structmem.md

Reference

• https://youtube.com/watch?v=RRgFBrhml-c
• https://github.com/contextfreecode/structmem

updated: https://godbolt.org/z/fqT37xGPM - add Dlang and fix c++ initialize_list

Not installed .NET

# D betterC 
Benchmark 1 (121 runs): ./structmem_rs
  measurement          mean ± σ            min … max           outliers         delta
  wall_time           673us ± 61.8us     549us … 1.13ms         11 ( 9%)        0%
  peak_rss           1.98MB ± 72.4KB    1.84MB … 2.10MB         39 (32%)        0%
  cpu_cycles          297K  ± 25.8K      279K  …  487K          12 (10%)        0%
  instructions        325K  ±  289       325K  …  326K           0 ( 0%)        0%
  cache_references   18.4K  ±  402      17.9K  … 22.0K           6 ( 5%)        0%
  cache_misses       6.89K  ±  104      6.59K  … 7.20K           4 ( 3%)        0%
  branch_misses      4.28K  ± 29.9      4.20K  … 4.36K           3 ( 2%)        0%
Benchmark 2 (168 runs): ./structmem_d
  measurement          mean ± σ            min … max           outliers         delta
  wall_time           565us ± 46.9us     453us …  717us          7 ( 4%)        ⚡- 16.1% ±  1.9%
  peak_rss           1.79MB ± 63.4KB    1.70MB … 1.84MB          0 ( 0%)        ⚡- 10.0% ±  0.8%
  cpu_cycles          217K  ± 14.1K      205K  …  308K          28 (17%)        ⚡- 26.8% ±  1.6%
  instructions        183K  ± 41.0       183K  …  183K           0 ( 0%)        ⚡- 43.7% ±  0.0%
  cache_references   14.5K  ±  282      13.7K  … 16.5K           8 ( 5%)        ⚡- 21.1% ±  0.4%
  cache_misses       5.46K  ± 95.9      5.25K  … 5.87K           7 ( 4%)        ⚡- 20.7% ±  0.3%
  branch_misses      3.29K  ± 23.6      3.18K  … 3.36K           4 ( 2%)        ⚡- 23.1% ±  0.1%
Benchmark 3 (178 runs): ./structmem_cpp
  measurement          mean ± σ            min … max           outliers         delta
  wall_time           533us ± 44.5us     404us …  628us          9 ( 5%)        ⚡- 20.9% ±  1.8%
  peak_rss           1.71MB ± 89.6KB    1.57MB … 1.84MB          0 ( 0%)        ⚡- 13.9% ±  1.0%
  cpu_cycles          219K  ± 12.3K      210K  …  285K          21 (12%)        ⚡- 26.0% ±  1.5%
  instructions        205K  ± 20.3       205K  …  205K           0 ( 0%)        ⚡- 36.9% ±  0.0%
  cache_references   13.8K  ±  217      13.0K  … 15.0K           4 ( 2%)        ⚡- 24.7% ±  0.4%
  cache_misses       5.70K  ± 95.5      5.33K  … 6.03K           6 ( 3%)        ⚡- 17.3% ±  0.3%
  branch_misses      3.33K  ± 42.1      3.12K  … 3.41K          16 ( 9%)        ⚡- 22.2% ±  0.2%

# NO D BetterC (nogc only)
Benchmark 1 (96 runs): ./structmem_rs
  measurement          mean ± σ            min … max           outliers         delta
  wall_time           680us ± 70.5us     540us … 1.01ms         15 (16%)        0%
  peak_rss           1.98MB ± 71.8KB    1.84MB … 2.10MB         31 (32%)        0%
  cpu_cycles          301K  ± 33.5K      280K  …  524K           7 ( 7%)        0%
  instructions        325K  ±  288       325K  …  326K           0 ( 0%)        0%
  cache_references   18.4K  ±  205      18.0K  … 19.4K           2 ( 2%)        0%
  cache_misses       6.92K  ±  108      6.69K  … 7.30K           4 ( 4%)        0%
  branch_misses      4.28K  ± 27.3      4.22K  … 4.34K           0 ( 0%)        0%
Benchmark 2 (134 runs): ./structmem_d
  measurement          mean ± σ            min … max           outliers         delta
  wall_time           716us ± 66.0us     571us … 1.12ms         19 (14%)        💩+  5.4% ±  2.6%
  peak_rss           2.41MB ± 88.1KB    2.23MB … 2.49MB          0 ( 0%)        💩+ 21.6% ±  1.1%
  cpu_cycles          313K  ± 20.3K      299K  …  432K          12 ( 9%)        💩+  4.3% ±  2.3%
  instructions        313K  ± 63.4       313K  …  313K           0 ( 0%)        ⚡-  3.7% ±  0.0%
  cache_references   23.7K  ±  230      23.2K  … 24.3K           0 ( 0%)        💩+ 29.1% ±  0.3%
  cache_misses       7.35K  ±  131      7.08K  … 7.73K           1 ( 1%)        💩+  6.3% ±  0.5%
  branch_misses      4.27K  ± 23.8      4.23K  … 4.34K           1 ( 1%)          -  0.1% ±  0.2%
Benchmark 3 (178 runs): ./structmem_cpp
  measurement          mean ± σ            min … max           outliers         delta
  wall_time           531us ± 44.0us     420us …  681us         12 ( 7%)        ⚡- 21.8% ±  2.0%
  peak_rss           1.72MB ± 89.9KB    1.57MB … 1.84MB          0 ( 0%)        ⚡- 13.2% ±  1.1%
  cpu_cycles          219K  ± 12.2K      208K  …  293K          29 (16%)        ⚡- 27.2% ±  1.8%
  instructions        205K  ± 19.4       205K  …  205K           0 ( 0%)        ⚡- 36.9% ±  0.0%
  cache_references   13.8K  ±  171      13.3K  … 14.4K           5 ( 3%)        ⚡- 24.9% ±  0.2%
  cache_misses       5.70K  ±  107      5.32K  … 6.13K           7 ( 4%)        ⚡- 17.6% ±  0.4%
  branch_misses      3.33K  ± 41.2      3.15K  … 3.39K          13 ( 7%)        ⚡- 22.0% ±  0.2%

SafeRefCounted uses malloc: https://dlang.org/library/std/typecons/safe_ref_counted.html (ldc2 get better performance instead dmd or gdc 14.x)
What software was used for the bench? poop

cc: @tjpalmer

Thanks! What optimizations did you use, by the way? It seemed at a glance with with heavy optimizations and allowing inlining that much of the work (including counting) was avoided in some cases. I was just avoiding that for purposes of looking at core semantics. I'm not sure how much can be easily optimized in realistic, complicated software.

And was your C++ fix just using parens instead of curlies? I didn't diff so far. Just looked at it.

What optimizations did you use, by the way?

# similar to -O2 or -O3
ldmd2 -O -release structmem.d -of=structmem_d # optional add -betterC forcing no DruntimeGC
clang++ -O3 strucmem.cpp -o structmem_cpp
rust - cargo build --release

And was your C++ fix just using parens instead of curlies? I didn't diff so far. Just looked at it.

clang get error on:

$ clang++ -O3 structmem.cpp -o structmem_cpp -std=c++23
structmem.cpp:34:13: error: no viable constructor or deduction guide for deduction of template arguments of 'initializer_list'
   34 |             std::initializer_list{6, 7}
      |             ^
/usr/bin/../lib64/gcc/x86_64-pc-linux-gnu/14.1.1/../../../../include/c++/14.1.1/initializer_list:60:17: note: candidate template ignored: could not match 'const_iterator' (aka 'const _E *') against 'int'
   60 |       constexpr initializer_list(const_iterator __a, size_type __l)
      |                 ^
/usr/bin/../lib64/gcc/x86_64-pc-linux-gnu/14.1.1/../../../../include/c++/14.1.1/initializer_list:45:11: note: candidate function template not viable: requires 1 argument, but 2 were provided
   45 |     class initializer_list
      |           ^~~~~~~~~~~~~~~~
/usr/bin/../lib64/gcc/x86_64-pc-linux-gnu/14.1.1/../../../../include/c++/14.1.1/initializer_list:64:17: note: candidate function template not viable: requires 0 arguments, but 2 were provided
   64 |       constexpr initializer_list() noexcept
      |                 ^
1 error generated.

Thanks much for the info and for the analysis! Out of curiosity, does ldc or clang do LTO with those settings, or would it matter?

does ldc or clang do LTO with those settings, or would it matter?

ldc2 have -flto=full or -flto=thin , similar to clang.
http://johanengelen.github.io/ldc/2016/11/10/Link-Time-Optimization-LDC.html

Note: When using clang you can also experiment with highlighting C++ABI. Results may differ between libstdc++ (gnu abi) and libc++ (llvm abi). -stdlib=libname

Thanks again!

After comment in D community about result w/o betterC. It confirms that without betterC flag SafeRefCounted(T) & RefCounted(T) use the GC to handle malloc/free.

https://discord.com/channels/242094594181955585/625407836473524246/1240702707032002702

See:

• https://github.com/dlang/phobos/blob/1a8d2eae06f0d37dd137517f51115dfc79276b0b/std/typecons.d#L7308-L7315
• https://github.com/dlang/phobos/blob/1a8d2eae06f0d37dd137517f51115dfc79276b0b/std/typecons.d#L7368-L7393

Mind if I link this gist from the video?

Mind if I link this gist from the video?

Feel free to share it.