Distinguish four support states
Flyology_SIMD.Features exposes the compiled and runtime-available backend states to applications. Project support also requires executed differential and code-generation evidence.
| Backend | Implemented | Compiled evidence | Executed evidence | CI configured |
|---|---|---|---|---|
| Scalar fallback | Yes | Darwin AArch64 and Linux x86-64 | Differential and AddressSanitizer | Linux x86-64, macOS AArch64 |
| AArch64 NEON, full 128-bit family | Yes | Darwin AArch64 | Differential and assembly audit | macOS AArch64 |
| x86-64 SSE2, full 128-bit family | Yes | Linux x86-64 | Differential, AddressSanitizer, assembly audit | Linux x86-64 |
| x86-64 AVX2 algorithms | Yes | Linux x86-64 | Differential, feature-gated execution, assembly audit | Linux x86-64 with runtime gate |
The public CI workflow is the authority for the current hosted result. A configured target becomes continuously tested only after its hosted job completes successfully.
Limit the compiler claim
Executed evidence uses GNAT FSF 16.1.0. GCC-based GNAT is the initial required compiler family. The project does not advertise unbuilt GNAT versions.
GNAT LLVM is an explicit compatibility target. It is not an implemented or tested backend because the current evidence does not verify equivalent intrinsic and assembly lowering.
Use the complete AArch64 family
The AArch64 backend implements all 128-bit signed, unsigned, and IEEE floating vector families with Advanced SIMD where the target supplies a suitable instruction.
Documented scalar composition preserves semantics for operations without a direct instruction or stable leaf. These include 64-bit integer multiplication, selected reductions, mask selection, and unordered floating comparison cases.
Advanced SIMD is architecturally available on AArch64. Available therefore does not require a separate NEON hardware probe.
Keep SSE2 baseline-safe
x86-64 uses SSE2 as its baseline. SSE2 handles arithmetic, bitwise operations, shifts, comparisons, compact masks, selection, shuffles, and full memory operations across the complete family.
Scalar composition preserves exact semantics where SSE2 lacks a direct operation. This includes wide saturation, ordered reductions, and floating number minimum and maximum.
AVX2 is optional and isolated. The runtime gate verifies AVX, OSXSAVE, XMM and YMM state in XCR0, and the AVX2 CPUID bit before entering the optional object.
Require checks without concealed failures
The workflow runs checked and release builds, deterministic differential tests, AddressSanitizer tail checks, and focused code-generation checks. The matrix includes scalar-only, AArch64 NEON, x86-64 SSE2, and runtime-gated AVX2 configurations.
The workflow has no continue-on-error. A source implementation, cross-compile, local execution, and hosted execution remain separate evidence even when all four concern the same target.
The project remains experimental until the public API, compiler matrix, and backend checks have accumulated broader continuous execution.