cmd/internal/obj, cmd/internal/ld, cmd/7l: external linking for darwin/arm64

Change-Id: I3b3f80791a1db4c2b7318f81a115972cd2237f02
Signed-off-by: Shenghou Ma <minux@golang.org>
Reviewed-on: https://go-review.googlesource.com/8781Reviewed-by: David Crawshaw <crawshaw@golang.org>

src/cmd/7l/asm.go

@@ -109,7 +109,82 @@ func elfsetupplt() {
 }
 
 func machoreloc1(r *ld.Reloc, sectoff int64) int {
+	var v uint32
+
+	rs := r.Xsym
+
+	// ld64 has a bug handling MACHO_ARM64_RELOC_UNSIGNED with !extern relocation.
+	// see cmd/internal/ld/data.go for details. The workarond is that don't use !extern
+	// UNSIGNED relocation at all.
+	if rs.Type == ld.SHOSTOBJ || r.Type == ld.R_CALLARM64 || r.Type == ld.R_ADDRARM64 || r.Type == ld.R_ADDR {
+		if rs.Dynid < 0 {
+			ld.Diag("reloc %d to non-macho symbol %s type=%d", r.Type, rs.Name, rs.Type)
+			return -1
+		}
+
+		v = uint32(rs.Dynid)
+		v |= 1 << 27 // external relocation
+	} else {
+		v = uint32((rs.Sect.(*ld.Section)).Extnum)
+		if v == 0 {
+			ld.Diag("reloc %d to symbol %s in non-macho section %s type=%d", r.Type, rs.Name, (rs.Sect.(*ld.Section)).Name, rs.Type)
+			return -1
+		}
+	}
+
+	switch r.Type {
+	default:
 		return -1
+
+	case ld.R_ADDR:
+		v |= ld.MACHO_ARM64_RELOC_UNSIGNED << 28
+
+	case ld.R_CALLARM64:
+		if r.Xadd != 0 {
+			ld.Diag("ld64 doesn't allow BR26 reloc with non-zero addend: %s+%d", rs.Name, r.Xadd)
+		}
+
+		v |= 1 << 24 // pc-relative bit
+		v |= ld.MACHO_ARM64_RELOC_BRANCH26 << 28
+
+	case ld.R_ADDRARM64:
+		r.Siz = 4
+		// Two relocation entries: MACHO_ARM64_RELOC_PAGEOFF12 MACHO_ARM64_RELOC_PAGE21
+		// if r.Xadd is non-zero, add two MACHO_ARM64_RELOC_ADDEND.
+		if r.Xadd != 0 {
+			ld.Thearch.Lput(uint32(sectoff + 4))
+			ld.Thearch.Lput((ld.MACHO_ARM64_RELOC_ADDEND << 28) | (2 << 25) | uint32(r.Xadd&0xffffff))
+		}
+		ld.Thearch.Lput(uint32(sectoff + 4))
+		ld.Thearch.Lput(v | (ld.MACHO_ARM64_RELOC_PAGEOFF12 << 28) | (2 << 25))
+		if r.Xadd != 0 {
+			ld.Thearch.Lput(uint32(sectoff))
+			ld.Thearch.Lput((ld.MACHO_ARM64_RELOC_ADDEND << 28) | (2 << 25) | uint32(r.Xadd&0xffffff))
+		}
+		v |= 1 << 24 // pc-relative bit
+		v |= ld.MACHO_ARM64_RELOC_PAGE21 << 28
+	}
+
+	switch r.Siz {
+	default:
+		return -1
+
+	case 1:
+		v |= 0 << 25
+
+	case 2:
+		v |= 1 << 25
+
+	case 4:
+		v |= 2 << 25
+
+	case 8:
+		v |= 3 << 25
+	}
+
+	ld.Thearch.Lput(uint32(sectoff))
+	ld.Thearch.Lput(v)
+	return 0
 }
 
 func archreloc(r *ld.Reloc, s *ld.LSym, val *int64) int {
@@ -121,18 +196,6 @@ func archreloc(r *ld.Reloc, s *ld.LSym, val *int64) int {
 		case ld.R_ADDRARM64:
 			r.Done = 0
 
-			// the first instruction is always at the lower address, this is endian neutral;
-			// but note that o0 and o1 should still use the target endian.
-			o0 := ld.Thelinkarch.ByteOrder.Uint32(s.P[r.Off : r.Off+4])
-			o1 := ld.Thelinkarch.ByteOrder.Uint32(s.P[r.Off+4 : r.Off+8])
-
-			// when laid out, the instruction order must always be o1, o2.
-			if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
-				*val = int64(o0)<<32 | int64(o1)
-			} else {
-				*val = int64(o1)<<32 | int64(o0)
-			}
-
 			// set up addend for eventual relocation via outer symbol.
 			rs := r.Sym
 			r.Xadd = r.Add
@@ -146,6 +209,34 @@ func archreloc(r *ld.Reloc, s *ld.LSym, val *int64) int {
 			}
 			r.Xsym = rs
 
+			// the first instruction is always at the lower address, this is endian neutral;
+			// but note that o0 and o1 should still use the target endian.
+			o0 := ld.Thelinkarch.ByteOrder.Uint32(s.P[r.Off : r.Off+4])
+			o1 := ld.Thelinkarch.ByteOrder.Uint32(s.P[r.Off+4 : r.Off+8])
+
+			// Note: ld64 currently has a bug that any non-zero addend for BR26 relocation
+			// will make the linking fail because it thinks the code is not PIC even though
+			// the BR26 relocation should be fully resolved at link time.
+			// That is the reason why the next if block is disabled. When the bug in ld64
+			// is fixed, we can enable this block and also enable duff's device in cmd/7g.
+			if false && ld.HEADTYPE == ld.Hdarwin {
+				// Mach-O wants the addend to be encoded in the instruction
+				// Note that although Mach-O supports ARM64_RELOC_ADDEND, it
+				// can only encode 24-bit of signed addend, but the instructions
+				// supports 33-bit of signed addend, so we always encode the
+				// addend in place.
+				o0 |= (uint32((r.Xadd>>12)&3) << 29) | (uint32((r.Xadd>>12>>2)&0x7ffff) << 5)
+				o1 |= uint32(r.Xadd&0xfff) << 10
+				r.Xadd = 0
+			}
+
+			// when laid out, the instruction order must always be o1, o2.
+			if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
+				*val = int64(o0)<<32 | int64(o1)
+			} else {
+				*val = int64(o1)<<32 | int64(o0)
+			}
+
 			return 0
 
 		case ld.R_CALLARM64:
@@ -217,6 +308,8 @@ func adddynlib(lib string) {
 			ld.Addstring(s, "")
 		}
 		ld.Elfwritedynent(ld.Linklookup(ld.Ctxt, ".dynamic", 0), ld.DT_NEEDED, uint64(ld.Addstring(s, lib)))
+	} else if ld.HEADTYPE == ld.Hdarwin {
+		ld.Machoadddynlib(lib)
 	} else {
 		ld.Diag("adddynlib: unsupported binary format")
 	}
@@ -258,6 +351,24 @@ func asmb() {
 	ld.Cseek(int64(ld.Segdata.Fileoff))
 	ld.Datblk(int64(ld.Segdata.Vaddr), int64(ld.Segdata.Filelen))
 
+	machlink := uint32(0)
+	if ld.HEADTYPE == ld.Hdarwin {
+		if ld.Debug['v'] != 0 {
+			fmt.Fprintf(&ld.Bso, "%5.2f dwarf\n", obj.Cputime())
+		}
+
+		if ld.Debug['w'] == 0 { // TODO(minux): enable DWARF Support
+			dwarfoff := uint32(ld.Rnd(int64(uint64(ld.HEADR)+ld.Segtext.Length), int64(ld.INITRND)) + ld.Rnd(int64(ld.Segdata.Filelen), int64(ld.INITRND)))
+			ld.Cseek(int64(dwarfoff))
+
+			ld.Segdwarf.Fileoff = uint64(ld.Cpos())
+			ld.Dwarfemitdebugsections()
+			ld.Segdwarf.Filelen = uint64(ld.Cpos()) - ld.Segdwarf.Fileoff
+		}
+
+		machlink = uint32(ld.Domacholink())
+	}
+
 	/* output symbol table */
 	ld.Symsize = 0
 
@@ -278,6 +389,9 @@ func asmb() {
 
 		case ld.Hplan9:
 			symo = uint32(ld.Segdata.Fileoff + ld.Segdata.Filelen)
+
+		case ld.Hdarwin:
+			symo = uint32(ld.Rnd(int64(uint64(ld.HEADR)+ld.Segtext.Filelen), int64(ld.INITRND)) + ld.Rnd(int64(ld.Segdata.Filelen), int64(ld.INITRND)) + int64(machlink))
 		}
 
 		ld.Cseek(int64(symo))
@@ -314,6 +428,11 @@ func asmb() {
 
 				ld.Cflush()
 			}
+
+		case ld.Hdarwin:
+			if ld.Linkmode == ld.LinkExternal {
+				ld.Machoemitreloc()
+			}
 		}
 	}
 
@@ -341,6 +460,9 @@ func asmb() {
 		ld.Hopenbsd,
 		ld.Hnacl:
 		ld.Asmbelf(int64(symo))
+
+	case ld.Hdarwin:
+		ld.Asmbmacho()
 	}
 
 	ld.Cflush()

src/cmd/7l/obj.go

@@ -88,6 +88,11 @@ func archinit() {
 		ld.Linkmode = ld.LinkInternal
 	}
 
+	// Darwin/arm64 only supports external linking
+	if ld.HEADTYPE == ld.Hdarwin {
+		ld.Linkmode = ld.LinkExternal
+	}
+
 	switch ld.HEADTYPE {
 	default:
 		if ld.Linkmode == ld.LinkAuto {
@@ -96,7 +101,7 @@ func archinit() {
 		if ld.Linkmode == ld.LinkExternal && obj.Getgoextlinkenabled() != "1" {
 			log.Fatalf("cannot use -linkmode=external with -H %s", ld.Headstr(int(ld.HEADTYPE)))
 		}
-	case ld.Hlinux:
+	case ld.Hlinux, ld.Hdarwin:
 		break
 	}
 
@@ -132,6 +137,20 @@ func archinit() {
 			ld.INITRND = 0x10000
 		}
 
+	case ld.Hdarwin: /* apple MACH */
+		ld.Debug['w'] = 1 // disable DWARF generation
+		ld.Machoinit()
+		ld.HEADR = ld.INITIAL_MACHO_HEADR
+		if ld.INITTEXT == -1 {
+			ld.INITTEXT = 4096 + int64(ld.HEADR)
+		}
+		if ld.INITDAT == -1 {
+			ld.INITDAT = 0
+		}
+		if ld.INITRND == -1 {
+			ld.INITRND = 4096
+		}
+
 	case ld.Hnacl:
 		ld.Elfinit()
 		ld.HEADR = 0x10000

src/cmd/internal/ld/data.go

@@ -451,9 +451,19 @@ func relocsym(s *LSym) {
 						o = 0
 					}
 				} else if HEADTYPE == Hdarwin {
+					// ld64 for arm64 has a bug where if the address pointed to by o exists in the
+					// symbol table (dynid >= 0), or is inside a symbol that exists in the symbol
+					// table, then it will add o twice into the relocated value.
+					// The workaround is that on arm64 don't ever add symaddr to o and always use
+					// extern relocation by requiring rs->dynid >= 0.
 					if rs.Type != SHOSTOBJ {
+						if Thearch.Thechar == '7' && rs.Dynid < 0 {
+							Diag("R_ADDR reloc to %s+%d is not supported on darwin/arm64", rs.Name, o)
+						}
+						if Thearch.Thechar != '7' {
 							o += Symaddr(rs)
 						}
+					}
 				} else if HEADTYPE == Hwindows {
 					// nothing to do
 				} else {

src/cmd/internal/ld/lib.go

@@ -475,7 +475,7 @@ func loadlib() {
 		// dependency problems when compiling natively (external linking requires
 		// runtime/cgo, runtime/cgo requires cmd/cgo, but cmd/cgo needs to be
 		// compiled using external linking.)
-		if Thearch.Thechar == '5' && HEADTYPE == Hdarwin && iscgo {
+		if (Thearch.Thechar == '5' || Thearch.Thechar == '7') && HEADTYPE == Hdarwin && iscgo {
 			Linkmode = LinkExternal
 		}
 	}
@@ -1599,6 +1599,7 @@ func genasmsym(put func(*LSym, string, int, int64, int64, int, *LSym)) {
 			STYPE,
 			SSTRING,
 			SGOSTRING,
+			SGOFUNC,
 			SWINDOWS:
 			if !s.Reachable {
 				continue

src/cmd/internal/ld/macho.go

@@ -63,6 +63,8 @@ const (
 	MACHO_CPU_ARM                 = 12
 	MACHO_SUBCPU_ARM              = 0
 	MACHO_SUBCPU_ARMV7            = 9
+	MACHO_CPU_ARM64               = 1<<24 | 12
+	MACHO_SUBCPU_ARM64_ALL        = 0
 	MACHO32SYMSIZE                = 12
 	MACHO64SYMSIZE                = 16
 	MACHO_X86_64_RELOC_UNSIGNED   = 0
@@ -76,6 +78,11 @@ const (
 	MACHO_X86_64_RELOC_SIGNED_4   = 8
 	MACHO_ARM_RELOC_VANILLA       = 0
 	MACHO_ARM_RELOC_BR24          = 5
+	MACHO_ARM64_RELOC_UNSIGNED    = 0
+	MACHO_ARM64_RELOC_BRANCH26    = 2
+	MACHO_ARM64_RELOC_PAGE21      = 3
+	MACHO_ARM64_RELOC_PAGEOFF12   = 4
+	MACHO_ARM64_RELOC_ADDEND      = 10
 	MACHO_GENERIC_RELOC_VANILLA   = 0
 	MACHO_FAKE_GOTPCREL           = 100
 )
@@ -125,7 +132,7 @@ var load_budget int = INITIAL_MACHO_HEADR - 2*1024
 func Machoinit() {
 	switch Thearch.Thechar {
 	// 64-bit architectures
-	case '6', '9':
+	case '6', '7', '9':
 		macho64 = true
 
 		// 32-bit architectures
@@ -349,7 +356,15 @@ func Machoadddynlib(lib string) {
 func machoshbits(mseg *MachoSeg, sect *Section, segname string) {
 	buf := "__" + strings.Replace(sect.Name[1:], ".", "_", -1)
 
-	msect := newMachoSect(mseg, buf, segname)
+	var msect *MachoSect
+	if Thearch.Thechar == '7' && sect.Rwx&1 == 0 {
+		// darwin/arm64 forbids absolute relocs in __TEXT, so if
+		// the section is not executable, put it in __DATA segment.
+		msect = newMachoSect(mseg, buf, "__DATA")
+	} else {
+		msect = newMachoSect(mseg, buf, segname)
+	}
+
 	if sect.Rellen > 0 {
 		msect.reloc = uint32(sect.Reloff)
 		msect.nreloc = uint32(sect.Rellen / 8)
@@ -416,6 +431,10 @@ func Asmbmacho() {
 		mh.cpu = MACHO_CPU_AMD64
 		mh.subcpu = MACHO_SUBCPU_X86
 
+	case '7':
+		mh.cpu = MACHO_CPU_ARM64
+		mh.subcpu = MACHO_SUBCPU_ARM64_ALL
+
 	case '8':
 		mh.cpu = MACHO_CPU_386
 		mh.subcpu = MACHO_SUBCPU_X86
@@ -487,7 +506,14 @@ func Asmbmacho() {
 			ml.data[0] = 4                       /* thread type */
 			ml.data[1] = 42                      /* word count */
 			ml.data[2+32] = uint32(Entryvalue()) /* start pc */
-			ml.data[2+32+1] = uint32(Entryvalue() >> 16 >> 16) // hide >>32 for 8l
+			ml.data[2+32+1] = uint32(Entryvalue() >> 32)
+
+		case '7':
+			ml := newMachoLoad(5, 68+2)          /* unix thread */
+			ml.data[0] = 6                       /* thread type */
+			ml.data[1] = 68                      /* word count */
+			ml.data[2+64] = uint32(Entryvalue()) /* start pc */
+			ml.data[2+64+1] = uint32(Entryvalue() >> 32)
 
 		case '8':
 			ml := newMachoLoad(5, 16+2)          /* unix thread */

src/cmd/internal/ld/sym.go

@@ -132,14 +132,17 @@ func linknew(arch *LinkArch) *Link {
 		default:
 			log.Fatalf("unknown thread-local storage offset for darwin/%s", ctxt.Arch.Name)
 
+		case '5':
+			ctxt.Tlsoffset = 0 // dummy value, not needed
+
 		case '6':
 			ctxt.Tlsoffset = 0x8a0
 
+		case '7':
+			ctxt.Tlsoffset = 0 // dummy value, not needed
+
 		case '8':
 			ctxt.Tlsoffset = 0x468
-
-		case '5':
-			ctxt.Tlsoffset = 0 // dummy value, not needed
 		}
 	}
 

src/cmd/internal/obj/sym.go

@@ -158,14 +158,17 @@ func Linknew(arch *LinkArch) *Link {
 		default:
 			log.Fatalf("unknown thread-local storage offset for darwin/%s", ctxt.Arch.Name)
 
+		case '5':
+			ctxt.Tlsoffset = 0 // dummy value, not needed
+
 		case '6':
 			ctxt.Tlsoffset = 0x8a0
 
+		case '7':
+			ctxt.Tlsoffset = 0 // dummy value, not needed
+
 		case '8':
 			ctxt.Tlsoffset = 0x468
-
-		case '5':
-			ctxt.Tlsoffset = 0 // dummy value, not needed
 		}
 	}