2 * mini-gc.c: GC interface for the mono JIT
5 * Zoltan Varga (vargaz@gmail.com)
7 * Copyright 2009 Novell, Inc (http://www.novell.com)
12 #include <mono/metadata/gc-internal.h>
17 #include <mono/metadata/gc-internal.h>
18 #include <mono/utils/mono-counters.h>
20 /* Contains state needed by the GC Map construction code */
23 * This contains information about stack slots initialized in the prolog, encoded using
24 * (slot_index << 16) | slot_type. The slot_index is relative to the CFA, i.e. 0
25 * means cfa+0, 1 means cfa-4/8, etc.
27 GSList *stack_slots_from_cfa;
28 /* Same for stack slots relative to the frame pointer */
29 GSList *stack_slots_from_fp;
31 /* Number of slots in the map */
33 /* The number of registers in the map */
35 /* Min and Max offsets of the stack frame relative to fp */
36 int min_offset, max_offset;
37 /* Same for the locals area */
38 int locals_min_offset, locals_max_offset;
40 /* The call sites where this frame can be stopped during GC */
41 GCCallSite **callsites;
42 /* The number of call sites */
46 * The width of the stack bitmaps in bytes. This is not equal to the bitmap width at
47 * runtime, since it includes columns which are 0.
49 int stack_bitmap_width;
51 * A bitmap whose width equals nslots, and whose height equals ncallsites.
52 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
55 guint8 *stack_ref_bitmap;
56 /* Same for SLOT_PIN */
57 guint8 *stack_pin_bitmap;
60 * Similar bitmaps for registers. These have width MONO_MAX_IREGS in bits.
63 guint8 *reg_ref_bitmap;
64 guint8 *reg_pin_bitmap;
67 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
70 #define DEBUG(s) do { s; } while (0)
71 #define DEBUG_ENABLED 1
77 #define DEBUG_GC_MAP(s) do { s; fflush (stdout); } while (0)
79 #define DEBUG_GC_MAP(s)
82 #define GC_BITS_PER_WORD (sizeof (gsize) * 8)
85 * Contains information collected during the conservative stack marking pass,
86 * used during the precise pass. This helps to avoid doing a stack walk twice, which
92 int frame_start_offset;
94 /* Relative to stack_start */
95 int reg_locations [MONO_MAX_IREGS];
99 int regs [MONO_MAX_IREGS];
103 /* Max number of frames stored in the TLS data */
104 #define MAX_FRAMES 50
107 * Per-thread data kept by this module. This is stored in the GC and passed to us as
108 * parameters, instead of being stored in a TLS variable, since during a collection,
109 * only the collection thread is active.
114 gboolean has_context;
115 MonoJitTlsData *jit_tls;
116 /* Number of frames collected during the !precise pass */
118 FrameInfo frames [MAX_FRAMES];
121 /* These are constant so don't store them in the GC Maps */
122 /* Number of registers stored in gc maps */
123 #define NREGS MONO_MAX_IREGS
127 * Contains information needed to mark a stack frame.
128 * This is a transient structure, created from a compressed representation on-demand.
132 * The offsets of the GC tracked area inside the stack frame relative to the frame pointer.
133 * This includes memory which is NOREF thus doesn't need GC maps.
138 * The offset relative to frame_offset where the the memory described by the GC maps
142 /* The number of stack slots in the map */
144 /* The frame pointer register */
146 /* The size of each callsite table entry */
147 guint8 callsite_entry_size;
148 guint has_pin_slots : 1;
149 guint has_ref_slots : 1;
150 guint has_ref_regs : 1;
151 guint has_pin_regs : 1;
153 /* The offsets below are into an external bitmaps array */
156 * A bitmap whose width is equal to bitmap_width, and whose height is equal to ncallsites.
157 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
160 guint32 stack_ref_bitmap_offset;
162 * Same for SLOT_PIN. It is possible that the same bit is set in both bitmaps at
163 * different callsites, if the slot starts out as PIN, and later changes to REF.
165 guint32 stack_pin_bitmap_offset;
168 * Corresponding bitmaps for registers
169 * These have width equal to the number of bits set in reg_ref_mask/reg_pin_mask.
170 * FIXME: Merge these with the normal bitmaps, i.e. reserve the first x slots for them ?
172 guint32 reg_pin_bitmap_offset;
173 guint32 reg_ref_bitmap_offset;
175 guint32 used_int_regs, reg_ref_mask, reg_pin_mask;
177 /* The number of bits set in the two masks above */
178 guint8 nref_regs, npin_regs;
181 * A bit array marking slots which contain refs.
182 * This is used only for debugging.
186 /* Callsite offsets */
187 /* These can take up a lot of space, so encode them compactly */
197 * A compressed version of GCMap. This is what gets stored in MonoJitInfo.
201 //guint8 encoded_size;
204 * The arrays below are embedded after the struct.
205 * Their address needs to be computed.
208 /* The fixed fields of the GCMap encoded using LEB128 */
209 guint8 encoded [MONO_ZERO_LEN_ARRAY];
211 /* An array of ncallsites entries, each entry is callsite_entry_size bytes long */
212 guint8 callsites [MONO_ZERO_LEN_ARRAY];
215 guint8 bitmaps [MONO_ZERO_LEN_ARRAY];
218 static int precise_frame_count [2], precise_frame_limit = -1;
219 static gboolean precise_frame_limit_inited;
225 int scanned_precisely;
226 int scanned_conservatively;
227 int scanned_registers;
235 int gc_callsites_size;
236 int gc_callsites8_size;
237 int gc_callsites16_size;
238 int gc_callsites32_size;
240 int gc_map_struct_size;
244 static JITGCStats stats;
246 // FIXME: Move these to a shared place
249 encode_uleb128 (guint32 value, guint8 *buf, guint8 **endbuf)
254 guint8 b = value & 0x7f;
256 if (value != 0) /* more bytes to come */
264 static G_GNUC_UNUSED void
265 encode_sleb128 (gint32 value, guint8 *buf, guint8 **endbuf)
268 gboolean negative = (value < 0);
276 /* the following is unnecessary if the
277 * implementation of >>= uses an arithmetic rather
278 * than logical shift for a signed left operand
282 value |= - (1 <<(size - 7));
283 /* sign bit of byte is second high order bit (0x40) */
284 if ((value == 0 && !(byte & 0x40)) ||
285 (value == -1 && (byte & 0x40)))
295 static inline guint32
296 decode_uleb128 (guint8 *buf, guint8 **endbuf)
306 res = res | (((int)(b & 0x7f)) << shift);
318 decode_sleb128 (guint8 *buf, guint8 **endbuf)
328 res = res | (((int)(b & 0x7f)) << shift);
331 if (shift < 32 && (b & 0x40))
332 res |= - (1 << shift);
343 encode_frame_reg (int frame_reg)
346 if (frame_reg == AMD64_RSP)
348 else if (frame_reg == AMD64_RBP)
353 g_assert_not_reached ();
358 decode_frame_reg (int encoded)
363 else if (encoded == 1)
368 g_assert_not_reached ();
374 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15, AMD64_RDI, AMD64_RSI };
376 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15 };
381 encode_regmask (guint32 regmask)
388 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
389 if (regmask & (1 << callee_saved_regs [i])) {
391 regmask -= (1 << callee_saved_regs [i]);
394 g_assert (regmask == 0);
403 decode_regmask (guint32 regmask)
410 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i)
411 if (regmask & (1 << i))
412 res |= (1 << callee_saved_regs [i]);
423 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
426 encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
430 encode_sleb128 (map->start_offset / sizeof (mgreg_t), buf, &buf);
431 encode_sleb128 (map->map_offset / sizeof (mgreg_t), buf, &buf);
432 encode_uleb128 (map->nslots, buf, &buf);
433 g_assert (map->callsite_entry_size <= 4);
434 freg = encode_frame_reg (map->frame_reg);
436 flags = (map->has_ref_slots ? 1 : 0) | (map->has_pin_slots ? 2 : 0) | (map->has_ref_regs ? 4 : 0) | (map->has_pin_regs ? 8 : 0) | ((map->callsite_entry_size - 1) << 4) | (freg << 6);
437 encode_uleb128 (flags, buf, &buf);
438 encode_uleb128 (encode_regmask (map->used_int_regs), buf, &buf);
439 if (map->has_ref_regs)
440 encode_uleb128 (encode_regmask (map->reg_ref_mask), buf, &buf);
441 if (map->has_pin_regs)
442 encode_uleb128 (encode_regmask (map->reg_pin_mask), buf, &buf);
443 encode_uleb128 (map->ncallsites, buf, &buf);
451 * Decode the encoded GC map representation in BUF and store the result into MAP.
454 decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
457 int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
460 map->start_offset = decode_sleb128 (buf, &buf) * sizeof (mgreg_t);
461 map->map_offset = decode_sleb128 (buf, &buf) * sizeof (mgreg_t);
462 map->nslots = decode_uleb128 (buf, &buf);
463 flags = decode_uleb128 (buf, &buf);
464 map->has_ref_slots = (flags & 1) ? 1 : 0;
465 map->has_pin_slots = (flags & 2) ? 1 : 0;
466 map->has_ref_regs = (flags & 4) ? 1 : 0;
467 map->has_pin_regs = (flags & 8) ? 1 : 0;
468 map->callsite_entry_size = ((flags >> 4) & 0x3) + 1;
470 map->frame_reg = decode_frame_reg (freg);
471 map->used_int_regs = decode_regmask (decode_uleb128 (buf, &buf));
472 if (map->has_ref_regs) {
473 map->reg_ref_mask = decode_regmask (decode_uleb128 (buf, &buf));
475 for (i = 0; i < NREGS; ++i)
476 if (map->reg_ref_mask & (1 << i))
480 if (map->has_pin_regs) {
481 map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
483 for (i = 0; i < NREGS; ++i)
484 if (map->reg_pin_mask & (1 << i))
488 map->ncallsites = decode_uleb128 (buf, &buf);
490 stack_bitmap_size = (ALIGN_TO (map->nslots, 8) / 8) * map->ncallsites;
491 reg_ref_bitmap_size = (ALIGN_TO (map->nref_regs, 8) / 8) * map->ncallsites;
492 reg_pin_bitmap_size = (ALIGN_TO (map->npin_regs, 8) / 8) * map->ncallsites;
494 map->stack_ref_bitmap_offset = offset;
495 if (map->has_ref_slots)
496 offset += stack_bitmap_size;
497 map->stack_pin_bitmap_offset = offset;
498 if (map->has_pin_slots)
499 offset += stack_bitmap_size;
500 map->reg_ref_bitmap_offset = offset;
501 if (map->has_ref_regs)
502 offset += reg_ref_bitmap_size;
503 map->reg_pin_bitmap_offset = offset;
504 if (map->has_pin_regs)
505 offset += reg_pin_bitmap_size;
511 thread_attach_func (void)
515 tls = g_new0 (TlsData, 1);
516 stats.tlsdata_size += sizeof (TlsData);
522 thread_detach_func (gpointer user_data)
524 TlsData *tls = user_data;
530 thread_suspend_func (gpointer user_data, void *sigctx)
532 TlsData *tls = user_data;
535 /* Happens during startup */
538 tls->lmf = mono_get_lmf ();
540 mono_arch_sigctx_to_monoctx (sigctx, &tls->ctx);
541 tls->has_context = TRUE;
543 tls->has_context = FALSE;
545 tls->jit_tls = TlsGetValue (mono_jit_tls_id);
548 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
551 set_bit (guint8 *bitmap, int width, int y, int x)
553 bitmap [(width * y) + (x / 8)] |= (1 << (x % 8));
557 clear_bit (guint8 *bitmap, int width, int y, int x)
559 bitmap [(width * y) + (x / 8)] &= ~(1 << (x % 8));
563 get_bit (guint8 *bitmap, int width, int y, int x)
565 return bitmap [(width * y) + (x / 8)] & (1 << (x % 8));
569 slot_type_to_string (GCSlotType type)
579 g_assert_not_reached ();
585 * conservatively_pass:
587 * Mark a thread stack conservatively and collect information needed by the precise pass.
590 conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
593 MonoContext ctx, new_ctx;
596 gboolean last = TRUE;
600 guint8* fp, *p, *frame_start, *frame_end;
601 int i, pc_offset, cindex, bitmap_width;
602 int scanned = 0, scanned_precisely, scanned_conservatively, scanned_registers;
604 StackFrameInfo frame;
605 mgreg_t *reg_locations [MONO_MAX_IREGS];
606 mgreg_t *new_reg_locations [MONO_MAX_IREGS];
609 guint32 precise_regmask;
611 /* tls == NULL can happen during startup */
612 if (mono_thread_internal_current () == NULL || !tls) {
613 mono_gc_conservatively_scan_area (stack_start, stack_end);
614 stats.scanned_stacks += stack_end - stack_start;
621 /* Number of bytes scanned based on GC map data */
623 /* Number of bytes scanned precisely based on GC map data */
624 scanned_precisely = 0;
625 /* Number of bytes scanned conservatively based on GC map data */
626 scanned_conservatively = 0;
627 /* Number of bytes scanned conservatively in register save areas */
628 scanned_registers = 0;
630 /* This is one past the last address which we have scanned */
631 stack_limit = stack_start;
633 if (!tls->has_context)
634 memset (&new_ctx, 0, sizeof (ctx));
636 memcpy (&new_ctx, &tls->ctx, sizeof (MonoContext));
638 memset (reg_locations, 0, sizeof (reg_locations));
639 memset (new_reg_locations, 0, sizeof (new_reg_locations));
644 memcpy (&ctx, &new_ctx, sizeof (ctx));
646 for (i = 0; i < MONO_MAX_IREGS; ++i) {
647 if (new_reg_locations [i]) {
649 * If the current frame saves the register, it means it might modify its
650 * value, thus the old location might not contain the same value, so
651 * we have to mark it conservatively.
652 * FIXME: This happens very often, due to:
653 * - outside the live intervals of the variables allocated to a register,
654 * we have to treat the register as PIN, since we don't know whenever it
655 * has the same value as in the caller, or a new dead value.
657 if (reg_locations [i]) {
658 DEBUG (printf ("\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
659 mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
660 scanned_registers += sizeof (mgreg_t);
663 reg_locations [i] = new_reg_locations [i];
665 DEBUG (printf ("\treg %s is at location %p.\n", mono_arch_regname (i), reg_locations [i]));
669 g_assert ((guint64)stack_limit % sizeof (mgreg_t) == 0);
671 #ifdef MONO_ARCH_HAVE_FIND_JIT_INFO_EXT
672 res = mono_find_jit_info_ext (frame.domain ? frame.domain : mono_domain_get (), tls->jit_tls, NULL, &ctx, &new_ctx, NULL, &lmf, new_reg_locations, &frame);
679 /* The last frame can be in any state so mark conservatively */
687 /* This happens with native-to-managed transitions */
688 if (!(MONO_CONTEXT_GET_IP (&ctx) >= ji->code_start && (guint8*)MONO_CONTEXT_GET_IP (&ctx) < (guint8*)ji->code_start + ji->code_size))
691 /* These frames are very problematic */
692 if (ji->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
693 DEBUG (char *fname = mono_method_full_name (ji->method, TRUE); printf ("Mark(0): %s+0x%x (%p)\n", fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx)); g_free (fname));
694 DEBUG (printf ("\tSkip.\n"));
699 /* FIXME: Some wrappers do not declare variables with the proper GC type */
700 if (ji->method->wrapper_type)
704 /* All the other frames are at a call site */
706 if (tls->nframes == MAX_FRAMES) {
708 * Can't save information since the array is full. So scan the rest of the
709 * stack conservatively.
714 /* Scan the frame of this method */
717 * A frame contains the following:
722 * - localloc-ed memory
724 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
725 g_assert (pc_offset >= 0);
730 DEBUG (char *fname = mono_method_full_name (ji->method, TRUE); printf ("Mark(0): %s+0x%x (%p)\n", fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx)); g_free (fname));
731 DEBUG (printf ("\tNo GC Map.\n"));
736 * Debugging aid to control the number of frames scanned precisely
738 if (!precise_frame_limit_inited) {
739 if (getenv ("MONO_PRECISE_COUNT"))
740 precise_frame_limit = atoi (getenv ("MONO_PRECISE_COUNT"));
741 precise_frame_limit_inited = TRUE;
744 if (precise_frame_limit != -1) {
745 if (precise_frame_count [FALSE] == precise_frame_limit)
746 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (ji->method, TRUE));
747 if (precise_frame_count [FALSE] > precise_frame_limit)
750 precise_frame_count [FALSE] ++;
752 /* Decode the encoded GC map */
754 memset (map, 0, sizeof (GCMap));
755 decode_gc_map (&emap->encoded [0], map, &p);
756 p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
757 map->callsites.offsets8 = p;
758 p += map->callsite_entry_size * map->ncallsites;
762 if (map->frame_reg == AMD64_RSP)
763 fp = (guint8*)ctx.rsp;
764 else if (map->frame_reg == AMD64_RBP)
765 fp = (guint8*)ctx.rbp;
767 g_assert_not_reached ();
770 g_assert_not_reached ();
773 frame_start = fp + map->start_offset + map->map_offset;
774 frame_end = fp + map->end_offset;
776 DEBUG (char *fname = mono_method_full_name (ji->method, TRUE); printf ("Mark(0): %s+0x%x (%p) limit=%p fp=%p frame=%p-%p (%d)\n", fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx), stack_limit, fp, frame_start, frame_end, (int)(frame_end - frame_start)); g_free (fname));
778 /* Find the callsite index */
779 if (ji->code_size < 256) {
780 for (i = 0; i < map->ncallsites; ++i)
781 /* ip points inside the call instruction */
782 if (map->callsites.offsets8 [i] == pc_offset + 1)
784 } else if (ji->code_size < 65536) {
785 // FIXME: Use a binary search
786 for (i = 0; i < map->ncallsites; ++i)
787 /* ip points inside the call instruction */
788 if (map->callsites.offsets16 [i] == pc_offset + 1)
791 // FIXME: Use a binary search
792 for (i = 0; i < map->ncallsites; ++i)
793 /* ip points inside the call instruction */
794 if (map->callsites.offsets32 [i] == pc_offset + 1)
797 if (i == map->ncallsites) {
798 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (ji->method, TRUE));
799 g_assert_not_reached ();
803 g_assert (frame_start >= stack_limit);
805 if (frame_start > stack_limit) {
806 /* This scans the previously skipped frames as well */
807 DEBUG (printf ("\tscan area %p-%p.\n", stack_limit, frame_start));
808 mono_gc_conservatively_scan_area (stack_limit, frame_start);
811 /* Mark stack slots */
812 if (map->has_pin_slots) {
813 int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
814 guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
819 for (i = 0; i < map->nslots; ++i) {
820 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
822 DEBUG (printf ("\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
823 mono_gc_conservatively_scan_area (p, p + sizeof (mgreg_t));
824 scanned_conservatively += sizeof (mgreg_t);
826 scanned_precisely += sizeof (mgreg_t);
828 p += sizeof (mgreg_t);
831 scanned_precisely += (map->nslots * sizeof (mgreg_t));
834 /* The area outside of start-end is NOREF */
835 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * sizeof (mgreg_t));
838 precise_regmask = map->used_int_regs;
839 if (map->has_pin_regs) {
840 int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
841 guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
843 for (i = 0; i < NREGS; ++i) {
844 if (!(map->used_int_regs & (1 << i)))
847 if (!(map->reg_pin_mask & (1 << i)))
850 if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
851 DEBUG (printf ("\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
852 precise_regmask &= ~(1 << i);
858 scanned += map->end_offset - map->start_offset;
860 g_assert (scanned == scanned_precisely + scanned_conservatively);
862 stack_limit = frame_end;
864 /* Save information for the precise pass */
865 fi = &tls->frames [tls->nframes];
866 fi->nslots = map->nslots;
867 bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
868 if (map->has_ref_slots)
869 fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
872 fi->frame_start_offset = frame_start - stack_start;
873 fi->nreg_locations = 0;
877 if (map->has_ref_regs) {
878 int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
879 guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
881 for (i = 0; i < NREGS; ++i) {
882 if (!(map->reg_ref_mask & (1 << i)))
885 if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
886 DEBUG (fi->regs [fi->nreg_locations] = i);
887 DEBUG (printf ("\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
888 fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
889 fi->nreg_locations ++;
896 * Clear locations of precisely stacked registers.
898 if (precise_regmask) {
899 for (i = 0; i < NREGS; ++i) {
900 if (precise_regmask & (1 << i)) {
902 * The method uses this register, and we have precise info for it.
903 * This means the location will be scanned precisely.
904 * Tell the code at the beginning of the loop that this location is
907 if (reg_locations [i])
908 DEBUG (printf ("\treg %s at location %p is precise.\n", mono_arch_regname (i), reg_locations [i]));
909 reg_locations [i] = NULL;
917 /* Scan the remaining register save locations */
918 for (i = 0; i < MONO_MAX_IREGS; ++i) {
919 if (reg_locations [i]) {
920 DEBUG (printf ("\tscan saved reg location %p.\n", reg_locations [i]));
921 mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
922 scanned_registers += sizeof (mgreg_t);
924 // FIXME: Is this needed ?
925 if (new_reg_locations [i]) {
926 DEBUG (printf ("\tscan saved reg location %p.\n", new_reg_locations [i]));
927 mono_gc_conservatively_scan_area (new_reg_locations [i], new_reg_locations [i] + sizeof (mgreg_t));
928 scanned_registers += sizeof (mgreg_t);
932 if (stack_limit < stack_end) {
933 DEBUG (printf ("\tscan area %p-%p.\n", stack_limit, stack_end));
934 mono_gc_conservatively_scan_area (stack_limit, stack_end);
937 DEBUG (printf ("Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
939 stats.scanned_stacks += stack_end - stack_start;
940 stats.scanned += scanned;
941 stats.scanned_precisely += scanned_precisely;
942 stats.scanned_conservatively += scanned_conservatively;
943 stats.scanned_registers += scanned_registers;
945 //mono_gc_conservatively_scan_area (stack_start, stack_end);
951 * Mark a thread stack precisely based on information saved during the conservative
955 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
964 for (findex = 0; findex < tls->nframes; findex ++) {
965 /* Load information saved by the !precise pass */
966 fi = &tls->frames [findex];
967 frame_start = stack_start + fi->frame_start_offset;
970 * FIXME: Add a function to mark using a bitmap, to avoid doing a
971 * call for each object.
974 /* Mark stack slots */
976 guint8 *ref_bitmap = fi->bitmap;
979 for (i = 0; i < fi->nslots; ++i) {
980 MonoObject **ptr = (MonoObject**)(frame_start + (i * sizeof (mgreg_t)));
982 live = ref_bitmap [i / 8] & (1 << (i % 8));
985 MonoObject *obj = *ptr;
987 DEBUG (printf ("\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
988 *ptr = mono_gc_scan_object (obj);
989 DEBUG (printf (" %p.\n", *ptr));
991 DEBUG (printf ("\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
996 * This is disabled because the pointer takes up a lot of space.
997 * Stack slots might be shared between ref and non-ref variables ?
999 if (map->ref_slots [i / 8] & (1 << (i % 8))) {
1000 DEBUG (printf ("\tref %s0x%x(fp)=%p: dead (%p)\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, *ptr));
1002 * Fail fast if the live range is incorrect, and
1003 * the JITted code tries to access this object
1012 /* Mark registers */
1015 * Registers are different from stack slots, they have no address where they
1016 * are stored. Instead, some frame below this frame in the stack saves them
1017 * in its prolog to the stack. We can mark this location precisely.
1019 for (i = 0; i < fi->nreg_locations; ++i) {
1021 * reg_locations [i] contains the address of the stack slot where
1022 * a reg was last saved, so mark that slot.
1024 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1025 MonoObject *obj = *ptr;
1028 DEBUG (printf ("\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
1029 *ptr = mono_gc_scan_object (obj);
1030 DEBUG (printf (" %p.\n", *ptr));
1032 DEBUG (printf ("\treg %s saved at %p: %p", mono_arch_regname (fi->regs [i]), ptr, obj));
1041 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1042 * call, and TRUE at the second. USER_DATA points to a TlsData
1043 * structure filled up by thread_suspend_func.
1046 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise)
1048 TlsData *tls = user_data;
1050 DEBUG (printf ("*** %s stack marking %p-%p ***\n", precise ? "Precise" : "Conservative", stack_start, stack_end));
1053 conservative_pass (tls, stack_start, stack_end);
1055 precise_pass (tls, stack_start, stack_end);
1059 mini_gc_init_gc_map (MonoCompile *cfg)
1061 if (COMPILE_LLVM (cfg))
1065 /* Debugging support */
1067 static int precise_count;
1070 if (getenv ("MONO_GCMAP_COUNT")) {
1071 if (precise_count == atoi (getenv ("MONO_GCMAP_COUNT")))
1072 printf ("LAST: %s\n", mono_method_full_name (cfg->method, TRUE));
1073 if (precise_count > atoi (getenv ("MONO_GCMAP_COUNT")))
1079 cfg->compute_gc_maps = TRUE;
1081 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1085 * mini_gc_set_slot_type_from_fp:
1087 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1088 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1089 * need to call this function for those slots.
1092 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1094 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1096 if (!cfg->compute_gc_maps)
1099 g_assert (slot_offset % sizeof (mgreg_t) == 0);
1101 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1105 * mini_gc_set_slot_type_from_cfa:
1107 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1108 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1109 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1110 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1111 * call this function for those slots.
1114 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1116 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1117 int slot = - (slot_offset / sizeof (mgreg_t));
1119 if (!cfg->compute_gc_maps)
1122 g_assert (slot_offset <= 0);
1123 g_assert (slot_offset % sizeof (mgreg_t) == 0);
1125 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1129 fp_offset_to_slot (MonoCompile *cfg, int offset)
1131 MonoCompileGC *gcfg = cfg->gc_info;
1133 return (offset - gcfg->min_offset) / sizeof (mgreg_t);
1137 slot_to_fp_offset (MonoCompile *cfg, int slot)
1139 MonoCompileGC *gcfg = cfg->gc_info;
1141 return (slot * sizeof (mgreg_t)) + gcfg->min_offset;
1145 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1147 g_assert (slot >= 0 && slot < gcfg->nslots);
1149 if (type == SLOT_PIN) {
1150 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1151 set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1152 } else if (type == SLOT_REF) {
1153 set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1154 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1155 } else if (type == SLOT_NOREF) {
1156 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1157 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1162 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1166 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1167 set_slot (gcfg, slot, cindex, type);
1171 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1175 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1176 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1177 if (callsite_offset >= from && callsite_offset < to)
1178 set_slot (gcfg, slot, cindex, type);
1183 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1185 g_assert (slot >= 0 && slot < gcfg->nregs);
1187 if (type == SLOT_PIN) {
1188 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1189 set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1190 } else if (type == SLOT_REF) {
1191 set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1192 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1193 } else if (type == SLOT_NOREF) {
1194 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1195 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1200 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1204 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1205 set_reg_slot (gcfg, slot, cindex, type);
1209 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1213 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1214 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1215 if (callsite_offset >= from && callsite_offset < to)
1216 set_reg_slot (gcfg, slot, cindex, type);
1221 process_spill_slots (MonoCompile *cfg)
1223 MonoCompileGC *gcfg = cfg->gc_info;
1228 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1229 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1230 for (l = bb->spill_slot_defs; l; l = l->next) {
1231 MonoInst *def = l->data;
1232 int spill_slot = def->inst_c0;
1233 int bank = def->inst_c1;
1234 int offset = cfg->spill_info [bank][spill_slot].offset;
1235 int slot = fp_offset_to_slot (cfg, offset);
1237 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1238 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1242 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1243 for (l = bb->spill_slot_defs; l; l = l->next) {
1244 MonoInst *def = l->data;
1245 int spill_slot = def->inst_c0;
1246 int bank = def->inst_c1;
1247 int offset = cfg->spill_info [bank][spill_slot].offset;
1248 int slot = fp_offset_to_slot (cfg, offset);
1251 if (bank == MONO_REG_INT_MP)
1257 * Extend the live interval for the GC tracked spill slots
1258 * defined in this bblock.
1259 * FIXME: This is not needed.
1261 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1263 if (cfg->verbose_level > 1)
1264 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
1268 /* Set fp spill slots to NOREF */
1269 for (i = 0; i < cfg->spill_info_len [MONO_REG_DOUBLE]; ++i) {
1270 int offset = cfg->spill_info [MONO_REG_DOUBLE][i].offset;
1276 slot = fp_offset_to_slot (cfg, offset);
1278 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1280 if (cfg->verbose_level > 1)
1281 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1284 /* Set int spill slots to NOREF */
1285 for (i = 0; i < cfg->spill_info_len [MONO_REG_INT]; ++i) {
1286 int offset = cfg->spill_info [MONO_REG_INT][i].offset;
1292 slot = fp_offset_to_slot (cfg, offset);
1294 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1295 if (cfg->verbose_level > 1)
1296 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1301 * process_other_slots:
1303 * Process stack slots registered using mini_gc_set_slot_type_... ().
1306 process_other_slots (MonoCompile *cfg)
1308 MonoCompileGC *gcfg = cfg->gc_info;
1311 /* Relative to the CFA */
1312 for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
1313 guint data = GPOINTER_TO_UINT (l->data);
1314 int cfa_slot = data >> 16;
1315 GCSlotType type = data & 0xff;
1319 * Map the cfa relative slot to an fp relative slot.
1320 * slot_addr == cfa - <cfa_slot>*4/8
1321 * fp + cfa_offset == cfa
1322 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1324 slot = (cfg->cfa_offset / sizeof (mgreg_t)) - cfa_slot - (gcfg->min_offset / sizeof (mgreg_t));
1326 set_slot_everywhere (gcfg, slot, type);
1328 if (cfg->verbose_level > 1) {
1329 int fp_offset = slot_to_fp_offset (cfg, slot);
1330 if (type == SLOT_NOREF)
1331 printf ("\tnoref slot at %s0x%x(fp) (slot = %d) (cfa - 0x%x)\n", fp_offset >= 0 ? "" : "-", ABS (fp_offset), slot, (int)(cfa_slot * sizeof (mgreg_t)));
1335 /* Relative to the FP */
1336 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1337 gint data = GPOINTER_TO_INT (l->data);
1338 int offset = data >> 16;
1339 GCSlotType type = data & 0xff;
1342 slot = fp_offset_to_slot (cfg, offset);
1344 set_slot_everywhere (gcfg, slot, type);
1346 /* Liveness for these slots is handled by process_spill_slots () */
1348 if (cfg->verbose_level > 1) {
1349 if (type == SLOT_REF)
1350 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
1356 process_variables (MonoCompile *cfg)
1358 MonoCompileGC *gcfg = cfg->gc_info;
1359 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1360 int i, locals_min_slot, locals_max_slot, cindex;
1364 int locals_min_offset = gcfg->locals_min_offset;
1365 int locals_max_offset = gcfg->locals_max_offset;
1367 /* Slots for locals are NOREF by default */
1368 locals_min_slot = (locals_min_offset - gcfg->min_offset) / sizeof (mgreg_t);
1369 locals_max_slot = (locals_max_offset - gcfg->min_offset) / sizeof (mgreg_t);
1370 for (i = locals_min_slot; i < locals_max_slot; ++i) {
1371 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1375 * Compute the offset where variables are initialized in the first bblock, if any.
1377 pc_offsets = g_new0 (int, cfg->next_vreg);
1379 bb = cfg->bb_entry->next_bb;
1380 MONO_BB_FOR_EACH_INS (bb, tmp) {
1381 if (tmp->opcode == OP_GC_LIVENESS_DEF) {
1382 int vreg = tmp->inst_c1;
1383 if (pc_offsets [vreg] == 0) {
1384 g_assert (tmp->backend.pc_offset > 0);
1385 pc_offsets [vreg] = tmp->backend.pc_offset;
1391 * Stack slots holding arguments are initialized in the prolog.
1392 * This means we can treat them alive for the whole method.
1394 for (i = 0; i < cfg->num_varinfo; i++) {
1395 MonoInst *ins = cfg->varinfo [i];
1396 MonoType *t = ins->inst_vtype;
1399 gboolean byref, is_this = FALSE;
1400 gboolean is_arg = i < cfg->locals_start;
1402 if (ins == cfg->ret)
1405 vmv = MONO_VARINFO (cfg, i);
1407 /* For some reason, 'this' is byref */
1408 if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
1409 t = &cfg->method->klass->byval_arg;
1415 if (ins->opcode == OP_REGVAR) {
1417 GCSlotType slot_type;
1419 t = mini_type_get_underlying_type (NULL, t);
1422 g_assert (hreg < MONO_MAX_IREGS);
1424 // FIXME: Add back this check
1426 if (is_arg && gcfg->reg_live_intervals [hreg]) {
1428 * FIXME: This argument shares a hreg with a local, we can't add the whole
1429 * method as a live interval, since it would overlap with the locals
1437 slot_type = SLOT_PIN;
1439 slot_type = MONO_TYPE_IS_REFERENCE (t) ? SLOT_REF : SLOT_NOREF;
1442 /* Live for the whole method */
1443 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1445 if (slot_type == SLOT_PIN) {
1446 /* These have no live interval, be conservative */
1447 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1450 * Unlike variables allocated to the stack, we generate liveness info
1451 * for noref vars in registers in mono_spill_global_vars (), because
1452 * knowing that a register doesn't contain a ref allows us to mark its save
1453 * locations precisely.
1455 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1456 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1457 set_reg_slot (gcfg, hreg, cindex, slot_type);
1461 if (cfg->verbose_level > 1) {
1462 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
1468 if (ins->opcode != OP_REGOFFSET)
1471 if (ins->inst_offset % sizeof (mgreg_t) != 0)
1474 if (is_arg && ins->inst_offset >= gcfg->max_offset)
1475 /* In parent frame */
1478 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1480 if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
1481 /* These do not get stored in the prolog */
1482 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1484 if (cfg->verbose_level > 1) {
1485 printf ("\tdead arg at fp%s0x%x (slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "+", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, pos, mono_type_full_name (ins->inst_vtype));
1490 if (MONO_TYPE_ISSTRUCT (t)) {
1492 gsize *bitmap = NULL;
1493 gboolean pin = FALSE;
1497 if (ins->backend.is_pinvoke)
1498 size = mono_class_native_size (ins->klass, NULL);
1500 size = mono_class_value_size (ins->klass, NULL);
1501 size_in_slots = ALIGN_TO (size, sizeof (mgreg_t)) / sizeof (mgreg_t);
1503 if (!ins->klass->has_references) {
1505 for (j = 0; j < size_in_slots; ++j)
1506 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1511 if (ins->klass->generic_container || mono_class_is_open_constructed_type (t)) {
1512 /* FIXME: Generic sharing */
1515 mono_class_compute_gc_descriptor (ins->klass);
1517 bitmap = mono_gc_get_bitmap_for_descr (ins->klass->gc_descr, &numbits);
1521 * Most vtypes are marked volatile because of the LDADDR instructions,
1522 * and they have no liveness information since they are decomposed
1523 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1524 * them during VZERO decomposition.
1526 if (!pc_offsets [vmv->vreg])
1530 if (ins->backend.is_pinvoke)
1534 if (pc_offsets [vmv->vreg]) {
1535 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1536 if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
1537 for (j = 0; j < numbits; ++j) {
1538 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1539 /* The descriptor is for the boxed object */
1540 set_slot (gcfg, (pos + j - (sizeof (MonoObject) / sizeof (gpointer))), cindex, pin ? SLOT_PIN : SLOT_REF);
1547 for (j = 0; j < size_in_slots; ++j) {
1548 set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
1554 if (cfg->verbose_level > 1) {
1555 printf ("\tvtype R%d at fp+0x%x-0x%x: %s\n", vmv->vreg, (int)ins->inst_offset, (int)(ins->inst_offset + (size / sizeof (mgreg_t))), mono_type_full_name (ins->inst_vtype));
1561 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1562 /* Vret addr etc. */
1567 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1569 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1570 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1571 set_slot (gcfg, pos, cindex, SLOT_PIN);
1573 if (cfg->verbose_level > 1)
1574 printf ("\tbyref at %s0x%x(fp) (R%d, slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, vmv->vreg, pos, mono_type_full_name (ins->inst_vtype));
1579 * This is currently disabled, but could be enabled to debug crashes.
1582 if (t->type == MONO_TYPE_I) {
1584 * Variables created in mono_handle_global_vregs have type I, but they
1585 * could hold GC refs since the vregs they were created from might not been
1586 * marked as holding a GC ref. So be conservative.
1588 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1593 t = mini_type_get_underlying_type (NULL, t);
1595 if (!MONO_TYPE_IS_REFERENCE (t)) {
1596 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1597 if (cfg->verbose_level > 1)
1598 printf ("\tnoref at %s0x%x(fp) (R%d, slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, vmv->vreg, pos, mono_type_full_name (ins->inst_vtype));
1602 /* 'this' is marked INDIRECT for gshared methods */
1603 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1605 * For volatile variables, treat them alive from the point they are
1606 * initialized in the first bblock until the end of the method.
1609 set_slot_everywhere (gcfg, pos, SLOT_REF);
1610 } else if (pc_offsets [vmv->vreg]) {
1611 set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
1612 set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
1614 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1616 if (cfg->verbose_level > 1)
1617 printf ("\tvolatile ref at %s0x%x(fp) (R%d, slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, vmv->vreg, pos, mono_type_full_name (ins->inst_vtype));
1622 /* Live for the whole method */
1623 set_slot_everywhere (gcfg, pos, SLOT_REF);
1625 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1626 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1627 set_slot (gcfg, pos, cindex, SLOT_REF);
1630 if (cfg->verbose_level > 1) {
1631 printf ("\tref at %s0x%x(fp) (R%d, slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)ins->inst_offset : (int)ins->inst_offset, vmv->vreg, pos, mono_type_full_name (ins->inst_vtype));
1635 g_free (pc_offsets);
1639 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1642 * Convert a sp relative offset to a slot index. This is
1643 * platform specific.
1646 /* fp = sp + offset */
1647 g_assert (cfg->frame_reg == AMD64_RBP);
1648 return (- cfg->arch.sp_fp_offset + sp_offset);
1656 type_to_gc_slot_type (MonoType *t)
1660 t = mini_type_get_underlying_type (NULL, t);
1661 if (MONO_TYPE_IS_REFERENCE (t))
1664 if (MONO_TYPE_ISSTRUCT (t)) {
1665 MonoClass *klass = mono_class_from_mono_type (t);
1666 if (!klass->has_references) {
1678 process_param_area_slots (MonoCompile *cfg)
1680 MonoCompileGC *gcfg = cfg->gc_info;
1685 * These slots are used for passing parameters during calls. They are sp relative, not
1686 * fp relative, so they are harder to handle.
1688 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1689 /* The distance between fp and sp is not constant */
1692 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1694 for (i = 0; i < gcfg->ncallsites; ++i) {
1695 GCCallSite *callsite = gcfg->callsites [i];
1698 for (l = callsite->param_slots; l; l = l->next) {
1699 MonoInst *def = l->data;
1700 int sp_offset = def->inst_offset;
1701 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1702 int slot = fp_offset_to_slot (cfg, fp_offset);
1704 g_assert (slot >= 0 && slot < gcfg->nslots);
1705 is_param [slot] = TRUE;
1709 /* All param area slots are noref by default */
1710 for (i = 0; i < gcfg->nslots; ++i) {
1712 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1715 for (i = 0; i < gcfg->ncallsites; ++i) {
1716 GCCallSite *callsite = gcfg->callsites [i];
1719 for (l = callsite->param_slots; l; l = l->next) {
1720 MonoInst *def = l->data;
1721 MonoType *t = def->inst_vtype;
1722 int sp_offset = def->inst_offset;
1723 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1724 int slot = fp_offset_to_slot (cfg, fp_offset);
1725 GCSlotType type = type_to_gc_slot_type (t);
1727 /* The slot is live between the def instruction and the call */
1728 set_slot_in_range (gcfg, slot, def->backend.pc_offset, callsite->pc_offset + 1, type);
1729 if (cfg->verbose_level > 1)
1730 printf ("\t%s param area slot at %s0x%x(fp)=0x%x(sp) (slot = %d) [0x%x-0x%x]\n", slot_type_to_string (type), fp_offset >= 0 ? "+" : "-", ABS (fp_offset), sp_offset, slot, def->backend.pc_offset, callsite->pc_offset + 1);
1736 compute_frame_size (MonoCompile *cfg)
1738 int i, locals_min_offset, locals_max_offset, cfa_min_offset, cfa_max_offset;
1739 int min_offset, max_offset;
1740 MonoCompileGC *gcfg = cfg->gc_info;
1741 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1744 /* Compute min/max offsets from the fp */
1748 locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, sizeof (mgreg_t));
1749 locals_max_offset = cfg->locals_max_stack_offset;
1751 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
1755 locals_min_offset = ALIGN_TO (locals_min_offset, sizeof (mgreg_t));
1756 locals_max_offset = ALIGN_TO (locals_max_offset, sizeof (mgreg_t));
1758 min_offset = locals_min_offset;
1759 max_offset = locals_max_offset;
1762 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1763 MonoInst *ins = cfg->args [i];
1765 if (ins->opcode == OP_REGOFFSET)
1766 min_offset = MIN (min_offset, ins->inst_offset);
1770 g_assert (cfg->frame_reg == cfg->cfa_reg);
1771 g_assert (cfg->cfa_offset > 0);
1773 cfa_max_offset = cfg->cfa_offset;
1775 min_offset = MIN (min_offset, cfa_min_offset);
1776 max_offset = MAX (max_offset, cfa_max_offset);
1778 /* Fp relative slots */
1779 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1780 gint data = GPOINTER_TO_INT (l->data);
1781 int offset = data >> 16;
1783 min_offset = MIN (min_offset, offset);
1787 if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
1788 int stack_offset = ALIGN_TO (cfg->stack_offset, sizeof (mgreg_t));
1789 min_offset = MIN (min_offset, (-stack_offset));
1792 /* Param area slots */
1794 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
1797 gcfg->min_offset = min_offset;
1798 gcfg->max_offset = max_offset;
1799 gcfg->locals_min_offset = locals_min_offset;
1800 gcfg->locals_max_offset = locals_max_offset;
1804 init_gcfg (MonoCompile *cfg)
1806 int i, nregs, nslots;
1807 MonoCompileGC *gcfg = cfg->gc_info;
1808 GCCallSite **callsites;
1817 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1818 ncallsites += g_slist_length (bb->gc_callsites);
1820 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
1822 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1823 for (l = bb->gc_callsites; l; l = l->next)
1824 callsites [i++] = l->data;
1827 /* The callsites should already be ordered by pc offset */
1828 for (i = 1; i < ncallsites; ++i)
1829 g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
1832 * The stack frame looks like this:
1834 * <fp + max_offset> == cfa -> <end of previous frame>
1835 * <other stack slots>
1837 * <other stack slots>
1838 * fp + min_offset ->
1843 if (cfg->verbose_level > 1)
1844 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
1846 nslots = (gcfg->max_offset - gcfg->min_offset) / sizeof (mgreg_t);
1849 gcfg->nslots = nslots;
1850 gcfg->nregs = nregs;
1851 gcfg->callsites = callsites;
1852 gcfg->ncallsites = ncallsites;
1853 gcfg->stack_bitmap_width = ALIGN_TO (nslots, 8) / 8;
1854 gcfg->reg_bitmap_width = ALIGN_TO (nregs, 8) / 8;
1855 gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
1856 gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
1857 gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
1858 gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
1860 /* All slots start out as PIN */
1861 memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * ncallsites);
1862 for (i = 0; i < nregs; ++i) {
1864 * By default, registers are NOREF.
1865 * It is possible for a callee to save them before being defined in this method,
1866 * but the saved value is dead too, so it doesn't need to be marked.
1868 if ((cfg->used_int_regs & (1 << i)))
1869 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
1875 create_map (MonoCompile *cfg)
1878 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
1880 int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
1881 int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
1883 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
1884 MonoCompileGC *gcfg = cfg->gc_info;
1885 GCCallSite **callsites;
1887 guint8 *bitmap, *bitmaps;
1888 guint32 reg_ref_mask, reg_pin_mask;
1889 gboolean has_finally;
1891 ncallsites = gcfg->ncallsites;
1892 nslots = gcfg->nslots;
1893 nregs = gcfg->nregs;
1894 callsites = gcfg->callsites;
1897 * FIXME: The calls to the finally clauses don't show up in the cfg. See
1898 * test_0_liveness_8 ().
1900 has_finally = FALSE;
1901 for (i = 0; i < cfg->header->num_clauses; ++i) {
1902 MonoExceptionClause *clause = &cfg->header->clauses [i];
1904 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1909 /* Treat every slot as pin everywhere */
1910 DEBUG (printf ("\tMethod has finally clauses, pessimizing live ranges.\n"));
1911 for (i = 0; i < nslots; ++i) {
1912 set_slot_everywhere (gcfg, i, SLOT_PIN);
1914 for (i = 0; i < nregs; ++i) {
1915 if (cfg->used_int_regs & (1 << i))
1916 set_reg_slot_everywhere (gcfg, i, SLOT_PIN);
1921 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
1922 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
1924 has_ref_slots = FALSE;
1925 has_pin_slots = FALSE;
1928 memset (ntypes, 0, sizeof (ntypes));
1929 for (i = 0; i < nslots; ++i) {
1930 gboolean has_ref = FALSE;
1931 gboolean has_pin = FALSE;
1933 for (j = 0; j < ncallsites; ++j) {
1934 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
1936 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
1941 has_ref_slots = TRUE;
1943 has_pin_slots = TRUE;
1946 ntypes [SLOT_REF] ++;
1948 ntypes [SLOT_PIN] ++;
1950 ntypes [SLOT_NOREF] ++;
1952 if (has_ref || has_pin) {
1959 start = end = nslots;
1961 g_assert (start != -1);
1962 g_assert (start < end);
1965 has_ref_regs = FALSE;
1966 has_pin_regs = FALSE;
1971 for (i = 0; i < nregs; ++i) {
1972 gboolean has_ref = FALSE;
1973 gboolean has_pin = FALSE;
1975 for (j = 0; j < ncallsites; ++j) {
1976 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i)) {
1981 for (j = 0; j < ncallsites; ++j) {
1982 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i)) {
1989 reg_ref_mask |= (1 << i);
1990 has_ref_regs = TRUE;
1994 reg_pin_mask |= (1 << i);
1995 has_pin_regs = TRUE;
2000 if (cfg->verbose_level > 1)
2001 printf ("Slots: %d Start: %d End: %d Refs: %d NoRefs: %d Pin: %d Callsites: %d\n", nslots, start, end, ntypes [SLOT_REF], ntypes [SLOT_NOREF], ntypes [SLOT_PIN], ncallsites);
2003 /* Create the GC Map */
2005 stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
2006 stack_bitmap_size = stack_bitmap_width * ncallsites;
2007 reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
2008 reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
2009 reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
2010 reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
2011 bitmaps_size = (has_ref_slots ? stack_bitmap_size : 0) + (has_pin_slots ? stack_bitmap_size : 0) + (has_ref_regs ? reg_ref_bitmap_size : 0) + (has_pin_regs ? reg_pin_bitmap_size : 0);
2013 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2015 map->frame_reg = cfg->frame_reg;
2016 map->start_offset = gcfg->min_offset;
2017 map->end_offset = gcfg->min_offset + (nslots * sizeof (mgreg_t));
2018 map->map_offset = start * sizeof (mgreg_t);
2019 map->nslots = end - start;
2020 map->has_ref_slots = has_ref_slots;
2021 map->has_pin_slots = has_pin_slots;
2022 map->has_ref_regs = has_ref_regs;
2023 map->has_pin_regs = has_pin_regs;
2024 g_assert (nregs < 32);
2025 map->used_int_regs = cfg->used_int_regs;
2026 map->reg_ref_mask = reg_ref_mask;
2027 map->reg_pin_mask = reg_pin_mask;
2028 map->nref_regs = nref_regs;
2029 map->npin_regs = npin_regs;
2031 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2034 if (has_ref_slots) {
2035 map->stack_ref_bitmap_offset = bitmaps_offset;
2036 bitmaps_offset += stack_bitmap_size;
2038 bitmap = &bitmaps [map->stack_ref_bitmap_offset];
2039 for (i = 0; i < nslots; ++i) {
2040 for (j = 0; j < ncallsites; ++j) {
2041 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
2042 set_bit (bitmap, stack_bitmap_width, j, i - start);
2046 if (has_pin_slots) {
2047 map->stack_pin_bitmap_offset = bitmaps_offset;
2048 bitmaps_offset += stack_bitmap_size;
2050 bitmap = &bitmaps [map->stack_pin_bitmap_offset];
2051 for (i = 0; i < nslots; ++i) {
2052 for (j = 0; j < ncallsites; ++j) {
2053 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
2054 set_bit (bitmap, stack_bitmap_width, j, i - start);
2059 map->reg_ref_bitmap_offset = bitmaps_offset;
2060 bitmaps_offset += reg_ref_bitmap_size;
2062 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2064 for (i = 0; i < nregs; ++i) {
2065 if (reg_ref_mask & (1 << i)) {
2066 for (j = 0; j < ncallsites; ++j) {
2067 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i))
2068 set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
2075 map->reg_pin_bitmap_offset = bitmaps_offset;
2076 bitmaps_offset += reg_pin_bitmap_size;
2078 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2080 for (i = 0; i < nregs; ++i) {
2081 if (reg_pin_mask & (1 << i)) {
2082 for (j = 0; j < ncallsites; ++j) {
2083 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i))
2084 set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
2092 map->ncallsites = ncallsites;
2093 if (cfg->code_len < 256)
2094 map->callsite_entry_size = 1;
2095 else if (cfg->code_len < 65536)
2096 map->callsite_entry_size = 2;
2098 map->callsite_entry_size = 4;
2100 /* Encode the GC Map */
2108 encode_gc_map (map, buf, &endbuf);
2109 g_assert (endbuf - buf < 256);
2111 encoded_size = endbuf - buf;
2112 alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
2114 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2115 //emap->ref_slots = map->ref_slots;
2117 /* Encoded fixed fields */
2118 p = &emap->encoded [0];
2119 //emap->encoded_size = encoded_size;
2120 memcpy (p, buf, encoded_size);
2123 /* Callsite table */
2124 p = (guint8*)ALIGN_TO ((guint64)p, map->callsite_entry_size);
2125 if (map->callsite_entry_size == 1) {
2126 guint8 *offsets = p;
2127 for (i = 0; i < ncallsites; ++i)
2128 offsets [i] = callsites [i]->pc_offset;
2129 stats.gc_callsites8_size += ncallsites * sizeof (guint8);
2130 } else if (map->callsite_entry_size == 2) {
2131 guint16 *offsets = (guint16*)p;
2132 for (i = 0; i < ncallsites; ++i)
2133 offsets [i] = callsites [i]->pc_offset;
2134 stats.gc_callsites16_size += ncallsites * sizeof (guint16);
2136 guint32 *offsets = (guint32*)p;
2137 for (i = 0; i < ncallsites; ++i)
2138 offsets [i] = callsites [i]->pc_offset;
2139 stats.gc_callsites32_size += ncallsites * sizeof (guint32);
2141 p += ncallsites * map->callsite_entry_size;
2144 memcpy (p, bitmaps, bitmaps_size);
2146 stats.gc_maps_size += alloc_size;
2147 stats.gc_callsites_size += ncallsites * map->callsite_entry_size;
2148 stats.gc_bitmaps_size += bitmaps_size;
2149 stats.gc_map_struct_size += sizeof (GCEncodedMap) + encoded_size;
2151 cfg->jit_info->gc_info = emap;
2154 stats.all_slots += nslots;
2155 stats.ref_slots += ntypes [SLOT_REF];
2156 stats.noref_slots += ntypes [SLOT_NOREF];
2157 stats.pin_slots += ntypes [SLOT_PIN];
2161 mini_gc_create_gc_map (MonoCompile *cfg)
2163 if (!cfg->compute_gc_maps)
2167 * During marking, all frames except the top frame are at a call site, and we mark the
2168 * top frame conservatively. This means that we only need to compute and record
2169 * GC maps for call sites.
2172 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2173 /* Without liveness info, the live ranges are not precise enough */
2176 mono_analyze_liveness_gc (cfg);
2178 compute_frame_size (cfg);
2182 process_spill_slots (cfg);
2183 process_other_slots (cfg);
2184 process_param_area_slots (cfg);
2185 process_variables (cfg);
2195 memset (&cb, 0, sizeof (cb));
2196 cb.thread_attach_func = thread_attach_func;
2197 cb.thread_detach_func = thread_detach_func;
2198 cb.thread_suspend_func = thread_suspend_func;
2199 /* Comment this out to disable precise stack marking */
2200 cb.thread_mark_func = thread_mark_func;
2201 mono_gc_set_gc_callbacks (&cb);
2203 mono_counters_register ("GC Maps size",
2204 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
2205 mono_counters_register ("GC Call Sites size",
2206 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
2207 mono_counters_register ("GC Bitmaps size",
2208 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
2209 mono_counters_register ("GC Map struct size",
2210 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
2211 mono_counters_register ("GC Call Sites encoded using 8 bits",
2212 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
2213 mono_counters_register ("GC Call Sites encoded using 16 bits",
2214 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
2215 mono_counters_register ("GC Call Sites encoded using 32 bits",
2216 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
2218 mono_counters_register ("GC Map slots (all)",
2219 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
2220 mono_counters_register ("GC Map slots (ref)",
2221 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.ref_slots);
2222 mono_counters_register ("GC Map slots (noref)",
2223 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.noref_slots);
2224 mono_counters_register ("GC Map slots (pin)",
2225 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
2227 mono_counters_register ("GC TLS Data size",
2228 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2230 mono_counters_register ("Stack space scanned (all)",
2231 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
2232 mono_counters_register ("Stack space scanned (using GC Maps)",
2233 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
2234 mono_counters_register ("Stack space scanned (precise)",
2235 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_precisely);
2236 mono_counters_register ("Stack space scanned (pin)",
2237 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_conservatively);
2238 mono_counters_register ("Stack space scanned (pin registers)",
2239 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_registers);
2250 mini_gc_init_gc_map (MonoCompile *cfg)
2255 mini_gc_create_gc_map (MonoCompile *cfg)
2260 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2265 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2274 * Set GC specific options in CFG.
2277 mini_gc_init_cfg (MonoCompile *cfg)
2279 if (mono_gc_is_moving ()) {
2280 cfg->disable_ref_noref_stack_slot_share = TRUE;
2281 cfg->gen_write_barriers = TRUE;
2284 mini_gc_init_gc_map (cfg);
2288 * Problems with the current code:
2289 * - the stack walk is slow
2290 * - vtypes/refs used in EH regions are treated conservatively
2291 * - if the code is finished, less pinning will be done, causing problems because
2292 * we promote all surviving objects to old-gen.
2296 * Ideas for creating smaller GC maps:
2297 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2299 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2301 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2302 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2303 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2304 * from an index entry.