func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if cnt < int64(4*gc.Widthptr) {
for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
p = appendpp(p, ppc64.AMOVD, obj.TYPE_REG, ppc64.REGZERO, 0, obj.TYPE_MEM, ppc64.REGSP, gc.Ctxt.FixedFrameSize()+frame+lo+i)
}
} else if cnt <= int64(128*gc.Widthptr) {
p = appendpp(p, ppc64.AADD, obj.TYPE_CONST, 0, gc.Ctxt.FixedFrameSize()+frame+lo-8, obj.TYPE_REG, ppc64.REGRT1, 0)
p.Reg = ppc64.REGSP
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
f := gc.Sysfunc("duffzero")
gc.Naddr(&p.To, f)
gc.Afunclit(&p.To, f)
p.To.Offset = 4 * (128 - cnt/int64(gc.Widthptr))
} else {
p = appendpp(p, ppc64.AMOVD, obj.TYPE_CONST, 0, gc.Ctxt.FixedFrameSize()+frame+lo-8, obj.TYPE_REG, ppc64.REGTMP, 0)
p = appendpp(p, ppc64.AADD, obj.TYPE_REG, ppc64.REGTMP, 0, obj.TYPE_REG, ppc64.REGRT1, 0)
p.Reg = ppc64.REGSP
p = appendpp(p, ppc64.AMOVD, obj.TYPE_CONST, 0, cnt, obj.TYPE_REG, ppc64.REGTMP, 0)
p = appendpp(p, ppc64.AADD, obj.TYPE_REG, ppc64.REGTMP, 0, obj.TYPE_REG, ppc64.REGRT2, 0)
p.Reg = ppc64.REGRT1
p = appendpp(p, ppc64.AMOVDU, obj.TYPE_REG, ppc64.REGZERO, 0, obj.TYPE_MEM, ppc64.REGRT1, int64(gc.Widthptr))
p1 := p
p = appendpp(p, ppc64.ACMP, obj.TYPE_REG, ppc64.REGRT1, 0, obj.TYPE_REG, ppc64.REGRT2, 0)
p = appendpp(p, ppc64.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
gc.Patch(p, p1)
}
return p
}
/*
* insert n into reg slot of p
*/
func raddr(n *gc.Node, p *obj.Prog) {
var a obj.Addr
gc.Naddr(&a, n)
if a.Type != obj.TYPE_REG {
if n != nil {
gc.Fatalf("bad in raddr: %v", gc.Oconv(int(n.Op), 0))
} else {
gc.Fatalf("bad in raddr: <null>")
}
p.Reg = 0
} else {
p.Reg = a.Reg
}
}
// copysub1 replaces v with s in p1->reg if f!=0 or indicates if it could if f==0.
// Returns 1 on failure to substitute (it always succeeds on mips).
func copysub1(p1 *obj.Prog, v *obj.Addr, s *obj.Addr, f int) int {
if f != 0 {
if copyau1(p1, v) {
p1.Reg = s.Reg
}
}
return 0
}
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if cnt < int64(4*gc.Widthptr) {
for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
p = appendpp(p, mips.AMOVV, obj.TYPE_REG, mips.REGZERO, 0, obj.TYPE_MEM, mips.REGSP, 8+frame+lo+i)
}
// TODO(dfc): https://golang.org/issue/12108
// If DUFFZERO is used inside a tail call (see genwrapper) it will
// overwrite the link register.
} else if false && cnt <= int64(128*gc.Widthptr) {
p = appendpp(p, mips.AADDV, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, mips.REGRT1, 0)
p.Reg = mips.REGSP
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
f := gc.Sysfunc("duffzero")
gc.Naddr(&p.To, f)
gc.Afunclit(&p.To, f)
p.To.Offset = 8 * (128 - cnt/int64(gc.Widthptr))
} else {
// ADDV $(8+frame+lo-8), SP, r1
// ADDV $cnt, r1, r2
// loop:
// MOVV R0, (Widthptr)r1
// ADDV $Widthptr, r1
// BNE r1, r2, loop
p = appendpp(p, mips.AADDV, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, mips.REGRT1, 0)
p.Reg = mips.REGSP
p = appendpp(p, mips.AADDV, obj.TYPE_CONST, 0, cnt, obj.TYPE_REG, mips.REGRT2, 0)
p.Reg = mips.REGRT1
p = appendpp(p, mips.AMOVV, obj.TYPE_REG, mips.REGZERO, 0, obj.TYPE_MEM, mips.REGRT1, int64(gc.Widthptr))
p1 := p
p = appendpp(p, mips.AADDV, obj.TYPE_CONST, 0, int64(gc.Widthptr), obj.TYPE_REG, mips.REGRT1, 0)
p = appendpp(p, mips.ABNE, obj.TYPE_REG, mips.REGRT1, 0, obj.TYPE_BRANCH, 0, 0)
p.Reg = mips.REGRT2
gc.Patch(p, p1)
}
return p
}
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64, r0 *uint32) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if *r0 == 0 {
p = appendpp(p, arm.AMOVW, obj.TYPE_CONST, 0, 0, obj.TYPE_REG, arm.REG_R0, 0)
*r0 = 1
}
if cnt < int64(4*gc.Widthptr) {
for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
p = appendpp(p, arm.AMOVW, obj.TYPE_REG, arm.REG_R0, 0, obj.TYPE_MEM, arm.REGSP, int32(4+frame+lo+i))
}
} else if !gc.Nacl && (cnt <= int64(128*gc.Widthptr)) {
p = appendpp(p, arm.AADD, obj.TYPE_CONST, 0, int32(4+frame+lo), obj.TYPE_REG, arm.REG_R1, 0)
p.Reg = arm.REGSP
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
f := gc.Sysfunc("duffzero")
gc.Naddr(&p.To, f)
gc.Afunclit(&p.To, f)
p.To.Offset = 4 * (128 - cnt/int64(gc.Widthptr))
} else {
p = appendpp(p, arm.AADD, obj.TYPE_CONST, 0, int32(4+frame+lo), obj.TYPE_REG, arm.REG_R1, 0)
p.Reg = arm.REGSP
p = appendpp(p, arm.AADD, obj.TYPE_CONST, 0, int32(cnt), obj.TYPE_REG, arm.REG_R2, 0)
p.Reg = arm.REG_R1
p = appendpp(p, arm.AMOVW, obj.TYPE_REG, arm.REG_R0, 0, obj.TYPE_MEM, arm.REG_R1, 4)
p1 := p
p.Scond |= arm.C_PBIT
p = appendpp(p, arm.ACMP, obj.TYPE_REG, arm.REG_R1, 0, obj.TYPE_NONE, 0, 0)
p.Reg = arm.REG_R2
p = appendpp(p, arm.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
gc.Patch(p, p1)
}
return p
}
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if cnt < int64(4*gc.Widthptr) {
for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGSP, 8+frame+lo+i)
}
} else if cnt <= int64(128*gc.Widthptr) && !darwin { // darwin ld64 cannot handle BR26 reloc with non-zero addend
p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGSP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
p = appendpp(p, arm64.AADD, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, arm64.REGRT1, 0)
p.Reg = arm64.REGRT1
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
f := gc.Sysfunc("duffzero")
gc.Naddr(&p.To, f)
gc.Afunclit(&p.To, f)
p.To.Offset = 4 * (128 - cnt/int64(gc.Widthptr))
} else {
p = appendpp(p, arm64.AMOVD, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, arm64.REGTMP, 0)
p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGSP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
p = appendpp(p, arm64.AADD, obj.TYPE_REG, arm64.REGTMP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
p.Reg = arm64.REGRT1
p = appendpp(p, arm64.AMOVD, obj.TYPE_CONST, 0, cnt, obj.TYPE_REG, arm64.REGTMP, 0)
p = appendpp(p, arm64.AADD, obj.TYPE_REG, arm64.REGTMP, 0, obj.TYPE_REG, arm64.REGRT2, 0)
p.Reg = arm64.REGRT1
p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGRT1, int64(gc.Widthptr))
p.Scond = arm64.C_XPRE
p1 := p
p = appendpp(p, arm64.ACMP, obj.TYPE_REG, arm64.REGRT1, 0, obj.TYPE_NONE, 0, 0)
p.Reg = arm64.REGRT2
p = appendpp(p, arm64.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
gc.Patch(p, p1)
}
return p
}
/*
* generate high multiply
* res = (nl * nr) >> wordsize
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Ullman < nr.Ullman {
nl, nr = nr, nl
}
t := nl.Type
w := int(t.Width * 8)
var n1 gc.Node
gc.Regalloc(&n1, t, res)
gc.Cgen(nl, &n1)
var n2 gc.Node
gc.Regalloc(&n2, t, nil)
gc.Cgen(nr, &n2)
switch gc.Simtype[t.Etype] {
case gc.TINT8,
gc.TINT16:
gins(optoas(gc.OMUL, t), &n2, &n1)
gshift(arm.AMOVW, &n1, arm.SHIFT_AR, int32(w), &n1)
case gc.TUINT8,
gc.TUINT16:
gins(optoas(gc.OMUL, t), &n2, &n1)
gshift(arm.AMOVW, &n1, arm.SHIFT_LR, int32(w), &n1)
// perform a long multiplication.
case gc.TINT32,
gc.TUINT32:
var p *obj.Prog
if gc.Issigned[t.Etype] {
p = gins(arm.AMULL, &n2, nil)
} else {
p = gins(arm.AMULLU, &n2, nil)
}
// n2 * n1 -> (n1 n2)
p.Reg = n1.Reg
p.To.Type = obj.TYPE_REGREG
p.To.Reg = n1.Reg
p.To.Offset = int64(n2.Reg)
default:
gc.Fatalf("cgen_hmul %v", t)
}
gc.Cgen(&n1, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
// If s==nil, copyu returns the set/use of v in p; otherwise, it
// modifies p to replace reads of v with reads of s and returns 0 for
// success or non-zero for failure.
//
// If s==nil, copy returns one of the following values:
// 1 if v only used
// 2 if v is set and used in one address (read-alter-rewrite;
// can't substitute)
// 3 if v is only set
// 4 if v is set in one address and used in another (so addresses
// can be rewritten independently)
// 0 otherwise (not touched)
func copyu(p *obj.Prog, v *obj.Addr, s *obj.Addr) int {
if p.From3Type() != obj.TYPE_NONE {
// never generates a from3
fmt.Printf("copyu: from3 (%v) not implemented\n", gc.Ctxt.Dconv(p.From3))
}
switch p.As {
default:
fmt.Printf("copyu: can't find %v\n", obj.Aconv(int(p.As)))
return 2
case obj.ANOP, /* read p->from, write p->to */
mips.AMOVV,
mips.AMOVF,
mips.AMOVD,
mips.AMOVH,
mips.AMOVHU,
mips.AMOVB,
mips.AMOVBU,
mips.AMOVW,
mips.AMOVWU,
mips.AMOVFD,
mips.AMOVDF,
mips.AMOVDW,
mips.AMOVWD,
mips.AMOVFW,
mips.AMOVWF,
mips.AMOVDV,
mips.AMOVVD,
mips.AMOVFV,
mips.AMOVVF,
mips.ATRUNCFV,
mips.ATRUNCDV,
mips.ATRUNCFW,
mips.ATRUNCDW:
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
// Update only indirect uses of v in p->to
if !copyas(&p.To, v) {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
}
return 0
}
if copyas(&p.To, v) {
// Fix up implicit from
if p.From.Type == obj.TYPE_NONE {
p.From = p.To
}
if copyau(&p.From, v) {
return 4
}
return 3
}
if copyau(&p.From, v) {
return 1
}
if copyau(&p.To, v) {
// p->to only indirectly uses v
return 1
}
return 0
case mips.ASGT, /* read p->from, read p->reg, write p->to */
mips.ASGTU,
mips.AADD,
mips.AADDU,
mips.ASUB,
mips.ASUBU,
mips.ASLL,
mips.ASRL,
mips.ASRA,
mips.AOR,
mips.ANOR,
mips.AAND,
mips.AXOR,
mips.AADDV,
mips.AADDVU,
mips.ASUBV,
mips.ASUBVU,
mips.ASLLV,
mips.ASRLV,
mips.ASRAV,
mips.AADDF,
mips.AADDD,
mips.ASUBF,
mips.ASUBD,
mips.AMULF,
mips.AMULD,
mips.ADIVF,
mips.ADIVD:
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
if copysub1(p, v, s, 1) != 0 {
return 1
}
// Update only indirect uses of v in p->to
if !copyas(&p.To, v) {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
}
return 0
}
if copyas(&p.To, v) {
if p.Reg == 0 {
// Fix up implicit reg (e.g., ADD
// R3,R4 -> ADD R3,R4,R4) so we can
// update reg and to separately.
p.Reg = p.To.Reg
}
if copyau(&p.From, v) {
return 4
}
if copyau1(p, v) {
return 4
}
return 3
}
if copyau(&p.From, v) {
return 1
}
if copyau1(p, v) {
return 1
}
if copyau(&p.To, v) {
return 1
}
return 0
case obj.ACHECKNIL, /* read p->from */
mips.ABEQ, /* read p->from, read p->reg */
mips.ABNE,
mips.ABGTZ,
mips.ABGEZ,
mips.ABLTZ,
mips.ABLEZ,
mips.ACMPEQD,
mips.ACMPEQF,
mips.ACMPGED,
mips.ACMPGEF,
mips.ACMPGTD,
mips.ACMPGTF,
mips.ABFPF,
mips.ABFPT,
mips.AMUL,
mips.AMULU,
mips.ADIV,
mips.ADIVU,
mips.AMULV,
mips.AMULVU,
mips.ADIVV,
mips.ADIVVU:
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
return copysub1(p, v, s, 1)
}
if copyau(&p.From, v) {
return 1
}
if copyau1(p, v) {
return 1
}
return 0
case mips.AJMP: /* read p->to */
if s != nil {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
return 0
}
if copyau(&p.To, v) {
return 1
}
return 0
case mips.ARET: /* funny */
if s != nil {
return 0
}
// All registers die at this point, so claim
// everything is set (and not used).
return 3
case mips.AJAL: /* funny */
if v.Type == obj.TYPE_REG {
// TODO(rsc): REG_R0 and REG_F0 used to be
// (when register numbers started at 0) exregoffset and exfregoffset,
// which are unset entirely.
// It's strange that this handles R0 and F0 differently from the other
// registers. Possible failure to optimize?
if mips.REG_R0 < v.Reg && v.Reg <= mips.REG_R31 {
return 2
}
if v.Reg == mips.REGARG {
return 2
}
if mips.REG_F0 < v.Reg && v.Reg <= mips.REG_F31 {
return 2
}
}
if p.From.Type == obj.TYPE_REG && v.Type == obj.TYPE_REG && p.From.Reg == v.Reg {
return 2
}
if s != nil {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
return 0
}
if copyau(&p.To, v) {
return 4
}
return 3
// R0 is zero, used by DUFFZERO, cannot be substituted.
// R1 is ptr to memory, used and set, cannot be substituted.
case obj.ADUFFZERO:
if v.Type == obj.TYPE_REG {
if v.Reg == 0 {
return 1
}
if v.Reg == 1 {
return 2
}
}
return 0
// R1, R2 are ptr to src, dst, used and set, cannot be substituted.
// R3 is scratch, set by DUFFCOPY, cannot be substituted.
case obj.ADUFFCOPY:
if v.Type == obj.TYPE_REG {
if v.Reg == 1 || v.Reg == 2 {
return 2
}
if v.Reg == 3 {
return 3
}
}
return 0
case obj.ATEXT: /* funny */
if v.Type == obj.TYPE_REG {
if v.Reg == mips.REGARG {
return 3
}
}
return 0
case obj.APCDATA,
obj.AFUNCDATA,
obj.AVARDEF,
obj.AVARKILL,
obj.AVARLIVE,
obj.AUSEFIELD:
return 0
}
}
// UNUSED
func peep(firstp *obj.Prog) {
g := (*gc.Graph)(gc.Flowstart(firstp, nil))
if g == nil {
return
}
gactive = 0
var r *gc.Flow
var p *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 {
/*
* elide shift into TYPE_SHIFT operand of subsequent instruction
*/
// if(shiftprop(r)) {
// excise(r);
// t++;
// break;
// }
case arm.ASLL,
arm.ASRL,
arm.ASRA:
break
case arm.AMOVB,
arm.AMOVH,
arm.AMOVW,
arm.AMOVF,
arm.AMOVD:
if regtyp(&p.From) {
if p.From.Type == p.To.Type && isfloatreg(&p.From) == isfloatreg(&p.To) {
if p.Scond == arm.C_SCOND_NONE {
if copyprop(g, r) {
excise(r)
t++
break
}
if subprop(r) && copyprop(g, r) {
excise(r)
t++
break
}
}
}
}
case arm.AMOVHS,
arm.AMOVHU,
arm.AMOVBS,
arm.AMOVBU:
if p.From.Type == obj.TYPE_REG {
if shortprop(r) {
t++
}
}
}
}
/*
if(p->scond == C_SCOND_NONE)
if(regtyp(&p->to))
if(isdconst(&p->from)) {
constprop(&p->from, &p->to, r->s1);
}
break;
*/
if t != 0 {
goto loop1
}
for r := (*gc.Flow)(g.Start); r != nil; r = r.Link {
p = r.Prog
switch p.As {
/*
* EOR -1,x,y => MVN x,y
*/
case arm.AEOR:
if isdconst(&p.From) && p.From.Offset == -1 {
p.As = arm.AMVN
p.From.Type = obj.TYPE_REG
if p.Reg != 0 {
p.From.Reg = p.Reg
} else {
p.From.Reg = p.To.Reg
}
p.Reg = 0
}
}
}
for r := (*gc.Flow)(g.Start); r != nil; r = r.Link {
p = r.Prog
switch p.As {
case arm.AMOVW,
arm.AMOVB,
arm.AMOVBS,
arm.AMOVBU:
if p.From.Type == obj.TYPE_MEM && p.From.Offset == 0 {
xtramodes(g, r, &p.From)
} else if p.To.Type == obj.TYPE_MEM && p.To.Offset == 0 {
xtramodes(g, r, &p.To)
} else {
continue
}
}
}
// case ACMP:
// /*
// * elide CMP $0,x if calculation of x can set condition codes
// */
// if(isdconst(&p->from) || p->from.offset != 0)
// continue;
// r2 = r->s1;
// if(r2 == nil)
// continue;
// t = r2->prog->as;
// switch(t) {
// default:
// continue;
// case ABEQ:
// case ABNE:
// case ABMI:
// case ABPL:
// break;
// case ABGE:
// t = ABPL;
// break;
// case ABLT:
// t = ABMI;
// break;
// case ABHI:
// t = ABNE;
// break;
// case ABLS:
// t = ABEQ;
// break;
// }
// r1 = r;
// do
// r1 = uniqp(r1);
// while (r1 != nil && r1->prog->as == ANOP);
// if(r1 == nil)
// continue;
// p1 = r1->prog;
// if(p1->to.type != TYPE_REG)
// continue;
// if(p1->to.reg != p->reg)
// if(!(p1->as == AMOVW && p1->from.type == TYPE_REG && p1->from.reg == p->reg))
// continue;
//
// switch(p1->as) {
// default:
// continue;
// case AMOVW:
// if(p1->from.type != TYPE_REG)
// continue;
// case AAND:
// case AEOR:
// case AORR:
// case ABIC:
// case AMVN:
// case ASUB:
// case ARSB:
// case AADD:
// case AADC:
// case ASBC:
// case ARSC:
// break;
// }
// p1->scond |= C_SBIT;
// r2->prog->as = t;
// excise(r);
// continue;
// predicate(g);
gc.Flowend(g)
}
/*
* ASLL x,y,w
* .. (not use w, not set x y w)
* AXXX w,a,b (a != w)
* .. (not use w)
* (set w)
* ----------- changed to
* ..
* AXXX (x<<y),a,b
* ..
*/
func shiftprop(r *gc.Flow) bool {
p := (*obj.Prog)(r.Prog)
if p.To.Type != obj.TYPE_REG {
if gc.Debug['P'] != 0 {
fmt.Printf("\tBOTCH: result not reg; FAILURE\n")
}
return false
}
n := int(int(p.To.Reg))
a := obj.Addr(obj.Addr{})
if p.Reg != 0 && p.Reg != p.To.Reg {
a.Type = obj.TYPE_REG
a.Reg = p.Reg
}
if gc.Debug['P'] != 0 {
fmt.Printf("shiftprop\n%v", p)
}
r1 := (*gc.Flow)(r)
var p1 *obj.Prog
for {
/* find first use of shift result; abort if shift operands or result are changed */
r1 = gc.Uniqs(r1)
if r1 == nil {
if gc.Debug['P'] != 0 {
fmt.Printf("\tbranch; FAILURE\n")
}
return false
}
if gc.Uniqp(r1) == nil {
if gc.Debug['P'] != 0 {
fmt.Printf("\tmerge; FAILURE\n")
}
return false
}
p1 = r1.Prog
if gc.Debug['P'] != 0 {
fmt.Printf("\n%v", p1)
}
switch copyu(p1, &p.To, nil) {
case 0: /* not used or set */
if (p.From.Type == obj.TYPE_REG && copyu(p1, &p.From, nil) > 1) || (a.Type == obj.TYPE_REG && copyu(p1, &a, nil) > 1) {
if gc.Debug['P'] != 0 {
fmt.Printf("\targs modified; FAILURE\n")
}
return false
}
continue
case 3: /* set, not used */
{
if gc.Debug['P'] != 0 {
fmt.Printf("\tBOTCH: noref; FAILURE\n")
}
return false
}
}
break
}
/* check whether substitution can be done */
switch p1.As {
default:
if gc.Debug['P'] != 0 {
fmt.Printf("\tnon-dpi; FAILURE\n")
}
return false
case arm.AAND,
arm.AEOR,
arm.AADD,
arm.AADC,
arm.AORR,
arm.ASUB,
arm.ASBC,
arm.ARSB,
arm.ARSC:
if int(p1.Reg) == n || (p1.Reg == 0 && p1.To.Type == obj.TYPE_REG && int(p1.To.Reg) == n) {
if p1.From.Type != obj.TYPE_REG {
if gc.Debug['P'] != 0 {
fmt.Printf("\tcan't swap; FAILURE\n")
}
return false
}
p1.Reg = p1.From.Reg
p1.From.Reg = int16(n)
switch p1.As {
case arm.ASUB:
p1.As = arm.ARSB
case arm.ARSB:
p1.As = arm.ASUB
case arm.ASBC:
p1.As = arm.ARSC
case arm.ARSC:
p1.As = arm.ASBC
}
if gc.Debug['P'] != 0 {
fmt.Printf("\t=>%v", p1)
}
}
fallthrough
case arm.ABIC,
arm.ATST,
arm.ACMP,
arm.ACMN:
if int(p1.Reg) == n {
if gc.Debug['P'] != 0 {
fmt.Printf("\tcan't swap; FAILURE\n")
}
return false
}
if p1.Reg == 0 && int(p1.To.Reg) == n {
if gc.Debug['P'] != 0 {
fmt.Printf("\tshift result used twice; FAILURE\n")
}
return false
}
// case AMVN:
if p1.From.Type == obj.TYPE_SHIFT {
if gc.Debug['P'] != 0 {
fmt.Printf("\tshift result used in shift; FAILURE\n")
}
return false
}
if p1.From.Type != obj.TYPE_REG || int(p1.From.Reg) != n {
if gc.Debug['P'] != 0 {
fmt.Printf("\tBOTCH: where is it used?; FAILURE\n")
}
return false
}
}
/* check whether shift result is used subsequently */
p2 := (*obj.Prog)(p1)
if int(p1.To.Reg) != n {
var p1 *obj.Prog
for {
r1 = gc.Uniqs(r1)
if r1 == nil {
if gc.Debug['P'] != 0 {
fmt.Printf("\tinconclusive; FAILURE\n")
}
return false
}
p1 = r1.Prog
if gc.Debug['P'] != 0 {
fmt.Printf("\n%v", p1)
}
switch copyu(p1, &p.To, nil) {
case 0: /* not used or set */
continue
case 3: /* set, not used */
break
default: /* used */
if gc.Debug['P'] != 0 {
fmt.Printf("\treused; FAILURE\n")
}
return false
}
break
}
}
/* make the substitution */
p2.From.Reg = 0
o := int(int(p.Reg))
if o == 0 {
o = int(p.To.Reg)
}
o &= 15
switch p.From.Type {
case obj.TYPE_CONST:
o |= int((p.From.Offset & 0x1f) << 7)
case obj.TYPE_REG:
o |= 1<<4 | (int(p.From.Reg)&15)<<8
}
switch p.As {
case arm.ASLL:
o |= 0 << 5
case arm.ASRL:
o |= 1 << 5
case arm.ASRA:
o |= 2 << 5
}
p2.From = obj.Addr{}
p2.From.Type = obj.TYPE_SHIFT
p2.From.Offset = int64(o)
if gc.Debug['P'] != 0 {
fmt.Printf("\t=>%v\tSUCCEED\n", p2)
}
return true
}
func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
// TODO(minux): add morestack short-cuts with small fixed frame-size.
ctxt.Cursym = cursym
if cursym.Text == nil || cursym.Text.Link == nil {
return
}
p := cursym.Text
textstksiz := p.To.Offset
if textstksiz == -8 {
// Compatibility hack.
p.From3.Offset |= obj.NOFRAME
textstksiz = 0
}
if textstksiz%8 != 0 {
ctxt.Diag("frame size %d not a multiple of 8", textstksiz)
}
if p.From3.Offset&obj.NOFRAME != 0 {
if textstksiz != 0 {
ctxt.Diag("NOFRAME functions must have a frame size of 0, not %d", textstksiz)
}
}
cursym.Args = p.To.Val.(int32)
cursym.Locals = int32(textstksiz)
/*
* find leaf subroutines
* strip NOPs
* expand RET
* expand BECOME pseudo
*/
if ctxt.Debugvlog != 0 {
fmt.Fprintf(ctxt.Bso, "%5.2f noops\n", obj.Cputime())
}
ctxt.Bso.Flush()
var q *obj.Prog
var q1 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
switch p.As {
/* too hard, just leave alone */
case obj.ATEXT:
q = p
p.Mark |= LABEL | LEAF | SYNC
if p.Link != nil {
p.Link.Mark |= LABEL
}
case ANOR:
q = p
if p.To.Type == obj.TYPE_REG {
if p.To.Reg == REGZERO {
p.Mark |= LABEL | SYNC
}
}
case ALWAR,
ASTWCCC,
AECIWX,
AECOWX,
AEIEIO,
AICBI,
AISYNC,
ATLBIE,
ATLBIEL,
ASLBIA,
ASLBIE,
ASLBMFEE,
ASLBMFEV,
ASLBMTE,
ADCBF,
ADCBI,
ADCBST,
ADCBT,
ADCBTST,
ADCBZ,
ASYNC,
ATLBSYNC,
APTESYNC,
ATW,
AWORD,
ARFI,
ARFCI,
ARFID,
AHRFID:
q = p
p.Mark |= LABEL | SYNC
continue
case AMOVW, AMOVWZ, AMOVD:
q = p
if p.From.Reg >= REG_SPECIAL || p.To.Reg >= REG_SPECIAL {
p.Mark |= LABEL | SYNC
}
continue
case AFABS,
AFABSCC,
AFADD,
AFADDCC,
AFCTIW,
AFCTIWCC,
AFCTIWZ,
AFCTIWZCC,
AFDIV,
AFDIVCC,
AFMADD,
AFMADDCC,
AFMOVD,
AFMOVDU,
/* case AFMOVDS: */
AFMOVS,
AFMOVSU,
/* case AFMOVSD: */
AFMSUB,
AFMSUBCC,
AFMUL,
AFMULCC,
AFNABS,
AFNABSCC,
AFNEG,
AFNEGCC,
AFNMADD,
AFNMADDCC,
AFNMSUB,
AFNMSUBCC,
AFRSP,
AFRSPCC,
AFSUB,
AFSUBCC:
q = p
p.Mark |= FLOAT
continue
case ABL,
ABCL,
obj.ADUFFZERO,
obj.ADUFFCOPY:
cursym.Text.Mark &^= LEAF
fallthrough
case ABC,
ABEQ,
ABGE,
ABGT,
ABLE,
ABLT,
ABNE,
ABR,
ABVC,
ABVS:
p.Mark |= BRANCH
q = p
q1 = p.Pcond
if q1 != nil {
for q1.As == obj.ANOP {
q1 = q1.Link
p.Pcond = q1
}
if q1.Mark&LEAF == 0 {
q1.Mark |= LABEL
}
} else {
p.Mark |= LABEL
}
q1 = p.Link
if q1 != nil {
q1.Mark |= LABEL
}
continue
case AFCMPO, AFCMPU:
q = p
p.Mark |= FCMP | FLOAT
continue
case obj.ARET:
q = p
if p.Link != nil {
p.Link.Mark |= LABEL
}
continue
case obj.ANOP:
q1 = p.Link
q.Link = q1 /* q is non-nop */
q1.Mark |= p.Mark
continue
default:
q = p
continue
}
}
autosize := int32(0)
var aoffset int
var mov int
var o int
var p1 *obj.Prog
var p2 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
o = int(p.As)
switch o {
case obj.ATEXT:
mov = AMOVD
aoffset = 0
autosize = int32(textstksiz)
if p.Mark&LEAF != 0 && autosize == 0 && p.From3.Offset&obj.NOFRAME == 0 {
// A leaf function with no locals has no frame.
p.From3.Offset |= obj.NOFRAME
}
if p.From3.Offset&obj.NOFRAME == 0 {
// If there is a stack frame at all, it includes
// space to save the LR.
autosize += int32(ctxt.FixedFrameSize())
}
p.To.Offset = int64(autosize)
q = p
if ctxt.Flag_shared != 0 && cursym.Name != "runtime.duffzero" && cursym.Name != "runtime.duffcopy" && cursym.Name != "runtime.stackBarrier" {
// When compiling Go into PIC, all functions must start
// with instructions to load the TOC pointer into r2:
//
// addis r2, r12, .TOC.-func@ha
// addi r2, r2, .TOC.-func@l+4
//
// We could probably skip this prologue in some situations
// but it's a bit subtle. However, it is both safe and
// necessary to leave the prologue off duffzero and
// duffcopy as we rely on being able to jump to a specific
// instruction offset for them, and stackBarrier is only
// ever called from an overwritten LR-save slot on the
// stack (when r12 will not be remotely the right thing)
// but fortunately does not access global data.
//
// These are AWORDS because there is no (afaict) way to
// generate the addis instruction except as part of the
// load of a large constant, and in that case there is no
// way to use r12 as the source.
q = obj.Appendp(ctxt, q)
q.As = AWORD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = 0x3c4c0000
q = obj.Appendp(ctxt, q)
q.As = AWORD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = 0x38420000
rel := obj.Addrel(ctxt.Cursym)
rel.Off = 0
rel.Siz = 8
rel.Sym = obj.Linklookup(ctxt, ".TOC.", 0)
rel.Type = obj.R_ADDRPOWER_PCREL
}
if cursym.Text.From3.Offset&obj.NOSPLIT == 0 {
q = stacksplit(ctxt, q, autosize) // emit split check
}
if autosize != 0 {
/* use MOVDU to adjust R1 when saving R31, if autosize is small */
if cursym.Text.Mark&LEAF == 0 && autosize >= -BIG && autosize <= BIG {
mov = AMOVDU
aoffset = int(-autosize)
} else {
q = obj.Appendp(ctxt, q)
q.As = AADD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(-autosize)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Spadj = +autosize
}
} else if cursym.Text.Mark&LEAF == 0 {
// A very few functions that do not return to their caller
// (e.g. gogo) are not identified as leaves but still have
// no frame.
cursym.Text.Mark |= LEAF
}
if cursym.Text.Mark&LEAF != 0 {
cursym.Leaf = 1
break
}
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_LR
q.To.Type = obj.TYPE_REG
q.To.Reg = REGTMP
q = obj.Appendp(ctxt, q)
q.As = int16(mov)
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_REG
q.From.Reg = REGTMP
q.To.Type = obj.TYPE_MEM
q.To.Offset = int64(aoffset)
q.To.Reg = REGSP
if q.As == AMOVDU {
q.Spadj = int32(-aoffset)
}
if ctxt.Flag_shared != 0 {
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R2
q.To.Type = obj.TYPE_MEM
q.To.Reg = REGSP
q.To.Offset = 24
}
if cursym.Text.From3.Offset&obj.WRAPPER != 0 {
// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
//
// MOVD g_panic(g), R3
// CMP R0, R3
// BEQ end
// MOVD panic_argp(R3), R4
// ADD $(autosize+8), R1, R5
// CMP R4, R5
// BNE end
// ADD $8, R1, R6
// MOVD R6, panic_argp(R3)
// end:
// NOP
//
// The NOP is needed to give the jumps somewhere to land.
// It is a liblink NOP, not a ppc64 NOP: it encodes to 0 instruction bytes.
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_MEM
q.From.Reg = REGG
q.From.Offset = 4 * int64(ctxt.Arch.Ptrsize) // G.panic
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R3
q = obj.Appendp(ctxt, q)
q.As = ACMP
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R0
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R3
q = obj.Appendp(ctxt, q)
q.As = ABEQ
q.To.Type = obj.TYPE_BRANCH
p1 = q
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_MEM
q.From.Reg = REG_R3
q.From.Offset = 0 // Panic.argp
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R4
q = obj.Appendp(ctxt, q)
q.As = AADD
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(autosize) + ctxt.FixedFrameSize()
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R5
q = obj.Appendp(ctxt, q)
q.As = ACMP
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R4
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R5
q = obj.Appendp(ctxt, q)
q.As = ABNE
q.To.Type = obj.TYPE_BRANCH
p2 = q
q = obj.Appendp(ctxt, q)
q.As = AADD
q.From.Type = obj.TYPE_CONST
q.From.Offset = ctxt.FixedFrameSize()
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R6
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R6
q.To.Type = obj.TYPE_MEM
q.To.Reg = REG_R3
q.To.Offset = 0 // Panic.argp
q = obj.Appendp(ctxt, q)
q.As = obj.ANOP
p1.Pcond = q
p2.Pcond = q
}
case obj.ARET:
if p.From.Type == obj.TYPE_CONST {
ctxt.Diag("using BECOME (%v) is not supported!", p)
break
}
retTarget := p.To.Sym
if cursym.Text.Mark&LEAF != 0 {
if autosize == 0 {
p.As = ABR
p.From = obj.Addr{}
if retTarget == nil {
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_LR
} else {
p.To.Type = obj.TYPE_BRANCH
p.To.Sym = retTarget
}
p.Mark |= BRANCH
break
}
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(autosize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REGSP
p.Spadj = -autosize
q = ctxt.NewProg()
q.As = ABR
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_LR
q.Mark |= BRANCH
q.Spadj = +autosize
q.Link = p.Link
p.Link = q
break
}
p.As = AMOVD
p.From.Type = obj.TYPE_MEM
p.From.Offset = 0
p.From.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REGTMP
q = ctxt.NewProg()
q.As = AMOVD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_REG
q.From.Reg = REGTMP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_LR
q.Link = p.Link
p.Link = q
p = q
if false {
// Debug bad returns
q = ctxt.NewProg()
q.As = AMOVD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_MEM
q.From.Offset = 0
q.From.Reg = REGTMP
q.To.Type = obj.TYPE_REG
q.To.Reg = REGTMP
q.Link = p.Link
p.Link = q
p = q
}
if autosize != 0 {
q = ctxt.NewProg()
q.As = AADD
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(autosize)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Spadj = -autosize
q.Link = p.Link
p.Link = q
}
q1 = ctxt.NewProg()
q1.As = ABR
q1.Lineno = p.Lineno
if retTarget == nil {
q1.To.Type = obj.TYPE_REG
q1.To.Reg = REG_LR
} else {
q1.To.Type = obj.TYPE_BRANCH
q1.To.Sym = retTarget
}
q1.Mark |= BRANCH
q1.Spadj = +autosize
q1.Link = q.Link
q.Link = q1
case AADD:
if p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP && p.From.Type == obj.TYPE_CONST {
p.Spadj = int32(-p.From.Offset)
}
}
}
}
/*
* return
* 1 if v only used (and substitute),
* 2 if read-alter-rewrite
* 3 if set
* 4 if set and used
* 0 otherwise (not touched)
*/
func copyu(p *obj.Prog, v *obj.Addr, s *obj.Addr) int {
switch p.As {
default:
fmt.Printf("copyu: can't find %v\n", obj.Aconv(int(p.As)))
return 2
case arm.AMOVM:
if v.Type != obj.TYPE_REG {
return 0
}
if p.From.Type == obj.TYPE_CONST { /* read reglist, read/rar */
if s != nil {
if p.From.Offset&(1<<uint(v.Reg)) != 0 {
return 1
}
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
return 0
}
if copyau(&p.To, v) {
if p.Scond&arm.C_WBIT != 0 {
return 2
}
return 1
}
if p.From.Offset&(1<<uint(v.Reg)) != 0 {
return 1 /* read/rar, write reglist */
}
} else {
if s != nil {
if p.To.Offset&(1<<uint(v.Reg)) != 0 {
return 1
}
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
return 0
}
if copyau(&p.From, v) {
if p.Scond&arm.C_WBIT != 0 {
return 2
}
if p.To.Offset&(1<<uint(v.Reg)) != 0 {
return 4
}
return 1
}
if p.To.Offset&(1<<uint(v.Reg)) != 0 {
return 3
}
}
return 0
case obj.ANOP, /* read,, write */
arm.ASQRTD,
arm.AMOVW,
arm.AMOVF,
arm.AMOVD,
arm.AMOVH,
arm.AMOVHS,
arm.AMOVHU,
arm.AMOVB,
arm.AMOVBS,
arm.AMOVBU,
arm.AMOVFW,
arm.AMOVWF,
arm.AMOVDW,
arm.AMOVWD,
arm.AMOVFD,
arm.AMOVDF:
if p.Scond&(arm.C_WBIT|arm.C_PBIT) != 0 {
if v.Type == obj.TYPE_REG {
if p.From.Type == obj.TYPE_MEM || p.From.Type == obj.TYPE_SHIFT {
if p.From.Reg == v.Reg {
return 2
}
} else {
if p.To.Reg == v.Reg {
return 2
}
}
}
}
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
if !copyas(&p.To, v) {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
}
return 0
}
if copyas(&p.To, v) {
if p.Scond != arm.C_SCOND_NONE {
return 2
}
if copyau(&p.From, v) {
return 4
}
return 3
}
if copyau(&p.From, v) {
return 1
}
if copyau(&p.To, v) {
return 1
}
return 0
case arm.AMULLU, /* read, read, write, write */
arm.AMULL,
arm.AMULA,
arm.AMVN:
return 2
case arm.AADD, /* read, read, write */
arm.AADC,
arm.ASUB,
arm.ASBC,
arm.ARSB,
arm.ASLL,
arm.ASRL,
arm.ASRA,
arm.AORR,
arm.AAND,
arm.AEOR,
arm.AMUL,
arm.AMULU,
arm.ADIV,
arm.ADIVU,
arm.AMOD,
arm.AMODU,
arm.AADDF,
arm.AADDD,
arm.ASUBF,
arm.ASUBD,
arm.AMULF,
arm.AMULD,
arm.ADIVF,
arm.ADIVD,
obj.ACHECKNIL,
/* read */
arm.ACMPF, /* read, read, */
arm.ACMPD,
arm.ACMP,
arm.ACMN,
arm.ATST:
/* read,, */
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
if copysub1(p, v, s, 1) != 0 {
return 1
}
if !copyas(&p.To, v) {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
}
return 0
}
if copyas(&p.To, v) {
if p.Scond != arm.C_SCOND_NONE {
return 2
}
if p.Reg == 0 {
p.Reg = p.To.Reg
}
if copyau(&p.From, v) {
return 4
}
if copyau1(p, v) {
return 4
}
return 3
}
if copyau(&p.From, v) {
return 1
}
if copyau1(p, v) {
return 1
}
if copyau(&p.To, v) {
return 1
}
return 0
case arm.ABEQ, /* read, read */
arm.ABNE,
arm.ABCS,
arm.ABHS,
arm.ABCC,
arm.ABLO,
arm.ABMI,
arm.ABPL,
arm.ABVS,
arm.ABVC,
arm.ABHI,
arm.ABLS,
arm.ABGE,
arm.ABLT,
arm.ABGT,
arm.ABLE:
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
return copysub1(p, v, s, 1)
}
if copyau(&p.From, v) {
return 1
}
if copyau1(p, v) {
return 1
}
return 0
case arm.AB: /* funny */
if s != nil {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
return 0
}
if copyau(&p.To, v) {
return 1
}
return 0
case obj.ARET: /* funny */
if s != nil {
return 1
}
return 3
case arm.ABL: /* funny */
if v.Type == obj.TYPE_REG {
// TODO(rsc): REG_R0 and REG_F0 used to be
// (when register numbers started at 0) exregoffset and exfregoffset,
// which are unset entirely.
// It's strange that this handles R0 and F0 differently from the other
// registers. Possible failure to optimize?
if arm.REG_R0 < v.Reg && v.Reg <= arm.REGEXT {
return 2
}
if v.Reg == arm.REGARG {
return 2
}
if arm.REG_F0 < v.Reg && v.Reg <= arm.FREGEXT {
return 2
}
}
if p.From.Type == obj.TYPE_REG && v.Type == obj.TYPE_REG && p.From.Reg == v.Reg {
return 2
}
if s != nil {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
return 0
}
if copyau(&p.To, v) {
return 4
}
return 3
// R0 is zero, used by DUFFZERO, cannot be substituted.
// R1 is ptr to memory, used and set, cannot be substituted.
case obj.ADUFFZERO:
if v.Type == obj.TYPE_REG {
if v.Reg == arm.REG_R0 {
return 1
}
if v.Reg == arm.REG_R0+1 {
return 2
}
}
return 0
// R0 is scratch, set by DUFFCOPY, cannot be substituted.
// R1, R2 areptr to src, dst, used and set, cannot be substituted.
case obj.ADUFFCOPY:
if v.Type == obj.TYPE_REG {
if v.Reg == arm.REG_R0 {
return 3
}
if v.Reg == arm.REG_R0+1 || v.Reg == arm.REG_R0+2 {
return 2
}
}
return 0
case obj.ATEXT: /* funny */
if v.Type == obj.TYPE_REG {
if v.Reg == arm.REGARG {
return 3
}
}
return 0
case obj.APCDATA,
obj.AFUNCDATA,
obj.AVARDEF,
obj.AVARKILL,
obj.AVARLIVE,
obj.AUSEFIELD:
return 0
}
}
// If s==nil, copyu returns the set/use of v in p; otherwise, it
// modifies p to replace reads of v with reads of s and returns 0 for
// success or non-zero for failure.
//
// If s==nil, copy returns one of the following values:
// 1 if v only used
// 2 if v is set and used in one address (read-alter-rewrite;
// can't substitute)
// 3 if v is only set
// 4 if v is set in one address and used in another (so addresses
// can be rewritten independently)
// 0 otherwise (not touched)
func copyu(p *obj.Prog, v *obj.Addr, s *obj.Addr) int {
if p.From3Type() != obj.TYPE_NONE {
// 7g never generates a from3
fmt.Printf("copyu: from3 (%v) not implemented\n", gc.Ctxt.Dconv(p.From3))
}
if p.RegTo2 != obj.REG_NONE {
// 7g never generates a to2
fmt.Printf("copyu: RegTo2 (%v) not implemented\n", obj.Rconv(int(p.RegTo2)))
}
switch p.As {
default:
fmt.Printf("copyu: can't find %v\n", obj.Aconv(int(p.As)))
return 2
case obj.ANOP, /* read p->from, write p->to */
arm64.ANEG,
arm64.AFNEGD,
arm64.AFNEGS,
arm64.AFSQRTD,
arm64.AFCVTZSD,
arm64.AFCVTZSS,
arm64.AFCVTZSDW,
arm64.AFCVTZSSW,
arm64.AFCVTZUD,
arm64.AFCVTZUS,
arm64.AFCVTZUDW,
arm64.AFCVTZUSW,
arm64.AFCVTSD,
arm64.AFCVTDS,
arm64.ASCVTFD,
arm64.ASCVTFS,
arm64.ASCVTFWD,
arm64.ASCVTFWS,
arm64.AUCVTFD,
arm64.AUCVTFS,
arm64.AUCVTFWD,
arm64.AUCVTFWS,
arm64.AMOVB,
arm64.AMOVBU,
arm64.AMOVH,
arm64.AMOVHU,
arm64.AMOVW,
arm64.AMOVWU,
arm64.AMOVD,
arm64.AFMOVS,
arm64.AFMOVD:
if p.Scond == 0 {
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
// Update only indirect uses of v in p->to
if !copyas(&p.To, v) {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
}
return 0
}
if copyas(&p.To, v) {
// Fix up implicit from
if p.From.Type == obj.TYPE_NONE {
p.From = p.To
}
if copyau(&p.From, v) {
return 4
}
return 3
}
if copyau(&p.From, v) {
return 1
}
if copyau(&p.To, v) {
// p->to only indirectly uses v
return 1
}
return 0
}
/* rar p->from, write p->to or read p->from, rar p->to */
if p.From.Type == obj.TYPE_MEM {
if copyas(&p.From, v) {
// No s!=nil check; need to fail
// anyway in that case
return 2
}
if s != nil {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
return 0
}
if copyas(&p.To, v) {
return 3
}
} else if p.To.Type == obj.TYPE_MEM {
if copyas(&p.To, v) {
return 2
}
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
return 0
}
if copyau(&p.From, v) {
return 1
}
} else {
fmt.Printf("copyu: bad %v\n", p)
}
return 0
case arm64.AADD, /* read p->from, read p->reg, write p->to */
arm64.ASUB,
arm64.AAND,
arm64.AORR,
arm64.AEOR,
arm64.AMUL,
arm64.ASMULL,
arm64.AUMULL,
arm64.ASMULH,
arm64.AUMULH,
arm64.ASDIV,
arm64.AUDIV,
arm64.ALSL,
arm64.ALSR,
arm64.AASR,
arm64.AFADDD,
arm64.AFADDS,
arm64.AFSUBD,
arm64.AFSUBS,
arm64.AFMULD,
arm64.AFMULS,
arm64.AFDIVD,
arm64.AFDIVS:
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
if copysub1(p, v, s, 1) != 0 {
return 1
}
// Update only indirect uses of v in p->to
if !copyas(&p.To, v) {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
}
return 0
}
if copyas(&p.To, v) {
if p.Reg == 0 {
// Fix up implicit reg (e.g., ADD
// R3,R4 -> ADD R3,R4,R4) so we can
// update reg and to separately.
p.Reg = p.To.Reg
}
if copyau(&p.From, v) {
return 4
}
if copyau1(p, v) {
return 4
}
return 3
}
if copyau(&p.From, v) {
return 1
}
if copyau1(p, v) {
return 1
}
if copyau(&p.To, v) {
return 1
}
return 0
case arm64.ABEQ,
arm64.ABNE,
arm64.ABGE,
arm64.ABLT,
arm64.ABGT,
arm64.ABLE,
arm64.ABLO,
arm64.ABLS,
arm64.ABHI,
arm64.ABHS:
return 0
case obj.ACHECKNIL, /* read p->from */
arm64.ACMP, /* read p->from, read p->reg */
arm64.AFCMPD,
arm64.AFCMPS:
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
return copysub1(p, v, s, 1)
}
if copyau(&p.From, v) {
return 1
}
if copyau1(p, v) {
return 1
}
return 0
case arm64.AB: /* read p->to */
if s != nil {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
return 0
}
if copyau(&p.To, v) {
return 1
}
return 0
case obj.ARET: /* funny */
if s != nil {
return 0
}
// All registers die at this point, so claim
// everything is set (and not used).
return 3
case arm64.ABL: /* funny */
if p.From.Type == obj.TYPE_REG && v.Type == obj.TYPE_REG && p.From.Reg == v.Reg {
return 2
}
if s != nil {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
return 0
}
if copyau(&p.To, v) {
return 4
}
return 3
// R31 is zero, used by DUFFZERO, cannot be substituted.
// R16 is ptr to memory, used and set, cannot be substituted.
case obj.ADUFFZERO:
if v.Type == obj.TYPE_REG {
if v.Reg == 31 {
return 1
}
if v.Reg == 16 {
return 2
}
}
return 0
// R16, R17 are ptr to src, dst, used and set, cannot be substituted.
// R27 is scratch, set by DUFFCOPY, cannot be substituted.
case obj.ADUFFCOPY:
if v.Type == obj.TYPE_REG {
if v.Reg == 16 || v.Reg == 17 {
return 2
}
if v.Reg == 27 {
return 3
}
}
return 0
case arm64.AHINT,
obj.ATEXT,
obj.APCDATA,
obj.AFUNCDATA,
obj.AVARDEF,
obj.AVARKILL,
obj.AVARLIVE,
obj.AUSEFIELD:
return 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 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 != 0 {
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 != 0 {
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
}
// If s==nil, copyu returns the set/use of v in p; otherwise, it
// modifies p to replace reads of v with reads of s and returns 0 for
// success or non-zero for failure.
//
// If s==nil, copy returns one of the following values:
// 1 if v only used
// 2 if v is set and used in one address (read-alter-rewrite;
// can't substitute)
// 3 if v is only set
// 4 if v is set in one address and used in another (so addresses
// can be rewritten independently)
// 0 otherwise (not touched)
func copyu(p *obj.Prog, v *obj.Addr, s *obj.Addr) int {
if p.From3Type() != obj.TYPE_NONE {
// 9g never generates a from3
fmt.Printf("copyu: from3 (%v) not implemented\n", gc.Ctxt.Dconv(p.From3))
}
switch p.As {
default:
fmt.Printf("copyu: can't find %v\n", obj.Aconv(int(p.As)))
return 2
case obj.ANOP, /* read p->from, write p->to */
ppc64.AMOVH,
ppc64.AMOVHZ,
ppc64.AMOVB,
ppc64.AMOVBZ,
ppc64.AMOVW,
ppc64.AMOVWZ,
ppc64.AMOVD,
ppc64.ANEG,
ppc64.ANEGCC,
ppc64.AADDME,
ppc64.AADDMECC,
ppc64.AADDZE,
ppc64.AADDZECC,
ppc64.ASUBME,
ppc64.ASUBMECC,
ppc64.ASUBZE,
ppc64.ASUBZECC,
ppc64.AFCTIW,
ppc64.AFCTIWZ,
ppc64.AFCTID,
ppc64.AFCTIDZ,
ppc64.AFCFID,
ppc64.AFCFIDCC,
ppc64.AFMOVS,
ppc64.AFMOVD,
ppc64.AFRSP,
ppc64.AFNEG,
ppc64.AFNEGCC:
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
// Update only indirect uses of v in p->to
if !copyas(&p.To, v) {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
}
return 0
}
if copyas(&p.To, v) {
// Fix up implicit from
if p.From.Type == obj.TYPE_NONE {
p.From = p.To
}
if copyau(&p.From, v) {
return 4
}
return 3
}
if copyau(&p.From, v) {
return 1
}
if copyau(&p.To, v) {
// p->to only indirectly uses v
return 1
}
return 0
case ppc64.AMOVBU, /* rar p->from, write p->to or read p->from, rar p->to */
ppc64.AMOVBZU,
ppc64.AMOVHU,
ppc64.AMOVHZU,
ppc64.AMOVWZU,
ppc64.AMOVDU:
if p.From.Type == obj.TYPE_MEM {
if copyas(&p.From, v) {
// No s!=nil check; need to fail
// anyway in that case
return 2
}
if s != nil {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
return 0
}
if copyas(&p.To, v) {
return 3
}
} else if p.To.Type == obj.TYPE_MEM {
if copyas(&p.To, v) {
return 2
}
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
return 0
}
if copyau(&p.From, v) {
return 1
}
} else {
fmt.Printf("copyu: bad %v\n", p)
}
return 0
case ppc64.ARLWMI, /* read p->from, read p->reg, rar p->to */
ppc64.ARLWMICC:
if copyas(&p.To, v) {
return 2
}
fallthrough
/* fall through */
case ppc64.AADD,
/* read p->from, read p->reg, write p->to */
ppc64.AADDC,
ppc64.AADDE,
ppc64.ASUB,
ppc64.ASLW,
ppc64.ASRW,
ppc64.ASRAW,
ppc64.ASLD,
ppc64.ASRD,
ppc64.ASRAD,
ppc64.AOR,
ppc64.AORCC,
ppc64.AORN,
ppc64.AORNCC,
ppc64.AAND,
ppc64.AANDCC,
ppc64.AANDN,
ppc64.AANDNCC,
ppc64.ANAND,
ppc64.ANANDCC,
ppc64.ANOR,
ppc64.ANORCC,
ppc64.AXOR,
ppc64.AMULHW,
ppc64.AMULHWU,
ppc64.AMULLW,
ppc64.AMULLD,
ppc64.ADIVW,
ppc64.ADIVD,
ppc64.ADIVWU,
ppc64.ADIVDU,
ppc64.AREM,
ppc64.AREMU,
ppc64.AREMD,
ppc64.AREMDU,
ppc64.ARLWNM,
ppc64.ARLWNMCC,
ppc64.AFADDS,
ppc64.AFADD,
ppc64.AFSUBS,
ppc64.AFSUB,
ppc64.AFMULS,
ppc64.AFMUL,
ppc64.AFDIVS,
ppc64.AFDIV:
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
if copysub1(p, v, s, 1) != 0 {
return 1
}
// Update only indirect uses of v in p->to
if !copyas(&p.To, v) {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
}
return 0
}
if copyas(&p.To, v) {
if p.Reg == 0 {
// Fix up implicit reg (e.g., ADD
// R3,R4 -> ADD R3,R4,R4) so we can
// update reg and to separately.
p.Reg = p.To.Reg
}
if copyau(&p.From, v) {
return 4
}
if copyau1(p, v) {
return 4
}
return 3
}
if copyau(&p.From, v) {
return 1
}
if copyau1(p, v) {
return 1
}
if copyau(&p.To, v) {
return 1
}
return 0
case ppc64.ABEQ,
ppc64.ABGT,
ppc64.ABGE,
ppc64.ABLT,
ppc64.ABLE,
ppc64.ABNE,
ppc64.ABVC,
ppc64.ABVS:
return 0
case obj.ACHECKNIL, /* read p->from */
ppc64.ACMP, /* read p->from, read p->to */
ppc64.ACMPU,
ppc64.ACMPW,
ppc64.ACMPWU,
ppc64.AFCMPO,
ppc64.AFCMPU:
if s != nil {
if copysub(&p.From, v, s, 1) != 0 {
return 1
}
return copysub(&p.To, v, s, 1)
}
if copyau(&p.From, v) {
return 1
}
if copyau(&p.To, v) {
return 1
}
return 0
// 9g never generates a branch to a GPR (this isn't
// even a normal instruction; liblink turns it in to a
// mov and a branch).
case ppc64.ABR: /* read p->to */
if s != nil {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
return 0
}
if copyau(&p.To, v) {
return 1
}
return 0
case obj.ARET: /* funny */
if s != nil {
return 0
}
// All registers die at this point, so claim
// everything is set (and not used).
return 3
case ppc64.ABL: /* funny */
if v.Type == obj.TYPE_REG {
// TODO(rsc): REG_R0 and REG_F0 used to be
// (when register numbers started at 0) exregoffset and exfregoffset,
// which are unset entirely.
// It's strange that this handles R0 and F0 differently from the other
// registers. Possible failure to optimize?
if ppc64.REG_R0 < v.Reg && v.Reg <= ppc64.REGEXT {
return 2
}
if v.Reg == ppc64.REGARG {
return 2
}
if ppc64.REG_F0 < v.Reg && v.Reg <= ppc64.FREGEXT {
return 2
}
}
if p.From.Type == obj.TYPE_REG && v.Type == obj.TYPE_REG && p.From.Reg == v.Reg {
return 2
}
if s != nil {
if copysub(&p.To, v, s, 1) != 0 {
return 1
}
return 0
}
if copyau(&p.To, v) {
return 4
}
return 3
// R0 is zero, used by DUFFZERO, cannot be substituted.
// R3 is ptr to memory, used and set, cannot be substituted.
case obj.ADUFFZERO:
if v.Type == obj.TYPE_REG {
if v.Reg == 0 {
return 1
}
if v.Reg == 3 {
return 2
}
}
return 0
// R3, R4 are ptr to src, dst, used and set, cannot be substituted.
// R5 is scratch, set by DUFFCOPY, cannot be substituted.
case obj.ADUFFCOPY:
if v.Type == obj.TYPE_REG {
if v.Reg == 3 || v.Reg == 4 {
return 2
}
if v.Reg == 5 {
return 3
}
}
return 0
case obj.ATEXT: /* funny */
if v.Type == obj.TYPE_REG {
if v.Reg == ppc64.REGARG {
return 3
}
}
return 0
case obj.APCDATA,
obj.AFUNCDATA,
obj.AVARDEF,
obj.AVARKILL,
obj.AVARLIVE,
obj.AUSEFIELD:
return 0
}
}
func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
// TODO(minux): add morestack short-cuts with small fixed frame-size.
ctxt.Cursym = cursym
// a switch for enabling/disabling instruction scheduling
nosched := true
if cursym.Text == nil || cursym.Text.Link == nil {
return
}
p := cursym.Text
textstksiz := p.To.Offset
cursym.Args = p.To.Val.(int32)
cursym.Locals = int32(textstksiz)
/*
* find leaf subroutines
* strip NOPs
* expand RET
* expand BECOME pseudo
*/
if ctxt.Debugvlog != 0 {
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
}
/* too hard, just leave alone */
case AMOVW,
AMOVV:
q = p
if p.To.Type == obj.TYPE_REG && p.To.Reg >= REG_SPECIAL {
p.Mark |= LABEL | SYNC
break
}
if p.From.Type == obj.TYPE_REG && p.From.Reg >= REG_SPECIAL {
p.Mark |= LABEL | SYNC
}
/* too hard, just leave alone */
case ASYSCALL,
AWORD,
ATLBWR,
ATLBWI,
ATLBP,
ATLBR:
q = p
p.Mark |= LABEL | SYNC
case ANOR:
q = p
if p.To.Type == obj.TYPE_REG {
if p.To.Reg == REGZERO {
p.Mark |= LABEL | SYNC
}
}
case ABGEZAL,
ABLTZAL,
AJAL,
obj.ADUFFZERO,
obj.ADUFFCOPY:
cursym.Text.Mark &^= LEAF
fallthrough
case AJMP,
ABEQ,
ABGEZ,
ABGTZ,
ABLEZ,
ABLTZ,
ABNE,
ABFPT, ABFPF:
if p.As == ABFPT || p.As == ABFPF {
// We don't treat ABFPT and ABFPF as branches here,
// so that we will always fill nop (0x0) in their
// delay slot during assembly.
// This is to workaround a kernel FPU emulator bug
// where it uses the user stack to simulate the
// instruction in the delay slot if it's not 0x0,
// and somehow that leads to SIGSEGV when the kernel
// jump to the stack.
p.Mark |= SYNC
} else {
p.Mark |= BRANCH
}
q = p
q1 = p.Pcond
if q1 != nil {
for q1.As == obj.ANOP {
q1 = q1.Link
p.Pcond = q1
}
if q1.Mark&LEAF == 0 {
q1.Mark |= LABEL
}
}
//else {
// p.Mark |= LABEL
//}
q1 = p.Link
if q1 != nil {
q1.Mark |= LABEL
}
continue
case ARET:
q = p
if p.Link != nil {
p.Link.Mark |= LABEL
}
continue
case obj.ANOP:
q1 = p.Link
q.Link = q1 /* q is non-nop */
q1.Mark |= p.Mark
continue
default:
q = p
continue
}
}
autosize := int32(0)
var o int
var p1 *obj.Prog
var p2 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
o = int(p.As)
switch o {
case obj.ATEXT:
autosize = int32(textstksiz + 8)
if (p.Mark&LEAF != 0) && autosize <= 8 {
autosize = 0
} else if autosize&4 != 0 {
autosize += 4
}
p.To.Offset = int64(autosize) - 8
if p.From3.Offset&obj.NOSPLIT == 0 {
p = stacksplit(ctxt, p, autosize) // emit split check
}
q = p
if autosize != 0 {
q = obj.Appendp(ctxt, p)
q.As = AADDV
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(-autosize)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Spadj = +autosize
} else if cursym.Text.Mark&LEAF == 0 {
if cursym.Text.From3.Offset&obj.NOSPLIT != 0 {
if ctxt.Debugvlog != 0 {
fmt.Fprintf(ctxt.Bso, "save suppressed in: %s\n", cursym.Name)
ctxt.Bso.Flush()
}
cursym.Text.Mark |= LEAF
}
}
if cursym.Text.Mark&LEAF != 0 {
cursym.Leaf = 1
break
}
q = obj.Appendp(ctxt, q)
q.As = AMOVV
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_REG
q.From.Reg = REGLINK
q.To.Type = obj.TYPE_MEM
q.To.Offset = int64(0)
q.To.Reg = REGSP
if cursym.Text.From3.Offset&obj.WRAPPER != 0 {
// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
//
// MOVV g_panic(g), R1
// BEQ R1, end
// MOVV panic_argp(R1), R2
// ADDV $(autosize+8), R29, R3
// BNE R2, R3, end
// ADDV $8, R29, R2
// MOVV R2, panic_argp(R1)
// end:
// NOP
//
// The NOP is needed to give the jumps somewhere to land.
// It is a liblink NOP, not an mips NOP: it encodes to 0 instruction bytes.
q = obj.Appendp(ctxt, q)
q.As = AMOVV
q.From.Type = obj.TYPE_MEM
q.From.Reg = REGG
q.From.Offset = 4 * int64(ctxt.Arch.Ptrsize) // G.panic
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R1
q = obj.Appendp(ctxt, q)
q.As = ABEQ
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R1
q.To.Type = obj.TYPE_BRANCH
q.Mark |= BRANCH
p1 = q
q = obj.Appendp(ctxt, q)
q.As = AMOVV
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 = AADDV
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(autosize) + 8
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R3
q = obj.Appendp(ctxt, q)
q.As = ABNE
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R2
q.Reg = REG_R3
q.To.Type = obj.TYPE_BRANCH
q.Mark |= BRANCH
p2 = q
q = obj.Appendp(ctxt, q)
q.As = AADDV
q.From.Type = obj.TYPE_CONST
q.From.Offset = 8
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R2
q = obj.Appendp(ctxt, q)
q.As = AMOVV
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R2
q.To.Type = obj.TYPE_MEM
q.To.Reg = REG_R1
q.To.Offset = 0 // Panic.argp
q = obj.Appendp(ctxt, q)
q.As = obj.ANOP
p1.Pcond = q
p2.Pcond = q
}
case ARET:
if p.From.Type == obj.TYPE_CONST {
ctxt.Diag("using BECOME (%v) is not supported!", p)
break
}
if p.To.Sym != nil { // retjmp
p.As = AJMP
p.To.Type = obj.TYPE_BRANCH
break
}
if cursym.Text.Mark&LEAF != 0 {
if autosize == 0 {
p.As = AJMP
p.From = obj.Addr{}
p.To.Type = obj.TYPE_MEM
p.To.Offset = 0
p.To.Reg = REGLINK
p.Mark |= BRANCH
break
}
p.As = AADDV
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(autosize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REGSP
p.Spadj = -autosize
q = ctxt.NewProg()
q.As = AJMP
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_MEM
q.To.Offset = 0
q.To.Reg = REGLINK
q.Mark |= BRANCH
q.Spadj = +autosize
q.Link = p.Link
p.Link = q
break
}
p.As = AMOVV
p.From.Type = obj.TYPE_MEM
p.From.Offset = 0
p.From.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
if false {
// Debug bad returns
q = ctxt.NewProg()
q.As = AMOVV
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_MEM
q.From.Offset = 0
q.From.Reg = REG_R4
q.To.Type = obj.TYPE_REG
q.To.Reg = REGTMP
q.Link = p.Link
p.Link = q
p = q
}
if autosize != 0 {
q = ctxt.NewProg()
q.As = AADDV
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(autosize)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Spadj = -autosize
q.Link = p.Link
p.Link = q
}
q1 = ctxt.NewProg()
q1.As = AJMP
q1.Lineno = p.Lineno
q1.To.Type = obj.TYPE_MEM
q1.To.Offset = 0
q1.To.Reg = REG_R4
q1.Mark |= BRANCH
q1.Spadj = +autosize
q1.Link = q.Link
q.Link = q1
case AADDV,
AADDVU:
if p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP && p.From.Type == obj.TYPE_CONST {
p.Spadj = int32(-p.From.Offset)
}
}
}
if nosched {
// if we don't do instruction scheduling, simply add
// NOP after each branch instruction.
for p = cursym.Text; p != nil; p = p.Link {
if p.Mark&BRANCH != 0 {
addnop(ctxt, p)
}
}
return
}
// instruction scheduling
q = nil // p - 1
q1 = cursym.Text // top of block
o = 0 // count of instructions
for p = cursym.Text; p != nil; p = p1 {
p1 = p.Link
o++
if p.Mark&NOSCHED != 0 {
if q1 != p {
sched(ctxt, q1, q)
}
for ; p != nil; p = p.Link {
if p.Mark&NOSCHED == 0 {
break
}
q = p
}
p1 = p
q1 = p
o = 0
continue
}
if p.Mark&(LABEL|SYNC) != 0 {
if q1 != p {
sched(ctxt, q1, q)
}
q1 = p
o = 1
}
if p.Mark&(BRANCH|SYNC) != 0 {
sched(ctxt, q1, p)
q1 = p1
o = 0
}
if o >= NSCHED {
sched(ctxt, q1, p)
q1 = p1
o = 0
}
q = p
}
}
func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32) *obj.Prog {
// MOVV g_stackguard(g), R1
p = obj.Appendp(ctxt, p)
p.As = AMOVV
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGG
p.From.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0
if ctxt.Cursym.Cfunc != 0 {
p.From.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1
}
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
var q *obj.Prog
if framesize <= obj.StackSmall {
// small stack: SP < stackguard
// AGTU SP, stackguard, R1
p = obj.Appendp(ctxt, p)
p.As = ASGTU
p.From.Type = obj.TYPE_REG
p.From.Reg = REGSP
p.Reg = REG_R1
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
} else if framesize <= obj.StackBig {
// large stack: SP-framesize < stackguard-StackSmall
// ADDV $-framesize, SP, R2
// SGTU R2, stackguard, R1
p = obj.Appendp(ctxt, p)
p.As = AADDV
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 = ASGTU
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R2
p.Reg = REG_R1
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
} 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 R1
// MOVV $StackPreempt, R2
// BEQ R1, R2, label-of-call-to-morestack
// ADDV $StackGuard, SP, R2
// SUBVU R1, R2
// MOVV $(framesize+(StackGuard-StackSmall)), R1
// SGTU R2, R1, R1
p = obj.Appendp(ctxt, p)
p.As = AMOVV
p.From.Type = obj.TYPE_CONST
p.From.Offset = obj.StackPreempt
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
q = p
p.As = ABEQ
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.Reg = REG_R2
p.To.Type = obj.TYPE_BRANCH
p.Mark |= BRANCH
p = obj.Appendp(ctxt, p)
p.As = AADDV
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 = ASUBVU
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 = AMOVV
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_R1
p = obj.Appendp(ctxt, p)
p.As = ASGTU
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R2
p.Reg = REG_R1
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
}
// q1: BNE R1, done
p = obj.Appendp(ctxt, p)
q1 := p
p.As = ABNE
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.To.Type = obj.TYPE_BRANCH
p.Mark |= BRANCH
// MOVV LINK, R3
p = obj.Appendp(ctxt, p)
p.As = AMOVV
p.From.Type = obj.TYPE_REG
p.From.Reg = REGLINK
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R3
if q != nil {
q.Pcond = p
p.Mark |= LABEL
}
// JAL runtime.morestack(SB)
p = obj.Appendp(ctxt, p)
p.As = AJAL
p.To.Type = obj.TYPE_BRANCH
if ctxt.Cursym.Cfunc != 0 {
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)
}
p.Mark |= BRANCH
// JMP start
p = obj.Appendp(ctxt, p)
p.As = AJMP
p.To.Type = obj.TYPE_BRANCH
p.Pcond = ctxt.Cursym.Text.Link
p.Mark |= BRANCH
// placeholder for q1's jump target
p = obj.Appendp(ctxt, p)
p.As = obj.ANOP // zero-width place holder
q1.Pcond = p
return p
}
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 != 0 {
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 != 0:
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 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 != 0 {
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 != 0:
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
}