- a * b + c is two roundings on every processor and --cpu: the IR has no fast-math flags, and clang
compiles it with -ffp-contract=on (checked in the assembly: never fused, also with FMA hardware)
- fma(a, b, c) / v.fma(b, c): one rounding, always (llvm.fma)
- muladd(a, b, c) / v.muladd(b, c): the fastest way here (llvm.fmuladd), what v.fma was before
- f32_bits / f32_from_bits / f64_bits / f64_from_bits: bitcasts, also of vectors
- f16_to_f32 / f32_to_f16 (LLVM half), bf16_to_f32 / f32_to_bf16 (integer ops, ties to even, NaN kept)
- docs (en, pl): what is guaranteed about floats, the new built-ins
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Psc42moPLwP43Cmb9oZqSR
- fNxM / iNxM / uNxM up to 512 bits, as LLVM vectors (<8 x float>): LLVM picks SSE, AVX2,
AVX-512 or NEON for them
- f32x8(x) in every lane, f32x8(a, ..., h), f32x8(list, i) loads, f32x8(i32x8 v) converts
- + - * / and the bit operators lane by lane, with a plain number on either side counting for every lane;
v[i], v[i] = x, v.len, == / != (every lane), print and str
- methods: sum min max dot abs sqrt floor ceil round fma store; min/max/abs/clamp lane by lane
- wide vectors (and structs holding them) are loaded and stored with align 16, what malloc and the stack give
- --cpu <name> (and cpu = "..." in finch.toml): native, x86-64-v3, apple-m1, ...
- debug info shows vectors; the language server completes their methods; Kate and VS Code highlight them
- f32 x = -1.5 works: a minus in front of a float written in the code keeps it such a number
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Psc42moPLwP43Cmb9oZqSR
