extras.c File Reference

Include dependency graph for extras.c:

Go to the source code of this file.

Functions
inline_static unsigned short	random_ushort (void)
inline_static int	profil_increment (long long value, int flags, long long excess, long long threshold)
static void	posix_profil (vptr_t address, PAPI_sprofil_t *prof, int flags, long long excess, long long threshold)
void	_papi_hwi_dispatch_profile (EventSetInfo_t *ESI, vptr_t pc, long long over, int profile_index)
int	_papi_hwi_dispatch_overflow_signal (void *papiContext, vptr_t address, int *isHardware, long long overflow_bit, int genOverflowBit, ThreadInfo_t **t, int cidx)
int	_papi_hwi_start_timer (int timer, int signal, int ns)
int	_papi_hwi_start_signal (int signal, int need_context, int cidx)
int	_papi_hwi_stop_signal (int signal)
int	_papi_hwi_stop_timer (int timer, int signal)

Variables
static unsigned int	_rnum = DEADBEEF
int	_papi_hwi_using_signal [PAPI_NSIG]

Function Documentation

_papi_hwi_dispatch_overflow_signal()

int _papi_hwi_dispatch_overflow_signal	(	void *	papiContext,
		vptr_t	address,
		int *	isHardware,
		long long	overflow_bit,
		int	genOverflowBit,
		ThreadInfo_t **	t,
		int	cidx
	)

Definition at line 216 of file extras.c.

{
    int retval, event_counter, i, overflow_flag, pos;
    int papi_index, j;
    int profile_index = 0;
    long long overflow_vector;
 
    long long temp[_papi_hwd[cidx]->cmp_info.num_cntrs], over;
    long long latest = 0;
    ThreadInfo_t *thread;
    EventSetInfo_t *ESI;
    _papi_hwi_context_t *ctx = ( _papi_hwi_context_t * ) papiContext;
 
    OVFDBG( "enter\n" );
 
    if ( *t )
        thread = *t;
    else
        *t = thread = _papi_hwi_lookup_thread( 0 );
 
    if ( thread != NULL ) {
        ESI = thread->running_eventset[cidx];
 
        if ( ( ESI == NULL ) || ( ( ESI->state & PAPI_OVERFLOWING ) == 0 ) ) {
            OVFDBG( "Either no eventset or eventset not set to overflow.\n" );
#ifdef ANY_THREAD_GETS_SIGNAL
            _papi_hwi_broadcast_signal( thread->tid );
#endif
            return ( PAPI_OK );
        }
 
        if ( ESI->CmpIdx != cidx )
            return ( PAPI_ENOCMP );
 
        if ( ESI->master != thread ) {
            PAPIERROR
                ( "eventset->thread %p vs. current thread %p mismatch",
                  ESI->master, thread );
            return ( PAPI_EBUG );
        }
 
        if ( isHardware ) {
            if ( ESI->overflow.flags & PAPI_OVERFLOW_HARDWARE ) {
                ESI->state |= PAPI_PAUSED;
                *isHardware = 1;
            } else
                *isHardware = 0;
        }
        /* Get the latest counter value */
        event_counter = ESI->overflow.event_counter;
 
        overflow_flag = 0;
        overflow_vector = 0;
 
        if ( !( ESI->overflow.flags & PAPI_OVERFLOW_HARDWARE ) ) {
            retval = _papi_hwi_read( thread->context[cidx], ESI, ESI->sw_stop );
            if ( retval < PAPI_OK )
                return ( retval );
            for ( i = 0; i < event_counter; i++ ) {
                papi_index = ESI->overflow.EventIndex[i];
                latest = ESI->sw_stop[papi_index];
                temp[i] = -1;
 
                if ( latest >= ( long long ) ESI->overflow.deadline[i] ) {
                    OVFDBG
                        ( "dispatch_overflow() latest %lld, deadline %lld, threshold %d\n",
                          latest, ESI->overflow.deadline[i],
                          ESI->overflow.threshold[i] );
                    pos = ESI->EventInfoArray[papi_index].pos[0];
                    overflow_vector ^= ( long long ) 1 << pos;
                    temp[i] = latest - ESI->overflow.deadline[i];
                    overflow_flag = 1;
                    /* adjust the deadline */
                    ESI->overflow.deadline[i] =
                        latest + ESI->overflow.threshold[i];
                }
            }
        } else if ( genOverflowBit ) {
            /* we had assumed the overflow event can't be derived event */
            papi_index = ESI->overflow.EventIndex[0];
 
            /* suppose the pos is the same as the counter number
             * (this is not true in Itanium, but itanium doesn't 
             * need us to generate the overflow bit
             */
            pos = ESI->EventInfoArray[papi_index].pos[0];
            overflow_vector = ( long long ) 1 << pos;
        } else
            overflow_vector = overflow_bit;
 
        if ( ( ESI->overflow.flags & PAPI_OVERFLOW_HARDWARE ) || overflow_flag ) {
            if ( ESI->state & PAPI_PROFILING ) {
                int k = 0;
                while ( overflow_vector ) {
                    i = ffsll( overflow_vector ) - 1;
                    for ( j = 0; j < event_counter; j++ ) {
                        papi_index = ESI->overflow.EventIndex[j];
                        /* This loop is here ONLY because Pentium 4 can have tagged *
                         * events that contain more than one counter without being  *
                         * derived. You've gotta scan all terms to make sure you    *
                         * find the one to profile. */
                        for ( k = 0, pos = 0; k < PAPI_EVENTS_IN_DERIVED_EVENT && pos >= 0;
                              k++ ) {
                            pos = ESI->EventInfoArray[papi_index].pos[k];
                            if ( i == pos ) {
                                profile_index = j;
                                goto foundit;
                            }
                        }
                    }
                    if ( j == event_counter ) {
                        PAPIERROR
                            ( "BUG! overflow_vector is 0, dropping interrupt" );
                        return ( PAPI_EBUG );
                    }
 
                  foundit:
                    if ( ( ESI->overflow.flags & PAPI_OVERFLOW_HARDWARE ) )
                        over = 0;
                    else
                        over = temp[profile_index];
                    _papi_hwi_dispatch_profile( ESI, address, over,
                                                profile_index );
                    overflow_vector ^= ( long long ) 1 << i;
                }
                /* do not use overflow_vector after this place */
            } else {
                ESI->overflow.handler( ESI->EventSetIndex, ( void * ) address,
                                       overflow_vector, ctx->ucontext );
            }
        }
        ESI->state &= ~( PAPI_PAUSED );
    }
#ifdef ANY_THREAD_GETS_SIGNAL
    else {
        OVFDBG( "I haven't been noticed by PAPI before\n" );
        _papi_hwi_broadcast_signal( ( *_papi_hwi_thread_id_fn ) (  ) );
    }
#endif
    return ( PAPI_OK );
}

Here is the call graph for this function:

Here is the caller graph for this function:

_papi_hwi_dispatch_profile()

void _papi_hwi_dispatch_profile	(	EventSetInfo_t *	ESI,
		vptr_t	pc,
		long long	over,
		int	profile_index
	)

Definition at line 165 of file extras.c.

{
    EventSetProfileInfo_t *profile = &ESI->profile;
    PAPI_sprofil_t *sprof;
    vptr_t offset = 0;
    vptr_t best_offset = 0;
    int count;
    int best_index = -1;
    int i;
 
    PRFDBG( "handled IP %p\n", pc );
 
    sprof = profile->prof[profile_index];
    count = profile->count[profile_index];
 
    for ( i = 0; i < count; i++ ) {
        offset = sprof[i].pr_off;
        if ( ( offset < pc ) && ( offset > best_offset ) ) {
            best_index = i;
            best_offset = offset;
        }
    }
 
    if ( best_index == -1 )
        best_index = 0;
 
    posix_profil( pc, &sprof[best_index], profile->flags, over,
                  profile->threshold[profile_index] );
}

Here is the call graph for this function:

Here is the caller graph for this function:

_papi_hwi_start_signal()

int _papi_hwi_start_signal	(	int	signal,
		int	need_context,
		int	cidx
	)

Definition at line 403 of file extras.c.

{
    struct sigaction action;
 
    _papi_hwi_lock( INTERNAL_LOCK );
    _papi_hwi_using_signal[signal]++;
    if ( _papi_hwi_using_signal[signal] - 1 ) {
        INTDBG( "_papi_hwi_using_signal is now %d\n",
                _papi_hwi_using_signal[signal] );
        _papi_hwi_unlock( INTERNAL_LOCK );
        return ( PAPI_OK );
    }
 
    memset( &action, 0x00, sizeof ( struct sigaction ) );
    action.sa_flags = SA_RESTART;
    action.sa_sigaction =
        ( void ( * )( int, siginfo_t *, void * ) ) _papi_hwd[cidx]->
        dispatch_timer;
    if ( need_context )
#if (defined(_BGL) /*|| defined (__bgp__)*/)
        action.sa_flags |= SIGPWR;
#else
        action.sa_flags |= SA_SIGINFO;
#endif
 
    INTDBG( "installing signal handler\n" );
    if ( sigaction( signal, &action, NULL ) < 0 ) {
        PAPIERROR( "sigaction errno %d", errno );
        _papi_hwi_unlock( INTERNAL_LOCK );
        return ( PAPI_ESYS );
    }
 
    INTDBG( "_papi_hwi_using_signal[%d] is now %d.\n", signal,
            _papi_hwi_using_signal[signal] );
    _papi_hwi_unlock( INTERNAL_LOCK );
 
    return ( PAPI_OK );
}

Here is the call graph for this function:

Here is the caller graph for this function:

_papi_hwi_start_timer()

int _papi_hwi_start_timer	(	int	timer,
		int	signal,
		int	ns
	)

Definition at line 368 of file extras.c.

{
    struct itimerval value;
    int us = ns / 1000;
 
    if ( us == 0 )
        us = 1;
 
#ifdef ANY_THREAD_GETS_SIGNAL
    _papi_hwi_lock( INTERNAL_LOCK );
    if ( ( _papi_hwi_using_signal[signal] - 1 ) ) {
        INTDBG( "itimer already installed\n" );
        _papi_hwi_unlock( INTERNAL_LOCK );
        return ( PAPI_OK );
    }
    _papi_hwi_unlock( INTERNAL_LOCK );
#else
    ( void ) signal;         /*unused */
#endif
 
    value.it_interval.tv_sec = 0;
    value.it_interval.tv_usec = us;
    value.it_value.tv_sec = 0;
    value.it_value.tv_usec = us;
 
    INTDBG( "Installing itimer %d, with %d us interval\n", timer, us );
    if ( setitimer( timer, &value, NULL ) < 0 ) {
        PAPIERROR( "setitimer errno %d", errno );
        return ( PAPI_ESYS );
    }
 
    return ( PAPI_OK );
}

Here is the call graph for this function:

Here is the caller graph for this function:

_papi_hwi_stop_signal()

int _papi_hwi_stop_signal

(

int

signal

)

Definition at line 443 of file extras.c.

{
    _papi_hwi_lock( INTERNAL_LOCK );
    if ( --_papi_hwi_using_signal[signal] == 0 ) {
        INTDBG( "removing signal handler\n" );
        if ( sigaction( signal, NULL, NULL ) == -1 ) {
            PAPIERROR( "sigaction errno %d", errno );
            _papi_hwi_unlock( INTERNAL_LOCK );
            return ( PAPI_ESYS );
        }
    }
 
    INTDBG( "_papi_hwi_using_signal[%d] is now %d\n", signal,
            _papi_hwi_using_signal[signal] );
    _papi_hwi_unlock( INTERNAL_LOCK );
 
    return ( PAPI_OK );
}

Here is the call graph for this function:

Here is the caller graph for this function:

_papi_hwi_stop_timer()

int _papi_hwi_stop_timer	(	int	timer,
		int	signal
	)

Definition at line 463 of file extras.c.

{
    struct itimerval value;
 
#ifdef ANY_THREAD_GETS_SIGNAL
    _papi_hwi_lock( INTERNAL_LOCK );
    if ( _papi_hwi_using_signal[signal] > 1 ) {
        INTDBG( "itimer in use by another thread\n" );
        _papi_hwi_unlock( INTERNAL_LOCK );
        return ( PAPI_OK );
    }
    _papi_hwi_unlock( INTERNAL_LOCK );
#else
    ( void ) signal;         /*unused */
#endif
 
    value.it_interval.tv_sec = 0;
    value.it_interval.tv_usec = 0;
    value.it_value.tv_sec = 0;
    value.it_value.tv_usec = 0;
 
    INTDBG( "turning off timer\n" );
    if ( setitimer( timer, &value, NULL ) == -1 ) {
        PAPIERROR( "setitimer errno %d", errno );
        return PAPI_ESYS;
    }
 
    return PAPI_OK;
}

Here is the call graph for this function:

Here is the caller graph for this function:

posix_profil()

static void posix_profil ( vptr_t  address, PAPI_sprofil_t *  prof, int  flags, long long  excess, long long  threshold  )

static

Definition at line 94 of file extras.c.

{
    unsigned short *buf16;
    unsigned int *buf32;
    unsigned long long *buf64;
    unsigned long indx;
    unsigned long long lloffset;
 
    /* SPECIAL CASE: if starting address is 0 and scale factor is 2
       then all counts go into first bin.
     */
    if ( ( prof->pr_off == 0 ) && ( prof->pr_scale == 0x2 ) )
        indx = 0;
    else {
        /* compute the profile buffer offset by:
           - subtracting the profiling base address from the pc address
           - multiplying by the scaling factor
           - dividing by max scale (65536, or 2^^16) 
           - dividing by implicit 2 (2^^1 for a total of 2^^17), for even addresses
           NOTE: 131072 is a valid scale value. It produces byte resolution of addresses
         */
        lloffset =
            ( unsigned long long ) ( ( address - prof->pr_off ) *
                                     prof->pr_scale );
        indx = ( unsigned long ) ( lloffset >> 17 );
    }
 
    /* confirm addresses within specified range */
    if ( address >= prof->pr_off ) {
        /* test first for 16-bit buckets; this should be the fast case */
        if ( flags & PAPI_PROFIL_BUCKET_16 ) {
            if ( ( indx * sizeof ( short ) ) < prof->pr_size ) {
                buf16 = (unsigned short *) prof->pr_base;
                buf16[indx] =
                    ( unsigned short ) ( ( unsigned short ) buf16[indx] +
                                         profil_increment( buf16[indx], flags,
                                                           excess,
                                                           threshold ) );
                PRFDBG( "posix_profil_16() bucket %lu = %u\n", indx,
                        buf16[indx] );
            }
        }
        /* next, look for the 32-bit case */
        else if ( flags & PAPI_PROFIL_BUCKET_32 ) {
            if ( ( indx * sizeof ( int ) ) < prof->pr_size ) {
                buf32 = (unsigned int *) prof->pr_base;
                buf32[indx] = ( unsigned int ) buf32[indx] +
                    ( unsigned int ) profil_increment( buf32[indx], flags,
                                                       excess, threshold );
                PRFDBG( "posix_profil_32() bucket %lu = %u\n", indx,
                        buf32[indx] );
            }
        }
        /* finally, fall through to the 64-bit case */
        else {
            if ( ( indx * sizeof ( long long ) ) < prof->pr_size ) {
                buf64 = (unsigned long long *) prof->pr_base;
                buf64[indx] = ( unsigned long long ) buf64[indx] +
                    ( unsigned long long ) profil_increment( ( long long )
                                                             buf64[indx], flags,
                                                             excess,
                                                             threshold );
                PRFDBG( "posix_profil_64() bucket %lu = %lld\n", indx,
                        buf64[indx] );
            }
        }
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

profil_increment()

inline_static int profil_increment	(	long long	value,
		int	flags,
		long long	excess,
		long long	threshold
	)

Definition at line 58 of file extras.c.

{
    int increment = 1;
 
    if ( flags == PAPI_PROFIL_POSIX ) {
        return ( 1 );
    }
 
    if ( flags & PAPI_PROFIL_RANDOM ) {
        if ( random_ushort(  ) <= ( USHRT_MAX / 4 ) )
            return ( 0 );
    }
 
    if ( flags & PAPI_PROFIL_COMPRESS ) {
        /* We're likely to ignore the sample if buf[address] gets big. */
        if ( random_ushort(  ) < value ) {
            return ( 0 );
        }
    }
 
    if ( flags & PAPI_PROFIL_WEIGHTED ) {   /* Increment is between 1 and 255 */
        if ( excess <= ( long long ) 1 )
            increment = 1;
        else if ( excess > threshold )
            increment = 255;
        else {
            threshold = threshold / ( long long ) 255;
            increment = ( int ) ( excess / threshold );
        }
    }
    return ( increment );
}

Here is the call graph for this function:

Here is the caller graph for this function:

random_ushort()

inline_static unsigned short random_ushort

(

void

)

Definition at line 46 of file extras.c.

{
    return ( unsigned short ) ( _rnum = 1664525 * _rnum + 1013904223 );
}

Here is the caller graph for this function:

Variable Documentation

_papi_hwi_using_signal

int _papi_hwi_using_signal[PAPI_NSIG]

Definition at line 365 of file extras.c.

_rnum

unsigned int _rnum = DEADBEEF

static

Definition at line 39 of file extras.c.