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.
5 // Garbage collector: type and heap bitmaps.
7 // Stack, data, and bss bitmaps
9 // Stack frames and global variables in the data and bss sections are described
10 // by 1-bit bitmaps in which 0 means uninteresting and 1 means live pointer
11 // to be visited during GC. The bits in each byte are consumed starting with
12 // the low bit: 1<<0, 1<<1, and so on.
16 // The allocated heap comes from a subset of the memory in the range [start, used),
17 // where start == mheap_.arena_start and used == mheap_.arena_used.
18 // The heap bitmap comprises 2 bits for each pointer-sized word in that range,
19 // stored in bytes indexed backward in memory from start.
20 // That is, the byte at address start-1 holds the 2-bit entries for the four words
21 // start through start+3*ptrSize, the byte at start-2 holds the entries for
22 // start+4*ptrSize through start+7*ptrSize, and so on.
24 // In each 2-bit entry, the lower bit holds the same information as in the 1-bit
25 // bitmaps: 0 means uninteresting and 1 means live pointer to be visited during GC.
26 // The meaning of the high bit depends on the position of the word being described
27 // in its allocated object. In all words *except* the second word, the
28 // high bit indicates that the object is still being described. In
29 // these words, if a bit pair with a high bit 0 is encountered, the
30 // low bit can also be assumed to be 0, and the object description is
31 // over. This 00 is called the ``dead'' encoding: it signals that the
32 // rest of the words in the object are uninteresting to the garbage
35 // In the second word, the high bit is the GC ``checkmarked'' bit (see below).
37 // The 2-bit entries are split when written into the byte, so that the top half
38 // of the byte contains 4 high bits and the bottom half contains 4 low (pointer)
40 // This form allows a copy from the 1-bit to the 4-bit form to keep the
41 // pointer bits contiguous, instead of having to space them out.
43 // The code makes use of the fact that the zero value for a heap bitmap
44 // has no live pointer bit set and is (depending on position), not used,
45 // not checkmarked, and is the dead encoding.
46 // These properties must be preserved when modifying the encoding.
48 // The bitmap for noscan spans is not maintained. Code must ensure
49 // that an object is scannable before consulting its bitmap by
50 // checking either the noscan bit in the span or by consulting its
51 // type's information.
55 // In a concurrent garbage collector, one worries about failing to mark
56 // a live object due to mutations without write barriers or bugs in the
57 // collector implementation. As a sanity check, the GC has a 'checkmark'
58 // mode that retraverses the object graph with the world stopped, to make
59 // sure that everything that should be marked is marked.
60 // In checkmark mode, in the heap bitmap, the high bit of the 2-bit entry
61 // for the second word of the object holds the checkmark bit.
62 // When not in checkmark mode, this bit is set to 1.
64 // The smallest possible allocation is 8 bytes. On a 32-bit machine, that
65 // means every allocated object has two words, so there is room for the
66 // checkmark bit. On a 64-bit machine, however, the 8-byte allocation is
67 // just one word, so the second bit pair is not available for encoding the
68 // checkmark. However, because non-pointer allocations are combined
69 // into larger 16-byte (maxTinySize) allocations, a plain 8-byte allocation
70 // must be a pointer, so the type bit in the first word is not actually needed.
71 // It is still used in general, except in checkmark the type bit is repurposed
72 // as the checkmark bit and then reinitialized (to 1) as the type bit when
79 "runtime/internal/atomic"
80 "runtime/internal/sys"
88 heapBitsShift
= 1 // shift offset between successive bitPointer or bitScan entries
89 heapBitmapScale
= sys
.PtrSize
* (8 / 2) // number of data bytes described by one heap bitmap byte
91 // all scan/pointer bits in a byte
92 bitScanAll
= bitScan | bitScan
<<heapBitsShift | bitScan
<<(2*heapBitsShift
) | bitScan
<<(3*heapBitsShift
)
93 bitPointerAll
= bitPointer | bitPointer
<<heapBitsShift | bitPointer
<<(2*heapBitsShift
) | bitPointer
<<(3*heapBitsShift
)
96 // addb returns the byte pointer p+n.
99 func addb(p
*byte, n
uintptr) *byte {
100 // Note: wrote out full expression instead of calling add(p, n)
101 // to reduce the number of temporaries generated by the
102 // compiler for this trivial expression during inlining.
103 return (*byte)(unsafe
.Pointer(uintptr(unsafe
.Pointer(p
)) + n
))
106 // subtractb returns the byte pointer p-n.
107 // subtractb is typically used when traversing the pointer tables referred to by hbits
108 // which are arranged in reverse order.
111 func subtractb(p
*byte, n
uintptr) *byte {
112 // Note: wrote out full expression instead of calling add(p, -n)
113 // to reduce the number of temporaries generated by the
114 // compiler for this trivial expression during inlining.
115 return (*byte)(unsafe
.Pointer(uintptr(unsafe
.Pointer(p
)) - n
))
118 // add1 returns the byte pointer p+1.
121 func add1(p
*byte) *byte {
122 // Note: wrote out full expression instead of calling addb(p, 1)
123 // to reduce the number of temporaries generated by the
124 // compiler for this trivial expression during inlining.
125 return (*byte)(unsafe
.Pointer(uintptr(unsafe
.Pointer(p
)) + 1))
128 // subtract1 returns the byte pointer p-1.
129 // subtract1 is typically used when traversing the pointer tables referred to by hbits
130 // which are arranged in reverse order.
133 // nosplit because it is used during write barriers and must not be preempted.
135 func subtract1(p
*byte) *byte {
136 // Note: wrote out full expression instead of calling subtractb(p, 1)
137 // to reduce the number of temporaries generated by the
138 // compiler for this trivial expression during inlining.
139 return (*byte)(unsafe
.Pointer(uintptr(unsafe
.Pointer(p
)) - 1))
142 // mapBits maps any additional bitmap memory needed for the new arena memory.
144 // Don't call this directly. Call mheap.setArenaUsed.
147 func (h
*mheap
) mapBits(arena_used
uintptr) {
148 // Caller has added extra mappings to the arena.
149 // Add extra mappings of bitmap words as needed.
150 // We allocate extra bitmap pieces in chunks of bitmapChunk.
151 const bitmapChunk
= 8192
153 n
:= (arena_used
- mheap_
.arena_start
) / heapBitmapScale
154 n
= round(n
, bitmapChunk
)
155 n
= round(n
, physPageSize
)
156 if h
.bitmap_mapped
>= n
{
160 sysMap(unsafe
.Pointer(h
.bitmap
-n
), n
-h
.bitmap_mapped
, h
.arena_reserved
, &memstats
.gc_sys
)
164 // heapBits provides access to the bitmap bits for a single heap word.
165 // The methods on heapBits take value receivers so that the compiler
166 // can more easily inline calls to those methods and registerize the
167 // struct fields independently.
168 type heapBits
struct {
173 // markBits provides access to the mark bit for an object in the heap.
174 // bytep points to the byte holding the mark bit.
175 // mask is a byte with a single bit set that can be &ed with *bytep
176 // to see if the bit has been set.
177 // *m.byte&m.mask != 0 indicates the mark bit is set.
178 // index can be used along with span information to generate
179 // the address of the object in the heap.
180 // We maintain one set of mark bits for allocation and one for
182 type markBits
struct {
189 func (s
*mspan
) allocBitsForIndex(allocBitIndex
uintptr) markBits
{
190 bytep
, mask
:= s
.allocBits
.bitp(allocBitIndex
)
191 return markBits
{bytep
, mask
, allocBitIndex
}
194 // refillaCache takes 8 bytes s.allocBits starting at whichByte
195 // and negates them so that ctz (count trailing zeros) instructions
196 // can be used. It then places these 8 bytes into the cached 64 bit
198 func (s
*mspan
) refillAllocCache(whichByte
uintptr) {
199 bytes
:= (*[8]uint8)(unsafe
.Pointer(s
.allocBits
.bytep(whichByte
)))
201 aCache |
= uint64(bytes
[0])
202 aCache |
= uint64(bytes
[1]) << (1 * 8)
203 aCache |
= uint64(bytes
[2]) << (2 * 8)
204 aCache |
= uint64(bytes
[3]) << (3 * 8)
205 aCache |
= uint64(bytes
[4]) << (4 * 8)
206 aCache |
= uint64(bytes
[5]) << (5 * 8)
207 aCache |
= uint64(bytes
[6]) << (6 * 8)
208 aCache |
= uint64(bytes
[7]) << (7 * 8)
209 s
.allocCache
= ^aCache
212 // nextFreeIndex returns the index of the next free object in s at
213 // or after s.freeindex.
214 // There are hardware instructions that can be used to make this
215 // faster if profiling warrants it.
216 func (s
*mspan
) nextFreeIndex() uintptr {
217 sfreeindex
:= s
.freeindex
219 if sfreeindex
== snelems
{
222 if sfreeindex
> snelems
{
223 throw("s.freeindex > s.nelems")
226 aCache
:= s
.allocCache
228 bitIndex
:= sys
.Ctz64(aCache
)
230 // Move index to start of next cached bits.
231 sfreeindex
= (sfreeindex
+ 64) &^ (64 - 1)
232 if sfreeindex
>= snelems
{
233 s
.freeindex
= snelems
236 whichByte
:= sfreeindex
/ 8
237 // Refill s.allocCache with the next 64 alloc bits.
238 s
.refillAllocCache(whichByte
)
239 aCache
= s
.allocCache
240 bitIndex
= sys
.Ctz64(aCache
)
241 // nothing available in cached bits
242 // grab the next 8 bytes and try again.
244 result
:= sfreeindex
+ uintptr(bitIndex
)
245 if result
>= snelems
{
246 s
.freeindex
= snelems
250 s
.allocCache
>>= uint(bitIndex
+ 1)
251 sfreeindex
= result
+ 1
253 if sfreeindex%64
== 0 && sfreeindex
!= snelems
{
254 // We just incremented s.freeindex so it isn't 0.
255 // As each 1 in s.allocCache was encountered and used for allocation
256 // it was shifted away. At this point s.allocCache contains all 0s.
257 // Refill s.allocCache so that it corresponds
258 // to the bits at s.allocBits starting at s.freeindex.
259 whichByte
:= sfreeindex
/ 8
260 s
.refillAllocCache(whichByte
)
262 s
.freeindex
= sfreeindex
266 // isFree returns whether the index'th object in s is unallocated.
267 func (s
*mspan
) isFree(index
uintptr) bool {
268 if index
< s
.freeindex
{
271 bytep
, mask
:= s
.allocBits
.bitp(index
)
272 return *bytep
&mask
== 0
275 func (s
*mspan
) objIndex(p
uintptr) uintptr {
276 byteOffset
:= p
- s
.base()
281 // s.baseMask is 0, elemsize is a power of two, so shift by s.divShift
282 return byteOffset
>> s
.divShift
284 return uintptr(((uint64(byteOffset
) >> s
.divShift
) * uint64(s
.divMul
)) >> s
.divShift2
)
287 func markBitsForAddr(p
uintptr) markBits
{
289 objIndex
:= s
.objIndex(p
)
290 return s
.markBitsForIndex(objIndex
)
293 func (s
*mspan
) markBitsForIndex(objIndex
uintptr) markBits
{
294 bytep
, mask
:= s
.gcmarkBits
.bitp(objIndex
)
295 return markBits
{bytep
, mask
, objIndex
}
298 func (s
*mspan
) markBitsForBase() markBits
{
299 return markBits
{(*uint8)(s
.gcmarkBits
), uint8(1), 0}
302 // isMarked reports whether mark bit m is set.
303 func (m markBits
) isMarked() bool {
304 return *m
.bytep
&m
.mask
!= 0
307 // setMarked sets the marked bit in the markbits, atomically. Some compilers
308 // are not able to inline atomic.Or8 function so if it appears as a hot spot consider
309 // inlining it manually.
310 func (m markBits
) setMarked() {
311 // Might be racing with other updates, so use atomic update always.
312 // We used to be clever here and use a non-atomic update in certain
313 // cases, but it's not worth the risk.
314 atomic
.Or8(m
.bytep
, m
.mask
)
317 // setMarkedNonAtomic sets the marked bit in the markbits, non-atomically.
318 func (m markBits
) setMarkedNonAtomic() {
322 // clearMarked clears the marked bit in the markbits, atomically.
323 func (m markBits
) clearMarked() {
324 // Might be racing with other updates, so use atomic update always.
325 // We used to be clever here and use a non-atomic update in certain
326 // cases, but it's not worth the risk.
327 atomic
.And8(m
.bytep
, ^m
.mask
)
330 // markBitsForSpan returns the markBits for the span base address base.
331 func markBitsForSpan(base
uintptr) (mbits markBits
) {
332 if base
< mheap_
.arena_start || base
>= mheap_
.arena_used
{
333 throw("markBitsForSpan: base out of range")
335 mbits
= markBitsForAddr(base
)
337 throw("markBitsForSpan: unaligned start")
342 // advance advances the markBits to the next object in the span.
343 func (m
*markBits
) advance() {
345 m
.bytep
= (*uint8)(unsafe
.Pointer(uintptr(unsafe
.Pointer(m
.bytep
)) + 1))
353 // heapBitsForAddr returns the heapBits for the address addr.
354 // The caller must have already checked that addr is in the range [mheap_.arena_start, mheap_.arena_used).
356 // nosplit because it is used during write barriers and must not be preempted.
358 func heapBitsForAddr(addr
uintptr) heapBits
{
359 // 2 bits per work, 4 pairs per byte, and a mask is hard coded.
360 off
:= (addr
- mheap_
.arena_start
) / sys
.PtrSize
361 return heapBits
{(*uint8)(unsafe
.Pointer(mheap_
.bitmap
- off
/4 - 1)), uint32(off
& 3)}
364 // heapBitsForSpan returns the heapBits for the span base address base.
365 func heapBitsForSpan(base
uintptr) (hbits heapBits
) {
366 if base
< mheap_
.arena_start || base
>= mheap_
.arena_used
{
367 print("runtime: base ", hex(base
), " not in range [", hex(mheap_
.arena_start
), ",", hex(mheap_
.arena_used
), ")\n")
368 throw("heapBitsForSpan: base out of range")
370 return heapBitsForAddr(base
)
373 // heapBitsForObject returns the base address for the heap object
374 // containing the address p, the heapBits for base,
375 // the object's span, and of the index of the object in s.
376 // If p does not point into a heap object,
378 // otherwise return the base of the object.
380 // For gccgo, the forStack parameter is true if the value came from the stack.
381 // The stack is collected conservatively and may contain invalid pointers.
383 // refBase and refOff optionally give the base address of the object
384 // in which the pointer p was found and the byte offset at which it
385 // was found. These are used for error reporting.
386 func heapBitsForObject(p
, refBase
, refOff
uintptr, forStack
bool) (base
uintptr, hbits heapBits
, s
*mspan
, objIndex
uintptr) {
387 arenaStart
:= mheap_
.arena_start
388 if p
< arenaStart || p
>= mheap_
.arena_used
{
391 off
:= p
- arenaStart
392 idx
:= off
>> _PageShift
393 // p points into the heap, but possibly to the middle of an object.
394 // Consult the span table to find the block beginning.
395 s
= mheap_
.spans
[idx
]
396 if s
== nil || p
< s
.base() || p
>= s
.limit || s
.state
!= mSpanInUse
{
397 if s
== nil || s
.state
== _MSpanManual || forStack
{
398 // If s is nil, the virtual address has never been part of the heap.
399 // This pointer may be to some mmap'd region, so we allow it.
400 // Pointers into stacks are also ok, the runtime manages these explicitly.
404 // The following ensures that we are rigorous about what data
405 // structures hold valid pointers.
406 if debug
.invalidptr
!= 0 {
407 // Typically this indicates an incorrect use
408 // of unsafe or cgo to store a bad pointer in
409 // the Go heap. It may also indicate a runtime
412 // TODO(austin): We could be more aggressive
413 // and detect pointers to unallocated objects
414 // in allocated spans.
416 print("runtime: pointer ", hex(p
))
417 if s
.state
!= mSpanInUse
{
418 print(" to unallocated span")
420 print(" to unused region of span")
422 print(" idx=", hex(idx
), " span.base()=", hex(s
.base()), " span.limit=", hex(s
.limit
), " span.state=", s
.state
, "\n")
424 print("runtime: found in object at *(", hex(refBase
), "+", hex(refOff
), ")\n")
425 gcDumpObject("object", refBase
, refOff
)
427 getg().m
.traceback
= 2
428 throw("found bad pointer in Go heap (incorrect use of unsafe or cgo?)")
434 // A span can be entered in mheap_.spans, and be set
435 // to mSpanInUse, before it is fully initialized.
436 // All we need in practice is allocBits and gcmarkBits,
437 // so make sure they are set.
438 if s
.allocBits
== nil || s
.gcmarkBits
== nil {
443 // If this span holds object of a power of 2 size, just mask off the bits to
444 // the interior of the object. Otherwise use the size to get the base.
446 // optimize for power of 2 sized objects.
448 base
= base
+ (p
-base
)&uintptr(s
.baseMask
)
449 objIndex
= (base
- s
.base()) >> s
.divShift
450 // base = p & s.baseMask is faster for small spans,
451 // but doesn't work for large spans.
452 // Overall, it's faster to use the more general computation above.
455 if p
-base
>= s
.elemsize
{
456 // n := (p - base) / s.elemsize, using division by multiplication
457 objIndex
= uintptr(p
-base
) >> s
.divShift
* uintptr(s
.divMul
) >> s
.divShift2
458 base
+= objIndex
* s
.elemsize
461 // Now that we know the actual base, compute heapBits to return to caller.
462 hbits
= heapBitsForAddr(base
)
466 // next returns the heapBits describing the next pointer-sized word in memory.
467 // That is, if h describes address p, h.next() describes p+ptrSize.
468 // Note that next does not modify h. The caller must record the result.
470 // nosplit because it is used during write barriers and must not be preempted.
472 func (h heapBits
) next() heapBits
{
473 if h
.shift
< 3*heapBitsShift
{
474 return heapBits
{h
.bitp
, h
.shift
+ heapBitsShift
}
476 return heapBits
{subtract1(h
.bitp
), 0}
479 // forward returns the heapBits describing n pointer-sized words ahead of h in memory.
480 // That is, if h describes address p, h.forward(n) describes p+n*ptrSize.
481 // h.forward(1) is equivalent to h.next(), just slower.
482 // Note that forward does not modify h. The caller must record the result.
483 // bits returns the heap bits for the current word.
484 func (h heapBits
) forward(n
uintptr) heapBits
{
485 n
+= uintptr(h
.shift
) / heapBitsShift
486 return heapBits
{subtractb(h
.bitp
, n
/4), uint32(n%4
) * heapBitsShift
}
489 // The caller can test morePointers and isPointer by &-ing with bitScan and bitPointer.
490 // The result includes in its higher bits the bits for subsequent words
491 // described by the same bitmap byte.
493 // nosplit because it is used during write barriers and must not be preempted.
495 func (h heapBits
) bits() uint32 {
496 // The (shift & 31) eliminates a test and conditional branch
497 // from the generated code.
498 return uint32(*h
.bitp
) >> (h
.shift
& 31)
501 // morePointers returns true if this word and all remaining words in this object
503 // h must not describe the second word of the object.
504 func (h heapBits
) morePointers() bool {
505 return h
.bits()&bitScan
!= 0
508 // isPointer reports whether the heap bits describe a pointer word.
510 // nosplit because it is used during write barriers and must not be preempted.
512 func (h heapBits
) isPointer() bool {
513 return h
.bits()&bitPointer
!= 0
516 // isCheckmarked reports whether the heap bits have the checkmarked bit set.
517 // It must be told how large the object at h is, because the encoding of the
518 // checkmark bit varies by size.
519 // h must describe the initial word of the object.
520 func (h heapBits
) isCheckmarked(size
uintptr) bool {
521 if size
== sys
.PtrSize
{
522 return (*h
.bitp
>>h
.shift
)&bitPointer
!= 0
524 // All multiword objects are 2-word aligned,
525 // so we know that the initial word's 2-bit pair
526 // and the second word's 2-bit pair are in the
527 // same heap bitmap byte, *h.bitp.
528 return (*h
.bitp
>>(heapBitsShift
+h
.shift
))&bitScan
!= 0
531 // setCheckmarked sets the checkmarked bit.
532 // It must be told how large the object at h is, because the encoding of the
533 // checkmark bit varies by size.
534 // h must describe the initial word of the object.
535 func (h heapBits
) setCheckmarked(size
uintptr) {
536 if size
== sys
.PtrSize
{
537 atomic
.Or8(h
.bitp
, bitPointer
<<h
.shift
)
540 atomic
.Or8(h
.bitp
, bitScan
<<(heapBitsShift
+h
.shift
))
543 // bulkBarrierPreWrite executes a write barrier
544 // for every pointer slot in the memory range [src, src+size),
545 // using pointer/scalar information from [dst, dst+size).
546 // This executes the write barriers necessary before a memmove.
547 // src, dst, and size must be pointer-aligned.
548 // The range [dst, dst+size) must lie within a single object.
549 // It does not perform the actual writes.
551 // As a special case, src == 0 indicates that this is being used for a
552 // memclr. bulkBarrierPreWrite will pass 0 for the src of each write
555 // Callers should call bulkBarrierPreWrite immediately before
556 // calling memmove(dst, src, size). This function is marked nosplit
557 // to avoid being preempted; the GC must not stop the goroutine
558 // between the memmove and the execution of the barriers.
559 // The caller is also responsible for cgo pointer checks if this
560 // may be writing Go pointers into non-Go memory.
562 // The pointer bitmap is not maintained for allocations containing
563 // no pointers at all; any caller of bulkBarrierPreWrite must first
564 // make sure the underlying allocation contains pointers, usually
565 // by checking typ.kind&kindNoPointers.
568 func bulkBarrierPreWrite(dst
, src
, size
uintptr) {
569 if (dst|src|size
)&(sys
.PtrSize
-1) != 0 {
570 throw("bulkBarrierPreWrite: unaligned arguments")
572 if !writeBarrier
.needed
{
576 // If dst is a global, use the data or BSS bitmaps to
577 // execute write barriers.
580 for i
:= 0; i
< roots
.count
; i
++ {
582 addr
:= uintptr(pr
.decl
)
583 if addr
<= dst
&& dst
< addr
+pr
.size
{
584 if dst
< addr
+pr
.ptrdata
{
585 bulkBarrierBitmap(dst
, src
, size
, dst
-addr
, pr
.gcdata
)
595 buf
:= &getg().m
.p
.ptr().wbBuf
596 h
:= heapBitsForAddr(dst
)
598 for i
:= uintptr(0); i
< size
; i
+= sys
.PtrSize
{
600 dstx
:= (*uintptr)(unsafe
.Pointer(dst
+ i
))
601 if !buf
.putFast(*dstx
, 0) {
608 for i
:= uintptr(0); i
< size
; i
+= sys
.PtrSize
{
610 dstx
:= (*uintptr)(unsafe
.Pointer(dst
+ i
))
611 srcx
:= (*uintptr)(unsafe
.Pointer(src
+ i
))
612 if !buf
.putFast(*dstx
, *srcx
) {
621 // bulkBarrierBitmap executes write barriers for copying from [src,
622 // src+size) to [dst, dst+size) using a 1-bit pointer bitmap. src is
623 // assumed to start maskOffset bytes into the data covered by the
624 // bitmap in bits (which may not be a multiple of 8).
626 // This is used by bulkBarrierPreWrite for writes to data and BSS.
629 func bulkBarrierBitmap(dst
, src
, size
, maskOffset
uintptr, bits
*uint8) {
630 word
:= maskOffset
/ sys
.PtrSize
631 bits
= addb(bits
, word
/8)
632 mask
:= uint8(1) << (word
% 8)
634 buf
:= &getg().m
.p
.ptr().wbBuf
635 for i
:= uintptr(0); i
< size
; i
+= sys
.PtrSize
{
646 dstx
:= (*uintptr)(unsafe
.Pointer(dst
+ i
))
648 if !buf
.putFast(*dstx
, 0) {
652 srcx
:= (*uintptr)(unsafe
.Pointer(src
+ i
))
653 if !buf
.putFast(*dstx
, *srcx
) {
662 // typeBitsBulkBarrier executes writebarrierptr_prewrite for every
663 // pointer that would be copied from [src, src+size) to [dst,
664 // dst+size) by a memmove using the type bitmap to locate those
667 // The type typ must correspond exactly to [src, src+size) and [dst, dst+size).
668 // dst, src, and size must be pointer-aligned.
669 // The type typ must have a plain bitmap, not a GC program.
670 // The only use of this function is in channel sends, and the
671 // 64 kB channel element limit takes care of this for us.
673 // Must not be preempted because it typically runs right before memmove,
674 // and the GC must observe them as an atomic action.
677 func typeBitsBulkBarrier(typ
*_type
, dst
, src
, size
uintptr) {
679 throw("runtime: typeBitsBulkBarrier without type")
681 if typ
.size
!= size
{
682 println("runtime: typeBitsBulkBarrier with type ", *typ
.string, " of size ", typ
.size
, " but memory size", size
)
683 throw("runtime: invalid typeBitsBulkBarrier")
685 if typ
.kind
&kindGCProg
!= 0 {
686 println("runtime: typeBitsBulkBarrier with type ", *typ
.string, " with GC prog")
687 throw("runtime: invalid typeBitsBulkBarrier")
689 if !writeBarrier
.needed
{
692 ptrmask
:= typ
.gcdata
694 for i
:= uintptr(0); i
< typ
.ptrdata
; i
+= sys
.PtrSize
{
695 if i
&(sys
.PtrSize
*8-1) == 0 {
696 bits
= uint32(*ptrmask
)
697 ptrmask
= addb(ptrmask
, 1)
702 dstx
:= (*uintptr)(unsafe
.Pointer(dst
+ i
))
703 srcx
:= (*uintptr)(unsafe
.Pointer(src
+ i
))
704 writebarrierptr_prewrite(dstx
, *srcx
)
709 // The methods operating on spans all require that h has been returned
710 // by heapBitsForSpan and that size, n, total are the span layout description
711 // returned by the mspan's layout method.
712 // If total > size*n, it means that there is extra leftover memory in the span,
713 // usually due to rounding.
715 // TODO(rsc): Perhaps introduce a different heapBitsSpan type.
717 // initSpan initializes the heap bitmap for a span.
718 // It clears all checkmark bits.
719 // If this is a span of pointer-sized objects, it initializes all
720 // words to pointer/scan.
721 // Otherwise, it initializes all words to scalar/dead.
722 func (h heapBits
) initSpan(s
*mspan
) {
723 size
, n
, total
:= s
.layout()
725 // Init the markbit structures
727 s
.allocCache
= ^uint64(0) // all 1s indicating all free.
731 s
.gcmarkBits
= newMarkBits(s
.nelems
)
732 s
.allocBits
= newAllocBits(s
.nelems
)
734 // Clear bits corresponding to objects.
735 if total%heapBitmapScale
!= 0 {
736 throw("initSpan: unaligned length")
738 nbyte
:= total
/ heapBitmapScale
739 if sys
.PtrSize
== 8 && size
== sys
.PtrSize
{
741 bitp
:= subtractb(end
, nbyte
-1)
743 *bitp
= bitPointerAll | bitScanAll
751 memclrNoHeapPointers(unsafe
.Pointer(subtractb(h
.bitp
, nbyte
-1)), nbyte
)
754 // initCheckmarkSpan initializes a span for being checkmarked.
755 // It clears the checkmark bits, which are set to 1 in normal operation.
756 func (h heapBits
) initCheckmarkSpan(size
, n
, total
uintptr) {
757 // The ptrSize == 8 is a compile-time constant false on 32-bit and eliminates this code entirely.
758 if sys
.PtrSize
== 8 && size
== sys
.PtrSize
{
759 // Checkmark bit is type bit, bottom bit of every 2-bit entry.
760 // Only possible on 64-bit system, since minimum size is 8.
761 // Must clear type bit (checkmark bit) of every word.
762 // The type bit is the lower of every two-bit pair.
764 for i
:= uintptr(0); i
< n
; i
+= 4 {
765 *bitp
&^= bitPointerAll
766 bitp
= subtract1(bitp
)
770 for i
:= uintptr(0); i
< n
; i
++ {
771 *h
.bitp
&^= bitScan
<< (heapBitsShift
+ h
.shift
)
772 h
= h
.forward(size
/ sys
.PtrSize
)
776 // clearCheckmarkSpan undoes all the checkmarking in a span.
777 // The actual checkmark bits are ignored, so the only work to do
778 // is to fix the pointer bits. (Pointer bits are ignored by scanobject
779 // but consulted by typedmemmove.)
780 func (h heapBits
) clearCheckmarkSpan(size
, n
, total
uintptr) {
781 // The ptrSize == 8 is a compile-time constant false on 32-bit and eliminates this code entirely.
782 if sys
.PtrSize
== 8 && size
== sys
.PtrSize
{
783 // Checkmark bit is type bit, bottom bit of every 2-bit entry.
784 // Only possible on 64-bit system, since minimum size is 8.
785 // Must clear type bit (checkmark bit) of every word.
786 // The type bit is the lower of every two-bit pair.
788 for i
:= uintptr(0); i
< n
; i
+= 4 {
789 *bitp |
= bitPointerAll
790 bitp
= subtract1(bitp
)
795 // oneBitCount is indexed by byte and produces the
796 // number of 1 bits in that byte. For example 128 has 1 bit set
797 // and oneBitCount[128] will holds 1.
798 var oneBitCount
= [256]uint8{
799 0, 1, 1, 2, 1, 2, 2, 3,
800 1, 2, 2, 3, 2, 3, 3, 4,
801 1, 2, 2, 3, 2, 3, 3, 4,
802 2, 3, 3, 4, 3, 4, 4, 5,
803 1, 2, 2, 3, 2, 3, 3, 4,
804 2, 3, 3, 4, 3, 4, 4, 5,
805 2, 3, 3, 4, 3, 4, 4, 5,
806 3, 4, 4, 5, 4, 5, 5, 6,
807 1, 2, 2, 3, 2, 3, 3, 4,
808 2, 3, 3, 4, 3, 4, 4, 5,
809 2, 3, 3, 4, 3, 4, 4, 5,
810 3, 4, 4, 5, 4, 5, 5, 6,
811 2, 3, 3, 4, 3, 4, 4, 5,
812 3, 4, 4, 5, 4, 5, 5, 6,
813 3, 4, 4, 5, 4, 5, 5, 6,
814 4, 5, 5, 6, 5, 6, 6, 7,
815 1, 2, 2, 3, 2, 3, 3, 4,
816 2, 3, 3, 4, 3, 4, 4, 5,
817 2, 3, 3, 4, 3, 4, 4, 5,
818 3, 4, 4, 5, 4, 5, 5, 6,
819 2, 3, 3, 4, 3, 4, 4, 5,
820 3, 4, 4, 5, 4, 5, 5, 6,
821 3, 4, 4, 5, 4, 5, 5, 6,
822 4, 5, 5, 6, 5, 6, 6, 7,
823 2, 3, 3, 4, 3, 4, 4, 5,
824 3, 4, 4, 5, 4, 5, 5, 6,
825 3, 4, 4, 5, 4, 5, 5, 6,
826 4, 5, 5, 6, 5, 6, 6, 7,
827 3, 4, 4, 5, 4, 5, 5, 6,
828 4, 5, 5, 6, 5, 6, 6, 7,
829 4, 5, 5, 6, 5, 6, 6, 7,
830 5, 6, 6, 7, 6, 7, 7, 8}
832 // countAlloc returns the number of objects allocated in span s by
833 // scanning the allocation bitmap.
834 // TODO:(rlh) Use popcount intrinsic.
835 func (s
*mspan
) countAlloc() int {
837 maxIndex
:= s
.nelems
/ 8
838 for i
:= uintptr(0); i
< maxIndex
; i
++ {
839 mrkBits
:= *s
.gcmarkBits
.bytep(i
)
840 count
+= int(oneBitCount
[mrkBits
])
842 if bitsInLastByte
:= s
.nelems
% 8; bitsInLastByte
!= 0 {
843 mrkBits
:= *s
.gcmarkBits
.bytep(maxIndex
)
844 mask
:= uint8((1 << bitsInLastByte
) - 1)
845 bits
:= mrkBits
& mask
846 count
+= int(oneBitCount
[bits
])
851 // heapBitsSetType records that the new allocation [x, x+size)
852 // holds in [x, x+dataSize) one or more values of type typ.
853 // (The number of values is given by dataSize / typ.size.)
854 // If dataSize < size, the fragment [x+dataSize, x+size) is
855 // recorded as non-pointer data.
856 // It is known that the type has pointers somewhere;
857 // malloc does not call heapBitsSetType when there are no pointers,
858 // because all free objects are marked as noscan during
859 // heapBitsSweepSpan.
861 // There can only be one allocation from a given span active at a time,
862 // and the bitmap for a span always falls on byte boundaries,
863 // so there are no write-write races for access to the heap bitmap.
864 // Hence, heapBitsSetType can access the bitmap without atomics.
866 // There can be read-write races between heapBitsSetType and things
867 // that read the heap bitmap like scanobject. However, since
868 // heapBitsSetType is only used for objects that have not yet been
869 // made reachable, readers will ignore bits being modified by this
870 // function. This does mean this function cannot transiently modify
871 // bits that belong to neighboring objects. Also, on weakly-ordered
872 // machines, callers must execute a store/store (publication) barrier
873 // between calling this function and making the object reachable.
874 func heapBitsSetType(x
, size
, dataSize
uintptr, typ
*_type
) {
875 const doubleCheck
= false // slow but helpful; enable to test modifications to this code
877 // dataSize is always size rounded up to the next malloc size class,
878 // except in the case of allocating a defer block, in which case
879 // size is sizeof(_defer{}) (at least 6 words) and dataSize may be
880 // arbitrarily larger.
882 // The checks for size == sys.PtrSize and size == 2*sys.PtrSize can therefore
883 // assume that dataSize == size without checking it explicitly.
885 if sys
.PtrSize
== 8 && size
== sys
.PtrSize
{
886 // It's one word and it has pointers, it must be a pointer.
887 // Since all allocated one-word objects are pointers
888 // (non-pointers are aggregated into tinySize allocations),
889 // initSpan sets the pointer bits for us. Nothing to do here.
891 h
:= heapBitsForAddr(x
)
893 throw("heapBitsSetType: pointer bit missing")
895 if !h
.morePointers() {
896 throw("heapBitsSetType: scan bit missing")
902 h
:= heapBitsForAddr(x
)
903 ptrmask
:= typ
.gcdata
// start of 1-bit pointer mask (or GC program, handled below)
905 // Heap bitmap bits for 2-word object are only 4 bits,
906 // so also shared with objects next to it.
907 // This is called out as a special case primarily for 32-bit systems,
908 // so that on 32-bit systems the code below can assume all objects
909 // are 4-word aligned (because they're all 16-byte aligned).
910 if size
== 2*sys
.PtrSize
{
911 if typ
.size
== sys
.PtrSize
{
912 // We're allocating a block big enough to hold two pointers.
913 // On 64-bit, that means the actual object must be two pointers,
914 // or else we'd have used the one-pointer-sized block.
915 // On 32-bit, however, this is the 8-byte block, the smallest one.
916 // So it could be that we're allocating one pointer and this was
917 // just the smallest block available. Distinguish by checking dataSize.
918 // (In general the number of instances of typ being allocated is
919 // dataSize/typ.size.)
920 if sys
.PtrSize
== 4 && dataSize
== sys
.PtrSize
{
921 // 1 pointer object. On 32-bit machines clear the bit for the
922 // unused second word.
923 *h
.bitp
&^= (bitPointer | bitScan |
((bitPointer | bitScan
) << heapBitsShift
)) << h
.shift
924 *h
.bitp |
= (bitPointer | bitScan
) << h
.shift
926 // 2-element slice of pointer.
927 *h
.bitp |
= (bitPointer | bitScan | bitPointer
<<heapBitsShift
) << h
.shift
931 // Otherwise typ.size must be 2*sys.PtrSize,
932 // and typ.kind&kindGCProg == 0.
934 if typ
.size
!= 2*sys
.PtrSize || typ
.kind
&kindGCProg
!= 0 {
935 print("runtime: heapBitsSetType size=", size
, " but typ.size=", typ
.size
, " gcprog=", typ
.kind
&kindGCProg
!= 0, "\n")
936 throw("heapBitsSetType")
939 b
:= uint32(*ptrmask
)
940 hb
:= (b
& 3) | bitScan
941 // bitPointer == 1, bitScan is 1 << 4, heapBitsShift is 1.
942 // 110011 is shifted h.shift and complemented.
943 // This clears out the bits that are about to be
944 // ored into *h.hbitp in the next instructions.
945 *h
.bitp
&^= (bitPointer | bitScan |
((bitPointer | bitScan
) << heapBitsShift
)) << h
.shift
946 *h
.bitp |
= uint8(hb
<< h
.shift
)
950 // Copy from 1-bit ptrmask into 2-bit bitmap.
951 // The basic approach is to use a single uintptr as a bit buffer,
952 // alternating between reloading the buffer and writing bitmap bytes.
953 // In general, one load can supply two bitmap byte writes.
954 // This is a lot of lines of code, but it compiles into relatively few
955 // machine instructions.
959 p
*byte // last ptrmask byte read
960 b
uintptr // ptrmask bits already loaded
961 nb
uintptr // number of bits in b at next read
962 endp
*byte // final ptrmask byte to read (then repeat)
963 endnb
uintptr // number of valid bits in *endp
964 pbits
uintptr // alternate source of bits
966 // Heap bitmap output.
967 w
uintptr // words processed
968 nw
uintptr // number of words to process
969 hbitp
*byte // next heap bitmap byte to write
970 hb
uintptr // bits being prepared for *hbitp
975 // Handle GC program. Delayed until this part of the code
976 // so that we can use the same double-checking mechanism
977 // as the 1-bit case. Nothing above could have encountered
978 // GC programs: the cases were all too small.
979 if typ
.kind
&kindGCProg
!= 0 {
980 heapBitsSetTypeGCProg(h
, typ
.ptrdata
, typ
.size
, dataSize
, size
, addb(typ
.gcdata
, 4))
982 // Double-check the heap bits written by GC program
983 // by running the GC program to create a 1-bit pointer mask
984 // and then jumping to the double-check code below.
985 // This doesn't catch bugs shared between the 1-bit and 4-bit
986 // GC program execution, but it does catch mistakes specific
987 // to just one of those and bugs in heapBitsSetTypeGCProg's
988 // implementation of arrays.
989 lock(&debugPtrmask
.lock
)
990 if debugPtrmask
.data
== nil {
991 debugPtrmask
.data
= (*byte)(persistentalloc(1<<20, 1, &memstats
.other_sys
))
993 ptrmask
= debugPtrmask
.data
994 runGCProg(addb(typ
.gcdata
, 4), nil, ptrmask
, 1)
1000 // Note about sizes:
1002 // typ.size is the number of words in the object,
1003 // and typ.ptrdata is the number of words in the prefix
1004 // of the object that contains pointers. That is, the final
1005 // typ.size - typ.ptrdata words contain no pointers.
1006 // This allows optimization of a common pattern where
1007 // an object has a small header followed by a large scalar
1008 // buffer. If we know the pointers are over, we don't have
1009 // to scan the buffer's heap bitmap at all.
1010 // The 1-bit ptrmasks are sized to contain only bits for
1011 // the typ.ptrdata prefix, zero padded out to a full byte
1012 // of bitmap. This code sets nw (below) so that heap bitmap
1013 // bits are only written for the typ.ptrdata prefix; if there is
1014 // more room in the allocated object, the next heap bitmap
1015 // entry is a 00, indicating that there are no more pointers
1016 // to scan. So only the ptrmask for the ptrdata bytes is needed.
1018 // Replicated copies are not as nice: if there is an array of
1019 // objects with scalar tails, all but the last tail does have to
1020 // be initialized, because there is no way to say "skip forward".
1021 // However, because of the possibility of a repeated type with
1022 // size not a multiple of 4 pointers (one heap bitmap byte),
1023 // the code already must handle the last ptrmask byte specially
1024 // by treating it as containing only the bits for endnb pointers,
1025 // where endnb <= 4. We represent large scalar tails that must
1026 // be expanded in the replication by setting endnb larger than 4.
1027 // This will have the effect of reading many bits out of b,
1028 // but once the real bits are shifted out, b will supply as many
1029 // zero bits as we try to read, which is exactly what we need.
1032 if typ
.size
< dataSize
{
1033 // Filling in bits for an array of typ.
1034 // Set up for repetition of ptrmask during main loop.
1035 // Note that ptrmask describes only a prefix of
1036 const maxBits
= sys
.PtrSize
*8 - 7
1037 if typ
.ptrdata
/sys
.PtrSize
<= maxBits
{
1038 // Entire ptrmask fits in uintptr with room for a byte fragment.
1039 // Load into pbits and never read from ptrmask again.
1040 // This is especially important when the ptrmask has
1041 // fewer than 8 bits in it; otherwise the reload in the middle
1042 // of the Phase 2 loop would itself need to loop to gather
1045 // Accumulate ptrmask into b.
1046 // ptrmask is sized to describe only typ.ptrdata, but we record
1047 // it as describing typ.size bytes, since all the high bits are zero.
1048 nb
= typ
.ptrdata
/ sys
.PtrSize
1049 for i
:= uintptr(0); i
< nb
; i
+= 8 {
1050 b |
= uintptr(*p
) << i
1053 nb
= typ
.size
/ sys
.PtrSize
1055 // Replicate ptrmask to fill entire pbits uintptr.
1056 // Doubling and truncating is fewer steps than
1057 // iterating by nb each time. (nb could be 1.)
1058 // Since we loaded typ.ptrdata/sys.PtrSize bits
1059 // but are pretending to have typ.size/sys.PtrSize,
1060 // there might be no replication necessary/possible.
1063 if nb
+nb
<= maxBits
{
1064 for endnb
<= sys
.PtrSize
*8 {
1065 pbits |
= pbits
<< endnb
1068 // Truncate to a multiple of original ptrmask.
1069 // Because nb+nb <= maxBits, nb fits in a byte.
1070 // Byte division is cheaper than uintptr division.
1071 endnb
= uintptr(maxBits
/byte(nb
)) * nb
1072 pbits
&= 1<<endnb
- 1
1077 // Clear p and endp as sentinel for using pbits.
1078 // Checked during Phase 2 loop.
1082 // Ptrmask is larger. Read it multiple times.
1083 n
:= (typ
.ptrdata
/sys
.PtrSize
+7)/8 - 1
1084 endp
= addb(ptrmask
, n
)
1085 endnb
= typ
.size
/sys
.PtrSize
- n
*8
1094 if typ
.size
== dataSize
{
1095 // Single entry: can stop once we reach the non-pointer data.
1096 nw
= typ
.ptrdata
/ sys
.PtrSize
1098 // Repeated instances of typ in an array.
1099 // Have to process first N-1 entries in full, but can stop
1100 // once we reach the non-pointer data in the final entry.
1101 nw
= ((dataSize
/typ
.size
-1)*typ
.size
+ typ
.ptrdata
) / sys
.PtrSize
1104 // No pointers! Caller was supposed to check.
1105 println("runtime: invalid type ", *typ
.string)
1106 throw("heapBitsSetType: called with non-pointer type")
1110 // Must write at least 2 words, because the "no scan"
1111 // encoding doesn't take effect until the third word.
1115 // Phase 1: Special case for leading byte (shift==0) or half-byte (shift==4).
1116 // The leading byte is special because it contains the bits for word 1,
1117 // which does not have the scan bit set.
1118 // The leading half-byte is special because it's a half a byte,
1119 // so we have to be careful with the bits already there.
1122 throw("heapBitsSetType: unexpected shift")
1125 // Ptrmask and heap bitmap are aligned.
1126 // Handle first byte of bitmap specially.
1128 // The first byte we write out covers the first four
1129 // words of the object. The scan/dead bit on the first
1130 // word must be set to scan since there are pointers
1131 // somewhere in the object. The scan/dead bit on the
1132 // second word is the checkmark, so we don't set it.
1133 // In all following words, we set the scan/dead
1134 // appropriately to indicate that the object contains
1135 // to the next 2-bit entry in the bitmap.
1137 // TODO: It doesn't matter if we set the checkmark, so
1138 // maybe this case isn't needed any more.
1139 hb
= b
& bitPointerAll
1140 hb |
= bitScan | bitScan
<<(2*heapBitsShift
) | bitScan
<<(3*heapBitsShift
)
1141 if w
+= 4; w
>= nw
{
1145 hbitp
= subtract1(hbitp
)
1149 case sys
.PtrSize
== 8 && h
.shift
== 2:
1150 // Ptrmask and heap bitmap are misaligned.
1151 // The bits for the first two words are in a byte shared
1152 // with another object, so we must be careful with the bits
1154 // We took care of 1-word and 2-word objects above,
1155 // so this is at least a 6-word object.
1156 hb
= (b
& (bitPointer | bitPointer
<<heapBitsShift
)) << (2 * heapBitsShift
)
1157 // This is not noscan, so set the scan bit in the
1159 hb |
= bitScan
<< (2 * heapBitsShift
)
1162 // Note: no bitScan for second word because that's
1164 *hbitp
&^= uint8((bitPointer | bitScan |
(bitPointer
<< heapBitsShift
)) << (2 * heapBitsShift
))
1166 hbitp
= subtract1(hbitp
)
1167 if w
+= 2; w
>= nw
{
1168 // We know that there is more data, because we handled 2-word objects above.
1169 // This must be at least a 6-word object. If we're out of pointer words,
1170 // mark no scan in next bitmap byte and finish.
1177 // Phase 2: Full bytes in bitmap, up to but not including write to last byte (full or partial) in bitmap.
1178 // The loop computes the bits for that last write but does not execute the write;
1179 // it leaves the bits in hb for processing by phase 3.
1180 // To avoid repeated adjustment of nb, we subtract out the 4 bits we're going to
1181 // use in the first half of the loop right now, and then we only adjust nb explicitly
1182 // if the 8 bits used by each iteration isn't balanced by 8 bits loaded mid-loop.
1185 // Emit bitmap byte.
1186 // b has at least nb+4 bits, with one exception:
1187 // if w+4 >= nw, then b has only nw-w bits,
1188 // but we'll stop at the break and then truncate
1189 // appropriately in Phase 3.
1190 hb
= b
& bitPointerAll
1192 if w
+= 4; w
>= nw
{
1196 hbitp
= subtract1(hbitp
)
1199 // Load more bits. b has nb right now.
1201 // Fast path: keep reading from ptrmask.
1202 // nb unmodified: we just loaded 8 bits,
1203 // and the next iteration will consume 8 bits,
1204 // leaving us with the same nb the next time we're here.
1206 b |
= uintptr(*p
) << nb
1209 // Reduce the number of bits in b.
1210 // This is important if we skipped
1211 // over a scalar tail, since nb could
1212 // be larger than the bit width of b.
1215 } else if p
== nil {
1216 // Almost as fast path: track bit count and refill from pbits.
1217 // For short repetitions.
1222 nb
-= 8 // for next iteration
1224 // Slow path: reached end of ptrmask.
1225 // Process final partial byte and rewind to start.
1226 b |
= uintptr(*p
) << nb
1229 b |
= uintptr(*ptrmask
) << nb
1237 // Emit bitmap byte.
1238 hb
= b
& bitPointerAll
1240 if w
+= 4; w
>= nw
{
1244 hbitp
= subtract1(hbitp
)
1249 // Phase 3: Write last byte or partial byte and zero the rest of the bitmap entries.
1251 // Counting the 4 entries in hb not yet written to memory,
1252 // there are more entries than possible pointer slots.
1253 // Discard the excess entries (can't be more than 3).
1254 mask
:= uintptr(1)<<(4-(w
-nw
)) - 1
1255 hb
&= mask | mask
<<4 // apply mask to both pointer bits and scan bits
1258 // Change nw from counting possibly-pointer words to total words in allocation.
1259 nw
= size
/ sys
.PtrSize
1261 // Write whole bitmap bytes.
1262 // The first is hb, the rest are zero.
1265 hbitp
= subtract1(hbitp
)
1266 hb
= 0 // for possible final half-byte below
1267 for w
+= 4; w
<= nw
; w
+= 4 {
1269 hbitp
= subtract1(hbitp
)
1273 // Write final partial bitmap byte if any.
1274 // We know w > nw, or else we'd still be in the loop above.
1275 // It can be bigger only due to the 4 entries in hb that it counts.
1276 // If w == nw+4 then there's nothing left to do: we wrote all nw entries
1277 // and can discard the 4 sitting in hb.
1278 // But if w == nw+2, we need to write first two in hb.
1279 // The byte is shared with the next object, so be careful with
1282 *hbitp
= *hbitp
&^(bitPointer|bitScan|
(bitPointer|bitScan
)<<heapBitsShift
) |
uint8(hb
)
1286 // Phase 4: all done, but perhaps double check.
1288 end
:= heapBitsForAddr(x
+ size
)
1289 if typ
.kind
&kindGCProg
== 0 && (hbitp
!= end
.bitp ||
(w
== nw
+2) != (end
.shift
== 2)) {
1290 println("ended at wrong bitmap byte for", *typ
.string, "x", dataSize
/typ
.size
)
1291 print("typ.size=", typ
.size
, " typ.ptrdata=", typ
.ptrdata
, " dataSize=", dataSize
, " size=", size
, "\n")
1292 print("w=", w
, " nw=", nw
, " b=", hex(b
), " nb=", nb
, " hb=", hex(hb
), "\n")
1293 h0
:= heapBitsForAddr(x
)
1294 print("initial bits h0.bitp=", h0
.bitp
, " h0.shift=", h0
.shift
, "\n")
1295 print("ended at hbitp=", hbitp
, " but next starts at bitp=", end
.bitp
, " shift=", end
.shift
, "\n")
1296 throw("bad heapBitsSetType")
1299 // Double-check that bits to be written were written correctly.
1300 // Does not check that other bits were not written, unfortunately.
1301 h
:= heapBitsForAddr(x
)
1302 nptr
:= typ
.ptrdata
/ sys
.PtrSize
1303 ndata
:= typ
.size
/ sys
.PtrSize
1304 count
:= dataSize
/ typ
.size
1305 totalptr
:= ((count
-1)*typ
.size
+ typ
.ptrdata
) / sys
.PtrSize
1306 for i
:= uintptr(0); i
< size
/sys
.PtrSize
; i
++ {
1308 var have
, want
uint8
1309 have
= (*h
.bitp
>> h
.shift
) & (bitPointer | bitScan
)
1311 want
= 0 // deadmarker
1312 if typ
.kind
&kindGCProg
!= 0 && i
< (totalptr
+3)/4*4 {
1316 if j
< nptr
&& (*addb(ptrmask
, j
/8)>>(j%8
))&1 != 0 {
1326 println("mismatch writing bits for", *typ
.string, "x", dataSize
/typ
.size
)
1327 print("typ.size=", typ
.size
, " typ.ptrdata=", typ
.ptrdata
, " dataSize=", dataSize
, " size=", size
, "\n")
1328 print("kindGCProg=", typ
.kind
&kindGCProg
!= 0, "\n")
1329 print("w=", w
, " nw=", nw
, " b=", hex(b
), " nb=", nb
, " hb=", hex(hb
), "\n")
1330 h0
:= heapBitsForAddr(x
)
1331 print("initial bits h0.bitp=", h0
.bitp
, " h0.shift=", h0
.shift
, "\n")
1332 print("current bits h.bitp=", h
.bitp
, " h.shift=", h
.shift
, " *h.bitp=", hex(*h
.bitp
), "\n")
1333 print("ptrmask=", ptrmask
, " p=", p
, " endp=", endp
, " endnb=", endnb
, " pbits=", hex(pbits
), " b=", hex(b
), " nb=", nb
, "\n")
1334 println("at word", i
, "offset", i
*sys
.PtrSize
, "have", have
, "want", want
)
1335 if typ
.kind
&kindGCProg
!= 0 {
1336 println("GC program:")
1337 dumpGCProg(addb(typ
.gcdata
, 4))
1339 throw("bad heapBitsSetType")
1343 if ptrmask
== debugPtrmask
.data
{
1344 unlock(&debugPtrmask
.lock
)
1349 var debugPtrmask
struct {
1354 // heapBitsSetTypeGCProg implements heapBitsSetType using a GC program.
1355 // progSize is the size of the memory described by the program.
1356 // elemSize is the size of the element that the GC program describes (a prefix of).
1357 // dataSize is the total size of the intended data, a multiple of elemSize.
1358 // allocSize is the total size of the allocated memory.
1360 // GC programs are only used for large allocations.
1361 // heapBitsSetType requires that allocSize is a multiple of 4 words,
1362 // so that the relevant bitmap bytes are not shared with surrounding
1364 func heapBitsSetTypeGCProg(h heapBits
, progSize
, elemSize
, dataSize
, allocSize
uintptr, prog
*byte) {
1365 if sys
.PtrSize
== 8 && allocSize
%(4*sys
.PtrSize
) != 0 {
1366 // Alignment will be wrong.
1367 throw("heapBitsSetTypeGCProg: small allocation")
1369 var totalBits
uintptr
1370 if elemSize
== dataSize
{
1371 totalBits
= runGCProg(prog
, nil, h
.bitp
, 2)
1372 if totalBits
*sys
.PtrSize
!= progSize
{
1373 println("runtime: heapBitsSetTypeGCProg: total bits", totalBits
, "but progSize", progSize
)
1374 throw("heapBitsSetTypeGCProg: unexpected bit count")
1377 count
:= dataSize
/ elemSize
1379 // Piece together program trailer to run after prog that does:
1381 // repeat(1, elemSize-progSize-1) // zeros to fill element size
1382 // repeat(elemSize, count-1) // repeat that element for count
1383 // This zero-pads the data remaining in the first element and then
1384 // repeats that first element to fill the array.
1385 var trailer
[40]byte // 3 varints (max 10 each) + some bytes
1387 if n
:= elemSize
/sys
.PtrSize
- progSize
/sys
.PtrSize
; n
> 0 {
1398 for ; n
>= 0x80; n
>>= 7 {
1399 trailer
[i
] = byte(n |
0x80)
1402 trailer
[i
] = byte(n
)
1406 // repeat(elemSize/ptrSize, count-1)
1409 n
:= elemSize
/ sys
.PtrSize
1410 for ; n
>= 0x80; n
>>= 7 {
1411 trailer
[i
] = byte(n |
0x80)
1414 trailer
[i
] = byte(n
)
1417 for ; n
>= 0x80; n
>>= 7 {
1418 trailer
[i
] = byte(n |
0x80)
1421 trailer
[i
] = byte(n
)
1426 runGCProg(prog
, &trailer
[0], h
.bitp
, 2)
1428 // Even though we filled in the full array just now,
1429 // record that we only filled in up to the ptrdata of the
1430 // last element. This will cause the code below to
1431 // memclr the dead section of the final array element,
1432 // so that scanobject can stop early in the final element.
1433 totalBits
= (elemSize
*(count
-1) + progSize
) / sys
.PtrSize
1435 endProg
:= unsafe
.Pointer(subtractb(h
.bitp
, (totalBits
+3)/4))
1436 endAlloc
:= unsafe
.Pointer(subtractb(h
.bitp
, allocSize
/heapBitmapScale
))
1437 memclrNoHeapPointers(add(endAlloc
, 1), uintptr(endProg
)-uintptr(endAlloc
))
1440 // progToPointerMask returns the 1-bit pointer mask output by the GC program prog.
1441 // size the size of the region described by prog, in bytes.
1442 // The resulting bitvector will have no more than size/sys.PtrSize bits.
1443 func progToPointerMask(prog
*byte, size
uintptr) bitvector
{
1444 n
:= (size
/sys
.PtrSize
+ 7) / 8
1445 x
:= (*[1 << 30]byte)(persistentalloc(n
+1, 1, &memstats
.buckhash_sys
))[:n
+1]
1446 x
[len(x
)-1] = 0xa1 // overflow check sentinel
1447 n
= runGCProg(prog
, nil, &x
[0], 1)
1448 if x
[len(x
)-1] != 0xa1 {
1449 throw("progToPointerMask: overflow")
1451 return bitvector
{int32(n
), &x
[0]}
1454 // Packed GC pointer bitmaps, aka GC programs.
1456 // For large types containing arrays, the type information has a
1457 // natural repetition that can be encoded to save space in the
1458 // binary and in the memory representation of the type information.
1460 // The encoding is a simple Lempel-Ziv style bytecode machine
1461 // with the following instructions:
1464 // 0nnnnnnn: emit n bits copied from the next (n+7)/8 bytes
1465 // 10000000 n c: repeat the previous n bits c times; n, c are varints
1466 // 1nnnnnnn c: repeat the previous n bits c times; c is a varint
1468 // runGCProg executes the GC program prog, and then trailer if non-nil,
1469 // writing to dst with entries of the given size.
1470 // If size == 1, dst is a 1-bit pointer mask laid out moving forward from dst.
1471 // If size == 2, dst is the 2-bit heap bitmap, and writes move backward
1472 // starting at dst (because the heap bitmap does). In this case, the caller guarantees
1473 // that only whole bytes in dst need to be written.
1475 // runGCProg returns the number of 1- or 2-bit entries written to memory.
1476 func runGCProg(prog
, trailer
, dst
*byte, size
int) uintptr {
1479 // Bits waiting to be written to memory.
1486 // Flush accumulated full bytes.
1487 // The rest of the loop assumes that nbits <= 7.
1488 for ; nbits
>= 8; nbits
-= 8 {
1494 v
:= bits
&bitPointerAll | bitScanAll
1496 dst
= subtract1(dst
)
1498 v
= bits
&bitPointerAll | bitScanAll
1500 dst
= subtract1(dst
)
1505 // Process one instruction.
1510 // Literal bits; n == 0 means end of program.
1512 // Program is over; continue in trailer if present.
1514 //println("trailer")
1522 //println("lit", n, dst)
1524 for i
:= uintptr(0); i
< nbyte
; i
++ {
1525 bits |
= uintptr(*p
) << nbits
1532 v
:= bits
&0xf | bitScanAll
1534 dst
= subtract1(dst
)
1536 v
= bits
&0xf | bitScanAll
1538 dst
= subtract1(dst
)
1543 bits |
= uintptr(*p
) << nbits
1550 // Repeat. If n == 0, it is encoded in a varint in the next bytes.
1552 for off
:= uint(0); ; off
+= 7 {
1555 n |
= (x
& 0x7F) << off
1562 // Count is encoded in a varint in the next bytes.
1564 for off
:= uint(0); ; off
+= 7 {
1567 c |
= (x
& 0x7F) << off
1572 c
*= n
// now total number of bits to copy
1574 // If the number of bits being repeated is small, load them
1575 // into a register and use that register for the entire loop
1576 // instead of repeatedly reading from memory.
1577 // Handling fewer than 8 bits here makes the general loop simpler.
1578 // The cutoff is sys.PtrSize*8 - 7 to guarantee that when we add
1579 // the pattern to a bit buffer holding at most 7 bits (a partial byte)
1580 // it will not overflow.
1582 const maxBits
= sys
.PtrSize
*8 - 7
1584 // Start with bits in output buffer.
1588 // If we need more bits, fetch them from memory.
1590 src
= subtract1(src
)
1593 pattern |
= uintptr(*src
)
1594 src
= subtract1(src
)
1601 pattern |
= uintptr(*src
) & 0xf
1607 // We started with the whole bit output buffer,
1608 // and then we loaded bits from whole bytes.
1609 // Either way, we might now have too many instead of too few.
1610 // Discard the extra.
1612 pattern
>>= npattern
- n
1616 // Replicate pattern to at most maxBits.
1618 // One bit being repeated.
1619 // If the bit is 1, make the pattern all 1s.
1620 // If the bit is 0, the pattern is already all 0s,
1621 // but we can claim that the number of bits
1622 // in the word is equal to the number we need (c),
1623 // because right shift of bits will zero fill.
1625 pattern
= 1<<maxBits
- 1
1633 if nb
+nb
<= maxBits
{
1634 // Double pattern until the whole uintptr is filled.
1635 for nb
<= sys
.PtrSize
*8 {
1639 // Trim away incomplete copy of original pattern in high bits.
1640 // TODO(rsc): Replace with table lookup or loop on systems without divide?
1641 nb
= maxBits
/ npattern
* npattern
1648 // Add pattern to bit buffer and flush bit buffer, c/npattern times.
1649 // Since pattern contains >8 bits, there will be full bytes to flush
1650 // on each iteration.
1651 for ; c
>= npattern
; c
-= npattern
{
1652 bits |
= pattern
<< nbits
1663 *dst
= uint8(bits
&0xf | bitScanAll
)
1664 dst
= subtract1(dst
)
1671 // Add final fragment to bit buffer.
1674 bits |
= pattern
<< nbits
1680 // Repeat; n too large to fit in a register.
1681 // Since nbits <= 7, we know the first few bytes of repeated data
1682 // are already written to memory.
1683 off
:= n
- nbits
// n > nbits because n > maxBits and nbits <= 7
1685 // Leading src fragment.
1686 src
= subtractb(src
, (off
+7)/8)
1687 if frag
:= off
& 7; frag
!= 0 {
1688 bits |
= uintptr(*src
) >> (8 - frag
) << nbits
1693 // Main loop: load one byte, write another.
1694 // The bits are rotating through the bit buffer.
1695 for i
:= c
/ 8; i
> 0; i
-- {
1696 bits |
= uintptr(*src
) << nbits
1702 // Final src fragment.
1704 bits |
= (uintptr(*src
) & (1<<c
- 1)) << nbits
1708 // Leading src fragment.
1709 src
= addb(src
, (off
+3)/4)
1710 if frag
:= off
& 3; frag
!= 0 {
1711 bits |
= (uintptr(*src
) & 0xf) >> (4 - frag
) << nbits
1712 src
= subtract1(src
)
1716 // Main loop: load one byte, write another.
1717 // The bits are rotating through the bit buffer.
1718 for i
:= c
/ 4; i
> 0; i
-- {
1719 bits |
= (uintptr(*src
) & 0xf) << nbits
1720 src
= subtract1(src
)
1721 *dst
= uint8(bits
&0xf | bitScanAll
)
1722 dst
= subtract1(dst
)
1725 // Final src fragment.
1727 bits |
= (uintptr(*src
) & (1<<c
- 1)) << nbits
1733 // Write any final bits out, using full-byte writes, even for the final byte.
1734 var totalBits
uintptr
1736 totalBits
= (uintptr(unsafe
.Pointer(dst
))-uintptr(unsafe
.Pointer(dstStart
)))*8 + nbits
1738 for ; nbits
> 0; nbits
-= 8 {
1744 totalBits
= (uintptr(unsafe
.Pointer(dstStart
))-uintptr(unsafe
.Pointer(dst
)))*4 + nbits
1746 for ; nbits
> 0; nbits
-= 4 {
1747 v
:= bits
&0xf | bitScanAll
1749 dst
= subtract1(dst
)
1756 func dumpGCProg(p
*byte) {
1762 print("\t", nptr
, " end\n")
1766 print("\t", nptr
, " lit ", x
, ":")
1768 for i
:= 0; i
< n
; i
++ {
1775 nbit
:= int(x
&^ 0x80)
1777 for nb
:= uint(0); ; nb
+= 7 {
1780 nbit |
= int(x
&0x7f) << nb
1787 for nb
:= uint(0); ; nb
+= 7 {
1790 count |
= int(x
&0x7f) << nb
1795 print("\t", nptr
, " repeat ", nbit
, " × ", count
, "\n")
1796 nptr
+= nbit
* count
1803 // gcbits returns the GC type info for x, for testing.
1804 // The result is the bitmap entries (0 or 1), one entry per byte.
1805 //go:linkname reflect_gcbits reflect.gcbits
1806 func reflect_gcbits(x
interface{}) []byte {
1808 typ
:= (*ptrtype
)(unsafe
.Pointer(efaceOf(&x
)._type
)).elem
1809 nptr
:= typ
.ptrdata
/ sys
.PtrSize
1810 for uintptr(len(ret
)) > nptr
&& ret
[len(ret
)-1] == 0 {
1811 ret
= ret
[:len(ret
)-1]
1816 // Returns GC type info for object p for testing.
1817 func getgcmask(ep
interface{}) (mask
[]byte) {
1824 for i
:= 0; i
< roots
.count
; i
++ {
1825 pr
:= roots
.roots
[i
]
1826 addr
:= uintptr(pr
.decl
)
1827 if addr
<= uintptr(p
) && uintptr(p
) < addr
+pr
.size
{
1828 n
:= (*ptrtype
)(unsafe
.Pointer(t
)).elem
.size
1829 mask
= make([]byte, n
/sys
.PtrSize
)
1830 copy(mask
, (*[1 << 29]uint8)(unsafe
.Pointer(pr
.gcdata
))[:pr
.ptrdata
])
1840 if mlookup(uintptr(p
), &base
, &n
, nil) != 0 {
1841 mask
= make([]byte, n
/sys
.PtrSize
)
1842 for i
:= uintptr(0); i
< n
; i
+= sys
.PtrSize
{
1843 hbits
:= heapBitsForAddr(base
+ i
)
1844 if hbits
.isPointer() {
1845 mask
[i
/sys
.PtrSize
] = 1
1847 if i
!= 1*sys
.PtrSize
&& !hbits
.morePointers() {
1848 mask
= mask
[:i
/sys
.PtrSize
]
1855 // otherwise, not something the GC knows about.
1856 // possibly read-only data, like malloc(0).
1857 // must not have pointers
1858 // For gccgo, may live on the stack, which is collected conservatively.