Files
prometeu-runtime/docs/vm-arch/ISA_CORE.md
T
bquarkz 2e041266f1
Intrepid/Prometeu/Runtime/pipeline/pr-master This commit looks good
fix(vm): enforce return boundaries and amortize budgeted growth
Validate and complete function-end branches through the shared return path. Handle integer DIV/MOD overflow with wrapping and charge CALL/RET refusals exactly once.

Amortize budgeted vector growth with exact-capacity fallback at quota limits. Add execution and accounting regressions, clarify ISA contracts, and include existing lint cleanups.
2026-10-09 06:35:51 +01:00

10 KiB
Raw Blame History

Prometeu Bytecode — Core ISA

Status: bytecode-level normative

This document defines the stable Core ISA surface for the Prometeu Virtual Machine at the bytecode level. It specifies instruction encoding, the stack evaluation model, and the instruction set currently available in the canonical opcode surface used by encoder, decoder, disassembler, assembler, verifier, and VM execution.

Machine boundary:

  • PROMETEU is not "just the VM". It is the broader fantasy console/handheld machine.
  • This document covers only the bytecode ISA of the VM subsystem embedded in that machine.

Authority rule:

  • This document is normative for bytecode-level encoding and opcode surface.
  • Runtime-wide invariants still live in ARCHITECTURE.md.
  • If a bytecode-level rule here conflicts with runtime architecture, the conflict must be resolved explicitly in both documents; neither should drift silently.

Encoding Rules

  • Endianness: Little‑endian.
  • Instruction layout: [opcode: u16][immediate: spec.imm_bytes].
  • Opcodes are defined in prometeu_bytecode::isa::core::CoreOpCode.
  • Immediate sizes and stack effects are defined by CoreOpCode::spec() returning CoreOpcodeSpec.
  • All jump immediates are absolute u32 byte offsets from the start of the current function.

Stack Machine Model

  • The VM is stack‑based. Unless noted, operands are taken from the top of the operand stack and results are pushed back.
  • Types at the bytecode level are represented by the Value enum; the VM may perform numeric promotion where appropriate (e.g., Int32 + Float -> Float).
  • Stack underflow is a trap (TRAP_STACK_UNDERFLOW).
  • Some operations may trap for other reasons (e.g., division by zero, invalid indices, type mismatches).

Instruction Set (Core)

  • Execution control:

    • NOP — no effect.
    • HALT — terminates execution (block terminator).
    • JMP u32 — unconditional absolute jump (block terminator).
    • JMP_IF_FALSE u32 — pops [bool], jumps if false.
    • JMP_IF_TRUE u32 — pops [bool], jumps if true.
    • TRAP — software trap/breakpoint (block terminator).
  • Stack manipulation:

    • PUSH_CONST u32 — load constant by index → _pushes [value].
    • PUSH_I64 i64, PUSH_F64 f64, PUSH_BOOL u8, PUSH_I32 i32 — push literals.
    • POP — pops 1.
    • POP_N u32 — pops N.
    • DUP — [x] -> [x, x].
    • SWAP — [a, b] -> [b, a].
  • Arithmetic:

    • ADD, SUB, MUL, DIV, MOD — binary numeric ops.
    • NEG — unary numeric negation.
    • Integer arithmetic wraps at the resulting carrier width. For DIV, the signed minimum divided by -1 returns that minimum; for MOD, the same operands return zero. Division or modulo by zero still traps. Mixed Int32/Int64 division promotes to Int64 before the operation.
  • Comparison and logic:

    • EQ, NEQ, LT, LTE, GT, GTE — comparisons → [bool].
    • AND, OR, NOT — boolean logic.
    • BIT_AND, BIT_OR, BIT_XOR, SHL, SHR — integer bit operations.
  • Variables:

    • GET_GLOBAL u32, SET_GLOBAL u32 — access global slots.
    • GET_LOCAL u32, SET_LOCAL u32 — access local slots (current frame).
  • Functions and scopes:

    • CALL u32 — call by function index; argument/result arity per function metadata.
    • RET — return from current function (block terminator).
    • MAKE_CLOSURE u32,u32 — create typed closure from (FunctionId, FnTypeId); cost 8+C.
    • CALL_CLOSURE u32 — invoke against expected FnTypeId; cost 6.
    • GET_CAPTURE u32 (0x57) — copy indexed complete capture group from current frame; cost 2+W.
    • CALL_TRAIT u32,u32 (0x58) — invoke TraitTypeId/method index through complete view; cost 6.
  • Retired execution operations:

    • 0x54–0x56 (SPAWN/YIELD/SLEEP) remain reserved and reject at load, including unreachable code. Diagnostic decoding may identify them; no execution or assembler emission is permitted.
  • System/Timing:

    • HOSTCALL u32 — PBX pre-load host binding call by SYSC table index; the loader must resolve and rewrite it before verification or execution.
    • SYSCALL u32 — final numeric platform call in the executable image; raw SYSCALL in PBX pre-load artifacts is rejected by the loader.
    • INTRINSIC u32 — final numeric VM-owned intrinsic call.
    • FRAME_SYNC — yield until the next frame boundary (e.g., vblank); explicit safepoint.

Host service arity is not encoded in the opcode itself. It is defined by resolved syscall metadata.

Example:

  • asset.load currently resolves with arg_slots = 2 and ret_slots = 2.
  • The canonical stack contract is asset_id, slot -> status, handle.
  • Callers do not provide an explicit asset kind; the runtime derives it from asset_table.
  • composer.bind_scene resolves with arg_slots = 1 and ret_slots = 1.
  • The canonical stack contract is bank_id -> status.
  • composer.emit_sprite resolves with arg_slots = 9 and ret_slots = 1.

Canonical Intrinsic Registry Artifact

  • Final intrinsic IDs and intrinsic stack metadata are published in INTRINSICS.csv.
  • This CSV is the ISA-scoped artifact intended to be consumed by compiler/tooling consumers such as ../studio.
  • Each row defines one canonical intrinsic identity and its final numeric ID.
  • canonical_name is the fully qualified intrinsic identity seen by compiler-side intrinsic pools.
  • arg_slots and ret_slots are the real stack effect contract for verifier/lowering consumers.
  • arg_layout and ret_layout use |-separated ABI atoms:
    • int
    • float
    • bool
    • builtin:<builtin_name>
  • Rows must remain unique by both (canonical_name, canonical_version) and final_id_hex / final_id_dec.
  • Rows must remain deterministically ordered by final ID.

For exact immediates and stack effects, see CoreOpCode::spec() which is the single source of truth used by the decoder, disassembler, and verifier.

Canonical Decoder Contract

  • The canonical decoder is prometeu_bytecode::decode_next(pc, bytes).
  • It uses the Core ISA spec to determine immediate size and the canonical next_pc.
  • Unknown opcodes produce deterministic UnknownOpcode errors; recognized retired opcodes produce structured load rejection before execution.

Module Boundary

  • Core ISA lives under prometeu_bytecode::isa::core and re‑exports:
    • CoreOpCode — the opcode enum of the core profile.
    • CoreOpcodeSpec and CoreOpCodeSpecExt — spec with imm_bytes, stack effects, and flags.
  • Consumers (encoder/decoder/disasm/verifier) should import from this module to avoid depending on internal layout.

Scope Notes

  • "Core ISA" in the current repository means the canonical opcode surface implemented by the runtime today.
  • It includes typed closures, trait calls, HOSTCALL patching semantics, INTRINSIC, and FRAME_SYNC.
  • It does not, by itself, define higher-level runtime policy such as crash taxonomy, firmware behavior, cartridge lifecycle, or host service organization. Those belong to the canonical runtime architecture and related specs.

FRAME_SYNC — Semantics and Placement (Bytecode Level)

  • Semantics:

    • FRAME_SYNC is a zero-operand instruction and does not modify the operand stack.
    • It marks a VM safepoint for threshold GC. In CoreOpcodeSpec this is exposed as spec.is_safepoint == true.
    • Execution returns to the host at the logical frame boundary, preserving the single context. Resumption continues at the next instruction with live frames/locals/environments. Successful last-root RET is a separate GC opportunity after result rooting and completes that entry.
  • Placement rules (representable and checkable):

    • FRAME_SYNC may appear anywhere inside a function body where normal instructions can appear. It is NOT a block terminator (spec.is_terminator == false).
    • Instruction boundaries are canonical: encoders/emitters must only place FRAME_SYNC at valid instruction PCs. The verifier already enforces “jump-to-boundary” and end-exclusive [start, end) function ranges using the canonical layout routine.
    • Entrypoints that represent a render/update loop SHOULD ensure at least one reachable FRAME_SYNC along every long-running path to provide deterministic safepoints for GC/frame handoff. This policy is semantic and may be enforced by higher-level tooling; at the bytecode level it is representable via spec.is_safepoint and can be counted by static analyzers.
  • Disassembly:

    • Disassemblers must print the mnemonic FRAME_SYNC verbatim for this opcode.
    • Tools MAY optionally annotate it as a safepoint in comments, e.g., FRAME_SYNC ; safepoint.
  • Verification notes:

    • The bytecode verifier treats FRAME_SYNC as a normal instruction with no stack effect and no control-flow targets. It is permitted before RET and between basic blocks; a reachable path must still terminate explicitly rather than fall through the function range. Jumps targeting the function end (pc == end) remain valid under the end-exclusive rule. A taken end branch completes the current call with the same result-height/type checks, frame release and root-return safepoint as RET. It consumes only the decoded branch's cycles, without an additional implicit RET charge, and never executes the neighboring function or range gap.

Registry instruction encodings

The Registry contract defines opcodes 0x0090–0x0096. Tooling recognizes their u32 TypeId immediates:

DYN_PACK 0
DYN_UNPACK 0
DYN_CAN_UNPACK 0
REF_IS_EXACT source=4, target=5
REF_IMPLEMENTS source=4, target=6
REF_TO_TRAIT source=4, target=6
REF_CAST source=6, target=4

Dynamic instructions occupy six bytes; reference instructions occupy ten bytes. Static opcode metadata must represent Registry-dependent slot effects as unresolved, never as zero. Resolve the semantic widths and immediate categories from the validated Registry. Encoding/decoding support alone does not authorize execution: a slot-only image/verifier must reject these instructions until typed-image integration is present.

Registry-backed collections

STRUCT_NEW (0x00A0), STRUCT_GET (0x00A1), STRUCT_SET (0x00A2), ARRAY_NEW (0x00A3), ARRAY_LEN (0x00A4), ARRAY_GET (0x00A5) and ARRAY_SET (0x00A6) require a validated Registry. Their exact immediates, full group effects, attempt costs and failure ordering are normative in the typed collection instruction table. These operations have no implicit conversion or safepoint. See the producer handoff for emission oracles.