func ginscmp(op gc.Op, t *gc.Type, n1, n2 *gc.Node, likely int) *obj.Prog {
if t.IsInteger() && n1.Op == gc.OLITERAL && n1.Int64() == 0 && n2.Op != gc.OLITERAL {
op = gc.Brrev(op)
n1, n2 = n2, n1
}
var r1, r2, g1, g2 gc.Node
gc.Regalloc(&r1, t, n1)
gc.Regalloc(&g1, n1.Type, &r1)
gc.Cgen(n1, &g1)
gmove(&g1, &r1)
if t.IsInteger() && n2.Op == gc.OLITERAL && n2.Int64() == 0 {
gins(arm.ACMP, &r1, n2)
} else {
gc.Regalloc(&r2, t, n2)
gc.Regalloc(&g2, n1.Type, &r2)
gc.Cgen(n2, &g2)
gmove(&g2, &r2)
gins(optoas(gc.OCMP, t), &r1, &r2)
gc.Regfree(&g2)
gc.Regfree(&r2)
}
gc.Regfree(&g1)
gc.Regfree(&r1)
return gc.Gbranch(optoas(op, t), nil, likely)
}
func ginscmp(op gc.Op, t *gc.Type, n1, n2 *gc.Node, likely int) *obj.Prog {
if t.IsInteger() && n1.Op == gc.OLITERAL && n2.Op != gc.OLITERAL {
// Reverse comparison to place constant last.
op = gc.Brrev(op)
n1, n2 = n2, n1
}
var r1, r2, g1, g2 gc.Node
gc.Regalloc(&r1, t, n1)
gc.Regalloc(&g1, n1.Type, &r1)
gc.Cgen(n1, &g1)
gmove(&g1, &r1)
if t.IsInteger() && gc.Isconst(n2, gc.CTINT) {
ginscon2(optoas(gc.OCMP, t), &r1, n2.Int64())
} else {
gc.Regalloc(&r2, t, n2)
gc.Regalloc(&g2, n1.Type, &r2)
gc.Cgen(n2, &g2)
gmove(&g2, &r2)
gcmp(optoas(gc.OCMP, t), &r1, &r2)
gc.Regfree(&g2)
gc.Regfree(&r2)
}
gc.Regfree(&g1)
gc.Regfree(&r1)
return gc.Gbranch(optoas(op, t), nil, likely)
}
/*
* 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)
}
// gins is called by the front end.
// It synthesizes some multiple-instruction sequences
// so the front end can stay simpler.
func gins(as obj.As, f, t *gc.Node) *obj.Prog {
if as >= obj.A_ARCHSPECIFIC {
if x, ok := f.IntLiteral(); ok {
ginscon(as, x, t)
return nil // caller must not use
}
}
return rawgins(as, f, t)
}
func intLiteral(n *gc.Node) (x int64, ok bool) {
switch {
case n == nil:
return
case gc.Isconst(n, gc.CTINT):
return n.Int64(), true
case gc.Isconst(n, gc.CTBOOL):
return int64(obj.Bool2int(n.Bool())), true
}
return
}
func restx(x *gc.Node, oldx *gc.Node) {
if oldx.Op != 0 {
x.Type = gc.Types[gc.TINT64]
gc.SetReg(int(x.Reg), int(oldx.Etype))
gmove(oldx, x)
gc.Regfree(oldx)
}
}
func restx(x *gc.Node, oldx *gc.Node) {
gc.Regfree(x)
if oldx.Op != 0 {
x.Type = gc.Types[gc.TINT32]
gmove(oldx, x)
}
}
func dotaddable(n *gc.Node, n1 *gc.Node) bool {
if n.Op != gc.ODOT {
return false
}
var oary [10]int64
var nn *gc.Node
o := gc.Dotoffset(n, oary[:], &nn)
if nn != nil && nn.Addable && o == 1 && oary[0] >= 0 {
*n1 = *nn
n1.Type = n.Type
n1.Xoffset += oary[0]
return true
}
return false
}
/*
* 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
// TODO(marvin): Fix Node.EType type union.
oldx.Etype = gc.EType(r) // squirrel away old r value
gc.SetReg(dr, 1)
}
}
/*
* n is a 64-bit value. fill in lo and hi to refer to its 32-bit halves.
*/
func split64(n *gc.Node, lo *gc.Node, hi *gc.Node) {
if !gc.Is64(n.Type) {
gc.Fatalf("split64 %v", n.Type)
}
if nsclean >= len(sclean) {
gc.Fatalf("split64 clean")
}
sclean[nsclean].Op = gc.OEMPTY
nsclean++
switch n.Op {
default:
switch n.Op {
default:
var n1 gc.Node
if !dotaddable(n, &n1) {
gc.Igen(n, &n1, nil)
sclean[nsclean-1] = n1
}
n = &n1
case gc.ONAME, gc.OINDREG:
// nothing
}
*lo = *n
*hi = *n
lo.Type = gc.Types[gc.TUINT32]
if n.Type.Etype == gc.TINT64 {
hi.Type = gc.Types[gc.TINT32]
} else {
hi.Type = gc.Types[gc.TUINT32]
}
hi.Xoffset += 4
case gc.OLITERAL:
var n1 gc.Node
n.Convconst(&n1, n.Type)
i := n1.Int64()
gc.Nodconst(lo, gc.Types[gc.TUINT32], int64(uint32(i)))
i >>= 32
if n.Type.Etype == gc.TINT64 {
gc.Nodconst(hi, gc.Types[gc.TINT32], int64(int32(i)))
} else {
gc.Nodconst(hi, gc.Types[gc.TUINT32], int64(uint32(i)))
}
}
}
func bignodes() {
if bignodes_did {
return
}
bignodes_did = true
gc.Nodconst(&zerof, gc.Types[gc.TINT64], 0)
zerof.Convconst(&zerof, gc.Types[gc.TFLOAT64])
var i big.Int
i.SetInt64(1)
i.Lsh(&i, 63)
var bigi gc.Node
gc.Nodconst(&bigi, gc.Types[gc.TUINT64], 0)
bigi.SetBigInt(&i)
bigi.Convconst(&two63f, gc.Types[gc.TFLOAT64])
gc.Nodconst(&bigi, gc.Types[gc.TUINT64], 0)
i.Lsh(&i, 1)
bigi.SetBigInt(&i)
bigi.Convconst(&two64f, gc.Types[gc.TFLOAT64])
}
/*
* 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) {
// 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 := false
if t.IsSigned() {
check = true
if gc.Isconst(nl, gc.CTINT) && nl.Int64() != -(1<<uint64(t.Width*8-1)) {
check = false
} else if gc.Isconst(nr, gc.CTINT) && nr.Int64() != -1 {
check = false
}
}
if t.Width < 4 {
if t.IsSigned() {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
check = false
}
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 {
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 !t.IsSigned() {
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 {
gc.Patch(p2, gc.Pc)
}
restx(&ax, &oldax)
}
/*
* 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", n.Op, res.Op)
}
switch n.Op {
default:
gc.Fatalf("cgen64 %v", n.Op)
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.Int64())
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.Int64())
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.Int64())
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.Int64())
hv := uint32(hi2.Int64())
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 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 && odst < 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--
}
}
}
/*
* 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, gc.FmtLong), gc.Nconv(t, gc.FmtLong))
}
ft := int(gc.Simsimtype(f.Type))
tt := int(gc.Simsimtype(t.Type))
cvt := 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 obj.As
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(mips.AMOVV, &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(mips.AMOVV, &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
}
}
// value -> value copy, first operand in memory.
// any floating point operand requires register
// src, so goto hard to copy to register first.
if gc.Ismem(f) && ft != tt && (gc.Isfloat[ft] || gc.Isfloat[tt]) {
cvt = gc.Types[ft]
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, gc.FmtLong), gc.Tconv(t.Type, gc.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 = mips.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 = mips.AMOVBU
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 = mips.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 = mips.AMOVHU
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 = mips.AMOVW
case gc.TINT32<<16 | gc.TUINT32, // same size
gc.TUINT32<<16 | gc.TUINT32,
gc.TINT64<<16 | gc.TUINT32, // truncate
gc.TUINT64<<16 | gc.TUINT32:
a = mips.AMOVWU
case gc.TINT64<<16 | gc.TINT64, // same size
gc.TINT64<<16 | gc.TUINT64,
gc.TUINT64<<16 | gc.TINT64,
gc.TUINT64<<16 | gc.TUINT64:
a = mips.AMOVV
/*
* 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 = mips.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 = mips.AMOVBU
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 = mips.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 = mips.AMOVHU
goto rdst
case gc.TINT32<<16 | gc.TINT64, // sign extend int32
gc.TINT32<<16 | gc.TUINT64:
a = mips.AMOVW
goto rdst
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
a = mips.AMOVWU
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()
gc.Regalloc(&r1, gc.Types[gc.TFLOAT64], nil)
gmove(f, &r1)
if tt == gc.TUINT64 {
gc.Regalloc(&r2, gc.Types[gc.TFLOAT64], nil)
gmove(&bigf, &r2)
gins3(mips.ACMPGED, &r1, &r2, nil)
p1 := gc.Gbranch(mips.ABFPF, nil, 0)
gins(mips.ASUBD, &r2, &r1)
gc.Patch(p1, gc.Pc)
gc.Regfree(&r2)
}
gc.Regalloc(&r2, gc.Types[gc.TINT64], t)
gins(mips.ATRUNCDV, &r1, &r1)
gins(mips.AMOVV, &r1, &r2)
gc.Regfree(&r1)
if tt == gc.TUINT64 {
p1 := gc.Gbranch(mips.ABFPF, nil, 0) // use FCR0 here again
gc.Nodreg(&r1, gc.Types[gc.TINT64], mips.REGTMP)
gmove(&bigi, &r1)
gins(mips.AADDVU, &r1, &r2)
gc.Patch(p1, gc.Pc)
}
gmove(&r2, t)
gc.Regfree(&r2)
return
//warn("gmove: convert int to float not implemented: %N -> %N\n", f, t);
//return;
// algorithm is:
// if small enough, use native int64 -> float64 conversion.
// otherwise, halve (x -> (x>>1)|(x&1)), 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 rtmp gc.Node
gc.Regalloc(&r1, gc.Types[gc.TINT64], nil)
gmove(f, &r1)
if ft == gc.TUINT64 {
gc.Nodreg(&rtmp, gc.Types[gc.TUINT64], mips.REGTMP)
gmove(&bigi, &rtmp)
gins(mips.AAND, &r1, &rtmp)
p1 := ginsbranch(mips.ABEQ, nil, &rtmp, nil, 0)
var r3 gc.Node
gc.Regalloc(&r3, gc.Types[gc.TUINT64], nil)
p2 := gins3(mips.AAND, nil, &r1, &r3)
p2.From.Type = obj.TYPE_CONST
p2.From.Offset = 1
p3 := gins(mips.ASRLV, nil, &r1)
p3.From.Type = obj.TYPE_CONST
p3.From.Offset = 1
gins(mips.AOR, &r3, &r1)
gc.Regfree(&r3)
gc.Patch(p1, gc.Pc)
}
gc.Regalloc(&r2, gc.Types[gc.TFLOAT64], t)
gins(mips.AMOVV, &r1, &r2)
gins(mips.AMOVVD, &r2, &r2)
gc.Regfree(&r1)
if ft == gc.TUINT64 {
p1 := ginsbranch(mips.ABEQ, nil, &rtmp, nil, 0)
gc.Nodreg(&r1, gc.Types[gc.TFLOAT64], mips.FREGTWO)
gins(mips.AMULD, &r1, &r2)
gc.Patch(p1, gc.Pc)
}
gmove(&r2, t)
gc.Regfree(&r2)
return
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32:
a = mips.AMOVF
case gc.TFLOAT64<<16 | gc.TFLOAT64:
a = mips.AMOVD
case gc.TFLOAT32<<16 | gc.TFLOAT64:
a = mips.AMOVFD
goto rdst
case gc.TFLOAT64<<16 | gc.TFLOAT32:
a = mips.AMOVDF
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
}
/*
* 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)
if op == gc.OLROT {
v := nr.Int64()
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
if w == 32 {
gc.Cgen(nl, &n1)
gshift(arm.AMOVW, &n1, arm.SHIFT_RR, int32(w)-int32(v), &n1)
} else {
var n2 gc.Node
gc.Regalloc(&n2, nl.Type, nil)
gc.Cgen(nl, &n2)
gshift(arm.AMOVW, &n2, arm.SHIFT_LL, int32(v), &n1)
gshift(arm.AORR, &n2, arm.SHIFT_LR, int32(w)-int32(v), &n1)
gc.Regfree(&n2)
// Ensure sign/zero-extended result.
gins(optoas(gc.OAS, nl.Type), &n1, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
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 == 0 {
} else // nothing to do
if sc >= uint64(nl.Type.Width*8) {
if op == gc.ORSH && nl.Type.IsSigned() {
gshift(arm.AMOVW, &n1, arm.SHIFT_AR, int32(w), &n1)
} else {
gins(arm.AEOR, &n1, &n1)
}
} else {
if op == gc.ORSH && nl.Type.IsSigned() {
gshift(arm.AMOVW, &n1, arm.SHIFT_AR, int32(sc), &n1)
} else if op == gc.ORSH {
gshift(arm.AMOVW, &n1, arm.SHIFT_LR, int32(sc), &n1) // OLSH
} else {
gshift(arm.AMOVW, &n1, arm.SHIFT_LL, int32(sc), &n1)
}
}
if w < 32 && op == gc.OLSH {
gins(optoas(gc.OAS, nl.Type), &n1, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
tr := nr.Type
var t gc.Node
var n1 gc.Node
var n2 gc.Node
var n3 gc.Node
if tr.Width > 4 {
var nt gc.Node
gc.Tempname(&nt, nr.Type)
if nl.Ullman >= nr.Ullman {
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
gc.Cgen(nr, &nt)
n1 = nt
} else {
gc.Cgen(nr, &nt)
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
}
var hi gc.Node
var lo gc.Node
split64(&nt, &lo, &hi)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], nil)
gc.Regalloc(&n3, gc.Types[gc.TUINT32], nil)
gmove(&lo, &n1)
gmove(&hi, &n3)
splitclean()
gins(arm.ATST, &n3, nil)
gc.Nodconst(&t, gc.Types[gc.TUINT32], int64(w))
p1 := gins(arm.AMOVW, &t, &n1)
p1.Scond = arm.C_SCOND_NE
tr = gc.Types[gc.TUINT32]
gc.Regfree(&n3)
} else {
if nl.Ullman >= nr.Ullman {
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
gc.Regalloc(&n1, nr.Type, nil)
gc.Cgen(nr, &n1)
} else {
gc.Regalloc(&n1, nr.Type, nil)
gc.Cgen(nr, &n1)
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
}
}
// test for shift being 0
gins(arm.ATST, &n1, nil)
p3 := gc.Gbranch(arm.ABEQ, nil, -1)
// test and fix up large shifts
// TODO: if(!bounded), don't emit some of this.
gc.Regalloc(&n3, tr, nil)
gc.Nodconst(&t, gc.Types[gc.TUINT32], int64(w))
gmove(&t, &n3)
gins(arm.ACMP, &n1, &n3)
if op == gc.ORSH {
var p1 *obj.Prog
var p2 *obj.Prog
if nl.Type.IsSigned() {
p1 = gshift(arm.AMOVW, &n2, arm.SHIFT_AR, int32(w)-1, &n2)
p2 = gregshift(arm.AMOVW, &n2, arm.SHIFT_AR, &n1, &n2)
} else {
p1 = gins(arm.AEOR, &n2, &n2)
p2 = gregshift(arm.AMOVW, &n2, arm.SHIFT_LR, &n1, &n2)
}
p1.Scond = arm.C_SCOND_HS
p2.Scond = arm.C_SCOND_LO
} else {
p1 := gins(arm.AEOR, &n2, &n2)
p2 := gregshift(arm.AMOVW, &n2, arm.SHIFT_LL, &n1, &n2)
p1.Scond = arm.C_SCOND_HS
p2.Scond = arm.C_SCOND_LO
}
gc.Regfree(&n3)
gc.Patch(p3, gc.Pc)
// Left-shift of smaller word must be sign/zero-extended.
if w < 32 && op == gc.OLSH {
gins(optoas(gc.OAS, nl.Type), &n2, &n2)
}
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 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 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
}
// cannot have two memory operands;
// except 64-bit, which always copies via registers anyway.
var a obj.As
var r1 gc.Node
if !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
switch tt {
default:
f.Convconst(&con, t.Type)
case gc.TINT16,
gc.TINT8:
var con gc.Node
f.Convconst(&con, gc.Types[gc.TINT32])
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(arm.AMOVW, &con, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
case gc.TUINT16,
gc.TUINT8:
var con gc.Node
f.Convconst(&con, gc.Types[gc.TUINT32])
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(arm.AMOVW, &con, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
f = &con
ft = gc.Simsimtype(con.Type)
// constants can't move directly to memory
if gc.Ismem(t) && !gc.Is64(t.Type) {
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:
// should not happen
gc.Fatalf("gmove %v -> %v", f, t)
return
/*
* integer copy and truncate
*/
case gc.TINT8<<16 | gc.TINT8: // same size
if !gc.Ismem(f) {
a = arm.AMOVB
break
}
fallthrough
case 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:
a = arm.AMOVBS
case gc.TUINT8<<16 | gc.TUINT8:
if !gc.Ismem(f) {
a = arm.AMOVB
break
}
fallthrough
case gc.TINT8<<16 | gc.TUINT8,
gc.TINT16<<16 | gc.TUINT8,
gc.TUINT16<<16 | gc.TUINT8,
gc.TINT32<<16 | gc.TUINT8,
gc.TUINT32<<16 | gc.TUINT8:
a = arm.AMOVBU
case gc.TINT64<<16 | gc.TINT8, // truncate low word
gc.TUINT64<<16 | gc.TINT8:
a = arm.AMOVBS
goto trunc64
case gc.TINT64<<16 | gc.TUINT8,
gc.TUINT64<<16 | gc.TUINT8:
a = arm.AMOVBU
goto trunc64
case gc.TINT16<<16 | gc.TINT16: // same size
if !gc.Ismem(f) {
a = arm.AMOVH
break
}
fallthrough
case gc.TUINT16<<16 | gc.TINT16,
gc.TINT32<<16 | gc.TINT16, // truncate
gc.TUINT32<<16 | gc.TINT16:
a = arm.AMOVHS
case gc.TUINT16<<16 | gc.TUINT16:
if !gc.Ismem(f) {
a = arm.AMOVH
break
}
fallthrough
case gc.TINT16<<16 | gc.TUINT16,
gc.TINT32<<16 | gc.TUINT16,
gc.TUINT32<<16 | gc.TUINT16:
a = arm.AMOVHU
case gc.TINT64<<16 | gc.TINT16, // truncate low word
gc.TUINT64<<16 | gc.TINT16:
a = arm.AMOVHS
goto trunc64
case gc.TINT64<<16 | gc.TUINT16,
gc.TUINT64<<16 | gc.TUINT16:
a = arm.AMOVHU
goto trunc64
case gc.TINT32<<16 | gc.TINT32, // same size
gc.TINT32<<16 | gc.TUINT32,
gc.TUINT32<<16 | gc.TINT32,
gc.TUINT32<<16 | gc.TUINT32:
a = arm.AMOVW
case gc.TINT64<<16 | gc.TINT32, // truncate
gc.TUINT64<<16 | gc.TINT32,
gc.TINT64<<16 | gc.TUINT32,
gc.TUINT64<<16 | gc.TUINT32:
var flo gc.Node
var fhi gc.Node
split64(f, &flo, &fhi)
var r1 gc.Node
gc.Regalloc(&r1, t.Type, nil)
gins(arm.AMOVW, &flo, &r1)
gins(arm.AMOVW, &r1, t)
gc.Regfree(&r1)
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)
var r1 gc.Node
gc.Regalloc(&r1, flo.Type, nil)
var r2 gc.Node
gc.Regalloc(&r2, fhi.Type, nil)
gins(arm.AMOVW, &flo, &r1)
gins(arm.AMOVW, &fhi, &r2)
gins(arm.AMOVW, &r1, &tlo)
gins(arm.AMOVW, &r2, &thi)
gc.Regfree(&r1)
gc.Regfree(&r2)
splitclean()
splitclean()
return
/*
* 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:
a = arm.AMOVBS
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,
gc.TUINT8<<16 | gc.TINT32,
gc.TUINT8<<16 | gc.TUINT32:
a = arm.AMOVBU
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 = arm.AMOVHS
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 = arm.AMOVHU
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 tlo gc.Node
var thi gc.Node
split64(t, &tlo, &thi)
var r1 gc.Node
gc.Regalloc(&r1, tlo.Type, nil)
var r2 gc.Node
gc.Regalloc(&r2, thi.Type, nil)
gmove(f, &r1)
p1 := gins(arm.AMOVW, &r1, &r2)
p1.From.Type = obj.TYPE_SHIFT
p1.From.Offset = 2<<5 | 31<<7 | int64(r1.Reg)&15 // r1->31
p1.From.Reg = 0
//print("gmove: %v\n", p1);
gins(arm.AMOVW, &r1, &tlo)
gins(arm.AMOVW, &r2, &thi)
gc.Regfree(&r1)
gc.Regfree(&r2)
splitclean()
return
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
var thi gc.Node
var tlo gc.Node
split64(t, &tlo, &thi)
gmove(f, &tlo)
var r1 gc.Node
gc.Regalloc(&r1, thi.Type, nil)
gins(arm.AMOVW, ncon(0), &r1)
gins(arm.AMOVW, &r1, &thi)
gc.Regfree(&r1)
splitclean()
return
// case CASE(TFLOAT64, TUINT64):
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT8,
gc.TFLOAT32<<16 | gc.TUINT8,
gc.TFLOAT32<<16 | gc.TINT16,
gc.TFLOAT32<<16 | gc.TUINT16,
gc.TFLOAT32<<16 | gc.TINT32,
gc.TFLOAT32<<16 | gc.TUINT32,
// case CASE(TFLOAT32, TUINT64):
gc.TFLOAT64<<16 | gc.TINT8,
gc.TFLOAT64<<16 | gc.TUINT8,
gc.TFLOAT64<<16 | gc.TINT16,
gc.TFLOAT64<<16 | gc.TUINT16,
gc.TFLOAT64<<16 | gc.TINT32,
gc.TFLOAT64<<16 | gc.TUINT32:
fa := arm.AMOVF
a := arm.AMOVFW
if ft == gc.TFLOAT64 {
fa = arm.AMOVD
a = arm.AMOVDW
}
ta := arm.AMOVW
switch tt {
case gc.TINT8:
ta = arm.AMOVBS
case gc.TUINT8:
ta = arm.AMOVBU
case gc.TINT16:
ta = arm.AMOVHS
case gc.TUINT16:
ta = arm.AMOVHU
}
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[ft], f)
var r2 gc.Node
gc.Regalloc(&r2, gc.Types[tt], t)
gins(fa, f, &r1) // load to fpu
p1 := gins(a, &r1, &r1) // convert to w
switch tt {
case gc.TUINT8,
gc.TUINT16,
gc.TUINT32:
p1.Scond |= arm.C_UBIT
}
gins(arm.AMOVW, &r1, &r2) // copy to cpu
gins(ta, &r2, t) // store
gc.Regfree(&r1)
gc.Regfree(&r2)
return
/*
* integer to float
*/
case gc.TINT8<<16 | gc.TFLOAT32,
gc.TUINT8<<16 | gc.TFLOAT32,
gc.TINT16<<16 | gc.TFLOAT32,
gc.TUINT16<<16 | gc.TFLOAT32,
gc.TINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TUINT8<<16 | gc.TFLOAT64,
gc.TINT16<<16 | gc.TFLOAT64,
gc.TUINT16<<16 | gc.TFLOAT64,
gc.TINT32<<16 | gc.TFLOAT64,
gc.TUINT32<<16 | gc.TFLOAT64:
fa := arm.AMOVW
switch ft {
case gc.TINT8:
fa = arm.AMOVBS
case gc.TUINT8:
fa = arm.AMOVBU
case gc.TINT16:
fa = arm.AMOVHS
case gc.TUINT16:
fa = arm.AMOVHU
}
a := arm.AMOVWF
ta := arm.AMOVF
if tt == gc.TFLOAT64 {
a = arm.AMOVWD
ta = arm.AMOVD
}
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[ft], f)
var r2 gc.Node
gc.Regalloc(&r2, gc.Types[tt], t)
gins(fa, f, &r1) // load to cpu
gins(arm.AMOVW, &r1, &r2) // copy to fpu
p1 := gins(a, &r2, &r2) // convert
switch ft {
case gc.TUINT8,
gc.TUINT16,
gc.TUINT32:
p1.Scond |= arm.C_UBIT
}
gins(ta, &r2, t) // store
gc.Regfree(&r1)
gc.Regfree(&r2)
return
case gc.TUINT64<<16 | gc.TFLOAT32,
gc.TUINT64<<16 | gc.TFLOAT64:
gc.Fatalf("gmove UINT64, TFLOAT not implemented")
return
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32:
a = arm.AMOVF
case gc.TFLOAT64<<16 | gc.TFLOAT64:
a = arm.AMOVD
case gc.TFLOAT32<<16 | gc.TFLOAT64:
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[gc.TFLOAT64], t)
gins(arm.AMOVF, f, &r1)
gins(arm.AMOVFD, &r1, &r1)
gins(arm.AMOVD, &r1, t)
gc.Regfree(&r1)
return
case gc.TFLOAT64<<16 | gc.TFLOAT32:
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[gc.TFLOAT64], t)
gins(arm.AMOVD, f, &r1)
gins(arm.AMOVDF, &r1, &r1)
gins(arm.AMOVF, &r1, t)
gc.Regfree(&r1)
return
}
gins(a, f, t)
return
// TODO(kaib): we almost always require a register dest anyway, this can probably be
// removed.
// 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
// truncate 64 bit integer
trunc64:
var fhi gc.Node
var flo gc.Node
split64(f, &flo, &fhi)
gc.Regalloc(&r1, t.Type, nil)
gins(a, &flo, &r1)
gins(a, &r1, t)
gc.Regfree(&r1)
splitclean()
return
}
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 obj.As
switch align {
default:
gc.Fatalf("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.Fatalf("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.Fatalf("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 && odst < 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 ; c > 0; c-- {
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)
}
/*
* generate code to compute address of n,
* a reference to a (perhaps nested) field inside
* an array or struct.
* return 0 on failure, 1 on success.
* on success, leaves usable address in a.
*
* caller is responsible for calling sudoclean
* after successful sudoaddable,
* to release the register used for a.
*/
func sudoaddable(as obj.As, n *gc.Node, a *obj.Addr) bool {
if n.Type == nil {
return false
}
*a = obj.Addr{}
switch n.Op {
case gc.OLITERAL:
if !gc.Isconst(n, gc.CTINT) {
break
}
v := n.Int64()
if v >= 32000 || v <= -32000 {
break
}
switch as {
default:
return false
case x86.AADDB,
x86.AADDW,
x86.AADDL,
x86.AADDQ,
x86.ASUBB,
x86.ASUBW,
x86.ASUBL,
x86.ASUBQ,
x86.AANDB,
x86.AANDW,
x86.AANDL,
x86.AANDQ,
x86.AORB,
x86.AORW,
x86.AORL,
x86.AORQ,
x86.AXORB,
x86.AXORW,
x86.AXORL,
x86.AXORQ,
x86.AINCB,
x86.AINCW,
x86.AINCL,
x86.AINCQ,
x86.ADECB,
x86.ADECW,
x86.ADECL,
x86.ADECQ,
x86.AMOVB,
x86.AMOVW,
x86.AMOVL,
x86.AMOVQ:
break
}
cleani += 2
reg := &clean[cleani-1]
reg1 := &clean[cleani-2]
reg.Op = gc.OEMPTY
reg1.Op = gc.OEMPTY
gc.Naddr(a, n)
return true
case gc.ODOT,
gc.ODOTPTR:
cleani += 2
reg := &clean[cleani-1]
reg1 := &clean[cleani-2]
reg.Op = gc.OEMPTY
reg1.Op = gc.OEMPTY
var nn *gc.Node
var oary [10]int64
o := gc.Dotoffset(n, oary[:], &nn)
if nn == nil {
sudoclean()
return false
}
if nn.Addable && o == 1 && oary[0] >= 0 {
// directly addressable set of DOTs
n1 := *nn
n1.Type = n.Type
n1.Xoffset += oary[0]
gc.Naddr(a, &n1)
return true
}
gc.Regalloc(reg, gc.Types[gc.Tptr], nil)
n1 := *reg
n1.Op = gc.OINDREG
if oary[0] >= 0 {
gc.Agen(nn, reg)
n1.Xoffset = oary[0]
} else {
gc.Cgen(nn, reg)
gc.Cgen_checknil(reg)
n1.Xoffset = -(oary[0] + 1)
}
for i := 1; i < o; i++ {
if oary[i] >= 0 {
gc.Fatalf("can't happen")
}
gins(movptr, &n1, reg)
gc.Cgen_checknil(reg)
n1.Xoffset = -(oary[i] + 1)
}
a.Type = obj.TYPE_NONE
a.Index = x86.REG_NONE
gc.Fixlargeoffset(&n1)
gc.Naddr(a, &n1)
return true
case gc.OINDEX:
return false
}
return false
}
/*
* 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) {
// 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 t.IsSigned() {
check = 1
if gc.Isconst(nl, gc.CTINT) && nl.Int64() != -(1<<uint64(t.Width*8-1)) {
check = 0
} else if gc.Isconst(nr, gc.CTINT) && nr.Int64() != -1 {
check = 0
}
}
if t.Width < 8 {
if t.IsSigned() {
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 = s390x.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.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)
}
}
/*
* 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 gc.Op, 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 := false
if t.IsSigned() {
check = true
if gc.Isconst(nl, gc.CTINT) && nl.Int64() != -1<<uint64(t.Width*8-1) {
check = false
} else if gc.Isconst(nr, gc.CTINT) && nr.Int64() != -1 {
check = false
}
}
if t.Width < 4 {
if t.IsSigned() {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
check = false
}
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 {
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 !t.IsSigned() {
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 {
gc.Patch(p2, gc.Pc)
}
}
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 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, gc.FmtLong), gc.Nconv(t, gc.FmtLong))
}
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
}
// cannot have two memory operands
var a obj.As
if 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 = tt // so big switch will choose a simple mov
// some constants can't move directly to memory.
if gc.Ismem(t) {
// float constants come from memory.
if gc.Isfloat[tt] {
goto hard
}
// 64-bit immediates are really 32-bit sign-extended
// unless moving into a register.
if gc.Isint[tt] {
if i := con.Int64(); int64(int32(i)) != i {
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.Dump("f", f)
gc.Dump("t", t)
gc.Fatalf("gmove %v -> %v", gc.Tconv(f.Type, gc.FmtLong), gc.Tconv(t.Type, gc.FmtLong))
/*
* 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,
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,
gc.TINT16<<16 | gc.TUINT8,
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 = x86.AMOVB
case gc.TINT16<<16 | gc.TINT16, // same size
gc.TINT16<<16 | gc.TUINT16,
gc.TUINT16<<16 | gc.TINT16,
gc.TUINT16<<16 | gc.TUINT16,
gc.TINT32<<16 | gc.TINT16,
// truncate
gc.TUINT32<<16 | gc.TINT16,
gc.TINT64<<16 | gc.TINT16,
gc.TUINT64<<16 | gc.TINT16,
gc.TINT32<<16 | gc.TUINT16,
gc.TUINT32<<16 | gc.TUINT16,
gc.TINT64<<16 | gc.TUINT16,
gc.TUINT64<<16 | gc.TUINT16:
a = x86.AMOVW
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:
a = x86.AMOVQL
case gc.TINT64<<16 | gc.TINT64, // same size
gc.TINT64<<16 | gc.TUINT64,
gc.TUINT64<<16 | gc.TINT64,
gc.TUINT64<<16 | gc.TUINT64:
a = x86.AMOVQ
/*
* 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,
gc.TINT8<<16 | gc.TUINT64:
a = x86.AMOVBQSX
goto rdst
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,
gc.TUINT8<<16 | gc.TUINT64:
a = x86.AMOVBQZX
goto rdst
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,
gc.TINT16<<16 | gc.TUINT64:
a = x86.AMOVWQSX
goto rdst
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,
gc.TUINT16<<16 | gc.TUINT64:
a = x86.AMOVWQZX
goto rdst
case gc.TINT32<<16 | gc.TINT64, // sign extend int32
gc.TINT32<<16 | gc.TUINT64:
a = x86.AMOVLQSX
goto rdst
// AMOVL into a register zeros the top of the register,
// so this is not always necessary, but if we rely on AMOVL
// the optimizer is almost certain to screw with us.
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
a = x86.AMOVLQZX
goto rdst
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT32:
a = x86.ACVTTSS2SL
goto rdst
case gc.TFLOAT64<<16 | gc.TINT32:
a = x86.ACVTTSD2SL
goto rdst
case gc.TFLOAT32<<16 | gc.TINT64:
a = x86.ACVTTSS2SQ
goto rdst
case gc.TFLOAT64<<16 | gc.TINT64:
a = x86.ACVTTSD2SQ
goto rdst
// convert via int32.
case 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:
cvt = gc.Types[gc.TINT32]
goto hard
// convert via int64.
case gc.TFLOAT32<<16 | gc.TUINT32,
gc.TFLOAT64<<16 | gc.TUINT32:
cvt = gc.Types[gc.TINT64]
goto hard
// algorithm is:
// if small enough, use native float64 -> int64 conversion.
// otherwise, subtract 2^63, convert, and add it back.
case gc.TFLOAT32<<16 | gc.TUINT64,
gc.TFLOAT64<<16 | gc.TUINT64:
a := x86.ACVTTSS2SQ
if ft == gc.TFLOAT64 {
a = x86.ACVTTSD2SQ
}
bignodes()
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[ft], nil)
var r2 gc.Node
gc.Regalloc(&r2, gc.Types[tt], t)
var r3 gc.Node
gc.Regalloc(&r3, gc.Types[ft], nil)
var r4 gc.Node
gc.Regalloc(&r4, gc.Types[tt], nil)
gins(optoas(gc.OAS, f.Type), f, &r1)
gins(optoas(gc.OCMP, f.Type), &bigf, &r1)
p1 := gc.Gbranch(optoas(gc.OLE, f.Type), nil, +1)
gins(a, &r1, &r2)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
gins(optoas(gc.OAS, f.Type), &bigf, &r3)
gins(optoas(gc.OSUB, f.Type), &r3, &r1)
gins(a, &r1, &r2)
gins(x86.AMOVQ, &bigi, &r4)
gins(x86.AXORQ, &r4, &r2)
gc.Patch(p2, gc.Pc)
gmove(&r2, t)
gc.Regfree(&r4)
gc.Regfree(&r3)
gc.Regfree(&r2)
gc.Regfree(&r1)
return
/*
* integer to float
*/
case gc.TINT32<<16 | gc.TFLOAT32:
a = x86.ACVTSL2SS
goto rdst
case gc.TINT32<<16 | gc.TFLOAT64:
a = x86.ACVTSL2SD
goto rdst
case gc.TINT64<<16 | gc.TFLOAT32:
a = x86.ACVTSQ2SS
goto rdst
case gc.TINT64<<16 | gc.TFLOAT64:
a = x86.ACVTSQ2SD
goto rdst
// convert via int32
case 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:
cvt = gc.Types[gc.TINT32]
goto hard
// convert via int64.
case gc.TUINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT64]
goto hard
// algorithm is:
// if small enough, use native int64 -> uint64 conversion.
// otherwise, halve (rounding to odd?), convert, and double.
case gc.TUINT64<<16 | gc.TFLOAT32,
gc.TUINT64<<16 | gc.TFLOAT64:
a := x86.ACVTSQ2SS
if tt == gc.TFLOAT64 {
a = x86.ACVTSQ2SD
}
var zero gc.Node
gc.Nodconst(&zero, gc.Types[gc.TUINT64], 0)
var one gc.Node
gc.Nodconst(&one, gc.Types[gc.TUINT64], 1)
var r1 gc.Node
gc.Regalloc(&r1, f.Type, f)
var r2 gc.Node
gc.Regalloc(&r2, t.Type, t)
var r3 gc.Node
gc.Regalloc(&r3, f.Type, nil)
var r4 gc.Node
gc.Regalloc(&r4, f.Type, nil)
gmove(f, &r1)
gins(x86.ACMPQ, &r1, &zero)
p1 := gc.Gbranch(x86.AJLT, nil, +1)
gins(a, &r1, &r2)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
gmove(&r1, &r3)
gins(x86.ASHRQ, &one, &r3)
gmove(&r1, &r4)
gins(x86.AANDL, &one, &r4)
gins(x86.AORQ, &r4, &r3)
gins(a, &r3, &r2)
gins(optoas(gc.OADD, t.Type), &r2, &r2)
gc.Patch(p2, gc.Pc)
gmove(&r2, t)
gc.Regfree(&r4)
gc.Regfree(&r3)
gc.Regfree(&r2)
gc.Regfree(&r1)
return
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32:
a = x86.AMOVSS
case gc.TFLOAT64<<16 | gc.TFLOAT64:
a = x86.AMOVSD
case gc.TFLOAT32<<16 | gc.TFLOAT64:
a = x86.ACVTSS2SD
goto rdst
case gc.TFLOAT64<<16 | gc.TFLOAT32:
a = x86.ACVTSD2SS
goto rdst
}
gins(a, f, t)
return
// requires register destination
rdst:
{
var r1 gc.Node
gc.Regalloc(&r1, t.Type, t)
gins(a, f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
// requires register intermediate
hard:
var r1 gc.Node
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
/*
* generate code to compute address of n,
* a reference to a (perhaps nested) field inside
* an array or struct.
* return 0 on failure, 1 on success.
* on success, leaves usable address in a.
*
* caller is responsible for calling sudoclean
* after successful sudoaddable,
* to release the register used for a.
*/
func sudoaddable(as obj.As, n *gc.Node, a *obj.Addr) bool {
if n.Type == nil {
return false
}
*a = obj.Addr{}
switch n.Op {
case gc.OLITERAL:
if !gc.Isconst(n, gc.CTINT) {
break
}
v := n.Int64()
if v >= 32000 || v <= -32000 {
break
}
switch as {
default:
return false
case arm.AADD,
arm.ASUB,
arm.AAND,
arm.AORR,
arm.AEOR,
arm.AMOVB,
arm.AMOVBS,
arm.AMOVBU,
arm.AMOVH,
arm.AMOVHS,
arm.AMOVHU,
arm.AMOVW:
break
}
cleani += 2
reg := &clean[cleani-1]
reg1 := &clean[cleani-2]
reg.Op = gc.OEMPTY
reg1.Op = gc.OEMPTY
gc.Naddr(a, n)
return true
case gc.ODOT,
gc.ODOTPTR:
cleani += 2
reg := &clean[cleani-1]
reg1 := &clean[cleani-2]
reg.Op = gc.OEMPTY
reg1.Op = gc.OEMPTY
var nn *gc.Node
var oary [10]int64
o := gc.Dotoffset(n, oary[:], &nn)
if nn == nil {
sudoclean()
return false
}
if nn.Addable && o == 1 && oary[0] >= 0 {
// directly addressable set of DOTs
n1 := *nn
n1.Type = n.Type
n1.Xoffset += oary[0]
gc.Naddr(a, &n1)
return true
}
gc.Regalloc(reg, gc.Types[gc.Tptr], nil)
n1 := *reg
n1.Op = gc.OINDREG
if oary[0] >= 0 {
gc.Agen(nn, reg)
n1.Xoffset = oary[0]
} else {
gc.Cgen(nn, reg)
gc.Cgen_checknil(reg)
n1.Xoffset = -(oary[0] + 1)
}
for i := 1; i < o; i++ {
if oary[i] >= 0 {
gc.Fatalf("can't happen")
}
gins(arm.AMOVW, &n1, reg)
gc.Cgen_checknil(reg)
n1.Xoffset = -(oary[i] + 1)
}
a.Type = obj.TYPE_NONE
a.Name = obj.NAME_NONE
n1.Type = n.Type
gc.Naddr(a, &n1)
return true
case gc.OINDEX:
return false
}
return false
}
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)
}
// 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)
}
func clearfat_tail(n1 *gc.Node, b int64) {
if b >= 16 && isPlan9 {
var z gc.Node
gc.Nodconst(&z, gc.Types[gc.TUINT64], 0)
q := b / 8
for ; q > 0; q-- {
n1.Type = z.Type
gins(x86.AMOVQ, &z, n1)
n1.Xoffset += 8
b -= 8
}
if b != 0 {
n1.Xoffset -= 8 - b
gins(x86.AMOVQ, &z, n1)
}
return
}
if b >= 16 {
var vec_zero gc.Node
gc.Regalloc(&vec_zero, gc.Types[gc.TFLOAT64], nil)
gins(x86.AXORPS, &vec_zero, &vec_zero)
for b >= 16 {
gins(x86.AMOVUPS, &vec_zero, n1)
n1.Xoffset += 16
b -= 16
}
// MOVUPS X0, off(base) is a few bytes shorter than MOV 0, off(base)
if b != 0 {
n1.Xoffset -= 16 - b
gins(x86.AMOVUPS, &vec_zero, n1)
}
gc.Regfree(&vec_zero)
return
}
// Write sequence of MOV 0, off(base) instead of using STOSQ.
// The hope is that although the code will be slightly longer,
// the MOVs will have no dependencies and pipeline better
// than the unrolled STOSQ loop.
var z gc.Node
gc.Nodconst(&z, gc.Types[gc.TUINT64], 0)
if b >= 8 {
n1.Type = z.Type
gins(x86.AMOVQ, &z, n1)
n1.Xoffset += 8
b -= 8
if b != 0 {
n1.Xoffset -= 8 - b
gins(x86.AMOVQ, &z, n1)
}
return
}
if b >= 4 {
gc.Nodconst(&z, gc.Types[gc.TUINT32], 0)
n1.Type = z.Type
gins(x86.AMOVL, &z, n1)
n1.Xoffset += 4
b -= 4
if b != 0 {
n1.Xoffset -= 4 - b
gins(x86.AMOVL, &z, n1)
}
return
}
if b >= 2 {
gc.Nodconst(&z, gc.Types[gc.TUINT16], 0)
n1.Type = z.Type
gins(x86.AMOVW, &z, n1)
n1.Xoffset += 2
b -= 2
}
gc.Nodconst(&z, gc.Types[gc.TUINT8], 0)
for b > 0 {
n1.Type = z.Type
gins(x86.AMOVB, &z, n1)
n1.Xoffset++
b--
}
}
/*
* 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 {
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 && 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)
}
/*
* 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.Int64())
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 && nl.Type.IsSigned() {
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)
}