func stacksplitPost(ctxt *obj.Link, p *obj.Prog, pPre *obj.Prog, pPreempt *obj.Prog) *obj.Prog {
// MOVD LR, R5
p = obj.Appendp(ctxt, p)
pPre.Pcond = p
p.As = AMOVD
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_LR
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R5
if pPreempt != nil {
pPreempt.Pcond = p
}
// BL runtime.morestack(SB)
p = obj.Appendp(ctxt, p)
p.As = ABL
p.To.Type = obj.TYPE_BRANCH
if ctxt.Cursym.Cfunc {
p.To.Sym = obj.Linklookup(ctxt, "runtime.morestackc", 0)
} else if ctxt.Cursym.Text.From3.Offset&obj.NEEDCTXT == 0 {
p.To.Sym = obj.Linklookup(ctxt, "runtime.morestack_noctxt", 0)
} else {
p.To.Sym = obj.Linklookup(ctxt, "runtime.morestack", 0)
}
// BR start
p = obj.Appendp(ctxt, p)
p.As = ABR
p.To.Type = obj.TYPE_BRANCH
p.Pcond = ctxt.Cursym.Text.Link
return p
}
// Called after regopt and peep have run.
// Expand CHECKNIL pseudo-op into actual nil pointer check.
func expandchecks(firstp *obj.Prog) {
var p1 *obj.Prog
var p2 *obj.Prog
for p := firstp; p != nil; p = p.Link {
if p.As != obj.ACHECKNIL {
continue
}
if gc.Debug_checknil != 0 && p.Lineno > 1 { // p->lineno==1 in generated wrappers
gc.Warnl(p.Lineno, "generated nil check")
}
// check is
// CMP arg, $0
// JNE 2(PC) (likely)
// MOV AX, 0
p1 = gc.Ctxt.NewProg()
p2 = gc.Ctxt.NewProg()
gc.Clearp(p1)
gc.Clearp(p2)
p1.Link = p2
p2.Link = p.Link
p.Link = p1
p1.Lineno = p.Lineno
p2.Lineno = p.Lineno
p1.Pc = 9999
p2.Pc = 9999
p.As = cmpptr
p.To.Type = obj.TYPE_CONST
p.To.Offset = 0
p1.As = x86.AJNE
p1.From.Type = obj.TYPE_CONST
p1.From.Offset = 1 // likely
p1.To.Type = obj.TYPE_BRANCH
p1.To.Val = p2.Link
// crash by write to memory address 0.
// if possible, since we know arg is 0, use 0(arg),
// which will be shorter to encode than plain 0.
p2.As = x86.AMOVL
p2.From.Type = obj.TYPE_REG
p2.From.Reg = x86.REG_AX
if regtyp(&p.From) {
p2.To.Type = obj.TYPE_MEM
p2.To.Reg = p.From.Reg
} else {
p2.To.Type = obj.TYPE_MEM
p2.To.Reg = x86.REG_NONE
}
p2.To.Offset = 0
}
}
func Prog(as obj.As) *obj.Prog {
var p *obj.Prog
if as == obj.AGLOBL {
if ddumped {
Fatalf("already dumped data")
}
if dpc == nil {
dpc = Ctxt.NewProg()
dfirst = dpc
}
p = dpc
dpc = Ctxt.NewProg()
p.Link = dpc
} else {
p = Pc
Pc = Ctxt.NewProg()
Clearp(Pc)
p.Link = Pc
}
if lineno == 0 && Debug['K'] != 0 {
Warn("prog: line 0")
}
p.As = as
p.Lineno = lineno
return p
}
func rewriteToPcrel(ctxt *obj.Link, p *obj.Prog) {
// RegTo2 is set on the instructions we insert here so they don't get
// processed twice.
if p.RegTo2 != 0 {
return
}
if p.As == obj.ATEXT || p.As == obj.AFUNCDATA || p.As == obj.ACALL || p.As == obj.ARET || p.As == obj.AJMP {
return
}
// Any Prog (aside from the above special cases) with an Addr with Name ==
// NAME_EXTERN, NAME_STATIC or NAME_GOTREF has a CALL __x86.get_pc_thunk.cx
// inserted before it.
isName := func(a *obj.Addr) bool {
if a.Sym == nil || (a.Type != obj.TYPE_MEM && a.Type != obj.TYPE_ADDR) || a.Reg != 0 {
return false
}
if a.Sym.Type == obj.STLSBSS {
return false
}
return a.Name == obj.NAME_EXTERN || a.Name == obj.NAME_STATIC || a.Name == obj.NAME_GOTREF
}
if isName(&p.From) && p.From.Type == obj.TYPE_ADDR {
// Handle things like "MOVL $sym, (SP)" or "PUSHL $sym" by rewriting
// to "MOVL $sym, CX; MOVL CX, (SP)" or "MOVL $sym, CX; PUSHL CX"
// respectively.
if p.To.Type != obj.TYPE_REG {
q := obj.Appendp(ctxt, p)
q.As = p.As
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_CX
q.To = p.To
p.As = AMOVL
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_CX
p.To.Sym = nil
p.To.Name = obj.NAME_NONE
}
}
if !isName(&p.From) && !isName(&p.To) && (p.From3 == nil || !isName(p.From3)) {
return
}
q := obj.Appendp(ctxt, p)
q.RegTo2 = 1
r := obj.Appendp(ctxt, q)
r.RegTo2 = 1
q.As = obj.ACALL
q.To.Sym = obj.Linklookup(ctxt, "__x86.get_pc_thunk.cx", 0)
q.To.Type = obj.TYPE_MEM
q.To.Name = obj.NAME_EXTERN
q.To.Sym.Local = true
r.As = p.As
r.Scond = p.Scond
r.From = p.From
r.From3 = p.From3
r.Reg = p.Reg
r.To = p.To
obj.Nopout(p)
}
// Called after regopt and peep have run.
// Expand CHECKNIL pseudo-op into actual nil pointer check.
func expandchecks(firstp *obj.Prog) {
var reg int
var p1 *obj.Prog
for p := firstp; p != nil; p = p.Link {
if p.As != obj.ACHECKNIL {
continue
}
if gc.Debug_checknil != 0 && p.Lineno > 1 { // p->lineno==1 in generated wrappers
gc.Warnl(p.Lineno, "generated nil check")
}
if p.From.Type != obj.TYPE_REG {
gc.Fatalf("invalid nil check %v", p)
}
reg = int(p.From.Reg)
// check is
// CMP arg, $0
// MOV.EQ arg, 0(arg)
p1 = gc.Ctxt.NewProg()
gc.Clearp(p1)
p1.Link = p.Link
p.Link = p1
p1.Lineno = p.Lineno
p1.Pc = 9999
p1.As = arm.AMOVW
p1.From.Type = obj.TYPE_REG
p1.From.Reg = int16(reg)
p1.To.Type = obj.TYPE_MEM
p1.To.Reg = int16(reg)
p1.To.Offset = 0
p1.Scond = arm.C_SCOND_EQ
p.As = arm.ACMP
p.From.Type = obj.TYPE_CONST
p.From.Reg = 0
p.From.Offset = 0
p.Reg = int16(reg)
}
}
// Called after regopt and peep have run.
// Expand CHECKNIL pseudo-op into actual nil pointer check.
func expandchecks(firstp *obj.Prog) {
var p1 *obj.Prog
for p := firstp; p != nil; p = p.Link {
if gc.Debug_checknil != 0 && gc.Ctxt.Debugvlog != 0 {
fmt.Printf("expandchecks: %v\n", p)
}
if p.As != obj.ACHECKNIL {
continue
}
if gc.Debug_checknil != 0 && p.Lineno > 1 { // p->lineno==1 in generated wrappers
gc.Warnl(p.Lineno, "generated nil check")
}
if p.From.Type != obj.TYPE_REG {
gc.Fatalf("invalid nil check %v\n", p)
}
// check is
// CBNZ arg, 2(PC)
// MOVD ZR, 0(arg)
p1 = gc.Ctxt.NewProg()
gc.Clearp(p1)
p1.Link = p.Link
p.Link = p1
p1.Lineno = p.Lineno
p1.Pc = 9999
p.As = arm64.ACBNZ
p.To.Type = obj.TYPE_BRANCH
p.To.Val = p1.Link
// crash by write to memory address 0.
p1.As = arm64.AMOVD
p1.From.Type = obj.TYPE_REG
p1.From.Reg = arm64.REGZERO
p1.To.Type = obj.TYPE_MEM
p1.To.Reg = p.From.Reg
p1.To.Offset = 0
}
}
/*
// instruction scheduling
if(debug['Q'] == 0)
return;
curtext = nil;
q = nil; // p - 1
q1 = firstp; // top of block
o = 0; // count of instructions
for(p = firstp; p != nil; p = p1) {
p1 = p->link;
o++;
if(p->mark & NOSCHED){
if(q1 != p){
sched(q1, q);
}
for(; p != nil; p = p->link){
if(!(p->mark & NOSCHED))
break;
q = p;
}
p1 = p;
q1 = p;
o = 0;
continue;
}
if(p->mark & (LABEL|SYNC)) {
if(q1 != p)
sched(q1, q);
q1 = p;
o = 1;
}
if(p->mark & (BRANCH|SYNC)) {
sched(q1, p);
q1 = p1;
o = 0;
}
if(o >= NSCHED) {
sched(q1, p);
q1 = p1;
o = 0;
}
q = p;
}
*/
func stacksplitPre(ctxt *obj.Link, p *obj.Prog, framesize int32) (*obj.Prog, *obj.Prog) {
var q *obj.Prog
// MOVD g_stackguard(g), R3
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGG
p.From.Offset = 2 * int64(ctxt.Arch.PtrSize) // G.stackguard0
if ctxt.Cursym.Cfunc {
p.From.Offset = 3 * int64(ctxt.Arch.PtrSize) // G.stackguard1
}
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R3
q = nil
if framesize <= obj.StackSmall {
// small stack: SP < stackguard
// CMP stackguard, SP
//p.To.Type = obj.TYPE_REG
//p.To.Reg = REGSP
// q1: BLT done
p = obj.Appendp(ctxt, p)
//q1 = p
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R3
p.Reg = REGSP
p.As = ACMPUBGE
p.To.Type = obj.TYPE_BRANCH
//p = obj.Appendp(ctxt, p)
//p.As = ACMPU
//p.From.Type = obj.TYPE_REG
//p.From.Reg = REG_R3
//p.To.Type = obj.TYPE_REG
//p.To.Reg = REGSP
//p = obj.Appendp(ctxt, p)
//p.As = ABGE
//p.To.Type = obj.TYPE_BRANCH
} else if framesize <= obj.StackBig {
// large stack: SP-framesize < stackguard-StackSmall
// ADD $-framesize, SP, R4
// CMP stackguard, R4
p = obj.Appendp(ctxt, p)
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(-framesize)
p.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
p = obj.Appendp(ctxt, p)
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R3
p.Reg = REG_R4
p.As = ACMPUBGE
p.To.Type = obj.TYPE_BRANCH
} else {
// Such a large stack we need to protect against wraparound.
// If SP is close to zero:
// SP-stackguard+StackGuard <= framesize + (StackGuard-StackSmall)
// The +StackGuard on both sides is required to keep the left side positive:
// SP is allowed to be slightly below stackguard. See stack.h.
//
// Preemption sets stackguard to StackPreempt, a very large value.
// That breaks the math above, so we have to check for that explicitly.
// // stackguard is R3
// CMP R3, $StackPreempt
// BEQ label-of-call-to-morestack
// ADD $StackGuard, SP, R4
// SUB R3, R4
// MOVD $(framesize+(StackGuard-StackSmall)), TEMP
// CMPUBGE TEMP, R4
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R3
p.To.Type = obj.TYPE_CONST
p.To.Offset = obj.StackPreempt
p = obj.Appendp(ctxt, p)
q = p
p.As = ABEQ
p.To.Type = obj.TYPE_BRANCH
p = obj.Appendp(ctxt, p)
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = obj.StackGuard
p.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
p = obj.Appendp(ctxt, p)
p.As = ASUB
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R3
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(framesize) + obj.StackGuard - obj.StackSmall
p.To.Type = obj.TYPE_REG
p.To.Reg = REGTMP
p = obj.Appendp(ctxt, p)
p.From.Type = obj.TYPE_REG
p.From.Reg = REGTMP
p.Reg = REG_R4
p.As = ACMPUBGE
p.To.Type = obj.TYPE_BRANCH
}
return p, q
}
// Called after regopt and peep have run.
// Expand CHECKNIL pseudo-op into actual nil pointer check.
func expandchecks(firstp *obj.Prog) {
var p1 *obj.Prog
var p2 *obj.Prog
for p := firstp; p != nil; p = p.Link {
if gc.Debug_checknil != 0 && gc.Ctxt.Debugvlog != 0 {
fmt.Printf("expandchecks: %v\n", p)
}
if p.As != obj.ACHECKNIL {
continue
}
if gc.Debug_checknil != 0 && p.Lineno > 1 { // p->lineno==1 in generated wrappers
gc.Warnl(p.Lineno, "generated nil check")
}
if p.From.Type != obj.TYPE_REG {
gc.Fatalf("invalid nil check %v\n", p)
}
/*
// check is
// TD $4, R0, arg (R0 is always zero)
// eqv. to:
// tdeq r0, arg
// NOTE: this needs special runtime support to make SIGTRAP recoverable.
reg = p->from.reg;
p->as = ATD;
p->from = p->to = p->from3 = zprog.from;
p->from.type = TYPE_CONST;
p->from.offset = 4;
p->from.reg = 0;
p->reg = REGZERO;
p->to.type = TYPE_REG;
p->to.reg = reg;
*/
// check is
// CMP arg, R0
// BNE 2(PC) [likely]
// MOVD R0, 0(R0)
p1 = gc.Ctxt.NewProg()
p2 = gc.Ctxt.NewProg()
gc.Clearp(p1)
gc.Clearp(p2)
p1.Link = p2
p2.Link = p.Link
p.Link = p1
p1.Lineno = p.Lineno
p2.Lineno = p.Lineno
p1.Pc = 9999
p2.Pc = 9999
p.As = ppc64.ACMP
p.To.Type = obj.TYPE_REG
p.To.Reg = ppc64.REGZERO
p1.As = ppc64.ABNE
//p1->from.type = TYPE_CONST;
//p1->from.offset = 1; // likely
p1.To.Type = obj.TYPE_BRANCH
p1.To.Val = p2.Link
// crash by write to memory address 0.
p2.As = ppc64.AMOVD
p2.From.Type = obj.TYPE_REG
p2.From.Reg = ppc64.REGZERO
p2.To.Type = obj.TYPE_MEM
p2.To.Reg = ppc64.REGZERO
p2.To.Offset = 0
}
}
func Clearp(p *obj.Prog) {
obj.Nopout(p)
p.As = obj.AEND
p.Pc = int64(pcloc)
pcloc++
}
func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32) *obj.Prog {
// MOVW g_stackguard(g), R1
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGG
p.From.Offset = 2 * int64(ctxt.Arch.PtrSize) // G.stackguard0
if ctxt.Cursym.Cfunc {
p.From.Offset = 3 * int64(ctxt.Arch.PtrSize) // G.stackguard1
}
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
if framesize <= obj.StackSmall {
// small stack: SP < stackguard
// CMP stackguard, SP
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.Reg = REGSP
} else if framesize <= obj.StackBig {
// large stack: SP-framesize < stackguard-StackSmall
// MOVW $-framesize(SP), R2
// CMP stackguard, R2
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_ADDR
p.From.Reg = REGSP
p.From.Offset = int64(-framesize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.Reg = REG_R2
} else {
// Such a large stack we need to protect against wraparound
// if SP is close to zero.
// SP-stackguard+StackGuard < framesize + (StackGuard-StackSmall)
// The +StackGuard on both sides is required to keep the left side positive:
// SP is allowed to be slightly below stackguard. See stack.h.
// CMP $StackPreempt, R1
// MOVW.NE $StackGuard(SP), R2
// SUB.NE R1, R2
// MOVW.NE $(framesize+(StackGuard-StackSmall)), R3
// CMP.NE R3, R2
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(uint32(obj.StackPreempt & (1<<32 - 1)))
p.Reg = REG_R1
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_ADDR
p.From.Reg = REGSP
p.From.Offset = obj.StackGuard
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p.Scond = C_SCOND_NE
p = obj.Appendp(ctxt, p)
p.As = ASUB
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p.Scond = C_SCOND_NE
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(framesize) + (obj.StackGuard - obj.StackSmall)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R3
p.Scond = C_SCOND_NE
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R3
p.Reg = REG_R2
p.Scond = C_SCOND_NE
}
// BLS call-to-morestack
bls := obj.Appendp(ctxt, p)
bls.As = ABLS
bls.To.Type = obj.TYPE_BRANCH
var last *obj.Prog
for last = ctxt.Cursym.Text; last.Link != nil; last = last.Link {
}
spfix := obj.Appendp(ctxt, last)
spfix.As = obj.ANOP
spfix.Spadj = -framesize
// MOVW LR, R3
movw := obj.Appendp(ctxt, spfix)
movw.As = AMOVW
movw.From.Type = obj.TYPE_REG
movw.From.Reg = REGLINK
movw.To.Type = obj.TYPE_REG
movw.To.Reg = REG_R3
bls.Pcond = movw
// BL runtime.morestack
call := obj.Appendp(ctxt, movw)
call.As = obj.ACALL
call.To.Type = obj.TYPE_BRANCH
morestack := "runtime.morestack"
switch {
case ctxt.Cursym.Cfunc:
morestack = "runtime.morestackc"
case ctxt.Cursym.Text.From3.Offset&obj.NEEDCTXT == 0:
morestack = "runtime.morestack_noctxt"
}
call.To.Sym = obj.Linklookup(ctxt, morestack, 0)
// B start
b := obj.Appendp(ctxt, call)
b.As = obj.AJMP
b.To.Type = obj.TYPE_BRANCH
b.Pcond = ctxt.Cursym.Text.Link
b.Spadj = +framesize
return bls
}
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 p1 *obj.Prog
var p2 *obj.Prog
var q2 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
o := 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 = true
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:
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 peep(firstp *obj.Prog) {
g := gc.Flowstart(firstp, nil)
if g == nil {
return
}
gactive = 0
// byte, word arithmetic elimination.
elimshortmov(g)
// constant propagation
// find MOV $con,R followed by
// another MOV $con,R without
// setting R in the interim
var p *obj.Prog
for r := g.Start; r != nil; r = r.Link {
p = r.Prog
switch p.As {
case x86.ALEAL:
if regtyp(&p.To) {
if p.From.Sym != nil {
if p.From.Index == x86.REG_NONE {
conprop(r)
}
}
}
case x86.AMOVB,
x86.AMOVW,
x86.AMOVL,
x86.AMOVSS,
x86.AMOVSD:
if regtyp(&p.To) {
if p.From.Type == obj.TYPE_CONST || p.From.Type == obj.TYPE_FCONST {
conprop(r)
}
}
}
}
var r1 *gc.Flow
var p1 *obj.Prog
var r *gc.Flow
var t int
loop1:
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
gc.Dumpit("loop1", g.Start, 0)
}
t = 0
for r = g.Start; r != nil; r = r.Link {
p = r.Prog
switch p.As {
case x86.AMOVL,
x86.AMOVSS,
x86.AMOVSD:
if regtyp(&p.To) {
if regtyp(&p.From) {
if copyprop(g, r) {
excise(r)
t++
} else if subprop(r) && copyprop(g, r) {
excise(r)
t++
}
}
}
case x86.AMOVBLZX,
x86.AMOVWLZX,
x86.AMOVBLSX,
x86.AMOVWLSX:
if regtyp(&p.To) {
r1 = rnops(gc.Uniqs(r))
if r1 != nil {
p1 = r1.Prog
if p.As == p1.As && p.To.Type == p1.From.Type && p.To.Reg == p1.From.Reg {
p1.As = x86.AMOVL
t++
}
}
}
case x86.AADDL,
x86.AADDW:
if p.From.Type != obj.TYPE_CONST || needc(p.Link) {
break
}
if p.From.Offset == -1 {
if p.As == x86.AADDL {
p.As = x86.ADECL
} else {
p.As = x86.ADECW
}
p.From = obj.Addr{}
break
}
if p.From.Offset == 1 {
if p.As == x86.AADDL {
p.As = x86.AINCL
} else {
p.As = x86.AINCW
}
p.From = obj.Addr{}
break
}
case x86.ASUBL,
x86.ASUBW:
if p.From.Type != obj.TYPE_CONST || needc(p.Link) {
break
}
if p.From.Offset == -1 {
if p.As == x86.ASUBL {
p.As = x86.AINCL
} else {
p.As = x86.AINCW
}
p.From = obj.Addr{}
break
}
if p.From.Offset == 1 {
if p.As == x86.ASUBL {
p.As = x86.ADECL
} else {
p.As = x86.ADECW
}
p.From = obj.Addr{}
break
}
}
}
if t != 0 {
goto loop1
}
// MOVSD removal.
// We never use packed registers, so a MOVSD between registers
// can be replaced by MOVAPD, which moves the pair of float64s
// instead of just the lower one. We only use the lower one, but
// the processor can do better if we do moves using both.
for r := g.Start; r != nil; r = r.Link {
p = r.Prog
if p.As == x86.AMOVSD {
if regtyp(&p.From) {
if regtyp(&p.To) {
p.As = x86.AMOVAPD
}
}
}
}
gc.Flowend(g)
}
func xfol(ctxt *obj.Link, p *obj.Prog, last **obj.Prog) {
var q *obj.Prog
var r *obj.Prog
var b obj.As
for p != nil {
a := 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
(*last).Pc = pc_cnt
pc_cnt += 1
p = p.Link
xfol(ctxt, p, last)
p = q
if p != nil && p.Mark&FOLL == 0 {
continue
}
return
}
if q != nil {
p.Mark |= FOLL
p = q
if p.Mark&FOLL == 0 {
continue
}
}
}
if p.Mark&FOLL != 0 {
q = p
for i := 0; i < 4; i, q = i+1, q.Link {
if q == *last || (q.Mark&NOSCHED != 0) {
break
}
b = 0 /* set */
a = q.As
if a == obj.ANOP {
i--
continue
}
if a != ABR && a != obj.ARET {
if q.Pcond == nil || (q.Pcond.Mark&FOLL != 0) {
continue
}
b = relinv(a)
if b == 0 {
continue
}
}
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
(*last).Pc = pc_cnt
pc_cnt += 1
continue
}
(*last).Link = r
*last = r
(*last).Pc = pc_cnt
pc_cnt += 1
if a == ABR || a == obj.ARET {
return
}
r.As = 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("can't happen 2\n")
}
return
}
}
a = ABR
q = ctxt.NewProg()
q.As = 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
(*last).Pc = pc_cnt
pc_cnt += 1
if a == ABR || a == obj.ARET {
if p.Mark&NOSCHED != 0 {
p = p.Link
continue
}
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
}
continue
}
}
p = p.Link
}
}
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.Family == sys.AMD64 || 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 && p.Mode == 32 {
rewriteToPcrel(ctxt, p)
}
}
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 {
for q := p; q != nil; q = q.Link {
if q.As == obj.ACALL {
goto noleaf
}
if (q.As == obj.ADUFFCOPY || q.As == obj.ADUFFZERO) && autoffset >= obj.StackSmall-8 {
goto noleaf
}
}
p.From3.Offset |= obj.NOSPLIT
noleaf:
}
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
} else {
// zero-byte stack adjustment.
// Insert a fake non-zero adjustment so that stkcheck can
// recognize the end of the stack-splitting prolog.
p = obj.Appendp(ctxt, p)
p.As = obj.ANOP
p.Spadj = int32(-ctxt.Arch.PtrSize)
p = obj.Appendp(ctxt, p)
p.As = obj.ANOP
p.Spadj = int32(ctxt.Arch.PtrSize)
}
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
}
var a int
var pcsize int
for ; p != nil; p = p.Link {
pcsize = int(p.Mode) / 8
a = int(p.From.Name)
if a == obj.NAME_AUTO {
p.From.Offset += int64(deltasp) - int64(bpsize)
}
if a == obj.NAME_PARAM {
p.From.Offset += int64(deltasp) + int64(pcsize)
}
if p.From3 != nil {
a = int(p.From3.Name)
if a == obj.NAME_AUTO {
p.From3.Offset += int64(deltasp) - int64(bpsize)
}
if a == obj.NAME_PARAM {
p.From3.Offset += int64(deltasp) + int64(pcsize)
}
}
a = int(p.To.Name)
if a == obj.NAME_AUTO {
p.To.Offset += int64(deltasp) - int64(bpsize)
}
if a == 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:
break
}
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
}
}
}
func indir_cx(ctxt *obj.Link, p *obj.Prog, a *obj.Addr) {
if ctxt.Headtype == obj.Hnacl && p.Mode == 64 {
a.Type = obj.TYPE_MEM
a.Reg = REG_R15
a.Index = REG_CX
a.Scale = 1
return
}
a.Type = obj.TYPE_MEM
a.Reg = REG_CX
}
// Append code to p to load g into cx.
// Overwrites p with the first instruction (no first appendp).
// Overwriting p is unusual but it lets use this in both the
// prologue (caller must call appendp first) and in the epilogue.
// Returns last new instruction.
func load_g_cx(ctxt *obj.Link, p *obj.Prog) *obj.Prog {
p.As = AMOVQ
if ctxt.Arch.PtrSize == 4 {
p.As = AMOVL
}
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_TLS
p.From.Offset = 0
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_CX
next := p.Link
progedit(ctxt, p)
for p.Link != next {
p = p.Link
}
if p.From.Index == REG_TLS {
p.From.Scale = 2
}
return p
}
// Append code to p to check for stack split.
// Appends to (does not overwrite) p.
// Assumes g is in CX.
// Returns last new instruction.
func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32, textarg int32) *obj.Prog {
cmp := ACMPQ
lea := ALEAQ
mov := AMOVQ
sub := ASUBQ
if ctxt.Headtype == obj.Hnacl || p.Mode == 32 {
cmp = ACMPL
lea = ALEAL
mov = AMOVL
sub = ASUBL
}
var q1 *obj.Prog
if framesize <= obj.StackSmall {
// small stack: SP <= stackguard
// CMPQ SP, stackguard
p = obj.Appendp(ctxt, p)
p.As = cmp
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_SP
indir_cx(ctxt, p, &p.To)
p.To.Offset = 2 * int64(ctxt.Arch.PtrSize) // G.stackguard0
if ctxt.Cursym.Cfunc {
p.To.Offset = 3 * int64(ctxt.Arch.PtrSize) // G.stackguard1
}
} else if framesize <= obj.StackBig {
// large stack: SP-framesize <= stackguard-StackSmall
// LEAQ -xxx(SP), AX
// CMPQ AX, stackguard
p = obj.Appendp(ctxt, p)
p.As = lea
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_SP
p.From.Offset = -(int64(framesize) - obj.StackSmall)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_AX
p = obj.Appendp(ctxt, p)
p.As = cmp
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_AX
indir_cx(ctxt, p, &p.To)
p.To.Offset = 2 * int64(ctxt.Arch.PtrSize) // G.stackguard0
if ctxt.Cursym.Cfunc {
p.To.Offset = 3 * int64(ctxt.Arch.PtrSize) // G.stackguard1
}
} else {
// Such a large stack we need to protect against wraparound.
// If SP is close to zero:
// SP-stackguard+StackGuard <= framesize + (StackGuard-StackSmall)
// The +StackGuard on both sides is required to keep the left side positive:
// SP is allowed to be slightly below stackguard. See stack.h.
//
// Preemption sets stackguard to StackPreempt, a very large value.
// That breaks the math above, so we have to check for that explicitly.
// MOVQ stackguard, CX
// CMPQ CX, $StackPreempt
// JEQ label-of-call-to-morestack
// LEAQ StackGuard(SP), AX
// SUBQ CX, AX
// CMPQ AX, $(framesize+(StackGuard-StackSmall))
p = obj.Appendp(ctxt, p)
p.As = mov
indir_cx(ctxt, p, &p.From)
p.From.Offset = 2 * int64(ctxt.Arch.PtrSize) // G.stackguard0
if ctxt.Cursym.Cfunc {
p.From.Offset = 3 * int64(ctxt.Arch.PtrSize) // G.stackguard1
}
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_SI
p = obj.Appendp(ctxt, p)
p.As = cmp
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_SI
p.To.Type = obj.TYPE_CONST
p.To.Offset = obj.StackPreempt
if p.Mode == 32 {
p.To.Offset = int64(uint32(obj.StackPreempt & (1<<32 - 1)))
}
p = obj.Appendp(ctxt, p)
p.As = AJEQ
p.To.Type = obj.TYPE_BRANCH
q1 = p
p = obj.Appendp(ctxt, p)
p.As = lea
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_SP
p.From.Offset = obj.StackGuard
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_AX
p = obj.Appendp(ctxt, p)
p.As = sub
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_SI
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_AX
p = obj.Appendp(ctxt, p)
p.As = cmp
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_AX
p.To.Type = obj.TYPE_CONST
p.To.Offset = int64(framesize) + (obj.StackGuard - obj.StackSmall)
}
// common
jls := obj.Appendp(ctxt, p)
jls.As = AJLS
jls.To.Type = obj.TYPE_BRANCH
var last *obj.Prog
for last = ctxt.Cursym.Text; last.Link != nil; last = last.Link {
}
spfix := obj.Appendp(ctxt, last)
spfix.As = obj.ANOP
spfix.Spadj = -framesize
call := obj.Appendp(ctxt, spfix)
call.Lineno = ctxt.Cursym.Text.Lineno
call.Mode = ctxt.Cursym.Text.Mode
call.As = obj.ACALL
call.To.Type = obj.TYPE_BRANCH
call.To.Name = obj.NAME_EXTERN
morestack := "runtime.morestack"
switch {
case ctxt.Cursym.Cfunc:
morestack = "runtime.morestackc"
case ctxt.Cursym.Text.From3Offset()&obj.NEEDCTXT == 0:
morestack = "runtime.morestack_noctxt"
}
call.To.Sym = obj.Linklookup(ctxt, morestack, 0)
// When compiling 386 code for dynamic linking, the call needs to be adjusted
// to follow PIC rules. This in turn can insert more instructions, so we need
// to keep track of the start of the call (where the jump will be to) and the
// end (which following instructions are appended to).
callend := call
progedit(ctxt, callend)
for ; callend.Link != nil; callend = callend.Link {
progedit(ctxt, callend.Link)
}
jmp := obj.Appendp(ctxt, callend)
jmp.As = obj.AJMP
jmp.To.Type = obj.TYPE_BRANCH
jmp.Pcond = ctxt.Cursym.Text.Link
jmp.Spadj = +framesize
jls.Pcond = call
if q1 != nil {
q1.Pcond = call
}
return jls
}
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 nofollow(a obj.As) bool {
switch a {
case obj.AJMP,
obj.ARET,
AIRETL,
AIRETQ,
AIRETW,
ARETFL,
ARETFQ,
ARETFW,
obj.AUNDEF:
return true
}
return false
}
func pushpop(a obj.As) bool {
switch a {
case APUSHL,
APUSHFL,
APUSHQ,
APUSHFQ,
APUSHW,
APUSHFW,
APOPL,
APOPFL,
APOPQ,
APOPFQ,
APOPW,
APOPFW:
return true
}
return false
}
func relinv(a obj.As) obj.As {
switch a {
case AJEQ:
return AJNE
case AJNE:
return AJEQ
case AJLE:
return AJGT
case AJLS:
return AJHI
case AJLT:
return AJGE
case AJMI:
return AJPL
case AJGE:
return AJLT
case AJPL:
return AJMI
case AJGT:
return AJLE
case AJHI:
return AJLS
case AJCS:
return AJCC
case AJCC:
return AJCS
case AJPS:
return AJPC
case AJPC:
return AJPS
case AJOS:
return AJOC
case AJOC:
return AJOS
}
log.Fatalf("unknown relation: %s", 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
}
var unaryDst = map[obj.As]bool{
ABSWAPL: true,
ABSWAPQ: true,
ACMPXCHG8B: true,
ADECB: true,
ADECL: true,
ADECQ: true,
ADECW: true,
AINCB: true,
AINCL: true,
AINCQ: true,
AINCW: true,
ANEGB: true,
ANEGL: true,
ANEGQ: true,
ANEGW: true,
ANOTB: true,
ANOTL: true,
ANOTQ: true,
ANOTW: true,
APOPL: true,
APOPQ: true,
APOPW: true,
ASETCC: true,
ASETCS: true,
ASETEQ: true,
ASETGE: true,
ASETGT: true,
ASETHI: true,
ASETLE: true,
ASETLS: true,
ASETLT: true,
ASETMI: true,
ASETNE: true,
ASETOC: true,
ASETOS: true,
ASETPC: true,
ASETPL: true,
ASETPS: true,
AFFREE: true,
AFLDENV: true,
AFSAVE: true,
AFSTCW: true,
AFSTENV: true,
AFSTSW: true,
AFXSAVE: true,
AFXSAVE64: true,
ASTMXCSR: true,
}
var Linkamd64 = obj.LinkArch{
Arch: sys.ArchAMD64,
Preprocess: preprocess,
Assemble: span6,
Follow: follow,
Progedit: progedit,
UnaryDst: unaryDst,
}
var Linkamd64p32 = obj.LinkArch{
Arch: sys.ArchAMD64P32,
Preprocess: preprocess,
Assemble: span6,
Follow: follow,
Progedit: progedit,
UnaryDst: unaryDst,
}
var Link386 = obj.LinkArch{
Arch: sys.Arch386,
Preprocess: preprocess,
Assemble: span6,
Follow: follow,
Progedit: progedit,
UnaryDst: unaryDst,
}
// 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 peep(firstp *obj.Prog) {
g := gc.Flowstart(firstp, nil)
if g == nil {
return
}
gactive = 0
// byte, word arithmetic elimination.
elimshortmov(g)
// constant propagation
// find MOV $con,R followed by
// another MOV $con,R without
// setting R in the interim
var p *obj.Prog
for r := g.Start; r != nil; r = r.Link {
p = r.Prog
switch p.As {
case x86.ALEAL,
x86.ALEAQ:
if regtyp(&p.To) {
if p.From.Sym != nil {
if p.From.Index == x86.REG_NONE {
conprop(r)
}
}
}
case x86.AMOVB,
x86.AMOVW,
x86.AMOVL,
x86.AMOVQ,
x86.AMOVSS,
x86.AMOVSD:
if regtyp(&p.To) {
if p.From.Type == obj.TYPE_CONST || p.From.Type == obj.TYPE_FCONST {
conprop(r)
}
}
}
}
var r *gc.Flow
var r1 *gc.Flow
var p1 *obj.Prog
var t int
loop1:
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
gc.Dumpit("loop1", g.Start, 0)
}
t = 0
for r = g.Start; r != nil; r = r.Link {
p = r.Prog
switch p.As {
case x86.AMOVL,
x86.AMOVQ,
x86.AMOVSS,
x86.AMOVSD:
if regtyp(&p.To) {
if regtyp(&p.From) {
if copyprop(g, r) {
excise(r)
t++
} else if subprop(r) && copyprop(g, r) {
excise(r)
t++
}
}
}
case x86.AMOVBLZX,
x86.AMOVWLZX,
x86.AMOVBLSX,
x86.AMOVWLSX:
if regtyp(&p.To) {
r1 = rnops(gc.Uniqs(r))
if r1 != nil {
p1 = r1.Prog
if p.As == p1.As && p.To.Type == p1.From.Type && p.To.Reg == p1.From.Reg {
p1.As = x86.AMOVL
t++
}
}
}
case x86.AMOVBQSX,
x86.AMOVBQZX,
x86.AMOVWQSX,
x86.AMOVWQZX,
x86.AMOVLQSX,
x86.AMOVLQZX,
x86.AMOVQL:
if regtyp(&p.To) {
r1 = rnops(gc.Uniqs(r))
if r1 != nil {
p1 = r1.Prog
if p.As == p1.As && p.To.Type == p1.From.Type && p.To.Reg == p1.From.Reg {
p1.As = x86.AMOVQ
t++
}
}
}
case x86.AADDL,
x86.AADDQ,
x86.AADDW:
if p.From.Type != obj.TYPE_CONST || needc(p.Link) {
break
}
if p.From.Offset == -1 {
if p.As == x86.AADDQ {
p.As = x86.ADECQ
} else if p.As == x86.AADDL {
p.As = x86.ADECL
} else {
p.As = x86.ADECW
}
p.From = obj.Addr{}
break
}
if p.From.Offset == 1 {
if p.As == x86.AADDQ {
p.As = x86.AINCQ
} else if p.As == x86.AADDL {
p.As = x86.AINCL
} else {
p.As = x86.AINCW
}
p.From = obj.Addr{}
break
}
case x86.ASUBL,
x86.ASUBQ,
x86.ASUBW:
if p.From.Type != obj.TYPE_CONST || needc(p.Link) {
break
}
if p.From.Offset == -1 {
if p.As == x86.ASUBQ {
p.As = x86.AINCQ
} else if p.As == x86.ASUBL {
p.As = x86.AINCL
} else {
p.As = x86.AINCW
}
p.From = obj.Addr{}
break
}
if p.From.Offset == 1 {
if p.As == x86.ASUBQ {
p.As = x86.ADECQ
} else if p.As == x86.ASUBL {
p.As = x86.ADECL
} else {
p.As = x86.ADECW
}
p.From = obj.Addr{}
break
}
}
}
if t != 0 {
goto loop1
}
// MOVLQZX removal.
// The MOVLQZX exists to avoid being confused for a
// MOVL that is just copying 32-bit data around during
// copyprop. Now that copyprop is done, remov MOVLQZX R1, R2
// if it is dominated by an earlier ADDL/MOVL/etc into R1 that
// will have already cleared the high bits.
//
// MOVSD removal.
// We never use packed registers, so a MOVSD between registers
// can be replaced by MOVAPD, which moves the pair of float64s
// instead of just the lower one. We only use the lower one, but
// the processor can do better if we do moves using both.
for r := g.Start; r != nil; r = r.Link {
p = r.Prog
if p.As == x86.AMOVLQZX {
if regtyp(&p.From) {
if p.From.Type == p.To.Type && p.From.Reg == p.To.Reg {
if prevl(r, p.From.Reg) {
excise(r)
}
}
}
}
if p.As == x86.AMOVSD {
if regtyp(&p.From) {
if regtyp(&p.To) {
p.As = x86.AMOVAPD
}
}
}
}
// load pipelining
// push any load from memory as early as possible
// to give it time to complete before use.
for r := g.Start; r != nil; r = r.Link {
p = r.Prog
switch p.As {
case x86.AMOVB,
x86.AMOVW,
x86.AMOVL,
x86.AMOVQ,
x86.AMOVLQZX:
if regtyp(&p.To) && !regconsttyp(&p.From) {
pushback(r)
}
}
}
gc.Flowend(g)
}
// movb elimination.
// movb is simulated by the linker
// when a register other than ax, bx, cx, dx
// is used, so rewrite to other instructions
// when possible. a movb into a register
// can smash the entire 32-bit register without
// causing any trouble.
//
// TODO: Using the Q forms here instead of the L forms
// seems unnecessary, and it makes the instructions longer.
func elimshortmov(g *gc.Graph) {
var p *obj.Prog
for r := g.Start; r != nil; r = r.Link {
p = r.Prog
if regtyp(&p.To) {
switch p.As {
case x86.AINCB,
x86.AINCW:
p.As = x86.AINCQ
case x86.ADECB,
x86.ADECW:
p.As = x86.ADECQ
case x86.ANEGB,
x86.ANEGW:
p.As = x86.ANEGQ
case x86.ANOTB,
x86.ANOTW:
p.As = x86.ANOTQ
}
if regtyp(&p.From) || p.From.Type == obj.TYPE_CONST {
// move or arithmetic into partial register.
// from another register or constant can be movl.
// we don't switch to 64-bit arithmetic if it can
// change how the carry bit is set (and the carry bit is needed).
switch p.As {
case x86.AMOVB,
x86.AMOVW:
p.As = x86.AMOVQ
case x86.AADDB,
x86.AADDW:
if !needc(p.Link) {
p.As = x86.AADDQ
}
case x86.ASUBB,
x86.ASUBW:
if !needc(p.Link) {
p.As = x86.ASUBQ
}
case x86.AMULB,
x86.AMULW:
p.As = x86.AMULQ
case x86.AIMULB,
x86.AIMULW:
p.As = x86.AIMULQ
case x86.AANDB,
x86.AANDW:
p.As = x86.AANDQ
case x86.AORB,
x86.AORW:
p.As = x86.AORQ
case x86.AXORB,
x86.AXORW:
p.As = x86.AXORQ
case x86.ASHLB,
x86.ASHLW:
p.As = x86.ASHLQ
}
} else if p.From.Type != obj.TYPE_REG {
// explicit zero extension, but don't
// do that if source is a byte register
// (only AH can occur and it's forbidden).
switch p.As {
case x86.AMOVB:
p.As = x86.AMOVBQZX
case x86.AMOVW:
p.As = x86.AMOVWQZX
}
}
}
}
}
// movb elimination.
// movb is simulated by the linker
// when a register other than ax, bx, cx, dx
// is used, so rewrite to other instructions
// when possible. a movb into a register
// can smash the entire 64-bit register without
// causing any trouble.
func elimshortmov(g *gc.Graph) {
var p *obj.Prog
for r := g.Start; r != nil; r = r.Link {
p = r.Prog
if regtyp(&p.To) {
switch p.As {
case x86.AINCB,
x86.AINCW:
p.As = x86.AINCL
case x86.ADECB,
x86.ADECW:
p.As = x86.ADECL
case x86.ANEGB,
x86.ANEGW:
p.As = x86.ANEGL
case x86.ANOTB,
x86.ANOTW:
p.As = x86.ANOTL
}
if regtyp(&p.From) || p.From.Type == obj.TYPE_CONST {
// move or arithmetic into partial register.
// from another register or constant can be movl.
// we don't switch to 32-bit arithmetic if it can
// change how the carry bit is set (and the carry bit is needed).
switch p.As {
case x86.AMOVB,
x86.AMOVW:
p.As = x86.AMOVL
case x86.AADDB,
x86.AADDW:
if !needc(p.Link) {
p.As = x86.AADDL
}
case x86.ASUBB,
x86.ASUBW:
if !needc(p.Link) {
p.As = x86.ASUBL
}
case x86.AMULB,
x86.AMULW:
p.As = x86.AMULL
case x86.AIMULB,
x86.AIMULW:
p.As = x86.AIMULL
case x86.AANDB,
x86.AANDW:
p.As = x86.AANDL
case x86.AORB,
x86.AORW:
p.As = x86.AORL
case x86.AXORB,
x86.AXORW:
p.As = x86.AXORL
case x86.ASHLB,
x86.ASHLW:
p.As = x86.ASHLL
}
} else {
// explicit zero extension
switch p.As {
case x86.AMOVB:
p.As = x86.AMOVBLZX
case x86.AMOVW:
p.As = x86.AMOVWLZX
}
}
}
}
}
// fuseMultiple merges memory loads and stores into load multiple and
// store multiple operations.
//
// Looks for this pattern (sequence of loads or stores):
// MOVD R1, 0(R15)
// MOVD R2, 8(R15)
// MOVD R3, 16(R15)
// Replaces with:
// STMG R1, R3, 0(R15)
func fuseMultiple(r *gc.Flow) int {
n := 0
var fused *obj.Prog
for ; r != nil; r = r.Link {
// If there is a branch into the instruction stream then
// we can't fuse into previous instructions.
if gc.Uniqp(r) == nil {
fused = nil
}
p := r.Prog
isStore := isGPR(&p.From) && isBDMem(&p.To)
isLoad := isGPR(&p.To) && isBDMem(&p.From)
// are we a candidate?
size := int64(0)
switch p.As {
default:
fused = nil
continue
case obj.ANOP:
// skip over nops
continue
case s390x.AMOVW, s390x.AMOVWZ:
size = 4
// TODO(mundaym): 32-bit load multiple is currently not supported
// as it requires sign/zero extension.
if !isStore {
fused = nil
continue
}
case s390x.AMOVD:
size = 8
if !isLoad && !isStore {
fused = nil
continue
}
}
// If we merge two loads/stores with different source/destination Nodes
// then we will lose a reference the second Node which means that the
// compiler might mark the Node as unused and free its slot on the stack.
// TODO(mundaym): allow this by adding a dummy reference to the Node.
if fused == nil ||
fused.From.Node != p.From.Node ||
fused.From.Type != p.From.Type ||
fused.To.Node != p.To.Node ||
fused.To.Type != p.To.Type {
fused = p
continue
}
// check two addresses
ca := func(a, b *obj.Addr, offset int64) bool {
return a.Reg == b.Reg && a.Offset+offset == b.Offset &&
a.Sym == b.Sym && a.Name == b.Name
}
switch fused.As {
default:
fused = p
case s390x.AMOVW, s390x.AMOVWZ:
if size == 4 && fused.From.Reg+1 == p.From.Reg && ca(&fused.To, &p.To, 4) {
fused.As = s390x.ASTMY
fused.Reg = p.From.Reg
excise(r)
n++
} else {
fused = p
}
case s390x.AMOVD:
if size == 8 && fused.From.Reg+1 == p.From.Reg && ca(&fused.To, &p.To, 8) {
fused.As = s390x.ASTMG
fused.Reg = p.From.Reg
excise(r)
n++
} else if size == 8 && fused.To.Reg+1 == p.To.Reg && ca(&fused.From, &p.From, 8) {
fused.As = s390x.ALMG
fused.Reg = fused.To.Reg
fused.To.Reg = p.To.Reg
excise(r)
n++
} else {
fused = p
}
case s390x.ASTMG, s390x.ASTMY:
if (fused.As == s390x.ASTMY && size != 4) ||
(fused.As == s390x.ASTMG && size != 8) {
fused = p
continue
}
offset := size * int64(fused.Reg-fused.From.Reg+1)
if fused.Reg+1 == p.From.Reg && ca(&fused.To, &p.To, offset) {
fused.Reg = p.From.Reg
excise(r)
n++
} else {
fused = p
}
case s390x.ALMG:
offset := 8 * int64(fused.To.Reg-fused.Reg+1)
if size == 8 && fused.To.Reg+1 == p.To.Reg && ca(&fused.From, &p.From, offset) {
fused.To.Reg = p.To.Reg
excise(r)
n++
} else {
fused = p
}
}
}
return n
}
// 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
// MOVW runtime.duffxxx@GOT, R9
// ADD $offset, R9
// CALL (R9)
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 = AMOVW
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_R9
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_R9
p2 := obj.Appendp(ctxt, p1)
p2.As = obj.ACALL
p2.To.Type = obj.TYPE_MEM
p2.To.Reg = REG_R9
return
}
// 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 {
// MOVW $sym, Rx becomes MOVW sym@GOT, Rx
// MOVW $sym+<off>, Rx becomes MOVW sym@GOT, Rx; ADD <off>, Rx
if p.As != AMOVW {
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 MOVW sym@GOT, R9; MOVx (R9), Ry
// MOVx Ry, sym becomes MOVW sym@GOT, R9; MOVx Ry, (R9)
// 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 = AMOVW
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_R9
p2.As = p.As
p2.From = p.From
p2.To = p.To
if p.From.Name == obj.NAME_EXTERN {
p2.From.Reg = REG_R9
p2.From.Name = obj.NAME_NONE
p2.From.Sym = nil
} else if p.To.Name == obj.NAME_EXTERN {
p2.To.Reg = REG_R9
p2.To.Name = obj.NAME_NONE
p2.To.Sym = nil
} else {
return
}
obj.Nopout(p)
}
func progedit(ctxt *obj.Link, p *obj.Prog) {
p.From.Class = 0
p.To.Class = 0
// $0 results in C_ZCON, which matches both C_REG and various
// C_xCON, however the C_REG cases in asmout don't expect a
// constant, so they will use the register fields and assemble
// a R0. To prevent that, rewrite $0 as ZR.
if p.From.Type == obj.TYPE_CONST && p.From.Offset == 0 {
p.From.Type = obj.TYPE_REG
p.From.Reg = REGZERO
}
if p.To.Type == obj.TYPE_CONST && p.To.Offset == 0 {
p.To.Type = obj.TYPE_REG
p.To.Reg = REGZERO
}
// Rewrite BR/BL to symbol as TYPE_BRANCH.
switch p.As {
case AB,
ABL,
obj.ARET,
obj.ADUFFZERO,
obj.ADUFFCOPY:
if p.To.Sym != nil {
p.To.Type = obj.TYPE_BRANCH
}
break
}
// Rewrite float constants to values stored in memory.
switch p.As {
case AFMOVS:
if p.From.Type == obj.TYPE_FCONST {
f32 := float32(p.From.Val.(float64))
i32 := math.Float32bits(f32)
literal := fmt.Sprintf("$f32.%08x", i32)
s := obj.Linklookup(ctxt, literal, 0)
s.Size = 4
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Sym.Local = true
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
case AFMOVD:
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)
s.Size = 8
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Sym.Local = true
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
break
}
// Rewrite negative immediates as positive immediates with
// complementary instruction.
switch p.As {
case AADD,
AADDW,
ASUB,
ASUBW,
ACMP,
ACMPW,
ACMN,
ACMNW:
if p.From.Type == obj.NAME_EXTERN && p.From.Offset < 0 {
p.From.Offset = -p.From.Offset
p.As = complements[p.As]
}
break
}
if ctxt.Flag_dynlink {
rewriteToUseGot(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 processors.
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 rewritten 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
}
}
if ctxt.Flag_dynlink {
rewriteToUseGot(ctxt, p)
}
}
func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32) *obj.Prog {
// MOV g_stackguard(g), R1
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGG
p.From.Offset = 2 * int64(ctxt.Arch.PtrSize) // G.stackguard0
if ctxt.Cursym.Cfunc {
p.From.Offset = 3 * int64(ctxt.Arch.PtrSize) // G.stackguard1
}
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
q := (*obj.Prog)(nil)
if framesize <= obj.StackSmall {
// small stack: SP < stackguard
// MOV SP, R2
// CMP stackguard, R2
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_REG
p.From.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.Reg = REG_R2
} else if framesize <= obj.StackBig {
// large stack: SP-framesize < stackguard-StackSmall
// SUB $framesize, SP, R2
// CMP stackguard, R2
p = obj.Appendp(ctxt, p)
p.As = ASUB
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(framesize)
p.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.Reg = REG_R2
} else {
// Such a large stack we need to protect against wraparound
// if SP is close to zero.
// SP-stackguard+StackGuard < framesize + (StackGuard-StackSmall)
// The +StackGuard on both sides is required to keep the left side positive:
// SP is allowed to be slightly below stackguard. See stack.h.
// CMP $StackPreempt, R1
// BEQ label_of_call_to_morestack
// ADD $StackGuard, SP, R2
// SUB R1, R2
// MOV $(framesize+(StackGuard-StackSmall)), R3
// CMP R3, R2
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_CONST
p.From.Offset = obj.StackPreempt
p.Reg = REG_R1
p = obj.Appendp(ctxt, p)
q = p
p.As = ABEQ
p.To.Type = obj.TYPE_BRANCH
p = obj.Appendp(ctxt, p)
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = obj.StackGuard
p.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = ASUB
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(framesize) + (obj.StackGuard - obj.StackSmall)
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_R3
p.Reg = REG_R2
}
// BLS do-morestack
bls := obj.Appendp(ctxt, p)
bls.As = ABLS
bls.To.Type = obj.TYPE_BRANCH
var last *obj.Prog
for last = ctxt.Cursym.Text; last.Link != nil; last = last.Link {
}
spfix := obj.Appendp(ctxt, last)
spfix.As = obj.ANOP
spfix.Spadj = -framesize
// MOV LR, R3
movlr := obj.Appendp(ctxt, spfix)
movlr.As = AMOVD
movlr.From.Type = obj.TYPE_REG
movlr.From.Reg = REGLINK
movlr.To.Type = obj.TYPE_REG
movlr.To.Reg = REG_R3
if q != nil {
q.Pcond = movlr
}
bls.Pcond = movlr
debug := movlr
if false {
debug = obj.Appendp(ctxt, debug)
debug.As = AMOVD
debug.From.Type = obj.TYPE_CONST
debug.From.Offset = int64(framesize)
debug.To.Type = obj.TYPE_REG
debug.To.Reg = REGTMP
}
// BL runtime.morestack(SB)
call := obj.Appendp(ctxt, debug)
call.As = ABL
call.To.Type = obj.TYPE_BRANCH
morestack := "runtime.morestack"
switch {
case ctxt.Cursym.Cfunc:
morestack = "runtime.morestackc"
case ctxt.Cursym.Text.From3.Offset&obj.NEEDCTXT == 0:
morestack = "runtime.morestack_noctxt"
}
call.To.Sym = obj.Linklookup(ctxt, morestack, 0)
// B start
jmp := obj.Appendp(ctxt, call)
jmp.As = AB
jmp.To.Type = obj.TYPE_BRANCH
jmp.Pcond = ctxt.Cursym.Text.Link
jmp.Spadj = +framesize
// placeholder for bls's jump target
// p = obj.Appendp(ctxt, p)
// p.As = obj.ANOP
return bls
}
func xfol(ctxt *obj.Link, p *obj.Prog, last **obj.Prog) {
var q *obj.Prog
var r *obj.Prog
var i int
loop:
if p == nil {
return
}
a := 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 = 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 = 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 = 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 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
*/
ctxt.Bso.Flush()
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 {
fmt.Fprintf(ctxt.Bso, "save suppressed in: %s\n", cursym.Text.From.Sym.Name)
}
ctxt.Bso.Flush()
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 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 ASYNC,
AWORD:
q = p
p.Mark |= LABEL | SYNC
continue
case AMOVW, AMOVWZ, AMOVD:
q = p
if p.From.Reg >= REG_RESERVED || p.To.Reg >= REG_RESERVED {
p.Mark |= LABEL | SYNC
}
continue
case AFABS,
AFADD,
AFDIV,
AFMADD,
AFMOVD,
AFMOVS,
AFMSUB,
AFMUL,
AFNABS,
AFNEG,
AFNMADD,
AFNMSUB,
ALEDBR,
ALDEBR,
AFSUB:
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,
ACMPBEQ,
ACMPBGE,
ACMPBGT,
ACMPBLE,
ACMPBLT,
ACMPBNE,
ACMPUBEQ,
ACMPUBGE,
ACMPUBGT,
ACMPUBLE,
ACMPUBLT,
ACMPUBNE:
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 p1 *obj.Prog
var p2 *obj.Prog
var pLast *obj.Prog
var pPre *obj.Prog
var pPreempt *obj.Prog
wasSplit := false
for p := cursym.Text; p != nil; p = p.Link {
pLast = p
switch p.As {
case obj.ATEXT:
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 p.From3.Offset&obj.NOSPLIT == 0 {
p, pPreempt = stacksplitPre(ctxt, p, autosize) // emit pre part of split check
pPre = p
wasSplit = true //need post part of split
}
if autosize != 0 {
q = obj.Appendp(ctxt, p)
q.As = AMOVD
q.From.Type = obj.TYPE_ADDR
q.From.Offset = int64(-autosize)
q.From.Reg = REGSP // not actually needed - REGSP is assumed if no reg is provided
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 = true
break
}
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_LR
q.To.Type = obj.TYPE_MEM
q.To.Reg = REGSP
q.To.Offset = 0
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 s390x 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 = obj.Appendp(ctxt, p)
q.As = ABR
q.From = obj.Addr{}
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_LR
q.Mark |= BRANCH
q.Spadj = autosize
break
}
p.As = AMOVD
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGSP
p.From.Offset = 0
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_LR
q = p
if autosize != 0 {
q = obj.Appendp(ctxt, q)
q.As = AADD
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 = obj.Appendp(ctxt, q)
q.As = ABR
q.From = obj.Addr{}
if retTarget == nil {
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_LR
} else {
q.To.Type = obj.TYPE_BRANCH
q.To.Sym = retTarget
}
q.Mark |= BRANCH
q.Spadj = autosize
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)
}
}
}
if wasSplit {
pLast = stacksplitPost(ctxt, pLast, pPre, pPreempt) // emit post part of split check
}
}
func progedit(ctxt *obj.Link, p *obj.Prog) {
p.From.Class = 0
p.To.Class = 0
// Rewrite BR/BL to symbol as TYPE_BRANCH.
switch p.As {
case ABR,
ABL,
obj.ARET,
obj.ADUFFZERO,
obj.ADUFFCOPY:
if p.To.Sym != nil {
p.To.Type = obj.TYPE_BRANCH
}
}
// Rewrite float constants to values stored in memory unless they are +0.
switch p.As {
case AFMOVS:
if p.From.Type == obj.TYPE_FCONST {
f32 := float32(p.From.Val.(float64))
i32 := math.Float32bits(f32)
if i32 == 0 { // +0
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.Sym.Local = true
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
case AFMOVD:
if p.From.Type == obj.TYPE_FCONST {
i64 := math.Float64bits(p.From.Val.(float64))
if i64 == 0 { // +0
break
}
literal := fmt.Sprintf("$f64.%016x", i64)
s := obj.Linklookup(ctxt, literal, 0)
s.Size = 8
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Sym.Local = true
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
// put constants not loadable by LOAD IMMEDIATE into memory
case AMOVD:
if p.From.Type == obj.TYPE_CONST {
val := p.From.Offset
if int64(int32(val)) != val &&
int64(uint32(val)) != val &&
int64(uint64(val)&(0xffffffff<<32)) != val {
literal := fmt.Sprintf("$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.Sym.Local = true
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
}
}
// Rewrite SUB constants into ADD.
switch p.As {
case ASUBC:
if p.From.Type == obj.TYPE_CONST {
p.From.Offset = -p.From.Offset
p.As = AADDC
}
case ASUB:
if p.From.Type == obj.TYPE_CONST {
p.From.Offset = -p.From.Offset
p.As = AADD
}
}
if ctxt.Flag_dynlink {
rewriteToUseGot(ctxt, p)
}
}
func peep(firstp *obj.Prog) {
g := gc.Flowstart(firstp, nil)
if g == nil {
return
}
gactive = 0
var p *obj.Prog
var r *gc.Flow
var t obj.As
loop1:
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
gc.Dumpit("loop1", g.Start, 0)
}
t = 0
for r = g.Start; r != nil; r = r.Link {
p = r.Prog
// TODO(austin) Handle smaller moves. arm and amd64
// distinguish between moves that moves that *must*
// sign/zero extend and moves that don't care so they
// can eliminate moves that don't care without
// breaking moves that do care. This might let us
// simplify or remove the next peep loop, too.
if p.As == ppc64.AMOVD || p.As == ppc64.AFMOVD {
if regtyp(&p.To) {
// Try to eliminate reg->reg moves
if regtyp(&p.From) {
if p.From.Type == p.To.Type {
if copyprop(r) {
excise(r)
t++
} else if subprop(r) && copyprop(r) {
excise(r)
t++
}
}
}
// Convert uses to $0 to uses of R0 and
// propagate R0
if regzer(&p.From) {
if p.To.Type == obj.TYPE_REG {
p.From.Type = obj.TYPE_REG
p.From.Reg = ppc64.REGZERO
if copyprop(r) {
excise(r)
t++
} else if subprop(r) && copyprop(r) {
excise(r)
t++
}
}
}
}
}
}
if t != 0 {
goto loop1
}
/*
* look for MOVB x,R; MOVB R,R (for small MOVs not handled above)
*/
var p1 *obj.Prog
var r1 *gc.Flow
for r := g.Start; r != nil; r = r.Link {
p = r.Prog
switch p.As {
default:
continue
case ppc64.AMOVH,
ppc64.AMOVHZ,
ppc64.AMOVB,
ppc64.AMOVBZ,
ppc64.AMOVW,
ppc64.AMOVWZ:
if p.To.Type != obj.TYPE_REG {
continue
}
}
r1 = r.Link
if r1 == nil {
continue
}
p1 = r1.Prog
if p1.As != p.As {
continue
}
if p1.From.Type != obj.TYPE_REG || p1.From.Reg != p.To.Reg {
continue
}
if p1.To.Type != obj.TYPE_REG || p1.To.Reg != p.To.Reg {
continue
}
excise(r1)
}
if gc.Debug['D'] > 1 {
goto ret /* allow following code improvement to be suppressed */
}
/*
* look for OP x,y,R; CMP R, $0 -> OPCC x,y,R
* when OP can set condition codes correctly
*/
for r := g.Start; r != nil; r = r.Link {
p = r.Prog
switch p.As {
case ppc64.ACMP,
ppc64.ACMPW: /* always safe? */
if !regzer(&p.To) {
continue
}
r1 = r.S1
if r1 == nil {
continue
}
switch r1.Prog.As {
default:
continue
/* the conditions can be complex and these are currently little used */
case ppc64.ABCL,
ppc64.ABC:
continue
case ppc64.ABEQ,
ppc64.ABGE,
ppc64.ABGT,
ppc64.ABLE,
ppc64.ABLT,
ppc64.ABNE,
ppc64.ABVC,
ppc64.ABVS:
break
}
r1 = r
for {
r1 = gc.Uniqp(r1)
if r1 == nil || r1.Prog.As != obj.ANOP {
break
}
}
if r1 == nil {
continue
}
p1 = r1.Prog
if p1.To.Type != obj.TYPE_REG || p1.To.Reg != p.From.Reg {
continue
}
switch p1.As {
/* irregular instructions */
case ppc64.ASUB,
ppc64.AADD,
ppc64.AXOR,
ppc64.AOR:
if p1.From.Type == obj.TYPE_CONST || p1.From.Type == obj.TYPE_ADDR {
continue
}
}
switch p1.As {
default:
continue
case ppc64.AMOVW,
ppc64.AMOVD:
if p1.From.Type != obj.TYPE_REG {
continue
}
continue
case ppc64.AANDCC,
ppc64.AANDNCC,
ppc64.AORCC,
ppc64.AORNCC,
ppc64.AXORCC,
ppc64.ASUBCC,
ppc64.ASUBECC,
ppc64.ASUBMECC,
ppc64.ASUBZECC,
ppc64.AADDCC,
ppc64.AADDCCC,
ppc64.AADDECC,
ppc64.AADDMECC,
ppc64.AADDZECC,
ppc64.ARLWMICC,
ppc64.ARLWNMCC,
/* don't deal with floating point instructions for now */
/*
case AFABS:
case AFADD:
case AFADDS:
case AFCTIW:
case AFCTIWZ:
case AFDIV:
case AFDIVS:
case AFMADD:
case AFMADDS:
case AFMOVD:
case AFMSUB:
case AFMSUBS:
case AFMUL:
case AFMULS:
case AFNABS:
case AFNEG:
case AFNMADD:
case AFNMADDS:
case AFNMSUB:
case AFNMSUBS:
case AFRSP:
case AFSUB:
case AFSUBS:
case ACNTLZW:
case AMTFSB0:
case AMTFSB1:
*/
ppc64.AADD,
ppc64.AADDV,
ppc64.AADDC,
ppc64.AADDCV,
ppc64.AADDME,
ppc64.AADDMEV,
ppc64.AADDE,
ppc64.AADDEV,
ppc64.AADDZE,
ppc64.AADDZEV,
ppc64.AAND,
ppc64.AANDN,
ppc64.ADIVW,
ppc64.ADIVWV,
ppc64.ADIVWU,
ppc64.ADIVWUV,
ppc64.ADIVD,
ppc64.ADIVDV,
ppc64.ADIVDU,
ppc64.ADIVDUV,
ppc64.AEQV,
ppc64.AEXTSB,
ppc64.AEXTSH,
ppc64.AEXTSW,
ppc64.AMULHW,
ppc64.AMULHWU,
ppc64.AMULLW,
ppc64.AMULLWV,
ppc64.AMULHD,
ppc64.AMULHDU,
ppc64.AMULLD,
ppc64.AMULLDV,
ppc64.ANAND,
ppc64.ANEG,
ppc64.ANEGV,
ppc64.ANOR,
ppc64.AOR,
ppc64.AORN,
ppc64.AREM,
ppc64.AREMV,
ppc64.AREMU,
ppc64.AREMUV,
ppc64.AREMD,
ppc64.AREMDV,
ppc64.AREMDU,
ppc64.AREMDUV,
ppc64.ARLWMI,
ppc64.ARLWNM,
ppc64.ASLW,
ppc64.ASRAW,
ppc64.ASRW,
ppc64.ASLD,
ppc64.ASRAD,
ppc64.ASRD,
ppc64.ASUB,
ppc64.ASUBV,
ppc64.ASUBC,
ppc64.ASUBCV,
ppc64.ASUBME,
ppc64.ASUBMEV,
ppc64.ASUBE,
ppc64.ASUBEV,
ppc64.ASUBZE,
ppc64.ASUBZEV,
ppc64.AXOR:
t = variant2as(p1.As, as2variant(p1.As)|V_CC)
}
if gc.Debug['D'] != 0 {
fmt.Printf("cmp %v; %v -> ", p1, p)
}
p1.As = t
if gc.Debug['D'] != 0 {
fmt.Printf("%v\n", p1)
}
excise(r)
continue
}
}
ret:
gc.Flowend(g)
}
/*
// instruction scheduling
if(debug['Q'] == 0)
return;
curtext = nil;
q = nil; // p - 1
q1 = firstp; // top of block
o = 0; // count of instructions
for(p = firstp; p != nil; p = p1) {
p1 = p->link;
o++;
if(p->mark & NOSCHED){
if(q1 != p){
sched(q1, q);
}
for(; p != nil; p = p->link){
if(!(p->mark & NOSCHED))
break;
q = p;
}
p1 = p;
q1 = p;
o = 0;
continue;
}
if(p->mark & (LABEL|SYNC)) {
if(q1 != p)
sched(q1, q);
q1 = p;
o = 1;
}
if(p->mark & (BRANCH|SYNC)) {
sched(q1, p);
q1 = p1;
o = 0;
}
if(o >= NSCHED) {
sched(q1, p);
q1 = p1;
o = 0;
}
q = p;
}
*/
func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32) *obj.Prog {
// MOVD g_stackguard(g), R3
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGG
p.From.Offset = 2 * int64(ctxt.Arch.PtrSize) // G.stackguard0
if ctxt.Cursym.Cfunc {
p.From.Offset = 3 * int64(ctxt.Arch.PtrSize) // G.stackguard1
}
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R3
var q *obj.Prog
if framesize <= obj.StackSmall {
// small stack: SP < stackguard
// CMP stackguard, SP
p = obj.Appendp(ctxt, p)
p.As = ACMPU
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R3
p.To.Type = obj.TYPE_REG
p.To.Reg = REGSP
} else if framesize <= obj.StackBig {
// large stack: SP-framesize < stackguard-StackSmall
// ADD $-framesize, SP, R4
// CMP stackguard, R4
p = obj.Appendp(ctxt, p)
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(-framesize)
p.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
p = obj.Appendp(ctxt, p)
p.As = ACMPU
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R3
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
} else {
// Such a large stack we need to protect against wraparound.
// If SP is close to zero:
// SP-stackguard+StackGuard <= framesize + (StackGuard-StackSmall)
// The +StackGuard on both sides is required to keep the left side positive:
// SP is allowed to be slightly below stackguard. See stack.h.
//
// Preemption sets stackguard to StackPreempt, a very large value.
// That breaks the math above, so we have to check for that explicitly.
// // stackguard is R3
// CMP R3, $StackPreempt
// BEQ label-of-call-to-morestack
// ADD $StackGuard, SP, R4
// SUB R3, R4
// MOVD $(framesize+(StackGuard-StackSmall)), R31
// CMPU R31, R4
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R3
p.To.Type = obj.TYPE_CONST
p.To.Offset = obj.StackPreempt
p = obj.Appendp(ctxt, p)
q = p
p.As = ABEQ
p.To.Type = obj.TYPE_BRANCH
p = obj.Appendp(ctxt, p)
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = obj.StackGuard
p.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
p = obj.Appendp(ctxt, p)
p.As = ASUB
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R3
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(framesize) + obj.StackGuard - obj.StackSmall
p.To.Type = obj.TYPE_REG
p.To.Reg = REGTMP
p = obj.Appendp(ctxt, p)
p.As = ACMPU
p.From.Type = obj.TYPE_REG
p.From.Reg = REGTMP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
}
// q1: BLT done
p = obj.Appendp(ctxt, p)
q1 := p
p.As = ABLT
p.To.Type = obj.TYPE_BRANCH
// MOVD LR, R5
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_LR
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R5
if q != nil {
q.Pcond = p
}
var morestacksym *obj.LSym
if ctxt.Cursym.Cfunc {
morestacksym = obj.Linklookup(ctxt, "runtime.morestackc", 0)
} else if ctxt.Cursym.Text.From3.Offset&obj.NEEDCTXT == 0 {
morestacksym = obj.Linklookup(ctxt, "runtime.morestack_noctxt", 0)
} else {
morestacksym = obj.Linklookup(ctxt, "runtime.morestack", 0)
}
if ctxt.Flag_dynlink {
// Avoid calling morestack via a PLT when dynamically linking. The
// PLT stubs generated by the system linker on ppc64le when "std r2,
// 24(r1)" to save the TOC pointer in their callers stack
// frame. Unfortunately (and necessarily) morestack is called before
// the function that calls it sets up its frame and so the PLT ends
// up smashing the saved TOC pointer for its caller's caller.
//
// According to the ABI documentation there is a mechanism to avoid
// the TOC save that the PLT stub does (put a R_PPC64_TOCSAVE
// relocation on the nop after the call to morestack) but at the time
// of writing it is not supported at all by gold and my attempt to
// use it with ld.bfd caused an internal linker error. So this hack
// seems preferable.
// MOVD $runtime.morestack(SB), R12
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_MEM
p.From.Sym = morestacksym
p.From.Name = obj.NAME_GOTREF
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R12
// MOVD R12, CTR
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R12
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_CTR
// BL CTR
p = obj.Appendp(ctxt, p)
p.As = obj.ACALL
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R12
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_CTR
} else {
// BL runtime.morestack(SB)
p = obj.Appendp(ctxt, p)
p.As = ABL
p.To.Type = obj.TYPE_BRANCH
p.To.Sym = morestacksym
}
// BR start
p = obj.Appendp(ctxt, p)
p.As = ABR
p.To.Type = obj.TYPE_BRANCH
p.Pcond = ctxt.Cursym.Text.Link
// placeholder for q1's jump target
p = obj.Appendp(ctxt, p)
p.As = obj.ANOP // zero-width place holder
q1.Pcond = p
return p
}