cmd/compile: print recursive types correctly

Change the type printer to take a map of types that we're currently
printing. When we happen upon a type that we're already in the middle
of printing, print a reference to it instead.

A reference to another type is built using the offset of the first
byte of that type's string representation in the result. To facilitate
that computation (and it's probably more efficient, regardless), we
print the type to a buffer as we go, and build the string at the end.

It would be nice to use string.Builder instead of bytes.Buffer, but
string.Builder wasn't around in Go 1.4, and we'd like to bootstrap
from that version.

Fixes #29312

Change-Id: I49d788c1fa20f770df7b2bae3b9979d990d54803
Reviewed-on: https://go-review.googlesource.com/c/go/+/214239Reviewed-by: Robert Griesemer <gri@golang.org>

src/cmd/compile/internal/gc/fmt.go

@@ -5,6 +5,7 @@
 package gc
 
 import (
+	"bytes"
 	"cmd/compile/internal/types"
 	"cmd/internal/src"
 	"fmt"
@@ -650,23 +651,55 @@ var basicnames = []string{
 	TBLANK:      "blank",
 }
 
-func typefmt(t *types.Type, flag FmtFlag, mode fmtMode, depth int) string {
+func tconv(t *types.Type, flag FmtFlag, mode fmtMode) string {
+	b := bytes.NewBuffer(make([]byte, 0, 64))
+	tconv2(b, t, flag, mode, nil)
+	return b.String()
+}
+
+// tconv2 writes a string representation of t to b.
+// flag and mode control exactly what is printed.
+// Any types x that are already in the visited map get printed as @%d where %d=visited[x].
+// See #16897 before changing the implementation of tconv.
+func tconv2(b *bytes.Buffer, t *types.Type, flag FmtFlag, mode fmtMode, visited map[*types.Type]int) {
+	if off, ok := visited[t]; ok {
+		// We've seen this type before, so we're trying to print it recursively.
+		// Print a reference to it instead.
+		fmt.Fprintf(b, "@%d", off)
+		return
+	}
 	if t == nil {
-		return "<T>"
+		b.WriteString("<T>")
+		return
+	}
+	if t.Etype == types.TSSA {
+		b.WriteString(t.Extra.(string))
+		return
+	}
+	if t.Etype == types.TTUPLE {
+		b.WriteString(t.FieldType(0).String())
+		b.WriteByte(',')
+		b.WriteString(t.FieldType(1).String())
+		return
 	}
 
+	flag, mode = flag.update(mode)
+	if mode == FTypeIdName {
+		flag |= FmtUnsigned
+	}
 	if t == types.Bytetype || t == types.Runetype {
 		// in %-T mode collapse rune and byte with their originals.
 		switch mode {
 		case FTypeIdName, FTypeId:
 			t = types.Types[t.Etype]
 		default:
-			return sconv(t.Sym, FmtShort, mode)
+			b.WriteString(sconv(t.Sym, FmtShort, mode))
+			return
 		}
 	}
-
 	if t == types.Errortype {
-		return "error"
+		b.WriteString("error")
+		return
 	}
 
 	// Unless the 'L' flag was specified, if the type has a name, just print that name.
@@ -675,161 +708,197 @@ func typefmt(t *types.Type, flag FmtFlag, mode fmtMode, depth int) string {
 		case FTypeId, FTypeIdName:
 			if flag&FmtShort != 0 {
 				if t.Vargen != 0 {
-					return mode.Sprintf("%v·%d", sconv(t.Sym, FmtShort, mode), t.Vargen)
+					fmt.Fprintf(b, "%s·%d", sconv(t.Sym, FmtShort, mode), t.Vargen)
+					return
 				}
-				return sconv(t.Sym, FmtShort, mode)
+				b.WriteString(sconv(t.Sym, FmtShort, mode))
+				return
 			}
 
 			if mode == FTypeIdName {
-				return sconv(t.Sym, FmtUnsigned, mode)
+				b.WriteString(sconv(t.Sym, FmtUnsigned, mode))
+				return
 			}
 
 			if t.Sym.Pkg == localpkg && t.Vargen != 0 {
-				return mode.Sprintf("%v·%d", t.Sym, t.Vargen)
+				b.WriteString(mode.Sprintf("%v·%d", t.Sym, t.Vargen))
+				return
 			}
 		}
 
-		return smodeString(t.Sym, mode)
+		b.WriteString(smodeString(t.Sym, mode))
+		return
 	}
 
 	if int(t.Etype) < len(basicnames) && basicnames[t.Etype] != "" {
+		var name string
 		switch t {
 		case types.Idealbool:
-			return "untyped bool"
+			name = "untyped bool"
 		case types.Idealstring:
-			return "untyped string"
+			name = "untyped string"
 		case types.Idealint:
-			return "untyped int"
+			name = "untyped int"
 		case types.Idealrune:
-			return "untyped rune"
+			name = "untyped rune"
 		case types.Idealfloat:
-			return "untyped float"
+			name = "untyped float"
 		case types.Idealcomplex:
-			return "untyped complex"
+			name = "untyped complex"
+		default:
+			name = basicnames[t.Etype]
 		}
-		return basicnames[t.Etype]
+		b.WriteString(name)
+		return
 	}
 
-	if mode == FDbg {
-		return t.Etype.String() + "-" + typefmt(t, flag, FErr, depth)
+	// At this point, we might call tconv2 recursively. Add the current type to the visited list so we don't
+	// try to print it recursively.
+	// We record the offset in the result buffer where the type's text starts. This offset serves as a reference
+	// point for any later references to the same type.
+	// Note that we remove the type from the visited map as soon as the recursive call is done.
+	// This prevents encoding types like map[*int]*int as map[*int]@4. (That encoding would work,
+	// but I'd like to use the @ notation only when strictly necessary.)
+	if visited == nil {
+		visited = map[*types.Type]int{}
 	}
+	visited[t] = b.Len()
+	defer delete(visited, t)
 
+	if mode == FDbg {
+		b.WriteString(t.Etype.String())
+		b.WriteByte('-')
+		tconv2(b, t, flag, FErr, visited)
+		return
+	}
 	switch t.Etype {
 	case TPTR:
+		b.WriteByte('*')
 		switch mode {
 		case FTypeId, FTypeIdName:
 			if flag&FmtShort != 0 {
-				return "*" + tconv(t.Elem(), FmtShort, mode, depth)
+				tconv2(b, t.Elem(), FmtShort, mode, visited)
+				return
 			}
 		}
-		return "*" + tmodeString(t.Elem(), mode, depth)
+		tconv2(b, t.Elem(), 0, mode, visited)
 
 	case TARRAY:
-		return "[" + strconv.FormatInt(t.NumElem(), 10) + "]" + tmodeString(t.Elem(), mode, depth)
+		b.WriteByte('[')
+		b.WriteString(strconv.FormatInt(t.NumElem(), 10))
+		b.WriteByte(']')
+		tconv2(b, t.Elem(), 0, mode, visited)
 
 	case TSLICE:
-		return "[]" + tmodeString(t.Elem(), mode, depth)
+		b.WriteString("[]")
+		tconv2(b, t.Elem(), 0, mode, visited)
 
 	case TCHAN:
 		switch t.ChanDir() {
 		case types.Crecv:
-			return "<-chan " + tmodeString(t.Elem(), mode, depth)
-
+			b.WriteString("<-chan ")
+			tconv2(b, t.Elem(), 0, mode, visited)
 		case types.Csend:
-			return "chan<- " + tmodeString(t.Elem(), mode, depth)
-		}
-
-		if t.Elem() != nil && t.Elem().IsChan() && t.Elem().Sym == nil && t.Elem().ChanDir() == types.Crecv {
-			return "chan (" + tmodeString(t.Elem(), mode, depth) + ")"
+			b.WriteString("chan<- ")
+			tconv2(b, t.Elem(), 0, mode, visited)
+		default:
+			b.WriteString("chan ")
+			if t.Elem() != nil && t.Elem().IsChan() && t.Elem().Sym == nil && t.Elem().ChanDir() == types.Crecv {
+				b.WriteByte('(')
+				tconv2(b, t.Elem(), 0, mode, visited)
+				b.WriteByte(')')
+			} else {
+				tconv2(b, t.Elem(), 0, mode, visited)
+			}
 		}
-		return "chan " + tmodeString(t.Elem(), mode, depth)
 
 	case TMAP:
-		return "map[" + tmodeString(t.Key(), mode, depth) + "]" + tmodeString(t.Elem(), mode, depth)
+		b.WriteString("map[")
+		tconv2(b, t.Key(), 0, mode, visited)
+		b.WriteByte(']')
+		tconv2(b, t.Elem(), 0, mode, visited)
 
 	case TINTER:
 		if t.IsEmptyInterface() {
-			return "interface {}"
+			b.WriteString("interface {}")
+			break
 		}
-		buf := make([]byte, 0, 64)
-		buf = append(buf, "interface {"...)
+		b.WriteString("interface {")
 		for i, f := range t.Fields().Slice() {
 			if i != 0 {
-				buf = append(buf, ';')
+				b.WriteByte(';')
 			}
-			buf = append(buf, ' ')
+			b.WriteByte(' ')
 			switch {
 			case f.Sym == nil:
 				// Check first that a symbol is defined for this type.
 				// Wrong interface definitions may have types lacking a symbol.
 				break
 			case types.IsExported(f.Sym.Name):
-				buf = append(buf, sconv(f.Sym, FmtShort, mode)...)
+				b.WriteString(sconv(f.Sym, FmtShort, mode))
 			default:
 				flag1 := FmtLeft
 				if flag&FmtUnsigned != 0 {
 					flag1 = FmtUnsigned
 				}
-				buf = append(buf, sconv(f.Sym, flag1, mode)...)
+				b.WriteString(sconv(f.Sym, flag1, mode))
 			}
-			buf = append(buf, tconv(f.Type, FmtShort, mode, depth)...)
+			tconv2(b, f.Type, FmtShort, mode, visited)
 		}
 		if t.NumFields() != 0 {
-			buf = append(buf, ' ')
+			b.WriteByte(' ')
 		}
-		buf = append(buf, '}')
-		return string(buf)
+		b.WriteByte('}')
 
 	case TFUNC:
-		buf := make([]byte, 0, 64)
 		if flag&FmtShort != 0 {
 			// no leading func
 		} else {
 			if t.Recv() != nil {
-				buf = append(buf, "method"...)
-				buf = append(buf, tmodeString(t.Recvs(), mode, depth)...)
-				buf = append(buf, ' ')
+				b.WriteString("method")
+				tconv2(b, t.Recvs(), 0, mode, visited)
+				b.WriteByte(' ')
 			}
-			buf = append(buf, "func"...)
+			b.WriteString("func")
 		}
-		buf = append(buf, tmodeString(t.Params(), mode, depth)...)
+		tconv2(b, t.Params(), 0, mode, visited)
 
 		switch t.NumResults() {
 		case 0:
 			// nothing to do
 
 		case 1:
-			buf = append(buf, ' ')
-			buf = append(buf, tmodeString(t.Results().Field(0).Type, mode, depth)...) // struct->field->field's type
+			b.WriteByte(' ')
+			tconv2(b, t.Results().Field(0).Type, 0, mode, visited) // struct->field->field's type
 
 		default:
-			buf = append(buf, ' ')
-			buf = append(buf, tmodeString(t.Results(), mode, depth)...)
+			b.WriteByte(' ')
+			tconv2(b, t.Results(), 0, mode, visited)
 		}
-		return string(buf)
 
 	case TSTRUCT:
 		if m := t.StructType().Map; m != nil {
 			mt := m.MapType()
 			// Format the bucket struct for map[x]y as map.bucket[x]y.
 			// This avoids a recursive print that generates very long names.
-			var subtype string
 			switch t {
 			case mt.Bucket:
-				subtype = "bucket"
+				b.WriteString("map.bucket[")
 			case mt.Hmap:
-				subtype = "hdr"
+				b.WriteString("map.hdr[")
 			case mt.Hiter:
-				subtype = "iter"
+				b.WriteString("map.iter[")
 			default:
 				Fatalf("unknown internal map type")
 			}
-			return fmt.Sprintf("map.%s[%s]%s", subtype, tmodeString(m.Key(), mode, depth), tmodeString(m.Elem(), mode, depth))
+			tconv2(b, m.Key(), 0, mode, visited)
+			b.WriteByte(']')
+			tconv2(b, m.Elem(), 0, mode, visited)
+			break
 		}
 
-		buf := make([]byte, 0, 64)
 		if funarg := t.StructType().Funarg; funarg != types.FunargNone {
-			buf = append(buf, '(')
+			b.WriteByte('(')
 			var flag1 FmtFlag
 			switch mode {
 			case FTypeId, FTypeIdName, FErr:
@@ -838,42 +907,42 @@ func typefmt(t *types.Type, flag FmtFlag, mode fmtMode, depth int) string {
 			}
 			for i, f := range t.Fields().Slice() {
 				if i != 0 {
-					buf = append(buf, ", "...)
+					b.WriteString(", ")
 				}
-				buf = append(buf, fldconv(f, flag1, mode, depth, funarg)...)
+				fldconv(b, f, flag1, mode, visited, funarg)
 			}
-			buf = append(buf, ')')
+			b.WriteByte(')')
 		} else {
-			buf = append(buf, "struct {"...)
+			b.WriteString("struct {")
 			for i, f := range t.Fields().Slice() {
 				if i != 0 {
-					buf = append(buf, ';')
+					b.WriteByte(';')
 				}
-				buf = append(buf, ' ')
-				buf = append(buf, fldconv(f, FmtLong, mode, depth, funarg)...)
+				b.WriteByte(' ')
+				fldconv(b, f, FmtLong, mode, visited, funarg)
 			}
 			if t.NumFields() != 0 {
-				buf = append(buf, ' ')
+				b.WriteByte(' ')
 			}
-			buf = append(buf, '}')
+			b.WriteByte('}')
 		}
-		return string(buf)
 
 	case TFORW:
+		b.WriteString("undefined")
 		if t.Sym != nil {
-			return "undefined " + smodeString(t.Sym, mode)
+			b.WriteByte(' ')
+			b.WriteString(smodeString(t.Sym, mode))
 		}
-		return "undefined"
 
 	case TUNSAFEPTR:
-		return "unsafe.Pointer"
+		b.WriteString("unsafe.Pointer")
 
 	case Txxx:
-		return "Txxx"
+		b.WriteString("Txxx")
+	default:
+		// Don't know how to handle - fall back to detailed prints.
+		b.WriteString(mode.Sprintf("%v <%v>", t.Etype, t.Sym))
 	}
-
-	// Don't know how to handle - fall back to detailed prints.
-	return mode.Sprintf("%v <%v>", t.Etype, t.Sym)
 }
 
 // Statements which may be rendered with a simplestmt as init.
@@ -1657,15 +1726,11 @@ func sconv(s *types.Sym, flag FmtFlag, mode fmtMode) string {
 	return symfmt(s, flag, mode)
 }
 
-func tmodeString(t *types.Type, mode fmtMode, depth int) string {
-	return tconv(t, 0, mode, depth)
-}
-
-func fldconv(f *types.Field, flag FmtFlag, mode fmtMode, depth int, funarg types.Funarg) string {
+func fldconv(b *bytes.Buffer, f *types.Field, flag FmtFlag, mode fmtMode, visited map[*types.Type]int, funarg types.Funarg) {
 	if f == nil {
-		return "<T>"
+		b.WriteString("<T>")
+		return
 	}
-
 	flag, mode = flag.update(mode)
 	if mode == FTypeIdName {
 		flag |= FmtUnsigned
@@ -1694,27 +1759,26 @@ func fldconv(f *types.Field, flag FmtFlag, mode fmtMode, depth int, funarg types
 		}
 	}
 
-	var typ string
+	if name != "" {
+		b.WriteString(name)
+		b.WriteString(" ")
+	}
+
 	if f.IsDDD() {
 		var et *types.Type
 		if f.Type != nil {
 			et = f.Type.Elem()
 		}
-		typ = "..." + tmodeString(et, mode, depth)
+		b.WriteString("...")
+		tconv2(b, et, 0, mode, visited)
 	} else {
-		typ = tmodeString(f.Type, mode, depth)
-	}
-
-	str := typ
-	if name != "" {
-		str = name + " " + typ
+		tconv2(b, f.Type, 0, mode, visited)
 	}
 
 	if flag&FmtShort == 0 && funarg == types.FunargNone && f.Note != "" {
-		str += " " + strconv.Quote(f.Note)
+		b.WriteString(" ")
+		b.WriteString(strconv.Quote(f.Note))
 	}
-
-	return str
 }
 
 // "%L"  print definition, not name
@@ -1722,58 +1786,12 @@ func fldconv(f *types.Field, flag FmtFlag, mode fmtMode, depth int, funarg types
 func typeFormat(t *types.Type, s fmt.State, verb rune, mode fmtMode) {
 	switch verb {
 	case 'v', 'S', 'L':
-		// This is an external entry point, so we pass depth 0 to tconv.
-		// See comments in Type.String.
-		fmt.Fprint(s, tconv(t, fmtFlag(s, verb), mode, 0))
-
+		fmt.Fprint(s, tconv(t, fmtFlag(s, verb), mode))
 	default:
 		fmt.Fprintf(s, "%%!%c(*Type=%p)", verb, t)
 	}
 }
 
-var deepTypes map[*types.Type]string
-
-// See #16897 before changing the implementation of tconv.
-func tconv(t *types.Type, flag FmtFlag, mode fmtMode, depth int) string {
-	if t == nil {
-		return "<T>"
-	}
-	if t.Etype == types.TSSA {
-		return t.Extra.(string)
-	}
-	if t.Etype == types.TTUPLE {
-		return t.FieldType(0).String() + "," + t.FieldType(1).String()
-	}
-
-	// Avoid endless recursion by setting an upper limit. This also
-	// limits the depths of valid composite types, but they are likely
-	// artificially created.
-	// TODO(gri) should have proper cycle detection here, eventually (issue #29312)
-	// For now, ensure that each of these really deep types are at least uniquely
-	// named, so that such types don't collide in the linker and thus allow security holes.
-	if depth > 250 {
-		if str := deepTypes[t]; str != "" {
-			return str
-		}
-		if deepTypes == nil {
-			deepTypes = map[*types.Type]string{}
-		}
-		id := len(deepTypes)
-		str := fmt.Sprintf("<...uniquetype_%d_in_%s>", id, curpkg().Path)
-		deepTypes[t] = str
-		return str
-	}
-
-	flag, mode = flag.update(mode)
-	if mode == FTypeIdName {
-		flag |= FmtUnsigned
-	}
-
-	str := typefmt(t, flag, mode, depth+1)
-
-	return str
-}
-
 func (n *Node) String() string                 { return fmt.Sprint(n) }
 func (n *Node) modeString(mode fmtMode) string { return mode.Sprint(n) }
 

src/cmd/compile/internal/gc/main.go

@@ -512,8 +512,8 @@ func Main(archInit func(*Arch)) {
 	types.Sconv = func(s *types.Sym, flag, mode int) string {
 		return sconv(s, FmtFlag(flag), fmtMode(mode))
 	}
-	types.Tconv = func(t *types.Type, flag, mode, depth int) string {
-		return tconv(t, FmtFlag(flag), fmtMode(mode), depth)
+	types.Tconv = func(t *types.Type, flag, mode int) string {
+		return tconv(t, FmtFlag(flag), fmtMode(mode))
 	}
 	types.FormatSym = func(sym *types.Sym, s fmt.State, verb rune, mode int) {
 		symFormat(sym, s, verb, fmtMode(mode))

src/cmd/compile/internal/ssa/export_test.go

@@ -153,7 +153,7 @@ func init() {
 	// TODO(josharian): move universe initialization to the types package,
 	// so this test setup can share it.
 
-	types.Tconv = func(t *types.Type, flag, mode, depth int) string {
+	types.Tconv = func(t *types.Type, flag, mode int) string {
 		return t.Etype.String()
 	}
 	types.Sconv = func(s *types.Sym, flag, mode int) string {

src/cmd/compile/internal/types/utils.go

@@ -19,7 +19,7 @@ var (
 	Dowidth     func(*Type)
 	Fatalf      func(string, ...interface{})
 	Sconv       func(*Sym, int, int) string       // orig: func sconv(s *Sym, flag FmtFlag, mode fmtMode) string
-	Tconv       func(*Type, int, int, int) string // orig: func tconv(t *Type, flag FmtFlag, mode fmtMode, depth int) string
+	Tconv       func(*Type, int, int) string      // orig: func tconv(t *Type, flag FmtFlag, mode fmtMode) string
 	FormatSym   func(*Sym, fmt.State, rune, int)  // orig: func symFormat(sym *Sym, s fmt.State, verb rune, mode fmtMode)
 	FormatType  func(*Type, fmt.State, rune, int) // orig: func typeFormat(t *Type, s fmt.State, verb rune, mode fmtMode)
 	TypeLinkSym func(*Type) *obj.LSym
@@ -39,25 +39,23 @@ func (sym *Sym) Format(s fmt.State, verb rune) {
 }
 
 func (t *Type) String() string {
-	// This is an external entry point, so we pass depth 0 to tconv.
 	// The implementation of tconv (including typefmt and fldconv)
-	// must take care not to use a type in a formatting string
-	// to avoid resetting the recursion counter.
-	return Tconv(t, 0, FErr, 0)
+	// must handle recursive types correctly.
+	return Tconv(t, 0, FErr)
 }
 
 // ShortString generates a short description of t.
 // It is used in autogenerated method names, reflection,
 // and itab names.
 func (t *Type) ShortString() string {
-	return Tconv(t, FmtLeft, FErr, 0)
+	return Tconv(t, FmtLeft, FErr)
 }
 
 // LongString generates a complete description of t.
 // It is useful for reflection,
 // or when a unique fingerprint or hash of a type is required.
 func (t *Type) LongString() string {
-	return Tconv(t, FmtLeft|FmtUnsigned, FErr, 0)
+	return Tconv(t, FmtLeft|FmtUnsigned, FErr)
 }
 
 func (t *Type) Format(s fmt.State, verb rune) {