FastRandomContext and InsecureRandomContext, and then removes the ad-hoc Shuffle functions, now that it appears that standard library std::shuffle has comparable performance. The other reason for keeping Shuffle, namely the fact that libstdc++ used self-move (which debug mode panics on) has been fixed as well (see #29625 (review)).
The following sections might be updated with supplementary metadata relevant to reviewers and maintainers.
For detailed information about the code coverage, see the test coverage report.
See the guideline for information on the review process.
| Type | Reviewers |
|---|---|
| ACK | hodlinator, dergoegge, achow101 |
If your review is incorrectly listed, please react with 👎 to this comment and the bot will ignore it on the next update.
I’ve also benchmarked Deniel Lemire’s nearly divisionless random integer generation (which libstdc++’s std::shuffle switched to), for randrange itself, but opted not to include this in this PR.
For InsecureRandomContext it’s unambiguously faster than the current code (but nothing performance-critical relies on InsecureRandomContext::randrange, so far). For FastRandomContext the advantage is there, but it’s more tricky to achieve (it either needs fast 128-bit integer multiplication, or needs specialization for 32-bit randrange where a 64-bit multiply suffices).
12+void BenchRandom_rand32(benchmark::Bench& bench, RNG&& rng) noexcept
13 {
14- FastRandomContext rng(true);
15 bench.run([&] {
16- rng.rand32();
17+ rng._rand32();
98+void InsecureRandom_randbits(benchmark::Bench& bench) { BenchRandom_randbits(bench, InsecureRandomContext(251438)); }
99+void InsecureRandom_randrange100(benchmark::Bench& bench) { BenchRandom_randrange<1000>(bench, InsecureRandomContext(251438)); }
100+void InsecureRandom_randrange1000(benchmark::Bench& bench) { BenchRandom_randrange<1000>(bench, InsecureRandomContext(251438)); }
101+void InsecureRandom_randrange1000000(benchmark::Bench& bench) { BenchRandom_randrange<1000000>(bench, InsecureRandomContext(251438)); }
102+void InsecureRandom_Shuffle100(benchmark::Bench& bench) { BenchRandom_Shuffle<1000>(bench, InsecureRandomContext(251438)); }
103+void InsecureRandom_stdshuffle100(benchmark::Bench& bench) { BenchRandom_stdshuffle<1000>(bench, InsecureRandomContext(251438)); }
251438 seems to be a fine hook for manhole covers. :)
Worth documenting the choice of seed initialization value?
94+
95+void InsecureRandom_rand64(benchmark::Bench& bench) { BenchRandom_rand64(bench, InsecureRandomContext(251438)); }
96+void InsecureRandom_rand32(benchmark::Bench& bench) { BenchRandom_rand32(bench, InsecureRandomContext(251438)); }
97+void InsecureRandom_randbool(benchmark::Bench& bench) { BenchRandom_randbool(bench, InsecureRandomContext(251438)); }
98+void InsecureRandom_randbits(benchmark::Bench& bench) { BenchRandom_randbits(bench, InsecureRandomContext(251438)); }
99+void InsecureRandom_randrange100(benchmark::Bench& bench) { BenchRandom_randrange<1000>(bench, InsecureRandomContext(251438)); }
<100>?
97+void InsecureRandom_randbool(benchmark::Bench& bench) { BenchRandom_randbool(bench, InsecureRandomContext(251438)); }
98+void InsecureRandom_randbits(benchmark::Bench& bench) { BenchRandom_randbits(bench, InsecureRandomContext(251438)); }
99+void InsecureRandom_randrange100(benchmark::Bench& bench) { BenchRandom_randrange<1000>(bench, InsecureRandomContext(251438)); }
100+void InsecureRandom_randrange1000(benchmark::Bench& bench) { BenchRandom_randrange<1000>(bench, InsecureRandomContext(251438)); }
101+void InsecureRandom_randrange1000000(benchmark::Bench& bench) { BenchRandom_randrange<1000000>(bench, InsecureRandomContext(251438)); }
102+void InsecureRandom_Shuffle100(benchmark::Bench& bench) { BenchRandom_Shuffle<1000>(bench, InsecureRandomContext(251438)); }
InsecureRandom_Shuffle100 combined with <1000> here as well.
std::shuffle controlled/controllable? If we don’t control it maybe it would be better to keep our custom one to ease reproducibility of test failures, despite it being slower.
Also rename the benchmark names to match the operation names
Benchmarks show it is no longer faster with modern standard C++ libraries,
and the debug-mode failure due to self-move has been fixed as well.
Ran benchmarks on da28a26aae3178fb7663efbe20bb650857ace775 with..
0$ src/bench/bench_bitcoin --filter=".*Random.*"
1
2| ns/number | number/s | err% | ins/number | cyc/number | IPC | bra/number | miss% | total | benchmark
3|--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
4| 12.84 | 77,883,564.55 | 0.2% | 103.12 | 47.37 | 2.177 | 15.66 | 4.2% | 0.01 | `FastRandom_Shuffle100`
5| 10.35 | 96,651,871.86 | 0.1% | 116.75 | 38.15 | 3.060 | 9.38 | 0.0% | 0.01 | `FastRandom_rand32`
6| 19.03 | 52,549,720.82 | 0.2% | 214.50 | 70.12 | 3.059 | 15.75 | 0.0% | 0.01 | `FastRandom_rand64`
7| 14.56 | 68,674,993.41 | 0.2% | 149.17 | 53.67 | 2.779 | 14.49 | 1.2% | 0.01 | `FastRandom_randbits`
8| 1.69 | 591,854,532.68 | 0.0% | 11.38 | 6.23 | 1.826 | 1.26 | 0.2% | 0.01 | `FastRandom_randbool`
9| 12.11 | 82,569,289.00 | 0.1% | 95.33 | 44.68 | 2.133 | 13.70 | 4.4% | 0.01 | `FastRandom_randrange100`
10| 13.55 | 73,808,837.05 | 0.6% | 108.46 | 49.99 | 2.170 | 14.63 | 4.4% | 0.01 | `FastRandom_randrange1000`
11| 17.80 | 56,164,658.15 | 0.1% | 156.81 | 65.65 | 2.389 | 18.39 | 3.3% | 0.20 | `FastRandom_randrange1000000`
12| 13.08 | 76,440,558.82 | 0.2% | 134.02 | 48.23 | 2.779 | 12.42 | 0.7% | 0.01 | `FastRandom_stdshuffle100`
13| 7.55 | 132,409,503.40 | 0.9% | 52.83 | 27.85 | 1.897 | 10.39 | 6.3% | 0.01 | `InsecureRandom_Shuffle100`
14| 1.70 | 589,139,570.89 | 0.2% | 19.50 | 6.25 | 3.118 | 1.50 | 0.0% | 0.01 | `InsecureRandom_rand32`
15| 1.11 | 901,858,137.81 | 0.0% | 7.00 | 4.09 | 1.711 | 0.00 | 0.0% | 0.01 | `InsecureRandom_rand64`
16| 2.10 | 475,829,327.28 | 0.4% | 20.89 | 7.75 | 2.697 | 2.48 | 0.0% | 0.01 | `InsecureRandom_randbits`
17| 1.40 | 714,972,427.11 | 0.1% | 12.25 | 5.16 | 2.374 | 1.02 | 0.0% | 0.01 | `InsecureRandom_randbool`
18| 6.46 | 154,870,239.99 | 0.3% | 39.00 | 23.80 | 1.639 | 8.14 | 7.7% | 0.01 | `InsecureRandom_randrange100`
19| 7.08 | 141,237,665.66 | 0.4% | 39.80 | 26.10 | 1.525 | 7.97 | 7.9% | 0.01 | `InsecureRandom_randrange1000`
20| 6.49 | 154,065,899.22 | 0.6% | 43.93 | 23.80 | 1.846 | 7.80 | 7.0% | 0.07 | `InsecureRandom_randrange1000000`
21| 3.61 | 276,805,356.11 | 0.1% | 34.38 | 13.33 | 2.580 | 3.98 | 0.3% | 0.01 | `InsecureRandom_stdshuffle100`
FastRandom - std::shuffle slower by ~2%.
InsecureRandom - std::shuffle is ~2.1x faster, the case for migrating is strong here.
0$ src/bench/bench_bitcoin --filter=".*Random.*"
1
2| ns/number | number/s | err% | ins/number | cyc/number | IPC | bra/number | miss% | total | benchmark
3|--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
4| 10.90 | 91,706,416.37 | 0.4% | 73.17 | 40.22 | 1.819 | 8.91 | 7.6% | 0.01 | `FastRandom_Shuffle100`
5| 10.29 | 97,154,304.49 | 0.1% | 115.31 | 37.96 | 3.037 | 8.94 | 0.0% | 0.01 | `FastRandom_rand32`
6| 18.97 | 52,727,508.94 | 0.4% | 210.63 | 69.79 | 3.018 | 15.88 | 0.0% | 0.01 | `FastRandom_rand64`
7| 11.56 | 86,507,143.42 | 0.5% | 122.79 | 42.62 | 2.881 | 10.55 | 0.4% | 0.01 | `FastRandom_randbits`
8| 1.68 | 594,607,147.44 | 0.1% | 12.32 | 6.20 | 1.987 | 1.25 | 0.2% | 0.01 | `FastRandom_randbool`
9| 9.99 | 100,126,193.36 | 0.4% | 61.84 | 36.83 | 1.679 | 6.91 | 9.1% | 0.01 | `FastRandom_randrange100`
10| 11.50 | 86,963,003.45 | 0.3% | 75.50 | 42.40 | 1.781 | 7.89 | 8.1% | 0.01 | `FastRandom_randrange1000`
11| 15.36 | 65,104,420.98 | 0.4% | 122.93 | 56.09 | 2.192 | 11.41 | 5.2% | 0.17 | `FastRandom_randrange1000000`
12| 14.70 | 68,050,253.74 | 0.0% | 138.64 | 54.21 | 2.558 | 11.02 | 0.1% | 0.01 | `FastRandom_stdshuffle100`
13| 8.32 | 120,167,896.55 | 0.4% | 51.58 | 30.66 | 1.682 | 7.84 | 8.3% | 0.01 | `InsecureRandom_Shuffle100`
14| 1.09 | 920,716,470.48 | 0.0% | 13.00 | 4.01 | 3.244 | 1.00 | 0.0% | 0.01 | `InsecureRandom_rand32`
15| 0.88 | 1,133,795,846.45 | 0.1% | 7.00 | 3.25 | 2.153 | 0.00 | 57.0% | 0.01 | `InsecureRandom_rand64`
16| 1.66 | 601,775,499.66 | 1.2% | 19.86 | 6.11 | 3.251 | 1.98 | 0.0% | 0.01 | `InsecureRandom_randbits`
17| 0.56 | 1,771,130,575.63 | 0.1% | 4.28 | 2.08 | 2.057 | 1.00 | 0.0% | 0.01 | `InsecureRandom_randbool`
18| 6.97 | 143,428,585.12 | 0.2% | 30.93 | 25.70 | 1.203 | 5.84 | 10.5% | 0.01 | `InsecureRandom_randrange100`
19| 7.25 | 137,895,833.08 | 0.2% | 32.18 | 26.75 | 1.203 | 5.78 | 10.6% | 0.01 | `InsecureRandom_randrange1000`
20| 7.37 | 135,732,504.89 | 0.4% | 37.62 | 26.87 | 1.400 | 5.81 | 9.2% | 0.08 | `InsecureRandom_randrange1000000`
21| 5.81 | 172,220,492.06 | 0.1% | 33.82 | 21.42 | 1.579 | 3.09 | 0.3% | 0.01 | `InsecureRandom_stdshuffle100`
FastRandom - std::shuffle shows a performance loss of 26%. :/
InsecureRandom - std::shuffle wins with ~1.4x speed multiplier.
(Side note: std::shuffle seems much better than Shuffle when it comes to branch misses in this configuration (already significantly better in Clang)).
0$ src/bench/bench_bitcoin --filter=".*Random.*"
1Warning, results might be unstable:
2* DEBUG defined
3
4Recommendations
5* Make sure you compile for Release
6
7| ns/number | number/s | err% | ins/number | cyc/number | IPC | bra/number | miss% | total | benchmark
8|--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
9| 96.61 | 10,350,962.32 | 0.4% | 844.51 | 356.24 | 2.371 | 124.47 | 0.5% | 0.01 | `FastRandom_Shuffle100`
10| 182.54 | 5,478,307.47 | 0.0% | 1,703.98 | 673.49 | 2.530 | 204.73 | 0.1% | 0.01 | `FastRandom_rand32`
11| 342.44 | 2,920,180.80 | 0.0% | 3,197.03 | 1,263.38 | 2.531 | 370.47 | 0.1% | 0.01 | `FastRandom_rand64`
12| 192.05 | 5,207,050.93 | 0.1% | 1,756.26 | 708.80 | 2.478 | 213.54 | 0.3% | 0.01 | `FastRandom_randbits`
13| 16.01 | 62,471,637.02 | 0.1% | 148.23 | 59.00 | 2.513 | 23.84 | 0.1% | 0.01 | `FastRandom_randbool`
14| 87.24 | 11,461,986.84 | 0.6% | 758.66 | 321.66 | 2.359 | 108.08 | 0.6% | 0.01 | `FastRandom_randrange100`
15| 110.16 | 9,077,576.97 | 0.5% | 965.24 | 405.83 | 2.378 | 132.06 | 0.5% | 0.01 | `FastRandom_randrange1000`
16| 186.39 | 5,365,223.88 | 0.1% | 1,676.54 | 687.82 | 2.437 | 215.45 | 0.3% | 2.05 | `FastRandom_randrange1000000`
17| 215.54 | 4,639,483.16 | 0.2% | 1,906.59 | 794.16 | 2.401 | 229.95 | 0.1% | 0.01 | `FastRandom_stdshuffle100`
18| 53.91 | 18,549,333.04 | 0.3% | 457.13 | 198.59 | 2.302 | 79.37 | 0.8% | 0.01 | `InsecureRandom_Shuffle100`
19| 22.03 | 45,393,295.79 | 0.1% | 215.50 | 81.27 | 2.652 | 31.50 | 0.0% | 0.01 | `InsecureRandom_rand32`
20| 22.72 | 44,019,773.42 | 0.0% | 220.00 | 83.84 | 2.624 | 24.00 | 0.0% | 0.01 | `InsecureRandom_rand64`
21| 27.54 | 36,307,742.75 | 0.6% | 244.77 | 101.11 | 2.421 | 37.63 | 1.0% | 0.01 | `InsecureRandom_randbits`
22| 10.95 | 91,358,394.49 | 0.1% | 101.70 | 40.40 | 2.518 | 18.42 | 0.1% | 0.01 | `InsecureRandom_randbool`
23| 44.06 | 22,693,987.66 | 0.3% | 371.40 | 162.65 | 2.283 | 63.01 | 0.9% | 0.01 | `InsecureRandom_randrange100`
24| 45.33 | 22,061,595.98 | 0.2% | 382.82 | 167.30 | 2.288 | 64.26 | 0.9% | 0.01 | `InsecureRandom_randrange1000`
25| 50.17 | 19,933,844.35 | 0.2% | 438.63 | 184.96 | 2.372 | 71.38 | 0.8% | 0.55 | `InsecureRandom_randrange1000000`
26| 52.96 | 18,880,998.91 | 0.2% | 418.53 | 195.25 | 2.144 | 56.76 | 0.0% | 0.01 | `InsecureRandom_stdshuffle100`
FastRandom - std::shuffle is less than half the speed. :/ A lot more instructions & branches.
InsecureRandom - Negligible difference.
InsecureRandom is only used in tests, so this would probably be a win for developers/testers/fuzzing but not really for end users. Maybe other systems/configurations show a clearer win.
Also ran make check & test/functional/test_runner.py on 6ecda04fefad980872c72fba89844393f5581120 in GCC debug without any failures. Confirming that the self-assign issue with std::shuffle doesn’t appear to be an issue. (Haven’t tried to create a configuration that triggers the panic though).
@hodlinator FWIW, my motivation here is just FastRandomContext non-debug performance. I did notice that std::shuffle is a bit slower than the custom Shuffle on my system too (AMD 5950X, GCC 13.2) , but not the huge (5x-10x) slowdown it once was. With 2% or 26% or even 50% slowdown, I don’t think it’s worth keeping a custom function around (I don’t think any of the Shuffle uses are that performance critical). Things like randbool and randrange are probably more relevant overall.
I do expect InsecureRandomContext will get used for non-testing purposes eventually, but that’s not a huge motivation now.
To explain what we’re seeing here, worked out for a shuffle of 100 elements:
Shuffle would use randrange 100 times, which for the ranges involved would invoke rand64 on average ~13.005 times (~195.433 times for 1000, ~416003.423 times for 1000000).std::shuffle would invoke rand64 ~100 times directly.In FastRandomContext, the cost of a rand64 is relevatively high, while for InsecureRandomContext it is very low. On the other hand, randrange itself has some overhead too. For FastRandomContext this randrange overhead is worth it as it is compensated by less of the expensive rand64 calls, while for InsecureRandomContext the reduction in rand64 cost is negligible.
If we really wanted, it would be possible to use the technique that std::shuffle uses for converting rand64s directly into integers of arbitrary range, in InsecureRandomContext::randrange, but it making it efficient would need a few different variants for different compilers/architectures. As there is no use case for efficient InsecureRandomContext::randrange right now that doesn’t seem worth it, but that can change.
For posterity, this is sufficient to make 32-bit InsecureRandomContext::randrange work significantly faster:
0+
1+ template<std::integral I>
2+ I randrange(I range) noexcept
3+ {
4+ if constexpr (std::numeric_limits<I>::digits <= 32) {
5+ uint32_t x = rand32();
6+ uint64_t m = uint64_t{x} * uint32_t(range);
7+ uint32_t l = uint32_t(m);
8+ if (l < uint32_t(range)) {
9+ uint32_t t = (0 - range) % range;
10+ while (l < t) {
11+ x = rand32();
12+ m = uint64_t{x} * uint32_t(range);
13+ l = uint32_t(m);
14+ }
15+ }
16+ return m >> 32;
17+ }
18+ return RandomMixin<InsecureRandomContext>::randrange(range);
19+ }
ACK 6ecda04fefad980872c72fba89844393f5581120
Thanks for taking the time to add valuable context!! (I read up a bit, understand you implemented a 32-bit version of Daniel Lemire’s nearly divisionless algo. Nice).
I did notice that
std::shuffleis a bit slower than the custom Shuffle on my system too (AMD 5950X, GCC 13.2)
Maybe worth adding some nuance to the PR summary which currently says “now that it appears that standard library std::shuffle beats it.”.
rand64() invocations
Shufflewould userandrange100 times, which for the ranges involved would invokerand64on average ~13.005 times
Agreed, verified with added instrumentation (rough reasoning through: range of 100 > 0b1100100 > 7 binary digits wide. 64 bits / 7 bits => ~9 randbits() calls per rand64()… with some extra randbits() because of if (ret <= maxval) in randrange() failing => 100 / 9 + extra = ~13).
std::shufflewould invokerand64~100 times directly.
nit: With the added instrumentation my recent testing shows it to be =50 times - as the ranges are small enough std::shuffle() will perform 2 swaps for each rand64(). This goes for both Clang & GCC.
As there is no use case for efficient
InsecureRandomContext::randrangeright now that doesn’t seem worth it, but that can change.
Seems like it could make fuzz-tests more efficient even now and hence free up resources for more test coverage. But agree it’s probably minor.
I’m also on GCC 13.2. CPU: Intel 9th gen Coffee Lake. Did some test with bigger ranges and performance seems to converge between Shuffle & std::shuffle for FastRandom.
0| ns/number | number/s | err% | ins/number | cyc/number | IPC | bra/number | miss% | total | benchmark
1|--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
2| 12.01 | 83,247,932.44 | 0.4% | 86.95 | 44.32 | 1.962 | 9.88 | 6.8% | 0.01 | `FastRandom_Shuffle1000`
3| 14.61 | 68,427,974.21 | 0.1% | 138.35 | 53.91 | 2.566 | 10.95 | 0.0% | 0.01 | `FastRandom_stdshuffle1000`
4| 9.65 | 103,602,289.99 | 0.3% | 52.55 | 35.61 | 1.475 | 7.78 | 7.9% | 0.01 | `InsecureRandom_Shuffle1000`
5| 5.67 | 176,274,334.46 | 0.0% | 33.53 | 20.92 | 1.603 | 3.01 | 0.0% | 0.01 | `InsecureRandom_stdshuffle1000`
6|--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
7| 13.99 | 71,471,645.86 | 0.4% | 107.03 | 51.56 | 2.076 | 11.44 | 6.2% | 0.01 | `FastRandom_Shuffle10000`
8| 14.68 | 68,116,488.28 | 0.1% | 138.32 | 54.15 | 2.554 | 10.94 | 0.0% | 0.01 | `FastRandom_stdshuffle10000`
9| 8.60 | 116,299,489.79 | 0.1% | 55.08 | 31.70 | 1.738 | 7.92 | 8.3% | 0.01 | `InsecureRandom_Shuffle10000`
10| 5.66 | 176,764,575.63 | 0.1% | 33.50 | 20.87 | 1.606 | 3.00 | 0.0% | 0.01 | `InsecureRandom_stdshuffle10000`
11|--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
12| 15.24 | 65,617,271.52 | 0.3% | 123.32 | 56.05 | 2.200 | 12.64 | 5.0% | 0.02 | `FastRandom_Shuffle100000`
13| 15.44 | 64,762,393.26 | 0.1% | 138.31 | 56.89 | 2.431 | 10.94 | 0.0% | 0.02 | `FastRandom_stdshuffle100000`
14| 8.41 | 118,884,436.01 | 0.3% | 56.50 | 30.95 | 1.826 | 7.93 | 7.4% | 0.01 | `InsecureRandom_Shuffle100000`
15| 5.79 | 172,767,668.95 | 0.0% | 33.50 | 21.35 | 1.569 | 3.00 | 0.0% | 0.01 | `InsecureRandom_stdshuffle100000`
16|--------------------:|--------------------:|--------:|----------------:|----------------:|-------:|---------------:|--------:|----------:|:----------
17| 16.44 | 60,831,272.72 | 0.4% | 134.57 | 60.60 | 2.220 | 13.41 | 4.4% | 0.18 | `FastRandom_Shuffle1000000`
18| 16.49 | 60,636,239.08 | 0.0% | 138.31 | 60.82 | 2.274 | 10.94 | 0.0% | 0.18 | `FastRandom_stdshuffle1000000`
19| 8.41 | 118,931,462.65 | 0.2% | 56.74 | 31.00 | 1.830 | 7.81 | 6.9% | 0.09 | `InsecureRandom_Shuffle1000000`
20| 6.23 | 160,482,706.30 | 0.6% | 33.50 | 22.98 | 1.458 | 3.00 | 0.0% | 0.07 | `InsecureRandom_stdshuffle1000000`
For fuzzing, just to keep it in mind, there is (obviously) no guarantee by the standard as to how std::shuffle behaves. See also https://en.cppreference.com/w/cpp/algorithm/random_shuffle#Notes
Note that the implementation is not dictated by the standard, so even if you use exactly the same RandomFunc or URBG (Uniform Random Number Generator) you may get different results with different standard library implementations.
So going forward, I guess one can only assume to surely reproduce a fuzz crash by using the exact same stdlib (and compiler). For context there is also a compiler behavior mismatch discussed in #29043.
So going forward, I guess one can only assume to surely reproduce a fuzz crash by using the exact same stdlib (and compiler). For context there is also a compiler behavior mismatch discussed in #29043.
This is a strong counter-point. I guess other stdlib-differences like std::unordered_map implementations already made this true though.
I don’t think there are any std::shuffle in the fuzz tests, though (apart from one in the random module’s tests itself).
It may make sense to add something like the old Shuffle back if we do need it, though.
I don’t think there are any
std::shufflein the fuzz tests, though?
I hope all code touched in this pull request is covered by fuzz tests. If not, then that should be fixed ;)
For example, joinpsbts is an RPC (not in the fuzz dir), but it is fuzzed, and it calls std::shuffle: https://maflcko.github.io/b-c-cov/fuzz.coverage/src/rpc/rawtransaction.cpp.gcov.html#L1793
I don’t have a strong feeling either way, because if an effort is made to have uniform fuzz behavior across compilers and stdlibs, it will probably be too hard.
I hope all code touched in this pull request is covered by fuzz tests. If not, then that should be fixed ;)
Hmm, for some reason I only considered calls to Shuffle / std::shuffle in the fuzz tests themselves. But you’re right, shuffles in the code being tested would also result in fuzz inconsistency between platforms.