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(nl.Type.Width)
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := w % 8 // bytes
q := w / 8 // dwords
var r0 gc.Node
gc.Nodreg(&r0, gc.Types[gc.TUINT64], arm64.REGZERO)
var dst gc.Node
// REGRT1 is reserved on arm64, see arm64/gsubr.go.
gc.Nodreg(&dst, gc.Types[gc.Tptr], arm64.REGRT1)
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 := 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.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)
}
}
// clearfat clears (i.e. replaces with zeros) the value pointed to by nl.
func clearfat(nl *gc.Node) {
if gc.Debug['g'] != 0 {
fmt.Printf("clearfat %v (%v, size: %d)\n", nl, nl.Type, nl.Type.Width)
}
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
var dst gc.Node
gc.Regalloc(&dst, gc.Types[gc.Tptr], nil)
gc.Agen(nl, &dst)
var boff int64
w := nl.Type.Width
if w > clearLoopCutoff {
// Generate a loop clearing 256 bytes per iteration using XCs.
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p := gins(s390x.AMOVD, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w - (w % 256)
p = gins(s390x.AXC, &dst, &dst)
p.From.Type = obj.TYPE_MEM
p.From.Offset = 0
p.To.Type = obj.TYPE_MEM
p.To.Offset = 0
p.From3 = new(obj.Addr)
p.From3.Offset = 256
p.From3.Type = obj.TYPE_CONST
pl := p
ginscon(s390x.AADD, 256, &dst)
gins(s390x.ACMP, &dst, &end)
gc.Patch(gc.Gbranch(s390x.ABNE, nil, 0), pl)
gc.Regfree(&end)
w = w % 256
}
// Generate instructions to clear the remaining memory.
for w > 0 {
n := w
// Can clear at most 256 bytes per instruction.
if n > 256 {
n = 256
}
switch n {
// Handle very small clears using moves.
case 8, 4, 2, 1:
ins := s390x.AMOVB
switch n {
case 8:
ins = s390x.AMOVD
case 4:
ins = s390x.AMOVW
case 2:
ins = s390x.AMOVH
}
p := gins(ins, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 0
p.To.Type = obj.TYPE_MEM
p.To.Offset = boff
// Handle clears that would require multiple moves with a XC.
default:
p := gins(s390x.AXC, &dst, &dst)
p.From.Type = obj.TYPE_MEM
p.From.Offset = boff
p.To.Type = obj.TYPE_MEM
p.To.Offset = boff
p.From3 = new(obj.Addr)
p.From3.Offset = n
p.From3.Type = obj.TYPE_CONST
}
boff += n
w -= n
}
gc.Regfree(&dst)
}
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
if nl.Type.Align < 4 {
q = 0
c = w
}
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. %v\n", p);
q--
}
}
if c > 4 {
// Loop to zero unaligned memory.
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(c)
p = gins(arm.AMOVB, &nz, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 1
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)
c = 0
}
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. %v\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(nl.Type.Width)
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := w % 8 // bytes
q := w / 8 // dwords
if gc.Reginuse(mips.REGRT1) {
gc.Fatalf("%v in use during clearfat", obj.Rconv(mips.REGRT1))
}
var r0 gc.Node
gc.Nodreg(&r0, gc.Types[gc.TUINT64], mips.REGZERO)
var dst gc.Node
gc.Nodreg(&dst, gc.Types[gc.Tptr], mips.REGRT1)
gc.Regrealloc(&dst)
gc.Agen(nl, &dst)
var boff uint64
if q > 128 {
p := gins(mips.ASUBV, 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(mips.AMOVV, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(q * 8)
p = gins(mips.AMOVV, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 8
pl := p
p = gins(mips.AADDV, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
gc.Patch(ginsbranch(mips.ABNE, nil, &dst, &end, 0), pl)
gc.Regfree(&end)
// The loop leaves R1 on the last zeroed dword
boff = 8
// 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 && q >= 4 {
p := gins(mips.ASUBV, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
f := gc.Sysfunc("duffzero")
p = gins(obj.ADUFFZERO, nil, f)
gc.Afunclit(&p.To, f)
// 8 and 128 = magic constants: see ../../runtime/asm_mips64x.s
p.To.Offset = int64(8 * (128 - q))
// duffzero leaves R1 on the last zeroed dword
boff = 8
} else {
var p *obj.Prog
for t := uint64(0); t < q; t++ {
p = gins(mips.AMOVV, &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(mips.AMOVB, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(t + boff)
}
gc.Regfree(&dst)
}
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
if q < 4 {
// Write sequence of MOV 0, off(base) instead of using STOSL.
// The hope is that although the code will be slightly longer,
// the MOVs will have no dependencies and pipeline better
// than the unrolled STOSL loop.
// NOTE: Must use agen, not igen, so that optimizer sees address
// being taken. We are not writing on field boundaries.
var n1 gc.Node
gc.Regalloc(&n1, gc.Types[gc.Tptr], nil)
gc.Agen(nl, &n1)
n1.Op = gc.OINDREG
var z gc.Node
gc.Nodconst(&z, gc.Types[gc.TUINT64], 0)
for ; q > 0; q-- {
n1.Type = z.Type
gins(x86.AMOVL, &z, &n1)
n1.Xoffset += 4
}
gc.Nodconst(&z, gc.Types[gc.TUINT8], 0)
for ; c > 0; c-- {
n1.Type = z.Type
gins(x86.AMOVB, &z, &n1)
n1.Xoffset++
}
gc.Regfree(&n1)
return
}
var n1 gc.Node
gc.Nodreg(&n1, gc.Types[gc.Tptr], x86.REG_DI)
gc.Agen(nl, &n1)
gconreg(x86.AMOVL, 0, x86.REG_AX)
if q > 128 || (q >= 4 && gc.Nacl) {
gconreg(x86.AMOVL, int64(q), x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.ASTOSL, nil, nil) // STOL AL,*(DI)+
} else if q >= 4 {
p := gins(obj.ADUFFZERO, nil, nil)
p.To.Type = obj.TYPE_ADDR
p.To.Sym = gc.Linksym(gc.Pkglookup("duffzero", gc.Runtimepkg))
// 1 and 128 = magic constants: see ../../runtime/asm_386.s
p.To.Offset = 1 * (128 - int64(q))
} else {
for q > 0 {
gins(x86.ASTOSL, nil, nil) // STOL AL,*(DI)+
q--
}
}
for c > 0 {
gins(x86.ASTOSB, nil, nil) // STOB AL,*(DI)+
c--
}
}
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(nl.Type.Width)
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := w % 8 // bytes
q := w / 8 // dwords
if gc.Reginuse(ppc64.REGRT1) {
gc.Fatalf("%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 := 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.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)
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
gc.Dump("\nclearfat", nl)
}
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
w := nl.Type.Width
if w > 1024 || (w >= 64 && (gc.Nacl || isPlan9)) {
var oldn1 gc.Node
var n1 gc.Node
savex(x86.REG_DI, &n1, &oldn1, nil, gc.Types[gc.Tptr])
gc.Agen(nl, &n1)
var ax gc.Node
var oldax gc.Node
savex(x86.REG_AX, &ax, &oldax, nil, gc.Types[gc.Tptr])
gconreg(x86.AMOVL, 0, x86.REG_AX)
gconreg(movptr, w/8, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.ASTOSQ, nil, nil) // STOQ AL,*(DI)+
if w%8 != 0 {
n1.Op = gc.OINDREG
clearfat_tail(&n1, w%8)
}
restx(&n1, &oldn1)
restx(&ax, &oldax)
return
}
if w >= 64 {
var oldn1 gc.Node
var n1 gc.Node
savex(x86.REG_DI, &n1, &oldn1, nil, gc.Types[gc.Tptr])
gc.Agen(nl, &n1)
var vec_zero gc.Node
var old_x0 gc.Node
savex(x86.REG_X0, &vec_zero, &old_x0, nil, gc.Types[gc.TFLOAT64])
gins(x86.AXORPS, &vec_zero, &vec_zero)
if di := dzDI(w); di != 0 {
gconreg(addptr, di, x86.REG_DI)
}
p := gins(obj.ADUFFZERO, nil, nil)
p.To.Type = obj.TYPE_ADDR
p.To.Sym = gc.Linksym(gc.Pkglookup("duffzero", gc.Runtimepkg))
p.To.Offset = dzOff(w)
if w%16 != 0 {
n1.Op = gc.OINDREG
n1.Xoffset -= 16 - w%16
gins(x86.AMOVUPS, &vec_zero, &n1)
}
restx(&vec_zero, &old_x0)
restx(&n1, &oldn1)
return
}
// NOTE: Must use agen, not igen, so that optimizer sees address
// being taken. We are not writing on field boundaries.
var n1 gc.Node
gc.Agenr(nl, &n1, nil)
n1.Op = gc.OINDREG
clearfat_tail(&n1, w)
gc.Regfree(&n1)
}