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)
8 * Copyright 2011 Xamarin, Inc (http://www.xamarin.com)
13 #include <mono/metadata/gc-internal.h>
16 #if defined(MONO_ARCH_GC_MAPS_SUPPORTED)
18 #include <mono/metadata/gc-internal.h>
19 #include <mono/utils/mono-counters.h>
21 #if SIZEOF_VOID_P == 4
22 typedef guint32 mword;
24 typedef guint64 mword;
27 #define SIZEOF_SLOT ((int)sizeof (mgreg_t))
29 #define GC_BITS_PER_WORD (sizeof (mword) * 8)
31 /* Contains state needed by the GC Map construction code */
34 * This contains information about stack slots initialized in the prolog, encoded using
35 * (slot_index << 16) | slot_type. The slot_index is relative to the CFA, i.e. 0
36 * means cfa+0, 1 means cfa-4/8, etc.
38 GSList *stack_slots_from_cfa;
39 /* Same for stack slots relative to the frame pointer */
40 GSList *stack_slots_from_fp;
42 /* Number of slots in the map */
44 /* The number of registers in the map */
46 /* Min and Max offsets of the stack frame relative to fp */
47 int min_offset, max_offset;
48 /* Same for the locals area */
49 int locals_min_offset, locals_max_offset;
51 /* The call sites where this frame can be stopped during GC */
52 GCCallSite **callsites;
53 /* The number of call sites */
57 * The width of the stack bitmaps in bytes. This is not equal to the bitmap width at
58 * runtime, since it includes columns which are 0.
60 int stack_bitmap_width;
62 * A bitmap whose width equals nslots, and whose height equals ncallsites.
63 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
66 guint8 *stack_ref_bitmap;
67 /* Same for SLOT_PIN */
68 guint8 *stack_pin_bitmap;
71 * Similar bitmaps for registers. These have width MONO_MAX_IREGS in bits.
74 guint8 *reg_ref_bitmap;
75 guint8 *reg_pin_bitmap;
78 #define ALIGN_TO(val,align) ((((mgreg_t)val) + ((align) - 1)) & ~((align) - 1))
83 /* We don't support debug levels, its all-or-nothing */
84 #define DEBUG(s) do { s; fflush (logfile); } while (0)
85 #define DEBUG_ENABLED 1
92 #define DEBUG_PRECISE(s) do { s; } while (0)
93 #define DEBUG_PRECISE_ENABLED
95 #define DEBUG_PRECISE(s)
99 * Contains information collected during the conservative stack marking pass,
100 * used during the precise pass. This helps to avoid doing a stack walk twice, which
106 int frame_start_offset;
108 /* Relative to stack_start */
109 int reg_locations [MONO_MAX_IREGS];
110 #ifdef DEBUG_PRECISE_ENABLED
113 int regs [MONO_MAX_IREGS];
117 /* Max number of frames stored in the TLS data */
118 #define MAX_FRAMES 50
121 * Per-thread data kept by this module. This is stored in the GC and passed to us as
122 * parameters, instead of being stored in a TLS variable, since during a collection,
123 * only the collection thread is active.
128 gboolean has_context;
129 MonoJitTlsData *jit_tls;
132 gpointer ref_to_track;
133 /* Number of frames collected during the !precise pass */
135 FrameInfo frames [MAX_FRAMES];
138 /* These are constant so don't store them in the GC Maps */
139 /* Number of registers stored in gc maps */
140 #define NREGS MONO_MAX_IREGS
144 * Contains information needed to mark a stack frame.
145 * This is a transient structure, created from a compressed representation on-demand.
149 * The offsets of the GC tracked area inside the stack frame relative to the frame pointer.
150 * This includes memory which is NOREF thus doesn't need GC maps.
155 * The offset relative to frame_offset where the the memory described by the GC maps
159 /* The number of stack slots in the map */
161 /* The frame pointer register */
163 /* The size of each callsite table entry */
164 guint8 callsite_entry_size;
165 guint has_pin_slots : 1;
166 guint has_ref_slots : 1;
167 guint has_ref_regs : 1;
168 guint has_pin_regs : 1;
170 /* The offsets below are into an external bitmaps array */
173 * A bitmap whose width is equal to bitmap_width, and whose height is equal to ncallsites.
174 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
177 guint32 stack_ref_bitmap_offset;
179 * Same for SLOT_PIN. It is possible that the same bit is set in both bitmaps at
180 * different callsites, if the slot starts out as PIN, and later changes to REF.
182 guint32 stack_pin_bitmap_offset;
185 * Corresponding bitmaps for registers
186 * These have width equal to the number of bits set in reg_ref_mask/reg_pin_mask.
187 * FIXME: Merge these with the normal bitmaps, i.e. reserve the first x slots for them ?
189 guint32 reg_pin_bitmap_offset;
190 guint32 reg_ref_bitmap_offset;
192 guint32 used_int_regs, reg_ref_mask, reg_pin_mask;
194 /* The number of bits set in the two masks above */
195 guint8 nref_regs, npin_regs;
198 * A bit array marking slots which contain refs.
199 * This is used only for debugging.
203 /* Callsite offsets */
204 /* These can take up a lot of space, so encode them compactly */
214 * A compressed version of GCMap. This is what gets stored in MonoJitInfo.
218 //guint8 encoded_size;
221 * The arrays below are embedded after the struct.
222 * Their address needs to be computed.
225 /* The fixed fields of the GCMap encoded using LEB128 */
226 guint8 encoded [MONO_ZERO_LEN_ARRAY];
228 /* An array of ncallsites entries, each entry is callsite_entry_size bytes long */
229 guint8 callsites [MONO_ZERO_LEN_ARRAY];
232 guint8 bitmaps [MONO_ZERO_LEN_ARRAY];
235 static int precise_frame_count [2], precise_frame_limit = -1;
236 static gboolean precise_frame_limit_inited;
242 int scanned_precisely;
243 int scanned_conservatively;
244 int scanned_registers;
254 int gc_callsites_size;
255 int gc_callsites8_size;
256 int gc_callsites16_size;
257 int gc_callsites32_size;
259 int gc_map_struct_size;
263 static JITGCStats stats;
265 static FILE *logfile;
267 // FIXME: Move these to a shared place
270 encode_uleb128 (guint32 value, guint8 *buf, guint8 **endbuf)
275 guint8 b = value & 0x7f;
277 if (value != 0) /* more bytes to come */
285 static G_GNUC_UNUSED void
286 encode_sleb128 (gint32 value, guint8 *buf, guint8 **endbuf)
289 gboolean negative = (value < 0);
297 /* the following is unnecessary if the
298 * implementation of >>= uses an arithmetic rather
299 * than logical shift for a signed left operand
303 value |= - (1 <<(size - 7));
304 /* sign bit of byte is second high order bit (0x40) */
305 if ((value == 0 && !(byte & 0x40)) ||
306 (value == -1 && (byte & 0x40)))
316 static inline guint32
317 decode_uleb128 (guint8 *buf, guint8 **endbuf)
327 res = res | (((int)(b & 0x7f)) << shift);
339 decode_sleb128 (guint8 *buf, guint8 **endbuf)
349 res = res | (((int)(b & 0x7f)) << shift);
352 if (shift < 32 && (b & 0x40))
353 res |= - (1 << shift);
364 encode_frame_reg (int frame_reg)
367 if (frame_reg == AMD64_RSP)
369 else if (frame_reg == AMD64_RBP)
371 #elif defined(TARGET_X86)
372 if (frame_reg == X86_EBP)
374 else if (frame_reg == X86_ESP)
376 #elif defined(TARGET_ARM)
377 if (frame_reg == ARMREG_SP)
379 else if (frame_reg == ARMREG_FP)
384 g_assert_not_reached ();
389 decode_frame_reg (int encoded)
394 else if (encoded == 1)
396 #elif defined(TARGET_X86)
399 else if (encoded == 1)
401 #elif defined(TARGET_ARM)
404 else if (encoded == 1)
409 g_assert_not_reached ();
415 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15, AMD64_RDI, AMD64_RSI };
417 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15 };
419 #elif defined(TARGET_X86)
420 static int callee_saved_regs [] = { X86_EBX, X86_ESI, X86_EDI };
421 #elif defined(TARGET_ARM)
422 static int callee_saved_regs [] = { ARMREG_V1, ARMREG_V2, ARMREG_V3, ARMREG_V4, ARMREG_V5, ARMREG_V7, ARMREG_FP };
426 encode_regmask (guint32 regmask)
432 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
433 if (regmask & (1 << callee_saved_regs [i])) {
435 regmask -= (1 << callee_saved_regs [i]);
438 g_assert (regmask == 0);
443 decode_regmask (guint32 regmask)
449 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i)
450 if (regmask & (1 << i))
451 res |= (1 << callee_saved_regs [i]);
458 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
461 encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
465 encode_sleb128 (map->start_offset / SIZEOF_SLOT, buf, &buf);
466 encode_sleb128 (map->end_offset / SIZEOF_SLOT, buf, &buf);
467 encode_sleb128 (map->map_offset / SIZEOF_SLOT, buf, &buf);
468 encode_uleb128 (map->nslots, buf, &buf);
469 g_assert (map->callsite_entry_size <= 4);
470 freg = encode_frame_reg (map->frame_reg);
472 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);
473 encode_uleb128 (flags, buf, &buf);
474 encode_uleb128 (encode_regmask (map->used_int_regs), buf, &buf);
475 if (map->has_ref_regs)
476 encode_uleb128 (encode_regmask (map->reg_ref_mask), buf, &buf);
477 if (map->has_pin_regs)
478 encode_uleb128 (encode_regmask (map->reg_pin_mask), buf, &buf);
479 encode_uleb128 (map->ncallsites, buf, &buf);
487 * Decode the encoded GC map representation in BUF and store the result into MAP.
490 decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
493 int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
496 map->start_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
497 map->end_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
498 map->map_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
499 map->nslots = decode_uleb128 (buf, &buf);
500 flags = decode_uleb128 (buf, &buf);
501 map->has_ref_slots = (flags & 1) ? 1 : 0;
502 map->has_pin_slots = (flags & 2) ? 1 : 0;
503 map->has_ref_regs = (flags & 4) ? 1 : 0;
504 map->has_pin_regs = (flags & 8) ? 1 : 0;
505 map->callsite_entry_size = ((flags >> 4) & 0x3) + 1;
507 map->frame_reg = decode_frame_reg (freg);
508 map->used_int_regs = decode_regmask (decode_uleb128 (buf, &buf));
509 if (map->has_ref_regs) {
510 map->reg_ref_mask = decode_regmask (decode_uleb128 (buf, &buf));
512 for (i = 0; i < NREGS; ++i)
513 if (map->reg_ref_mask & (1 << i))
517 if (map->has_pin_regs) {
518 map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
520 for (i = 0; i < NREGS; ++i)
521 if (map->reg_pin_mask & (1 << i))
525 map->ncallsites = decode_uleb128 (buf, &buf);
527 stack_bitmap_size = (ALIGN_TO (map->nslots, 8) / 8) * map->ncallsites;
528 reg_ref_bitmap_size = (ALIGN_TO (map->nref_regs, 8) / 8) * map->ncallsites;
529 reg_pin_bitmap_size = (ALIGN_TO (map->npin_regs, 8) / 8) * map->ncallsites;
531 map->stack_ref_bitmap_offset = offset;
532 if (map->has_ref_slots)
533 offset += stack_bitmap_size;
534 map->stack_pin_bitmap_offset = offset;
535 if (map->has_pin_slots)
536 offset += stack_bitmap_size;
537 map->reg_ref_bitmap_offset = offset;
538 if (map->has_ref_regs)
539 offset += reg_ref_bitmap_size;
540 map->reg_pin_bitmap_offset = offset;
541 if (map->has_pin_regs)
542 offset += reg_pin_bitmap_size;
548 thread_attach_func (void)
552 tls = g_new0 (TlsData, 1);
553 tls->tid = GetCurrentThreadId ();
554 stats.tlsdata_size += sizeof (TlsData);
560 thread_detach_func (gpointer user_data)
562 TlsData *tls = user_data;
568 thread_suspend_func (gpointer user_data, void *sigctx)
570 TlsData *tls = user_data;
573 /* Happens during startup */
576 tls->lmf = mono_get_lmf ();
578 mono_arch_sigctx_to_monoctx (sigctx, &tls->ctx);
579 tls->has_context = TRUE;
581 tls->has_context = FALSE;
583 tls->jit_tls = mono_native_tls_get_value (mono_jit_tls_id);
586 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
589 set_bit (guint8 *bitmap, int width, int y, int x)
591 bitmap [(width * y) + (x / 8)] |= (1 << (x % 8));
595 clear_bit (guint8 *bitmap, int width, int y, int x)
597 bitmap [(width * y) + (x / 8)] &= ~(1 << (x % 8));
601 get_bit (guint8 *bitmap, int width, int y, int x)
603 return bitmap [(width * y) + (x / 8)] & (1 << (x % 8));
607 slot_type_to_string (GCSlotType type)
617 g_assert_not_reached ();
622 static inline mgreg_t
623 get_frame_pointer (MonoContext *ctx, int frame_reg)
625 #if defined(TARGET_AMD64)
626 if (frame_reg == AMD64_RSP)
628 else if (frame_reg == AMD64_RBP)
630 #elif defined(TARGET_X86)
631 if (frame_reg == X86_ESP)
633 else if (frame_reg == X86_EBP)
635 #elif defined(TARGET_ARM)
636 if (frame_reg == ARMREG_SP)
637 return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
638 else if (frame_reg == ARMREG_FP)
639 return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
641 g_assert_not_reached ();
646 * conservatively_pass:
648 * Mark a thread stack conservatively and collect information needed by the precise pass.
651 conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
654 MonoContext ctx, new_ctx;
657 gboolean last = TRUE;
661 guint8* fp, *p, *real_frame_start, *frame_start, *frame_end;
662 int i, pc_offset, cindex, bitmap_width;
663 int scanned = 0, scanned_precisely, scanned_conservatively, scanned_registers;
665 StackFrameInfo frame;
666 mgreg_t *reg_locations [MONO_MAX_IREGS];
667 mgreg_t *new_reg_locations [MONO_MAX_IREGS];
670 guint32 precise_regmask;
674 tls->ref_to_track = NULL;
677 /* tls == NULL can happen during startup */
678 if (mono_thread_internal_current () == NULL || !tls) {
679 mono_gc_conservatively_scan_area (stack_start, stack_end);
680 stats.scanned_stacks += stack_end - stack_start;
687 /* Number of bytes scanned based on GC map data */
689 /* Number of bytes scanned precisely based on GC map data */
690 scanned_precisely = 0;
691 /* Number of bytes scanned conservatively based on GC map data */
692 scanned_conservatively = 0;
693 /* Number of bytes scanned conservatively in register save areas */
694 scanned_registers = 0;
696 /* This is one past the last address which we have scanned */
697 stack_limit = stack_start;
699 if (!tls->has_context)
700 memset (&new_ctx, 0, sizeof (ctx));
702 memcpy (&new_ctx, &tls->ctx, sizeof (MonoContext));
704 memset (reg_locations, 0, sizeof (reg_locations));
705 memset (new_reg_locations, 0, sizeof (new_reg_locations));
708 memcpy (&ctx, &new_ctx, sizeof (ctx));
710 for (i = 0; i < MONO_MAX_IREGS; ++i) {
711 if (new_reg_locations [i]) {
713 * If the current frame saves the register, it means it might modify its
714 * value, thus the old location might not contain the same value, so
715 * we have to mark it conservatively.
717 if (reg_locations [i]) {
718 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
719 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
720 scanned_registers += SIZEOF_SLOT;
723 reg_locations [i] = new_reg_locations [i];
725 DEBUG (fprintf (logfile, "\treg %s is now at location %p.\n", mono_arch_regname (i), reg_locations [i]));
729 g_assert ((mgreg_t)stack_limit % SIZEOF_SLOT == 0);
731 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);
737 if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
739 * These frames are problematic for several reasons:
740 * - they are unwound through an LMF, and we have no precise register tracking for those.
741 * - the LMF might not contain a precise ip, so we can't compute the call site.
742 * - the LMF only unwinds to the wrapper frame, so we get these methods twice.
744 DEBUG (fprintf (logfile, "Mark(0): <Managed-to-native transition>\n"));
745 for (i = 0; i < MONO_MAX_IREGS; ++i) {
746 if (reg_locations [i]) {
747 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
748 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
749 scanned_registers += SIZEOF_SLOT;
751 reg_locations [i] = NULL;
752 new_reg_locations [i] = NULL;
758 /* The last frame can be in any state so mark conservatively */
761 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));
763 DEBUG (fprintf (logfile, "\t <Last frame>\n"));
768 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
770 /* These frames are very problematic */
771 if (ji->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
772 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));
773 DEBUG (fprintf (logfile, "\tSkip.\n"));
777 /* All the other frames are at a call site */
779 if (tls->nframes == MAX_FRAMES) {
781 * Can't save information since the array is full. So scan the rest of the
782 * stack conservatively.
784 DEBUG (fprintf (logfile, "Mark (0): Frame stack full.\n"));
788 /* Scan the frame of this method */
791 * A frame contains the following:
796 * - localloc-ed memory
798 g_assert (pc_offset >= 0);
803 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));
804 DEBUG (fprintf (logfile, "\tNo GC Map.\n"));
808 /* The embedded callsite table requires this */
809 g_assert (((mgreg_t)emap % 4) == 0);
812 * Debugging aid to control the number of frames scanned precisely
814 if (!precise_frame_limit_inited) {
815 if (getenv ("MONO_PRECISE_COUNT"))
816 precise_frame_limit = atoi (getenv ("MONO_PRECISE_COUNT"));
817 precise_frame_limit_inited = TRUE;
820 if (precise_frame_limit != -1) {
821 if (precise_frame_count [FALSE] == precise_frame_limit)
822 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (ji->method, TRUE));
823 if (precise_frame_count [FALSE] > precise_frame_limit)
826 precise_frame_count [FALSE] ++;
828 /* Decode the encoded GC map */
830 memset (map, 0, sizeof (GCMap));
831 decode_gc_map (&emap->encoded [0], map, &p);
832 p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
833 map->callsites.offsets8 = p;
834 p += map->callsite_entry_size * map->ncallsites;
837 fp = (guint8*)get_frame_pointer (&ctx, map->frame_reg);
839 real_frame_start = fp + map->start_offset;
840 frame_start = fp + map->start_offset + map->map_offset;
841 frame_end = fp + map->end_offset;
843 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));
845 /* Find the callsite index */
846 if (map->callsite_entry_size == 1) {
847 for (i = 0; i < map->ncallsites; ++i)
848 /* ip points inside the call instruction */
849 if (map->callsites.offsets8 [i] == pc_offset + 1)
851 } else if (map->callsite_entry_size == 2) {
852 // FIXME: Use a binary search
853 for (i = 0; i < map->ncallsites; ++i)
854 /* ip points inside the call instruction */
855 if (map->callsites.offsets16 [i] == pc_offset + 1)
858 // FIXME: Use a binary search
859 for (i = 0; i < map->ncallsites; ++i)
860 /* ip points inside the call instruction */
861 if (map->callsites.offsets32 [i] == pc_offset + 1)
864 if (i == map->ncallsites) {
865 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (ji->method, TRUE));
866 g_assert_not_reached ();
871 * This is not neccessary true on x86 because frames have a different size at each
874 //g_assert (real_frame_start >= stack_limit);
876 if (real_frame_start > stack_limit) {
877 /* This scans the previously skipped frames as well */
878 DEBUG (fprintf (logfile, "\tscan area %p-%p (%d).\n", stack_limit, real_frame_start, (int)(real_frame_start - stack_limit)));
879 mono_gc_conservatively_scan_area (stack_limit, real_frame_start);
880 stats.scanned_other += real_frame_start - stack_limit;
883 /* Mark stack slots */
884 if (map->has_pin_slots) {
885 int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
886 guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
891 for (i = 0; i < map->nslots; ++i) {
892 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
894 DEBUG (fprintf (logfile, "\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
895 mono_gc_conservatively_scan_area (p, p + SIZEOF_SLOT);
896 scanned_conservatively += SIZEOF_SLOT;
898 scanned_precisely += SIZEOF_SLOT;
903 scanned_precisely += (map->nslots * SIZEOF_SLOT);
906 /* The area outside of start-end is NOREF */
907 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * SIZEOF_SLOT);
910 precise_regmask = map->used_int_regs | (1 << map->frame_reg);
911 if (map->has_pin_regs) {
912 int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
913 guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
915 for (i = 0; i < NREGS; ++i) {
916 if (!(map->used_int_regs & (1 << i)))
919 if (!(map->reg_pin_mask & (1 << i)))
922 if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
923 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
924 precise_regmask &= ~(1 << i);
930 scanned += map->end_offset - map->start_offset;
932 g_assert (scanned == scanned_precisely + scanned_conservatively);
934 stack_limit = frame_end;
936 /* Save information for the precise pass */
937 fi = &tls->frames [tls->nframes];
938 fi->nslots = map->nslots;
939 bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
940 if (map->has_ref_slots)
941 fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
944 fi->frame_start_offset = frame_start - stack_start;
945 fi->nreg_locations = 0;
946 DEBUG_PRECISE (fi->ji = ji);
947 DEBUG_PRECISE (fi->fp = fp);
949 if (map->has_ref_regs) {
950 int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
951 guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
953 for (i = 0; i < NREGS; ++i) {
954 if (!(map->reg_ref_mask & (1 << i)))
957 if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
958 DEBUG_PRECISE (fi->regs [fi->nreg_locations] = i);
959 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
960 fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
961 fi->nreg_locations ++;
968 * Clear locations of precisely stacked registers.
970 if (precise_regmask) {
971 for (i = 0; i < NREGS; ++i) {
972 if (precise_regmask & (1 << i)) {
974 * The method uses this register, and we have precise info for it.
975 * This means the location will be scanned precisely.
976 * Tell the code at the beginning of the loop that this location is
979 if (reg_locations [i])
980 DEBUG (fprintf (logfile, "\treg %s at location %p (==%p) is precise.\n", mono_arch_regname (i), reg_locations [i], (gpointer)*reg_locations [i]));
981 reg_locations [i] = NULL;
989 /* Scan the remaining register save locations */
990 for (i = 0; i < MONO_MAX_IREGS; ++i) {
991 if (reg_locations [i]) {
992 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", reg_locations [i]));
993 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
994 scanned_registers += SIZEOF_SLOT;
996 if (new_reg_locations [i]) {
997 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", new_reg_locations [i]));
998 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
999 scanned_registers += SIZEOF_SLOT;
1003 if (stack_limit < stack_end) {
1004 DEBUG (fprintf (logfile, "\tscan remaining stack %p-%p (%d).\n", stack_limit, stack_end, (int)(stack_end - stack_limit)));
1005 mono_gc_conservatively_scan_area (stack_limit, stack_end);
1006 stats.scanned_native += stack_end - stack_limit;
1009 DEBUG (fprintf (logfile, "Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
1011 stats.scanned_stacks += stack_end - stack_start;
1012 stats.scanned += scanned;
1013 stats.scanned_precisely += scanned_precisely;
1014 stats.scanned_conservatively += scanned_conservatively;
1015 stats.scanned_registers += scanned_registers;
1017 //mono_gc_conservatively_scan_area (stack_start, stack_end);
1023 * Mark a thread stack precisely based on information saved during the conservative
1027 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
1031 guint8 *frame_start;
1036 for (findex = 0; findex < tls->nframes; findex ++) {
1037 /* Load information saved by the !precise pass */
1038 fi = &tls->frames [findex];
1039 frame_start = stack_start + fi->frame_start_offset;
1041 DEBUG (char *fname = mono_method_full_name (fi->ji->method, TRUE); fprintf (logfile, "Mark(1): %s\n", fname); g_free (fname));
1044 * FIXME: Add a function to mark using a bitmap, to avoid doing a
1045 * call for each object.
1048 /* Mark stack slots */
1050 guint8 *ref_bitmap = fi->bitmap;
1053 for (i = 0; i < fi->nslots; ++i) {
1054 MonoObject **ptr = (MonoObject**)(frame_start + (i * SIZEOF_SLOT));
1056 live = ref_bitmap [i / 8] & (1 << (i % 8));
1059 MonoObject *obj = *ptr;
1061 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1062 *ptr = mono_gc_scan_object (obj);
1063 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1065 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1070 * This is disabled because the pointer takes up a lot of space.
1071 * Stack slots might be shared between ref and non-ref variables ?
1073 if (map->ref_slots [i / 8] & (1 << (i % 8))) {
1074 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));
1076 * Fail fast if the live range is incorrect, and
1077 * the JITted code tries to access this object
1086 /* Mark registers */
1089 * Registers are different from stack slots, they have no address where they
1090 * are stored. Instead, some frame below this frame in the stack saves them
1091 * in its prolog to the stack. We can mark this location precisely.
1093 for (i = 0; i < fi->nreg_locations; ++i) {
1095 * reg_locations [i] contains the address of the stack slot where
1096 * a reg was last saved, so mark that slot.
1098 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1099 MonoObject *obj = *ptr;
1102 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
1103 *ptr = mono_gc_scan_object (obj);
1104 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1106 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p\n", mono_arch_regname (fi->regs [i]), ptr, obj));
1112 * Debugging aid to check for missed refs.
1114 if (tls->ref_to_track) {
1117 for (p = (mgreg_t*)stack_start; p < (mgreg_t*)stack_end; ++p)
1118 if (*p == (mgreg_t)tls->ref_to_track)
1119 printf ("REF AT %p.\n", p);
1126 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1127 * call, and TRUE at the second. USER_DATA points to a TlsData
1128 * structure filled up by thread_suspend_func.
1131 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise)
1133 TlsData *tls = user_data;
1135 DEBUG (fprintf (logfile, "****************************************\n"));
1136 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));
1137 DEBUG (fprintf (logfile, "****************************************\n"));
1140 conservative_pass (tls, stack_start, stack_end);
1142 precise_pass (tls, stack_start, stack_end);
1146 mini_gc_init_gc_map (MonoCompile *cfg)
1148 if (COMPILE_LLVM (cfg))
1151 if (!mono_gc_is_moving ())
1154 if (!cfg->compile_aot && !mono_gc_precise_stack_mark_enabled ())
1158 /* Debugging support */
1160 static int precise_count;
1163 if (getenv ("MONO_GCMAP_COUNT")) {
1164 if (precise_count == atoi (getenv ("MONO_GCMAP_COUNT")))
1165 printf ("LAST: %s\n", mono_method_full_name (cfg->method, TRUE));
1166 if (precise_count > atoi (getenv ("MONO_GCMAP_COUNT")))
1172 cfg->compute_gc_maps = TRUE;
1174 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1178 * mini_gc_set_slot_type_from_fp:
1180 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1181 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1182 * need to call this function for those slots.
1185 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1187 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1189 if (!cfg->compute_gc_maps)
1192 g_assert (slot_offset % SIZEOF_SLOT == 0);
1194 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1198 * mini_gc_set_slot_type_from_cfa:
1200 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1201 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1202 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1203 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1204 * call this function for those slots.
1207 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1209 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1210 int slot = - (slot_offset / SIZEOF_SLOT);
1212 if (!cfg->compute_gc_maps)
1215 g_assert (slot_offset <= 0);
1216 g_assert (slot_offset % SIZEOF_SLOT == 0);
1218 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1222 fp_offset_to_slot (MonoCompile *cfg, int offset)
1224 MonoCompileGC *gcfg = cfg->gc_info;
1226 return (offset - gcfg->min_offset) / SIZEOF_SLOT;
1230 slot_to_fp_offset (MonoCompile *cfg, int slot)
1232 MonoCompileGC *gcfg = cfg->gc_info;
1234 return (slot * SIZEOF_SLOT) + gcfg->min_offset;
1238 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1240 g_assert (slot >= 0 && slot < gcfg->nslots);
1242 if (type == SLOT_PIN) {
1243 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1244 set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1245 } else if (type == SLOT_REF) {
1246 set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1247 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1248 } else if (type == SLOT_NOREF) {
1249 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1250 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1255 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1259 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1260 set_slot (gcfg, slot, cindex, type);
1264 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1268 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1269 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1270 if (callsite_offset >= from && callsite_offset < to)
1271 set_slot (gcfg, slot, cindex, type);
1276 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1278 g_assert (slot >= 0 && slot < gcfg->nregs);
1280 if (type == SLOT_PIN) {
1281 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1282 set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1283 } else if (type == SLOT_REF) {
1284 set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1285 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1286 } else if (type == SLOT_NOREF) {
1287 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1288 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1293 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1297 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1298 set_reg_slot (gcfg, slot, cindex, type);
1302 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1306 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1307 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1308 if (callsite_offset >= from && callsite_offset < to)
1309 set_reg_slot (gcfg, slot, cindex, type);
1314 process_spill_slots (MonoCompile *cfg)
1316 MonoCompileGC *gcfg = cfg->gc_info;
1321 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1322 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1323 for (l = bb->spill_slot_defs; l; l = l->next) {
1324 MonoInst *def = l->data;
1325 int spill_slot = def->inst_c0;
1326 int bank = def->inst_c1;
1327 int offset = cfg->spill_info [bank][spill_slot].offset;
1328 int slot = fp_offset_to_slot (cfg, offset);
1330 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1331 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1335 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1336 for (l = bb->spill_slot_defs; l; l = l->next) {
1337 MonoInst *def = l->data;
1338 int spill_slot = def->inst_c0;
1339 int bank = def->inst_c1;
1340 int offset = cfg->spill_info [bank][spill_slot].offset;
1341 int slot = fp_offset_to_slot (cfg, offset);
1344 if (bank == MONO_REG_INT_MP)
1350 * Extend the live interval for the GC tracked spill slots
1351 * defined in this bblock.
1352 * FIXME: This is not needed.
1354 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1356 if (cfg->verbose_level > 1)
1357 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
1361 /* Set fp spill slots to NOREF */
1362 for (i = 0; i < cfg->spill_info_len [MONO_REG_DOUBLE]; ++i) {
1363 int offset = cfg->spill_info [MONO_REG_DOUBLE][i].offset;
1369 slot = fp_offset_to_slot (cfg, offset);
1371 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1373 if (cfg->verbose_level > 1)
1374 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1377 /* Set int spill slots to NOREF */
1378 for (i = 0; i < cfg->spill_info_len [MONO_REG_INT]; ++i) {
1379 int offset = cfg->spill_info [MONO_REG_INT][i].offset;
1385 slot = fp_offset_to_slot (cfg, offset);
1387 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1388 if (cfg->verbose_level > 1)
1389 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1394 * process_other_slots:
1396 * Process stack slots registered using mini_gc_set_slot_type_... ().
1399 process_other_slots (MonoCompile *cfg)
1401 MonoCompileGC *gcfg = cfg->gc_info;
1404 /* Relative to the CFA */
1405 for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
1406 guint data = GPOINTER_TO_UINT (l->data);
1407 int cfa_slot = data >> 16;
1408 GCSlotType type = data & 0xff;
1412 * Map the cfa relative slot to an fp relative slot.
1413 * slot_addr == cfa - <cfa_slot>*4/8
1414 * fp + cfa_offset == cfa
1415 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1417 slot = (cfg->cfa_offset / SIZEOF_SLOT) - cfa_slot - (gcfg->min_offset / SIZEOF_SLOT);
1419 set_slot_everywhere (gcfg, slot, type);
1421 if (cfg->verbose_level > 1) {
1422 int fp_offset = slot_to_fp_offset (cfg, slot);
1423 if (type == SLOT_NOREF)
1424 printf ("\tnoref slot at %s0x%x(fp) (slot = %d) (cfa - 0x%x)\n", fp_offset >= 0 ? "" : "-", ABS (fp_offset), slot, (int)(cfa_slot * SIZEOF_SLOT));
1428 /* Relative to the FP */
1429 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1430 gint data = GPOINTER_TO_INT (l->data);
1431 int offset = data >> 16;
1432 GCSlotType type = data & 0xff;
1435 slot = fp_offset_to_slot (cfg, offset);
1437 set_slot_everywhere (gcfg, slot, type);
1439 /* Liveness for these slots is handled by process_spill_slots () */
1441 if (cfg->verbose_level > 1) {
1442 if (type == SLOT_REF)
1443 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
1444 else if (type == SLOT_NOREF)
1445 printf ("\tnoref slot at 0x%x(fp) (slot = %d)\n", offset, slot);
1451 process_variables (MonoCompile *cfg)
1453 MonoCompileGC *gcfg = cfg->gc_info;
1454 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1455 int i, locals_min_slot, locals_max_slot, cindex;
1459 int locals_min_offset = gcfg->locals_min_offset;
1460 int locals_max_offset = gcfg->locals_max_offset;
1462 /* Slots for locals are NOREF by default */
1463 locals_min_slot = (locals_min_offset - gcfg->min_offset) / SIZEOF_SLOT;
1464 locals_max_slot = (locals_max_offset - gcfg->min_offset) / SIZEOF_SLOT;
1465 for (i = locals_min_slot; i < locals_max_slot; ++i) {
1466 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1470 * Compute the offset where variables are initialized in the first bblock, if any.
1472 pc_offsets = g_new0 (int, cfg->next_vreg);
1474 bb = cfg->bb_entry->next_bb;
1475 MONO_BB_FOR_EACH_INS (bb, tmp) {
1476 if (tmp->opcode == OP_GC_LIVENESS_DEF) {
1477 int vreg = tmp->inst_c1;
1478 if (pc_offsets [vreg] == 0) {
1479 g_assert (tmp->backend.pc_offset > 0);
1480 pc_offsets [vreg] = tmp->backend.pc_offset;
1486 * Stack slots holding arguments are initialized in the prolog.
1487 * This means we can treat them alive for the whole method.
1489 for (i = 0; i < cfg->num_varinfo; i++) {
1490 MonoInst *ins = cfg->varinfo [i];
1491 MonoType *t = ins->inst_vtype;
1494 gboolean byref, is_this = FALSE;
1495 gboolean is_arg = i < cfg->locals_start;
1497 if (ins == cfg->ret)
1500 vmv = MONO_VARINFO (cfg, i);
1502 /* For some reason, 'this' is byref */
1503 if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
1504 t = &cfg->method->klass->byval_arg;
1510 if (ins->opcode == OP_REGVAR) {
1512 GCSlotType slot_type;
1514 t = mini_type_get_underlying_type (NULL, t);
1517 g_assert (hreg < MONO_MAX_IREGS);
1520 slot_type = SLOT_PIN;
1522 slot_type = mini_type_is_reference (cfg, t) ? SLOT_REF : SLOT_NOREF;
1524 if (slot_type == SLOT_PIN) {
1525 /* These have no live interval, be conservative */
1526 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1529 * Unlike variables allocated to the stack, we generate liveness info
1530 * for noref vars in registers in mono_spill_global_vars (), because
1531 * knowing that a register doesn't contain a ref allows us to mark its save
1532 * locations precisely.
1534 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1535 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1536 set_reg_slot (gcfg, hreg, cindex, slot_type);
1539 if (cfg->verbose_level > 1) {
1540 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
1546 if (ins->opcode != OP_REGOFFSET)
1549 if (ins->inst_offset % SIZEOF_SLOT != 0)
1552 if (is_arg && ins->inst_offset >= gcfg->max_offset)
1553 /* In parent frame */
1556 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1558 if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
1559 /* These do not get stored in the prolog */
1560 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1562 if (cfg->verbose_level > 1) {
1563 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));
1568 if (MONO_TYPE_ISSTRUCT (t)) {
1570 gsize *bitmap = NULL;
1571 gboolean pin = FALSE;
1575 if (ins->backend.is_pinvoke)
1576 size = mono_class_native_size (ins->klass, NULL);
1578 size = mono_class_value_size (ins->klass, NULL);
1579 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
1581 if (!ins->klass->has_references) {
1583 for (j = 0; j < size_in_slots; ++j)
1584 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1589 if (ins->klass->generic_container || mono_class_is_open_constructed_type (t)) {
1590 /* FIXME: Generic sharing */
1593 mono_class_compute_gc_descriptor (ins->klass);
1595 bitmap = mono_gc_get_bitmap_for_descr (ins->klass->gc_descr, &numbits);
1600 * Most vtypes are marked volatile because of the LDADDR instructions,
1601 * and they have no liveness information since they are decomposed
1602 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1603 * them during VZERO decomposition.
1605 if (!pc_offsets [vmv->vreg])
1609 if (ins->backend.is_pinvoke)
1612 if (cfg->verbose_level > 1)
1613 printf ("\tvtype R%d at fp+0x%x-0x%x: %s\n", vmv->vreg, (int)ins->inst_offset, (int)(ins->inst_offset + (size / SIZEOF_SLOT)), mono_type_full_name (ins->inst_vtype));
1616 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1617 if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
1618 for (j = 0; j < numbits; ++j) {
1619 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1620 /* The descriptor is for the boxed object */
1621 set_slot (gcfg, (pos + j - (sizeof (MonoObject) / SIZEOF_SLOT)), cindex, pin ? SLOT_PIN : SLOT_REF);
1627 if (cfg->verbose_level > 1) {
1628 for (j = 0; j < numbits; ++j) {
1629 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD)))
1630 printf ("\t\t%s slot at 0x%x(fp) (slot = %d)\n", pin ? "pin" : "ref", (int)(ins->inst_offset + (j * SIZEOF_SLOT)), (int)(pos + j - (sizeof (MonoObject) / SIZEOF_SLOT)));
1634 if (cfg->verbose_level > 1)
1635 printf ("\t\tpinned\n");
1636 for (j = 0; j < size_in_slots; ++j) {
1637 set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
1646 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1647 /* Vret addr etc. */
1652 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1654 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1655 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1656 set_slot (gcfg, pos, cindex, SLOT_PIN);
1658 if (cfg->verbose_level > 1)
1659 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));
1664 * This is currently disabled, but could be enabled to debug crashes.
1667 if (t->type == MONO_TYPE_I) {
1669 * Variables created in mono_handle_global_vregs have type I, but they
1670 * could hold GC refs since the vregs they were created from might not been
1671 * marked as holding a GC ref. So be conservative.
1673 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1678 t = mini_type_get_underlying_type (NULL, t);
1680 if (!mini_type_is_reference (cfg, t)) {
1681 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1682 if (cfg->verbose_level > 1)
1683 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));
1684 if (!t->byref && sizeof (mgreg_t) == 4 && (t->type == MONO_TYPE_I8 || t->type == MONO_TYPE_U8 || t->type == MONO_TYPE_R8)) {
1685 set_slot_everywhere (gcfg, pos + 1, SLOT_NOREF);
1686 if (cfg->verbose_level > 1)
1687 printf ("\tnoref at %s0x%x(fp) (R%d, slot = %d): %s\n", ins->inst_offset < 0 ? "-" : "", (ins->inst_offset < 0) ? -(int)(ins->inst_offset + 4) : (int)ins->inst_offset + 4, vmv->vreg, pos + 1, mono_type_full_name (ins->inst_vtype));
1692 /* 'this' is marked INDIRECT for gshared methods */
1693 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1695 * For volatile variables, treat them alive from the point they are
1696 * initialized in the first bblock until the end of the method.
1699 set_slot_everywhere (gcfg, pos, SLOT_REF);
1700 } else if (pc_offsets [vmv->vreg]) {
1701 set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
1702 set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
1704 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1706 if (cfg->verbose_level > 1)
1707 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));
1712 /* Live for the whole method */
1713 set_slot_everywhere (gcfg, pos, SLOT_REF);
1715 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1716 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1717 set_slot (gcfg, pos, cindex, SLOT_REF);
1720 if (cfg->verbose_level > 1) {
1721 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));
1725 g_free (pc_offsets);
1729 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1732 * Convert a sp relative offset to a slot index. This is
1733 * platform specific.
1736 /* fp = sp + offset */
1737 g_assert (cfg->frame_reg == AMD64_RBP);
1738 return (- cfg->arch.sp_fp_offset + sp_offset);
1739 #elif defined(TARGET_X86)
1740 /* The offset is computed from the sp at the start of the call sequence */
1741 g_assert (cfg->frame_reg == X86_EBP);
1742 return (- cfg->arch.sp_fp_offset - sp_offset);
1750 type_to_gc_slot_type (MonoCompile *cfg, MonoType *t)
1754 t = mini_type_get_underlying_type (NULL, t);
1755 if (mini_type_is_reference (cfg, t))
1758 if (MONO_TYPE_ISSTRUCT (t)) {
1759 MonoClass *klass = mono_class_from_mono_type (t);
1760 if (!klass->has_references) {
1772 process_param_area_slots (MonoCompile *cfg)
1774 MonoCompileGC *gcfg = cfg->gc_info;
1779 * These slots are used for passing parameters during calls. They are sp relative, not
1780 * fp relative, so they are harder to handle.
1782 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1783 /* The distance between fp and sp is not constant */
1786 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1788 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1789 GCCallSite *callsite = gcfg->callsites [cindex];
1792 for (l = callsite->param_slots; l; l = l->next) {
1793 MonoInst *def = l->data;
1794 MonoType *t = def->inst_vtype;
1795 int sp_offset = def->inst_offset;
1796 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1797 int slot = fp_offset_to_slot (cfg, fp_offset);
1801 if (MONO_TYPE_ISSTRUCT (t)) {
1802 size = mini_type_stack_size_full (cfg->generic_sharing_context, t, &align, FALSE);
1804 size = sizeof (mgreg_t);
1807 for (i = 0; i < size / sizeof (mgreg_t); ++i) {
1808 g_assert (slot + i >= 0 && slot + i < gcfg->nslots);
1809 is_param [slot + i] = TRUE;
1814 /* All param area slots are noref by default */
1815 for (i = 0; i < gcfg->nslots; ++i) {
1817 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1820 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1821 GCCallSite *callsite = gcfg->callsites [cindex];
1824 for (l = callsite->param_slots; l; l = l->next) {
1825 MonoInst *def = l->data;
1826 MonoType *t = def->inst_vtype;
1827 int sp_offset = def->inst_offset;
1828 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1829 int slot = fp_offset_to_slot (cfg, fp_offset);
1830 GCSlotType type = type_to_gc_slot_type (cfg, t);
1832 if (MONO_TYPE_ISSTRUCT (t)) {
1837 size = mini_type_stack_size_full (cfg->generic_sharing_context, t, &align, FALSE);
1838 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
1840 for (i = 0; i < size_in_slots; ++i) {
1841 set_slot_in_range (gcfg, slot + i, def->backend.pc_offset, callsite->pc_offset + 1, type);
1843 if (cfg->verbose_level > 1)
1844 printf ("\t%s param area slots at %s0x%x(fp)=0x%x(sp) (slot = %d-%d) [0x%x-0x%x]\n", slot_type_to_string (type), fp_offset >= 0 ? "+" : "-", ABS (fp_offset), sp_offset, slot, slot + (size / (int)sizeof (mgreg_t)), def->backend.pc_offset, callsite->pc_offset + 1);
1846 /* The slot is live between the def instruction and the call */
1847 set_slot_in_range (gcfg, slot, def->backend.pc_offset, callsite->pc_offset + 1, type);
1848 if (cfg->verbose_level > 1)
1849 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);
1856 process_finally_clauses (MonoCompile *cfg)
1858 MonoCompileGC *gcfg = cfg->gc_info;
1859 GCCallSite **callsites;
1861 gboolean has_finally;
1862 int i, j, nslots, nregs;
1864 ncallsites = gcfg->ncallsites;
1865 nslots = gcfg->nslots;
1866 nregs = gcfg->nregs;
1867 callsites = gcfg->callsites;
1870 * The calls to the finally clauses don't show up in the cfg. See
1871 * test_0_liveness_8 ().
1872 * Variables accessed inside the finally clause are already marked VOLATILE by
1873 * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
1874 * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
1876 has_finally = FALSE;
1877 for (i = 0; i < cfg->header->num_clauses; ++i) {
1878 MonoExceptionClause *clause = &cfg->header->clauses [i];
1880 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1885 if (cfg->verbose_level > 1)
1886 printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
1887 for (j = 0; j < ncallsites; ++j) {
1888 MonoBasicBlock *bb = callsites [j]->bb;
1889 MonoExceptionClause *clause;
1890 gboolean is_in_finally = FALSE;
1892 for (i = 0; i < cfg->header->num_clauses; ++i) {
1893 clause = &cfg->header->clauses [i];
1895 if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
1896 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1897 is_in_finally = TRUE;
1903 if (is_in_finally) {
1904 for (i = 0; i < nslots; ++i)
1905 set_slot (gcfg, i, j, SLOT_PIN);
1906 for (i = 0; i < nregs; ++i)
1907 set_reg_slot (gcfg, i, j, SLOT_PIN);
1914 compute_frame_size (MonoCompile *cfg)
1916 int i, locals_min_offset, locals_max_offset, cfa_min_offset, cfa_max_offset;
1917 int min_offset, max_offset;
1918 MonoCompileGC *gcfg = cfg->gc_info;
1919 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1922 /* Compute min/max offsets from the fp */
1925 #if defined(TARGET_AMD64) || defined(TARGET_X86) || defined(TARGET_ARM)
1926 locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, SIZEOF_SLOT);
1927 locals_max_offset = cfg->locals_max_stack_offset;
1929 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
1933 locals_min_offset = ALIGN_TO (locals_min_offset, SIZEOF_SLOT);
1934 locals_max_offset = ALIGN_TO (locals_max_offset, SIZEOF_SLOT);
1936 min_offset = locals_min_offset;
1937 max_offset = locals_max_offset;
1940 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
1941 MonoInst *ins = cfg->args [i];
1943 if (ins->opcode == OP_REGOFFSET)
1944 min_offset = MIN (min_offset, ins->inst_offset);
1948 g_assert (cfg->frame_reg == cfg->cfa_reg);
1949 g_assert (cfg->cfa_offset > 0);
1951 cfa_max_offset = cfg->cfa_offset;
1953 min_offset = MIN (min_offset, cfa_min_offset);
1954 max_offset = MAX (max_offset, cfa_max_offset);
1956 /* Fp relative slots */
1957 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1958 gint data = GPOINTER_TO_INT (l->data);
1959 int offset = data >> 16;
1961 min_offset = MIN (min_offset, offset);
1965 if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
1966 int stack_offset = ALIGN_TO (cfg->stack_offset, SIZEOF_SLOT);
1967 min_offset = MIN (min_offset, (-stack_offset));
1970 /* Param area slots */
1972 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
1973 #elif defined(TARGET_X86)
1974 min_offset = MIN (min_offset, - (cfg->arch.sp_fp_offset + cfg->arch.param_area_size));
1975 #elif defined(TARGET_ARM)
1981 gcfg->min_offset = min_offset;
1982 gcfg->max_offset = max_offset;
1983 gcfg->locals_min_offset = locals_min_offset;
1984 gcfg->locals_max_offset = locals_max_offset;
1988 init_gcfg (MonoCompile *cfg)
1990 int i, nregs, nslots;
1991 MonoCompileGC *gcfg = cfg->gc_info;
1992 GCCallSite **callsites;
2001 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2002 ncallsites += g_slist_length (bb->gc_callsites);
2004 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
2006 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2007 for (l = bb->gc_callsites; l; l = l->next)
2008 callsites [i++] = l->data;
2011 /* The callsites should already be ordered by pc offset */
2012 for (i = 1; i < ncallsites; ++i)
2013 g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
2016 * The stack frame looks like this:
2018 * <fp + max_offset> == cfa -> <end of previous frame>
2019 * <other stack slots>
2021 * <other stack slots>
2022 * fp + min_offset ->
2027 if (cfg->verbose_level > 1)
2028 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
2030 nslots = (gcfg->max_offset - gcfg->min_offset) / SIZEOF_SLOT;
2033 gcfg->nslots = nslots;
2034 gcfg->nregs = nregs;
2035 gcfg->callsites = callsites;
2036 gcfg->ncallsites = ncallsites;
2037 gcfg->stack_bitmap_width = ALIGN_TO (nslots, 8) / 8;
2038 gcfg->reg_bitmap_width = ALIGN_TO (nregs, 8) / 8;
2039 gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
2040 gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
2041 gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
2042 gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
2044 /* All slots start out as PIN */
2045 memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * ncallsites);
2046 for (i = 0; i < nregs; ++i) {
2048 * By default, registers are NOREF.
2049 * It is possible for a callee to save them before being defined in this method,
2050 * but the saved value is dead too, so it doesn't need to be marked.
2052 if ((cfg->used_int_regs & (1 << i)))
2053 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
2058 create_map (MonoCompile *cfg)
2061 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
2063 int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
2064 int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
2066 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
2067 MonoCompileGC *gcfg = cfg->gc_info;
2068 GCCallSite **callsites;
2070 guint8 *bitmap, *bitmaps;
2071 guint32 reg_ref_mask, reg_pin_mask;
2073 ncallsites = gcfg->ncallsites;
2074 nslots = gcfg->nslots;
2075 nregs = gcfg->nregs;
2076 callsites = gcfg->callsites;
2079 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
2080 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
2082 has_ref_slots = FALSE;
2083 has_pin_slots = FALSE;
2086 memset (ntypes, 0, sizeof (ntypes));
2087 for (i = 0; i < nslots; ++i) {
2088 gboolean has_ref = FALSE;
2089 gboolean has_pin = FALSE;
2091 for (j = 0; j < ncallsites; ++j) {
2092 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
2094 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
2099 has_ref_slots = TRUE;
2101 has_pin_slots = TRUE;
2104 ntypes [SLOT_REF] ++;
2106 ntypes [SLOT_PIN] ++;
2108 ntypes [SLOT_NOREF] ++;
2110 if (has_ref || has_pin) {
2117 start = end = nslots;
2119 g_assert (start != -1);
2120 g_assert (start < end);
2123 has_ref_regs = FALSE;
2124 has_pin_regs = FALSE;
2129 for (i = 0; i < nregs; ++i) {
2130 gboolean has_ref = FALSE;
2131 gboolean has_pin = FALSE;
2133 if (!(cfg->used_int_regs & (1 << i)))
2136 for (j = 0; j < ncallsites; ++j) {
2137 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i)) {
2142 for (j = 0; j < ncallsites; ++j) {
2143 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i)) {
2150 reg_ref_mask |= (1 << i);
2151 has_ref_regs = TRUE;
2155 reg_pin_mask |= (1 << i);
2156 has_pin_regs = TRUE;
2161 if (cfg->verbose_level > 1)
2162 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);
2164 /* Create the GC Map */
2166 stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
2167 stack_bitmap_size = stack_bitmap_width * ncallsites;
2168 reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
2169 reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
2170 reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
2171 reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
2172 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);
2174 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2176 map->frame_reg = cfg->frame_reg;
2177 map->start_offset = gcfg->min_offset;
2178 map->end_offset = gcfg->min_offset + (nslots * SIZEOF_SLOT);
2179 map->map_offset = start * SIZEOF_SLOT;
2180 map->nslots = end - start;
2181 map->has_ref_slots = has_ref_slots;
2182 map->has_pin_slots = has_pin_slots;
2183 map->has_ref_regs = has_ref_regs;
2184 map->has_pin_regs = has_pin_regs;
2185 g_assert (nregs < 32);
2186 map->used_int_regs = cfg->used_int_regs;
2187 map->reg_ref_mask = reg_ref_mask;
2188 map->reg_pin_mask = reg_pin_mask;
2189 map->nref_regs = nref_regs;
2190 map->npin_regs = npin_regs;
2192 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2195 if (has_ref_slots) {
2196 map->stack_ref_bitmap_offset = bitmaps_offset;
2197 bitmaps_offset += stack_bitmap_size;
2199 bitmap = &bitmaps [map->stack_ref_bitmap_offset];
2200 for (i = 0; i < nslots; ++i) {
2201 for (j = 0; j < ncallsites; ++j) {
2202 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
2203 set_bit (bitmap, stack_bitmap_width, j, i - start);
2207 if (has_pin_slots) {
2208 map->stack_pin_bitmap_offset = bitmaps_offset;
2209 bitmaps_offset += stack_bitmap_size;
2211 bitmap = &bitmaps [map->stack_pin_bitmap_offset];
2212 for (i = 0; i < nslots; ++i) {
2213 for (j = 0; j < ncallsites; ++j) {
2214 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
2215 set_bit (bitmap, stack_bitmap_width, j, i - start);
2220 map->reg_ref_bitmap_offset = bitmaps_offset;
2221 bitmaps_offset += reg_ref_bitmap_size;
2223 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2225 for (i = 0; i < nregs; ++i) {
2226 if (reg_ref_mask & (1 << i)) {
2227 for (j = 0; j < ncallsites; ++j) {
2228 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i))
2229 set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
2236 map->reg_pin_bitmap_offset = bitmaps_offset;
2237 bitmaps_offset += reg_pin_bitmap_size;
2239 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2241 for (i = 0; i < nregs; ++i) {
2242 if (reg_pin_mask & (1 << i)) {
2243 for (j = 0; j < ncallsites; ++j) {
2244 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i))
2245 set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
2253 map->ncallsites = ncallsites;
2254 if (cfg->code_len < 256)
2255 map->callsite_entry_size = 1;
2256 else if (cfg->code_len < 65536)
2257 map->callsite_entry_size = 2;
2259 map->callsite_entry_size = 4;
2261 /* Encode the GC Map */
2269 encode_gc_map (map, buf, &endbuf);
2270 g_assert (endbuf - buf < 256);
2272 encoded_size = endbuf - buf;
2273 alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
2275 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2276 //emap->ref_slots = map->ref_slots;
2278 /* Encoded fixed fields */
2279 p = &emap->encoded [0];
2280 //emap->encoded_size = encoded_size;
2281 memcpy (p, buf, encoded_size);
2284 /* Callsite table */
2285 p = (guint8*)ALIGN_TO ((mgreg_t)p, map->callsite_entry_size);
2286 if (map->callsite_entry_size == 1) {
2287 guint8 *offsets = p;
2288 for (i = 0; i < ncallsites; ++i)
2289 offsets [i] = callsites [i]->pc_offset;
2290 stats.gc_callsites8_size += ncallsites * sizeof (guint8);
2291 } else if (map->callsite_entry_size == 2) {
2292 guint16 *offsets = (guint16*)p;
2293 for (i = 0; i < ncallsites; ++i)
2294 offsets [i] = callsites [i]->pc_offset;
2295 stats.gc_callsites16_size += ncallsites * sizeof (guint16);
2297 guint32 *offsets = (guint32*)p;
2298 for (i = 0; i < ncallsites; ++i)
2299 offsets [i] = callsites [i]->pc_offset;
2300 stats.gc_callsites32_size += ncallsites * sizeof (guint32);
2302 p += ncallsites * map->callsite_entry_size;
2305 memcpy (p, bitmaps, bitmaps_size);
2308 g_assert ((guint8*)p - (guint8*)emap <= alloc_size);
2310 stats.gc_maps_size += alloc_size;
2311 stats.gc_callsites_size += ncallsites * map->callsite_entry_size;
2312 stats.gc_bitmaps_size += bitmaps_size;
2313 stats.gc_map_struct_size += sizeof (GCEncodedMap) + encoded_size;
2315 cfg->jit_info->gc_info = emap;
2317 cfg->gc_map = (guint8*)emap;
2318 cfg->gc_map_size = alloc_size;
2321 stats.all_slots += nslots;
2322 stats.ref_slots += ntypes [SLOT_REF];
2323 stats.noref_slots += ntypes [SLOT_NOREF];
2324 stats.pin_slots += ntypes [SLOT_PIN];
2328 mini_gc_create_gc_map (MonoCompile *cfg)
2330 if (!cfg->compute_gc_maps)
2334 * During marking, all frames except the top frame are at a call site, and we mark the
2335 * top frame conservatively. This means that we only need to compute and record
2336 * GC maps for call sites.
2339 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2340 /* Without liveness info, the live ranges are not precise enough */
2343 mono_analyze_liveness_gc (cfg);
2345 compute_frame_size (cfg);
2349 process_spill_slots (cfg);
2350 process_other_slots (cfg);
2351 process_param_area_slots (cfg);
2352 process_variables (cfg);
2353 process_finally_clauses (cfg);
2359 parse_debug_options (void)
2364 env = getenv ("MONO_GCMAP_DEBUG");
2368 opts = g_strsplit (env, ",", -1);
2369 for (ptr = opts; ptr && *ptr; ptr ++) {
2370 /* No options yet */
2371 fprintf (stderr, "Invalid format for the MONO_GCMAP_DEBUG env variable: '%s'\n", env);
2382 memset (&cb, 0, sizeof (cb));
2383 cb.thread_attach_func = thread_attach_func;
2384 cb.thread_detach_func = thread_detach_func;
2385 cb.thread_suspend_func = thread_suspend_func;
2386 /* Comment this out to disable precise stack marking */
2387 cb.thread_mark_func = thread_mark_func;
2388 mono_gc_set_gc_callbacks (&cb);
2390 logfile = mono_gc_get_logfile ();
2392 parse_debug_options ();
2394 mono_counters_register ("GC Maps size",
2395 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
2396 mono_counters_register ("GC Call Sites size",
2397 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
2398 mono_counters_register ("GC Bitmaps size",
2399 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
2400 mono_counters_register ("GC Map struct size",
2401 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
2402 mono_counters_register ("GC Call Sites encoded using 8 bits",
2403 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
2404 mono_counters_register ("GC Call Sites encoded using 16 bits",
2405 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
2406 mono_counters_register ("GC Call Sites encoded using 32 bits",
2407 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
2409 mono_counters_register ("GC Map slots (all)",
2410 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
2411 mono_counters_register ("GC Map slots (ref)",
2412 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.ref_slots);
2413 mono_counters_register ("GC Map slots (noref)",
2414 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.noref_slots);
2415 mono_counters_register ("GC Map slots (pin)",
2416 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
2418 mono_counters_register ("GC TLS Data size",
2419 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2421 mono_counters_register ("Stack space scanned (all)",
2422 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
2423 mono_counters_register ("Stack space scanned (native)",
2424 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_native);
2425 mono_counters_register ("Stack space scanned (other)",
2426 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_other);
2427 mono_counters_register ("Stack space scanned (using GC Maps)",
2428 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
2429 mono_counters_register ("Stack space scanned (precise)",
2430 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_precisely);
2431 mono_counters_register ("Stack space scanned (pin)",
2432 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_conservatively);
2433 mono_counters_register ("Stack space scanned (pin registers)",
2434 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_registers);
2445 mini_gc_init_gc_map (MonoCompile *cfg)
2450 mini_gc_create_gc_map (MonoCompile *cfg)
2455 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2460 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2469 * Set GC specific options in CFG.
2472 mini_gc_init_cfg (MonoCompile *cfg)
2474 if (mono_gc_is_moving ()) {
2475 cfg->disable_ref_noref_stack_slot_share = TRUE;
2476 cfg->gen_write_barriers = TRUE;
2479 mini_gc_init_gc_map (cfg);
2483 * Problems with the current code:
2484 * - the stack walk is slow
2485 * - vtypes/refs used in EH regions are treated conservatively
2486 * - if the code is finished, less pinning will be done, causing problems because
2487 * we promote all surviving objects to old-gen.
2488 * - the unwind code can't handle a method stopped inside a finally region, it thinks the caller is
2489 * another method, but in reality it is either the exception handling code or the CALL_HANDLER opcode.
2490 * This manifests in "Unable to find ip offset x in callsite list" assertions.
2491 * - the unwind code also can't handle frames which are in the epilog, since the unwind info is not
2496 * Ideas for creating smaller GC maps:
2497 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2499 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2501 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2502 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2503 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2504 * from an index entry.