Source file src/testing/benchmark.go

     1  // Copyright 2009 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  package testing
     6  
     7  import (
     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  
    25  func 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  
    31  var (
    32  	matchBenchmarks *string
    33  	benchmarkMemory *bool
    34  
    35  	benchTime = durationOrCountFlag{d: 1 * time.Second} // changed during test of testing package
    36  )
    37  
    38  type durationOrCountFlag struct {
    39  	d         time.Duration
    40  	n         int
    41  	allowZero bool
    42  }
    43  
    44  func (f *durationOrCountFlag) String() string {
    45  	if f.n > 0 {
    46  		return fmt.Sprintf("%dx", f.n)
    47  	}
    48  	return f.d.String()
    49  }
    50  
    51  func (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 nil
    59  	}
    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 nil
    66  }
    67  
    68  // Global lock to ensure only one benchmark runs at a time.
    69  var benchmarkLock sync.Mutex
    70  
    71  // Used for every benchmark for measuring memory.
    72  var memStats runtime.MemStats
    73  
    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.
    76  type InternalBenchmark struct {
    77  	Name string
    78  	F    func(b *B)
    79  }
    80  
    81  // B is a type passed to [Benchmark] functions to manage benchmark
    82  // timing and control the number of iterations.
    83  //
    84  // A benchmark ends when its Benchmark function returns or calls any of the methods
    85  // [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 execution
    91  // and dumped to standard output when done. Unlike in tests, benchmark logs
    92  // are always printed, so as not to hide output whose existence may be
    93  // affecting benchmark results.
    94  type B struct {
    95  	common
    96  	importPath       string // import path of the package containing the benchmark
    97  	bstate           *benchState
    98  	N                int
    99  	previousN        int           // number of iterations in the previous run
   100  	previousDuration time.Duration // total duration of the previous run
   101  	benchFunc        func(b *B)
   102  	benchTime        durationOrCountFlag
   103  	bytes            int64
   104  	missingBytes     bool // one of the subbenchmarks does not have bytes set.
   105  	timerOn          bool
   106  	showAllocResult  bool
   107  	result           BenchmarkResult
   108  	parallelism      int // RunParallel creates parallelism*GOMAXPROCS goroutines
   109  	// The initial states of memStats.Mallocs and memStats.TotalAlloc.
   110  	startAllocs uint64
   111  	startBytes  uint64
   112  	// The net total of this test after being run.
   113  	netAllocs uint64
   114  	netBytes  uint64
   115  	// Extra metrics collected by ReportMetric.
   116  	extra map[string]float64
   117  
   118  	// loop tracks the state of B.Loop
   119  	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 can
   122  		// do reporting based on it, but we avoid exposing it until then.
   123  		n uint64
   124  		// i is the current Loop iteration. It's strictly monotonically
   125  		// increasing toward n.
   126  		//
   127  		// The high bit is used to poison the Loop fast path and fall back to
   128  		// the slow path.
   129  		i uint64
   130  
   131  		done bool // set when B.Loop return false
   132  	}
   133  }
   134  
   135  // StartTimer starts timing a test. This function is called automatically
   136  // before a benchmark starts, but it can also be used to resume timing after
   137  // a call to [B.StopTimer].
   138  func (b *B) StartTimer() {
   139  	if !b.timerOn {
   140  		runtime.ReadMemStats(&memStats)
   141  		b.startAllocs = memStats.Mallocs
   142  		b.startBytes = memStats.TotalAlloc
   143  		b.start = highPrecisionTimeNow()
   144  		b.timerOn = true
   145  		b.loop.i &^= loopPoisonTimer
   146  	}
   147  }
   148  
   149  // StopTimer stops timing a test. This can be used to pause the timer
   150  // while performing steps that you don't want to measure.
   151  func (b *B) StopTimer() {
   152  	if b.timerOn {
   153  		b.duration += highPrecisionTimeSince(b.start)
   154  		runtime.ReadMemStats(&memStats)
   155  		b.netAllocs += memStats.Mallocs - b.startAllocs
   156  		b.netBytes += memStats.TotalAlloc - b.startBytes
   157  		b.timerOn = false
   158  		// If we hit B.Loop with the timer stopped, fail.
   159  		b.loop.i |= loopPoisonTimer
   160  	}
   161  }
   162  
   163  // ResetTimer zeroes the elapsed benchmark time and memory allocation counters
   164  // and deletes user-reported metrics.
   165  // It does not affect whether the timer is running.
   166  func (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.Mallocs
   177  		b.startBytes = memStats.TotalAlloc
   178  		b.start = highPrecisionTimeNow()
   179  	}
   180  	b.duration = 0
   181  	b.netAllocs = 0
   182  	b.netBytes = 0
   183  }
   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.
   187  func (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 the
   191  // benchmark function that calls ReportAllocs.
   192  func (b *B) ReportAllocs() {
   193  	b.showAllocResult = true
   194  }
   195  
   196  // runN runs a single benchmark for the specified number of iterations.
   197  func (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 run
   206  	// by clearing garbage from previous runs.
   207  	runtime.GC()
   208  	b.resetRaces()
   209  	b.N = n
   210  	b.loop.n = 0
   211  	b.loop.i = 0
   212  	b.loop.done = false
   213  	b.ctx = ctx
   214  	b.cancelCtx = cancelCtx
   215  
   216  	b.parallelism = 1
   217  	b.ResetTimer()
   218  	b.StartTimer()
   219  	b.benchFunc(b)
   220  	b.StopTimer()
   221  	b.previousN = n
   222  	b.previousDuration = b.duration
   223  
   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 more
   230  // iterations of this benchmarks should be run.
   231  func (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 normally
   240  		// or by FailNow's runtime.Goexit.
   241  		defer func() {
   242  			b.signal <- true
   243  		}()
   244  
   245  		b.runN(1)
   246  	}()
   247  	<-b.signal
   248  	if b.failed {
   249  		fmt.Fprintf(b.w, "%s--- FAIL: %s\n%s", b.chatty.prefix(), b.name, b.output)
   250  		return false
   251  	}
   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.finished
   256  	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 false
   267  	}
   268  	return true
   269  }
   270  
   271  var labelsOnce sync.Once
   272  
   273  // run executes the benchmark in a separate goroutine, including all of its
   274  // subbenchmarks. b must not have subbenchmarks.
   275  func (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.bench
   288  		b.bstate.processBench(b) // Must call doBench.
   289  	} else {
   290  		// Running func Benchmark.
   291  		b.doBench()
   292  	}
   293  }
   294  
   295  func (b *B) doBench() BenchmarkResult {
   296  	go b.launch()
   297  	<-b.signal
   298  	return b.result
   299  }
   300  
   301  func predictN(goalns int64, prevIters int64, prevns int64, last int64) int {
   302  	if prevns == 0 {
   303  		// Round up to dodge divide by zero. See https://go.dev/issue/70709.
   304  		prevns = 1
   305  	}
   306  
   307  	// Order of operations matters.
   308  	// For very fast benchmarks, prevIters ~= prevns.
   309  	// If you divide first, you get 0 or 1,
   310  	// which can hide an order of magnitude in execution time.
   311  	// So multiply first, then divide.
   312  	n := goalns * prevIters / prevns
   313  	// Run more iterations than we think we'll need (1.2x).
   314  	n += n / 5
   315  	// Don't grow too fast in case we had timing errors previously.
   316  	n = min(n, 100*last)
   317  	// Be sure to run at least one more than last time.
   318  	n = max(n, last+1)
   319  	// Don't run more than 1e9 times. (This also keeps n in int range on 32 bit platforms.)
   320  	n = min(n, 1e9)
   321  	return int(n)
   322  }
   323  
   324  // launch launches the benchmark function. It gradually increases the number
   325  // of benchmark iterations until the benchmark runs for the requested benchtime.
   326  // launch is run by the doBench function as a separate goroutine.
   327  // run1 must have been called on b.
   328  func (b *B) launch() {
   329  	// Signal that we're done whether we return normally
   330  	// or by FailNow's runtime.Goexit.
   331  	defer func() {
   332  		b.signal <- true
   333  	}()
   334  
   335  	// b.Loop does its own ramp-up logic so we just need to run it once.
   336  	// If b.loop.n is non zero, it means b.Loop has already run.
   337  	if b.loop.n == 0 {
   338  		// Run the benchmark for at least the specified amount of time.
   339  		if b.benchTime.n > 0 {
   340  			// We already ran a single iteration in run1.
   341  			// If -benchtime=1x was requested, use that result.
   342  			// See https://golang.org/issue/32051.
   343  			if b.benchTime.n > 1 {
   344  				b.runN(b.benchTime.n)
   345  			}
   346  		} else {
   347  			d := b.benchTime.d
   348  			for n := int64(1); !b.failed && b.duration < d && n < 1e9; {
   349  				last := n
   350  				// Predict required iterations.
   351  				goalns := d.Nanoseconds()
   352  				prevIters := int64(b.N)
   353  				n = int64(predictN(goalns, prevIters, b.duration.Nanoseconds(), last))
   354  				b.runN(int(n))
   355  			}
   356  		}
   357  	}
   358  	b.result = BenchmarkResult{b.N, b.duration, b.bytes, b.netAllocs, b.netBytes, b.extra}
   359  }
   360  
   361  // Elapsed returns the measured elapsed time of the benchmark.
   362  // The duration reported by Elapsed matches the one measured by
   363  // [B.StartTimer], [B.StopTimer], and [B.ResetTimer].
   364  func (b *B) Elapsed() time.Duration {
   365  	d := b.duration
   366  	if b.timerOn {
   367  		d += highPrecisionTimeSince(b.start)
   368  	}
   369  	return d
   370  }
   371  
   372  // ReportMetric adds "n unit" to the reported benchmark results.
   373  // If the metric is per-iteration, the caller should divide by b.N,
   374  // and by convention units should end in "/op".
   375  // ReportMetric overrides any previously reported value for the same unit.
   376  // ReportMetric panics if unit is the empty string or if unit contains
   377  // any whitespace.
   378  // If unit is a unit normally reported by the benchmark framework itself
   379  // (such as "allocs/op"), ReportMetric will override that metric.
   380  // Setting "ns/op" to 0 will suppress that built-in metric.
   381  func (b *B) ReportMetric(n float64, unit string) {
   382  	if unit == "" {
   383  		panic("metric unit must not be empty")
   384  	}
   385  	if strings.IndexFunc(unit, unicode.IsSpace) >= 0 {
   386  		panic("metric unit must not contain whitespace")
   387  	}
   388  	b.extra[unit] = n
   389  }
   390  
   391  func (b *B) stopOrScaleBLoop() bool {
   392  	t := b.Elapsed()
   393  	if t >= b.benchTime.d {
   394  		// We've reached the target
   395  		return false
   396  	}
   397  	// Loop scaling
   398  	goalns := b.benchTime.d.Nanoseconds()
   399  	prevIters := int64(b.loop.n)
   400  	b.loop.n = uint64(predictN(goalns, prevIters, t.Nanoseconds(), prevIters))
   401  	if b.loop.n&loopPoisonMask != 0 {
   402  		// The iteration count should never get this high, but if it did we'd be
   403  		// in big trouble.
   404  		panic("loop iteration target overflow")
   405  	}
   406  	return true
   407  }
   408  
   409  func (b *B) loopSlowPath() bool {
   410  	// Consistency checks
   411  	if !b.timerOn {
   412  		b.Fatal("B.Loop called with timer stopped")
   413  	}
   414  	if b.loop.i&loopPoisonMask != 0 {
   415  		panic(fmt.Sprintf("unknown loop stop condition: %#x", b.loop.i))
   416  	}
   417  
   418  	if b.loop.n == 0 {
   419  		// It's the first call to b.Loop() in the benchmark function.
   420  		if b.benchTime.n > 0 {
   421  			// Fixed iteration count.
   422  			b.loop.n = uint64(b.benchTime.n)
   423  		} else {
   424  			// Initialize target to 1 to kick start loop scaling.
   425  			b.loop.n = 1
   426  		}
   427  		// Within a b.Loop loop, we don't use b.N (to avoid confusion).
   428  		b.N = 0
   429  		b.ResetTimer()
   430  
   431  		// Start the next iteration.
   432  		b.loop.i++
   433  		return true
   434  	}
   435  
   436  	// Should we keep iterating?
   437  	var more bool
   438  	if b.benchTime.n > 0 {
   439  		// The iteration count is fixed, so we should have run this many and now
   440  		// be done.
   441  		if b.loop.i != uint64(b.benchTime.n) {
   442  			// We shouldn't be able to reach the slow path in this case.
   443  			panic(fmt.Sprintf("iteration count %d < fixed target %d", b.loop.i, b.benchTime.n))
   444  		}
   445  		more = false
   446  	} else {
   447  		// Handle fixed time case
   448  		more = b.stopOrScaleBLoop()
   449  	}
   450  	if !more {
   451  		b.StopTimer()
   452  		// Commit iteration count
   453  		b.N = int(b.loop.n)
   454  		b.loop.done = true
   455  		return false
   456  	}
   457  
   458  	// Start the next iteration.
   459  	b.loop.i++
   460  	return true
   461  }
   462  
   463  // Loop returns true as long as the benchmark should continue running.
   464  //
   465  // A typical benchmark is structured like:
   466  //
   467  //	func Benchmark(b *testing.B) {
   468  //		... setup ...
   469  //		for b.Loop() {
   470  //			... code to measure ...
   471  //		}
   472  //		... cleanup ...
   473  //	}
   474  //
   475  // Loop resets the benchmark timer the first time it is called in a benchmark,
   476  // so any setup performed prior to starting the benchmark loop does not count
   477  // toward the benchmark measurement. Likewise, when it returns false, it stops
   478  // the timer so cleanup code is not measured.
   479  //
   480  // Within the body of a "for b.Loop() { ... }" loop, arguments to and
   481  // results from function calls within the loop are kept alive, preventing
   482  // the compiler from fully optimizing away the loop body. Currently, this is
   483  // implemented by disabling inlining of functions called in a b.Loop loop.
   484  // This applies only to calls syntactically between the curly braces of the loop,
   485  // and the loop condition must be written exactly as "b.Loop()". Optimizations
   486  // are performed as usual in any functions called by the loop.
   487  //
   488  // After Loop returns false, b.N contains the total number of iterations that
   489  // ran, so the benchmark may use b.N to compute other average metrics.
   490  //
   491  // Prior to the introduction of Loop, benchmarks were expected to contain an
   492  // explicit loop from 0 to b.N. Benchmarks should either use Loop or contain a
   493  // loop to b.N, but not both. Loop offers more automatic management of the
   494  // benchmark timer, and runs each benchmark function only once per measurement,
   495  // whereas b.N-based benchmarks must run the benchmark function (and any
   496  // associated setup and cleanup) several times.
   497  func (b *B) Loop() bool {
   498  	// This is written such that the fast path is as fast as possible and can be
   499  	// inlined.
   500  	//
   501  	// There are three cases where we'll fall out of the fast path:
   502  	//
   503  	// - On the first call, both i and n are 0.
   504  	//
   505  	// - If the loop reaches the n'th iteration, then i == n and we need
   506  	//   to figure out the new target iteration count or if we're done.
   507  	//
   508  	// - If the timer is stopped, it poisons the top bit of i so the slow
   509  	//   path can do consistency checks and fail.
   510  	if b.loop.i < b.loop.n {
   511  		b.loop.i++
   512  		return true
   513  	}
   514  	return b.loopSlowPath()
   515  }
   516  
   517  // The loopPoison constants can be OR'd into B.loop.i to cause it to fall back
   518  // to the slow path.
   519  const (
   520  	loopPoisonTimer = uint64(1 << (63 - iota))
   521  	// If necessary, add more poison bits here.
   522  
   523  	// loopPoisonMask is the set of all loop poison bits. (iota-1) is the index
   524  	// of the bit we just set, from which we recreate that bit mask. We subtract
   525  	// 1 to set all of the bits below that bit, then complement the result to
   526  	// get the mask. Sorry, not sorry.
   527  	loopPoisonMask = ^uint64((1 << (63 - (iota - 1))) - 1)
   528  )
   529  
   530  // BenchmarkResult contains the results of a benchmark run.
   531  type BenchmarkResult struct {
   532  	N         int           // The number of iterations.
   533  	T         time.Duration // The total time taken.
   534  	Bytes     int64         // Bytes processed in one iteration.
   535  	MemAllocs uint64        // The total number of memory allocations.
   536  	MemBytes  uint64        // The total number of bytes allocated.
   537  
   538  	// Extra records additional metrics reported by ReportMetric.
   539  	Extra map[string]float64
   540  }
   541  
   542  // NsPerOp returns the "ns/op" metric.
   543  func (r BenchmarkResult) NsPerOp() int64 {
   544  	if v, ok := r.Extra["ns/op"]; ok {
   545  		return int64(v)
   546  	}
   547  	if r.N <= 0 {
   548  		return 0
   549  	}
   550  	return r.T.Nanoseconds() / int64(r.N)
   551  }
   552  
   553  // mbPerSec returns the "MB/s" metric.
   554  func (r BenchmarkResult) mbPerSec() float64 {
   555  	if v, ok := r.Extra["MB/s"]; ok {
   556  		return v
   557  	}
   558  	if r.Bytes <= 0 || r.T <= 0 || r.N <= 0 {
   559  		return 0
   560  	}
   561  	return (float64(r.Bytes) * float64(r.N) / 1e6) / r.T.Seconds()
   562  }
   563  
   564  // AllocsPerOp returns the "allocs/op" metric,
   565  // which is calculated as r.MemAllocs / r.N.
   566  func (r BenchmarkResult) AllocsPerOp() int64 {
   567  	if v, ok := r.Extra["allocs/op"]; ok {
   568  		return int64(v)
   569  	}
   570  	if r.N <= 0 {
   571  		return 0
   572  	}
   573  	return int64(r.MemAllocs) / int64(r.N)
   574  }
   575  
   576  // AllocedBytesPerOp returns the "B/op" metric,
   577  // which is calculated as r.MemBytes / r.N.
   578  func (r BenchmarkResult) AllocedBytesPerOp() int64 {
   579  	if v, ok := r.Extra["B/op"]; ok {
   580  		return int64(v)
   581  	}
   582  	if r.N <= 0 {
   583  		return 0
   584  	}
   585  	return int64(r.MemBytes) / int64(r.N)
   586  }
   587  
   588  // String returns a summary of the benchmark results.
   589  // It follows the benchmark result line format from
   590  // https://golang.org/design/14313-benchmark-format, not including the
   591  // benchmark name.
   592  // Extra metrics override built-in metrics of the same name.
   593  // String does not include allocs/op or B/op, since those are reported
   594  // by [BenchmarkResult.MemString].
   595  func (r BenchmarkResult) String() string {
   596  	buf := new(strings.Builder)
   597  	fmt.Fprintf(buf, "%8d", r.N)
   598  
   599  	// Get ns/op as a float.
   600  	ns, ok := r.Extra["ns/op"]
   601  	if !ok {
   602  		ns = float64(r.T.Nanoseconds()) / float64(r.N)
   603  	}
   604  	if ns != 0 {
   605  		buf.WriteByte('\t')
   606  		prettyPrint(buf, ns, "ns/op")
   607  	}
   608  
   609  	if mbs := r.mbPerSec(); mbs != 0 {
   610  		fmt.Fprintf(buf, "\t%7.2f MB/s", mbs)
   611  	}
   612  
   613  	// Print extra metrics that aren't represented in the standard
   614  	// metrics.
   615  	var extraKeys []string
   616  	for k := range r.Extra {
   617  		switch k {
   618  		case "ns/op", "MB/s", "B/op", "allocs/op":
   619  			// Built-in metrics reported elsewhere.
   620  			continue
   621  		}
   622  		extraKeys = append(extraKeys, k)
   623  	}
   624  	slices.Sort(extraKeys)
   625  	for _, k := range extraKeys {
   626  		buf.WriteByte('\t')
   627  		prettyPrint(buf, r.Extra[k], k)
   628  	}
   629  	return buf.String()
   630  }
   631  
   632  func prettyPrint(w io.Writer, x float64, unit string) {
   633  	// Print all numbers with 10 places before the decimal point
   634  	// and small numbers with four sig figs. Field widths are
   635  	// chosen to fit the whole part in 10 places while aligning
   636  	// the decimal point of all fractional formats.
   637  	var format string
   638  	switch y := math.Abs(x); {
   639  	case y == 0 || y >= 999.95:
   640  		format = "%10.0f %s"
   641  	case y >= 99.995:
   642  		format = "%12.1f %s"
   643  	case y >= 9.9995:
   644  		format = "%13.2f %s"
   645  	case y >= 0.99995:
   646  		format = "%14.3f %s"
   647  	case y >= 0.099995:
   648  		format = "%15.4f %s"
   649  	case y >= 0.0099995:
   650  		format = "%16.5f %s"
   651  	case y >= 0.00099995:
   652  		format = "%17.6f %s"
   653  	default:
   654  		format = "%18.7f %s"
   655  	}
   656  	fmt.Fprintf(w, format, x, unit)
   657  }
   658  
   659  // MemString returns r.AllocedBytesPerOp and r.AllocsPerOp in the same format as 'go test'.
   660  func (r BenchmarkResult) MemString() string {
   661  	return fmt.Sprintf("%8d B/op\t%8d allocs/op",
   662  		r.AllocedBytesPerOp(), r.AllocsPerOp())
   663  }
   664  
   665  // benchmarkName returns full name of benchmark including procs suffix.
   666  func benchmarkName(name string, n int) string {
   667  	if n != 1 {
   668  		return fmt.Sprintf("%s-%d", name, n)
   669  	}
   670  	return name
   671  }
   672  
   673  type benchState struct {
   674  	match *matcher
   675  
   676  	maxLen int // The largest recorded benchmark name.
   677  	extLen int // Maximum extension length.
   678  }
   679  
   680  // RunBenchmarks is an internal function but exported because it is cross-package;
   681  // it is part of the implementation of the "go test" command.
   682  func RunBenchmarks(matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) {
   683  	runBenchmarks("", matchString, benchmarks)
   684  }
   685  
   686  func runBenchmarks(importPath string, matchString func(pat, str string) (bool, error), benchmarks []InternalBenchmark) bool {
   687  	// If no flag was specified, don't run benchmarks.
   688  	if len(*matchBenchmarks) == 0 {
   689  		return true
   690  	}
   691  	// Collect matching benchmarks and determine longest name.
   692  	maxprocs := 1
   693  	for _, procs := range cpuList {
   694  		if procs > maxprocs {
   695  			maxprocs = procs
   696  		}
   697  	}
   698  	bstate := &benchState{
   699  		match:  newMatcher(matchString, *matchBenchmarks, "-test.bench", *skip),
   700  		extLen: len(benchmarkName("", maxprocs)),
   701  	}
   702  	var bs []InternalBenchmark
   703  	for _, Benchmark := range benchmarks {
   704  		if _, matched, _ := bstate.match.fullName(nil, Benchmark.Name); matched {
   705  			bs = append(bs, Benchmark)
   706  			benchName := benchmarkName(Benchmark.Name, maxprocs)
   707  			if l := len(benchName) + bstate.extLen + 1; l > bstate.maxLen {
   708  				bstate.maxLen = l
   709  			}
   710  		}
   711  	}
   712  	main := &B{
   713  		common: common{
   714  			name:  "Main",
   715  			w:     os.Stdout,
   716  			bench: true,
   717  		},
   718  		importPath: importPath,
   719  		benchFunc: func(b *B) {
   720  			for _, Benchmark := range bs {
   721  				b.Run(Benchmark.Name, Benchmark.F)
   722  			}
   723  		},
   724  		benchTime: benchTime,
   725  		bstate:    bstate,
   726  	}
   727  	if Verbose() {
   728  		main.chatty = newChattyPrinter(main.w)
   729  	}
   730  	main.runN(1)
   731  	return !main.failed
   732  }
   733  
   734  // processBench runs bench b for the configured CPU counts and prints the results.
   735  func (s *benchState) processBench(b *B) {
   736  	for i, procs := range cpuList {
   737  		for j := uint(0); j < *count; j++ {
   738  			runtime.GOMAXPROCS(procs)
   739  			benchName := benchmarkName(b.name, procs)
   740  
   741  			// If it's chatty, we've already printed this information.
   742  			if b.chatty == nil {
   743  				fmt.Fprintf(b.w, "%-*s\t", s.maxLen, benchName)
   744  			}
   745  			// Recompute the running time for all but the first iteration.
   746  			if i > 0 || j > 0 {
   747  				b = &B{
   748  					common: common{
   749  						signal: make(chan bool),
   750  						name:   b.name,
   751  						w:      b.w,
   752  						chatty: b.chatty,
   753  						bench:  true,
   754  					},
   755  					benchFunc: b.benchFunc,
   756  					benchTime: b.benchTime,
   757  				}
   758  				b.setOutputWriter()
   759  				b.run1()
   760  			}
   761  			r := b.doBench()
   762  			if b.failed {
   763  				// The output could be very long here, but probably isn't.
   764  				// We print it all, regardless, because we don't want to trim the reason
   765  				// the benchmark failed.
   766  				fmt.Fprintf(b.w, "%s--- FAIL: %s\n%s", b.chatty.prefix(), benchName, b.output)
   767  				continue
   768  			}
   769  			results := r.String()
   770  			if b.chatty != nil {
   771  				fmt.Fprintf(b.w, "%-*s\t", s.maxLen, benchName)
   772  			}
   773  			if *benchmarkMemory || b.showAllocResult {
   774  				results += "\t" + r.MemString()
   775  			}
   776  			fmt.Fprintln(b.w, results)
   777  			// Unlike with tests, we ignore the -chatty flag and always print output for
   778  			// benchmarks since the output generation time will skew the results.
   779  			if len(b.output) > 0 {
   780  				b.trimOutput()
   781  				fmt.Fprintf(b.w, "%s--- BENCH: %s\n%s", b.chatty.prefix(), benchName, b.output)
   782  			}
   783  			if p := runtime.GOMAXPROCS(-1); p != procs {
   784  				fmt.Fprintf(os.Stderr, "testing: %s left GOMAXPROCS set to %d\n", benchName, p)
   785  			}
   786  			if b.chatty != nil && b.chatty.json {
   787  				b.chatty.Updatef("", "=== NAME  %s\n", "")
   788  			}
   789  		}
   790  	}
   791  }
   792  
   793  // If hideStdoutForTesting is true, Run does not print the benchName.
   794  // This avoids a spurious print during 'go test' on package testing itself,
   795  // which invokes b.Run in its own tests (see sub_test.go).
   796  var hideStdoutForTesting = false
   797  
   798  // Run benchmarks f as a subbenchmark with the given name. It reports
   799  // whether there were any failures.
   800  //
   801  // A subbenchmark is like any other benchmark. A benchmark that calls Run at
   802  // least once will not be measured itself and will be called once with N=1.
   803  func (b *B) Run(name string, f func(b *B)) bool {
   804  	// Since b has subbenchmarks, we will no longer run it as a benchmark itself.
   805  	// Release the lock and acquire it on exit to ensure locks stay paired.
   806  	b.hasSub.Store(true)
   807  	benchmarkLock.Unlock()
   808  	defer benchmarkLock.Lock()
   809  
   810  	benchName, ok, partial := b.name, true, false
   811  	if b.bstate != nil {
   812  		benchName, ok, partial = b.bstate.match.fullName(&b.common, name)
   813  	}
   814  	if !ok {
   815  		return true
   816  	}
   817  	var pc [maxStackLen]uintptr
   818  	n := runtime.Callers(2, pc[:])
   819  	sub := &B{
   820  		common: common{
   821  			signal:  make(chan bool),
   822  			name:    benchName,
   823  			parent:  &b.common,
   824  			level:   b.level + 1,
   825  			creator: pc[:n],
   826  			w:       b.w,
   827  			chatty:  b.chatty,
   828  			bench:   true,
   829  		},
   830  		importPath: b.importPath,
   831  		benchFunc:  f,
   832  		benchTime:  b.benchTime,
   833  		bstate:     b.bstate,
   834  	}
   835  	sub.setOutputWriter()
   836  	if partial {
   837  		// Partial name match, like -bench=X/Y matching BenchmarkX.
   838  		// Only process sub-benchmarks, if any.
   839  		sub.hasSub.Store(true)
   840  	}
   841  
   842  	if b.chatty != nil {
   843  		labelsOnce.Do(func() {
   844  			fmt.Printf("goos: %s\n", runtime.GOOS)
   845  			fmt.Printf("goarch: %s\n", runtime.GOARCH)
   846  			if b.importPath != "" {
   847  				fmt.Printf("pkg: %s\n", b.importPath)
   848  			}
   849  			if cpu := sysinfo.CPUName(); cpu != "" {
   850  				fmt.Printf("cpu: %s\n", cpu)
   851  			}
   852  		})
   853  
   854  		if !hideStdoutForTesting {
   855  			if b.chatty.json {
   856  				b.chatty.Updatef(benchName, "=== RUN   %s\n", benchName)
   857  			}
   858  			fmt.Println(benchName)
   859  		}
   860  	}
   861  
   862  	if sub.run1() {
   863  		sub.run()
   864  	}
   865  	b.add(sub.result)
   866  	return !sub.failed
   867  }
   868  
   869  // add simulates running benchmarks in sequence in a single iteration. It is
   870  // used to give some meaningful results in case func Benchmark is used in
   871  // combination with Run.
   872  func (b *B) add(other BenchmarkResult) {
   873  	r := &b.result
   874  	// The aggregated BenchmarkResults resemble running all subbenchmarks as
   875  	// in sequence in a single benchmark.
   876  	r.N = 1
   877  	r.T += time.Duration(other.NsPerOp())
   878  	if other.Bytes == 0 {
   879  		// Summing Bytes is meaningless in aggregate if not all subbenchmarks
   880  		// set it.
   881  		b.missingBytes = true
   882  		r.Bytes = 0
   883  	}
   884  	if !b.missingBytes {
   885  		r.Bytes += other.Bytes
   886  	}
   887  	r.MemAllocs += uint64(other.AllocsPerOp())
   888  	r.MemBytes += uint64(other.AllocedBytesPerOp())
   889  }
   890  
   891  // trimOutput shortens the output from a benchmark, which can be very long.
   892  func (b *B) trimOutput() {
   893  	// The output is likely to appear multiple times because the benchmark
   894  	// is run multiple times, but at least it will be seen. This is not a big deal
   895  	// because benchmarks rarely print, but just in case, we trim it if it's too long.
   896  	const maxNewlines = 10
   897  	for nlCount, j := 0, 0; j < len(b.output); j++ {
   898  		if b.output[j] == '\n' {
   899  			nlCount++
   900  			if nlCount >= maxNewlines {
   901  				b.output = append(b.output[:j], "\n\t... [output truncated]\n"...)
   902  				break
   903  			}
   904  		}
   905  	}
   906  }
   907  
   908  // A PB is used by RunParallel for running parallel benchmarks.
   909  type PB struct {
   910  	globalN *atomic.Uint64 // shared between all worker goroutines iteration counter
   911  	grain   uint64         // acquire that many iterations from globalN at once
   912  	cache   uint64         // local cache of acquired iterations
   913  	bN      uint64         // total number of iterations to execute (b.N)
   914  }
   915  
   916  // Next reports whether there are more iterations to execute.
   917  func (pb *PB) Next() bool {
   918  	if pb.cache == 0 {
   919  		n := pb.globalN.Add(pb.grain)
   920  		if n <= pb.bN {
   921  			pb.cache = pb.grain
   922  		} else if n < pb.bN+pb.grain {
   923  			pb.cache = pb.bN + pb.grain - n
   924  		} else {
   925  			return false
   926  		}
   927  	}
   928  	pb.cache--
   929  	return true
   930  }
   931  
   932  // RunParallel runs a benchmark in parallel.
   933  // It creates multiple goroutines and distributes b.N iterations among them.
   934  // The number of goroutines defaults to GOMAXPROCS. To increase parallelism for
   935  // non-CPU-bound benchmarks, call [B.SetParallelism] before RunParallel.
   936  // RunParallel is usually used with the go test -cpu flag.
   937  //
   938  // The body function will be run in each goroutine. It should set up any
   939  // goroutine-local state and then iterate until pb.Next returns false.
   940  // It should not use the [B.StartTimer], [B.StopTimer], or [B.ResetTimer] functions,
   941  // because they have global effect. It should also not call [B.Run].
   942  //
   943  // RunParallel reports ns/op values as wall time for the benchmark as a whole,
   944  // not the sum of wall time or CPU time over each parallel goroutine.
   945  func (b *B) RunParallel(body func(*PB)) {
   946  	if b.N == 0 {
   947  		return // Nothing to do when probing.
   948  	}
   949  	// Calculate grain size as number of iterations that take ~100µs.
   950  	// 100µs is enough to amortize the overhead and provide sufficient
   951  	// dynamic load balancing.
   952  	grain := uint64(0)
   953  	if b.previousN > 0 && b.previousDuration > 0 {
   954  		grain = 1e5 * uint64(b.previousN) / uint64(b.previousDuration)
   955  	}
   956  	if grain < 1 {
   957  		grain = 1
   958  	}
   959  	// We expect the inner loop and function call to take at least 10ns,
   960  	// so do not do more than 100µs/10ns=1e4 iterations.
   961  	if grain > 1e4 {
   962  		grain = 1e4
   963  	}
   964  
   965  	var n atomic.Uint64
   966  	numProcs := b.parallelism * runtime.GOMAXPROCS(0)
   967  	var wg sync.WaitGroup
   968  	wg.Add(numProcs)
   969  	for p := 0; p < numProcs; p++ {
   970  		go func() {
   971  			defer wg.Done()
   972  			pb := &PB{
   973  				globalN: &n,
   974  				grain:   grain,
   975  				bN:      uint64(b.N),
   976  			}
   977  			body(pb)
   978  		}()
   979  	}
   980  	wg.Wait()
   981  	if n.Load() <= uint64(b.N) && !b.Failed() {
   982  		b.Fatal("RunParallel: body exited without pb.Next() == false")
   983  	}
   984  }
   985  
   986  // SetParallelism sets the number of goroutines used by [B.RunParallel] to p*GOMAXPROCS.
   987  // There is usually no need to call SetParallelism for CPU-bound benchmarks.
   988  // If p is less than 1, this call will have no effect.
   989  func (b *B) SetParallelism(p int) {
   990  	if p >= 1 {
   991  		b.parallelism = p
   992  	}
   993  }
   994  
   995  // Benchmark benchmarks a single function. It is useful for creating
   996  // custom benchmarks that do not use the "go test" command.
   997  //
   998  // If f depends on testing flags, then [Init] must be used to register
   999  // those flags before calling Benchmark and before calling [flag.Parse].
  1000  //
  1001  // If f calls Run, the result will be an estimate of running all its
  1002  // subbenchmarks that don't call Run in sequence in a single benchmark.
  1003  func Benchmark(f func(b *B)) BenchmarkResult {
  1004  	b := &B{
  1005  		common: common{
  1006  			signal: make(chan bool),
  1007  			w:      discard{},
  1008  		},
  1009  		benchFunc: f,
  1010  		benchTime: benchTime,
  1011  	}
  1012  	b.setOutputWriter()
  1013  	if b.run1() {
  1014  		b.run()
  1015  	}
  1016  	return b.result
  1017  }
  1018  
  1019  type discard struct{}
  1020  
  1021  func (discard) Write(b []byte) (n int, err error) { return len(b), nil }
  1022  

View as plain text