3 * GC interface for the mono JIT
6 * Zoltan Varga (vargaz@gmail.com)
8 * Copyright 2009 Novell, Inc (http://www.novell.com)
9 * Copyright 2011 Xamarin, Inc (http://www.xamarin.com)
10 * Licensed under the MIT license. See LICENSE file in the project root for full license information.
15 #include <mono/metadata/gc-internals.h>
18 get_provenance (StackFrameInfo *frame, MonoContext *ctx, gpointer data)
20 MonoJitInfo *ji = frame->ji;
24 method = jinfo_get_method (ji);
25 if (method->wrapper_type != MONO_WRAPPER_NONE)
27 *(gpointer *)data = method;
32 get_provenance_func (void)
34 gpointer provenance = NULL;
35 mono_walk_stack (get_provenance, MONO_UNWIND_DEFAULT, (gpointer)&provenance);
40 //#if defined(MONO_ARCH_GC_MAPS_SUPPORTED)
42 #include <mono/metadata/sgen-conf.h>
43 #include <mono/metadata/gc-internals.h>
44 #include <mono/utils/mono-counters.h>
45 #include <mono/utils/unlocked.h>
47 #define SIZEOF_SLOT ((int)sizeof (mgreg_t))
49 #define GC_BITS_PER_WORD (sizeof (mword) * 8)
51 /* Contains state needed by the GC Map construction code */
54 * This contains information about stack slots initialized in the prolog, encoded using
55 * (slot_index << 16) | slot_type. The slot_index is relative to the CFA, i.e. 0
56 * means cfa+0, 1 means cfa-4/8, etc.
58 GSList *stack_slots_from_cfa;
59 /* Same for stack slots relative to the frame pointer */
60 GSList *stack_slots_from_fp;
62 /* Number of slots in the map */
64 /* The number of registers in the map */
66 /* Min and Max offsets of the stack frame relative to fp */
67 int min_offset, max_offset;
68 /* Same for the locals area */
69 int locals_min_offset, locals_max_offset;
71 /* The call sites where this frame can be stopped during GC */
72 GCCallSite **callsites;
73 /* The number of call sites */
77 * The width of the stack bitmaps in bytes. This is not equal to the bitmap width at
78 * runtime, since it includes columns which are 0.
80 int stack_bitmap_width;
82 * A bitmap whose width equals nslots, and whose height equals ncallsites.
83 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
86 guint8 *stack_ref_bitmap;
87 /* Same for SLOT_PIN */
88 guint8 *stack_pin_bitmap;
91 * Similar bitmaps for registers. These have width MONO_MAX_IREGS in bits.
94 guint8 *reg_ref_bitmap;
95 guint8 *reg_pin_bitmap;
98 #define ALIGN_TO(val,align) ((((mgreg_t)val) + ((align) - 1)) & ~((align) - 1))
103 /* We don't support debug levels, its all-or-nothing */
104 #define DEBUG(s) do { s; fflush (logfile); } while (0)
105 #define DEBUG_ENABLED 1
112 #define DEBUG_PRECISE(s) do { s; } while (0)
113 #define DEBUG_PRECISE_ENABLED
115 #define DEBUG_PRECISE(s)
119 * Contains information collected during the conservative stack marking pass,
120 * used during the precise pass. This helps to avoid doing a stack walk twice, which
126 int frame_start_offset;
128 /* Relative to stack_start */
129 int reg_locations [MONO_MAX_IREGS];
130 #ifdef DEBUG_PRECISE_ENABLED
133 int regs [MONO_MAX_IREGS];
137 /* Max number of frames stored in the TLS data */
138 #define MAX_FRAMES 50
141 * Per-thread data kept by this module. This is stored in the GC and passed to us as
142 * parameters, instead of being stored in a TLS variable, since during a collection,
143 * only the collection thread is active.
146 MonoThreadUnwindState unwind_state;
147 MonoThreadInfo *info;
150 gpointer ref_to_track;
151 /* Number of frames collected during the !precise pass */
153 FrameInfo frames [MAX_FRAMES];
156 /* These are constant so don't store them in the GC Maps */
157 /* Number of registers stored in gc maps */
158 #define NREGS MONO_MAX_IREGS
162 * Contains information needed to mark a stack frame.
163 * This is a transient structure, created from a compressed representation on-demand.
167 * The offsets of the GC tracked area inside the stack frame relative to the frame pointer.
168 * This includes memory which is NOREF thus doesn't need GC maps.
173 * The offset relative to frame_offset where the the memory described by the GC maps
177 /* The number of stack slots in the map */
179 /* The frame pointer register */
181 /* The size of each callsite table entry */
182 guint8 callsite_entry_size;
183 guint has_pin_slots : 1;
184 guint has_ref_slots : 1;
185 guint has_ref_regs : 1;
186 guint has_pin_regs : 1;
188 /* The offsets below are into an external bitmaps array */
191 * A bitmap whose width is equal to bitmap_width, and whose height is equal to ncallsites.
192 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
195 guint32 stack_ref_bitmap_offset;
197 * Same for SLOT_PIN. It is possible that the same bit is set in both bitmaps at
198 * different callsites, if the slot starts out as PIN, and later changes to REF.
200 guint32 stack_pin_bitmap_offset;
203 * Corresponding bitmaps for registers
204 * These have width equal to the number of bits set in reg_ref_mask/reg_pin_mask.
205 * FIXME: Merge these with the normal bitmaps, i.e. reserve the first x slots for them ?
207 guint32 reg_pin_bitmap_offset;
208 guint32 reg_ref_bitmap_offset;
210 guint32 used_int_regs, reg_ref_mask, reg_pin_mask;
212 /* The number of bits set in the two masks above */
213 guint8 nref_regs, npin_regs;
216 * A bit array marking slots which contain refs.
217 * This is used only for debugging.
221 /* Callsite offsets */
222 /* These can take up a lot of space, so encode them compactly */
232 * A compressed version of GCMap. This is what gets stored in MonoJitInfo.
236 //guint8 encoded_size;
239 * The arrays below are embedded after the struct.
240 * Their address needs to be computed.
243 /* The fixed fields of the GCMap encoded using LEB128 */
244 guint8 encoded [MONO_ZERO_LEN_ARRAY];
246 /* An array of ncallsites entries, each entry is callsite_entry_size bytes long */
247 guint8 callsites [MONO_ZERO_LEN_ARRAY];
250 guint8 bitmaps [MONO_ZERO_LEN_ARRAY];
253 static int precise_frame_count [2], precise_frame_limit = -1;
254 static gboolean precise_frame_limit_inited;
258 gint32 scanned_stacks;
260 gint32 scanned_precisely;
261 gint32 scanned_conservatively;
262 gint32 scanned_registers;
263 gint32 scanned_native;
264 gint32 scanned_other;
272 gint32 gc_callsites_size;
273 gint32 gc_callsites8_size;
274 gint32 gc_callsites16_size;
275 gint32 gc_callsites32_size;
276 gint32 gc_bitmaps_size;
277 gint32 gc_map_struct_size;
281 static JITGCStats stats;
283 static FILE *logfile;
285 static gboolean enable_gc_maps_for_aot;
288 mini_gc_enable_gc_maps_for_aot (void)
290 enable_gc_maps_for_aot = TRUE;
293 // FIXME: Move these to a shared place
296 encode_uleb128 (guint32 value, guint8 *buf, guint8 **endbuf)
301 guint8 b = value & 0x7f;
303 if (value != 0) /* more bytes to come */
311 static G_GNUC_UNUSED void
312 encode_sleb128 (gint32 value, guint8 *buf, guint8 **endbuf)
315 gboolean negative = (value < 0);
323 /* the following is unnecessary if the
324 * implementation of >>= uses an arithmetic rather
325 * than logical shift for a signed left operand
329 value |= - (1 <<(size - 7));
330 /* sign bit of byte is second high order bit (0x40) */
331 if ((value == 0 && !(byte & 0x40)) ||
332 (value == -1 && (byte & 0x40)))
342 static inline guint32
343 decode_uleb128 (guint8 *buf, guint8 **endbuf)
353 res = res | (((int)(b & 0x7f)) << shift);
365 decode_sleb128 (guint8 *buf, guint8 **endbuf)
375 res = res | (((int)(b & 0x7f)) << shift);
378 if (shift < 32 && (b & 0x40))
379 res |= - (1 << shift);
390 encode_frame_reg (int frame_reg)
393 if (frame_reg == AMD64_RSP)
395 else if (frame_reg == AMD64_RBP)
397 #elif defined(TARGET_X86)
398 if (frame_reg == X86_EBP)
400 else if (frame_reg == X86_ESP)
402 #elif defined(TARGET_ARM)
403 if (frame_reg == ARMREG_SP)
405 else if (frame_reg == ARMREG_FP)
407 #elif defined(TARGET_S390X)
408 if (frame_reg == S390_SP)
410 else if (frame_reg == S390_FP)
415 g_assert_not_reached ();
420 decode_frame_reg (int encoded)
425 else if (encoded == 1)
427 #elif defined(TARGET_X86)
430 else if (encoded == 1)
432 #elif defined(TARGET_ARM)
435 else if (encoded == 1)
437 #elif defined(TARGET_S390X)
440 else if (encoded == 1)
445 g_assert_not_reached ();
451 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15, AMD64_RDI, AMD64_RSI };
453 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15 };
455 #elif defined(TARGET_X86)
456 static int callee_saved_regs [] = { X86_EBX, X86_ESI, X86_EDI };
457 #elif defined(TARGET_ARM)
458 static int callee_saved_regs [] = { ARMREG_V1, ARMREG_V2, ARMREG_V3, ARMREG_V4, ARMREG_V5, ARMREG_V7, ARMREG_FP };
459 #elif defined(TARGET_ARM64)
461 static int callee_saved_regs [] = { };
462 #elif defined(TARGET_S390X)
463 static int callee_saved_regs [] = { s390_r6, s390_r7, s390_r8, s390_r9, s390_r10, s390_r11, s390_r12, s390_r13, s390_r14 };
464 #elif defined(TARGET_POWERPC64) && _CALL_ELF == 2
465 static int callee_saved_regs [] = {
466 ppc_r13, ppc_r14, ppc_r15, ppc_r16,
467 ppc_r17, ppc_r18, ppc_r19, ppc_r20,
468 ppc_r21, ppc_r22, ppc_r23, ppc_r24,
469 ppc_r25, ppc_r26, ppc_r27, ppc_r28,
470 ppc_r29, ppc_r30, ppc_r31 };
471 #elif defined(TARGET_POWERPC)
472 static int callee_saved_regs [] = { ppc_r6, ppc_r7, ppc_r8, ppc_r9, ppc_r10, ppc_r11, ppc_r12, ppc_r13, ppc_r14 };
476 encode_regmask (guint32 regmask)
482 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
483 if (regmask & (1 << callee_saved_regs [i])) {
485 regmask -= (1 << callee_saved_regs [i]);
488 g_assert (regmask == 0);
493 decode_regmask (guint32 regmask)
499 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i)
500 if (regmask & (1 << i))
501 res |= (1 << callee_saved_regs [i]);
508 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
511 encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
515 encode_sleb128 (map->start_offset / SIZEOF_SLOT, buf, &buf);
516 encode_sleb128 (map->end_offset / SIZEOF_SLOT, buf, &buf);
517 encode_sleb128 (map->map_offset / SIZEOF_SLOT, buf, &buf);
518 encode_uleb128 (map->nslots, buf, &buf);
519 g_assert (map->callsite_entry_size <= 4);
520 freg = encode_frame_reg (map->frame_reg);
522 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);
523 encode_uleb128 (flags, buf, &buf);
524 encode_uleb128 (encode_regmask (map->used_int_regs), buf, &buf);
525 if (map->has_ref_regs)
526 encode_uleb128 (encode_regmask (map->reg_ref_mask), buf, &buf);
527 if (map->has_pin_regs)
528 encode_uleb128 (encode_regmask (map->reg_pin_mask), buf, &buf);
529 encode_uleb128 (map->ncallsites, buf, &buf);
537 * Decode the encoded GC map representation in BUF and store the result into MAP.
540 decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
543 int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
546 map->start_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
547 map->end_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
548 map->map_offset = decode_sleb128 (buf, &buf) * SIZEOF_SLOT;
549 map->nslots = decode_uleb128 (buf, &buf);
550 flags = decode_uleb128 (buf, &buf);
551 map->has_ref_slots = (flags & 1) ? 1 : 0;
552 map->has_pin_slots = (flags & 2) ? 1 : 0;
553 map->has_ref_regs = (flags & 4) ? 1 : 0;
554 map->has_pin_regs = (flags & 8) ? 1 : 0;
555 map->callsite_entry_size = ((flags >> 4) & 0x3) + 1;
557 map->frame_reg = decode_frame_reg (freg);
558 map->used_int_regs = decode_regmask (decode_uleb128 (buf, &buf));
559 if (map->has_ref_regs) {
560 map->reg_ref_mask = decode_regmask (decode_uleb128 (buf, &buf));
562 for (i = 0; i < NREGS; ++i)
563 if (map->reg_ref_mask & (1 << i))
567 if (map->has_pin_regs) {
568 map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
570 for (i = 0; i < NREGS; ++i)
571 if (map->reg_pin_mask & (1 << i))
575 map->ncallsites = decode_uleb128 (buf, &buf);
577 stack_bitmap_size = (ALIGN_TO (map->nslots, 8) / 8) * map->ncallsites;
578 reg_ref_bitmap_size = (ALIGN_TO (map->nref_regs, 8) / 8) * map->ncallsites;
579 reg_pin_bitmap_size = (ALIGN_TO (map->npin_regs, 8) / 8) * map->ncallsites;
581 map->stack_ref_bitmap_offset = offset;
582 if (map->has_ref_slots)
583 offset += stack_bitmap_size;
584 map->stack_pin_bitmap_offset = offset;
585 if (map->has_pin_slots)
586 offset += stack_bitmap_size;
587 map->reg_ref_bitmap_offset = offset;
588 if (map->has_ref_regs)
589 offset += reg_ref_bitmap_size;
590 map->reg_pin_bitmap_offset = offset;
591 if (map->has_pin_regs)
592 offset += reg_pin_bitmap_size;
598 thread_attach_func (void)
602 tls = g_new0 (TlsData, 1);
603 tls->tid = mono_native_thread_id_get ();
604 tls->info = mono_thread_info_current ();
605 UnlockedAdd (&stats.tlsdata_size, sizeof (TlsData));
611 thread_detach_func (gpointer user_data)
613 TlsData *tls = user_data;
619 thread_suspend_func (gpointer user_data, void *sigctx, MonoContext *ctx)
621 TlsData *tls = user_data;
624 /* Happens during startup */
628 if (tls->tid != mono_native_thread_id_get ()) {
629 /* Happens on osx because threads are not suspended using signals */
634 g_assert (tls->info);
638 res = mono_thread_state_init_from_handle (&tls->unwind_state, tls->info);
641 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF] = mono_get_lmf ();
643 #ifdef MONO_ARCH_HAVE_SIGCTX_TO_MONOCTX
644 mono_sigctx_to_monoctx (sigctx, &tls->unwind_state.ctx);
645 tls->unwind_state.valid = TRUE;
647 tls->unwind_state.valid = FALSE;
650 memcpy (&tls->unwind_state.ctx, ctx, sizeof (MonoContext));
651 tls->unwind_state.valid = TRUE;
653 tls->unwind_state.valid = FALSE;
655 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_JIT_TLS] = mono_tls_get_jit_tls ();
656 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN] = mono_domain_get ();
659 if (!tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN]) {
660 /* Happens during startup */
661 tls->unwind_state.valid = FALSE;
666 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
669 set_bit (guint8 *bitmap, int width, int y, int x)
671 bitmap [(width * y) + (x / 8)] |= (1 << (x % 8));
675 clear_bit (guint8 *bitmap, int width, int y, int x)
677 bitmap [(width * y) + (x / 8)] &= ~(1 << (x % 8));
681 get_bit (guint8 *bitmap, int width, int y, int x)
683 return bitmap [(width * y) + (x / 8)] & (1 << (x % 8));
687 slot_type_to_string (GCSlotType type)
697 g_assert_not_reached ();
702 static inline mgreg_t
703 get_frame_pointer (MonoContext *ctx, int frame_reg)
705 #if defined(TARGET_AMD64)
706 if (frame_reg == AMD64_RSP)
708 else if (frame_reg == AMD64_RBP)
710 #elif defined(TARGET_X86)
711 if (frame_reg == X86_ESP)
713 else if (frame_reg == X86_EBP)
715 #elif defined(TARGET_ARM)
716 if (frame_reg == ARMREG_SP)
717 return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
718 else if (frame_reg == ARMREG_FP)
719 return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
720 #elif defined(TARGET_S390X)
721 if (frame_reg == S390_SP)
722 return (mgreg_t)MONO_CONTEXT_GET_SP (ctx);
723 else if (frame_reg == S390_FP)
724 return (mgreg_t)MONO_CONTEXT_GET_BP (ctx);
726 g_assert_not_reached ();
731 * conservatively_pass:
733 * Mark a thread stack conservatively and collect information needed by the precise pass.
736 conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
740 MonoContext ctx, new_ctx;
743 gboolean last = TRUE;
747 guint8* fp, *p, *real_frame_start, *frame_start, *frame_end;
748 int i, pc_offset, cindex, bitmap_width;
749 int scanned = 0, scanned_precisely, scanned_conservatively, scanned_registers;
751 StackFrameInfo frame;
752 mgreg_t *reg_locations [MONO_MAX_IREGS];
753 mgreg_t *new_reg_locations [MONO_MAX_IREGS];
756 guint32 precise_regmask;
760 tls->ref_to_track = NULL;
763 /* tls == NULL can happen during startup */
764 if (mono_thread_internal_current () == NULL || !tls) {
765 mono_gc_conservatively_scan_area (stack_start, stack_end);
766 UnlockedAdd (&stats.scanned_stacks, stack_end - stack_start);
770 lmf = tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF];
773 /* Number of bytes scanned based on GC map data */
775 /* Number of bytes scanned precisely based on GC map data */
776 scanned_precisely = 0;
777 /* Number of bytes scanned conservatively based on GC map data */
778 scanned_conservatively = 0;
779 /* Number of bytes scanned conservatively in register save areas */
780 scanned_registers = 0;
782 /* This is one past the last address which we have scanned */
783 stack_limit = stack_start;
785 if (!tls->unwind_state.valid)
786 memset (&new_ctx, 0, sizeof (ctx));
788 memcpy (&new_ctx, &tls->unwind_state.ctx, sizeof (MonoContext));
790 memset (reg_locations, 0, sizeof (reg_locations));
791 memset (new_reg_locations, 0, sizeof (new_reg_locations));
794 if (!tls->unwind_state.valid)
797 memcpy (&ctx, &new_ctx, sizeof (ctx));
799 for (i = 0; i < MONO_MAX_IREGS; ++i) {
800 if (new_reg_locations [i]) {
802 * If the current frame saves the register, it means it might modify its
803 * value, thus the old location might not contain the same value, so
804 * we have to mark it conservatively.
806 if (reg_locations [i]) {
807 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
808 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
809 scanned_registers += SIZEOF_SLOT;
812 reg_locations [i] = new_reg_locations [i];
814 DEBUG (fprintf (logfile, "\treg %s is now at location %p.\n", mono_arch_regname (i), reg_locations [i]));
818 g_assert ((mgreg_t)stack_limit % SIZEOF_SLOT == 0);
820 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);
826 // FIXME: For skipped frames, scan the param area of the parent frame conservatively ?
827 // FIXME: trampolines
829 if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
831 * These frames are problematic for several reasons:
832 * - they are unwound through an LMF, and we have no precise register tracking for those.
833 * - the LMF might not contain a precise ip, so we can't compute the call site.
834 * - the LMF only unwinds to the wrapper frame, so we get these methods twice.
836 DEBUG (fprintf (logfile, "Mark(0): <Managed-to-native transition>\n"));
837 for (i = 0; i < MONO_MAX_IREGS; ++i) {
838 if (reg_locations [i]) {
839 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
840 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
841 scanned_registers += SIZEOF_SLOT;
843 reg_locations [i] = NULL;
844 new_reg_locations [i] = NULL;
851 method = jinfo_get_method (ji);
855 /* The last frame can be in any state so mark conservatively */
858 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));
860 DEBUG (fprintf (logfile, "\t <Last frame>\n"));
863 * new_reg_locations is not precise when a method is interrupted during its epilog, so clear it.
865 for (i = 0; i < MONO_MAX_IREGS; ++i) {
866 if (reg_locations [i]) {
867 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), reg_locations [i]));
868 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
869 scanned_registers += SIZEOF_SLOT;
871 if (new_reg_locations [i]) {
872 DEBUG (fprintf (logfile, "\tscan saved reg %s location %p.\n", mono_arch_regname (i), new_reg_locations [i]));
873 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
874 scanned_registers += SIZEOF_SLOT;
876 reg_locations [i] = NULL;
877 new_reg_locations [i] = NULL;
882 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
884 /* These frames are very problematic */
885 if (method->wrapper_type == MONO_WRAPPER_MANAGED_TO_NATIVE) {
886 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));
887 DEBUG (fprintf (logfile, "\tSkip.\n"));
891 /* All the other frames are at a call site */
893 if (tls->nframes == MAX_FRAMES) {
895 * Can't save information since the array is full. So scan the rest of the
896 * stack conservatively.
898 DEBUG (fprintf (logfile, "Mark (0): Frame stack full.\n"));
902 /* Scan the frame of this method */
905 * A frame contains the following:
910 * - localloc-ed memory
912 g_assert (pc_offset >= 0);
917 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));
918 DEBUG (fprintf (logfile, "\tNo GC Map.\n"));
922 /* The embedded callsite table requires this */
923 g_assert (((mgreg_t)emap % 4) == 0);
926 * Debugging aid to control the number of frames scanned precisely
928 if (!precise_frame_limit_inited) {
929 char *mono_precise_count = g_getenv ("MONO_PRECISE_COUNT");
930 if (mono_precise_count) {
931 precise_frame_limit = atoi (mono_precise_count);
932 g_free (mono_precise_count);
934 precise_frame_limit_inited = TRUE;
937 if (precise_frame_limit != -1) {
938 if (precise_frame_count [FALSE] == precise_frame_limit)
939 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (method, TRUE));
940 if (precise_frame_count [FALSE] > precise_frame_limit)
943 precise_frame_count [FALSE] ++;
945 /* Decode the encoded GC map */
947 memset (map, 0, sizeof (GCMap));
948 decode_gc_map (&emap->encoded [0], map, &p);
949 p = (guint8*)ALIGN_TO (p, map->callsite_entry_size);
950 map->callsites.offsets8 = p;
951 p += map->callsite_entry_size * map->ncallsites;
954 fp = (guint8*)get_frame_pointer (&ctx, map->frame_reg);
956 real_frame_start = fp + map->start_offset;
957 frame_start = fp + map->start_offset + map->map_offset;
958 frame_end = fp + map->end_offset;
960 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));
962 /* Find the callsite index */
963 if (map->callsite_entry_size == 1) {
964 for (i = 0; i < map->ncallsites; ++i)
965 /* ip points inside the call instruction */
966 if (map->callsites.offsets8 [i] == pc_offset + 1)
968 } else if (map->callsite_entry_size == 2) {
969 // FIXME: Use a binary search
970 for (i = 0; i < map->ncallsites; ++i)
971 /* ip points inside the call instruction */
972 if (map->callsites.offsets16 [i] == pc_offset + 1)
975 // FIXME: Use a binary search
976 for (i = 0; i < map->ncallsites; ++i)
977 /* ip points inside the call instruction */
978 if (map->callsites.offsets32 [i] == pc_offset + 1)
981 if (i == map->ncallsites) {
982 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset + 1, mono_method_full_name (method, TRUE));
983 g_assert_not_reached ();
988 * This is not neccessary true on x86 because frames have a different size at each
991 //g_assert (real_frame_start >= stack_limit);
993 if (real_frame_start > stack_limit) {
994 /* This scans the previously skipped frames as well */
995 DEBUG (fprintf (logfile, "\tscan area %p-%p (%d).\n", stack_limit, real_frame_start, (int)(real_frame_start - stack_limit)));
996 mono_gc_conservatively_scan_area (stack_limit, real_frame_start);
997 UnlockedAdd (&stats.scanned_other, real_frame_start - stack_limit);
1000 /* Mark stack slots */
1001 if (map->has_pin_slots) {
1002 int bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
1003 guint8 *pin_bitmap = &bitmaps [map->stack_pin_bitmap_offset + (bitmap_width * cindex)];
1008 for (i = 0; i < map->nslots; ++i) {
1009 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
1011 DEBUG (fprintf (logfile, "\tscan slot %s0x%x(fp)=%p.\n", (guint8*)p > (guint8*)fp ? "" : "-", ABS ((int)((gssize)p - (gssize)fp)), p));
1012 mono_gc_conservatively_scan_area (p, p + SIZEOF_SLOT);
1013 scanned_conservatively += SIZEOF_SLOT;
1015 scanned_precisely += SIZEOF_SLOT;
1020 scanned_precisely += (map->nslots * SIZEOF_SLOT);
1023 /* The area outside of start-end is NOREF */
1024 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * SIZEOF_SLOT);
1026 /* Mark registers */
1027 precise_regmask = map->used_int_regs | (1 << map->frame_reg);
1028 if (map->has_pin_regs) {
1029 int bitmap_width = ALIGN_TO (map->npin_regs, 8) / 8;
1030 guint8 *pin_bitmap = &bitmaps [map->reg_pin_bitmap_offset + (bitmap_width * cindex)];
1032 for (i = 0; i < NREGS; ++i) {
1033 if (!(map->used_int_regs & (1 << i)))
1036 if (!(map->reg_pin_mask & (1 << i)))
1039 if (pin_bitmap [bindex / 8] & (1 << (bindex % 8))) {
1040 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i), reg_locations [i]));
1041 precise_regmask &= ~(1 << i);
1047 scanned += map->end_offset - map->start_offset;
1049 g_assert (scanned == scanned_precisely + scanned_conservatively);
1051 stack_limit = frame_end;
1053 /* Save information for the precise pass */
1054 fi = &tls->frames [tls->nframes];
1055 fi->nslots = map->nslots;
1056 bitmap_width = ALIGN_TO (map->nslots, 8) / 8;
1057 if (map->has_ref_slots)
1058 fi->bitmap = &bitmaps [map->stack_ref_bitmap_offset + (bitmap_width * cindex)];
1061 fi->frame_start_offset = frame_start - stack_start;
1062 fi->nreg_locations = 0;
1063 DEBUG_PRECISE (fi->ji = ji);
1064 DEBUG_PRECISE (fi->fp = fp);
1066 if (map->has_ref_regs) {
1067 int bitmap_width = ALIGN_TO (map->nref_regs, 8) / 8;
1068 guint8 *ref_bitmap = &bitmaps [map->reg_ref_bitmap_offset + (bitmap_width * cindex)];
1070 for (i = 0; i < NREGS; ++i) {
1071 if (!(map->reg_ref_mask & (1 << i)))
1074 if (reg_locations [i] && (ref_bitmap [bindex / 8] & (1 << (bindex % 8)))) {
1075 DEBUG_PRECISE (fi->regs [fi->nreg_locations] = i);
1076 DEBUG (fprintf (logfile, "\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i), reg_locations [i]));
1077 fi->reg_locations [fi->nreg_locations] = (guint8*)reg_locations [i] - stack_start;
1078 fi->nreg_locations ++;
1085 * Clear locations of precisely tracked registers.
1087 if (precise_regmask) {
1088 for (i = 0; i < NREGS; ++i) {
1089 if (precise_regmask & (1 << i)) {
1091 * The method uses this register, and we have precise info for it.
1092 * This means the location will be scanned precisely.
1093 * Tell the code at the beginning of the loop that this location is
1096 if (reg_locations [i])
1097 DEBUG (fprintf (logfile, "\treg %s at location %p (==%p) is precise.\n", mono_arch_regname (i), reg_locations [i], (gpointer)*reg_locations [i]));
1098 reg_locations [i] = NULL;
1106 /* Scan the remaining register save locations */
1107 for (i = 0; i < MONO_MAX_IREGS; ++i) {
1108 if (reg_locations [i]) {
1109 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", reg_locations [i]));
1110 mono_gc_conservatively_scan_area (reg_locations [i], (char*)reg_locations [i] + SIZEOF_SLOT);
1111 scanned_registers += SIZEOF_SLOT;
1113 if (new_reg_locations [i]) {
1114 DEBUG (fprintf (logfile, "\tscan saved reg location %p.\n", new_reg_locations [i]));
1115 mono_gc_conservatively_scan_area (new_reg_locations [i], (char*)new_reg_locations [i] + SIZEOF_SLOT);
1116 scanned_registers += SIZEOF_SLOT;
1120 if (stack_limit < stack_end) {
1121 DEBUG (fprintf (logfile, "\tscan remaining stack %p-%p (%d).\n", stack_limit, stack_end, (int)(stack_end - stack_limit)));
1122 mono_gc_conservatively_scan_area (stack_limit, stack_end);
1123 UnlockedAdd (&stats.scanned_native, stack_end - stack_limit);
1126 DEBUG (fprintf (logfile, "Marked %d bytes, p=%d,c=%d out of %d.\n", scanned, scanned_precisely, scanned_conservatively, (int)(stack_end - stack_start)));
1128 UnlockedAdd (&stats.scanned_stacks, stack_end - stack_start);
1129 UnlockedAdd (&stats.scanned, scanned);
1130 UnlockedAdd (&stats.scanned_precisely, scanned_precisely);
1131 UnlockedAdd (&stats.scanned_conservatively, scanned_conservatively);
1132 UnlockedAdd (&stats.scanned_registers, scanned_registers);
1134 //mono_gc_conservatively_scan_area (stack_start, stack_end);
1140 * Mark a thread stack precisely based on information saved during the conservative
1144 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end, void *gc_data)
1148 guint8 *frame_start;
1153 if (!tls->unwind_state.valid)
1156 for (findex = 0; findex < tls->nframes; findex ++) {
1157 /* Load information saved by the !precise pass */
1158 fi = &tls->frames [findex];
1159 frame_start = stack_start + fi->frame_start_offset;
1161 DEBUG (char *fname = mono_method_full_name (jinfo_get_method (fi->ji), TRUE); fprintf (logfile, "Mark(1): %s\n", fname); g_free (fname));
1164 * FIXME: Add a function to mark using a bitmap, to avoid doing a
1165 * call for each object.
1168 /* Mark stack slots */
1170 guint8 *ref_bitmap = fi->bitmap;
1173 for (i = 0; i < fi->nslots; ++i) {
1174 MonoObject **ptr = (MonoObject**)(frame_start + (i * SIZEOF_SLOT));
1176 live = ref_bitmap [i / 8] & (1 << (i % 8));
1179 MonoObject *obj = *ptr;
1181 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p ->", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1182 *ptr = mono_gc_scan_object (obj, gc_data);
1183 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1185 DEBUG (fprintf (logfile, "\tref %s0x%x(fp)=%p: %p.\n", (guint8*)ptr >= (guint8*)fi->fp ? "" : "-", ABS ((int)((gssize)ptr - (gssize)fi->fp)), ptr, obj));
1190 * This is disabled because the pointer takes up a lot of space.
1191 * Stack slots might be shared between ref and non-ref variables ?
1193 if (map->ref_slots [i / 8] & (1 << (i % 8))) {
1194 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));
1196 * Fail fast if the live range is incorrect, and
1197 * the JITted code tries to access this object
1206 /* Mark registers */
1209 * Registers are different from stack slots, they have no address where they
1210 * are stored. Instead, some frame below this frame in the stack saves them
1211 * in its prolog to the stack. We can mark this location precisely.
1213 for (i = 0; i < fi->nreg_locations; ++i) {
1215 * reg_locations [i] contains the address of the stack slot where
1216 * a reg was last saved, so mark that slot.
1218 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1219 MonoObject *obj = *ptr;
1222 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p ->", mono_arch_regname (fi->regs [i]), ptr, obj));
1223 *ptr = mono_gc_scan_object (obj, gc_data);
1224 DEBUG (fprintf (logfile, " %p.\n", *ptr));
1226 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p\n", mono_arch_regname (fi->regs [i]), ptr, obj));
1232 * Debugging aid to check for missed refs.
1234 if (tls->ref_to_track) {
1237 for (p = (mgreg_t*)stack_start; p < (mgreg_t*)stack_end; ++p)
1238 if (*p == (mgreg_t)tls->ref_to_track)
1239 printf ("REF AT %p.\n", p);
1246 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1247 * call, and TRUE at the second. USER_DATA points to a TlsData
1248 * structure filled up by thread_suspend_func.
1251 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise, void *gc_data)
1253 TlsData *tls = user_data;
1255 DEBUG (fprintf (logfile, "****************************************\n"));
1256 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));
1257 DEBUG (fprintf (logfile, "****************************************\n"));
1260 conservative_pass (tls, stack_start, stack_end);
1262 precise_pass (tls, stack_start, stack_end, gc_data);
1268 mini_gc_init_gc_map (MonoCompile *cfg)
1270 if (COMPILE_LLVM (cfg))
1273 if (!mono_gc_is_moving ())
1276 if (cfg->compile_aot) {
1277 if (!enable_gc_maps_for_aot)
1279 } else if (!mono_gc_precise_stack_mark_enabled ())
1283 /* Debugging support */
1285 static int precise_count;
1288 char *mono_gcmap_count = g_getenv ("MONO_GCMAP_COUNT");
1289 if (mono_gcmap_count) {
1290 int count = atoi (mono_gcmap_count);
1291 g_free (mono_gcmap_count);
1292 if (precise_count == count)
1293 printf ("LAST: %s\n", mono_method_full_name (cfg->method, TRUE));
1294 if (precise_count > count)
1300 cfg->compute_gc_maps = TRUE;
1302 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1306 * mini_gc_set_slot_type_from_fp:
1308 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1309 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1310 * need to call this function for those slots.
1313 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1315 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1317 if (!cfg->compute_gc_maps)
1320 g_assert (slot_offset % SIZEOF_SLOT == 0);
1322 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1326 * mini_gc_set_slot_type_from_cfa:
1328 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1329 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1330 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1331 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1332 * call this function for those slots.
1335 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1337 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1338 int slot = - (slot_offset / SIZEOF_SLOT);
1340 if (!cfg->compute_gc_maps)
1343 g_assert (slot_offset <= 0);
1344 g_assert (slot_offset % SIZEOF_SLOT == 0);
1346 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1350 fp_offset_to_slot (MonoCompile *cfg, int offset)
1352 MonoCompileGC *gcfg = cfg->gc_info;
1354 return (offset - gcfg->min_offset) / SIZEOF_SLOT;
1358 slot_to_fp_offset (MonoCompile *cfg, int slot)
1360 MonoCompileGC *gcfg = cfg->gc_info;
1362 return (slot * SIZEOF_SLOT) + gcfg->min_offset;
1365 static inline MONO_ALWAYS_INLINE void
1366 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1368 g_assert (slot >= 0 && slot < gcfg->nslots);
1370 if (type == SLOT_PIN) {
1371 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1372 set_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1373 } else if (type == SLOT_REF) {
1374 set_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1375 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1376 } else if (type == SLOT_NOREF) {
1377 clear_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1378 clear_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, slot, callsite_index);
1383 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1386 guint8 *ref_bitmap, *pin_bitmap;
1391 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1392 set_slot (gcfg, slot, cindex, type);
1394 ref_bitmap = gcfg->stack_ref_bitmap;
1395 pin_bitmap = gcfg->stack_pin_bitmap;
1396 width = gcfg->stack_bitmap_width;
1399 if (type == SLOT_PIN) {
1400 memset (ref_bitmap + pos, 0, width);
1401 memset (pin_bitmap + pos, 0xff, width);
1402 } else if (type == SLOT_REF) {
1403 memset (ref_bitmap + pos, 0xff, width);
1404 memset (pin_bitmap + pos, 0, width);
1405 } else if (type == SLOT_NOREF) {
1406 memset (ref_bitmap + pos, 0, width);
1407 memset (pin_bitmap + pos, 0, width);
1412 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1416 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1417 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1418 if (callsite_offset >= from && callsite_offset < to)
1419 set_slot (gcfg, slot, cindex, type);
1424 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1426 g_assert (slot >= 0 && slot < gcfg->nregs);
1428 if (type == SLOT_PIN) {
1429 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1430 set_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1431 } else if (type == SLOT_REF) {
1432 set_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1433 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1434 } else if (type == SLOT_NOREF) {
1435 clear_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1436 clear_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, slot, callsite_index);
1441 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1445 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1446 set_reg_slot (gcfg, slot, cindex, type);
1450 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1454 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1455 int callsite_offset = gcfg->callsites [cindex]->pc_offset;
1456 if (callsite_offset >= from && callsite_offset < to)
1457 set_reg_slot (gcfg, slot, cindex, type);
1462 process_spill_slots (MonoCompile *cfg)
1464 MonoCompileGC *gcfg = cfg->gc_info;
1469 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1470 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1471 for (l = bb->spill_slot_defs; l; l = l->next) {
1472 MonoInst *def = l->data;
1473 int spill_slot = def->inst_c0;
1474 int bank = def->inst_c1;
1475 int offset = cfg->spill_info [bank][spill_slot].offset;
1476 int slot = fp_offset_to_slot (cfg, offset);
1478 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1479 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1483 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
1484 for (l = bb->spill_slot_defs; l; l = l->next) {
1485 MonoInst *def = l->data;
1486 int spill_slot = def->inst_c0;
1487 int bank = def->inst_c1;
1488 int offset = cfg->spill_info [bank][spill_slot].offset;
1489 int slot = fp_offset_to_slot (cfg, offset);
1492 if (bank == MONO_REG_INT_MP)
1498 * Extend the live interval for the GC tracked spill slots
1499 * defined in this bblock.
1500 * FIXME: This is not needed.
1502 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1504 if (cfg->verbose_level > 1)
1505 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type), offset >= 0 ? "" : "-", ABS (offset), slot);
1509 /* Set fp spill slots to NOREF */
1510 for (i = 0; i < cfg->spill_info_len [MONO_REG_DOUBLE]; ++i) {
1511 int offset = cfg->spill_info [MONO_REG_DOUBLE][i].offset;
1517 slot = fp_offset_to_slot (cfg, offset);
1519 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1521 if (cfg->verbose_level > 1)
1522 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1525 /* Set int spill slots to NOREF */
1526 for (i = 0; i < cfg->spill_info_len [MONO_REG_INT]; ++i) {
1527 int offset = cfg->spill_info [MONO_REG_INT][i].offset;
1533 slot = fp_offset_to_slot (cfg, offset);
1535 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1536 if (cfg->verbose_level > 1)
1537 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1542 * process_other_slots:
1544 * Process stack slots registered using mini_gc_set_slot_type_... ().
1547 process_other_slots (MonoCompile *cfg)
1549 MonoCompileGC *gcfg = cfg->gc_info;
1552 /* Relative to the CFA */
1553 for (l = gcfg->stack_slots_from_cfa; l; l = l->next) {
1554 guint data = GPOINTER_TO_UINT (l->data);
1555 int cfa_slot = data >> 16;
1556 GCSlotType type = data & 0xff;
1560 * Map the cfa relative slot to an fp relative slot.
1561 * slot_addr == cfa - <cfa_slot>*4/8
1562 * fp + cfa_offset == cfa
1563 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1565 slot = (cfg->cfa_offset / SIZEOF_SLOT) - cfa_slot - (gcfg->min_offset / SIZEOF_SLOT);
1567 set_slot_everywhere (gcfg, slot, type);
1569 if (cfg->verbose_level > 1) {
1570 int fp_offset = slot_to_fp_offset (cfg, slot);
1571 if (type == SLOT_NOREF)
1572 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));
1576 /* Relative to the FP */
1577 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
1578 gint data = GPOINTER_TO_INT (l->data);
1579 int offset = data >> 16;
1580 GCSlotType type = data & 0xff;
1583 slot = fp_offset_to_slot (cfg, offset);
1585 set_slot_everywhere (gcfg, slot, type);
1587 /* Liveness for these slots is handled by process_spill_slots () */
1589 if (cfg->verbose_level > 1) {
1590 if (type == SLOT_REF)
1591 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset, slot);
1592 else if (type == SLOT_NOREF)
1593 printf ("\tnoref slot at 0x%x(fp) (slot = %d)\n", offset, slot);
1599 get_vtype_bitmap (MonoType *t, int *numbits)
1601 MonoClass *klass = mono_class_from_mono_type (t);
1603 if (klass->generic_container || mono_class_is_open_constructed_type (t)) {
1604 /* FIXME: Generic sharing */
1607 mono_class_compute_gc_descriptor (klass);
1609 return mono_gc_get_bitmap_for_descr (klass->gc_descr, numbits);
1613 static inline const char*
1614 get_offset_sign (int offset)
1616 return offset < 0 ? "-" : "+";
1620 get_offset_val (int offset)
1622 return offset < 0 ? (- offset) : offset;
1626 process_variables (MonoCompile *cfg)
1628 MonoCompileGC *gcfg = cfg->gc_info;
1629 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1630 int i, locals_min_slot, locals_max_slot, cindex;
1634 int locals_min_offset = gcfg->locals_min_offset;
1635 int locals_max_offset = gcfg->locals_max_offset;
1637 /* Slots for locals are NOREF by default */
1638 locals_min_slot = (locals_min_offset - gcfg->min_offset) / SIZEOF_SLOT;
1639 locals_max_slot = (locals_max_offset - gcfg->min_offset) / SIZEOF_SLOT;
1640 for (i = locals_min_slot; i < locals_max_slot; ++i) {
1641 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1645 * Compute the offset where variables are initialized in the first bblock, if any.
1647 pc_offsets = g_new0 (int, cfg->next_vreg);
1649 bb = cfg->bb_entry->next_bb;
1650 MONO_BB_FOR_EACH_INS (bb, tmp) {
1651 if (tmp->opcode == OP_GC_LIVENESS_DEF) {
1652 int vreg = tmp->inst_c1;
1653 if (pc_offsets [vreg] == 0) {
1654 g_assert (tmp->backend.pc_offset > 0);
1655 pc_offsets [vreg] = tmp->backend.pc_offset;
1661 * Stack slots holding arguments are initialized in the prolog.
1662 * This means we can treat them alive for the whole method.
1664 for (i = 0; i < cfg->num_varinfo; i++) {
1665 MonoInst *ins = cfg->varinfo [i];
1666 MonoType *t = ins->inst_vtype;
1669 gboolean byref, is_this = FALSE;
1670 gboolean is_arg = i < cfg->locals_start;
1672 if (ins == cfg->ret) {
1673 if (!(ins->opcode == OP_REGOFFSET && MONO_TYPE_ISSTRUCT (t)))
1677 vmv = MONO_VARINFO (cfg, i);
1679 /* For some reason, 'this' is byref */
1680 if (sig->hasthis && ins == cfg->args [0] && !cfg->method->klass->valuetype) {
1681 t = &cfg->method->klass->byval_arg;
1687 if (ins->opcode == OP_REGVAR) {
1689 GCSlotType slot_type;
1691 t = mini_get_underlying_type (t);
1694 g_assert (hreg < MONO_MAX_IREGS);
1697 slot_type = SLOT_PIN;
1699 slot_type = mini_type_is_reference (t) ? SLOT_REF : SLOT_NOREF;
1701 if (slot_type == SLOT_PIN) {
1702 /* These have no live interval, be conservative */
1703 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1706 * Unlike variables allocated to the stack, we generate liveness info
1707 * for noref vars in registers in mono_spill_global_vars (), because
1708 * knowing that a register doesn't contain a ref allows us to mark its save
1709 * locations precisely.
1711 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1712 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1713 set_reg_slot (gcfg, hreg, cindex, slot_type);
1716 if (cfg->verbose_level > 1) {
1717 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type), is_arg ? "arg " : "", mono_arch_regname (hreg), vmv->vreg);
1723 if (ins->opcode != OP_REGOFFSET)
1726 if (ins->inst_offset % SIZEOF_SLOT != 0)
1729 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1731 if (is_arg && ins->flags & MONO_INST_IS_DEAD) {
1732 /* These do not get stored in the prolog */
1733 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1735 if (cfg->verbose_level > 1) {
1736 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));
1741 if (MONO_TYPE_ISSTRUCT (t)) {
1743 gsize *bitmap = NULL;
1744 gboolean pin = FALSE;
1748 if (ins->backend.is_pinvoke)
1749 size = mono_class_native_size (ins->klass, NULL);
1751 size = mono_class_value_size (ins->klass, NULL);
1752 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
1754 if (cfg->verbose_level > 1)
1755 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));
1757 if (!ins->klass->has_references) {
1759 for (j = 0; j < size_in_slots; ++j)
1760 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1765 bitmap = get_vtype_bitmap (t, &numbits);
1770 * Most vtypes are marked volatile because of the LDADDR instructions,
1771 * and they have no liveness information since they are decomposed
1772 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1773 * them during VZERO decomposition.
1776 if (!pc_offsets [vmv->vreg])
1779 if (ins->backend.is_pinvoke)
1784 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1785 if (gcfg->callsites [cindex]->pc_offset > pc_offsets [vmv->vreg]) {
1786 for (j = 0; j < numbits; ++j) {
1787 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD))) {
1788 /* The descriptor is for the boxed object */
1789 set_slot (gcfg, (pos + j - (sizeof (MonoObject) / SIZEOF_SLOT)), cindex, pin ? SLOT_PIN : SLOT_REF);
1795 if (cfg->verbose_level > 1) {
1796 for (j = 0; j < numbits; ++j) {
1797 if (bitmap [j / GC_BITS_PER_WORD] & ((gsize)1 << (j % GC_BITS_PER_WORD)))
1798 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)));
1802 if (cfg->verbose_level > 1)
1803 printf ("\t\tpinned\n");
1804 for (j = 0; j < size_in_slots; ++j) {
1805 set_slot_everywhere (gcfg, pos + j, SLOT_PIN);
1814 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1815 /* Vret addr etc. */
1820 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1822 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1823 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1824 set_slot (gcfg, pos, cindex, SLOT_PIN);
1826 if (cfg->verbose_level > 1)
1827 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));
1832 * This is currently disabled, but could be enabled to debug crashes.
1835 if (t->type == MONO_TYPE_I) {
1837 * Variables created in mono_handle_global_vregs have type I, but they
1838 * could hold GC refs since the vregs they were created from might not been
1839 * marked as holding a GC ref. So be conservative.
1841 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1846 t = mini_get_underlying_type (t);
1848 if (!mini_type_is_reference (t)) {
1849 set_slot_everywhere (gcfg, pos, SLOT_NOREF);
1850 if (cfg->verbose_level > 1)
1851 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));
1852 if (!t->byref && sizeof (mgreg_t) == 4 && (t->type == MONO_TYPE_I8 || t->type == MONO_TYPE_U8 || t->type == MONO_TYPE_R8)) {
1853 set_slot_everywhere (gcfg, pos + 1, SLOT_NOREF);
1854 if (cfg->verbose_level > 1)
1855 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));
1860 /* 'this' is marked INDIRECT for gshared methods */
1861 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1863 * For volatile variables, treat them alive from the point they are
1864 * initialized in the first bblock until the end of the method.
1867 set_slot_everywhere (gcfg, pos, SLOT_REF);
1868 } else if (pc_offsets [vmv->vreg]) {
1869 set_slot_in_range (gcfg, pos, 0, pc_offsets [vmv->vreg], SLOT_PIN);
1870 set_slot_in_range (gcfg, pos, pc_offsets [vmv->vreg], cfg->code_size, SLOT_REF);
1872 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1874 if (cfg->verbose_level > 1)
1875 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));
1880 /* Live for the whole method */
1881 set_slot_everywhere (gcfg, pos, SLOT_REF);
1883 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1884 if (gcfg->callsites [cindex]->liveness [i / 8] & (1 << (i % 8)))
1885 set_slot (gcfg, pos, cindex, SLOT_REF);
1888 if (cfg->verbose_level > 1) {
1889 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));
1893 g_free (pc_offsets);
1897 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1900 * Convert a sp relative offset to a slot index. This is
1901 * platform specific.
1904 /* fp = sp + offset */
1905 g_assert (cfg->frame_reg == AMD64_RBP);
1906 return (- cfg->arch.sp_fp_offset + sp_offset);
1907 #elif defined(TARGET_X86)
1908 /* The offset is computed from the sp at the start of the call sequence */
1909 g_assert (cfg->frame_reg == X86_EBP);
1910 #ifdef MONO_X86_NO_PUSHES
1911 return (- cfg->arch.sp_fp_offset + sp_offset);
1913 return (- cfg->arch.sp_fp_offset - sp_offset);
1922 process_param_area_slots (MonoCompile *cfg)
1924 MonoCompileGC *gcfg = cfg->gc_info;
1929 * These slots are used for passing parameters during calls. They are sp relative, not
1930 * fp relative, so they are harder to handle.
1932 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1933 /* The distance between fp and sp is not constant */
1936 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1938 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1939 GCCallSite *callsite = gcfg->callsites [cindex];
1942 for (l = callsite->param_slots; l; l = l->next) {
1943 MonoInst *def = l->data;
1944 MonoType *t = def->inst_vtype;
1945 int sp_offset = def->inst_offset;
1946 int fp_offset = sp_offset_to_fp_offset (cfg, sp_offset);
1947 int slot = fp_offset_to_slot (cfg, fp_offset);
1951 if (MONO_TYPE_ISSTRUCT (t)) {
1952 size = mini_type_stack_size_full (t, &align, FALSE);
1954 size = sizeof (mgreg_t);
1957 for (i = 0; i < size / sizeof (mgreg_t); ++i) {
1958 g_assert (slot + i >= 0 && slot + i < gcfg->nslots);
1959 is_param [slot + i] = TRUE;
1964 /* All param area slots are noref by default */
1965 for (i = 0; i < gcfg->nslots; ++i) {
1967 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1971 * We treat param area slots as being part of the callee's frame, to be able to handle tail calls which overwrite
1972 * the argument area of the caller.
1977 process_finally_clauses (MonoCompile *cfg)
1979 MonoCompileGC *gcfg = cfg->gc_info;
1980 GCCallSite **callsites;
1982 gboolean has_finally;
1983 int i, j, nslots, nregs;
1985 ncallsites = gcfg->ncallsites;
1986 nslots = gcfg->nslots;
1987 nregs = gcfg->nregs;
1988 callsites = gcfg->callsites;
1991 * The calls to the finally clauses don't show up in the cfg. See
1992 * test_0_liveness_8 ().
1993 * Variables accessed inside the finally clause are already marked VOLATILE by
1994 * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
1995 * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
1997 has_finally = FALSE;
1998 for (i = 0; i < cfg->header->num_clauses; ++i) {
1999 MonoExceptionClause *clause = &cfg->header->clauses [i];
2001 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
2006 if (cfg->verbose_level > 1)
2007 printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
2008 for (j = 0; j < ncallsites; ++j) {
2009 MonoBasicBlock *bb = callsites [j]->bb;
2010 MonoExceptionClause *clause;
2011 gboolean is_in_finally = FALSE;
2013 for (i = 0; i < cfg->header->num_clauses; ++i) {
2014 clause = &cfg->header->clauses [i];
2016 if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
2017 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
2018 is_in_finally = TRUE;
2024 if (is_in_finally) {
2025 for (i = 0; i < nslots; ++i)
2026 set_slot (gcfg, i, j, SLOT_PIN);
2027 for (i = 0; i < nregs; ++i)
2028 set_reg_slot (gcfg, i, j, SLOT_PIN);
2035 compute_frame_size (MonoCompile *cfg)
2037 int i, locals_min_offset, locals_max_offset, cfa_min_offset, cfa_max_offset;
2038 int min_offset, max_offset;
2039 MonoCompileGC *gcfg = cfg->gc_info;
2040 MonoMethodSignature *sig = mono_method_signature (cfg->method);
2043 /* Compute min/max offsets from the fp */
2046 #if defined(TARGET_AMD64) || defined(TARGET_X86) || defined(TARGET_ARM) || defined(TARGET_S390X)
2047 locals_min_offset = ALIGN_TO (cfg->locals_min_stack_offset, SIZEOF_SLOT);
2048 locals_max_offset = cfg->locals_max_stack_offset;
2050 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
2054 locals_min_offset = ALIGN_TO (locals_min_offset, SIZEOF_SLOT);
2055 locals_max_offset = ALIGN_TO (locals_max_offset, SIZEOF_SLOT);
2057 min_offset = locals_min_offset;
2058 max_offset = locals_max_offset;
2061 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
2062 MonoInst *ins = cfg->args [i];
2064 if (ins->opcode == OP_REGOFFSET) {
2065 int size, size_in_slots;
2066 size = mini_type_stack_size_full (ins->inst_vtype, NULL, ins->backend.is_pinvoke);
2067 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
2069 min_offset = MIN (min_offset, ins->inst_offset);
2070 max_offset = MAX ((int)max_offset, (int)(ins->inst_offset + (size_in_slots * SIZEOF_SLOT)));
2075 g_assert (cfg->frame_reg == cfg->cfa_reg);
2076 g_assert (cfg->cfa_offset > 0);
2078 cfa_max_offset = cfg->cfa_offset;
2080 min_offset = MIN (min_offset, cfa_min_offset);
2081 max_offset = MAX (max_offset, cfa_max_offset);
2083 /* Fp relative slots */
2084 for (l = gcfg->stack_slots_from_fp; l; l = l->next) {
2085 gint data = GPOINTER_TO_INT (l->data);
2086 int offset = data >> 16;
2088 min_offset = MIN (min_offset, offset);
2092 if (!(cfg->flags & MONO_CFG_HAS_SPILLUP)) {
2093 int stack_offset = ALIGN_TO (cfg->stack_offset, SIZEOF_SLOT);
2094 min_offset = MIN (min_offset, (-stack_offset));
2097 /* Param area slots */
2099 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
2100 #elif defined(TARGET_X86)
2101 #ifdef MONO_X86_NO_PUSHES
2102 min_offset = MIN (min_offset, -cfg->arch.sp_fp_offset);
2104 min_offset = MIN (min_offset, - (cfg->arch.sp_fp_offset + cfg->arch.param_area_size));
2106 #elif defined(TARGET_ARM)
2108 #elif defined(TARGET_s390X)
2114 gcfg->min_offset = min_offset;
2115 gcfg->max_offset = max_offset;
2116 gcfg->locals_min_offset = locals_min_offset;
2117 gcfg->locals_max_offset = locals_max_offset;
2121 init_gcfg (MonoCompile *cfg)
2123 int i, nregs, nslots;
2124 MonoCompileGC *gcfg = cfg->gc_info;
2125 GCCallSite **callsites;
2134 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2135 ncallsites += g_slist_length (bb->gc_callsites);
2137 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
2139 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2140 for (l = bb->gc_callsites; l; l = l->next)
2141 callsites [i++] = l->data;
2144 /* The callsites should already be ordered by pc offset */
2145 for (i = 1; i < ncallsites; ++i)
2146 g_assert (callsites [i - 1]->pc_offset < callsites [i]->pc_offset);
2149 * The stack frame looks like this:
2151 * <fp + max_offset> == cfa -> <end of previous frame>
2152 * <other stack slots>
2154 * <other stack slots>
2155 * fp + min_offset ->
2160 if (cfg->verbose_level > 1)
2161 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg->method, TRUE), gcfg->min_offset, gcfg->max_offset);
2163 nslots = (gcfg->max_offset - gcfg->min_offset) / SIZEOF_SLOT;
2166 gcfg->nslots = nslots;
2167 gcfg->nregs = nregs;
2168 gcfg->callsites = callsites;
2169 gcfg->ncallsites = ncallsites;
2170 gcfg->stack_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
2171 gcfg->reg_bitmap_width = ALIGN_TO (ncallsites, 8) / 8;
2172 gcfg->stack_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
2173 gcfg->stack_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->stack_bitmap_width * nslots);
2174 gcfg->reg_ref_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
2175 gcfg->reg_pin_bitmap = mono_mempool_alloc0 (cfg->mempool, gcfg->reg_bitmap_width * nregs);
2177 /* All slots start out as PIN */
2178 memset (gcfg->stack_pin_bitmap, 0xff, gcfg->stack_bitmap_width * nregs);
2179 for (i = 0; i < nregs; ++i) {
2181 * By default, registers are NOREF.
2182 * It is possible for a callee to save them before being defined in this method,
2183 * but the saved value is dead too, so it doesn't need to be marked.
2185 if ((cfg->used_int_regs & (1 << i)))
2186 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
2190 static inline gboolean
2191 has_bit_set (guint8 *bitmap, int width, int slot)
2194 int pos = width * slot;
2196 for (i = 0; i < width; ++i) {
2197 if (bitmap [pos + i])
2204 create_map (MonoCompile *cfg)
2207 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
2209 int stack_bitmap_width, stack_bitmap_size, reg_ref_bitmap_width, reg_ref_bitmap_size;
2210 int reg_pin_bitmap_width, reg_pin_bitmap_size, bindex;
2212 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
2213 MonoCompileGC *gcfg = cfg->gc_info;
2214 GCCallSite **callsites;
2216 guint8 *bitmap, *bitmaps;
2217 guint32 reg_ref_mask, reg_pin_mask;
2219 ncallsites = gcfg->ncallsites;
2220 nslots = gcfg->nslots;
2221 nregs = gcfg->nregs;
2222 callsites = gcfg->callsites;
2225 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
2226 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
2228 has_ref_slots = FALSE;
2229 has_pin_slots = FALSE;
2232 memset (ntypes, 0, sizeof (ntypes));
2233 for (i = 0; i < nslots; ++i) {
2234 gboolean has_ref = FALSE;
2235 gboolean has_pin = FALSE;
2237 if (has_bit_set (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i))
2239 if (has_bit_set (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i))
2243 has_ref_slots = TRUE;
2245 has_pin_slots = TRUE;
2248 ntypes [SLOT_REF] ++;
2250 ntypes [SLOT_PIN] ++;
2252 ntypes [SLOT_NOREF] ++;
2254 if (has_ref || has_pin) {
2261 start = end = nslots;
2263 g_assert (start != -1);
2264 g_assert (start < end);
2267 has_ref_regs = FALSE;
2268 has_pin_regs = FALSE;
2273 for (i = 0; i < nregs; ++i) {
2274 gboolean has_ref = FALSE;
2275 gboolean has_pin = FALSE;
2277 if (!(cfg->used_int_regs & (1 << i)))
2280 if (has_bit_set (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i))
2282 if (has_bit_set (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i))
2286 reg_ref_mask |= (1 << i);
2287 has_ref_regs = TRUE;
2291 reg_pin_mask |= (1 << i);
2292 has_pin_regs = TRUE;
2297 if (cfg->verbose_level > 1)
2298 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);
2300 /* Create the GC Map */
2302 /* The work bitmaps have one row for each slot, since this is how we access them during construction */
2303 stack_bitmap_width = ALIGN_TO (end - start, 8) / 8;
2304 stack_bitmap_size = stack_bitmap_width * ncallsites;
2305 reg_ref_bitmap_width = ALIGN_TO (nref_regs, 8) / 8;
2306 reg_ref_bitmap_size = reg_ref_bitmap_width * ncallsites;
2307 reg_pin_bitmap_width = ALIGN_TO (npin_regs, 8) / 8;
2308 reg_pin_bitmap_size = reg_pin_bitmap_width * ncallsites;
2309 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);
2311 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2313 map->frame_reg = cfg->frame_reg;
2314 map->start_offset = gcfg->min_offset;
2315 map->end_offset = gcfg->min_offset + (nslots * SIZEOF_SLOT);
2316 map->map_offset = start * SIZEOF_SLOT;
2317 map->nslots = end - start;
2318 map->has_ref_slots = has_ref_slots;
2319 map->has_pin_slots = has_pin_slots;
2320 map->has_ref_regs = has_ref_regs;
2321 map->has_pin_regs = has_pin_regs;
2322 g_assert (nregs < 32);
2323 map->used_int_regs = cfg->used_int_regs;
2324 map->reg_ref_mask = reg_ref_mask;
2325 map->reg_pin_mask = reg_pin_mask;
2326 map->nref_regs = nref_regs;
2327 map->npin_regs = npin_regs;
2329 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2332 if (has_ref_slots) {
2333 map->stack_ref_bitmap_offset = bitmaps_offset;
2334 bitmaps_offset += stack_bitmap_size;
2336 bitmap = &bitmaps [map->stack_ref_bitmap_offset];
2337 for (i = 0; i < nslots; ++i) {
2338 for (j = 0; j < ncallsites; ++j) {
2339 if (get_bit (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i, j))
2340 set_bit (bitmap, stack_bitmap_width, j, i - start);
2344 if (has_pin_slots) {
2345 map->stack_pin_bitmap_offset = bitmaps_offset;
2346 bitmaps_offset += stack_bitmap_size;
2348 bitmap = &bitmaps [map->stack_pin_bitmap_offset];
2349 for (i = 0; i < nslots; ++i) {
2350 for (j = 0; j < ncallsites; ++j) {
2351 if (get_bit (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i, j))
2352 set_bit (bitmap, stack_bitmap_width, j, i - start);
2357 map->reg_ref_bitmap_offset = bitmaps_offset;
2358 bitmaps_offset += reg_ref_bitmap_size;
2360 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2362 for (i = 0; i < nregs; ++i) {
2363 if (reg_ref_mask & (1 << i)) {
2364 for (j = 0; j < ncallsites; ++j) {
2365 if (get_bit (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i, j))
2366 set_bit (bitmap, reg_ref_bitmap_width, j, bindex);
2373 map->reg_pin_bitmap_offset = bitmaps_offset;
2374 bitmaps_offset += reg_pin_bitmap_size;
2376 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2378 for (i = 0; i < nregs; ++i) {
2379 if (reg_pin_mask & (1 << i)) {
2380 for (j = 0; j < ncallsites; ++j) {
2381 if (get_bit (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i, j))
2382 set_bit (bitmap, reg_pin_bitmap_width, j, bindex);
2390 map->ncallsites = ncallsites;
2391 if (cfg->code_len < 256)
2392 map->callsite_entry_size = 1;
2393 else if (cfg->code_len < 65536)
2394 map->callsite_entry_size = 2;
2396 map->callsite_entry_size = 4;
2398 /* Encode the GC Map */
2406 encode_gc_map (map, buf, &endbuf);
2407 g_assert (endbuf - buf < 256);
2409 encoded_size = endbuf - buf;
2410 alloc_size = sizeof (GCEncodedMap) + ALIGN_TO (encoded_size, map->callsite_entry_size) + (map->callsite_entry_size * map->ncallsites) + bitmaps_size;
2412 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2413 //emap->ref_slots = map->ref_slots;
2415 /* Encoded fixed fields */
2416 p = &emap->encoded [0];
2417 //emap->encoded_size = encoded_size;
2418 memcpy (p, buf, encoded_size);
2421 /* Callsite table */
2422 p = (guint8*)ALIGN_TO ((mgreg_t)p, map->callsite_entry_size);
2423 if (map->callsite_entry_size == 1) {
2424 guint8 *offsets = p;
2425 for (i = 0; i < ncallsites; ++i)
2426 offsets [i] = callsites [i]->pc_offset;
2427 UnlockedAdd (&stats.gc_callsites8_size, ncallsites * sizeof (guint8));
2428 } else if (map->callsite_entry_size == 2) {
2429 guint16 *offsets = (guint16*)p;
2430 for (i = 0; i < ncallsites; ++i)
2431 offsets [i] = callsites [i]->pc_offset;
2432 UnlockedAdd (&stats.gc_callsites16_size, ncallsites * sizeof (guint16));
2434 guint32 *offsets = (guint32*)p;
2435 for (i = 0; i < ncallsites; ++i)
2436 offsets [i] = callsites [i]->pc_offset;
2437 UnlockedAdd (&stats.gc_callsites32_size, ncallsites * sizeof (guint32));
2439 p += ncallsites * map->callsite_entry_size;
2442 memcpy (p, bitmaps, bitmaps_size);
2445 g_assert ((guint8*)p - (guint8*)emap <= alloc_size);
2447 UnlockedAdd (&stats.gc_maps_size, alloc_size);
2448 UnlockedAdd (&stats.gc_callsites_size, ncallsites * map->callsite_entry_size);
2449 UnlockedAdd (&stats.gc_bitmaps_size, bitmaps_size);
2450 UnlockedAdd (&stats.gc_map_struct_size, sizeof (GCEncodedMap) + encoded_size);
2452 cfg->jit_info->gc_info = emap;
2454 cfg->gc_map = (guint8*)emap;
2455 cfg->gc_map_size = alloc_size;
2458 UnlockedAdd (&stats.all_slots, nslots);
2459 UnlockedAdd (&stats.ref_slots, ntypes [SLOT_REF]);
2460 UnlockedAdd (&stats.noref_slots, ntypes [SLOT_NOREF]);
2461 UnlockedAdd (&stats.pin_slots, ntypes [SLOT_PIN]);
2465 mini_gc_create_gc_map (MonoCompile *cfg)
2467 if (!cfg->compute_gc_maps)
2471 * During marking, all frames except the top frame are at a call site, and we mark the
2472 * top frame conservatively. This means that we only need to compute and record
2473 * GC maps for call sites.
2476 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2477 /* Without liveness info, the live ranges are not precise enough */
2480 mono_analyze_liveness_gc (cfg);
2482 compute_frame_size (cfg);
2486 process_spill_slots (cfg);
2487 process_other_slots (cfg);
2488 process_param_area_slots (cfg);
2489 process_variables (cfg);
2490 process_finally_clauses (cfg);
2495 #endif /* DISABLE_JIT */
2498 parse_debug_options (void)
2503 env = g_getenv ("MONO_GCMAP_DEBUG");
2507 opts = g_strsplit (env, ",", -1);
2508 for (ptr = opts; ptr && *ptr; ptr ++) {
2509 /* No options yet */
2510 fprintf (stderr, "Invalid format for the MONO_GCMAP_DEBUG env variable: '%s'\n", env);
2522 memset (&cb, 0, sizeof (cb));
2523 cb.thread_attach_func = thread_attach_func;
2524 cb.thread_detach_func = thread_detach_func;
2525 cb.thread_suspend_func = thread_suspend_func;
2526 /* Comment this out to disable precise stack marking */
2527 cb.thread_mark_func = thread_mark_func;
2528 cb.get_provenance_func = get_provenance_func;
2529 mono_gc_set_gc_callbacks (&cb);
2531 logfile = mono_gc_get_logfile ();
2533 parse_debug_options ();
2535 mono_counters_register ("GC Maps size",
2536 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_maps_size);
2537 mono_counters_register ("GC Call Sites size",
2538 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites_size);
2539 mono_counters_register ("GC Bitmaps size",
2540 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_bitmaps_size);
2541 mono_counters_register ("GC Map struct size",
2542 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_map_struct_size);
2543 mono_counters_register ("GC Call Sites encoded using 8 bits",
2544 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites8_size);
2545 mono_counters_register ("GC Call Sites encoded using 16 bits",
2546 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites16_size);
2547 mono_counters_register ("GC Call Sites encoded using 32 bits",
2548 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.gc_callsites32_size);
2550 mono_counters_register ("GC Map slots (all)",
2551 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.all_slots);
2552 mono_counters_register ("GC Map slots (ref)",
2553 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.ref_slots);
2554 mono_counters_register ("GC Map slots (noref)",
2555 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.noref_slots);
2556 mono_counters_register ("GC Map slots (pin)",
2557 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.pin_slots);
2559 mono_counters_register ("GC TLS Data size",
2560 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2562 mono_counters_register ("Stack space scanned (all)",
2563 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_stacks);
2564 mono_counters_register ("Stack space scanned (native)",
2565 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_native);
2566 mono_counters_register ("Stack space scanned (other)",
2567 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_other);
2568 mono_counters_register ("Stack space scanned (using GC Maps)",
2569 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned);
2570 mono_counters_register ("Stack space scanned (precise)",
2571 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_precisely);
2572 mono_counters_register ("Stack space scanned (pin)",
2573 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_conservatively);
2574 mono_counters_register ("Stack space scanned (pin registers)",
2575 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.scanned_registers);
2581 mini_gc_enable_gc_maps_for_aot (void)
2589 memset (&cb, 0, sizeof (cb));
2590 cb.get_provenance_func = get_provenance_func;
2591 mono_gc_set_gc_callbacks (&cb);
2597 mini_gc_init_gc_map (MonoCompile *cfg)
2602 mini_gc_create_gc_map (MonoCompile *cfg)
2607 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2612 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2616 #endif /* DISABLE_JIT */
2625 * Set GC specific options in CFG.
2628 mini_gc_init_cfg (MonoCompile *cfg)
2630 if (mono_gc_is_moving ()) {
2631 cfg->disable_ref_noref_stack_slot_share = TRUE;
2632 cfg->gen_write_barriers = TRUE;
2635 mini_gc_init_gc_map (cfg);
2638 #endif /* DISABLE_JIT */
2641 * Problems with the current code:
2642 * - the stack walk is slow
2643 * - vtypes/refs used in EH regions are treated conservatively
2644 * - if the code is finished, less pinning will be done, causing problems because
2645 * we promote all surviving objects to old-gen.
2646 * - the unwind code can't handle a method stopped inside a finally region, it thinks the caller is
2647 * another method, but in reality it is either the exception handling code or the CALL_HANDLER opcode.
2648 * This manifests in "Unable to find ip offset x in callsite list" assertions.
2649 * - the unwind code also can't handle frames which are in the epilog, since the unwind info is not
2654 * Ideas for creating smaller GC maps:
2655 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2657 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2659 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2660 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2661 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2662 * from an index entry.