[AntonOresten]

Julia 1.14 lets a primitive type declare fewer bits than it stores (JuliaLang/julia#61359), e.g. primitive type Float4 4 end, which occupies a byte and has Core.bitsizeof(Float4) == 4. JuliaLang/julia#62852, also in 1.14, rounds sizeof of such types up to their alignment, so a 24-bit primitive now takes 4 bytes.

Changes

• bitwidth fallback: Core.bitsizeof(T) where it exists (@static if isdefined(Core, :bitsizeof)), else 8 * sizeof(T) as before. These types need no overload, and byte-multiple types like a 24-bit primitive keep their real width; the old fallback reports 32 on 1.14. Bool, Pad, NArray and NarrowTuple keep their own methods.
• raw_bits(x) / from_raw_bits(T, bits) (new, public): reinterpret rejects these types (bitcast requires equal bit sizes), so these zero-extend and truncate them instead, and fall back to reinterpret for everything else.
• NArray, NarrowTuple: use the helpers. On 1.14, NVector and NarrowTuple of such types threw.
• NarrowArray: reinterpret also refuses arrays of these types, so the constructor, copy, T.(narr) and dest .= x route them through raw bits (one temporary). Other types take the same reinterpret path as before. copy now builds concrete SVector{L,T}s instead of map(SArray, …), which keeps it inferable on the new path.

Testing

• test/unused_bits.jl runs only where Core.bitsizeof exists, covering native 4-, 6- and 24-bit types across the scalar helpers, NVector, NarrowTuple and NarrowArray. It uses @inferred because GPU broadcasts need these to infer (an earlier from_raw_bits.(T, x) with a static parameter T did not).
• Local: 274/274 tests pass on 1.14.0-alpha1 and nightly; 229/229 on 1.13 and 1.10.
• GPU (GB10, 1.14, CuArray): construction, copy, T.(narr) and Narrow(T).(d) work for a native 4-bit type, and copy/UInt8.(…) for Bool. I didn't run the opt-in cuTile tests locally; they don't use the changed paths.
• Codegen for byte-sized types is unchanged: the new condition folds at compile time.

Not in this PR

dest .= cuarray into a GPU-backed NarrowArray throws a kernel exception. It does the same on main with Bool: GPUArrays' copyto! takes over and calls setindex! on the NarrowArray. It needs a method that's aware of GPUArrays, so I left it for a separate change.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 81.25000% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.47%. Comparing base (cf69542) to head (54da004).

Files with missing lines	Patch %	Lines
src/NarrowArray.jl	66.66%	5 Missing ⚠️
src/BitPacking.jl	88.88%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #17      +/-   ##
==========================================
- Coverage   98.44%   97.47%   -0.98%     
==========================================
  Files           7        7              
  Lines         578      593      +15     
==========================================
+ Hits          569      578       +9     
- Misses          9       15       +6     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.