mono/metadata/sgen-alloc.c

   1 /*
   2  * sgen-alloc.c: Object allocation routines + managed allocators
   3  *
   4  * Author:
   5  *      Paolo Molaro (lupus@ximian.com)
   6  *  Rodrigo Kumpera (kumpera@gmail.com)
   7  *
   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
  12  *
  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;
  16  *
  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.
  21  *
  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.
  25  */
  26
  27 /*
  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
  37  */
  38
  39 #include "config.h"
  40 #ifdef HAVE_SGEN_GC
  41
  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"
  51
  52 #define ALIGN_UP                SGEN_ALIGN_UP
  53 #define ALLOC_ALIGN             SGEN_ALLOC_ALIGN
  54 #define MAX_SMALL_OBJ_SIZE      SGEN_MAX_SMALL_OBJ_SIZE
  55 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
  56
  57 #define OPDEF(a,b,c,d,e,f,g,h,i,j) \
  58         a = i,
  59
  60 enum {
  61 #include "mono/cil/opcode.def"
  62         CEE_LAST
  63 };
  64
  65 #undef OPDEF
  66
  67 static gboolean use_managed_allocator = TRUE;
  68
  69 #ifdef HEAVY_STATISTICS
  70 static guint64 stat_objects_alloced = 0;
  71 static guint64 stat_bytes_alloced = 0;
  72 static guint64 stat_bytes_alloced_los = 0;
  73
  74 #endif
  75
  76 /*
  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
  80  * be allocated.
  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.
  84  */
  85
  86 /*
  87  * FIXME: What is faster, a TLS variable pointing to a structure, or separate TLS
  88  * variables for next+temp_end ?
  89  */
  90 #ifdef HAVE_KW_THREAD
  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;
  97 #endif
  98
  99 #ifdef HAVE_KW_THREAD
 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
 104 #else
 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)
 109 #endif
 110
 111 static void*
 112 alloc_degraded (MonoVTable *vtable, size_t size, gboolean for_mature)
 113 {
 114         static int last_major_gc_warned = -1;
 115         static int num_degraded = 0;
 116
 117         void *p;
 118
 119         if (!for_mature) {
 120                 if (last_major_gc_warned < gc_stats.major_gc_count) {
 121                         ++num_degraded;
 122                         if (num_degraded == 1 || num_degraded == 3)
 123                                 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_GC, "Warning: Degraded allocation.  Consider increasing nursery-size if the warning persists.");
 124                         else if (num_degraded == 10)
 125                                 mono_trace (G_LOG_LEVEL_INFO, MONO_TRACE_GC, "Warning: Repeated degraded allocation.  Consider increasing nursery-size.");
 126                         last_major_gc_warned = gc_stats.major_gc_count;
 127                 }
 128                 SGEN_ATOMIC_ADD_P (degraded_mode, size);
 129                 sgen_ensure_free_space (size);
 130         } else {
 131                 if (sgen_need_major_collection (size))
 132                         sgen_perform_collection (size, GENERATION_OLD, "mature allocation failure", !for_mature);
 133         }
 134
 135
 136         p = major_collector.alloc_degraded (vtable, size);
 137
 138         if (for_mature) {
 139                 MONO_GC_MAJOR_OBJ_ALLOC_MATURE ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
 140         } else {
 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);
 143         }
 144
 145         return p;
 146 }
 147
 148 static void
 149 zero_tlab_if_necessary (void *p, size_t size)
 150 {
 151         if (nursery_clear_policy == CLEAR_AT_TLAB_CREATION || nursery_clear_policy == CLEAR_AT_TLAB_CREATION_DEBUG) {
 152                 memset (p, 0, size);
 153         } else {
 154                 /*
 155                  * This function is called for all allocations in
 156                  * TLABs.  TLABs originate from fragments, which are
 157                  * initialized to be faux arrays.  The remainder of
 158                  * the fragments are zeroed out at initialization for
 159                  * CLEAR_AT_GC, so here we just need to make sure that
 160                  * the array header is zeroed.  Since we don't know
 161                  * whether we're called for the start of a fragment or
 162                  * for somewhere in between, we zero in any case, just
 163                  * to make sure.
 164                  */
 165
 166                 if (size >= sizeof (MonoArray))
 167                         memset (p, 0, sizeof (MonoArray));
 168                 else {
 169                         static guint8 zeros [sizeof (MonoArray)];
 170
 171                         SGEN_ASSERT (0, !memcmp (p, zeros, size), "TLAB segment must be zeroed out.");
 172                 }
 173         }
 174 }
 175
 176 /*
 177  * Provide a variant that takes just the vtable for small fixed-size objects.
 178  * The aligned size is already computed and stored in vt->gc_descr.
 179  * Note: every SGEN_SCAN_START_SIZE or so we are given the chance to do some special
 180  * processing. We can keep track of where objects start, for example,
 181  * so when we scan the thread stacks for pinned objects, we can start
 182  * a search for the pinned object in SGEN_SCAN_START_SIZE chunks.
 183  */
 184 static void*
 185 mono_gc_alloc_obj_nolock (MonoVTable *vtable, size_t size)
 186 {
 187         /* FIXME: handle OOM */
 188         void **p;
 189         char *new_next;
 190         size_t real_size = size;
 191         TLAB_ACCESS_INIT;
 192
 193         CANARIFY_SIZE(size);
 194
 195         HEAVY_STAT (++stat_objects_alloced);
 196         if (real_size <= SGEN_MAX_SMALL_OBJ_SIZE)
 197                 HEAVY_STAT (stat_bytes_alloced += size);
 198         else
 199                 HEAVY_STAT (stat_bytes_alloced_los += size);
 200
 201         size = ALIGN_UP (size);
 202
 203         g_assert (vtable->gc_descr);
 204
 205         if (G_UNLIKELY (has_per_allocation_action)) {
 206                 static int alloc_count;
 207                 int current_alloc = InterlockedIncrement (&alloc_count);
 208
 209                 if (collect_before_allocs) {
 210                         if (((current_alloc % collect_before_allocs) == 0) && nursery_section) {
 211                                 sgen_perform_collection (0, GENERATION_NURSERY, "collect-before-alloc-triggered", TRUE);
 212                                 if (!degraded_mode && sgen_can_alloc_size (size) && real_size <= SGEN_MAX_SMALL_OBJ_SIZE) {
 213                                         // FIXME:
 214                                         g_assert_not_reached ();
 215                                 }
 216                         }
 217                 } else if (verify_before_allocs) {
 218                         if ((current_alloc % verify_before_allocs) == 0)
 219                                 sgen_check_whole_heap_stw ();
 220                 }
 221         }
 222
 223         /*
 224          * We must already have the lock here instead of after the
 225          * fast path because we might be interrupted in the fast path
 226          * (after confirming that new_next < TLAB_TEMP_END) by the GC,
 227          * and we'll end up allocating an object in a fragment which
 228          * no longer belongs to us.
 229          *
 230          * The managed allocator does not do this, but it's treated
 231          * specially by the world-stopping code.
 232          */
 233
 234         if (real_size > SGEN_MAX_SMALL_OBJ_SIZE) {
 235                 p = sgen_los_alloc_large_inner (vtable, ALIGN_UP (real_size));
 236         } else {
 237                 /* tlab_next and tlab_temp_end are TLS vars so accessing them might be expensive */
 238
 239                 p = (void**)TLAB_NEXT;
 240                 /* FIXME: handle overflow */
 241                 new_next = (char*)p + size;
 242                 TLAB_NEXT = new_next;
 243
 244                 if (G_LIKELY (new_next < TLAB_TEMP_END)) {
 245                         /* Fast path */
 246
 247                         /*
 248                          * FIXME: We might need a memory barrier here so the change to tlab_next is
 249                          * visible before the vtable store.
 250                          */
 251
 252                         CANARIFY_ALLOC(p,real_size);
 253                         SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, vtable->klass->name, size);
 254                         binary_protocol_alloc (p , vtable, size);
 255                         if (G_UNLIKELY (MONO_GC_NURSERY_OBJ_ALLOC_ENABLED ()))
 256                                 MONO_GC_NURSERY_OBJ_ALLOC ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
 257                         g_assert (*p == NULL);
 258                         mono_atomic_store_seq (p, vtable);
 259
 260                         return p;
 261                 }
 262
 263                 /* Slow path */
 264
 265                 /* there are two cases: the object is too big or we run out of space in the TLAB */
 266                 /* we also reach here when the thread does its first allocation after a minor
 267                  * collection, since the tlab_ variables are initialized to NULL.
 268                  * there can be another case (from ORP), if we cooperate with the runtime a bit:
 269                  * objects that need finalizers can have the high bit set in their size
 270                  * so the above check fails and we can readily add the object to the queue.
 271                  * This avoids taking again the GC lock when registering, but this is moot when
 272                  * doing thread-local allocation, so it may not be a good idea.
 273                  */
 274                 if (TLAB_NEXT >= TLAB_REAL_END) {
 275                         int available_in_tlab;
 276                         /*
 277                          * Run out of space in the TLAB. When this happens, some amount of space
 278                          * remains in the TLAB, but not enough to satisfy the current allocation
 279                          * request. Currently, we retire the TLAB in all cases, later we could
 280                          * keep it if the remaining space is above a treshold, and satisfy the
 281                          * allocation directly from the nursery.
 282                          */
 283                         TLAB_NEXT -= size;
 284                         /* when running in degraded mode, we continue allocing that way
 285                          * for a while, to decrease the number of useless nursery collections.
 286                          */
 287                         if (degraded_mode && degraded_mode < DEFAULT_NURSERY_SIZE)
 288                                 return alloc_degraded (vtable, size, FALSE);
 289
 290                         available_in_tlab = (int)(TLAB_REAL_END - TLAB_NEXT);//We'll never have tlabs > 2Gb
 291                         if (size > tlab_size || available_in_tlab > SGEN_MAX_NURSERY_WASTE) {
 292                                 /* Allocate directly from the nursery */
 293                                 p = sgen_nursery_alloc (size);
 294                                 if (!p) {
 295                                         /*
 296                                          * We couldn't allocate from the nursery, so we try
 297                                          * collecting.  Even after the collection, we might
 298                                          * still not have enough memory to allocate the
 299                                          * object.  The reason will most likely be that we've
 300                                          * run out of memory, but there is the theoretical
 301                                          * possibility that other threads might have consumed
 302                                          * the freed up memory ahead of us.
 303                                          *
 304                                          * What we do in this case is allocate degraded, i.e.,
 305                                          * from the major heap.
 306                                          *
 307                                          * Ideally we'd like to detect the case of other
 308                                          * threads allocating ahead of us and loop (if we
 309                                          * always loop we will loop endlessly in the case of
 310                                          * OOM).
 311                                          */
 312                                         sgen_ensure_free_space (real_size);
 313                                         if (!degraded_mode)
 314                                                 p = sgen_nursery_alloc (size);
 315                                 }
 316                                 if (!p)
 317                                         return alloc_degraded (vtable, size, FALSE);
 318
 319                                 zero_tlab_if_necessary (p, size);
 320                         } else {
 321                                 size_t alloc_size = 0;
 322                                 if (TLAB_START)
 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);
 325
 326                                 p = sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
 327                                 if (!p) {
 328                                         /* See comment above in similar case. */
 329                                         sgen_ensure_free_space (tlab_size);
 330                                         if (!degraded_mode)
 331                                                 p = sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
 332                                 }
 333                                 if (!p)
 334                                         return alloc_degraded (vtable, size, FALSE);
 335
 336                                 /* Allocate a new TLAB from the current nursery fragment */
 337                                 TLAB_START = (char*)p;
 338                                 TLAB_NEXT = TLAB_START;
 339                                 TLAB_REAL_END = TLAB_START + alloc_size;
 340                                 TLAB_TEMP_END = TLAB_START + MIN (SGEN_SCAN_START_SIZE, alloc_size);
 341
 342                                 zero_tlab_if_necessary (TLAB_START, alloc_size);
 343
 344                                 /* Allocate from the TLAB */
 345                                 p = (void*)TLAB_NEXT;
 346                                 TLAB_NEXT += size;
 347                                 sgen_set_nursery_scan_start ((char*)p);
 348                         }
 349                 } else {
 350                         /* Reached tlab_temp_end */
 351
 352                         /* record the scan start so we can find pinned objects more easily */
 353                         sgen_set_nursery_scan_start ((char*)p);
 354                         /* we just bump tlab_temp_end as well */
 355                         TLAB_TEMP_END = MIN (TLAB_REAL_END, TLAB_NEXT + SGEN_SCAN_START_SIZE);
 356                         SGEN_LOG (5, "Expanding local alloc: %p-%p", TLAB_NEXT, TLAB_TEMP_END);
 357                 }
 358                 CANARIFY_ALLOC(p,real_size);
 359         }
 360
 361         if (G_LIKELY (p)) {
 362                 SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, vtable->klass->name, size);
 363                 binary_protocol_alloc (p, vtable, size);
 364                 if (G_UNLIKELY (MONO_GC_MAJOR_OBJ_ALLOC_LARGE_ENABLED ()|| MONO_GC_NURSERY_OBJ_ALLOC_ENABLED ())) {
 365                         if (real_size > SGEN_MAX_SMALL_OBJ_SIZE)
 366                                 MONO_GC_MAJOR_OBJ_ALLOC_LARGE ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
 367                         else
 368                                 MONO_GC_NURSERY_OBJ_ALLOC ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
 369                 }
 370                 mono_atomic_store_seq (p, vtable);
 371         }
 372
 373         return p;
 374 }
 375
 376 static void*
 377 mono_gc_try_alloc_obj_nolock (MonoVTable *vtable, size_t size)
 378 {
 379         void **p;
 380         char *new_next;
 381         size_t real_size = size;
 382         TLAB_ACCESS_INIT;
 383
 384         CANARIFY_SIZE(size);
 385
 386         size = ALIGN_UP (size);
 387         SGEN_ASSERT (9, real_size >= sizeof (MonoObject), "Object too small");
 388
 389         g_assert (vtable->gc_descr);
 390         if (real_size > SGEN_MAX_SMALL_OBJ_SIZE)
 391                 return NULL;
 392
 393         if (G_UNLIKELY (size > tlab_size)) {
 394                 /* Allocate directly from the nursery */
 395                 p = sgen_nursery_alloc (size);
 396                 if (!p)
 397                         return NULL;
 398                 sgen_set_nursery_scan_start ((char*)p);
 399
 400                 /*FIXME we should use weak memory ops here. Should help specially on x86. */
 401                 zero_tlab_if_necessary (p, size);
 402         } else {
 403                 int available_in_tlab;
 404                 char *real_end;
 405                 /* tlab_next and tlab_temp_end are TLS vars so accessing them might be expensive */
 406
 407                 p = (void**)TLAB_NEXT;
 408                 /* FIXME: handle overflow */
 409                 new_next = (char*)p + size;
 410
 411                 real_end = TLAB_REAL_END;
 412                 available_in_tlab = (int)(real_end - (char*)p);//We'll never have tlabs > 2Gb
 413
 414                 if (G_LIKELY (new_next < real_end)) {
 415                         TLAB_NEXT = new_next;
 416
 417                         /* Second case, we overflowed temp end */
 418                         if (G_UNLIKELY (new_next >= TLAB_TEMP_END)) {
 419                                 sgen_set_nursery_scan_start (new_next);
 420                                 /* we just bump tlab_temp_end as well */
 421                                 TLAB_TEMP_END = MIN (TLAB_REAL_END, TLAB_NEXT + SGEN_SCAN_START_SIZE);
 422                                 SGEN_LOG (5, "Expanding local alloc: %p-%p", TLAB_NEXT, TLAB_TEMP_END);
 423                         }
 424                 } else if (available_in_tlab > SGEN_MAX_NURSERY_WASTE) {
 425                         /* Allocate directly from the nursery */
 426                         p = sgen_nursery_alloc (size);
 427                         if (!p)
 428                                 return NULL;
 429
 430                         zero_tlab_if_necessary (p, size);
 431                 } else {
 432                         size_t alloc_size = 0;
 433
 434                         sgen_nursery_retire_region (p, available_in_tlab);
 435                         new_next = sgen_nursery_alloc_range (tlab_size, size, &alloc_size);
 436                         p = (void**)new_next;
 437                         if (!p)
 438                                 return NULL;
 439
 440                         TLAB_START = (char*)new_next;
 441                         TLAB_NEXT = new_next + size;
 442                         TLAB_REAL_END = new_next + alloc_size;
 443                         TLAB_TEMP_END = new_next + MIN (SGEN_SCAN_START_SIZE, alloc_size);
 444                         sgen_set_nursery_scan_start ((char*)p);
 445
 446                         zero_tlab_if_necessary (new_next, alloc_size);
 447
 448                         MONO_GC_NURSERY_TLAB_ALLOC ((mword)new_next, alloc_size);
 449                 }
 450         }
 451
 452         HEAVY_STAT (++stat_objects_alloced);
 453         HEAVY_STAT (stat_bytes_alloced += size);
 454
 455         CANARIFY_ALLOC(p,real_size);
 456         SGEN_LOG (6, "Allocated object %p, vtable: %p (%s), size: %zd", p, vtable, vtable->klass->name, size);
 457         binary_protocol_alloc (p, vtable, size);
 458         if (G_UNLIKELY (MONO_GC_NURSERY_OBJ_ALLOC_ENABLED ()))
 459                 MONO_GC_NURSERY_OBJ_ALLOC ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
 460         g_assert (*p == NULL); /* FIXME disable this in non debug builds */
 461
 462         mono_atomic_store_seq (p, vtable);
 463
 464         return p;
 465 }
 466
 467 void*
 468 mono_gc_alloc_obj (MonoVTable *vtable, size_t size)
 469 {
 470         void *res;
 471         TLAB_ACCESS_INIT;
 472
 473         if (!SGEN_CAN_ALIGN_UP (size))
 474                 return NULL;
 475
 476 #ifndef DISABLE_CRITICAL_REGION
 477
 478         if (G_UNLIKELY (has_per_allocation_action)) {
 479                 static int alloc_count;
 480                 int current_alloc = InterlockedIncrement (&alloc_count);
 481
 482                 if (verify_before_allocs) {
 483                         if ((current_alloc % verify_before_allocs) == 0)
 484                                 sgen_check_whole_heap_stw ();
 485                 }
 486                 if (collect_before_allocs) {
 487                         if (((current_alloc % collect_before_allocs) == 0) && nursery_section) {
 488                                 LOCK_GC;
 489                                 sgen_perform_collection (0, GENERATION_NURSERY, "collect-before-alloc-triggered", TRUE);
 490                                 UNLOCK_GC;
 491                         }
 492                 }
 493         }
 494
 495         ENTER_CRITICAL_REGION;
 496         res = mono_gc_try_alloc_obj_nolock (vtable, size);
 497         if (res) {
 498                 EXIT_CRITICAL_REGION;
 499                 return res;
 500         }
 501         EXIT_CRITICAL_REGION;
 502 #endif
 503         LOCK_GC;
 504         res = mono_gc_alloc_obj_nolock (vtable, size);
 505         UNLOCK_GC;
 506         if (G_UNLIKELY (!res))
 507                 return mono_gc_out_of_memory (size);
 508         return res;
 509 }
 510
 511 void*
 512 mono_gc_alloc_vector (MonoVTable *vtable, size_t size, uintptr_t max_length)
 513 {
 514         MonoArray *arr;
 515         TLAB_ACCESS_INIT;
 516
 517         if (!SGEN_CAN_ALIGN_UP (size))
 518                 return NULL;
 519
 520 #ifndef DISABLE_CRITICAL_REGION
 521         ENTER_CRITICAL_REGION;
 522         arr = mono_gc_try_alloc_obj_nolock (vtable, size);
 523         if (arr) {
 524                 /*This doesn't require fencing since EXIT_CRITICAL_REGION already does it for us*/
 525                 arr->max_length = (mono_array_size_t)max_length;
 526                 EXIT_CRITICAL_REGION;
 527                 return arr;
 528         }
 529         EXIT_CRITICAL_REGION;
 530 #endif
 531
 532         LOCK_GC;
 533
 534         arr = mono_gc_alloc_obj_nolock (vtable, size);
 535         if (G_UNLIKELY (!arr)) {
 536                 UNLOCK_GC;
 537                 return mono_gc_out_of_memory (size);
 538         }
 539
 540         arr->max_length = (mono_array_size_t)max_length;
 541
 542         UNLOCK_GC;
 543
 544         return arr;
 545 }
 546
 547 void*
 548 mono_gc_alloc_array (MonoVTable *vtable, size_t size, uintptr_t max_length, uintptr_t bounds_size)
 549 {
 550         MonoArray *arr;
 551         MonoArrayBounds *bounds;
 552         TLAB_ACCESS_INIT;
 553
 554         if (!SGEN_CAN_ALIGN_UP (size))
 555                 return NULL;
 556
 557 #ifndef DISABLE_CRITICAL_REGION
 558         ENTER_CRITICAL_REGION;
 559         arr = mono_gc_try_alloc_obj_nolock (vtable, size);
 560         if (arr) {
 561                 /*This doesn't require fencing since EXIT_CRITICAL_REGION already does it for us*/
 562                 arr->max_length = (mono_array_size_t)max_length;
 563
 564                 bounds = (MonoArrayBounds*)((char*)arr + size - bounds_size);
 565                 arr->bounds = bounds;
 566                 EXIT_CRITICAL_REGION;
 567                 return arr;
 568         }
 569         EXIT_CRITICAL_REGION;
 570 #endif
 571
 572         LOCK_GC;
 573
 574         arr = mono_gc_alloc_obj_nolock (vtable, size);
 575         if (G_UNLIKELY (!arr)) {
 576                 UNLOCK_GC;
 577                 return mono_gc_out_of_memory (size);
 578         }
 579
 580         arr->max_length = (mono_array_size_t)max_length;
 581
 582         bounds = (MonoArrayBounds*)((char*)arr + size - bounds_size);
 583         arr->bounds = bounds;
 584
 585         UNLOCK_GC;
 586
 587         return arr;
 588 }
 589
 590 void*
 591 mono_gc_alloc_string (MonoVTable *vtable, size_t size, gint32 len)
 592 {
 593         MonoString *str;
 594         TLAB_ACCESS_INIT;
 595
 596         if (!SGEN_CAN_ALIGN_UP (size))
 597                 return NULL;
 598
 599 #ifndef DISABLE_CRITICAL_REGION
 600         ENTER_CRITICAL_REGION;
 601         str = mono_gc_try_alloc_obj_nolock (vtable, size);
 602         if (str) {
 603                 /*This doesn't require fencing since EXIT_CRITICAL_REGION already does it for us*/
 604                 str->length = len;
 605                 EXIT_CRITICAL_REGION;
 606                 return str;
 607         }
 608         EXIT_CRITICAL_REGION;
 609 #endif
 610
 611         LOCK_GC;
 612
 613         str = mono_gc_alloc_obj_nolock (vtable, size);
 614         if (G_UNLIKELY (!str)) {
 615                 UNLOCK_GC;
 616                 return mono_gc_out_of_memory (size);
 617         }
 618
 619         str->length = len;
 620
 621         UNLOCK_GC;
 622
 623         return str;
 624 }
 625
 626 /*
 627  * To be used for interned strings and possibly MonoThread, reflection handles.
 628  * We may want to explicitly free these objects.
 629  */
 630 void*
 631 mono_gc_alloc_pinned_obj (MonoVTable *vtable, size_t size)
 632 {
 633         void **p;
 634
 635         if (!SGEN_CAN_ALIGN_UP (size))
 636                 return NULL;
 637         size = ALIGN_UP (size);
 638
 639         LOCK_GC;
 640
 641         if (size > SGEN_MAX_SMALL_OBJ_SIZE) {
 642                 /* large objects are always pinned anyway */
 643                 p = sgen_los_alloc_large_inner (vtable, size);
 644         } else {
 645                 SGEN_ASSERT (9, vtable->klass->inited, "class %s:%s is not initialized", vtable->klass->name_space, vtable->klass->name);
 646                 p = major_collector.alloc_small_pinned_obj (vtable, size, SGEN_VTABLE_HAS_REFERENCES (vtable));
 647         }
 648         if (G_LIKELY (p)) {
 649                 SGEN_LOG (6, "Allocated pinned object %p, vtable: %p (%s), size: %zd", p, vtable, vtable->klass->name, size);
 650                 if (size > SGEN_MAX_SMALL_OBJ_SIZE)
 651                         MONO_GC_MAJOR_OBJ_ALLOC_LARGE ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
 652                 else
 653                         MONO_GC_MAJOR_OBJ_ALLOC_PINNED ((mword)p, size, vtable->klass->name_space, vtable->klass->name);
 654                 binary_protocol_alloc_pinned (p, vtable, size);
 655         }
 656         UNLOCK_GC;
 657         return p;
 658 }
 659
 660 void*
 661 mono_gc_alloc_mature (MonoVTable *vtable)
 662 {
 663         void **res;
 664         size_t size = vtable->klass->instance_size;
 665
 666         if (!SGEN_CAN_ALIGN_UP (size))
 667                 return NULL;
 668         size = ALIGN_UP (size);
 669
 670         LOCK_GC;
 671         res = alloc_degraded (vtable, size, TRUE);
 672         UNLOCK_GC;
 673         if (G_UNLIKELY (vtable->klass->has_finalize))
 674                 mono_object_register_finalizer ((MonoObject*)res);
 675
 676         return res;
 677 }
 678
 679 void*
 680 mono_gc_alloc_fixed (size_t size, void *descr)
 681 {
 682         /* FIXME: do a single allocation */
 683         void *res = calloc (1, size);
 684         if (!res)
 685                 return NULL;
 686         if (!mono_gc_register_root (res, size, descr)) {
 687                 free (res);
 688                 res = NULL;
 689         }
 690         return res;
 691 }
 692
 693 void
 694 mono_gc_free_fixed (void* addr)
 695 {
 696         mono_gc_deregister_root (addr);
 697         free (addr);
 698 }
 699
 700 void
 701 sgen_init_tlab_info (SgenThreadInfo* info)
 702 {
 703 #ifndef HAVE_KW_THREAD
 704         SgenThreadInfo *__thread_info__ = info;
 705 #endif
 706
 707         info->tlab_start_addr = &TLAB_START;
 708         info->tlab_next_addr = &TLAB_NEXT;
 709         info->tlab_temp_end_addr = &TLAB_TEMP_END;
 710         info->tlab_real_end_addr = &TLAB_REAL_END;
 711
 712 #ifdef HAVE_KW_THREAD
 713         tlab_next_addr = &tlab_next;
 714 #endif
 715 }
 716
 717 /*
 718  * Clear the thread local TLAB variables for all threads.
 719  */
 720 void
 721 sgen_clear_tlabs (void)
 722 {
 723         SgenThreadInfo *info;
 724
 725         FOREACH_THREAD (info) {
 726                 /* A new TLAB will be allocated when the thread does its first allocation */
 727                 *info->tlab_start_addr = NULL;
 728                 *info->tlab_next_addr = NULL;
 729                 *info->tlab_temp_end_addr = NULL;
 730                 *info->tlab_real_end_addr = NULL;
 731         } END_FOREACH_THREAD
 732 }
 733
 734 static MonoMethod* alloc_method_cache [ATYPE_NUM];
 735
 736 #ifdef MANAGED_ALLOCATION
 737 /* FIXME: Do this in the JIT, where specialized allocation sequences can be created
 738  * for each class. This is currently not easy to do, as it is hard to generate basic
 739  * blocks + branches, but it is easy with the linear IL codebase.
 740  *
 741  * For this to work we'd need to solve the TLAB race, first.  Now we
 742  * require the allocator to be in a few known methods to make sure
 743  * that they are executed atomically via the restart mechanism.
 744  */
 745 static MonoMethod*
 746 create_allocator (int atype)
 747 {
 748         int p_var, size_var;
 749         guint32 slowpath_branch, max_size_branch;
 750         MonoMethodBuilder *mb;
 751         MonoMethod *res;
 752         MonoMethodSignature *csig;
 753         static gboolean registered = FALSE;
 754         int tlab_next_addr_var, new_next_var;
 755         int num_params, i;
 756         const char *name = NULL;
 757         AllocatorWrapperInfo *info;
 758
 759 #ifdef HAVE_KW_THREAD
 760         int tlab_next_addr_offset = -1;
 761         int tlab_temp_end_offset = -1;
 762
 763         MONO_THREAD_VAR_OFFSET (tlab_next_addr, tlab_next_addr_offset);
 764         MONO_THREAD_VAR_OFFSET (tlab_temp_end, tlab_temp_end_offset);
 765
 766         mono_tls_key_set_offset (TLS_KEY_SGEN_TLAB_NEXT_ADDR, tlab_next_addr_offset);
 767         mono_tls_key_set_offset (TLS_KEY_SGEN_TLAB_TEMP_END, tlab_temp_end_offset);
 768
 769         g_assert (tlab_next_addr_offset != -1);
 770         g_assert (tlab_temp_end_offset != -1);
 771 #endif
 772
 773         if (!registered) {
 774                 mono_register_jit_icall (mono_gc_alloc_obj, "mono_gc_alloc_obj", mono_create_icall_signature ("object ptr int"), FALSE);
 775                 mono_register_jit_icall (mono_gc_alloc_vector, "mono_gc_alloc_vector", mono_create_icall_signature ("object ptr int int"), FALSE);
 776                 mono_register_jit_icall (mono_gc_alloc_string, "mono_gc_alloc_string", mono_create_icall_signature ("object ptr int int32"), FALSE);
 777                 registered = TRUE;
 778         }
 779
 780         if (atype == ATYPE_SMALL) {
 781                 num_params = 1;
 782                 name = "AllocSmall";
 783         } else if (atype == ATYPE_NORMAL) {
 784                 num_params = 1;
 785                 name = "Alloc";
 786         } else if (atype == ATYPE_VECTOR) {
 787                 num_params = 2;
 788                 name = "AllocVector";
 789         } else if (atype == ATYPE_STRING) {
 790                 num_params = 2;
 791                 name = "AllocString";
 792         } else {
 793                 g_assert_not_reached ();
 794         }
 795
 796         csig = mono_metadata_signature_alloc (mono_defaults.corlib, num_params);
 797         if (atype == ATYPE_STRING) {
 798                 csig->ret = &mono_defaults.string_class->byval_arg;
 799                 csig->params [0] = &mono_defaults.int_class->byval_arg;
 800                 csig->params [1] = &mono_defaults.int32_class->byval_arg;
 801         } else {
 802                 csig->ret = &mono_defaults.object_class->byval_arg;
 803                 for (i = 0; i < num_params; ++i)
 804                         csig->params [i] = &mono_defaults.int_class->byval_arg;
 805         }
 806
 807         mb = mono_mb_new (mono_defaults.object_class, name, MONO_WRAPPER_ALLOC);
 808
 809 #ifndef DISABLE_JIT
 810         size_var = mono_mb_add_local (mb, &mono_defaults.int_class->byval_arg);
 811         if (atype == ATYPE_NORMAL || atype == ATYPE_SMALL) {
 812                 /* size = vtable->klass->instance_size; */
 813                 mono_mb_emit_ldarg (mb, 0);
 814                 mono_mb_emit_icon (mb, MONO_STRUCT_OFFSET (MonoVTable, klass));
 815                 mono_mb_emit_byte (mb, CEE_ADD);
 816                 mono_mb_emit_byte (mb, CEE_LDIND_I);
 817                 mono_mb_emit_icon (mb, MONO_STRUCT_OFFSET (MonoClass, instance_size));
 818                 mono_mb_emit_byte (mb, CEE_ADD);
 819                 /* FIXME: assert instance_size stays a 4 byte integer */
 820                 mono_mb_emit_byte (mb, CEE_LDIND_U4);
 821                 mono_mb_emit_byte (mb, CEE_CONV_I);
 822                 mono_mb_emit_stloc (mb, size_var);
 823         } else if (atype == ATYPE_VECTOR) {
 824                 MonoExceptionClause *clause;
 825                 int pos, pos_leave, pos_error;
 826                 MonoClass *oom_exc_class;
 827                 MonoMethod *ctor;
 828
 829                 /*
 830                  * n > MONO_ARRAY_MAX_INDEX => OutOfMemoryException
 831                  * n < 0                    => OverflowException
 832                  *
 833                  * We can do an unsigned comparison to catch both cases, then in the error
 834                  * case compare signed to distinguish between them.
 835                  */
 836                 mono_mb_emit_ldarg (mb, 1);
 837                 mono_mb_emit_icon (mb, MONO_ARRAY_MAX_INDEX);
 838                 mono_mb_emit_byte (mb, CEE_CONV_U);
 839                 pos = mono_mb_emit_short_branch (mb, CEE_BLE_UN_S);
 840
 841                 mono_mb_emit_byte (mb, MONO_CUSTOM_PREFIX);
 842                 mono_mb_emit_byte (mb, CEE_MONO_NOT_TAKEN);
 843                 mono_mb_emit_ldarg (mb, 1);
 844                 mono_mb_emit_icon (mb, 0);
 845                 pos_error = mono_mb_emit_short_branch (mb, CEE_BLT_S);
 846                 mono_mb_emit_exception (mb, "OutOfMemoryException", NULL);
 847                 mono_mb_patch_short_branch (mb, pos_error);
 848                 mono_mb_emit_exception (mb, "OverflowException", NULL);
 849
 850                 mono_mb_patch_short_branch (mb, pos);
 851
 852                 clause = mono_image_alloc0 (mono_defaults.corlib, sizeof (MonoExceptionClause));
 853                 clause->try_offset = mono_mb_get_label (mb);
 854
 855                 /* vtable->klass->sizes.element_size */
 856                 mono_mb_emit_ldarg (mb, 0);
 857                 mono_mb_emit_icon (mb, MONO_STRUCT_OFFSET (MonoVTable, klass));
 858                 mono_mb_emit_byte (mb, CEE_ADD);
 859                 mono_mb_emit_byte (mb, CEE_LDIND_I);
 860                 mono_mb_emit_icon (mb, MONO_STRUCT_OFFSET (MonoClass, sizes));
 861                 mono_mb_emit_byte (mb, CEE_ADD);
 862                 mono_mb_emit_byte (mb, CEE_LDIND_U4);
 863                 mono_mb_emit_byte (mb, CEE_CONV_I);
 864
 865                 /* * n */
 866                 mono_mb_emit_ldarg (mb, 1);
 867                 mono_mb_emit_byte (mb, CEE_MUL_OVF_UN);
 868                 /* + sizeof (MonoArray) */
 869                 mono_mb_emit_icon (mb, sizeof (MonoArray));
 870                 mono_mb_emit_byte (mb, CEE_ADD_OVF_UN);
 871                 mono_mb_emit_stloc (mb, size_var);
 872
 873                 pos_leave = mono_mb_emit_branch (mb, CEE_LEAVE);
 874
 875                 /* catch */
 876                 clause->flags = MONO_EXCEPTION_CLAUSE_NONE;
 877                 clause->try_len = mono_mb_get_pos (mb) - clause->try_offset;
 878                 clause->data.catch_class = mono_class_from_name (mono_defaults.corlib,
 879                                 "System", "OverflowException");
 880                 g_assert (clause->data.catch_class);
 881                 clause->handler_offset = mono_mb_get_label (mb);
 882
 883                 oom_exc_class = mono_class_from_name (mono_defaults.corlib,
 884                                 "System", "OutOfMemoryException");
 885                 g_assert (oom_exc_class);
 886                 ctor = mono_class_get_method_from_name (oom_exc_class, ".ctor", 0);
 887                 g_assert (ctor);
 888
 889                 mono_mb_emit_byte (mb, CEE_POP);
 890                 mono_mb_emit_op (mb, CEE_NEWOBJ, ctor);
 891                 mono_mb_emit_byte (mb, CEE_THROW);
 892
 893                 clause->handler_len = mono_mb_get_pos (mb) - clause->handler_offset;
 894                 mono_mb_set_clauses (mb, 1, clause);
 895                 mono_mb_patch_branch (mb, pos_leave);
 896                 /* end catch */
 897         } else if (atype == ATYPE_STRING) {
 898                 int pos;
 899
 900                 /*
 901                  * a string allocator method takes the args: (vtable, len)
 902                  *
 903                  * bytes = offsetof (MonoString, chars) + ((len + 1) * 2)
 904                  *
 905                  * condition:
 906                  *
 907                  * bytes <= INT32_MAX - (SGEN_ALLOC_ALIGN - 1)
 908                  *
 909                  * therefore:
 910                  *
 911                  * offsetof (MonoString, chars) + ((len + 1) * 2) <= INT32_MAX - (SGEN_ALLOC_ALIGN - 1)
 912                  * len <= (INT32_MAX - (SGEN_ALLOC_ALIGN - 1) - offsetof (MonoString, chars)) / 2 - 1
 913                  */
 914                 mono_mb_emit_ldarg (mb, 1);
 915                 mono_mb_emit_icon (mb, (INT32_MAX - (SGEN_ALLOC_ALIGN - 1) - MONO_STRUCT_OFFSET (MonoString, chars)) / 2 - 1);
 916                 pos = mono_mb_emit_short_branch (mb, MONO_CEE_BLE_UN_S);
 917
 918                 mono_mb_emit_byte (mb, MONO_CUSTOM_PREFIX);
 919                 mono_mb_emit_byte (mb, CEE_MONO_NOT_TAKEN);
 920                 mono_mb_emit_exception (mb, "OutOfMemoryException", NULL);
 921                 mono_mb_patch_short_branch (mb, pos);
 922
 923                 mono_mb_emit_ldarg (mb, 1);
 924                 mono_mb_emit_icon (mb, 1);
 925                 mono_mb_emit_byte (mb, MONO_CEE_SHL);
 926                 //WE manually fold the above + 2 here
 927                 mono_mb_emit_icon (mb, MONO_STRUCT_OFFSET (MonoString, chars) + 2);
 928                 mono_mb_emit_byte (mb, CEE_ADD);
 929                 mono_mb_emit_stloc (mb, size_var);
 930         } else {
 931                 g_assert_not_reached ();
 932         }
 933
 934         /* size += ALLOC_ALIGN - 1; */
 935         mono_mb_emit_ldloc (mb, size_var);
 936         mono_mb_emit_icon (mb, ALLOC_ALIGN - 1);
 937         mono_mb_emit_byte (mb, CEE_ADD);
 938         /* size &= ~(ALLOC_ALIGN - 1); */
 939         mono_mb_emit_icon (mb, ~(ALLOC_ALIGN - 1));
 940         mono_mb_emit_byte (mb, CEE_AND);
 941         mono_mb_emit_stloc (mb, size_var);
 942
 943         /* if (size > MAX_SMALL_OBJ_SIZE) goto slowpath */
 944         if (atype != ATYPE_SMALL) {
 945                 mono_mb_emit_ldloc (mb, size_var);
 946                 mono_mb_emit_icon (mb, MAX_SMALL_OBJ_SIZE);
 947                 max_size_branch = mono_mb_emit_short_branch (mb, MONO_CEE_BGT_UN_S);
 948         }
 949
 950         /*
 951          * We need to modify tlab_next, but the JIT only supports reading, so we read
 952          * another tls var holding its address instead.
 953          */
 954
 955         /* tlab_next_addr (local) = tlab_next_addr (TLS var) */
 956         tlab_next_addr_var = mono_mb_add_local (mb, &mono_defaults.int_class->byval_arg);
 957         EMIT_TLS_ACCESS_NEXT_ADDR (mb);
 958         mono_mb_emit_stloc (mb, tlab_next_addr_var);
 959
 960         /* p = (void**)tlab_next; */
 961         p_var = mono_mb_add_local (mb, &mono_defaults.int_class->byval_arg);
 962         mono_mb_emit_ldloc (mb, tlab_next_addr_var);
 963         mono_mb_emit_byte (mb, CEE_LDIND_I);
 964         mono_mb_emit_stloc (mb, p_var);
 965
 966         /* new_next = (char*)p + size; */
 967         new_next_var = mono_mb_add_local (mb, &mono_defaults.int_class->byval_arg);
 968         mono_mb_emit_ldloc (mb, p_var);
 969         mono_mb_emit_ldloc (mb, size_var);
 970         mono_mb_emit_byte (mb, CEE_CONV_I);
 971         mono_mb_emit_byte (mb, CEE_ADD);
 972         mono_mb_emit_stloc (mb, new_next_var);
 973
 974         /* if (G_LIKELY (new_next < tlab_temp_end)) */
 975         mono_mb_emit_ldloc (mb, new_next_var);
 976         EMIT_TLS_ACCESS_TEMP_END (mb);
 977         slowpath_branch = mono_mb_emit_short_branch (mb, MONO_CEE_BLT_UN_S);
 978
 979         /* Slowpath */
 980         if (atype != ATYPE_SMALL)
 981                 mono_mb_patch_short_branch (mb, max_size_branch);
 982
 983         mono_mb_emit_byte (mb, MONO_CUSTOM_PREFIX);
 984         mono_mb_emit_byte (mb, CEE_MONO_NOT_TAKEN);
 985
 986         /* FIXME: mono_gc_alloc_obj takes a 'size_t' as an argument, not an int32 */
 987         mono_mb_emit_ldarg (mb, 0);
 988         mono_mb_emit_ldloc (mb, size_var);
 989         if (atype == ATYPE_NORMAL || atype == ATYPE_SMALL) {
 990                 mono_mb_emit_icall (mb, mono_gc_alloc_obj);
 991         } else if (atype == ATYPE_VECTOR) {
 992                 mono_mb_emit_ldarg (mb, 1);
 993                 mono_mb_emit_icall (mb, mono_gc_alloc_vector);
 994         } else if (atype == ATYPE_STRING) {
 995                 mono_mb_emit_ldarg (mb, 1);
 996                 mono_mb_emit_icall (mb, mono_gc_alloc_string);
 997         } else {
 998                 g_assert_not_reached ();
 999         }
1000         mono_mb_emit_byte (mb, CEE_RET);
1001
1002         /* Fastpath */
1003         mono_mb_patch_short_branch (mb, slowpath_branch);
1004
1005         /* FIXME: Memory barrier */
1006
1007         /* tlab_next = new_next */
1008         mono_mb_emit_ldloc (mb, tlab_next_addr_var);
1009         mono_mb_emit_ldloc (mb, new_next_var);
1010         mono_mb_emit_byte (mb, CEE_STIND_I);
1011
1012         /*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. */
1013         mono_mb_emit_byte ((mb), MONO_CUSTOM_PREFIX);
1014         mono_mb_emit_op (mb, CEE_MONO_MEMORY_BARRIER, (gpointer)StoreStoreBarrier);
1015
1016         /* *p = vtable; */
1017         mono_mb_emit_ldloc (mb, p_var);
1018         mono_mb_emit_ldarg (mb, 0);
1019         mono_mb_emit_byte (mb, CEE_STIND_I);
1020
1021         if (atype == ATYPE_VECTOR) {
1022                 /* arr->max_length = max_length; */
1023                 mono_mb_emit_ldloc (mb, p_var);
1024                 mono_mb_emit_ldflda (mb, MONO_STRUCT_OFFSET (MonoArray, max_length));
1025                 mono_mb_emit_ldarg (mb, 1);
1026 #ifdef MONO_BIG_ARRAYS
1027                 mono_mb_emit_byte (mb, CEE_STIND_I);
1028 #else
1029                 mono_mb_emit_byte (mb, CEE_STIND_I4);
1030 #endif
1031         } else  if (atype == ATYPE_STRING) {
1032                 /* need to set length and clear the last char */
1033                 /* s->length = len; */
1034                 mono_mb_emit_ldloc (mb, p_var);
1035                 mono_mb_emit_icon (mb, MONO_STRUCT_OFFSET (MonoString, length));
1036                 mono_mb_emit_byte (mb, MONO_CEE_ADD);
1037                 mono_mb_emit_ldarg (mb, 1);
1038                 mono_mb_emit_byte (mb, MONO_CEE_STIND_I4);
1039                 /* s->chars [len] = 0; */
1040                 mono_mb_emit_ldloc (mb, p_var);
1041                 mono_mb_emit_ldloc (mb, size_var);
1042                 mono_mb_emit_icon (mb, 2);
1043                 mono_mb_emit_byte (mb, MONO_CEE_SUB);
1044                 mono_mb_emit_byte (mb, MONO_CEE_ADD);
1045                 mono_mb_emit_icon (mb, 0);
1046                 mono_mb_emit_byte (mb, MONO_CEE_STIND_I2);
1047         }
1048
1049         /*
1050         We must make sure both vtable and max_length are globaly visible before returning to managed land.
1051         */
1052         mono_mb_emit_byte ((mb), MONO_CUSTOM_PREFIX);
1053         mono_mb_emit_op (mb, CEE_MONO_MEMORY_BARRIER, (gpointer)StoreStoreBarrier);
1054
1055         /* return p */
1056         mono_mb_emit_ldloc (mb, p_var);
1057         mono_mb_emit_byte (mb, CEE_RET);
1058 #endif
1059
1060         res = mono_mb_create_method (mb, csig, 8);
1061         mono_mb_free (mb);
1062         mono_method_get_header (res)->init_locals = FALSE;
1063
1064         info = mono_image_alloc0 (mono_defaults.corlib, sizeof (AllocatorWrapperInfo));
1065         info->gc_name = "sgen";
1066         info->alloc_type = atype;
1067         mono_marshal_set_wrapper_info (res, info);
1068
1069         return res;
1070 }
1071 #endif
1072
1073 /*
1074  * Generate an allocator method implementing the fast path of mono_gc_alloc_obj ().
1075  * The signature of the called method is:
1076  *      object allocate (MonoVTable *vtable)
1077  */
1078 MonoMethod*
1079 mono_gc_get_managed_allocator (MonoClass *klass, gboolean for_box)
1080 {
1081 #ifdef MANAGED_ALLOCATION
1082
1083 #ifdef HAVE_KW_THREAD
1084         int tlab_next_offset = -1;
1085         int tlab_temp_end_offset = -1;
1086         MONO_THREAD_VAR_OFFSET (tlab_next, tlab_next_offset);
1087         MONO_THREAD_VAR_OFFSET (tlab_temp_end, tlab_temp_end_offset);
1088
1089         if (tlab_next_offset == -1 || tlab_temp_end_offset == -1)
1090                 return NULL;
1091 #endif
1092         if (collect_before_allocs)
1093                 return NULL;
1094         if (!mono_runtime_has_tls_get ())
1095                 return NULL;
1096         if (klass->instance_size > tlab_size)
1097                 return NULL;
1098
1099         if (klass->has_finalize || mono_class_is_marshalbyref (klass) || (mono_profiler_get_events () & MONO_PROFILE_ALLOCATIONS))
1100                 return NULL;
1101         if (klass->rank)
1102                 return NULL;
1103         if (klass->byval_arg.type == MONO_TYPE_STRING)
1104                 return mono_gc_get_managed_allocator_by_type (ATYPE_STRING);
1105         /* Generic classes have dynamic field and can go above MAX_SMALL_OBJ_SIZE. */
1106         if (ALIGN_TO (klass->instance_size, ALLOC_ALIGN) < MAX_SMALL_OBJ_SIZE && !mono_class_is_open_constructed_type (&klass->byval_arg))
1107                 return mono_gc_get_managed_allocator_by_type (ATYPE_SMALL);
1108         else
1109                 return mono_gc_get_managed_allocator_by_type (ATYPE_NORMAL);
1110 #else
1111         return NULL;
1112 #endif
1113 }
1114
1115 MonoMethod*
1116 mono_gc_get_managed_array_allocator (MonoClass *klass)
1117 {
1118 #ifdef MANAGED_ALLOCATION
1119 #ifdef HAVE_KW_THREAD
1120         int tlab_next_offset = -1;
1121         int tlab_temp_end_offset = -1;
1122         MONO_THREAD_VAR_OFFSET (tlab_next, tlab_next_offset);
1123         MONO_THREAD_VAR_OFFSET (tlab_temp_end, tlab_temp_end_offset);
1124
1125         if (tlab_next_offset == -1 || tlab_temp_end_offset == -1)
1126                 return NULL;
1127 #endif
1128
1129         if (klass->rank != 1)
1130                 return NULL;
1131         if (!mono_runtime_has_tls_get ())
1132                 return NULL;
1133         if (mono_profiler_get_events () & MONO_PROFILE_ALLOCATIONS)
1134                 return NULL;
1135         if (has_per_allocation_action)
1136                 return NULL;
1137         g_assert (!mono_class_has_finalizer (klass) && !mono_class_is_marshalbyref (klass));
1138
1139         return mono_gc_get_managed_allocator_by_type (ATYPE_VECTOR);
1140 #else
1141         return NULL;
1142 #endif
1143 }
1144
1145 void
1146 sgen_set_use_managed_allocator (gboolean flag)
1147 {
1148         use_managed_allocator = flag;
1149 }
1150
1151 MonoMethod*
1152 mono_gc_get_managed_allocator_by_type (int atype)
1153 {
1154 #ifdef MANAGED_ALLOCATION
1155         MonoMethod *res;
1156
1157         if (!use_managed_allocator)
1158                 return NULL;
1159
1160         if (!mono_runtime_has_tls_get ())
1161                 return NULL;
1162
1163         res = alloc_method_cache [atype];
1164         if (res)
1165                 return res;
1166
1167         res = create_allocator (atype);
1168         LOCK_GC;
1169         if (alloc_method_cache [atype]) {
1170                 mono_free_method (res);
1171                 res = alloc_method_cache [atype];
1172         } else {
1173                 mono_memory_barrier ();
1174                 alloc_method_cache [atype] = res;
1175         }
1176         UNLOCK_GC;
1177
1178         return res;
1179 #else
1180         return NULL;
1181 #endif
1182 }
1183
1184 guint32
1185 mono_gc_get_managed_allocator_types (void)
1186 {
1187         return ATYPE_NUM;
1188 }
1189
1190 gboolean
1191 sgen_is_managed_allocator (MonoMethod *method)
1192 {
1193         int i;
1194
1195         for (i = 0; i < ATYPE_NUM; ++i)
1196                 if (method == alloc_method_cache [i])
1197                         return TRUE;
1198         return FALSE;
1199 }
1200
1201 gboolean
1202 sgen_has_managed_allocator (void)
1203 {
1204         int i;
1205
1206         for (i = 0; i < ATYPE_NUM; ++i)
1207                 if (alloc_method_cache [i])
1208                         return TRUE;
1209         return FALSE;
1210 }
1211
1212 #ifdef HEAVY_STATISTICS
1213 void
1214 sgen_alloc_init_heavy_stats (void)
1215 {
1216         mono_counters_register ("# objects allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_objects_alloced);
1217         mono_counters_register ("bytes allocated", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_bytes_alloced);
1218         mono_counters_register ("bytes allocated in LOS", MONO_COUNTER_GC | MONO_COUNTER_ULONG, &stat_bytes_alloced_los);
1219 }
1220 #endif
1221
1222 #endif /*HAVE_SGEN_GC*/