func archreloc(r *ld.Reloc, s *ld.LSym, val *int64) int {
if ld.Linkmode == ld.LinkExternal {
// TODO(minux): translate R_ADDRPOWER and R_CALLPOWER into standard ELF relocations.
// R_ADDRPOWER corresponds to R_PPC_ADDR16_HA and R_PPC_ADDR16_LO.
// R_CALLPOWER corresponds to R_PPC_REL24.
return -1
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(r.Sym) + r.Add - ld.Symaddr(ld.Linklookup(ld.Ctxt, ".got", 0))
return 0
case obj.R_ADDRPOWER, obj.R_ADDRPOWER_DS:
return archrelocaddr(r, s, val)
case obj.R_CALLPOWER:
// Bits 6 through 29 = (S + A - P) >> 2
t := ld.Symaddr(r.Sym) + r.Add - (s.Value + int64(r.Off))
if t&3 != 0 {
ld.Ctxt.Diag("relocation for %s+%d is not aligned: %d", r.Sym.Name, r.Off, t)
}
if int64(int32(t<<6)>>6) != t {
// TODO(austin) This can happen if text > 32M.
// Add a call trampoline to .text in that case.
ld.Ctxt.Diag("relocation for %s+%d is too big: %d", r.Sym.Name, r.Off, t)
}
*val |= int64(uint32(t) &^ 0xfc000003)
return 0
case obj.R_POWER_TOC: // S + A - .TOC.
*val = ld.Symaddr(r.Sym) + r.Add - symtoc(s)
return 0
case obj.R_POWER_TLS_LE:
// The thread pointer points 0x7000 bytes after the start of the the
// thread local storage area as documented in section "3.7.2 TLS
// Runtime Handling" of "Power Architecture 64-Bit ELF V2 ABI
// Specification".
v := r.Sym.Value - 0x7000
if int64(int16(v)) != v {
ld.Diag("TLS offset out of range %d", v)
}
*val = (*val &^ 0xffff) | (v & 0xffff)
return 0
}
return -1
}
// Convert the direct jump relocation r to refer to a trampoline if the target is too far
func trampoline(ctxt *ld.Link, r *ld.Reloc, s *ld.Symbol) {
switch r.Type {
case obj.R_CALLARM:
// r.Add is the instruction
// low 24-bit encodes the target address
t := (ld.Symaddr(r.Sym) + int64(signext24(r.Add&0xffffff)*4) - (s.Value + int64(r.Off))) / 4
if t > 0x7fffff || t < -0x800000 || (*ld.FlagDebugTramp > 1 && s.File != r.Sym.File) {
// direct call too far, need to insert trampoline.
// look up existing trampolines first. if we found one within the range
// of direct call, we can reuse it. otherwise create a new one.
offset := (signext24(r.Add&0xffffff) + 2) * 4
var tramp *ld.Symbol
for i := 0; ; i++ {
name := r.Sym.Name + fmt.Sprintf("%+d-tramp%d", offset, i)
tramp = ctxt.Syms.Lookup(name, int(r.Sym.Version))
if tramp.Type == obj.SDYNIMPORT {
// don't reuse trampoline defined in other module
continue
}
if tramp.Value == 0 {
// either the trampoline does not exist -- we need to create one,
// or found one the address which is not assigned -- this will be
// laid down immediately after the current function. use this one.
break
}
t = (ld.Symaddr(tramp) - 8 - (s.Value + int64(r.Off))) / 4
if t >= -0x800000 && t < 0x7fffff {
// found an existing trampoline that is not too far
// we can just use it
break
}
}
if tramp.Type == 0 {
// trampoline does not exist, create one
ctxt.AddTramp(tramp)
if ctxt.DynlinkingGo() {
if immrot(uint32(offset)) == 0 {
ld.Errorf(s, "odd offset in dynlink direct call: %v+%d", r.Sym, offset)
}
gentrampdyn(tramp, r.Sym, int64(offset))
} else if ld.Buildmode == ld.BuildmodeCArchive || ld.Buildmode == ld.BuildmodeCShared || ld.Buildmode == ld.BuildmodePIE {
gentramppic(tramp, r.Sym, int64(offset))
} else {
gentramp(tramp, r.Sym, int64(offset))
}
}
// modify reloc to point to tramp, which will be resolved later
r.Sym = tramp
r.Add = r.Add&0xff000000 | 0xfffffe // clear the offset embedded in the instruction
r.Done = 0
}
default:
ld.Errorf(s, "trampoline called with non-jump reloc: %v", r.Type)
}
}
// Convert the direct jump relocation r to refer to a trampoline if the target is too far
func trampoline(ctxt *ld.Link, r *ld.Reloc, s *ld.Symbol) {
switch r.Type {
case obj.R_CALLARM:
// r.Add is the instruction
// low 24-bit encodes the target address
t := (ld.Symaddr(r.Sym) + int64(signext24(r.Add&0xffffff)*4) - (s.Value + int64(r.Off))) / 4
if t > 0x7fffff || t < -0x800000 || (*ld.FlagDebugTramp > 1 && s.File != r.Sym.File) {
// direct call too far, need to insert trampoline
offset := (signext24(r.Add&0xffffff) + 2) * 4
var tramp *ld.Symbol
for i := 0; ; i++ {
name := r.Sym.Name + fmt.Sprintf("%+d-tramp%d", offset, i)
tramp = ctxt.Syms.Lookup(name, int(r.Sym.Version))
if tramp.Value == 0 {
// either the trampoline does not exist -- we need to create one,
// or found one the address which is not assigned -- this will be
// laid down immediately after the current function. use this one.
break
}
t = (ld.Symaddr(tramp) - 8 - (s.Value + int64(r.Off))) / 4
if t >= -0x800000 && t < 0x7fffff {
// found an existing trampoline that is not too far
// we can just use it
break
}
}
if tramp.Type == 0 {
// trampoline does not exist, create one
ctxt.AddTramp(tramp)
tramp.Size = 12 // 3 instructions
tramp.P = make([]byte, tramp.Size)
t = ld.Symaddr(r.Sym) + int64(offset)
o1 := uint32(0xe5900000 | 11<<12 | 15<<16) // MOVW (R15), R11 // R15 is actual pc + 8
o2 := uint32(0xe12fff10 | 11) // JMP (R11)
o3 := uint32(t) // WORD $target
ld.SysArch.ByteOrder.PutUint32(tramp.P, o1)
ld.SysArch.ByteOrder.PutUint32(tramp.P[4:], o2)
ld.SysArch.ByteOrder.PutUint32(tramp.P[8:], o3)
}
// modify reloc to point to tramp, which will be resolved later
r.Sym = tramp
r.Add = r.Add&0xff000000 | 0xfffffe // clear the offset embedded in the instruction
r.Done = 0
}
default:
ld.Errorf(s, "trampoline called with non-jump reloc: %v", r.Type)
}
}
func archrelocaddr(r *ld.Reloc, s *ld.Symbol, val *int64) int {
var o1, o2 uint32
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
o1 = uint32(*val >> 32)
o2 = uint32(*val)
} else {
o1 = uint32(*val)
o2 = uint32(*val >> 32)
}
// We are spreading a 31-bit address across two instructions, putting the
// high (adjusted) part in the low 16 bits of the first instruction and the
// low part in the low 16 bits of the second instruction, or, in the DS case,
// bits 15-2 (inclusive) of the address into bits 15-2 of the second
// instruction (it is an error in this case if the low 2 bits of the address
// are non-zero).
t := ld.Symaddr(r.Sym) + r.Add
if t < 0 || t >= 1<<31 {
ld.Ctxt.Diag("relocation for %s is too big (>=2G): %d", s.Name, ld.Symaddr(r.Sym))
}
if t&0x8000 != 0 {
t += 0x10000
}
switch r.Type {
case obj.R_ADDRPOWER:
o1 |= (uint32(t) >> 16) & 0xffff
o2 |= uint32(t) & 0xffff
case obj.R_ADDRPOWER_DS:
o1 |= (uint32(t) >> 16) & 0xffff
if t&3 != 0 {
ld.Ctxt.Diag("bad DS reloc for %s: %d", s.Name, ld.Symaddr(r.Sym))
}
o2 |= uint32(t) & 0xfffc
default:
return -1
}
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
*val = int64(o1)<<32 | int64(o2)
} else {
*val = int64(o2)<<32 | int64(o1)
}
return 0
}
func archreloc(r *ld.Reloc, s *ld.LSym, val *int64) int {
if ld.Linkmode == ld.LinkExternal {
return -1
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(r.Sym) + r.Add - ld.Symaddr(ld.Linklookup(ld.Ctxt, ".got", 0))
return 0
}
return -1
}
func archreloc(ctxt *ld.Link, r *ld.Reloc, s *ld.Symbol, val *int64) int {
if ld.Linkmode == ld.LinkExternal {
return -1
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(r.Sym) + r.Add - ld.Symaddr(ctxt.Syms.Lookup(".got", 0))
return 0
}
return -1
}
func archreloc(r *ld.Reloc, s *ld.LSym, val *int64) int {
if ld.Linkmode == ld.LinkExternal {
return -1
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(r.Sym) + r.Add - ld.Symaddr(ld.Linklookup(ld.Ctxt, ".got", 0))
return 0
case obj.R_ADDRMIPS:
t := ld.Symaddr(r.Sym) + r.Add
if t >= 1<<32 || t < -1<<32 {
ld.Diag("program too large, address relocation = %v", t)
}
// the first instruction is always at the lower address, this is endian neutral;
// but note that o1 and o2 should still use the target endian.
o1 := ld.Thelinkarch.ByteOrder.Uint32(s.P[r.Off:])
o2 := ld.Thelinkarch.ByteOrder.Uint32(s.P[r.Off+4:])
o1 = o1&0xffff0000 | uint32(t>>16)&0xffff
o2 = o2&0xffff0000 | uint32(t)&0xffff
// when laid out, the instruction order must always be o1, o2.
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
*val = int64(o1)<<32 | int64(o2)
} else {
*val = int64(o2)<<32 | int64(o1)
}
return 0
case obj.R_CALLMIPS,
obj.R_JMPMIPS:
// Low 26 bits = (S + A) >> 2
t := ld.Symaddr(r.Sym) + r.Add
o1 := ld.Thelinkarch.ByteOrder.Uint32(s.P[r.Off:])
*val = int64(o1&0xfc000000 | uint32(t>>2)&^0xfc000000)
return 0
}
return -1
}
func archrelocaddr(r *ld.Reloc, s *ld.LSym, val *int64) int {
var o1, o2 uint32
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
o1 = uint32(*val >> 32)
o2 = uint32(*val)
} else {
o1 = uint32(*val)
o2 = uint32(*val >> 32)
}
// We are inserting an address into two instructions: adrp and
// then either addi or a load.
address := ld.Symaddr(r.Sym) + r.Add
pgaddress := (address &^ 0xfff) - ((s.Value + int64(r.Off)) &^ 0xfff)
if pgaddress < -1<<31 || pgaddress >= 1<<31 {
ld.Ctxt.Diag("relocation for %s is too big (>=2G): %d", s.Name, pgaddress)
}
pgoff := uint32(address & 0xfff)
o1 |= uint32((((pgaddress >> 12) & 3) << 29) | (((pgaddress >> 12 >> 2) & 0x7ffff) << 5))
switch r.Type {
case obj.R_ADDRARM64, obj.R_ARM64_LOAD8:
o2 |= pgoff << 10
case obj.R_ARM64_LOAD16:
if pgoff&0x1 != 0 {
ld.Diag("offset for 16-byte load/store has unaligned value %d", pgoff)
}
o2 |= pgoff << 9
case obj.R_ARM64_LOAD32:
if pgoff&0x3 != 0 {
ld.Diag("offset for 32-byte load/store has unaligned value %d", pgoff)
}
o2 |= pgoff << 8
case obj.R_ARM64_LOAD64:
if pgoff&0x7 != 0 {
ld.Diag("offset for 64-byte load/store has unaligned value %d", pgoff)
}
o2 |= pgoff << 7
default:
return -1
}
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
*val = int64(o1)<<32 | int64(o2)
} else {
*val = int64(o2)<<32 | int64(o1)
}
return 0
}
func archreloc(ctxt *ld.Link, r *ld.Reloc, s *ld.Symbol, val *int64) int {
if ld.Linkmode == ld.LinkExternal {
return -1
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(r.Sym) + r.Add - ld.Symaddr(ctxt.Syms.Lookup(".got", 0))
return 0
case obj.R_ADDRMIPS,
obj.R_ADDRMIPSU:
t := ld.Symaddr(r.Sym) + r.Add
o1 := ld.SysArch.ByteOrder.Uint32(s.P[r.Off:])
if r.Type == obj.R_ADDRMIPS {
*val = int64(o1&0xffff0000 | uint32(t)&0xffff)
} else {
*val = int64(o1&0xffff0000 | uint32((t+1<<15)>>16)&0xffff)
}
return 0
case obj.R_CALLMIPS,
obj.R_JMPMIPS:
// Low 26 bits = (S + A) >> 2
t := ld.Symaddr(r.Sym) + r.Add
o1 := ld.SysArch.ByteOrder.Uint32(s.P[r.Off:])
*val = int64(o1&0xfc000000 | uint32(t>>2)&^0xfc000000)
return 0
}
return -1
}
// generate a trampoline to target+offset
func gentramp(tramp, target *ld.Symbol, offset int64) {
tramp.Size = 12 // 3 instructions
tramp.P = make([]byte, tramp.Size)
t := ld.Symaddr(target) + int64(offset)
o1 := uint32(0xe5900000 | 11<<12 | 15<<16) // MOVW (R15), R11 // R15 is actual pc + 8
o2 := uint32(0xe12fff10 | 11) // JMP (R11)
o3 := uint32(t) // WORD $target
ld.SysArch.ByteOrder.PutUint32(tramp.P, o1)
ld.SysArch.ByteOrder.PutUint32(tramp.P[4:], o2)
ld.SysArch.ByteOrder.PutUint32(tramp.P[8:], o3)
if ld.Linkmode == ld.LinkExternal {
r := ld.Addrel(tramp)
r.Off = 8
r.Type = obj.R_ADDR
r.Siz = 4
r.Sym = target
r.Add = offset
}
}
func archreloc(r *ld.Reloc, s *ld.Symbol, val *int64) int {
if ld.Linkmode == ld.LinkExternal {
switch r.Type {
default:
return -1
case obj.R_POWER_TLS, obj.R_POWER_TLS_LE, obj.R_POWER_TLS_IE:
r.Done = 0
// check Outer is nil, Type is TLSBSS?
r.Xadd = r.Add
r.Xsym = r.Sym
return 0
case obj.R_ADDRPOWER,
obj.R_ADDRPOWER_DS,
obj.R_ADDRPOWER_TOCREL,
obj.R_ADDRPOWER_TOCREL_DS,
obj.R_ADDRPOWER_GOT,
obj.R_ADDRPOWER_PCREL:
r.Done = 0
// set up addend for eventual relocation via outer symbol.
rs := r.Sym
r.Xadd = r.Add
for rs.Outer != nil {
r.Xadd += ld.Symaddr(rs) - ld.Symaddr(rs.Outer)
rs = rs.Outer
}
if rs.Type != obj.SHOSTOBJ && rs.Type != obj.SDYNIMPORT && rs.Sect == nil {
ld.Diag("missing section for %s", rs.Name)
}
r.Xsym = rs
return 0
case obj.R_CALLPOWER:
r.Done = 0
r.Xsym = r.Sym
r.Xadd = r.Add
return 0
}
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(r.Sym) + r.Add - ld.Symaddr(ld.Linklookup(ld.Ctxt, ".got", 0))
return 0
case obj.R_ADDRPOWER, obj.R_ADDRPOWER_DS:
return archrelocaddr(r, s, val)
case obj.R_CALLPOWER:
// Bits 6 through 29 = (S + A - P) >> 2
t := ld.Symaddr(r.Sym) + r.Add - (s.Value + int64(r.Off))
if t&3 != 0 {
ld.Ctxt.Diag("relocation for %s+%d is not aligned: %d", r.Sym.Name, r.Off, t)
}
if int64(int32(t<<6)>>6) != t {
// TODO(austin) This can happen if text > 32M.
// Add a call trampoline to .text in that case.
ld.Ctxt.Diag("relocation for %s+%d is too big: %d", r.Sym.Name, r.Off, t)
}
*val |= int64(uint32(t) &^ 0xfc000003)
return 0
case obj.R_POWER_TOC: // S + A - .TOC.
*val = ld.Symaddr(r.Sym) + r.Add - symtoc(s)
return 0
case obj.R_POWER_TLS_LE:
// The thread pointer points 0x7000 bytes after the start of the the
// thread local storage area as documented in section "3.7.2 TLS
// Runtime Handling" of "Power Architecture 64-Bit ELF V2 ABI
// Specification".
v := r.Sym.Value - 0x7000
if int64(int16(v)) != v {
ld.Diag("TLS offset out of range %d", v)
}
*val = (*val &^ 0xffff) | (v & 0xffff)
return 0
}
return -1
}
func archreloc(ctxt *ld.Link, r *ld.Reloc, s *ld.Symbol, val *int64) int {
if ld.Linkmode == ld.LinkExternal {
switch r.Type {
default:
return -1
case obj.R_ADDRMIPS,
obj.R_ADDRMIPSU:
r.Done = 0
// set up addend for eventual relocation via outer symbol.
rs := r.Sym
r.Xadd = r.Add
for rs.Outer != nil {
r.Xadd += ld.Symaddr(ctxt, rs) - ld.Symaddr(ctxt, rs.Outer)
rs = rs.Outer
}
if rs.Type != obj.SHOSTOBJ && rs.Type != obj.SDYNIMPORT && rs.Sect == nil {
ctxt.Diag("missing section for %s", rs.Name)
}
r.Xsym = rs
return 0
case obj.R_ADDRMIPSTLS,
obj.R_CALLMIPS,
obj.R_JMPMIPS:
r.Done = 0
r.Xsym = r.Sym
r.Xadd = r.Add
return 0
}
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(ctxt, r.Sym) + r.Add - ld.Symaddr(ctxt, ld.Linklookup(ctxt, ".got", 0))
return 0
case obj.R_ADDRMIPS,
obj.R_ADDRMIPSU:
t := ld.Symaddr(ctxt, r.Sym) + r.Add
o1 := ld.SysArch.ByteOrder.Uint32(s.P[r.Off:])
if r.Type == obj.R_ADDRMIPS {
*val = int64(o1&0xffff0000 | uint32(t)&0xffff)
} else {
*val = int64(o1&0xffff0000 | uint32((t+1<<15)>>16)&0xffff)
}
return 0
case obj.R_ADDRMIPSTLS:
// thread pointer is at 0x7000 offset from the start of TLS data area
t := ld.Symaddr(ctxt, r.Sym) + r.Add - 0x7000
if t < -32768 || t >= 32678 {
ctxt.Diag("TLS offset out of range %d", t)
}
o1 := ld.SysArch.ByteOrder.Uint32(s.P[r.Off:])
*val = int64(o1&0xffff0000 | uint32(t)&0xffff)
return 0
case obj.R_CALLMIPS,
obj.R_JMPMIPS:
// Low 26 bits = (S + A) >> 2
t := ld.Symaddr(ctxt, r.Sym) + r.Add
o1 := ld.SysArch.ByteOrder.Uint32(s.P[r.Off:])
*val = int64(o1&0xfc000000 | uint32(t>>2)&^0xfc000000)
return 0
}
return -1
}
func archreloc(ctxt *ld.Link, r *ld.Reloc, s *ld.Symbol, val *int64) int {
if ld.Linkmode == ld.LinkExternal {
switch r.Type {
case obj.R_CALLARM:
r.Done = 0
// set up addend for eventual relocation via outer symbol.
rs := r.Sym
r.Xadd = r.Add
if r.Xadd&0x800000 != 0 {
r.Xadd |= ^0xffffff
}
r.Xadd *= 4
for rs.Outer != nil {
r.Xadd += ld.Symaddr(ctxt, rs) - ld.Symaddr(ctxt, rs.Outer)
rs = rs.Outer
}
if rs.Type != obj.SHOSTOBJ && rs.Type != obj.SDYNIMPORT && rs.Sect == nil {
ctxt.Diag("missing section for %s", rs.Name)
}
r.Xsym = rs
// ld64 for arm seems to want the symbol table to contain offset
// into the section rather than pseudo virtual address that contains
// the section load address.
// we need to compensate that by removing the instruction's address
// from addend.
if ld.HEADTYPE == obj.Hdarwin {
r.Xadd -= ld.Symaddr(ctxt, s) + int64(r.Off)
}
*val = int64(braddoff(int32(0xff000000&uint32(r.Add)), int32(0xffffff&uint32(r.Xadd/4))))
return 0
}
return -1
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(ctxt, r.Sym) + r.Add - ld.Symaddr(ctxt, ld.Linklookup(ctxt, ".got", 0))
return 0
// The following three arch specific relocations are only for generation of
// Linux/ARM ELF's PLT entry (3 assembler instruction)
case obj.R_PLT0: // add ip, pc, #0xXX00000
if ld.Symaddr(ctxt, ld.Linklookup(ctxt, ".got.plt", 0)) < ld.Symaddr(ctxt, ld.Linklookup(ctxt, ".plt", 0)) {
ctxt.Diag(".got.plt should be placed after .plt section.")
}
*val = 0xe28fc600 + (0xff & (int64(uint32(ld.Symaddr(ctxt, r.Sym)-(ld.Symaddr(ctxt, ld.Linklookup(ctxt, ".plt", 0))+int64(r.Off))+r.Add)) >> 20))
return 0
case obj.R_PLT1: // add ip, ip, #0xYY000
*val = 0xe28cca00 + (0xff & (int64(uint32(ld.Symaddr(ctxt, r.Sym)-(ld.Symaddr(ctxt, ld.Linklookup(ctxt, ".plt", 0))+int64(r.Off))+r.Add+4)) >> 12))
return 0
case obj.R_PLT2: // ldr pc, [ip, #0xZZZ]!
*val = 0xe5bcf000 + (0xfff & int64(uint32(ld.Symaddr(ctxt, r.Sym)-(ld.Symaddr(ctxt, ld.Linklookup(ctxt, ".plt", 0))+int64(r.Off))+r.Add+8)))
return 0
case obj.R_CALLARM: // bl XXXXXX or b YYYYYY
*val = int64(braddoff(int32(0xff000000&uint32(r.Add)), int32(0xffffff&uint32((ld.Symaddr(ctxt, r.Sym)+int64((uint32(r.Add))*4)-(s.Value+int64(r.Off)))/4))))
return 0
}
return -1
}
func machoreloc1(ctxt *ld.Link, r *ld.Reloc, sectoff int64) int {
var v uint32
rs := r.Xsym
if r.Type == obj.R_PCREL {
if rs.Type == obj.SHOSTOBJ {
ctxt.Diag("pc-relative relocation of external symbol is not supported")
return -1
}
if r.Siz != 4 {
return -1
}
// emit a pair of "scattered" relocations that
// resolve to the difference of section addresses of
// the symbol and the instruction
// this value is added to the field being relocated
o1 := uint32(sectoff)
o1 |= 1 << 31 // scattered bit
o1 |= ld.MACHO_ARM_RELOC_SECTDIFF << 24
o1 |= 2 << 28 // size = 4
o2 := uint32(0)
o2 |= 1 << 31 // scattered bit
o2 |= ld.MACHO_ARM_RELOC_PAIR << 24
o2 |= 2 << 28 // size = 4
ld.Thearch.Lput(o1)
ld.Thearch.Lput(uint32(ld.Symaddr(ctxt, rs)))
ld.Thearch.Lput(o2)
ld.Thearch.Lput(uint32(ctxt.Cursym.Value + int64(r.Off)))
return 0
}
if rs.Type == obj.SHOSTOBJ || r.Type == obj.R_CALLARM {
if rs.Dynid < 0 {
ctxt.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.Extnum)
if v == 0 {
ctxt.Diag("reloc %d to symbol %s in non-macho section %s type=%d", r.Type, rs.Name, rs.Sect.Name, rs.Type)
return -1
}
}
switch r.Type {
default:
return -1
case obj.R_ADDR:
v |= ld.MACHO_GENERIC_RELOC_VANILLA << 28
case obj.R_CALLARM:
v |= 1 << 24 // pc-relative bit
v |= ld.MACHO_ARM_RELOC_BR24 << 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 {
if ld.Linkmode == ld.LinkExternal {
switch r.Type {
default:
return -1
case obj.R_ARM64_GOTPCREL:
var o1, o2 uint32
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
o1 = uint32(*val >> 32)
o2 = uint32(*val)
} else {
o1 = uint32(*val)
o2 = uint32(*val >> 32)
}
// Any relocation against a function symbol is redirected to
// be against a local symbol instead (see putelfsym in
// symtab.go) but unfortunately the system linker was buggy
// when confronted with a R_AARCH64_ADR_GOT_PAGE relocation
// against a local symbol until May 2015
// (https://sourceware.org/bugzilla/show_bug.cgi?id=18270). So
// we convert the adrp; ld64 + R_ARM64_GOTPCREL into adrp;
// add + R_ADDRARM64.
if !(r.Sym.Version != 0 || (r.Sym.Type&obj.SHIDDEN != 0) || r.Sym.Attr.Local()) && r.Sym.Type == obj.STEXT && ld.DynlinkingGo() {
if o2&0xffc00000 != 0xf9400000 {
ld.Ctxt.Diag("R_ARM64_GOTPCREL against unexpected instruction %x", o2)
}
o2 = 0x91000000 | (o2 & 0x000003ff)
r.Type = obj.R_ADDRARM64
}
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
*val = int64(o1)<<32 | int64(o2)
} else {
*val = int64(o2)<<32 | int64(o1)
}
fallthrough
case obj.R_ADDRARM64:
r.Done = 0
// set up addend for eventual relocation via outer symbol.
rs := r.Sym
r.Xadd = r.Add
for rs.Outer != nil {
r.Xadd += ld.Symaddr(rs) - ld.Symaddr(rs.Outer)
rs = rs.Outer
}
if rs.Type != obj.SHOSTOBJ && rs.Type != obj.SDYNIMPORT && rs.Sect == nil {
ld.Diag("missing section for %s", rs.Name)
}
r.Xsym = rs
// 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 == obj.Hdarwin {
var o0, o1 uint32
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
o0 = uint32(*val >> 32)
o1 = uint32(*val)
} else {
o0 = uint32(*val)
o1 = uint32(*val >> 32)
}
// 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 obj.R_CALLARM64,
obj.R_ARM64_TLS_LE,
obj.R_ARM64_TLS_IE:
r.Done = 0
r.Xsym = r.Sym
r.Xadd = r.Add
return 0
}
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(r.Sym) + r.Add - ld.Symaddr(ld.Linklookup(ld.Ctxt, ".got", 0))
return 0
case obj.R_ADDRARM64:
t := ld.Symaddr(r.Sym) + r.Add - ((s.Value + int64(r.Off)) &^ 0xfff)
if t >= 1<<32 || t < -1<<32 {
ld.Diag("program too large, address relocation distance = %d", t)
}
var o0, o1 uint32
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
o0 = uint32(*val >> 32)
o1 = uint32(*val)
} else {
o0 = uint32(*val)
o1 = uint32(*val >> 32)
}
o0 |= (uint32((t>>12)&3) << 29) | (uint32((t>>12>>2)&0x7ffff) << 5)
o1 |= uint32(t&0xfff) << 10
// 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 obj.R_ARM64_TLS_LE:
r.Done = 0
if ld.HEADTYPE != obj.Hlinux {
ld.Diag("TLS reloc on unsupported OS %s", ld.Headstr(int(ld.HEADTYPE)))
}
// The TCB is two pointers. This is not documented anywhere, but is
// de facto part of the ABI.
v := r.Sym.Value + int64(2*ld.Thearch.Ptrsize)
if v < 0 || v >= 32678 {
ld.Diag("TLS offset out of range %d", v)
}
*val |= v << 5
return 0
case obj.R_CALLARM64:
t := (ld.Symaddr(r.Sym) + r.Add) - (s.Value + int64(r.Off))
if t >= 1<<27 || t < -1<<27 {
ld.Diag("program too large, call relocation distance = %d", t)
}
*val |= (t >> 2) & 0x03ffffff
return 0
}
return -1
}
func archreloc(r *ld.Reloc, s *ld.LSym, val *int64) int {
if ld.Linkmode == ld.LinkExternal {
switch r.Type {
default:
return -1
case obj.R_ADDRARM64:
r.Done = 0
// set up addend for eventual relocation via outer symbol.
rs := r.Sym
r.Xadd = r.Add
for rs.Outer != nil {
r.Xadd += ld.Symaddr(rs) - ld.Symaddr(rs.Outer)
rs = rs.Outer
}
if rs.Type != obj.SHOSTOBJ && rs.Sect == nil {
ld.Diag("missing section for %s", rs.Name)
}
r.Xsym = rs
// 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 == obj.Hdarwin {
var o0, o1 uint32
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
o0 = uint32(*val >> 32)
o1 = uint32(*val)
} else {
o0 = uint32(*val)
o1 = uint32(*val >> 32)
}
// 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 obj.R_CALLARM64,
obj.R_ARM64_TLS_LE,
obj.R_ARM64_TLS_IE:
r.Done = 0
r.Xsym = r.Sym
r.Xadd = r.Add
return 0
}
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(r.Sym) + r.Add - ld.Symaddr(ld.Linklookup(ld.Ctxt, ".got", 0))
return 0
case obj.R_ADDRARM64:
t := ld.Symaddr(r.Sym) + r.Add - ((s.Value + int64(r.Off)) &^ 0xfff)
if t >= 1<<32 || t < -1<<32 {
ld.Diag("program too large, address relocation distance = %d", t)
}
var o0, o1 uint32
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
o0 = uint32(*val >> 32)
o1 = uint32(*val)
} else {
o0 = uint32(*val)
o1 = uint32(*val >> 32)
}
o0 |= (uint32((t>>12)&3) << 29) | (uint32((t>>12>>2)&0x7ffff) << 5)
o1 |= uint32(t&0xfff) << 10
// 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 obj.R_ARM64_TLS_LE:
r.Done = 0
if ld.HEADTYPE != obj.Hlinux {
ld.Diag("TLS reloc on unsupported OS %s", ld.Headstr(int(ld.HEADTYPE)))
}
// The TCB is two pointers. This is not documented anywhere, but is
// de facto part of the ABI.
v := r.Sym.Value + int64(2*ld.Thearch.Ptrsize)
if v < 0 || v >= 32678 {
ld.Diag("TLS offset out of range %d", v)
}
*val |= v << 5
return 0
case obj.R_CALLARM64:
t := (ld.Symaddr(r.Sym) + r.Add) - (s.Value + int64(r.Off))
if t >= 1<<27 || t < -1<<27 {
ld.Diag("program too large, call relocation distance = %d", t)
}
*val |= (t >> 2) & 0x03ffffff
return 0
}
return -1
}
func archreloc(r *ld.Reloc, s *ld.LSym, val *int64) int {
if ld.Linkmode == ld.LinkExternal {
switch r.Type {
default:
return -1
case obj.R_ADDRARM64:
r.Done = 0
// set up addend for eventual relocation via outer symbol.
rs := r.Sym
r.Xadd = r.Add
for rs.Outer != nil {
r.Xadd += ld.Symaddr(rs) - ld.Symaddr(rs.Outer)
rs = rs.Outer
}
if rs.Type != obj.SHOSTOBJ && rs.Sect == nil {
ld.Diag("missing section for %s", rs.Name)
}
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 == obj.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 obj.R_CALLARM64:
r.Done = 0
r.Xsym = r.Sym
*val = int64(0xfc000000 & uint32(r.Add))
r.Xadd = int64((uint32(r.Add) &^ 0xfc000000) * 4)
r.Add = 0
return 0
}
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(r.Sym) + r.Add - ld.Symaddr(ld.Linklookup(ld.Ctxt, ".got", 0))
return 0
case obj.R_ADDRARM64:
t := ld.Symaddr(r.Sym) + r.Add - ((s.Value + int64(r.Off)) &^ 0xfff)
if t >= 1<<32 || t < -1<<32 {
ld.Diag("program too large, address relocation distance = %d", t)
}
// 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])
o0 |= (uint32((t>>12)&3) << 29) | (uint32((t>>12>>2)&0x7ffff) << 5)
o1 |= uint32(t&0xfff) << 10
// 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 obj.R_CALLARM64:
*val = int64((0xfc000000 & uint32(r.Add)) | uint32((ld.Symaddr(r.Sym)+r.Add*4-(s.Value+int64(r.Off)))/4))
return 0
}
return -1
}
func archreloc(r *ld.Reloc, s *ld.LSym, val *int64) int {
if ld.Linkmode == ld.LinkExternal {
// TODO(minux): translate R_ADDRPOWER and R_CALLPOWER into standard ELF relocations.
// R_ADDRPOWER corresponds to R_PPC_ADDR16_HA and R_PPC_ADDR16_LO.
// R_CALLPOWER corresponds to R_PPC_REL24.
return -1
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(r.Sym) + r.Add - ld.Symaddr(ld.Linklookup(ld.Ctxt, ".got", 0))
return 0
case obj.R_ADDRPOWER:
// We are spreading a 31-bit address across two instructions,
// putting the high (adjusted) part in the low 16 bits of the
// first instruction and the low part in the low 16 bits of the
// second instruction.
t := ld.Symaddr(r.Sym) + r.Add
if t < 0 || t >= 1<<31 {
ld.Ctxt.Diag("relocation for %s is too big (>=2G): %d", s.Name, ld.Symaddr(r.Sym))
}
var o1, o2 uint32
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
o1 = uint32(*val >> 32)
o2 = uint32(*val)
} else {
o1 = uint32(*val)
o2 = uint32(*val >> 32)
}
if t&0x8000 != 0 {
t += 0x10000
}
// There is an almost-bug here. When R_ADDRPOWER is relocating a
// load, the two instructions are addi and then a load. addi and
// almost all loads are "D-form" instructions, which have a
// 16-bit immediate in the lower 16-bits of the instruction
// word. But the load doubleword instruction is a "DS-form"
// instruction: the immediate only occupies bits 16-29 of the
// instruction and is implicity padded with zeros on the
// right. The reason the belows isn't a bug is because we only
// ever use immediates that have zeros on in their lower bits
// with ld, and we combine the immediate with | so bits 30 and
// 31 are preserved.
o1 |= (uint32(t) >> 16) & 0xffff
o2 |= uint32(t) & 0xffff
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
*val = int64(o1)<<32 | int64(o2)
} else {
*val = int64(o2)<<32 | int64(o1)
}
return 0
case obj.R_CALLPOWER:
// Bits 6 through 29 = (S + A - P) >> 2
t := ld.Symaddr(r.Sym) + r.Add - (s.Value + int64(r.Off))
if t&3 != 0 {
ld.Ctxt.Diag("relocation for %s+%d is not aligned: %d", r.Sym.Name, r.Off, t)
}
if int64(int32(t<<6)>>6) != t {
// TODO(austin) This can happen if text > 32M.
// Add a call trampoline to .text in that case.
ld.Ctxt.Diag("relocation for %s+%d is too big: %d", r.Sym.Name, r.Off, t)
}
*val |= int64(uint32(t) &^ 0xfc000003)
return 0
case obj.R_POWER_TOC: // S + A - .TOC.
*val = ld.Symaddr(r.Sym) + r.Add - symtoc(s)
return 0
case obj.R_POWER_TLS_LE:
// The thread pointer points 0x7000 bytes after the start of the the
// thread local storage area as documented in section "3.7.2 TLS
// Runtime Handling" of "Power Architecture 64-Bit ELF V2 ABI
// Specification".
v := r.Sym.Value - 0x7000
if int64(int16(v)) != v {
ld.Diag("TLS offset out of range %d", v)
}
*val = (*val &^ 0xffff) | (v & 0xffff)
return 0
}
return -1
}
func archreloc(r *ld.Reloc, s *ld.LSym, val *int64) int {
if ld.Linkmode == ld.LinkExternal {
// TODO(minux): translate R_ADDRPOWER and R_CALLPOWER into standard ELF relocations.
// R_ADDRPOWER corresponds to R_PPC_ADDR16_HA and R_PPC_ADDR16_LO.
// R_CALLPOWER corresponds to R_PPC_REL24.
return -1
}
switch r.Type {
case obj.R_CONST:
*val = r.Add
return 0
case obj.R_GOTOFF:
*val = ld.Symaddr(r.Sym) + r.Add - ld.Symaddr(ld.Linklookup(ld.Ctxt, ".got", 0))
return 0
case obj.R_ADDRPOWER:
// r->add is two ppc64 instructions holding an immediate 32-bit constant.
// We want to add r->sym's address to that constant.
// The encoding of the immediate x<<16 + y,
// where x is the low 16 bits of the first instruction and y is the low 16
// bits of the second. Both x and y are signed (int16, not uint16).
o1 := uint32(r.Add >> 32)
o2 := uint32(r.Add)
t := ld.Symaddr(r.Sym)
if t < 0 {
ld.Ctxt.Diag("relocation for %s is too big (>=2G): %d", s.Name, ld.Symaddr(r.Sym))
}
t += int64((o1&0xffff)<<16 + uint32(int32(o2)<<16>>16))
if t&0x8000 != 0 {
t += 0x10000
}
o1 = o1&0xffff0000 | (uint32(t)>>16)&0xffff
o2 = o2&0xffff0000 | uint32(t)&0xffff
// when laid out, the instruction order must always be o1, o2.
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
*val = int64(o1)<<32 | int64(o2)
} else {
*val = int64(o2)<<32 | int64(o1)
}
return 0
case obj.R_CALLPOWER:
// Bits 6 through 29 = (S + A - P) >> 2
var o1 uint32
if ld.Ctxt.Arch.ByteOrder == binary.BigEndian {
o1 = ld.Be32(s.P[r.Off:])
} else {
o1 = ld.Le32(s.P[r.Off:])
}
t := ld.Symaddr(r.Sym) + r.Add - (s.Value + int64(r.Off))
if t&3 != 0 {
ld.Ctxt.Diag("relocation for %s+%d is not aligned: %d", r.Sym.Name, r.Off, t)
}
if int64(int32(t<<6)>>6) != t {
// TODO(austin) This can happen if text > 32M.
// Add a call trampoline to .text in that case.
ld.Ctxt.Diag("relocation for %s+%d is too big: %d", r.Sym.Name, r.Off, t)
}
*val = int64(o1&0xfc000003 | uint32(t)&^0xfc000003)
return 0
case obj.R_POWER_TOC: // S + A - .TOC.
*val = ld.Symaddr(r.Sym) + r.Add - symtoc(s)
return 0
}
return -1
}
// resolve direct jump relocation r in s, and add trampoline if necessary
func trampoline(ctxt *ld.Link, r *ld.Reloc, s *ld.Symbol) {
t := ld.Symaddr(r.Sym) + r.Add - (s.Value + int64(r.Off))
switch r.Type {
case obj.R_CALLPOWER:
// If branch offset is too far then create a trampoline.
if int64(int32(t<<6)>>6) != t || (*ld.FlagDebugTramp > 1 && s.File != r.Sym.File) {
var tramp *ld.Symbol
for i := 0; ; i++ {
// Using r.Add as part of the name is significant in functions like duffzero where the call
// target is at some offset within the function. Calls to duff+8 and duff+256 must appear as
// distinct trampolines.
name := r.Sym.Name
if r.Add == 0 {
name = name + fmt.Sprintf("-tramp%d", i)
} else {
name = name + fmt.Sprintf("%+x-tramp%d", r.Add, i)
}
// Look up the trampoline in case it already exists
tramp = ctxt.Syms.Lookup(name, int(r.Sym.Version))
if tramp.Value == 0 {
break
}
t = ld.Symaddr(tramp) + r.Add - (s.Value + int64(r.Off))
// If the offset of the trampoline that has been found is within range, use it.
if int64(int32(t<<6)>>6) == t {
break
}
}
if tramp.Type == 0 {
ctxt.AddTramp(tramp)
tramp.Size = 16 // 4 instructions
tramp.P = make([]byte, tramp.Size)
t = ld.Symaddr(r.Sym) + r.Add
f := t & 0xffff0000
o1 := uint32(0x3fe00000 | (f >> 16)) // lis r31,trampaddr hi (r31 is temp reg)
f = t & 0xffff
o2 := uint32(0x63ff0000 | f) // ori r31,trampaddr lo
o3 := uint32(0x7fe903a6) // mtctr
o4 := uint32(0x4e800420) // bctr
ld.SysArch.ByteOrder.PutUint32(tramp.P, o1)
ld.SysArch.ByteOrder.PutUint32(tramp.P[4:], o2)
ld.SysArch.ByteOrder.PutUint32(tramp.P[8:], o3)
ld.SysArch.ByteOrder.PutUint32(tramp.P[12:], o4)
}
r.Sym = tramp
r.Add = 0 // This was folded into the trampoline target address
r.Done = 0
}
default:
ld.Errorf(s, "trampoline called with non-jump reloc: %v", r.Type)
}
}