CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

Unreleased

0.1.16 - 2024-11-25

Added

• Thread pool for reusing threads across multi-threaded benchmarks. The result is that when running Divan benchmarks under a sampling profiler, the profiler's output will be cleaner and easier to understand. (#37)

• Track the maximum number of allocations during a benchmark.

Changed

• Make private Arg::get trait method not take self, so that text editors don't recommend using it. (#59)

• Cache BenchOptions using LazyLock instead of OnceLock, saving space and simplifying the implementation.

0.1.15 - 2024-10-31

Added

• CyclesCount counter to display cycle throughput as Hertz.

• Track the maximum number of bytes allocated during a benchmark.

Removed

• Remove has_cpuid polyfill due to it no longer being planned for Rust, since CPUID is assumed to be available on all old x86 Rust targets.

Fixed

• List generic benchmark type parameter A<4> before A<32>. (#64)

• Improve precision by using f64 when calculating allocation count and sizes for the median samples.

• Multi-thread allocation counting in sum_alloc_tallies on macOS was loading a null pointer instead of the pointer initialized by sync_threads.

Changes

• Sort all output benchmark names naturally instead of lexicographically.

• Internally reuse &[&str] slice for args names.

• Subtract overhead of AllocProfiler from timings. Now that Divan also tracks the maximum bytes allocated, the overhead was apparent in timings.

• Simplify ThreadAllocInfo::clear.

• Move measured loop overhead from SharedContext to global OnceLock.

• Macros no longer rely on std being re-exported by Divan. Instead they use ::std or ::core to greatly simplify code. Although this is technically a breaking change, it is extremely unlikely to do extern crate std as x.

0.1.14 - 2024-02-17

Fixed

• Set correct field in Divan::max_time. (#45)

Changes

• Improve args documentation by relating it to using Bencher.

• Define BytesCount::of_iter in terms of BytesCount::of_many.

0.1.13 - 2024-02-09

Fixed

• Missing update to divan-macros dependency.

0.1.12 - 2024-02-09

Added

• Display args option values with Debug instead if ToString is not implemented.

  This makes it simple to use enums with derived Debug:

  #[derive(Debug)]
  enum Arg { A, B }
  
  #[divan::bench(args = [Arg::A, Arg::B])]
  fn bench_args(arg: &Arg) {
      ...
  }

• Documentation of when to use black_box in benchmarks.

0.1.11 - 2024-01-20

Fixed

• Sorting negative args numbers.

0.1.10 - 2024-01-20

Fixed

• Sort args numbers like consts.

0.1.9 - 2024-01-20

Added

• args option for providing runtime arguments to benchmarks:

  #[divan::bench(args = [1, 2, 3])]
  fn args_list(arg: usize) { ... }
  
  #[divan::bench(args = 1..=3)]
  fn args_range(arg: usize) { ... }
  
  const ARGS: &[usize] = [1, 2, 3];
  
  #[divan::bench(args = ARGS)]
  fn args_const(arg: usize) { ... }

  This option may be preferred over the similar consts option because:

  • It is compatible with more types, only requiring that the argument type implements Any, Copy, Send, Sync, and ToString. Copy is not needed if the argument is used through a reference.
  • It does not increase compile times, unlike consts which needs to generate new code for each constant used.

0.1.8 - 2023-12-19

Changes

• Reduce AllocProfiler footprint from 6-10ns to 1-2ns:

  • Thread-local values are now exclusively owned by their threads and are no longer kept in a global list. This enables some optimizations:

    • Performing faster unsynchronized arithmetic.

    • Removing one level of pointer indirection by storing the thread-local value entirely inline in thread_local!, rather than storing a pointer to a globally-shared instance.

    • Compiler emits SIMD arithmetic for x86_64 using paddq.

  • Improved thread-local lookup on x86_64 macOS by using a static lookup key instead of a dynamic key from pthread_key_create. Key 11 is used because it is reserved for Windows.

    The dyn_thread_local crate feature disables this optimization. This is recommended if your code or another dependency uses the same static key.

Fixed

• Remove unused allocations if AllocProfiler is not active as the global allocator.

0.1.7 - 2023-12-13

Changes

• Improve AllocProfiler implementation documentation.

• Limit AllocProfiler mean count outputs to 4 significant digits to not be very wide and for consistency with other outputs.

0.1.6 - 2023-12-13

Added

• AllocProfiler allocator that tracks allocation counts and sizes during benchmarks.

0.1.5 - 2023-12-05

Added

• black_box_drop convenience function for black_box + drop. This is useful when benchmarking a lazy Iterator to completion with for_each:

  #[divan::bench]
  fn parse_iter() {
      let input: &str = // ...
  
      Parser::new(input)
          .for_each(divan::black_box_drop);
  }

0.1.4 - 2023-12-02

Added

• From implementations for counters on references to u8–u64 and usize, such as From<&u64> and From<&&u64>. This allows for doing:

  bencher
      .with_inputs(|| { ... })
      .input_counter(ItemsCount::from)
      .bench_values(|n| { ... });

• Bencher::count_inputs_as<C> method to convert inputs to a Counter:

  bencher
      .with_inputs(|| -> usize {
          // ...
      })
      .count_inputs_as::<ItemsCount>()
      .bench_values(|n| -> Vec<usize> {
          (0..n).collect()
      });

0.1.3 - 2023-11-21

Added

• Convenience shorthand options for #[divan::bench] and #[divan::bench_group] counters:

  • bytes_count for counter = BytesCount::from(n)
  • chars_count for counter = CharsCount::from(n)
  • items_count for counter = ItemsCount::from(n)

• Support for NetBSD, DragonFly BSD, and Haiku OS by using pre-main.

• Set global thread counts using:

  • Divan::threads
  • --threads A B C... CLI arg
  • DIVAN_THREADS=A,B,C env var

  The following example will benchmark across 2, 4, and available parallelism thread counts:

  DIVAN_THREADS=0,2,4 cargo bench -q -p examples --bench atomic

• Set global Counters at runtime using:

  • Divan::counter
  • Divan::items_count
  • Divan::bytes_count
  • Divan::chars_count
  • --items-count N CLI arg
  • --bytes-count N CLI arg
  • --chars-count N CLI arg
  • DIVAN_ITEMS_COUNT=N env var
  • DIVAN_BYTES_COUNT=N env var
  • DIVAN_CHARS_COUNT=N env var

• From<C> for ItemsCount, BytesCount, and CharsCount where C is u8–u64 or usize (via CountUInt internally). This provides an alternative to the new constructor.

• BytesCount::of_many method similar to BytesCount::of, but with a parameter by which to multiply the size of the type.

• BytesCount::u64, BytesCount::f64, and similar methods based on BytesCount::of_many.

Removed

• black_box inside benchmark loop when deferring Drop of outputs. This is now done after the loop.

• linkme dependency in favor of pre-main to register benchmarks and benchmark groups. This is generally be more portable and reliable.

Changed

• Now calling black_box at the end of the benchmark loop when deferring use of inputs or Drop of outputs.

0.1.2 - 2023-10-28

Fixed

• Multi-threaded benchmarks being spread across CPUs, instead of pinning the main thread to CPU 0 and having all threads inherit the main thread's affinity.

0.1.1 - 2023-10-25

Fixed

• Fix using LLD as linker for Linux by using the same pre-main approach as Windows.

0.1.0 - 2023-10-04

Initial release. See blog post.