This component accumulates profiling information from a target CPU and generates a gprof-compatible histogram file.
This component emulates profiling routines embedded in target programs, as if they were compiled with gcc -pg.
| Behaviors | |
|---|---|
| configuration |
This component needs to be configured with the target-component relation. The first element in the relation will be used as the target of profiling. The value-attribute setting contains the name of the target component's attribute that will be collected. The bucket-size attribute configures the width of the histogram buckets. It is invalid to change the bucket size if any samples have been accumulated (if value-count is nonzero). The output-file setting gives the name of the eventual profiling output file. The output-file-endianness setting determines the endianness of the gprof formatted output. |
| data gathering |
Whenever the sample pin is driven, this component takes a single sample of the configured target component's value attribute. The attribute is interpreted as a numeric string, and the resulting number is accumulated in a big histogram, in the appropriate bucket. The bucket is chosen by masking the number into bucket-size-wide buckets. If the target component is unset, or its target attribute does not result in a valid numeric string, no sample is accumulated. Alternately, if the cg-caller and then the cg-callee pins are driven, the values driven are interpreted as the caller and the callee PC addresses of a subroutine call event. These events are recorded in a dynamic call graph. If a sample was collected, and falls between the limit-min and limit-max attributes' bounds, value-count is incremented, and value-min and value-max registers are updated. |
| resetting |
Whenever the reset pin is driven, the entire accumulated histogram and associated counters are discarded and a new count will be started. It is not necessary to drive this pin at simulation startup. |
| profile generation | Whenever the store pin is driven, this component dumps the entire accumulated histogram and call graph into a gprof (version 1) formatted file. The file's endianness is chosen based on the target-file-endianness setting. If it is unset, the target component's endian attribute will be queried and used if valid. The histogram and call graph are not automatically cleared after the store operation. This makes it possible to save multiple generational profiling files to track history of a program cumulatively. If the reset pin is triggered after every store, then separate intervals of the program's execution may be profiled. |
| SID Conventions | ||
|---|---|---|
| supervisory | supported | - |
| save/restore | not supported | - |
| triggerpoints | not supported | - |
| detect-reentrancy | not supported | - |
| category | supported | pinsetting |
Related components
This component is typically connected to a CPU, and told to monitor the CPU's program counter (pc) register. Sampling may be driven by any activity in SID, such as bus traffic (see hw-probe-bus), or host or target schedulers (see sid-sched-*). Perhaps the simplest way is to take a CPU PC sample every time the CPU has completed a batch of instructions.
new sw-profile-gprof gprof
new SOME_KIND_OF_CPU cpu
relate gprof target-component cpu
connect-pin shutdown-manager output-5 -> gprof store
connect-pin foo output -> cpu step!
connect-pin foo output -> gprof sample
# Set sampling interval by number of CPU instructions.
set cpu step-insn-count 127
Host system
The amount of data tracked by this component may be very large, if bucket-size is small and the target value has a large number of distinct values. Internally, a map or hash_map is used to track each bucket, so it does not matter how sparse the target value domain is. However, at store time, each potential bucket in the histogram, from value-min to value-max, is recorded, so sparse values may generate large blocks of zeroes in the output.
Since this component generates a histogram-type gprof file instead of the arc-type file that -pg-compiled target programs usually generate, gprof must be run with an option such as --flat-profile. Call graph information is not collected.
| pins | |||
|---|---|---|---|
| name | direction | legalvalues | behaviors |
| reset | in | any | resetting |
| sample | in | any | data gathering |
| cg-caller | in | any | data gathering |
| cg-callee | in | any | data gathering |
| store | in | any | profile generation |
| attributes | |||||
|---|---|---|---|---|---|
| name | category | legal values | default value | behaviors | |
| value-count | register | decimal string | - | data gathering | |
| value-min | register | decimal string | - | data gathering | |
| value-max | register | decimal string | - | data gathering | |
| limit-min | setting | decimal string | 0 | data gathering | |
| limit-max | setting | decimal string | infinity | data gathering | |
| value-attribute | setting | name | 'pc' | data gathering | |
| output-file | setting | file name | 'gmon.out' | configuration | |
| output-file-endianness | setting | 0/1/2/little/big/unknown | unknown | configuration | |
| reset | pin | - | - | resetting | |
| sample | pin | - | - | data gathering | |
| cg-caller | pin | - | - | data gathering | |
| cg-callee | pin | - | - | data gathering | |
| store | pin | - | - | profile generation | |