cmd/internal/obj, cmd/link: put all symbol data in one contiguous section

Another object file change, gives a reasonable improvement:

name       old s/op   new s/op   delta
LinkCmdGo  0.46 ± 3%  0.44 ± 9%  -3.34%  (p=0.000 n=98+82)
LinkJuju   4.09 ± 4%  3.92 ± 5%  -4.30%  (p=0.000 n=98+99)

I guess the data section could be mmap-ed instead of read, I haven't tried
that.

Change-Id: I959eee470a05526ab1579e3f5d3ede41c16c954f
Reviewed-on: https://go-review.googlesource.com/20928
Run-TryBot: Michael Hudson-Doyle <michael.hudson@canonical.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Crawshaw <crawshaw@golang.org>

src/cmd/internal/goobj/read.go

@@ -241,6 +241,7 @@ type objReader struct {
 	f          io.ReadSeeker
 	err        error
 	offset     int64
+	dataOffset int64
 	limit      int64
 	tmp        [256]byte
 	pkg        string
@@ -416,8 +417,8 @@ func (r *objReader) readRef() {
 // readData reads a data reference from the input file.
 func (r *objReader) readData() Data {
 	n := r.readInt()
-	d := Data{Offset: r.offset, Size: int64(n)}
-	r.skip(int64(n))
+	d := Data{Offset: r.dataOffset, Size: int64(n)}
+	r.dataOffset += int64(n)
 	return d
 }
 
@@ -610,6 +611,10 @@ func (r *objReader) parseObject(prefix []byte) error {
 		r.readRef()
 	}
 
+	dataLength := r.readInt()
+	r.dataOffset = r.offset
+	r.skip(int64(dataLength))
+
 	// Symbols.
 	for {
 		if b := r.readByte(); b != 0xfe {

src/cmd/internal/obj/objfile.go

@@ -19,8 +19,10 @@
 //	- byte 1 - version number
 //	- sequence of strings giving dependencies (imported packages)
 //	- empty string (marks end of sequence)
-//	- sequence of sybol references used by the defined symbols
+//	- sequence of symbol references used by the defined symbols
 //	- byte 0xff (marks end of sequence)
+//	- integer (length of following data)
+//	- data, the content of the defined symbols
 //	- sequence of defined symbols
 //	- byte 0xff (marks end of sequence)
 //	- magic footer: "\xff\xffgo13ld"
@@ -96,9 +98,6 @@
 //
 // TODO(rsc): The file format is good for a first pass but needs work.
 //	- There are SymID in the object file that should really just be strings.
-//	- The actual symbol memory images are interlaced with the symbol
-//	  metadata. They should be separated, to reduce the I/O required to
-//	  load just the metadata.
 
 package obj
 
@@ -319,15 +318,42 @@ func Writeobjfile(ctxt *Link, b *Biobuf) {
 	}
 	wrstring(b, "")
 
+	var dataLength int64
 	// Emit symbol references.
 	for _, s := range ctxt.Text {
 		writerefs(ctxt, b, s)
+		dataLength += int64(len(s.P))
+
+		pc := s.Pcln
+		dataLength += int64(len(pc.Pcsp.P))
+		dataLength += int64(len(pc.Pcfile.P))
+		dataLength += int64(len(pc.Pcline.P))
+		for i := 0; i < len(pc.Pcdata); i++ {
+			dataLength += int64(len(pc.Pcdata[i].P))
+		}
 	}
 	for _, s := range ctxt.Data {
 		writerefs(ctxt, b, s)
+		dataLength += int64(len(s.P))
 	}
 	Bputc(b, 0xff)
 
+	// Write data block
+	wrint(b, dataLength)
+	for _, s := range ctxt.Text {
+		b.w.Write(s.P)
+		pc := s.Pcln
+		b.w.Write(pc.Pcsp.P)
+		b.w.Write(pc.Pcfile.P)
+		b.w.Write(pc.Pcline.P)
+		for i := 0; i < len(pc.Pcdata); i++ {
+			b.w.Write(pc.Pcdata[i].P)
+		}
+	}
+	for _, s := range ctxt.Data {
+		b.w.Write(s.P)
+	}
+
 	// Emit symbols.
 	for _, s := range ctxt.Text {
 		writesym(ctxt, b, s)
@@ -480,7 +506,7 @@ func writesym(ctxt *Link, b *Biobuf, s *LSym) {
 	wrint(b, flags)
 	wrint(b, s.Size)
 	wrsym(b, s.Gotype)
-	wrdata(b, s.P)
+	wrint(b, int64(len(s.P)))
 
 	wrint(b, int64(len(s.R)))
 	var r *Reloc
@@ -521,12 +547,12 @@ func writesym(ctxt *Link, b *Biobuf, s *LSym) {
 		}
 
 		pc := s.Pcln
-		wrdata(b, pc.Pcsp.P)
-		wrdata(b, pc.Pcfile.P)
-		wrdata(b, pc.Pcline.P)
+		wrint(b, int64(len(pc.Pcsp.P)))
+		wrint(b, int64(len(pc.Pcfile.P)))
+		wrint(b, int64(len(pc.Pcline.P)))
 		wrint(b, int64(len(pc.Pcdata)))
 		for i := 0; i < len(pc.Pcdata); i++ {
-			wrdata(b, pc.Pcdata[i].P)
+			wrint(b, int64(len(pc.Pcdata[i].P)))
 		}
 		wrint(b, int64(len(pc.Funcdataoff)))
 		for i := 0; i < len(pc.Funcdataoff); i++ {
@@ -564,11 +590,6 @@ func wrstring(b *Biobuf, s string) {
 	b.w.WriteString(s)
 }
 
-func wrdata(b *Biobuf, v []byte) {
-	wrint(b, int64(len(v)))
-	b.Write(v)
-}
-
 func wrsym(b *Biobuf, s *LSym) {
 	if s == nil {
 		wrint(b, 0)

src/cmd/link/internal/ld/objfile.go

@@ -23,6 +23,8 @@ package ld
 //	- empty string (marks end of sequence)
 //	- sequence of sybol references used by the defined symbols
 //	- byte 0xff (marks end of sequence)
+//	- integer (length of following data)
+//	- data, the content of the defined symbols
 //	- sequence of defined symbols
 //	- byte 0xff (marks end of sequence)
 //	- magic footer: "\xff\xffgo13ld"
@@ -98,9 +100,6 @@ package ld
 //
 // TODO(rsc): The file format is good for a first pass but needs work.
 //	- There are SymID in the object file that should really just be strings.
-//	- The actual symbol memory images are interlaced with the symbol
-//	  metadata. They should be separated, to reduce the I/O required to
-//	  load just the metadata.
 
 import (
 	"bytes"
@@ -151,6 +150,10 @@ func ldobjfile(ctxt *Link, f *obj.Biobuf, pkg string, length int64, pn string) {
 		readref(ctxt, f, pkg, pn)
 	}
 
+	dataLength := rdint64(f)
+	data := make([]byte, dataLength)
+	obj.Bread(f, data)
+
 	for {
 		c, err := f.Peek(1)
 		if err != nil {
@@ -159,7 +162,7 @@ func ldobjfile(ctxt *Link, f *obj.Biobuf, pkg string, length int64, pn string) {
 		if c[0] == 0xff {
 			break
 		}
-		readsym(ctxt, f, pkg, pn)
+		readsym(ctxt, f, &data, pkg, pn)
 	}
 
 	buf = [8]uint8{}
@@ -173,7 +176,7 @@ func ldobjfile(ctxt *Link, f *obj.Biobuf, pkg string, length int64, pn string) {
 	}
 }
 
-func readsym(ctxt *Link, f *obj.Biobuf, pkg string, pn string) {
+func readsym(ctxt *Link, f *obj.Biobuf, buf *[]byte, pkg string, pn string) {
 	if obj.Bgetc(f) != 0xfe {
 		log.Fatalf("readsym out of sync")
 	}
@@ -184,7 +187,7 @@ func readsym(ctxt *Link, f *obj.Biobuf, pkg string, pn string) {
 	local := flags&2 != 0
 	size := rdint(f)
 	typ := rdsym(ctxt, f, pkg)
-	data := rddata(f)
+	data := rddata(f, buf)
 	nreloc := rdint(f)
 
 	var dup *LSym
@@ -283,14 +286,14 @@ overwrite:
 
 		s.Pcln = new(Pcln)
 		pc := s.Pcln
-		pc.Pcsp.P = rddata(f)
-		pc.Pcfile.P = rddata(f)
-		pc.Pcline.P = rddata(f)
+		pc.Pcsp.P = rddata(f, buf)
+		pc.Pcfile.P = rddata(f, buf)
+		pc.Pcline.P = rddata(f, buf)
 		n = rdint(f)
 		pc.Pcdata = make([]Pcdata, n)
 		pc.Npcdata = n
 		for i := 0; i < n; i++ {
-			pc.Pcdata[i].P = rddata(f)
+			pc.Pcdata[i].P = rddata(f, buf)
 		}
 		n = rdint(f)
 		pc.Funcdata = make([]*LSym, n)
@@ -482,23 +485,10 @@ func rdstring(f *obj.Biobuf) string {
 	return string(rdBuf[:n])
 }
 
-const rddataBufMax = 1 << 14
-
-var rddataBuf = make([]byte, rddataBufMax)
-
-func rddata(f *obj.Biobuf) []byte {
-	var p []byte
+func rddata(f *obj.Biobuf, buf *[]byte) []byte {
 	n := rdint(f)
-	if n > rddataBufMax {
-		p = make([]byte, n)
-	} else {
-		if len(rddataBuf) < n {
-			rddataBuf = make([]byte, rddataBufMax)
-		}
-		p = rddataBuf[:n:n]
-		rddataBuf = rddataBuf[n:]
-	}
-	obj.Bread(f, p)
+	p := (*buf)[:n:n]
+	*buf = (*buf)[n:]
 	return p
 }