mono/profiler/decode.c

   1 /*
   2  * decode.c: mprof-report program source: decode and analyze the log profiler data
   3  *
   4  * Author:
   5  *   Paolo Molaro (lupus@ximian.com)
   6  *
   7  * Copyright 2010 Novell, Inc (http://www.novell.com)
   8  */
   9 #include <config.h>
  10 #include "utils.c"
  11 #include "proflog.h"
  12 #include <string.h>
  13 #include <assert.h>
  14 #include <stdio.h>
  15 #if !defined(__APPLE__)
  16 #include <malloc.h>
  17 #endif
  18 #include <unistd.h>
  19 #include <stdlib.h>
  20 #if defined (HAVE_SYS_ZLIB)
  21 #include <zlib.h>
  22 #endif
  23 #include <mono/metadata/profiler.h>
  24 #include <mono/metadata/object.h>
  25 #include <mono/metadata/debug-helpers.h>
  26
  27 #define HASH_SIZE 9371
  28 #define SMALL_HASH_SIZE 31
  29
  30 static int debug = 0;
  31 static int collect_traces = 0;
  32 static int show_traces = 0;
  33 static int trace_max = 6;
  34 static int verbose = 0;
  35 static uintptr_t *tracked_objects = 0;
  36 static int num_tracked_objects = 0;
  37 static uintptr_t thread_filter = 0;
  38 static uint64_t find_size = 0;
  39 static const char* find_name = NULL;
  40 static uint64_t time_from = 0;
  41 static uint64_t time_to = 0xffffffffffffffffULL;
  42 static uint64_t startup_time = 0;
  43 static FILE* outfile = NULL;
  44
  45 static int32_t
  46 read_int32 (unsigned char *p)
  47 {
  48         int32_t value = *p++;
  49         value |= (*p++) << 8;
  50         value |= (*p++) << 16;
  51         value |= (uint32_t)(*p++) << 24;
  52         return value;
  53 }
  54
  55 static int64_t
  56 read_int64 (unsigned char *p)
  57 {
  58         uint64_t value = *p++;
  59         value |= (*p++) << 8;
  60         value |= (*p++) << 16;
  61         value |= (uint64_t)(*p++) << 24;
  62         value |= (uint64_t)(*p++) << 32;
  63         value |= (uint64_t)(*p++) << 40;
  64         value |= (uint64_t)(*p++) << 48;
  65         value |= (uint64_t)(*p++) << 54;
  66         return value;
  67 }
  68
  69 static char*
  70 pstrdup (const char *s)
  71 {
  72         int len = strlen (s) + 1;
  73         char *p = malloc (len);
  74         memcpy (p, s, len);
  75         return p;
  76 }
  77
  78 static int num_images;
  79 typedef struct _ImageDesc ImageDesc;
  80 struct _ImageDesc {
  81         ImageDesc *next;
  82         intptr_t image;
  83         char *filename;
  84 };
  85
  86 static ImageDesc* image_hash [SMALL_HASH_SIZE] = {0};
  87
  88 static void
  89 add_image (intptr_t image, char *name)
  90 {
  91         int slot = ((image >> 2) & 0xffff) % SMALL_HASH_SIZE;
  92         ImageDesc *cd = malloc (sizeof (ImageDesc));
  93         cd->image = image;
  94         cd->filename = pstrdup (name);
  95         cd->next = image_hash [slot];
  96         image_hash [slot] = cd;
  97         num_images++;
  98 }
  99
 100 typedef struct _BackTrace BackTrace;
 101 typedef struct {
 102         uint64_t count;
 103         BackTrace *bt;
 104 } CallContext;
 105
 106 typedef struct {
 107         int count;
 108         int size;
 109         CallContext *traces;
 110 } TraceDesc;
 111
 112 typedef struct _ClassDesc ClassDesc;
 113 struct _ClassDesc {
 114         ClassDesc *next;
 115         intptr_t klass;
 116         char *name;
 117         intptr_t allocs;
 118         uint64_t alloc_size;
 119         TraceDesc traces;
 120 };
 121
 122 static ClassDesc* class_hash [HASH_SIZE] = {0};
 123 static int num_classes = 0;
 124
 125 static ClassDesc*
 126 add_class (intptr_t klass, const char *name)
 127 {
 128         int slot = ((klass >> 2) & 0xffff) % HASH_SIZE;
 129         ClassDesc *cd;
 130         cd = class_hash [slot];
 131         while (cd && cd->klass != klass)
 132                 cd = cd->next;
 133         /* we resolved an unknown class (unless we had the code unloaded) */
 134         if (cd) {
 135                 /*printf ("resolved unknown: %s\n", name);*/
 136                 free (cd->name);
 137                 cd->name = pstrdup (name);
 138                 return cd;
 139         }
 140         cd = calloc (sizeof (ClassDesc), 1);
 141         cd->klass = klass;
 142         cd->name = pstrdup (name);
 143         cd->next = class_hash [slot];
 144         cd->allocs = 0;
 145         cd->alloc_size = 0;
 146         cd->traces.count = 0;
 147         cd->traces.size = 0;
 148         cd->traces.traces = NULL;
 149         class_hash [slot] = cd;
 150         num_classes++;
 151         return cd;
 152 }
 153
 154 static ClassDesc *
 155 lookup_class (intptr_t klass)
 156 {
 157         int slot = ((klass >> 2) & 0xffff) % HASH_SIZE;
 158         ClassDesc *cd = class_hash [slot];
 159         while (cd && cd->klass != klass)
 160                 cd = cd->next;
 161         if (!cd)
 162                 return add_class (klass, "unresolved class");
 163         return cd;
 164 }
 165
 166 typedef struct _MethodDesc MethodDesc;
 167 struct _MethodDesc {
 168         MethodDesc *next;
 169         intptr_t method;
 170         char *name;
 171         intptr_t code;
 172         int len;
 173         int recurse_count;
 174         uint64_t calls;
 175         uint64_t total_time;
 176         uint64_t callee_time;
 177         uint64_t self_time;
 178         TraceDesc traces;
 179 };
 180
 181 static MethodDesc* method_hash [HASH_SIZE] = {0};
 182 static int num_methods = 0;
 183
 184 static MethodDesc*
 185 add_method (intptr_t method, const char *name, intptr_t code, int len)
 186 {
 187         int slot = ((method >> 2) & 0xffff) % HASH_SIZE;
 188         MethodDesc *cd;
 189         cd = method_hash [slot];
 190         while (cd && cd->method != method)
 191                 cd = cd->next;
 192         /* we resolved an unknown method (unless we had the code unloaded) */
 193         if (cd) {
 194                 cd->code = code;
 195                 cd->len = len;
 196                 /*printf ("resolved unknown: %s\n", name);*/
 197                 free (cd->name);
 198                 cd->name = pstrdup (name);
 199                 return cd;
 200         }
 201         cd = calloc (sizeof (MethodDesc), 1);
 202         cd->method = method;
 203         cd->name = pstrdup (name);
 204         cd->code = code;
 205         cd->len = len;
 206         cd->calls = 0;
 207         cd->total_time = 0;
 208         cd->traces.count = 0;
 209         cd->traces.size = 0;
 210         cd->traces.traces = NULL;
 211         cd->next = method_hash [slot];
 212         method_hash [slot] = cd;
 213         num_methods++;
 214         return cd;
 215 }
 216
 217 static MethodDesc *
 218 lookup_method (intptr_t method)
 219 {
 220         int slot = ((method >> 2) & 0xffff) % HASH_SIZE;
 221         MethodDesc *cd = method_hash [slot];
 222         while (cd && cd->method != method)
 223                 cd = cd->next;
 224         if (!cd)
 225                 return add_method (method, "unknown method", 0, 0);
 226         return cd;
 227 }
 228
 229 typedef struct _HeapClassDesc HeapClassDesc;
 230 typedef struct {
 231         HeapClassDesc *klass;
 232         uint64_t count;
 233 } HeapClassRevRef;
 234
 235 struct _HeapClassDesc {
 236         ClassDesc *klass;
 237         int64_t count;
 238         int64_t total_size;
 239         HeapClassRevRef *rev_hash;
 240         int rev_hash_size;
 241         int rev_count;
 242 };
 243
 244 static int
 245 add_rev_class_hashed (HeapClassRevRef *rev_hash, uintptr_t size, HeapClassDesc *hklass, uint64_t value)
 246 {
 247         uintptr_t i;
 248         uintptr_t start_pos;
 249         start_pos = (hklass->klass->klass >> 2) % size;
 250         assert (start_pos < size);
 251         i = start_pos;
 252         do {
 253                 if (rev_hash [i].klass == hklass) {
 254                         rev_hash [i].count += value;
 255                         return 0;
 256                 } else if (!rev_hash [i].klass) {
 257                         rev_hash [i].klass = hklass;
 258                         rev_hash [i].count += value;
 259                         start_pos = 0;
 260                         for (i = 0; i < size; ++i)
 261                                 if (rev_hash [i].klass && rev_hash [i].klass->klass == hklass->klass)
 262                                         start_pos ++;
 263                         assert (start_pos == 1);
 264                         return 1;
 265                 }
 266                 /* wrap around */
 267                 if (++i == size)
 268                         i = 0;
 269         } while (i != start_pos);
 270         /* should not happen */
 271         printf ("failed revref store\n");
 272         return 0;
 273 }
 274
 275 static void
 276 add_heap_class_rev (HeapClassDesc *from, HeapClassDesc *to)
 277 {
 278         uintptr_t i;
 279         if (to->rev_count * 2 >= to->rev_hash_size) {
 280                 HeapClassRevRef *n;
 281                 uintptr_t old_size = to->rev_hash_size;
 282                 to->rev_hash_size *= 2;
 283                 if (to->rev_hash_size == 0)
 284                         to->rev_hash_size = 4;
 285                 n = calloc (sizeof (HeapClassRevRef) * to->rev_hash_size, 1);
 286                 for (i = 0; i < old_size; ++i) {
 287                         if (to->rev_hash [i].klass)
 288                                 add_rev_class_hashed (n, to->rev_hash_size, to->rev_hash [i].klass, to->rev_hash [i].count);
 289                 }
 290                 if (to->rev_hash)
 291                         free (to->rev_hash);
 292                 to->rev_hash = n;
 293         }
 294         to->rev_count += add_rev_class_hashed (to->rev_hash, to->rev_hash_size, from, 1);
 295 }
 296
 297 typedef struct {
 298         uintptr_t objaddr;
 299         HeapClassDesc *hklass;
 300         uintptr_t num_refs;
 301         uintptr_t refs [0];
 302 } HeapObjectDesc;
 303
 304 typedef struct _HeapShot HeapShot;
 305 struct _HeapShot {
 306         HeapShot *next;
 307         uint64_t timestamp;
 308         int class_count;
 309         int hash_size;
 310         HeapClassDesc **class_hash;
 311         HeapClassDesc **sorted;
 312         HeapObjectDesc **objects_hash;
 313         uintptr_t objects_count;
 314         uintptr_t objects_hash_size;
 315 };
 316
 317 static HeapShot *heap_shots = NULL;
 318 static int num_heap_shots = 0;
 319
 320 static HeapShot*
 321 new_heap_shot (uint64_t timestamp)
 322 {
 323         HeapShot *hs = calloc (sizeof (HeapShot), 1);
 324         hs->hash_size = 4;
 325         hs->class_hash = calloc (sizeof (void*), hs->hash_size);
 326         hs->timestamp = timestamp;
 327         num_heap_shots++;
 328         hs->next = heap_shots;
 329         heap_shots = hs;
 330         return hs;
 331 }
 332
 333 static HeapClassDesc*
 334 heap_class_lookup (HeapShot *hs, ClassDesc *klass)
 335 {
 336         int i;
 337         unsigned int start_pos;
 338         start_pos = ((uintptr_t)klass->klass >> 2) % hs->hash_size;
 339         i = start_pos;
 340         do {
 341                 HeapClassDesc* cd = hs->class_hash [i];
 342                 if (!cd)
 343                         return NULL;
 344                 if (cd->klass == klass)
 345                         return cd;
 346                 /* wrap around */
 347                 if (++i == hs->hash_size)
 348                         i = 0;
 349         } while (i != start_pos);
 350         return NULL;
 351 }
 352
 353 static int
 354 add_heap_hashed (HeapClassDesc **hash, HeapClassDesc **retv, uintptr_t hsize, ClassDesc *klass, uint64_t size, uint64_t count)
 355 {
 356         uintptr_t i;
 357         uintptr_t start_pos;
 358         start_pos = ((uintptr_t)klass->klass >> 2) % hsize;
 359         i = start_pos;
 360         do {
 361                 if (hash [i] && hash [i]->klass == klass) {
 362                         hash [i]->total_size += size;
 363                         hash [i]->count += count;
 364                         *retv = hash [i];
 365                         return 0;
 366                 } else if (!hash [i]) {
 367                         if (*retv) {
 368                                 hash [i] = *retv;
 369                                 return 1;
 370                         }
 371                         hash [i] = calloc (sizeof (HeapClassDesc), 1);
 372                         hash [i]->klass = klass;
 373                         hash [i]->total_size += size;
 374                         hash [i]->count += count;
 375                         *retv = hash [i];
 376                         return 1;
 377                 }
 378                 /* wrap around */
 379                 if (++i == hsize)
 380                         i = 0;
 381         } while (i != start_pos);
 382         /* should not happen */
 383         printf ("failed heap class store\n");
 384         return 0;
 385 }
 386
 387 static HeapClassDesc*
 388 add_heap_shot_class (HeapShot *hs, ClassDesc *klass, uint64_t size)
 389 {
 390         HeapClassDesc *res;
 391         int i;
 392         if (hs->class_count * 2 >= hs->hash_size) {
 393                 HeapClassDesc **n;
 394                 int old_size = hs->hash_size;
 395                 hs->hash_size *= 2;
 396                 if (hs->hash_size == 0)
 397                         hs->hash_size = 4;
 398                 n = calloc (sizeof (void*) * hs->hash_size, 1);
 399                 for (i = 0; i < old_size; ++i) {
 400                         res = hs->class_hash [i];
 401                         if (hs->class_hash [i])
 402                                 add_heap_hashed (n, &res, hs->hash_size, hs->class_hash [i]->klass, hs->class_hash [i]->total_size, hs->class_hash [i]->count);
 403                 }
 404                 if (hs->class_hash)
 405                         free (hs->class_hash);
 406                 hs->class_hash = n;
 407         }
 408         res = NULL;
 409         hs->class_count += add_heap_hashed (hs->class_hash, &res, hs->hash_size, klass, size, 1);
 410         //if (res->count == 1)
 411         //      printf ("added heap class: %s\n", res->klass->name);
 412         return res;
 413 }
 414
 415 static HeapObjectDesc*
 416 alloc_heap_obj (uintptr_t objaddr, HeapClassDesc *hklass, uintptr_t num_refs)
 417 {
 418         HeapObjectDesc* ho = calloc (sizeof (HeapObjectDesc) + num_refs * sizeof (uintptr_t), 1);
 419         ho->objaddr = objaddr;
 420         ho->hklass = hklass;
 421         ho->num_refs = num_refs;
 422         return ho;
 423 }
 424
 425 static uintptr_t
 426 heap_shot_find_obj_slot (HeapShot *hs, uintptr_t objaddr)
 427 {
 428         uintptr_t i;
 429         uintptr_t start_pos;
 430         HeapObjectDesc **hash = hs->objects_hash;
 431         start_pos = ((uintptr_t)objaddr >> 3) % hs->objects_hash_size;
 432         i = start_pos;
 433         do {
 434                 if (hash [i] && hash [i]->objaddr == objaddr) {
 435                         return i;
 436                 } else if (!hash [i]) {
 437                         break; /* fail */
 438                 }
 439                 /* wrap around */
 440                 if (++i == hs->objects_hash_size)
 441                         i = 0;
 442         } while (i != start_pos);
 443         /* should not happen */
 444         printf ("failed heap obj update\n");
 445         return -1;
 446 }
 447
 448 static HeapObjectDesc*
 449 heap_shot_obj_add_refs (HeapShot *hs, uintptr_t objaddr, uintptr_t num, uintptr_t *ref_offset)
 450 {
 451         HeapObjectDesc **hash = hs->objects_hash;
 452         uintptr_t i = heap_shot_find_obj_slot (hs, objaddr);
 453         if (i >= 0) {
 454                 HeapObjectDesc* ho = alloc_heap_obj (objaddr, hash [i]->hklass, hash [i]->num_refs + num);
 455                 *ref_offset = hash [i]->num_refs;
 456                 memcpy (ho->refs, hash [i]->refs, hash [i]->num_refs * sizeof (uintptr_t));
 457                 free (hash [i]);
 458                 hash [i] = ho;
 459                 return ho;
 460         }
 461         /* should not happen */
 462         printf ("failed heap obj update\n");
 463         return NULL;
 464
 465 }
 466
 467 static uintptr_t
 468 add_heap_hashed_obj (HeapObjectDesc **hash, uintptr_t hsize, HeapObjectDesc *obj)
 469 {
 470         uintptr_t i;
 471         uintptr_t start_pos;
 472         start_pos = ((uintptr_t)obj->objaddr >> 3) % hsize;
 473         i = start_pos;
 474         do {
 475                 if (hash [i] && hash [i]->objaddr == obj->objaddr) {
 476                         printf ("duplicate object!\n");
 477                         return 0;
 478                 } else if (!hash [i]) {
 479                         hash [i] = obj;
 480                         return 1;
 481                 }
 482                 /* wrap around */
 483                 if (++i == hsize)
 484                         i = 0;
 485         } while (i != start_pos);
 486         /* should not happen */
 487         printf ("failed heap obj store\n");
 488         return 0;
 489 }
 490
 491 static void
 492 add_heap_shot_obj (HeapShot *hs, HeapObjectDesc *obj)
 493 {
 494         uintptr_t i;
 495         if (hs->objects_count * 2 >= hs->objects_hash_size) {
 496                 HeapObjectDesc **n;
 497                 uintptr_t old_size = hs->objects_hash_size;
 498                 hs->objects_hash_size *= 2;
 499                 if (hs->objects_hash_size == 0)
 500                         hs->objects_hash_size = 4;
 501                 n = calloc (sizeof (void*) * hs->objects_hash_size, 1);
 502                 for (i = 0; i < old_size; ++i) {
 503                         if (hs->objects_hash [i])
 504                                 add_heap_hashed_obj (n, hs->objects_hash_size, hs->objects_hash [i]);
 505                 }
 506                 if (hs->objects_hash)
 507                         free (hs->objects_hash);
 508                 hs->objects_hash = n;
 509         }
 510         hs->objects_count += add_heap_hashed_obj (hs->objects_hash, hs->objects_hash_size, obj);
 511 }
 512
 513 static void
 514 heap_shot_resolve_reverse_refs (HeapShot *hs)
 515 {
 516         uintptr_t i;
 517         for (i = 0; i < hs->objects_hash_size; ++i) {
 518                 uintptr_t r;
 519                 HeapObjectDesc *ho = hs->objects_hash [i];
 520                 if (!ho)
 521                         continue;
 522                 for (r = 0; r < ho->num_refs; ++r) {
 523                         uintptr_t oi = heap_shot_find_obj_slot (hs, ho->refs [r]);
 524                         add_heap_class_rev (ho->hklass, hs->objects_hash [oi]->hklass);
 525                 }
 526         }
 527 }
 528
 529 static void
 530 heap_shot_free_objects (HeapShot *hs)
 531 {
 532         uintptr_t i;
 533         for (i = 0; i < hs->objects_hash_size; ++i) {
 534                 HeapObjectDesc *ho = hs->objects_hash [i];
 535                 if (ho)
 536                         free (ho);
 537         }
 538         if (hs->objects_hash)
 539                 free (hs->objects_hash);
 540         hs->objects_hash = NULL;
 541         hs->objects_hash_size = 0;
 542         hs->objects_count = 0;
 543 }
 544
 545 struct _BackTrace {
 546         BackTrace *next;
 547         unsigned int hash;
 548         int count;
 549         int id;
 550         MethodDesc *methods [1];
 551 };
 552
 553 static BackTrace *backtrace_hash [HASH_SIZE];
 554 static BackTrace **backtraces = NULL;
 555 static int num_backtraces = 0;
 556 static int next_backtrace = 0;
 557
 558 static int
 559 hash_backtrace (int count, MethodDesc **methods)
 560 {
 561         int hash = count;
 562         int i;
 563         for (i = 0; i < count; ++i) {
 564                 hash = (hash << 5) - hash + methods [i]->method;
 565         }
 566         return hash;
 567 }
 568
 569 static int
 570 compare_backtrace (BackTrace *bt, int count, MethodDesc **methods)
 571 {
 572         int i;
 573         if (bt->count != count)
 574                 return 0;
 575         for (i = 0; i < count; ++i)
 576                 if (methods [i] != bt->methods [i])
 577                         return 0;
 578         return 1;
 579 }
 580
 581 static BackTrace*
 582 add_backtrace (int count, MethodDesc **methods)
 583 {
 584         int hash = hash_backtrace (count, methods);
 585         int slot = (hash & 0xffff) % HASH_SIZE;
 586         BackTrace *bt = backtrace_hash [slot];
 587         while (bt) {
 588                 if (bt->hash == hash && compare_backtrace (bt, count, methods))
 589                         return bt;
 590                 bt = bt->next;
 591         }
 592         bt = malloc (sizeof (BackTrace) + ((count - 1) * sizeof (void*)));
 593         bt->next = backtrace_hash [slot];
 594         backtrace_hash [slot] = bt;
 595         if (next_backtrace == num_backtraces) {
 596                 num_backtraces *= 2;
 597                 if (!num_backtraces)
 598                         num_backtraces = 16;
 599                 backtraces = realloc (backtraces, sizeof (void*) * num_backtraces);
 600         }
 601         bt->id = next_backtrace++;
 602         backtraces [bt->id] = bt;
 603         bt->count = count;
 604         bt->hash = hash;
 605         for (slot = 0; slot < count; ++slot)
 606                 bt->methods [slot] = methods [slot];
 607
 608         return bt;
 609 }
 610
 611 typedef struct _MonitorDesc MonitorDesc;
 612 typedef struct _ThreadContext ThreadContext;
 613
 614 typedef struct {
 615         FILE *file;
 616 #if defined (HAVE_SYS_ZLIB)
 617         gzFile *gzfile;
 618 #endif
 619         unsigned char *buf;
 620         int size;
 621         int data_version;
 622         int version_major;
 623         int version_minor;
 624         int timer_overhead;
 625         int pid;
 626         uint64_t startup_time;
 627         ThreadContext *threads;
 628         ThreadContext *current;
 629 } ProfContext;
 630
 631 struct _ThreadContext {
 632         ThreadContext *next;
 633         intptr_t thread_id;
 634         char *name;
 635         /* emulated stack */
 636         MethodDesc **stack;
 637         uint64_t *time_stack;
 638         uint64_t *callee_time_stack;
 639         uint64_t last_time;
 640         uint64_t contention_start;
 641         MonitorDesc *monitor;
 642         int stack_size;
 643         int stack_id;
 644         HeapShot *current_heap_shot;
 645 };
 646
 647 static void
 648 ensure_buffer (ProfContext *ctx, int size)
 649 {
 650         if (ctx->size < size) {
 651                 ctx->buf = realloc (ctx->buf, size);
 652                 ctx->size = size;
 653         }
 654 }
 655
 656 static int
 657 load_data (ProfContext *ctx, int size)
 658 {
 659         ensure_buffer (ctx, size);
 660 #if defined (HAVE_SYS_ZLIB)
 661         if (ctx->gzfile)
 662                 return gzread (ctx->gzfile, ctx->buf, size) == size;
 663         else
 664 #endif
 665                 return fread (ctx->buf, size, 1, ctx->file);
 666 }
 667
 668 static ThreadContext*
 669 get_thread (ProfContext *ctx, intptr_t thread_id)
 670 {
 671         ThreadContext *thread;
 672         if (ctx->current && ctx->current->thread_id == thread_id)
 673                 return ctx->current;
 674         thread = ctx->threads;
 675         while (thread) {
 676                 if (thread->thread_id == thread_id) {
 677                         return thread;
 678                 }
 679                 thread = thread->next;
 680         }
 681         thread = calloc (sizeof (ThreadContext), 1);
 682         thread->next = ctx->threads;
 683         ctx->threads = thread;
 684         thread->thread_id = thread_id;
 685         thread->last_time = 0;
 686         thread->stack_id = 0;
 687         thread->stack_size = 32;
 688         thread->stack = malloc (thread->stack_size * sizeof (void*));
 689         thread->time_stack = malloc (thread->stack_size * sizeof (uint64_t));
 690         thread->callee_time_stack = malloc (thread->stack_size * sizeof (uint64_t));
 691         return thread;
 692 }
 693
 694 static ThreadContext*
 695 load_thread (ProfContext *ctx, intptr_t thread_id)
 696 {
 697         ThreadContext *thread = get_thread (ctx, thread_id);
 698         ctx->current = thread;
 699         return thread;
 700 }
 701
 702 static void
 703 ensure_thread_stack (ThreadContext *thread)
 704 {
 705         if (thread->stack_id == thread->stack_size) {
 706                 thread->stack_size *= 2;
 707                 thread->stack = realloc (thread->stack, thread->stack_size * sizeof (void*));
 708                 thread->time_stack = realloc (thread->time_stack, thread->stack_size * sizeof (uint64_t));
 709                 thread->callee_time_stack = realloc (thread->callee_time_stack, thread->stack_size * sizeof (uint64_t));
 710         }
 711 }
 712
 713 static int
 714 add_trace_hashed (CallContext *traces, int size, BackTrace *bt, uint64_t value)
 715 {
 716         int i;
 717         unsigned int start_pos;
 718         start_pos = bt->hash % size;
 719         i = start_pos;
 720         do {
 721                 if (traces [i].bt == bt) {
 722                         traces [i].count += value;
 723                         return 0;
 724                 } else if (!traces [i].bt) {
 725                         traces [i].bt = bt;
 726                         traces [i].count += value;
 727                         return 1;
 728                 }
 729                 /* wrap around */
 730                 if (++i == size)
 731                         i = 0;
 732         } while (i != start_pos);
 733         /* should not happen */
 734         printf ("failed trace store\n");
 735         return 0;
 736 }
 737
 738 static void
 739 add_trace_bt (BackTrace *bt, TraceDesc *trace, uint64_t value)
 740 {
 741         int i;
 742         if (!collect_traces)
 743                 return;
 744         if (trace->count * 2 >= trace->size) {
 745                 CallContext *n;
 746                 int old_size = trace->size;
 747                 trace->size *= 2;
 748                 if (trace->size == 0)
 749                         trace->size = 4;
 750                 n = calloc (sizeof (CallContext) * trace->size, 1);
 751                 for (i = 0; i < old_size; ++i) {
 752                         if (trace->traces [i].bt)
 753                                 add_trace_hashed (n, trace->size, trace->traces [i].bt, trace->traces [i].count);
 754                 }
 755                 if (trace->traces)
 756                         free (trace->traces);
 757                 trace->traces = n;
 758         }
 759         trace->count += add_trace_hashed (trace->traces, trace->size, bt, value);
 760 }
 761
 762 static BackTrace*
 763 add_trace_thread (ThreadContext *thread, TraceDesc *trace, uint64_t value)
 764 {
 765         BackTrace *bt;
 766         int count = thread->stack_id;
 767         if (!collect_traces)
 768                 return NULL;
 769         if (count > trace_max)
 770                 count = trace_max;
 771         bt = add_backtrace (count, thread->stack + thread->stack_id - count);
 772         add_trace_bt (bt, trace, value);
 773         return bt;
 774 }
 775
 776 static BackTrace*
 777 add_trace_methods (MethodDesc **methods, int count, TraceDesc *trace, uint64_t value)
 778 {
 779         BackTrace *bt;
 780         if (!collect_traces)
 781                 return NULL;
 782         if (count > trace_max)
 783                 count = trace_max;
 784         bt = add_backtrace (count, methods);
 785         add_trace_bt (bt, trace, value);
 786         return bt;
 787 }
 788
 789 static int
 790 compare_callc (const void *a, const void *b)
 791 {
 792         const CallContext *A = a;
 793         const CallContext *B = b;
 794         if (B->count == A->count)
 795                 return 0;
 796         if (B->count < A->count)
 797                 return -1;
 798         return 1;
 799 }
 800
 801 static void
 802 sort_context_array (TraceDesc* traces)
 803 {
 804         int i, j;
 805         for (i = 0, j = 0; i < traces->size; ++i) {
 806                 if (traces->traces [i].bt) {
 807                         traces->traces [j].bt = traces->traces [i].bt;
 808                         traces->traces [j].count = traces->traces [i].count;
 809                         j++;
 810                 }
 811         }
 812         qsort (traces->traces, traces->count, sizeof (CallContext), compare_callc);
 813 }
 814
 815 static void
 816 push_method (ThreadContext *thread, MethodDesc *method, uint64_t timestamp)
 817 {
 818         ensure_thread_stack (thread);
 819         thread->time_stack [thread->stack_id] = timestamp;
 820         thread->callee_time_stack [thread->stack_id] = 0;
 821         thread->stack [thread->stack_id++] = method;
 822         method->recurse_count++;
 823 }
 824
 825 static void
 826 pop_method (ThreadContext *thread, MethodDesc *method, uint64_t timestamp)
 827 {
 828         method->recurse_count--;
 829         if (thread->stack_id > 0 && thread->stack [thread->stack_id - 1] == method) {
 830                 uint64_t tdiff;
 831                 thread->stack_id--;
 832                 method->calls++;
 833                 if (timestamp < thread->time_stack [thread->stack_id])
 834                         fprintf (outfile, "time went backwards for %s\n", method->name);
 835                 tdiff = timestamp - thread->time_stack [thread->stack_id];
 836                 if (thread->callee_time_stack [thread->stack_id] > tdiff)
 837                         fprintf (outfile, "callee time bigger for %s\n", method->name);
 838                 method->self_time += tdiff - thread->callee_time_stack [thread->stack_id];
 839                 method->callee_time += thread->callee_time_stack [thread->stack_id];
 840                 if (thread->stack_id)
 841                         thread->callee_time_stack [thread->stack_id - 1] += tdiff;
 842                 //fprintf (outfile, "method %s took %d\n", method->name, (int)(tdiff/1000));
 843         } else {
 844                 fprintf (outfile, "unmatched leave at stack pos: %d for method %s\n", thread->stack_id, method->name);
 845         }
 846 }
 847
 848 typedef struct {
 849         uint64_t start_time; /* move this per-thread? */
 850         uint64_t total_time;
 851         uint64_t max_time;
 852         int count;
 853 } GCDesc;
 854 static GCDesc gc_info [3];
 855 static uint64_t max_heap_size;
 856 static uint64_t gc_object_moves;
 857 static int gc_resizes;
 858 typedef struct {
 859         uint64_t created;
 860         uint64_t destroyed;
 861         uint64_t max_live;
 862         TraceDesc traces;
 863 } HandleInfo;
 864 static HandleInfo handle_info [4];
 865
 866 static const char*
 867 gc_event_name (int ev)
 868 {
 869         switch (ev) {
 870         case MONO_GC_EVENT_START: return "start";
 871         case MONO_GC_EVENT_MARK_START: return "mark start";
 872         case MONO_GC_EVENT_MARK_END: return "mark end";
 873         case MONO_GC_EVENT_RECLAIM_START: return "reclaim start";
 874         case MONO_GC_EVENT_RECLAIM_END: return "reclaim end";
 875         case MONO_GC_EVENT_END: return "end";
 876         case MONO_GC_EVENT_PRE_STOP_WORLD: return "pre stop";
 877         case MONO_GC_EVENT_POST_STOP_WORLD: return "post stop";
 878         case MONO_GC_EVENT_PRE_START_WORLD: return "pre start";
 879         case MONO_GC_EVENT_POST_START_WORLD: return "post start";
 880         default:
 881                 return "unknown";
 882         }
 883 }
 884
 885 static uint64_t clause_summary [MONO_EXCEPTION_CLAUSE_FAULT + 1];
 886 static uint64_t throw_count = 0;
 887 static TraceDesc exc_traces;
 888
 889 static const char*
 890 clause_name (int type)
 891 {
 892         switch (type) {
 893         case MONO_EXCEPTION_CLAUSE_NONE: return "catch";
 894         case MONO_EXCEPTION_CLAUSE_FILTER: return "filter";
 895         case MONO_EXCEPTION_CLAUSE_FINALLY: return "finally";
 896         case MONO_EXCEPTION_CLAUSE_FAULT: return "fault";
 897         default: return "invalid";
 898         }
 899 }
 900
 901 static uint64_t monitor_contention;
 902 static uint64_t monitor_failed;
 903 static uint64_t monitor_acquired;
 904
 905 struct _MonitorDesc {
 906         MonitorDesc *next;
 907         uintptr_t objid;
 908         uintptr_t contentions;
 909         uint64_t wait_time;
 910         uint64_t max_wait_time;
 911         TraceDesc traces;
 912 };
 913
 914 static MonitorDesc* monitor_hash [SMALL_HASH_SIZE] = {0};
 915 static int num_monitors = 0;
 916
 917 static MonitorDesc*
 918 lookup_monitor (uintptr_t objid)
 919 {
 920         int slot = ((objid >> 3) & 0xffff) % SMALL_HASH_SIZE;
 921         MonitorDesc *cd = monitor_hash [slot];
 922         while (cd && cd->objid != objid)
 923                 cd = cd->next;
 924         if (!cd) {
 925                 cd = calloc (sizeof (MonitorDesc), 1);
 926                 cd->objid = objid;
 927                 cd->next = monitor_hash [slot];
 928                 monitor_hash [slot] = cd;
 929                 num_monitors++;
 930         }
 931         return cd;
 932 }
 933
 934 static const char*
 935 monitor_ev_name (int ev)
 936 {
 937         switch (ev) {
 938         case MONO_PROFILER_MONITOR_CONTENTION: return "contended";
 939         case MONO_PROFILER_MONITOR_DONE: return "acquired";
 940         case MONO_PROFILER_MONITOR_FAIL: return "not taken";
 941         default: return "invalid";
 942         }
 943 }
 944
 945 static const char*
 946 get_handle_name (int htype)
 947 {
 948         switch (htype) {
 949         case 0: return "weak";
 950         case 1: return "weaktrack";
 951         case 2: return "normal";
 952         case 3: return "pinned";
 953         default: return "unknown";
 954         }
 955 }
 956
 957 static MethodDesc**
 958 decode_bt (MethodDesc** sframes, int *size, unsigned char *p, unsigned char **endp, intptr_t ptr_base)
 959 {
 960         MethodDesc **frames;
 961         int i;
 962         int flags = decode_uleb128 (p, &p);
 963         int count = decode_uleb128 (p, &p);
 964         if (flags != 0)
 965                 return NULL;
 966         if (count > *size)
 967                 frames = malloc (count * sizeof (void*));
 968         else
 969                 frames = sframes;
 970         for (i = 0; i < count; ++i) {
 971                 intptr_t ptrdiff = decode_sleb128 (p, &p);
 972                 frames [i] = lookup_method (ptr_base + ptrdiff);
 973         }
 974         *size = count;
 975         *endp = p;
 976         return frames;
 977 }
 978
 979 static void
 980 tracked_creation (uintptr_t obj, ClassDesc *cd, uint64_t size, BackTrace *bt, uint64_t timestamp)
 981 {
 982         int i;
 983         for (i = 0; i < num_tracked_objects; ++i) {
 984                 if (tracked_objects [i] != obj)
 985                         continue;
 986                 fprintf (outfile, "Object %p created (%s, %llu bytes) at %.3f secs.\n", (void*)obj, cd->name, size, (timestamp - startup_time)/1000000000.0);
 987                 if (bt && bt->count) {
 988                         int k;
 989                         for (k = 0; k < bt->count; ++k)
 990                                 fprintf (outfile, "\t%s\n", bt->methods [k]->name);
 991                 }
 992         }
 993 }
 994
 995 static void
 996 track_handle (uintptr_t obj, int htype, uint32_t handle)
 997 {
 998         int i;
 999         for (i = 0; i < num_tracked_objects; ++i) {
1000                 if (tracked_objects [i] == obj)
1001                         fprintf (outfile, "Object %p referenced from handle %u\n", (void*)obj, handle);
1002         }
1003 }
1004
1005 static void
1006 track_move (uintptr_t src, uintptr_t dst)
1007 {
1008         int i;
1009         for (i = 0; i < num_tracked_objects; ++i) {
1010                 if (tracked_objects [i] == src)
1011                         fprintf (outfile, "Object %p moved to %p\n", (void*)src, (void*)dst);
1012                 else if (tracked_objects [i] == dst)
1013                         fprintf (outfile, "Object %p moved from %p\n", (void*)dst, (void*)src);
1014         }
1015 }
1016
1017 static void
1018 track_obj_reference (uintptr_t obj, uintptr_t parent, ClassDesc *cd)
1019 {
1020         int i;
1021         for (i = 0; i < num_tracked_objects; ++i) {
1022                 if (tracked_objects [i] == obj)
1023                         fprintf (outfile, "Object %p referenced from %p (%s).\n", (void*)obj, (void*)parent, cd->name);
1024         }
1025 }
1026
1027 static void
1028 found_object (uintptr_t obj)
1029 {
1030         num_tracked_objects ++;
1031         tracked_objects = realloc (tracked_objects, num_tracked_objects * sizeof (tracked_objects [0]));
1032         tracked_objects [num_tracked_objects - 1] = obj;
1033 }
1034
1035 #define OBJ_ADDR(diff) ((obj_base + diff) << 3)
1036 #define LOG_TIME(base,diff) /*fprintf("outfile, time %llu + %llu near offset %d\n", base, diff, p - ctx->buf)*/
1037
1038 static int
1039 decode_buffer (ProfContext *ctx)
1040 {
1041         unsigned char *p;
1042         unsigned char *end;
1043         intptr_t thread_id;
1044         intptr_t ptr_base;
1045         intptr_t obj_base;
1046         intptr_t method_base;
1047         uint64_t time_base;
1048         uint64_t file_offset;
1049         int len, i;
1050         ThreadContext *thread;
1051
1052 #ifdef HAVE_SYS_ZLIB
1053         if (ctx->gzfile)
1054                 file_offset = gztell (ctx->gzfile);
1055         else
1056 #endif
1057                 file_offset = ftell (ctx->file);
1058         if (!load_data (ctx, 48))
1059                 return 0;
1060         p = ctx->buf;
1061         if (read_int32 (p) != BUF_ID)
1062                 return 0;
1063         len = read_int32 (p + 4);
1064         time_base = read_int64 (p + 8);
1065         ptr_base = read_int64 (p + 16);
1066         obj_base = read_int64 (p + 24);
1067         thread_id = read_int64 (p + 32);
1068         method_base = read_int64 (p + 40);
1069         if (debug)
1070                 fprintf (outfile, "buf: thread:%x, len: %d, time: %llu, file offset: %llu\n", thread_id, len, time_base, file_offset);
1071         thread = load_thread (ctx, thread_id);
1072         if (!load_data (ctx, len))
1073                 return 0;
1074         if (!startup_time) {
1075                 startup_time = time_base;
1076                 if (time_from) {
1077                         time_from += startup_time;
1078                         time_to += startup_time;
1079                 }
1080                 if (!thread->name)
1081                         thread->name = pstrdup ("Main");
1082         }
1083         for (i = 0; i < thread->stack_id; ++i)
1084                 thread->stack [i]->recurse_count++;
1085         p = ctx->buf;
1086         end = p + len;
1087         while (p < end) {
1088                 switch (*p & 0xf) {
1089                 case TYPE_GC: {
1090                         int subtype = *p & 0xf0;
1091                         uint64_t tdiff = decode_uleb128 (p + 1, &p);
1092                         LOG_TIME (time_base, tdiff);
1093                         time_base += tdiff;
1094                         if (subtype == TYPE_GC_RESIZE) {
1095                                 uint64_t new_size = decode_uleb128 (p, &p);
1096                                 if (debug)
1097                                         fprintf (outfile, "gc heap resized to %llu\n", new_size);
1098                                 gc_resizes++;
1099                                 if (new_size > max_heap_size)
1100                                         max_heap_size = new_size;
1101                         } else if (subtype == TYPE_GC_EVENT) {
1102                                 uint64_t ev = decode_uleb128 (p, &p);
1103                                 int gen = decode_uleb128 (p, &p);
1104                                 if (debug)
1105                                         fprintf (outfile, "gc event for gen%d: %s at %llu\n", gen - 1, gc_event_name (ev), time_base);
1106                                 if (gen > 2) {
1107                                         fprintf (outfile, "incorrect gc gen: %d\n", gen);
1108                                         break;
1109                                 }
1110                                 if (ev == MONO_GC_EVENT_START) {
1111                                         gc_info [gen].start_time = time_base;
1112                                         gc_info [gen].count++;
1113                                 } else if (ev == MONO_GC_EVENT_END) {
1114                                         tdiff = time_base - gc_info [gen].start_time;
1115                                         gc_info [gen].total_time += tdiff;
1116                                         if (tdiff > gc_info [gen].max_time)
1117                                                 gc_info [gen].max_time = tdiff;
1118                                 }
1119                         } else if (subtype == TYPE_GC_MOVE) {
1120                                 int j, num = decode_uleb128 (p, &p);
1121                                 gc_object_moves += num / 2;
1122                                 for (j = 0; j < num; j += 2) {
1123                                         intptr_t obj1diff = decode_sleb128 (p, &p);
1124                                         intptr_t obj2diff = decode_sleb128 (p, &p);
1125                                         if (num_tracked_objects)
1126                                                 track_move (OBJ_ADDR (obj1diff), OBJ_ADDR (obj2diff));
1127                                         if (debug) {
1128                                                 fprintf (outfile, "moved obj %p to %p\n", (void*)OBJ_ADDR (obj1diff), (void*)OBJ_ADDR (obj2diff));
1129                                         }
1130                                 }
1131                         } else if (subtype == TYPE_GC_HANDLE_CREATED) {
1132                                 int htype = decode_uleb128 (p, &p);
1133                                 uint32_t handle = decode_uleb128 (p, &p);
1134                                 intptr_t objdiff = decode_sleb128 (p, &p);
1135                                 if (htype > 3)
1136                                         return 0;
1137                                 handle_info [htype].created++;
1138                                 add_trace_thread (thread, &handle_info [htype].traces, 1);
1139                                 /* FIXME: we don't take into account timing here */
1140                                 if (handle_info [htype].created > handle_info [htype].max_live)
1141                                         handle_info [htype].max_live = handle_info [htype].created;
1142                                 if (num_tracked_objects)
1143                                         track_handle (OBJ_ADDR (objdiff), htype, handle);
1144                                 if (debug)
1145                                         fprintf (outfile, "handle (%s) %u created for object %p\n", get_handle_name (htype), handle, (void*)OBJ_ADDR (objdiff));
1146                         } else if (subtype == TYPE_GC_HANDLE_DESTROYED) {
1147                                 int htype = decode_uleb128 (p, &p);
1148                                 uint32_t handle = decode_uleb128 (p, &p);
1149                                 if (htype > 3)
1150                                         return 0;
1151                                 handle_info [htype].created--;
1152                                 if (debug)
1153                                         fprintf (outfile, "handle (%s) %u destroyed\n", get_handle_name (htype), handle);
1154                         }
1155                         break;
1156                 }
1157                 case TYPE_METADATA: {
1158                         int error = *p & TYPE_LOAD_ERR;
1159                         uint64_t tdiff = decode_uleb128 (p + 1, &p);
1160                         int mtype = *p++;
1161                         intptr_t ptrdiff = decode_sleb128 (p, &p);
1162                         LOG_TIME (time_base, tdiff);
1163                         time_base += tdiff;
1164                         if (mtype == TYPE_CLASS) {
1165                                 intptr_t imptrdiff = decode_sleb128 (p, &p);
1166                                 uint64_t flags = decode_uleb128 (p, &p);
1167                                 if (flags) {
1168                                         fprintf (outfile, "non-zero flags in class\n");
1169                                         return 0;
1170                                 }
1171                                 if (debug)
1172                                         fprintf (outfile, "loaded class %p (%s in %p) at %llu\n", (void*)(ptr_base + ptrdiff), p, (void*)(ptr_base + imptrdiff), time_base);
1173                                 if (!error)
1174                                         add_class (ptr_base + ptrdiff, (char*)p);
1175                                 while (*p) p++;
1176                                 p++;
1177                         } else if (mtype == TYPE_IMAGE) {
1178                                 uint64_t flags = decode_uleb128 (p, &p);
1179                                 if (flags) {
1180                                         fprintf (outfile, "non-zero flags in image\n");
1181                                         return 0;
1182                                 }
1183                                 if (debug)
1184                                         fprintf (outfile, "loaded image %p (%s) at %llu\n", (void*)(ptr_base + ptrdiff), p, time_base);
1185                                 if (!error)
1186                                         add_image (ptr_base + ptrdiff, (char*)p);
1187                                 while (*p) p++;
1188                                 p++;
1189                         } else if (mtype == TYPE_THREAD) {
1190                                 ThreadContext *nt;
1191                                 uint64_t flags = decode_uleb128 (p, &p);
1192                                 if (flags) {
1193                                         fprintf (outfile, "non-zero flags in thread\n");
1194                                         return 0;
1195                                 }
1196                                 nt = get_thread (ctx, ptr_base * ptrdiff);
1197                                 nt->name = pstrdup ((char*)p);
1198                                 if (debug)
1199                                         fprintf (outfile, "thread %p named: %s\n", (void*)(ptr_base + ptrdiff), p);
1200                                 while (*p) p++;
1201                                 p++;
1202                         }
1203                         break;
1204                 }
1205                 case TYPE_ALLOC: {
1206                         int has_bt = *p & TYPE_ALLOC_BT;
1207                         uint64_t tdiff = decode_uleb128 (p + 1, &p);
1208                         intptr_t ptrdiff = decode_sleb128 (p, &p);
1209                         intptr_t objdiff = decode_sleb128 (p, &p);
1210                         uint64_t len;
1211                         int num_bt = 0;
1212                         MethodDesc* sframes [8];
1213                         MethodDesc** frames = sframes;
1214                         ClassDesc *cd = lookup_class (ptr_base + ptrdiff);
1215                         len = decode_uleb128 (p, &p);
1216                         LOG_TIME (time_base, tdiff);
1217                         time_base += tdiff;
1218                         if (debug)
1219                                 fprintf (outfile, "alloced object %p, size %llu (%s) at %llu\n", (void*)OBJ_ADDR (objdiff), len, lookup_class (ptr_base + ptrdiff)->name, time_base);
1220                         if (has_bt) {
1221                                 num_bt = 8;
1222                                 frames = decode_bt (sframes, &num_bt, p, &p, ptr_base);
1223                                 if (!frames) {
1224                                         fprintf (outfile, "Cannot load backtrace\n");
1225                                         return 0;
1226                                 }
1227                         }
1228                         if ((thread_filter && thread_filter == thread->thread_id) || (time_base >= time_from && time_base < time_to)) {
1229                                 BackTrace *bt;
1230                                 cd->allocs++;
1231                                 cd->alloc_size += len;
1232                                 if (has_bt)
1233                                         bt = add_trace_methods (frames, num_bt, &cd->traces, len);
1234                                 else
1235                                         bt = add_trace_thread (thread, &cd->traces, len);
1236                                 if (find_size && len >= find_size) {
1237                                         if (!find_name || strstr (cd->name, find_name))
1238                                                 found_object (OBJ_ADDR (objdiff));
1239                                 } else if (!find_size && find_name && strstr (cd->name, find_name)) {
1240                                         found_object (OBJ_ADDR (objdiff));
1241                                 }
1242                                 if (num_tracked_objects)
1243                                         tracked_creation (OBJ_ADDR (objdiff), cd, len, bt, time_base);
1244                         }
1245                         if (frames != sframes)
1246                                 free (frames);
1247                         break;
1248                 }
1249                 case TYPE_METHOD: {
1250                         int subtype = *p & 0xf0;
1251                         uint64_t tdiff = decode_uleb128 (p + 1, &p);
1252                         int64_t ptrdiff = decode_sleb128 (p, &p);
1253                         LOG_TIME (time_base, tdiff);
1254                         time_base += tdiff;
1255                         method_base += ptrdiff;
1256                         if (subtype == TYPE_JIT) {
1257                                 intptr_t codediff = decode_sleb128 (p, &p);
1258                                 int codelen = decode_uleb128 (p, &p);
1259                                 if (debug)
1260                                         fprintf (outfile, "jitted method %p (%s), size: %d\n", (void*)(method_base), p, codelen);
1261                                 add_method (method_base, (char*)p, ptr_base + codediff, codelen);
1262                                 while (*p) p++;
1263                                 p++;
1264                         } else {
1265                                 MethodDesc *method;
1266                                 if ((thread_filter && thread_filter != thread->thread_id))
1267                                         break;
1268                                 method = lookup_method (method_base);
1269                                 if (subtype == TYPE_ENTER) {
1270                                         add_trace_thread (thread, &method->traces, 1);
1271                                         push_method (thread, method, time_base);
1272                                 } else {
1273                                         pop_method (thread, method, time_base);
1274                                 }
1275                                 if (debug)
1276                                         fprintf (outfile, "%s method %s\n", subtype == TYPE_ENTER? "enter": subtype == TYPE_EXC_LEAVE? "exleave": "leave", method->name);
1277                         }
1278                         break;
1279                 }
1280                 case TYPE_HEAP: {
1281                         int subtype = *p & 0xf0;
1282                         if (subtype == TYPE_HEAP_OBJECT) {
1283                                 HeapObjectDesc *ho;
1284                                 int i;
1285                                 intptr_t objdiff = decode_sleb128 (p + 1, &p);
1286                                 intptr_t ptrdiff = decode_sleb128 (p, &p);
1287                                 uint64_t size = decode_uleb128 (p, &p);
1288                                 uintptr_t num = decode_uleb128 (p, &p);
1289                                 uintptr_t ref_offset;
1290                                 ClassDesc *cd = lookup_class (ptr_base + ptrdiff);
1291                                 if (size) {
1292                                         HeapClassDesc *hcd = add_heap_shot_class (thread->current_heap_shot, cd, size);
1293                                         if (collect_traces) {
1294                                                 ho = alloc_heap_obj (OBJ_ADDR (objdiff), hcd, num);
1295                                                 add_heap_shot_obj (thread->current_heap_shot, ho);
1296                                                 ref_offset = 0;
1297                                         }
1298                                 } else {
1299                                         if (collect_traces)
1300                                                 ho = heap_shot_obj_add_refs (thread->current_heap_shot, OBJ_ADDR (objdiff), num, &ref_offset);
1301                                 }
1302                                 for (i = 0; i < num; ++i) {
1303                                         /* FIXME: use object distance to measure how good
1304                                          * the GC is at keeping related objects close
1305                                          */
1306                                         intptr_t obj1diff = decode_sleb128 (p, &p);
1307                                         if (collect_traces)
1308                                                 ho->refs [ref_offset + i] = OBJ_ADDR (obj1diff);
1309                                         if (num_tracked_objects)
1310                                                 track_obj_reference (OBJ_ADDR (obj1diff), OBJ_ADDR (objdiff), cd);
1311                                 }
1312                                 if (debug && size)
1313                                         fprintf (outfile, "traced object %p, size %llu (%s), refs: %d\n", (void*)OBJ_ADDR (objdiff), size, cd->name, num);
1314                         } else if (subtype == TYPE_HEAP_END) {
1315                                 uint64_t tdiff = decode_uleb128 (p + 1, &p);
1316                                 LOG_TIME (time_base, tdiff);
1317                                 time_base += tdiff;
1318                                 if (debug)
1319                                         fprintf (outfile, "heap shot end\n");
1320                                 if (collect_traces) {
1321                                         heap_shot_resolve_reverse_refs (thread->current_heap_shot);
1322                                         heap_shot_free_objects (thread->current_heap_shot);
1323                                 }
1324                                 thread->current_heap_shot = NULL;
1325                         } else if (subtype == TYPE_HEAP_START) {
1326                                 uint64_t tdiff = decode_uleb128 (p + 1, &p);
1327                                 LOG_TIME (time_base, tdiff);
1328                                 time_base += tdiff;
1329                                 if (debug)
1330                                         fprintf (outfile, "heap shot start\n");
1331                                 thread->current_heap_shot = new_heap_shot (time_base);
1332                         }
1333                         break;
1334                 }
1335                 case TYPE_MONITOR: {
1336                         int event = (*p >> 4) & 0x3;
1337                         int has_bt = *p & TYPE_MONITOR_BT;
1338                         uint64_t tdiff = decode_uleb128 (p + 1, &p);
1339                         intptr_t objdiff = decode_sleb128 (p, &p);
1340                         MethodDesc* sframes [8];
1341                         MethodDesc** frames = sframes;
1342                         int record;
1343                         int num_bt = 0;
1344                         LOG_TIME (time_base, tdiff);
1345                         time_base += tdiff;
1346                         record = (!thread_filter || thread_filter == thread->thread_id);
1347                         if (event == MONO_PROFILER_MONITOR_CONTENTION) {
1348                                 MonitorDesc *mdesc = lookup_monitor (OBJ_ADDR (objdiff));
1349                                 if (record) {
1350                                         monitor_contention++;
1351                                         mdesc->contentions++;
1352                                         thread->monitor = mdesc;
1353                                         thread->contention_start = time_base;
1354                                 }
1355                                 if (has_bt) {
1356                                         num_bt = 8;
1357                                         frames = decode_bt (sframes, &num_bt, p, &p, ptr_base);
1358                                         if (!frames) {
1359                                                 fprintf (outfile, "Cannot load backtrace\n");
1360                                                 return 0;
1361                                         }
1362                                         if (record)
1363                                                 add_trace_methods (frames, num_bt, &mdesc->traces, 1);
1364                                 } else {
1365                                         if (record)
1366                                                 add_trace_thread (thread, &mdesc->traces, 1);
1367                                 }
1368                         } else if (event == MONO_PROFILER_MONITOR_FAIL) {
1369                                 if (record) {
1370                                         monitor_failed++;
1371                                         if (thread->monitor && thread->contention_start) {
1372                                                 uint64_t wait_time = time_base - thread->contention_start;
1373                                                 if (wait_time > thread->monitor->max_wait_time)
1374                                                         thread->monitor->max_wait_time = wait_time;
1375                                                 thread->monitor->wait_time += wait_time;
1376                                                 thread->monitor = NULL;
1377                                                 thread->contention_start = 0;
1378                                         }
1379                                 }
1380                         } else if (event == MONO_PROFILER_MONITOR_DONE) {
1381                                 if (record) {
1382                                         monitor_acquired++;
1383                                         if (thread->monitor && thread->contention_start) {
1384                                                 uint64_t wait_time = time_base - thread->contention_start;
1385                                                 if (wait_time > thread->monitor->max_wait_time)
1386                                                         thread->monitor->max_wait_time = wait_time;
1387                                                 thread->monitor->wait_time += wait_time;
1388                                                 thread->monitor = NULL;
1389                                                 thread->contention_start = 0;
1390                                         }
1391                                 }
1392                         }
1393                         if (debug)
1394                                 fprintf (outfile, "monitor %s for object %p\n", monitor_ev_name (event), (void*)OBJ_ADDR (objdiff));
1395                         if (frames != sframes)
1396                                 free (frames);
1397                         break;
1398                 }
1399                 case TYPE_EXCEPTION: {
1400                         int subtype = *p & 0x70;
1401                         int has_bt = *p & TYPE_EXCEPTION_BT;
1402                         uint64_t tdiff = decode_uleb128 (p + 1, &p);
1403                         MethodDesc* sframes [8];
1404                         MethodDesc** frames = sframes;
1405                         int record;
1406                         LOG_TIME (time_base, tdiff);
1407                         time_base += tdiff;
1408                         record = (!thread_filter || thread_filter == thread->thread_id);
1409                         if (subtype == TYPE_CLAUSE) {
1410                                 int clause_type = decode_uleb128 (p, &p);
1411                                 int clause_num = decode_uleb128 (p, &p);
1412                                 int64_t ptrdiff = decode_sleb128 (p, &p);
1413                                 method_base += ptrdiff;
1414                                 if (record)
1415                                         clause_summary [clause_type]++;
1416                                 if (debug)
1417                                         fprintf (outfile, "clause %s (%d) in method %s\n", clause_name (clause_type), clause_num, lookup_method (method_base)->name);
1418                         } else {
1419                                 intptr_t objdiff = decode_sleb128 (p, &p);
1420                                 if (record)
1421                                         throw_count++;
1422                                 if (has_bt) {
1423                                         has_bt = 8;
1424                                         frames = decode_bt (sframes, &has_bt, p, &p, ptr_base);
1425                                         if (!frames) {
1426                                                 fprintf (outfile, "Cannot load backtrace\n");
1427                                                 return 0;
1428                                         }
1429                                         if (record)
1430                                                 add_trace_methods (frames, has_bt, &exc_traces, 1);
1431                                 } else {
1432                                         if (record)
1433                                                 add_trace_thread (thread, &exc_traces, 1);
1434                                 }
1435                                 if (frames != sframes)
1436                                         free (frames);
1437                                 if (debug)
1438                                         fprintf (outfile, "throw %p\n", (void*)OBJ_ADDR (objdiff));
1439                         }
1440                         break;
1441                 }
1442                 default:
1443                         fprintf (outfile, "unhandled profiler event: 0x%x at file offset: %llu + %d (len: %d\n)\n", *p, file_offset, p - ctx->buf, len);
1444                         exit (1);
1445                 }
1446         }
1447         thread->last_time = time_base;
1448         for (i = 0; i < thread->stack_id; ++i)
1449                 thread->stack [i]->recurse_count = 0;
1450         return 1;
1451 }
1452
1453 static ProfContext*
1454 load_file (char *name)
1455 {
1456         unsigned char *p;
1457         ProfContext *ctx = calloc (sizeof (ProfContext), 1);
1458         if (strcmp (name, "-") == 0)
1459                 ctx->file = stdin;
1460         else
1461                 ctx->file = fopen (name, "rb");
1462         if (!ctx->file) {
1463                 printf ("Cannot open file: %s\n", name);
1464                 exit (1);
1465         }
1466 #if defined (HAVE_SYS_ZLIB)
1467         ctx->gzfile = gzdopen (fileno (ctx->file), "rb");
1468 #endif
1469         if (!load_data (ctx, 32))
1470                 return NULL;
1471         p = ctx->buf;
1472         if (read_int32 (p) != LOG_HEADER_ID || p [6] != LOG_DATA_VERSION)
1473                 return NULL;
1474         ctx->version_major = p [4];
1475         ctx->version_minor = p [5];
1476         ctx->data_version = p [6];
1477         /* reading 64 bit files on 32 bit systems not supported yet */
1478         if (p [7] > sizeof (void*))
1479                 return NULL;
1480         if (read_int32 (p + 20)) /* flags must be 0 */
1481                 return NULL;
1482         ctx->startup_time = read_int64 (p + 8);
1483         ctx->timer_overhead = read_int32 (p + 16);
1484         ctx->pid = read_int32 (p + 24);
1485         return ctx;
1486 }
1487
1488 enum {
1489         ALLOC_SORT_BYTES,
1490         ALLOC_SORT_COUNT
1491 };
1492 static int alloc_sort_mode = ALLOC_SORT_BYTES;
1493
1494 static int
1495 compare_class (const void *a, const void *b)
1496 {
1497         ClassDesc *const*A = a;
1498         ClassDesc *const*B = b;
1499         uint64_t vala, valb;
1500         if (alloc_sort_mode == ALLOC_SORT_BYTES) {
1501                 vala = (*A)->alloc_size;
1502                 valb = (*B)->alloc_size;
1503         } else {
1504                 vala = (*A)->allocs;
1505                 valb = (*B)->allocs;
1506         }
1507         if (valb == vala)
1508                 return 0;
1509         if (valb < vala)
1510                 return -1;
1511         return 1;
1512 }
1513
1514 static void
1515 dump_header (ProfContext *ctx)
1516 {
1517         time_t st = ctx->startup_time / 1000;
1518         char *t = ctime (&st);
1519         fprintf (outfile, "\nMono log profiler data\n");
1520         fprintf (outfile, "\tProfiler version: %d.%d\n", ctx->version_major, ctx->version_minor);
1521         fprintf (outfile, "\tData version: %d\n", ctx->data_version);
1522         fprintf (outfile, "\tMean timer overhead: %d nanoseconds\n", ctx->timer_overhead);
1523         fprintf (outfile, "\tProgram startup: %s", t);
1524         if (ctx->pid)
1525                 fprintf (outfile, "\tProgram ID: %d\n", ctx->pid);
1526 }
1527
1528 static void
1529 dump_traces (TraceDesc *traces, const char *desc)
1530 {
1531         int j;
1532         if (!show_traces)
1533                 return;
1534         if (!traces->count)
1535                 return;
1536         sort_context_array (traces);
1537         for (j = 0; j < traces->count; ++j) {
1538                 int k;
1539                 BackTrace *bt;
1540                 bt = traces->traces [j].bt;
1541                 if (!bt->count)
1542                         continue;
1543                 fprintf (outfile, "\t%llu %s from:\n", traces->traces [j].count, desc);
1544                 for (k = 0; k < bt->count; ++k)
1545                         fprintf (outfile, "\t\t%s\n", bt->methods [k]->name);
1546         }
1547 }
1548
1549 static void
1550 dump_threads (ProfContext *ctx)
1551 {
1552         ThreadContext *thread;
1553         fprintf (outfile, "\nThread summary\n");
1554         for (thread = ctx->threads; thread; thread = thread->next) {
1555                 fprintf (outfile, "\tThread: %p, name: \"%s\"\n", (void*)thread->thread_id, thread->name? thread->name: "");
1556         }
1557 }
1558
1559 static void
1560 dump_exceptions (void)
1561 {
1562         int i;
1563         fprintf (outfile, "\nException summary\n");
1564         fprintf (outfile, "\tThrows: %llu\n", throw_count);
1565         dump_traces (&exc_traces, "throws");
1566         for (i = 0; i <= MONO_EXCEPTION_CLAUSE_FAULT; ++i) {
1567                 if (!clause_summary [i])
1568                         continue;
1569                 fprintf (outfile, "\tExecuted %s clauses: %llu\n", clause_name (i), clause_summary [i]);
1570         }
1571 }
1572
1573 static int
1574 compare_monitor (const void *a, const void *b)
1575 {
1576         MonitorDesc *const*A = a;
1577         MonitorDesc *const*B = b;
1578         if ((*B)->wait_time == (*A)->wait_time)
1579                 return 0;
1580         if ((*B)->wait_time < (*A)->wait_time)
1581                 return -1;
1582         return 1;
1583 }
1584
1585 static void
1586 dump_monitors (void)
1587 {
1588         MonitorDesc **monitors;
1589         int i, j;
1590         if (!num_monitors)
1591                 return;
1592         monitors = malloc (sizeof (void*) * num_monitors);
1593         for (i = 0, j = 0; i < SMALL_HASH_SIZE; ++i) {
1594                 MonitorDesc *mdesc = monitor_hash [i];
1595                 while (mdesc) {
1596                         monitors [j++] = mdesc;
1597                         mdesc = mdesc->next;
1598                 }
1599         }
1600         qsort (monitors, num_monitors, sizeof (void*), compare_monitor);
1601         fprintf (outfile, "\nMonitor lock summary\n");
1602         for (i = 0; i < num_monitors; ++i) {
1603                 MonitorDesc *mdesc = monitors [i];
1604                 fprintf (outfile, "\tLock object %p: %d contentions\n", (void*)mdesc->objid, (int)mdesc->contentions);
1605                 fprintf (outfile, "\t\t%.6f secs total wait time, %.6f max, %.6f average\n",
1606                         mdesc->wait_time/1000000000.0, mdesc->max_wait_time/1000000000.0, mdesc->wait_time/1000000000.0/mdesc->contentions);
1607                 dump_traces (&mdesc->traces, "contentions");
1608         }
1609         fprintf (outfile, "\tLock contentions: %llu\n", monitor_contention);
1610         fprintf (outfile, "\tLock acquired: %llu\n", monitor_acquired);
1611         fprintf (outfile, "\tLock failures: %llu\n", monitor_failed);
1612 }
1613
1614 static void
1615 dump_gcs (void)
1616 {
1617         int i;
1618         fprintf (outfile, "\nGC summary\n");
1619         fprintf (outfile, "\tGC resizes: %d\n", gc_resizes);
1620         fprintf (outfile, "\tMax heap size: %llu\n", max_heap_size);
1621         fprintf (outfile, "\tObject moves: %llu\n", gc_object_moves);
1622         for (i = 0; i < 3; ++i) {
1623                 if (!gc_info [i].count)
1624                         continue;
1625                 fprintf (outfile, "\tGen%d collections: %d, max time: %lluus, total time: %lluus, average: %lluus\n",
1626                         i, gc_info [i].count, gc_info [i].max_time / 1000, gc_info [i].total_time / 1000,
1627                         gc_info [i].total_time / gc_info [i].count / 1000);
1628         }
1629         for (i = 0; i < 3; ++i) {
1630                 if (!handle_info [i].max_live)
1631                         continue;
1632                 fprintf (outfile, "\tGC handles %s: created: %llu, destroyed: %llu, max: %llu\n",
1633                         get_handle_name (i), handle_info [i].created, handle_info [i].destroyed, handle_info [i].max_live);
1634                 dump_traces (&handle_info [i].traces, "created");
1635         }
1636 }
1637
1638 static void
1639 dump_allocations (void)
1640 {
1641         int i, c;
1642         intptr_t allocs = 0;
1643         uint64_t size = 0;
1644         int header_done = 0;
1645         ClassDesc **classes = malloc (num_classes * sizeof (void*));
1646         ClassDesc *cd;
1647         c = 0;
1648         for (i = 0; i < HASH_SIZE; ++i) {
1649                 cd = class_hash [i];
1650                 while (cd) {
1651                         classes [c++] = cd;
1652                         cd = cd->next;
1653                 }
1654         }
1655         qsort (classes, num_classes, sizeof (void*), compare_class);
1656         for (i = 0; i < num_classes; ++i) {
1657                 cd = classes [i];
1658                 if (!cd->allocs)
1659                         continue;
1660                 allocs += cd->allocs;
1661                 size += cd->alloc_size;
1662                 if (!header_done++) {
1663                         fprintf (outfile, "\nAllocation summary\n");
1664                         fprintf (outfile, "%10s %10s %8s Type name\n", "Bytes", "Count", "Average");
1665                 }
1666                 fprintf (outfile, "%10llu %10d %8llu %s\n", cd->alloc_size, cd->allocs, cd->alloc_size / cd->allocs, cd->name);
1667                 dump_traces (&cd->traces, "bytes");
1668         }
1669         if (allocs)
1670                 fprintf (outfile, "Total memory allocated: %llu bytes in %d objects\n", size, allocs);
1671 }
1672
1673 enum {
1674         METHOD_SORT_TOTAL,
1675         METHOD_SORT_SELF,
1676         METHOD_SORT_CALLS
1677 };
1678
1679 static int method_sort_mode = METHOD_SORT_TOTAL;
1680
1681 static int
1682 compare_method (const void *a, const void *b)
1683 {
1684         MethodDesc *const*A = a;
1685         MethodDesc *const*B = b;
1686         uint64_t vala, valb;
1687         if (method_sort_mode == METHOD_SORT_SELF) {
1688                 vala = (*A)->self_time;
1689                 valb = (*B)->self_time;
1690         } else if (method_sort_mode == METHOD_SORT_CALLS) {
1691                 vala = (*A)->calls;
1692                 valb = (*B)->calls;
1693         } else {
1694                 vala = (*A)->total_time;
1695                 valb = (*B)->total_time;
1696         }
1697         if (vala == valb)
1698                 return 0;
1699         if (valb < vala)
1700                 return -1;
1701         return 1;
1702 }
1703
1704 static void
1705 dump_metadata (void)
1706 {
1707         fprintf (outfile, "\nMetadata summary\n");
1708         fprintf (outfile, "\tLoaded images: %d\n", num_images);
1709         if (verbose) {
1710                 ImageDesc *image;
1711                 int i;
1712                 for (i = 0; i < SMALL_HASH_SIZE; ++i) {
1713                         image = image_hash [i];
1714                         while (image) {
1715                                 fprintf (outfile, "\t\t%s\n", image->filename);
1716                                 image = image->next;
1717                         }
1718                 }
1719         }
1720
1721 }
1722
1723 static void
1724 dump_methods (void)
1725 {
1726         int i, c;
1727         uint64_t calls = 0;
1728         int header_done = 0;
1729         MethodDesc **methods = malloc (num_methods * sizeof (void*));
1730         MethodDesc *cd;
1731         c = 0;
1732         for (i = 0; i < HASH_SIZE; ++i) {
1733                 cd = method_hash [i];
1734                 while (cd) {
1735                         cd->total_time = cd->self_time + cd->callee_time;
1736                         methods [c++] = cd;
1737                         cd = cd->next;
1738                 }
1739         }
1740         qsort (methods, num_methods, sizeof (void*), compare_method);
1741         for (i = 0; i < num_methods; ++i) {
1742                 uint64_t msecs;
1743                 uint64_t smsecs;
1744                 cd = methods [i];
1745                 if (!cd->calls)
1746                         continue;
1747                 calls += cd->calls;
1748                 msecs = cd->total_time / 1000000;
1749                 smsecs = (cd->total_time - cd->callee_time) / 1000000;
1750                 if (!msecs && !verbose)
1751                         continue;
1752                 if (!header_done++) {
1753                         fprintf (outfile, "\nMethod call summary\n");
1754                         fprintf (outfile, "%8s %8s %10s Method name\n", "Total(ms)", "Self(ms)", "Calls");
1755                 }
1756                 fprintf (outfile, "%8llu %8llu %10llu %s\n", msecs, smsecs, cd->calls, cd->name);
1757                 dump_traces (&cd->traces, "calls");
1758         }
1759         if (calls)
1760                 fprintf (outfile, "Total calls: %llu\n", calls);
1761 }
1762
1763 static int
1764 compare_heap_class (const void *a, const void *b)
1765 {
1766         HeapClassDesc *const*A = a;
1767         HeapClassDesc *const*B = b;
1768         uint64_t vala, valb;
1769         if (alloc_sort_mode == ALLOC_SORT_BYTES) {
1770                 vala = (*A)->total_size;
1771                 valb = (*B)->total_size;
1772         } else {
1773                 vala = (*A)->count;
1774                 valb = (*B)->count;
1775         }
1776         if (valb == vala)
1777                 return 0;
1778         if (valb < vala)
1779                 return -1;
1780         return 1;
1781 }
1782
1783 static int
1784 compare_rev_class (const void *a, const void *b)
1785 {
1786         const HeapClassRevRef *A = a;
1787         const HeapClassRevRef *B = b;
1788         if (B->count == A->count)
1789                 return 0;
1790         if (B->count < A->count)
1791                 return -1;
1792         return 1;
1793 }
1794
1795 static void
1796 dump_rev_claases (HeapClassRevRef *revs, int count)
1797 {
1798         int j;
1799         if (!show_traces)
1800                 return;
1801         if (!count)
1802                 return;
1803         for (j = 0; j < count; ++j) {
1804                 HeapClassDesc *cd = revs [j].klass;
1805                 fprintf (outfile, "\t\t%llu references from: %s\n", revs [j].count, cd->klass->name);
1806         }
1807 }
1808
1809 static void
1810 heap_shot_summary (HeapShot *hs, int hs_num, HeapShot *last_hs)
1811 {
1812         uint64_t size = 0;
1813         uint64_t count = 0;
1814         int ccount = 0;
1815         int i;
1816         HeapClassDesc *cd;
1817         HeapClassDesc **sorted;
1818         sorted = malloc (sizeof (void*) * hs->class_count);
1819         for (i = 0; i < hs->hash_size; ++i) {
1820                 cd = hs->class_hash [i];
1821                 if (!cd)
1822                         continue;
1823                 count += cd->count;
1824                 size += cd->total_size;
1825                 sorted [ccount++] = cd;
1826         }
1827         hs->sorted = sorted;
1828         qsort (sorted, ccount, sizeof (void*), compare_heap_class);
1829         fprintf (outfile, "\n\tHeap shot %d at %.3f secs: size: %llu, object count: %llu, class count: %d\n",
1830                 hs_num, (hs->timestamp - startup_time)/1000000000.0, size, count, ccount);
1831         if (!verbose && ccount > 30)
1832                 ccount = 30;
1833         fprintf (outfile, "\t%10s %10s %8s Class name\n", "Bytes", "Count", "Average");
1834         for (i = 0; i < ccount; ++i) {
1835                 HeapClassRevRef *rev_sorted;
1836                 int j, k;
1837                 HeapClassDesc *ocd = NULL;
1838                 cd = sorted [i];
1839                 if (last_hs)
1840                         ocd = heap_class_lookup (last_hs, cd->klass);
1841                 fprintf (outfile, "\t%10llu %10llu %8llu %s", cd->total_size, cd->count, cd->total_size / cd->count, cd->klass->name);
1842                 if (ocd) {
1843                         int64_t bdiff = cd->total_size - ocd->total_size;
1844                         int64_t cdiff = cd->count - ocd->count;
1845                         fprintf (outfile, " (bytes: %+lld, count: %+lld)\n", bdiff, cdiff);
1846                 } else {
1847                         fprintf (outfile, "\n");
1848                 }
1849                 if (!collect_traces)
1850                         continue;
1851                 rev_sorted = malloc (cd->rev_count * sizeof (HeapClassRevRef));
1852                 k = 0;
1853                 for (j = 0; j < cd->rev_hash_size; ++j) {
1854                         if (cd->rev_hash [j].klass)
1855                                 rev_sorted [k++] = cd->rev_hash [j];
1856                 }
1857                 assert (cd->rev_count == k);
1858                 qsort (rev_sorted, cd->rev_count, sizeof (HeapClassRevRef), compare_rev_class);
1859                 dump_rev_claases (rev_sorted, cd->rev_count);
1860                 free (rev_sorted);
1861         }
1862         free (sorted);
1863 }
1864
1865 static int
1866 compare_heap_shots (const void *a, const void *b)
1867 {
1868         HeapShot *const*A = a;
1869         HeapShot *const*B = b;
1870         if ((*B)->timestamp == (*A)->timestamp)
1871                 return 0;
1872         if ((*B)->timestamp > (*A)->timestamp)
1873                 return -1;
1874         return 1;
1875 }
1876
1877 static void
1878 dump_heap_shots (void)
1879 {
1880         HeapShot **hs_sorted;
1881         HeapShot *hs;
1882         HeapShot *last_hs = NULL;
1883         int i;
1884         if (!heap_shots)
1885                 return;
1886         hs_sorted = malloc (num_heap_shots * sizeof (void*));
1887         fprintf (outfile, "\nHeap shot summary\n");
1888         i = 0;
1889         for (hs = heap_shots; hs; hs = hs->next)
1890                 hs_sorted [i++] = hs;
1891         qsort (hs_sorted, num_heap_shots, sizeof (void*), compare_heap_shots);
1892         for (i = 0; i < num_heap_shots; ++i) {
1893                 hs = hs_sorted [i];
1894                 heap_shot_summary (hs, i, last_hs);
1895                 last_hs = hs;
1896         }
1897 }
1898
1899 static void
1900 flush_context (ProfContext *ctx)
1901 {
1902         ThreadContext *thread;
1903         /* FIXME: sometimes there are leftovers: indagate */
1904         for (thread = ctx->threads; thread; thread = thread->next) {
1905                 while (thread->stack_id) {
1906                         if (debug)
1907                                 fprintf (outfile, "thread %p has %d items on stack\n", (void*)thread->thread_id, thread->stack_id);
1908                         pop_method (thread, thread->stack [thread->stack_id - 1], thread->last_time);
1909                 }
1910         }
1911 }
1912
1913 static const char *reports = "header,gc,alloc,call,metadata,exception,monitor,thread,heapshot";
1914
1915 static const char*
1916 match_option (const char *p, const char *opt)
1917 {
1918         int len = strlen (opt);
1919         if (strncmp (p, opt, len) == 0) {
1920                 if (p [len] == ',')
1921                         len++;
1922                 return p + len;
1923         }
1924         return p;
1925 }
1926
1927 static int
1928 print_reports (ProfContext *ctx, const char *reps, int parse_only)
1929 {
1930         const char *opt;
1931         const char *p;
1932         for (p = reps; *p; p = opt) {
1933                 if ((opt = match_option (p, "header")) != p) {
1934                         if (!parse_only)
1935                                 dump_header (ctx);
1936                         continue;
1937                 }
1938                 if ((opt = match_option (p, "thread")) != p) {
1939                         if (!parse_only)
1940                                 dump_threads (ctx);
1941                         continue;
1942                 }
1943                 if ((opt = match_option (p, "gc")) != p) {
1944                         if (!parse_only)
1945                                 dump_gcs ();
1946                         continue;
1947                 }
1948                 if ((opt = match_option (p, "alloc")) != p) {
1949                         if (!parse_only)
1950                                 dump_allocations ();
1951                         continue;
1952                 }
1953                 if ((opt = match_option (p, "call")) != p) {
1954                         if (!parse_only)
1955                                 dump_methods ();
1956                         continue;
1957                 }
1958                 if ((opt = match_option (p, "metadata")) != p) {
1959                         if (!parse_only)
1960                                 dump_metadata ();
1961                         continue;
1962                 }
1963                 if ((opt = match_option (p, "exception")) != p) {
1964                         if (!parse_only)
1965                                 dump_exceptions ();
1966                         continue;
1967                 }
1968                 if ((opt = match_option (p, "monitor")) != p) {
1969                         if (!parse_only)
1970                                 dump_monitors ();
1971                         continue;
1972                 }
1973                 if ((opt = match_option (p, "heapshot")) != p) {
1974                         if (!parse_only)
1975                                 dump_heap_shots ();
1976                         continue;
1977                 }
1978                 return 0;
1979         }
1980         return 1;
1981 }
1982
1983 static int
1984 add_find_spec (const char *p)
1985 {
1986         if (p [0] == 'S' && p [1] == ':') {
1987                 char *vale;
1988                 find_size = strtoul (p + 2, &vale, 10);
1989                 return 1;
1990         } else if (p [0] == 'T' && p [1] == ':') {
1991                 find_name = p + 2;
1992                 return 1;
1993         }
1994         return 0;
1995 }
1996
1997 static void
1998 usage (void)
1999 {
2000         printf ("Mono log profiler report version %d.%d\n", LOG_VERSION_MAJOR, LOG_VERSION_MINOR);
2001         printf ("Usage: mprof-report [OPTIONS] FILENAME\n");
2002         printf ("FILENAME can be '-' to read from standard input.\n");
2003         printf ("Options:\n");
2004         printf ("\t--help               display this help\n");
2005         printf ("\t--out=FILE           write to FILE instead of stdout\n");
2006         printf ("\t--traces             collect and show backtraces\n");
2007         printf ("\t--maxframes=NUM      limit backtraces to NUM entries\n");
2008         printf ("\t--reports=R1[,R2...] print the specified reports. Defaults are:\n");
2009         printf ("\t                     %s\n", reports);
2010         printf ("\t--method-sort=MODE   sort methods according to MODE: total, self, calls\n");
2011         printf ("\t--alloc-sort=MODE    sort allocations according to MODE: bytes, count\n");
2012         printf ("\t--track=OB1[,OB2...] track what happens to objects OBJ1, O2 etc.\n");
2013         printf ("\t--find=FINDSPEC      find and track objects matching FINFSPEC, where FINDSPEC is:\n");
2014         printf ("\t                     S:minimum_size or T:partial_name\n");
2015         printf ("\t--thread=THREADID    consider just the data for thread THREADID\n");
2016         printf ("\t--time=FROM-TO       consider data FROM seconds from startup up to TO seconds\n");
2017         printf ("\t--verbose            increase verbosity level\n");
2018         printf ("\t--debug              display decoding debug info for mprof-report devs\n");
2019 }
2020
2021 int
2022 main (int argc, char *argv[])
2023 {
2024         ProfContext *ctx;
2025         int i;
2026         outfile = stdout;
2027         for (i = 1; i < argc; ++i) {
2028                 if (strcmp ("--debug", argv [i]) == 0) {
2029                         debug++;
2030                 } else if (strcmp ("--help", argv [i]) == 0) {
2031                         usage ();
2032                         return 0;
2033                 } else if (strncmp ("--alloc-sort=", argv [i], 13) == 0) {
2034                         const char *val = argv [i] + 13;
2035                         if (strcmp (val, "bytes") == 0) {
2036                                 alloc_sort_mode = ALLOC_SORT_BYTES;
2037                         } else if (strcmp (val, "count") == 0) {
2038                                 alloc_sort_mode = ALLOC_SORT_COUNT;
2039                         } else {
2040                                 usage ();
2041                                 return 1;
2042                         }
2043                 } else if (strncmp ("--method-sort=", argv [i], 14) == 0) {
2044                         const char *val = argv [i] + 14;
2045                         if (strcmp (val, "total") == 0) {
2046                                 method_sort_mode = METHOD_SORT_TOTAL;
2047                         } else if (strcmp (val, "self") == 0) {
2048                                 method_sort_mode = METHOD_SORT_SELF;
2049                         } else if (strcmp (val, "calls") == 0) {
2050                                 method_sort_mode = METHOD_SORT_CALLS;
2051                         } else {
2052                                 usage ();
2053                                 return 1;
2054                         }
2055                 } else if (strncmp ("--reports=", argv [i], 10) == 0) {
2056                         const char *val = argv [i] + 10;
2057                         if (!print_reports (NULL, val, 1)) {
2058                                 usage ();
2059                                 return 1;
2060                         }
2061                         reports = val;
2062                 } else if (strncmp ("--out=", argv [i], 6) == 0) {
2063                         const char *val = argv [i] + 6;
2064                         outfile = fopen (val, "w");
2065                         if (!outfile) {
2066                                 printf ("Cannot open output file: %s\n", val);
2067                                 return 1;
2068                         }
2069                 } else if (strncmp ("--maxframes=", argv [i], 12) == 0) {
2070                         const char *val = argv [i] + 12;
2071                         char *vale;
2072                         trace_max = strtoul (val, &vale, 10);
2073                 } else if (strncmp ("--find=", argv [i], 7) == 0) {
2074                         const char *val = argv [i] + 7;
2075                         if (!add_find_spec (val)) {
2076                                 usage ();
2077                                 return 1;
2078                         }
2079                 } else if (strncmp ("--track=", argv [i], 8) == 0) {
2080                         const char *val = argv [i] + 8;
2081                         char *vale;
2082                         while (*val) {
2083                                 uintptr_t tracked_obj;
2084                                 if (*val == ',') {
2085                                         val++;
2086                                         continue;
2087                                 }
2088                                 tracked_obj = strtoul (val, &vale, 0);
2089                                 found_object (tracked_obj);
2090                                 val = vale;
2091                         }
2092                 } else if (strncmp ("--thread=", argv [i], 9) == 0) {
2093                         const char *val = argv [i] + 9;
2094                         char *vale;
2095                         thread_filter = strtoul (val, &vale, 0);
2096                 } else if (strncmp ("--time=", argv [i], 7) == 0) {
2097                         char *val = pstrdup (argv [i] + 7);
2098                         double from_secs, to_secs;
2099                         char *top = strchr (val, '-');
2100                         if (!top) {
2101                                 usage ();
2102                                 return 1;
2103                         }
2104                         *top++ = 0;
2105                         from_secs = atof (val);
2106                         to_secs = atof (top);
2107                         free (val);
2108                         if (from_secs > to_secs) {
2109                                 usage ();
2110                                 return 1;
2111                         }
2112                         time_from = from_secs * 1000000000;
2113                         time_to = to_secs * 1000000000;
2114                 } else if (strcmp ("--verbose", argv [i]) == 0) {
2115                         verbose++;
2116                 } else if (strcmp ("--traces", argv [i]) == 0) {
2117                         show_traces = 1;
2118                         collect_traces = 1;
2119                 } else {
2120                         break;
2121                 }
2122         }
2123         if (i >= argc) {
2124                 usage ();
2125                 return 2;
2126         }
2127         ctx = load_file (argv [i]);
2128         if (!ctx) {
2129                 printf ("Not a log profiler data file (or unsupported version).\n");
2130                 return 1;
2131         }
2132         while (decode_buffer (ctx));
2133         flush_context (ctx);
2134         if (num_tracked_objects)
2135                 return 0;
2136         print_reports (ctx, reports, 0);
2137         return 0;
2138 }
2139