C++17 bindings (uses "modern" data types for option,result,resource)

Cargo.toml

@@ -41,6 +41,7 @@ wasm-compose = "0.234.0"
 
 wit-bindgen-core = { path = 'crates/core', version = '0.42.1' }
 wit-bindgen-c = { path = 'crates/c', version = '0.42.1' }
+wit-bindgen-cpp = { path = 'crates/cpp', version = '0.42.1' }
 wit-bindgen-rust = { path = "crates/rust", version = "0.42.1" }
 wit-bindgen-csharp = { path = 'crates/csharp', version = '0.42.1' }
 wit-bindgen-markdown = { path = 'crates/markdown', version = '0.42.1' }
@@ -57,6 +58,7 @@ clap = { workspace = true, features = ['wrap_help'] }
 wit-bindgen-core = { workspace = true }
 wit-bindgen-rust = { workspace = true, features = ['clap'], optional = true }
 wit-bindgen-c = { workspace = true, features = ['clap'], optional = true }
+wit-bindgen-cpp = { workspace = true, features = ['clap'], optional = true }
 wit-bindgen-markdown = { workspace = true, features = ['clap'], optional = true }
 wit-bindgen-moonbit = { workspace = true, features = ['clap'], optional = true }
 wit-bindgen-csharp = { workspace = true, features = ['clap'], optional = true }
@@ -72,10 +74,12 @@ default = [
   'markdown',
   'go',
   'csharp',
+  'cpp',
   'moonbit',
   'async',
 ]
 c = ['dep:wit-bindgen-c']
+cpp = ['dep:wit-bindgen-cpp']
 rust = ['dep:wit-bindgen-rust']
 markdown = ['dep:wit-bindgen-markdown']
 go = []

README.md

@@ -385,6 +385,14 @@ Then, you can generate the bindings for your project:
 wit-bindgen-go generate <path-to-wit-pkg>
 ```
 
+### Guest: C++-17+
+
+The cpp crate contains code to generate C++ code which uses the std types 
+optional, string, string_view, vector, expected to represent generic
+WIT types.
+
+This relies on wasi-SDK for guest compilation.
+
 ### Guest: MoonBit
 
 MoonBit can be compiled to WebAssembly using [its toolchain](https://moonbitlang.com/download):

ci/publish.rs

@@ -19,6 +19,7 @@ use std::time::Duration;
 const CRATES_TO_PUBLISH: &[&str] = &[
     "wit-bindgen-core",
     "wit-bindgen-c",
+    "wit-bindgen-cpp",
     "wit-bindgen-rust",
     "wit-bindgen-csharp",
     "wit-bindgen-markdown",

crates/c/src/lib.rs

@@ -1,4 +1,4 @@
-mod component_type_object;
+pub mod component_type_object;
 
 use anyhow::Result;
 use heck::*;
@@ -3864,7 +3864,7 @@ impl Source {
     }
 }
 
-fn wasm_type(ty: WasmType) -> &'static str {
+pub fn wasm_type(ty: WasmType) -> &'static str {
     match ty {
         WasmType::I32 => "int32_t",
         WasmType::I64 => "int64_t",

crates/cpp/Cargo.toml

@@ -0,0 +1,27 @@
+[package]
+name = "wit-bindgen-cpp"
+authors = ["Christof Petig <[email protected]>"]
+version = "0.42.1"
+edition.workspace = true
+repository = 'https://github.com/cpetig/wit-bindgen'
+license = "Apache-2.0 WITH LLVM-exception"
+description = """
+C++ guest and host binding generator for WIT and the component model.
+"""
+
+[lib]
+doctest = false
+test = false
+
+[dependencies]
+wit-bindgen-core = { workspace = true }
+wit-component = { workspace = true }
+wasm-encoder = { workspace = true }
+wasm-metadata = { workspace = true }
+wit-bindgen-c = { workspace = true }
+anyhow = { workspace = true }
+heck = { workspace = true }
+clap = { workspace = true, optional = true }
+
+[dev-dependencies]
+test-helpers = { path = '../test-helpers' }

crates/cpp/DESIGN.md

@@ -0,0 +1,98 @@
+# Type mapping
+
+| Code | Environment |
+| --- | --- |
+| G-- | guest side |
+| H-- | host side |
+| -I- | guest-import (guest calls) |
+| -E- | guest-export (host calls) |
+| --A | argument |
+| --R | result |
+| --S | in struct |
+
+| mode | |
+| --- | --- |
+| v | passed by value |
+| t | owernership transferred |
+| p | cabi_post_ cleans up |
+
+| API | | | ABI | |
+| --- | --- | --- | --- | --- |
+| 🕸 | old | | 📘 | canonical |
+| 💎 | new | | 🪞 | symmetric |
+
+| Code | mode | WIT Type | Rust type | C++ Type | Lower | Reason |
+| --- | --- | --- | --- | --- | --- | --- |
+| GIA | v | string | &str[^1] | string_view (17) | addr, len | |
+| | | list | &[T] | wit::span [^5] | addr, len | |
+| | | tuple | (...) | std::tuple | 0, 1, ...| |
+| | | tuple<string, list> | (&str, &[T]) | std::tuple<...> | a,l,a,l |
+| | | record{string, list} | &T | T const& | a,l,a,l |
+| | | large-struct (>16 args) | &T | T const& | &t |
+| | | result<string,list> | Result<&str, &[]> | std::expected<string_view, span> | d,a,l |
+| | | option\<string> | Option\<&str> | optional<string_view> const& | d,a,l|
+| | | list\<resrc> | &[\&Resrc]? | vector<string_view> const& | a,l|
+| GIR | t | string | String | wit::string[^2] | &(addr, len) [^8] | |
+| | | list | Vec | wit::vector | &(a,l) |
+| | | result<string,list> | Result<String, Vec> | std::expected<wit::string, wit::vector> | &(d,a,l) |
+| GEA | t | string | String | 🕸 wit::string&& | addr, len |
+| | | | | 💎 string_view | |
+| | | result<string,list> | Result<String, Vec> | 🕸 std::expected<wit::string, wit::vector>&& | d,a,l |
+| | | | | 💎 std::expected<string_view, wit::span> | |
+| GER | p | string | String | wit::string (or std?) | 📘 -> &(a,l) cabi_post_N:P/I#F [^7] |
+| | | | | | 🪞 &(a,l) |
+| | | result<string,list> | Result<String, Vec> | std::expected<wit::string, wit::vector> | 📘 -> &(d,a,l) cabi_post |
+| --S | ? | string | String | wit::string | addr, len |
+| HIA | v | string | | string_view | a,l |
+| HIR | t | string | | wit::string[^3] | &(a,l) |
+| HEA | t | string | | 🕸 wit::string[^4] | a,l |
+| | | | | 💎 string_view [^6] | |
+| HER | p | string | | 🕸 wit::guest_owned<string_view> | 📘 -> &(a,l) |
+| | | | | 💎 wit::string [^6] | 🪞 &(a,l) |
+
+[^1]: The host never frees memory (is never passed ownership)!
+
+[^2]: A wit::string is identical to the canonical representation, so it can be part of structures. On the guest a wit::string owns the memory and frees it after use.
+On the host a wit::string can be constructed(=allocated) with an exec_env argument. Thus, without an exec_env a wit::string on the host is inaccessible.
+Complex (non-POD) struct elements on the host will need exec_env to decode or construct.
+
+[^3]: A wit::string requires exec_env inside the host implementation. ~~Perhaps a flexible type (either std::string or wit::string would be possible), or make this a generation option?~~ std::string requires a copy, wit::string requires passing exec_env to the method (which is necessary for methods anyway).
+
+[^4]: A host side wit::string doesn't own the data (not free in dtor), thus no move semantics.
+
+[^5]: std::span requires C++-20, this alias should give minimal functionality with older compiler targets.
+
+[^6]: Not implemented, for now symmetric is priority
+
+[^7]: Here the callee (guest) allocates the memory for the set on its side
+
+[^8]: Caller passes address of the return object as argument
+
+## [Symmetric ABI](https://github.com/WebAssembly/component-model/issues/386)
+
+The idea is to directly connect (link) components to each other.
+
+Thus imported and exported functions and resources need to be compatible
+at the ABI level.
+
+For now for functions the guest import convention is used in both directions:
+
+- The imported function ABI is used with the following properties
+
+  - (unchanged) List and string arguments are passed as Views, no free
+    required, lifetime is constrained until the end of the call
+
+  - (unchanged) Owned resources in arguments or results pass ownership
+    to the callee
+
+  - (unchanged) If there are too many (>1) flat results, a local
+    uninitialized ret_area is passed via the last argument
+
+  - (unchanged) Returned objects are owned.
+    For functional safety, i.e. avoiding all
+    allocations in the hot path, the hope is with [#385](https://github.com/WebAssembly/component-model/issues/385).
+
+- The imported resource ABI is used also for exporting
+  with one modification:
+
+   Resource IDs become usize, so you can optimize the resource table away.

crates/cpp/helper-types/wit-common.h

@@ -0,0 +1,75 @@
+#pragma once
+
+#include <assert.h>
+#include <map>
+#include <optional>
+#include <stddef.h> // size_t
+#include <stdint.h>
+#if __cplusplus > 202001L
+#include <span>
+#else
+#include <vector>
+#endif
+
+namespace wit {
+#if __cplusplus > 202001L
+using std::span;
+#else
+/// Minimal span (vector view) implementation for older C++ environments
+template <class T> class span {
+  T const *address;
+  size_t length;
+
+public:
+  T const *data() const { return address; }
+  size_t size() const { return length; }
+
+  typedef T const *const_iterator;
+
+  const_iterator begin() const { return address; }
+  const_iterator end() const { return address + length; }
+  bool empty() const { return !length; }
+  T const &operator[](size_t index) const { return address[index]; }
+  span(T *a, size_t l) : address(a), length(l) {}
+  // create from any compatible vector (borrows data!)
+  template <class U>
+  span(std::vector<U> const &vec) : address(vec.data()), length(vec.size()) {}
+};
+#endif
+
+/// @brief Helper class to map between IDs and resources
+/// @tparam R Type of the Resource
+template <class R> class ResourceTable {
+  static std::map<int32_t, R> resources;
+
+public:
+  static R *lookup_resource(int32_t id) {
+    auto result = resources.find(id);
+    return result == resources.end() ? nullptr : &result->second;
+  }
+  static int32_t store_resource(R &&value) {
+    auto last = resources.rbegin();
+    int32_t id = last == resources.rend() ? 0 : last->first + 1;
+    resources.insert(std::pair<int32_t, R>(id, std::move(value)));
+    return id;
+  }
+  static std::optional<R> remove_resource(int32_t id) {
+    auto iter = resources.find(id);
+    std::optional<R> result;
+    if (iter != resources.end()) {
+      result = std::move(iter->second);
+      resources.erase(iter);
+    }
+    return std::move(result);
+  }
+};
+
+/// @brief Replaces void in the error position of a result
+struct Void {};
+
+template<class To, class From>
+constexpr To bit_cast(const From& from) noexcept {
+  union Bits { From from; To to; };
+  Bits b; b.from = from; return b.to;
+}
+} // namespace wit