2 * sgen-alloc.c: Object allocation routines + managed allocators
5 * Paolo Molaro (lupus@ximian.com)
6 * Rodrigo Kumpera (kumpera@gmail.com)
8 * Copyright 2005-2011 Novell, Inc (http://www.novell.com)
9 * Copyright 2011 Xamarin Inc (http://www.xamarin.com)
10 * Copyright 2011 Xamarin, Inc.
11 * Copyright (C) 2012 Xamarin Inc
13 * This library is free software; you can redistribute it and/or
14 * modify it under the terms of the GNU Library General Public
15 * License 2.0 as published by the Free Software Foundation;
17 * This library is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
20 * Library General Public License for more details.
22 * You should have received a copy of the GNU Library General Public
23 * License 2.0 along with this library; if not, write to the Free
24 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
28 * ######################################################################
29 * ######## Object allocation
30 * ######################################################################
31 * This section of code deals with allocating memory for objects.
32 * There are several ways:
33 * *) allocate large objects
34 * *) allocate normal objects
35 * *) fast lock-free allocation
36 * *) allocation of pinned objects
42 #include "metadata/sgen-gc.h"
43 #include "metadata/sgen-protocol.h"
44 #include "metadata/sgen-memory-governor.h"
45 #include "metadata/profiler-private.h"
46 #include "metadata/marshal.h"
47 #include "metadata/method-builder.h"
48 #include "metadata/abi-details.h"
49 #include "utils/mono-memory-model.h"
50 #include "utils/mono-counters.h"
52 #define ALIGN_UP SGEN_ALIGN_UP
53 #define ALLOC_ALIGN SGEN_ALLOC_ALIGN
54 #define ALLOC_ALIGN_BITS SGEN_ALLOC_ALIGN_BITS
55 #define MAX_SMALL_OBJ_SIZE SGEN_MAX_SMALL_OBJ_SIZE
56 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
58 #define OPDEF(a,b,c,d,e,f,g,h,i,j) \
62 #include "mono/cil/opcode.def"
68 static gboolean use_managed_allocator = TRUE;
70 #ifdef HEAVY_STATISTICS
71 static long long stat_objects_alloced = 0;
72 static long long stat_bytes_alloced = 0;
73 static long long stat_bytes_alloced_los = 0;
78 * Allocation is done from a Thread Local Allocation Buffer (TLAB). TLABs are allocated
79 * from nursery fragments.
80 * tlab_next is the pointer to the space inside the TLAB where the next object will
82 * tlab_temp_end is the pointer to the end of the temporary space reserved for
83 * the allocation: it allows us to set the scan starts at reasonable intervals.
84 * tlab_real_end points to the end of the TLAB.
88 * FIXME: What is faster, a TLS variable pointing to a structure, or separate TLS
89 * variables for next+temp_end ?
92 static __thread char *tlab_start;
93 static __thread char *tlab_next;
94 static __thread char *tlab_temp_end;
95 static __thread char *tlab_real_end;
96 /* Used by the managed allocator/wbarrier */
97 static __thread char **tlab_next_addr;
100 #ifdef HAVE_KW_THREAD
101 #define TLAB_START tlab_start
102 #define TLAB_NEXT tlab_next
103 #define TLAB_TEMP_END tlab_temp_end
104 #define TLAB_REAL_END tlab_real_end
106 #define TLAB_START (__thread_info__->tlab_start)
107 #define TLAB_NEXT (__thread_info__->tlab_next)
108 #define TLAB_TEMP_END (__thread_info__->tlab_temp_end)
109 #define TLAB_REAL_END (__thread_info__->tlab_real_end)
113 alloc_degraded (MonoVTable *vtable, size_t size, gboolean for_mature)
115 static int last_major_gc_warned = -1;
116 static int num_degraded = 0;
121 if (last_major_gc_warned < gc_stats.major_gc_count) {
123 if (num_degraded == 1 || num_degraded == 3)
124 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_GC, "Warning: Degraded allocation. Consider increasing nursery-size if the warning persists.");
125 else if (num_degraded == 10)
126 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_GC, "Warning: Repeated degraded allocation. Consider increasing nursery-size.");
127 last_major_gc_warned = gc_stats.major_gc_count;
129 SGEN_ATOMIC_ADD_P (degraded_mode, size);
130 sgen_ensure_free_space (size);
132 if (sgen_need_major_collection (size))
133 sgen_perform_collection (size, GENERATION_OLD, "mature allocation failure", !for_mature);
137 p = major_collector.alloc_degraded (vtable, size);
140 MONO_GC_MAJOR_OBJ_ALLOC_MATURE ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
142 binary_protocol_alloc_degraded (p, vtable, size);
143 MONO_GC_MAJOR_OBJ_ALLOC_DEGRADED ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
150 zero_tlab_if_necessary (void *p, size_t size)
152 if (nursery_clear_policy == CLEAR_AT_TLAB_CREATION || nursery_clear_policy == CLEAR_AT_TLAB_CREATION_DEBUG) {
156 * This function is called for all allocations in
157 * TLABs. TLABs originate from fragments, which are
158 * initialized to be faux arrays. The remainder of
159 * the fragments are zeroed out at initialization for
160 * CLEAR_AT_GC, so here we just need to make sure that
161 * the array header is zeroed. Since we don't know
162 * whether we're called for the start of a fragment or
163 * for somewhere in between, we zero in any case, just
167 if (size >= sizeof (MonoArray))
168 memset (p, 0, sizeof (MonoArray));
170 static guint8 zeros [sizeof (MonoArray)];
172 SGEN_ASSERT (0, !memcmp (p, zeros, size), "TLAB segment must be zeroed out.");
178 * Provide a variant that takes just the vtable for small fixed-size objects.
179 * The aligned size is already computed and stored in vt->gc_descr.
180 * Note: every SGEN_SCAN_START_SIZE or so we are given the chance to do some special
181 * processing. We can keep track of where objects start, for example,
182 * so when we scan the thread stacks for pinned objects, we can start
183 * a search for the pinned object in SGEN_SCAN_START_SIZE chunks.
186 mono_gc_alloc_obj_nolock (MonoVTable *vtable, size_t size)
188 /* FIXME: handle OOM */
192 size_t real_size = size;
196 HEAVY_STAT (++stat_objects_alloced);
197 if (real_size <= SGEN_MAX_SMALL_OBJ_SIZE)
198 HEAVY_STAT (stat_bytes_alloced += size);
200 HEAVY_STAT (stat_bytes_alloced_los += size);
202 size = ALIGN_UP (size);
204 g_assert (vtable->gc_descr);
206 if (G_UNLIKELY (has_per_allocation_action)) {
207 static int alloc_count;
208 int current_alloc = InterlockedIncrement (&alloc_count);
210 if (collect_before_allocs) {
211 if (((current_alloc % collect_before_allocs) == 0) && nursery_section) {
212 sgen_perform_collection (0, GENERATION_NURSERY, "collect-before-alloc-triggered", TRUE);
213 if (!degraded_mode && sgen_can_alloc_size (size) && real_size <= SGEN_MAX_SMALL_OBJ_SIZE) {
215 g_assert_not_reached ();
218 } else if (verify_before_allocs) {
219 if ((current_alloc % verify_before_allocs) == 0)
220 sgen_check_whole_heap_stw ();
225 * We must already have the lock here instead of after the
226 * fast path because we might be interrupted in the fast path
227 * (after confirming that new_next < TLAB_TEMP_END) by the GC,
228 * and we'll end up allocating an object in a fragment which
229 * no longer belongs to us.
231 * The managed allocator does not do this, but it's treated
232 * specially by the world-stopping code.
235 if (real_size > SGEN_MAX_SMALL_OBJ_SIZE) {
236 p = sgen_los_alloc_large_inner (vtable, ALIGN_UP (real_size));
238 /* tlab_next and tlab_temp_end are TLS vars so accessing them might be expensive */
240 p = (void**)TLAB_NEXT;
241 /* FIXME: handle overflow */
242 new_next = (char*)p + size;
243 TLAB_NEXT = new_next;
245 if (G_LIKELY (new_next < TLAB_TEMP_END)) {
249 * FIXME: We might need a memory barrier here so the change to tlab_next is
250 * visible before the vtable store.
253 CANARIFY_ALLOC(p,real_size);
254 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, vtable->klass->name, size);
255 binary_protocol_alloc (p , vtable, size);
256 if (G_UNLIKELY (MONO_GC_NURSERY_OBJ_ALLOC_ENABLED ()))
257 MONO_GC_NURSERY_OBJ_ALLOC ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
258 g_assert (*p == NULL);
259 mono_atomic_store_seq (p, vtable);
266 /* there are two cases: the object is too big or we run out of space in the TLAB */
267 /* we also reach here when the thread does its first allocation after a minor
268 * collection, since the tlab_ variables are initialized to NULL.
269 * there can be another case (from ORP), if we cooperate with the runtime a bit:
270 * objects that need finalizers can have the high bit set in their size
271 * so the above check fails and we can readily add the object to the queue.
272 * This avoids taking again the GC lock when registering, but this is moot when
273 * doing thread-local allocation, so it may not be a good idea.
275 if (TLAB_NEXT >= TLAB_REAL_END) {
276 int available_in_tlab;
278 * Run out of space in the TLAB. When this happens, some amount of space
279 * remains in the TLAB, but not enough to satisfy the current allocation
280 * request. Currently, we retire the TLAB in all cases, later we could
281 * keep it if the remaining space is above a treshold, and satisfy the
282 * allocation directly from the nursery.
285 /* when running in degraded mode, we continue allocing that way
286 * for a while, to decrease the number of useless nursery collections.
288 if (degraded_mode && degraded_mode < DEFAULT_NURSERY_SIZE)
289 return alloc_degraded (vtable, size, FALSE);
291 available_in_tlab = (int)(TLAB_REAL_END - TLAB_NEXT);//We'll never have tlabs > 2Gb
292 if (size > tlab_size || available_in_tlab > SGEN_MAX_NURSERY_WASTE) {
293 /* Allocate directly from the nursery */
294 p = sgen_nursery_alloc (size);
297 * We couldn't allocate from the nursery, so we try
298 * collecting. Even after the collection, we might
299 * still not have enough memory to allocate the
300 * object. The reason will most likely be that we've
301 * run out of memory, but there is the theoretical
302 * possibility that other threads might have consumed
303 * the freed up memory ahead of us, so doing another
304 * collection and trying again might actually help.
305 * Of course the same thing might happen again.
307 * Ideally we'd like to detect that case and loop (if
308 * we always loop we will loop endlessly in the case of
309 * OOM). What we do here is give up right away.
311 sgen_ensure_free_space (real_size);
313 return alloc_degraded (vtable, size, FALSE);
315 p = sgen_nursery_alloc (size);
317 SGEN_ASSERT (0, p, "Out of memory");
319 zero_tlab_if_necessary (p, size);
321 size_t alloc_size = 0;
323 SGEN_LOG (3, "Retire TLAB: %p-%p [%ld]", TLAB_START, TLAB_REAL_END, (long)(TLAB_REAL_END - TLAB_NEXT - size));
324 sgen_nursery_retire_region (p, available_in_tlab);
326 p = sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
328 /* See comment above in similar case. */
329 sgen_ensure_free_space (tlab_size);
331 return alloc_degraded (vtable, size, FALSE);
333 p = sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
335 SGEN_ASSERT (0, p, "Out of memory");
337 /* Allocate a new TLAB from the current nursery fragment */
338 TLAB_START = (char*)p;
339 TLAB_NEXT = TLAB_START;
340 TLAB_REAL_END = TLAB_START + alloc_size;
341 TLAB_TEMP_END = TLAB_START + MIN (SGEN_SCAN_START_SIZE, alloc_size);
343 zero_tlab_if_necessary (TLAB_START, alloc_size);
345 /* Allocate from the TLAB */
346 p = (void*)TLAB_NEXT;
348 sgen_set_nursery_scan_start ((char*)p);
351 /* Reached tlab_temp_end */
353 /* record the scan start so we can find pinned objects more easily */
354 sgen_set_nursery_scan_start ((char*)p);
355 /* we just bump tlab_temp_end as well */
356 TLAB_TEMP_END = MIN (TLAB_REAL_END, TLAB_NEXT + SGEN_SCAN_START_SIZE);
357 SGEN_LOG (5, "Expanding local alloc: %p-%p", TLAB_NEXT, TLAB_TEMP_END);
359 CANARIFY_ALLOC(p,real_size);
363 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, vtable->klass->name, size);
364 binary_protocol_alloc (p, vtable, size);
365 if (G_UNLIKELY (MONO_GC_MAJOR_OBJ_ALLOC_LARGE_ENABLED ()|| MONO_GC_NURSERY_OBJ_ALLOC_ENABLED ())) {
366 if (real_size > SGEN_MAX_SMALL_OBJ_SIZE)
367 MONO_GC_MAJOR_OBJ_ALLOC_LARGE ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
369 MONO_GC_NURSERY_OBJ_ALLOC ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
371 mono_atomic_store_seq (p, vtable);
378 mono_gc_try_alloc_obj_nolock (MonoVTable *vtable, size_t size)
383 size_t real_size = size;
387 size = ALIGN_UP (size);
388 SGEN_ASSERT (9, real_size >= sizeof (MonoObject), "Object too small");
390 g_assert (vtable->gc_descr);
391 if (real_size > SGEN_MAX_SMALL_OBJ_SIZE)
394 if (G_UNLIKELY (size > tlab_size)) {
395 /* Allocate directly from the nursery */
396 p = sgen_nursery_alloc (size);
399 sgen_set_nursery_scan_start ((char*)p);
401 /*FIXME we should use weak memory ops here. Should help specially on x86. */
402 zero_tlab_if_necessary (p, size);
404 int available_in_tlab;
406 /* tlab_next and tlab_temp_end are TLS vars so accessing them might be expensive */
408 p = (void**)TLAB_NEXT;
409 /* FIXME: handle overflow */
410 new_next = (char*)p + size;
412 real_end = TLAB_REAL_END;
413 available_in_tlab = (int)(real_end - (char*)p);//We'll never have tlabs > 2Gb
415 if (G_LIKELY (new_next < real_end)) {
416 TLAB_NEXT = new_next;
418 /* Second case, we overflowed temp end */
419 if (G_UNLIKELY (new_next >= TLAB_TEMP_END)) {
420 sgen_set_nursery_scan_start (new_next);
421 /* we just bump tlab_temp_end as well */
422 TLAB_TEMP_END = MIN (TLAB_REAL_END, TLAB_NEXT + SGEN_SCAN_START_SIZE);
423 SGEN_LOG (5, "Expanding local alloc: %p-%p", TLAB_NEXT, TLAB_TEMP_END);
425 } else if (available_in_tlab > SGEN_MAX_NURSERY_WASTE) {
426 /* Allocate directly from the nursery */
427 p = sgen_nursery_alloc (size);
431 zero_tlab_if_necessary (p, size);
433 size_t alloc_size = 0;
435 sgen_nursery_retire_region (p, available_in_tlab);
436 new_next = sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
437 p = (void**)new_next;
441 TLAB_START = (char*)new_next;
442 TLAB_NEXT = new_next + size;
443 TLAB_REAL_END = new_next + alloc_size;
444 TLAB_TEMP_END = new_next + MIN (SGEN_SCAN_START_SIZE, alloc_size);
445 sgen_set_nursery_scan_start ((char*)p);
447 zero_tlab_if_necessary (new_next, alloc_size);
449 MONO_GC_NURSERY_TLAB_ALLOC ((mword)new_next, alloc_size);
453 HEAVY_STAT (++stat_objects_alloced);
454 HEAVY_STAT (stat_bytes_alloced += size);
456 CANARIFY_ALLOC(p,real_size);
457 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, vtable->klass->name, size);
458 binary_protocol_alloc (p, vtable, size);
459 if (G_UNLIKELY (MONO_GC_NURSERY_OBJ_ALLOC_ENABLED ()))
460 MONO_GC_NURSERY_OBJ_ALLOC ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
461 g_assert (*p == NULL); /* FIXME disable this in non debug builds */
463 mono_atomic_store_seq (p, vtable);
469 mono_gc_alloc_obj (MonoVTable *vtable, size_t size)
474 if (!SGEN_CAN_ALIGN_UP (size))
477 #ifndef DISABLE_CRITICAL_REGION
479 if (G_UNLIKELY (has_per_allocation_action)) {
480 static int alloc_count;
481 int current_alloc = InterlockedIncrement (&alloc_count);
483 if (verify_before_allocs) {
484 if ((current_alloc % verify_before_allocs) == 0)
485 sgen_check_whole_heap_stw ();
487 if (collect_before_allocs) {
488 if (((current_alloc % collect_before_allocs) == 0) && nursery_section) {
490 sgen_perform_collection (0, GENERATION_NURSERY, "collect-before-alloc-triggered", TRUE);
496 ENTER_CRITICAL_REGION;
497 res = mono_gc_try_alloc_obj_nolock (vtable, size);
499 EXIT_CRITICAL_REGION;
502 EXIT_CRITICAL_REGION;
505 res = mono_gc_alloc_obj_nolock (vtable, size);
507 if (G_UNLIKELY (!res))
508 return mono_gc_out_of_memory (size);
513 mono_gc_alloc_vector (MonoVTable *vtable, size_t size, uintptr_t max_length)
518 if (!SGEN_CAN_ALIGN_UP (size))
521 #ifndef DISABLE_CRITICAL_REGION
522 ENTER_CRITICAL_REGION;
523 arr = mono_gc_try_alloc_obj_nolock (vtable, size);
525 /*This doesn't require fencing since EXIT_CRITICAL_REGION already does it for us*/
526 arr->max_length = (mono_array_size_t)max_length;
527 EXIT_CRITICAL_REGION;
530 EXIT_CRITICAL_REGION;
535 arr = mono_gc_alloc_obj_nolock (vtable, size);
536 if (G_UNLIKELY (!arr)) {
538 return mono_gc_out_of_memory (size);
541 arr->max_length = (mono_array_size_t)max_length;
549 mono_gc_alloc_array (MonoVTable *vtable, size_t size, uintptr_t max_length, uintptr_t bounds_size)
552 MonoArrayBounds *bounds;
555 if (!SGEN_CAN_ALIGN_UP (size))
558 #ifndef DISABLE_CRITICAL_REGION
559 ENTER_CRITICAL_REGION;
560 arr = mono_gc_try_alloc_obj_nolock (vtable, size);
562 /*This doesn't require fencing since EXIT_CRITICAL_REGION already does it for us*/
563 arr->max_length = (mono_array_size_t)max_length;
565 bounds = (MonoArrayBounds*)((char*)arr + size - bounds_size);
566 arr->bounds = bounds;
567 EXIT_CRITICAL_REGION;
570 EXIT_CRITICAL_REGION;
575 arr = mono_gc_alloc_obj_nolock (vtable, size);
576 if (G_UNLIKELY (!arr)) {
578 return mono_gc_out_of_memory (size);
581 arr->max_length = (mono_array_size_t)max_length;
583 bounds = (MonoArrayBounds*)((char*)arr + size - bounds_size);
584 arr->bounds = bounds;
592 mono_gc_alloc_string (MonoVTable *vtable, size_t size, gint32 len)
597 if (!SGEN_CAN_ALIGN_UP (size))
600 #ifndef DISABLE_CRITICAL_REGION
601 ENTER_CRITICAL_REGION;
602 str = mono_gc_try_alloc_obj_nolock (vtable, size);
604 /*This doesn't require fencing since EXIT_CRITICAL_REGION already does it for us*/
606 EXIT_CRITICAL_REGION;
609 EXIT_CRITICAL_REGION;
614 str = mono_gc_alloc_obj_nolock (vtable, size);
615 if (G_UNLIKELY (!str)) {
617 return mono_gc_out_of_memory (size);
628 * To be used for interned strings and possibly MonoThread, reflection handles.
629 * We may want to explicitly free these objects.
632 mono_gc_alloc_pinned_obj (MonoVTable *vtable, size_t size)
636 if (!SGEN_CAN_ALIGN_UP (size))
638 size = ALIGN_UP (size);
642 if (size > SGEN_MAX_SMALL_OBJ_SIZE) {
643 /* large objects are always pinned anyway */
644 p = sgen_los_alloc_large_inner (vtable, size);
646 SGEN_ASSERT (9, vtable->klass->inited, "class %s:%s is not initialized", vtable->klass->name_space, vtable->klass->name);
647 p = major_collector.alloc_small_pinned_obj (vtable, size, SGEN_VTABLE_HAS_REFERENCES (vtable));
650 SGEN_LOG (6, "Allocated pinned object %p, vtable: %p (%s), size: %zd", p, vtable, vtable->klass->name, size);
651 if (size > SGEN_MAX_SMALL_OBJ_SIZE)
652 MONO_GC_MAJOR_OBJ_ALLOC_LARGE ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
654 MONO_GC_MAJOR_OBJ_ALLOC_PINNED ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
655 binary_protocol_alloc_pinned (p, vtable, size);
662 mono_gc_alloc_mature (MonoVTable *vtable)
665 size_t size = vtable->klass->instance_size;
667 if (!SGEN_CAN_ALIGN_UP (size))
669 size = ALIGN_UP (size);
672 res = alloc_degraded (vtable, size, TRUE);
674 if (G_UNLIKELY (vtable->klass->has_finalize))
675 mono_object_register_finalizer ((MonoObject*)res);
681 mono_gc_alloc_fixed (size_t size, void *descr)
683 /* FIXME: do a single allocation */
684 void *res = calloc (1, size);
687 if (!mono_gc_register_root (res, size, descr)) {
695 mono_gc_free_fixed (void* addr)
697 mono_gc_deregister_root (addr);
702 sgen_init_tlab_info (SgenThreadInfo* info)
704 #ifndef HAVE_KW_THREAD
705 SgenThreadInfo *__thread_info__ = info;
708 info->tlab_start_addr = &TLAB_START;
709 info->tlab_next_addr = &TLAB_NEXT;
710 info->tlab_temp_end_addr = &TLAB_TEMP_END;
711 info->tlab_real_end_addr = &TLAB_REAL_END;
713 #ifdef HAVE_KW_THREAD
714 tlab_next_addr = &tlab_next;
719 * Clear the thread local TLAB variables for all threads.
722 sgen_clear_tlabs (void)
724 SgenThreadInfo *info;
726 FOREACH_THREAD (info) {
727 /* A new TLAB will be allocated when the thread does its first allocation */
728 *info->tlab_start_addr = NULL;
729 *info->tlab_next_addr = NULL;
730 *info->tlab_temp_end_addr = NULL;
731 *info->tlab_real_end_addr = NULL;
735 static MonoMethod* alloc_method_cache [ATYPE_NUM];
737 #ifdef MANAGED_ALLOCATION
738 /* FIXME: Do this in the JIT, where specialized allocation sequences can be created
739 * for each class. This is currently not easy to do, as it is hard to generate basic
740 * blocks + branches, but it is easy with the linear IL codebase.
742 * For this to work we'd need to solve the TLAB race, first. Now we
743 * require the allocator to be in a few known methods to make sure
744 * that they are executed atomically via the restart mechanism.
747 create_allocator (int atype)
750 guint32 slowpath_branch, max_size_branch;
751 MonoMethodBuilder *mb;
753 MonoMethodSignature *csig;
754 static gboolean registered = FALSE;
755 int tlab_next_addr_var, new_next_var;
757 const char *name = NULL;
758 AllocatorWrapperInfo *info;
760 #ifdef HAVE_KW_THREAD
761 int tlab_next_addr_offset = -1;
762 int tlab_temp_end_offset = -1;
764 MONO_THREAD_VAR_OFFSET (tlab_next_addr, tlab_next_addr_offset);
765 MONO_THREAD_VAR_OFFSET (tlab_temp_end, tlab_temp_end_offset);
767 mono_tls_key_set_offset (TLS_KEY_SGEN_TLAB_NEXT_ADDR, tlab_next_addr_offset);
768 mono_tls_key_set_offset (TLS_KEY_SGEN_TLAB_TEMP_END, tlab_temp_end_offset);
770 g_assert (tlab_next_addr_offset != -1);
771 g_assert (tlab_temp_end_offset != -1);
775 mono_register_jit_icall (mono_gc_alloc_obj, "mono_gc_alloc_obj", mono_create_icall_signature ("object ptr int"), FALSE);
776 mono_register_jit_icall (mono_gc_alloc_vector, "mono_gc_alloc_vector", mono_create_icall_signature ("object ptr int int"), FALSE);
777 mono_register_jit_icall (mono_gc_alloc_string, "mono_gc_alloc_string", mono_create_icall_signature ("object ptr int int32"), FALSE);
781 if (atype == ATYPE_SMALL) {
784 } else if (atype == ATYPE_NORMAL) {
787 } else if (atype == ATYPE_VECTOR) {
789 name = "AllocVector";
790 } else if (atype == ATYPE_STRING) {
792 name = "AllocString";
794 g_assert_not_reached ();
797 csig = mono_metadata_signature_alloc (mono_defaults.corlib, num_params);
798 if (atype == ATYPE_STRING) {
799 csig->ret = &mono_defaults.string_class->byval_arg;
800 csig->params [0] = &mono_defaults.int_class->byval_arg;
801 csig->params [1] = &mono_defaults.int32_class->byval_arg;
803 csig->ret = &mono_defaults.object_class->byval_arg;
804 for (i = 0; i < num_params; ++i)
805 csig->params [i] = &mono_defaults.int_class->byval_arg;
808 mb = mono_mb_new (mono_defaults.object_class, name, MONO_WRAPPER_ALLOC);
811 size_var = mono_mb_add_local (mb, &mono_defaults.int_class->byval_arg);
812 if (atype == ATYPE_NORMAL || atype == ATYPE_SMALL) {
813 /* size = vtable->klass->instance_size; */
814 mono_mb_emit_ldarg (mb, 0);
815 mono_mb_emit_icon (mb, MONO_STRUCT_OFFSET (MonoVTable, klass));
816 mono_mb_emit_byte (mb, CEE_ADD);
817 mono_mb_emit_byte (mb, CEE_LDIND_I);
818 mono_mb_emit_icon (mb, MONO_STRUCT_OFFSET (MonoClass, instance_size));
819 mono_mb_emit_byte (mb, CEE_ADD);
820 /* FIXME: assert instance_size stays a 4 byte integer */
821 mono_mb_emit_byte (mb, CEE_LDIND_U4);
822 mono_mb_emit_byte (mb, CEE_CONV_I);
823 mono_mb_emit_stloc (mb, size_var);
824 } else if (atype == ATYPE_VECTOR) {
825 MonoExceptionClause *clause;
826 int pos, pos_leave, pos_error;
827 MonoClass *oom_exc_class;
831 * n > MONO_ARRAY_MAX_INDEX => OutOfMemoryException
832 * n < 0 => OverflowException
834 * We can do an unsigned comparison to catch both cases, then in the error
835 * case compare signed to distinguish between them.
837 mono_mb_emit_ldarg (mb, 1);
838 mono_mb_emit_icon (mb, MONO_ARRAY_MAX_INDEX);
839 mono_mb_emit_byte (mb, CEE_CONV_U);
840 pos = mono_mb_emit_short_branch (mb, CEE_BLE_UN_S);
842 mono_mb_emit_byte (mb, MONO_CUSTOM_PREFIX);
843 mono_mb_emit_byte (mb, CEE_MONO_NOT_TAKEN);
844 mono_mb_emit_ldarg (mb, 1);
845 mono_mb_emit_icon (mb, 0);
846 pos_error = mono_mb_emit_short_branch (mb, CEE_BLT_S);
847 mono_mb_emit_exception (mb, "OutOfMemoryException", NULL);
848 mono_mb_patch_short_branch (mb, pos_error);
849 mono_mb_emit_exception (mb, "OverflowException", NULL);
851 mono_mb_patch_short_branch (mb, pos);
853 clause = mono_image_alloc0 (mono_defaults.corlib, sizeof (MonoExceptionClause));
854 clause->try_offset = mono_mb_get_label (mb);
856 /* vtable->klass->sizes.element_size */
857 mono_mb_emit_ldarg (mb, 0);
858 mono_mb_emit_icon (mb, MONO_STRUCT_OFFSET (MonoVTable, klass));
859 mono_mb_emit_byte (mb, CEE_ADD);
860 mono_mb_emit_byte (mb, CEE_LDIND_I);
861 mono_mb_emit_icon (mb, MONO_STRUCT_OFFSET (MonoClass, sizes));
862 mono_mb_emit_byte (mb, CEE_ADD);
863 mono_mb_emit_byte (mb, CEE_LDIND_U4);
864 mono_mb_emit_byte (mb, CEE_CONV_I);
867 mono_mb_emit_ldarg (mb, 1);
868 mono_mb_emit_byte (mb, CEE_MUL_OVF_UN);
869 /* + sizeof (MonoArray) */
870 mono_mb_emit_icon (mb, sizeof (MonoArray));
871 mono_mb_emit_byte (mb, CEE_ADD_OVF_UN);
872 mono_mb_emit_stloc (mb, size_var);
874 pos_leave = mono_mb_emit_branch (mb, CEE_LEAVE);
877 clause->flags = MONO_EXCEPTION_CLAUSE_NONE;
878 clause->try_len = mono_mb_get_pos (mb) - clause->try_offset;
879 clause->data.catch_class = mono_class_from_name (mono_defaults.corlib,
880 "System", "OverflowException");
881 g_assert (clause->data.catch_class);
882 clause->handler_offset = mono_mb_get_label (mb);
884 oom_exc_class = mono_class_from_name (mono_defaults.corlib,
885 "System", "OutOfMemoryException");
886 g_assert (oom_exc_class);
887 ctor = mono_class_get_method_from_name (oom_exc_class, ".ctor", 0);
890 mono_mb_emit_byte (mb, CEE_POP);
891 mono_mb_emit_op (mb, CEE_NEWOBJ, ctor);
892 mono_mb_emit_byte (mb, CEE_THROW);
894 clause->handler_len = mono_mb_get_pos (mb) - clause->handler_offset;
895 mono_mb_set_clauses (mb, 1, clause);
896 mono_mb_patch_branch (mb, pos_leave);
898 } else if (atype == ATYPE_STRING) {
902 * a string allocator method takes the args: (vtable, len)
904 * bytes = offsetof (MonoString, chars) + ((len + 1) * 2)
908 * bytes <= INT32_MAX - (SGEN_ALLOC_ALIGN - 1)
912 * offsetof (MonoString, chars) + ((len + 1) * 2) <= INT32_MAX - (SGEN_ALLOC_ALIGN - 1)
913 * len <= (INT32_MAX - (SGEN_ALLOC_ALIGN - 1) - offsetof (MonoString, chars)) / 2 - 1
915 mono_mb_emit_ldarg (mb, 1);
916 mono_mb_emit_icon (mb, (INT32_MAX - (SGEN_ALLOC_ALIGN - 1) - MONO_STRUCT_OFFSET (MonoString, chars)) / 2 - 1);
917 pos = mono_mb_emit_short_branch (mb, MONO_CEE_BLE_UN_S);
919 mono_mb_emit_byte (mb, MONO_CUSTOM_PREFIX);
920 mono_mb_emit_byte (mb, CEE_MONO_NOT_TAKEN);
921 mono_mb_emit_exception (mb, "OutOfMemoryException", NULL);
922 mono_mb_patch_short_branch (mb, pos);
924 mono_mb_emit_ldarg (mb, 1);
925 mono_mb_emit_icon (mb, 1);
926 mono_mb_emit_byte (mb, MONO_CEE_SHL);
927 //WE manually fold the above + 2 here
928 mono_mb_emit_icon (mb, MONO_STRUCT_OFFSET (MonoString, chars) + 2);
929 mono_mb_emit_byte (mb, CEE_ADD);
930 mono_mb_emit_stloc (mb, size_var);
932 g_assert_not_reached ();
935 /* size += ALLOC_ALIGN - 1; */
936 mono_mb_emit_ldloc (mb, size_var);
937 mono_mb_emit_icon (mb, ALLOC_ALIGN - 1);
938 mono_mb_emit_byte (mb, CEE_ADD);
939 /* size &= ~(ALLOC_ALIGN - 1); */
940 mono_mb_emit_icon (mb, ~(ALLOC_ALIGN - 1));
941 mono_mb_emit_byte (mb, CEE_AND);
942 mono_mb_emit_stloc (mb, size_var);
944 /* if (size > MAX_SMALL_OBJ_SIZE) goto slowpath */
945 if (atype != ATYPE_SMALL) {
946 mono_mb_emit_ldloc (mb, size_var);
947 mono_mb_emit_icon (mb, MAX_SMALL_OBJ_SIZE);
948 max_size_branch = mono_mb_emit_short_branch (mb, MONO_CEE_BGT_UN_S);
952 * We need to modify tlab_next, but the JIT only supports reading, so we read
953 * another tls var holding its address instead.
956 /* tlab_next_addr (local) = tlab_next_addr (TLS var) */
957 tlab_next_addr_var = mono_mb_add_local (mb, &mono_defaults.int_class->byval_arg);
958 EMIT_TLS_ACCESS_NEXT_ADDR (mb);
959 mono_mb_emit_stloc (mb, tlab_next_addr_var);
961 /* p = (void**)tlab_next; */
962 p_var = mono_mb_add_local (mb, &mono_defaults.int_class->byval_arg);
963 mono_mb_emit_ldloc (mb, tlab_next_addr_var);
964 mono_mb_emit_byte (mb, CEE_LDIND_I);
965 mono_mb_emit_stloc (mb, p_var);
967 /* new_next = (char*)p + size; */
968 new_next_var = mono_mb_add_local (mb, &mono_defaults.int_class->byval_arg);
969 mono_mb_emit_ldloc (mb, p_var);
970 mono_mb_emit_ldloc (mb, size_var);
971 mono_mb_emit_byte (mb, CEE_CONV_I);
972 mono_mb_emit_byte (mb, CEE_ADD);
973 mono_mb_emit_stloc (mb, new_next_var);
975 /* if (G_LIKELY (new_next < tlab_temp_end)) */
976 mono_mb_emit_ldloc (mb, new_next_var);
977 EMIT_TLS_ACCESS_TEMP_END (mb);
978 slowpath_branch = mono_mb_emit_short_branch (mb, MONO_CEE_BLT_UN_S);
981 if (atype != ATYPE_SMALL)
982 mono_mb_patch_short_branch (mb, max_size_branch);
984 mono_mb_emit_byte (mb, MONO_CUSTOM_PREFIX);
985 mono_mb_emit_byte (mb, CEE_MONO_NOT_TAKEN);
987 /* FIXME: mono_gc_alloc_obj takes a 'size_t' as an argument, not an int32 */
988 mono_mb_emit_ldarg (mb, 0);
989 mono_mb_emit_ldloc (mb, size_var);
990 if (atype == ATYPE_NORMAL || atype == ATYPE_SMALL) {
991 mono_mb_emit_icall (mb, mono_gc_alloc_obj);
992 } else if (atype == ATYPE_VECTOR) {
993 mono_mb_emit_ldarg (mb, 1);
994 mono_mb_emit_icall (mb, mono_gc_alloc_vector);
995 } else if (atype == ATYPE_STRING) {
996 mono_mb_emit_ldarg (mb, 1);
997 mono_mb_emit_icall (mb, mono_gc_alloc_string);
999 g_assert_not_reached ();
1001 mono_mb_emit_byte (mb, CEE_RET);
1004 mono_mb_patch_short_branch (mb, slowpath_branch);
1006 /* FIXME: Memory barrier */
1008 /* tlab_next = new_next */
1009 mono_mb_emit_ldloc (mb, tlab_next_addr_var);
1010 mono_mb_emit_ldloc (mb, new_next_var);
1011 mono_mb_emit_byte (mb, CEE_STIND_I);
1013 /*The tlab store must be visible before the the vtable store. This could be replaced with a DDS but doing it with IL would be tricky. */
1014 mono_mb_emit_byte ((mb), MONO_CUSTOM_PREFIX);
1015 mono_mb_emit_op (mb, CEE_MONO_MEMORY_BARRIER, (gpointer)StoreStoreBarrier);
1018 mono_mb_emit_ldloc (mb, p_var);
1019 mono_mb_emit_ldarg (mb, 0);
1020 mono_mb_emit_byte (mb, CEE_STIND_I);
1022 if (atype == ATYPE_VECTOR) {
1023 /* arr->max_length = max_length; */
1024 mono_mb_emit_ldloc (mb, p_var);
1025 mono_mb_emit_ldflda (mb, MONO_STRUCT_OFFSET (MonoArray, max_length));
1026 mono_mb_emit_ldarg (mb, 1);
1027 #ifdef MONO_BIG_ARRAYS
1028 mono_mb_emit_byte (mb, CEE_STIND_I);
1030 mono_mb_emit_byte (mb, CEE_STIND_I4);
1032 } else if (atype == ATYPE_STRING) {
1033 /* need to set length and clear the last char */
1034 /* s->length = len; */
1035 mono_mb_emit_ldloc (mb, p_var);
1036 mono_mb_emit_icon (mb, MONO_STRUCT_OFFSET (MonoString, length));
1037 mono_mb_emit_byte (mb, MONO_CEE_ADD);
1038 mono_mb_emit_ldarg (mb, 1);
1039 mono_mb_emit_byte (mb, MONO_CEE_STIND_I4);
1040 /* s->chars [len] = 0; */
1041 mono_mb_emit_ldloc (mb, p_var);
1042 mono_mb_emit_ldloc (mb, size_var);
1043 mono_mb_emit_icon (mb, 2);
1044 mono_mb_emit_byte (mb, MONO_CEE_SUB);
1045 mono_mb_emit_byte (mb, MONO_CEE_ADD);
1046 mono_mb_emit_icon (mb, 0);
1047 mono_mb_emit_byte (mb, MONO_CEE_STIND_I2);
1051 We must make sure both vtable and max_length are globaly visible before returning to managed land.
1053 mono_mb_emit_byte ((mb), MONO_CUSTOM_PREFIX);
1054 mono_mb_emit_op (mb, CEE_MONO_MEMORY_BARRIER, (gpointer)StoreStoreBarrier);
1057 mono_mb_emit_ldloc (mb, p_var);
1058 mono_mb_emit_byte (mb, CEE_RET);
1061 res = mono_mb_create_method (mb, csig, 8);
1063 mono_method_get_header (res)->init_locals = FALSE;
1065 info = mono_image_alloc0 (mono_defaults.corlib, sizeof (AllocatorWrapperInfo));
1066 info->gc_name = "sgen";
1067 info->alloc_type = atype;
1068 mono_marshal_set_wrapper_info (res, info);
1075 * Generate an allocator method implementing the fast path of mono_gc_alloc_obj ().
1076 * The signature of the called method is:
1077 * object allocate (MonoVTable *vtable)
1080 mono_gc_get_managed_allocator (MonoClass *klass, gboolean for_box)
1082 #ifdef MANAGED_ALLOCATION
1084 #ifdef HAVE_KW_THREAD
1085 int tlab_next_offset = -1;
1086 int tlab_temp_end_offset = -1;
1087 MONO_THREAD_VAR_OFFSET (tlab_next, tlab_next_offset);
1088 MONO_THREAD_VAR_OFFSET (tlab_temp_end, tlab_temp_end_offset);
1090 if (tlab_next_offset == -1 || tlab_temp_end_offset == -1)
1093 if (collect_before_allocs)
1095 if (!mono_runtime_has_tls_get ())
1097 if (klass->instance_size > tlab_size)
1100 if (klass->has_finalize || mono_class_is_marshalbyref (klass) || (mono_profiler_get_events () & MONO_PROFILE_ALLOCATIONS))
1104 if (klass->byval_arg.type == MONO_TYPE_STRING)
1105 return mono_gc_get_managed_allocator_by_type (ATYPE_STRING);
1106 /* Generic classes have dynamic field and can go above MAX_SMALL_OBJ_SIZE. */
1107 if (ALIGN_TO (klass->instance_size, ALLOC_ALIGN) < MAX_SMALL_OBJ_SIZE && !mono_class_is_open_constructed_type (&klass->byval_arg))
1108 return mono_gc_get_managed_allocator_by_type (ATYPE_SMALL);
1110 return mono_gc_get_managed_allocator_by_type (ATYPE_NORMAL);
1117 mono_gc_get_managed_array_allocator (MonoClass *klass)
1119 #ifdef MANAGED_ALLOCATION
1120 #ifdef HAVE_KW_THREAD
1121 int tlab_next_offset = -1;
1122 int tlab_temp_end_offset = -1;
1123 MONO_THREAD_VAR_OFFSET (tlab_next, tlab_next_offset);
1124 MONO_THREAD_VAR_OFFSET (tlab_temp_end, tlab_temp_end_offset);
1126 if (tlab_next_offset == -1 || tlab_temp_end_offset == -1)
1130 if (klass->rank != 1)
1132 if (!mono_runtime_has_tls_get ())
1134 if (mono_profiler_get_events () & MONO_PROFILE_ALLOCATIONS)
1136 if (has_per_allocation_action)
1138 g_assert (!mono_class_has_finalizer (klass) && !mono_class_is_marshalbyref (klass));
1140 return mono_gc_get_managed_allocator_by_type (ATYPE_VECTOR);
1147 sgen_set_use_managed_allocator (gboolean flag)
1149 use_managed_allocator = flag;
1153 mono_gc_get_managed_allocator_by_type (int atype)
1155 #ifdef MANAGED_ALLOCATION
1158 if (!use_managed_allocator)
1161 if (!mono_runtime_has_tls_get ())
1164 res = alloc_method_cache [atype];
1168 res = create_allocator (atype);
1170 if (alloc_method_cache [atype]) {
1171 mono_free_method (res);
1172 res = alloc_method_cache [atype];
1174 mono_memory_barrier ();
1175 alloc_method_cache [atype] = res;
1186 mono_gc_get_managed_allocator_types (void)
1192 sgen_is_managed_allocator (MonoMethod *method)
1196 for (i = 0; i < ATYPE_NUM; ++i)
1197 if (method == alloc_method_cache [i])
1203 sgen_has_managed_allocator (void)
1207 for (i = 0; i < ATYPE_NUM; ++i)
1208 if (alloc_method_cache [i])
1213 #ifdef HEAVY_STATISTICS
1215 sgen_alloc_init_heavy_stats (void)
1217 mono_counters_register ("# objects allocated", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_objects_alloced);
1218 mono_counters_register ("bytes allocated", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_bytes_alloced);
1219 mono_counters_register ("bytes allocated in LOS", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_bytes_alloced_los);
1223 #endif /*HAVE_SGEN_GC*/