cmd/link: support PIE mode with internal link on linux arm64

This CL improves internal link to provide basic support for cgo and PIE:
1, add support for GOT, PLT and GOTPLT.
2, add support for following ELF relocation types which have been used by std
   packages:
     R_AARCH64_ADR_GOT_PAGE
     R_AARCH64_LD64_GOT_LO12_NC
     R_AARCH64_ADR_PREL_PG_HI21
     R_AARCH64_ADD_ABS_LO12_NC
     R_AARCH64_LDST8_ABS_LO12_NC
     R_AARCH64_LDST32_ABS_LO12_NC
     R_AARCH64_LDST64_ABS_LO12_NC
     R_AARCH64_JUMP26
     R_AARCH64_ABS64
     R_AARCH64_PREL32
     R_AARCH64_PREL64

With this change, Go toolchain can be built in internal linking mode, and
pure Go programs can be built with PIE mode in internal linking mode on arm64.

Updates #10373
The prototype of this CL is contributed by Wei Xiao <wei.xiao@arm.com>

Change-Id: I2253923c69e855fd1524d54def309a961dce6247
Reviewed-on: https://go-review.googlesource.com/c/go/+/163579Reviewed-by: Cherry Zhang <cherryyz@google.com>
Run-TryBot: Cherry Zhang <cherryyz@google.com>

src/cmd/dist/test.go

@@ -569,7 +569,7 @@ func (t *tester) registerTests() {
 	}
 
 	// Test internal linking of PIE binaries where it is supported.
-	if goos == "linux" && goarch == "amd64" {
+	if goos == "linux" && (goarch == "amd64" || goarch == "arm64") {
 		t.tests = append(t.tests, distTest{
 			name:    "pie_internal",
 			heading: "internal linking of -buildmode=pie",
@@ -910,7 +910,7 @@ func (t *tester) internalLink() bool {
 	// Internally linking cgo is incomplete on some architectures.
 	// https://golang.org/issue/10373
 	// https://golang.org/issue/14449
-	if goarch == "arm64" || goarch == "mips64" || goarch == "mips64le" || goarch == "mips" || goarch == "mipsle" {
+	if goarch == "mips64" || goarch == "mips64le" || goarch == "mips" || goarch == "mipsle" {
 		return false
 	}
 	if goos == "aix" {
@@ -1030,7 +1030,7 @@ func (t *tester) cgoTest(dt *distTest) error {
 		"dragonfly-amd64",
 		"freebsd-386", "freebsd-amd64", "freebsd-arm",
 		"linux-386", "linux-amd64", "linux-arm", "linux-ppc64le", "linux-s390x",
-		"netbsd-386", "netbsd-amd64":
+		"netbsd-386", "netbsd-amd64", "linux-arm64":
 
 		cmd := t.addCmd(dt, "misc/cgo/test", t.goTest())
 		cmd.Env = append(os.Environ(), "GOFLAGS=-ldflags=-linkmode=external")

src/cmd/internal/objabi/reloctype.go

@@ -132,6 +132,23 @@ const (
 	// slot of the referenced symbol.
 	R_ARM64_GOTPCREL
 
+	// R_ARM64_GOT resolves a GOT-relative instruction sequence, usually an adrp
+	// followed by another ld instruction.
+	R_ARM64_GOT
+
+	// R_ARM64_PCREL resolves a PC-relative addresses instruction sequence, usually an
+	// adrp followed by another add instruction.
+	R_ARM64_PCREL
+
+	// R_ARM64_LDST8 sets a LD/ST immediate value to bits [11:0] of a local address.
+	R_ARM64_LDST8
+
+	// R_ARM64_LDST32 sets a LD/ST immediate value to bits [11:2] of a local address.
+	R_ARM64_LDST32
+
+	// R_ARM64_LDST64 sets a LD/ST immediate value to bits [11:3] of a local address.
+	R_ARM64_LDST64
+
 	// PPC64.
 
 	// R_POWER_TLS_LE is used to implement the "local exec" model for tls

src/cmd/internal/objabi/reloctype_string.go

-// Code generated by "stringer -type=RelocType"; DO NOT EDIT.
+// Code generated by "stringer -type RelocType"; DO NOT EDIT.
 
 package objabi
 
 import "strconv"
 
-const _RelocType_name = "R_ADDRR_ADDRPOWERR_ADDRARM64R_ADDRMIPSR_ADDROFFR_WEAKADDROFFR_SIZER_CALLR_CALLARMR_CALLARM64R_CALLINDR_CALLPOWERR_CALLMIPSR_CONSTR_PCRELR_TLS_LER_TLS_IER_GOTOFFR_PLT0R_PLT1R_PLT2R_USEFIELDR_USETYPER_METHODOFFR_POWER_TOCR_GOTPCRELR_JMPMIPSR_DWARFSECREFR_DWARFFILEREFR_ARM64_TLS_LER_ARM64_TLS_IER_ARM64_GOTPCRELR_POWER_TLS_LER_POWER_TLS_IER_POWER_TLSR_ADDRPOWER_DSR_ADDRPOWER_GOTR_ADDRPOWER_PCRELR_ADDRPOWER_TOCRELR_ADDRPOWER_TOCREL_DSR_PCRELDBLR_ADDRMIPSUR_ADDRMIPSTLSR_ADDRCUOFFR_WASMIMPORT"
+const _RelocType_name = "R_ADDRR_ADDRPOWERR_ADDRARM64R_ADDRMIPSR_ADDROFFR_WEAKADDROFFR_SIZER_CALLR_CALLARMR_CALLARM64R_CALLINDR_CALLPOWERR_CALLMIPSR_CONSTR_PCRELR_TLS_LER_TLS_IER_GOTOFFR_PLT0R_PLT1R_PLT2R_USEFIELDR_USETYPER_METHODOFFR_POWER_TOCR_GOTPCRELR_JMPMIPSR_DWARFSECREFR_DWARFFILEREFR_ARM64_TLS_LER_ARM64_TLS_IER_ARM64_GOTPCRELR_ARM64_GOTR_ARM64_PCRELR_ARM64_LDST8R_ARM64_LDST32R_ARM64_LDST64R_POWER_TLS_LER_POWER_TLS_IER_POWER_TLSR_ADDRPOWER_DSR_ADDRPOWER_GOTR_ADDRPOWER_PCRELR_ADDRPOWER_TOCRELR_ADDRPOWER_TOCREL_DSR_PCRELDBLR_ADDRMIPSUR_ADDRMIPSTLSR_ADDRCUOFFR_WASMIMPORTR_XCOFFREF"
 
-var _RelocType_index = [...]uint16{0, 6, 17, 28, 38, 47, 60, 66, 72, 81, 92, 101, 112, 122, 129, 136, 144, 152, 160, 166, 172, 178, 188, 197, 208, 219, 229, 238, 251, 265, 279, 293, 309, 323, 337, 348, 362, 377, 394, 412, 433, 443, 454, 467, 478, 490}
+var _RelocType_index = [...]uint16{0, 6, 17, 28, 38, 47, 60, 66, 72, 81, 92, 101, 112, 122, 129, 136, 144, 152, 160, 166, 172, 178, 188, 197, 208, 219, 229, 238, 251, 265, 279, 293, 309, 320, 333, 346, 360, 374, 388, 402, 413, 427, 442, 459, 477, 498, 508, 519, 532, 543, 555, 565}
 
 func (i RelocType) String() string {
 	i -= 1

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

@@ -92,11 +92,110 @@ func gentext(ctxt *ld.Link) {
 	initarray_entry.AddAddr(ctxt.Arch, initfunc)
 }
 
-// adddynrel implements just enough to support external linking to
-// the system libc functions used by the runtime.
 func adddynrel(ctxt *ld.Link, s *sym.Symbol, r *sym.Reloc) bool {
 	targ := r.Sym
 
+	switch r.Type {
+	default:
+		if r.Type >= objabi.ElfRelocOffset {
+			ld.Errorf(s, "unexpected relocation type %d (%s)", r.Type, sym.RelocName(ctxt.Arch, r.Type))
+			return false
+		}
+
+	// Handle relocations found in ELF object files.
+	case objabi.ElfRelocOffset + objabi.RelocType(elf.R_AARCH64_PREL32):
+		if targ.Type == sym.SDYNIMPORT {
+			ld.Errorf(s, "unexpected R_AARCH64_PREL32 relocation for dynamic symbol %s", targ.Name)
+		}
+		// TODO(mwhudson): the test of VisibilityHidden here probably doesn't make
+		// sense and should be removed when someone has thought about it properly.
+		if (targ.Type == 0 || targ.Type == sym.SXREF) && !targ.Attr.VisibilityHidden() {
+			ld.Errorf(s, "unknown symbol %s in pcrel", targ.Name)
+		}
+		r.Type = objabi.R_PCREL
+		r.Add += 4
+		return true
+
+	case objabi.ElfRelocOffset + objabi.RelocType(elf.R_AARCH64_PREL64):
+		if targ.Type == sym.SDYNIMPORT {
+			ld.Errorf(s, "unexpected R_AARCH64_PREL64 relocation for dynamic symbol %s", targ.Name)
+		}
+		if targ.Type == 0 || targ.Type == sym.SXREF {
+			ld.Errorf(s, "unknown symbol %s in pcrel", targ.Name)
+		}
+		r.Type = objabi.R_PCREL
+		r.Add += 8
+		return true
+
+	case objabi.ElfRelocOffset + objabi.RelocType(elf.R_AARCH64_CALL26),
+		objabi.ElfRelocOffset + objabi.RelocType(elf.R_AARCH64_JUMP26):
+		if targ.Type == sym.SDYNIMPORT {
+			addpltsym(ctxt, targ)
+			r.Sym = ctxt.Syms.Lookup(".plt", 0)
+			r.Add += int64(targ.Plt())
+		}
+		if (targ.Type == 0 || targ.Type == sym.SXREF) && !targ.Attr.VisibilityHidden() {
+			ld.Errorf(s, "unknown symbol %s in callarm64", targ.Name)
+		}
+		r.Type = objabi.R_CALLARM64
+		return true
+
+	case objabi.ElfRelocOffset + objabi.RelocType(elf.R_AARCH64_ADR_GOT_PAGE),
+		objabi.ElfRelocOffset + objabi.RelocType(elf.R_AARCH64_LD64_GOT_LO12_NC):
+		if targ.Type != sym.SDYNIMPORT {
+			// have symbol
+			// TODO: turn LDR of GOT entry into ADR of symbol itself
+		}
+
+		// fall back to using GOT
+		// TODO: just needs relocation, no need to put in .dynsym
+		addgotsym(ctxt, targ)
+
+		r.Type = objabi.R_ARM64_GOT
+		r.Sym = ctxt.Syms.Lookup(".got", 0)
+		r.Add += int64(targ.Got())
+		return true
+
+	case objabi.ElfRelocOffset + objabi.RelocType(elf.R_AARCH64_ADR_PREL_PG_HI21),
+		objabi.ElfRelocOffset + objabi.RelocType(elf.R_AARCH64_ADD_ABS_LO12_NC):
+		if targ.Type == sym.SDYNIMPORT {
+			ld.Errorf(s, "unexpected relocation for dynamic symbol %s", targ.Name)
+		}
+		if targ.Type == 0 || targ.Type == sym.SXREF {
+			ld.Errorf(s, "unknown symbol %s", targ.Name)
+		}
+		r.Type = objabi.R_ARM64_PCREL
+		return true
+
+	case objabi.ElfRelocOffset + objabi.RelocType(elf.R_AARCH64_ABS64):
+		if targ.Type == sym.SDYNIMPORT {
+			ld.Errorf(s, "unexpected R_AARCH64_ABS64 relocation for dynamic symbol %s", targ.Name)
+		}
+		r.Type = objabi.R_ADDR
+		return true
+
+	case objabi.ElfRelocOffset + objabi.RelocType(elf.R_AARCH64_LDST8_ABS_LO12_NC):
+		if targ.Type == sym.SDYNIMPORT {
+			ld.Errorf(s, "unexpected relocation for dynamic symbol %s", targ.Name)
+		}
+		r.Type = objabi.R_ARM64_LDST8
+		return true
+
+	case objabi.ElfRelocOffset + objabi.RelocType(elf.R_AARCH64_LDST32_ABS_LO12_NC):
+		if targ.Type == sym.SDYNIMPORT {
+			ld.Errorf(s, "unexpected relocation for dynamic symbol %s", targ.Name)
+		}
+		r.Type = objabi.R_ARM64_LDST32
+		return true
+
+	case objabi.ElfRelocOffset + objabi.RelocType(elf.R_AARCH64_LDST64_ABS_LO12_NC):
+		if targ.Type == sym.SDYNIMPORT {
+			ld.Errorf(s, "unexpected relocation for dynamic symbol %s", targ.Name)
+		}
+		r.Type = objabi.R_ARM64_LDST64
+		return true
+	}
+
 	switch r.Type {
 	case objabi.R_CALL,
 		objabi.R_PCREL,
@@ -109,8 +208,85 @@ func adddynrel(ctxt *ld.Link, s *sym.Symbol, r *sym.Reloc) bool {
 			// External linker will do this relocation.
 			return true
 		}
+
+	case objabi.R_ADDR:
+		if s.Type == sym.STEXT && ctxt.IsELF {
+			// The code is asking for the address of an external
+			// function. We provide it with the address of the
+			// correspondent GOT symbol.
+			addgotsym(ctxt, targ)
+
+			r.Sym = ctxt.Syms.Lookup(".got", 0)
+			r.Add += int64(targ.Got())
+			return true
+		}
+
+		// Process dynamic relocations for the data sections.
+		if ctxt.BuildMode == ld.BuildModePIE && ctxt.LinkMode == ld.LinkInternal {
+			// When internally linking, generate dynamic relocations
+			// for all typical R_ADDR relocations. The exception
+			// are those R_ADDR that are created as part of generating
+			// the dynamic relocations and must be resolved statically.
+			//
+			// There are three phases relevant to understanding this:
+			//
+			//	dodata()  // we are here
+			//	address() // symbol address assignment
+			//	reloc()   // resolution of static R_ADDR relocs
+			//
+			// At this point symbol addresses have not been
+			// assigned yet (as the final size of the .rela section
+			// will affect the addresses), and so we cannot write
+			// the Elf64_Rela.r_offset now. Instead we delay it
+			// until after the 'address' phase of the linker is
+			// complete. We do this via Addaddrplus, which creates
+			// a new R_ADDR relocation which will be resolved in
+			// the 'reloc' phase.
+			//
+			// These synthetic static R_ADDR relocs must be skipped
+			// now, or else we will be caught in an infinite loop
+			// of generating synthetic relocs for our synthetic
+			// relocs.
+			//
+			// Furthermore, the rela sections contain dynamic
+			// relocations with R_ADDR relocations on
+			// Elf64_Rela.r_offset. This field should contain the
+			// symbol offset as determined by reloc(), not the
+			// final dynamically linked address as a dynamic
+			// relocation would provide.
+			switch s.Name {
+			case ".dynsym", ".rela", ".rela.plt", ".got.plt", ".dynamic":
+				return false
+			}
+		} else {
+			// Either internally linking a static executable,
+			// in which case we can resolve these relocations
+			// statically in the 'reloc' phase, or externally
+			// linking, in which case the relocation will be
+			// prepared in the 'reloc' phase and passed to the
+			// external linker in the 'asmb' phase.
+			if s.Type != sym.SDATA && s.Type != sym.SRODATA {
+				break
+			}
+		}
+
+		if ctxt.IsELF {
+			// TODO: We generate a R_AARCH64_ABS64 relocation for every R_ADDR, even
+			// though it would be more efficient (for the dynamic linker) if we
+			// generated R_AARCH64_RELATIVE instead.
+			ld.Adddynsym(ctxt, targ)
+			rela := ctxt.Syms.Lookup(".rela", 0)
+			rela.AddAddrPlus(ctxt.Arch, s, int64(r.Off))
+			if r.Siz == 8 {
+				rela.AddUint64(ctxt.Arch, ld.ELF64_R_INFO(uint32(targ.Dynid), uint32(elf.R_AARCH64_ABS64)))
+			} else {
+				ld.Errorf(s, "unexpected relocation for dynamic symbol %s", targ.Name)
+			}
+			rela.AddUint64(ctxt.Arch, uint64(r.Add))
+			r.Type = objabi.ElfRelocOffset // ignore during relocsym
+			return true
+		}
 	}
-	log.Fatalf("adddynrel not implemented for relocation type %d (%s)", r.Type, sym.RelocName(ctxt.Arch, r.Type))
 	return false
 }
 
@@ -160,11 +336,6 @@ func elfreloc1(ctxt *ld.Link, r *sym.Reloc, sectoff int64) bool {
 	return true
 }
 
-func elfsetupplt(ctxt *ld.Link) {
-	// TODO(aram)
-	return
-}
-
 func machoreloc1(arch *sys.Arch, out *ld.OutBuf, s *sym.Symbol, r *sym.Reloc, sectoff int64) bool {
 	var v uint32
 
@@ -331,8 +502,10 @@ func archreloc(ctxt *ld.Link, r *sym.Reloc, s *sym.Symbol, val int64) (int64, bo
 	switch r.Type {
 	case objabi.R_CONST:
 		return r.Add, true
+
 	case objabi.R_GOTOFF:
 		return ld.Symaddr(r.Sym) + r.Add - ld.Symaddr(ctxt.Syms.Lookup(".got", 0)), true
+
 	case objabi.R_ADDRARM64:
 		t := ld.Symaddr(r.Sym) + r.Add - ((s.Value + int64(r.Off)) &^ 0xfff)
 		if t >= 1<<32 || t < -1<<32 {
@@ -357,6 +530,7 @@ func archreloc(ctxt *ld.Link, r *sym.Reloc, s *sym.Symbol, val int64) (int64, bo
 			return int64(o0)<<32 | int64(o1), true
 		}
 		return int64(o1)<<32 | int64(o0), true
+
 	case objabi.R_ARM64_TLS_LE:
 		r.Done = false
 		if ctxt.HeadType == objabi.Hdarwin {
@@ -369,12 +543,128 @@ func archreloc(ctxt *ld.Link, r *sym.Reloc, s *sym.Symbol, val int64) (int64, bo
 			ld.Errorf(s, "TLS offset out of range %d", v)
 		}
 		return val | (v << 5), true
+
+	case objabi.R_ARM64_TLS_IE:
+		if ctxt.BuildMode == ld.BuildModePIE && ctxt.IsELF {
+			// We are linking the final executable, so we
+			// can optimize any TLS IE relocation to LE.
+			r.Done = false
+			if ctxt.HeadType != objabi.Hlinux {
+				ld.Errorf(s, "TLS reloc on unsupported OS %v", ctxt.HeadType)
+			}
+
+			// The TCB is two pointers. This is not documented anywhere, but is
+			// de facto part of the ABI.
+			v := ld.Symaddr(r.Sym) + int64(2*ctxt.Arch.PtrSize) + r.Add
+			if v < 0 || v >= 32678 {
+				ld.Errorf(s, "TLS offset out of range %d", v)
+			}
+
+			var o0, o1 uint32
+			if ctxt.Arch.ByteOrder == binary.BigEndian {
+				o0 = uint32(val >> 32)
+				o1 = uint32(val)
+			} else {
+				o0 = uint32(val)
+				o1 = uint32(val >> 32)
+			}
+
+			// R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21
+			// turn ADRP to MOVZ
+			o0 = 0xd2a00000 | uint32(o0&0x1f) | (uint32((v>>16)&0xffff) << 5)
+			// R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC
+			// turn LD64 to MOVK
+			if v&3 != 0 {
+				ld.Errorf(s, "invalid address: %x for relocation type: R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC", v)
+			}
+			o1 = 0xf2800000 | uint32(o1&0x1f) | (uint32(v&0xffff) << 5)
+
+			// when laid out, the instruction order must always be o0, o1.
+			if ctxt.Arch.ByteOrder == binary.BigEndian {
+				return int64(o0)<<32 | int64(o1), true
+			}
+			return int64(o1)<<32 | int64(o0), true
+		} else {
+			log.Fatalf("cannot handle R_ARM64_TLS_IE (sym %s) when linking internally", s.Name)
+		}
+
 	case objabi.R_CALLARM64:
-		t := (ld.Symaddr(r.Sym) + r.Add) - (s.Value + int64(r.Off))
+		var t int64
+		if r.Sym.Type == sym.SDYNIMPORT {
+			t = (ld.Symaddr(ctxt.Syms.Lookup(".plt", 0)) + r.Add) - (s.Value + int64(r.Off))
+		} else {
+			t = (ld.Symaddr(r.Sym) + r.Add) - (s.Value + int64(r.Off))
+		}
 		if t >= 1<<27 || t < -1<<27 {
 			ld.Errorf(s, "program too large, call relocation distance = %d", t)
 		}
 		return val | ((t >> 2) & 0x03ffffff), true
+
+	case objabi.R_ARM64_GOT:
+		if s.P[r.Off+3]&0x9f == 0x90 {
+			// R_AARCH64_ADR_GOT_PAGE
+			// patch instruction: adrp
+			t := ld.Symaddr(r.Sym) + r.Add - ((s.Value + int64(r.Off)) &^ 0xfff)
+			if t >= 1<<32 || t < -1<<32 {
+				ld.Errorf(s, "program too large, address relocation distance = %d", t)
+			}
+			var o0 uint32
+			o0 |= (uint32((t>>12)&3) << 29) | (uint32((t>>12>>2)&0x7ffff) << 5)
+			return val | int64(o0), true
+		} else if s.P[r.Off+3] == 0xf9 {
+			// R_AARCH64_LD64_GOT_LO12_NC
+			// patch instruction: ldr
+			t := ld.Symaddr(r.Sym) + r.Add - ((s.Value + int64(r.Off)) &^ 0xfff)
+			if t&7 != 0 {
+				ld.Errorf(s, "invalid address: %x for relocation type: R_AARCH64_LD64_GOT_LO12_NC", t)
+			}
+			var o1 uint32
+			o1 |= uint32(t&0xfff) << (10 - 3)
+			return val | int64(uint64(o1)), true
+		} else {
+			ld.Errorf(s, "unsupported instruction for %v R_GOTARM64", s.P[r.Off:r.Off+4])
+		}
+
+	case objabi.R_ARM64_PCREL:
+		if s.P[r.Off+3]&0x9f == 0x90 {
+			// R_AARCH64_ADR_PREL_PG_HI21
+			// patch instruction: adrp
+			t := ld.Symaddr(r.Sym) + r.Add - ((s.Value + int64(r.Off)) &^ 0xfff)
+			if t >= 1<<32 || t < -1<<32 {
+				ld.Errorf(s, "program too large, address relocation distance = %d", t)
+			}
+			o0 := (uint32((t>>12)&3) << 29) | (uint32((t>>12>>2)&0x7ffff) << 5)
+			return val | int64(o0), true
+		} else if s.P[r.Off+3]&0x91 == 0x91 {
+			// R_AARCH64_ADD_ABS_LO12_NC
+			// patch instruction: add
+			t := ld.Symaddr(r.Sym) + r.Add - ((s.Value + int64(r.Off)) &^ 0xfff)
+			o1 := uint32(t&0xfff) << 10
+			return val | int64(o1), true
+		} else {
+			ld.Errorf(s, "unsupported instruction for %v R_PCRELARM64", s.P[r.Off:r.Off+4])
+		}
+
+	case objabi.R_ARM64_LDST8:
+		t := ld.Symaddr(r.Sym) + r.Add - ((s.Value + int64(r.Off)) &^ 0xfff)
+		o0 := uint32(t&0xfff) << 10
+		return val | int64(o0), true
+
+	case objabi.R_ARM64_LDST32:
+		t := ld.Symaddr(r.Sym) + r.Add - ((s.Value + int64(r.Off)) &^ 0xfff)
+		if t&3 != 0 {
+			ld.Errorf(s, "invalid address: %x for relocation type: R_AARCH64_LDST32_ABS_LO12_NC", t)
+		}
+		o0 := (uint32(t&0xfff) >> 2) << 10
+		return val | int64(o0), true
+
+	case objabi.R_ARM64_LDST64:
+		t := ld.Symaddr(r.Sym) + r.Add - ((s.Value + int64(r.Off)) &^ 0xfff)
+		if t&7 != 0 {
+			ld.Errorf(s, "invalid address: %x for relocation type: R_AARCH64_LDST64_ABS_LO12_NC", t)
+		}
+		o0 := (uint32(t&0xfff) >> 3) << 10
+		return val | int64(o0), true
 	}
 
 	return val, false
@@ -385,6 +675,118 @@ func archrelocvariant(ctxt *ld.Link, r *sym.Reloc, s *sym.Symbol, t int64) int64
 	return -1
 }
 
+func elfsetupplt(ctxt *ld.Link) {
+	plt := ctxt.Syms.Lookup(".plt", 0)
+	gotplt := ctxt.Syms.Lookup(".got.plt", 0)
+	if plt.Size == 0 {
+		// stp     x16, x30, [sp, #-16]!
+		// identifying information
+		plt.AddUint32(ctxt.Arch, 0xa9bf7bf0)
+
+		// the following two instructions (adrp + ldr) load *got[2] into x17
+		// adrp    x16, &got[0]
+		plt.AddAddrPlus4(gotplt, 16)
+		plt.SetUint32(ctxt.Arch, plt.Size-4, 0x90000010)
+		plt.R[len(plt.R)-1].Type = objabi.R_ARM64_GOT
+
+		// <imm> is the offset value of &got[2] to &got[0], the same below
+		// ldr     x17, [x16, <imm>]
+		plt.AddAddrPlus4(gotplt, 16)
+		plt.SetUint32(ctxt.Arch, plt.Size-4, 0xf9400211)
+		plt.R[len(plt.R)-1].Type = objabi.R_ARM64_GOT
+
+		// add     x16, x16, <imm>
+		plt.AddAddrPlus4(gotplt, 16)
+		plt.SetUint32(ctxt.Arch, plt.Size-4, 0x91000210)
+		plt.R[len(plt.R)-1].Type = objabi.R_ARM64_PCREL
+
+		// br      x17
+		plt.AddUint32(ctxt.Arch, 0xd61f0220)
+
+		// 3 nop for place holder
+		plt.AddUint32(ctxt.Arch, 0xd503201f)
+		plt.AddUint32(ctxt.Arch, 0xd503201f)
+		plt.AddUint32(ctxt.Arch, 0xd503201f)
+
+		// check gotplt.size == 0
+		if gotplt.Size != 0 {
+			ld.Errorf(gotplt, "got.plt is not empty at the very beginning")
+		}
+		gotplt.AddAddrPlus(ctxt.Arch, ctxt.Syms.Lookup(".dynamic", 0), 0)
+
+		gotplt.AddUint64(ctxt.Arch, 0)
+		gotplt.AddUint64(ctxt.Arch, 0)
+	}
+}
+
+func addpltsym(ctxt *ld.Link, s *sym.Symbol) {
+	if s.Plt() >= 0 {
+		return
+	}
+
+	ld.Adddynsym(ctxt, s)
+
+	if ctxt.IsELF {
+		plt := ctxt.Syms.Lookup(".plt", 0)
+		gotplt := ctxt.Syms.Lookup(".got.plt", 0)
+		rela := ctxt.Syms.Lookup(".rela.plt", 0)
+		if plt.Size == 0 {
+			elfsetupplt(ctxt)
+		}
+
+		// adrp    x16, &got.plt[0]
+		plt.AddAddrPlus4(gotplt, gotplt.Size)
+		plt.SetUint32(ctxt.Arch, plt.Size-4, 0x90000010)
+		plt.R[len(plt.R)-1].Type = objabi.R_ARM64_GOT
+
+		// <offset> is the offset value of &got.plt[n] to &got.plt[0]
+		// ldr     x17, [x16, <offset>]
+		plt.AddAddrPlus4(gotplt, gotplt.Size)
+		plt.SetUint32(ctxt.Arch, plt.Size-4, 0xf9400211)
+		plt.R[len(plt.R)-1].Type = objabi.R_ARM64_GOT
+
+		// add     x16, x16, <offset>
+		plt.AddAddrPlus4(gotplt, gotplt.Size)
+		plt.SetUint32(ctxt.Arch, plt.Size-4, 0x91000210)
+		plt.R[len(plt.R)-1].Type = objabi.R_ARM64_PCREL
+
+		// br      x17
+		plt.AddUint32(ctxt.Arch, 0xd61f0220)
+
+		// add to got.plt: pointer to plt[0]
+		gotplt.AddAddrPlus(ctxt.Arch, plt, 0)
+
+		// rela
+		rela.AddAddrPlus(ctxt.Arch, gotplt, gotplt.Size-8)
+		rela.AddUint64(ctxt.Arch, ld.ELF64_R_INFO(uint32(s.Dynid), uint32(elf.R_AARCH64_JUMP_SLOT)))
+		rela.AddUint64(ctxt.Arch, 0)
+
+		s.SetPlt(int32(plt.Size - 16))
+	} else {
+		ld.Errorf(s, "addpltsym: unsupported binary format")
+	}
+}
+
+func addgotsym(ctxt *ld.Link, s *sym.Symbol) {
+	if s.Got() >= 0 {
+		return
+	}
+
+	ld.Adddynsym(ctxt, s)
+	got := ctxt.Syms.Lookup(".got", 0)
+	s.SetGot(int32(got.Size))
+	got.AddUint64(ctxt.Arch, 0)
+
+	if ctxt.IsELF {
+		rela := ctxt.Syms.Lookup(".rela", 0)
+		rela.AddAddrPlus(ctxt.Arch, got, int64(s.Got()))
+		rela.AddUint64(ctxt.Arch, ld.ELF64_R_INFO(uint32(s.Dynid), uint32(elf.R_AARCH64_GLOB_DAT)))
+		rela.AddUint64(ctxt.Arch, 0)
+	} else {
+		ld.Errorf(s, "addgotsym: unsupported binary format")
+	}
+}
+
 func asmb(ctxt *ld.Link) {
 	if ctxt.Debugvlog != 0 {
 		ctxt.Logf("%5.2f asmb\n", ld.Cputime())

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

@@ -184,17 +184,16 @@ func mustLinkExternal(ctxt *Link) (res bool, reason string) {
 	}
 
 	// Internally linking cgo is incomplete on some architectures.
-	// https://golang.org/issue/10373
 	// https://golang.org/issue/14449
 	// https://golang.org/issue/21961
-	if iscgo && ctxt.Arch.InFamily(sys.ARM64, sys.MIPS64, sys.MIPS, sys.PPC64) {
+	if iscgo && ctxt.Arch.InFamily(sys.MIPS64, sys.MIPS, sys.PPC64) {
 		return true, objabi.GOARCH + " does not support internal cgo"
 	}
 
 	// When the race flag is set, the LLVM tsan relocatable file is linked
 	// into the final binary, which means external linking is required because
 	// internal linking does not support it.
-	if *flagRace && ctxt.Arch.InFamily(sys.PPC64, sys.ARM64) {
+	if *flagRace && ctxt.Arch.InFamily(sys.PPC64) {
 		return true, "race on " + objabi.GOARCH
 	}
 
@@ -206,7 +205,7 @@ func mustLinkExternal(ctxt *Link) (res bool, reason string) {
 		return true, "buildmode=c-shared"
 	case BuildModePIE:
 		switch objabi.GOOS + "/" + objabi.GOARCH {
-		case "linux/amd64":
+		case "linux/amd64", "linux/arm64":
 		default:
 			// Internal linking does not support TLS_IE.
 			return true, "buildmode=pie"

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

@@ -2492,7 +2492,7 @@ func (ctxt *Link) callgraph() {
 			if r.Sym == nil {
 				continue
 			}
-			if (r.Type == objabi.R_CALL || r.Type == objabi.R_CALLARM || r.Type == objabi.R_CALLPOWER || r.Type == objabi.R_CALLMIPS) && r.Sym.Type == sym.STEXT {
+			if (r.Type == objabi.R_CALL || r.Type == objabi.R_CALLARM || r.Type == objabi.R_CALLARM64 || r.Type == objabi.R_CALLPOWER || r.Type == objabi.R_CALLMIPS) && r.Sym.Type == sym.STEXT {
 				ctxt.Logf("%s calls %s\n", s.Name, r.Sym.Name)
 			}
 		}

src/cmd/link/internal/loadelf/ldelf.go

@@ -794,6 +794,13 @@ func Load(arch *sys.Arch, syms *sym.Symbols, f *bio.Reader, pkg string, length i
 				continue
 			}
 
+			if strings.HasPrefix(elfsym.name, "$d") && elfsym.type_ == 0 && sect.name == ".debug_frame" {
+				// "$d" is a marker, not a real symbol.
+				// This happens with gcc on ARM64.
+				// See https://sourceware.org/bugzilla/show_bug.cgi?id=21809
+				continue
+			}
+
 			if strings.HasPrefix(elfsym.name, ".LASF") { // gcc on s390x does this
 				continue
 			}
@@ -1064,8 +1071,8 @@ func readelfsym(arch *sys.Arch, syms *sym.Symbols, elfobj *ElfObj, i int, elfsym
 			}
 
 		case ElfSymBindLocal:
-			if arch.Family == sys.ARM && (strings.HasPrefix(elfsym.name, "$a") || strings.HasPrefix(elfsym.name, "$d")) {
-				// binutils for arm generate these mapping
+			if (arch.Family == sys.ARM || arch.Family == sys.ARM64) && (strings.HasPrefix(elfsym.name, "$a") || strings.HasPrefix(elfsym.name, "$d") || strings.HasPrefix(elfsym.name, "$x")) {
+				// binutils for arm and arm64 generate these mapping
 				// symbols, ignore these
 				break
 			}
@@ -1127,81 +1134,94 @@ func relSize(arch *sys.Arch, pn string, elftype uint32) (uint8, error) {
 	const (
 		AMD64 = uint32(sys.AMD64)
 		ARM   = uint32(sys.ARM)
+		ARM64 = uint32(sys.ARM64)
 		I386  = uint32(sys.I386)
 		PPC64 = uint32(sys.PPC64)
 		S390X = uint32(sys.S390X)
 	)
 
-	switch uint32(arch.Family) | elftype<<24 {
+	switch uint32(arch.Family) | elftype<<16 {
 	default:
 		return 0, fmt.Errorf("%s: unknown relocation type %d; compiled without -fpic?", pn, elftype)
 
-	case S390X | uint32(elf.R_390_8)<<24:
+	case S390X | uint32(elf.R_390_8)<<16:
 		return 1, nil
 
-	case PPC64 | uint32(elf.R_PPC64_TOC16)<<24,
-		PPC64 | uint32(elf.R_PPC64_TOC16_LO)<<24,
-		PPC64 | uint32(elf.R_PPC64_TOC16_HI)<<24,
-		PPC64 | uint32(elf.R_PPC64_TOC16_HA)<<24,
-		PPC64 | uint32(elf.R_PPC64_TOC16_DS)<<24,
-		PPC64 | uint32(elf.R_PPC64_TOC16_LO_DS)<<24,
-		PPC64 | uint32(elf.R_PPC64_REL16_LO)<<24,
-		PPC64 | uint32(elf.R_PPC64_REL16_HI)<<24,
-		PPC64 | uint32(elf.R_PPC64_REL16_HA)<<24,
-		S390X | uint32(elf.R_390_16)<<24,
-		S390X | uint32(elf.R_390_GOT16)<<24,
-		S390X | uint32(elf.R_390_PC16)<<24,
-		S390X | uint32(elf.R_390_PC16DBL)<<24,
-		S390X | uint32(elf.R_390_PLT16DBL)<<24:
+	case PPC64 | uint32(elf.R_PPC64_TOC16)<<16,
+		PPC64 | uint32(elf.R_PPC64_TOC16_LO)<<16,
+		PPC64 | uint32(elf.R_PPC64_TOC16_HI)<<16,
+		PPC64 | uint32(elf.R_PPC64_TOC16_HA)<<16,
+		PPC64 | uint32(elf.R_PPC64_TOC16_DS)<<16,
+		PPC64 | uint32(elf.R_PPC64_TOC16_LO_DS)<<16,
+		PPC64 | uint32(elf.R_PPC64_REL16_LO)<<16,
+		PPC64 | uint32(elf.R_PPC64_REL16_HI)<<16,
+		PPC64 | uint32(elf.R_PPC64_REL16_HA)<<16,
+		S390X | uint32(elf.R_390_16)<<16,
+		S390X | uint32(elf.R_390_GOT16)<<16,
+		S390X | uint32(elf.R_390_PC16)<<16,
+		S390X | uint32(elf.R_390_PC16DBL)<<16,
+		S390X | uint32(elf.R_390_PLT16DBL)<<16:
 		return 2, nil
 
-	case ARM | uint32(elf.R_ARM_ABS32)<<24,
-		ARM | uint32(elf.R_ARM_GOT32)<<24,
-		ARM | uint32(elf.R_ARM_PLT32)<<24,
-		ARM | uint32(elf.R_ARM_GOTOFF)<<24,
-		ARM | uint32(elf.R_ARM_GOTPC)<<24,
-		ARM | uint32(elf.R_ARM_THM_PC22)<<24,
-		ARM | uint32(elf.R_ARM_REL32)<<24,
-		ARM | uint32(elf.R_ARM_CALL)<<24,
-		ARM | uint32(elf.R_ARM_V4BX)<<24,
-		ARM | uint32(elf.R_ARM_GOT_PREL)<<24,
-		ARM | uint32(elf.R_ARM_PC24)<<24,
-		ARM | uint32(elf.R_ARM_JUMP24)<<24,
-		AMD64 | uint32(elf.R_X86_64_PC32)<<24,
-		AMD64 | uint32(elf.R_X86_64_PLT32)<<24,
-		AMD64 | uint32(elf.R_X86_64_GOTPCREL)<<24,
-		AMD64 | uint32(elf.R_X86_64_GOTPCRELX)<<24,
-		AMD64 | uint32(elf.R_X86_64_REX_GOTPCRELX)<<24,
-		I386 | uint32(elf.R_386_32)<<24,
-		I386 | uint32(elf.R_386_PC32)<<24,
-		I386 | uint32(elf.R_386_GOT32)<<24,
-		I386 | uint32(elf.R_386_PLT32)<<24,
-		I386 | uint32(elf.R_386_GOTOFF)<<24,
-		I386 | uint32(elf.R_386_GOTPC)<<24,
-		I386 | uint32(elf.R_386_GOT32X)<<24,
-		PPC64 | uint32(elf.R_PPC64_REL24)<<24,
-		PPC64 | uint32(elf.R_PPC_REL32)<<24,
-		S390X | uint32(elf.R_390_32)<<24,
-		S390X | uint32(elf.R_390_PC32)<<24,
-		S390X | uint32(elf.R_390_GOT32)<<24,
-		S390X | uint32(elf.R_390_PLT32)<<24,
-		S390X | uint32(elf.R_390_PC32DBL)<<24,
-		S390X | uint32(elf.R_390_PLT32DBL)<<24,
-		S390X | uint32(elf.R_390_GOTPCDBL)<<24,
-		S390X | uint32(elf.R_390_GOTENT)<<24:
+	case ARM | uint32(elf.R_ARM_ABS32)<<16,
+		ARM | uint32(elf.R_ARM_GOT32)<<16,
+		ARM | uint32(elf.R_ARM_PLT32)<<16,
+		ARM | uint32(elf.R_ARM_GOTOFF)<<16,
+		ARM | uint32(elf.R_ARM_GOTPC)<<16,
+		ARM | uint32(elf.R_ARM_THM_PC22)<<16,
+		ARM | uint32(elf.R_ARM_REL32)<<16,
+		ARM | uint32(elf.R_ARM_CALL)<<16,
+		ARM | uint32(elf.R_ARM_V4BX)<<16,
+		ARM | uint32(elf.R_ARM_GOT_PREL)<<16,
+		ARM | uint32(elf.R_ARM_PC24)<<16,
+		ARM | uint32(elf.R_ARM_JUMP24)<<16,
+		ARM64 | uint32(elf.R_AARCH64_CALL26)<<16,
+		ARM64 | uint32(elf.R_AARCH64_ADR_GOT_PAGE)<<16,
+		ARM64 | uint32(elf.R_AARCH64_LD64_GOT_LO12_NC)<<16,
+		ARM64 | uint32(elf.R_AARCH64_ADR_PREL_PG_HI21)<<16,
+		ARM64 | uint32(elf.R_AARCH64_ADD_ABS_LO12_NC)<<16,
+		ARM64 | uint32(elf.R_AARCH64_LDST8_ABS_LO12_NC)<<16,
+		ARM64 | uint32(elf.R_AARCH64_LDST32_ABS_LO12_NC)<<16,
+		ARM64 | uint32(elf.R_AARCH64_LDST64_ABS_LO12_NC)<<16,
+		ARM64 | uint32(elf.R_AARCH64_PREL32)<<16,
+		ARM64 | uint32(elf.R_AARCH64_JUMP26)<<16,
+		AMD64 | uint32(elf.R_X86_64_PC32)<<16,
+		AMD64 | uint32(elf.R_X86_64_PLT32)<<16,
+		AMD64 | uint32(elf.R_X86_64_GOTPCREL)<<16,
+		AMD64 | uint32(elf.R_X86_64_GOTPCRELX)<<16,
+		AMD64 | uint32(elf.R_X86_64_REX_GOTPCRELX)<<16,
+		I386 | uint32(elf.R_386_32)<<16,
+		I386 | uint32(elf.R_386_PC32)<<16,
+		I386 | uint32(elf.R_386_GOT32)<<16,
+		I386 | uint32(elf.R_386_PLT32)<<16,
+		I386 | uint32(elf.R_386_GOTOFF)<<16,
+		I386 | uint32(elf.R_386_GOTPC)<<16,
+		I386 | uint32(elf.R_386_GOT32X)<<16,
+		PPC64 | uint32(elf.R_PPC64_REL24)<<16,
+		PPC64 | uint32(elf.R_PPC_REL32)<<16,
+		S390X | uint32(elf.R_390_32)<<16,
+		S390X | uint32(elf.R_390_PC32)<<16,
+		S390X | uint32(elf.R_390_GOT32)<<16,
+		S390X | uint32(elf.R_390_PLT32)<<16,
+		S390X | uint32(elf.R_390_PC32DBL)<<16,
+		S390X | uint32(elf.R_390_PLT32DBL)<<16,
+		S390X | uint32(elf.R_390_GOTPCDBL)<<16,
+		S390X | uint32(elf.R_390_GOTENT)<<16:
 		return 4, nil
 
-	case AMD64 | uint32(elf.R_X86_64_64)<<24,
-		AMD64 | uint32(elf.R_X86_64_PC64)<<24,
-		PPC64 | uint32(elf.R_PPC64_ADDR64)<<24,
-		S390X | uint32(elf.R_390_GLOB_DAT)<<24,
-		S390X | uint32(elf.R_390_RELATIVE)<<24,
-		S390X | uint32(elf.R_390_GOTOFF)<<24,
-		S390X | uint32(elf.R_390_GOTPC)<<24,
-		S390X | uint32(elf.R_390_64)<<24,
-		S390X | uint32(elf.R_390_PC64)<<24,
-		S390X | uint32(elf.R_390_GOT64)<<24,
-		S390X | uint32(elf.R_390_PLT64)<<24:
+	case AMD64 | uint32(elf.R_X86_64_64)<<16,
+		AMD64 | uint32(elf.R_X86_64_PC64)<<16,
+		ARM64 | uint32(elf.R_AARCH64_ABS64)<<16,
+		ARM64 | uint32(elf.R_AARCH64_PREL64)<<16,
+		PPC64 | uint32(elf.R_PPC64_ADDR64)<<16,
+		S390X | uint32(elf.R_390_GLOB_DAT)<<16,
+		S390X | uint32(elf.R_390_RELATIVE)<<16,
+		S390X | uint32(elf.R_390_GOTOFF)<<16,
+		S390X | uint32(elf.R_390_GOTPC)<<16,
+		S390X | uint32(elf.R_390_64)<<16,
+		S390X | uint32(elf.R_390_PC64)<<16,
+		S390X | uint32(elf.R_390_GOT64)<<16,
+		S390X | uint32(elf.R_390_PLT64)<<16:
 		return 8, nil
 	}
 }