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)
9 * Licensed under the MIT license. See LICENSE file in the project root for full license information.
14 #include <mono/metadata/gc-internals.h>
17 get_provenance (StackFrameInfo *frame, MonoContext *ctx, gpointer data)
19 MonoJitInfo *ji = frame->ji;
23 method = jinfo_get_method (ji);
24 if (method->wrapper_type != MONO_WRAPPER_NONE)
26 *(gpointer *)data = method;
31 get_provenance_func (void)
33 gpointer provenance = NULL;
34 mono_walk_stack (get_provenance, MONO_UNWIND_DEFAULT, (gpointer)&provenance);
39 //#if defined(MONO_ARCH_GC_MAPS_SUPPORTED)
41 #include <mono/metadata/sgen-conf.h>
42 #include <mono/metadata/gc-internals.h>
43 #include <mono/utils/mono-counters.h>
45 #define SIZEOF_SLOT ((int)sizeof (mgreg_t))
47 #define GC_BITS_PER_WORD (sizeof (mword) * 8)
49 /* Contains state needed by the GC Map construction code */
52 * This contains information about stack slots initialized in the prolog, encoded using
53 * (slot_index << 16) | slot_type. The slot_index is relative to the CFA, i.e. 0
54 * means cfa+0, 1 means cfa-4/8, etc.
56 GSList *stack_slots_from_cfa;
57 /* Same for stack slots relative to the frame pointer */
58 GSList *stack_slots_from_fp;
60 /* Number of slots in the map */
62 /* The number of registers in the map */
64 /* Min and Max offsets of the stack frame relative to fp */
65 int min_offset, max_offset;
66 /* Same for the locals area */
67 int locals_min_offset, locals_max_offset;
69 /* The call sites where this frame can be stopped during GC */
70 GCCallSite **callsites;
71 /* The number of call sites */
75 * The width of the stack bitmaps in bytes. This is not equal to the bitmap width at
76 * runtime, since it includes columns which are 0.
78 int stack_bitmap_width;
80 * A bitmap whose width equals nslots, and whose height equals ncallsites.
81 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
84 guint8 *stack_ref_bitmap;
85 /* Same for SLOT_PIN */
86 guint8 *stack_pin_bitmap;
89 * Similar bitmaps for registers. These have width MONO_MAX_IREGS in bits.
92 guint8 *reg_ref_bitmap;
93 guint8 *reg_pin_bitmap;
96 #define ALIGN_TO(val,align) ((((mgreg_t)val) + ((align) - 1)) & ~((align) - 1))
101 /* We don't support debug levels, its all-or-nothing */
102 #define DEBUG(s) do { s; fflush (logfile); } while (0)
103 #define DEBUG_ENABLED 1
110 #define DEBUG_PRECISE(s) do { s; } while (0)
111 #define DEBUG_PRECISE_ENABLED
113 #define DEBUG_PRECISE(s)
117 * Contains information collected during the conservative stack marking pass,
118 * used during the precise pass. This helps to avoid doing a stack walk twice, which
124 int frame_start_offset;
126 /* Relative to stack_start */
127 int reg_locations [MONO_MAX_IREGS];
128 #ifdef DEBUG_PRECISE_ENABLED
131 int regs [MONO_MAX_IREGS];
135 /* Max number of frames stored in the TLS data */
136 #define MAX_FRAMES 50
139 * Per-thread data kept by this module. This is stored in the GC and passed to us as
140 * parameters, instead of being stored in a TLS variable, since during a collection,
141 * only the collection thread is active.
144 MonoThreadUnwindState unwind_state;
145 MonoThreadInfo *info;
148 gpointer ref_to_track;
149 /* Number of frames collected during the !precise pass */
151 FrameInfo frames [MAX_FRAMES];
154 /* These are constant so don't store them in the GC Maps */
155 /* Number of registers stored in gc maps */
156 #define NREGS MONO_MAX_IREGS
160 * Contains information needed to mark a stack frame.
161 * This is a transient structure, created from a compressed representation on-demand.
165 * The offsets of the GC tracked area inside the stack frame relative to the frame pointer.
166 * This includes memory which is NOREF thus doesn't need GC maps.
171 * The offset relative to frame_offset where the the memory described by the GC maps
175 /* The number of stack slots in the map */
177 /* The frame pointer register */
179 /* The size of each callsite table entry */
180 guint8 callsite_entry_size;
181 guint has_pin_slots : 1;
182 guint has_ref_slots : 1;
183 guint has_ref_regs : 1;
184 guint has_pin_regs : 1;
186 /* The offsets below are into an external bitmaps array */
189 * A bitmap whose width is equal to bitmap_width, and whose height is equal to ncallsites.
190 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
193 guint32 stack_ref_bitmap_offset;
195 * Same for SLOT_PIN. It is possible that the same bit is set in both bitmaps at
196 * different callsites, if the slot starts out as PIN, and later changes to REF.
198 guint32 stack_pin_bitmap_offset;
201 * Corresponding bitmaps for registers
202 * These have width equal to the number of bits set in reg_ref_mask/reg_pin_mask.
203 * FIXME: Merge these with the normal bitmaps, i.e. reserve the first x slots for them ?
205 guint32 reg_pin_bitmap_offset;
206 guint32 reg_ref_bitmap_offset;
208 guint32 used_int_regs, reg_ref_mask, reg_pin_mask;
210 /* The number of bits set in the two masks above */
211 guint8 nref_regs, npin_regs;
214 * A bit array marking slots which contain refs.
215 * This is used only for debugging.
219 /* Callsite offsets */
220 /* These can take up a lot of space, so encode them compactly */
230 * A compressed version of GCMap. This is what gets stored in MonoJitInfo.
234 //guint8 encoded_size;
237 * The arrays below are embedded after the struct.
238 * Their address needs to be computed.
241 /* The fixed fields of the GCMap encoded using LEB128 */
242 guint8 encoded [MONO_ZERO_LEN_ARRAY];
244 /* An array of ncallsites entries, each entry is callsite_entry_size bytes long */
245 guint8 callsites [MONO_ZERO_LEN_ARRAY];
248 guint8 bitmaps [MONO_ZERO_LEN_ARRAY];
251 static int precise_frame_count [2], precise_frame_limit = -1;
252 static gboolean precise_frame_limit_inited;
258 int scanned_precisely;
259 int scanned_conservatively;
260 int scanned_registers;
270 int gc_callsites_size;
271 int gc_callsites8_size;
272 int gc_callsites16_size;
273 int gc_callsites32_size;
275 int gc_map_struct_size;
279 static JITGCStats stats;
281 static FILE *logfile;
283 static gboolean enable_gc_maps_for_aot;
286 mini_gc_enable_gc_maps_for_aot (void)
288 enable_gc_maps_for_aot = TRUE;
291 // FIXME: Move these to a shared place
294 encode_uleb128 (guint32 value, guint8 *buf, guint8 **endbuf)
299 guint8 b = value & 0x7f;
301 if (value != 0) /* more bytes to come */
309 static G_GNUC_UNUSED void
310 encode_sleb128 (gint32 value, guint8 *buf, guint8 **endbuf)
313 gboolean negative = (value < 0);
321 /* the following is unnecessary if the
322 * implementation of >>= uses an arithmetic rather
323 * than logical shift for a signed left operand
327 value |= - (1 <<(size - 7));
328 /* sign bit of byte is second high order bit (0x40) */
329 if ((value == 0 && !(byte & 0x40)) ||
330 (value == -1 && (byte & 0x40)))
340 static inline guint32
341 decode_uleb128 (guint8 *buf, guint8 **endbuf)
351 res = res | (((int)(b & 0x7f)) << shift);
363 decode_sleb128 (guint8 *buf, guint8 **endbuf)
373 res = res | (((int)(b & 0x7f)) << shift);
376 if (shift < 32 && (b & 0x40))
377 res |= - (1 << shift);
388 encode_frame_reg (int frame_reg)
391 if (frame_reg == AMD64_RSP)
393 else if (frame_reg == AMD64_RBP)
395 #elif defined(TARGET_X86)
396 if (frame_reg == X86_EBP)
398 else if (frame_reg == X86_ESP)
400 #elif defined(TARGET_ARM)
401 if (frame_reg == ARMREG_SP)
403 else if (frame_reg == ARMREG_FP)
405 #elif defined(TARGET_S390X)
406 if (frame_reg == S390_SP)
408 else if (frame_reg == S390_FP)
413 g_assert_not_reached ();
418 decode_frame_reg (int encoded)
423 else if (encoded == 1)
425 #elif defined(TARGET_X86)
428 else if (encoded == 1)
430 #elif defined(TARGET_ARM)
433 else if (encoded == 1)
435 #elif defined(TARGET_S390X)
438 else if (encoded == 1)
443 g_assert_not_reached ();
449 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15, AMD64_RDI, AMD64_RSI };
451 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15 };
453 #elif defined(TARGET_X86)
454 static int callee_saved_regs [] = { X86_EBX, X86_ESI, X86_EDI };
455 #elif defined(TARGET_ARM)
456 static int callee_saved_regs [] = { ARMREG_V1, ARMREG_V2, ARMREG_V3, ARMREG_V4, ARMREG_V5, ARMREG_V7, ARMREG_FP };
457 #elif defined(TARGET_ARM64)
459 static int callee_saved_regs [] = { };
460 #elif defined(TARGET_S390X)
461 static int callee_saved_regs [] = { s390_r6, s390_r7, s390_r8, s390_r9, s390_r10, s390_r11, s390_r12, s390_r13, s390_r14 };
462 #elif defined(TARGET_POWERPC64) && _CALL_ELF == 2
463 static int callee_saved_regs [] = {
464 ppc_r13, ppc_r14, ppc_r15, ppc_r16,
465 ppc_r17, ppc_r18, ppc_r19, ppc_r20,
466 ppc_r21, ppc_r22, ppc_r23, ppc_r24,
467 ppc_r25, ppc_r26, ppc_r27, ppc_r28,
468 ppc_r29, ppc_r30, ppc_r31 };
469 #elif defined(TARGET_POWERPC)
470 static int callee_saved_regs [] = { ppc_r6, ppc_r7, ppc_r8, ppc_r9, ppc_r10, ppc_r11, ppc_r12, ppc_r13, ppc_r14 };
474 encode_regmask (guint32 regmask)
480 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
481 if (regmask & (1 << callee_saved_regs [i])) {
483 regmask -= (1 << callee_saved_regs [i]);
486 g_assert (regmask == 0);
491 decode_regmask (guint32 regmask)
497 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i)
498 if (regmask & (1 << i))
499 res |= (1 << callee_saved_regs [i]);
506 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
509 encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
513 encode_sleb128 (map->start_offset / SIZEOF_SLOT, buf, &buf);
514 encode_sleb128 (map->end_offset / SIZEOF_SLOT, buf, &buf);
515 encode_sleb128 (map->map_offset / SIZEOF_SLOT, buf, &buf);
516 encode_uleb128 (map->nslots, buf, &buf);
517 g_assert (map->callsite_entry_size <= 4);
518 freg = encode_frame_reg (map->frame_reg);
520 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);
521 encode_uleb128 (flags, buf, &buf);
522 encode_uleb128 (encode_regmask (map->used_int_regs), buf, &buf);
523 if (map->has_ref_regs)
524 encode_uleb128 (encode_regmask (map->reg_ref_mask), buf, &buf);
525 if (map->has_pin_regs)
526 encode_uleb128 (encode_regmask (map->reg_pin_mask), buf, &buf);
527 encode_uleb128 (map->ncallsites, buf, &buf);
535 * Decode the encoded GC map representation in BUF and store the result into MAP.
538 decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
541 int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
544 map->start_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
545 map->end_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
546 map->map_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
547 map->nslots = decode_uleb128 (buf, &buf);
548 flags = decode_uleb128 (buf, &buf);
549 map->has_ref_slots = (flags & 1) ? 1 : 0;
550 map->has_pin_slots = (flags & 2) ? 1 : 0;
551 map->has_ref_regs = (flags & 4) ? 1 : 0;
552 map->has_pin_regs = (flags & 8) ? 1 : 0;
553 map->callsite_entry_size = ((flags >> 4) & 0x3) + 1;
555 map->frame_reg = decode_frame_reg (freg);
556 map->used_int_regs = decode_regmask (decode_uleb128 (buf, &buf));
557 if (map->has_ref_regs) {
558 map->reg_ref_mask = decode_regmask (decode_uleb128 (buf, &buf));
560 for (i = 0; i < NREGS; ++i)
561 if (map->reg_ref_mask & (1 << i))
565 if (map->has_pin_regs) {
566 map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
568 for (i = 0; i < NREGS; ++i)
569 if (map->reg_pin_mask & (1 << i))
573 map->ncallsites = decode_uleb128 (buf, &buf);
575 stack_bitmap_size = (ALIGN_TO (map->nslots, 8) / 8) * map->ncallsites;
576 reg_ref_bitmap_size = (ALIGN_TO (map->nref_regs, 8) / 8) * map->ncallsites;
577 reg_pin_bitmap_size = (ALIGN_TO (map->npin_regs, 8) / 8) * map->ncallsites;
579 map->stack_ref_bitmap_offset = offset;
580 if (map->has_ref_slots)
581 offset += stack_bitmap_size;
582 map->stack_pin_bitmap_offset = offset;
583 if (map->has_pin_slots)
584 offset += stack_bitmap_size;
585 map->reg_ref_bitmap_offset = offset;
586 if (map->has_ref_regs)
587 offset += reg_ref_bitmap_size;
588 map->reg_pin_bitmap_offset = offset;
589 if (map->has_pin_regs)
590 offset += reg_pin_bitmap_size;
596 thread_attach_func (void)
600 tls = g_new0 (TlsData, 1);
601 tls->tid = mono_native_thread_id_get ();
602 tls->info = mono_thread_info_current ();
603 stats.tlsdata_size += sizeof (TlsData);
609 thread_detach_func (gpointer user_data)
611 TlsData *tls = user_data;
617 thread_suspend_func (gpointer user_data, void *sigctx, MonoContext *ctx)
619 TlsData *tls = user_data;
622 /* Happens during startup */
626 if (tls->tid != mono_native_thread_id_get ()) {
627 /* Happens on osx because threads are not suspended using signals */
632 g_assert (tls->info);
636 res = mono_thread_state_init_from_handle (&tls->unwind_state, tls->info);
639 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF] = mono_get_lmf ();
641 #ifdef MONO_ARCH_HAVE_SIGCTX_TO_MONOCTX
642 mono_sigctx_to_monoctx (sigctx, &tls->unwind_state.ctx);
643 tls->unwind_state.valid = TRUE;
645 tls->unwind_state.valid = FALSE;
648 memcpy (&tls->unwind_state.ctx, ctx, sizeof (MonoContext));
649 tls->unwind_state.valid = TRUE;
651 tls->unwind_state.valid = FALSE;
653 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_JIT_TLS] = mono_tls_get_jit_tls ();
654 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN] = mono_domain_get ();
657 if (!tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN]) {
658 /* Happens during startup */
659 tls->unwind_state.valid = FALSE;
664 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
667 set_bit (guint8 *bitmap, int width, int y, int x)
669 bitmap [(width * y) + (x / 8)] |= (1 << (x % 8));
673 clear_bit (guint8 *bitmap, int width, int y, int x)
675 bitmap [(width * y) + (x / 8)] &= ~(1 << (x % 8));
679 get_bit (guint8 *bitmap, int width, int y, int x)
681 return bitmap [(width * y) + (x / 8)] & (1 << (x % 8));
685 slot_type_to_string (GCSlotType type)
695 g_assert_not_reached ();
700 static inline mgreg_t
701 get_frame_pointer (MonoContext *ctx, int frame_reg)
703 #if defined(TARGET_AMD64)
704 if (frame_reg == AMD64_RSP)
706 else if (frame_reg == AMD64_RBP)
708 #elif defined(TARGET_X86)
709 if (frame_reg == X86_ESP)
711 else if (frame_reg == X86_EBP)
713 #elif defined(TARGET_ARM)
714 if (frame_reg == ARMREG_SP)
715 return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
716 else if (frame_reg == ARMREG_FP)
717 return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
718 #elif defined(TARGET_S390X)
719 if (frame_reg == S390_SP)
720 return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
721 else if (frame_reg == S390_FP)
722 return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
724 g_assert_not_reached ();
729 * conservatively_pass:
731 * Mark a thread stack conservatively and collect information needed by the precise pass.
734 conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
738 MonoContext ctx, new_ctx;
741 gboolean last = TRUE;
745 guint8* fp, *p, *real_frame_start, *frame_start, *frame_end;
746 int i, pc_offset, cindex, bitmap_width;
747 int scanned = 0, scanned_precisely, scanned_conservatively, scanned_registers;
749 StackFrameInfo frame;
750 mgreg_t *reg_locations [MONO_MAX_IREGS];
751 mgreg_t *new_reg_locations [MONO_MAX_IREGS];
754 guint32 precise_regmask;
758 tls->ref_to_track = NULL;
761 /* tls == NULL can happen during startup */
762 if (mono_thread_internal_current () == NULL || !tls) {
763 mono_gc_conservatively_scan_area (stack_start, stack_end);
764 stats.scanned_stacks += stack_end - stack_start;
768 lmf = tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF];
771 /* Number of bytes scanned based on GC map data */
773 /* Number of bytes scanned precisely based on GC map data */
774 scanned_precisely = 0;
775 /* Number of bytes scanned conservatively based on GC map data */
776 scanned_conservatively = 0;
777 /* Number of bytes scanned conservatively in register save areas */
778 scanned_registers = 0;
780 /* This is one past the last address which we have scanned */
781 stack_limit = stack_start;
783 if (!tls->unwind_state.valid)
784 memset (&new_ctx, 0, sizeof (ctx));
786 memcpy (&new_ctx, &tls->unwind_state.ctx, sizeof (MonoContext));
788 memset (reg_locations, 0, sizeof (reg_locations));
789 memset (new_reg_locations, 0, sizeof (new_reg_locations));
792 if (!tls->unwind_state.valid)
795 memcpy (&ctx, &new_ctx, sizeof (ctx));
797 for (i = 0; i < MONO_MAX_IREGS; ++i) {
798 if (new_reg_locations [i]) {
800 * If the current frame saves the register, it means it might modify its
801 * value, thus the old location might not contain the same value, so
802 * we have to mark it conservatively.
804 if (reg_locations [i]) {
805 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
806 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
807 scanned_registers += SIZEOF_SLOT;
810 reg_locations [i] = new_reg_locations [i];
812 DEBUG (fprintf (logfile, "\treg %s is now at location %p.\n", mono_arch_regname (i), reg_locations [i]));
816 g_assert ((mgreg_t)stack_limit % SIZEOF_SLOT == 0);
818 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);
824 // FIXME: For skipped frames, scan the param area of the parent frame conservatively ?
825 // FIXME: trampolines
827 if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
829 * These frames are problematic for several reasons:
830 * - they are unwound through an LMF, and we have no precise register tracking for those.
831 * - the LMF might not contain a precise ip, so we can't compute the call site.
832 * - the LMF only unwinds to the wrapper frame, so we get these methods twice.
834 DEBUG (fprintf (logfile, "Mark(0): <Managed-to-native transition>\n"));
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 reg_locations [i] = NULL;
842 new_reg_locations [i] = NULL;
849 method = jinfo_get_method (ji);
853 /* The last frame can be in any state so mark conservatively */
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));
858 DEBUG (fprintf (logfile, "\t <Last frame>\n"));
861 * new_reg_locations is not precise when a method is interrupted during its epilog, so clear it.
863 for (i = 0; i < MONO_MAX_IREGS; ++i) {
864 if (reg_locations [i]) {
865 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
866 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
867 scanned_registers += SIZEOF_SLOT;
869 if (new_reg_locations [i]) {
870 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), new_reg_locations [i]));
871 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
872 scanned_registers += SIZEOF_SLOT;
874 reg_locations [i] = NULL;
875 new_reg_locations [i] = NULL;
880 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
882 /* These frames are very problematic */
883 if (method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
884 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));
885 DEBUG (fprintf (logfile, "\tSkip.\n"));
889 /* All the other frames are at a call site */
891 if (tls->nframes == MAX_FRAMES) {
893 * Can't save information since the array is full. So scan the rest of the
894 * stack conservatively.
896 DEBUG (fprintf (logfile, "Mark (0): Frame stack full.\n"));
900 /* Scan the frame of this method */
903 * A frame contains the following:
908 * - localloc-ed memory
910 g_assert (pc_offset >= 0);
915 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));
916 DEBUG (fprintf (logfile, "\tNo GC Map.\n"));
920 /* The embedded callsite table requires this */
921 g_assert (((mgreg_t)emap % 4) == 0);
924 * Debugging aid to control the number of frames scanned precisely
926 if (!precise_frame_limit_inited) {
927 if (g_getenv ("MONO_PRECISE_COUNT"))
928 precise_frame_limit = atoi (g_getenv ("MONO_PRECISE_COUNT"));
929 precise_frame_limit_inited = TRUE;
932 if (precise_frame_limit != -1) {
933 if (precise_frame_count [FALSE] == precise_frame_limit)
934 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (method, TRUE));
935 if (precise_frame_count [FALSE] > precise_frame_limit)
938 precise_frame_count [FALSE] ++;
940 /* Decode the encoded GC map */
942 memset (map, 0, sizeof (GCMap));
943 decode_gc_map (&emap->encoded [0], map, &p);
944 p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
945 map->callsites.offsets8 = p;
946 p += map->callsite_entry_size * map->ncallsites;
949 fp = (guint8*)get_frame_pointer (&ctx, map->frame_reg);
951 real_frame_start = fp + map->start_offset;
952 frame_start = fp + map->start_offset + map->map_offset;
953 frame_end = fp + map->end_offset;
955 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));
957 /* Find the callsite index */
958 if (map->callsite_entry_size == 1) {
959 for (i = 0; i < map->ncallsites; ++i)
960 /* ip points inside the call instruction */
961 if (map->callsites.offsets8 [i] == pc_offset + 1)
963 } else if (map->callsite_entry_size == 2) {
964 // FIXME: Use a binary search
965 for (i = 0; i < map->ncallsites; ++i)
966 /* ip points inside the call instruction */
967 if (map->callsites.offsets16 [i] == pc_offset + 1)
970 // FIXME: Use a binary search
971 for (i = 0; i < map->ncallsites; ++i)
972 /* ip points inside the call instruction */
973 if (map->callsites.offsets32 [i] == pc_offset + 1)
976 if (i == map->ncallsites) {
977 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (method, TRUE));
978 g_assert_not_reached ();
983 * This is not neccessary true on x86 because frames have a different size at each
986 //g_assert (real_frame_start >= stack_limit);
988 if (real_frame_start > stack_limit) {
989 /* This scans the previously skipped frames as well */
990 DEBUG (fprintf (logfile, "\tscan area %p-%p (%d).\n", stack_limit, real_frame_start, (int)(real_frame_start - stack_limit)));
991 mono_gc_conservatively_scan_area (stack_limit, real_frame_start);
992 stats.scanned_other += real_frame_start - stack_limit;
995 /* Mark stack slots */
996 if (map->has_pin_slots) {
997 int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
998 guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
1003 for (i = 0; i < map->nslots; ++i) {
1004 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
1006 DEBUG (fprintf (logfile, "\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
1007 mono_gc_conservatively_scan_area (p, p + SIZEOF_SLOT);
1008 scanned_conservatively += SIZEOF_SLOT;
1010 scanned_precisely += SIZEOF_SLOT;
1015 scanned_precisely += (map->nslots * SIZEOF_SLOT);
1018 /* The area outside of start-end is NOREF */
1019 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * SIZEOF_SLOT);
1021 /* Mark registers */
1022 precise_regmask = map->used_int_regs | (1 << map->frame_reg);
1023 if (map->has_pin_regs) {
1024 int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
1025 guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
1027 for (i = 0; i < NREGS; ++i) {
1028 if (!(map->used_int_regs & (1 << i)))
1031 if (!(map->reg_pin_mask & (1 << i)))
1034 if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
1035 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
1036 precise_regmask &= ~(1 << i);
1042 scanned += map->end_offset - map->start_offset;
1044 g_assert (scanned == scanned_precisely + scanned_conservatively);
1046 stack_limit = frame_end;
1048 /* Save information for the precise pass */
1049 fi = &tls->frames [tls->nframes];
1050 fi->nslots = map->nslots;
1051 bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
1052 if (map->has_ref_slots)
1053 fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
1056 fi->frame_start_offset = frame_start - stack_start;
1057 fi->nreg_locations = 0;
1058 DEBUG_PRECISE (fi->ji = ji);
1059 DEBUG_PRECISE (fi->fp = fp);
1061 if (map->has_ref_regs) {
1062 int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
1063 guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
1065 for (i = 0; i < NREGS; ++i) {
1066 if (!(map->reg_ref_mask & (1 << i)))
1069 if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
1070 DEBUG_PRECISE (fi->regs [fi->nreg_locations] = i);
1071 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
1072 fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
1073 fi->nreg_locations ++;
1080 * Clear locations of precisely tracked registers.
1082 if (precise_regmask) {
1083 for (i = 0; i < NREGS; ++i) {
1084 if (precise_regmask & (1 << i)) {
1086 * The method uses this register, and we have precise info for it.
1087 * This means the location will be scanned precisely.
1088 * Tell the code at the beginning of the loop that this location is
1091 if (reg_locations [i])
1092 DEBUG (fprintf (logfile, "\treg %s at location %p (==%p) is precise.\n", mono_arch_regname (i), reg_locations [i], (gpointer)*reg_locations [i]));
1093 reg_locations [i] = NULL;
1101 /* Scan the remaining register save locations */
1102 for (i = 0; i < MONO_MAX_IREGS; ++i) {
1103 if (reg_locations [i]) {
1104 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", reg_locations [i]));
1105 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
1106 scanned_registers += SIZEOF_SLOT;
1108 if (new_reg_locations [i]) {
1109 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", new_reg_locations [i]));
1110 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
1111 scanned_registers += SIZEOF_SLOT;
1115 if (stack_limit < stack_end) {
1116 DEBUG (fprintf (logfile, "\tscan remaining stack %p-%p (%d).\n", stack_limit, stack_end, (int)(stack_end - stack_limit)));
1117 mono_gc_conservatively_scan_area (stack_limit, stack_end);
1118 stats.scanned_native += stack_end - stack_limit;
1121 DEBUG (fprintf (logfile, "Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
1123 stats.scanned_stacks += stack_end - stack_start;
1124 stats.scanned += scanned;
1125 stats.scanned_precisely += scanned_precisely;
1126 stats.scanned_conservatively += scanned_conservatively;
1127 stats.scanned_registers += scanned_registers;
1129 //mono_gc_conservatively_scan_area (stack_start, stack_end);
1135 * Mark a thread stack precisely based on information saved during the conservative
1139 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end, void *gc_data)
1143 guint8 *frame_start;
1148 if (!tls->unwind_state.valid)
1151 for (findex = 0; findex < tls->nframes; findex ++) {
1152 /* Load information saved by the !precise pass */
1153 fi = &tls->frames [findex];
1154 frame_start = stack_start + fi->frame_start_offset;
1156 DEBUG (char *fname = mono_method_full_name (jinfo_get_method (fi->ji), TRUE); fprintf (logfile, "Mark(1): %s\n", fname); g_free (fname));
1159 * FIXME: Add a function to mark using a bitmap, to avoid doing a
1160 * call for each object.
1163 /* Mark stack slots */
1165 guint8 *ref_bitmap = fi->bitmap;
1168 for (i = 0; i < fi->nslots; ++i) {
1169 MonoObject **ptr = (MonoObject**)(frame_start + (i * SIZEOF_SLOT));
1171 live = ref_bitmap [i / 8] & (1 << (i % 8));
1174 MonoObject *obj = *ptr;
1176 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1177 *ptr = mono_gc_scan_object (obj, gc_data);
1178 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1180 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1185 * This is disabled because the pointer takes up a lot of space.
1186 * Stack slots might be shared between ref and non-ref variables ?
1188 if (map->ref_slots [i / 8] & (1 << (i % 8))) {
1189 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));
1191 * Fail fast if the live range is incorrect, and
1192 * the JITted code tries to access this object
1201 /* Mark registers */
1204 * Registers are different from stack slots, they have no address where they
1205 * are stored. Instead, some frame below this frame in the stack saves them
1206 * in its prolog to the stack. We can mark this location precisely.
1208 for (i = 0; i < fi->nreg_locations; ++i) {
1210 * reg_locations [i] contains the address of the stack slot where
1211 * a reg was last saved, so mark that slot.
1213 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1214 MonoObject *obj = *ptr;
1217 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
1218 *ptr = mono_gc_scan_object (obj, gc_data);
1219 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1221 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p\n", mono_arch_regname (fi->regs [i]), ptr, obj));
1227 * Debugging aid to check for missed refs.
1229 if (tls->ref_to_track) {
1232 for (p = (mgreg_t*)stack_start; p < (mgreg_t*)stack_end; ++p)
1233 if (*p == (mgreg_t)tls->ref_to_track)
1234 printf ("REF AT %p.\n", p);
1241 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1242 * call, and TRUE at the second. USER_DATA points to a TlsData
1243 * structure filled up by thread_suspend_func.
1246 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise, void *gc_data)
1248 TlsData *tls = user_data;
1250 DEBUG (fprintf (logfile, "****************************************\n"));
1251 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));
1252 DEBUG (fprintf (logfile, "****************************************\n"));
1255 conservative_pass (tls, stack_start, stack_end);
1257 precise_pass (tls, stack_start, stack_end, gc_data);
1263 mini_gc_init_gc_map (MonoCompile *cfg)
1265 if (COMPILE_LLVM (cfg))
1268 if (!mono_gc_is_moving ())
1271 if (cfg->compile_aot) {
1272 if (!enable_gc_maps_for_aot)
1274 } else if (!mono_gc_precise_stack_mark_enabled ())
1278 /* Debugging support */
1280 static int precise_count;
1283 if (g_getenv ("MONO_GCMAP_COUNT")) {
1284 if (precise_count == atoi (g_getenv ("MONO_GCMAP_COUNT")))
1285 printf ("LAST: %s\n", mono_method_full_name (cfg->method, TRUE));
1286 if (precise_count > atoi (g_getenv ("MONO_GCMAP_COUNT")))
1292 cfg->compute_gc_maps = TRUE;
1294 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1298 * mini_gc_set_slot_type_from_fp:
1300 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1301 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1302 * need to call this function for those slots.
1305 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1307 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1309 if (!cfg->compute_gc_maps)
1312 g_assert (slot_offset % SIZEOF_SLOT == 0);
1314 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1318 * mini_gc_set_slot_type_from_cfa:
1320 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1321 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1322 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1323 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1324 * call this function for those slots.
1327 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1329 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1330 int slot = - (slot_offset / SIZEOF_SLOT);
1332 if (!cfg->compute_gc_maps)
1335 g_assert (slot_offset <= 0);
1336 g_assert (slot_offset % SIZEOF_SLOT == 0);
1338 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1342 fp_offset_to_slot (MonoCompile *cfg, int offset)
1344 MonoCompileGC *gcfg = cfg->gc_info;
1346 return (offset - gcfg->min_offset) / SIZEOF_SLOT;
1350 slot_to_fp_offset (MonoCompile *cfg, int slot)
1352 MonoCompileGC *gcfg = cfg->gc_info;
1354 return (slot * SIZEOF_SLOT) + gcfg->min_offset;
1357 static inline MONO_ALWAYS_INLINE void
1358 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1360 g_assert (slot >= 0 && slot < gcfg->nslots);
1362 if (type == SLOT_PIN) {
1363 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1364 set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1365 } else if (type == SLOT_REF) {
1366 set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1367 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1368 } else if (type == SLOT_NOREF) {
1369 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1370 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1375 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1378 guint8 *ref_bitmap, *pin_bitmap;
1383 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1384 set_slot (gcfg, slot, cindex, type);
1386 ref_bitmap = gcfg->stack_ref_bitmap;
1387 pin_bitmap = gcfg->stack_pin_bitmap;
1388 width = gcfg->stack_bitmap_width;
1391 if (type == SLOT_PIN) {
1392 memset (ref_bitmap + pos, 0, width);
1393 memset (pin_bitmap + pos, 0xff, width);
1394 } else if (type == SLOT_REF) {
1395 memset (ref_bitmap + pos, 0xff, width);
1396 memset (pin_bitmap + pos, 0, width);
1397 } else if (type == SLOT_NOREF) {
1398 memset (ref_bitmap + pos, 0, width);
1399 memset (pin_bitmap + pos, 0, width);
1404 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1408 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1409 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1410 if (callsite_offset >= from && callsite_offset < to)
1411 set_slot (gcfg, slot, cindex, type);
1416 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1418 g_assert (slot >= 0 && slot < gcfg->nregs);
1420 if (type == SLOT_PIN) {
1421 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1422 set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1423 } else if (type == SLOT_REF) {
1424 set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1425 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1426 } else if (type == SLOT_NOREF) {
1427 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1428 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1433 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1437 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1438 set_reg_slot (gcfg, slot, cindex, type);
1442 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1446 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1447 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1448 if (callsite_offset >= from && callsite_offset < to)
1449 set_reg_slot (gcfg, slot, cindex, type);
1454 process_spill_slots (MonoCompile *cfg)
1456 MonoCompileGC *gcfg = cfg->gc_info;
1461 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1462 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1463 for (l = bb->spill_slot_defs; l; l = l->next) {
1464 MonoInst *def = l->data;
1465 int spill_slot = def->inst_c0;
1466 int bank = def->inst_c1;
1467 int offset = cfg->spill_info [bank][spill_slot].offset;
1468 int slot = fp_offset_to_slot (cfg, offset);
1470 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1471 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1475 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1476 for (l = bb->spill_slot_defs; l; l = l->next) {
1477 MonoInst *def = l->data;
1478 int spill_slot = def->inst_c0;
1479 int bank = def->inst_c1;
1480 int offset = cfg->spill_info [bank][spill_slot].offset;
1481 int slot = fp_offset_to_slot (cfg, offset);
1484 if (bank == MONO_REG_INT_MP)
1490 * Extend the live interval for the GC tracked spill slots
1491 * defined in this bblock.
1492 * FIXME: This is not needed.
1494 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1496 if (cfg->verbose_level > 1)
1497 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
1501 /* Set fp spill slots to NOREF */
1502 for (i = 0; i < cfg->spill_info_len [MONO_REG_DOUBLE]; ++i) {
1503 int offset = cfg->spill_info [MONO_REG_DOUBLE][i].offset;
1509 slot = fp_offset_to_slot (cfg, offset);
1511 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1513 if (cfg->verbose_level > 1)
1514 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1517 /* Set int spill slots to NOREF */
1518 for (i = 0; i < cfg->spill_info_len [MONO_REG_INT]; ++i) {
1519 int offset = cfg->spill_info [MONO_REG_INT][i].offset;
1525 slot = fp_offset_to_slot (cfg, offset);
1527 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1528 if (cfg->verbose_level > 1)
1529 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1534 * process_other_slots:
1536 * Process stack slots registered using mini_gc_set_slot_type_... ().
1539 process_other_slots (MonoCompile *cfg)
1541 MonoCompileGC *gcfg = cfg->gc_info;
1544 /* Relative to the CFA */
1545 for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
1546 guint data = GPOINTER_TO_UINT (l->data);
1547 int cfa_slot = data >> 16;
1548 GCSlotType type = data & 0xff;
1552 * Map the cfa relative slot to an fp relative slot.
1553 * slot_addr == cfa - <cfa_slot>*4/8
1554 * fp + cfa_offset == cfa
1555 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1557 slot = (cfg->cfa_offset / SIZEOF_SLOT) - cfa_slot - (gcfg->min_offset / SIZEOF_SLOT);
1559 set_slot_everywhere (gcfg, slot, type);
1561 if (cfg->verbose_level > 1) {
1562 int fp_offset = slot_to_fp_offset (cfg, slot);
1563 if (type == SLOT_NOREF)
1564 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));
1568 /* Relative to the FP */
1569 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1570 gint data = GPOINTER_TO_INT (l->data);
1571 int offset = data >> 16;
1572 GCSlotType type = data & 0xff;
1575 slot = fp_offset_to_slot (cfg, offset);
1577 set_slot_everywhere (gcfg, slot, type);
1579 /* Liveness for these slots is handled by process_spill_slots () */
1581 if (cfg->verbose_level > 1) {
1582 if (type == SLOT_REF)
1583 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
1584 else if (type == SLOT_NOREF)
1585 printf ("\tnoref slot at 0x%x(fp) (slot = %d)\n", offset, slot);
1591 get_vtype_bitmap (MonoType *t, int *numbits)
1593 MonoClass *klass = mono_class_from_mono_type (t);
1595 if (klass->generic_container || mono_class_is_open_constructed_type (t)) {
1596 /* FIXME: Generic sharing */
1599 mono_class_compute_gc_descriptor (klass);
1601 return mono_gc_get_bitmap_for_descr (klass->gc_descr, numbits);
1605 static inline const char*
1606 get_offset_sign (int offset)
1608 return offset < 0 ? "-" : "+";
1612 get_offset_val (int offset)
1614 return offset < 0 ? (- offset) : offset;
1618 process_variables (MonoCompile *cfg)
1620 MonoCompileGC *gcfg = cfg->gc_info;
1621 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1622 int i, locals_min_slot, locals_max_slot, cindex;
1626 int locals_min_offset = gcfg->locals_min_offset;
1627 int locals_max_offset = gcfg->locals_max_offset;
1629 /* Slots for locals are NOREF by default */
1630 locals_min_slot = (locals_min_offset - gcfg->min_offset) / SIZEOF_SLOT;
1631 locals_max_slot = (locals_max_offset - gcfg->min_offset) / SIZEOF_SLOT;
1632 for (i = locals_min_slot; i < locals_max_slot; ++i) {
1633 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1637 * Compute the offset where variables are initialized in the first bblock, if any.
1639 pc_offsets = g_new0 (int, cfg->next_vreg);
1641 bb = cfg->bb_entry->next_bb;
1642 MONO_BB_FOR_EACH_INS (bb, tmp) {
1643 if (tmp->opcode == OP_GC_LIVENESS_DEF) {
1644 int vreg = tmp->inst_c1;
1645 if (pc_offsets [vreg] == 0) {
1646 g_assert (tmp->backend.pc_offset > 0);
1647 pc_offsets [vreg] = tmp->backend.pc_offset;
1653 * Stack slots holding arguments are initialized in the prolog.
1654 * This means we can treat them alive for the whole method.
1656 for (i = 0; i < cfg->num_varinfo; i++) {
1657 MonoInst *ins = cfg->varinfo [i];
1658 MonoType *t = ins->inst_vtype;
1661 gboolean byref, is_this = FALSE;
1662 gboolean is_arg = i < cfg->locals_start;
1664 if (ins == cfg->ret) {
1665 if (!(ins->opcode == OP_REGOFFSET && MONO_TYPE_ISSTRUCT (t)))
1669 vmv = MONO_VARINFO (cfg, i);
1671 /* For some reason, 'this' is byref */
1672 if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
1673 t = &cfg->method->klass->byval_arg;
1679 if (ins->opcode == OP_REGVAR) {
1681 GCSlotType slot_type;
1683 t = mini_get_underlying_type (t);
1686 g_assert (hreg < MONO_MAX_IREGS);
1689 slot_type = SLOT_PIN;
1691 slot_type = mini_type_is_reference (t) ? SLOT_REF : SLOT_NOREF;
1693 if (slot_type == SLOT_PIN) {
1694 /* These have no live interval, be conservative */
1695 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1698 * Unlike variables allocated to the stack, we generate liveness info
1699 * for noref vars in registers in mono_spill_global_vars (), because
1700 * knowing that a register doesn't contain a ref allows us to mark its save
1701 * locations precisely.
1703 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1704 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1705 set_reg_slot (gcfg, hreg, cindex, slot_type);
1708 if (cfg->verbose_level > 1) {
1709 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
1715 if (ins->opcode != OP_REGOFFSET)
1718 if (ins->inst_offset % SIZEOF_SLOT != 0)
1721 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1723 if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
1724 /* These do not get stored in the prolog */
1725 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1727 if (cfg->verbose_level > 1) {
1728 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));
1733 if (MONO_TYPE_ISSTRUCT (t)) {
1735 gsize *bitmap = NULL;
1736 gboolean pin = FALSE;
1740 if (ins->backend.is_pinvoke)
1741 size = mono_class_native_size (ins->klass, NULL);
1743 size = mono_class_value_size (ins->klass, NULL);
1744 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
1746 if (cfg->verbose_level > 1)
1747 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));
1749 if (!ins->klass->has_references) {
1751 for (j = 0; j < size_in_slots; ++j)
1752 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1757 bitmap = get_vtype_bitmap (t, &numbits);
1762 * Most vtypes are marked volatile because of the LDADDR instructions,
1763 * and they have no liveness information since they are decomposed
1764 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1765 * them during VZERO decomposition.
1768 if (!pc_offsets [vmv->vreg])
1771 if (ins->backend.is_pinvoke)
1776 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1777 if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
1778 for (j = 0; j < numbits; ++j) {
1779 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1780 /* The descriptor is for the boxed object */
1781 set_slot (gcfg, (pos + j - (sizeof (MonoObject) / SIZEOF_SLOT)), cindex, pin ? SLOT_PIN : SLOT_REF);
1787 if (cfg->verbose_level > 1) {
1788 for (j = 0; j < numbits; ++j) {
1789 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD)))
1790 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)));
1794 if (cfg->verbose_level > 1)
1795 printf ("\t\tpinned\n");
1796 for (j = 0; j < size_in_slots; ++j) {
1797 set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
1806 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1807 /* Vret addr etc. */
1812 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1814 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1815 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1816 set_slot (gcfg, pos, cindex, SLOT_PIN);
1818 if (cfg->verbose_level > 1)
1819 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));
1824 * This is currently disabled, but could be enabled to debug crashes.
1827 if (t->type == MONO_TYPE_I) {
1829 * Variables created in mono_handle_global_vregs have type I, but they
1830 * could hold GC refs since the vregs they were created from might not been
1831 * marked as holding a GC ref. So be conservative.
1833 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1838 t = mini_get_underlying_type (t);
1840 if (!mini_type_is_reference (t)) {
1841 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1842 if (cfg->verbose_level > 1)
1843 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));
1844 if (!t->byref && sizeof (mgreg_t) == 4 && (t->type == MONO_TYPE_I8 || t->type == MONO_TYPE_U8 || t->type == MONO_TYPE_R8)) {
1845 set_slot_everywhere (gcfg, pos + 1, SLOT_NOREF);
1846 if (cfg->verbose_level > 1)
1847 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));
1852 /* 'this' is marked INDIRECT for gshared methods */
1853 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1855 * For volatile variables, treat them alive from the point they are
1856 * initialized in the first bblock until the end of the method.
1859 set_slot_everywhere (gcfg, pos, SLOT_REF);
1860 } else if (pc_offsets [vmv->vreg]) {
1861 set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
1862 set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
1864 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1866 if (cfg->verbose_level > 1)
1867 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));
1872 /* Live for the whole method */
1873 set_slot_everywhere (gcfg, pos, SLOT_REF);
1875 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1876 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1877 set_slot (gcfg, pos, cindex, SLOT_REF);
1880 if (cfg->verbose_level > 1) {
1881 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));
1885 g_free (pc_offsets);
1889 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1892 * Convert a sp relative offset to a slot index. This is
1893 * platform specific.
1896 /* fp = sp + offset */
1897 g_assert (cfg->frame_reg == AMD64_RBP);
1898 return (- cfg->arch.sp_fp_offset + sp_offset);
1899 #elif defined(TARGET_X86)
1900 /* The offset is computed from the sp at the start of the call sequence */
1901 g_assert (cfg->frame_reg == X86_EBP);
1902 #ifdef MONO_X86_NO_PUSHES
1903 return (- cfg->arch.sp_fp_offset + sp_offset);
1905 return (- cfg->arch.sp_fp_offset - sp_offset);
1914 process_param_area_slots (MonoCompile *cfg)
1916 MonoCompileGC *gcfg = cfg->gc_info;
1921 * These slots are used for passing parameters during calls. They are sp relative, not
1922 * fp relative, so they are harder to handle.
1924 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1925 /* The distance between fp and sp is not constant */
1928 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1930 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1931 GCCallSite *callsite = gcfg->callsites [cindex];
1934 for (l = callsite->param_slots; l; l = l->next) {
1935 MonoInst *def = l->data;
1936 MonoType *t = def->inst_vtype;
1937 int sp_offset = def->inst_offset;
1938 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1939 int slot = fp_offset_to_slot (cfg, fp_offset);
1943 if (MONO_TYPE_ISSTRUCT (t)) {
1944 size = mini_type_stack_size_full (t, &align, FALSE);
1946 size = sizeof (mgreg_t);
1949 for (i = 0; i < size / sizeof (mgreg_t); ++i) {
1950 g_assert (slot + i >= 0 && slot + i < gcfg->nslots);
1951 is_param [slot + i] = TRUE;
1956 /* All param area slots are noref by default */
1957 for (i = 0; i < gcfg->nslots; ++i) {
1959 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1963 * We treat param area slots as being part of the callee's frame, to be able to handle tail calls which overwrite
1964 * the argument area of the caller.
1969 process_finally_clauses (MonoCompile *cfg)
1971 MonoCompileGC *gcfg = cfg->gc_info;
1972 GCCallSite **callsites;
1974 gboolean has_finally;
1975 int i, j, nslots, nregs;
1977 ncallsites = gcfg->ncallsites;
1978 nslots = gcfg->nslots;
1979 nregs = gcfg->nregs;
1980 callsites = gcfg->callsites;
1983 * The calls to the finally clauses don't show up in the cfg. See
1984 * test_0_liveness_8 ().
1985 * Variables accessed inside the finally clause are already marked VOLATILE by
1986 * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
1987 * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
1989 has_finally = FALSE;
1990 for (i = 0; i < cfg->header->num_clauses; ++i) {
1991 MonoExceptionClause *clause = &cfg->header->clauses [i];
1993 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
1998 if (cfg->verbose_level > 1)
1999 printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
2000 for (j = 0; j < ncallsites; ++j) {
2001 MonoBasicBlock *bb = callsites [j]->bb;
2002 MonoExceptionClause *clause;
2003 gboolean is_in_finally = FALSE;
2005 for (i = 0; i < cfg->header->num_clauses; ++i) {
2006 clause = &cfg->header->clauses [i];
2008 if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
2009 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
2010 is_in_finally = TRUE;
2016 if (is_in_finally) {
2017 for (i = 0; i < nslots; ++i)
2018 set_slot (gcfg, i, j, SLOT_PIN);
2019 for (i = 0; i < nregs; ++i)
2020 set_reg_slot (gcfg, i, j, SLOT_PIN);
2027 compute_frame_size (MonoCompile *cfg)
2029 int i, locals_min_offset, locals_max_offset, cfa_min_offset, cfa_max_offset;
2030 int min_offset, max_offset;
2031 MonoCompileGC *gcfg = cfg->gc_info;
2032 MonoMethodSignature *sig = mono_method_signature (cfg->method);
2035 /* Compute min/max offsets from the fp */
2038 #if defined(TARGET_AMD64) || defined(TARGET_X86) || defined(TARGET_ARM) || defined(TARGET_S390X)
2039 locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, SIZEOF_SLOT);
2040 locals_max_offset = cfg->locals_max_stack_offset;
2042 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
2046 locals_min_offset = ALIGN_TO (locals_min_offset, SIZEOF_SLOT);
2047 locals_max_offset = ALIGN_TO (locals_max_offset, SIZEOF_SLOT);
2049 min_offset = locals_min_offset;
2050 max_offset = locals_max_offset;
2053 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
2054 MonoInst *ins = cfg->args [i];
2056 if (ins->opcode == OP_REGOFFSET) {
2057 int size, size_in_slots;
2058 size = mini_type_stack_size_full (ins->inst_vtype, NULL, ins->backend.is_pinvoke);
2059 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
2061 min_offset = MIN (min_offset, ins->inst_offset);
2062 max_offset = MAX ((int)max_offset, (int)(ins->inst_offset + (size_in_slots * SIZEOF_SLOT)));
2067 g_assert (cfg->frame_reg == cfg->cfa_reg);
2068 g_assert (cfg->cfa_offset > 0);
2070 cfa_max_offset = cfg->cfa_offset;
2072 min_offset = MIN (min_offset, cfa_min_offset);
2073 max_offset = MAX (max_offset, cfa_max_offset);
2075 /* Fp relative slots */
2076 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
2077 gint data = GPOINTER_TO_INT (l->data);
2078 int offset = data >> 16;
2080 min_offset = MIN (min_offset, offset);
2084 if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
2085 int stack_offset = ALIGN_TO (cfg->stack_offset, SIZEOF_SLOT);
2086 min_offset = MIN (min_offset, (-stack_offset));
2089 /* Param area slots */
2091 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
2092 #elif defined(TARGET_X86)
2093 #ifdef MONO_X86_NO_PUSHES
2094 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
2096 min_offset = MIN (min_offset, - (cfg->arch.sp_fp_offset + cfg->arch.param_area_size));
2098 #elif defined(TARGET_ARM)
2100 #elif defined(TARGET_s390X)
2106 gcfg->min_offset = min_offset;
2107 gcfg->max_offset = max_offset;
2108 gcfg->locals_min_offset = locals_min_offset;
2109 gcfg->locals_max_offset = locals_max_offset;
2113 init_gcfg (MonoCompile *cfg)
2115 int i, nregs, nslots;
2116 MonoCompileGC *gcfg = cfg->gc_info;
2117 GCCallSite **callsites;
2126 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2127 ncallsites += g_slist_length (bb->gc_callsites);
2129 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
2131 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2132 for (l = bb->gc_callsites; l; l = l->next)
2133 callsites [i++] = l->data;
2136 /* The callsites should already be ordered by pc offset */
2137 for (i = 1; i < ncallsites; ++i)
2138 g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
2141 * The stack frame looks like this:
2143 * <fp + max_offset> == cfa -> <end of previous frame>
2144 * <other stack slots>
2146 * <other stack slots>
2147 * fp + min_offset ->
2152 if (cfg->verbose_level > 1)
2153 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
2155 nslots = (gcfg->max_offset - gcfg->min_offset) / SIZEOF_SLOT;
2158 gcfg->nslots = nslots;
2159 gcfg->nregs = nregs;
2160 gcfg->callsites = callsites;
2161 gcfg->ncallsites = ncallsites;
2162 gcfg->stack_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
2163 gcfg->reg_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
2164 gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
2165 gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
2166 gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
2167 gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
2169 /* All slots start out as PIN */
2170 memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * nregs);
2171 for (i = 0; i < nregs; ++i) {
2173 * By default, registers are NOREF.
2174 * It is possible for a callee to save them before being defined in this method,
2175 * but the saved value is dead too, so it doesn't need to be marked.
2177 if ((cfg->used_int_regs & (1 << i)))
2178 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
2182 static inline gboolean
2183 has_bit_set (guint8 *bitmap, int width, int slot)
2186 int pos = width * slot;
2188 for (i = 0; i < width; ++i) {
2189 if (bitmap [pos + i])
2196 create_map (MonoCompile *cfg)
2199 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
2201 int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
2202 int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
2204 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
2205 MonoCompileGC *gcfg = cfg->gc_info;
2206 GCCallSite **callsites;
2208 guint8 *bitmap, *bitmaps;
2209 guint32 reg_ref_mask, reg_pin_mask;
2211 ncallsites = gcfg->ncallsites;
2212 nslots = gcfg->nslots;
2213 nregs = gcfg->nregs;
2214 callsites = gcfg->callsites;
2217 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
2218 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
2220 has_ref_slots = FALSE;
2221 has_pin_slots = FALSE;
2224 memset (ntypes, 0, sizeof (ntypes));
2225 for (i = 0; i < nslots; ++i) {
2226 gboolean has_ref = FALSE;
2227 gboolean has_pin = FALSE;
2229 if (has_bit_set (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i))
2231 if (has_bit_set (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i))
2235 has_ref_slots = TRUE;
2237 has_pin_slots = TRUE;
2240 ntypes [SLOT_REF] ++;
2242 ntypes [SLOT_PIN] ++;
2244 ntypes [SLOT_NOREF] ++;
2246 if (has_ref || has_pin) {
2253 start = end = nslots;
2255 g_assert (start != -1);
2256 g_assert (start < end);
2259 has_ref_regs = FALSE;
2260 has_pin_regs = FALSE;
2265 for (i = 0; i < nregs; ++i) {
2266 gboolean has_ref = FALSE;
2267 gboolean has_pin = FALSE;
2269 if (!(cfg->used_int_regs & (1 << i)))
2272 if (has_bit_set (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i))
2274 if (has_bit_set (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i))
2278 reg_ref_mask |= (1 << i);
2279 has_ref_regs = TRUE;
2283 reg_pin_mask |= (1 << i);
2284 has_pin_regs = TRUE;
2289 if (cfg->verbose_level > 1)
2290 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);
2292 /* Create the GC Map */
2294 /* The work bitmaps have one row for each slot, since this is how we access them during construction */
2295 stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
2296 stack_bitmap_size = stack_bitmap_width * ncallsites;
2297 reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
2298 reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
2299 reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
2300 reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
2301 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);
2303 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2305 map->frame_reg = cfg->frame_reg;
2306 map->start_offset = gcfg->min_offset;
2307 map->end_offset = gcfg->min_offset + (nslots * SIZEOF_SLOT);
2308 map->map_offset = start * SIZEOF_SLOT;
2309 map->nslots = end - start;
2310 map->has_ref_slots = has_ref_slots;
2311 map->has_pin_slots = has_pin_slots;
2312 map->has_ref_regs = has_ref_regs;
2313 map->has_pin_regs = has_pin_regs;
2314 g_assert (nregs < 32);
2315 map->used_int_regs = cfg->used_int_regs;
2316 map->reg_ref_mask = reg_ref_mask;
2317 map->reg_pin_mask = reg_pin_mask;
2318 map->nref_regs = nref_regs;
2319 map->npin_regs = npin_regs;
2321 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2324 if (has_ref_slots) {
2325 map->stack_ref_bitmap_offset = bitmaps_offset;
2326 bitmaps_offset += stack_bitmap_size;
2328 bitmap = &bitmaps [map->stack_ref_bitmap_offset];
2329 for (i = 0; i < nslots; ++i) {
2330 for (j = 0; j < ncallsites; ++j) {
2331 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i, j))
2332 set_bit (bitmap, stack_bitmap_width, j, i - start);
2336 if (has_pin_slots) {
2337 map->stack_pin_bitmap_offset = bitmaps_offset;
2338 bitmaps_offset += stack_bitmap_size;
2340 bitmap = &bitmaps [map->stack_pin_bitmap_offset];
2341 for (i = 0; i < nslots; ++i) {
2342 for (j = 0; j < ncallsites; ++j) {
2343 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i, j))
2344 set_bit (bitmap, stack_bitmap_width, j, i - start);
2349 map->reg_ref_bitmap_offset = bitmaps_offset;
2350 bitmaps_offset += reg_ref_bitmap_size;
2352 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2354 for (i = 0; i < nregs; ++i) {
2355 if (reg_ref_mask & (1 << i)) {
2356 for (j = 0; j < ncallsites; ++j) {
2357 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i, j))
2358 set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
2365 map->reg_pin_bitmap_offset = bitmaps_offset;
2366 bitmaps_offset += reg_pin_bitmap_size;
2368 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2370 for (i = 0; i < nregs; ++i) {
2371 if (reg_pin_mask & (1 << i)) {
2372 for (j = 0; j < ncallsites; ++j) {
2373 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i, j))
2374 set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
2382 map->ncallsites = ncallsites;
2383 if (cfg->code_len < 256)
2384 map->callsite_entry_size = 1;
2385 else if (cfg->code_len < 65536)
2386 map->callsite_entry_size = 2;
2388 map->callsite_entry_size = 4;
2390 /* Encode the GC Map */
2398 encode_gc_map (map, buf, &endbuf);
2399 g_assert (endbuf - buf < 256);
2401 encoded_size = endbuf - buf;
2402 alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
2404 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2405 //emap->ref_slots = map->ref_slots;
2407 /* Encoded fixed fields */
2408 p = &emap->encoded [0];
2409 //emap->encoded_size = encoded_size;
2410 memcpy (p, buf, encoded_size);
2413 /* Callsite table */
2414 p = (guint8*)ALIGN_TO ((mgreg_t)p, map->callsite_entry_size);
2415 if (map->callsite_entry_size == 1) {
2416 guint8 *offsets = p;
2417 for (i = 0; i < ncallsites; ++i)
2418 offsets [i] = callsites [i]->pc_offset;
2419 stats.gc_callsites8_size += ncallsites * sizeof (guint8);
2420 } else if (map->callsite_entry_size == 2) {
2421 guint16 *offsets = (guint16*)p;
2422 for (i = 0; i < ncallsites; ++i)
2423 offsets [i] = callsites [i]->pc_offset;
2424 stats.gc_callsites16_size += ncallsites * sizeof (guint16);
2426 guint32 *offsets = (guint32*)p;
2427 for (i = 0; i < ncallsites; ++i)
2428 offsets [i] = callsites [i]->pc_offset;
2429 stats.gc_callsites32_size += ncallsites * sizeof (guint32);
2431 p += ncallsites * map->callsite_entry_size;
2434 memcpy (p, bitmaps, bitmaps_size);
2437 g_assert ((guint8*)p - (guint8*)emap <= alloc_size);
2439 stats.gc_maps_size += alloc_size;
2440 stats.gc_callsites_size += ncallsites * map->callsite_entry_size;
2441 stats.gc_bitmaps_size += bitmaps_size;
2442 stats.gc_map_struct_size += sizeof (GCEncodedMap) + encoded_size;
2444 cfg->jit_info->gc_info = emap;
2446 cfg->gc_map = (guint8*)emap;
2447 cfg->gc_map_size = alloc_size;
2450 stats.all_slots += nslots;
2451 stats.ref_slots += ntypes [SLOT_REF];
2452 stats.noref_slots += ntypes [SLOT_NOREF];
2453 stats.pin_slots += ntypes [SLOT_PIN];
2457 mini_gc_create_gc_map (MonoCompile *cfg)
2459 if (!cfg->compute_gc_maps)
2463 * During marking, all frames except the top frame are at a call site, and we mark the
2464 * top frame conservatively. This means that we only need to compute and record
2465 * GC maps for call sites.
2468 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2469 /* Without liveness info, the live ranges are not precise enough */
2472 mono_analyze_liveness_gc (cfg);
2474 compute_frame_size (cfg);
2478 process_spill_slots (cfg);
2479 process_other_slots (cfg);
2480 process_param_area_slots (cfg);
2481 process_variables (cfg);
2482 process_finally_clauses (cfg);
2487 #endif /* DISABLE_JIT */
2490 parse_debug_options (void)
2495 env = g_getenv ("MONO_GCMAP_DEBUG");
2499 opts = g_strsplit (env, ",", -1);
2500 for (ptr = opts; ptr && *ptr; ptr ++) {
2501 /* No options yet */
2502 fprintf (stderr, "Invalid format for the MONO_GCMAP_DEBUG env variable: '%s'\n", env);
2513 memset (&cb, 0, sizeof (cb));
2514 cb.thread_attach_func = thread_attach_func;
2515 cb.thread_detach_func = thread_detach_func;
2516 cb.thread_suspend_func = thread_suspend_func;
2517 /* Comment this out to disable precise stack marking */
2518 cb.thread_mark_func = thread_mark_func;
2519 cb.get_provenance_func = get_provenance_func;
2520 mono_gc_set_gc_callbacks (&cb);
2522 logfile = mono_gc_get_logfile ();
2524 parse_debug_options ();
2526 mono_counters_register ("GC Maps size",
2527 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
2528 mono_counters_register ("GC Call Sites size",
2529 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
2530 mono_counters_register ("GC Bitmaps size",
2531 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
2532 mono_counters_register ("GC Map struct size",
2533 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
2534 mono_counters_register ("GC Call Sites encoded using 8 bits",
2535 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
2536 mono_counters_register ("GC Call Sites encoded using 16 bits",
2537 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
2538 mono_counters_register ("GC Call Sites encoded using 32 bits",
2539 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
2541 mono_counters_register ("GC Map slots (all)",
2542 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
2543 mono_counters_register ("GC Map slots (ref)",
2544 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.ref_slots);
2545 mono_counters_register ("GC Map slots (noref)",
2546 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.noref_slots);
2547 mono_counters_register ("GC Map slots (pin)",
2548 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
2550 mono_counters_register ("GC TLS Data size",
2551 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2553 mono_counters_register ("Stack space scanned (all)",
2554 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
2555 mono_counters_register ("Stack space scanned (native)",
2556 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_native);
2557 mono_counters_register ("Stack space scanned (other)",
2558 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_other);
2559 mono_counters_register ("Stack space scanned (using GC Maps)",
2560 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
2561 mono_counters_register ("Stack space scanned (precise)",
2562 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_precisely);
2563 mono_counters_register ("Stack space scanned (pin)",
2564 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_conservatively);
2565 mono_counters_register ("Stack space scanned (pin registers)",
2566 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_registers);
2572 mini_gc_enable_gc_maps_for_aot (void)
2580 memset (&cb, 0, sizeof (cb));
2581 cb.get_provenance_func = get_provenance_func;
2582 mono_gc_set_gc_callbacks (&cb);
2588 mini_gc_init_gc_map (MonoCompile *cfg)
2593 mini_gc_create_gc_map (MonoCompile *cfg)
2598 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2603 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2607 #endif /* DISABLE_JIT */
2616 * Set GC specific options in CFG.
2619 mini_gc_init_cfg (MonoCompile *cfg)
2621 if (mono_gc_is_moving ()) {
2622 cfg->disable_ref_noref_stack_slot_share = TRUE;
2623 cfg->gen_write_barriers = TRUE;
2626 mini_gc_init_gc_map (cfg);
2629 #endif /* DISABLE_JIT */
2632 * Problems with the current code:
2633 * - the stack walk is slow
2634 * - vtypes/refs used in EH regions are treated conservatively
2635 * - if the code is finished, less pinning will be done, causing problems because
2636 * we promote all surviving objects to old-gen.
2637 * - the unwind code can't handle a method stopped inside a finally region, it thinks the caller is
2638 * another method, but in reality it is either the exception handling code or the CALL_HANDLER opcode.
2639 * This manifests in "Unable to find ip offset x in callsite list" assertions.
2640 * - the unwind code also can't handle frames which are in the epilog, since the unwind info is not
2645 * Ideas for creating smaller GC maps:
2646 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2648 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2650 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2651 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2652 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2653 * from an index entry.