diff options
Diffstat (limited to 'libbindgen/src/ir')
-rw-r--r-- | libbindgen/src/ir/annotations.rs | 157 | ||||
-rw-r--r-- | libbindgen/src/ir/comp.rs | 871 | ||||
-rw-r--r-- | libbindgen/src/ir/context.rs | 1055 | ||||
-rw-r--r-- | libbindgen/src/ir/enum_ty.rs | 137 | ||||
-rw-r--r-- | libbindgen/src/ir/function.rs | 282 | ||||
-rw-r--r-- | libbindgen/src/ir/int.rs | 93 | ||||
-rw-r--r-- | libbindgen/src/ir/item.rs | 1172 | ||||
-rw-r--r-- | libbindgen/src/ir/item_kind.rs | 114 | ||||
-rw-r--r-- | libbindgen/src/ir/layout.rs | 34 | ||||
-rw-r--r-- | libbindgen/src/ir/mod.rs | 18 | ||||
-rw-r--r-- | libbindgen/src/ir/module.rs | 61 | ||||
-rw-r--r-- | libbindgen/src/ir/ty.rs | 869 | ||||
-rw-r--r-- | libbindgen/src/ir/type_collector.rs | 22 | ||||
-rw-r--r-- | libbindgen/src/ir/var.rs | 246 |
14 files changed, 5131 insertions, 0 deletions
diff --git a/libbindgen/src/ir/annotations.rs b/libbindgen/src/ir/annotations.rs new file mode 100644 index 00000000..58308d6d --- /dev/null +++ b/libbindgen/src/ir/annotations.rs @@ -0,0 +1,157 @@ +//! Types and functions related to bindgen annotation comments. +//! +//! Users can add annotations in doc comments to types that they would like to +//! replace other types with, mark as opaque, etc. This module deals with all of +//! that stuff. + +use clang; + +/// What kind of accessor should we provide for a field? +#[derive(Copy, PartialEq, Clone, Debug)] +pub enum FieldAccessorKind { + /// No accessor. + None, + /// Plain accessor. + Regular, + /// Unsafe accessor. + Unsafe, + /// Immutable accessor. + Immutable, +} + +/// Annotations for a given item, or a field. +#[derive(Clone, PartialEq, Debug)] +pub struct Annotations { + /// Whether this item is marked as opaque. Only applies to types. + opaque: bool, + /// Whether this item should be hidden from the output. Only applies to + /// types. + hide: bool, + /// Whether this type should be replaced by another. The name must be the + /// canonical name that that type would get. + use_instead_of: Option<String>, + /// Manually disable deriving copy/clone on this type. Only applies to + /// struct or union types. + disallow_copy: bool, + /// Whether fields should be marked as private or not. You can set this on + /// structs (it will apply to all the fields), or individual fields. + private_fields: Option<bool>, + /// The kind of accessor this field will have. Also can be applied to + /// structs so all the fields inside share it by default. + accessor_kind: Option<FieldAccessorKind>, +} + +fn parse_accessor(s: &str) -> FieldAccessorKind { + match s { + "false" => FieldAccessorKind::None, + "unsafe" => FieldAccessorKind::Unsafe, + "immutable" => FieldAccessorKind::Immutable, + _ => FieldAccessorKind::Regular, + } +} + +impl Default for Annotations { + fn default() -> Self { + Annotations { + opaque: false, + hide: false, + use_instead_of: None, + disallow_copy: false, + private_fields: None, + accessor_kind: None, + } + } +} + +impl Annotations { + /// Construct new annotations for the given cursor and its bindgen comments + /// (if any). + pub fn new(cursor: &clang::Cursor) -> Option<Annotations> { + let mut anno = Annotations::default(); + let mut matched_one = false; + anno.parse(&cursor.comment(), &mut matched_one); + + if matched_one { Some(anno) } else { None } + } + + /// Should this type be hidden? + pub fn hide(&self) -> bool { + self.hide + } + + /// Should this type be opaque? + pub fn opaque(&self) -> bool { + self.opaque + } + + /// For a given type, indicates the type it should replace. + /// + /// For example, in the following code: + /// + /// ```cpp + /// + /// /** <div rustbindgen replaces="Bar"></div> */ + /// struct Foo { int x; }; + /// + /// struct Bar { char foo; }; + /// ``` + /// + /// the generated code would look something like: + /// + /// ``` + /// /** <div rustbindgen replaces="Bar"></div> */ + /// struct Bar { + /// x: ::std::os::raw::c_int, + /// }; + /// ``` + /// + /// That is, code for `Foo` is used to generate `Bar`. + pub fn use_instead_of(&self) -> Option<&str> { + self.use_instead_of.as_ref().map(|s| &**s) + } + + /// Should we avoid implementing the `Copy` trait? + pub fn disallow_copy(&self) -> bool { + self.disallow_copy + } + + /// Should the fields be private? + pub fn private_fields(&self) -> Option<bool> { + self.private_fields + } + + /// What kind of accessors should we provide for this type's fields? + pub fn accessor_kind(&self) -> Option<FieldAccessorKind> { + self.accessor_kind + } + + fn parse(&mut self, comment: &clang::Comment, matched: &mut bool) { + use clangll::CXComment_HTMLStartTag; + if comment.kind() == CXComment_HTMLStartTag && + comment.get_tag_name() == "div" && + comment.get_tag_attrs() + .next() + .map_or(false, |attr| attr.name == "rustbindgen") { + *matched = true; + for attr in comment.get_tag_attrs() { + match attr.name.as_str() { + "opaque" => self.opaque = true, + "hide" => self.hide = true, + "nocopy" => self.disallow_copy = true, + "replaces" => self.use_instead_of = Some(attr.value), + "private" => { + self.private_fields = Some(attr.value != "false") + } + "accessor" => { + self.accessor_kind = Some(parse_accessor(&attr.value)) + } + _ => {} + } + } + } + + for child in comment.get_children() { + self.parse(&child, matched); + } + } +} diff --git a/libbindgen/src/ir/comp.rs b/libbindgen/src/ir/comp.rs new file mode 100644 index 00000000..d19d1209 --- /dev/null +++ b/libbindgen/src/ir/comp.rs @@ -0,0 +1,871 @@ +//! Compound types (unions and structs) in our intermediate representation. + +use clang; +use parse::{ClangItemParser, ParseError}; +use std::cell::Cell; +use std::cmp; +use super::annotations::Annotations; +use super::context::{BindgenContext, ItemId}; +use super::item::Item; +use super::layout::Layout; +use super::ty::{RUST_DERIVE_IN_ARRAY_LIMIT, Type}; +use super::type_collector::{ItemSet, TypeCollector}; + +/// The kind of compound type. +#[derive(Debug, Copy, Clone, PartialEq)] +pub enum CompKind { + /// A struct. + Struct, + /// A union. + Union, +} + +/// The kind of C++ method. +#[derive(Debug, Copy, Clone, PartialEq)] +pub enum MethodKind { + /// A static method. + Static, + /// A normal method. + Normal, + /// A virtual method. + Virtual, +} + +/// A struct representing a C++ method, either static, normal, or virtual. +#[derive(Debug)] +pub struct Method { + kind: MethodKind, + /// The signature of the method. Take into account this is not a `Type` + /// item, but a `Function` one. + /// + /// This is tricky and probably this field should be renamed. + signature: ItemId, + is_const: bool, +} + +impl Method { + /// Construct a new `Method`. + fn new(kind: MethodKind, signature: ItemId, is_const: bool) -> Self { + Method { + kind: kind, + signature: signature, + is_const: is_const, + } + } + + /// What kind of method is this? + pub fn kind(&self) -> MethodKind { + self.kind + } + + /// Is this a virtual method? + pub fn is_virtual(&self) -> bool { + self.kind == MethodKind::Virtual + } + + /// Is this a static method? + pub fn is_static(&self) -> bool { + self.kind == MethodKind::Static + } + + /// Get the `ItemId` for the `Function` signature for this method. + pub fn signature(&self) -> ItemId { + self.signature + } + + /// Is this a const qualified method? + pub fn is_const(&self) -> bool { + self.is_const + } +} + +/// A struct representing a C++ field. +#[derive(Clone, Debug)] +pub struct Field { + /// The name of the field, empty if it's an unnamed bitfield width. + name: Option<String>, + /// The inner type. + ty: ItemId, + /// The doc comment on the field if any. + comment: Option<String>, + /// Annotations for this field, or the default. + annotations: Annotations, + /// If this field is a bitfield, and how many bits does it contain if it is. + bitfield: Option<u32>, + /// If the C++ field is marked as `mutable` + mutable: bool, +} + +impl Field { + /// Construct a new `Field`. + pub fn new(name: Option<String>, + ty: ItemId, + comment: Option<String>, + annotations: Option<Annotations>, + bitfield: Option<u32>, + mutable: bool) + -> Field { + Field { + name: name, + ty: ty, + comment: comment, + annotations: annotations.unwrap_or_default(), + bitfield: bitfield, + mutable: mutable, + } + } + + /// Get the name of this field. + pub fn name(&self) -> Option<&str> { + self.name.as_ref().map(|n| &**n) + } + + /// Get the type of this field. + pub fn ty(&self) -> ItemId { + self.ty + } + + /// Get the comment for this field. + pub fn comment(&self) -> Option<&str> { + self.comment.as_ref().map(|c| &**c) + } + + /// If this is a bitfield, how many bits does it need? + pub fn bitfield(&self) -> Option<u32> { + self.bitfield + } + + /// Is this field marked as `mutable`? + pub fn is_mutable(&self) -> bool { + self.mutable + } + + /// Get the annotations for this field. + pub fn annotations(&self) -> &Annotations { + &self.annotations + } +} + +/// A compound type. +/// +/// Either a struct or union, a compound type is built up from the combination +/// of fields which also are associated with their own (potentially compound) +/// type. +#[derive(Debug)] +pub struct CompInfo { + /// Whether this is a struct or a union. + kind: CompKind, + + /// The members of this struct or union. + fields: Vec<Field>, + + /// The template parameters of this class. These are non-concrete, and + /// should always be a Type(TypeKind::Named(name)), but still they need to + /// be registered with an unique type id in the context. + template_args: Vec<ItemId>, + + /// The method declarations inside this class, if in C++ mode. + methods: Vec<Method>, + + /// Vector of classes this one inherits from. + base_members: Vec<ItemId>, + + /// The parent reference template if any. + ref_template: Option<ItemId>, + + /// The inner types that were declared inside this class, in something like: + /// + /// class Foo { + /// typedef int FooTy; + /// struct Bar { + /// int baz; + /// }; + /// } + /// + /// static Foo::Bar const = {3}; + inner_types: Vec<ItemId>, + + /// Set of static constants declared inside this class. + inner_vars: Vec<ItemId>, + + /// Whether this type should generate an vtable (TODO: Should be able to + /// look at the virtual methods and ditch this field). + has_vtable: bool, + + /// Whether this type has destructor. + has_destructor: bool, + + /// Whether this type has a base type with more than one member. + /// + /// TODO: We should be able to compute this. + has_nonempty_base: bool, + + /// If this type has a template parameter which is not a type (e.g.: a + /// size_t) + has_non_type_template_params: bool, + + /// Whether this struct layout is packed. + packed: bool, + + /// Whether this struct is anonymous. + is_anonymous: bool, + + /// Used to know if we've found an opaque attribute that could cause us to + /// generate a type with invalid layout. This is explicitly used to avoid us + /// generating bad alignments when parsing types like max_align_t. + /// + /// It's not clear what the behavior should be here, if generating the item + /// and pray, or behave as an opaque type. + found_unknown_attr: bool, + + /// Used to detect if we've run in a can_derive_debug cycle while cycling + /// around the template arguments. + detect_derive_debug_cycle: Cell<bool>, + + /// Used to detect if we've run in a has_destructor cycle while cycling + /// around the template arguments. + detect_has_destructor_cycle: Cell<bool>, +} + +impl CompInfo { + /// Construct a new compound type. + pub fn new(kind: CompKind) -> Self { + CompInfo { + kind: kind, + fields: vec![], + template_args: vec![], + methods: vec![], + base_members: vec![], + ref_template: None, + inner_types: vec![], + inner_vars: vec![], + has_vtable: false, + has_destructor: false, + has_nonempty_base: false, + has_non_type_template_params: false, + packed: false, + is_anonymous: false, + found_unknown_attr: false, + detect_derive_debug_cycle: Cell::new(false), + detect_has_destructor_cycle: Cell::new(false), + } + } + + /// Can we derive the `Debug` trait for this compound type? + pub fn can_derive_debug(&self, + ctx: &BindgenContext, + layout: Option<Layout>) + -> bool { + // We can reach here recursively via template parameters of a member, + // for example. + if self.detect_derive_debug_cycle.get() { + warn!("Derive debug cycle detected!"); + return true; + } + + if self.kind == CompKind::Union { + if ctx.options().unstable_rust { + return false; + } + + let layout = layout.unwrap_or_else(Layout::zero); + let size_divisor = cmp::max(1, layout.align); + return layout.size / size_divisor <= RUST_DERIVE_IN_ARRAY_LIMIT; + } + + self.detect_derive_debug_cycle.set(true); + + let can_derive_debug = { + self.base_members + .iter() + .all(|ty| ctx.resolve_type(*ty).can_derive_debug(ctx)) && + self.template_args + .iter() + .all(|ty| ctx.resolve_type(*ty).can_derive_debug(ctx)) && + self.fields + .iter() + .all(|f| ctx.resolve_type(f.ty).can_derive_debug(ctx)) && + self.ref_template.map_or(true, |template| { + ctx.resolve_type(template).can_derive_debug(ctx) + }) + }; + + self.detect_derive_debug_cycle.set(false); + + can_derive_debug + } + + /// Is this compound type unsized? + pub fn is_unsized(&self, ctx: &BindgenContext) -> bool { + !self.has_vtable(ctx) && self.fields.is_empty() && + self.base_members.iter().all(|base| { + ctx.resolve_type(*base).canonical_type(ctx).is_unsized(ctx) + }) && + self.ref_template + .map_or(true, |template| ctx.resolve_type(template).is_unsized(ctx)) + } + + /// Does this compound type have a destructor? + pub fn has_destructor(&self, ctx: &BindgenContext) -> bool { + if self.detect_has_destructor_cycle.get() { + warn!("Cycle detected looking for destructors"); + // Assume no destructor, since we don't have an explicit one. + return false; + } + + self.detect_has_destructor_cycle.set(true); + + let has_destructor = self.has_destructor || + match self.kind { + CompKind::Union => false, + CompKind::Struct => { + // NB: We can't rely on a type with type parameters + // not having destructor. + // + // This is unfortunate, but... + self.ref_template.as_ref().map_or(false, |t| { + ctx.resolve_type(*t).has_destructor(ctx) + }) || + self.template_args + .iter() + .any(|t| ctx.resolve_type(*t).has_destructor(ctx)) || + self.base_members + .iter() + .any(|t| ctx.resolve_type(*t).has_destructor(ctx)) || + self.fields.iter().any(|field| { + ctx.resolve_type(field.ty) + .has_destructor(ctx) + }) + } + }; + + self.detect_has_destructor_cycle.set(false); + + has_destructor + } + + /// Can we derive the `Copy` trait for this type? + pub fn can_derive_copy(&self, ctx: &BindgenContext, item: &Item) -> bool { + // NOTE: Take into account that while unions in C and C++ are copied by + // default, the may have an explicit destructor in C++, so we can't + // defer this check just for the union case. + if self.has_destructor(ctx) { + return false; + } + + if self.kind == CompKind::Union { + if !ctx.options().unstable_rust { + return true; + } + + // https://github.com/rust-lang/rust/issues/36640 + if !self.template_args.is_empty() || self.ref_template.is_some() || + !item.applicable_template_args(ctx).is_empty() { + return false; + } + } + + // With template args, use a safe subset of the types, + // since copyability depends on the types itself. + self.ref_template + .as_ref() + .map_or(true, |t| ctx.resolve_item(*t).can_derive_copy(ctx)) && + self.base_members + .iter() + .all(|t| ctx.resolve_item(*t).can_derive_copy(ctx)) && + self.fields.iter().all(|field| { + ctx.resolve_item(field.ty) + .can_derive_copy(ctx) + }) + } + + /// Is this type a template specialization? + pub fn is_template_specialization(&self) -> bool { + self.ref_template.is_some() + } + + /// Get the template declaration this specialization is specializing. + pub fn specialized_template(&self) -> Option<ItemId> { + self.ref_template + } + + /// Compute the layout of this type. + /// + /// This is called as a fallback under some circumstances where LLVM doesn't + /// give us the correct layout. + /// + /// If we're a union without known layout, we try to compute it from our + /// members. This is not ideal, but clang fails to report the size for these + /// kind of unions, see test/headers/template_union.hpp + pub fn layout(&self, ctx: &BindgenContext) -> Option<Layout> { + use std::cmp; + + // We can't do better than clang here, sorry. + if self.kind == CompKind::Struct { + return None; + } + + let mut max_size = 0; + let mut max_align = 0; + for field in &self.fields { + let field_layout = ctx.resolve_type(field.ty) + .layout(ctx); + + if let Some(layout) = field_layout { + max_size = cmp::max(max_size, layout.size); + max_align = cmp::max(max_align, layout.align); + } + } + + Some(Layout::new(max_size, max_align)) + } + + /// Get this type's set of fields. + pub fn fields(&self) -> &[Field] { + &self.fields + } + + /// Get this type's set of free template arguments. Empty if this is not a + /// template. + pub fn template_args(&self) -> &[ItemId] { + &self.template_args + } + + /// Does this type have any template parameters that aren't types + /// (e.g. int)? + pub fn has_non_type_template_params(&self) -> bool { + self.has_non_type_template_params + } + + /// Does this type have a virtual table? + pub fn has_vtable(&self, ctx: &BindgenContext) -> bool { + self.has_vtable || + self.base_members().iter().any(|base| { + ctx.resolve_type(*base) + .has_vtable(ctx) + }) || + self.ref_template.map_or(false, |template| { + ctx.resolve_type(template).has_vtable(ctx) + }) + } + + /// Get this type's set of methods. + pub fn methods(&self) -> &[Method] { + &self.methods + } + + /// What kind of compound type is this? + pub fn kind(&self) -> CompKind { + self.kind + } + + /// The set of types that this one inherits from. + pub fn base_members(&self) -> &[ItemId] { + &self.base_members + } + + /// Construct a new compound type from a Clang type. + pub fn from_ty(potential_id: ItemId, + ty: &clang::Type, + location: Option<clang::Cursor>, + ctx: &mut BindgenContext) + -> Result<Self, ParseError> { + use clangll::*; + // Sigh... For class templates we want the location, for + // specialisations, we want the declaration... So just try both. + // + // TODO: Yeah, this code reads really bad. + let mut cursor = ty.declaration(); + let mut kind = Self::kind_from_cursor(&cursor); + if kind.is_err() { + if let Some(location) = location { + kind = Self::kind_from_cursor(&location); + cursor = location; + } + } + + let kind = try!(kind); + + debug!("CompInfo::from_ty({:?}, {:?})", kind, cursor); + + let mut ci = CompInfo::new(kind); + ci.is_anonymous = cursor.is_anonymous(); + ci.template_args = match ty.template_args() { + // In forward declarations and not specializations, + // etc, they are in + // the ast, we'll meet them in + // CXCursor_TemplateTypeParameter + None => vec![], + Some(arg_types) => { + let num_arg_types = arg_types.len(); + + let args = arg_types.filter(|t| t.kind() != CXType_Invalid) + .map(|t| Item::from_ty_or_ref(t, None, None, ctx)) + .collect::<Vec<_>>(); + + if args.len() != num_arg_types { + ci.has_non_type_template_params = true; + warn!("warning: Template parameter is not a type"); + } + + args + } + }; + + ci.ref_template = cursor.specialized() + .and_then(|c| Item::parse(c, None, ctx).ok()); + + let mut maybe_anonymous_struct_field = None; + cursor.visit(|cur| { + if cur.kind() != CXCursor_FieldDecl { + if let Some((ty, _)) = maybe_anonymous_struct_field { + let field = Field::new(None, ty, None, None, None, false); + ci.fields.push(field); + } + maybe_anonymous_struct_field = None; + } + + match cur.kind() { + CXCursor_FieldDecl => { + match maybe_anonymous_struct_field.take() { + Some((ty, clang_ty)) => { + let mut used = false; + cur.visit(|child| { + if child.cur_type() == clang_ty { + used = true; + } + CXChildVisit_Continue + }); + if !used { + let field = Field::new(None, + ty, + None, + None, + None, + false); + ci.fields.push(field); + } + } + None => {} + } + + let bit_width = cur.bit_width(); + let field_type = Item::from_ty_or_ref(cur.cur_type(), + Some(cur), + Some(potential_id), + ctx); + + let comment = cur.raw_comment(); + let annotations = Annotations::new(&cur); + let name = cur.spelling(); + let is_mutable = cursor.is_mutable_field(); + + // Name can be empty if there are bitfields, for example, + // see tests/headers/struct_with_bitfields.h + assert!(!name.is_empty() || bit_width.is_some(), + "Empty field name?"); + + let name = if name.is_empty() { None } else { Some(name) }; + + let field = Field::new(name, + field_type, + comment, + annotations, + bit_width, + is_mutable); + ci.fields.push(field); + + // No we look for things like attributes and stuff. + cur.visit(|cur| { + if cur.kind() == CXCursor_UnexposedAttr { + ci.found_unknown_attr = true; + } + CXChildVisit_Continue + }); + + } + CXCursor_UnexposedAttr => { + ci.found_unknown_attr = true; + } + CXCursor_EnumDecl | + CXCursor_TypeAliasDecl | + CXCursor_TypedefDecl | + CXCursor_StructDecl | + CXCursor_UnionDecl | + CXCursor_ClassTemplate | + CXCursor_ClassDecl => { + let inner = Item::parse(cur, Some(potential_id), ctx) + .expect("Inner ClassDecl"); + if !ci.inner_types.contains(&inner) { + ci.inner_types.push(inner); + } + // A declaration of an union or a struct without name could + // also be an unnamed field, unfortunately. + if cur.spelling().is_empty() && + cur.kind() != CXCursor_EnumDecl { + let ty = cur.cur_type(); + maybe_anonymous_struct_field = Some((inner, ty)); + } + } + CXCursor_PackedAttr => { + ci.packed = true; + } + CXCursor_TemplateTypeParameter => { + // Yes! You can arrive here with an empty template parameter + // name! Awesome, isn't it? + // + // see tests/headers/empty_template_param_name.hpp + if cur.spelling().is_empty() { + return CXChildVisit_Continue; + } + + let default_type = Item::from_ty(&cur.cur_type(), + Some(cur), + Some(potential_id), + ctx) + .ok(); + let param = Item::named_type(cur.spelling(), + default_type, + potential_id, + ctx); + ci.template_args.push(param); + } + CXCursor_CXXBaseSpecifier => { + if !ci.has_vtable { + ci.has_vtable = cur.is_virtual_base(); + } + let type_id = + Item::from_ty(&cur.cur_type(), Some(cur), None, ctx) + .expect("BaseSpecifier"); + ci.base_members.push(type_id); + } + CXCursor_CXXMethod => { + let is_virtual = cur.method_is_virtual(); + let is_static = cur.method_is_static(); + debug_assert!(!(is_static && is_virtual), "How?"); + + if !ci.has_vtable { + ci.has_vtable = is_virtual; + } + + let linkage = cur.linkage(); + if linkage != CXLinkage_External { + return CXChildVisit_Continue; + } + + if cur.access_specifier() == CX_CXXPrivate { + return CXChildVisit_Continue; + } + + let visibility = cur.visibility(); + if visibility != CXVisibility_Default { + return CXChildVisit_Continue; + } + + if cur.is_inlined_function() { + return CXChildVisit_Continue; + } + + let spelling = cur.spelling(); + if spelling.starts_with("operator") { + return CXChildVisit_Continue; + } + + // This used to not be here, but then I tried generating + // stylo bindings with this (without path filters), and + // cried a lot with a method in gfx/Point.h + // (ToUnknownPoint), that somehow was causing the same type + // to be inserted in the map two times. + // + // I couldn't make a reduced test case, but anyway... + // Methods of template functions not only use to be inlined, + // but also instantiated, and we wouldn't be able to call + // them, so just bail out. + if !ci.template_args.is_empty() { + return CXChildVisit_Continue; + } + + // NB: This gets us an owned `Function`, not a + // `FunctionSig`. + let method_signature = + Item::parse(cur, Some(potential_id), ctx) + .expect("CXXMethod"); + + let is_const = cur.method_is_const(); + let method_kind = if is_static { + MethodKind::Static + } else if is_virtual { + MethodKind::Virtual + } else { + MethodKind::Normal + }; + + let method = + Method::new(method_kind, method_signature, is_const); + + ci.methods.push(method); + } + CXCursor_Destructor => { + if cur.method_is_virtual() { + // FIXME: Push to the method list? + ci.has_vtable = true; + } + ci.has_destructor = true; + } + CXCursor_NonTypeTemplateParameter => { + ci.has_non_type_template_params = true; + } + CXCursor_VarDecl => { + let linkage = cur.linkage(); + if linkage != CXLinkage_External && + linkage != CXLinkage_UniqueExternal { + return CXChildVisit_Continue; + } + + let visibility = cur.visibility(); + if visibility != CXVisibility_Default { + return CXChildVisit_Continue; + } + + let item = Item::parse(cur, Some(potential_id), ctx) + .expect("VarDecl"); + ci.inner_vars.push(item); + } + // Intentionally not handled + CXCursor_CXXAccessSpecifier | + CXCursor_CXXFinalAttr | + CXCursor_Constructor | + CXCursor_FunctionTemplate | + CXCursor_ConversionFunction => {} + _ => { + warn!("unhandled comp member `{}` (kind {}) in `{}` ({})", + cur.spelling(), + cur.kind(), + cursor.spelling(), + cur.location()); + } + } + CXChildVisit_Continue + }); + + if let Some((ty, _)) = maybe_anonymous_struct_field { + let field = Field::new(None, ty, None, None, None, false); + ci.fields.push(field); + } + + Ok(ci) + } + + fn kind_from_cursor(cursor: &clang::Cursor) + -> Result<CompKind, ParseError> { + use clangll::*; + Ok(match cursor.kind() { + CXCursor_UnionDecl => CompKind::Union, + CXCursor_ClassDecl | + CXCursor_StructDecl => CompKind::Struct, + CXCursor_CXXBaseSpecifier | + CXCursor_ClassTemplatePartialSpecialization | + CXCursor_ClassTemplate => { + match cursor.template_kind() { + CXCursor_UnionDecl => CompKind::Union, + _ => CompKind::Struct, + } + } + _ => { + warn!("Unknown kind for comp type: {:?}", cursor); + return Err(ParseError::Continue); + } + }) + } + + /// Do any of the types that participate in this type's "signature" use the + /// named type `ty`? + /// + /// See also documentation for `ir::Item::signature_contains_named_type`. + pub fn signature_contains_named_type(&self, + ctx: &BindgenContext, + ty: &Type) + -> bool { + // We don't generate these, so rather don't make the codegen step to + // think we got it covered. + if self.has_non_type_template_params() { + return false; + } + self.template_args.iter().any(|arg| { + ctx.resolve_type(*arg) + .signature_contains_named_type(ctx, ty) + }) + } + + /// Get the set of types that were declared within this compound type + /// (e.g. nested class definitions). + pub fn inner_types(&self) -> &[ItemId] { + &self.inner_types + } + + /// Get the set of static variables declared within this compound type. + pub fn inner_vars(&self) -> &[ItemId] { + &self.inner_vars + } + + /// Have we found a field with an opaque type that could potentially mess up + /// the layout of this compound type? + pub fn found_unknown_attr(&self) -> bool { + self.found_unknown_attr + } + + /// Is this compound type packed? + pub fn packed(&self) -> bool { + self.packed + } + + /// Returns whether this type needs an explicit vtable because it has + /// virtual methods and none of its base classes has already a vtable. + pub fn needs_explicit_vtable(&self, ctx: &BindgenContext) -> bool { + self.has_vtable(ctx) && + !self.base_members.iter().any(|base| { + // NB: Ideally, we could rely in all these types being `comp`, and + // life would be beautiful. + // + // Unfortunately, given the way we implement --match-pat, and also + // that you can inherit from templated types, we need to handle + // other cases here too. + ctx.resolve_type(*base) + .canonical_type(ctx) + .as_comp() + .map_or(false, |ci| ci.has_vtable(ctx)) + }) + } +} + +impl TypeCollector for CompInfo { + type Extra = Item; + + fn collect_types(&self, + context: &BindgenContext, + types: &mut ItemSet, + item: &Item) { + if let Some(template) = self.specialized_template() { + types.insert(template); + } + + let applicable_template_args = item.applicable_template_args(context); + for arg in applicable_template_args { + types.insert(arg); + } + + for &base in self.base_members() { + types.insert(base); + } + + for field in self.fields() { + types.insert(field.ty()); + } + + for &ty in self.inner_types() { + types.insert(ty); + } + + // FIXME(emilio): Methods, VTable? + } +} diff --git a/libbindgen/src/ir/context.rs b/libbindgen/src/ir/context.rs new file mode 100644 index 00000000..85721978 --- /dev/null +++ b/libbindgen/src/ir/context.rs @@ -0,0 +1,1055 @@ +//! Common context that is passed around during parsing and codegen. + +use BindgenOptions; +use cexpr; +use clang::{self, Cursor}; +use parse::ClangItemParser; +use std::borrow::Cow; +use std::cell::Cell; +use std::collections::{HashMap, hash_map}; +use std::collections::btree_map::{self, BTreeMap}; +use std::fmt; +use super::int::IntKind; +use super::item::{Item, ItemCanonicalName}; +use super::item_kind::ItemKind; +use super::module::Module; +use super::ty::{FloatKind, Type, TypeKind}; +use super::type_collector::{ItemSet, TypeCollector}; +use syntax::ast::Ident; +use syntax::codemap::{DUMMY_SP, Span}; +use syntax::ext::base::ExtCtxt; + +/// A single identifier for an item. +/// +/// TODO: Build stronger abstractions on top of this, like TypeId(ItemId)? +#[derive(Debug, Copy, Clone, PartialEq, Eq, PartialOrd, Ord, Hash)] +pub struct ItemId(usize); + +impl ItemId { + /// Get a numeric representation of this id. + pub fn as_usize(&self) -> usize { + self.0 + } +} + +/// A key used to index a resolved type, so we only process it once. +/// +/// This is almost always a USR string (an unique identifier generated by +/// clang), but it can also be the canonical declaration if the type is unnamed, +/// in which case clang may generate the same USR for multiple nested unnamed +/// types. +#[derive(Eq, PartialEq, Hash, Debug)] +enum TypeKey { + USR(String), + Declaration(Cursor), +} + +// This is just convenience to avoid creating a manual debug impl for the +// context. +struct GenContext<'ctx>(ExtCtxt<'ctx>); + +impl<'ctx> fmt::Debug for GenContext<'ctx> { + fn fmt(&self, fmt: &mut fmt::Formatter) -> fmt::Result { + write!(fmt, "GenContext {{ ... }}") + } +} + +/// A context used during parsing and generation of structs. +#[derive(Debug)] +pub struct BindgenContext<'ctx> { + /// The map of all the items parsed so far. + /// + /// It's a BTreeMap because we want the keys to be sorted to have consistent + /// output. + items: BTreeMap<ItemId, Item>, + + /// The next item id to use during this bindings regeneration. + next_item_id: ItemId, + + /// Clang USR to type map. This is needed to be able to associate types with + /// item ids during parsing. + types: HashMap<TypeKey, ItemId>, + + /// A cursor to module map. Similar reason than above. + modules: HashMap<Cursor, ItemId>, + + /// The root module, this is guaranteed to be an item of kind Module. + root_module: ItemId, + + /// Current module being traversed. + current_module: ItemId, + + /// A stack with the current type declarations and types we're parsing. This + /// is needed to avoid infinite recursion when parsing a type like: + /// + /// struct c { struct c* next; }; + /// + /// This means effectively, that a type has a potential ID before knowing if + /// it's a correct type. But that's not important in practice. + /// + /// We could also use the `types` HashMap, but my intention with it is that + /// only valid types and declarations end up there, and this could + /// potentially break that assumption. + /// + /// FIXME: Should not be public, though... meh. + pub currently_parsed_types: Vec<(Cursor, ItemId)>, + + /// A HashSet with all the already parsed macro names. This is done to avoid + /// hard errors while parsing duplicated macros, as well to allow macro + /// expression parsing. + parsed_macros: HashMap<Vec<u8>, cexpr::expr::EvalResult>, + + /// The active replacements collected from replaces="xxx" annotations. + replacements: HashMap<String, ItemId>, + + collected_typerefs: bool, + + /// Dummy structures for code generation. + gen_ctx: Option<&'ctx GenContext<'ctx>>, + span: Span, + + /// The clang index for parsing. + index: clang::Index, + + /// The translation unit for parsing. + translation_unit: clang::TranslationUnit, + + /// The options given by the user via cli or other medium. + options: BindgenOptions, + + /// Whether a bindgen complex was generated + generated_bindegen_complex: Cell<bool>, +} + +impl<'ctx> BindgenContext<'ctx> { + /// Construct the context for the given `options`. + pub fn new(options: BindgenOptions) -> Self { + use clangll; + + let index = clang::Index::new(false, true); + + let parse_options = + clangll::CXTranslationUnit_DetailedPreprocessingRecord; + let translation_unit = + clang::TranslationUnit::parse(&index, + "", + &options.clang_args, + &[], + parse_options) + .expect("TranslationUnit::parse"); + + let root_module = Self::build_root_module(ItemId(0)); + let mut me = BindgenContext { + items: Default::default(), + types: Default::default(), + modules: Default::default(), + next_item_id: ItemId(1), + root_module: root_module.id(), + current_module: root_module.id(), + currently_parsed_types: vec![], + parsed_macros: Default::default(), + replacements: Default::default(), + collected_typerefs: false, + gen_ctx: None, + span: DUMMY_SP, + index: index, + translation_unit: translation_unit, + options: options, + generated_bindegen_complex: Cell::new(false), + }; + + me.add_item(root_module, None, None); + + me + } + + /// Define a new item. + /// + /// This inserts it into the internal items set, and its type into the + /// internal types set. + pub fn add_item(&mut self, + item: Item, + declaration: Option<Cursor>, + location: Option<Cursor>) { + use clangll::{CXCursor_ClassTemplate, + CXCursor_ClassTemplatePartialSpecialization}; + debug!("BindgenContext::add_item({:?}, declaration: {:?}, loc: {:?}", + item, + declaration, + location); + debug_assert!(declaration.is_some() || !item.kind().is_type() || + item.kind().expect_type().is_builtin_or_named(), + "Adding a type without declaration?"); + + let id = item.id(); + let is_type = item.kind().is_type(); + let is_unnamed = is_type && item.expect_type().name().is_none(); + let old_item = self.items.insert(id, item); + assert!(old_item.is_none(), "Inserted type twice?"); + + // Unnamed items can have an USR, but they can't be referenced from + // other sites explicitly and the USR can match if the unnamed items are + // nested, so don't bother tracking them. + if is_type && declaration.is_some() { + let mut declaration = declaration.unwrap(); + if !declaration.is_valid() { + if let Some(location) = location { + if location.kind() == CXCursor_ClassTemplate || + location.kind() == + CXCursor_ClassTemplatePartialSpecialization { + declaration = location; + } + } + } + declaration = declaration.canonical(); + if !declaration.is_valid() { + // This could happen, for example, with types like `int*` or + // similar. + // + // Fortunately, we don't care about those types being + // duplicated, so we can just ignore them. + debug!("Invalid declaration {:?} found for type {:?}", + declaration, + self.items.get(&id).unwrap().kind().expect_type()); + return; + } + + let key = if is_unnamed { + TypeKey::Declaration(declaration) + } else if let Some(usr) = declaration.usr() { + TypeKey::USR(usr) + } else { + error!("Valid declaration with no USR: {:?}, {:?}", + declaration, + location); + return; + }; + + let old = self.types.insert(key, id); + debug_assert_eq!(old, None); + } + } + + // TODO: Move all this syntax crap to other part of the code. + + /// Given that we are in the codegen phase, get the syntex context. + pub fn ext_cx(&self) -> &ExtCtxt<'ctx> { + &self.gen_ctx.expect("Not in gen phase").0 + } + + /// Given that we are in the codegen phase, get the current syntex span. + pub fn span(&self) -> Span { + self.span + } + + /// Mangles a name so it doesn't conflict with any keyword. + pub fn rust_mangle<'a>(&self, name: &'a str) -> Cow<'a, str> { + use syntax::parse::token; + let ident = self.rust_ident_raw(name); + let token = token::Ident(ident); + if token.is_any_keyword() || name.contains("@") || + name.contains("?") || name.contains("$") || + "bool" == name { + let mut s = name.to_owned(); + s = s.replace("@", "_"); + s = s.replace("?", "_"); + s = s.replace("$", "_"); + s.push_str("_"); + return Cow::Owned(s); + } + Cow::Borrowed(name) + } + + /// Returns a mangled name as a rust identifier. + pub fn rust_ident(&self, name: &str) -> Ident { + self.rust_ident_raw(&self.rust_mangle(name)) + } + + /// Returns a mangled name as a rust identifier. + pub fn rust_ident_raw(&self, name: &str) -> Ident { + self.ext_cx().ident_of(name) + } + + /// Iterate over all items that have been defined. + pub fn items<'a>(&'a self) -> btree_map::Iter<'a, ItemId, Item> { + self.items.iter() + } + + /// Have we collected all unresolved type references yet? + pub fn collected_typerefs(&self) -> bool { + self.collected_typerefs + } + + /// Gather all the unresolved type references. + fn collect_typerefs + (&mut self) + -> Vec<(ItemId, clang::Type, Option<clang::Cursor>, Option<ItemId>)> { + debug_assert!(!self.collected_typerefs); + self.collected_typerefs = true; + let mut typerefs = vec![]; + for (id, ref mut item) in &mut self.items { + let kind = item.kind(); + let ty = match kind.as_type() { + Some(ty) => ty, + None => continue, + }; + + match *ty.kind() { + TypeKind::UnresolvedTypeRef(ref ty, loc, parent_id) => { + typerefs.push((*id, ty.clone(), loc, parent_id)); + } + _ => {} + }; + } + typerefs + } + + /// Collect all of our unresolved type references and resolve them. + fn resolve_typerefs(&mut self) { + let typerefs = self.collect_typerefs(); + + for (id, ty, loc, parent_id) in typerefs { + let _resolved = { + let resolved = Item::from_ty(&ty, loc, parent_id, self) + .expect("What happened?"); + let mut item = self.items.get_mut(&id).unwrap(); + + *item.kind_mut().as_type_mut().unwrap().kind_mut() = + TypeKind::ResolvedTypeRef(resolved); + resolved + }; + + // Something in the STL is trolling me. I don't need this assertion + // right now, but worth investigating properly once this lands. + // + // debug_assert!(self.items.get(&resolved).is_some(), "How?"); + } + } + + /// Iterate over all items and replace any item that has been named in a + /// `replaces="SomeType"` annotation with the replacement type. + fn process_replacements(&mut self) { + if self.replacements.is_empty() { + debug!("No replacements to process"); + return; + } + + // FIXME: This is linear, but the replaces="xxx" annotation was already + // there, and for better or worse it's useful, sigh... + // + // We leverage the ResolvedTypeRef thing, though, which is cool :P. + + let mut replacements = vec![]; + + for (id, item) in self.items.iter() { + // Calls to `canonical_name` are expensive, so eagerly filter out + // items that cannot be replaced. + let ty = match item.kind().as_type() { + Some(ty) => ty, + None => continue, + }; + + match *ty.kind() { + TypeKind::Comp(ref ci) if !ci.is_template_specialization() => {} + TypeKind::TemplateAlias(_, _) | + TypeKind::Alias(_, _) => {} + _ => continue, + } + + let name = item.real_canonical_name(self, + self.options() + .enable_cxx_namespaces, + true); + let replacement = self.replacements.get(&name); + + if let Some(replacement) = replacement { + if replacement != id { + // We set this just after parsing the annotation. It's + // very unlikely, but this can happen. + if self.items.get(replacement).is_some() { + replacements.push((*id, *replacement)); + } + } + } + } + + for (id, replacement) in replacements { + debug!("Replacing {:?} with {:?}", id, replacement); + + let mut item = self.items.get_mut(&id).unwrap(); + *item.kind_mut().as_type_mut().unwrap().kind_mut() = + TypeKind::ResolvedTypeRef(replacement); + } + } + + /// Enter the code generation phase, invoke the given callback `cb`, and + /// leave the code generation phase. + pub fn gen<F, Out>(&mut self, cb: F) -> Out + where F: FnOnce(&Self) -> Out, + { + use syntax::ext::expand::ExpansionConfig; + use syntax::codemap::{ExpnInfo, MacroBang, NameAndSpan}; + use syntax::ext::base; + use syntax::parse; + use std::mem; + + let cfg = ExpansionConfig::default("xxx".to_owned()); + let sess = parse::ParseSess::new(); + let mut loader = base::DummyResolver; + let mut ctx = + GenContext(base::ExtCtxt::new(&sess, vec![], cfg, &mut loader)); + + ctx.0.bt_push(ExpnInfo { + call_site: self.span, + callee: NameAndSpan { + format: MacroBang(parse::token::intern("")), + allow_internal_unstable: false, + span: None, + }, + }); + + // FIXME: This is evil, we should move code generation to use a wrapper + // of BindgenContext instead, I guess. Even though we know it's fine + // because we remove it before the end of this function. + self.gen_ctx = Some(unsafe { mem::transmute(&ctx) }); + + if !self.collected_typerefs() { + self.resolve_typerefs(); + self.process_replacements(); + } + + let ret = cb(self); + self.gen_ctx = None; + ret + } + + // This deserves a comment. Builtin types don't get a valid declaration, so + // we can't add it to the cursor->type map. + // + // That being said, they're not generated anyway, and are few, so the + // duplication and special-casing is fine. + // + // If at some point we care about the memory here, probably a map TypeKind + // -> builtin type ItemId would be the best to improve that. + fn add_builtin_item(&mut self, item: Item) { + debug!("add_builtin_item: item = {:?}", item); + debug_assert!(item.kind().is_type()); + let id = item.id(); + let old_item = self.items.insert(id, item); + assert!(old_item.is_none(), "Inserted type twice?"); + } + + fn build_root_module(id: ItemId) -> Item { + let module = Module::new(Some("root".into())); + Item::new(id, None, None, id, ItemKind::Module(module)) + } + + /// Get the root module. + pub fn root_module(&self) -> ItemId { + self.root_module + } + + /// Resolve the given `ItemId` as a type. + /// + /// Panics if there is no item for the given `ItemId` or if the resolved + /// item is not a `Type`. + pub fn resolve_type(&self, type_id: ItemId) -> &Type { + self.items.get(&type_id).unwrap().kind().expect_type() + } + + /// Resolve the given `ItemId` as a type, or `None` if there is no item with + /// the given id. + /// + /// Panics if the id resolves to an item that is not a type. + pub fn safe_resolve_type(&self, type_id: ItemId) -> Option<&Type> { + self.items.get(&type_id).map(|t| t.kind().expect_type()) + } + + /// Resolve the given `ItemId` into an `Item`, or `None` if no such item + /// exists. + pub fn resolve_item_fallible(&self, item_id: ItemId) -> Option<&Item> { + self.items.get(&item_id) + } + + /// Resolve the given `ItemId` into an `Item`. + /// + /// Panics if the given id does not resolve to any item. + pub fn resolve_item(&self, item_id: ItemId) -> &Item { + match self.items.get(&item_id) { + Some(item) => item, + None => panic!("Not an item: {:?}", item_id), + } + } + + /// Get the current module. + pub fn current_module(&self) -> ItemId { + self.current_module + } + + /// This is one of the hackiest methods in all the parsing code. This method + /// is used to allow having templates with another argument names instead of + /// the canonical ones. + /// + /// This is surprisingly difficult to do with libclang, due to the fact that + /// partial template specializations don't provide explicit template + /// argument information. + /// + /// The only way to do this as far as I know, is inspecting manually the + /// AST, looking for TypeRefs inside. This, unfortunately, doesn't work for + /// more complex cases, see the comment on the assertion below. + /// + /// To see an example of what this handles: + /// + /// ```c++ + /// template<typename T> + /// class Incomplete { + /// T p; + /// }; + /// + /// template<typename U> + /// class Foo { + /// Incomplete<U> bar; + /// }; + /// ``` + fn build_template_wrapper(&mut self, + with_id: ItemId, + wrapping: ItemId, + parent_id: ItemId, + ty: &clang::Type, + location: clang::Cursor, + declaration: clang::Cursor) + -> ItemId { + use clangll::*; + let mut args = vec![]; + location.visit(|c| { + if c.kind() == CXCursor_TypeRef { + // The `with_id` id will potentially end up unused if we give up + // on this type (for example, its a tricky partial template + // specialization), so if we pass `with_id` as the parent, it is + // potentially a dangling reference. Instead, use the canonical + // template declaration as the parent. It is already parsed and + // has a known-resolvable `ItemId`. + let new_ty = Item::from_ty_or_ref(c.cur_type(), + Some(c), + Some(wrapping), + self); + args.push(new_ty); + } + CXChildVisit_Continue + }); + + let item = { + let wrapping_type = self.resolve_type(wrapping); + if let TypeKind::Comp(ref ci) = *wrapping_type.kind() { + let old_args = ci.template_args(); + + // The following assertion actually fails with partial template + // specialization. But as far as I know there's no way at all to + // grab the specialized types from neither the AST or libclang, + // which sucks. The same happens for specialized type alias + // template declarations, where we have that ugly hack up there. + // + // This flaw was already on the old parser, but I now think it + // has no clear solution (apart from patching libclang to + // somehow expose them, of course). + // + // For an easy example in which there's no way at all of getting + // the `int` type, except manually parsing the spelling: + // + // template<typename T, typename U> + // class Incomplete { + // T d; + // U p; + // }; + // + // template<typename U> + // class Foo { + // Incomplete<U, int> bar; + // }; + // + // debug_assert_eq!(old_args.len(), args.len()); + // + // That being said, this is not so common, so just error! and + // hope for the best, returning the previous type, who knows. + if old_args.len() != args.len() { + error!("Found partial template specialization, \ + expect dragons!"); + return wrapping; + } + } else { + assert_eq!(declaration.kind(), + ::clangll::CXCursor_TypeAliasTemplateDecl, + "Expected wrappable type"); + } + + let type_kind = TypeKind::TemplateRef(wrapping, args); + let name = ty.spelling(); + let name = if name.is_empty() { None } else { Some(name) }; + let ty = Type::new(name, + ty.fallible_layout().ok(), + type_kind, + ty.is_const()); + Item::new(with_id, None, None, parent_id, ItemKind::Type(ty)) + }; + + // Bypass all the validations in add_item explicitly. + debug!("build_template_wrapper: inserting item: {:?}", item); + debug_assert!(with_id == item.id()); + self.items.insert(with_id, item); + with_id + } + + /// Looks up for an already resolved type, either because it's builtin, or + /// because we already have it in the map. + pub fn builtin_or_resolved_ty(&mut self, + with_id: ItemId, + parent_id: Option<ItemId>, + ty: &clang::Type, + location: Option<clang::Cursor>) + -> Option<ItemId> { + use clangll::{CXCursor_ClassTemplate, + CXCursor_ClassTemplatePartialSpecialization, + CXCursor_TypeAliasTemplateDecl, + CXCursor_TypeRef}; + debug!("builtin_or_resolved_ty: {:?}, {:?}, {:?}", + ty, + location, + parent_id); + let mut declaration = ty.declaration(); + if !declaration.is_valid() { + if let Some(location) = location { + if location.kind() == CXCursor_ClassTemplate || + location.kind() == + CXCursor_ClassTemplatePartialSpecialization { + declaration = location; + } + } + } + let canonical_declaration = declaration.canonical(); + if canonical_declaration.is_valid() { + let id = self.types + .get(&TypeKey::Declaration(canonical_declaration)) + .map(|id| *id) + .or_else(|| { + canonical_declaration.usr() + .and_then(|usr| self.types.get(&TypeKey::USR(usr))) + .map(|id| *id) + }); + if let Some(id) = id { + debug!("Already resolved ty {:?}, {:?}, {:?} {:?}", + id, + declaration, + ty, + location); + + // If the declaration existed, we *might* be done, but it's not + // the case for class templates, where the template arguments + // may vary. + // + // In this case, we create a TemplateRef with the new template + // arguments, pointing to the canonical template. + // + // Note that we only do it if parent_id is some, and we have a + // location for building the new arguments, the template + // argument names don't matter in the global context. + if (declaration.kind() == CXCursor_ClassTemplate || + declaration.kind() == + CXCursor_ClassTemplatePartialSpecialization || + declaration.kind() == CXCursor_TypeAliasTemplateDecl) && + *ty != canonical_declaration.cur_type() && + location.is_some() && + parent_id.is_some() { + // For specialized type aliases, there's no way to get the + // template parameters as of this writing (for a struct + // specialization we wouldn't be in this branch anyway). + // + // Explicitly return `None` if there aren't any + // unspecialized parameters (contains any `TypeRef`) so we + // resolve the canonical type if there is one and it's + // exposed. + // + // This is _tricky_, I know :( + if declaration.kind() == CXCursor_TypeAliasTemplateDecl && + !location.unwrap().contains_cursor(CXCursor_TypeRef) && + ty.canonical_type().is_valid_and_exposed() { + return None; + } + + return Some(self.build_template_wrapper(with_id, + id, + parent_id.unwrap(), + ty, + location.unwrap(), + declaration)); + } + + return Some(self.build_ty_wrapper(with_id, id, parent_id, ty)); + } + } + + debug!("Not resolved, maybe builtin?"); + + // Else, build it. + self.build_builtin_ty(ty, declaration) + } + + // This is unfortunately a lot of bloat, but is needed to properly track + // constness et. al. + // + // We should probably make the constness tracking separate, so it doesn't + // bloat that much, but hey, we already bloat the heck out of builtin types. + fn build_ty_wrapper(&mut self, + with_id: ItemId, + wrapped_id: ItemId, + parent_id: Option<ItemId>, + ty: &clang::Type) + -> ItemId { + let spelling = ty.spelling(); + let is_const = ty.is_const(); + let layout = ty.fallible_layout().ok(); + let type_kind = TypeKind::ResolvedTypeRef(wrapped_id); + let ty = Type::new(Some(spelling), layout, type_kind, is_const); + let item = Item::new(with_id, + None, + None, + parent_id.unwrap_or(self.current_module), + ItemKind::Type(ty)); + self.add_builtin_item(item); + with_id + } + + /// Returns the next item id to be used for an item. + pub fn next_item_id(&mut self) -> ItemId { + let ret = self.next_item_id; + self.next_item_id = ItemId(self.next_item_id.0 + 1); + ret + } + + fn build_builtin_ty(&mut self, + ty: &clang::Type, + _declaration: Cursor) + -> Option<ItemId> { + use clangll::*; + let type_kind = match ty.kind() { + CXType_NullPtr => TypeKind::NullPtr, + CXType_Void => TypeKind::Void, + CXType_Bool => TypeKind::Int(IntKind::Bool), + CXType_Int => TypeKind::Int(IntKind::Int), + CXType_UInt => TypeKind::Int(IntKind::UInt), + CXType_SChar | CXType_Char_S => TypeKind::Int(IntKind::Char), + CXType_UChar | CXType_Char_U => TypeKind::Int(IntKind::UChar), + CXType_Short => TypeKind::Int(IntKind::Short), + CXType_UShort => TypeKind::Int(IntKind::UShort), + CXType_WChar | CXType_Char16 => TypeKind::Int(IntKind::U16), + CXType_Char32 => TypeKind::Int(IntKind::U32), + CXType_Long => TypeKind::Int(IntKind::Long), + CXType_ULong => TypeKind::Int(IntKind::ULong), + CXType_LongLong => TypeKind::Int(IntKind::LongLong), + CXType_ULongLong => TypeKind::Int(IntKind::ULongLong), + CXType_Int128 => TypeKind::Int(IntKind::I128), + CXType_UInt128 => TypeKind::Int(IntKind::U128), + CXType_Float => TypeKind::Float(FloatKind::Float), + CXType_Double => TypeKind::Float(FloatKind::Double), + CXType_LongDouble => TypeKind::Float(FloatKind::LongDouble), + CXType_Float128 => TypeKind::Float(FloatKind::Float128), + CXType_Complex => { + let float_type = ty.elem_type() + .expect("Not able to resolve complex type?"); + let float_kind = match float_type.kind() { + CXType_Float => FloatKind::Float, + CXType_Double => FloatKind::Double, + CXType_LongDouble => FloatKind::LongDouble, + _ => panic!("Non floating-type complex?"), + }; + TypeKind::Complex(float_kind) + } + _ => return None, + }; + + let spelling = ty.spelling(); + let is_const = ty.is_const(); + let layout = ty.fallible_layout().ok(); + let ty = Type::new(Some(spelling), layout, type_kind, is_const); + let id = self.next_item_id(); + let item = + Item::new(id, None, None, self.root_module, ItemKind::Type(ty)); + self.add_builtin_item(item); + Some(id) + } + + /// Get the current Clang translation unit that is being processed. + pub fn translation_unit(&self) -> &clang::TranslationUnit { + &self.translation_unit + } + + /// Have we parsed the macro named `macro_name` already? + pub fn parsed_macro(&self, macro_name: &[u8]) -> bool { + self.parsed_macros.contains_key(macro_name) + } + + /// Get the currently parsed macros. + pub fn parsed_macros(&self) -> &HashMap<Vec<u8>, cexpr::expr::EvalResult> { + debug_assert!(!self.in_codegen_phase()); + &self.parsed_macros + } + + /// Mark the macro named `macro_name` as parsed. + pub fn note_parsed_macro(&mut self, + id: Vec<u8>, + value: cexpr::expr::EvalResult) { + self.parsed_macros.insert(id, value); + } + + /// Are we in the codegen phase? + pub fn in_codegen_phase(&self) -> bool { + self.gen_ctx.is_some() + } + + /// Mark the type with the given `name` as replaced by the type with id + /// `potential_ty`. + /// + /// Replacement types are declared using the `replaces="xxx"` annotation, + /// and implies that the original type is hidden. + pub fn replace(&mut self, name: &str, potential_ty: ItemId) { + match self.replacements.entry(name.into()) { + hash_map::Entry::Vacant(entry) => { + debug!("Defining replacement for {} as {:?}", + name, + potential_ty); + entry.insert(potential_ty); + } + hash_map::Entry::Occupied(occupied) => { + warn!("Replacement for {} already defined as {:?}; \ + ignoring duplicate replacement definition as {:?}}}", + name, + occupied.get(), + potential_ty); + } + } + } + + /// Is the item with the given `name` hidden? Or is the item with the given + /// `name` and `id` replaced by another type, and effectively hidden? + pub fn hidden_by_name(&self, name: &str, id: ItemId) -> bool { + debug_assert!(self.in_codegen_phase(), + "You're not supposed to call this yet"); + self.options.hidden_types.contains(name) || + self.is_replaced_type(name, id) + } + + /// Has the item with the given `name` and `id` been replaced by another + /// type? + pub fn is_replaced_type(&self, name: &str, id: ItemId) -> bool { + match self.replacements.get(name) { + Some(replaced_by) if *replaced_by != id => true, + _ => false, + } + } + + /// Is the type with the given `name` marked as opaque? + pub fn opaque_by_name(&self, name: &str) -> bool { + debug_assert!(self.in_codegen_phase(), + "You're not supposed to call this yet"); + self.options.opaque_types.contains(name) + } + + /// Get the options used to configure this bindgen context. + pub fn options(&self) -> &BindgenOptions { + &self.options + } + + /// Given a CXCursor_Namespace cursor, return the item id of the + /// corresponding module, or create one on the fly. + pub fn module(&mut self, cursor: clang::Cursor) -> ItemId { + use clangll::*; + assert!(cursor.kind() == CXCursor_Namespace, "Be a nice person"); + let cursor = cursor.canonical(); + if let Some(id) = self.modules.get(&cursor) { + return *id; + } + + let module_id = self.next_item_id(); + let module_name = self.translation_unit + .tokens(&cursor) + .and_then(|tokens| { + if tokens.len() <= 1 { + None + } else { + match &*tokens[1].spelling { + "{" => None, + s => Some(s.to_owned()), + } + } + }); + + let module = Module::new(module_name); + let module = Item::new(module_id, + None, + None, + self.current_module, + ItemKind::Module(module)); + + self.add_item(module, None, None); + + module_id + } + + /// Start traversing the module with the given `module_id`, invoke the + /// callback `cb`, and then return to traversing the original module. + pub fn with_module<F>(&mut self, module_id: ItemId, cb: F) + where F: FnOnce(&mut Self, &mut Vec<ItemId>), + { + debug_assert!(self.resolve_item(module_id).kind().is_module(), "Wat"); + + let previous_id = self.current_module; + self.current_module = module_id; + + let mut children = vec![]; + cb(self, &mut children); + + self.items + .get_mut(&module_id) + .unwrap() + .as_module_mut() + .expect("Not a module?") + .children_mut() + .extend(children.into_iter()); + + self.current_module = previous_id; + } + + /// Iterate over all (explicitly or transitively) whitelisted items. + /// + /// If no items are explicitly whitelisted, then all items are considered + /// whitelisted. + pub fn whitelisted_items<'me>(&'me self) + -> WhitelistedItemsIter<'me, 'ctx> { + assert!(self.in_codegen_phase()); + assert!(self.current_module == self.root_module); + + let roots = self.items() + .filter(|&(_, item)| { + // If nothing is explicitly whitelisted, then everything is fair + // game. + if self.options().whitelisted_types.is_empty() && + self.options().whitelisted_functions.is_empty() && + self.options().whitelisted_vars.is_empty() { + return true; + } + + let name = item.canonical_name(self); + match *item.kind() { + ItemKind::Module(..) => false, + ItemKind::Function(_) => { + self.options().whitelisted_functions.matches(&name) + } + ItemKind::Var(_) => { + self.options().whitelisted_vars.matches(&name) + } + ItemKind::Type(ref ty) => { + if self.options().whitelisted_types.matches(&name) { + return true; + } + + // Unnamed top-level enums are special and we whitelist + // them via the `whitelisted_vars` filter, since they're + // effectively top-level constants, and there's no way + // for them to be referenced consistently. + if let TypeKind::Enum(ref enum_) = *ty.kind() { + if ty.name().is_none() && + enum_.variants().iter().any(|variant| { + self.options() + .whitelisted_vars + .matches(&variant.name()) + }) { + return true; + } + } + + false + } + } + }) + .map(|(&id, _)| id); + + let seen: ItemSet = roots.collect(); + + // The .rev() preserves the expected ordering traversal, resulting in + // more stable-ish bindgen-generated names for anonymous types (like + // unions). + let to_iterate = seen.iter().cloned().rev().collect(); + + WhitelistedItemsIter { + ctx: self, + seen: seen, + to_iterate: to_iterate, + } + } + + /// Convenient method for getting the prefix to use for most traits in + /// codegen depending on the `use_core` option. + pub fn trait_prefix(&self) -> Ident { + if self.options().use_core { + self.rust_ident_raw("core") + } else { + self.rust_ident_raw("std") + } + } + + /// Call if a binden complex is generated + pub fn generated_bindegen_complex(&self) { + self.generated_bindegen_complex.set(true) + } + + /// Whether we need to generate the binden complex type + pub fn need_bindegen_complex_type(&self) -> bool { + self.generated_bindegen_complex.get() + } +} + +/// An iterator over whitelisted items. +/// +/// See `BindgenContext::whitelisted_items` for more information. +pub struct WhitelistedItemsIter<'ctx, 'gen> + where 'gen: 'ctx, +{ + ctx: &'ctx BindgenContext<'gen>, + + // The set of whitelisted items we have seen. If you think of traversing + // whitelisted items like GC tracing, this is the mark bits, and contains + // both black and gray items. + seen: ItemSet, + + // The set of whitelisted items that we have seen but have yet to iterate + // over and collect transitive references from. To return to the GC analogy, + // this is the mark stack, containing the set of gray items which we have + // not finished tracing yet. + to_iterate: Vec<ItemId>, +} + +impl<'ctx, 'gen> Iterator for WhitelistedItemsIter<'ctx, 'gen> + where 'gen: 'ctx, +{ + type Item = ItemId; + + fn next(&mut self) -> Option<Self::Item> { + let id = match self.to_iterate.pop() { + None => return None, + Some(id) => id, + }; + + debug_assert!(self.seen.contains(&id)); + debug_assert!(self.ctx.items.contains_key(&id)); + + let mut sub_types = ItemSet::new(); + id.collect_types(self.ctx, &mut sub_types, &()); + + for id in sub_types { + if self.seen.insert(id) { + self.to_iterate.push(id); + } + } + + Some(id) + } +} diff --git a/libbindgen/src/ir/enum_ty.rs b/libbindgen/src/ir/enum_ty.rs new file mode 100644 index 00000000..6085833d --- /dev/null +++ b/libbindgen/src/ir/enum_ty.rs @@ -0,0 +1,137 @@ +//! Intermediate representation for C/C++ enumerations. + +use clang; +use parse::{ClangItemParser, ParseError}; +use super::context::{BindgenContext, ItemId}; +use super::item::Item; +use super::ty::TypeKind; + +/// A C/C++ enumeration. +#[derive(Debug)] +pub struct Enum { + /// The representation used for this enum; it should be an `IntKind` type or + /// an alias to one. + /// + /// It's `None` if the enum is a forward declaration and isn't defined + /// anywhere else, see `tests/headers/func_ptr_in_struct.h`. + repr: Option<ItemId>, + + /// The different variants, with explicit values. + variants: Vec<EnumVariant>, +} + +impl Enum { + /// Construct a new `Enum` with the given representation and variants. + pub fn new(repr: Option<ItemId>, variants: Vec<EnumVariant>) -> Self { + Enum { + repr: repr, + variants: variants, + } + } + + /// Get this enumeration's representation. + pub fn repr(&self) -> Option<ItemId> { + self.repr + } + + /// Get this enumeration's variants. + pub fn variants(&self) -> &[EnumVariant] { + &self.variants + } + + /// Construct an enumeration from the given Clang type. + pub fn from_ty(ty: &clang::Type, + ctx: &mut BindgenContext) + -> Result<Self, ParseError> { + use clangll::*; + if ty.kind() != CXType_Enum { + return Err(ParseError::Continue); + } + + let declaration = ty.declaration().canonical(); + let repr = Item::from_ty(&declaration.enum_type(), None, None, ctx) + .ok(); + let mut variants = vec![]; + + let is_signed = match repr { + Some(repr) => { + let repr_type = ctx.resolve_type(repr); + match *repr_type.canonical_type(ctx).kind() { + TypeKind::Int(ref int_kind) => int_kind.is_signed(), + ref other => { + panic!("Since when enums can be non-integers? {:?}", + other) + } + } + } + // Assume signedness since the default type by the C + // standard is an + // int. + None => true, + }; + + declaration.visit(|cursor| { + if cursor.kind() == CXCursor_EnumConstantDecl { + let value = if is_signed { + cursor.enum_val_signed().map(EnumVariantValue::Signed) + } else { + cursor.enum_val_unsigned().map(EnumVariantValue::Unsigned) + }; + if let Some(val) = value { + let name = cursor.spelling(); + let comment = cursor.raw_comment(); + variants.push(EnumVariant::new(name, comment, val)); + } + } + CXChildVisit_Continue + }); + Ok(Enum::new(repr, variants)) + } +} + +/// A single enum variant, to be contained only in an enum. +#[derive(Debug)] +pub struct EnumVariant { + /// The name of the variant. + name: String, + + /// An optional doc comment. + comment: Option<String>, + + /// The integer value of the variant. + val: EnumVariantValue, +} + +/// A constant value assigned to an enumeration variant. +#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)] +pub enum EnumVariantValue { + /// A signed constant. + Signed(i64), + + /// An unsigned constant. + Unsigned(u64), +} + +impl EnumVariant { + /// Construct a new enumeration variant from the given parts. + pub fn new(name: String, + comment: Option<String>, + val: EnumVariantValue) + -> Self { + EnumVariant { + name: name, + comment: comment, + val: val, + } + } + + /// Get this variant's name. + pub fn name(&self) -> &str { + &self.name + } + + /// Get this variant's value. + pub fn val(&self) -> EnumVariantValue { + self.val + } +} diff --git a/libbindgen/src/ir/function.rs b/libbindgen/src/ir/function.rs new file mode 100644 index 00000000..eacb6c0e --- /dev/null +++ b/libbindgen/src/ir/function.rs @@ -0,0 +1,282 @@ +//! Intermediate representation for C/C++ functions and methods. + +use clang; +use clangll::Enum_CXCallingConv; +use parse::{ClangItemParser, ClangSubItemParser, ParseError, ParseResult}; +use super::context::{BindgenContext, ItemId}; +use super::item::Item; +use super::ty::TypeKind; +use super::type_collector::{ItemSet, TypeCollector}; +use syntax::abi; + +/// A function declaration, with a signature, arguments, and argument names. +/// +/// The argument names vector must be the same length as the ones in the +/// signature. +#[derive(Debug)] +pub struct Function { + /// The name of this function. + name: String, + + /// The mangled name, that is, the symbol. + mangled_name: Option<String>, + + /// The id pointing to the current function signature. + signature: ItemId, + + /// The doc comment on the function, if any. + comment: Option<String>, +} + +impl Function { + /// Construct a new function. + pub fn new(name: String, + mangled_name: Option<String>, + sig: ItemId, + comment: Option<String>) + -> Self { + Function { + name: name, + mangled_name: mangled_name, + signature: sig, + comment: comment, + } + } + + /// Get this function's name. + pub fn name(&self) -> &str { + &self.name + } + + /// Get this function's name. + pub fn mangled_name(&self) -> Option<&str> { + self.mangled_name.as_ref().map(|n| &**n) + } + + /// Get this function's signature. + pub fn signature(&self) -> ItemId { + self.signature + } +} + +/// A function signature. +#[derive(Debug)] +pub struct FunctionSig { + /// The return type of the function. + return_type: ItemId, + + /// The type of the arguments, optionally with the name of the argument when + /// declared. + argument_types: Vec<(Option<String>, ItemId)>, + + /// Whether this function is variadic. + is_variadic: bool, + + /// The ABI of this function. + abi: abi::Abi, +} + +fn get_abi(cc: Enum_CXCallingConv) -> abi::Abi { + use clangll::*; + match cc { + CXCallingConv_Default => abi::Abi::C, + CXCallingConv_C => abi::Abi::C, + CXCallingConv_X86StdCall => abi::Abi::Stdcall, + CXCallingConv_X86FastCall => abi::Abi::Fastcall, + CXCallingConv_AAPCS => abi::Abi::Aapcs, + CXCallingConv_X86_64Win64 => abi::Abi::Win64, + other => panic!("unsupported calling convention: {}", other), + } +} + +/// Get the mangled name for the cursor's referent. +pub fn cursor_mangling(cursor: &clang::Cursor) -> Option<String> { + // We early return here because libclang may crash in some case + // if we pass in a variable inside a partial specialized template. + // See servo/rust-bindgen#67. + if cursor.is_in_non_fully_specialized_template() { + return None; + } + + let mut mangling = cursor.mangling(); + if mangling.is_empty() { + return None; + } + + // Try to undo backend linkage munging (prepended _, generally) + if cfg!(target_os = "macos") { + mangling.remove(0); + } + + Some(mangling) +} + +impl FunctionSig { + /// Construct a new function signature. + pub fn new(return_type: ItemId, + arguments: Vec<(Option<String>, ItemId)>, + is_variadic: bool, + abi: abi::Abi) + -> Self { + FunctionSig { + return_type: return_type, + argument_types: arguments, + is_variadic: is_variadic, + abi: abi, + } + } + + /// Construct a new function signature from the given Clang type. + pub fn from_ty(ty: &clang::Type, + cursor: &clang::Cursor, + ctx: &mut BindgenContext) + -> Result<Self, ParseError> { + use clangll::*; + debug!("FunctionSig::from_ty {:?} {:?}", ty, cursor); + + // Don't parse operatorxx functions in C++ + let spelling = cursor.spelling(); + if spelling.starts_with("operator") { + return Err(ParseError::Continue); + } + + let cursor = if cursor.is_valid() { + *cursor + } else { + ty.declaration() + }; + let mut args: Vec<_> = match cursor.kind() { + CXCursor_FunctionDecl | + CXCursor_CXXMethod => { + // For CXCursor_FunctionDecl, cursor.args() is the reliable way + // to get parameter names and types. + cursor.args() + .iter() + .map(|arg| { + let arg_ty = arg.cur_type(); + let name = arg.spelling(); + let name = + if name.is_empty() { None } else { Some(name) }; + let ty = Item::from_ty(&arg_ty, Some(*arg), None, ctx) + .expect("Argument?"); + (name, ty) + }) + .collect() + } + _ => { + // For non-CXCursor_FunctionDecl, visiting the cursor's children + // is the only reliable way to get parameter names. + let mut args = vec![]; + cursor.visit(|c| { + if c.kind() == CXCursor_ParmDecl { + let ty = + Item::from_ty(&c.cur_type(), Some(c), None, ctx) + .expect("ParmDecl?"); + let name = c.spelling(); + let name = + if name.is_empty() { None } else { Some(name) }; + args.push((name, ty)); + } + CXChildVisit_Continue + }); + args + } + }; + + if cursor.kind() == CXCursor_CXXMethod { + let is_const = cursor.method_is_const(); + let is_virtual = cursor.method_is_virtual(); + let is_static = cursor.method_is_static(); + if !is_static && !is_virtual { + let class = Item::parse(cursor.semantic_parent(), None, ctx) + .expect("Expected to parse the class"); + let ptr = + Item::builtin_type(TypeKind::Pointer(class), is_const, ctx); + args.insert(0, (Some("this".into()), ptr)); + } else if is_virtual { + let void = Item::builtin_type(TypeKind::Void, false, ctx); + let ptr = + Item::builtin_type(TypeKind::Pointer(void), false, ctx); + args.insert(0, (Some("this".into()), ptr)); + } + } + + let ty_ret_type = try!(ty.ret_type().ok_or(ParseError::Continue)); + let ret = try!(Item::from_ty(&ty_ret_type, None, None, ctx)); + let abi = get_abi(ty.call_conv()); + + Ok(Self::new(ret, args, ty.is_variadic(), abi)) + } + + /// Get this function signature's return type. + pub fn return_type(&self) -> ItemId { + self.return_type + } + + /// Get this function signature's argument (name, type) pairs. + pub fn argument_types(&self) -> &[(Option<String>, ItemId)] { + &self.argument_types + } + + /// Get this function signature's ABI. + pub fn abi(&self) -> abi::Abi { + self.abi + } + + /// Is this function signature variadic? + pub fn is_variadic(&self) -> bool { + // Clang reports some functions as variadic when they *might* be + // variadic. We do the argument check because rust doesn't codegen well + // variadic functions without an initial argument. + self.is_variadic && !self.argument_types.is_empty() + } +} + +impl ClangSubItemParser for Function { + fn parse(cursor: clang::Cursor, + context: &mut BindgenContext) + -> Result<ParseResult<Self>, ParseError> { + use clangll::*; + match cursor.kind() { + CXCursor_FunctionDecl | + CXCursor_CXXMethod => {} + _ => return Err(ParseError::Continue), + }; + + debug!("Function::parse({:?}, {:?})", cursor, cursor.cur_type()); + + // Grab the signature using Item::from_ty. + let sig = try!(Item::from_ty(&cursor.cur_type(), + Some(cursor), + None, + context)); + + let name = cursor.spelling(); + assert!(!name.is_empty(), "Empty function name?"); + + let mut mangled_name = cursor_mangling(&cursor); + if mangled_name.as_ref() == Some(&name) { + mangled_name = None; + } + + let comment = cursor.raw_comment(); + + let function = Self::new(name, mangled_name, sig, comment); + Ok(ParseResult::New(function, Some(cursor))) + } +} + +impl TypeCollector for FunctionSig { + type Extra = Item; + + fn collect_types(&self, + _context: &BindgenContext, + types: &mut ItemSet, + _item: &Item) { + types.insert(self.return_type()); + + for &(_, ty) in self.argument_types() { + types.insert(ty); + } + } +} diff --git a/libbindgen/src/ir/int.rs b/libbindgen/src/ir/int.rs new file mode 100644 index 00000000..2d85db83 --- /dev/null +++ b/libbindgen/src/ir/int.rs @@ -0,0 +1,93 @@ +//! Intermediate representation for integral types. + +/// Which integral type are we dealing with? +#[derive(Debug, Copy, Clone, PartialEq, Eq, Hash, PartialOrd, Ord)] +pub enum IntKind { + /// A `bool`. + Bool, + + /// A `char`. + Char, + + /// An `unsigned char`. + UChar, + + /// A `short`. + Short, + + /// An `unsigned short`. + UShort, + + /// An `int`. + Int, + + /// An `unsigned int`. + UInt, + + /// A `long`. + Long, + + /// An `unsigned long`. + ULong, + + /// A `long long`. + LongLong, + + /// An `unsigned long long`. + ULongLong, + + /// A 8-bit signed integer. + I8, + + /// A 8-bit unsigned integer. + U8, + + /// A 16-bit signed integer. + I16, + + /// Either a `char16_t` or a `wchar_t`. + U16, + + /// A 32-bit signed integer. + I32, + + /// A 32-bit unsigned integer. + U32, + + /// A 64-bit signed integer. + I64, + + /// A 64-bit unsigned integer. + U64, + + /// An `int128_t` + I128, + + /// A `uint128_t`. + U128, + + /// A custom integer type, used to allow custom macro types depending on + /// range. + Custom { + /// The name of the type, which would be used without modification. + name: &'static str, + /// Whether the type is signed or not. + is_signed: bool, + }, +} + +impl IntKind { + /// Is this integral type signed? + pub fn is_signed(&self) -> bool { + use self::IntKind::*; + match *self { + Bool | UChar | UShort | UInt | ULong | ULongLong | U8 | U16 | + U32 | U64 | U128 => false, + + Char | Short | Int | Long | LongLong | I8 | I16 | I32 | I64 | + I128 => true, + + Custom { is_signed, .. } => is_signed, + } + } +} diff --git a/libbindgen/src/ir/item.rs b/libbindgen/src/ir/item.rs new file mode 100644 index 00000000..1f05f92f --- /dev/null +++ b/libbindgen/src/ir/item.rs @@ -0,0 +1,1172 @@ +//! Bindgen's core intermediate representation type. + +use clang; +use parse::{ClangItemParser, ClangSubItemParser, ParseError, ParseResult}; +use regex::Regex; +use std::cell::{Cell, RefCell}; +use super::annotations::Annotations; +use super::context::{BindgenContext, ItemId}; +use super::function::Function; +use super::item_kind::ItemKind; +use super::module::Module; +use super::ty::{Type, TypeKind}; +use super::type_collector::{ItemSet, TypeCollector}; + +/// A trait to get the canonical name from an item. +/// +/// This is the trait that will eventually isolate all the logic related to name +/// mangling and that kind of stuff. +/// +/// This assumes no nested paths, at some point I'll have to make it a more +/// complex thing. +/// +/// This name is required to be safe for Rust, that is, is not expected to +/// return any rust keyword from here. +pub trait ItemCanonicalName { + /// Get the canonical name for this item. + fn canonical_name(&self, ctx: &BindgenContext) -> String; +} + +/// The same, but specifies the path that needs to be followed to reach an item. +/// +/// To contrast with canonical_name, here's an example: +/// +/// ```c++ +/// namespace foo { +/// const BAR = 3; +/// } +/// ``` +/// +/// For bar, the canonical path is `vec!["foo", "BAR"]`, while the canonical +/// name is just `"BAR"`. +pub trait ItemCanonicalPath { + /// Get the canonical path for this item. + fn canonical_path(&self, ctx: &BindgenContext) -> Vec<String>; +} + +/// A trait for iterating over an item and its parents and up its ancestor chain +/// up to (but not including) the implicit root module. +pub trait ItemAncestors { + /// Get an iterable over this item's ancestors. + fn ancestors<'a, 'b>(&self, + ctx: &'a BindgenContext<'b>) + -> ItemAncestorsIter<'a, 'b>; +} + +/// An iterator over an item and its ancestors. +pub struct ItemAncestorsIter<'a, 'b> + where 'b: 'a, +{ + item: ItemId, + ctx: &'a BindgenContext<'b>, +} + +impl<'a, 'b> Iterator for ItemAncestorsIter<'a, 'b> + where 'b: 'a, +{ + type Item = ItemId; + + fn next(&mut self) -> Option<Self::Item> { + let item = self.ctx.resolve_item(self.item); + if item.parent_id() == self.item { + None + } else { + self.item = item.parent_id(); + Some(item.id()) + } + } +} + +// Pure convenience +impl ItemCanonicalName for ItemId { + fn canonical_name(&self, ctx: &BindgenContext) -> String { + debug_assert!(ctx.in_codegen_phase(), + "You're not supposed to call this yet"); + ctx.resolve_item(*self).canonical_name(ctx) + } +} + +impl ItemCanonicalPath for ItemId { + fn canonical_path(&self, ctx: &BindgenContext) -> Vec<String> { + debug_assert!(ctx.in_codegen_phase(), + "You're not supposed to call this yet"); + ctx.resolve_item(*self).canonical_path(ctx) + } +} + +impl ItemAncestors for ItemId { + fn ancestors<'a, 'b>(&self, + ctx: &'a BindgenContext<'b>) + -> ItemAncestorsIter<'a, 'b> { + ItemAncestorsIter { + item: *self, + ctx: ctx, + } + } +} + +impl ItemAncestors for Item { + fn ancestors<'a, 'b>(&self, + ctx: &'a BindgenContext<'b>) + -> ItemAncestorsIter<'a, 'b> { + self.id().ancestors(ctx) + } +} + +impl TypeCollector for ItemId { + type Extra = (); + + fn collect_types(&self, + ctx: &BindgenContext, + types: &mut ItemSet, + extra: &()) { + ctx.resolve_item(*self).collect_types(ctx, types, extra); + } +} + +impl TypeCollector for Item { + type Extra = (); + + fn collect_types(&self, + ctx: &BindgenContext, + types: &mut ItemSet, + _extra: &()) { + if self.is_hidden(ctx) || types.contains(&self.id()) { + return; + } + + match *self.kind() { + ItemKind::Type(ref ty) => { + if !self.is_opaque(ctx) { + ty.collect_types(ctx, types, self); + } + } + _ => {} // FIXME. + } + } +} + +/// An item is the base of the bindgen representation, it can be either a +/// module, a type, a function, or a variable (see `ItemKind` for more +/// information). +/// +/// Items refer to each other by `ItemId`. Every item has its parent's +/// id. Depending on the kind of item this is, it may also refer to other items, +/// such as a compound type item referring to other types. Collectively, these +/// references form a graph. +/// +/// The entry-point to this graph is the "root module": a meta-item used to hold +/// all top-level items. +/// +/// An item may have a comment, and annotations (see the `annotations` module). +/// +/// Note that even though we parse all the types of annotations in comments, not +/// all of them apply to every item. Those rules are described in the +/// `annotations` module. +#[derive(Debug)] +pub struct Item { + /// This item's id. + id: ItemId, + + /// The item's local id, unique only amongst its siblings. Only used for + /// anonymous items. + /// + /// Lazily initialized in local_id(). + /// + /// Note that only structs, unions, and enums get a local type id. In any + /// case this is an implementation detail. + local_id: Cell<Option<usize>>, + + /// The next local id to use for a child.. + next_child_local_id: Cell<usize>, + + /// A cached copy of the canonical name, as returned by `canonical_name`. + /// + /// This is a fairly used operation during codegen so this makes bindgen + /// considerably faster in those cases. + canonical_name_cache: RefCell<Option<String>>, + + /// A doc comment over the item, if any. + comment: Option<String>, + /// Annotations extracted from the doc comment, or the default ones + /// otherwise. + annotations: Annotations, + /// An item's parent id. This will most likely be a class where this item + /// was declared, or a module, etc. + /// + /// All the items have a parent, except the root module, in which case the + /// parent id is its own id. + parent_id: ItemId, + /// The item kind. + kind: ItemKind, +} + +impl Item { + /// Construct a new `Item`. + pub fn new(id: ItemId, + comment: Option<String>, + annotations: Option<Annotations>, + parent_id: ItemId, + kind: ItemKind) + -> Self { + debug_assert!(id != parent_id || kind.is_module()); + Item { + id: id, + local_id: Cell::new(None), + next_child_local_id: Cell::new(1), + canonical_name_cache: RefCell::new(None), + parent_id: parent_id, + comment: comment, + annotations: annotations.unwrap_or_default(), + kind: kind, + } + } + + /// Get this `Item`'s identifier. + pub fn id(&self) -> ItemId { + self.id + } + + /// Get this `Item`'s parent's identifier. + /// + /// For the root module, the parent's ID is its own ID. + pub fn parent_id(&self) -> ItemId { + self.parent_id + } + + /// Get this `Item`'s comment, if it has any. + pub fn comment(&self) -> Option<&str> { + self.comment.as_ref().map(|c| &**c) + } + + /// What kind of item is this? + pub fn kind(&self) -> &ItemKind { + &self.kind + } + + /// Get a mutable reference to this item's kind. + pub fn kind_mut(&mut self) -> &mut ItemKind { + &mut self.kind + } + + /// Get an identifier that differentiates this item from its siblings. + /// + /// This should stay relatively stable in the face of code motion outside or + /// below this item's lexical scope, meaning that this can be useful for + /// generating relatively stable identifiers within a scope. + pub fn local_id(&self, ctx: &BindgenContext) -> usize { + if self.local_id.get().is_none() { + let parent = ctx.resolve_item(self.parent_id); + let local_id = parent.next_child_local_id.get(); + parent.next_child_local_id.set(local_id + 1); + self.local_id.set(Some(local_id)); + } + self.local_id.get().unwrap() + } + + /// Returns whether this item is a top-level item, from the point of view of + /// bindgen. + /// + /// This point of view changes depending on whether namespaces are enabled + /// or not. That way, in the following example: + /// + /// ```c++ + /// namespace foo { + /// static int var; + /// } + /// ``` + /// + /// `var` would be a toplevel item if namespaces are disabled, but won't if + /// they aren't. + /// + /// This function is used to determine when the codegen phase should call + /// `codegen` on an item, since any item that is not top-level will be + /// generated by its parent. + pub fn is_toplevel(&self, ctx: &BindgenContext) -> bool { + // FIXME: Workaround for some types falling behind when parsing weird + // stl classes, for example. + if ctx.options().enable_cxx_namespaces && self.kind().is_module() && + self.id() != ctx.root_module() { + return false; + } + + let mut parent = self.parent_id; + loop { + let parent_item = match ctx.resolve_item_fallible(parent) { + Some(item) => item, + None => return false, + }; + + if parent_item.id() == ctx.root_module() { + return true; + } else if ctx.options().enable_cxx_namespaces || + !parent_item.kind().is_module() { + return false; + } + + parent = parent_item.parent_id(); + } + } + + /// Get a reference to this item's underlying `Type`. Panic if this is some + /// other kind of item. + pub fn expect_type(&self) -> &Type { + self.kind().expect_type() + } + + /// Get a reference to this item's underlying `Type`, or `None` if this is + /// some other kind of item. + pub fn as_type(&self) -> Option<&Type> { + self.kind().as_type() + } + + /// Get a reference to this item's underlying `Function`. Panic if this is + /// some other kind of item. + pub fn expect_function(&self) -> &Function { + self.kind().expect_function() + } + + /// Checks whether an item contains in its "type signature" some named type. + /// + /// This function is used to avoid unused template parameter errors in Rust + /// when generating typedef declarations, and also to know whether we need + /// to generate a `PhantomData` member for a template parameter. + /// + /// For example, in code like the following: + /// + /// ```c++ + /// template<typename T, typename U> + /// struct Foo { + /// T bar; + /// + /// struct Baz { + /// U bas; + /// }; + /// }; + /// ``` + /// + /// Both `Foo` and `Baz` contain both `T` and `U` template parameters in + /// their signature: + /// + /// * `Foo<T, U>` + /// * `Bar<T, U>` + /// + /// But the Rust structure for `Foo` would look like: + /// + /// ```rust + /// struct Foo<T, U> { + /// bar: T, + /// _phantom0: ::std::marker::PhantomData<U>, + /// } + /// ``` + /// + /// because none of its member fields contained the `U` type in the + /// signature. Similarly, `Bar` would contain a `PhantomData<T>` type, for + /// the same reason. + /// + /// Note that this is somewhat similar to `applicable_template_args`, but + /// this also takes into account other kind of types, like arrays, + /// (`[T; 40]`), pointers: `*mut T`, etc... + /// + /// Normally we could do this check just in the `Type` kind, but we also + /// need to check the `applicable_template_args` more generally, since we + /// could need a type transitively from our parent, see the test added in + /// commit 2a3f93074dd2898669dbbce6e97e5cc4405d7cb1. + /// + /// It's kind of unfortunate (in the sense that it's a sort of complex + /// process), but I think it should get all the cases. + fn signature_contains_named_type(&self, + ctx: &BindgenContext, + ty: &Type) + -> bool { + debug_assert!(ty.is_named()); + self.expect_type().signature_contains_named_type(ctx, ty) || + self.applicable_template_args(ctx).iter().any(|template| { + ctx.resolve_type(*template).signature_contains_named_type(ctx, ty) + }) + } + + /// Returns the template arguments that apply to a struct. This is a concept + /// needed because of type declarations inside templates, for example: + /// + /// ```c++ + /// template<typename T> + /// class Foo { + /// typedef T element_type; + /// typedef int Bar; + /// + /// template<typename U> + /// class Baz { + /// }; + /// }; + /// ``` + /// + /// In this case, the applicable template arguments for the different types + /// would be: + /// + /// * `Foo`: [`T`] + /// * `Foo::element_type`: [`T`] + /// * `Foo::Bar`: [`T`] + /// * `Foo::Baz`: [`T`, `U`] + /// + /// You might notice that we can't generate something like: + /// + /// ```rust,ignore + /// type Foo_Bar<T> = ::std::os::raw::c_int; + /// ``` + /// + /// since that would be invalid Rust. Still, conceptually, `Bar` *could* use + /// the template parameter type `T`, and that's exactly what this method + /// represents. The unused template parameters get stripped in the + /// `signature_contains_named_type` check. + pub fn applicable_template_args(&self, + ctx: &BindgenContext) + -> Vec<ItemId> { + let ty = match *self.kind() { + ItemKind::Type(ref ty) => ty, + _ => return vec![], + }; + + fn parent_contains(ctx: &BindgenContext, + parent_template_args: &[ItemId], + item: ItemId) + -> bool { + let item_ty = ctx.resolve_type(item); + parent_template_args.iter().any(|parent_item| { + let parent_ty = ctx.resolve_type(*parent_item); + match (parent_ty.kind(), item_ty.kind()) { + (&TypeKind::Named(ref n, _), + &TypeKind::Named(ref i, _)) => n == i, + _ => false, + } + }) + } + + match *ty.kind() { + TypeKind::Named(..) => vec![self.id()], + TypeKind::Array(inner, _) | + TypeKind::Pointer(inner) | + TypeKind::Reference(inner) | + TypeKind::ResolvedTypeRef(inner) => { + ctx.resolve_item(inner).applicable_template_args(ctx) + } + TypeKind::Alias(_, inner) => { + let parent_args = ctx.resolve_item(self.parent_id()) + .applicable_template_args(ctx); + let inner = ctx.resolve_item(inner); + + // Avoid unused type parameters, sigh. + parent_args.iter() + .cloned() + .filter(|arg| { + let arg = ctx.resolve_type(*arg); + arg.is_named() && + inner.signature_contains_named_type(ctx, arg) + }) + .collect() + } + // XXX Is this completely correct? Partial template specialization + // is hard anyways, sigh... + TypeKind::TemplateAlias(_, ref args) | + TypeKind::TemplateRef(_, ref args) => args.clone(), + // In a template specialization we've got all we want. + TypeKind::Comp(ref ci) if ci.is_template_specialization() => { + ci.template_args().iter().cloned().collect() + } + TypeKind::Comp(ref ci) => { + let mut parent_template_args = + ctx.resolve_item(self.parent_id()) + .applicable_template_args(ctx); + + for ty in ci.template_args() { + if !parent_contains(ctx, &parent_template_args, *ty) { + parent_template_args.push(*ty); + } + } + + parent_template_args + } + _ => vec![], + } + } + + fn is_module(&self) -> bool { + match self.kind { + ItemKind::Module(..) => true, + _ => false, + } + } + + /// Get this item's annotations. + pub fn annotations(&self) -> &Annotations { + &self.annotations + } + + /// Whether this item should be hidden. + /// + /// This may be due to either annotations or to other kind of configuration. + pub fn is_hidden(&self, ctx: &BindgenContext) -> bool { + debug_assert!(ctx.in_codegen_phase(), + "You're not supposed to call this yet"); + self.annotations.hide() || + ctx.hidden_by_name(&self.real_canonical_name(ctx, false, true), self.id) + } + + /// Is this item opaque? + pub fn is_opaque(&self, ctx: &BindgenContext) -> bool { + debug_assert!(ctx.in_codegen_phase(), + "You're not supposed to call this yet"); + self.annotations.opaque() || + ctx.opaque_by_name(&self.real_canonical_name(ctx, false, true)) + } + + /// Is this a reference to another type? + pub fn is_type_ref(&self) -> bool { + self.as_type().map_or(false, |ty| ty.is_type_ref()) + } + + /// Get the canonical name without taking into account the replaces + /// annotation. + /// + /// This is the base logic used to implement hiding and replacing via + /// annotations, and also to implement proper name mangling. + /// + /// The idea is that each generated type in the same "level" (read: module + /// or namespace) has a unique canonical name. + /// + /// This name should be derived from the immutable state contained in the + /// type and the parent chain, since it should be consistent. + pub fn real_canonical_name(&self, + ctx: &BindgenContext, + count_namespaces: bool, + for_name_checking: bool) + -> String { + let base_name = match *self.kind() { + ItemKind::Type(ref ty) => { + match *ty.kind() { + // If we're a template specialization, our name is our + // parent's. + TypeKind::Comp(ref ci) + if ci.is_template_specialization() => { + return ci.specialized_template().unwrap() + .canonical_name(ctx); + }, + // Same as above + TypeKind::ResolvedTypeRef(inner) | + TypeKind::TemplateRef(inner, _) => { + return inner.canonical_name(ctx); + } + // If we're a named type, we don't need to mangle it, and we + // should be able to assert we're not top level. + TypeKind::Named(ref name, _) => { + return name.to_owned(); + } + // We call codegen on the inner type, but we do not want + // this alias's name to appear in the canonical name just + // because it is in the inner type's parent chain, so we use + // an empty base name. + // + // Note that this would be incorrect if this type could be + // referenced from, let's say, a member variable, but in + // that case the referenced type is the inner alias, so + // we're good there. If we wouldn't, a more complex solution + // would be needed. + TypeKind::TemplateAlias(inner, _) => { + if for_name_checking { + return ctx.resolve_item(inner) + .real_canonical_name(ctx, + count_namespaces, + false); + } + Some("") + } + // Else use the proper name, or fallback to a name with an + // id. + _ => { + ty.name() + } + }.map(ToOwned::to_owned) + } + ItemKind::Function(ref fun) => { + let mut base = fun.name().to_owned(); + + // We might need to deduplicate if we're a method. + let parent = ctx.resolve_item(self.parent_id()); + if let ItemKind::Type(ref ty) = *parent.kind() { + if let TypeKind::Comp(ref ci) = *ty.kind() { + let mut count = 0; + let mut found = false; + for method in ci.methods() { + if method.signature() == self.id() { + found = true; + break; + } + let fun = ctx.resolve_item(method.signature()) + .expect_function(); + if fun.name() == base { + count += 1; + } + } + + assert!(found, "Method not found?"); + if count != 0 { + base.push_str(&count.to_string()); + } + } + } + Some(base) + } + ItemKind::Var(ref var) => Some(var.name().to_owned()), + ItemKind::Module(ref module) => { + module.name().map(ToOwned::to_owned) + } + }; + + let parent = ctx.resolve_item(self.parent_id()); + let parent_is_namespace = parent.is_module(); + + if self.is_toplevel(ctx) || (parent_is_namespace && count_namespaces) { + let base_name = self.make_exposed_name(None, base_name, ctx); + return ctx.rust_mangle(&base_name).into_owned(); + } + + // TODO: allow modification of the mangling functions, maybe even per + // item type? + let parent_name = parent.canonical_name(ctx); + self.make_exposed_name(Some(parent_name), base_name, ctx) + } + + fn exposed_id(&self, ctx: &BindgenContext) -> String { + // Only use local ids for enums, classes, structs and union types. All + // other items use their global id. + let ty_kind = self.kind().as_type().map(|t| t.kind()); + if let Some(ty_kind) = ty_kind { + match *ty_kind { + TypeKind::Comp(..) | + TypeKind::Enum(..) => return self.local_id(ctx).to_string(), + _ => {} + } + } + + // Note that this `id_` prefix prevents (really unlikely) collisions + // between the global id and the local id of an item with the same + // parent. + format!("id_{}", self.id().as_usize()) + } + + fn make_exposed_name(&self, + parent_name: Option<String>, + base_name: Option<String>, + ctx: &BindgenContext) + -> String { + lazy_static! { + static ref RE_ENDS_WITH_BINDGEN_TY: Regex = + Regex::new(r"_bindgen_ty(_\d+)+$").unwrap(); + + static ref RE_ENDS_WITH_BINDGEN_MOD: Regex = + Regex::new(r"_bindgen_mod(_\d+)+$").unwrap(); + } + + let (re, kind) = match *self.kind() { + ItemKind::Module(..) => (&*RE_ENDS_WITH_BINDGEN_MOD, "mod"), + _ => (&*RE_ENDS_WITH_BINDGEN_TY, "ty"), + }; + + let parent_name = + parent_name.and_then(|n| if n.is_empty() { None } else { Some(n) }); + match (parent_name, base_name) { + (Some(parent), Some(base)) => format!("{}_{}", parent, base), + (Some(parent), None) => { + if re.is_match(parent.as_str()) { + format!("{}_{}", parent, self.exposed_id(ctx)) + } else { + format!("{}__bindgen_{}_{}", + parent, + kind, + self.exposed_id(ctx)) + } + } + (None, Some(base)) => base, + (None, None) => { + format!("_bindgen_{}_{}", kind, self.exposed_id(ctx)) + } + } + } + + /// Get a mutable reference to this item's `Module`, or `None` if this is + /// not a `Module` item. + pub fn as_module_mut(&mut self) -> Option<&mut Module> { + match self.kind { + ItemKind::Module(ref mut module) => Some(module), + _ => None, + } + } + + /// Can we derive an implementation of the `Copy` trait for this type? + pub fn can_derive_copy(&self, ctx: &BindgenContext) -> bool { + self.expect_type().can_derive_copy(ctx, self) + } + + /// Can we derive an implementation of the `Copy` trait for an array of this + /// type? + /// + /// See `Type::can_derive_copy_in_array` for details. + pub fn can_derive_copy_in_array(&self, ctx: &BindgenContext) -> bool { + self.expect_type().can_derive_copy_in_array(ctx, self) + } +} + +impl ClangItemParser for Item { + fn builtin_type(kind: TypeKind, + is_const: bool, + ctx: &mut BindgenContext) + -> ItemId { + // Feel free to add more here, I'm just lazy. + match kind { + TypeKind::Void | + TypeKind::Int(..) | + TypeKind::Pointer(..) | + TypeKind::Float(..) => {} + _ => panic!("Unsupported builtin type"), + } + + let ty = Type::new(None, None, kind, is_const); + let id = ctx.next_item_id(); + let module = ctx.root_module(); + ctx.add_item(Item::new(id, None, None, module, ItemKind::Type(ty)), + None, + None); + id + } + + + fn parse(cursor: clang::Cursor, + parent_id: Option<ItemId>, + ctx: &mut BindgenContext) + -> Result<ItemId, ParseError> { + use ir::function::Function; + use ir::module::Module; + use ir::var::Var; + use clangll::*; + + if !cursor.is_valid() { + return Err(ParseError::Continue); + } + + let comment = cursor.raw_comment(); + let annotations = Annotations::new(&cursor); + + let current_module = ctx.current_module(); + let relevant_parent_id = parent_id.unwrap_or(current_module); + + macro_rules! try_parse { + ($what:ident) => { + match $what::parse(cursor, ctx) { + Ok(ParseResult::New(item, declaration)) => { + let id = ctx.next_item_id(); + + ctx.add_item(Item::new(id, comment, annotations, + relevant_parent_id, + ItemKind::$what(item)), + declaration, + Some(cursor)); + return Ok(id); + } + Ok(ParseResult::AlreadyResolved(id)) => { + return Ok(id); + } + Err(ParseError::Recurse) => return Err(ParseError::Recurse), + Err(ParseError::Continue) => {}, + } + } + } + + try_parse!(Module); + + // NOTE: Is extremely important to parse functions and vars **before** + // types. Otherwise we can parse a function declaration as a type + // (which is legal), and lose functions to generate. + // + // In general, I'm not totally confident this split between + // ItemKind::Function and TypeKind::FunctionSig is totally worth it, but + // I guess we can try. + try_parse!(Function); + try_parse!(Var); + + // Types are sort of special, so to avoid parsing template classes + // twice, handle them separately. + { + let applicable_cursor = cursor.definition().unwrap_or(cursor); + match Self::from_ty(&applicable_cursor.cur_type(), + Some(applicable_cursor), + parent_id, + ctx) { + Ok(ty) => return Ok(ty), + Err(ParseError::Recurse) => return Err(ParseError::Recurse), + Err(ParseError::Continue) => {} + } + } + + // Guess how does clang treat extern "C" blocks? + if cursor.kind() == CXCursor_UnexposedDecl { + Err(ParseError::Recurse) + } else { + // We whitelist cursors here known to be unhandled, to prevent being + // too noisy about this. + match cursor.kind() { + CXCursor_MacroDefinition | + CXCursor_MacroExpansion | + CXCursor_UsingDeclaration | + CXCursor_StaticAssert | + CXCursor_InclusionDirective => { + debug!("Unhandled cursor kind {:?}: {:?}", + cursor.kind(), + cursor); + } + _ => { + error!("Unhandled cursor kind {:?}: {:?}", + cursor.kind(), + cursor); + } + } + + Err(ParseError::Continue) + } + } + + fn from_ty_or_ref(ty: clang::Type, + location: Option<clang::Cursor>, + parent_id: Option<ItemId>, + ctx: &mut BindgenContext) + -> ItemId { + let id = ctx.next_item_id(); + Self::from_ty_or_ref_with_id(id, + ty, + location, + parent_id, + ctx) + } + + /// Parse a C++ type. If we find a reference to a type that has not been + /// defined yet, use `UnresolvedTypeRef` as a placeholder. + /// + /// This logic is needed to avoid parsing items with the incorrect parent + /// and it's sort of complex to explain, so I'll just point to + /// `tests/headers/typeref.hpp` to see the kind of constructs that forced + /// this. + /// + /// Typerefs are resolved once parsing is completely done, see + /// `BindgenContext::resolve_typerefs`. + fn from_ty_or_ref_with_id(potential_id: ItemId, + ty: clang::Type, + location: Option<clang::Cursor>, + parent_id: Option<ItemId>, + ctx: &mut BindgenContext) + -> ItemId { + debug!("from_ty_or_ref_with_id: {:?} {:?}, {:?}, {:?}", + potential_id, + ty, + location, + parent_id); + + if ctx.collected_typerefs() { + debug!("refs already collected, resolving directly"); + return Self::from_ty_with_id(potential_id, + &ty, + location, + parent_id, + ctx) + .expect("Unable to resolve type"); + } + + if let Some(ty) = ctx.builtin_or_resolved_ty(potential_id, + parent_id, &ty, + location) { + debug!("{:?} already resolved: {:?}", ty, location); + return ty; + } + + debug!("New unresolved type reference: {:?}, {:?}", ty, location); + + let is_const = ty.is_const(); + let kind = TypeKind::UnresolvedTypeRef(ty, location, parent_id); + let current_module = ctx.current_module(); + ctx.add_item(Item::new(potential_id, + None, + None, + parent_id.unwrap_or(current_module), + ItemKind::Type(Type::new(None, + None, + kind, + is_const))), + Some(clang::Cursor::null()), + None); + potential_id + } + + + fn from_ty(ty: &clang::Type, + location: Option<clang::Cursor>, + parent_id: Option<ItemId>, + ctx: &mut BindgenContext) + -> Result<ItemId, ParseError> { + let id = ctx.next_item_id(); + Self::from_ty_with_id(id, ty, location, parent_id, ctx) + } + + /// This is one of the trickiest methods you'll find (probably along with + /// some of the ones that handle templates in `BindgenContext`). + /// + /// This method parses a type, given the potential id of that type (if + /// parsing it was correct), an optional location we're scanning, which is + /// critical some times to obtain information, an optional parent item id, + /// that will, if it's `None`, become the current module id, and the + /// context. + fn from_ty_with_id(id: ItemId, + ty: &clang::Type, + location: Option<clang::Cursor>, + parent_id: Option<ItemId>, + ctx: &mut BindgenContext) + -> Result<ItemId, ParseError> { + use clangll::*; + + let decl = { + let decl = ty.declaration(); + decl.definition().unwrap_or(decl) + }; + + let comment = decl.raw_comment() + .or_else(|| location.as_ref().and_then(|l| l.raw_comment())); + let annotations = Annotations::new(&decl) + .or_else(|| location.as_ref().and_then(|l| Annotations::new(l))); + + if let Some(ref annotations) = annotations { + if let Some(ref replaced) = annotations.use_instead_of() { + ctx.replace(replaced, id); + } + } + + if let Some(ty) = + ctx.builtin_or_resolved_ty(id, parent_id, ty, location) { + return Ok(ty); + } + + // First, check we're not recursing. + let mut valid_decl = decl.kind() != CXCursor_NoDeclFound; + let declaration_to_look_for = if valid_decl { + decl.canonical() + } else if location.is_some() && + location.unwrap().kind() == + CXCursor_ClassTemplate { + valid_decl = true; + location.unwrap() + } else { + decl + }; + + if valid_decl { + if let Some(&(_, item_id)) = ctx.currently_parsed_types + .iter() + .find(|&&(d, _)| d == declaration_to_look_for) { + debug!("Avoiding recursion parsing type: {:?}", ty); + return Ok(item_id); + } + } + + let current_module = ctx.current_module(); + if valid_decl { + ctx.currently_parsed_types.push((declaration_to_look_for, id)); + } + + let result = Type::from_clang_ty(id, ty, location, parent_id, ctx); + let relevant_parent_id = parent_id.unwrap_or(current_module); + let ret = match result { + Ok(ParseResult::AlreadyResolved(ty)) => Ok(ty), + Ok(ParseResult::New(item, declaration)) => { + ctx.add_item(Item::new(id, + comment, + annotations, + relevant_parent_id, + ItemKind::Type(item)), + declaration, + location); + Ok(id) + } + Err(ParseError::Continue) => Err(ParseError::Continue), + Err(ParseError::Recurse) => { + debug!("Item::from_ty recursing in the ast"); + let mut result = Err(ParseError::Recurse); + if let Some(ref location) = location { + // Need to pop here, otherwise we'll get stuck. + // + // TODO: Find a nicer interface, really. Also, the + // declaration_to_look_for suspiciously shares a lot of + // logic with ir::context, so we should refactor that. + if valid_decl { + let (popped_decl, _) = + ctx.currently_parsed_types.pop().unwrap(); + assert_eq!(popped_decl, declaration_to_look_for); + } + + location.visit(|cur| { + use clangll::*; + result = Item::from_ty_with_id(id, + ty, + Some(cur), + parent_id, + ctx); + match result { + Ok(..) => CXChildVisit_Break, + Err(ParseError::Recurse) => CXChildVisit_Recurse, + Err(ParseError::Continue) => CXChildVisit_Continue, + } + }); + + if valid_decl { + ctx.currently_parsed_types + .push((declaration_to_look_for, id)); + } + } + // If we have recursed into the AST all we know, and we still + // haven't found what we've got, let's just make a named type. + // + // This is what happens with some template members, for example. + // + // FIXME: Maybe we should restrict this to things with parent? + // It's harmless, but if we restrict that, then + // tests/headers/nsStyleAutoArray.hpp crashes. + if let Err(ParseError::Recurse) = result { + warn!("Unknown type, assuming named template type: id = {:?}; spelling = {}", + id, + ty.spelling()); + Ok(Self::named_type_with_id(id, + ty.spelling(), + None, + relevant_parent_id, + ctx)) + } else { + result + } + } + }; + + if valid_decl { + let (popped_decl, _) = ctx.currently_parsed_types.pop().unwrap(); + assert_eq!(popped_decl, declaration_to_look_for); + } + + ret + } + + /// A named type is a template parameter, e.g., the "T" in Foo<T>. They're + /// always local so it's the only exception when there's no declaration for + /// a type. + /// + /// It must have an id, and must not be the current module id. Ideally we + /// could assert the parent id is a Comp(..) type, but that info isn't + /// available yet. + fn named_type_with_id<S>(id: ItemId, + name: S, + default: Option<ItemId>, + parent_id: ItemId, + ctx: &mut BindgenContext) + -> ItemId + where S: Into<String>, + { + // see tests/headers/const_tparam.hpp + // and tests/headers/variadic_tname.hpp + let name = name.into().replace("const ", "").replace(".", ""); + + ctx.add_item(Item::new(id, + None, + None, + parent_id, + ItemKind::Type(Type::named(name, default))), + None, + None); + + id + } + + fn named_type<S>(name: S, + default: Option<ItemId>, + parent_id: ItemId, + ctx: &mut BindgenContext) + -> ItemId + where S: Into<String>, + { + let id = ctx.next_item_id(); + Self::named_type_with_id(id, name, default, parent_id, ctx) + } +} + +impl ItemCanonicalName for Item { + fn canonical_name(&self, ctx: &BindgenContext) -> String { + debug_assert!(ctx.in_codegen_phase(), + "You're not supposed to call this yet"); + if let Some(other_canon_type) = self.annotations.use_instead_of() { + return other_canon_type.to_owned(); + } + if self.canonical_name_cache.borrow().is_none() { + *self.canonical_name_cache.borrow_mut() = + Some(self.real_canonical_name(ctx, + ctx.options() + .enable_cxx_namespaces, + false)); + } + return self.canonical_name_cache.borrow().as_ref().unwrap().clone(); + } +} + +impl ItemCanonicalPath for Item { + fn canonical_path(&self, ctx: &BindgenContext) -> Vec<String> { + if !ctx.options().enable_cxx_namespaces { + return vec![self.canonical_name(ctx)]; + } + + if self.id() == ctx.root_module() { + match self.kind { + ItemKind::Module(ref module) => { + return vec![module.name().unwrap().into()] + } + _ => panic!("Something has wrong horribly wrong"), + } + } + + // TODO: This duplicates too much logic with real_canonical_name. + if let ItemKind::Type(ref ty) = *self.kind() { + match *ty.kind() { + TypeKind::Comp(ref ci) if ci.is_template_specialization() => { + return ci.specialized_template() + .unwrap() + .canonical_path(ctx); + } + TypeKind::ResolvedTypeRef(inner) | + TypeKind::TemplateRef(inner, _) => { + return inner.canonical_path(ctx); + } + TypeKind::Named(ref name, _) => { + return vec![name.clone()]; + } + _ => {} + } + } + + let mut parent_path = self.parent_id().canonical_path(&ctx); + if parent_path.last() + .map_or(false, |parent_name| parent_name.is_empty()) { + // This only happens (or should only happen) when we're an alias, + // and our parent is a templated alias, in which case the last + // component of the path will be empty. + let is_alias = match *self.expect_type().kind() { + TypeKind::Alias(..) => true, + _ => false, + }; + debug_assert!(is_alias, "How can this ever happen?"); + parent_path.pop().unwrap(); + } + parent_path.push(self.real_canonical_name(ctx, true, false)); + + parent_path + } +} diff --git a/libbindgen/src/ir/item_kind.rs b/libbindgen/src/ir/item_kind.rs new file mode 100644 index 00000000..d9e4690c --- /dev/null +++ b/libbindgen/src/ir/item_kind.rs @@ -0,0 +1,114 @@ +//! Different variants of an `Item` in our intermediate representation. + +use super::function::Function; +use super::module::Module; +use super::ty::Type; +use super::var::Var; + +/// A item we parse and translate. +#[derive(Debug)] +pub enum ItemKind { + /// A module, created implicitly once (the root module), or via C++ + /// namespaces. + Module(Module), + + /// A type declared in any of the multiple ways it can be declared. + Type(Type), + + /// A function or method declaration. + Function(Function), + + /// A variable declaration, most likely a static. + Var(Var), +} + +impl ItemKind { + /// Get a reference to this `ItemKind`'s underying `Module`, or `None` if it + /// is some other kind. + pub fn as_module(&self) -> Option<&Module> { + match *self { + ItemKind::Module(ref module) => Some(module), + _ => None, + } + } + + /// Is this a module? + pub fn is_module(&self) -> bool { + self.as_module().is_some() + } + + /// Get a reference to this `ItemKind`'s underying `Module`, or panic if it + /// is some other kind. + pub fn expect_module(&self) -> &Module { + self.as_module().expect("Not a module") + } + + /// Get a reference to this `ItemKind`'s underying `Function`, or `None` if + /// it is some other kind. + pub fn as_function(&self) -> Option<&Function> { + match *self { + ItemKind::Function(ref func) => Some(func), + _ => None, + } + } + + /// Is this a function? + pub fn is_function(&self) -> bool { + self.as_function().is_some() + } + + /// Get a reference to this `ItemKind`'s underying `Function`, or panic if + /// it is some other kind. + pub fn expect_function(&self) -> &Function { + self.as_function().expect("Not a function") + } + + /// Get a reference to this `ItemKind`'s underying `Type`, or `None` if + /// it is some other kind. + pub fn as_type(&self) -> Option<&Type> { + match *self { + ItemKind::Type(ref ty) => Some(ty), + _ => None, + } + } + + /// Get a mutable reference to this `ItemKind`'s underying `Type`, or `None` + /// if it is some other kind. + pub fn as_type_mut(&mut self) -> Option<&mut Type> { + match *self { + ItemKind::Type(ref mut ty) => Some(ty), + _ => None, + } + } + + /// Is this a type? + pub fn is_type(&self) -> bool { + self.as_type().is_some() + } + + /// Get a reference to this `ItemKind`'s underying `Type`, or panic if it is + /// some other kind. + pub fn expect_type(&self) -> &Type { + self.as_type().expect("Not a type") + } + + /// Get a reference to this `ItemKind`'s underying `Var`, or `None` if it is + /// some other kind. + pub fn as_var(&self) -> Option<&Var> { + match *self { + ItemKind::Var(ref v) => Some(v), + _ => None, + } + } + + /// Is this a variable? + pub fn is_var(&self) -> bool { + self.as_var().is_some() + } + + /// Get a reference to this `ItemKind`'s underying `Var`, or panic if it is + /// some other kind. + pub fn expect_var(&self) -> &Var { + self.as_var().expect("Not a var") + } +} diff --git a/libbindgen/src/ir/layout.rs b/libbindgen/src/ir/layout.rs new file mode 100644 index 00000000..3ac4a5f4 --- /dev/null +++ b/libbindgen/src/ir/layout.rs @@ -0,0 +1,34 @@ +//! Intermediate representation for the physical layout of some type. + +/// A type that represents the struct layout of a type. +#[derive(Debug, Clone, Copy)] +pub struct Layout { + /// The size (in bytes) of this layout. + pub size: usize, + /// The alignment (in bytes) of this layout. + pub align: usize, + /// Whether this layout's members are packed or not. + pub packed: bool, +} + +impl Layout { + /// Construct a new `Layout` with the given `size` and `align`. It is not + /// packed. + pub fn new(size: usize, align: usize) -> Self { + Layout { + size: size, + align: align, + packed: false, + } + } + + /// Is this a zero-sized layout? + pub fn is_zero(&self) -> bool { + self.size == 0 && self.align == 0 + } + + /// Construct a zero-sized layout. + pub fn zero() -> Self { + Self::new(0, 0) + } +} diff --git a/libbindgen/src/ir/mod.rs b/libbindgen/src/ir/mod.rs new file mode 100644 index 00000000..3c658a4a --- /dev/null +++ b/libbindgen/src/ir/mod.rs @@ -0,0 +1,18 @@ +//! The ir module defines bindgen's intermediate representation. +//! +//! Parsing C/C++ generates the IR, while code generation outputs Rust code from +//! the IR. + +pub mod annotations; +pub mod comp; +pub mod context; +pub mod enum_ty; +pub mod function; +pub mod int; +pub mod item; +pub mod item_kind; +pub mod layout; +pub mod module; +pub mod ty; +pub mod type_collector; +pub mod var; diff --git a/libbindgen/src/ir/module.rs b/libbindgen/src/ir/module.rs new file mode 100644 index 00000000..c5d8cfa7 --- /dev/null +++ b/libbindgen/src/ir/module.rs @@ -0,0 +1,61 @@ +//! Intermediate representation for modules (AKA C++ namespaces). + +use clang; +use parse::{ClangSubItemParser, ParseError, ParseResult}; +use parse_one; +use super::context::{BindgenContext, ItemId}; + +/// A module, as in, a C++ namespace. +#[derive(Clone, Debug)] +pub struct Module { + /// The name of the module, or none if it's anonymous. + name: Option<String>, + /// The children of this module, just here for convenience. + children_ids: Vec<ItemId>, +} + +impl Module { + /// Construct a new `Module`. + pub fn new(name: Option<String>) -> Self { + Module { + name: name, + children_ids: vec![], + } + } + + /// Get this module's name. + pub fn name(&self) -> Option<&str> { + self.name.as_ref().map(|n| &**n) + } + + /// Get a mutable reference to this module's children. + pub fn children_mut(&mut self) -> &mut Vec<ItemId> { + &mut self.children_ids + } + + /// Get this module's children. + pub fn children(&self) -> &[ItemId] { + &self.children_ids + } +} + +impl ClangSubItemParser for Module { + fn parse(cursor: clang::Cursor, + ctx: &mut BindgenContext) + -> Result<ParseResult<Self>, ParseError> { + use clangll::*; + match cursor.kind() { + CXCursor_Namespace => { + let module_id = ctx.module(cursor); + ctx.with_module(module_id, |ctx, children| { + cursor.visit(|cursor| { + parse_one(ctx, cursor, Some(module_id), children) + }) + }); + + Ok(ParseResult::AlreadyResolved(module_id)) + } + _ => Err(ParseError::Continue), + } + } +} diff --git a/libbindgen/src/ir/ty.rs b/libbindgen/src/ir/ty.rs new file mode 100644 index 00000000..34af2db5 --- /dev/null +++ b/libbindgen/src/ir/ty.rs @@ -0,0 +1,869 @@ +//! Everything related to types in our intermediate representation. + +use clang::{self, Cursor}; +use parse::{ClangItemParser, ParseError, ParseResult}; +use super::comp::CompInfo; +use super::context::{BindgenContext, ItemId}; +use super::enum_ty::Enum; +use super::function::FunctionSig; +use super::int::IntKind; +use super::item::Item; +use super::layout::Layout; +use super::type_collector::{ItemSet, TypeCollector}; + +/// The base representation of a type in bindgen. +/// +/// A type has an optional name, which if present cannot be empty, a `layout` +/// (size, alignment and packedness) if known, a `Kind`, which determines which +/// kind of type it is, and whether the type is const. +#[derive(Debug)] +pub struct Type { + /// The name of the type, or None if it was an unnamed struct or union. + name: Option<String>, + /// The layout of the type, if known. + layout: Option<Layout>, + /// The inner kind of the type + kind: TypeKind, + /// Whether this type is const-qualified. + is_const: bool, +} + +/// The maximum number of items in an array for which Rust implements common +/// traits, and so if we have a type containing an array with more than this +/// many items, we won't be able to derive common traits on that type. +/// +/// We need type-level integers yesterday :'( +pub const RUST_DERIVE_IN_ARRAY_LIMIT: usize = 32; + +impl Type { + /// Get the underlying `CompInfo` for this type, or `None` if this is some + /// other kind of type. + pub fn as_comp(&self) -> Option<&CompInfo> { + match self.kind { + TypeKind::Comp(ref ci) => Some(ci), + _ => None, + } + } + + /// Construct a new `Type`. + pub fn new(name: Option<String>, + layout: Option<Layout>, + kind: TypeKind, + is_const: bool) + -> Self { + Type { + name: name, + layout: layout, + kind: kind, + is_const: is_const, + } + } + + /// Which kind of type is this? + pub fn kind(&self) -> &TypeKind { + &self.kind + } + + /// Get a mutable reference to this type's kind. + pub fn kind_mut(&mut self) -> &mut TypeKind { + &mut self.kind + } + + /// Get this type's name. + pub fn name(&self) -> Option<&str> { + self.name.as_ref().map(|name| &**name) + } + + /// Is this a compound type? + pub fn is_comp(&self) -> bool { + match self.kind { + TypeKind::Comp(..) => true, + _ => false, + } + } + + /// Is this a named type? + pub fn is_named(&self) -> bool { + match self.kind { + TypeKind::Named(..) => true, + _ => false, + } + } + + /// Is this a function type? + pub fn is_function(&self) -> bool { + match self.kind { + TypeKind::Function(..) => true, + _ => false, + } + } + + /// Is this either a builtin or named type? + pub fn is_builtin_or_named(&self) -> bool { + match self.kind { + TypeKind::Void | + TypeKind::NullPtr | + TypeKind::Function(..) | + TypeKind::Array(..) | + TypeKind::Reference(..) | + TypeKind::Pointer(..) | + TypeKind::BlockPointer | + TypeKind::Int(..) | + TypeKind::Float(..) | + TypeKind::Named(..) => true, + _ => false, + } + } + + /// Creates a new named type, with name `name`. + pub fn named(name: String, default: Option<ItemId>) -> Self { + assert!(!name.is_empty()); + // TODO: stop duplicating the name, it's stupid. + let kind = TypeKind::Named(name.clone(), default); + Self::new(Some(name), None, kind, false) + } + + /// Is this an integer type? + pub fn is_integer(&self) -> bool { + match self.kind { + TypeKind::Int(..) => true, + _ => false, + } + } + + /// Is this a `const` qualified type? + pub fn is_const(&self) -> bool { + self.is_const + } + + /// Is this a reference to another type? + pub fn is_type_ref(&self) -> bool { + match self.kind { + TypeKind::ResolvedTypeRef(_) | + TypeKind::UnresolvedTypeRef(_, _, _) => true, + _ => false, + } + } + + /// What is the layout of this type? + pub fn layout(&self, ctx: &BindgenContext) -> Option<Layout> { + use std::mem; + + self.layout.or_else(|| { + match self.kind { + TypeKind::Comp(ref ci) => ci.layout(ctx), + // FIXME(emilio): This is a hack for anonymous union templates. + // Use the actual pointer size! + TypeKind::Pointer(..) | + TypeKind::BlockPointer => { + Some(Layout::new(mem::size_of::<*mut ()>(), + mem::align_of::<*mut ()>())) + } + TypeKind::ResolvedTypeRef(inner) => { + ctx.resolve_type(inner).layout(ctx) + } + _ => None, + } + }) + } + + /// Wether we can derive rust's `Debug` annotation in Rust. This should + /// ideally be a no-op that just returns `true`, but instead needs to be a + /// recursive method that checks whether all the proper members can derive + /// debug or not, because of the limit rust has on 32 items as max in the + /// array. + pub fn can_derive_debug(&self, ctx: &BindgenContext) -> bool { + match self.kind { + TypeKind::Array(t, len) => { + len <= RUST_DERIVE_IN_ARRAY_LIMIT && + ctx.resolve_type(t).can_derive_debug(ctx) + } + TypeKind::ResolvedTypeRef(t) | + TypeKind::TemplateAlias(t, _) | + TypeKind::Alias(_, t) => ctx.resolve_type(t).can_derive_debug(ctx), + TypeKind::Comp(ref info) => { + info.can_derive_debug(ctx, self.layout(ctx)) + } + _ => true, + } + } + + /// For some reason, deriving copies of an array of a type that is not known + /// to be copy is a compile error. e.g.: + /// + /// ```rust + /// #[derive(Copy, Clone)] + /// struct A<T> { + /// member: T, + /// } + /// ``` + /// + /// is fine, while: + /// + /// ```rust,ignore + /// #[derive(Copy, Clone)] + /// struct A<T> { + /// member: [T; 1], + /// } + /// ``` + /// + /// is an error. + /// + /// That's the whole point of the existence of `can_derive_copy_in_array`. + pub fn can_derive_copy_in_array(&self, + ctx: &BindgenContext, + item: &Item) + -> bool { + match self.kind { + TypeKind::ResolvedTypeRef(t) | + TypeKind::TemplateAlias(t, _) | + TypeKind::Alias(_, t) | + TypeKind::Array(t, _) => { + ctx.resolve_item(t) + .can_derive_copy_in_array(ctx) + } + TypeKind::Named(..) => false, + _ => self.can_derive_copy(ctx, item), + } + } + + /// Wether we'd be able to derive the `Copy` trait in Rust or not. Same + /// rationale than `can_derive_debug`. + pub fn can_derive_copy(&self, ctx: &BindgenContext, item: &Item) -> bool { + match self.kind { + TypeKind::Array(t, len) => { + len <= RUST_DERIVE_IN_ARRAY_LIMIT && + ctx.resolve_item(t).can_derive_copy_in_array(ctx) + } + TypeKind::ResolvedTypeRef(t) | + TypeKind::TemplateAlias(t, _) | + TypeKind::TemplateRef(t, _) | + TypeKind::Alias(_, t) => ctx.resolve_item(t).can_derive_copy(ctx), + TypeKind::Comp(ref info) => info.can_derive_copy(ctx, item), + _ => true, + } + } + + /// Whether this type has a vtable. + pub fn has_vtable(&self, ctx: &BindgenContext) -> bool { + // FIXME: Can we do something about template parameters? Huh... + match self.kind { + TypeKind::TemplateRef(t, _) | + TypeKind::TemplateAlias(t, _) | + TypeKind::Alias(_, t) | + TypeKind::ResolvedTypeRef(t) => ctx.resolve_type(t).has_vtable(ctx), + TypeKind::Comp(ref info) => info.has_vtable(ctx), + _ => false, + } + + } + + /// Returns whether this type has a destructor. + pub fn has_destructor(&self, ctx: &BindgenContext) -> bool { + match self.kind { + TypeKind::TemplateRef(t, _) | + TypeKind::TemplateAlias(t, _) | + TypeKind::Alias(_, t) | + TypeKind::ResolvedTypeRef(t) => { + ctx.resolve_type(t).has_destructor(ctx) + } + TypeKind::Comp(ref info) => info.has_destructor(ctx), + _ => false, + } + } + + /// See the comment in `Item::signature_contains_named_type`. + pub fn signature_contains_named_type(&self, + ctx: &BindgenContext, + ty: &Type) + -> bool { + debug_assert!(ty.is_named()); + let name = match *ty.kind() { + TypeKind::Named(ref name, _) => name, + _ => unreachable!(), + }; + + match self.kind { + TypeKind::Named(ref this_name, _) => this_name == name, + TypeKind::ResolvedTypeRef(t) | + TypeKind::Array(t, _) | + TypeKind::Pointer(t) | + TypeKind::Alias(_, t) => { + ctx.resolve_type(t) + .signature_contains_named_type(ctx, ty) + } + TypeKind::Function(ref sig) => { + sig.argument_types().iter().any(|&(_, arg)| { + ctx.resolve_type(arg) + .signature_contains_named_type(ctx, ty) + }) || + ctx.resolve_type(sig.return_type()) + .signature_contains_named_type(ctx, ty) + } + TypeKind::TemplateAlias(_, ref template_args) | + TypeKind::TemplateRef(_, ref template_args) => { + template_args.iter().any(|arg| { + ctx.resolve_type(*arg) + .signature_contains_named_type(ctx, ty) + }) + } + TypeKind::Comp(ref ci) => ci.signature_contains_named_type(ctx, ty), + _ => false, + } + } + + /// See safe_canonical_type. + pub fn canonical_type<'tr>(&'tr self, + ctx: &'tr BindgenContext) + -> &'tr Type { + self.safe_canonical_type(ctx) + .expect("Should have been resolved after parsing!") + } + + /// Returns the canonical type of this type, that is, the "inner type". + /// + /// For example, for a `typedef`, the canonical type would be the + /// `typedef`ed type, for a template specialization, would be the template + /// its specializing, and so on. Return None if the type is unresolved. + pub fn safe_canonical_type<'tr>(&'tr self, + ctx: &'tr BindgenContext) + -> Option<&'tr Type> { + match self.kind { + TypeKind::Named(..) | + TypeKind::Array(..) | + TypeKind::Comp(..) | + TypeKind::Int(..) | + TypeKind::Float(..) | + TypeKind::Complex(..) | + TypeKind::Function(..) | + TypeKind::Enum(..) | + TypeKind::Reference(..) | + TypeKind::Void | + TypeKind::NullPtr | + TypeKind::BlockPointer | + TypeKind::Pointer(..) => Some(self), + + TypeKind::ResolvedTypeRef(inner) | + TypeKind::Alias(_, inner) | + TypeKind::TemplateAlias(inner, _) | + TypeKind::TemplateRef(inner, _) => { + ctx.resolve_type(inner).safe_canonical_type(ctx) + } + + TypeKind::UnresolvedTypeRef(..) => None, + } + } +} + +/// The kind of float this type represents. +#[derive(Debug, Copy, Clone, PartialEq)] +pub enum FloatKind { + /// A `float`. + Float, + /// A `double`. + Double, + /// A `long double`. + LongDouble, + /// A `__float128`. + Float128, +} + +/// The different kinds of types that we can parse. +#[derive(Debug)] +pub enum TypeKind { + /// The void type. + Void, + + /// The `nullptr_t` type. + NullPtr, + + /// A compound type, that is, a class, struct, or union. + Comp(CompInfo), + + /// An integer type, of a given kind. `bool` and `char` are also considered + /// integers. + Int(IntKind), + + /// A floating point type. + Float(FloatKind), + + /// A complex floating point type. + Complex(FloatKind), + + /// A type alias, with a name, that points to another type. + Alias(String, ItemId), + + /// A templated alias, pointing to an inner `Alias` type, with template + /// parameters. + TemplateAlias(ItemId, Vec<ItemId>), + + /// An array of a type and a lenght. + Array(ItemId, usize), + + /// A function type, with a given signature. + Function(FunctionSig), + + /// An `enum` type. + Enum(Enum), + + /// A pointer to a type. The bool field represents whether it's const or + /// not. + Pointer(ItemId), + + /// A pointer to an Apple block. + BlockPointer, + + /// A reference to a type, as in: int& foo(). + Reference(ItemId), + + /// A reference to a template, with different template parameter names. To + /// see why this is needed, check out the creation of this variant in + /// `Type::from_clang_ty`. + TemplateRef(ItemId, Vec<ItemId>), + + /// A reference to a yet-to-resolve type. This stores the clang cursor + /// itself, and postpones its resolution. + /// + /// These are gone in a phase after parsing where these are mapped to + /// already known types, and are converted to ResolvedTypeRef. + /// + /// see tests/headers/typeref.hpp to see somewhere where this is a problem. + UnresolvedTypeRef(clang::Type, + Option<clang::Cursor>, + /* parent_id */ + Option<ItemId>), + + /// An indirection to another type. + /// + /// These are generated after we resolve a forward declaration, or when we + /// replace one type with another. + ResolvedTypeRef(ItemId), + + /// A named type, that is, a template parameter, with an optional default + /// type. + Named(String, Option<ItemId>), +} + +impl Type { + /// Whether this type is unsized, that is, has no members. This is used to + /// derive whether we should generate a dummy `_address` field for structs, + /// to comply to the C and C++ layouts, that specify that every type needs + /// to be addressable. + pub fn is_unsized(&self, ctx: &BindgenContext) -> bool { + debug_assert!(ctx.in_codegen_phase(), "Not yet"); + + match self.kind { + TypeKind::Void => true, + TypeKind::Comp(ref ci) => ci.is_unsized(ctx), + TypeKind::Array(inner, size) => { + size == 0 || ctx.resolve_type(inner).is_unsized(ctx) + } + TypeKind::ResolvedTypeRef(inner) | + TypeKind::Alias(_, inner) | + TypeKind::TemplateAlias(inner, _) | + TypeKind::TemplateRef(inner, _) => { + ctx.resolve_type(inner).is_unsized(ctx) + } + TypeKind::Named(..) | + TypeKind::Int(..) | + TypeKind::Float(..) | + TypeKind::Complex(..) | + TypeKind::Function(..) | + TypeKind::Enum(..) | + TypeKind::Reference(..) | + TypeKind::NullPtr | + TypeKind::BlockPointer | + TypeKind::Pointer(..) => false, + + TypeKind::UnresolvedTypeRef(..) => { + unreachable!("Should have been resolved after parsing!"); + } + } + } + + /// This is another of the nasty methods. This one is the one that takes + /// care of the core logic of converting a clang type to a `Type`. + /// + /// It's sort of nasty and full of special-casing, but hopefully the + /// comments in every special case justify why they're there. + pub fn from_clang_ty(potential_id: ItemId, + ty: &clang::Type, + location: Option<Cursor>, + parent_id: Option<ItemId>, + ctx: &mut BindgenContext) + -> Result<ParseResult<Self>, ParseError> { + use clangll::*; + { + let already_resolved = + ctx.builtin_or_resolved_ty(potential_id, + parent_id, + ty, + location); + if let Some(ty) = already_resolved { + debug!("{:?} already resolved: {:?}", ty, location); + return Ok(ParseResult::AlreadyResolved(ty)); + } + } + + let layout = ty.fallible_layout().ok(); + let cursor = ty.declaration(); + let mut name = cursor.spelling(); + + debug!("from_clang_ty: {:?}, ty: {:?}, loc: {:?}", + potential_id, + ty, + location); + debug!("currently_parsed_types: {:?}", ctx.currently_parsed_types); + + let canonical_ty = ty.canonical_type(); + let kind = match ty.kind() { + CXType_Unexposed if *ty != canonical_ty && + canonical_ty.kind() != CXType_Invalid => { + debug!("Looking for canonical type: {:?}", canonical_ty); + return Self::from_clang_ty(potential_id, + &canonical_ty, + location, + parent_id, + ctx); + } + CXType_Unexposed | CXType_Invalid => { + // For some reason Clang doesn't give us any hint in some + // situations where we should generate a function pointer (see + // tests/headers/func_ptr_in_struct.h), so we do a guess here + // trying to see if it has a valid return type. + if ty.ret_type().is_some() { + let signature = try!(FunctionSig::from_ty(ty, + &location.unwrap_or(cursor), + ctx)); + TypeKind::Function(signature) + // Same here, with template specialisations we can safely + // assume this is a Comp(..) + } else if ty.template_args().map_or(false, |x| x.len() > 0) { + debug!("Template specialization: {:?}", ty); + let complex = + CompInfo::from_ty(potential_id, ty, location, ctx) + .expect("C'mon"); + TypeKind::Comp(complex) + } else if let Some(location) = location { + match location.kind() { + CXCursor_ClassTemplatePartialSpecialization | + CXCursor_CXXBaseSpecifier | + CXCursor_ClassTemplate => { + if location.kind() == CXCursor_CXXBaseSpecifier { + // In the case we're parsing a base specifier + // inside an unexposed or invalid type, it means + // that we're parsing one of two things: + // + // * A template parameter. + // * A complex class that isn't exposed. + // + // This means, unfortunately, that there's no + // good way to differentiate between them. + // + // Probably we could try to look at the + // declaration and complicate more this logic, + // but we'll keep it simple... if it's a valid + // C++ identifier, we'll consider it as a + // template parameter. + // + // This is because: + // + // * We expect every other base that is a + // proper identifier (that is, a simple + // struct/union declaration), to be exposed, + // so this path can't be reached in that + // case. + // + // * Quite conveniently, complex base + // specifiers preserve their full names (that + // is: Foo<T> instead of Foo). We can take + // advantage of this. + // + // If we find some edge case where this doesn't + // work (which I guess is unlikely, see the + // different test cases[1][2][3][4]), we'd need + // to find more creative ways of differentiating + // these two cases. + // + // [1]: inherit_named.hpp + // [2]: forward-inherit-struct-with-fields.hpp + // [3]: forward-inherit-struct.hpp + // [4]: inherit-namespaced.hpp + if location.spelling() + .chars() + .all(|c| c.is_alphanumeric() || c == '_') { + return Err(ParseError::Recurse); + } + } else { + name = location.spelling(); + } + let complex = CompInfo::from_ty(potential_id, + ty, + Some(location), + ctx) + .expect("C'mon"); + TypeKind::Comp(complex) + } + CXCursor_TypeAliasTemplateDecl => { + debug!("TypeAliasTemplateDecl"); + + // We need to manually unwind this one. + let mut inner = Err(ParseError::Continue); + let mut args = vec![]; + + location.visit(|cur| { + match cur.kind() { + CXCursor_TypeAliasDecl => { + debug_assert!(cur.cur_type().kind() == + CXType_Typedef); + inner = + Item::from_ty(&cur.cur_type(), + Some(cur), + Some(potential_id), + ctx); + } + CXCursor_TemplateTypeParameter => { + // See the comment in src/ir/comp.rs + // about the same situation. + if cur.spelling().is_empty() { + return CXChildVisit_Continue; + } + + let default_type = + Item::from_ty(&cur.cur_type(), + Some(cur), + Some(potential_id), + ctx) + .ok(); + let param = + Item::named_type(cur.spelling(), + default_type, + potential_id, + ctx); + args.push(param); + } + _ => {} + } + CXChildVisit_Continue + }); + + if inner.is_err() { + error!("Failed to parse templated alias {:?}", + location); + return Err(ParseError::Continue); + } + + // NB: `args` may be empty here (if for example the + // template parameters are constants). + // + // We can't reject it here then because inner points + // to `potential_id` now, so either we remove + // `inner` and return an error, or carry on. + // + // In this case, we just carry on, since it seems + // easier if than removing every possible reference + // to `item` from `ctx`, and it doesn't give any + // problems that we didn't have anyway. + TypeKind::TemplateAlias(inner.unwrap(), args) + } + CXCursor_TemplateRef => { + let referenced = location.referenced().expect("expected value, got none"); + let referenced_ty = referenced.cur_type(); + let referenced_declaration = + Some(referenced_ty.declaration()); + + return Self::from_clang_ty(potential_id, + &referenced_ty, + referenced_declaration, + parent_id, + ctx); + } + CXCursor_TypeRef => { + let referenced = location.referenced().expect("expected value, got none"); + let referenced_ty = referenced.cur_type(); + let referenced_declaration = + Some(referenced_ty.declaration()); + + let item = + Item::from_ty_or_ref_with_id( + potential_id, + referenced_ty, + referenced_declaration, + parent_id, + ctx); + return Ok(ParseResult::AlreadyResolved(item)); + } + _ => { + if ty.kind() == CXType_Unexposed { + warn!("Unexposed type {:?}, recursing inside, \ + loc: {:?}", + ty, + location); + return Err(ParseError::Recurse); + } + + // If the type name is empty we're probably + // over-recursing to find a template parameter name + // or something like that, so just don't be too + // noisy with it since it causes confusion, see for + // example the discussion in: + // + // https://github.com/jamesmunns/teensy3-rs/issues/9 + if !ty.spelling().is_empty() { + error!("invalid type {:?}", ty); + } else { + warn!("invalid type {:?}", ty); + } + return Err(ParseError::Continue); + } + } + } else { + // TODO: Don't duplicate this! + if ty.kind() == CXType_Unexposed { + warn!("Unexposed type {:?}, recursing inside", ty); + return Err(ParseError::Recurse); + } + + if !ty.spelling().is_empty() { + error!("invalid type {:?}", ty); + } else { + warn!("invalid type {:?}", ty); + } + return Err(ParseError::Continue); + } + } + // NOTE: We don't resolve pointers eagerly because the pointee type + // might not have been parsed, and if it contains templates or + // something else we might get confused, see the comment inside + // TypeRef. + // + // We might need to, though, if the context is already in the + // process of resolving them. + CXType_MemberPointer | + CXType_Pointer => { + let inner = Item::from_ty_or_ref(ty.pointee_type().unwrap(), + location, + parent_id, + ctx); + TypeKind::Pointer(inner) + } + CXType_BlockPointer => TypeKind::BlockPointer, + // XXX: RValueReference is most likely wrong, but I don't think we + // can even add bindings for that, so huh. + CXType_RValueReference | + CXType_LValueReference => { + let inner = Item::from_ty_or_ref(ty.pointee_type().unwrap(), + location, + parent_id, + ctx); + TypeKind::Reference(inner) + } + // XXX DependentSizedArray is wrong + CXType_VariableArray | + CXType_DependentSizedArray | + CXType_IncompleteArray => { + let inner = Item::from_ty(ty.elem_type().as_ref().unwrap(), + location, + parent_id, + ctx) + .expect("Not able to resolve array element?"); + TypeKind::Pointer(inner) + } + CXType_FunctionNoProto | + CXType_FunctionProto => { + let signature = try!(FunctionSig::from_ty(ty, + &location.unwrap_or(cursor), + ctx)); + TypeKind::Function(signature) + } + CXType_Typedef => { + let inner = cursor.typedef_type(); + let inner = + Item::from_ty_or_ref(inner, location, parent_id, ctx); + TypeKind::Alias(ty.spelling(), inner) + } + CXType_Enum => { + let enum_ = Enum::from_ty(ty, ctx).expect("Not an enum?"); + TypeKind::Enum(enum_) + } + CXType_Record => { + let complex = + CompInfo::from_ty(potential_id, ty, location, ctx) + .expect("Not a complex type?"); + TypeKind::Comp(complex) + } + // FIXME: We stub vectors as arrays since in 99% of the cases the + // layout is going to be correct, and there's no way we can generate + // vector types properly in Rust for now. + // + // That being said, that should be fixed eventually. + CXType_Vector | + CXType_ConstantArray => { + let inner = Item::from_ty(ty.elem_type().as_ref().unwrap(), + location, + parent_id, + ctx) + .expect("Not able to resolve array element?"); + TypeKind::Array(inner, ty.num_elements().unwrap()) + } + #[cfg(not(feature="llvm_stable"))] + CXType_Elaborated => { + return Self::from_clang_ty(potential_id, + &ty.named(), + location, + parent_id, + ctx); + } + _ => { + error!("unsupported type: kind = {:?}; ty = {:?}; at {:?}", + ty.kind(), + ty, + location); + return Err(ParseError::Continue); + } + }; + + let name = if name.is_empty() { None } else { Some(name) }; + let is_const = ty.is_const(); + + let ty = Type::new(name, layout, kind, is_const); + // TODO: maybe declaration.canonical()? + Ok(ParseResult::New(ty, Some(cursor.canonical()))) + } +} + +impl TypeCollector for Type { + type Extra = Item; + + fn collect_types(&self, + context: &BindgenContext, + types: &mut ItemSet, + item: &Item) { + match *self.kind() { + TypeKind::Pointer(inner) | + TypeKind::Reference(inner) | + TypeKind::Array(inner, _) | + TypeKind::TemplateAlias(inner, _) | + TypeKind::Alias(_, inner) | + TypeKind::Named(_, Some(inner)) | + TypeKind::ResolvedTypeRef(inner) => { + types.insert(inner); + } + + TypeKind::TemplateRef(inner, ref template_args) => { + types.insert(inner); + for &item in template_args { + types.insert(item); + } + } + TypeKind::Comp(ref ci) => ci.collect_types(context, types, item), + TypeKind::Function(ref sig) => { + sig.collect_types(context, types, item) + } + // FIXME: Pending types! + ref other @ _ => { + debug!("<Type as TypeCollector>::collect_types: Ignoring: {:?}", other); + } + } + } +} diff --git a/libbindgen/src/ir/type_collector.rs b/libbindgen/src/ir/type_collector.rs new file mode 100644 index 00000000..0f10152d --- /dev/null +++ b/libbindgen/src/ir/type_collector.rs @@ -0,0 +1,22 @@ +//! Collecting type items. + +use std::collections::BTreeSet; +use super::context::{BindgenContext, ItemId}; + +/// A set of items. +pub type ItemSet = BTreeSet<ItemId>; + +/// Collect all the type items referenced by this item. +pub trait TypeCollector { + /// If a particular type needs extra information beyond what it has in + /// `self` and `context` to find its referenced type items, its + /// implementation can define this associated type, forcing callers to pass + /// the needed information through. + type Extra; + + /// Add each type item referenced by `self` into the `types` set. + fn collect_types(&self, + context: &BindgenContext, + types: &mut ItemSet, + extra: &Self::Extra); +} diff --git a/libbindgen/src/ir/var.rs b/libbindgen/src/ir/var.rs new file mode 100644 index 00000000..d0c4d9ca --- /dev/null +++ b/libbindgen/src/ir/var.rs @@ -0,0 +1,246 @@ +//! Intermediate representation of variables. + +use cexpr; +use clang; +use parse::{ClangItemParser, ClangSubItemParser, ParseError, ParseResult}; +use std::num::Wrapping; +use super::context::{BindgenContext, ItemId}; +use super::function::cursor_mangling; +use super::int::IntKind; +use super::item::Item; +use super::ty::TypeKind; + +/// A `Var` is our intermediate representation of a variable. +#[derive(Debug)] +pub struct Var { + /// The name of the variable. + name: String, + /// The mangled name of the variable. + mangled_name: Option<String>, + /// The type of the variable. + ty: ItemId, + /// TODO: support non-integer constants? + /// The integer value of the variable. + val: Option<i64>, + /// Whether this variable is const. + is_const: bool, +} + +impl Var { + /// Construct a new `Var`. + pub fn new(name: String, + mangled: Option<String>, + ty: ItemId, + val: Option<i64>, + is_const: bool) + -> Var { + assert!(!name.is_empty()); + Var { + name: name, + mangled_name: mangled, + ty: ty, + val: val, + is_const: is_const, + } + } + + /// Is this variable `const` qualified? + pub fn is_const(&self) -> bool { + self.is_const + } + + /// The value of this constant variable, if any. + pub fn val(&self) -> Option<i64> { + self.val + } + + /// Get this variable's type. + pub fn ty(&self) -> ItemId { + self.ty + } + + /// Get this variable's name. + pub fn name(&self) -> &str { + &self.name + } + + /// Get this variable's mangled name. + pub fn mangled_name(&self) -> Option<&str> { + self.mangled_name.as_ref().map(|n| &**n) + } +} + +impl ClangSubItemParser for Var { + fn parse(cursor: clang::Cursor, + ctx: &mut BindgenContext) + -> Result<ParseResult<Self>, ParseError> { + use clangll::*; + use cexpr::expr::EvalResult; + match cursor.kind() { + CXCursor_MacroDefinition => { + let value = parse_macro(ctx, &cursor, ctx.translation_unit()); + + let (id, value) = match value { + Some(v) => v, + None => return Err(ParseError::Continue), + }; + + assert!(!id.is_empty(), "Empty macro name?"); + + let previously_defined = ctx.parsed_macro(&id); + + // NB: It's important to "note" the macro even if the result is + // not an integer, otherwise we might loose other kind of + // derived macros. + ctx.note_parsed_macro(id.clone(), value.clone()); + + if previously_defined { + let name = String::from_utf8(id).unwrap(); + warn!("Duplicated macro definition: {}", name); + return Err(ParseError::Continue); + } + + // NOTE: Unwrapping, here and above, is safe, because the + // identifier of a token comes straight from clang, and we + // enforce utf8 there, so we should have already panicked at + // this point. + let name = String::from_utf8(id).unwrap(); + let (int_kind, val) = match value { + // TODO(emilio): Handle the non-invalid ones! + EvalResult::Float(..) | + EvalResult::Char(..) | + EvalResult::Str(..) | + EvalResult::Invalid => return Err(ParseError::Continue), + + EvalResult::Int(Wrapping(value)) => { + let kind = ctx.options() + .type_chooser + .as_ref() + .and_then(|c| c.int_macro(&name, value)) + .unwrap_or_else(|| { + if value < 0 { + if value < i32::min_value() as i64 { + IntKind::LongLong + } else { + IntKind::Int + } + } else if value > u32::max_value() as i64 { + IntKind::ULongLong + } else { + IntKind::UInt + } + }); + + (kind, value) + } + }; + + let ty = Item::builtin_type(TypeKind::Int(int_kind), true, ctx); + + Ok(ParseResult::New(Var::new(name, None, ty, Some(val), true), + Some(cursor))) + } + CXCursor_VarDecl => { + let name = cursor.spelling(); + if name.is_empty() { + warn!("Empty constant name?"); + return Err(ParseError::Continue); + } + + let ty = cursor.cur_type(); + + // XXX this is redundant, remove! + let is_const = ty.is_const(); + + let ty = Item::from_ty(&ty, Some(cursor), None, ctx) + .expect("Unable to resolve constant type?"); + + // Note: Ty might not be totally resolved yet, see + // tests/headers/inner_const.hpp + // + // That's fine because in that case we know it's not a literal. + let value = ctx.safe_resolve_type(ty) + .and_then(|t| t.safe_canonical_type(ctx)) + .and_then(|t| if t.is_integer() { Some(t) } else { None }) + .and_then(|_| { + get_integer_literal_from_cursor(&cursor, + ctx.translation_unit()) + }); + + let mangling = cursor_mangling(&cursor); + + let var = Var::new(name, mangling, ty, value, is_const); + Ok(ParseResult::New(var, Some(cursor))) + + } + _ => { + /* TODO */ + Err(ParseError::Continue) + } + } + } +} + +/// Try and parse a macro using all the macros parsed until now. +fn parse_macro(ctx: &BindgenContext, + cursor: &clang::Cursor, + unit: &clang::TranslationUnit) + -> Option<(Vec<u8>, cexpr::expr::EvalResult)> { + use cexpr::{expr, nom}; + + let cexpr_tokens = match unit.cexpr_tokens(cursor) { + None => return None, + Some(tokens) => tokens, + }; + + let parser = expr::IdentifierParser::new(ctx.parsed_macros()); + let result = parser.macro_definition(&cexpr_tokens); + + match result { + nom::IResult::Done(_, (id, val)) => Some((id.into(), val)), + _ => None, + } +} + +fn parse_int_literal_tokens(cursor: &clang::Cursor, + unit: &clang::TranslationUnit) + -> Option<i64> { + use cexpr::{expr, nom}; + use cexpr::expr::EvalResult; + + let cexpr_tokens = match unit.cexpr_tokens(cursor) { + None => return None, + Some(tokens) => tokens, + }; + + // TODO(emilio): We can try to parse other kinds of literals. + match expr::expr(&cexpr_tokens) { + nom::IResult::Done(_, EvalResult::Int(Wrapping(val))) => Some(val), + _ => None, + } +} + +fn get_integer_literal_from_cursor(cursor: &clang::Cursor, + unit: &clang::TranslationUnit) + -> Option<i64> { + use clangll::*; + let mut value = None; + cursor.visit(|c| { + match c.kind() { + CXCursor_IntegerLiteral | + CXCursor_UnaryOperator => { + value = parse_int_literal_tokens(&c, unit); + } + CXCursor_UnexposedExpr => { + value = get_integer_literal_from_cursor(&c, unit); + } + _ => (), + } + if value.is_some() { + CXChildVisit_Break + } else { + CXChildVisit_Continue + } + }); + value +} |