1 files changed, 487 insertions, 67 deletions

diff --git a/pdf/pdf.go b/pdf/pdf.go
index 4f1f5e1..1fcdaa4 100644
--- a/pdf/pdf.go
+++ b/pdf/pdf.go
@@ -1,5 +1,5 @@
 //
-// Copyright (c) 2018 - 2020, Přemysl Eric Janouch <p@janouch.name>
+// Copyright (c) 2018 - 2024, Přemysl Eric Janouch <p@janouch.name>
 //
 // Permission to use, copy, modify, and/or distribute this software for any
 // purpose with or without fee is hereby granted.
@@ -18,6 +18,8 @@ package pdf
 
 import (
 	"bytes"
+	"compress/zlib"
+	"encoding/binary"
 	"encoding/hex"
 	"errors"
 	"fmt"
@@ -59,20 +61,22 @@ const (
 	// higher-level objects
 	Array
 	Dict
+	Stream
 	Indirect
 	Reference
 )
 
-// Object is a PDF token/object thingy.  Objects may be composed either from
+// Object is a PDF token/object thingy. Objects may be composed either from
 // one or a sequence of tokens. The PDF Reference doesn't actually speak
-// of tokens.
+// of tokens, though ISO 32000-1:2008 does.
 type Object struct {
 	Kind ObjectKind
 
 	String        string            // Comment/Keyword/Name/String
 	Number        float64           // Bool, Numeric
 	Array         []Object          // Array, Indirect
-	Dict          map[string]Object // Dict, in the future also Stream
+	Dict          map[string]Object // Dict, Stream
+	Stream        []byte            // Stream
 	N, Generation uint              // Indirect, Reference
 }
 
@@ -118,6 +122,13 @@ func NewDict(d map[string]Object) Object {
 	return Object{Kind: Dict, Dict: d}
 }
 
+func NewStream(d map[string]Object, s []byte) Object {
+	if d == nil {
+		d = make(map[string]Object)
+	}
+	return Object{Kind: Stream, Dict: d, Stream: s}
+}
+
 func NewIndirect(o Object, n, generation uint) Object {
 	return Object{Kind: Indirect, N: n, Generation: generation,
 		Array: []Object{o}}
@@ -458,6 +469,10 @@ func (o *Object) Serialize() string {
 			fmt.Fprint(b, " /", k, " ", v.Serialize())
 		}
 		return "<<" + b.String() + " >>"
+	case Stream:
+		d := NewDict(o.Dict)
+		d.Dict["Length"] = NewNumeric(float64(len(o.Stream)))
+		return d.Serialize() + "\nstream\n" + string(o.Stream) + "\nendstream"
 	case Indirect:
 		return fmt.Sprintf("%d %d obj\n%s\nendobj", o.N, o.Generation,
 			o.Array[0].Serialize())
@@ -471,8 +486,9 @@ func (o *Object) Serialize() string {
 // -----------------------------------------------------------------------------
 
 type ref struct {
-	offset     int64 // file offset or N of the next free entry
+	offset     int64 // file offset, or N of the next free entry, or index
 	generation uint  // object generation
+	compressed *uint // PDF 1.5: N of the containing compressed object
 	nonfree    bool  // whether this N is taken (for a good zero value)
 }
 
@@ -497,6 +513,65 @@ type Updater struct {
 	Trailer map[string]Object
 }
 
+// ListIndirect returns the whole cross-reference table as Reference Objects.
+func (u *Updater) ListIndirect() []Object {
+	result := []Object{}
+	for i := 0; i < len(u.xref); i++ {
+		if u.xref[i].nonfree {
+			result = append(result, NewReference(uint(i), u.xref[i].generation))
+		}
+	}
+	return result
+}
+
+func (u *Updater) parseStream(lex *Lexer, stack *[]Object) (Object, error) {
+	lenStack := len(*stack)
+	if lenStack < 1 {
+		return newError("missing stream dictionary")
+	}
+	dict := (*stack)[lenStack-1]
+	if dict.Kind != Dict {
+		return newError("stream not preceded by a dictionary")
+	}
+
+	*stack = (*stack)[:lenStack-1]
+	length, ok := dict.Dict["Length"]
+	if !ok {
+		return newError("missing stream Length")
+	}
+	length, err := u.Dereference(length)
+	if err != nil {
+		return length, err
+	}
+	if !length.IsUint() || length.Number > math.MaxInt {
+		return newError("stream Length not an unsigned integer")
+	}
+
+	// Expect exactly one newline.
+	if nl, err := lex.Next(); err != nil {
+		return nl, err
+	} else if nl.Kind != NL {
+		return newError("stream does not start with a newline")
+	}
+
+	size := int(length.Number)
+	if len(lex.P) < size {
+		return newError("stream is longer than the document")
+	}
+
+	dict.Kind = Stream
+	dict.Stream = lex.P[:size]
+	lex.P = lex.P[size:]
+
+	// Skip any number of trailing newlines or comments.
+	if end, err := u.parse(lex, stack); err != nil {
+		return end, err
+	} else if end.Kind != Keyword || end.String != "endstream" {
+		return newError("improperly terminated stream")
+	}
+	return dict, nil
+}
+
 func (u *Updater) parseIndirect(lex *Lexer, stack *[]Object) (Object, error) {
 	lenStack := len(*stack)
 	if lenStack < 2 {
@@ -590,15 +665,11 @@ func (u *Updater) parse(lex *Lexer, stack *[]Object) (Object, error) {
 		}
 		return NewDict(dict), nil
 	case Keyword:
-		// Appears in the document body, typically needs
-		// to access the cross-reference table.
-		//
-		// TODO(p): Use the xref to read /Length etc. once we
-		// actually need to read such objects; presumably
-		// streams can use the Object.String member.
 		switch token.String {
 		case "stream":
-			return newError("streams are not supported yet")
+			// Appears in the document body,
+			// typically needs to access the cross-reference table.
+			return u.parseStream(lex, stack)
 		case "obj":
 			return u.parseIndirect(lex, stack)
 		case "R":
@@ -610,16 +681,159 @@ func (u *Updater) parse(lex *Lexer, stack *[]Object) (Object, error) {
 	}
 }
 
-func (u *Updater) loadXref(lex *Lexer, loadedEntries map[uint]struct{}) error {
+func (u *Updater) loadXrefEntry(
+	n uint, r ref, loadedEntries map[uint]struct{}) {
+	if _, ok := loadedEntries[n]; ok {
+		return
+	}
+	if lenXref := uint(len(u.xref)); n >= lenXref {
+		u.xref = append(u.xref, make([]ref, n-lenXref+1)...)
+	}
+	loadedEntries[n] = struct{}{}
+
+	u.xref[n] = r
+}
+
+func (u *Updater) loadXrefStream(
+	lex *Lexer, stack []Object, loadedEntries map[uint]struct{}) (
+	Object, error) {
+	var object Object
+	for {
+		var err error
+		if object, err = u.parse(lex, &stack); err != nil {
+			return New(End), fmt.Errorf("invalid xref table: %s", err)
+		} else if object.Kind == End {
+			return newError("invalid xref table")
+		}
+
+		// For the sake of simplicity, keep stacking until we find an object.
+		if object.Kind == Indirect {
+			break
+		}
+
+		stack = append(stack, object)
+	}
+
+	// ISO 32000-2:2020 7.5.8.2 Cross-reference stream dictionary
+	stream := object.Array[0]
+	if stream.Kind != Stream {
+		return newError("invalid xref table")
+	}
+	if typ, ok := stream.Dict["Type"]; !ok ||
+		typ.Kind != Name || typ.String != "XRef" {
+		return newError("invalid xref stream")
+	}
+
+	data, err := u.GetStreamData(stream)
+	if err != nil {
+		return New(End), fmt.Errorf("invalid xref stream: %s", err)
+	}
+
+	size, ok := stream.Dict["Size"]
+	if !ok || !size.IsUint() || size.Number <= 0 {
+		return newError("invalid or missing cross-reference stream Size")
+	}
+
+	type pair struct{ start, count uint }
+	pairs := []pair{}
+	if index, ok := stream.Dict["Index"]; !ok {
+		pairs = append(pairs, pair{0, uint(size.Number)})
+	} else {
+		if index.Kind != Array || len(index.Array)%2 != 0 {
+			return newError("invalid cross-reference stream Index")
+		}
+
+		a := index.Array
+		for i := 0; i < len(a); i += 2 {
+			if !a[i].IsUint() || !a[i+1].IsUint() {
+				return newError("invalid cross-reference stream Index")
+			}
+			pairs = append(pairs, pair{uint(a[i].Number), uint(a[i+1].Number)})
+		}
+	}
+
+	w, ok := stream.Dict["W"]
+	if !ok || w.Kind != Array || len(w.Array) != 3 ||
+		!w.Array[0].IsUint() || !w.Array[1].IsUint() || !w.Array[2].IsUint() {
+		return newError("invalid or missing cross-reference stream W")
+	}
+
+	w1 := uint(w.Array[0].Number)
+	w2 := uint(w.Array[1].Number)
+	w3 := uint(w.Array[2].Number)
+	if w2 == 0 {
+		return newError("invalid cross-reference stream W")
+	}
+
+	unit := w1 + w2 + w3
+	if uint(len(data))%unit != 0 {
+		return newError("invalid cross-reference stream length")
+	}
+
+	readField := func(data []byte, width uint) (uint, []byte) {
+		var n uint
+		for ; width != 0; width-- {
+			n = n<<8 | uint(data[0])
+			data = data[1:]
+		}
+		return n, data
+	}
+
+	// ISO 32000-2:2020 7.5.8.3 Cross-reference stream data
+	for _, pair := range pairs {
+		for i := uint(0); i < pair.count; i++ {
+			if uint(len(data)) < unit {
+				return newError("premature cross-reference stream EOF")
+			}
+
+			var f1, f2, f3 uint = 1, 0, 0
+			if w1 > 0 {
+				f1, data = readField(data, w1)
+			}
+			f2, data = readField(data, w2)
+			if w3 > 0 {
+				f3, data = readField(data, w3)
+			}
+
+			var r ref
+			switch f1 {
+			case 0:
+				r.offset = int64(f2)
+				r.generation = f3
+			case 1:
+				r.offset = int64(f2)
+				r.generation = f3
+				r.nonfree = true
+			case 2:
+				r.offset = int64(f3)
+				r.compressed = &f2
+				r.nonfree = true
+			default:
+				// TODO(p): It should be treated as a reference to
+				// the null object. We can't currently represent that.
+				return newError("unsupported cross-reference stream contents")
+			}
+
+			u.loadXrefEntry(pair.start+i, r, loadedEntries)
+		}
+	}
+
+	stream.Kind = Dict
+	stream.Stream = nil
+	return stream, nil
+}
+
+func (u *Updater) loadXref(lex *Lexer, loadedEntries map[uint]struct{}) (
+	Object, error) {
 	var throwawayStack []Object
-	if keyword, _ := u.parse(lex,
-		&throwawayStack); keyword.Kind != Keyword || keyword.String != "xref" {
-		return errors.New("invalid xref table")
+	if object, _ := u.parse(lex,
+		&throwawayStack); object.Kind != Keyword || object.String != "xref" {
+		return u.loadXrefStream(lex, []Object{object}, loadedEntries)
 	}
 	for {
 		object, _ := u.parse(lex, &throwawayStack)
 		if object.Kind == End {
-			return errors.New("unexpected EOF while looking for the trailer")
+			return newError("unexpected EOF while looking for the trailer")
 		}
 		if object.Kind == Keyword && object.String == "trailer" {
 			break
@@ -627,7 +841,7 @@ func (u *Updater) loadXref(lex *Lexer, loadedEntries map[uint]struct{}) error {
 
 		second, _ := u.parse(lex, &throwawayStack)
 		if !object.IsUint() || !second.IsUint() {
-			return errors.New("invalid xref section header")
+			return newError("invalid xref section header")
 		}
 
 		start, count := uint(object.Number), uint(second.Number)
@@ -639,33 +853,29 @@ func (u *Updater) loadXref(lex *Lexer, loadedEntries map[uint]struct{}) error {
 				off.Number > float64(len(u.Document)) ||
 				!gen.IsInteger() || gen.Number < 0 || gen.Number > 65535 ||
 				key.Kind != Keyword {
-				return errors.New("invalid xref entry")
+				return newError("invalid xref entry")
 			}
 
 			free := true
 			if key.String == "n" {
 				free = false
 			} else if key.String != "f" {
-				return errors.New("invalid xref entry")
-			}
-
-			n := start + i
-			if _, ok := loadedEntries[n]; ok {
-				continue
+				return newError("invalid xref entry")
 			}
-			if lenXref := uint(len(u.xref)); n >= lenXref {
-				u.xref = append(u.xref, make([]ref, n-lenXref+1)...)
-			}
-			loadedEntries[n] = struct{}{}
 
-			u.xref[n] = ref{
+			u.loadXrefEntry(start+i, ref{
 				offset:     int64(off.Number),
 				generation: uint(gen.Number),
 				nonfree:    !free,
-			}
+			}, loadedEntries)
 		}
 	}
-	return nil
+
+	trailer, _ := u.parse(lex, &throwawayStack)
+	if trailer.Kind != Dict {
+		return newError("invalid trailer dictionary")
+	}
+	return trailer, nil
 }
 
 // -----------------------------------------------------------------------------
@@ -695,7 +905,6 @@ func NewUpdater(document []byte) (*Updater, error) {
 	loadedXrefs := make(map[int64]struct{})
 	loadedEntries := make(map[uint]struct{})
 
-	var throwawayStack []Object
 	for {
 		if _, ok := loadedXrefs[xrefOffset]; ok {
 			return nil, errors.New("circular xref offsets")
@@ -705,24 +914,26 @@ func NewUpdater(document []byte) (*Updater, error) {
 		}
 
 		lex := Lexer{u.Document[xrefOffset:]}
-		if err := u.loadXref(&lex, loadedEntries); err != nil {
+		trailer, err := u.loadXref(&lex, loadedEntries)
+		if err != nil {
 			return nil, err
 		}
 
-		trailer, _ := u.parse(&lex, &throwawayStack)
-		if trailer.Kind != Dict {
-			return nil, errors.New("invalid trailer dictionary")
-		}
 		if len(loadedXrefs) == 0 {
 			u.Trailer = trailer.Dict
 		}
 		loadedXrefs[xrefOffset] = struct{}{}
 
+		// TODO(p): Descend into XRefStm here first, if present,
+		// which is also a linked list.
+
+		// We allow for mixed cross-reference tables and streams
+		// within a single Prev list, although this should never occur.
 		prevOffset, ok := trailer.Dict["Prev"]
 		if !ok {
 			break
 		}
-		// FIXME: We don't check for size_t over or underflow.
+		// FIXME: Do not read offsets and sizes as floating point numbers.
 		if !prevOffset.IsInteger() {
 			return nil, errors.New("invalid Prev offset")
 		}
@@ -764,18 +975,115 @@ func (u *Updater) Version(root *Object) int {
 	return 0
 }
 
+func (u *Updater) getFromObjStm(nObjStm, n uint) (Object, error) {
+	if nObjStm == n {
+		return newError("ObjStm recursion")
+	}
+
+	stream, err := u.Get(nObjStm, 0)
+	if err != nil {
+		return stream, err
+	}
+	if stream.Kind != Stream {
+		return newError("invalid ObjStm")
+	}
+	if typ, ok := stream.Dict["Type"]; !ok ||
+		typ.Kind != Name || typ.String != "ObjStm" {
+		return newError("invalid ObjStm")
+	}
+
+	data, err := u.GetStreamData(stream)
+	if err != nil {
+		return New(End), fmt.Errorf("invalid ObjStm: %s", err)
+	}
+	entryN, ok := stream.Dict["N"]
+	if !ok || !entryN.IsUint() || entryN.Number <= 0 {
+		return newError("invalid ObjStm N")
+	}
+	entryFirst, ok := stream.Dict["First"]
+	if !ok || !entryFirst.IsUint() || entryFirst.Number <= 0 {
+		return newError("invalid ObjStm First")
+	}
+
+	// NOTE: This means descending into that stream if n is not found here.
+	// It is meant to be an object reference.
+	if extends, ok := stream.Dict["Extends"]; ok && extends.Kind != Nil {
+		return newError("ObjStm extensions are unsupported")
+	}
+
+	count := uint(entryN.Number)
+	first := uint(entryFirst.Number)
+	if first > uint(len(data)) {
+		return newError("invalid ObjStm First")
+	}
+
+	lex1 := Lexer{data[:first]}
+	data = data[first:]
+
+	type pair struct{ n, offset uint }
+	pairs := []pair{}
+	for i := uint(0); i < count; i++ {
+		var throwawayStack []Object
+		objN, _ := u.parse(&lex1, &throwawayStack)
+		objOffset, _ := u.parse(&lex1, &throwawayStack)
+		if !objN.IsUint() || !objOffset.IsUint() {
+			return newError("invalid ObjStm pairs")
+		}
+		pairs = append(pairs, pair{uint(objN.Number), uint(objOffset.Number)})
+	}
+	for i, pair := range pairs {
+		if pair.offset > uint(len(data)) ||
+			i > 0 && pairs[i-1].offset >= pair.offset {
+			return newError("invalid ObjStm pairs")
+		}
+	}
+
+	for i, pair := range pairs {
+		if pair.n != n {
+			continue
+		}
+
+		if i+1 < len(pairs) {
+			data = data[pair.offset:pairs[i+1].offset]
+		} else {
+			data = data[pair.offset:]
+		}
+
+		lex2 := Lexer{data}
+		var stack []Object
+		for {
+			object, err := u.parse(&lex2, &stack)
+			if err != nil {
+				return object, err
+			} else if object.Kind == End {
+				break
+			} else {
+				stack = append(stack, object)
+			}
+		}
+		if len(stack) == 0 {
+			return newError("empty ObjStm object")
+		}
+		return stack[0], nil
+	}
+	return newError("object not found in ObjStm")
+}
+
 // Get retrieves an object by its number and generation--may return
 // Nil or End with an error.
-//
-// TODO(p): We should fix all uses of this not to eat the error.
 func (u *Updater) Get(n, generation uint) (Object, error) {
 	if n >= u.xrefSize {
 		return New(Nil), nil
 	}
 
 	ref := u.xref[n]
-	if !ref.nonfree || ref.generation != generation ||
-		ref.offset >= int64(len(u.Document)) {
+	if !ref.nonfree || ref.generation != generation {
+		return New(Nil), nil
+	}
+
+	if ref.compressed != nil {
+		return u.getFromObjStm(*ref.compressed, n)
+	} else if ref.offset >= int64(len(u.Document)) {
 		return New(Nil), nil
 	}
 
@@ -796,6 +1104,14 @@ func (u *Updater) Get(n, generation uint) (Object, error) {
 	}
 }
 
+// Derefence dereferences Reference objects, and passes the other kinds through.
+func (u *Updater) Dereference(o Object) (Object, error) {
+	if o.Kind != Reference {
+		return o, nil
+	}
+	return u.Get(o.N, o.Generation)
+}
+
 // Allocate allocates a new object number.
 func (u *Updater) Allocate() uint {
 	n := u.xrefSize
@@ -822,8 +1138,8 @@ type BytesWriter interface {
 	WriteString(s string) (n int, err error)
 }
 
-// Update appends an updated object to the end of the document. The fill
-// callback must write exactly one PDF object.
+// Update appends an updated object to the end of the document.
+// The fill callback must write exactly one PDF object.
 func (u *Updater) Update(n uint, fill func(buf BytesWriter)) {
 	oldRef := u.xref[n]
 	u.updated[n] = struct{}{}
@@ -843,20 +1159,62 @@ func (u *Updater) Update(n uint, fill func(buf BytesWriter)) {
 	u.Document = buf.Bytes()
 }
 
-// FlushUpdates writes an updated cross-reference table and trailer.
-func (u *Updater) FlushUpdates() {
-	updated := make([]uint, 0, len(u.updated))
-	for n := range u.updated {
-		updated = append(updated, n)
+func (u *Updater) flushXRefStm(updated []uint, buf *bytes.Buffer) {
+	// The cross-reference stream has to point to itself.
+	// XXX: We only duplicate Update code here due to how we currently buffer.
+	n := u.Allocate()
+	updated = append(updated, n)
+
+	u.updated[n] = struct{}{}
+	u.xref[n] = ref{
+		offset:     int64(buf.Len() + 1),
+		generation: 0,
+		nonfree:    true,
 	}
-	sort.Slice(updated, func(i, j int) bool {
-		return updated[i] < updated[j]
+
+	index, b := []Object{}, []byte{}
+	write := func(f1 byte, f2, f3 uint64) {
+		b = append(b, f1)
+		b = binary.BigEndian.AppendUint64(b, f2)
+		b = binary.BigEndian.AppendUint64(b, f3)
+	}
+	for i := 0; i < len(updated); {
+		start, stop := updated[i], updated[i]+1
+		for i++; i < len(updated) && updated[i] == stop; i++ {
+			stop++
+		}
+
+		index = append(index,
+			NewNumeric(float64(start)), NewNumeric(float64(stop-start)))
+		for ; start < stop; start++ {
+			ref := u.xref[start]
+			if ref.compressed != nil {
+				write(2, uint64(*ref.compressed), uint64(ref.offset))
+			} else if ref.nonfree {
+				write(1, uint64(ref.offset), uint64(ref.generation))
+			} else {
+				write(0, uint64(ref.offset), uint64(ref.generation))
+			}
+		}
+	}
+
+	u.Trailer["Size"] = NewNumeric(float64(u.xrefSize))
+	u.Trailer["Index"] = NewArray(index)
+	u.Trailer["W"] = NewArray([]Object{
+		NewNumeric(1), NewNumeric(8), NewNumeric(8),
 	})
 
-	buf := bytes.NewBuffer(u.Document)
-	startXref := buf.Len() + 1
-	buf.WriteString("\nxref\n")
+	for _, key := range []string{
+		"Filter", "DecodeParms", "F", "FFilter", "FDecodeParms", "DL"} {
+		delete(u.Trailer, key)
+	}
 
+	stream := NewStream(u.Trailer, b)
+	fmt.Fprintf(buf, "\n%d 0 obj\n%s\nendobj", n, stream.Serialize())
+}
+
+func (u *Updater) flushXRefTable(updated []uint, buf *bytes.Buffer) {
+	buf.WriteString("\nxref\n")
 	for i := 0; i < len(updated); {
 		start, stop := updated[i], updated[i]+1
 		for i++; i < len(updated) && updated[i] == stop; i++ {
@@ -865,8 +1223,9 @@ func (u *Updater) FlushUpdates() {
 
 		fmt.Fprintf(buf, "%d %d\n", start, stop-start)
 		for ; start < stop; start++ {
+			// XXX: We should warn about any object streams here.
 			ref := u.xref[start]
-			if ref.nonfree {
+			if ref.nonfree && ref.compressed == nil {
 				fmt.Fprintf(buf, "%010d %05d n \n", ref.offset, ref.generation)
 			} else {
 				fmt.Fprintf(buf, "%010d %05d f \n", ref.offset, ref.generation)
@@ -883,10 +1242,38 @@ func (u *Updater) FlushUpdates() {
 
 	u.Trailer["Size"] = NewNumeric(float64(u.xrefSize))
 	trailer := NewDict(u.Trailer)
+	fmt.Fprintf(buf, "trailer\n%s", trailer.Serialize())
+}
+
+// FlushUpdates writes an updated cross-reference table and trailer, or stream.
+func (u *Updater) FlushUpdates() {
+	updated := make([]uint, 0, len(u.updated))
+	for n := range u.updated {
+		updated = append(updated, n)
+	}
+	sort.Slice(updated, func(i, j int) bool {
+		return updated[i] < updated[j]
+	})
+
+	// It does not seem to be possible to upgrade a PDF file
+	// from trailer dictionaries to cross-reference streams,
+	// so keep continuity either way.
+	//
+	// (Downgrading from cross-reference streams using XRefStm would not
+	// create a true hybrid-reference file, although it should work.)
+	buf := bytes.NewBuffer(u.Document)
+	startXref := buf.Len() + 1 /* '\n' */
+	if typ, _ := u.Trailer["Type"]; typ.Kind == Name && typ.String == "XRef" {
+		u.flushXRefStm(updated, buf)
+	} else {
+		u.flushXRefTable(updated, buf)
+	}
 
-	fmt.Fprintf(buf, "trailer\n%s\nstartxref\n%d\n%%%%EOF\n",
-		trailer.Serialize(), startXref)
+	fmt.Fprintf(buf, "\nstartxref\n%d\n%%%%EOF\n", startXref)
 	u.Document = buf.Bytes()
+	u.updated = make(map[uint]struct{})
+
+	u.Trailer["Prev"] = NewNumeric(float64(startXref))
 }
 
 // -----------------------------------------------------------------------------
@@ -904,17 +1291,47 @@ func NewDate(ts time.Time) Object {
 	return NewString(string(buf))
 }
 
+// GetStreamData returns the actual data stored in a stream object,
+// applying any filters.
+func (u *Updater) GetStreamData(stream Object) ([]byte, error) {
+	if f, ok := stream.Dict["F"]; ok && f.Kind != Nil {
+		return nil, errors.New("stream data in other files are unsupported")
+	}
+
+	// Support just enough to decode a common cross-reference stream.
+	if filter, ok := stream.Dict["Filter"]; !ok {
+		return stream.Stream, nil
+	} else if filter.Kind != Name || filter.String != "FlateDecode" {
+		return nil, errors.New("unsupported stream Filter")
+	}
+
+	// TODO(p): Support << /Columns N /Predictor 12 >>
+	// which usually appears in files with cross-reference streams.
+	if parms, ok := stream.Dict["DecodeParms"]; ok && parms.Kind != Nil {
+		return nil, errors.New("DecodeParms are not supported")
+	}
+
+	r, err := zlib.NewReader(bytes.NewReader(stream.Stream))
+	if err != nil {
+		return nil, err
+	}
+
+	var b bytes.Buffer
+	_, err = b.ReadFrom(r)
+	return b.Bytes(), err
+}
+
 // GetFirstPage retrieves the first page of the given page (sub)tree reference,
 // or returns a Nil object if unsuccessful.
-func (u *Updater) GetFirstPage(nodeN, nodeGeneration uint) Object {
-	obj, _ := u.Get(nodeN, nodeGeneration)
-	if obj.Kind != Dict {
+func (u *Updater) GetFirstPage(node Object) Object {
+	obj, err := u.Dereference(node)
+	if err != nil || obj.Kind != Dict {
 		return New(Nil)
 	}
 
 	// Out of convenience; these aren't filled normally.
-	obj.N = nodeN
-	obj.Generation = nodeGeneration
+	obj.N = node.N
+	obj.Generation = node.Generation
 
 	if typ, ok := obj.Dict["Type"]; !ok || typ.Kind != Name {
 		return New(Nil)
@@ -934,7 +1351,7 @@ func (u *Updater) GetFirstPage(nodeN, nodeGeneration uint) Object {
 	}
 
 	// XXX: Nothing prevents us from recursing in an evil circular graph.
-	return u.GetFirstPage(kids.Array[0].N, kids.Array[0].Generation)
+	return u.GetFirstPage(kids.Array[0])
 }
 
 // -----------------------------------------------------------------------------
@@ -1128,7 +1545,10 @@ func Sign(document []byte, key crypto.PrivateKey, certs []*x509.Certificate,
 	if !ok || rootRef.Kind != Reference {
 		return nil, errors.New("trailer does not contain a reference to Root")
 	}
-	root, _ := pdf.Get(rootRef.N, rootRef.Generation)
+	root, err := pdf.Dereference(rootRef)
+	if err != nil {
+		return nil, fmt.Errorf("Root dictionary retrieval failed: %s", err)
+	}
 	if root.Kind != Dict {
 		return nil, errors.New("invalid Root dictionary reference")
 	}
@@ -1182,7 +1602,7 @@ func Sign(document []byte, key crypto.PrivateKey, certs []*x509.Certificate,
 	if !ok || pagesRef.Kind != Reference {
 		return nil, errors.New("invalid Pages reference")
 	}
-	page := pdf.GetFirstPage(pagesRef.N, pagesRef.Generation)
+	page := pdf.GetFirstPage(pagesRef)
 	if page.Kind != Dict {
 		return nil, errors.New("invalid or unsupported page tree")
 	}
@@ -1204,7 +1624,7 @@ func Sign(document []byte, key crypto.PrivateKey, certs []*x509.Certificate,
 	})
 
 	// 8.6.1 Interactive Form Dictionary
-	if _, ok := root.Dict["AcroForm"]; ok {
+	if acroform, ok := root.Dict["AcroForm"]; ok && acroform.Kind != Nil {
 		return nil, errors.New("the document already contains forms, " +
 			"they would be overwritten")
 	}