func zerorange(p *obj.Prog, frame int64, lo int64, hi int64, ax *uint32) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if *ax == 0 {
p = appendpp(p, x86.AMOVL, obj.TYPE_CONST, 0, 0, obj.TYPE_REG, x86.REG_AX, 0)
*ax = 1
}
if cnt <= int64(4*gc.Widthreg) {
for i := int64(0); i < cnt; i += int64(gc.Widthreg) {
p = appendpp(p, x86.AMOVL, obj.TYPE_REG, x86.REG_AX, 0, obj.TYPE_MEM, x86.REG_SP, frame+lo+i)
}
} else if !gc.Nacl && cnt <= int64(128*gc.Widthreg) {
p = appendpp(p, x86.ALEAL, obj.TYPE_MEM, x86.REG_SP, frame+lo, obj.TYPE_REG, x86.REG_DI, 0)
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_ADDR, 0, 1*(128-cnt/int64(gc.Widthreg)))
p.To.Sym = gc.Linksym(gc.Pkglookup("duffzero", gc.Runtimepkg))
} else {
p = appendpp(p, x86.AMOVL, obj.TYPE_CONST, 0, cnt/int64(gc.Widthreg), obj.TYPE_REG, x86.REG_CX, 0)
p = appendpp(p, x86.ALEAL, obj.TYPE_MEM, x86.REG_SP, frame+lo, obj.TYPE_REG, x86.REG_DI, 0)
p = appendpp(p, x86.AREP, obj.TYPE_NONE, 0, 0, obj.TYPE_NONE, 0, 0)
p = appendpp(p, x86.ASTOSL, obj.TYPE_NONE, 0, 0, obj.TYPE_NONE, 0, 0)
}
return p
}
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 || (gc.Nacl && 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 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)
}
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64, ax *uint32, x0 *uint32) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if cnt%int64(gc.Widthreg) != 0 {
// should only happen with nacl
if cnt%int64(gc.Widthptr) != 0 {
gc.Fatalf("zerorange count not a multiple of widthptr %d", cnt)
}
if *ax == 0 {
p = appendpp(p, x86.AMOVQ, obj.TYPE_CONST, 0, 0, obj.TYPE_REG, x86.REG_AX, 0)
*ax = 1
}
p = appendpp(p, x86.AMOVL, obj.TYPE_REG, x86.REG_AX, 0, obj.TYPE_MEM, x86.REG_SP, frame+lo)
lo += int64(gc.Widthptr)
cnt -= int64(gc.Widthptr)
}
if cnt == 8 {
if *ax == 0 {
p = appendpp(p, x86.AMOVQ, obj.TYPE_CONST, 0, 0, obj.TYPE_REG, x86.REG_AX, 0)
*ax = 1
}
p = appendpp(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_AX, 0, obj.TYPE_MEM, x86.REG_SP, frame+lo)
} else if cnt <= int64(8*gc.Widthreg) {
if *x0 == 0 {
p = appendpp(p, x86.AXORPS, obj.TYPE_REG, x86.REG_X0, 0, obj.TYPE_REG, x86.REG_X0, 0)
*x0 = 1
}
for i := int64(0); i < cnt/16; i++ {
p = appendpp(p, x86.AMOVUPS, obj.TYPE_REG, x86.REG_X0, 0, obj.TYPE_MEM, x86.REG_SP, frame+lo+i*16)
}
if cnt%16 != 0 {
p = appendpp(p, x86.AMOVUPS, obj.TYPE_REG, x86.REG_X0, 0, obj.TYPE_MEM, x86.REG_SP, frame+lo+cnt-int64(16))
}
} else if !gc.Nacl && (cnt <= int64(128*gc.Widthreg)) {
if *x0 == 0 {
p = appendpp(p, x86.AXORPS, obj.TYPE_REG, x86.REG_X0, 0, obj.TYPE_REG, x86.REG_X0, 0)
*x0 = 1
}
p = appendpp(p, leaptr, obj.TYPE_MEM, x86.REG_SP, frame+lo+dzDI(cnt), obj.TYPE_REG, x86.REG_DI, 0)
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_ADDR, 0, dzOff(cnt))
p.To.Sym = gc.Linksym(gc.Pkglookup("duffzero", gc.Runtimepkg))
if cnt%16 != 0 {
p = appendpp(p, x86.AMOVUPS, obj.TYPE_REG, x86.REG_X0, 0, obj.TYPE_MEM, x86.REG_DI, -int64(8))
}
} else {
if *ax == 0 {
p = appendpp(p, x86.AMOVQ, obj.TYPE_CONST, 0, 0, obj.TYPE_REG, x86.REG_AX, 0)
*ax = 1
}
p = appendpp(p, x86.AMOVQ, obj.TYPE_CONST, 0, cnt/int64(gc.Widthreg), obj.TYPE_REG, x86.REG_CX, 0)
p = appendpp(p, leaptr, obj.TYPE_MEM, x86.REG_SP, frame+lo, obj.TYPE_REG, x86.REG_DI, 0)
p = appendpp(p, x86.AREP, obj.TYPE_NONE, 0, 0, obj.TYPE_NONE, 0, 0)
p = appendpp(p, x86.ASTOSQ, obj.TYPE_NONE, 0, 0, obj.TYPE_NONE, 0, 0)
}
return p
}
func blockcopy(n, res *gc.Node, osrc, odst, w int64) {
var dst gc.Node
gc.Nodreg(&dst, gc.Types[gc.Tptr], x86.REG_DI)
var src gc.Node
gc.Nodreg(&src, gc.Types[gc.Tptr], x86.REG_SI)
var tsrc gc.Node
gc.Tempname(&tsrc, gc.Types[gc.Tptr])
var tdst gc.Node
gc.Tempname(&tdst, gc.Types[gc.Tptr])
if !n.Addable {
gc.Agen(n, &tsrc)
}
if !res.Addable {
gc.Agen(res, &tdst)
}
if n.Addable {
gc.Agen(n, &src)
} else {
gmove(&tsrc, &src)
}
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
if res.Addable {
gc.Agen(res, &dst)
} else {
gmove(&tdst, &dst)
}
c := int32(w % 4) // bytes
q := int32(w / 4) // doublewords
// if we are copying forward on the stack and
// the src and dst overlap, then reverse direction
if osrc < odst && int64(odst) < int64(osrc)+w {
// reverse direction
gins(x86.ASTD, nil, nil) // set direction flag
if c > 0 {
gconreg(x86.AADDL, w-1, x86.REG_SI)
gconreg(x86.AADDL, w-1, x86.REG_DI)
gconreg(x86.AMOVL, int64(c), x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSB, nil, nil) // MOVB *(SI)-,*(DI)-
}
if q > 0 {
if c > 0 {
gconreg(x86.AADDL, -3, x86.REG_SI)
gconreg(x86.AADDL, -3, x86.REG_DI)
} else {
gconreg(x86.AADDL, w-4, x86.REG_SI)
gconreg(x86.AADDL, w-4, x86.REG_DI)
}
gconreg(x86.AMOVL, int64(q), x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSL, nil, nil) // MOVL *(SI)-,*(DI)-
}
// we leave with the flag clear
gins(x86.ACLD, nil, nil)
} else {
gins(x86.ACLD, nil, nil) // paranoia. TODO(rsc): remove?
// normal direction
if q > 128 || (q >= 4 && gc.Nacl) {
gconreg(x86.AMOVL, int64(q), x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSL, nil, nil) // MOVL *(SI)+,*(DI)+
} else if q >= 4 {
p := gins(obj.ADUFFCOPY, nil, nil)
p.To.Type = obj.TYPE_ADDR
p.To.Sym = gc.Linksym(gc.Pkglookup("duffcopy", gc.Runtimepkg))
// 10 and 128 = magic constants: see ../../runtime/asm_386.s
p.To.Offset = 10 * (128 - int64(q))
} else if !gc.Nacl && c == 0 {
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TINT32], x86.REG_CX)
// We don't need the MOVSL side-effect of updating SI and DI,
// and issuing a sequence of MOVLs directly is faster.
src.Op = gc.OINDREG
dst.Op = gc.OINDREG
for q > 0 {
gmove(&src, &cx) // MOVL x+(SI),CX
gmove(&cx, &dst) // MOVL CX,x+(DI)
src.Xoffset += 4
dst.Xoffset += 4
q--
}
} else {
for q > 0 {
gins(x86.AMOVSL, nil, nil) // MOVL *(SI)+,*(DI)+
q--
}
}
for c > 0 {
gins(x86.AMOVSB, nil, nil) // MOVB *(SI)+,*(DI)+
c--
}
}
}
func blockcopy(n, ns *gc.Node, osrc, odst, w int64) {
var noddi gc.Node
gc.Nodreg(&noddi, gc.Types[gc.Tptr], x86.REG_DI)
var nodsi gc.Node
gc.Nodreg(&nodsi, gc.Types[gc.Tptr], x86.REG_SI)
var nodl gc.Node
var nodr gc.Node
if n.Ullman >= ns.Ullman {
gc.Agenr(n, &nodr, &nodsi)
if ns.Op == gc.ONAME {
gc.Gvardef(ns)
}
gc.Agenr(ns, &nodl, &noddi)
} else {
if ns.Op == gc.ONAME {
gc.Gvardef(ns)
}
gc.Agenr(ns, &nodl, &noddi)
gc.Agenr(n, &nodr, &nodsi)
}
if nodl.Reg != x86.REG_DI {
gmove(&nodl, &noddi)
}
if nodr.Reg != x86.REG_SI {
gmove(&nodr, &nodsi)
}
gc.Regfree(&nodl)
gc.Regfree(&nodr)
c := w % 8 // bytes
q := w / 8 // quads
var oldcx gc.Node
var cx gc.Node
savex(x86.REG_CX, &cx, &oldcx, nil, gc.Types[gc.TINT64])
// if we are copying forward on the stack and
// the src and dst overlap, then reverse direction
if osrc < odst && odst < osrc+w {
// reverse direction
gins(x86.ASTD, nil, nil) // set direction flag
if c > 0 {
gconreg(addptr, w-1, x86.REG_SI)
gconreg(addptr, w-1, x86.REG_DI)
gconreg(movptr, c, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSB, nil, nil) // MOVB *(SI)-,*(DI)-
}
if q > 0 {
if c > 0 {
gconreg(addptr, -7, x86.REG_SI)
gconreg(addptr, -7, x86.REG_DI)
} else {
gconreg(addptr, w-8, x86.REG_SI)
gconreg(addptr, w-8, x86.REG_DI)
}
gconreg(movptr, q, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)-,*(DI)-
}
// we leave with the flag clear
gins(x86.ACLD, nil, nil)
} else {
// normal direction
if q > 128 || (gc.Nacl && q >= 4) || (obj.Getgoos() == "plan9" && q >= 4) {
gconreg(movptr, q, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)+,*(DI)+
} else if q >= 4 {
var oldx0 gc.Node
var x0 gc.Node
savex(x86.REG_X0, &x0, &oldx0, nil, gc.Types[gc.TFLOAT64])
p := gins(obj.ADUFFCOPY, nil, nil)
p.To.Type = obj.TYPE_ADDR
p.To.Sym = gc.Linksym(gc.Pkglookup("duffcopy", gc.Runtimepkg))
// 64 blocks taking 14 bytes each
// see ../../../../runtime/mkduff.go
p.To.Offset = 14 * (64 - q/2)
restx(&x0, &oldx0)
if q%2 != 0 {
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)+,*(DI)+
}
} else if !gc.Nacl && c == 0 {
// We don't need the MOVSQ side-effect of updating SI and DI,
// and issuing a sequence of MOVQs directly is faster.
nodsi.Op = gc.OINDREG
noddi.Op = gc.OINDREG
for q > 0 {
gmove(&nodsi, &cx) // MOVQ x+(SI),CX
gmove(&cx, &noddi) // MOVQ CX,x+(DI)
nodsi.Xoffset += 8
noddi.Xoffset += 8
q--
}
} else {
for q > 0 {
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)+,*(DI)+
q--
}
}
// copy the remaining c bytes
if w < 4 || c <= 1 || (odst < osrc && osrc < odst+w) {
for c > 0 {
gins(x86.AMOVSB, nil, nil) // MOVB *(SI)+,*(DI)+
c--
}
} else if w < 8 || c <= 4 {
nodsi.Op = gc.OINDREG
noddi.Op = gc.OINDREG
cx.Type = gc.Types[gc.TINT32]
nodsi.Type = gc.Types[gc.TINT32]
noddi.Type = gc.Types[gc.TINT32]
if c > 4 {
nodsi.Xoffset = 0
noddi.Xoffset = 0
gmove(&nodsi, &cx)
gmove(&cx, &noddi)
}
nodsi.Xoffset = c - 4
noddi.Xoffset = c - 4
gmove(&nodsi, &cx)
gmove(&cx, &noddi)
} else {
nodsi.Op = gc.OINDREG
noddi.Op = gc.OINDREG
cx.Type = gc.Types[gc.TINT64]
nodsi.Type = gc.Types[gc.TINT64]
noddi.Type = gc.Types[gc.TINT64]
nodsi.Xoffset = c - 8
noddi.Xoffset = c - 8
gmove(&nodsi, &cx)
gmove(&cx, &noddi)
}
}
restx(&cx, &oldcx)
}
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--
}
}