Black Lives Matter. Support the Equal Justice Initiative.

Source file src/runtime/mbitmap.go

Documentation: runtime

     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  // Garbage collector: type and heap bitmaps.
     6  //
     7  // Stack, data, and bss bitmaps
     8  //
     9  // Stack frames and global variables in the data and bss sections are
    10  // described by bitmaps with 1 bit per pointer-sized word. A "1" bit
    11  // means the word is a live pointer to be visited by the GC (referred to
    12  // as "pointer"). A "0" bit means the word should be ignored by GC
    13  // (referred to as "scalar", though it could be a dead pointer value).
    14  //
    15  // Heap bitmap
    16  //
    17  // The heap bitmap comprises 2 bits for each pointer-sized word in the heap,
    18  // stored in the heapArena metadata backing each heap arena.
    19  // That is, if ha is the heapArena for the arena starting a start,
    20  // then ha.bitmap[0] holds the 2-bit entries for the four words start
    21  // through start+3*ptrSize, ha.bitmap[1] holds the entries for
    22  // start+4*ptrSize through start+7*ptrSize, and so on.
    23  //
    24  // In each 2-bit entry, the lower bit is a pointer/scalar bit, just
    25  // like in the stack/data bitmaps described above. The upper bit
    26  // indicates scan/dead: a "1" value ("scan") indicates that there may
    27  // be pointers in later words of the allocation, and a "0" value
    28  // ("dead") indicates there are no more pointers in the allocation. If
    29  // the upper bit is 0, the lower bit must also be 0, and this
    30  // indicates scanning can ignore the rest of the allocation.
    31  //
    32  // The 2-bit entries are split when written into the byte, so that the top half
    33  // of the byte contains 4 high (scan) bits and the bottom half contains 4 low
    34  // (pointer) bits. This form allows a copy from the 1-bit to the 4-bit form to
    35  // keep the pointer bits contiguous, instead of having to space them out.
    36  //
    37  // The code makes use of the fact that the zero value for a heap
    38  // bitmap means scalar/dead. This property must be preserved when
    39  // modifying the encoding.
    40  //
    41  // The bitmap for noscan spans is not maintained. Code must ensure
    42  // that an object is scannable before consulting its bitmap by
    43  // checking either the noscan bit in the span or by consulting its
    44  // type's information.
    45  
    46  package runtime
    47  
    48  import (
    49  	"runtime/internal/atomic"
    50  	"runtime/internal/sys"
    51  	"unsafe"
    52  )
    53  
    54  const (
    55  	bitPointer = 1 << 0
    56  	bitScan    = 1 << 4
    57  
    58  	heapBitsShift      = 1     // shift offset between successive bitPointer or bitScan entries
    59  	wordsPerBitmapByte = 8 / 2 // heap words described by one bitmap byte
    60  
    61  	// all scan/pointer bits in a byte
    62  	bitScanAll    = bitScan | bitScan<<heapBitsShift | bitScan<<(2*heapBitsShift) | bitScan<<(3*heapBitsShift)
    63  	bitPointerAll = bitPointer | bitPointer<<heapBitsShift | bitPointer<<(2*heapBitsShift) | bitPointer<<(3*heapBitsShift)
    64  )
    65  
    66  // addb returns the byte pointer p+n.
    67  //go:nowritebarrier
    68  //go:nosplit
    69  func addb(p *byte, n uintptr) *byte {
    70  	// Note: wrote out full expression instead of calling add(p, n)
    71  	// to reduce the number of temporaries generated by the
    72  	// compiler for this trivial expression during inlining.
    73  	return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) + n))
    74  }
    75  
    76  // subtractb returns the byte pointer p-n.
    77  //go:nowritebarrier
    78  //go:nosplit
    79  func subtractb(p *byte, n uintptr) *byte {
    80  	// Note: wrote out full expression instead of calling add(p, -n)
    81  	// to reduce the number of temporaries generated by the
    82  	// compiler for this trivial expression during inlining.
    83  	return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) - n))
    84  }
    85  
    86  // add1 returns the byte pointer p+1.
    87  //go:nowritebarrier
    88  //go:nosplit
    89  func add1(p *byte) *byte {
    90  	// Note: wrote out full expression instead of calling addb(p, 1)
    91  	// to reduce the number of temporaries generated by the
    92  	// compiler for this trivial expression during inlining.
    93  	return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) + 1))
    94  }
    95  
    96  // subtract1 returns the byte pointer p-1.
    97  //go:nowritebarrier
    98  //
    99  // nosplit because it is used during write barriers and must not be preempted.
   100  //go:nosplit
   101  func subtract1(p *byte) *byte {
   102  	// Note: wrote out full expression instead of calling subtractb(p, 1)
   103  	// to reduce the number of temporaries generated by the
   104  	// compiler for this trivial expression during inlining.
   105  	return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) - 1))
   106  }
   107  
   108  // heapBits provides access to the bitmap bits for a single heap word.
   109  // The methods on heapBits take value receivers so that the compiler
   110  // can more easily inline calls to those methods and registerize the
   111  // struct fields independently.
   112  type heapBits struct {
   113  	bitp  *uint8
   114  	shift uint32
   115  	arena uint32 // Index of heap arena containing bitp
   116  	last  *uint8 // Last byte arena's bitmap
   117  }
   118  
   119  // Make the compiler check that heapBits.arena is large enough to hold
   120  // the maximum arena frame number.
   121  var _ = heapBits{arena: (1<<heapAddrBits)/heapArenaBytes - 1}
   122  
   123  // markBits provides access to the mark bit for an object in the heap.
   124  // bytep points to the byte holding the mark bit.
   125  // mask is a byte with a single bit set that can be &ed with *bytep
   126  // to see if the bit has been set.
   127  // *m.byte&m.mask != 0 indicates the mark bit is set.
   128  // index can be used along with span information to generate
   129  // the address of the object in the heap.
   130  // We maintain one set of mark bits for allocation and one for
   131  // marking purposes.
   132  type markBits struct {
   133  	bytep *uint8
   134  	mask  uint8
   135  	index uintptr
   136  }
   137  
   138  //go:nosplit
   139  func (s *mspan) allocBitsForIndex(allocBitIndex uintptr) markBits {
   140  	bytep, mask := s.allocBits.bitp(allocBitIndex)
   141  	return markBits{bytep, mask, allocBitIndex}
   142  }
   143  
   144  // refillAllocCache takes 8 bytes s.allocBits starting at whichByte
   145  // and negates them so that ctz (count trailing zeros) instructions
   146  // can be used. It then places these 8 bytes into the cached 64 bit
   147  // s.allocCache.
   148  func (s *mspan) refillAllocCache(whichByte uintptr) {
   149  	bytes := (*[8]uint8)(unsafe.Pointer(s.allocBits.bytep(whichByte)))
   150  	aCache := uint64(0)
   151  	aCache |= uint64(bytes[0])
   152  	aCache |= uint64(bytes[1]) << (1 * 8)
   153  	aCache |= uint64(bytes[2]) << (2 * 8)
   154  	aCache |= uint64(bytes[3]) << (3 * 8)
   155  	aCache |= uint64(bytes[4]) << (4 * 8)
   156  	aCache |= uint64(bytes[5]) << (5 * 8)
   157  	aCache |= uint64(bytes[6]) << (6 * 8)
   158  	aCache |= uint64(bytes[7]) << (7 * 8)
   159  	s.allocCache = ^aCache
   160  }
   161  
   162  // nextFreeIndex returns the index of the next free object in s at
   163  // or after s.freeindex.
   164  // There are hardware instructions that can be used to make this
   165  // faster if profiling warrants it.
   166  func (s *mspan) nextFreeIndex() uintptr {
   167  	sfreeindex := s.freeindex
   168  	snelems := s.nelems
   169  	if sfreeindex == snelems {
   170  		return sfreeindex
   171  	}
   172  	if sfreeindex > snelems {
   173  		throw("s.freeindex > s.nelems")
   174  	}
   175  
   176  	aCache := s.allocCache
   177  
   178  	bitIndex := sys.Ctz64(aCache)
   179  	for bitIndex == 64 {
   180  		// Move index to start of next cached bits.
   181  		sfreeindex = (sfreeindex + 64) &^ (64 - 1)
   182  		if sfreeindex >= snelems {
   183  			s.freeindex = snelems
   184  			return snelems
   185  		}
   186  		whichByte := sfreeindex / 8
   187  		// Refill s.allocCache with the next 64 alloc bits.
   188  		s.refillAllocCache(whichByte)
   189  		aCache = s.allocCache
   190  		bitIndex = sys.Ctz64(aCache)
   191  		// nothing available in cached bits
   192  		// grab the next 8 bytes and try again.
   193  	}
   194  	result := sfreeindex + uintptr(bitIndex)
   195  	if result >= snelems {
   196  		s.freeindex = snelems
   197  		return snelems
   198  	}
   199  
   200  	s.allocCache >>= uint(bitIndex + 1)
   201  	sfreeindex = result + 1
   202  
   203  	if sfreeindex%64 == 0 && sfreeindex != snelems {
   204  		// We just incremented s.freeindex so it isn't 0.
   205  		// As each 1 in s.allocCache was encountered and used for allocation
   206  		// it was shifted away. At this point s.allocCache contains all 0s.
   207  		// Refill s.allocCache so that it corresponds
   208  		// to the bits at s.allocBits starting at s.freeindex.
   209  		whichByte := sfreeindex / 8
   210  		s.refillAllocCache(whichByte)
   211  	}
   212  	s.freeindex = sfreeindex
   213  	return result
   214  }
   215  
   216  // isFree reports whether the index'th object in s is unallocated.
   217  //
   218  // The caller must ensure s.state is mSpanInUse, and there must have
   219  // been no preemption points since ensuring this (which could allow a
   220  // GC transition, which would allow the state to change).
   221  func (s *mspan) isFree(index uintptr) bool {
   222  	if index < s.freeindex {
   223  		return false
   224  	}
   225  	bytep, mask := s.allocBits.bitp(index)
   226  	return *bytep&mask == 0
   227  }
   228  
   229  func (s *mspan) objIndex(p uintptr) uintptr {
   230  	byteOffset := p - s.base()
   231  	if byteOffset == 0 {
   232  		return 0
   233  	}
   234  	if s.baseMask != 0 {
   235  		// s.baseMask is non-0, elemsize is a power of two, so shift by s.divShift
   236  		return byteOffset >> s.divShift
   237  	}
   238  	return uintptr(((uint64(byteOffset) >> s.divShift) * uint64(s.divMul)) >> s.divShift2)
   239  }
   240  
   241  func markBitsForAddr(p uintptr) markBits {
   242  	s := spanOf(p)
   243  	objIndex := s.objIndex(p)
   244  	return s.markBitsForIndex(objIndex)
   245  }
   246  
   247  func (s *mspan) markBitsForIndex(objIndex uintptr) markBits {
   248  	bytep, mask := s.gcmarkBits.bitp(objIndex)
   249  	return markBits{bytep, mask, objIndex}
   250  }
   251  
   252  func (s *mspan) markBitsForBase() markBits {
   253  	return markBits{(*uint8)(s.gcmarkBits), uint8(1), 0}
   254  }
   255  
   256  // isMarked reports whether mark bit m is set.
   257  func (m markBits) isMarked() bool {
   258  	return *m.bytep&m.mask != 0
   259  }
   260  
   261  // setMarked sets the marked bit in the markbits, atomically.
   262  func (m markBits) setMarked() {
   263  	// Might be racing with other updates, so use atomic update always.
   264  	// We used to be clever here and use a non-atomic update in certain
   265  	// cases, but it's not worth the risk.
   266  	atomic.Or8(m.bytep, m.mask)
   267  }
   268  
   269  // setMarkedNonAtomic sets the marked bit in the markbits, non-atomically.
   270  func (m markBits) setMarkedNonAtomic() {
   271  	*m.bytep |= m.mask
   272  }
   273  
   274  // clearMarked clears the marked bit in the markbits, atomically.
   275  func (m markBits) clearMarked() {
   276  	// Might be racing with other updates, so use atomic update always.
   277  	// We used to be clever here and use a non-atomic update in certain
   278  	// cases, but it's not worth the risk.
   279  	atomic.And8(m.bytep, ^m.mask)
   280  }
   281  
   282  // markBitsForSpan returns the markBits for the span base address base.
   283  func markBitsForSpan(base uintptr) (mbits markBits) {
   284  	mbits = markBitsForAddr(base)
   285  	if mbits.mask != 1 {
   286  		throw("markBitsForSpan: unaligned start")
   287  	}
   288  	return mbits
   289  }
   290  
   291  // advance advances the markBits to the next object in the span.
   292  func (m *markBits) advance() {
   293  	if m.mask == 1<<7 {
   294  		m.bytep = (*uint8)(unsafe.Pointer(uintptr(unsafe.Pointer(m.bytep)) + 1))
   295  		m.mask = 1
   296  	} else {
   297  		m.mask = m.mask << 1
   298  	}
   299  	m.index++
   300  }
   301  
   302  // heapBitsForAddr returns the heapBits for the address addr.
   303  // The caller must ensure addr is in an allocated span.
   304  // In particular, be careful not to point past the end of an object.
   305  //
   306  // nosplit because it is used during write barriers and must not be preempted.
   307  //go:nosplit
   308  func heapBitsForAddr(addr uintptr) (h heapBits) {
   309  	// 2 bits per word, 4 pairs per byte, and a mask is hard coded.
   310  	arena := arenaIndex(addr)
   311  	ha := mheap_.arenas[arena.l1()][arena.l2()]
   312  	// The compiler uses a load for nil checking ha, but in this
   313  	// case we'll almost never hit that cache line again, so it
   314  	// makes more sense to do a value check.
   315  	if ha == nil {
   316  		// addr is not in the heap. Return nil heapBits, which
   317  		// we expect to crash in the caller.
   318  		return
   319  	}
   320  	h.bitp = &ha.bitmap[(addr/(sys.PtrSize*4))%heapArenaBitmapBytes]
   321  	h.shift = uint32((addr / sys.PtrSize) & 3)
   322  	h.arena = uint32(arena)
   323  	h.last = &ha.bitmap[len(ha.bitmap)-1]
   324  	return
   325  }
   326  
   327  // badPointer throws bad pointer in heap panic.
   328  func badPointer(s *mspan, p, refBase, refOff uintptr) {
   329  	// Typically this indicates an incorrect use
   330  	// of unsafe or cgo to store a bad pointer in
   331  	// the Go heap. It may also indicate a runtime
   332  	// bug.
   333  	//
   334  	// TODO(austin): We could be more aggressive
   335  	// and detect pointers to unallocated objects
   336  	// in allocated spans.
   337  	printlock()
   338  	print("runtime: pointer ", hex(p))
   339  	state := s.state.get()
   340  	if state != mSpanInUse {
   341  		print(" to unallocated span")
   342  	} else {
   343  		print(" to unused region of span")
   344  	}
   345  	print(" span.base()=", hex(s.base()), " span.limit=", hex(s.limit), " span.state=", state, "\n")
   346  	if refBase != 0 {
   347  		print("runtime: found in object at *(", hex(refBase), "+", hex(refOff), ")\n")
   348  		gcDumpObject("object", refBase, refOff)
   349  	}
   350  	getg().m.traceback = 2
   351  	throw("found bad pointer in Go heap (incorrect use of unsafe or cgo?)")
   352  }
   353  
   354  // findObject returns the base address for the heap object containing
   355  // the address p, the object's span, and the index of the object in s.
   356  // If p does not point into a heap object, it returns base == 0.
   357  //
   358  // If p points is an invalid heap pointer and debug.invalidptr != 0,
   359  // findObject panics.
   360  //
   361  // refBase and refOff optionally give the base address of the object
   362  // in which the pointer p was found and the byte offset at which it
   363  // was found. These are used for error reporting.
   364  //
   365  // It is nosplit so it is safe for p to be a pointer to the current goroutine's stack.
   366  // Since p is a uintptr, it would not be adjusted if the stack were to move.
   367  //go:nosplit
   368  func findObject(p, refBase, refOff uintptr) (base uintptr, s *mspan, objIndex uintptr) {
   369  	s = spanOf(p)
   370  	// If s is nil, the virtual address has never been part of the heap.
   371  	// This pointer may be to some mmap'd region, so we allow it.
   372  	if s == nil {
   373  		return
   374  	}
   375  	// If p is a bad pointer, it may not be in s's bounds.
   376  	//
   377  	// Check s.state to synchronize with span initialization
   378  	// before checking other fields. See also spanOfHeap.
   379  	if state := s.state.get(); state != mSpanInUse || p < s.base() || p >= s.limit {
   380  		// Pointers into stacks are also ok, the runtime manages these explicitly.
   381  		if state == mSpanManual {
   382  			return
   383  		}
   384  		// The following ensures that we are rigorous about what data
   385  		// structures hold valid pointers.
   386  		if debug.invalidptr != 0 {
   387  			badPointer(s, p, refBase, refOff)
   388  		}
   389  		return
   390  	}
   391  	// If this span holds object of a power of 2 size, just mask off the bits to
   392  	// the interior of the object. Otherwise use the size to get the base.
   393  	if s.baseMask != 0 {
   394  		// optimize for power of 2 sized objects.
   395  		base = s.base()
   396  		base = base + (p-base)&uintptr(s.baseMask)
   397  		objIndex = (base - s.base()) >> s.divShift
   398  		// base = p & s.baseMask is faster for small spans,
   399  		// but doesn't work for large spans.
   400  		// Overall, it's faster to use the more general computation above.
   401  	} else {
   402  		base = s.base()
   403  		if p-base >= s.elemsize {
   404  			// n := (p - base) / s.elemsize, using division by multiplication
   405  			objIndex = uintptr(p-base) >> s.divShift * uintptr(s.divMul) >> s.divShift2
   406  			base += objIndex * s.elemsize
   407  		}
   408  	}
   409  	return
   410  }
   411  
   412  // next returns the heapBits describing the next pointer-sized word in memory.
   413  // That is, if h describes address p, h.next() describes p+ptrSize.
   414  // Note that next does not modify h. The caller must record the result.
   415  //
   416  // nosplit because it is used during write barriers and must not be preempted.
   417  //go:nosplit
   418  func (h heapBits) next() heapBits {
   419  	if h.shift < 3*heapBitsShift {
   420  		h.shift += heapBitsShift
   421  	} else if h.bitp != h.last {
   422  		h.bitp, h.shift = add1(h.bitp), 0
   423  	} else {
   424  		// Move to the next arena.
   425  		return h.nextArena()
   426  	}
   427  	return h
   428  }
   429  
   430  // nextArena advances h to the beginning of the next heap arena.
   431  //
   432  // This is a slow-path helper to next. gc's inliner knows that
   433  // heapBits.next can be inlined even though it calls this. This is
   434  // marked noinline so it doesn't get inlined into next and cause next
   435  // to be too big to inline.
   436  //
   437  //go:nosplit
   438  //go:noinline
   439  func (h heapBits) nextArena() heapBits {
   440  	h.arena++
   441  	ai := arenaIdx(h.arena)
   442  	l2 := mheap_.arenas[ai.l1()]
   443  	if l2 == nil {
   444  		// We just passed the end of the object, which
   445  		// was also the end of the heap. Poison h. It
   446  		// should never be dereferenced at this point.
   447  		return heapBits{}
   448  	}
   449  	ha := l2[ai.l2()]
   450  	if ha == nil {
   451  		return heapBits{}
   452  	}
   453  	h.bitp, h.shift = &ha.bitmap[0], 0
   454  	h.last = &ha.bitmap[len(ha.bitmap)-1]
   455  	return h
   456  }
   457  
   458  // forward returns the heapBits describing n pointer-sized words ahead of h in memory.
   459  // That is, if h describes address p, h.forward(n) describes p+n*ptrSize.
   460  // h.forward(1) is equivalent to h.next(), just slower.
   461  // Note that forward does not modify h. The caller must record the result.
   462  // bits returns the heap bits for the current word.
   463  //go:nosplit
   464  func (h heapBits) forward(n uintptr) heapBits {
   465  	n += uintptr(h.shift) / heapBitsShift
   466  	nbitp := uintptr(unsafe.Pointer(h.bitp)) + n/4
   467  	h.shift = uint32(n%4) * heapBitsShift
   468  	if nbitp <= uintptr(unsafe.Pointer(h.last)) {
   469  		h.bitp = (*uint8)(unsafe.Pointer(nbitp))
   470  		return h
   471  	}
   472  
   473  	// We're in a new heap arena.
   474  	past := nbitp - (uintptr(unsafe.Pointer(h.last)) + 1)
   475  	h.arena += 1 + uint32(past/heapArenaBitmapBytes)
   476  	ai := arenaIdx(h.arena)
   477  	if l2 := mheap_.arenas[ai.l1()]; l2 != nil && l2[ai.l2()] != nil {
   478  		a := l2[ai.l2()]
   479  		h.bitp = &a.bitmap[past%heapArenaBitmapBytes]
   480  		h.last = &a.bitmap[len(a.bitmap)-1]
   481  	} else {
   482  		h.bitp, h.last = nil, nil
   483  	}
   484  	return h
   485  }
   486  
   487  // forwardOrBoundary is like forward, but stops at boundaries between
   488  // contiguous sections of the bitmap. It returns the number of words
   489  // advanced over, which will be <= n.
   490  func (h heapBits) forwardOrBoundary(n uintptr) (heapBits, uintptr) {
   491  	maxn := 4 * ((uintptr(unsafe.Pointer(h.last)) + 1) - uintptr(unsafe.Pointer(h.bitp)))
   492  	if n > maxn {
   493  		n = maxn
   494  	}
   495  	return h.forward(n), n
   496  }
   497  
   498  // The caller can test morePointers and isPointer by &-ing with bitScan and bitPointer.
   499  // The result includes in its higher bits the bits for subsequent words
   500  // described by the same bitmap byte.
   501  //
   502  // nosplit because it is used during write barriers and must not be preempted.
   503  //go:nosplit
   504  func (h heapBits) bits() uint32 {
   505  	// The (shift & 31) eliminates a test and conditional branch
   506  	// from the generated code.
   507  	return uint32(*h.bitp) >> (h.shift & 31)
   508  }
   509  
   510  // morePointers reports whether this word and all remaining words in this object
   511  // are scalars.
   512  // h must not describe the second word of the object.
   513  func (h heapBits) morePointers() bool {
   514  	return h.bits()&bitScan != 0
   515  }
   516  
   517  // isPointer reports whether the heap bits describe a pointer word.
   518  //
   519  // nosplit because it is used during write barriers and must not be preempted.
   520  //go:nosplit
   521  func (h heapBits) isPointer() bool {
   522  	return h.bits()&bitPointer != 0
   523  }
   524  
   525  // bulkBarrierPreWrite executes a write barrier
   526  // for every pointer slot in the memory range [src, src+size),
   527  // using pointer/scalar information from [dst, dst+size).
   528  // This executes the write barriers necessary before a memmove.
   529  // src, dst, and size must be pointer-aligned.
   530  // The range [dst, dst+size) must lie within a single object.
   531  // It does not perform the actual writes.
   532  //
   533  // As a special case, src == 0 indicates that this is being used for a
   534  // memclr. bulkBarrierPreWrite will pass 0 for the src of each write
   535  // barrier.
   536  //
   537  // Callers should call bulkBarrierPreWrite immediately before
   538  // calling memmove(dst, src, size). This function is marked nosplit
   539  // to avoid being preempted; the GC must not stop the goroutine
   540  // between the memmove and the execution of the barriers.
   541  // The caller is also responsible for cgo pointer checks if this
   542  // may be writing Go pointers into non-Go memory.
   543  //
   544  // The pointer bitmap is not maintained for allocations containing
   545  // no pointers at all; any caller of bulkBarrierPreWrite must first
   546  // make sure the underlying allocation contains pointers, usually
   547  // by checking typ.ptrdata.
   548  //
   549  // Callers must perform cgo checks if writeBarrier.cgo.
   550  //
   551  //go:nosplit
   552  func bulkBarrierPreWrite(dst, src, size uintptr) {
   553  	if (dst|src|size)&(sys.PtrSize-1) != 0 {
   554  		throw("bulkBarrierPreWrite: unaligned arguments")
   555  	}
   556  	if !writeBarrier.needed {
   557  		return
   558  	}
   559  	if s := spanOf(dst); s == nil {
   560  		// If dst is a global, use the data or BSS bitmaps to
   561  		// execute write barriers.
   562  		for _, datap := range activeModules() {
   563  			if datap.data <= dst && dst < datap.edata {
   564  				bulkBarrierBitmap(dst, src, size, dst-datap.data, datap.gcdatamask.bytedata)
   565  				return
   566  			}
   567  		}
   568  		for _, datap := range activeModules() {
   569  			if datap.bss <= dst && dst < datap.ebss {
   570  				bulkBarrierBitmap(dst, src, size, dst-datap.bss, datap.gcbssmask.bytedata)
   571  				return
   572  			}
   573  		}
   574  		return
   575  	} else if s.state.get() != mSpanInUse || dst < s.base() || s.limit <= dst {
   576  		// dst was heap memory at some point, but isn't now.
   577  		// It can't be a global. It must be either our stack,
   578  		// or in the case of direct channel sends, it could be
   579  		// another stack. Either way, no need for barriers.
   580  		// This will also catch if dst is in a freed span,
   581  		// though that should never have.
   582  		return
   583  	}
   584  
   585  	buf := &getg().m.p.ptr().wbBuf
   586  	h := heapBitsForAddr(dst)
   587  	if src == 0 {
   588  		for i := uintptr(0); i < size; i += sys.PtrSize {
   589  			if h.isPointer() {
   590  				dstx := (*uintptr)(unsafe.Pointer(dst + i))
   591  				if !buf.putFast(*dstx, 0) {
   592  					wbBufFlush(nil, 0)
   593  				}
   594  			}
   595  			h = h.next()
   596  		}
   597  	} else {
   598  		for i := uintptr(0); i < size; i += sys.PtrSize {
   599  			if h.isPointer() {
   600  				dstx := (*uintptr)(unsafe.Pointer(dst + i))
   601  				srcx := (*uintptr)(unsafe.Pointer(src + i))
   602  				if !buf.putFast(*dstx, *srcx) {
   603  					wbBufFlush(nil, 0)
   604  				}
   605  			}
   606  			h = h.next()
   607  		}
   608  	}
   609  }
   610  
   611  // bulkBarrierPreWriteSrcOnly is like bulkBarrierPreWrite but
   612  // does not execute write barriers for [dst, dst+size).
   613  //
   614  // In addition to the requirements of bulkBarrierPreWrite
   615  // callers need to ensure [dst, dst+size) is zeroed.
   616  //
   617  // This is used for special cases where e.g. dst was just
   618  // created and zeroed with malloc.
   619  //go:nosplit
   620  func bulkBarrierPreWriteSrcOnly(dst, src, size uintptr) {
   621  	if (dst|src|size)&(sys.PtrSize-1) != 0 {
   622  		throw("bulkBarrierPreWrite: unaligned arguments")
   623  	}
   624  	if !writeBarrier.needed {
   625  		return
   626  	}
   627  	buf := &getg().m.p.ptr().wbBuf
   628  	h := heapBitsForAddr(dst)
   629  	for i := uintptr(0); i < size; i += sys.PtrSize {
   630  		if h.isPointer() {
   631  			srcx := (*uintptr)(unsafe.Pointer(src + i))
   632  			if !buf.putFast(0, *srcx) {
   633  				wbBufFlush(nil, 0)
   634  			}
   635  		}
   636  		h = h.next()
   637  	}
   638  }
   639  
   640  // bulkBarrierBitmap executes write barriers for copying from [src,
   641  // src+size) to [dst, dst+size) using a 1-bit pointer bitmap. src is
   642  // assumed to start maskOffset bytes into the data covered by the
   643  // bitmap in bits (which may not be a multiple of 8).
   644  //
   645  // This is used by bulkBarrierPreWrite for writes to data and BSS.
   646  //
   647  //go:nosplit
   648  func bulkBarrierBitmap(dst, src, size, maskOffset uintptr, bits *uint8) {
   649  	word := maskOffset / sys.PtrSize
   650  	bits = addb(bits, word/8)
   651  	mask := uint8(1) << (word % 8)
   652  
   653  	buf := &getg().m.p.ptr().wbBuf
   654  	for i := uintptr(0); i < size; i += sys.PtrSize {
   655  		if mask == 0 {
   656  			bits = addb(bits, 1)
   657  			if *bits == 0 {
   658  				// Skip 8 words.
   659  				i += 7 * sys.PtrSize
   660  				continue
   661  			}
   662  			mask = 1
   663  		}
   664  		if *bits&mask != 0 {
   665  			dstx := (*uintptr)(unsafe.Pointer(dst + i))
   666  			if src == 0 {
   667  				if !buf.putFast(*dstx, 0) {
   668  					wbBufFlush(nil, 0)
   669  				}
   670  			} else {
   671  				srcx := (*uintptr)(unsafe.Pointer(src + i))
   672  				if !buf.putFast(*dstx, *srcx) {
   673  					wbBufFlush(nil, 0)
   674  				}
   675  			}
   676  		}
   677  		mask <<= 1
   678  	}
   679  }
   680  
   681  // typeBitsBulkBarrier executes a write barrier for every
   682  // pointer that would be copied from [src, src+size) to [dst,
   683  // dst+size) by a memmove using the type bitmap to locate those
   684  // pointer slots.
   685  //
   686  // The type typ must correspond exactly to [src, src+size) and [dst, dst+size).
   687  // dst, src, and size must be pointer-aligned.
   688  // The type typ must have a plain bitmap, not a GC program.
   689  // The only use of this function is in channel sends, and the
   690  // 64 kB channel element limit takes care of this for us.
   691  //
   692  // Must not be preempted because it typically runs right before memmove,
   693  // and the GC must observe them as an atomic action.
   694  //
   695  // Callers must perform cgo checks if writeBarrier.cgo.
   696  //
   697  //go:nosplit
   698  func typeBitsBulkBarrier(typ *_type, dst, src, size uintptr) {
   699  	if typ == nil {
   700  		throw("runtime: typeBitsBulkBarrier without type")
   701  	}
   702  	if typ.size != size {
   703  		println("runtime: typeBitsBulkBarrier with type ", typ.string(), " of size ", typ.size, " but memory size", size)
   704  		throw("runtime: invalid typeBitsBulkBarrier")
   705  	}
   706  	if typ.kind&kindGCProg != 0 {
   707  		println("runtime: typeBitsBulkBarrier with type ", typ.string(), " with GC prog")
   708  		throw("runtime: invalid typeBitsBulkBarrier")
   709  	}
   710  	if !writeBarrier.needed {
   711  		return
   712  	}
   713  	ptrmask := typ.gcdata
   714  	buf := &getg().m.p.ptr().wbBuf
   715  	var bits uint32
   716  	for i := uintptr(0); i < typ.ptrdata; i += sys.PtrSize {
   717  		if i&(sys.PtrSize*8-1) == 0 {
   718  			bits = uint32(*ptrmask)
   719  			ptrmask = addb(ptrmask, 1)
   720  		} else {
   721  			bits = bits >> 1
   722  		}
   723  		if bits&1 != 0 {
   724  			dstx := (*uintptr)(unsafe.Pointer(dst + i))
   725  			srcx := (*uintptr)(unsafe.Pointer(src + i))
   726  			if !buf.putFast(*dstx, *srcx) {
   727  				wbBufFlush(nil, 0)
   728  			}
   729  		}
   730  	}
   731  }
   732  
   733  // The methods operating on spans all require that h has been returned
   734  // by heapBitsForSpan and that size, n, total are the span layout description
   735  // returned by the mspan's layout method.
   736  // If total > size*n, it means that there is extra leftover memory in the span,
   737  // usually due to rounding.
   738  //
   739  // TODO(rsc): Perhaps introduce a different heapBitsSpan type.
   740  
   741  // initSpan initializes the heap bitmap for a span.
   742  // If this is a span of pointer-sized objects, it initializes all
   743  // words to pointer/scan.
   744  // Otherwise, it initializes all words to scalar/dead.
   745  func (h heapBits) initSpan(s *mspan) {
   746  	// Clear bits corresponding to objects.
   747  	nw := (s.npages << _PageShift) / sys.PtrSize
   748  	if nw%wordsPerBitmapByte != 0 {
   749  		throw("initSpan: unaligned length")
   750  	}
   751  	if h.shift != 0 {
   752  		throw("initSpan: unaligned base")
   753  	}
   754  	isPtrs := sys.PtrSize == 8 && s.elemsize == sys.PtrSize
   755  	for nw > 0 {
   756  		hNext, anw := h.forwardOrBoundary(nw)
   757  		nbyte := anw / wordsPerBitmapByte
   758  		if isPtrs {
   759  			bitp := h.bitp
   760  			for i := uintptr(0); i < nbyte; i++ {
   761  				*bitp = bitPointerAll | bitScanAll
   762  				bitp = add1(bitp)
   763  			}
   764  		} else {
   765  			memclrNoHeapPointers(unsafe.Pointer(h.bitp), nbyte)
   766  		}
   767  		h = hNext
   768  		nw -= anw
   769  	}
   770  }
   771  
   772  // countAlloc returns the number of objects allocated in span s by
   773  // scanning the allocation bitmap.
   774  func (s *mspan) countAlloc() int {
   775  	count := 0
   776  	bytes := divRoundUp(s.nelems, 8)
   777  	// Iterate over each 8-byte chunk and count allocations
   778  	// with an intrinsic. Note that newMarkBits guarantees that
   779  	// gcmarkBits will be 8-byte aligned, so we don't have to
   780  	// worry about edge cases, irrelevant bits will simply be zero.
   781  	for i := uintptr(0); i < bytes; i += 8 {
   782  		// Extract 64 bits from the byte pointer and get a OnesCount.
   783  		// Note that the unsafe cast here doesn't preserve endianness,
   784  		// but that's OK. We only care about how many bits are 1, not
   785  		// about the order we discover them in.
   786  		mrkBits := *(*uint64)(unsafe.Pointer(s.gcmarkBits.bytep(i)))
   787  		count += sys.OnesCount64(mrkBits)
   788  	}
   789  	return count
   790  }
   791  
   792  // heapBitsSetType records that the new allocation [x, x+size)
   793  // holds in [x, x+dataSize) one or more values of type typ.
   794  // (The number of values is given by dataSize / typ.size.)
   795  // If dataSize < size, the fragment [x+dataSize, x+size) is
   796  // recorded as non-pointer data.
   797  // It is known that the type has pointers somewhere;
   798  // malloc does not call heapBitsSetType when there are no pointers,
   799  // because all free objects are marked as noscan during
   800  // heapBitsSweepSpan.
   801  //
   802  // There can only be one allocation from a given span active at a time,
   803  // and the bitmap for a span always falls on byte boundaries,
   804  // so there are no write-write races for access to the heap bitmap.
   805  // Hence, heapBitsSetType can access the bitmap without atomics.
   806  //
   807  // There can be read-write races between heapBitsSetType and things
   808  // that read the heap bitmap like scanobject. However, since
   809  // heapBitsSetType is only used for objects that have not yet been
   810  // made reachable, readers will ignore bits being modified by this
   811  // function. This does mean this function cannot transiently modify
   812  // bits that belong to neighboring objects. Also, on weakly-ordered
   813  // machines, callers must execute a store/store (publication) barrier
   814  // between calling this function and making the object reachable.
   815  func heapBitsSetType(x, size, dataSize uintptr, typ *_type) {
   816  	const doubleCheck = false // slow but helpful; enable to test modifications to this code
   817  
   818  	const (
   819  		mask1 = bitPointer | bitScan                        // 00010001
   820  		mask2 = bitPointer | bitScan | mask1<<heapBitsShift // 00110011
   821  		mask3 = bitPointer | bitScan | mask2<<heapBitsShift // 01110111
   822  	)
   823  
   824  	// dataSize is always size rounded up to the next malloc size class,
   825  	// except in the case of allocating a defer block, in which case
   826  	// size is sizeof(_defer{}) (at least 6 words) and dataSize may be
   827  	// arbitrarily larger.
   828  	//
   829  	// The checks for size == sys.PtrSize and size == 2*sys.PtrSize can therefore
   830  	// assume that dataSize == size without checking it explicitly.
   831  
   832  	if sys.PtrSize == 8 && size == sys.PtrSize {
   833  		// It's one word and it has pointers, it must be a pointer.
   834  		// Since all allocated one-word objects are pointers
   835  		// (non-pointers are aggregated into tinySize allocations),
   836  		// initSpan sets the pointer bits for us. Nothing to do here.
   837  		if doubleCheck {
   838  			h := heapBitsForAddr(x)
   839  			if !h.isPointer() {
   840  				throw("heapBitsSetType: pointer bit missing")
   841  			}
   842  			if !h.morePointers() {
   843  				throw("heapBitsSetType: scan bit missing")
   844  			}
   845  		}
   846  		return
   847  	}
   848  
   849  	h := heapBitsForAddr(x)
   850  	ptrmask := typ.gcdata // start of 1-bit pointer mask (or GC program, handled below)
   851  
   852  	// 2-word objects only have 4 bitmap bits and 3-word objects only have 6 bitmap bits.
   853  	// Therefore, these objects share a heap bitmap byte with the objects next to them.
   854  	// These are called out as a special case primarily so the code below can assume all
   855  	// objects are at least 4 words long and that their bitmaps start either at the beginning
   856  	// of a bitmap byte, or half-way in (h.shift of 0 and 2 respectively).
   857  
   858  	if size == 2*sys.PtrSize {
   859  		if typ.size == sys.PtrSize {
   860  			// We're allocating a block big enough to hold two pointers.
   861  			// On 64-bit, that means the actual object must be two pointers,
   862  			// or else we'd have used the one-pointer-sized block.
   863  			// On 32-bit, however, this is the 8-byte block, the smallest one.
   864  			// So it could be that we're allocating one pointer and this was
   865  			// just the smallest block available. Distinguish by checking dataSize.
   866  			// (In general the number of instances of typ being allocated is
   867  			// dataSize/typ.size.)
   868  			if sys.PtrSize == 4 && dataSize == sys.PtrSize {
   869  				// 1 pointer object. On 32-bit machines clear the bit for the
   870  				// unused second word.
   871  				*h.bitp &^= (bitPointer | bitScan | (bitPointer|bitScan)<<heapBitsShift) << h.shift
   872  				*h.bitp |= (bitPointer | bitScan) << h.shift
   873  			} else {
   874  				// 2-element array of pointer.
   875  				*h.bitp |= (bitPointer | bitScan | (bitPointer|bitScan)<<heapBitsShift) << h.shift
   876  			}
   877  			return
   878  		}
   879  		// Otherwise typ.size must be 2*sys.PtrSize,
   880  		// and typ.kind&kindGCProg == 0.
   881  		if doubleCheck {
   882  			if typ.size != 2*sys.PtrSize || typ.kind&kindGCProg != 0 {
   883  				print("runtime: heapBitsSetType size=", size, " but typ.size=", typ.size, " gcprog=", typ.kind&kindGCProg != 0, "\n")
   884  				throw("heapBitsSetType")
   885  			}
   886  		}
   887  		b := uint32(*ptrmask)
   888  		hb := b & 3
   889  		hb |= bitScanAll & ((bitScan << (typ.ptrdata / sys.PtrSize)) - 1)
   890  		// Clear the bits for this object so we can set the
   891  		// appropriate ones.
   892  		*h.bitp &^= (bitPointer | bitScan | ((bitPointer | bitScan) << heapBitsShift)) << h.shift
   893  		*h.bitp |= uint8(hb << h.shift)
   894  		return
   895  	} else if size == 3*sys.PtrSize {
   896  		b := uint8(*ptrmask)
   897  		if doubleCheck {
   898  			if b == 0 {
   899  				println("runtime: invalid type ", typ.string())
   900  				throw("heapBitsSetType: called with non-pointer type")
   901  			}
   902  			if sys.PtrSize != 8 {
   903  				throw("heapBitsSetType: unexpected 3 pointer wide size class on 32 bit")
   904  			}
   905  			if typ.kind&kindGCProg != 0 {
   906  				throw("heapBitsSetType: unexpected GC prog for 3 pointer wide size class")
   907  			}
   908  			if typ.size == 2*sys.PtrSize {
   909  				print("runtime: heapBitsSetType size=", size, " but typ.size=", typ.size, "\n")
   910  				throw("heapBitsSetType: inconsistent object sizes")
   911  			}
   912  		}
   913  		if typ.size == sys.PtrSize {
   914  			// The type contains a pointer otherwise heapBitsSetType wouldn't have been called.
   915  			// Since the type is only 1 pointer wide and contains a pointer, its gcdata must be exactly 1.
   916  			if doubleCheck && *typ.gcdata != 1 {
   917  				print("runtime: heapBitsSetType size=", size, " typ.size=", typ.size, "but *typ.gcdata", *typ.gcdata, "\n")
   918  				throw("heapBitsSetType: unexpected gcdata for 1 pointer wide type size in 3 pointer wide size class")
   919  			}
   920  			// 3 element array of pointers. Unrolling ptrmask 3 times into p yields 00000111.
   921  			b = 7
   922  		}
   923  
   924  		hb := b & 7
   925  		// Set bitScan bits for all pointers.
   926  		hb |= hb << wordsPerBitmapByte
   927  		// First bitScan bit is always set since the type contains pointers.
   928  		hb |= bitScan
   929  		// Second bitScan bit needs to also be set if the third bitScan bit is set.
   930  		hb |= hb & (bitScan << (2 * heapBitsShift)) >> 1
   931  
   932  		// For h.shift > 1 heap bits cross a byte boundary and need to be written part
   933  		// to h.bitp and part to the next h.bitp.
   934  		switch h.shift {
   935  		case 0:
   936  			*h.bitp &^= mask3 << 0
   937  			*h.bitp |= hb << 0
   938  		case 1:
   939  			*h.bitp &^= mask3 << 1
   940  			*h.bitp |= hb << 1
   941  		case 2:
   942  			*h.bitp &^= mask2 << 2
   943  			*h.bitp |= (hb & mask2) << 2
   944  			// Two words written to the first byte.
   945  			// Advance two words to get to the next byte.
   946  			h = h.next().next()
   947  			*h.bitp &^= mask1
   948  			*h.bitp |= (hb >> 2) & mask1
   949  		case 3:
   950  			*h.bitp &^= mask1 << 3
   951  			*h.bitp |= (hb & mask1) << 3
   952  			// One word written to the first byte.
   953  			// Advance one word to get to the next byte.
   954  			h = h.next()
   955  			*h.bitp &^= mask2
   956  			*h.bitp |= (hb >> 1) & mask2
   957  		}
   958  		return
   959  	}
   960  
   961  	// Copy from 1-bit ptrmask into 2-bit bitmap.
   962  	// The basic approach is to use a single uintptr as a bit buffer,
   963  	// alternating between reloading the buffer and writing bitmap bytes.
   964  	// In general, one load can supply two bitmap byte writes.
   965  	// This is a lot of lines of code, but it compiles into relatively few
   966  	// machine instructions.
   967  
   968  	outOfPlace := false
   969  	if arenaIndex(x+size-1) != arenaIdx(h.arena) || (doubleCheck && fastrand()%2 == 0) {
   970  		// This object spans heap arenas, so the bitmap may be
   971  		// discontiguous. Unroll it into the object instead
   972  		// and then copy it out.
   973  		//
   974  		// In doubleCheck mode, we randomly do this anyway to
   975  		// stress test the bitmap copying path.
   976  		outOfPlace = true
   977  		h.bitp = (*uint8)(unsafe.Pointer(x))
   978  		h.last = nil
   979  	}
   980  
   981  	var (
   982  		// Ptrmask input.
   983  		p     *byte   // last ptrmask byte read
   984  		b     uintptr // ptrmask bits already loaded
   985  		nb    uintptr // number of bits in b at next read
   986  		endp  *byte   // final ptrmask byte to read (then repeat)
   987  		endnb uintptr // number of valid bits in *endp
   988  		pbits uintptr // alternate source of bits
   989  
   990  		// Heap bitmap output.
   991  		w     uintptr // words processed
   992  		nw    uintptr // number of words to process
   993  		hbitp *byte   // next heap bitmap byte to write
   994  		hb    uintptr // bits being prepared for *hbitp
   995  	)
   996  
   997  	hbitp = h.bitp
   998  
   999  	// Handle GC program. Delayed until this part of the code
  1000  	// so that we can use the same double-checking mechanism
  1001  	// as the 1-bit case. Nothing above could have encountered
  1002  	// GC programs: the cases were all too small.
  1003  	if typ.kind&kindGCProg != 0 {
  1004  		heapBitsSetTypeGCProg(h, typ.ptrdata, typ.size, dataSize, size, addb(typ.gcdata, 4))
  1005  		if doubleCheck {
  1006  			// Double-check the heap bits written by GC program
  1007  			// by running the GC program to create a 1-bit pointer mask
  1008  			// and then jumping to the double-check code below.
  1009  			// This doesn't catch bugs shared between the 1-bit and 4-bit
  1010  			// GC program execution, but it does catch mistakes specific
  1011  			// to just one of those and bugs in heapBitsSetTypeGCProg's
  1012  			// implementation of arrays.
  1013  			lock(&debugPtrmask.lock)
  1014  			if debugPtrmask.data == nil {
  1015  				debugPtrmask.data = (*byte)(persistentalloc(1<<20, 1, &memstats.other_sys))
  1016  			}
  1017  			ptrmask = debugPtrmask.data
  1018  			runGCProg(addb(typ.gcdata, 4), nil, ptrmask, 1)
  1019  		}
  1020  		goto Phase4
  1021  	}
  1022  
  1023  	// Note about sizes:
  1024  	//
  1025  	// typ.size is the number of words in the object,
  1026  	// and typ.ptrdata is the number of words in the prefix
  1027  	// of the object that contains pointers. That is, the final
  1028  	// typ.size - typ.ptrdata words contain no pointers.
  1029  	// This allows optimization of a common pattern where
  1030  	// an object has a small header followed by a large scalar
  1031  	// buffer. If we know the pointers are over, we don't have
  1032  	// to scan the buffer's heap bitmap at all.
  1033  	// The 1-bit ptrmasks are sized to contain only bits for
  1034  	// the typ.ptrdata prefix, zero padded out to a full byte
  1035  	// of bitmap. This code sets nw (below) so that heap bitmap
  1036  	// bits are only written for the typ.ptrdata prefix; if there is
  1037  	// more room in the allocated object, the next heap bitmap
  1038  	// entry is a 00, indicating that there are no more pointers
  1039  	// to scan. So only the ptrmask for the ptrdata bytes is needed.
  1040  	//
  1041  	// Replicated copies are not as nice: if there is an array of
  1042  	// objects with scalar tails, all but the last tail does have to
  1043  	// be initialized, because there is no way to say "skip forward".
  1044  	// However, because of the possibility of a repeated type with
  1045  	// size not a multiple of 4 pointers (one heap bitmap byte),
  1046  	// the code already must handle the last ptrmask byte specially
  1047  	// by treating it as containing only the bits for endnb pointers,
  1048  	// where endnb <= 4. We represent large scalar tails that must
  1049  	// be expanded in the replication by setting endnb larger than 4.
  1050  	// This will have the effect of reading many bits out of b,
  1051  	// but once the real bits are shifted out, b will supply as many
  1052  	// zero bits as we try to read, which is exactly what we need.
  1053  
  1054  	p = ptrmask
  1055  	if typ.size < dataSize {
  1056  		// Filling in bits for an array of typ.
  1057  		// Set up for repetition of ptrmask during main loop.
  1058  		// Note that ptrmask describes only a prefix of
  1059  		const maxBits = sys.PtrSize*8 - 7
  1060  		if typ.ptrdata/sys.PtrSize <= maxBits {
  1061  			// Entire ptrmask fits in uintptr with room for a byte fragment.
  1062  			// Load into pbits and never read from ptrmask again.
  1063  			// This is especially important when the ptrmask has
  1064  			// fewer than 8 bits in it; otherwise the reload in the middle
  1065  			// of the Phase 2 loop would itself need to loop to gather
  1066  			// at least 8 bits.
  1067  
  1068  			// Accumulate ptrmask into b.
  1069  			// ptrmask is sized to describe only typ.ptrdata, but we record
  1070  			// it as describing typ.size bytes, since all the high bits are zero.
  1071  			nb = typ.ptrdata / sys.PtrSize
  1072  			for i := uintptr(0); i < nb; i += 8 {
  1073  				b |= uintptr(*p) << i
  1074  				p = add1(p)
  1075  			}
  1076  			nb = typ.size / sys.PtrSize
  1077  
  1078  			// Replicate ptrmask to fill entire pbits uintptr.
  1079  			// Doubling and truncating is fewer steps than
  1080  			// iterating by nb each time. (nb could be 1.)
  1081  			// Since we loaded typ.ptrdata/sys.PtrSize bits
  1082  			// but are pretending to have typ.size/sys.PtrSize,
  1083  			// there might be no replication necessary/possible.
  1084  			pbits = b
  1085  			endnb = nb
  1086  			if nb+nb <= maxBits {
  1087  				for endnb <= sys.PtrSize*8 {
  1088  					pbits |= pbits << endnb
  1089  					endnb += endnb
  1090  				}
  1091  				// Truncate to a multiple of original ptrmask.
  1092  				// Because nb+nb <= maxBits, nb fits in a byte.
  1093  				// Byte division is cheaper than uintptr division.
  1094  				endnb = uintptr(maxBits/byte(nb)) * nb
  1095  				pbits &= 1<<endnb - 1
  1096  				b = pbits
  1097  				nb = endnb
  1098  			}
  1099  
  1100  			// Clear p and endp as sentinel for using pbits.
  1101  			// Checked during Phase 2 loop.
  1102  			p = nil
  1103  			endp = nil
  1104  		} else {
  1105  			// Ptrmask is larger. Read it multiple times.
  1106  			n := (typ.ptrdata/sys.PtrSize+7)/8 - 1
  1107  			endp = addb(ptrmask, n)
  1108  			endnb = typ.size/sys.PtrSize - n*8
  1109  		}
  1110  	}
  1111  	if p != nil {
  1112  		b = uintptr(*p)
  1113  		p = add1(p)
  1114  		nb = 8
  1115  	}
  1116  
  1117  	if typ.size == dataSize {
  1118  		// Single entry: can stop once we reach the non-pointer data.
  1119  		nw = typ.ptrdata / sys.PtrSize
  1120  	} else {
  1121  		// Repeated instances of typ in an array.
  1122  		// Have to process first N-1 entries in full, but can stop
  1123  		// once we reach the non-pointer data in the final entry.
  1124  		nw = ((dataSize/typ.size-1)*typ.size + typ.ptrdata) / sys.PtrSize
  1125  	}
  1126  	if nw == 0 {
  1127  		// No pointers! Caller was supposed to check.
  1128  		println("runtime: invalid type ", typ.string())
  1129  		throw("heapBitsSetType: called with non-pointer type")
  1130  		return
  1131  	}
  1132  
  1133  	// Phase 1: Special case for leading byte (shift==0) or half-byte (shift==2).
  1134  	// The leading byte is special because it contains the bits for word 1,
  1135  	// which does not have the scan bit set.
  1136  	// The leading half-byte is special because it's a half a byte,
  1137  	// so we have to be careful with the bits already there.
  1138  	switch {
  1139  	default:
  1140  		throw("heapBitsSetType: unexpected shift")
  1141  
  1142  	case h.shift == 0:
  1143  		// Ptrmask and heap bitmap are aligned.
  1144  		//
  1145  		// This is a fast path for small objects.
  1146  		//
  1147  		// The first byte we write out covers the first four
  1148  		// words of the object. The scan/dead bit on the first
  1149  		// word must be set to scan since there are pointers
  1150  		// somewhere in the object.
  1151  		// In all following words, we set the scan/dead
  1152  		// appropriately to indicate that the object continues
  1153  		// to the next 2-bit entry in the bitmap.
  1154  		//
  1155  		// We set four bits at a time here, but if the object
  1156  		// is fewer than four words, phase 3 will clear
  1157  		// unnecessary bits.
  1158  		hb = b & bitPointerAll
  1159  		hb |= bitScanAll
  1160  		if w += 4; w >= nw {
  1161  			goto Phase3
  1162  		}
  1163  		*hbitp = uint8(hb)
  1164  		hbitp = add1(hbitp)
  1165  		b >>= 4
  1166  		nb -= 4
  1167  
  1168  	case h.shift == 2:
  1169  		// Ptrmask and heap bitmap are misaligned.
  1170  		//
  1171  		// On 32 bit architectures only the 6-word object that corresponds
  1172  		// to a 24 bytes size class can start with h.shift of 2 here since
  1173  		// all other non 16 byte aligned size classes have been handled by
  1174  		// special code paths at the beginning of heapBitsSetType on 32 bit.
  1175  		//
  1176  		// Many size classes are only 16 byte aligned. On 64 bit architectures
  1177  		// this results in a heap bitmap position starting with a h.shift of 2.
  1178  		//
  1179  		// The bits for the first two words are in a byte shared
  1180  		// with another object, so we must be careful with the bits
  1181  		// already there.
  1182  		//
  1183  		// We took care of 1-word, 2-word, and 3-word objects above,
  1184  		// so this is at least a 6-word object.
  1185  		hb = (b & (bitPointer | bitPointer<<heapBitsShift)) << (2 * heapBitsShift)
  1186  		hb |= bitScan << (2 * heapBitsShift)
  1187  		if nw > 1 {
  1188  			hb |= bitScan << (3 * heapBitsShift)
  1189  		}
  1190  		b >>= 2
  1191  		nb -= 2
  1192  		*hbitp &^= uint8((bitPointer | bitScan | ((bitPointer | bitScan) << heapBitsShift)) << (2 * heapBitsShift))
  1193  		*hbitp |= uint8(hb)
  1194  		hbitp = add1(hbitp)
  1195  		if w += 2; w >= nw {
  1196  			// We know that there is more data, because we handled 2-word and 3-word objects above.
  1197  			// This must be at least a 6-word object. If we're out of pointer words,
  1198  			// mark no scan in next bitmap byte and finish.
  1199  			hb = 0
  1200  			w += 4
  1201  			goto Phase3
  1202  		}
  1203  	}
  1204  
  1205  	// Phase 2: Full bytes in bitmap, up to but not including write to last byte (full or partial) in bitmap.
  1206  	// The loop computes the bits for that last write but does not execute the write;
  1207  	// it leaves the bits in hb for processing by phase 3.
  1208  	// To avoid repeated adjustment of nb, we subtract out the 4 bits we're going to
  1209  	// use in the first half of the loop right now, and then we only adjust nb explicitly
  1210  	// if the 8 bits used by each iteration isn't balanced by 8 bits loaded mid-loop.
  1211  	nb -= 4
  1212  	for {
  1213  		// Emit bitmap byte.
  1214  		// b has at least nb+4 bits, with one exception:
  1215  		// if w+4 >= nw, then b has only nw-w bits,
  1216  		// but we'll stop at the break and then truncate
  1217  		// appropriately in Phase 3.
  1218  		hb = b & bitPointerAll
  1219  		hb |= bitScanAll
  1220  		if w += 4; w >= nw {
  1221  			break
  1222  		}
  1223  		*hbitp = uint8(hb)
  1224  		hbitp = add1(hbitp)
  1225  		b >>= 4
  1226  
  1227  		// Load more bits. b has nb right now.
  1228  		if p != endp {
  1229  			// Fast path: keep reading from ptrmask.
  1230  			// nb unmodified: we just loaded 8 bits,
  1231  			// and the next iteration will consume 8 bits,
  1232  			// leaving us with the same nb the next time we're here.
  1233  			if nb < 8 {
  1234  				b |= uintptr(*p) << nb
  1235  				p = add1(p)
  1236  			} else {
  1237  				// Reduce the number of bits in b.
  1238  				// This is important if we skipped
  1239  				// over a scalar tail, since nb could
  1240  				// be larger than the bit width of b.
  1241  				nb -= 8
  1242  			}
  1243  		} else if p == nil {
  1244  			// Almost as fast path: track bit count and refill from pbits.
  1245  			// For short repetitions.
  1246  			if nb < 8 {
  1247  				b |= pbits << nb
  1248  				nb += endnb
  1249  			}
  1250  			nb -= 8 // for next iteration
  1251  		} else {
  1252  			// Slow path: reached end of ptrmask.
  1253  			// Process final partial byte and rewind to start.
  1254  			b |= uintptr(*p) << nb
  1255  			nb += endnb
  1256  			if nb < 8 {
  1257  				b |= uintptr(*ptrmask) << nb
  1258  				p = add1(ptrmask)
  1259  			} else {
  1260  				nb -= 8
  1261  				p = ptrmask
  1262  			}
  1263  		}
  1264  
  1265  		// Emit bitmap byte.
  1266  		hb = b & bitPointerAll
  1267  		hb |= bitScanAll
  1268  		if w += 4; w >= nw {
  1269  			break
  1270  		}
  1271  		*hbitp = uint8(hb)
  1272  		hbitp = add1(hbitp)
  1273  		b >>= 4
  1274  	}
  1275  
  1276  Phase3:
  1277  	// Phase 3: Write last byte or partial byte and zero the rest of the bitmap entries.
  1278  	if w > nw {
  1279  		// Counting the 4 entries in hb not yet written to memory,
  1280  		// there are more entries than possible pointer slots.
  1281  		// Discard the excess entries (can't be more than 3).
  1282  		mask := uintptr(1)<<(4-(w-nw)) - 1
  1283  		hb &= mask | mask<<4 // apply mask to both pointer bits and scan bits
  1284  	}
  1285  
  1286  	// Change nw from counting possibly-pointer words to total words in allocation.
  1287  	nw = size / sys.PtrSize
  1288  
  1289  	// Write whole bitmap bytes.
  1290  	// The first is hb, the rest are zero.
  1291  	if w <= nw {
  1292  		*hbitp = uint8(hb)
  1293  		hbitp = add1(hbitp)
  1294  		hb = 0 // for possible final half-byte below
  1295  		for w += 4; w <= nw; w += 4 {
  1296  			*hbitp = 0
  1297  			hbitp = add1(hbitp)
  1298  		}
  1299  	}
  1300  
  1301  	// Write final partial bitmap byte if any.
  1302  	// We know w > nw, or else we'd still be in the loop above.
  1303  	// It can be bigger only due to the 4 entries in hb that it counts.
  1304  	// If w == nw+4 then there's nothing left to do: we wrote all nw entries
  1305  	// and can discard the 4 sitting in hb.
  1306  	// But if w == nw+2, we need to write first two in hb.
  1307  	// The byte is shared with the next object, so be careful with
  1308  	// existing bits.
  1309  	if w == nw+2 {
  1310  		*hbitp = *hbitp&^(bitPointer|bitScan|(bitPointer|bitScan)<<heapBitsShift) | uint8(hb)
  1311  	}
  1312  
  1313  Phase4:
  1314  	// Phase 4: Copy unrolled bitmap to per-arena bitmaps, if necessary.
  1315  	if outOfPlace {
  1316  		// TODO: We could probably make this faster by
  1317  		// handling [x+dataSize, x+size) specially.
  1318  		h := heapBitsForAddr(x)
  1319  		// cnw is the number of heap words, or bit pairs
  1320  		// remaining (like nw above).
  1321  		cnw := size / sys.PtrSize
  1322  		src := (*uint8)(unsafe.Pointer(x))
  1323  		// We know the first and last byte of the bitmap are
  1324  		// not the same, but it's still possible for small
  1325  		// objects span arenas, so it may share bitmap bytes
  1326  		// with neighboring objects.
  1327  		//
  1328  		// Handle the first byte specially if it's shared. See
  1329  		// Phase 1 for why this is the only special case we need.
  1330  		if doubleCheck {
  1331  			if !(h.shift == 0 || h.shift == 2) {
  1332  				print("x=", x, " size=", size, " cnw=", h.shift, "\n")
  1333  				throw("bad start shift")
  1334  			}
  1335  		}
  1336  		if h.shift == 2 {
  1337  			*h.bitp = *h.bitp&^((bitPointer|bitScan|(bitPointer|bitScan)<<heapBitsShift)<<(2*heapBitsShift)) | *src
  1338  			h = h.next().next()
  1339  			cnw -= 2
  1340  			src = addb(src, 1)
  1341  		}
  1342  		// We're now byte aligned. Copy out to per-arena
  1343  		// bitmaps until the last byte (which may again be
  1344  		// partial).
  1345  		for cnw >= 4 {
  1346  			// This loop processes four words at a time,
  1347  			// so round cnw down accordingly.
  1348  			hNext, words := h.forwardOrBoundary(cnw / 4 * 4)
  1349  
  1350  			// n is the number of bitmap bytes to copy.
  1351  			n := words / 4
  1352  			memmove(unsafe.Pointer(h.bitp), unsafe.Pointer(src), n)
  1353  			cnw -= words
  1354  			h = hNext
  1355  			src = addb(src, n)
  1356  		}
  1357  		if doubleCheck && h.shift != 0 {
  1358  			print("cnw=", cnw, " h.shift=", h.shift, "\n")
  1359  			throw("bad shift after block copy")
  1360  		}
  1361  		// Handle the last byte if it's shared.
  1362  		if cnw == 2 {
  1363  			*h.bitp = *h.bitp&^(bitPointer|bitScan|(bitPointer|bitScan)<<heapBitsShift) | *src
  1364  			src = addb(src, 1)
  1365  			h = h.next().next()
  1366  		}
  1367  		if doubleCheck {
  1368  			if uintptr(unsafe.Pointer(src)) > x+size {
  1369  				throw("copy exceeded object size")
  1370  			}
  1371  			if !(cnw == 0 || cnw == 2) {
  1372  				print("x=", x, " size=", size, " cnw=", cnw, "\n")
  1373  				throw("bad number of remaining words")
  1374  			}
  1375  			// Set up hbitp so doubleCheck code below can check it.
  1376  			hbitp = h.bitp
  1377  		}
  1378  		// Zero the object where we wrote the bitmap.
  1379  		memclrNoHeapPointers(unsafe.Pointer(x), uintptr(unsafe.Pointer(src))-x)
  1380  	}
  1381  
  1382  	// Double check the whole bitmap.
  1383  	if doubleCheck {
  1384  		// x+size may not point to the heap, so back up one
  1385  		// word and then advance it the way we do above.
  1386  		end := heapBitsForAddr(x + size - sys.PtrSize)
  1387  		if outOfPlace {
  1388  			// In out-of-place copying, we just advance
  1389  			// using next.
  1390  			end = end.next()
  1391  		} else {
  1392  			// Don't use next because that may advance to
  1393  			// the next arena and the in-place logic
  1394  			// doesn't do that.
  1395  			end.shift += heapBitsShift
  1396  			if end.shift == 4*heapBitsShift {
  1397  				end.bitp, end.shift = add1(end.bitp), 0
  1398  			}
  1399  		}
  1400  		if typ.kind&kindGCProg == 0 && (hbitp != end.bitp || (w == nw+2) != (end.shift == 2)) {
  1401  			println("ended at wrong bitmap byte for", typ.string(), "x", dataSize/typ.size)
  1402  			print("typ.size=", typ.size, " typ.ptrdata=", typ.ptrdata, " dataSize=", dataSize, " size=", size, "\n")
  1403  			print("w=", w, " nw=", nw, " b=", hex(b), " nb=", nb, " hb=", hex(hb), "\n")
  1404  			h0 := heapBitsForAddr(x)
  1405  			print("initial bits h0.bitp=", h0.bitp, " h0.shift=", h0.shift, "\n")
  1406  			print("ended at hbitp=", hbitp, " but next starts at bitp=", end.bitp, " shift=", end.shift, "\n")
  1407  			throw("bad heapBitsSetType")
  1408  		}
  1409  
  1410  		// Double-check that bits to be written were written correctly.
  1411  		// Does not check that other bits were not written, unfortunately.
  1412  		h := heapBitsForAddr(x)
  1413  		nptr := typ.ptrdata / sys.PtrSize
  1414  		ndata := typ.size / sys.PtrSize
  1415  		count := dataSize / typ.size
  1416  		totalptr := ((count-1)*typ.size + typ.ptrdata) / sys.PtrSize
  1417  		for i := uintptr(0); i < size/sys.PtrSize; i++ {
  1418  			j := i % ndata
  1419  			var have, want uint8
  1420  			have = (*h.bitp >> h.shift) & (bitPointer | bitScan)
  1421  			if i >= totalptr {
  1422  				if typ.kind&kindGCProg != 0 && i < (totalptr+3)/4*4 {
  1423  					// heapBitsSetTypeGCProg always fills
  1424  					// in full nibbles of bitScan.
  1425  					want = bitScan
  1426  				}
  1427  			} else {
  1428  				if j < nptr && (*addb(ptrmask, j/8)>>(j%8))&1 != 0 {
  1429  					want |= bitPointer
  1430  				}
  1431  				want |= bitScan
  1432  			}
  1433  			if have != want {
  1434  				println("mismatch writing bits for", typ.string(), "x", dataSize/typ.size)
  1435  				print("typ.size=", typ.size, " typ.ptrdata=", typ.ptrdata, " dataSize=", dataSize, " size=", size, "\n")
  1436  				print("kindGCProg=", typ.kind&kindGCProg != 0, " outOfPlace=", outOfPlace, "\n")
  1437  				print("w=", w, " nw=", nw, " b=", hex(b), " nb=", nb, " hb=", hex(hb), "\n")
  1438  				h0 := heapBitsForAddr(x)
  1439  				print("initial bits h0.bitp=", h0.bitp, " h0.shift=", h0.shift, "\n")
  1440  				print("current bits h.bitp=", h.bitp, " h.shift=", h.shift, " *h.bitp=", hex(*h.bitp), "\n")
  1441  				print("ptrmask=", ptrmask, " p=", p, " endp=", endp, " endnb=", endnb, " pbits=", hex(pbits), " b=", hex(b), " nb=", nb, "\n")
  1442  				println("at word", i, "offset", i*sys.PtrSize, "have", hex(have), "want", hex(want))
  1443  				if typ.kind&kindGCProg != 0 {
  1444  					println("GC program:")
  1445  					dumpGCProg(addb(typ.gcdata, 4))
  1446  				}
  1447  				throw("bad heapBitsSetType")
  1448  			}
  1449  			h = h.next()
  1450  		}
  1451  		if ptrmask == debugPtrmask.data {
  1452  			unlock(&debugPtrmask.lock)
  1453  		}
  1454  	}
  1455  }
  1456  
  1457  var debugPtrmask struct {
  1458  	lock mutex
  1459  	data *byte
  1460  }
  1461  
  1462  // heapBitsSetTypeGCProg implements heapBitsSetType using a GC program.
  1463  // progSize is the size of the memory described by the program.
  1464  // elemSize is the size of the element that the GC program describes (a prefix of).
  1465  // dataSize is the total size of the intended data, a multiple of elemSize.
  1466  // allocSize is the total size of the allocated memory.
  1467  //
  1468  // GC programs are only used for large allocations.
  1469  // heapBitsSetType requires that allocSize is a multiple of 4 words,
  1470  // so that the relevant bitmap bytes are not shared with surrounding
  1471  // objects.
  1472  func heapBitsSetTypeGCProg(h heapBits, progSize, elemSize, dataSize, allocSize uintptr, prog *byte) {
  1473  	if sys.PtrSize == 8 && allocSize%(4*sys.PtrSize) != 0 {
  1474  		// Alignment will be wrong.
  1475  		throw("heapBitsSetTypeGCProg: small allocation")
  1476  	}
  1477  	var totalBits uintptr
  1478  	if elemSize == dataSize {
  1479  		totalBits = runGCProg(prog, nil, h.bitp, 2)
  1480  		if totalBits*sys.PtrSize != progSize {
  1481  			println("runtime: heapBitsSetTypeGCProg: total bits", totalBits, "but progSize", progSize)
  1482  			throw("heapBitsSetTypeGCProg: unexpected bit count")
  1483  		}
  1484  	} else {
  1485  		count := dataSize / elemSize
  1486  
  1487  		// Piece together program trailer to run after prog that does:
  1488  		//	literal(0)
  1489  		//	repeat(1, elemSize-progSize-1) // zeros to fill element size
  1490  		//	repeat(elemSize, count-1) // repeat that element for count
  1491  		// This zero-pads the data remaining in the first element and then
  1492  		// repeats that first element to fill the array.
  1493  		var trailer [40]byte // 3 varints (max 10 each) + some bytes
  1494  		i := 0
  1495  		if n := elemSize/sys.PtrSize - progSize/sys.PtrSize; n > 0 {
  1496  			// literal(0)
  1497  			trailer[i] = 0x01
  1498  			i++
  1499  			trailer[i] = 0
  1500  			i++
  1501  			if n > 1 {
  1502  				// repeat(1, n-1)
  1503  				trailer[i] = 0x81
  1504  				i++
  1505  				n--
  1506  				for ; n >= 0x80; n >>= 7 {
  1507  					trailer[i] = byte(n | 0x80)
  1508  					i++
  1509  				}
  1510  				trailer[i] = byte(n)
  1511  				i++
  1512  			}
  1513  		}
  1514  		// repeat(elemSize/ptrSize, count-1)
  1515  		trailer[i] = 0x80
  1516  		i++
  1517  		n := elemSize / sys.PtrSize
  1518  		for ; n >= 0x80; n >>= 7 {
  1519  			trailer[i] = byte(n | 0x80)
  1520  			i++
  1521  		}
  1522  		trailer[i] = byte(n)
  1523  		i++
  1524  		n = count - 1
  1525  		for ; n >= 0x80; n >>= 7 {
  1526  			trailer[i] = byte(n | 0x80)
  1527  			i++
  1528  		}
  1529  		trailer[i] = byte(n)
  1530  		i++
  1531  		trailer[i] = 0
  1532  		i++
  1533  
  1534  		runGCProg(prog, &trailer[0], h.bitp, 2)
  1535  
  1536  		// Even though we filled in the full array just now,
  1537  		// record that we only filled in up to the ptrdata of the
  1538  		// last element. This will cause the code below to
  1539  		// memclr the dead section of the final array element,
  1540  		// so that scanobject can stop early in the final element.
  1541  		totalBits = (elemSize*(count-1) + progSize) / sys.PtrSize
  1542  	}
  1543  	endProg := unsafe.Pointer(addb(h.bitp, (totalBits+3)/4))
  1544  	endAlloc := unsafe.Pointer(addb(h.bitp, allocSize/sys.PtrSize/wordsPerBitmapByte))
  1545  	memclrNoHeapPointers(endProg, uintptr(endAlloc)-uintptr(endProg))
  1546  }
  1547  
  1548  // progToPointerMask returns the 1-bit pointer mask output by the GC program prog.
  1549  // size the size of the region described by prog, in bytes.
  1550  // The resulting bitvector will have no more than size/sys.PtrSize bits.
  1551  func progToPointerMask(prog *byte, size uintptr) bitvector {
  1552  	n := (size/sys.PtrSize + 7) / 8
  1553  	x := (*[1 << 30]byte)(persistentalloc(n+1, 1, &memstats.buckhash_sys))[:n+1]
  1554  	x[len(x)-1] = 0xa1 // overflow check sentinel
  1555  	n = runGCProg(prog, nil, &x[0], 1)
  1556  	if x[len(x)-1] != 0xa1 {
  1557  		throw("progToPointerMask: overflow")
  1558  	}
  1559  	return bitvector{int32(n), &x[0]}
  1560  }
  1561  
  1562  // Packed GC pointer bitmaps, aka GC programs.
  1563  //
  1564  // For large types containing arrays, the type information has a
  1565  // natural repetition that can be encoded to save space in the
  1566  // binary and in the memory representation of the type information.
  1567  //
  1568  // The encoding is a simple Lempel-Ziv style bytecode machine
  1569  // with the following instructions:
  1570  //
  1571  //	00000000: stop
  1572  //	0nnnnnnn: emit n bits copied from the next (n+7)/8 bytes
  1573  //	10000000 n c: repeat the previous n bits c times; n, c are varints
  1574  //	1nnnnnnn c: repeat the previous n bits c times; c is a varint
  1575  
  1576  // runGCProg executes the GC program prog, and then trailer if non-nil,
  1577  // writing to dst with entries of the given size.
  1578  // If size == 1, dst is a 1-bit pointer mask laid out moving forward from dst.
  1579  // If size == 2, dst is the 2-bit heap bitmap, and writes move backward
  1580  // starting at dst (because the heap bitmap does). In this case, the caller guarantees
  1581  // that only whole bytes in dst need to be written.
  1582  //
  1583  // runGCProg returns the number of 1- or 2-bit entries written to memory.
  1584  func runGCProg(prog, trailer, dst *byte, size int) uintptr {
  1585  	dstStart := dst
  1586  
  1587  	// Bits waiting to be written to memory.
  1588  	var bits uintptr
  1589  	var nbits uintptr
  1590  
  1591  	p := prog
  1592  Run:
  1593  	for {
  1594  		// Flush accumulated full bytes.
  1595  		// The rest of the loop assumes that nbits <= 7.
  1596  		for ; nbits >= 8; nbits -= 8 {
  1597  			if size == 1 {
  1598  				*dst = uint8(bits)
  1599  				dst = add1(dst)
  1600  				bits >>= 8
  1601  			} else {
  1602  				v := bits&bitPointerAll | bitScanAll
  1603  				*dst = uint8(v)
  1604  				dst = add1(dst)
  1605  				bits >>= 4
  1606  				v = bits&bitPointerAll | bitScanAll
  1607  				*dst = uint8(v)
  1608  				dst = add1(dst)
  1609  				bits >>= 4
  1610  			}
  1611  		}
  1612  
  1613  		// Process one instruction.
  1614  		inst := uintptr(*p)
  1615  		p = add1(p)
  1616  		n := inst & 0x7F
  1617  		if inst&0x80 == 0 {
  1618  			// Literal bits; n == 0 means end of program.
  1619  			if n == 0 {
  1620  				// Program is over; continue in trailer if present.
  1621  				if trailer != nil {
  1622  					p = trailer
  1623  					trailer = nil
  1624  					continue
  1625  				}
  1626  				break Run
  1627  			}
  1628  			nbyte := n / 8
  1629  			for i := uintptr(0); i < nbyte; i++ {
  1630  				bits |= uintptr(*p) << nbits
  1631  				p = add1(p)
  1632  				if size == 1 {
  1633  					*dst = uint8(bits)
  1634  					dst = add1(dst)
  1635  					bits >>= 8
  1636  				} else {
  1637  					v := bits&0xf | bitScanAll
  1638  					*dst = uint8(v)
  1639  					dst = add1(dst)
  1640  					bits >>= 4
  1641  					v = bits&0xf | bitScanAll
  1642  					*dst = uint8(v)
  1643  					dst = add1(dst)
  1644  					bits >>= 4
  1645  				}
  1646  			}
  1647  			if n %= 8; n > 0 {
  1648  				bits |= uintptr(*p) << nbits
  1649  				p = add1(p)
  1650  				nbits += n
  1651  			}
  1652  			continue Run
  1653  		}
  1654  
  1655  		// Repeat. If n == 0, it is encoded in a varint in the next bytes.
  1656  		if n == 0 {
  1657  			for off := uint(0); ; off += 7 {
  1658  				x := uintptr(*p)
  1659  				p = add1(p)
  1660  				n |= (x & 0x7F) << off
  1661  				if x&0x80 == 0 {
  1662  					break
  1663  				}
  1664  			}
  1665  		}
  1666  
  1667  		// Count is encoded in a varint in the next bytes.
  1668  		c := uintptr(0)
  1669  		for off := uint(0); ; off += 7 {
  1670  			x := uintptr(*p)
  1671  			p = add1(p)
  1672  			c |= (x & 0x7F) << off
  1673  			if x&0x80 == 0 {
  1674  				break
  1675  			}
  1676  		}
  1677  		c *= n // now total number of bits to copy
  1678  
  1679  		// If the number of bits being repeated is small, load them
  1680  		// into a register and use that register for the entire loop
  1681  		// instead of repeatedly reading from memory.
  1682  		// Handling fewer than 8 bits here makes the general loop simpler.
  1683  		// The cutoff is sys.PtrSize*8 - 7 to guarantee that when we add
  1684  		// the pattern to a bit buffer holding at most 7 bits (a partial byte)
  1685  		// it will not overflow.
  1686  		src := dst
  1687  		const maxBits = sys.PtrSize*8 - 7
  1688  		if n <= maxBits {
  1689  			// Start with bits in output buffer.
  1690  			pattern := bits
  1691  			npattern := nbits
  1692  
  1693  			// If we need more bits, fetch them from memory.
  1694  			if size == 1 {
  1695  				src = subtract1(src)
  1696  				for npattern < n {
  1697  					pattern <<= 8
  1698  					pattern |= uintptr(*src)
  1699  					src = subtract1(src)
  1700  					npattern += 8
  1701  				}
  1702  			} else {
  1703  				src = subtract1(src)
  1704  				for npattern < n {
  1705  					pattern <<= 4
  1706  					pattern |= uintptr(*src) & 0xf
  1707  					src = subtract1(src)
  1708  					npattern += 4
  1709  				}
  1710  			}
  1711  
  1712  			// We started with the whole bit output buffer,
  1713  			// and then we loaded bits from whole bytes.
  1714  			// Either way, we might now have too many instead of too few.
  1715  			// Discard the extra.
  1716  			if npattern > n {
  1717  				pattern >>= npattern - n
  1718  				npattern = n
  1719  			}
  1720  
  1721  			// Replicate pattern to at most maxBits.
  1722  			if npattern == 1 {
  1723  				// One bit being repeated.
  1724  				// If the bit is 1, make the pattern all 1s.
  1725  				// If the bit is 0, the pattern is already all 0s,
  1726  				// but we can claim that the number of bits
  1727  				// in the word is equal to the number we need (c),
  1728  				// because right shift of bits will zero fill.
  1729  				if pattern == 1 {
  1730  					pattern = 1<<maxBits - 1
  1731  					npattern = maxBits
  1732  				} else {
  1733  					npattern = c
  1734  				}
  1735  			} else {
  1736  				b := pattern
  1737  				nb := npattern
  1738  				if nb+nb <= maxBits {
  1739  					// Double pattern until the whole uintptr is filled.
  1740  					for nb <= sys.PtrSize*8 {
  1741  						b |= b << nb
  1742  						nb += nb
  1743  					}
  1744  					// Trim away incomplete copy of original pattern in high bits.
  1745  					// TODO(rsc): Replace with table lookup or loop on systems without divide?
  1746  					nb = maxBits / npattern * npattern
  1747  					b &= 1<<nb - 1
  1748  					pattern = b
  1749  					npattern = nb
  1750  				}
  1751  			}
  1752  
  1753  			// Add pattern to bit buffer and flush bit buffer, c/npattern times.
  1754  			// Since pattern contains >8 bits, there will be full bytes to flush
  1755  			// on each iteration.
  1756  			for ; c >= npattern; c -= npattern {
  1757  				bits |= pattern << nbits
  1758  				nbits += npattern
  1759  				if size == 1 {
  1760  					for nbits >= 8 {
  1761  						*dst = uint8(bits)
  1762  						dst = add1(dst)
  1763  						bits >>= 8
  1764  						nbits -= 8
  1765  					}
  1766  				} else {
  1767  					for nbits >= 4 {
  1768  						*dst = uint8(bits&0xf | bitScanAll)
  1769  						dst = add1(dst)
  1770  						bits >>= 4
  1771  						nbits -= 4
  1772  					}
  1773  				}
  1774  			}
  1775  
  1776  			// Add final fragment to bit buffer.
  1777  			if c > 0 {
  1778  				pattern &= 1<<c - 1
  1779  				bits |= pattern << nbits
  1780  				nbits += c
  1781  			}
  1782  			continue Run
  1783  		}
  1784  
  1785  		// Repeat; n too large to fit in a register.
  1786  		// Since nbits <= 7, we know the first few bytes of repeated data
  1787  		// are already written to memory.
  1788  		off := n - nbits // n > nbits because n > maxBits and nbits <= 7
  1789  		if size == 1 {
  1790  			// Leading src fragment.
  1791  			src = subtractb(src, (off+7)/8)
  1792  			if frag := off & 7; frag != 0 {
  1793  				bits |= uintptr(*src) >> (8 - frag) << nbits
  1794  				src = add1(src)
  1795  				nbits += frag
  1796  				c -= frag
  1797  			}
  1798  			// Main loop: load one byte, write another.
  1799  			// The bits are rotating through the bit buffer.
  1800  			for i := c / 8; i > 0; i-- {
  1801  				bits |= uintptr(*src) << nbits
  1802  				src = add1(src)
  1803  				*dst = uint8(bits)
  1804  				dst = add1(dst)
  1805  				bits >>= 8
  1806  			}
  1807  			// Final src fragment.
  1808  			if c %= 8; c > 0 {
  1809  				bits |= (uintptr(*src) & (1<<c - 1)) << nbits
  1810  				nbits += c
  1811  			}
  1812  		} else {
  1813  			// Leading src fragment.
  1814  			src = subtractb(src, (off+3)/4)
  1815  			if frag := off & 3; frag != 0 {
  1816  				bits |= (uintptr(*src) & 0xf) >> (4 - frag) << nbits
  1817  				src = add1(src)
  1818  				nbits += frag
  1819  				c -= frag
  1820  			}
  1821  			// Main loop: load one byte, write another.
  1822  			// The bits are rotating through the bit buffer.
  1823  			for i := c / 4; i > 0; i-- {
  1824  				bits |= (uintptr(*src) & 0xf) << nbits
  1825  				src = add1(src)
  1826  				*dst = uint8(bits&0xf | bitScanAll)
  1827  				dst = add1(dst)
  1828  				bits >>= 4
  1829  			}
  1830  			// Final src fragment.
  1831  			if c %= 4; c > 0 {
  1832  				bits |= (uintptr(*src) & (1<<c - 1)) << nbits
  1833  				nbits += c
  1834  			}
  1835  		}
  1836  	}
  1837  
  1838  	// Write any final bits out, using full-byte writes, even for the final byte.
  1839  	var totalBits uintptr
  1840  	if size == 1 {
  1841  		totalBits = (uintptr(unsafe.Pointer(dst))-uintptr(unsafe.Pointer(dstStart)))*8 + nbits
  1842  		nbits += -nbits & 7
  1843  		for ; nbits > 0; nbits -= 8 {
  1844  			*dst = uint8(bits)
  1845  			dst = add1(dst)
  1846  			bits >>= 8
  1847  		}
  1848  	} else {
  1849  		totalBits = (uintptr(unsafe.Pointer(dst))-uintptr(unsafe.Pointer(dstStart)))*4 + nbits
  1850  		nbits += -nbits & 3
  1851  		for ; nbits > 0; nbits -= 4 {
  1852  			v := bits&0xf | bitScanAll
  1853  			*dst = uint8(v)
  1854  			dst = add1(dst)
  1855  			bits >>= 4
  1856  		}
  1857  	}
  1858  	return totalBits
  1859  }
  1860  
  1861  // materializeGCProg allocates space for the (1-bit) pointer bitmask
  1862  // for an object of size ptrdata.  Then it fills that space with the
  1863  // pointer bitmask specified by the program prog.
  1864  // The bitmask starts at s.startAddr.
  1865  // The result must be deallocated with dematerializeGCProg.
  1866  func materializeGCProg(ptrdata uintptr, prog *byte) *mspan {
  1867  	// Each word of ptrdata needs one bit in the bitmap.
  1868  	bitmapBytes := divRoundUp(ptrdata, 8*sys.PtrSize)
  1869  	// Compute the number of pages needed for bitmapBytes.
  1870  	pages := divRoundUp(bitmapBytes, pageSize)
  1871  	s := mheap_.allocManual(pages, spanAllocPtrScalarBits)
  1872  	runGCProg(addb(prog, 4), nil, (*byte)(unsafe.Pointer(s.startAddr)), 1)
  1873  	return s
  1874  }
  1875  func dematerializeGCProg(s *mspan) {
  1876  	mheap_.freeManual(s, spanAllocPtrScalarBits)
  1877  }
  1878  
  1879  func dumpGCProg(p *byte) {
  1880  	nptr := 0
  1881  	for {
  1882  		x := *p
  1883  		p = add1(p)
  1884  		if x == 0 {
  1885  			print("\t", nptr, " end\n")
  1886  			break
  1887  		}
  1888  		if x&0x80 == 0 {
  1889  			print("\t", nptr, " lit ", x, ":")
  1890  			n := int(x+7) / 8
  1891  			for i := 0; i < n; i++ {
  1892  				print(" ", hex(*p))
  1893  				p = add1(p)
  1894  			}
  1895  			print("\n")
  1896  			nptr += int(x)
  1897  		} else {
  1898  			nbit := int(x &^ 0x80)
  1899  			if nbit == 0 {
  1900  				for nb := uint(0); ; nb += 7 {
  1901  					x := *p
  1902  					p = add1(p)
  1903  					nbit |= int(x&0x7f) << nb
  1904  					if x&0x80 == 0 {
  1905  						break
  1906  					}
  1907  				}
  1908  			}
  1909  			count := 0
  1910  			for nb := uint(0); ; nb += 7 {
  1911  				x := *p
  1912  				p = add1(p)
  1913  				count |= int(x&0x7f) << nb
  1914  				if x&0x80 == 0 {
  1915  					break
  1916  				}
  1917  			}
  1918  			print("\t", nptr, " repeat ", nbit, " × ", count, "\n")
  1919  			nptr += nbit * count
  1920  		}
  1921  	}
  1922  }
  1923  
  1924  // Testing.
  1925  
  1926  func getgcmaskcb(frame *stkframe, ctxt unsafe.Pointer) bool {
  1927  	target := (*stkframe)(ctxt)
  1928  	if frame.sp <= target.sp && target.sp < frame.varp {
  1929  		*target = *frame
  1930  		return false
  1931  	}
  1932  	return true
  1933  }
  1934  
  1935  // gcbits returns the GC type info for x, for testing.
  1936  // The result is the bitmap entries (0 or 1), one entry per byte.
  1937  //go:linkname reflect_gcbits reflect.gcbits
  1938  func reflect_gcbits(x interface{}) []byte {
  1939  	ret := getgcmask(x)
  1940  	typ := (*ptrtype)(unsafe.Pointer(efaceOf(&x)._type)).elem
  1941  	nptr := typ.ptrdata / sys.PtrSize
  1942  	for uintptr(len(ret)) > nptr && ret[len(ret)-1] == 0 {
  1943  		ret = ret[:len(ret)-1]
  1944  	}
  1945  	return ret
  1946  }
  1947  
  1948  // Returns GC type info for the pointer stored in ep for testing.
  1949  // If ep points to the stack, only static live information will be returned
  1950  // (i.e. not for objects which are only dynamically live stack objects).
  1951  func getgcmask(ep interface{}) (mask []byte) {
  1952  	e := *efaceOf(&ep)
  1953  	p := e.data
  1954  	t := e._type
  1955  	// data or bss
  1956  	for _, datap := range activeModules() {
  1957  		// data
  1958  		if datap.data <= uintptr(p) && uintptr(p) < datap.edata {
  1959  			bitmap := datap.gcdatamask.bytedata
  1960  			n := (*ptrtype)(unsafe.Pointer(t)).elem.size
  1961  			mask = make([]byte, n/sys.PtrSize)
  1962  			for i := uintptr(0); i < n; i += sys.PtrSize {
  1963  				off := (uintptr(p) + i - datap.data) / sys.PtrSize
  1964  				mask[i/sys.PtrSize] = (*addb(bitmap, off/8) >> (off % 8)) & 1
  1965  			}
  1966  			return
  1967  		}
  1968  
  1969  		// bss
  1970  		if datap.bss <= uintptr(p) && uintptr(p) < datap.ebss {
  1971  			bitmap := datap.gcbssmask.bytedata
  1972  			n := (*ptrtype)(unsafe.Pointer(t)).elem.size
  1973  			mask = make([]byte, n/sys.PtrSize)
  1974  			for i := uintptr(0); i < n; i += sys.PtrSize {
  1975  				off := (uintptr(p) + i - datap.bss) / sys.PtrSize
  1976  				mask[i/sys.PtrSize] = (*addb(bitmap, off/8) >> (off % 8)) & 1
  1977  			}
  1978  			return
  1979  		}
  1980  	}
  1981  
  1982  	// heap
  1983  	if base, s, _ := findObject(uintptr(p), 0, 0); base != 0 {
  1984  		hbits := heapBitsForAddr(base)
  1985  		n := s.elemsize
  1986  		mask = make([]byte, n/sys.PtrSize)
  1987  		for i := uintptr(0); i < n; i += sys.PtrSize {
  1988  			if hbits.isPointer() {
  1989  				mask[i/sys.PtrSize] = 1
  1990  			}
  1991  			if !hbits.morePointers() {
  1992  				mask = mask[:i/sys.PtrSize]
  1993  				break
  1994  			}
  1995  			hbits = hbits.next()
  1996  		}
  1997  		return
  1998  	}
  1999  
  2000  	// stack
  2001  	if _g_ := getg(); _g_.m.curg.stack.lo <= uintptr(p) && uintptr(p) < _g_.m.curg.stack.hi {
  2002  		var frame stkframe
  2003  		frame.sp = uintptr(p)
  2004  		_g_ := getg()
  2005  		gentraceback(_g_.m.curg.sched.pc, _g_.m.curg.sched.sp, 0, _g_.m.curg, 0, nil, 1000, getgcmaskcb, noescape(unsafe.Pointer(&frame)), 0)
  2006  		if frame.fn.valid() {
  2007  			locals, _, _ := getStackMap(&frame, nil, false)
  2008  			if locals.n == 0 {
  2009  				return
  2010  			}
  2011  			size := uintptr(locals.n) * sys.PtrSize
  2012  			n := (*ptrtype)(unsafe.Pointer(t)).elem.size
  2013  			mask = make([]byte, n/sys.PtrSize)
  2014  			for i := uintptr(0); i < n; i += sys.PtrSize {
  2015  				off := (uintptr(p) + i - frame.varp + size) / sys.PtrSize
  2016  				mask[i/sys.PtrSize] = locals.ptrbit(off)
  2017  			}
  2018  		}
  2019  		return
  2020  	}
  2021  
  2022  	// otherwise, not something the GC knows about.
  2023  	// possibly read-only data, like malloc(0).
  2024  	// must not have pointers
  2025  	return
  2026  }
  2027  

View as plain text