[dev.typeparams] go/types: merge instance and Named to eliminate sanitization

Storing temporary syntactic information using an *instance type forces
us to be careful not to leak references to *instance in the checker
output. This is complex and error prone, as types are written in many
places during type checking.

Instead, temporarily pin the necessary syntactic information directly to
the Named type during the type checking pass. This allows us to avoid
having to sanitize references.

This includes a couple of small, unrelated changes that were made in the
process of debugging:
 - eliminate the expandf indirection: it is no longer necessary
 - include type parameters when printing objects

For #46151

Change-Id: I767e35b289f2fea512a168997af0f861cd242175
Reviewed-on: https://go-review.googlesource.com/c/go/+/335929
Trust: Robert Findley <rfindley@google.com>
Trust: Robert Griesemer <gri@golang.org>
Run-TryBot: Robert Findley <rfindley@google.com>
TryBot-Result: Go Bot <gobot@golang.org>
Reviewed-by: Robert Griesemer <gri@golang.org>

src/go/types/check.go

@@ -273,10 +273,6 @@ func (check *Checker) checkFiles(files []*ast.File) (err error) {
 
 	check.recordUntyped()
 
-	if check.Info != nil {
-		sanitizeInfo(check.Info)
-	}
-
 	check.pkg.complete = true
 
 	// no longer needed - release memory

src/go/types/decl.go

@@ -317,6 +317,7 @@ func (check *Checker) validType(typ Type, path []Object) typeInfo {
 		}
 
 	case *Named:
+		t.complete()
 		// don't touch the type if it is from a different package or the Universe scope
 		// (doing so would lead to a race condition - was issue #35049)
 		if t.obj.pkg != check.pkg {
@@ -349,9 +350,6 @@ func (check *Checker) validType(typ Type, path []Object) typeInfo {
 			panic("internal error: cycle start not found")
 		}
 		return t.info
-
-	case *instance:
-		return check.validType(t.expand(), path)
 	}
 
 	return valid
@@ -607,6 +605,7 @@ func (check *Checker) typeDecl(obj *TypeName, tdecl *ast.TypeSpec, def *Named) {
 
 	// determine underlying type of named
 	named.fromRHS = check.definedType(tdecl.Type, named)
+	assert(named.fromRHS != nil)
 
 	// The underlying type of named may be itself a named type that is
 	// incomplete:
@@ -685,7 +684,8 @@ func (check *Checker) boundType(e ast.Expr) Type {
 
 	bound := check.typ(e)
 	check.later(func() {
-		if _, ok := under(bound).(*Interface); !ok && bound != Typ[Invalid] {
+		u := under(bound)
+		if _, ok := u.(*Interface); !ok && u != Typ[Invalid] {
 			check.errorf(e, _Todo, "%s is not an interface", bound)
 		}
 	})

src/go/types/infer.go

@@ -337,9 +337,6 @@ func (w *tpWalker) isParameterized(typ Type) (res bool) {
 		// t must be one of w.tparams
 		return t.index < len(w.tparams) && w.tparams[t.index].typ == t
 
-	case *instance:
-		return w.isParameterizedList(t.targs)
-
 	default:
 		unreachable()
 	}

src/go/types/instance.go

@@ -4,56 +4,39 @@
 
 package types
 
+// TODO(rfindley): move this code to named.go.
+
 import "go/token"
 
-// An instance represents an instantiated generic type syntactically
+// instance holds a Checker along with syntactic information
-// (without expanding the instantiation). Type instances appear only
+// information, for use in lazy instantiation.
-// during type-checking and are replaced by their fully instantiated
-// (expanded) types before the end of type-checking.
 type instance struct {
-	check   *Checker    // for lazy instantiation
+	check   *Checker
 	pos     token.Pos   // position of type instantiation; for error reporting only
-	base    *Named      // parameterized type to be instantiated
-	targs   []Type      // type arguments
 	posList []token.Pos // position of each targ; for error reporting only
 	verify  bool        // if set, constraint satisfaction is verified
-	value   Type        // base[targs...] after instantiation or Typ[Invalid]; nil if not yet set
 }
 
-// expand returns the instantiated (= expanded) type of t.
+// complete ensures that the underlying type of n is instantiated.
-// The result is either an instantiated *Named type, or
+// The underlying type will be Typ[Invalid] if there was an error.
-// Typ[Invalid] if there was an error.
+// TODO(rfindley): expand would be a better name for this method, but conflicts
-func (t *instance) expand() Type {
+// with the existing concept of lazy expansion. Need to reconcile this.
-	v := t.value
+func (n *Named) complete() {
-	if v == nil {
+	if n.instance != nil && len(n.targs) > 0 && n.underlying == nil {
-		v = t.check.Instantiate(t.pos, t.base, t.targs, t.posList, t.verify)
+		check := n.instance.check
-		if v == nil {
+		inst := check.instantiate(n.instance.pos, n.orig.underlying, n.tparams, n.targs, n.instance.posList, n.instance.verify)
-			v = Typ[Invalid]
+		n.underlying = inst
-		}
+		n.fromRHS = inst
-		t.value = v
+		n.methods = n.orig.methods
 	}
-	// After instantiation we must have an invalid or a *Named type.
-	if debug && v != Typ[Invalid] {
-		_ = v.(*Named)
-	}
-	return v
 }
 
 // expand expands a type instance into its instantiated
 // type and leaves all other types alone. expand does
 // not recurse.
 func expand(typ Type) Type {
-	if t, _ := typ.(*instance); t != nil {
+	if t, _ := typ.(*Named); t != nil {
-		return t.expand()
+		t.complete()
 	}
 	return typ
 }
-
-// expandf is set to expand.
-// Call expandf when calling expand causes compile-time cycle error.
-var expandf func(Type) Type
-
-func init() { expandf = expand }
-
-func (t *instance) Underlying() Type { return t }
-func (t *instance) String() string   { return TypeString(t, nil) }

src/go/types/instantiate.go

@@ -25,29 +25,6 @@ import (
 // Any methods attached to a *Named are simply copied; they are not
 // instantiated.
 func (check *Checker) Instantiate(pos token.Pos, typ Type, targs []Type, posList []token.Pos, verify bool) (res Type) {
-	if verify && check == nil {
-		panic("cannot have nil receiver if verify is set")
-	}
-
-	if check != nil && trace {
-		check.trace(pos, "-- instantiating %s with %s", typ, typeListString(targs))
-		check.indent++
-		defer func() {
-			check.indent--
-			var under Type
-			if res != nil {
-				// Calling under() here may lead to endless instantiations.
-				// Test case: type T[P any] T[P]
-				// TODO(gri) investigate if that's a bug or to be expected.
-				under = res.Underlying()
-			}
-			check.trace(pos, "=> %s (under = %s)", res, under)
-		}()
-	}
-
-	assert(len(posList) <= len(targs))
-
-	// TODO(gri) What is better here: work with TypeParams, or work with TypeNames?
 	var tparams []*TypeName
 	switch t := typ.(type) {
 	case *Named:
@@ -77,6 +54,10 @@ func (check *Checker) Instantiate(pos token.Pos, typ Type, targs []Type, posList
 		panic(fmt.Sprintf("%v: cannot instantiate %v", pos, typ))
 	}
 
+	return check.instantiate(pos, typ, tparams, targs, posList, verify)
+}
+
+func (check *Checker) instantiate(pos token.Pos, typ Type, tparams []*TypeName, targs []Type, posList []token.Pos, verify bool) (res Type) {
 	// the number of supplied types must match the number of type parameters
 	if len(targs) != len(tparams) {
 		// TODO(gri) provide better error message
@@ -86,6 +67,29 @@ func (check *Checker) Instantiate(pos token.Pos, typ Type, targs []Type, posList
 		}
 		panic(fmt.Sprintf("%v: got %d arguments but %d type parameters", pos, len(targs), len(tparams)))
 	}
+	if verify && check == nil {
+		panic("cannot have nil receiver if verify is set")
+	}
+
+	if check != nil && trace {
+		check.trace(pos, "-- instantiating %s with %s", typ, typeListString(targs))
+		check.indent++
+		defer func() {
+			check.indent--
+			var under Type
+			if res != nil {
+				// Calling under() here may lead to endless instantiations.
+				// Test case: type T[P any] T[P]
+				// TODO(gri) investigate if that's a bug or to be expected.
+				under = res.Underlying()
+			}
+			check.trace(pos, "=> %s (under = %s)", res, under)
+		}()
+	}
+
+	assert(len(posList) <= len(targs))
+
+	// TODO(gri) What is better here: work with TypeParams, or work with TypeNames?
 
 	if len(tparams) == 0 {
 		return typ // nothing to do (minor optimization)
@@ -120,15 +124,26 @@ func (check *Checker) InstantiateLazy(pos token.Pos, typ Type, targs []Type, pos
 	if base == nil {
 		panic(fmt.Sprintf("%v: cannot instantiate %v", pos, typ))
 	}
+	h := instantiatedHash(base, targs)
+	if check != nil {
+		if named := check.typMap[h]; named != nil {
+			return named
+		}
+	}
 
-	return &instance{
+	tname := NewTypeName(pos, base.obj.pkg, base.obj.name, nil)
+	named := check.newNamed(tname, base, nil, base.tparams, base.methods) // methods are instantiated lazily
+	named.targs = targs
+	named.instance = &instance{
 		check:   check,
 		pos:     pos,
-		base:    base,
-		targs:   targs,
 		posList: posList,
 		verify:  verify,
 	}
+	if check != nil {
+		check.typMap[h] = named
+	}
+	return named
 }
 
 // satisfies reports whether the type argument targ satisfies the constraint of type parameter

src/go/types/named.go

@@ -10,7 +10,7 @@ import "sync"
 
 // A Named represents a named (defined) type.
 type Named struct {
-	check      *Checker    // for Named.under implementation; nilled once under has been called
+	instance   *instance   // syntactic information for lazy instantiation
 	info       typeInfo    // for cycle detection
 	obj        *TypeName   // corresponding declared object
 	orig       *Named      // original, uninstantiated type
@@ -65,7 +65,13 @@ func (t *Named) expand() *Named {
 
 // newNamed is like NewNamed but with a *Checker receiver and additional orig argument.
 func (check *Checker) newNamed(obj *TypeName, orig *Named, underlying Type, tparams []*TypeName, methods []*Func) *Named {
-	typ := &Named{check: check, obj: obj, orig: orig, fromRHS: underlying, underlying: underlying, tparams: tparams, methods: methods}
+	var inst *instance
+	if check != nil {
+		inst = &instance{
+			check: check,
+		}
+	}
+	typ := &Named{instance: inst, obj: obj, orig: orig, fromRHS: underlying, underlying: underlying, tparams: tparams, methods: methods}
 	if typ.orig == nil {
 		typ.orig = typ
 	}
@@ -83,10 +89,10 @@ func (check *Checker) newNamed(obj *TypeName, orig *Named, underlying Type, tpar
 	if check != nil {
 		check.later(func() {
 			switch typ.under().(type) {
-			case *Named, *instance:
+			case *Named:
 				panic("internal error: unexpanded underlying type")
 			}
-			typ.check = nil
+			typ.instance = nil
 		})
 	}
 	return typ
@@ -153,6 +159,8 @@ func (t *Named) String() string   { return TypeString(t, nil) }
 // is detected, the result is Typ[Invalid]. If a cycle is detected and
 // n0.check != nil, the cycle is reported.
 func (n0 *Named) under() Type {
+	n0.complete()
+
 	u := n0.Underlying()
 
 	if u == Typ[Invalid] {
@@ -168,17 +176,17 @@ func (n0 *Named) under() Type {
 	default:
 		// common case
 		return u
-	case *Named, *instance:
+	case *Named:
 		// handled below
 	}
 
-	if n0.check == nil {
+	if n0.instance == nil || n0.instance.check == nil {
 		panic("internal error: Named.check == nil but type is incomplete")
 	}
 
 	// Invariant: after this point n0 as well as any named types in its
 	// underlying chain should be set up when this function exits.
-	check := n0.check
+	check := n0.instance.check
 
 	// If we can't expand u at this point, it is invalid.
 	n := asNamed(u)
@@ -199,12 +207,8 @@ func (n0 *Named) under() Type {
 		var n1 *Named
 		switch u1 := u.(type) {
 		case *Named:
+			u1.complete()
 			n1 = u1
-		case *instance:
-			n1, _ = u1.expand().(*Named)
-			if n1 == nil {
-				u = Typ[Invalid]
-			}
 		}
 		if n1 == nil {
 			break // end of chain

src/go/types/object.go

@@ -429,6 +429,9 @@ func writeObject(buf *bytes.Buffer, obj Object, qf Qualifier) {
 		if _, ok := typ.(*Basic); ok {
 			return
 		}
+		if named, _ := typ.(*Named); named != nil && len(named.tparams) > 0 {
+			writeTParamList(buf, named.tparams, qf, nil)
+		}
 		if tname.IsAlias() {
 			buf.WriteString(" =")
 		} else {

src/go/types/predicates.go

@@ -10,7 +10,7 @@ package types
 // isNamed may be called with types that are not fully set up.
 func isNamed(typ Type) bool {
 	switch typ.(type) {
-	case *Basic, *Named, *TypeParam, *instance:
+	case *Basic, *Named, *TypeParam:
 		return true
 	}
 	return false
@@ -159,8 +159,8 @@ func (p *ifacePair) identical(q *ifacePair) bool {
 // For changes to this code the corresponding changes should be made to unifier.nify.
 func identical(x, y Type, cmpTags bool, p *ifacePair) bool {
 	// types must be expanded for comparison
-	x = expandf(x)
+	x = expand(x)
-	y = expandf(y)
+	y = expand(y)
 
 	if x == y {
 		return true

src/go/types/sanitize.go

@@ -1,206 +0,0 @@
-// Copyright 2020 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 types
-
-// sanitizeInfo walks the types contained in info to ensure that all instances
-// are expanded.
-//
-// This includes some objects that may be shared across concurrent
-// type-checking passes (such as those in the universe scope), so we are
-// careful here not to write types that are already sanitized. This avoids a
-// data race as any shared types should already be sanitized.
-func sanitizeInfo(info *Info) {
-	var s sanitizer = make(map[Type]Type)
-
-	// Note: Some map entries are not references.
-	// If modified, they must be assigned back.
-
-	for e, tv := range info.Types {
-		if typ := s.typ(tv.Type); typ != tv.Type {
-			tv.Type = typ
-			info.Types[e] = tv
-		}
-	}
-
-	inferred := info.Inferred
-	for e, inf := range inferred {
-		changed := false
-		for i, targ := range inf.TArgs {
-			if typ := s.typ(targ); typ != targ {
-				inf.TArgs[i] = typ
-				changed = true
-			}
-		}
-		if typ := s.typ(inf.Sig); typ != inf.Sig {
-			inf.Sig = typ.(*Signature)
-			changed = true
-		}
-		if changed {
-			inferred[e] = inf
-		}
-	}
-
-	for _, obj := range info.Defs {
-		if obj != nil {
-			if typ := s.typ(obj.Type()); typ != obj.Type() {
-				obj.setType(typ)
-			}
-		}
-	}
-
-	for _, obj := range info.Uses {
-		if obj != nil {
-			if typ := s.typ(obj.Type()); typ != obj.Type() {
-				obj.setType(typ)
-			}
-		}
-	}
-
-	// TODO(gri) sanitize as needed
-	// - info.Implicits
-	// - info.Selections
-	// - info.Scopes
-	// - info.InitOrder
-}
-
-type sanitizer map[Type]Type
-
-func (s sanitizer) typ(typ Type) Type {
-	if typ == nil {
-		return nil
-	}
-
-	if t, found := s[typ]; found {
-		return t
-	}
-	s[typ] = typ
-
-	switch t := typ.(type) {
-	case *Basic, *top:
-		// nothing to do
-
-	case *Array:
-		if elem := s.typ(t.elem); elem != t.elem {
-			t.elem = elem
-		}
-
-	case *Slice:
-		if elem := s.typ(t.elem); elem != t.elem {
-			t.elem = elem
-		}
-
-	case *Struct:
-		s.varList(t.fields)
-
-	case *Pointer:
-		if base := s.typ(t.base); base != t.base {
-			t.base = base
-		}
-
-	case *Tuple:
-		s.tuple(t)
-
-	case *Signature:
-		s.var_(t.recv)
-		s.tuple(t.params)
-		s.tuple(t.results)
-
-	case *Union:
-		s.typeList(t.types)
-
-	case *Interface:
-		s.funcList(t.methods)
-		s.typeList(t.embeddeds)
-		// TODO(gri) do we need to sanitize type sets?
-		tset := t.typeSet()
-		s.funcList(tset.methods)
-		if types := s.typ(tset.types); types != tset.types {
-			tset.types = types
-		}
-
-	case *Map:
-		if key := s.typ(t.key); key != t.key {
-			t.key = key
-		}
-		if elem := s.typ(t.elem); elem != t.elem {
-			t.elem = elem
-		}
-
-	case *Chan:
-		if elem := s.typ(t.elem); elem != t.elem {
-			t.elem = elem
-		}
-
-	case *Named:
-		if debug && t.check != nil {
-			panic("internal error: Named.check != nil")
-		}
-		t.expand()
-		if orig := s.typ(t.fromRHS); orig != t.fromRHS {
-			t.fromRHS = orig
-		}
-		if under := s.typ(t.underlying); under != t.underlying {
-			t.underlying = under
-		}
-		s.typeList(t.targs)
-		s.funcList(t.methods)
-
-	case *TypeParam:
-		if bound := s.typ(t.bound); bound != t.bound {
-			t.bound = bound
-		}
-
-	case *instance:
-		typ = t.expand()
-		s[t] = typ
-
-	default:
-		panic("unimplemented")
-	}
-
-	return typ
-}
-
-func (s sanitizer) var_(v *Var) {
-	if v != nil {
-		if typ := s.typ(v.typ); typ != v.typ {
-			v.typ = typ
-		}
-	}
-}
-
-func (s sanitizer) varList(list []*Var) {
-	for _, v := range list {
-		s.var_(v)
-	}
-}
-
-func (s sanitizer) tuple(t *Tuple) {
-	if t != nil {
-		s.varList(t.vars)
-	}
-}
-
-func (s sanitizer) func_(f *Func) {
-	if f != nil {
-		if typ := s.typ(f.typ); typ != f.typ {
-			f.typ = typ
-		}
-	}
-}
-
-func (s sanitizer) funcList(list []*Func) {
-	for _, f := range list {
-		s.func_(f)
-	}
-}
-
-func (s sanitizer) typeList(list []Type) {
-	for i, t := range list {
-		if typ := s.typ(t); typ != t {
-			list[i] = typ
-		}
-	}
-}

src/go/types/sizeof_test.go

@@ -32,7 +32,6 @@ func TestSizeof(t *testing.T) {
 		{Chan{}, 12, 24},
 		{Named{}, 84, 160},
 		{TypeParam{}, 28, 48},
-		{instance{}, 48, 96},
 		{top{}, 0, 0},
 
 		// Objects

src/go/types/subst.go

@@ -29,12 +29,7 @@ func makeSubstMap(tpars []*TypeName, targs []Type) *substMap {
 	assert(len(tpars) == len(targs))
 	proj := make(map[*TypeParam]Type, len(tpars))
 	for i, tpar := range tpars {
-		// We must expand type arguments otherwise *instance
+		proj[tpar.typ.(*TypeParam)] = targs[i]
-		// types end up as components in composite types.
-		// TODO(gri) explain why this causes problems, if it does
-		targ := expand(targs[i]) // possibly nil
-		targs[i] = targ
-		proj[tpar.typ.(*TypeParam)] = targ
 	}
 	return &substMap{targs, proj}
 }
@@ -86,6 +81,7 @@ func (check *Checker) subst(pos token.Pos, typ Type, smap *substMap) Type {
 		// for recursive types (example: type T[P any] *T[P]).
 		subst.typMap = make(map[string]*Named)
 	}
+
 	return subst.typ(typ)
 }
 
@@ -248,10 +244,13 @@ func (subst *subster) typ(typ Type) Type {
 		named := subst.check.newNamed(tname, t, t.Underlying(), t.TParams(), t.methods) // method signatures are updated lazily
 		named.targs = newTargs
 		subst.typMap[h] = named
+		t.complete() // must happen after typMap update to avoid infinite recursion
 
 		// do the substitution
 		dump(">>> subst %s with %s (new: %s)", t.underlying, subst.smap, newTargs)
 		named.underlying = subst.typOrNil(t.Underlying())
+		dump(">>> underlying: %v", named.underlying)
+		assert(named.underlying != nil)
 		named.fromRHS = named.underlying // for cycle detection (Checker.validType)
 
 		return named
@@ -259,10 +258,6 @@ func (subst *subster) typ(typ Type) Type {
 	case *TypeParam:
 		return subst.smap.lookup(t)
 
-	case *instance:
-		// TODO(gri) can we avoid the expansion here and just substitute the type parameters?
-		return subst.typ(t.expand())
-
 	default:
 		panic("unimplemented")
 	}

src/go/types/testdata/check/issues.go2

@@ -81,8 +81,10 @@ func (u T2[U]) Add1() U {
     return u.s + 1
 }
 
+// TODO(rfindley): we should probably report an error here as well, not
+//                 just when the type is first instantiated.
 func NewT2[U any]() T2[U /* ERROR U has no type constraints */ ] {
-    return T2[U /* ERROR U has no type constraints */ ]{}
+    return T2[U]{}
 }
 
 func _() {

src/go/types/typeparam.go

@@ -24,7 +24,8 @@ type TypeParam struct {
 	id    uint64    // unique id, for debugging only
 	obj   *TypeName // corresponding type name
 	index int       // type parameter index in source order, starting at 0
-	bound Type      // *Named or *Interface; underlying type is always *Interface
+	// TODO(rfindley): this could also be Typ[Invalid]. Verify that this is handled correctly.
+	bound Type // *Named or *Interface; underlying type is always *Interface
 }
 
 // NewTypeParam returns a new TypeParam.  bound can be nil (and set later).

src/go/types/typestring.go

@@ -295,13 +295,6 @@ func writeType(buf *bytes.Buffer, typ Type, qf Qualifier, visited []Type) {
 		}
 		buf.WriteString(s + subscript(t.id))
 
-	case *instance:
-		buf.WriteByte(instanceMarker) // indicate "non-evaluated" syntactic instance
-		writeTypeName(buf, t.base.obj, qf)
-		buf.WriteByte('[')
-		writeTypeList(buf, t.targs, qf, visited)
-		buf.WriteByte(']')
-
 	case *top:
 		buf.WriteString("⊤")
 

src/go/types/typexpr.go

@@ -436,7 +436,7 @@ func (check *Checker) instantiatedType(x ast.Expr, targsx []ast.Expr, def *Named
 	// make sure we check instantiation works at least once
 	// and that the resulting type is valid
 	check.later(func() {
-		t := typ.(*instance).expand()
+		t := expand(typ)
 		check.validType(t, nil)
 	})