func nacladdr(ctxt *obj.Link, p *obj.Prog, a *obj.Addr) {
if p.As == ALEAL || p.As == ALEAQ {
return
}
if a.Reg == REG_BP {
ctxt.Diag("invalid address: %v", p)
return
}
if a.Reg == REG_TLS {
a.Reg = REG_BP
}
if a.Type == obj.TYPE_MEM && a.Name == obj.NAME_NONE {
switch a.Reg {
// all ok
case REG_BP, REG_SP, REG_R15:
break
default:
if a.Index != REG_NONE {
ctxt.Diag("invalid address %v", p)
}
a.Index = a.Reg
if a.Index != REG_NONE {
a.Scale = 1
}
a.Reg = REG_R15
}
}
}
func follow(ctxt *obj.Link, s *obj.LSym) {
ctxt.Cursym = s
firstp := ctxt.NewProg()
lastp := firstp
xfol(ctxt, s.Text, &lastp)
lastp.Link = nil
s.Text = firstp.Link
}
func initdiv(ctxt *obj.Link) {
if ctxt.Sym_div != nil {
return
}
ctxt.Sym_div = obj.Linklookup(ctxt, "_div", 0)
ctxt.Sym_divu = obj.Linklookup(ctxt, "_divu", 0)
ctxt.Sym_mod = obj.Linklookup(ctxt, "_mod", 0)
ctxt.Sym_modu = obj.Linklookup(ctxt, "_modu", 0)
}
func oplook(ctxt *obj.Link, p *obj.Prog) *Optab {
if oprange[AOR&obj.AMask].start == nil {
buildop(ctxt)
}
a1 := int(p.Optab)
if a1 != 0 {
return &optab[a1-1:][0]
}
a1 = int(p.From.Class)
if a1 == 0 {
a1 = aclass(ctxt, &p.From) + 1
p.From.Class = int8(a1)
}
a1--
a3 := int(p.To.Class)
if a3 == 0 {
a3 = aclass(ctxt, &p.To) + 1
p.To.Class = int8(a3)
}
a3--
a2 := C_NONE
if p.Reg != 0 {
a2 = C_REG
}
//print("oplook %P %d %d %d\n", p, a1, a2, a3);
r0 := p.As & obj.AMask
o := oprange[r0].start
if o == nil {
o = oprange[r0].stop /* just generate an error */
}
e := oprange[r0].stop
c1 := xcmp[a1][:]
c3 := xcmp[a3][:]
for ; -cap(o) < -cap(e); o = o[1:] {
if int(o[0].a2) == a2 {
if c1[o[0].a1] != 0 {
if c3[o[0].a3] != 0 {
p.Optab = uint16((-cap(o) + cap(optab)) + 1)
return &o[0]
}
}
}
}
ctxt.Diag("illegal combination %v %v %v %v", obj.Aconv(int(p.As)), DRconv(a1), DRconv(a2), DRconv(a3))
prasm(p)
if o == nil {
o = optab
}
return &o[0]
}
func addnop(ctxt *obj.Link, p *obj.Prog) {
q := ctxt.NewProg()
// we want to use the canonical NOP (SLL $0,R0,R0) here,
// however, as the assembler will always replace $0
// as R0, we have to resort to manually encode the SLL
// instruction as WORD $0.
q.As = AWORD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Name = obj.NAME_NONE
q.From.Offset = 0
q.Link = p.Link
p.Link = q
}
func oplook(ctxt *obj.Link, p *obj.Prog) *Optab {
if oprange[AOR&obj.AMask] == nil {
buildop(ctxt)
}
a1 := int(p.Optab)
if a1 != 0 {
return &optab[a1-1]
}
a1 = int(p.From.Class)
if a1 == 0 {
a1 = aclass(ctxt, &p.From) + 1
p.From.Class = int8(a1)
}
a1--
a3 := int(p.To.Class)
if a3 == 0 {
a3 = aclass(ctxt, &p.To) + 1
p.To.Class = int8(a3)
}
a3--
a2 := C_NONE
if p.Reg != 0 {
a2 = C_REG
}
//print("oplook %P %d %d %d\n", p, a1, a2, a3);
ops := oprange[p.As&obj.AMask]
c1 := &xcmp[a1]
c3 := &xcmp[a3]
for i := range ops {
op := &ops[i]
if int(op.a2) == a2 && c1[op.a1] && c3[op.a3] {
p.Optab = uint16(cap(optab) - cap(ops) + i + 1)
return op
}
}
ctxt.Diag("illegal combination %v %v %v %v", obj.Aconv(p.As), DRconv(a1), DRconv(a2), DRconv(a3))
prasm(p)
if ops == nil {
ops = optab
}
return &ops[0]
}
func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
if ctxt.Headtype == obj.Hplan9 && ctxt.Plan9privates == nil {
ctxt.Plan9privates = obj.Linklookup(ctxt, "_privates", 0)
}
ctxt.Cursym = cursym
if cursym.Text == nil || cursym.Text.Link == nil {
return
}
p := cursym.Text
autoffset := int32(p.To.Offset)
if autoffset < 0 {
autoffset = 0
}
var bpsize int
if p.Mode == 64 && ctxt.Framepointer_enabled && autoffset > 0 && p.From3.Offset&obj.NOFRAME == 0 {
// Make room for to save a base pointer. If autoffset == 0,
// this might do something special like a tail jump to
// another function, so in that case we omit this.
bpsize = ctxt.Arch.PtrSize
autoffset += int32(bpsize)
p.To.Offset += int64(bpsize)
} else {
bpsize = 0
}
textarg := int64(p.To.Val.(int32))
cursym.Args = int32(textarg)
cursym.Locals = int32(p.To.Offset)
// TODO(rsc): Remove.
if p.Mode == 32 && cursym.Locals < 0 {
cursym.Locals = 0
}
// TODO(rsc): Remove 'p.Mode == 64 &&'.
if p.Mode == 64 && autoffset < obj.StackSmall && p.From3Offset()&obj.NOSPLIT == 0 {
leaf := true
LeafSearch:
for q := p; q != nil; q = q.Link {
switch q.As {
case obj.ACALL:
// Treat common runtime calls that take no arguments
// the same as duffcopy and duffzero.
if !isZeroArgRuntimeCall(q.To.Sym) {
leaf = false
break LeafSearch
}
fallthrough
case obj.ADUFFCOPY, obj.ADUFFZERO:
if autoffset >= obj.StackSmall-8 {
leaf = false
break LeafSearch
}
}
}
if leaf {
p.From3.Offset |= obj.NOSPLIT
}
}
if p.From3Offset()&obj.NOSPLIT == 0 || p.From3Offset()&obj.WRAPPER != 0 {
p = obj.Appendp(ctxt, p)
p = load_g_cx(ctxt, p) // load g into CX
}
if cursym.Text.From3Offset()&obj.NOSPLIT == 0 {
p = stacksplit(ctxt, p, autoffset, int32(textarg)) // emit split check
}
if autoffset != 0 {
if autoffset%int32(ctxt.Arch.RegSize) != 0 {
ctxt.Diag("unaligned stack size %d", autoffset)
}
p = obj.Appendp(ctxt, p)
p.As = AADJSP
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(autoffset)
p.Spadj = autoffset
}
deltasp := autoffset
if bpsize > 0 {
// Save caller's BP
p = obj.Appendp(ctxt, p)
p.As = AMOVQ
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_BP
p.To.Type = obj.TYPE_MEM
p.To.Reg = REG_SP
p.To.Scale = 1
p.To.Offset = int64(autoffset) - int64(bpsize)
// Move current frame to BP
p = obj.Appendp(ctxt, p)
p.As = ALEAQ
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_SP
p.From.Scale = 1
p.From.Offset = int64(autoffset) - int64(bpsize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_BP
}
if cursym.Text.From3Offset()&obj.WRAPPER != 0 {
// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
//
// MOVQ g_panic(CX), BX
// TESTQ BX, BX
// JEQ end
// LEAQ (autoffset+8)(SP), DI
// CMPQ panic_argp(BX), DI
// JNE end
// MOVQ SP, panic_argp(BX)
// end:
// NOP
//
// The NOP is needed to give the jumps somewhere to land.
// It is a liblink NOP, not an x86 NOP: it encodes to 0 instruction bytes.
p = obj.Appendp(ctxt, p)
p.As = AMOVQ
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_CX
p.From.Offset = 4 * int64(ctxt.Arch.PtrSize) // G.panic
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_BX
if ctxt.Headtype == obj.Hnacl && p.Mode == 64 {
p.As = AMOVL
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_R15
p.From.Scale = 1
p.From.Index = REG_CX
}
if p.Mode == 32 {
p.As = AMOVL
}
p = obj.Appendp(ctxt, p)
p.As = ATESTQ
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_BX
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_BX
if ctxt.Headtype == obj.Hnacl || p.Mode == 32 {
p.As = ATESTL
}
p = obj.Appendp(ctxt, p)
p.As = AJEQ
p.To.Type = obj.TYPE_BRANCH
p1 := p
p = obj.Appendp(ctxt, p)
p.As = ALEAQ
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_SP
p.From.Offset = int64(autoffset) + int64(ctxt.Arch.RegSize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_DI
if ctxt.Headtype == obj.Hnacl || p.Mode == 32 {
p.As = ALEAL
}
p = obj.Appendp(ctxt, p)
p.As = ACMPQ
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_BX
p.From.Offset = 0 // Panic.argp
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_DI
if ctxt.Headtype == obj.Hnacl && p.Mode == 64 {
p.As = ACMPL
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_R15
p.From.Scale = 1
p.From.Index = REG_BX
}
if p.Mode == 32 {
p.As = ACMPL
}
p = obj.Appendp(ctxt, p)
p.As = AJNE
p.To.Type = obj.TYPE_BRANCH
p2 := p
p = obj.Appendp(ctxt, p)
p.As = AMOVQ
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_SP
p.To.Type = obj.TYPE_MEM
p.To.Reg = REG_BX
p.To.Offset = 0 // Panic.argp
if ctxt.Headtype == obj.Hnacl && p.Mode == 64 {
p.As = AMOVL
p.To.Type = obj.TYPE_MEM
p.To.Reg = REG_R15
p.To.Scale = 1
p.To.Index = REG_BX
}
if p.Mode == 32 {
p.As = AMOVL
}
p = obj.Appendp(ctxt, p)
p.As = obj.ANOP
p1.Pcond = p
p2.Pcond = p
}
for ; p != nil; p = p.Link {
pcsize := int(p.Mode) / 8
switch p.From.Name {
case obj.NAME_AUTO:
p.From.Offset += int64(deltasp) - int64(bpsize)
case obj.NAME_PARAM:
p.From.Offset += int64(deltasp) + int64(pcsize)
}
if p.From3 != nil {
switch p.From3.Name {
case obj.NAME_AUTO:
p.From3.Offset += int64(deltasp) - int64(bpsize)
case obj.NAME_PARAM:
p.From3.Offset += int64(deltasp) + int64(pcsize)
}
}
switch p.To.Name {
case obj.NAME_AUTO:
p.To.Offset += int64(deltasp) - int64(bpsize)
case obj.NAME_PARAM:
p.To.Offset += int64(deltasp) + int64(pcsize)
}
switch p.As {
default:
continue
case APUSHL, APUSHFL:
deltasp += 4
p.Spadj = 4
continue
case APUSHQ, APUSHFQ:
deltasp += 8
p.Spadj = 8
continue
case APUSHW, APUSHFW:
deltasp += 2
p.Spadj = 2
continue
case APOPL, APOPFL:
deltasp -= 4
p.Spadj = -4
continue
case APOPQ, APOPFQ:
deltasp -= 8
p.Spadj = -8
continue
case APOPW, APOPFW:
deltasp -= 2
p.Spadj = -2
continue
case obj.ARET:
// do nothing
}
if autoffset != deltasp {
ctxt.Diag("unbalanced PUSH/POP")
}
if autoffset != 0 {
if bpsize > 0 {
// Restore caller's BP
p.As = AMOVQ
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_SP
p.From.Scale = 1
p.From.Offset = int64(autoffset) - int64(bpsize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_BP
p = obj.Appendp(ctxt, p)
}
p.As = AADJSP
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(-autoffset)
p.Spadj = -autoffset
p = obj.Appendp(ctxt, p)
p.As = obj.ARET
// If there are instructions following
// this ARET, they come from a branch
// with the same stackframe, so undo
// the cleanup.
p.Spadj = +autoffset
}
if p.To.Sym != nil { // retjmp
p.As = obj.AJMP
}
}
}
// Rewrite p, if necessary, to access global data via the global offset table.
func rewriteToUseGot(ctxt *obj.Link, p *obj.Prog) {
var add, lea, mov obj.As
var reg int16
if p.Mode == 64 {
add = AADDQ
lea = ALEAQ
mov = AMOVQ
reg = REG_R15
} else {
add = AADDL
lea = ALEAL
mov = AMOVL
reg = REG_CX
}
if p.As == obj.ADUFFCOPY || p.As == obj.ADUFFZERO {
// ADUFFxxx $offset
// becomes
// $MOV runtime.duffxxx@GOT, $reg
// $ADD $offset, $reg
// CALL $reg
var sym *obj.LSym
if p.As == obj.ADUFFZERO {
sym = obj.Linklookup(ctxt, "runtime.duffzero", 0)
} else {
sym = obj.Linklookup(ctxt, "runtime.duffcopy", 0)
}
offset := p.To.Offset
p.As = mov
p.From.Type = obj.TYPE_MEM
p.From.Name = obj.NAME_GOTREF
p.From.Sym = sym
p.To.Type = obj.TYPE_REG
p.To.Reg = reg
p.To.Offset = 0
p.To.Sym = nil
p1 := obj.Appendp(ctxt, p)
p1.As = add
p1.From.Type = obj.TYPE_CONST
p1.From.Offset = offset
p1.To.Type = obj.TYPE_REG
p1.To.Reg = reg
p2 := obj.Appendp(ctxt, p1)
p2.As = obj.ACALL
p2.To.Type = obj.TYPE_REG
p2.To.Reg = reg
}
// We only care about global data: NAME_EXTERN means a global
// symbol in the Go sense, and p.Sym.Local is true for a few
// internally defined symbols.
if p.As == lea && p.From.Type == obj.TYPE_MEM && p.From.Name == obj.NAME_EXTERN && !p.From.Sym.Local {
// $LEA sym, Rx becomes $MOV $sym, Rx which will be rewritten below
p.As = mov
p.From.Type = obj.TYPE_ADDR
}
if p.From.Type == obj.TYPE_ADDR && p.From.Name == obj.NAME_EXTERN && !p.From.Sym.Local {
// $MOV $sym, Rx becomes $MOV sym@GOT, Rx
// $MOV $sym+<off>, Rx becomes $MOV sym@GOT, Rx; $LEA <off>(Rx), Rx
// On 386 only, more complicated things like PUSHL $sym become $MOV sym@GOT, CX; PUSHL CX
cmplxdest := false
pAs := p.As
var dest obj.Addr
if p.To.Type != obj.TYPE_REG || pAs != mov {
if p.Mode == 64 {
ctxt.Diag("do not know how to handle LEA-type insn to non-register in %v with -dynlink", p)
}
cmplxdest = true
dest = p.To
p.As = mov
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_CX
p.To.Sym = nil
p.To.Name = obj.NAME_NONE
}
p.From.Type = obj.TYPE_MEM
p.From.Name = obj.NAME_GOTREF
q := p
if p.From.Offset != 0 {
q = obj.Appendp(ctxt, p)
q.As = lea
q.From.Type = obj.TYPE_MEM
q.From.Reg = p.To.Reg
q.From.Offset = p.From.Offset
q.To = p.To
p.From.Offset = 0
}
if cmplxdest {
q = obj.Appendp(ctxt, q)
q.As = pAs
q.To = dest
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_CX
}
}
if p.From3 != nil && p.From3.Name == obj.NAME_EXTERN {
ctxt.Diag("don't know how to handle %v with -dynlink", p)
}
var source *obj.Addr
// MOVx sym, Ry becomes $MOV sym@GOT, R15; MOVx (R15), Ry
// MOVx Ry, sym becomes $MOV sym@GOT, R15; MOVx Ry, (R15)
// An addition may be inserted between the two MOVs if there is an offset.
if p.From.Name == obj.NAME_EXTERN && !p.From.Sym.Local {
if p.To.Name == obj.NAME_EXTERN && !p.To.Sym.Local {
ctxt.Diag("cannot handle NAME_EXTERN on both sides in %v with -dynlink", p)
}
source = &p.From
} else if p.To.Name == obj.NAME_EXTERN && !p.To.Sym.Local {
source = &p.To
} else {
return
}
if p.As == obj.ACALL {
// When dynlinking on 386, almost any call might end up being a call
// to a PLT, so make sure the GOT pointer is loaded into BX.
// RegTo2 is set on the replacement call insn to stop it being
// processed when it is in turn passed to progedit.
if p.Mode == 64 || (p.To.Sym != nil && p.To.Sym.Local) || p.RegTo2 != 0 {
return
}
p1 := obj.Appendp(ctxt, p)
p2 := obj.Appendp(ctxt, p1)
p1.As = ALEAL
p1.From.Type = obj.TYPE_MEM
p1.From.Name = obj.NAME_STATIC
p1.From.Sym = obj.Linklookup(ctxt, "_GLOBAL_OFFSET_TABLE_", 0)
p1.To.Type = obj.TYPE_REG
p1.To.Reg = REG_BX
p2.As = p.As
p2.Scond = p.Scond
p2.From = p.From
p2.From3 = p.From3
p2.Reg = p.Reg
p2.To = p.To
// p.To.Type was set to TYPE_BRANCH above, but that makes checkaddr
// in ../pass.go complain, so set it back to TYPE_MEM here, until p2
// itself gets passed to progedit.
p2.To.Type = obj.TYPE_MEM
p2.RegTo2 = 1
obj.Nopout(p)
return
}
if p.As == obj.ATEXT || p.As == obj.AFUNCDATA || p.As == obj.ARET || p.As == obj.AJMP {
return
}
if source.Type != obj.TYPE_MEM {
ctxt.Diag("don't know how to handle %v with -dynlink", p)
}
p1 := obj.Appendp(ctxt, p)
p2 := obj.Appendp(ctxt, p1)
p1.As = mov
p1.From.Type = obj.TYPE_MEM
p1.From.Sym = source.Sym
p1.From.Name = obj.NAME_GOTREF
p1.To.Type = obj.TYPE_REG
p1.To.Reg = reg
p2.As = p.As
p2.From = p.From
p2.To = p.To
if p.From.Name == obj.NAME_EXTERN {
p2.From.Reg = reg
p2.From.Name = obj.NAME_NONE
p2.From.Sym = nil
} else if p.To.Name == obj.NAME_EXTERN {
p2.To.Reg = reg
p2.To.Name = obj.NAME_NONE
p2.To.Sym = nil
} else {
return
}
obj.Nopout(p)
}
func xfol(ctxt *obj.Link, p *obj.Prog, last **obj.Prog) {
var q *obj.Prog
var r *obj.Prog
var a int
var i int
loop:
if p == nil {
return
}
a = int(p.As)
if a == AB {
q = p.Pcond
if q != nil && q.As != obj.ATEXT {
p.Mark |= FOLL
p = q
if p.Mark&FOLL == 0 {
goto loop
}
}
}
if p.Mark&FOLL != 0 {
i = 0
q = p
for ; i < 4; i, q = i+1, q.Link {
if q == *last || q == nil {
break
}
a = int(q.As)
if a == obj.ANOP {
i--
continue
}
if a == AB || (a == obj.ARET && q.Scond == C_SCOND_NONE) || a == ARFE || a == obj.AUNDEF {
goto copy
}
if q.Pcond == nil || (q.Pcond.Mark&FOLL != 0) {
continue
}
if a != ABEQ && a != ABNE {
continue
}
copy:
for {
r = ctxt.NewProg()
*r = *p
if r.Mark&FOLL == 0 {
fmt.Printf("can't happen 1\n")
}
r.Mark |= FOLL
if p != q {
p = p.Link
(*last).Link = r
*last = r
continue
}
(*last).Link = r
*last = r
if a == AB || (a == obj.ARET && q.Scond == C_SCOND_NONE) || a == ARFE || a == obj.AUNDEF {
return
}
r.As = ABNE
if a == ABNE {
r.As = ABEQ
}
r.Pcond = p.Link
r.Link = p.Pcond
if r.Link.Mark&FOLL == 0 {
xfol(ctxt, r.Link, last)
}
if r.Pcond.Mark&FOLL == 0 {
fmt.Printf("can't happen 2\n")
}
return
}
}
a = AB
q = ctxt.NewProg()
q.As = int16(a)
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_BRANCH
q.To.Offset = p.Pc
q.Pcond = p
p = q
}
p.Mark |= FOLL
(*last).Link = p
*last = p
if a == AB || (a == obj.ARET && p.Scond == C_SCOND_NONE) || a == ARFE || a == obj.AUNDEF {
return
}
if p.Pcond != nil {
if a != ABL && a != ABX && p.Link != nil {
q = obj.Brchain(ctxt, p.Link)
if a != obj.ATEXT {
if q != nil && (q.Mark&FOLL != 0) {
p.As = int16(relinv(a))
p.Link = p.Pcond
p.Pcond = q
}
}
xfol(ctxt, p.Link, last)
q = obj.Brchain(ctxt, p.Pcond)
if q == nil {
q = p.Pcond
}
if q.Mark&FOLL != 0 {
p.Pcond = q
return
}
p = q
goto loop
}
}
p = p.Link
goto loop
}
func aclass(ctxt *obj.Link, a *obj.Addr) int {
switch a.Type {
case obj.TYPE_NONE:
return C_NONE
case obj.TYPE_REG:
if REG_R0 <= a.Reg && a.Reg <= REG_R31 {
return C_REG
}
if REG_F0 <= a.Reg && a.Reg <= REG_F31 {
return C_FREG
}
if REG_M0 <= a.Reg && a.Reg <= REG_M31 {
return C_MREG
}
if REG_FCR0 <= a.Reg && a.Reg <= REG_FCR31 {
return C_FCREG
}
if a.Reg == REG_LO {
return C_LO
}
if a.Reg == REG_HI {
return C_HI
}
return C_GOK
case obj.TYPE_MEM:
switch a.Name {
case obj.NAME_EXTERN,
obj.NAME_STATIC:
if a.Sym == nil {
break
}
ctxt.Instoffset = a.Offset
if a.Sym != nil { // use relocation
return C_ADDR
}
return C_LEXT
case obj.NAME_AUTO:
ctxt.Instoffset = int64(ctxt.Autosize) + a.Offset
if ctxt.Instoffset >= -BIG && ctxt.Instoffset < BIG {
return C_SAUTO
}
return C_LAUTO
case obj.NAME_PARAM:
ctxt.Instoffset = int64(ctxt.Autosize) + a.Offset + 8
if ctxt.Instoffset >= -BIG && ctxt.Instoffset < BIG {
return C_SAUTO
}
return C_LAUTO
case obj.NAME_NONE:
ctxt.Instoffset = a.Offset
if ctxt.Instoffset == 0 {
return C_ZOREG
}
if ctxt.Instoffset >= -BIG && ctxt.Instoffset < BIG {
return C_SOREG
}
return C_LOREG
}
return C_GOK
case obj.TYPE_TEXTSIZE:
return C_TEXTSIZE
case obj.TYPE_CONST,
obj.TYPE_ADDR:
switch a.Name {
case obj.NAME_NONE:
ctxt.Instoffset = a.Offset
if a.Reg != 0 {
if -BIG <= ctxt.Instoffset && ctxt.Instoffset <= BIG {
return C_SACON
}
if isint32(ctxt.Instoffset) {
return C_LACON
}
return C_DACON
}
goto consize
case obj.NAME_EXTERN,
obj.NAME_STATIC:
s := a.Sym
if s == nil {
break
}
if s.Type == obj.SCONST {
ctxt.Instoffset = s.Value + a.Offset
goto consize
}
ctxt.Instoffset = s.Value + a.Offset
/* not sure why this barfs */
return C_LCON
case obj.NAME_AUTO:
ctxt.Instoffset = int64(ctxt.Autosize) + a.Offset
if ctxt.Instoffset >= -BIG && ctxt.Instoffset < BIG {
return C_SACON
}
return C_LACON
case obj.NAME_PARAM:
ctxt.Instoffset = int64(ctxt.Autosize) + a.Offset + 8
if ctxt.Instoffset >= -BIG && ctxt.Instoffset < BIG {
return C_SACON
}
return C_LACON
}
return C_GOK
consize:
if ctxt.Instoffset >= 0 {
if ctxt.Instoffset == 0 {
return C_ZCON
}
if ctxt.Instoffset <= 0x7fff {
return C_SCON
}
if ctxt.Instoffset <= 0xffff {
return C_ANDCON
}
if ctxt.Instoffset&0xffff == 0 && isuint32(uint64(ctxt.Instoffset)) { /* && (instoffset & (1<<31)) == 0) */
return C_UCON
}
if isint32(ctxt.Instoffset) || isuint32(uint64(ctxt.Instoffset)) {
return C_LCON
}
return C_LCON // C_DCON
}
if ctxt.Instoffset >= -0x8000 {
return C_ADDCON
}
if ctxt.Instoffset&0xffff == 0 && isint32(ctxt.Instoffset) {
return C_UCON
}
if isint32(ctxt.Instoffset) {
return C_LCON
}
return C_LCON // C_DCON
case obj.TYPE_BRANCH:
return C_SBRA
}
return C_GOK
}
func opirr(ctxt *obj.Link, a int) uint32 {
switch a {
case AADD:
return SP(1, 0)
case AADDU:
return SP(1, 1)
case ASGT:
return SP(1, 2)
case ASGTU:
return SP(1, 3)
case AAND:
return SP(1, 4)
case AOR:
return SP(1, 5)
case AXOR:
return SP(1, 6)
case ALAST:
return SP(1, 7) /* lui */
case ASLL:
return OP(0, 0)
case ASRL:
return OP(0, 2)
case ASRA:
return OP(0, 3)
case AADDV:
return SP(3, 0)
case AADDVU:
return SP(3, 1)
case AJMP:
return SP(0, 2)
case AJAL,
obj.ADUFFZERO,
obj.ADUFFCOPY:
return SP(0, 3)
case ABEQ:
return SP(0, 4)
case ABEQ + ALAST:
return SP(2, 4) /* likely */
case ABNE:
return SP(0, 5)
case ABNE + ALAST:
return SP(2, 5) /* likely */
case ABGEZ:
return SP(0, 1) | BCOND(0, 1)
case ABGEZ + ALAST:
return SP(0, 1) | BCOND(0, 3) /* likely */
case ABGEZAL:
return SP(0, 1) | BCOND(2, 1)
case ABGEZAL + ALAST:
return SP(0, 1) | BCOND(2, 3) /* likely */
case ABGTZ:
return SP(0, 7)
case ABGTZ + ALAST:
return SP(2, 7) /* likely */
case ABLEZ:
return SP(0, 6)
case ABLEZ + ALAST:
return SP(2, 6) /* likely */
case ABLTZ:
return SP(0, 1) | BCOND(0, 0)
case ABLTZ + ALAST:
return SP(0, 1) | BCOND(0, 2) /* likely */
case ABLTZAL:
return SP(0, 1) | BCOND(2, 0)
case ABLTZAL + ALAST:
return SP(0, 1) | BCOND(2, 2) /* likely */
case ABFPT:
return SP(2, 1) | (257 << 16)
case ABFPT + ALAST:
return SP(2, 1) | (259 << 16) /* likely */
case ABFPF:
return SP(2, 1) | (256 << 16)
case ABFPF + ALAST:
return SP(2, 1) | (258 << 16) /* likely */
case AMOVB,
AMOVBU:
return SP(5, 0)
case AMOVH,
AMOVHU:
return SP(5, 1)
case AMOVW,
AMOVWU:
return SP(5, 3)
case AMOVV:
return SP(7, 7)
case AMOVF:
return SP(7, 1)
case AMOVD:
return SP(7, 5)
case AMOVWL:
return SP(5, 2)
case AMOVWR:
return SP(5, 6)
case AMOVVL:
return SP(5, 4)
case AMOVVR:
return SP(5, 5)
case ABREAK:
return SP(5, 7)
case AMOVWL + ALAST:
return SP(4, 2)
case AMOVWR + ALAST:
return SP(4, 6)
case AMOVVL + ALAST:
return SP(3, 2)
case AMOVVR + ALAST:
return SP(3, 3)
case AMOVB + ALAST:
return SP(4, 0)
case AMOVBU + ALAST:
return SP(4, 4)
case AMOVH + ALAST:
return SP(4, 1)
case AMOVHU + ALAST:
return SP(4, 5)
case AMOVW + ALAST:
return SP(4, 3)
case AMOVWU + ALAST:
return SP(4, 7)
case AMOVV + ALAST:
return SP(6, 7)
case AMOVF + ALAST:
return SP(6, 1)
case AMOVD + ALAST:
return SP(6, 5)
case ASLLV:
return OP(7, 0)
case ASRLV:
return OP(7, 2)
case ASRAV:
return OP(7, 3)
case ASLLV + ALAST:
return OP(7, 4)
case ASRLV + ALAST:
return OP(7, 6)
case ASRAV + ALAST:
return OP(7, 7)
}
if a >= ALAST {
ctxt.Diag("bad irr opcode %v+ALAST", obj.Aconv(a-ALAST))
} else {
ctxt.Diag("bad irr opcode %v", obj.Aconv(a))
}
return 0
}
func vregoff(ctxt *obj.Link, a *obj.Addr) int64 {
ctxt.Instoffset = 0
aclass(ctxt, a)
return ctxt.Instoffset
}
func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
autosize := int32(0)
ctxt.Cursym = cursym
if cursym.Text == nil || cursym.Text.Link == nil {
return
}
softfloat(ctxt, cursym)
p := cursym.Text
autoffset := int32(p.To.Offset)
if autoffset < 0 {
autoffset = 0
}
cursym.Locals = autoffset
cursym.Args = p.To.Val.(int32)
/*
* find leaf subroutines
* strip NOPs
* expand RET
* expand BECOME pseudo
*/
var q1 *obj.Prog
var q *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
switch p.As {
case obj.ATEXT:
p.Mark |= LEAF
case obj.ARET:
break
case ADIV, ADIVU, AMOD, AMODU:
q = p
if ctxt.Sym_div == nil {
initdiv(ctxt)
}
cursym.Text.Mark &^= LEAF
continue
case obj.ANOP:
q1 = p.Link
q.Link = q1 /* q is non-nop */
if q1 != nil {
q1.Mark |= p.Mark
}
continue
case ABL,
ABX,
obj.ADUFFZERO,
obj.ADUFFCOPY:
cursym.Text.Mark &^= LEAF
fallthrough
case AB,
ABEQ,
ABNE,
ABCS,
ABHS,
ABCC,
ABLO,
ABMI,
ABPL,
ABVS,
ABVC,
ABHI,
ABLS,
ABGE,
ABLT,
ABGT,
ABLE:
q1 = p.Pcond
if q1 != nil {
for q1.As == obj.ANOP {
q1 = q1.Link
p.Pcond = q1
}
}
}
q = p
}
var o int
var p1 *obj.Prog
var p2 *obj.Prog
var q2 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
o = int(p.As)
switch o {
case obj.ATEXT:
autosize = int32(p.To.Offset + 4)
if autosize <= 4 {
if cursym.Text.Mark&LEAF != 0 {
p.To.Offset = -4
autosize = 0
}
}
if autosize == 0 && cursym.Text.Mark&LEAF == 0 {
if ctxt.Debugvlog != 0 {
fmt.Fprintf(ctxt.Bso, "save suppressed in: %s\n", cursym.Name)
ctxt.Bso.Flush()
}
cursym.Text.Mark |= LEAF
}
if cursym.Text.Mark&LEAF != 0 {
cursym.Leaf = 1
if autosize == 0 {
break
}
}
if p.From3.Offset&obj.NOSPLIT == 0 {
p = stacksplit(ctxt, p, autosize) // emit split check
}
// MOVW.W R14,$-autosize(SP)
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.Scond |= C_WBIT
p.From.Type = obj.TYPE_REG
p.From.Reg = REGLINK
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(-autosize)
p.To.Reg = REGSP
p.Spadj = autosize
if cursym.Text.From3.Offset&obj.WRAPPER != 0 {
// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
//
// MOVW g_panic(g), R1
// CMP $0, R1
// B.EQ end
// MOVW panic_argp(R1), R2
// ADD $(autosize+4), R13, R3
// CMP R2, R3
// B.NE end
// ADD $4, R13, R4
// MOVW R4, panic_argp(R1)
// end:
// NOP
//
// The NOP is needed to give the jumps somewhere to land.
// It is a liblink NOP, not an ARM NOP: it encodes to 0 instruction bytes.
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGG
p.From.Offset = 4 * int64(ctxt.Arch.Ptrsize) // G.panic
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_CONST
p.From.Offset = 0
p.Reg = REG_R1
p = obj.Appendp(ctxt, p)
p.As = ABEQ
p.To.Type = obj.TYPE_BRANCH
p1 = p
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_R1
p.From.Offset = 0 // Panic.argp
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(autosize) + 4
p.Reg = REG_R13
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R3
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R2
p.Reg = REG_R3
p = obj.Appendp(ctxt, p)
p.As = ABNE
p.To.Type = obj.TYPE_BRANCH
p2 = p
p = obj.Appendp(ctxt, p)
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = 4
p.Reg = REG_R13
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R4
p.To.Type = obj.TYPE_MEM
p.To.Reg = REG_R1
p.To.Offset = 0 // Panic.argp
p = obj.Appendp(ctxt, p)
p.As = obj.ANOP
p1.Pcond = p
p2.Pcond = p
}
case obj.ARET:
obj.Nocache(p)
if cursym.Text.Mark&LEAF != 0 {
if autosize == 0 {
p.As = AB
p.From = obj.Addr{}
if p.To.Sym != nil { // retjmp
p.To.Type = obj.TYPE_BRANCH
} else {
p.To.Type = obj.TYPE_MEM
p.To.Offset = 0
p.To.Reg = REGLINK
}
break
}
}
p.As = AMOVW
p.Scond |= C_PBIT
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(autosize)
p.From.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REGPC
// If there are instructions following
// this ARET, they come from a branch
// with the same stackframe, so no spadj.
if p.To.Sym != nil { // retjmp
p.To.Reg = REGLINK
q2 = obj.Appendp(ctxt, p)
q2.As = AB
q2.To.Type = obj.TYPE_BRANCH
q2.To.Sym = p.To.Sym
p.To.Sym = nil
p = q2
}
case AADD:
if p.From.Type == obj.TYPE_CONST && p.From.Reg == 0 && p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP {
p.Spadj = int32(-p.From.Offset)
}
case ASUB:
if p.From.Type == obj.TYPE_CONST && p.From.Reg == 0 && p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP {
p.Spadj = int32(p.From.Offset)
}
case ADIV, ADIVU, AMOD, AMODU:
if cursym.Text.From3.Offset&obj.NOSPLIT != 0 {
ctxt.Diag("cannot divide in NOSPLIT function")
}
if ctxt.Debugdivmod != 0 {
break
}
if p.From.Type != obj.TYPE_REG {
break
}
if p.To.Type != obj.TYPE_REG {
break
}
// Make copy because we overwrite p below.
q1 := *p
if q1.Reg == REGTMP || q1.Reg == 0 && q1.To.Reg == REGTMP {
ctxt.Diag("div already using REGTMP: %v", p)
}
/* MOV m(g),REGTMP */
p.As = AMOVW
p.Lineno = q1.Lineno
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGG
p.From.Offset = 6 * 4 // offset of g.m
p.Reg = 0
p.To.Type = obj.TYPE_REG
p.To.Reg = REGTMP
/* MOV a,m_divmod(REGTMP) */
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.Lineno = q1.Lineno
p.From.Type = obj.TYPE_REG
p.From.Reg = q1.From.Reg
p.To.Type = obj.TYPE_MEM
p.To.Reg = REGTMP
p.To.Offset = 8 * 4 // offset of m.divmod
/* MOV b,REGTMP */
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.Lineno = q1.Lineno
p.From.Type = obj.TYPE_REG
p.From.Reg = q1.Reg
if q1.Reg == 0 {
p.From.Reg = q1.To.Reg
}
p.To.Type = obj.TYPE_REG
p.To.Reg = REGTMP
p.To.Offset = 0
/* CALL appropriate */
p = obj.Appendp(ctxt, p)
p.As = ABL
p.Lineno = q1.Lineno
p.To.Type = obj.TYPE_BRANCH
switch o {
case ADIV:
p.To.Sym = ctxt.Sym_div
case ADIVU:
p.To.Sym = ctxt.Sym_divu
case AMOD:
p.To.Sym = ctxt.Sym_mod
case AMODU:
p.To.Sym = ctxt.Sym_modu
}
/* MOV REGTMP, b */
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.Lineno = q1.Lineno
p.From.Type = obj.TYPE_REG
p.From.Reg = REGTMP
p.From.Offset = 0
p.To.Type = obj.TYPE_REG
p.To.Reg = q1.To.Reg
case AMOVW:
if (p.Scond&C_WBIT != 0) && p.To.Type == obj.TYPE_MEM && p.To.Reg == REGSP {
p.Spadj = int32(-p.To.Offset)
}
if (p.Scond&C_PBIT != 0) && p.From.Type == obj.TYPE_MEM && p.From.Reg == REGSP && p.To.Reg != REGPC {
p.Spadj = int32(-p.From.Offset)
}
if p.From.Type == obj.TYPE_ADDR && p.From.Reg == REGSP && p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP {
p.Spadj = int32(-p.From.Offset)
}
}
}
}
func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
// TODO(minux): add morestack short-cuts with small fixed frame-size.
ctxt.Cursym = cursym
if cursym.Text == nil || cursym.Text.Link == nil {
return
}
p := cursym.Text
textstksiz := p.To.Offset
if textstksiz == -8 {
// Compatibility hack.
p.From3.Offset |= obj.NOFRAME
textstksiz = 0
}
if textstksiz%8 != 0 {
ctxt.Diag("frame size %d not a multiple of 8", textstksiz)
}
if p.From3.Offset&obj.NOFRAME != 0 {
if textstksiz != 0 {
ctxt.Diag("NOFRAME functions must have a frame size of 0, not %d", textstksiz)
}
}
cursym.Args = p.To.Val.(int32)
cursym.Locals = int32(textstksiz)
/*
* find leaf subroutines
* strip NOPs
* expand RET
* expand BECOME pseudo
*/
if ctxt.Debugvlog != 0 {
fmt.Fprintf(ctxt.Bso, "%5.2f noops\n", obj.Cputime())
}
ctxt.Bso.Flush()
var q *obj.Prog
var q1 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
switch p.As {
/* too hard, just leave alone */
case obj.ATEXT:
q = p
p.Mark |= LABEL | LEAF | SYNC
if p.Link != nil {
p.Link.Mark |= LABEL
}
case ANOR:
q = p
if p.To.Type == obj.TYPE_REG {
if p.To.Reg == REGZERO {
p.Mark |= LABEL | SYNC
}
}
case ALWAR,
ASTWCCC,
AECIWX,
AECOWX,
AEIEIO,
AICBI,
AISYNC,
ATLBIE,
ATLBIEL,
ASLBIA,
ASLBIE,
ASLBMFEE,
ASLBMFEV,
ASLBMTE,
ADCBF,
ADCBI,
ADCBST,
ADCBT,
ADCBTST,
ADCBZ,
ASYNC,
ATLBSYNC,
APTESYNC,
ATW,
AWORD,
ARFI,
ARFCI,
ARFID,
AHRFID:
q = p
p.Mark |= LABEL | SYNC
continue
case AMOVW, AMOVWZ, AMOVD:
q = p
if p.From.Reg >= REG_SPECIAL || p.To.Reg >= REG_SPECIAL {
p.Mark |= LABEL | SYNC
}
continue
case AFABS,
AFABSCC,
AFADD,
AFADDCC,
AFCTIW,
AFCTIWCC,
AFCTIWZ,
AFCTIWZCC,
AFDIV,
AFDIVCC,
AFMADD,
AFMADDCC,
AFMOVD,
AFMOVDU,
/* case AFMOVDS: */
AFMOVS,
AFMOVSU,
/* case AFMOVSD: */
AFMSUB,
AFMSUBCC,
AFMUL,
AFMULCC,
AFNABS,
AFNABSCC,
AFNEG,
AFNEGCC,
AFNMADD,
AFNMADDCC,
AFNMSUB,
AFNMSUBCC,
AFRSP,
AFRSPCC,
AFSUB,
AFSUBCC:
q = p
p.Mark |= FLOAT
continue
case ABL,
ABCL,
obj.ADUFFZERO,
obj.ADUFFCOPY:
cursym.Text.Mark &^= LEAF
fallthrough
case ABC,
ABEQ,
ABGE,
ABGT,
ABLE,
ABLT,
ABNE,
ABR,
ABVC,
ABVS:
p.Mark |= BRANCH
q = p
q1 = p.Pcond
if q1 != nil {
for q1.As == obj.ANOP {
q1 = q1.Link
p.Pcond = q1
}
if q1.Mark&LEAF == 0 {
q1.Mark |= LABEL
}
} else {
p.Mark |= LABEL
}
q1 = p.Link
if q1 != nil {
q1.Mark |= LABEL
}
continue
case AFCMPO, AFCMPU:
q = p
p.Mark |= FCMP | FLOAT
continue
case obj.ARET:
q = p
if p.Link != nil {
p.Link.Mark |= LABEL
}
continue
case obj.ANOP:
q1 = p.Link
q.Link = q1 /* q is non-nop */
q1.Mark |= p.Mark
continue
default:
q = p
continue
}
}
autosize := int32(0)
var aoffset int
var mov int
var o int
var p1 *obj.Prog
var p2 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
o = int(p.As)
switch o {
case obj.ATEXT:
mov = AMOVD
aoffset = 0
autosize = int32(textstksiz)
if p.Mark&LEAF != 0 && autosize == 0 && p.From3.Offset&obj.NOFRAME == 0 {
// A leaf function with no locals has no frame.
p.From3.Offset |= obj.NOFRAME
}
if p.From3.Offset&obj.NOFRAME == 0 {
// If there is a stack frame at all, it includes
// space to save the LR.
autosize += int32(ctxt.FixedFrameSize())
}
p.To.Offset = int64(autosize)
q = p
if ctxt.Flag_shared != 0 && cursym.Name != "runtime.duffzero" && cursym.Name != "runtime.duffcopy" {
// When compiling Go into PIC, all functions must start
// with instructions to load the TOC pointer into r2:
//
// addis r2, r12, .TOC.-func@ha
// addi r2, r2, .TOC.-func@l+4
//
// We could probably skip this prologue in some situations
// but it's a bit subtle. However, it is both safe and
// necessary to leave the prologue off duffzero and
// duffcopy as we rely on being able to jump to a specific
// instruction offset for them.
//
// These are AWORDS because there is no (afaict) way to
// generate the addis instruction except as part of the
// load of a large constant, and in that case there is no
// way to use r12 as the source.
q = obj.Appendp(ctxt, q)
q.As = AWORD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = 0x3c4c0000
q = obj.Appendp(ctxt, q)
q.As = AWORD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = 0x38420000
rel := obj.Addrel(ctxt.Cursym)
rel.Off = 0
rel.Siz = 8
rel.Sym = obj.Linklookup(ctxt, ".TOC.", 0)
rel.Type = obj.R_ADDRPOWER_PCREL
}
if cursym.Text.From3.Offset&obj.NOSPLIT == 0 {
q = stacksplit(ctxt, q, autosize) // emit split check
}
if autosize != 0 {
/* use MOVDU to adjust R1 when saving R31, if autosize is small */
if cursym.Text.Mark&LEAF == 0 && autosize >= -BIG && autosize <= BIG {
mov = AMOVDU
aoffset = int(-autosize)
} else {
q = obj.Appendp(ctxt, q)
q.As = AADD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(-autosize)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Spadj = +autosize
}
} else if cursym.Text.Mark&LEAF == 0 {
// A very few functions that do not return to their caller
// (e.g. gogo) are not identified as leaves but still have
// no frame.
cursym.Text.Mark |= LEAF
}
if cursym.Text.Mark&LEAF != 0 {
cursym.Leaf = 1
break
}
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_LR
q.To.Type = obj.TYPE_REG
q.To.Reg = REGTMP
q = obj.Appendp(ctxt, q)
q.As = int16(mov)
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_REG
q.From.Reg = REGTMP
q.To.Type = obj.TYPE_MEM
q.To.Offset = int64(aoffset)
q.To.Reg = REGSP
if q.As == AMOVDU {
q.Spadj = int32(-aoffset)
}
if ctxt.Flag_shared != 0 {
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R2
q.To.Type = obj.TYPE_MEM
q.To.Reg = REGSP
q.To.Offset = 24
}
if cursym.Text.From3.Offset&obj.WRAPPER != 0 {
// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
//
// MOVD g_panic(g), R3
// CMP R0, R3
// BEQ end
// MOVD panic_argp(R3), R4
// ADD $(autosize+8), R1, R5
// CMP R4, R5
// BNE end
// ADD $8, R1, R6
// MOVD R6, panic_argp(R3)
// end:
// NOP
//
// The NOP is needed to give the jumps somewhere to land.
// It is a liblink NOP, not a ppc64 NOP: it encodes to 0 instruction bytes.
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_MEM
q.From.Reg = REGG
q.From.Offset = 4 * int64(ctxt.Arch.Ptrsize) // G.panic
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R3
q = obj.Appendp(ctxt, q)
q.As = ACMP
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R0
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R3
q = obj.Appendp(ctxt, q)
q.As = ABEQ
q.To.Type = obj.TYPE_BRANCH
p1 = q
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_MEM
q.From.Reg = REG_R3
q.From.Offset = 0 // Panic.argp
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R4
q = obj.Appendp(ctxt, q)
q.As = AADD
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(autosize) + ctxt.FixedFrameSize()
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R5
q = obj.Appendp(ctxt, q)
q.As = ACMP
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R4
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R5
q = obj.Appendp(ctxt, q)
q.As = ABNE
q.To.Type = obj.TYPE_BRANCH
p2 = q
q = obj.Appendp(ctxt, q)
q.As = AADD
q.From.Type = obj.TYPE_CONST
q.From.Offset = ctxt.FixedFrameSize()
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R6
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R6
q.To.Type = obj.TYPE_MEM
q.To.Reg = REG_R3
q.To.Offset = 0 // Panic.argp
q = obj.Appendp(ctxt, q)
q.As = obj.ANOP
p1.Pcond = q
p2.Pcond = q
}
case obj.ARET:
if p.From.Type == obj.TYPE_CONST {
ctxt.Diag("using BECOME (%v) is not supported!", p)
break
}
retTarget := p.To.Sym
if cursym.Text.Mark&LEAF != 0 {
if autosize == 0 {
p.As = ABR
p.From = obj.Addr{}
if retTarget == nil {
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_LR
} else {
p.To.Type = obj.TYPE_BRANCH
p.To.Sym = retTarget
}
p.Mark |= BRANCH
break
}
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(autosize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REGSP
p.Spadj = -autosize
q = ctxt.NewProg()
q.As = ABR
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_LR
q.Mark |= BRANCH
q.Spadj = +autosize
q.Link = p.Link
p.Link = q
break
}
p.As = AMOVD
p.From.Type = obj.TYPE_MEM
p.From.Offset = 0
p.From.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REGTMP
q = ctxt.NewProg()
q.As = AMOVD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_REG
q.From.Reg = REGTMP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_LR
q.Link = p.Link
p.Link = q
p = q
if false {
// Debug bad returns
q = ctxt.NewProg()
q.As = AMOVD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_MEM
q.From.Offset = 0
q.From.Reg = REGTMP
q.To.Type = obj.TYPE_REG
q.To.Reg = REGTMP
q.Link = p.Link
p.Link = q
p = q
}
if autosize != 0 {
q = ctxt.NewProg()
q.As = AADD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(autosize)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Spadj = -autosize
q.Link = p.Link
p.Link = q
}
q1 = ctxt.NewProg()
q1.As = ABR
q1.Lineno = p.Lineno
if retTarget == nil {
q1.To.Type = obj.TYPE_REG
q1.To.Reg = REG_LR
} else {
q1.To.Type = obj.TYPE_BRANCH
q1.To.Sym = retTarget
}
q1.Mark |= BRANCH
q1.Spadj = +autosize
q1.Link = q.Link
q.Link = q1
case AADD:
if p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP && p.From.Type == obj.TYPE_CONST {
p.Spadj = int32(-p.From.Offset)
}
}
}
}
// Rewrite p, if necessary, to access global data via the global offset table.
func rewriteToUseGot(ctxt *obj.Link, p *obj.Prog) {
if p.As == obj.ADUFFCOPY || p.As == obj.ADUFFZERO {
// ADUFFxxx $offset
// becomes
// MOVD runtime.duffxxx@GOT, R12
// ADD $offset, R12
// MOVD R12, CTR
// BL (CTR)
var sym *obj.LSym
if p.As == obj.ADUFFZERO {
sym = obj.Linklookup(ctxt, "runtime.duffzero", 0)
} else {
sym = obj.Linklookup(ctxt, "runtime.duffcopy", 0)
}
offset := p.To.Offset
p.As = AMOVD
p.From.Type = obj.TYPE_MEM
p.From.Name = obj.NAME_GOTREF
p.From.Sym = sym
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R12
p.To.Name = obj.NAME_NONE
p.To.Offset = 0
p.To.Sym = nil
p1 := obj.Appendp(ctxt, p)
p1.As = AADD
p1.From.Type = obj.TYPE_CONST
p1.From.Offset = offset
p1.To.Type = obj.TYPE_REG
p1.To.Reg = REG_R12
p2 := obj.Appendp(ctxt, p1)
p2.As = AMOVD
p2.From.Type = obj.TYPE_REG
p2.From.Reg = REG_R12
p2.To.Type = obj.TYPE_REG
p2.To.Reg = REG_CTR
p3 := obj.Appendp(ctxt, p2)
p3.As = obj.ACALL
p3.From.Type = obj.TYPE_REG
p3.From.Reg = REG_R12
p3.To.Type = obj.TYPE_REG
p3.To.Reg = REG_CTR
}
// We only care about global data: NAME_EXTERN means a global
// symbol in the Go sense, and p.Sym.Local is true for a few
// internally defined symbols.
if p.From.Type == obj.TYPE_ADDR && p.From.Name == obj.NAME_EXTERN && !p.From.Sym.Local {
// MOVD $sym, Rx becomes MOVD sym@GOT, Rx
// MOVD $sym+<off>, Rx becomes MOVD sym@GOT, Rx; ADD <off>, Rx
if p.As != AMOVD {
ctxt.Diag("do not know how to handle TYPE_ADDR in %v with -dynlink", p)
}
if p.To.Type != obj.TYPE_REG {
ctxt.Diag("do not know how to handle LEAQ-type insn to non-register in %v with -dynlink", p)
}
p.From.Type = obj.TYPE_MEM
p.From.Name = obj.NAME_GOTREF
if p.From.Offset != 0 {
q := obj.Appendp(ctxt, p)
q.As = AADD
q.From.Type = obj.TYPE_CONST
q.From.Offset = p.From.Offset
q.To = p.To
p.From.Offset = 0
}
}
if p.From3 != nil && p.From3.Name == obj.NAME_EXTERN {
ctxt.Diag("don't know how to handle %v with -dynlink", p)
}
var source *obj.Addr
// MOVx sym, Ry becomes MOVD sym@GOT, REGTMP; MOVx (REGTMP), Ry
// MOVx Ry, sym becomes MOVD sym@GOT, REGTMP; MOVx Ry, (REGTMP)
// An addition may be inserted between the two MOVs if there is an offset.
if p.From.Name == obj.NAME_EXTERN && !p.From.Sym.Local {
if p.To.Name == obj.NAME_EXTERN && !p.To.Sym.Local {
ctxt.Diag("cannot handle NAME_EXTERN on both sides in %v with -dynlink", p)
}
source = &p.From
} else if p.To.Name == obj.NAME_EXTERN && !p.To.Sym.Local {
source = &p.To
} else {
return
}
if p.As == obj.ATEXT || p.As == obj.AFUNCDATA || p.As == obj.ACALL || p.As == obj.ARET || p.As == obj.AJMP {
return
}
if source.Sym.Type == obj.STLSBSS {
return
}
if source.Type != obj.TYPE_MEM {
ctxt.Diag("don't know how to handle %v with -dynlink", p)
}
p1 := obj.Appendp(ctxt, p)
p2 := obj.Appendp(ctxt, p1)
p1.As = AMOVD
p1.From.Type = obj.TYPE_MEM
p1.From.Sym = source.Sym
p1.From.Name = obj.NAME_GOTREF
p1.To.Type = obj.TYPE_REG
p1.To.Reg = REGTMP
p2.As = p.As
p2.From = p.From
p2.To = p.To
if p.From.Name == obj.NAME_EXTERN {
p2.From.Reg = REGTMP
p2.From.Name = obj.NAME_NONE
p2.From.Sym = nil
} else if p.To.Name == obj.NAME_EXTERN {
p2.To.Reg = REGTMP
p2.To.Name = obj.NAME_NONE
p2.To.Sym = nil
} else {
return
}
obj.Nopout(p)
}
func xfol(ctxt *obj.Link, p *obj.Prog, last **obj.Prog) {
var q *obj.Prog
var r *obj.Prog
var a int
var b int
var i int
loop:
if p == nil {
return
}
a = int(p.As)
if a == ABR {
q = p.Pcond
if (p.Mark&NOSCHED != 0) || q != nil && (q.Mark&NOSCHED != 0) {
p.Mark |= FOLL
(*last).Link = p
*last = p
p = p.Link
xfol(ctxt, p, last)
p = q
if p != nil && p.Mark&FOLL == 0 {
goto loop
}
return
}
if q != nil {
p.Mark |= FOLL
p = q
if p.Mark&FOLL == 0 {
goto loop
}
}
}
if p.Mark&FOLL != 0 {
i = 0
q = p
for ; i < 4; i, q = i+1, q.Link {
if q == *last || (q.Mark&NOSCHED != 0) {
break
}
b = 0 /* set */
a = int(q.As)
if a == obj.ANOP {
i--
continue
}
if a == ABR || a == obj.ARET || a == ARFI || a == ARFCI || a == ARFID || a == AHRFID {
goto copy
}
if q.Pcond == nil || (q.Pcond.Mark&FOLL != 0) {
continue
}
b = relinv(a)
if b == 0 {
continue
}
copy:
for {
r = ctxt.NewProg()
*r = *p
if r.Mark&FOLL == 0 {
fmt.Printf("cant happen 1\n")
}
r.Mark |= FOLL
if p != q {
p = p.Link
(*last).Link = r
*last = r
continue
}
(*last).Link = r
*last = r
if a == ABR || a == obj.ARET || a == ARFI || a == ARFCI || a == ARFID || a == AHRFID {
return
}
r.As = int16(b)
r.Pcond = p.Link
r.Link = p.Pcond
if r.Link.Mark&FOLL == 0 {
xfol(ctxt, r.Link, last)
}
if r.Pcond.Mark&FOLL == 0 {
fmt.Printf("cant happen 2\n")
}
return
}
}
a = ABR
q = ctxt.NewProg()
q.As = int16(a)
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_BRANCH
q.To.Offset = p.Pc
q.Pcond = p
p = q
}
p.Mark |= FOLL
(*last).Link = p
*last = p
if a == ABR || a == obj.ARET || a == ARFI || a == ARFCI || a == ARFID || a == AHRFID {
if p.Mark&NOSCHED != 0 {
p = p.Link
goto loop
}
return
}
if p.Pcond != nil {
if a != ABL && p.Link != nil {
xfol(ctxt, p.Link, last)
p = p.Pcond
if p == nil || (p.Mark&FOLL != 0) {
return
}
goto loop
}
}
p = p.Link
goto loop
}
func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
ctxt.Cursym = cursym
if cursym.Text == nil || cursym.Text.Link == nil {
return
}
p := cursym.Text
textstksiz := p.To.Offset
aoffset := int32(textstksiz)
cursym.Args = p.To.Val.(int32)
cursym.Locals = int32(textstksiz)
/*
* find leaf subroutines
* strip NOPs
* expand RET
*/
q := (*obj.Prog)(nil)
var q1 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
switch p.As {
case obj.ATEXT:
p.Mark |= LEAF
case obj.ARET:
break
case obj.ANOP:
q1 = p.Link
q.Link = q1 /* q is non-nop */
q1.Mark |= p.Mark
continue
case ABL,
obj.ADUFFZERO,
obj.ADUFFCOPY:
cursym.Text.Mark &^= LEAF
fallthrough
case ACBNZ,
ACBZ,
ACBNZW,
ACBZW,
ATBZ,
ATBNZ,
AB,
ABEQ,
ABNE,
ABCS,
ABHS,
ABCC,
ABLO,
ABMI,
ABPL,
ABVS,
ABVC,
ABHI,
ABLS,
ABGE,
ABLT,
ABGT,
ABLE,
AADR, /* strange */
AADRP:
q1 = p.Pcond
if q1 != nil {
for q1.As == obj.ANOP {
q1 = q1.Link
p.Pcond = q1
}
}
break
}
q = p
}
var q2 *obj.Prog
var retjmp *obj.LSym
for p := cursym.Text; p != nil; p = p.Link {
o := p.As
switch o {
case obj.ATEXT:
cursym.Text = p
if textstksiz < 0 {
ctxt.Autosize = 0
} else {
ctxt.Autosize = int32(textstksiz + 8)
}
if (cursym.Text.Mark&LEAF != 0) && ctxt.Autosize <= 8 {
ctxt.Autosize = 0
} else if ctxt.Autosize&(16-1) != 0 {
// The frame includes an LR.
// If the frame size is 8, it's only an LR,
// so there's no potential for breaking references to
// local variables by growing the frame size,
// because there are no local variables.
// But otherwise, if there is a non-empty locals section,
// the author of the code is responsible for making sure
// that the frame size is 8 mod 16.
if ctxt.Autosize == 8 {
ctxt.Autosize += 8
cursym.Locals += 8
} else {
ctxt.Diag("%v: unaligned frame size %d - must be 8 mod 16 (or 0)", p, ctxt.Autosize-8)
}
}
p.To.Offset = int64(ctxt.Autosize) - 8
if ctxt.Autosize == 0 && !(cursym.Text.Mark&LEAF != 0) {
if ctxt.Debugvlog != 0 {
ctxt.Logf("save suppressed in: %s\n", cursym.Text.From.Sym.Name)
}
cursym.Text.Mark |= LEAF
}
if !(p.From3.Offset&obj.NOSPLIT != 0) {
p = stacksplit(ctxt, p, ctxt.Autosize) // emit split check
}
aoffset = ctxt.Autosize
if aoffset > 0xF0 {
aoffset = 0xF0
}
if cursym.Text.Mark&LEAF != 0 {
cursym.Leaf = true
if ctxt.Autosize == 0 {
break
}
aoffset = 0
}
q = p
if ctxt.Autosize > aoffset {
q = ctxt.NewProg()
q.As = ASUB
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(ctxt.Autosize) - int64(aoffset)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Spadj = int32(q.From.Offset)
q.Link = p.Link
p.Link = q
if cursym.Text.Mark&LEAF != 0 {
break
}
}
q1 = ctxt.NewProg()
q1.As = AMOVD
q1.Lineno = p.Lineno
q1.From.Type = obj.TYPE_REG
q1.From.Reg = REGLINK
q1.To.Type = obj.TYPE_MEM
q1.Scond = C_XPRE
q1.To.Offset = int64(-aoffset)
q1.To.Reg = REGSP
q1.Link = q.Link
q1.Spadj = aoffset
q.Link = q1
if cursym.Text.From3.Offset&obj.WRAPPER != 0 {
// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
//
// MOV g_panic(g), R1
// CMP ZR, R1
// BEQ end
// MOV panic_argp(R1), R2
// ADD $(autosize+8), RSP, R3
// CMP R2, R3
// BNE end
// ADD $8, RSP, R4
// MOVD R4, panic_argp(R1)
// end:
// NOP
//
// The NOP is needed to give the jumps somewhere to land.
// It is a liblink NOP, not a ARM64 NOP: it encodes to 0 instruction bytes.
q = q1
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_MEM
q.From.Reg = REGG
q.From.Offset = 4 * int64(ctxt.Arch.PtrSize) // G.panic
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R1
q = obj.Appendp(ctxt, q)
q.As = ACMP
q.From.Type = obj.TYPE_REG
q.From.Reg = REGZERO
q.Reg = REG_R1
q = obj.Appendp(ctxt, q)
q.As = ABEQ
q.To.Type = obj.TYPE_BRANCH
q1 = q
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_MEM
q.From.Reg = REG_R1
q.From.Offset = 0 // Panic.argp
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R2
q = obj.Appendp(ctxt, q)
q.As = AADD
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(ctxt.Autosize) + 8
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R3
q = obj.Appendp(ctxt, q)
q.As = ACMP
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R2
q.Reg = REG_R3
q = obj.Appendp(ctxt, q)
q.As = ABNE
q.To.Type = obj.TYPE_BRANCH
q2 = q
q = obj.Appendp(ctxt, q)
q.As = AADD
q.From.Type = obj.TYPE_CONST
q.From.Offset = 8
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R4
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R4
q.To.Type = obj.TYPE_MEM
q.To.Reg = REG_R1
q.To.Offset = 0 // Panic.argp
q = obj.Appendp(ctxt, q)
q.As = obj.ANOP
q1.Pcond = q
q2.Pcond = q
}
case obj.ARET:
nocache(p)
if p.From.Type == obj.TYPE_CONST {
ctxt.Diag("using BECOME (%v) is not supported!", p)
break
}
retjmp = p.To.Sym
p.To = obj.Addr{}
if cursym.Text.Mark&LEAF != 0 {
if ctxt.Autosize != 0 {
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(ctxt.Autosize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REGSP
p.Spadj = -ctxt.Autosize
}
} else {
/* want write-back pre-indexed SP+autosize -> SP, loading REGLINK*/
aoffset = ctxt.Autosize
if aoffset > 0xF0 {
aoffset = 0xF0
}
p.As = AMOVD
p.From.Type = obj.TYPE_MEM
p.Scond = C_XPOST
p.From.Offset = int64(aoffset)
p.From.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REGLINK
p.Spadj = -aoffset
if ctxt.Autosize > aoffset {
q = ctxt.NewProg()
q.As = AADD
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(ctxt.Autosize) - int64(aoffset)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Link = p.Link
q.Spadj = int32(-q.From.Offset)
q.Lineno = p.Lineno
p.Link = q
p = q
}
}
if p.As != obj.ARET {
q = ctxt.NewProg()
q.Lineno = p.Lineno
q.Link = p.Link
p.Link = q
p = q
}
if retjmp != nil { // retjmp
p.As = AB
p.To.Type = obj.TYPE_BRANCH
p.To.Sym = retjmp
p.Spadj = +ctxt.Autosize
break
}
p.As = obj.ARET
p.To.Type = obj.TYPE_MEM
p.To.Offset = 0
p.To.Reg = REGLINK
p.Spadj = +ctxt.Autosize
case AADD, ASUB:
if p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP && p.From.Type == obj.TYPE_CONST {
if p.As == AADD {
p.Spadj = int32(-p.From.Offset)
} else {
p.Spadj = int32(+p.From.Offset)
}
}
break
}
}
}
func opirr(ctxt *obj.Link, a obj.As) uint32 {
switch a {
case AADD:
return SP(1, 0)
case AADDU:
return SP(1, 1)
case ASGT:
return SP(1, 2)
case ASGTU:
return SP(1, 3)
case AAND:
return SP(1, 4)
case AOR:
return SP(1, 5)
case AXOR:
return SP(1, 6)
case ALUI:
return SP(1, 7)
case ASLL:
return OP(0, 0)
case ASRL:
return OP(0, 2)
case ASRA:
return OP(0, 3)
case AADDV:
return SP(3, 0)
case AADDVU:
return SP(3, 1)
case AJMP:
return SP(0, 2)
case AJAL,
obj.ADUFFZERO,
obj.ADUFFCOPY:
return SP(0, 3)
case ABEQ:
return SP(0, 4)
case -ABEQ:
return SP(2, 4) /* likely */
case ABNE:
return SP(0, 5)
case -ABNE:
return SP(2, 5) /* likely */
case ABGEZ:
return SP(0, 1) | BCOND(0, 1)
case -ABGEZ:
return SP(0, 1) | BCOND(0, 3) /* likely */
case ABGEZAL:
return SP(0, 1) | BCOND(2, 1)
case -ABGEZAL:
return SP(0, 1) | BCOND(2, 3) /* likely */
case ABGTZ:
return SP(0, 7)
case -ABGTZ:
return SP(2, 7) /* likely */
case ABLEZ:
return SP(0, 6)
case -ABLEZ:
return SP(2, 6) /* likely */
case ABLTZ:
return SP(0, 1) | BCOND(0, 0)
case -ABLTZ:
return SP(0, 1) | BCOND(0, 2) /* likely */
case ABLTZAL:
return SP(0, 1) | BCOND(2, 0)
case -ABLTZAL:
return SP(0, 1) | BCOND(2, 2) /* likely */
case ABFPT:
return SP(2, 1) | (257 << 16)
case -ABFPT:
return SP(2, 1) | (259 << 16) /* likely */
case ABFPF:
return SP(2, 1) | (256 << 16)
case -ABFPF:
return SP(2, 1) | (258 << 16) /* likely */
case AMOVB,
AMOVBU:
return SP(5, 0)
case AMOVH,
AMOVHU:
return SP(5, 1)
case AMOVW,
AMOVWU:
return SP(5, 3)
case AMOVV:
return SP(7, 7)
case AMOVF:
return SP(7, 1)
case AMOVD:
return SP(7, 5)
case AMOVWL:
return SP(5, 2)
case AMOVWR:
return SP(5, 6)
case AMOVVL:
return SP(5, 4)
case AMOVVR:
return SP(5, 5)
case ABREAK:
return SP(5, 7)
case -AMOVWL:
return SP(4, 2)
case -AMOVWR:
return SP(4, 6)
case -AMOVVL:
return SP(3, 2)
case -AMOVVR:
return SP(3, 3)
case -AMOVB:
return SP(4, 0)
case -AMOVBU:
return SP(4, 4)
case -AMOVH:
return SP(4, 1)
case -AMOVHU:
return SP(4, 5)
case -AMOVW:
return SP(4, 3)
case -AMOVWU:
return SP(4, 7)
case -AMOVV:
return SP(6, 7)
case -AMOVF:
return SP(6, 1)
case -AMOVD:
return SP(6, 5)
case ASLLV:
return OP(7, 0)
case ASRLV:
return OP(7, 2)
case ASRAV:
return OP(7, 3)
case -ASLLV:
return OP(7, 4)
case -ASRLV:
return OP(7, 6)
case -ASRAV:
return OP(7, 7)
case ATEQ:
return OP(6, 4)
case ATNE:
return OP(6, 6)
case -ALL:
return SP(6, 0)
case ASC:
return SP(7, 0)
}
if a < 0 {
ctxt.Diag("bad irr opcode -%v", -a)
} else {
ctxt.Diag("bad irr opcode %v", a)
}
return 0
}
func asmout(ctxt *obj.Link, p *obj.Prog, o *Optab, out []uint32) {
o1 := uint32(0)
o2 := uint32(0)
o3 := uint32(0)
o4 := uint32(0)
add := AADDU
if ctxt.Mode&Mips64 != 0 {
add = AADDVU
}
switch o.type_ {
default:
ctxt.Diag("unknown type %d %v", o.type_)
prasm(p)
case 0: /* pseudo ops */
break
case 1: /* mov r1,r2 ==> OR r1,r0,r2 */
a := AOR
if p.As == AMOVW && ctxt.Mode&Mips64 != 0 {
a = AADDU // sign-extended to high 32 bits
}
o1 = OP_RRR(oprrr(ctxt, a), uint32(REGZERO), uint32(p.From.Reg), uint32(p.To.Reg))
case 2: /* add/sub r1,[r2],r3 */
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o1 = OP_RRR(oprrr(ctxt, p.As), uint32(p.From.Reg), uint32(r), uint32(p.To.Reg))
case 3: /* mov $soreg, r ==> or/add $i,o,r */
v := regoff(ctxt, &p.From)
r := int(p.From.Reg)
if r == 0 {
r = int(o.param)
}
a := add
if o.a1 == C_ANDCON {
a = AOR
}
o1 = OP_IRR(opirr(ctxt, a), uint32(v), uint32(r), uint32(p.To.Reg))
case 4: /* add $scon,[r1],r2 */
v := regoff(ctxt, &p.From)
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o1 = OP_IRR(opirr(ctxt, p.As), uint32(v), uint32(r), uint32(p.To.Reg))
case 5: /* syscall */
o1 = oprrr(ctxt, p.As)
case 6: /* beq r1,[r2],sbra */
v := int32(0)
if p.Pcond == nil {
v = int32(-4) >> 2
} else {
v = int32(p.Pcond.Pc-p.Pc-4) >> 2
}
if (v<<16)>>16 != v {
ctxt.Diag("short branch too far\n%v", p)
}
o1 = OP_IRR(opirr(ctxt, p.As), uint32(v), uint32(p.From.Reg), uint32(p.Reg))
// for ABFPT and ABFPF only: always fill delay slot with 0
// see comments in func preprocess for details.
o2 = 0
case 7: /* mov r, soreg ==> sw o(r) */
r := int(p.To.Reg)
if r == 0 {
r = int(o.param)
}
v := regoff(ctxt, &p.To)
o1 = OP_IRR(opirr(ctxt, p.As), uint32(v), uint32(r), uint32(p.From.Reg))
case 8: /* mov soreg, r ==> lw o(r) */
r := int(p.From.Reg)
if r == 0 {
r = int(o.param)
}
v := regoff(ctxt, &p.From)
o1 = OP_IRR(opirr(ctxt, -p.As), uint32(v), uint32(r), uint32(p.To.Reg))
case 9: /* sll r1,[r2],r3 */
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o1 = OP_RRR(oprrr(ctxt, p.As), uint32(r), uint32(p.From.Reg), uint32(p.To.Reg))
case 10: /* add $con,[r1],r2 ==> mov $con, t; add t,[r1],r2 */
v := regoff(ctxt, &p.From)
a := AOR
if v < 0 {
a = AADDU
}
o1 = OP_IRR(opirr(ctxt, a), uint32(v), uint32(0), uint32(REGTMP))
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o2 = OP_RRR(oprrr(ctxt, p.As), uint32(REGTMP), uint32(r), uint32(p.To.Reg))
case 11: /* jmp lbra */
v := int32(0)
if aclass(ctxt, &p.To) == C_SBRA && p.To.Sym == nil && p.As == AJMP {
// use PC-relative branch for short branches
// BEQ R0, R0, sbra
if p.Pcond == nil {
v = int32(-4) >> 2
} else {
v = int32(p.Pcond.Pc-p.Pc-4) >> 2
}
if (v<<16)>>16 == v {
o1 = OP_IRR(opirr(ctxt, ABEQ), uint32(v), uint32(REGZERO), uint32(REGZERO))
break
}
}
if p.Pcond == nil {
v = int32(p.Pc) >> 2
} else {
v = int32(p.Pcond.Pc) >> 2
}
o1 = OP_JMP(opirr(ctxt, p.As), uint32(v))
if p.To.Sym == nil {
p.To.Sym = ctxt.Cursym.Text.From.Sym
p.To.Offset = p.Pcond.Pc
}
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc)
rel.Siz = 4
rel.Sym = p.To.Sym
rel.Add = p.To.Offset
if p.As == AJAL {
rel.Type = obj.R_CALLMIPS
} else {
rel.Type = obj.R_JMPMIPS
}
case 12: /* movbs r,r */
v := 16
if p.As == AMOVB {
v = 24
}
o1 = OP_SRR(opirr(ctxt, ASLL), uint32(v), uint32(p.From.Reg), uint32(p.To.Reg))
o2 = OP_SRR(opirr(ctxt, ASRA), uint32(v), uint32(p.To.Reg), uint32(p.To.Reg))
case 13: /* movbu r,r */
if p.As == AMOVBU {
o1 = OP_IRR(opirr(ctxt, AAND), uint32(0xff), uint32(p.From.Reg), uint32(p.To.Reg))
} else {
o1 = OP_IRR(opirr(ctxt, AAND), uint32(0xffff), uint32(p.From.Reg), uint32(p.To.Reg))
}
case 14: /* movwu r,r */
o1 = OP_SRR(opirr(ctxt, -ASLLV), uint32(0), uint32(p.From.Reg), uint32(p.To.Reg))
o2 = OP_SRR(opirr(ctxt, -ASRLV), uint32(0), uint32(p.To.Reg), uint32(p.To.Reg))
case 15: /* teq $c r,r */
v := regoff(ctxt, &p.From)
r := int(p.Reg)
if r == 0 {
r = REGZERO
}
/* only use 10 bits of trap code */
o1 = OP_IRR(opirr(ctxt, p.As), (uint32(v)&0x3FF)<<6, uint32(p.Reg), uint32(p.To.Reg))
case 16: /* sll $c,[r1],r2 */
v := regoff(ctxt, &p.From)
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
/* OP_SRR will use only the low 5 bits of the shift value */
if v >= 32 && vshift(p.As) {
o1 = OP_SRR(opirr(ctxt, -p.As), uint32(v-32), uint32(r), uint32(p.To.Reg))
} else {
o1 = OP_SRR(opirr(ctxt, p.As), uint32(v), uint32(r), uint32(p.To.Reg))
}
case 17:
o1 = OP_RRR(oprrr(ctxt, p.As), uint32(REGZERO), uint32(p.From.Reg), uint32(p.To.Reg))
case 18: /* jmp [r1],0(r2) */
r := int(p.Reg)
if r == 0 {
r = int(o.param)
}
o1 = OP_RRR(oprrr(ctxt, p.As), uint32(0), uint32(p.To.Reg), uint32(r))
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc)
rel.Siz = 0
rel.Type = obj.R_CALLIND
case 19: /* mov $lcon,r ==> lu+or */
v := regoff(ctxt, &p.From)
o1 = OP_IRR(opirr(ctxt, ALUI), uint32(v>>16), uint32(REGZERO), uint32(p.To.Reg))
o2 = OP_IRR(opirr(ctxt, AOR), uint32(v), uint32(p.To.Reg), uint32(p.To.Reg))
case 20: /* mov lo/hi,r */
a := OP(2, 0) /* mfhi */
if p.From.Reg == REG_LO {
a = OP(2, 2) /* mflo */
}
o1 = OP_RRR(a, uint32(REGZERO), uint32(REGZERO), uint32(p.To.Reg))
case 21: /* mov r,lo/hi */
a := OP(2, 1) /* mthi */
if p.To.Reg == REG_LO {
a = OP(2, 3) /* mtlo */
}
o1 = OP_RRR(a, uint32(REGZERO), uint32(p.From.Reg), uint32(REGZERO))
case 22: /* mul r1,r2 [r3]*/
if p.To.Reg != 0 {
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
a := SP(3, 4) | 2 /* mul */
o1 = OP_RRR(a, uint32(p.From.Reg), uint32(r), uint32(p.To.Reg))
} else {
o1 = OP_RRR(oprrr(ctxt, p.As), uint32(p.From.Reg), uint32(p.Reg), uint32(REGZERO))
}
case 23: /* add $lcon,r1,r2 ==> lu+or+add */
v := regoff(ctxt, &p.From)
o1 = OP_IRR(opirr(ctxt, ALUI), uint32(v>>16), uint32(REGZERO), uint32(REGTMP))
o2 = OP_IRR(opirr(ctxt, AOR), uint32(v), uint32(REGTMP), uint32(REGTMP))
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o3 = OP_RRR(oprrr(ctxt, p.As), uint32(REGTMP), uint32(r), uint32(p.To.Reg))
case 24: /* mov $ucon,r ==> lu r */
v := regoff(ctxt, &p.From)
o1 = OP_IRR(opirr(ctxt, ALUI), uint32(v>>16), uint32(REGZERO), uint32(p.To.Reg))
case 25: /* add/and $ucon,[r1],r2 ==> lu $con,t; add t,[r1],r2 */
v := regoff(ctxt, &p.From)
o1 = OP_IRR(opirr(ctxt, ALUI), uint32(v>>16), uint32(REGZERO), uint32(REGTMP))
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o2 = OP_RRR(oprrr(ctxt, p.As), uint32(REGTMP), uint32(r), uint32(p.To.Reg))
case 26: /* mov $lsext/auto/oreg,r ==> lu+or+add */
v := regoff(ctxt, &p.From)
o1 = OP_IRR(opirr(ctxt, ALUI), uint32(v>>16), uint32(REGZERO), uint32(REGTMP))
o2 = OP_IRR(opirr(ctxt, AOR), uint32(v), uint32(REGTMP), uint32(REGTMP))
r := int(p.From.Reg)
if r == 0 {
r = int(o.param)
}
o3 = OP_RRR(oprrr(ctxt, add), uint32(REGTMP), uint32(r), uint32(p.To.Reg))
case 27: /* mov [sl]ext/auto/oreg,fr ==> lwc1 o(r) */
v := regoff(ctxt, &p.From)
r := int(p.From.Reg)
if r == 0 {
r = int(o.param)
}
a := -AMOVF
if p.As == AMOVD {
a = -AMOVD
}
switch o.size {
case 12:
o1 = OP_IRR(opirr(ctxt, ALUI), uint32((v+1<<15)>>16), uint32(REGZERO), uint32(REGTMP))
o2 = OP_RRR(oprrr(ctxt, add), uint32(r), uint32(REGTMP), uint32(REGTMP))
o3 = OP_IRR(opirr(ctxt, a), uint32(v), uint32(REGTMP), uint32(p.To.Reg))
case 4:
o1 = OP_IRR(opirr(ctxt, a), uint32(v), uint32(r), uint32(p.To.Reg))
}
case 28: /* mov fr,[sl]ext/auto/oreg ==> swc1 o(r) */
v := regoff(ctxt, &p.To)
r := int(p.To.Reg)
if r == 0 {
r = int(o.param)
}
a := AMOVF
if p.As == AMOVD {
a = AMOVD
}
switch o.size {
case 12:
o1 = OP_IRR(opirr(ctxt, ALUI), uint32((v+1<<15)>>16), uint32(REGZERO), uint32(REGTMP))
o2 = OP_RRR(oprrr(ctxt, add), uint32(r), uint32(REGTMP), uint32(REGTMP))
o3 = OP_IRR(opirr(ctxt, a), uint32(v), uint32(REGTMP), uint32(p.From.Reg))
case 4:
o1 = OP_IRR(opirr(ctxt, a), uint32(v), uint32(r), uint32(p.From.Reg))
}
case 30: /* movw r,fr */
a := SP(2, 1) | (4 << 21) /* mtc1 */
o1 = OP_RRR(a, uint32(p.From.Reg), uint32(0), uint32(p.To.Reg))
case 31: /* movw fr,r */
a := SP(2, 1) | (0 << 21) /* mtc1 */
o1 = OP_RRR(a, uint32(p.To.Reg), uint32(0), uint32(p.From.Reg))
case 32: /* fadd fr1,[fr2],fr3 */
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o1 = OP_FRRR(oprrr(ctxt, p.As), uint32(p.From.Reg), uint32(r), uint32(p.To.Reg))
case 33: /* fabs fr1, fr3 */
o1 = OP_FRRR(oprrr(ctxt, p.As), uint32(0), uint32(p.From.Reg), uint32(p.To.Reg))
case 34: /* mov $con,fr ==> or/add $i,t; mov t,fr */
v := regoff(ctxt, &p.From)
a := AADDU
if o.a1 == C_ANDCON {
a = AOR
}
o1 = OP_IRR(opirr(ctxt, a), uint32(v), uint32(0), uint32(REGTMP))
o2 = OP_RRR(SP(2, 1)|(4<<21), uint32(REGTMP), uint32(0), uint32(p.To.Reg)) /* mtc1 */
case 35: /* mov r,lext/auto/oreg ==> sw o(REGTMP) */
v := regoff(ctxt, &p.To)
r := int(p.To.Reg)
if r == 0 {
r = int(o.param)
}
o1 = OP_IRR(opirr(ctxt, ALUI), uint32((v+1<<15)>>16), uint32(REGZERO), uint32(REGTMP))
o2 = OP_RRR(oprrr(ctxt, add), uint32(r), uint32(REGTMP), uint32(REGTMP))
o3 = OP_IRR(opirr(ctxt, p.As), uint32(v), uint32(REGTMP), uint32(p.From.Reg))
case 36: /* mov lext/auto/oreg,r ==> lw o(REGTMP) */
v := regoff(ctxt, &p.From)
r := int(p.From.Reg)
if r == 0 {
r = int(o.param)
}
o1 = OP_IRR(opirr(ctxt, ALUI), uint32((v+1<<15)>>16), uint32(REGZERO), uint32(REGTMP))
o2 = OP_RRR(oprrr(ctxt, add), uint32(r), uint32(REGTMP), uint32(REGTMP))
o3 = OP_IRR(opirr(ctxt, -p.As), uint32(v), uint32(REGTMP), uint32(p.To.Reg))
case 37: /* movw r,mr */
a := SP(2, 0) | (4 << 21) /* mtc0 */
if p.As == AMOVV {
a = SP(2, 0) | (5 << 21) /* dmtc0 */
}
o1 = OP_RRR(a, uint32(p.From.Reg), uint32(0), uint32(p.To.Reg))
case 38: /* movw mr,r */
a := SP(2, 0) | (0 << 21) /* mfc0 */
if p.As == AMOVV {
a = SP(2, 0) | (1 << 21) /* dmfc0 */
}
o1 = OP_RRR(a, uint32(p.To.Reg), uint32(0), uint32(p.From.Reg))
case 40: /* word */
o1 = uint32(regoff(ctxt, &p.From))
case 41: /* movw f,fcr */
o1 = OP_RRR(SP(2, 1)|(2<<21), uint32(REGZERO), uint32(0), uint32(p.To.Reg)) /* mfcc1 */
o2 = OP_RRR(SP(2, 1)|(6<<21), uint32(p.From.Reg), uint32(0), uint32(p.To.Reg)) /* mtcc1 */
case 42: /* movw fcr,r */
o1 = OP_RRR(SP(2, 1)|(2<<21), uint32(p.To.Reg), uint32(0), uint32(p.From.Reg)) /* mfcc1 */
case 47: /* movv r,fr */
a := SP(2, 1) | (5 << 21) /* dmtc1 */
o1 = OP_RRR(a, uint32(p.From.Reg), uint32(0), uint32(p.To.Reg))
case 48: /* movv fr,r */
a := SP(2, 1) | (1 << 21) /* dmtc1 */
o1 = OP_RRR(a, uint32(p.To.Reg), uint32(0), uint32(p.From.Reg))
case 49: /* undef */
o1 = 52 /* trap -- teq r0, r0 */
/* relocation operations */
case 50: /* mov r,addr ==> lu + add REGSB, REGTMP + sw o(REGTMP) */
o1 = OP_IRR(opirr(ctxt, ALUI), uint32(0), uint32(REGZERO), uint32(REGTMP))
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc)
rel.Siz = 4
rel.Sym = p.To.Sym
rel.Add = p.To.Offset
rel.Type = obj.R_ADDRMIPSU
o2 = OP_IRR(opirr(ctxt, p.As), uint32(0), uint32(REGTMP), uint32(p.From.Reg))
rel2 := obj.Addrel(ctxt.Cursym)
rel2.Off = int32(ctxt.Pc + 4)
rel2.Siz = 4
rel2.Sym = p.To.Sym
rel2.Add = p.To.Offset
rel2.Type = obj.R_ADDRMIPS
if o.size == 12 {
o3 = o2
o2 = OP_RRR(oprrr(ctxt, AADDVU), uint32(REGSB), uint32(REGTMP), uint32(REGTMP))
rel2.Off += 4
}
case 51: /* mov addr,r ==> lu + add REGSB, REGTMP + lw o(REGTMP) */
o1 = OP_IRR(opirr(ctxt, ALUI), uint32(0), uint32(REGZERO), uint32(REGTMP))
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc)
rel.Siz = 4
rel.Sym = p.From.Sym
rel.Add = p.From.Offset
rel.Type = obj.R_ADDRMIPSU
o2 = OP_IRR(opirr(ctxt, -p.As), uint32(0), uint32(REGTMP), uint32(p.To.Reg))
rel2 := obj.Addrel(ctxt.Cursym)
rel2.Off = int32(ctxt.Pc + 4)
rel2.Siz = 4
rel2.Sym = p.From.Sym
rel2.Add = p.From.Offset
rel2.Type = obj.R_ADDRMIPS
if o.size == 12 {
o3 = o2
o2 = OP_RRR(oprrr(ctxt, AADDVU), uint32(REGSB), uint32(REGTMP), uint32(REGTMP))
rel2.Off += 4
}
case 52: /* mov $lext, r ==> lu + add REGSB, r + add */
o1 = OP_IRR(opirr(ctxt, ALUI), uint32(0), uint32(REGZERO), uint32(p.To.Reg))
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc)
rel.Siz = 4
rel.Sym = p.From.Sym
rel.Add = p.From.Offset
rel.Type = obj.R_ADDRMIPSU
o2 = OP_IRR(opirr(ctxt, add), uint32(0), uint32(p.To.Reg), uint32(p.To.Reg))
rel2 := obj.Addrel(ctxt.Cursym)
rel2.Off = int32(ctxt.Pc + 4)
rel2.Siz = 4
rel2.Sym = p.From.Sym
rel2.Add = p.From.Offset
rel2.Type = obj.R_ADDRMIPS
if o.size == 12 {
o3 = o2
o2 = OP_RRR(oprrr(ctxt, AADDVU), uint32(REGSB), uint32(p.To.Reg), uint32(p.To.Reg))
rel2.Off += 4
}
case 53: /* mov r, tlsvar ==> rdhwr + sw o(r3) */
// clobbers R3 !
// load thread pointer with RDHWR, R3 is used for fast kernel emulation on Linux
o1 = (037<<26 + 073) | (29 << 11) | (3 << 16) // rdhwr $29, r3
o2 = OP_IRR(opirr(ctxt, p.As), uint32(0), uint32(REG_R3), uint32(p.From.Reg))
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc + 4)
rel.Siz = 4
rel.Sym = p.To.Sym
rel.Add = p.To.Offset
rel.Type = obj.R_ADDRMIPSTLS
case 54: /* mov tlsvar, r ==> rdhwr + lw o(r3) */
// clobbers R3 !
o1 = (037<<26 + 073) | (29 << 11) | (3 << 16) // rdhwr $29, r3
o2 = OP_IRR(opirr(ctxt, -p.As), uint32(0), uint32(REG_R3), uint32(p.To.Reg))
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc + 4)
rel.Siz = 4
rel.Sym = p.From.Sym
rel.Add = p.From.Offset
rel.Type = obj.R_ADDRMIPSTLS
case 55: /* mov $tlsvar, r ==> rdhwr + add */
// clobbers R3 !
o1 = (037<<26 + 073) | (29 << 11) | (3 << 16) // rdhwr $29, r3
o2 = OP_IRR(opirr(ctxt, add), uint32(0), uint32(REG_R3), uint32(p.To.Reg))
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc + 4)
rel.Siz = 4
rel.Sym = p.From.Sym
rel.Add = p.From.Offset
rel.Type = obj.R_ADDRMIPSTLS
}
out[0] = o1
out[1] = o2
out[2] = o3
out[3] = o4
return
}
func progedit(ctxt *obj.Link, p *obj.Prog) {
// Maintain information about code generation mode.
if ctxt.Mode == 0 {
ctxt.Mode = ctxt.Arch.Regsize * 8
}
p.Mode = int8(ctxt.Mode)
switch p.As {
case AMODE:
if p.From.Type == obj.TYPE_CONST || (p.From.Type == obj.TYPE_MEM && p.From.Reg == REG_NONE) {
switch int(p.From.Offset) {
case 16, 32, 64:
ctxt.Mode = int(p.From.Offset)
}
}
obj.Nopout(p)
}
// Thread-local storage references use the TLS pseudo-register.
// As a register, TLS refers to the thread-local storage base, and it
// can only be loaded into another register:
//
// MOVQ TLS, AX
//
// An offset from the thread-local storage base is written off(reg)(TLS*1).
// Semantically it is off(reg), but the (TLS*1) annotation marks this as
// indexing from the loaded TLS base. This emits a relocation so that
// if the linker needs to adjust the offset, it can. For example:
//
// MOVQ TLS, AX
// MOVQ 0(AX)(TLS*1), CX // load g into CX
//
// On systems that support direct access to the TLS memory, this
// pair of instructions can be reduced to a direct TLS memory reference:
//
// MOVQ 0(TLS), CX // load g into CX
//
// The 2-instruction and 1-instruction forms correspond to the two code
// sequences for loading a TLS variable in the local exec model given in "ELF
// Handling For Thread-Local Storage".
//
// We apply this rewrite on systems that support the 1-instruction form.
// The decision is made using only the operating system and the -shared flag,
// not the link mode. If some link modes on a particular operating system
// require the 2-instruction form, then all builds for that operating system
// will use the 2-instruction form, so that the link mode decision can be
// delayed to link time.
//
// In this way, all supported systems use identical instructions to
// access TLS, and they are rewritten appropriately first here in
// liblink and then finally using relocations in the linker.
//
// When -shared is passed, we leave the code in the 2-instruction form but
// assemble (and relocate) them in different ways to generate the initial
// exec code sequence. It's a bit of a fluke that this is possible without
// rewriting the instructions more comprehensively, and it only does because
// we only support a single TLS variable (g).
if CanUse1InsnTLS(ctxt) {
// Reduce 2-instruction sequence to 1-instruction sequence.
// Sequences like
// MOVQ TLS, BX
// ... off(BX)(TLS*1) ...
// become
// NOP
// ... off(TLS) ...
//
// TODO(rsc): Remove the Hsolaris special case. It exists only to
// guarantee we are producing byte-identical binaries as before this code.
// But it should be unnecessary.
if (p.As == AMOVQ || p.As == AMOVL) && p.From.Type == obj.TYPE_REG && p.From.Reg == REG_TLS && p.To.Type == obj.TYPE_REG && REG_AX <= p.To.Reg && p.To.Reg <= REG_R15 && ctxt.Headtype != obj.Hsolaris {
obj.Nopout(p)
}
if p.From.Type == obj.TYPE_MEM && p.From.Index == REG_TLS && REG_AX <= p.From.Reg && p.From.Reg <= REG_R15 {
p.From.Reg = REG_TLS
p.From.Scale = 0
p.From.Index = REG_NONE
}
if p.To.Type == obj.TYPE_MEM && p.To.Index == REG_TLS && REG_AX <= p.To.Reg && p.To.Reg <= REG_R15 {
p.To.Reg = REG_TLS
p.To.Scale = 0
p.To.Index = REG_NONE
}
} else {
// load_g_cx, below, always inserts the 1-instruction sequence. Rewrite it
// as the 2-instruction sequence if necessary.
// MOVQ 0(TLS), BX
// becomes
// MOVQ TLS, BX
// MOVQ 0(BX)(TLS*1), BX
if (p.As == AMOVQ || p.As == AMOVL) && p.From.Type == obj.TYPE_MEM && p.From.Reg == REG_TLS && p.To.Type == obj.TYPE_REG && REG_AX <= p.To.Reg && p.To.Reg <= REG_R15 {
q := obj.Appendp(ctxt, p)
q.As = p.As
q.From = p.From
q.From.Type = obj.TYPE_MEM
q.From.Reg = p.To.Reg
q.From.Index = REG_TLS
q.From.Scale = 2 // TODO: use 1
q.To = p.To
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_TLS
p.From.Index = REG_NONE
p.From.Offset = 0
}
}
// TODO: Remove.
if ctxt.Headtype == obj.Hwindows && p.Mode == 64 || ctxt.Headtype == obj.Hplan9 {
if p.From.Scale == 1 && p.From.Index == REG_TLS {
p.From.Scale = 2
}
if p.To.Scale == 1 && p.To.Index == REG_TLS {
p.To.Scale = 2
}
}
// Rewrite 0 to $0 in 3rd argument to CMPPS etc.
// That's what the tables expect.
switch p.As {
case ACMPPD, ACMPPS, ACMPSD, ACMPSS:
if p.To.Type == obj.TYPE_MEM && p.To.Name == obj.NAME_NONE && p.To.Reg == REG_NONE && p.To.Index == REG_NONE && p.To.Sym == nil {
p.To.Type = obj.TYPE_CONST
}
}
// Rewrite CALL/JMP/RET to symbol as TYPE_BRANCH.
switch p.As {
case obj.ACALL, obj.AJMP, obj.ARET:
if p.To.Type == obj.TYPE_MEM && (p.To.Name == obj.NAME_EXTERN || p.To.Name == obj.NAME_STATIC) && p.To.Sym != nil {
p.To.Type = obj.TYPE_BRANCH
}
}
// Rewrite MOVL/MOVQ $XXX(FP/SP) as LEAL/LEAQ.
if p.From.Type == obj.TYPE_ADDR && (ctxt.Arch.Thechar == '6' || p.From.Name != obj.NAME_EXTERN && p.From.Name != obj.NAME_STATIC) {
switch p.As {
case AMOVL:
p.As = ALEAL
p.From.Type = obj.TYPE_MEM
case AMOVQ:
p.As = ALEAQ
p.From.Type = obj.TYPE_MEM
}
}
if ctxt.Headtype == obj.Hnacl && p.Mode == 64 {
if p.From3 != nil {
nacladdr(ctxt, p, p.From3)
}
nacladdr(ctxt, p, &p.From)
nacladdr(ctxt, p, &p.To)
}
// Rewrite float constants to values stored in memory.
switch p.As {
// Convert AMOVSS $(0), Xx to AXORPS Xx, Xx
case AMOVSS:
if p.From.Type == obj.TYPE_FCONST {
// f == 0 can't be used here due to -0, so use Float64bits
if f := p.From.Val.(float64); math.Float64bits(f) == 0 {
if p.To.Type == obj.TYPE_REG && REG_X0 <= p.To.Reg && p.To.Reg <= REG_X15 {
p.As = AXORPS
p.From = p.To
break
}
}
}
fallthrough
case AFMOVF,
AFADDF,
AFSUBF,
AFSUBRF,
AFMULF,
AFDIVF,
AFDIVRF,
AFCOMF,
AFCOMFP,
AADDSS,
ASUBSS,
AMULSS,
ADIVSS,
ACOMISS,
AUCOMISS:
if p.From.Type == obj.TYPE_FCONST {
f32 := float32(p.From.Val.(float64))
i32 := math.Float32bits(f32)
literal := fmt.Sprintf("$f32.%08x", i32)
s := obj.Linklookup(ctxt, literal, 0)
p.From.Type = obj.TYPE_MEM
p.From.Name = obj.NAME_EXTERN
p.From.Sym = s
p.From.Sym.Local = true
p.From.Offset = 0
}
case AMOVSD:
// Convert AMOVSD $(0), Xx to AXORPS Xx, Xx
if p.From.Type == obj.TYPE_FCONST {
// f == 0 can't be used here due to -0, so use Float64bits
if f := p.From.Val.(float64); math.Float64bits(f) == 0 {
if p.To.Type == obj.TYPE_REG && REG_X0 <= p.To.Reg && p.To.Reg <= REG_X15 {
p.As = AXORPS
p.From = p.To
break
}
}
}
fallthrough
case AFMOVD,
AFADDD,
AFSUBD,
AFSUBRD,
AFMULD,
AFDIVD,
AFDIVRD,
AFCOMD,
AFCOMDP,
AADDSD,
ASUBSD,
AMULSD,
ADIVSD,
ACOMISD,
AUCOMISD:
if p.From.Type == obj.TYPE_FCONST {
i64 := math.Float64bits(p.From.Val.(float64))
literal := fmt.Sprintf("$f64.%016x", i64)
s := obj.Linklookup(ctxt, literal, 0)
p.From.Type = obj.TYPE_MEM
p.From.Name = obj.NAME_EXTERN
p.From.Sym = s
p.From.Sym.Local = true
p.From.Offset = 0
}
}
if ctxt.Flag_dynlink {
rewriteToUseGot(ctxt, p)
}
if ctxt.Flag_shared != 0 && p.Mode == 32 {
rewriteToPcrel(ctxt, p)
}
}
func progedit(ctxt *obj.Link, p *obj.Prog) {
p.From.Class = 0
p.To.Class = 0
// Rewrite B/BL to symbol as TYPE_BRANCH.
switch p.As {
case AB,
ABL,
obj.ADUFFZERO,
obj.ADUFFCOPY:
if p.To.Type == obj.TYPE_MEM && (p.To.Name == obj.NAME_EXTERN || p.To.Name == obj.NAME_STATIC) && p.To.Sym != nil {
p.To.Type = obj.TYPE_BRANCH
}
}
// Replace TLS register fetches on older ARM procesors.
switch p.As {
// Treat MRC 15, 0, <reg>, C13, C0, 3 specially.
case AMRC:
if p.To.Offset&0xffff0fff == 0xee1d0f70 {
// Because the instruction might be rewriten to a BL which returns in R0
// the register must be zero.
if p.To.Offset&0xf000 != 0 {
ctxt.Diag("%v: TLS MRC instruction must write to R0 as it might get translated into a BL instruction", p.Line())
}
if ctxt.Goarm < 7 {
// Replace it with BL runtime.read_tls_fallback(SB) for ARM CPUs that lack the tls extension.
if progedit_tlsfallback == nil {
progedit_tlsfallback = obj.Linklookup(ctxt, "runtime.read_tls_fallback", 0)
}
// MOVW LR, R11
p.As = AMOVW
p.From.Type = obj.TYPE_REG
p.From.Reg = REGLINK
p.To.Type = obj.TYPE_REG
p.To.Reg = REGTMP
// BL runtime.read_tls_fallback(SB)
p = obj.Appendp(ctxt, p)
p.As = ABL
p.To.Type = obj.TYPE_BRANCH
p.To.Sym = progedit_tlsfallback
p.To.Offset = 0
// MOVW R11, LR
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_REG
p.From.Reg = REGTMP
p.To.Type = obj.TYPE_REG
p.To.Reg = REGLINK
break
}
}
// Otherwise, MRC/MCR instructions need no further treatment.
p.As = AWORD
}
// Rewrite float constants to values stored in memory.
switch p.As {
case AMOVF:
if p.From.Type == obj.TYPE_FCONST && chipfloat5(ctxt, p.From.Val.(float64)) < 0 && (chipzero5(ctxt, p.From.Val.(float64)) < 0 || p.Scond&C_SCOND != C_SCOND_NONE) {
f32 := float32(p.From.Val.(float64))
i32 := math.Float32bits(f32)
literal := fmt.Sprintf("$f32.%08x", i32)
s := obj.Linklookup(ctxt, literal, 0)
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
case AMOVD:
if p.From.Type == obj.TYPE_FCONST && chipfloat5(ctxt, p.From.Val.(float64)) < 0 && (chipzero5(ctxt, p.From.Val.(float64)) < 0 || p.Scond&C_SCOND != C_SCOND_NONE) {
i64 := math.Float64bits(p.From.Val.(float64))
literal := fmt.Sprintf("$f64.%016x", i64)
s := obj.Linklookup(ctxt, literal, 0)
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
}
}
func buildop(ctxt *obj.Link) {
var n int
for i := 0; i < C_NCLASS; i++ {
for n = 0; n < C_NCLASS; n++ {
if cmp(n, i) {
xcmp[i][n] = 1
}
}
}
for n = 0; optab[n].as != obj.AXXX; n++ {
}
sort.Sort(ocmp(optab[:n]))
for i := 0; i < n; i++ {
r := optab[i].as
r0 := r & obj.AMask
oprange[r0].start = optab[i:]
for optab[i].as == r {
i++
}
oprange[r0].stop = optab[i:]
i--
switch r {
default:
ctxt.Diag("unknown op in build: %v", obj.Aconv(int(r)))
log.Fatalf("bad code")
case AABSF:
opset(AMOVFD, r0)
opset(AMOVDF, r0)
opset(AMOVWF, r0)
opset(AMOVFW, r0)
opset(AMOVWD, r0)
opset(AMOVDW, r0)
opset(ANEGF, r0)
opset(ANEGD, r0)
opset(AABSD, r0)
opset(ATRUNCDW, r0)
opset(ATRUNCFW, r0)
opset(ATRUNCDV, r0)
opset(ATRUNCFV, r0)
opset(AMOVVF, r0)
opset(AMOVFV, r0)
opset(AMOVVD, r0)
opset(AMOVDV, r0)
case AADD:
opset(ASGT, r0)
opset(ASGTU, r0)
opset(AADDU, r0)
opset(AADDV, r0)
opset(AADDVU, r0)
case AADDF:
opset(ADIVF, r0)
opset(ADIVD, r0)
opset(AMULF, r0)
opset(AMULD, r0)
opset(ASUBF, r0)
opset(ASUBD, r0)
opset(AADDD, r0)
case AAND:
opset(AOR, r0)
opset(AXOR, r0)
case ABEQ:
opset(ABNE, r0)
case ABLEZ:
opset(ABGEZ, r0)
opset(ABGEZAL, r0)
opset(ABLTZ, r0)
opset(ABLTZAL, r0)
opset(ABGTZ, r0)
case AMOVB:
opset(AMOVH, r0)
case AMOVBU:
opset(AMOVHU, r0)
case AMUL:
opset(AREM, r0)
opset(AREMU, r0)
opset(ADIVU, r0)
opset(AMULU, r0)
opset(ADIV, r0)
opset(ADIVV, r0)
opset(ADIVVU, r0)
opset(AMULV, r0)
opset(AMULVU, r0)
opset(AREMV, r0)
opset(AREMVU, r0)
case ASLL:
opset(ASRL, r0)
opset(ASRA, r0)
opset(ASLLV, r0)
opset(ASRAV, r0)
opset(ASRLV, r0)
case ASUB:
opset(ASUBU, r0)
opset(ASUBV, r0)
opset(ASUBVU, r0)
opset(ANOR, r0)
case ASYSCALL:
opset(ATLBP, r0)
opset(ATLBR, r0)
opset(ATLBWI, r0)
opset(ATLBWR, r0)
case ACMPEQF:
opset(ACMPGTF, r0)
opset(ACMPGTD, r0)
opset(ACMPGEF, r0)
opset(ACMPGED, r0)
opset(ACMPEQD, r0)
case ABFPT:
opset(ABFPF, r0)
case AMOVWL:
opset(AMOVWR, r0)
opset(AMOVVR, r0)
opset(AMOVVL, r0)
case AMOVW,
AMOVD,
AMOVF,
AMOVV,
ABREAK,
ARFE,
AJAL,
AJMP,
AMOVWU,
AWORD,
obj.ANOP,
obj.ATEXT,
obj.AUNDEF,
obj.AUSEFIELD,
obj.AFUNCDATA,
obj.APCDATA,
obj.ADUFFZERO,
obj.ADUFFCOPY:
break
}
}
}
func oprrr(ctxt *obj.Link, a int) uint32 {
switch a {
case AADD:
return OP(4, 0)
case AADDU:
return OP(4, 1)
case ASGT:
return OP(5, 2)
case ASGTU:
return OP(5, 3)
case AAND:
return OP(4, 4)
case AOR:
return OP(4, 5)
case AXOR:
return OP(4, 6)
case ASUB:
return OP(4, 2)
case ASUBU:
return OP(4, 3)
case ANOR:
return OP(4, 7)
case ASLL:
return OP(0, 4)
case ASRL:
return OP(0, 6)
case ASRA:
return OP(0, 7)
case ASLLV:
return OP(2, 4)
case ASRLV:
return OP(2, 6)
case ASRAV:
return OP(2, 7)
case AADDV:
return OP(5, 4)
case AADDVU:
return OP(5, 5)
case ASUBV:
return OP(5, 6)
case ASUBVU:
return OP(5, 7)
case AREM,
ADIV:
return OP(3, 2)
case AREMU,
ADIVU:
return OP(3, 3)
case AMUL:
return OP(3, 0)
case AMULU:
return OP(3, 1)
case AREMV,
ADIVV:
return OP(3, 6)
case AREMVU,
ADIVVU:
return OP(3, 7)
case AMULV:
return OP(3, 4)
case AMULVU:
return OP(3, 5)
case AJMP:
return OP(1, 0)
case AJAL:
return OP(1, 1)
case ABREAK:
return OP(1, 5)
case ASYSCALL:
return OP(1, 4)
case ATLBP:
return MMU(1, 0)
case ATLBR:
return MMU(0, 1)
case ATLBWI:
return MMU(0, 2)
case ATLBWR:
return MMU(0, 6)
case ARFE:
return MMU(2, 0)
case ADIVF:
return FPF(0, 3)
case ADIVD:
return FPD(0, 3)
case AMULF:
return FPF(0, 2)
case AMULD:
return FPD(0, 2)
case ASUBF:
return FPF(0, 1)
case ASUBD:
return FPD(0, 1)
case AADDF:
return FPF(0, 0)
case AADDD:
return FPD(0, 0)
case ATRUNCFV:
return FPF(1, 1)
case ATRUNCDV:
return FPD(1, 1)
case ATRUNCFW:
return FPF(1, 5)
case ATRUNCDW:
return FPD(1, 5)
case AMOVFV:
return FPF(4, 5)
case AMOVDV:
return FPD(4, 5)
case AMOVVF:
return FPV(4, 0)
case AMOVVD:
return FPV(4, 1)
case AMOVFW:
return FPF(4, 4)
case AMOVDW:
return FPD(4, 4)
case AMOVWF:
return FPW(4, 0)
case AMOVDF:
return FPD(4, 0)
case AMOVWD:
return FPW(4, 1)
case AMOVFD:
return FPF(4, 1)
case AABSF:
return FPF(0, 5)
case AABSD:
return FPD(0, 5)
case AMOVF:
return FPF(0, 6)
case AMOVD:
return FPD(0, 6)
case ANEGF:
return FPF(0, 7)
case ANEGD:
return FPD(0, 7)
case ACMPEQF:
return FPF(6, 2)
case ACMPEQD:
return FPD(6, 2)
case ACMPGTF:
return FPF(7, 4)
case ACMPGTD:
return FPD(7, 4)
case ACMPGEF:
return FPF(7, 6)
case ACMPGED:
return FPD(7, 6)
}
if a >= ALAST {
ctxt.Diag("bad rrr opcode %v+ALAST", obj.Aconv(a-ALAST))
} else {
ctxt.Diag("bad rrr opcode %v", obj.Aconv(a))
}
return 0
}
func xfol(ctxt *obj.Link, p *obj.Prog, last **obj.Prog) {
var q *obj.Prog
var i int
var a obj.As
loop:
if p == nil {
return
}
if p.As == obj.AJMP {
q = p.Pcond
if q != nil && q.As != obj.ATEXT {
/* mark instruction as done and continue layout at target of jump */
p.Mark |= DONE
p = q
if p.Mark&DONE == 0 {
goto loop
}
}
}
if p.Mark&DONE != 0 {
/*
* p goes here, but already used it elsewhere.
* copy up to 4 instructions or else branch to other copy.
*/
i = 0
q = p
for ; i < 4; i, q = i+1, q.Link {
if q == nil {
break
}
if q == *last {
break
}
a = q.As
if a == obj.ANOP {
i--
continue
}
if nofollow(a) || pushpop(a) {
break // NOTE(rsc): arm does goto copy
}
if q.Pcond == nil || q.Pcond.Mark&DONE != 0 {
continue
}
if a == obj.ACALL || a == ALOOP {
continue
}
for {
if p.As == obj.ANOP {
p = p.Link
continue
}
q = obj.Copyp(ctxt, p)
p = p.Link
q.Mark |= DONE
(*last).Link = q
*last = q
if q.As != a || q.Pcond == nil || q.Pcond.Mark&DONE != 0 {
continue
}
q.As = relinv(q.As)
p = q.Pcond
q.Pcond = q.Link
q.Link = p
xfol(ctxt, q.Link, last)
p = q.Link
if p.Mark&DONE != 0 {
return
}
goto loop
/* */
}
}
q = ctxt.NewProg()
q.As = obj.AJMP
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_BRANCH
q.To.Offset = p.Pc
q.Pcond = p
p = q
}
/* emit p */
p.Mark |= DONE
(*last).Link = p
*last = p
a = p.As
/* continue loop with what comes after p */
if nofollow(a) {
return
}
if p.Pcond != nil && a != obj.ACALL {
/*
* some kind of conditional branch.
* recurse to follow one path.
* continue loop on the other.
*/
q = obj.Brchain(ctxt, p.Pcond)
if q != nil {
p.Pcond = q
}
q = obj.Brchain(ctxt, p.Link)
if q != nil {
p.Link = q
}
if p.From.Type == obj.TYPE_CONST {
if p.From.Offset == 1 {
/*
* expect conditional jump to be taken.
* rewrite so that's the fall-through case.
*/
p.As = relinv(a)
q = p.Link
p.Link = p.Pcond
p.Pcond = q
}
} else {
q = p.Link
if q.Mark&DONE != 0 {
if a != ALOOP {
p.As = relinv(a)
p.Link = p.Pcond
p.Pcond = q
}
}
}
xfol(ctxt, p.Link, last)
if p.Pcond.Mark&DONE != 0 {
return
}
p = p.Pcond
goto loop
}
p = p.Link
goto loop
}
func span0(ctxt *obj.Link, cursym *obj.LSym) {
p := cursym.Text
if p == nil || p.Link == nil { // handle external functions and ELF section symbols
return
}
ctxt.Cursym = cursym
ctxt.Autosize = int32(p.To.Offset + 8)
if oprange[AOR&obj.AMask].start == nil {
buildop(ctxt)
}
c := int64(0)
p.Pc = c
var m int
var o *Optab
for p = p.Link; p != nil; p = p.Link {
ctxt.Curp = p
p.Pc = c
o = oplook(ctxt, p)
m = int(o.size)
if m == 0 {
if p.As != obj.ANOP && p.As != obj.AFUNCDATA && p.As != obj.APCDATA && p.As != obj.AUSEFIELD {
ctxt.Diag("zero-width instruction\n%v", p)
}
continue
}
c += int64(m)
}
cursym.Size = c
/*
* if any procedure is large enough to
* generate a large SBRA branch, then
* generate extra passes putting branches
* around jmps to fix. this is rare.
*/
bflag := 1
var otxt int64
var q *obj.Prog
for bflag != 0 {
if ctxt.Debugvlog != 0 {
fmt.Fprintf(ctxt.Bso, "%5.2f span1\n", obj.Cputime())
}
bflag = 0
c = 0
for p = cursym.Text.Link; p != nil; p = p.Link {
p.Pc = c
o = oplook(ctxt, p)
// very large conditional branches
if o.type_ == 6 && p.Pcond != nil {
otxt = p.Pcond.Pc - c
if otxt < -(1<<17)+10 || otxt >= (1<<17)-10 {
q = ctxt.NewProg()
q.Link = p.Link
p.Link = q
q.As = AJMP
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_BRANCH
q.Pcond = p.Pcond
p.Pcond = q
q = ctxt.NewProg()
q.Link = p.Link
p.Link = q
q.As = AJMP
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_BRANCH
q.Pcond = q.Link.Link
addnop(ctxt, p.Link)
addnop(ctxt, p)
bflag = 1
}
}
m = int(o.size)
if m == 0 {
if p.As != obj.ANOP && p.As != obj.AFUNCDATA && p.As != obj.APCDATA && p.As != obj.AUSEFIELD {
ctxt.Diag("zero-width instruction\n%v", p)
}
continue
}
c += int64(m)
}
cursym.Size = c
}
c += -c & (FuncAlign - 1)
cursym.Size = c
/*
* lay out the code, emitting code and data relocations.
*/
obj.Symgrow(ctxt, cursym, cursym.Size)
bp := cursym.P
var i int32
var out [4]uint32
for p := cursym.Text.Link; p != nil; p = p.Link {
ctxt.Pc = p.Pc
ctxt.Curp = p
o = oplook(ctxt, p)
if int(o.size) > 4*len(out) {
log.Fatalf("out array in span0 is too small, need at least %d for %v", o.size/4, p)
}
asmout(ctxt, p, o, out[:])
for i = 0; i < int32(o.size/4); i++ {
ctxt.Arch.ByteOrder.PutUint32(bp, out[i])
bp = bp[4:]
}
}
}
func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
// TODO(minux): add morestack short-cuts with small fixed frame-size.
ctxt.Cursym = cursym
// a switch for enabling/disabling instruction scheduling
nosched := true
if cursym.Text == nil || cursym.Text.Link == nil {
return
}
p := cursym.Text
textstksiz := p.To.Offset
cursym.Args = p.To.Val.(int32)
cursym.Locals = int32(textstksiz)
/*
* find leaf subroutines
* strip NOPs
* expand RET
* expand BECOME pseudo
*/
if ctxt.Debugvlog != 0 {
ctxt.Logf("%5.2f noops\n", obj.Cputime())
}
var q *obj.Prog
var q1 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
switch p.As {
/* too hard, just leave alone */
case obj.ATEXT:
q = p
p.Mark |= LABEL | LEAF | SYNC
if p.Link != nil {
p.Link.Mark |= LABEL
}
/* too hard, just leave alone */
case AMOVW,
AMOVV:
q = p
if p.To.Type == obj.TYPE_REG && p.To.Reg >= REG_SPECIAL {
p.Mark |= LABEL | SYNC
break
}
if p.From.Type == obj.TYPE_REG && p.From.Reg >= REG_SPECIAL {
p.Mark |= LABEL | SYNC
}
/* too hard, just leave alone */
case ASYSCALL,
AWORD,
ATLBWR,
ATLBWI,
ATLBP,
ATLBR:
q = p
p.Mark |= LABEL | SYNC
case ANOR:
q = p
if p.To.Type == obj.TYPE_REG {
if p.To.Reg == REGZERO {
p.Mark |= LABEL | SYNC
}
}
case ABGEZAL,
ABLTZAL,
AJAL,
obj.ADUFFZERO,
obj.ADUFFCOPY:
cursym.Text.Mark &^= LEAF
fallthrough
case AJMP,
ABEQ,
ABGEZ,
ABGTZ,
ABLEZ,
ABLTZ,
ABNE,
ABFPT, ABFPF:
if p.As == ABFPT || p.As == ABFPF {
// We don't treat ABFPT and ABFPF as branches here,
// so that we will always fill nop (0x0) in their
// delay slot during assembly.
// This is to workaround a kernel FPU emulator bug
// where it uses the user stack to simulate the
// instruction in the delay slot if it's not 0x0,
// and somehow that leads to SIGSEGV when the kernel
// jump to the stack.
p.Mark |= SYNC
} else {
p.Mark |= BRANCH
}
q = p
q1 = p.Pcond
if q1 != nil {
for q1.As == obj.ANOP {
q1 = q1.Link
p.Pcond = q1
}
if q1.Mark&LEAF == 0 {
q1.Mark |= LABEL
}
}
//else {
// p.Mark |= LABEL
//}
q1 = p.Link
if q1 != nil {
q1.Mark |= LABEL
}
continue
case ARET:
q = p
if p.Link != nil {
p.Link.Mark |= LABEL
}
continue
case obj.ANOP:
q1 = p.Link
q.Link = q1 /* q is non-nop */
q1.Mark |= p.Mark
continue
default:
q = p
continue
}
}
var mov, add obj.As
if ctxt.Mode&Mips64 != 0 {
add = AADDV
mov = AMOVV
} else {
add = AADDU
mov = AMOVW
}
autosize := int32(0)
var p1 *obj.Prog
var p2 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
o := p.As
switch o {
case obj.ATEXT:
autosize = int32(textstksiz + ctxt.FixedFrameSize())
if (p.Mark&LEAF != 0) && autosize <= int32(ctxt.FixedFrameSize()) {
autosize = 0
} else if autosize&4 != 0 && ctxt.Mode&Mips64 != 0 {
autosize += 4
}
p.To.Offset = int64(autosize) - ctxt.FixedFrameSize()
if p.From3.Offset&obj.NOSPLIT == 0 {
p = stacksplit(ctxt, p, autosize) // emit split check
}
q = p
if autosize != 0 {
// Make sure to save link register for non-empty frame, even if
// it is a leaf function, so that traceback works.
// Store link register before decrement SP, so if a signal comes
// during the execution of the function prologue, the traceback
// code will not see a half-updated stack frame.
q = obj.Appendp(ctxt, q)
q.As = mov
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_REG
q.From.Reg = REGLINK
q.To.Type = obj.TYPE_MEM
q.To.Offset = int64(-autosize)
q.To.Reg = REGSP
q = obj.Appendp(ctxt, q)
q.As = add
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(-autosize)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Spadj = +autosize
} else if cursym.Text.Mark&LEAF == 0 {
if cursym.Text.From3.Offset&obj.NOSPLIT != 0 {
if ctxt.Debugvlog != 0 {
ctxt.Logf("save suppressed in: %s\n", cursym.Name)
}
cursym.Text.Mark |= LEAF
}
}
if cursym.Text.Mark&LEAF != 0 {
cursym.Set(obj.AttrLeaf, true)
break
}
if cursym.Text.From3.Offset&obj.WRAPPER != 0 {
// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
//
// MOV g_panic(g), R1
// BEQ R1, end
// MOV panic_argp(R1), R2
// ADD $(autosize+FIXED_FRAME), R29, R3
// BNE R2, R3, end
// ADD $FIXED_FRAME, R29, R2
// MOV R2, panic_argp(R1)
// end:
// NOP
//
// The NOP is needed to give the jumps somewhere to land.
// It is a liblink NOP, not an mips NOP: it encodes to 0 instruction bytes.
q = obj.Appendp(ctxt, q)
q.As = mov
q.From.Type = obj.TYPE_MEM
q.From.Reg = REGG
q.From.Offset = 4 * int64(ctxt.Arch.PtrSize) // G.panic
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R1
q = obj.Appendp(ctxt, q)
q.As = ABEQ
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R1
q.To.Type = obj.TYPE_BRANCH
q.Mark |= BRANCH
p1 = q
q = obj.Appendp(ctxt, q)
q.As = mov
q.From.Type = obj.TYPE_MEM
q.From.Reg = REG_R1
q.From.Offset = 0 // Panic.argp
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R2
q = obj.Appendp(ctxt, q)
q.As = add
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(autosize) + ctxt.FixedFrameSize()
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R3
q = obj.Appendp(ctxt, q)
q.As = ABNE
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R2
q.Reg = REG_R3
q.To.Type = obj.TYPE_BRANCH
q.Mark |= BRANCH
p2 = q
q = obj.Appendp(ctxt, q)
q.As = add
q.From.Type = obj.TYPE_CONST
q.From.Offset = ctxt.FixedFrameSize()
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R2
q = obj.Appendp(ctxt, q)
q.As = mov
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R2
q.To.Type = obj.TYPE_MEM
q.To.Reg = REG_R1
q.To.Offset = 0 // Panic.argp
q = obj.Appendp(ctxt, q)
q.As = obj.ANOP
p1.Pcond = q
p2.Pcond = q
}
case ARET:
if p.From.Type == obj.TYPE_CONST {
ctxt.Diag("using BECOME (%v) is not supported!", p)
break
}
retSym := p.To.Sym
p.To.Name = obj.NAME_NONE // clear fields as we may modify p to other instruction
p.To.Sym = nil
if cursym.Text.Mark&LEAF != 0 {
if autosize == 0 {
p.As = AJMP
p.From = obj.Addr{}
if retSym != nil { // retjmp
p.To.Type = obj.TYPE_BRANCH
p.To.Name = obj.NAME_EXTERN
p.To.Sym = retSym
} else {
p.To.Type = obj.TYPE_MEM
p.To.Reg = REGLINK
p.To.Offset = 0
}
p.Mark |= BRANCH
break
}
p.As = add
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(autosize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REGSP
p.Spadj = -autosize
q = ctxt.NewProg()
q.As = AJMP
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_MEM
q.To.Offset = 0
q.To.Reg = REGLINK
q.Mark |= BRANCH
q.Spadj = +autosize
q.Link = p.Link
p.Link = q
break
}
p.As = mov
p.From.Type = obj.TYPE_MEM
p.From.Offset = 0
p.From.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
if retSym != nil { // retjmp from non-leaf, need to restore LINK register
p.To.Reg = REGLINK
}
if autosize != 0 {
q = ctxt.NewProg()
q.As = add
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(autosize)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Spadj = -autosize
q.Link = p.Link
p.Link = q
}
q1 = ctxt.NewProg()
q1.As = AJMP
q1.Lineno = p.Lineno
if retSym != nil { // retjmp
q1.To.Type = obj.TYPE_BRANCH
q1.To.Name = obj.NAME_EXTERN
q1.To.Sym = retSym
} else {
q1.To.Type = obj.TYPE_MEM
q1.To.Offset = 0
q1.To.Reg = REG_R4
}
q1.Mark |= BRANCH
q1.Spadj = +autosize
q1.Link = q.Link
q.Link = q1
case AADD,
AADDU,
AADDV,
AADDVU:
if p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP && p.From.Type == obj.TYPE_CONST {
p.Spadj = int32(-p.From.Offset)
}
}
}
if nosched {
// if we don't do instruction scheduling, simply add
// NOP after each branch instruction.
for p = cursym.Text; p != nil; p = p.Link {
if p.Mark&BRANCH != 0 {
addnop(ctxt, p)
}
}
return
}
// instruction scheduling
q = nil // p - 1
q1 = cursym.Text // top of block
o := 0 // count of instructions
for p = cursym.Text; p != nil; p = p1 {
p1 = p.Link
o++
if p.Mark&NOSCHED != 0 {
if q1 != p {
sched(ctxt, q1, q)
}
for ; p != nil; p = p.Link {
if p.Mark&NOSCHED == 0 {
break
}
q = p
}
p1 = p
q1 = p
o = 0
continue
}
if p.Mark&(LABEL|SYNC) != 0 {
if q1 != p {
sched(ctxt, q1, q)
}
q1 = p
o = 1
}
if p.Mark&(BRANCH|SYNC) != 0 {
sched(ctxt, q1, p)
q1 = p1
o = 0
}
if o >= NSCHED {
sched(ctxt, q1, p)
q1 = p1
o = 0
}
q = p
}
}
func softfloat(ctxt *obj.Link, cursym *obj.LSym) {
if ctxt.Goarm > 5 {
return
}
symsfloat := obj.Linklookup(ctxt, "_sfloat", 0)
wasfloat := 0
for p := cursym.Text; p != nil; p = p.Link {
if p.Pcond != nil {
p.Pcond.Mark |= LABEL
}
}
var next *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
switch p.As {
case AMOVW:
if isfloatreg(&p.To) || isfloatreg(&p.From) {
goto soft
}
goto notsoft
case AMOVWD,
AMOVWF,
AMOVDW,
AMOVFW,
AMOVFD,
AMOVDF,
AMOVF,
AMOVD,
ACMPF,
ACMPD,
AADDF,
AADDD,
ASUBF,
ASUBD,
AMULF,
AMULD,
ADIVF,
ADIVD,
ASQRTF,
ASQRTD,
AABSF,
AABSD:
goto soft
default:
goto notsoft
}
soft:
if wasfloat == 0 || (p.Mark&LABEL != 0) {
next = ctxt.NewProg()
*next = *p
// BL _sfloat(SB)
*p = obj.Prog{}
p.Ctxt = ctxt
p.Link = next
p.As = ABL
p.To.Type = obj.TYPE_BRANCH
p.To.Sym = symsfloat
p.Lineno = next.Lineno
p = next
wasfloat = 1
}
continue
notsoft:
wasfloat = 0
}
}
func progedit(ctxt *obj.Link, p *obj.Prog) {
// Maintain information about code generation mode.
if ctxt.Mode == 0 {
switch ctxt.Arch.Family {
default:
ctxt.Diag("unsupported arch family")
case sys.MIPS:
ctxt.Mode = Mips32
case sys.MIPS64:
ctxt.Mode = Mips64
}
}
p.From.Class = 0
p.To.Class = 0
// Rewrite JMP/JAL to symbol as TYPE_BRANCH.
switch p.As {
case AJMP,
AJAL,
ARET,
obj.ADUFFZERO,
obj.ADUFFCOPY:
if p.To.Sym != nil {
p.To.Type = obj.TYPE_BRANCH
}
}
// Rewrite float constants to values stored in memory.
switch p.As {
case AMOVF:
if p.From.Type == obj.TYPE_FCONST {
f32 := float32(p.From.Val.(float64))
i32 := math.Float32bits(f32)
if i32 == 0 {
p.As = AMOVW
p.From.Type = obj.TYPE_REG
p.From.Reg = REGZERO
break
}
literal := fmt.Sprintf("$f32.%08x", i32)
s := obj.Linklookup(ctxt, literal, 0)
s.Size = 4
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
case AMOVD:
if p.From.Type == obj.TYPE_FCONST {
i64 := math.Float64bits(p.From.Val.(float64))
if i64 == 0 && ctxt.Mode&Mips64 != 0 {
p.As = AMOVV
p.From.Type = obj.TYPE_REG
p.From.Reg = REGZERO
break
}
literal := fmt.Sprintf("$f64.%016x", i64)
s := obj.Linklookup(ctxt, literal, 0)
s.Size = 8
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
// Put >32-bit constants in memory and load them
case AMOVV:
if p.From.Type == obj.TYPE_CONST && p.From.Name == obj.NAME_NONE && p.From.Reg == 0 && int64(int32(p.From.Offset)) != p.From.Offset {
literal := fmt.Sprintf("$i64.%016x", uint64(p.From.Offset))
s := obj.Linklookup(ctxt, literal, 0)
s.Size = 8
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
}
// Rewrite SUB constants into ADD.
switch p.As {
case ASUB:
if p.From.Type == obj.TYPE_CONST {
p.From.Offset = -p.From.Offset
p.As = AADD
}
case ASUBU:
if p.From.Type == obj.TYPE_CONST {
p.From.Offset = -p.From.Offset
p.As = AADDU
}
case ASUBV:
if p.From.Type == obj.TYPE_CONST {
p.From.Offset = -p.From.Offset
p.As = AADDV
}
case ASUBVU:
if p.From.Type == obj.TYPE_CONST {
p.From.Offset = -p.From.Offset
p.As = AADDVU
}
}
}
func asmout(ctxt *obj.Link, p *obj.Prog, o *Optab, out []uint32) {
o1 := uint32(0)
o2 := uint32(0)
o3 := uint32(0)
o4 := uint32(0)
switch o.type_ {
default:
ctxt.Diag("unknown type %d %v", o.type_)
prasm(p)
case 0: /* pseudo ops */
break
case 1: /* mov r1,r2 ==> OR r1,r0,r2 */
o1 = OP_RRR(oprrr(ctxt, AOR), uint32(p.From.Reg), uint32(REGZERO), uint32(p.To.Reg))
case 2: /* add/sub r1,[r2],r3 */
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o1 = OP_RRR(oprrr(ctxt, int(p.As)), uint32(p.From.Reg), uint32(r), uint32(p.To.Reg))
case 3: /* mov $soreg, r ==> or/add $i,o,r */
v := regoff(ctxt, &p.From)
r := int(p.From.Reg)
if r == 0 {
r = int(o.param)
}
a := AADDVU
if o.a1 == C_ANDCON {
a = AOR
}
o1 = OP_IRR(opirr(ctxt, a), uint32(v), uint32(r), uint32(p.To.Reg))
case 4: /* add $scon,[r1],r2 */
v := regoff(ctxt, &p.From)
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o1 = OP_IRR(opirr(ctxt, int(p.As)), uint32(v), uint32(r), uint32(p.To.Reg))
case 5: /* syscall */
o1 = uint32(oprrr(ctxt, int(p.As)))
case 6: /* beq r1,[r2],sbra */
v := int32(0)
if p.Pcond == nil {
v = int32(-4) >> 2
} else {
v = int32(p.Pcond.Pc-p.Pc-4) >> 2
}
if (v<<16)>>16 != v {
ctxt.Diag("short branch too far\n%v", p)
}
o1 = OP_IRR(opirr(ctxt, int(p.As)), uint32(v), uint32(p.From.Reg), uint32(p.Reg))
// for ABFPT and ABFPF only: always fill delay slot with 0
// see comments in func preprocess for details.
o2 = 0
case 7: /* mov r, soreg ==> sw o(r) */
r := int(p.To.Reg)
if r == 0 {
r = int(o.param)
}
v := regoff(ctxt, &p.To)
o1 = OP_IRR(opirr(ctxt, int(p.As)), uint32(v), uint32(r), uint32(p.From.Reg))
case 8: /* mov soreg, r ==> lw o(r) */
r := int(p.From.Reg)
if r == 0 {
r = int(o.param)
}
v := regoff(ctxt, &p.From)
o1 = OP_IRR(opirr(ctxt, int(p.As)+ALAST), uint32(v), uint32(r), uint32(p.To.Reg))
case 9: /* sll r1,[r2],r3 */
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o1 = OP_RRR(oprrr(ctxt, int(p.As)), uint32(r), uint32(p.From.Reg), uint32(p.To.Reg))
case 10: /* add $con,[r1],r2 ==> mov $con, t; add t,[r1],r2 */
v := regoff(ctxt, &p.From)
a := AOR
if v < 0 {
a = AADDU
}
o1 = OP_IRR(opirr(ctxt, a), uint32(v), uint32(0), uint32(REGTMP))
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o2 = OP_RRR(oprrr(ctxt, int(p.As)), uint32(REGTMP), uint32(r), uint32(p.To.Reg))
case 11: /* jmp lbra */
v := int32(0)
if aclass(ctxt, &p.To) == C_SBRA && p.To.Sym == nil && p.As == AJMP {
// use PC-relative branch for short branches
// BEQ R0, R0, sbra
if p.Pcond == nil {
v = int32(-4) >> 2
} else {
v = int32(p.Pcond.Pc-p.Pc-4) >> 2
}
if (v<<16)>>16 == v {
o1 = OP_IRR(opirr(ctxt, ABEQ), uint32(v), uint32(REGZERO), uint32(REGZERO))
break
}
}
if p.Pcond == nil {
v = int32(p.Pc) >> 2
} else {
v = int32(p.Pcond.Pc) >> 2
}
o1 = OP_JMP(opirr(ctxt, int(p.As)), uint32(v))
if p.To.Sym == nil {
p.To.Sym = ctxt.Cursym.Text.From.Sym
p.To.Offset = p.Pcond.Pc
}
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc)
rel.Siz = 4
rel.Sym = p.To.Sym
rel.Add = p.To.Offset
if p.As == AJAL {
rel.Type = obj.R_CALLMIPS
} else {
rel.Type = obj.R_JMPMIPS
}
case 12: /* movbs r,r */
v := 16
if p.As == AMOVB {
v = 24
}
o1 = OP_SRR(opirr(ctxt, ASLL), uint32(v), uint32(p.From.Reg), uint32(p.To.Reg))
o2 = OP_SRR(opirr(ctxt, ASRA), uint32(v), uint32(p.To.Reg), uint32(p.To.Reg))
case 13: /* movbu r,r */
if p.As == AMOVBU {
o1 = OP_IRR(opirr(ctxt, AAND), uint32(0xff), uint32(p.From.Reg), uint32(p.To.Reg))
} else {
o1 = OP_IRR(opirr(ctxt, AAND), uint32(0xffff), uint32(p.From.Reg), uint32(p.To.Reg))
}
case 14: /* movwu r,r */
o1 = OP_SRR(opirr(ctxt, ASLLV+ALAST), uint32(0), uint32(p.From.Reg), uint32(p.To.Reg))
if p.As == AMOVWU {
o2 = OP_SRR(opirr(ctxt, ASRLV+ALAST), uint32(0), uint32(p.To.Reg), uint32(p.To.Reg))
} else {
o2 = OP_SRR(opirr(ctxt, ASRAV+ALAST), uint32(0), uint32(p.To.Reg), uint32(p.To.Reg))
}
case 16: /* sll $c,[r1],r2 */
v := regoff(ctxt, &p.From)
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
/* OP_SRR will use only the low 5 bits of the shift value */
if v >= 32 && vshift(p.As) {
o1 = OP_SRR(opirr(ctxt, int(p.As)+ALAST), uint32(v-32), uint32(r), uint32(p.To.Reg))
} else {
o1 = OP_SRR(opirr(ctxt, int(p.As)), uint32(v), uint32(r), uint32(p.To.Reg))
}
case 18: /* jmp [r1],0(r2) */
r := int(p.Reg)
if r == 0 {
r = int(o.param)
}
o1 = OP_RRR(oprrr(ctxt, int(p.As)), uint32(0), uint32(p.To.Reg), uint32(r))
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc)
rel.Siz = 0
rel.Type = obj.R_CALLIND
case 19: /* mov $lcon,r ==> lu+or */
v := regoff(ctxt, &p.From)
o1 = OP_IRR(opirr(ctxt, ALAST), uint32(v>>16), uint32(REGZERO), uint32(p.To.Reg))
o2 = OP_IRR(opirr(ctxt, AOR), uint32(v), uint32(p.To.Reg), uint32(p.To.Reg))
if p.From.Sym != nil {
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc)
rel.Siz = 8
rel.Sym = p.From.Sym
rel.Add = p.From.Offset
rel.Type = obj.R_ADDRMIPS
}
case 20: /* mov lo/hi,r */
a := OP(2, 0) /* mfhi */
if p.From.Reg == REG_LO {
a = OP(2, 2) /* mflo */
}
o1 = OP_RRR(a, uint32(REGZERO), uint32(REGZERO), uint32(p.To.Reg))
case 21: /* mov r,lo/hi */
a := OP(2, 1) /* mthi */
if p.To.Reg == REG_LO {
a = OP(2, 3) /* mtlo */
}
o1 = OP_RRR(a, uint32(REGZERO), uint32(p.From.Reg), uint32(REGZERO))
case 22: /* mul r1,r2 */
o1 = OP_RRR(oprrr(ctxt, int(p.As)), uint32(p.From.Reg), uint32(p.Reg), uint32(REGZERO))
case 23: /* add $lcon,r1,r2 ==> lu+or+add */
v := regoff(ctxt, &p.From)
o1 = OP_IRR(opirr(ctxt, ALAST), uint32(v>>16), uint32(REGZERO), uint32(REGTMP))
o2 = OP_IRR(opirr(ctxt, AOR), uint32(v), uint32(REGTMP), uint32(REGTMP))
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o3 = OP_RRR(oprrr(ctxt, int(p.As)), uint32(REGTMP), uint32(r), uint32(p.To.Reg))
case 24: /* mov $ucon,r ==> lu r */
v := regoff(ctxt, &p.From)
o1 = OP_IRR(opirr(ctxt, ALAST), uint32(v>>16), uint32(REGZERO), uint32(p.To.Reg))
case 25: /* add/and $ucon,[r1],r2 ==> lu $con,t; add t,[r1],r2 */
v := regoff(ctxt, &p.From)
o1 = OP_IRR(opirr(ctxt, ALAST), uint32(v>>16), uint32(REGZERO), uint32(REGTMP))
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o2 = OP_RRR(oprrr(ctxt, int(p.As)), uint32(REGTMP), uint32(r), uint32(p.To.Reg))
case 26: /* mov $lsext/auto/oreg,r ==> lu+or+add */
v := regoff(ctxt, &p.From)
o1 = OP_IRR(opirr(ctxt, ALAST), uint32(v>>16), uint32(REGZERO), uint32(REGTMP))
o2 = OP_IRR(opirr(ctxt, AOR), uint32(v), uint32(REGTMP), uint32(REGTMP))
r := int(p.From.Reg)
if r == 0 {
r = int(o.param)
}
o3 = OP_RRR(oprrr(ctxt, AADDVU), uint32(REGTMP), uint32(r), uint32(p.To.Reg))
case 27: /* mov [sl]ext/auto/oreg,fr ==> lwc1 o(r) */
v := regoff(ctxt, &p.From)
r := int(p.From.Reg)
if r == 0 {
r = int(o.param)
}
a := AMOVF + ALAST
if p.As == AMOVD {
a = AMOVD + ALAST
}
switch o.size {
case 16:
o1 = OP_IRR(opirr(ctxt, ALAST), uint32(v>>16), uint32(REGZERO), uint32(REGTMP))
o2 = OP_IRR(opirr(ctxt, AOR), uint32(v), uint32(REGTMP), uint32(REGTMP))
o3 = OP_RRR(opirr(ctxt, AADDVU), uint32(r), uint32(REGTMP), uint32(REGTMP))
o4 = OP_IRR(opirr(ctxt, a), uint32(0), uint32(r), uint32(p.To.Reg))
case 4:
o1 = OP_IRR(opirr(ctxt, a), uint32(v), uint32(r), uint32(p.To.Reg))
}
case 28: /* mov fr,[sl]ext/auto/oreg ==> swc1 o(r) */
v := regoff(ctxt, &p.To)
r := int(p.To.Reg)
if r == 0 {
r = int(o.param)
}
a := AMOVF
if p.As == AMOVD {
a = AMOVD
}
switch o.size {
case 16:
o1 = OP_IRR(opirr(ctxt, ALAST), uint32(v>>16), uint32(REGZERO), uint32(REGTMP))
o2 = OP_IRR(opirr(ctxt, AOR), uint32(v), uint32(REGTMP), uint32(REGTMP))
o3 = OP_RRR(opirr(ctxt, AADDVU), uint32(r), uint32(REGTMP), uint32(REGTMP))
o4 = OP_IRR(opirr(ctxt, a), uint32(0), uint32(REGTMP), uint32(p.From.Reg))
case 4:
o1 = OP_IRR(opirr(ctxt, a), uint32(v), uint32(r), uint32(p.From.Reg))
}
case 30: /* movw r,fr */
a := SP(2, 1) | (4 << 21) /* mtc1 */
o1 = OP_RRR(a, uint32(p.From.Reg), uint32(0), uint32(p.To.Reg))
case 31: /* movw fr,r */
a := SP(2, 1) | (0 << 21) /* mtc1 */
o1 = OP_RRR(a, uint32(p.To.Reg), uint32(0), uint32(p.From.Reg))
case 32: /* fadd fr1,[fr2],fr3 */
r := int(p.Reg)
if r == 0 {
r = int(p.To.Reg)
}
o1 = OP_FRRR(oprrr(ctxt, int(p.As)), uint32(p.From.Reg), uint32(r), uint32(p.To.Reg))
case 33: /* fabs fr1, fr3 */
o1 = OP_FRRR(oprrr(ctxt, int(p.As)), uint32(0), uint32(p.From.Reg), uint32(p.To.Reg))
case 34: /* mov $con,fr ==> or/add $i,t; mov t,fr */
v := regoff(ctxt, &p.From)
a := AADDU
if o.a1 == C_ANDCON {
a = AOR
}
o1 = OP_IRR(opirr(ctxt, a), uint32(v), uint32(0), uint32(REGTMP))
o2 = OP_RRR(SP(2, 1)|(4<<21), uint32(REGTMP), uint32(0), uint32(p.To.Reg)) /* mtc1 */
case 35: /* mov r,lext/auto/oreg ==> sw o(r) */
v := regoff(ctxt, &p.To)
r := int(p.To.Reg)
if r == 0 {
r = int(o.param)
}
o1 = OP_IRR(opirr(ctxt, ALAST), uint32(v>>16), uint32(REGZERO), uint32(REGTMP))
o2 = OP_IRR(opirr(ctxt, AOR), uint32(v), uint32(REGTMP), uint32(REGTMP))
o3 = OP_RRR(oprrr(ctxt, AADDVU), uint32(r), uint32(REGTMP), uint32(REGTMP))
o4 = OP_IRR(opirr(ctxt, int(p.As)), uint32(0), uint32(REGTMP), uint32(p.From.Reg))
case 36: /* mov lext/auto/oreg,r ==> lw o(r30) */
v := regoff(ctxt, &p.From)
r := int(p.From.Reg)
if r == 0 {
r = int(o.param)
}
o1 = OP_IRR(opirr(ctxt, ALAST), uint32(v>>16), uint32(REGZERO), uint32(REGTMP))
o2 = OP_IRR(opirr(ctxt, AOR), uint32(v), uint32(REGTMP), uint32(REGTMP))
o3 = OP_RRR(oprrr(ctxt, AADDVU), uint32(r), uint32(REGTMP), uint32(REGTMP))
o4 = OP_IRR(opirr(ctxt, int(p.As)+ALAST), uint32(0), uint32(REGTMP), uint32(p.To.Reg))
case 37: /* movw r,mr */
a := SP(2, 0) | (4 << 21) /* mtc0 */
if p.As == AMOVV {
a = SP(2, 0) | (5 << 21) /* dmtc0 */
}
o1 = OP_RRR(a, uint32(p.From.Reg), uint32(0), uint32(p.To.Reg))
case 38: /* movw mr,r */
a := SP(2, 0) | (0 << 21) /* mfc0 */
if p.As == AMOVV {
a = SP(2, 0) | (1 << 21) /* dmfc0 */
}
o1 = OP_RRR(a, uint32(p.To.Reg), uint32(0), uint32(p.From.Reg))
case 40: /* word */
o1 = uint32(regoff(ctxt, &p.From))
case 41: /* movw f,fcr */
o1 = OP_RRR(SP(2, 1)|(2<<21), uint32(REGZERO), uint32(0), uint32(p.To.Reg)) /* mfcc1 */
o2 = OP_RRR(SP(2, 1)|(6<<21), uint32(p.From.Reg), uint32(0), uint32(p.To.Reg)) /* mtcc1 */
case 42: /* movw fcr,r */
o1 = OP_RRR(SP(2, 1)|(2<<21), uint32(p.To.Reg), uint32(0), uint32(p.From.Reg)) /* mfcc1 */
case 47: /* movv r,fr */
a := SP(2, 1) | (5 << 21) /* dmtc1 */
o1 = OP_RRR(a, uint32(p.From.Reg), uint32(0), uint32(p.To.Reg))
case 48: /* movv fr,r */
a := SP(2, 1) | (1 << 21) /* dmtc1 */
o1 = OP_RRR(a, uint32(p.To.Reg), uint32(0), uint32(p.From.Reg))
case 49: /* undef */
o1 = 8 /* JMP (R0) */
/* relocation operations */
case 50: /* mov r,addr ==> lu + or + sw (REGTMP) */
o1 = OP_IRR(opirr(ctxt, ALAST), uint32(0), uint32(REGZERO), uint32(REGTMP))
o2 = OP_IRR(opirr(ctxt, AOR), uint32(0), uint32(REGTMP), uint32(REGTMP))
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc)
rel.Siz = 8
rel.Sym = p.To.Sym
rel.Add = p.To.Offset
rel.Type = obj.R_ADDRMIPS
o3 = OP_IRR(opirr(ctxt, int(p.As)), uint32(0), uint32(REGTMP), uint32(p.From.Reg))
case 51: /* mov addr,r ==> lu + or + lw (REGTMP) */
o1 = OP_IRR(opirr(ctxt, ALAST), uint32(0), uint32(REGZERO), uint32(REGTMP))
o2 = OP_IRR(opirr(ctxt, AOR), uint32(0), uint32(REGTMP), uint32(REGTMP))
rel := obj.Addrel(ctxt.Cursym)
rel.Off = int32(ctxt.Pc)
rel.Siz = 8
rel.Sym = p.From.Sym
rel.Add = p.From.Offset
rel.Type = obj.R_ADDRMIPS
o3 = OP_IRR(opirr(ctxt, int(p.As)+ALAST), uint32(0), uint32(REGTMP), uint32(p.To.Reg))
}
out[0] = o1
out[1] = o2
out[2] = o3
out[3] = o4
return
}
func progedit(ctxt *obj.Link, p *obj.Prog) {
p.From.Class = 0
p.To.Class = 0
// Rewrite B/BL to symbol as TYPE_BRANCH.
switch p.As {
case AB,
ABL,
obj.ADUFFZERO,
obj.ADUFFCOPY:
if p.To.Type == obj.TYPE_MEM && (p.To.Name == obj.NAME_EXTERN || p.To.Name == obj.NAME_STATIC) && p.To.Sym != nil {
p.To.Type = obj.TYPE_BRANCH
}
}
// Replace TLS register fetches on older ARM procesors.
switch p.As {
// Treat MRC 15, 0, <reg>, C13, C0, 3 specially.
case AMRC:
if p.To.Offset&0xffff0fff == 0xee1d0f70 {
// Because the instruction might be rewriten to a BL which returns in R0
// the register must be zero.
if p.To.Offset&0xf000 != 0 {
ctxt.Diag("%v: TLS MRC instruction must write to R0 as it might get translated into a BL instruction", p.Line())
}
if ctxt.Goarm < 7 {
// Replace it with BL runtime.read_tls_fallback(SB) for ARM CPUs that lack the tls extension.
if progedit_tlsfallback == nil {
progedit_tlsfallback = obj.Linklookup(ctxt, "runtime.read_tls_fallback", 0)
}
// MOVW LR, R11
p.As = AMOVW
p.From.Type = obj.TYPE_REG
p.From.Reg = REGLINK
p.To.Type = obj.TYPE_REG
p.To.Reg = REGTMP
// BL runtime.read_tls_fallback(SB)
p = obj.Appendp(ctxt, p)
p.As = ABL
p.To.Type = obj.TYPE_BRANCH
p.To.Sym = progedit_tlsfallback
p.To.Offset = 0
// MOVW R11, LR
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_REG
p.From.Reg = REGTMP
p.To.Type = obj.TYPE_REG
p.To.Reg = REGLINK
break
}
}
// Otherwise, MRC/MCR instructions need no further treatment.
p.As = AWORD
}
// Rewrite float constants to values stored in memory.
switch p.As {
case AMOVF:
if p.From.Type == obj.TYPE_FCONST && chipfloat5(ctxt, p.From.U.Dval) < 0 && (chipzero5(ctxt, p.From.U.Dval) < 0 || p.Scond&C_SCOND != C_SCOND_NONE) {
f32 := float32(p.From.U.Dval)
i32 := math.Float32bits(f32)
literal := fmt.Sprintf("$f32.%08x", i32)
s := obj.Linklookup(ctxt, literal, 0)
if s.Type == 0 {
s.Type = obj.SRODATA
obj.Adduint32(ctxt, s, i32)
s.Reachable = 0
}
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
case AMOVD:
if p.From.Type == obj.TYPE_FCONST && chipfloat5(ctxt, p.From.U.Dval) < 0 && (chipzero5(ctxt, p.From.U.Dval) < 0 || p.Scond&C_SCOND != C_SCOND_NONE) {
i64 := math.Float64bits(p.From.U.Dval)
literal := fmt.Sprintf("$f64.%016x", i64)
s := obj.Linklookup(ctxt, literal, 0)
if s.Type == 0 {
s.Type = obj.SRODATA
obj.Adduint64(ctxt, s, i64)
s.Reachable = 0
}
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
}
if ctxt.Flag_shared != 0 {
// Shared libraries use R_ARM_TLS_IE32 instead of
// R_ARM_TLS_LE32, replacing the link time constant TLS offset in
// runtime.tlsg with an address to a GOT entry containing the
// offset. Rewrite $runtime.tlsg(SB) to runtime.tlsg(SB) to
// compensate.
if ctxt.Tlsg == nil {
ctxt.Tlsg = obj.Linklookup(ctxt, "runtime.tlsg", 0)
}
if p.From.Type == obj.TYPE_ADDR && p.From.Name == obj.NAME_EXTERN && p.From.Sym == ctxt.Tlsg {
p.From.Type = obj.TYPE_MEM
}
if p.To.Type == obj.TYPE_ADDR && p.To.Name == obj.NAME_EXTERN && p.To.Sym == ctxt.Tlsg {
p.To.Type = obj.TYPE_MEM
}
}
}