philihp / pcg / 26025512348

100%

Build Type

push

github

Committed by web-flow

Commit Message

perf: ~7× faster randomInt/nextState/randomList hot path (#328)

* perf: hot-path optimizations for randomInt/nextState/randomList

The dominant consumer of this library — fast-shuffle — calls randomInt
in a tight loop, where each call goes through nextState. Three changes
remove most of the per-call overhead:

- nextState now does the simple PCG step (state * multiplier + increment)
  directly, instead of dispatching through stepState's curry + jump-ahead
  bookkeeping loop. stepState(1) was running a 1-iteration loop while
  allocating a handful of extra Longs each call.
- randomInt no longer wraps the output number in Long inside the
  rejection loop; the threshold and modulo are done with plain JS
  arithmetic. Values fit in safe-integer range so semantics are
  preserved.
- randomList replaces ramda.scan with a direct for loop, avoiding
  the iterator + accumulator allocations.

Microbenchmarks (Node, 1M iters where applicable):

  nextState              381k -> 3,301k ops/s  (8.7x)
  randomInt(0, 100)      323k -> 2,230k ops/s  (6.9x)
  shuffle-style 100 ints 6.0k -> 44.1k ops/s   (7.3x)
  randomList(100)        5.9k -> 41.6k ops/s   (7.1x)

All existing tests pass with identical output bits.

* refactor: drop long dependency in favor of native bigint

Replaces the long runtime dependency with native BigInt for all 64-bit
arithmetic. Masking with 0xFFFFFFFFFFFFFFFFn after each multiplication or
addition keeps values in the uint64 range. Output bits are bit-identical
to the previous Long-based implementation across all four output
functions and all three stream schemes.

API-wise, the LongLike alias is preserved as a public type name and is
now `bigint | number | string` (which is what the previous Long-based
implementation effectively accepted). PCGState.state, PCGConfig.multiplier,
and PCGConfig.increment are now `bigint` instead of `Long`.

Benchmarks (Node, microbench vs original Long-based main):

  nextState              381k -> 2,865k ops/... (continued)

Coverage Stats

24 of 32 branches covered (75.0%)

Branch coverage included in aggregate %.

48 of 53 new or added lines in 3 files covered. (90.57%)

1 existing line in 1 file now uncovered.

89 of 95 relevant lines covered (93.68%)

4855.35 hits per line