/*
* generate array index into res.
* n might be any size; res is 32-bit.
* returns Prog* to patch to panic call.
*/
func cgenindex(n *gc.Node, res *gc.Node, bounded bool) *obj.Prog {
if !gc.Is64(n.Type) {
gc.Cgen(n, res)
return nil
}
var tmp gc.Node
gc.Tempname(&tmp, gc.Types[gc.TINT64])
gc.Cgen(n, &tmp)
var lo gc.Node
var hi gc.Node
split64(&tmp, &lo, &hi)
gmove(&lo, res)
if bounded {
splitclean()
return nil
}
var n1 gc.Node
gc.Regalloc(&n1, gc.Types[gc.TINT32], nil)
var n2 gc.Node
gc.Regalloc(&n2, gc.Types[gc.TINT32], nil)
var zero gc.Node
gc.Nodconst(&zero, gc.Types[gc.TINT32], 0)
gmove(&hi, &n1)
gmove(&zero, &n2)
gins(arm.ACMP, &n1, &n2)
gc.Regfree(&n2)
gc.Regfree(&n1)
splitclean()
return gc.Gbranch(arm.ABNE, nil, -1)
}
/*
* 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 gc.Op, nl *gc.Node, nr *gc.Node, res *gc.Node) bool {
if optoas(op, nl.Type) != x86.AIMULB {
return false
}
// copy from byte to full registers
t := gc.Types[gc.TUINT32]
if gc.Issigned[nl.Type.Etype] {
t = gc.Types[gc.TINT32]
}
// largest ullman on left.
if nl.Ullman < nr.Ullman {
nl, nr = nr, nl
}
var nt gc.Node
gc.Tempname(&nt, nl.Type)
gc.Cgen(nl, &nt)
var n1 gc.Node
gc.Regalloc(&n1, t, res)
gc.Cgen(nr, &n1)
var n2 gc.Node
gc.Regalloc(&n2, t, nil)
gmove(&nt, &n2)
a := optoas(op, t)
gins(a, &n2, &n1)
gc.Regfree(&n2)
gmove(&n1, res)
gc.Regfree(&n1)
return true
}
/*
* 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 gc.Issigned[t.Etype] {
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)
}
func ginscmp(op gc.Op, t *gc.Type, n1, n2 *gc.Node, likely int) *obj.Prog {
if gc.Isint[t.Etype] && n1.Op == gc.OLITERAL && n1.Int() == 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 gc.Isint[t.Etype] && n2.Op == gc.OLITERAL && n2.Int() == 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 gc.Isint[t.Etype] && 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 gc.Isint[t.Etype] && gc.Isconst(n2, gc.CTINT) {
ginscon2(optoas(gc.OCMP, t), &r1, n2.Int())
} else {
gc.Regalloc(&r2, t, n2)
gc.Regalloc(&g2, n1.Type, &r2)
gc.Cgen(n2, &g2)
gmove(&g2, &r2)
rawgins(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 high multiply
* res = (nl * nr) >> wordsize
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Ullman < nr.Ullman {
nl, nr = nr, nl
}
t := nl.Type
w := int(t.Width * 8)
var n1 gc.Node
gc.Regalloc(&n1, t, res)
gc.Cgen(nl, &n1)
var n2 gc.Node
gc.Regalloc(&n2, t, nil)
gc.Cgen(nr, &n2)
switch gc.Simtype[t.Etype] {
case gc.TINT8,
gc.TINT16:
gins(optoas(gc.OMUL, t), &n2, &n1)
gshift(arm.AMOVW, &n1, arm.SHIFT_AR, int32(w), &n1)
case gc.TUINT8,
gc.TUINT16:
gins(optoas(gc.OMUL, t), &n2, &n1)
gshift(arm.AMOVW, &n1, arm.SHIFT_LR, int32(w), &n1)
// perform a long multiplication.
case gc.TINT32,
gc.TUINT32:
var p *obj.Prog
if gc.Issigned[t.Etype] {
p = gins(arm.AMULL, &n2, nil)
} else {
p = gins(arm.AMULLU, &n2, nil)
}
// n2 * n1 -> (n1 n2)
p.Reg = n1.Reg
p.To.Type = obj.TYPE_REGREG
p.To.Reg = n1.Reg
p.To.Offset = int64(n2.Reg)
default:
gc.Fatalf("cgen_hmul %v", t)
}
gc.Cgen(&n1, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
/*
* generate
* as $c, n
*/
func ginscon(as int, c int64, n2 *gc.Node) {
var n1 gc.Node
switch as {
case x86.AADDL,
x86.AMOVL,
x86.ALEAL:
gc.Nodconst(&n1, gc.Types[gc.TINT32], c)
default:
gc.Nodconst(&n1, gc.Types[gc.TINT64], c)
}
if as != x86.AMOVQ && (c < -(1<<31) || c >= 1<<31) {
// cannot have 64-bit immediate in ADD, etc.
// instead, MOV into register first.
var ntmp gc.Node
gc.Regalloc(&ntmp, gc.Types[gc.TINT64], nil)
gins(x86.AMOVQ, &n1, &ntmp)
gins(as, &ntmp, n2)
gc.Regfree(&ntmp)
return
}
gins(as, &n1, n2)
}
/*
* generate
* as n, $c (CMP/CMPU)
*/
func ginscon2(as int, n2 *gc.Node, c int64) {
var n1 gc.Node
gc.Nodconst(&n1, gc.Types[gc.TINT64], c)
switch as {
default:
gc.Fatalf("ginscon2")
case ppc64.ACMP:
if -ppc64.BIG <= c && c <= ppc64.BIG {
rawgins(as, n2, &n1)
return
}
case ppc64.ACMPU:
if 0 <= c && c <= 2*ppc64.BIG {
rawgins(as, n2, &n1)
return
}
}
// MOV n1 into register first
var ntmp gc.Node
gc.Regalloc(&ntmp, gc.Types[gc.TINT64], nil)
rawgins(ppc64.AMOVD, &n1, &ntmp)
rawgins(as, n2, &ntmp)
gc.Regfree(&ntmp)
}
func ginscmp(op gc.Op, t *gc.Type, n1, n2 *gc.Node, likely int) *obj.Prog {
if gc.Isint[t.Etype] || t.Etype == gc.Tptr {
if (n1.Op == gc.OLITERAL || n1.Op == gc.OADDR && n1.Left.Op == gc.ONAME) && n2.Op != gc.OLITERAL {
// Reverse comparison to place constant (including address constant) last.
op = gc.Brrev(op)
n1, n2 = n2, n1
}
}
// General case.
var r1, r2, g1, g2 gc.Node
// A special case to make write barriers more efficient.
// Comparing the first field of a named struct can be done directly.
base := n1
if n1.Op == gc.ODOT && n1.Left.Type.Etype == gc.TSTRUCT && n1.Left.Type.Type.Sym == n1.Right.Sym {
base = n1.Left
}
if base.Op == gc.ONAME && base.Class&gc.PHEAP == 0 || n1.Op == gc.OINDREG {
r1 = *n1
} else {
gc.Regalloc(&r1, t, n1)
gc.Regalloc(&g1, n1.Type, &r1)
gc.Cgen(n1, &g1)
gmove(&g1, &r1)
}
if n2.Op == gc.OLITERAL && gc.Isint[t.Etype] || n2.Op == gc.OADDR && n2.Left.Op == gc.ONAME && n2.Left.Class == gc.PEXTERN {
r2 = *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)
if r1.Op == gc.OREGISTER {
gc.Regfree(&g1)
gc.Regfree(&r1)
}
if r2.Op == gc.OREGISTER {
gc.Regfree(&g2)
gc.Regfree(&r2)
}
return gc.Gbranch(optoas(op, t), nil, likely)
}
/*
* generate
* as $c, n
*/
func ginscon(as int, c int64, n *gc.Node) {
var n1 gc.Node
gc.Nodconst(&n1, gc.Types[gc.TINT32], c)
var n2 gc.Node
gc.Regalloc(&n2, gc.Types[gc.TINT32], nil)
gmove(&n1, &n2)
gins(as, &n2, n)
gc.Regfree(&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 gc.Op, 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 {
nl, nr = nr, nl
}
// 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
}
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)
}
// res = runtime.getg()
func getg(res *gc.Node) {
var n1 gc.Node
gc.Regalloc(&n1, res.Type, res)
mov := optoas(gc.OAS, gc.Types[gc.Tptr])
p := gins(mov, nil, &n1)
p.From.Type = obj.TYPE_REG
p.From.Reg = x86.REG_TLS
p = gins(mov, nil, &n1)
p.From = p.To
p.From.Type = obj.TYPE_MEM
p.From.Index = x86.REG_TLS
p.From.Scale = 1
gmove(&n1, res)
gc.Regfree(&n1)
}
/*
* 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)
}
}
/*
* generate
* as $c, n
*/
func ginscon(as int, c int64, n2 *gc.Node) {
var n1 gc.Node
gc.Nodconst(&n1, gc.Types[gc.TINT64], c)
if as != ppc64.AMOVD && (c < -ppc64.BIG || c > ppc64.BIG) || n2.Op != gc.OREGISTER || as == ppc64.AMULLD {
// cannot have more than 16-bit of immediate in ADD, etc.
// instead, MOV into register first.
var ntmp gc.Node
gc.Regalloc(&ntmp, gc.Types[gc.TINT64], nil)
rawgins(ppc64.AMOVD, &n1, &ntmp)
rawgins(as, &ntmp, n2)
gc.Regfree(&ntmp)
return
}
rawgins(as, &n1, n2)
}
/*
* 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)
}
/*
* 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 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)
}
}
/*
* 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 := int(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
}
// 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)
gcmp(optoas(gc.OCMP, tcount), &n1, &n3)
p1 := (*obj.Prog)(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)
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
/*
* 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 := false
if gc.Issigned[t.Etype] {
check = true
if gc.Isconst(nl, gc.CTINT) && nl.Int() != -(1<<uint64(t.Width*8-1)) {
check = false
} else if gc.Isconst(nr, gc.CTINT) && nr.Int() != -1 {
check = false
}
}
if t.Width < 8 {
if gc.Issigned[t.Etype] {
t = gc.Types[gc.TINT64]
} else {
t = gc.Types[gc.TUINT64]
}
check = false
}
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.Reg = arm64.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 {
var nm1 gc.Node
gc.Nodconst(&nm1, t, -1)
gcmp(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), &tl, &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 {
gc.Patch(p2, gc.Pc)
}
}
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(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 := (*obj.Prog)(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.Node)(gc.Sysfunc("duffzero"))
p = gins(obj.ADUFFZERO, nil, f)
gc.Afunclit(&p.To, f)
// 8 and 128 = magic constants: see ../../runtime/asm_mips64x.s
p.To.Offset = int64(8 * (128 - q))
// duffzero leaves R1 on the last zeroed dword
boff = 8
} else {
var p *obj.Prog
for t := uint64(0); t < q; t++ {
p = gins(mips.AMOVV, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(8 * t)
}
boff = 8 * q
}
var p *obj.Prog
for t := uint64(0); t < c; t++ {
p = gins(mips.AMOVB, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(t + boff)
}
gc.Regfree(&dst)
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
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.Fatalf("%v in use during clearfat", obj.Rconv(ppc64.REGRT1))
}
var r0 gc.Node
gc.Nodreg(&r0, gc.Types[gc.TUINT64], ppc64.REGZERO)
var dst gc.Node
gc.Nodreg(&dst, gc.Types[gc.Tptr], ppc64.REGRT1)
gc.Regrealloc(&dst)
gc.Agen(nl, &dst)
var boff uint64
if q > 128 {
p := gins(ppc64.ASUB, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p = gins(ppc64.AMOVD, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(q * 8)
p = gins(ppc64.AMOVDU, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 8
pl := (*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)
}
/*
* generate comparison of nl, nr, both 64-bit.
* nl is memory; nr is constant or memory.
*/
func cmp64(nl *gc.Node, nr *gc.Node, op gc.Op, likely int, to *obj.Prog) {
var lo1 gc.Node
var hi1 gc.Node
var lo2 gc.Node
var hi2 gc.Node
var r1 gc.Node
var r2 gc.Node
split64(nl, &lo1, &hi1)
split64(nr, &lo2, &hi2)
// compare most significant word;
// if they differ, we're done.
t := hi1.Type
gc.Regalloc(&r1, gc.Types[gc.TINT32], nil)
gc.Regalloc(&r2, gc.Types[gc.TINT32], nil)
gins(arm.AMOVW, &hi1, &r1)
gins(arm.AMOVW, &hi2, &r2)
gins(arm.ACMP, &r1, &r2)
gc.Regfree(&r1)
gc.Regfree(&r2)
var br *obj.Prog
switch op {
default:
gc.Fatalf("cmp64 %v %v", gc.Oconv(int(op), 0), t)
// cmp hi
// bne L
// cmp lo
// beq to
// L:
case gc.OEQ:
br = gc.Gbranch(arm.ABNE, nil, -likely)
// cmp hi
// bne to
// cmp lo
// bne to
case gc.ONE:
gc.Patch(gc.Gbranch(arm.ABNE, nil, likely), to)
// cmp hi
// bgt to
// blt L
// cmp lo
// bge to (or bgt to)
// L:
case gc.OGE,
gc.OGT:
gc.Patch(gc.Gbranch(optoas(gc.OGT, t), nil, likely), to)
br = gc.Gbranch(optoas(gc.OLT, t), nil, -likely)
// cmp hi
// blt to
// bgt L
// cmp lo
// ble to (or jlt to)
// L:
case gc.OLE,
gc.OLT:
gc.Patch(gc.Gbranch(optoas(gc.OLT, t), nil, likely), to)
br = gc.Gbranch(optoas(gc.OGT, t), nil, -likely)
}
// compare least significant word
t = lo1.Type
gc.Regalloc(&r1, gc.Types[gc.TINT32], nil)
gc.Regalloc(&r2, gc.Types[gc.TINT32], nil)
gins(arm.AMOVW, &lo1, &r1)
gins(arm.AMOVW, &lo2, &r2)
gins(arm.ACMP, &r1, &r2)
gc.Regfree(&r1)
gc.Regfree(&r2)
// jump again
gc.Patch(gc.Gbranch(optoas(op, t), nil, likely), to)
// point first branch down here if appropriate
if br != nil {
gc.Patch(br, gc.Pc)
}
splitclean()
splitclean()
}
/*
* 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))
}
l := n.Left
var t1 gc.Node
if !l.Addable {
gc.Tempname(&t1, l.Type)
gc.Cgen(l, &t1)
l = &t1
}
var hi1 gc.Node
var lo1 gc.Node
split64(l, &lo1, &hi1)
switch n.Op {
default:
gc.Fatalf("cgen64 %v", gc.Oconv(int(n.Op), 0))
case gc.OMINUS:
var lo2 gc.Node
var hi2 gc.Node
split64(res, &lo2, &hi2)
gc.Regalloc(&t1, lo1.Type, nil)
var al gc.Node
gc.Regalloc(&al, lo1.Type, nil)
var ah gc.Node
gc.Regalloc(&ah, hi1.Type, nil)
gins(arm.AMOVW, &lo1, &al)
gins(arm.AMOVW, &hi1, &ah)
gmove(ncon(0), &t1)
p1 := gins(arm.ASUB, &al, &t1)
p1.Scond |= arm.C_SBIT
gins(arm.AMOVW, &t1, &lo2)
gmove(ncon(0), &t1)
gins(arm.ASBC, &ah, &t1)
gins(arm.AMOVW, &t1, &hi2)
gc.Regfree(&t1)
gc.Regfree(&al)
gc.Regfree(&ah)
splitclean()
splitclean()
return
case gc.OCOM:
gc.Regalloc(&t1, lo1.Type, nil)
gmove(ncon(^uint32(0)), &t1)
var lo2 gc.Node
var hi2 gc.Node
split64(res, &lo2, &hi2)
var n1 gc.Node
gc.Regalloc(&n1, lo1.Type, nil)
gins(arm.AMOVW, &lo1, &n1)
gins(arm.AEOR, &t1, &n1)
gins(arm.AMOVW, &n1, &lo2)
gins(arm.AMOVW, &hi1, &n1)
gins(arm.AEOR, &t1, &n1)
gins(arm.AMOVW, &n1, &hi2)
gc.Regfree(&t1)
gc.Regfree(&n1)
splitclean()
splitclean()
return
// binary operators.
// common setup below.
case gc.OADD,
gc.OSUB,
gc.OMUL,
gc.OLSH,
gc.ORSH,
gc.OAND,
gc.OOR,
gc.OXOR,
gc.OLROT:
break
}
// setup for binary operators
r := n.Right
if r != nil && !r.Addable {
var t2 gc.Node
gc.Tempname(&t2, r.Type)
gc.Cgen(r, &t2)
r = &t2
}
var hi2 gc.Node
var lo2 gc.Node
if gc.Is64(r.Type) {
split64(r, &lo2, &hi2)
}
var al gc.Node
gc.Regalloc(&al, lo1.Type, nil)
var ah gc.Node
gc.Regalloc(&ah, hi1.Type, nil)
// Do op. Leave result in ah:al.
switch n.Op {
default:
gc.Fatalf("cgen64: not implemented: %v\n", n)
// TODO: Constants
case gc.OADD:
var bl gc.Node
gc.Regalloc(&bl, gc.Types[gc.TPTR32], nil)
var bh gc.Node
gc.Regalloc(&bh, gc.Types[gc.TPTR32], nil)
gins(arm.AMOVW, &hi1, &ah)
gins(arm.AMOVW, &lo1, &al)
gins(arm.AMOVW, &hi2, &bh)
gins(arm.AMOVW, &lo2, &bl)
p1 := gins(arm.AADD, &bl, &al)
p1.Scond |= arm.C_SBIT
gins(arm.AADC, &bh, &ah)
gc.Regfree(&bl)
gc.Regfree(&bh)
// TODO: Constants.
case gc.OSUB:
var bl gc.Node
gc.Regalloc(&bl, gc.Types[gc.TPTR32], nil)
var bh gc.Node
gc.Regalloc(&bh, gc.Types[gc.TPTR32], nil)
gins(arm.AMOVW, &lo1, &al)
gins(arm.AMOVW, &hi1, &ah)
gins(arm.AMOVW, &lo2, &bl)
gins(arm.AMOVW, &hi2, &bh)
p1 := gins(arm.ASUB, &bl, &al)
p1.Scond |= arm.C_SBIT
gins(arm.ASBC, &bh, &ah)
gc.Regfree(&bl)
gc.Regfree(&bh)
// TODO(kaib): this can be done with 4 regs and does not need 6
case gc.OMUL:
var bl gc.Node
gc.Regalloc(&bl, gc.Types[gc.TPTR32], nil)
var bh gc.Node
gc.Regalloc(&bh, gc.Types[gc.TPTR32], nil)
var cl gc.Node
gc.Regalloc(&cl, gc.Types[gc.TPTR32], nil)
var ch gc.Node
gc.Regalloc(&ch, gc.Types[gc.TPTR32], nil)
// load args into bh:bl and bh:bl.
gins(arm.AMOVW, &hi1, &bh)
gins(arm.AMOVW, &lo1, &bl)
gins(arm.AMOVW, &hi2, &ch)
gins(arm.AMOVW, &lo2, &cl)
// bl * cl -> ah al
p1 := gins(arm.AMULLU, nil, nil)
p1.From.Type = obj.TYPE_REG
p1.From.Reg = bl.Reg
p1.Reg = cl.Reg
p1.To.Type = obj.TYPE_REGREG
p1.To.Reg = ah.Reg
p1.To.Offset = int64(al.Reg)
//print("%v\n", p1);
// bl * ch + ah -> ah
p1 = gins(arm.AMULA, nil, nil)
p1.From.Type = obj.TYPE_REG
p1.From.Reg = bl.Reg
p1.Reg = ch.Reg
p1.To.Type = obj.TYPE_REGREG2
p1.To.Reg = ah.Reg
p1.To.Offset = int64(ah.Reg)
//print("%v\n", p1);
// bh * cl + ah -> ah
p1 = gins(arm.AMULA, nil, nil)
p1.From.Type = obj.TYPE_REG
p1.From.Reg = bh.Reg
p1.Reg = cl.Reg
p1.To.Type = obj.TYPE_REGREG2
p1.To.Reg = ah.Reg
p1.To.Offset = int64(ah.Reg)
//print("%v\n", p1);
gc.Regfree(&bh)
gc.Regfree(&bl)
gc.Regfree(&ch)
gc.Regfree(&cl)
// 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())
var bl gc.Node
gc.Regalloc(&bl, lo1.Type, nil)
var bh gc.Node
gc.Regalloc(&bh, hi1.Type, nil)
if v >= 32 {
// reverse during load to do the first 32 bits of rotate
v -= 32
gins(arm.AMOVW, &hi1, &bl)
gins(arm.AMOVW, &lo1, &bh)
} else {
gins(arm.AMOVW, &hi1, &bh)
gins(arm.AMOVW, &lo1, &bl)
}
if v == 0 {
gins(arm.AMOVW, &bh, &ah)
gins(arm.AMOVW, &bl, &al)
} else {
// rotate by 1 <= v <= 31
// MOVW bl<<v, al
// MOVW bh<<v, ah
// OR bl>>(32-v), ah
// OR bh>>(32-v), al
gshift(arm.AMOVW, &bl, arm.SHIFT_LL, int32(v), &al)
gshift(arm.AMOVW, &bh, arm.SHIFT_LL, int32(v), &ah)
gshift(arm.AORR, &bl, arm.SHIFT_LR, int32(32-v), &ah)
gshift(arm.AORR, &bh, arm.SHIFT_LR, int32(32-v), &al)
}
gc.Regfree(&bl)
gc.Regfree(&bh)
case gc.OLSH:
var bl gc.Node
gc.Regalloc(&bl, lo1.Type, nil)
var bh gc.Node
gc.Regalloc(&bh, hi1.Type, nil)
gins(arm.AMOVW, &hi1, &bh)
gins(arm.AMOVW, &lo1, &bl)
var p6 *obj.Prog
var s gc.Node
var n1 gc.Node
var creg gc.Node
var p1 *obj.Prog
var p2 *obj.Prog
var p3 *obj.Prog
var p4 *obj.Prog
var p5 *obj.Prog
if r.Op == gc.OLITERAL {
v := uint64(r.Int())
if v >= 64 {
// TODO(kaib): replace with gins(AMOVW, nodintconst(0), &al)
// here and below (verify it optimizes to EOR)
gins(arm.AEOR, &al, &al)
gins(arm.AEOR, &ah, &ah)
} else if v > 32 {
gins(arm.AEOR, &al, &al)
// MOVW bl<<(v-32), ah
gshift(arm.AMOVW, &bl, arm.SHIFT_LL, int32(v-32), &ah)
} else if v == 32 {
gins(arm.AEOR, &al, &al)
gins(arm.AMOVW, &bl, &ah)
} else if v > 0 {
// MOVW bl<<v, al
gshift(arm.AMOVW, &bl, arm.SHIFT_LL, int32(v), &al)
// MOVW bh<<v, ah
gshift(arm.AMOVW, &bh, arm.SHIFT_LL, int32(v), &ah)
// OR bl>>(32-v), ah
gshift(arm.AORR, &bl, arm.SHIFT_LR, int32(32-v), &ah)
} else {
gins(arm.AMOVW, &bl, &al)
gins(arm.AMOVW, &bh, &ah)
}
goto olsh_break
}
gc.Regalloc(&s, gc.Types[gc.TUINT32], nil)
gc.Regalloc(&creg, gc.Types[gc.TUINT32], nil)
if gc.Is64(r.Type) {
// shift is >= 1<<32
var cl gc.Node
var ch gc.Node
split64(r, &cl, &ch)
gmove(&ch, &s)
gins(arm.ATST, &s, nil)
p6 = gc.Gbranch(arm.ABNE, nil, 0)
gmove(&cl, &s)
splitclean()
} else {
gmove(r, &s)
p6 = nil
}
gins(arm.ATST, &s, nil)
// shift == 0
p1 = gins(arm.AMOVW, &bl, &al)
p1.Scond = arm.C_SCOND_EQ
p1 = gins(arm.AMOVW, &bh, &ah)
p1.Scond = arm.C_SCOND_EQ
p2 = gc.Gbranch(arm.ABEQ, nil, 0)
// shift is < 32
gc.Nodconst(&n1, gc.Types[gc.TUINT32], 32)
gmove(&n1, &creg)
gins(arm.ACMP, &s, &creg)
// MOVW.LO bl<<s, al
p1 = gregshift(arm.AMOVW, &bl, arm.SHIFT_LL, &s, &al)
p1.Scond = arm.C_SCOND_LO
// MOVW.LO bh<<s, ah
p1 = gregshift(arm.AMOVW, &bh, arm.SHIFT_LL, &s, &ah)
p1.Scond = arm.C_SCOND_LO
// SUB.LO s, creg
p1 = gins(arm.ASUB, &s, &creg)
p1.Scond = arm.C_SCOND_LO
// OR.LO bl>>creg, ah
p1 = gregshift(arm.AORR, &bl, arm.SHIFT_LR, &creg, &ah)
p1.Scond = arm.C_SCOND_LO
// BLO end
p3 = gc.Gbranch(arm.ABLO, nil, 0)
// shift == 32
p1 = gins(arm.AEOR, &al, &al)
p1.Scond = arm.C_SCOND_EQ
p1 = gins(arm.AMOVW, &bl, &ah)
p1.Scond = arm.C_SCOND_EQ
p4 = gc.Gbranch(arm.ABEQ, nil, 0)
// shift is < 64
gc.Nodconst(&n1, gc.Types[gc.TUINT32], 64)
gmove(&n1, &creg)
gins(arm.ACMP, &s, &creg)
// EOR.LO al, al
p1 = gins(arm.AEOR, &al, &al)
p1.Scond = arm.C_SCOND_LO
// MOVW.LO creg>>1, creg
p1 = gshift(arm.AMOVW, &creg, arm.SHIFT_LR, 1, &creg)
p1.Scond = arm.C_SCOND_LO
// SUB.LO creg, s
p1 = gins(arm.ASUB, &creg, &s)
p1.Scond = arm.C_SCOND_LO
// MOVW bl<<s, ah
p1 = gregshift(arm.AMOVW, &bl, arm.SHIFT_LL, &s, &ah)
p1.Scond = arm.C_SCOND_LO
p5 = gc.Gbranch(arm.ABLO, nil, 0)
// shift >= 64
if p6 != nil {
gc.Patch(p6, gc.Pc)
}
gins(arm.AEOR, &al, &al)
gins(arm.AEOR, &ah, &ah)
gc.Patch(p2, gc.Pc)
gc.Patch(p3, gc.Pc)
gc.Patch(p4, gc.Pc)
gc.Patch(p5, gc.Pc)
gc.Regfree(&s)
gc.Regfree(&creg)
olsh_break:
gc.Regfree(&bl)
gc.Regfree(&bh)
case gc.ORSH:
var bl gc.Node
gc.Regalloc(&bl, lo1.Type, nil)
var bh gc.Node
gc.Regalloc(&bh, hi1.Type, nil)
gins(arm.AMOVW, &hi1, &bh)
gins(arm.AMOVW, &lo1, &bl)
var p4 *obj.Prog
var p5 *obj.Prog
var n1 gc.Node
var p6 *obj.Prog
var s gc.Node
var p1 *obj.Prog
var p2 *obj.Prog
var creg gc.Node
var p3 *obj.Prog
if r.Op == gc.OLITERAL {
v := uint64(r.Int())
if v >= 64 {
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->31, al
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &al)
// MOVW bh->31, ah
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
} else {
gins(arm.AEOR, &al, &al)
gins(arm.AEOR, &ah, &ah)
}
} else if v > 32 {
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->(v-32), al
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, int32(v-32), &al)
// MOVW bh->31, ah
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
} else {
// MOVW bh>>(v-32), al
gshift(arm.AMOVW, &bh, arm.SHIFT_LR, int32(v-32), &al)
gins(arm.AEOR, &ah, &ah)
}
} else if v == 32 {
gins(arm.AMOVW, &bh, &al)
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->31, ah
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
} else {
gins(arm.AEOR, &ah, &ah)
}
} else if v > 0 {
// MOVW bl>>v, al
gshift(arm.AMOVW, &bl, arm.SHIFT_LR, int32(v), &al)
// OR bh<<(32-v), al
gshift(arm.AORR, &bh, arm.SHIFT_LL, int32(32-v), &al)
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->v, ah
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, int32(v), &ah)
} else {
// MOVW bh>>v, ah
gshift(arm.AMOVW, &bh, arm.SHIFT_LR, int32(v), &ah)
}
} else {
gins(arm.AMOVW, &bl, &al)
gins(arm.AMOVW, &bh, &ah)
}
goto orsh_break
}
gc.Regalloc(&s, gc.Types[gc.TUINT32], nil)
gc.Regalloc(&creg, gc.Types[gc.TUINT32], nil)
if gc.Is64(r.Type) {
// shift is >= 1<<32
var ch gc.Node
var cl gc.Node
split64(r, &cl, &ch)
gmove(&ch, &s)
gins(arm.ATST, &s, nil)
var p1 *obj.Prog
if bh.Type.Etype == gc.TINT32 {
p1 = gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
} else {
p1 = gins(arm.AEOR, &ah, &ah)
}
p1.Scond = arm.C_SCOND_NE
p6 = gc.Gbranch(arm.ABNE, nil, 0)
gmove(&cl, &s)
splitclean()
} else {
gmove(r, &s)
p6 = nil
}
gins(arm.ATST, &s, nil)
// shift == 0
p1 = gins(arm.AMOVW, &bl, &al)
p1.Scond = arm.C_SCOND_EQ
p1 = gins(arm.AMOVW, &bh, &ah)
p1.Scond = arm.C_SCOND_EQ
p2 = gc.Gbranch(arm.ABEQ, nil, 0)
// check if shift is < 32
gc.Nodconst(&n1, gc.Types[gc.TUINT32], 32)
gmove(&n1, &creg)
gins(arm.ACMP, &s, &creg)
// MOVW.LO bl>>s, al
p1 = gregshift(arm.AMOVW, &bl, arm.SHIFT_LR, &s, &al)
p1.Scond = arm.C_SCOND_LO
// SUB.LO s,creg
p1 = gins(arm.ASUB, &s, &creg)
p1.Scond = arm.C_SCOND_LO
// OR.LO bh<<(32-s), al
p1 = gregshift(arm.AORR, &bh, arm.SHIFT_LL, &creg, &al)
p1.Scond = arm.C_SCOND_LO
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->s, ah
p1 = gregshift(arm.AMOVW, &bh, arm.SHIFT_AR, &s, &ah)
} else {
// MOVW bh>>s, ah
p1 = gregshift(arm.AMOVW, &bh, arm.SHIFT_LR, &s, &ah)
}
p1.Scond = arm.C_SCOND_LO
// BLO end
p3 = gc.Gbranch(arm.ABLO, nil, 0)
// shift == 32
p1 = gins(arm.AMOVW, &bh, &al)
p1.Scond = arm.C_SCOND_EQ
if bh.Type.Etype == gc.TINT32 {
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
} else {
gins(arm.AEOR, &ah, &ah)
}
p4 = gc.Gbranch(arm.ABEQ, nil, 0)
// check if shift is < 64
gc.Nodconst(&n1, gc.Types[gc.TUINT32], 64)
gmove(&n1, &creg)
gins(arm.ACMP, &s, &creg)
// MOVW.LO creg>>1, creg
p1 = gshift(arm.AMOVW, &creg, arm.SHIFT_LR, 1, &creg)
p1.Scond = arm.C_SCOND_LO
// SUB.LO creg, s
p1 = gins(arm.ASUB, &creg, &s)
p1.Scond = arm.C_SCOND_LO
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->(s-32), al
p1 := gregshift(arm.AMOVW, &bh, arm.SHIFT_AR, &s, &al)
p1.Scond = arm.C_SCOND_LO
} else {
// MOVW bh>>(v-32), al
p1 := gregshift(arm.AMOVW, &bh, arm.SHIFT_LR, &s, &al)
p1.Scond = arm.C_SCOND_LO
}
// BLO end
p5 = gc.Gbranch(arm.ABLO, nil, 0)
// s >= 64
if p6 != nil {
gc.Patch(p6, gc.Pc)
}
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->31, al
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &al)
} else {
gins(arm.AEOR, &al, &al)
}
gc.Patch(p2, gc.Pc)
gc.Patch(p3, gc.Pc)
gc.Patch(p4, gc.Pc)
gc.Patch(p5, gc.Pc)
gc.Regfree(&s)
gc.Regfree(&creg)
orsh_break:
gc.Regfree(&bl)
gc.Regfree(&bh)
// TODO(kaib): literal optimizations
// make constant the right side (it usually is anyway).
// if(lo1.op == OLITERAL) {
// nswap(&lo1, &lo2);
// nswap(&hi1, &hi2);
// }
// if(lo2.op == OLITERAL) {
// // special cases for constants.
// lv = mpgetfix(lo2.val.u.xval);
// hv = mpgetfix(hi2.val.u.xval);
// splitclean(); // right side
// split64(res, &lo2, &hi2);
// switch(n->op) {
// case OXOR:
// gmove(&lo1, &lo2);
// gmove(&hi1, &hi2);
// switch(lv) {
// case 0:
// break;
// case 0xffffffffu:
// gins(ANOTL, N, &lo2);
// break;
// default:
// gins(AXORL, ncon(lv), &lo2);
// break;
// }
// switch(hv) {
// case 0:
// break;
// case 0xffffffffu:
// gins(ANOTL, N, &hi2);
// break;
// default:
// gins(AXORL, ncon(hv), &hi2);
// break;
// }
// break;
// case OAND:
// switch(lv) {
// case 0:
// gins(AMOVL, ncon(0), &lo2);
// break;
// default:
// gmove(&lo1, &lo2);
// if(lv != 0xffffffffu)
// gins(AANDL, ncon(lv), &lo2);
// break;
// }
// switch(hv) {
// case 0:
// gins(AMOVL, ncon(0), &hi2);
// break;
// default:
// gmove(&hi1, &hi2);
// if(hv != 0xffffffffu)
// gins(AANDL, ncon(hv), &hi2);
// break;
// }
// break;
// case OOR:
// switch(lv) {
// case 0:
// gmove(&lo1, &lo2);
// break;
// case 0xffffffffu:
// gins(AMOVL, ncon(0xffffffffu), &lo2);
// break;
// default:
// gmove(&lo1, &lo2);
// gins(AORL, ncon(lv), &lo2);
// break;
// }
// switch(hv) {
// case 0:
// gmove(&hi1, &hi2);
// break;
// case 0xffffffffu:
// gins(AMOVL, ncon(0xffffffffu), &hi2);
// break;
// default:
// gmove(&hi1, &hi2);
// gins(AORL, ncon(hv), &hi2);
// break;
// }
// break;
// }
// splitclean();
// splitclean();
// goto out;
// }
case gc.OXOR,
gc.OAND,
gc.OOR:
var n1 gc.Node
gc.Regalloc(&n1, lo1.Type, nil)
gins(arm.AMOVW, &lo1, &al)
gins(arm.AMOVW, &hi1, &ah)
gins(arm.AMOVW, &lo2, &n1)
gins(optoas(n.Op, lo1.Type), &n1, &al)
gins(arm.AMOVW, &hi2, &n1)
gins(optoas(n.Op, lo1.Type), &n1, &ah)
gc.Regfree(&n1)
}
if gc.Is64(r.Type) {
splitclean()
}
splitclean()
split64(res, &lo1, &hi1)
gins(arm.AMOVW, &al, &lo1)
gins(arm.AMOVW, &ah, &hi1)
splitclean()
//out:
gc.Regfree(&al)
gc.Regfree(&ah)
}
func clearfat_tail(n1 *gc.Node, b int64) {
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--
}
}
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 int
switch align {
default:
gc.Fatalf("sgen: invalid alignment %d for %v", align, n.Type)
case 1:
op = ppc64.AMOVBU
case 2:
op = ppc64.AMOVHU
case 4:
op = ppc64.AMOVWZU // there is no lwau, only lwaux
case 8:
op = ppc64.AMOVDU
}
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 we are copying forward on the stack and
// the src and dst overlap, then reverse direction
dir := align
if osrc < odst && int64(odst) < int64(osrc)+w {
dir = -dir
}
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(ppc64.AMOVD, &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.Tptr], nil)
// set up end marker
var nend gc.Node
// move src and dest to the end of block if necessary
if dir < 0 {
if c >= 4 {
gc.Regalloc(&nend, gc.Types[gc.Tptr], nil)
gins(ppc64.AMOVD, &src, &nend)
}
p := gins(ppc64.AADD, nil, &src)
p.From.Type = obj.TYPE_CONST
p.From.Offset = w
p = gins(ppc64.AADD, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = w
} else {
p := gins(ppc64.AADD, nil, &src)
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(-dir)
p = gins(ppc64.AADD, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(-dir)
if c >= 4 {
gc.Regalloc(&nend, gc.Types[gc.Tptr], nil)
p := gins(ppc64.AMOVD, &src, &nend)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w
}
}
// move
// TODO: enable duffcopy for larger copies.
if c >= 4 {
p := gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
ploop := p
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
p = gins(ppc64.ACMP, &src, &nend)
gc.Patch(gc.Gbranch(ppc64.ABNE, nil, 0), ploop)
gc.Regfree(&nend)
} else {
// TODO(austin): Instead of generating ADD $-8,R8; ADD
// $-8,R7; n*(MOVDU 8(R8),R9; MOVDU R9,8(R7);) just
// generate the offsets directly and eliminate the
// ADDs. That will produce shorter, more
// pipeline-able code.
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 = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
}
}
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 int, 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.Int()
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 = obj.TYPE_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 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 gc.Issigned[t.Etype] {
check = true
if gc.Isconst(nl, gc.CTINT) && nl.Int() != -(1<<uint64(t.Width*8-1)) {
check = false
} else if gc.Isconst(nr, gc.CTINT) && nr.Int() != -1 {
check = false
}
}
if t.Width < 4 {
if gc.Issigned[t.Etype] {
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 !gc.Issigned[t.Etype] {
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)
}
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 int
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 && int64(odst) < int64(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 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 := 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 int
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.Int(); 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.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.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
}