std.ArrayHashMap: base linear_scan_max on cache line size

[andrewrk]

I have a hunch that this will be a better default. Will post perf data points shortly.

[AndrewKraevskii]

Why not? It seems to produce identical result

Suggested change

	const linear_scan_max = @as(comptime_int, @max(1, @as(comptime_int, @min(
	std.atomic.cache_line / @as(comptime_int, @max(1, @sizeOf(Hash))),
	std.atomic.cache_line / @as(comptime_int, @max(1, @sizeOf(K))),
	std.atomic.cache_line / @as(comptime_int, @max(1, @sizeOf(V))),
	))));
	const linear_scan_max: comptime_int = @max(1, @min(
	std.atomic.cache_line / @max(1, @sizeOf(Hash)),
	std.atomic.cache_line / @max(1, @sizeOf(K)),
	std.atomic.cache_line / @max(1, @sizeOf(V)),
	));

[andrewrk]

compile error because @max and @min (incorrectly) return a sized integer even when the value is comptime-known

[mlugg]

comptime_int is an awkward type which is fundamentally at odds with the range refinement we want @min/@max to do -- our saviour will hopefully be #3806

[andrewrk]

I disagree; comptime_int is basically a ranged integer with the min and max set to the same value. Strongly believe that @max and @min should return comptime_int when the value is comptime known given that #3806 is not implemented

[mlugg]

That would lead to a major behavioral inconsistency between runtime and comptime.

EDIT: oh, to be fair, I guess @min/@max already has this issue, so disregard that point. TBH, its range refinement is necessarily kind of weird in the absence of #3806.

[andrewrk]

Pushed an update to only consider hash and key sizes since that's what is iterated over when doing linear scans. Performance seems to be insignificant difference compared to status quo.

build zig with itself:

Benchmark 1 (149 runs): master/bin/zig build-exe ../test/standalone/simple/hello_world/hello.zig -target wasm32-wasi -fno-llvm -fno-lld -fno-compiler-rt
  measurement          mean ± σ            min … max           outliers         delta
  wall_time          33.5ms ± 3.91ms    24.3ms … 44.5ms          0 ( 0%)        0%
  peak_rss           94.5MB ±  745KB    93.0MB … 96.6MB          6 ( 4%)        0%
  cpu_cycles         51.4M  ±  888K     48.8M  … 53.6M           3 ( 2%)        0%
  instructions       69.4M  ± 7.09K     69.3M  … 69.4M           1 ( 1%)        0%
  cache_references   3.62M  ± 31.2K     3.54M  … 3.71M           2 ( 1%)        0%
  cache_misses        537K  ± 17.2K      499K  …  581K           0 ( 0%)        0%
  branch_misses       355K  ± 2.75K      347K  …  365K           1 ( 1%)        0%
Benchmark 2 (153 runs): branch/bin/zig build-exe ../test/standalone/simple/hello_world/hello.zig -target wasm32-wasi -fno-llvm -fno-lld -fno-compiler-rt
  measurement          mean ± σ            min … max           outliers         delta
  wall_time          32.7ms ± 3.85ms    24.9ms … 43.4ms          0 ( 0%)          -  2.2% ±  2.6%
  peak_rss           94.4MB ±  670KB    92.7MB … 96.7MB          3 ( 2%)          -  0.1% ±  0.2%
  cpu_cycles         51.5M  ±  968K     49.5M  … 55.3M           7 ( 5%)          +  0.3% ±  0.4%
  instructions       69.3M  ± 7.11K     69.3M  … 69.4M           1 ( 1%)          -  0.1% ±  0.0%
  cache_references   3.61M  ± 33.7K     3.52M  … 3.73M           1 ( 1%)          -  0.4% ±  0.2%
  cache_misses        532K  ± 17.7K      495K  …  583K           1 ( 1%)          -  0.8% ±  0.7%
  branch_misses       357K  ± 3.67K      349K  …  379K           2 ( 1%)          +  0.7% ±  0.2%

wasm linker hello world (wasm linker uses a lot of array hash map):

Benchmark 1 (148 runs): master/bin/zig build-exe ../test/standalone/simple/hello_world/hello.zig -target wasm32-wasi -fno-llvm -fno-lld -fno-compiler-rt
  measurement          mean ± σ            min … max           outliers         delta
  wall_time          33.7ms ± 4.61ms    22.3ms … 43.4ms          0 ( 0%)        0%
  peak_rss           94.4MB ±  676KB    92.9MB … 96.2MB          3 ( 2%)        0%
  cpu_cycles         51.2M  ±  985K     48.7M  … 53.9M           6 ( 4%)        0%
  instructions       69.4M  ± 6.37K     69.3M  … 69.4M           1 ( 1%)        0%
  cache_references   3.63M  ± 30.4K     3.57M  … 3.73M           3 ( 2%)        0%
  cache_misses        537K  ± 15.1K      505K  …  589K           2 ( 1%)        0%
  branch_misses       355K  ± 2.80K      350K  …  365K           2 ( 1%)        0%
Benchmark 2 (153 runs): branch/bin/zig build-exe ../test/standalone/simple/hello_world/hello.zig -target wasm32-wasi -fno-llvm -fno-lld -fno-compiler-rt
  measurement          mean ± σ            min … max           outliers         delta
  wall_time          32.8ms ± 4.13ms    22.2ms … 44.9ms          0 ( 0%)          -  2.8% ±  2.9%
  peak_rss           94.4MB ±  689KB    92.7MB … 96.1MB          2 ( 1%)          -  0.0% ±  0.2%
  cpu_cycles         51.5M  ±  972K     49.2M  … 55.8M           2 ( 1%)          +  0.5% ±  0.4%
  instructions       69.3M  ± 6.05K     69.3M  … 69.3M           0 ( 0%)          -  0.1% ±  0.0%
  cache_references   3.61M  ± 30.1K     3.54M  … 3.71M           2 ( 1%)          -  0.5% ±  0.2%
  cache_misses        532K  ± 17.7K      504K  …  582K           1 ( 1%)          -  0.9% ±  0.7%
  branch_misses       357K  ± 3.02K      350K  …  367K           2 ( 1%)          +  0.6% ±  0.2%

I'm going to keep the change because it feels good 🌈