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, 8+frame+lo+i)
}
} else if cnt <= int64(128*gc.Widthptr) {
p = appendpp(p, ppc64.AADD, obj.TYPE_CONST, 0, 8+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, 8+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
}
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 division.
* caller must set:
* ax = allocated AX register
* dx = allocated DX register
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op int, nl *gc.Node, nr *gc.Node, res *gc.Node, ax *gc.Node, dx *gc.Node) {
// Have to be careful about handling
// most negative int divided by -1 correctly.
// The hardware will trap.
// Also the byte divide instruction needs AH,
// which we otherwise don't have to deal with.
// Easiest way to avoid for int8, int16: use int32.
// For int32 and int64, use explicit test.
// Could use int64 hw for int32.
t := nl.Type
t0 := t
check := 0
if gc.Issigned[t.Etype] {
check = 1
if gc.Isconst(nl, gc.CTINT) && gc.Mpgetfix(nl.Val.U.Xval) != -1<<uint64(t.Width*8-1) {
check = 0
} else if gc.Isconst(nr, gc.CTINT) && gc.Mpgetfix(nr.Val.U.Xval) != -1 {
check = 0
}
}
if t.Width < 4 {
if gc.Issigned[t.Etype] {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
check = 0
}
var t1 gc.Node
gc.Tempname(&t1, t)
var t2 gc.Node
gc.Tempname(&t2, t)
if t0 != t {
var t3 gc.Node
gc.Tempname(&t3, t0)
var t4 gc.Node
gc.Tempname(&t4, t0)
gc.Cgen(nl, &t3)
gc.Cgen(nr, &t4)
// Convert.
gmove(&t3, &t1)
gmove(&t4, &t2)
} else {
gc.Cgen(nl, &t1)
gc.Cgen(nr, &t2)
}
var n1 gc.Node
if !gc.Samereg(ax, res) && !gc.Samereg(dx, res) {
gc.Regalloc(&n1, t, res)
} else {
gc.Regalloc(&n1, t, nil)
}
gmove(&t2, &n1)
gmove(&t1, ax)
var p2 *obj.Prog
var n4 gc.Node
if gc.Nacl {
// Native Client does not relay the divide-by-zero trap
// to the executing program, so we must insert a check
// for ourselves.
gc.Nodconst(&n4, t, 0)
gins(optoas(gc.OCMP, t), &n1, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if panicdiv == nil {
panicdiv = gc.Sysfunc("panicdivide")
}
gc.Ginscall(panicdiv, -1)
gc.Patch(p1, gc.Pc)
}
if check != 0 {
gc.Nodconst(&n4, t, -1)
gins(optoas(gc.OCMP, t), &n1, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if op == gc.ODIV {
// a / (-1) is -a.
gins(optoas(gc.OMINUS, t), nil, ax)
gmove(ax, res)
} else {
// a % (-1) is 0.
gc.Nodconst(&n4, t, 0)
gmove(&n4, res)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
}
if !gc.Issigned[t.Etype] {
var nz gc.Node
gc.Nodconst(&nz, t, 0)
gmove(&nz, dx)
} else {
gins(optoas(gc.OEXTEND, t), nil, nil)
}
gins(optoas(op, t), &n1, nil)
gc.Regfree(&n1)
if op == gc.ODIV {
gmove(ax, res)
} else {
gmove(dx, res)
}
if check != 0 {
gc.Patch(p2, gc.Pc)
}
}
/*
* generate division.
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op int, nl *gc.Node, nr *gc.Node, res *gc.Node) {
// Have to be careful about handling
// most negative int divided by -1 correctly.
// The hardware will trap.
// Also the byte divide instruction needs AH,
// which we otherwise don't have to deal with.
// Easiest way to avoid for int8, int16: use int32.
// For int32 and int64, use explicit test.
// Could use int64 hw for int32.
t := nl.Type
t0 := t
check := 0
if gc.Issigned[t.Etype] {
check = 1
if gc.Isconst(nl, gc.CTINT) && gc.Mpgetfix(nl.Val.U.Xval) != -(1<<uint64(t.Width*8-1)) {
check = 0
} else if gc.Isconst(nr, gc.CTINT) && gc.Mpgetfix(nr.Val.U.Xval) != -1 {
check = 0
}
}
if t.Width < 4 {
if gc.Issigned[t.Etype] {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
check = 0
}
a := optoas(op, t)
var n3 gc.Node
gc.Regalloc(&n3, t0, nil)
var ax gc.Node
var oldax gc.Node
if nl.Ullman >= nr.Ullman {
savex(x86.REG_AX, &ax, &oldax, res, t0)
gc.Cgen(nl, &ax)
gc.Regalloc(&ax, t0, &ax) // mark ax live during cgen
gc.Cgen(nr, &n3)
gc.Regfree(&ax)
} else {
gc.Cgen(nr, &n3)
savex(x86.REG_AX, &ax, &oldax, res, t0)
gc.Cgen(nl, &ax)
}
if t != t0 {
// Convert
ax1 := ax
n31 := n3
ax.Type = t
n3.Type = t
gmove(&ax1, &ax)
gmove(&n31, &n3)
}
var n4 gc.Node
if gc.Nacl {
// Native Client does not relay the divide-by-zero trap
// to the executing program, so we must insert a check
// for ourselves.
gc.Nodconst(&n4, t, 0)
gins(optoas(gc.OCMP, t), &n3, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if panicdiv == nil {
panicdiv = gc.Sysfunc("panicdivide")
}
gc.Ginscall(panicdiv, -1)
gc.Patch(p1, gc.Pc)
}
var p2 *obj.Prog
if check != 0 {
gc.Nodconst(&n4, t, -1)
gins(optoas(gc.OCMP, t), &n3, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if op == gc.ODIV {
// a / (-1) is -a.
gins(optoas(gc.OMINUS, t), nil, &ax)
gmove(&ax, res)
} else {
// a % (-1) is 0.
gc.Nodconst(&n4, t, 0)
gmove(&n4, res)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
}
var olddx gc.Node
var dx gc.Node
savex(x86.REG_DX, &dx, &olddx, res, t)
if !gc.Issigned[t.Etype] {
gc.Nodconst(&n4, t, 0)
gmove(&n4, &dx)
} else {
gins(optoas(gc.OEXTEND, t), nil, nil)
}
gins(a, &n3, nil)
gc.Regfree(&n3)
if op == gc.ODIV {
gmove(&ax, res)
} else {
gmove(&dx, res)
}
restx(&dx, &olddx)
if check != 0 {
gc.Patch(p2, gc.Pc)
}
restx(&ax, &oldax)
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
fmt.Printf("clearfat %v (%v, size: %d)\n", gc.Nconv(nl, 0), gc.Tconv(nl.Type, 0), nl.Type.Width)
}
w := uint64(uint64(nl.Type.Width))
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := uint64(w % 8) // bytes
q := uint64(w / 8) // dwords
if gc.Reginuse(ppc64.REGRT1) {
gc.Fatal("%v in use during clearfat", obj.Rconv(ppc64.REGRT1))
}
var r0 gc.Node
gc.Nodreg(&r0, gc.Types[gc.TUINT64], ppc64.REGZERO)
var dst gc.Node
gc.Nodreg(&dst, gc.Types[gc.Tptr], ppc64.REGRT1)
gc.Regrealloc(&dst)
gc.Agen(nl, &dst)
var boff uint64
if q > 128 {
p := gins(ppc64.ASUB, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p = gins(ppc64.AMOVD, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(q * 8)
p = gins(ppc64.AMOVDU, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 8
pl := (*obj.Prog)(p)
p = gins(ppc64.ACMP, &dst, &end)
gc.Patch(gc.Gbranch(ppc64.ABNE, nil, 0), pl)
gc.Regfree(&end)
// The loop leaves R3 on the last zeroed dword
boff = 8
} else if q >= 4 {
p := gins(ppc64.ASUB, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
f := (*gc.Node)(gc.Sysfunc("duffzero"))
p = gins(obj.ADUFFZERO, nil, f)
gc.Afunclit(&p.To, f)
// 4 and 128 = magic constants: see ../../runtime/asm_ppc64x.s
p.To.Offset = int64(4 * (128 - q))
// duffzero leaves R3 on the last zeroed dword
boff = 8
} else {
var p *obj.Prog
for t := uint64(0); t < q; t++ {
p = gins(ppc64.AMOVD, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(8 * t)
}
boff = 8 * q
}
var p *obj.Prog
for t := uint64(0); t < c; t++ {
p = gins(ppc64.AMOVB, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(t + boff)
}
gc.Regfree(&dst)
}
/*
* generate division.
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op int, nl *gc.Node, nr *gc.Node, res *gc.Node) {
// Have to be careful about handling
// most negative int divided by -1 correctly.
// The hardware will generate undefined result.
// Also need to explicitly trap on division on zero,
// the hardware will silently generate undefined result.
// DIVW will leave unpredicable result in higher 32-bit,
// so always use DIVD/DIVDU.
t := nl.Type
t0 := t
check := 0
if gc.Issigned[t.Etype] {
check = 1
if gc.Isconst(nl, gc.CTINT) && gc.Mpgetfix(nl.Val.U.Xval) != -(1<<uint64(t.Width*8-1)) {
check = 0
} else if gc.Isconst(nr, gc.CTINT) && gc.Mpgetfix(nr.Val.U.Xval) != -1 {
check = 0
}
}
if t.Width < 8 {
if gc.Issigned[t.Etype] {
t = gc.Types[gc.TINT64]
} else {
t = gc.Types[gc.TUINT64]
}
check = 0
}
a := optoas(gc.ODIV, t)
var tl gc.Node
gc.Regalloc(&tl, t0, nil)
var tr gc.Node
gc.Regalloc(&tr, t0, nil)
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &tl)
gc.Cgen(nr, &tr)
} else {
gc.Cgen(nr, &tr)
gc.Cgen(nl, &tl)
}
if t != t0 {
// Convert
tl2 := tl
tr2 := tr
tl.Type = t
tr.Type = t
gmove(&tl2, &tl)
gmove(&tr2, &tr)
}
// Handle divide-by-zero panic.
p1 := gins(optoas(gc.OCMP, t), &tr, nil)
p1.To.Type = obj.TYPE_REG
p1.To.Reg = ppc64.REGZERO
p1 = gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if panicdiv == nil {
panicdiv = gc.Sysfunc("panicdivide")
}
gc.Ginscall(panicdiv, -1)
gc.Patch(p1, gc.Pc)
var p2 *obj.Prog
if check != 0 {
var nm1 gc.Node
gc.Nodconst(&nm1, t, -1)
gins(optoas(gc.OCMP, t), &tr, &nm1)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if op == gc.ODIV {
// a / (-1) is -a.
gins(optoas(gc.OMINUS, t), nil, &tl)
gmove(&tl, res)
} else {
// a % (-1) is 0.
var nz gc.Node
gc.Nodconst(&nz, t, 0)
gmove(&nz, res)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
}
p1 = gins(a, &tr, &tl)
if op == gc.ODIV {
gc.Regfree(&tr)
gmove(&tl, res)
} else {
// A%B = A-(A/B*B)
var tm gc.Node
gc.Regalloc(&tm, t, nil)
// patch div to use the 3 register form
// TODO(minux): add gins3?
p1.Reg = p1.To.Reg
p1.To.Reg = tm.Val.U.Reg
gins(optoas(gc.OMUL, t), &tr, &tm)
gc.Regfree(&tr)
gins(optoas(gc.OSUB, t), &tm, &tl)
gc.Regfree(&tm)
gmove(&tl, res)
}
gc.Regfree(&tl)
if check != 0 {
gc.Patch(p2, gc.Pc)
}
}
func blockcopy(n, res *gc.Node, osrc, odst, w int64) {
// determine alignment.
// want to avoid unaligned access, so have to use
// smaller operations for less aligned types.
// for example moving [4]byte must use 4 MOVB not 1 MOVW.
align := int(n.Type.Align)
var op int
switch align {
default:
gc.Fatal("sgen: invalid alignment %d for %v", align, n.Type)
case 1:
op = arm.AMOVB
case 2:
op = arm.AMOVH
case 4:
op = arm.AMOVW
}
if w%int64(align) != 0 {
gc.Fatal("sgen: unaligned size %d (align=%d) for %v", w, align, n.Type)
}
c := int32(w / int64(align))
if osrc%int64(align) != 0 || odst%int64(align) != 0 {
gc.Fatal("sgen: unaligned offset src %d or dst %d (align %d)", osrc, odst, align)
}
// if we are copying forward on the stack and
// the src and dst overlap, then reverse direction
dir := align
if osrc < odst && int64(odst) < int64(osrc)+w {
dir = -dir
}
if op == arm.AMOVW && !gc.Nacl && dir > 0 && c >= 4 && c <= 128 {
var r0 gc.Node
r0.Op = gc.OREGISTER
r0.Reg = arm.REG_R0
var r1 gc.Node
r1.Op = gc.OREGISTER
r1.Reg = arm.REG_R0 + 1
var r2 gc.Node
r2.Op = gc.OREGISTER
r2.Reg = arm.REG_R0 + 2
var src gc.Node
gc.Regalloc(&src, gc.Types[gc.Tptr], &r1)
var dst gc.Node
gc.Regalloc(&dst, gc.Types[gc.Tptr], &r2)
if n.Ullman >= res.Ullman {
// eval n first
gc.Agen(n, &src)
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agen(res, &dst)
} else {
// eval res first
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agen(res, &dst)
gc.Agen(n, &src)
}
var tmp gc.Node
gc.Regalloc(&tmp, gc.Types[gc.Tptr], &r0)
f := gc.Sysfunc("duffcopy")
p := gins(obj.ADUFFCOPY, nil, f)
gc.Afunclit(&p.To, f)
// 8 and 128 = magic constants: see ../../runtime/asm_arm.s
p.To.Offset = 8 * (128 - int64(c))
gc.Regfree(&tmp)
gc.Regfree(&src)
gc.Regfree(&dst)
return
}
var dst gc.Node
var src gc.Node
if n.Ullman >= res.Ullman {
gc.Agenr(n, &dst, res) // temporarily use dst
gc.Regalloc(&src, gc.Types[gc.Tptr], nil)
gins(arm.AMOVW, &dst, &src)
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agen(res, &dst)
} else {
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agenr(res, &dst, res)
gc.Agenr(n, &src, nil)
}
var tmp gc.Node
gc.Regalloc(&tmp, gc.Types[gc.TUINT32], nil)
// set up end marker
var nend gc.Node
if c >= 4 {
gc.Regalloc(&nend, gc.Types[gc.TUINT32], nil)
p := gins(arm.AMOVW, &src, &nend)
p.From.Type = obj.TYPE_ADDR
if dir < 0 {
p.From.Offset = int64(dir)
} else {
p.From.Offset = w
}
}
// move src and dest to the end of block if necessary
if dir < 0 {
p := gins(arm.AMOVW, &src, &src)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w + int64(dir)
p = gins(arm.AMOVW, &dst, &dst)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w + int64(dir)
}
// move
if c >= 4 {
p := gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
p.Scond |= arm.C_PBIT
ploop := p
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
p.Scond |= arm.C_PBIT
p = gins(arm.ACMP, &src, nil)
raddr(&nend, p)
gc.Patch(gc.Gbranch(arm.ABNE, nil, 0), ploop)
gc.Regfree(&nend)
} else {
var p *obj.Prog
for {
tmp14 := c
c--
if tmp14 <= 0 {
break
}
p = gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
p.Scond |= arm.C_PBIT
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
p.Scond |= arm.C_PBIT
}
}
gc.Regfree(&dst)
gc.Regfree(&src)
gc.Regfree(&tmp)
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
gc.Dump("\nclearfat", nl)
}
w := uint32(nl.Type.Width)
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := w % 4 // bytes
q := w / 4 // quads
var r0 gc.Node
r0.Op = gc.OREGISTER
r0.Reg = arm.REG_R0
var r1 gc.Node
r1.Op = gc.OREGISTER
r1.Reg = arm.REG_R1
var dst gc.Node
gc.Regalloc(&dst, gc.Types[gc.Tptr], &r1)
gc.Agen(nl, &dst)
var nc gc.Node
gc.Nodconst(&nc, gc.Types[gc.TUINT32], 0)
var nz gc.Node
gc.Regalloc(&nz, gc.Types[gc.TUINT32], &r0)
gc.Cgen(&nc, &nz)
if q > 128 {
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p := gins(arm.AMOVW, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(q) * 4
p = gins(arm.AMOVW, &nz, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 4
p.Scond |= arm.C_PBIT
pl := p
p = gins(arm.ACMP, &dst, nil)
raddr(&end, p)
gc.Patch(gc.Gbranch(arm.ABNE, nil, 0), pl)
gc.Regfree(&end)
} else if q >= 4 && !gc.Nacl {
f := gc.Sysfunc("duffzero")
p := gins(obj.ADUFFZERO, nil, f)
gc.Afunclit(&p.To, f)
// 4 and 128 = magic constants: see ../../runtime/asm_arm.s
p.To.Offset = 4 * (128 - int64(q))
} else {
var p *obj.Prog
for q > 0 {
p = gins(arm.AMOVW, &nz, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 4
p.Scond |= arm.C_PBIT
//print("1. %P\n", p);
q--
}
}
var p *obj.Prog
for c > 0 {
p = gins(arm.AMOVB, &nz, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 1
p.Scond |= arm.C_PBIT
//print("2. %P\n", p);
c--
}
gc.Regfree(&dst)
gc.Regfree(&nz)
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
fmt.Printf("clearfat %v (%v, size: %d)\n", nl, nl.Type, nl.Type.Width)
}
w := uint64(uint64(nl.Type.Width))
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := uint64(w % 8) // bytes
q := uint64(w / 8) // dwords
if reg[arm64.REGRT1-arm64.REG_R0] > 0 {
gc.Fatal("R%d in use during clearfat", arm64.REGRT1-arm64.REG_R0)
}
var r0 gc.Node
gc.Nodreg(&r0, gc.Types[gc.TUINT64], arm64.REGZERO)
var dst gc.Node
gc.Nodreg(&dst, gc.Types[gc.Tptr], arm64.REGRT1)
reg[arm64.REGRT1-arm64.REG_R0]++
gc.Agen(nl, &dst)
var boff uint64
if q > 128 {
p := gins(arm64.ASUB, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p = gins(arm64.AMOVD, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(q * 8)
p = gins(arm64.AMOVD, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 8
p.Scond = arm64.C_XPRE
pl := (*obj.Prog)(p)
p = gcmp(arm64.ACMP, &dst, &end)
gc.Patch(gc.Gbranch(arm64.ABNE, nil, 0), pl)
gc.Regfree(&end)
// The loop leaves R16 on the last zeroed dword
boff = 8
} else if q >= 4 && !darwin { // darwin ld64 cannot handle BR26 reloc with non-zero addend
p := gins(arm64.ASUB, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
f := (*gc.Node)(gc.Sysfunc("duffzero"))
p = gins(obj.ADUFFZERO, nil, f)
gc.Afunclit(&p.To, f)
// 4 and 128 = magic constants: see ../../runtime/asm_arm64x.s
p.To.Offset = int64(4 * (128 - q))
// duffzero leaves R16 on the last zeroed dword
boff = 8
} else {
var p *obj.Prog
for t := uint64(0); t < q; t++ {
p = gins(arm64.AMOVD, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(8 * t)
}
boff = 8 * q
}
var p *obj.Prog
for t := uint64(0); t < c; t++ {
p = gins(arm64.AMOVB, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(t + boff)
}
reg[arm64.REGRT1-arm64.REG_R0]--
}
/*
* block copy:
* memmove(&res, &n, w);
* NB: character copy assumed little endian architecture
*/
func sgen(n *gc.Node, res *gc.Node, w int64) {
if gc.Debug['g'] != 0 {
fmt.Printf("\nsgen w=%d\n", w)
gc.Dump("r", n)
gc.Dump("res", res)
}
if n.Ullman >= gc.UINF && res.Ullman >= gc.UINF {
gc.Fatal("sgen UINF")
}
if w < 0 || int64(int32(w)) != w {
gc.Fatal("sgen copy %d", w)
}
if n.Type == nil {
gc.Fatal("sgen: missing type")
}
if w == 0 {
// evaluate side effects only.
var dst gc.Node
regalloc(&dst, gc.Types[gc.Tptr], nil)
agen(res, &dst)
agen(n, &dst)
regfree(&dst)
return
}
// If copying .args, that's all the results, so record definition sites
// for them for the liveness analysis.
if res.Op == gc.ONAME && res.Sym.Name == ".args" {
for l := gc.Curfn.Dcl; l != nil; l = l.Next {
if l.N.Class == gc.PPARAMOUT {
gc.Gvardef(l.N)
}
}
}
// Avoid taking the address for simple enough types.
if componentgen(n, res) {
return
}
// determine alignment.
// want to avoid unaligned access, so have to use
// smaller operations for less aligned types.
// for example moving [4]byte must use 4 MOVB not 1 MOVW.
align := int(n.Type.Align)
var op int
switch align {
default:
gc.Fatal("sgen: invalid alignment %d for %v", align, gc.Tconv(n.Type, 0))
case 1:
op = arm.AMOVB
case 2:
op = arm.AMOVH
case 4:
op = arm.AMOVW
}
if w%int64(align) != 0 {
gc.Fatal("sgen: unaligned size %d (align=%d) for %v", w, align, gc.Tconv(n.Type, 0))
}
c := int32(w / int64(align))
// offset on the stack
osrc := stkof(n)
odst := stkof(res)
if osrc != -1000 && odst != -1000 && (osrc == 1000 || odst == 1000) {
// osrc and odst both on stack, and at least one is in
// an unknown position. Could generate code to test
// for forward/backward copy, but instead just copy
// to a temporary location first.
var tmp gc.Node
gc.Tempname(&tmp, n.Type)
sgen(n, &tmp, w)
sgen(&tmp, res, w)
return
}
if osrc%int32(align) != 0 || odst%int32(align) != 0 {
gc.Fatal("sgen: unaligned offset src %d or dst %d (align %d)", osrc, odst, align)
}
// if we are copying forward on the stack and
// the src and dst overlap, then reverse direction
dir := align
if osrc < odst && int64(odst) < int64(osrc)+w {
dir = -dir
}
if op == arm.AMOVW && !gc.Nacl && dir > 0 && c >= 4 && c <= 128 {
var r0 gc.Node
r0.Op = gc.OREGISTER
r0.Val.U.Reg = REGALLOC_R0
var r1 gc.Node
r1.Op = gc.OREGISTER
r1.Val.U.Reg = REGALLOC_R0 + 1
var r2 gc.Node
r2.Op = gc.OREGISTER
r2.Val.U.Reg = REGALLOC_R0 + 2
var src gc.Node
regalloc(&src, gc.Types[gc.Tptr], &r1)
var dst gc.Node
regalloc(&dst, gc.Types[gc.Tptr], &r2)
if n.Ullman >= res.Ullman {
// eval n first
agen(n, &src)
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
agen(res, &dst)
} else {
// eval res first
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
agen(res, &dst)
agen(n, &src)
}
var tmp gc.Node
regalloc(&tmp, gc.Types[gc.Tptr], &r0)
f := gc.Sysfunc("duffcopy")
p := gins(obj.ADUFFCOPY, nil, f)
gc.Afunclit(&p.To, f)
// 8 and 128 = magic constants: see ../../runtime/asm_arm.s
p.To.Offset = 8 * (128 - int64(c))
regfree(&tmp)
regfree(&src)
regfree(&dst)
return
}
var dst gc.Node
var src gc.Node
if n.Ullman >= res.Ullman {
agenr(n, &dst, res) // temporarily use dst
regalloc(&src, gc.Types[gc.Tptr], nil)
gins(arm.AMOVW, &dst, &src)
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
agen(res, &dst)
} else {
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
agenr(res, &dst, res)
agenr(n, &src, nil)
}
var tmp gc.Node
regalloc(&tmp, gc.Types[gc.TUINT32], nil)
// set up end marker
var nend gc.Node
if c >= 4 {
regalloc(&nend, gc.Types[gc.TUINT32], nil)
p := gins(arm.AMOVW, &src, &nend)
p.From.Type = obj.TYPE_ADDR
if dir < 0 {
p.From.Offset = int64(dir)
} else {
p.From.Offset = w
}
}
// move src and dest to the end of block if necessary
if dir < 0 {
p := gins(arm.AMOVW, &src, &src)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w + int64(dir)
p = gins(arm.AMOVW, &dst, &dst)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w + int64(dir)
}
// move
if c >= 4 {
p := gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
p.Scond |= arm.C_PBIT
ploop := p
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
p.Scond |= arm.C_PBIT
p = gins(arm.ACMP, &src, nil)
raddr(&nend, p)
gc.Patch(gc.Gbranch(arm.ABNE, nil, 0), ploop)
regfree(&nend)
} else {
var p *obj.Prog
for {
tmp14 := c
c--
if tmp14 <= 0 {
break
}
p = gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
p.Scond |= arm.C_PBIT
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
p.Scond |= arm.C_PBIT
}
}
regfree(&dst)
regfree(&src)
regfree(&tmp)
}