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.
653 if (reg_locations [i]) {
654 DEBUG (printf ("\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
655 mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
656 scanned_registers += sizeof (mgreg_t);
659 reg_locations [i] = new_reg_locations [i];
661 DEBUG (printf ("\treg %s is at location %p.\n", mono_arch_regname (i), reg_locations [i]));
665 g_assert ((guint64)stack_limit % sizeof (mgreg_t) == 0);
667 #ifdef MONO_ARCH_HAVE_FIND_JIT_INFO_EXT
668 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);
675 /* The last frame can be in any state so mark conservatively */
683 /* This happens with native-to-managed transitions */
684 if (!(MONO_CONTEXT_GET_IP (&ctx) >= ji->code_start && (guint8*)MONO_CONTEXT_GET_IP (&ctx) < (guint8*)ji->code_start + ji->code_size))
687 /* These frames are very problematic */
688 if (ji->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
689 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));
690 DEBUG (printf ("\tSkip.\n"));
694 /* All the other frames are at a call site */
696 if (tls->nframes == MAX_FRAMES) {
698 * Can't save information since the array is full. So scan the rest of the
699 * stack conservatively.
704 /* Scan the frame of this method */
707 * A frame contains the following:
712 * - localloc-ed memory
714 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
715 g_assert (pc_offset >= 0);
720 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));
721 DEBUG (printf ("\tNo GC Map.\n"));
726 * Debugging aid to control the number of frames scanned precisely
728 if (!precise_frame_limit_inited) {
729 if (getenv ("MONO_PRECISE_COUNT"))
730 precise_frame_limit = atoi (getenv ("MONO_PRECISE_COUNT"));
731 precise_frame_limit_inited = TRUE;
734 if (precise_frame_limit != -1) {
735 if (precise_frame_count [FALSE] == precise_frame_limit)
736 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (ji->method, TRUE));
737 if (precise_frame_count [FALSE] > precise_frame_limit)
740 precise_frame_count [FALSE] ++;
742 /* Decode the encoded GC map */
744 memset (map, 0, sizeof (GCMap));
745 decode_gc_map (&emap->encoded [0], map, &p);
746 p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
747 map->callsites.offsets8 = p;
748 p += map->callsite_entry_size * map->ncallsites;
752 if (map->frame_reg == AMD64_RSP)
753 fp = (guint8*)ctx.rsp;
754 else if (map->frame_reg == AMD64_RBP)
755 fp = (guint8*)ctx.rbp;
757 g_assert_not_reached ();
760 g_assert_not_reached ();
763 frame_start = fp + map->start_offset + map->map_offset;
764 frame_end = fp + map->end_offset;
766 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));
768 /* Find the callsite index */
769 if (ji->code_size < 256) {
770 for (i = 0; i < map->ncallsites; ++i)
771 /* ip points inside the call instruction */
772 if (map->callsites.offsets8 [i] == pc_offset + 1)
774 } else if (ji->code_size < 65536) {
775 // FIXME: Use a binary search
776 for (i = 0; i < map->ncallsites; ++i)
777 /* ip points inside the call instruction */
778 if (map->callsites.offsets16 [i] == pc_offset + 1)
781 // FIXME: Use a binary search
782 for (i = 0; i < map->ncallsites; ++i)
783 /* ip points inside the call instruction */
784 if (map->callsites.offsets32 [i] == pc_offset + 1)
787 if (i == map->ncallsites) {
788 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (ji->method, TRUE));
789 g_assert_not_reached ();
793 g_assert (frame_start >= stack_limit);
795 if (frame_start > stack_limit) {
796 /* This scans the previously skipped frames as well */
797 DEBUG (printf ("\tscan area %p-%p.\n", stack_limit, frame_start));
798 mono_gc_conservatively_scan_area (stack_limit, frame_start);
801 /* Mark stack slots */
802 if (map->has_pin_slots) {
803 int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
804 guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
809 for (i = 0; i < map->nslots; ++i) {
810 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
812 DEBUG (printf ("\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
813 mono_gc_conservatively_scan_area (p, p + sizeof (mgreg_t));
814 scanned_conservatively += sizeof (mgreg_t);
816 scanned_precisely += sizeof (mgreg_t);
818 p += sizeof (mgreg_t);
821 scanned_precisely += (map->nslots * sizeof (mgreg_t));
824 /* The area outside of start-end is NOREF */
825 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * sizeof (mgreg_t));
828 precise_regmask = map->used_int_regs | (1 << map->frame_reg);
829 if (map->has_pin_regs) {
830 int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
831 guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
833 for (i = 0; i < NREGS; ++i) {
834 if (!(map->used_int_regs & (1 << i)))
837 if (!(map->reg_pin_mask & (1 << i)))
840 if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
841 DEBUG (printf ("\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
842 precise_regmask &= ~(1 << i);
848 scanned += map->end_offset - map->start_offset;
850 g_assert (scanned == scanned_precisely + scanned_conservatively);
852 stack_limit = frame_end;
854 /* Save information for the precise pass */
855 fi = &tls->frames [tls->nframes];
856 fi->nslots = map->nslots;
857 bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
858 if (map->has_ref_slots)
859 fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
862 fi->frame_start_offset = frame_start - stack_start;
863 fi->nreg_locations = 0;
867 if (map->has_ref_regs) {
868 int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
869 guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
871 for (i = 0; i < NREGS; ++i) {
872 if (!(map->reg_ref_mask & (1 << i)))
875 if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
876 DEBUG (fi->regs [fi->nreg_locations] = i);
877 DEBUG (printf ("\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
878 fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
879 fi->nreg_locations ++;
886 * Clear locations of precisely stacked registers.
888 if (precise_regmask) {
889 for (i = 0; i < NREGS; ++i) {
890 if (precise_regmask & (1 << i)) {
892 * The method uses this register, and we have precise info for it.
893 * This means the location will be scanned precisely.
894 * Tell the code at the beginning of the loop that this location is
897 if (reg_locations [i])
898 DEBUG (printf ("\treg %s at location %p is precise.\n", mono_arch_regname (i), reg_locations [i]));
899 reg_locations [i] = NULL;
907 /* Scan the remaining register save locations */
908 for (i = 0; i < MONO_MAX_IREGS; ++i) {
909 if (reg_locations [i]) {
910 DEBUG (printf ("\tscan saved reg location %p.\n", reg_locations [i]));
911 mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
912 scanned_registers += sizeof (mgreg_t);
914 if (new_reg_locations [i]) {
915 DEBUG (printf ("\tscan saved reg location %p.\n", new_reg_locations [i]));
916 mono_gc_conservatively_scan_area (new_reg_locations [i], new_reg_locations [i] + sizeof (mgreg_t));
917 scanned_registers += sizeof (mgreg_t);
921 if (stack_limit < stack_end) {
922 DEBUG (printf ("\tscan area %p-%p.\n", stack_limit, stack_end));
923 mono_gc_conservatively_scan_area (stack_limit, stack_end);
926 DEBUG (printf ("Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
928 stats.scanned_stacks += stack_end - stack_start;
929 stats.scanned += scanned;
930 stats.scanned_precisely += scanned_precisely;
931 stats.scanned_conservatively += scanned_conservatively;
932 stats.scanned_registers += scanned_registers;
934 //mono_gc_conservatively_scan_area (stack_start, stack_end);
940 * Mark a thread stack precisely based on information saved during the conservative
944 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
953 for (findex = 0; findex < tls->nframes; findex ++) {
954 /* Load information saved by the !precise pass */
955 fi = &tls->frames [findex];
956 frame_start = stack_start + fi->frame_start_offset;
959 * FIXME: Add a function to mark using a bitmap, to avoid doing a
960 * call for each object.
963 /* Mark stack slots */
965 guint8 *ref_bitmap = fi->bitmap;
968 for (i = 0; i < fi->nslots; ++i) {
969 MonoObject **ptr = (MonoObject**)(frame_start + (i * sizeof (mgreg_t)));
971 live = ref_bitmap [i / 8] & (1 << (i % 8));
974 MonoObject *obj = *ptr;
976 DEBUG (printf ("\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
977 *ptr = mono_gc_scan_object (obj);
978 DEBUG (printf (" %p.\n", *ptr));
980 DEBUG (printf ("\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
985 * This is disabled because the pointer takes up a lot of space.
986 * Stack slots might be shared between ref and non-ref variables ?
988 if (map->ref_slots [i / 8] & (1 << (i % 8))) {
989 DEBUG (printf ("\tref %s0x%x(fp)=%p: dead (%p)\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, *ptr));
991 * Fail fast if the live range is incorrect, and
992 * the JITted code tries to access this object
1001 /* Mark registers */
1004 * Registers are different from stack slots, they have no address where they
1005 * are stored. Instead, some frame below this frame in the stack saves them
1006 * in its prolog to the stack. We can mark this location precisely.
1008 for (i = 0; i < fi->nreg_locations; ++i) {
1010 * reg_locations [i] contains the address of the stack slot where
1011 * a reg was last saved, so mark that slot.
1013 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1014 MonoObject *obj = *ptr;
1017 DEBUG (printf ("\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
1018 *ptr = mono_gc_scan_object (obj);
1019 DEBUG (printf (" %p.\n", *ptr));
1021 DEBUG (printf ("\treg %s saved at %p: %p", mono_arch_regname (fi->regs [i]), ptr, obj));
1030 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1031 * call, and TRUE at the second. USER_DATA points to a TlsData
1032 * structure filled up by thread_suspend_func.
1035 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise)
1037 TlsData *tls = user_data;
1039 DEBUG (printf ("*** %s stack marking %p-%p ***\n", precise ? "Precise" : "Conservative", stack_start, stack_end));
1042 conservative_pass (tls, stack_start, stack_end);
1044 precise_pass (tls, stack_start, stack_end);
1048 mini_gc_init_gc_map (MonoCompile *cfg)
1050 if (COMPILE_LLVM (cfg))
1054 /* Debugging support */
1056 static int precise_count;
1059 if (getenv ("MONO_GCMAP_COUNT")) {
1060 if (precise_count == atoi (getenv ("MONO_GCMAP_COUNT")))
1061 printf ("LAST: %s\n", mono_method_full_name (cfg->method, TRUE));
1062 if (precise_count > atoi (getenv ("MONO_GCMAP_COUNT")))
1068 cfg->compute_gc_maps = TRUE;
1070 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1074 * mini_gc_set_slot_type_from_fp:
1076 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1077 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1078 * need to call this function for those slots.
1081 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1083 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1085 if (!cfg->compute_gc_maps)
1088 g_assert (slot_offset % sizeof (mgreg_t) == 0);
1090 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1094 * mini_gc_set_slot_type_from_cfa:
1096 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1097 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1098 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1099 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1100 * call this function for those slots.
1103 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1105 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1106 int slot = - (slot_offset / sizeof (mgreg_t));
1108 if (!cfg->compute_gc_maps)
1111 g_assert (slot_offset <= 0);
1112 g_assert (slot_offset % sizeof (mgreg_t) == 0);
1114 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1118 fp_offset_to_slot (MonoCompile *cfg, int offset)
1120 MonoCompileGC *gcfg = cfg->gc_info;
1122 return (offset - gcfg->min_offset) / sizeof (mgreg_t);
1126 slot_to_fp_offset (MonoCompile *cfg, int slot)
1128 MonoCompileGC *gcfg = cfg->gc_info;
1130 return (slot * sizeof (mgreg_t)) + gcfg->min_offset;
1134 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1136 g_assert (slot >= 0 && slot < gcfg->nslots);
1138 if (type == SLOT_PIN) {
1139 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1140 set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1141 } else if (type == SLOT_REF) {
1142 set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1143 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1144 } else if (type == SLOT_NOREF) {
1145 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1146 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1151 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1155 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1156 set_slot (gcfg, slot, cindex, type);
1160 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1164 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1165 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1166 if (callsite_offset >= from && callsite_offset < to)
1167 set_slot (gcfg, slot, cindex, type);
1172 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1174 g_assert (slot >= 0 && slot < gcfg->nregs);
1176 if (type == SLOT_PIN) {
1177 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1178 set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1179 } else if (type == SLOT_REF) {
1180 set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1181 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1182 } else if (type == SLOT_NOREF) {
1183 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1184 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1189 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1193 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1194 set_reg_slot (gcfg, slot, cindex, type);
1198 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1202 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1203 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1204 if (callsite_offset >= from && callsite_offset < to)
1205 set_reg_slot (gcfg, slot, cindex, type);
1210 process_spill_slots (MonoCompile *cfg)
1212 MonoCompileGC *gcfg = cfg->gc_info;
1217 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1218 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1219 for (l = bb->spill_slot_defs; l; l = l->next) {
1220 MonoInst *def = l->data;
1221 int spill_slot = def->inst_c0;
1222 int bank = def->inst_c1;
1223 int offset = cfg->spill_info [bank][spill_slot].offset;
1224 int slot = fp_offset_to_slot (cfg, offset);
1226 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1227 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1231 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1232 for (l = bb->spill_slot_defs; l; l = l->next) {
1233 MonoInst *def = l->data;
1234 int spill_slot = def->inst_c0;
1235 int bank = def->inst_c1;
1236 int offset = cfg->spill_info [bank][spill_slot].offset;
1237 int slot = fp_offset_to_slot (cfg, offset);
1240 if (bank == MONO_REG_INT_MP)
1246 * Extend the live interval for the GC tracked spill slots
1247 * defined in this bblock.
1248 * FIXME: This is not needed.
1250 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1252 if (cfg->verbose_level > 1)
1253 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
1257 /* Set fp spill slots to NOREF */
1258 for (i = 0; i < cfg->spill_info_len [MONO_REG_DOUBLE]; ++i) {
1259 int offset = cfg->spill_info [MONO_REG_DOUBLE][i].offset;
1265 slot = fp_offset_to_slot (cfg, offset);
1267 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1269 if (cfg->verbose_level > 1)
1270 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1273 /* Set int spill slots to NOREF */
1274 for (i = 0; i < cfg->spill_info_len [MONO_REG_INT]; ++i) {
1275 int offset = cfg->spill_info [MONO_REG_INT][i].offset;
1281 slot = fp_offset_to_slot (cfg, offset);
1283 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1284 if (cfg->verbose_level > 1)
1285 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1290 * process_other_slots:
1292 * Process stack slots registered using mini_gc_set_slot_type_... ().
1295 process_other_slots (MonoCompile *cfg)
1297 MonoCompileGC *gcfg = cfg->gc_info;
1300 /* Relative to the CFA */
1301 for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
1302 guint data = GPOINTER_TO_UINT (l->data);
1303 int cfa_slot = data >> 16;
1304 GCSlotType type = data & 0xff;
1308 * Map the cfa relative slot to an fp relative slot.
1309 * slot_addr == cfa - <cfa_slot>*4/8
1310 * fp + cfa_offset == cfa
1311 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1313 slot = (cfg->cfa_offset / sizeof (mgreg_t)) - cfa_slot - (gcfg->min_offset / sizeof (mgreg_t));
1315 set_slot_everywhere (gcfg, slot, type);
1317 if (cfg->verbose_level > 1) {
1318 int fp_offset = slot_to_fp_offset (cfg, slot);
1319 if (type == SLOT_NOREF)
1320 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)));
1324 /* Relative to the FP */
1325 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1326 gint data = GPOINTER_TO_INT (l->data);
1327 int offset = data >> 16;
1328 GCSlotType type = data & 0xff;
1331 slot = fp_offset_to_slot (cfg, offset);
1333 set_slot_everywhere (gcfg, slot, type);
1335 /* Liveness for these slots is handled by process_spill_slots () */
1337 if (cfg->verbose_level > 1) {
1338 if (type == SLOT_REF)
1339 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
1345 process_variables (MonoCompile *cfg)
1347 MonoCompileGC *gcfg = cfg->gc_info;
1348 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1349 int i, locals_min_slot, locals_max_slot, cindex;
1353 int locals_min_offset = gcfg->locals_min_offset;
1354 int locals_max_offset = gcfg->locals_max_offset;
1356 /* Slots for locals are NOREF by default */
1357 locals_min_slot = (locals_min_offset - gcfg->min_offset) / sizeof (mgreg_t);
1358 locals_max_slot = (locals_max_offset - gcfg->min_offset) / sizeof (mgreg_t);
1359 for (i = locals_min_slot; i < locals_max_slot; ++i) {
1360 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1364 * Compute the offset where variables are initialized in the first bblock, if any.
1366 pc_offsets = g_new0 (int, cfg->next_vreg);
1368 bb = cfg->bb_entry->next_bb;
1369 MONO_BB_FOR_EACH_INS (bb, tmp) {
1370 if (tmp->opcode == OP_GC_LIVENESS_DEF) {
1371 int vreg = tmp->inst_c1;
1372 if (pc_offsets [vreg] == 0) {
1373 g_assert (tmp->backend.pc_offset > 0);
1374 pc_offsets [vreg] = tmp->backend.pc_offset;
1380 * Stack slots holding arguments are initialized in the prolog.
1381 * This means we can treat them alive for the whole method.
1383 for (i = 0; i < cfg->num_varinfo; i++) {
1384 MonoInst *ins = cfg->varinfo [i];
1385 MonoType *t = ins->inst_vtype;
1388 gboolean byref, is_this = FALSE;
1389 gboolean is_arg = i < cfg->locals_start;
1391 if (ins == cfg->ret)
1394 vmv = MONO_VARINFO (cfg, i);
1396 /* For some reason, 'this' is byref */
1397 if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
1398 t = &cfg->method->klass->byval_arg;
1404 if (ins->opcode == OP_REGVAR) {
1406 GCSlotType slot_type;
1408 t = mini_type_get_underlying_type (NULL, t);
1411 g_assert (hreg < MONO_MAX_IREGS);
1413 // FIXME: Add back this check
1415 if (is_arg && gcfg->reg_live_intervals [hreg]) {
1417 * FIXME: This argument shares a hreg with a local, we can't add the whole
1418 * method as a live interval, since it would overlap with the locals
1426 slot_type = SLOT_PIN;
1428 slot_type = MONO_TYPE_IS_REFERENCE (t) ? SLOT_REF : SLOT_NOREF;
1431 /* Live for the whole method */
1432 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1434 if (slot_type == SLOT_PIN) {
1435 /* These have no live interval, be conservative */
1436 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1439 * Unlike variables allocated to the stack, we generate liveness info
1440 * for noref vars in registers in mono_spill_global_vars (), because
1441 * knowing that a register doesn't contain a ref allows us to mark its save
1442 * locations precisely.
1444 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1445 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1446 set_reg_slot (gcfg, hreg, cindex, slot_type);
1450 if (cfg->verbose_level > 1) {
1451 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
1457 if (ins->opcode != OP_REGOFFSET)
1460 if (ins->inst_offset % sizeof (mgreg_t) != 0)
1463 if (is_arg && ins->inst_offset >= gcfg->max_offset)
1464 /* In parent frame */
1467 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1469 if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
1470 /* These do not get stored in the prolog */
1471 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1473 if (cfg->verbose_level > 1) {
1474 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));
1479 if (MONO_TYPE_ISSTRUCT (t)) {
1481 gsize *bitmap = NULL;
1482 gboolean pin = FALSE;
1486 if (ins->backend.is_pinvoke)
1487 size = mono_class_native_size (ins->klass, NULL);
1489 size = mono_class_value_size (ins->klass, NULL);
1490 size_in_slots = ALIGN_TO (size, sizeof (mgreg_t)) / sizeof (mgreg_t);
1492 if (!ins->klass->has_references) {
1494 for (j = 0; j < size_in_slots; ++j)
1495 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1500 if (ins->klass->generic_container || mono_class_is_open_constructed_type (t)) {
1501 /* FIXME: Generic sharing */
1504 mono_class_compute_gc_descriptor (ins->klass);
1506 bitmap = mono_gc_get_bitmap_for_descr (ins->klass->gc_descr, &numbits);
1510 * Most vtypes are marked volatile because of the LDADDR instructions,
1511 * and they have no liveness information since they are decomposed
1512 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1513 * them during VZERO decomposition.
1515 if (!pc_offsets [vmv->vreg])
1519 if (ins->backend.is_pinvoke)
1523 if (pc_offsets [vmv->vreg]) {
1524 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1525 if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
1526 for (j = 0; j < numbits; ++j) {
1527 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1528 /* The descriptor is for the boxed object */
1529 set_slot (gcfg, (pos + j - (sizeof (MonoObject) / sizeof (gpointer))), cindex, pin ? SLOT_PIN : SLOT_REF);
1536 for (j = 0; j < size_in_slots; ++j) {
1537 set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
1543 if (cfg->verbose_level > 1) {
1544 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));
1550 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1551 /* Vret addr etc. */
1556 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1558 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1559 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1560 set_slot (gcfg, pos, cindex, SLOT_PIN);
1562 if (cfg->verbose_level > 1)
1563 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));
1568 * This is currently disabled, but could be enabled to debug crashes.
1571 if (t->type == MONO_TYPE_I) {
1573 * Variables created in mono_handle_global_vregs have type I, but they
1574 * could hold GC refs since the vregs they were created from might not been
1575 * marked as holding a GC ref. So be conservative.
1577 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1582 t = mini_type_get_underlying_type (NULL, t);
1584 if (!MONO_TYPE_IS_REFERENCE (t)) {
1585 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1586 if (cfg->verbose_level > 1)
1587 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));
1591 /* 'this' is marked INDIRECT for gshared methods */
1592 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1594 * For volatile variables, treat them alive from the point they are
1595 * initialized in the first bblock until the end of the method.
1598 set_slot_everywhere (gcfg, pos, SLOT_REF);
1599 } else if (pc_offsets [vmv->vreg]) {
1600 set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
1601 set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
1603 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1605 if (cfg->verbose_level > 1)
1606 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));
1611 /* Live for the whole method */
1612 set_slot_everywhere (gcfg, pos, SLOT_REF);
1614 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1615 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1616 set_slot (gcfg, pos, cindex, SLOT_REF);
1619 if (cfg->verbose_level > 1) {
1620 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));
1624 g_free (pc_offsets);
1628 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1631 * Convert a sp relative offset to a slot index. This is
1632 * platform specific.
1635 /* fp = sp + offset */
1636 g_assert (cfg->frame_reg == AMD64_RBP);
1637 return (- cfg->arch.sp_fp_offset + sp_offset);
1645 type_to_gc_slot_type (MonoType *t)
1649 t = mini_type_get_underlying_type (NULL, t);
1650 if (MONO_TYPE_IS_REFERENCE (t))
1653 if (MONO_TYPE_ISSTRUCT (t)) {
1654 MonoClass *klass = mono_class_from_mono_type (t);
1655 if (!klass->has_references) {
1667 process_param_area_slots (MonoCompile *cfg)
1669 MonoCompileGC *gcfg = cfg->gc_info;
1674 * These slots are used for passing parameters during calls. They are sp relative, not
1675 * fp relative, so they are harder to handle.
1677 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1678 /* The distance between fp and sp is not constant */
1681 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1683 for (i = 0; i < gcfg->ncallsites; ++i) {
1684 GCCallSite *callsite = gcfg->callsites [i];
1687 for (l = callsite->param_slots; l; l = l->next) {
1688 MonoInst *def = l->data;
1689 int sp_offset = def->inst_offset;
1690 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1691 int slot = fp_offset_to_slot (cfg, fp_offset);
1693 g_assert (slot >= 0 && slot < gcfg->nslots);
1694 is_param [slot] = TRUE;
1698 /* All param area slots are noref by default */
1699 for (i = 0; i < gcfg->nslots; ++i) {
1701 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1704 for (i = 0; i < gcfg->ncallsites; ++i) {
1705 GCCallSite *callsite = gcfg->callsites [i];
1708 for (l = callsite->param_slots; l; l = l->next) {
1709 MonoInst *def = l->data;
1710 MonoType *t = def->inst_vtype;
1711 int sp_offset = def->inst_offset;
1712 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1713 int slot = fp_offset_to_slot (cfg, fp_offset);
1714 GCSlotType type = type_to_gc_slot_type (t);
1716 /* The slot is live between the def instruction and the call */
1717 set_slot_in_range (gcfg, slot, def->backend.pc_offset, callsite->pc_offset + 1, type);
1718 if (cfg->verbose_level > 1)
1719 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);
1725 process_finally_clauses (MonoCompile *cfg)
1727 MonoCompileGC *gcfg = cfg->gc_info;
1728 GCCallSite **callsites;
1730 gboolean has_finally;
1731 int i, j, nslots, nregs;
1733 ncallsites = gcfg->ncallsites;
1734 nslots = gcfg->nslots;
1735 nregs = gcfg->nregs;
1736 callsites = gcfg->callsites;
1739 * The calls to the finally clauses don't show up in the cfg. See
1740 * test_0_liveness_8 ().
1741 * Variables accessed inside the finally clause are already marked VOLATILE by
1742 * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
1743 * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
1745 has_finally = FALSE;
1746 for (i = 0; i < cfg->header->num_clauses; ++i) {
1747 MonoExceptionClause *clause = &cfg->header->clauses [i];
1749 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1754 if (cfg->verbose_level > 1)
1755 printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
1756 for (j = 0; j < ncallsites; ++j) {
1757 MonoBasicBlock *bb = callsites [j]->bb;
1758 MonoExceptionClause *clause;
1759 gboolean is_in_finally = FALSE;
1761 for (i = 0; i < cfg->header->num_clauses; ++i) {
1762 clause = &cfg->header->clauses [i];
1764 if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
1765 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1766 is_in_finally = TRUE;
1772 if (is_in_finally) {
1773 for (i = 0; i < nslots; ++i)
1774 set_slot (gcfg, i, j, SLOT_PIN);
1775 for (i = 0; i < nregs; ++i)
1776 set_reg_slot (gcfg, i, j, SLOT_PIN);
1783 compute_frame_size (MonoCompile *cfg)
1785 int i, locals_min_offset, locals_max_offset, cfa_min_offset, cfa_max_offset;
1786 int min_offset, max_offset;
1787 MonoCompileGC *gcfg = cfg->gc_info;
1788 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1791 /* Compute min/max offsets from the fp */
1795 locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, sizeof (mgreg_t));
1796 locals_max_offset = cfg->locals_max_stack_offset;
1798 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
1802 locals_min_offset = ALIGN_TO (locals_min_offset, sizeof (mgreg_t));
1803 locals_max_offset = ALIGN_TO (locals_max_offset, sizeof (mgreg_t));
1805 min_offset = locals_min_offset;
1806 max_offset = locals_max_offset;
1809 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1810 MonoInst *ins = cfg->args [i];
1812 if (ins->opcode == OP_REGOFFSET)
1813 min_offset = MIN (min_offset, ins->inst_offset);
1817 g_assert (cfg->frame_reg == cfg->cfa_reg);
1818 g_assert (cfg->cfa_offset > 0);
1820 cfa_max_offset = cfg->cfa_offset;
1822 min_offset = MIN (min_offset, cfa_min_offset);
1823 max_offset = MAX (max_offset, cfa_max_offset);
1825 /* Fp relative slots */
1826 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1827 gint data = GPOINTER_TO_INT (l->data);
1828 int offset = data >> 16;
1830 min_offset = MIN (min_offset, offset);
1834 if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
1835 int stack_offset = ALIGN_TO (cfg->stack_offset, sizeof (mgreg_t));
1836 min_offset = MIN (min_offset, (-stack_offset));
1839 /* Param area slots */
1841 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
1844 gcfg->min_offset = min_offset;
1845 gcfg->max_offset = max_offset;
1846 gcfg->locals_min_offset = locals_min_offset;
1847 gcfg->locals_max_offset = locals_max_offset;
1851 init_gcfg (MonoCompile *cfg)
1853 int i, nregs, nslots;
1854 MonoCompileGC *gcfg = cfg->gc_info;
1855 GCCallSite **callsites;
1864 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1865 ncallsites += g_slist_length (bb->gc_callsites);
1867 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
1869 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1870 for (l = bb->gc_callsites; l; l = l->next)
1871 callsites [i++] = l->data;
1874 /* The callsites should already be ordered by pc offset */
1875 for (i = 1; i < ncallsites; ++i)
1876 g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
1879 * The stack frame looks like this:
1881 * <fp + max_offset> == cfa -> <end of previous frame>
1882 * <other stack slots>
1884 * <other stack slots>
1885 * fp + min_offset ->
1890 if (cfg->verbose_level > 1)
1891 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
1893 nslots = (gcfg->max_offset - gcfg->min_offset) / sizeof (mgreg_t);
1896 gcfg->nslots = nslots;
1897 gcfg->nregs = nregs;
1898 gcfg->callsites = callsites;
1899 gcfg->ncallsites = ncallsites;
1900 gcfg->stack_bitmap_width = ALIGN_TO (nslots, 8) / 8;
1901 gcfg->reg_bitmap_width = ALIGN_TO (nregs, 8) / 8;
1902 gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
1903 gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
1904 gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
1905 gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
1907 /* All slots start out as PIN */
1908 memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * ncallsites);
1909 for (i = 0; i < nregs; ++i) {
1911 * By default, registers are NOREF.
1912 * It is possible for a callee to save them before being defined in this method,
1913 * but the saved value is dead too, so it doesn't need to be marked.
1915 if ((cfg->used_int_regs & (1 << i)))
1916 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
1921 create_map (MonoCompile *cfg)
1924 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
1926 int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
1927 int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
1929 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
1930 MonoCompileGC *gcfg = cfg->gc_info;
1931 GCCallSite **callsites;
1933 guint8 *bitmap, *bitmaps;
1934 guint32 reg_ref_mask, reg_pin_mask;
1936 ncallsites = gcfg->ncallsites;
1937 nslots = gcfg->nslots;
1938 nregs = gcfg->nregs;
1939 callsites = gcfg->callsites;
1942 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
1943 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
1945 has_ref_slots = FALSE;
1946 has_pin_slots = FALSE;
1949 memset (ntypes, 0, sizeof (ntypes));
1950 for (i = 0; i < nslots; ++i) {
1951 gboolean has_ref = FALSE;
1952 gboolean has_pin = FALSE;
1954 for (j = 0; j < ncallsites; ++j) {
1955 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
1957 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
1962 has_ref_slots = TRUE;
1964 has_pin_slots = TRUE;
1967 ntypes [SLOT_REF] ++;
1969 ntypes [SLOT_PIN] ++;
1971 ntypes [SLOT_NOREF] ++;
1973 if (has_ref || has_pin) {
1980 start = end = nslots;
1982 g_assert (start != -1);
1983 g_assert (start < end);
1986 has_ref_regs = FALSE;
1987 has_pin_regs = FALSE;
1992 for (i = 0; i < nregs; ++i) {
1993 gboolean has_ref = FALSE;
1994 gboolean has_pin = FALSE;
1996 if (!(cfg->used_int_regs & (1 << i)))
1999 for (j = 0; j < ncallsites; ++j) {
2000 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i)) {
2005 for (j = 0; j < ncallsites; ++j) {
2006 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i)) {
2013 reg_ref_mask |= (1 << i);
2014 has_ref_regs = TRUE;
2018 reg_pin_mask |= (1 << i);
2019 has_pin_regs = TRUE;
2024 if (cfg->verbose_level > 1)
2025 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);
2027 /* Create the GC Map */
2029 stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
2030 stack_bitmap_size = stack_bitmap_width * ncallsites;
2031 reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
2032 reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
2033 reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
2034 reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
2035 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);
2037 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2039 map->frame_reg = cfg->frame_reg;
2040 map->start_offset = gcfg->min_offset;
2041 map->end_offset = gcfg->min_offset + (nslots * sizeof (mgreg_t));
2042 map->map_offset = start * sizeof (mgreg_t);
2043 map->nslots = end - start;
2044 map->has_ref_slots = has_ref_slots;
2045 map->has_pin_slots = has_pin_slots;
2046 map->has_ref_regs = has_ref_regs;
2047 map->has_pin_regs = has_pin_regs;
2048 g_assert (nregs < 32);
2049 map->used_int_regs = cfg->used_int_regs;
2050 map->reg_ref_mask = reg_ref_mask;
2051 map->reg_pin_mask = reg_pin_mask;
2052 map->nref_regs = nref_regs;
2053 map->npin_regs = npin_regs;
2055 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2058 if (has_ref_slots) {
2059 map->stack_ref_bitmap_offset = bitmaps_offset;
2060 bitmaps_offset += stack_bitmap_size;
2062 bitmap = &bitmaps [map->stack_ref_bitmap_offset];
2063 for (i = 0; i < nslots; ++i) {
2064 for (j = 0; j < ncallsites; ++j) {
2065 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
2066 set_bit (bitmap, stack_bitmap_width, j, i - start);
2070 if (has_pin_slots) {
2071 map->stack_pin_bitmap_offset = bitmaps_offset;
2072 bitmaps_offset += stack_bitmap_size;
2074 bitmap = &bitmaps [map->stack_pin_bitmap_offset];
2075 for (i = 0; i < nslots; ++i) {
2076 for (j = 0; j < ncallsites; ++j) {
2077 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
2078 set_bit (bitmap, stack_bitmap_width, j, i - start);
2083 map->reg_ref_bitmap_offset = bitmaps_offset;
2084 bitmaps_offset += reg_ref_bitmap_size;
2086 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2088 for (i = 0; i < nregs; ++i) {
2089 if (reg_ref_mask & (1 << i)) {
2090 for (j = 0; j < ncallsites; ++j) {
2091 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i))
2092 set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
2099 map->reg_pin_bitmap_offset = bitmaps_offset;
2100 bitmaps_offset += reg_pin_bitmap_size;
2102 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2104 for (i = 0; i < nregs; ++i) {
2105 if (reg_pin_mask & (1 << i)) {
2106 for (j = 0; j < ncallsites; ++j) {
2107 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i))
2108 set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
2116 map->ncallsites = ncallsites;
2117 if (cfg->code_len < 256)
2118 map->callsite_entry_size = 1;
2119 else if (cfg->code_len < 65536)
2120 map->callsite_entry_size = 2;
2122 map->callsite_entry_size = 4;
2124 /* Encode the GC Map */
2132 encode_gc_map (map, buf, &endbuf);
2133 g_assert (endbuf - buf < 256);
2135 encoded_size = endbuf - buf;
2136 alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
2138 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2139 //emap->ref_slots = map->ref_slots;
2141 /* Encoded fixed fields */
2142 p = &emap->encoded [0];
2143 //emap->encoded_size = encoded_size;
2144 memcpy (p, buf, encoded_size);
2147 /* Callsite table */
2148 p = (guint8*)ALIGN_TO ((guint64)p, map->callsite_entry_size);
2149 if (map->callsite_entry_size == 1) {
2150 guint8 *offsets = p;
2151 for (i = 0; i < ncallsites; ++i)
2152 offsets [i] = callsites [i]->pc_offset;
2153 stats.gc_callsites8_size += ncallsites * sizeof (guint8);
2154 } else if (map->callsite_entry_size == 2) {
2155 guint16 *offsets = (guint16*)p;
2156 for (i = 0; i < ncallsites; ++i)
2157 offsets [i] = callsites [i]->pc_offset;
2158 stats.gc_callsites16_size += ncallsites * sizeof (guint16);
2160 guint32 *offsets = (guint32*)p;
2161 for (i = 0; i < ncallsites; ++i)
2162 offsets [i] = callsites [i]->pc_offset;
2163 stats.gc_callsites32_size += ncallsites * sizeof (guint32);
2165 p += ncallsites * map->callsite_entry_size;
2168 memcpy (p, bitmaps, bitmaps_size);
2170 stats.gc_maps_size += alloc_size;
2171 stats.gc_callsites_size += ncallsites * map->callsite_entry_size;
2172 stats.gc_bitmaps_size += bitmaps_size;
2173 stats.gc_map_struct_size += sizeof (GCEncodedMap) + encoded_size;
2175 cfg->jit_info->gc_info = emap;
2178 stats.all_slots += nslots;
2179 stats.ref_slots += ntypes [SLOT_REF];
2180 stats.noref_slots += ntypes [SLOT_NOREF];
2181 stats.pin_slots += ntypes [SLOT_PIN];
2185 mini_gc_create_gc_map (MonoCompile *cfg)
2187 if (!cfg->compute_gc_maps)
2191 * During marking, all frames except the top frame are at a call site, and we mark the
2192 * top frame conservatively. This means that we only need to compute and record
2193 * GC maps for call sites.
2196 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2197 /* Without liveness info, the live ranges are not precise enough */
2200 mono_analyze_liveness_gc (cfg);
2202 compute_frame_size (cfg);
2206 process_spill_slots (cfg);
2207 process_other_slots (cfg);
2208 process_param_area_slots (cfg);
2209 process_variables (cfg);
2210 process_finally_clauses (cfg);
2220 memset (&cb, 0, sizeof (cb));
2221 cb.thread_attach_func = thread_attach_func;
2222 cb.thread_detach_func = thread_detach_func;
2223 cb.thread_suspend_func = thread_suspend_func;
2224 /* Comment this out to disable precise stack marking */
2225 cb.thread_mark_func = thread_mark_func;
2226 mono_gc_set_gc_callbacks (&cb);
2228 mono_counters_register ("GC Maps size",
2229 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
2230 mono_counters_register ("GC Call Sites size",
2231 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
2232 mono_counters_register ("GC Bitmaps size",
2233 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
2234 mono_counters_register ("GC Map struct size",
2235 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
2236 mono_counters_register ("GC Call Sites encoded using 8 bits",
2237 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
2238 mono_counters_register ("GC Call Sites encoded using 16 bits",
2239 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
2240 mono_counters_register ("GC Call Sites encoded using 32 bits",
2241 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
2243 mono_counters_register ("GC Map slots (all)",
2244 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
2245 mono_counters_register ("GC Map slots (ref)",
2246 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.ref_slots);
2247 mono_counters_register ("GC Map slots (noref)",
2248 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.noref_slots);
2249 mono_counters_register ("GC Map slots (pin)",
2250 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
2252 mono_counters_register ("GC TLS Data size",
2253 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2255 mono_counters_register ("Stack space scanned (all)",
2256 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
2257 mono_counters_register ("Stack space scanned (using GC Maps)",
2258 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
2259 mono_counters_register ("Stack space scanned (precise)",
2260 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_precisely);
2261 mono_counters_register ("Stack space scanned (pin)",
2262 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_conservatively);
2263 mono_counters_register ("Stack space scanned (pin registers)",
2264 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_registers);
2275 mini_gc_init_gc_map (MonoCompile *cfg)
2280 mini_gc_create_gc_map (MonoCompile *cfg)
2285 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2290 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2299 * Set GC specific options in CFG.
2302 mini_gc_init_cfg (MonoCompile *cfg)
2304 if (mono_gc_is_moving ()) {
2305 cfg->disable_ref_noref_stack_slot_share = TRUE;
2306 cfg->gen_write_barriers = TRUE;
2309 mini_gc_init_gc_map (cfg);
2313 * Problems with the current code:
2314 * - the stack walk is slow
2315 * - vtypes/refs used in EH regions are treated conservatively
2316 * - if the code is finished, less pinning will be done, causing problems because
2317 * we promote all surviving objects to old-gen.
2321 * Ideas for creating smaller GC maps:
2322 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2324 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2326 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2327 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2328 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2329 * from an index entry.