloc statistics for this benchmark.190// It is equivalent to setting -test.benchmem, but it only affects the191// benchmark function that calls ReportAllocs.192func (b *B) ReportAllocs() {193	b.showAllocResult = true194}195 196// runN runs a single benchmark for the specified number of iterations.197func (b *B) runN(n int) {198	benchmarkLock.Lock()199	defer benchmarkLock.Unlock()200	ctx, cancelCtx := context.WithCancel(context.Background())201	defer func() {202		b.runCleanup(normalPanic)203		b.checkRaces()204	}()205	// Try to get a comparable environment for each run206	// by clearing garbage from previous runs.207	runtime.GC()208	b.resetRaces()209	b.N = n210	b.loop.n = 0211	b.loop.i = 0212	b.loop.done = false213	b.ctx = ctx214	b.cancelCtx = cancelCtx215 216	b.parallelism = 1217	b.ResetTimer()218	b.StartTimer()219	b.benchFunc(b)220	b.StopTimer()221	b.previousN = n222	b.previousDuration = b.duration223 224	if b.loop.n > 0 && !b.loop.done && !b.failed {225		b.Error("benchmark function returned without B.Loop() == false (break or return in loop?)")226	}227}228 229// run1 runs the first iteration of benchFunc. It reports whether more230// iterations of this benchmarks should be run.231func (b *B) run1() bool {232	if bstate := b.bstate; bstate != nil {233		// Extend maxLen, if needed.234		if n := len(b.name) + bstate.extLen + 1; n > bstate.maxLen {235			bstate.maxLen = n + 8 // Add additional slack to avoid too many jumps in size.236		}237	}238	go func() {239		// Signal that we're done whether we return normally240		// or by FailNow's runtime.Goexit.241		defer func() {242			b.signal <- true243		}()244 245		b.runN(1)246	}()247	<-b.signal248	if b.failed {249		fmt.Fprintf(b.w, "%s--- FAIL: %s\n%s", b.chatty.prefix(), b.name, b.output)250		return false251	}252	// Only print the output if we know we are not going to proceed.253	// Otherwise it is printed in processBench.254	b.mu.RLock()255	finished := b.finished256	b.mu.RUnlock()257	if b.hasSub.Load() || finished {258		tag := "BENCH"259		if b.skipped {260			tag = "SKIP"261		}262		if b.chatty != nil && (len(b.output) > 0 || finished) {263			b.trimOutput()264			fmt.Fprintf(b.w, "%s--- %s: %s\n%s", b.chatty.prefix(), tag, b.name, b.output)265		}266		return false267	}268	return true269}270 271var labelsOnce sync.Once272 273// run executes the benchmark in a separate goroutine, including all of its274// subbenchmarks. b must not have subbenchmarks.275func (b *B) run() {276	labelsOnce.Do(func() {277		fmt.Fprintf(b.w, "goos: %s\n", runtime.GOOS)278		fmt.Fprintf(b.w, "goarch: %s\n", runtime.GOARCH)279		if b.importPath != "" {280			fmt.Fprintf(b.w, "pkg: %s\n", b.importPath)281		}282		if cpu := sysinfo.CPUName(); cpu != "" {283			fmt.Fprintf(b.w, "cpu: %s\n", cpu)284		}285	})286	if b.bstate != nil {287		// Running go test --test.bench288		b.bstate.processBench(b) // Must call doBench.289	} else {290		// Running func Benchmark.291		b.doBench()292	}293}294 295func (b *B) doBench() BenchmarkResult {296	go b.launch()297	<-b.signal298	return b.result299}300 301// Don't run more than 1e9 times. (This also keeps n in int range on 32 bit platforms.)302const maxBenchPredictIters = 1_000_000_000303 304func predictN(goalns int64, prevIters int64, prevns int64, last int64) int {305	if prevns == 0 {306		// Round up to dodge divide by zero. See https://go.dev/issue/70709.307		prevns = 1308	}309 310	// Order of operations matters.311	// For very fast benchmarks, prevIters ~= prevns.312	// If you divide first, you get 0 or 1,313	// which can hide an order of magnitude in execution time.314	// So multiply first, then divide.315	n := goalns * prevIters / prevns316	// Run more iterations than we think we'll need (1.2x).317	n += n / 5318	// Don't grow too fast in case we had timing errors previously.319	n = min(n, 100*last)320	// Be sure to run at least one more than last time.321	n = max(n, last+1)322	// Don't run more than 1e9 times. (This also keeps n in int range on 32 bit platforms.)323	n = min(n, maxBenchPredictIters)324	return int(n)325}326 327// launch launches the benchmark function. It gradually increases the number328// of benchmark iterations until the benchmark runs for the requested benchtime.329// launch is run by the doBench function as a separate goroutine.330// run1 must have been called on b.331func (b *B) launch() {332	// Signal that we're done whether we return normally333	// or by FailNow's runtime.Goexit.334	defer func() {335		b.signal <- true336	}()337 338	// b.Loop does its own ramp-up logic so we just need to run it once.339	// If b.loop.n is non zero, it means b.Loop has already run.340	if b.loop.n == 0 {341		// Run the benchmark for at least the specified amount of time.342		if b.benchTime.n > 0 {343			// We already ran a single iteration in run1.344			// If -benchtime=1x was requested, use that result.345			// See https://golang.org/issue/32051.346			if b.benchTime.n > 1 {347				b.runN(b.benchTime.n)348			}349		} else {350			d := b.benchTime.d351			for n := int64(1); !b.failed && b.duration < d && n < 1e9; {352				last := n353				// Predict required iterations.354				goalns := d.Nanoseconds()355				prevIters := int64(b.N)356				n = int64(predictN(goalns, prevIters, b.duration.Nanoseconds(), last))357				b.runN(int(n))358			}359		}360	}361	b.result = BenchmarkResult{b.N, b.duration, b.bytes, b.netAllocs, b.netBytes, b.extra}362}363 364// Elapsed returns the measured elapsed time of the benchmark.365// The duration reported by Elapsed matches the one measured by366// [B.StartTimer], [B.StopTimer], and [B.ResetTimer].367func (b *B) Elapsed() time.Duration {368	d := b.duration369	if b.timerOn {370		d += highPrecisionTimeSince(b.start)371	}372	return d373}374 375// ReportMetric adds "n unit" to the reported benchmark results.376// If the metric is per-iteration, the caller should divide by b.N,377// and by convention units should end in "/op".378// ReportMetric overrides any previously reported value for the same unit.379// ReportMetric panics if unit is the empty string or if unit contains380// any whitespace.381// If unit is a unit normally reported by the benchmark framework itself382// (such as "allocs/op"), ReportMetric will override that metric.383// Setting "ns/op" to 0 will suppress that built-in metric.384func (b *B) ReportMetric(n float64, unit string) {385	if unit == "" {386		panic("metric unit must not be empty")387	}388	if strings.IndexFunc(unit, unicode.IsSpace) >= 0 {389		panic("metric unit must not contain whitespace")390	}391	b.extra[unit] = n392}393 394func (b *B) stopOrScaleBLoop() bool {395	t := b.Elapsed()396	if t >= b.benchTime.d {397		// We've reached the target398		return false399	}400	// Loop scaling401	goalns := b.benchTime.d.Nanoseconds()402	prevIters := int64(b.loop.n)403	b.loop.n = uint64(predictN(goalns, prevIters, t.Nanoseconds(), prevIters))404	if b.loop.n&loopPoisonMask != 0 {405		// The iteration count should never get this high, but if it did we'd be406		// in big trouble.407		panic("loop iteration target overflow")408	}409	// predictN may have capped the number of iterations; make sure to410	// terminate if we've already hit that cap.411	return uint64(prevIters) < b.loop.n412}413 414func (b *B) loopSlowPath() bool {415	// Consistency checks416	if !b.timerOn {417		b.Fatal("B.Loop called with timer stopped")418	}419	if b.loop.i&loopPoisonMask != 0 {420		panic(fmt.Sprintf("unknown loop stop condition: %#x", b.loop.i))421	}422 423	if b.loop.n == 0 {424		// It's the first call to b.Loop() in the benchmark function.425		if b.benchTime.n > 0 {426			// Fixed iteration count.427			b.loop.n = uint64(b.benchTime.n)428		} else {429			// Initialize target to 1 to kick start loop scaling.430			b.loop.n = 1431		}432		// Within a b.Loop loop, we don't use b.N (to avoid confusion).433		b.N = 0434		b.ResetTimer()435 436		// Start the next iteration.437		b.loop.i++438		return true439	}440 441	// Should we keep iterating?442	var more bool443	if b.benchTime.n > 0 {444		// The iteration count is fixed, so we should have run this many and now445		// be done.446		if b.loop.i != uint64(b.benchTime.n) {447			// We shouldn't be able to reach the slow path in this case.448			panic(fmt.Sprintf("iteration count %d < fixed target %d", b.loop.i, b.benchTime.n))449		}450		more = false451	} else {452		// Handle fixed time case453		more = b.stopOrScaleBLoop()454	}455	if !more {456		b.StopTimer()457		// Commit iteration count458		b.N = int(b.loop.n)459		b.loop.done = true460		return false461	}462 463	// Start the next iteration.464	b.loop.i++465	return true466}467 468// Loop returns true as long as the benchmark should continue running.469//470// A typical benchmark is structured like:471//472//	func Benchmark(b *testing.B) {473//		... setup ...474//		for b.Loop() {475//			... code to measure ...476//		}477//		... cleanup ...478//	}479//480// Loop resets the benchmark timer the first time it is called in a benchmark,481// so any setup performed prior to starting the benchmark loop does not count482// toward the benchmark measurement. Likewise, when it returns false, it stops483// the timer so cleanup code is not measured.484//485// Within the body of a "for b.Loop() { ... }" loop, arguments to and486// results from function calls and assigned variables within the loop are kept487// alive, preventing the compiler from fully optimizing away the loop body.488// Currently, this is implemented as a compiler transformation that wraps such489// variables with a runtime.KeepAlive intrinsic call. This applies only to490// statements syntactically between the curly braces of the loop, and the loop491// condition must be written exactly as "b.Loop()".492//493// After Loop returns false, b.N contains the total number of iterations that494// ran, so the benchmark may use b.N to compute other average metrics.495//496// Prior to the introduction of Loop, benchmarks were expected to contain an497// explicit loop from 0 to b.N. Benchmarks should either use Loop or contain a498// loop to b.N, but not both. Loop offers more automatic management of the499// benchmark timer, and runs each benchmark function only once per measurement,500// whereas b.N-based benchmarks must run the benchmark function (and any501// associated setup and cleanup) several times.502func (b *B) Loop() bool {503	// This is written such that the fast path is as fast as possible and can be504	// inlined.505	//506	// There are three cases where we'll fall out of the fast path:507	//508	// - On the first call, both i and n are 0.509	//510	// - If the loop reaches the n'th iteration, then i == n and we need511	//   to figure out the new target iteration count or if we're done.512	//513	// - If the timer is stopped, it poisons the top bit of i so the slow514	//   path can do consistency checks and fail.515	if b.loop.i < b.loop.n {516		b.loop.i++517		return true518	}519	return b.loopSlowPath()520}521 522// The loopPoison constants can be OR'd into B.loop.i to cause it to fall back523// to the slow path.524const (525	loopPoisonTimer = uint64(1 << (63 - iota))526	// If necessary, add more poison bits here.527 528	// loopPoisonMask is the set of all loop poison bits. (iota-1) is the index529	// of the bit we just set, from which we recreate that bit mask. We subtract530	// 1 to set all of the bits below that bit, then complement the result to531	// get the mask. Sorry, not sorry.532	loopPoisonMask = ^uint64((1 << (63 - (iota - 1))) - 1)533)534 535// BenchmarkResult contains the results of a benchmark run.536type BenchmarkResult struct {537	N         int           // The number of iterations.538	T         time.Duration // The total time taken.539	Bytes     int64         // Bytes processed in one iteration.540	MemAllocs uint64        // The total number of memory allocations.541	MemBytes  uint64        // The total number of bytes allocated.542 543	// Extra records additional metrics reported by ReportMetric.544	Extra map[string]float64545}546 547// NsPerOp returns the "ns/op" metric.548func (r BenchmarkResult) NsPerOp() int64 {549	if v, ok := r.Extra["ns/op"]; ok {550		return int64(v)551	}552	if r.N <= 0 {553		return 0554	}555	return r.T.Nanoseconds() / int64(r.N)556}557 558// mbPerSec returns the "MB/s" metric.559func (r BenchmarkResult) mbPerSec() float64 {560	if v, ok := r.Extra["MB/s"]; ok {561		return v562	}563	if r.Bytes <= 0 || r.T <= 0 || r.N <= 0 {564		return 0565	}566	return (float64(r.Bytes) * float64(r.N) / 1e6) / r.T.Seconds()567}568 569// AllocsPerOp returns the "allocs/op" metric,570// which is calculated as r.MemAllocs / r.N.571func (r BenchmarkResult) AllocsPerOp() int64 {572	if v, ok := r.Extra["allocs/op"]; ok {573		return int64(v)574	}575	if r.N <= 0 {576		return 0577	}578	return int64(r.MemAllocs) / int64(r.N)579}580 581// AllocedBytesPerOp returns the "B/op" metric,582// which is calculated as r.MemBytes / r.N.583func (r BenchmarkResult) AllocedBytesPerOp() int64 {584	if v, ok := r.Extra["B/op"]; ok {585		return int64(v)586	}587	if r.N <= 0 {588		return 0589	}590	return int64(r.MemBytes) / int64(r.N)591}592 593// String returns a summary of the benchmark results.594// It follows the benchmark result line format from595// https://golang.org/design/14313-benchmark-format, not including the596// benchmark name.597// Extra metrics override built-in metrics of the same name.598// String does not include allocs/op or B/op, since those are reported599// by [BenchmarkResult.MemString].600func (r BenchmarkResult) String() string {601	buf := new(strings.Builder)602	fmt.Fprintf(buf, "%8d", r.N)603 604	// Get ns/op as a float.605	ns, ok := r.Extra["ns/op"]606	if !ok {607		ns = float64(r.T.Nanoseconds()) / float64(r.N)608	}609	if ns != 0 {610		buf.WriteByte('\t')611		prettyPrint(buf, ns, "ns/op")612	}613 614	if mbs := r.mbPerSec(); mbs != 0 {615		fmt.Fprintf(buf, "\t%7.2f MB/s", mbs)616	}617 618	// Print extra metrics that aren't represented in the standard619	// metrics.620	var extraKeys []string621	for k := range r.Extra {622		switch k {623		case "ns/op", "MB/s", "B/op", "allocs/op":624			// Built-in metrics reported elsewhere.625			continue626		}627		extraKeys = append(extraKeys, k)628	}629	slices.Sort(extraKeys)630	for _, k := range extraKeys {631		buf.WriteByte('\t')632		prettyPrint(buf, r.Extra[k], k)633	}634	return buf.String()635}636 637func prettyPrint(w io.Writer, x float64, unit string) {638	// Print all numbers with 10 places before the decimal point639	// and small numbers with four sig figs. Field widths are640	// chosen to fit the whole part in 10 places while aligning641	// the decimal point of all fractional formats.642	var format string643	switch y := math.Abs(x); {644	case y == 0 || y >= 999.95:645		format = "%10.0f %s"646	case y >= 99.995:647		format = "%12.1f %s"648	case y >= 9.9995:649		format = "%13.2f %s"650	case y >= 0.99995:651		format = "%14.3f %s"652	case y >= 0.099995:653		format = "%15.4f %s"654	case y >= 0.0099995:655		format = "%16.5f %s"656	case y >= 0.00099995:657		format = "%17.6f %s"658	default:659		format = "%18.7f %s"660	}661	fmt.Fprintf(w, format, x, unit)662}663 664// MemString returns r.AllocedBytesPerOp and r.AllocsPerOp in the same format as 'go test'.665func (r BenchmarkResult) MemString() string {666	return fmt.Sprintf("%8d B/op\t%8d allocs/op",667		r.AllocedBytesPerOp(), r.AllocsPerOp())668}669 670// benchmarkName returns full name of benchmark including procs suffix.671func benchmarkName(name string, n int) string {672	if n != 1 {673		return fmt.Sprintf("%s-%d", name, n)674	}675	return name676}677 678type benchState struct {679	match *matcher680 681	maxLen int // The largest recorded benchmark name.682	extLen int // Maximum extension length.683}684 685// RunBenchmarks is an internal function but exported because it is cross-package;686// it is part of the implementation of the "go test" command.687func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {688	runBenchmarks("", matchString, benchmarks)689}690 691func runBenchmarks(importPath string, matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) bool {692	// If no flag was specified, don't run benchmarks.693	if len(*matchBenchmarks) == 0 {694		return true695	}696	// Collect matching benchmarks and determine longest name.697	maxprocs := 1698	for _, procs := range cpuList {699		if procs > maxprocs {700			maxprocs = procs701		}702	}703	bstate := &benchState{704		match:  newMatcher(matchString, *matchBenchmarks, "-test.bench", *skip),705		extLen: len(benchmarkName("", maxprocs)),706	}707	var bs []InternalBenchmark708	for _, Benchmark := range benchmarks {709		if _, matched, _ := bstate.match.fullName(nil, Benchmark.Name); matched {710			bs = append(bs, Benchmark)711			benchName := benchmarkName(Benchmark.Name, maxprocs)712			if l := len(benchName) + bstate.extLen + 1; l > bstate.maxLen {713				bstate.maxLen = l714			}715		}716	}717	main := &B{718		common: common{719			name:  "Main",720			w:     os.Stdout,721			bench: true,722		},723		importPath: importPath,724		benchFunc: func(b *B) {725			for _, Benchmark := range bs {726				b.Run(Benchmark.Name, Benchmark.F)727			}728		},729		benchTime: benchTime,730		bstate:    bstate,731	}732	if Verbose() {733		main.chatty = newChattyPrinter(main.w)734	}735	main.runN(1)736	return !main.failed737}738 739// processBench runs bench b for the configured CPU counts and prints the results.740func (s *benchState) processBench(b *B) {741	for i, procs := range cpuList {742		for j := uint(0); j < *count; j++ {743			runtime.GOMAXPROCS(procs)744			benchName := benchmarkName(b.name, procs)745 746			// If it's chatty, we've already printed this information.747			if b.chatty == nil {748				fmt.Fprintf(b.w, "%-*s\t", s.maxLen, benchName)749			}750			// Recompute the running time for all but the first iteration.751			if i > 0 || j > 0 {752				b = &B{753					common: common{754						signal: make(chan bool),755						name:   b.name,756						w:      b.w,757						chatty: b.chatty,758						bench:  true,759					},760					benchFunc: b.benchFunc,761					benchTime: b.benchTime,762				}763				b.setOutputWriter()764				b.run1()765			}766			r := b.doBench()767			if b.failed {768				// The output could be very long here, but probably isn't.769				// We print it all, regardless, because we don't want to trim the reason770				// the benchmark failed.771				fmt.Fprintf(b.w, "%s--- FAIL: %s\n%s", b.chatty.prefix(), benchName, b.output)772				continue773			}774			results := r.String()775			if b.chatty != nil {776				fmt.Fprintf(b.w, "%-*s\t", s.maxLen, benchName)777			}778			if *benchmarkMemory || b.showAllocResult {779				results += "\t" + r.MemString()780			}781			fmt.Fprintln(b.w, results)782			// Unlike with tests, we ignore the -chatty flag and always print output for783			// benchmarks since the output generation time will skew the results.784			if len(b.output) > 0 {785				b.trimOutput()786				fmt.Fprintf(b.w, "%s--- BENCH: %s\n%s", b.chatty.prefix(), benchName, b.output)787			}788			if p := runtime.GOMAXPROCS(-1); p != procs {789				fmt.Fprintf(os.Stderr, "testing: %s left GOMAXPROCS set to %d\n", benchName, p)790			}791			if b.chatty != nil && b.chatty.json {792				b.chatty.Updatef("", "=== NAME  %s\n", "")793			}794		}795	}796}797 798// If hideStdoutForTesting is true, Run does not print the benchName.799// This avoids a spurious print during 'go test' on package testing itself,800// which invokes b.Run in its own tests (see sub_test.go).801var hideStdoutForTesting = false802 803// Run benchmarks f as a subbenchmark with the given name. It reports804// whether there were any failures.805//806// A subbenchmark is like any other benchmark. A benchmark that calls Run at807// least once will not be measured itself and will be called once with N=1.808func (b *B) Run(name string, f func(b *B)) bool {809	// Since b has subbenchmarks, we will no longer run it as a benchmark itself.810	// Release the lock and acquire it on exit to ensure locks stay paired.811	b.hasSub.Store(true)812	benchmarkLock.Unlock()813	defer benchmarkLock.Lock()814 815	benchName, ok, partial := b.name, true, false816	if b.bstate != nil {817		benchName, ok, partial = b.bstate.match.fullName(&b.common, name)818	}819	if !ok {820		return true821	}822	var pc [maxStackLen]uintptr823	n := runtime.Callers(2, pc[:])824	sub := &B{825		common: common{826			signal:  make(chan bool),827			name:    benchName,828			parent:  &b.common,829			level:   b.level + 1,830			creator: pc[:n],831			w:       b.w,832			chatty:  b.chatty,833			bench:   true,834		},835		importPath: b.importPath,836		benchFunc:  f,837		benchTime:  b.benchTime,838		bstate:     b.bstate,839	}840	sub.setOutputWriter()841	if partial {842		// Partial name match, like -bench=X/Y matching BenchmarkX.843		// Only process sub-benchmarks, if any.844		sub.hasSub.Store(true)845	}846 847	if b.chatty != nil {848		labelsOnce.Do(func() {849			fmt.Printf("goos: %s\n", runtime.GOOS)850			fmt.Printf("goarch: %s\n", runtime.GOARCH)851			if b.importPath != "" {852				fmt.Printf("pkg: %s\n", b.importPath)853			}854			if cpu := sysinfo.CPUName(); cpu != "" {855				fmt.Printf("cpu: %s\n", cpu)856			}857		})858 859		if !hideStdoutForTesting {860			if b.chatty.json {861				b.chatty.Updatef(benchName, "=== RUN   %s\n", benchName)862			}863			fmt.Println(benchName)864		}865	}866 867	if sub.run1() {868		sub.run()869	}870	b.add(sub.result)871	return !sub.failed872}873 874// add simulates running benchmarks in sequence in a single iteration. It is875// used to give some meaningful results in case func Benchmark is used in876// combination with Run.877func (b *B) add(other BenchmarkResult) {878	r := &b.result879	// The aggregated BenchmarkResults resemble running all subbenchmarks as880	// in sequence in a single benchmark.881	r.N = 1882	r.T += time.Duration(other.NsPerOp())883	if other.Bytes == 0 {884		// Summing Bytes is meaningless in aggregate if not all subbenchmarks885		// set it.886		b.missingBytes = true887		r.Bytes = 0888	}889	if !b.missingBytes {890		r.Bytes += other.Bytes891	}892	r.MemAllocs += uint64(other.AllocsPerOp())893	r.MemBytes += uint64(other.AllocedBytesPerOp())894}895 896// trimOutput shortens the output from a benchmark, which can be very long.897func (b *B) trimOutput() {898	// The output is likely to appear multiple times because the benchmark899	// is run multiple times, but at least it will be seen. This is not a big deal900	// because benchmarks rarely print, but just in case, we trim it if it's too long.901	const maxNewlines = 10902	for nlCount, j := 0, 0; j < len(b.output); j++ {903		if b.output[j] == '\n' {904			nlCount++905			if nlCount >= maxNewlines {906				b.output = append(b.output[:j], "\n\t... [output truncated]\n"...)907				break908			}909		}910	}911}912 913// A PB is used by RunParallel for running parallel benchmarks.914type PB struct {915	globalN *atomic.Uint64 // shared between all worker goroutines iteration counter916	grain   uint64         // acquire that many iterations from globalN at once917	cache   uint64         // local cache of acquired iterations918	bN      uint64         // total number of iterations to execute (b.N)919}920 921// Next reports whether there are more iterations to execute.922func (pb *PB) Next() bool {923	if pb.cache == 0 {924		n := pb.globalN.Add(pb.grain)925		if n <= pb.bN {926			pb.cache = pb.grain927		} else if n < pb.bN+pb.grain {928			pb.cache = pb.bN + pb.grain - n929		} else {930			return false931		}932	}933	pb.cache--934	return true935}936 937// RunParallel runs a benchmark in parallel.938// It creates multiple goroutines and distributes b.N iterations among them.939// The number of goroutines defaults to GOMAXPROCS. To increase parallelism for940// non-CPU-bound benchmarks, call [B.SetParallelism] before RunParallel.941// RunParallel is usually used with the go test -cpu flag.942//943// The body function will be run in each goroutine. It should set up any944// goroutine-local state and then iterate until pb.Next returns false.945// It should not use the [B.StartTimer], [B.StopTimer], or [B.ResetTimer] functions,946// because they have global effect. It should also not call [B.Run].947//948// RunParallel reports ns/op values as wall time for the benchmark as a whole,949// not the sum of wall time or CPU time over each parallel goroutine.950func (b *B) RunParallel(body func(*PB)) {951	if b.N == 0 {952		return // Nothing to do when probing.953	}954	// Calculate grain size as number of iterations that take ~100µs.955	// 100µs is enough to amortize the overhead and provide sufficient956	// dynamic load balancing.957	grain := uint64(0)958	if b.previousN > 0 && b.previousDuration > 0 {959		grain = 1e5 * uint64(b.previousN) / uint64(b.previousDuration)960	}961	if grain < 1 {962		grain = 1963	}964	// We expect the inner loop and function call to take at least 10ns,965	// so do not do more than 100µs/10ns=1e4 iterations.966	if grain > 1e4 {967		grain = 1e4968	}969 970	var n atomic.Uint64971	numProcs := b.parallelism * runtime.GOMAXPROCS(0)972	var wg sync.WaitGroup973	wg.Add(numProcs)974	for p := 0; p < numProcs; p++ {975		go func() {976			defer wg.Done()977			pb := &PB{978				globalN: &n,979				grain:   grain,980				bN:      uint64(b.N),981			}982			body(pb)983		}()984	}985	wg.Wait()986	if n.Load() <= uint64(b.N) && !b.Failed() {987		b.Fatal("RunParallel: body exited without pb.Next() == false")988	}989}990 991// SetParallelism sets the number of goroutines used by [B.RunParallel] to p*GOMAXPROCS.992// There is usually no need to call SetParallelism for CPU-bound benchmarks.993// If p is less than 1, this call will have no effect.994func (b *B) SetParallelism(p int) {995	if p >= 1 {996		b.parallelism = p997	}998}999 1000// Benchmark benchmarks a single function. It is useful for creating1001// custom benchmarks that do not use the "go test" command.1002//1003// If f depends on testing flags, then [Init] must be used to register1004// those flags before calling Benchmark and before calling [flag.Parse].1005//1006// If f calls Run, the result will be an estimate of running all its1007// subbenchmarks that don't call Run in sequence in a single benchmark.1008func Benchmark(f func(b *B)) BenchmarkResult {1009	b := &B{1010		common: common{1011			signal: make(chan bool),1012			w:      discard{},1013		},1014		benchFunc: f,1015		benchTime: benchTime,1016	}1017	b.setOutputWriter()1018	if b.run1() {1019		b.run()1020	}1021	return b.result1022}1023 1024type discard struct{}1025 1026func (discard) Write(b []byte) (n int, err error) { return len(b), nil }1027