/*
* 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 division.
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op gc.Op, nl *gc.Node, nr *gc.Node, res *gc.Node) {
t := nl.Type
t0 := t
if t.Width < 8 {
if t.IsSigned() {
t = gc.Types[gc.TINT64]
} else {
t = gc.Types[gc.TUINT64]
}
}
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 := ginsbranch(mips.ABNE, nil, &tr, nil, 0)
if panicdiv == nil {
panicdiv = gc.Sysfunc("panicdivide")
}
gc.Ginscall(panicdiv, -1)
gc.Patch(p1, gc.Pc)
gins3(a, &tr, &tl, nil)
gc.Regfree(&tr)
if op == gc.ODIV {
var lo gc.Node
gc.Nodreg(&lo, gc.Types[gc.TUINT64], mips.REG_LO)
gins(mips.AMOVV, &lo, &tl)
} else { // remainder in REG_HI
var hi gc.Node
gc.Nodreg(&hi, gc.Types[gc.TUINT64], mips.REG_HI)
gins(mips.AMOVV, &hi, &tl)
}
gmove(&tl, res)
gc.Regfree(&tl)
}
/*
* 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 {
nl, nr = nr, nl
}
var n1 gc.Node
gc.Cgenr(nl, &n1, res)
var n2 gc.Node
gc.Cgenr(nr, &n2, nil)
var ax, oldax, dx, olddx gc.Node
savex(x86.REG_AX, &ax, &oldax, res, gc.Types[gc.TUINT64])
savex(x86.REG_DX, &dx, &olddx, res, gc.Types[gc.TUINT64])
gmove(&n1, &ax)
gins(a, &n2, nil)
gc.Regfree(&n2)
gc.Regfree(&n1)
if t.Width == 1 {
// byte multiply behaves differently.
var byteAH, byteDX gc.Node
gc.Nodreg(&byteAH, t, x86.REG_AH)
gc.Nodreg(&byteDX, t, x86.REG_DX)
gmove(&byteAH, &byteDX)
}
gmove(&dx, res)
restx(&ax, &oldax)
restx(&dx, &olddx)
}
/*
* generate high multiply:
* res = (nl*nr) >> width
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
// largest ullman on left.
if nl.Ullman < nr.Ullman {
nl, nr = nr, nl
}
t := nl.Type
w := t.Width * 8
var n1 gc.Node
gc.Cgenr(nl, &n1, res)
var n2 gc.Node
gc.Cgenr(nr, &n2, nil)
switch gc.Simtype[t.Etype] {
case gc.TINT8,
gc.TINT16,
gc.TINT32:
gins3(optoas(gc.OMUL, t), &n2, &n1, nil)
var lo gc.Node
gc.Nodreg(&lo, gc.Types[gc.TUINT64], mips.REG_LO)
gins(mips.AMOVV, &lo, &n1)
p := gins(mips.ASRAV, nil, &n1)
p.From.Type = obj.TYPE_CONST
p.From.Offset = w
case gc.TUINT8,
gc.TUINT16,
gc.TUINT32:
gins3(optoas(gc.OMUL, t), &n2, &n1, nil)
var lo gc.Node
gc.Nodreg(&lo, gc.Types[gc.TUINT64], mips.REG_LO)
gins(mips.AMOVV, &lo, &n1)
p := gins(mips.ASRLV, nil, &n1)
p.From.Type = obj.TYPE_CONST
p.From.Offset = w
case gc.TINT64,
gc.TUINT64:
if t.IsSigned() {
gins3(mips.AMULV, &n2, &n1, nil)
} else {
gins3(mips.AMULVU, &n2, &n1, nil)
}
var hi gc.Node
gc.Nodreg(&hi, gc.Types[gc.TUINT64], mips.REG_HI)
gins(mips.AMOVV, &hi, &n1)
default:
gc.Fatalf("cgen_hmul %v", t)
}
gc.Cgen(&n1, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
/*
* 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
* as $c, reg
*/
func gconreg(as int, c int64, reg int) {
var n1 gc.Node
var n2 gc.Node
gc.Nodconst(&n1, gc.Types[gc.TINT64], c)
gc.Nodreg(&n2, gc.Types[gc.TINT64], reg)
gins(as, &n1, &n2)
}
/*
* 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
}
// floating-point. 387 (not SSE2)
func cgen_float387(n *gc.Node, res *gc.Node) {
var f0 gc.Node
var f1 gc.Node
nl := n.Left
nr := n.Right
gc.Nodreg(&f0, nl.Type, x86.REG_F0)
gc.Nodreg(&f1, n.Type, x86.REG_F0+1)
if nr != nil {
// binary
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &f0)
if nr.Addable {
gins(foptoas(n.Op, n.Type, 0), nr, &f0)
} else {
gc.Cgen(nr, &f0)
gins(foptoas(n.Op, n.Type, Fpop), &f0, &f1)
}
} else {
gc.Cgen(nr, &f0)
if nl.Addable {
gins(foptoas(n.Op, n.Type, Frev), nl, &f0)
} else {
gc.Cgen(nl, &f0)
gins(foptoas(n.Op, n.Type, Frev|Fpop), &f0, &f1)
}
}
gmove(&f0, res)
return
}
// unary
gc.Cgen(nl, &f0)
if n.Op != gc.OCONV && n.Op != gc.OPLUS {
gins(foptoas(n.Op, n.Type, 0), nil, nil)
}
gmove(&f0, res)
return
}
/*
* generate high multiply:
* res = (nl*nr) >> width
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
var n1 gc.Node
var n2 gc.Node
var ax gc.Node
var dx gc.Node
t := nl.Type
a := optoas(gc.OHMUL, t)
// gen nl in n1.
gc.Tempname(&n1, t)
gc.Cgen(nl, &n1)
// gen nr in n2.
gc.Regalloc(&n2, t, res)
gc.Cgen(nr, &n2)
// multiply.
gc.Nodreg(&ax, t, x86.REG_AX)
gmove(&n2, &ax)
gins(a, &n1, nil)
gc.Regfree(&n2)
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 high multiply:
* res = (nl*nr) >> width
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
var n1 gc.Node
var n2 gc.Node
t := nl.Type
a := optoas(gc.OHMUL, t)
// gen nl in n1.
gc.Tempname(&n1, t)
gc.Cgen(nl, &n1)
// gen nr in n2.
gc.Regalloc(&n2, t, res)
gc.Cgen(nr, &n2)
var ax, oldax, dx, olddx gc.Node
savex(x86.REG_AX, &ax, &oldax, res, gc.Types[gc.TUINT32])
savex(x86.REG_DX, &dx, &olddx, res, gc.Types[gc.TUINT32])
gmove(&n2, &ax)
gins(a, &n1, nil)
gc.Regfree(&n2)
if t.Width == 1 {
// byte multiply behaves differently.
var byteAH, byteDX gc.Node
gc.Nodreg(&byteAH, t, x86.REG_AH)
gc.Nodreg(&byteDX, t, x86.REG_DX)
gmove(&byteAH, &byteDX)
}
gmove(&dx, res)
restx(&ax, &oldax)
restx(&dx, &olddx)
}
func savex(dr int, x *gc.Node, oldx *gc.Node, res *gc.Node, t *gc.Type) {
r := gc.GetReg(dr)
gc.Nodreg(x, gc.Types[gc.TINT32], dr)
// save current ax and dx if they are live
// and not the destination
*oldx = gc.Node{}
if r > 0 && !gc.Samereg(x, res) {
gc.Tempname(oldx, gc.Types[gc.TINT32])
gmove(x, oldx)
}
gc.Regalloc(x, t, x)
}
/*
* 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 := uint8(gc.GetReg(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.Etype = r // squirrel away old r value
gc.SetReg(dr, 1)
}
}
/*
* attempt to generate 64-bit
* res = n
* return 1 on success, 0 if op not handled.
*/
func cgen64(n *gc.Node, res *gc.Node) {
if res.Op != gc.OINDREG && res.Op != gc.ONAME {
gc.Dump("n", n)
gc.Dump("res", res)
gc.Fatalf("cgen64 %v of %v", gc.Oconv(int(n.Op), 0), gc.Oconv(int(res.Op), 0))
}
switch n.Op {
default:
gc.Fatalf("cgen64 %v", gc.Oconv(int(n.Op), 0))
case gc.OMINUS:
gc.Cgen(n.Left, res)
var hi1 gc.Node
var lo1 gc.Node
split64(res, &lo1, &hi1)
gins(x86.ANEGL, nil, &lo1)
gins(x86.AADCL, ncon(0), &hi1)
gins(x86.ANEGL, nil, &hi1)
splitclean()
return
case gc.OCOM:
gc.Cgen(n.Left, res)
var lo1 gc.Node
var hi1 gc.Node
split64(res, &lo1, &hi1)
gins(x86.ANOTL, nil, &lo1)
gins(x86.ANOTL, nil, &hi1)
splitclean()
return
// binary operators.
// common setup below.
case gc.OADD,
gc.OSUB,
gc.OMUL,
gc.OLROT,
gc.OLSH,
gc.ORSH,
gc.OAND,
gc.OOR,
gc.OXOR:
break
}
l := n.Left
r := n.Right
if !l.Addable {
var t1 gc.Node
gc.Tempname(&t1, l.Type)
gc.Cgen(l, &t1)
l = &t1
}
if r != nil && !r.Addable {
var t2 gc.Node
gc.Tempname(&t2, r.Type)
gc.Cgen(r, &t2)
r = &t2
}
var ax gc.Node
gc.Nodreg(&ax, gc.Types[gc.TINT32], x86.REG_AX)
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TINT32], x86.REG_CX)
var dx gc.Node
gc.Nodreg(&dx, gc.Types[gc.TINT32], x86.REG_DX)
// Setup for binary operation.
var hi1 gc.Node
var lo1 gc.Node
split64(l, &lo1, &hi1)
var lo2 gc.Node
var hi2 gc.Node
if gc.Is64(r.Type) {
split64(r, &lo2, &hi2)
}
// Do op. Leave result in DX:AX.
switch n.Op {
// TODO: Constants
case gc.OADD:
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(x86.AADDL, &lo2, &ax)
gins(x86.AADCL, &hi2, &dx)
// TODO: Constants.
case gc.OSUB:
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(x86.ASUBL, &lo2, &ax)
gins(x86.ASBBL, &hi2, &dx)
case gc.OMUL:
// let's call the next three EX, FX and GX
var ex, fx, gx gc.Node
gc.Regalloc(&ex, gc.Types[gc.TPTR32], nil)
gc.Regalloc(&fx, gc.Types[gc.TPTR32], nil)
gc.Regalloc(&gx, gc.Types[gc.TPTR32], nil)
// load args into DX:AX and EX:GX.
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(x86.AMOVL, &lo2, &gx)
gins(x86.AMOVL, &hi2, &ex)
// if DX and EX are zero, use 32 x 32 -> 64 unsigned multiply.
gins(x86.AMOVL, &dx, &fx)
gins(x86.AORL, &ex, &fx)
p1 := gc.Gbranch(x86.AJNE, nil, 0)
gins(x86.AMULL, &gx, nil) // implicit &ax
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// full 64x64 -> 64, from 32x32 -> 64.
gins(x86.AIMULL, &gx, &dx)
gins(x86.AMOVL, &ax, &fx)
gins(x86.AIMULL, &ex, &fx)
gins(x86.AADDL, &dx, &fx)
gins(x86.AMOVL, &gx, &dx)
gins(x86.AMULL, &dx, nil) // implicit &ax
gins(x86.AADDL, &fx, &dx)
gc.Patch(p2, gc.Pc)
gc.Regfree(&ex)
gc.Regfree(&fx)
gc.Regfree(&gx)
// We only rotate by a constant c in [0,64).
// if c >= 32:
// lo, hi = hi, lo
// c -= 32
// if c == 0:
// no-op
// else:
// t = hi
// shld hi:lo, c
// shld lo:t, c
case gc.OLROT:
v := uint64(r.Int())
if v >= 32 {
// reverse during load to do the first 32 bits of rotate
v -= 32
gins(x86.AMOVL, &lo1, &dx)
gins(x86.AMOVL, &hi1, &ax)
} else {
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
}
if v == 0 {
} else // done
{
gins(x86.AMOVL, &dx, &cx)
p1 := gins(x86.ASHLL, ncon(uint32(v)), &dx)
p1.From.Index = x86.REG_AX // double-width shift
p1.From.Scale = 0
p1 = gins(x86.ASHLL, ncon(uint32(v)), &ax)
p1.From.Index = x86.REG_CX // double-width shift
p1.From.Scale = 0
}
case gc.OLSH:
if r.Op == gc.OLITERAL {
v := uint64(r.Int())
if v >= 64 {
if gc.Is64(r.Type) {
splitclean()
}
splitclean()
split64(res, &lo2, &hi2)
gins(x86.AMOVL, ncon(0), &lo2)
gins(x86.AMOVL, ncon(0), &hi2)
splitclean()
return
}
if v >= 32 {
if gc.Is64(r.Type) {
splitclean()
}
split64(res, &lo2, &hi2)
gmove(&lo1, &hi2)
if v > 32 {
gins(x86.ASHLL, ncon(uint32(v-32)), &hi2)
}
gins(x86.AMOVL, ncon(0), &lo2)
splitclean()
splitclean()
return
}
// general shift
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
p1 := gins(x86.ASHLL, ncon(uint32(v)), &dx)
p1.From.Index = x86.REG_AX // double-width shift
p1.From.Scale = 0
gins(x86.ASHLL, ncon(uint32(v)), &ax)
break
}
// load value into DX:AX.
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
// load shift value into register.
// if high bits are set, zero value.
var p1 *obj.Prog
if gc.Is64(r.Type) {
gins(x86.ACMPL, &hi2, ncon(0))
p1 = gc.Gbranch(x86.AJNE, nil, +1)
gins(x86.AMOVL, &lo2, &cx)
} else {
cx.Type = gc.Types[gc.TUINT32]
gmove(r, &cx)
}
// if shift count is >=64, zero value
gins(x86.ACMPL, &cx, ncon(64))
p2 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
if p1 != nil {
gc.Patch(p1, gc.Pc)
}
gins(x86.AXORL, &dx, &dx)
gins(x86.AXORL, &ax, &ax)
gc.Patch(p2, gc.Pc)
// if shift count is >= 32, zero low.
gins(x86.ACMPL, &cx, ncon(32))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
gins(x86.AMOVL, &ax, &dx)
gins(x86.ASHLL, &cx, &dx) // SHLL only uses bottom 5 bits of count
gins(x86.AXORL, &ax, &ax)
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// general shift
p1 = gins(x86.ASHLL, &cx, &dx)
p1.From.Index = x86.REG_AX // double-width shift
p1.From.Scale = 0
gins(x86.ASHLL, &cx, &ax)
gc.Patch(p2, gc.Pc)
case gc.ORSH:
if r.Op == gc.OLITERAL {
v := uint64(r.Int())
if v >= 64 {
if gc.Is64(r.Type) {
splitclean()
}
splitclean()
split64(res, &lo2, &hi2)
if hi1.Type.Etype == gc.TINT32 {
gmove(&hi1, &lo2)
gins(x86.ASARL, ncon(31), &lo2)
gmove(&hi1, &hi2)
gins(x86.ASARL, ncon(31), &hi2)
} else {
gins(x86.AMOVL, ncon(0), &lo2)
gins(x86.AMOVL, ncon(0), &hi2)
}
splitclean()
return
}
if v >= 32 {
if gc.Is64(r.Type) {
splitclean()
}
split64(res, &lo2, &hi2)
gmove(&hi1, &lo2)
if v > 32 {
gins(optoas(gc.ORSH, hi1.Type), ncon(uint32(v-32)), &lo2)
}
if hi1.Type.Etype == gc.TINT32 {
gmove(&hi1, &hi2)
gins(x86.ASARL, ncon(31), &hi2)
} else {
gins(x86.AMOVL, ncon(0), &hi2)
}
splitclean()
splitclean()
return
}
// general shift
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
p1 := gins(x86.ASHRL, ncon(uint32(v)), &ax)
p1.From.Index = x86.REG_DX // double-width shift
p1.From.Scale = 0
gins(optoas(gc.ORSH, hi1.Type), ncon(uint32(v)), &dx)
break
}
// load value into DX:AX.
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
// load shift value into register.
// if high bits are set, zero value.
var p1 *obj.Prog
if gc.Is64(r.Type) {
gins(x86.ACMPL, &hi2, ncon(0))
p1 = gc.Gbranch(x86.AJNE, nil, +1)
gins(x86.AMOVL, &lo2, &cx)
} else {
cx.Type = gc.Types[gc.TUINT32]
gmove(r, &cx)
}
// if shift count is >=64, zero or sign-extend value
gins(x86.ACMPL, &cx, ncon(64))
p2 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
if p1 != nil {
gc.Patch(p1, gc.Pc)
}
if hi1.Type.Etype == gc.TINT32 {
gins(x86.ASARL, ncon(31), &dx)
gins(x86.AMOVL, &dx, &ax)
} else {
gins(x86.AXORL, &dx, &dx)
gins(x86.AXORL, &ax, &ax)
}
gc.Patch(p2, gc.Pc)
// if shift count is >= 32, sign-extend hi.
gins(x86.ACMPL, &cx, ncon(32))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
gins(x86.AMOVL, &dx, &ax)
if hi1.Type.Etype == gc.TINT32 {
gins(x86.ASARL, &cx, &ax) // SARL only uses bottom 5 bits of count
gins(x86.ASARL, ncon(31), &dx)
} else {
gins(x86.ASHRL, &cx, &ax)
gins(x86.AXORL, &dx, &dx)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// general shift
p1 = gins(x86.ASHRL, &cx, &ax)
p1.From.Index = x86.REG_DX // double-width shift
p1.From.Scale = 0
gins(optoas(gc.ORSH, hi1.Type), &cx, &dx)
gc.Patch(p2, gc.Pc)
// make constant the right side (it usually is anyway).
case gc.OXOR,
gc.OAND,
gc.OOR:
if lo1.Op == gc.OLITERAL {
nswap(&lo1, &lo2)
nswap(&hi1, &hi2)
}
if lo2.Op == gc.OLITERAL {
// special cases for constants.
lv := uint32(lo2.Int())
hv := uint32(hi2.Int())
splitclean() // right side
split64(res, &lo2, &hi2)
switch n.Op {
case gc.OXOR:
gmove(&lo1, &lo2)
gmove(&hi1, &hi2)
switch lv {
case 0:
break
case 0xffffffff:
gins(x86.ANOTL, nil, &lo2)
default:
gins(x86.AXORL, ncon(lv), &lo2)
}
switch hv {
case 0:
break
case 0xffffffff:
gins(x86.ANOTL, nil, &hi2)
default:
gins(x86.AXORL, ncon(hv), &hi2)
}
case gc.OAND:
switch lv {
case 0:
gins(x86.AMOVL, ncon(0), &lo2)
default:
gmove(&lo1, &lo2)
if lv != 0xffffffff {
gins(x86.AANDL, ncon(lv), &lo2)
}
}
switch hv {
case 0:
gins(x86.AMOVL, ncon(0), &hi2)
default:
gmove(&hi1, &hi2)
if hv != 0xffffffff {
gins(x86.AANDL, ncon(hv), &hi2)
}
}
case gc.OOR:
switch lv {
case 0:
gmove(&lo1, &lo2)
case 0xffffffff:
gins(x86.AMOVL, ncon(0xffffffff), &lo2)
default:
gmove(&lo1, &lo2)
gins(x86.AORL, ncon(lv), &lo2)
}
switch hv {
case 0:
gmove(&hi1, &hi2)
case 0xffffffff:
gins(x86.AMOVL, ncon(0xffffffff), &hi2)
default:
gmove(&hi1, &hi2)
gins(x86.AORL, ncon(hv), &hi2)
}
}
splitclean()
splitclean()
return
}
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(optoas(n.Op, lo1.Type), &lo2, &ax)
gins(optoas(n.Op, lo1.Type), &hi2, &dx)
}
if gc.Is64(r.Type) {
splitclean()
}
splitclean()
split64(res, &lo1, &hi1)
gins(x86.AMOVL, &ax, &lo1)
gins(x86.AMOVL, &dx, &hi1)
splitclean()
}
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 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)
}
func gmove(f *gc.Node, t *gc.Node) {
if gc.Debug['M'] != 0 {
fmt.Printf("gmove %v -> %v\n", f, t)
}
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
cvt := t.Type
if gc.Iscomplex[ft] || gc.Iscomplex[tt] {
gc.Complexmove(f, t)
return
}
if gc.Isfloat[ft] || gc.Isfloat[tt] {
floatmove(f, t)
return
}
// cannot have two integer memory operands;
// except 64-bit, which always copies via registers anyway.
var r1 gc.Node
var a int
if gc.Isint[ft] && gc.Isint[tt] && !gc.Is64(f.Type) && !gc.Is64(t.Type) && gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
f.Convconst(&con, t.Type)
f = &con
ft = gc.Simsimtype(con.Type)
}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
// should not happen
gc.Fatalf("gmove %v -> %v", f, t)
return
/*
* integer copy and truncate
*/
case gc.TINT8<<16 | gc.TINT8, // same size
gc.TINT8<<16 | gc.TUINT8,
gc.TUINT8<<16 | gc.TINT8,
gc.TUINT8<<16 | gc.TUINT8:
a = x86.AMOVB
case gc.TINT16<<16 | gc.TINT8, // truncate
gc.TUINT16<<16 | gc.TINT8,
gc.TINT32<<16 | gc.TINT8,
gc.TUINT32<<16 | gc.TINT8,
gc.TINT16<<16 | gc.TUINT8,
gc.TUINT16<<16 | gc.TUINT8,
gc.TINT32<<16 | gc.TUINT8,
gc.TUINT32<<16 | gc.TUINT8:
a = x86.AMOVB
goto rsrc
case gc.TINT64<<16 | gc.TINT8, // truncate low word
gc.TUINT64<<16 | gc.TINT8,
gc.TINT64<<16 | gc.TUINT8,
gc.TUINT64<<16 | gc.TUINT8:
var flo gc.Node
var fhi gc.Node
split64(f, &flo, &fhi)
var r1 gc.Node
gc.Nodreg(&r1, t.Type, x86.REG_AX)
gmove(&flo, &r1)
gins(x86.AMOVB, &r1, t)
splitclean()
return
case gc.TINT16<<16 | gc.TINT16, // same size
gc.TINT16<<16 | gc.TUINT16,
gc.TUINT16<<16 | gc.TINT16,
gc.TUINT16<<16 | gc.TUINT16:
a = x86.AMOVW
case gc.TINT32<<16 | gc.TINT16, // truncate
gc.TUINT32<<16 | gc.TINT16,
gc.TINT32<<16 | gc.TUINT16,
gc.TUINT32<<16 | gc.TUINT16:
a = x86.AMOVW
goto rsrc
case gc.TINT64<<16 | gc.TINT16, // truncate low word
gc.TUINT64<<16 | gc.TINT16,
gc.TINT64<<16 | gc.TUINT16,
gc.TUINT64<<16 | gc.TUINT16:
var flo gc.Node
var fhi gc.Node
split64(f, &flo, &fhi)
var r1 gc.Node
gc.Nodreg(&r1, t.Type, x86.REG_AX)
gmove(&flo, &r1)
gins(x86.AMOVW, &r1, t)
splitclean()
return
case gc.TINT32<<16 | gc.TINT32, // same size
gc.TINT32<<16 | gc.TUINT32,
gc.TUINT32<<16 | gc.TINT32,
gc.TUINT32<<16 | gc.TUINT32:
a = x86.AMOVL
case gc.TINT64<<16 | gc.TINT32, // truncate
gc.TUINT64<<16 | gc.TINT32,
gc.TINT64<<16 | gc.TUINT32,
gc.TUINT64<<16 | gc.TUINT32:
var fhi gc.Node
var flo gc.Node
split64(f, &flo, &fhi)
var r1 gc.Node
gc.Nodreg(&r1, t.Type, x86.REG_AX)
gmove(&flo, &r1)
gins(x86.AMOVL, &r1, t)
splitclean()
return
case gc.TINT64<<16 | gc.TINT64, // same size
gc.TINT64<<16 | gc.TUINT64,
gc.TUINT64<<16 | gc.TINT64,
gc.TUINT64<<16 | gc.TUINT64:
var fhi gc.Node
var flo gc.Node
split64(f, &flo, &fhi)
var tlo gc.Node
var thi gc.Node
split64(t, &tlo, &thi)
if f.Op == gc.OLITERAL {
gins(x86.AMOVL, &flo, &tlo)
gins(x86.AMOVL, &fhi, &thi)
} else {
// Implementation of conversion-free x = y for int64 or uint64 x.
// This is generated by the code that copies small values out of closures,
// and that code has DX live, so avoid DX and just use AX twice.
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[gc.TUINT32], x86.REG_AX)
gins(x86.AMOVL, &flo, &r1)
gins(x86.AMOVL, &r1, &tlo)
gins(x86.AMOVL, &fhi, &r1)
gins(x86.AMOVL, &r1, &thi)
}
splitclean()
splitclean()
return
/*
* integer up-conversions
*/
case gc.TINT8<<16 | gc.TINT16, // sign extend int8
gc.TINT8<<16 | gc.TUINT16:
a = x86.AMOVBWSX
goto rdst
case gc.TINT8<<16 | gc.TINT32,
gc.TINT8<<16 | gc.TUINT32:
a = x86.AMOVBLSX
goto rdst
case gc.TINT8<<16 | gc.TINT64, // convert via int32
gc.TINT8<<16 | gc.TUINT64:
cvt = gc.Types[gc.TINT32]
goto hard
case gc.TUINT8<<16 | gc.TINT16, // zero extend uint8
gc.TUINT8<<16 | gc.TUINT16:
a = x86.AMOVBWZX
goto rdst
case gc.TUINT8<<16 | gc.TINT32,
gc.TUINT8<<16 | gc.TUINT32:
a = x86.AMOVBLZX
goto rdst
case gc.TUINT8<<16 | gc.TINT64, // convert via uint32
gc.TUINT8<<16 | gc.TUINT64:
cvt = gc.Types[gc.TUINT32]
goto hard
case gc.TINT16<<16 | gc.TINT32, // sign extend int16
gc.TINT16<<16 | gc.TUINT32:
a = x86.AMOVWLSX
goto rdst
case gc.TINT16<<16 | gc.TINT64, // convert via int32
gc.TINT16<<16 | gc.TUINT64:
cvt = gc.Types[gc.TINT32]
goto hard
case gc.TUINT16<<16 | gc.TINT32, // zero extend uint16
gc.TUINT16<<16 | gc.TUINT32:
a = x86.AMOVWLZX
goto rdst
case gc.TUINT16<<16 | gc.TINT64, // convert via uint32
gc.TUINT16<<16 | gc.TUINT64:
cvt = gc.Types[gc.TUINT32]
goto hard
case gc.TINT32<<16 | gc.TINT64, // sign extend int32
gc.TINT32<<16 | gc.TUINT64:
var thi gc.Node
var tlo gc.Node
split64(t, &tlo, &thi)
var flo gc.Node
gc.Nodreg(&flo, tlo.Type, x86.REG_AX)
var fhi gc.Node
gc.Nodreg(&fhi, thi.Type, x86.REG_DX)
gmove(f, &flo)
gins(x86.ACDQ, nil, nil)
gins(x86.AMOVL, &flo, &tlo)
gins(x86.AMOVL, &fhi, &thi)
splitclean()
return
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
var tlo gc.Node
var thi gc.Node
split64(t, &tlo, &thi)
gmove(f, &tlo)
gins(x86.AMOVL, ncon(0), &thi)
splitclean()
return
}
gins(a, f, t)
return
// requires register source
rsrc:
gc.Regalloc(&r1, f.Type, t)
gmove(f, &r1)
gins(a, &r1, t)
gc.Regfree(&r1)
return
// requires register destination
rdst:
{
gc.Regalloc(&r1, t.Type, t)
gins(a, f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
/*
* generate move:
* t = f
* hard part is conversions.
*/
func gmove(f *gc.Node, t *gc.Node) {
if gc.Debug['M'] != 0 {
fmt.Printf("gmove %v -> %v\n", gc.Nconv(f, obj.FmtLong), gc.Nconv(t, obj.FmtLong))
}
ft := int(gc.Simsimtype(f.Type))
tt := int(gc.Simsimtype(t.Type))
cvt := (*gc.Type)(t.Type)
if gc.Iscomplex[ft] || gc.Iscomplex[tt] {
gc.Complexmove(f, t)
return
}
// cannot have two memory operands
var r2 gc.Node
var r1 gc.Node
var a int
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
switch tt {
default:
f.Convconst(&con, t.Type)
case gc.TINT32,
gc.TINT16,
gc.TINT8:
var con gc.Node
f.Convconst(&con, gc.Types[gc.TINT64])
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(ppc64.AMOVD, &con, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
case gc.TUINT32,
gc.TUINT16,
gc.TUINT8:
var con gc.Node
f.Convconst(&con, gc.Types[gc.TUINT64])
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(ppc64.AMOVD, &con, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
f = &con
ft = tt // so big switch will choose a simple mov
// constants can't move directly to memory.
if gc.Ismem(t) {
goto hard
}
}
// float constants come from memory.
//if(isfloat[tt])
// goto hard;
// 64-bit immediates are also from memory.
//if(isint[tt])
// goto hard;
//// 64-bit immediates are really 32-bit sign-extended
//// unless moving into a register.
//if(isint[tt]) {
// if(mpcmpfixfix(con.val.u.xval, minintval[TINT32]) < 0)
// goto hard;
// if(mpcmpfixfix(con.val.u.xval, maxintval[TINT32]) > 0)
// goto hard;
//}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
gc.Fatalf("gmove %v -> %v", gc.Tconv(f.Type, obj.FmtLong), gc.Tconv(t.Type, obj.FmtLong))
/*
* integer copy and truncate
*/
case gc.TINT8<<16 | gc.TINT8, // same size
gc.TUINT8<<16 | gc.TINT8,
gc.TINT16<<16 | gc.TINT8,
// truncate
gc.TUINT16<<16 | gc.TINT8,
gc.TINT32<<16 | gc.TINT8,
gc.TUINT32<<16 | gc.TINT8,
gc.TINT64<<16 | gc.TINT8,
gc.TUINT64<<16 | gc.TINT8:
a = ppc64.AMOVB
case gc.TINT8<<16 | gc.TUINT8, // same size
gc.TUINT8<<16 | gc.TUINT8,
gc.TINT16<<16 | gc.TUINT8,
// truncate
gc.TUINT16<<16 | gc.TUINT8,
gc.TINT32<<16 | gc.TUINT8,
gc.TUINT32<<16 | gc.TUINT8,
gc.TINT64<<16 | gc.TUINT8,
gc.TUINT64<<16 | gc.TUINT8:
a = ppc64.AMOVBZ
case gc.TINT16<<16 | gc.TINT16, // same size
gc.TUINT16<<16 | gc.TINT16,
gc.TINT32<<16 | gc.TINT16,
// truncate
gc.TUINT32<<16 | gc.TINT16,
gc.TINT64<<16 | gc.TINT16,
gc.TUINT64<<16 | gc.TINT16:
a = ppc64.AMOVH
case gc.TINT16<<16 | gc.TUINT16, // same size
gc.TUINT16<<16 | gc.TUINT16,
gc.TINT32<<16 | gc.TUINT16,
// truncate
gc.TUINT32<<16 | gc.TUINT16,
gc.TINT64<<16 | gc.TUINT16,
gc.TUINT64<<16 | gc.TUINT16:
a = ppc64.AMOVHZ
case gc.TINT32<<16 | gc.TINT32, // same size
gc.TUINT32<<16 | gc.TINT32,
gc.TINT64<<16 | gc.TINT32,
// truncate
gc.TUINT64<<16 | gc.TINT32:
a = ppc64.AMOVW
case gc.TINT32<<16 | gc.TUINT32, // same size
gc.TUINT32<<16 | gc.TUINT32,
gc.TINT64<<16 | gc.TUINT32,
gc.TUINT64<<16 | gc.TUINT32:
a = ppc64.AMOVWZ
case gc.TINT64<<16 | gc.TINT64, // same size
gc.TINT64<<16 | gc.TUINT64,
gc.TUINT64<<16 | gc.TINT64,
gc.TUINT64<<16 | gc.TUINT64:
a = ppc64.AMOVD
/*
* integer up-conversions
*/
case gc.TINT8<<16 | gc.TINT16, // sign extend int8
gc.TINT8<<16 | gc.TUINT16,
gc.TINT8<<16 | gc.TINT32,
gc.TINT8<<16 | gc.TUINT32,
gc.TINT8<<16 | gc.TINT64,
gc.TINT8<<16 | gc.TUINT64:
a = ppc64.AMOVB
goto rdst
case gc.TUINT8<<16 | gc.TINT16, // zero extend uint8
gc.TUINT8<<16 | gc.TUINT16,
gc.TUINT8<<16 | gc.TINT32,
gc.TUINT8<<16 | gc.TUINT32,
gc.TUINT8<<16 | gc.TINT64,
gc.TUINT8<<16 | gc.TUINT64:
a = ppc64.AMOVBZ
goto rdst
case gc.TINT16<<16 | gc.TINT32, // sign extend int16
gc.TINT16<<16 | gc.TUINT32,
gc.TINT16<<16 | gc.TINT64,
gc.TINT16<<16 | gc.TUINT64:
a = ppc64.AMOVH
goto rdst
case gc.TUINT16<<16 | gc.TINT32, // zero extend uint16
gc.TUINT16<<16 | gc.TUINT32,
gc.TUINT16<<16 | gc.TINT64,
gc.TUINT16<<16 | gc.TUINT64:
a = ppc64.AMOVHZ
goto rdst
case gc.TINT32<<16 | gc.TINT64, // sign extend int32
gc.TINT32<<16 | gc.TUINT64:
a = ppc64.AMOVW
goto rdst
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
a = ppc64.AMOVWZ
goto rdst
//warn("gmove: convert float to int not implemented: %N -> %N\n", f, t);
//return;
// algorithm is:
// if small enough, use native float64 -> int64 conversion.
// otherwise, subtract 2^63, convert, and add it back.
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT32,
gc.TFLOAT64<<16 | gc.TINT32,
gc.TFLOAT32<<16 | gc.TINT64,
gc.TFLOAT64<<16 | gc.TINT64,
gc.TFLOAT32<<16 | gc.TINT16,
gc.TFLOAT32<<16 | gc.TINT8,
gc.TFLOAT32<<16 | gc.TUINT16,
gc.TFLOAT32<<16 | gc.TUINT8,
gc.TFLOAT64<<16 | gc.TINT16,
gc.TFLOAT64<<16 | gc.TINT8,
gc.TFLOAT64<<16 | gc.TUINT16,
gc.TFLOAT64<<16 | gc.TUINT8,
gc.TFLOAT32<<16 | gc.TUINT32,
gc.TFLOAT64<<16 | gc.TUINT32,
gc.TFLOAT32<<16 | gc.TUINT64,
gc.TFLOAT64<<16 | gc.TUINT64:
bignodes()
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[ft], f)
gmove(f, &r1)
if tt == gc.TUINT64 {
gc.Regalloc(&r2, gc.Types[gc.TFLOAT64], nil)
gmove(&bigf, &r2)
gins(ppc64.AFCMPU, &r1, &r2)
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TFLOAT64]), nil, +1))
gins(ppc64.AFSUB, &r2, &r1)
gc.Patch(p1, gc.Pc)
gc.Regfree(&r2)
}
gc.Regalloc(&r2, gc.Types[gc.TFLOAT64], nil)
var r3 gc.Node
gc.Regalloc(&r3, gc.Types[gc.TINT64], t)
gins(ppc64.AFCTIDZ, &r1, &r2)
p1 := (*obj.Prog)(gins(ppc64.AFMOVD, &r2, nil))
p1.To.Type = obj.TYPE_MEM
p1.To.Reg = ppc64.REGSP
p1.To.Offset = -8
p1 = gins(ppc64.AMOVD, nil, &r3)
p1.From.Type = obj.TYPE_MEM
p1.From.Reg = ppc64.REGSP
p1.From.Offset = -8
gc.Regfree(&r2)
gc.Regfree(&r1)
if tt == gc.TUINT64 {
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TFLOAT64]), nil, +1)) // use CR0 here again
gc.Nodreg(&r1, gc.Types[gc.TINT64], ppc64.REGTMP)
gins(ppc64.AMOVD, &bigi, &r1)
gins(ppc64.AADD, &r1, &r3)
gc.Patch(p1, gc.Pc)
}
gmove(&r3, t)
gc.Regfree(&r3)
return
//warn("gmove: convert int to float not implemented: %N -> %N\n", f, t);
//return;
// algorithm is:
// if small enough, use native int64 -> uint64 conversion.
// otherwise, halve (rounding to odd?), convert, and double.
/*
* integer to float
*/
case gc.TINT32<<16 | gc.TFLOAT32,
gc.TINT32<<16 | gc.TFLOAT64,
gc.TINT64<<16 | gc.TFLOAT32,
gc.TINT64<<16 | gc.TFLOAT64,
gc.TINT16<<16 | gc.TFLOAT32,
gc.TINT16<<16 | gc.TFLOAT64,
gc.TINT8<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TUINT16<<16 | gc.TFLOAT32,
gc.TUINT16<<16 | gc.TFLOAT64,
gc.TUINT8<<16 | gc.TFLOAT32,
gc.TUINT8<<16 | gc.TFLOAT64,
gc.TUINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT64,
gc.TUINT64<<16 | gc.TFLOAT32,
gc.TUINT64<<16 | gc.TFLOAT64:
bignodes()
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[gc.TINT64], nil)
gmove(f, &r1)
if ft == gc.TUINT64 {
gc.Nodreg(&r2, gc.Types[gc.TUINT64], ppc64.REGTMP)
gmove(&bigi, &r2)
gins(ppc64.ACMPU, &r1, &r2)
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT64]), nil, +1))
p2 := (*obj.Prog)(gins(ppc64.ASRD, nil, &r1))
p2.From.Type = obj.TYPE_CONST
p2.From.Offset = 1
gc.Patch(p1, gc.Pc)
}
gc.Regalloc(&r2, gc.Types[gc.TFLOAT64], t)
p1 := (*obj.Prog)(gins(ppc64.AMOVD, &r1, nil))
p1.To.Type = obj.TYPE_MEM
p1.To.Reg = ppc64.REGSP
p1.To.Offset = -8
p1 = gins(ppc64.AFMOVD, nil, &r2)
p1.From.Type = obj.TYPE_MEM
p1.From.Reg = ppc64.REGSP
p1.From.Offset = -8
gins(ppc64.AFCFID, &r2, &r2)
gc.Regfree(&r1)
if ft == gc.TUINT64 {
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT64]), nil, +1)) // use CR0 here again
gc.Nodreg(&r1, gc.Types[gc.TFLOAT64], ppc64.FREGTWO)
gins(ppc64.AFMUL, &r1, &r2)
gc.Patch(p1, gc.Pc)
}
gmove(&r2, t)
gc.Regfree(&r2)
return
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32:
a = ppc64.AFMOVS
case gc.TFLOAT64<<16 | gc.TFLOAT64:
a = ppc64.AFMOVD
case gc.TFLOAT32<<16 | gc.TFLOAT64:
a = ppc64.AFMOVS
goto rdst
case gc.TFLOAT64<<16 | gc.TFLOAT32:
a = ppc64.AFRSP
goto rdst
}
gins(a, f, t)
return
// requires register destination
rdst:
{
gc.Regalloc(&r1, t.Type, t)
gins(a, f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
func ginsnop() {
var reg gc.Node
gc.Nodreg(®, gc.Types[gc.TINT], x86.REG_AX)
gins(x86.AXCHGL, ®, ®)
}
func bgen_float(n *gc.Node, wantTrue bool, likely int, to *obj.Prog) {
nl := n.Left
nr := n.Right
op := n.Op
if !wantTrue {
// brcom is not valid on floats when NaN is involved.
p1 := gc.Gbranch(obj.AJMP, nil, 0)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// No need to avoid re-genning ninit.
bgen_float(n, true, -likely, p2)
gc.Patch(gc.Gbranch(obj.AJMP, nil, 0), to)
gc.Patch(p2, gc.Pc)
return
}
if gc.Thearch.Use387 {
op = gc.Brrev(op) // because the args are stacked
if op == gc.OGE || op == gc.OGT {
// only < and <= work right with NaN; reverse if needed
nl, nr = nr, nl
op = gc.Brrev(op)
}
var ax, n2, tmp gc.Node
gc.Nodreg(&tmp, nr.Type, x86.REG_F0)
gc.Nodreg(&n2, nr.Type, x86.REG_F0+1)
gc.Nodreg(&ax, gc.Types[gc.TUINT16], x86.REG_AX)
if gc.Simsimtype(nr.Type) == gc.TFLOAT64 {
if nl.Ullman > nr.Ullman {
gc.Cgen(nl, &tmp)
gc.Cgen(nr, &tmp)
gins(x86.AFXCHD, &tmp, &n2)
} else {
gc.Cgen(nr, &tmp)
gc.Cgen(nl, &tmp)
}
gins(x86.AFUCOMPP, &tmp, &n2)
} else {
// TODO(rsc): The moves back and forth to memory
// here are for truncating the value to 32 bits.
// This handles 32-bit comparison but presumably
// all the other ops have the same problem.
// We need to figure out what the right general
// solution is, besides telling people to use float64.
var t1 gc.Node
gc.Tempname(&t1, gc.Types[gc.TFLOAT32])
var t2 gc.Node
gc.Tempname(&t2, gc.Types[gc.TFLOAT32])
gc.Cgen(nr, &t1)
gc.Cgen(nl, &t2)
gmove(&t2, &tmp)
gins(x86.AFCOMFP, &t1, &tmp)
}
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
} else {
// Not 387
if !nl.Addable {
nl = gc.CgenTemp(nl)
}
if !nr.Addable {
nr = gc.CgenTemp(nr)
}
var n2 gc.Node
gc.Regalloc(&n2, nr.Type, nil)
gmove(nr, &n2)
nr = &n2
if nl.Op != gc.OREGISTER {
var n3 gc.Node
gc.Regalloc(&n3, nl.Type, nil)
gmove(nl, &n3)
nl = &n3
}
if op == gc.OGE || op == gc.OGT {
// only < and <= work right with NopN; reverse if needed
nl, nr = nr, nl
op = gc.Brrev(op)
}
gins(foptoas(gc.OCMP, nr.Type, 0), nl, nr)
if nl.Op == gc.OREGISTER {
gc.Regfree(nl)
}
gc.Regfree(nr)
}
switch op {
case gc.OEQ:
// neither NE nor P
p1 := gc.Gbranch(x86.AJNE, nil, -likely)
p2 := gc.Gbranch(x86.AJPS, nil, -likely)
gc.Patch(gc.Gbranch(obj.AJMP, nil, 0), to)
gc.Patch(p1, gc.Pc)
gc.Patch(p2, gc.Pc)
case gc.ONE:
// either NE or P
gc.Patch(gc.Gbranch(x86.AJNE, nil, likely), to)
gc.Patch(gc.Gbranch(x86.AJPS, nil, likely), to)
default:
gc.Patch(gc.Gbranch(optoas(op, nr.Type), nil, likely), to)
}
}
func floatmove_387(f *gc.Node, t *gc.Node) {
var r1 gc.Node
var a int
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
cvt := t.Type
switch uint32(ft)<<16 | uint32(tt) {
default:
goto fatal
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT16,
gc.TFLOAT32<<16 | gc.TINT32,
gc.TFLOAT32<<16 | gc.TINT64,
gc.TFLOAT64<<16 | gc.TINT16,
gc.TFLOAT64<<16 | gc.TINT32,
gc.TFLOAT64<<16 | gc.TINT64:
if t.Op == gc.OREGISTER {
goto hardmem
}
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[ft], x86.REG_F0)
if f.Op != gc.OREGISTER {
if ft == gc.TFLOAT32 {
gins(x86.AFMOVF, f, &r1)
} else {
gins(x86.AFMOVD, f, &r1)
}
}
// set round to zero mode during conversion
var t1 gc.Node
memname(&t1, gc.Types[gc.TUINT16])
var t2 gc.Node
memname(&t2, gc.Types[gc.TUINT16])
gins(x86.AFSTCW, nil, &t1)
gins(x86.AMOVW, ncon(0xf7f), &t2)
gins(x86.AFLDCW, &t2, nil)
if tt == gc.TINT16 {
gins(x86.AFMOVWP, &r1, t)
} else if tt == gc.TINT32 {
gins(x86.AFMOVLP, &r1, t)
} else {
gins(x86.AFMOVVP, &r1, t)
}
gins(x86.AFLDCW, &t1, nil)
return
// convert via int32.
case gc.TFLOAT32<<16 | gc.TINT8,
gc.TFLOAT32<<16 | gc.TUINT16,
gc.TFLOAT32<<16 | gc.TUINT8,
gc.TFLOAT64<<16 | gc.TINT8,
gc.TFLOAT64<<16 | gc.TUINT16,
gc.TFLOAT64<<16 | gc.TUINT8:
var t1 gc.Node
gc.Tempname(&t1, gc.Types[gc.TINT32])
gmove(f, &t1)
switch tt {
default:
gc.Fatalf("gmove %v", t)
case gc.TINT8:
gins(x86.ACMPL, &t1, ncon(-0x80&(1<<32-1)))
p1 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TINT32]), nil, -1)
gins(x86.ACMPL, &t1, ncon(0x7f))
p2 := gc.Gbranch(optoas(gc.OGT, gc.Types[gc.TINT32]), nil, -1)
p3 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
gc.Patch(p2, gc.Pc)
gmove(ncon(-0x80&(1<<32-1)), &t1)
gc.Patch(p3, gc.Pc)
gmove(&t1, t)
case gc.TUINT8:
gins(x86.ATESTL, ncon(0xffffff00), &t1)
p1 := gc.Gbranch(x86.AJEQ, nil, +1)
gins(x86.AMOVL, ncon(0), &t1)
gc.Patch(p1, gc.Pc)
gmove(&t1, t)
case gc.TUINT16:
gins(x86.ATESTL, ncon(0xffff0000), &t1)
p1 := gc.Gbranch(x86.AJEQ, nil, +1)
gins(x86.AMOVL, ncon(0), &t1)
gc.Patch(p1, gc.Pc)
gmove(&t1, t)
}
return
// convert via int64.
case gc.TFLOAT32<<16 | gc.TUINT32,
gc.TFLOAT64<<16 | gc.TUINT32:
cvt = gc.Types[gc.TINT64]
goto hardmem
/*
* integer to float
*/
case gc.TINT16<<16 | gc.TFLOAT32,
gc.TINT16<<16 | gc.TFLOAT64,
gc.TINT32<<16 | gc.TFLOAT32,
gc.TINT32<<16 | gc.TFLOAT64,
gc.TINT64<<16 | gc.TFLOAT32,
gc.TINT64<<16 | gc.TFLOAT64:
if t.Op != gc.OREGISTER {
goto hard
}
if f.Op == gc.OREGISTER {
cvt = f.Type
goto hardmem
}
switch ft {
case gc.TINT16:
a = x86.AFMOVW
case gc.TINT32:
a = x86.AFMOVL
default:
a = x86.AFMOVV
}
// convert via int32 memory
case gc.TINT8<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TUINT16<<16 | gc.TFLOAT32,
gc.TUINT16<<16 | gc.TFLOAT64,
gc.TUINT8<<16 | gc.TFLOAT32,
gc.TUINT8<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT32]
goto hardmem
// convert via int64 memory
case gc.TUINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT64]
goto hardmem
// The way the code generator uses floating-point
// registers, a move from F0 to F0 is intended as a no-op.
// On the x86, it's not: it pushes a second copy of F0
// on the floating point stack. So toss it away here.
// Also, F0 is the *only* register we ever evaluate
// into, so we should only see register/register as F0/F0.
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32,
gc.TFLOAT64<<16 | gc.TFLOAT64:
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
if f.Op == gc.OREGISTER && t.Op == gc.OREGISTER {
if f.Reg != x86.REG_F0 || t.Reg != x86.REG_F0 {
goto fatal
}
return
}
a = x86.AFMOVF
if ft == gc.TFLOAT64 {
a = x86.AFMOVD
}
if gc.Ismem(t) {
if f.Op != gc.OREGISTER || f.Reg != x86.REG_F0 {
gc.Fatalf("gmove %v", f)
}
a = x86.AFMOVFP
if ft == gc.TFLOAT64 {
a = x86.AFMOVDP
}
}
case gc.TFLOAT32<<16 | gc.TFLOAT64:
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
if f.Op == gc.OREGISTER && t.Op == gc.OREGISTER {
if f.Reg != x86.REG_F0 || t.Reg != x86.REG_F0 {
goto fatal
}
return
}
if f.Op == gc.OREGISTER {
gins(x86.AFMOVDP, f, t)
} else {
gins(x86.AFMOVF, f, t)
}
return
case gc.TFLOAT64<<16 | gc.TFLOAT32:
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
if f.Op == gc.OREGISTER && t.Op == gc.OREGISTER {
var r1 gc.Node
gc.Tempname(&r1, gc.Types[gc.TFLOAT32])
gins(x86.AFMOVFP, f, &r1)
gins(x86.AFMOVF, &r1, t)
return
}
if f.Op == gc.OREGISTER {
gins(x86.AFMOVFP, f, t)
} else {
gins(x86.AFMOVD, f, t)
}
return
}
gins(a, f, t)
return
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
// requires memory intermediate
hardmem:
gc.Tempname(&r1, cvt)
gmove(f, &r1)
gmove(&r1, t)
return
// should not happen
fatal:
gc.Fatalf("gmove %v -> %v", gc.Nconv(f, obj.FmtLong), gc.Nconv(t, obj.FmtLong))
return
}
func floatmove(f *gc.Node, t *gc.Node) {
var r1 gc.Node
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
cvt := t.Type
// cannot have two floating point memory operands.
if gc.Isfloat[ft] && gc.Isfloat[tt] && gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
f.Convconst(&con, t.Type)
f = &con
ft = gc.Simsimtype(con.Type)
// some constants can't move directly to memory.
if gc.Ismem(t) {
// float constants come from memory.
if gc.Isfloat[tt] {
goto hard
}
}
}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
if gc.Thearch.Use387 {
floatmove_387(f, t)
} else {
floatmove_sse(f, t)
}
return
// float to very long integer.
case gc.TFLOAT32<<16 | gc.TINT64,
gc.TFLOAT64<<16 | gc.TINT64:
if f.Op == gc.OREGISTER {
cvt = f.Type
goto hardmem
}
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[ft], x86.REG_F0)
if ft == gc.TFLOAT32 {
gins(x86.AFMOVF, f, &r1)
} else {
gins(x86.AFMOVD, f, &r1)
}
// set round to zero mode during conversion
var t1 gc.Node
memname(&t1, gc.Types[gc.TUINT16])
var t2 gc.Node
memname(&t2, gc.Types[gc.TUINT16])
gins(x86.AFSTCW, nil, &t1)
gins(x86.AMOVW, ncon(0xf7f), &t2)
gins(x86.AFLDCW, &t2, nil)
if tt == gc.TINT16 {
gins(x86.AFMOVWP, &r1, t)
} else if tt == gc.TINT32 {
gins(x86.AFMOVLP, &r1, t)
} else {
gins(x86.AFMOVVP, &r1, t)
}
gins(x86.AFLDCW, &t1, nil)
return
case gc.TFLOAT32<<16 | gc.TUINT64,
gc.TFLOAT64<<16 | gc.TUINT64:
if !gc.Ismem(f) {
cvt = f.Type
goto hardmem
}
bignodes()
var f0 gc.Node
gc.Nodreg(&f0, gc.Types[ft], x86.REG_F0)
var f1 gc.Node
gc.Nodreg(&f1, gc.Types[ft], x86.REG_F0+1)
var ax gc.Node
gc.Nodreg(&ax, gc.Types[gc.TUINT16], x86.REG_AX)
if ft == gc.TFLOAT32 {
gins(x86.AFMOVF, f, &f0)
} else {
gins(x86.AFMOVD, f, &f0)
}
// if 0 > v { answer = 0 }
gins(x86.AFMOVD, &zerof, &f0)
gins(x86.AFUCOMP, &f0, &f1)
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
p1 := gc.Gbranch(optoas(gc.OGT, gc.Types[tt]), nil, 0)
// if 1<<64 <= v { answer = 0 too }
gins(x86.AFMOVD, &two64f, &f0)
gins(x86.AFUCOMP, &f0, &f1)
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
p2 := gc.Gbranch(optoas(gc.OGT, gc.Types[tt]), nil, 0)
gc.Patch(p1, gc.Pc)
gins(x86.AFMOVVP, &f0, t) // don't care about t, but will pop the stack
var thi gc.Node
var tlo gc.Node
split64(t, &tlo, &thi)
gins(x86.AMOVL, ncon(0), &tlo)
gins(x86.AMOVL, ncon(0), &thi)
splitclean()
p1 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p2, gc.Pc)
// in range; algorithm is:
// if small enough, use native float64 -> int64 conversion.
// otherwise, subtract 2^63, convert, and add it back.
// set round to zero mode during conversion
var t1 gc.Node
memname(&t1, gc.Types[gc.TUINT16])
var t2 gc.Node
memname(&t2, gc.Types[gc.TUINT16])
gins(x86.AFSTCW, nil, &t1)
gins(x86.AMOVW, ncon(0xf7f), &t2)
gins(x86.AFLDCW, &t2, nil)
// actual work
gins(x86.AFMOVD, &two63f, &f0)
gins(x86.AFUCOMP, &f0, &f1)
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
p2 = gc.Gbranch(optoas(gc.OLE, gc.Types[tt]), nil, 0)
gins(x86.AFMOVVP, &f0, t)
p3 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p2, gc.Pc)
gins(x86.AFMOVD, &two63f, &f0)
gins(x86.AFSUBDP, &f0, &f1)
gins(x86.AFMOVVP, &f0, t)
split64(t, &tlo, &thi)
gins(x86.AXORL, ncon(0x80000000), &thi) // + 2^63
gc.Patch(p3, gc.Pc)
splitclean()
// restore rounding mode
gins(x86.AFLDCW, &t1, nil)
gc.Patch(p1, gc.Pc)
return
/*
* integer to float
*/
case gc.TINT64<<16 | gc.TFLOAT32,
gc.TINT64<<16 | gc.TFLOAT64:
if t.Op == gc.OREGISTER {
goto hardmem
}
var f0 gc.Node
gc.Nodreg(&f0, t.Type, x86.REG_F0)
gins(x86.AFMOVV, f, &f0)
if tt == gc.TFLOAT32 {
gins(x86.AFMOVFP, &f0, t)
} else {
gins(x86.AFMOVDP, &f0, t)
}
return
// algorithm is:
// if small enough, use native int64 -> float64 conversion.
// otherwise, halve (rounding to odd?), convert, and double.
case gc.TUINT64<<16 | gc.TFLOAT32,
gc.TUINT64<<16 | gc.TFLOAT64:
var ax gc.Node
gc.Nodreg(&ax, gc.Types[gc.TUINT32], x86.REG_AX)
var dx gc.Node
gc.Nodreg(&dx, gc.Types[gc.TUINT32], x86.REG_DX)
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TUINT32], x86.REG_CX)
var t1 gc.Node
gc.Tempname(&t1, f.Type)
var tlo gc.Node
var thi gc.Node
split64(&t1, &tlo, &thi)
gmove(f, &t1)
gins(x86.ACMPL, &thi, ncon(0))
p1 := gc.Gbranch(x86.AJLT, nil, 0)
// native
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[tt], x86.REG_F0)
gins(x86.AFMOVV, &t1, &r1)
if tt == gc.TFLOAT32 {
gins(x86.AFMOVFP, &r1, t)
} else {
gins(x86.AFMOVDP, &r1, t)
}
p2 := gc.Gbranch(obj.AJMP, nil, 0)
// simulated
gc.Patch(p1, gc.Pc)
gmove(&tlo, &ax)
gmove(&thi, &dx)
p1 = gins(x86.ASHRL, ncon(1), &ax)
p1.From.Index = x86.REG_DX // double-width shift DX -> AX
p1.From.Scale = 0
gins(x86.AMOVL, ncon(0), &cx)
gins(x86.ASETCC, nil, &cx)
gins(x86.AORL, &cx, &ax)
gins(x86.ASHRL, ncon(1), &dx)
gmove(&dx, &thi)
gmove(&ax, &tlo)
gc.Nodreg(&r1, gc.Types[tt], x86.REG_F0)
var r2 gc.Node
gc.Nodreg(&r2, gc.Types[tt], x86.REG_F0+1)
gins(x86.AFMOVV, &t1, &r1)
gins(x86.AFMOVD, &r1, &r1)
gins(x86.AFADDDP, &r1, &r2)
if tt == gc.TFLOAT32 {
gins(x86.AFMOVFP, &r1, t)
} else {
gins(x86.AFMOVDP, &r1, t)
}
gc.Patch(p2, gc.Pc)
splitclean()
return
}
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
// requires memory intermediate
hardmem:
gc.Tempname(&r1, cvt)
gmove(f, &r1)
gmove(&r1, t)
return
}
func ginsnop() {
var reg gc.Node
gc.Nodreg(®, gc.Types[gc.TINT], mips.REG_R0)
gins(mips.ANOR, ®, ®)
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op gc.Op, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Type.Width > 4 {
gc.Fatalf("cgen_shift %v", nl.Type)
}
w := int(nl.Type.Width * 8)
a := optoas(op, nl.Type)
if nr.Op == gc.OLITERAL {
var n2 gc.Node
gc.Tempname(&n2, nl.Type)
gc.Cgen(nl, &n2)
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gmove(&n2, &n1)
sc := uint64(nr.Int())
if sc >= uint64(nl.Type.Width*8) {
// large shift gets 2 shifts by width-1
gins(a, ncon(uint32(w)-1), &n1)
gins(a, ncon(uint32(w)-1), &n1)
} else {
gins(a, nr, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
var oldcx gc.Node
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TUINT32], x86.REG_CX)
if gc.GetReg(x86.REG_CX) > 1 && !gc.Samereg(&cx, res) {
gc.Tempname(&oldcx, gc.Types[gc.TUINT32])
gmove(&cx, &oldcx)
}
var n1 gc.Node
var nt gc.Node
if nr.Type.Width > 4 {
gc.Tempname(&nt, nr.Type)
n1 = nt
} else {
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, nr.Type, &n1) // to hold the shift type in CX
}
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)
} else {
gc.Cgen(nr, &n1)
gc.Cgen(nl, &n2)
}
// test and fix up large shifts
if bounded {
if nr.Type.Width > 4 {
// delayed reg alloc
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], &n1) // to hold the shift type in CX
var lo gc.Node
var hi gc.Node
split64(&nt, &lo, &hi)
gmove(&lo, &n1)
splitclean()
}
} else {
var p1 *obj.Prog
if nr.Type.Width > 4 {
// delayed reg alloc
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], &n1) // to hold the shift type in CX
var lo gc.Node
var hi gc.Node
split64(&nt, &lo, &hi)
gmove(&lo, &n1)
gins(optoas(gc.OCMP, gc.Types[gc.TUINT32]), &hi, ncon(0))
p2 := gc.Gbranch(optoas(gc.ONE, gc.Types[gc.TUINT32]), nil, +1)
gins(optoas(gc.OCMP, gc.Types[gc.TUINT32]), &n1, ncon(uint32(w)))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
splitclean()
gc.Patch(p2, gc.Pc)
} else {
gins(optoas(gc.OCMP, nr.Type), &n1, ncon(uint32(w)))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
}
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gins(a, ncon(uint32(w)-1), &n2)
} else {
gmove(ncon(0), &n2)
}
gc.Patch(p1, gc.Pc)
}
gins(a, &n1, &n2)
if oldcx.Op != 0 {
gmove(&oldcx, &cx)
}
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
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--
}
}
/*
* 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(nr.Int())
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 := gc.GetReg(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)
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op gc.Op, 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(nr.Int64())
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
}
// 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]
}
var n1 gc.Node
gc.Regalloc(&n1, nr.Type, nil) // to hold the shift type in CX
var n3 gc.Node
gc.Regalloc(&n3, tcount, &n1) // to clear high bits of CX
var n2 gc.Node
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 {
var rtmp gc.Node
gc.Nodreg(&rtmp, tcount, mips.REGTMP)
gc.Nodconst(&n3, tcount, nl.Type.Width*8)
gins3(mips.ASGTU, &n3, &n1, &rtmp)
p1 := ginsbranch(mips.ABNE, nil, &rtmp, nil, 0)
if op == gc.ORSH && nl.Type.IsSigned() {
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)
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
func ginsnop() {
var reg gc.Node
gc.Nodreg(®, gc.Types[gc.TINT], ppc64.REG_R0)
gins(ppc64.AOR, ®, ®)
}
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)
}
// res = runtime.getg()
func getg(res *gc.Node) {
var n1 gc.Node
gc.Nodreg(&n1, res.Type, ppc64.REGG)
gmove(&n1, res)
}
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
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 := (*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)
}
}