v1.0

• Removed Deprecated Header Files: All legacy .h forwarding headers (e.g., perfcpp/sampler.h, perfcpp/event_counter.h) have been removed. Use .hpp files instead.
• Richer Branch Stack Entries: Each entry in a branch stack sample can now carry additional hardware-reported metadata (see the sampling documentation):
  • Classification (Linux 4.15+): the type of branch instruction, i.e., conditional, unconditional, call, return, syscall, and more.
  • Speculation Result (Linux 6.1+): whether the branch was on the correct or wrong speculative path.
  • Privilege Level (Linux 6.1+): whether the branch executed in user space, kernel space, or at hypervisor level.
• New IBS Sample Fields (AMD IBS PMU only): Three additional fields decoded from IBS raw data, none of which are accessible via the standard perf_event_open interface (see the sampling documentation):
  • Op Cache Miss (record.instruction_execution().cache()->is_op_miss()): indicates the fetch missed the op cache (decoded instruction cache), even if the L1 instruction cache hit. Enabled via sampler.values().instruction_cache(true). (IBS Fetch PMU)
  • iTLB Refill Latency (record.instruction_execution().latency().itlb_refill()): cycles to refill the instruction TLB after a miss. Only set when an iTLB miss occurred. Enabled via sampler.values().instruction_latency(true). (IBS Fetch PMU)
  • Is Microcode (record.instruction_execution().is_microcode()): indicates the sampled op was dispatched from the microcode ROM sequencer. Enabled via sampler.values().instruction_type(true). (IBS Op PMU)
• Configurable Sample Timestamp Clock: SampleConfig::clock(perf::Clock) selects the POSIX clock used for timestamps, making them directly comparable to clock_gettime() values without offset calibration. Note: Clock::Realtime and Clock::Boottime are rejected by the kernel for hardware PMU events on x86. Use Clock::Monotonic or Clock::MonotonicRaw instead. See the sampling documentation.
• Removed Deprecated Config Setters: The Config::is_pinned(bool) and Config::is_debug(bool) setters (deprecated since v0.13.0) have been removed. Use pinned(bool) and debug(bool) instead.
• Renamed Sample Fields: Two recording fields have clearer names. The old names still work but are deprecated and will be removed in v2.0:
  • values().sample_id(true) replaces values().id(true) and now matches the record.metadata().sample_id() accessor. The CSV column was renamed from id to sample_id as well.
  • values().data_access_misaligned(true) and record.data_access().is_misaligned() replace the *_misalign_penalty() variants.
• CMake Package Support: Fixed the installation process. Installed builds now ship a CMake package config, so find_package(perf-cpp) works without manual setup.
• Raised Compiler Requirements: Building now requires GCC 11 and newer or Clang 14 and newer (for std::from_chars support).
• Extended Event Configuration: Added config3 and config4 fields for raw event configuration.
• Tests on ARM: x86-only tests (Sampler, live events) are skipped on other architectures, so the test suite now runs on ARM (e.g., Raspberry Pi).
• Bugfixes:
  • Live counter reads no longer return 0 when they race a concurrent kernel update (seqlock retry in the mmap buffer).
  • The flamegraph export counted every group of identical call stacks one sample too many.
  • Symbol resolver: ELF virtual addresses were treated as file offsets, which produced wrong symbols. Modules that cannot be opened no longer abort resolution, and modules mapped with shared permissions are now resolved.
  • Replaced select() with poll() in the sample buffer to avoid a possible stack buffer overflow when monitoring many file descriptors.
  • EventCounter counted fixed hardware events incorrectly when forming groups.
  • CPU detection: Intel Comet Lake was misclassified as 12th generation (it is 10th), and Intel Rocket Lake was misclassified as well.
  • Hardware counter detection applied a wrong offset for the counter consumed by the NMI watchdog.
  • The perf file export wrote the logical memory address and sample ID in the wrong order.
  • Sampler: fixed an offset when recording performance counter values with samples, multiple IBS decoding bugs, and a missing vendor check for IBS triggers.
  • Event-file parsing no longer depends on the system locale and avoids undefined behavior in the tokenizer.
  • The JSON export now escapes special characters in counter and metric names.
  • Register values are reported as unsigned 64-bit integers.