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>
15 #if defined(HAVE_SGEN_GC) && defined(MONO_ARCH_GC_MAPS_SUPPORTED)
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) ((((mgreg_t)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 * Contains information collected during the conservative stack marking pass,
88 * used during the precise pass. This helps to avoid doing a stack walk twice, which
94 int frame_start_offset;
96 /* Relative to stack_start */
97 int reg_locations [MONO_MAX_IREGS];
98 #ifdef DEBUG_PRECISE_ENABLED
101 int regs [MONO_MAX_IREGS];
105 /* Max number of frames stored in the TLS data */
106 #define MAX_FRAMES 50
109 * Per-thread data kept by this module. This is stored in the GC and passed to us as
110 * parameters, instead of being stored in a TLS variable, since during a collection,
111 * only the collection thread is active.
116 gboolean has_context;
117 MonoJitTlsData *jit_tls;
120 gpointer ref_to_track;
121 /* Number of frames collected during the !precise pass */
123 FrameInfo frames [MAX_FRAMES];
126 /* These are constant so don't store them in the GC Maps */
127 /* Number of registers stored in gc maps */
128 #define NREGS MONO_MAX_IREGS
132 * Contains information needed to mark a stack frame.
133 * This is a transient structure, created from a compressed representation on-demand.
137 * The offsets of the GC tracked area inside the stack frame relative to the frame pointer.
138 * This includes memory which is NOREF thus doesn't need GC maps.
143 * The offset relative to frame_offset where the the memory described by the GC maps
147 /* The number of stack slots in the map */
149 /* The frame pointer register */
151 /* The size of each callsite table entry */
152 guint8 callsite_entry_size;
153 guint has_pin_slots : 1;
154 guint has_ref_slots : 1;
155 guint has_ref_regs : 1;
156 guint has_pin_regs : 1;
158 /* The offsets below are into an external bitmaps array */
161 * A bitmap whose width is equal to bitmap_width, and whose height is equal to ncallsites.
162 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
165 guint32 stack_ref_bitmap_offset;
167 * Same for SLOT_PIN. It is possible that the same bit is set in both bitmaps at
168 * different callsites, if the slot starts out as PIN, and later changes to REF.
170 guint32 stack_pin_bitmap_offset;
173 * Corresponding bitmaps for registers
174 * These have width equal to the number of bits set in reg_ref_mask/reg_pin_mask.
175 * FIXME: Merge these with the normal bitmaps, i.e. reserve the first x slots for them ?
177 guint32 reg_pin_bitmap_offset;
178 guint32 reg_ref_bitmap_offset;
180 guint32 used_int_regs, reg_ref_mask, reg_pin_mask;
182 /* The number of bits set in the two masks above */
183 guint8 nref_regs, npin_regs;
186 * A bit array marking slots which contain refs.
187 * This is used only for debugging.
191 /* Callsite offsets */
192 /* These can take up a lot of space, so encode them compactly */
202 * A compressed version of GCMap. This is what gets stored in MonoJitInfo.
206 //guint8 encoded_size;
209 * The arrays below are embedded after the struct.
210 * Their address needs to be computed.
213 /* The fixed fields of the GCMap encoded using LEB128 */
214 guint8 encoded [MONO_ZERO_LEN_ARRAY];
216 /* An array of ncallsites entries, each entry is callsite_entry_size bytes long */
217 guint8 callsites [MONO_ZERO_LEN_ARRAY];
220 guint8 bitmaps [MONO_ZERO_LEN_ARRAY];
223 static int precise_frame_count [2], precise_frame_limit = -1;
224 static gboolean precise_frame_limit_inited;
230 int scanned_precisely;
231 int scanned_conservatively;
232 int scanned_registers;
242 int gc_callsites_size;
243 int gc_callsites8_size;
244 int gc_callsites16_size;
245 int gc_callsites32_size;
247 int gc_map_struct_size;
251 static JITGCStats stats;
253 static FILE *logfile;
255 // FIXME: Move these to a shared place
258 encode_uleb128 (guint32 value, guint8 *buf, guint8 **endbuf)
263 guint8 b = value & 0x7f;
265 if (value != 0) /* more bytes to come */
273 static G_GNUC_UNUSED void
274 encode_sleb128 (gint32 value, guint8 *buf, guint8 **endbuf)
277 gboolean negative = (value < 0);
285 /* the following is unnecessary if the
286 * implementation of >>= uses an arithmetic rather
287 * than logical shift for a signed left operand
291 value |= - (1 <<(size - 7));
292 /* sign bit of byte is second high order bit (0x40) */
293 if ((value == 0 && !(byte & 0x40)) ||
294 (value == -1 && (byte & 0x40)))
304 static inline guint32
305 decode_uleb128 (guint8 *buf, guint8 **endbuf)
315 res = res | (((int)(b & 0x7f)) << shift);
327 decode_sleb128 (guint8 *buf, guint8 **endbuf)
337 res = res | (((int)(b & 0x7f)) << shift);
340 if (shift < 32 && (b & 0x40))
341 res |= - (1 << shift);
352 encode_frame_reg (int frame_reg)
355 if (frame_reg == AMD64_RSP)
357 else if (frame_reg == AMD64_RBP)
359 #elif defined(TARGET_X86)
360 if (frame_reg == X86_EBP)
362 else if (frame_reg == X86_ESP)
367 g_assert_not_reached ();
372 decode_frame_reg (int encoded)
377 else if (encoded == 1)
379 #elif defined(TARGET_X86)
382 else if (encoded == 1)
387 g_assert_not_reached ();
393 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15, AMD64_RDI, AMD64_RSI };
395 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15 };
397 #elif defined(TARGET_X86)
398 static int callee_saved_regs [] = { X86_EBX, X86_ESI, X86_EDI };
402 encode_regmask (guint32 regmask)
408 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
409 if (regmask & (1 << callee_saved_regs [i])) {
411 regmask -= (1 << callee_saved_regs [i]);
414 g_assert (regmask == 0);
419 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]);
434 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
437 encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
441 encode_sleb128 (map->start_offset / sizeof (mgreg_t), buf, &buf);
442 encode_sleb128 (map->end_offset / sizeof (mgreg_t), buf, &buf);
443 encode_sleb128 (map->map_offset / sizeof (mgreg_t), buf, &buf);
444 encode_uleb128 (map->nslots, buf, &buf);
445 g_assert (map->callsite_entry_size <= 4);
446 freg = encode_frame_reg (map->frame_reg);
448 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);
449 encode_uleb128 (flags, buf, &buf);
450 encode_uleb128 (encode_regmask (map->used_int_regs), buf, &buf);
451 if (map->has_ref_regs)
452 encode_uleb128 (encode_regmask (map->reg_ref_mask), buf, &buf);
453 if (map->has_pin_regs)
454 encode_uleb128 (encode_regmask (map->reg_pin_mask), buf, &buf);
455 encode_uleb128 (map->ncallsites, buf, &buf);
463 * Decode the encoded GC map representation in BUF and store the result into MAP.
466 decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
469 int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
472 map->start_offset = decode_sleb128 (buf, &buf) * sizeof (mgreg_t);
473 map->end_offset = decode_sleb128 (buf, &buf) * sizeof (mgreg_t);
474 map->map_offset = decode_sleb128 (buf, &buf) * sizeof (mgreg_t);
475 map->nslots = decode_uleb128 (buf, &buf);
476 flags = decode_uleb128 (buf, &buf);
477 map->has_ref_slots = (flags & 1) ? 1 : 0;
478 map->has_pin_slots = (flags & 2) ? 1 : 0;
479 map->has_ref_regs = (flags & 4) ? 1 : 0;
480 map->has_pin_regs = (flags & 8) ? 1 : 0;
481 map->callsite_entry_size = ((flags >> 4) & 0x3) + 1;
483 map->frame_reg = decode_frame_reg (freg);
484 map->used_int_regs = decode_regmask (decode_uleb128 (buf, &buf));
485 if (map->has_ref_regs) {
486 map->reg_ref_mask = decode_regmask (decode_uleb128 (buf, &buf));
488 for (i = 0; i < NREGS; ++i)
489 if (map->reg_ref_mask & (1 << i))
493 if (map->has_pin_regs) {
494 map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
496 for (i = 0; i < NREGS; ++i)
497 if (map->reg_pin_mask & (1 << i))
501 map->ncallsites = decode_uleb128 (buf, &buf);
503 stack_bitmap_size = (ALIGN_TO (map->nslots, 8) / 8) * map->ncallsites;
504 reg_ref_bitmap_size = (ALIGN_TO (map->nref_regs, 8) / 8) * map->ncallsites;
505 reg_pin_bitmap_size = (ALIGN_TO (map->npin_regs, 8) / 8) * map->ncallsites;
507 map->stack_ref_bitmap_offset = offset;
508 if (map->has_ref_slots)
509 offset += stack_bitmap_size;
510 map->stack_pin_bitmap_offset = offset;
511 if (map->has_pin_slots)
512 offset += stack_bitmap_size;
513 map->reg_ref_bitmap_offset = offset;
514 if (map->has_ref_regs)
515 offset += reg_ref_bitmap_size;
516 map->reg_pin_bitmap_offset = offset;
517 if (map->has_pin_regs)
518 offset += reg_pin_bitmap_size;
524 thread_attach_func (void)
528 tls = g_new0 (TlsData, 1);
529 tls->tid = GetCurrentThreadId ();
530 stats.tlsdata_size += sizeof (TlsData);
536 thread_detach_func (gpointer user_data)
538 TlsData *tls = user_data;
544 thread_suspend_func (gpointer user_data, void *sigctx)
546 TlsData *tls = user_data;
549 /* Happens during startup */
552 tls->lmf = mono_get_lmf ();
554 mono_arch_sigctx_to_monoctx (sigctx, &tls->ctx);
555 tls->has_context = TRUE;
557 tls->has_context = FALSE;
559 tls->jit_tls = TlsGetValue (mono_jit_tls_id);
562 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
565 set_bit (guint8 *bitmap, int width, int y, int x)
567 bitmap [(width * y) + (x / 8)] |= (1 << (x % 8));
571 clear_bit (guint8 *bitmap, int width, int y, int x)
573 bitmap [(width * y) + (x / 8)] &= ~(1 << (x % 8));
577 get_bit (guint8 *bitmap, int width, int y, int x)
579 return bitmap [(width * y) + (x / 8)] & (1 << (x % 8));
583 slot_type_to_string (GCSlotType type)
593 g_assert_not_reached ();
598 static inline mgreg_t
599 get_frame_pointer (MonoContext *ctx, int frame_reg)
601 #if defined(TARGET_AMD64)
602 if (frame_reg == AMD64_RSP)
604 else if (frame_reg == AMD64_RBP)
606 #elif defined(TARGET_X86)
607 if (frame_reg == X86_ESP)
609 else if (frame_reg == X86_EBP)
612 g_assert_not_reached ();
617 * conservatively_pass:
619 * Mark a thread stack conservatively and collect information needed by the precise pass.
622 conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
625 MonoContext ctx, new_ctx;
628 gboolean last = TRUE;
632 guint8* fp, *p, *real_frame_start, *frame_start, *frame_end;
633 int i, pc_offset, cindex, bitmap_width;
634 int scanned = 0, scanned_precisely, scanned_conservatively, scanned_registers;
636 StackFrameInfo frame;
637 mgreg_t *reg_locations [MONO_MAX_IREGS];
638 mgreg_t *new_reg_locations [MONO_MAX_IREGS];
641 guint32 precise_regmask;
645 tls->ref_to_track = NULL;
648 /* tls == NULL can happen during startup */
649 if (mono_thread_internal_current () == NULL || !tls) {
650 mono_gc_conservatively_scan_area (stack_start, stack_end);
651 stats.scanned_stacks += stack_end - stack_start;
658 /* Number of bytes scanned based on GC map data */
660 /* Number of bytes scanned precisely based on GC map data */
661 scanned_precisely = 0;
662 /* Number of bytes scanned conservatively based on GC map data */
663 scanned_conservatively = 0;
664 /* Number of bytes scanned conservatively in register save areas */
665 scanned_registers = 0;
667 /* This is one past the last address which we have scanned */
668 stack_limit = stack_start;
670 if (!tls->has_context)
671 memset (&new_ctx, 0, sizeof (ctx));
673 memcpy (&new_ctx, &tls->ctx, sizeof (MonoContext));
675 memset (reg_locations, 0, sizeof (reg_locations));
676 memset (new_reg_locations, 0, sizeof (new_reg_locations));
679 memcpy (&ctx, &new_ctx, sizeof (ctx));
681 for (i = 0; i < MONO_MAX_IREGS; ++i) {
682 if (new_reg_locations [i]) {
684 * If the current frame saves the register, it means it might modify its
685 * value, thus the old location might not contain the same value, so
686 * we have to mark it conservatively.
688 if (reg_locations [i]) {
689 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
690 mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
691 scanned_registers += sizeof (mgreg_t);
694 reg_locations [i] = new_reg_locations [i];
696 DEBUG (fprintf (logfile, "\treg %s is now at location %p.\n", mono_arch_regname (i), reg_locations [i]));
700 g_assert ((mgreg_t)stack_limit % sizeof (mgreg_t) == 0);
702 #ifdef MONO_ARCH_HAVE_FIND_JIT_INFO_EXT
703 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);
712 if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
714 * These frames are problematic for several reasons:
715 * - they are unwound through an LMF, and we have no precise register tracking for those.
716 * - the LMF might not contain a precise ip, so we can't compute the call site.
717 * - the LMF only unwinds to the wrapper frame, so we get these methods twice.
719 DEBUG (fprintf (logfile, "Mark(0): <Managed-to-native transition>\n"));
720 for (i = 0; i < MONO_MAX_IREGS; ++i) {
721 if (reg_locations [i]) {
722 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
723 mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
724 scanned_registers += sizeof (mgreg_t);
726 reg_locations [i] = NULL;
727 new_reg_locations [i] = NULL;
732 /* The last frame can be in any state so mark conservatively */
735 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));
737 DEBUG (fprintf (logfile, "\t <Last frame>\n"));
742 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
744 /* These frames are very problematic */
745 if (ji->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
746 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));
747 DEBUG (fprintf (logfile, "\tSkip.\n"));
751 /* All the other frames are at a call site */
753 if (tls->nframes == MAX_FRAMES) {
755 * Can't save information since the array is full. So scan the rest of the
756 * stack conservatively.
758 DEBUG (fprintf (logfile, "Mark (0): Frame stack full.\n"));
762 /* Scan the frame of this method */
765 * A frame contains the following:
770 * - localloc-ed memory
772 g_assert (pc_offset >= 0);
777 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));
778 DEBUG (fprintf (logfile, "\tNo GC Map.\n"));
782 /* The embedded callsite table requires this */
783 g_assert (((mgreg_t)emap % 4) == 0);
786 * Debugging aid to control the number of frames scanned precisely
788 if (!precise_frame_limit_inited) {
789 if (getenv ("MONO_PRECISE_COUNT"))
790 precise_frame_limit = atoi (getenv ("MONO_PRECISE_COUNT"));
791 precise_frame_limit_inited = TRUE;
794 if (precise_frame_limit != -1) {
795 if (precise_frame_count [FALSE] == precise_frame_limit)
796 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (ji->method, TRUE));
797 if (precise_frame_count [FALSE] > precise_frame_limit)
800 precise_frame_count [FALSE] ++;
802 /* Decode the encoded GC map */
804 memset (map, 0, sizeof (GCMap));
805 decode_gc_map (&emap->encoded [0], map, &p);
806 p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
807 map->callsites.offsets8 = p;
808 p += map->callsite_entry_size * map->ncallsites;
811 fp = (guint8*)get_frame_pointer (&ctx, map->frame_reg);
813 real_frame_start = fp + map->start_offset;
814 frame_start = fp + map->start_offset + map->map_offset;
815 frame_end = fp + map->end_offset;
817 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));
819 /* Find the callsite index */
820 if (map->callsite_entry_size == 1) {
821 for (i = 0; i < map->ncallsites; ++i)
822 /* ip points inside the call instruction */
823 if (map->callsites.offsets8 [i] == pc_offset + 1)
825 } else if (map->callsite_entry_size == 2) {
826 // FIXME: Use a binary search
827 for (i = 0; i < map->ncallsites; ++i)
828 /* ip points inside the call instruction */
829 if (map->callsites.offsets16 [i] == pc_offset + 1)
832 // FIXME: Use a binary search
833 for (i = 0; i < map->ncallsites; ++i)
834 /* ip points inside the call instruction */
835 if (map->callsites.offsets32 [i] == pc_offset + 1)
838 if (i == map->ncallsites) {
839 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (ji->method, TRUE));
840 g_assert_not_reached ();
844 g_assert (real_frame_start >= stack_limit);
846 if (real_frame_start > stack_limit) {
847 /* This scans the previously skipped frames as well */
848 DEBUG (fprintf (logfile, "\tscan area %p-%p (%d).\n", stack_limit, real_frame_start, (int)(real_frame_start - stack_limit)));
849 mono_gc_conservatively_scan_area (stack_limit, real_frame_start);
850 stats.scanned_other += real_frame_start - stack_limit;
853 /* Mark stack slots */
854 if (map->has_pin_slots) {
855 int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
856 guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
861 for (i = 0; i < map->nslots; ++i) {
862 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
864 DEBUG (fprintf (logfile, "\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
865 mono_gc_conservatively_scan_area (p, p + sizeof (mgreg_t));
866 scanned_conservatively += sizeof (mgreg_t);
868 scanned_precisely += sizeof (mgreg_t);
870 p += sizeof (mgreg_t);
873 scanned_precisely += (map->nslots * sizeof (mgreg_t));
876 /* The area outside of start-end is NOREF */
877 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * sizeof (mgreg_t));
880 precise_regmask = map->used_int_regs | (1 << map->frame_reg);
881 if (map->has_pin_regs) {
882 int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
883 guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
885 for (i = 0; i < NREGS; ++i) {
886 if (!(map->used_int_regs & (1 << i)))
889 if (!(map->reg_pin_mask & (1 << i)))
892 if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
893 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
894 precise_regmask &= ~(1 << i);
900 scanned += map->end_offset - map->start_offset;
902 g_assert (scanned == scanned_precisely + scanned_conservatively);
904 stack_limit = frame_end;
906 /* Save information for the precise pass */
907 fi = &tls->frames [tls->nframes];
908 fi->nslots = map->nslots;
909 bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
910 if (map->has_ref_slots)
911 fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
914 fi->frame_start_offset = frame_start - stack_start;
915 fi->nreg_locations = 0;
916 DEBUG_PRECISE (fi->ji = ji);
917 DEBUG_PRECISE (fi->fp = fp);
919 if (map->has_ref_regs) {
920 int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
921 guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
923 for (i = 0; i < NREGS; ++i) {
924 if (!(map->reg_ref_mask & (1 << i)))
927 if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
928 DEBUG_PRECISE (fi->regs [fi->nreg_locations] = i);
929 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
930 fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
931 fi->nreg_locations ++;
938 * Clear locations of precisely stacked registers.
940 if (precise_regmask) {
941 for (i = 0; i < NREGS; ++i) {
942 if (precise_regmask & (1 << i)) {
944 * The method uses this register, and we have precise info for it.
945 * This means the location will be scanned precisely.
946 * Tell the code at the beginning of the loop that this location is
949 if (reg_locations [i])
950 DEBUG (fprintf (logfile, "\treg %s at location %p (==%p) is precise.\n", mono_arch_regname (i), reg_locations [i], (gpointer)*reg_locations [i]));
951 reg_locations [i] = NULL;
959 /* Scan the remaining register save locations */
960 for (i = 0; i < MONO_MAX_IREGS; ++i) {
961 if (reg_locations [i]) {
962 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", reg_locations [i]));
963 mono_gc_conservatively_scan_area (reg_locations [i], reg_locations [i] + sizeof (mgreg_t));
964 scanned_registers += sizeof (mgreg_t);
966 if (new_reg_locations [i]) {
967 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", new_reg_locations [i]));
968 mono_gc_conservatively_scan_area (new_reg_locations [i], new_reg_locations [i] + sizeof (mgreg_t));
969 scanned_registers += sizeof (mgreg_t);
973 if (stack_limit < stack_end) {
974 DEBUG (fprintf (logfile, "\tscan remaining stack %p-%p (%d).\n", stack_limit, stack_end, (int)(stack_end - stack_limit)));
975 mono_gc_conservatively_scan_area (stack_limit, stack_end);
976 stats.scanned_native += stack_end - stack_limit;
979 DEBUG (fprintf (logfile, "Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
981 stats.scanned_stacks += stack_end - stack_start;
982 stats.scanned += scanned;
983 stats.scanned_precisely += scanned_precisely;
984 stats.scanned_conservatively += scanned_conservatively;
985 stats.scanned_registers += scanned_registers;
987 //mono_gc_conservatively_scan_area (stack_start, stack_end);
993 * Mark a thread stack precisely based on information saved during the conservative
997 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
1001 guint8 *frame_start;
1006 for (findex = 0; findex < tls->nframes; findex ++) {
1007 /* Load information saved by the !precise pass */
1008 fi = &tls->frames [findex];
1009 frame_start = stack_start + fi->frame_start_offset;
1011 DEBUG (char *fname = mono_method_full_name (fi->ji->method, TRUE); fprintf (logfile, "Mark(1): %s\n", fname); g_free (fname));
1014 * FIXME: Add a function to mark using a bitmap, to avoid doing a
1015 * call for each object.
1018 /* Mark stack slots */
1020 guint8 *ref_bitmap = fi->bitmap;
1023 for (i = 0; i < fi->nslots; ++i) {
1024 MonoObject **ptr = (MonoObject**)(frame_start + (i * sizeof (mgreg_t)));
1026 live = ref_bitmap [i / 8] & (1 << (i % 8));
1029 MonoObject *obj = *ptr;
1031 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1032 *ptr = mono_gc_scan_object (obj);
1033 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1035 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1040 * This is disabled because the pointer takes up a lot of space.
1041 * Stack slots might be shared between ref and non-ref variables ?
1043 if (map->ref_slots [i / 8] & (1 << (i % 8))) {
1044 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));
1046 * Fail fast if the live range is incorrect, and
1047 * the JITted code tries to access this object
1056 /* Mark registers */
1059 * Registers are different from stack slots, they have no address where they
1060 * are stored. Instead, some frame below this frame in the stack saves them
1061 * in its prolog to the stack. We can mark this location precisely.
1063 for (i = 0; i < fi->nreg_locations; ++i) {
1065 * reg_locations [i] contains the address of the stack slot where
1066 * a reg was last saved, so mark that slot.
1068 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1069 MonoObject *obj = *ptr;
1072 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
1073 *ptr = mono_gc_scan_object (obj);
1074 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1076 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p\n", mono_arch_regname (fi->regs [i]), ptr, obj));
1082 * Debugging aid to check for missed refs.
1084 if (tls->ref_to_track) {
1087 for (p = (mgreg_t*)stack_start; p < (mgreg_t*)stack_end; ++p)
1088 if (*p == (mgreg_t)tls->ref_to_track)
1089 printf ("REF AT %p.\n", p);
1096 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1097 * call, and TRUE at the second. USER_DATA points to a TlsData
1098 * structure filled up by thread_suspend_func.
1101 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise)
1103 TlsData *tls = user_data;
1105 DEBUG (fprintf (logfile, "****************************************\n"));
1106 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));
1107 DEBUG (fprintf (logfile, "****************************************\n"));
1110 conservative_pass (tls, stack_start, stack_end);
1112 precise_pass (tls, stack_start, stack_end);
1116 mini_gc_init_gc_map (MonoCompile *cfg)
1118 if (COMPILE_LLVM (cfg))
1122 /* Debugging support */
1124 static int precise_count;
1127 if (getenv ("MONO_GCMAP_COUNT")) {
1128 if (precise_count == atoi (getenv ("MONO_GCMAP_COUNT")))
1129 printf ("LAST: %s\n", mono_method_full_name (cfg->method, TRUE));
1130 if (precise_count > atoi (getenv ("MONO_GCMAP_COUNT")))
1136 cfg->compute_gc_maps = TRUE;
1138 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1142 * mini_gc_set_slot_type_from_fp:
1144 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1145 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1146 * need to call this function for those slots.
1149 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1151 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1153 if (!cfg->compute_gc_maps)
1156 g_assert (slot_offset % sizeof (mgreg_t) == 0);
1158 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1162 * mini_gc_set_slot_type_from_cfa:
1164 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1165 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1166 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1167 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1168 * call this function for those slots.
1171 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1173 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1174 int slot = - (slot_offset / sizeof (mgreg_t));
1176 if (!cfg->compute_gc_maps)
1179 g_assert (slot_offset <= 0);
1180 g_assert (slot_offset % sizeof (mgreg_t) == 0);
1182 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1186 fp_offset_to_slot (MonoCompile *cfg, int offset)
1188 MonoCompileGC *gcfg = cfg->gc_info;
1190 return (offset - gcfg->min_offset) / sizeof (mgreg_t);
1194 slot_to_fp_offset (MonoCompile *cfg, int slot)
1196 MonoCompileGC *gcfg = cfg->gc_info;
1198 return (slot * sizeof (mgreg_t)) + gcfg->min_offset;
1202 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1204 g_assert (slot >= 0 && slot < gcfg->nslots);
1206 if (type == SLOT_PIN) {
1207 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1208 set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1209 } else if (type == SLOT_REF) {
1210 set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1211 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1212 } else if (type == SLOT_NOREF) {
1213 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1214 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1219 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1223 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1224 set_slot (gcfg, slot, cindex, type);
1228 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1232 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1233 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1234 if (callsite_offset >= from && callsite_offset < to)
1235 set_slot (gcfg, slot, cindex, type);
1240 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1242 g_assert (slot >= 0 && slot < gcfg->nregs);
1244 if (type == SLOT_PIN) {
1245 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1246 set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1247 } else if (type == SLOT_REF) {
1248 set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1249 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1250 } else if (type == SLOT_NOREF) {
1251 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1252 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1257 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1261 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1262 set_reg_slot (gcfg, slot, cindex, type);
1266 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1270 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1271 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1272 if (callsite_offset >= from && callsite_offset < to)
1273 set_reg_slot (gcfg, slot, cindex, type);
1278 process_spill_slots (MonoCompile *cfg)
1280 MonoCompileGC *gcfg = cfg->gc_info;
1285 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1286 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1287 for (l = bb->spill_slot_defs; l; l = l->next) {
1288 MonoInst *def = l->data;
1289 int spill_slot = def->inst_c0;
1290 int bank = def->inst_c1;
1291 int offset = cfg->spill_info [bank][spill_slot].offset;
1292 int slot = fp_offset_to_slot (cfg, offset);
1294 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1295 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1299 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1300 for (l = bb->spill_slot_defs; l; l = l->next) {
1301 MonoInst *def = l->data;
1302 int spill_slot = def->inst_c0;
1303 int bank = def->inst_c1;
1304 int offset = cfg->spill_info [bank][spill_slot].offset;
1305 int slot = fp_offset_to_slot (cfg, offset);
1308 if (bank == MONO_REG_INT_MP)
1314 * Extend the live interval for the GC tracked spill slots
1315 * defined in this bblock.
1316 * FIXME: This is not needed.
1318 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1320 if (cfg->verbose_level > 1)
1321 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
1325 /* Set fp spill slots to NOREF */
1326 for (i = 0; i < cfg->spill_info_len [MONO_REG_DOUBLE]; ++i) {
1327 int offset = cfg->spill_info [MONO_REG_DOUBLE][i].offset;
1333 slot = fp_offset_to_slot (cfg, offset);
1335 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1337 if (cfg->verbose_level > 1)
1338 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1341 /* Set int spill slots to NOREF */
1342 for (i = 0; i < cfg->spill_info_len [MONO_REG_INT]; ++i) {
1343 int offset = cfg->spill_info [MONO_REG_INT][i].offset;
1349 slot = fp_offset_to_slot (cfg, offset);
1351 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1352 if (cfg->verbose_level > 1)
1353 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1358 * process_other_slots:
1360 * Process stack slots registered using mini_gc_set_slot_type_... ().
1363 process_other_slots (MonoCompile *cfg)
1365 MonoCompileGC *gcfg = cfg->gc_info;
1368 /* Relative to the CFA */
1369 for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
1370 guint data = GPOINTER_TO_UINT (l->data);
1371 int cfa_slot = data >> 16;
1372 GCSlotType type = data & 0xff;
1376 * Map the cfa relative slot to an fp relative slot.
1377 * slot_addr == cfa - <cfa_slot>*4/8
1378 * fp + cfa_offset == cfa
1379 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1381 slot = (cfg->cfa_offset / sizeof (mgreg_t)) - cfa_slot - (gcfg->min_offset / sizeof (mgreg_t));
1383 set_slot_everywhere (gcfg, slot, type);
1385 if (cfg->verbose_level > 1) {
1386 int fp_offset = slot_to_fp_offset (cfg, slot);
1387 if (type == SLOT_NOREF)
1388 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)));
1392 /* Relative to the FP */
1393 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1394 gint data = GPOINTER_TO_INT (l->data);
1395 int offset = data >> 16;
1396 GCSlotType type = data & 0xff;
1399 slot = fp_offset_to_slot (cfg, offset);
1401 set_slot_everywhere (gcfg, slot, type);
1403 /* Liveness for these slots is handled by process_spill_slots () */
1405 if (cfg->verbose_level > 1) {
1406 if (type == SLOT_REF)
1407 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
1413 process_variables (MonoCompile *cfg)
1415 MonoCompileGC *gcfg = cfg->gc_info;
1416 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1417 int i, locals_min_slot, locals_max_slot, cindex;
1421 int locals_min_offset = gcfg->locals_min_offset;
1422 int locals_max_offset = gcfg->locals_max_offset;
1424 /* Slots for locals are NOREF by default */
1425 locals_min_slot = (locals_min_offset - gcfg->min_offset) / sizeof (mgreg_t);
1426 locals_max_slot = (locals_max_offset - gcfg->min_offset) / sizeof (mgreg_t);
1427 for (i = locals_min_slot; i < locals_max_slot; ++i) {
1428 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1432 * Compute the offset where variables are initialized in the first bblock, if any.
1434 pc_offsets = g_new0 (int, cfg->next_vreg);
1436 bb = cfg->bb_entry->next_bb;
1437 MONO_BB_FOR_EACH_INS (bb, tmp) {
1438 if (tmp->opcode == OP_GC_LIVENESS_DEF) {
1439 int vreg = tmp->inst_c1;
1440 if (pc_offsets [vreg] == 0) {
1441 g_assert (tmp->backend.pc_offset > 0);
1442 pc_offsets [vreg] = tmp->backend.pc_offset;
1448 * Stack slots holding arguments are initialized in the prolog.
1449 * This means we can treat them alive for the whole method.
1451 for (i = 0; i < cfg->num_varinfo; i++) {
1452 MonoInst *ins = cfg->varinfo [i];
1453 MonoType *t = ins->inst_vtype;
1456 gboolean byref, is_this = FALSE;
1457 gboolean is_arg = i < cfg->locals_start;
1459 if (ins == cfg->ret)
1462 vmv = MONO_VARINFO (cfg, i);
1464 /* For some reason, 'this' is byref */
1465 if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
1466 t = &cfg->method->klass->byval_arg;
1472 if (ins->opcode == OP_REGVAR) {
1474 GCSlotType slot_type;
1476 t = mini_type_get_underlying_type (NULL, t);
1479 g_assert (hreg < MONO_MAX_IREGS);
1482 slot_type = SLOT_PIN;
1484 slot_type = MONO_TYPE_IS_REFERENCE (t) ? SLOT_REF : SLOT_NOREF;
1486 if (slot_type == SLOT_PIN) {
1487 /* These have no live interval, be conservative */
1488 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1491 * Unlike variables allocated to the stack, we generate liveness info
1492 * for noref vars in registers in mono_spill_global_vars (), because
1493 * knowing that a register doesn't contain a ref allows us to mark its save
1494 * locations precisely.
1496 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1497 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1498 set_reg_slot (gcfg, hreg, cindex, slot_type);
1501 if (cfg->verbose_level > 1) {
1502 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
1508 if (ins->opcode != OP_REGOFFSET)
1511 if (ins->inst_offset % sizeof (mgreg_t) != 0)
1514 if (is_arg && ins->inst_offset >= gcfg->max_offset)
1515 /* In parent frame */
1518 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1520 if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
1521 /* These do not get stored in the prolog */
1522 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1524 if (cfg->verbose_level > 1) {
1525 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));
1530 if (MONO_TYPE_ISSTRUCT (t)) {
1532 gsize *bitmap = NULL;
1533 gboolean pin = FALSE;
1537 if (ins->backend.is_pinvoke)
1538 size = mono_class_native_size (ins->klass, NULL);
1540 size = mono_class_value_size (ins->klass, NULL);
1541 size_in_slots = ALIGN_TO (size, sizeof (mgreg_t)) / sizeof (mgreg_t);
1543 if (!ins->klass->has_references) {
1545 for (j = 0; j < size_in_slots; ++j)
1546 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1551 if (ins->klass->generic_container || mono_class_is_open_constructed_type (t)) {
1552 /* FIXME: Generic sharing */
1555 mono_class_compute_gc_descriptor (ins->klass);
1557 bitmap = mono_gc_get_bitmap_for_descr (ins->klass->gc_descr, &numbits);
1561 * Most vtypes are marked volatile because of the LDADDR instructions,
1562 * and they have no liveness information since they are decomposed
1563 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1564 * them during VZERO decomposition.
1566 if (!pc_offsets [vmv->vreg])
1570 if (ins->backend.is_pinvoke)
1573 if (cfg->verbose_level > 1)
1574 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));
1577 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1578 if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
1579 for (j = 0; j < numbits; ++j) {
1580 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1581 /* The descriptor is for the boxed object */
1582 set_slot (gcfg, (pos + j - (sizeof (MonoObject) / sizeof (mgreg_t))), cindex, pin ? SLOT_PIN : SLOT_REF);
1588 if (cfg->verbose_level > 1) {
1589 for (j = 0; j < numbits; ++j) {
1590 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD)))
1591 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))));
1595 if (cfg->verbose_level > 1)
1596 printf ("\t\tpinned\n");
1597 for (j = 0; j < size_in_slots; ++j) {
1598 set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
1607 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1608 /* Vret addr etc. */
1613 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1615 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1616 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1617 set_slot (gcfg, pos, cindex, SLOT_PIN);
1619 if (cfg->verbose_level > 1)
1620 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));
1625 * This is currently disabled, but could be enabled to debug crashes.
1628 if (t->type == MONO_TYPE_I) {
1630 * Variables created in mono_handle_global_vregs have type I, but they
1631 * could hold GC refs since the vregs they were created from might not been
1632 * marked as holding a GC ref. So be conservative.
1634 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1639 t = mini_type_get_underlying_type (NULL, t);
1641 if (!MONO_TYPE_IS_REFERENCE (t)) {
1642 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1643 if (cfg->verbose_level > 1)
1644 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));
1648 /* 'this' is marked INDIRECT for gshared methods */
1649 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1651 * For volatile variables, treat them alive from the point they are
1652 * initialized in the first bblock until the end of the method.
1655 set_slot_everywhere (gcfg, pos, SLOT_REF);
1656 } else if (pc_offsets [vmv->vreg]) {
1657 set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
1658 set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
1660 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1662 if (cfg->verbose_level > 1)
1663 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));
1668 /* Live for the whole method */
1669 set_slot_everywhere (gcfg, pos, SLOT_REF);
1671 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1672 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1673 set_slot (gcfg, pos, cindex, SLOT_REF);
1676 if (cfg->verbose_level > 1) {
1677 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));
1681 g_free (pc_offsets);
1685 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1688 * Convert a sp relative offset to a slot index. This is
1689 * platform specific.
1692 /* fp = sp + offset */
1693 g_assert (cfg->frame_reg == AMD64_RBP);
1694 return (- cfg->arch.sp_fp_offset + sp_offset);
1702 type_to_gc_slot_type (MonoType *t)
1706 t = mini_type_get_underlying_type (NULL, t);
1707 if (MONO_TYPE_IS_REFERENCE (t))
1710 if (MONO_TYPE_ISSTRUCT (t)) {
1711 MonoClass *klass = mono_class_from_mono_type (t);
1712 if (!klass->has_references) {
1724 process_param_area_slots (MonoCompile *cfg)
1726 MonoCompileGC *gcfg = cfg->gc_info;
1731 * These slots are used for passing parameters during calls. They are sp relative, not
1732 * fp relative, so they are harder to handle.
1734 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1735 /* The distance between fp and sp is not constant */
1738 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1740 for (i = 0; i < gcfg->ncallsites; ++i) {
1741 GCCallSite *callsite = gcfg->callsites [i];
1744 for (l = callsite->param_slots; l; l = l->next) {
1745 MonoInst *def = l->data;
1746 int sp_offset = def->inst_offset;
1747 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1748 int slot = fp_offset_to_slot (cfg, fp_offset);
1750 g_assert (slot >= 0 && slot < gcfg->nslots);
1751 is_param [slot] = TRUE;
1755 /* All param area slots are noref by default */
1756 for (i = 0; i < gcfg->nslots; ++i) {
1758 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1761 for (i = 0; i < gcfg->ncallsites; ++i) {
1762 GCCallSite *callsite = gcfg->callsites [i];
1765 for (l = callsite->param_slots; l; l = l->next) {
1766 MonoInst *def = l->data;
1767 MonoType *t = def->inst_vtype;
1768 int sp_offset = def->inst_offset;
1769 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1770 int slot = fp_offset_to_slot (cfg, fp_offset);
1771 GCSlotType type = type_to_gc_slot_type (t);
1773 /* The slot is live between the def instruction and the call */
1774 set_slot_in_range (gcfg, slot, def->backend.pc_offset, callsite->pc_offset + 1, type);
1775 if (cfg->verbose_level > 1)
1776 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);
1782 process_finally_clauses (MonoCompile *cfg)
1784 MonoCompileGC *gcfg = cfg->gc_info;
1785 GCCallSite **callsites;
1787 gboolean has_finally;
1788 int i, j, nslots, nregs;
1790 ncallsites = gcfg->ncallsites;
1791 nslots = gcfg->nslots;
1792 nregs = gcfg->nregs;
1793 callsites = gcfg->callsites;
1796 * The calls to the finally clauses don't show up in the cfg. See
1797 * test_0_liveness_8 ().
1798 * Variables accessed inside the finally clause are already marked VOLATILE by
1799 * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
1800 * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
1802 has_finally = FALSE;
1803 for (i = 0; i < cfg->header->num_clauses; ++i) {
1804 MonoExceptionClause *clause = &cfg->header->clauses [i];
1806 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1811 if (cfg->verbose_level > 1)
1812 printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
1813 for (j = 0; j < ncallsites; ++j) {
1814 MonoBasicBlock *bb = callsites [j]->bb;
1815 MonoExceptionClause *clause;
1816 gboolean is_in_finally = FALSE;
1818 for (i = 0; i < cfg->header->num_clauses; ++i) {
1819 clause = &cfg->header->clauses [i];
1821 if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
1822 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1823 is_in_finally = TRUE;
1829 if (is_in_finally) {
1830 for (i = 0; i < nslots; ++i)
1831 set_slot (gcfg, i, j, SLOT_PIN);
1832 for (i = 0; i < nregs; ++i)
1833 set_reg_slot (gcfg, i, j, SLOT_PIN);
1840 compute_frame_size (MonoCompile *cfg)
1842 int i, locals_min_offset, locals_max_offset, cfa_min_offset, cfa_max_offset;
1843 int min_offset, max_offset;
1844 MonoCompileGC *gcfg = cfg->gc_info;
1845 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1848 /* Compute min/max offsets from the fp */
1851 #if defined(TARGET_AMD64) || defined(TARGET_X86)
1852 locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, sizeof (mgreg_t));
1853 locals_max_offset = cfg->locals_max_stack_offset;
1855 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
1859 locals_min_offset = ALIGN_TO (locals_min_offset, sizeof (mgreg_t));
1860 locals_max_offset = ALIGN_TO (locals_max_offset, sizeof (mgreg_t));
1862 min_offset = locals_min_offset;
1863 max_offset = locals_max_offset;
1866 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1867 MonoInst *ins = cfg->args [i];
1869 if (ins->opcode == OP_REGOFFSET)
1870 min_offset = MIN (min_offset, ins->inst_offset);
1874 g_assert (cfg->frame_reg == cfg->cfa_reg);
1875 g_assert (cfg->cfa_offset > 0);
1877 cfa_max_offset = cfg->cfa_offset;
1879 min_offset = MIN (min_offset, cfa_min_offset);
1880 max_offset = MAX (max_offset, cfa_max_offset);
1882 /* Fp relative slots */
1883 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1884 gint data = GPOINTER_TO_INT (l->data);
1885 int offset = data >> 16;
1887 min_offset = MIN (min_offset, offset);
1891 if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
1892 int stack_offset = ALIGN_TO (cfg->stack_offset, sizeof (mgreg_t));
1893 min_offset = MIN (min_offset, (-stack_offset));
1896 /* Param area slots */
1898 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
1901 gcfg->min_offset = min_offset;
1902 gcfg->max_offset = max_offset;
1903 gcfg->locals_min_offset = locals_min_offset;
1904 gcfg->locals_max_offset = locals_max_offset;
1908 init_gcfg (MonoCompile *cfg)
1910 int i, nregs, nslots;
1911 MonoCompileGC *gcfg = cfg->gc_info;
1912 GCCallSite **callsites;
1921 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1922 ncallsites += g_slist_length (bb->gc_callsites);
1924 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
1926 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1927 for (l = bb->gc_callsites; l; l = l->next)
1928 callsites [i++] = l->data;
1931 /* The callsites should already be ordered by pc offset */
1932 for (i = 1; i < ncallsites; ++i)
1933 g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
1936 * The stack frame looks like this:
1938 * <fp + max_offset> == cfa -> <end of previous frame>
1939 * <other stack slots>
1941 * <other stack slots>
1942 * fp + min_offset ->
1947 if (cfg->verbose_level > 1)
1948 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
1950 nslots = (gcfg->max_offset - gcfg->min_offset) / sizeof (mgreg_t);
1953 gcfg->nslots = nslots;
1954 gcfg->nregs = nregs;
1955 gcfg->callsites = callsites;
1956 gcfg->ncallsites = ncallsites;
1957 gcfg->stack_bitmap_width = ALIGN_TO (nslots, 8) / 8;
1958 gcfg->reg_bitmap_width = ALIGN_TO (nregs, 8) / 8;
1959 gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
1960 gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
1961 gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
1962 gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
1964 /* All slots start out as PIN */
1965 memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * ncallsites);
1966 for (i = 0; i < nregs; ++i) {
1968 * By default, registers are NOREF.
1969 * It is possible for a callee to save them before being defined in this method,
1970 * but the saved value is dead too, so it doesn't need to be marked.
1972 if ((cfg->used_int_regs & (1 << i)))
1973 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
1978 create_map (MonoCompile *cfg)
1981 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
1983 int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
1984 int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
1986 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
1987 MonoCompileGC *gcfg = cfg->gc_info;
1988 GCCallSite **callsites;
1990 guint8 *bitmap, *bitmaps;
1991 guint32 reg_ref_mask, reg_pin_mask;
1993 ncallsites = gcfg->ncallsites;
1994 nslots = gcfg->nslots;
1995 nregs = gcfg->nregs;
1996 callsites = gcfg->callsites;
1999 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
2000 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
2002 has_ref_slots = FALSE;
2003 has_pin_slots = FALSE;
2006 memset (ntypes, 0, sizeof (ntypes));
2007 for (i = 0; i < nslots; ++i) {
2008 gboolean has_ref = FALSE;
2009 gboolean has_pin = FALSE;
2011 for (j = 0; j < ncallsites; ++j) {
2012 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
2014 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
2019 has_ref_slots = TRUE;
2021 has_pin_slots = TRUE;
2024 ntypes [SLOT_REF] ++;
2026 ntypes [SLOT_PIN] ++;
2028 ntypes [SLOT_NOREF] ++;
2030 if (has_ref || has_pin) {
2037 start = end = nslots;
2039 g_assert (start != -1);
2040 g_assert (start < end);
2043 has_ref_regs = FALSE;
2044 has_pin_regs = FALSE;
2049 for (i = 0; i < nregs; ++i) {
2050 gboolean has_ref = FALSE;
2051 gboolean has_pin = FALSE;
2053 if (!(cfg->used_int_regs & (1 << i)))
2056 for (j = 0; j < ncallsites; ++j) {
2057 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i)) {
2062 for (j = 0; j < ncallsites; ++j) {
2063 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i)) {
2070 reg_ref_mask |= (1 << i);
2071 has_ref_regs = TRUE;
2075 reg_pin_mask |= (1 << i);
2076 has_pin_regs = TRUE;
2081 if (cfg->verbose_level > 1)
2082 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);
2084 /* Create the GC Map */
2086 stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
2087 stack_bitmap_size = stack_bitmap_width * ncallsites;
2088 reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
2089 reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
2090 reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
2091 reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
2092 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);
2094 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2096 map->frame_reg = cfg->frame_reg;
2097 map->start_offset = gcfg->min_offset;
2098 map->end_offset = gcfg->min_offset + (nslots * sizeof (mgreg_t));
2099 map->map_offset = start * sizeof (mgreg_t);
2100 map->nslots = end - start;
2101 map->has_ref_slots = has_ref_slots;
2102 map->has_pin_slots = has_pin_slots;
2103 map->has_ref_regs = has_ref_regs;
2104 map->has_pin_regs = has_pin_regs;
2105 g_assert (nregs < 32);
2106 map->used_int_regs = cfg->used_int_regs;
2107 map->reg_ref_mask = reg_ref_mask;
2108 map->reg_pin_mask = reg_pin_mask;
2109 map->nref_regs = nref_regs;
2110 map->npin_regs = npin_regs;
2112 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2115 if (has_ref_slots) {
2116 map->stack_ref_bitmap_offset = bitmaps_offset;
2117 bitmaps_offset += stack_bitmap_size;
2119 bitmap = &bitmaps [map->stack_ref_bitmap_offset];
2120 for (i = 0; i < nslots; ++i) {
2121 for (j = 0; j < ncallsites; ++j) {
2122 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
2123 set_bit (bitmap, stack_bitmap_width, j, i - start);
2127 if (has_pin_slots) {
2128 map->stack_pin_bitmap_offset = bitmaps_offset;
2129 bitmaps_offset += stack_bitmap_size;
2131 bitmap = &bitmaps [map->stack_pin_bitmap_offset];
2132 for (i = 0; i < nslots; ++i) {
2133 for (j = 0; j < ncallsites; ++j) {
2134 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
2135 set_bit (bitmap, stack_bitmap_width, j, i - start);
2140 map->reg_ref_bitmap_offset = bitmaps_offset;
2141 bitmaps_offset += reg_ref_bitmap_size;
2143 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2145 for (i = 0; i < nregs; ++i) {
2146 if (reg_ref_mask & (1 << i)) {
2147 for (j = 0; j < ncallsites; ++j) {
2148 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i))
2149 set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
2156 map->reg_pin_bitmap_offset = bitmaps_offset;
2157 bitmaps_offset += reg_pin_bitmap_size;
2159 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2161 for (i = 0; i < nregs; ++i) {
2162 if (reg_pin_mask & (1 << i)) {
2163 for (j = 0; j < ncallsites; ++j) {
2164 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i))
2165 set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
2173 map->ncallsites = ncallsites;
2174 if (cfg->code_len < 256)
2175 map->callsite_entry_size = 1;
2176 else if (cfg->code_len < 65536)
2177 map->callsite_entry_size = 2;
2179 map->callsite_entry_size = 4;
2181 /* Encode the GC Map */
2189 encode_gc_map (map, buf, &endbuf);
2190 g_assert (endbuf - buf < 256);
2192 encoded_size = endbuf - buf;
2193 alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
2195 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2196 //emap->ref_slots = map->ref_slots;
2198 /* Encoded fixed fields */
2199 p = &emap->encoded [0];
2200 //emap->encoded_size = encoded_size;
2201 memcpy (p, buf, encoded_size);
2204 /* Callsite table */
2205 p = (guint8*)ALIGN_TO ((mgreg_t)p, map->callsite_entry_size);
2206 if (map->callsite_entry_size == 1) {
2207 guint8 *offsets = p;
2208 for (i = 0; i < ncallsites; ++i)
2209 offsets [i] = callsites [i]->pc_offset;
2210 stats.gc_callsites8_size += ncallsites * sizeof (guint8);
2211 } else if (map->callsite_entry_size == 2) {
2212 guint16 *offsets = (guint16*)p;
2213 for (i = 0; i < ncallsites; ++i)
2214 offsets [i] = callsites [i]->pc_offset;
2215 stats.gc_callsites16_size += ncallsites * sizeof (guint16);
2217 guint32 *offsets = (guint32*)p;
2218 for (i = 0; i < ncallsites; ++i)
2219 offsets [i] = callsites [i]->pc_offset;
2220 stats.gc_callsites32_size += ncallsites * sizeof (guint32);
2222 p += ncallsites * map->callsite_entry_size;
2225 memcpy (p, bitmaps, bitmaps_size);
2228 g_assert ((guint8*)p - (guint8*)emap <= alloc_size);
2230 stats.gc_maps_size += alloc_size;
2231 stats.gc_callsites_size += ncallsites * map->callsite_entry_size;
2232 stats.gc_bitmaps_size += bitmaps_size;
2233 stats.gc_map_struct_size += sizeof (GCEncodedMap) + encoded_size;
2235 cfg->jit_info->gc_info = emap;
2237 cfg->gc_map = (guint8*)emap;
2238 cfg->gc_map_size = alloc_size;
2241 stats.all_slots += nslots;
2242 stats.ref_slots += ntypes [SLOT_REF];
2243 stats.noref_slots += ntypes [SLOT_NOREF];
2244 stats.pin_slots += ntypes [SLOT_PIN];
2248 mini_gc_create_gc_map (MonoCompile *cfg)
2250 if (!cfg->compute_gc_maps)
2254 * During marking, all frames except the top frame are at a call site, and we mark the
2255 * top frame conservatively. This means that we only need to compute and record
2256 * GC maps for call sites.
2259 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2260 /* Without liveness info, the live ranges are not precise enough */
2263 mono_analyze_liveness_gc (cfg);
2265 compute_frame_size (cfg);
2269 process_spill_slots (cfg);
2270 process_other_slots (cfg);
2271 process_param_area_slots (cfg);
2272 process_variables (cfg);
2273 process_finally_clauses (cfg);
2279 parse_debug_options (void)
2284 env = getenv ("MONO_GCMAP_DEBUG");
2288 opts = g_strsplit (env, ",", -1);
2289 for (ptr = opts; ptr && *ptr; ptr ++) {
2290 /* No options yet */
2291 fprintf (stderr, "Invalid format for the MONO_GCMAP_DEBUG env variable: '%s'\n", env);
2302 memset (&cb, 0, sizeof (cb));
2303 cb.thread_attach_func = thread_attach_func;
2304 cb.thread_detach_func = thread_detach_func;
2305 cb.thread_suspend_func = thread_suspend_func;
2306 /* Comment this out to disable precise stack marking */
2307 cb.thread_mark_func = thread_mark_func;
2308 mono_gc_set_gc_callbacks (&cb);
2310 logfile = mono_sgen_get_logfile ();
2312 parse_debug_options ();
2314 mono_counters_register ("GC Maps size",
2315 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
2316 mono_counters_register ("GC Call Sites size",
2317 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
2318 mono_counters_register ("GC Bitmaps size",
2319 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
2320 mono_counters_register ("GC Map struct size",
2321 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
2322 mono_counters_register ("GC Call Sites encoded using 8 bits",
2323 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
2324 mono_counters_register ("GC Call Sites encoded using 16 bits",
2325 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
2326 mono_counters_register ("GC Call Sites encoded using 32 bits",
2327 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
2329 mono_counters_register ("GC Map slots (all)",
2330 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
2331 mono_counters_register ("GC Map slots (ref)",
2332 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.ref_slots);
2333 mono_counters_register ("GC Map slots (noref)",
2334 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.noref_slots);
2335 mono_counters_register ("GC Map slots (pin)",
2336 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
2338 mono_counters_register ("GC TLS Data size",
2339 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2341 mono_counters_register ("Stack space scanned (all)",
2342 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
2343 mono_counters_register ("Stack space scanned (native)",
2344 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_native);
2345 mono_counters_register ("Stack space scanned (other)",
2346 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_other);
2347 mono_counters_register ("Stack space scanned (using GC Maps)",
2348 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
2349 mono_counters_register ("Stack space scanned (precise)",
2350 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_precisely);
2351 mono_counters_register ("Stack space scanned (pin)",
2352 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_conservatively);
2353 mono_counters_register ("Stack space scanned (pin registers)",
2354 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_registers);
2365 mini_gc_init_gc_map (MonoCompile *cfg)
2370 mini_gc_create_gc_map (MonoCompile *cfg)
2375 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2380 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2389 * Set GC specific options in CFG.
2392 mini_gc_init_cfg (MonoCompile *cfg)
2394 if (mono_gc_is_moving ()) {
2395 cfg->disable_ref_noref_stack_slot_share = TRUE;
2396 cfg->gen_write_barriers = TRUE;
2399 mini_gc_init_gc_map (cfg);
2403 * Problems with the current code:
2404 * - the stack walk is slow
2405 * - vtypes/refs used in EH regions are treated conservatively
2406 * - if the code is finished, less pinning will be done, causing problems because
2407 * we promote all surviving objects to old-gen.
2408 * - the unwind code can't handle a method stopped inside a finally region, it thinks the caller is
2409 * another method, but in reality it is either the exception handling code or the CALL_HANDLER opcode.
2410 * This manifests in "Unable to find ip offset x in callsite list" assertions.
2411 * - the unwind code also can't handle frames which are in the epilog, since the unwind info is not
2416 * Ideas for creating smaller GC maps:
2417 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2419 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2421 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2422 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2423 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2424 * from an index entry.