Revision history for RoundingFiasco

1.0.0.0 -- 2025-11-13

• theme: this release focuses on easy performance improvements
• breaking API changes:
  • sign functions renamed
    • i32_arithmic_signum_f64 to i32_trinary_signum_f64 (spelling, align with WebAssembly proposal)
    • i32_arithmic_signum_f64# to i32_trinary_signum_f64# (spelling, align with WebAssembly proposal)
    • i32_arithmic_signum_f32 to i32_trinary_signum_f32 (spelling, align with WebAssembly proposal)
    • i32_arithmic_signum_f32# to i32_trinary_signum_f32# (spelling, align with WebAssembly proposal)
    • i32_sign_bit_f64 to i32_binary_signum_f64 (align with the trinary_signum naming)
    • i32_sign_bit_f64# to i32_binary_signum_f64# (align with the trinary_signum naming)
    • i32_sign_bit_f32 to i32_binary_signum_f32 (align with the trinary_signum naming)
    • i32_sign_bit_f32# to i32_binary_signum_f32# (align with the trinary_signum naming)
  • * operators renamed
    • f32_multiplicate_ceil to f32_multiply_ceil (vocabulary)
    • f64_multiplicate_ceil to f64_multiply_ceil (vocabulary)
    • f32_multiplicate_floor to f32_multiply_floor (vocabulary)
    • f64_multiplicate_floor to f64_multiply_floor (vocabulary)
    • f32_multiplicate_truncate to f32_multiply_truncate (vocabulary)
    • f64_multiplicate_truncate to f64_multiply_truncate (vocabulary)
• minimum required GHC version:
  • bump to 9.10.3
    • remove the deprecation warnings of usage of: stgDoubleToWord64, stgFloatToWord32, stgWord64ToDouble, and castWord32ToFloat#
    • while avoiding preprocessor macros that would conflict with a possible future Safe Haskell adoption
• performance:
  • remove superfluous sanitizing check in f32_predecessorIEEE#
  • manually inline low-level word literals
  • aggressive inlining throughout the codebase
    • performance improvement seen by more than 10%
  • replace debug error messages for initial development with undefined so that they do not increase the binaries with strings that are never supposed to be used
  • the reason for the major A-version bump (see Package Versioning Policy (PVP)) is to indicate that every performance improvement can have unintended effects on user usecases especially for performance critical numerical code
• testing:
  • add promotion and demotion tests for NaN handling
  • add more thorough test cases for successorIEEE and predecessorIEEE regarding negative zero
• fix typos and add more documentation
• forward compatibility:
  • explicitly import from GHC.Exts, GHC.Float, GHC.Int, and Data.Word instead of implicitly

0.1.0.0 -- 2024-04-08

• theme: this release focuses on initial software implementation and its correctness
• floats: single, double
• rounding variants: ceil, floor, truncate
• operators: +-*/√
• conversions: promotion, demotion
• helpers: arithmic_signum, sign_bit, successor, predecessor