diff --git a/src/cmd/link/internal/amd64/asm.go b/src/cmd/link/internal/amd64/asm.go
index 24134adc27..f3b3d703b5 100644
--- a/src/cmd/link/internal/amd64/asm.go
+++ b/src/cmd/link/internal/amd64/asm.go
@@ -841,7 +841,7 @@ func asmb2(ctxt *ld.Link) {
 	}
 }
 
-func tlsIEtoLE(s *sym.Symbol, off, size int) {
+func tlsIEtoLE(P []byte, off, size int) {
 	// Transform the PC-relative instruction into a constant load.
 	// That is,
 	//
@@ -852,7 +852,7 @@ func tlsIEtoLE(s *sym.Symbol, off, size int) {
 	if off < 3 {
 		log.Fatal("R_X86_64_GOTTPOFF reloc not preceded by MOVQ or ADDQ instruction")
 	}
-	op := s.P[off-3 : off]
+	op := P[off-3 : off]
 	reg := op[2] >> 3
 
 	if op[1] == 0x8b || reg == 4 {
diff --git a/src/cmd/link/internal/ld/data.go b/src/cmd/link/internal/ld/data.go
index d44d590623..44e9b884ff 100644
--- a/src/cmd/link/internal/ld/data.go
+++ b/src/cmd/link/internal/ld/data.go
@@ -271,7 +271,6 @@ func relocsym(target *Target, ldr *loader.Loader, err *ErrorReporter, syms *Arch
 				log.Fatalf("unexpected R_TLS_LE relocation for %v", target.HeadType)
 			}
 		case objabi.R_TLS_IE:
-			panic("not implemented")
 			//if target.IsExternal() && target.IsElf() {
 			//	r.Done = false
 			//	if r.Sym == nil {
@@ -285,17 +284,17 @@ func relocsym(target *Target, ldr *loader.Loader, err *ErrorReporter, syms *Arch
 			//	}
 			//	break
 			//}
-			//if target.IsPIE() && target.IsElf() {
-			//	// We are linking the final executable, so we
-			//	// can optimize any TLS IE relocation to LE.
-			//	if thearch.TLSIEtoLE == nil {
-			//		log.Fatalf("internal linking of TLS IE not supported on %v", target.Arch.Family)
-			//	}
-			//	thearch.TLSIEtoLE(ldr, s, int(off), int(siz))
-			//	o = int64(syms.Tlsoffset)
-			//} else {
-			//	log.Fatalf("cannot handle R_TLS_IE (sym %s) when linking internally", ldr.SymName(s))
-			//}
+			if target.IsPIE() && target.IsElf() {
+				// We are linking the final executable, so we
+				// can optimize any TLS IE relocation to LE.
+				if thearch.TLSIEtoLE == nil {
+					log.Fatalf("internal linking of TLS IE not supported on %v", target.Arch.Family)
+				}
+				thearch.TLSIEtoLE(P, int(off), int(siz))
+				o = int64(syms.Tlsoffset)
+			} else {
+				log.Fatalf("cannot handle R_TLS_IE (sym %s) when linking internally", ldr.SymName(s))
+			}
 		case objabi.R_ADDR:
 			//if target.IsExternal() && r.Sym.Type != sym.SCONST {
 			//	r.Done = false
@@ -562,22 +561,22 @@ func (ctxt *Link) reloc() {
 	wg.Add(3)
 	go func() {
 		if !ctxt.IsWasm() { // On Wasm, text relocations are applied in Asmb2.
-			for _, s := range ctxt.Textp {
-				relocsym2(target, ldr, reporter, syms, s)
+			for _, s := range ctxt.Textp2 {
+				relocsym(target, ldr, reporter, syms, s, ldr.OutData(s))
 			}
 		}
 		wg.Done()
 	}()
 	go func() {
-		for _, s := range ctxt.datap {
-			relocsym2(target, ldr, reporter, syms, s)
+		for _, s := range ctxt.datap2 {
+			relocsym(target, ldr, reporter, syms, s, ldr.OutData(s))
 		}
 		wg.Done()
 	}()
 	go func() {
-		for _, si := range dwarfp {
+		for _, si := range dwarfp2 {
 			for _, s := range si.syms {
-				relocsym2(target, ldr, reporter, syms, s)
+				relocsym(target, ldr, reporter, syms, s, ldr.OutData(s))
 			}
 		}
 		wg.Done()
diff --git a/src/cmd/link/internal/ld/data2.go b/src/cmd/link/internal/ld/data2.go
index b9c8cd83de..d4503a4b0a 100644
--- a/src/cmd/link/internal/ld/data2.go
+++ b/src/cmd/link/internal/ld/data2.go
@@ -11,6 +11,7 @@ import (
 	"fmt"
 	"log"
 	"strings"
+	"sync"
 )
 
 // Temporary dumping around for sym.Symbol version of helper
@@ -199,7 +200,7 @@ func relocsym2(target *Target, ldr *loader.Loader, err *ErrorReporter, syms *Arc
 				if thearch.TLSIEtoLE == nil {
 					log.Fatalf("internal linking of TLS IE not supported on %v", target.Arch.Family)
 				}
-				thearch.TLSIEtoLE(s, int(off), int(r.Siz))
+				thearch.TLSIEtoLE(s.P, int(off), int(r.Siz))
 				o = int64(syms.Tlsoffset)
 				// TODO: o += r.Add when !target.IsAmd64()?
 				// Why do we treat r.Add differently on AMD64?
@@ -483,3 +484,35 @@ func relocsym2(target *Target, ldr *loader.Loader, err *ErrorReporter, syms *Arc
 		}
 	}
 }
+
+func (ctxt *Link) reloc2() {
+	var wg sync.WaitGroup
+	target := &ctxt.Target
+	ldr := ctxt.loader
+	reporter := &ctxt.ErrorReporter
+	syms := &ctxt.ArchSyms
+	wg.Add(3)
+	go func() {
+		if !ctxt.IsWasm() { // On Wasm, text relocations are applied in Asmb2.
+			for _, s := range ctxt.Textp {
+				relocsym2(target, ldr, reporter, syms, s)
+			}
+		}
+		wg.Done()
+	}()
+	go func() {
+		for _, s := range ctxt.datap {
+			relocsym2(target, ldr, reporter, syms, s)
+		}
+		wg.Done()
+	}()
+	go func() {
+		for _, si := range dwarfp {
+			for _, s := range si.syms {
+				relocsym2(target, ldr, reporter, syms, s)
+			}
+		}
+		wg.Done()
+	}()
+	wg.Wait()
+}
diff --git a/src/cmd/link/internal/ld/lib.go b/src/cmd/link/internal/ld/lib.go
index e1f89858b6..618faf2233 100644
--- a/src/cmd/link/internal/ld/lib.go
+++ b/src/cmd/link/internal/ld/lib.go
@@ -282,7 +282,7 @@ type Arch struct {
 	// This is possible when a TLS IE relocation refers to a local
 	// symbol in an executable, which is typical when internally
 	// linking PIE binaries.
-	TLSIEtoLE func(s *sym.Symbol, off, size int)
+	TLSIEtoLE func(P []byte, off, size int)
 
 	// optional override for assignAddress
 	AssignAddress func(ldr *loader.Loader, sect *sym.Section, n int, s loader.Sym, va uint64, isTramp bool) (*sym.Section, int, uint64)
@@ -2824,10 +2824,9 @@ func addToTextp(ctxt *Link) {
 	ctxt.Textp = textp
 }
 
-func (ctxt *Link) loadlibfull(symGroupType []sym.SymKind) {
-
+func (ctxt *Link) loadlibfull(symGroupType []sym.SymKind, needReloc bool) {
 	// Load full symbol contents, resolve indexed references.
-	ctxt.loader.LoadFull(ctxt.Arch, ctxt.Syms)
+	ctxt.loader.LoadFull(ctxt.Arch, ctxt.Syms, needReloc)
 
 	// Convert ctxt.Moduledata2 to ctxt.Moduledata, etc
 	if ctxt.Moduledata2 != 0 {
diff --git a/src/cmd/link/internal/ld/main.go b/src/cmd/link/internal/ld/main.go
index a4526230c9..84f40d9b81 100644
--- a/src/cmd/link/internal/ld/main.go
+++ b/src/cmd/link/internal/ld/main.go
@@ -325,10 +325,22 @@ func Main(arch *sys.Arch, theArch Arch) {
 	bench.Start("Asmb")
 	ctxt.loader.InitOutData()
 	thearch.Asmb(ctxt, ctxt.loader)
-	bench.Start("loadlibfull")
-	ctxt.loadlibfull(symGroupType) // XXX do it here for now
-	bench.Start("reloc")
-	ctxt.reloc()
+
+	newreloc := ctxt.IsInternal() && ctxt.IsAMD64()
+	if newreloc {
+		bench.Start("reloc")
+		ctxt.reloc()
+		bench.Start("loadlibfull")
+		// We don't need relocations at this point.
+		// An exception is Windows, see pe.go:addPEBaseRelocSym
+		needReloc := ctxt.IsWindows()
+		ctxt.loadlibfull(symGroupType, needReloc) // XXX do it here for now
+	} else {
+		bench.Start("loadlibfull")
+		ctxt.loadlibfull(symGroupType, true) // XXX do it here for now
+		bench.Start("reloc")
+		ctxt.reloc2()
+	}
 	bench.Start("Asmb2")
 	thearch.Asmb2(ctxt)
 
diff --git a/src/cmd/link/internal/loader/loader.go b/src/cmd/link/internal/loader/loader.go
index 775a2cb522..2abd0e60e1 100644
--- a/src/cmd/link/internal/loader/loader.go
+++ b/src/cmd/link/internal/loader/loader.go
@@ -2016,7 +2016,7 @@ func (l *Loader) preprocess(arch *sys.Arch, s Sym, name string) {
 }
 
 // Load full contents.
-func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols) {
+func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols, needReloc bool) {
 	// create all Symbols first.
 	l.growSyms(l.NSym())
 	l.growSects(l.NSym())
@@ -2049,7 +2049,9 @@ func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols) {
 	}
 
 	// allocate a single large slab of relocations for all live symbols
-	l.relocBatch = make([]sym.Reloc, nr)
+	if needReloc {
+		l.relocBatch = make([]sym.Reloc, nr)
+	}
 
 	// convert payload-based external symbols into sym.Symbol-based
 	for _, i := range toConvert {
@@ -2062,11 +2064,13 @@ func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols) {
 		s.Size = pp.size
 
 		// Copy relocations
-		batch := l.relocBatch
-		s.R = batch[:len(pp.relocs):len(pp.relocs)]
-		l.relocBatch = batch[len(pp.relocs):]
-		relocs := l.Relocs(i)
-		l.convertRelocations(i, &relocs, s, false)
+		if needReloc {
+			batch := l.relocBatch
+			s.R = batch[:len(pp.relocs):len(pp.relocs)]
+			l.relocBatch = batch[len(pp.relocs):]
+			relocs := l.Relocs(i)
+			l.convertRelocations(i, &relocs, s, false)
+		}
 
 		// Copy data
 		s.P = pp.data
@@ -2077,7 +2081,7 @@ func (l *Loader) LoadFull(arch *sys.Arch, syms *sym.Symbols) {
 
 	// load contents of defined symbols
 	for _, o := range l.objs[1:] {
-		loadObjFull(l, o.r)
+		loadObjFull(l, o.r, needReloc)
 	}
 
 	// Note: resolution of ABI aliases is now also handled in
@@ -2598,7 +2602,7 @@ func (l *Loader) CreateStaticSym(name string) Sym {
 	return l.newExtSym(name, l.anonVersion)
 }
 
-func loadObjFull(l *Loader, r *oReader) {
+func loadObjFull(l *Loader, r *oReader, needReloc bool) {
 	for i, n := 0, r.NSym()+r.NNonpkgdef(); i < n; i++ {
 		// A symbol may be a dup or overwritten. In this case, its
 		// content will actually be provided by a different object
@@ -2623,11 +2627,13 @@ func loadObjFull(l *Loader, r *oReader) {
 		s.P = l.OutData(gi)
 
 		// Relocs
-		relocs := l.relocs(r, i)
-		batch := l.relocBatch
-		s.R = batch[:relocs.Count():relocs.Count()]
-		l.relocBatch = batch[relocs.Count():]
-		l.convertRelocations(gi, &relocs, s, false)
+		if needReloc {
+			relocs := l.relocs(r, i)
+			batch := l.relocBatch
+			s.R = batch[:relocs.Count():relocs.Count()]
+			l.relocBatch = batch[relocs.Count():]
+			l.convertRelocations(gi, &relocs, s, false)
+		}
 
 		// Aux symbol info
 		auxs := r.Auxs(i)