[dev.link] cmd/link: make asmb parallel across all platforms

Numbers for x86:

Asmb                      54.8ms ±15%    12.1ms ±28%   -78.02%  (p=0.000 n=10+10)
TotalTime                  545ms ± 3%     515ms ± 4%    -5.42%  (p=0.000 n=10+10)

Change-Id: I8fc17f7ccdf83fbd70cf7d27b44f41f9db9c9922
Reviewed-on: https://go-review.googlesource.com/c/go/+/224419
Run-TryBot: Jeremy Faller <jeremy@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Than McIntosh <thanm@google.com>
Reviewed-by: Cherry Zhang <cherryyz@google.com>

src/cmd/link/internal/amd64/asm.go

@@ -705,19 +705,20 @@ func asmb(ctxt *ld.Link) {
 
 	for _, sect := range ld.Segtext.Sections[1:] {
 		offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
-		ld.WriteParallel(&wg, ld.Datblk2, ctxt, offset, sect.Vaddr, sect.Length)
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, offset, sect.Vaddr, sect.Length)
 	}
 
 	if ld.Segrodata.Filelen > 0 {
-		ld.WriteParallel(&wg, ld.Datblk2, ctxt, ld.Segrodata.Fileoff, ld.Segrodata.Vaddr, ld.Segrodata.Filelen)
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrodata.Fileoff, ld.Segrodata.Vaddr, ld.Segrodata.Filelen)
 	}
+
 	if ld.Segrelrodata.Filelen > 0 {
-		ld.WriteParallel(&wg, ld.Datblk2, ctxt, ld.Segrelrodata.Fileoff, ld.Segrelrodata.Vaddr, ld.Segrelrodata.Filelen)
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrelrodata.Fileoff, ld.Segrelrodata.Vaddr, ld.Segrelrodata.Filelen)
 	}
 
-	ld.WriteParallel(&wg, ld.Datblk2, ctxt, ld.Segdata.Fileoff, ld.Segdata.Vaddr, ld.Segdata.Filelen)
+	ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segdata.Fileoff, ld.Segdata.Vaddr, ld.Segdata.Filelen)
 
-	ld.WriteParallel(&wg, ld.Dwarfblk2, ctxt, ld.Segdwarf.Fileoff, ld.Segdwarf.Vaddr, ld.Segdwarf.Filelen)
+	ld.WriteParallel(&wg, ld.Dwarfblk, ctxt, ld.Segdwarf.Fileoff, ld.Segdwarf.Vaddr, ld.Segdwarf.Filelen)
 
 	wg.Wait()
 }

src/cmd/link/internal/arm/asm.go

@@ -39,6 +39,7 @@ import (
 	"debug/elf"
 	"fmt"
 	"log"
+	"sync"
 )
 
 // This assembler:
@@ -763,28 +764,28 @@ func asmb(ctxt *ld.Link) {
 		ld.Asmbelfsetup()
 	}
 
+	var wg sync.WaitGroup
 	sect := ld.Segtext.Sections[0]
-	ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-	ld.Codeblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
-	for _, sect = range ld.Segtext.Sections[1:] {
-		ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-		ld.Datblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
+	offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+	ld.WriteParallel(&wg, ld.Codeblk, ctxt, offset, sect.Vaddr, sect.Length)
+
+	for _, sect := range ld.Segtext.Sections[1:] {
+		offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, offset, sect.Vaddr, sect.Length)
 	}
 
 	if ld.Segrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrodata.Vaddr), int64(ld.Segrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrodata.Fileoff, ld.Segrodata.Vaddr, ld.Segrodata.Filelen)
 	}
+
 	if ld.Segrelrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrelrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrelrodata.Vaddr), int64(ld.Segrelrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrelrodata.Fileoff, ld.Segrelrodata.Vaddr, ld.Segrelrodata.Filelen)
 	}
 
-	ctxt.Out.SeekSet(int64(ld.Segdata.Fileoff))
-	ld.Datblk(ctxt, int64(ld.Segdata.Vaddr), int64(ld.Segdata.Filelen))
+	ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segdata.Fileoff, ld.Segdata.Vaddr, ld.Segdata.Filelen)
 
-	ctxt.Out.SeekSet(int64(ld.Segdwarf.Fileoff))
-	ld.Dwarfblk(ctxt, int64(ld.Segdwarf.Vaddr), int64(ld.Segdwarf.Filelen))
+	ld.WriteParallel(&wg, ld.Dwarfblk, ctxt, ld.Segdwarf.Fileoff, ld.Segdwarf.Vaddr, ld.Segdwarf.Filelen)
+	wg.Wait()
 }
 
 func asmb2(ctxt *ld.Link) {

src/cmd/link/internal/arm64/asm.go

@@ -39,6 +39,7 @@ import (
 	"debug/elf"
 	"fmt"
 	"log"
+	"sync"
 )
 
 func gentext(ctxt *ld.Link) {
@@ -824,28 +825,28 @@ func asmb(ctxt *ld.Link) {
 		ld.Asmbelfsetup()
 	}
 
+	var wg sync.WaitGroup
 	sect := ld.Segtext.Sections[0]
-	ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-	ld.Codeblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
-	for _, sect = range ld.Segtext.Sections[1:] {
-		ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-		ld.Datblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
+	offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+	ld.WriteParallel(&wg, ld.Codeblk, ctxt, offset, sect.Vaddr, sect.Length)
+
+	for _, sect := range ld.Segtext.Sections[1:] {
+		offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, offset, sect.Vaddr, sect.Length)
 	}
 
 	if ld.Segrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrodata.Vaddr), int64(ld.Segrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrodata.Fileoff, ld.Segrodata.Vaddr, ld.Segrodata.Filelen)
 	}
+
 	if ld.Segrelrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrelrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrelrodata.Vaddr), int64(ld.Segrelrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrelrodata.Fileoff, ld.Segrelrodata.Vaddr, ld.Segrelrodata.Filelen)
 	}
 
-	ctxt.Out.SeekSet(int64(ld.Segdata.Fileoff))
-	ld.Datblk(ctxt, int64(ld.Segdata.Vaddr), int64(ld.Segdata.Filelen))
+	ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segdata.Fileoff, ld.Segdata.Vaddr, ld.Segdata.Filelen)
 
-	ctxt.Out.SeekSet(int64(ld.Segdwarf.Fileoff))
-	ld.Dwarfblk(ctxt, int64(ld.Segdwarf.Vaddr), int64(ld.Segdwarf.Filelen))
+	ld.WriteParallel(&wg, ld.Dwarfblk, ctxt, ld.Segdwarf.Fileoff, ld.Segdwarf.Vaddr, ld.Segdwarf.Filelen)
+	wg.Wait()
 }
 
 func asmb2(ctxt *ld.Link) {

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

@@ -729,8 +729,8 @@ func dynreloc(ctxt *Link, data *[sym.SXREF][]*sym.Symbol) {
 	}
 }
 
-func Codeblk(ctxt *Link, addr int64, size int64) {
-	CodeblkPad(ctxt, ctxt.Out, addr, size, zeros[:])
+func Codeblk(ctxt *Link, out *OutBuf, addr int64, size int64) {
+	CodeblkPad(ctxt, out, addr, size, zeros[:])
 }
 
 func CodeblkPad(ctxt *Link, out *OutBuf, addr int64, size int64, pad []byte) {
@@ -926,11 +926,7 @@ func WriteParallel(wg *sync.WaitGroup, fn writeFn, ctxt *Link, seek, vaddr, leng
 	}
 }
 
-func Datblk(ctxt *Link, addr int64, size int64) {
-	writeDatblkToOutBuf(ctxt, ctxt.Out, addr, size)
-}
-
-func Datblk2(ctxt *Link, out *OutBuf, addr, size int64) {
+func Datblk(ctxt *Link, out *OutBuf, addr, size int64) {
 	writeDatblkToOutBuf(ctxt, out, addr, size)
 }
 
@@ -1017,7 +1013,7 @@ func writeDatblkToOutBuf(ctxt *Link, out *OutBuf, addr int64, size int64) {
 	ctxt.Logf("\t%.8x|\n", uint(eaddr))
 }
 
-func Dwarfblk2(ctxt *Link, out *OutBuf, addr int64, size int64) {
+func Dwarfblk(ctxt *Link, out *OutBuf, addr int64, size int64) {
 	if *flagA {
 		ctxt.Logf("dwarfblk [%#x,%#x) at offset %#x\n", addr, addr+size, ctxt.Out.Offset())
 	}
@@ -1025,14 +1021,6 @@ func Dwarfblk2(ctxt *Link, out *OutBuf, addr int64, size int64) {
 	writeBlocks(out, ctxt.outSem, dwarfp, addr, size, zeros[:])
 }
 
-func Dwarfblk(ctxt *Link, addr int64, size int64) {
-	if *flagA {
-		ctxt.Logf("dwarfblk [%#x,%#x) at offset %#x\n", addr, addr+size, ctxt.Out.Offset())
-	}
-
-	writeBlock(ctxt.Out, dwarfp, addr, size, zeros[:])
-}
-
 var zeros [512]byte
 
 var (

src/cmd/link/internal/mips/asm.go

@@ -39,6 +39,7 @@ import (
 	"debug/elf"
 	"fmt"
 	"log"
+	"sync"
 )
 
 func gentext(ctxt *ld.Link) {
@@ -168,24 +169,24 @@ func asmb(ctxt *ld.Link) {
 		ld.Asmbelfsetup()
 	}
 
+	var wg sync.WaitGroup
 	sect := ld.Segtext.Sections[0]
-	ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-	ld.Codeblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
+	offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+	ld.WriteParallel(&wg, ld.Codeblk, ctxt, offset, sect.Vaddr, sect.Length)
+
 	for _, sect = range ld.Segtext.Sections[1:] {
-		ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-		ld.Datblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
+		offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, offset, sect.Vaddr, sect.Length)
 	}
 
 	if ld.Segrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrodata.Vaddr), int64(ld.Segrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrodata.Fileoff, ld.Segrodata.Vaddr, ld.Segrodata.Filelen)
 	}
 
-	ctxt.Out.SeekSet(int64(ld.Segdata.Fileoff))
-	ld.Datblk(ctxt, int64(ld.Segdata.Vaddr), int64(ld.Segdata.Filelen))
+	ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segdata.Fileoff, ld.Segdata.Vaddr, ld.Segdata.Filelen)
 
-	ctxt.Out.SeekSet(int64(ld.Segdwarf.Fileoff))
-	ld.Dwarfblk(ctxt, int64(ld.Segdwarf.Vaddr), int64(ld.Segdwarf.Filelen))
+	ld.WriteParallel(&wg, ld.Dwarfblk, ctxt, ld.Segdwarf.Fileoff, ld.Segdwarf.Vaddr, ld.Segdwarf.Filelen)
+	wg.Wait()
 }
 
 func asmb2(ctxt *ld.Link) {

src/cmd/link/internal/mips64/asm.go

@@ -39,6 +39,7 @@ import (
 	"debug/elf"
 	"fmt"
 	"log"
+	"sync"
 )
 
 func gentext(ctxt *ld.Link) {}
@@ -174,28 +175,28 @@ func asmb(ctxt *ld.Link) {
 		ld.Asmbelfsetup()
 	}
 
+	var wg sync.WaitGroup
 	sect := ld.Segtext.Sections[0]
-	ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-	ld.Codeblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
-	for _, sect = range ld.Segtext.Sections[1:] {
-		ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-		ld.Datblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
+	offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+	ld.WriteParallel(&wg, ld.Codeblk, ctxt, offset, sect.Vaddr, sect.Length)
+
+	for _, sect := range ld.Segtext.Sections[1:] {
+		offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, offset, sect.Vaddr, sect.Length)
 	}
 
 	if ld.Segrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrodata.Vaddr), int64(ld.Segrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrodata.Fileoff, ld.Segrodata.Vaddr, ld.Segrodata.Filelen)
 	}
+
 	if ld.Segrelrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrelrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrelrodata.Vaddr), int64(ld.Segrelrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrelrodata.Fileoff, ld.Segrelrodata.Vaddr, ld.Segrelrodata.Filelen)
 	}
 
-	ctxt.Out.SeekSet(int64(ld.Segdata.Fileoff))
-	ld.Datblk(ctxt, int64(ld.Segdata.Vaddr), int64(ld.Segdata.Filelen))
+	ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segdata.Fileoff, ld.Segdata.Vaddr, ld.Segdata.Filelen)
 
-	ctxt.Out.SeekSet(int64(ld.Segdwarf.Fileoff))
-	ld.Dwarfblk(ctxt, int64(ld.Segdwarf.Vaddr), int64(ld.Segdwarf.Filelen))
+	ld.WriteParallel(&wg, ld.Dwarfblk, ctxt, ld.Segdwarf.Fileoff, ld.Segdwarf.Vaddr, ld.Segdwarf.Filelen)
+	wg.Wait()
 }
 
 func asmb2(ctxt *ld.Link) {

src/cmd/link/internal/ppc64/asm.go

@@ -41,6 +41,7 @@ import (
 	"fmt"
 	"log"
 	"strings"
+	"sync"
 )
 
 func genplt(ctxt *ld.Link) {
@@ -1065,30 +1066,34 @@ func asmb(ctxt *ld.Link) {
 		ld.Asmbelfsetup()
 	}
 
+	var wg sync.WaitGroup
 	for _, sect := range ld.Segtext.Sections {
-		ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
+		offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
 		// Handle additional text sections with Codeblk
 		if sect.Name == ".text" {
-			ld.Codeblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
+			ld.WriteParallel(&wg, ld.Codeblk, ctxt, offset, sect.Vaddr, sect.Length)
 		} else {
-			ld.Datblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
+			ld.WriteParallel(&wg, ld.Datblk, ctxt, offset, sect.Vaddr, sect.Length)
 		}
 	}
 
+	for _, sect := range ld.Segtext.Sections[1:] {
+		offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, offset, sect.Vaddr, sect.Length)
+	}
+
 	if ld.Segrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrodata.Vaddr), int64(ld.Segrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrodata.Fileoff, ld.Segrodata.Vaddr, ld.Segrodata.Filelen)
 	}
+
 	if ld.Segrelrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrelrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrelrodata.Vaddr), int64(ld.Segrelrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrelrodata.Fileoff, ld.Segrelrodata.Vaddr, ld.Segrelrodata.Filelen)
 	}
 
-	ctxt.Out.SeekSet(int64(ld.Segdata.Fileoff))
-	ld.Datblk(ctxt, int64(ld.Segdata.Vaddr), int64(ld.Segdata.Filelen))
+	ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segdata.Fileoff, ld.Segdata.Vaddr, ld.Segdata.Filelen)
 
-	ctxt.Out.SeekSet(int64(ld.Segdwarf.Fileoff))
-	ld.Dwarfblk(ctxt, int64(ld.Segdwarf.Vaddr), int64(ld.Segdwarf.Filelen))
+	ld.WriteParallel(&wg, ld.Dwarfblk, ctxt, ld.Segdwarf.Fileoff, ld.Segdwarf.Vaddr, ld.Segdwarf.Filelen)
+	wg.Wait()
 }
 
 func asmb2(ctxt *ld.Link) {

src/cmd/link/internal/riscv64/asm.go

@@ -13,6 +13,7 @@ import (
 	"cmd/link/internal/sym"
 	"fmt"
 	"log"
+	"sync"
 )
 
 func gentext(ctxt *ld.Link) {
@@ -102,28 +103,28 @@ func asmb(ctxt *ld.Link) {
 		ld.Asmbelfsetup()
 	}
 
+	var wg sync.WaitGroup
 	sect := ld.Segtext.Sections[0]
-	ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-	ld.Codeblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
-	for _, sect = range ld.Segtext.Sections[1:] {
-		ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-		ld.Datblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
+	offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+	ld.WriteParallel(&wg, ld.Codeblk, ctxt, offset, sect.Vaddr, sect.Length)
+
+	for _, sect := range ld.Segtext.Sections[1:] {
+		offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, offset, sect.Vaddr, sect.Length)
 	}
 
 	if ld.Segrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrodata.Vaddr), int64(ld.Segrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrodata.Fileoff, ld.Segrodata.Vaddr, ld.Segrodata.Filelen)
 	}
+
 	if ld.Segrelrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrelrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrelrodata.Vaddr), int64(ld.Segrelrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrelrodata.Fileoff, ld.Segrelrodata.Vaddr, ld.Segrelrodata.Filelen)
 	}
 
-	ctxt.Out.SeekSet(int64(ld.Segdata.Fileoff))
-	ld.Datblk(ctxt, int64(ld.Segdata.Vaddr), int64(ld.Segdata.Filelen))
+	ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segdata.Fileoff, ld.Segdata.Vaddr, ld.Segdata.Filelen)
 
-	ctxt.Out.SeekSet(int64(ld.Segdwarf.Fileoff))
-	ld.Dwarfblk(ctxt, int64(ld.Segdwarf.Vaddr), int64(ld.Segdwarf.Filelen))
+	ld.WriteParallel(&wg, ld.Dwarfblk, ctxt, ld.Segdwarf.Fileoff, ld.Segdwarf.Vaddr, ld.Segdwarf.Filelen)
+	wg.Wait()
 }
 
 func asmb2(ctxt *ld.Link) {

src/cmd/link/internal/s390x/asm.go

@@ -38,6 +38,7 @@ import (
 	"cmd/link/internal/sym"
 	"debug/elf"
 	"fmt"
+	"sync"
 )
 
 // gentext generates assembly to append the local moduledata to the global
@@ -506,28 +507,28 @@ func asmb(ctxt *ld.Link) {
 		ld.Asmbelfsetup()
 	}
 
+	var wg sync.WaitGroup
 	sect := ld.Segtext.Sections[0]
-	ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-	ld.Codeblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
-	for _, sect = range ld.Segtext.Sections[1:] {
-		ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-		ld.Datblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
+	offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+	ld.WriteParallel(&wg, ld.Codeblk, ctxt, offset, sect.Vaddr, sect.Length)
+
+	for _, sect := range ld.Segtext.Sections[1:] {
+		offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, offset, sect.Vaddr, sect.Length)
 	}
 
 	if ld.Segrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrodata.Vaddr), int64(ld.Segrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrodata.Fileoff, ld.Segrodata.Vaddr, ld.Segrodata.Filelen)
 	}
+
 	if ld.Segrelrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrelrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrelrodata.Vaddr), int64(ld.Segrelrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrelrodata.Fileoff, ld.Segrelrodata.Vaddr, ld.Segrelrodata.Filelen)
 	}
 
-	ctxt.Out.SeekSet(int64(ld.Segdata.Fileoff))
-	ld.Datblk(ctxt, int64(ld.Segdata.Vaddr), int64(ld.Segdata.Filelen))
+	ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segdata.Fileoff, ld.Segdata.Vaddr, ld.Segdata.Filelen)
 
-	ctxt.Out.SeekSet(int64(ld.Segdwarf.Fileoff))
-	ld.Dwarfblk(ctxt, int64(ld.Segdwarf.Vaddr), int64(ld.Segdwarf.Filelen))
+	ld.WriteParallel(&wg, ld.Dwarfblk, ctxt, ld.Segdwarf.Fileoff, ld.Segdwarf.Vaddr, ld.Segdwarf.Filelen)
+	wg.Wait()
 }
 
 func asmb2(ctxt *ld.Link) {

src/cmd/link/internal/x86/asm.go

@@ -38,6 +38,7 @@ import (
 	"cmd/link/internal/sym"
 	"debug/elf"
 	"log"
+	"sync"
 )
 
 // Append 4 bytes to s and create a R_CALL relocation targeting t to fill them in.
@@ -623,29 +624,31 @@ func asmb(ctxt *ld.Link) {
 		ld.Asmbelfsetup()
 	}
 
+	var wg sync.WaitGroup
 	sect := ld.Segtext.Sections[0]
-	ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-	// 0xCC is INT $3 - breakpoint instruction
-	ld.CodeblkPad(ctxt, ctxt.Out, int64(sect.Vaddr), int64(sect.Length), []byte{0xCC})
-	for _, sect = range ld.Segtext.Sections[1:] {
-		ctxt.Out.SeekSet(int64(sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff))
-		ld.Datblk(ctxt, int64(sect.Vaddr), int64(sect.Length))
+	offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+	f := func(ctxt *ld.Link, out *ld.OutBuf, start, length int64) {
+		ld.CodeblkPad(ctxt, out, start, length, []byte{0xCC})
+	}
+	ld.WriteParallel(&wg, f, ctxt, offset, sect.Vaddr, sect.Length)
+
+	for _, sect := range ld.Segtext.Sections[1:] {
+		offset := sect.Vaddr - ld.Segtext.Vaddr + ld.Segtext.Fileoff
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, offset, sect.Vaddr, sect.Length)
 	}
 
 	if ld.Segrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrodata.Vaddr), int64(ld.Segrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrodata.Fileoff, ld.Segrodata.Vaddr, ld.Segrodata.Filelen)
 	}
+
 	if ld.Segrelrodata.Filelen > 0 {
-		ctxt.Out.SeekSet(int64(ld.Segrelrodata.Fileoff))
-		ld.Datblk(ctxt, int64(ld.Segrelrodata.Vaddr), int64(ld.Segrelrodata.Filelen))
+		ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segrelrodata.Fileoff, ld.Segrelrodata.Vaddr, ld.Segrelrodata.Filelen)
 	}
 
-	ctxt.Out.SeekSet(int64(ld.Segdata.Fileoff))
-	ld.Datblk(ctxt, int64(ld.Segdata.Vaddr), int64(ld.Segdata.Filelen))
+	ld.WriteParallel(&wg, ld.Datblk, ctxt, ld.Segdata.Fileoff, ld.Segdata.Vaddr, ld.Segdata.Filelen)
 
-	ctxt.Out.SeekSet(int64(ld.Segdwarf.Fileoff))
-	ld.Dwarfblk(ctxt, int64(ld.Segdwarf.Vaddr), int64(ld.Segdwarf.Filelen))
+	ld.WriteParallel(&wg, ld.Dwarfblk, ctxt, ld.Segdwarf.Fileoff, ld.Segdwarf.Vaddr, ld.Segdwarf.Filelen)
+	wg.Wait()
 }
 
 func asmb2(ctxt *ld.Link) {