/*
* generate high multiply:
* res = (nl*nr) >> width
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
t := nl.Type
a := optoas(gc.OHMUL, t)
if nl.Ullman < nr.Ullman {
tmp := nl
nl = nr
nr = tmp
}
var n1 gc.Node
gc.Cgenr(nl, &n1, res)
var n2 gc.Node
gc.Cgenr(nr, &n2, nil)
var ax gc.Node
gc.Nodreg(&ax, t, x86.REG_AX)
gmove(&n1, &ax)
gins(a, &n2, nil)
gc.Regfree(&n2)
gc.Regfree(&n1)
var dx gc.Node
if t.Width == 1 {
// byte multiply behaves differently.
gc.Nodreg(&ax, t, x86.REG_AH)
gc.Nodreg(&dx, t, x86.REG_DX)
gmove(&ax, &dx)
}
gc.Nodreg(&dx, t, x86.REG_DX)
gmove(&dx, res)
}
/*
* generate 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)
}
func ginscmp(op int, 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 int, 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)
}
func ginscmp(op int, t *gc.Type, n1, n2 *gc.Node, likely int) *obj.Prog {
if gc.Isint[t.Etype] && n1.Op == gc.OLITERAL && gc.Smallintconst(n1) && n2.Op != gc.OLITERAL {
// Reverse comparison to place constant last.
op = gc.Brrev(op)
n1, n2 = n2, n1
}
// General case.
var r1, r2, g1, g2 gc.Node
if n1.Op == gc.ONAME && n1.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] && gc.Smallintconst(n2) {
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)
}
func sudoclean() {
if clean[cleani-1].Op != gc.OEMPTY {
gc.Regfree(&clean[cleani-1])
}
if clean[cleani-2].Op != gc.OEMPTY {
gc.Regfree(&clean[cleani-2])
}
cleani -= 2
}
/*
* generate high multiply
* res = (nl * nr) >> wordsize
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Ullman < nr.Ullman {
tmp := nl
nl = nr
nr = tmp
}
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.Fatal("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.Fatal("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)
}
/*
* generate high multiply:
* res = (nl*nr) >> width
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
// largest ullman on left.
if nl.Ullman < nr.Ullman {
tmp := (*gc.Node)(nl)
nl = nr
nr = tmp
}
t := (*gc.Type)(nl.Type)
w := int(int(t.Width * 8))
var n1 gc.Node
gc.Cgenr(nl, &n1, res)
var n2 gc.Node
gc.Cgenr(nr, &n2, nil)
switch gc.Simtype[t.Etype] {
case gc.TINT8,
gc.TINT16,
gc.TINT32:
gins(optoas(gc.OMUL, t), &n2, &n1)
p := (*obj.Prog)(gins(ppc64.ASRAD, nil, &n1))
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(w)
case gc.TUINT8,
gc.TUINT16,
gc.TUINT32:
gins(optoas(gc.OMUL, t), &n2, &n1)
p := (*obj.Prog)(gins(ppc64.ASRD, nil, &n1))
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(w)
case gc.TINT64,
gc.TUINT64:
if gc.Issigned[t.Etype] {
gins(ppc64.AMULHD, &n2, &n1)
} else {
gins(ppc64.AMULHDU, &n2, &n1)
}
default:
gc.Fatal("cgen_hmul %v", gc.Tconv(t, 0))
}
gc.Cgen(&n1, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
func restx(x *gc.Node, oldx *gc.Node) {
if oldx.Op != 0 {
x.Type = gc.Types[gc.TINT64]
reg[x.Reg] = uint8(oldx.Ostk)
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 splitclean() {
if nsclean <= 0 {
gc.Fatal("splitclean")
}
nsclean--
if sclean[nsclean].Op != gc.OEMPTY {
gc.Regfree(&sclean[nsclean])
}
}
/*
* 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 int, nl *gc.Node, nr *gc.Node, res *gc.Node) bool {
if optoas(op, nl.Type) != x86.AIMULB {
return false
}
// largest ullman on left.
if nl.Ullman < nr.Ullman {
tmp := nl
nl = nr
nr = tmp
}
// generate operands in "8-bit" registers.
var n1b gc.Node
gc.Regalloc(&n1b, nl.Type, res)
gc.Cgen(nl, &n1b)
var n2b gc.Node
gc.Regalloc(&n2b, nr.Type, nil)
gc.Cgen(nr, &n2b)
// perform full-width multiplication.
t := gc.Types[gc.TUINT64]
if gc.Issigned[nl.Type.Etype] {
t = gc.Types[gc.TINT64]
}
var n1 gc.Node
gc.Nodreg(&n1, t, int(n1b.Reg))
var n2 gc.Node
gc.Nodreg(&n2, t, int(n2b.Reg))
a := optoas(op, t)
gins(a, &n2, &n1)
// truncate.
gmove(&n1, res)
gc.Regfree(&n1b)
gc.Regfree(&n2b)
return true
}
func gencmp0(n *gc.Node, t *gc.Type, o int, likely int, to *obj.Prog) {
var n1 gc.Node
gc.Regalloc(&n1, t, nil)
gc.Cgen(n, &n1)
a := optoas(gc.OCMP, t)
if a != arm.ACMP {
var n2 gc.Node
gc.Nodconst(&n2, t, 0)
var n3 gc.Node
gc.Regalloc(&n3, t, nil)
gmove(&n2, &n3)
gins(a, &n1, &n3)
gc.Regfree(&n3)
} else {
gins(arm.ATST, &n1, nil)
}
a = optoas(o, t)
gc.Patch(gc.Gbranch(a, t, likely), to)
gc.Regfree(&n1)
}
// 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)
}
/*
* generate byte multiply:
* res = nl * nr
* there is no 2-operand byte multiply instruction so
* we do a full-width multiplication and truncate afterwards.
*/
func cgen_bmul(op int, nl *gc.Node, nr *gc.Node, res *gc.Node) bool {
if optoas(op, nl.Type) != x86.AIMULB {
return false
}
// 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 {
tmp := nl
nl = nr
nr = tmp
}
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
* 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
var ax gc.Node
var dx gc.Node
t := nl.Type
a := optoas(gc.OHMUL, t)
// gen nl in n1.
gc.Tempname(&n1, t)
gc.Cgen(nl, &n1)
// gen nr in n2.
gc.Regalloc(&n2, t, res)
gc.Cgen(nr, &n2)
// multiply.
gc.Nodreg(&ax, t, x86.REG_AX)
gmove(&n2, &ax)
gins(a, &n1, nil)
gc.Regfree(&n2)
if t.Width == 1 {
// byte multiply behaves differently.
gc.Nodreg(&ax, t, x86.REG_AH)
gc.Nodreg(&dx, t, x86.REG_DX)
gmove(&ax, &dx)
}
gc.Nodreg(&dx, t, x86.REG_DX)
gmove(&dx, res)
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op int, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Type.Width > 4 {
gc.Fatal("cgen_shift %v", gc.Tconv(nl.Type, 0))
}
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(gc.Mpgetfix(nr.Val.U.Xval))
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 reg[x86.REG_CX] > 1 && !gc.Samereg(&cx, res) {
gc.Tempname(&oldcx, gc.Types[gc.TUINT32])
gmove(&cx, &oldcx)
}
var n1 gc.Node
var nt gc.Node
if nr.Type.Width > 4 {
gc.Tempname(&nt, nr.Type)
n1 = nt
} else {
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, nr.Type, &n1) // to hold the shift type in CX
}
var n2 gc.Node
if gc.Samereg(&cx, res) {
gc.Regalloc(&n2, nl.Type, nil)
} else {
gc.Regalloc(&n2, nl.Type, res)
}
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &n2)
gc.Cgen(nr, &n1)
} else {
gc.Cgen(nr, &n1)
gc.Cgen(nl, &n2)
}
// test and fix up large shifts
if bounded {
if nr.Type.Width > 4 {
// delayed reg alloc
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], &n1) // to hold the shift type in CX
var lo gc.Node
var hi gc.Node
split64(&nt, &lo, &hi)
gmove(&lo, &n1)
splitclean()
}
} else {
var p1 *obj.Prog
if nr.Type.Width > 4 {
// delayed reg alloc
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], &n1) // to hold the shift type in CX
var lo gc.Node
var hi gc.Node
split64(&nt, &lo, &hi)
gmove(&lo, &n1)
gins(optoas(gc.OCMP, gc.Types[gc.TUINT32]), &hi, ncon(0))
p2 := gc.Gbranch(optoas(gc.ONE, gc.Types[gc.TUINT32]), nil, +1)
gins(optoas(gc.OCMP, gc.Types[gc.TUINT32]), &n1, ncon(uint32(w)))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
splitclean()
gc.Patch(p2, gc.Pc)
} else {
gins(optoas(gc.OCMP, nr.Type), &n1, ncon(uint32(w)))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
}
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gins(a, ncon(uint32(w)-1), &n2)
} else {
gmove(ncon(0), &n2)
}
gc.Patch(p1, gc.Pc)
}
gins(a, &n1, &n2)
if oldcx.Op != 0 {
gmove(&oldcx, &cx)
}
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
func gmove(f *gc.Node, t *gc.Node) {
if gc.Debug['M'] != 0 {
fmt.Printf("gmove %v -> %v\n", gc.Nconv(f, 0), gc.Nconv(t, 0))
}
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
cvt := t.Type
if gc.Iscomplex[ft] || gc.Iscomplex[tt] {
gc.Complexmove(f, t)
return
}
if gc.Isfloat[ft] || gc.Isfloat[tt] {
floatmove(f, t)
return
}
// cannot have two integer memory operands;
// except 64-bit, which always copies via registers anyway.
var r1 gc.Node
var a int
if gc.Isint[ft] && gc.Isint[tt] && !gc.Is64(f.Type) && !gc.Is64(t.Type) && gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
gc.Convconst(&con, t.Type, &f.Val)
f = &con
ft = gc.Simsimtype(con.Type)
}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
// should not happen
gc.Fatal("gmove %v -> %v", gc.Nconv(f, 0), gc.Nconv(t, 0))
return
/*
* integer copy and truncate
*/
case gc.TINT8<<16 | gc.TINT8, // same size
gc.TINT8<<16 | gc.TUINT8,
gc.TUINT8<<16 | gc.TINT8,
gc.TUINT8<<16 | gc.TUINT8:
a = x86.AMOVB
case gc.TINT16<<16 | gc.TINT8, // truncate
gc.TUINT16<<16 | gc.TINT8,
gc.TINT32<<16 | gc.TINT8,
gc.TUINT32<<16 | gc.TINT8,
gc.TINT16<<16 | gc.TUINT8,
gc.TUINT16<<16 | gc.TUINT8,
gc.TINT32<<16 | gc.TUINT8,
gc.TUINT32<<16 | gc.TUINT8:
a = x86.AMOVB
goto rsrc
case gc.TINT64<<16 | gc.TINT8, // truncate low word
gc.TUINT64<<16 | gc.TINT8,
gc.TINT64<<16 | gc.TUINT8,
gc.TUINT64<<16 | gc.TUINT8:
var flo gc.Node
var fhi gc.Node
split64(f, &flo, &fhi)
var r1 gc.Node
gc.Nodreg(&r1, t.Type, x86.REG_AX)
gmove(&flo, &r1)
gins(x86.AMOVB, &r1, t)
splitclean()
return
case gc.TINT16<<16 | gc.TINT16, // same size
gc.TINT16<<16 | gc.TUINT16,
gc.TUINT16<<16 | gc.TINT16,
gc.TUINT16<<16 | gc.TUINT16:
a = x86.AMOVW
case gc.TINT32<<16 | gc.TINT16, // truncate
gc.TUINT32<<16 | gc.TINT16,
gc.TINT32<<16 | gc.TUINT16,
gc.TUINT32<<16 | gc.TUINT16:
a = x86.AMOVW
goto rsrc
case gc.TINT64<<16 | gc.TINT16, // truncate low word
gc.TUINT64<<16 | gc.TINT16,
gc.TINT64<<16 | gc.TUINT16,
gc.TUINT64<<16 | gc.TUINT16:
var flo gc.Node
var fhi gc.Node
split64(f, &flo, &fhi)
var r1 gc.Node
gc.Nodreg(&r1, t.Type, x86.REG_AX)
gmove(&flo, &r1)
gins(x86.AMOVW, &r1, t)
splitclean()
return
case gc.TINT32<<16 | gc.TINT32, // same size
gc.TINT32<<16 | gc.TUINT32,
gc.TUINT32<<16 | gc.TINT32,
gc.TUINT32<<16 | gc.TUINT32:
a = x86.AMOVL
case gc.TINT64<<16 | gc.TINT32, // truncate
gc.TUINT64<<16 | gc.TINT32,
gc.TINT64<<16 | gc.TUINT32,
gc.TUINT64<<16 | gc.TUINT32:
var fhi gc.Node
var flo gc.Node
split64(f, &flo, &fhi)
var r1 gc.Node
gc.Nodreg(&r1, t.Type, x86.REG_AX)
gmove(&flo, &r1)
gins(x86.AMOVL, &r1, t)
splitclean()
return
case gc.TINT64<<16 | gc.TINT64, // same size
gc.TINT64<<16 | gc.TUINT64,
gc.TUINT64<<16 | gc.TINT64,
gc.TUINT64<<16 | gc.TUINT64:
var fhi gc.Node
var flo gc.Node
split64(f, &flo, &fhi)
var tlo gc.Node
var thi gc.Node
split64(t, &tlo, &thi)
if f.Op == gc.OLITERAL {
gins(x86.AMOVL, &flo, &tlo)
gins(x86.AMOVL, &fhi, &thi)
} else {
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[gc.TUINT32], x86.REG_AX)
var r2 gc.Node
gc.Nodreg(&r2, gc.Types[gc.TUINT32], x86.REG_DX)
gins(x86.AMOVL, &flo, &r1)
gins(x86.AMOVL, &fhi, &r2)
gins(x86.AMOVL, &r1, &tlo)
gins(x86.AMOVL, &r2, &thi)
}
splitclean()
splitclean()
return
/*
* integer up-conversions
*/
case gc.TINT8<<16 | gc.TINT16, // sign extend int8
gc.TINT8<<16 | gc.TUINT16:
a = x86.AMOVBWSX
goto rdst
case gc.TINT8<<16 | gc.TINT32,
gc.TINT8<<16 | gc.TUINT32:
a = x86.AMOVBLSX
goto rdst
case gc.TINT8<<16 | gc.TINT64, // convert via int32
gc.TINT8<<16 | gc.TUINT64:
cvt = gc.Types[gc.TINT32]
goto hard
case gc.TUINT8<<16 | gc.TINT16, // zero extend uint8
gc.TUINT8<<16 | gc.TUINT16:
a = x86.AMOVBWZX
goto rdst
case gc.TUINT8<<16 | gc.TINT32,
gc.TUINT8<<16 | gc.TUINT32:
a = x86.AMOVBLZX
goto rdst
case gc.TUINT8<<16 | gc.TINT64, // convert via uint32
gc.TUINT8<<16 | gc.TUINT64:
cvt = gc.Types[gc.TUINT32]
goto hard
case gc.TINT16<<16 | gc.TINT32, // sign extend int16
gc.TINT16<<16 | gc.TUINT32:
a = x86.AMOVWLSX
goto rdst
case gc.TINT16<<16 | gc.TINT64, // convert via int32
gc.TINT16<<16 | gc.TUINT64:
cvt = gc.Types[gc.TINT32]
goto hard
case gc.TUINT16<<16 | gc.TINT32, // zero extend uint16
gc.TUINT16<<16 | gc.TUINT32:
a = x86.AMOVWLZX
goto rdst
case gc.TUINT16<<16 | gc.TINT64, // convert via uint32
gc.TUINT16<<16 | gc.TUINT64:
cvt = gc.Types[gc.TUINT32]
goto hard
case gc.TINT32<<16 | gc.TINT64, // sign extend int32
gc.TINT32<<16 | gc.TUINT64:
var thi gc.Node
var tlo gc.Node
split64(t, &tlo, &thi)
var flo gc.Node
gc.Nodreg(&flo, tlo.Type, x86.REG_AX)
var fhi gc.Node
gc.Nodreg(&fhi, thi.Type, x86.REG_DX)
gmove(f, &flo)
gins(x86.ACDQ, nil, nil)
gins(x86.AMOVL, &flo, &tlo)
gins(x86.AMOVL, &fhi, &thi)
splitclean()
return
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
var tlo gc.Node
var thi gc.Node
split64(t, &tlo, &thi)
gmove(f, &tlo)
gins(x86.AMOVL, ncon(0), &thi)
splitclean()
return
}
gins(a, f, t)
return
// requires register source
rsrc:
gc.Regalloc(&r1, f.Type, t)
gmove(f, &r1)
gins(a, &r1, t)
gc.Regfree(&r1)
return
// requires register destination
rdst:
{
gc.Regalloc(&r1, t.Type, t)
gins(a, f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
/*
* generate 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 int, 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 := 0
if gc.Issigned[t.Etype] {
check = 1
if gc.Isconst(nl, gc.CTINT) && gc.Mpgetfix(nl.Val.U.Xval) != -1<<uint64(t.Width*8-1) {
check = 0
} else if gc.Isconst(nr, gc.CTINT) && gc.Mpgetfix(nr.Val.U.Xval) != -1 {
check = 0
}
}
if t.Width < 4 {
if gc.Issigned[t.Etype] {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
check = 0
}
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 != 0 {
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 !gc.Issigned[t.Etype] {
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 != 0 {
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 {
tmp14 := q
q--
if tmp14 <= 0 {
break
}
n1.Type = z.Type
gins(x86.AMOVL, &z, &n1)
n1.Xoffset += 4
}
gc.Nodconst(&z, gc.Types[gc.TUINT8], 0)
for {
tmp15 := c
c--
if tmp15 <= 0 {
break
}
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--
}
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
gc.Dump("\nclearfat", nl)
}
w := nl.Type.Width
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := w % 8 // bytes
q := w / 8 // quads
if q < 4 {
// 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.
// 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
var z gc.Node
gc.Nodconst(&z, gc.Types[gc.TUINT64], 0)
for {
tmp14 := q
q--
if tmp14 <= 0 {
break
}
n1.Type = z.Type
gins(x86.AMOVQ, &z, &n1)
n1.Xoffset += 8
}
if c >= 4 {
gc.Nodconst(&z, gc.Types[gc.TUINT32], 0)
n1.Type = z.Type
gins(x86.AMOVL, &z, &n1)
n1.Xoffset += 4
c -= 4
}
gc.Nodconst(&z, gc.Types[gc.TUINT8], 0)
for {
tmp15 := c
c--
if tmp15 <= 0 {
break
}
n1.Type = z.Type
gins(x86.AMOVB, &z, &n1)
n1.Xoffset++
}
gc.Regfree(&n1)
return
}
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)
if q > 128 || gc.Nacl {
gconreg(movptr, q, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.ASTOSQ, nil, nil) // STOQ AL,*(DI)+
} else {
if di := dzDI(q); 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(q)
}
z := ax
di := n1
if w >= 8 && c >= 4 {
di.Op = gc.OINDREG
z.Type = gc.Types[gc.TINT64]
di.Type = z.Type
p := gins(x86.AMOVQ, &z, &di)
p.To.Scale = 1
p.To.Offset = c - 8
} else if c >= 4 {
di.Op = gc.OINDREG
z.Type = gc.Types[gc.TINT32]
di.Type = z.Type
gins(x86.AMOVL, &z, &di)
if c > 4 {
p := gins(x86.AMOVL, &z, &di)
p.To.Scale = 1
p.To.Offset = c - 4
}
} else {
for c > 0 {
gins(x86.ASTOSB, nil, nil) // STOB AL,*(DI)+
c--
}
}
restx(&n1, &oldn1)
restx(&ax, &oldax)
}
/*
* 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:
gc.Convconst(&con, t.Type, &f.Val)
case gc.TINT32,
gc.TINT16,
gc.TINT8:
var con gc.Node
gc.Convconst(&con, gc.Types[gc.TINT64], &f.Val)
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
gc.Convconst(&con, gc.Types[gc.TUINT64], &f.Val)
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.Fatal("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
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op int, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
a := optoas(op, nl.Type)
if nr.Op == gc.OLITERAL {
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gc.Cgen(nl, &n1)
sc := uint64(gc.Mpgetfix(nr.Val.U.Xval))
if sc >= uint64(nl.Type.Width*8) {
// large shift gets 2 shifts by width-1
var n3 gc.Node
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n1)
gins(a, &n3, &n1)
} else {
gins(a, nr, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
if nl.Ullman >= gc.UINF {
var n4 gc.Node
gc.Tempname(&n4, nl.Type)
gc.Cgen(nl, &n4)
nl = &n4
}
if nr.Ullman >= gc.UINF {
var n5 gc.Node
gc.Tempname(&n5, nr.Type)
gc.Cgen(nr, &n5)
nr = &n5
}
rcx := int(reg[x86.REG_CX])
var n1 gc.Node
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
// Allow either uint32 or uint64 as shift type,
// to avoid unnecessary conversion from uint32 to uint64
// just to do the comparison.
tcount := gc.Types[gc.Simtype[nr.Type.Etype]]
if tcount.Etype < gc.TUINT32 {
tcount = gc.Types[gc.TUINT32]
}
gc.Regalloc(&n1, nr.Type, &n1) // to hold the shift type in CX
var n3 gc.Node
gc.Regalloc(&n3, tcount, &n1) // to clear high bits of CX
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TUINT64], x86.REG_CX)
var oldcx gc.Node
if rcx > 0 && !gc.Samereg(&cx, res) {
gc.Regalloc(&oldcx, gc.Types[gc.TUINT64], nil)
gmove(&cx, &oldcx)
}
cx.Type = tcount
var n2 gc.Node
if gc.Samereg(&cx, res) {
gc.Regalloc(&n2, nl.Type, nil)
} else {
gc.Regalloc(&n2, nl.Type, res)
}
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &n2)
gc.Cgen(nr, &n1)
gmove(&n1, &n3)
} else {
gc.Cgen(nr, &n1)
gmove(&n1, &n3)
gc.Cgen(nl, &n2)
}
gc.Regfree(&n3)
// test and fix up large shifts
if !bounded {
gc.Nodconst(&n3, tcount, nl.Type.Width*8)
gins(optoas(gc.OCMP, tcount), &n1, &n3)
p1 := gc.Gbranch(optoas(gc.OLT, tcount), nil, +1)
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n2)
} else {
gc.Nodconst(&n3, nl.Type, 0)
gmove(&n3, &n2)
}
gc.Patch(p1, gc.Pc)
}
gins(a, &n1, &n2)
if oldcx.Op != 0 {
cx.Type = gc.Types[gc.TUINT64]
gmove(&oldcx, &cx)
gc.Regfree(&oldcx)
}
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
func cgen_floatsse(n *gc.Node, res *gc.Node) {
var a int
nl := n.Left
nr := n.Right
switch n.Op {
default:
gc.Dump("cgen_floatsse", n)
gc.Fatal("cgen_floatsse %v", gc.Oconv(int(n.Op), 0))
return
case gc.OMINUS,
gc.OCOM:
nr = gc.Nodintconst(-1)
gc.Convlit(&nr, n.Type)
a = foptoas(gc.OMUL, nl.Type, 0)
goto sbop
// symmetric binary
case gc.OADD,
gc.OMUL:
a = foptoas(int(n.Op), nl.Type, 0)
goto sbop
// asymmetric binary
case gc.OSUB,
gc.OMOD,
gc.ODIV:
a = foptoas(int(n.Op), nl.Type, 0)
goto abop
}
sbop: // symmetric binary
if nl.Ullman < nr.Ullman || nl.Op == gc.OLITERAL {
r := nl
nl = nr
nr = r
}
abop: // asymmetric binary
if nl.Ullman >= nr.Ullman {
var nt gc.Node
gc.Tempname(&nt, nl.Type)
gc.Cgen(nl, &nt)
var n2 gc.Node
gc.Mgen(nr, &n2, nil)
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gmove(&nt, &n1)
gins(a, &n2, &n1)
gmove(&n1, res)
gc.Regfree(&n1)
gc.Mfree(&n2)
} else {
var n2 gc.Node
gc.Regalloc(&n2, nr.Type, res)
gc.Cgen(nr, &n2)
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, nil)
gc.Cgen(nl, &n1)
gins(a, &n2, &n1)
gc.Regfree(&n2)
gmove(&n1, res)
gc.Regfree(&n1)
}
return
}
func bgen_float(n *gc.Node, true_ int, likely int, to *obj.Prog) {
nl := n.Left
nr := n.Right
a := int(n.Op)
if true_ == 0 {
// brcom is not valid on floats when NaN is involved.
p1 := gc.Gbranch(obj.AJMP, nil, 0)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// No need to avoid re-genning ninit.
bgen_float(n, 1, -likely, p2)
gc.Patch(gc.Gbranch(obj.AJMP, nil, 0), to)
gc.Patch(p2, gc.Pc)
return
}
var tmp gc.Node
var et int
var n2 gc.Node
var ax gc.Node
if !gc.Thearch.Use387 {
if nl.Addable == 0 {
var n1 gc.Node
gc.Tempname(&n1, nl.Type)
gc.Cgen(nl, &n1)
nl = &n1
}
if nr.Addable == 0 {
var tmp gc.Node
gc.Tempname(&tmp, nr.Type)
gc.Cgen(nr, &tmp)
nr = &tmp
}
var n2 gc.Node
gc.Regalloc(&n2, nr.Type, nil)
gmove(nr, &n2)
nr = &n2
if nl.Op != gc.OREGISTER {
var n3 gc.Node
gc.Regalloc(&n3, nl.Type, nil)
gmove(nl, &n3)
nl = &n3
}
if a == gc.OGE || a == gc.OGT {
// only < and <= work right with NaN; reverse if needed
r := nr
nr = nl
nl = r
a = gc.Brrev(a)
}
gins(foptoas(gc.OCMP, nr.Type, 0), nl, nr)
if nl.Op == gc.OREGISTER {
gc.Regfree(nl)
}
gc.Regfree(nr)
goto ret
} else {
goto x87
}
x87:
a = gc.Brrev(a) // because the args are stacked
if a == gc.OGE || a == gc.OGT {
// only < and <= work right with NaN; reverse if needed
r := nr
nr = nl
nl = r
a = gc.Brrev(a)
}
gc.Nodreg(&tmp, nr.Type, x86.REG_F0)
gc.Nodreg(&n2, nr.Type, x86.REG_F0+1)
gc.Nodreg(&ax, gc.Types[gc.TUINT16], x86.REG_AX)
et = gc.Simsimtype(nr.Type)
if et == gc.TFLOAT64 {
if nl.Ullman > nr.Ullman {
gc.Cgen(nl, &tmp)
gc.Cgen(nr, &tmp)
gins(x86.AFXCHD, &tmp, &n2)
} else {
gc.Cgen(nr, &tmp)
gc.Cgen(nl, &tmp)
}
gins(x86.AFUCOMIP, &tmp, &n2)
gins(x86.AFMOVDP, &tmp, &tmp) // annoying pop but still better than STSW+SAHF
} else {
// TODO(rsc): The moves back and forth to memory
// here are for truncating the value to 32 bits.
// This handles 32-bit comparison but presumably
// all the other ops have the same problem.
// We need to figure out what the right general
// solution is, besides telling people to use float64.
var t1 gc.Node
gc.Tempname(&t1, gc.Types[gc.TFLOAT32])
var t2 gc.Node
gc.Tempname(&t2, gc.Types[gc.TFLOAT32])
gc.Cgen(nr, &t1)
gc.Cgen(nl, &t2)
gmove(&t2, &tmp)
gins(x86.AFCOMFP, &t1, &tmp)
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
}
goto ret
ret:
if a == gc.OEQ {
// neither NE nor P
p1 := gc.Gbranch(x86.AJNE, nil, -likely)
p2 := gc.Gbranch(x86.AJPS, nil, -likely)
gc.Patch(gc.Gbranch(obj.AJMP, nil, 0), to)
gc.Patch(p1, gc.Pc)
gc.Patch(p2, gc.Pc)
} else if a == gc.ONE {
// either NE or P
gc.Patch(gc.Gbranch(x86.AJNE, nil, likely), to)
gc.Patch(gc.Gbranch(x86.AJPS, nil, likely), to)
} else {
gc.Patch(gc.Gbranch(optoas(a, nr.Type), nil, likely), to)
}
}
/*
* generate division.
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op int, 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 := 0
if gc.Issigned[t.Etype] {
check = 1
if gc.Isconst(nl, gc.CTINT) && gc.Mpgetfix(nl.Val.U.Xval) != -(1<<uint64(t.Width*8-1)) {
check = 0
} else if gc.Isconst(nr, gc.CTINT) && gc.Mpgetfix(nr.Val.U.Xval) != -1 {
check = 0
}
}
if t.Width < 4 {
if gc.Issigned[t.Etype] {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
check = 0
}
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 != 0 {
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 != 0 {
gc.Patch(p2, gc.Pc)
}
restx(&ax, &oldax)
}
