github.com/lovishpuri/go-40569/src@v0.0.0-20230519171745-f8623e7c56cf/runtime/map.go (about)

     1  // Copyright 2014 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 runtime
     6  
     7  // This file contains the implementation of Go's map type.
     8  //
     9  // A map is just a hash table. The data is arranged
    10  // into an array of buckets. Each bucket contains up to
    11  // 8 key/elem pairs. The low-order bits of the hash are
    12  // used to select a bucket. Each bucket contains a few
    13  // high-order bits of each hash to distinguish the entries
    14  // within a single bucket.
    15  //
    16  // If more than 8 keys hash to a bucket, we chain on
    17  // extra buckets.
    18  //
    19  // When the hashtable grows, we allocate a new array
    20  // of buckets twice as big. Buckets are incrementally
    21  // copied from the old bucket array to the new bucket array.
    22  //
    23  // Map iterators walk through the array of buckets and
    24  // return the keys in walk order (bucket #, then overflow
    25  // chain order, then bucket index).  To maintain iteration
    26  // semantics, we never move keys within their bucket (if
    27  // we did, keys might be returned 0 or 2 times).  When
    28  // growing the table, iterators remain iterating through the
    29  // old table and must check the new table if the bucket
    30  // they are iterating through has been moved ("evacuated")
    31  // to the new table.
    32  
    33  // Picking loadFactor: too large and we have lots of overflow
    34  // buckets, too small and we waste a lot of space. I wrote
    35  // a simple program to check some stats for different loads:
    36  // (64-bit, 8 byte keys and elems)
    37  //  loadFactor    %overflow  bytes/entry     hitprobe    missprobe
    38  //        4.00         2.13        20.77         3.00         4.00
    39  //        4.50         4.05        17.30         3.25         4.50
    40  //        5.00         6.85        14.77         3.50         5.00
    41  //        5.50        10.55        12.94         3.75         5.50
    42  //        6.00        15.27        11.67         4.00         6.00
    43  //        6.50        20.90        10.79         4.25         6.50
    44  //        7.00        27.14        10.15         4.50         7.00
    45  //        7.50        34.03         9.73         4.75         7.50
    46  //        8.00        41.10         9.40         5.00         8.00
    47  //
    48  // %overflow   = percentage of buckets which have an overflow bucket
    49  // bytes/entry = overhead bytes used per key/elem pair
    50  // hitprobe    = # of entries to check when looking up a present key
    51  // missprobe   = # of entries to check when looking up an absent key
    52  //
    53  // Keep in mind this data is for maximally loaded tables, i.e. just
    54  // before the table grows. Typical tables will be somewhat less loaded.
    55  
    56  import (
    57  	"internal/abi"
    58  	"internal/goarch"
    59  	"runtime/internal/atomic"
    60  	"runtime/internal/math"
    61  	"unsafe"
    62  )
    63  
    64  const (
    65  	// Maximum number of key/elem pairs a bucket can hold.
    66  	bucketCntBits = abi.MapBucketCountBits
    67  	bucketCnt     = abi.MapBucketCount
    68  
    69  	// Maximum average load of a bucket that triggers growth is bucketCnt*13/16 (about 80% full)
    70  	// Because of minimum alignment rules, bucketCnt is known to be at least 8.
    71  	// Represent as loadFactorNum/loadFactorDen, to allow integer math.
    72  	loadFactorDen = 2
    73  	loadFactorNum = (bucketCnt * 13 / 16) * loadFactorDen
    74  
    75  	// Maximum key or elem size to keep inline (instead of mallocing per element).
    76  	// Must fit in a uint8.
    77  	// Fast versions cannot handle big elems - the cutoff size for
    78  	// fast versions in cmd/compile/internal/gc/walk.go must be at most this elem.
    79  	maxKeySize  = abi.MapMaxKeyBytes
    80  	maxElemSize = abi.MapMaxElemBytes
    81  
    82  	// data offset should be the size of the bmap struct, but needs to be
    83  	// aligned correctly. For amd64p32 this means 64-bit alignment
    84  	// even though pointers are 32 bit.
    85  	dataOffset = unsafe.Offsetof(struct {
    86  		b bmap
    87  		v int64
    88  	}{}.v)
    89  
    90  	// Possible tophash values. We reserve a few possibilities for special marks.
    91  	// Each bucket (including its overflow buckets, if any) will have either all or none of its
    92  	// entries in the evacuated* states (except during the evacuate() method, which only happens
    93  	// during map writes and thus no one else can observe the map during that time).
    94  	emptyRest      = 0 // this cell is empty, and there are no more non-empty cells at higher indexes or overflows.
    95  	emptyOne       = 1 // this cell is empty
    96  	evacuatedX     = 2 // key/elem is valid.  Entry has been evacuated to first half of larger table.
    97  	evacuatedY     = 3 // same as above, but evacuated to second half of larger table.
    98  	evacuatedEmpty = 4 // cell is empty, bucket is evacuated.
    99  	minTopHash     = 5 // minimum tophash for a normal filled cell.
   100  
   101  	// flags
   102  	iterator     = 1 // there may be an iterator using buckets
   103  	oldIterator  = 2 // there may be an iterator using oldbuckets
   104  	hashWriting  = 4 // a goroutine is writing to the map
   105  	sameSizeGrow = 8 // the current map growth is to a new map of the same size
   106  
   107  	// sentinel bucket ID for iterator checks
   108  	noCheck = 1<<(8*goarch.PtrSize) - 1
   109  )
   110  
   111  // isEmpty reports whether the given tophash array entry represents an empty bucket entry.
   112  func isEmpty(x uint8) bool {
   113  	return x <= emptyOne
   114  }
   115  
   116  // A header for a Go map.
   117  type hmap struct {
   118  	// Note: the format of the hmap is also encoded in cmd/compile/internal/reflectdata/reflect.go.
   119  	// Make sure this stays in sync with the compiler's definition.
   120  	count     int // # live cells == size of map.  Must be first (used by len() builtin)
   121  	flags     uint8
   122  	B         uint8  // log_2 of # of buckets (can hold up to loadFactor * 2^B items)
   123  	noverflow uint16 // approximate number of overflow buckets; see incrnoverflow for details
   124  	hash0     uint32 // hash seed
   125  
   126  	buckets    unsafe.Pointer // array of 2^B Buckets. may be nil if count==0.
   127  	oldbuckets unsafe.Pointer // previous bucket array of half the size, non-nil only when growing
   128  	nevacuate  uintptr        // progress counter for evacuation (buckets less than this have been evacuated)
   129  
   130  	extra *mapextra // optional fields
   131  }
   132  
   133  // mapextra holds fields that are not present on all maps.
   134  type mapextra struct {
   135  	// If both key and elem do not contain pointers and are inline, then we mark bucket
   136  	// type as containing no pointers. This avoids scanning such maps.
   137  	// However, bmap.overflow is a pointer. In order to keep overflow buckets
   138  	// alive, we store pointers to all overflow buckets in hmap.extra.overflow and hmap.extra.oldoverflow.
   139  	// overflow and oldoverflow are only used if key and elem do not contain pointers.
   140  	// overflow contains overflow buckets for hmap.buckets.
   141  	// oldoverflow contains overflow buckets for hmap.oldbuckets.
   142  	// The indirection allows to store a pointer to the slice in hiter.
   143  	overflow    *[]*bmap
   144  	oldoverflow *[]*bmap
   145  
   146  	// nextOverflow holds a pointer to a free overflow bucket.
   147  	nextOverflow *bmap
   148  }
   149  
   150  // A bucket for a Go map.
   151  type bmap struct {
   152  	// tophash generally contains the top byte of the hash value
   153  	// for each key in this bucket. If tophash[0] < minTopHash,
   154  	// tophash[0] is a bucket evacuation state instead.
   155  	tophash [bucketCnt]uint8
   156  	// Followed by bucketCnt keys and then bucketCnt elems.
   157  	// NOTE: packing all the keys together and then all the elems together makes the
   158  	// code a bit more complicated than alternating key/elem/key/elem/... but it allows
   159  	// us to eliminate padding which would be needed for, e.g., map[int64]int8.
   160  	// Followed by an overflow pointer.
   161  }
   162  
   163  // A hash iteration structure.
   164  // If you modify hiter, also change cmd/compile/internal/reflectdata/reflect.go
   165  // and reflect/value.go to match the layout of this structure.
   166  type hiter struct {
   167  	key         unsafe.Pointer // Must be in first position.  Write nil to indicate iteration end (see cmd/compile/internal/walk/range.go).
   168  	elem        unsafe.Pointer // Must be in second position (see cmd/compile/internal/walk/range.go).
   169  	t           *maptype
   170  	h           *hmap
   171  	buckets     unsafe.Pointer // bucket ptr at hash_iter initialization time
   172  	bptr        *bmap          // current bucket
   173  	overflow    *[]*bmap       // keeps overflow buckets of hmap.buckets alive
   174  	oldoverflow *[]*bmap       // keeps overflow buckets of hmap.oldbuckets alive
   175  	startBucket uintptr        // bucket iteration started at
   176  	offset      uint8          // intra-bucket offset to start from during iteration (should be big enough to hold bucketCnt-1)
   177  	wrapped     bool           // already wrapped around from end of bucket array to beginning
   178  	B           uint8
   179  	i           uint8
   180  	bucket      uintptr
   181  	checkBucket uintptr
   182  }
   183  
   184  // bucketShift returns 1<<b, optimized for code generation.
   185  func bucketShift(b uint8) uintptr {
   186  	// Masking the shift amount allows overflow checks to be elided.
   187  	return uintptr(1) << (b & (goarch.PtrSize*8 - 1))
   188  }
   189  
   190  // bucketMask returns 1<<b - 1, optimized for code generation.
   191  func bucketMask(b uint8) uintptr {
   192  	return bucketShift(b) - 1
   193  }
   194  
   195  // tophash calculates the tophash value for hash.
   196  func tophash(hash uintptr) uint8 {
   197  	top := uint8(hash >> (goarch.PtrSize*8 - 8))
   198  	if top < minTopHash {
   199  		top += minTopHash
   200  	}
   201  	return top
   202  }
   203  
   204  func evacuated(b *bmap) bool {
   205  	h := b.tophash[0]
   206  	return h > emptyOne && h < minTopHash
   207  }
   208  
   209  func (b *bmap) overflow(t *maptype) *bmap {
   210  	return *(**bmap)(add(unsafe.Pointer(b), uintptr(t.BucketSize)-goarch.PtrSize))
   211  }
   212  
   213  func (b *bmap) setoverflow(t *maptype, ovf *bmap) {
   214  	*(**bmap)(add(unsafe.Pointer(b), uintptr(t.BucketSize)-goarch.PtrSize)) = ovf
   215  }
   216  
   217  func (b *bmap) keys() unsafe.Pointer {
   218  	return add(unsafe.Pointer(b), dataOffset)
   219  }
   220  
   221  // incrnoverflow increments h.noverflow.
   222  // noverflow counts the number of overflow buckets.
   223  // This is used to trigger same-size map growth.
   224  // See also tooManyOverflowBuckets.
   225  // To keep hmap small, noverflow is a uint16.
   226  // When there are few buckets, noverflow is an exact count.
   227  // When there are many buckets, noverflow is an approximate count.
   228  func (h *hmap) incrnoverflow() {
   229  	// We trigger same-size map growth if there are
   230  	// as many overflow buckets as buckets.
   231  	// We need to be able to count to 1<<h.B.
   232  	if h.B < 16 {
   233  		h.noverflow++
   234  		return
   235  	}
   236  	// Increment with probability 1/(1<<(h.B-15)).
   237  	// When we reach 1<<15 - 1, we will have approximately
   238  	// as many overflow buckets as buckets.
   239  	mask := uint32(1)<<(h.B-15) - 1
   240  	// Example: if h.B == 18, then mask == 7,
   241  	// and fastrand & 7 == 0 with probability 1/8.
   242  	if fastrand()&mask == 0 {
   243  		h.noverflow++
   244  	}
   245  }
   246  
   247  func (h *hmap) newoverflow(t *maptype, b *bmap) *bmap {
   248  	var ovf *bmap
   249  	if h.extra != nil && h.extra.nextOverflow != nil {
   250  		// We have preallocated overflow buckets available.
   251  		// See makeBucketArray for more details.
   252  		ovf = h.extra.nextOverflow
   253  		if ovf.overflow(t) == nil {
   254  			// We're not at the end of the preallocated overflow buckets. Bump the pointer.
   255  			h.extra.nextOverflow = (*bmap)(add(unsafe.Pointer(ovf), uintptr(t.BucketSize)))
   256  		} else {
   257  			// This is the last preallocated overflow bucket.
   258  			// Reset the overflow pointer on this bucket,
   259  			// which was set to a non-nil sentinel value.
   260  			ovf.setoverflow(t, nil)
   261  			h.extra.nextOverflow = nil
   262  		}
   263  	} else {
   264  		ovf = (*bmap)(newobject(t.Bucket))
   265  	}
   266  	h.incrnoverflow()
   267  	if t.Bucket.PtrBytes == 0 {
   268  		h.createOverflow()
   269  		*h.extra.overflow = append(*h.extra.overflow, ovf)
   270  	}
   271  	b.setoverflow(t, ovf)
   272  	return ovf
   273  }
   274  
   275  func (h *hmap) createOverflow() {
   276  	if h.extra == nil {
   277  		h.extra = new(mapextra)
   278  	}
   279  	if h.extra.overflow == nil {
   280  		h.extra.overflow = new([]*bmap)
   281  	}
   282  }
   283  
   284  func makemap64(t *maptype, hint int64, h *hmap) *hmap {
   285  	if int64(int(hint)) != hint {
   286  		hint = 0
   287  	}
   288  	return makemap(t, int(hint), h)
   289  }
   290  
   291  // makemap_small implements Go map creation for make(map[k]v) and
   292  // make(map[k]v, hint) when hint is known to be at most bucketCnt
   293  // at compile time and the map needs to be allocated on the heap.
   294  func makemap_small() *hmap {
   295  	h := new(hmap)
   296  	h.hash0 = fastrand()
   297  	return h
   298  }
   299  
   300  // makemap implements Go map creation for make(map[k]v, hint).
   301  // If the compiler has determined that the map or the first bucket
   302  // can be created on the stack, h and/or bucket may be non-nil.
   303  // If h != nil, the map can be created directly in h.
   304  // If h.buckets != nil, bucket pointed to can be used as the first bucket.
   305  func makemap(t *maptype, hint int, h *hmap) *hmap {
   306  	mem, overflow := math.MulUintptr(uintptr(hint), t.Bucket.Size_)
   307  	if overflow || mem > maxAlloc {
   308  		hint = 0
   309  	}
   310  
   311  	// initialize Hmap
   312  	if h == nil {
   313  		h = new(hmap)
   314  	}
   315  	h.hash0 = fastrand()
   316  
   317  	// Find the size parameter B which will hold the requested # of elements.
   318  	// For hint < 0 overLoadFactor returns false since hint < bucketCnt.
   319  	B := uint8(0)
   320  	for overLoadFactor(hint, B) {
   321  		B++
   322  	}
   323  	h.B = B
   324  
   325  	// allocate initial hash table
   326  	// if B == 0, the buckets field is allocated lazily later (in mapassign)
   327  	// If hint is large zeroing this memory could take a while.
   328  	if h.B != 0 {
   329  		var nextOverflow *bmap
   330  		h.buckets, nextOverflow = makeBucketArray(t, h.B, nil)
   331  		if nextOverflow != nil {
   332  			h.extra = new(mapextra)
   333  			h.extra.nextOverflow = nextOverflow
   334  		}
   335  	}
   336  
   337  	return h
   338  }
   339  
   340  // makeBucketArray initializes a backing array for map buckets.
   341  // 1<<b is the minimum number of buckets to allocate.
   342  // dirtyalloc should either be nil or a bucket array previously
   343  // allocated by makeBucketArray with the same t and b parameters.
   344  // If dirtyalloc is nil a new backing array will be alloced and
   345  // otherwise dirtyalloc will be cleared and reused as backing array.
   346  func makeBucketArray(t *maptype, b uint8, dirtyalloc unsafe.Pointer) (buckets unsafe.Pointer, nextOverflow *bmap) {
   347  	base := bucketShift(b)
   348  	nbuckets := base
   349  	// For small b, overflow buckets are unlikely.
   350  	// Avoid the overhead of the calculation.
   351  	if b >= 4 {
   352  		// Add on the estimated number of overflow buckets
   353  		// required to insert the median number of elements
   354  		// used with this value of b.
   355  		nbuckets += bucketShift(b - 4)
   356  		sz := t.Bucket.Size_ * nbuckets
   357  		up := roundupsize(sz)
   358  		if up != sz {
   359  			nbuckets = up / t.Bucket.Size_
   360  		}
   361  	}
   362  
   363  	if dirtyalloc == nil {
   364  		buckets = newarray(t.Bucket, int(nbuckets))
   365  	} else {
   366  		// dirtyalloc was previously generated by
   367  		// the above newarray(t.Bucket, int(nbuckets))
   368  		// but may not be empty.
   369  		buckets = dirtyalloc
   370  		size := t.Bucket.Size_ * nbuckets
   371  		if t.Bucket.PtrBytes != 0 {
   372  			memclrHasPointers(buckets, size)
   373  		} else {
   374  			memclrNoHeapPointers(buckets, size)
   375  		}
   376  	}
   377  
   378  	if base != nbuckets {
   379  		// We preallocated some overflow buckets.
   380  		// To keep the overhead of tracking these overflow buckets to a minimum,
   381  		// we use the convention that if a preallocated overflow bucket's overflow
   382  		// pointer is nil, then there are more available by bumping the pointer.
   383  		// We need a safe non-nil pointer for the last overflow bucket; just use buckets.
   384  		nextOverflow = (*bmap)(add(buckets, base*uintptr(t.BucketSize)))
   385  		last := (*bmap)(add(buckets, (nbuckets-1)*uintptr(t.BucketSize)))
   386  		last.setoverflow(t, (*bmap)(buckets))
   387  	}
   388  	return buckets, nextOverflow
   389  }
   390  
   391  // mapaccess1 returns a pointer to h[key].  Never returns nil, instead
   392  // it will return a reference to the zero object for the elem type if
   393  // the key is not in the map.
   394  // NOTE: The returned pointer may keep the whole map live, so don't
   395  // hold onto it for very long.
   396  func mapaccess1(t *maptype, h *hmap, key unsafe.Pointer) unsafe.Pointer {
   397  	if raceenabled && h != nil {
   398  		callerpc := getcallerpc()
   399  		pc := abi.FuncPCABIInternal(mapaccess1)
   400  		racereadpc(unsafe.Pointer(h), callerpc, pc)
   401  		raceReadObjectPC(t.Key, key, callerpc, pc)
   402  	}
   403  	if msanenabled && h != nil {
   404  		msanread(key, t.Key.Size_)
   405  	}
   406  	if asanenabled && h != nil {
   407  		asanread(key, t.Key.Size_)
   408  	}
   409  	if h == nil || h.count == 0 {
   410  		if t.HashMightPanic() {
   411  			t.Hasher(key, 0) // see issue 23734
   412  		}
   413  		return unsafe.Pointer(&zeroVal[0])
   414  	}
   415  	if h.flags&hashWriting != 0 {
   416  		fatal("concurrent map read and map write")
   417  	}
   418  	hash := t.Hasher(key, uintptr(h.hash0))
   419  	m := bucketMask(h.B)
   420  	b := (*bmap)(add(h.buckets, (hash&m)*uintptr(t.BucketSize)))
   421  	if c := h.oldbuckets; c != nil {
   422  		if !h.sameSizeGrow() {
   423  			// There used to be half as many buckets; mask down one more power of two.
   424  			m >>= 1
   425  		}
   426  		oldb := (*bmap)(add(c, (hash&m)*uintptr(t.BucketSize)))
   427  		if !evacuated(oldb) {
   428  			b = oldb
   429  		}
   430  	}
   431  	top := tophash(hash)
   432  bucketloop:
   433  	for ; b != nil; b = b.overflow(t) {
   434  		for i := uintptr(0); i < bucketCnt; i++ {
   435  			if b.tophash[i] != top {
   436  				if b.tophash[i] == emptyRest {
   437  					break bucketloop
   438  				}
   439  				continue
   440  			}
   441  			k := add(unsafe.Pointer(b), dataOffset+i*uintptr(t.KeySize))
   442  			if t.IndirectKey() {
   443  				k = *((*unsafe.Pointer)(k))
   444  			}
   445  			if t.Key.Equal(key, k) {
   446  				e := add(unsafe.Pointer(b), dataOffset+bucketCnt*uintptr(t.KeySize)+i*uintptr(t.ValueSize))
   447  				if t.IndirectElem() {
   448  					e = *((*unsafe.Pointer)(e))
   449  				}
   450  				return e
   451  			}
   452  		}
   453  	}
   454  	return unsafe.Pointer(&zeroVal[0])
   455  }
   456  
   457  func mapaccess2(t *maptype, h *hmap, key unsafe.Pointer) (unsafe.Pointer, bool) {
   458  	if raceenabled && h != nil {
   459  		callerpc := getcallerpc()
   460  		pc := abi.FuncPCABIInternal(mapaccess2)
   461  		racereadpc(unsafe.Pointer(h), callerpc, pc)
   462  		raceReadObjectPC(t.Key, key, callerpc, pc)
   463  	}
   464  	if msanenabled && h != nil {
   465  		msanread(key, t.Key.Size_)
   466  	}
   467  	if asanenabled && h != nil {
   468  		asanread(key, t.Key.Size_)
   469  	}
   470  	if h == nil || h.count == 0 {
   471  		if t.HashMightPanic() {
   472  			t.Hasher(key, 0) // see issue 23734
   473  		}
   474  		return unsafe.Pointer(&zeroVal[0]), false
   475  	}
   476  	if h.flags&hashWriting != 0 {
   477  		fatal("concurrent map read and map write")
   478  	}
   479  	hash := t.Hasher(key, uintptr(h.hash0))
   480  	m := bucketMask(h.B)
   481  	b := (*bmap)(add(h.buckets, (hash&m)*uintptr(t.BucketSize)))
   482  	if c := h.oldbuckets; c != nil {
   483  		if !h.sameSizeGrow() {
   484  			// There used to be half as many buckets; mask down one more power of two.
   485  			m >>= 1
   486  		}
   487  		oldb := (*bmap)(add(c, (hash&m)*uintptr(t.BucketSize)))
   488  		if !evacuated(oldb) {
   489  			b = oldb
   490  		}
   491  	}
   492  	top := tophash(hash)
   493  bucketloop:
   494  	for ; b != nil; b = b.overflow(t) {
   495  		for i := uintptr(0); i < bucketCnt; i++ {
   496  			if b.tophash[i] != top {
   497  				if b.tophash[i] == emptyRest {
   498  					break bucketloop
   499  				}
   500  				continue
   501  			}
   502  			k := add(unsafe.Pointer(b), dataOffset+i*uintptr(t.KeySize))
   503  			if t.IndirectKey() {
   504  				k = *((*unsafe.Pointer)(k))
   505  			}
   506  			if t.Key.Equal(key, k) {
   507  				e := add(unsafe.Pointer(b), dataOffset+bucketCnt*uintptr(t.KeySize)+i*uintptr(t.ValueSize))
   508  				if t.IndirectElem() {
   509  					e = *((*unsafe.Pointer)(e))
   510  				}
   511  				return e, true
   512  			}
   513  		}
   514  	}
   515  	return unsafe.Pointer(&zeroVal[0]), false
   516  }
   517  
   518  // returns both key and elem. Used by map iterator.
   519  func mapaccessK(t *maptype, h *hmap, key unsafe.Pointer) (unsafe.Pointer, unsafe.Pointer) {
   520  	if h == nil || h.count == 0 {
   521  		return nil, nil
   522  	}
   523  	hash := t.Hasher(key, uintptr(h.hash0))
   524  	m := bucketMask(h.B)
   525  	b := (*bmap)(add(h.buckets, (hash&m)*uintptr(t.BucketSize)))
   526  	if c := h.oldbuckets; c != nil {
   527  		if !h.sameSizeGrow() {
   528  			// There used to be half as many buckets; mask down one more power of two.
   529  			m >>= 1
   530  		}
   531  		oldb := (*bmap)(add(c, (hash&m)*uintptr(t.BucketSize)))
   532  		if !evacuated(oldb) {
   533  			b = oldb
   534  		}
   535  	}
   536  	top := tophash(hash)
   537  bucketloop:
   538  	for ; b != nil; b = b.overflow(t) {
   539  		for i := uintptr(0); i < bucketCnt; i++ {
   540  			if b.tophash[i] != top {
   541  				if b.tophash[i] == emptyRest {
   542  					break bucketloop
   543  				}
   544  				continue
   545  			}
   546  			k := add(unsafe.Pointer(b), dataOffset+i*uintptr(t.KeySize))
   547  			if t.IndirectKey() {
   548  				k = *((*unsafe.Pointer)(k))
   549  			}
   550  			if t.Key.Equal(key, k) {
   551  				e := add(unsafe.Pointer(b), dataOffset+bucketCnt*uintptr(t.KeySize)+i*uintptr(t.ValueSize))
   552  				if t.IndirectElem() {
   553  					e = *((*unsafe.Pointer)(e))
   554  				}
   555  				return k, e
   556  			}
   557  		}
   558  	}
   559  	return nil, nil
   560  }
   561  
   562  func mapaccess1_fat(t *maptype, h *hmap, key, zero unsafe.Pointer) unsafe.Pointer {
   563  	e := mapaccess1(t, h, key)
   564  	if e == unsafe.Pointer(&zeroVal[0]) {
   565  		return zero
   566  	}
   567  	return e
   568  }
   569  
   570  func mapaccess2_fat(t *maptype, h *hmap, key, zero unsafe.Pointer) (unsafe.Pointer, bool) {
   571  	e := mapaccess1(t, h, key)
   572  	if e == unsafe.Pointer(&zeroVal[0]) {
   573  		return zero, false
   574  	}
   575  	return e, true
   576  }
   577  
   578  // Like mapaccess, but allocates a slot for the key if it is not present in the map.
   579  func mapassign(t *maptype, h *hmap, key unsafe.Pointer) unsafe.Pointer {
   580  	if h == nil {
   581  		panic(plainError("assignment to entry in nil map"))
   582  	}
   583  	if raceenabled {
   584  		callerpc := getcallerpc()
   585  		pc := abi.FuncPCABIInternal(mapassign)
   586  		racewritepc(unsafe.Pointer(h), callerpc, pc)
   587  		raceReadObjectPC(t.Key, key, callerpc, pc)
   588  	}
   589  	if msanenabled {
   590  		msanread(key, t.Key.Size_)
   591  	}
   592  	if asanenabled {
   593  		asanread(key, t.Key.Size_)
   594  	}
   595  	if h.flags&hashWriting != 0 {
   596  		fatal("concurrent map writes")
   597  	}
   598  	hash := t.Hasher(key, uintptr(h.hash0))
   599  
   600  	// Set hashWriting after calling t.hasher, since t.hasher may panic,
   601  	// in which case we have not actually done a write.
   602  	h.flags ^= hashWriting
   603  
   604  	if h.buckets == nil {
   605  		h.buckets = newobject(t.Bucket) // newarray(t.Bucket, 1)
   606  	}
   607  
   608  again:
   609  	bucket := hash & bucketMask(h.B)
   610  	if h.growing() {
   611  		growWork(t, h, bucket)
   612  	}
   613  	b := (*bmap)(add(h.buckets, bucket*uintptr(t.BucketSize)))
   614  	top := tophash(hash)
   615  
   616  	var inserti *uint8
   617  	var insertk unsafe.Pointer
   618  	var elem unsafe.Pointer
   619  bucketloop:
   620  	for {
   621  		for i := uintptr(0); i < bucketCnt; i++ {
   622  			if b.tophash[i] != top {
   623  				if isEmpty(b.tophash[i]) && inserti == nil {
   624  					inserti = &b.tophash[i]
   625  					insertk = add(unsafe.Pointer(b), dataOffset+i*uintptr(t.KeySize))
   626  					elem = add(unsafe.Pointer(b), dataOffset+bucketCnt*uintptr(t.KeySize)+i*uintptr(t.ValueSize))
   627  				}
   628  				if b.tophash[i] == emptyRest {
   629  					break bucketloop
   630  				}
   631  				continue
   632  			}
   633  			k := add(unsafe.Pointer(b), dataOffset+i*uintptr(t.KeySize))
   634  			if t.IndirectKey() {
   635  				k = *((*unsafe.Pointer)(k))
   636  			}
   637  			if !t.Key.Equal(key, k) {
   638  				continue
   639  			}
   640  			// already have a mapping for key. Update it.
   641  			if t.NeedKeyUpdate() {
   642  				typedmemmove(t.Key, k, key)
   643  			}
   644  			elem = add(unsafe.Pointer(b), dataOffset+bucketCnt*uintptr(t.KeySize)+i*uintptr(t.ValueSize))
   645  			goto done
   646  		}
   647  		ovf := b.overflow(t)
   648  		if ovf == nil {
   649  			break
   650  		}
   651  		b = ovf
   652  	}
   653  
   654  	// Did not find mapping for key. Allocate new cell & add entry.
   655  
   656  	// If we hit the max load factor or we have too many overflow buckets,
   657  	// and we're not already in the middle of growing, start growing.
   658  	if !h.growing() && (overLoadFactor(h.count+1, h.B) || tooManyOverflowBuckets(h.noverflow, h.B)) {
   659  		hashGrow(t, h)
   660  		goto again // Growing the table invalidates everything, so try again
   661  	}
   662  
   663  	if inserti == nil {
   664  		// The current bucket and all the overflow buckets connected to it are full, allocate a new one.
   665  		newb := h.newoverflow(t, b)
   666  		inserti = &newb.tophash[0]
   667  		insertk = add(unsafe.Pointer(newb), dataOffset)
   668  		elem = add(insertk, bucketCnt*uintptr(t.KeySize))
   669  	}
   670  
   671  	// store new key/elem at insert position
   672  	if t.IndirectKey() {
   673  		kmem := newobject(t.Key)
   674  		*(*unsafe.Pointer)(insertk) = kmem
   675  		insertk = kmem
   676  	}
   677  	if t.IndirectElem() {
   678  		vmem := newobject(t.Elem)
   679  		*(*unsafe.Pointer)(elem) = vmem
   680  	}
   681  	typedmemmove(t.Key, insertk, key)
   682  	*inserti = top
   683  	h.count++
   684  
   685  done:
   686  	if h.flags&hashWriting == 0 {
   687  		fatal("concurrent map writes")
   688  	}
   689  	h.flags &^= hashWriting
   690  	if t.IndirectElem() {
   691  		elem = *((*unsafe.Pointer)(elem))
   692  	}
   693  	return elem
   694  }
   695  
   696  func mapdelete(t *maptype, h *hmap, key unsafe.Pointer) {
   697  	if raceenabled && h != nil {
   698  		callerpc := getcallerpc()
   699  		pc := abi.FuncPCABIInternal(mapdelete)
   700  		racewritepc(unsafe.Pointer(h), callerpc, pc)
   701  		raceReadObjectPC(t.Key, key, callerpc, pc)
   702  	}
   703  	if msanenabled && h != nil {
   704  		msanread(key, t.Key.Size_)
   705  	}
   706  	if asanenabled && h != nil {
   707  		asanread(key, t.Key.Size_)
   708  	}
   709  	if h == nil || h.count == 0 {
   710  		if t.HashMightPanic() {
   711  			t.Hasher(key, 0) // see issue 23734
   712  		}
   713  		return
   714  	}
   715  	if h.flags&hashWriting != 0 {
   716  		fatal("concurrent map writes")
   717  	}
   718  
   719  	hash := t.Hasher(key, uintptr(h.hash0))
   720  
   721  	// Set hashWriting after calling t.hasher, since t.hasher may panic,
   722  	// in which case we have not actually done a write (delete).
   723  	h.flags ^= hashWriting
   724  
   725  	bucket := hash & bucketMask(h.B)
   726  	if h.growing() {
   727  		growWork(t, h, bucket)
   728  	}
   729  	b := (*bmap)(add(h.buckets, bucket*uintptr(t.BucketSize)))
   730  	bOrig := b
   731  	top := tophash(hash)
   732  search:
   733  	for ; b != nil; b = b.overflow(t) {
   734  		for i := uintptr(0); i < bucketCnt; i++ {
   735  			if b.tophash[i] != top {
   736  				if b.tophash[i] == emptyRest {
   737  					break search
   738  				}
   739  				continue
   740  			}
   741  			k := add(unsafe.Pointer(b), dataOffset+i*uintptr(t.KeySize))
   742  			k2 := k
   743  			if t.IndirectKey() {
   744  				k2 = *((*unsafe.Pointer)(k2))
   745  			}
   746  			if !t.Key.Equal(key, k2) {
   747  				continue
   748  			}
   749  			// Only clear key if there are pointers in it.
   750  			if t.IndirectKey() {
   751  				*(*unsafe.Pointer)(k) = nil
   752  			} else if t.Key.PtrBytes != 0 {
   753  				memclrHasPointers(k, t.Key.Size_)
   754  			}
   755  			e := add(unsafe.Pointer(b), dataOffset+bucketCnt*uintptr(t.KeySize)+i*uintptr(t.ValueSize))
   756  			if t.IndirectElem() {
   757  				*(*unsafe.Pointer)(e) = nil
   758  			} else if t.Elem.PtrBytes != 0 {
   759  				memclrHasPointers(e, t.Elem.Size_)
   760  			} else {
   761  				memclrNoHeapPointers(e, t.Elem.Size_)
   762  			}
   763  			b.tophash[i] = emptyOne
   764  			// If the bucket now ends in a bunch of emptyOne states,
   765  			// change those to emptyRest states.
   766  			// It would be nice to make this a separate function, but
   767  			// for loops are not currently inlineable.
   768  			if i == bucketCnt-1 {
   769  				if b.overflow(t) != nil && b.overflow(t).tophash[0] != emptyRest {
   770  					goto notLast
   771  				}
   772  			} else {
   773  				if b.tophash[i+1] != emptyRest {
   774  					goto notLast
   775  				}
   776  			}
   777  			for {
   778  				b.tophash[i] = emptyRest
   779  				if i == 0 {
   780  					if b == bOrig {
   781  						break // beginning of initial bucket, we're done.
   782  					}
   783  					// Find previous bucket, continue at its last entry.
   784  					c := b
   785  					for b = bOrig; b.overflow(t) != c; b = b.overflow(t) {
   786  					}
   787  					i = bucketCnt - 1
   788  				} else {
   789  					i--
   790  				}
   791  				if b.tophash[i] != emptyOne {
   792  					break
   793  				}
   794  			}
   795  		notLast:
   796  			h.count--
   797  			// Reset the hash seed to make it more difficult for attackers to
   798  			// repeatedly trigger hash collisions. See issue 25237.
   799  			if h.count == 0 {
   800  				h.hash0 = fastrand()
   801  			}
   802  			break search
   803  		}
   804  	}
   805  
   806  	if h.flags&hashWriting == 0 {
   807  		fatal("concurrent map writes")
   808  	}
   809  	h.flags &^= hashWriting
   810  }
   811  
   812  // mapiterinit initializes the hiter struct used for ranging over maps.
   813  // The hiter struct pointed to by 'it' is allocated on the stack
   814  // by the compilers order pass or on the heap by reflect_mapiterinit.
   815  // Both need to have zeroed hiter since the struct contains pointers.
   816  func mapiterinit(t *maptype, h *hmap, it *hiter) {
   817  	if raceenabled && h != nil {
   818  		callerpc := getcallerpc()
   819  		racereadpc(unsafe.Pointer(h), callerpc, abi.FuncPCABIInternal(mapiterinit))
   820  	}
   821  
   822  	it.t = t
   823  	if h == nil || h.count == 0 {
   824  		return
   825  	}
   826  
   827  	if unsafe.Sizeof(hiter{})/goarch.PtrSize != 12 {
   828  		throw("hash_iter size incorrect") // see cmd/compile/internal/reflectdata/reflect.go
   829  	}
   830  	it.h = h
   831  
   832  	// grab snapshot of bucket state
   833  	it.B = h.B
   834  	it.buckets = h.buckets
   835  	if t.Bucket.PtrBytes == 0 {
   836  		// Allocate the current slice and remember pointers to both current and old.
   837  		// This preserves all relevant overflow buckets alive even if
   838  		// the table grows and/or overflow buckets are added to the table
   839  		// while we are iterating.
   840  		h.createOverflow()
   841  		it.overflow = h.extra.overflow
   842  		it.oldoverflow = h.extra.oldoverflow
   843  	}
   844  
   845  	// decide where to start
   846  	var r uintptr
   847  	if h.B > 31-bucketCntBits {
   848  		r = uintptr(fastrand64())
   849  	} else {
   850  		r = uintptr(fastrand())
   851  	}
   852  	it.startBucket = r & bucketMask(h.B)
   853  	it.offset = uint8(r >> h.B & (bucketCnt - 1))
   854  
   855  	// iterator state
   856  	it.bucket = it.startBucket
   857  
   858  	// Remember we have an iterator.
   859  	// Can run concurrently with another mapiterinit().
   860  	if old := h.flags; old&(iterator|oldIterator) != iterator|oldIterator {
   861  		atomic.Or8(&h.flags, iterator|oldIterator)
   862  	}
   863  
   864  	mapiternext(it)
   865  }
   866  
   867  func mapiternext(it *hiter) {
   868  	h := it.h
   869  	if raceenabled {
   870  		callerpc := getcallerpc()
   871  		racereadpc(unsafe.Pointer(h), callerpc, abi.FuncPCABIInternal(mapiternext))
   872  	}
   873  	if h.flags&hashWriting != 0 {
   874  		fatal("concurrent map iteration and map write")
   875  	}
   876  	t := it.t
   877  	bucket := it.bucket
   878  	b := it.bptr
   879  	i := it.i
   880  	checkBucket := it.checkBucket
   881  
   882  next:
   883  	if b == nil {
   884  		if bucket == it.startBucket && it.wrapped {
   885  			// end of iteration
   886  			it.key = nil
   887  			it.elem = nil
   888  			return
   889  		}
   890  		if h.growing() && it.B == h.B {
   891  			// Iterator was started in the middle of a grow, and the grow isn't done yet.
   892  			// If the bucket we're looking at hasn't been filled in yet (i.e. the old
   893  			// bucket hasn't been evacuated) then we need to iterate through the old
   894  			// bucket and only return the ones that will be migrated to this bucket.
   895  			oldbucket := bucket & it.h.oldbucketmask()
   896  			b = (*bmap)(add(h.oldbuckets, oldbucket*uintptr(t.BucketSize)))
   897  			if !evacuated(b) {
   898  				checkBucket = bucket
   899  			} else {
   900  				b = (*bmap)(add(it.buckets, bucket*uintptr(t.BucketSize)))
   901  				checkBucket = noCheck
   902  			}
   903  		} else {
   904  			b = (*bmap)(add(it.buckets, bucket*uintptr(t.BucketSize)))
   905  			checkBucket = noCheck
   906  		}
   907  		bucket++
   908  		if bucket == bucketShift(it.B) {
   909  			bucket = 0
   910  			it.wrapped = true
   911  		}
   912  		i = 0
   913  	}
   914  	for ; i < bucketCnt; i++ {
   915  		offi := (i + it.offset) & (bucketCnt - 1)
   916  		if isEmpty(b.tophash[offi]) || b.tophash[offi] == evacuatedEmpty {
   917  			// TODO: emptyRest is hard to use here, as we start iterating
   918  			// in the middle of a bucket. It's feasible, just tricky.
   919  			continue
   920  		}
   921  		k := add(unsafe.Pointer(b), dataOffset+uintptr(offi)*uintptr(t.KeySize))
   922  		if t.IndirectKey() {
   923  			k = *((*unsafe.Pointer)(k))
   924  		}
   925  		e := add(unsafe.Pointer(b), dataOffset+bucketCnt*uintptr(t.KeySize)+uintptr(offi)*uintptr(t.ValueSize))
   926  		if checkBucket != noCheck && !h.sameSizeGrow() {
   927  			// Special case: iterator was started during a grow to a larger size
   928  			// and the grow is not done yet. We're working on a bucket whose
   929  			// oldbucket has not been evacuated yet. Or at least, it wasn't
   930  			// evacuated when we started the bucket. So we're iterating
   931  			// through the oldbucket, skipping any keys that will go
   932  			// to the other new bucket (each oldbucket expands to two
   933  			// buckets during a grow).
   934  			if t.ReflexiveKey() || t.Key.Equal(k, k) {
   935  				// If the item in the oldbucket is not destined for
   936  				// the current new bucket in the iteration, skip it.
   937  				hash := t.Hasher(k, uintptr(h.hash0))
   938  				if hash&bucketMask(it.B) != checkBucket {
   939  					continue
   940  				}
   941  			} else {
   942  				// Hash isn't repeatable if k != k (NaNs).  We need a
   943  				// repeatable and randomish choice of which direction
   944  				// to send NaNs during evacuation. We'll use the low
   945  				// bit of tophash to decide which way NaNs go.
   946  				// NOTE: this case is why we need two evacuate tophash
   947  				// values, evacuatedX and evacuatedY, that differ in
   948  				// their low bit.
   949  				if checkBucket>>(it.B-1) != uintptr(b.tophash[offi]&1) {
   950  					continue
   951  				}
   952  			}
   953  		}
   954  		if (b.tophash[offi] != evacuatedX && b.tophash[offi] != evacuatedY) ||
   955  			!(t.ReflexiveKey() || t.Key.Equal(k, k)) {
   956  			// This is the golden data, we can return it.
   957  			// OR
   958  			// key!=key, so the entry can't be deleted or updated, so we can just return it.
   959  			// That's lucky for us because when key!=key we can't look it up successfully.
   960  			it.key = k
   961  			if t.IndirectElem() {
   962  				e = *((*unsafe.Pointer)(e))
   963  			}
   964  			it.elem = e
   965  		} else {
   966  			// The hash table has grown since the iterator was started.
   967  			// The golden data for this key is now somewhere else.
   968  			// Check the current hash table for the data.
   969  			// This code handles the case where the key
   970  			// has been deleted, updated, or deleted and reinserted.
   971  			// NOTE: we need to regrab the key as it has potentially been
   972  			// updated to an equal() but not identical key (e.g. +0.0 vs -0.0).
   973  			rk, re := mapaccessK(t, h, k)
   974  			if rk == nil {
   975  				continue // key has been deleted
   976  			}
   977  			it.key = rk
   978  			it.elem = re
   979  		}
   980  		it.bucket = bucket
   981  		if it.bptr != b { // avoid unnecessary write barrier; see issue 14921
   982  			it.bptr = b
   983  		}
   984  		it.i = i + 1
   985  		it.checkBucket = checkBucket
   986  		return
   987  	}
   988  	b = b.overflow(t)
   989  	i = 0
   990  	goto next
   991  }
   992  
   993  // mapclear deletes all keys from a map.
   994  func mapclear(t *maptype, h *hmap) {
   995  	if raceenabled && h != nil {
   996  		callerpc := getcallerpc()
   997  		pc := abi.FuncPCABIInternal(mapclear)
   998  		racewritepc(unsafe.Pointer(h), callerpc, pc)
   999  	}
  1000  
  1001  	if h == nil || h.count == 0 {
  1002  		return
  1003  	}
  1004  
  1005  	if h.flags&hashWriting != 0 {
  1006  		fatal("concurrent map writes")
  1007  	}
  1008  
  1009  	h.flags ^= hashWriting
  1010  
  1011  	// Mark buckets empty, so existing iterators can be terminated, see issue #59411.
  1012  	markBucketsEmpty := func(bucket unsafe.Pointer, mask uintptr) {
  1013  		for i := uintptr(0); i <= mask; i++ {
  1014  			b := (*bmap)(add(bucket, i*uintptr(t.BucketSize)))
  1015  			for ; b != nil; b = b.overflow(t) {
  1016  				for i := uintptr(0); i < bucketCnt; i++ {
  1017  					b.tophash[i] = emptyRest
  1018  				}
  1019  			}
  1020  		}
  1021  	}
  1022  	markBucketsEmpty(h.buckets, bucketMask(h.B))
  1023  	if oldBuckets := h.oldbuckets; oldBuckets != nil {
  1024  		markBucketsEmpty(oldBuckets, h.oldbucketmask())
  1025  	}
  1026  
  1027  	h.flags &^= sameSizeGrow
  1028  	h.oldbuckets = nil
  1029  	h.nevacuate = 0
  1030  	h.noverflow = 0
  1031  	h.count = 0
  1032  
  1033  	// Reset the hash seed to make it more difficult for attackers to
  1034  	// repeatedly trigger hash collisions. See issue 25237.
  1035  	h.hash0 = fastrand()
  1036  
  1037  	// Keep the mapextra allocation but clear any extra information.
  1038  	if h.extra != nil {
  1039  		*h.extra = mapextra{}
  1040  	}
  1041  
  1042  	// makeBucketArray clears the memory pointed to by h.buckets
  1043  	// and recovers any overflow buckets by generating them
  1044  	// as if h.buckets was newly alloced.
  1045  	_, nextOverflow := makeBucketArray(t, h.B, h.buckets)
  1046  	if nextOverflow != nil {
  1047  		// If overflow buckets are created then h.extra
  1048  		// will have been allocated during initial bucket creation.
  1049  		h.extra.nextOverflow = nextOverflow
  1050  	}
  1051  
  1052  	if h.flags&hashWriting == 0 {
  1053  		fatal("concurrent map writes")
  1054  	}
  1055  	h.flags &^= hashWriting
  1056  }
  1057  
  1058  func hashGrow(t *maptype, h *hmap) {
  1059  	// If we've hit the load factor, get bigger.
  1060  	// Otherwise, there are too many overflow buckets,
  1061  	// so keep the same number of buckets and "grow" laterally.
  1062  	bigger := uint8(1)
  1063  	if !overLoadFactor(h.count+1, h.B) {
  1064  		bigger = 0
  1065  		h.flags |= sameSizeGrow
  1066  	}
  1067  	oldbuckets := h.buckets
  1068  	newbuckets, nextOverflow := makeBucketArray(t, h.B+bigger, nil)
  1069  
  1070  	flags := h.flags &^ (iterator | oldIterator)
  1071  	if h.flags&iterator != 0 {
  1072  		flags |= oldIterator
  1073  	}
  1074  	// commit the grow (atomic wrt gc)
  1075  	h.B += bigger
  1076  	h.flags = flags
  1077  	h.oldbuckets = oldbuckets
  1078  	h.buckets = newbuckets
  1079  	h.nevacuate = 0
  1080  	h.noverflow = 0
  1081  
  1082  	if h.extra != nil && h.extra.overflow != nil {
  1083  		// Promote current overflow buckets to the old generation.
  1084  		if h.extra.oldoverflow != nil {
  1085  			throw("oldoverflow is not nil")
  1086  		}
  1087  		h.extra.oldoverflow = h.extra.overflow
  1088  		h.extra.overflow = nil
  1089  	}
  1090  	if nextOverflow != nil {
  1091  		if h.extra == nil {
  1092  			h.extra = new(mapextra)
  1093  		}
  1094  		h.extra.nextOverflow = nextOverflow
  1095  	}
  1096  
  1097  	// the actual copying of the hash table data is done incrementally
  1098  	// by growWork() and evacuate().
  1099  }
  1100  
  1101  // overLoadFactor reports whether count items placed in 1<<B buckets is over loadFactor.
  1102  func overLoadFactor(count int, B uint8) bool {
  1103  	return count > bucketCnt && uintptr(count) > loadFactorNum*(bucketShift(B)/loadFactorDen)
  1104  }
  1105  
  1106  // tooManyOverflowBuckets reports whether noverflow buckets is too many for a map with 1<<B buckets.
  1107  // Note that most of these overflow buckets must be in sparse use;
  1108  // if use was dense, then we'd have already triggered regular map growth.
  1109  func tooManyOverflowBuckets(noverflow uint16, B uint8) bool {
  1110  	// If the threshold is too low, we do extraneous work.
  1111  	// If the threshold is too high, maps that grow and shrink can hold on to lots of unused memory.
  1112  	// "too many" means (approximately) as many overflow buckets as regular buckets.
  1113  	// See incrnoverflow for more details.
  1114  	if B > 15 {
  1115  		B = 15
  1116  	}
  1117  	// The compiler doesn't see here that B < 16; mask B to generate shorter shift code.
  1118  	return noverflow >= uint16(1)<<(B&15)
  1119  }
  1120  
  1121  // growing reports whether h is growing. The growth may be to the same size or bigger.
  1122  func (h *hmap) growing() bool {
  1123  	return h.oldbuckets != nil
  1124  }
  1125  
  1126  // sameSizeGrow reports whether the current growth is to a map of the same size.
  1127  func (h *hmap) sameSizeGrow() bool {
  1128  	return h.flags&sameSizeGrow != 0
  1129  }
  1130  
  1131  // noldbuckets calculates the number of buckets prior to the current map growth.
  1132  func (h *hmap) noldbuckets() uintptr {
  1133  	oldB := h.B
  1134  	if !h.sameSizeGrow() {
  1135  		oldB--
  1136  	}
  1137  	return bucketShift(oldB)
  1138  }
  1139  
  1140  // oldbucketmask provides a mask that can be applied to calculate n % noldbuckets().
  1141  func (h *hmap) oldbucketmask() uintptr {
  1142  	return h.noldbuckets() - 1
  1143  }
  1144  
  1145  func growWork(t *maptype, h *hmap, bucket uintptr) {
  1146  	// make sure we evacuate the oldbucket corresponding
  1147  	// to the bucket we're about to use
  1148  	evacuate(t, h, bucket&h.oldbucketmask())
  1149  
  1150  	// evacuate one more oldbucket to make progress on growing
  1151  	if h.growing() {
  1152  		evacuate(t, h, h.nevacuate)
  1153  	}
  1154  }
  1155  
  1156  func bucketEvacuated(t *maptype, h *hmap, bucket uintptr) bool {
  1157  	b := (*bmap)(add(h.oldbuckets, bucket*uintptr(t.BucketSize)))
  1158  	return evacuated(b)
  1159  }
  1160  
  1161  // evacDst is an evacuation destination.
  1162  type evacDst struct {
  1163  	b *bmap          // current destination bucket
  1164  	i int            // key/elem index into b
  1165  	k unsafe.Pointer // pointer to current key storage
  1166  	e unsafe.Pointer // pointer to current elem storage
  1167  }
  1168  
  1169  func evacuate(t *maptype, h *hmap, oldbucket uintptr) {
  1170  	b := (*bmap)(add(h.oldbuckets, oldbucket*uintptr(t.BucketSize)))
  1171  	newbit := h.noldbuckets()
  1172  	if !evacuated(b) {
  1173  		// TODO: reuse overflow buckets instead of using new ones, if there
  1174  		// is no iterator using the old buckets.  (If !oldIterator.)
  1175  
  1176  		// xy contains the x and y (low and high) evacuation destinations.
  1177  		var xy [2]evacDst
  1178  		x := &xy[0]
  1179  		x.b = (*bmap)(add(h.buckets, oldbucket*uintptr(t.BucketSize)))
  1180  		x.k = add(unsafe.Pointer(x.b), dataOffset)
  1181  		x.e = add(x.k, bucketCnt*uintptr(t.KeySize))
  1182  
  1183  		if !h.sameSizeGrow() {
  1184  			// Only calculate y pointers if we're growing bigger.
  1185  			// Otherwise GC can see bad pointers.
  1186  			y := &xy[1]
  1187  			y.b = (*bmap)(add(h.buckets, (oldbucket+newbit)*uintptr(t.BucketSize)))
  1188  			y.k = add(unsafe.Pointer(y.b), dataOffset)
  1189  			y.e = add(y.k, bucketCnt*uintptr(t.KeySize))
  1190  		}
  1191  
  1192  		for ; b != nil; b = b.overflow(t) {
  1193  			k := add(unsafe.Pointer(b), dataOffset)
  1194  			e := add(k, bucketCnt*uintptr(t.KeySize))
  1195  			for i := 0; i < bucketCnt; i, k, e = i+1, add(k, uintptr(t.KeySize)), add(e, uintptr(t.ValueSize)) {
  1196  				top := b.tophash[i]
  1197  				if isEmpty(top) {
  1198  					b.tophash[i] = evacuatedEmpty
  1199  					continue
  1200  				}
  1201  				if top < minTopHash {
  1202  					throw("bad map state")
  1203  				}
  1204  				k2 := k
  1205  				if t.IndirectKey() {
  1206  					k2 = *((*unsafe.Pointer)(k2))
  1207  				}
  1208  				var useY uint8
  1209  				if !h.sameSizeGrow() {
  1210  					// Compute hash to make our evacuation decision (whether we need
  1211  					// to send this key/elem to bucket x or bucket y).
  1212  					hash := t.Hasher(k2, uintptr(h.hash0))
  1213  					if h.flags&iterator != 0 && !t.ReflexiveKey() && !t.Key.Equal(k2, k2) {
  1214  						// If key != key (NaNs), then the hash could be (and probably
  1215  						// will be) entirely different from the old hash. Moreover,
  1216  						// it isn't reproducible. Reproducibility is required in the
  1217  						// presence of iterators, as our evacuation decision must
  1218  						// match whatever decision the iterator made.
  1219  						// Fortunately, we have the freedom to send these keys either
  1220  						// way. Also, tophash is meaningless for these kinds of keys.
  1221  						// We let the low bit of tophash drive the evacuation decision.
  1222  						// We recompute a new random tophash for the next level so
  1223  						// these keys will get evenly distributed across all buckets
  1224  						// after multiple grows.
  1225  						useY = top & 1
  1226  						top = tophash(hash)
  1227  					} else {
  1228  						if hash&newbit != 0 {
  1229  							useY = 1
  1230  						}
  1231  					}
  1232  				}
  1233  
  1234  				if evacuatedX+1 != evacuatedY || evacuatedX^1 != evacuatedY {
  1235  					throw("bad evacuatedN")
  1236  				}
  1237  
  1238  				b.tophash[i] = evacuatedX + useY // evacuatedX + 1 == evacuatedY
  1239  				dst := &xy[useY]                 // evacuation destination
  1240  
  1241  				if dst.i == bucketCnt {
  1242  					dst.b = h.newoverflow(t, dst.b)
  1243  					dst.i = 0
  1244  					dst.k = add(unsafe.Pointer(dst.b), dataOffset)
  1245  					dst.e = add(dst.k, bucketCnt*uintptr(t.KeySize))
  1246  				}
  1247  				dst.b.tophash[dst.i&(bucketCnt-1)] = top // mask dst.i as an optimization, to avoid a bounds check
  1248  				if t.IndirectKey() {
  1249  					*(*unsafe.Pointer)(dst.k) = k2 // copy pointer
  1250  				} else {
  1251  					typedmemmove(t.Key, dst.k, k) // copy elem
  1252  				}
  1253  				if t.IndirectElem() {
  1254  					*(*unsafe.Pointer)(dst.e) = *(*unsafe.Pointer)(e)
  1255  				} else {
  1256  					typedmemmove(t.Elem, dst.e, e)
  1257  				}
  1258  				dst.i++
  1259  				// These updates might push these pointers past the end of the
  1260  				// key or elem arrays.  That's ok, as we have the overflow pointer
  1261  				// at the end of the bucket to protect against pointing past the
  1262  				// end of the bucket.
  1263  				dst.k = add(dst.k, uintptr(t.KeySize))
  1264  				dst.e = add(dst.e, uintptr(t.ValueSize))
  1265  			}
  1266  		}
  1267  		// Unlink the overflow buckets & clear key/elem to help GC.
  1268  		if h.flags&oldIterator == 0 && t.Bucket.PtrBytes != 0 {
  1269  			b := add(h.oldbuckets, oldbucket*uintptr(t.BucketSize))
  1270  			// Preserve b.tophash because the evacuation
  1271  			// state is maintained there.
  1272  			ptr := add(b, dataOffset)
  1273  			n := uintptr(t.BucketSize) - dataOffset
  1274  			memclrHasPointers(ptr, n)
  1275  		}
  1276  	}
  1277  
  1278  	if oldbucket == h.nevacuate {
  1279  		advanceEvacuationMark(h, t, newbit)
  1280  	}
  1281  }
  1282  
  1283  func advanceEvacuationMark(h *hmap, t *maptype, newbit uintptr) {
  1284  	h.nevacuate++
  1285  	// Experiments suggest that 1024 is overkill by at least an order of magnitude.
  1286  	// Put it in there as a safeguard anyway, to ensure O(1) behavior.
  1287  	stop := h.nevacuate + 1024
  1288  	if stop > newbit {
  1289  		stop = newbit
  1290  	}
  1291  	for h.nevacuate != stop && bucketEvacuated(t, h, h.nevacuate) {
  1292  		h.nevacuate++
  1293  	}
  1294  	if h.nevacuate == newbit { // newbit == # of oldbuckets
  1295  		// Growing is all done. Free old main bucket array.
  1296  		h.oldbuckets = nil
  1297  		// Can discard old overflow buckets as well.
  1298  		// If they are still referenced by an iterator,
  1299  		// then the iterator holds a pointers to the slice.
  1300  		if h.extra != nil {
  1301  			h.extra.oldoverflow = nil
  1302  		}
  1303  		h.flags &^= sameSizeGrow
  1304  	}
  1305  }
  1306  
  1307  // Reflect stubs. Called from ../reflect/asm_*.s
  1308  
  1309  //go:linkname reflect_makemap reflect.makemap
  1310  func reflect_makemap(t *maptype, cap int) *hmap {
  1311  	// Check invariants and reflects math.
  1312  	if t.Key.Equal == nil {
  1313  		throw("runtime.reflect_makemap: unsupported map key type")
  1314  	}
  1315  	if t.Key.Size_ > maxKeySize && (!t.IndirectKey() || t.KeySize != uint8(goarch.PtrSize)) ||
  1316  		t.Key.Size_ <= maxKeySize && (t.IndirectKey() || t.KeySize != uint8(t.Key.Size_)) {
  1317  		throw("key size wrong")
  1318  	}
  1319  	if t.Elem.Size_ > maxElemSize && (!t.IndirectElem() || t.ValueSize != uint8(goarch.PtrSize)) ||
  1320  		t.Elem.Size_ <= maxElemSize && (t.IndirectElem() || t.ValueSize != uint8(t.Elem.Size_)) {
  1321  		throw("elem size wrong")
  1322  	}
  1323  	if t.Key.Align_ > bucketCnt {
  1324  		throw("key align too big")
  1325  	}
  1326  	if t.Elem.Align_ > bucketCnt {
  1327  		throw("elem align too big")
  1328  	}
  1329  	if t.Key.Size_%uintptr(t.Key.Align_) != 0 {
  1330  		throw("key size not a multiple of key align")
  1331  	}
  1332  	if t.Elem.Size_%uintptr(t.Elem.Align_) != 0 {
  1333  		throw("elem size not a multiple of elem align")
  1334  	}
  1335  	if bucketCnt < 8 {
  1336  		throw("bucketsize too small for proper alignment")
  1337  	}
  1338  	if dataOffset%uintptr(t.Key.Align_) != 0 {
  1339  		throw("need padding in bucket (key)")
  1340  	}
  1341  	if dataOffset%uintptr(t.Elem.Align_) != 0 {
  1342  		throw("need padding in bucket (elem)")
  1343  	}
  1344  
  1345  	return makemap(t, cap, nil)
  1346  }
  1347  
  1348  //go:linkname reflect_mapaccess reflect.mapaccess
  1349  func reflect_mapaccess(t *maptype, h *hmap, key unsafe.Pointer) unsafe.Pointer {
  1350  	elem, ok := mapaccess2(t, h, key)
  1351  	if !ok {
  1352  		// reflect wants nil for a missing element
  1353  		elem = nil
  1354  	}
  1355  	return elem
  1356  }
  1357  
  1358  //go:linkname reflect_mapaccess_faststr reflect.mapaccess_faststr
  1359  func reflect_mapaccess_faststr(t *maptype, h *hmap, key string) unsafe.Pointer {
  1360  	elem, ok := mapaccess2_faststr(t, h, key)
  1361  	if !ok {
  1362  		// reflect wants nil for a missing element
  1363  		elem = nil
  1364  	}
  1365  	return elem
  1366  }
  1367  
  1368  //go:linkname reflect_mapassign reflect.mapassign0
  1369  func reflect_mapassign(t *maptype, h *hmap, key unsafe.Pointer, elem unsafe.Pointer) {
  1370  	p := mapassign(t, h, key)
  1371  	typedmemmove(t.Elem, p, elem)
  1372  }
  1373  
  1374  //go:linkname reflect_mapassign_faststr reflect.mapassign_faststr0
  1375  func reflect_mapassign_faststr(t *maptype, h *hmap, key string, elem unsafe.Pointer) {
  1376  	p := mapassign_faststr(t, h, key)
  1377  	typedmemmove(t.Elem, p, elem)
  1378  }
  1379  
  1380  //go:linkname reflect_mapdelete reflect.mapdelete
  1381  func reflect_mapdelete(t *maptype, h *hmap, key unsafe.Pointer) {
  1382  	mapdelete(t, h, key)
  1383  }
  1384  
  1385  //go:linkname reflect_mapdelete_faststr reflect.mapdelete_faststr
  1386  func reflect_mapdelete_faststr(t *maptype, h *hmap, key string) {
  1387  	mapdelete_faststr(t, h, key)
  1388  }
  1389  
  1390  //go:linkname reflect_mapiterinit reflect.mapiterinit
  1391  func reflect_mapiterinit(t *maptype, h *hmap, it *hiter) {
  1392  	mapiterinit(t, h, it)
  1393  }
  1394  
  1395  //go:linkname reflect_mapiternext reflect.mapiternext
  1396  func reflect_mapiternext(it *hiter) {
  1397  	mapiternext(it)
  1398  }
  1399  
  1400  //go:linkname reflect_mapiterkey reflect.mapiterkey
  1401  func reflect_mapiterkey(it *hiter) unsafe.Pointer {
  1402  	return it.key
  1403  }
  1404  
  1405  //go:linkname reflect_mapiterelem reflect.mapiterelem
  1406  func reflect_mapiterelem(it *hiter) unsafe.Pointer {
  1407  	return it.elem
  1408  }
  1409  
  1410  //go:linkname reflect_maplen reflect.maplen
  1411  func reflect_maplen(h *hmap) int {
  1412  	if h == nil {
  1413  		return 0
  1414  	}
  1415  	if raceenabled {
  1416  		callerpc := getcallerpc()
  1417  		racereadpc(unsafe.Pointer(h), callerpc, abi.FuncPCABIInternal(reflect_maplen))
  1418  	}
  1419  	return h.count
  1420  }
  1421  
  1422  //go:linkname reflect_mapclear reflect.mapclear
  1423  func reflect_mapclear(t *maptype, h *hmap) {
  1424  	mapclear(t, h)
  1425  }
  1426  
  1427  //go:linkname reflectlite_maplen internal/reflectlite.maplen
  1428  func reflectlite_maplen(h *hmap) int {
  1429  	if h == nil {
  1430  		return 0
  1431  	}
  1432  	if raceenabled {
  1433  		callerpc := getcallerpc()
  1434  		racereadpc(unsafe.Pointer(h), callerpc, abi.FuncPCABIInternal(reflect_maplen))
  1435  	}
  1436  	return h.count
  1437  }
  1438  
  1439  const maxZero = 1024 // must match value in reflect/value.go:maxZero cmd/compile/internal/gc/walk.go:zeroValSize
  1440  var zeroVal [maxZero]byte
  1441  
  1442  // mapinitnoop is a no-op function known the Go linker; if a given global
  1443  // map (of the right size) is determined to be dead, the linker will
  1444  // rewrite the relocation (from the package init func) from the outlined
  1445  // map init function to this symbol. Defined in assembly so as to avoid
  1446  // complications with instrumentation (coverage, etc).
  1447  func mapinitnoop()
  1448  
  1449  // mapclone for implementing maps.Clone
  1450  //
  1451  //go:linkname mapclone maps.clone
  1452  func mapclone(m any) any {
  1453  	e := efaceOf(&m)
  1454  	e.data = unsafe.Pointer(mapclone2((*maptype)(unsafe.Pointer(e._type)), (*hmap)(e.data)))
  1455  	return m
  1456  }
  1457  
  1458  // moveToBmap moves a bucket from src to dst. It returns the destination bucket or new destination bucket if it overflows
  1459  // and the pos that the next key/value will be written, if pos == bucketCnt means needs to written in overflow bucket.
  1460  func moveToBmap(t *maptype, h *hmap, dst *bmap, pos int, src *bmap) (*bmap, int) {
  1461  	for i := 0; i < bucketCnt; i++ {
  1462  		if isEmpty(src.tophash[i]) {
  1463  			continue
  1464  		}
  1465  
  1466  		for ; pos < bucketCnt; pos++ {
  1467  			if isEmpty(dst.tophash[pos]) {
  1468  				break
  1469  			}
  1470  		}
  1471  
  1472  		if pos == bucketCnt {
  1473  			dst = h.newoverflow(t, dst)
  1474  			pos = 0
  1475  		}
  1476  
  1477  		srcK := add(unsafe.Pointer(src), dataOffset+uintptr(i)*uintptr(t.KeySize))
  1478  		srcEle := add(unsafe.Pointer(src), dataOffset+bucketCnt*uintptr(t.KeySize)+uintptr(i)*uintptr(t.ValueSize))
  1479  		dstK := add(unsafe.Pointer(dst), dataOffset+uintptr(pos)*uintptr(t.KeySize))
  1480  		dstEle := add(unsafe.Pointer(dst), dataOffset+bucketCnt*uintptr(t.KeySize)+uintptr(pos)*uintptr(t.ValueSize))
  1481  
  1482  		dst.tophash[pos] = src.tophash[i]
  1483  		if t.IndirectKey() {
  1484  			*(*unsafe.Pointer)(dstK) = *(*unsafe.Pointer)(srcK)
  1485  		} else {
  1486  			typedmemmove(t.Key, dstK, srcK)
  1487  		}
  1488  		if t.IndirectElem() {
  1489  			*(*unsafe.Pointer)(dstEle) = *(*unsafe.Pointer)(srcEle)
  1490  		} else {
  1491  			typedmemmove(t.Elem, dstEle, srcEle)
  1492  		}
  1493  		pos++
  1494  		h.count++
  1495  	}
  1496  	return dst, pos
  1497  }
  1498  
  1499  func mapclone2(t *maptype, src *hmap) *hmap {
  1500  	dst := makemap(t, src.count, nil)
  1501  	dst.hash0 = src.hash0
  1502  	dst.nevacuate = 0
  1503  	//flags do not need to be copied here, just like a new map has no flags.
  1504  
  1505  	if src.count == 0 {
  1506  		return dst
  1507  	}
  1508  
  1509  	if src.flags&hashWriting != 0 {
  1510  		fatal("concurrent map clone and map write")
  1511  	}
  1512  
  1513  	if src.B == 0 {
  1514  		dst.buckets = newobject(t.Bucket)
  1515  		dst.count = src.count
  1516  		typedmemmove(t.Bucket, dst.buckets, src.buckets)
  1517  		return dst
  1518  	}
  1519  
  1520  	//src.B != 0
  1521  	if dst.B == 0 {
  1522  		dst.buckets = newobject(t.Bucket)
  1523  	}
  1524  	dstArraySize := int(bucketShift(dst.B))
  1525  	srcArraySize := int(bucketShift(src.B))
  1526  	for i := 0; i < dstArraySize; i++ {
  1527  		dstBmap := (*bmap)(add(dst.buckets, uintptr(i*int(t.BucketSize))))
  1528  		pos := 0
  1529  		for j := 0; j < srcArraySize; j += dstArraySize {
  1530  			srcBmap := (*bmap)(add(src.buckets, uintptr((i+j)*int(t.BucketSize))))
  1531  			for srcBmap != nil {
  1532  				dstBmap, pos = moveToBmap(t, dst, dstBmap, pos, srcBmap)
  1533  				srcBmap = srcBmap.overflow(t)
  1534  			}
  1535  		}
  1536  	}
  1537  
  1538  	if src.oldbuckets == nil {
  1539  		return dst
  1540  	}
  1541  
  1542  	oldB := src.B
  1543  	srcOldbuckets := src.oldbuckets
  1544  	if !src.sameSizeGrow() {
  1545  		oldB--
  1546  	}
  1547  	oldSrcArraySize := int(bucketShift(oldB))
  1548  
  1549  	for i := 0; i < oldSrcArraySize; i++ {
  1550  		srcBmap := (*bmap)(add(srcOldbuckets, uintptr(i*int(t.BucketSize))))
  1551  		if evacuated(srcBmap) {
  1552  			continue
  1553  		}
  1554  
  1555  		if oldB >= dst.B { // main bucket bits in dst is less than oldB bits in src
  1556  			dstBmap := (*bmap)(add(dst.buckets, uintptr(i)&bucketMask(dst.B)))
  1557  			for dstBmap.overflow(t) != nil {
  1558  				dstBmap = dstBmap.overflow(t)
  1559  			}
  1560  			pos := 0
  1561  			for srcBmap != nil {
  1562  				dstBmap, pos = moveToBmap(t, dst, dstBmap, pos, srcBmap)
  1563  				srcBmap = srcBmap.overflow(t)
  1564  			}
  1565  			continue
  1566  		}
  1567  
  1568  		for srcBmap != nil {
  1569  			// move from oldBlucket to new bucket
  1570  			for i := uintptr(0); i < bucketCnt; i++ {
  1571  				if isEmpty(srcBmap.tophash[i]) {
  1572  					continue
  1573  				}
  1574  
  1575  				if src.flags&hashWriting != 0 {
  1576  					fatal("concurrent map clone and map write")
  1577  				}
  1578  
  1579  				srcK := add(unsafe.Pointer(srcBmap), dataOffset+i*uintptr(t.KeySize))
  1580  				if t.IndirectKey() {
  1581  					srcK = *((*unsafe.Pointer)(srcK))
  1582  				}
  1583  
  1584  				srcEle := add(unsafe.Pointer(srcBmap), dataOffset+bucketCnt*uintptr(t.KeySize)+i*uintptr(t.ValueSize))
  1585  				if t.IndirectElem() {
  1586  					srcEle = *((*unsafe.Pointer)(srcEle))
  1587  				}
  1588  				dstEle := mapassign(t, dst, srcK)
  1589  				typedmemmove(t.Elem, dstEle, srcEle)
  1590  			}
  1591  			srcBmap = srcBmap.overflow(t)
  1592  		}
  1593  	}
  1594  	return dst
  1595  }
  1596  
  1597  // keys for implementing maps.keys
  1598  //
  1599  //go:linkname keys maps.keys
  1600  func keys(m any, p unsafe.Pointer) {
  1601  	e := efaceOf(&m)
  1602  	t := (*maptype)(unsafe.Pointer(e._type))
  1603  	h := (*hmap)(e.data)
  1604  
  1605  	if h == nil || h.count == 0 {
  1606  		return
  1607  	}
  1608  	s := (*slice)(p)
  1609  	r := int(fastrand())
  1610  	offset := uint8(r >> h.B & (bucketCnt - 1))
  1611  	if h.B == 0 {
  1612  		copyKeys(t, h, (*bmap)(h.buckets), s, offset)
  1613  		return
  1614  	}
  1615  	arraySize := int(bucketShift(h.B))
  1616  	buckets := h.buckets
  1617  	for i := 0; i < arraySize; i++ {
  1618  		bucket := (i + r) & (arraySize - 1)
  1619  		b := (*bmap)(add(buckets, uintptr(bucket)*uintptr(t.BucketSize)))
  1620  		copyKeys(t, h, b, s, offset)
  1621  	}
  1622  
  1623  	if h.growing() {
  1624  		oldArraySize := int(h.noldbuckets())
  1625  		for i := 0; i < oldArraySize; i++ {
  1626  			bucket := (i + r) & (oldArraySize - 1)
  1627  			b := (*bmap)(add(h.oldbuckets, uintptr(bucket)*uintptr(t.BucketSize)))
  1628  			if evacuated(b) {
  1629  				continue
  1630  			}
  1631  			copyKeys(t, h, b, s, offset)
  1632  		}
  1633  	}
  1634  	return
  1635  }
  1636  
  1637  func copyKeys(t *maptype, h *hmap, b *bmap, s *slice, offset uint8) {
  1638  	for b != nil {
  1639  		for i := uintptr(0); i < bucketCnt; i++ {
  1640  			offi := (i + uintptr(offset)) & (bucketCnt - 1)
  1641  			if isEmpty(b.tophash[offi]) {
  1642  				continue
  1643  			}
  1644  			if h.flags&hashWriting != 0 {
  1645  				fatal("concurrent map read and map write")
  1646  			}
  1647  			k := add(unsafe.Pointer(b), dataOffset+offi*uintptr(t.KeySize))
  1648  			if t.IndirectKey() {
  1649  				k = *((*unsafe.Pointer)(k))
  1650  			}
  1651  			if s.len >= s.cap {
  1652  				fatal("concurrent map read and map write")
  1653  			}
  1654  			typedmemmove(t.Key, add(s.array, uintptr(s.len)*uintptr(t.KeySize)), k)
  1655  			s.len++
  1656  		}
  1657  		b = b.overflow(t)
  1658  	}
  1659  }