gopls/internal/lsp/cache/parse.go

// Copyright 2019 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

package cache

import (
	"bytes"
	"context"
	"fmt"
	"go/ast"
	"go/parser"
	"go/scanner"
	"go/token"
	"go/types"
	"path/filepath"
	"reflect"
	"strconv"
	"strings"

	"golang.org/x/tools/gopls/internal/lsp/protocol"
	"golang.org/x/tools/gopls/internal/lsp/safetoken"
	"golang.org/x/tools/gopls/internal/lsp/source"
	"golang.org/x/tools/internal/diff"
	"golang.org/x/tools/internal/event"
	"golang.org/x/tools/internal/event/tag"
	"golang.org/x/tools/internal/memoize"
)

// parseKey uniquely identifies a parsed Go file.
type parseKey struct {
	file source.FileIdentity
	mode source.ParseMode
}

// ParseGo parses the file whose contents are provided by fh, using a cache.
// The resulting tree may have be fixed up.
//
// Token position information will be added to the snapshot's FileSet.
//
// The parser mode must not be ParseExported: that mode is used during
// type checking to destructively trim the tree to reduce work,
// which is not safe for values from a shared cache.
// TODO(adonovan): opt: shouldn't parseGoImpl do the trimming?
// Then we can cache the result since it would never change.
func (s *snapshot) ParseGo(ctx context.Context, fh source.FileHandle, mode source.ParseMode) (*source.ParsedGoFile, error) {
	if mode == source.ParseExported {
		panic("only type checking should use Exported")
	}

	key := parseKey{
		file: fh.FileIdentity(),
		mode: mode,
	}

	s.mu.Lock()
	entry, hit := s.parsedGoFiles.Get(key)
	s.mu.Unlock()

	// cache miss?
	if !hit {
		handle, release := s.store.Promise(key, func(ctx context.Context, arg interface{}) interface{} {
			parsed, err := parseGoImpl(ctx, arg.(*snapshot).FileSet(), fh, mode)
			return parseGoResult{parsed, err}
		})

		s.mu.Lock()
		// Check cache again in case another thread got there first.
		if prev, ok := s.parsedGoFiles.Get(key); ok {
			entry = prev
			release()
		} else {
			entry = handle
			s.parsedGoFiles.Set(key, entry, func(_, _ interface{}) { release() })
		}
		s.mu.Unlock()
	}

	// Await result.
	v, err := s.awaitPromise(ctx, entry.(*memoize.Promise))
	if err != nil {
		return nil, err
	}
	res := v.(parseGoResult)
	return res.parsed, res.err
}

// peekParseGoLocked peeks at the cache used by ParseGo but does not
// populate it or wait for other threads to do so. On cache hit, it returns
// the cache result of parseGoImpl; otherwise it returns (nil, nil).
func (s *snapshot) peekParseGoLocked(fh source.FileHandle, mode source.ParseMode) (*source.ParsedGoFile, error) {
	entry, hit := s.parsedGoFiles.Get(parseKey{fh.FileIdentity(), mode})
	if !hit {
		return nil, nil // no-one has requested this file
	}
	v := entry.(*memoize.Promise).Cached()
	if v == nil {
		return nil, nil // parsing is still in progress
	}
	res := v.(parseGoResult)
	return res.parsed, res.err
}

// parseGoResult holds the result of a call to parseGoImpl.
type parseGoResult struct {
	parsed *source.ParsedGoFile
	err    error
}

// parseGoImpl parses the Go source file whose content is provided by fh.
func parseGoImpl(ctx context.Context, fset *token.FileSet, fh source.FileHandle, mode source.ParseMode) (*source.ParsedGoFile, error) {
	ctx, done := event.Start(ctx, "cache.parseGo", tag.File.Of(fh.URI().Filename()))
	defer done()

	ext := filepath.Ext(fh.URI().Filename())
	if ext != ".go" && ext != "" { // files generated by cgo have no extension
		return nil, fmt.Errorf("cannot parse non-Go file %s", fh.URI())
	}
	src, err := fh.Read()
	if err != nil {
		return nil, err
	}

	parserMode := parser.AllErrors | parser.ParseComments
	if mode == source.ParseHeader {
		parserMode = parser.ImportsOnly | parser.ParseComments
	}

	file, err := parser.ParseFile(fset, fh.URI().Filename(), src, parserMode)
	var parseErr scanner.ErrorList
	if err != nil {
		// We passed a byte slice, so the only possible error is a parse error.
		parseErr = err.(scanner.ErrorList)
	}

	tok := fset.File(file.Pos())
	if tok == nil {
		// file.Pos is the location of the package declaration (issue #53202). If there was
		// none, we can't find the token.File that ParseFile created, and we
		// have no choice but to recreate it.
		tok = fset.AddFile(fh.URI().Filename(), -1, len(src))
		tok.SetLinesForContent(src)
	}

	fixed := false
	// If there were parse errors, attempt to fix them up.
	if parseErr != nil {
		// Fix any badly parsed parts of the AST.
		fixed = fixAST(ctx, file, tok, src)

		for i := 0; i < 10; i++ {
			// Fix certain syntax errors that render the file unparseable.
			newSrc := fixSrc(file, tok, src)
			if newSrc == nil {
				break
			}

			// If we thought there was something to fix 10 times in a row,
			// it is likely we got stuck in a loop somehow. Log out a diff
			// of the last changes we made to aid in debugging.
			if i == 9 {
				unified := diff.Unified("before", "after", string(src), string(newSrc))
				event.Log(ctx, fmt.Sprintf("fixSrc loop - last diff:\n%v", unified), tag.File.Of(tok.Name()))
			}

			newFile, _ := parser.ParseFile(fset, fh.URI().Filename(), newSrc, parserMode)
			if newFile != nil {
				// Maintain the original parseError so we don't try formatting the doctored file.
				file = newFile
				src = newSrc
				tok = fset.File(file.Pos())

				fixed = fixAST(ctx, file, tok, src)
			}
		}
	}

	return &source.ParsedGoFile{
		URI:   fh.URI(),
		Mode:  mode,
		Src:   src,
		Fixed: fixed,
		File:  file,
		Tok:   tok,
		Mapper: &protocol.ColumnMapper{
			URI:     fh.URI(),
			TokFile: tok,
			Content: src,
		},
		ParseErr: parseErr,
	}, nil
}

// An unexportedFilter removes as much unexported AST from a set of Files as possible.
type unexportedFilter struct {
	uses map[string]bool
}

// Filter records uses of unexported identifiers and filters out all other
// unexported declarations.
func (f *unexportedFilter) Filter(files []*ast.File) {
	// Iterate to fixed point -- unexported types can include other unexported types.
	oldLen := len(f.uses)
	for {
		for _, file := range files {
			f.recordUses(file)
		}
		if len(f.uses) == oldLen {
			break
		}
		oldLen = len(f.uses)
	}

	for _, file := range files {
		var newDecls []ast.Decl
		for _, decl := range file.Decls {
			if f.filterDecl(decl) {
				newDecls = append(newDecls, decl)
			}
		}
		file.Decls = newDecls
		file.Scope = nil
		file.Unresolved = nil
		file.Comments = nil
		trimAST(file)
	}
}

func (f *unexportedFilter) keep(ident *ast.Ident) bool {
	return ast.IsExported(ident.Name) || f.uses[ident.Name]
}

func (f *unexportedFilter) filterDecl(decl ast.Decl) bool {
	switch decl := decl.(type) {
	case *ast.FuncDecl:
		if ident := source.RecvIdent(decl.Recv); ident != nil && !f.keep(ident) {
			return false
		}
		return f.keep(decl.Name)
	case *ast.GenDecl:
		if decl.Tok == token.CONST {
			// Constants can involve iota, and iota is hard to deal with.
			return true
		}
		var newSpecs []ast.Spec
		for _, spec := range decl.Specs {
			if f.filterSpec(spec) {
				newSpecs = append(newSpecs, spec)
			}
		}
		decl.Specs = newSpecs
		return len(newSpecs) != 0
	case *ast.BadDecl:
		return false
	}
	panic(fmt.Sprintf("unknown ast.Decl %T", decl))
}

func (f *unexportedFilter) filterSpec(spec ast.Spec) bool {
	switch spec := spec.(type) {
	case *ast.ImportSpec:
		return true
	case *ast.ValueSpec:
		var newNames []*ast.Ident
		for _, name := range spec.Names {
			if f.keep(name) {
				newNames = append(newNames, name)
			}
		}
		spec.Names = newNames
		return len(spec.Names) != 0
	case *ast.TypeSpec:
		if !f.keep(spec.Name) {
			return false
		}
		switch typ := spec.Type.(type) {
		case *ast.StructType:
			// In practice this no longer filters anything;
			// see comment at StructType case in recordUses.
			f.filterFieldList(typ.Fields)
		case *ast.InterfaceType:
			f.filterFieldList(typ.Methods)
		}
		return true
	}
	panic(fmt.Sprintf("unknown ast.Spec %T", spec))
}

func (f *unexportedFilter) filterFieldList(fields *ast.FieldList) {
	var newFields []*ast.Field
	for _, field := range fields.List {
		if len(field.Names) == 0 {
			// Keep embedded fields: they can export methods and fields.
			newFields = append(newFields, field)
		}
		for _, name := range field.Names {
			if f.keep(name) {
				newFields = append(newFields, field)
				break
			}
		}
	}
	fields.List = newFields
}

func (f *unexportedFilter) recordUses(file *ast.File) {
	for _, decl := range file.Decls {
		switch decl := decl.(type) {
		case *ast.FuncDecl:
			// Ignore methods on dropped types.
			if ident := source.RecvIdent(decl.Recv); ident != nil && !f.keep(ident) {
				break
			}
			// Ignore functions with dropped names.
			if !f.keep(decl.Name) {
				break
			}
			f.recordFuncType(decl.Type)
		case *ast.GenDecl:
			for _, spec := range decl.Specs {
				switch spec := spec.(type) {
				case *ast.ValueSpec:
					for i, name := range spec.Names {
						// Don't mess with constants -- iota is hard.
						if f.keep(name) || decl.Tok == token.CONST {
							f.recordIdents(spec.Type)
							if len(spec.Values) > i {
								f.recordIdents(spec.Values[i])
							}
						}
					}
				case *ast.TypeSpec:
					switch typ := spec.Type.(type) {
					case *ast.StructType:
						// We used to trim unexported fields but this
						// had observable consequences. For example,
						// the 'fieldalignment' analyzer would compute
						// incorrect diagnostics from the size and
						// offsets, and the UI hover information for
						// types was inaccurate. So now we keep them.
						if typ.Fields != nil {
							for _, field := range typ.Fields.List {
								f.recordIdents(field.Type)
							}
						}
					case *ast.InterfaceType:
						f.recordInterfaceMethodUses(typ.Methods)
					}
				}
			}
		}
	}
}

// recordIdents records unexported identifiers in an Expr in uses.
// These may be types, e.g. in map[key]value, function names, e.g. in foo(),
// or simple variable references. References that will be discarded, such
// as those in function literal bodies, are ignored.
func (f *unexportedFilter) recordIdents(x ast.Expr) {
	ast.Inspect(x, func(n ast.Node) bool {
		if n == nil {
			return false
		}
		if complit, ok := n.(*ast.CompositeLit); ok {
			// We clear out composite literal contents; just record their type.
			f.recordIdents(complit.Type)
			return false
		}
		if flit, ok := n.(*ast.FuncLit); ok {
			f.recordFuncType(flit.Type)
			return false
		}
		if ident, ok := n.(*ast.Ident); ok && !ast.IsExported(ident.Name) {
			f.uses[ident.Name] = true
		}
		return true
	})
}

// recordFuncType records the types mentioned by a function type.
func (f *unexportedFilter) recordFuncType(fn *ast.FuncType) {
	// Parameter and result types of retained functions need to be retained.
	if fn.Params != nil {
		for _, field := range fn.Params.List {
			f.recordIdents(field.Type)
		}
	}
	if fn.Results != nil {
		for _, field := range fn.Results.List {
			f.recordIdents(field.Type)
		}
	}
}

// recordInterfaceMethodUses records unexported identifiers used in interface methods.
func (f *unexportedFilter) recordInterfaceMethodUses(methods *ast.FieldList) {
	if methods != nil {
		for _, method := range methods.List {
			if len(method.Names) == 0 {
				// I, pkg.I, I[T] -- embedded interface:
				// may contribute exported names.
				f.recordIdents(method.Type)
			} else if ft, ok := method.Type.(*ast.FuncType); ok {
				// f(T) -- ordinary interface method:
				// needs all its types retained.
				f.recordFuncType(ft)
			}
		}
	}
}

// ProcessErrors records additional uses from errors, returning the new uses
// and any unexpected errors.
func (f *unexportedFilter) ProcessErrors(errors []types.Error) (map[string]bool, []types.Error) {
	var unexpected []types.Error
	missing := map[string]bool{}
	for _, err := range errors {
		if strings.Contains(err.Msg, "missing return") {
			continue
		}
		const undeclared = "undeclared name: "
		if strings.HasPrefix(err.Msg, undeclared) {
			missing[strings.TrimPrefix(err.Msg, undeclared)] = true
			f.uses[strings.TrimPrefix(err.Msg, undeclared)] = true
			continue
		}
		unexpected = append(unexpected, err)
	}
	return missing, unexpected
}

// trimAST clears any part of the AST not relevant to type checking
// the package-level declarations.
func trimAST(file *ast.File) {
	// Eliminate bodies of top-level functions, methods, inits.
	for _, decl := range file.Decls {
		if fn, ok := decl.(*ast.FuncDecl); ok {
			fn.Body = nil
		}
	}

	// Simplify remaining declarations.
	ast.Inspect(file, func(n ast.Node) bool {
		switch n := n.(type) {
		case *ast.FuncLit:
			// Eliminate bodies of literal functions.
			// func() { ... } => func() {}
			n.Body.List = nil
		case *ast.CompositeLit:
			// types.Info.Types for long slice/array literals are particularly
			// expensive. Try to clear them out: T{e, ..., e} => T{}
			at, ok := n.Type.(*ast.ArrayType)
			if !ok {
				// Map or struct literal: no harm removing all its fields.
				n.Elts = nil
				break
			}

			// Removing the elements from an ellipsis array changes its type.
			// Try to set the length explicitly so we can continue.
			//  [...]T{e, ..., e} => [3]T[]{}
			if _, ok := at.Len.(*ast.Ellipsis); ok {
				length, ok := arrayLength(n)
				if !ok {
					break
				}
				at.Len = &ast.BasicLit{
					Kind:     token.INT,
					Value:    fmt.Sprint(length),
					ValuePos: at.Len.Pos(),
				}
			}
			n.Elts = nil
		}
		return true
	})
}

// arrayLength returns the length of some simple forms of ellipsis array literal.
// Notably, it handles the tables in golang.org/x/text.
func arrayLength(array *ast.CompositeLit) (int, bool) {
	litVal := func(expr ast.Expr) (int, bool) {
		lit, ok := expr.(*ast.BasicLit)
		if !ok {
			return 0, false
		}
		val, err := strconv.ParseInt(lit.Value, 10, 64)
		if err != nil {
			return 0, false
		}
		return int(val), true
	}
	largestKey := -1
	for _, elt := range array.Elts {
		kve, ok := elt.(*ast.KeyValueExpr)
		if !ok {
			continue
		}
		switch key := kve.Key.(type) {
		case *ast.BasicLit:
			if val, ok := litVal(key); ok && largestKey < val {
				largestKey = val
			}
		case *ast.BinaryExpr:
			// golang.org/x/text uses subtraction (and only subtraction) in its indices.
			if key.Op != token.SUB {
				break
			}
			x, ok := litVal(key.X)
			if !ok {
				break
			}
			y, ok := litVal(key.Y)
			if !ok {
				break
			}
			if val := x - y; largestKey < val {
				largestKey = val
			}
		}
	}
	if largestKey != -1 {
		return largestKey + 1, true
	}
	return len(array.Elts), true
}

// fixAST inspects the AST and potentially modifies any *ast.BadStmts so that it can be
// type-checked more effectively.
//
// If fixAST returns true, the resulting AST is considered "fixed", meaning
// positions have been mangled, and type checker errors may not make sense.
func fixAST(ctx context.Context, n ast.Node, tok *token.File, src []byte) (fixed bool) {
	var err error
	walkASTWithParent(n, func(n, parent ast.Node) bool {
		switch n := n.(type) {
		case *ast.BadStmt:
			if fixed = fixDeferOrGoStmt(n, parent, tok, src); fixed {
				// Recursively fix in our fixed node.
				_ = fixAST(ctx, parent, tok, src)
			} else {
				err = fmt.Errorf("unable to parse defer or go from *ast.BadStmt: %v", err)
			}
			return false
		case *ast.BadExpr:
			if fixed = fixArrayType(n, parent, tok, src); fixed {
				// Recursively fix in our fixed node.
				_ = fixAST(ctx, parent, tok, src)
				return false
			}

			// Fix cases where parser interprets if/for/switch "init"
			// statement as "cond" expression, e.g.:
			//
			//   // "i := foo" is init statement, not condition.
			//   for i := foo
			//
			fixInitStmt(n, parent, tok, src)

			return false
		case *ast.SelectorExpr:
			// Fix cases where a keyword prefix results in a phantom "_" selector, e.g.:
			//
			//   foo.var<> // want to complete to "foo.variance"
			//
			fixPhantomSelector(n, tok, src)
			return true

		case *ast.BlockStmt:
			switch parent.(type) {
			case *ast.SwitchStmt, *ast.TypeSwitchStmt, *ast.SelectStmt:
				// Adjust closing curly brace of empty switch/select
				// statements so we can complete inside them.
				fixEmptySwitch(n, tok, src)
			}

			return true
		default:
			return true
		}
	})
	return fixed
}

// walkASTWithParent walks the AST rooted at n. The semantics are
// similar to ast.Inspect except it does not call f(nil).
func walkASTWithParent(n ast.Node, f func(n ast.Node, parent ast.Node) bool) {
	var ancestors []ast.Node
	ast.Inspect(n, func(n ast.Node) (recurse bool) {
		defer func() {
			if recurse {
				ancestors = append(ancestors, n)
			}
		}()

		if n == nil {
			ancestors = ancestors[:len(ancestors)-1]
			return false
		}

		var parent ast.Node
		if len(ancestors) > 0 {
			parent = ancestors[len(ancestors)-1]
		}

		return f(n, parent)
	})
}

// fixSrc attempts to modify the file's source code to fix certain
// syntax errors that leave the rest of the file unparsed.
func fixSrc(f *ast.File, tf *token.File, src []byte) (newSrc []byte) {
	walkASTWithParent(f, func(n, parent ast.Node) bool {
		if newSrc != nil {
			return false
		}

		switch n := n.(type) {
		case *ast.BlockStmt:
			newSrc = fixMissingCurlies(f, n, parent, tf, src)
		case *ast.SelectorExpr:
			newSrc = fixDanglingSelector(n, tf, src)
		}

		return newSrc == nil
	})

	return newSrc
}

// fixMissingCurlies adds in curly braces for block statements that
// are missing curly braces. For example:
//
//	if foo
//
// becomes
//
//	if foo {}
func fixMissingCurlies(f *ast.File, b *ast.BlockStmt, parent ast.Node, tok *token.File, src []byte) []byte {
	// If the "{" is already in the source code, there isn't anything to
	// fix since we aren't missing curlies.
	if b.Lbrace.IsValid() {
		braceOffset, err := safetoken.Offset(tok, b.Lbrace)
		if err != nil {
			return nil
		}
		if braceOffset < len(src) && src[braceOffset] == '{' {
			return nil
		}
	}

	parentLine := tok.Line(parent.Pos())

	if parentLine >= tok.LineCount() {
		// If we are the last line in the file, no need to fix anything.
		return nil
	}

	// Insert curlies at the end of parent's starting line. The parent
	// is the statement that contains the block, e.g. *ast.IfStmt. The
	// block's Pos()/End() can't be relied upon because they are based
	// on the (missing) curly braces. We assume the statement is a
	// single line for now and try sticking the curly braces at the end.
	insertPos := tok.LineStart(parentLine+1) - 1

	// Scootch position backwards until it's not in a comment. For example:
	//
	// if foo<> // some amazing comment |
	// someOtherCode()
	//
	// insertPos will be located at "|", so we back it out of the comment.
	didSomething := true
	for didSomething {
		didSomething = false
		for _, c := range f.Comments {
			if c.Pos() < insertPos && insertPos <= c.End() {
				insertPos = c.Pos()
				didSomething = true
			}
		}
	}

	// Bail out if line doesn't end in an ident or ".". This is to avoid
	// cases like below where we end up making things worse by adding
	// curlies:
	//
	//   if foo &&
	//     bar<>
	switch precedingToken(insertPos, tok, src) {
	case token.IDENT, token.PERIOD:
		// ok
	default:
		return nil
	}

	var buf bytes.Buffer
	buf.Grow(len(src) + 3)
	offset, err := safetoken.Offset(tok, insertPos)
	if err != nil {
		return nil
	}
	buf.Write(src[:offset])

	// Detect if we need to insert a semicolon to fix "for" loop situations like:
	//
	//   for i := foo(); foo<>
	//
	// Just adding curlies is not sufficient to make things parse well.
	if fs, ok := parent.(*ast.ForStmt); ok {
		if _, ok := fs.Cond.(*ast.BadExpr); !ok {
			if xs, ok := fs.Post.(*ast.ExprStmt); ok {
				if _, ok := xs.X.(*ast.BadExpr); ok {
					buf.WriteByte(';')
				}
			}
		}
	}

	// Insert "{}" at insertPos.
	buf.WriteByte('{')
	buf.WriteByte('}')
	buf.Write(src[offset:])
	return buf.Bytes()
}

// fixEmptySwitch moves empty switch/select statements' closing curly
// brace down one line. This allows us to properly detect incomplete
// "case" and "default" keywords as inside the switch statement. For
// example:
//
//	switch {
//	def<>
//	}
//
// gets parsed like:
//
//	switch {
//	}
//
// Later we manually pull out the "def" token, but we need to detect
// that our "<>" position is inside the switch block. To do that we
// move the curly brace so it looks like:
//
//	switch {
//
//	}
func fixEmptySwitch(body *ast.BlockStmt, tok *token.File, src []byte) {
	// We only care about empty switch statements.
	if len(body.List) > 0 || !body.Rbrace.IsValid() {
		return
	}

	// If the right brace is actually in the source code at the
	// specified position, don't mess with it.
	braceOffset, err := safetoken.Offset(tok, body.Rbrace)
	if err != nil {
		return
	}
	if braceOffset < len(src) && src[braceOffset] == '}' {
		return
	}

	braceLine := tok.Line(body.Rbrace)
	if braceLine >= tok.LineCount() {
		// If we are the last line in the file, no need to fix anything.
		return
	}

	// Move the right brace down one line.
	body.Rbrace = tok.LineStart(braceLine + 1)
}

// fixDanglingSelector inserts real "_" selector expressions in place
// of phantom "_" selectors. For example:
//
//	func _() {
//		x.<>
//	}
//
// var x struct { i int }
//
// To fix completion at "<>", we insert a real "_" after the "." so the
// following declaration of "x" can be parsed and type checked
// normally.
func fixDanglingSelector(s *ast.SelectorExpr, tf *token.File, src []byte) []byte {
	if !isPhantomUnderscore(s.Sel, tf, src) {
		return nil
	}

	if !s.X.End().IsValid() {
		return nil
	}

	insertOffset, err := safetoken.Offset(tf, s.X.End())
	if err != nil {
		return nil
	}
	// Insert directly after the selector's ".".
	insertOffset++
	if src[insertOffset-1] != '.' {
		return nil
	}

	var buf bytes.Buffer
	buf.Grow(len(src) + 1)
	buf.Write(src[:insertOffset])
	buf.WriteByte('_')
	buf.Write(src[insertOffset:])
	return buf.Bytes()
}

// fixPhantomSelector tries to fix selector expressions with phantom
// "_" selectors. In particular, we check if the selector is a
// keyword, and if so we swap in an *ast.Ident with the keyword text. For example:
//
// foo.var
//
// yields a "_" selector instead of "var" since "var" is a keyword.
//
// TODO(rfindley): should this constitute an ast 'fix'?
func fixPhantomSelector(sel *ast.SelectorExpr, tf *token.File, src []byte) {
	if !isPhantomUnderscore(sel.Sel, tf, src) {
		return
	}

	// Only consider selectors directly abutting the selector ".". This
	// avoids false positives in cases like:
	//
	//   foo. // don't think "var" is our selector
	//   var bar = 123
	//
	if sel.Sel.Pos() != sel.X.End()+1 {
		return
	}

	maybeKeyword := readKeyword(sel.Sel.Pos(), tf, src)
	if maybeKeyword == "" {
		return
	}

	replaceNode(sel, sel.Sel, &ast.Ident{
		Name:    maybeKeyword,
		NamePos: sel.Sel.Pos(),
	})
}

// isPhantomUnderscore reports whether the given ident is a phantom
// underscore. The parser sometimes inserts phantom underscores when
// it encounters otherwise unparseable situations.
func isPhantomUnderscore(id *ast.Ident, tok *token.File, src []byte) bool {
	if id == nil || id.Name != "_" {
		return false
	}

	// Phantom underscore means the underscore is not actually in the
	// program text.
	offset, err := safetoken.Offset(tok, id.Pos())
	if err != nil {
		return false
	}
	return len(src) <= offset || src[offset] != '_'
}

// fixInitStmt fixes cases where the parser misinterprets an
// if/for/switch "init" statement as the "cond" conditional. In cases
// like "if i := 0" the user hasn't typed the semicolon yet so the
// parser is looking for the conditional expression. However, "i := 0"
// are not valid expressions, so we get a BadExpr.
//
// fixInitStmt returns valid AST for the original source.
func fixInitStmt(bad *ast.BadExpr, parent ast.Node, tok *token.File, src []byte) {
	if !bad.Pos().IsValid() || !bad.End().IsValid() {
		return
	}

	// Try to extract a statement from the BadExpr.
	start, err := safetoken.Offset(tok, bad.Pos())
	if err != nil {
		return
	}
	end, err := safetoken.Offset(tok, bad.End()-1)
	if err != nil {
		return
	}
	stmtBytes := src[start : end+1]
	stmt, err := parseStmt(bad.Pos(), stmtBytes)
	if err != nil {
		return
	}

	// If the parent statement doesn't already have an "init" statement,
	// move the extracted statement into the "init" field and insert a
	// dummy expression into the required "cond" field.
	switch p := parent.(type) {
	case *ast.IfStmt:
		if p.Init != nil {
			return
		}
		p.Init = stmt
		p.Cond = &ast.Ident{
			Name:    "_",
			NamePos: stmt.End(),
		}
	case *ast.ForStmt:
		if p.Init != nil {
			return
		}
		p.Init = stmt
		p.Cond = &ast.Ident{
			Name:    "_",
			NamePos: stmt.End(),
		}
	case *ast.SwitchStmt:
		if p.Init != nil {
			return
		}
		p.Init = stmt
		p.Tag = nil
	}
}

// readKeyword reads the keyword starting at pos, if any.
func readKeyword(pos token.Pos, tok *token.File, src []byte) string {
	var kwBytes []byte
	offset, err := safetoken.Offset(tok, pos)
	if err != nil {
		return ""
	}
	for i := offset; i < len(src); i++ {
		// Use a simplified identifier check since keywords are always lowercase ASCII.
		if src[i] < 'a' || src[i] > 'z' {
			break
		}
		kwBytes = append(kwBytes, src[i])

		// Stop search at arbitrarily chosen too-long-for-a-keyword length.
		if len(kwBytes) > 15 {
			return ""
		}
	}

	if kw := string(kwBytes); token.Lookup(kw).IsKeyword() {
		return kw
	}

	return ""
}

// fixArrayType tries to parse an *ast.BadExpr into an *ast.ArrayType.
// go/parser often turns lone array types like "[]int" into BadExprs
// if it isn't expecting a type.
func fixArrayType(bad *ast.BadExpr, parent ast.Node, tok *token.File, src []byte) bool {
	// Our expected input is a bad expression that looks like "[]someExpr".

	from := bad.Pos()
	to := bad.End()

	if !from.IsValid() || !to.IsValid() {
		return false
	}

	exprBytes := make([]byte, 0, int(to-from)+3)
	// Avoid doing tok.Offset(to) since that panics if badExpr ends at EOF.
	// It also panics if the position is not in the range of the file, and
	// badExprs may not necessarily have good positions, so check first.
	fromOffset, err := safetoken.Offset(tok, from)
	if err != nil {
		return false
	}
	toOffset, err := safetoken.Offset(tok, to-1)
	if err != nil {
		return false
	}
	exprBytes = append(exprBytes, src[fromOffset:toOffset+1]...)
	exprBytes = bytes.TrimSpace(exprBytes)

	// If our expression ends in "]" (e.g. "[]"), add a phantom selector
	// so we can complete directly after the "[]".
	if len(exprBytes) > 0 && exprBytes[len(exprBytes)-1] == ']' {
		exprBytes = append(exprBytes, '_')
	}

	// Add "{}" to turn our ArrayType into a CompositeLit. This is to
	// handle the case of "[...]int" where we must make it a composite
	// literal to be parseable.
	exprBytes = append(exprBytes, '{', '}')

	expr, err := parseExpr(from, exprBytes)
	if err != nil {
		return false
	}

	cl, _ := expr.(*ast.CompositeLit)
	if cl == nil {
		return false
	}

	at, _ := cl.Type.(*ast.ArrayType)
	if at == nil {
		return false
	}

	return replaceNode(parent, bad, at)
}

// precedingToken scans src to find the token preceding pos.
func precedingToken(pos token.Pos, tok *token.File, src []byte) token.Token {
	s := &scanner.Scanner{}
	s.Init(tok, src, nil, 0)

	var lastTok token.Token
	for {
		p, t, _ := s.Scan()
		if t == token.EOF || p >= pos {
			break
		}

		lastTok = t
	}
	return lastTok
}

// fixDeferOrGoStmt tries to parse an *ast.BadStmt into a defer or a go statement.
//
// go/parser packages a statement of the form "defer x." as an *ast.BadStmt because
// it does not include a call expression. This means that go/types skips type-checking
// this statement entirely, and we can't use the type information when completing.
// Here, we try to generate a fake *ast.DeferStmt or *ast.GoStmt to put into the AST,
// instead of the *ast.BadStmt.
func fixDeferOrGoStmt(bad *ast.BadStmt, parent ast.Node, tok *token.File, src []byte) bool {
	// Check if we have a bad statement containing either a "go" or "defer".
	s := &scanner.Scanner{}
	s.Init(tok, src, nil, 0)

	var (
		pos token.Pos
		tkn token.Token
	)
	for {
		if tkn == token.EOF {
			return false
		}
		if pos >= bad.From {
			break
		}
		pos, tkn, _ = s.Scan()
	}

	var stmt ast.Stmt
	switch tkn {
	case token.DEFER:
		stmt = &ast.DeferStmt{
			Defer: pos,
		}
	case token.GO:
		stmt = &ast.GoStmt{
			Go: pos,
		}
	default:
		return false
	}

	var (
		from, to, last   token.Pos
		lastToken        token.Token
		braceDepth       int
		phantomSelectors []token.Pos
	)
FindTo:
	for {
		to, tkn, _ = s.Scan()

		if from == token.NoPos {
			from = to
		}

		switch tkn {
		case token.EOF:
			break FindTo
		case token.SEMICOLON:
			// If we aren't in nested braces, end of statement means
			// end of expression.
			if braceDepth == 0 {
				break FindTo
			}
		case token.LBRACE:
			braceDepth++
		}

		// This handles the common dangling selector case. For example in
		//
		// defer fmt.
		// y := 1
		//
		// we notice the dangling period and end our expression.
		//
		// If the previous token was a "." and we are looking at a "}",
		// the period is likely a dangling selector and needs a phantom
		// "_". Likewise if the current token is on a different line than
		// the period, the period is likely a dangling selector.
		if lastToken == token.PERIOD && (tkn == token.RBRACE || tok.Line(to) > tok.Line(last)) {
			// Insert phantom "_" selector after the dangling ".".
			phantomSelectors = append(phantomSelectors, last+1)
			// If we aren't in a block then end the expression after the ".".
			if braceDepth == 0 {
				to = last + 1
				break
			}
		}

		lastToken = tkn
		last = to

		switch tkn {
		case token.RBRACE:
			braceDepth--
			if braceDepth <= 0 {
				if braceDepth == 0 {
					// +1 to include the "}" itself.
					to += 1
				}
				break FindTo
			}
		}
	}

	fromOffset, err := safetoken.Offset(tok, from)
	if err != nil {
		return false
	}
	if !from.IsValid() || fromOffset >= len(src) {
		return false
	}

	toOffset, err := safetoken.Offset(tok, to)
	if err != nil {
		return false
	}
	if !to.IsValid() || toOffset >= len(src) {
		return false
	}

	// Insert any phantom selectors needed to prevent dangling "." from messing
	// up the AST.
	exprBytes := make([]byte, 0, int(to-from)+len(phantomSelectors))
	for i, b := range src[fromOffset:toOffset] {
		if len(phantomSelectors) > 0 && from+token.Pos(i) == phantomSelectors[0] {
			exprBytes = append(exprBytes, '_')
			phantomSelectors = phantomSelectors[1:]
		}
		exprBytes = append(exprBytes, b)
	}

	if len(phantomSelectors) > 0 {
		exprBytes = append(exprBytes, '_')
	}

	expr, err := parseExpr(from, exprBytes)
	if err != nil {
		return false
	}

	// Package the expression into a fake *ast.CallExpr and re-insert
	// into the function.
	call := &ast.CallExpr{
		Fun:    expr,
		Lparen: to,
		Rparen: to,
	}

	switch stmt := stmt.(type) {
	case *ast.DeferStmt:
		stmt.Call = call
	case *ast.GoStmt:
		stmt.Call = call
	}

	return replaceNode(parent, bad, stmt)
}

// parseStmt parses the statement in src and updates its position to
// start at pos.
func parseStmt(pos token.Pos, src []byte) (ast.Stmt, error) {
	// Wrap our expression to make it a valid Go file we can pass to ParseFile.
	fileSrc := bytes.Join([][]byte{
		[]byte("package fake;func _(){"),
		src,
		[]byte("}"),
	}, nil)

	// Use ParseFile instead of ParseExpr because ParseFile has
	// best-effort behavior, whereas ParseExpr fails hard on any error.
	fakeFile, err := parser.ParseFile(token.NewFileSet(), "", fileSrc, 0)
	if fakeFile == nil {
		return nil, fmt.Errorf("error reading fake file source: %v", err)
	}

	// Extract our expression node from inside the fake file.
	if len(fakeFile.Decls) == 0 {
		return nil, fmt.Errorf("error parsing fake file: %v", err)
	}

	fakeDecl, _ := fakeFile.Decls[0].(*ast.FuncDecl)
	if fakeDecl == nil || len(fakeDecl.Body.List) == 0 {
		return nil, fmt.Errorf("no statement in %s: %v", src, err)
	}

	stmt := fakeDecl.Body.List[0]

	// parser.ParseFile returns undefined positions.
	// Adjust them for the current file.
	offsetPositions(stmt, pos-1-(stmt.Pos()-1))

	return stmt, nil
}

// parseExpr parses the expression in src and updates its position to
// start at pos.
func parseExpr(pos token.Pos, src []byte) (ast.Expr, error) {
	stmt, err := parseStmt(pos, src)
	if err != nil {
		return nil, err
	}

	exprStmt, ok := stmt.(*ast.ExprStmt)
	if !ok {
		return nil, fmt.Errorf("no expr in %s: %v", src, err)
	}

	return exprStmt.X, nil
}

var tokenPosType = reflect.TypeOf(token.NoPos)

// offsetPositions applies an offset to the positions in an ast.Node.
func offsetPositions(n ast.Node, offset token.Pos) {
	ast.Inspect(n, func(n ast.Node) bool {
		if n == nil {
			return false
		}

		v := reflect.ValueOf(n).Elem()

		switch v.Kind() {
		case reflect.Struct:
			for i := 0; i < v.NumField(); i++ {
				f := v.Field(i)
				if f.Type() != tokenPosType {
					continue
				}

				if !f.CanSet() {
					continue
				}

				// Don't offset invalid positions: they should stay invalid.
				if !token.Pos(f.Int()).IsValid() {
					continue
				}

				f.SetInt(f.Int() + int64(offset))
			}
		}

		return true
	})
}

// replaceNode updates parent's child oldChild to be newChild. It
// returns whether it replaced successfully.
func replaceNode(parent, oldChild, newChild ast.Node) bool {
	if parent == nil || oldChild == nil || newChild == nil {
		return false
	}

	parentVal := reflect.ValueOf(parent).Elem()
	if parentVal.Kind() != reflect.Struct {
		return false
	}

	newChildVal := reflect.ValueOf(newChild)

	tryReplace := func(v reflect.Value) bool {
		if !v.CanSet() || !v.CanInterface() {
			return false
		}

		// If the existing value is oldChild, we found our child. Make
		// sure our newChild is assignable and then make the swap.
		if v.Interface() == oldChild && newChildVal.Type().AssignableTo(v.Type()) {
			v.Set(newChildVal)
			return true
		}

		return false
	}

	// Loop over parent's struct fields.
	for i := 0; i < parentVal.NumField(); i++ {
		f := parentVal.Field(i)

		switch f.Kind() {
		// Check interface and pointer fields.
		case reflect.Interface, reflect.Ptr:
			if tryReplace(f) {
				return true
			}

		// Search through any slice fields.
		case reflect.Slice:
			for i := 0; i < f.Len(); i++ {
				if tryReplace(f.Index(i)) {
					return true
				}
			}
		}
	}

	return false
}