Rewrite as layered formats over one engine 0.1.0 steered behaviour through two enums, `NanStyle` and `SubnormalStyle`, which turned out to be mis-factored: all six subnormal-encode branches carried an `N != FNUZ` fix-up, 10 of 12 `FNUZ` sites were zero-row policy rather than NaN policy, and `Fast`/`Reserved` decoded row 0 with normal semantics while `MIN_EXP` assumed it was subnormal. The enums also could not express the no-NaN OCP MX formats, which is the concrete driver for this change. Replace them with a chain of format policies, each reinterpreting one part of the bit space in terms of the layer below it: `ScientificFormat` reads every magnitude as (1 + m/2**M)·2**(e-B); `FiniteFormat` spends row 0 on zero and subnormals; `IeeeFormat`, `FnFormat`, and `FnuzFormat` each reserve their own code points on top of that. A layer only uses the one below -- static calls and a template parameter, no inheritance and no contained object -- so adding a format is a policy of about forty lines, not a class with thirty members. The engine `Minifloat` consults only orthogonal, format-declared properties, all behind `if constexpr`. Public types become `Finite`, `IEEE`, `FN`, and `FNUZ`, plus aliases for LLVM APFloat's set of formats at 16 bits or fewer. `Finite` is what the MX types need and is the layer the other three build on. `FNUZ` now defaults to a bias one greater than `default_bias(E)`, matching LLVM. The 301 macro-generated typedefs are gone. Three latent bugs surfaced while porting, each verified against 0.1.0 first and now covered by `TestWideExponentRange`: - `to_double()` rebuilt garbage bits for formats whose exponent range exceeds double's; `IEEE<12, 3>::from_bits(2047 << 3)` read as -3.999 rather than 1. Routing the inexact case through `ldexp` fixes it and deletes the bit-hacking fallback entirely. - Constructing such a format from a host zero or subnormal read the zero exponent field as an ordinary exponent, so `IEEE<12, 3>{0.0F}` returned the maximum finite value. - The same expression shifted a negative value left, which is undefined. NaN is now detected with `std::isnan` instead of `x != x` (#3). The header ends at 847 lines, down from 873, and the -O2 hot paths for E4M3, E5M2, E4M3FN, and E4M3FNUZ compile to the same instruction counts as 0.1.0 with no calls. The test suite re-points its checkers at types instead of `` and gains a per-format oracle that checks every bit pattern of 21 types against textbook decoding written independently of the engine. Version becomes 0.2.0-dev; CMake carries no pre-release suffix, so the project version reads 0.2.0 and the changelog records the -dev state. Co-Authored-By: Claude Opus 5 (1M context)

Uh oh!

Use std::isnan(x) instead of x != x for checking if x is NaN #3

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Use `std::isnan(x)` instead of `x != x` for checking if x is NaN #3