Add documentation for ColsRef macro

Author: Avaneesh-axiom

crates/circuits/primitives/derive/Cargo.toml

@@ -12,7 +12,7 @@ license.workspace = true
 proc-macro = true
 
 [dependencies]
-syn = { version = "2.0", features = ["parsing"] }
+syn = { version = "2.0", features = ["parsing", "extra-traits"] }
 quote = "1.0"
 itertools = { workspace = true, default-features = true }
 proc-macro2 = "1.0"

crates/circuits/primitives/derive/src/cols_ref/README.md

@@ -0,0 +1,112 @@
+# ColsRef macro
+
+The `ColsRef` procedural macro is used in constraint generation to create column structs that have dynamic sizes.
+
+Note: this macro was originally created for use in the SHA-2 VM extension, where we reuse the same constraint generation code for three different circuits (SHA-256, SHA-512, and SHA-384).
+See the [SHA-2 VM extension](../../../../../../extensions/sha2/circuit/src/sha2_chip/air.rs) for an example of how to use the `ColsRef` macro to reuse constraint generation code over multiple circuits.
+
+## Overview
+
+As an illustrative example, consider the following columns struct:
+```rust
+struct ExampleCols<T, const N: usize> {
+    arr: [T; N],
+    sum: T,
+}
+```
+Let's say we want to constrain `sum` to be the sum of the elements of `arr`, and `N` can be either 5 or 10.
+We can define a trait that stores the config parameters.
+```rust
+pub trait ExampleConfig {
+    const N: usize;
+}
+```
+and then implement it for the two different configs.
+```rust
+pub struct ExampleConfigImplA;
+impl ExampleConfig for ExampleConfigImplA {
+    const N: usize = 5;
+}
+pub struct ExampleConfigImplB;
+impl ExampleConfig for ExampleConfigImplB {
+    const N: usize = 10;
+}
+```
+Then we can use the `ColsRef` macro like this
+```rust
+#[derive(ColsRef)]
+#[config(ExampleConfig)]
+struct ExampleCols<T, const N: usize> {
+    arr: [T; N],
+    sum: T,
+}
+```
+which will generate a columns struct that uses references to the fields.
+```rust
+struct ExampleColsRef<'a, T, const N: usize> {
+    arr: &'a [T; N],
+    sum: &'a T,
+}
+```
+The `ColsRef` macro will also generate a `from` method that takes a slice of the correct length and returns an instance of the columns struct.
+The `from` method is parameterized by a struct that implements the `ExampleConfig` trait, and it uses the associated constants to determine how to split the input slice into the fields of the columns struct.
+
+So, the constraint generation code can be written as
+```rust
+impl<AB: InteractionBuilder, C: ExampleConfig> Air<AB> for ExampleAir<C> {
+    fn eval(&self, builder: &mut AB) {
+        let main = builder.main();
+        let (local, _) = (main.row_slice(0), main.row_slice(1));
+        let local_cols = ExampleColsRef::<AB::Var>::from::<C>(&local[..C::N + 1]);
+        let sum = local_cols.arr.iter().sum();
+        builder.assert_eq(local_cols.sum, sum);
+    }
+}
+```
+Notes:
+- the `arr` and `sum` fields of `ExampleColsRef` are references to the elements of the `local` slice.
+- the name, `N`, of the const generic parameter must match the name of the associated constant `N` in the `ExampleConfig` trait.
+
+The `ColsRef` macro also generates a `ExampleColsRefMut` struct that stores mutable references to the fields, for use in trace generation.
+
+The `ColsRef` macro supports more than just variable-length array fields.
+The field types can also be:
+- any type that derives `AlignedBorrow` via `#[derive(AlignedBorrow)]`
+- any type that derives `ColsRef` via `#[derive(ColsRef)]`
+- (possibly nested) arrays of `T` or (possibly nested) arrays of a type that derives `AlignedBorrow`
+
+Note that we currently do not support arrays of types that derive `ColsRef`.
+
+## Specification
+
+Annotating a struct named `ExampleCols` with `#[derive(ColsRef)]` and `#[config(ExampleConfig)]` produces two structs, `ExampleColsRef` and `ExampleColsRefMut`.
+- we assume `ExampleCols` has exactly one generic type parameter, typically named `T`, and any number of const generic parameters. Each const generic parameter must have a name that matches an associated constant in the `ExampleConfig` trait
+
+The fields of `ExampleColsRef` have the same names as the fields of `ExampleCols`, but their types are transformed as follows:
+- type `T` becomes `&T`
+- type `[T; LEN]` becomes `&ArrayView1<T>` (see [ndarray](https://docs.rs/ndarray/latest/ndarray/index.html)) where `LEN` is an associated constant in `ExampleConfig`
+    - the `ExampleColsRef::from` method will correctly infer the length of the array from the config
+- fields with names that end in `Cols` are assumed to be a columns struct that derives `ColsRef` and are transformed into the appropriate `ColsRef` type recursively
+    - one restriction is that any nested `ColsRef` type must have the same config as the outer `ColsRef` type
+- fields that are annotated with `#[aligned_borrow]` are assumed to derive `AlignedBorrow` and are borrowed from the input slice. The new type is a reference to the `AlignedBorrow` type
+- nested arrays of `U` become `&ArrayViewX<U>` where `X` is the number of dimensions in the nested array type
+    - `U` can be either the generic type `T` or a type that derives `AlignedBorrow`. In the latter case, the field must be annotated with `#[aligned_borrow]`
+    - the `ArrayViewX` type provides a `X`-dimensional view into the row slice 
+
+The fields of `ExampleColsRefMut` are almost the same as the fields of `ExampleColsRef`, but they are mutable references.
+- the `ArrayViewMutX` type is used instead of `ArrayViewX` for the array fields.
+- fields that derive `ColsRef` are transformed into the appropriate `ColsRefMut` type recursively.
+
+Each of the `ExampleColsRef` and `ExampleColsRefMut` types has the following methods implemented:
+```rust
+// Takes a slice of the correct length and returns an instance of the columns struct.
+pub const fn from<C: ExampleConfig>(slice: &[T]) -> Self;
+// Returns the number of cells in the struct
+pub const fn width<C: ExampleConfig>() -> usize;
+```
+Note that the `width` method on both structs returns the same value.
+
+Additionally, the `ExampleColsRef` struct has a `from_mut` method that takes a `ExampleColsRefMut` and returns a `ExampleColsRef`.
+This may be useful in trace generation to pass a `ExampleColsRefMut` to a function that expects a `ExampleColsRef`.
+
+See the [tests](../../tests/test_cols_ref.rs) for concrete examples of how the `ColsRef` macro handles each of the supported field types.

crates/circuits/primitives/derive/src/cols_ref.rs renamed to crates/circuits/primitives/derive/src/cols_ref/mod.rs

@@ -1,3 +1,5 @@
+extern crate proc_macro;
+
 use itertools::Itertools;
 use quote::{format_ident, quote};
 use syn::{parse_quote, DeriveInput};
@@ -36,16 +38,21 @@ pub fn cols_ref_impl(
 
     match data {
         syn::Data::Struct(data_struct) => {
+            // Process the fields of the struct, transforming the types for use in ColsRef struct
             let const_field_infos: Vec<FieldInfo> = data_struct
                 .fields
                 .iter()
                 .map(|f| get_const_cols_ref_fields(f, generic_type, &const_generics))
                 .collect::<Result<Vec<_>, String>>()
                 .map_err(|e| format!("Failed to process fields. {}", e))?;
 
+            // The ColsRef struct is named by appending `Ref` to the struct name
             let const_cols_ref_name = syn::Ident::new(&format!("{}Ref", ident), ident.span());
+
+            // the args to the `from` method will be different for the ColsRef and ColsRefMut structs
             let from_args = quote! { slice: &'a [#generic_type] };
 
+            // Package all the necessary information to generate the ColsRef struct
             let struct_info = StructInfo {
                 name: const_cols_ref_name,
                 vis: vis.clone(),
@@ -56,20 +63,27 @@ pub fn cols_ref_impl(
                 derive_clone: true,
             };
 
+            // Generate the ColsRef struct
             let const_cols_ref_struct = make_struct(struct_info.clone(), &config);
 
+            // Generate the `from_mut` method for the ColsRef struct
             let from_mut_impl = make_from_mut(struct_info, &config)?;
 
+            // Process the fields of the struct, transforming the types for use in ColsRefMut struct
             let mut_field_infos: Vec<FieldInfo> = data_struct
                 .fields
                 .iter()
                 .map(|f| get_mut_cols_ref_fields(f, generic_type, &const_generics))
                 .collect::<Result<Vec<_>, String>>()
                 .map_err(|e| format!("Failed to process fields. {}", e))?;
 
+            // The ColsRefMut struct is named by appending `RefMut` to the struct name
             let mut_cols_ref_name = syn::Ident::new(&format!("{}RefMut", ident), ident.span());
+
+            // the args to the `from` method will be different for the ColsRef and ColsRefMut structs
             let from_args = quote! { slice: &'a mut [#generic_type] };
 
+            // Package all the necessary information to generate the ColsRefMut struct
             let struct_info = StructInfo {
                 name: mut_cols_ref_name,
                 vis,
@@ -80,6 +94,7 @@ pub fn cols_ref_impl(
                 derive_clone: false,
             };
 
+            // Generate the ColsRefMut struct
             let mut_cols_ref_struct = make_struct(struct_info, &config);
 
             Ok(quote! {
@@ -103,6 +118,12 @@ struct StructInfo {
     derive_clone: bool,
 }
 
+// Generate the ColsRef and ColsRefMut structs, depending on the value of `struct_info`
+// This function is meant to reduce code duplication between the code needed to generate the two structs
+// Notable differences between the two structs are:
+//   - the types of the fields
+//   - ColsRef derives Clone, but ColsRefMut cannot (since it stores mutable references)
+//   - the `from` method parameter is a reference to a slice for ColsRef and a mutable reference to a slice for ColsRefMut
 fn make_struct(struct_info: StructInfo, config: &proc_macro2::Ident) -> proc_macro2::TokenStream {
     let StructInfo {
         name,
@@ -147,7 +168,6 @@ fn make_struct(struct_info: StructInfo, config: &proc_macro2::Ident) -> proc_mac
                 }
             }
 
-            // TODO: make this return the size in bytes (to support fields of constant size)
             // returns number of cells in the struct (where each cell has type T)
             pub const fn width<C: #config>() -> usize {
                 0 #( + #length_exprs )*
@@ -156,6 +176,7 @@ fn make_struct(struct_info: StructInfo, config: &proc_macro2::Ident) -> proc_mac
     }
 }
 
+// Generate the `from_mut` method for the ColsRef struct
 fn make_from_mut(
     struct_info: StructInfo,
     config: &proc_macro2::Ident,
@@ -183,21 +204,25 @@ fn make_from_mut(
             let is_array = matches!(f.ty, syn::Type::Array(_));
 
             if is_array {
+                // calling view() on ArrayViewMut returns an ArrayView
                 Ok(quote! {
                     other.#ident.view()
                 })
             } else if derives_aligned_borrow {
+                // implicitly converts a mutable reference to an immutable reference, so leave the field value unchanged
                 Ok(quote! {
                     other.#ident
                 })
             } else if is_columns_struct(&f.ty) {
                 // lifetime 'b is used in from_mut to allow more flexible lifetime of return value
                 let cols_ref_type =
                     get_const_cols_ref_type(&f.ty, &generic_type, parse_quote! { 'b });
+                // Recursively call `from_mut` on the ColsRef field
                 Ok(quote! {
                     <#cols_ref_type>::from_mut::<C>(&other.#ident)
                 })
             } else if is_generic_type(&f.ty, &generic_type) {
+                // implicitly converts a mutable reference to an immutable reference, so leave the field value unchanged
                 Ok(quote! {
                     &other.#ident
                 })
@@ -230,6 +255,8 @@ fn make_from_mut(
     })
 }
 
+// Information about a field that is used to generate the ColsRef and ColsRefMut structs
+// See the `make_struct` function to see how this information is used
 #[derive(Debug, Clone)]
 struct FieldInfo {
     // type for struct definition
@@ -257,13 +284,8 @@ fn get_const_cols_ref_fields(
         .iter()
         .any(|attr| attr.path().is_ident("aligned_borrow"));
 
-    let has_plain_array_attribute = f.attrs.iter().any(|attr| attr.path().is_ident("array"));
     let is_array = matches!(f.ty, syn::Type::Array(_));
 
-    if has_plain_array_attribute && !is_array {
-        panic!("field marked with `plain_array` attribute must be an array");
-    }
-
     if is_array {
         let ArrayInfo { dims, elem_type } = get_array_info(&f.ty, const_generics);
         debug_assert!(
@@ -286,33 +308,7 @@ fn get_const_cols_ref_fields(
             })
             .collect_vec();
 
-        if has_plain_array_attribute {
-            Err("unsupported currently".to_string())
-            /*
-            debug_assert!(
-                dims.len() == 1,
-                "field marked with `plain_array` attribute must be a 1D array"
-            );
-
-            let length_expr = quote! {
-                1 #(* #dim_exprs)*
-            };
-
-            Ok(FieldInfo {
-                ty: parse_quote! {
-                    & #f.ty
-                },
-                length_expr: length_expr.clone(),
-                prepare_subslice: quote! {
-                    let (#slice_var, slice) = slice.split_at(#length_expr);
-                    let #slice_var = #slice_var.try_into().unwrap();
-                },
-                initializer: quote! {
-                    #slice_var
-                },
-            })
-            */
-        } else if derives_aligned_borrow {
+        if derives_aligned_borrow {
             let length_expr = quote! {
                 <#elem_type>::width() #(* #dim_exprs)*
             };
@@ -552,6 +548,8 @@ fn get_mut_cols_ref_fields(
     }
 }
 
+// Helper functions
+
 fn is_columns_struct(ty: &syn::Type) -> bool {
     if let syn::Type::Path(type_path) = ty {
         type_path
@@ -666,6 +664,8 @@ fn get_elem_type(ty: &syn::Type) -> syn::Type {
     }
 }
 
+// Get a vector of the dimensions of the array
+// Each dimension is either a constant generic or a literal integer value
 fn get_dims(ty: &syn::Type, const_generics: &[&syn::Ident]) -> Vec<Dimension> {
     get_dims_impl(ty, const_generics)
         .into_iter()

crates/circuits/primitives/derive/src/lib.rs

@@ -402,8 +402,8 @@ pub fn bytes_stateful_derive(input: TokenStream) -> TokenStream {
     }
 }
 
-#[proc_macro_derive(ColsRef, attributes(aligned_borrow, config, plain_array))]
-pub fn cols_ref(input: TokenStream) -> TokenStream {
+#[proc_macro_derive(ColsRef, attributes(aligned_borrow, config))]
+pub fn cols_ref_derive(input: TokenStream) -> TokenStream {
     let derive_input: DeriveInput = parse_macro_input!(input as DeriveInput);
 
     let config = derive_input