Answers are grounded in this paper and cite their sources.
From the earliest interpreters—IVSYS (1965) and APL/360 (1966), both written in FORTRAN IV—to the present day, APL implementations have consistently interposed a software layer between the notation and the processor. Dyalog APL compiles to bytecode; GNU APL interprets an AST. Even Co-dfns, an APL compiler written in APL that generates native CPU and GPU code, retains an intermediate representation: it parses source into an IR, optimises the IR, then emits machine code, and is tied to the Dyalog dialect. This paper takes a different view: APL glyphs are themselves mnemonics for vector processor instructions, and the underlying architecture is irrelevant so long as it provides SIMD extensions. Each glyph corresponds to 1–5 native SIMD instructions. We present μAPL, a JIT compiler for an APL subset on x86-64/AVX-512 that demonstrates direct symbol-to-instruction translation with no intermediate representation—no AST, no bytecode, no IR. The prototype supports 10 operators (8 vector primitives plus 2 meta-operators), branching, user-defined functions, nested arrays, and a peel+tail mechanism designed for arbitrary-length vectors. We further describe an architecture for porting μAPL to SPARC V9/VIS—and by extension to ARM SVE/SVE2 and RISC-V V—and outline a three-stage bootstrap strategy: a minimal assembler written in the APL subset compiles a larger subset, which in turn compiles a full, architecture-independent APL environment. APL can serve as a macro-assembler for vector processors; direct translation is feasible, works across architectures, and eliminates vendor lock-in.
No takes yet. Share an insight, caveat, or question.
Lex Shilov (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: