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  // divideByElemSize returns n/s.elemsize.
   230  // n must be within [0, s.npages*_PageSize),
   231  // or may be exactly s.npages*_PageSize
   232  // if s.elemsize is from sizeclasses.go.
   233  func (s *mspan) divideByElemSize(n uintptr) uintptr {
   234  	const doubleCheck = false
   235  
   236  	// See explanation in mksizeclasses.go's computeDivMagic.
   237  	q := uintptr((uint64(n) * uint64(s.divMul)) >> 32)
   238  
   239  	if doubleCheck && q != n/s.elemsize {
   240  		println(n, "/", s.elemsize, "should be", n/s.elemsize, "but got", q)
   241  		throw("bad magic division")
   242  	}
   243  	return q
   244  }
   245  
   246  func (s *mspan) objIndex(p uintptr) uintptr {
   247  	return s.divideByElemSize(p - s.base())
   248  }
   249  
   250  func markBitsForAddr(p uintptr) markBits {
   251  	s := spanOf(p)
   252  	objIndex := s.objIndex(p)
   253  	return s.markBitsForIndex(objIndex)
   254  }
   255  
   256  func (s *mspan) markBitsForIndex(objIndex uintptr) markBits {
   257  	bytep, mask := s.gcmarkBits.bitp(objIndex)
   258  	return markBits{bytep, mask, objIndex}
   259  }
   260  
   261  func (s *mspan) markBitsForBase() markBits {
   262  	return markBits{(*uint8)(s.gcmarkBits), uint8(1), 0}
   263  }
   264  
   265  // isMarked reports whether mark bit m is set.
   266  func (m markBits) isMarked() bool {
   267  	return *m.bytep&m.mask != 0
   268  }
   269  
   270  // setMarked sets the marked bit in the markbits, atomically.
   271  func (m markBits) setMarked() {
   272  	// Might be racing with other updates, so use atomic update always.
   273  	// We used to be clever here and use a non-atomic update in certain
   274  	// cases, but it's not worth the risk.
   275  	atomic.Or8(m.bytep, m.mask)
   276  }
   277  
   278  // setMarkedNonAtomic sets the marked bit in the markbits, non-atomically.
   279  func (m markBits) setMarkedNonAtomic() {
   280  	*m.bytep |= m.mask
   281  }
   282  
   283  // clearMarked clears the marked bit in the markbits, atomically.
   284  func (m markBits) clearMarked() {
   285  	// Might be racing with other updates, so use atomic update always.
   286  	// We used to be clever here and use a non-atomic update in certain
   287  	// cases, but it's not worth the risk.
   288  	atomic.And8(m.bytep, ^m.mask)
   289  }
   290  
   291  // markBitsForSpan returns the markBits for the span base address base.
   292  func markBitsForSpan(base uintptr) (mbits markBits) {
   293  	mbits = markBitsForAddr(base)
   294  	if mbits.mask != 1 {
   295  		throw("markBitsForSpan: unaligned start")
   296  	}
   297  	return mbits
   298  }
   299  
   300  // advance advances the markBits to the next object in the span.
   301  func (m *markBits) advance() {
   302  	if m.mask == 1<<7 {
   303  		m.bytep = (*uint8)(unsafe.Pointer(uintptr(unsafe.Pointer(m.bytep)) + 1))
   304  		m.mask = 1
   305  	} else {
   306  		m.mask = m.mask << 1
   307  	}
   308  	m.index++
   309  }
   310  
   311  // heapBitsForAddr returns the heapBits for the address addr.
   312  // The caller must ensure addr is in an allocated span.
   313  // In particular, be careful not to point past the end of an object.
   314  //
   315  // nosplit because it is used during write barriers and must not be preempted.
   316  //go:nosplit
   317  func heapBitsForAddr(addr uintptr) (h heapBits) {
   318  	// 2 bits per word, 4 pairs per byte, and a mask is hard coded.
   319  	arena := arenaIndex(addr)
   320  	ha := mheap_.arenas[arena.l1()][arena.l2()]
   321  	// The compiler uses a load for nil checking ha, but in this
   322  	// case we'll almost never hit that cache line again, so it
   323  	// makes more sense to do a value check.
   324  	if ha == nil {
   325  		// addr is not in the heap. Return nil heapBits, which
   326  		// we expect to crash in the caller.
   327  		return
   328  	}
   329  	h.bitp = &ha.bitmap[(addr/(sys.PtrSize*4))%heapArenaBitmapBytes]
   330  	h.shift = uint32((addr / sys.PtrSize) & 3)
   331  	h.arena = uint32(arena)
   332  	h.last = &ha.bitmap[len(ha.bitmap)-1]
   333  	return
   334  }
   335  
   336  // clobberdeadPtr is a special value that is used by the compiler to
   337  // clobber dead stack slots, when -clobberdead flag is set.
   338  const clobberdeadPtr = uintptr(0xdeaddead | 0xdeaddead<<((^uintptr(0)>>63)*32))
   339  
   340  // badPointer throws bad pointer in heap panic.
   341  func badPointer(s *mspan, p, refBase, refOff uintptr) {
   342  	// Typically this indicates an incorrect use
   343  	// of unsafe or cgo to store a bad pointer in
   344  	// the Go heap. It may also indicate a runtime
   345  	// bug.
   346  	//
   347  	// TODO(austin): We could be more aggressive
   348  	// and detect pointers to unallocated objects
   349  	// in allocated spans.
   350  	printlock()
   351  	print("runtime: pointer ", hex(p))
   352  	if s != nil {
   353  		state := s.state.get()
   354  		if state != mSpanInUse {
   355  			print(" to unallocated span")
   356  		} else {
   357  			print(" to unused region of span")
   358  		}
   359  		print(" span.base()=", hex(s.base()), " span.limit=", hex(s.limit), " span.state=", state)
   360  	}
   361  	print("\n")
   362  	if refBase != 0 {
   363  		print("runtime: found in object at *(", hex(refBase), "+", hex(refOff), ")\n")
   364  		gcDumpObject("object", refBase, refOff)
   365  	}
   366  	getg().m.traceback = 2
   367  	throw("found bad pointer in Go heap (incorrect use of unsafe or cgo?)")
   368  }
   369  
   370  // findObject returns the base address for the heap object containing
   371  // the address p, the object's span, and the index of the object in s.
   372  // If p does not point into a heap object, it returns base == 0.
   373  //
   374  // If p points is an invalid heap pointer and debug.invalidptr != 0,
   375  // findObject panics.
   376  //
   377  // refBase and refOff optionally give the base address of the object
   378  // in which the pointer p was found and the byte offset at which it
   379  // was found. These are used for error reporting.
   380  //
   381  // It is nosplit so it is safe for p to be a pointer to the current goroutine's stack.
   382  // Since p is a uintptr, it would not be adjusted if the stack were to move.
   383  //go:nosplit
   384  func findObject(p, refBase, refOff uintptr) (base uintptr, s *mspan, objIndex uintptr) {
   385  	s = spanOf(p)
   386  	// If s is nil, the virtual address has never been part of the heap.
   387  	// This pointer may be to some mmap'd region, so we allow it.
   388  	if s == nil {
   389  		if GOARCH == "amd64" && p == clobberdeadPtr && debug.invalidptr != 0 {
   390  			// Crash if clobberdeadPtr is seen. Only on AMD64 for now, as
   391  			// it is the only platform where compiler's clobberdead mode is
   392  			// implemented. On AMD64 clobberdeadPtr cannot be a valid address.
   393  			badPointer(s, p, refBase, refOff)
   394  		}
   395  		return
   396  	}
   397  	// If p is a bad pointer, it may not be in s's bounds.
   398  	//
   399  	// Check s.state to synchronize with span initialization
   400  	// before checking other fields. See also spanOfHeap.
   401  	if state := s.state.get(); state != mSpanInUse || p < s.base() || p >= s.limit {
   402  		// Pointers into stacks are also ok, the runtime manages these explicitly.
   403  		if state == mSpanManual {
   404  			return
   405  		}
   406  		// The following ensures that we are rigorous about what data
   407  		// structures hold valid pointers.
   408  		if debug.invalidptr != 0 {
   409  			badPointer(s, p, refBase, refOff)
   410  		}
   411  		return
   412  	}
   413  
   414  	objIndex = s.objIndex(p)
   415  	base = s.base() + objIndex*s.elemsize
   416  	return
   417  }
   418  
   419  // next returns the heapBits describing the next pointer-sized word in memory.
   420  // That is, if h describes address p, h.next() describes p+ptrSize.
   421  // Note that next does not modify h. The caller must record the result.
   422  //
   423  // nosplit because it is used during write barriers and must not be preempted.
   424  //go:nosplit
   425  func (h heapBits) next() heapBits {
   426  	if h.shift < 3*heapBitsShift {
   427  		h.shift += heapBitsShift
   428  	} else if h.bitp != h.last {
   429  		h.bitp, h.shift = add1(h.bitp), 0
   430  	} else {
   431  		// Move to the next arena.
   432  		return h.nextArena()
   433  	}
   434  	return h
   435  }
   436  
   437  // nextArena advances h to the beginning of the next heap arena.
   438  //
   439  // This is a slow-path helper to next. gc's inliner knows that
   440  // heapBits.next can be inlined even though it calls this. This is
   441  // marked noinline so it doesn't get inlined into next and cause next
   442  // to be too big to inline.
   443  //
   444  //go:nosplit
   445  //go:noinline
   446  func (h heapBits) nextArena() heapBits {
   447  	h.arena++
   448  	ai := arenaIdx(h.arena)
   449  	l2 := mheap_.arenas[ai.l1()]
   450  	if l2 == nil {
   451  		// We just passed the end of the object, which
   452  		// was also the end of the heap. Poison h. It
   453  		// should never be dereferenced at this point.
   454  		return heapBits{}
   455  	}
   456  	ha := l2[ai.l2()]
   457  	if ha == nil {
   458  		return heapBits{}
   459  	}
   460  	h.bitp, h.shift = &ha.bitmap[0], 0
   461  	h.last = &ha.bitmap[len(ha.bitmap)-1]
   462  	return h
   463  }
   464  
   465  // forward returns the heapBits describing n pointer-sized words ahead of h in memory.
   466  // That is, if h describes address p, h.forward(n) describes p+n*ptrSize.
   467  // h.forward(1) is equivalent to h.next(), just slower.
   468  // Note that forward does not modify h. The caller must record the result.
   469  // bits returns the heap bits for the current word.
   470  //go:nosplit
   471  func (h heapBits) forward(n uintptr) heapBits {
   472  	n += uintptr(h.shift) / heapBitsShift
   473  	nbitp := uintptr(unsafe.Pointer(h.bitp)) + n/4
   474  	h.shift = uint32(n%4) * heapBitsShift
   475  	if nbitp <= uintptr(unsafe.Pointer(h.last)) {
   476  		h.bitp = (*uint8)(unsafe.Pointer(nbitp))
   477  		return h
   478  	}
   479  
   480  	// We're in a new heap arena.
   481  	past := nbitp - (uintptr(unsafe.Pointer(h.last)) + 1)
   482  	h.arena += 1 + uint32(past/heapArenaBitmapBytes)
   483  	ai := arenaIdx(h.arena)
   484  	if l2 := mheap_.arenas[ai.l1()]; l2 != nil && l2[ai.l2()] != nil {
   485  		a := l2[ai.l2()]
   486  		h.bitp = &a.bitmap[past%heapArenaBitmapBytes]
   487  		h.last = &a.bitmap[len(a.bitmap)-1]
   488  	} else {
   489  		h.bitp, h.last = nil, nil
   490  	}
   491  	return h
   492  }
   493  
   494  // forwardOrBoundary is like forward, but stops at boundaries between
   495  // contiguous sections of the bitmap. It returns the number of words
   496  // advanced over, which will be <= n.
   497  func (h heapBits) forwardOrBoundary(n uintptr) (heapBits, uintptr) {
   498  	maxn := 4 * ((uintptr(unsafe.Pointer(h.last)) + 1) - uintptr(unsafe.Pointer(h.bitp)))
   499  	if n > maxn {
   500  		n = maxn
   501  	}
   502  	return h.forward(n), n
   503  }
   504  
   505  // The caller can test morePointers and isPointer by &-ing with bitScan and bitPointer.
   506  // The result includes in its higher bits the bits for subsequent words
   507  // described by the same bitmap byte.
   508  //
   509  // nosplit because it is used during write barriers and must not be preempted.
   510  //go:nosplit
   511  func (h heapBits) bits() uint32 {
   512  	// The (shift & 31) eliminates a test and conditional branch
   513  	// from the generated code.
   514  	return uint32(*h.bitp) >> (h.shift & 31)
   515  }
   516  
   517  // morePointers reports whether this word and all remaining words in this object
   518  // are scalars.
   519  // h must not describe the second word of the object.
   520  func (h heapBits) morePointers() bool {
   521  	return h.bits()&bitScan != 0
   522  }
   523  
   524  // isPointer reports whether the heap bits describe a pointer word.
   525  //
   526  // nosplit because it is used during write barriers and must not be preempted.
   527  //go:nosplit
   528  func (h heapBits) isPointer() bool {
   529  	return h.bits()&bitPointer != 0
   530  }
   531  
   532  // bulkBarrierPreWrite executes a write barrier
   533  // for every pointer slot in the memory range [src, src+size),
   534  // using pointer/scalar information from [dst, dst+size).
   535  // This executes the write barriers necessary before a memmove.
   536  // src, dst, and size must be pointer-aligned.
   537  // The range [dst, dst+size) must lie within a single object.
   538  // It does not perform the actual writes.
   539  //
   540  // As a special case, src == 0 indicates that this is being used for a
   541  // memclr. bulkBarrierPreWrite will pass 0 for the src of each write
   542  // barrier.
   543  //
   544  // Callers should call bulkBarrierPreWrite immediately before
   545  // calling memmove(dst, src, size). This function is marked nosplit
   546  // to avoid being preempted; the GC must not stop the goroutine
   547  // between the memmove and the execution of the barriers.
   548  // The caller is also responsible for cgo pointer checks if this
   549  // may be writing Go pointers into non-Go memory.
   550  //
   551  // The pointer bitmap is not maintained for allocations containing
   552  // no pointers at all; any caller of bulkBarrierPreWrite must first
   553  // make sure the underlying allocation contains pointers, usually
   554  // by checking typ.ptrdata.
   555  //
   556  // Callers must perform cgo checks if writeBarrier.cgo.
   557  //
   558  //go:nosplit
   559  func bulkBarrierPreWrite(dst, src, size uintptr) {
   560  	if (dst|src|size)&(sys.PtrSize-1) != 0 {
   561  		throw("bulkBarrierPreWrite: unaligned arguments")
   562  	}
   563  	if !writeBarrier.needed {
   564  		return
   565  	}
   566  	if s := spanOf(dst); s == nil {
   567  		// If dst is a global, use the data or BSS bitmaps to
   568  		// execute write barriers.
   569  		for _, datap := range activeModules() {
   570  			if datap.data <= dst && dst < datap.edata {
   571  				bulkBarrierBitmap(dst, src, size, dst-datap.data, datap.gcdatamask.bytedata)
   572  				return
   573  			}
   574  		}
   575  		for _, datap := range activeModules() {
   576  			if datap.bss <= dst && dst < datap.ebss {
   577  				bulkBarrierBitmap(dst, src, size, dst-datap.bss, datap.gcbssmask.bytedata)
   578  				return
   579  			}
   580  		}
   581  		return
   582  	} else if s.state.get() != mSpanInUse || dst < s.base() || s.limit <= dst {
   583  		// dst was heap memory at some point, but isn't now.
   584  		// It can't be a global. It must be either our stack,
   585  		// or in the case of direct channel sends, it could be
   586  		// another stack. Either way, no need for barriers.
   587  		// This will also catch if dst is in a freed span,
   588  		// though that should never have.
   589  		return
   590  	}
   591  
   592  	buf := &getg().m.p.ptr().wbBuf
   593  	h := heapBitsForAddr(dst)
   594  	if src == 0 {
   595  		for i := uintptr(0); i < size; i += sys.PtrSize {
   596  			if h.isPointer() {
   597  				dstx := (*uintptr)(unsafe.Pointer(dst + i))
   598  				if !buf.putFast(*dstx, 0) {
   599  					wbBufFlush(nil, 0)
   600  				}
   601  			}
   602  			h = h.next()
   603  		}
   604  	} else {
   605  		for i := uintptr(0); i < size; i += sys.PtrSize {
   606  			if h.isPointer() {
   607  				dstx := (*uintptr)(unsafe.Pointer(dst + i))
   608  				srcx := (*uintptr)(unsafe.Pointer(src + i))
   609  				if !buf.putFast(*dstx, *srcx) {
   610  					wbBufFlush(nil, 0)
   611  				}
   612  			}
   613  			h = h.next()
   614  		}
   615  	}
   616  }
   617  
   618  // bulkBarrierPreWriteSrcOnly is like bulkBarrierPreWrite but
   619  // does not execute write barriers for [dst, dst+size).
   620  //
   621  // In addition to the requirements of bulkBarrierPreWrite
   622  // callers need to ensure [dst, dst+size) is zeroed.
   623  //
   624  // This is used for special cases where e.g. dst was just
   625  // created and zeroed with malloc.
   626  //go:nosplit
   627  func bulkBarrierPreWriteSrcOnly(dst, src, size uintptr) {
   628  	if (dst|src|size)&(sys.PtrSize-1) != 0 {
   629  		throw("bulkBarrierPreWrite: unaligned arguments")
   630  	}
   631  	if !writeBarrier.needed {
   632  		return
   633  	}
   634  	buf := &getg().m.p.ptr().wbBuf
   635  	h := heapBitsForAddr(dst)
   636  	for i := uintptr(0); i < size; i += sys.PtrSize {
   637  		if h.isPointer() {
   638  			srcx := (*uintptr)(unsafe.Pointer(src + i))
   639  			if !buf.putFast(0, *srcx) {
   640  				wbBufFlush(nil, 0)
   641  			}
   642  		}
   643  		h = h.next()
   644  	}
   645  }
   646  
   647  // bulkBarrierBitmap executes write barriers for copying from [src,
   648  // src+size) to [dst, dst+size) using a 1-bit pointer bitmap. src is
   649  // assumed to start maskOffset bytes into the data covered by the
   650  // bitmap in bits (which may not be a multiple of 8).
   651  //
   652  // This is used by bulkBarrierPreWrite for writes to data and BSS.
   653  //
   654  //go:nosplit
   655  func bulkBarrierBitmap(dst, src, size, maskOffset uintptr, bits *uint8) {
   656  	word := maskOffset / sys.PtrSize
   657  	bits = addb(bits, word/8)
   658  	mask := uint8(1) << (word % 8)
   659  
   660  	buf := &getg().m.p.ptr().wbBuf
   661  	for i := uintptr(0); i < size; i += sys.PtrSize {
   662  		if mask == 0 {
   663  			bits = addb(bits, 1)
   664  			if *bits == 0 {
   665  				// Skip 8 words.
   666  				i += 7 * sys.PtrSize
   667  				continue
   668  			}
   669  			mask = 1
   670  		}
   671  		if *bits&mask != 0 {
   672  			dstx := (*uintptr)(unsafe.Pointer(dst + i))
   673  			if src == 0 {
   674  				if !buf.putFast(*dstx, 0) {
   675  					wbBufFlush(nil, 0)
   676  				}
   677  			} else {
   678  				srcx := (*uintptr)(unsafe.Pointer(src + i))
   679  				if !buf.putFast(*dstx, *srcx) {
   680  					wbBufFlush(nil, 0)
   681  				}
   682  			}
   683  		}
   684  		mask <<= 1
   685  	}
   686  }
   687  
   688  // typeBitsBulkBarrier executes a write barrier for every
   689  // pointer that would be copied from [src, src+size) to [dst,
   690  // dst+size) by a memmove using the type bitmap to locate those
   691  // pointer slots.
   692  //
   693  // The type typ must correspond exactly to [src, src+size) and [dst, dst+size).
   694  // dst, src, and size must be pointer-aligned.
   695  // The type typ must have a plain bitmap, not a GC program.
   696  // The only use of this function is in channel sends, and the
   697  // 64 kB channel element limit takes care of this for us.
   698  //
   699  // Must not be preempted because it typically runs right before memmove,
   700  // and the GC must observe them as an atomic action.
   701  //
   702  // Callers must perform cgo checks if writeBarrier.cgo.
   703  //
   704  //go:nosplit
   705  func typeBitsBulkBarrier(typ *_type, dst, src, size uintptr) {
   706  	if typ == nil {
   707  		throw("runtime: typeBitsBulkBarrier without type")
   708  	}
   709  	if typ.size != size {
   710  		println("runtime: typeBitsBulkBarrier with type ", typ.string(), " of size ", typ.size, " but memory size", size)
   711  		throw("runtime: invalid typeBitsBulkBarrier")
   712  	}
   713  	if typ.kind&kindGCProg != 0 {
   714  		println("runtime: typeBitsBulkBarrier with type ", typ.string(), " with GC prog")
   715  		throw("runtime: invalid typeBitsBulkBarrier")
   716  	}
   717  	if !writeBarrier.needed {
   718  		return
   719  	}
   720  	ptrmask := typ.gcdata
   721  	buf := &getg().m.p.ptr().wbBuf
   722  	var bits uint32
   723  	for i := uintptr(0); i < typ.ptrdata; i += sys.PtrSize {
   724  		if i&(sys.PtrSize*8-1) == 0 {
   725  			bits = uint32(*ptrmask)
   726  			ptrmask = addb(ptrmask, 1)
   727  		} else {
   728  			bits = bits >> 1
   729  		}
   730  		if bits&1 != 0 {
   731  			dstx := (*uintptr)(unsafe.Pointer(dst + i))
   732  			srcx := (*uintptr)(unsafe.Pointer(src + i))
   733  			if !buf.putFast(*dstx, *srcx) {
   734  				wbBufFlush(nil, 0)
   735  			}
   736  		}
   737  	}
   738  }
   739  
   740  // The methods operating on spans all require that h has been returned
   741  // by heapBitsForSpan and that size, n, total are the span layout description
   742  // returned by the mspan's layout method.
   743  // If total > size*n, it means that there is extra leftover memory in the span,
   744  // usually due to rounding.
   745  //
   746  // TODO(rsc): Perhaps introduce a different heapBitsSpan type.
   747  
   748  // initSpan initializes the heap bitmap for a span.
   749  // If this is a span of pointer-sized objects, it initializes all
   750  // words to pointer/scan.
   751  // Otherwise, it initializes all words to scalar/dead.
   752  func (h heapBits) initSpan(s *mspan) {
   753  	// Clear bits corresponding to objects.
   754  	nw := (s.npages << _PageShift) / sys.PtrSize
   755  	if nw%wordsPerBitmapByte != 0 {
   756  		throw("initSpan: unaligned length")
   757  	}
   758  	if h.shift != 0 {
   759  		throw("initSpan: unaligned base")
   760  	}
   761  	isPtrs := sys.PtrSize == 8 && s.elemsize == sys.PtrSize
   762  	for nw > 0 {
   763  		hNext, anw := h.forwardOrBoundary(nw)
   764  		nbyte := anw / wordsPerBitmapByte
   765  		if isPtrs {
   766  			bitp := h.bitp
   767  			for i := uintptr(0); i < nbyte; i++ {
   768  				*bitp = bitPointerAll | bitScanAll
   769  				bitp = add1(bitp)
   770  			}
   771  		} else {
   772  			memclrNoHeapPointers(unsafe.Pointer(h.bitp), nbyte)
   773  		}
   774  		h = hNext
   775  		nw -= anw
   776  	}
   777  }
   778  
   779  // countAlloc returns the number of objects allocated in span s by
   780  // scanning the allocation bitmap.
   781  func (s *mspan) countAlloc() int {
   782  	count := 0
   783  	bytes := divRoundUp(s.nelems, 8)
   784  	// Iterate over each 8-byte chunk and count allocations
   785  	// with an intrinsic. Note that newMarkBits guarantees that
   786  	// gcmarkBits will be 8-byte aligned, so we don't have to
   787  	// worry about edge cases, irrelevant bits will simply be zero.
   788  	for i := uintptr(0); i < bytes; i += 8 {
   789  		// Extract 64 bits from the byte pointer and get a OnesCount.
   790  		// Note that the unsafe cast here doesn't preserve endianness,
   791  		// but that's OK. We only care about how many bits are 1, not
   792  		// about the order we discover them in.
   793  		mrkBits := *(*uint64)(unsafe.Pointer(s.gcmarkBits.bytep(i)))
   794  		count += sys.OnesCount64(mrkBits)
   795  	}
   796  	return count
   797  }
   798  
   799  // heapBitsSetType records that the new allocation [x, x+size)
   800  // holds in [x, x+dataSize) one or more values of type typ.
   801  // (The number of values is given by dataSize / typ.size.)
   802  // If dataSize < size, the fragment [x+dataSize, x+size) is
   803  // recorded as non-pointer data.
   804  // It is known that the type has pointers somewhere;
   805  // malloc does not call heapBitsSetType when there are no pointers,
   806  // because all free objects are marked as noscan during
   807  // heapBitsSweepSpan.
   808  //
   809  // There can only be one allocation from a given span active at a time,
   810  // and the bitmap for a span always falls on byte boundaries,
   811  // so there are no write-write races for access to the heap bitmap.
   812  // Hence, heapBitsSetType can access the bitmap without atomics.
   813  //
   814  // There can be read-write races between heapBitsSetType and things
   815  // that read the heap bitmap like scanobject. However, since
   816  // heapBitsSetType is only used for objects that have not yet been
   817  // made reachable, readers will ignore bits being modified by this
   818  // function. This does mean this function cannot transiently modify
   819  // bits that belong to neighboring objects. Also, on weakly-ordered
   820  // machines, callers must execute a store/store (publication) barrier
   821  // between calling this function and making the object reachable.
   822  func heapBitsSetType(x, size, dataSize uintptr, typ *_type) {
   823  	const doubleCheck = false // slow but helpful; enable to test modifications to this code
   824  
   825  	const (
   826  		mask1 = bitPointer | bitScan                        // 00010001
   827  		mask2 = bitPointer | bitScan | mask1<<heapBitsShift // 00110011
   828  		mask3 = bitPointer | bitScan | mask2<<heapBitsShift // 01110111
   829  	)
   830  
   831  	// dataSize is always size rounded up to the next malloc size class,
   832  	// except in the case of allocating a defer block, in which case
   833  	// size is sizeof(_defer{}) (at least 6 words) and dataSize may be
   834  	// arbitrarily larger.
   835  	//
   836  	// The checks for size == sys.PtrSize and size == 2*sys.PtrSize can therefore
   837  	// assume that dataSize == size without checking it explicitly.
   838  
   839  	if sys.PtrSize == 8 && size == sys.PtrSize {
   840  		// It's one word and it has pointers, it must be a pointer.
   841  		// Since all allocated one-word objects are pointers
   842  		// (non-pointers are aggregated into tinySize allocations),
   843  		// initSpan sets the pointer bits for us. Nothing to do here.
   844  		if doubleCheck {
   845  			h := heapBitsForAddr(x)
   846  			if !h.isPointer() {
   847  				throw("heapBitsSetType: pointer bit missing")
   848  			}
   849  			if !h.morePointers() {
   850  				throw("heapBitsSetType: scan bit missing")
   851  			}
   852  		}
   853  		return
   854  	}
   855  
   856  	h := heapBitsForAddr(x)
   857  	ptrmask := typ.gcdata // start of 1-bit pointer mask (or GC program, handled below)
   858  
   859  	// 2-word objects only have 4 bitmap bits and 3-word objects only have 6 bitmap bits.
   860  	// Therefore, these objects share a heap bitmap byte with the objects next to them.
   861  	// These are called out as a special case primarily so the code below can assume all
   862  	// objects are at least 4 words long and that their bitmaps start either at the beginning
   863  	// of a bitmap byte, or half-way in (h.shift of 0 and 2 respectively).
   864  
   865  	if size == 2*sys.PtrSize {
   866  		if typ.size == sys.PtrSize {
   867  			// We're allocating a block big enough to hold two pointers.
   868  			// On 64-bit, that means the actual object must be two pointers,
   869  			// or else we'd have used the one-pointer-sized block.
   870  			// On 32-bit, however, this is the 8-byte block, the smallest one.
   871  			// So it could be that we're allocating one pointer and this was
   872  			// just the smallest block available. Distinguish by checking dataSize.
   873  			// (In general the number of instances of typ being allocated is
   874  			// dataSize/typ.size.)
   875  			if sys.PtrSize == 4 && dataSize == sys.PtrSize {
   876  				// 1 pointer object. On 32-bit machines clear the bit for the
   877  				// unused second word.
   878  				*h.bitp &^= (bitPointer | bitScan | (bitPointer|bitScan)<<heapBitsShift) << h.shift
   879  				*h.bitp |= (bitPointer | bitScan) << h.shift
   880  			} else {
   881  				// 2-element array of pointer.
   882  				*h.bitp |= (bitPointer | bitScan | (bitPointer|bitScan)<<heapBitsShift) << h.shift
   883  			}
   884  			return
   885  		}
   886  		// Otherwise typ.size must be 2*sys.PtrSize,
   887  		// and typ.kind&kindGCProg == 0.
   888  		if doubleCheck {
   889  			if typ.size != 2*sys.PtrSize || typ.kind&kindGCProg != 0 {
   890  				print("runtime: heapBitsSetType size=", size, " but typ.size=", typ.size, " gcprog=", typ.kind&kindGCProg != 0, "\n")
   891  				throw("heapBitsSetType")
   892  			}
   893  		}
   894  		b := uint32(*ptrmask)
   895  		hb := b & 3
   896  		hb |= bitScanAll & ((bitScan << (typ.ptrdata / sys.PtrSize)) - 1)
   897  		// Clear the bits for this object so we can set the
   898  		// appropriate ones.
   899  		*h.bitp &^= (bitPointer | bitScan | ((bitPointer | bitScan) << heapBitsShift)) << h.shift
   900  		*h.bitp |= uint8(hb << h.shift)
   901  		return
   902  	} else if size == 3*sys.PtrSize {
   903  		b := uint8(*ptrmask)
   904  		if doubleCheck {
   905  			if b == 0 {
   906  				println("runtime: invalid type ", typ.string())
   907  				throw("heapBitsSetType: called with non-pointer type")
   908  			}
   909  			if sys.PtrSize != 8 {
   910  				throw("heapBitsSetType: unexpected 3 pointer wide size class on 32 bit")
   911  			}
   912  			if typ.kind&kindGCProg != 0 {
   913  				throw("heapBitsSetType: unexpected GC prog for 3 pointer wide size class")
   914  			}
   915  			if typ.size == 2*sys.PtrSize {
   916  				print("runtime: heapBitsSetType size=", size, " but typ.size=", typ.size, "\n")
   917  				throw("heapBitsSetType: inconsistent object sizes")
   918  			}
   919  		}
   920  		if typ.size == sys.PtrSize {
   921  			// The type contains a pointer otherwise heapBitsSetType wouldn't have been called.
   922  			// Since the type is only 1 pointer wide and contains a pointer, its gcdata must be exactly 1.
   923  			if doubleCheck && *typ.gcdata != 1 {
   924  				print("runtime: heapBitsSetType size=", size, " typ.size=", typ.size, "but *typ.gcdata", *typ.gcdata, "\n")
   925  				throw("heapBitsSetType: unexpected gcdata for 1 pointer wide type size in 3 pointer wide size class")
   926  			}
   927  			// 3 element array of pointers. Unrolling ptrmask 3 times into p yields 00000111.
   928  			b = 7
   929  		}
   930  
   931  		hb := b & 7
   932  		// Set bitScan bits for all pointers.
   933  		hb |= hb << wordsPerBitmapByte
   934  		// First bitScan bit is always set since the type contains pointers.
   935  		hb |= bitScan
   936  		// Second bitScan bit needs to also be set if the third bitScan bit is set.
   937  		hb |= hb & (bitScan << (2 * heapBitsShift)) >> 1
   938  
   939  		// For h.shift > 1 heap bits cross a byte boundary and need to be written part
   940  		// to h.bitp and part to the next h.bitp.
   941  		switch h.shift {
   942  		case 0:
   943  			*h.bitp &^= mask3 << 0
   944  			*h.bitp |= hb << 0
   945  		case 1:
   946  			*h.bitp &^= mask3 << 1
   947  			*h.bitp |= hb << 1
   948  		case 2:
   949  			*h.bitp &^= mask2 << 2
   950  			*h.bitp |= (hb & mask2) << 2
   951  			// Two words written to the first byte.
   952  			// Advance two words to get to the next byte.
   953  			h = h.next().next()
   954  			*h.bitp &^= mask1
   955  			*h.bitp |= (hb >> 2) & mask1
   956  		case 3:
   957  			*h.bitp &^= mask1 << 3
   958  			*h.bitp |= (hb & mask1) << 3
   959  			// One word written to the first byte.
   960  			// Advance one word to get to the next byte.
   961  			h = h.next()
   962  			*h.bitp &^= mask2
   963  			*h.bitp |= (hb >> 1) & mask2
   964  		}
   965  		return
   966  	}
   967  
   968  	// Copy from 1-bit ptrmask into 2-bit bitmap.
   969  	// The basic approach is to use a single uintptr as a bit buffer,
   970  	// alternating between reloading the buffer and writing bitmap bytes.
   971  	// In general, one load can supply two bitmap byte writes.
   972  	// This is a lot of lines of code, but it compiles into relatively few
   973  	// machine instructions.
   974  
   975  	outOfPlace := false
   976  	if arenaIndex(x+size-1) != arenaIdx(h.arena) || (doubleCheck && fastrand()%2 == 0) {
   977  		// This object spans heap arenas, so the bitmap may be
   978  		// discontiguous. Unroll it into the object instead
   979  		// and then copy it out.
   980  		//
   981  		// In doubleCheck mode, we randomly do this anyway to
   982  		// stress test the bitmap copying path.
   983  		outOfPlace = true
   984  		h.bitp = (*uint8)(unsafe.Pointer(x))
   985  		h.last = nil
   986  	}
   987  
   988  	var (
   989  		// Ptrmask input.
   990  		p     *byte   // last ptrmask byte read
   991  		b     uintptr // ptrmask bits already loaded
   992  		nb    uintptr // number of bits in b at next read
   993  		endp  *byte   // final ptrmask byte to read (then repeat)
   994  		endnb uintptr // number of valid bits in *endp
   995  		pbits uintptr // alternate source of bits
   996  
   997  		// Heap bitmap output.
   998  		w     uintptr // words processed
   999  		nw    uintptr // number of words to process
  1000  		hbitp *byte   // next heap bitmap byte to write
  1001  		hb    uintptr // bits being prepared for *hbitp
  1002  	)
  1003  
  1004  	hbitp = h.bitp
  1005  
  1006  	// Handle GC program. Delayed until this part of the code
  1007  	// so that we can use the same double-checking mechanism
  1008  	// as the 1-bit case. Nothing above could have encountered
  1009  	// GC programs: the cases were all too small.
  1010  	if typ.kind&kindGCProg != 0 {
  1011  		heapBitsSetTypeGCProg(h, typ.ptrdata, typ.size, dataSize, size, addb(typ.gcdata, 4))
  1012  		if doubleCheck {
  1013  			// Double-check the heap bits written by GC program
  1014  			// by running the GC program to create a 1-bit pointer mask
  1015  			// and then jumping to the double-check code below.
  1016  			// This doesn't catch bugs shared between the 1-bit and 4-bit
  1017  			// GC program execution, but it does catch mistakes specific
  1018  			// to just one of those and bugs in heapBitsSetTypeGCProg's
  1019  			// implementation of arrays.
  1020  			lock(&debugPtrmask.lock)
  1021  			if debugPtrmask.data == nil {
  1022  				debugPtrmask.data = (*byte)(persistentalloc(1<<20, 1, &memstats.other_sys))
  1023  			}
  1024  			ptrmask = debugPtrmask.data
  1025  			runGCProg(addb(typ.gcdata, 4), nil, ptrmask, 1)
  1026  		}
  1027  		goto Phase4
  1028  	}
  1029  
  1030  	// Note about sizes:
  1031  	//
  1032  	// typ.size is the number of words in the object,
  1033  	// and typ.ptrdata is the number of words in the prefix
  1034  	// of the object that contains pointers. That is, the final
  1035  	// typ.size - typ.ptrdata words contain no pointers.
  1036  	// This allows optimization of a common pattern where
  1037  	// an object has a small header followed by a large scalar
  1038  	// buffer. If we know the pointers are over, we don't have
  1039  	// to scan the buffer's heap bitmap at all.
  1040  	// The 1-bit ptrmasks are sized to contain only bits for
  1041  	// the typ.ptrdata prefix, zero padded out to a full byte
  1042  	// of bitmap. This code sets nw (below) so that heap bitmap
  1043  	// bits are only written for the typ.ptrdata prefix; if there is
  1044  	// more room in the allocated object, the next heap bitmap
  1045  	// entry is a 00, indicating that there are no more pointers
  1046  	// to scan. So only the ptrmask for the ptrdata bytes is needed.
  1047  	//
  1048  	// Replicated copies are not as nice: if there is an array of
  1049  	// objects with scalar tails, all but the last tail does have to
  1050  	// be initialized, because there is no way to say "skip forward".
  1051  	// However, because of the possibility of a repeated type with
  1052  	// size not a multiple of 4 pointers (one heap bitmap byte),
  1053  	// the code already must handle the last ptrmask byte specially
  1054  	// by treating it as containing only the bits for endnb pointers,
  1055  	// where endnb <= 4. We represent large scalar tails that must
  1056  	// be expanded in the replication by setting endnb larger than 4.
  1057  	// This will have the effect of reading many bits out of b,
  1058  	// but once the real bits are shifted out, b will supply as many
  1059  	// zero bits as we try to read, which is exactly what we need.
  1060  
  1061  	p = ptrmask
  1062  	if typ.size < dataSize {
  1063  		// Filling in bits for an array of typ.
  1064  		// Set up for repetition of ptrmask during main loop.
  1065  		// Note that ptrmask describes only a prefix of
  1066  		const maxBits = sys.PtrSize*8 - 7
  1067  		if typ.ptrdata/sys.PtrSize <= maxBits {
  1068  			// Entire ptrmask fits in uintptr with room for a byte fragment.
  1069  			// Load into pbits and never read from ptrmask again.
  1070  			// This is especially important when the ptrmask has
  1071  			// fewer than 8 bits in it; otherwise the reload in the middle
  1072  			// of the Phase 2 loop would itself need to loop to gather
  1073  			// at least 8 bits.
  1074  
  1075  			// Accumulate ptrmask into b.
  1076  			// ptrmask is sized to describe only typ.ptrdata, but we record
  1077  			// it as describing typ.size bytes, since all the high bits are zero.
  1078  			nb = typ.ptrdata / sys.PtrSize
  1079  			for i := uintptr(0); i < nb; i += 8 {
  1080  				b |= uintptr(*p) << i
  1081  				p = add1(p)
  1082  			}
  1083  			nb = typ.size / sys.PtrSize
  1084  
  1085  			// Replicate ptrmask to fill entire pbits uintptr.
  1086  			// Doubling and truncating is fewer steps than
  1087  			// iterating by nb each time. (nb could be 1.)
  1088  			// Since we loaded typ.ptrdata/sys.PtrSize bits
  1089  			// but are pretending to have typ.size/sys.PtrSize,
  1090  			// there might be no replication necessary/possible.
  1091  			pbits = b
  1092  			endnb = nb
  1093  			if nb+nb <= maxBits {
  1094  				for endnb <= sys.PtrSize*8 {
  1095  					pbits |= pbits << endnb
  1096  					endnb += endnb
  1097  				}
  1098  				// Truncate to a multiple of original ptrmask.
  1099  				// Because nb+nb <= maxBits, nb fits in a byte.
  1100  				// Byte division is cheaper than uintptr division.
  1101  				endnb = uintptr(maxBits/byte(nb)) * nb
  1102  				pbits &= 1<<endnb - 1
  1103  				b = pbits
  1104  				nb = endnb
  1105  			}
  1106  
  1107  			// Clear p and endp as sentinel for using pbits.
  1108  			// Checked during Phase 2 loop.
  1109  			p = nil
  1110  			endp = nil
  1111  		} else {
  1112  			// Ptrmask is larger. Read it multiple times.
  1113  			n := (typ.ptrdata/sys.PtrSize+7)/8 - 1
  1114  			endp = addb(ptrmask, n)
  1115  			endnb = typ.size/sys.PtrSize - n*8
  1116  		}
  1117  	}
  1118  	if p != nil {
  1119  		b = uintptr(*p)
  1120  		p = add1(p)
  1121  		nb = 8
  1122  	}
  1123  
  1124  	if typ.size == dataSize {
  1125  		// Single entry: can stop once we reach the non-pointer data.
  1126  		nw = typ.ptrdata / sys.PtrSize
  1127  	} else {
  1128  		// Repeated instances of typ in an array.
  1129  		// Have to process first N-1 entries in full, but can stop
  1130  		// once we reach the non-pointer data in the final entry.
  1131  		nw = ((dataSize/typ.size-1)*typ.size + typ.ptrdata) / sys.PtrSize
  1132  	}
  1133  	if nw == 0 {
  1134  		// No pointers! Caller was supposed to check.
  1135  		println("runtime: invalid type ", typ.string())
  1136  		throw("heapBitsSetType: called with non-pointer type")
  1137  		return
  1138  	}
  1139  
  1140  	// Phase 1: Special case for leading byte (shift==0) or half-byte (shift==2).
  1141  	// The leading byte is special because it contains the bits for word 1,
  1142  	// which does not have the scan bit set.
  1143  	// The leading half-byte is special because it's a half a byte,
  1144  	// so we have to be careful with the bits already there.
  1145  	switch {
  1146  	default:
  1147  		throw("heapBitsSetType: unexpected shift")
  1148  
  1149  	case h.shift == 0:
  1150  		// Ptrmask and heap bitmap are aligned.
  1151  		//
  1152  		// This is a fast path for small objects.
  1153  		//
  1154  		// The first byte we write out covers the first four
  1155  		// words of the object. The scan/dead bit on the first
  1156  		// word must be set to scan since there are pointers
  1157  		// somewhere in the object.
  1158  		// In all following words, we set the scan/dead
  1159  		// appropriately to indicate that the object continues
  1160  		// to the next 2-bit entry in the bitmap.
  1161  		//
  1162  		// We set four bits at a time here, but if the object
  1163  		// is fewer than four words, phase 3 will clear
  1164  		// unnecessary bits.
  1165  		hb = b & bitPointerAll
  1166  		hb |= bitScanAll
  1167  		if w += 4; w >= nw {
  1168  			goto Phase3
  1169  		}
  1170  		*hbitp = uint8(hb)
  1171  		hbitp = add1(hbitp)
  1172  		b >>= 4
  1173  		nb -= 4
  1174  
  1175  	case h.shift == 2:
  1176  		// Ptrmask and heap bitmap are misaligned.
  1177  		//
  1178  		// On 32 bit architectures only the 6-word object that corresponds
  1179  		// to a 24 bytes size class can start with h.shift of 2 here since
  1180  		// all other non 16 byte aligned size classes have been handled by
  1181  		// special code paths at the beginning of heapBitsSetType on 32 bit.
  1182  		//
  1183  		// Many size classes are only 16 byte aligned. On 64 bit architectures
  1184  		// this results in a heap bitmap position starting with a h.shift of 2.
  1185  		//
  1186  		// The bits for the first two words are in a byte shared
  1187  		// with another object, so we must be careful with the bits
  1188  		// already there.
  1189  		//
  1190  		// We took care of 1-word, 2-word, and 3-word objects above,
  1191  		// so this is at least a 6-word object.
  1192  		hb = (b & (bitPointer | bitPointer<<heapBitsShift)) << (2 * heapBitsShift)
  1193  		hb |= bitScan << (2 * heapBitsShift)
  1194  		if nw > 1 {
  1195  			hb |= bitScan << (3 * heapBitsShift)
  1196  		}
  1197  		b >>= 2
  1198  		nb -= 2
  1199  		*hbitp &^= uint8((bitPointer | bitScan | ((bitPointer | bitScan) << heapBitsShift)) << (2 * heapBitsShift))
  1200  		*hbitp |= uint8(hb)
  1201  		hbitp = add1(hbitp)
  1202  		if w += 2; w >= nw {
  1203  			// We know that there is more data, because we handled 2-word and 3-word objects above.
  1204  			// This must be at least a 6-word object. If we're out of pointer words,
  1205  			// mark no scan in next bitmap byte and finish.
  1206  			hb = 0
  1207  			w += 4
  1208  			goto Phase3
  1209  		}
  1210  	}
  1211  
  1212  	// Phase 2: Full bytes in bitmap, up to but not including write to last byte (full or partial) in bitmap.
  1213  	// The loop computes the bits for that last write but does not execute the write;
  1214  	// it leaves the bits in hb for processing by phase 3.
  1215  	// To avoid repeated adjustment of nb, we subtract out the 4 bits we're going to
  1216  	// use in the first half of the loop right now, and then we only adjust nb explicitly
  1217  	// if the 8 bits used by each iteration isn't balanced by 8 bits loaded mid-loop.
  1218  	nb -= 4
  1219  	for {
  1220  		// Emit bitmap byte.
  1221  		// b has at least nb+4 bits, with one exception:
  1222  		// if w+4 >= nw, then b has only nw-w bits,
  1223  		// but we'll stop at the break and then truncate
  1224  		// appropriately in Phase 3.
  1225  		hb = b & bitPointerAll
  1226  		hb |= bitScanAll
  1227  		if w += 4; w >= nw {
  1228  			break
  1229  		}
  1230  		*hbitp = uint8(hb)
  1231  		hbitp = add1(hbitp)
  1232  		b >>= 4
  1233  
  1234  		// Load more bits. b has nb right now.
  1235  		if p != endp {
  1236  			// Fast path: keep reading from ptrmask.
  1237  			// nb unmodified: we just loaded 8 bits,
  1238  			// and the next iteration will consume 8 bits,
  1239  			// leaving us with the same nb the next time we're here.
  1240  			if nb < 8 {
  1241  				b |= uintptr(*p) << nb
  1242  				p = add1(p)
  1243  			} else {
  1244  				// Reduce the number of bits in b.
  1245  				// This is important if we skipped
  1246  				// over a scalar tail, since nb could
  1247  				// be larger than the bit width of b.
  1248  				nb -= 8
  1249  			}
  1250  		} else if p == nil {
  1251  			// Almost as fast path: track bit count and refill from pbits.
  1252  			// For short repetitions.
  1253  			if nb < 8 {
  1254  				b |= pbits << nb
  1255  				nb += endnb
  1256  			}
  1257  			nb -= 8 // for next iteration
  1258  		} else {
  1259  			// Slow path: reached end of ptrmask.
  1260  			// Process final partial byte and rewind to start.
  1261  			b |= uintptr(*p) << nb
  1262  			nb += endnb
  1263  			if nb < 8 {
  1264  				b |= uintptr(*ptrmask) << nb
  1265  				p = add1(ptrmask)
  1266  			} else {
  1267  				nb -= 8
  1268  				p = ptrmask
  1269  			}
  1270  		}
  1271  
  1272  		// Emit bitmap byte.
  1273  		hb = b & bitPointerAll
  1274  		hb |= bitScanAll
  1275  		if w += 4; w >= nw {
  1276  			break
  1277  		}
  1278  		*hbitp = uint8(hb)
  1279  		hbitp = add1(hbitp)
  1280  		b >>= 4
  1281  	}
  1282  
  1283  Phase3:
  1284  	// Phase 3: Write last byte or partial byte and zero the rest of the bitmap entries.
  1285  	if w > nw {
  1286  		// Counting the 4 entries in hb not yet written to memory,
  1287  		// there are more entries than possible pointer slots.
  1288  		// Discard the excess entries (can't be more than 3).
  1289  		mask := uintptr(1)<<(4-(w-nw)) - 1
  1290  		hb &= mask | mask<<4 // apply mask to both pointer bits and scan bits
  1291  	}
  1292  
  1293  	// Change nw from counting possibly-pointer words to total words in allocation.
  1294  	nw = size / sys.PtrSize
  1295  
  1296  	// Write whole bitmap bytes.
  1297  	// The first is hb, the rest are zero.
  1298  	if w <= nw {
  1299  		*hbitp = uint8(hb)
  1300  		hbitp = add1(hbitp)
  1301  		hb = 0 // for possible final half-byte below
  1302  		for w += 4; w <= nw; w += 4 {
  1303  			*hbitp = 0
  1304  			hbitp = add1(hbitp)
  1305  		}
  1306  	}
  1307  
  1308  	// Write final partial bitmap byte if any.
  1309  	// We know w > nw, or else we'd still be in the loop above.
  1310  	// It can be bigger only due to the 4 entries in hb that it counts.
  1311  	// If w == nw+4 then there's nothing left to do: we wrote all nw entries
  1312  	// and can discard the 4 sitting in hb.
  1313  	// But if w == nw+2, we need to write first two in hb.
  1314  	// The byte is shared with the next object, so be careful with
  1315  	// existing bits.
  1316  	if w == nw+2 {
  1317  		*hbitp = *hbitp&^(bitPointer|bitScan|(bitPointer|bitScan)<<heapBitsShift) | uint8(hb)
  1318  	}
  1319  
  1320  Phase4:
  1321  	// Phase 4: Copy unrolled bitmap to per-arena bitmaps, if necessary.
  1322  	if outOfPlace {
  1323  		// TODO: We could probably make this faster by
  1324  		// handling [x+dataSize, x+size) specially.
  1325  		h := heapBitsForAddr(x)
  1326  		// cnw is the number of heap words, or bit pairs
  1327  		// remaining (like nw above).
  1328  		cnw := size / sys.PtrSize
  1329  		src := (*uint8)(unsafe.Pointer(x))
  1330  		// We know the first and last byte of the bitmap are
  1331  		// not the same, but it's still possible for small
  1332  		// objects span arenas, so it may share bitmap bytes
  1333  		// with neighboring objects.
  1334  		//
  1335  		// Handle the first byte specially if it's shared. See
  1336  		// Phase 1 for why this is the only special case we need.
  1337  		if doubleCheck {
  1338  			if !(h.shift == 0 || h.shift == 2) {
  1339  				print("x=", x, " size=", size, " cnw=", h.shift, "\n")
  1340  				throw("bad start shift")
  1341  			}
  1342  		}
  1343  		if h.shift == 2 {
  1344  			*h.bitp = *h.bitp&^((bitPointer|bitScan|(bitPointer|bitScan)<<heapBitsShift)<<(2*heapBitsShift)) | *src
  1345  			h = h.next().next()
  1346  			cnw -= 2
  1347  			src = addb(src, 1)
  1348  		}
  1349  		// We're now byte aligned. Copy out to per-arena
  1350  		// bitmaps until the last byte (which may again be
  1351  		// partial).
  1352  		for cnw >= 4 {
  1353  			// This loop processes four words at a time,
  1354  			// so round cnw down accordingly.
  1355  			hNext, words := h.forwardOrBoundary(cnw / 4 * 4)
  1356  
  1357  			// n is the number of bitmap bytes to copy.
  1358  			n := words / 4
  1359  			memmove(unsafe.Pointer(h.bitp), unsafe.Pointer(src), n)
  1360  			cnw -= words
  1361  			h = hNext
  1362  			src = addb(src, n)
  1363  		}
  1364  		if doubleCheck && h.shift != 0 {
  1365  			print("cnw=", cnw, " h.shift=", h.shift, "\n")
  1366  			throw("bad shift after block copy")
  1367  		}
  1368  		// Handle the last byte if it's shared.
  1369  		if cnw == 2 {
  1370  			*h.bitp = *h.bitp&^(bitPointer|bitScan|(bitPointer|bitScan)<<heapBitsShift) | *src
  1371  			src = addb(src, 1)
  1372  			h = h.next().next()
  1373  		}
  1374  		if doubleCheck {
  1375  			if uintptr(unsafe.Pointer(src)) > x+size {
  1376  				throw("copy exceeded object size")
  1377  			}
  1378  			if !(cnw == 0 || cnw == 2) {
  1379  				print("x=", x, " size=", size, " cnw=", cnw, "\n")
  1380  				throw("bad number of remaining words")
  1381  			}
  1382  			// Set up hbitp so doubleCheck code below can check it.
  1383  			hbitp = h.bitp
  1384  		}
  1385  		// Zero the object where we wrote the bitmap.
  1386  		memclrNoHeapPointers(unsafe.Pointer(x), uintptr(unsafe.Pointer(src))-x)
  1387  	}
  1388  
  1389  	// Double check the whole bitmap.
  1390  	if doubleCheck {
  1391  		// x+size may not point to the heap, so back up one
  1392  		// word and then advance it the way we do above.
  1393  		end := heapBitsForAddr(x + size - sys.PtrSize)
  1394  		if outOfPlace {
  1395  			// In out-of-place copying, we just advance
  1396  			// using next.
  1397  			end = end.next()
  1398  		} else {
  1399  			// Don't use next because that may advance to
  1400  			// the next arena and the in-place logic
  1401  			// doesn't do that.
  1402  			end.shift += heapBitsShift
  1403  			if end.shift == 4*heapBitsShift {
  1404  				end.bitp, end.shift = add1(end.bitp), 0
  1405  			}
  1406  		}
  1407  		if typ.kind&kindGCProg == 0 && (hbitp != end.bitp || (w == nw+2) != (end.shift == 2)) {
  1408  			println("ended at wrong bitmap byte for", typ.string(), "x", dataSize/typ.size)
  1409  			print("typ.size=", typ.size, " typ.ptrdata=", typ.ptrdata, " dataSize=", dataSize, " size=", size, "\n")
  1410  			print("w=", w, " nw=", nw, " b=", hex(b), " nb=", nb, " hb=", hex(hb), "\n")
  1411  			h0 := heapBitsForAddr(x)
  1412  			print("initial bits h0.bitp=", h0.bitp, " h0.shift=", h0.shift, "\n")
  1413  			print("ended at hbitp=", hbitp, " but next starts at bitp=", end.bitp, " shift=", end.shift, "\n")
  1414  			throw("bad heapBitsSetType")
  1415  		}
  1416  
  1417  		// Double-check that bits to be written were written correctly.
  1418  		// Does not check that other bits were not written, unfortunately.
  1419  		h := heapBitsForAddr(x)
  1420  		nptr := typ.ptrdata / sys.PtrSize
  1421  		ndata := typ.size / sys.PtrSize
  1422  		count := dataSize / typ.size
  1423  		totalptr := ((count-1)*typ.size + typ.ptrdata) / sys.PtrSize
  1424  		for i := uintptr(0); i < size/sys.PtrSize; i++ {
  1425  			j := i % ndata
  1426  			var have, want uint8
  1427  			have = (*h.bitp >> h.shift) & (bitPointer | bitScan)
  1428  			if i >= totalptr {
  1429  				if typ.kind&kindGCProg != 0 && i < (totalptr+3)/4*4 {
  1430  					// heapBitsSetTypeGCProg always fills
  1431  					// in full nibbles of bitScan.
  1432  					want = bitScan
  1433  				}
  1434  			} else {
  1435  				if j < nptr && (*addb(ptrmask, j/8)>>(j%8))&1 != 0 {
  1436  					want |= bitPointer
  1437  				}
  1438  				want |= bitScan
  1439  			}
  1440  			if have != want {
  1441  				println("mismatch writing bits for", typ.string(), "x", dataSize/typ.size)
  1442  				print("typ.size=", typ.size, " typ.ptrdata=", typ.ptrdata, " dataSize=", dataSize, " size=", size, "\n")
  1443  				print("kindGCProg=", typ.kind&kindGCProg != 0, " outOfPlace=", outOfPlace, "\n")
  1444  				print("w=", w, " nw=", nw, " b=", hex(b), " nb=", nb, " hb=", hex(hb), "\n")
  1445  				h0 := heapBitsForAddr(x)
  1446  				print("initial bits h0.bitp=", h0.bitp, " h0.shift=", h0.shift, "\n")
  1447  				print("current bits h.bitp=", h.bitp, " h.shift=", h.shift, " *h.bitp=", hex(*h.bitp), "\n")
  1448  				print("ptrmask=", ptrmask, " p=", p, " endp=", endp, " endnb=", endnb, " pbits=", hex(pbits), " b=", hex(b), " nb=", nb, "\n")
  1449  				println("at word", i, "offset", i*sys.PtrSize, "have", hex(have), "want", hex(want))
  1450  				if typ.kind&kindGCProg != 0 {
  1451  					println("GC program:")
  1452  					dumpGCProg(addb(typ.gcdata, 4))
  1453  				}
  1454  				throw("bad heapBitsSetType")
  1455  			}
  1456  			h = h.next()
  1457  		}
  1458  		if ptrmask == debugPtrmask.data {
  1459  			unlock(&debugPtrmask.lock)
  1460  		}
  1461  	}
  1462  }
  1463  
  1464  var debugPtrmask struct {
  1465  	lock mutex
  1466  	data *byte
  1467  }
  1468  
  1469  // heapBitsSetTypeGCProg implements heapBitsSetType using a GC program.
  1470  // progSize is the size of the memory described by the program.
  1471  // elemSize is the size of the element that the GC program describes (a prefix of).
  1472  // dataSize is the total size of the intended data, a multiple of elemSize.
  1473  // allocSize is the total size of the allocated memory.
  1474  //
  1475  // GC programs are only used for large allocations.
  1476  // heapBitsSetType requires that allocSize is a multiple of 4 words,
  1477  // so that the relevant bitmap bytes are not shared with surrounding
  1478  // objects.
  1479  func heapBitsSetTypeGCProg(h heapBits, progSize, elemSize, dataSize, allocSize uintptr, prog *byte) {
  1480  	if sys.PtrSize == 8 && allocSize%(4*sys.PtrSize) != 0 {
  1481  		// Alignment will be wrong.
  1482  		throw("heapBitsSetTypeGCProg: small allocation")
  1483  	}
  1484  	var totalBits uintptr
  1485  	if elemSize == dataSize {
  1486  		totalBits = runGCProg(prog, nil, h.bitp, 2)
  1487  		if totalBits*sys.PtrSize != progSize {
  1488  			println("runtime: heapBitsSetTypeGCProg: total bits", totalBits, "but progSize", progSize)
  1489  			throw("heapBitsSetTypeGCProg: unexpected bit count")
  1490  		}
  1491  	} else {
  1492  		count := dataSize / elemSize
  1493  
  1494  		// Piece together program trailer to run after prog that does:
  1495  		//	literal(0)
  1496  		//	repeat(1, elemSize-progSize-1) // zeros to fill element size
  1497  		//	repeat(elemSize, count-1) // repeat that element for count
  1498  		// This zero-pads the data remaining in the first element and then
  1499  		// repeats that first element to fill the array.
  1500  		var trailer [40]byte // 3 varints (max 10 each) + some bytes
  1501  		i := 0
  1502  		if n := elemSize/sys.PtrSize - progSize/sys.PtrSize; n > 0 {
  1503  			// literal(0)
  1504  			trailer[i] = 0x01
  1505  			i++
  1506  			trailer[i] = 0
  1507  			i++
  1508  			if n > 1 {
  1509  				// repeat(1, n-1)
  1510  				trailer[i] = 0x81
  1511  				i++
  1512  				n--
  1513  				for ; n >= 0x80; n >>= 7 {
  1514  					trailer[i] = byte(n | 0x80)
  1515  					i++
  1516  				}
  1517  				trailer[i] = byte(n)
  1518  				i++
  1519  			}
  1520  		}
  1521  		// repeat(elemSize/ptrSize, count-1)
  1522  		trailer[i] = 0x80
  1523  		i++
  1524  		n := elemSize / sys.PtrSize
  1525  		for ; n >= 0x80; n >>= 7 {
  1526  			trailer[i] = byte(n | 0x80)
  1527  			i++
  1528  		}
  1529  		trailer[i] = byte(n)
  1530  		i++
  1531  		n = count - 1
  1532  		for ; n >= 0x80; n >>= 7 {
  1533  			trailer[i] = byte(n | 0x80)
  1534  			i++
  1535  		}
  1536  		trailer[i] = byte(n)
  1537  		i++
  1538  		trailer[i] = 0
  1539  		i++
  1540  
  1541  		runGCProg(prog, &trailer[0], h.bitp, 2)
  1542  
  1543  		// Even though we filled in the full array just now,
  1544  		// record that we only filled in up to the ptrdata of the
  1545  		// last element. This will cause the code below to
  1546  		// memclr the dead section of the final array element,
  1547  		// so that scanobject can stop early in the final element.
  1548  		totalBits = (elemSize*(count-1) + progSize) / sys.PtrSize
  1549  	}
  1550  	endProg := unsafe.Pointer(addb(h.bitp, (totalBits+3)/4))
  1551  	endAlloc := unsafe.Pointer(addb(h.bitp, allocSize/sys.PtrSize/wordsPerBitmapByte))
  1552  	memclrNoHeapPointers(endProg, uintptr(endAlloc)-uintptr(endProg))
  1553  }
  1554  
  1555  // progToPointerMask returns the 1-bit pointer mask output by the GC program prog.
  1556  // size the size of the region described by prog, in bytes.
  1557  // The resulting bitvector will have no more than size/sys.PtrSize bits.
  1558  func progToPointerMask(prog *byte, size uintptr) bitvector {
  1559  	n := (size/sys.PtrSize + 7) / 8
  1560  	x := (*[1 << 30]byte)(persistentalloc(n+1, 1, &memstats.buckhash_sys))[:n+1]
  1561  	x[len(x)-1] = 0xa1 // overflow check sentinel
  1562  	n = runGCProg(prog, nil, &x[0], 1)
  1563  	if x[len(x)-1] != 0xa1 {
  1564  		throw("progToPointerMask: overflow")
  1565  	}
  1566  	return bitvector{int32(n), &x[0]}
  1567  }
  1568  
  1569  // Packed GC pointer bitmaps, aka GC programs.
  1570  //
  1571  // For large types containing arrays, the type information has a
  1572  // natural repetition that can be encoded to save space in the
  1573  // binary and in the memory representation of the type information.
  1574  //
  1575  // The encoding is a simple Lempel-Ziv style bytecode machine
  1576  // with the following instructions:
  1577  //
  1578  //	00000000: stop
  1579  //	0nnnnnnn: emit n bits copied from the next (n+7)/8 bytes
  1580  //	10000000 n c: repeat the previous n bits c times; n, c are varints
  1581  //	1nnnnnnn c: repeat the previous n bits c times; c is a varint
  1582  
  1583  // runGCProg executes the GC program prog, and then trailer if non-nil,
  1584  // writing to dst with entries of the given size.
  1585  // If size == 1, dst is a 1-bit pointer mask laid out moving forward from dst.
  1586  // If size == 2, dst is the 2-bit heap bitmap, and writes move backward
  1587  // starting at dst (because the heap bitmap does). In this case, the caller guarantees
  1588  // that only whole bytes in dst need to be written.
  1589  //
  1590  // runGCProg returns the number of 1- or 2-bit entries written to memory.
  1591  func runGCProg(prog, trailer, dst *byte, size int) uintptr {
  1592  	dstStart := dst
  1593  
  1594  	// Bits waiting to be written to memory.
  1595  	var bits uintptr
  1596  	var nbits uintptr
  1597  
  1598  	p := prog
  1599  Run:
  1600  	for {
  1601  		// Flush accumulated full bytes.
  1602  		// The rest of the loop assumes that nbits <= 7.
  1603  		for ; nbits >= 8; nbits -= 8 {
  1604  			if size == 1 {
  1605  				*dst = uint8(bits)
  1606  				dst = add1(dst)
  1607  				bits >>= 8
  1608  			} else {
  1609  				v := bits&bitPointerAll | bitScanAll
  1610  				*dst = uint8(v)
  1611  				dst = add1(dst)
  1612  				bits >>= 4
  1613  				v = bits&bitPointerAll | bitScanAll
  1614  				*dst = uint8(v)
  1615  				dst = add1(dst)
  1616  				bits >>= 4
  1617  			}
  1618  		}
  1619  
  1620  		// Process one instruction.
  1621  		inst := uintptr(*p)
  1622  		p = add1(p)
  1623  		n := inst & 0x7F
  1624  		if inst&0x80 == 0 {
  1625  			// Literal bits; n == 0 means end of program.
  1626  			if n == 0 {
  1627  				// Program is over; continue in trailer if present.
  1628  				if trailer != nil {
  1629  					p = trailer
  1630  					trailer = nil
  1631  					continue
  1632  				}
  1633  				break Run
  1634  			}
  1635  			nbyte := n / 8
  1636  			for i := uintptr(0); i < nbyte; i++ {
  1637  				bits |= uintptr(*p) << nbits
  1638  				p = add1(p)
  1639  				if size == 1 {
  1640  					*dst = uint8(bits)
  1641  					dst = add1(dst)
  1642  					bits >>= 8
  1643  				} else {
  1644  					v := bits&0xf | bitScanAll
  1645  					*dst = uint8(v)
  1646  					dst = add1(dst)
  1647  					bits >>= 4
  1648  					v = bits&0xf | bitScanAll
  1649  					*dst = uint8(v)
  1650  					dst = add1(dst)
  1651  					bits >>= 4
  1652  				}
  1653  			}
  1654  			if n %= 8; n > 0 {
  1655  				bits |= uintptr(*p) << nbits
  1656  				p = add1(p)
  1657  				nbits += n
  1658  			}
  1659  			continue Run
  1660  		}
  1661  
  1662  		// Repeat. If n == 0, it is encoded in a varint in the next bytes.
  1663  		if n == 0 {
  1664  			for off := uint(0); ; off += 7 {
  1665  				x := uintptr(*p)
  1666  				p = add1(p)
  1667  				n |= (x & 0x7F) << off
  1668  				if x&0x80 == 0 {
  1669  					break
  1670  				}
  1671  			}
  1672  		}
  1673  
  1674  		// Count is encoded in a varint in the next bytes.
  1675  		c := uintptr(0)
  1676  		for off := uint(0); ; off += 7 {
  1677  			x := uintptr(*p)
  1678  			p = add1(p)
  1679  			c |= (x & 0x7F) << off
  1680  			if x&0x80 == 0 {
  1681  				break
  1682  			}
  1683  		}
  1684  		c *= n // now total number of bits to copy
  1685  
  1686  		// If the number of bits being repeated is small, load them
  1687  		// into a register and use that register for the entire loop
  1688  		// instead of repeatedly reading from memory.
  1689  		// Handling fewer than 8 bits here makes the general loop simpler.
  1690  		// The cutoff is sys.PtrSize*8 - 7 to guarantee that when we add
  1691  		// the pattern to a bit buffer holding at most 7 bits (a partial byte)
  1692  		// it will not overflow.
  1693  		src := dst
  1694  		const maxBits = sys.PtrSize*8 - 7
  1695  		if n <= maxBits {
  1696  			// Start with bits in output buffer.
  1697  			pattern := bits
  1698  			npattern := nbits
  1699  
  1700  			// If we need more bits, fetch them from memory.
  1701  			if size == 1 {
  1702  				src = subtract1(src)
  1703  				for npattern < n {
  1704  					pattern <<= 8
  1705  					pattern |= uintptr(*src)
  1706  					src = subtract1(src)
  1707  					npattern += 8
  1708  				}
  1709  			} else {
  1710  				src = subtract1(src)
  1711  				for npattern < n {
  1712  					pattern <<= 4
  1713  					pattern |= uintptr(*src) & 0xf
  1714  					src = subtract1(src)
  1715  					npattern += 4
  1716  				}
  1717  			}
  1718  
  1719  			// We started with the whole bit output buffer,
  1720  			// and then we loaded bits from whole bytes.
  1721  			// Either way, we might now have too many instead of too few.
  1722  			// Discard the extra.
  1723  			if npattern > n {
  1724  				pattern >>= npattern - n
  1725  				npattern = n
  1726  			}
  1727  
  1728  			// Replicate pattern to at most maxBits.
  1729  			if npattern == 1 {
  1730  				// One bit being repeated.
  1731  				// If the bit is 1, make the pattern all 1s.
  1732  				// If the bit is 0, the pattern is already all 0s,
  1733  				// but we can claim that the number of bits
  1734  				// in the word is equal to the number we need (c),
  1735  				// because right shift of bits will zero fill.
  1736  				if pattern == 1 {
  1737  					pattern = 1<<maxBits - 1
  1738  					npattern = maxBits
  1739  				} else {
  1740  					npattern = c
  1741  				}
  1742  			} else {
  1743  				b := pattern
  1744  				nb := npattern
  1745  				if nb+nb <= maxBits {
  1746  					// Double pattern until the whole uintptr is filled.
  1747  					for nb <= sys.PtrSize*8 {
  1748  						b |= b << nb
  1749  						nb += nb
  1750  					}
  1751  					// Trim away incomplete copy of original pattern in high bits.
  1752  					// TODO(rsc): Replace with table lookup or loop on systems without divide?
  1753  					nb = maxBits / npattern * npattern
  1754  					b &= 1<<nb - 1
  1755  					pattern = b
  1756  					npattern = nb
  1757  				}
  1758  			}
  1759  
  1760  			// Add pattern to bit buffer and flush bit buffer, c/npattern times.
  1761  			// Since pattern contains >8 bits, there will be full bytes to flush
  1762  			// on each iteration.
  1763  			for ; c >= npattern; c -= npattern {
  1764  				bits |= pattern << nbits
  1765  				nbits += npattern
  1766  				if size == 1 {
  1767  					for nbits >= 8 {
  1768  						*dst = uint8(bits)
  1769  						dst = add1(dst)
  1770  						bits >>= 8
  1771  						nbits -= 8
  1772  					}
  1773  				} else {
  1774  					for nbits >= 4 {
  1775  						*dst = uint8(bits&0xf | bitScanAll)
  1776  						dst = add1(dst)
  1777  						bits >>= 4
  1778  						nbits -= 4
  1779  					}
  1780  				}
  1781  			}
  1782  
  1783  			// Add final fragment to bit buffer.
  1784  			if c > 0 {
  1785  				pattern &= 1<<c - 1
  1786  				bits |= pattern << nbits
  1787  				nbits += c
  1788  			}
  1789  			continue Run
  1790  		}
  1791  
  1792  		// Repeat; n too large to fit in a register.
  1793  		// Since nbits <= 7, we know the first few bytes of repeated data
  1794  		// are already written to memory.
  1795  		off := n - nbits // n > nbits because n > maxBits and nbits <= 7
  1796  		if size == 1 {
  1797  			// Leading src fragment.
  1798  			src = subtractb(src, (off+7)/8)
  1799  			if frag := off & 7; frag != 0 {
  1800  				bits |= uintptr(*src) >> (8 - frag) << nbits
  1801  				src = add1(src)
  1802  				nbits += frag
  1803  				c -= frag
  1804  			}
  1805  			// Main loop: load one byte, write another.
  1806  			// The bits are rotating through the bit buffer.
  1807  			for i := c / 8; i > 0; i-- {
  1808  				bits |= uintptr(*src) << nbits
  1809  				src = add1(src)
  1810  				*dst = uint8(bits)
  1811  				dst = add1(dst)
  1812  				bits >>= 8
  1813  			}
  1814  			// Final src fragment.
  1815  			if c %= 8; c > 0 {
  1816  				bits |= (uintptr(*src) & (1<<c - 1)) << nbits
  1817  				nbits += c
  1818  			}
  1819  		} else {
  1820  			// Leading src fragment.
  1821  			src = subtractb(src, (off+3)/4)
  1822  			if frag := off & 3; frag != 0 {
  1823  				bits |= (uintptr(*src) & 0xf) >> (4 - frag) << nbits
  1824  				src = add1(src)
  1825  				nbits += frag
  1826  				c -= frag
  1827  			}
  1828  			// Main loop: load one byte, write another.
  1829  			// The bits are rotating through the bit buffer.
  1830  			for i := c / 4; i > 0; i-- {
  1831  				bits |= (uintptr(*src) & 0xf) << nbits
  1832  				src = add1(src)
  1833  				*dst = uint8(bits&0xf | bitScanAll)
  1834  				dst = add1(dst)
  1835  				bits >>= 4
  1836  			}
  1837  			// Final src fragment.
  1838  			if c %= 4; c > 0 {
  1839  				bits |= (uintptr(*src) & (1<<c - 1)) << nbits
  1840  				nbits += c
  1841  			}
  1842  		}
  1843  	}
  1844  
  1845  	// Write any final bits out, using full-byte writes, even for the final byte.
  1846  	var totalBits uintptr
  1847  	if size == 1 {
  1848  		totalBits = (uintptr(unsafe.Pointer(dst))-uintptr(unsafe.Pointer(dstStart)))*8 + nbits
  1849  		nbits += -nbits & 7
  1850  		for ; nbits > 0; nbits -= 8 {
  1851  			*dst = uint8(bits)
  1852  			dst = add1(dst)
  1853  			bits >>= 8
  1854  		}
  1855  	} else {
  1856  		totalBits = (uintptr(unsafe.Pointer(dst))-uintptr(unsafe.Pointer(dstStart)))*4 + nbits
  1857  		nbits += -nbits & 3
  1858  		for ; nbits > 0; nbits -= 4 {
  1859  			v := bits&0xf | bitScanAll
  1860  			*dst = uint8(v)
  1861  			dst = add1(dst)
  1862  			bits >>= 4
  1863  		}
  1864  	}
  1865  	return totalBits
  1866  }
  1867  
  1868  // materializeGCProg allocates space for the (1-bit) pointer bitmask
  1869  // for an object of size ptrdata.  Then it fills that space with the
  1870  // pointer bitmask specified by the program prog.
  1871  // The bitmask starts at s.startAddr.
  1872  // The result must be deallocated with dematerializeGCProg.
  1873  func materializeGCProg(ptrdata uintptr, prog *byte) *mspan {
  1874  	// Each word of ptrdata needs one bit in the bitmap.
  1875  	bitmapBytes := divRoundUp(ptrdata, 8*sys.PtrSize)
  1876  	// Compute the number of pages needed for bitmapBytes.
  1877  	pages := divRoundUp(bitmapBytes, pageSize)
  1878  	s := mheap_.allocManual(pages, spanAllocPtrScalarBits)
  1879  	runGCProg(addb(prog, 4), nil, (*byte)(unsafe.Pointer(s.startAddr)), 1)
  1880  	return s
  1881  }
  1882  func dematerializeGCProg(s *mspan) {
  1883  	mheap_.freeManual(s, spanAllocPtrScalarBits)
  1884  }
  1885  
  1886  func dumpGCProg(p *byte) {
  1887  	nptr := 0
  1888  	for {
  1889  		x := *p
  1890  		p = add1(p)
  1891  		if x == 0 {
  1892  			print("\t", nptr, " end\n")
  1893  			break
  1894  		}
  1895  		if x&0x80 == 0 {
  1896  			print("\t", nptr, " lit ", x, ":")
  1897  			n := int(x+7) / 8
  1898  			for i := 0; i < n; i++ {
  1899  				print(" ", hex(*p))
  1900  				p = add1(p)
  1901  			}
  1902  			print("\n")
  1903  			nptr += int(x)
  1904  		} else {
  1905  			nbit := int(x &^ 0x80)
  1906  			if nbit == 0 {
  1907  				for nb := uint(0); ; nb += 7 {
  1908  					x := *p
  1909  					p = add1(p)
  1910  					nbit |= int(x&0x7f) << nb
  1911  					if x&0x80 == 0 {
  1912  						break
  1913  					}
  1914  				}
  1915  			}
  1916  			count := 0
  1917  			for nb := uint(0); ; nb += 7 {
  1918  				x := *p
  1919  				p = add1(p)
  1920  				count |= int(x&0x7f) << nb
  1921  				if x&0x80 == 0 {
  1922  					break
  1923  				}
  1924  			}
  1925  			print("\t", nptr, " repeat ", nbit, " × ", count, "\n")
  1926  			nptr += nbit * count
  1927  		}
  1928  	}
  1929  }
  1930  
  1931  // Testing.
  1932  
  1933  func getgcmaskcb(frame *stkframe, ctxt unsafe.Pointer) bool {
  1934  	target := (*stkframe)(ctxt)
  1935  	if frame.sp <= target.sp && target.sp < frame.varp {
  1936  		*target = *frame
  1937  		return false
  1938  	}
  1939  	return true
  1940  }
  1941  
  1942  // gcbits returns the GC type info for x, for testing.
  1943  // The result is the bitmap entries (0 or 1), one entry per byte.
  1944  //go:linkname reflect_gcbits reflect.gcbits
  1945  func reflect_gcbits(x interface{}) []byte {
  1946  	ret := getgcmask(x)
  1947  	typ := (*ptrtype)(unsafe.Pointer(efaceOf(&x)._type)).elem
  1948  	nptr := typ.ptrdata / sys.PtrSize
  1949  	for uintptr(len(ret)) > nptr && ret[len(ret)-1] == 0 {
  1950  		ret = ret[:len(ret)-1]
  1951  	}
  1952  	return ret
  1953  }
  1954  
  1955  // Returns GC type info for the pointer stored in ep for testing.
  1956  // If ep points to the stack, only static live information will be returned
  1957  // (i.e. not for objects which are only dynamically live stack objects).
  1958  func getgcmask(ep interface{}) (mask []byte) {
  1959  	e := *efaceOf(&ep)
  1960  	p := e.data
  1961  	t := e._type
  1962  	// data or bss
  1963  	for _, datap := range activeModules() {
  1964  		// data
  1965  		if datap.data <= uintptr(p) && uintptr(p) < datap.edata {
  1966  			bitmap := datap.gcdatamask.bytedata
  1967  			n := (*ptrtype)(unsafe.Pointer(t)).elem.size
  1968  			mask = make([]byte, n/sys.PtrSize)
  1969  			for i := uintptr(0); i < n; i += sys.PtrSize {
  1970  				off := (uintptr(p) + i - datap.data) / sys.PtrSize
  1971  				mask[i/sys.PtrSize] = (*addb(bitmap, off/8) >> (off % 8)) & 1
  1972  			}
  1973  			return
  1974  		}
  1975  
  1976  		// bss
  1977  		if datap.bss <= uintptr(p) && uintptr(p) < datap.ebss {
  1978  			bitmap := datap.gcbssmask.bytedata
  1979  			n := (*ptrtype)(unsafe.Pointer(t)).elem.size
  1980  			mask = make([]byte, n/sys.PtrSize)
  1981  			for i := uintptr(0); i < n; i += sys.PtrSize {
  1982  				off := (uintptr(p) + i - datap.bss) / sys.PtrSize
  1983  				mask[i/sys.PtrSize] = (*addb(bitmap, off/8) >> (off % 8)) & 1
  1984  			}
  1985  			return
  1986  		}
  1987  	}
  1988  
  1989  	// heap
  1990  	if base, s, _ := findObject(uintptr(p), 0, 0); base != 0 {
  1991  		hbits := heapBitsForAddr(base)
  1992  		n := s.elemsize
  1993  		mask = make([]byte, n/sys.PtrSize)
  1994  		for i := uintptr(0); i < n; i += sys.PtrSize {
  1995  			if hbits.isPointer() {
  1996  				mask[i/sys.PtrSize] = 1
  1997  			}
  1998  			if !hbits.morePointers() {
  1999  				mask = mask[:i/sys.PtrSize]
  2000  				break
  2001  			}
  2002  			hbits = hbits.next()
  2003  		}
  2004  		return
  2005  	}
  2006  
  2007  	// stack
  2008  	if _g_ := getg(); _g_.m.curg.stack.lo <= uintptr(p) && uintptr(p) < _g_.m.curg.stack.hi {
  2009  		var frame stkframe
  2010  		frame.sp = uintptr(p)
  2011  		_g_ := getg()
  2012  		gentraceback(_g_.m.curg.sched.pc, _g_.m.curg.sched.sp, 0, _g_.m.curg, 0, nil, 1000, getgcmaskcb, noescape(unsafe.Pointer(&frame)), 0)
  2013  		if frame.fn.valid() {
  2014  			locals, _, _ := getStackMap(&frame, nil, false)
  2015  			if locals.n == 0 {
  2016  				return
  2017  			}
  2018  			size := uintptr(locals.n) * sys.PtrSize
  2019  			n := (*ptrtype)(unsafe.Pointer(t)).elem.size
  2020  			mask = make([]byte, n/sys.PtrSize)
  2021  			for i := uintptr(0); i < n; i += sys.PtrSize {
  2022  				off := (uintptr(p) + i - frame.varp + size) / sys.PtrSize
  2023  				mask[i/sys.PtrSize] = locals.ptrbit(off)
  2024  			}
  2025  		}
  2026  		return
  2027  	}
  2028  
  2029  	// otherwise, not something the GC knows about.
  2030  	// possibly read-only data, like malloc(0).
  2031  	// must not have pointers
  2032  	return
  2033  }
  2034  

View as plain text