bexport.go

// Copyright 2015 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.

// Binary package export.

/*
1) Export data encoding principles:

The export data is a serialized description of the graph of exported
"objects": constants, types, variables, and functions. Aliases may be
directly reexported, and unaliased types may be indirectly reexported
(as part of the type of a directly exported object). More generally,
objects referred to from inlined function bodies can be reexported.
We need to know which package declares these reexported objects, and
therefore packages are also part of the export graph.

The roots of the graph are two lists of objects. The 1st list (phase 1,
see Export) contains all objects that are exported at the package level.
These objects are the full representation of the package's API, and they
are the only information a platform-independent tool (e.g., go/types)
needs to know to type-check against a package.

The 2nd list of objects contains all objects referred to from exported
inlined function bodies. These objects are needed by the compiler to
make sense of the function bodies; the exact list contents are compiler-
specific.

Finally, the export data contains a list of representations for inlined
function bodies. The format of this representation is compiler specific.

The graph is serialized in in-order fashion, starting with the roots.
Each object in the graph is serialized by writing its fields sequentially.
If the field is a pointer to another object, that object is serialized in
place, recursively. Otherwise the field is written in place. Non-pointer
fields are all encoded as integer or string values.

Some objects (packages, types) may be referred to more than once. When
reaching an object that was not serialized before, an integer _index_
is assigned to it, starting at 0. In this case, the encoding starts
with an integer _tag_ < 0. The tag value indicates the kind of object
that follows and that this is the first time that we see this object.
If the object was already serialized, the encoding is simply the object
index >= 0. An importer can trivially determine if an object needs to
be read in for the first time (tag < 0) and entered into the respective
object table, or if the object was seen already (index >= 0), in which
case the index is used to look up the object in the respective table.

Before exporting or importing, the type tables are populated with the
predeclared types (int, string, error, unsafe.Pointer, etc.). This way
they are automatically encoded with a known and fixed type index.

2) Encoding format:

The export data starts with two newline-terminated strings: a version
string and either an empty string, or "debug", when emitting the debug
format. These strings are followed by version-specific encoding options.

(The Go1.7 version starts with a couple of bytes specifying the format.
That format encoding is no longer used but is supported to avoid spurious
errors when importing old installed package files.)

This header is followed by the package object for the exported package,
two lists of objects, and the list of inlined function bodies.

The encoding of objects is straight-forward: Constants, variables, and
functions start with their name, type, and possibly a value. Named types
record their name and package so that they can be canonicalized: If the
same type was imported before via another import, the importer must use
the previously imported type pointer so that we have exactly one version
(i.e., one pointer) for each named type (and read but discard the current
type encoding). Unnamed types simply encode their respective fields.
Aliases are encoded starting with their name followed by the qualified
identifier denoting the original (aliased) object, which was exported
earlier.

In the encoding, some lists start with the list length. Some lists are
terminated with an end marker (usually for lists where we may not know
the length a priori).

Integers use variable-length encoding for compact representation.

Strings are canonicalized similar to objects that may occur multiple times:
If the string was exported already, it is represented by its index only.
Otherwise, the export data starts with the negative string length (negative,
so we can distinguish from string index), followed by the string bytes.
The empty string is mapped to index 0. (The initial format string is an
exception; it is encoded as the string bytes followed by a newline).

The exporter and importer are completely symmetric in implementation: For
each encoding routine there is a matching and symmetric decoding routine.
This symmetry makes it very easy to change or extend the format: If a new
field needs to be encoded, a symmetric change can be made to exporter and
importer.

3) Making changes to the encoding format:

Any change to the encoding format requires a respective change in the
exporter below and a corresponding symmetric change to the importer in
bimport.go.

Furthermore, it requires a corresponding change to go/internal/gcimporter
and golang.org/x/tools/go/gcimporter15. Changes to the latter must preserve
compatibility with both the last release of the compiler, and with the
corresponding compiler at tip. That change is necessarily more involved,
as it must switch based on the version number in the export data file.

It is recommended to turn on debugFormat temporarily when working on format
changes as it will help finding encoding/decoding inconsistencies quickly.
*/

package gc

import (
	"bufio"
	"bytes"
	"cmd/compile/internal/types"
	"encoding/binary"
	"fmt"
	"math/big"
	"sort"
	"strings"
)

// If debugFormat is set, each integer and string value is preceded by a marker
// and position information in the encoding. This mechanism permits an importer
// to recognize immediately when it is out of sync. The importer recognizes this
// mode automatically (i.e., it can import export data produced with debugging
// support even if debugFormat is not set at the time of import). This mode will
// lead to massively larger export data (by a factor of 2 to 3) and should only
// be enabled during development and debugging.
//
// NOTE: This flag is the first flag to enable if importing dies because of
// (suspected) format errors, and whenever a change is made to the format.
const debugFormat = false // default: false

// Current export format version. Increase with each format change.
// 5: improved position encoding efficiency (issue 20080, CL 41619)
// 4: type name objects support type aliases, uses aliasTag
// 3: Go1.8 encoding (same as version 2, aliasTag defined but never used)
// 2: removed unused bool in ODCL export (compiler only)
// 1: header format change (more regular), export package for _ struct fields
// 0: Go1.7 encoding
const exportVersion = 5

// exportInlined enables the export of inlined function bodies and related
// dependencies. The compiler should work w/o any loss of functionality with
// the flag disabled, but the generated code will lose access to inlined
// function bodies across packages, leading to performance bugs.
// Leave for debugging.
const exportInlined = true // default: true

// trackAllTypes enables cycle tracking for all types, not just named
// types. The existing compiler invariants assume that unnamed types
// that are not completely set up are not used, or else there are spurious
// errors.
// If disabled, only named types are tracked, possibly leading to slightly
// less efficient encoding in rare cases. It also prevents the export of
// some corner-case type declarations (but those were not handled correctly
// with the former textual export format either).
// Note that when a type is only seen once, as many unnamed types are,
// it is less efficient to track it, since we then also record an index for it.
// See CLs 41622 and 41623 for some data and discussion.
// TODO(gri) enable selectively and remove once issues caused by it are fixed
const trackAllTypes = false

type exporter struct {
	out *bufio.Writer

	// object -> index maps, indexed in order of serialization
	strIndex  map[string]int
	pathIndex map[string]int
	pkgIndex  map[*types.Pkg]int
	typIndex  map[*types.Type]int
	funcList  []*Func

	marked map[*types.Type]bool // types already seen by markType

	// position encoding
	posInfoFormat bool
	prevFile      string
	prevLine      int

	// debugging support
	written int // bytes written
	indent  int // for p.trace
	trace   bool
}

// export writes the exportlist for localpkg to out and returns the number of bytes written.
func export(out *bufio.Writer, trace bool) int {
	p := exporter{
		out:           out,
		strIndex:      map[string]int{"": 0}, // empty string is mapped to 0
		pathIndex:     map[string]int{"": 0}, // empty path is mapped to 0
		pkgIndex:      make(map[*types.Pkg]int),
		typIndex:      make(map[*types.Type]int),
		posInfoFormat: true,
		trace:         trace,
	}

	// write version info
	// The version string must start with "version %d" where %d is the version
	// number. Additional debugging information may follow after a blank; that
	// text is ignored by the importer.
	p.rawStringln(fmt.Sprintf("version %d", exportVersion))
	var debug string
	if debugFormat {
		debug = "debug"
	}
	p.rawStringln(debug) // cannot use p.bool since it's affected by debugFormat; also want to see this clearly
	p.bool(trackAllTypes)
	p.bool(p.posInfoFormat)

	// --- generic export data ---

	// populate type map with predeclared "known" types
	predecl := predeclared()
	for index, typ := range predecl {
		p.typIndex[typ] = index
	}
	if len(p.typIndex) != len(predecl) {
		Fatalf("exporter: duplicate entries in type map?")
	}

	// write package data
	if localpkg.Path != "" {
		Fatalf("exporter: local package path not empty: %q", localpkg.Path)
	}
	p.pkg(localpkg)
	if p.trace {
		p.tracef("\n")
	}

	// Mark all inlineable functions that the importer could call.
	// This is done by tracking down all inlineable methods
	// reachable from exported types.
	p.marked = make(map[*types.Type]bool)
	for _, n := range exportlist {
		sym := n.Sym
		if sym.Exported() {
			// Closures are added to exportlist, but with Exported
			// already set. The export code below skips over them, so
			// we have to here as well.
			// TODO(mdempsky): Investigate why. This seems suspicious.
			continue
		}
		p.markType(asNode(sym.Def).Type)
	}
	p.marked = nil

	// export objects
	//
	// First, export all exported (package-level) objects; i.e., all objects
	// in the current exportlist. These objects represent all information
	// required to import this package and type-check against it; i.e., this
	// is the platform-independent export data. The format is generic in the
	// sense that different compilers can use the same representation.
	//
	// During this first phase, more objects may be added to the exportlist
	// (due to inlined function bodies and their dependencies). Export those
	// objects in a second phase. That data is platform-specific as it depends
	// on the inlining decisions of the compiler and the representation of the
	// inlined function bodies.

	// remember initial exportlist length
	var numglobals = len(exportlist)

	// Phase 1: Export objects in _current_ exportlist; exported objects at
	//          package level.
	// Use range since we want to ignore objects added to exportlist during
	// this phase.
	objcount := 0
	for _, n := range exportlist {
		sym := n.Sym

		if sym.Exported() {
			continue
		}
		sym.SetExported(true)

		// TODO(gri) Closures have dots in their names;
		// e.g., TestFloatZeroValue.func1 in math/big tests.
		if strings.Contains(sym.Name, ".") {
			Fatalf("exporter: unexpected symbol: %v", sym)
		}

		if sym.Def == nil {
			Fatalf("exporter: unknown export symbol: %v", sym)
		}

		// TODO(gri) Optimization: Probably worthwhile collecting
		// long runs of constants and export them "in bulk" (saving
		// tags and types, and making import faster).

		if p.trace {
			p.tracef("\n")
		}
		p.obj(sym)
		objcount++
	}

	// indicate end of list
	if p.trace {
		p.tracef("\n")
	}
	p.tag(endTag)

	// for self-verification only (redundant)
	p.int(objcount)

	// --- compiler-specific export data ---

	if p.trace {
		p.tracef("\n--- compiler-specific export data ---\n[ ")
		if p.indent != 0 {
			Fatalf("exporter: incorrect indentation")
		}
	}

	// write compiler-specific flags
	if p.trace {
		p.tracef("\n")
	}

	// Phase 2: Export objects added to exportlist during phase 1.
	// Don't use range since exportlist may grow during this phase
	// and we want to export all remaining objects.
	objcount = 0
	for i := numglobals; exportInlined && i < len(exportlist); i++ {
		n := exportlist[i]
		sym := n.Sym

		// TODO(gri) The rest of this loop body is identical with
		// the loop body above. Leave alone for now since there
		// are different optimization opportunities, but factor
		// eventually.

		if sym.Exported() {
			continue
		}
		sym.SetExported(true)

		// TODO(gri) Closures have dots in their names;
		// e.g., TestFloatZeroValue.func1 in math/big tests.
		if strings.Contains(sym.Name, ".") {
			Fatalf("exporter: unexpected symbol: %v", sym)
		}

		if sym.Def == nil {
			Fatalf("exporter: unknown export symbol: %v", sym)
		}

		// TODO(gri) Optimization: Probably worthwhile collecting
		// long runs of constants and export them "in bulk" (saving
		// tags and types, and making import faster).

		if p.trace {
			p.tracef("\n")
		}

		if IsAlias(sym) {
			Fatalf("exporter: unexpected type alias %v in inlined function body", sym)
		}

		p.obj(sym)
		objcount++
	}

	// indicate end of list
	if p.trace {
		p.tracef("\n")
	}
	p.tag(endTag)

	// for self-verification only (redundant)
	p.int(objcount)

	// --- inlined function bodies ---

	if p.trace {
		p.tracef("\n--- inlined function bodies ---\n")
		if p.indent != 0 {
			Fatalf("exporter: incorrect indentation")
		}
	}

	// write inlineable function bodies
	// Don't use range since funcList may grow.
	objcount = 0
	for i := 0; i < len(p.funcList); i++ {
		if f := p.funcList[i]; f != nil {
			// function has inlineable body:
			// write index and body
			if p.trace {
				p.tracef("\n----\nfunc { %#v }\n", f.Inl)
			}
			p.int(i)
			p.int(int(f.InlCost))
			p.stmtList(f.Inl)
			if p.trace {
				p.tracef("\n")
			}
			objcount++
		}
	}

	// indicate end of list
	if p.trace {
		p.tracef("\n")
	}
	p.int(-1) // invalid index terminates list

	// for self-verification only (redundant)
	p.int(objcount)

	if p.trace {
		p.tracef("\n--- end ---\n")
	}

	// --- end of export data ---

	return p.written
}

func (p *exporter) pkg(pkg *types.Pkg) {
	if pkg == nil {
		Fatalf("exporter: unexpected nil pkg")
	}

	// if we saw the package before, write its index (>= 0)
	if i, ok := p.pkgIndex[pkg]; ok {
		p.index('P', i)
		return
	}

	// otherwise, remember the package, write the package tag (< 0) and package data
	if p.trace {
		p.tracef("P%d = { ", len(p.pkgIndex))
		defer p.tracef("} ")
	}
	p.pkgIndex[pkg] = len(p.pkgIndex)

	p.tag(packageTag)
	p.string(pkg.Name)
	p.path(pkg.Path)
}

func unidealType(typ *types.Type, val Val) *types.Type {
	// Untyped (ideal) constants get their own type. This decouples
	// the constant type from the encoding of the constant value.
	if typ == nil || typ.IsUntyped() {
		typ = untype(val.Ctype())
	}
	return typ
}

// markType recursively visits types reachable from t to identify
// functions whose inline bodies may be needed.
func (p *exporter) markType(t *types.Type) {
	if p.marked[t] {
		return
	}
	p.marked[t] = true

	// If this is a named type, mark all of its associated
	// methods. Skip interface types because t.Methods contains
	// only their unexpanded method set (i.e., exclusive of
	// interface embeddings), and the switch statement below
	// handles their full method set.
	if t.Sym != nil && t.Etype != TINTER {
		for _, m := range t.Methods().Slice() {
			if exportname(m.Sym.Name) {
				p.markType(m.Type)
			}
		}
	}

	// Recursively mark any types that can be produced given a
	// value of type t: dereferencing a pointer; indexing an
	// array, slice, or map; receiving from a channel; accessing a
	// struct field or interface method; or calling a function.
	//
	// Notably, we don't mark map key or function parameter types,
	// because the user already needs some way to construct values
	// of those types.
	//
	// It's not critical for correctness that this algorithm is
	// perfect. Worst case, we might miss opportunities to inline
	// some function calls in downstream packages.
	switch t.Etype {
	case TPTR32, TPTR64, TARRAY, TSLICE, TCHAN:
		p.markType(t.Elem())

	case TMAP:
		p.markType(t.Val())

	case TSTRUCT:
		for _, f := range t.FieldSlice() {
			if exportname(f.Sym.Name) || f.Embedded != 0 {
				p.markType(f.Type)
			}
		}

	case TFUNC:
		// If t is the type of a function or method, then
		// t.Nname() is its ONAME. Mark its inline body and
		// any recursively called functions for export.
		inlFlood(asNode(t.Nname()))

		for _, f := range t.Results().FieldSlice() {
			p.markType(f.Type)
		}

	case TINTER:
		for _, f := range t.FieldSlice() {
			if exportname(f.Sym.Name) {
				p.markType(f.Type)
			}
		}
	}
}

func (p *exporter) obj(sym *types.Sym) {
	// Exported objects may be from different packages because they
	// may be re-exported via an exported alias or as dependencies in
	// exported inlined function bodies. Thus, exported object names
	// must be fully qualified.
	//
	// (This can only happen for aliased objects or during phase 2
	// (exportInlined enabled) of object export. Unaliased Objects
	// exported in phase 1 (compiler-indendepent objects) are by
	// definition only the objects from the current package and not
	// pulled in via inlined function bodies. In that case the package
	// qualifier is not needed. Possible space optimization.)

	n := asNode(sym.Def)
	switch n.Op {
	case OLITERAL:
		// constant
		// TODO(gri) determine if we need the typecheck call here
		n = typecheck(n, Erv)
		if n == nil || n.Op != OLITERAL {
			Fatalf("exporter: dumpexportconst: oconst nil: %v", sym)
		}

		p.tag(constTag)
		p.pos(n)
		// TODO(gri) In inlined functions, constants are used directly
		// so they should never occur as re-exported objects. We may
		// not need the qualified name here. See also comment above.
		// Possible space optimization.
		p.qualifiedName(sym)
		p.typ(unidealType(n.Type, n.Val()))
		p.value(n.Val())

	case OTYPE:
		// named type
		t := n.Type
		if t.Etype == TFORW {
			Fatalf("exporter: export of incomplete type %v", sym)
		}

		if IsAlias(sym) {
			p.tag(aliasTag)
			p.pos(n)
			p.qualifiedName(sym)
		} else {
			p.tag(typeTag)
		}
		p.typ(t)

	case ONAME:
		// variable or function
		n = typecheck(n, Erv|Ecall)
		if n == nil || n.Type == nil {
			Fatalf("exporter: variable/function exported but not defined: %v", sym)
		}

		if n.Type.Etype == TFUNC && n.Class() == PFUNC {
			// function
			p.tag(funcTag)
			p.pos(n)
			p.qualifiedName(sym)

			sig := asNode(sym.Def).Type
			inlineable := isInlineable(asNode(sym.Def))

			p.paramList(sig.Params(), inlineable)
			p.paramList(sig.Results(), inlineable)

			var f *Func
			if inlineable && asNode(sym.Def).Func.ExportInline() {
				f = asNode(sym.Def).Func
				// TODO(gri) re-examine reexportdeplist:
				// Because we can trivially export types
				// in-place, we don't need to collect types
				// inside function bodies in the exportlist.
				// With an adjusted reexportdeplist used only
				// by the binary exporter, we can also avoid
				// the global exportlist.
				reexportdeplist(f.Inl)
			}
			p.funcList = append(p.funcList, f)
		} else {
			// variable
			p.tag(varTag)
			p.pos(n)
			p.qualifiedName(sym)
			p.typ(asNode(sym.Def).Type)
		}

	default:
		Fatalf("exporter: unexpected export symbol: %v %v", n.Op, sym)
	}
}

// deltaNewFile is a magic line delta offset indicating a new file.
// We use -64 because it is rare; see issue 20080 and CL 41619.
// -64 is the smallest int that fits in a single byte as a varint.
const deltaNewFile = -64

func (p *exporter) pos(n *Node) {
	if !p.posInfoFormat {
		return
	}

	file, line := fileLine(n)
	if file == p.prevFile {
		// common case: write line delta
		// delta == deltaNewFile means different file
		// if the actual line delta is deltaNewFile,
		// follow up with a negative int to indicate that.
		// only non-negative ints can follow deltaNewFile
		// when writing a new file.
		delta := line - p.prevLine
		p.int(delta)
		if delta == deltaNewFile {
			p.int(-1) // -1 means no file change
		}
	} else {
		// different file
		p.int(deltaNewFile)
		p.int(line) // line >= 0
		p.path(file)
		p.prevFile = file
	}
	p.prevLine = line
}

func (p *exporter) path(s string) {
	if i, ok := p.pathIndex[s]; ok {
		// Note: Using p.index(i) here requires the use of p.tag(-len(c)) below
		//       to get matching debug markers ('t'). But in trace mode p.tag
		//       assumes that the tag argument is a valid tag that can be looked
		//       up in the tagString list, rather then some arbitrary slice length.
		//       Use p.int instead.
		p.int(i) // i >= 0
		return
	}
	p.pathIndex[s] = len(p.pathIndex)
	c := strings.Split(s, "/")
	p.int(-len(c)) // -len(c) < 0
	for _, x := range c {
		p.string(x)
	}
}

func fileLine(n *Node) (file string, line int) {
	if n != nil {
		pos := Ctxt.PosTable.Pos(n.Pos)
		file = pos.Base().AbsFilename()
		line = int(pos.RelLine())
	}
	return
}

func isInlineable(n *Node) bool {
	if exportInlined && n != nil && n.Func != nil {
		// When lazily typechecking inlined bodies, some
		// re-exported ones may not have been typechecked yet.
		// Currently that can leave unresolved ONONAMEs in
		// import-dot-ed packages in the wrong package.
		//
		// TODO(mdempsky): Having the ExportInline check here
		// instead of the outer if statement means we end up
		// exporting parameter names even for functions whose
		// inline body won't be exported by this package. This
		// is currently necessary because we might first
		// import a function/method from a package where it
		// doesn't need to be re-exported, and then from a
		// package where it does. If this happens, we'll need
		// the parameter names.
		//
		// We could initially do without the parameter names,
		// and then fill them in when importing the inline
		// body. But parameter names are attached to the
		// function type, and modifying types after the fact
		// is a little sketchy.
		if Debug_typecheckinl == 0 && n.Func.ExportInline() {
			typecheckinl(n)
		}
		return true
	}
	return false
}

func (p *exporter) typ(t *types.Type) {
	if t == nil {
		Fatalf("exporter: nil type")
	}

	// Possible optimization: Anonymous pointer types *T where
	// T is a named type are common. We could canonicalize all
	// such types *T to a single type PT = *T. This would lead
	// to at most one *T entry in typIndex, and all future *T's
	// would be encoded as the respective index directly. Would
	// save 1 byte (pointerTag) per *T and reduce the typIndex
	// size (at the cost of a canonicalization map). We can do
	// this later, without encoding format change.

	// if we saw the type before, write its index (>= 0)
	if i, ok := p.typIndex[t]; ok {
		p.index('T', i)
		return
	}

	// otherwise, remember the type, write the type tag (< 0) and type data
	if trackAllTypes {
		if p.trace {
			p.tracef("T%d = {>\n", len(p.typIndex))
			defer p.tracef("<\n} ")
		}
		p.typIndex[t] = len(p.typIndex)
	}

	// pick off named types
	if tsym := t.Sym; tsym != nil {
		if !trackAllTypes {
			// if we don't track all types, track named types now
			p.typIndex[t] = len(p.typIndex)
		}

		// Predeclared types should have been found in the type map.
		if t.Orig == t {
			Fatalf("exporter: predeclared type missing from type map?")
		}

		n := typenod(t)
		if n.Type != t {
			Fatalf("exporter: named type definition incorrectly set up")
		}

		p.tag(namedTag)
		p.pos(n)
		p.qualifiedName(tsym)

		// write underlying type
		p.typ(t.Orig)

		// interfaces don't have associated methods
		if t.Orig.IsInterface() {
			return
		}

		// sort methods for reproducible export format
		// TODO(gri) Determine if they are already sorted
		// in which case we can drop this step.
		var methods []*types.Field
		methods = append(methods, t.Methods().Slice()...)
		sort.Sort(methodbyname(methods))
		p.int(len(methods))

		if p.trace && len(methods) > 0 {
			p.tracef("associated methods {>")
		}

		for _, m := range methods {
			if p.trace {
				p.tracef("\n")
			}
			if strings.Contains(m.Sym.Name, ".") {
				Fatalf("invalid symbol name: %s (%v)", m.Sym.Name, m.Sym)
			}

			p.pos(asNode(m.Nname))
			p.fieldSym(m.Sym, false)

			sig := m.Type
			mfn := asNode(sig.FuncType().Nname)
			inlineable := isInlineable(mfn)

			p.paramList(sig.Recvs(), inlineable)
			p.paramList(sig.Params(), inlineable)
			p.paramList(sig.Results(), inlineable)
			p.bool(m.Nointerface()) // record go:nointerface pragma value (see also #16243)

			var f *Func
			if inlineable && mfn.Func.ExportInline() {
				f = mfn.Func
				reexportdeplist(mfn.Func.Inl)
			}
			p.funcList = append(p.funcList, f)
		}

		if p.trace && len(methods) > 0 {
			p.tracef("<\n} ")
		}

		return
	}

	// otherwise we have a type literal
	switch t.Etype {
	case TARRAY:
		if t.IsDDDArray() {
			Fatalf("array bounds should be known at export time: %v", t)
		}
		p.tag(arrayTag)
		p.int64(t.NumElem())
		p.typ(t.Elem())

	case TSLICE:
		p.tag(sliceTag)
		p.typ(t.Elem())

	case TDDDFIELD:
		// see p.param use of TDDDFIELD
		p.tag(dddTag)
		p.typ(t.DDDField())

	case TSTRUCT:
		p.tag(structTag)
		p.fieldList(t)

	case TPTR32, TPTR64: // could use Tptr but these are constants
		p.tag(pointerTag)
		p.typ(t.Elem())

	case TFUNC:
		p.tag(signatureTag)
		p.paramList(t.Params(), false)
		p.paramList(t.Results(), false)

	case TINTER:
		p.tag(interfaceTag)
		p.methodList(t)

	case TMAP:
		p.tag(mapTag)
		p.typ(t.Key())
		p.typ(t.Val())

	case TCHAN:
		p.tag(chanTag)
		p.int(int(t.ChanDir()))
		p.typ(t.Elem())

	default:
		Fatalf("exporter: unexpected type: %v (Etype = %d)", t, t.Etype)
	}
}

func (p *exporter) qualifiedName(sym *types.Sym) {
	p.string(sym.Name)
	p.pkg(sym.Pkg)
}

func (p *exporter) fieldList(t *types.Type) {
	if p.trace && t.NumFields() > 0 {
		p.tracef("fields {>")
		defer p.tracef("<\n} ")
	}

	p.int(t.NumFields())
	for _, f := range t.Fields().Slice() {
		if p.trace {
			p.tracef("\n")
		}
		p.field(f)
	}
}

func (p *exporter) field(f *types.Field) {
	p.pos(asNode(f.Nname))
	p.fieldName(f)
	p.typ(f.Type)
	p.string(f.Note)
}

func (p *exporter) methodList(t *types.Type) {
	var embeddeds, methods []*types.Field

	for _, m := range t.Methods().Slice() {
		if m.Sym != nil {
			methods = append(methods, m)
		} else {
			embeddeds = append(embeddeds, m)
		}
	}

	if p.trace && len(embeddeds) > 0 {
		p.tracef("embeddeds {>")
	}
	p.int(len(embeddeds))
	for _, m := range embeddeds {
		if p.trace {
			p.tracef("\n")
		}
		p.pos(asNode(m.Nname))
		p.typ(m.Type)
	}
	if p.trace && len(embeddeds) > 0 {
		p.tracef("<\n} ")
	}

	if p.trace && len(methods) > 0 {
		p.tracef("methods {>")
	}
	p.int(len(methods))
	for _, m := range methods {
		if p.trace {
			p.tracef("\n")
		}
		p.method(m)
	}
	if p.trace && len(methods) > 0 {
		p.tracef("<\n} ")
	}
}

func (p *exporter) method(m *types.Field) {
	p.pos(asNode(m.Nname))
	p.methodName(m.Sym)
	p.paramList(m.Type.Params(), false)
	p.paramList(m.Type.Results(), false)
}

func (p *exporter) fieldName(t *types.Field) {
	name := t.Sym.Name
	if t.Embedded != 0 {
		// anonymous field - we distinguish between 3 cases:
		// 1) field name matches base type name and is exported
		// 2) field name matches base type name and is not exported
		// 3) field name doesn't match base type name (alias name)
		bname := basetypeName(t.Type)
		if name == bname {
			if exportname(name) {
				name = "" // 1) we don't need to know the field name or package
			} else {
				name = "?" // 2) use unexported name "?" to force package export
			}
		} else {
			// 3) indicate alias and export name as is
			// (this requires an extra "@" but this is a rare case)
			p.string("@")
		}
	}
	p.string(name)
	if name != "" && !exportname(name) {
		p.pkg(t.Sym.Pkg)
	}
}

// methodName is like qualifiedName but it doesn't record the package for exported names.
func (p *exporter) methodName(sym *types.Sym) {
	p.string(sym.Name)
	if !exportname(sym.Name) {
		p.pkg(sym.Pkg)
	}
}

func basetypeName(t *types.Type) string {
	s := t.Sym
	if s == nil && t.IsPtr() {
		s = t.Elem().Sym // deref
	}
	if s != nil {
		return s.Name
	}
	return "" // unnamed type
}

func (p *exporter) paramList(params *types.Type, numbered bool) {
	if !params.IsFuncArgStruct() {
		Fatalf("exporter: parameter list expected")
	}

	// use negative length to indicate unnamed parameters
	// (look at the first parameter only since either all
	// names are present or all are absent)
	//
	// TODO(gri) If we don't have an exported function
	// body, the parameter names are irrelevant for the
	// compiler (though they may be of use for other tools).
	// Possible space optimization.
	n := params.NumFields()
	if n > 0 && parName(params.Field(0), numbered) == "" {
		n = -n
	}
	p.int(n)
	for _, q := range params.Fields().Slice() {
		p.param(q, n, numbered)
	}
}

func (p *exporter) param(q *types.Field, n int, numbered bool) {
	t := q.Type
	if q.Isddd() {
		// create a fake type to encode ... just for the p.typ call
		t = types.NewDDDField(t.Elem())
	}
	p.typ(t)
	if n > 0 {
		name := parName(q, numbered)
		if name == "" {
			// Sometimes we see an empty name even for n > 0.
			// This appears to happen for interface methods
			// with _ (blank) parameter names. Make sure we
			// have a proper name and package so we don't crash
			// during import (see also issue #15470).
			// (parName uses "" instead of "?" as in fmt.go)
			// TODO(gri) review parameter name encoding
			name = "_"
		}
		p.string(name)
		if name != "_" {
			// Because of (re-)exported inlined functions
			// the importpkg may not be the package to which this
			// function (and thus its parameter) belongs. We need to
			// supply the parameter package here. We need the package
			// when the function is inlined so we can properly resolve
			// the name. The _ (blank) parameter cannot be accessed, so
			// we don't need to export a package.
			//
			// TODO(gri) This is compiler-specific. Try using importpkg
			// here and then update the symbols if we find an inlined
			// body only. Otherwise, the parameter name is ignored and
			// the package doesn't matter. This would remove an int
			// (likely 1 byte) for each named parameter.
			p.pkg(q.Sym.Pkg)
		}
	}
	// TODO(gri) This is compiler-specific (escape info).
	// Move into compiler-specific section eventually?
	// (Not having escape info causes tests to fail, e.g. runtime GCInfoTest)
	p.string(q.Note)
}

func parName(f *types.Field, numbered bool) string {
	s := f.Sym
	if s == nil {
		return ""
	}

	// Take the name from the original, lest we substituted it with ~r%d or ~b%d.
	// ~r%d is a (formerly) unnamed result.
	if asNode(f.Nname) != nil {
		if asNode(f.Nname).Orig == nil {
			return "" // s = nil
		}
		s = asNode(f.Nname).Orig.Sym
		if s != nil && s.Name[0] == '~' {
			if s.Name[1] == 'r' { // originally an unnamed result
				return "" // s = nil
			} else if s.Name[1] == 'b' { // originally the blank identifier _
				return "_" // belongs to localpkg
			}
		}
	}

	if s == nil {
		return ""
	}

	// print symbol with Vargen number or not as desired
	name := s.Name
	if strings.Contains(name, ".") {
		Fatalf("invalid symbol name: %s", name)
	}

	// Functions that can be inlined use numbered parameters so we can distinguish them
	// from other names in their context after inlining (i.e., the parameter numbering
	// is a form of parameter rewriting). See issue 4326 for an example and test case.
	if numbered {
		if !strings.Contains(name, "·") && asNode(f.Nname) != nil && asNode(f.Nname).Name != nil && asNode(f.Nname).Name.Vargen > 0 {
			name = fmt.Sprintf("%s·%d", name, asNode(f.Nname).Name.Vargen) // append Vargen
		}
	} else {
		if i := strings.Index(name, "·"); i > 0 {
			name = name[:i] // cut off Vargen
		}
	}
	return name
}

func (p *exporter) value(x Val) {
	if p.trace {
		p.tracef("= ")
	}

	switch x := x.U.(type) {
	case bool:
		tag := falseTag
		if x {
			tag = trueTag
		}
		p.tag(tag)

	case *Mpint:
		if minintval[TINT64].Cmp(x) <= 0 && x.Cmp(maxintval[TINT64]) <= 0 {
			// common case: x fits into an int64 - use compact encoding
			p.tag(int64Tag)
			p.int64(x.Int64())
			return
		}
		// uncommon case: large x - use float encoding
		// (powers of 2 will be encoded efficiently with exponent)
		f := newMpflt()
		f.SetInt(x)
		p.tag(floatTag)
		p.float(f)

	case *Mpflt:
		p.tag(floatTag)
		p.float(x)

	case *Mpcplx:
		p.tag(complexTag)
		p.float(&x.Real)
		p.float(&x.Imag)

	case string:
		p.tag(stringTag)
		p.string(x)

	case *NilVal:
		// not a constant but used in exported function bodies
		p.tag(nilTag)

	default:
		Fatalf("exporter: unexpected value %v (%T)", x, x)
	}
}

func (p *exporter) float(x *Mpflt) {
	// extract sign (there is no -0)
	f := &x.Val
	sign := f.Sign()
	if sign == 0 {
		// x == 0
		p.int(0)
		return
	}
	// x != 0

	// extract exponent such that 0.5 <= m < 1.0
	var m big.Float
	exp := f.MantExp(&m)

	// extract mantissa as *big.Int
	// - set exponent large enough so mant satisfies mant.IsInt()
	// - get *big.Int from mant
	m.SetMantExp(&m, int(m.MinPrec()))
	mant, acc := m.Int(nil)
	if acc != big.Exact {
		Fatalf("exporter: internal error")
	}

	p.int(sign)
	p.int(exp)
	p.string(string(mant.Bytes()))
}

// ----------------------------------------------------------------------------
// Inlined function bodies

// Approach: More or less closely follow what fmt.go is doing for FExp mode
// but instead of emitting the information textually, emit the node tree in
// binary form.

// TODO(gri) Improve tracing output. The current format is difficult to read.

// stmtList may emit more (or fewer) than len(list) nodes.
func (p *exporter) stmtList(list Nodes) {
	if p.trace {
		if list.Len() == 0 {
			p.tracef("{}")
		} else {
			p.tracef("{>")
			defer p.tracef("<\n}")
		}
	}

	for _, n := range list.Slice() {
		if p.trace {
			p.tracef("\n")
		}
		// TODO inlining produces expressions with ninits. we can't export these yet.
		// (from fmt.go:1461ff)
		if opprec[n.Op] < 0 {
			p.stmt(n)
		} else {
			p.expr(n)
		}
	}

	p.op(OEND)
}

func (p *exporter) exprList(list Nodes) {
	if p.trace {
		if list.Len() == 0 {
			p.tracef("{}")
		} else {
			p.tracef("{>")
			defer p.tracef("<\n}")
		}
	}

	for _, n := range list.Slice() {
		if p.trace {
			p.tracef("\n")
		}
		p.expr(n)
	}

	p.op(OEND)
}

func (p *exporter) elemList(list Nodes) {
	if p.trace {
		p.tracef("[ ")
	}
	p.int(list.Len())
	if p.trace {
		if list.Len() == 0 {
			p.tracef("] {}")
		} else {
			p.tracef("] {>")
			defer p.tracef("<\n}")
		}
	}

	for _, n := range list.Slice() {
		if p.trace {
			p.tracef("\n")
		}
		p.fieldSym(n.Sym, false)
		p.expr(n.Left)
	}
}

func (p *exporter) expr(n *Node) {
	if p.trace {
		p.tracef("( ")
		defer p.tracef(") ")
	}

	// from nodefmt (fmt.go)
	//
	// nodefmt reverts nodes back to their original - we don't need to do
	// it because we are not bound to produce valid Go syntax when exporting
	//
	// if (fmtmode != FExp || n.Op != OLITERAL) && n.Orig != nil {
	// 	n = n.Orig
	// }

	// from exprfmt (fmt.go)
	for n != nil && n.Implicit() && (n.Op == OIND || n.Op == OADDR) {
		n = n.Left
	}

	switch op := n.Op; op {
	// expressions
	// (somewhat closely following the structure of exprfmt in fmt.go)
	case OPAREN:
		p.expr(n.Left) // unparen

	// case ODDDARG:
	//	unimplemented - handled by default case

	case OLITERAL:
		if n.Val().Ctype() == CTNIL && n.Orig != nil && n.Orig != n {
			p.expr(n.Orig)
			break
		}
		p.op(OLITERAL)
		p.pos(n)
		p.typ(unidealType(n.Type, n.Val()))
		p.value(n.Val())

	case ONAME:
		// Special case: explicit name of func (*T) method(...) is turned into pkg.(*T).method,
		// but for export, this should be rendered as (*pkg.T).meth.
		// These nodes have the special property that they are names with a left OTYPE and a right ONAME.
		if n.isMethodExpression() {
			p.op(OXDOT)
			p.pos(n)
			p.expr(n.Left) // n.Left.Op == OTYPE
			p.fieldSym(n.Right.Sym, true)
			break
		}

		p.op(ONAME)
		p.pos(n)
		p.sym(n)

	// case OPACK, ONONAME:
	// 	should have been resolved by typechecking - handled by default case

	case OTYPE:
		p.op(OTYPE)
		p.pos(n)
		p.typ(n.Type)

	// case OTARRAY, OTMAP, OTCHAN, OTSTRUCT, OTINTER, OTFUNC:
	// 	should have been resolved by typechecking - handled by default case

	// case OCLOSURE:
	//	unimplemented - handled by default case

	// case OCOMPLIT:
	// 	should have been resolved by typechecking - handled by default case

	case OPTRLIT:
		p.op(OPTRLIT)
		p.pos(n)
		p.expr(n.Left)
		p.bool(n.Implicit())

	case OSTRUCTLIT:
		p.op(OSTRUCTLIT)
		p.pos(n)
		p.typ(n.Type)
		p.elemList(n.List) // special handling of field names

	case OARRAYLIT, OSLICELIT, OMAPLIT:
		p.op(OCOMPLIT)
		p.pos(n)
		p.typ(n.Type)
		p.exprList(n.List)

	case OKEY:
		p.op(OKEY)
		p.pos(n)
		p.exprsOrNil(n.Left, n.Right)

	// case OSTRUCTKEY:
	//	unreachable - handled in case OSTRUCTLIT by elemList

	// case OCALLPART:
	//	unimplemented - handled by default case

	case OXDOT, ODOT, ODOTPTR, ODOTINTER, ODOTMETH:
		p.op(OXDOT)
		p.pos(n)
		p.expr(n.Left)
		p.fieldSym(n.Sym, true)

	case ODOTTYPE, ODOTTYPE2:
		p.op(ODOTTYPE)
		p.pos(n)
		p.expr(n.Left)
		p.typ(n.Type)

	case OINDEX, OINDEXMAP:
		p.op(OINDEX)
		p.pos(n)
		p.expr(n.Left)
		p.expr(n.Right)

	case OSLICE, OSLICESTR, OSLICEARR:
		p.op(OSLICE)
		p.pos(n)
		p.expr(n.Left)
		low, high, _ := n.SliceBounds()
		p.exprsOrNil(low, high)

	case OSLICE3, OSLICE3ARR:
		p.op(OSLICE3)
		p.pos(n)
		p.expr(n.Left)
		low, high, max := n.SliceBounds()
		p.exprsOrNil(low, high)
		p.expr(max)

	case OCOPY, OCOMPLEX:
		// treated like other builtin calls (see e.g., OREAL)
		p.op(op)
		p.pos(n)
		p.expr(n.Left)
		p.expr(n.Right)
		p.op(OEND)

	case OCONV, OCONVIFACE, OCONVNOP, OARRAYBYTESTR, OARRAYRUNESTR, OSTRARRAYBYTE, OSTRARRAYRUNE, ORUNESTR:
		p.op(OCONV)
		p.pos(n)
		p.expr(n.Left)
		p.typ(n.Type)

	case OREAL, OIMAG, OAPPEND, OCAP, OCLOSE, ODELETE, OLEN, OMAKE, ONEW, OPANIC, ORECOVER, OPRINT, OPRINTN:
		p.op(op)
		p.pos(n)
		if n.Left != nil {
			p.expr(n.Left)
			p.op(OEND)
		} else {
			p.exprList(n.List) // emits terminating OEND
		}
		// only append() calls may contain '...' arguments
		if op == OAPPEND {
			p.bool(n.Isddd())
		} else if n.Isddd() {
			Fatalf("exporter: unexpected '...' with %v call", op)
		}

	case OCALL, OCALLFUNC, OCALLMETH, OCALLINTER, OGETG:
		p.op(OCALL)
		p.pos(n)
		p.expr(n.Left)
		p.exprList(n.List)
		p.bool(n.Isddd())

	case OMAKEMAP, OMAKECHAN, OMAKESLICE:
		p.op(op) // must keep separate from OMAKE for importer
		p.pos(n)
		p.typ(n.Type)
		switch {
		default:
			// empty list
			p.op(OEND)
		case n.List.Len() != 0: // pre-typecheck
			p.exprList(n.List) // emits terminating OEND
		case n.Right != nil:
			p.expr(n.Left)
			p.expr(n.Right)
			p.op(OEND)
		case n.Left != nil && (n.Op == OMAKESLICE || !n.Left.Type.IsUntyped()):
			p.expr(n.Left)
			p.op(OEND)
		}

	// unary expressions
	case OPLUS, OMINUS, OADDR, OCOM, OIND, ONOT, ORECV:
		p.op(op)
		p.pos(n)
		p.expr(n.Left)

	// binary expressions
	case OADD, OAND, OANDAND, OANDNOT, ODIV, OEQ, OGE, OGT, OLE, OLT,
		OLSH, OMOD, OMUL, ONE, OOR, OOROR, ORSH, OSEND, OSUB, OXOR:
		p.op(op)
		p.pos(n)
		p.expr(n.Left)
		p.expr(n.Right)

	case OADDSTR:
		p.op(OADDSTR)
		p.pos(n)
		p.exprList(n.List)

	case OCMPSTR, OCMPIFACE:
		p.op(Op(n.Etype))
		p.pos(n)
		p.expr(n.Left)
		p.expr(n.Right)

	case ODCLCONST:
		// if exporting, DCLCONST should just be removed as its usage
		// has already been replaced with literals
		// TODO(gri) these should not be exported in the first place
		// TODO(gri) why is this considered an expression in fmt.go?
		p.op(ODCLCONST)
		p.pos(n)

	default:
		Fatalf("cannot export %v (%d) node\n"+
			"==> please file an issue and assign to gri@\n", n.Op, int(n.Op))
	}
}

// Caution: stmt will emit more than one node for statement nodes n that have a non-empty
// n.Ninit and where n cannot have a natural init section (such as in "if", "for", etc.).
func (p *exporter) stmt(n *Node) {
	if p.trace {
		p.tracef("( ")
		defer p.tracef(") ")
	}

	if n.Ninit.Len() > 0 && !stmtwithinit(n.Op) {
		if p.trace {
			p.tracef("( /* Ninits */ ")
		}

		// can't use stmtList here since we don't want the final OEND
		for _, n := range n.Ninit.Slice() {
			p.stmt(n)
		}

		if p.trace {
			p.tracef(") ")
		}
	}

	switch op := n.Op; op {
	case ODCL:
		p.op(ODCL)
		p.pos(n)
		p.sym(n.Left)
		p.typ(n.Left.Type)

	// case ODCLFIELD:
	//	unimplemented - handled by default case

	case OAS:
		// Don't export "v = <N>" initializing statements, hope they're always
		// preceded by the DCL which will be re-parsed and typecheck to reproduce
		// the "v = <N>" again.
		if n.Right != nil {
			p.op(OAS)
			p.pos(n)
			p.expr(n.Left)
			p.expr(n.Right)
		}

	case OASOP:
		p.op(OASOP)
		p.pos(n)
		p.int(int(n.Etype))
		p.expr(n.Left)
		if p.bool(!n.Implicit()) {
			p.expr(n.Right)
		}

	case OAS2, OAS2DOTTYPE, OAS2FUNC, OAS2MAPR, OAS2RECV:
		p.op(OAS2)
		p.pos(n)
		p.exprList(n.List)
		p.exprList(n.Rlist)

	case ORETURN:
		p.op(ORETURN)
		p.pos(n)
		p.exprList(n.List)

	// case ORETJMP:
	// 	unreachable - generated by compiler for trampolin routines

	case OPROC, ODEFER:
		p.op(op)
		p.pos(n)
		p.expr(n.Left)

	case OIF:
		p.op(OIF)
		p.pos(n)
		p.stmtList(n.Ninit)
		p.expr(n.Left)
		p.stmtList(n.Nbody)
		p.stmtList(n.Rlist)

	case OFOR:
		p.op(OFOR)
		p.pos(n)
		p.stmtList(n.Ninit)
		p.exprsOrNil(n.Left, n.Right)
		p.stmtList(n.Nbody)

	case ORANGE:
		p.op(ORANGE)
		p.pos(n)
		p.stmtList(n.List)
		p.expr(n.Right)
		p.stmtList(n.Nbody)

	case OSELECT, OSWITCH:
		p.op(op)
		p.pos(n)
		p.stmtList(n.Ninit)
		p.exprsOrNil(n.Left, nil)
		p.stmtList(n.List)

	case OCASE, OXCASE:
		p.op(OXCASE)
		p.pos(n)
		p.stmtList(n.List)
		p.stmtList(n.Nbody)

	case OFALL:
		p.op(OFALL)
		p.pos(n)

	case OBREAK, OCONTINUE:
		p.op(op)
		p.pos(n)
		p.exprsOrNil(n.Left, nil)

	case OEMPTY:
		// nothing to emit

	case OGOTO, OLABEL:
		p.op(op)
		p.pos(n)
		p.expr(n.Left)

	default:
		Fatalf("exporter: CANNOT EXPORT: %v\nPlease notify gri@\n", n.Op)
	}
}

func (p *exporter) exprsOrNil(a, b *Node) {
	ab := 0
	if a != nil {
		ab |= 1
	}
	if b != nil {
		ab |= 2
	}
	p.int(ab)
	if ab&1 != 0 {
		p.expr(a)
	}
	if ab&2 != 0 {
		p.expr(b)
	}
}

func (p *exporter) fieldSym(s *types.Sym, short bool) {
	name := s.Name

	// remove leading "type." in method names ("(T).m" -> "m")
	if short {
		if i := strings.LastIndex(name, "."); i >= 0 {
			name = name[i+1:]
		}
	}

	// we should never see a _ (blank) here - these are accessible ("read") fields
	// TODO(gri) can we assert this with an explicit check?
	p.string(name)
	if !exportname(name) {
		p.pkg(s.Pkg)
	}
}

// sym must encode the _ (blank) identifier as a single string "_" since
// encoding for some nodes is based on this assumption (e.g. ONAME nodes).
func (p *exporter) sym(n *Node) {
	s := n.Sym
	if s.Pkg != nil {
		if len(s.Name) > 0 && s.Name[0] == '.' {
			Fatalf("exporter: exporting synthetic symbol %s", s.Name)
		}
	}

	if p.trace {
		p.tracef("{ SYM ")
		defer p.tracef("} ")
	}

	name := s.Name

	// remove leading "type." in method names ("(T).m" -> "m")
	if i := strings.LastIndex(name, "."); i >= 0 {
		name = name[i+1:]
	}

	if strings.Contains(name, "·") && n.Name.Vargen > 0 {
		Fatalf("exporter: unexpected · in symbol name")
	}

	if i := n.Name.Vargen; i > 0 {
		name = fmt.Sprintf("%s·%d", name, i)
	}

	p.string(name)
	if name != "_" {
		p.pkg(s.Pkg)
	}
	// Fixes issue #18167.
	p.string(s.Linkname)
}

func (p *exporter) bool(b bool) bool {
	if p.trace {
		p.tracef("[")
		defer p.tracef("= %v] ", b)
	}

	x := 0
	if b {
		x = 1
	}
	p.int(x)
	return b
}

func (p *exporter) op(op Op) {
	if p.trace {
		p.tracef("[")
		defer p.tracef("= %v] ", op)
	}

	p.int(int(op))
}

// ----------------------------------------------------------------------------
// Low-level encoders

func (p *exporter) index(marker byte, index int) {
	if index < 0 {
		Fatalf("exporter: invalid index < 0")
	}
	if debugFormat {
		p.marker('t')
	}
	if p.trace {
		p.tracef("%c%d ", marker, index)
	}
	p.rawInt64(int64(index))
}

func (p *exporter) tag(tag int) {
	if tag >= 0 {
		Fatalf("exporter: invalid tag >= 0")
	}
	if debugFormat {
		p.marker('t')
	}
	if p.trace {
		p.tracef("%s ", tagString[-tag])
	}
	p.rawInt64(int64(tag))
}

func (p *exporter) int(x int) {
	p.int64(int64(x))
}

func (p *exporter) int64(x int64) {
	if debugFormat {
		p.marker('i')
	}
	if p.trace {
		p.tracef("%d ", x)
	}
	p.rawInt64(x)
}

func (p *exporter) string(s string) {
	if debugFormat {
		p.marker('s')
	}
	if p.trace {
		p.tracef("%q ", s)
	}
	// if we saw the string before, write its index (>= 0)
	// (the empty string is mapped to 0)
	if i, ok := p.strIndex[s]; ok {
		p.rawInt64(int64(i))
		return
	}
	// otherwise, remember string and write its negative length and bytes
	p.strIndex[s] = len(p.strIndex)
	p.rawInt64(-int64(len(s)))
	for i := 0; i < len(s); i++ {
		p.rawByte(s[i])
	}
}

// marker emits a marker byte and position information which makes
// it easy for a reader to detect if it is "out of sync". Used only
// if debugFormat is set.
func (p *exporter) marker(m byte) {
	p.rawByte(m)
	// Uncomment this for help tracking down the location
	// of an incorrect marker when running in debugFormat.
	// if p.trace {
	// 	p.tracef("#%d ", p.written)
	// }
	p.rawInt64(int64(p.written))
}

// rawInt64 should only be used by low-level encoders.
func (p *exporter) rawInt64(x int64) {
	var tmp [binary.MaxVarintLen64]byte
	n := binary.PutVarint(tmp[:], x)
	for i := 0; i < n; i++ {
		p.rawByte(tmp[i])
	}
}

// rawStringln should only be used to emit the initial version string.
func (p *exporter) rawStringln(s string) {
	for i := 0; i < len(s); i++ {
		p.rawByte(s[i])
	}
	p.rawByte('\n')
}

// rawByte is the bottleneck interface to write to p.out.
// rawByte escapes b as follows (any encoding does that
// hides '$'):
//
//	'$'  => '|' 'S'
//	'|'  => '|' '|'
//
// Necessary so other tools can find the end of the
// export data by searching for "$$".
// rawByte should only be used by low-level encoders.
func (p *exporter) rawByte(b byte) {
	switch b {
	case '$':
		// write '$' as '|' 'S'
		b = 'S'
		fallthrough
	case '|':
		// write '|' as '|' '|'
		p.out.WriteByte('|')
		p.written++
	}
	p.out.WriteByte(b)
	p.written++
}

// tracef is like fmt.Printf but it rewrites the format string
// to take care of indentation.
func (p *exporter) tracef(format string, args ...interface{}) {
	if strings.ContainsAny(format, "<>\n") {
		var buf bytes.Buffer
		for i := 0; i < len(format); i++ {
			// no need to deal with runes
			ch := format[i]
			switch ch {
			case '>':
				p.indent++
				continue
			case '<':
				p.indent--
				continue
			}
			buf.WriteByte(ch)
			if ch == '\n' {
				for j := p.indent; j > 0; j-- {
					buf.WriteString(".  ")
				}
			}
		}
		format = buf.String()
	}
	fmt.Printf(format, args...)
}

// ----------------------------------------------------------------------------
// Export format

// Tags. Must be < 0.
const (
	// Objects
	packageTag = -(iota + 1)
	constTag
	typeTag
	varTag
	funcTag
	endTag

	// Types
	namedTag
	arrayTag
	sliceTag
	dddTag
	structTag
	pointerTag
	signatureTag
	interfaceTag
	mapTag
	chanTag

	// Values
	falseTag
	trueTag
	int64Tag
	floatTag
	fractionTag // not used by gc
	complexTag
	stringTag
	nilTag
	unknownTag // not used by gc (only appears in packages with errors)

	// Type aliases
	aliasTag
)

// Debugging support.
// (tagString is only used when tracing is enabled)
var tagString = [...]string{
	// Objects
	-packageTag: "package",
	-constTag:   "const",
	-typeTag:    "type",
	-varTag:     "var",
	-funcTag:    "func",
	-endTag:     "end",

	// Types
	-namedTag:     "named type",
	-arrayTag:     "array",
	-sliceTag:     "slice",
	-dddTag:       "ddd",
	-structTag:    "struct",
	-pointerTag:   "pointer",
	-signatureTag: "signature",
	-interfaceTag: "interface",
	-mapTag:       "map",
	-chanTag:      "chan",

	// Values
	-falseTag:    "false",
	-trueTag:     "true",
	-int64Tag:    "int64",
	-floatTag:    "float",
	-fractionTag: "fraction",
	-complexTag:  "complex",
	-stringTag:   "string",
	-nilTag:      "nil",
	-unknownTag:  "unknown",

	// Type aliases
	-aliasTag: "alias",
}

// untype returns the "pseudo" untyped type for a Ctype (import/export use only).
// (we can't use an pre-initialized array because we must be sure all types are
// set up)
func untype(ctype Ctype) *types.Type {
	switch ctype {
	case CTINT:
		return types.Idealint
	case CTRUNE:
		return types.Idealrune
	case CTFLT:
		return types.Idealfloat
	case CTCPLX:
		return types.Idealcomplex
	case CTSTR:
		return types.Idealstring
	case CTBOOL:
		return types.Idealbool
	case CTNIL:
		return types.Types[TNIL]
	}
	Fatalf("exporter: unknown Ctype")
	return nil
}

var predecl []*types.Type // initialized lazily

func predeclared() []*types.Type {
	if predecl == nil {
		// initialize lazily to be sure that all
		// elements have been initialized before
		predecl = []*types.Type{
			// basic types
			types.Types[TBOOL],
			types.Types[TINT],
			types.Types[TINT8],
			types.Types[TINT16],
			types.Types[TINT32],
			types.Types[TINT64],
			types.Types[TUINT],
			types.Types[TUINT8],
			types.Types[TUINT16],
			types.Types[TUINT32],
			types.Types[TUINT64],
			types.Types[TUINTPTR],
			types.Types[TFLOAT32],
			types.Types[TFLOAT64],
			types.Types[TCOMPLEX64],
			types.Types[TCOMPLEX128],
			types.Types[TSTRING],

			// basic type aliases
			types.Bytetype,
			types.Runetype,

			// error
			types.Errortype,

			// untyped types
			untype(CTBOOL),
			untype(CTINT),
			untype(CTRUNE),
			untype(CTFLT),
			untype(CTCPLX),
			untype(CTSTR),
			untype(CTNIL),

			// package unsafe
			types.Types[TUNSAFEPTR],

			// invalid type (package contains errors)
			types.Types[Txxx],

			// any type, for builtin export data
			types.Types[TANY],
		}
	}
	return predecl
}