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 "utils/mono-memory-model.h"
49 #include "utils/mono-counters.h"
51 #define ALIGN_UP SGEN_ALIGN_UP
52 #define ALLOC_ALIGN SGEN_ALLOC_ALIGN
53 #define ALLOC_ALIGN_BITS SGEN_ALLOC_ALIGN_BITS
54 #define MAX_SMALL_OBJ_SIZE SGEN_MAX_SMALL_OBJ_SIZE
55 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
57 #define OPDEF(a,b,c,d,e,f,g,h,i,j) \
61 #include "mono/cil/opcode.def"
67 static gboolean use_managed_allocator = TRUE;
69 #ifdef HEAVY_STATISTICS
70 static long long stat_objects_alloced = 0;
71 static long long stat_bytes_alloced = 0;
72 static long long stat_bytes_alloced_los = 0;
77 * Allocation is done from a Thread Local Allocation Buffer (TLAB). TLABs are allocated
78 * from nursery fragments.
79 * tlab_next is the pointer to the space inside the TLAB where the next object will
81 * tlab_temp_end is the pointer to the end of the temporary space reserved for
82 * the allocation: it allows us to set the scan starts at reasonable intervals.
83 * tlab_real_end points to the end of the TLAB.
87 * FIXME: What is faster, a TLS variable pointing to a structure, or separate TLS
88 * variables for next+temp_end ?
91 static __thread char *tlab_start;
92 static __thread char *tlab_next;
93 static __thread char *tlab_temp_end;
94 static __thread char *tlab_real_end;
95 /* Used by the managed allocator/wbarrier */
96 static __thread char **tlab_next_addr;
100 #define TLAB_START tlab_start
101 #define TLAB_NEXT tlab_next
102 #define TLAB_TEMP_END tlab_temp_end
103 #define TLAB_REAL_END tlab_real_end
105 #define TLAB_START (__thread_info__->tlab_start)
106 #define TLAB_NEXT (__thread_info__->tlab_next)
107 #define TLAB_TEMP_END (__thread_info__->tlab_temp_end)
108 #define TLAB_REAL_END (__thread_info__->tlab_real_end)
112 alloc_degraded (MonoVTable *vtable, size_t size, gboolean for_mature)
114 static int last_major_gc_warned = -1;
115 static int num_degraded = 0;
120 if (last_major_gc_warned < stat_major_gcs) {
122 if (num_degraded == 1 || num_degraded == 3)
123 fprintf (stderr, "Warning: Degraded allocation. Consider increasing nursery-size if the warning persists.\n");
124 else if (num_degraded == 10)
125 fprintf (stderr, "Warning: Repeated degraded allocation. Consider increasing nursery-size.\n");
126 last_major_gc_warned = stat_major_gcs;
128 InterlockedExchangeAdd (°raded_mode, size);
129 sgen_ensure_free_space (size);
131 if (sgen_need_major_collection (size))
132 sgen_perform_collection (size, GENERATION_OLD, "mature allocation failure");
136 p = major_collector.alloc_degraded (vtable, size);
139 MONO_GC_MAJOR_OBJ_ALLOC_MATURE ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
141 binary_protocol_alloc_degraded (p, vtable, size);
142 MONO_GC_MAJOR_OBJ_ALLOC_DEGRADED ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
149 * Provide a variant that takes just the vtable for small fixed-size objects.
150 * The aligned size is already computed and stored in vt->gc_descr.
151 * Note: every SGEN_SCAN_START_SIZE or so we are given the chance to do some special
152 * processing. We can keep track of where objects start, for example,
153 * so when we scan the thread stacks for pinned objects, we can start
154 * a search for the pinned object in SGEN_SCAN_START_SIZE chunks.
157 mono_gc_alloc_obj_nolock (MonoVTable *vtable, size_t size)
159 /* FIXME: handle OOM */
164 HEAVY_STAT (++stat_objects_alloced);
165 if (size <= SGEN_MAX_SMALL_OBJ_SIZE)
166 HEAVY_STAT (stat_bytes_alloced += size);
168 HEAVY_STAT (stat_bytes_alloced_los += size);
170 size = ALIGN_UP (size);
172 g_assert (vtable->gc_descr);
174 if (G_UNLIKELY (has_per_allocation_action)) {
175 static int alloc_count;
176 int current_alloc = InterlockedIncrement (&alloc_count);
178 if (collect_before_allocs) {
179 if (((current_alloc % collect_before_allocs) == 0) && nursery_section) {
180 sgen_perform_collection (0, GENERATION_NURSERY, "collect-before-alloc-triggered");
181 if (!degraded_mode && sgen_can_alloc_size (size) && size <= SGEN_MAX_SMALL_OBJ_SIZE) {
183 g_assert_not_reached ();
186 } else if (verify_before_allocs) {
187 if ((current_alloc % verify_before_allocs) == 0)
188 sgen_check_whole_heap_stw ();
193 * We must already have the lock here instead of after the
194 * fast path because we might be interrupted in the fast path
195 * (after confirming that new_next < TLAB_TEMP_END) by the GC,
196 * and we'll end up allocating an object in a fragment which
197 * no longer belongs to us.
199 * The managed allocator does not do this, but it's treated
200 * specially by the world-stopping code.
203 if (size > SGEN_MAX_SMALL_OBJ_SIZE) {
204 p = sgen_los_alloc_large_inner (vtable, size);
206 /* tlab_next and tlab_temp_end are TLS vars so accessing them might be expensive */
208 p = (void**)TLAB_NEXT;
209 /* FIXME: handle overflow */
210 new_next = (char*)p + size;
211 TLAB_NEXT = new_next;
213 if (G_LIKELY (new_next < TLAB_TEMP_END)) {
217 * FIXME: We might need a memory barrier here so the change to tlab_next is
218 * visible before the vtable store.
221 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, vtable->klass->name, size);
222 binary_protocol_alloc (p , vtable, size);
223 if (G_UNLIKELY (MONO_GC_NURSERY_OBJ_ALLOC_ENABLED ()))
224 MONO_GC_NURSERY_OBJ_ALLOC ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
225 g_assert (*p == NULL);
226 mono_atomic_store_seq (p, vtable);
233 /* there are two cases: the object is too big or we run out of space in the TLAB */
234 /* we also reach here when the thread does its first allocation after a minor
235 * collection, since the tlab_ variables are initialized to NULL.
236 * there can be another case (from ORP), if we cooperate with the runtime a bit:
237 * objects that need finalizers can have the high bit set in their size
238 * so the above check fails and we can readily add the object to the queue.
239 * This avoids taking again the GC lock when registering, but this is moot when
240 * doing thread-local allocation, so it may not be a good idea.
242 if (TLAB_NEXT >= TLAB_REAL_END) {
243 int available_in_tlab;
245 * Run out of space in the TLAB. When this happens, some amount of space
246 * remains in the TLAB, but not enough to satisfy the current allocation
247 * request. Currently, we retire the TLAB in all cases, later we could
248 * keep it if the remaining space is above a treshold, and satisfy the
249 * allocation directly from the nursery.
252 /* when running in degraded mode, we continue allocing that way
253 * for a while, to decrease the number of useless nursery collections.
255 if (degraded_mode && degraded_mode < DEFAULT_NURSERY_SIZE)
256 return alloc_degraded (vtable, size, FALSE);
258 available_in_tlab = TLAB_REAL_END - TLAB_NEXT;
259 if (size > tlab_size || available_in_tlab > SGEN_MAX_NURSERY_WASTE) {
260 /* Allocate directly from the nursery */
262 p = sgen_nursery_alloc (size);
264 sgen_ensure_free_space (size);
266 return alloc_degraded (vtable, size, FALSE);
268 p = sgen_nursery_alloc (size);
276 if (nursery_clear_policy == CLEAR_AT_TLAB_CREATION) {
280 size_t alloc_size = 0;
282 SGEN_LOG (3, "Retire TLAB: %p-%p [%ld]", TLAB_START, TLAB_REAL_END, (long)(TLAB_REAL_END - TLAB_NEXT - size));
283 sgen_nursery_retire_region (p, available_in_tlab);
286 p = sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
288 sgen_ensure_free_space (tlab_size);
290 return alloc_degraded (vtable, size, FALSE);
292 p = sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
301 /* Allocate a new TLAB from the current nursery fragment */
302 TLAB_START = (char*)p;
303 TLAB_NEXT = TLAB_START;
304 TLAB_REAL_END = TLAB_START + alloc_size;
305 TLAB_TEMP_END = TLAB_START + MIN (SGEN_SCAN_START_SIZE, alloc_size);
307 if (nursery_clear_policy == CLEAR_AT_TLAB_CREATION) {
308 memset (TLAB_START, 0, alloc_size);
311 /* Allocate from the TLAB */
312 p = (void*)TLAB_NEXT;
314 sgen_set_nursery_scan_start ((char*)p);
317 /* Reached tlab_temp_end */
319 /* record the scan start so we can find pinned objects more easily */
320 sgen_set_nursery_scan_start ((char*)p);
321 /* we just bump tlab_temp_end as well */
322 TLAB_TEMP_END = MIN (TLAB_REAL_END, TLAB_NEXT + SGEN_SCAN_START_SIZE);
323 SGEN_LOG (5, "Expanding local alloc: %p-%p", TLAB_NEXT, TLAB_TEMP_END);
328 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, vtable->klass->name, size);
329 binary_protocol_alloc (p, vtable, size);
330 if (G_UNLIKELY (MONO_GC_MAJOR_OBJ_ALLOC_LARGE_ENABLED ()|| MONO_GC_NURSERY_OBJ_ALLOC_ENABLED ())) {
331 if (size > SGEN_MAX_SMALL_OBJ_SIZE)
332 MONO_GC_MAJOR_OBJ_ALLOC_LARGE ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
334 MONO_GC_NURSERY_OBJ_ALLOC ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
336 mono_atomic_store_seq (p, vtable);
343 mono_gc_try_alloc_obj_nolock (MonoVTable *vtable, size_t size)
349 size = ALIGN_UP (size);
351 g_assert (vtable->gc_descr);
352 if (size > SGEN_MAX_SMALL_OBJ_SIZE)
355 if (G_UNLIKELY (size > tlab_size)) {
356 /* Allocate directly from the nursery */
357 p = sgen_nursery_alloc (size);
360 sgen_set_nursery_scan_start ((char*)p);
362 /*FIXME we should use weak memory ops here. Should help specially on x86. */
363 if (nursery_clear_policy == CLEAR_AT_TLAB_CREATION)
366 int available_in_tlab;
368 /* tlab_next and tlab_temp_end are TLS vars so accessing them might be expensive */
370 p = (void**)TLAB_NEXT;
371 /* FIXME: handle overflow */
372 new_next = (char*)p + size;
374 real_end = TLAB_REAL_END;
375 available_in_tlab = real_end - (char*)p;
377 if (G_LIKELY (new_next < real_end)) {
378 TLAB_NEXT = new_next;
380 /* Second case, we overflowed temp end */
381 if (G_UNLIKELY (new_next >= TLAB_TEMP_END)) {
382 sgen_set_nursery_scan_start (new_next);
383 /* we just bump tlab_temp_end as well */
384 TLAB_TEMP_END = MIN (TLAB_REAL_END, TLAB_NEXT + SGEN_SCAN_START_SIZE);
385 SGEN_LOG (5, "Expanding local alloc: %p-%p", TLAB_NEXT, TLAB_TEMP_END);
387 } else if (available_in_tlab > SGEN_MAX_NURSERY_WASTE) {
388 /* Allocate directly from the nursery */
389 p = sgen_nursery_alloc (size);
393 if (nursery_clear_policy == CLEAR_AT_TLAB_CREATION)
396 size_t alloc_size = 0;
398 sgen_nursery_retire_region (p, available_in_tlab);
399 new_next = sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
400 p = (void**)new_next;
404 TLAB_START = (char*)new_next;
405 TLAB_NEXT = new_next + size;
406 TLAB_REAL_END = new_next + alloc_size;
407 TLAB_TEMP_END = new_next + MIN (SGEN_SCAN_START_SIZE, alloc_size);
408 sgen_set_nursery_scan_start ((char*)p);
410 if (nursery_clear_policy == CLEAR_AT_TLAB_CREATION)
411 memset (new_next, 0, alloc_size);
413 MONO_GC_NURSERY_TLAB_ALLOC ((mword)new_next, alloc_size);
417 HEAVY_STAT (++stat_objects_alloced);
418 HEAVY_STAT (stat_bytes_alloced += size);
420 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, vtable->klass->name, size);
421 binary_protocol_alloc (p, vtable, size);
422 if (G_UNLIKELY (MONO_GC_NURSERY_OBJ_ALLOC_ENABLED ()))
423 MONO_GC_NURSERY_OBJ_ALLOC ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
424 g_assert (*p == NULL); /* FIXME disable this in non debug builds */
426 mono_atomic_store_seq (p, vtable);
432 mono_gc_alloc_obj (MonoVTable *vtable, size_t size)
435 #ifndef DISABLE_CRITICAL_REGION
437 ENTER_CRITICAL_REGION;
438 res = mono_gc_try_alloc_obj_nolock (vtable, size);
440 EXIT_CRITICAL_REGION;
443 EXIT_CRITICAL_REGION;
446 res = mono_gc_alloc_obj_nolock (vtable, size);
448 if (G_UNLIKELY (!res))
449 return mono_gc_out_of_memory (size);
454 mono_gc_alloc_vector (MonoVTable *vtable, size_t size, uintptr_t max_length)
457 #ifndef DISABLE_CRITICAL_REGION
459 ENTER_CRITICAL_REGION;
460 arr = mono_gc_try_alloc_obj_nolock (vtable, size);
462 /*This doesn't require fencing since EXIT_CRITICAL_REGION already does it for us*/
463 arr->max_length = max_length;
464 EXIT_CRITICAL_REGION;
467 EXIT_CRITICAL_REGION;
472 arr = mono_gc_alloc_obj_nolock (vtable, size);
473 if (G_UNLIKELY (!arr)) {
475 return mono_gc_out_of_memory (size);
478 arr->max_length = max_length;
486 mono_gc_alloc_array (MonoVTable *vtable, size_t size, uintptr_t max_length, uintptr_t bounds_size)
489 MonoArrayBounds *bounds;
491 #ifndef DISABLE_CRITICAL_REGION
493 ENTER_CRITICAL_REGION;
494 arr = mono_gc_try_alloc_obj_nolock (vtable, size);
496 /*This doesn't require fencing since EXIT_CRITICAL_REGION already does it for us*/
497 arr->max_length = max_length;
499 bounds = (MonoArrayBounds*)((char*)arr + size - bounds_size);
500 arr->bounds = bounds;
501 EXIT_CRITICAL_REGION;
504 EXIT_CRITICAL_REGION;
509 arr = mono_gc_alloc_obj_nolock (vtable, size);
510 if (G_UNLIKELY (!arr)) {
512 return mono_gc_out_of_memory (size);
515 arr->max_length = max_length;
517 bounds = (MonoArrayBounds*)((char*)arr + size - bounds_size);
518 arr->bounds = bounds;
526 mono_gc_alloc_string (MonoVTable *vtable, size_t size, gint32 len)
529 #ifndef DISABLE_CRITICAL_REGION
531 ENTER_CRITICAL_REGION;
532 str = mono_gc_try_alloc_obj_nolock (vtable, size);
534 /*This doesn't require fencing since EXIT_CRITICAL_REGION already does it for us*/
536 EXIT_CRITICAL_REGION;
539 EXIT_CRITICAL_REGION;
544 str = mono_gc_alloc_obj_nolock (vtable, size);
545 if (G_UNLIKELY (!str)) {
547 return mono_gc_out_of_memory (size);
558 * To be used for interned strings and possibly MonoThread, reflection handles.
559 * We may want to explicitly free these objects.
562 mono_gc_alloc_pinned_obj (MonoVTable *vtable, size_t size)
565 size = ALIGN_UP (size);
568 if (size > SGEN_MAX_SMALL_OBJ_SIZE) {
569 /* large objects are always pinned anyway */
570 p = sgen_los_alloc_large_inner (vtable, size);
572 SGEN_ASSERT (9, vtable->klass->inited, "class %s:%s is not initialized", vtable->klass->name_space, vtable->klass->name);
573 p = major_collector.alloc_small_pinned_obj (size, SGEN_VTABLE_HAS_REFERENCES (vtable));
576 SGEN_LOG (6, "Allocated pinned object %p, vtable: %p (%s), size: %zd", p, vtable, vtable->klass->name, size);
577 if (size > SGEN_MAX_SMALL_OBJ_SIZE)
578 MONO_GC_MAJOR_OBJ_ALLOC_LARGE ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
580 MONO_GC_MAJOR_OBJ_ALLOC_PINNED ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
581 binary_protocol_alloc_pinned (p, vtable, size);
582 mono_atomic_store_seq (p, vtable);
589 mono_gc_alloc_mature (MonoVTable *vtable)
592 size_t size = ALIGN_UP (vtable->klass->instance_size);
594 res = alloc_degraded (vtable, size, TRUE);
595 mono_atomic_store_seq (res, vtable);
597 if (G_UNLIKELY (vtable->klass->has_finalize))
598 mono_object_register_finalizer ((MonoObject*)res);
604 mono_gc_alloc_fixed (size_t size, void *descr)
606 /* FIXME: do a single allocation */
607 void *res = calloc (1, size);
610 if (!mono_gc_register_root (res, size, descr)) {
618 mono_gc_free_fixed (void* addr)
620 mono_gc_deregister_root (addr);
625 sgen_init_tlab_info (SgenThreadInfo* info)
627 #ifndef HAVE_KW_THREAD
628 SgenThreadInfo *__thread_info__ = info;
631 info->tlab_start_addr = &TLAB_START;
632 info->tlab_next_addr = &TLAB_NEXT;
633 info->tlab_temp_end_addr = &TLAB_TEMP_END;
634 info->tlab_real_end_addr = &TLAB_REAL_END;
636 #ifdef HAVE_KW_THREAD
637 tlab_next_addr = &tlab_next;
642 * Clear the thread local TLAB variables for all threads.
645 sgen_clear_tlabs (void)
647 SgenThreadInfo *info;
649 FOREACH_THREAD (info) {
650 /* A new TLAB will be allocated when the thread does its first allocation */
651 *info->tlab_start_addr = NULL;
652 *info->tlab_next_addr = NULL;
653 *info->tlab_temp_end_addr = NULL;
654 *info->tlab_real_end_addr = NULL;
658 static MonoMethod* alloc_method_cache [ATYPE_NUM];
660 #ifdef MANAGED_ALLOCATION
661 /* FIXME: Do this in the JIT, where specialized allocation sequences can be created
662 * for each class. This is currently not easy to do, as it is hard to generate basic
663 * blocks + branches, but it is easy with the linear IL codebase.
665 * For this to work we'd need to solve the TLAB race, first. Now we
666 * require the allocator to be in a few known methods to make sure
667 * that they are executed atomically via the restart mechanism.
670 create_allocator (int atype)
673 guint32 slowpath_branch, max_size_branch;
674 MonoMethodBuilder *mb;
676 MonoMethodSignature *csig;
677 static gboolean registered = FALSE;
678 int tlab_next_addr_var, new_next_var;
680 const char *name = NULL;
681 AllocatorWrapperInfo *info;
683 #ifdef HAVE_KW_THREAD
684 int tlab_next_addr_offset = -1;
685 int tlab_temp_end_offset = -1;
687 MONO_THREAD_VAR_OFFSET (tlab_next_addr, tlab_next_addr_offset);
688 MONO_THREAD_VAR_OFFSET (tlab_temp_end, tlab_temp_end_offset);
690 g_assert (tlab_next_addr_offset != -1);
691 g_assert (tlab_temp_end_offset != -1);
695 mono_register_jit_icall (mono_gc_alloc_obj, "mono_gc_alloc_obj", mono_create_icall_signature ("object ptr int"), FALSE);
696 mono_register_jit_icall (mono_gc_alloc_vector, "mono_gc_alloc_vector", mono_create_icall_signature ("object ptr int int"), FALSE);
700 if (atype == ATYPE_SMALL) {
703 } else if (atype == ATYPE_NORMAL) {
706 } else if (atype == ATYPE_VECTOR) {
708 name = "AllocVector";
710 g_assert_not_reached ();
713 csig = mono_metadata_signature_alloc (mono_defaults.corlib, num_params);
714 csig->ret = &mono_defaults.object_class->byval_arg;
715 for (i = 0; i < num_params; ++i)
716 csig->params [i] = &mono_defaults.int_class->byval_arg;
718 mb = mono_mb_new (mono_defaults.object_class, name, MONO_WRAPPER_ALLOC);
719 size_var = mono_mb_add_local (mb, &mono_defaults.int32_class->byval_arg);
720 if (atype == ATYPE_NORMAL || atype == ATYPE_SMALL) {
721 /* size = vtable->klass->instance_size; */
722 mono_mb_emit_ldarg (mb, 0);
723 mono_mb_emit_icon (mb, G_STRUCT_OFFSET (MonoVTable, klass));
724 mono_mb_emit_byte (mb, CEE_ADD);
725 mono_mb_emit_byte (mb, CEE_LDIND_I);
726 mono_mb_emit_icon (mb, G_STRUCT_OFFSET (MonoClass, instance_size));
727 mono_mb_emit_byte (mb, CEE_ADD);
728 /* FIXME: assert instance_size stays a 4 byte integer */
729 mono_mb_emit_byte (mb, CEE_LDIND_U4);
730 mono_mb_emit_stloc (mb, size_var);
731 } else if (atype == ATYPE_VECTOR) {
732 MonoExceptionClause *clause;
734 MonoClass *oom_exc_class;
737 /* n > MONO_ARRAY_MAX_INDEX -> OverflowException */
738 mono_mb_emit_ldarg (mb, 1);
739 mono_mb_emit_icon (mb, MONO_ARRAY_MAX_INDEX);
740 pos = mono_mb_emit_short_branch (mb, CEE_BLE_UN_S);
741 mono_mb_emit_exception (mb, "OverflowException", NULL);
742 mono_mb_patch_short_branch (mb, pos);
744 clause = mono_image_alloc0 (mono_defaults.corlib, sizeof (MonoExceptionClause));
745 clause->try_offset = mono_mb_get_label (mb);
747 /* vtable->klass->sizes.element_size */
748 mono_mb_emit_ldarg (mb, 0);
749 mono_mb_emit_icon (mb, G_STRUCT_OFFSET (MonoVTable, klass));
750 mono_mb_emit_byte (mb, CEE_ADD);
751 mono_mb_emit_byte (mb, CEE_LDIND_I);
752 mono_mb_emit_icon (mb, G_STRUCT_OFFSET (MonoClass, sizes.element_size));
753 mono_mb_emit_byte (mb, CEE_ADD);
754 mono_mb_emit_byte (mb, CEE_LDIND_U4);
757 mono_mb_emit_ldarg (mb, 1);
758 mono_mb_emit_byte (mb, CEE_MUL_OVF_UN);
759 /* + sizeof (MonoArray) */
760 mono_mb_emit_icon (mb, sizeof (MonoArray));
761 mono_mb_emit_byte (mb, CEE_ADD_OVF_UN);
762 mono_mb_emit_stloc (mb, size_var);
764 pos_leave = mono_mb_emit_branch (mb, CEE_LEAVE);
767 clause->flags = MONO_EXCEPTION_CLAUSE_NONE;
768 clause->try_len = mono_mb_get_pos (mb) - clause->try_offset;
769 clause->data.catch_class = mono_class_from_name (mono_defaults.corlib,
770 "System", "OverflowException");
771 g_assert (clause->data.catch_class);
772 clause->handler_offset = mono_mb_get_label (mb);
774 oom_exc_class = mono_class_from_name (mono_defaults.corlib,
775 "System", "OutOfMemoryException");
776 g_assert (oom_exc_class);
777 ctor = mono_class_get_method_from_name (oom_exc_class, ".ctor", 0);
780 mono_mb_emit_byte (mb, CEE_POP);
781 mono_mb_emit_op (mb, CEE_NEWOBJ, ctor);
782 mono_mb_emit_byte (mb, CEE_THROW);
784 clause->handler_len = mono_mb_get_pos (mb) - clause->handler_offset;
785 mono_mb_set_clauses (mb, 1, clause);
786 mono_mb_patch_branch (mb, pos_leave);
789 g_assert_not_reached ();
792 /* size += ALLOC_ALIGN - 1; */
793 mono_mb_emit_ldloc (mb, size_var);
794 mono_mb_emit_icon (mb, ALLOC_ALIGN - 1);
795 mono_mb_emit_byte (mb, CEE_ADD);
796 /* size &= ~(ALLOC_ALIGN - 1); */
797 mono_mb_emit_icon (mb, ~(ALLOC_ALIGN - 1));
798 mono_mb_emit_byte (mb, CEE_AND);
799 mono_mb_emit_stloc (mb, size_var);
801 /* if (size > MAX_SMALL_OBJ_SIZE) goto slowpath */
802 if (atype != ATYPE_SMALL) {
803 mono_mb_emit_ldloc (mb, size_var);
804 mono_mb_emit_icon (mb, MAX_SMALL_OBJ_SIZE);
805 max_size_branch = mono_mb_emit_short_branch (mb, MONO_CEE_BGT_UN_S);
809 * We need to modify tlab_next, but the JIT only supports reading, so we read
810 * another tls var holding its address instead.
813 /* tlab_next_addr (local) = tlab_next_addr (TLS var) */
814 tlab_next_addr_var = mono_mb_add_local (mb, &mono_defaults.int_class->byval_arg);
815 EMIT_TLS_ACCESS (mb, tlab_next_addr, tlab_next_addr_offset);
816 mono_mb_emit_stloc (mb, tlab_next_addr_var);
818 /* p = (void**)tlab_next; */
819 p_var = mono_mb_add_local (mb, &mono_defaults.int_class->byval_arg);
820 mono_mb_emit_ldloc (mb, tlab_next_addr_var);
821 mono_mb_emit_byte (mb, CEE_LDIND_I);
822 mono_mb_emit_stloc (mb, p_var);
824 /* new_next = (char*)p + size; */
825 new_next_var = mono_mb_add_local (mb, &mono_defaults.int_class->byval_arg);
826 mono_mb_emit_ldloc (mb, p_var);
827 mono_mb_emit_ldloc (mb, size_var);
828 mono_mb_emit_byte (mb, CEE_CONV_I);
829 mono_mb_emit_byte (mb, CEE_ADD);
830 mono_mb_emit_stloc (mb, new_next_var);
832 /* if (G_LIKELY (new_next < tlab_temp_end)) */
833 mono_mb_emit_ldloc (mb, new_next_var);
834 EMIT_TLS_ACCESS (mb, tlab_temp_end, tlab_temp_end_offset);
835 slowpath_branch = mono_mb_emit_short_branch (mb, MONO_CEE_BLT_UN_S);
838 if (atype != ATYPE_SMALL)
839 mono_mb_patch_short_branch (mb, max_size_branch);
841 mono_mb_emit_byte (mb, MONO_CUSTOM_PREFIX);
842 mono_mb_emit_byte (mb, CEE_MONO_NOT_TAKEN);
844 /* FIXME: mono_gc_alloc_obj takes a 'size_t' as an argument, not an int32 */
845 mono_mb_emit_ldarg (mb, 0);
846 mono_mb_emit_ldloc (mb, size_var);
847 if (atype == ATYPE_NORMAL || atype == ATYPE_SMALL) {
848 mono_mb_emit_icall (mb, mono_gc_alloc_obj);
849 } else if (atype == ATYPE_VECTOR) {
850 mono_mb_emit_ldarg (mb, 1);
851 mono_mb_emit_icall (mb, mono_gc_alloc_vector);
853 g_assert_not_reached ();
855 mono_mb_emit_byte (mb, CEE_RET);
858 mono_mb_patch_short_branch (mb, slowpath_branch);
860 /* FIXME: Memory barrier */
862 /* tlab_next = new_next */
863 mono_mb_emit_ldloc (mb, tlab_next_addr_var);
864 mono_mb_emit_ldloc (mb, new_next_var);
865 mono_mb_emit_byte (mb, CEE_STIND_I);
867 /*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. */
868 mono_mb_emit_byte ((mb), MONO_CUSTOM_PREFIX);
869 mono_mb_emit_op (mb, CEE_MONO_MEMORY_BARRIER, StoreStoreBarrier);
872 mono_mb_emit_ldloc (mb, p_var);
873 mono_mb_emit_ldarg (mb, 0);
874 mono_mb_emit_byte (mb, CEE_STIND_I);
876 if (atype == ATYPE_VECTOR) {
877 /* arr->max_length = max_length; */
878 mono_mb_emit_ldloc (mb, p_var);
879 mono_mb_emit_ldflda (mb, G_STRUCT_OFFSET (MonoArray, max_length));
880 mono_mb_emit_ldarg (mb, 1);
881 #ifdef MONO_BIG_ARRAYS
882 mono_mb_emit_byte (mb, CEE_STIND_I);
884 mono_mb_emit_byte (mb, CEE_STIND_I4);
889 We must make sure both vtable and max_length are globaly visible before returning to managed land.
891 mono_mb_emit_byte ((mb), MONO_CUSTOM_PREFIX);
892 mono_mb_emit_op (mb, CEE_MONO_MEMORY_BARRIER, StoreStoreBarrier);
895 mono_mb_emit_ldloc (mb, p_var);
896 mono_mb_emit_byte (mb, CEE_RET);
898 res = mono_mb_create_method (mb, csig, 8);
900 mono_method_get_header (res)->init_locals = FALSE;
902 info = mono_image_alloc0 (mono_defaults.corlib, sizeof (AllocatorWrapperInfo));
903 info->gc_name = "sgen";
904 info->alloc_type = atype;
905 mono_marshal_set_wrapper_info (res, info);
912 * Generate an allocator method implementing the fast path of mono_gc_alloc_obj ().
913 * The signature of the called method is:
914 * object allocate (MonoVTable *vtable)
917 mono_gc_get_managed_allocator (MonoVTable *vtable, gboolean for_box)
919 #ifdef MANAGED_ALLOCATION
920 MonoClass *klass = vtable->klass;
922 #ifdef HAVE_KW_THREAD
923 int tlab_next_offset = -1;
924 int tlab_temp_end_offset = -1;
925 MONO_THREAD_VAR_OFFSET (tlab_next, tlab_next_offset);
926 MONO_THREAD_VAR_OFFSET (tlab_temp_end, tlab_temp_end_offset);
928 if (tlab_next_offset == -1 || tlab_temp_end_offset == -1)
932 if (!mono_runtime_has_tls_get ())
934 if (klass->instance_size > tlab_size)
936 if (klass->has_finalize || klass->marshalbyref || (mono_profiler_get_events () & MONO_PROFILE_ALLOCATIONS))
940 if (klass->byval_arg.type == MONO_TYPE_STRING)
942 if (collect_before_allocs)
945 if (ALIGN_TO (klass->instance_size, ALLOC_ALIGN) < MAX_SMALL_OBJ_SIZE)
946 return mono_gc_get_managed_allocator_by_type (ATYPE_SMALL);
948 return mono_gc_get_managed_allocator_by_type (ATYPE_NORMAL);
955 mono_gc_get_managed_array_allocator (MonoVTable *vtable, int rank)
957 #ifdef MANAGED_ALLOCATION
958 MonoClass *klass = vtable->klass;
960 #ifdef HAVE_KW_THREAD
961 int tlab_next_offset = -1;
962 int tlab_temp_end_offset = -1;
963 MONO_THREAD_VAR_OFFSET (tlab_next, tlab_next_offset);
964 MONO_THREAD_VAR_OFFSET (tlab_temp_end, tlab_temp_end_offset);
966 if (tlab_next_offset == -1 || tlab_temp_end_offset == -1)
972 if (!mono_runtime_has_tls_get ())
974 if (mono_profiler_get_events () & MONO_PROFILE_ALLOCATIONS)
976 if (has_per_allocation_action)
978 g_assert (!mono_class_has_finalizer (klass) && !klass->marshalbyref);
980 return mono_gc_get_managed_allocator_by_type (ATYPE_VECTOR);
987 sgen_set_use_managed_allocator (gboolean flag)
989 use_managed_allocator = flag;
993 mono_gc_get_managed_allocator_by_type (int atype)
995 #ifdef MANAGED_ALLOCATION
998 if (!use_managed_allocator)
1001 if (!mono_runtime_has_tls_get ())
1004 mono_loader_lock ();
1005 res = alloc_method_cache [atype];
1007 res = alloc_method_cache [atype] = create_allocator (atype);
1008 mono_loader_unlock ();
1016 mono_gc_get_managed_allocator_types (void)
1022 sgen_is_managed_allocator (MonoMethod *method)
1026 for (i = 0; i < ATYPE_NUM; ++i)
1027 if (method == alloc_method_cache [i])
1033 sgen_has_managed_allocator (void)
1037 for (i = 0; i < ATYPE_NUM; ++i)
1038 if (alloc_method_cache [i])
1043 #ifdef HEAVY_STATISTICS
1045 sgen_alloc_init_heavy_stats (void)
1047 mono_counters_register ("# objects allocated", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_objects_alloced);
1048 mono_counters_register ("bytes allocated", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_bytes_alloced);
1049 mono_counters_register ("bytes allocated in LOS", MONO_COUNTER_GC | MONO_COUNTER_LONG, &stat_bytes_alloced_los);
1053 #endif /*HAVE_SGEN_GC*/