/*
* generate high multiply:
* res = (nl*nr) >> width
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
t := nl.Type
a := optoas(gc.OHMUL, t)
if nl.Ullman < nr.Ullman {
tmp := nl
nl = nr
nr = tmp
}
var n1 gc.Node
gc.Cgenr(nl, &n1, res)
var n2 gc.Node
gc.Cgenr(nr, &n2, nil)
var ax gc.Node
gc.Nodreg(&ax, t, x86.REG_AX)
gmove(&n1, &ax)
gins(a, &n2, nil)
gc.Regfree(&n2)
gc.Regfree(&n1)
var dx gc.Node
if t.Width == 1 {
// byte multiply behaves differently.
gc.Nodreg(&ax, t, x86.REG_AH)
gc.Nodreg(&dx, t, x86.REG_DX)
gmove(&ax, &dx)
}
gc.Nodreg(&dx, t, x86.REG_DX)
gmove(&dx, res)
}
/*
* generate
* as $c, reg
*/
func gconreg(as int, c int64, reg int) {
var nr gc.Node
switch as {
case x86.AADDL,
x86.AMOVL,
x86.ALEAL:
gc.Nodreg(&nr, gc.Types[gc.TINT32], reg)
default:
gc.Nodreg(&nr, gc.Types[gc.TINT64], reg)
}
ginscon(as, c, &nr)
}
func ginsnop() {
// This is actually not the x86 NOP anymore,
// but at the point where it gets used, AX is dead
// so it's okay if we lose the high bits.
var reg gc.Node
gc.Nodreg(®, gc.Types[gc.TINT], x86.REG_AX)
gins(x86.AXCHGL, ®, ®)
}
/*
* generate byte multiply:
* res = nl * nr
* there is no 2-operand byte multiply instruction so
* we do a full-width multiplication and truncate afterwards.
*/
func cgen_bmul(op int, nl *gc.Node, nr *gc.Node, res *gc.Node) bool {
if optoas(op, nl.Type) != x86.AIMULB {
return false
}
// largest ullman on left.
if nl.Ullman < nr.Ullman {
tmp := nl
nl = nr
nr = tmp
}
// generate operands in "8-bit" registers.
var n1b gc.Node
gc.Regalloc(&n1b, nl.Type, res)
gc.Cgen(nl, &n1b)
var n2b gc.Node
gc.Regalloc(&n2b, nr.Type, nil)
gc.Cgen(nr, &n2b)
// perform full-width multiplication.
t := gc.Types[gc.TUINT64]
if gc.Issigned[nl.Type.Etype] {
t = gc.Types[gc.TINT64]
}
var n1 gc.Node
gc.Nodreg(&n1, t, int(n1b.Reg))
var n2 gc.Node
gc.Nodreg(&n2, t, int(n2b.Reg))
a := optoas(op, t)
gins(a, &n2, &n1)
// truncate.
gmove(&n1, res)
gc.Regfree(&n1b)
gc.Regfree(&n2b)
return true
}
/*
* register dr is one of the special ones (AX, CX, DI, SI, etc.).
* we need to use it. if it is already allocated as a temporary
* (r > 1; can only happen if a routine like sgen passed a
* special as cgen's res and then cgen used regalloc to reuse
* it as its own temporary), then move it for now to another
* register. caller must call restx to move it back.
* the move is not necessary if dr == res, because res is
* known to be dead.
*/
func savex(dr int, x *gc.Node, oldx *gc.Node, res *gc.Node, t *gc.Type) {
r := int(reg[dr])
// save current ax and dx if they are live
// and not the destination
*oldx = gc.Node{}
gc.Nodreg(x, t, dr)
if r > 1 && !gc.Samereg(x, res) {
gc.Regalloc(oldx, gc.Types[gc.TINT64], nil)
x.Type = gc.Types[gc.TINT64]
gmove(x, oldx)
x.Type = t
oldx.Ostk = int32(r) // squirrel away old r value
reg[dr] = 1
}
}
func stackcopy(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) {
gconreg(movptr, q, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSQ, nil, nil) // MOVQ *(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))
// 14 and 128 = magic constants: see ../../runtime/asm_amd64.s
p.To.Offset = 14 * (128 - q)
} 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)
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op int, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
a := optoas(op, nl.Type)
if nr.Op == gc.OLITERAL {
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gc.Cgen(nl, &n1)
sc := uint64(gc.Mpgetfix(nr.Val.U.Xval))
if sc >= uint64(nl.Type.Width*8) {
// large shift gets 2 shifts by width-1
var n3 gc.Node
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n1)
gins(a, &n3, &n1)
} else {
gins(a, nr, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
if nl.Ullman >= gc.UINF {
var n4 gc.Node
gc.Tempname(&n4, nl.Type)
gc.Cgen(nl, &n4)
nl = &n4
}
if nr.Ullman >= gc.UINF {
var n5 gc.Node
gc.Tempname(&n5, nr.Type)
gc.Cgen(nr, &n5)
nr = &n5
}
rcx := int(reg[x86.REG_CX])
var n1 gc.Node
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
// Allow either uint32 or uint64 as shift type,
// to avoid unnecessary conversion from uint32 to uint64
// just to do the comparison.
tcount := gc.Types[gc.Simtype[nr.Type.Etype]]
if tcount.Etype < gc.TUINT32 {
tcount = gc.Types[gc.TUINT32]
}
gc.Regalloc(&n1, nr.Type, &n1) // to hold the shift type in CX
var n3 gc.Node
gc.Regalloc(&n3, tcount, &n1) // to clear high bits of CX
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TUINT64], x86.REG_CX)
var oldcx gc.Node
if rcx > 0 && !gc.Samereg(&cx, res) {
gc.Regalloc(&oldcx, gc.Types[gc.TUINT64], nil)
gmove(&cx, &oldcx)
}
cx.Type = tcount
var n2 gc.Node
if gc.Samereg(&cx, res) {
gc.Regalloc(&n2, nl.Type, nil)
} else {
gc.Regalloc(&n2, nl.Type, res)
}
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &n2)
gc.Cgen(nr, &n1)
gmove(&n1, &n3)
} else {
gc.Cgen(nr, &n1)
gmove(&n1, &n3)
gc.Cgen(nl, &n2)
}
gc.Regfree(&n3)
// test and fix up large shifts
if !bounded {
gc.Nodconst(&n3, tcount, nl.Type.Width*8)
gins(optoas(gc.OCMP, tcount), &n1, &n3)
p1 := gc.Gbranch(optoas(gc.OLT, tcount), nil, +1)
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n2)
} else {
gc.Nodconst(&n3, nl.Type, 0)
gmove(&n3, &n2)
}
gc.Patch(p1, gc.Pc)
}
gins(a, &n1, &n2)
if oldcx.Op != 0 {
cx.Type = gc.Types[gc.TUINT64]
gmove(&oldcx, &cx)
gc.Regfree(&oldcx)
}
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}