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_ARM64)
441 static int callee_saved_regs [] = { };
442 #elif defined(TARGET_S390X)
443 static int callee_saved_regs [] = { s390_r6, s390_r7, s390_r8, s390_r9, s390_r10, s390_r11, s390_r12, s390_r13, s390_r14 };
447 encode_regmask (guint32 regmask)
453 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
454 if (regmask & (1 << callee_saved_regs [i])) {
456 regmask -= (1 << callee_saved_regs [i]);
459 g_assert (regmask == 0);
464 decode_regmask (guint32 regmask)
470 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i)
471 if (regmask & (1 << i))
472 res |= (1 << callee_saved_regs [i]);
479 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
482 encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
486 encode_sleb128 (map->start_offset / SIZEOF_SLOT, buf, &buf);
487 encode_sleb128 (map->end_offset / SIZEOF_SLOT, buf, &buf);
488 encode_sleb128 (map->map_offset / SIZEOF_SLOT, buf, &buf);
489 encode_uleb128 (map->nslots, buf, &buf);
490 g_assert (map->callsite_entry_size <= 4);
491 freg = encode_frame_reg (map->frame_reg);
493 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);
494 encode_uleb128 (flags, buf, &buf);
495 encode_uleb128 (encode_regmask (map->used_int_regs), buf, &buf);
496 if (map->has_ref_regs)
497 encode_uleb128 (encode_regmask (map->reg_ref_mask), buf, &buf);
498 if (map->has_pin_regs)
499 encode_uleb128 (encode_regmask (map->reg_pin_mask), buf, &buf);
500 encode_uleb128 (map->ncallsites, buf, &buf);
508 * Decode the encoded GC map representation in BUF and store the result into MAP.
511 decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
514 int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
517 map->start_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
518 map->end_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
519 map->map_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
520 map->nslots = decode_uleb128 (buf, &buf);
521 flags = decode_uleb128 (buf, &buf);
522 map->has_ref_slots = (flags & 1) ? 1 : 0;
523 map->has_pin_slots = (flags & 2) ? 1 : 0;
524 map->has_ref_regs = (flags & 4) ? 1 : 0;
525 map->has_pin_regs = (flags & 8) ? 1 : 0;
526 map->callsite_entry_size = ((flags >> 4) & 0x3) + 1;
528 map->frame_reg = decode_frame_reg (freg);
529 map->used_int_regs = decode_regmask (decode_uleb128 (buf, &buf));
530 if (map->has_ref_regs) {
531 map->reg_ref_mask = decode_regmask (decode_uleb128 (buf, &buf));
533 for (i = 0; i < NREGS; ++i)
534 if (map->reg_ref_mask & (1 << i))
538 if (map->has_pin_regs) {
539 map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
541 for (i = 0; i < NREGS; ++i)
542 if (map->reg_pin_mask & (1 << i))
546 map->ncallsites = decode_uleb128 (buf, &buf);
548 stack_bitmap_size = (ALIGN_TO (map->nslots, 8) / 8) * map->ncallsites;
549 reg_ref_bitmap_size = (ALIGN_TO (map->nref_regs, 8) / 8) * map->ncallsites;
550 reg_pin_bitmap_size = (ALIGN_TO (map->npin_regs, 8) / 8) * map->ncallsites;
552 map->stack_ref_bitmap_offset = offset;
553 if (map->has_ref_slots)
554 offset += stack_bitmap_size;
555 map->stack_pin_bitmap_offset = offset;
556 if (map->has_pin_slots)
557 offset += stack_bitmap_size;
558 map->reg_ref_bitmap_offset = offset;
559 if (map->has_ref_regs)
560 offset += reg_ref_bitmap_size;
561 map->reg_pin_bitmap_offset = offset;
562 if (map->has_pin_regs)
563 offset += reg_pin_bitmap_size;
569 thread_attach_func (void)
573 tls = g_new0 (TlsData, 1);
574 tls->tid = GetCurrentThreadId ();
575 tls->info = mono_thread_info_current ();
576 stats.tlsdata_size += sizeof (TlsData);
582 thread_detach_func (gpointer user_data)
584 TlsData *tls = user_data;
590 thread_suspend_func (gpointer user_data, void *sigctx, MonoContext *ctx)
592 TlsData *tls = user_data;
595 /* Happens during startup */
599 if (tls->tid != GetCurrentThreadId ()) {
600 /* Happens on osx because threads are not suspended using signals */
605 g_assert (tls->info);
609 res = mono_thread_state_init_from_handle (&tls->unwind_state, tls->info);
612 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF] = mono_get_lmf ();
614 #ifdef MONO_ARCH_HAVE_SIGCTX_TO_MONOCTX
615 mono_arch_sigctx_to_monoctx (sigctx, &tls->unwind_state.ctx);
616 tls->unwind_state.valid = TRUE;
618 tls->unwind_state.valid = FALSE;
621 memcpy (&tls->unwind_state.ctx, ctx, sizeof (MonoContext));
622 tls->unwind_state.valid = TRUE;
624 tls->unwind_state.valid = FALSE;
626 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_JIT_TLS] = mono_native_tls_get_value (mono_jit_tls_id);
627 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN] = mono_domain_get ();
630 if (!tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN]) {
631 /* Happens during startup */
632 tls->unwind_state.valid = FALSE;
637 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
640 set_bit (guint8 *bitmap, int width, int y, int x)
642 bitmap [(width * y) + (x / 8)] |= (1 << (x % 8));
646 clear_bit (guint8 *bitmap, int width, int y, int x)
648 bitmap [(width * y) + (x / 8)] &= ~(1 << (x % 8));
652 get_bit (guint8 *bitmap, int width, int y, int x)
654 return bitmap [(width * y) + (x / 8)] & (1 << (x % 8));
658 slot_type_to_string (GCSlotType type)
668 g_assert_not_reached ();
673 static inline mgreg_t
674 get_frame_pointer (MonoContext *ctx, int frame_reg)
676 #if defined(TARGET_AMD64)
677 if (frame_reg == AMD64_RSP)
679 else if (frame_reg == AMD64_RBP)
681 #elif defined(TARGET_X86)
682 if (frame_reg == X86_ESP)
684 else if (frame_reg == X86_EBP)
686 #elif defined(TARGET_ARM)
687 if (frame_reg == ARMREG_SP)
688 return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
689 else if (frame_reg == ARMREG_FP)
690 return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
691 #elif defined(TARGET_S390X)
692 if (frame_reg == S390_SP)
693 return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
694 else if (frame_reg == S390_FP)
695 return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
697 g_assert_not_reached ();
702 * conservatively_pass:
704 * Mark a thread stack conservatively and collect information needed by the precise pass.
707 conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
711 MonoContext ctx, new_ctx;
714 gboolean last = TRUE;
718 guint8* fp, *p, *real_frame_start, *frame_start, *frame_end;
719 int i, pc_offset, cindex, bitmap_width;
720 int scanned = 0, scanned_precisely, scanned_conservatively, scanned_registers;
722 StackFrameInfo frame;
723 mgreg_t *reg_locations [MONO_MAX_IREGS];
724 mgreg_t *new_reg_locations [MONO_MAX_IREGS];
727 guint32 precise_regmask;
731 tls->ref_to_track = NULL;
734 /* tls == NULL can happen during startup */
735 if (mono_thread_internal_current () == NULL || !tls) {
736 mono_gc_conservatively_scan_area (stack_start, stack_end);
737 stats.scanned_stacks += stack_end - stack_start;
741 lmf = tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF];
744 /* Number of bytes scanned based on GC map data */
746 /* Number of bytes scanned precisely based on GC map data */
747 scanned_precisely = 0;
748 /* Number of bytes scanned conservatively based on GC map data */
749 scanned_conservatively = 0;
750 /* Number of bytes scanned conservatively in register save areas */
751 scanned_registers = 0;
753 /* This is one past the last address which we have scanned */
754 stack_limit = stack_start;
756 if (!tls->unwind_state.valid)
757 memset (&new_ctx, 0, sizeof (ctx));
759 memcpy (&new_ctx, &tls->unwind_state.ctx, sizeof (MonoContext));
761 memset (reg_locations, 0, sizeof (reg_locations));
762 memset (new_reg_locations, 0, sizeof (new_reg_locations));
765 if (!tls->unwind_state.valid)
768 memcpy (&ctx, &new_ctx, sizeof (ctx));
770 for (i = 0; i < MONO_MAX_IREGS; ++i) {
771 if (new_reg_locations [i]) {
773 * If the current frame saves the register, it means it might modify its
774 * value, thus the old location might not contain the same value, so
775 * we have to mark it conservatively.
777 if (reg_locations [i]) {
778 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
779 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
780 scanned_registers += SIZEOF_SLOT;
783 reg_locations [i] = new_reg_locations [i];
785 DEBUG (fprintf (logfile, "\treg %s is now at location %p.\n", mono_arch_regname (i), reg_locations [i]));
789 g_assert ((mgreg_t)stack_limit % SIZEOF_SLOT == 0);
791 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);
797 // FIXME: For skipped frames, scan the param area of the parent frame conservatively ?
799 if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
801 * These frames are problematic for several reasons:
802 * - they are unwound through an LMF, and we have no precise register tracking for those.
803 * - the LMF might not contain a precise ip, so we can't compute the call site.
804 * - the LMF only unwinds to the wrapper frame, so we get these methods twice.
806 DEBUG (fprintf (logfile, "Mark(0): <Managed-to-native transition>\n"));
807 for (i = 0; i < MONO_MAX_IREGS; ++i) {
808 if (reg_locations [i]) {
809 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
810 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
811 scanned_registers += SIZEOF_SLOT;
813 reg_locations [i] = NULL;
814 new_reg_locations [i] = NULL;
821 method = jinfo_get_method (ji);
825 /* The last frame can be in any state so mark conservatively */
828 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));
830 DEBUG (fprintf (logfile, "\t <Last frame>\n"));
833 * new_reg_locations is not precise when a method is interrupted during its epilog, so clear it.
835 for (i = 0; i < MONO_MAX_IREGS; ++i) {
836 if (reg_locations [i]) {
837 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
838 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
839 scanned_registers += SIZEOF_SLOT;
841 if (new_reg_locations [i]) {
842 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), new_reg_locations [i]));
843 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
844 scanned_registers += SIZEOF_SLOT;
846 reg_locations [i] = NULL;
847 new_reg_locations [i] = NULL;
852 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
854 /* These frames are very problematic */
855 if (method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
856 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));
857 DEBUG (fprintf (logfile, "\tSkip.\n"));
861 /* All the other frames are at a call site */
863 if (tls->nframes == MAX_FRAMES) {
865 * Can't save information since the array is full. So scan the rest of the
866 * stack conservatively.
868 DEBUG (fprintf (logfile, "Mark (0): Frame stack full.\n"));
872 /* Scan the frame of this method */
875 * A frame contains the following:
880 * - localloc-ed memory
882 g_assert (pc_offset >= 0);
887 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));
888 DEBUG (fprintf (logfile, "\tNo GC Map.\n"));
892 /* The embedded callsite table requires this */
893 g_assert (((mgreg_t)emap % 4) == 0);
896 * Debugging aid to control the number of frames scanned precisely
898 if (!precise_frame_limit_inited) {
899 if (g_getenv ("MONO_PRECISE_COUNT"))
900 precise_frame_limit = atoi (g_getenv ("MONO_PRECISE_COUNT"));
901 precise_frame_limit_inited = TRUE;
904 if (precise_frame_limit != -1) {
905 if (precise_frame_count [FALSE] == precise_frame_limit)
906 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (method, TRUE));
907 if (precise_frame_count [FALSE] > precise_frame_limit)
910 precise_frame_count [FALSE] ++;
912 /* Decode the encoded GC map */
914 memset (map, 0, sizeof (GCMap));
915 decode_gc_map (&emap->encoded [0], map, &p);
916 p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
917 map->callsites.offsets8 = p;
918 p += map->callsite_entry_size * map->ncallsites;
921 fp = (guint8*)get_frame_pointer (&ctx, map->frame_reg);
923 real_frame_start = fp + map->start_offset;
924 frame_start = fp + map->start_offset + map->map_offset;
925 frame_end = fp + map->end_offset;
927 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));
929 /* Find the callsite index */
930 if (map->callsite_entry_size == 1) {
931 for (i = 0; i < map->ncallsites; ++i)
932 /* ip points inside the call instruction */
933 if (map->callsites.offsets8 [i] == pc_offset + 1)
935 } else if (map->callsite_entry_size == 2) {
936 // FIXME: Use a binary search
937 for (i = 0; i < map->ncallsites; ++i)
938 /* ip points inside the call instruction */
939 if (map->callsites.offsets16 [i] == pc_offset + 1)
942 // FIXME: Use a binary search
943 for (i = 0; i < map->ncallsites; ++i)
944 /* ip points inside the call instruction */
945 if (map->callsites.offsets32 [i] == pc_offset + 1)
948 if (i == map->ncallsites) {
949 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (method, TRUE));
950 g_assert_not_reached ();
955 * This is not neccessary true on x86 because frames have a different size at each
958 //g_assert (real_frame_start >= stack_limit);
960 if (real_frame_start > stack_limit) {
961 /* This scans the previously skipped frames as well */
962 DEBUG (fprintf (logfile, "\tscan area %p-%p (%d).\n", stack_limit, real_frame_start, (int)(real_frame_start - stack_limit)));
963 mono_gc_conservatively_scan_area (stack_limit, real_frame_start);
964 stats.scanned_other += real_frame_start - stack_limit;
967 /* Mark stack slots */
968 if (map->has_pin_slots) {
969 int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
970 guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
975 for (i = 0; i < map->nslots; ++i) {
976 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
978 DEBUG (fprintf (logfile, "\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
979 mono_gc_conservatively_scan_area (p, p + SIZEOF_SLOT);
980 scanned_conservatively += SIZEOF_SLOT;
982 scanned_precisely += SIZEOF_SLOT;
987 scanned_precisely += (map->nslots * SIZEOF_SLOT);
990 /* The area outside of start-end is NOREF */
991 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * SIZEOF_SLOT);
994 precise_regmask = map->used_int_regs | (1 << map->frame_reg);
995 if (map->has_pin_regs) {
996 int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
997 guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
999 for (i = 0; i < NREGS; ++i) {
1000 if (!(map->used_int_regs & (1 << i)))
1003 if (!(map->reg_pin_mask & (1 << i)))
1006 if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
1007 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
1008 precise_regmask &= ~(1 << i);
1014 scanned += map->end_offset - map->start_offset;
1016 g_assert (scanned == scanned_precisely + scanned_conservatively);
1018 stack_limit = frame_end;
1020 /* Save information for the precise pass */
1021 fi = &tls->frames [tls->nframes];
1022 fi->nslots = map->nslots;
1023 bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
1024 if (map->has_ref_slots)
1025 fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
1028 fi->frame_start_offset = frame_start - stack_start;
1029 fi->nreg_locations = 0;
1030 DEBUG_PRECISE (fi->ji = ji);
1031 DEBUG_PRECISE (fi->fp = fp);
1033 if (map->has_ref_regs) {
1034 int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
1035 guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
1037 for (i = 0; i < NREGS; ++i) {
1038 if (!(map->reg_ref_mask & (1 << i)))
1041 if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
1042 DEBUG_PRECISE (fi->regs [fi->nreg_locations] = i);
1043 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
1044 fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
1045 fi->nreg_locations ++;
1052 * Clear locations of precisely tracked registers.
1054 if (precise_regmask) {
1055 for (i = 0; i < NREGS; ++i) {
1056 if (precise_regmask & (1 << i)) {
1058 * The method uses this register, and we have precise info for it.
1059 * This means the location will be scanned precisely.
1060 * Tell the code at the beginning of the loop that this location is
1063 if (reg_locations [i])
1064 DEBUG (fprintf (logfile, "\treg %s at location %p (==%p) is precise.\n", mono_arch_regname (i), reg_locations [i], (gpointer)*reg_locations [i]));
1065 reg_locations [i] = NULL;
1073 /* Scan the remaining register save locations */
1074 for (i = 0; i < MONO_MAX_IREGS; ++i) {
1075 if (reg_locations [i]) {
1076 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", reg_locations [i]));
1077 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
1078 scanned_registers += SIZEOF_SLOT;
1080 if (new_reg_locations [i]) {
1081 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", new_reg_locations [i]));
1082 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
1083 scanned_registers += SIZEOF_SLOT;
1087 if (stack_limit < stack_end) {
1088 DEBUG (fprintf (logfile, "\tscan remaining stack %p-%p (%d).\n", stack_limit, stack_end, (int)(stack_end - stack_limit)));
1089 mono_gc_conservatively_scan_area (stack_limit, stack_end);
1090 stats.scanned_native += stack_end - stack_limit;
1093 DEBUG (fprintf (logfile, "Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
1095 stats.scanned_stacks += stack_end - stack_start;
1096 stats.scanned += scanned;
1097 stats.scanned_precisely += scanned_precisely;
1098 stats.scanned_conservatively += scanned_conservatively;
1099 stats.scanned_registers += scanned_registers;
1101 //mono_gc_conservatively_scan_area (stack_start, stack_end);
1107 * Mark a thread stack precisely based on information saved during the conservative
1111 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
1115 guint8 *frame_start;
1120 if (!tls->unwind_state.valid)
1123 for (findex = 0; findex < tls->nframes; findex ++) {
1124 /* Load information saved by the !precise pass */
1125 fi = &tls->frames [findex];
1126 frame_start = stack_start + fi->frame_start_offset;
1128 DEBUG (char *fname = mono_method_full_name (jinfo_get_method (fi->ji), TRUE); fprintf (logfile, "Mark(1): %s\n", fname); g_free (fname));
1131 * FIXME: Add a function to mark using a bitmap, to avoid doing a
1132 * call for each object.
1135 /* Mark stack slots */
1137 guint8 *ref_bitmap = fi->bitmap;
1140 for (i = 0; i < fi->nslots; ++i) {
1141 MonoObject **ptr = (MonoObject**)(frame_start + (i * SIZEOF_SLOT));
1143 live = ref_bitmap [i / 8] & (1 << (i % 8));
1146 MonoObject *obj = *ptr;
1148 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1149 *ptr = mono_gc_scan_object (obj);
1150 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1152 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1157 * This is disabled because the pointer takes up a lot of space.
1158 * Stack slots might be shared between ref and non-ref variables ?
1160 if (map->ref_slots [i / 8] & (1 << (i % 8))) {
1161 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));
1163 * Fail fast if the live range is incorrect, and
1164 * the JITted code tries to access this object
1173 /* Mark registers */
1176 * Registers are different from stack slots, they have no address where they
1177 * are stored. Instead, some frame below this frame in the stack saves them
1178 * in its prolog to the stack. We can mark this location precisely.
1180 for (i = 0; i < fi->nreg_locations; ++i) {
1182 * reg_locations [i] contains the address of the stack slot where
1183 * a reg was last saved, so mark that slot.
1185 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1186 MonoObject *obj = *ptr;
1189 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
1190 *ptr = mono_gc_scan_object (obj);
1191 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1193 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p\n", mono_arch_regname (fi->regs [i]), ptr, obj));
1199 * Debugging aid to check for missed refs.
1201 if (tls->ref_to_track) {
1204 for (p = (mgreg_t*)stack_start; p < (mgreg_t*)stack_end; ++p)
1205 if (*p == (mgreg_t)tls->ref_to_track)
1206 printf ("REF AT %p.\n", p);
1213 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1214 * call, and TRUE at the second. USER_DATA points to a TlsData
1215 * structure filled up by thread_suspend_func.
1218 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise)
1220 TlsData *tls = user_data;
1222 DEBUG (fprintf (logfile, "****************************************\n"));
1223 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));
1224 DEBUG (fprintf (logfile, "****************************************\n"));
1227 conservative_pass (tls, stack_start, stack_end);
1229 precise_pass (tls, stack_start, stack_end);
1235 mini_gc_init_gc_map (MonoCompile *cfg)
1237 if (COMPILE_LLVM (cfg))
1240 if (!mono_gc_is_moving ())
1243 if (cfg->compile_aot) {
1244 if (!enable_gc_maps_for_aot)
1246 } else if (!mono_gc_precise_stack_mark_enabled ())
1250 /* Debugging support */
1252 static int precise_count;
1255 if (g_getenv ("MONO_GCMAP_COUNT")) {
1256 if (precise_count == atoi (g_getenv ("MONO_GCMAP_COUNT")))
1257 printf ("LAST: %s\n", mono_method_full_name (cfg->method, TRUE));
1258 if (precise_count > atoi (g_getenv ("MONO_GCMAP_COUNT")))
1264 cfg->compute_gc_maps = TRUE;
1266 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1270 * mini_gc_set_slot_type_from_fp:
1272 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1273 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1274 * need to call this function for those slots.
1277 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1279 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1281 if (!cfg->compute_gc_maps)
1284 g_assert (slot_offset % SIZEOF_SLOT == 0);
1286 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1290 * mini_gc_set_slot_type_from_cfa:
1292 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1293 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1294 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1295 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1296 * call this function for those slots.
1299 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1301 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1302 int slot = - (slot_offset / SIZEOF_SLOT);
1304 if (!cfg->compute_gc_maps)
1307 g_assert (slot_offset <= 0);
1308 g_assert (slot_offset % SIZEOF_SLOT == 0);
1310 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1314 fp_offset_to_slot (MonoCompile *cfg, int offset)
1316 MonoCompileGC *gcfg = cfg->gc_info;
1318 return (offset - gcfg->min_offset) / SIZEOF_SLOT;
1322 slot_to_fp_offset (MonoCompile *cfg, int slot)
1324 MonoCompileGC *gcfg = cfg->gc_info;
1326 return (slot * SIZEOF_SLOT) + gcfg->min_offset;
1329 static inline MONO_ALWAYS_INLINE void
1330 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1332 g_assert (slot >= 0 && slot < gcfg->nslots);
1334 if (type == SLOT_PIN) {
1335 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1336 set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1337 } else if (type == SLOT_REF) {
1338 set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1339 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1340 } else if (type == SLOT_NOREF) {
1341 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1342 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1347 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1350 guint8 *ref_bitmap, *pin_bitmap;
1355 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1356 set_slot (gcfg, slot, cindex, type);
1358 ref_bitmap = gcfg->stack_ref_bitmap;
1359 pin_bitmap = gcfg->stack_pin_bitmap;
1360 width = gcfg->stack_bitmap_width;
1363 if (type == SLOT_PIN) {
1364 memset (ref_bitmap + pos, 0, width);
1365 memset (pin_bitmap + pos, 0xff, width);
1366 } else if (type == SLOT_REF) {
1367 memset (ref_bitmap + pos, 0xff, width);
1368 memset (pin_bitmap + pos, 0, width);
1369 } else if (type == SLOT_NOREF) {
1370 memset (ref_bitmap + pos, 0, width);
1371 memset (pin_bitmap + pos, 0, width);
1376 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1380 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1381 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1382 if (callsite_offset >= from && callsite_offset < to)
1383 set_slot (gcfg, slot, cindex, type);
1388 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1390 g_assert (slot >= 0 && slot < gcfg->nregs);
1392 if (type == SLOT_PIN) {
1393 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1394 set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1395 } else if (type == SLOT_REF) {
1396 set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1397 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1398 } else if (type == SLOT_NOREF) {
1399 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1400 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1405 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1409 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1410 set_reg_slot (gcfg, slot, cindex, type);
1414 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1418 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1419 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1420 if (callsite_offset >= from && callsite_offset < to)
1421 set_reg_slot (gcfg, slot, cindex, type);
1426 process_spill_slots (MonoCompile *cfg)
1428 MonoCompileGC *gcfg = cfg->gc_info;
1433 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1434 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1435 for (l = bb->spill_slot_defs; l; l = l->next) {
1436 MonoInst *def = l->data;
1437 int spill_slot = def->inst_c0;
1438 int bank = def->inst_c1;
1439 int offset = cfg->spill_info [bank][spill_slot].offset;
1440 int slot = fp_offset_to_slot (cfg, offset);
1442 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1443 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1447 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1448 for (l = bb->spill_slot_defs; l; l = l->next) {
1449 MonoInst *def = l->data;
1450 int spill_slot = def->inst_c0;
1451 int bank = def->inst_c1;
1452 int offset = cfg->spill_info [bank][spill_slot].offset;
1453 int slot = fp_offset_to_slot (cfg, offset);
1456 if (bank == MONO_REG_INT_MP)
1462 * Extend the live interval for the GC tracked spill slots
1463 * defined in this bblock.
1464 * FIXME: This is not needed.
1466 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1468 if (cfg->verbose_level > 1)
1469 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
1473 /* Set fp spill slots to NOREF */
1474 for (i = 0; i < cfg->spill_info_len [MONO_REG_DOUBLE]; ++i) {
1475 int offset = cfg->spill_info [MONO_REG_DOUBLE][i].offset;
1481 slot = fp_offset_to_slot (cfg, offset);
1483 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1485 if (cfg->verbose_level > 1)
1486 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1489 /* Set int spill slots to NOREF */
1490 for (i = 0; i < cfg->spill_info_len [MONO_REG_INT]; ++i) {
1491 int offset = cfg->spill_info [MONO_REG_INT][i].offset;
1497 slot = fp_offset_to_slot (cfg, offset);
1499 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1500 if (cfg->verbose_level > 1)
1501 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1506 * process_other_slots:
1508 * Process stack slots registered using mini_gc_set_slot_type_... ().
1511 process_other_slots (MonoCompile *cfg)
1513 MonoCompileGC *gcfg = cfg->gc_info;
1516 /* Relative to the CFA */
1517 for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
1518 guint data = GPOINTER_TO_UINT (l->data);
1519 int cfa_slot = data >> 16;
1520 GCSlotType type = data & 0xff;
1524 * Map the cfa relative slot to an fp relative slot.
1525 * slot_addr == cfa - <cfa_slot>*4/8
1526 * fp + cfa_offset == cfa
1527 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1529 slot = (cfg->cfa_offset / SIZEOF_SLOT) - cfa_slot - (gcfg->min_offset / SIZEOF_SLOT);
1531 set_slot_everywhere (gcfg, slot, type);
1533 if (cfg->verbose_level > 1) {
1534 int fp_offset = slot_to_fp_offset (cfg, slot);
1535 if (type == SLOT_NOREF)
1536 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));
1540 /* Relative to the FP */
1541 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1542 gint data = GPOINTER_TO_INT (l->data);
1543 int offset = data >> 16;
1544 GCSlotType type = data & 0xff;
1547 slot = fp_offset_to_slot (cfg, offset);
1549 set_slot_everywhere (gcfg, slot, type);
1551 /* Liveness for these slots is handled by process_spill_slots () */
1553 if (cfg->verbose_level > 1) {
1554 if (type == SLOT_REF)
1555 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
1556 else if (type == SLOT_NOREF)
1557 printf ("\tnoref slot at 0x%x(fp) (slot = %d)\n", offset, slot);
1563 get_vtype_bitmap (MonoType *t, int *numbits)
1565 MonoClass *klass = mono_class_from_mono_type (t);
1567 if (klass->generic_container || mono_class_is_open_constructed_type (t)) {
1568 /* FIXME: Generic sharing */
1571 mono_class_compute_gc_descriptor (klass);
1573 return mono_gc_get_bitmap_for_descr (klass->gc_descr, numbits);
1577 static inline const char*
1578 get_offset_sign (int offset)
1580 return offset < 0 ? "-" : "+";
1584 get_offset_val (int offset)
1586 return offset < 0 ? (- offset) : offset;
1590 process_variables (MonoCompile *cfg)
1592 MonoCompileGC *gcfg = cfg->gc_info;
1593 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1594 int i, locals_min_slot, locals_max_slot, cindex;
1598 int locals_min_offset = gcfg->locals_min_offset;
1599 int locals_max_offset = gcfg->locals_max_offset;
1601 /* Slots for locals are NOREF by default */
1602 locals_min_slot = (locals_min_offset - gcfg->min_offset) / SIZEOF_SLOT;
1603 locals_max_slot = (locals_max_offset - gcfg->min_offset) / SIZEOF_SLOT;
1604 for (i = locals_min_slot; i < locals_max_slot; ++i) {
1605 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1609 * Compute the offset where variables are initialized in the first bblock, if any.
1611 pc_offsets = g_new0 (int, cfg->next_vreg);
1613 bb = cfg->bb_entry->next_bb;
1614 MONO_BB_FOR_EACH_INS (bb, tmp) {
1615 if (tmp->opcode == OP_GC_LIVENESS_DEF) {
1616 int vreg = tmp->inst_c1;
1617 if (pc_offsets [vreg] == 0) {
1618 g_assert (tmp->backend.pc_offset > 0);
1619 pc_offsets [vreg] = tmp->backend.pc_offset;
1625 * Stack slots holding arguments are initialized in the prolog.
1626 * This means we can treat them alive for the whole method.
1628 for (i = 0; i < cfg->num_varinfo; i++) {
1629 MonoInst *ins = cfg->varinfo [i];
1630 MonoType *t = ins->inst_vtype;
1633 gboolean byref, is_this = FALSE;
1634 gboolean is_arg = i < cfg->locals_start;
1636 if (ins == cfg->ret) {
1637 if (!(ins->opcode == OP_REGOFFSET && MONO_TYPE_ISSTRUCT (t)))
1641 vmv = MONO_VARINFO (cfg, i);
1643 /* For some reason, 'this' is byref */
1644 if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
1645 t = &cfg->method->klass->byval_arg;
1651 if (ins->opcode == OP_REGVAR) {
1653 GCSlotType slot_type;
1655 t = mini_type_get_underlying_type (NULL, t);
1658 g_assert (hreg < MONO_MAX_IREGS);
1661 slot_type = SLOT_PIN;
1663 slot_type = mini_type_is_reference (cfg, t) ? SLOT_REF : SLOT_NOREF;
1665 if (slot_type == SLOT_PIN) {
1666 /* These have no live interval, be conservative */
1667 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1670 * Unlike variables allocated to the stack, we generate liveness info
1671 * for noref vars in registers in mono_spill_global_vars (), because
1672 * knowing that a register doesn't contain a ref allows us to mark its save
1673 * locations precisely.
1675 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1676 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1677 set_reg_slot (gcfg, hreg, cindex, slot_type);
1680 if (cfg->verbose_level > 1) {
1681 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
1687 if (ins->opcode != OP_REGOFFSET)
1690 if (ins->inst_offset % SIZEOF_SLOT != 0)
1693 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1695 if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
1696 /* These do not get stored in the prolog */
1697 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1699 if (cfg->verbose_level > 1) {
1700 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));
1705 if (MONO_TYPE_ISSTRUCT (t)) {
1707 gsize *bitmap = NULL;
1708 gboolean pin = FALSE;
1712 if (ins->backend.is_pinvoke)
1713 size = mono_class_native_size (ins->klass, NULL);
1715 size = mono_class_value_size (ins->klass, NULL);
1716 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
1718 if (cfg->verbose_level > 1)
1719 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));
1721 if (!ins->klass->has_references) {
1723 for (j = 0; j < size_in_slots; ++j)
1724 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1729 bitmap = get_vtype_bitmap (t, &numbits);
1734 * Most vtypes are marked volatile because of the LDADDR instructions,
1735 * and they have no liveness information since they are decomposed
1736 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1737 * them during VZERO decomposition.
1740 if (!pc_offsets [vmv->vreg])
1743 if (ins->backend.is_pinvoke)
1748 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1749 if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
1750 for (j = 0; j < numbits; ++j) {
1751 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1752 /* The descriptor is for the boxed object */
1753 set_slot (gcfg, (pos + j - (sizeof (MonoObject) / SIZEOF_SLOT)), cindex, pin ? SLOT_PIN : SLOT_REF);
1759 if (cfg->verbose_level > 1) {
1760 for (j = 0; j < numbits; ++j) {
1761 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD)))
1762 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)));
1766 if (cfg->verbose_level > 1)
1767 printf ("\t\tpinned\n");
1768 for (j = 0; j < size_in_slots; ++j) {
1769 set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
1778 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1779 /* Vret addr etc. */
1784 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1786 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1787 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1788 set_slot (gcfg, pos, cindex, SLOT_PIN);
1790 if (cfg->verbose_level > 1)
1791 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));
1796 * This is currently disabled, but could be enabled to debug crashes.
1799 if (t->type == MONO_TYPE_I) {
1801 * Variables created in mono_handle_global_vregs have type I, but they
1802 * could hold GC refs since the vregs they were created from might not been
1803 * marked as holding a GC ref. So be conservative.
1805 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1810 t = mini_type_get_underlying_type (NULL, t);
1812 if (!mini_type_is_reference (cfg, t)) {
1813 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1814 if (cfg->verbose_level > 1)
1815 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));
1816 if (!t->byref && sizeof (mgreg_t) == 4 && (t->type == MONO_TYPE_I8 || t->type == MONO_TYPE_U8 || t->type == MONO_TYPE_R8)) {
1817 set_slot_everywhere (gcfg, pos + 1, SLOT_NOREF);
1818 if (cfg->verbose_level > 1)
1819 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));
1824 /* 'this' is marked INDIRECT for gshared methods */
1825 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1827 * For volatile variables, treat them alive from the point they are
1828 * initialized in the first bblock until the end of the method.
1831 set_slot_everywhere (gcfg, pos, SLOT_REF);
1832 } else if (pc_offsets [vmv->vreg]) {
1833 set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
1834 set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
1836 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1838 if (cfg->verbose_level > 1)
1839 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));
1844 /* Live for the whole method */
1845 set_slot_everywhere (gcfg, pos, SLOT_REF);
1847 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1848 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1849 set_slot (gcfg, pos, cindex, SLOT_REF);
1852 if (cfg->verbose_level > 1) {
1853 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));
1857 g_free (pc_offsets);
1861 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1864 * Convert a sp relative offset to a slot index. This is
1865 * platform specific.
1868 /* fp = sp + offset */
1869 g_assert (cfg->frame_reg == AMD64_RBP);
1870 return (- cfg->arch.sp_fp_offset + sp_offset);
1871 #elif defined(TARGET_X86)
1872 /* The offset is computed from the sp at the start of the call sequence */
1873 g_assert (cfg->frame_reg == X86_EBP);
1874 #ifdef MONO_X86_NO_PUSHES
1875 return (- cfg->arch.sp_fp_offset + sp_offset);
1877 return (- cfg->arch.sp_fp_offset - sp_offset);
1886 process_param_area_slots (MonoCompile *cfg)
1888 MonoCompileGC *gcfg = cfg->gc_info;
1893 * These slots are used for passing parameters during calls. They are sp relative, not
1894 * fp relative, so they are harder to handle.
1896 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1897 /* The distance between fp and sp is not constant */
1900 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1902 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1903 GCCallSite *callsite = gcfg->callsites [cindex];
1906 for (l = callsite->param_slots; l; l = l->next) {
1907 MonoInst *def = l->data;
1908 MonoType *t = def->inst_vtype;
1909 int sp_offset = def->inst_offset;
1910 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1911 int slot = fp_offset_to_slot (cfg, fp_offset);
1915 if (MONO_TYPE_ISSTRUCT (t)) {
1916 size = mini_type_stack_size_full (cfg->generic_sharing_context, t, &align, FALSE);
1918 size = sizeof (mgreg_t);
1921 for (i = 0; i < size / sizeof (mgreg_t); ++i) {
1922 g_assert (slot + i >= 0 && slot + i < gcfg->nslots);
1923 is_param [slot + i] = TRUE;
1928 /* All param area slots are noref by default */
1929 for (i = 0; i < gcfg->nslots; ++i) {
1931 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1935 * We treat param area slots as being part of the callee's frame, to be able to handle tail calls which overwrite
1936 * the argument area of the caller.
1941 process_finally_clauses (MonoCompile *cfg)
1943 MonoCompileGC *gcfg = cfg->gc_info;
1944 GCCallSite **callsites;
1946 gboolean has_finally;
1947 int i, j, nslots, nregs;
1949 ncallsites = gcfg->ncallsites;
1950 nslots = gcfg->nslots;
1951 nregs = gcfg->nregs;
1952 callsites = gcfg->callsites;
1955 * The calls to the finally clauses don't show up in the cfg. See
1956 * test_0_liveness_8 ().
1957 * Variables accessed inside the finally clause are already marked VOLATILE by
1958 * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
1959 * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
1961 has_finally = FALSE;
1962 for (i = 0; i < cfg->header->num_clauses; ++i) {
1963 MonoExceptionClause *clause = &cfg->header->clauses [i];
1965 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1970 if (cfg->verbose_level > 1)
1971 printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
1972 for (j = 0; j < ncallsites; ++j) {
1973 MonoBasicBlock *bb = callsites [j]->bb;
1974 MonoExceptionClause *clause;
1975 gboolean is_in_finally = FALSE;
1977 for (i = 0; i < cfg->header->num_clauses; ++i) {
1978 clause = &cfg->header->clauses [i];
1980 if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
1981 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1982 is_in_finally = TRUE;
1988 if (is_in_finally) {
1989 for (i = 0; i < nslots; ++i)
1990 set_slot (gcfg, i, j, SLOT_PIN);
1991 for (i = 0; i < nregs; ++i)
1992 set_reg_slot (gcfg, i, j, SLOT_PIN);
1999 compute_frame_size (MonoCompile *cfg)
2001 int i, locals_min_offset, locals_max_offset, cfa_min_offset, cfa_max_offset;
2002 int min_offset, max_offset;
2003 MonoCompileGC *gcfg = cfg->gc_info;
2004 MonoMethodSignature *sig = mono_method_signature (cfg->method);
2007 /* Compute min/max offsets from the fp */
2010 #if defined(TARGET_AMD64) || defined(TARGET_X86) || defined(TARGET_ARM) || defined(TARGET_S390X)
2011 locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, SIZEOF_SLOT);
2012 locals_max_offset = cfg->locals_max_stack_offset;
2014 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
2018 locals_min_offset = ALIGN_TO (locals_min_offset, SIZEOF_SLOT);
2019 locals_max_offset = ALIGN_TO (locals_max_offset, SIZEOF_SLOT);
2021 min_offset = locals_min_offset;
2022 max_offset = locals_max_offset;
2025 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
2026 MonoInst *ins = cfg->args [i];
2028 if (ins->opcode == OP_REGOFFSET) {
2029 int size, size_in_slots;
2030 size = mini_type_stack_size_full (cfg->generic_sharing_context, ins->inst_vtype, NULL, ins->backend.is_pinvoke);
2031 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
2033 min_offset = MIN (min_offset, ins->inst_offset);
2034 max_offset = MAX ((int)max_offset, (int)(ins->inst_offset + (size_in_slots * SIZEOF_SLOT)));
2039 g_assert (cfg->frame_reg == cfg->cfa_reg);
2040 g_assert (cfg->cfa_offset > 0);
2042 cfa_max_offset = cfg->cfa_offset;
2044 min_offset = MIN (min_offset, cfa_min_offset);
2045 max_offset = MAX (max_offset, cfa_max_offset);
2047 /* Fp relative slots */
2048 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
2049 gint data = GPOINTER_TO_INT (l->data);
2050 int offset = data >> 16;
2052 min_offset = MIN (min_offset, offset);
2056 if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
2057 int stack_offset = ALIGN_TO (cfg->stack_offset, SIZEOF_SLOT);
2058 min_offset = MIN (min_offset, (-stack_offset));
2061 /* Param area slots */
2063 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
2064 #elif defined(TARGET_X86)
2065 #ifdef MONO_X86_NO_PUSHES
2066 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
2068 min_offset = MIN (min_offset, - (cfg->arch.sp_fp_offset + cfg->arch.param_area_size));
2070 #elif defined(TARGET_ARM)
2072 #elif defined(TARGET_s390X)
2078 gcfg->min_offset = min_offset;
2079 gcfg->max_offset = max_offset;
2080 gcfg->locals_min_offset = locals_min_offset;
2081 gcfg->locals_max_offset = locals_max_offset;
2085 init_gcfg (MonoCompile *cfg)
2087 int i, nregs, nslots;
2088 MonoCompileGC *gcfg = cfg->gc_info;
2089 GCCallSite **callsites;
2098 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2099 ncallsites += g_slist_length (bb->gc_callsites);
2101 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
2103 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2104 for (l = bb->gc_callsites; l; l = l->next)
2105 callsites [i++] = l->data;
2108 /* The callsites should already be ordered by pc offset */
2109 for (i = 1; i < ncallsites; ++i)
2110 g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
2113 * The stack frame looks like this:
2115 * <fp + max_offset> == cfa -> <end of previous frame>
2116 * <other stack slots>
2118 * <other stack slots>
2119 * fp + min_offset ->
2124 if (cfg->verbose_level > 1)
2125 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
2127 nslots = (gcfg->max_offset - gcfg->min_offset) / SIZEOF_SLOT;
2130 gcfg->nslots = nslots;
2131 gcfg->nregs = nregs;
2132 gcfg->callsites = callsites;
2133 gcfg->ncallsites = ncallsites;
2134 gcfg->stack_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
2135 gcfg->reg_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
2136 gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
2137 gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
2138 gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
2139 gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
2141 /* All slots start out as PIN */
2142 memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * nregs);
2143 for (i = 0; i < nregs; ++i) {
2145 * By default, registers are NOREF.
2146 * It is possible for a callee to save them before being defined in this method,
2147 * but the saved value is dead too, so it doesn't need to be marked.
2149 if ((cfg->used_int_regs & (1 << i)))
2150 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
2154 static inline gboolean
2155 has_bit_set (guint8 *bitmap, int width, int slot)
2158 int pos = width * slot;
2160 for (i = 0; i < width; ++i) {
2161 if (bitmap [pos + i])
2168 create_map (MonoCompile *cfg)
2171 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
2173 int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
2174 int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
2176 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
2177 MonoCompileGC *gcfg = cfg->gc_info;
2178 GCCallSite **callsites;
2180 guint8 *bitmap, *bitmaps;
2181 guint32 reg_ref_mask, reg_pin_mask;
2183 ncallsites = gcfg->ncallsites;
2184 nslots = gcfg->nslots;
2185 nregs = gcfg->nregs;
2186 callsites = gcfg->callsites;
2189 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
2190 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
2192 has_ref_slots = FALSE;
2193 has_pin_slots = FALSE;
2196 memset (ntypes, 0, sizeof (ntypes));
2197 for (i = 0; i < nslots; ++i) {
2198 gboolean has_ref = FALSE;
2199 gboolean has_pin = FALSE;
2201 if (has_bit_set (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i))
2203 if (has_bit_set (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i))
2207 has_ref_slots = TRUE;
2209 has_pin_slots = TRUE;
2212 ntypes [SLOT_REF] ++;
2214 ntypes [SLOT_PIN] ++;
2216 ntypes [SLOT_NOREF] ++;
2218 if (has_ref || has_pin) {
2225 start = end = nslots;
2227 g_assert (start != -1);
2228 g_assert (start < end);
2231 has_ref_regs = FALSE;
2232 has_pin_regs = FALSE;
2237 for (i = 0; i < nregs; ++i) {
2238 gboolean has_ref = FALSE;
2239 gboolean has_pin = FALSE;
2241 if (!(cfg->used_int_regs & (1 << i)))
2244 if (has_bit_set (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i))
2246 if (has_bit_set (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i))
2250 reg_ref_mask |= (1 << i);
2251 has_ref_regs = TRUE;
2255 reg_pin_mask |= (1 << i);
2256 has_pin_regs = TRUE;
2261 if (cfg->verbose_level > 1)
2262 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);
2264 /* Create the GC Map */
2266 /* The work bitmaps have one row for each slot, since this is how we access them during construction */
2267 stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
2268 stack_bitmap_size = stack_bitmap_width * ncallsites;
2269 reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
2270 reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
2271 reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
2272 reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
2273 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);
2275 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2277 map->frame_reg = cfg->frame_reg;
2278 map->start_offset = gcfg->min_offset;
2279 map->end_offset = gcfg->min_offset + (nslots * SIZEOF_SLOT);
2280 map->map_offset = start * SIZEOF_SLOT;
2281 map->nslots = end - start;
2282 map->has_ref_slots = has_ref_slots;
2283 map->has_pin_slots = has_pin_slots;
2284 map->has_ref_regs = has_ref_regs;
2285 map->has_pin_regs = has_pin_regs;
2286 g_assert (nregs < 32);
2287 map->used_int_regs = cfg->used_int_regs;
2288 map->reg_ref_mask = reg_ref_mask;
2289 map->reg_pin_mask = reg_pin_mask;
2290 map->nref_regs = nref_regs;
2291 map->npin_regs = npin_regs;
2293 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2296 if (has_ref_slots) {
2297 map->stack_ref_bitmap_offset = bitmaps_offset;
2298 bitmaps_offset += stack_bitmap_size;
2300 bitmap = &bitmaps [map->stack_ref_bitmap_offset];
2301 for (i = 0; i < nslots; ++i) {
2302 for (j = 0; j < ncallsites; ++j) {
2303 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i, j))
2304 set_bit (bitmap, stack_bitmap_width, j, i - start);
2308 if (has_pin_slots) {
2309 map->stack_pin_bitmap_offset = bitmaps_offset;
2310 bitmaps_offset += stack_bitmap_size;
2312 bitmap = &bitmaps [map->stack_pin_bitmap_offset];
2313 for (i = 0; i < nslots; ++i) {
2314 for (j = 0; j < ncallsites; ++j) {
2315 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i, j))
2316 set_bit (bitmap, stack_bitmap_width, j, i - start);
2321 map->reg_ref_bitmap_offset = bitmaps_offset;
2322 bitmaps_offset += reg_ref_bitmap_size;
2324 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2326 for (i = 0; i < nregs; ++i) {
2327 if (reg_ref_mask & (1 << i)) {
2328 for (j = 0; j < ncallsites; ++j) {
2329 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i, j))
2330 set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
2337 map->reg_pin_bitmap_offset = bitmaps_offset;
2338 bitmaps_offset += reg_pin_bitmap_size;
2340 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2342 for (i = 0; i < nregs; ++i) {
2343 if (reg_pin_mask & (1 << i)) {
2344 for (j = 0; j < ncallsites; ++j) {
2345 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i, j))
2346 set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
2354 map->ncallsites = ncallsites;
2355 if (cfg->code_len < 256)
2356 map->callsite_entry_size = 1;
2357 else if (cfg->code_len < 65536)
2358 map->callsite_entry_size = 2;
2360 map->callsite_entry_size = 4;
2362 /* Encode the GC Map */
2370 encode_gc_map (map, buf, &endbuf);
2371 g_assert (endbuf - buf < 256);
2373 encoded_size = endbuf - buf;
2374 alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
2376 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2377 //emap->ref_slots = map->ref_slots;
2379 /* Encoded fixed fields */
2380 p = &emap->encoded [0];
2381 //emap->encoded_size = encoded_size;
2382 memcpy (p, buf, encoded_size);
2385 /* Callsite table */
2386 p = (guint8*)ALIGN_TO ((mgreg_t)p, map->callsite_entry_size);
2387 if (map->callsite_entry_size == 1) {
2388 guint8 *offsets = p;
2389 for (i = 0; i < ncallsites; ++i)
2390 offsets [i] = callsites [i]->pc_offset;
2391 stats.gc_callsites8_size += ncallsites * sizeof (guint8);
2392 } else if (map->callsite_entry_size == 2) {
2393 guint16 *offsets = (guint16*)p;
2394 for (i = 0; i < ncallsites; ++i)
2395 offsets [i] = callsites [i]->pc_offset;
2396 stats.gc_callsites16_size += ncallsites * sizeof (guint16);
2398 guint32 *offsets = (guint32*)p;
2399 for (i = 0; i < ncallsites; ++i)
2400 offsets [i] = callsites [i]->pc_offset;
2401 stats.gc_callsites32_size += ncallsites * sizeof (guint32);
2403 p += ncallsites * map->callsite_entry_size;
2406 memcpy (p, bitmaps, bitmaps_size);
2409 g_assert ((guint8*)p - (guint8*)emap <= alloc_size);
2411 stats.gc_maps_size += alloc_size;
2412 stats.gc_callsites_size += ncallsites * map->callsite_entry_size;
2413 stats.gc_bitmaps_size += bitmaps_size;
2414 stats.gc_map_struct_size += sizeof (GCEncodedMap) + encoded_size;
2416 cfg->jit_info->gc_info = emap;
2418 cfg->gc_map = (guint8*)emap;
2419 cfg->gc_map_size = alloc_size;
2422 stats.all_slots += nslots;
2423 stats.ref_slots += ntypes [SLOT_REF];
2424 stats.noref_slots += ntypes [SLOT_NOREF];
2425 stats.pin_slots += ntypes [SLOT_PIN];
2429 mini_gc_create_gc_map (MonoCompile *cfg)
2431 if (!cfg->compute_gc_maps)
2435 * During marking, all frames except the top frame are at a call site, and we mark the
2436 * top frame conservatively. This means that we only need to compute and record
2437 * GC maps for call sites.
2440 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2441 /* Without liveness info, the live ranges are not precise enough */
2444 mono_analyze_liveness_gc (cfg);
2446 compute_frame_size (cfg);
2450 process_spill_slots (cfg);
2451 process_other_slots (cfg);
2452 process_param_area_slots (cfg);
2453 process_variables (cfg);
2454 process_finally_clauses (cfg);
2459 #endif /* DISABLE_JIT */
2462 parse_debug_options (void)
2467 env = g_getenv ("MONO_GCMAP_DEBUG");
2471 opts = g_strsplit (env, ",", -1);
2472 for (ptr = opts; ptr && *ptr; ptr ++) {
2473 /* No options yet */
2474 fprintf (stderr, "Invalid format for the MONO_GCMAP_DEBUG env variable: '%s'\n", env);
2485 memset (&cb, 0, sizeof (cb));
2486 cb.thread_attach_func = thread_attach_func;
2487 cb.thread_detach_func = thread_detach_func;
2488 cb.thread_suspend_func = thread_suspend_func;
2489 /* Comment this out to disable precise stack marking */
2490 cb.thread_mark_func = thread_mark_func;
2491 mono_gc_set_gc_callbacks (&cb);
2493 logfile = mono_gc_get_logfile ();
2495 parse_debug_options ();
2497 mono_counters_register ("GC Maps size",
2498 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
2499 mono_counters_register ("GC Call Sites size",
2500 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
2501 mono_counters_register ("GC Bitmaps size",
2502 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
2503 mono_counters_register ("GC Map struct size",
2504 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
2505 mono_counters_register ("GC Call Sites encoded using 8 bits",
2506 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
2507 mono_counters_register ("GC Call Sites encoded using 16 bits",
2508 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
2509 mono_counters_register ("GC Call Sites encoded using 32 bits",
2510 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
2512 mono_counters_register ("GC Map slots (all)",
2513 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
2514 mono_counters_register ("GC Map slots (ref)",
2515 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.ref_slots);
2516 mono_counters_register ("GC Map slots (noref)",
2517 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.noref_slots);
2518 mono_counters_register ("GC Map slots (pin)",
2519 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
2521 mono_counters_register ("GC TLS Data size",
2522 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2524 mono_counters_register ("Stack space scanned (all)",
2525 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
2526 mono_counters_register ("Stack space scanned (native)",
2527 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_native);
2528 mono_counters_register ("Stack space scanned (other)",
2529 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_other);
2530 mono_counters_register ("Stack space scanned (using GC Maps)",
2531 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
2532 mono_counters_register ("Stack space scanned (precise)",
2533 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_precisely);
2534 mono_counters_register ("Stack space scanned (pin)",
2535 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_conservatively);
2536 mono_counters_register ("Stack space scanned (pin registers)",
2537 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_registers);
2543 mini_gc_enable_gc_maps_for_aot (void)
2555 mini_gc_init_gc_map (MonoCompile *cfg)
2560 mini_gc_create_gc_map (MonoCompile *cfg)
2565 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2570 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2574 #endif /* DISABLE_JIT */
2583 * Set GC specific options in CFG.
2586 mini_gc_init_cfg (MonoCompile *cfg)
2588 if (mono_gc_is_moving ()) {
2589 cfg->disable_ref_noref_stack_slot_share = TRUE;
2590 cfg->gen_write_barriers = TRUE;
2593 mini_gc_init_gc_map (cfg);
2596 #endif /* DISABLE_JIT */
2599 * Problems with the current code:
2600 * - the stack walk is slow
2601 * - vtypes/refs used in EH regions are treated conservatively
2602 * - if the code is finished, less pinning will be done, causing problems because
2603 * we promote all surviving objects to old-gen.
2604 * - the unwind code can't handle a method stopped inside a finally region, it thinks the caller is
2605 * another method, but in reality it is either the exception handling code or the CALL_HANDLER opcode.
2606 * This manifests in "Unable to find ip offset x in callsite list" assertions.
2607 * - the unwind code also can't handle frames which are in the epilog, since the unwind info is not
2612 * Ideas for creating smaller GC maps:
2613 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2615 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2617 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2618 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2619 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2620 * from an index entry.