blob: e4b42dc5a36b9cecf55fbba727a167ed9dc99e45 [file] [log] [blame]
// Copyright 2017 The Go Authors. All rights reserved.
// Use of this src code is governed by a BSD-style
// license that can be found in the LICENSE file.
package types
import (
"bytes"
"fmt"
"go/ast"
"go/token"
)
// This file implements the collection of an interface's methods
// without relying on partially computed types of methods or interfaces
// for interface types declared at the package level.
//
// Because interfaces must not embed themselves, directly or indirectly,
// the method set of a valid interface can always be computed independent
// of any cycles that might exist via method signatures (see also issue #18395).
//
// Except for blank method name and interface cycle errors, no errors
// are reported. Affected methods or embedded interfaces are silently
// dropped. Subsequent type-checking of the interface will check
// signatures and embedded interfaces and report errors at that time.
//
// Only infoFromTypeLit should be called directly from code outside this file
// to compute an ifaceInfo.
// ifaceInfo describes the method set for an interface.
// The zero value for an ifaceInfo is a ready-to-use ifaceInfo representing
// the empty interface.
type ifaceInfo struct {
explicits int // number of explicitly declared methods
methods []*methodInfo // all methods, starting with explicitly declared ones in source order
}
// emptyIfaceInfo represents the ifaceInfo for the empty interface.
var emptyIfaceInfo ifaceInfo
func (info *ifaceInfo) String() string {
var buf bytes.Buffer
fmt.Fprintf(&buf, "interface{")
for i, m := range info.methods {
if i > 0 {
fmt.Fprint(&buf, " ")
}
fmt.Fprint(&buf, m)
}
fmt.Fprintf(&buf, "}")
return buf.String()
}
// methodInfo represents an interface method.
// At least one of src or fun must be non-nil.
// (Methods declared in the current package have a non-nil scope
// and src, and eventually a non-nil fun field; imported and pre-
// declared methods have a nil scope and src, and only a non-nil
// fun field.)
type methodInfo struct {
scope *Scope // scope of interface method; or nil
src *ast.Field // syntax tree representation of interface method; or nil
fun *Func // corresponding fully type-checked method type; or nil
}
func (info *methodInfo) String() string {
if info.fun != nil {
return info.fun.name
}
return info.src.Names[0].Name
}
func (info *methodInfo) Pos() token.Pos {
if info.fun != nil {
return info.fun.Pos()
}
return info.src.Pos()
}
func (info *methodInfo) id(pkg *Package) string {
if info.fun != nil {
return info.fun.Id()
}
return Id(pkg, info.src.Names[0].Name)
}
// A methodInfoSet maps method ids to methodInfos.
// It is used to determine duplicate declarations.
// (A methodInfo set is the equivalent of an objset
// but for methodInfos rather than Objects.)
type methodInfoSet map[string]*methodInfo
// insert attempts to insert an method m into the method set s.
// If s already contains an alternative method alt with
// the same name, insert leaves s unchanged and returns alt.
// Otherwise it inserts m and returns nil.
func (s *methodInfoSet) insert(pkg *Package, m *methodInfo) *methodInfo {
id := m.id(pkg)
if alt := (*s)[id]; alt != nil {
return alt
}
if *s == nil {
*s = make(methodInfoSet)
}
(*s)[id] = m
return nil
}
// like Checker.declareInSet but for method infos.
func (check *Checker) declareInMethodSet(mset *methodInfoSet, pos token.Pos, m *methodInfo) bool {
if alt := mset.insert(check.pkg, m); alt != nil {
check.errorf(pos, "%s redeclared", m)
check.reportAltMethod(alt)
return false
}
return true
}
// like Checker.reportAltDecl but for method infos.
func (check *Checker) reportAltMethod(m *methodInfo) {
if pos := m.Pos(); pos.IsValid() {
// We use "other" rather than "previous" here because
// the first declaration seen may not be textually
// earlier in the source.
check.errorf(pos, "\tother declaration of %s", m) // secondary error, \t indented
}
}
// infoFromTypeLit computes the method set for the given interface iface
// declared in scope.
// If a corresponding type name exists (tname != nil), it is used for
// cycle detection and to cache the method set.
// The result is the method set, or nil if there is a cycle via embedded
// interfaces. A non-nil result doesn't mean that there were no errors,
// but they were either reported (e.g., blank methods), or will be found
// (again) when computing the interface's type.
// If tname is not nil it must be the last element in path.
func (check *Checker) infoFromTypeLit(scope *Scope, iface *ast.InterfaceType, tname *TypeName, path []*TypeName) (info *ifaceInfo) {
assert(iface != nil)
// lazy-allocate interfaces map
if check.interfaces == nil {
check.interfaces = make(map[*TypeName]*ifaceInfo)
}
if trace {
check.trace(iface.Pos(), "-- collect methods for %v (path = %s, objPath = %s)", iface, pathString(path), check.pathString())
check.indent++
defer func() {
check.indent--
check.trace(iface.Pos(), "=> %s", info)
}()
}
// If the interface is named, check if we computed info already.
//
// This is not simply an optimization; we may run into stack
// overflow with recursive interface declarations. Example:
//
// type T interface {
// m() interface { T }
// }
//
// (Since recursive definitions can only be expressed via names,
// it is sufficient to track named interfaces here.)
//
// While at it, use the same mechanism to detect cycles. (We still
// have the path-based cycle check because we want to report the
// entire cycle if present.)
if tname != nil {
assert(path[len(path)-1] == tname) // tname must be last path element
var found bool
if info, found = check.interfaces[tname]; found {
if info == nil {
// We have a cycle and use check.cycle to report it.
// We are guaranteed that check.cycle also finds the
// cycle because when infoFromTypeLit is called, any
// tname that's already in check.interfaces was also
// added to the path. (But the converse is not true:
// A non-nil tname is always the last element in path.)
ok := check.cycle(tname, path, true)
assert(ok)
}
return
}
check.interfaces[tname] = nil // computation started but not complete
}
if iface.Methods.List == nil {
// fast track for empty interface
info = &emptyIfaceInfo
} else {
// (syntactically) non-empty interface
info = new(ifaceInfo)
// collect explicitly declared methods and embedded interfaces
var mset methodInfoSet
var embeddeds []*ifaceInfo
var positions []token.Pos // entries correspond to positions of embeddeds; used for error reporting
for _, f := range iface.Methods.List {
if len(f.Names) > 0 {
// We have a method with name f.Names[0].
// (The parser ensures that there's only one method
// and we don't care if a constructed AST has more.)
// spec: "As with all method sets, in an interface type,
// each method must have a unique non-blank name."
if name := f.Names[0]; name.Name == "_" {
check.errorf(name.Pos(), "invalid method name _")
continue // ignore
}
m := &methodInfo{scope: scope, src: f}
if check.declareInMethodSet(&mset, f.Pos(), m) {
info.methods = append(info.methods, m)
}
} else {
// We have an embedded interface and f.Type is its
// (possibly qualified) embedded type name. Collect
// it if it's a valid interface.
var e *ifaceInfo
switch ename := f.Type.(type) {
case *ast.Ident:
e = check.infoFromTypeName(scope, ename, path)
case *ast.SelectorExpr:
e = check.infoFromQualifiedTypeName(scope, ename)
default:
// The parser makes sure we only see one of the above.
// Constructed ASTs may contain other (invalid) nodes;
// we simply ignore them. The full type-checking pass
// will report those as errors later.
}
if e != nil {
embeddeds = append(embeddeds, e)
positions = append(positions, f.Type.Pos())
}
}
}
info.explicits = len(info.methods)
// collect methods of embedded interfaces
for i, e := range embeddeds {
pos := positions[i] // position of type name of embedded interface
for _, m := range e.methods {
if check.declareInMethodSet(&mset, pos, m) {
info.methods = append(info.methods, m)
}
}
}
}
// mark check.interfaces as complete
assert(info != nil)
if tname != nil {
check.interfaces[tname] = info
}
return
}
// infoFromTypeName computes the method set for the given type name
// which must denote a type whose underlying type is an interface.
// The same result qualifications apply as for infoFromTypeLit.
// infoFromTypeName should only be called from infoFromTypeLit.
func (check *Checker) infoFromTypeName(scope *Scope, name *ast.Ident, path []*TypeName) *ifaceInfo {
// A single call of infoFromTypeName handles a sequence of (possibly
// recursive) type declarations connected via unqualified type names.
// Each type declaration leading to another typename causes a "tail call"
// (goto) of this function. The general scenario looks like this:
//
// ...
// type Pn T // previous declarations leading to T, path = [..., Pn]
// type T interface { T0; ... } // T0 leads to call of infoFromTypeName
//
// // infoFromTypeName(name = T0, path = [..., Pn, T])
// type T0 T1 // path = [..., Pn, T, T0]
// type T1 T2 <-+ // path = [..., Pn, T, T0, T1]
// type T2 ... | // path = [..., Pn, T, T0, T1, T2]
// type Tn T1 --+ // path = [..., Pn, T, T0, T1, T2, Tn] and T1 is in path => cycle
//
// infoFromTypeName returns nil when such a cycle is detected. But in
// contrast to cycles involving interfaces, we must not report the
// error for "type name only" cycles because they will be found again
// during type-checking of embedded interfaces. Reporting those cycles
// here would lead to double reporting. Cycles involving embedding are
// not reported again later because type-checking of interfaces relies
// on the ifaceInfos computed here which are cycle-free by design.
//
// Remember the path length to detect "type name only" cycles.
start := len(path)
typenameLoop:
// name must be a type name denoting a type whose underlying type is an interface
_, obj := scope.LookupParent(name.Name, check.pos)
if obj == nil {
return nil
}
tname, _ := obj.(*TypeName)
if tname == nil {
return nil
}
// We have a type name. It may be predeclared (error type),
// imported (dot import), or declared by a type declaration.
// It may not be an interface (e.g., predeclared type int).
// Resolve it by analyzing each possible case.
// Abort but don't report an error if we have a "type name only"
// cycle (see big function comment).
if check.cycle(tname, path[start:], false) {
return nil
}
// Abort and report an error if we have a general cycle.
if check.cycle(tname, path, true) {
return nil
}
path = append(path, tname)
// If tname is a package-level type declaration, it must be
// in the objMap. Follow the RHS of that declaration if so.
// The RHS may be a literal type (likely case), or another
// (possibly parenthesized and/or qualified) type name.
// (The declaration may be an alias declaration, but it
// doesn't matter for the purpose of determining the under-
// lying interface.)
if decl := check.objMap[tname]; decl != nil {
switch typ := unparen(decl.typ).(type) {
case *ast.Ident:
// type tname T
name = typ
goto typenameLoop
case *ast.SelectorExpr:
// type tname p.T
return check.infoFromQualifiedTypeName(decl.file, typ)
case *ast.InterfaceType:
// type tname interface{...}
return check.infoFromTypeLit(decl.file, typ, tname, path)
}
// type tname X // and X is not an interface type
return nil
}
// If tname is not a package-level declaration, in a well-typed
// program it should be a predeclared (error type), imported (dot
// import), or function local declaration. Either way, it should
// have been fully declared before use, except if there is a direct
// cycle, and direct cycles will be caught above. Also, the denoted
// type should be an interface (e.g., int is not an interface).
if typ := tname.typ; typ != nil {
// typ should be an interface
if ityp, _ := typ.Underlying().(*Interface); ityp != nil {
return infoFromType(ityp)
}
}
// In all other cases we have some error.
return nil
}
// infoFromQualifiedTypeName computes the method set for the given qualified type name, or nil.
func (check *Checker) infoFromQualifiedTypeName(scope *Scope, qname *ast.SelectorExpr) *ifaceInfo {
// see also Checker.selector
name, _ := qname.X.(*ast.Ident)
if name == nil {
return nil
}
_, obj1 := scope.LookupParent(name.Name, check.pos)
if obj1 == nil {
return nil
}
pname, _ := obj1.(*PkgName)
if pname == nil {
return nil
}
assert(pname.pkg == check.pkg)
obj2 := pname.imported.scope.Lookup(qname.Sel.Name)
if obj2 == nil || !obj2.Exported() {
return nil
}
tname, _ := obj2.(*TypeName)
if tname == nil {
return nil
}
ityp, _ := tname.typ.Underlying().(*Interface)
if ityp == nil {
return nil
}
return infoFromType(ityp)
}
// infoFromType computes the method set for the given interface type.
// The result is never nil.
func infoFromType(typ *Interface) *ifaceInfo {
assert(typ.allMethods != nil) // typ must be completely set up
// fast track for empty interface
n := len(typ.allMethods)
if n == 0 {
return &emptyIfaceInfo
}
info := new(ifaceInfo)
info.explicits = len(typ.methods)
info.methods = make([]*methodInfo, n)
// If there are no embedded interfaces, simply collect the
// explicitly declared methods (optimization of common case).
if len(typ.methods) == n {
for i, m := range typ.methods {
info.methods[i] = &methodInfo{fun: m}
}
return info
}
// Interface types have a separate list for explicitly declared methods
// which shares its methods with the list of all (explicitly declared or
// embedded) methods. Collect all methods in a set so we can separate
// the embedded methods from the explicitly declared ones.
all := make(map[*Func]bool, n)
for _, m := range typ.allMethods {
all[m] = true
}
assert(len(all) == n) // methods must be unique
// collect explicitly declared methods
info.methods = make([]*methodInfo, n)
for i, m := range typ.methods {
info.methods[i] = &methodInfo{fun: m}
delete(all, m)
}
// collect remaining (embedded) methods
i := len(typ.methods)
for m := range all {
info.methods[i] = &methodInfo{fun: m}
i++
}
assert(i == n)
return info
}