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.
126 MonoThreadUnwindState unwind_state;
127 MonoThreadInfo *info;
130 gpointer ref_to_track;
131 /* Number of frames collected during the !precise pass */
133 FrameInfo frames [MAX_FRAMES];
136 /* These are constant so don't store them in the GC Maps */
137 /* Number of registers stored in gc maps */
138 #define NREGS MONO_MAX_IREGS
142 * Contains information needed to mark a stack frame.
143 * This is a transient structure, created from a compressed representation on-demand.
147 * The offsets of the GC tracked area inside the stack frame relative to the frame pointer.
148 * This includes memory which is NOREF thus doesn't need GC maps.
153 * The offset relative to frame_offset where the the memory described by the GC maps
157 /* The number of stack slots in the map */
159 /* The frame pointer register */
161 /* The size of each callsite table entry */
162 guint8 callsite_entry_size;
163 guint has_pin_slots : 1;
164 guint has_ref_slots : 1;
165 guint has_ref_regs : 1;
166 guint has_pin_regs : 1;
168 /* The offsets below are into an external bitmaps array */
171 * A bitmap whose width is equal to bitmap_width, and whose height is equal to ncallsites.
172 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
175 guint32 stack_ref_bitmap_offset;
177 * Same for SLOT_PIN. It is possible that the same bit is set in both bitmaps at
178 * different callsites, if the slot starts out as PIN, and later changes to REF.
180 guint32 stack_pin_bitmap_offset;
183 * Corresponding bitmaps for registers
184 * These have width equal to the number of bits set in reg_ref_mask/reg_pin_mask.
185 * FIXME: Merge these with the normal bitmaps, i.e. reserve the first x slots for them ?
187 guint32 reg_pin_bitmap_offset;
188 guint32 reg_ref_bitmap_offset;
190 guint32 used_int_regs, reg_ref_mask, reg_pin_mask;
192 /* The number of bits set in the two masks above */
193 guint8 nref_regs, npin_regs;
196 * A bit array marking slots which contain refs.
197 * This is used only for debugging.
201 /* Callsite offsets */
202 /* These can take up a lot of space, so encode them compactly */
212 * A compressed version of GCMap. This is what gets stored in MonoJitInfo.
216 //guint8 encoded_size;
219 * The arrays below are embedded after the struct.
220 * Their address needs to be computed.
223 /* The fixed fields of the GCMap encoded using LEB128 */
224 guint8 encoded [MONO_ZERO_LEN_ARRAY];
226 /* An array of ncallsites entries, each entry is callsite_entry_size bytes long */
227 guint8 callsites [MONO_ZERO_LEN_ARRAY];
230 guint8 bitmaps [MONO_ZERO_LEN_ARRAY];
233 static int precise_frame_count [2], precise_frame_limit = -1;
234 static gboolean precise_frame_limit_inited;
240 int scanned_precisely;
241 int scanned_conservatively;
242 int scanned_registers;
252 int gc_callsites_size;
253 int gc_callsites8_size;
254 int gc_callsites16_size;
255 int gc_callsites32_size;
257 int gc_map_struct_size;
261 static JITGCStats stats;
263 static FILE *logfile;
265 // FIXME: Move these to a shared place
268 encode_uleb128 (guint32 value, guint8 *buf, guint8 **endbuf)
273 guint8 b = value & 0x7f;
275 if (value != 0) /* more bytes to come */
283 static G_GNUC_UNUSED void
284 encode_sleb128 (gint32 value, guint8 *buf, guint8 **endbuf)
287 gboolean negative = (value < 0);
295 /* the following is unnecessary if the
296 * implementation of >>= uses an arithmetic rather
297 * than logical shift for a signed left operand
301 value |= - (1 <<(size - 7));
302 /* sign bit of byte is second high order bit (0x40) */
303 if ((value == 0 && !(byte & 0x40)) ||
304 (value == -1 && (byte & 0x40)))
314 static inline guint32
315 decode_uleb128 (guint8 *buf, guint8 **endbuf)
325 res = res | (((int)(b & 0x7f)) << shift);
337 decode_sleb128 (guint8 *buf, guint8 **endbuf)
347 res = res | (((int)(b & 0x7f)) << shift);
350 if (shift < 32 && (b & 0x40))
351 res |= - (1 << shift);
362 encode_frame_reg (int frame_reg)
365 if (frame_reg == AMD64_RSP)
367 else if (frame_reg == AMD64_RBP)
369 #elif defined(TARGET_X86)
370 if (frame_reg == X86_EBP)
372 else if (frame_reg == X86_ESP)
374 #elif defined(TARGET_ARM)
375 if (frame_reg == ARMREG_SP)
377 else if (frame_reg == ARMREG_FP)
382 g_assert_not_reached ();
387 decode_frame_reg (int encoded)
392 else if (encoded == 1)
394 #elif defined(TARGET_X86)
397 else if (encoded == 1)
399 #elif defined(TARGET_ARM)
402 else if (encoded == 1)
407 g_assert_not_reached ();
413 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15, AMD64_RDI, AMD64_RSI };
415 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15 };
417 #elif defined(TARGET_X86)
418 static int callee_saved_regs [] = { X86_EBX, X86_ESI, X86_EDI };
419 #elif defined(TARGET_ARM)
420 static int callee_saved_regs [] = { ARMREG_V1, ARMREG_V2, ARMREG_V3, ARMREG_V4, ARMREG_V5, ARMREG_V7, ARMREG_FP };
424 encode_regmask (guint32 regmask)
430 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
431 if (regmask & (1 << callee_saved_regs [i])) {
433 regmask -= (1 << callee_saved_regs [i]);
436 g_assert (regmask == 0);
441 decode_regmask (guint32 regmask)
447 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i)
448 if (regmask & (1 << i))
449 res |= (1 << callee_saved_regs [i]);
456 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
459 encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
463 encode_sleb128 (map->start_offset / SIZEOF_SLOT, buf, &buf);
464 encode_sleb128 (map->end_offset / SIZEOF_SLOT, buf, &buf);
465 encode_sleb128 (map->map_offset / SIZEOF_SLOT, buf, &buf);
466 encode_uleb128 (map->nslots, buf, &buf);
467 g_assert (map->callsite_entry_size <= 4);
468 freg = encode_frame_reg (map->frame_reg);
470 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);
471 encode_uleb128 (flags, buf, &buf);
472 encode_uleb128 (encode_regmask (map->used_int_regs), buf, &buf);
473 if (map->has_ref_regs)
474 encode_uleb128 (encode_regmask (map->reg_ref_mask), buf, &buf);
475 if (map->has_pin_regs)
476 encode_uleb128 (encode_regmask (map->reg_pin_mask), buf, &buf);
477 encode_uleb128 (map->ncallsites, buf, &buf);
485 * Decode the encoded GC map representation in BUF and store the result into MAP.
488 decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
491 int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
494 map->start_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
495 map->end_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
496 map->map_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
497 map->nslots = decode_uleb128 (buf, &buf);
498 flags = decode_uleb128 (buf, &buf);
499 map->has_ref_slots = (flags & 1) ? 1 : 0;
500 map->has_pin_slots = (flags & 2) ? 1 : 0;
501 map->has_ref_regs = (flags & 4) ? 1 : 0;
502 map->has_pin_regs = (flags & 8) ? 1 : 0;
503 map->callsite_entry_size = ((flags >> 4) & 0x3) + 1;
505 map->frame_reg = decode_frame_reg (freg);
506 map->used_int_regs = decode_regmask (decode_uleb128 (buf, &buf));
507 if (map->has_ref_regs) {
508 map->reg_ref_mask = decode_regmask (decode_uleb128 (buf, &buf));
510 for (i = 0; i < NREGS; ++i)
511 if (map->reg_ref_mask & (1 << i))
515 if (map->has_pin_regs) {
516 map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
518 for (i = 0; i < NREGS; ++i)
519 if (map->reg_pin_mask & (1 << i))
523 map->ncallsites = decode_uleb128 (buf, &buf);
525 stack_bitmap_size = (ALIGN_TO (map->nslots, 8) / 8) * map->ncallsites;
526 reg_ref_bitmap_size = (ALIGN_TO (map->nref_regs, 8) / 8) * map->ncallsites;
527 reg_pin_bitmap_size = (ALIGN_TO (map->npin_regs, 8) / 8) * map->ncallsites;
529 map->stack_ref_bitmap_offset = offset;
530 if (map->has_ref_slots)
531 offset += stack_bitmap_size;
532 map->stack_pin_bitmap_offset = offset;
533 if (map->has_pin_slots)
534 offset += stack_bitmap_size;
535 map->reg_ref_bitmap_offset = offset;
536 if (map->has_ref_regs)
537 offset += reg_ref_bitmap_size;
538 map->reg_pin_bitmap_offset = offset;
539 if (map->has_pin_regs)
540 offset += reg_pin_bitmap_size;
546 thread_attach_func (void)
550 tls = g_new0 (TlsData, 1);
551 tls->tid = GetCurrentThreadId ();
552 tls->info = mono_thread_info_current ();
553 stats.tlsdata_size += sizeof (TlsData);
559 thread_detach_func (gpointer user_data)
561 TlsData *tls = user_data;
567 thread_suspend_func (gpointer user_data, void *sigctx)
569 TlsData *tls = user_data;
572 /* Happens during startup */
573 tls->unwind_state.valid = FALSE;
577 if (tls->tid != GetCurrentThreadId ()) {
578 /* Happens on osx because threads are not suspended using signals */
581 g_assert (tls->info);
582 res = mono_thread_state_init_from_handle (&tls->unwind_state, (MonoNativeThreadId)tls->tid, tls->info->native_handle);
584 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF] = mono_get_lmf ();
586 mono_arch_sigctx_to_monoctx (sigctx, &tls->unwind_state.ctx);
587 tls->unwind_state.valid = TRUE;
589 tls->unwind_state.valid = FALSE;
591 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_JIT_TLS] = mono_native_tls_get_value (mono_jit_tls_id);
592 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN] = mono_domain_get ();
595 if (!tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN]) {
596 /* Happens during startup */
597 tls->unwind_state.valid = FALSE;
602 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
605 set_bit (guint8 *bitmap, int width, int y, int x)
607 bitmap [(width * y) + (x / 8)] |= (1 << (x % 8));
611 clear_bit (guint8 *bitmap, int width, int y, int x)
613 bitmap [(width * y) + (x / 8)] &= ~(1 << (x % 8));
617 get_bit (guint8 *bitmap, int width, int y, int x)
619 return bitmap [(width * y) + (x / 8)] & (1 << (x % 8));
623 slot_type_to_string (GCSlotType type)
633 g_assert_not_reached ();
638 static inline mgreg_t
639 get_frame_pointer (MonoContext *ctx, int frame_reg)
641 #if defined(TARGET_AMD64)
642 if (frame_reg == AMD64_RSP)
644 else if (frame_reg == AMD64_RBP)
646 #elif defined(TARGET_X86)
647 if (frame_reg == X86_ESP)
649 else if (frame_reg == X86_EBP)
651 #elif defined(TARGET_ARM)
652 if (frame_reg == ARMREG_SP)
653 return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
654 else if (frame_reg == ARMREG_FP)
655 return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
657 g_assert_not_reached ();
662 * conservatively_pass:
664 * Mark a thread stack conservatively and collect information needed by the precise pass.
667 conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
670 MonoContext ctx, new_ctx;
673 gboolean last = TRUE;
677 guint8* fp, *p, *real_frame_start, *frame_start, *frame_end;
678 int i, pc_offset, cindex, bitmap_width;
679 int scanned = 0, scanned_precisely, scanned_conservatively, scanned_registers;
681 StackFrameInfo frame;
682 mgreg_t *reg_locations [MONO_MAX_IREGS];
683 mgreg_t *new_reg_locations [MONO_MAX_IREGS];
686 guint32 precise_regmask;
690 tls->ref_to_track = NULL;
693 /* tls == NULL can happen during startup */
694 if (mono_thread_internal_current () == NULL || !tls) {
695 mono_gc_conservatively_scan_area (stack_start, stack_end);
696 stats.scanned_stacks += stack_end - stack_start;
700 lmf = tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF];
703 /* Number of bytes scanned based on GC map data */
705 /* Number of bytes scanned precisely based on GC map data */
706 scanned_precisely = 0;
707 /* Number of bytes scanned conservatively based on GC map data */
708 scanned_conservatively = 0;
709 /* Number of bytes scanned conservatively in register save areas */
710 scanned_registers = 0;
712 /* This is one past the last address which we have scanned */
713 stack_limit = stack_start;
715 if (!tls->unwind_state.valid)
716 memset (&new_ctx, 0, sizeof (ctx));
718 memcpy (&new_ctx, &tls->unwind_state.ctx, sizeof (MonoContext));
720 memset (reg_locations, 0, sizeof (reg_locations));
721 memset (new_reg_locations, 0, sizeof (new_reg_locations));
724 if (!tls->unwind_state.valid)
727 memcpy (&ctx, &new_ctx, sizeof (ctx));
729 for (i = 0; i < MONO_MAX_IREGS; ++i) {
730 if (new_reg_locations [i]) {
732 * If the current frame saves the register, it means it might modify its
733 * value, thus the old location might not contain the same value, so
734 * we have to mark it conservatively.
736 if (reg_locations [i]) {
737 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
738 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
739 scanned_registers += SIZEOF_SLOT;
742 reg_locations [i] = new_reg_locations [i];
744 DEBUG (fprintf (logfile, "\treg %s is now at location %p.\n", mono_arch_regname (i), reg_locations [i]));
748 g_assert ((mgreg_t)stack_limit % SIZEOF_SLOT == 0);
750 res = mono_find_jit_info_ext (frame.domain ? frame.domain : tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN], tls->unwind_state.unwind_data [MONO_UNWIND_DATA_JIT_TLS], NULL, &ctx, &new_ctx, NULL, &lmf, new_reg_locations, &frame);
756 if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
758 * These frames are problematic for several reasons:
759 * - they are unwound through an LMF, and we have no precise register tracking for those.
760 * - the LMF might not contain a precise ip, so we can't compute the call site.
761 * - the LMF only unwinds to the wrapper frame, so we get these methods twice.
763 DEBUG (fprintf (logfile, "Mark(0): <Managed-to-native transition>\n"));
764 for (i = 0; i < MONO_MAX_IREGS; ++i) {
765 if (reg_locations [i]) {
766 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
767 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
768 scanned_registers += SIZEOF_SLOT;
770 reg_locations [i] = NULL;
771 new_reg_locations [i] = NULL;
777 /* The last frame can be in any state so mark conservatively */
780 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));
782 DEBUG (fprintf (logfile, "\t <Last frame>\n"));
787 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
789 /* These frames are very problematic */
790 if (ji->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
791 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));
792 DEBUG (fprintf (logfile, "\tSkip.\n"));
796 /* All the other frames are at a call site */
798 if (tls->nframes == MAX_FRAMES) {
800 * Can't save information since the array is full. So scan the rest of the
801 * stack conservatively.
803 DEBUG (fprintf (logfile, "Mark (0): Frame stack full.\n"));
807 /* Scan the frame of this method */
810 * A frame contains the following:
815 * - localloc-ed memory
817 g_assert (pc_offset >= 0);
822 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));
823 DEBUG (fprintf (logfile, "\tNo GC Map.\n"));
827 /* The embedded callsite table requires this */
828 g_assert (((mgreg_t)emap % 4) == 0);
831 * Debugging aid to control the number of frames scanned precisely
833 if (!precise_frame_limit_inited) {
834 if (getenv ("MONO_PRECISE_COUNT"))
835 precise_frame_limit = atoi (getenv ("MONO_PRECISE_COUNT"));
836 precise_frame_limit_inited = TRUE;
839 if (precise_frame_limit != -1) {
840 if (precise_frame_count [FALSE] == precise_frame_limit)
841 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (ji->method, TRUE));
842 if (precise_frame_count [FALSE] > precise_frame_limit)
845 precise_frame_count [FALSE] ++;
847 /* Decode the encoded GC map */
849 memset (map, 0, sizeof (GCMap));
850 decode_gc_map (&emap->encoded [0], map, &p);
851 p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
852 map->callsites.offsets8 = p;
853 p += map->callsite_entry_size * map->ncallsites;
856 fp = (guint8*)get_frame_pointer (&ctx, map->frame_reg);
858 real_frame_start = fp + map->start_offset;
859 frame_start = fp + map->start_offset + map->map_offset;
860 frame_end = fp + map->end_offset;
862 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));
864 /* Find the callsite index */
865 if (map->callsite_entry_size == 1) {
866 for (i = 0; i < map->ncallsites; ++i)
867 /* ip points inside the call instruction */
868 if (map->callsites.offsets8 [i] == pc_offset + 1)
870 } else if (map->callsite_entry_size == 2) {
871 // FIXME: Use a binary search
872 for (i = 0; i < map->ncallsites; ++i)
873 /* ip points inside the call instruction */
874 if (map->callsites.offsets16 [i] == pc_offset + 1)
877 // FIXME: Use a binary search
878 for (i = 0; i < map->ncallsites; ++i)
879 /* ip points inside the call instruction */
880 if (map->callsites.offsets32 [i] == pc_offset + 1)
883 if (i == map->ncallsites) {
884 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (ji->method, TRUE));
885 g_assert_not_reached ();
890 * This is not neccessary true on x86 because frames have a different size at each
893 //g_assert (real_frame_start >= stack_limit);
895 if (real_frame_start > stack_limit) {
896 /* This scans the previously skipped frames as well */
897 DEBUG (fprintf (logfile, "\tscan area %p-%p (%d).\n", stack_limit, real_frame_start, (int)(real_frame_start - stack_limit)));
898 mono_gc_conservatively_scan_area (stack_limit, real_frame_start);
899 stats.scanned_other += real_frame_start - stack_limit;
902 /* Mark stack slots */
903 if (map->has_pin_slots) {
904 int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
905 guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
910 for (i = 0; i < map->nslots; ++i) {
911 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
913 DEBUG (fprintf (logfile, "\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
914 mono_gc_conservatively_scan_area (p, p + SIZEOF_SLOT);
915 scanned_conservatively += SIZEOF_SLOT;
917 scanned_precisely += SIZEOF_SLOT;
922 scanned_precisely += (map->nslots * SIZEOF_SLOT);
925 /* The area outside of start-end is NOREF */
926 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * SIZEOF_SLOT);
929 precise_regmask = map->used_int_regs | (1 << map->frame_reg);
930 if (map->has_pin_regs) {
931 int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
932 guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
934 for (i = 0; i < NREGS; ++i) {
935 if (!(map->used_int_regs & (1 << i)))
938 if (!(map->reg_pin_mask & (1 << i)))
941 if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
942 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
943 precise_regmask &= ~(1 << i);
949 scanned += map->end_offset - map->start_offset;
951 g_assert (scanned == scanned_precisely + scanned_conservatively);
953 stack_limit = frame_end;
955 /* Save information for the precise pass */
956 fi = &tls->frames [tls->nframes];
957 fi->nslots = map->nslots;
958 bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
959 if (map->has_ref_slots)
960 fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
963 fi->frame_start_offset = frame_start - stack_start;
964 fi->nreg_locations = 0;
965 DEBUG_PRECISE (fi->ji = ji);
966 DEBUG_PRECISE (fi->fp = fp);
968 if (map->has_ref_regs) {
969 int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
970 guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
972 for (i = 0; i < NREGS; ++i) {
973 if (!(map->reg_ref_mask & (1 << i)))
976 if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
977 DEBUG_PRECISE (fi->regs [fi->nreg_locations] = i);
978 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
979 fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
980 fi->nreg_locations ++;
987 * Clear locations of precisely stacked registers.
989 if (precise_regmask) {
990 for (i = 0; i < NREGS; ++i) {
991 if (precise_regmask & (1 << i)) {
993 * The method uses this register, and we have precise info for it.
994 * This means the location will be scanned precisely.
995 * Tell the code at the beginning of the loop that this location is
998 if (reg_locations [i])
999 DEBUG (fprintf (logfile, "\treg %s at location %p (==%p) is precise.\n", mono_arch_regname (i), reg_locations [i], (gpointer)*reg_locations [i]));
1000 reg_locations [i] = NULL;
1008 /* Scan the remaining register save locations */
1009 for (i = 0; i < MONO_MAX_IREGS; ++i) {
1010 if (reg_locations [i]) {
1011 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", reg_locations [i]));
1012 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
1013 scanned_registers += SIZEOF_SLOT;
1015 if (new_reg_locations [i]) {
1016 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", new_reg_locations [i]));
1017 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
1018 scanned_registers += SIZEOF_SLOT;
1022 if (stack_limit < stack_end) {
1023 DEBUG (fprintf (logfile, "\tscan remaining stack %p-%p (%d).\n", stack_limit, stack_end, (int)(stack_end - stack_limit)));
1024 mono_gc_conservatively_scan_area (stack_limit, stack_end);
1025 stats.scanned_native += stack_end - stack_limit;
1028 DEBUG (fprintf (logfile, "Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
1030 stats.scanned_stacks += stack_end - stack_start;
1031 stats.scanned += scanned;
1032 stats.scanned_precisely += scanned_precisely;
1033 stats.scanned_conservatively += scanned_conservatively;
1034 stats.scanned_registers += scanned_registers;
1036 //mono_gc_conservatively_scan_area (stack_start, stack_end);
1042 * Mark a thread stack precisely based on information saved during the conservative
1046 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
1050 guint8 *frame_start;
1055 if (!tls->unwind_state.valid)
1058 for (findex = 0; findex < tls->nframes; findex ++) {
1059 /* Load information saved by the !precise pass */
1060 fi = &tls->frames [findex];
1061 frame_start = stack_start + fi->frame_start_offset;
1063 DEBUG (char *fname = mono_method_full_name (fi->ji->method, TRUE); fprintf (logfile, "Mark(1): %s\n", fname); g_free (fname));
1066 * FIXME: Add a function to mark using a bitmap, to avoid doing a
1067 * call for each object.
1070 /* Mark stack slots */
1072 guint8 *ref_bitmap = fi->bitmap;
1075 for (i = 0; i < fi->nslots; ++i) {
1076 MonoObject **ptr = (MonoObject**)(frame_start + (i * SIZEOF_SLOT));
1078 live = ref_bitmap [i / 8] & (1 << (i % 8));
1081 MonoObject *obj = *ptr;
1083 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1084 *ptr = mono_gc_scan_object (obj);
1085 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1087 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1092 * This is disabled because the pointer takes up a lot of space.
1093 * Stack slots might be shared between ref and non-ref variables ?
1095 if (map->ref_slots [i / 8] & (1 << (i % 8))) {
1096 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));
1098 * Fail fast if the live range is incorrect, and
1099 * the JITted code tries to access this object
1108 /* Mark registers */
1111 * Registers are different from stack slots, they have no address where they
1112 * are stored. Instead, some frame below this frame in the stack saves them
1113 * in its prolog to the stack. We can mark this location precisely.
1115 for (i = 0; i < fi->nreg_locations; ++i) {
1117 * reg_locations [i] contains the address of the stack slot where
1118 * a reg was last saved, so mark that slot.
1120 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1121 MonoObject *obj = *ptr;
1124 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
1125 *ptr = mono_gc_scan_object (obj);
1126 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1128 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p\n", mono_arch_regname (fi->regs [i]), ptr, obj));
1134 * Debugging aid to check for missed refs.
1136 if (tls->ref_to_track) {
1139 for (p = (mgreg_t*)stack_start; p < (mgreg_t*)stack_end; ++p)
1140 if (*p == (mgreg_t)tls->ref_to_track)
1141 printf ("REF AT %p.\n", p);
1148 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1149 * call, and TRUE at the second. USER_DATA points to a TlsData
1150 * structure filled up by thread_suspend_func.
1153 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise)
1155 TlsData *tls = user_data;
1157 DEBUG (fprintf (logfile, "****************************************\n"));
1158 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));
1159 DEBUG (fprintf (logfile, "****************************************\n"));
1162 conservative_pass (tls, stack_start, stack_end);
1164 precise_pass (tls, stack_start, stack_end);
1168 mini_gc_init_gc_map (MonoCompile *cfg)
1170 if (COMPILE_LLVM (cfg))
1173 if (!mono_gc_is_moving ())
1176 if (!cfg->compile_aot && !mono_gc_precise_stack_mark_enabled ())
1180 /* Debugging support */
1182 static int precise_count;
1185 if (getenv ("MONO_GCMAP_COUNT")) {
1186 if (precise_count == atoi (getenv ("MONO_GCMAP_COUNT")))
1187 printf ("LAST: %s\n", mono_method_full_name (cfg->method, TRUE));
1188 if (precise_count > atoi (getenv ("MONO_GCMAP_COUNT")))
1194 cfg->compute_gc_maps = TRUE;
1196 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1200 * mini_gc_set_slot_type_from_fp:
1202 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1203 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1204 * need to call this function for those slots.
1207 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1209 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1211 if (!cfg->compute_gc_maps)
1214 g_assert (slot_offset % SIZEOF_SLOT == 0);
1216 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1220 * mini_gc_set_slot_type_from_cfa:
1222 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1223 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1224 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1225 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1226 * call this function for those slots.
1229 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1231 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1232 int slot = - (slot_offset / SIZEOF_SLOT);
1234 if (!cfg->compute_gc_maps)
1237 g_assert (slot_offset <= 0);
1238 g_assert (slot_offset % SIZEOF_SLOT == 0);
1240 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1244 fp_offset_to_slot (MonoCompile *cfg, int offset)
1246 MonoCompileGC *gcfg = cfg->gc_info;
1248 return (offset - gcfg->min_offset) / SIZEOF_SLOT;
1252 slot_to_fp_offset (MonoCompile *cfg, int slot)
1254 MonoCompileGC *gcfg = cfg->gc_info;
1256 return (slot * SIZEOF_SLOT) + gcfg->min_offset;
1260 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1262 g_assert (slot >= 0 && slot < gcfg->nslots);
1264 if (type == SLOT_PIN) {
1265 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1266 set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1267 } else if (type == SLOT_REF) {
1268 set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1269 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1270 } else if (type == SLOT_NOREF) {
1271 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1272 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, callsite_index, slot);
1277 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1281 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1282 set_slot (gcfg, slot, cindex, type);
1286 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1290 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1291 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1292 if (callsite_offset >= from && callsite_offset < to)
1293 set_slot (gcfg, slot, cindex, type);
1298 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1300 g_assert (slot >= 0 && slot < gcfg->nregs);
1302 if (type == SLOT_PIN) {
1303 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1304 set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1305 } else if (type == SLOT_REF) {
1306 set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1307 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1308 } else if (type == SLOT_NOREF) {
1309 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1310 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, callsite_index, slot);
1315 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1319 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1320 set_reg_slot (gcfg, slot, cindex, type);
1324 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1328 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1329 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1330 if (callsite_offset >= from && callsite_offset < to)
1331 set_reg_slot (gcfg, slot, cindex, type);
1336 process_spill_slots (MonoCompile *cfg)
1338 MonoCompileGC *gcfg = cfg->gc_info;
1343 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1344 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1345 for (l = bb->spill_slot_defs; l; l = l->next) {
1346 MonoInst *def = l->data;
1347 int spill_slot = def->inst_c0;
1348 int bank = def->inst_c1;
1349 int offset = cfg->spill_info [bank][spill_slot].offset;
1350 int slot = fp_offset_to_slot (cfg, offset);
1352 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1353 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1357 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1358 for (l = bb->spill_slot_defs; l; l = l->next) {
1359 MonoInst *def = l->data;
1360 int spill_slot = def->inst_c0;
1361 int bank = def->inst_c1;
1362 int offset = cfg->spill_info [bank][spill_slot].offset;
1363 int slot = fp_offset_to_slot (cfg, offset);
1366 if (bank == MONO_REG_INT_MP)
1372 * Extend the live interval for the GC tracked spill slots
1373 * defined in this bblock.
1374 * FIXME: This is not needed.
1376 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1378 if (cfg->verbose_level > 1)
1379 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
1383 /* Set fp spill slots to NOREF */
1384 for (i = 0; i < cfg->spill_info_len [MONO_REG_DOUBLE]; ++i) {
1385 int offset = cfg->spill_info [MONO_REG_DOUBLE][i].offset;
1391 slot = fp_offset_to_slot (cfg, offset);
1393 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1395 if (cfg->verbose_level > 1)
1396 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1399 /* Set int spill slots to NOREF */
1400 for (i = 0; i < cfg->spill_info_len [MONO_REG_INT]; ++i) {
1401 int offset = cfg->spill_info [MONO_REG_INT][i].offset;
1407 slot = fp_offset_to_slot (cfg, offset);
1409 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1410 if (cfg->verbose_level > 1)
1411 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1416 * process_other_slots:
1418 * Process stack slots registered using mini_gc_set_slot_type_... ().
1421 process_other_slots (MonoCompile *cfg)
1423 MonoCompileGC *gcfg = cfg->gc_info;
1426 /* Relative to the CFA */
1427 for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
1428 guint data = GPOINTER_TO_UINT (l->data);
1429 int cfa_slot = data >> 16;
1430 GCSlotType type = data & 0xff;
1434 * Map the cfa relative slot to an fp relative slot.
1435 * slot_addr == cfa - <cfa_slot>*4/8
1436 * fp + cfa_offset == cfa
1437 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1439 slot = (cfg->cfa_offset / SIZEOF_SLOT) - cfa_slot - (gcfg->min_offset / SIZEOF_SLOT);
1441 set_slot_everywhere (gcfg, slot, type);
1443 if (cfg->verbose_level > 1) {
1444 int fp_offset = slot_to_fp_offset (cfg, slot);
1445 if (type == SLOT_NOREF)
1446 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));
1450 /* Relative to the FP */
1451 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1452 gint data = GPOINTER_TO_INT (l->data);
1453 int offset = data >> 16;
1454 GCSlotType type = data & 0xff;
1457 slot = fp_offset_to_slot (cfg, offset);
1459 set_slot_everywhere (gcfg, slot, type);
1461 /* Liveness for these slots is handled by process_spill_slots () */
1463 if (cfg->verbose_level > 1) {
1464 if (type == SLOT_REF)
1465 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
1466 else if (type == SLOT_NOREF)
1467 printf ("\tnoref slot at 0x%x(fp) (slot = %d)\n", offset, slot);
1473 get_vtype_bitmap (MonoType *t, int *numbits)
1475 MonoClass *klass = mono_class_from_mono_type (t);
1477 if (klass->generic_container || mono_class_is_open_constructed_type (t)) {
1478 /* FIXME: Generic sharing */
1481 mono_class_compute_gc_descriptor (klass);
1483 return mono_gc_get_bitmap_for_descr (klass->gc_descr, numbits);
1487 static inline const char*
1488 get_offset_sign (int offset)
1490 return offset < 0 ? "-" : "+";
1494 get_offset_val (int offset)
1496 return offset < 0 ? (- offset) : offset;
1500 process_variables (MonoCompile *cfg)
1502 MonoCompileGC *gcfg = cfg->gc_info;
1503 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1504 int i, locals_min_slot, locals_max_slot, cindex;
1508 int locals_min_offset = gcfg->locals_min_offset;
1509 int locals_max_offset = gcfg->locals_max_offset;
1511 /* Slots for locals are NOREF by default */
1512 locals_min_slot = (locals_min_offset - gcfg->min_offset) / SIZEOF_SLOT;
1513 locals_max_slot = (locals_max_offset - gcfg->min_offset) / SIZEOF_SLOT;
1514 for (i = locals_min_slot; i < locals_max_slot; ++i) {
1515 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1519 * Compute the offset where variables are initialized in the first bblock, if any.
1521 pc_offsets = g_new0 (int, cfg->next_vreg);
1523 bb = cfg->bb_entry->next_bb;
1524 MONO_BB_FOR_EACH_INS (bb, tmp) {
1525 if (tmp->opcode == OP_GC_LIVENESS_DEF) {
1526 int vreg = tmp->inst_c1;
1527 if (pc_offsets [vreg] == 0) {
1528 g_assert (tmp->backend.pc_offset > 0);
1529 pc_offsets [vreg] = tmp->backend.pc_offset;
1535 * Stack slots holding arguments are initialized in the prolog.
1536 * This means we can treat them alive for the whole method.
1538 for (i = 0; i < cfg->num_varinfo; i++) {
1539 MonoInst *ins = cfg->varinfo [i];
1540 MonoType *t = ins->inst_vtype;
1543 gboolean byref, is_this = FALSE;
1544 gboolean is_arg = i < cfg->locals_start;
1546 if (ins == cfg->ret) {
1547 if (!(ins->opcode == OP_REGOFFSET && MONO_TYPE_ISSTRUCT (t)))
1551 vmv = MONO_VARINFO (cfg, i);
1553 /* For some reason, 'this' is byref */
1554 if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
1555 t = &cfg->method->klass->byval_arg;
1561 if (ins->opcode == OP_REGVAR) {
1563 GCSlotType slot_type;
1565 t = mini_type_get_underlying_type (NULL, t);
1568 g_assert (hreg < MONO_MAX_IREGS);
1571 slot_type = SLOT_PIN;
1573 slot_type = mini_type_is_reference (cfg, t) ? SLOT_REF : SLOT_NOREF;
1575 if (slot_type == SLOT_PIN) {
1576 /* These have no live interval, be conservative */
1577 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1580 * Unlike variables allocated to the stack, we generate liveness info
1581 * for noref vars in registers in mono_spill_global_vars (), because
1582 * knowing that a register doesn't contain a ref allows us to mark its save
1583 * locations precisely.
1585 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1586 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1587 set_reg_slot (gcfg, hreg, cindex, slot_type);
1590 if (cfg->verbose_level > 1) {
1591 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
1597 if (ins->opcode != OP_REGOFFSET)
1600 if (ins->inst_offset % SIZEOF_SLOT != 0)
1603 if (is_arg && ins->inst_offset >= gcfg->max_offset)
1604 /* In parent frame */
1607 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1609 if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
1610 /* These do not get stored in the prolog */
1611 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1613 if (cfg->verbose_level > 1) {
1614 printf ("\tdead arg at fp%s0x%x (slot = %d): %s\n", get_offset_sign (ins->inst_offset), get_offset_val (ins->inst_offset), pos, mono_type_full_name (ins->inst_vtype));
1619 if (MONO_TYPE_ISSTRUCT (t)) {
1621 gsize *bitmap = NULL;
1622 gboolean pin = FALSE;
1626 if (ins->backend.is_pinvoke)
1627 size = mono_class_native_size (ins->klass, NULL);
1629 size = mono_class_value_size (ins->klass, NULL);
1630 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
1632 if (cfg->verbose_level > 1)
1633 printf ("\tvtype R%d at %s0x%x(fp)-%s0x%x(fp) (slot %d-%d): %s\n", vmv->vreg, get_offset_sign (ins->inst_offset), get_offset_val (ins->inst_offset), get_offset_sign (ins->inst_offset), get_offset_val (ins->inst_offset + (size_in_slots * SIZEOF_SLOT)), pos, pos + size_in_slots, mono_type_full_name (ins->inst_vtype));
1635 if (!ins->klass->has_references) {
1637 for (j = 0; j < size_in_slots; ++j)
1638 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1643 bitmap = get_vtype_bitmap (t, &numbits);
1648 * Most vtypes are marked volatile because of the LDADDR instructions,
1649 * and they have no liveness information since they are decomposed
1650 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1651 * them during VZERO decomposition.
1653 if (!pc_offsets [vmv->vreg])
1656 if (ins->backend.is_pinvoke)
1660 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1661 if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
1662 for (j = 0; j < numbits; ++j) {
1663 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1664 /* The descriptor is for the boxed object */
1665 set_slot (gcfg, (pos + j - (sizeof (MonoObject) / SIZEOF_SLOT)), cindex, pin ? SLOT_PIN : SLOT_REF);
1671 if (cfg->verbose_level > 1) {
1672 for (j = 0; j < numbits; ++j) {
1673 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD)))
1674 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)));
1678 if (cfg->verbose_level > 1)
1679 printf ("\t\tpinned\n");
1680 for (j = 0; j < size_in_slots; ++j) {
1681 set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
1690 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1691 /* Vret addr etc. */
1696 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1698 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1699 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1700 set_slot (gcfg, pos, cindex, SLOT_PIN);
1702 if (cfg->verbose_level > 1)
1703 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));
1708 * This is currently disabled, but could be enabled to debug crashes.
1711 if (t->type == MONO_TYPE_I) {
1713 * Variables created in mono_handle_global_vregs have type I, but they
1714 * could hold GC refs since the vregs they were created from might not been
1715 * marked as holding a GC ref. So be conservative.
1717 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1722 t = mini_type_get_underlying_type (NULL, t);
1724 if (!mini_type_is_reference (cfg, t)) {
1725 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1726 if (cfg->verbose_level > 1)
1727 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));
1728 if (!t->byref && sizeof (mgreg_t) == 4 && (t->type == MONO_TYPE_I8 || t->type == MONO_TYPE_U8 || t->type == MONO_TYPE_R8)) {
1729 set_slot_everywhere (gcfg, pos + 1, SLOT_NOREF);
1730 if (cfg->verbose_level > 1)
1731 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));
1736 /* 'this' is marked INDIRECT for gshared methods */
1737 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1739 * For volatile variables, treat them alive from the point they are
1740 * initialized in the first bblock until the end of the method.
1743 set_slot_everywhere (gcfg, pos, SLOT_REF);
1744 } else if (pc_offsets [vmv->vreg]) {
1745 set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
1746 set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
1748 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1750 if (cfg->verbose_level > 1)
1751 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));
1756 /* Live for the whole method */
1757 set_slot_everywhere (gcfg, pos, SLOT_REF);
1759 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1760 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1761 set_slot (gcfg, pos, cindex, SLOT_REF);
1764 if (cfg->verbose_level > 1) {
1765 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));
1769 g_free (pc_offsets);
1773 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1776 * Convert a sp relative offset to a slot index. This is
1777 * platform specific.
1780 /* fp = sp + offset */
1781 g_assert (cfg->frame_reg == AMD64_RBP);
1782 return (- cfg->arch.sp_fp_offset + sp_offset);
1783 #elif defined(TARGET_X86)
1784 /* The offset is computed from the sp at the start of the call sequence */
1785 g_assert (cfg->frame_reg == X86_EBP);
1786 return (- cfg->arch.sp_fp_offset - sp_offset);
1794 type_to_gc_slot_type (MonoCompile *cfg, MonoType *t)
1798 t = mini_type_get_underlying_type (NULL, t);
1799 if (mini_type_is_reference (cfg, t))
1802 if (MONO_TYPE_ISSTRUCT (t)) {
1803 MonoClass *klass = mono_class_from_mono_type (t);
1804 if (!klass->has_references) {
1816 process_param_area_slots (MonoCompile *cfg)
1818 MonoCompileGC *gcfg = cfg->gc_info;
1823 * These slots are used for passing parameters during calls. They are sp relative, not
1824 * fp relative, so they are harder to handle.
1826 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1827 /* The distance between fp and sp is not constant */
1830 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1832 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1833 GCCallSite *callsite = gcfg->callsites [cindex];
1836 for (l = callsite->param_slots; l; l = l->next) {
1837 MonoInst *def = l->data;
1838 MonoType *t = def->inst_vtype;
1839 int sp_offset = def->inst_offset;
1840 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1841 int slot = fp_offset_to_slot (cfg, fp_offset);
1845 if (MONO_TYPE_ISSTRUCT (t)) {
1846 size = mini_type_stack_size_full (cfg->generic_sharing_context, t, &align, FALSE);
1848 size = sizeof (mgreg_t);
1851 for (i = 0; i < size / sizeof (mgreg_t); ++i) {
1852 g_assert (slot + i >= 0 && slot + i < gcfg->nslots);
1853 is_param [slot + i] = TRUE;
1858 /* All param area slots are noref by default */
1859 for (i = 0; i < gcfg->nslots; ++i) {
1861 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1864 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1865 GCCallSite *callsite = gcfg->callsites [cindex];
1868 for (l = callsite->param_slots; l; l = l->next) {
1869 MonoInst *def = l->data;
1870 MonoType *t = def->inst_vtype;
1871 int sp_offset = def->inst_offset;
1872 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1873 int slot = fp_offset_to_slot (cfg, fp_offset);
1874 GCSlotType type = type_to_gc_slot_type (cfg, t);
1876 if (MONO_TYPE_ISSTRUCT (t)) {
1883 size = mini_type_stack_size_full (cfg->generic_sharing_context, t, &align, FALSE);
1884 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
1886 bitmap = get_vtype_bitmap (t, &numbits);
1887 if (type == SLOT_NOREF || !bitmap) {
1888 for (i = 0; i < size_in_slots; ++i) {
1889 set_slot_in_range (gcfg, slot + i, def->backend.pc_offset, callsite->pc_offset + 1, type);
1891 if (cfg->verbose_level > 1)
1892 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), get_offset_sign (fp_offset), get_offset_val (fp_offset), sp_offset, slot, slot + (size / (int)sizeof (mgreg_t)), def->backend.pc_offset, callsite->pc_offset + 1);
1894 for (j = 0; j < numbits; ++j) {
1895 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1896 /* The descriptor is for the boxed object */
1897 set_slot (gcfg, (slot + j - (sizeof (MonoObject) / SIZEOF_SLOT)), cindex, SLOT_REF);
1900 if (cfg->verbose_level > 1)
1901 printf ("\tvtype param area slots at %s0x%x(fp)=0x%x(sp) (slot = %d-%d) [0x%x-0x%x]\n", get_offset_sign (fp_offset), get_offset_val (fp_offset), sp_offset, slot, slot + (size / (int)sizeof (mgreg_t)), def->backend.pc_offset, callsite->pc_offset + 1);
1905 /* The slot is live between the def instruction and the call */
1906 set_slot_in_range (gcfg, slot, def->backend.pc_offset, callsite->pc_offset + 1, type);
1907 if (cfg->verbose_level > 1)
1908 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), get_offset_sign (fp_offset), get_offset_val (fp_offset), sp_offset, slot, def->backend.pc_offset, callsite->pc_offset + 1);
1915 process_finally_clauses (MonoCompile *cfg)
1917 MonoCompileGC *gcfg = cfg->gc_info;
1918 GCCallSite **callsites;
1920 gboolean has_finally;
1921 int i, j, nslots, nregs;
1923 ncallsites = gcfg->ncallsites;
1924 nslots = gcfg->nslots;
1925 nregs = gcfg->nregs;
1926 callsites = gcfg->callsites;
1929 * The calls to the finally clauses don't show up in the cfg. See
1930 * test_0_liveness_8 ().
1931 * Variables accessed inside the finally clause are already marked VOLATILE by
1932 * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
1933 * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
1935 has_finally = FALSE;
1936 for (i = 0; i < cfg->header->num_clauses; ++i) {
1937 MonoExceptionClause *clause = &cfg->header->clauses [i];
1939 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1944 if (cfg->verbose_level > 1)
1945 printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
1946 for (j = 0; j < ncallsites; ++j) {
1947 MonoBasicBlock *bb = callsites [j]->bb;
1948 MonoExceptionClause *clause;
1949 gboolean is_in_finally = FALSE;
1951 for (i = 0; i < cfg->header->num_clauses; ++i) {
1952 clause = &cfg->header->clauses [i];
1954 if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
1955 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1956 is_in_finally = TRUE;
1962 if (is_in_finally) {
1963 for (i = 0; i < nslots; ++i)
1964 set_slot (gcfg, i, j, SLOT_PIN);
1965 for (i = 0; i < nregs; ++i)
1966 set_reg_slot (gcfg, i, j, SLOT_PIN);
1973 compute_frame_size (MonoCompile *cfg)
1975 int i, locals_min_offset, locals_max_offset, cfa_min_offset, cfa_max_offset;
1976 int min_offset, max_offset;
1977 MonoCompileGC *gcfg = cfg->gc_info;
1978 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1981 /* Compute min/max offsets from the fp */
1984 #if defined(TARGET_AMD64) || defined(TARGET_X86) || defined(TARGET_ARM)
1985 locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, SIZEOF_SLOT);
1986 locals_max_offset = cfg->locals_max_stack_offset;
1988 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
1992 locals_min_offset = ALIGN_TO (locals_min_offset, SIZEOF_SLOT);
1993 locals_max_offset = ALIGN_TO (locals_max_offset, SIZEOF_SLOT);
1995 min_offset = locals_min_offset;
1996 max_offset = locals_max_offset;
1999 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
2000 MonoInst *ins = cfg->args [i];
2002 if (ins->opcode == OP_REGOFFSET)
2003 min_offset = MIN (min_offset, ins->inst_offset);
2007 g_assert (cfg->frame_reg == cfg->cfa_reg);
2008 g_assert (cfg->cfa_offset > 0);
2010 cfa_max_offset = cfg->cfa_offset;
2012 min_offset = MIN (min_offset, cfa_min_offset);
2013 max_offset = MAX (max_offset, cfa_max_offset);
2015 /* Fp relative slots */
2016 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
2017 gint data = GPOINTER_TO_INT (l->data);
2018 int offset = data >> 16;
2020 min_offset = MIN (min_offset, offset);
2024 if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
2025 int stack_offset = ALIGN_TO (cfg->stack_offset, SIZEOF_SLOT);
2026 min_offset = MIN (min_offset, (-stack_offset));
2029 /* Param area slots */
2031 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
2032 #elif defined(TARGET_X86)
2033 min_offset = MIN (min_offset, - (cfg->arch.sp_fp_offset + cfg->arch.param_area_size));
2034 #elif defined(TARGET_ARM)
2040 gcfg->min_offset = min_offset;
2041 gcfg->max_offset = max_offset;
2042 gcfg->locals_min_offset = locals_min_offset;
2043 gcfg->locals_max_offset = locals_max_offset;
2047 init_gcfg (MonoCompile *cfg)
2049 int i, nregs, nslots;
2050 MonoCompileGC *gcfg = cfg->gc_info;
2051 GCCallSite **callsites;
2060 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2061 ncallsites += g_slist_length (bb->gc_callsites);
2063 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
2065 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2066 for (l = bb->gc_callsites; l; l = l->next)
2067 callsites [i++] = l->data;
2070 /* The callsites should already be ordered by pc offset */
2071 for (i = 1; i < ncallsites; ++i)
2072 g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
2075 * The stack frame looks like this:
2077 * <fp + max_offset> == cfa -> <end of previous frame>
2078 * <other stack slots>
2080 * <other stack slots>
2081 * fp + min_offset ->
2086 if (cfg->verbose_level > 1)
2087 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
2089 nslots = (gcfg->max_offset - gcfg->min_offset) / SIZEOF_SLOT;
2092 gcfg->nslots = nslots;
2093 gcfg->nregs = nregs;
2094 gcfg->callsites = callsites;
2095 gcfg->ncallsites = ncallsites;
2096 gcfg->stack_bitmap_width = ALIGN_TO (nslots, 8) / 8;
2097 gcfg->reg_bitmap_width = ALIGN_TO (nregs, 8) / 8;
2098 gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
2099 gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * ncallsites);
2100 gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
2101 gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * ncallsites);
2103 /* All slots start out as PIN */
2104 memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * ncallsites);
2105 for (i = 0; i < nregs; ++i) {
2107 * By default, registers are NOREF.
2108 * It is possible for a callee to save them before being defined in this method,
2109 * but the saved value is dead too, so it doesn't need to be marked.
2111 if ((cfg->used_int_regs & (1 << i)))
2112 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
2117 create_map (MonoCompile *cfg)
2120 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
2122 int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
2123 int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
2125 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
2126 MonoCompileGC *gcfg = cfg->gc_info;
2127 GCCallSite **callsites;
2129 guint8 *bitmap, *bitmaps;
2130 guint32 reg_ref_mask, reg_pin_mask;
2132 ncallsites = gcfg->ncallsites;
2133 nslots = gcfg->nslots;
2134 nregs = gcfg->nregs;
2135 callsites = gcfg->callsites;
2138 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
2139 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
2141 has_ref_slots = FALSE;
2142 has_pin_slots = FALSE;
2145 memset (ntypes, 0, sizeof (ntypes));
2146 for (i = 0; i < nslots; ++i) {
2147 gboolean has_ref = FALSE;
2148 gboolean has_pin = FALSE;
2150 for (j = 0; j < ncallsites; ++j) {
2151 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
2153 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
2158 has_ref_slots = TRUE;
2160 has_pin_slots = TRUE;
2163 ntypes [SLOT_REF] ++;
2165 ntypes [SLOT_PIN] ++;
2167 ntypes [SLOT_NOREF] ++;
2169 if (has_ref || has_pin) {
2176 start = end = nslots;
2178 g_assert (start != -1);
2179 g_assert (start < end);
2182 has_ref_regs = FALSE;
2183 has_pin_regs = FALSE;
2188 for (i = 0; i < nregs; ++i) {
2189 gboolean has_ref = FALSE;
2190 gboolean has_pin = FALSE;
2192 if (!(cfg->used_int_regs & (1 << i)))
2195 for (j = 0; j < ncallsites; ++j) {
2196 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i)) {
2201 for (j = 0; j < ncallsites; ++j) {
2202 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i)) {
2209 reg_ref_mask |= (1 << i);
2210 has_ref_regs = TRUE;
2214 reg_pin_mask |= (1 << i);
2215 has_pin_regs = TRUE;
2220 if (cfg->verbose_level > 1)
2221 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);
2223 /* Create the GC Map */
2225 stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
2226 stack_bitmap_size = stack_bitmap_width * ncallsites;
2227 reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
2228 reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
2229 reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
2230 reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
2231 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);
2233 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2235 map->frame_reg = cfg->frame_reg;
2236 map->start_offset = gcfg->min_offset;
2237 map->end_offset = gcfg->min_offset + (nslots * SIZEOF_SLOT);
2238 map->map_offset = start * SIZEOF_SLOT;
2239 map->nslots = end - start;
2240 map->has_ref_slots = has_ref_slots;
2241 map->has_pin_slots = has_pin_slots;
2242 map->has_ref_regs = has_ref_regs;
2243 map->has_pin_regs = has_pin_regs;
2244 g_assert (nregs < 32);
2245 map->used_int_regs = cfg->used_int_regs;
2246 map->reg_ref_mask = reg_ref_mask;
2247 map->reg_pin_mask = reg_pin_mask;
2248 map->nref_regs = nref_regs;
2249 map->npin_regs = npin_regs;
2251 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2254 if (has_ref_slots) {
2255 map->stack_ref_bitmap_offset = bitmaps_offset;
2256 bitmaps_offset += stack_bitmap_size;
2258 bitmap = &bitmaps [map->stack_ref_bitmap_offset];
2259 for (i = 0; i < nslots; ++i) {
2260 for (j = 0; j < ncallsites; ++j) {
2261 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, j, i))
2262 set_bit (bitmap, stack_bitmap_width, j, i - start);
2266 if (has_pin_slots) {
2267 map->stack_pin_bitmap_offset = bitmaps_offset;
2268 bitmaps_offset += stack_bitmap_size;
2270 bitmap = &bitmaps [map->stack_pin_bitmap_offset];
2271 for (i = 0; i < nslots; ++i) {
2272 for (j = 0; j < ncallsites; ++j) {
2273 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, j, i))
2274 set_bit (bitmap, stack_bitmap_width, j, i - start);
2279 map->reg_ref_bitmap_offset = bitmaps_offset;
2280 bitmaps_offset += reg_ref_bitmap_size;
2282 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2284 for (i = 0; i < nregs; ++i) {
2285 if (reg_ref_mask & (1 << i)) {
2286 for (j = 0; j < ncallsites; ++j) {
2287 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, j, i))
2288 set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
2295 map->reg_pin_bitmap_offset = bitmaps_offset;
2296 bitmaps_offset += reg_pin_bitmap_size;
2298 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2300 for (i = 0; i < nregs; ++i) {
2301 if (reg_pin_mask & (1 << i)) {
2302 for (j = 0; j < ncallsites; ++j) {
2303 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, j, i))
2304 set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
2312 map->ncallsites = ncallsites;
2313 if (cfg->code_len < 256)
2314 map->callsite_entry_size = 1;
2315 else if (cfg->code_len < 65536)
2316 map->callsite_entry_size = 2;
2318 map->callsite_entry_size = 4;
2320 /* Encode the GC Map */
2328 encode_gc_map (map, buf, &endbuf);
2329 g_assert (endbuf - buf < 256);
2331 encoded_size = endbuf - buf;
2332 alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
2334 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2335 //emap->ref_slots = map->ref_slots;
2337 /* Encoded fixed fields */
2338 p = &emap->encoded [0];
2339 //emap->encoded_size = encoded_size;
2340 memcpy (p, buf, encoded_size);
2343 /* Callsite table */
2344 p = (guint8*)ALIGN_TO ((mgreg_t)p, map->callsite_entry_size);
2345 if (map->callsite_entry_size == 1) {
2346 guint8 *offsets = p;
2347 for (i = 0; i < ncallsites; ++i)
2348 offsets [i] = callsites [i]->pc_offset;
2349 stats.gc_callsites8_size += ncallsites * sizeof (guint8);
2350 } else if (map->callsite_entry_size == 2) {
2351 guint16 *offsets = (guint16*)p;
2352 for (i = 0; i < ncallsites; ++i)
2353 offsets [i] = callsites [i]->pc_offset;
2354 stats.gc_callsites16_size += ncallsites * sizeof (guint16);
2356 guint32 *offsets = (guint32*)p;
2357 for (i = 0; i < ncallsites; ++i)
2358 offsets [i] = callsites [i]->pc_offset;
2359 stats.gc_callsites32_size += ncallsites * sizeof (guint32);
2361 p += ncallsites * map->callsite_entry_size;
2364 memcpy (p, bitmaps, bitmaps_size);
2367 g_assert ((guint8*)p - (guint8*)emap <= alloc_size);
2369 stats.gc_maps_size += alloc_size;
2370 stats.gc_callsites_size += ncallsites * map->callsite_entry_size;
2371 stats.gc_bitmaps_size += bitmaps_size;
2372 stats.gc_map_struct_size += sizeof (GCEncodedMap) + encoded_size;
2374 cfg->jit_info->gc_info = emap;
2376 cfg->gc_map = (guint8*)emap;
2377 cfg->gc_map_size = alloc_size;
2380 stats.all_slots += nslots;
2381 stats.ref_slots += ntypes [SLOT_REF];
2382 stats.noref_slots += ntypes [SLOT_NOREF];
2383 stats.pin_slots += ntypes [SLOT_PIN];
2387 mini_gc_create_gc_map (MonoCompile *cfg)
2389 if (!cfg->compute_gc_maps)
2393 * During marking, all frames except the top frame are at a call site, and we mark the
2394 * top frame conservatively. This means that we only need to compute and record
2395 * GC maps for call sites.
2398 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2399 /* Without liveness info, the live ranges are not precise enough */
2402 mono_analyze_liveness_gc (cfg);
2404 compute_frame_size (cfg);
2408 process_spill_slots (cfg);
2409 process_other_slots (cfg);
2410 process_param_area_slots (cfg);
2411 process_variables (cfg);
2412 process_finally_clauses (cfg);
2418 parse_debug_options (void)
2423 env = getenv ("MONO_GCMAP_DEBUG");
2427 opts = g_strsplit (env, ",", -1);
2428 for (ptr = opts; ptr && *ptr; ptr ++) {
2429 /* No options yet */
2430 fprintf (stderr, "Invalid format for the MONO_GCMAP_DEBUG env variable: '%s'\n", env);
2441 memset (&cb, 0, sizeof (cb));
2442 cb.thread_attach_func = thread_attach_func;
2443 cb.thread_detach_func = thread_detach_func;
2444 cb.thread_suspend_func = thread_suspend_func;
2445 /* Comment this out to disable precise stack marking */
2446 cb.thread_mark_func = thread_mark_func;
2447 mono_gc_set_gc_callbacks (&cb);
2449 logfile = mono_gc_get_logfile ();
2451 parse_debug_options ();
2453 mono_counters_register ("GC Maps size",
2454 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
2455 mono_counters_register ("GC Call Sites size",
2456 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
2457 mono_counters_register ("GC Bitmaps size",
2458 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
2459 mono_counters_register ("GC Map struct size",
2460 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
2461 mono_counters_register ("GC Call Sites encoded using 8 bits",
2462 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
2463 mono_counters_register ("GC Call Sites encoded using 16 bits",
2464 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
2465 mono_counters_register ("GC Call Sites encoded using 32 bits",
2466 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
2468 mono_counters_register ("GC Map slots (all)",
2469 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
2470 mono_counters_register ("GC Map slots (ref)",
2471 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.ref_slots);
2472 mono_counters_register ("GC Map slots (noref)",
2473 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.noref_slots);
2474 mono_counters_register ("GC Map slots (pin)",
2475 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
2477 mono_counters_register ("GC TLS Data size",
2478 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2480 mono_counters_register ("Stack space scanned (all)",
2481 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
2482 mono_counters_register ("Stack space scanned (native)",
2483 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_native);
2484 mono_counters_register ("Stack space scanned (other)",
2485 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_other);
2486 mono_counters_register ("Stack space scanned (using GC Maps)",
2487 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
2488 mono_counters_register ("Stack space scanned (precise)",
2489 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_precisely);
2490 mono_counters_register ("Stack space scanned (pin)",
2491 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_conservatively);
2492 mono_counters_register ("Stack space scanned (pin registers)",
2493 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_registers);
2504 mini_gc_init_gc_map (MonoCompile *cfg)
2509 mini_gc_create_gc_map (MonoCompile *cfg)
2514 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2519 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2528 * Set GC specific options in CFG.
2531 mini_gc_init_cfg (MonoCompile *cfg)
2533 if (mono_gc_is_moving ()) {
2534 cfg->disable_ref_noref_stack_slot_share = TRUE;
2535 cfg->gen_write_barriers = TRUE;
2538 mini_gc_init_gc_map (cfg);
2542 * Problems with the current code:
2543 * - the stack walk is slow
2544 * - vtypes/refs used in EH regions are treated conservatively
2545 * - if the code is finished, less pinning will be done, causing problems because
2546 * we promote all surviving objects to old-gen.
2547 * - the unwind code can't handle a method stopped inside a finally region, it thinks the caller is
2548 * another method, but in reality it is either the exception handling code or the CALL_HANDLER opcode.
2549 * This manifests in "Unable to find ip offset x in callsite list" assertions.
2550 * - the unwind code also can't handle frames which are in the epilog, since the unwind info is not
2555 * Ideas for creating smaller GC maps:
2556 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2558 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2560 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2561 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2562 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2563 * from an index entry.