Home | History | Annotate | Download | only in callgrind
      1 /*--------------------------------------------------------------------*/
      2 /*--- Callgrind                                                    ---*/
      3 /*---                                                     global.h ---*/
      4 /*--- (C) 2004, 2005 Josef Weidendorfer                            ---*/
      5 /*--------------------------------------------------------------------*/
      6 
      7 #ifndef CLG_GLOBAL
      8 #define CLG_GLOBAL
      9 
     10 #include "pub_tool_basics.h"
     11 #include "pub_tool_vki.h"
     12 #include "pub_tool_debuginfo.h"
     13 #include "pub_tool_libcbase.h"
     14 #include "pub_tool_libcassert.h"
     15 #include "pub_tool_libcfile.h"
     16 #include "pub_tool_libcprint.h"
     17 #include "pub_tool_libcproc.h"
     18 #include "pub_tool_machine.h"
     19 #include "pub_tool_mallocfree.h"
     20 #include "pub_tool_options.h"
     21 #include "pub_tool_tooliface.h"
     22 #include "pub_tool_xarray.h"
     23 #include "pub_tool_clientstate.h"
     24 #include "pub_tool_machine.h"      // VG_(fnptr_to_fnentry)
     25 
     26 #include "events.h" // defines CLG_ macro
     27 #include "costs.h"
     28 
     29 
     30 /*------------------------------------------------------------*/
     31 /*--- Calltree compile options                            --- */
     32 /*------------------------------------------------------------*/
     33 
     34 /* Enable debug output */
     35 #define CLG_ENABLE_DEBUG 1
     36 
     37 /* Enable experimental features? */
     38 #define CLG_EXPERIMENTAL 0
     39 
     40 /* Syscall Timing in microseconds?
     41  * (define to 0 if you get compile errors) */
     42 #define CLG_MICROSYSTIME 0
     43 
     44 /* Set to 1 if you want full sanity checks for JCC */
     45 #define JCC_CHECK 0
     46 
     47 
     48 
     49 /*------------------------------------------------------------*/
     50 /*--- Command line options                                 ---*/
     51 /*------------------------------------------------------------*/
     52 
     53 #define DEFAULT_OUTFORMAT   "callgrind.out.%p"
     54 #define DEFAULT_COMMANDNAME "callgrind.cmd"
     55 #define DEFAULT_RESULTNAME  "callgrind.res"
     56 #define DEFAULT_INFONAME    "callgrind.info"
     57 
     58 typedef struct _CommandLineOptions CommandLineOptions;
     59 struct _CommandLineOptions {
     60 
     61   /* Dump format options */
     62   Char* out_format;      /* Format string for callgrind output file name */
     63   Bool combine_dumps;    /* Dump trace parts into same file? */
     64   Bool compress_strings;
     65   Bool compress_events;
     66   Bool compress_pos;
     67   Bool mangle_names;
     68   Bool compress_mangled;
     69   Bool dump_line;
     70   Bool dump_instr;
     71   Bool dump_bb;
     72   Bool dump_bbs;         /* Dump basic block information? */
     73 
     74   /* Dump generation options */
     75   ULong dump_every_bb;     /* Dump every xxx BBs. */
     76 
     77   /* Collection options */
     78   Bool separate_threads; /* Separate threads in dump? */
     79   Int  separate_callers; /* Separate dependent on how many callers? */
     80   Int  separate_recursions; /* Max level of recursions to separate */
     81   Bool skip_plt;         /* Skip functions in PLT section? */
     82   Bool skip_direct_recursion; /* Increment direct recursions the level? */
     83 
     84   Bool collect_atstart;  /* Start in collecting state ? */
     85   Bool collect_jumps;    /* Collect (cond.) jumps in functions ? */
     86 
     87   Bool collect_alloc;    /* Collect size of allocated memory */
     88   Bool collect_systime;  /* Collect time for system calls */
     89 
     90   Bool collect_bus;      /* Collect global bus events */
     91 
     92   /* Instrument options */
     93   Bool instrument_atstart;  /* Instrument at start? */
     94   Bool simulate_cache;      /* Call into cache simulator ? */
     95   Bool simulate_branch;     /* Call into branch prediction simulator ? */
     96 
     97   /* Call graph generation */
     98   Bool pop_on_jump;       /* Handle a jump between functions as ret+call */
     99 
    100 #if CLG_ENABLE_DEBUG
    101   Int   verbose;
    102   ULong verbose_start;
    103 #endif
    104 };
    105 
    106 /*------------------------------------------------------------*/
    107 /*--- Constants                                            ---*/
    108 /*------------------------------------------------------------*/
    109 
    110 
    111 /* According to IA-32 Intel Architecture Software Developer's Manual: Vol 2 */
    112 #define MAX_x86_INSTR_SIZE              16
    113 
    114 /* Minimum cache line size allowed */
    115 #define MIN_LINE_SIZE   16
    116 
    117 /* Size of various buffers used for storing strings */
    118 #define FILENAME_LEN                    VKI_PATH_MAX
    119 #define FN_NAME_LEN                    4096 /* for C++ code :-) */
    120 #define OBJ_NAME_LEN                    256
    121 #define COSTS_LEN                       512 /* at least 17x 64bit values */
    122 #define BUF_LEN                         512
    123 #define COMMIFY_BUF_LEN                 128
    124 #define RESULTS_BUF_LEN                 256
    125 #define LINE_BUF_LEN                     64
    126 
    127 
    128 /* Convenience macros */
    129 
    130 /* Use this only when size of sprintf args are known to fit into
    131  * given buffer; for strings of unknown length, use WRITE_STR below
    132  */
    133 #define WRITE_SPRINTF(fd, zz_buf, fmt, args...) \
    134    do { Int len = VG_(sprintf)(zz_buf, fmt, ## args); \
    135         VG_(write)(fd, (void*)zz_buf, len); \
    136    } while (0)
    137 
    138 #define WRITE_STR(fd, str) \
    139    do { if (str) { Int len = VG_(strlen)(str); \
    140         VG_(write)(fd, (void*)str, len); } \
    141         else VG_(write)(fd, "(null)", 6); \
    142    } while (0)
    143 
    144 #define WRITE_STR2(fd, str1, str2) \
    145    do { if (str1) { Int len = VG_(strlen)(str1); \
    146         VG_(write)(fd, (void*)str1, len); } \
    147         else VG_(write)(fd, "(null)", 6); \
    148 	if (str2) { Int len = VG_(strlen)(str2); \
    149         VG_(write)(fd, (void*)str2, len); } \
    150         else VG_(write)(fd, "(null)", 6); \
    151    } while (0)
    152 
    153 #define WRITE_STR3(fd, str1, str2, str3) \
    154    do { if (str1) { Int len = VG_(strlen)(str1); \
    155         VG_(write)(fd, (void*)str1, len); } \
    156         else VG_(write)(fd, "(null)", 6); \
    157         if (str2) { Int len = VG_(strlen)(str2); \
    158         VG_(write)(fd, (void*)str2, len); } \
    159         else VG_(write)(fd, "(null)", 6); \
    160         if (str3) { Int len = VG_(strlen)(str3); \
    161         VG_(write)(fd, (void*)str3, len); } \
    162         else VG_(write)(fd, "(null)", 6); \
    163    } while (0)
    164 
    165 
    166 /*------------------------------------------------------------*/
    167 /*--- Statistics                                           ---*/
    168 /*------------------------------------------------------------*/
    169 
    170 typedef struct _Statistics Statistics;
    171 struct _Statistics {
    172   ULong call_counter;
    173   ULong jcnd_counter;
    174   ULong jump_counter;
    175   ULong rec_call_counter;
    176   ULong ret_counter;
    177   ULong bb_executions;
    178 
    179   Int  context_counter;
    180   Int  bb_retranslations;
    181 
    182   Int  distinct_objs;
    183   Int  distinct_files;
    184   Int  distinct_fns;
    185   Int  distinct_contexts;
    186   Int  distinct_bbs;
    187   Int  distinct_jccs;
    188   Int  distinct_bbccs;
    189   Int  distinct_instrs;
    190   Int  distinct_skips;
    191 
    192   Int  bb_hash_resizes;
    193   Int  bbcc_hash_resizes;
    194   Int  jcc_hash_resizes;
    195   Int  cxt_hash_resizes;
    196   Int  fn_array_resizes;
    197   Int  call_stack_resizes;
    198   Int  fn_stack_resizes;
    199 
    200   Int  full_debug_BBs;
    201   Int  file_line_debug_BBs;
    202   Int  fn_name_debug_BBs;
    203   Int  no_debug_BBs;
    204   Int  bbcc_lru_misses;
    205   Int  jcc_lru_misses;
    206   Int  cxt_lru_misses;
    207   Int  bbcc_clones;
    208 };
    209 
    210 
    211 /*------------------------------------------------------------*/
    212 /*--- Structure declarations                               ---*/
    213 /*------------------------------------------------------------*/
    214 
    215 typedef struct _Context     Context;
    216 typedef struct _CC          CC;
    217 typedef struct _BB          BB;
    218 typedef struct _Skipped     Skipped;
    219 typedef struct _BBCC        BBCC;
    220 typedef struct _jCC         jCC;
    221 typedef struct _fCC         fCC;
    222 typedef struct _fn_node     fn_node;
    223 typedef struct _file_node   file_node;
    224 typedef struct _obj_node    obj_node;
    225 typedef struct _fn_config   fn_config;
    226 typedef struct _call_entry  call_entry;
    227 typedef struct _thread_info thread_info;
    228 
    229 /* Costs of event sets. Aliases to arrays of 64-bit values */
    230 typedef ULong* SimCost;  /* All events the simulator can produce */
    231 typedef ULong* UserCost;
    232 typedef ULong* FullCost; /* Simulator + User */
    233 
    234 
    235 /* JmpCall cost center
    236  * for subroutine call (from->bb->jmp_addr => to->bb->addr)
    237  *
    238  * Each BB has at most one CALL instruction. The list of JCC from
    239  * this call is a pointer to the list head (stored in BBCC), and
    240  * <next_from> in the JCC struct.
    241  *
    242  * For fast lookup, JCCs are reachable with a hash table, keyed by
    243  * the (from_bbcc,to) pair. <next_hash> is used for the JCC chain
    244  * of one hash table entry.
    245  *
    246  * Cost <sum> holds event counts for already returned executions.
    247  * <last> are the event counters at last enter of the subroutine.
    248  * <sum> is updated on returning from the subroutine by
    249  * adding the diff of <last> and current event counters to <sum>.
    250  *
    251  * After updating, <last> is set to current event counters. Thus,
    252  * events are not counted twice for recursive calls (TODO: True?)
    253  */
    254 #define JmpNone (Ijk_Boring+30)
    255 #define JmpCond (Ijk_Boring+31)
    256 
    257 struct _jCC {
    258   Int  jmpkind;     /* JmpCall, JmpBoring, JmpCond */
    259   jCC* next_hash;   /* for hash entry chain */
    260   jCC* next_from;   /* next JCC from a BBCC */
    261   BBCC *from, *to;  /* call arc from/to this BBCC */
    262   UInt jmp;         /* jump no. in source */
    263 
    264   ULong call_counter; /* no wraparound with 64 bit */
    265 
    266   FullCost cost; /* simulator + user counters */
    267 };
    268 
    269 
    270 /*
    271  * Info for one instruction of a basic block.
    272  */
    273 typedef struct _InstrInfo InstrInfo;
    274 struct _InstrInfo {
    275   UInt instr_offset;
    276   UInt instr_size;
    277   UInt cost_offset;
    278   EventSet* eventset;
    279 };
    280 
    281 
    282 /*
    283  * Info for a conditional jump in a basic block
    284  */
    285 typedef struct _CJmpInfo CJmpInfo;
    286 struct _CJmpInfo {
    287     UInt instr; /* instruction index in this basic block */
    288     Bool skip;   /* Cond.Jumps to next instruction should be ignored */
    289 };
    290 
    291 
    292 /**
    293  * An instrumented basic block (BB).
    294  *
    295  * BBs are put into a resizable hash to allow for fast detection if a
    296  * BB is to be retranslated but cost info is already available.
    297  * The key for a BB is a (object, offset) tupel making it independent
    298  * from possibly multiple mappings of the same ELF object.
    299  *
    300  * At the beginning of each instrumented BB,
    301  * a call to setup_bbcc(), specifying a pointer to the
    302  * according BB structure, is added.
    303  *
    304  * As cost of a BB has to be distinguished depending on the context,
    305  * multiple cost centers for one BB (struct BBCC) exist and the according
    306  * BBCC is set by setup_bbcc.
    307  */
    308 struct _BB {
    309   obj_node*  obj;         /* ELF object of BB */
    310   PtrdiffT   offset;      /* offset of BB in ELF object file */
    311   BB*        next;       /* chaining for a hash entry */
    312 
    313   VgSectKind sect_kind;  /* section of this BB, e.g. PLT */
    314   UInt       instr_count;
    315 
    316   /* filled by CLG_(get_fn_node) if debug info is available */
    317   fn_node*   fn;          /* debug info for this BB */
    318   UInt       line;
    319   Bool       is_entry;    /* True if this BB is a function entry */
    320 
    321   BBCC*      bbcc_list;  /* BBCCs for same BB (see next_bbcc in BBCC) */
    322   BBCC*      last_bbcc;  /* Temporary: Cached for faster access (LRU) */
    323 
    324   /* filled by CLG_(instrument) if not seen before */
    325   UInt       cjmp_count;  /* number of conditional exits */
    326   CJmpInfo*  jmp;         /* array of info for condition jumps,
    327 			   * allocated directly after this struct */
    328   Int        jmpkind;    /* remember jump kind of final exit */
    329   Bool       cjmp_inverted; /* condition of last cond.jump can be inverted by VEX */
    330 
    331   UInt       instr_len;
    332   UInt       cost_count;
    333   InstrInfo  instr[0];   /* info on instruction sizes and costs */
    334 };
    335 
    336 
    337 
    338 /**
    339  * Function context
    340  *
    341  * Basic blocks are always executed in the scope of a context.
    342  * A function context is a list of function nodes representing
    343  * the call chain to the current context: I.e. fn[0] is the
    344  * function we are currently in, fn[1] has called fn[0], and so on.
    345  * Recursion levels are used for fn[0].
    346  *
    347  * To get a unique number for a full execution context, use
    348  *  rec_index = min(<fn->rec_separation>,<active>) - 1;
    349  *  unique_no = <number> + rec_index
    350  *
    351  * For each Context, recursion index and BB, there can be a BBCC.
    352  */
    353 struct _Context {
    354     UInt size;        // number of function dependencies
    355     UInt base_number; // for context compression & dump array
    356     Context* next;    // entry chaining for hash
    357     UWord hash;       // for faster lookup...
    358     fn_node* fn[0];
    359 };
    360 
    361 
    362 /*
    363  * Info for a conditional jump in a basic block
    364  */
    365 typedef struct _JmpData JmpData;
    366 struct _JmpData {
    367     ULong ecounter; /* number of times the BB was left at this exit */
    368     jCC*  jcc_list;  /* JCCs for Cond.Jumps from this exit */
    369 };
    370 
    371 
    372 /*
    373  * Basic Block Cost Center
    374  *
    375  * On demand, multiple BBCCs will be created for the same BB
    376  * dependend on command line options and:
    377  * - current function (it's possible that a BB is executed in the
    378  *   context of different functions, e.g. in manual assembler/PLT)
    379  * - current thread ID
    380  * - position where current function is called from
    381  * - recursion level of current function
    382  *
    383  * The cost centres for the instructions of a basic block are
    384  * stored in a contiguous array.
    385  * They are distinguishable by their tag field.
    386  */
    387 struct _BBCC {
    388     BB*      bb;           /* BB for this cost center */
    389 
    390     Context* cxt;          /* execution context of this BBCC */
    391     ThreadId tid;          /* only for assertion check purpose */
    392     UInt     rec_index;    /* Recursion index in rec->bbcc for this bbcc */
    393     BBCC**   rec_array;    /* Variable sized array of pointers to
    394 			    * recursion BBCCs. Shared. */
    395     ULong    ret_counter;  /* how often returned from jccs of this bbcc;
    396 			    * used to check if a dump for this BBCC is needed */
    397 
    398     BBCC*    next_bbcc;    /* Chain of BBCCs for same BB */
    399     BBCC*    lru_next_bbcc; /* BBCC executed next the last time */
    400 
    401     jCC*     lru_from_jcc; /* Temporary: Cached for faster access (LRU) */
    402     jCC*     lru_to_jcc;   /* Temporary: Cached for faster access (LRU) */
    403     FullCost skipped;      /* cost for skipped functions called from
    404 			    * jmp_addr. Allocated lazy */
    405 
    406     BBCC*    next;         /* entry chain in hash */
    407     ULong*   cost;         /* start of 64bit costs for this BBCC */
    408     ULong    ecounter_sum; /* execution counter for first instruction of BB */
    409     JmpData  jmp[0];
    410 };
    411 
    412 
    413 /* the <number> of fn_node, file_node and obj_node are for compressed dumping
    414  * and a index into the dump boolean table and fn_info_table
    415  */
    416 
    417 struct _fn_node {
    418   Char*      name;
    419   UInt       number;
    420   Context*   last_cxt; /* LRU info */
    421   Context*   pure_cxt; /* the context with only the function itself */
    422   file_node* file;     /* reverse mapping for 2nd hash */
    423   fn_node* next;
    424 
    425   Bool dump_before :1;
    426   Bool dump_after :1;
    427   Bool zero_before :1;
    428   Bool toggle_collect :1;
    429   Bool skip :1;
    430   Bool pop_on_jump : 1;
    431 
    432   Bool is_malloc :1;
    433   Bool is_realloc :1;
    434   Bool is_free :1;
    435 
    436   Int  group;
    437   Int  separate_callers;
    438   Int  separate_recursions;
    439 #if CLG_ENABLE_DEBUG
    440   Int  verbosity; /* Stores old verbosity level while in function */
    441 #endif
    442 };
    443 
    444 /* Quite arbitrary fixed hash sizes */
    445 
    446 #define   N_OBJ_ENTRIES         47
    447 #define  N_FILE_ENTRIES         53
    448 #define    N_FN_ENTRIES         87
    449 #define N_BBCC2_ENTRIES         37
    450 
    451 struct _file_node {
    452    Char*      name;
    453    fn_node*   fns[N_FN_ENTRIES];
    454    UInt       number;
    455    obj_node*  obj;
    456    file_node* next;
    457 };
    458 
    459 /* If an object is dlopened multiple times, we hope that <name> is unique;
    460  * <start> and <offset> can change with each dlopen, and <start> is
    461  * zero when object is unmapped (possible at dump time).
    462  */
    463 struct _obj_node {
    464    Char*      name;
    465    UInt       last_slash_pos;
    466 
    467    Addr       start;  /* Start address of text segment mapping */
    468    SizeT      size;   /* Length of mapping */
    469    PtrdiffT   offset; /* Offset between symbol address and file offset */
    470 
    471    file_node* files[N_FILE_ENTRIES];
    472    UInt       number;
    473    obj_node*  next;
    474 };
    475 
    476 /* an entry in the callstack
    477  *
    478  * <nonskipped> is 0 if the function called is not skipped (usual case).
    479  * Otherwise, it is the last non-skipped BBCC. This one gets all
    480  * the calls to non-skipped functions and all costs in skipped
    481  * instructions.
    482  */
    483 struct _call_entry {
    484     jCC* jcc;           /* jCC for this call */
    485     FullCost enter_cost; /* cost event counters at entering frame */
    486     Addr sp;            /* stack pointer directly after call */
    487     Addr ret_addr;      /* address to which to return to
    488 			 * is 0 on a simulated call */
    489     BBCC* nonskipped;   /* see above */
    490     Context* cxt;       /* context before call */
    491     Int fn_sp;          /* function stack index before call */
    492 };
    493 
    494 
    495 /*
    496  * Execution state of main thread or a running signal handler in
    497  * a thread while interrupted by another signal handler.
    498  * As there's no scheduling among running signal handlers of one thread,
    499  * we only need a subset of a full thread state:
    500  * - event counter
    501  * - collect state
    502  * - last BB, last jump kind, last nonskipped BB
    503  * - callstack pointer for sanity checking and correct unwinding
    504  *   after exit
    505  */
    506 typedef struct _exec_state exec_state;
    507 struct _exec_state {
    508 
    509   /* the signum of the handler, 0 for main thread context
    510    */
    511   Int sig;
    512 
    513   /* the old call stack pointer at entering the signal handler */
    514   Int orig_sp;
    515 
    516   FullCost cost;
    517   Bool     collect;
    518   Context* cxt;
    519 
    520   Int   jmps_passed; /* number of conditional jumps passed in last BB */
    521   BBCC* bbcc;      /* last BB executed */
    522   BBCC* nonskipped;
    523 
    524   Int call_stack_bottom; /* Index into fn_stack */
    525 };
    526 
    527 /* Global state structures */
    528 typedef struct _bb_hash bb_hash;
    529 struct _bb_hash {
    530   UInt size, entries;
    531   BB** table;
    532 };
    533 
    534 typedef struct _cxt_hash cxt_hash;
    535 struct _cxt_hash {
    536   UInt size, entries;
    537   Context** table;
    538 };
    539 
    540 /* Thread specific state structures, i.e. parts of a thread state.
    541  * There are variables for the current state of each part,
    542  * on which a thread state is copied at thread switch.
    543  */
    544 typedef struct _bbcc_hash bbcc_hash;
    545 struct _bbcc_hash {
    546   UInt size, entries;
    547   BBCC** table;
    548 };
    549 
    550 typedef struct _jcc_hash jcc_hash;
    551 struct _jcc_hash {
    552   UInt size, entries;
    553   jCC** table;
    554   jCC* spontaneous;
    555 };
    556 
    557 typedef struct _fn_array fn_array;
    558 struct _fn_array {
    559   UInt size;
    560   UInt* array;
    561 };
    562 
    563 typedef struct _call_stack call_stack;
    564 struct _call_stack {
    565   UInt size;
    566   Int sp;
    567   call_entry* entry;
    568 };
    569 
    570 typedef struct _fn_stack fn_stack;
    571 struct _fn_stack {
    572   UInt size;
    573   fn_node **bottom, **top;
    574 };
    575 
    576 /* The maximum number of simultaneous running signal handlers per thread.
    577  * This is the number of execution states storable in a thread.
    578  */
    579 #define MAX_SIGHANDLERS 10
    580 
    581 typedef struct _exec_stack exec_stack;
    582 struct _exec_stack {
    583   Int sp; /* > 0 if a handler is running */
    584   exec_state* entry[MAX_SIGHANDLERS];
    585 };
    586 
    587 /* Thread State
    588  *
    589  * This structure stores thread specific info while a thread is *not*
    590  * running. See function switch_thread() for save/restore on thread switch.
    591  *
    592  * If --separate-threads=no, BBCCs and JCCs can be shared by all threads, i.e.
    593  * only structures of thread 1 are used.
    594  * This involves variables fn_info_table, bbcc_table and jcc_table.
    595  */
    596 struct _thread_info {
    597 
    598   /* state */
    599   fn_stack fns;       /* function stack */
    600   call_stack calls;   /* context call arc stack */
    601   exec_stack states;  /* execution states interrupted by signals */
    602 
    603   /* dump statistics */
    604   FullCost lastdump_cost;    /* Cost at last dump */
    605   FullCost sighandler_cost;
    606 
    607   /* thread specific data structure containers */
    608   fn_array fn_active;
    609   jcc_hash jccs;
    610   bbcc_hash bbccs;
    611 };
    612 
    613 /* Structs used for dumping */
    614 
    615 /* Address position inside of a BBCC:
    616  * This includes
    617  * - the address offset from the BB start address
    618  * - file/line from debug info for that address (can change inside a BB)
    619  */
    620 typedef struct _AddrPos AddrPos;
    621 struct _AddrPos {
    622     Addr addr;
    623     Addr bb_addr;
    624     file_node* file;
    625     UInt line;
    626 };
    627 
    628 /* a simulator cost entity that can be written out in one line */
    629 typedef struct _AddrCost AddrCost;
    630 struct _AddrCost {
    631     AddrPos p;
    632     SimCost cost;
    633 };
    634 
    635 /* A function in an execution context */
    636 typedef struct _FnPos FnPos;
    637 struct _FnPos {
    638     file_node* file;
    639     fn_node* fn;
    640     obj_node* obj;
    641     Context* cxt;
    642     int rec_index;
    643     UInt line;
    644 };
    645 
    646 /*------------------------------------------------------------*/
    647 /*--- Cache simulator interface                            ---*/
    648 /*------------------------------------------------------------*/
    649 
    650 struct cachesim_if
    651 {
    652     void (*print_opts)(void);
    653     Bool (*parse_opt)(Char* arg);
    654     void (*post_clo_init)(void);
    655     void (*clear)(void);
    656     void (*getdesc)(Char* buf);
    657     void (*printstat)(Int,Int,Int);
    658     void (*add_icost)(SimCost, BBCC*, InstrInfo*, ULong);
    659     void (*finish)(void);
    660 
    661     void (*log_1I0D)(InstrInfo*) VG_REGPARM(1);
    662     void (*log_2I0D)(InstrInfo*, InstrInfo*) VG_REGPARM(2);
    663     void (*log_3I0D)(InstrInfo*, InstrInfo*, InstrInfo*) VG_REGPARM(3);
    664 
    665     void (*log_1I1Dr)(InstrInfo*, Addr, Word) VG_REGPARM(3);
    666     void (*log_1I1Dw)(InstrInfo*, Addr, Word) VG_REGPARM(3);
    667 
    668     void (*log_0I1Dr)(InstrInfo*, Addr, Word) VG_REGPARM(3);
    669     void (*log_0I1Dw)(InstrInfo*, Addr, Word) VG_REGPARM(3);
    670 
    671     // function names of helpers (for debugging generated code)
    672     Char *log_1I0D_name, *log_2I0D_name, *log_3I0D_name;
    673     Char *log_1I1Dr_name, *log_1I1Dw_name;
    674     Char *log_0I1Dr_name, *log_0I1Dw_name;
    675 };
    676 
    677 // set by setup_bbcc at start of every BB, and needed by log_* helpers
    678 extern Addr   CLG_(bb_base);
    679 extern ULong* CLG_(cost_base);
    680 
    681 // Event groups
    682 #define EG_USE   0
    683 #define EG_IR    1
    684 #define EG_DR    2
    685 #define EG_DW    3
    686 #define EG_BC    4
    687 #define EG_BI    5
    688 #define EG_BUS   6
    689 #define EG_ALLOC 7
    690 #define EG_SYS   8
    691 
    692 struct event_sets {
    693     EventSet *base, *full;
    694 };
    695 extern struct event_sets CLG_(sets);
    696 
    697 #define fullOffset(group) (CLG_(sets).full->offset[group])
    698 
    699 
    700 /*------------------------------------------------------------*/
    701 /*--- Functions                                            ---*/
    702 /*------------------------------------------------------------*/
    703 
    704 /* from clo.c */
    705 
    706 void CLG_(set_clo_defaults)(void);
    707 void CLG_(update_fn_config)(fn_node*);
    708 Bool CLG_(process_cmd_line_option)(Char*);
    709 void CLG_(print_usage)(void);
    710 void CLG_(print_debug_usage)(void);
    711 
    712 /* from sim.c */
    713 extern struct cachesim_if CLG_(cachesim);
    714 void CLG_(init_eventsets)(void);
    715 
    716 /* from main.c */
    717 Bool CLG_(get_debug_info)(Addr, Char filename[FILENAME_LEN],
    718 			 Char fn_name[FN_NAME_LEN], UInt*, DebugInfo**);
    719 void CLG_(collectBlockInfo)(IRSB* bbIn, UInt*, UInt*, Bool*);
    720 void CLG_(set_instrument_state)(Char*,Bool);
    721 void CLG_(dump_profile)(Char* trigger,Bool only_current_thread);
    722 void CLG_(zero_all_cost)(Bool only_current_thread);
    723 Int CLG_(get_dump_counter)(void);
    724 void CLG_(fini)(Int exitcode);
    725 
    726 /* from command.c */
    727 void CLG_(init_command)(void);
    728 void CLG_(check_command)(void);
    729 void CLG_(finish_command)(void);
    730 
    731 /* from bb.c */
    732 void CLG_(init_bb_hash)(void);
    733 bb_hash* CLG_(get_bb_hash)(void);
    734 BB*  CLG_(get_bb)(Addr addr, IRSB* bb_in, Bool *seen_before);
    735 void CLG_(delete_bb)(Addr addr);
    736 
    737 static __inline__ Addr bb_addr(BB* bb)
    738  { return bb->offset + bb->obj->offset; }
    739 static __inline__ Addr bb_jmpaddr(BB* bb)
    740  { UInt off = (bb->instr_count > 0) ? bb->instr[bb->instr_count-1].instr_offset : 0;
    741    return off + bb->offset + bb->obj->offset; }
    742 
    743 /* from fn.c */
    744 void CLG_(init_fn_array)(fn_array*);
    745 void CLG_(copy_current_fn_array)(fn_array* dst);
    746 fn_array* CLG_(get_current_fn_array)(void);
    747 void CLG_(set_current_fn_array)(fn_array*);
    748 UInt* CLG_(get_fn_entry)(Int n);
    749 
    750 void      CLG_(init_obj_table)(void);
    751 obj_node* CLG_(get_obj_node)(DebugInfo* si);
    752 file_node* CLG_(get_file_node)(obj_node*, Char* filename);
    753 fn_node*  CLG_(get_fn_node)(BB* bb);
    754 
    755 /* from bbcc.c */
    756 void CLG_(init_bbcc_hash)(bbcc_hash* bbccs);
    757 void CLG_(copy_current_bbcc_hash)(bbcc_hash* dst);
    758 bbcc_hash* CLG_(get_current_bbcc_hash)(void);
    759 void CLG_(set_current_bbcc_hash)(bbcc_hash*);
    760 void CLG_(forall_bbccs)(void (*func)(BBCC*));
    761 void CLG_(zero_bbcc)(BBCC* bbcc);
    762 BBCC* CLG_(get_bbcc)(BB* bb);
    763 BBCC* CLG_(clone_bbcc)(BBCC* orig, Context* cxt, Int rec_index);
    764 void CLG_(setup_bbcc)(BB* bb) VG_REGPARM(1);
    765 
    766 
    767 /* from jumps.c */
    768 void CLG_(init_jcc_hash)(jcc_hash*);
    769 void CLG_(copy_current_jcc_hash)(jcc_hash* dst);
    770 jcc_hash* CLG_(get_current_jcc_hash)(void);
    771 void CLG_(set_current_jcc_hash)(jcc_hash*);
    772 jCC* CLG_(get_jcc)(BBCC* from, UInt, BBCC* to);
    773 
    774 /* from callstack.c */
    775 void CLG_(init_call_stack)(call_stack*);
    776 void CLG_(copy_current_call_stack)(call_stack* dst);
    777 void CLG_(set_current_call_stack)(call_stack*);
    778 call_entry* CLG_(get_call_entry)(Int n);
    779 
    780 void CLG_(push_call_stack)(BBCC* from, UInt jmp, BBCC* to, Addr sp, Bool skip);
    781 void CLG_(pop_call_stack)(void);
    782 Int CLG_(unwind_call_stack)(Addr sp, Int);
    783 
    784 /* from context.c */
    785 void CLG_(init_fn_stack)(fn_stack*);
    786 void CLG_(copy_current_fn_stack)(fn_stack*);
    787 fn_stack* CLG_(get_current_fn_stack)(void);
    788 void CLG_(set_current_fn_stack)(fn_stack*);
    789 
    790 void CLG_(init_cxt_table)(void);
    791 cxt_hash* CLG_(get_cxt_hash)(void);
    792 Context* CLG_(get_cxt)(fn_node** fn);
    793 void CLG_(push_cxt)(fn_node* fn);
    794 
    795 /* from threads.c */
    796 void CLG_(init_threads)(void);
    797 thread_info** CLG_(get_threads)(void);
    798 thread_info* CLG_(get_current_thread)(void);
    799 void CLG_(switch_thread)(ThreadId tid);
    800 void CLG_(forall_threads)(void (*func)(thread_info*));
    801 void CLG_(run_thread)(ThreadId tid);
    802 
    803 void CLG_(init_exec_state)(exec_state* es);
    804 void CLG_(init_exec_stack)(exec_stack*);
    805 void CLG_(copy_current_exec_stack)(exec_stack*);
    806 void CLG_(set_current_exec_stack)(exec_stack*);
    807 void CLG_(pre_signal)(ThreadId tid, Int sigNum, Bool alt_stack);
    808 void CLG_(post_signal)(ThreadId tid, Int sigNum);
    809 void CLG_(run_post_signal_on_call_stack_bottom)(void);
    810 
    811 /* from dump.c */
    812 extern FullCost CLG_(total_cost);
    813 void CLG_(init_dumps)(void);
    814 Char* CLG_(get_out_file)(void);
    815 Char* CLG_(get_out_directory)(void);
    816 
    817 /*------------------------------------------------------------*/
    818 /*--- Exported global variables                            ---*/
    819 /*------------------------------------------------------------*/
    820 
    821 extern CommandLineOptions CLG_(clo);
    822 extern Statistics CLG_(stat);
    823 extern EventMapping* CLG_(dumpmap);
    824 
    825 /* Function active counter array, indexed by function number */
    826 extern UInt* CLG_(fn_active_array);
    827 extern Bool CLG_(instrument_state);
    828 
    829 extern call_stack CLG_(current_call_stack);
    830 extern fn_stack   CLG_(current_fn_stack);
    831 extern exec_state CLG_(current_state);
    832 extern ThreadId   CLG_(current_tid);
    833 
    834 
    835 /*------------------------------------------------------------*/
    836 /*--- Debug output                                         ---*/
    837 /*------------------------------------------------------------*/
    838 
    839 #if CLG_ENABLE_DEBUG
    840 
    841 #define CLG_DEBUGIF(x) \
    842   if ( (CLG_(clo).verbose >x) && \
    843        (CLG_(stat).bb_executions >= CLG_(clo).verbose_start))
    844 
    845 #define CLG_DEBUG(x,format,args...)   \
    846     CLG_DEBUGIF(x) {                  \
    847       CLG_(print_bbno)();	      \
    848       VG_(printf)(format,##args);     \
    849     }
    850 
    851 #define CLG_ASSERT(cond)              \
    852     if (!(cond)) {                    \
    853       CLG_(print_context)();          \
    854       CLG_(print_bbno)();	      \
    855       tl_assert(cond);                \
    856      }
    857 
    858 #else
    859 #define CLG_DEBUGIF(x) if (0)
    860 #define CLG_DEBUG(x...) {}
    861 #define CLG_ASSERT(cond) tl_assert(cond);
    862 #endif
    863 
    864 /* from debug.c */
    865 void CLG_(print_bbno)(void);
    866 void CLG_(print_context)(void);
    867 void CLG_(print_jcc)(int s, jCC* jcc);
    868 void CLG_(print_bbcc)(int s, BBCC* bbcc);
    869 void CLG_(print_bbcc_fn)(BBCC* bbcc);
    870 void CLG_(print_execstate)(int s, exec_state* es);
    871 void CLG_(print_eventset)(int s, EventSet* es);
    872 void CLG_(print_cost)(int s, EventSet*, ULong* cost);
    873 void CLG_(print_bb)(int s, BB* bb);
    874 void CLG_(print_bbcc_cost)(int s, BBCC*);
    875 void CLG_(print_cxt)(int s, Context* cxt, int rec_index);
    876 void CLG_(print_short_jcc)(jCC* jcc);
    877 void CLG_(print_stackentry)(int s, int sp);
    878 void CLG_(print_addr)(Addr addr);
    879 void CLG_(print_addr_ln)(Addr addr);
    880 
    881 void* CLG_(malloc)(HChar* cc, UWord s, char* f);
    882 void* CLG_(free)(void* p, char* f);
    883 #if 0
    884 #define CLG_MALLOC(_cc,x) CLG_(malloc)((_cc),x,__FUNCTION__)
    885 #define CLG_FREE(p)       CLG_(free)(p,__FUNCTION__)
    886 #else
    887 #define CLG_MALLOC(_cc,x) VG_(malloc)((_cc),x)
    888 #define CLG_FREE(p)       VG_(free)(p)
    889 #endif
    890 
    891 #endif /* CLG_GLOBAL */
    892