int
pfm_list_supported_pmus(int (*pf)(const char *fmt,...))
{
pfm_pmu_support_t **p = pmus;
char *name;
if (pf == NULL) return PFMLIB_ERR_INVAL;
(*pf)("supported PMU models: ");
for (p = pmus; *p; p++) {
(*pf)("[%s] ", pmu_names[(*p)->pmu_type]);
}
pfm_get_pmu_name(&name);
(*pf)("\ndetected host PMU: %s\n", name);
return PFMLIB_SUCCESS;
}
int
pfm_force_pmu(int type)
{
pfm_pmu_support_t **p = pmus;
char *name;
while (*p) {
if ((*p)->pmu_type == type) goto found;
p++;
}
return PFMLIB_ERR_NOTSUPP;
found:
/* verify that this is valid */
if ((*p)->cpu_detect() != PFMLIB_SUCCESS) return PFMLIB_ERR_NOTSUPP;
pfm_config.current = *p;
pfm_get_pmu_name(&name);
pfm_current = *p;
return PFMLIB_SUCCESS;
}
int
_papi_libpfm_init(papi_vector_t *my_vector, int cidx) {
int retval;
unsigned int ncnt;
unsigned int version;
char pmu_name[PAPI_MIN_STR_LEN];
/* The following checks the version of the PFM library
against the version PAPI linked to... */
SUBDBG( "pfm_initialize()\n" );
if ( ( retval = pfm_initialize( ) ) != PFMLIB_SUCCESS ) {
PAPIERROR( "pfm_initialize(): %s", pfm_strerror( retval ) );
return PAPI_ESYS;
}
/* Get the libpfm3 version */
SUBDBG( "pfm_get_version(%p)\n", &version );
if ( pfm_get_version( &version ) != PFMLIB_SUCCESS ) {
PAPIERROR( "pfm_get_version(%p): %s", version, pfm_strerror( retval ) );
return PAPI_ESYS;
}
/* Set the version */
sprintf( my_vector->cmp_info.support_version, "%d.%d",
PFM_VERSION_MAJOR( version ), PFM_VERSION_MINOR( version ) );
/* Complain if the compiled-against version doesn't match current version */
if ( PFM_VERSION_MAJOR( version ) != PFM_VERSION_MAJOR( PFMLIB_VERSION ) ) {
PAPIERROR( "Version mismatch of libpfm: compiled %#x vs. installed %#x\n",
PFM_VERSION_MAJOR( PFMLIB_VERSION ),
PFM_VERSION_MAJOR( version ) );
return PAPI_ESYS;
}
/* Always initialize globals dynamically to handle forks properly. */
_perfmon2_pfm_pmu_type = -1;
/* Opened once for all threads. */
SUBDBG( "pfm_get_pmu_type(%p)\n", &_perfmon2_pfm_pmu_type );
if ( pfm_get_pmu_type( &_perfmon2_pfm_pmu_type ) != PFMLIB_SUCCESS ) {
PAPIERROR( "pfm_get_pmu_type(%p): %s", _perfmon2_pfm_pmu_type,
pfm_strerror( retval ) );
return PAPI_ESYS;
}
pmu_name[0] = '\0';
if ( pfm_get_pmu_name( pmu_name, PAPI_MIN_STR_LEN ) != PFMLIB_SUCCESS ) {
PAPIERROR( "pfm_get_pmu_name(%p,%d): %s", pmu_name, PAPI_MIN_STR_LEN,
pfm_strerror( retval ) );
return PAPI_ESYS;
}
SUBDBG( "PMU is a %s, type %d\n", pmu_name, _perfmon2_pfm_pmu_type );
/* Setup presets */
retval = _papi_load_preset_table( pmu_name, _perfmon2_pfm_pmu_type, cidx );
if ( retval )
return retval;
/* Fill in cmp_info */
SUBDBG( "pfm_get_num_events(%p)\n", &ncnt );
if ( ( retval = pfm_get_num_events( &ncnt ) ) != PFMLIB_SUCCESS ) {
PAPIERROR( "pfm_get_num_events(%p): %s\n", &ncnt,
pfm_strerror( retval ) );
return PAPI_ESYS;
}
SUBDBG( "pfm_get_num_events: %d\n", ncnt );
my_vector->cmp_info.num_native_events = ncnt;
num_native_events = ncnt;
pfm_get_num_counters( ( unsigned int * ) &my_vector->cmp_info.num_cntrs );
SUBDBG( "pfm_get_num_counters: %d\n", my_vector->cmp_info.num_cntrs );
if ( _papi_hwi_system_info.hw_info.vendor == PAPI_VENDOR_INTEL ) {
/* Pentium4 */
if ( _papi_hwi_system_info.hw_info.cpuid_family == 15 ) {
PAPI_NATIVE_EVENT_AND_MASK = 0x000000ff;
PAPI_NATIVE_UMASK_AND_MASK = 0x0fffff00;
PAPI_NATIVE_UMASK_SHIFT = 8;
/* Itanium2 */
} else if ( _papi_hwi_system_info.hw_info.cpuid_family == 31 ||
_papi_hwi_system_info.hw_info.cpuid_family == 32 ) {
PAPI_NATIVE_EVENT_AND_MASK = 0x00000fff;
PAPI_NATIVE_UMASK_AND_MASK = 0x0ffff000;
PAPI_NATIVE_UMASK_SHIFT = 12;
}
}
return PAPI_OK;
}
int
main(void)
{
int ret;
int type = 0;
pid_t pid = getpid();
pfmlib_ita2_param_t ita_param;
pfarg_reg_t pd[NUM_PMDS];
pfarg_context_t ctx[1];
pfmlib_options_t pfmlib_options;
struct sigaction act;
/*
* Initialize pfm library (required before we can use it)
*/
if (pfm_initialize() != PFMLIB_SUCCESS) {
fatal_error("Can't initialize library\n");
}
/*
* Let's make sure we run this on the right CPU
*/
pfm_get_pmu_type(&type);
if (type != PFMLIB_ITANIUM2_PMU) {
char *model;
pfm_get_pmu_name(&model);
fatal_error("this program does not work with %s PMU\n", model);
}
/*
* Install the overflow handler (SIGPROF)
*/
memset(&act, 0, sizeof(act));
act.sa_handler = (sig_t)overflow_handler;
sigaction (SIGPROF, &act, 0);
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* set to 1 for debug */
pfmlib_options.pfm_verbose = 0; /* set to 1 for debug */
pfm_set_options(&pfmlib_options);
memset(pd, 0, sizeof(pd));
memset(ctx, 0, sizeof(ctx));
/*
* prepare parameters to library. we don't use any Itanium
* specific features here. so the pfp_model is NULL.
*/
memset(&evt,0, sizeof(evt));
memset(&ita_param,0, sizeof(ita_param));
/*
* because we use a model specific feature, we must initialize the
* model specific pfmlib parameter structure and link it to the
* common structure.
* The magic number is a simple mechanism used by the library to check
* that the model specific data structure is decent. You must set it manually
* otherwise the model specific feature won't work.
*/
ita_param.pfp_magic = PFMLIB_ITA2_PARAM_MAGIC;
evt.pfp_model = &ita_param;
/*
* Before calling pfm_find_dispatch(), we must specify what kind
* of branches we want to capture. We are interesteed in all the mispredicted branches,
* therefore we program we set the various fields of the BTB config to:
*/
ita_param.pfp_ita2_btb.btb_used = 1;
ita_param.pfp_ita2_btb.btb_ds = 0;
ita_param.pfp_ita2_btb.btb_tm = 0x3;
ita_param.pfp_ita2_btb.btb_ptm = 0x3;
ita_param.pfp_ita2_btb.btb_ppm = 0x3;
ita_param.pfp_ita2_btb.btb_brt = 0x0;
ita_param.pfp_ita2_btb.btb_plm = PFM_PLM3;
/*
* To count the number of occurence of this instruction, we must
* program a counting monitor with the IA64_TAGGED_INST_RETIRED_PMC8
* event.
*/
if (pfm_find_event_byname("BRANCH_EVENT", &evt.pfp_events[0].event) != PFMLIB_SUCCESS) {
fatal_error("cannot find event BRANCH_EVENT\n");
}
/*
* set the (global) privilege mode:
* PFM_PLM3 : user level only
*/
evt.pfp_dfl_plm = PFM_PLM3;
/*
* how many counters we use
*/
evt.pfp_event_count = 1;
/*
* let the library figure out the values for the PMCS
*/
if ((ret=pfm_dispatch_events(&evt)) != PFMLIB_SUCCESS) {
fatal_error("cannot configure events: %s\n", pfm_strerror(ret));
}
/*
* for this example, we will get notified ONLY when the sampling
* buffer is full. The monitoring is not to be inherited
* in derived tasks
*/
ctx[0].ctx_flags = PFM_FL_INHERIT_NONE;
ctx[0].ctx_notify_pid = getpid();
ctx[0].ctx_smpl_entries = SMPL_BUF_NENTRIES;
ctx[0].ctx_smpl_regs[0] = smpl_regs = BTB_REGS_MASK;
/*
* now create the context for self monitoring/per-task
*/
if (perfmonctl(pid, PFM_CREATE_CONTEXT, ctx, 1) == -1 ) {
if (errno == ENOSYS) {
fatal_error("Your kernel does not have performance monitoring support!\n");
}
fatal_error("Can't create PFM context %s\n", strerror(errno));
}
printf("Sampling buffer mapped at %p\n", ctx[0].ctx_smpl_vaddr);
smpl_vaddr = ctx[0].ctx_smpl_vaddr;
/*
* Must be done before any PMD/PMD calls (unfreeze PMU). Initialize
* PMC/PMD to safe values. psr.up is cleared.
*/
if (perfmonctl(pid, PFM_ENABLE, NULL, 0) == -1) {
fatal_error("perfmonctl error PFM_ENABLE errno %d\n",errno);
}
/*
* indicate we want notification when buffer is full
*/
evt.pfp_pc[0].reg_flags |= PFM_REGFL_OVFL_NOTIFY;
/*
* Now prepare the argument to initialize the PMD and the sampling period
*/
pd[0].reg_num = evt.pfp_pc[0].reg_num;
pd[0].reg_value = (~0UL) - SMPL_PERIOD +1;
pd[0].reg_long_reset = (~0UL) - SMPL_PERIOD +1;
pd[0].reg_short_reset = (~0UL) - SMPL_PERIOD +1;
/*
* When our counter overflows, we want to BTB index to be reset, so that we keep
* in sync. This is required to make it possible to interpret pmd16 on overflow
* to avoid repeating the same branch several times.
*/
evt.pfp_pc[0].reg_reset_pmds[0] = M_PMD(16);
/*
* reset pmd16, short and long reset value are set to zero as well
*/
pd[1].reg_num = 16;
pd[1].reg_value = 0UL;
/*
* Now program the registers
*
* We don't use the save variable to indicate the number of elements passed to
* the kernel because, as we said earlier, pc may contain more elements than
* the number of events we specified, i.e., contains more thann coutning monitors.
*/
if (perfmonctl(pid, PFM_WRITE_PMCS, evt.pfp_pc, evt.pfp_pc_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMCS errno %d\n",errno);
}
if (perfmonctl(pid, PFM_WRITE_PMDS, pd, 2) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMDS errno %d\n",errno);
}
/*
* Let's roll now.
*/
do_test(100000);
/*
* We must call the processing routine to cover the last entries recorded
* in the sampling buffer, i.e. which may not be full
*/
process_smpl_buffer();
/*
* let's stop this now
*/
if (perfmonctl(pid, PFM_DESTROY_CONTEXT, NULL, 0) == -1) {
fatal_error("perfmonctl error PFM_DESTROY errno %d\n",errno);
}
return 0;
}
int
main(void)
{
pfmlib_input_param_t inp;
pfmlib_output_param_t outp;
pfmlib_ita2_input_param_t ita2_inp;
pfarg_reg_t pd[NUM_PMDS];
pfarg_reg_t pc[NUM_PMCS];
pfarg_context_t ctx[1];
pfarg_load_t load_args;
pfmlib_options_t pfmlib_options;
int ret;
int type = 0;
int id;
unsigned int i;
char name[MAX_EVT_NAME_LEN];
/*
* Initialize pfm library (required before we can use it)
*/
if (pfm_initialize() != PFMLIB_SUCCESS) {
fatal_error("Can't initialize library\n");
}
/*
* Let's make sure we run this on the right CPU
*/
pfm_get_pmu_type(&type);
if (type != PFMLIB_ITANIUM2_PMU) {
char model[MAX_PMU_NAME_LEN];
pfm_get_pmu_name(model, MAX_PMU_NAME_LEN);
fatal_error("this program does not work with the %s PMU\n", model);
}
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* set to 1 for debug */
pfmlib_options.pfm_verbose = 0; /* set to 1 for verbose */
pfm_set_options(&pfmlib_options);
memset(pd, 0, sizeof(pd));
memset(pc, 0, sizeof(pc));
memset(ctx, 0, sizeof(ctx));
memset(&load_args, 0, sizeof(load_args));
memset(&inp,0, sizeof(inp));
memset(&outp,0, sizeof(outp));
memset(&ita2_inp,0, sizeof(ita2_inp));
/*
* We indicate that we are using the PMC8 opcode matcher. This is required
* otherwise the library add PMC8 to the list of PMC to pogram during
* pfm_dispatch_events().
*/
ita2_inp.pfp_ita2_pmc8.opcm_used = 1;
/*
* We want to match all the br.cloop in our test function.
* This branch is an IP-relative branch for which the major
* opcode (bits [40-37]=4) and the btype field is 5 (which represents
* bits[6-8]) so it is included in the match/mask fields of PMC8.
* It is necessarily in a B slot.
*
* We don't care which operands are used with br.cloop therefore
* the mask field of pmc8 is set such that only the 4 bits of the
* opcode and 3 bits of btype must match exactly. This is accomplished by
* clearing the top 4 bits and bits [6-8] of the mask field and setting the
* remaining bits. Similarly, the match field only has the opcode value and btype
* set according to the encoding of br.cloop, the
* remaining bits are zero. Bit 60 of PMC8 is set to indicate
* that we look only in B slots (this is the only possibility for
* this instruction anyway).
*
* So the binary representation of the value for PMC8 is as follows:
*
* 6666555555555544444444443333333333222222222211111111110000000000
* 3210987654321098765432109876543210987654321098765432109876543210
* ----------------------------------------------------------------
* 0001010000000000000000101000000000000011111111111111000111111000
*
* which yields a value of 0x1400028003fff1f8.
*
* Depending on the level of optimization to compile this code, it may
* be that the count reported could be zero, if the compiler uses a br.cond
* instead of br.cloop.
*
*
* The 0x1 sets the ig_ad field to make sure we ignore any range restriction.
* Also bit 2 must always be set
*/
ita2_inp.pfp_ita2_pmc8.pmc_val = 0x1400028003fff1fa;
/*
* To count the number of occurence of this instruction, we must
* program a counting monitor with the IA64_TAGGED_INST_RETIRED_PMC8
* event.
*/
if (pfm_find_full_event("IA64_TAGGED_INST_RETIRED_IBRP0_PMC8", &inp.pfp_events[0]) != PFMLIB_SUCCESS) {
fatal_error("cannot find event IA64_TAGGED_INST_RETIRED_IBRP0_PMC8\n");
}
/*
* set the privilege mode:
* PFM_PLM3 : user level only
*/
inp.pfp_dfl_plm = PFM_PLM3;
/*
* how many counters we use
*/
inp.pfp_event_count = 1;
/*
* let the library figure out the values for the PMCS
*/
if ((ret=pfm_dispatch_events(&inp, &ita2_inp, &outp, NULL)) != PFMLIB_SUCCESS) {
fatal_error("cannot configure events: %s\n", pfm_strerror(ret));
}
/*
* now create the context for self monitoring/per-task
*/
if (perfmonctl(0, PFM_CREATE_CONTEXT, ctx, 1) == -1 ) {
if (errno == ENOSYS) {
fatal_error("Your kernel does not have performance monitoring support!\n");
}
fatal_error("Can't create PFM context %s\n", strerror(errno));
}
/*
* extract the unique identifier for our context, a regular file descriptor
*/
id = ctx[0].ctx_fd;
/*
* Now prepare the argument to initialize the PMDs and PMCS.
* We must pfp_pmc_count to determine the number of PMC to intialize.
* We must use pfp_event_count to determine the number of PMD to initialize.
* Some events causes extra PMCs to be used, so pfp_pmc_count may be >= pfp_event_count.
*
* This step is new compared to libpfm-2.x. It is necessary because the library no
* longer knows about the kernel data structures.
*/
for (i=0; i < outp.pfp_pmc_count; i++) {
pc[i].reg_num = outp.pfp_pmcs[i].reg_num;
pc[i].reg_value = outp.pfp_pmcs[i].reg_value;
}
/*
* the PMC controlling the event ALWAYS come first, that's why this loop
* is safe even when extra PMC are needed to support a particular event.
*/
for (i=0; i < inp.pfp_event_count; i++) {
pd[i].reg_num = pc[i].reg_num;
}
printf("event_count=%d id=%d\n", inp.pfp_event_count, id);
/*
* Now program the registers
*
* We don't use the save variable to indicate the number of elements passed to
* the kernel because, as we said earlier, pc may contain more elements than
* the number of events we specified, i.e., contains more thann coutning monitors.
*/
if (perfmonctl(id, PFM_WRITE_PMCS, pc, outp.pfp_pmc_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMCS errno %d\n",errno);
}
if (perfmonctl(id, PFM_WRITE_PMDS, pd, inp.pfp_event_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMDS errno %d\n",errno);
}
/*
* now we load (i.e., attach) the context to ourself
*/
load_args.load_pid = getpid();
if (perfmonctl(id, PFM_LOAD_CONTEXT, &load_args, 1) == -1) {
fatal_error("perfmonctl error PFM_LOAD_CONTEXT errno %d\n",errno);
}
/*
* Let's roll now.
*/
pfm_self_start(id);
do_test(100UL);
pfm_self_stop(id);
/*
* now read the results
*/
if (perfmonctl(id, PFM_READ_PMDS, pd, inp.pfp_event_count) == -1) {
fatal_error("perfmonctl error READ_PMDS errno %d\n",errno);
}
/*
* print the results
*/
pfm_get_full_event_name(&inp.pfp_events[0], name, MAX_EVT_NAME_LEN);
printf("PMD%u %20lu %s\n",
pd[0].reg_num,
pd[0].reg_value,
name);
if (pd[0].reg_value != 0)
printf("compiler used br.cloop\n");
else
printf("compiler did not use br.cloop\n");
/*
* let's stop this now
*/
close(id);
return 0;
}
int
main(void)
{
int ret;
int type = 0;
char *name;
pid_t pid = getpid();
pfmlib_param_t evt;
pfmlib_ita2_param_t ita2_param;
pfarg_reg_t pd[NUM_PMDS];
pfarg_context_t ctx[1];
pfmlib_options_t pfmlib_options;
/*
* Initialize pfm library (required before we can use it)
*/
if (pfm_initialize() != PFMLIB_SUCCESS) {
fatal_error("Can't initialize library\n");
}
/*
* Let's make sure we run this on the right CPU
*/
pfm_get_pmu_type(&type);
if (type != PFMLIB_ITANIUM2_PMU) {
char *model;
pfm_get_pmu_name(&model);
fatal_error("this program does not work with the %s PMU\n", model);
}
/*
* pass options to library (optional)
*/
memset(&pfmlib_options, 0, sizeof(pfmlib_options));
pfmlib_options.pfm_debug = 0; /* set to 1 for debug */
pfmlib_options.pfm_verbose = 0; /* set to 1 for verbose */
pfm_set_options(&pfmlib_options);
memset(pd, 0, sizeof(pd));
memset(ctx, 0, sizeof(ctx));
memset(&evt,0, sizeof(evt));
memset(&ita2_param,0, sizeof(ita2_param));
/*
* because we use a model specific feature, we must initialize the
* model specific pfmlib parameter structure and link it to the
* common structure.
* The magic number is a simple mechanism used by the library to check
* that the model specific data structure is decent. You must set it manually
* otherwise the model specific feature won't work.
*/
ita2_param.pfp_magic = PFMLIB_ITA2_PARAM_MAGIC;
evt.pfp_model = &ita2_param;
/*
* We indicate that we are using the PMC8 opcode matcher. This is required
* otherwise the library add PMC8 to the list of PMC to pogram during
* pfm_dispatch_events().
*/
ita2_param.pfp_ita2_pmc8.opcm_used = 1;
/*
* We want to match all the br.cloop in our test function.
* This branch is an IP-relative branch for which the major
* opcode (bits [40-37]=4) and the btype field is 5 (which represents
* bits[6-8]) so it is included in the match/mask fields of PMC8.
* It is necessarily in a B slot.
*
* We don't care which operands are used with br.cloop therefore
* the mask field of pmc8 is set such that only the 4 bits of the
* opcode and 3 bits of btype must match exactly. This is accomplished by
* clearing the top 4 bits and bits [6-8] of the mask field and setting the
* remaining bits. Similarly, the match field only has the opcode value and btype
* set according to the encoding of br.cloop, the
* remaining bits are zero. Bit 60 of PMC8 is set to indicate
* that we look only in B slots (this is the only possibility for
* this instruction anyway).
*
* So the binary representation of the value for PMC8 is as follows:
*
* 6666555555555544444444443333333333222222222211111111110000000000
* 3210987654321098765432109876543210987654321098765432109876543210
* ----------------------------------------------------------------
* 0001010000000000000000101000000000000011111111111111000111111000
*
* which yields a value of 0x1400028003fff1f8.
*
* Depending on the level of optimization to compile this code, it may
* be that the count reported could be zero, if the compiler uses a br.cond
* instead of br.cloop.
*
*
* The 0x1 sets the ig_ad field to make sure we ignore any range restriction.
* Also bit 2 must always be set
*/
ita2_param.pfp_ita2_pmc8.pmc_val = 0x1400028003fff1fa;
/*
* To count the number of occurence of this instruction, we must
* program a counting monitor with the IA64_TAGGED_INST_RETIRED_PMC8
* event.
*/
if (pfm_find_event_byname("IA64_TAGGED_INST_RETIRED_IBRP0_PMC8", &evt.pfp_events[0].event) != PFMLIB_SUCCESS) {
fatal_error("cannot find event IA64_TAGGED_INST_RETIRED_IBRP0_PMC8\n");
}
/*
* set the privilege mode:
* PFM_PLM3 : user level only
*/
evt.pfp_dfl_plm = PFM_PLM3;
/*
* how many counters we use
*/
evt.pfp_event_count = 1;
/*
* let the library figure out the values for the PMCS
*/
if ((ret=pfm_dispatch_events(&evt)) != PFMLIB_SUCCESS) {
fatal_error("cannot configure events: %s\n", pfm_strerror(ret));
}
/*
* for this example, we have decided not to get notified
* on counter overflows and the monitoring is not to be inherited
* in derived tasks
*/
ctx[0].ctx_flags = PFM_FL_INHERIT_NONE;
/*
* now create the context for self monitoring/per-task
*/
if (perfmonctl(pid, PFM_CREATE_CONTEXT, ctx, 1) == -1 ) {
if (errno == ENOSYS) {
fatal_error("Your kernel does not have performance monitoring support!\n");
}
fatal_error("Can't create PFM context %s\n", strerror(errno));
}
/*
* Must be done before any PMD/PMD calls (unfreeze PMU). Initialize
* PMC/PMD to safe values. psr.up is cleared.
*/
if (perfmonctl(pid, PFM_ENABLE, NULL, 0) == -1) {
fatal_error("perfmonctl error PFM_ENABLE errno %d\n",errno);
}
/*
* Now prepare the argument to initialize the PMD.
*/
pd[0].reg_num = evt.pfp_pc[0].reg_num;
/*
* Now program the registers
*
* We don't use the save variable to indicate the number of elements passed to
* the kernel because, as we said earlier, pc may contain more elements than
* the number of events we specified, i.e., contains more thann coutning monitors.
*/
if (perfmonctl(pid, PFM_WRITE_PMCS, evt.pfp_pc, evt.pfp_pc_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMCS errno %d\n",errno);
}
if (perfmonctl(pid, PFM_WRITE_PMDS, pd, evt.pfp_event_count) == -1) {
fatal_error("perfmonctl error PFM_WRITE_PMDS errno %d\n",errno);
}
/*
* Let's roll now.
*/
pfm_start();
do_test(100UL);
pfm_stop();
/*
* now read the results
*/
if (perfmonctl(pid, PFM_READ_PMDS, pd, evt.pfp_event_count) == -1) {
fatal_error("perfmonctl error READ_PMDS errno %d\n",errno);
}
/*
* print the results
*/
pfm_get_event_name(evt.pfp_events[0].event, &name);
printf("PMD%u %20lu %s\n",
pd[0].reg_num,
pd[0].reg_value,
name);
if (pd[0].reg_value != 0)
printf("compiler used br.cloop\n");
else
printf("compiler did not use br.cloop\n");
/*
* let's stop this now
*/
if (perfmonctl(pid, PFM_DESTROY_CONTEXT, NULL, 0) == -1) {
fatal_error("perfmonctl error PFM_DESTROY errno %d\n",errno);
}
return 0;
}
int
main(int argc, char **argv)
{
static char *argv_all[2] = { ".*", NULL };
char *endptr = NULL;
char **args;
int c, match;
regex_t preg;
char model[MAX_PMU_NAME_LEN];
while ((c=getopt(argc, argv,"hsm:")) != -1) {
switch(c) {
case 's':
options.sort = 1;
break;
case 'm':
options.mask = strtoull(optarg, &endptr, 16);
if (*endptr)
fatal_error("mask must be in hexadecimal\n");
break;
case 'h':
usage();
exit(0);
default:
fatal_error("unknown error");
}
}
if (pfm_initialize() != PFMLIB_SUCCESS)
fatal_error("PMU model not supported by library\n");
if (options.mask == 0)
options.mask = ~0;
if (optind == argc) {
args = argv_all;
} else {
args = argv + optind;
}
pfm_get_max_event_name_len(&max_len);
name = malloc(max_len+1);
if (name == NULL)
fatal_error("cannot allocate name buffer\n");
if (argc == 1)
*argv = ".*"; /* match everything */
else
++argv;
pfm_get_pmu_name(model, MAX_PMU_NAME_LEN);
printf("PMU model: %s\n", model);
while(*args) {
if (regcomp(&preg, *args, REG_ICASE|REG_NOSUB))
fatal_error("error in regular expression for event \"%s\"", *argv);
if (options.sort)
match = show_info_sorted(&preg);
else
match = show_info(&preg);
if (match == 0)
fatal_error("event %s not found", *args);
args++;
}
regfree(&preg);
free(name);
return 0;
}
