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/sgen-gc.h>
18 #include <mono/metadata/gc-internal.h>
19 #include <mono/utils/mono-counters.h>
21 /* Contains state needed by the GC Map construction code */
24 * This contains information about stack slots initialized in the prolog, encoded using
25 * (slot_index << 16) | slot_type. The slot_index is relative to the CFA, i.e. 0
26 * means cfa+0, 1 means cfa-4/8, etc.
28 GSList *stack_slots_from_cfa;
29 /* Same for stack slots relative to the frame pointer */
30 GSList *stack_slots_from_fp;
32 /* Number of slots in the map */
34 /* The number of registers in the map */
36 /* Min and Max offsets of the stack frame relative to fp */
37 int min_offset, max_offset;
38 /* Same for the locals area */
39 int locals_min_offset, locals_max_offset;
41 /* The call sites where this frame can be stopped during GC */
42 GCCallSite **callsites;
43 /* The number of call sites */
47 * The width of the stack bitmaps in bytes. This is not equal to the bitmap width at
48 * runtime, since it includes columns which are 0.
50 int stack_bitmap_width;
52 * A bitmap whose width equals nslots, and whose height equals ncallsites.
53 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
56 guint8 *stack_ref_bitmap;
57 /* Same for SLOT_PIN */
58 guint8 *stack_pin_bitmap;
61 * Similar bitmaps for registers. These have width MONO_MAX_IREGS in bits.
64 guint8 *reg_ref_bitmap;
65 guint8 *reg_pin_bitmap;
68 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
71 /* We don't support debug levels, its all-or-nothing */
72 #define DEBUG(s) do { s; fflush (logfile); } while (0)
73 #define DEBUG_ENABLED 1
80 #define DEBUG_PRECISE(s) do { s; } while (0)
81 #define DEBUG_PRECISE_ENABLED
83 #define DEBUG_PRECISE(s)
87 #define DEBUG_GC_MAP(s) do { s; fflush (stdout); } while (0)
89 #define DEBUG_GC_MAP(s)
93 * Contains information collected during the conservative stack marking pass,
94 * used during the precise pass. This helps to avoid doing a stack walk twice, which
100 int frame_start_offset;
102 /* Relative to stack_start */
103 int reg_locations [MONO_MAX_IREGS];
104 #ifdef DEBUG_PRECISE_ENABLED
107 int regs [MONO_MAX_IREGS];
111 /* Max number of frames stored in the TLS data */
112 #define MAX_FRAMES 50
115 * Per-thread data kept by this module. This is stored in the GC and passed to us as
116 * parameters, instead of being stored in a TLS variable, since during a collection,
117 * only the collection thread is active.
122 gboolean has_context;
123 MonoJitTlsData *jit_tls;
126 gpointer ref_to_track;
127 /* Number of frames collected during the !precise pass */
129 FrameInfo frames [MAX_FRAMES];
132 /* These are constant so don't store them in the GC Maps */
133 /* Number of registers stored in gc maps */
134 #define NREGS MONO_MAX_IREGS
138 * Contains information needed to mark a stack frame.
139 * This is a transient structure, created from a compressed representation on-demand.
143 * The offsets of the GC tracked area inside the stack frame relative to the frame pointer.
144 * This includes memory which is NOREF thus doesn't need GC maps.
149 * The offset relative to frame_offset where the the memory described by the GC maps
153 /* The number of stack slots in the map */
155 /* The frame pointer register */
157 /* The size of each callsite table entry */
158 guint8 callsite_entry_size;
159 guint has_pin_slots : 1;
160 guint has_ref_slots : 1;
161 guint has_ref_regs : 1;
162 guint has_pin_regs : 1;
164 /* The offsets below are into an external bitmaps array */
167 * A bitmap whose width is equal to bitmap_width, and whose height is equal to ncallsites.
168 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
171 guint32 stack_ref_bitmap_offset;
173 * Same for SLOT_PIN. It is possible that the same bit is set in both bitmaps at
174 * different callsites, if the slot starts out as PIN, and later changes to REF.
176 guint32 stack_pin_bitmap_offset;
179 * Corresponding bitmaps for registers
180 * These have width equal to the number of bits set in reg_ref_mask/reg_pin_mask.
181 * FIXME: Merge these with the normal bitmaps, i.e. reserve the first x slots for them ?
183 guint32 reg_pin_bitmap_offset;
184 guint32 reg_ref_bitmap_offset;
186 guint32 used_int_regs, reg_ref_mask, reg_pin_mask;
188 /* The number of bits set in the two masks above */
189 guint8 nref_regs, npin_regs;
192 * A bit array marking slots which contain refs.
193 * This is used only for debugging.
197 /* Callsite offsets */
198 /* These can take up a lot of space, so encode them compactly */
208 * A compressed version of GCMap. This is what gets stored in MonoJitInfo.
212 //guint8 encoded_size;
215 * The arrays below are embedded after the struct.
216 * Their address needs to be computed.
219 /* The fixed fields of the GCMap encoded using LEB128 */
220 guint8 encoded [MONO_ZERO_LEN_ARRAY];
222 /* An array of ncallsites entries, each entry is callsite_entry_size bytes long */
223 guint8 callsites [MONO_ZERO_LEN_ARRAY];
226 guint8 bitmaps [MONO_ZERO_LEN_ARRAY];
229 static int precise_frame_count [2], precise_frame_limit = -1;
230 static gboolean precise_frame_limit_inited;
236 int scanned_precisely;
237 int scanned_conservatively;
238 int scanned_registers;
248 int gc_callsites_size;
249 int gc_callsites8_size;
250 int gc_callsites16_size;
251 int gc_callsites32_size;
253 int gc_map_struct_size;
257 static JITGCStats stats;
259 static FILE *logfile;
261 // FIXME: Move these to a shared place
264 encode_uleb128 (guint32 value, guint8 *buf, guint8 **endbuf)
269 guint8 b = value & 0x7f;
271 if (value != 0) /* more bytes to come */
279 static G_GNUC_UNUSED void
280 encode_sleb128 (gint32 value, guint8 *buf, guint8 **endbuf)
283 gboolean negative = (value < 0);
291 /* the following is unnecessary if the
292 * implementation of >>= uses an arithmetic rather
293 * than logical shift for a signed left operand
297 value |= - (1 <<(size - 7));
298 /* sign bit of byte is second high order bit (0x40) */
299 if ((value == 0 && !(byte & 0x40)) ||
300 (value == -1 && (byte & 0x40)))
310 static inline guint32
311 decode_uleb128 (guint8 *buf, guint8 **endbuf)
321 res = res | (((int)(b & 0x7f)) << shift);
333 decode_sleb128 (guint8 *buf, guint8 **endbuf)
343 res = res | (((int)(b & 0x7f)) << shift);
346 if (shift < 32 && (b & 0x40))
347 res |= - (1 << shift);
358 encode_frame_reg (int frame_reg)
361 if (frame_reg == AMD64_RSP)
363 else if (frame_reg == AMD64_RBP)
368 g_assert_not_reached ();
373 decode_frame_reg (int encoded)
378 else if (encoded == 1)
383 g_assert_not_reached ();
389 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15, AMD64_RDI, AMD64_RSI };
391 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15 };
396 encode_regmask (guint32 regmask)
403 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
404 if (regmask & (1 << callee_saved_regs [i])) {
406 regmask -= (1 << callee_saved_regs [i]);
409 g_assert (regmask == 0);
418 decode_regmask (guint32 regmask)
425 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i)
426 if (regmask & (1 << i))
427 res |= (1 << callee_saved_regs [i]);
438 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
441 encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
445 encode_sleb128 (map->start_offset / sizeof (mgreg_t), buf, &buf);
446 encode_sleb128 (map->end_offset / sizeof (mgreg_t), buf, &buf);
447 encode_sleb128 (map->map_offset / sizeof (mgreg_t), buf, &buf);
448 encode_uleb128 (map->nslots, buf, &buf);
449 g_assert (map->callsite_entry_size <= 4);
450 freg = encode_frame_reg (map->frame_reg);
452 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);
453 encode_uleb128 (flags, buf, &buf);
454 encode_uleb128 (encode_regmask (map->used_int_regs), buf, &buf);
455 if (map->has_ref_regs)
456 encode_uleb128 (encode_regmask (map->reg_ref_mask), buf, &buf);
457 if (map->has_pin_regs)
458 encode_uleb128 (encode_regmask (map->reg_pin_mask), buf, &buf);
459 encode_uleb128 (map->ncallsites, buf, &buf);
467 * Decode the encoded GC map representation in BUF and store the result into MAP.
470 decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
473 int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
476 map->start_offset = decode_sleb128 (buf, &buf) * sizeof (mgreg_t);
477 map->end_offset = decode_sleb128 (buf, &buf) * sizeof (mgreg_t);
478 map->map_offset = decode_sleb128 (buf, &buf) * sizeof (mgreg_t);
479 map->nslots = decode_uleb128 (buf, &buf);
480 flags = decode_uleb128 (buf, &buf);
481 map->has_ref_slots = (flags & 1) ? 1 : 0;
482 map->has_pin_slots = (flags & 2) ? 1 : 0;
483 map->has_ref_regs = (flags & 4) ? 1 : 0;
484 map->has_pin_regs = (flags & 8) ? 1 : 0;
485 map->callsite_entry_size = ((flags >> 4) & 0x3) + 1;
487 map->frame_reg = decode_frame_reg (freg);
488 map->used_int_regs = decode_regmask (decode_uleb128 (buf, &buf));
489 if (map->has_ref_regs) {
490 map->reg_ref_mask = decode_regmask (decode_uleb128 (buf, &buf));
492 for (i = 0; i < NREGS; ++i)
493 if (map->reg_ref_mask & (1 << i))
497 if (map->has_pin_regs) {
498 map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
500 for (i = 0; i < NREGS; ++i)
501 if (map->reg_pin_mask & (1 << i))
505 map->ncallsites = decode_uleb128 (buf, &buf);
507 stack_bitmap_size = (ALIGN_TO (map->nslots, 8) / 8) * map->ncallsites;
508 reg_ref_bitmap_size = (ALIGN_TO (map->nref_regs, 8) / 8) * map->ncallsites;
509 reg_pin_bitmap_size = (ALIGN_TO (map->npin_regs, 8) / 8) * map->ncallsites;
511 map->stack_ref_bitmap_offset = offset;
512 if (map->has_ref_slots)
513 offset += stack_bitmap_size;
514 map->stack_pin_bitmap_offset = offset;
515 if (map->has_pin_slots)
516 offset += stack_bitmap_size;
517 map->reg_ref_bitmap_offset = offset;
518 if (map->has_ref_regs)
519 offset += reg_ref_bitmap_size;
520 map->reg_pin_bitmap_offset = offset;
521 if (map->has_pin_regs)
522 offset += reg_pin_bitmap_size;
528 thread_attach_func (void)
532 tls = g_new0 (TlsData, 1);
533 tls->tid = GetCurrentThreadId ();
534 stats.tlsdata_size += sizeof (TlsData);
540 thread_detach_func (gpointer user_data)
542 TlsData *tls = user_data;
548 thread_suspend_func (gpointer user_data, void *sigctx)
550 TlsData *tls = user_data;
553 /* Happens during startup */
556 tls->lmf = mono_get_lmf ();
558 mono_arch_sigctx_to_monoctx (sigctx, &tls->ctx);
559 tls->has_context = TRUE;
561 tls->has_context = FALSE;
563 tls->jit_tls = TlsGetValue (mono_jit_tls_id);
566 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
569 set_bit (guint8 *bitmap, int width, int y, int x)
571 bitmap [(width * y) + (x / 8)] |= (1 << (x % 8));
575 clear_bit (guint8 *bitmap, int width, int y, int x)
577 bitmap [(width * y) + (x / 8)] &= ~(1 << (x % 8));
581 get_bit (guint8 *bitmap, int width, int y, int x)
583 return bitmap [(width * y) + (x / 8)] & (1 << (x % 8));
587 slot_type_to_string (GCSlotType type)
597 g_assert_not_reached ();
603 * conservatively_pass:
605 * Mark a thread stack conservatively and collect information needed by the precise pass.
608 conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
611 MonoContext ctx, new_ctx;
614 gboolean last = TRUE;
618 guint8* fp, *p, *real_frame_start, *frame_start, *frame_end;
619 int i, pc_offset, cindex, bitmap_width;
620 int scanned = 0, scanned_precisely, scanned_conservatively, scanned_registers;
622 StackFrameInfo frame;
623 mgreg_t *reg_locations [MONO_MAX_IREGS];
624 mgreg_t *new_reg_locations [MONO_MAX_IREGS];
627 guint32 precise_regmask;
631 tls->ref_to_track = NULL;
634 /* tls == NULL can happen during startup */
635 if (mono_thread_internal_current () == NULL || !tls) {
636 mono_gc_conservatively_scan_area (stack_start, stack_end);
637 stats.scanned_stacks += stack_end - stack_start;
644 /* Number of bytes scanned based on GC map data */
646 /* Number of bytes scanned precisely based on GC map data */
647 scanned_precisely = 0;
648 /* Number of bytes scanned conservatively based on GC map data */
649 scanned_conservatively = 0;
650 /* Number of bytes scanned conservatively in register save areas */
651 scanned_registers = 0;
653 /* This is one past the last address which we have scanned */
654 stack_limit = stack_start;
656 if (!tls->has_context)
657 memset (&new_ctx, 0, sizeof (ctx));
659 memcpy (&new_ctx, &tls->ctx, sizeof (MonoContext));
661 memset (reg_locations, 0, sizeof (reg_locations));
662 memset (new_reg_locations, 0, sizeof (new_reg_locations));
665 memcpy (&ctx, &new_ctx, sizeof (ctx));
667 for (i = 0; i < MONO_MAX_IREGS; ++i) {
668 if (new_reg_locations [i]) {
670 * If the current frame saves the register, it means it might modify its
671 * value, thus the old location might not contain the same value, so
672 * we have to mark it conservatively.
674 if (reg_locations [i]) {
675 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
676 mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
677 scanned_registers += sizeof (mgreg_t);
680 reg_locations [i] = new_reg_locations [i];
682 DEBUG (fprintf (logfile, "\treg %s is now at location %p.\n", mono_arch_regname (i), reg_locations [i]));
686 g_assert ((guint64)stack_limit % sizeof (mgreg_t) == 0);
688 #ifdef MONO_ARCH_HAVE_FIND_JIT_INFO_EXT
689 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);
696 /* The last frame can be in any state so mark conservatively */
703 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
705 if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
706 /* These frames are unwound through an LMF, and we have no precise register tracking for those */
707 DEBUG (fprintf (logfile, "Mark(0): <Managed-to-native transition>\n"));
708 for (i = 0; i < MONO_MAX_IREGS; ++i) {
709 if (reg_locations [i]) {
710 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
711 mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
712 scanned_registers += sizeof (mgreg_t);
714 reg_locations [i] = NULL;
715 new_reg_locations [i] = NULL;
720 /* These frames are very problematic */
721 if (ji->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
722 DEBUG (char *fname = mono_method_full_name (ji->method, TRUE); fprintf (logfile, "Mark(0): %s+0x%x (%p)\n", fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx)); g_free (fname));
723 DEBUG (fprintf (logfile, "\tSkip.\n"));
727 /* All the other frames are at a call site */
729 if (tls->nframes == MAX_FRAMES) {
731 * Can't save information since the array is full. So scan the rest of the
732 * stack conservatively.
734 DEBUG (fprintf (logfile, "Mark (0): Frame stack full.\n"));
738 /* Scan the frame of this method */
741 * A frame contains the following:
746 * - localloc-ed memory
748 g_assert (pc_offset >= 0);
753 DEBUG (char *fname = mono_method_full_name (ji->method, TRUE); fprintf (logfile, "Mark(0): %s+0x%x (%p)\n", fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx)); g_free (fname));
754 DEBUG (fprintf (logfile, "\tNo GC Map.\n"));
759 * Debugging aid to control the number of frames scanned precisely
761 if (!precise_frame_limit_inited) {
762 if (getenv ("MONO_PRECISE_COUNT"))
763 precise_frame_limit = atoi (getenv ("MONO_PRECISE_COUNT"));
764 precise_frame_limit_inited = TRUE;
767 if (precise_frame_limit != -1) {
768 if (precise_frame_count [FALSE] == precise_frame_limit)
769 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (ji->method, TRUE));
770 if (precise_frame_count [FALSE] > precise_frame_limit)
773 precise_frame_count [FALSE] ++;
775 /* Decode the encoded GC map */
777 memset (map, 0, sizeof (GCMap));
778 decode_gc_map (&emap->encoded [0], map, &p);
779 p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
780 map->callsites.offsets8 = p;
781 p += map->callsite_entry_size * map->ncallsites;
785 if (map->frame_reg == AMD64_RSP)
786 fp = (guint8*)ctx.rsp;
787 else if (map->frame_reg == AMD64_RBP)
788 fp = (guint8*)ctx.rbp;
790 g_assert_not_reached ();
793 g_assert_not_reached ();
796 real_frame_start = fp + map->start_offset;
797 frame_start = fp + map->start_offset + map->map_offset;
798 frame_end = fp + map->end_offset;
800 DEBUG (char *fname = mono_method_full_name (ji->method, TRUE); fprintf (logfile, "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));
802 /* Find the callsite index */
803 if (ji->code_size < 256) {
804 for (i = 0; i < map->ncallsites; ++i)
805 /* ip points inside the call instruction */
806 if (map->callsites.offsets8 [i] == pc_offset + 1)
808 } else if (ji->code_size < 65536) {
809 // FIXME: Use a binary search
810 for (i = 0; i < map->ncallsites; ++i)
811 /* ip points inside the call instruction */
812 if (map->callsites.offsets16 [i] == pc_offset + 1)
815 // FIXME: Use a binary search
816 for (i = 0; i < map->ncallsites; ++i)
817 /* ip points inside the call instruction */
818 if (map->callsites.offsets32 [i] == pc_offset + 1)
821 if (i == map->ncallsites) {
822 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (ji->method, TRUE));
823 g_assert_not_reached ();
827 g_assert (real_frame_start >= stack_limit);
829 if (real_frame_start > stack_limit) {
830 /* This scans the previously skipped frames as well */
831 DEBUG (fprintf (logfile, "\tscan area %p-%p (%d).\n", stack_limit, real_frame_start, (int)(real_frame_start - stack_limit)));
832 mono_gc_conservatively_scan_area (stack_limit, real_frame_start);
833 stats.scanned_other += real_frame_start - stack_limit;
836 /* Mark stack slots */
837 if (map->has_pin_slots) {
838 int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
839 guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
844 for (i = 0; i < map->nslots; ++i) {
845 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
847 DEBUG (fprintf (logfile, "\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
848 mono_gc_conservatively_scan_area (p, p + sizeof (mgreg_t));
849 scanned_conservatively += sizeof (mgreg_t);
851 scanned_precisely += sizeof (mgreg_t);
853 p += sizeof (mgreg_t);
856 scanned_precisely += (map->nslots * sizeof (mgreg_t));
859 /* The area outside of start-end is NOREF */
860 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * sizeof (mgreg_t));
863 precise_regmask = map->used_int_regs | (1 << map->frame_reg);
864 if (map->has_pin_regs) {
865 int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
866 guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
868 for (i = 0; i < NREGS; ++i) {
869 if (!(map->used_int_regs & (1 << i)))
872 if (!(map->reg_pin_mask & (1 << i)))
875 if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
876 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
877 precise_regmask &= ~(1 << i);
883 scanned += map->end_offset - map->start_offset;
885 g_assert (scanned == scanned_precisely + scanned_conservatively);
887 stack_limit = frame_end;
889 /* Save information for the precise pass */
890 fi = &tls->frames [tls->nframes];
891 fi->nslots = map->nslots;
892 bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
893 if (map->has_ref_slots)
894 fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
897 fi->frame_start_offset = frame_start - stack_start;
898 fi->nreg_locations = 0;
899 DEBUG_PRECISE (fi->ji = ji);
900 DEBUG_PRECISE (fi->fp = fp);
902 if (map->has_ref_regs) {
903 int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
904 guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
906 for (i = 0; i < NREGS; ++i) {
907 if (!(map->reg_ref_mask & (1 << i)))
910 if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
911 DEBUG_PRECISE (fi->regs [fi->nreg_locations] = i);
912 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
913 fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
914 fi->nreg_locations ++;
921 * Clear locations of precisely stacked registers.
923 if (precise_regmask) {
924 for (i = 0; i < NREGS; ++i) {
925 if (precise_regmask & (1 << i)) {
927 * The method uses this register, and we have precise info for it.
928 * This means the location will be scanned precisely.
929 * Tell the code at the beginning of the loop that this location is
932 if (reg_locations [i])
933 DEBUG (fprintf (logfile, "\treg %s at location %p (==%p) is precise.\n", mono_arch_regname (i), reg_locations [i], (gpointer)*reg_locations [i]));
934 reg_locations [i] = NULL;
942 /* Scan the remaining register save locations */
943 for (i = 0; i < MONO_MAX_IREGS; ++i) {
944 if (reg_locations [i]) {
945 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", reg_locations [i]));
946 mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
947 scanned_registers += sizeof (mgreg_t);
949 if (new_reg_locations [i]) {
950 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", new_reg_locations [i]));
951 mono_gc_conservatively_scan_area (new_reg_locations [i], new_reg_locations [i] + sizeof (mgreg_t));
952 scanned_registers += sizeof (mgreg_t);
956 if (stack_limit < stack_end) {
957 DEBUG (fprintf (logfile, "\tscan remaining stack %p-%p (%d).\n", stack_limit, stack_end, (int)(stack_end - stack_limit)));
958 mono_gc_conservatively_scan_area (stack_limit, stack_end);
959 stats.scanned_native += stack_end - stack_limit;
962 DEBUG (fprintf (logfile, "Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
964 stats.scanned_stacks += stack_end - stack_start;
965 stats.scanned += scanned;
966 stats.scanned_precisely += scanned_precisely;
967 stats.scanned_conservatively += scanned_conservatively;
968 stats.scanned_registers += scanned_registers;
970 //mono_gc_conservatively_scan_area (stack_start, stack_end);
976 * Mark a thread stack precisely based on information saved during the conservative
980 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
989 for (findex = 0; findex < tls->nframes; findex ++) {
990 /* Load information saved by the !precise pass */
991 fi = &tls->frames [findex];
992 frame_start = stack_start + fi->frame_start_offset;
994 DEBUG (char *fname = mono_method_full_name (fi->ji->method, TRUE); fprintf (logfile, "Mark(1): %s\n", fname); g_free (fname));
997 * FIXME: Add a function to mark using a bitmap, to avoid doing a
998 * call for each object.
1001 /* Mark stack slots */
1003 guint8 *ref_bitmap = fi->bitmap;
1006 for (i = 0; i < fi->nslots; ++i) {
1007 MonoObject **ptr = (MonoObject**)(frame_start + (i * sizeof (mgreg_t)));
1009 live = ref_bitmap [i / 8] & (1 << (i % 8));
1012 MonoObject *obj = *ptr;
1014 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1015 *ptr = mono_gc_scan_object (obj);
1016 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1018 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1023 * This is disabled because the pointer takes up a lot of space.
1024 * Stack slots might be shared between ref and non-ref variables ?
1026 if (map->ref_slots [i / 8] & (1 << (i % 8))) {
1027 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: dead (%p)\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, *ptr));
1029 * Fail fast if the live range is incorrect, and
1030 * the JITted code tries to access this object
1039 /* Mark registers */
1042 * Registers are different from stack slots, they have no address where they
1043 * are stored. Instead, some frame below this frame in the stack saves them
1044 * in its prolog to the stack. We can mark this location precisely.
1046 for (i = 0; i < fi->nreg_locations; ++i) {
1048 * reg_locations [i] contains the address of the stack slot where
1049 * a reg was last saved, so mark that slot.
1051 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1052 MonoObject *obj = *ptr;
1055 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
1056 *ptr = mono_gc_scan_object (obj);
1057 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1059 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p\n", mono_arch_regname (fi->regs [i]), ptr, obj));
1065 * Debugging aid to check for missed refs.
1067 if (tls->ref_to_track) {
1070 for (p = (mgreg_t*)stack_start; p < (mgreg_t*)stack_end; ++p)
1071 if (*p == (mgreg_t)tls->ref_to_track)
1072 printf ("REF AT %p.\n", p);
1079 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1080 * call, and TRUE at the second. USER_DATA points to a TlsData
1081 * structure filled up by thread_suspend_func.
1084 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise)
1086 TlsData *tls = user_data;
1088 DEBUG (fprintf (logfile, "****************************************\n"));
1089 DEBUG (fprintf (logfile, "*** %s stack marking for thread %p (%p-%p) ***\n", precise ? "Precise" : "Conservative", tls ? GUINT_TO_POINTER (tls->tid) : NULL, stack_start, stack_end));
1090 DEBUG (fprintf (logfile, "****************************************\n"));
1093 conservative_pass (tls, stack_start, stack_end);
1095 precise_pass (tls, stack_start, stack_end);
1099 mini_gc_init_gc_map (MonoCompile *cfg)
1101 if (COMPILE_LLVM (cfg))
1105 /* Debugging support */
1107 static int precise_count;
1110 if (getenv ("MONO_GCMAP_COUNT")) {
1111 if (precise_count == atoi (getenv ("MONO_GCMAP_COUNT")))
1112 printf ("LAST: %s\n", mono_method_full_name (cfg->method, TRUE));
1113 if (precise_count > atoi (getenv ("MONO_GCMAP_COUNT")))
1119 cfg->compute_gc_maps = TRUE;
1121 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1125 * mini_gc_set_slot_type_from_fp:
1127 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1128 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1129 * need to call this function for those slots.
1132 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1134 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1136 if (!cfg->compute_gc_maps)
1139 g_assert (slot_offset % sizeof (mgreg_t) == 0);
1141 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1145 * mini_gc_set_slot_type_from_cfa:
1147 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1148 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1149 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1150 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1151 * call this function for those slots.
1154 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1156 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1157 int slot = - (slot_offset / sizeof (mgreg_t));
1159 if (!cfg->compute_gc_maps)
1162 g_assert (slot_offset <= 0);
1163 g_assert (slot_offset % sizeof (mgreg_t) == 0);
1165 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1169 fp_offset_to_slot (MonoCompile *cfg, int offset)
1171 MonoCompileGC *gcfg = cfg->gc_info;
1173 return (offset - gcfg->min_offset) / sizeof (mgreg_t);
1177 slot_to_fp_offset (MonoCompile *cfg, int slot)
1179 MonoCompileGC *gcfg = cfg->gc_info;
1181 return (slot * sizeof (mgreg_t)) + gcfg->min_offset;
1185 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1187 g_assert (slot >= 0 && slot < gcfg->nslots);
1189 if (type == SLOT_PIN) {
1190 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1191 set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1192 } else if (type == SLOT_REF) {
1193 set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1194 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1195 } else if (type == SLOT_NOREF) {
1196 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1197 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1202 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1206 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1207 set_slot (gcfg, slot, cindex, type);
1211 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1215 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1216 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1217 if (callsite_offset >= from && callsite_offset < to)
1218 set_slot (gcfg, slot, cindex, type);
1223 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1225 g_assert (slot >= 0 && slot < gcfg->nregs);
1227 if (type == SLOT_PIN) {
1228 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1229 set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1230 } else if (type == SLOT_REF) {
1231 set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1232 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1233 } else if (type == SLOT_NOREF) {
1234 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1235 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1240 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1244 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1245 set_reg_slot (gcfg, slot, cindex, type);
1249 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1253 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1254 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1255 if (callsite_offset >= from && callsite_offset < to)
1256 set_reg_slot (gcfg, slot, cindex, type);
1261 process_spill_slots (MonoCompile *cfg)
1263 MonoCompileGC *gcfg = cfg->gc_info;
1268 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1269 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1270 for (l = bb->spill_slot_defs; l; l = l->next) {
1271 MonoInst *def = l->data;
1272 int spill_slot = def->inst_c0;
1273 int bank = def->inst_c1;
1274 int offset = cfg->spill_info [bank][spill_slot].offset;
1275 int slot = fp_offset_to_slot (cfg, offset);
1277 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1278 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1282 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1283 for (l = bb->spill_slot_defs; l; l = l->next) {
1284 MonoInst *def = l->data;
1285 int spill_slot = def->inst_c0;
1286 int bank = def->inst_c1;
1287 int offset = cfg->spill_info [bank][spill_slot].offset;
1288 int slot = fp_offset_to_slot (cfg, offset);
1291 if (bank == MONO_REG_INT_MP)
1297 * Extend the live interval for the GC tracked spill slots
1298 * defined in this bblock.
1299 * FIXME: This is not needed.
1301 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1303 if (cfg->verbose_level > 1)
1304 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
1308 /* Set fp spill slots to NOREF */
1309 for (i = 0; i < cfg->spill_info_len [MONO_REG_DOUBLE]; ++i) {
1310 int offset = cfg->spill_info [MONO_REG_DOUBLE][i].offset;
1316 slot = fp_offset_to_slot (cfg, offset);
1318 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1320 if (cfg->verbose_level > 1)
1321 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1324 /* Set int spill slots to NOREF */
1325 for (i = 0; i < cfg->spill_info_len [MONO_REG_INT]; ++i) {
1326 int offset = cfg->spill_info [MONO_REG_INT][i].offset;
1332 slot = fp_offset_to_slot (cfg, offset);
1334 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1335 if (cfg->verbose_level > 1)
1336 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1341 * process_other_slots:
1343 * Process stack slots registered using mini_gc_set_slot_type_... ().
1346 process_other_slots (MonoCompile *cfg)
1348 MonoCompileGC *gcfg = cfg->gc_info;
1351 /* Relative to the CFA */
1352 for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
1353 guint data = GPOINTER_TO_UINT (l->data);
1354 int cfa_slot = data >> 16;
1355 GCSlotType type = data & 0xff;
1359 * Map the cfa relative slot to an fp relative slot.
1360 * slot_addr == cfa - <cfa_slot>*4/8
1361 * fp + cfa_offset == cfa
1362 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1364 slot = (cfg->cfa_offset / sizeof (mgreg_t)) - cfa_slot - (gcfg->min_offset / sizeof (mgreg_t));
1366 set_slot_everywhere (gcfg, slot, type);
1368 if (cfg->verbose_level > 1) {
1369 int fp_offset = slot_to_fp_offset (cfg, slot);
1370 if (type == SLOT_NOREF)
1371 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)));
1375 /* Relative to the FP */
1376 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1377 gint data = GPOINTER_TO_INT (l->data);
1378 int offset = data >> 16;
1379 GCSlotType type = data & 0xff;
1382 slot = fp_offset_to_slot (cfg, offset);
1384 set_slot_everywhere (gcfg, slot, type);
1386 /* Liveness for these slots is handled by process_spill_slots () */
1388 if (cfg->verbose_level > 1) {
1389 if (type == SLOT_REF)
1390 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
1396 process_variables (MonoCompile *cfg)
1398 MonoCompileGC *gcfg = cfg->gc_info;
1399 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1400 int i, locals_min_slot, locals_max_slot, cindex;
1404 int locals_min_offset = gcfg->locals_min_offset;
1405 int locals_max_offset = gcfg->locals_max_offset;
1407 /* Slots for locals are NOREF by default */
1408 locals_min_slot = (locals_min_offset - gcfg->min_offset) / sizeof (mgreg_t);
1409 locals_max_slot = (locals_max_offset - gcfg->min_offset) / sizeof (mgreg_t);
1410 for (i = locals_min_slot; i < locals_max_slot; ++i) {
1411 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1415 * Compute the offset where variables are initialized in the first bblock, if any.
1417 pc_offsets = g_new0 (int, cfg->next_vreg);
1419 bb = cfg->bb_entry->next_bb;
1420 MONO_BB_FOR_EACH_INS (bb, tmp) {
1421 if (tmp->opcode == OP_GC_LIVENESS_DEF) {
1422 int vreg = tmp->inst_c1;
1423 if (pc_offsets [vreg] == 0) {
1424 g_assert (tmp->backend.pc_offset > 0);
1425 pc_offsets [vreg] = tmp->backend.pc_offset;
1431 * Stack slots holding arguments are initialized in the prolog.
1432 * This means we can treat them alive for the whole method.
1434 for (i = 0; i < cfg->num_varinfo; i++) {
1435 MonoInst *ins = cfg->varinfo [i];
1436 MonoType *t = ins->inst_vtype;
1439 gboolean byref, is_this = FALSE;
1440 gboolean is_arg = i < cfg->locals_start;
1442 if (ins == cfg->ret)
1445 vmv = MONO_VARINFO (cfg, i);
1447 /* For some reason, 'this' is byref */
1448 if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
1449 t = &cfg->method->klass->byval_arg;
1455 if (ins->opcode == OP_REGVAR) {
1457 GCSlotType slot_type;
1459 t = mini_type_get_underlying_type (NULL, t);
1462 g_assert (hreg < MONO_MAX_IREGS);
1465 slot_type = SLOT_PIN;
1467 slot_type = MONO_TYPE_IS_REFERENCE (t) ? SLOT_REF : SLOT_NOREF;
1469 if (slot_type == SLOT_PIN) {
1470 /* These have no live interval, be conservative */
1471 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1474 * Unlike variables allocated to the stack, we generate liveness info
1475 * for noref vars in registers in mono_spill_global_vars (), because
1476 * knowing that a register doesn't contain a ref allows us to mark its save
1477 * locations precisely.
1479 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1480 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1481 set_reg_slot (gcfg, hreg, cindex, slot_type);
1484 if (cfg->verbose_level > 1) {
1485 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
1491 if (ins->opcode != OP_REGOFFSET)
1494 if (ins->inst_offset % sizeof (mgreg_t) != 0)
1497 if (is_arg && ins->inst_offset >= gcfg->max_offset)
1498 /* In parent frame */
1501 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1503 if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
1504 /* These do not get stored in the prolog */
1505 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1507 if (cfg->verbose_level > 1) {
1508 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));
1513 if (MONO_TYPE_ISSTRUCT (t)) {
1515 gsize *bitmap = NULL;
1516 gboolean pin = FALSE;
1520 if (ins->backend.is_pinvoke)
1521 size = mono_class_native_size (ins->klass, NULL);
1523 size = mono_class_value_size (ins->klass, NULL);
1524 size_in_slots = ALIGN_TO (size, sizeof (mgreg_t)) / sizeof (mgreg_t);
1526 if (!ins->klass->has_references) {
1528 for (j = 0; j < size_in_slots; ++j)
1529 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1534 if (ins->klass->generic_container || mono_class_is_open_constructed_type (t)) {
1535 /* FIXME: Generic sharing */
1538 mono_class_compute_gc_descriptor (ins->klass);
1540 bitmap = mono_gc_get_bitmap_for_descr (ins->klass->gc_descr, &numbits);
1544 * Most vtypes are marked volatile because of the LDADDR instructions,
1545 * and they have no liveness information since they are decomposed
1546 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1547 * them during VZERO decomposition.
1549 if (!pc_offsets [vmv->vreg])
1553 if (ins->backend.is_pinvoke)
1556 if (cfg->verbose_level > 1)
1557 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));
1560 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1561 if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
1562 for (j = 0; j < numbits; ++j) {
1563 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1564 /* The descriptor is for the boxed object */
1565 set_slot (gcfg, (pos + j - (sizeof (MonoObject) / sizeof (mgreg_t))), cindex, pin ? SLOT_PIN : SLOT_REF);
1571 if (cfg->verbose_level > 1) {
1572 for (j = 0; j < numbits; ++j) {
1573 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD)))
1574 printf ("\t\t%s slot at 0x%x(fp) (slot = %d)\n", pin ? "pin" : "ref", (int)(ins->inst_offset + (j * sizeof (mgreg_t))), (int)(pos + j - (sizeof (MonoObject) / sizeof (mgreg_t))));
1578 if (cfg->verbose_level > 1)
1579 printf ("\t\tpinned\n");
1580 for (j = 0; j < size_in_slots; ++j) {
1581 set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
1590 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1591 /* Vret addr etc. */
1596 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1598 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1599 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1600 set_slot (gcfg, pos, cindex, SLOT_PIN);
1602 if (cfg->verbose_level > 1)
1603 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));
1608 * This is currently disabled, but could be enabled to debug crashes.
1611 if (t->type == MONO_TYPE_I) {
1613 * Variables created in mono_handle_global_vregs have type I, but they
1614 * could hold GC refs since the vregs they were created from might not been
1615 * marked as holding a GC ref. So be conservative.
1617 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1622 t = mini_type_get_underlying_type (NULL, t);
1624 if (!MONO_TYPE_IS_REFERENCE (t)) {
1625 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1626 if (cfg->verbose_level > 1)
1627 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));
1631 /* 'this' is marked INDIRECT for gshared methods */
1632 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1634 * For volatile variables, treat them alive from the point they are
1635 * initialized in the first bblock until the end of the method.
1638 set_slot_everywhere (gcfg, pos, SLOT_REF);
1639 } else if (pc_offsets [vmv->vreg]) {
1640 set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
1641 set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
1643 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1645 if (cfg->verbose_level > 1)
1646 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));
1651 /* Live for the whole method */
1652 set_slot_everywhere (gcfg, pos, SLOT_REF);
1654 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1655 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1656 set_slot (gcfg, pos, cindex, SLOT_REF);
1659 if (cfg->verbose_level > 1) {
1660 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));
1664 g_free (pc_offsets);
1668 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1671 * Convert a sp relative offset to a slot index. This is
1672 * platform specific.
1675 /* fp = sp + offset */
1676 g_assert (cfg->frame_reg == AMD64_RBP);
1677 return (- cfg->arch.sp_fp_offset + sp_offset);
1685 type_to_gc_slot_type (MonoType *t)
1689 t = mini_type_get_underlying_type (NULL, t);
1690 if (MONO_TYPE_IS_REFERENCE (t))
1693 if (MONO_TYPE_ISSTRUCT (t)) {
1694 MonoClass *klass = mono_class_from_mono_type (t);
1695 if (!klass->has_references) {
1707 process_param_area_slots (MonoCompile *cfg)
1709 MonoCompileGC *gcfg = cfg->gc_info;
1714 * These slots are used for passing parameters during calls. They are sp relative, not
1715 * fp relative, so they are harder to handle.
1717 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1718 /* The distance between fp and sp is not constant */
1721 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1723 for (i = 0; i < gcfg->ncallsites; ++i) {
1724 GCCallSite *callsite = gcfg->callsites [i];
1727 for (l = callsite->param_slots; l; l = l->next) {
1728 MonoInst *def = l->data;
1729 int sp_offset = def->inst_offset;
1730 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1731 int slot = fp_offset_to_slot (cfg, fp_offset);
1733 g_assert (slot >= 0 && slot < gcfg->nslots);
1734 is_param [slot] = TRUE;
1738 /* All param area slots are noref by default */
1739 for (i = 0; i < gcfg->nslots; ++i) {
1741 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1744 for (i = 0; i < gcfg->ncallsites; ++i) {
1745 GCCallSite *callsite = gcfg->callsites [i];
1748 for (l = callsite->param_slots; l; l = l->next) {
1749 MonoInst *def = l->data;
1750 MonoType *t = def->inst_vtype;
1751 int sp_offset = def->inst_offset;
1752 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1753 int slot = fp_offset_to_slot (cfg, fp_offset);
1754 GCSlotType type = type_to_gc_slot_type (t);
1756 /* The slot is live between the def instruction and the call */
1757 set_slot_in_range (gcfg, slot, def->backend.pc_offset, callsite->pc_offset + 1, type);
1758 if (cfg->verbose_level > 1)
1759 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);
1765 process_finally_clauses (MonoCompile *cfg)
1767 MonoCompileGC *gcfg = cfg->gc_info;
1768 GCCallSite **callsites;
1770 gboolean has_finally;
1771 int i, j, nslots, nregs;
1773 ncallsites = gcfg->ncallsites;
1774 nslots = gcfg->nslots;
1775 nregs = gcfg->nregs;
1776 callsites = gcfg->callsites;
1779 * The calls to the finally clauses don't show up in the cfg. See
1780 * test_0_liveness_8 ().
1781 * Variables accessed inside the finally clause are already marked VOLATILE by
1782 * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
1783 * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
1785 has_finally = FALSE;
1786 for (i = 0; i < cfg->header->num_clauses; ++i) {
1787 MonoExceptionClause *clause = &cfg->header->clauses [i];
1789 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1794 if (cfg->verbose_level > 1)
1795 printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
1796 for (j = 0; j < ncallsites; ++j) {
1797 MonoBasicBlock *bb = callsites [j]->bb;
1798 MonoExceptionClause *clause;
1799 gboolean is_in_finally = FALSE;
1801 for (i = 0; i < cfg->header->num_clauses; ++i) {
1802 clause = &cfg->header->clauses [i];
1804 if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
1805 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1806 is_in_finally = TRUE;
1812 if (is_in_finally) {
1813 for (i = 0; i < nslots; ++i)
1814 set_slot (gcfg, i, j, SLOT_PIN);
1815 for (i = 0; i < nregs; ++i)
1816 set_reg_slot (gcfg, i, j, SLOT_PIN);
1823 compute_frame_size (MonoCompile *cfg)
1825 int i, locals_min_offset, locals_max_offset, cfa_min_offset, cfa_max_offset;
1826 int min_offset, max_offset;
1827 MonoCompileGC *gcfg = cfg->gc_info;
1828 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1831 /* Compute min/max offsets from the fp */
1835 locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, sizeof (mgreg_t));
1836 locals_max_offset = cfg->locals_max_stack_offset;
1838 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
1842 locals_min_offset = ALIGN_TO (locals_min_offset, sizeof (mgreg_t));
1843 locals_max_offset = ALIGN_TO (locals_max_offset, sizeof (mgreg_t));
1845 min_offset = locals_min_offset;
1846 max_offset = locals_max_offset;
1849 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1850 MonoInst *ins = cfg->args [i];
1852 if (ins->opcode == OP_REGOFFSET)
1853 min_offset = MIN (min_offset, ins->inst_offset);
1857 g_assert (cfg->frame_reg == cfg->cfa_reg);
1858 g_assert (cfg->cfa_offset > 0);
1860 cfa_max_offset = cfg->cfa_offset;
1862 min_offset = MIN (min_offset, cfa_min_offset);
1863 max_offset = MAX (max_offset, cfa_max_offset);
1865 /* Fp relative slots */
1866 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1867 gint data = GPOINTER_TO_INT (l->data);
1868 int offset = data >> 16;
1870 min_offset = MIN (min_offset, offset);
1874 if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
1875 int stack_offset = ALIGN_TO (cfg->stack_offset, sizeof (mgreg_t));
1876 min_offset = MIN (min_offset, (-stack_offset));
1879 /* Param area slots */
1881 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
1884 gcfg->min_offset = min_offset;
1885 gcfg->max_offset = max_offset;
1886 gcfg->locals_min_offset = locals_min_offset;
1887 gcfg->locals_max_offset = locals_max_offset;
1891 init_gcfg (MonoCompile *cfg)
1893 int i, nregs, nslots;
1894 MonoCompileGC *gcfg = cfg->gc_info;
1895 GCCallSite **callsites;
1904 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1905 ncallsites += g_slist_length (bb->gc_callsites);
1907 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
1909 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1910 for (l = bb->gc_callsites; l; l = l->next)
1911 callsites [i++] = l->data;
1914 /* The callsites should already be ordered by pc offset */
1915 for (i = 1; i < ncallsites; ++i)
1916 g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
1919 * The stack frame looks like this:
1921 * <fp + max_offset> == cfa -> <end of previous frame>
1922 * <other stack slots>
1924 * <other stack slots>
1925 * fp + min_offset ->
1930 if (cfg->verbose_level > 1)
1931 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
1933 nslots = (gcfg->max_offset - gcfg->min_offset) / sizeof (mgreg_t);
1936 gcfg->nslots = nslots;
1937 gcfg->nregs = nregs;
1938 gcfg->callsites = callsites;
1939 gcfg->ncallsites = ncallsites;
1940 gcfg->stack_bitmap_width = ALIGN_TO (nslots, 8) / 8;
1941 gcfg->reg_bitmap_width = ALIGN_TO (nregs, 8) / 8;
1942 gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
1943 gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
1944 gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
1945 gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
1947 /* All slots start out as PIN */
1948 memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * ncallsites);
1949 for (i = 0; i < nregs; ++i) {
1951 * By default, registers are NOREF.
1952 * It is possible for a callee to save them before being defined in this method,
1953 * but the saved value is dead too, so it doesn't need to be marked.
1955 if ((cfg->used_int_regs & (1 << i)))
1956 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
1961 create_map (MonoCompile *cfg)
1964 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
1966 int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
1967 int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
1969 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
1970 MonoCompileGC *gcfg = cfg->gc_info;
1971 GCCallSite **callsites;
1973 guint8 *bitmap, *bitmaps;
1974 guint32 reg_ref_mask, reg_pin_mask;
1976 ncallsites = gcfg->ncallsites;
1977 nslots = gcfg->nslots;
1978 nregs = gcfg->nregs;
1979 callsites = gcfg->callsites;
1982 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
1983 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
1985 has_ref_slots = FALSE;
1986 has_pin_slots = FALSE;
1989 memset (ntypes, 0, sizeof (ntypes));
1990 for (i = 0; i < nslots; ++i) {
1991 gboolean has_ref = FALSE;
1992 gboolean has_pin = FALSE;
1994 for (j = 0; j < ncallsites; ++j) {
1995 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
1997 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
2002 has_ref_slots = TRUE;
2004 has_pin_slots = TRUE;
2007 ntypes [SLOT_REF] ++;
2009 ntypes [SLOT_PIN] ++;
2011 ntypes [SLOT_NOREF] ++;
2013 if (has_ref || has_pin) {
2020 start = end = nslots;
2022 g_assert (start != -1);
2023 g_assert (start < end);
2026 has_ref_regs = FALSE;
2027 has_pin_regs = FALSE;
2032 for (i = 0; i < nregs; ++i) {
2033 gboolean has_ref = FALSE;
2034 gboolean has_pin = FALSE;
2036 if (!(cfg->used_int_regs & (1 << i)))
2039 for (j = 0; j < ncallsites; ++j) {
2040 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i)) {
2045 for (j = 0; j < ncallsites; ++j) {
2046 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i)) {
2053 reg_ref_mask |= (1 << i);
2054 has_ref_regs = TRUE;
2058 reg_pin_mask |= (1 << i);
2059 has_pin_regs = TRUE;
2064 if (cfg->verbose_level > 1)
2065 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);
2067 /* Create the GC Map */
2069 stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
2070 stack_bitmap_size = stack_bitmap_width * ncallsites;
2071 reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
2072 reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
2073 reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
2074 reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
2075 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);
2077 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2079 map->frame_reg = cfg->frame_reg;
2080 map->start_offset = gcfg->min_offset;
2081 map->end_offset = gcfg->min_offset + (nslots * sizeof (mgreg_t));
2082 map->map_offset = start * sizeof (mgreg_t);
2083 map->nslots = end - start;
2084 map->has_ref_slots = has_ref_slots;
2085 map->has_pin_slots = has_pin_slots;
2086 map->has_ref_regs = has_ref_regs;
2087 map->has_pin_regs = has_pin_regs;
2088 g_assert (nregs < 32);
2089 map->used_int_regs = cfg->used_int_regs;
2090 map->reg_ref_mask = reg_ref_mask;
2091 map->reg_pin_mask = reg_pin_mask;
2092 map->nref_regs = nref_regs;
2093 map->npin_regs = npin_regs;
2095 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2098 if (has_ref_slots) {
2099 map->stack_ref_bitmap_offset = bitmaps_offset;
2100 bitmaps_offset += stack_bitmap_size;
2102 bitmap = &bitmaps [map->stack_ref_bitmap_offset];
2103 for (i = 0; i < nslots; ++i) {
2104 for (j = 0; j < ncallsites; ++j) {
2105 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
2106 set_bit (bitmap, stack_bitmap_width, j, i - start);
2110 if (has_pin_slots) {
2111 map->stack_pin_bitmap_offset = bitmaps_offset;
2112 bitmaps_offset += stack_bitmap_size;
2114 bitmap = &bitmaps [map->stack_pin_bitmap_offset];
2115 for (i = 0; i < nslots; ++i) {
2116 for (j = 0; j < ncallsites; ++j) {
2117 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
2118 set_bit (bitmap, stack_bitmap_width, j, i - start);
2123 map->reg_ref_bitmap_offset = bitmaps_offset;
2124 bitmaps_offset += reg_ref_bitmap_size;
2126 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2128 for (i = 0; i < nregs; ++i) {
2129 if (reg_ref_mask & (1 << i)) {
2130 for (j = 0; j < ncallsites; ++j) {
2131 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i))
2132 set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
2139 map->reg_pin_bitmap_offset = bitmaps_offset;
2140 bitmaps_offset += reg_pin_bitmap_size;
2142 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2144 for (i = 0; i < nregs; ++i) {
2145 if (reg_pin_mask & (1 << i)) {
2146 for (j = 0; j < ncallsites; ++j) {
2147 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i))
2148 set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
2156 map->ncallsites = ncallsites;
2157 if (cfg->code_len < 256)
2158 map->callsite_entry_size = 1;
2159 else if (cfg->code_len < 65536)
2160 map->callsite_entry_size = 2;
2162 map->callsite_entry_size = 4;
2164 /* Encode the GC Map */
2172 encode_gc_map (map, buf, &endbuf);
2173 g_assert (endbuf - buf < 256);
2175 encoded_size = endbuf - buf;
2176 alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
2178 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2179 //emap->ref_slots = map->ref_slots;
2181 /* Encoded fixed fields */
2182 p = &emap->encoded [0];
2183 //emap->encoded_size = encoded_size;
2184 memcpy (p, buf, encoded_size);
2187 /* Callsite table */
2188 p = (guint8*)ALIGN_TO ((guint64)p, map->callsite_entry_size);
2189 if (map->callsite_entry_size == 1) {
2190 guint8 *offsets = p;
2191 for (i = 0; i < ncallsites; ++i)
2192 offsets [i] = callsites [i]->pc_offset;
2193 stats.gc_callsites8_size += ncallsites * sizeof (guint8);
2194 } else if (map->callsite_entry_size == 2) {
2195 guint16 *offsets = (guint16*)p;
2196 for (i = 0; i < ncallsites; ++i)
2197 offsets [i] = callsites [i]->pc_offset;
2198 stats.gc_callsites16_size += ncallsites * sizeof (guint16);
2200 guint32 *offsets = (guint32*)p;
2201 for (i = 0; i < ncallsites; ++i)
2202 offsets [i] = callsites [i]->pc_offset;
2203 stats.gc_callsites32_size += ncallsites * sizeof (guint32);
2205 p += ncallsites * map->callsite_entry_size;
2208 memcpy (p, bitmaps, bitmaps_size);
2210 stats.gc_maps_size += alloc_size;
2211 stats.gc_callsites_size += ncallsites * map->callsite_entry_size;
2212 stats.gc_bitmaps_size += bitmaps_size;
2213 stats.gc_map_struct_size += sizeof (GCEncodedMap) + encoded_size;
2215 cfg->jit_info->gc_info = emap;
2218 stats.all_slots += nslots;
2219 stats.ref_slots += ntypes [SLOT_REF];
2220 stats.noref_slots += ntypes [SLOT_NOREF];
2221 stats.pin_slots += ntypes [SLOT_PIN];
2225 mini_gc_create_gc_map (MonoCompile *cfg)
2227 if (!cfg->compute_gc_maps)
2231 * During marking, all frames except the top frame are at a call site, and we mark the
2232 * top frame conservatively. This means that we only need to compute and record
2233 * GC maps for call sites.
2236 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2237 /* Without liveness info, the live ranges are not precise enough */
2240 mono_analyze_liveness_gc (cfg);
2242 compute_frame_size (cfg);
2246 process_spill_slots (cfg);
2247 process_other_slots (cfg);
2248 process_param_area_slots (cfg);
2249 process_variables (cfg);
2250 process_finally_clauses (cfg);
2256 parse_debug_options (void)
2261 env = getenv ("MONO_GCMAP_DEBUG");
2265 opts = g_strsplit (env, ",", -1);
2266 for (ptr = opts; ptr && *ptr; ptr ++) {
2267 /* No options yet */
2268 fprintf (stderr, "Invalid format for the MONO_GCMAP_DEBUG env variable: '%s'\n", env);
2279 memset (&cb, 0, sizeof (cb));
2280 cb.thread_attach_func = thread_attach_func;
2281 cb.thread_detach_func = thread_detach_func;
2282 cb.thread_suspend_func = thread_suspend_func;
2283 /* Comment this out to disable precise stack marking */
2284 cb.thread_mark_func = thread_mark_func;
2285 mono_gc_set_gc_callbacks (&cb);
2287 logfile = mono_sgen_get_logfile ();
2289 parse_debug_options ();
2291 mono_counters_register ("GC Maps size",
2292 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
2293 mono_counters_register ("GC Call Sites size",
2294 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
2295 mono_counters_register ("GC Bitmaps size",
2296 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
2297 mono_counters_register ("GC Map struct size",
2298 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
2299 mono_counters_register ("GC Call Sites encoded using 8 bits",
2300 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
2301 mono_counters_register ("GC Call Sites encoded using 16 bits",
2302 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
2303 mono_counters_register ("GC Call Sites encoded using 32 bits",
2304 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
2306 mono_counters_register ("GC Map slots (all)",
2307 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
2308 mono_counters_register ("GC Map slots (ref)",
2309 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.ref_slots);
2310 mono_counters_register ("GC Map slots (noref)",
2311 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.noref_slots);
2312 mono_counters_register ("GC Map slots (pin)",
2313 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
2315 mono_counters_register ("GC TLS Data size",
2316 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2318 mono_counters_register ("Stack space scanned (all)",
2319 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
2320 mono_counters_register ("Stack space scanned (native)",
2321 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_native);
2322 mono_counters_register ("Stack space scanned (other)",
2323 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_other);
2324 mono_counters_register ("Stack space scanned (using GC Maps)",
2325 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
2326 mono_counters_register ("Stack space scanned (precise)",
2327 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_precisely);
2328 mono_counters_register ("Stack space scanned (pin)",
2329 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_conservatively);
2330 mono_counters_register ("Stack space scanned (pin registers)",
2331 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_registers);
2342 mini_gc_init_gc_map (MonoCompile *cfg)
2347 mini_gc_create_gc_map (MonoCompile *cfg)
2352 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2357 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2366 * Set GC specific options in CFG.
2369 mini_gc_init_cfg (MonoCompile *cfg)
2371 if (mono_gc_is_moving ()) {
2372 cfg->disable_ref_noref_stack_slot_share = TRUE;
2373 cfg->gen_write_barriers = TRUE;
2376 mini_gc_init_gc_map (cfg);
2380 * Problems with the current code:
2381 * - the stack walk is slow
2382 * - vtypes/refs used in EH regions are treated conservatively
2383 * - if the code is finished, less pinning will be done, causing problems because
2384 * we promote all surviving objects to old-gen.
2385 * - the unwind code can't handle a method stopped inside a finally region, it thinks the caller is
2386 * another method, but in reality it is either the exception handling code or the CALL_HANDLER opcode.
2387 * This manifests in "Unable to find ip offset x in callsite list" assertions.
2388 * - the unwind code also can't handle frames which are in the epilog, since the unwind info is not
2393 * Ideas for creating smaller GC maps:
2394 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2396 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2398 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2399 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2400 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2401 * from an index entry.