/*
* 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)
}
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
* 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)
}
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)
}
}
/*
* 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
}
// 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)
}
func stackcopy(n, ns *gc.Node, osrc, odst, w int64) {
var noddi gc.Node
gc.Nodreg(&noddi, gc.Types[gc.Tptr], x86.REG_DI)
var nodsi gc.Node
gc.Nodreg(&nodsi, gc.Types[gc.Tptr], x86.REG_SI)
var nodl gc.Node
var nodr gc.Node
if n.Ullman >= ns.Ullman {
gc.Agenr(n, &nodr, &nodsi)
if ns.Op == gc.ONAME {
gc.Gvardef(ns)
}
gc.Agenr(ns, &nodl, &noddi)
} else {
if ns.Op == gc.ONAME {
gc.Gvardef(ns)
}
gc.Agenr(ns, &nodl, &noddi)
gc.Agenr(n, &nodr, &nodsi)
}
if nodl.Reg != x86.REG_DI {
gmove(&nodl, &noddi)
}
if nodr.Reg != x86.REG_SI {
gmove(&nodr, &nodsi)
}
gc.Regfree(&nodl)
gc.Regfree(&nodr)
c := w % 8 // bytes
q := w / 8 // quads
var oldcx gc.Node
var cx gc.Node
savex(x86.REG_CX, &cx, &oldcx, nil, gc.Types[gc.TINT64])
// if we are copying forward on the stack and
// the src and dst overlap, then reverse direction
if osrc < odst && odst < osrc+w {
// reverse direction
gins(x86.ASTD, nil, nil) // set direction flag
if c > 0 {
gconreg(addptr, w-1, x86.REG_SI)
gconreg(addptr, w-1, x86.REG_DI)
gconreg(movptr, c, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSB, nil, nil) // MOVB *(SI)-,*(DI)-
}
if q > 0 {
if c > 0 {
gconreg(addptr, -7, x86.REG_SI)
gconreg(addptr, -7, x86.REG_DI)
} else {
gconreg(addptr, w-8, x86.REG_SI)
gconreg(addptr, w-8, x86.REG_DI)
}
gconreg(movptr, q, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)-,*(DI)-
}
// we leave with the flag clear
gins(x86.ACLD, nil, nil)
} else {
// normal direction
if q > 128 || (gc.Nacl && q >= 4) {
gconreg(movptr, q, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)+,*(DI)+
} else if q >= 4 {
p := gins(obj.ADUFFCOPY, nil, nil)
p.To.Type = obj.TYPE_ADDR
p.To.Sym = gc.Linksym(gc.Pkglookup("duffcopy", gc.Runtimepkg))
// 14 and 128 = magic constants: see ../../runtime/asm_amd64.s
p.To.Offset = 14 * (128 - q)
} else if !gc.Nacl && c == 0 {
// We don't need the MOVSQ side-effect of updating SI and DI,
// and issuing a sequence of MOVQs directly is faster.
nodsi.Op = gc.OINDREG
noddi.Op = gc.OINDREG
for q > 0 {
gmove(&nodsi, &cx) // MOVQ x+(SI),CX
gmove(&cx, &noddi) // MOVQ CX,x+(DI)
nodsi.Xoffset += 8
noddi.Xoffset += 8
q--
}
} else {
for q > 0 {
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)+,*(DI)+
q--
}
}
// copy the remaining c bytes
if w < 4 || c <= 1 || (odst < osrc && osrc < odst+w) {
for c > 0 {
gins(x86.AMOVSB, nil, nil) // MOVB *(SI)+,*(DI)+
c--
}
} else if w < 8 || c <= 4 {
nodsi.Op = gc.OINDREG
noddi.Op = gc.OINDREG
cx.Type = gc.Types[gc.TINT32]
nodsi.Type = gc.Types[gc.TINT32]
noddi.Type = gc.Types[gc.TINT32]
if c > 4 {
nodsi.Xoffset = 0
noddi.Xoffset = 0
gmove(&nodsi, &cx)
gmove(&cx, &noddi)
}
nodsi.Xoffset = c - 4
noddi.Xoffset = c - 4
gmove(&nodsi, &cx)
gmove(&cx, &noddi)
} else {
nodsi.Op = gc.OINDREG
noddi.Op = gc.OINDREG
cx.Type = gc.Types[gc.TINT64]
nodsi.Type = gc.Types[gc.TINT64]
noddi.Type = gc.Types[gc.TINT64]
nodsi.Xoffset = c - 8
noddi.Xoffset = c - 8
gmove(&nodsi, &cx)
gmove(&cx, &noddi)
}
}
restx(&cx, &oldcx)
}
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 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)
}
/*
* 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)
}
/*
* 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
gc.Convconst(&con, t.Type, &f.Val)
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 gc.Mpcmpfixfix(con.Val.U.Xval, gc.Minintval[gc.TINT32]) < 0 {
goto hard
}
if gc.Mpcmpfixfix(con.Val.U.Xval, gc.Maxintval[gc.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.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
}