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/sgen-conf.h>
19 #include <mono/metadata/gc-internal.h>
20 #include <mono/utils/mono-counters.h>
22 #define SIZEOF_SLOT ((int)sizeof (mgreg_t))
24 #define GC_BITS_PER_WORD (sizeof (mword) * 8)
26 /* Contains state needed by the GC Map construction code */
29 * This contains information about stack slots initialized in the prolog, encoded using
30 * (slot_index << 16) | slot_type. The slot_index is relative to the CFA, i.e. 0
31 * means cfa+0, 1 means cfa-4/8, etc.
33 GSList *stack_slots_from_cfa;
34 /* Same for stack slots relative to the frame pointer */
35 GSList *stack_slots_from_fp;
37 /* Number of slots in the map */
39 /* The number of registers in the map */
41 /* Min and Max offsets of the stack frame relative to fp */
42 int min_offset, max_offset;
43 /* Same for the locals area */
44 int locals_min_offset, locals_max_offset;
46 /* The call sites where this frame can be stopped during GC */
47 GCCallSite **callsites;
48 /* The number of call sites */
52 * The width of the stack bitmaps in bytes. This is not equal to the bitmap width at
53 * runtime, since it includes columns which are 0.
55 int stack_bitmap_width;
57 * A bitmap whose width equals nslots, and whose height equals ncallsites.
58 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
61 guint8 *stack_ref_bitmap;
62 /* Same for SLOT_PIN */
63 guint8 *stack_pin_bitmap;
66 * Similar bitmaps for registers. These have width MONO_MAX_IREGS in bits.
69 guint8 *reg_ref_bitmap;
70 guint8 *reg_pin_bitmap;
73 #define ALIGN_TO(val,align) ((((mgreg_t)val) + ((align) - 1)) & ~((align) - 1))
78 /* We don't support debug levels, its all-or-nothing */
79 #define DEBUG(s) do { s; fflush (logfile); } while (0)
80 #define DEBUG_ENABLED 1
87 #define DEBUG_PRECISE(s) do { s; } while (0)
88 #define DEBUG_PRECISE_ENABLED
90 #define DEBUG_PRECISE(s)
94 * Contains information collected during the conservative stack marking pass,
95 * used during the precise pass. This helps to avoid doing a stack walk twice, which
101 int frame_start_offset;
103 /* Relative to stack_start */
104 int reg_locations [MONO_MAX_IREGS];
105 #ifdef DEBUG_PRECISE_ENABLED
108 int regs [MONO_MAX_IREGS];
112 /* Max number of frames stored in the TLS data */
113 #define MAX_FRAMES 50
116 * Per-thread data kept by this module. This is stored in the GC and passed to us as
117 * parameters, instead of being stored in a TLS variable, since during a collection,
118 * only the collection thread is active.
121 MonoThreadUnwindState unwind_state;
122 MonoThreadInfo *info;
125 gpointer ref_to_track;
126 /* Number of frames collected during the !precise pass */
128 FrameInfo frames [MAX_FRAMES];
131 /* These are constant so don't store them in the GC Maps */
132 /* Number of registers stored in gc maps */
133 #define NREGS MONO_MAX_IREGS
137 * Contains information needed to mark a stack frame.
138 * This is a transient structure, created from a compressed representation on-demand.
142 * The offsets of the GC tracked area inside the stack frame relative to the frame pointer.
143 * This includes memory which is NOREF thus doesn't need GC maps.
148 * The offset relative to frame_offset where the the memory described by the GC maps
152 /* The number of stack slots in the map */
154 /* The frame pointer register */
156 /* The size of each callsite table entry */
157 guint8 callsite_entry_size;
158 guint has_pin_slots : 1;
159 guint has_ref_slots : 1;
160 guint has_ref_regs : 1;
161 guint has_pin_regs : 1;
163 /* The offsets below are into an external bitmaps array */
166 * A bitmap whose width is equal to bitmap_width, and whose height is equal to ncallsites.
167 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
170 guint32 stack_ref_bitmap_offset;
172 * Same for SLOT_PIN. It is possible that the same bit is set in both bitmaps at
173 * different callsites, if the slot starts out as PIN, and later changes to REF.
175 guint32 stack_pin_bitmap_offset;
178 * Corresponding bitmaps for registers
179 * These have width equal to the number of bits set in reg_ref_mask/reg_pin_mask.
180 * FIXME: Merge these with the normal bitmaps, i.e. reserve the first x slots for them ?
182 guint32 reg_pin_bitmap_offset;
183 guint32 reg_ref_bitmap_offset;
185 guint32 used_int_regs, reg_ref_mask, reg_pin_mask;
187 /* The number of bits set in the two masks above */
188 guint8 nref_regs, npin_regs;
191 * A bit array marking slots which contain refs.
192 * This is used only for debugging.
196 /* Callsite offsets */
197 /* These can take up a lot of space, so encode them compactly */
207 * A compressed version of GCMap. This is what gets stored in MonoJitInfo.
211 //guint8 encoded_size;
214 * The arrays below are embedded after the struct.
215 * Their address needs to be computed.
218 /* The fixed fields of the GCMap encoded using LEB128 */
219 guint8 encoded [MONO_ZERO_LEN_ARRAY];
221 /* An array of ncallsites entries, each entry is callsite_entry_size bytes long */
222 guint8 callsites [MONO_ZERO_LEN_ARRAY];
225 guint8 bitmaps [MONO_ZERO_LEN_ARRAY];
228 static int precise_frame_count [2], precise_frame_limit = -1;
229 static gboolean precise_frame_limit_inited;
235 int scanned_precisely;
236 int scanned_conservatively;
237 int scanned_registers;
247 int gc_callsites_size;
248 int gc_callsites8_size;
249 int gc_callsites16_size;
250 int gc_callsites32_size;
252 int gc_map_struct_size;
256 static JITGCStats stats;
258 static FILE *logfile;
260 static gboolean enable_gc_maps_for_aot;
263 mini_gc_enable_gc_maps_for_aot (void)
265 enable_gc_maps_for_aot = TRUE;
268 // FIXME: Move these to a shared place
271 encode_uleb128 (guint32 value, guint8 *buf, guint8 **endbuf)
276 guint8 b = value & 0x7f;
278 if (value != 0) /* more bytes to come */
286 static G_GNUC_UNUSED void
287 encode_sleb128 (gint32 value, guint8 *buf, guint8 **endbuf)
290 gboolean negative = (value < 0);
298 /* the following is unnecessary if the
299 * implementation of >>= uses an arithmetic rather
300 * than logical shift for a signed left operand
304 value |= - (1 <<(size - 7));
305 /* sign bit of byte is second high order bit (0x40) */
306 if ((value == 0 && !(byte & 0x40)) ||
307 (value == -1 && (byte & 0x40)))
317 static inline guint32
318 decode_uleb128 (guint8 *buf, guint8 **endbuf)
328 res = res | (((int)(b & 0x7f)) << shift);
340 decode_sleb128 (guint8 *buf, guint8 **endbuf)
350 res = res | (((int)(b & 0x7f)) << shift);
353 if (shift < 32 && (b & 0x40))
354 res |= - (1 << shift);
365 encode_frame_reg (int frame_reg)
368 if (frame_reg == AMD64_RSP)
370 else if (frame_reg == AMD64_RBP)
372 #elif defined(TARGET_X86)
373 if (frame_reg == X86_EBP)
375 else if (frame_reg == X86_ESP)
377 #elif defined(TARGET_ARM)
378 if (frame_reg == ARMREG_SP)
380 else if (frame_reg == ARMREG_FP)
382 #elif defined(TARGET_S390X)
383 if (frame_reg == S390_SP)
385 else if (frame_reg == S390_FP)
390 g_assert_not_reached ();
395 decode_frame_reg (int encoded)
400 else if (encoded == 1)
402 #elif defined(TARGET_X86)
405 else if (encoded == 1)
407 #elif defined(TARGET_ARM)
410 else if (encoded == 1)
412 #elif defined(TARGET_S390X)
415 else if (encoded == 1)
420 g_assert_not_reached ();
426 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15, AMD64_RDI, AMD64_RSI };
428 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15 };
430 #elif defined(TARGET_X86)
431 static int callee_saved_regs [] = { X86_EBX, X86_ESI, X86_EDI };
432 #elif defined(TARGET_ARM)
433 static int callee_saved_regs [] = { ARMREG_V1, ARMREG_V2, ARMREG_V3, ARMREG_V4, ARMREG_V5, ARMREG_V7, ARMREG_FP };
434 #elif defined(TARGET_ARM64)
436 static int callee_saved_regs [] = { };
437 #elif defined(TARGET_S390X)
438 static int callee_saved_regs [] = { s390_r6, s390_r7, s390_r8, s390_r9, s390_r10, s390_r11, s390_r12, s390_r13, s390_r14 };
442 encode_regmask (guint32 regmask)
448 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
449 if (regmask & (1 << callee_saved_regs [i])) {
451 regmask -= (1 << callee_saved_regs [i]);
454 g_assert (regmask == 0);
459 decode_regmask (guint32 regmask)
465 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i)
466 if (regmask & (1 << i))
467 res |= (1 << callee_saved_regs [i]);
474 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
477 encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
481 encode_sleb128 (map->start_offset / SIZEOF_SLOT, buf, &buf);
482 encode_sleb128 (map->end_offset / SIZEOF_SLOT, buf, &buf);
483 encode_sleb128 (map->map_offset / SIZEOF_SLOT, buf, &buf);
484 encode_uleb128 (map->nslots, buf, &buf);
485 g_assert (map->callsite_entry_size <= 4);
486 freg = encode_frame_reg (map->frame_reg);
488 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);
489 encode_uleb128 (flags, buf, &buf);
490 encode_uleb128 (encode_regmask (map->used_int_regs), buf, &buf);
491 if (map->has_ref_regs)
492 encode_uleb128 (encode_regmask (map->reg_ref_mask), buf, &buf);
493 if (map->has_pin_regs)
494 encode_uleb128 (encode_regmask (map->reg_pin_mask), buf, &buf);
495 encode_uleb128 (map->ncallsites, buf, &buf);
503 * Decode the encoded GC map representation in BUF and store the result into MAP.
506 decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
509 int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
512 map->start_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
513 map->end_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
514 map->map_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
515 map->nslots = decode_uleb128 (buf, &buf);
516 flags = decode_uleb128 (buf, &buf);
517 map->has_ref_slots = (flags & 1) ? 1 : 0;
518 map->has_pin_slots = (flags & 2) ? 1 : 0;
519 map->has_ref_regs = (flags & 4) ? 1 : 0;
520 map->has_pin_regs = (flags & 8) ? 1 : 0;
521 map->callsite_entry_size = ((flags >> 4) & 0x3) + 1;
523 map->frame_reg = decode_frame_reg (freg);
524 map->used_int_regs = decode_regmask (decode_uleb128 (buf, &buf));
525 if (map->has_ref_regs) {
526 map->reg_ref_mask = decode_regmask (decode_uleb128 (buf, &buf));
528 for (i = 0; i < NREGS; ++i)
529 if (map->reg_ref_mask & (1 << i))
533 if (map->has_pin_regs) {
534 map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
536 for (i = 0; i < NREGS; ++i)
537 if (map->reg_pin_mask & (1 << i))
541 map->ncallsites = decode_uleb128 (buf, &buf);
543 stack_bitmap_size = (ALIGN_TO (map->nslots, 8) / 8) * map->ncallsites;
544 reg_ref_bitmap_size = (ALIGN_TO (map->nref_regs, 8) / 8) * map->ncallsites;
545 reg_pin_bitmap_size = (ALIGN_TO (map->npin_regs, 8) / 8) * map->ncallsites;
547 map->stack_ref_bitmap_offset = offset;
548 if (map->has_ref_slots)
549 offset += stack_bitmap_size;
550 map->stack_pin_bitmap_offset = offset;
551 if (map->has_pin_slots)
552 offset += stack_bitmap_size;
553 map->reg_ref_bitmap_offset = offset;
554 if (map->has_ref_regs)
555 offset += reg_ref_bitmap_size;
556 map->reg_pin_bitmap_offset = offset;
557 if (map->has_pin_regs)
558 offset += reg_pin_bitmap_size;
564 thread_attach_func (void)
568 tls = g_new0 (TlsData, 1);
569 tls->tid = GetCurrentThreadId ();
570 tls->info = mono_thread_info_current ();
571 stats.tlsdata_size += sizeof (TlsData);
577 thread_detach_func (gpointer user_data)
579 TlsData *tls = user_data;
585 thread_suspend_func (gpointer user_data, void *sigctx, MonoContext *ctx)
587 TlsData *tls = user_data;
590 /* Happens during startup */
594 if (tls->tid != GetCurrentThreadId ()) {
595 /* 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, tls->info);
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)
706 MonoContext ctx, new_ctx;
709 gboolean last = TRUE;
713 guint8* fp, *p, *real_frame_start, *frame_start, *frame_end;
714 int i, pc_offset, cindex, bitmap_width;
715 int scanned = 0, scanned_precisely, scanned_conservatively, scanned_registers;
717 StackFrameInfo frame;
718 mgreg_t *reg_locations [MONO_MAX_IREGS];
719 mgreg_t *new_reg_locations [MONO_MAX_IREGS];
722 guint32 precise_regmask;
726 tls->ref_to_track = NULL;
729 /* tls == NULL can happen during startup */
730 if (mono_thread_internal_current () == NULL || !tls) {
731 mono_gc_conservatively_scan_area (stack_start, stack_end);
732 stats.scanned_stacks += stack_end - stack_start;
736 lmf = tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF];
739 /* Number of bytes scanned based on GC map data */
741 /* Number of bytes scanned precisely based on GC map data */
742 scanned_precisely = 0;
743 /* Number of bytes scanned conservatively based on GC map data */
744 scanned_conservatively = 0;
745 /* Number of bytes scanned conservatively in register save areas */
746 scanned_registers = 0;
748 /* This is one past the last address which we have scanned */
749 stack_limit = stack_start;
751 if (!tls->unwind_state.valid)
752 memset (&new_ctx, 0, sizeof (ctx));
754 memcpy (&new_ctx, &tls->unwind_state.ctx, sizeof (MonoContext));
756 memset (reg_locations, 0, sizeof (reg_locations));
757 memset (new_reg_locations, 0, sizeof (new_reg_locations));
760 if (!tls->unwind_state.valid)
763 memcpy (&ctx, &new_ctx, sizeof (ctx));
765 for (i = 0; i < MONO_MAX_IREGS; ++i) {
766 if (new_reg_locations [i]) {
768 * If the current frame saves the register, it means it might modify its
769 * value, thus the old location might not contain the same value, so
770 * we have to mark it conservatively.
772 if (reg_locations [i]) {
773 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
774 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
775 scanned_registers += SIZEOF_SLOT;
778 reg_locations [i] = new_reg_locations [i];
780 DEBUG (fprintf (logfile, "\treg %s is now at location %p.\n", mono_arch_regname (i), reg_locations [i]));
784 g_assert ((mgreg_t)stack_limit % SIZEOF_SLOT == 0);
786 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);
792 // FIXME: For skipped frames, scan the param area of the parent frame conservatively ?
794 if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
796 * These frames are problematic for several reasons:
797 * - they are unwound through an LMF, and we have no precise register tracking for those.
798 * - the LMF might not contain a precise ip, so we can't compute the call site.
799 * - the LMF only unwinds to the wrapper frame, so we get these methods twice.
801 DEBUG (fprintf (logfile, "Mark(0): <Managed-to-native transition>\n"));
802 for (i = 0; i < MONO_MAX_IREGS; ++i) {
803 if (reg_locations [i]) {
804 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
805 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
806 scanned_registers += SIZEOF_SLOT;
808 reg_locations [i] = NULL;
809 new_reg_locations [i] = NULL;
816 method = jinfo_get_method (ji);
820 /* The last frame can be in any state so mark conservatively */
823 DEBUG (char *fname = mono_method_full_name (method, TRUE); fprintf (logfile, "Mark(0): %s+0x%x (%p)\n", fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx)); g_free (fname));
825 DEBUG (fprintf (logfile, "\t <Last frame>\n"));
828 * new_reg_locations is not precise when a method is interrupted during its epilog, so clear it.
830 for (i = 0; i < MONO_MAX_IREGS; ++i) {
831 if (reg_locations [i]) {
832 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
833 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
834 scanned_registers += SIZEOF_SLOT;
836 if (new_reg_locations [i]) {
837 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), new_reg_locations [i]));
838 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
839 scanned_registers += SIZEOF_SLOT;
841 reg_locations [i] = NULL;
842 new_reg_locations [i] = NULL;
847 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
849 /* These frames are very problematic */
850 if (method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
851 DEBUG (char *fname = mono_method_full_name (method, TRUE); fprintf (logfile, "Mark(0): %s+0x%x (%p)\n", fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx)); g_free (fname));
852 DEBUG (fprintf (logfile, "\tSkip.\n"));
856 /* All the other frames are at a call site */
858 if (tls->nframes == MAX_FRAMES) {
860 * Can't save information since the array is full. So scan the rest of the
861 * stack conservatively.
863 DEBUG (fprintf (logfile, "Mark (0): Frame stack full.\n"));
867 /* Scan the frame of this method */
870 * A frame contains the following:
875 * - localloc-ed memory
877 g_assert (pc_offset >= 0);
882 DEBUG (char *fname = mono_method_full_name (jinfo_get_method (ji), TRUE); fprintf (logfile, "Mark(0): %s+0x%x (%p)\n", fname, pc_offset, (gpointer)MONO_CONTEXT_GET_IP (&ctx)); g_free (fname));
883 DEBUG (fprintf (logfile, "\tNo GC Map.\n"));
887 /* The embedded callsite table requires this */
888 g_assert (((mgreg_t)emap % 4) == 0);
891 * Debugging aid to control the number of frames scanned precisely
893 if (!precise_frame_limit_inited) {
894 if (g_getenv ("MONO_PRECISE_COUNT"))
895 precise_frame_limit = atoi (g_getenv ("MONO_PRECISE_COUNT"));
896 precise_frame_limit_inited = TRUE;
899 if (precise_frame_limit != -1) {
900 if (precise_frame_count [FALSE] == precise_frame_limit)
901 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (method, TRUE));
902 if (precise_frame_count [FALSE] > precise_frame_limit)
905 precise_frame_count [FALSE] ++;
907 /* Decode the encoded GC map */
909 memset (map, 0, sizeof (GCMap));
910 decode_gc_map (&emap->encoded [0], map, &p);
911 p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
912 map->callsites.offsets8 = p;
913 p += map->callsite_entry_size * map->ncallsites;
916 fp = (guint8*)get_frame_pointer (&ctx, map->frame_reg);
918 real_frame_start = fp + map->start_offset;
919 frame_start = fp + map->start_offset + map->map_offset;
920 frame_end = fp + map->end_offset;
922 DEBUG (char *fname = mono_method_full_name (jinfo_get_method (ji), 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));
924 /* Find the callsite index */
925 if (map->callsite_entry_size == 1) {
926 for (i = 0; i < map->ncallsites; ++i)
927 /* ip points inside the call instruction */
928 if (map->callsites.offsets8 [i] == pc_offset + 1)
930 } else if (map->callsite_entry_size == 2) {
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.offsets16 [i] == pc_offset + 1)
937 // FIXME: Use a binary search
938 for (i = 0; i < map->ncallsites; ++i)
939 /* ip points inside the call instruction */
940 if (map->callsites.offsets32 [i] == pc_offset + 1)
943 if (i == map->ncallsites) {
944 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (method, TRUE));
945 g_assert_not_reached ();
950 * This is not neccessary true on x86 because frames have a different size at each
953 //g_assert (real_frame_start >= stack_limit);
955 if (real_frame_start > stack_limit) {
956 /* This scans the previously skipped frames as well */
957 DEBUG (fprintf (logfile, "\tscan area %p-%p (%d).\n", stack_limit, real_frame_start, (int)(real_frame_start - stack_limit)));
958 mono_gc_conservatively_scan_area (stack_limit, real_frame_start);
959 stats.scanned_other += real_frame_start - stack_limit;
962 /* Mark stack slots */
963 if (map->has_pin_slots) {
964 int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
965 guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
970 for (i = 0; i < map->nslots; ++i) {
971 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
973 DEBUG (fprintf (logfile, "\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
974 mono_gc_conservatively_scan_area (p, p + SIZEOF_SLOT);
975 scanned_conservatively += SIZEOF_SLOT;
977 scanned_precisely += SIZEOF_SLOT;
982 scanned_precisely += (map->nslots * SIZEOF_SLOT);
985 /* The area outside of start-end is NOREF */
986 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * SIZEOF_SLOT);
989 precise_regmask = map->used_int_regs | (1 << map->frame_reg);
990 if (map->has_pin_regs) {
991 int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
992 guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
994 for (i = 0; i < NREGS; ++i) {
995 if (!(map->used_int_regs & (1 << i)))
998 if (!(map->reg_pin_mask & (1 << i)))
1001 if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
1002 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
1003 precise_regmask &= ~(1 << i);
1009 scanned += map->end_offset - map->start_offset;
1011 g_assert (scanned == scanned_precisely + scanned_conservatively);
1013 stack_limit = frame_end;
1015 /* Save information for the precise pass */
1016 fi = &tls->frames [tls->nframes];
1017 fi->nslots = map->nslots;
1018 bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
1019 if (map->has_ref_slots)
1020 fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
1023 fi->frame_start_offset = frame_start - stack_start;
1024 fi->nreg_locations = 0;
1025 DEBUG_PRECISE (fi->ji = ji);
1026 DEBUG_PRECISE (fi->fp = fp);
1028 if (map->has_ref_regs) {
1029 int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
1030 guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
1032 for (i = 0; i < NREGS; ++i) {
1033 if (!(map->reg_ref_mask & (1 << i)))
1036 if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
1037 DEBUG_PRECISE (fi->regs [fi->nreg_locations] = i);
1038 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
1039 fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
1040 fi->nreg_locations ++;
1047 * Clear locations of precisely tracked registers.
1049 if (precise_regmask) {
1050 for (i = 0; i < NREGS; ++i) {
1051 if (precise_regmask & (1 << i)) {
1053 * The method uses this register, and we have precise info for it.
1054 * This means the location will be scanned precisely.
1055 * Tell the code at the beginning of the loop that this location is
1058 if (reg_locations [i])
1059 DEBUG (fprintf (logfile, "\treg %s at location %p (==%p) is precise.\n", mono_arch_regname (i), reg_locations [i], (gpointer)*reg_locations [i]));
1060 reg_locations [i] = NULL;
1068 /* Scan the remaining register save locations */
1069 for (i = 0; i < MONO_MAX_IREGS; ++i) {
1070 if (reg_locations [i]) {
1071 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", reg_locations [i]));
1072 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
1073 scanned_registers += SIZEOF_SLOT;
1075 if (new_reg_locations [i]) {
1076 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", new_reg_locations [i]));
1077 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
1078 scanned_registers += SIZEOF_SLOT;
1082 if (stack_limit < stack_end) {
1083 DEBUG (fprintf (logfile, "\tscan remaining stack %p-%p (%d).\n", stack_limit, stack_end, (int)(stack_end - stack_limit)));
1084 mono_gc_conservatively_scan_area (stack_limit, stack_end);
1085 stats.scanned_native += stack_end - stack_limit;
1088 DEBUG (fprintf (logfile, "Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
1090 stats.scanned_stacks += stack_end - stack_start;
1091 stats.scanned += scanned;
1092 stats.scanned_precisely += scanned_precisely;
1093 stats.scanned_conservatively += scanned_conservatively;
1094 stats.scanned_registers += scanned_registers;
1096 //mono_gc_conservatively_scan_area (stack_start, stack_end);
1102 * Mark a thread stack precisely based on information saved during the conservative
1106 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end, void *gc_data)
1110 guint8 *frame_start;
1115 if (!tls->unwind_state.valid)
1118 for (findex = 0; findex < tls->nframes; findex ++) {
1119 /* Load information saved by the !precise pass */
1120 fi = &tls->frames [findex];
1121 frame_start = stack_start + fi->frame_start_offset;
1123 DEBUG (char *fname = mono_method_full_name (jinfo_get_method (fi->ji), TRUE); fprintf (logfile, "Mark(1): %s\n", fname); g_free (fname));
1126 * FIXME: Add a function to mark using a bitmap, to avoid doing a
1127 * call for each object.
1130 /* Mark stack slots */
1132 guint8 *ref_bitmap = fi->bitmap;
1135 for (i = 0; i < fi->nslots; ++i) {
1136 MonoObject **ptr = (MonoObject**)(frame_start + (i * SIZEOF_SLOT));
1138 live = ref_bitmap [i / 8] & (1 << (i % 8));
1141 MonoObject *obj = *ptr;
1143 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1144 *ptr = mono_gc_scan_object (obj, gc_data);
1145 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1147 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1152 * This is disabled because the pointer takes up a lot of space.
1153 * Stack slots might be shared between ref and non-ref variables ?
1155 if (map->ref_slots [i / 8] & (1 << (i % 8))) {
1156 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));
1158 * Fail fast if the live range is incorrect, and
1159 * the JITted code tries to access this object
1168 /* Mark registers */
1171 * Registers are different from stack slots, they have no address where they
1172 * are stored. Instead, some frame below this frame in the stack saves them
1173 * in its prolog to the stack. We can mark this location precisely.
1175 for (i = 0; i < fi->nreg_locations; ++i) {
1177 * reg_locations [i] contains the address of the stack slot where
1178 * a reg was last saved, so mark that slot.
1180 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1181 MonoObject *obj = *ptr;
1184 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
1185 *ptr = mono_gc_scan_object (obj, gc_data);
1186 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1188 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p\n", mono_arch_regname (fi->regs [i]), ptr, obj));
1194 * Debugging aid to check for missed refs.
1196 if (tls->ref_to_track) {
1199 for (p = (mgreg_t*)stack_start; p < (mgreg_t*)stack_end; ++p)
1200 if (*p == (mgreg_t)tls->ref_to_track)
1201 printf ("REF AT %p.\n", p);
1208 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1209 * call, and TRUE at the second. USER_DATA points to a TlsData
1210 * structure filled up by thread_suspend_func.
1213 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise, void *gc_data)
1215 TlsData *tls = user_data;
1217 DEBUG (fprintf (logfile, "****************************************\n"));
1218 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));
1219 DEBUG (fprintf (logfile, "****************************************\n"));
1222 conservative_pass (tls, stack_start, stack_end);
1224 precise_pass (tls, stack_start, stack_end, gc_data);
1230 mini_gc_init_gc_map (MonoCompile *cfg)
1232 if (COMPILE_LLVM (cfg))
1235 if (!mono_gc_is_moving ())
1238 if (cfg->compile_aot) {
1239 if (!enable_gc_maps_for_aot)
1241 } else if (!mono_gc_precise_stack_mark_enabled ())
1245 /* Debugging support */
1247 static int precise_count;
1250 if (g_getenv ("MONO_GCMAP_COUNT")) {
1251 if (precise_count == atoi (g_getenv ("MONO_GCMAP_COUNT")))
1252 printf ("LAST: %s\n", mono_method_full_name (cfg->method, TRUE));
1253 if (precise_count > atoi (g_getenv ("MONO_GCMAP_COUNT")))
1259 cfg->compute_gc_maps = TRUE;
1261 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1265 * mini_gc_set_slot_type_from_fp:
1267 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1268 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1269 * need to call this function for those slots.
1272 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1274 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1276 if (!cfg->compute_gc_maps)
1279 g_assert (slot_offset % SIZEOF_SLOT == 0);
1281 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1285 * mini_gc_set_slot_type_from_cfa:
1287 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1288 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1289 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1290 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1291 * call this function for those slots.
1294 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1296 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1297 int slot = - (slot_offset / SIZEOF_SLOT);
1299 if (!cfg->compute_gc_maps)
1302 g_assert (slot_offset <= 0);
1303 g_assert (slot_offset % SIZEOF_SLOT == 0);
1305 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1309 fp_offset_to_slot (MonoCompile *cfg, int offset)
1311 MonoCompileGC *gcfg = cfg->gc_info;
1313 return (offset - gcfg->min_offset) / SIZEOF_SLOT;
1317 slot_to_fp_offset (MonoCompile *cfg, int slot)
1319 MonoCompileGC *gcfg = cfg->gc_info;
1321 return (slot * SIZEOF_SLOT) + gcfg->min_offset;
1324 static inline MONO_ALWAYS_INLINE void
1325 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1327 g_assert (slot >= 0 && slot < gcfg->nslots);
1329 if (type == SLOT_PIN) {
1330 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1331 set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1332 } else if (type == SLOT_REF) {
1333 set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1334 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1335 } else if (type == SLOT_NOREF) {
1336 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1337 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1342 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1345 guint8 *ref_bitmap, *pin_bitmap;
1350 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1351 set_slot (gcfg, slot, cindex, type);
1353 ref_bitmap = gcfg->stack_ref_bitmap;
1354 pin_bitmap = gcfg->stack_pin_bitmap;
1355 width = gcfg->stack_bitmap_width;
1358 if (type == SLOT_PIN) {
1359 memset (ref_bitmap + pos, 0, width);
1360 memset (pin_bitmap + pos, 0xff, width);
1361 } else if (type == SLOT_REF) {
1362 memset (ref_bitmap + pos, 0xff, width);
1363 memset (pin_bitmap + pos, 0, width);
1364 } else if (type == SLOT_NOREF) {
1365 memset (ref_bitmap + pos, 0, width);
1366 memset (pin_bitmap + pos, 0, width);
1371 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1375 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1376 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1377 if (callsite_offset >= from && callsite_offset < to)
1378 set_slot (gcfg, slot, cindex, type);
1383 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1385 g_assert (slot >= 0 && slot < gcfg->nregs);
1387 if (type == SLOT_PIN) {
1388 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1389 set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1390 } else if (type == SLOT_REF) {
1391 set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1392 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1393 } else if (type == SLOT_NOREF) {
1394 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1395 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1400 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1404 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1405 set_reg_slot (gcfg, slot, cindex, type);
1409 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1413 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1414 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1415 if (callsite_offset >= from && callsite_offset < to)
1416 set_reg_slot (gcfg, slot, cindex, type);
1421 process_spill_slots (MonoCompile *cfg)
1423 MonoCompileGC *gcfg = cfg->gc_info;
1428 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1429 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1430 for (l = bb->spill_slot_defs; l; l = l->next) {
1431 MonoInst *def = l->data;
1432 int spill_slot = def->inst_c0;
1433 int bank = def->inst_c1;
1434 int offset = cfg->spill_info [bank][spill_slot].offset;
1435 int slot = fp_offset_to_slot (cfg, offset);
1437 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1438 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1442 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1443 for (l = bb->spill_slot_defs; l; l = l->next) {
1444 MonoInst *def = l->data;
1445 int spill_slot = def->inst_c0;
1446 int bank = def->inst_c1;
1447 int offset = cfg->spill_info [bank][spill_slot].offset;
1448 int slot = fp_offset_to_slot (cfg, offset);
1451 if (bank == MONO_REG_INT_MP)
1457 * Extend the live interval for the GC tracked spill slots
1458 * defined in this bblock.
1459 * FIXME: This is not needed.
1461 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1463 if (cfg->verbose_level > 1)
1464 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
1468 /* Set fp spill slots to NOREF */
1469 for (i = 0; i < cfg->spill_info_len [MONO_REG_DOUBLE]; ++i) {
1470 int offset = cfg->spill_info [MONO_REG_DOUBLE][i].offset;
1476 slot = fp_offset_to_slot (cfg, offset);
1478 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1480 if (cfg->verbose_level > 1)
1481 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1484 /* Set int spill slots to NOREF */
1485 for (i = 0; i < cfg->spill_info_len [MONO_REG_INT]; ++i) {
1486 int offset = cfg->spill_info [MONO_REG_INT][i].offset;
1492 slot = fp_offset_to_slot (cfg, offset);
1494 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1495 if (cfg->verbose_level > 1)
1496 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1501 * process_other_slots:
1503 * Process stack slots registered using mini_gc_set_slot_type_... ().
1506 process_other_slots (MonoCompile *cfg)
1508 MonoCompileGC *gcfg = cfg->gc_info;
1511 /* Relative to the CFA */
1512 for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
1513 guint data = GPOINTER_TO_UINT (l->data);
1514 int cfa_slot = data >> 16;
1515 GCSlotType type = data & 0xff;
1519 * Map the cfa relative slot to an fp relative slot.
1520 * slot_addr == cfa - <cfa_slot>*4/8
1521 * fp + cfa_offset == cfa
1522 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1524 slot = (cfg->cfa_offset / SIZEOF_SLOT) - cfa_slot - (gcfg->min_offset / SIZEOF_SLOT);
1526 set_slot_everywhere (gcfg, slot, type);
1528 if (cfg->verbose_level > 1) {
1529 int fp_offset = slot_to_fp_offset (cfg, slot);
1530 if (type == SLOT_NOREF)
1531 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));
1535 /* Relative to the FP */
1536 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1537 gint data = GPOINTER_TO_INT (l->data);
1538 int offset = data >> 16;
1539 GCSlotType type = data & 0xff;
1542 slot = fp_offset_to_slot (cfg, offset);
1544 set_slot_everywhere (gcfg, slot, type);
1546 /* Liveness for these slots is handled by process_spill_slots () */
1548 if (cfg->verbose_level > 1) {
1549 if (type == SLOT_REF)
1550 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
1551 else if (type == SLOT_NOREF)
1552 printf ("\tnoref slot at 0x%x(fp) (slot = %d)\n", offset, slot);
1558 get_vtype_bitmap (MonoType *t, int *numbits)
1560 MonoClass *klass = mono_class_from_mono_type (t);
1562 if (klass->generic_container || mono_class_is_open_constructed_type (t)) {
1563 /* FIXME: Generic sharing */
1566 mono_class_compute_gc_descriptor (klass);
1568 return mono_gc_get_bitmap_for_descr (klass->gc_descr, numbits);
1572 static inline const char*
1573 get_offset_sign (int offset)
1575 return offset < 0 ? "-" : "+";
1579 get_offset_val (int offset)
1581 return offset < 0 ? (- offset) : offset;
1585 process_variables (MonoCompile *cfg)
1587 MonoCompileGC *gcfg = cfg->gc_info;
1588 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1589 int i, locals_min_slot, locals_max_slot, cindex;
1593 int locals_min_offset = gcfg->locals_min_offset;
1594 int locals_max_offset = gcfg->locals_max_offset;
1596 /* Slots for locals are NOREF by default */
1597 locals_min_slot = (locals_min_offset - gcfg->min_offset) / SIZEOF_SLOT;
1598 locals_max_slot = (locals_max_offset - gcfg->min_offset) / SIZEOF_SLOT;
1599 for (i = locals_min_slot; i < locals_max_slot; ++i) {
1600 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1604 * Compute the offset where variables are initialized in the first bblock, if any.
1606 pc_offsets = g_new0 (int, cfg->next_vreg);
1608 bb = cfg->bb_entry->next_bb;
1609 MONO_BB_FOR_EACH_INS (bb, tmp) {
1610 if (tmp->opcode == OP_GC_LIVENESS_DEF) {
1611 int vreg = tmp->inst_c1;
1612 if (pc_offsets [vreg] == 0) {
1613 g_assert (tmp->backend.pc_offset > 0);
1614 pc_offsets [vreg] = tmp->backend.pc_offset;
1620 * Stack slots holding arguments are initialized in the prolog.
1621 * This means we can treat them alive for the whole method.
1623 for (i = 0; i < cfg->num_varinfo; i++) {
1624 MonoInst *ins = cfg->varinfo [i];
1625 MonoType *t = ins->inst_vtype;
1628 gboolean byref, is_this = FALSE;
1629 gboolean is_arg = i < cfg->locals_start;
1631 if (ins == cfg->ret) {
1632 if (!(ins->opcode == OP_REGOFFSET && MONO_TYPE_ISSTRUCT (t)))
1636 vmv = MONO_VARINFO (cfg, i);
1638 /* For some reason, 'this' is byref */
1639 if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
1640 t = &cfg->method->klass->byval_arg;
1646 if (ins->opcode == OP_REGVAR) {
1648 GCSlotType slot_type;
1650 t = mini_type_get_underlying_type (NULL, t);
1653 g_assert (hreg < MONO_MAX_IREGS);
1656 slot_type = SLOT_PIN;
1658 slot_type = mini_type_is_reference (cfg, t) ? SLOT_REF : SLOT_NOREF;
1660 if (slot_type == SLOT_PIN) {
1661 /* These have no live interval, be conservative */
1662 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1665 * Unlike variables allocated to the stack, we generate liveness info
1666 * for noref vars in registers in mono_spill_global_vars (), because
1667 * knowing that a register doesn't contain a ref allows us to mark its save
1668 * locations precisely.
1670 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1671 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1672 set_reg_slot (gcfg, hreg, cindex, slot_type);
1675 if (cfg->verbose_level > 1) {
1676 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
1682 if (ins->opcode != OP_REGOFFSET)
1685 if (ins->inst_offset % SIZEOF_SLOT != 0)
1688 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1690 if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
1691 /* These do not get stored in the prolog */
1692 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1694 if (cfg->verbose_level > 1) {
1695 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));
1700 if (MONO_TYPE_ISSTRUCT (t)) {
1702 gsize *bitmap = NULL;
1703 gboolean pin = FALSE;
1707 if (ins->backend.is_pinvoke)
1708 size = mono_class_native_size (ins->klass, NULL);
1710 size = mono_class_value_size (ins->klass, NULL);
1711 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
1713 if (cfg->verbose_level > 1)
1714 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));
1716 if (!ins->klass->has_references) {
1718 for (j = 0; j < size_in_slots; ++j)
1719 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1724 bitmap = get_vtype_bitmap (t, &numbits);
1729 * Most vtypes are marked volatile because of the LDADDR instructions,
1730 * and they have no liveness information since they are decomposed
1731 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1732 * them during VZERO decomposition.
1735 if (!pc_offsets [vmv->vreg])
1738 if (ins->backend.is_pinvoke)
1743 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1744 if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
1745 for (j = 0; j < numbits; ++j) {
1746 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1747 /* The descriptor is for the boxed object */
1748 set_slot (gcfg, (pos + j - (sizeof (MonoObject) / SIZEOF_SLOT)), cindex, pin ? SLOT_PIN : SLOT_REF);
1754 if (cfg->verbose_level > 1) {
1755 for (j = 0; j < numbits; ++j) {
1756 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD)))
1757 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)));
1761 if (cfg->verbose_level > 1)
1762 printf ("\t\tpinned\n");
1763 for (j = 0; j < size_in_slots; ++j) {
1764 set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
1773 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1774 /* Vret addr etc. */
1779 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1781 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1782 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1783 set_slot (gcfg, pos, cindex, SLOT_PIN);
1785 if (cfg->verbose_level > 1)
1786 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));
1791 * This is currently disabled, but could be enabled to debug crashes.
1794 if (t->type == MONO_TYPE_I) {
1796 * Variables created in mono_handle_global_vregs have type I, but they
1797 * could hold GC refs since the vregs they were created from might not been
1798 * marked as holding a GC ref. So be conservative.
1800 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1805 t = mini_type_get_underlying_type (NULL, t);
1807 if (!mini_type_is_reference (cfg, t)) {
1808 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1809 if (cfg->verbose_level > 1)
1810 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));
1811 if (!t->byref && sizeof (mgreg_t) == 4 && (t->type == MONO_TYPE_I8 || t->type == MONO_TYPE_U8 || t->type == MONO_TYPE_R8)) {
1812 set_slot_everywhere (gcfg, pos + 1, SLOT_NOREF);
1813 if (cfg->verbose_level > 1)
1814 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));
1819 /* 'this' is marked INDIRECT for gshared methods */
1820 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1822 * For volatile variables, treat them alive from the point they are
1823 * initialized in the first bblock until the end of the method.
1826 set_slot_everywhere (gcfg, pos, SLOT_REF);
1827 } else if (pc_offsets [vmv->vreg]) {
1828 set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
1829 set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
1831 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1833 if (cfg->verbose_level > 1)
1834 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));
1839 /* Live for the whole method */
1840 set_slot_everywhere (gcfg, pos, SLOT_REF);
1842 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1843 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1844 set_slot (gcfg, pos, cindex, SLOT_REF);
1847 if (cfg->verbose_level > 1) {
1848 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));
1852 g_free (pc_offsets);
1856 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1859 * Convert a sp relative offset to a slot index. This is
1860 * platform specific.
1863 /* fp = sp + offset */
1864 g_assert (cfg->frame_reg == AMD64_RBP);
1865 return (- cfg->arch.sp_fp_offset + sp_offset);
1866 #elif defined(TARGET_X86)
1867 /* The offset is computed from the sp at the start of the call sequence */
1868 g_assert (cfg->frame_reg == X86_EBP);
1869 #ifdef MONO_X86_NO_PUSHES
1870 return (- cfg->arch.sp_fp_offset + sp_offset);
1872 return (- cfg->arch.sp_fp_offset - sp_offset);
1881 process_param_area_slots (MonoCompile *cfg)
1883 MonoCompileGC *gcfg = cfg->gc_info;
1888 * These slots are used for passing parameters during calls. They are sp relative, not
1889 * fp relative, so they are harder to handle.
1891 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1892 /* The distance between fp and sp is not constant */
1895 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1897 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1898 GCCallSite *callsite = gcfg->callsites [cindex];
1901 for (l = callsite->param_slots; l; l = l->next) {
1902 MonoInst *def = l->data;
1903 MonoType *t = def->inst_vtype;
1904 int sp_offset = def->inst_offset;
1905 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1906 int slot = fp_offset_to_slot (cfg, fp_offset);
1910 if (MONO_TYPE_ISSTRUCT (t)) {
1911 size = mini_type_stack_size_full (cfg->generic_sharing_context, t, &align, FALSE);
1913 size = sizeof (mgreg_t);
1916 for (i = 0; i < size / sizeof (mgreg_t); ++i) {
1917 g_assert (slot + i >= 0 && slot + i < gcfg->nslots);
1918 is_param [slot + i] = TRUE;
1923 /* All param area slots are noref by default */
1924 for (i = 0; i < gcfg->nslots; ++i) {
1926 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1930 * We treat param area slots as being part of the callee's frame, to be able to handle tail calls which overwrite
1931 * the argument area of the caller.
1936 process_finally_clauses (MonoCompile *cfg)
1938 MonoCompileGC *gcfg = cfg->gc_info;
1939 GCCallSite **callsites;
1941 gboolean has_finally;
1942 int i, j, nslots, nregs;
1944 ncallsites = gcfg->ncallsites;
1945 nslots = gcfg->nslots;
1946 nregs = gcfg->nregs;
1947 callsites = gcfg->callsites;
1950 * The calls to the finally clauses don't show up in the cfg. See
1951 * test_0_liveness_8 ().
1952 * Variables accessed inside the finally clause are already marked VOLATILE by
1953 * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
1954 * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
1956 has_finally = FALSE;
1957 for (i = 0; i < cfg->header->num_clauses; ++i) {
1958 MonoExceptionClause *clause = &cfg->header->clauses [i];
1960 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1965 if (cfg->verbose_level > 1)
1966 printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
1967 for (j = 0; j < ncallsites; ++j) {
1968 MonoBasicBlock *bb = callsites [j]->bb;
1969 MonoExceptionClause *clause;
1970 gboolean is_in_finally = FALSE;
1972 for (i = 0; i < cfg->header->num_clauses; ++i) {
1973 clause = &cfg->header->clauses [i];
1975 if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
1976 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1977 is_in_finally = TRUE;
1983 if (is_in_finally) {
1984 for (i = 0; i < nslots; ++i)
1985 set_slot (gcfg, i, j, SLOT_PIN);
1986 for (i = 0; i < nregs; ++i)
1987 set_reg_slot (gcfg, i, j, SLOT_PIN);
1994 compute_frame_size (MonoCompile *cfg)
1996 int i, locals_min_offset, locals_max_offset, cfa_min_offset, cfa_max_offset;
1997 int min_offset, max_offset;
1998 MonoCompileGC *gcfg = cfg->gc_info;
1999 MonoMethodSignature *sig = mono_method_signature (cfg->method);
2002 /* Compute min/max offsets from the fp */
2005 #if defined(TARGET_AMD64) || defined(TARGET_X86) || defined(TARGET_ARM) || defined(TARGET_S390X)
2006 locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, SIZEOF_SLOT);
2007 locals_max_offset = cfg->locals_max_stack_offset;
2009 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
2013 locals_min_offset = ALIGN_TO (locals_min_offset, SIZEOF_SLOT);
2014 locals_max_offset = ALIGN_TO (locals_max_offset, SIZEOF_SLOT);
2016 min_offset = locals_min_offset;
2017 max_offset = locals_max_offset;
2020 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
2021 MonoInst *ins = cfg->args [i];
2023 if (ins->opcode == OP_REGOFFSET) {
2024 int size, size_in_slots;
2025 size = mini_type_stack_size_full (cfg->generic_sharing_context, ins->inst_vtype, NULL, ins->backend.is_pinvoke);
2026 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
2028 min_offset = MIN (min_offset, ins->inst_offset);
2029 max_offset = MAX ((int)max_offset, (int)(ins->inst_offset + (size_in_slots * SIZEOF_SLOT)));
2034 g_assert (cfg->frame_reg == cfg->cfa_reg);
2035 g_assert (cfg->cfa_offset > 0);
2037 cfa_max_offset = cfg->cfa_offset;
2039 min_offset = MIN (min_offset, cfa_min_offset);
2040 max_offset = MAX (max_offset, cfa_max_offset);
2042 /* Fp relative slots */
2043 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
2044 gint data = GPOINTER_TO_INT (l->data);
2045 int offset = data >> 16;
2047 min_offset = MIN (min_offset, offset);
2051 if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
2052 int stack_offset = ALIGN_TO (cfg->stack_offset, SIZEOF_SLOT);
2053 min_offset = MIN (min_offset, (-stack_offset));
2056 /* Param area slots */
2058 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
2059 #elif defined(TARGET_X86)
2060 #ifdef MONO_X86_NO_PUSHES
2061 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
2063 min_offset = MIN (min_offset, - (cfg->arch.sp_fp_offset + cfg->arch.param_area_size));
2065 #elif defined(TARGET_ARM)
2067 #elif defined(TARGET_s390X)
2073 gcfg->min_offset = min_offset;
2074 gcfg->max_offset = max_offset;
2075 gcfg->locals_min_offset = locals_min_offset;
2076 gcfg->locals_max_offset = locals_max_offset;
2080 init_gcfg (MonoCompile *cfg)
2082 int i, nregs, nslots;
2083 MonoCompileGC *gcfg = cfg->gc_info;
2084 GCCallSite **callsites;
2093 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2094 ncallsites += g_slist_length (bb->gc_callsites);
2096 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
2098 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2099 for (l = bb->gc_callsites; l; l = l->next)
2100 callsites [i++] = l->data;
2103 /* The callsites should already be ordered by pc offset */
2104 for (i = 1; i < ncallsites; ++i)
2105 g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
2108 * The stack frame looks like this:
2110 * <fp + max_offset> == cfa -> <end of previous frame>
2111 * <other stack slots>
2113 * <other stack slots>
2114 * fp + min_offset ->
2119 if (cfg->verbose_level > 1)
2120 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
2122 nslots = (gcfg->max_offset - gcfg->min_offset) / SIZEOF_SLOT;
2125 gcfg->nslots = nslots;
2126 gcfg->nregs = nregs;
2127 gcfg->callsites = callsites;
2128 gcfg->ncallsites = ncallsites;
2129 gcfg->stack_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
2130 gcfg->reg_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
2131 gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
2132 gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
2133 gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
2134 gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
2136 /* All slots start out as PIN */
2137 memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * nregs);
2138 for (i = 0; i < nregs; ++i) {
2140 * By default, registers are NOREF.
2141 * It is possible for a callee to save them before being defined in this method,
2142 * but the saved value is dead too, so it doesn't need to be marked.
2144 if ((cfg->used_int_regs & (1 << i)))
2145 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
2149 static inline gboolean
2150 has_bit_set (guint8 *bitmap, int width, int slot)
2153 int pos = width * slot;
2155 for (i = 0; i < width; ++i) {
2156 if (bitmap [pos + i])
2163 create_map (MonoCompile *cfg)
2166 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
2168 int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
2169 int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
2171 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
2172 MonoCompileGC *gcfg = cfg->gc_info;
2173 GCCallSite **callsites;
2175 guint8 *bitmap, *bitmaps;
2176 guint32 reg_ref_mask, reg_pin_mask;
2178 ncallsites = gcfg->ncallsites;
2179 nslots = gcfg->nslots;
2180 nregs = gcfg->nregs;
2181 callsites = gcfg->callsites;
2184 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
2185 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
2187 has_ref_slots = FALSE;
2188 has_pin_slots = FALSE;
2191 memset (ntypes, 0, sizeof (ntypes));
2192 for (i = 0; i < nslots; ++i) {
2193 gboolean has_ref = FALSE;
2194 gboolean has_pin = FALSE;
2196 if (has_bit_set (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i))
2198 if (has_bit_set (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i))
2202 has_ref_slots = TRUE;
2204 has_pin_slots = TRUE;
2207 ntypes [SLOT_REF] ++;
2209 ntypes [SLOT_PIN] ++;
2211 ntypes [SLOT_NOREF] ++;
2213 if (has_ref || has_pin) {
2220 start = end = nslots;
2222 g_assert (start != -1);
2223 g_assert (start < end);
2226 has_ref_regs = FALSE;
2227 has_pin_regs = FALSE;
2232 for (i = 0; i < nregs; ++i) {
2233 gboolean has_ref = FALSE;
2234 gboolean has_pin = FALSE;
2236 if (!(cfg->used_int_regs & (1 << i)))
2239 if (has_bit_set (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i))
2241 if (has_bit_set (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i))
2245 reg_ref_mask |= (1 << i);
2246 has_ref_regs = TRUE;
2250 reg_pin_mask |= (1 << i);
2251 has_pin_regs = TRUE;
2256 if (cfg->verbose_level > 1)
2257 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);
2259 /* Create the GC Map */
2261 /* The work bitmaps have one row for each slot, since this is how we access them during construction */
2262 stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
2263 stack_bitmap_size = stack_bitmap_width * ncallsites;
2264 reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
2265 reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
2266 reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
2267 reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
2268 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);
2270 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2272 map->frame_reg = cfg->frame_reg;
2273 map->start_offset = gcfg->min_offset;
2274 map->end_offset = gcfg->min_offset + (nslots * SIZEOF_SLOT);
2275 map->map_offset = start * SIZEOF_SLOT;
2276 map->nslots = end - start;
2277 map->has_ref_slots = has_ref_slots;
2278 map->has_pin_slots = has_pin_slots;
2279 map->has_ref_regs = has_ref_regs;
2280 map->has_pin_regs = has_pin_regs;
2281 g_assert (nregs < 32);
2282 map->used_int_regs = cfg->used_int_regs;
2283 map->reg_ref_mask = reg_ref_mask;
2284 map->reg_pin_mask = reg_pin_mask;
2285 map->nref_regs = nref_regs;
2286 map->npin_regs = npin_regs;
2288 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2291 if (has_ref_slots) {
2292 map->stack_ref_bitmap_offset = bitmaps_offset;
2293 bitmaps_offset += stack_bitmap_size;
2295 bitmap = &bitmaps [map->stack_ref_bitmap_offset];
2296 for (i = 0; i < nslots; ++i) {
2297 for (j = 0; j < ncallsites; ++j) {
2298 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i, j))
2299 set_bit (bitmap, stack_bitmap_width, j, i - start);
2303 if (has_pin_slots) {
2304 map->stack_pin_bitmap_offset = bitmaps_offset;
2305 bitmaps_offset += stack_bitmap_size;
2307 bitmap = &bitmaps [map->stack_pin_bitmap_offset];
2308 for (i = 0; i < nslots; ++i) {
2309 for (j = 0; j < ncallsites; ++j) {
2310 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i, j))
2311 set_bit (bitmap, stack_bitmap_width, j, i - start);
2316 map->reg_ref_bitmap_offset = bitmaps_offset;
2317 bitmaps_offset += reg_ref_bitmap_size;
2319 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2321 for (i = 0; i < nregs; ++i) {
2322 if (reg_ref_mask & (1 << i)) {
2323 for (j = 0; j < ncallsites; ++j) {
2324 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i, j))
2325 set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
2332 map->reg_pin_bitmap_offset = bitmaps_offset;
2333 bitmaps_offset += reg_pin_bitmap_size;
2335 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2337 for (i = 0; i < nregs; ++i) {
2338 if (reg_pin_mask & (1 << i)) {
2339 for (j = 0; j < ncallsites; ++j) {
2340 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i, j))
2341 set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
2349 map->ncallsites = ncallsites;
2350 if (cfg->code_len < 256)
2351 map->callsite_entry_size = 1;
2352 else if (cfg->code_len < 65536)
2353 map->callsite_entry_size = 2;
2355 map->callsite_entry_size = 4;
2357 /* Encode the GC Map */
2365 encode_gc_map (map, buf, &endbuf);
2366 g_assert (endbuf - buf < 256);
2368 encoded_size = endbuf - buf;
2369 alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
2371 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2372 //emap->ref_slots = map->ref_slots;
2374 /* Encoded fixed fields */
2375 p = &emap->encoded [0];
2376 //emap->encoded_size = encoded_size;
2377 memcpy (p, buf, encoded_size);
2380 /* Callsite table */
2381 p = (guint8*)ALIGN_TO ((mgreg_t)p, map->callsite_entry_size);
2382 if (map->callsite_entry_size == 1) {
2383 guint8 *offsets = p;
2384 for (i = 0; i < ncallsites; ++i)
2385 offsets [i] = callsites [i]->pc_offset;
2386 stats.gc_callsites8_size += ncallsites * sizeof (guint8);
2387 } else if (map->callsite_entry_size == 2) {
2388 guint16 *offsets = (guint16*)p;
2389 for (i = 0; i < ncallsites; ++i)
2390 offsets [i] = callsites [i]->pc_offset;
2391 stats.gc_callsites16_size += ncallsites * sizeof (guint16);
2393 guint32 *offsets = (guint32*)p;
2394 for (i = 0; i < ncallsites; ++i)
2395 offsets [i] = callsites [i]->pc_offset;
2396 stats.gc_callsites32_size += ncallsites * sizeof (guint32);
2398 p += ncallsites * map->callsite_entry_size;
2401 memcpy (p, bitmaps, bitmaps_size);
2404 g_assert ((guint8*)p - (guint8*)emap <= alloc_size);
2406 stats.gc_maps_size += alloc_size;
2407 stats.gc_callsites_size += ncallsites * map->callsite_entry_size;
2408 stats.gc_bitmaps_size += bitmaps_size;
2409 stats.gc_map_struct_size += sizeof (GCEncodedMap) + encoded_size;
2411 cfg->jit_info->gc_info = emap;
2413 cfg->gc_map = (guint8*)emap;
2414 cfg->gc_map_size = alloc_size;
2417 stats.all_slots += nslots;
2418 stats.ref_slots += ntypes [SLOT_REF];
2419 stats.noref_slots += ntypes [SLOT_NOREF];
2420 stats.pin_slots += ntypes [SLOT_PIN];
2424 mini_gc_create_gc_map (MonoCompile *cfg)
2426 if (!cfg->compute_gc_maps)
2430 * During marking, all frames except the top frame are at a call site, and we mark the
2431 * top frame conservatively. This means that we only need to compute and record
2432 * GC maps for call sites.
2435 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2436 /* Without liveness info, the live ranges are not precise enough */
2439 mono_analyze_liveness_gc (cfg);
2441 compute_frame_size (cfg);
2445 process_spill_slots (cfg);
2446 process_other_slots (cfg);
2447 process_param_area_slots (cfg);
2448 process_variables (cfg);
2449 process_finally_clauses (cfg);
2454 #endif /* DISABLE_JIT */
2457 parse_debug_options (void)
2462 env = g_getenv ("MONO_GCMAP_DEBUG");
2466 opts = g_strsplit (env, ",", -1);
2467 for (ptr = opts; ptr && *ptr; ptr ++) {
2468 /* No options yet */
2469 fprintf (stderr, "Invalid format for the MONO_GCMAP_DEBUG env variable: '%s'\n", env);
2480 memset (&cb, 0, sizeof (cb));
2481 cb.thread_attach_func = thread_attach_func;
2482 cb.thread_detach_func = thread_detach_func;
2483 cb.thread_suspend_func = thread_suspend_func;
2484 /* Comment this out to disable precise stack marking */
2485 cb.thread_mark_func = thread_mark_func;
2486 mono_gc_set_gc_callbacks (&cb);
2488 logfile = mono_gc_get_logfile ();
2490 parse_debug_options ();
2492 mono_counters_register ("GC Maps size",
2493 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
2494 mono_counters_register ("GC Call Sites size",
2495 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
2496 mono_counters_register ("GC Bitmaps size",
2497 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
2498 mono_counters_register ("GC Map struct size",
2499 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
2500 mono_counters_register ("GC Call Sites encoded using 8 bits",
2501 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
2502 mono_counters_register ("GC Call Sites encoded using 16 bits",
2503 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
2504 mono_counters_register ("GC Call Sites encoded using 32 bits",
2505 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
2507 mono_counters_register ("GC Map slots (all)",
2508 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
2509 mono_counters_register ("GC Map slots (ref)",
2510 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.ref_slots);
2511 mono_counters_register ("GC Map slots (noref)",
2512 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.noref_slots);
2513 mono_counters_register ("GC Map slots (pin)",
2514 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
2516 mono_counters_register ("GC TLS Data size",
2517 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2519 mono_counters_register ("Stack space scanned (all)",
2520 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
2521 mono_counters_register ("Stack space scanned (native)",
2522 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_native);
2523 mono_counters_register ("Stack space scanned (other)",
2524 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_other);
2525 mono_counters_register ("Stack space scanned (using GC Maps)",
2526 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
2527 mono_counters_register ("Stack space scanned (precise)",
2528 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_precisely);
2529 mono_counters_register ("Stack space scanned (pin)",
2530 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_conservatively);
2531 mono_counters_register ("Stack space scanned (pin registers)",
2532 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_registers);
2538 mini_gc_enable_gc_maps_for_aot (void)
2550 mini_gc_init_gc_map (MonoCompile *cfg)
2555 mini_gc_create_gc_map (MonoCompile *cfg)
2560 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2565 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2569 #endif /* DISABLE_JIT */
2578 * Set GC specific options in CFG.
2581 mini_gc_init_cfg (MonoCompile *cfg)
2583 if (mono_gc_is_moving ()) {
2584 cfg->disable_ref_noref_stack_slot_share = TRUE;
2585 cfg->gen_write_barriers = TRUE;
2588 mini_gc_init_gc_map (cfg);
2591 #endif /* DISABLE_JIT */
2594 * Problems with the current code:
2595 * - the stack walk is slow
2596 * - vtypes/refs used in EH regions are treated conservatively
2597 * - if the code is finished, less pinning will be done, causing problems because
2598 * we promote all surviving objects to old-gen.
2599 * - the unwind code can't handle a method stopped inside a finally region, it thinks the caller is
2600 * another method, but in reality it is either the exception handling code or the CALL_HANDLER opcode.
2601 * This manifests in "Unable to find ip offset x in callsite list" assertions.
2602 * - the unwind code also can't handle frames which are in the epilog, since the unwind info is not
2607 * Ideas for creating smaller GC maps:
2608 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2610 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2612 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2613 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2614 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2615 * from an index entry.