pulley-interpreter 28.0.0

The Pulley interpreter, its bytecode definition, encoder, decoder, and etc...
Documentation

A Bytecode Alliance project

About

Pulley is a portable bytecode and fast interpreter for use in Wasmtime.

Pulley's primary goal is portability and its secondary goal is fast interpretation.

Pulley is not intended to be a simple reference interpreter, support dynamically switching to just-in-time compiled code, or even to be the very fastest interpreter in the world.

For more details on Pulley's motivation, goals, and non-goals, see the Bytecode Alliance RFC that originally proposed Pulley.

Status

Pulley is very much still a work in progress! Expect the details of the bytecode to change, instructions to appear and disappear, and APIs to be overhauled.

Example

Here is the disassembly of f(a, b) = a + b in Pulley today:

       0: 2f                              push_frame
       1: 12 00 04                        xadd32 x0, x0, x1
       4: 30                              pop_frame
       5: 00                              ret

Note that there are a number of things that could be improved here:

  • We could avoid allocating and deallocating a stack frame because this function's body doesn't use any stack slots.

As mentioned above, Pulley is very much a work in progress.

Principles

What follows are some general, incomplete, and sometimes-conflicting principles that we try and follow when designing the Pulley bytecode format and its interpreter:

  • The bytecode should be simple and fast to decode in software. For example, we should avoid overly-complicated bitpacking, and only reach for that kind of thing when benchmarks and profiles show it to be of benefit.

  • The interpreter never materializes enum Instruction { .. } values. Instead, it decodes immediates and operands as needed in each opcode handler. This avoids constructing unnecessary temporary storage and branching on opcode multiple times.

  • Because we never materialize enum Instruction { .. } values, we don't have to worry about unused padding or one very-large instruction inflating the size of all the rest of our small instructions. To put it concisely: we can lean into a variable-length encoding where some instructions require only a single byte and others require many. This helps keep the bytecode compact and cache-efficient.

  • We lean into defining super-instructions (sometimes called "macro ops") that perform the work of multiple operations in a single instruction. The more work we do in each turn of the interpreter loop the less we are impacted by its overhead. Additionally, Cranelift, as the primary Pulley bytecode producer, can leverage ISLE lowering patterns to easily identify opportunities for emitting super-instructions.

  • We do not, in general, define sub-opcodes. There should be only one branch, on the initial opcode, when evaluating any given instruction. For example, we do not have a generic load instruction that is followed by a sub-opcode to discriminate between different addressing modes. Instead, we have many different kinds of load instructions, one for each of our addressing modes.

    The one exception is the split between regular and extended ops. Regular ops are a single u8 opcode, where 255 is reserved for all extended ops, and a u16 opcode follows after the 255 regular opcode. This keeps the most common instructions extra small, and provides a pressure release valve for defining an unbounded number of additional, colder, ops.

  • We strive to cut down on boilerplate as much as possible, and try to avoid matching on every opcode repeatedly throughout the whole code base. We do this via heavy macro_rules usage where we define the bytecode inside a higher-order macro and then automatically derive a disassembler, decoder, encoder, etc... from that definition. This also avoids any kind of drift where the encoder and decoder get out of sync with each other, for example.