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)
12 #include <mono/metadata/gc-internal.h>
15 #if defined(MONO_ARCH_GC_MAPS_SUPPORTED)
17 #include <mono/metadata/gc-internal.h>
18 #include <mono/utils/mono-counters.h>
20 #if SIZEOF_VOID_P == 4
21 typedef guint32 mword
;
23 typedef guint64 mword
;
26 #define GC_BITS_PER_WORD (sizeof (mword) * 8)
28 /* Contains state needed by the GC Map construction code */
31 * This contains information about stack slots initialized in the prolog, encoded using
32 * (slot_index << 16) | slot_type. The slot_index is relative to the CFA, i.e. 0
33 * means cfa+0, 1 means cfa-4/8, etc.
35 GSList
*stack_slots_from_cfa
;
36 /* Same for stack slots relative to the frame pointer */
37 GSList
*stack_slots_from_fp
;
39 /* Number of slots in the map */
41 /* The number of registers in the map */
43 /* Min and Max offsets of the stack frame relative to fp */
44 int min_offset
, max_offset
;
45 /* Same for the locals area */
46 int locals_min_offset
, locals_max_offset
;
48 /* The call sites where this frame can be stopped during GC */
49 GCCallSite
**callsites
;
50 /* The number of call sites */
54 * The width of the stack bitmaps in bytes. This is not equal to the bitmap width at
55 * runtime, since it includes columns which are 0.
57 int stack_bitmap_width
;
59 * A bitmap whose width equals nslots, and whose height equals ncallsites.
60 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
63 guint8
*stack_ref_bitmap
;
64 /* Same for SLOT_PIN */
65 guint8
*stack_pin_bitmap
;
68 * Similar bitmaps for registers. These have width MONO_MAX_IREGS in bits.
71 guint8
*reg_ref_bitmap
;
72 guint8
*reg_pin_bitmap
;
75 #define ALIGN_TO(val,align) ((((mgreg_t)val) + ((align) - 1)) & ~((align) - 1))
80 /* We don't support debug levels, its all-or-nothing */
81 #define DEBUG(s) do { s; fflush (logfile); } while (0)
82 #define DEBUG_ENABLED 1
89 #define DEBUG_PRECISE(s) do { s; } while (0)
90 #define DEBUG_PRECISE_ENABLED
92 #define DEBUG_PRECISE(s)
96 * Contains information collected during the conservative stack marking pass,
97 * used during the precise pass. This helps to avoid doing a stack walk twice, which
103 int frame_start_offset
;
105 /* Relative to stack_start */
106 int reg_locations
[MONO_MAX_IREGS
];
107 #ifdef DEBUG_PRECISE_ENABLED
110 int regs
[MONO_MAX_IREGS
];
114 /* Max number of frames stored in the TLS data */
115 #define MAX_FRAMES 50
118 * Per-thread data kept by this module. This is stored in the GC and passed to us as
119 * parameters, instead of being stored in a TLS variable, since during a collection,
120 * only the collection thread is active.
125 gboolean has_context
;
126 MonoJitTlsData
*jit_tls
;
129 gpointer ref_to_track
;
130 /* Number of frames collected during the !precise pass */
132 FrameInfo frames
[MAX_FRAMES
];
135 /* These are constant so don't store them in the GC Maps */
136 /* Number of registers stored in gc maps */
137 #define NREGS MONO_MAX_IREGS
141 * Contains information needed to mark a stack frame.
142 * This is a transient structure, created from a compressed representation on-demand.
146 * The offsets of the GC tracked area inside the stack frame relative to the frame pointer.
147 * This includes memory which is NOREF thus doesn't need GC maps.
152 * The offset relative to frame_offset where the the memory described by the GC maps
156 /* The number of stack slots in the map */
158 /* The frame pointer register */
160 /* The size of each callsite table entry */
161 guint8 callsite_entry_size
;
162 guint has_pin_slots
: 1;
163 guint has_ref_slots
: 1;
164 guint has_ref_regs
: 1;
165 guint has_pin_regs
: 1;
167 /* The offsets below are into an external bitmaps array */
170 * A bitmap whose width is equal to bitmap_width, and whose height is equal to ncallsites.
171 * The bitmap contains a 1 if the corresponding stack slot has type SLOT_REF at the
174 guint32 stack_ref_bitmap_offset
;
176 * Same for SLOT_PIN. It is possible that the same bit is set in both bitmaps at
177 * different callsites, if the slot starts out as PIN, and later changes to REF.
179 guint32 stack_pin_bitmap_offset
;
182 * Corresponding bitmaps for registers
183 * These have width equal to the number of bits set in reg_ref_mask/reg_pin_mask.
184 * FIXME: Merge these with the normal bitmaps, i.e. reserve the first x slots for them ?
186 guint32 reg_pin_bitmap_offset
;
187 guint32 reg_ref_bitmap_offset
;
189 guint32 used_int_regs
, reg_ref_mask
, reg_pin_mask
;
191 /* The number of bits set in the two masks above */
192 guint8 nref_regs
, npin_regs
;
195 * A bit array marking slots which contain refs.
196 * This is used only for debugging.
200 /* Callsite offsets */
201 /* These can take up a lot of space, so encode them compactly */
211 * A compressed version of GCMap. This is what gets stored in MonoJitInfo.
215 //guint8 encoded_size;
218 * The arrays below are embedded after the struct.
219 * Their address needs to be computed.
222 /* The fixed fields of the GCMap encoded using LEB128 */
223 guint8 encoded
[MONO_ZERO_LEN_ARRAY
];
225 /* An array of ncallsites entries, each entry is callsite_entry_size bytes long */
226 guint8 callsites
[MONO_ZERO_LEN_ARRAY
];
229 guint8 bitmaps
[MONO_ZERO_LEN_ARRAY
];
232 static int precise_frame_count
[2], precise_frame_limit
= -1;
233 static gboolean precise_frame_limit_inited
;
239 int scanned_precisely
;
240 int scanned_conservatively
;
241 int scanned_registers
;
251 int gc_callsites_size
;
252 int gc_callsites8_size
;
253 int gc_callsites16_size
;
254 int gc_callsites32_size
;
256 int gc_map_struct_size
;
260 static JITGCStats stats
;
262 static FILE *logfile
;
264 // FIXME: Move these to a shared place
267 encode_uleb128 (guint32 value
, guint8
*buf
, guint8
**endbuf
)
272 guint8 b
= value
& 0x7f;
274 if (value
!= 0) /* more bytes to come */
282 static G_GNUC_UNUSED
void
283 encode_sleb128 (gint32 value
, guint8
*buf
, guint8
**endbuf
)
286 gboolean negative
= (value
< 0);
294 /* the following is unnecessary if the
295 * implementation of >>= uses an arithmetic rather
296 * than logical shift for a signed left operand
300 value
|= - (1 <<(size
- 7));
301 /* sign bit of byte is second high order bit (0x40) */
302 if ((value
== 0 && !(byte
& 0x40)) ||
303 (value
== -1 && (byte
& 0x40)))
313 static inline guint32
314 decode_uleb128 (guint8
*buf
, guint8
**endbuf
)
324 res
= res
| (((int)(b
& 0x7f)) << shift
);
336 decode_sleb128 (guint8
*buf
, guint8
**endbuf
)
346 res
= res
| (((int)(b
& 0x7f)) << shift
);
349 if (shift
< 32 && (b
& 0x40))
350 res
|= - (1 << shift
);
361 encode_frame_reg (int frame_reg
)
364 if (frame_reg
== AMD64_RSP
)
366 else if (frame_reg
== AMD64_RBP
)
368 #elif defined(TARGET_X86)
369 if (frame_reg
== X86_EBP
)
371 else if (frame_reg
== X86_ESP
)
376 g_assert_not_reached ();
381 decode_frame_reg (int encoded
)
386 else if (encoded
== 1)
388 #elif defined(TARGET_X86)
391 else if (encoded
== 1)
396 g_assert_not_reached ();
402 static int callee_saved_regs
[] = { AMD64_RBP
, AMD64_RBX
, AMD64_R12
, AMD64_R13
, AMD64_R14
, AMD64_R15
, AMD64_RDI
, AMD64_RSI
};
404 static int callee_saved_regs
[] = { AMD64_RBP
, AMD64_RBX
, AMD64_R12
, AMD64_R13
, AMD64_R14
, AMD64_R15
};
406 #elif defined(TARGET_X86)
407 static int callee_saved_regs
[] = { X86_EBX
, X86_ESI
, X86_EDI
};
411 encode_regmask (guint32 regmask
)
417 for (i
= 0; i
< sizeof (callee_saved_regs
) / sizeof (int); ++i
) {
418 if (regmask
& (1 << callee_saved_regs
[i
])) {
420 regmask
-= (1 << callee_saved_regs
[i
]);
423 g_assert (regmask
== 0);
428 decode_regmask (guint32 regmask
)
434 for (i
= 0; i
< sizeof (callee_saved_regs
) / sizeof (int); ++i
)
435 if (regmask
& (1 << i
))
436 res
|= (1 << callee_saved_regs
[i
]);
443 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
446 encode_gc_map (GCMap
*map
, guint8
*buf
, guint8
**endbuf
)
450 encode_sleb128 (map
->start_offset
/ sizeof (mgreg_t
), buf
, &buf
);
451 encode_sleb128 (map
->end_offset
/ sizeof (mgreg_t
), buf
, &buf
);
452 encode_sleb128 (map
->map_offset
/ sizeof (mgreg_t
), buf
, &buf
);
453 encode_uleb128 (map
->nslots
, buf
, &buf
);
454 g_assert (map
->callsite_entry_size
<= 4);
455 freg
= encode_frame_reg (map
->frame_reg
);
457 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);
458 encode_uleb128 (flags
, buf
, &buf
);
459 encode_uleb128 (encode_regmask (map
->used_int_regs
), buf
, &buf
);
460 if (map
->has_ref_regs
)
461 encode_uleb128 (encode_regmask (map
->reg_ref_mask
), buf
, &buf
);
462 if (map
->has_pin_regs
)
463 encode_uleb128 (encode_regmask (map
->reg_pin_mask
), buf
, &buf
);
464 encode_uleb128 (map
->ncallsites
, buf
, &buf
);
472 * Decode the encoded GC map representation in BUF and store the result into MAP.
475 decode_gc_map (guint8
*buf
, GCMap
*map
, guint8
**endbuf
)
478 int stack_bitmap_size
, reg_ref_bitmap_size
, reg_pin_bitmap_size
, offset
, freg
;
481 map
->start_offset
= decode_sleb128 (buf
, &buf
) * sizeof (mgreg_t
);
482 map
->end_offset
= decode_sleb128 (buf
, &buf
) * sizeof (mgreg_t
);
483 map
->map_offset
= decode_sleb128 (buf
, &buf
) * sizeof (mgreg_t
);
484 map
->nslots
= decode_uleb128 (buf
, &buf
);
485 flags
= decode_uleb128 (buf
, &buf
);
486 map
->has_ref_slots
= (flags
& 1) ? 1 : 0;
487 map
->has_pin_slots
= (flags
& 2) ? 1 : 0;
488 map
->has_ref_regs
= (flags
& 4) ? 1 : 0;
489 map
->has_pin_regs
= (flags
& 8) ? 1 : 0;
490 map
->callsite_entry_size
= ((flags
>> 4) & 0x3) + 1;
492 map
->frame_reg
= decode_frame_reg (freg
);
493 map
->used_int_regs
= decode_regmask (decode_uleb128 (buf
, &buf
));
494 if (map
->has_ref_regs
) {
495 map
->reg_ref_mask
= decode_regmask (decode_uleb128 (buf
, &buf
));
497 for (i
= 0; i
< NREGS
; ++i
)
498 if (map
->reg_ref_mask
& (1 << i
))
502 if (map
->has_pin_regs
) {
503 map
->reg_pin_mask
= decode_regmask (decode_uleb128 (buf
, &buf
));
505 for (i
= 0; i
< NREGS
; ++i
)
506 if (map
->reg_pin_mask
& (1 << i
))
510 map
->ncallsites
= decode_uleb128 (buf
, &buf
);
512 stack_bitmap_size
= (ALIGN_TO (map
->nslots
, 8) / 8) * map
->ncallsites
;
513 reg_ref_bitmap_size
= (ALIGN_TO (map
->nref_regs
, 8) / 8) * map
->ncallsites
;
514 reg_pin_bitmap_size
= (ALIGN_TO (map
->npin_regs
, 8) / 8) * map
->ncallsites
;
516 map
->stack_ref_bitmap_offset
= offset
;
517 if (map
->has_ref_slots
)
518 offset
+= stack_bitmap_size
;
519 map
->stack_pin_bitmap_offset
= offset
;
520 if (map
->has_pin_slots
)
521 offset
+= stack_bitmap_size
;
522 map
->reg_ref_bitmap_offset
= offset
;
523 if (map
->has_ref_regs
)
524 offset
+= reg_ref_bitmap_size
;
525 map
->reg_pin_bitmap_offset
= offset
;
526 if (map
->has_pin_regs
)
527 offset
+= reg_pin_bitmap_size
;
533 thread_attach_func (void)
537 tls
= g_new0 (TlsData
, 1);
538 tls
->tid
= GetCurrentThreadId ();
539 stats
.tlsdata_size
+= sizeof (TlsData
);
545 thread_detach_func (gpointer user_data
)
547 TlsData
*tls
= user_data
;
553 thread_suspend_func (gpointer user_data
, void *sigctx
)
555 TlsData
*tls
= user_data
;
558 /* Happens during startup */
561 tls
->lmf
= mono_get_lmf ();
563 mono_arch_sigctx_to_monoctx (sigctx
, &tls
->ctx
);
564 tls
->has_context
= TRUE
;
566 tls
->has_context
= FALSE
;
568 tls
->jit_tls
= TlsGetValue (mono_jit_tls_id
);
571 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
574 set_bit (guint8
*bitmap
, int width
, int y
, int x
)
576 bitmap
[(width
* y
) + (x
/ 8)] |= (1 << (x
% 8));
580 clear_bit (guint8
*bitmap
, int width
, int y
, int x
)
582 bitmap
[(width
* y
) + (x
/ 8)] &= ~(1 << (x
% 8));
586 get_bit (guint8
*bitmap
, int width
, int y
, int x
)
588 return bitmap
[(width
* y
) + (x
/ 8)] & (1 << (x
% 8));
592 slot_type_to_string (GCSlotType type
)
602 g_assert_not_reached ();
607 static inline mgreg_t
608 get_frame_pointer (MonoContext
*ctx
, int frame_reg
)
610 #if defined(TARGET_AMD64)
611 if (frame_reg
== AMD64_RSP
)
613 else if (frame_reg
== AMD64_RBP
)
615 #elif defined(TARGET_X86)
616 if (frame_reg
== X86_ESP
)
618 else if (frame_reg
== X86_EBP
)
621 g_assert_not_reached ();
626 * conservatively_pass:
628 * Mark a thread stack conservatively and collect information needed by the precise pass.
631 conservative_pass (TlsData
*tls
, guint8
*stack_start
, guint8
*stack_end
)
634 MonoContext ctx
, new_ctx
;
637 gboolean last
= TRUE
;
641 guint8
* fp
, *p
, *real_frame_start
, *frame_start
, *frame_end
;
642 int i
, pc_offset
, cindex
, bitmap_width
;
643 int scanned
= 0, scanned_precisely
, scanned_conservatively
, scanned_registers
;
645 StackFrameInfo frame
;
646 mgreg_t
*reg_locations
[MONO_MAX_IREGS
];
647 mgreg_t
*new_reg_locations
[MONO_MAX_IREGS
];
650 guint32 precise_regmask
;
654 tls
->ref_to_track
= NULL
;
657 /* tls == NULL can happen during startup */
658 if (mono_thread_internal_current () == NULL
|| !tls
) {
659 mono_gc_conservatively_scan_area (stack_start
, stack_end
);
660 stats
.scanned_stacks
+= stack_end
- stack_start
;
667 /* Number of bytes scanned based on GC map data */
669 /* Number of bytes scanned precisely based on GC map data */
670 scanned_precisely
= 0;
671 /* Number of bytes scanned conservatively based on GC map data */
672 scanned_conservatively
= 0;
673 /* Number of bytes scanned conservatively in register save areas */
674 scanned_registers
= 0;
676 /* This is one past the last address which we have scanned */
677 stack_limit
= stack_start
;
679 if (!tls
->has_context
)
680 memset (&new_ctx
, 0, sizeof (ctx
));
682 memcpy (&new_ctx
, &tls
->ctx
, sizeof (MonoContext
));
684 memset (reg_locations
, 0, sizeof (reg_locations
));
685 memset (new_reg_locations
, 0, sizeof (new_reg_locations
));
688 memcpy (&ctx
, &new_ctx
, sizeof (ctx
));
690 for (i
= 0; i
< MONO_MAX_IREGS
; ++i
) {
691 if (new_reg_locations
[i
]) {
693 * If the current frame saves the register, it means it might modify its
694 * value, thus the old location might not contain the same value, so
695 * we have to mark it conservatively.
697 if (reg_locations
[i
]) {
698 DEBUG (fprintf (logfile
, "\tscan saved reg %s location %p.\n", mono_arch_regname (i
), reg_locations
[i
]));
699 mono_gc_conservatively_scan_area (reg_locations
[i
], reg_locations
[i
] + sizeof (mgreg_t
));
700 scanned_registers
+= sizeof (mgreg_t
);
703 reg_locations
[i
] = new_reg_locations
[i
];
705 DEBUG (fprintf (logfile
, "\treg %s is now at location %p.\n", mono_arch_regname (i
), reg_locations
[i
]));
709 g_assert ((mgreg_t
)stack_limit
% sizeof (mgreg_t
) == 0);
711 res
= mono_find_jit_info_ext (frame
.domain
? frame
.domain
: mono_domain_get (), tls
->jit_tls
, NULL
, &ctx
, &new_ctx
, NULL
, &lmf
, new_reg_locations
, &frame
);
717 if (frame
.type
== FRAME_TYPE_MANAGED_TO_NATIVE
) {
719 * These frames are problematic for several reasons:
720 * - they are unwound through an LMF, and we have no precise register tracking for those.
721 * - the LMF might not contain a precise ip, so we can't compute the call site.
722 * - the LMF only unwinds to the wrapper frame, so we get these methods twice.
724 DEBUG (fprintf (logfile
, "Mark(0): <Managed-to-native transition>\n"));
725 for (i
= 0; i
< MONO_MAX_IREGS
; ++i
) {
726 if (reg_locations
[i
]) {
727 DEBUG (fprintf (logfile
, "\tscan saved reg %s location %p.\n", mono_arch_regname (i
), reg_locations
[i
]));
728 mono_gc_conservatively_scan_area (reg_locations
[i
], reg_locations
[i
] + sizeof (mgreg_t
));
729 scanned_registers
+= sizeof (mgreg_t
);
731 reg_locations
[i
] = NULL
;
732 new_reg_locations
[i
] = NULL
;
738 /* The last frame can be in any state so mark conservatively */
741 DEBUG (char *fname
= mono_method_full_name (ji
->method
, TRUE
); fprintf (logfile
, "Mark(0): %s+0x%x (%p)\n", fname
, pc_offset
, (gpointer
)MONO_CONTEXT_GET_IP (&ctx
)); g_free (fname
));
743 DEBUG (fprintf (logfile
, "\t <Last frame>\n"));
748 pc_offset
= (guint8
*)MONO_CONTEXT_GET_IP (&ctx
) - (guint8
*)ji
->code_start
;
750 /* These frames are very problematic */
751 if (ji
->method
->wrapper_type
== MONO_WRAPPER_MANAGED_TO_NATIVE
) {
752 DEBUG (char *fname
= mono_method_full_name (ji
->method
, TRUE
); fprintf (logfile
, "Mark(0): %s+0x%x (%p)\n", fname
, pc_offset
, (gpointer
)MONO_CONTEXT_GET_IP (&ctx
)); g_free (fname
));
753 DEBUG (fprintf (logfile
, "\tSkip.\n"));
757 /* All the other frames are at a call site */
759 if (tls
->nframes
== MAX_FRAMES
) {
761 * Can't save information since the array is full. So scan the rest of the
762 * stack conservatively.
764 DEBUG (fprintf (logfile
, "Mark (0): Frame stack full.\n"));
768 /* Scan the frame of this method */
771 * A frame contains the following:
776 * - localloc-ed memory
778 g_assert (pc_offset
>= 0);
783 DEBUG (char *fname
= mono_method_full_name (ji
->method
, TRUE
); fprintf (logfile
, "Mark(0): %s+0x%x (%p)\n", fname
, pc_offset
, (gpointer
)MONO_CONTEXT_GET_IP (&ctx
)); g_free (fname
));
784 DEBUG (fprintf (logfile
, "\tNo GC Map.\n"));
788 /* The embedded callsite table requires this */
789 g_assert (((mgreg_t
)emap
% 4) == 0);
792 * Debugging aid to control the number of frames scanned precisely
794 if (!precise_frame_limit_inited
) {
795 if (getenv ("MONO_PRECISE_COUNT"))
796 precise_frame_limit
= atoi (getenv ("MONO_PRECISE_COUNT"));
797 precise_frame_limit_inited
= TRUE
;
800 if (precise_frame_limit
!= -1) {
801 if (precise_frame_count
[FALSE
] == precise_frame_limit
)
802 printf ("LAST PRECISE FRAME: %s\n", mono_method_full_name (ji
->method
, TRUE
));
803 if (precise_frame_count
[FALSE
] > precise_frame_limit
)
806 precise_frame_count
[FALSE
] ++;
808 /* Decode the encoded GC map */
810 memset (map
, 0, sizeof (GCMap
));
811 decode_gc_map (&emap
->encoded
[0], map
, &p
);
812 p
= (guint8
*)ALIGN_TO (p
, map
->callsite_entry_size
);
813 map
->callsites
.offsets8
= p
;
814 p
+= map
->callsite_entry_size
* map
->ncallsites
;
817 fp
= (guint8
*)get_frame_pointer (&ctx
, map
->frame_reg
);
819 real_frame_start
= fp
+ map
->start_offset
;
820 frame_start
= fp
+ map
->start_offset
+ map
->map_offset
;
821 frame_end
= fp
+ map
->end_offset
;
823 DEBUG (char *fname
= mono_method_full_name (ji
->method
, 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
));
825 /* Find the callsite index */
826 if (map
->callsite_entry_size
== 1) {
827 for (i
= 0; i
< map
->ncallsites
; ++i
)
828 /* ip points inside the call instruction */
829 if (map
->callsites
.offsets8
[i
] == pc_offset
+ 1)
831 } else if (map
->callsite_entry_size
== 2) {
832 // FIXME: Use a binary search
833 for (i
= 0; i
< map
->ncallsites
; ++i
)
834 /* ip points inside the call instruction */
835 if (map
->callsites
.offsets16
[i
] == pc_offset
+ 1)
838 // FIXME: Use a binary search
839 for (i
= 0; i
< map
->ncallsites
; ++i
)
840 /* ip points inside the call instruction */
841 if (map
->callsites
.offsets32
[i
] == pc_offset
+ 1)
844 if (i
== map
->ncallsites
) {
845 printf ("Unable to find ip offset 0x%x in callsite list of %s.\n", pc_offset
+ 1, mono_method_full_name (ji
->method
, TRUE
));
846 g_assert_not_reached ();
850 g_assert (real_frame_start
>= stack_limit
);
852 if (real_frame_start
> stack_limit
) {
853 /* This scans the previously skipped frames as well */
854 DEBUG (fprintf (logfile
, "\tscan area %p-%p (%d).\n", stack_limit
, real_frame_start
, (int)(real_frame_start
- stack_limit
)));
855 mono_gc_conservatively_scan_area (stack_limit
, real_frame_start
);
856 stats
.scanned_other
+= real_frame_start
- stack_limit
;
859 /* Mark stack slots */
860 if (map
->has_pin_slots
) {
861 int bitmap_width
= ALIGN_TO (map
->nslots
, 8) / 8;
862 guint8
*pin_bitmap
= &bitmaps
[map
->stack_pin_bitmap_offset
+ (bitmap_width
* cindex
)];
867 for (i
= 0; i
< map
->nslots
; ++i
) {
868 pinned
= pin_bitmap
[i
/ 8] & (1 << (i
% 8));
870 DEBUG (fprintf (logfile
, "\tscan slot %s0x%x(fp)=%p.\n", (guint8
*)p
> (guint8
*)fp
? "" : "-", ABS ((int)((gssize
)p
- (gssize
)fp
)), p
));
871 mono_gc_conservatively_scan_area (p
, p
+ sizeof (mgreg_t
));
872 scanned_conservatively
+= sizeof (mgreg_t
);
874 scanned_precisely
+= sizeof (mgreg_t
);
876 p
+= sizeof (mgreg_t
);
879 scanned_precisely
+= (map
->nslots
* sizeof (mgreg_t
));
882 /* The area outside of start-end is NOREF */
883 scanned_precisely
+= (map
->end_offset
- map
->start_offset
) - (map
->nslots
* sizeof (mgreg_t
));
886 precise_regmask
= map
->used_int_regs
| (1 << map
->frame_reg
);
887 if (map
->has_pin_regs
) {
888 int bitmap_width
= ALIGN_TO (map
->npin_regs
, 8) / 8;
889 guint8
*pin_bitmap
= &bitmaps
[map
->reg_pin_bitmap_offset
+ (bitmap_width
* cindex
)];
891 for (i
= 0; i
< NREGS
; ++i
) {
892 if (!(map
->used_int_regs
& (1 << i
)))
895 if (!(map
->reg_pin_mask
& (1 << i
)))
898 if (pin_bitmap
[bindex
/ 8] & (1 << (bindex
% 8))) {
899 DEBUG (fprintf (logfile
, "\treg %s saved at 0x%p is pinning.\n", mono_arch_regname (i
), reg_locations
[i
]));
900 precise_regmask
&= ~(1 << i
);
906 scanned
+= map
->end_offset
- map
->start_offset
;
908 g_assert (scanned
== scanned_precisely
+ scanned_conservatively
);
910 stack_limit
= frame_end
;
912 /* Save information for the precise pass */
913 fi
= &tls
->frames
[tls
->nframes
];
914 fi
->nslots
= map
->nslots
;
915 bitmap_width
= ALIGN_TO (map
->nslots
, 8) / 8;
916 if (map
->has_ref_slots
)
917 fi
->bitmap
= &bitmaps
[map
->stack_ref_bitmap_offset
+ (bitmap_width
* cindex
)];
920 fi
->frame_start_offset
= frame_start
- stack_start
;
921 fi
->nreg_locations
= 0;
922 DEBUG_PRECISE (fi
->ji
= ji
);
923 DEBUG_PRECISE (fi
->fp
= fp
);
925 if (map
->has_ref_regs
) {
926 int bitmap_width
= ALIGN_TO (map
->nref_regs
, 8) / 8;
927 guint8
*ref_bitmap
= &bitmaps
[map
->reg_ref_bitmap_offset
+ (bitmap_width
* cindex
)];
929 for (i
= 0; i
< NREGS
; ++i
) {
930 if (!(map
->reg_ref_mask
& (1 << i
)))
933 if (reg_locations
[i
] && (ref_bitmap
[bindex
/ 8] & (1 << (bindex
% 8)))) {
934 DEBUG_PRECISE (fi
->regs
[fi
->nreg_locations
] = i
);
935 DEBUG (fprintf (logfile
, "\treg %s saved at 0x%p is ref.\n", mono_arch_regname (i
), reg_locations
[i
]));
936 fi
->reg_locations
[fi
->nreg_locations
] = (guint8
*)reg_locations
[i
] - stack_start
;
937 fi
->nreg_locations
++;
944 * Clear locations of precisely stacked registers.
946 if (precise_regmask
) {
947 for (i
= 0; i
< NREGS
; ++i
) {
948 if (precise_regmask
& (1 << i
)) {
950 * The method uses this register, and we have precise info for it.
951 * This means the location will be scanned precisely.
952 * Tell the code at the beginning of the loop that this location is
955 if (reg_locations
[i
])
956 DEBUG (fprintf (logfile
, "\treg %s at location %p (==%p) is precise.\n", mono_arch_regname (i
), reg_locations
[i
], (gpointer
)*reg_locations
[i
]));
957 reg_locations
[i
] = NULL
;
965 /* Scan the remaining register save locations */
966 for (i
= 0; i
< MONO_MAX_IREGS
; ++i
) {
967 if (reg_locations
[i
]) {
968 DEBUG (fprintf (logfile
, "\tscan saved reg location %p.\n", reg_locations
[i
]));
969 mono_gc_conservatively_scan_area (reg_locations
[i
], reg_locations
[i
] + sizeof (mgreg_t
));
970 scanned_registers
+= sizeof (mgreg_t
);
972 if (new_reg_locations
[i
]) {
973 DEBUG (fprintf (logfile
, "\tscan saved reg location %p.\n", new_reg_locations
[i
]));
974 mono_gc_conservatively_scan_area (new_reg_locations
[i
], new_reg_locations
[i
] + sizeof (mgreg_t
));
975 scanned_registers
+= sizeof (mgreg_t
);
979 if (stack_limit
< stack_end
) {
980 DEBUG (fprintf (logfile
, "\tscan remaining stack %p-%p (%d).\n", stack_limit
, stack_end
, (int)(stack_end
- stack_limit
)));
981 mono_gc_conservatively_scan_area (stack_limit
, stack_end
);
982 stats
.scanned_native
+= stack_end
- stack_limit
;
985 DEBUG (fprintf (logfile
, "Marked %d bytes, p=%d,c=%d out of %d.\n", scanned
, scanned_precisely
, scanned_conservatively
, (int)(stack_end
- stack_start
)));
987 stats
.scanned_stacks
+= stack_end
- stack_start
;
988 stats
.scanned
+= scanned
;
989 stats
.scanned_precisely
+= scanned_precisely
;
990 stats
.scanned_conservatively
+= scanned_conservatively
;
991 stats
.scanned_registers
+= scanned_registers
;
993 //mono_gc_conservatively_scan_area (stack_start, stack_end);
999 * Mark a thread stack precisely based on information saved during the conservative
1003 precise_pass (TlsData
*tls
, guint8
*stack_start
, guint8
*stack_end
)
1007 guint8
*frame_start
;
1012 for (findex
= 0; findex
< tls
->nframes
; findex
++) {
1013 /* Load information saved by the !precise pass */
1014 fi
= &tls
->frames
[findex
];
1015 frame_start
= stack_start
+ fi
->frame_start_offset
;
1017 DEBUG (char *fname
= mono_method_full_name (fi
->ji
->method
, TRUE
); fprintf (logfile
, "Mark(1): %s\n", fname
); g_free (fname
));
1020 * FIXME: Add a function to mark using a bitmap, to avoid doing a
1021 * call for each object.
1024 /* Mark stack slots */
1026 guint8
*ref_bitmap
= fi
->bitmap
;
1029 for (i
= 0; i
< fi
->nslots
; ++i
) {
1030 MonoObject
**ptr
= (MonoObject
**)(frame_start
+ (i
* sizeof (mgreg_t
)));
1032 live
= ref_bitmap
[i
/ 8] & (1 << (i
% 8));
1035 MonoObject
*obj
= *ptr
;
1037 DEBUG (fprintf (logfile
, "\tref %s0x%x(fp)=%p: %p ->", (guint8
*)ptr
>= (guint8
*)fi
->fp
? "" : "-", ABS ((int)((gssize
)ptr
- (gssize
)fi
->fp
)), ptr
, obj
));
1038 *ptr
= mono_gc_scan_object (obj
);
1039 DEBUG (fprintf (logfile
, " %p.\n", *ptr
));
1041 DEBUG (fprintf (logfile
, "\tref %s0x%x(fp)=%p: %p.\n", (guint8
*)ptr
>= (guint8
*)fi
->fp
? "" : "-", ABS ((int)((gssize
)ptr
- (gssize
)fi
->fp
)), ptr
, obj
));
1046 * This is disabled because the pointer takes up a lot of space.
1047 * Stack slots might be shared between ref and non-ref variables ?
1049 if (map
->ref_slots
[i
/ 8] & (1 << (i
% 8))) {
1050 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
));
1052 * Fail fast if the live range is incorrect, and
1053 * the JITted code tries to access this object
1062 /* Mark registers */
1065 * Registers are different from stack slots, they have no address where they
1066 * are stored. Instead, some frame below this frame in the stack saves them
1067 * in its prolog to the stack. We can mark this location precisely.
1069 for (i
= 0; i
< fi
->nreg_locations
; ++i
) {
1071 * reg_locations [i] contains the address of the stack slot where
1072 * a reg was last saved, so mark that slot.
1074 MonoObject
**ptr
= (MonoObject
**)((guint8
*)stack_start
+ fi
->reg_locations
[i
]);
1075 MonoObject
*obj
= *ptr
;
1078 DEBUG (fprintf (logfile
, "\treg %s saved at %p: %p ->", mono_arch_regname (fi
->regs
[i
]), ptr
, obj
));
1079 *ptr
= mono_gc_scan_object (obj
);
1080 DEBUG (fprintf (logfile
, " %p.\n", *ptr
));
1082 DEBUG (fprintf (logfile
, "\treg %s saved at %p: %p\n", mono_arch_regname (fi
->regs
[i
]), ptr
, obj
));
1088 * Debugging aid to check for missed refs.
1090 if (tls
->ref_to_track
) {
1093 for (p
= (mgreg_t
*)stack_start
; p
< (mgreg_t
*)stack_end
; ++p
)
1094 if (*p
== (mgreg_t
)tls
->ref_to_track
)
1095 printf ("REF AT %p.\n", p
);
1102 * This is called by the GC twice to mark a thread stack. PRECISE is FALSE at the first
1103 * call, and TRUE at the second. USER_DATA points to a TlsData
1104 * structure filled up by thread_suspend_func.
1107 thread_mark_func (gpointer user_data
, guint8
*stack_start
, guint8
*stack_end
, gboolean precise
)
1109 TlsData
*tls
= user_data
;
1111 DEBUG (fprintf (logfile
, "****************************************\n"));
1112 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
));
1113 DEBUG (fprintf (logfile
, "****************************************\n"));
1116 conservative_pass (tls
, stack_start
, stack_end
);
1118 precise_pass (tls
, stack_start
, stack_end
);
1122 mini_gc_init_gc_map (MonoCompile
*cfg
)
1124 if (COMPILE_LLVM (cfg
))
1127 if (!mono_gc_is_moving ())
1130 if (!cfg
->compile_aot
&& !mono_gc_precise_stack_mark_enabled ())
1134 /* Debugging support */
1136 static int precise_count
;
1139 if (getenv ("MONO_GCMAP_COUNT")) {
1140 if (precise_count
== atoi (getenv ("MONO_GCMAP_COUNT")))
1141 printf ("LAST: %s\n", mono_method_full_name (cfg
->method
, TRUE
));
1142 if (precise_count
> atoi (getenv ("MONO_GCMAP_COUNT")))
1148 cfg
->compute_gc_maps
= TRUE
;
1150 cfg
->gc_info
= mono_mempool_alloc0 (cfg
->mempool
, sizeof (MonoCompileGC
));
1154 * mini_gc_set_slot_type_from_fp:
1156 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1157 * relative to the frame pointer. By default, all stack slots are type PIN, so there is no
1158 * need to call this function for those slots.
1161 mini_gc_set_slot_type_from_fp (MonoCompile
*cfg
, int slot_offset
, GCSlotType type
)
1163 MonoCompileGC
*gcfg
= (MonoCompileGC
*)cfg
->gc_info
;
1165 if (!cfg
->compute_gc_maps
)
1168 g_assert (slot_offset
% sizeof (mgreg_t
) == 0);
1170 gcfg
->stack_slots_from_fp
= g_slist_prepend_mempool (cfg
->mempool
, gcfg
->stack_slots_from_fp
, GINT_TO_POINTER (((slot_offset
) << 16) | type
));
1174 * mini_gc_set_slot_type_from_cfa:
1176 * Set the GC slot type of the stack slot identified by SLOT_OFFSET, which should be
1177 * relative to the DWARF CFA value. This should be called from mono_arch_emit_prolog ().
1178 * If type is STACK_REF, the slot is assumed to be live from the end of the prolog until
1179 * the end of the method. By default, all stack slots are type PIN, so there is no need to
1180 * call this function for those slots.
1183 mini_gc_set_slot_type_from_cfa (MonoCompile
*cfg
, int slot_offset
, GCSlotType type
)
1185 MonoCompileGC
*gcfg
= (MonoCompileGC
*)cfg
->gc_info
;
1186 int slot
= - (slot_offset
/ sizeof (mgreg_t
));
1188 if (!cfg
->compute_gc_maps
)
1191 g_assert (slot_offset
<= 0);
1192 g_assert (slot_offset
% sizeof (mgreg_t
) == 0);
1194 gcfg
->stack_slots_from_cfa
= g_slist_prepend_mempool (cfg
->mempool
, gcfg
->stack_slots_from_cfa
, GUINT_TO_POINTER (((slot
) << 16) | type
));
1198 fp_offset_to_slot (MonoCompile
*cfg
, int offset
)
1200 MonoCompileGC
*gcfg
= cfg
->gc_info
;
1202 return (offset
- gcfg
->min_offset
) / sizeof (mgreg_t
);
1206 slot_to_fp_offset (MonoCompile
*cfg
, int slot
)
1208 MonoCompileGC
*gcfg
= cfg
->gc_info
;
1210 return (slot
* sizeof (mgreg_t
)) + gcfg
->min_offset
;
1214 set_slot (MonoCompileGC
*gcfg
, int slot
, int callsite_index
, GCSlotType type
)
1216 g_assert (slot
>= 0 && slot
< gcfg
->nslots
);
1218 if (type
== SLOT_PIN
) {
1219 clear_bit (gcfg
->stack_ref_bitmap
, gcfg
->stack_bitmap_width
, callsite_index
, slot
);
1220 set_bit (gcfg
->stack_pin_bitmap
, gcfg
->stack_bitmap_width
, callsite_index
, slot
);
1221 } else if (type
== SLOT_REF
) {
1222 set_bit (gcfg
->stack_ref_bitmap
, gcfg
->stack_bitmap_width
, callsite_index
, slot
);
1223 clear_bit (gcfg
->stack_pin_bitmap
, gcfg
->stack_bitmap_width
, callsite_index
, slot
);
1224 } else if (type
== SLOT_NOREF
) {
1225 clear_bit (gcfg
->stack_ref_bitmap
, gcfg
->stack_bitmap_width
, callsite_index
, slot
);
1226 clear_bit (gcfg
->stack_pin_bitmap
, gcfg
->stack_bitmap_width
, callsite_index
, slot
);
1231 set_slot_everywhere (MonoCompileGC
*gcfg
, int slot
, GCSlotType type
)
1235 for (cindex
= 0; cindex
< gcfg
->ncallsites
; ++cindex
)
1236 set_slot (gcfg
, slot
, cindex
, type
);
1240 set_slot_in_range (MonoCompileGC
*gcfg
, int slot
, int from
, int to
, GCSlotType type
)
1244 for (cindex
= 0; cindex
< gcfg
->ncallsites
; ++cindex
) {
1245 int callsite_offset
= gcfg
->callsites
[cindex
]->pc_offset
;
1246 if (callsite_offset
>= from
&& callsite_offset
< to
)
1247 set_slot (gcfg
, slot
, cindex
, type
);
1252 set_reg_slot (MonoCompileGC
*gcfg
, int slot
, int callsite_index
, GCSlotType type
)
1254 g_assert (slot
>= 0 && slot
< gcfg
->nregs
);
1256 if (type
== SLOT_PIN
) {
1257 clear_bit (gcfg
->reg_ref_bitmap
, gcfg
->reg_bitmap_width
, callsite_index
, slot
);
1258 set_bit (gcfg
->reg_pin_bitmap
, gcfg
->reg_bitmap_width
, callsite_index
, slot
);
1259 } else if (type
== SLOT_REF
) {
1260 set_bit (gcfg
->reg_ref_bitmap
, gcfg
->reg_bitmap_width
, callsite_index
, slot
);
1261 clear_bit (gcfg
->reg_pin_bitmap
, gcfg
->reg_bitmap_width
, callsite_index
, slot
);
1262 } else if (type
== SLOT_NOREF
) {
1263 clear_bit (gcfg
->reg_ref_bitmap
, gcfg
->reg_bitmap_width
, callsite_index
, slot
);
1264 clear_bit (gcfg
->reg_pin_bitmap
, gcfg
->reg_bitmap_width
, callsite_index
, slot
);
1269 set_reg_slot_everywhere (MonoCompileGC
*gcfg
, int slot
, GCSlotType type
)
1273 for (cindex
= 0; cindex
< gcfg
->ncallsites
; ++cindex
)
1274 set_reg_slot (gcfg
, slot
, cindex
, type
);
1278 set_reg_slot_in_range (MonoCompileGC
*gcfg
, int slot
, int from
, int to
, GCSlotType type
)
1282 for (cindex
= 0; cindex
< gcfg
->ncallsites
; ++cindex
) {
1283 int callsite_offset
= gcfg
->callsites
[cindex
]->pc_offset
;
1284 if (callsite_offset
>= from
&& callsite_offset
< to
)
1285 set_reg_slot (gcfg
, slot
, cindex
, type
);
1290 process_spill_slots (MonoCompile
*cfg
)
1292 MonoCompileGC
*gcfg
= cfg
->gc_info
;
1297 /* Mark all ref/pin spill slots as NOREF by default outside of their live range */
1298 for (bb
= cfg
->bb_entry
; bb
; bb
= bb
->next_bb
) {
1299 for (l
= bb
->spill_slot_defs
; l
; l
= l
->next
) {
1300 MonoInst
*def
= l
->data
;
1301 int spill_slot
= def
->inst_c0
;
1302 int bank
= def
->inst_c1
;
1303 int offset
= cfg
->spill_info
[bank
][spill_slot
].offset
;
1304 int slot
= fp_offset_to_slot (cfg
, offset
);
1306 if (bank
== MONO_REG_INT_MP
|| bank
== MONO_REG_INT_REF
)
1307 set_slot_everywhere (gcfg
, slot
, SLOT_NOREF
);
1311 for (bb
= cfg
->bb_entry
; bb
; bb
= bb
->next_bb
) {
1312 for (l
= bb
->spill_slot_defs
; l
; l
= l
->next
) {
1313 MonoInst
*def
= l
->data
;
1314 int spill_slot
= def
->inst_c0
;
1315 int bank
= def
->inst_c1
;
1316 int offset
= cfg
->spill_info
[bank
][spill_slot
].offset
;
1317 int slot
= fp_offset_to_slot (cfg
, offset
);
1320 if (bank
== MONO_REG_INT_MP
)
1326 * Extend the live interval for the GC tracked spill slots
1327 * defined in this bblock.
1328 * FIXME: This is not needed.
1330 set_slot_in_range (gcfg
, slot
, def
->backend
.pc_offset
, bb
->native_offset
+ bb
->native_length
, type
);
1332 if (cfg
->verbose_level
> 1)
1333 printf ("\t%s spill slot at %s0x%x(fp) (slot = %d)\n", slot_type_to_string (type
), offset
>= 0 ? "" : "-", ABS (offset
), slot
);
1337 /* Set fp spill slots to NOREF */
1338 for (i
= 0; i
< cfg
->spill_info_len
[MONO_REG_DOUBLE
]; ++i
) {
1339 int offset
= cfg
->spill_info
[MONO_REG_DOUBLE
][i
].offset
;
1345 slot
= fp_offset_to_slot (cfg
, offset
);
1347 set_slot_everywhere (gcfg
, slot
, SLOT_NOREF
);
1349 if (cfg
->verbose_level
> 1)
1350 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset
>= 0 ? "" : "-", ABS (offset
), slot
);
1353 /* Set int spill slots to NOREF */
1354 for (i
= 0; i
< cfg
->spill_info_len
[MONO_REG_INT
]; ++i
) {
1355 int offset
= cfg
->spill_info
[MONO_REG_INT
][i
].offset
;
1361 slot
= fp_offset_to_slot (cfg
, offset
);
1363 set_slot_everywhere (gcfg
, slot
, SLOT_NOREF
);
1364 if (cfg
->verbose_level
> 1)
1365 printf ("\tint spill slot at %s0x%x(fp) (slot = %d)\n", offset
>= 0 ? "" : "-", ABS (offset
), slot
);
1370 * process_other_slots:
1372 * Process stack slots registered using mini_gc_set_slot_type_... ().
1375 process_other_slots (MonoCompile
*cfg
)
1377 MonoCompileGC
*gcfg
= cfg
->gc_info
;
1380 /* Relative to the CFA */
1381 for (l
= gcfg
->stack_slots_from_cfa
; l
; l
= l
->next
) {
1382 guint data
= GPOINTER_TO_UINT (l
->data
);
1383 int cfa_slot
= data
>> 16;
1384 GCSlotType type
= data
& 0xff;
1388 * Map the cfa relative slot to an fp relative slot.
1389 * slot_addr == cfa - <cfa_slot>*4/8
1390 * fp + cfa_offset == cfa
1391 * -> slot_addr == fp + (cfa_offset - <cfa_slot>*4/8)
1393 slot
= (cfg
->cfa_offset
/ sizeof (mgreg_t
)) - cfa_slot
- (gcfg
->min_offset
/ sizeof (mgreg_t
));
1395 set_slot_everywhere (gcfg
, slot
, type
);
1397 if (cfg
->verbose_level
> 1) {
1398 int fp_offset
= slot_to_fp_offset (cfg
, slot
);
1399 if (type
== SLOT_NOREF
)
1400 printf ("\tnoref slot at %s0x%x(fp) (slot = %d) (cfa - 0x%x)\n", fp_offset
>= 0 ? "" : "-", ABS (fp_offset
), slot
, (int)(cfa_slot
* sizeof (mgreg_t
)));
1404 /* Relative to the FP */
1405 for (l
= gcfg
->stack_slots_from_fp
; l
; l
= l
->next
) {
1406 gint data
= GPOINTER_TO_INT (l
->data
);
1407 int offset
= data
>> 16;
1408 GCSlotType type
= data
& 0xff;
1411 slot
= fp_offset_to_slot (cfg
, offset
);
1413 set_slot_everywhere (gcfg
, slot
, type
);
1415 /* Liveness for these slots is handled by process_spill_slots () */
1417 if (cfg
->verbose_level
> 1) {
1418 if (type
== SLOT_REF
)
1419 printf ("\tref slot at fp+0x%x (slot = %d)\n", offset
, slot
);
1425 process_variables (MonoCompile
*cfg
)
1427 MonoCompileGC
*gcfg
= cfg
->gc_info
;
1428 MonoMethodSignature
*sig
= mono_method_signature (cfg
->method
);
1429 int i
, locals_min_slot
, locals_max_slot
, cindex
;
1433 int locals_min_offset
= gcfg
->locals_min_offset
;
1434 int locals_max_offset
= gcfg
->locals_max_offset
;
1436 /* Slots for locals are NOREF by default */
1437 locals_min_slot
= (locals_min_offset
- gcfg
->min_offset
) / sizeof (mgreg_t
);
1438 locals_max_slot
= (locals_max_offset
- gcfg
->min_offset
) / sizeof (mgreg_t
);
1439 for (i
= locals_min_slot
; i
< locals_max_slot
; ++i
) {
1440 set_slot_everywhere (gcfg
, i
, SLOT_NOREF
);
1444 * Compute the offset where variables are initialized in the first bblock, if any.
1446 pc_offsets
= g_new0 (int, cfg
->next_vreg
);
1448 bb
= cfg
->bb_entry
->next_bb
;
1449 MONO_BB_FOR_EACH_INS (bb
, tmp
) {
1450 if (tmp
->opcode
== OP_GC_LIVENESS_DEF
) {
1451 int vreg
= tmp
->inst_c1
;
1452 if (pc_offsets
[vreg
] == 0) {
1453 g_assert (tmp
->backend
.pc_offset
> 0);
1454 pc_offsets
[vreg
] = tmp
->backend
.pc_offset
;
1460 * Stack slots holding arguments are initialized in the prolog.
1461 * This means we can treat them alive for the whole method.
1463 for (i
= 0; i
< cfg
->num_varinfo
; i
++) {
1464 MonoInst
*ins
= cfg
->varinfo
[i
];
1465 MonoType
*t
= ins
->inst_vtype
;
1468 gboolean byref
, is_this
= FALSE
;
1469 gboolean is_arg
= i
< cfg
->locals_start
;
1471 if (ins
== cfg
->ret
)
1474 vmv
= MONO_VARINFO (cfg
, i
);
1476 /* For some reason, 'this' is byref */
1477 if (sig
->hasthis
&& ins
== cfg
->args
[0] && !cfg
->method
->klass
->valuetype
) {
1478 t
= &cfg
->method
->klass
->byval_arg
;
1484 if (ins
->opcode
== OP_REGVAR
) {
1486 GCSlotType slot_type
;
1488 t
= mini_type_get_underlying_type (NULL
, t
);
1491 g_assert (hreg
< MONO_MAX_IREGS
);
1494 slot_type
= SLOT_PIN
;
1496 slot_type
= mini_type_is_reference (cfg
, t
) ? SLOT_REF
: SLOT_NOREF
;
1498 if (slot_type
== SLOT_PIN
) {
1499 /* These have no live interval, be conservative */
1500 set_reg_slot_everywhere (gcfg
, hreg
, slot_type
);
1503 * Unlike variables allocated to the stack, we generate liveness info
1504 * for noref vars in registers in mono_spill_global_vars (), because
1505 * knowing that a register doesn't contain a ref allows us to mark its save
1506 * locations precisely.
1508 for (cindex
= 0; cindex
< gcfg
->ncallsites
; ++cindex
)
1509 if (gcfg
->callsites
[cindex
]->liveness
[i
/ 8] & (1 << (i
% 8)))
1510 set_reg_slot (gcfg
, hreg
, cindex
, slot_type
);
1513 if (cfg
->verbose_level
> 1) {
1514 printf ("\t%s %sreg %s(R%d)\n", slot_type_to_string (slot_type
), is_arg
? "arg " : "", mono_arch_regname (hreg
), vmv
->vreg
);
1520 if (ins
->opcode
!= OP_REGOFFSET
)
1523 if (ins
->inst_offset
% sizeof (mgreg_t
) != 0)
1526 if (is_arg
&& ins
->inst_offset
>= gcfg
->max_offset
)
1527 /* In parent frame */
1530 pos
= fp_offset_to_slot (cfg
, ins
->inst_offset
);
1532 if (is_arg
&& ins
->flags
& MONO_INST_IS_DEAD
) {
1533 /* These do not get stored in the prolog */
1534 set_slot_everywhere (gcfg
, pos
, SLOT_NOREF
);
1536 if (cfg
->verbose_level
> 1) {
1537 printf ("\tdead arg at fp%s0x%x (slot = %d): %s\n", ins
->inst_offset
< 0 ? "-" : "+", (ins
->inst_offset
< 0) ? -(int)ins
->inst_offset
: (int)ins
->inst_offset
, pos
, mono_type_full_name (ins
->inst_vtype
));
1542 if (MONO_TYPE_ISSTRUCT (t
)) {
1544 gsize
*bitmap
= NULL
;
1545 gboolean pin
= FALSE
;
1549 if (ins
->backend
.is_pinvoke
)
1550 size
= mono_class_native_size (ins
->klass
, NULL
);
1552 size
= mono_class_value_size (ins
->klass
, NULL
);
1553 size_in_slots
= ALIGN_TO (size
, sizeof (mgreg_t
)) / sizeof (mgreg_t
);
1555 if (!ins
->klass
->has_references
) {
1557 for (j
= 0; j
< size_in_slots
; ++j
)
1558 set_slot_everywhere (gcfg
, pos
+ j
, SLOT_NOREF
);
1563 if (ins
->klass
->generic_container
|| mono_class_is_open_constructed_type (t
)) {
1564 /* FIXME: Generic sharing */
1567 mono_class_compute_gc_descriptor (ins
->klass
);
1569 bitmap
= mono_gc_get_bitmap_for_descr (ins
->klass
->gc_descr
, &numbits
);
1574 * Most vtypes are marked volatile because of the LDADDR instructions,
1575 * and they have no liveness information since they are decomposed
1576 * before the liveness pass. We emit OP_GC_LIVENESS_DEF instructions for
1577 * them during VZERO decomposition.
1579 if (!pc_offsets
[vmv
->vreg
])
1583 if (ins
->backend
.is_pinvoke
)
1586 if (cfg
->verbose_level
> 1)
1587 printf ("\tvtype R%d at fp+0x%x-0x%x: %s\n", vmv
->vreg
, (int)ins
->inst_offset
, (int)(ins
->inst_offset
+ (size
/ sizeof (mgreg_t
))), mono_type_full_name (ins
->inst_vtype
));
1590 for (cindex
= 0; cindex
< gcfg
->ncallsites
; ++cindex
) {
1591 if (gcfg
->callsites
[cindex
]->pc_offset
> pc_offsets
[vmv
->vreg
]) {
1592 for (j
= 0; j
< numbits
; ++j
) {
1593 if (bitmap
[j
/ GC_BITS_PER_WORD
] & ((gsize
)1 << (j
% GC_BITS_PER_WORD
))) {
1594 /* The descriptor is for the boxed object */
1595 set_slot (gcfg
, (pos
+ j
- (sizeof (MonoObject
) / sizeof (mgreg_t
))), cindex
, pin
? SLOT_PIN
: SLOT_REF
);
1601 if (cfg
->verbose_level
> 1) {
1602 for (j
= 0; j
< numbits
; ++j
) {
1603 if (bitmap
[j
/ GC_BITS_PER_WORD
] & ((gsize
)1 << (j
% GC_BITS_PER_WORD
)))
1604 printf ("\t\t%s slot at 0x%x(fp) (slot = %d)\n", pin
? "pin" : "ref", (int)(ins
->inst_offset
+ (j
* sizeof (mgreg_t
))), (int)(pos
+ j
- (sizeof (MonoObject
) / sizeof (mgreg_t
))));
1608 if (cfg
->verbose_level
> 1)
1609 printf ("\t\tpinned\n");
1610 for (j
= 0; j
< size_in_slots
; ++j
) {
1611 set_slot_everywhere (gcfg
, pos
+ j
, SLOT_PIN
);
1620 if (!is_arg
&& (ins
->inst_offset
< gcfg
->min_offset
|| ins
->inst_offset
>= gcfg
->max_offset
))
1621 /* Vret addr etc. */
1626 set_slot_everywhere (gcfg
, pos
, SLOT_PIN
);
1628 for (cindex
= 0; cindex
< gcfg
->ncallsites
; ++cindex
)
1629 if (gcfg
->callsites
[cindex
]->liveness
[i
/ 8] & (1 << (i
% 8)))
1630 set_slot (gcfg
, pos
, cindex
, SLOT_PIN
);
1632 if (cfg
->verbose_level
> 1)
1633 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
));
1638 * This is currently disabled, but could be enabled to debug crashes.
1641 if (t
->type
== MONO_TYPE_I
) {
1643 * Variables created in mono_handle_global_vregs have type I, but they
1644 * could hold GC refs since the vregs they were created from might not been
1645 * marked as holding a GC ref. So be conservative.
1647 set_slot_everywhere (gcfg
, pos
, SLOT_PIN
);
1652 t
= mini_type_get_underlying_type (NULL
, t
);
1654 if (!mini_type_is_reference (cfg
, t
)) {
1655 set_slot_everywhere (gcfg
, pos
, SLOT_NOREF
);
1656 if (cfg
->verbose_level
> 1)
1657 printf ("\tnoref 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
));
1661 /* 'this' is marked INDIRECT for gshared methods */
1662 if (ins
->flags
& (MONO_INST_VOLATILE
| MONO_INST_INDIRECT
) && !is_this
) {
1664 * For volatile variables, treat them alive from the point they are
1665 * initialized in the first bblock until the end of the method.
1668 set_slot_everywhere (gcfg
, pos
, SLOT_REF
);
1669 } else if (pc_offsets
[vmv
->vreg
]) {
1670 set_slot_in_range (gcfg
, pos
, 0, pc_offsets
[vmv
->vreg
], SLOT_PIN
);
1671 set_slot_in_range (gcfg
, pos
, pc_offsets
[vmv
->vreg
], cfg
->code_size
, SLOT_REF
);
1673 set_slot_everywhere (gcfg
, pos
, SLOT_PIN
);
1675 if (cfg
->verbose_level
> 1)
1676 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
));
1681 /* Live for the whole method */
1682 set_slot_everywhere (gcfg
, pos
, SLOT_REF
);
1684 for (cindex
= 0; cindex
< gcfg
->ncallsites
; ++cindex
)
1685 if (gcfg
->callsites
[cindex
]->liveness
[i
/ 8] & (1 << (i
% 8)))
1686 set_slot (gcfg
, pos
, cindex
, SLOT_REF
);
1689 if (cfg
->verbose_level
> 1) {
1690 printf ("\tref 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
));
1694 g_free (pc_offsets
);
1698 sp_offset_to_fp_offset (MonoCompile
*cfg
, int sp_offset
)
1701 * Convert a sp relative offset to a slot index. This is
1702 * platform specific.
1705 /* fp = sp + offset */
1706 g_assert (cfg
->frame_reg
== AMD64_RBP
);
1707 return (- cfg
->arch
.sp_fp_offset
+ sp_offset
);
1715 type_to_gc_slot_type (MonoCompile
*cfg
, MonoType
*t
)
1719 t
= mini_type_get_underlying_type (NULL
, t
);
1720 if (mini_type_is_reference (cfg
, t
))
1723 if (MONO_TYPE_ISSTRUCT (t
)) {
1724 MonoClass
*klass
= mono_class_from_mono_type (t
);
1725 if (!klass
->has_references
) {
1737 process_param_area_slots (MonoCompile
*cfg
)
1739 MonoCompileGC
*gcfg
= cfg
->gc_info
;
1744 * These slots are used for passing parameters during calls. They are sp relative, not
1745 * fp relative, so they are harder to handle.
1747 if (cfg
->flags
& MONO_CFG_HAS_ALLOCA
)
1748 /* The distance between fp and sp is not constant */
1751 is_param
= mono_mempool_alloc0 (cfg
->mempool
, gcfg
->nslots
* sizeof (gboolean
));
1753 for (i
= 0; i
< gcfg
->ncallsites
; ++i
) {
1754 GCCallSite
*callsite
= gcfg
->callsites
[i
];
1757 for (l
= callsite
->param_slots
; l
; l
= l
->next
) {
1758 MonoInst
*def
= l
->data
;
1759 int sp_offset
= def
->inst_offset
;
1760 int fp_offset
= sp_offset_to_fp_offset (cfg
, sp_offset
);
1761 int slot
= fp_offset_to_slot (cfg
, fp_offset
);
1763 g_assert (slot
>= 0 && slot
< gcfg
->nslots
);
1764 is_param
[slot
] = TRUE
;
1768 /* All param area slots are noref by default */
1769 for (i
= 0; i
< gcfg
->nslots
; ++i
) {
1771 set_slot_everywhere (gcfg
, i
, SLOT_NOREF
);
1774 for (i
= 0; i
< gcfg
->ncallsites
; ++i
) {
1775 GCCallSite
*callsite
= gcfg
->callsites
[i
];
1778 for (l
= callsite
->param_slots
; l
; l
= l
->next
) {
1779 MonoInst
*def
= l
->data
;
1780 MonoType
*t
= def
->inst_vtype
;
1781 int sp_offset
= def
->inst_offset
;
1782 int fp_offset
= sp_offset_to_fp_offset (cfg
, sp_offset
);
1783 int slot
= fp_offset_to_slot (cfg
, fp_offset
);
1784 GCSlotType type
= type_to_gc_slot_type (cfg
, t
);
1786 /* The slot is live between the def instruction and the call */
1787 set_slot_in_range (gcfg
, slot
, def
->backend
.pc_offset
, callsite
->pc_offset
+ 1, type
);
1788 if (cfg
->verbose_level
> 1)
1789 printf ("\t%s param area slot at %s0x%x(fp)=0x%x(sp) (slot = %d) [0x%x-0x%x]\n", slot_type_to_string (type
), fp_offset
>= 0 ? "+" : "-", ABS (fp_offset
), sp_offset
, slot
, def
->backend
.pc_offset
, callsite
->pc_offset
+ 1);
1795 process_finally_clauses (MonoCompile
*cfg
)
1797 MonoCompileGC
*gcfg
= cfg
->gc_info
;
1798 GCCallSite
**callsites
;
1800 gboolean has_finally
;
1801 int i
, j
, nslots
, nregs
;
1803 ncallsites
= gcfg
->ncallsites
;
1804 nslots
= gcfg
->nslots
;
1805 nregs
= gcfg
->nregs
;
1806 callsites
= gcfg
->callsites
;
1809 * The calls to the finally clauses don't show up in the cfg. See
1810 * test_0_liveness_8 ().
1811 * Variables accessed inside the finally clause are already marked VOLATILE by
1812 * mono_liveness_handle_exception_clauses (). Variables not accessed inside the finally clause have
1813 * correct liveness outside the finally clause. So mark them PIN inside the finally clauses.
1815 has_finally
= FALSE
;
1816 for (i
= 0; i
< cfg
->header
->num_clauses
; ++i
) {
1817 MonoExceptionClause
*clause
= &cfg
->header
->clauses
[i
];
1819 if (clause
->flags
== MONO_EXCEPTION_CLAUSE_FINALLY
) {
1824 if (cfg
->verbose_level
> 1)
1825 printf ("\tMethod has finally clauses, pessimizing live ranges.\n");
1826 for (j
= 0; j
< ncallsites
; ++j
) {
1827 MonoBasicBlock
*bb
= callsites
[j
]->bb
;
1828 MonoExceptionClause
*clause
;
1829 gboolean is_in_finally
= FALSE
;
1831 for (i
= 0; i
< cfg
->header
->num_clauses
; ++i
) {
1832 clause
= &cfg
->header
->clauses
[i
];
1834 if (MONO_OFFSET_IN_HANDLER (clause
, bb
->real_offset
)) {
1835 if (clause
->flags
== MONO_EXCEPTION_CLAUSE_FINALLY
) {
1836 is_in_finally
= TRUE
;
1842 if (is_in_finally
) {
1843 for (i
= 0; i
< nslots
; ++i
)
1844 set_slot (gcfg
, i
, j
, SLOT_PIN
);
1845 for (i
= 0; i
< nregs
; ++i
)
1846 set_reg_slot (gcfg
, i
, j
, SLOT_PIN
);
1853 compute_frame_size (MonoCompile
*cfg
)
1855 int i
, locals_min_offset
, locals_max_offset
, cfa_min_offset
, cfa_max_offset
;
1856 int min_offset
, max_offset
;
1857 MonoCompileGC
*gcfg
= cfg
->gc_info
;
1858 MonoMethodSignature
*sig
= mono_method_signature (cfg
->method
);
1861 /* Compute min/max offsets from the fp */
1864 #if defined(TARGET_AMD64) || defined(TARGET_X86)
1865 locals_min_offset
= ALIGN_TO (cfg
->locals_min_stack_offset
, sizeof (mgreg_t
));
1866 locals_max_offset
= cfg
->locals_max_stack_offset
;
1868 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
1872 locals_min_offset
= ALIGN_TO (locals_min_offset
, sizeof (mgreg_t
));
1873 locals_max_offset
= ALIGN_TO (locals_max_offset
, sizeof (mgreg_t
));
1875 min_offset
= locals_min_offset
;
1876 max_offset
= locals_max_offset
;
1879 for (i
= 0; i
< sig
->param_count
+ sig
->hasthis
; ++i
) {
1880 MonoInst
*ins
= cfg
->args
[i
];
1882 if (ins
->opcode
== OP_REGOFFSET
)
1883 min_offset
= MIN (min_offset
, ins
->inst_offset
);
1887 g_assert (cfg
->frame_reg
== cfg
->cfa_reg
);
1888 g_assert (cfg
->cfa_offset
> 0);
1890 cfa_max_offset
= cfg
->cfa_offset
;
1892 min_offset
= MIN (min_offset
, cfa_min_offset
);
1893 max_offset
= MAX (max_offset
, cfa_max_offset
);
1895 /* Fp relative slots */
1896 for (l
= gcfg
->stack_slots_from_fp
; l
; l
= l
->next
) {
1897 gint data
= GPOINTER_TO_INT (l
->data
);
1898 int offset
= data
>> 16;
1900 min_offset
= MIN (min_offset
, offset
);
1904 if (!(cfg
->flags
& MONO_CFG_HAS_SPILLUP
)) {
1905 int stack_offset
= ALIGN_TO (cfg
->stack_offset
, sizeof (mgreg_t
));
1906 min_offset
= MIN (min_offset
, (-stack_offset
));
1909 /* Param area slots */
1911 min_offset
= MIN (min_offset
, -cfg
->arch
.sp_fp_offset
);
1914 gcfg
->min_offset
= min_offset
;
1915 gcfg
->max_offset
= max_offset
;
1916 gcfg
->locals_min_offset
= locals_min_offset
;
1917 gcfg
->locals_max_offset
= locals_max_offset
;
1921 init_gcfg (MonoCompile
*cfg
)
1923 int i
, nregs
, nslots
;
1924 MonoCompileGC
*gcfg
= cfg
->gc_info
;
1925 GCCallSite
**callsites
;
1934 for (bb
= cfg
->bb_entry
; bb
; bb
= bb
->next_bb
) {
1935 ncallsites
+= g_slist_length (bb
->gc_callsites
);
1937 callsites
= mono_mempool_alloc0 (cfg
->mempool
, ncallsites
* sizeof (GCCallSite
*));
1939 for (bb
= cfg
->bb_entry
; bb
; bb
= bb
->next_bb
) {
1940 for (l
= bb
->gc_callsites
; l
; l
= l
->next
)
1941 callsites
[i
++] = l
->data
;
1944 /* The callsites should already be ordered by pc offset */
1945 for (i
= 1; i
< ncallsites
; ++i
)
1946 g_assert (callsites
[i
- 1]->pc_offset
< callsites
[i
]->pc_offset
);
1949 * The stack frame looks like this:
1951 * <fp + max_offset> == cfa -> <end of previous frame>
1952 * <other stack slots>
1954 * <other stack slots>
1955 * fp + min_offset ->
1960 if (cfg
->verbose_level
> 1)
1961 printf ("GC Map for %s: 0x%x-0x%x\n", mono_method_full_name (cfg
->method
, TRUE
), gcfg
->min_offset
, gcfg
->max_offset
);
1963 nslots
= (gcfg
->max_offset
- gcfg
->min_offset
) / sizeof (mgreg_t
);
1966 gcfg
->nslots
= nslots
;
1967 gcfg
->nregs
= nregs
;
1968 gcfg
->callsites
= callsites
;
1969 gcfg
->ncallsites
= ncallsites
;
1970 gcfg
->stack_bitmap_width
= ALIGN_TO (nslots
, 8) / 8;
1971 gcfg
->reg_bitmap_width
= ALIGN_TO (nregs
, 8) / 8;
1972 gcfg
->stack_ref_bitmap
= mono_mempool_alloc0 (cfg
->mempool
, gcfg
->stack_bitmap_width
* ncallsites
);
1973 gcfg
->stack_pin_bitmap
= mono_mempool_alloc0 (cfg
->mempool
, gcfg
->stack_bitmap_width
* ncallsites
);
1974 gcfg
->reg_ref_bitmap
= mono_mempool_alloc0 (cfg
->mempool
, gcfg
->reg_bitmap_width
* ncallsites
);
1975 gcfg
->reg_pin_bitmap
= mono_mempool_alloc0 (cfg
->mempool
, gcfg
->reg_bitmap_width
* ncallsites
);
1977 /* All slots start out as PIN */
1978 memset (gcfg
->stack_pin_bitmap
, 0xff, gcfg
->stack_bitmap_width
* ncallsites
);
1979 for (i
= 0; i
< nregs
; ++i
) {
1981 * By default, registers are NOREF.
1982 * It is possible for a callee to save them before being defined in this method,
1983 * but the saved value is dead too, so it doesn't need to be marked.
1985 if ((cfg
->used_int_regs
& (1 << i
)))
1986 set_reg_slot_everywhere (gcfg
, i
, SLOT_NOREF
);
1991 create_map (MonoCompile
*cfg
)
1994 int i
, j
, nregs
, nslots
, nref_regs
, npin_regs
, alloc_size
, bitmaps_size
, bitmaps_offset
;
1996 int stack_bitmap_width
, stack_bitmap_size
, reg_ref_bitmap_width
, reg_ref_bitmap_size
;
1997 int reg_pin_bitmap_width
, reg_pin_bitmap_size
, bindex
;
1999 gboolean has_ref_slots
, has_pin_slots
, has_ref_regs
, has_pin_regs
;
2000 MonoCompileGC
*gcfg
= cfg
->gc_info
;
2001 GCCallSite
**callsites
;
2003 guint8
*bitmap
, *bitmaps
;
2004 guint32 reg_ref_mask
, reg_pin_mask
;
2006 ncallsites
= gcfg
->ncallsites
;
2007 nslots
= gcfg
->nslots
;
2008 nregs
= gcfg
->nregs
;
2009 callsites
= gcfg
->callsites
;
2012 * Compute the real size of the bitmap i.e. ignore NOREF columns at the beginning and at
2013 * the end. Also, compute whenever the map needs ref/pin bitmaps, and collect stats.
2015 has_ref_slots
= FALSE
;
2016 has_pin_slots
= FALSE
;
2019 memset (ntypes
, 0, sizeof (ntypes
));
2020 for (i
= 0; i
< nslots
; ++i
) {
2021 gboolean has_ref
= FALSE
;
2022 gboolean has_pin
= FALSE
;
2024 for (j
= 0; j
< ncallsites
; ++j
) {
2025 if (get_bit (gcfg
->stack_pin_bitmap
, gcfg
->stack_bitmap_width
, j
, i
))
2027 if (get_bit (gcfg
->stack_ref_bitmap
, gcfg
->stack_bitmap_width
, j
, i
))
2032 has_ref_slots
= TRUE
;
2034 has_pin_slots
= TRUE
;
2037 ntypes
[SLOT_REF
] ++;
2039 ntypes
[SLOT_PIN
] ++;
2041 ntypes
[SLOT_NOREF
] ++;
2043 if (has_ref
|| has_pin
) {
2050 start
= end
= nslots
;
2052 g_assert (start
!= -1);
2053 g_assert (start
< end
);
2056 has_ref_regs
= FALSE
;
2057 has_pin_regs
= FALSE
;
2062 for (i
= 0; i
< nregs
; ++i
) {
2063 gboolean has_ref
= FALSE
;
2064 gboolean has_pin
= FALSE
;
2066 if (!(cfg
->used_int_regs
& (1 << i
)))
2069 for (j
= 0; j
< ncallsites
; ++j
) {
2070 if (get_bit (gcfg
->reg_ref_bitmap
, gcfg
->reg_bitmap_width
, j
, i
)) {
2075 for (j
= 0; j
< ncallsites
; ++j
) {
2076 if (get_bit (gcfg
->reg_pin_bitmap
, gcfg
->reg_bitmap_width
, j
, i
)) {
2083 reg_ref_mask
|= (1 << i
);
2084 has_ref_regs
= TRUE
;
2088 reg_pin_mask
|= (1 << i
);
2089 has_pin_regs
= TRUE
;
2094 if (cfg
->verbose_level
> 1)
2095 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
);
2097 /* Create the GC Map */
2099 stack_bitmap_width
= ALIGN_TO (end
- start
, 8) / 8;
2100 stack_bitmap_size
= stack_bitmap_width
* ncallsites
;
2101 reg_ref_bitmap_width
= ALIGN_TO (nref_regs
, 8) / 8;
2102 reg_ref_bitmap_size
= reg_ref_bitmap_width
* ncallsites
;
2103 reg_pin_bitmap_width
= ALIGN_TO (npin_regs
, 8) / 8;
2104 reg_pin_bitmap_size
= reg_pin_bitmap_width
* ncallsites
;
2105 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);
2107 map
= mono_mempool_alloc0 (cfg
->mempool
, sizeof (GCMap
));
2109 map
->frame_reg
= cfg
->frame_reg
;
2110 map
->start_offset
= gcfg
->min_offset
;
2111 map
->end_offset
= gcfg
->min_offset
+ (nslots
* sizeof (mgreg_t
));
2112 map
->map_offset
= start
* sizeof (mgreg_t
);
2113 map
->nslots
= end
- start
;
2114 map
->has_ref_slots
= has_ref_slots
;
2115 map
->has_pin_slots
= has_pin_slots
;
2116 map
->has_ref_regs
= has_ref_regs
;
2117 map
->has_pin_regs
= has_pin_regs
;
2118 g_assert (nregs
< 32);
2119 map
->used_int_regs
= cfg
->used_int_regs
;
2120 map
->reg_ref_mask
= reg_ref_mask
;
2121 map
->reg_pin_mask
= reg_pin_mask
;
2122 map
->nref_regs
= nref_regs
;
2123 map
->npin_regs
= npin_regs
;
2125 bitmaps
= mono_mempool_alloc0 (cfg
->mempool
, bitmaps_size
);
2128 if (has_ref_slots
) {
2129 map
->stack_ref_bitmap_offset
= bitmaps_offset
;
2130 bitmaps_offset
+= stack_bitmap_size
;
2132 bitmap
= &bitmaps
[map
->stack_ref_bitmap_offset
];
2133 for (i
= 0; i
< nslots
; ++i
) {
2134 for (j
= 0; j
< ncallsites
; ++j
) {
2135 if (get_bit (gcfg
->stack_ref_bitmap
, gcfg
->stack_bitmap_width
, j
, i
))
2136 set_bit (bitmap
, stack_bitmap_width
, j
, i
- start
);
2140 if (has_pin_slots
) {
2141 map
->stack_pin_bitmap_offset
= bitmaps_offset
;
2142 bitmaps_offset
+= stack_bitmap_size
;
2144 bitmap
= &bitmaps
[map
->stack_pin_bitmap_offset
];
2145 for (i
= 0; i
< nslots
; ++i
) {
2146 for (j
= 0; j
< ncallsites
; ++j
) {
2147 if (get_bit (gcfg
->stack_pin_bitmap
, gcfg
->stack_bitmap_width
, j
, i
))
2148 set_bit (bitmap
, stack_bitmap_width
, j
, i
- start
);
2153 map
->reg_ref_bitmap_offset
= bitmaps_offset
;
2154 bitmaps_offset
+= reg_ref_bitmap_size
;
2156 bitmap
= &bitmaps
[map
->reg_ref_bitmap_offset
];
2158 for (i
= 0; i
< nregs
; ++i
) {
2159 if (reg_ref_mask
& (1 << i
)) {
2160 for (j
= 0; j
< ncallsites
; ++j
) {
2161 if (get_bit (gcfg
->reg_ref_bitmap
, gcfg
->reg_bitmap_width
, j
, i
))
2162 set_bit (bitmap
, reg_ref_bitmap_width
, j
, bindex
);
2169 map
->reg_pin_bitmap_offset
= bitmaps_offset
;
2170 bitmaps_offset
+= reg_pin_bitmap_size
;
2172 bitmap
= &bitmaps
[map
->reg_pin_bitmap_offset
];
2174 for (i
= 0; i
< nregs
; ++i
) {
2175 if (reg_pin_mask
& (1 << i
)) {
2176 for (j
= 0; j
< ncallsites
; ++j
) {
2177 if (get_bit (gcfg
->reg_pin_bitmap
, gcfg
->reg_bitmap_width
, j
, i
))
2178 set_bit (bitmap
, reg_pin_bitmap_width
, j
, bindex
);
2186 map
->ncallsites
= ncallsites
;
2187 if (cfg
->code_len
< 256)
2188 map
->callsite_entry_size
= 1;
2189 else if (cfg
->code_len
< 65536)
2190 map
->callsite_entry_size
= 2;
2192 map
->callsite_entry_size
= 4;
2194 /* Encode the GC Map */
2202 encode_gc_map (map
, buf
, &endbuf
);
2203 g_assert (endbuf
- buf
< 256);
2205 encoded_size
= endbuf
- buf
;
2206 alloc_size
= sizeof (GCEncodedMap
) + ALIGN_TO (encoded_size
, map
->callsite_entry_size
) + (map
->callsite_entry_size
* map
->ncallsites
) + bitmaps_size
;
2208 emap
= mono_domain_alloc0 (cfg
->domain
, alloc_size
);
2209 //emap->ref_slots = map->ref_slots;
2211 /* Encoded fixed fields */
2212 p
= &emap
->encoded
[0];
2213 //emap->encoded_size = encoded_size;
2214 memcpy (p
, buf
, encoded_size
);
2217 /* Callsite table */
2218 p
= (guint8
*)ALIGN_TO ((mgreg_t
)p
, map
->callsite_entry_size
);
2219 if (map
->callsite_entry_size
== 1) {
2220 guint8
*offsets
= p
;
2221 for (i
= 0; i
< ncallsites
; ++i
)
2222 offsets
[i
] = callsites
[i
]->pc_offset
;
2223 stats
.gc_callsites8_size
+= ncallsites
* sizeof (guint8
);
2224 } else if (map
->callsite_entry_size
== 2) {
2225 guint16
*offsets
= (guint16
*)p
;
2226 for (i
= 0; i
< ncallsites
; ++i
)
2227 offsets
[i
] = callsites
[i
]->pc_offset
;
2228 stats
.gc_callsites16_size
+= ncallsites
* sizeof (guint16
);
2230 guint32
*offsets
= (guint32
*)p
;
2231 for (i
= 0; i
< ncallsites
; ++i
)
2232 offsets
[i
] = callsites
[i
]->pc_offset
;
2233 stats
.gc_callsites32_size
+= ncallsites
* sizeof (guint32
);
2235 p
+= ncallsites
* map
->callsite_entry_size
;
2238 memcpy (p
, bitmaps
, bitmaps_size
);
2241 g_assert ((guint8
*)p
- (guint8
*)emap
<= alloc_size
);
2243 stats
.gc_maps_size
+= alloc_size
;
2244 stats
.gc_callsites_size
+= ncallsites
* map
->callsite_entry_size
;
2245 stats
.gc_bitmaps_size
+= bitmaps_size
;
2246 stats
.gc_map_struct_size
+= sizeof (GCEncodedMap
) + encoded_size
;
2248 cfg
->jit_info
->gc_info
= emap
;
2250 cfg
->gc_map
= (guint8
*)emap
;
2251 cfg
->gc_map_size
= alloc_size
;
2254 stats
.all_slots
+= nslots
;
2255 stats
.ref_slots
+= ntypes
[SLOT_REF
];
2256 stats
.noref_slots
+= ntypes
[SLOT_NOREF
];
2257 stats
.pin_slots
+= ntypes
[SLOT_PIN
];
2261 mini_gc_create_gc_map (MonoCompile
*cfg
)
2263 if (!cfg
->compute_gc_maps
)
2267 * During marking, all frames except the top frame are at a call site, and we mark the
2268 * top frame conservatively. This means that we only need to compute and record
2269 * GC maps for call sites.
2272 if (!(cfg
->comp_done
& MONO_COMP_LIVENESS
))
2273 /* Without liveness info, the live ranges are not precise enough */
2276 mono_analyze_liveness_gc (cfg
);
2278 compute_frame_size (cfg
);
2282 process_spill_slots (cfg
);
2283 process_other_slots (cfg
);
2284 process_param_area_slots (cfg
);
2285 process_variables (cfg
);
2286 process_finally_clauses (cfg
);
2292 parse_debug_options (void)
2297 env
= getenv ("MONO_GCMAP_DEBUG");
2301 opts
= g_strsplit (env
, ",", -1);
2302 for (ptr
= opts
; ptr
&& *ptr
; ptr
++) {
2303 /* No options yet */
2304 fprintf (stderr
, "Invalid format for the MONO_GCMAP_DEBUG env variable: '%s'\n", env
);
2315 memset (&cb
, 0, sizeof (cb
));
2316 cb
.thread_attach_func
= thread_attach_func
;
2317 cb
.thread_detach_func
= thread_detach_func
;
2318 cb
.thread_suspend_func
= thread_suspend_func
;
2319 /* Comment this out to disable precise stack marking */
2320 cb
.thread_mark_func
= thread_mark_func
;
2321 mono_gc_set_gc_callbacks (&cb
);
2323 logfile
= mono_gc_get_logfile ();
2325 parse_debug_options ();
2327 mono_counters_register ("GC Maps size",
2328 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.gc_maps_size
);
2329 mono_counters_register ("GC Call Sites size",
2330 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.gc_callsites_size
);
2331 mono_counters_register ("GC Bitmaps size",
2332 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.gc_bitmaps_size
);
2333 mono_counters_register ("GC Map struct size",
2334 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.gc_map_struct_size
);
2335 mono_counters_register ("GC Call Sites encoded using 8 bits",
2336 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.gc_callsites8_size
);
2337 mono_counters_register ("GC Call Sites encoded using 16 bits",
2338 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.gc_callsites16_size
);
2339 mono_counters_register ("GC Call Sites encoded using 32 bits",
2340 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.gc_callsites32_size
);
2342 mono_counters_register ("GC Map slots (all)",
2343 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.all_slots
);
2344 mono_counters_register ("GC Map slots (ref)",
2345 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.ref_slots
);
2346 mono_counters_register ("GC Map slots (noref)",
2347 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.noref_slots
);
2348 mono_counters_register ("GC Map slots (pin)",
2349 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.pin_slots
);
2351 mono_counters_register ("GC TLS Data size",
2352 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.tlsdata_size
);
2354 mono_counters_register ("Stack space scanned (all)",
2355 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.scanned_stacks
);
2356 mono_counters_register ("Stack space scanned (native)",
2357 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.scanned_native
);
2358 mono_counters_register ("Stack space scanned (other)",
2359 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.scanned_other
);
2360 mono_counters_register ("Stack space scanned (using GC Maps)",
2361 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.scanned
);
2362 mono_counters_register ("Stack space scanned (precise)",
2363 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.scanned_precisely
);
2364 mono_counters_register ("Stack space scanned (pin)",
2365 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.scanned_conservatively
);
2366 mono_counters_register ("Stack space scanned (pin registers)",
2367 MONO_COUNTER_GC
| MONO_COUNTER_INT
, &stats
.scanned_registers
);
2378 mini_gc_init_gc_map (MonoCompile
*cfg
)
2383 mini_gc_create_gc_map (MonoCompile
*cfg
)
2388 mini_gc_set_slot_type_from_fp (MonoCompile
*cfg
, int slot_offset
, GCSlotType type
)
2393 mini_gc_set_slot_type_from_cfa (MonoCompile
*cfg
, int slot_offset
, GCSlotType type
)
2402 * Set GC specific options in CFG.
2405 mini_gc_init_cfg (MonoCompile
*cfg
)
2407 if (mono_gc_is_moving ()) {
2408 cfg
->disable_ref_noref_stack_slot_share
= TRUE
;
2409 cfg
->gen_write_barriers
= TRUE
;
2412 mini_gc_init_gc_map (cfg
);
2416 * Problems with the current code:
2417 * - the stack walk is slow
2418 * - vtypes/refs used in EH regions are treated conservatively
2419 * - if the code is finished, less pinning will be done, causing problems because
2420 * we promote all surviving objects to old-gen.
2421 * - the unwind code can't handle a method stopped inside a finally region, it thinks the caller is
2422 * another method, but in reality it is either the exception handling code or the CALL_HANDLER opcode.
2423 * This manifests in "Unable to find ip offset x in callsite list" assertions.
2424 * - the unwind code also can't handle frames which are in the epilog, since the unwind info is not
2429 * Ideas for creating smaller GC maps:
2430 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2432 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2434 * - if the bitmap width is not a multiple of 8, the remaining bits are wasted.
2435 * - group ref and non-ref stack slots together in mono_allocate_stack_slots ().
2436 * - add an index for the callsite table so that each entry can be encoded as a 1 byte difference
2437 * from an index entry.