1 /**
2 * \file
3 * GC interface for the mono JIT
4 *
5 * Author:
6 * Zoltan Varga (vargaz@gmail.com)
7 *
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.
11 */
12
13 #include "config.h"
14 #include "mini-gc.h"
15 #include <mono/metadata/gc-internals.h>
16
17 static gboolean
get_provenance(StackFrameInfo * frame,MonoContext * ctx,gpointer data)18 get_provenance (StackFrameInfo *frame, MonoContext *ctx, gpointer data)
19 {
20 MonoJitInfo *ji = frame->ji;
21 MonoMethod *method;
22 if (!ji)
23 return FALSE;
24 method = jinfo_get_method (ji);
25 if (method->wrapper_type != MONO_WRAPPER_NONE)
26 return FALSE;
27 *(gpointer *)data = method;
28 return TRUE;
29 }
30
31 static gpointer
get_provenance_func(void)32 get_provenance_func (void)
33 {
34 gpointer provenance = NULL;
35 mono_walk_stack (get_provenance, MONO_UNWIND_DEFAULT, (gpointer)&provenance);
36 return provenance;
37 }
38
39 #if 0
40 //#if defined(MONO_ARCH_GC_MAPS_SUPPORTED)
41
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>
46
47 #define SIZEOF_SLOT ((int)sizeof (mgreg_t))
48
49 #define GC_BITS_PER_WORD (sizeof (mword) * 8)
50
51 /* Contains state needed by the GC Map construction code */
52 typedef struct {
53 /*
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.
57 */
58 GSList *stack_slots_from_cfa;
59 /* Same for stack slots relative to the frame pointer */
60 GSList *stack_slots_from_fp;
61
62 /* Number of slots in the map */
63 int nslots;
64 /* The number of registers in the map */
65 int nregs;
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;
70
71 /* The call sites where this frame can be stopped during GC */
72 GCCallSite **callsites;
73 /* The number of call sites */
74 int ncallsites;
75
76 /*
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.
79 */
80 int stack_bitmap_width;
81 /*
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
84 * given callsite.
85 */
86 guint8 *stack_ref_bitmap;
87 /* Same for SLOT_PIN */
88 guint8 *stack_pin_bitmap;
89
90 /*
91 * Similar bitmaps for registers. These have width MONO_MAX_IREGS in bits.
92 */
93 int reg_bitmap_width;
94 guint8 *reg_ref_bitmap;
95 guint8 *reg_pin_bitmap;
96 } MonoCompileGC;
97
98 #define ALIGN_TO(val,align) ((((mgreg_t)val) + ((align) - 1)) & ~((align) - 1))
99
100 #undef DEBUG
101
102 #if 0
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
106 #else
107 #define DEBUG(s)
108 #endif
109
110 #ifdef DEBUG_ENABLED
111 //#if 1
112 #define DEBUG_PRECISE(s) do { s; } while (0)
113 #define DEBUG_PRECISE_ENABLED
114 #else
115 #define DEBUG_PRECISE(s)
116 #endif
117
118 /*
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
121 * is expensive.
122 */
123 typedef struct {
124 guint8 *bitmap;
125 int nslots;
126 int frame_start_offset;
127 int nreg_locations;
128 /* Relative to stack_start */
129 int reg_locations [MONO_MAX_IREGS];
130 #ifdef DEBUG_PRECISE_ENABLED
131 MonoJitInfo *ji;
132 gpointer fp;
133 int regs [MONO_MAX_IREGS];
134 #endif
135 } FrameInfo;
136
137 /* Max number of frames stored in the TLS data */
138 #define MAX_FRAMES 50
139
140 /*
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.
144 */
145 typedef struct {
146 MonoThreadUnwindState unwind_state;
147 MonoThreadInfo *info;
148 /* For debugging */
149 mgreg_t tid;
150 gpointer ref_to_track;
151 /* Number of frames collected during the !precise pass */
152 int nframes;
153 FrameInfo frames [MAX_FRAMES];
154 } TlsData;
155
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
159
160 /*
161 * The GC Map itself.
162 * Contains information needed to mark a stack frame.
163 * This is a transient structure, created from a compressed representation on-demand.
164 */
165 typedef struct {
166 /*
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.
169 */
170 int start_offset;
171 int end_offset;
172 /*
173 * The offset relative to frame_offset where the the memory described by the GC maps
174 * begins.
175 */
176 int map_offset;
177 /* The number of stack slots in the map */
178 int nslots;
179 /* The frame pointer register */
180 guint8 frame_reg;
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;
187
188 /* The offsets below are into an external bitmaps array */
189
190 /*
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
193 * given callsite.
194 */
195 guint32 stack_ref_bitmap_offset;
196 /*
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.
199 */
200 guint32 stack_pin_bitmap_offset;
201
202 /*
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 ?
206 */
207 guint32 reg_pin_bitmap_offset;
208 guint32 reg_ref_bitmap_offset;
209
210 guint32 used_int_regs, reg_ref_mask, reg_pin_mask;
211
212 /* The number of bits set in the two masks above */
213 guint8 nref_regs, npin_regs;
214
215 /*
216 * A bit array marking slots which contain refs.
217 * This is used only for debugging.
218 */
219 //guint8 *ref_slots;
220
221 /* Callsite offsets */
222 /* These can take up a lot of space, so encode them compactly */
223 union {
224 guint8 *offsets8;
225 guint16 *offsets16;
226 guint32 *offsets32;
227 } callsites;
228 int ncallsites;
229 } GCMap;
230
231 /*
232 * A compressed version of GCMap. This is what gets stored in MonoJitInfo.
233 */
234 typedef struct {
235 //guint8 *ref_slots;
236 //guint8 encoded_size;
237
238 /*
239 * The arrays below are embedded after the struct.
240 * Their address needs to be computed.
241 */
242
243 /* The fixed fields of the GCMap encoded using LEB128 */
244 guint8 encoded [MONO_ZERO_LEN_ARRAY];
245
246 /* An array of ncallsites entries, each entry is callsite_entry_size bytes long */
247 guint8 callsites [MONO_ZERO_LEN_ARRAY];
248
249 /* The GC bitmaps */
250 guint8 bitmaps [MONO_ZERO_LEN_ARRAY];
251 } GCEncodedMap;
252
253 static int precise_frame_count [2], precise_frame_limit = -1;
254 static gboolean precise_frame_limit_inited;
255
256 /* Stats */
257 typedef struct {
258 gint32 scanned_stacks;
259 gint32 scanned;
260 gint32 scanned_precisely;
261 gint32 scanned_conservatively;
262 gint32 scanned_registers;
263 gint32 scanned_native;
264 gint32 scanned_other;
265
266 gint32 all_slots;
267 gint32 noref_slots;
268 gint32 ref_slots;
269 gint32 pin_slots;
270
271 gint32 gc_maps_size;
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;
278 gint32 tlsdata_size;
279 } JITGCStats;
280
281 static JITGCStats stats;
282
283 static FILE *logfile;
284
285 static gboolean enable_gc_maps_for_aot;
286
287 void
288 mini_gc_enable_gc_maps_for_aot (void)
289 {
290 enable_gc_maps_for_aot = TRUE;
291 }
292
293 // FIXME: Move these to a shared place
294
295 static inline void
296 encode_uleb128 (guint32 value, guint8 *buf, guint8 **endbuf)
297 {
298 guint8 *p = buf;
299
300 do {
301 guint8 b = value & 0x7f;
302 value >>= 7;
303 if (value != 0) /* more bytes to come */
304 b |= 0x80;
305 *p ++ = b;
306 } while (value);
307
308 *endbuf = p;
309 }
310
311 static G_GNUC_UNUSED void
312 encode_sleb128 (gint32 value, guint8 *buf, guint8 **endbuf)
313 {
314 gboolean more = 1;
315 gboolean negative = (value < 0);
316 guint32 size = 32;
317 guint8 byte;
318 guint8 *p = buf;
319
320 while (more) {
321 byte = value & 0x7f;
322 value >>= 7;
323 /* the following is unnecessary if the
324 * implementation of >>= uses an arithmetic rather
325 * than logical shift for a signed left operand
326 */
327 if (negative)
328 /* sign extend */
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)))
333 more = 0;
334 else
335 byte |= 0x80;
336 *p ++= byte;
337 }
338
339 *endbuf = p;
340 }
341
342 static inline guint32
343 decode_uleb128 (guint8 *buf, guint8 **endbuf)
344 {
345 guint8 *p = buf;
346 guint32 res = 0;
347 int shift = 0;
348
349 while (TRUE) {
350 guint8 b = *p;
351 p ++;
352
353 res = res | (((int)(b & 0x7f)) << shift);
354 if (!(b & 0x80))
355 break;
356 shift += 7;
357 }
358
359 *endbuf = p;
360
361 return res;
362 }
363
364 static inline gint32
365 decode_sleb128 (guint8 *buf, guint8 **endbuf)
366 {
367 guint8 *p = buf;
368 gint32 res = 0;
369 int shift = 0;
370
371 while (TRUE) {
372 guint8 b = *p;
373 p ++;
374
375 res = res | (((int)(b & 0x7f)) << shift);
376 shift += 7;
377 if (!(b & 0x80)) {
378 if (shift < 32 && (b & 0x40))
379 res |= - (1 << shift);
380 break;
381 }
382 }
383
384 *endbuf = p;
385
386 return res;
387 }
388
389 static int
390 encode_frame_reg (int frame_reg)
391 {
392 #ifdef TARGET_AMD64
393 if (frame_reg == AMD64_RSP)
394 return 0;
395 else if (frame_reg == AMD64_RBP)
396 return 1;
397 #elif defined(TARGET_X86)
398 if (frame_reg == X86_EBP)
399 return 0;
400 else if (frame_reg == X86_ESP)
401 return 1;
402 #elif defined(TARGET_ARM)
403 if (frame_reg == ARMREG_SP)
404 return 0;
405 else if (frame_reg == ARMREG_FP)
406 return 1;
407 #elif defined(TARGET_S390X)
408 if (frame_reg == S390_SP)
409 return 0;
410 else if (frame_reg == S390_FP)
411 return 1;
412 #else
413 NOT_IMPLEMENTED;
414 #endif
415 g_assert_not_reached ();
416 return -1;
417 }
418
419 static int
420 decode_frame_reg (int encoded)
421 {
422 #ifdef TARGET_AMD64
423 if (encoded == 0)
424 return AMD64_RSP;
425 else if (encoded == 1)
426 return AMD64_RBP;
427 #elif defined(TARGET_X86)
428 if (encoded == 0)
429 return X86_EBP;
430 else if (encoded == 1)
431 return X86_ESP;
432 #elif defined(TARGET_ARM)
433 if (encoded == 0)
434 return ARMREG_SP;
435 else if (encoded == 1)
436 return ARMREG_FP;
437 #elif defined(TARGET_S390X)
438 if (encoded == 0)
439 return S390_SP;
440 else if (encoded == 1)
441 return S390_FP;
442 #else
443 NOT_IMPLEMENTED;
444 #endif
445 g_assert_not_reached ();
446 return -1;
447 }
448
449 #ifdef TARGET_AMD64
450 #ifdef HOST_WIN32
451 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15, AMD64_RDI, AMD64_RSI };
452 #else
453 static int callee_saved_regs [] = { AMD64_RBP, AMD64_RBX, AMD64_R12, AMD64_R13, AMD64_R14, AMD64_R15 };
454 #endif
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)
460 // FIXME:
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 };
473 #endif
474
475 static guint32
476 encode_regmask (guint32 regmask)
477 {
478 int i;
479 guint32 res;
480
481 res = 0;
482 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i) {
483 if (regmask & (1 << callee_saved_regs [i])) {
484 res |= (1 << i);
485 regmask -= (1 << callee_saved_regs [i]);
486 }
487 }
488 g_assert (regmask == 0);
489 return res;
490 }
491
492 static guint32
493 decode_regmask (guint32 regmask)
494 {
495 int i;
496 guint32 res;
497
498 res = 0;
499 for (i = 0; i < sizeof (callee_saved_regs) / sizeof (int); ++i)
500 if (regmask & (1 << i))
501 res |= (1 << callee_saved_regs [i]);
502 return res;
503 }
504
505 /*
506 * encode_gc_map:
507 *
508 * Encode the fixed fields of MAP into a buffer pointed to by BUF.
509 */
510 static void
511 encode_gc_map (GCMap *map, guint8 *buf, guint8 **endbuf)
512 {
513 guint32 flags, freg;
514
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);
521 g_assert (freg < 2);
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);
530
531 *endbuf = buf;
532 }
533
534 /*
535 * decode_gc_map:
536 *
537 * Decode the encoded GC map representation in BUF and store the result into MAP.
538 */
539 static void
540 decode_gc_map (guint8 *buf, GCMap *map, guint8 **endbuf)
541 {
542 guint32 flags;
543 int stack_bitmap_size, reg_ref_bitmap_size, reg_pin_bitmap_size, offset, freg;
544 int i, n;
545
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;
556 freg = flags >> 6;
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));
561 n = 0;
562 for (i = 0; i < NREGS; ++i)
563 if (map->reg_ref_mask & (1 << i))
564 n ++;
565 map->nref_regs = n;
566 }
567 if (map->has_pin_regs) {
568 map->reg_pin_mask = decode_regmask (decode_uleb128 (buf, &buf));
569 n = 0;
570 for (i = 0; i < NREGS; ++i)
571 if (map->reg_pin_mask & (1 << i))
572 n ++;
573 map->npin_regs = n;
574 }
575 map->ncallsites = decode_uleb128 (buf, &buf);
576
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;
580 offset = 0;
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;
593
594 *endbuf = buf;
595 }
596
597 static gpointer
598 thread_attach_func (void)
599 {
600 TlsData *tls;
601
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));
606
607 return tls;
608 }
609
610 static void
611 thread_detach_func (gpointer user_data)
612 {
613 TlsData *tls = user_data;
614
615 g_free (tls);
616 }
617
618 static void
619 thread_suspend_func (gpointer user_data, void *sigctx, MonoContext *ctx)
620 {
621 TlsData *tls = user_data;
622
623 if (!tls) {
624 /* Happens during startup */
625 return;
626 }
627
628 if (tls->tid != mono_native_thread_id_get ()) {
629 /* Happens on osx because threads are not suspended using signals */
630 #ifndef TARGET_WIN32
631 gboolean res;
632 #endif
633
634 g_assert (tls->info);
635 #ifdef TARGET_WIN32
636 return;
637 #else
638 res = mono_thread_state_init_from_handle (&tls->unwind_state, tls->info, NULL);
639 #endif
640 } else {
641 tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF] = mono_get_lmf ();
642 if (sigctx) {
643 #ifdef MONO_ARCH_HAVE_SIGCTX_TO_MONOCTX
644 mono_sigctx_to_monoctx (sigctx, &tls->unwind_state.ctx);
645 tls->unwind_state.valid = TRUE;
646 #else
647 tls->unwind_state.valid = FALSE;
648 #endif
649 } else if (ctx) {
650 memcpy (&tls->unwind_state.ctx, ctx, sizeof (MonoContext));
651 tls->unwind_state.valid = TRUE;
652 } else {
653 tls->unwind_state.valid = FALSE;
654 }
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 ();
657 }
658
659 if (!tls->unwind_state.unwind_data [MONO_UNWIND_DATA_DOMAIN]) {
660 /* Happens during startup */
661 tls->unwind_state.valid = FALSE;
662 return;
663 }
664 }
665
666 #define DEAD_REF ((gpointer)(gssize)0x2a2a2a2a2a2a2a2aULL)
667
668 static inline void
669 set_bit (guint8 *bitmap, int width, int y, int x)
670 {
671 bitmap [(width * y) + (x / 8)] |= (1 << (x % 8));
672 }
673
674 static inline void
675 clear_bit (guint8 *bitmap, int width, int y, int x)
676 {
677 bitmap [(width * y) + (x / 8)] &= ~(1 << (x % 8));
678 }
679
680 static inline int
681 get_bit (guint8 *bitmap, int width, int y, int x)
682 {
683 return bitmap [(width * y) + (x / 8)] & (1 << (x % 8));
684 }
685
686 static const char*
687 slot_type_to_string (GCSlotType type)
688 {
689 switch (type) {
690 case SLOT_REF:
691 return "ref";
692 case SLOT_NOREF:
693 return "noref";
694 case SLOT_PIN:
695 return "pin";
696 default:
697 g_assert_not_reached ();
698 return NULL;
699 }
700 }
701
702 static inline mgreg_t
703 get_frame_pointer (MonoContext *ctx, int frame_reg)
704 {
705 #if defined(TARGET_AMD64)
706 if (frame_reg == AMD64_RSP)
707 return ctx->rsp;
708 else if (frame_reg == AMD64_RBP)
709 return ctx->rbp;
710 #elif defined(TARGET_X86)
711 if (frame_reg == X86_ESP)
712 return ctx->esp;
713 else if (frame_reg == X86_EBP)
714 return ctx->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);
725 #endif
726 g_assert_not_reached ();
727 return 0;
728 }
729
730 /*
731 * conservatively_pass:
732 *
733 * Mark a thread stack conservatively and collect information needed by the precise pass.
734 */
735 static void
736 conservative_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end)
737 {
738 MonoJitInfo *ji;
739 MonoMethod *method;
740 MonoContext ctx, new_ctx;
741 MonoLMF *lmf;
742 guint8 *stack_limit;
743 gboolean last = TRUE;
744 GCMap *map;
745 GCMap map_tmp;
746 GCEncodedMap *emap;
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;
750 gboolean res;
751 StackFrameInfo frame;
752 mgreg_t *reg_locations [MONO_MAX_IREGS];
753 mgreg_t *new_reg_locations [MONO_MAX_IREGS];
754 guint8 *bitmaps;
755 FrameInfo *fi;
756 guint32 precise_regmask;
757
758 if (tls) {
759 tls->nframes = 0;
760 tls->ref_to_track = NULL;
761 }
762
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);
767 return;
768 }
769
770 lmf = tls->unwind_state.unwind_data [MONO_UNWIND_DATA_LMF];
771 frame.domain = NULL;
772
773 /* Number of bytes scanned based on GC map data */
774 scanned = 0;
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;
781
782 /* This is one past the last address which we have scanned */
783 stack_limit = stack_start;
784
785 if (!tls->unwind_state.valid)
786 memset (&new_ctx, 0, sizeof (ctx));
787 else
788 memcpy (&new_ctx, &tls->unwind_state.ctx, sizeof (MonoContext));
789
790 memset (reg_locations, 0, sizeof (reg_locations));
791 memset (new_reg_locations, 0, sizeof (new_reg_locations));
792
793 while (TRUE) {
794 if (!tls->unwind_state.valid)
795 break;
796
797 memcpy (&ctx, &new_ctx, sizeof (ctx));
798
799 for (i = 0; i < MONO_MAX_IREGS; ++i) {
800 if (new_reg_locations [i]) {
801 /*
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.
805 */
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;
810 }
811
812 reg_locations [i] = new_reg_locations [i];
813
814 DEBUG (fprintf (logfile, "\treg %s is now at location %p.\n", mono_arch_regname (i), reg_locations [i]));
815 }
816 }
817
818 g_assert ((mgreg_t)stack_limit % SIZEOF_SLOT == 0);
819
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);
821 if (!res)
822 break;
823
824 ji = frame.ji;
825
826 // FIXME: For skipped frames, scan the param area of the parent frame conservatively ?
827 // FIXME: trampolines
828
829 if (frame.type == FRAME_TYPE_MANAGED_TO_NATIVE) {
830 /*
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.
835 */
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;
842 }
843 reg_locations [i] = NULL;
844 new_reg_locations [i] = NULL;
845 }
846 ctx = new_ctx;
847 continue;
848 }
849
850 if (ji)
851 method = jinfo_get_method (ji);
852 else
853 method = NULL;
854
855 /* The last frame can be in any state so mark conservatively */
856 if (last) {
857 if (ji) {
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));
859 }
860 DEBUG (fprintf (logfile, "\t <Last frame>\n"));
861 last = FALSE;
862 /*
863 * new_reg_locations is not precise when a method is interrupted during its epilog, so clear it.
864 */
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;
870 }
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;
875 }
876 reg_locations [i] = NULL;
877 new_reg_locations [i] = NULL;
878 }
879 continue;
880 }
881
882 pc_offset = (guint8*)MONO_CONTEXT_GET_IP (&ctx) - (guint8*)ji->code_start;
883
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"));
888 continue;
889 }
890
891 /* All the other frames are at a call site */
892
893 if (tls->nframes == MAX_FRAMES) {
894 /*
895 * Can't save information since the array is full. So scan the rest of the
896 * stack conservatively.
897 */
898 DEBUG (fprintf (logfile, "Mark (0): Frame stack full.\n"));
899 break;
900 }
901
902 /* Scan the frame of this method */
903
904 /*
905 * A frame contains the following:
906 * - saved registers
907 * - saved args
908 * - locals
909 * - spill area
910 * - localloc-ed memory
911 */
912 g_assert (pc_offset >= 0);
913
914 emap = ji->gc_info;
915
916 if (!emap) {
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"));
919 continue;
920 }
921
922 /* The embedded callsite table requires this */
923 g_assert (((mgreg_t)emap % 4) == 0);
924
925 /*
926 * Debugging aid to control the number of frames scanned precisely
927 */
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);
933 }
934 precise_frame_limit_inited = TRUE;
935 }
936
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)
941 continue;
942 }
943 precise_frame_count [FALSE] ++;
944
945 /* Decode the encoded GC map */
946 map = &map_tmp;
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;
952 bitmaps = p;
953
954 fp = (guint8*)get_frame_pointer (&ctx, map->frame_reg);
955
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;
959
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));
961
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)
967 break;
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)
973 break;
974 } else {
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)
979 break;
980 }
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 ();
984 }
985 cindex = i;
986
987 /*
988 * This is not neccessary true on x86 because frames have a different size at each
989 * call site.
990 */
991 //g_assert (real_frame_start >= stack_limit);
992
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);
998 }
999
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)];
1004 guint8 *p;
1005 gboolean pinned;
1006
1007 p = frame_start;
1008 for (i = 0; i < map->nslots; ++i) {
1009 pinned = pin_bitmap [i / 8] & (1 << (i % 8));
1010 if (pinned) {
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;
1014 } else {
1015 scanned_precisely += SIZEOF_SLOT;
1016 }
1017 p += SIZEOF_SLOT;
1018 }
1019 } else {
1020 scanned_precisely += (map->nslots * SIZEOF_SLOT);
1021 }
1022
1023 /* The area outside of start-end is NOREF */
1024 scanned_precisely += (map->end_offset - map->start_offset) - (map->nslots * SIZEOF_SLOT);
1025
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)];
1031 int bindex = 0;
1032 for (i = 0; i < NREGS; ++i) {
1033 if (!(map->used_int_regs & (1 << i)))
1034 continue;
1035
1036 if (!(map->reg_pin_mask & (1 << i)))
1037 continue;
1038
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);
1042 }
1043 bindex ++;
1044 }
1045 }
1046
1047 scanned += map->end_offset - map->start_offset;
1048
1049 g_assert (scanned == scanned_precisely + scanned_conservatively);
1050
1051 stack_limit = frame_end;
1052
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)];
1059 else
1060 fi->bitmap = NULL;
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);
1065
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)];
1069 int bindex = 0;
1070 for (i = 0; i < NREGS; ++i) {
1071 if (!(map->reg_ref_mask & (1 << i)))
1072 continue;
1073
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 ++;
1079 }
1080 bindex ++;
1081 }
1082 }
1083
1084 /*
1085 * Clear locations of precisely tracked registers.
1086 */
1087 if (precise_regmask) {
1088 for (i = 0; i < NREGS; ++i) {
1089 if (precise_regmask & (1 << i)) {
1090 /*
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
1094 * processed.
1095 */
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;
1099 }
1100 }
1101 }
1102
1103 tls->nframes ++;
1104 }
1105
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;
1112 }
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;
1117 }
1118 }
1119
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);
1124 }
1125
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)));
1127
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);
1133
1134 //mono_gc_conservatively_scan_area (stack_start, stack_end);
1135 }
1136
1137 /*
1138 * precise_pass:
1139 *
1140 * Mark a thread stack precisely based on information saved during the conservative
1141 * pass.
1142 */
1143 static void
1144 precise_pass (TlsData *tls, guint8 *stack_start, guint8 *stack_end, void *gc_data)
1145 {
1146 int findex, i;
1147 FrameInfo *fi;
1148 guint8 *frame_start;
1149
1150 if (!tls)
1151 return;
1152
1153 if (!tls->unwind_state.valid)
1154 return;
1155
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;
1160
1161 DEBUG (char *fname = mono_method_full_name (jinfo_get_method (fi->ji), TRUE); fprintf (logfile, "Mark(1): %s\n", fname); g_free (fname));
1162
1163 /*
1164 * FIXME: Add a function to mark using a bitmap, to avoid doing a
1165 * call for each object.
1166 */
1167
1168 /* Mark stack slots */
1169 if (fi->bitmap) {
1170 guint8 *ref_bitmap = fi->bitmap;
1171 gboolean live;
1172
1173 for (i = 0; i < fi->nslots; ++i) {
1174 MonoObject **ptr = (MonoObject**)(frame_start + (i * SIZEOF_SLOT));
1175
1176 live = ref_bitmap [i / 8] & (1 << (i % 8));
1177
1178 if (live) {
1179 MonoObject *obj = *ptr;
1180 if (obj) {
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));
1184 } else {
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));
1186 }
1187 } else {
1188 #if 0
1189 /*
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 ?
1192 */
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));
1195 /*
1196 * Fail fast if the live range is incorrect, and
1197 * the JITted code tries to access this object
1198 */
1199 *ptr = DEAD_REF;
1200 }
1201 #endif
1202 }
1203 }
1204 }
1205
1206 /* Mark registers */
1207
1208 /*
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.
1212 */
1213 for (i = 0; i < fi->nreg_locations; ++i) {
1214 /*
1215 * reg_locations [i] contains the address of the stack slot where
1216 * a reg was last saved, so mark that slot.
1217 */
1218 MonoObject **ptr = (MonoObject**)((guint8*)stack_start + fi->reg_locations [i]);
1219 MonoObject *obj = *ptr;
1220
1221 if (obj) {
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));
1225 } else {
1226 DEBUG (fprintf (logfile, "\treg %s saved at %p: %p\n", mono_arch_regname (fi->regs [i]), ptr, obj));
1227 }
1228 }
1229 }
1230
1231 /*
1232 * Debugging aid to check for missed refs.
1233 */
1234 if (tls->ref_to_track) {
1235 mgreg_t *p;
1236
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);
1240 }
1241 }
1242
1243 /*
1244 * thread_mark_func:
1245 *
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.
1249 */
1250 static void
1251 thread_mark_func (gpointer user_data, guint8 *stack_start, guint8 *stack_end, gboolean precise, void *gc_data)
1252 {
1253 TlsData *tls = user_data;
1254
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"));
1258
1259 if (!precise)
1260 conservative_pass (tls, stack_start, stack_end);
1261 else
1262 precise_pass (tls, stack_start, stack_end, gc_data);
1263 }
1264
1265 #ifndef DISABLE_JIT
1266
1267 static void
1268 mini_gc_init_gc_map (MonoCompile *cfg)
1269 {
1270 if (COMPILE_LLVM (cfg))
1271 return;
1272
1273 if (!mono_gc_is_moving ())
1274 return;
1275
1276 if (cfg->compile_aot) {
1277 if (!enable_gc_maps_for_aot)
1278 return;
1279 } else if (!mono_gc_precise_stack_mark_enabled ())
1280 return;
1281
1282 #if 1
1283 /* Debugging support */
1284 {
1285 static int precise_count;
1286
1287 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)
1295 return;
1296 }
1297 }
1298 #endif
1299
1300 cfg->compute_gc_maps = TRUE;
1301
1302 cfg->gc_info = mono_mempool_alloc0 (cfg->mempool, sizeof (MonoCompileGC));
1303 }
1304
1305 /*
1306 * mini_gc_set_slot_type_from_fp:
1307 *
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.
1311 */
1312 void
1313 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
1314 {
1315 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1316
1317 if (!cfg->compute_gc_maps)
1318 return;
1319
1320 g_assert (slot_offset % SIZEOF_SLOT == 0);
1321
1322 gcfg->stack_slots_from_fp = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_fp, GINT_TO_POINTER (((slot_offset) << 16) | type));
1323 }
1324
1325 /*
1326 * mini_gc_set_slot_type_from_cfa:
1327 *
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.
1333 */
1334 void
1335 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
1336 {
1337 MonoCompileGC *gcfg = (MonoCompileGC*)cfg->gc_info;
1338 int slot = - (slot_offset / SIZEOF_SLOT);
1339
1340 if (!cfg->compute_gc_maps)
1341 return;
1342
1343 g_assert (slot_offset <= 0);
1344 g_assert (slot_offset % SIZEOF_SLOT == 0);
1345
1346 gcfg->stack_slots_from_cfa = g_slist_prepend_mempool (cfg->mempool, gcfg->stack_slots_from_cfa, GUINT_TO_POINTER (((slot) << 16) | type));
1347 }
1348
1349 static inline int
1350 fp_offset_to_slot (MonoCompile *cfg, int offset)
1351 {
1352 MonoCompileGC *gcfg = cfg->gc_info;
1353
1354 return (offset - gcfg->min_offset) / SIZEOF_SLOT;
1355 }
1356
1357 static inline int
1358 slot_to_fp_offset (MonoCompile *cfg, int slot)
1359 {
1360 MonoCompileGC *gcfg = cfg->gc_info;
1361
1362 return (slot * SIZEOF_SLOT) + gcfg->min_offset;
1363 }
1364
1365 static inline MONO_ALWAYS_INLINE void
1366 set_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1367 {
1368 g_assert (slot >= 0 && slot < gcfg->nslots);
1369
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);
1379 }
1380 }
1381
1382 static inline void
1383 set_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1384 {
1385 int width, pos;
1386 guint8 *ref_bitmap, *pin_bitmap;
1387
1388 /*
1389 int cindex;
1390
1391 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1392 set_slot (gcfg, slot, cindex, type);
1393 */
1394 ref_bitmap = gcfg->stack_ref_bitmap;
1395 pin_bitmap = gcfg->stack_pin_bitmap;
1396 width = gcfg->stack_bitmap_width;
1397 pos = width * slot;
1398
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);
1408 }
1409 }
1410
1411 static inline void
1412 set_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1413 {
1414 int cindex;
1415
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);
1420 }
1421 }
1422
1423 static inline void
1424 set_reg_slot (MonoCompileGC *gcfg, int slot, int callsite_index, GCSlotType type)
1425 {
1426 g_assert (slot >= 0 && slot < gcfg->nregs);
1427
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);
1437 }
1438 }
1439
1440 static inline void
1441 set_reg_slot_everywhere (MonoCompileGC *gcfg, int slot, GCSlotType type)
1442 {
1443 int cindex;
1444
1445 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex)
1446 set_reg_slot (gcfg, slot, cindex, type);
1447 }
1448
1449 static inline void
1450 set_reg_slot_in_range (MonoCompileGC *gcfg, int slot, int from, int to, GCSlotType type)
1451 {
1452 int cindex;
1453
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);
1458 }
1459 }
1460
1461 static void
1462 process_spill_slots (MonoCompile *cfg)
1463 {
1464 MonoCompileGC *gcfg = cfg->gc_info;
1465 MonoBasicBlock *bb;
1466 GSList *l;
1467 int i;
1468
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);
1477
1478 if (bank == MONO_REG_INT_MP || bank == MONO_REG_INT_REF)
1479 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1480 }
1481 }
1482
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);
1490 GCSlotType type;
1491
1492 if (bank == MONO_REG_INT_MP)
1493 type = SLOT_PIN;
1494 else
1495 type = SLOT_REF;
1496
1497 /*
1498 * Extend the live interval for the GC tracked spill slots
1499 * defined in this bblock.
1500 * FIXME: This is not needed.
1501 */
1502 set_slot_in_range (gcfg, slot, def->backend.pc_offset, bb->native_offset + bb->native_length, type);
1503
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);
1506 }
1507 }
1508
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;
1512 int slot;
1513
1514 if (offset == -1)
1515 continue;
1516
1517 slot = fp_offset_to_slot (cfg, offset);
1518
1519 set_slot_everywhere (gcfg, slot, SLOT_NOREF);
1520 /* FIXME: 32 bit */
1521 if (cfg->verbose_level > 1)
1522 printf ("\tfp spill slot at %s0x%x(fp) (slot = %d)\n", offset >= 0 ? "" : "-", ABS (offset), slot);
1523 }
1524
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;
1528 int slot;
1529
1530 if (offset == -1)
1531 continue;
1532
1533 slot = fp_offset_to_slot (cfg, offset);
1534
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);
1538 }
1539 }
1540
1541 /*
1542 * process_other_slots:
1543 *
1544 * Process stack slots registered using mini_gc_set_slot_type_... ().
1545 */
1546 static void
1547 process_other_slots (MonoCompile *cfg)
1548 {
1549 MonoCompileGC *gcfg = cfg->gc_info;
1550 GSList *l;
1551
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;
1557 int slot;
1558
1559 /*
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)
1564 */
1565 slot = (cfg->cfa_offset / SIZEOF_SLOT) - cfa_slot - (gcfg->min_offset / SIZEOF_SLOT);
1566
1567 set_slot_everywhere (gcfg, slot, type);
1568
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));
1573 }
1574 }
1575
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;
1581 int slot;
1582
1583 slot = fp_offset_to_slot (cfg, offset);
1584
1585 set_slot_everywhere (gcfg, slot, type);
1586
1587 /* Liveness for these slots is handled by process_spill_slots () */
1588
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);
1594 }
1595 }
1596 }
1597
1598 static gsize*
1599 get_vtype_bitmap (MonoType *t, int *numbits)
1600 {
1601 MonoClass *klass = mono_class_from_mono_type (t);
1602
1603 if (klass->generic_container || mono_class_is_open_constructed_type (t)) {
1604 /* FIXME: Generic sharing */
1605 return NULL;
1606 } else {
1607 mono_class_compute_gc_descriptor (klass);
1608
1609 return mono_gc_get_bitmap_for_descr (klass->gc_descr, numbits);
1610 }
1611 }
1612
1613 static inline const char*
1614 get_offset_sign (int offset)
1615 {
1616 return offset < 0 ? "-" : "+";
1617 }
1618
1619 static inline int
1620 get_offset_val (int offset)
1621 {
1622 return offset < 0 ? (- offset) : offset;
1623 }
1624
1625 static void
1626 process_variables (MonoCompile *cfg)
1627 {
1628 MonoCompileGC *gcfg = cfg->gc_info;
1629 MonoMethodSignature *sig = mono_method_signature (cfg->method);
1630 int i, locals_min_slot, locals_max_slot, cindex;
1631 MonoBasicBlock *bb;
1632 MonoInst *tmp;
1633 int *pc_offsets;
1634 int locals_min_offset = gcfg->locals_min_offset;
1635 int locals_max_offset = gcfg->locals_max_offset;
1636
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);
1642 }
1643
1644 /*
1645 * Compute the offset where variables are initialized in the first bblock, if any.
1646 */
1647 pc_offsets = g_new0 (int, cfg->next_vreg);
1648
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;
1656 }
1657 }
1658 }
1659
1660 /*
1661 * Stack slots holding arguments are initialized in the prolog.
1662 * This means we can treat them alive for the whole method.
1663 */
1664 for (i = 0; i < cfg->num_varinfo; i++) {
1665 MonoInst *ins = cfg->varinfo [i];
1666 MonoType *t = ins->inst_vtype;
1667 MonoMethodVar *vmv;
1668 guint32 pos;
1669 gboolean byref, is_this = FALSE;
1670 gboolean is_arg = i < cfg->locals_start;
1671
1672 if (ins == cfg->ret) {
1673 if (!(ins->opcode == OP_REGOFFSET && MONO_TYPE_ISSTRUCT (t)))
1674 continue;
1675 }
1676
1677 vmv = MONO_VARINFO (cfg, i);
1678
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;
1682 is_this = TRUE;
1683 }
1684
1685 byref = t->byref;
1686
1687 if (ins->opcode == OP_REGVAR) {
1688 int hreg;
1689 GCSlotType slot_type;
1690
1691 t = mini_get_underlying_type (t);
1692
1693 hreg = ins->dreg;
1694 g_assert (hreg < MONO_MAX_IREGS);
1695
1696 if (byref)
1697 slot_type = SLOT_PIN;
1698 else
1699 slot_type = mini_type_is_reference (t) ? SLOT_REF : SLOT_NOREF;
1700
1701 if (slot_type == SLOT_PIN) {
1702 /* These have no live interval, be conservative */
1703 set_reg_slot_everywhere (gcfg, hreg, slot_type);
1704 } else {
1705 /*
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.
1710 */
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);
1714 }
1715
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);
1718 }
1719
1720 continue;
1721 }
1722
1723 if (ins->opcode != OP_REGOFFSET)
1724 continue;
1725
1726 if (ins->inst_offset % SIZEOF_SLOT != 0)
1727 continue;
1728
1729 pos = fp_offset_to_slot (cfg, ins->inst_offset);
1730
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);
1734
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));
1737 }
1738 continue;
1739 }
1740
1741 if (MONO_TYPE_ISSTRUCT (t)) {
1742 int numbits = 0, j;
1743 gsize *bitmap = NULL;
1744 gboolean pin = FALSE;
1745 int size;
1746 int size_in_slots;
1747
1748 if (ins->backend.is_pinvoke)
1749 size = mono_class_native_size (ins->klass, NULL);
1750 else
1751 size = mono_class_value_size (ins->klass, NULL);
1752 size_in_slots = ALIGN_TO (size, SIZEOF_SLOT) / SIZEOF_SLOT;
1753
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));
1756
1757 if (!ins->klass->has_references) {
1758 if (is_arg) {
1759 for (j = 0; j < size_in_slots; ++j)
1760 set_slot_everywhere (gcfg, pos + j, SLOT_NOREF);
1761 }
1762 continue;
1763 }
1764
1765 bitmap = get_vtype_bitmap (t, &numbits);
1766 if (!bitmap)
1767 pin = TRUE;
1768
1769 /*
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.
1774 */
1775 if (!is_arg) {
1776 if (!pc_offsets [vmv->vreg])
1777 pin = TRUE;
1778
1779 if (ins->backend.is_pinvoke)
1780 pin = TRUE;
1781 }
1782
1783 if (bitmap) {
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);
1790 }
1791 }
1792 }
1793 }
1794
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)));
1799 }
1800 }
1801 } else {
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);
1806 }
1807 }
1808
1809 g_free (bitmap);
1810
1811 continue;
1812 }
1813
1814 if (!is_arg && (ins->inst_offset < gcfg->min_offset || ins->inst_offset >= gcfg->max_offset))
1815 /* Vret addr etc. */
1816 continue;
1817
1818 if (t->byref) {
1819 if (is_arg) {
1820 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1821 } else {
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);
1825 }
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));
1828 continue;
1829 }
1830
1831 /*
1832 * This is currently disabled, but could be enabled to debug crashes.
1833 */
1834 #if 0
1835 if (t->type == MONO_TYPE_I) {
1836 /*
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.
1840 */
1841 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1842 continue;
1843 }
1844 #endif
1845
1846 t = mini_get_underlying_type (t);
1847
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));
1856 }
1857 continue;
1858 }
1859
1860 /* 'this' is marked INDIRECT for gshared methods */
1861 if (ins->flags & (MONO_INST_VOLATILE | MONO_INST_INDIRECT) && !is_this) {
1862 /*
1863 * For volatile variables, treat them alive from the point they are
1864 * initialized in the first bblock until the end of the method.
1865 */
1866 if (is_arg) {
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);
1871 } else {
1872 set_slot_everywhere (gcfg, pos, SLOT_PIN);
1873 }
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));
1876 continue;
1877 }
1878
1879 if (is_arg) {
1880 /* Live for the whole method */
1881 set_slot_everywhere (gcfg, pos, SLOT_REF);
1882 } else {
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);
1886 }
1887
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));
1890 }
1891 }
1892
1893 g_free (pc_offsets);
1894 }
1895
1896 static int
1897 sp_offset_to_fp_offset (MonoCompile *cfg, int sp_offset)
1898 {
1899 /*
1900 * Convert a sp relative offset to a slot index. This is
1901 * platform specific.
1902 */
1903 #ifdef TARGET_AMD64
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);
1912 #else
1913 return (- cfg->arch.sp_fp_offset - sp_offset);
1914 #endif
1915 #else
1916 NOT_IMPLEMENTED;
1917 return -1;
1918 #endif
1919 }
1920
1921 static void
1922 process_param_area_slots (MonoCompile *cfg)
1923 {
1924 MonoCompileGC *gcfg = cfg->gc_info;
1925 int cindex, i;
1926 gboolean *is_param;
1927
1928 /*
1929 * These slots are used for passing parameters during calls. They are sp relative, not
1930 * fp relative, so they are harder to handle.
1931 */
1932 if (cfg->flags & MONO_CFG_HAS_ALLOCA)
1933 /* The distance between fp and sp is not constant */
1934 return;
1935
1936 is_param = mono_mempool_alloc0 (cfg->mempool, gcfg->nslots * sizeof (gboolean));
1937
1938 for (cindex = 0; cindex < gcfg->ncallsites; ++cindex) {
1939 GCCallSite *callsite = gcfg->callsites [cindex];
1940 GSList *l;
1941
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);
1948 guint32 align;
1949 guint32 size;
1950
1951 if (MONO_TYPE_ISSTRUCT (t)) {
1952 size = mini_type_stack_size_full (t, &align, FALSE);
1953 } else {
1954 size = sizeof (mgreg_t);
1955 }
1956
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;
1960 }
1961 }
1962 }
1963
1964 /* All param area slots are noref by default */
1965 for (i = 0; i < gcfg->nslots; ++i) {
1966 if (is_param [i])
1967 set_slot_everywhere (gcfg, i, SLOT_NOREF);
1968 }
1969
1970 /*
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.
1973 */
1974 }
1975
1976 static void
1977 process_finally_clauses (MonoCompile *cfg)
1978 {
1979 MonoCompileGC *gcfg = cfg->gc_info;
1980 GCCallSite **callsites;
1981 int ncallsites;
1982 gboolean has_finally;
1983 int i, j, nslots, nregs;
1984
1985 ncallsites = gcfg->ncallsites;
1986 nslots = gcfg->nslots;
1987 nregs = gcfg->nregs;
1988 callsites = gcfg->callsites;
1989
1990 /*
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.
1996 */
1997 has_finally = FALSE;
1998 for (i = 0; i < cfg->header->num_clauses; ++i) {
1999 MonoExceptionClause *clause = &cfg->header->clauses [i];
2000
2001 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
2002 has_finally = TRUE;
2003 }
2004 }
2005 if (has_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;
2012
2013 for (i = 0; i < cfg->header->num_clauses; ++i) {
2014 clause = &cfg->header->clauses [i];
2015
2016 if (MONO_OFFSET_IN_HANDLER (clause, bb->real_offset)) {
2017 if (clause->flags == MONO_EXCEPTION_CLAUSE_FINALLY) {
2018 is_in_finally = TRUE;
2019 break;
2020 }
2021 }
2022 }
2023
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);
2029 }
2030 }
2031 }
2032 }
2033
2034 static void
2035 compute_frame_size (MonoCompile *cfg)
2036 {
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);
2041 GSList *l;
2042
2043 /* Compute min/max offsets from the fp */
2044
2045 /* Locals */
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;
2049 #else
2050 /* min/max stack offset needs to be computed in mono_arch_allocate_vars () */
2051 NOT_IMPLEMENTED;
2052 #endif
2053
2054 locals_min_offset = ALIGN_TO (locals_min_offset, SIZEOF_SLOT);
2055 locals_max_offset = ALIGN_TO (locals_max_offset, SIZEOF_SLOT);
2056
2057 min_offset = locals_min_offset;
2058 max_offset = locals_max_offset;
2059
2060 /* Arguments */
2061 for (i = 0; i < sig->param_count + sig->hasthis; ++i) {
2062 MonoInst *ins = cfg->args [i];
2063
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;
2068
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)));
2071 }
2072 }
2073
2074 /* Cfa slots */
2075 g_assert (cfg->frame_reg == cfg->cfa_reg);
2076 g_assert (cfg->cfa_offset > 0);
2077 cfa_min_offset = 0;
2078 cfa_max_offset = cfg->cfa_offset;
2079
2080 min_offset = MIN (min_offset, cfa_min_offset);
2081 max_offset = MAX (max_offset, cfa_max_offset);
2082
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;
2087
2088 min_offset = MIN (min_offset, offset);
2089 }
2090
2091 /* Spill slots */
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));
2095 }
2096
2097 /* Param area slots */
2098 #ifdef TARGET_AMD64
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);
2103 #else
2104 min_offset = MIN (min_offset, - (cfg->arch.sp_fp_offset + cfg->arch.param_area_size));
2105 #endif
2106 #elif defined(TARGET_ARM)
2107 // FIXME:
2108 #elif defined(TARGET_s390X)
2109 // FIXME:
2110 #else
2111 NOT_IMPLEMENTED;
2112 #endif
2113
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;
2118 }
2119
2120 static void
2121 init_gcfg (MonoCompile *cfg)
2122 {
2123 int i, nregs, nslots;
2124 MonoCompileGC *gcfg = cfg->gc_info;
2125 GCCallSite **callsites;
2126 int ncallsites;
2127 MonoBasicBlock *bb;
2128 GSList *l;
2129
2130 /*
2131 * Collect callsites
2132 */
2133 ncallsites = 0;
2134 for (bb = cfg->bb_entry; bb; bb = bb->next_bb) {
2135 ncallsites += g_slist_length (bb->gc_callsites);
2136 }
2137 callsites = mono_mempool_alloc0 (cfg->mempool, ncallsites * sizeof (GCCallSite*));
2138 i = 0;
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;
2142 }
2143
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);
2147
2148 /*
2149 * The stack frame looks like this:
2150 *
2151 * <fp + max_offset> == cfa -> <end of previous frame>
2152 * <other stack slots>
2153 * <locals>
2154 * <other stack slots>
2155 * fp + min_offset ->
2156 * ...
2157 * fp ->
2158 */
2159
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);
2162
2163 nslots = (gcfg->max_offset - gcfg->min_offset) / SIZEOF_SLOT;
2164 nregs = NREGS;
2165
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);
2176
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) {
2180 /*
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.
2184 */
2185 if ((cfg->used_int_regs & (1 << i)))
2186 set_reg_slot_everywhere (gcfg, i, SLOT_NOREF);
2187 }
2188 }
2189
2190 static inline gboolean
2191 has_bit_set (guint8 *bitmap, int width, int slot)
2192 {
2193 int i;
2194 int pos = width * slot;
2195
2196 for (i = 0; i < width; ++i) {
2197 if (bitmap [pos + i])
2198 break;
2199 }
2200 return i < width;
2201 }
2202
2203 static void
2204 create_map (MonoCompile *cfg)
2205 {
2206 GCMap *map;
2207 int i, j, nregs, nslots, nref_regs, npin_regs, alloc_size, bitmaps_size, bitmaps_offset;
2208 int ntypes [16];
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;
2211 int start, end;
2212 gboolean has_ref_slots, has_pin_slots, has_ref_regs, has_pin_regs;
2213 MonoCompileGC *gcfg = cfg->gc_info;
2214 GCCallSite **callsites;
2215 int ncallsites;
2216 guint8 *bitmap, *bitmaps;
2217 guint32 reg_ref_mask, reg_pin_mask;
2218
2219 ncallsites = gcfg->ncallsites;
2220 nslots = gcfg->nslots;
2221 nregs = gcfg->nregs;
2222 callsites = gcfg->callsites;
2223
2224 /*
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.
2227 */
2228 has_ref_slots = FALSE;
2229 has_pin_slots = FALSE;
2230 start = -1;
2231 end = -1;
2232 memset (ntypes, 0, sizeof (ntypes));
2233 for (i = 0; i < nslots; ++i) {
2234 gboolean has_ref = FALSE;
2235 gboolean has_pin = FALSE;
2236
2237 if (has_bit_set (gcfg->stack_pin_bitmap, gcfg->stack_bitmap_width, i))
2238 has_pin = TRUE;
2239 if (has_bit_set (gcfg->stack_ref_bitmap, gcfg->stack_bitmap_width, i))
2240 has_ref = TRUE;
2241
2242 if (has_ref)
2243 has_ref_slots = TRUE;
2244 if (has_pin)
2245 has_pin_slots = TRUE;
2246
2247 if (has_ref)
2248 ntypes [SLOT_REF] ++;
2249 else if (has_pin)
2250 ntypes [SLOT_PIN] ++;
2251 else
2252 ntypes [SLOT_NOREF] ++;
2253
2254 if (has_ref || has_pin) {
2255 if (start == -1)
2256 start = i;
2257 end = i + 1;
2258 }
2259 }
2260 if (start == -1) {
2261 start = end = nslots;
2262 } else {
2263 g_assert (start != -1);
2264 g_assert (start < end);
2265 }
2266
2267 has_ref_regs = FALSE;
2268 has_pin_regs = FALSE;
2269 reg_ref_mask = 0;
2270 reg_pin_mask = 0;
2271 nref_regs = 0;
2272 npin_regs = 0;
2273 for (i = 0; i < nregs; ++i) {
2274 gboolean has_ref = FALSE;
2275 gboolean has_pin = FALSE;
2276
2277 if (!(cfg->used_int_regs & (1 << i)))
2278 continue;
2279
2280 if (has_bit_set (gcfg->reg_pin_bitmap, gcfg->reg_bitmap_width, i))
2281 has_pin = TRUE;
2282 if (has_bit_set (gcfg->reg_ref_bitmap, gcfg->reg_bitmap_width, i))
2283 has_ref = TRUE;
2284
2285 if (has_ref) {
2286 reg_ref_mask |= (1 << i);
2287 has_ref_regs = TRUE;
2288 nref_regs ++;
2289 }
2290 if (has_pin) {
2291 reg_pin_mask |= (1 << i);
2292 has_pin_regs = TRUE;
2293 npin_regs ++;
2294 }
2295 }
2296
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);
2299
2300 /* Create the GC Map */
2301
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);
2310
2311 map = mono_mempool_alloc0 (cfg->mempool, sizeof (GCMap));
2312
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;
2328
2329 bitmaps = mono_mempool_alloc0 (cfg->mempool, bitmaps_size);
2330
2331 bitmaps_offset = 0;
2332 if (has_ref_slots) {
2333 map->stack_ref_bitmap_offset = bitmaps_offset;
2334 bitmaps_offset += stack_bitmap_size;
2335
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);
2341 }
2342 }
2343 }
2344 if (has_pin_slots) {
2345 map->stack_pin_bitmap_offset = bitmaps_offset;
2346 bitmaps_offset += stack_bitmap_size;
2347
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);
2353 }
2354 }
2355 }
2356 if (has_ref_regs) {
2357 map->reg_ref_bitmap_offset = bitmaps_offset;
2358 bitmaps_offset += reg_ref_bitmap_size;
2359
2360 bitmap = &bitmaps [map->reg_ref_bitmap_offset];
2361 bindex = 0;
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);
2367 }
2368 bindex ++;
2369 }
2370 }
2371 }
2372 if (has_pin_regs) {
2373 map->reg_pin_bitmap_offset = bitmaps_offset;
2374 bitmaps_offset += reg_pin_bitmap_size;
2375
2376 bitmap = &bitmaps [map->reg_pin_bitmap_offset];
2377 bindex = 0;
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);
2383 }
2384 bindex ++;
2385 }
2386 }
2387 }
2388
2389 /* Call sites */
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;
2395 else
2396 map->callsite_entry_size = 4;
2397
2398 /* Encode the GC Map */
2399 {
2400 guint8 buf [256];
2401 guint8 *endbuf;
2402 GCEncodedMap *emap;
2403 int encoded_size;
2404 guint8 *p;
2405
2406 encode_gc_map (map, buf, &endbuf);
2407 g_assert (endbuf - buf < 256);
2408
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;
2411
2412 emap = mono_domain_alloc0 (cfg->domain, alloc_size);
2413 //emap->ref_slots = map->ref_slots;
2414
2415 /* Encoded fixed fields */
2416 p = &emap->encoded [0];
2417 //emap->encoded_size = encoded_size;
2418 memcpy (p, buf, encoded_size);
2419 p += encoded_size;
2420
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));
2433 } else {
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));
2438 }
2439 p += ncallsites * map->callsite_entry_size;
2440
2441 /* Bitmaps */
2442 memcpy (p, bitmaps, bitmaps_size);
2443 p += bitmaps_size;
2444
2445 g_assert ((guint8*)p - (guint8*)emap <= alloc_size);
2446
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);
2451
2452 cfg->jit_info->gc_info = emap;
2453
2454 cfg->gc_map = (guint8*)emap;
2455 cfg->gc_map_size = alloc_size;
2456 }
2457
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]);
2462 }
2463
2464 void
2465 mini_gc_create_gc_map (MonoCompile *cfg)
2466 {
2467 if (!cfg->compute_gc_maps)
2468 return;
2469
2470 /*
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.
2474 */
2475
2476 if (!(cfg->comp_done & MONO_COMP_LIVENESS))
2477 /* Without liveness info, the live ranges are not precise enough */
2478 return;
2479
2480 mono_analyze_liveness_gc (cfg);
2481
2482 compute_frame_size (cfg);
2483
2484 init_gcfg (cfg);
2485
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);
2491
2492 create_map (cfg);
2493 }
2494
2495 #endif /* DISABLE_JIT */
2496
2497 static void
2498 parse_debug_options (void)
2499 {
2500 char **opts, **ptr;
2501 const char *env;
2502
2503 env = g_getenv ("MONO_GCMAP_DEBUG");
2504 if (!env)
2505 return;
2506
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);
2511 exit (1);
2512 }
2513 g_strfreev (opts);
2514 g_free (env);
2515 }
2516
2517 void
2518 mini_gc_init (void)
2519 {
2520 MonoGCCallbacks cb;
2521
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);
2530
2531 logfile = mono_gc_get_logfile ();
2532
2533 parse_debug_options ();
2534
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);
2549
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);
2558
2559 mono_counters_register ("GC TLS Data size",
2560 MONO_COUNTER_GC | MONO_COUNTER_INT, &stats.tlsdata_size);
2561
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);
2576 }
2577
2578 #else
2579
2580 void
mini_gc_enable_gc_maps_for_aot(void)2581 mini_gc_enable_gc_maps_for_aot (void)
2582 {
2583 }
2584
2585 void
mini_gc_init(void)2586 mini_gc_init (void)
2587 {
2588 MonoGCCallbacks cb;
2589 memset (&cb, 0, sizeof (cb));
2590 cb.get_provenance_func = get_provenance_func;
2591 mono_gc_set_gc_callbacks (&cb);
2592 }
2593
2594 #ifndef DISABLE_JIT
2595
2596 static void
mini_gc_init_gc_map(MonoCompile * cfg)2597 mini_gc_init_gc_map (MonoCompile *cfg)
2598 {
2599 }
2600
2601 void
mini_gc_create_gc_map(MonoCompile * cfg)2602 mini_gc_create_gc_map (MonoCompile *cfg)
2603 {
2604 }
2605
2606 void
mini_gc_set_slot_type_from_fp(MonoCompile * cfg,int slot_offset,GCSlotType type)2607 mini_gc_set_slot_type_from_fp (MonoCompile *cfg, int slot_offset, GCSlotType type)
2608 {
2609 }
2610
2611 void
mini_gc_set_slot_type_from_cfa(MonoCompile * cfg,int slot_offset,GCSlotType type)2612 mini_gc_set_slot_type_from_cfa (MonoCompile *cfg, int slot_offset, GCSlotType type)
2613 {
2614 }
2615
2616 #endif /* DISABLE_JIT */
2617
2618 #endif
2619
2620 #ifndef DISABLE_JIT
2621
2622 /*
2623 * mini_gc_init_cfg:
2624 *
2625 * Set GC specific options in CFG.
2626 */
2627 void
mini_gc_init_cfg(MonoCompile * cfg)2628 mini_gc_init_cfg (MonoCompile *cfg)
2629 {
2630 if (mono_gc_is_moving ()) {
2631 cfg->disable_ref_noref_stack_slot_share = TRUE;
2632 cfg->gen_write_barriers = TRUE;
2633 }
2634
2635 mini_gc_init_gc_map (cfg);
2636 }
2637
2638 #endif /* DISABLE_JIT */
2639
2640 /*
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
2650 * precise there.
2651 */
2652
2653 /*
2654 * Ideas for creating smaller GC maps:
2655 * - remove empty columns from the bitmaps. This requires adding a mask bit array for
2656 * each bitmap.
2657 * - merge reg and stack slot bitmaps, so the unused bits at the end of the reg bitmap are
2658 * not wasted.
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.
2663 */
2664