feat: introduce experimental packaging format #278

Previous to this commit, we were expecting the rodata segments to be encoded in a specific order, and placed on the advice stack before anything else. With this commit, we now simply expect to the advice map to contain each segment keyed by its commitment hash, which we then move to the advice stack on demand, and immediately pipe to memory. This means the order of the segments no longer matters, and the advice stack is not sensitive to codegen changes or other influences which might perturb the advice stack or otherwise disrupt our assumptions. It also sets the stage for us to be able to initialize rodata after a context switch, as at that point the advice stack will be in an unknown condition, and using the advice map gives us certainty that we can arrange to have exactly what we need on the advice stack, when we need it. Additionally, I've updated the `midenc debug` input config file, as well as the usage documentation to reflect this. The last related change to this, will be emitting the rodata segments to disk in a convenient form, so that when the compiler emits the program, it also emits the segments alongside it, making it convenient to run the debugger against that program (or via the VM directly).

This commit implements a basic package format for Miden programs and libraries, with the metadata needed to simplify using them together. The format is experimental, and can be changed at any time, however it does use a header that lets us identify whether a package was produced with the same specification or not. This also modifies the compiler and debugger to emit and consume, respectively, the package format as their primary artifact. The debugger uses this, for example, to initialize the VM with the necessary libraries and rodata segments required by the program. The package format is described using the structures in `codegen/masm/src/packaging/package.rs`, and is encoded using a tiny, but efficient binary format from the `bitcode` crate, using `serde` to facilitate the lowering of our high-level types to that format. This made it relatively straightforward to implement the format without having to mess around with the low-level byte representation. The `bitcode` crate is designed with stability in mind, though because we are using it via `serde`, the stability is subject to being violated if `serde` changes something that causes `bitcode` to serialize a type differently. For now, this is a non-issue, longer term we might want to actually implement `bitcode::{Encode, Decode}` for our types if we want to gain more stability.

Some recent changes caused the `midenc compile` command to emit no outputs when `--emit` was not provided. Also fixed: * Moved some unnecessary debug logs to trace level * Added more debug logging to the compiler to aid in troubleshooting * Fixed serialization of packages involving more complex inputs which were hitting an issue with fields marked as skippable via serde; this is not supported via bitcode, so those attributes were removed

…ompatible

It turns out we were assuming that Rust was laying out its memory using the default LLVM layout, but rustc actually specifies a larger shadow stack. We were not properly handling the defined memory in Wasm modules parsed by the frontend, and so we ended up incorrectly placing global variables and other data in the middle of memory that rust was using for either the shadow stack or static data. Because of the size of the shadow stack region reserved by Rust, this was not causing any corruption in our tests, but it was causing other assumptions to be invalid, and miscompilations were the result. This commit treats the Wasm memory for a parsed module as reserved for use by the code in that module, which causes midenc to place any global variables it defines after that region, and computes the start of the heap as immediately following any such data. It does so by propagating the page size and reserved number of pages through the IR and to the backend, taking that information into account as needed.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: introduce experimental packaging format #278

feat: introduce experimental packaging format #278

Commits on Aug 27, 2024