Source file src/runtime/mbitmap.go

     1  // Copyright 2009 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  // 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  	"internal/goarch"
    50  	"runtime/internal/atomic"
    51  	"runtime/internal/sys"
    52  	"unsafe"
    53  )
    54  
    55  const (
    56  	bitPointer = 1 << 0
    57  	bitScan    = 1 << 4
    58  
    59  	heapBitsShift      = 1     // shift offset between successive bitPointer or bitScan entries
    60  	wordsPerBitmapByte = 8 / 2 // heap words described by one bitmap byte
    61  
    62  	// all scan/pointer bits in a byte
    63  	bitScanAll    = bitScan | bitScan<<heapBitsShift | bitScan<<(2*heapBitsShift) | bitScan<<(3*heapBitsShift)
    64  	bitPointerAll = bitPointer | bitPointer<<heapBitsShift | bitPointer<<(2*heapBitsShift) | bitPointer<<(3*heapBitsShift)
    65  )
    66  
    67  // addb returns the byte pointer p+n.
    68  //go:nowritebarrier
    69  //go:nosplit
    70  func addb(p *byte, n uintptr) *byte {
    71  	// Note: wrote out full expression instead of calling add(p, n)
    72  	// to reduce the number of temporaries generated by the
    73  	// compiler for this trivial expression during inlining.
    74  	return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) + n))
    75  }
    76  
    77  // subtractb returns the byte pointer p-n.
    78  //go:nowritebarrier
    79  //go:nosplit
    80  func subtractb(p *byte, n uintptr) *byte {
    81  	// Note: wrote out full expression instead of calling add(p, -n)
    82  	// to reduce the number of temporaries generated by the
    83  	// compiler for this trivial expression during inlining.
    84  	return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) - n))
    85  }
    86  
    87  // add1 returns the byte pointer p+1.
    88  //go:nowritebarrier
    89  //go:nosplit
    90  func add1(p *byte) *byte {
    91  	// Note: wrote out full expression instead of calling addb(p, 1)
    92  	// to reduce the number of temporaries generated by the
    93  	// compiler for this trivial expression during inlining.
    94  	return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) + 1))
    95  }
    96  
    97  // subtract1 returns the byte pointer p-1.
    98  //go:nowritebarrier
    99  //
   100  // nosplit because it is used during write barriers and must not be preempted.
   101  //go:nosplit
   102  func subtract1(p *byte) *byte {
   103  	// Note: wrote out full expression instead of calling subtractb(p, 1)
   104  	// to reduce the number of temporaries generated by the
   105  	// compiler for this trivial expression during inlining.
   106  	return (*byte)(unsafe.Pointer(uintptr(unsafe.Pointer(p)) - 1))
   107  }
   108  
   109  // heapBits provides access to the bitmap bits for a single heap word.
   110  // The methods on heapBits take value receivers so that the compiler
   111  // can more easily inline calls to those methods and registerize the
   112  // struct fields independently.
   113  type heapBits struct {
   114  	bitp  *uint8
   115  	shift uint32
   116  	arena uint32 // Index of heap arena containing bitp
   117  	last  *uint8 // Last byte arena's bitmap
   118  }
   119  
   120  // Make the compiler check that heapBits.arena is large enough to hold
   121  // the maximum arena frame number.
   122  var _ = heapBits{arena: (1<<heapAddrBits)/heapArenaBytes - 1}
   123  
   124  // markBits provides access to the mark bit for an object in the heap.
   125  // bytep points to the byte holding the mark bit.
   126  // mask is a byte with a single bit set that can be &ed with *bytep
   127  // to see if the bit has been set.
   128  // *m.byte&m.mask != 0 indicates the mark bit is set.
   129  // index can be used along with span information to generate
   130  // the address of the object in the heap.
   131  // We maintain one set of mark bits for allocation and one for
   132  // marking purposes.
   133  type markBits struct {
   134  	bytep *uint8
   135  	mask  uint8
   136  	index uintptr
   137  }
   138  
   139  //go:nosplit
   140  func (s *mspan) allocBitsForIndex(allocBitIndex uintptr) markBits {
   141  	bytep, mask := s.allocBits.bitp(allocBitIndex)
   142  	return markBits{bytep, mask, allocBitIndex}
   143  }
   144  
   145  // refillAllocCache takes 8 bytes s.allocBits starting at whichByte
   146  // and negates them so that ctz (count trailing zeros) instructions
   147  // can be used. It then places these 8 bytes into the cached 64 bit
   148  // s.allocCache.
   149  func (s *mspan) refillAllocCache(whichByte uintptr) {
   150  	bytes := (*[8]uint8)(unsafe.Pointer(s.allocBits.bytep(whichByte)))
   151  	aCache := uint64(0)
   152  	aCache |= uint64(bytes[0])
   153  	aCache |= uint64(bytes[1]) << (1 * 8)
   154  	aCache |= uint64(bytes[2]) << (2 * 8)
   155  	aCache |= uint64(bytes[3]) << (3 * 8)
   156  	aCache |= uint64(bytes[4]) << (4 * 8)
   157  	aCache |= uint64(bytes[5]) << (5 * 8)
   158  	aCache |= uint64(bytes[6]) << (6 * 8)
   159  	aCache |= uint64(bytes[7]) << (7 * 8)
   160  	s.allocCache = ^aCache
   161  }
   162  
   163  // nextFreeIndex returns the index of the next free object in s at
   164  // or after s.freeindex.
   165  // There are hardware instructions that can be used to make this
   166  // faster if profiling warrants it.
   167  func (s *mspan) nextFreeIndex() uintptr {
   168  	sfreeindex := s.freeindex
   169  	snelems := s.nelems
   170  	if sfreeindex == snelems {
   171  		return sfreeindex
   172  	}
   173  	if sfreeindex > snelems {
   174  		throw("s.freeindex > s.nelems")
   175  	}
   176  
   177  	aCache := s.allocCache
   178  
   179  	bitIndex := sys.Ctz64(aCache)
   180  	for bitIndex == 64 {
   181  		// Move index to start of next cached bits.
   182  		sfreeindex = (sfreeindex + 64) &^ (64 - 1)
   183  		if sfreeindex >= snelems {
   184  			s.freeindex = snelems
   185  			return snelems
   186  		}
   187  		whichByte := sfreeindex / 8
   188  		// Refill s.allocCache with the next 64 alloc bits.
   189  		s.refillAllocCache(whichByte)
   190  		aCache = s.allocCache
   191  		bitIndex = sys.Ctz64(aCache)
   192  		// nothing available in cached bits
   193  		// grab the next 8 bytes and try again.
   194  	}
   195  	result := sfreeindex + uintptr(bitIndex)
   196  	if result >= snelems {
   197  		s.freeindex = snelems
   198  		return snelems
   199  	}
   200  
   201  	s.allocCache >>= uint(bitIndex + 1)
   202  	sfreeindex = result + 1
   203  
   204  	if sfreeindex%64 == 0 && sfreeindex != snelems {
   205  		// We just incremented s.freeindex so it isn't 0.
   206  		// As each 1 in s.allocCache was encountered and used for allocation
   207  		// it was shifted away. At this point s.allocCache contains all 0s.
   208  		// Refill s.allocCache so that it corresponds
   209  		// to the bits at s.allocBits starting at s.freeindex.
   210  		whichByte := sfreeindex / 8
   211  		s.refillAllocCache(whichByte)
   212  	}
   213  	s.freeindex = sfreeindex
   214  	return result
   215  }
   216  
   217  // isFree reports whether the index'th object in s is unallocated.
   218  //
   219  // The caller must ensure s.state is mSpanInUse, and there must have
   220  // been no preemption points since ensuring this (which could allow a
   221  // GC transition, which would allow the state to change).
   222  func (s *mspan) isFree(index uintptr) bool {
   223  	if index < s.freeindex {
   224  		return false
   225  	}
   226  	bytep, mask := s.allocBits.bitp(index)
   227  	return *bytep&mask == 0
   228  }
   229  
   230  // divideByElemSize returns n/s.elemsize.
   231  // n must be within [0, s.npages*_PageSize),
   232  // or may be exactly s.npages*_PageSize
   233  // if s.elemsize is from sizeclasses.go.
   234  func (s *mspan) divideByElemSize(n uintptr) uintptr {
   235  	const doubleCheck = false
   236  
   237  	// See explanation in mksizeclasses.go's computeDivMagic.
   238  	q := uintptr((uint64(n) * uint64(s.divMul)) >> 32)
   239  
   240  	if doubleCheck && q != n/s.elemsize {
   241  		println(n, "/", s.elemsize, "should be", n/s.elemsize, "but got", q)
   242  		throw("bad magic division")
   243  	}
   244  	return q
   245  }
   246  
   247  func (s *mspan) objIndex(p uintptr) uintptr {
   248  	return s.divideByElemSize(p - s.base())
   249  }
   250  
   251  func markBitsForAddr(p uintptr) markBits {
   252  	s := spanOf(p)
   253  	objIndex := s.objIndex(p)
   254  	return s.markBitsForIndex(objIndex)
   255  }
   256  
   257  func (s *mspan) markBitsForIndex(objIndex uintptr) markBits {
   258  	bytep, mask := s.gcmarkBits.bitp(objIndex)
   259  	return markBits{bytep, mask, objIndex}
   260  }
   261  
   262  func (s *mspan) markBitsForBase() markBits {
   263  	return markBits{(*uint8)(s.gcmarkBits), uint8(1), 0}
   264  }
   265  
   266  // isMarked reports whether mark bit m is set.
   267  func (m markBits) isMarked() bool {
   268  	return *m.bytep&m.mask != 0
   269  }
   270  
   271  // setMarked sets the marked bit in the markbits, atomically.
   272  func (m markBits) setMarked() {
   273  	// Might be racing with other updates, so use atomic update always.
   274  	// We used to be clever here and use a non-atomic update in certain
   275  	// cases, but it's not worth the risk.
   276  	atomic.Or8(m.bytep, m.mask)
   277  }
   278  
   279  // setMarkedNonAtomic sets the marked bit in the markbits, non-atomically.
   280  func (m markBits) setMarkedNonAtomic() {
   281  	*m.bytep |= m.mask
   282  }
   283  
   284  // clearMarked clears the marked bit in the markbits, atomically.
   285  func (m markBits) clearMarked() {
   286  	// Might be racing with other updates, so use atomic update always.
   287  	// We used to be clever here and use a non-atomic update in certain
   288  	// cases, but it's not worth the risk.
   289  	atomic.And8(m.bytep, ^m.mask)
   290  }
   291  
   292  // markBitsForSpan returns the markBits for the span base address base.
   293  func markBitsForSpan(base uintptr) (mbits markBits) {
   294  	mbits = markBitsForAddr(base)
   295  	if mbits.mask != 1 {
   296  		throw("markBitsForSpan: unaligned start")
   297  	}
   298  	return mbits
   299  }
   300  
   301  // advance advances the markBits to the next object in the span.
   302  func (m *markBits) advance() {
   303  	if m.mask == 1<<7 {
   304  		m.bytep = (*uint8)(unsafe.Pointer(uintptr(unsafe.Pointer(m.bytep)) + 1))
   305  		m.mask = 1
   306  	} else {
   307  		m.mask = m.mask << 1
   308  	}
   309  	m.index++
   310  }
   311  
   312  // heapBitsForAddr returns the heapBits for the address addr.
   313  // The caller must ensure addr is in an allocated span.
   314  // In particular, be careful not to point past the end of an object.
   315  //
   316  // nosplit because it is used during write barriers and must not be preempted.
   317  //go:nosplit
   318  func heapBitsForAddr(addr uintptr) (h heapBits) {
   319  	// 2 bits per word, 4 pairs per byte, and a mask is hard coded.
   320  	arena := arenaIndex(addr)
   321  	ha := mheap_.arenas[arena.l1()][arena.l2()]
   322  	// The compiler uses a load for nil checking ha, but in this
   323  	// case we'll almost never hit that cache line again, so it
   324  	// makes more sense to do a value check.
   325  	if ha == nil {
   326  		// addr is not in the heap. Return nil heapBits, which
   327  		// we expect to crash in the caller.
   328  		return
   329  	}
   330  	h.bitp = &ha.bitmap[(addr/(goarch.PtrSize*4))%heapArenaBitmapBytes]
   331  	h.shift = uint32((addr / goarch.PtrSize) & 3)
   332  	h.arena = uint32(arena)
   333  	h.last = &ha.bitmap[len(ha.bitmap)-1]
   334  	return
   335  }
   336  
   337  // clobberdeadPtr is a special value that is used by the compiler to
   338  // clobber dead stack slots, when -clobberdead flag is set.
   339  const clobberdeadPtr = uintptr(0xdeaddead | 0xdeaddead<<((^uintptr(0)>>63)*32))
   340  
   341  // badPointer throws bad pointer in heap panic.
   342  func badPointer(s *mspan, p, refBase, refOff uintptr) {
   343  	// Typically this indicates an incorrect use
   344  	// of unsafe or cgo to store a bad pointer in
   345  	// the Go heap. It may also indicate a runtime
   346  	// bug.
   347  	//
   348  	// TODO(austin): We could be more aggressive
   349  	// and detect pointers to unallocated objects
   350  	// in allocated spans.
   351  	printlock()
   352  	print("runtime: pointer ", hex(p))
   353  	if s != nil {
   354  		state := s.state.get()
   355  		if state != mSpanInUse {
   356  			print(" to unallocated span")
   357  		} else {
   358  			print(" to unused region of span")
   359  		}
   360  		print(" span.base()=", hex(s.base()), " span.limit=", hex(s.limit), " span.state=", state)
   361  	}
   362  	print("\n")
   363  	if refBase != 0 {
   364  		print("runtime: found in object at *(", hex(refBase), "+", hex(refOff), ")\n")
   365  		gcDumpObject("object", refBase, refOff)
   366  	}
   367  	getg().m.traceback = 2
   368  	throw("found bad pointer in Go heap (incorrect use of unsafe or cgo?)")
   369  }
   370  
   371  // findObject returns the base address for the heap object containing
   372  // the address p, the object's span, and the index of the object in s.
   373  // If p does not point into a heap object, it returns base == 0.
   374  //
   375  // If p points is an invalid heap pointer and debug.invalidptr != 0,
   376  // findObject panics.
   377  //
   378  // refBase and refOff optionally give the base address of the object
   379  // in which the pointer p was found and the byte offset at which it
   380  // was found. These are used for error reporting.
   381  //
   382  // It is nosplit so it is safe for p to be a pointer to the current goroutine's stack.
   383  // Since p is a uintptr, it would not be adjusted if the stack were to move.
   384  //go:nosplit
   385  func findObject(p, refBase, refOff uintptr) (base uintptr, s *mspan, objIndex uintptr) {
   386  	s = spanOf(p)
   387  	// If s is nil, the virtual address has never been part of the heap.
   388  	// This pointer may be to some mmap'd region, so we allow it.
   389  	if s == nil {
   390  		if (GOARCH == "amd64" || GOARCH == "arm64") && p == clobberdeadPtr && debug.invalidptr != 0 {
   391  			// Crash if clobberdeadPtr is seen. Only on AMD64 and ARM64 for now,
   392  			// as they are the only platform where compiler's clobberdead mode is
   393  			// implemented. On these platforms clobberdeadPtr cannot be a valid address.
   394  			badPointer(s, p, refBase, refOff)
   395  		}
   396  		return
   397  	}
   398  	// If p is a bad pointer, it may not be in s's bounds.
   399  	//
   400  	// Check s.state to synchronize with span initialization
   401  	// before checking other fields. See also spanOfHeap.
   402  	if state := s.state.get(); state != mSpanInUse || p < s.base() || p >= s.limit {
   403  		// Pointers into stacks are also ok, the runtime manages these explicitly.
   404  		if state == mSpanManual {
   405  			return
   406  		}
   407  		// The following ensures that we are rigorous about what data
   408  		// structures hold valid pointers.
   409  		if debug.invalidptr != 0 {
   410  			badPointer(s, p, refBase, refOff)
   411  		}
   412  		return
   413  	}
   414  
   415  	objIndex = s.objIndex(p)
   416  	base = s.base() + objIndex*s.elemsize
   417  	return
   418  }
   419  
   420  // verifyNotInHeapPtr reports whether converting the not-in-heap pointer into a unsafe.Pointer is ok.
   421  //go:linkname reflect_verifyNotInHeapPtr reflect.verifyNotInHeapPtr
   422  func reflect_verifyNotInHeapPtr(p uintptr) bool {
   423  	// Conversion to a pointer is ok as long as findObject above does not call badPointer.
   424  	// Since we're already promised that p doesn't point into the heap, just disallow heap
   425  	// pointers and the special clobbered pointer.
   426  	return spanOf(p) == nil && p != clobberdeadPtr
   427  }
   428  
   429  // next returns the heapBits describing the next pointer-sized word in memory.
   430  // That is, if h describes address p, h.next() describes p+ptrSize.
   431  // Note that next does not modify h. The caller must record the result.
   432  //
   433  // nosplit because it is used during write barriers and must not be preempted.
   434  //go:nosplit
   435  func (h heapBits) next() heapBits {
   436  	if h.shift < 3*heapBitsShift {
   437  		h.shift += heapBitsShift
   438  	} else if h.bitp != h.last {
   439  		h.bitp, h.shift = add1(h.bitp), 0
   440  	} else {
   441  		// Move to the next arena.
   442  		return h.nextArena()
   443  	}
   444  	return h
   445  }
   446  
   447  // nextArena advances h to the beginning of the next heap arena.
   448  //
   449  // This is a slow-path helper to next. gc's inliner knows that
   450  // heapBits.next can be inlined even though it calls this. This is
   451  // marked noinline so it doesn't get inlined into next and cause next
   452  // to be too big to inline.
   453  //
   454  //go:nosplit
   455  //go:noinline
   456  func (h heapBits) nextArena() heapBits {
   457  	h.arena++
   458  	ai := arenaIdx(h.arena)
   459  	l2 := mheap_.arenas[ai.l1()]
   460  	if l2 == nil {
   461  		// We just passed the end of the object, which
   462  		// was also the end of the heap. Poison h. It
   463  		// should never be dereferenced at this point.
   464  		return heapBits{}
   465  	}
   466  	ha := l2[ai.l2()]
   467  	if ha == nil {
   468  		return heapBits{}
   469  	}
   470  	h.bitp, h.shift = &ha.bitmap[0], 0
   471  	h.last = &ha.bitmap[len(ha.bitmap)-1]
   472  	return h
   473  }
   474  
   475  // forward returns the heapBits describing n pointer-sized words ahead of h in memory.
   476  // That is, if h describes address p, h.forward(n) describes p+n*ptrSize.
   477  // h.forward(1) is equivalent to h.next(), just slower.
   478  // Note that forward does not modify h. The caller must record the result.
   479  // bits returns the heap bits for the current word.
   480  //go:nosplit
   481  func (h heapBits) forward(n uintptr) heapBits {
   482  	n += uintptr(h.shift) / heapBitsShift
   483  	nbitp := uintptr(unsafe.Pointer(h.bitp)) + n/4
   484  	h.shift = uint32(n%4) * heapBitsShift
   485  	if nbitp <= uintptr(unsafe.Pointer(h.last)) {
   486  		h.bitp = (*uint8)(unsafe.Pointer(nbitp))
   487  		return h
   488  	}
   489  
   490  	// We're in a new heap arena.
   491  	past := nbitp - (uintptr(unsafe.Pointer(h.last)) + 1)
   492  	h.arena += 1 + uint32(past/heapArenaBitmapBytes)
   493  	ai := arenaIdx(h.arena)
   494  	if l2 := mheap_.arenas[ai.l1()]; l2 != nil && l2[ai.l2()] != nil {
   495  		a := l2[ai.l2()]
   496  		h.bitp = &a.bitmap[past%heapArenaBitmapBytes]
   497  		h.last = &a.bitmap[len(a.bitmap)-1]
   498  	} else {
   499  		h.bitp, h.last = nil, nil
   500  	}
   501  	return h
   502  }
   503  
   504  // forwardOrBoundary is like forward, but stops at boundaries between
   505  // contiguous sections of the bitmap. It returns the number of words
   506  // advanced over, which will be <= n.
   507  func (h heapBits) forwardOrBoundary(n uintptr) (heapBits, uintptr) {
   508  	maxn := 4 * ((uintptr(unsafe.Pointer(h.last)) + 1) - uintptr(unsafe.Pointer(h.bitp)))
   509  	if n > maxn {
   510  		n = maxn
   511  	}
   512  	return h.forward(n), n
   513  }
   514  
   515  // The caller can test morePointers and isPointer by &-ing with bitScan and bitPointer.
   516  // The result includes in its higher bits the bits for subsequent words
   517  // described by the same bitmap byte.
   518  //
   519  // nosplit because it is used during write barriers and must not be preempted.
   520  //go:nosplit
   521  func (h heapBits) bits() uint32 {
   522  	// The (shift & 31) eliminates a test and conditional branch
   523  	// from the generated code.
   524  	return uint32(*h.bitp) >> (h.shift & 31)
   525  }
   526  
   527  // morePointers reports whether this word and all remaining words in this object
   528  // are scalars.
   529  // h must not describe the second word of the object.
   530  func (h heapBits) morePointers() bool {
   531  	return h.bits()&bitScan != 0
   532  }
   533  
   534  // isPointer reports whether the heap bits describe a pointer word.
   535  //
   536  // nosplit because it is used during write barriers and must not be preempted.
   537  //go:nosplit
   538  func (h heapBits) isPointer() bool {
   539  	return h.bits()&bitPointer != 0
   540  }
   541  
   542  // bulkBarrierPreWrite executes a write barrier
   543  // for every pointer slot in the memory range [src, src+size),
   544  // using pointer/scalar information from [dst, dst+size).
   545  // This executes the write barriers necessary before a memmove.
   546  // src, dst, and size must be pointer-aligned.
   547  // The range [dst, dst+size) must lie within a single object.
   548  // It does not perform the actual writes.
   549  //
   550  // As a special case, src == 0 indicates that this is being used for a
   551  // memclr. bulkBarrierPreWrite will pass 0 for the src of each write
   552  // barrier.
   553  //
   554  // Callers should call bulkBarrierPreWrite immediately before
   555  // calling memmove(dst, src, size). This function is marked nosplit
   556  // to avoid being preempted; the GC must not stop the goroutine
   557  // between the memmove and the execution of the barriers.
   558  // The caller is also responsible for cgo pointer checks if this
   559  // may be writing Go pointers into non-Go memory.
   560  //
   561  // The pointer bitmap is not maintained for allocations containing
   562  // no pointers at all; any caller of bulkBarrierPreWrite must first
   563  // make sure the underlying allocation contains pointers, usually
   564  // by checking typ.ptrdata.
   565  //
   566  // Callers must perform cgo checks if writeBarrier.cgo.
   567  //
   568  //go:nosplit
   569  func bulkBarrierPreWrite(dst, src, size uintptr) {
   570  	if (dst|src|size)&(goarch.PtrSize-1) != 0 {
   571  		throw("bulkBarrierPreWrite: unaligned arguments")
   572  	}
   573  	if !writeBarrier.needed {
   574  		return
   575  	}
   576  	if s := spanOf(dst); s == nil {
   577  		// If dst is a global, use the data or BSS bitmaps to
   578  		// execute write barriers.
   579  		for _, datap := range activeModules() {
   580  			if datap.data <= dst && dst < datap.edata {
   581  				bulkBarrierBitmap(dst, src, size, dst-datap.data, datap.gcdatamask.bytedata)
   582  				return
   583  			}
   584  		}
   585  		for _, datap := range activeModules() {
   586  			if datap.bss <= dst && dst < datap.ebss {
   587  				bulkBarrierBitmap(dst, src, size, dst-datap.bss, datap.gcbssmask.bytedata)
   588  				return
   589  			}
   590  		}
   591  		return
   592  	} else if s.state.get() != mSpanInUse || dst < s.base() || s.limit <= dst {
   593  		// dst was heap memory at some point, but isn't now.
   594  		// It can't be a global. It must be either our stack,
   595  		// or in the case of direct channel sends, it could be
   596  		// another stack. Either way, no need for barriers.
   597  		// This will also catch if dst is in a freed span,
   598  		// though that should never have.
   599  		return
   600  	}
   601  
   602  	buf := &getg().m.p.ptr().wbBuf
   603  	h := heapBitsForAddr(dst)
   604  	if src == 0 {
   605  		for i := uintptr(0); i < size; i += goarch.PtrSize {
   606  			if h.isPointer() {
   607  				dstx := (*uintptr)(unsafe.Pointer(dst + i))
   608  				if !buf.putFast(*dstx, 0) {
   609  					wbBufFlush(nil, 0)
   610  				}
   611  			}
   612  			h = h.next()
   613  		}
   614  	} else {
   615  		for i := uintptr(0); i < size; i += goarch.PtrSize {
   616  			if h.isPointer() {
   617  				dstx := (*uintptr)(unsafe.Pointer(dst + i))
   618  				srcx := (*uintptr)(unsafe.Pointer(src + i))
   619  				if !buf.putFast(*dstx, *srcx) {
   620  					wbBufFlush(nil, 0)
   621  				}
   622  			}
   623  			h = h.next()
   624  		}
   625  	}
   626  }
   627  
   628  // bulkBarrierPreWriteSrcOnly is like bulkBarrierPreWrite but
   629  // does not execute write barriers for [dst, dst+size).
   630  //
   631  // In addition to the requirements of bulkBarrierPreWrite
   632  // callers need to ensure [dst, dst+size) is zeroed.
   633  //
   634  // This is used for special cases where e.g. dst was just
   635  // created and zeroed with malloc.
   636  //go:nosplit
   637  func bulkBarrierPreWriteSrcOnly(dst, src, size uintptr) {
   638  	if (dst|src|size)&(goarch.PtrSize-1) != 0 {
   639  		throw("bulkBarrierPreWrite: unaligned arguments")
   640  	}
   641  	if !writeBarrier.needed {
   642  		return
   643  	}
   644  	buf := &getg().m.p.ptr().wbBuf
   645  	h := heapBitsForAddr(dst)
   646  	for i := uintptr(0); i < size; i += goarch.PtrSize {
   647  		if h.isPointer() {
   648  			srcx := (*uintptr)(unsafe.Pointer(src + i))
   649  			if !buf.putFast(0, *srcx) {
   650  				wbBufFlush(nil, 0)
   651  			}
   652  		}
   653  		h = h.next()
   654  	}
   655  }
   656  
   657  // bulkBarrierBitmap executes write barriers for copying from [src,
   658  // src+size) to [dst, dst+size) using a 1-bit pointer bitmap. src is
   659  // assumed to start maskOffset bytes into the data covered by the
   660  // bitmap in bits (which may not be a multiple of 8).
   661  //
   662  // This is used by bulkBarrierPreWrite for writes to data and BSS.
   663  //
   664  //go:nosplit
   665  func bulkBarrierBitmap(dst, src, size, maskOffset uintptr, bits *uint8) {
   666  	word := maskOffset / goarch.PtrSize
   667  	bits = addb(bits, word/8)
   668  	mask := uint8(1) << (word % 8)
   669  
   670  	buf := &getg().m.p.ptr().wbBuf
   671  	for i := uintptr(0); i < size; i += goarch.PtrSize {
   672  		if mask == 0 {
   673  			bits = addb(bits, 1)
   674  			if *bits == 0 {
   675  				// Skip 8 words.
   676  				i += 7 * goarch.PtrSize
   677  				continue
   678  			}
   679  			mask = 1
   680  		}
   681  		if *bits&mask != 0 {
   682  			dstx := (*uintptr)(unsafe.Pointer(dst + i))
   683  			if src == 0 {
   684  				if !buf.putFast(*dstx, 0) {
   685  					wbBufFlush(nil, 0)
   686  				}
   687  			} else {
   688  				srcx := (*uintptr)(unsafe.Pointer(src + i))
   689  				if !buf.putFast(*dstx, *srcx) {
   690  					wbBufFlush(nil, 0)
   691  				}
   692  			}
   693  		}
   694  		mask <<= 1
   695  	}
   696  }
   697  
   698  // typeBitsBulkBarrier executes a write barrier for every
   699  // pointer that would be copied from [src, src+size) to [dst,
   700  // dst+size) by a memmove using the type bitmap to locate those
   701  // pointer slots.
   702  //
   703  // The type typ must correspond exactly to [src, src+size) and [dst, dst+size).
   704  // dst, src, and size must be pointer-aligned.
   705  // The type typ must have a plain bitmap, not a GC program.
   706  // The only use of this function is in channel sends, and the
   707  // 64 kB channel element limit takes care of this for us.
   708  //
   709  // Must not be preempted because it typically runs right before memmove,
   710  // and the GC must observe them as an atomic action.
   711  //
   712  // Callers must perform cgo checks if writeBarrier.cgo.
   713  //
   714  //go:nosplit
   715  func typeBitsBulkBarrier(typ *_type, dst, src, size uintptr) {
   716  	if typ == nil {
   717  		throw("runtime: typeBitsBulkBarrier without type")
   718  	}
   719  	if typ.size != size {
   720  		println("runtime: typeBitsBulkBarrier with type ", typ.string(), " of size ", typ.size, " but memory size", size)
   721  		throw("runtime: invalid typeBitsBulkBarrier")
   722  	}
   723  	if typ.kind&kindGCProg != 0 {
   724  		println("runtime: typeBitsBulkBarrier with type ", typ.string(), " with GC prog")
   725  		throw("runtime: invalid typeBitsBulkBarrier")
   726  	}
   727  	if !writeBarrier.needed {
   728  		return
   729  	}
   730  	ptrmask := typ.gcdata
   731  	buf := &getg().m.p.ptr().wbBuf
   732  	var bits uint32
   733  	for i := uintptr(0); i < typ.ptrdata; i += goarch.PtrSize {
   734  		if i&(goarch.PtrSize*8-1) == 0 {
   735  			bits = uint32(*ptrmask)
   736  			ptrmask = addb(ptrmask, 1)
   737  		} else {
   738  			bits = bits >> 1
   739  		}
   740  		if bits&1 != 0 {
   741  			dstx := (*uintptr)(unsafe.Pointer(dst + i))
   742  			srcx := (*uintptr)(unsafe.Pointer(src + i))
   743  			if !buf.putFast(*dstx, *srcx) {
   744  				wbBufFlush(nil, 0)
   745  			}
   746  		}
   747  	}
   748  }
   749  
   750  // The methods operating on spans all require that h has been returned
   751  // by heapBitsForSpan and that size, n, total are the span layout description
   752  // returned by the mspan's layout method.
   753  // If total > size*n, it means that there is extra leftover memory in the span,
   754  // usually due to rounding.
   755  //
   756  // TODO(rsc): Perhaps introduce a different heapBitsSpan type.
   757  
   758  // initSpan initializes the heap bitmap for a span.
   759  // If this is a span of pointer-sized objects, it initializes all
   760  // words to pointer/scan.
   761  // Otherwise, it initializes all words to scalar/dead.
   762  func (h heapBits) initSpan(s *mspan) {
   763  	// Clear bits corresponding to objects.
   764  	nw := (s.npages << _PageShift) / goarch.PtrSize
   765  	if nw%wordsPerBitmapByte != 0 {
   766  		throw("initSpan: unaligned length")
   767  	}
   768  	if h.shift != 0 {
   769  		throw("initSpan: unaligned base")
   770  	}
   771  	isPtrs := goarch.PtrSize == 8 && s.elemsize == goarch.PtrSize
   772  	for nw > 0 {
   773  		hNext, anw := h.forwardOrBoundary(nw)
   774  		nbyte := anw / wordsPerBitmapByte
   775  		if isPtrs {
   776  			bitp := h.bitp
   777  			for i := uintptr(0); i < nbyte; i++ {
   778  				*bitp = bitPointerAll | bitScanAll
   779  				bitp = add1(bitp)
   780  			}
   781  		} else {
   782  			memclrNoHeapPointers(unsafe.Pointer(h.bitp), nbyte)
   783  		}
   784  		h = hNext
   785  		nw -= anw
   786  	}
   787  }
   788  
   789  // countAlloc returns the number of objects allocated in span s by
   790  // scanning the allocation bitmap.
   791  func (s *mspan) countAlloc() int {
   792  	count := 0
   793  	bytes := divRoundUp(s.nelems, 8)
   794  	// Iterate over each 8-byte chunk and count allocations
   795  	// with an intrinsic. Note that newMarkBits guarantees that
   796  	// gcmarkBits will be 8-byte aligned, so we don't have to
   797  	// worry about edge cases, irrelevant bits will simply be zero.
   798  	for i := uintptr(0); i < bytes; i += 8 {
   799  		// Extract 64 bits from the byte pointer and get a OnesCount.
   800  		// Note that the unsafe cast here doesn't preserve endianness,
   801  		// but that's OK. We only care about how many bits are 1, not
   802  		// about the order we discover them in.
   803  		mrkBits := *(*uint64)(unsafe.Pointer(s.gcmarkBits.bytep(i)))
   804  		count += sys.OnesCount64(mrkBits)
   805  	}
   806  	return count
   807  }
   808  
   809  // heapBitsSetType records that the new allocation [x, x+size)
   810  // holds in [x, x+dataSize) one or more values of type typ.
   811  // (The number of values is given by dataSize / typ.size.)
   812  // If dataSize < size, the fragment [x+dataSize, x+size) is
   813  // recorded as non-pointer data.
   814  // It is known that the type has pointers somewhere;
   815  // malloc does not call heapBitsSetType when there are no pointers,
   816  // because all free objects are marked as noscan during
   817  // heapBitsSweepSpan.
   818  //
   819  // There can only be one allocation from a given span active at a time,
   820  // and the bitmap for a span always falls on byte boundaries,
   821  // so there are no write-write races for access to the heap bitmap.
   822  // Hence, heapBitsSetType can access the bitmap without atomics.
   823  //
   824  // There can be read-write races between heapBitsSetType and things
   825  // that read the heap bitmap like scanobject. However, since
   826  // heapBitsSetType is only used for objects that have not yet been
   827  // made reachable, readers will ignore bits being modified by this
   828  // function. This does mean this function cannot transiently modify
   829  // bits that belong to neighboring objects. Also, on weakly-ordered
   830  // machines, callers must execute a store/store (publication) barrier
   831  // between calling this function and making the object reachable.
   832  func heapBitsSetType(x, size, dataSize uintptr, typ *_type) {
   833  	const doubleCheck = false // slow but helpful; enable to test modifications to this code
   834  
   835  	const (
   836  		mask1 = bitPointer | bitScan                        // 00010001
   837  		mask2 = bitPointer | bitScan | mask1<<heapBitsShift // 00110011
   838  		mask3 = bitPointer | bitScan | mask2<<heapBitsShift // 01110111
   839  	)
   840  
   841  	// dataSize is always size rounded up to the next malloc size class,
   842  	// except in the case of allocating a defer block, in which case
   843  	// size is sizeof(_defer{}) (at least 6 words) and dataSize may be
   844  	// arbitrarily larger.
   845  	//
   846  	// The checks for size == goarch.PtrSize and size == 2*goarch.PtrSize can therefore
   847  	// assume that dataSize == size without checking it explicitly.
   848  
   849  	if goarch.PtrSize == 8 && size == goarch.PtrSize {
   850  		// It's one word and it has pointers, it must be a pointer.
   851  		// Since all allocated one-word objects are pointers
   852  		// (non-pointers are aggregated into tinySize allocations),
   853  		// initSpan sets the pointer bits for us. Nothing to do here.
   854  		if doubleCheck {
   855  			h := heapBitsForAddr(x)
   856  			if !h.isPointer() {
   857  				throw("heapBitsSetType: pointer bit missing")
   858  			}
   859  			if !h.morePointers() {
   860  				throw("heapBitsSetType: scan bit missing")
   861  			}
   862  		}
   863  		return
   864  	}
   865  
   866  	h := heapBitsForAddr(x)
   867  	ptrmask := typ.gcdata // start of 1-bit pointer mask (or GC program, handled below)
   868  
   869  	// 2-word objects only have 4 bitmap bits and 3-word objects only have 6 bitmap bits.
   870  	// Therefore, these objects share a heap bitmap byte with the objects next to them.
   871  	// These are called out as a special case primarily so the code below can assume all
   872  	// objects are at least 4 words long and that their bitmaps start either at the beginning
   873  	// of a bitmap byte, or half-way in (h.shift of 0 and 2 respectively).
   874  
   875  	if size == 2*goarch.PtrSize {
   876  		if typ.size == goarch.PtrSize {
   877  			// We're allocating a block big enough to hold two pointers.
   878  			// On 64-bit, that means the actual object must be two pointers,
   879  			// or else we'd have used the one-pointer-sized block.
   880  			// On 32-bit, however, this is the 8-byte block, the smallest one.
   881  			// So it could be that we're allocating one pointer and this was
   882  			// just the smallest block available. Distinguish by checking dataSize.
   883  			// (In general the number of instances of typ being allocated is
   884  			// dataSize/typ.size.)
   885  			if goarch.PtrSize == 4 && dataSize == goarch.PtrSize {
   886  				// 1 pointer object. On 32-bit machines clear the bit for the
   887  				// unused second word.
   888  				*h.bitp &^= (bitPointer | bitScan | (bitPointer|bitScan)<<heapBitsShift) << h.shift
   889  				*h.bitp |= (bitPointer | bitScan) << h.shift
   890  			} else {
   891  				// 2-element array of pointer.
   892  				*h.bitp |= (bitPointer | bitScan | (bitPointer|bitScan)<<heapBitsShift) << h.shift
   893  			}
   894  			return
   895  		}
   896  		// Otherwise typ.size must be 2*goarch.PtrSize,
   897  		// and typ.kind&kindGCProg == 0.
   898  		if doubleCheck {
   899  			if typ.size != 2*goarch.PtrSize || typ.kind&kindGCProg != 0 {
   900  				print("runtime: heapBitsSetType size=", size, " but typ.size=", typ.size, " gcprog=", typ.kind&kindGCProg != 0, "\n")
   901  				throw("heapBitsSetType")
   902  			}
   903  		}
   904  		b := uint32(*ptrmask)
   905  		hb := b & 3
   906  		hb |= bitScanAll & ((bitScan << (typ.ptrdata / goarch.PtrSize)) - 1)
   907  		// Clear the bits for this object so we can set the
   908  		// appropriate ones.
   909  		*h.bitp &^= (bitPointer | bitScan | ((bitPointer | bitScan) << heapBitsShift)) << h.shift
   910  		*h.bitp |= uint8(hb << h.shift)
   911  		return
   912  	} else if size == 3*goarch.PtrSize {
   913  		b := uint8(*ptrmask)
   914  		if doubleCheck {
   915  			if b == 0 {
   916  				println("runtime: invalid type ", typ.string())
   917  				throw("heapBitsSetType: called with non-pointer type")
   918  			}
   919  			if goarch.PtrSize != 8 {
   920  				throw("heapBitsSetType: unexpected 3 pointer wide size class on 32 bit")
   921  			}
   922  			if typ.kind&kindGCProg != 0 {
   923  				throw("heapBitsSetType: unexpected GC prog for 3 pointer wide size class")
   924  			}
   925  			if typ.size == 2*goarch.PtrSize {
   926  				print("runtime: heapBitsSetType size=", size, " but typ.size=", typ.size, "\n")
   927  				throw("heapBitsSetType: inconsistent object sizes")
   928  			}
   929  		}
   930  		if typ.size == goarch.PtrSize {
   931  			// The type contains a pointer otherwise heapBitsSetType wouldn't have been called.
   932  			// Since the type is only 1 pointer wide and contains a pointer, its gcdata must be exactly 1.
   933  			if doubleCheck && *typ.gcdata != 1 {
   934  				print("runtime: heapBitsSetType size=", size, " typ.size=", typ.size, "but *typ.gcdata", *typ.gcdata, "\n")
   935  				throw("heapBitsSetType: unexpected gcdata for 1 pointer wide type size in 3 pointer wide size class")
   936  			}
   937  			// 3 element array of pointers. Unrolling ptrmask 3 times into p yields 00000111.
   938  			b = 7
   939  		}
   940  
   941  		hb := b & 7
   942  		// Set bitScan bits for all pointers.
   943  		hb |= hb << wordsPerBitmapByte
   944  		// First bitScan bit is always set since the type contains pointers.
   945  		hb |= bitScan
   946  		// Second bitScan bit needs to also be set if the third bitScan bit is set.
   947  		hb |= hb & (bitScan << (2 * heapBitsShift)) >> 1
   948  
   949  		// For h.shift > 1 heap bits cross a byte boundary and need to be written part
   950  		// to h.bitp and part to the next h.bitp.
   951  		switch h.shift {
   952  		case 0:
   953  			*h.bitp &^= mask3 << 0
   954  			*h.bitp |= hb << 0
   955  		case 1:
   956  			*h.bitp &^= mask3 << 1
   957  			*h.bitp |= hb << 1
   958  		case 2:
   959  			*h.bitp &^= mask2 << 2
   960  			*h.bitp |= (hb & mask2) << 2
   961  			// Two words written to the first byte.
   962  			// Advance two words to get to the next byte.
   963  			h = h.next().next()
   964  			*h.bitp &^= mask1
   965  			*h.bitp |= (hb >> 2) & mask1
   966  		case 3:
   967  			*h.bitp &^= mask1 << 3
   968  			*h.bitp |= (hb & mask1) << 3
   969  			// One word written to the first byte.
   970  			// Advance one word to get to the next byte.
   971  			h = h.next()
   972  			*h.bitp &^= mask2
   973  			*h.bitp |= (hb >> 1) & mask2
   974  		}
   975  		return
   976  	}
   977  
   978  	// Copy from 1-bit ptrmask into 2-bit bitmap.
   979  	// The basic approach is to use a single uintptr as a bit buffer,
   980  	// alternating between reloading the buffer and writing bitmap bytes.
   981  	// In general, one load can supply two bitmap byte writes.
   982  	// This is a lot of lines of code, but it compiles into relatively few
   983  	// machine instructions.
   984  
   985  	outOfPlace := false
   986  	if arenaIndex(x+size-1) != arenaIdx(h.arena) || (doubleCheck && fastrandn(2) == 0) {
   987  		// This object spans heap arenas, so the bitmap may be
   988  		// discontiguous. Unroll it into the object instead
   989  		// and then copy it out.
   990  		//
   991  		// In doubleCheck mode, we randomly do this anyway to
   992  		// stress test the bitmap copying path.
   993  		outOfPlace = true
   994  		h.bitp = (*uint8)(unsafe.Pointer(x))
   995  		h.last = nil
   996  	}
   997  
   998  	var (
   999  		// Ptrmask input.
  1000  		p     *byte   // last ptrmask byte read
  1001  		b     uintptr // ptrmask bits already loaded
  1002  		nb    uintptr // number of bits in b at next read
  1003  		endp  *byte   // final ptrmask byte to read (then repeat)
  1004  		endnb uintptr // number of valid bits in *endp
  1005  		pbits uintptr // alternate source of bits
  1006  
  1007  		// Heap bitmap output.
  1008  		w     uintptr // words processed
  1009  		nw    uintptr // number of words to process
  1010  		hbitp *byte   // next heap bitmap byte to write
  1011  		hb    uintptr // bits being prepared for *hbitp
  1012  	)
  1013  
  1014  	hbitp = h.bitp
  1015  
  1016  	// Handle GC program. Delayed until this part of the code
  1017  	// so that we can use the same double-checking mechanism
  1018  	// as the 1-bit case. Nothing above could have encountered
  1019  	// GC programs: the cases were all too small.
  1020  	if typ.kind&kindGCProg != 0 {
  1021  		heapBitsSetTypeGCProg(h, typ.ptrdata, typ.size, dataSize, size, addb(typ.gcdata, 4))
  1022  		if doubleCheck {
  1023  			// Double-check the heap bits written by GC program
  1024  			// by running the GC program to create a 1-bit pointer mask
  1025  			// and then jumping to the double-check code below.
  1026  			// This doesn't catch bugs shared between the 1-bit and 4-bit
  1027  			// GC program execution, but it does catch mistakes specific
  1028  			// to just one of those and bugs in heapBitsSetTypeGCProg's
  1029  			// implementation of arrays.
  1030  			lock(&debugPtrmask.lock)
  1031  			if debugPtrmask.data == nil {
  1032  				debugPtrmask.data = (*byte)(persistentalloc(1<<20, 1, &memstats.other_sys))
  1033  			}
  1034  			ptrmask = debugPtrmask.data
  1035  			runGCProg(addb(typ.gcdata, 4), nil, ptrmask, 1)
  1036  		}
  1037  		goto Phase4
  1038  	}
  1039  
  1040  	// Note about sizes:
  1041  	//
  1042  	// typ.size is the number of words in the object,
  1043  	// and typ.ptrdata is the number of words in the prefix
  1044  	// of the object that contains pointers. That is, the final
  1045  	// typ.size - typ.ptrdata words contain no pointers.
  1046  	// This allows optimization of a common pattern where
  1047  	// an object has a small header followed by a large scalar
  1048  	// buffer. If we know the pointers are over, we don't have
  1049  	// to scan the buffer's heap bitmap at all.
  1050  	// The 1-bit ptrmasks are sized to contain only bits for
  1051  	// the typ.ptrdata prefix, zero padded out to a full byte
  1052  	// of bitmap. This code sets nw (below) so that heap bitmap
  1053  	// bits are only written for the typ.ptrdata prefix; if there is
  1054  	// more room in the allocated object, the next heap bitmap
  1055  	// entry is a 00, indicating that there are no more pointers
  1056  	// to scan. So only the ptrmask for the ptrdata bytes is needed.
  1057  	//
  1058  	// Replicated copies are not as nice: if there is an array of
  1059  	// objects with scalar tails, all but the last tail does have to
  1060  	// be initialized, because there is no way to say "skip forward".
  1061  	// However, because of the possibility of a repeated type with
  1062  	// size not a multiple of 4 pointers (one heap bitmap byte),
  1063  	// the code already must handle the last ptrmask byte specially
  1064  	// by treating it as containing only the bits for endnb pointers,
  1065  	// where endnb <= 4. We represent large scalar tails that must
  1066  	// be expanded in the replication by setting endnb larger than 4.
  1067  	// This will have the effect of reading many bits out of b,
  1068  	// but once the real bits are shifted out, b will supply as many
  1069  	// zero bits as we try to read, which is exactly what we need.
  1070  
  1071  	p = ptrmask
  1072  	if typ.size < dataSize {
  1073  		// Filling in bits for an array of typ.
  1074  		// Set up for repetition of ptrmask during main loop.
  1075  		// Note that ptrmask describes only a prefix of
  1076  		const maxBits = goarch.PtrSize*8 - 7
  1077  		if typ.ptrdata/goarch.PtrSize <= maxBits {
  1078  			// Entire ptrmask fits in uintptr with room for a byte fragment.
  1079  			// Load into pbits and never read from ptrmask again.
  1080  			// This is especially important when the ptrmask has
  1081  			// fewer than 8 bits in it; otherwise the reload in the middle
  1082  			// of the Phase 2 loop would itself need to loop to gather
  1083  			// at least 8 bits.
  1084  
  1085  			// Accumulate ptrmask into b.
  1086  			// ptrmask is sized to describe only typ.ptrdata, but we record
  1087  			// it as describing typ.size bytes, since all the high bits are zero.
  1088  			nb = typ.ptrdata / goarch.PtrSize
  1089  			for i := uintptr(0); i < nb; i += 8 {
  1090  				b |= uintptr(*p) << i
  1091  				p = add1(p)
  1092  			}
  1093  			nb = typ.size / goarch.PtrSize
  1094  
  1095  			// Replicate ptrmask to fill entire pbits uintptr.
  1096  			// Doubling and truncating is fewer steps than
  1097  			// iterating by nb each time. (nb could be 1.)
  1098  			// Since we loaded typ.ptrdata/goarch.PtrSize bits
  1099  			// but are pretending to have typ.size/goarch.PtrSize,
  1100  			// there might be no replication necessary/possible.
  1101  			pbits = b
  1102  			endnb = nb
  1103  			if nb+nb <= maxBits {
  1104  				for endnb <= goarch.PtrSize*8 {
  1105  					pbits |= pbits << endnb
  1106  					endnb += endnb
  1107  				}
  1108  				// Truncate to a multiple of original ptrmask.
  1109  				// Because nb+nb <= maxBits, nb fits in a byte.
  1110  				// Byte division is cheaper than uintptr division.
  1111  				endnb = uintptr(maxBits/byte(nb)) * nb
  1112  				pbits &= 1<<endnb - 1
  1113  				b = pbits
  1114  				nb = endnb
  1115  			}
  1116  
  1117  			// Clear p and endp as sentinel for using pbits.
  1118  			// Checked during Phase 2 loop.
  1119  			p = nil
  1120  			endp = nil
  1121  		} else {
  1122  			// Ptrmask is larger. Read it multiple times.
  1123  			n := (typ.ptrdata/goarch.PtrSize+7)/8 - 1
  1124  			endp = addb(ptrmask, n)
  1125  			endnb = typ.size/goarch.PtrSize - n*8
  1126  		}
  1127  	}
  1128  	if p != nil {
  1129  		b = uintptr(*p)
  1130  		p = add1(p)
  1131  		nb = 8
  1132  	}
  1133  
  1134  	if typ.size == dataSize {
  1135  		// Single entry: can stop once we reach the non-pointer data.
  1136  		nw = typ.ptrdata / goarch.PtrSize
  1137  	} else {
  1138  		// Repeated instances of typ in an array.
  1139  		// Have to process first N-1 entries in full, but can stop
  1140  		// once we reach the non-pointer data in the final entry.
  1141  		nw = ((dataSize/typ.size-1)*typ.size + typ.ptrdata) / goarch.PtrSize
  1142  	}
  1143  	if nw == 0 {
  1144  		// No pointers! Caller was supposed to check.
  1145  		println("runtime: invalid type ", typ.string())
  1146  		throw("heapBitsSetType: called with non-pointer type")
  1147  		return
  1148  	}
  1149  
  1150  	// Phase 1: Special case for leading byte (shift==0) or half-byte (shift==2).
  1151  	// The leading byte is special because it contains the bits for word 1,
  1152  	// which does not have the scan bit set.
  1153  	// The leading half-byte is special because it's a half a byte,
  1154  	// so we have to be careful with the bits already there.
  1155  	switch {
  1156  	default:
  1157  		throw("heapBitsSetType: unexpected shift")
  1158  
  1159  	case h.shift == 0:
  1160  		// Ptrmask and heap bitmap are aligned.
  1161  		//
  1162  		// This is a fast path for small objects.
  1163  		//
  1164  		// The first byte we write out covers the first four
  1165  		// words of the object. The scan/dead bit on the first
  1166  		// word must be set to scan since there are pointers
  1167  		// somewhere in the object.
  1168  		// In all following words, we set the scan/dead
  1169  		// appropriately to indicate that the object continues
  1170  		// to the next 2-bit entry in the bitmap.
  1171  		//
  1172  		// We set four bits at a time here, but if the object
  1173  		// is fewer than four words, phase 3 will clear
  1174  		// unnecessary bits.
  1175  		hb = b & bitPointerAll
  1176  		hb |= bitScanAll
  1177  		if w += 4; w >= nw {
  1178  			goto Phase3
  1179  		}
  1180  		*hbitp = uint8(hb)
  1181  		hbitp = add1(hbitp)
  1182  		b >>= 4
  1183  		nb -= 4
  1184  
  1185  	case h.shift == 2:
  1186  		// Ptrmask and heap bitmap are misaligned.
  1187  		//
  1188  		// On 32 bit architectures only the 6-word object that corresponds
  1189  		// to a 24 bytes size class can start with h.shift of 2 here since
  1190  		// all other non 16 byte aligned size classes have been handled by
  1191  		// special code paths at the beginning of heapBitsSetType on 32 bit.
  1192  		//
  1193  		// Many size classes are only 16 byte aligned. On 64 bit architectures
  1194  		// this results in a heap bitmap position starting with a h.shift of 2.
  1195  		//
  1196  		// The bits for the first two words are in a byte shared
  1197  		// with another object, so we must be careful with the bits
  1198  		// already there.
  1199  		//
  1200  		// We took care of 1-word, 2-word, and 3-word objects above,
  1201  		// so this is at least a 6-word object.
  1202  		hb = (b & (bitPointer | bitPointer<<heapBitsShift)) << (2 * heapBitsShift)
  1203  		hb |= bitScan << (2 * heapBitsShift)
  1204  		if nw > 1 {
  1205  			hb |= bitScan << (3 * heapBitsShift)
  1206  		}
  1207  		b >>= 2
  1208  		nb -= 2
  1209  		*hbitp &^= uint8((bitPointer | bitScan | ((bitPointer | bitScan) << heapBitsShift)) << (2 * heapBitsShift))
  1210  		*hbitp |= uint8(hb)
  1211  		hbitp = add1(hbitp)
  1212  		if w += 2; w >= nw {
  1213  			// We know that there is more data, because we handled 2-word and 3-word objects above.
  1214  			// This must be at least a 6-word object. If we're out of pointer words,
  1215  			// mark no scan in next bitmap byte and finish.
  1216  			hb = 0
  1217  			w += 4
  1218  			goto Phase3
  1219  		}
  1220  	}
  1221  
  1222  	// Phase 2: Full bytes in bitmap, up to but not including write to last byte (full or partial) in bitmap.
  1223  	// The loop computes the bits for that last write but does not execute the write;
  1224  	// it leaves the bits in hb for processing by phase 3.
  1225  	// To avoid repeated adjustment of nb, we subtract out the 4 bits we're going to
  1226  	// use in the first half of the loop right now, and then we only adjust nb explicitly
  1227  	// if the 8 bits used by each iteration isn't balanced by 8 bits loaded mid-loop.
  1228  	nb -= 4
  1229  	for {
  1230  		// Emit bitmap byte.
  1231  		// b has at least nb+4 bits, with one exception:
  1232  		// if w+4 >= nw, then b has only nw-w bits,
  1233  		// but we'll stop at the break and then truncate
  1234  		// appropriately in Phase 3.
  1235  		hb = b & bitPointerAll
  1236  		hb |= bitScanAll
  1237  		if w += 4; w >= nw {
  1238  			break
  1239  		}
  1240  		*hbitp = uint8(hb)
  1241  		hbitp = add1(hbitp)
  1242  		b >>= 4
  1243  
  1244  		// Load more bits. b has nb right now.
  1245  		if p != endp {
  1246  			// Fast path: keep reading from ptrmask.
  1247  			// nb unmodified: we just loaded 8 bits,
  1248  			// and the next iteration will consume 8 bits,
  1249  			// leaving us with the same nb the next time we're here.
  1250  			if nb < 8 {
  1251  				b |= uintptr(*p) << nb
  1252  				p = add1(p)
  1253  			} else {
  1254  				// Reduce the number of bits in b.
  1255  				// This is important if we skipped
  1256  				// over a scalar tail, since nb could
  1257  				// be larger than the bit width of b.
  1258  				nb -= 8
  1259  			}
  1260  		} else if p == nil {
  1261  			// Almost as fast path: track bit count and refill from pbits.
  1262  			// For short repetitions.
  1263  			if nb < 8 {
  1264  				b |= pbits << nb
  1265  				nb += endnb
  1266  			}
  1267  			nb -= 8 // for next iteration
  1268  		} else {
  1269  			// Slow path: reached end of ptrmask.
  1270  			// Process final partial byte and rewind to start.
  1271  			b |= uintptr(*p) << nb
  1272  			nb += endnb
  1273  			if nb < 8 {
  1274  				b |= uintptr(*ptrmask) << nb
  1275  				p = add1(ptrmask)
  1276  			} else {
  1277  				nb -= 8
  1278  				p = ptrmask
  1279  			}
  1280  		}
  1281  
  1282  		// Emit bitmap byte.
  1283  		hb = b & bitPointerAll
  1284  		hb |= bitScanAll
  1285  		if w += 4; w >= nw {
  1286  			break
  1287  		}
  1288  		*hbitp = uint8(hb)
  1289  		hbitp = add1(hbitp)
  1290  		b >>= 4
  1291  	}
  1292  
  1293  Phase3:
  1294  	// Phase 3: Write last byte or partial byte and zero the rest of the bitmap entries.
  1295  	if w > nw {
  1296  		// Counting the 4 entries in hb not yet written to memory,
  1297  		// there are more entries than possible pointer slots.
  1298  		// Discard the excess entries (can't be more than 3).
  1299  		mask := uintptr(1)<<(4-(w-nw)) - 1
  1300  		hb &= mask | mask<<4 // apply mask to both pointer bits and scan bits
  1301  	}
  1302  
  1303  	// Change nw from counting possibly-pointer words to total words in allocation.
  1304  	nw = size / goarch.PtrSize
  1305  
  1306  	// Write whole bitmap bytes.
  1307  	// The first is hb, the rest are zero.
  1308  	if w <= nw {
  1309  		*hbitp = uint8(hb)
  1310  		hbitp = add1(hbitp)
  1311  		hb = 0 // for possible final half-byte below
  1312  		for w += 4; w <= nw; w += 4 {
  1313  			*hbitp = 0
  1314  			hbitp = add1(hbitp)
  1315  		}
  1316  	}
  1317  
  1318  	// Write final partial bitmap byte if any.
  1319  	// We know w > nw, or else we'd still be in the loop above.
  1320  	// It can be bigger only due to the 4 entries in hb that it counts.
  1321  	// If w == nw+4 then there's nothing left to do: we wrote all nw entries
  1322  	// and can discard the 4 sitting in hb.
  1323  	// But if w == nw+2, we need to write first two in hb.
  1324  	// The byte is shared with the next object, so be careful with
  1325  	// existing bits.
  1326  	if w == nw+2 {
  1327  		*hbitp = *hbitp&^(bitPointer|bitScan|(bitPointer|bitScan)<<heapBitsShift) | uint8(hb)
  1328  	}
  1329  
  1330  Phase4:
  1331  	// Phase 4: Copy unrolled bitmap to per-arena bitmaps, if necessary.
  1332  	if outOfPlace {
  1333  		// TODO: We could probably make this faster by
  1334  		// handling [x+dataSize, x+size) specially.
  1335  		h := heapBitsForAddr(x)
  1336  		// cnw is the number of heap words, or bit pairs
  1337  		// remaining (like nw above).
  1338  		cnw := size / goarch.PtrSize
  1339  		src := (*uint8)(unsafe.Pointer(x))
  1340  		// We know the first and last byte of the bitmap are
  1341  		// not the same, but it's still possible for small
  1342  		// objects span arenas, so it may share bitmap bytes
  1343  		// with neighboring objects.
  1344  		//
  1345  		// Handle the first byte specially if it's shared. See
  1346  		// Phase 1 for why this is the only special case we need.
  1347  		if doubleCheck {
  1348  			if !(h.shift == 0 || h.shift == 2) {
  1349  				print("x=", x, " size=", size, " cnw=", h.shift, "\n")
  1350  				throw("bad start shift")
  1351  			}
  1352  		}
  1353  		if h.shift == 2 {
  1354  			*h.bitp = *h.bitp&^((bitPointer|bitScan|(bitPointer|bitScan)<<heapBitsShift)<<(2*heapBitsShift)) | *src
  1355  			h = h.next().next()
  1356  			cnw -= 2
  1357  			src = addb(src, 1)
  1358  		}
  1359  		// We're now byte aligned. Copy out to per-arena
  1360  		// bitmaps until the last byte (which may again be
  1361  		// partial).
  1362  		for cnw >= 4 {
  1363  			// This loop processes four words at a time,
  1364  			// so round cnw down accordingly.
  1365  			hNext, words := h.forwardOrBoundary(cnw / 4 * 4)
  1366  
  1367  			// n is the number of bitmap bytes to copy.
  1368  			n := words / 4
  1369  			memmove(unsafe.Pointer(h.bitp), unsafe.Pointer(src), n)
  1370  			cnw -= words
  1371  			h = hNext
  1372  			src = addb(src, n)
  1373  		}
  1374  		if doubleCheck && h.shift != 0 {
  1375  			print("cnw=", cnw, " h.shift=", h.shift, "\n")
  1376  			throw("bad shift after block copy")
  1377  		}
  1378  		// Handle the last byte if it's shared.
  1379  		if cnw == 2 {
  1380  			*h.bitp = *h.bitp&^(bitPointer|bitScan|(bitPointer|bitScan)<<heapBitsShift) | *src
  1381  			src = addb(src, 1)
  1382  			h = h.next().next()
  1383  		}
  1384  		if doubleCheck {
  1385  			if uintptr(unsafe.Pointer(src)) > x+size {
  1386  				throw("copy exceeded object size")
  1387  			}
  1388  			if !(cnw == 0 || cnw == 2) {
  1389  				print("x=", x, " size=", size, " cnw=", cnw, "\n")
  1390  				throw("bad number of remaining words")
  1391  			}
  1392  			// Set up hbitp so doubleCheck code below can check it.
  1393  			hbitp = h.bitp
  1394  		}
  1395  		// Zero the object where we wrote the bitmap.
  1396  		memclrNoHeapPointers(unsafe.Pointer(x), uintptr(unsafe.Pointer(src))-x)
  1397  	}
  1398  
  1399  	// Double check the whole bitmap.
  1400  	if doubleCheck {
  1401  		// x+size may not point to the heap, so back up one
  1402  		// word and then advance it the way we do above.
  1403  		end := heapBitsForAddr(x + size - goarch.PtrSize)
  1404  		if outOfPlace {
  1405  			// In out-of-place copying, we just advance
  1406  			// using next.
  1407  			end = end.next()
  1408  		} else {
  1409  			// Don't use next because that may advance to
  1410  			// the next arena and the in-place logic
  1411  			// doesn't do that.
  1412  			end.shift += heapBitsShift
  1413  			if end.shift == 4*heapBitsShift {
  1414  				end.bitp, end.shift = add1(end.bitp), 0
  1415  			}
  1416  		}
  1417  		if typ.kind&kindGCProg == 0 && (hbitp != end.bitp || (w == nw+2) != (end.shift == 2)) {
  1418  			println("ended at wrong bitmap byte for", typ.string(), "x", dataSize/typ.size)
  1419  			print("typ.size=", typ.size, " typ.ptrdata=", typ.ptrdata, " dataSize=", dataSize, " size=", size, "\n")
  1420  			print("w=", w, " nw=", nw, " b=", hex(b), " nb=", nb, " hb=", hex(hb), "\n")
  1421  			h0 := heapBitsForAddr(x)
  1422  			print("initial bits h0.bitp=", h0.bitp, " h0.shift=", h0.shift, "\n")
  1423  			print("ended at hbitp=", hbitp, " but next starts at bitp=", end.bitp, " shift=", end.shift, "\n")
  1424  			throw("bad heapBitsSetType")
  1425  		}
  1426  
  1427  		// Double-check that bits to be written were written correctly.
  1428  		// Does not check that other bits were not written, unfortunately.
  1429  		h := heapBitsForAddr(x)
  1430  		nptr := typ.ptrdata / goarch.PtrSize
  1431  		ndata := typ.size / goarch.PtrSize
  1432  		count := dataSize / typ.size
  1433  		totalptr := ((count-1)*typ.size + typ.ptrdata) / goarch.PtrSize
  1434  		for i := uintptr(0); i < size/goarch.PtrSize; i++ {
  1435  			j := i % ndata
  1436  			var have, want uint8
  1437  			have = (*h.bitp >> h.shift) & (bitPointer | bitScan)
  1438  			if i >= totalptr {
  1439  				if typ.kind&kindGCProg != 0 && i < (totalptr+3)/4*4 {
  1440  					// heapBitsSetTypeGCProg always fills
  1441  					// in full nibbles of bitScan.
  1442  					want = bitScan
  1443  				}
  1444  			} else {
  1445  				if j < nptr && (*addb(ptrmask, j/8)>>(j%8))&1 != 0 {
  1446  					want |= bitPointer
  1447  				}
  1448  				want |= bitScan
  1449  			}
  1450  			if have != want {
  1451  				println("mismatch writing bits for", typ.string(), "x", dataSize/typ.size)
  1452  				print("typ.size=", typ.size, " typ.ptrdata=", typ.ptrdata, " dataSize=", dataSize, " size=", size, "\n")
  1453  				print("kindGCProg=", typ.kind&kindGCProg != 0, " outOfPlace=", outOfPlace, "\n")
  1454  				print("w=", w, " nw=", nw, " b=", hex(b), " nb=", nb, " hb=", hex(hb), "\n")
  1455  				h0 := heapBitsForAddr(x)
  1456  				print("initial bits h0.bitp=", h0.bitp, " h0.shift=", h0.shift, "\n")
  1457  				print("current bits h.bitp=", h.bitp, " h.shift=", h.shift, " *h.bitp=", hex(*h.bitp), "\n")
  1458  				print("ptrmask=", ptrmask, " p=", p, " endp=", endp, " endnb=", endnb, " pbits=", hex(pbits), " b=", hex(b), " nb=", nb, "\n")
  1459  				println("at word", i, "offset", i*goarch.PtrSize, "have", hex(have), "want", hex(want))
  1460  				if typ.kind&kindGCProg != 0 {
  1461  					println("GC program:")
  1462  					dumpGCProg(addb(typ.gcdata, 4))
  1463  				}
  1464  				throw("bad heapBitsSetType")
  1465  			}
  1466  			h = h.next()
  1467  		}
  1468  		if ptrmask == debugPtrmask.data {
  1469  			unlock(&debugPtrmask.lock)
  1470  		}
  1471  	}
  1472  }
  1473  
  1474  var debugPtrmask struct {
  1475  	lock mutex
  1476  	data *byte
  1477  }
  1478  
  1479  // heapBitsSetTypeGCProg implements heapBitsSetType using a GC program.
  1480  // progSize is the size of the memory described by the program.
  1481  // elemSize is the size of the element that the GC program describes (a prefix of).
  1482  // dataSize is the total size of the intended data, a multiple of elemSize.
  1483  // allocSize is the total size of the allocated memory.
  1484  //
  1485  // GC programs are only used for large allocations.
  1486  // heapBitsSetType requires that allocSize is a multiple of 4 words,
  1487  // so that the relevant bitmap bytes are not shared with surrounding
  1488  // objects.
  1489  func heapBitsSetTypeGCProg(h heapBits, progSize, elemSize, dataSize, allocSize uintptr, prog *byte) {
  1490  	if goarch.PtrSize == 8 && allocSize%(4*goarch.PtrSize) != 0 {
  1491  		// Alignment will be wrong.
  1492  		throw("heapBitsSetTypeGCProg: small allocation")
  1493  	}
  1494  	var totalBits uintptr
  1495  	if elemSize == dataSize {
  1496  		totalBits = runGCProg(prog, nil, h.bitp, 2)
  1497  		if totalBits*goarch.PtrSize != progSize {
  1498  			println("runtime: heapBitsSetTypeGCProg: total bits", totalBits, "but progSize", progSize)
  1499  			throw("heapBitsSetTypeGCProg: unexpected bit count")
  1500  		}
  1501  	} else {
  1502  		count := dataSize / elemSize
  1503  
  1504  		// Piece together program trailer to run after prog that does:
  1505  		//	literal(0)
  1506  		//	repeat(1, elemSize-progSize-1) // zeros to fill element size
  1507  		//	repeat(elemSize, count-1) // repeat that element for count
  1508  		// This zero-pads the data remaining in the first element and then
  1509  		// repeats that first element to fill the array.
  1510  		var trailer [40]byte // 3 varints (max 10 each) + some bytes
  1511  		i := 0
  1512  		if n := elemSize/goarch.PtrSize - progSize/goarch.PtrSize; n > 0 {
  1513  			// literal(0)
  1514  			trailer[i] = 0x01
  1515  			i++
  1516  			trailer[i] = 0
  1517  			i++
  1518  			if n > 1 {
  1519  				// repeat(1, n-1)
  1520  				trailer[i] = 0x81
  1521  				i++
  1522  				n--
  1523  				for ; n >= 0x80; n >>= 7 {
  1524  					trailer[i] = byte(n | 0x80)
  1525  					i++
  1526  				}
  1527  				trailer[i] = byte(n)
  1528  				i++
  1529  			}
  1530  		}
  1531  		// repeat(elemSize/ptrSize, count-1)
  1532  		trailer[i] = 0x80
  1533  		i++
  1534  		n := elemSize / goarch.PtrSize
  1535  		for ; n >= 0x80; n >>= 7 {
  1536  			trailer[i] = byte(n | 0x80)
  1537  			i++
  1538  		}
  1539  		trailer[i] = byte(n)
  1540  		i++
  1541  		n = count - 1
  1542  		for ; n >= 0x80; n >>= 7 {
  1543  			trailer[i] = byte(n | 0x80)
  1544  			i++
  1545  		}
  1546  		trailer[i] = byte(n)
  1547  		i++
  1548  		trailer[i] = 0
  1549  		i++
  1550  
  1551  		runGCProg(prog, &trailer[0], h.bitp, 2)
  1552  
  1553  		// Even though we filled in the full array just now,
  1554  		// record that we only filled in up to the ptrdata of the
  1555  		// last element. This will cause the code below to
  1556  		// memclr the dead section of the final array element,
  1557  		// so that scanobject can stop early in the final element.
  1558  		totalBits = (elemSize*(count-1) + progSize) / goarch.PtrSize
  1559  	}
  1560  	endProg := unsafe.Pointer(addb(h.bitp, (totalBits+3)/4))
  1561  	endAlloc := unsafe.Pointer(addb(h.bitp, allocSize/goarch.PtrSize/wordsPerBitmapByte))
  1562  	memclrNoHeapPointers(endProg, uintptr(endAlloc)-uintptr(endProg))
  1563  }
  1564  
  1565  // progToPointerMask returns the 1-bit pointer mask output by the GC program prog.
  1566  // size the size of the region described by prog, in bytes.
  1567  // The resulting bitvector will have no more than size/goarch.PtrSize bits.
  1568  func progToPointerMask(prog *byte, size uintptr) bitvector {
  1569  	n := (size/goarch.PtrSize + 7) / 8
  1570  	x := (*[1 << 30]byte)(persistentalloc(n+1, 1, &memstats.buckhash_sys))[:n+1]
  1571  	x[len(x)-1] = 0xa1 // overflow check sentinel
  1572  	n = runGCProg(prog, nil, &x[0], 1)
  1573  	if x[len(x)-1] != 0xa1 {
  1574  		throw("progToPointerMask: overflow")
  1575  	}
  1576  	return bitvector{int32(n), &x[0]}
  1577  }
  1578  
  1579  // Packed GC pointer bitmaps, aka GC programs.
  1580  //
  1581  // For large types containing arrays, the type information has a
  1582  // natural repetition that can be encoded to save space in the
  1583  // binary and in the memory representation of the type information.
  1584  //
  1585  // The encoding is a simple Lempel-Ziv style bytecode machine
  1586  // with the following instructions:
  1587  //
  1588  //	00000000: stop
  1589  //	0nnnnnnn: emit n bits copied from the next (n+7)/8 bytes
  1590  //	10000000 n c: repeat the previous n bits c times; n, c are varints
  1591  //	1nnnnnnn c: repeat the previous n bits c times; c is a varint
  1592  
  1593  // runGCProg executes the GC program prog, and then trailer if non-nil,
  1594  // writing to dst with entries of the given size.
  1595  // If size == 1, dst is a 1-bit pointer mask laid out moving forward from dst.
  1596  // If size == 2, dst is the 2-bit heap bitmap, and writes move backward
  1597  // starting at dst (because the heap bitmap does). In this case, the caller guarantees
  1598  // that only whole bytes in dst need to be written.
  1599  //
  1600  // runGCProg returns the number of 1- or 2-bit entries written to memory.
  1601  func runGCProg(prog, trailer, dst *byte, size int) uintptr {
  1602  	dstStart := dst
  1603  
  1604  	// Bits waiting to be written to memory.
  1605  	var bits uintptr
  1606  	var nbits uintptr
  1607  
  1608  	p := prog
  1609  Run:
  1610  	for {
  1611  		// Flush accumulated full bytes.
  1612  		// The rest of the loop assumes that nbits <= 7.
  1613  		for ; nbits >= 8; nbits -= 8 {
  1614  			if size == 1 {
  1615  				*dst = uint8(bits)
  1616  				dst = add1(dst)
  1617  				bits >>= 8
  1618  			} else {
  1619  				v := bits&bitPointerAll | bitScanAll
  1620  				*dst = uint8(v)
  1621  				dst = add1(dst)
  1622  				bits >>= 4
  1623  				v = bits&bitPointerAll | bitScanAll
  1624  				*dst = uint8(v)
  1625  				dst = add1(dst)
  1626  				bits >>= 4
  1627  			}
  1628  		}
  1629  
  1630  		// Process one instruction.
  1631  		inst := uintptr(*p)
  1632  		p = add1(p)
  1633  		n := inst & 0x7F
  1634  		if inst&0x80 == 0 {
  1635  			// Literal bits; n == 0 means end of program.
  1636  			if n == 0 {
  1637  				// Program is over; continue in trailer if present.
  1638  				if trailer != nil {
  1639  					p = trailer
  1640  					trailer = nil
  1641  					continue
  1642  				}
  1643  				break Run
  1644  			}
  1645  			nbyte := n / 8
  1646  			for i := uintptr(0); i < nbyte; i++ {
  1647  				bits |= uintptr(*p) << nbits
  1648  				p = add1(p)
  1649  				if size == 1 {
  1650  					*dst = uint8(bits)
  1651  					dst = add1(dst)
  1652  					bits >>= 8
  1653  				} else {
  1654  					v := bits&0xf | bitScanAll
  1655  					*dst = uint8(v)
  1656  					dst = add1(dst)
  1657  					bits >>= 4
  1658  					v = bits&0xf | bitScanAll
  1659  					*dst = uint8(v)
  1660  					dst = add1(dst)
  1661  					bits >>= 4
  1662  				}
  1663  			}
  1664  			if n %= 8; n > 0 {
  1665  				bits |= uintptr(*p) << nbits
  1666  				p = add1(p)
  1667  				nbits += n
  1668  			}
  1669  			continue Run
  1670  		}
  1671  
  1672  		// Repeat. If n == 0, it is encoded in a varint in the next bytes.
  1673  		if n == 0 {
  1674  			for off := uint(0); ; off += 7 {
  1675  				x := uintptr(*p)
  1676  				p = add1(p)
  1677  				n |= (x & 0x7F) << off
  1678  				if x&0x80 == 0 {
  1679  					break
  1680  				}
  1681  			}
  1682  		}
  1683  
  1684  		// Count is encoded in a varint in the next bytes.
  1685  		c := uintptr(0)
  1686  		for off := uint(0); ; off += 7 {
  1687  			x := uintptr(*p)
  1688  			p = add1(p)
  1689  			c |= (x & 0x7F) << off
  1690  			if x&0x80 == 0 {
  1691  				break
  1692  			}
  1693  		}
  1694  		c *= n // now total number of bits to copy
  1695  
  1696  		// If the number of bits being repeated is small, load them
  1697  		// into a register and use that register for the entire loop
  1698  		// instead of repeatedly reading from memory.
  1699  		// Handling fewer than 8 bits here makes the general loop simpler.
  1700  		// The cutoff is goarch.PtrSize*8 - 7 to guarantee that when we add
  1701  		// the pattern to a bit buffer holding at most 7 bits (a partial byte)
  1702  		// it will not overflow.
  1703  		src := dst
  1704  		const maxBits = goarch.PtrSize*8 - 7
  1705  		if n <= maxBits {
  1706  			// Start with bits in output buffer.
  1707  			pattern := bits
  1708  			npattern := nbits
  1709  
  1710  			// If we need more bits, fetch them from memory.
  1711  			if size == 1 {
  1712  				src = subtract1(src)
  1713  				for npattern < n {
  1714  					pattern <<= 8
  1715  					pattern |= uintptr(*src)
  1716  					src = subtract1(src)
  1717  					npattern += 8
  1718  				}
  1719  			} else {
  1720  				src = subtract1(src)
  1721  				for npattern < n {
  1722  					pattern <<= 4
  1723  					pattern |= uintptr(*src) & 0xf
  1724  					src = subtract1(src)
  1725  					npattern += 4
  1726  				}
  1727  			}
  1728  
  1729  			// We started with the whole bit output buffer,
  1730  			// and then we loaded bits from whole bytes.
  1731  			// Either way, we might now have too many instead of too few.
  1732  			// Discard the extra.
  1733  			if npattern > n {
  1734  				pattern >>= npattern - n
  1735  				npattern = n
  1736  			}
  1737  
  1738  			// Replicate pattern to at most maxBits.
  1739  			if npattern == 1 {
  1740  				// One bit being repeated.
  1741  				// If the bit is 1, make the pattern all 1s.
  1742  				// If the bit is 0, the pattern is already all 0s,
  1743  				// but we can claim that the number of bits
  1744  				// in the word is equal to the number we need (c),
  1745  				// because right shift of bits will zero fill.
  1746  				if pattern == 1 {
  1747  					pattern = 1<<maxBits - 1
  1748  					npattern = maxBits
  1749  				} else {
  1750  					npattern = c
  1751  				}
  1752  			} else {
  1753  				b := pattern
  1754  				nb := npattern
  1755  				if nb+nb <= maxBits {
  1756  					// Double pattern until the whole uintptr is filled.
  1757  					for nb <= goarch.PtrSize*8 {
  1758  						b |= b << nb
  1759  						nb += nb
  1760  					}
  1761  					// Trim away incomplete copy of original pattern in high bits.
  1762  					// TODO(rsc): Replace with table lookup or loop on systems without divide?
  1763  					nb = maxBits / npattern * npattern
  1764  					b &= 1<<nb - 1
  1765  					pattern = b
  1766  					npattern = nb
  1767  				}
  1768  			}
  1769  
  1770  			// Add pattern to bit buffer and flush bit buffer, c/npattern times.
  1771  			// Since pattern contains >8 bits, there will be full bytes to flush
  1772  			// on each iteration.
  1773  			for ; c >= npattern; c -= npattern {
  1774  				bits |= pattern << nbits
  1775  				nbits += npattern
  1776  				if size == 1 {
  1777  					for nbits >= 8 {
  1778  						*dst = uint8(bits)
  1779  						dst = add1(dst)
  1780  						bits >>= 8
  1781  						nbits -= 8
  1782  					}
  1783  				} else {
  1784  					for nbits >= 4 {
  1785  						*dst = uint8(bits&0xf | bitScanAll)
  1786  						dst = add1(dst)
  1787  						bits >>= 4
  1788  						nbits -= 4
  1789  					}
  1790  				}
  1791  			}
  1792  
  1793  			// Add final fragment to bit buffer.
  1794  			if c > 0 {
  1795  				pattern &= 1<<c - 1
  1796  				bits |= pattern << nbits
  1797  				nbits += c
  1798  			}
  1799  			continue Run
  1800  		}
  1801  
  1802  		// Repeat; n too large to fit in a register.
  1803  		// Since nbits <= 7, we know the first few bytes of repeated data
  1804  		// are already written to memory.
  1805  		off := n - nbits // n > nbits because n > maxBits and nbits <= 7
  1806  		if size == 1 {
  1807  			// Leading src fragment.
  1808  			src = subtractb(src, (off+7)/8)
  1809  			if frag := off & 7; frag != 0 {
  1810  				bits |= uintptr(*src) >> (8 - frag) << nbits
  1811  				src = add1(src)
  1812  				nbits += frag
  1813  				c -= frag
  1814  			}
  1815  			// Main loop: load one byte, write another.
  1816  			// The bits are rotating through the bit buffer.
  1817  			for i := c / 8; i > 0; i-- {
  1818  				bits |= uintptr(*src) << nbits
  1819  				src = add1(src)
  1820  				*dst = uint8(bits)
  1821  				dst = add1(dst)
  1822  				bits >>= 8
  1823  			}
  1824  			// Final src fragment.
  1825  			if c %= 8; c > 0 {
  1826  				bits |= (uintptr(*src) & (1<<c - 1)) << nbits
  1827  				nbits += c
  1828  			}
  1829  		} else {
  1830  			// Leading src fragment.
  1831  			src = subtractb(src, (off+3)/4)
  1832  			if frag := off & 3; frag != 0 {
  1833  				bits |= (uintptr(*src) & 0xf) >> (4 - frag) << nbits
  1834  				src = add1(src)
  1835  				nbits += frag
  1836  				c -= frag
  1837  			}
  1838  			// Main loop: load one byte, write another.
  1839  			// The bits are rotating through the bit buffer.
  1840  			for i := c / 4; i > 0; i-- {
  1841  				bits |= (uintptr(*src) & 0xf) << nbits
  1842  				src = add1(src)
  1843  				*dst = uint8(bits&0xf | bitScanAll)
  1844  				dst = add1(dst)
  1845  				bits >>= 4
  1846  			}
  1847  			// Final src fragment.
  1848  			if c %= 4; c > 0 {
  1849  				bits |= (uintptr(*src) & (1<<c - 1)) << nbits
  1850  				nbits += c
  1851  			}
  1852  		}
  1853  	}
  1854  
  1855  	// Write any final bits out, using full-byte writes, even for the final byte.
  1856  	var totalBits uintptr
  1857  	if size == 1 {
  1858  		totalBits = (uintptr(unsafe.Pointer(dst))-uintptr(unsafe.Pointer(dstStart)))*8 + nbits
  1859  		nbits += -nbits & 7
  1860  		for ; nbits > 0; nbits -= 8 {
  1861  			*dst = uint8(bits)
  1862  			dst = add1(dst)
  1863  			bits >>= 8
  1864  		}
  1865  	} else {
  1866  		totalBits = (uintptr(unsafe.Pointer(dst))-uintptr(unsafe.Pointer(dstStart)))*4 + nbits
  1867  		nbits += -nbits & 3
  1868  		for ; nbits > 0; nbits -= 4 {
  1869  			v := bits&0xf | bitScanAll
  1870  			*dst = uint8(v)
  1871  			dst = add1(dst)
  1872  			bits >>= 4
  1873  		}
  1874  	}
  1875  	return totalBits
  1876  }
  1877  
  1878  // materializeGCProg allocates space for the (1-bit) pointer bitmask
  1879  // for an object of size ptrdata.  Then it fills that space with the
  1880  // pointer bitmask specified by the program prog.
  1881  // The bitmask starts at s.startAddr.
  1882  // The result must be deallocated with dematerializeGCProg.
  1883  func materializeGCProg(ptrdata uintptr, prog *byte) *mspan {
  1884  	// Each word of ptrdata needs one bit in the bitmap.
  1885  	bitmapBytes := divRoundUp(ptrdata, 8*goarch.PtrSize)
  1886  	// Compute the number of pages needed for bitmapBytes.
  1887  	pages := divRoundUp(bitmapBytes, pageSize)
  1888  	s := mheap_.allocManual(pages, spanAllocPtrScalarBits)
  1889  	runGCProg(addb(prog, 4), nil, (*byte)(unsafe.Pointer(s.startAddr)), 1)
  1890  	return s
  1891  }
  1892  func dematerializeGCProg(s *mspan) {
  1893  	mheap_.freeManual(s, spanAllocPtrScalarBits)
  1894  }
  1895  
  1896  func dumpGCProg(p *byte) {
  1897  	nptr := 0
  1898  	for {
  1899  		x := *p
  1900  		p = add1(p)
  1901  		if x == 0 {
  1902  			print("\t", nptr, " end\n")
  1903  			break
  1904  		}
  1905  		if x&0x80 == 0 {
  1906  			print("\t", nptr, " lit ", x, ":")
  1907  			n := int(x+7) / 8
  1908  			for i := 0; i < n; i++ {
  1909  				print(" ", hex(*p))
  1910  				p = add1(p)
  1911  			}
  1912  			print("\n")
  1913  			nptr += int(x)
  1914  		} else {
  1915  			nbit := int(x &^ 0x80)
  1916  			if nbit == 0 {
  1917  				for nb := uint(0); ; nb += 7 {
  1918  					x := *p
  1919  					p = add1(p)
  1920  					nbit |= int(x&0x7f) << nb
  1921  					if x&0x80 == 0 {
  1922  						break
  1923  					}
  1924  				}
  1925  			}
  1926  			count := 0
  1927  			for nb := uint(0); ; nb += 7 {
  1928  				x := *p
  1929  				p = add1(p)
  1930  				count |= int(x&0x7f) << nb
  1931  				if x&0x80 == 0 {
  1932  					break
  1933  				}
  1934  			}
  1935  			print("\t", nptr, " repeat ", nbit, " × ", count, "\n")
  1936  			nptr += nbit * count
  1937  		}
  1938  	}
  1939  }
  1940  
  1941  // Testing.
  1942  
  1943  func getgcmaskcb(frame *stkframe, ctxt unsafe.Pointer) bool {
  1944  	target := (*stkframe)(ctxt)
  1945  	if frame.sp <= target.sp && target.sp < frame.varp {
  1946  		*target = *frame
  1947  		return false
  1948  	}
  1949  	return true
  1950  }
  1951  
  1952  // gcbits returns the GC type info for x, for testing.
  1953  // The result is the bitmap entries (0 or 1), one entry per byte.
  1954  //go:linkname reflect_gcbits reflect.gcbits
  1955  func reflect_gcbits(x any) []byte {
  1956  	ret := getgcmask(x)
  1957  	typ := (*ptrtype)(unsafe.Pointer(efaceOf(&x)._type)).elem
  1958  	nptr := typ.ptrdata / goarch.PtrSize
  1959  	for uintptr(len(ret)) > nptr && ret[len(ret)-1] == 0 {
  1960  		ret = ret[:len(ret)-1]
  1961  	}
  1962  	return ret
  1963  }
  1964  
  1965  // Returns GC type info for the pointer stored in ep for testing.
  1966  // If ep points to the stack, only static live information will be returned
  1967  // (i.e. not for objects which are only dynamically live stack objects).
  1968  func getgcmask(ep any) (mask []byte) {
  1969  	e := *efaceOf(&ep)
  1970  	p := e.data
  1971  	t := e._type
  1972  	// data or bss
  1973  	for _, datap := range activeModules() {
  1974  		// data
  1975  		if datap.data <= uintptr(p) && uintptr(p) < datap.edata {
  1976  			bitmap := datap.gcdatamask.bytedata
  1977  			n := (*ptrtype)(unsafe.Pointer(t)).elem.size
  1978  			mask = make([]byte, n/goarch.PtrSize)
  1979  			for i := uintptr(0); i < n; i += goarch.PtrSize {
  1980  				off := (uintptr(p) + i - datap.data) / goarch.PtrSize
  1981  				mask[i/goarch.PtrSize] = (*addb(bitmap, off/8) >> (off % 8)) & 1
  1982  			}
  1983  			return
  1984  		}
  1985  
  1986  		// bss
  1987  		if datap.bss <= uintptr(p) && uintptr(p) < datap.ebss {
  1988  			bitmap := datap.gcbssmask.bytedata
  1989  			n := (*ptrtype)(unsafe.Pointer(t)).elem.size
  1990  			mask = make([]byte, n/goarch.PtrSize)
  1991  			for i := uintptr(0); i < n; i += goarch.PtrSize {
  1992  				off := (uintptr(p) + i - datap.bss) / goarch.PtrSize
  1993  				mask[i/goarch.PtrSize] = (*addb(bitmap, off/8) >> (off % 8)) & 1
  1994  			}
  1995  			return
  1996  		}
  1997  	}
  1998  
  1999  	// heap
  2000  	if base, s, _ := findObject(uintptr(p), 0, 0); base != 0 {
  2001  		hbits := heapBitsForAddr(base)
  2002  		n := s.elemsize
  2003  		mask = make([]byte, n/goarch.PtrSize)
  2004  		for i := uintptr(0); i < n; i += goarch.PtrSize {
  2005  			if hbits.isPointer() {
  2006  				mask[i/goarch.PtrSize] = 1
  2007  			}
  2008  			if !hbits.morePointers() {
  2009  				mask = mask[:i/goarch.PtrSize]
  2010  				break
  2011  			}
  2012  			hbits = hbits.next()
  2013  		}
  2014  		return
  2015  	}
  2016  
  2017  	// stack
  2018  	if _g_ := getg(); _g_.m.curg.stack.lo <= uintptr(p) && uintptr(p) < _g_.m.curg.stack.hi {
  2019  		var frame stkframe
  2020  		frame.sp = uintptr(p)
  2021  		_g_ := getg()
  2022  		gentraceback(_g_.m.curg.sched.pc, _g_.m.curg.sched.sp, 0, _g_.m.curg, 0, nil, 1000, getgcmaskcb, noescape(unsafe.Pointer(&frame)), 0)
  2023  		if frame.fn.valid() {
  2024  			locals, _, _ := getStackMap(&frame, nil, false)
  2025  			if locals.n == 0 {
  2026  				return
  2027  			}
  2028  			size := uintptr(locals.n) * goarch.PtrSize
  2029  			n := (*ptrtype)(unsafe.Pointer(t)).elem.size
  2030  			mask = make([]byte, n/goarch.PtrSize)
  2031  			for i := uintptr(0); i < n; i += goarch.PtrSize {
  2032  				off := (uintptr(p) + i - frame.varp + size) / goarch.PtrSize
  2033  				mask[i/goarch.PtrSize] = locals.ptrbit(off)
  2034  			}
  2035  		}
  2036  		return
  2037  	}
  2038  
  2039  	// otherwise, not something the GC knows about.
  2040  	// possibly read-only data, like malloc(0).
  2041  	// must not have pointers
  2042  	return
  2043  }
  2044  

View as plain text