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 static gboolean enable_gc_maps_for_aot;
268 mini_gc_enable_gc_maps_for_aot (void)
270 enable_gc_maps_for_aot = TRUE;
273 // FIXME: Move these to a shared place
276 encode_uleb128 (guint32 value, guint8 *buf, guint8 **endbuf)
281 guint8 b = value & 0x7f;
283 if (value != 0) /* more bytes to come */
291 static G_GNUC_UNUSED void
292 encode_sleb128 (gint32 value, guint8 *buf, guint8 **endbuf)
295 gboolean negative = (value < 0);
303 /* the following is unnecessary if the
304 * implementation of >>= uses an arithmetic rather
305 * than logical shift for a signed left operand
309 value |= - (1 <<(size - 7));
310 /* sign bit of byte is second high order bit (0x40) */
311 if ((value == 0 && !(byte & 0x40)) ||
312 (value == -1 && (byte & 0x40)))
322 static inline guint32
323 decode_uleb128 (guint8 *buf, guint8 **endbuf)
333 res = res | (((int)(b & 0x7f)) << shift);
345 decode_sleb128 (guint8 *buf, guint8 **endbuf)
355 res = res | (((int)(b & 0x7f)) << shift);
358 if (shift < 32 && (b & 0x40))
359 res |= - (1 << shift);
370 encode_frame_reg (int frame_reg)
373 if (frame_reg == AMD64_RSP)
375 else if (frame_reg == AMD64_RBP)
377 #elif defined(TARGET_X86)
378 if (frame_reg == X86_EBP)
380 else if (frame_reg == X86_ESP)
382 #elif defined(TARGET_ARM)
383 if (frame_reg == ARMREG_SP)
385 else if (frame_reg == ARMREG_FP)
387 #elif defined(TARGET_S390X)
388 if (frame_reg == S390_SP)
390 else if (frame_reg == S390_FP)
395 g_assert_not_reached ();
400 decode_frame_reg (int encoded)
405 else if (encoded == 1)
407 #elif defined(TARGET_X86)
410 else if (encoded == 1)
412 #elif defined(TARGET_ARM)
415 else if (encoded == 1)
417 #elif defined(TARGET_S390X)
420 else if (encoded == 1)
425 g_assert_not_reached ();
431 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15, AMD64_RDI, AMD64_RSI };
433 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15 };
435 #elif defined(TARGET_X86)
436 static int callee_saved_regs [] = { X86_EBX, X86_ESI, X86_EDI };
437 #elif defined(TARGET_ARM)
438 static int callee_saved_regs [] = { ARMREG_V1, ARMREG_V2, ARMREG_V3, ARMREG_V4, ARMREG_V5, ARMREG_V7, ARMREG_FP };
439 #elif defined(TARGET_S390X)
440 static int callee_saved_regs [] = { s390_r6, s390_r7, s390_r8, s390_r9, s390_r10, s390_r11, s390_r12, s390_r13, s390_r14 };
444 encode_regmask (guint32 regmask)
450 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
451 if (regmask & (1 << callee_saved_regs [i])) {
453 regmask -= (1 << callee_saved_regs [i]);
456 g_assert (regmask == 0);
461 decode_regmask (guint32 regmask)
467 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i)
468 if (regmask & (1 << i))
469 res |= (1 << callee_saved_regs [i]);
476 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
479 encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
483 encode_sleb128 (map->start_offset / SIZEOF_SLOT, buf, &buf);
484 encode_sleb128 (map->end_offset / SIZEOF_SLOT, buf, &buf);
485 encode_sleb128 (map->map_offset / SIZEOF_SLOT, buf, &buf);
486 encode_uleb128 (map->nslots, buf, &buf);
487 g_assert (map->callsite_entry_size <= 4);
488 freg = encode_frame_reg (map->frame_reg);
490 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);
491 encode_uleb128 (flags, buf, &buf);
492 encode_uleb128 (encode_regmask (map->used_int_regs), buf, &buf);
493 if (map->has_ref_regs)
494 encode_uleb128 (encode_regmask (map->reg_ref_mask), buf, &buf);
495 if (map->has_pin_regs)
496 encode_uleb128 (encode_regmask (map->reg_pin_mask), buf, &buf);
497 encode_uleb128 (map->ncallsites, buf, &buf);
505 * Decode the encoded GC map representation in BUF and store the result into MAP.
508 decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
511 int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
514 map->start_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
515 map->end_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
516 map->map_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
517 map->nslots = decode_uleb128 (buf, &buf);
518 flags = decode_uleb128 (buf, &buf);
519 map->has_ref_slots = (flags & 1) ? 1 : 0;
520 map->has_pin_slots = (flags & 2) ? 1 : 0;
521 map->has_ref_regs = (flags & 4) ? 1 : 0;
522 map->has_pin_regs = (flags & 8) ? 1 : 0;
523 map->callsite_entry_size = ((flags >> 4) & 0x3) + 1;
525 map->frame_reg = decode_frame_reg (freg);
526 map->used_int_regs = decode_regmask (decode_uleb128 (buf, &buf));
527 if (map->has_ref_regs) {
528 map->reg_ref_mask = decode_regmask (decode_uleb128 (buf, &buf));
530 for (i = 0; i < NREGS; ++i)
531 if (map->reg_ref_mask & (1 << i))
535 if (map->has_pin_regs) {
536 map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
538 for (i = 0; i < NREGS; ++i)
539 if (map->reg_pin_mask & (1 << i))
543 map->ncallsites = decode_uleb128 (buf, &buf);
545 stack_bitmap_size = (ALIGN_TO (map->nslots, 8) / 8) * map->ncallsites;
546 reg_ref_bitmap_size = (ALIGN_TO (map->nref_regs, 8) / 8) * map->ncallsites;
547 reg_pin_bitmap_size = (ALIGN_TO (map->npin_regs, 8) / 8) * map->ncallsites;
549 map->stack_ref_bitmap_offset = offset;
550 if (map->has_ref_slots)
551 offset += stack_bitmap_size;
552 map->stack_pin_bitmap_offset = offset;
553 if (map->has_pin_slots)
554 offset += stack_bitmap_size;
555 map->reg_ref_bitmap_offset = offset;
556 if (map->has_ref_regs)
557 offset += reg_ref_bitmap_size;
558 map->reg_pin_bitmap_offset = offset;
559 if (map->has_pin_regs)
560 offset += reg_pin_bitmap_size;
566 thread_attach_func (void)
570 tls = g_new0 (TlsData, 1);
571 tls->tid = GetCurrentThreadId ();
572 tls->info = mono_thread_info_current ();
573 stats.tlsdata_size += sizeof (TlsData);
579 thread_detach_func (gpointer user_data)
581 TlsData *tls = user_data;
587 thread_suspend_func (gpointer user_data, void *sigctx, MonoContext *ctx)
589 TlsData *tls = user_data;
592 /* Happens during startup */
596 if (tls->tid != GetCurrentThreadId ()) {
597 /* Happens on osx because threads are not suspended using signals */
600 g_assert (tls->info);
604 res = mono_thread_state_init_from_handle (&tls->unwind_state, (MonoNativeThreadId)tls->tid, tls->info->native_handle);
607 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF] = mono_get_lmf ();
609 #ifdef MONO_ARCH_HAVE_SIGCTX_TO_MONOCTX
610 mono_arch_sigctx_to_monoctx (sigctx, &tls->unwind_state.ctx);
611 tls->unwind_state.valid = TRUE;
613 tls->unwind_state.valid = FALSE;
616 memcpy (&tls->unwind_state.ctx, ctx, sizeof (MonoContext));
617 tls->unwind_state.valid = TRUE;
619 tls->unwind_state.valid = FALSE;
621 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_JIT_TLS] = mono_native_tls_get_value (mono_jit_tls_id);
622 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN] = mono_domain_get ();
625 if (!tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN]) {
626 /* Happens during startup */
627 tls->unwind_state.valid = FALSE;
632 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
635 set_bit (guint8 *bitmap, int width, int y, int x)
637 bitmap [(width * y) + (x / 8)] |= (1 << (x % 8));
641 clear_bit (guint8 *bitmap, int width, int y, int x)
643 bitmap [(width * y) + (x / 8)] &= ~(1 << (x % 8));
647 get_bit (guint8 *bitmap, int width, int y, int x)
649 return bitmap [(width * y) + (x / 8)] & (1 << (x % 8));
653 slot_type_to_string (GCSlotType type)
663 g_assert_not_reached ();
668 static inline mgreg_t
669 get_frame_pointer (MonoContext *ctx, int frame_reg)
671 #if defined(TARGET_AMD64)
672 if (frame_reg == AMD64_RSP)
674 else if (frame_reg == AMD64_RBP)
676 #elif defined(TARGET_X86)
677 if (frame_reg == X86_ESP)
679 else if (frame_reg == X86_EBP)
681 #elif defined(TARGET_ARM)
682 if (frame_reg == ARMREG_SP)
683 return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
684 else if (frame_reg == ARMREG_FP)
685 return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
686 #elif defined(TARGET_S390X)
687 if (frame_reg == S390_SP)
688 return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
689 else if (frame_reg == S390_FP)
690 return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
692 g_assert_not_reached ();
697 * conservatively_pass:
699 * Mark a thread stack conservatively and collect information needed by the precise pass.
702 conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
705 MonoContext ctx, new_ctx;
708 gboolean last = TRUE;
712 guint8* fp, *p, *real_frame_start, *frame_start, *frame_end;
713 int i, pc_offset, cindex, bitmap_width;
714 int scanned = 0, scanned_precisely, scanned_conservatively, scanned_registers;
716 StackFrameInfo frame;
717 mgreg_t *reg_locations [MONO_MAX_IREGS];
718 mgreg_t *new_reg_locations [MONO_MAX_IREGS];
721 guint32 precise_regmask;
725 tls->ref_to_track = NULL;
728 /* tls == NULL can happen during startup */
729 if (mono_thread_internal_current () == NULL || !tls) {
730 mono_gc_conservatively_scan_area (stack_start, stack_end);
731 stats.scanned_stacks += stack_end - stack_start;
735 lmf = tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF];
738 /* Number of bytes scanned based on GC map data */
740 /* Number of bytes scanned precisely based on GC map data */
741 scanned_precisely = 0;
742 /* Number of bytes scanned conservatively based on GC map data */
743 scanned_conservatively = 0;
744 /* Number of bytes scanned conservatively in register save areas */
745 scanned_registers = 0;
747 /* This is one past the last address which we have scanned */
748 stack_limit = stack_start;
750 if (!tls->unwind_state.valid)
751 memset (&new_ctx, 0, sizeof (ctx));
753 memcpy (&new_ctx, &tls->unwind_state.ctx, sizeof (MonoContext));
755 memset (reg_locations, 0, sizeof (reg_locations));
756 memset (new_reg_locations, 0, sizeof (new_reg_locations));
759 if (!tls->unwind_state.valid)
762 memcpy (&ctx, &new_ctx, sizeof (ctx));
764 for (i = 0; i < MONO_MAX_IREGS; ++i) {
765 if (new_reg_locations [i]) {
767 * If the current frame saves the register, it means it might modify its
768 * value, thus the old location might not contain the same value, so
769 * we have to mark it conservatively.
771 if (reg_locations [i]) {
772 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
773 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
774 scanned_registers += SIZEOF_SLOT;
777 reg_locations [i] = new_reg_locations [i];
779 DEBUG (fprintf (logfile, "\treg %s is now at location %p.\n", mono_arch_regname (i), reg_locations [i]));
783 g_assert ((mgreg_t)stack_limit % SIZEOF_SLOT == 0);
785 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);
791 // FIXME: For skipped frames, scan the param area of the parent frame conservatively ?
793 if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
795 * These frames are problematic for several reasons:
796 * - they are unwound through an LMF, and we have no precise register tracking for those.
797 * - the LMF might not contain a precise ip, so we can't compute the call site.
798 * - the LMF only unwinds to the wrapper frame, so we get these methods twice.
800 DEBUG (fprintf (logfile, "Mark(0): <Managed-to-native transition>\n"));
801 for (i = 0; i < MONO_MAX_IREGS; ++i) {
802 if (reg_locations [i]) {
803 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
804 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
805 scanned_registers += SIZEOF_SLOT;
807 reg_locations [i] = NULL;
808 new_reg_locations [i] = NULL;
814 /* The last frame can be in any state so mark conservatively */
817 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));
819 DEBUG (fprintf (logfile, "\t <Last frame>\n"));
822 * new_reg_locations is not precise when a method is interrupted during its epilog, so clear it.
824 for (i = 0; i < MONO_MAX_IREGS; ++i) {
825 if (reg_locations [i]) {
826 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
827 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
828 scanned_registers += SIZEOF_SLOT;
830 if (new_reg_locations [i]) {
831 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), new_reg_locations [i]));
832 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
833 scanned_registers += SIZEOF_SLOT;
835 reg_locations [i] = NULL;
836 new_reg_locations [i] = NULL;
841 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
843 /* These frames are very problematic */
844 if (ji->method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
845 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));
846 DEBUG (fprintf (logfile, "\tSkip.\n"));
850 /* All the other frames are at a call site */
852 if (tls->nframes == MAX_FRAMES) {
854 * Can't save information since the array is full. So scan the rest of the
855 * stack conservatively.
857 DEBUG (fprintf (logfile, "Mark (0): Frame stack full.\n"));
861 /* Scan the frame of this method */
864 * A frame contains the following:
869 * - localloc-ed memory
871 g_assert (pc_offset >= 0);
876 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));
877 DEBUG (fprintf (logfile, "\tNo GC Map.\n"));
881 /* The embedded callsite table requires this */
882 g_assert (((mgreg_t)emap % 4) == 0);
885 * Debugging aid to control the number of frames scanned precisely
887 if (!precise_frame_limit_inited) {
888 if (g_getenv ("MONO_PRECISE_COUNT"))
889 precise_frame_limit = atoi (g_getenv ("MONO_PRECISE_COUNT"));
890 precise_frame_limit_inited = TRUE;
893 if (precise_frame_limit != -1) {
894 if (precise_frame_count [FALSE] == precise_frame_limit)
895 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (ji->method, TRUE));
896 if (precise_frame_count [FALSE] > precise_frame_limit)
899 precise_frame_count [FALSE] ++;
901 /* Decode the encoded GC map */
903 memset (map, 0, sizeof (GCMap));
904 decode_gc_map (&emap->encoded [0], map, &p);
905 p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
906 map->callsites.offsets8 = p;
907 p += map->callsite_entry_size * map->ncallsites;
910 fp = (guint8*)get_frame_pointer (&ctx, map->frame_reg);
912 real_frame_start = fp + map->start_offset;
913 frame_start = fp + map->start_offset + map->map_offset;
914 frame_end = fp + map->end_offset;
916 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));
918 /* Find the callsite index */
919 if (map->callsite_entry_size == 1) {
920 for (i = 0; i < map->ncallsites; ++i)
921 /* ip points inside the call instruction */
922 if (map->callsites.offsets8 [i] == pc_offset + 1)
924 } else if (map->callsite_entry_size == 2) {
925 // FIXME: Use a binary search
926 for (i = 0; i < map->ncallsites; ++i)
927 /* ip points inside the call instruction */
928 if (map->callsites.offsets16 [i] == pc_offset + 1)
931 // FIXME: Use a binary search
932 for (i = 0; i < map->ncallsites; ++i)
933 /* ip points inside the call instruction */
934 if (map->callsites.offsets32 [i] == pc_offset + 1)
937 if (i == map->ncallsites) {
938 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (ji->method, TRUE));
939 g_assert_not_reached ();
944 * This is not neccessary true on x86 because frames have a different size at each
947 //g_assert (real_frame_start >= stack_limit);
949 if (real_frame_start > stack_limit) {
950 /* This scans the previously skipped frames as well */
951 DEBUG (fprintf (logfile, "\tscan area %p-%p (%d).\n", stack_limit, real_frame_start, (int)(real_frame_start - stack_limit)));
952 mono_gc_conservatively_scan_area (stack_limit, real_frame_start);
953 stats.scanned_other += real_frame_start - stack_limit;
956 /* Mark stack slots */
957 if (map->has_pin_slots) {
958 int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
959 guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
964 for (i = 0; i < map->nslots; ++i) {
965 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
967 DEBUG (fprintf (logfile, "\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
968 mono_gc_conservatively_scan_area (p, p + SIZEOF_SLOT);
969 scanned_conservatively += SIZEOF_SLOT;
971 scanned_precisely += SIZEOF_SLOT;
976 scanned_precisely += (map->nslots * SIZEOF_SLOT);
979 /* The area outside of start-end is NOREF */
980 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * SIZEOF_SLOT);
983 precise_regmask = map->used_int_regs | (1 << map->frame_reg);
984 if (map->has_pin_regs) {
985 int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
986 guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
988 for (i = 0; i < NREGS; ++i) {
989 if (!(map->used_int_regs & (1 << i)))
992 if (!(map->reg_pin_mask & (1 << i)))
995 if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
996 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
997 precise_regmask &= ~(1 << i);
1003 scanned += map->end_offset - map->start_offset;
1005 g_assert (scanned == scanned_precisely + scanned_conservatively);
1007 stack_limit = frame_end;
1009 /* Save information for the precise pass */
1010 fi = &tls->frames [tls->nframes];
1011 fi->nslots = map->nslots;
1012 bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
1013 if (map->has_ref_slots)
1014 fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
1017 fi->frame_start_offset = frame_start - stack_start;
1018 fi->nreg_locations = 0;
1019 DEBUG_PRECISE (fi->ji = ji);
1020 DEBUG_PRECISE (fi->fp = fp);
1022 if (map->has_ref_regs) {
1023 int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
1024 guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
1026 for (i = 0; i < NREGS; ++i) {
1027 if (!(map->reg_ref_mask & (1 << i)))
1030 if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
1031 DEBUG_PRECISE (fi->regs [fi->nreg_locations] = i);
1032 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
1033 fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
1034 fi->nreg_locations ++;
1041 * Clear locations of precisely tracked registers.
1043 if (precise_regmask) {
1044 for (i = 0; i < NREGS; ++i) {
1045 if (precise_regmask & (1 << i)) {
1047 * The method uses this register, and we have precise info for it.
1048 * This means the location will be scanned precisely.
1049 * Tell the code at the beginning of the loop that this location is
1052 if (reg_locations [i])
1053 DEBUG (fprintf (logfile, "\treg %s at location %p (==%p) is precise.\n", mono_arch_regname (i), reg_locations [i], (gpointer)*reg_locations [i]));
1054 reg_locations [i] = NULL;
1062 /* Scan the remaining register save locations */
1063 for (i = 0; i < MONO_MAX_IREGS; ++i) {
1064 if (reg_locations [i]) {
1065 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", reg_locations [i]));
1066 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
1067 scanned_registers += SIZEOF_SLOT;
1069 if (new_reg_locations [i]) {
1070 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", new_reg_locations [i]));
1071 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
1072 scanned_registers += SIZEOF_SLOT;
1076 if (stack_limit < stack_end) {
1077 DEBUG (fprintf (logfile, "\tscan remaining stack %p-%p (%d).\n", stack_limit, stack_end, (int)(stack_end - stack_limit)));
1078 mono_gc_conservatively_scan_area (stack_limit, stack_end);
1079 stats.scanned_native += stack_end - stack_limit;
1082 DEBUG (fprintf (logfile, "Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
1084 stats.scanned_stacks += stack_end - stack_start;
1085 stats.scanned += scanned;
1086 stats.scanned_precisely += scanned_precisely;
1087 stats.scanned_conservatively += scanned_conservatively;
1088 stats.scanned_registers += scanned_registers;
1090 //mono_gc_conservatively_scan_area (stack_start, stack_end);
1096 * Mark a thread stack precisely based on information saved during the conservative
1100 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
1104 guint8 *frame_start;
1109 if (!tls->unwind_state.valid)
1112 for (findex = 0; findex < tls->nframes; findex ++) {
1113 /* Load information saved by the !precise pass */
1114 fi = &tls->frames [findex];
1115 frame_start = stack_start + fi->frame_start_offset;
1117 DEBUG (char *fname = mono_method_full_name (fi->ji->method, TRUE); fprintf (logfile, "Mark(1): %s\n", fname); g_free (fname));
1120 * FIXME: Add a function to mark using a bitmap, to avoid doing a
1121 * call for each object.
1124 /* Mark stack slots */
1126 guint8 *ref_bitmap = fi->bitmap;
1129 for (i = 0; i < fi->nslots; ++i) {
1130 MonoObject **ptr = (MonoObject**)(frame_start + (i * SIZEOF_SLOT));
1132 live = ref_bitmap [i / 8] & (1 << (i % 8));
1135 MonoObject *obj = *ptr;
1137 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1138 *ptr = mono_gc_scan_object (obj);
1139 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1141 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1146 * This is disabled because the pointer takes up a lot of space.
1147 * Stack slots might be shared between ref and non-ref variables ?
1149 if (map->ref_slots [i / 8] & (1 << (i % 8))) {
1150 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));
1152 * Fail fast if the live range is incorrect, and
1153 * the JITted code tries to access this object
1162 /* Mark registers */
1165 * Registers are different from stack slots, they have no address where they
1166 * are stored. Instead, some frame below this frame in the stack saves them
1167 * in its prolog to the stack. We can mark this location precisely.
1169 for (i = 0; i < fi->nreg_locations; ++i) {
1171 * reg_locations [i] contains the address of the stack slot where
1172 * a reg was last saved, so mark that slot.
1174 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1175 MonoObject *obj = *ptr;
1178 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
1179 *ptr = mono_gc_scan_object (obj);
1180 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1182 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p\n", mono_arch_regname (fi->regs [i]), ptr, obj));
1188 * Debugging aid to check for missed refs.
1190 if (tls->ref_to_track) {
1193 for (p = (mgreg_t*)stack_start; p < (mgreg_t*)stack_end; ++p)
1194 if (*p == (mgreg_t)tls->ref_to_track)
1195 printf ("REF AT %p.\n", p);
1202 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1203 * call, and TRUE at the second. USER_DATA points to a TlsData
1204 * structure filled up by thread_suspend_func.
1207 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise)
1209 TlsData *tls = user_data;
1211 DEBUG (fprintf (logfile, "****************************************\n"));
1212 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));
1213 DEBUG (fprintf (logfile, "****************************************\n"));
1216 conservative_pass (tls, stack_start, stack_end);
1218 precise_pass (tls, stack_start, stack_end);
1224 mini_gc_init_gc_map (MonoCompile *cfg)
1226 if (COMPILE_LLVM (cfg))
1229 if (!mono_gc_is_moving ())
1232 if (cfg->compile_aot) {
1233 if (!enable_gc_maps_for_aot)
1235 } else if (!mono_gc_precise_stack_mark_enabled ())
1239 /* Debugging support */
1241 static int precise_count;
1244 if (g_getenv ("MONO_GCMAP_COUNT")) {
1245 if (precise_count == atoi (g_getenv ("MONO_GCMAP_COUNT")))
1246 printf ("LAST: %s\n", mono_method_full_name (cfg->method, TRUE));
1247 if (precise_count > atoi (g_getenv ("MONO_GCMAP_COUNT")))
1253 cfg->compute_gc_maps = TRUE;
1255 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1259 * mini_gc_set_slot_type_from_fp:
1261 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1262 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1263 * need to call this function for those slots.
1266 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1268 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1270 if (!cfg->compute_gc_maps)
1273 g_assert (slot_offset % SIZEOF_SLOT == 0);
1275 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1279 * mini_gc_set_slot_type_from_cfa:
1281 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1282 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1283 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1284 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1285 * call this function for those slots.
1288 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1290 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1291 int slot = - (slot_offset / SIZEOF_SLOT);
1293 if (!cfg->compute_gc_maps)
1296 g_assert (slot_offset <= 0);
1297 g_assert (slot_offset % SIZEOF_SLOT == 0);
1299 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1303 fp_offset_to_slot (MonoCompile *cfg, int offset)
1305 MonoCompileGC *gcfg = cfg->gc_info;
1307 return (offset - gcfg->min_offset) / SIZEOF_SLOT;
1311 slot_to_fp_offset (MonoCompile *cfg, int slot)
1313 MonoCompileGC *gcfg = cfg->gc_info;
1315 return (slot * SIZEOF_SLOT) + gcfg->min_offset;
1318 static inline MONO_ALWAYS_INLINE void
1319 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1321 g_assert (slot >= 0 && slot < gcfg->nslots);
1323 if (type == SLOT_PIN) {
1324 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1325 set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1326 } else if (type == SLOT_REF) {
1327 set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1328 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1329 } else if (type == SLOT_NOREF) {
1330 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1331 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1336 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1339 guint8 *ref_bitmap, *pin_bitmap;
1344 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1345 set_slot (gcfg, slot, cindex, type);
1347 ref_bitmap = gcfg->stack_ref_bitmap;
1348 pin_bitmap = gcfg->stack_pin_bitmap;
1349 width = gcfg->stack_bitmap_width;
1352 if (type == SLOT_PIN) {
1353 memset (ref_bitmap + pos, 0, width);
1354 memset (pin_bitmap + pos, 0xff, width);
1355 } else if (type == SLOT_REF) {
1356 memset (ref_bitmap + pos, 0xff, width);
1357 memset (pin_bitmap + pos, 0, width);
1358 } else if (type == SLOT_NOREF) {
1359 memset (ref_bitmap + pos, 0, width);
1360 memset (pin_bitmap + pos, 0, width);
1365 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1369 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1370 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1371 if (callsite_offset >= from && callsite_offset < to)
1372 set_slot (gcfg, slot, cindex, type);
1377 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1379 g_assert (slot >= 0 && slot < gcfg->nregs);
1381 if (type == SLOT_PIN) {
1382 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1383 set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1384 } else if (type == SLOT_REF) {
1385 set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1386 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1387 } else if (type == SLOT_NOREF) {
1388 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1389 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1394 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1398 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1399 set_reg_slot (gcfg, slot, cindex, type);
1403 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1407 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1408 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1409 if (callsite_offset >= from && callsite_offset < to)
1410 set_reg_slot (gcfg, slot, cindex, type);
1415 process_spill_slots (MonoCompile *cfg)
1417 MonoCompileGC *gcfg = cfg->gc_info;
1422 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1423 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1424 for (l = bb->spill_slot_defs; l; l = l->next) {
1425 MonoInst *def = l->data;
1426 int spill_slot = def->inst_c0;
1427 int bank = def->inst_c1;
1428 int offset = cfg->spill_info [bank][spill_slot].offset;
1429 int slot = fp_offset_to_slot (cfg, offset);
1431 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1432 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1436 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1437 for (l = bb->spill_slot_defs; l; l = l->next) {
1438 MonoInst *def = l->data;
1439 int spill_slot = def->inst_c0;
1440 int bank = def->inst_c1;
1441 int offset = cfg->spill_info [bank][spill_slot].offset;
1442 int slot = fp_offset_to_slot (cfg, offset);
1445 if (bank == MONO_REG_INT_MP)
1451 * Extend the live interval for the GC tracked spill slots
1452 * defined in this bblock.
1453 * FIXME: This is not needed.
1455 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1457 if (cfg->verbose_level > 1)
1458 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
1462 /* Set fp spill slots to NOREF */
1463 for (i = 0; i < cfg->spill_info_len [MONO_REG_DOUBLE]; ++i) {
1464 int offset = cfg->spill_info [MONO_REG_DOUBLE][i].offset;
1470 slot = fp_offset_to_slot (cfg, offset);
1472 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1474 if (cfg->verbose_level > 1)
1475 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1478 /* Set int spill slots to NOREF */
1479 for (i = 0; i < cfg->spill_info_len [MONO_REG_INT]; ++i) {
1480 int offset = cfg->spill_info [MONO_REG_INT][i].offset;
1486 slot = fp_offset_to_slot (cfg, offset);
1488 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1489 if (cfg->verbose_level > 1)
1490 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1495 * process_other_slots:
1497 * Process stack slots registered using mini_gc_set_slot_type_... ().
1500 process_other_slots (MonoCompile *cfg)
1502 MonoCompileGC *gcfg = cfg->gc_info;
1505 /* Relative to the CFA */
1506 for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
1507 guint data = GPOINTER_TO_UINT (l->data);
1508 int cfa_slot = data >> 16;
1509 GCSlotType type = data & 0xff;
1513 * Map the cfa relative slot to an fp relative slot.
1514 * slot_addr == cfa - <cfa_slot>*4/8
1515 * fp + cfa_offset == cfa
1516 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1518 slot = (cfg->cfa_offset / SIZEOF_SLOT) - cfa_slot - (gcfg->min_offset / SIZEOF_SLOT);
1520 set_slot_everywhere (gcfg, slot, type);
1522 if (cfg->verbose_level > 1) {
1523 int fp_offset = slot_to_fp_offset (cfg, slot);
1524 if (type == SLOT_NOREF)
1525 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));
1529 /* Relative to the FP */
1530 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1531 gint data = GPOINTER_TO_INT (l->data);
1532 int offset = data >> 16;
1533 GCSlotType type = data & 0xff;
1536 slot = fp_offset_to_slot (cfg, offset);
1538 set_slot_everywhere (gcfg, slot, type);
1540 /* Liveness for these slots is handled by process_spill_slots () */
1542 if (cfg->verbose_level > 1) {
1543 if (type == SLOT_REF)
1544 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
1545 else if (type == SLOT_NOREF)
1546 printf ("\tnoref slot at 0x%x(fp) (slot = %d)\n", offset, slot);
1552 get_vtype_bitmap (MonoType *t, int *numbits)
1554 MonoClass *klass = mono_class_from_mono_type (t);
1556 if (klass->generic_container || mono_class_is_open_constructed_type (t)) {
1557 /* FIXME: Generic sharing */
1560 mono_class_compute_gc_descriptor (klass);
1562 return mono_gc_get_bitmap_for_descr (klass->gc_descr, numbits);
1566 static inline const char*
1567 get_offset_sign (int offset)
1569 return offset < 0 ? "-" : "+";
1573 get_offset_val (int offset)
1575 return offset < 0 ? (- offset) : offset;
1579 process_variables (MonoCompile *cfg)
1581 MonoCompileGC *gcfg = cfg->gc_info;
1582 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1583 int i, locals_min_slot, locals_max_slot, cindex;
1587 int locals_min_offset = gcfg->locals_min_offset;
1588 int locals_max_offset = gcfg->locals_max_offset;
1590 /* Slots for locals are NOREF by default */
1591 locals_min_slot = (locals_min_offset - gcfg->min_offset) / SIZEOF_SLOT;
1592 locals_max_slot = (locals_max_offset - gcfg->min_offset) / SIZEOF_SLOT;
1593 for (i = locals_min_slot; i < locals_max_slot; ++i) {
1594 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1598 * Compute the offset where variables are initialized in the first bblock, if any.
1600 pc_offsets = g_new0 (int, cfg->next_vreg);
1602 bb = cfg->bb_entry->next_bb;
1603 MONO_BB_FOR_EACH_INS (bb, tmp) {
1604 if (tmp->opcode == OP_GC_LIVENESS_DEF) {
1605 int vreg = tmp->inst_c1;
1606 if (pc_offsets [vreg] == 0) {
1607 g_assert (tmp->backend.pc_offset > 0);
1608 pc_offsets [vreg] = tmp->backend.pc_offset;
1614 * Stack slots holding arguments are initialized in the prolog.
1615 * This means we can treat them alive for the whole method.
1617 for (i = 0; i < cfg->num_varinfo; i++) {
1618 MonoInst *ins = cfg->varinfo [i];
1619 MonoType *t = ins->inst_vtype;
1622 gboolean byref, is_this = FALSE;
1623 gboolean is_arg = i < cfg->locals_start;
1625 if (ins == cfg->ret) {
1626 if (!(ins->opcode == OP_REGOFFSET && MONO_TYPE_ISSTRUCT (t)))
1630 vmv = MONO_VARINFO (cfg, i);
1632 /* For some reason, 'this' is byref */
1633 if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
1634 t = &cfg->method->klass->byval_arg;
1640 if (ins->opcode == OP_REGVAR) {
1642 GCSlotType slot_type;
1644 t = mini_type_get_underlying_type (NULL, t);
1647 g_assert (hreg < MONO_MAX_IREGS);
1650 slot_type = SLOT_PIN;
1652 slot_type = mini_type_is_reference (cfg, t) ? SLOT_REF : SLOT_NOREF;
1654 if (slot_type == SLOT_PIN) {
1655 /* These have no live interval, be conservative */
1656 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1659 * Unlike variables allocated to the stack, we generate liveness info
1660 * for noref vars in registers in mono_spill_global_vars (), because
1661 * knowing that a register doesn't contain a ref allows us to mark its save
1662 * locations precisely.
1664 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1665 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1666 set_reg_slot (gcfg, hreg, cindex, slot_type);
1669 if (cfg->verbose_level > 1) {
1670 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
1676 if (ins->opcode != OP_REGOFFSET)
1679 if (ins->inst_offset % SIZEOF_SLOT != 0)
1682 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1684 if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
1685 /* These do not get stored in the prolog */
1686 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1688 if (cfg->verbose_level > 1) {
1689 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));
1694 if (MONO_TYPE_ISSTRUCT (t)) {
1696 gsize *bitmap = NULL;
1697 gboolean pin = FALSE;
1701 if (ins->backend.is_pinvoke)
1702 size = mono_class_native_size (ins->klass, NULL);
1704 size = mono_class_value_size (ins->klass, NULL);
1705 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
1707 if (cfg->verbose_level > 1)
1708 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));
1710 if (!ins->klass->has_references) {
1712 for (j = 0; j < size_in_slots; ++j)
1713 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1718 bitmap = get_vtype_bitmap (t, &numbits);
1723 * Most vtypes are marked volatile because of the LDADDR instructions,
1724 * and they have no liveness information since they are decomposed
1725 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1726 * them during VZERO decomposition.
1729 if (!pc_offsets [vmv->vreg])
1732 if (ins->backend.is_pinvoke)
1737 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1738 if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
1739 for (j = 0; j < numbits; ++j) {
1740 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1741 /* The descriptor is for the boxed object */
1742 set_slot (gcfg, (pos + j - (sizeof (MonoObject) / SIZEOF_SLOT)), cindex, pin ? SLOT_PIN : SLOT_REF);
1748 if (cfg->verbose_level > 1) {
1749 for (j = 0; j < numbits; ++j) {
1750 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD)))
1751 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)));
1755 if (cfg->verbose_level > 1)
1756 printf ("\t\tpinned\n");
1757 for (j = 0; j < size_in_slots; ++j) {
1758 set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
1767 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1768 /* Vret addr etc. */
1773 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1775 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1776 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1777 set_slot (gcfg, pos, cindex, SLOT_PIN);
1779 if (cfg->verbose_level > 1)
1780 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));
1785 * This is currently disabled, but could be enabled to debug crashes.
1788 if (t->type == MONO_TYPE_I) {
1790 * Variables created in mono_handle_global_vregs have type I, but they
1791 * could hold GC refs since the vregs they were created from might not been
1792 * marked as holding a GC ref. So be conservative.
1794 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1799 t = mini_type_get_underlying_type (NULL, t);
1801 if (!mini_type_is_reference (cfg, t)) {
1802 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1803 if (cfg->verbose_level > 1)
1804 printf ("\tnoref%s at %s0x%x(fp) (R%d, slot = %d): %s\n", (is_arg ? " arg" : ""), 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));
1805 if (!t->byref && sizeof (mgreg_t) == 4 && (t->type == MONO_TYPE_I8 || t->type == MONO_TYPE_U8 || t->type == MONO_TYPE_R8)) {
1806 set_slot_everywhere (gcfg, pos + 1, SLOT_NOREF);
1807 if (cfg->verbose_level > 1)
1808 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));
1813 /* 'this' is marked INDIRECT for gshared methods */
1814 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1816 * For volatile variables, treat them alive from the point they are
1817 * initialized in the first bblock until the end of the method.
1820 set_slot_everywhere (gcfg, pos, SLOT_REF);
1821 } else if (pc_offsets [vmv->vreg]) {
1822 set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
1823 set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
1825 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1827 if (cfg->verbose_level > 1)
1828 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));
1833 /* Live for the whole method */
1834 set_slot_everywhere (gcfg, pos, SLOT_REF);
1836 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1837 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1838 set_slot (gcfg, pos, cindex, SLOT_REF);
1841 if (cfg->verbose_level > 1) {
1842 printf ("\tref%s at %s0x%x(fp) (R%d, slot = %d): %s\n", (is_arg ? " arg" : ""), 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));
1846 g_free (pc_offsets);
1850 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1853 * Convert a sp relative offset to a slot index. This is
1854 * platform specific.
1857 /* fp = sp + offset */
1858 g_assert (cfg->frame_reg == AMD64_RBP);
1859 return (- cfg->arch.sp_fp_offset + sp_offset);
1860 #elif defined(TARGET_X86)
1861 /* The offset is computed from the sp at the start of the call sequence */
1862 g_assert (cfg->frame_reg == X86_EBP);
1863 return (- cfg->arch.sp_fp_offset - sp_offset);
1871 process_param_area_slots (MonoCompile *cfg)
1873 MonoCompileGC *gcfg = cfg->gc_info;
1878 * These slots are used for passing parameters during calls. They are sp relative, not
1879 * fp relative, so they are harder to handle.
1881 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1882 /* The distance between fp and sp is not constant */
1885 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1887 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1888 GCCallSite *callsite = gcfg->callsites [cindex];
1891 for (l = callsite->param_slots; l; l = l->next) {
1892 MonoInst *def = l->data;
1893 MonoType *t = def->inst_vtype;
1894 int sp_offset = def->inst_offset;
1895 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1896 int slot = fp_offset_to_slot (cfg, fp_offset);
1900 if (MONO_TYPE_ISSTRUCT (t)) {
1901 size = mini_type_stack_size_full (cfg->generic_sharing_context, t, &align, FALSE);
1903 size = sizeof (mgreg_t);
1906 for (i = 0; i < size / sizeof (mgreg_t); ++i) {
1907 g_assert (slot + i >= 0 && slot + i < gcfg->nslots);
1908 is_param [slot + i] = TRUE;
1913 /* All param area slots are noref by default */
1914 for (i = 0; i < gcfg->nslots; ++i) {
1916 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1920 * We treat param area slots as being part of the callee's frame, to be able to handle tail calls which overwrite
1921 * the argument area of the caller.
1926 process_finally_clauses (MonoCompile *cfg)
1928 MonoCompileGC *gcfg = cfg->gc_info;
1929 GCCallSite **callsites;
1931 gboolean has_finally;
1932 int i, j, nslots, nregs;
1934 ncallsites = gcfg->ncallsites;
1935 nslots = gcfg->nslots;
1936 nregs = gcfg->nregs;
1937 callsites = gcfg->callsites;
1940 * The calls to the finally clauses don't show up in the cfg. See
1941 * test_0_liveness_8 ().
1942 * Variables accessed inside the finally clause are already marked VOLATILE by
1943 * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
1944 * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
1946 has_finally = FALSE;
1947 for (i = 0; i < cfg->header->num_clauses; ++i) {
1948 MonoExceptionClause *clause = &cfg->header->clauses [i];
1950 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1955 if (cfg->verbose_level > 1)
1956 printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
1957 for (j = 0; j < ncallsites; ++j) {
1958 MonoBasicBlock *bb = callsites [j]->bb;
1959 MonoExceptionClause *clause;
1960 gboolean is_in_finally = FALSE;
1962 for (i = 0; i < cfg->header->num_clauses; ++i) {
1963 clause = &cfg->header->clauses [i];
1965 if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
1966 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1967 is_in_finally = TRUE;
1973 if (is_in_finally) {
1974 for (i = 0; i < nslots; ++i)
1975 set_slot (gcfg, i, j, SLOT_PIN);
1976 for (i = 0; i < nregs; ++i)
1977 set_reg_slot (gcfg, i, j, SLOT_PIN);
1984 compute_frame_size (MonoCompile *cfg)
1986 int i, locals_min_offset, locals_max_offset, cfa_min_offset, cfa_max_offset;
1987 int min_offset, max_offset;
1988 MonoCompileGC *gcfg = cfg->gc_info;
1989 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1992 /* Compute min/max offsets from the fp */
1995 #if defined(TARGET_AMD64) || defined(TARGET_X86) || defined(TARGET_ARM) || defined(TARGET_S390X)
1996 locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, SIZEOF_SLOT);
1997 locals_max_offset = cfg->locals_max_stack_offset;
1999 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
2003 locals_min_offset = ALIGN_TO (locals_min_offset, SIZEOF_SLOT);
2004 locals_max_offset = ALIGN_TO (locals_max_offset, SIZEOF_SLOT);
2006 min_offset = locals_min_offset;
2007 max_offset = locals_max_offset;
2010 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
2011 MonoInst *ins = cfg->args [i];
2013 if (ins->opcode == OP_REGOFFSET) {
2014 int size, size_in_slots;
2015 size = mini_type_stack_size_full (cfg->generic_sharing_context, ins->inst_vtype, NULL, ins->backend.is_pinvoke);
2016 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
2018 min_offset = MIN (min_offset, ins->inst_offset);
2019 max_offset = MAX ((int)max_offset, (int)(ins->inst_offset + (size_in_slots * SIZEOF_SLOT)));
2024 g_assert (cfg->frame_reg == cfg->cfa_reg);
2025 g_assert (cfg->cfa_offset > 0);
2027 cfa_max_offset = cfg->cfa_offset;
2029 min_offset = MIN (min_offset, cfa_min_offset);
2030 max_offset = MAX (max_offset, cfa_max_offset);
2032 /* Fp relative slots */
2033 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
2034 gint data = GPOINTER_TO_INT (l->data);
2035 int offset = data >> 16;
2037 min_offset = MIN (min_offset, offset);
2041 if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
2042 int stack_offset = ALIGN_TO (cfg->stack_offset, SIZEOF_SLOT);
2043 min_offset = MIN (min_offset, (-stack_offset));
2046 /* Param area slots */
2048 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
2049 #elif defined(TARGET_X86)
2050 min_offset = MIN (min_offset, - (cfg->arch.sp_fp_offset + cfg->arch.param_area_size));
2051 #elif defined(TARGET_ARM)
2053 #elif defined(TARGET_s390X)
2059 gcfg->min_offset = min_offset;
2060 gcfg->max_offset = max_offset;
2061 gcfg->locals_min_offset = locals_min_offset;
2062 gcfg->locals_max_offset = locals_max_offset;
2066 init_gcfg (MonoCompile *cfg)
2068 int i, nregs, nslots;
2069 MonoCompileGC *gcfg = cfg->gc_info;
2070 GCCallSite **callsites;
2079 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2080 ncallsites += g_slist_length (bb->gc_callsites);
2082 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
2084 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2085 for (l = bb->gc_callsites; l; l = l->next)
2086 callsites [i++] = l->data;
2089 /* The callsites should already be ordered by pc offset */
2090 for (i = 1; i < ncallsites; ++i)
2091 g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
2094 * The stack frame looks like this:
2096 * <fp + max_offset> == cfa -> <end of previous frame>
2097 * <other stack slots>
2099 * <other stack slots>
2100 * fp + min_offset ->
2105 if (cfg->verbose_level > 1)
2106 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
2108 nslots = (gcfg->max_offset - gcfg->min_offset) / SIZEOF_SLOT;
2111 gcfg->nslots = nslots;
2112 gcfg->nregs = nregs;
2113 gcfg->callsites = callsites;
2114 gcfg->ncallsites = ncallsites;
2115 gcfg->stack_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
2116 gcfg->reg_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
2117 gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
2118 gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
2119 gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
2120 gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
2122 /* All slots start out as PIN */
2123 memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * nregs);
2124 for (i = 0; i < nregs; ++i) {
2126 * By default, registers are NOREF.
2127 * It is possible for a callee to save them before being defined in this method,
2128 * but the saved value is dead too, so it doesn't need to be marked.
2130 if ((cfg->used_int_regs & (1 << i)))
2131 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
2135 static inline gboolean
2136 has_bit_set (guint8 *bitmap, int width, int slot)
2139 int pos = width * slot;
2141 for (i = 0; i < width; ++i) {
2142 if (bitmap [pos + i])
2149 create_map (MonoCompile *cfg)
2152 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
2154 int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
2155 int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
2157 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
2158 MonoCompileGC *gcfg = cfg->gc_info;
2159 GCCallSite **callsites;
2161 guint8 *bitmap, *bitmaps;
2162 guint32 reg_ref_mask, reg_pin_mask;
2164 ncallsites = gcfg->ncallsites;
2165 nslots = gcfg->nslots;
2166 nregs = gcfg->nregs;
2167 callsites = gcfg->callsites;
2170 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
2171 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
2173 has_ref_slots = FALSE;
2174 has_pin_slots = FALSE;
2177 memset (ntypes, 0, sizeof (ntypes));
2178 for (i = 0; i < nslots; ++i) {
2179 gboolean has_ref = FALSE;
2180 gboolean has_pin = FALSE;
2182 if (has_bit_set (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i))
2184 if (has_bit_set (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i))
2188 has_ref_slots = TRUE;
2190 has_pin_slots = TRUE;
2193 ntypes [SLOT_REF] ++;
2195 ntypes [SLOT_PIN] ++;
2197 ntypes [SLOT_NOREF] ++;
2199 if (has_ref || has_pin) {
2206 start = end = nslots;
2208 g_assert (start != -1);
2209 g_assert (start < end);
2212 has_ref_regs = FALSE;
2213 has_pin_regs = FALSE;
2218 for (i = 0; i < nregs; ++i) {
2219 gboolean has_ref = FALSE;
2220 gboolean has_pin = FALSE;
2222 if (!(cfg->used_int_regs & (1 << i)))
2225 if (has_bit_set (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i))
2227 if (has_bit_set (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i))
2231 reg_ref_mask |= (1 << i);
2232 has_ref_regs = TRUE;
2236 reg_pin_mask |= (1 << i);
2237 has_pin_regs = TRUE;
2242 if (cfg->verbose_level > 1)
2243 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);
2245 /* Create the GC Map */
2247 /* The work bitmaps have one row for each slot, since this is how we access them during construction */
2248 stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
2249 stack_bitmap_size = stack_bitmap_width * ncallsites;
2250 reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
2251 reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
2252 reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
2253 reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
2254 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);
2256 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2258 map->frame_reg = cfg->frame_reg;
2259 map->start_offset = gcfg->min_offset;
2260 map->end_offset = gcfg->min_offset + (nslots * SIZEOF_SLOT);
2261 map->map_offset = start * SIZEOF_SLOT;
2262 map->nslots = end - start;
2263 map->has_ref_slots = has_ref_slots;
2264 map->has_pin_slots = has_pin_slots;
2265 map->has_ref_regs = has_ref_regs;
2266 map->has_pin_regs = has_pin_regs;
2267 g_assert (nregs < 32);
2268 map->used_int_regs = cfg->used_int_regs;
2269 map->reg_ref_mask = reg_ref_mask;
2270 map->reg_pin_mask = reg_pin_mask;
2271 map->nref_regs = nref_regs;
2272 map->npin_regs = npin_regs;
2274 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2277 if (has_ref_slots) {
2278 map->stack_ref_bitmap_offset = bitmaps_offset;
2279 bitmaps_offset += stack_bitmap_size;
2281 bitmap = &bitmaps [map->stack_ref_bitmap_offset];
2282 for (i = 0; i < nslots; ++i) {
2283 for (j = 0; j < ncallsites; ++j) {
2284 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i, j))
2285 set_bit (bitmap, stack_bitmap_width, j, i - start);
2289 if (has_pin_slots) {
2290 map->stack_pin_bitmap_offset = bitmaps_offset;
2291 bitmaps_offset += stack_bitmap_size;
2293 bitmap = &bitmaps [map->stack_pin_bitmap_offset];
2294 for (i = 0; i < nslots; ++i) {
2295 for (j = 0; j < ncallsites; ++j) {
2296 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i, j))
2297 set_bit (bitmap, stack_bitmap_width, j, i - start);
2302 map->reg_ref_bitmap_offset = bitmaps_offset;
2303 bitmaps_offset += reg_ref_bitmap_size;
2305 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2307 for (i = 0; i < nregs; ++i) {
2308 if (reg_ref_mask & (1 << i)) {
2309 for (j = 0; j < ncallsites; ++j) {
2310 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i, j))
2311 set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
2318 map->reg_pin_bitmap_offset = bitmaps_offset;
2319 bitmaps_offset += reg_pin_bitmap_size;
2321 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2323 for (i = 0; i < nregs; ++i) {
2324 if (reg_pin_mask & (1 << i)) {
2325 for (j = 0; j < ncallsites; ++j) {
2326 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i, j))
2327 set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
2335 map->ncallsites = ncallsites;
2336 if (cfg->code_len < 256)
2337 map->callsite_entry_size = 1;
2338 else if (cfg->code_len < 65536)
2339 map->callsite_entry_size = 2;
2341 map->callsite_entry_size = 4;
2343 /* Encode the GC Map */
2351 encode_gc_map (map, buf, &endbuf);
2352 g_assert (endbuf - buf < 256);
2354 encoded_size = endbuf - buf;
2355 alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
2357 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2358 //emap->ref_slots = map->ref_slots;
2360 /* Encoded fixed fields */
2361 p = &emap->encoded [0];
2362 //emap->encoded_size = encoded_size;
2363 memcpy (p, buf, encoded_size);
2366 /* Callsite table */
2367 p = (guint8*)ALIGN_TO ((mgreg_t)p, map->callsite_entry_size);
2368 if (map->callsite_entry_size == 1) {
2369 guint8 *offsets = p;
2370 for (i = 0; i < ncallsites; ++i)
2371 offsets [i] = callsites [i]->pc_offset;
2372 stats.gc_callsites8_size += ncallsites * sizeof (guint8);
2373 } else if (map->callsite_entry_size == 2) {
2374 guint16 *offsets = (guint16*)p;
2375 for (i = 0; i < ncallsites; ++i)
2376 offsets [i] = callsites [i]->pc_offset;
2377 stats.gc_callsites16_size += ncallsites * sizeof (guint16);
2379 guint32 *offsets = (guint32*)p;
2380 for (i = 0; i < ncallsites; ++i)
2381 offsets [i] = callsites [i]->pc_offset;
2382 stats.gc_callsites32_size += ncallsites * sizeof (guint32);
2384 p += ncallsites * map->callsite_entry_size;
2387 memcpy (p, bitmaps, bitmaps_size);
2390 g_assert ((guint8*)p - (guint8*)emap <= alloc_size);
2392 stats.gc_maps_size += alloc_size;
2393 stats.gc_callsites_size += ncallsites * map->callsite_entry_size;
2394 stats.gc_bitmaps_size += bitmaps_size;
2395 stats.gc_map_struct_size += sizeof (GCEncodedMap) + encoded_size;
2397 cfg->jit_info->gc_info = emap;
2399 cfg->gc_map = (guint8*)emap;
2400 cfg->gc_map_size = alloc_size;
2403 stats.all_slots += nslots;
2404 stats.ref_slots += ntypes [SLOT_REF];
2405 stats.noref_slots += ntypes [SLOT_NOREF];
2406 stats.pin_slots += ntypes [SLOT_PIN];
2410 mini_gc_create_gc_map (MonoCompile *cfg)
2412 if (!cfg->compute_gc_maps)
2416 * During marking, all frames except the top frame are at a call site, and we mark the
2417 * top frame conservatively. This means that we only need to compute and record
2418 * GC maps for call sites.
2421 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2422 /* Without liveness info, the live ranges are not precise enough */
2425 mono_analyze_liveness_gc (cfg);
2427 compute_frame_size (cfg);
2431 process_spill_slots (cfg);
2432 process_other_slots (cfg);
2433 process_param_area_slots (cfg);
2434 process_variables (cfg);
2435 process_finally_clauses (cfg);
2440 #endif /* DISABLE_JIT */
2443 parse_debug_options (void)
2448 env = g_getenv ("MONO_GCMAP_DEBUG");
2452 opts = g_strsplit (env, ",", -1);
2453 for (ptr = opts; ptr && *ptr; ptr ++) {
2454 /* No options yet */
2455 fprintf (stderr, "Invalid format for the MONO_GCMAP_DEBUG env variable: '%s'\n", env);
2466 memset (&cb, 0, sizeof (cb));
2467 cb.thread_attach_func = thread_attach_func;
2468 cb.thread_detach_func = thread_detach_func;
2469 cb.thread_suspend_func = thread_suspend_func;
2470 /* Comment this out to disable precise stack marking */
2471 cb.thread_mark_func = thread_mark_func;
2472 mono_gc_set_gc_callbacks (&cb);
2474 logfile = mono_gc_get_logfile ();
2476 parse_debug_options ();
2478 mono_counters_register ("GC Maps size",
2479 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
2480 mono_counters_register ("GC Call Sites size",
2481 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
2482 mono_counters_register ("GC Bitmaps size",
2483 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
2484 mono_counters_register ("GC Map struct size",
2485 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
2486 mono_counters_register ("GC Call Sites encoded using 8 bits",
2487 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
2488 mono_counters_register ("GC Call Sites encoded using 16 bits",
2489 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
2490 mono_counters_register ("GC Call Sites encoded using 32 bits",
2491 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
2493 mono_counters_register ("GC Map slots (all)",
2494 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
2495 mono_counters_register ("GC Map slots (ref)",
2496 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.ref_slots);
2497 mono_counters_register ("GC Map slots (noref)",
2498 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.noref_slots);
2499 mono_counters_register ("GC Map slots (pin)",
2500 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
2502 mono_counters_register ("GC TLS Data size",
2503 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2505 mono_counters_register ("Stack space scanned (all)",
2506 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
2507 mono_counters_register ("Stack space scanned (native)",
2508 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_native);
2509 mono_counters_register ("Stack space scanned (other)",
2510 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_other);
2511 mono_counters_register ("Stack space scanned (using GC Maps)",
2512 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
2513 mono_counters_register ("Stack space scanned (precise)",
2514 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_precisely);
2515 mono_counters_register ("Stack space scanned (pin)",
2516 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_conservatively);
2517 mono_counters_register ("Stack space scanned (pin registers)",
2518 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_registers);
2524 mini_gc_enable_gc_maps_for_aot (void)
2536 mini_gc_init_gc_map (MonoCompile *cfg)
2541 mini_gc_create_gc_map (MonoCompile *cfg)
2546 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2551 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2555 #endif /* DISABLE_JIT */
2564 * Set GC specific options in CFG.
2567 mini_gc_init_cfg (MonoCompile *cfg)
2569 if (mono_gc_is_moving ()) {
2570 cfg->disable_ref_noref_stack_slot_share = TRUE;
2571 cfg->gen_write_barriers = TRUE;
2574 mini_gc_init_gc_map (cfg);
2577 #endif /* DISABLE_JIT */
2580 * Problems with the current code:
2581 * - the stack walk is slow
2582 * - vtypes/refs used in EH regions are treated conservatively
2583 * - if the code is finished, less pinning will be done, causing problems because
2584 * we promote all surviving objects to old-gen.
2585 * - the unwind code can't handle a method stopped inside a finally region, it thinks the caller is
2586 * another method, but in reality it is either the exception handling code or the CALL_HANDLER opcode.
2587 * This manifests in "Unable to find ip offset x in callsite list" assertions.
2588 * - the unwind code also can't handle frames which are in the epilog, since the unwind info is not
2593 * Ideas for creating smaller GC maps:
2594 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2596 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2598 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2599 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2600 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2601 * from an index entry.