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 };
439 #elif defined(TARGET_POWERPC64) && _CALL_ELF == 2
440 static int callee_saved_regs [] = {
441 ppc_r13, ppc_r14, ppc_r15, ppc_r16,
442 ppc_r17, ppc_r18, ppc_r19, ppc_r20,
443 ppc_r21, ppc_r22, ppc_r23, ppc_r24,
444 ppc_r25, ppc_r26, ppc_r27, ppc_r28,
445 ppc_r29, ppc_r30, ppc_r31 };
446 #elif defined(TARGET_POWERPC)
447 static int callee_saved_regs [] = { ppc_r6, ppc_r7, ppc_r8, ppc_r9, ppc_r10, ppc_r11, ppc_r12, ppc_r13, ppc_r14 };
451 encode_regmask (guint32 regmask)
457 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
458 if (regmask & (1 << callee_saved_regs [i])) {
460 regmask -= (1 << callee_saved_regs [i]);
463 g_assert (regmask == 0);
468 decode_regmask (guint32 regmask)
474 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i)
475 if (regmask & (1 << i))
476 res |= (1 << callee_saved_regs [i]);
483 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
486 encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
490 encode_sleb128 (map->start_offset / SIZEOF_SLOT, buf, &buf);
491 encode_sleb128 (map->end_offset / SIZEOF_SLOT, buf, &buf);
492 encode_sleb128 (map->map_offset / SIZEOF_SLOT, buf, &buf);
493 encode_uleb128 (map->nslots, buf, &buf);
494 g_assert (map->callsite_entry_size <= 4);
495 freg = encode_frame_reg (map->frame_reg);
497 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);
498 encode_uleb128 (flags, buf, &buf);
499 encode_uleb128 (encode_regmask (map->used_int_regs), buf, &buf);
500 if (map->has_ref_regs)
501 encode_uleb128 (encode_regmask (map->reg_ref_mask), buf, &buf);
502 if (map->has_pin_regs)
503 encode_uleb128 (encode_regmask (map->reg_pin_mask), buf, &buf);
504 encode_uleb128 (map->ncallsites, buf, &buf);
512 * Decode the encoded GC map representation in BUF and store the result into MAP.
515 decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
518 int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
521 map->start_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
522 map->end_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
523 map->map_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
524 map->nslots = decode_uleb128 (buf, &buf);
525 flags = decode_uleb128 (buf, &buf);
526 map->has_ref_slots = (flags & 1) ? 1 : 0;
527 map->has_pin_slots = (flags & 2) ? 1 : 0;
528 map->has_ref_regs = (flags & 4) ? 1 : 0;
529 map->has_pin_regs = (flags & 8) ? 1 : 0;
530 map->callsite_entry_size = ((flags >> 4) & 0x3) + 1;
532 map->frame_reg = decode_frame_reg (freg);
533 map->used_int_regs = decode_regmask (decode_uleb128 (buf, &buf));
534 if (map->has_ref_regs) {
535 map->reg_ref_mask = decode_regmask (decode_uleb128 (buf, &buf));
537 for (i = 0; i < NREGS; ++i)
538 if (map->reg_ref_mask & (1 << i))
542 if (map->has_pin_regs) {
543 map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
545 for (i = 0; i < NREGS; ++i)
546 if (map->reg_pin_mask & (1 << i))
550 map->ncallsites = decode_uleb128 (buf, &buf);
552 stack_bitmap_size = (ALIGN_TO (map->nslots, 8) / 8) * map->ncallsites;
553 reg_ref_bitmap_size = (ALIGN_TO (map->nref_regs, 8) / 8) * map->ncallsites;
554 reg_pin_bitmap_size = (ALIGN_TO (map->npin_regs, 8) / 8) * map->ncallsites;
556 map->stack_ref_bitmap_offset = offset;
557 if (map->has_ref_slots)
558 offset += stack_bitmap_size;
559 map->stack_pin_bitmap_offset = offset;
560 if (map->has_pin_slots)
561 offset += stack_bitmap_size;
562 map->reg_ref_bitmap_offset = offset;
563 if (map->has_ref_regs)
564 offset += reg_ref_bitmap_size;
565 map->reg_pin_bitmap_offset = offset;
566 if (map->has_pin_regs)
567 offset += reg_pin_bitmap_size;
573 thread_attach_func (void)
577 tls = g_new0 (TlsData, 1);
578 tls->tid = GetCurrentThreadId ();
579 tls->info = mono_thread_info_current ();
580 stats.tlsdata_size += sizeof (TlsData);
586 thread_detach_func (gpointer user_data)
588 TlsData *tls = user_data;
594 thread_suspend_func (gpointer user_data, void *sigctx, MonoContext *ctx)
596 TlsData *tls = user_data;
599 /* Happens during startup */
603 if (tls->tid != GetCurrentThreadId ()) {
604 /* Happens on osx because threads are not suspended using signals */
609 g_assert (tls->info);
613 res = mono_thread_state_init_from_handle (&tls->unwind_state, tls->info);
616 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF] = mono_get_lmf ();
618 #ifdef MONO_ARCH_HAVE_SIGCTX_TO_MONOCTX
619 mono_sigctx_to_monoctx (sigctx, &tls->unwind_state.ctx);
620 tls->unwind_state.valid = TRUE;
622 tls->unwind_state.valid = FALSE;
625 memcpy (&tls->unwind_state.ctx, ctx, sizeof (MonoContext));
626 tls->unwind_state.valid = TRUE;
628 tls->unwind_state.valid = FALSE;
630 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_JIT_TLS] = mono_native_tls_get_value (mono_jit_tls_id);
631 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN] = mono_domain_get ();
634 if (!tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN]) {
635 /* Happens during startup */
636 tls->unwind_state.valid = FALSE;
641 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
644 set_bit (guint8 *bitmap, int width, int y, int x)
646 bitmap [(width * y) + (x / 8)] |= (1 << (x % 8));
650 clear_bit (guint8 *bitmap, int width, int y, int x)
652 bitmap [(width * y) + (x / 8)] &= ~(1 << (x % 8));
656 get_bit (guint8 *bitmap, int width, int y, int x)
658 return bitmap [(width * y) + (x / 8)] & (1 << (x % 8));
662 slot_type_to_string (GCSlotType type)
672 g_assert_not_reached ();
677 static inline mgreg_t
678 get_frame_pointer (MonoContext *ctx, int frame_reg)
680 #if defined(TARGET_AMD64)
681 if (frame_reg == AMD64_RSP)
683 else if (frame_reg == AMD64_RBP)
685 #elif defined(TARGET_X86)
686 if (frame_reg == X86_ESP)
688 else if (frame_reg == X86_EBP)
690 #elif defined(TARGET_ARM)
691 if (frame_reg == ARMREG_SP)
692 return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
693 else if (frame_reg == ARMREG_FP)
694 return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
695 #elif defined(TARGET_S390X)
696 if (frame_reg == S390_SP)
697 return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
698 else if (frame_reg == S390_FP)
699 return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
701 g_assert_not_reached ();
706 * conservatively_pass:
708 * Mark a thread stack conservatively and collect information needed by the precise pass.
711 conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
715 MonoContext ctx, new_ctx;
718 gboolean last = TRUE;
722 guint8* fp, *p, *real_frame_start, *frame_start, *frame_end;
723 int i, pc_offset, cindex, bitmap_width;
724 int scanned = 0, scanned_precisely, scanned_conservatively, scanned_registers;
726 StackFrameInfo frame;
727 mgreg_t *reg_locations [MONO_MAX_IREGS];
728 mgreg_t *new_reg_locations [MONO_MAX_IREGS];
731 guint32 precise_regmask;
735 tls->ref_to_track = NULL;
738 /* tls == NULL can happen during startup */
739 if (mono_thread_internal_current () == NULL || !tls) {
740 mono_gc_conservatively_scan_area (stack_start, stack_end);
741 stats.scanned_stacks += stack_end - stack_start;
745 lmf = tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF];
748 /* Number of bytes scanned based on GC map data */
750 /* Number of bytes scanned precisely based on GC map data */
751 scanned_precisely = 0;
752 /* Number of bytes scanned conservatively based on GC map data */
753 scanned_conservatively = 0;
754 /* Number of bytes scanned conservatively in register save areas */
755 scanned_registers = 0;
757 /* This is one past the last address which we have scanned */
758 stack_limit = stack_start;
760 if (!tls->unwind_state.valid)
761 memset (&new_ctx, 0, sizeof (ctx));
763 memcpy (&new_ctx, &tls->unwind_state.ctx, sizeof (MonoContext));
765 memset (reg_locations, 0, sizeof (reg_locations));
766 memset (new_reg_locations, 0, sizeof (new_reg_locations));
769 if (!tls->unwind_state.valid)
772 memcpy (&ctx, &new_ctx, sizeof (ctx));
774 for (i = 0; i < MONO_MAX_IREGS; ++i) {
775 if (new_reg_locations [i]) {
777 * If the current frame saves the register, it means it might modify its
778 * value, thus the old location might not contain the same value, so
779 * we have to mark it conservatively.
781 if (reg_locations [i]) {
782 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
783 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
784 scanned_registers += SIZEOF_SLOT;
787 reg_locations [i] = new_reg_locations [i];
789 DEBUG (fprintf (logfile, "\treg %s is now at location %p.\n", mono_arch_regname (i), reg_locations [i]));
793 g_assert ((mgreg_t)stack_limit % SIZEOF_SLOT == 0);
795 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);
801 // FIXME: For skipped frames, scan the param area of the parent frame conservatively ?
803 if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
805 * These frames are problematic for several reasons:
806 * - they are unwound through an LMF, and we have no precise register tracking for those.
807 * - the LMF might not contain a precise ip, so we can't compute the call site.
808 * - the LMF only unwinds to the wrapper frame, so we get these methods twice.
810 DEBUG (fprintf (logfile, "Mark(0): <Managed-to-native transition>\n"));
811 for (i = 0; i < MONO_MAX_IREGS; ++i) {
812 if (reg_locations [i]) {
813 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
814 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
815 scanned_registers += SIZEOF_SLOT;
817 reg_locations [i] = NULL;
818 new_reg_locations [i] = NULL;
825 method = jinfo_get_method (ji);
829 /* The last frame can be in any state so mark conservatively */
832 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));
834 DEBUG (fprintf (logfile, "\t <Last frame>\n"));
837 * new_reg_locations is not precise when a method is interrupted during its epilog, so clear it.
839 for (i = 0; i < MONO_MAX_IREGS; ++i) {
840 if (reg_locations [i]) {
841 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
842 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
843 scanned_registers += SIZEOF_SLOT;
845 if (new_reg_locations [i]) {
846 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), new_reg_locations [i]));
847 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
848 scanned_registers += SIZEOF_SLOT;
850 reg_locations [i] = NULL;
851 new_reg_locations [i] = NULL;
856 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
858 /* These frames are very problematic */
859 if (method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
860 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));
861 DEBUG (fprintf (logfile, "\tSkip.\n"));
865 /* All the other frames are at a call site */
867 if (tls->nframes == MAX_FRAMES) {
869 * Can't save information since the array is full. So scan the rest of the
870 * stack conservatively.
872 DEBUG (fprintf (logfile, "Mark (0): Frame stack full.\n"));
876 /* Scan the frame of this method */
879 * A frame contains the following:
884 * - localloc-ed memory
886 g_assert (pc_offset >= 0);
891 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));
892 DEBUG (fprintf (logfile, "\tNo GC Map.\n"));
896 /* The embedded callsite table requires this */
897 g_assert (((mgreg_t)emap % 4) == 0);
900 * Debugging aid to control the number of frames scanned precisely
902 if (!precise_frame_limit_inited) {
903 if (g_getenv ("MONO_PRECISE_COUNT"))
904 precise_frame_limit = atoi (g_getenv ("MONO_PRECISE_COUNT"));
905 precise_frame_limit_inited = TRUE;
908 if (precise_frame_limit != -1) {
909 if (precise_frame_count [FALSE] == precise_frame_limit)
910 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (method, TRUE));
911 if (precise_frame_count [FALSE] > precise_frame_limit)
914 precise_frame_count [FALSE] ++;
916 /* Decode the encoded GC map */
918 memset (map, 0, sizeof (GCMap));
919 decode_gc_map (&emap->encoded [0], map, &p);
920 p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
921 map->callsites.offsets8 = p;
922 p += map->callsite_entry_size * map->ncallsites;
925 fp = (guint8*)get_frame_pointer (&ctx, map->frame_reg);
927 real_frame_start = fp + map->start_offset;
928 frame_start = fp + map->start_offset + map->map_offset;
929 frame_end = fp + map->end_offset;
931 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));
933 /* Find the callsite index */
934 if (map->callsite_entry_size == 1) {
935 for (i = 0; i < map->ncallsites; ++i)
936 /* ip points inside the call instruction */
937 if (map->callsites.offsets8 [i] == pc_offset + 1)
939 } else if (map->callsite_entry_size == 2) {
940 // FIXME: Use a binary search
941 for (i = 0; i < map->ncallsites; ++i)
942 /* ip points inside the call instruction */
943 if (map->callsites.offsets16 [i] == pc_offset + 1)
946 // FIXME: Use a binary search
947 for (i = 0; i < map->ncallsites; ++i)
948 /* ip points inside the call instruction */
949 if (map->callsites.offsets32 [i] == pc_offset + 1)
952 if (i == map->ncallsites) {
953 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (method, TRUE));
954 g_assert_not_reached ();
959 * This is not neccessary true on x86 because frames have a different size at each
962 //g_assert (real_frame_start >= stack_limit);
964 if (real_frame_start > stack_limit) {
965 /* This scans the previously skipped frames as well */
966 DEBUG (fprintf (logfile, "\tscan area %p-%p (%d).\n", stack_limit, real_frame_start, (int)(real_frame_start - stack_limit)));
967 mono_gc_conservatively_scan_area (stack_limit, real_frame_start);
968 stats.scanned_other += real_frame_start - stack_limit;
971 /* Mark stack slots */
972 if (map->has_pin_slots) {
973 int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
974 guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
979 for (i = 0; i < map->nslots; ++i) {
980 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
982 DEBUG (fprintf (logfile, "\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
983 mono_gc_conservatively_scan_area (p, p + SIZEOF_SLOT);
984 scanned_conservatively += SIZEOF_SLOT;
986 scanned_precisely += SIZEOF_SLOT;
991 scanned_precisely += (map->nslots * SIZEOF_SLOT);
994 /* The area outside of start-end is NOREF */
995 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * SIZEOF_SLOT);
998 precise_regmask = map->used_int_regs | (1 << map->frame_reg);
999 if (map->has_pin_regs) {
1000 int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
1001 guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
1003 for (i = 0; i < NREGS; ++i) {
1004 if (!(map->used_int_regs & (1 << i)))
1007 if (!(map->reg_pin_mask & (1 << i)))
1010 if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
1011 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
1012 precise_regmask &= ~(1 << i);
1018 scanned += map->end_offset - map->start_offset;
1020 g_assert (scanned == scanned_precisely + scanned_conservatively);
1022 stack_limit = frame_end;
1024 /* Save information for the precise pass */
1025 fi = &tls->frames [tls->nframes];
1026 fi->nslots = map->nslots;
1027 bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
1028 if (map->has_ref_slots)
1029 fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
1032 fi->frame_start_offset = frame_start - stack_start;
1033 fi->nreg_locations = 0;
1034 DEBUG_PRECISE (fi->ji = ji);
1035 DEBUG_PRECISE (fi->fp = fp);
1037 if (map->has_ref_regs) {
1038 int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
1039 guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
1041 for (i = 0; i < NREGS; ++i) {
1042 if (!(map->reg_ref_mask & (1 << i)))
1045 if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
1046 DEBUG_PRECISE (fi->regs [fi->nreg_locations] = i);
1047 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
1048 fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
1049 fi->nreg_locations ++;
1056 * Clear locations of precisely tracked registers.
1058 if (precise_regmask) {
1059 for (i = 0; i < NREGS; ++i) {
1060 if (precise_regmask & (1 << i)) {
1062 * The method uses this register, and we have precise info for it.
1063 * This means the location will be scanned precisely.
1064 * Tell the code at the beginning of the loop that this location is
1067 if (reg_locations [i])
1068 DEBUG (fprintf (logfile, "\treg %s at location %p (==%p) is precise.\n", mono_arch_regname (i), reg_locations [i], (gpointer)*reg_locations [i]));
1069 reg_locations [i] = NULL;
1077 /* Scan the remaining register save locations */
1078 for (i = 0; i < MONO_MAX_IREGS; ++i) {
1079 if (reg_locations [i]) {
1080 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", reg_locations [i]));
1081 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
1082 scanned_registers += SIZEOF_SLOT;
1084 if (new_reg_locations [i]) {
1085 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", new_reg_locations [i]));
1086 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
1087 scanned_registers += SIZEOF_SLOT;
1091 if (stack_limit < stack_end) {
1092 DEBUG (fprintf (logfile, "\tscan remaining stack %p-%p (%d).\n", stack_limit, stack_end, (int)(stack_end - stack_limit)));
1093 mono_gc_conservatively_scan_area (stack_limit, stack_end);
1094 stats.scanned_native += stack_end - stack_limit;
1097 DEBUG (fprintf (logfile, "Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
1099 stats.scanned_stacks += stack_end - stack_start;
1100 stats.scanned += scanned;
1101 stats.scanned_precisely += scanned_precisely;
1102 stats.scanned_conservatively += scanned_conservatively;
1103 stats.scanned_registers += scanned_registers;
1105 //mono_gc_conservatively_scan_area (stack_start, stack_end);
1111 * Mark a thread stack precisely based on information saved during the conservative
1115 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end, void *gc_data)
1119 guint8 *frame_start;
1124 if (!tls->unwind_state.valid)
1127 for (findex = 0; findex < tls->nframes; findex ++) {
1128 /* Load information saved by the !precise pass */
1129 fi = &tls->frames [findex];
1130 frame_start = stack_start + fi->frame_start_offset;
1132 DEBUG (char *fname = mono_method_full_name (jinfo_get_method (fi->ji), TRUE); fprintf (logfile, "Mark(1): %s\n", fname); g_free (fname));
1135 * FIXME: Add a function to mark using a bitmap, to avoid doing a
1136 * call for each object.
1139 /* Mark stack slots */
1141 guint8 *ref_bitmap = fi->bitmap;
1144 for (i = 0; i < fi->nslots; ++i) {
1145 MonoObject **ptr = (MonoObject**)(frame_start + (i * SIZEOF_SLOT));
1147 live = ref_bitmap [i / 8] & (1 << (i % 8));
1150 MonoObject *obj = *ptr;
1152 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1153 *ptr = mono_gc_scan_object (obj, gc_data);
1154 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1156 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1161 * This is disabled because the pointer takes up a lot of space.
1162 * Stack slots might be shared between ref and non-ref variables ?
1164 if (map->ref_slots [i / 8] & (1 << (i % 8))) {
1165 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));
1167 * Fail fast if the live range is incorrect, and
1168 * the JITted code tries to access this object
1177 /* Mark registers */
1180 * Registers are different from stack slots, they have no address where they
1181 * are stored. Instead, some frame below this frame in the stack saves them
1182 * in its prolog to the stack. We can mark this location precisely.
1184 for (i = 0; i < fi->nreg_locations; ++i) {
1186 * reg_locations [i] contains the address of the stack slot where
1187 * a reg was last saved, so mark that slot.
1189 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1190 MonoObject *obj = *ptr;
1193 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
1194 *ptr = mono_gc_scan_object (obj, gc_data);
1195 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1197 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p\n", mono_arch_regname (fi->regs [i]), ptr, obj));
1203 * Debugging aid to check for missed refs.
1205 if (tls->ref_to_track) {
1208 for (p = (mgreg_t*)stack_start; p < (mgreg_t*)stack_end; ++p)
1209 if (*p == (mgreg_t)tls->ref_to_track)
1210 printf ("REF AT %p.\n", p);
1217 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1218 * call, and TRUE at the second. USER_DATA points to a TlsData
1219 * structure filled up by thread_suspend_func.
1222 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise, void *gc_data)
1224 TlsData *tls = user_data;
1226 DEBUG (fprintf (logfile, "****************************************\n"));
1227 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));
1228 DEBUG (fprintf (logfile, "****************************************\n"));
1231 conservative_pass (tls, stack_start, stack_end);
1233 precise_pass (tls, stack_start, stack_end, gc_data);
1239 mini_gc_init_gc_map (MonoCompile *cfg)
1241 if (COMPILE_LLVM (cfg))
1244 if (!mono_gc_is_moving ())
1247 if (cfg->compile_aot) {
1248 if (!enable_gc_maps_for_aot)
1250 } else if (!mono_gc_precise_stack_mark_enabled ())
1254 /* Debugging support */
1256 static int precise_count;
1259 if (g_getenv ("MONO_GCMAP_COUNT")) {
1260 if (precise_count == atoi (g_getenv ("MONO_GCMAP_COUNT")))
1261 printf ("LAST: %s\n", mono_method_full_name (cfg->method, TRUE));
1262 if (precise_count > atoi (g_getenv ("MONO_GCMAP_COUNT")))
1268 cfg->compute_gc_maps = TRUE;
1270 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1274 * mini_gc_set_slot_type_from_fp:
1276 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1277 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1278 * need to call this function for those slots.
1281 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1283 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1285 if (!cfg->compute_gc_maps)
1288 g_assert (slot_offset % SIZEOF_SLOT == 0);
1290 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1294 * mini_gc_set_slot_type_from_cfa:
1296 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1297 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1298 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1299 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1300 * call this function for those slots.
1303 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1305 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1306 int slot = - (slot_offset / SIZEOF_SLOT);
1308 if (!cfg->compute_gc_maps)
1311 g_assert (slot_offset <= 0);
1312 g_assert (slot_offset % SIZEOF_SLOT == 0);
1314 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1318 fp_offset_to_slot (MonoCompile *cfg, int offset)
1320 MonoCompileGC *gcfg = cfg->gc_info;
1322 return (offset - gcfg->min_offset) / SIZEOF_SLOT;
1326 slot_to_fp_offset (MonoCompile *cfg, int slot)
1328 MonoCompileGC *gcfg = cfg->gc_info;
1330 return (slot * SIZEOF_SLOT) + gcfg->min_offset;
1333 static inline MONO_ALWAYS_INLINE void
1334 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1336 g_assert (slot >= 0 && slot < gcfg->nslots);
1338 if (type == SLOT_PIN) {
1339 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1340 set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1341 } else if (type == SLOT_REF) {
1342 set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1343 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1344 } else if (type == SLOT_NOREF) {
1345 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1346 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1351 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1354 guint8 *ref_bitmap, *pin_bitmap;
1359 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1360 set_slot (gcfg, slot, cindex, type);
1362 ref_bitmap = gcfg->stack_ref_bitmap;
1363 pin_bitmap = gcfg->stack_pin_bitmap;
1364 width = gcfg->stack_bitmap_width;
1367 if (type == SLOT_PIN) {
1368 memset (ref_bitmap + pos, 0, width);
1369 memset (pin_bitmap + pos, 0xff, width);
1370 } else if (type == SLOT_REF) {
1371 memset (ref_bitmap + pos, 0xff, width);
1372 memset (pin_bitmap + pos, 0, width);
1373 } else if (type == SLOT_NOREF) {
1374 memset (ref_bitmap + pos, 0, width);
1375 memset (pin_bitmap + pos, 0, width);
1380 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1384 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1385 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1386 if (callsite_offset >= from && callsite_offset < to)
1387 set_slot (gcfg, slot, cindex, type);
1392 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1394 g_assert (slot >= 0 && slot < gcfg->nregs);
1396 if (type == SLOT_PIN) {
1397 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1398 set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1399 } else if (type == SLOT_REF) {
1400 set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1401 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1402 } else if (type == SLOT_NOREF) {
1403 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1404 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1409 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1413 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1414 set_reg_slot (gcfg, slot, cindex, type);
1418 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1422 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1423 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1424 if (callsite_offset >= from && callsite_offset < to)
1425 set_reg_slot (gcfg, slot, cindex, type);
1430 process_spill_slots (MonoCompile *cfg)
1432 MonoCompileGC *gcfg = cfg->gc_info;
1437 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1438 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1439 for (l = bb->spill_slot_defs; l; l = l->next) {
1440 MonoInst *def = l->data;
1441 int spill_slot = def->inst_c0;
1442 int bank = def->inst_c1;
1443 int offset = cfg->spill_info [bank][spill_slot].offset;
1444 int slot = fp_offset_to_slot (cfg, offset);
1446 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1447 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1451 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1452 for (l = bb->spill_slot_defs; l; l = l->next) {
1453 MonoInst *def = l->data;
1454 int spill_slot = def->inst_c0;
1455 int bank = def->inst_c1;
1456 int offset = cfg->spill_info [bank][spill_slot].offset;
1457 int slot = fp_offset_to_slot (cfg, offset);
1460 if (bank == MONO_REG_INT_MP)
1466 * Extend the live interval for the GC tracked spill slots
1467 * defined in this bblock.
1468 * FIXME: This is not needed.
1470 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1472 if (cfg->verbose_level > 1)
1473 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
1477 /* Set fp spill slots to NOREF */
1478 for (i = 0; i < cfg->spill_info_len [MONO_REG_DOUBLE]; ++i) {
1479 int offset = cfg->spill_info [MONO_REG_DOUBLE][i].offset;
1485 slot = fp_offset_to_slot (cfg, offset);
1487 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1489 if (cfg->verbose_level > 1)
1490 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1493 /* Set int spill slots to NOREF */
1494 for (i = 0; i < cfg->spill_info_len [MONO_REG_INT]; ++i) {
1495 int offset = cfg->spill_info [MONO_REG_INT][i].offset;
1501 slot = fp_offset_to_slot (cfg, offset);
1503 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1504 if (cfg->verbose_level > 1)
1505 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1510 * process_other_slots:
1512 * Process stack slots registered using mini_gc_set_slot_type_... ().
1515 process_other_slots (MonoCompile *cfg)
1517 MonoCompileGC *gcfg = cfg->gc_info;
1520 /* Relative to the CFA */
1521 for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
1522 guint data = GPOINTER_TO_UINT (l->data);
1523 int cfa_slot = data >> 16;
1524 GCSlotType type = data & 0xff;
1528 * Map the cfa relative slot to an fp relative slot.
1529 * slot_addr == cfa - <cfa_slot>*4/8
1530 * fp + cfa_offset == cfa
1531 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1533 slot = (cfg->cfa_offset / SIZEOF_SLOT) - cfa_slot - (gcfg->min_offset / SIZEOF_SLOT);
1535 set_slot_everywhere (gcfg, slot, type);
1537 if (cfg->verbose_level > 1) {
1538 int fp_offset = slot_to_fp_offset (cfg, slot);
1539 if (type == SLOT_NOREF)
1540 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));
1544 /* Relative to the FP */
1545 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1546 gint data = GPOINTER_TO_INT (l->data);
1547 int offset = data >> 16;
1548 GCSlotType type = data & 0xff;
1551 slot = fp_offset_to_slot (cfg, offset);
1553 set_slot_everywhere (gcfg, slot, type);
1555 /* Liveness for these slots is handled by process_spill_slots () */
1557 if (cfg->verbose_level > 1) {
1558 if (type == SLOT_REF)
1559 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
1560 else if (type == SLOT_NOREF)
1561 printf ("\tnoref slot at 0x%x(fp) (slot = %d)\n", offset, slot);
1567 get_vtype_bitmap (MonoType *t, int *numbits)
1569 MonoClass *klass = mono_class_from_mono_type (t);
1571 if (klass->generic_container || mono_class_is_open_constructed_type (t)) {
1572 /* FIXME: Generic sharing */
1575 mono_class_compute_gc_descriptor (klass);
1577 return mono_gc_get_bitmap_for_descr (klass->gc_descr, numbits);
1581 static inline const char*
1582 get_offset_sign (int offset)
1584 return offset < 0 ? "-" : "+";
1588 get_offset_val (int offset)
1590 return offset < 0 ? (- offset) : offset;
1594 process_variables (MonoCompile *cfg)
1596 MonoCompileGC *gcfg = cfg->gc_info;
1597 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1598 int i, locals_min_slot, locals_max_slot, cindex;
1602 int locals_min_offset = gcfg->locals_min_offset;
1603 int locals_max_offset = gcfg->locals_max_offset;
1605 /* Slots for locals are NOREF by default */
1606 locals_min_slot = (locals_min_offset - gcfg->min_offset) / SIZEOF_SLOT;
1607 locals_max_slot = (locals_max_offset - gcfg->min_offset) / SIZEOF_SLOT;
1608 for (i = locals_min_slot; i < locals_max_slot; ++i) {
1609 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1613 * Compute the offset where variables are initialized in the first bblock, if any.
1615 pc_offsets = g_new0 (int, cfg->next_vreg);
1617 bb = cfg->bb_entry->next_bb;
1618 MONO_BB_FOR_EACH_INS (bb, tmp) {
1619 if (tmp->opcode == OP_GC_LIVENESS_DEF) {
1620 int vreg = tmp->inst_c1;
1621 if (pc_offsets [vreg] == 0) {
1622 g_assert (tmp->backend.pc_offset > 0);
1623 pc_offsets [vreg] = tmp->backend.pc_offset;
1629 * Stack slots holding arguments are initialized in the prolog.
1630 * This means we can treat them alive for the whole method.
1632 for (i = 0; i < cfg->num_varinfo; i++) {
1633 MonoInst *ins = cfg->varinfo [i];
1634 MonoType *t = ins->inst_vtype;
1637 gboolean byref, is_this = FALSE;
1638 gboolean is_arg = i < cfg->locals_start;
1640 if (ins == cfg->ret) {
1641 if (!(ins->opcode == OP_REGOFFSET && MONO_TYPE_ISSTRUCT (t)))
1645 vmv = MONO_VARINFO (cfg, i);
1647 /* For some reason, 'this' is byref */
1648 if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
1649 t = &cfg->method->klass->byval_arg;
1655 if (ins->opcode == OP_REGVAR) {
1657 GCSlotType slot_type;
1659 t = mini_type_get_underlying_type (NULL, t);
1662 g_assert (hreg < MONO_MAX_IREGS);
1665 slot_type = SLOT_PIN;
1667 slot_type = mini_type_is_reference (cfg, t) ? SLOT_REF : SLOT_NOREF;
1669 if (slot_type == SLOT_PIN) {
1670 /* These have no live interval, be conservative */
1671 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1674 * Unlike variables allocated to the stack, we generate liveness info
1675 * for noref vars in registers in mono_spill_global_vars (), because
1676 * knowing that a register doesn't contain a ref allows us to mark its save
1677 * locations precisely.
1679 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1680 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1681 set_reg_slot (gcfg, hreg, cindex, slot_type);
1684 if (cfg->verbose_level > 1) {
1685 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
1691 if (ins->opcode != OP_REGOFFSET)
1694 if (ins->inst_offset % SIZEOF_SLOT != 0)
1697 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1699 if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
1700 /* These do not get stored in the prolog */
1701 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1703 if (cfg->verbose_level > 1) {
1704 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));
1709 if (MONO_TYPE_ISSTRUCT (t)) {
1711 gsize *bitmap = NULL;
1712 gboolean pin = FALSE;
1716 if (ins->backend.is_pinvoke)
1717 size = mono_class_native_size (ins->klass, NULL);
1719 size = mono_class_value_size (ins->klass, NULL);
1720 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
1722 if (cfg->verbose_level > 1)
1723 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));
1725 if (!ins->klass->has_references) {
1727 for (j = 0; j < size_in_slots; ++j)
1728 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1733 bitmap = get_vtype_bitmap (t, &numbits);
1738 * Most vtypes are marked volatile because of the LDADDR instructions,
1739 * and they have no liveness information since they are decomposed
1740 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1741 * them during VZERO decomposition.
1744 if (!pc_offsets [vmv->vreg])
1747 if (ins->backend.is_pinvoke)
1752 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1753 if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
1754 for (j = 0; j < numbits; ++j) {
1755 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1756 /* The descriptor is for the boxed object */
1757 set_slot (gcfg, (pos + j - (sizeof (MonoObject) / SIZEOF_SLOT)), cindex, pin ? SLOT_PIN : SLOT_REF);
1763 if (cfg->verbose_level > 1) {
1764 for (j = 0; j < numbits; ++j) {
1765 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD)))
1766 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)));
1770 if (cfg->verbose_level > 1)
1771 printf ("\t\tpinned\n");
1772 for (j = 0; j < size_in_slots; ++j) {
1773 set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
1782 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1783 /* Vret addr etc. */
1788 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1790 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1791 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1792 set_slot (gcfg, pos, cindex, SLOT_PIN);
1794 if (cfg->verbose_level > 1)
1795 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));
1800 * This is currently disabled, but could be enabled to debug crashes.
1803 if (t->type == MONO_TYPE_I) {
1805 * Variables created in mono_handle_global_vregs have type I, but they
1806 * could hold GC refs since the vregs they were created from might not been
1807 * marked as holding a GC ref. So be conservative.
1809 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1814 t = mini_type_get_underlying_type (NULL, t);
1816 if (!mini_type_is_reference (cfg, t)) {
1817 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1818 if (cfg->verbose_level > 1)
1819 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));
1820 if (!t->byref && sizeof (mgreg_t) == 4 && (t->type == MONO_TYPE_I8 || t->type == MONO_TYPE_U8 || t->type == MONO_TYPE_R8)) {
1821 set_slot_everywhere (gcfg, pos + 1, SLOT_NOREF);
1822 if (cfg->verbose_level > 1)
1823 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));
1828 /* 'this' is marked INDIRECT for gshared methods */
1829 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1831 * For volatile variables, treat them alive from the point they are
1832 * initialized in the first bblock until the end of the method.
1835 set_slot_everywhere (gcfg, pos, SLOT_REF);
1836 } else if (pc_offsets [vmv->vreg]) {
1837 set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
1838 set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
1840 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1842 if (cfg->verbose_level > 1)
1843 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));
1848 /* Live for the whole method */
1849 set_slot_everywhere (gcfg, pos, SLOT_REF);
1851 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1852 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1853 set_slot (gcfg, pos, cindex, SLOT_REF);
1856 if (cfg->verbose_level > 1) {
1857 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));
1861 g_free (pc_offsets);
1865 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1868 * Convert a sp relative offset to a slot index. This is
1869 * platform specific.
1872 /* fp = sp + offset */
1873 g_assert (cfg->frame_reg == AMD64_RBP);
1874 return (- cfg->arch.sp_fp_offset + sp_offset);
1875 #elif defined(TARGET_X86)
1876 /* The offset is computed from the sp at the start of the call sequence */
1877 g_assert (cfg->frame_reg == X86_EBP);
1878 #ifdef MONO_X86_NO_PUSHES
1879 return (- cfg->arch.sp_fp_offset + sp_offset);
1881 return (- cfg->arch.sp_fp_offset - sp_offset);
1890 process_param_area_slots (MonoCompile *cfg)
1892 MonoCompileGC *gcfg = cfg->gc_info;
1897 * These slots are used for passing parameters during calls. They are sp relative, not
1898 * fp relative, so they are harder to handle.
1900 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1901 /* The distance between fp and sp is not constant */
1904 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1906 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1907 GCCallSite *callsite = gcfg->callsites [cindex];
1910 for (l = callsite->param_slots; l; l = l->next) {
1911 MonoInst *def = l->data;
1912 MonoType *t = def->inst_vtype;
1913 int sp_offset = def->inst_offset;
1914 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1915 int slot = fp_offset_to_slot (cfg, fp_offset);
1919 if (MONO_TYPE_ISSTRUCT (t)) {
1920 size = mini_type_stack_size_full (cfg->generic_sharing_context, t, &align, FALSE);
1922 size = sizeof (mgreg_t);
1925 for (i = 0; i < size / sizeof (mgreg_t); ++i) {
1926 g_assert (slot + i >= 0 && slot + i < gcfg->nslots);
1927 is_param [slot + i] = TRUE;
1932 /* All param area slots are noref by default */
1933 for (i = 0; i < gcfg->nslots; ++i) {
1935 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1939 * We treat param area slots as being part of the callee's frame, to be able to handle tail calls which overwrite
1940 * the argument area of the caller.
1945 process_finally_clauses (MonoCompile *cfg)
1947 MonoCompileGC *gcfg = cfg->gc_info;
1948 GCCallSite **callsites;
1950 gboolean has_finally;
1951 int i, j, nslots, nregs;
1953 ncallsites = gcfg->ncallsites;
1954 nslots = gcfg->nslots;
1955 nregs = gcfg->nregs;
1956 callsites = gcfg->callsites;
1959 * The calls to the finally clauses don't show up in the cfg. See
1960 * test_0_liveness_8 ().
1961 * Variables accessed inside the finally clause are already marked VOLATILE by
1962 * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
1963 * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
1965 has_finally = FALSE;
1966 for (i = 0; i < cfg->header->num_clauses; ++i) {
1967 MonoExceptionClause *clause = &cfg->header->clauses [i];
1969 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1974 if (cfg->verbose_level > 1)
1975 printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
1976 for (j = 0; j < ncallsites; ++j) {
1977 MonoBasicBlock *bb = callsites [j]->bb;
1978 MonoExceptionClause *clause;
1979 gboolean is_in_finally = FALSE;
1981 for (i = 0; i < cfg->header->num_clauses; ++i) {
1982 clause = &cfg->header->clauses [i];
1984 if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
1985 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1986 is_in_finally = TRUE;
1992 if (is_in_finally) {
1993 for (i = 0; i < nslots; ++i)
1994 set_slot (gcfg, i, j, SLOT_PIN);
1995 for (i = 0; i < nregs; ++i)
1996 set_reg_slot (gcfg, i, j, SLOT_PIN);
2003 compute_frame_size (MonoCompile *cfg)
2005 int i, locals_min_offset, locals_max_offset, cfa_min_offset, cfa_max_offset;
2006 int min_offset, max_offset;
2007 MonoCompileGC *gcfg = cfg->gc_info;
2008 MonoMethodSignature *sig = mono_method_signature (cfg->method);
2011 /* Compute min/max offsets from the fp */
2014 #if defined(TARGET_AMD64) || defined(TARGET_X86) || defined(TARGET_ARM) || defined(TARGET_S390X)
2015 locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, SIZEOF_SLOT);
2016 locals_max_offset = cfg->locals_max_stack_offset;
2018 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
2022 locals_min_offset = ALIGN_TO (locals_min_offset, SIZEOF_SLOT);
2023 locals_max_offset = ALIGN_TO (locals_max_offset, SIZEOF_SLOT);
2025 min_offset = locals_min_offset;
2026 max_offset = locals_max_offset;
2029 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
2030 MonoInst *ins = cfg->args [i];
2032 if (ins->opcode == OP_REGOFFSET) {
2033 int size, size_in_slots;
2034 size = mini_type_stack_size_full (cfg->generic_sharing_context, ins->inst_vtype, NULL, ins->backend.is_pinvoke);
2035 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
2037 min_offset = MIN (min_offset, ins->inst_offset);
2038 max_offset = MAX ((int)max_offset, (int)(ins->inst_offset + (size_in_slots * SIZEOF_SLOT)));
2043 g_assert (cfg->frame_reg == cfg->cfa_reg);
2044 g_assert (cfg->cfa_offset > 0);
2046 cfa_max_offset = cfg->cfa_offset;
2048 min_offset = MIN (min_offset, cfa_min_offset);
2049 max_offset = MAX (max_offset, cfa_max_offset);
2051 /* Fp relative slots */
2052 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
2053 gint data = GPOINTER_TO_INT (l->data);
2054 int offset = data >> 16;
2056 min_offset = MIN (min_offset, offset);
2060 if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
2061 int stack_offset = ALIGN_TO (cfg->stack_offset, SIZEOF_SLOT);
2062 min_offset = MIN (min_offset, (-stack_offset));
2065 /* Param area slots */
2067 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
2068 #elif defined(TARGET_X86)
2069 #ifdef MONO_X86_NO_PUSHES
2070 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
2072 min_offset = MIN (min_offset, - (cfg->arch.sp_fp_offset + cfg->arch.param_area_size));
2074 #elif defined(TARGET_ARM)
2076 #elif defined(TARGET_s390X)
2082 gcfg->min_offset = min_offset;
2083 gcfg->max_offset = max_offset;
2084 gcfg->locals_min_offset = locals_min_offset;
2085 gcfg->locals_max_offset = locals_max_offset;
2089 init_gcfg (MonoCompile *cfg)
2091 int i, nregs, nslots;
2092 MonoCompileGC *gcfg = cfg->gc_info;
2093 GCCallSite **callsites;
2102 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2103 ncallsites += g_slist_length (bb->gc_callsites);
2105 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
2107 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2108 for (l = bb->gc_callsites; l; l = l->next)
2109 callsites [i++] = l->data;
2112 /* The callsites should already be ordered by pc offset */
2113 for (i = 1; i < ncallsites; ++i)
2114 g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
2117 * The stack frame looks like this:
2119 * <fp + max_offset> == cfa -> <end of previous frame>
2120 * <other stack slots>
2122 * <other stack slots>
2123 * fp + min_offset ->
2128 if (cfg->verbose_level > 1)
2129 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
2131 nslots = (gcfg->max_offset - gcfg->min_offset) / SIZEOF_SLOT;
2134 gcfg->nslots = nslots;
2135 gcfg->nregs = nregs;
2136 gcfg->callsites = callsites;
2137 gcfg->ncallsites = ncallsites;
2138 gcfg->stack_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
2139 gcfg->reg_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
2140 gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
2141 gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
2142 gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
2143 gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
2145 /* All slots start out as PIN */
2146 memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * nregs);
2147 for (i = 0; i < nregs; ++i) {
2149 * By default, registers are NOREF.
2150 * It is possible for a callee to save them before being defined in this method,
2151 * but the saved value is dead too, so it doesn't need to be marked.
2153 if ((cfg->used_int_regs & (1 << i)))
2154 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
2158 static inline gboolean
2159 has_bit_set (guint8 *bitmap, int width, int slot)
2162 int pos = width * slot;
2164 for (i = 0; i < width; ++i) {
2165 if (bitmap [pos + i])
2172 create_map (MonoCompile *cfg)
2175 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
2177 int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
2178 int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
2180 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
2181 MonoCompileGC *gcfg = cfg->gc_info;
2182 GCCallSite **callsites;
2184 guint8 *bitmap, *bitmaps;
2185 guint32 reg_ref_mask, reg_pin_mask;
2187 ncallsites = gcfg->ncallsites;
2188 nslots = gcfg->nslots;
2189 nregs = gcfg->nregs;
2190 callsites = gcfg->callsites;
2193 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
2194 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
2196 has_ref_slots = FALSE;
2197 has_pin_slots = FALSE;
2200 memset (ntypes, 0, sizeof (ntypes));
2201 for (i = 0; i < nslots; ++i) {
2202 gboolean has_ref = FALSE;
2203 gboolean has_pin = FALSE;
2205 if (has_bit_set (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i))
2207 if (has_bit_set (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i))
2211 has_ref_slots = TRUE;
2213 has_pin_slots = TRUE;
2216 ntypes [SLOT_REF] ++;
2218 ntypes [SLOT_PIN] ++;
2220 ntypes [SLOT_NOREF] ++;
2222 if (has_ref || has_pin) {
2229 start = end = nslots;
2231 g_assert (start != -1);
2232 g_assert (start < end);
2235 has_ref_regs = FALSE;
2236 has_pin_regs = FALSE;
2241 for (i = 0; i < nregs; ++i) {
2242 gboolean has_ref = FALSE;
2243 gboolean has_pin = FALSE;
2245 if (!(cfg->used_int_regs & (1 << i)))
2248 if (has_bit_set (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i))
2250 if (has_bit_set (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i))
2254 reg_ref_mask |= (1 << i);
2255 has_ref_regs = TRUE;
2259 reg_pin_mask |= (1 << i);
2260 has_pin_regs = TRUE;
2265 if (cfg->verbose_level > 1)
2266 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);
2268 /* Create the GC Map */
2270 /* The work bitmaps have one row for each slot, since this is how we access them during construction */
2271 stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
2272 stack_bitmap_size = stack_bitmap_width * ncallsites;
2273 reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
2274 reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
2275 reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
2276 reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
2277 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);
2279 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2281 map->frame_reg = cfg->frame_reg;
2282 map->start_offset = gcfg->min_offset;
2283 map->end_offset = gcfg->min_offset + (nslots * SIZEOF_SLOT);
2284 map->map_offset = start * SIZEOF_SLOT;
2285 map->nslots = end - start;
2286 map->has_ref_slots = has_ref_slots;
2287 map->has_pin_slots = has_pin_slots;
2288 map->has_ref_regs = has_ref_regs;
2289 map->has_pin_regs = has_pin_regs;
2290 g_assert (nregs < 32);
2291 map->used_int_regs = cfg->used_int_regs;
2292 map->reg_ref_mask = reg_ref_mask;
2293 map->reg_pin_mask = reg_pin_mask;
2294 map->nref_regs = nref_regs;
2295 map->npin_regs = npin_regs;
2297 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2300 if (has_ref_slots) {
2301 map->stack_ref_bitmap_offset = bitmaps_offset;
2302 bitmaps_offset += stack_bitmap_size;
2304 bitmap = &bitmaps [map->stack_ref_bitmap_offset];
2305 for (i = 0; i < nslots; ++i) {
2306 for (j = 0; j < ncallsites; ++j) {
2307 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i, j))
2308 set_bit (bitmap, stack_bitmap_width, j, i - start);
2312 if (has_pin_slots) {
2313 map->stack_pin_bitmap_offset = bitmaps_offset;
2314 bitmaps_offset += stack_bitmap_size;
2316 bitmap = &bitmaps [map->stack_pin_bitmap_offset];
2317 for (i = 0; i < nslots; ++i) {
2318 for (j = 0; j < ncallsites; ++j) {
2319 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i, j))
2320 set_bit (bitmap, stack_bitmap_width, j, i - start);
2325 map->reg_ref_bitmap_offset = bitmaps_offset;
2326 bitmaps_offset += reg_ref_bitmap_size;
2328 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2330 for (i = 0; i < nregs; ++i) {
2331 if (reg_ref_mask & (1 << i)) {
2332 for (j = 0; j < ncallsites; ++j) {
2333 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i, j))
2334 set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
2341 map->reg_pin_bitmap_offset = bitmaps_offset;
2342 bitmaps_offset += reg_pin_bitmap_size;
2344 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2346 for (i = 0; i < nregs; ++i) {
2347 if (reg_pin_mask & (1 << i)) {
2348 for (j = 0; j < ncallsites; ++j) {
2349 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i, j))
2350 set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
2358 map->ncallsites = ncallsites;
2359 if (cfg->code_len < 256)
2360 map->callsite_entry_size = 1;
2361 else if (cfg->code_len < 65536)
2362 map->callsite_entry_size = 2;
2364 map->callsite_entry_size = 4;
2366 /* Encode the GC Map */
2374 encode_gc_map (map, buf, &endbuf);
2375 g_assert (endbuf - buf < 256);
2377 encoded_size = endbuf - buf;
2378 alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
2380 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2381 //emap->ref_slots = map->ref_slots;
2383 /* Encoded fixed fields */
2384 p = &emap->encoded [0];
2385 //emap->encoded_size = encoded_size;
2386 memcpy (p, buf, encoded_size);
2389 /* Callsite table */
2390 p = (guint8*)ALIGN_TO ((mgreg_t)p, map->callsite_entry_size);
2391 if (map->callsite_entry_size == 1) {
2392 guint8 *offsets = p;
2393 for (i = 0; i < ncallsites; ++i)
2394 offsets [i] = callsites [i]->pc_offset;
2395 stats.gc_callsites8_size += ncallsites * sizeof (guint8);
2396 } else if (map->callsite_entry_size == 2) {
2397 guint16 *offsets = (guint16*)p;
2398 for (i = 0; i < ncallsites; ++i)
2399 offsets [i] = callsites [i]->pc_offset;
2400 stats.gc_callsites16_size += ncallsites * sizeof (guint16);
2402 guint32 *offsets = (guint32*)p;
2403 for (i = 0; i < ncallsites; ++i)
2404 offsets [i] = callsites [i]->pc_offset;
2405 stats.gc_callsites32_size += ncallsites * sizeof (guint32);
2407 p += ncallsites * map->callsite_entry_size;
2410 memcpy (p, bitmaps, bitmaps_size);
2413 g_assert ((guint8*)p - (guint8*)emap <= alloc_size);
2415 stats.gc_maps_size += alloc_size;
2416 stats.gc_callsites_size += ncallsites * map->callsite_entry_size;
2417 stats.gc_bitmaps_size += bitmaps_size;
2418 stats.gc_map_struct_size += sizeof (GCEncodedMap) + encoded_size;
2420 cfg->jit_info->gc_info = emap;
2422 cfg->gc_map = (guint8*)emap;
2423 cfg->gc_map_size = alloc_size;
2426 stats.all_slots += nslots;
2427 stats.ref_slots += ntypes [SLOT_REF];
2428 stats.noref_slots += ntypes [SLOT_NOREF];
2429 stats.pin_slots += ntypes [SLOT_PIN];
2433 mini_gc_create_gc_map (MonoCompile *cfg)
2435 if (!cfg->compute_gc_maps)
2439 * During marking, all frames except the top frame are at a call site, and we mark the
2440 * top frame conservatively. This means that we only need to compute and record
2441 * GC maps for call sites.
2444 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2445 /* Without liveness info, the live ranges are not precise enough */
2448 mono_analyze_liveness_gc (cfg);
2450 compute_frame_size (cfg);
2454 process_spill_slots (cfg);
2455 process_other_slots (cfg);
2456 process_param_area_slots (cfg);
2457 process_variables (cfg);
2458 process_finally_clauses (cfg);
2463 #endif /* DISABLE_JIT */
2466 parse_debug_options (void)
2471 env = g_getenv ("MONO_GCMAP_DEBUG");
2475 opts = g_strsplit (env, ",", -1);
2476 for (ptr = opts; ptr && *ptr; ptr ++) {
2477 /* No options yet */
2478 fprintf (stderr, "Invalid format for the MONO_GCMAP_DEBUG env variable: '%s'\n", env);
2489 memset (&cb, 0, sizeof (cb));
2490 cb.thread_attach_func = thread_attach_func;
2491 cb.thread_detach_func = thread_detach_func;
2492 cb.thread_suspend_func = thread_suspend_func;
2493 /* Comment this out to disable precise stack marking */
2494 cb.thread_mark_func = thread_mark_func;
2495 mono_gc_set_gc_callbacks (&cb);
2497 logfile = mono_gc_get_logfile ();
2499 parse_debug_options ();
2501 mono_counters_register ("GC Maps size",
2502 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
2503 mono_counters_register ("GC Call Sites size",
2504 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
2505 mono_counters_register ("GC Bitmaps size",
2506 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
2507 mono_counters_register ("GC Map struct size",
2508 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
2509 mono_counters_register ("GC Call Sites encoded using 8 bits",
2510 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
2511 mono_counters_register ("GC Call Sites encoded using 16 bits",
2512 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
2513 mono_counters_register ("GC Call Sites encoded using 32 bits",
2514 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
2516 mono_counters_register ("GC Map slots (all)",
2517 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
2518 mono_counters_register ("GC Map slots (ref)",
2519 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.ref_slots);
2520 mono_counters_register ("GC Map slots (noref)",
2521 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.noref_slots);
2522 mono_counters_register ("GC Map slots (pin)",
2523 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
2525 mono_counters_register ("GC TLS Data size",
2526 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2528 mono_counters_register ("Stack space scanned (all)",
2529 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
2530 mono_counters_register ("Stack space scanned (native)",
2531 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_native);
2532 mono_counters_register ("Stack space scanned (other)",
2533 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_other);
2534 mono_counters_register ("Stack space scanned (using GC Maps)",
2535 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
2536 mono_counters_register ("Stack space scanned (precise)",
2537 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_precisely);
2538 mono_counters_register ("Stack space scanned (pin)",
2539 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_conservatively);
2540 mono_counters_register ("Stack space scanned (pin registers)",
2541 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_registers);
2547 mini_gc_enable_gc_maps_for_aot (void)
2559 mini_gc_init_gc_map (MonoCompile *cfg)
2564 mini_gc_create_gc_map (MonoCompile *cfg)
2569 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2574 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2578 #endif /* DISABLE_JIT */
2587 * Set GC specific options in CFG.
2590 mini_gc_init_cfg (MonoCompile *cfg)
2592 if (mono_gc_is_moving ()) {
2593 cfg->disable_ref_noref_stack_slot_share = TRUE;
2594 cfg->gen_write_barriers = TRUE;
2597 mini_gc_init_gc_map (cfg);
2600 #endif /* DISABLE_JIT */
2603 * Problems with the current code:
2604 * - the stack walk is slow
2605 * - vtypes/refs used in EH regions are treated conservatively
2606 * - if the code is finished, less pinning will be done, causing problems because
2607 * we promote all surviving objects to old-gen.
2608 * - the unwind code can't handle a method stopped inside a finally region, it thinks the caller is
2609 * another method, but in reality it is either the exception handling code or the CALL_HANDLER opcode.
2610 * This manifests in "Unable to find ip offset x in callsite list" assertions.
2611 * - the unwind code also can't handle frames which are in the epilog, since the unwind info is not
2616 * Ideas for creating smaller GC maps:
2617 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2619 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2621 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2622 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2623 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2624 * from an index entry.