Intel Skylake supports reporting the time in cycles a branch in the LBR took, to give a rough indication of the basic block performance. Export the cycle information in the branch_info structure. This can be done by just reusing some currently zero padding. This is just the generic header change. The architecture still needs to fill it in. There's no attempt to convert to real time, as we really want cycles here.
Showing the changes to include/uapi (1 of 1 file). Show all
Unified · Split
| @@ -951,6 +951,7 @@ union perf_mem_data_src { · 950 unmodified lines | ||
| 951 | 951 | * |
| 952 | 952 | * in_tx: running in a hardware transaction |
| 953 | 953 | * abort: aborting a hardware transaction |
| 954 | + * cycles: cycles from last branch (or 0 if not supported) | |
| 954 | 955 | */ |
| 955 | 956 | struct perf_branch_entry { |
| 956 | 957 | __u64 from; |
| @@ -959,7 +960,8 @@ struct perf_branch_entry { · 2 unmodified lines | ||
| 959 | 960 | predicted:1,/* target predicted */ |
| 960 | 961 | in_tx:1, /* in transaction */ |
| 961 | 962 | abort:1, /* transaction abort */ |
| 962 | − reserved:60; | |
| 963 | + cycles:16, /* cycle count to last branch */ | |
| 964 | + reserved:44; | |
| 963 | 965 | }; |
| 964 | 966 | |
| 965 | 967 | #endif /* _UAPI_LINUX_PERF_EVENT_H */ |