func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if cnt < int64(4*gc.Widthptr) {
for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
p = appendpp(p, ppc64.AMOVD, obj.TYPE_REG, ppc64.REGZERO, 0, obj.TYPE_MEM, ppc64.REGSP, gc.Ctxt.FixedFrameSize()+frame+lo+i)
}
} else if cnt <= int64(128*gc.Widthptr) {
p = appendpp(p, ppc64.AADD, obj.TYPE_CONST, 0, gc.Ctxt.FixedFrameSize()+frame+lo-8, obj.TYPE_REG, ppc64.REGRT1, 0)
p.Reg = ppc64.REGSP
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
f := gc.Sysfunc("duffzero")
gc.Naddr(&p.To, f)
gc.Afunclit(&p.To, f)
p.To.Offset = 4 * (128 - cnt/int64(gc.Widthptr))
} else {
p = appendpp(p, ppc64.AMOVD, obj.TYPE_CONST, 0, gc.Ctxt.FixedFrameSize()+frame+lo-8, obj.TYPE_REG, ppc64.REGTMP, 0)
p = appendpp(p, ppc64.AADD, obj.TYPE_REG, ppc64.REGTMP, 0, obj.TYPE_REG, ppc64.REGRT1, 0)
p.Reg = ppc64.REGSP
p = appendpp(p, ppc64.AMOVD, obj.TYPE_CONST, 0, cnt, obj.TYPE_REG, ppc64.REGTMP, 0)
p = appendpp(p, ppc64.AADD, obj.TYPE_REG, ppc64.REGTMP, 0, obj.TYPE_REG, ppc64.REGRT2, 0)
p.Reg = ppc64.REGRT1
p = appendpp(p, ppc64.AMOVDU, obj.TYPE_REG, ppc64.REGZERO, 0, obj.TYPE_MEM, ppc64.REGRT1, int64(gc.Widthptr))
p1 := p
p = appendpp(p, ppc64.ACMP, obj.TYPE_REG, ppc64.REGRT1, 0, obj.TYPE_REG, ppc64.REGRT2, 0)
p = appendpp(p, ppc64.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
gc.Patch(p, p1)
}
return p
}
/*
* generate division.
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op gc.Op, nl *gc.Node, nr *gc.Node, res *gc.Node) {
t := nl.Type
t0 := t
if t.Width < 8 {
if t.IsSigned() {
t = gc.Types[gc.TINT64]
} else {
t = gc.Types[gc.TUINT64]
}
}
a := optoas(gc.ODIV, t)
var tl gc.Node
gc.Regalloc(&tl, t0, nil)
var tr gc.Node
gc.Regalloc(&tr, t0, nil)
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &tl)
gc.Cgen(nr, &tr)
} else {
gc.Cgen(nr, &tr)
gc.Cgen(nl, &tl)
}
if t != t0 {
// Convert
tl2 := tl
tr2 := tr
tl.Type = t
tr.Type = t
gmove(&tl2, &tl)
gmove(&tr2, &tr)
}
// Handle divide-by-zero panic.
p1 := ginsbranch(mips.ABNE, nil, &tr, nil, 0)
if panicdiv == nil {
panicdiv = gc.Sysfunc("panicdivide")
}
gc.Ginscall(panicdiv, -1)
gc.Patch(p1, gc.Pc)
gins3(a, &tr, &tl, nil)
gc.Regfree(&tr)
if op == gc.ODIV {
var lo gc.Node
gc.Nodreg(&lo, gc.Types[gc.TUINT64], mips.REG_LO)
gins(mips.AMOVV, &lo, &tl)
} else { // remainder in REG_HI
var hi gc.Node
gc.Nodreg(&hi, gc.Types[gc.TUINT64], mips.REG_HI)
gins(mips.AMOVV, &hi, &tl)
}
gmove(&tl, res)
gc.Regfree(&tl)
}
/*
* generate floating-point operation.
*/
func cgen_float(n *gc.Node, res *gc.Node) {
nl := n.Left
switch n.Op {
case gc.OEQ,
gc.ONE,
gc.OLT,
gc.OLE,
gc.OGE:
p1 := gc.Gbranch(obj.AJMP, nil, 0)
p2 := gc.Pc
gmove(gc.Nodbool(true), res)
p3 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
gc.Bgen(n, true, 0, p2)
gmove(gc.Nodbool(false), res)
gc.Patch(p3, gc.Pc)
return
case gc.OPLUS:
gc.Cgen(nl, res)
return
case gc.OCONV:
if gc.Eqtype(n.Type, nl.Type) || gc.Noconv(n.Type, nl.Type) {
gc.Cgen(nl, res)
return
}
var n2 gc.Node
gc.Tempname(&n2, n.Type)
var n1 gc.Node
gc.Mgen(nl, &n1, res)
gmove(&n1, &n2)
gmove(&n2, res)
gc.Mfree(&n1)
return
}
if gc.Thearch.Use387 {
cgen_float387(n, res)
} else {
cgen_floatsse(n, res)
}
}
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if cnt < int64(4*gc.Widthptr) {
for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
p = appendpp(p, mips.AMOVV, obj.TYPE_REG, mips.REGZERO, 0, obj.TYPE_MEM, mips.REGSP, 8+frame+lo+i)
}
// 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 && cnt <= int64(128*gc.Widthptr) {
p = appendpp(p, mips.AADDV, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, mips.REGRT1, 0)
p.Reg = mips.REGSP
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
f := gc.Sysfunc("duffzero")
gc.Naddr(&p.To, f)
gc.Afunclit(&p.To, f)
p.To.Offset = 8 * (128 - cnt/int64(gc.Widthptr))
} else {
// ADDV $(8+frame+lo-8), SP, r1
// ADDV $cnt, r1, r2
// loop:
// MOVV R0, (Widthptr)r1
// ADDV $Widthptr, r1
// BNE r1, r2, loop
p = appendpp(p, mips.AADDV, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, mips.REGRT1, 0)
p.Reg = mips.REGSP
p = appendpp(p, mips.AADDV, obj.TYPE_CONST, 0, cnt, obj.TYPE_REG, mips.REGRT2, 0)
p.Reg = mips.REGRT1
p = appendpp(p, mips.AMOVV, obj.TYPE_REG, mips.REGZERO, 0, obj.TYPE_MEM, mips.REGRT1, int64(gc.Widthptr))
p1 := p
p = appendpp(p, mips.AADDV, obj.TYPE_CONST, 0, int64(gc.Widthptr), obj.TYPE_REG, mips.REGRT1, 0)
p = appendpp(p, mips.ABNE, obj.TYPE_REG, mips.REGRT1, 0, obj.TYPE_BRANCH, 0, 0)
p.Reg = mips.REGRT2
gc.Patch(p, p1)
}
return p
}
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64, r0 *uint32) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if *r0 == 0 {
p = appendpp(p, arm.AMOVW, obj.TYPE_CONST, 0, 0, obj.TYPE_REG, arm.REG_R0, 0)
*r0 = 1
}
if cnt < int64(4*gc.Widthptr) {
for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
p = appendpp(p, arm.AMOVW, obj.TYPE_REG, arm.REG_R0, 0, obj.TYPE_MEM, arm.REGSP, int32(4+frame+lo+i))
}
} else if !gc.Nacl && (cnt <= int64(128*gc.Widthptr)) {
p = appendpp(p, arm.AADD, obj.TYPE_CONST, 0, int32(4+frame+lo), obj.TYPE_REG, arm.REG_R1, 0)
p.Reg = arm.REGSP
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
f := gc.Sysfunc("duffzero")
gc.Naddr(&p.To, f)
gc.Afunclit(&p.To, f)
p.To.Offset = 4 * (128 - cnt/int64(gc.Widthptr))
} else {
p = appendpp(p, arm.AADD, obj.TYPE_CONST, 0, int32(4+frame+lo), obj.TYPE_REG, arm.REG_R1, 0)
p.Reg = arm.REGSP
p = appendpp(p, arm.AADD, obj.TYPE_CONST, 0, int32(cnt), obj.TYPE_REG, arm.REG_R2, 0)
p.Reg = arm.REG_R1
p = appendpp(p, arm.AMOVW, obj.TYPE_REG, arm.REG_R0, 0, obj.TYPE_MEM, arm.REG_R1, 4)
p1 := p
p.Scond |= arm.C_PBIT
p = appendpp(p, arm.ACMP, obj.TYPE_REG, arm.REG_R1, 0, obj.TYPE_NONE, 0, 0)
p.Reg = arm.REG_R2
p = appendpp(p, arm.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
gc.Patch(p, p1)
}
return p
}
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if cnt < int64(4*gc.Widthptr) {
for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGSP, 8+frame+lo+i)
}
} else if cnt <= int64(128*gc.Widthptr) && !darwin { // darwin ld64 cannot handle BR26 reloc with non-zero addend
p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGSP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
p = appendpp(p, arm64.AADD, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, arm64.REGRT1, 0)
p.Reg = arm64.REGRT1
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
f := gc.Sysfunc("duffzero")
gc.Naddr(&p.To, f)
gc.Afunclit(&p.To, f)
p.To.Offset = 4 * (128 - cnt/int64(gc.Widthptr))
} else {
p = appendpp(p, arm64.AMOVD, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, arm64.REGTMP, 0)
p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGSP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
p = appendpp(p, arm64.AADD, obj.TYPE_REG, arm64.REGTMP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
p.Reg = arm64.REGRT1
p = appendpp(p, arm64.AMOVD, obj.TYPE_CONST, 0, cnt, obj.TYPE_REG, arm64.REGTMP, 0)
p = appendpp(p, arm64.AADD, obj.TYPE_REG, arm64.REGTMP, 0, obj.TYPE_REG, arm64.REGRT2, 0)
p.Reg = arm64.REGRT1
p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGRT1, int64(gc.Widthptr))
p.Scond = arm64.C_XPRE
p1 := p
p = appendpp(p, arm64.ACMP, obj.TYPE_REG, arm64.REGRT1, 0, obj.TYPE_NONE, 0, 0)
p.Reg = arm64.REGRT2
p = appendpp(p, arm64.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
gc.Patch(p, p1)
}
return p
}
/*
* generate move:
* t = f
* hard part is conversions.
*/
func gmove(f *gc.Node, t *gc.Node) {
if gc.Debug['M'] != 0 {
fmt.Printf("gmove %v -> %v\n", gc.Nconv(f, gc.FmtLong), gc.Nconv(t, gc.FmtLong))
}
ft := int(gc.Simsimtype(f.Type))
tt := int(gc.Simsimtype(t.Type))
cvt := t.Type
if gc.Iscomplex[ft] || gc.Iscomplex[tt] {
gc.Complexmove(f, t)
return
}
// cannot have two memory operands
var r2 gc.Node
var r1 gc.Node
var a obj.As
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
switch tt {
default:
f.Convconst(&con, t.Type)
case gc.TINT32,
gc.TINT16,
gc.TINT8:
var con gc.Node
f.Convconst(&con, gc.Types[gc.TINT64])
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(mips.AMOVV, &con, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
case gc.TUINT32,
gc.TUINT16,
gc.TUINT8:
var con gc.Node
f.Convconst(&con, gc.Types[gc.TUINT64])
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(mips.AMOVV, &con, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
f = &con
ft = tt // so big switch will choose a simple mov
// constants can't move directly to memory.
if gc.Ismem(t) {
goto hard
}
}
// value -> value copy, first operand in memory.
// any floating point operand requires register
// src, so goto hard to copy to register first.
if gc.Ismem(f) && ft != tt && (gc.Isfloat[ft] || gc.Isfloat[tt]) {
cvt = gc.Types[ft]
goto hard
}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
gc.Fatalf("gmove %v -> %v", gc.Tconv(f.Type, gc.FmtLong), gc.Tconv(t.Type, gc.FmtLong))
/*
* integer copy and truncate
*/
case gc.TINT8<<16 | gc.TINT8, // same size
gc.TUINT8<<16 | gc.TINT8,
gc.TINT16<<16 | gc.TINT8, // truncate
gc.TUINT16<<16 | gc.TINT8,
gc.TINT32<<16 | gc.TINT8,
gc.TUINT32<<16 | gc.TINT8,
gc.TINT64<<16 | gc.TINT8,
gc.TUINT64<<16 | gc.TINT8:
a = mips.AMOVB
case gc.TINT8<<16 | gc.TUINT8, // same size
gc.TUINT8<<16 | gc.TUINT8,
gc.TINT16<<16 | gc.TUINT8, // truncate
gc.TUINT16<<16 | gc.TUINT8,
gc.TINT32<<16 | gc.TUINT8,
gc.TUINT32<<16 | gc.TUINT8,
gc.TINT64<<16 | gc.TUINT8,
gc.TUINT64<<16 | gc.TUINT8:
a = mips.AMOVBU
case gc.TINT16<<16 | gc.TINT16, // same size
gc.TUINT16<<16 | gc.TINT16,
gc.TINT32<<16 | gc.TINT16, // truncate
gc.TUINT32<<16 | gc.TINT16,
gc.TINT64<<16 | gc.TINT16,
gc.TUINT64<<16 | gc.TINT16:
a = mips.AMOVH
case gc.TINT16<<16 | gc.TUINT16, // same size
gc.TUINT16<<16 | gc.TUINT16,
gc.TINT32<<16 | gc.TUINT16, // truncate
gc.TUINT32<<16 | gc.TUINT16,
gc.TINT64<<16 | gc.TUINT16,
gc.TUINT64<<16 | gc.TUINT16:
a = mips.AMOVHU
case gc.TINT32<<16 | gc.TINT32, // same size
gc.TUINT32<<16 | gc.TINT32,
gc.TINT64<<16 | gc.TINT32, // truncate
gc.TUINT64<<16 | gc.TINT32:
a = mips.AMOVW
case gc.TINT32<<16 | gc.TUINT32, // same size
gc.TUINT32<<16 | gc.TUINT32,
gc.TINT64<<16 | gc.TUINT32, // truncate
gc.TUINT64<<16 | gc.TUINT32:
a = mips.AMOVWU
case gc.TINT64<<16 | gc.TINT64, // same size
gc.TINT64<<16 | gc.TUINT64,
gc.TUINT64<<16 | gc.TINT64,
gc.TUINT64<<16 | gc.TUINT64:
a = mips.AMOVV
/*
* integer up-conversions
*/
case gc.TINT8<<16 | gc.TINT16, // sign extend int8
gc.TINT8<<16 | gc.TUINT16,
gc.TINT8<<16 | gc.TINT32,
gc.TINT8<<16 | gc.TUINT32,
gc.TINT8<<16 | gc.TINT64,
gc.TINT8<<16 | gc.TUINT64:
a = mips.AMOVB
goto rdst
case gc.TUINT8<<16 | gc.TINT16, // zero extend uint8
gc.TUINT8<<16 | gc.TUINT16,
gc.TUINT8<<16 | gc.TINT32,
gc.TUINT8<<16 | gc.TUINT32,
gc.TUINT8<<16 | gc.TINT64,
gc.TUINT8<<16 | gc.TUINT64:
a = mips.AMOVBU
goto rdst
case gc.TINT16<<16 | gc.TINT32, // sign extend int16
gc.TINT16<<16 | gc.TUINT32,
gc.TINT16<<16 | gc.TINT64,
gc.TINT16<<16 | gc.TUINT64:
a = mips.AMOVH
goto rdst
case gc.TUINT16<<16 | gc.TINT32, // zero extend uint16
gc.TUINT16<<16 | gc.TUINT32,
gc.TUINT16<<16 | gc.TINT64,
gc.TUINT16<<16 | gc.TUINT64:
a = mips.AMOVHU
goto rdst
case gc.TINT32<<16 | gc.TINT64, // sign extend int32
gc.TINT32<<16 | gc.TUINT64:
a = mips.AMOVW
goto rdst
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
a = mips.AMOVWU
goto rdst
//warn("gmove: convert float to int not implemented: %N -> %N\n", f, t);
//return;
// algorithm is:
// if small enough, use native float64 -> int64 conversion.
// otherwise, subtract 2^63, convert, and add it back.
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT32,
gc.TFLOAT64<<16 | gc.TINT32,
gc.TFLOAT32<<16 | gc.TINT64,
gc.TFLOAT64<<16 | gc.TINT64,
gc.TFLOAT32<<16 | gc.TINT16,
gc.TFLOAT32<<16 | gc.TINT8,
gc.TFLOAT32<<16 | gc.TUINT16,
gc.TFLOAT32<<16 | gc.TUINT8,
gc.TFLOAT64<<16 | gc.TINT16,
gc.TFLOAT64<<16 | gc.TINT8,
gc.TFLOAT64<<16 | gc.TUINT16,
gc.TFLOAT64<<16 | gc.TUINT8,
gc.TFLOAT32<<16 | gc.TUINT32,
gc.TFLOAT64<<16 | gc.TUINT32,
gc.TFLOAT32<<16 | gc.TUINT64,
gc.TFLOAT64<<16 | gc.TUINT64:
bignodes()
gc.Regalloc(&r1, gc.Types[gc.TFLOAT64], nil)
gmove(f, &r1)
if tt == gc.TUINT64 {
gc.Regalloc(&r2, gc.Types[gc.TFLOAT64], nil)
gmove(&bigf, &r2)
gins3(mips.ACMPGED, &r1, &r2, nil)
p1 := gc.Gbranch(mips.ABFPF, nil, 0)
gins(mips.ASUBD, &r2, &r1)
gc.Patch(p1, gc.Pc)
gc.Regfree(&r2)
}
gc.Regalloc(&r2, gc.Types[gc.TINT64], t)
gins(mips.ATRUNCDV, &r1, &r1)
gins(mips.AMOVV, &r1, &r2)
gc.Regfree(&r1)
if tt == gc.TUINT64 {
p1 := gc.Gbranch(mips.ABFPF, nil, 0) // use FCR0 here again
gc.Nodreg(&r1, gc.Types[gc.TINT64], mips.REGTMP)
gmove(&bigi, &r1)
gins(mips.AADDVU, &r1, &r2)
gc.Patch(p1, gc.Pc)
}
gmove(&r2, t)
gc.Regfree(&r2)
return
//warn("gmove: convert int to float not implemented: %N -> %N\n", f, t);
//return;
// algorithm is:
// if small enough, use native int64 -> float64 conversion.
// otherwise, halve (x -> (x>>1)|(x&1)), convert, and double.
/*
* integer to float
*/
case gc.TINT32<<16 | gc.TFLOAT32,
gc.TINT32<<16 | gc.TFLOAT64,
gc.TINT64<<16 | gc.TFLOAT32,
gc.TINT64<<16 | gc.TFLOAT64,
gc.TINT16<<16 | gc.TFLOAT32,
gc.TINT16<<16 | gc.TFLOAT64,
gc.TINT8<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TUINT16<<16 | gc.TFLOAT32,
gc.TUINT16<<16 | gc.TFLOAT64,
gc.TUINT8<<16 | gc.TFLOAT32,
gc.TUINT8<<16 | gc.TFLOAT64,
gc.TUINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT64,
gc.TUINT64<<16 | gc.TFLOAT32,
gc.TUINT64<<16 | gc.TFLOAT64:
bignodes()
var rtmp gc.Node
gc.Regalloc(&r1, gc.Types[gc.TINT64], nil)
gmove(f, &r1)
if ft == gc.TUINT64 {
gc.Nodreg(&rtmp, gc.Types[gc.TUINT64], mips.REGTMP)
gmove(&bigi, &rtmp)
gins(mips.AAND, &r1, &rtmp)
p1 := ginsbranch(mips.ABEQ, nil, &rtmp, nil, 0)
var r3 gc.Node
gc.Regalloc(&r3, gc.Types[gc.TUINT64], nil)
p2 := gins3(mips.AAND, nil, &r1, &r3)
p2.From.Type = obj.TYPE_CONST
p2.From.Offset = 1
p3 := gins(mips.ASRLV, nil, &r1)
p3.From.Type = obj.TYPE_CONST
p3.From.Offset = 1
gins(mips.AOR, &r3, &r1)
gc.Regfree(&r3)
gc.Patch(p1, gc.Pc)
}
gc.Regalloc(&r2, gc.Types[gc.TFLOAT64], t)
gins(mips.AMOVV, &r1, &r2)
gins(mips.AMOVVD, &r2, &r2)
gc.Regfree(&r1)
if ft == gc.TUINT64 {
p1 := ginsbranch(mips.ABEQ, nil, &rtmp, nil, 0)
gc.Nodreg(&r1, gc.Types[gc.TFLOAT64], mips.FREGTWO)
gins(mips.AMULD, &r1, &r2)
gc.Patch(p1, gc.Pc)
}
gmove(&r2, t)
gc.Regfree(&r2)
return
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32:
a = mips.AMOVF
case gc.TFLOAT64<<16 | gc.TFLOAT64:
a = mips.AMOVD
case gc.TFLOAT32<<16 | gc.TFLOAT64:
a = mips.AMOVFD
goto rdst
case gc.TFLOAT64<<16 | gc.TFLOAT32:
a = mips.AMOVDF
goto rdst
}
gins(a, f, t)
return
// requires register destination
rdst:
{
gc.Regalloc(&r1, t.Type, t)
gins(a, f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
gc.Dump("\nclearfat", nl)
}
w := uint32(nl.Type.Width)
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := w % 4 // bytes
q := w / 4 // quads
if nl.Type.Align < 4 {
q = 0
c = w
}
var r0 gc.Node
r0.Op = gc.OREGISTER
r0.Reg = arm.REG_R0
var r1 gc.Node
r1.Op = gc.OREGISTER
r1.Reg = arm.REG_R1
var dst gc.Node
gc.Regalloc(&dst, gc.Types[gc.Tptr], &r1)
gc.Agen(nl, &dst)
var nc gc.Node
gc.Nodconst(&nc, gc.Types[gc.TUINT32], 0)
var nz gc.Node
gc.Regalloc(&nz, gc.Types[gc.TUINT32], &r0)
gc.Cgen(&nc, &nz)
if q > 128 {
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p := gins(arm.AMOVW, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(q) * 4
p = gins(arm.AMOVW, &nz, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 4
p.Scond |= arm.C_PBIT
pl := p
p = gins(arm.ACMP, &dst, nil)
raddr(&end, p)
gc.Patch(gc.Gbranch(arm.ABNE, nil, 0), pl)
gc.Regfree(&end)
} else if q >= 4 && !gc.Nacl {
f := gc.Sysfunc("duffzero")
p := gins(obj.ADUFFZERO, nil, f)
gc.Afunclit(&p.To, f)
// 4 and 128 = magic constants: see ../../runtime/asm_arm.s
p.To.Offset = 4 * (128 - int64(q))
} else {
var p *obj.Prog
for q > 0 {
p = gins(arm.AMOVW, &nz, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 4
p.Scond |= arm.C_PBIT
//print("1. %v\n", p);
q--
}
}
if c > 4 {
// Loop to zero unaligned memory.
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p := gins(arm.AMOVW, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(c)
p = gins(arm.AMOVB, &nz, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 1
p.Scond |= arm.C_PBIT
pl := p
p = gins(arm.ACMP, &dst, nil)
raddr(&end, p)
gc.Patch(gc.Gbranch(arm.ABNE, nil, 0), pl)
gc.Regfree(&end)
c = 0
}
var p *obj.Prog
for c > 0 {
p = gins(arm.AMOVB, &nz, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 1
p.Scond |= arm.C_PBIT
//print("2. %v\n", p);
c--
}
gc.Regfree(&dst)
gc.Regfree(&nz)
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op gc.Op, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Type.Width > 4 {
gc.Fatalf("cgen_shift %v", nl.Type)
}
w := int(nl.Type.Width * 8)
if op == gc.OLROT {
v := nr.Int64()
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
if w == 32 {
gc.Cgen(nl, &n1)
gshift(arm.AMOVW, &n1, arm.SHIFT_RR, int32(w)-int32(v), &n1)
} else {
var n2 gc.Node
gc.Regalloc(&n2, nl.Type, nil)
gc.Cgen(nl, &n2)
gshift(arm.AMOVW, &n2, arm.SHIFT_LL, int32(v), &n1)
gshift(arm.AORR, &n2, arm.SHIFT_LR, int32(w)-int32(v), &n1)
gc.Regfree(&n2)
// Ensure sign/zero-extended result.
gins(optoas(gc.OAS, nl.Type), &n1, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
if nr.Op == gc.OLITERAL {
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gc.Cgen(nl, &n1)
sc := uint64(nr.Int64())
if sc == 0 {
} else // nothing to do
if sc >= uint64(nl.Type.Width*8) {
if op == gc.ORSH && nl.Type.IsSigned() {
gshift(arm.AMOVW, &n1, arm.SHIFT_AR, int32(w), &n1)
} else {
gins(arm.AEOR, &n1, &n1)
}
} else {
if op == gc.ORSH && nl.Type.IsSigned() {
gshift(arm.AMOVW, &n1, arm.SHIFT_AR, int32(sc), &n1)
} else if op == gc.ORSH {
gshift(arm.AMOVW, &n1, arm.SHIFT_LR, int32(sc), &n1) // OLSH
} else {
gshift(arm.AMOVW, &n1, arm.SHIFT_LL, int32(sc), &n1)
}
}
if w < 32 && op == gc.OLSH {
gins(optoas(gc.OAS, nl.Type), &n1, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
tr := nr.Type
var t gc.Node
var n1 gc.Node
var n2 gc.Node
var n3 gc.Node
if tr.Width > 4 {
var nt gc.Node
gc.Tempname(&nt, nr.Type)
if nl.Ullman >= nr.Ullman {
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
gc.Cgen(nr, &nt)
n1 = nt
} else {
gc.Cgen(nr, &nt)
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
}
var hi gc.Node
var lo gc.Node
split64(&nt, &lo, &hi)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], nil)
gc.Regalloc(&n3, gc.Types[gc.TUINT32], nil)
gmove(&lo, &n1)
gmove(&hi, &n3)
splitclean()
gins(arm.ATST, &n3, nil)
gc.Nodconst(&t, gc.Types[gc.TUINT32], int64(w))
p1 := gins(arm.AMOVW, &t, &n1)
p1.Scond = arm.C_SCOND_NE
tr = gc.Types[gc.TUINT32]
gc.Regfree(&n3)
} else {
if nl.Ullman >= nr.Ullman {
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
gc.Regalloc(&n1, nr.Type, nil)
gc.Cgen(nr, &n1)
} else {
gc.Regalloc(&n1, nr.Type, nil)
gc.Cgen(nr, &n1)
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
}
}
// test for shift being 0
gins(arm.ATST, &n1, nil)
p3 := gc.Gbranch(arm.ABEQ, nil, -1)
// test and fix up large shifts
// TODO: if(!bounded), don't emit some of this.
gc.Regalloc(&n3, tr, nil)
gc.Nodconst(&t, gc.Types[gc.TUINT32], int64(w))
gmove(&t, &n3)
gins(arm.ACMP, &n1, &n3)
if op == gc.ORSH {
var p1 *obj.Prog
var p2 *obj.Prog
if nl.Type.IsSigned() {
p1 = gshift(arm.AMOVW, &n2, arm.SHIFT_AR, int32(w)-1, &n2)
p2 = gregshift(arm.AMOVW, &n2, arm.SHIFT_AR, &n1, &n2)
} else {
p1 = gins(arm.AEOR, &n2, &n2)
p2 = gregshift(arm.AMOVW, &n2, arm.SHIFT_LR, &n1, &n2)
}
p1.Scond = arm.C_SCOND_HS
p2.Scond = arm.C_SCOND_LO
} else {
p1 := gins(arm.AEOR, &n2, &n2)
p2 := gregshift(arm.AMOVW, &n2, arm.SHIFT_LL, &n1, &n2)
p1.Scond = arm.C_SCOND_HS
p2.Scond = arm.C_SCOND_LO
}
gc.Regfree(&n3)
gc.Patch(p3, gc.Pc)
// Left-shift of smaller word must be sign/zero-extended.
if w < 32 && op == gc.OLSH {
gins(optoas(gc.OAS, nl.Type), &n2, &n2)
}
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
func floatmove(f *gc.Node, t *gc.Node) {
var r1 gc.Node
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
cvt := t.Type
// cannot have two floating point memory operands.
if gc.Isfloat[ft] && gc.Isfloat[tt] && gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
f.Convconst(&con, t.Type)
f = &con
ft = gc.Simsimtype(con.Type)
// some constants can't move directly to memory.
if gc.Ismem(t) {
// float constants come from memory.
if gc.Isfloat[tt] {
goto hard
}
}
}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
if gc.Thearch.Use387 {
floatmove_387(f, t)
} else {
floatmove_sse(f, t)
}
return
// float to very long integer.
case gc.TFLOAT32<<16 | gc.TINT64,
gc.TFLOAT64<<16 | gc.TINT64:
if f.Op == gc.OREGISTER {
cvt = f.Type
goto hardmem
}
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[ft], x86.REG_F0)
if ft == gc.TFLOAT32 {
gins(x86.AFMOVF, f, &r1)
} else {
gins(x86.AFMOVD, f, &r1)
}
// set round to zero mode during conversion
var t1 gc.Node
memname(&t1, gc.Types[gc.TUINT16])
var t2 gc.Node
memname(&t2, gc.Types[gc.TUINT16])
gins(x86.AFSTCW, nil, &t1)
gins(x86.AMOVW, ncon(0xf7f), &t2)
gins(x86.AFLDCW, &t2, nil)
if tt == gc.TINT16 {
gins(x86.AFMOVWP, &r1, t)
} else if tt == gc.TINT32 {
gins(x86.AFMOVLP, &r1, t)
} else {
gins(x86.AFMOVVP, &r1, t)
}
gins(x86.AFLDCW, &t1, nil)
return
case gc.TFLOAT32<<16 | gc.TUINT64,
gc.TFLOAT64<<16 | gc.TUINT64:
if !gc.Ismem(f) {
cvt = f.Type
goto hardmem
}
bignodes()
var f0 gc.Node
gc.Nodreg(&f0, gc.Types[ft], x86.REG_F0)
var f1 gc.Node
gc.Nodreg(&f1, gc.Types[ft], x86.REG_F0+1)
var ax gc.Node
gc.Nodreg(&ax, gc.Types[gc.TUINT16], x86.REG_AX)
if ft == gc.TFLOAT32 {
gins(x86.AFMOVF, f, &f0)
} else {
gins(x86.AFMOVD, f, &f0)
}
// if 0 > v { answer = 0 }
gins(x86.AFMOVD, &zerof, &f0)
gins(x86.AFUCOMP, &f0, &f1)
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
p1 := gc.Gbranch(optoas(gc.OGT, gc.Types[tt]), nil, 0)
// if 1<<64 <= v { answer = 0 too }
gins(x86.AFMOVD, &two64f, &f0)
gins(x86.AFUCOMP, &f0, &f1)
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
p2 := gc.Gbranch(optoas(gc.OGT, gc.Types[tt]), nil, 0)
gc.Patch(p1, gc.Pc)
gins(x86.AFMOVVP, &f0, t) // don't care about t, but will pop the stack
var thi gc.Node
var tlo gc.Node
split64(t, &tlo, &thi)
gins(x86.AMOVL, ncon(0), &tlo)
gins(x86.AMOVL, ncon(0), &thi)
splitclean()
p1 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p2, gc.Pc)
// in range; algorithm is:
// if small enough, use native float64 -> int64 conversion.
// otherwise, subtract 2^63, convert, and add it back.
// set round to zero mode during conversion
var t1 gc.Node
memname(&t1, gc.Types[gc.TUINT16])
var t2 gc.Node
memname(&t2, gc.Types[gc.TUINT16])
gins(x86.AFSTCW, nil, &t1)
gins(x86.AMOVW, ncon(0xf7f), &t2)
gins(x86.AFLDCW, &t2, nil)
// actual work
gins(x86.AFMOVD, &two63f, &f0)
gins(x86.AFUCOMP, &f0, &f1)
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
p2 = gc.Gbranch(optoas(gc.OLE, gc.Types[tt]), nil, 0)
gins(x86.AFMOVVP, &f0, t)
p3 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p2, gc.Pc)
gins(x86.AFMOVD, &two63f, &f0)
gins(x86.AFSUBDP, &f0, &f1)
gins(x86.AFMOVVP, &f0, t)
split64(t, &tlo, &thi)
gins(x86.AXORL, ncon(0x80000000), &thi) // + 2^63
gc.Patch(p3, gc.Pc)
splitclean()
// restore rounding mode
gins(x86.AFLDCW, &t1, nil)
gc.Patch(p1, gc.Pc)
return
/*
* integer to float
*/
case gc.TINT64<<16 | gc.TFLOAT32,
gc.TINT64<<16 | gc.TFLOAT64:
if t.Op == gc.OREGISTER {
goto hardmem
}
var f0 gc.Node
gc.Nodreg(&f0, t.Type, x86.REG_F0)
gins(x86.AFMOVV, f, &f0)
if tt == gc.TFLOAT32 {
gins(x86.AFMOVFP, &f0, t)
} else {
gins(x86.AFMOVDP, &f0, t)
}
return
// algorithm is:
// if small enough, use native int64 -> float64 conversion.
// otherwise, halve (rounding to odd?), convert, and double.
case gc.TUINT64<<16 | gc.TFLOAT32,
gc.TUINT64<<16 | gc.TFLOAT64:
var ax gc.Node
gc.Nodreg(&ax, gc.Types[gc.TUINT32], x86.REG_AX)
var dx gc.Node
gc.Nodreg(&dx, gc.Types[gc.TUINT32], x86.REG_DX)
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TUINT32], x86.REG_CX)
var t1 gc.Node
gc.Tempname(&t1, f.Type)
var tlo gc.Node
var thi gc.Node
split64(&t1, &tlo, &thi)
gmove(f, &t1)
gins(x86.ACMPL, &thi, ncon(0))
p1 := gc.Gbranch(x86.AJLT, nil, 0)
// native
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[tt], x86.REG_F0)
gins(x86.AFMOVV, &t1, &r1)
if tt == gc.TFLOAT32 {
gins(x86.AFMOVFP, &r1, t)
} else {
gins(x86.AFMOVDP, &r1, t)
}
p2 := gc.Gbranch(obj.AJMP, nil, 0)
// simulated
gc.Patch(p1, gc.Pc)
gmove(&tlo, &ax)
gmove(&thi, &dx)
p1 = gins(x86.ASHRL, ncon(1), &ax)
p1.From.Index = x86.REG_DX // double-width shift DX -> AX
p1.From.Scale = 0
gins(x86.AMOVL, ncon(0), &cx)
gins(x86.ASETCC, nil, &cx)
gins(x86.AORL, &cx, &ax)
gins(x86.ASHRL, ncon(1), &dx)
gmove(&dx, &thi)
gmove(&ax, &tlo)
gc.Nodreg(&r1, gc.Types[tt], x86.REG_F0)
var r2 gc.Node
gc.Nodreg(&r2, gc.Types[tt], x86.REG_F0+1)
gins(x86.AFMOVV, &t1, &r1)
gins(x86.AFMOVD, &r1, &r1)
gins(x86.AFADDDP, &r1, &r2)
if tt == gc.TFLOAT32 {
gins(x86.AFMOVFP, &r1, t)
} else {
gins(x86.AFMOVDP, &r1, t)
}
gc.Patch(p2, gc.Pc)
splitclean()
return
}
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
// requires memory intermediate
hardmem:
gc.Tempname(&r1, cvt)
gmove(f, &r1)
gmove(&r1, t)
return
}
/*
* generate division.
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op gc.Op, nl *gc.Node, nr *gc.Node, res *gc.Node) {
// Have to be careful about handling
// most negative int divided by -1 correctly.
// The hardware will trap.
// Also the byte divide instruction needs AH,
// which we otherwise don't have to deal with.
// Easiest way to avoid for int8, int16: use int32.
// For int32 and int64, use explicit test.
// Could use int64 hw for int32.
t := nl.Type
t0 := t
check := false
if t.IsSigned() {
check = true
if gc.Isconst(nl, gc.CTINT) && nl.Int64() != -(1<<uint64(t.Width*8-1)) {
check = false
} else if gc.Isconst(nr, gc.CTINT) && nr.Int64() != -1 {
check = false
}
}
if t.Width < 4 {
if t.IsSigned() {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
check = false
}
a := optoas(op, t)
var n3 gc.Node
gc.Regalloc(&n3, t0, nil)
var ax gc.Node
var oldax gc.Node
if nl.Ullman >= nr.Ullman {
savex(x86.REG_AX, &ax, &oldax, res, t0)
gc.Cgen(nl, &ax)
gc.Regalloc(&ax, t0, &ax) // mark ax live during cgen
gc.Cgen(nr, &n3)
gc.Regfree(&ax)
} else {
gc.Cgen(nr, &n3)
savex(x86.REG_AX, &ax, &oldax, res, t0)
gc.Cgen(nl, &ax)
}
if t != t0 {
// Convert
ax1 := ax
n31 := n3
ax.Type = t
n3.Type = t
gmove(&ax1, &ax)
gmove(&n31, &n3)
}
var n4 gc.Node
if gc.Nacl {
// Native Client does not relay the divide-by-zero trap
// to the executing program, so we must insert a check
// for ourselves.
gc.Nodconst(&n4, t, 0)
gins(optoas(gc.OCMP, t), &n3, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if panicdiv == nil {
panicdiv = gc.Sysfunc("panicdivide")
}
gc.Ginscall(panicdiv, -1)
gc.Patch(p1, gc.Pc)
}
var p2 *obj.Prog
if check {
gc.Nodconst(&n4, t, -1)
gins(optoas(gc.OCMP, t), &n3, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if op == gc.ODIV {
// a / (-1) is -a.
gins(optoas(gc.OMINUS, t), nil, &ax)
gmove(&ax, res)
} else {
// a % (-1) is 0.
gc.Nodconst(&n4, t, 0)
gmove(&n4, res)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
}
var olddx gc.Node
var dx gc.Node
savex(x86.REG_DX, &dx, &olddx, res, t)
if !t.IsSigned() {
gc.Nodconst(&n4, t, 0)
gmove(&n4, &dx)
} else {
gins(optoas(gc.OEXTEND, t), nil, nil)
}
gins(a, &n3, nil)
gc.Regfree(&n3)
if op == gc.ODIV {
gmove(&ax, res)
} else {
gmove(&dx, res)
}
restx(&dx, &olddx)
if check {
gc.Patch(p2, gc.Pc)
}
restx(&ax, &oldax)
}
func bgen_float(n *gc.Node, wantTrue bool, likely int, to *obj.Prog) {
nl := n.Left
nr := n.Right
op := n.Op
if !wantTrue {
// brcom is not valid on floats when NaN is involved.
p1 := gc.Gbranch(obj.AJMP, nil, 0)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// No need to avoid re-genning ninit.
bgen_float(n, true, -likely, p2)
gc.Patch(gc.Gbranch(obj.AJMP, nil, 0), to)
gc.Patch(p2, gc.Pc)
return
}
if gc.Thearch.Use387 {
op = gc.Brrev(op) // because the args are stacked
if op == gc.OGE || op == gc.OGT {
// only < and <= work right with NaN; reverse if needed
nl, nr = nr, nl
op = gc.Brrev(op)
}
var ax, n2, tmp gc.Node
gc.Nodreg(&tmp, nr.Type, x86.REG_F0)
gc.Nodreg(&n2, nr.Type, x86.REG_F0+1)
gc.Nodreg(&ax, gc.Types[gc.TUINT16], x86.REG_AX)
if gc.Simsimtype(nr.Type) == gc.TFLOAT64 {
if nl.Ullman > nr.Ullman {
gc.Cgen(nl, &tmp)
gc.Cgen(nr, &tmp)
gins(x86.AFXCHD, &tmp, &n2)
} else {
gc.Cgen(nr, &tmp)
gc.Cgen(nl, &tmp)
}
gins(x86.AFUCOMPP, &tmp, &n2)
} else {
// TODO(rsc): The moves back and forth to memory
// here are for truncating the value to 32 bits.
// This handles 32-bit comparison but presumably
// all the other ops have the same problem.
// We need to figure out what the right general
// solution is, besides telling people to use float64.
var t1 gc.Node
gc.Tempname(&t1, gc.Types[gc.TFLOAT32])
var t2 gc.Node
gc.Tempname(&t2, gc.Types[gc.TFLOAT32])
gc.Cgen(nr, &t1)
gc.Cgen(nl, &t2)
gmove(&t2, &tmp)
gins(x86.AFCOMFP, &t1, &tmp)
}
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
} else {
// Not 387
if !nl.Addable {
nl = gc.CgenTemp(nl)
}
if !nr.Addable {
nr = gc.CgenTemp(nr)
}
var n2 gc.Node
gc.Regalloc(&n2, nr.Type, nil)
gmove(nr, &n2)
nr = &n2
if nl.Op != gc.OREGISTER {
var n3 gc.Node
gc.Regalloc(&n3, nl.Type, nil)
gmove(nl, &n3)
nl = &n3
}
if op == gc.OGE || op == gc.OGT {
// only < and <= work right with NopN; reverse if needed
nl, nr = nr, nl
op = gc.Brrev(op)
}
gins(foptoas(gc.OCMP, nr.Type, 0), nl, nr)
if nl.Op == gc.OREGISTER {
gc.Regfree(nl)
}
gc.Regfree(nr)
}
switch op {
case gc.OEQ:
// neither NE nor P
p1 := gc.Gbranch(x86.AJNE, nil, -likely)
p2 := gc.Gbranch(x86.AJPS, nil, -likely)
gc.Patch(gc.Gbranch(obj.AJMP, nil, 0), to)
gc.Patch(p1, gc.Pc)
gc.Patch(p2, gc.Pc)
case gc.ONE:
// either NE or P
gc.Patch(gc.Gbranch(x86.AJNE, nil, likely), to)
gc.Patch(gc.Gbranch(x86.AJPS, nil, likely), to)
default:
gc.Patch(gc.Gbranch(optoas(op, nr.Type), nil, likely), to)
}
}
/*
* 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 rr gc.Node
split64(nl, &lo1, &hi1)
split64(nr, &lo2, &hi2)
// compare most significant word;
// if they differ, we're done.
t := hi1.Type
if nl.Op == gc.OLITERAL || nr.Op == gc.OLITERAL {
gins(x86.ACMPL, &hi1, &hi2)
} else {
gc.Regalloc(&rr, gc.Types[gc.TINT32], nil)
gins(x86.AMOVL, &hi1, &rr)
gins(x86.ACMPL, &rr, &hi2)
gc.Regfree(&rr)
}
var br *obj.Prog
switch op {
default:
gc.Fatalf("cmp64 %v %v", op, t)
// cmp hi
// jne L
// cmp lo
// jeq to
// L:
case gc.OEQ:
br = gc.Gbranch(x86.AJNE, nil, -likely)
// cmp hi
// jne to
// cmp lo
// jne to
case gc.ONE:
gc.Patch(gc.Gbranch(x86.AJNE, nil, likely), to)
// cmp hi
// jgt to
// jlt L
// cmp lo
// jge to (or jgt 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
// jlt to
// jgt L
// cmp lo
// jle 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
if nl.Op == gc.OLITERAL || nr.Op == gc.OLITERAL {
gins(x86.ACMPL, &lo1, &lo2)
} else {
gc.Regalloc(&rr, gc.Types[gc.TINT32], nil)
gins(x86.AMOVL, &lo1, &rr)
gins(x86.ACMPL, &rr, &lo2)
gc.Regfree(&rr)
}
// 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()
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op gc.Op, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Type.Width > 4 {
gc.Fatalf("cgen_shift %v", nl.Type)
}
w := int(nl.Type.Width * 8)
a := optoas(op, nl.Type)
if nr.Op == gc.OLITERAL {
var n2 gc.Node
gc.Tempname(&n2, nl.Type)
gc.Cgen(nl, &n2)
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gmove(&n2, &n1)
sc := uint64(nr.Int64())
if sc >= uint64(nl.Type.Width*8) {
// large shift gets 2 shifts by width-1
gins(a, ncon(uint32(w)-1), &n1)
gins(a, ncon(uint32(w)-1), &n1)
} else {
gins(a, nr, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
var oldcx gc.Node
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TUINT32], x86.REG_CX)
if gc.GetReg(x86.REG_CX) > 1 && !gc.Samereg(&cx, res) {
gc.Tempname(&oldcx, gc.Types[gc.TUINT32])
gmove(&cx, &oldcx)
}
var n1 gc.Node
var nt gc.Node
if nr.Type.Width > 4 {
gc.Tempname(&nt, nr.Type)
n1 = nt
} else {
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, nr.Type, &n1) // to hold the shift type in CX
}
var n2 gc.Node
if gc.Samereg(&cx, res) {
gc.Regalloc(&n2, nl.Type, nil)
} else {
gc.Regalloc(&n2, nl.Type, res)
}
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &n2)
gc.Cgen(nr, &n1)
} else {
gc.Cgen(nr, &n1)
gc.Cgen(nl, &n2)
}
// test and fix up large shifts
if bounded {
if nr.Type.Width > 4 {
// delayed reg alloc
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], &n1) // to hold the shift type in CX
var lo gc.Node
var hi gc.Node
split64(&nt, &lo, &hi)
gmove(&lo, &n1)
splitclean()
}
} else {
var p1 *obj.Prog
if nr.Type.Width > 4 {
// delayed reg alloc
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], &n1) // to hold the shift type in CX
var lo gc.Node
var hi gc.Node
split64(&nt, &lo, &hi)
gmove(&lo, &n1)
gins(optoas(gc.OCMP, gc.Types[gc.TUINT32]), &hi, ncon(0))
p2 := gc.Gbranch(optoas(gc.ONE, gc.Types[gc.TUINT32]), nil, +1)
gins(optoas(gc.OCMP, gc.Types[gc.TUINT32]), &n1, ncon(uint32(w)))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
splitclean()
gc.Patch(p2, gc.Pc)
} else {
gins(optoas(gc.OCMP, nr.Type), &n1, ncon(uint32(w)))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
}
if op == gc.ORSH && nl.Type.IsSigned() {
gins(a, ncon(uint32(w)-1), &n2)
} else {
gmove(ncon(0), &n2)
}
gc.Patch(p1, gc.Pc)
}
gins(a, &n1, &n2)
if oldcx.Op != 0 {
gmove(&oldcx, &cx)
}
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
func floatmove_387(f *gc.Node, t *gc.Node) {
var r1 gc.Node
var a obj.As
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
cvt := t.Type
switch uint32(ft)<<16 | uint32(tt) {
default:
goto fatal
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT16,
gc.TFLOAT32<<16 | gc.TINT32,
gc.TFLOAT32<<16 | gc.TINT64,
gc.TFLOAT64<<16 | gc.TINT16,
gc.TFLOAT64<<16 | gc.TINT32,
gc.TFLOAT64<<16 | gc.TINT64:
if t.Op == gc.OREGISTER {
goto hardmem
}
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[ft], x86.REG_F0)
if f.Op != gc.OREGISTER {
if ft == gc.TFLOAT32 {
gins(x86.AFMOVF, f, &r1)
} else {
gins(x86.AFMOVD, f, &r1)
}
}
// set round to zero mode during conversion
var t1 gc.Node
memname(&t1, gc.Types[gc.TUINT16])
var t2 gc.Node
memname(&t2, gc.Types[gc.TUINT16])
gins(x86.AFSTCW, nil, &t1)
gins(x86.AMOVW, ncon(0xf7f), &t2)
gins(x86.AFLDCW, &t2, nil)
if tt == gc.TINT16 {
gins(x86.AFMOVWP, &r1, t)
} else if tt == gc.TINT32 {
gins(x86.AFMOVLP, &r1, t)
} else {
gins(x86.AFMOVVP, &r1, t)
}
gins(x86.AFLDCW, &t1, nil)
return
// convert via int32.
case gc.TFLOAT32<<16 | gc.TINT8,
gc.TFLOAT32<<16 | gc.TUINT16,
gc.TFLOAT32<<16 | gc.TUINT8,
gc.TFLOAT64<<16 | gc.TINT8,
gc.TFLOAT64<<16 | gc.TUINT16,
gc.TFLOAT64<<16 | gc.TUINT8:
var t1 gc.Node
gc.Tempname(&t1, gc.Types[gc.TINT32])
gmove(f, &t1)
switch tt {
default:
gc.Fatalf("gmove %v", t)
case gc.TINT8:
gins(x86.ACMPL, &t1, ncon(-0x80&(1<<32-1)))
p1 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TINT32]), nil, -1)
gins(x86.ACMPL, &t1, ncon(0x7f))
p2 := gc.Gbranch(optoas(gc.OGT, gc.Types[gc.TINT32]), nil, -1)
p3 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
gc.Patch(p2, gc.Pc)
gmove(ncon(-0x80&(1<<32-1)), &t1)
gc.Patch(p3, gc.Pc)
gmove(&t1, t)
case gc.TUINT8:
gins(x86.ATESTL, ncon(0xffffff00), &t1)
p1 := gc.Gbranch(x86.AJEQ, nil, +1)
gins(x86.AMOVL, ncon(0), &t1)
gc.Patch(p1, gc.Pc)
gmove(&t1, t)
case gc.TUINT16:
gins(x86.ATESTL, ncon(0xffff0000), &t1)
p1 := gc.Gbranch(x86.AJEQ, nil, +1)
gins(x86.AMOVL, ncon(0), &t1)
gc.Patch(p1, gc.Pc)
gmove(&t1, t)
}
return
// convert via int64.
case gc.TFLOAT32<<16 | gc.TUINT32,
gc.TFLOAT64<<16 | gc.TUINT32:
cvt = gc.Types[gc.TINT64]
goto hardmem
/*
* integer to float
*/
case gc.TINT16<<16 | gc.TFLOAT32,
gc.TINT16<<16 | gc.TFLOAT64,
gc.TINT32<<16 | gc.TFLOAT32,
gc.TINT32<<16 | gc.TFLOAT64,
gc.TINT64<<16 | gc.TFLOAT32,
gc.TINT64<<16 | gc.TFLOAT64:
if t.Op != gc.OREGISTER {
goto hard
}
if f.Op == gc.OREGISTER {
cvt = f.Type
goto hardmem
}
switch ft {
case gc.TINT16:
a = x86.AFMOVW
case gc.TINT32:
a = x86.AFMOVL
default:
a = x86.AFMOVV
}
// convert via int32 memory
case gc.TINT8<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TUINT16<<16 | gc.TFLOAT32,
gc.TUINT16<<16 | gc.TFLOAT64,
gc.TUINT8<<16 | gc.TFLOAT32,
gc.TUINT8<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT32]
goto hardmem
// convert via int64 memory
case gc.TUINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT64]
goto hardmem
// The way the code generator uses floating-point
// registers, a move from F0 to F0 is intended as a no-op.
// On the x86, it's not: it pushes a second copy of F0
// on the floating point stack. So toss it away here.
// Also, F0 is the *only* register we ever evaluate
// into, so we should only see register/register as F0/F0.
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32,
gc.TFLOAT64<<16 | gc.TFLOAT64:
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
if f.Op == gc.OREGISTER && t.Op == gc.OREGISTER {
if f.Reg != x86.REG_F0 || t.Reg != x86.REG_F0 {
goto fatal
}
return
}
a = x86.AFMOVF
if ft == gc.TFLOAT64 {
a = x86.AFMOVD
}
if gc.Ismem(t) {
if f.Op != gc.OREGISTER || f.Reg != x86.REG_F0 {
gc.Fatalf("gmove %v", f)
}
a = x86.AFMOVFP
if ft == gc.TFLOAT64 {
a = x86.AFMOVDP
}
}
case gc.TFLOAT32<<16 | gc.TFLOAT64:
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
if f.Op == gc.OREGISTER && t.Op == gc.OREGISTER {
if f.Reg != x86.REG_F0 || t.Reg != x86.REG_F0 {
goto fatal
}
return
}
if f.Op == gc.OREGISTER {
gins(x86.AFMOVDP, f, t)
} else {
gins(x86.AFMOVF, f, t)
}
return
case gc.TFLOAT64<<16 | gc.TFLOAT32:
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
if f.Op == gc.OREGISTER && t.Op == gc.OREGISTER {
var r1 gc.Node
gc.Tempname(&r1, gc.Types[gc.TFLOAT32])
gins(x86.AFMOVFP, f, &r1)
gins(x86.AFMOVF, &r1, t)
return
}
if f.Op == gc.OREGISTER {
gins(x86.AFMOVFP, f, t)
} else {
gins(x86.AFMOVD, f, t)
}
return
}
gins(a, f, t)
return
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
// requires memory intermediate
hardmem:
gc.Tempname(&r1, cvt)
gmove(f, &r1)
gmove(&r1, t)
return
// should not happen
fatal:
gc.Fatalf("gmove %v -> %v", gc.Nconv(f, gc.FmtLong), gc.Nconv(t, gc.FmtLong))
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 gc.Op, nl *gc.Node, nr *gc.Node, res *gc.Node, ax *gc.Node, dx *gc.Node) {
// Have to be careful about handling
// most negative int divided by -1 correctly.
// The hardware will trap.
// Also the byte divide instruction needs AH,
// which we otherwise don't have to deal with.
// Easiest way to avoid for int8, int16: use int32.
// For int32 and int64, use explicit test.
// Could use int64 hw for int32.
t := nl.Type
t0 := t
check := false
if t.IsSigned() {
check = true
if gc.Isconst(nl, gc.CTINT) && nl.Int64() != -1<<uint64(t.Width*8-1) {
check = false
} else if gc.Isconst(nr, gc.CTINT) && nr.Int64() != -1 {
check = false
}
}
if t.Width < 4 {
if t.IsSigned() {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
check = false
}
var t1 gc.Node
gc.Tempname(&t1, t)
var t2 gc.Node
gc.Tempname(&t2, t)
if t0 != t {
var t3 gc.Node
gc.Tempname(&t3, t0)
var t4 gc.Node
gc.Tempname(&t4, t0)
gc.Cgen(nl, &t3)
gc.Cgen(nr, &t4)
// Convert.
gmove(&t3, &t1)
gmove(&t4, &t2)
} else {
gc.Cgen(nl, &t1)
gc.Cgen(nr, &t2)
}
var n1 gc.Node
if !gc.Samereg(ax, res) && !gc.Samereg(dx, res) {
gc.Regalloc(&n1, t, res)
} else {
gc.Regalloc(&n1, t, nil)
}
gmove(&t2, &n1)
gmove(&t1, ax)
var p2 *obj.Prog
var n4 gc.Node
if gc.Nacl {
// Native Client does not relay the divide-by-zero trap
// to the executing program, so we must insert a check
// for ourselves.
gc.Nodconst(&n4, t, 0)
gins(optoas(gc.OCMP, t), &n1, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if panicdiv == nil {
panicdiv = gc.Sysfunc("panicdivide")
}
gc.Ginscall(panicdiv, -1)
gc.Patch(p1, gc.Pc)
}
if check {
gc.Nodconst(&n4, t, -1)
gins(optoas(gc.OCMP, t), &n1, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if op == gc.ODIV {
// a / (-1) is -a.
gins(optoas(gc.OMINUS, t), nil, ax)
gmove(ax, res)
} else {
// a % (-1) is 0.
gc.Nodconst(&n4, t, 0)
gmove(&n4, res)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
}
if !t.IsSigned() {
var nz gc.Node
gc.Nodconst(&nz, t, 0)
gmove(&nz, dx)
} else {
gins(optoas(gc.OEXTEND, t), nil, nil)
}
gins(optoas(op, t), &n1, nil)
gc.Regfree(&n1)
if op == gc.ODIV {
gmove(ax, res)
} else {
gmove(dx, res)
}
if check {
gc.Patch(p2, gc.Pc)
}
}
/*
* generate move:
* t = f
* hard part is conversions.
*/
func gmove(f *gc.Node, t *gc.Node) {
if gc.Debug['M'] != 0 {
fmt.Printf("gmove %v -> %v\n", gc.Nconv(f, gc.FmtLong), gc.Nconv(t, gc.FmtLong))
}
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
cvt := t.Type
if gc.Iscomplex[ft] || gc.Iscomplex[tt] {
gc.Complexmove(f, t)
return
}
// cannot have two memory operands
var a obj.As
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
f.Convconst(&con, t.Type)
f = &con
ft = tt // so big switch will choose a simple mov
// some constants can't move directly to memory.
if gc.Ismem(t) {
// float constants come from memory.
if gc.Isfloat[tt] {
goto hard
}
// 64-bit immediates are really 32-bit sign-extended
// unless moving into a register.
if gc.Isint[tt] {
if i := con.Int64(); int64(int32(i)) != i {
goto hard
}
}
}
}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
gc.Dump("f", f)
gc.Dump("t", t)
gc.Fatalf("gmove %v -> %v", gc.Tconv(f.Type, gc.FmtLong), gc.Tconv(t.Type, gc.FmtLong))
/*
* integer copy and truncate
*/
case gc.TINT8<<16 | gc.TINT8, // same size
gc.TINT8<<16 | gc.TUINT8,
gc.TUINT8<<16 | gc.TINT8,
gc.TUINT8<<16 | gc.TUINT8,
gc.TINT16<<16 | gc.TINT8,
// truncate
gc.TUINT16<<16 | gc.TINT8,
gc.TINT32<<16 | gc.TINT8,
gc.TUINT32<<16 | gc.TINT8,
gc.TINT64<<16 | gc.TINT8,
gc.TUINT64<<16 | gc.TINT8,
gc.TINT16<<16 | gc.TUINT8,
gc.TUINT16<<16 | gc.TUINT8,
gc.TINT32<<16 | gc.TUINT8,
gc.TUINT32<<16 | gc.TUINT8,
gc.TINT64<<16 | gc.TUINT8,
gc.TUINT64<<16 | gc.TUINT8:
a = x86.AMOVB
case gc.TINT16<<16 | gc.TINT16, // same size
gc.TINT16<<16 | gc.TUINT16,
gc.TUINT16<<16 | gc.TINT16,
gc.TUINT16<<16 | gc.TUINT16,
gc.TINT32<<16 | gc.TINT16,
// truncate
gc.TUINT32<<16 | gc.TINT16,
gc.TINT64<<16 | gc.TINT16,
gc.TUINT64<<16 | gc.TINT16,
gc.TINT32<<16 | gc.TUINT16,
gc.TUINT32<<16 | gc.TUINT16,
gc.TINT64<<16 | gc.TUINT16,
gc.TUINT64<<16 | gc.TUINT16:
a = x86.AMOVW
case gc.TINT32<<16 | gc.TINT32, // same size
gc.TINT32<<16 | gc.TUINT32,
gc.TUINT32<<16 | gc.TINT32,
gc.TUINT32<<16 | gc.TUINT32:
a = x86.AMOVL
case gc.TINT64<<16 | gc.TINT32, // truncate
gc.TUINT64<<16 | gc.TINT32,
gc.TINT64<<16 | gc.TUINT32,
gc.TUINT64<<16 | gc.TUINT32:
a = x86.AMOVQL
case gc.TINT64<<16 | gc.TINT64, // same size
gc.TINT64<<16 | gc.TUINT64,
gc.TUINT64<<16 | gc.TINT64,
gc.TUINT64<<16 | gc.TUINT64:
a = x86.AMOVQ
/*
* integer up-conversions
*/
case gc.TINT8<<16 | gc.TINT16, // sign extend int8
gc.TINT8<<16 | gc.TUINT16:
a = x86.AMOVBWSX
goto rdst
case gc.TINT8<<16 | gc.TINT32,
gc.TINT8<<16 | gc.TUINT32:
a = x86.AMOVBLSX
goto rdst
case gc.TINT8<<16 | gc.TINT64,
gc.TINT8<<16 | gc.TUINT64:
a = x86.AMOVBQSX
goto rdst
case gc.TUINT8<<16 | gc.TINT16, // zero extend uint8
gc.TUINT8<<16 | gc.TUINT16:
a = x86.AMOVBWZX
goto rdst
case gc.TUINT8<<16 | gc.TINT32,
gc.TUINT8<<16 | gc.TUINT32:
a = x86.AMOVBLZX
goto rdst
case gc.TUINT8<<16 | gc.TINT64,
gc.TUINT8<<16 | gc.TUINT64:
a = x86.AMOVBQZX
goto rdst
case gc.TINT16<<16 | gc.TINT32, // sign extend int16
gc.TINT16<<16 | gc.TUINT32:
a = x86.AMOVWLSX
goto rdst
case gc.TINT16<<16 | gc.TINT64,
gc.TINT16<<16 | gc.TUINT64:
a = x86.AMOVWQSX
goto rdst
case gc.TUINT16<<16 | gc.TINT32, // zero extend uint16
gc.TUINT16<<16 | gc.TUINT32:
a = x86.AMOVWLZX
goto rdst
case gc.TUINT16<<16 | gc.TINT64,
gc.TUINT16<<16 | gc.TUINT64:
a = x86.AMOVWQZX
goto rdst
case gc.TINT32<<16 | gc.TINT64, // sign extend int32
gc.TINT32<<16 | gc.TUINT64:
a = x86.AMOVLQSX
goto rdst
// AMOVL into a register zeros the top of the register,
// so this is not always necessary, but if we rely on AMOVL
// the optimizer is almost certain to screw with us.
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
a = x86.AMOVLQZX
goto rdst
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT32:
a = x86.ACVTTSS2SL
goto rdst
case gc.TFLOAT64<<16 | gc.TINT32:
a = x86.ACVTTSD2SL
goto rdst
case gc.TFLOAT32<<16 | gc.TINT64:
a = x86.ACVTTSS2SQ
goto rdst
case gc.TFLOAT64<<16 | gc.TINT64:
a = x86.ACVTTSD2SQ
goto rdst
// convert via int32.
case gc.TFLOAT32<<16 | gc.TINT16,
gc.TFLOAT32<<16 | gc.TINT8,
gc.TFLOAT32<<16 | gc.TUINT16,
gc.TFLOAT32<<16 | gc.TUINT8,
gc.TFLOAT64<<16 | gc.TINT16,
gc.TFLOAT64<<16 | gc.TINT8,
gc.TFLOAT64<<16 | gc.TUINT16,
gc.TFLOAT64<<16 | gc.TUINT8:
cvt = gc.Types[gc.TINT32]
goto hard
// convert via int64.
case gc.TFLOAT32<<16 | gc.TUINT32,
gc.TFLOAT64<<16 | gc.TUINT32:
cvt = gc.Types[gc.TINT64]
goto hard
// algorithm is:
// if small enough, use native float64 -> int64 conversion.
// otherwise, subtract 2^63, convert, and add it back.
case gc.TFLOAT32<<16 | gc.TUINT64,
gc.TFLOAT64<<16 | gc.TUINT64:
a := x86.ACVTTSS2SQ
if ft == gc.TFLOAT64 {
a = x86.ACVTTSD2SQ
}
bignodes()
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[ft], nil)
var r2 gc.Node
gc.Regalloc(&r2, gc.Types[tt], t)
var r3 gc.Node
gc.Regalloc(&r3, gc.Types[ft], nil)
var r4 gc.Node
gc.Regalloc(&r4, gc.Types[tt], nil)
gins(optoas(gc.OAS, f.Type), f, &r1)
gins(optoas(gc.OCMP, f.Type), &bigf, &r1)
p1 := gc.Gbranch(optoas(gc.OLE, f.Type), nil, +1)
gins(a, &r1, &r2)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
gins(optoas(gc.OAS, f.Type), &bigf, &r3)
gins(optoas(gc.OSUB, f.Type), &r3, &r1)
gins(a, &r1, &r2)
gins(x86.AMOVQ, &bigi, &r4)
gins(x86.AXORQ, &r4, &r2)
gc.Patch(p2, gc.Pc)
gmove(&r2, t)
gc.Regfree(&r4)
gc.Regfree(&r3)
gc.Regfree(&r2)
gc.Regfree(&r1)
return
/*
* integer to float
*/
case gc.TINT32<<16 | gc.TFLOAT32:
a = x86.ACVTSL2SS
goto rdst
case gc.TINT32<<16 | gc.TFLOAT64:
a = x86.ACVTSL2SD
goto rdst
case gc.TINT64<<16 | gc.TFLOAT32:
a = x86.ACVTSQ2SS
goto rdst
case gc.TINT64<<16 | gc.TFLOAT64:
a = x86.ACVTSQ2SD
goto rdst
// convert via int32
case gc.TINT16<<16 | gc.TFLOAT32,
gc.TINT16<<16 | gc.TFLOAT64,
gc.TINT8<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TUINT16<<16 | gc.TFLOAT32,
gc.TUINT16<<16 | gc.TFLOAT64,
gc.TUINT8<<16 | gc.TFLOAT32,
gc.TUINT8<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT32]
goto hard
// convert via int64.
case gc.TUINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT64]
goto hard
// algorithm is:
// if small enough, use native int64 -> uint64 conversion.
// otherwise, halve (rounding to odd?), convert, and double.
case gc.TUINT64<<16 | gc.TFLOAT32,
gc.TUINT64<<16 | gc.TFLOAT64:
a := x86.ACVTSQ2SS
if tt == gc.TFLOAT64 {
a = x86.ACVTSQ2SD
}
var zero gc.Node
gc.Nodconst(&zero, gc.Types[gc.TUINT64], 0)
var one gc.Node
gc.Nodconst(&one, gc.Types[gc.TUINT64], 1)
var r1 gc.Node
gc.Regalloc(&r1, f.Type, f)
var r2 gc.Node
gc.Regalloc(&r2, t.Type, t)
var r3 gc.Node
gc.Regalloc(&r3, f.Type, nil)
var r4 gc.Node
gc.Regalloc(&r4, f.Type, nil)
gmove(f, &r1)
gins(x86.ACMPQ, &r1, &zero)
p1 := gc.Gbranch(x86.AJLT, nil, +1)
gins(a, &r1, &r2)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
gmove(&r1, &r3)
gins(x86.ASHRQ, &one, &r3)
gmove(&r1, &r4)
gins(x86.AANDL, &one, &r4)
gins(x86.AORQ, &r4, &r3)
gins(a, &r3, &r2)
gins(optoas(gc.OADD, t.Type), &r2, &r2)
gc.Patch(p2, gc.Pc)
gmove(&r2, t)
gc.Regfree(&r4)
gc.Regfree(&r3)
gc.Regfree(&r2)
gc.Regfree(&r1)
return
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32:
a = x86.AMOVSS
case gc.TFLOAT64<<16 | gc.TFLOAT64:
a = x86.AMOVSD
case gc.TFLOAT32<<16 | gc.TFLOAT64:
a = x86.ACVTSS2SD
goto rdst
case gc.TFLOAT64<<16 | gc.TFLOAT32:
a = x86.ACVTSD2SS
goto rdst
}
gins(a, f, t)
return
// requires register destination
rdst:
{
var r1 gc.Node
gc.Regalloc(&r1, t.Type, t)
gins(a, f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
// requires register intermediate
hard:
var r1 gc.Node
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
func blockcopy(n, res *gc.Node, osrc, odst, w int64) {
// determine alignment.
// want to avoid unaligned access, so have to use
// smaller operations for less aligned types.
// for example moving [4]byte must use 4 MOVB not 1 MOVW.
align := int(n.Type.Align)
var op obj.As
switch align {
default:
gc.Fatalf("sgen: invalid alignment %d for %v", align, n.Type)
case 1:
op = arm.AMOVB
case 2:
op = arm.AMOVH
case 4:
op = arm.AMOVW
}
if w%int64(align) != 0 {
gc.Fatalf("sgen: unaligned size %d (align=%d) for %v", w, align, n.Type)
}
c := int32(w / int64(align))
if osrc%int64(align) != 0 || odst%int64(align) != 0 {
gc.Fatalf("sgen: unaligned offset src %d or dst %d (align %d)", osrc, odst, align)
}
// if we are copying forward on the stack and
// the src and dst overlap, then reverse direction
dir := align
if osrc < odst && odst < osrc+w {
dir = -dir
}
if op == arm.AMOVW && !gc.Nacl && dir > 0 && c >= 4 && c <= 128 {
var r0 gc.Node
r0.Op = gc.OREGISTER
r0.Reg = arm.REG_R0
var r1 gc.Node
r1.Op = gc.OREGISTER
r1.Reg = arm.REG_R0 + 1
var r2 gc.Node
r2.Op = gc.OREGISTER
r2.Reg = arm.REG_R0 + 2
var src gc.Node
gc.Regalloc(&src, gc.Types[gc.Tptr], &r1)
var dst gc.Node
gc.Regalloc(&dst, gc.Types[gc.Tptr], &r2)
if n.Ullman >= res.Ullman {
// eval n first
gc.Agen(n, &src)
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agen(res, &dst)
} else {
// eval res first
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agen(res, &dst)
gc.Agen(n, &src)
}
var tmp gc.Node
gc.Regalloc(&tmp, gc.Types[gc.Tptr], &r0)
f := gc.Sysfunc("duffcopy")
p := gins(obj.ADUFFCOPY, nil, f)
gc.Afunclit(&p.To, f)
// 8 and 128 = magic constants: see ../../runtime/asm_arm.s
p.To.Offset = 8 * (128 - int64(c))
gc.Regfree(&tmp)
gc.Regfree(&src)
gc.Regfree(&dst)
return
}
var dst gc.Node
var src gc.Node
if n.Ullman >= res.Ullman {
gc.Agenr(n, &dst, res) // temporarily use dst
gc.Regalloc(&src, gc.Types[gc.Tptr], nil)
gins(arm.AMOVW, &dst, &src)
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agen(res, &dst)
} else {
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agenr(res, &dst, res)
gc.Agenr(n, &src, nil)
}
var tmp gc.Node
gc.Regalloc(&tmp, gc.Types[gc.TUINT32], nil)
// set up end marker
var nend gc.Node
if c >= 4 {
gc.Regalloc(&nend, gc.Types[gc.TUINT32], nil)
p := gins(arm.AMOVW, &src, &nend)
p.From.Type = obj.TYPE_ADDR
if dir < 0 {
p.From.Offset = int64(dir)
} else {
p.From.Offset = w
}
}
// move src and dest to the end of block if necessary
if dir < 0 {
p := gins(arm.AMOVW, &src, &src)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w + int64(dir)
p = gins(arm.AMOVW, &dst, &dst)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w + int64(dir)
}
// move
if c >= 4 {
p := gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
p.Scond |= arm.C_PBIT
ploop := p
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
p.Scond |= arm.C_PBIT
p = gins(arm.ACMP, &src, nil)
raddr(&nend, p)
gc.Patch(gc.Gbranch(arm.ABNE, nil, 0), ploop)
gc.Regfree(&nend)
} else {
var p *obj.Prog
for ; c > 0; c-- {
p = gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
p.Scond |= arm.C_PBIT
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
p.Scond |= arm.C_PBIT
}
}
gc.Regfree(&dst)
gc.Regfree(&src)
gc.Regfree(&tmp)
}
/*
* attempt to generate 64-bit
* res = n
* return 1 on success, 0 if op not handled.
*/
func cgen64(n *gc.Node, res *gc.Node) {
if res.Op != gc.OINDREG && res.Op != gc.ONAME {
gc.Dump("n", n)
gc.Dump("res", res)
gc.Fatalf("cgen64 %v of %v", n.Op, res.Op)
}
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", n.Op)
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.Int64())
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.Int64())
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.Int64())
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)
}
/*
* generate division.
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op gc.Op, nl *gc.Node, nr *gc.Node, res *gc.Node) {
// Have to be careful about handling
// most negative int divided by -1 correctly.
// The hardware will generate undefined result.
// Also need to explicitly trap on division on zero,
// the hardware will silently generate undefined result.
// DIVW will leave unpredicable result in higher 32-bit,
// so always use DIVD/DIVDU.
t := nl.Type
t0 := t
check := 0
if t.IsSigned() {
check = 1
if gc.Isconst(nl, gc.CTINT) && nl.Int64() != -(1<<uint64(t.Width*8-1)) {
check = 0
} else if gc.Isconst(nr, gc.CTINT) && nr.Int64() != -1 {
check = 0
}
}
if t.Width < 8 {
if t.IsSigned() {
t = gc.Types[gc.TINT64]
} else {
t = gc.Types[gc.TUINT64]
}
check = 0
}
a := optoas(gc.ODIV, t)
var tl gc.Node
gc.Regalloc(&tl, t0, nil)
var tr gc.Node
gc.Regalloc(&tr, t0, nil)
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &tl)
gc.Cgen(nr, &tr)
} else {
gc.Cgen(nr, &tr)
gc.Cgen(nl, &tl)
}
if t != t0 {
// Convert
tl2 := tl
tr2 := tr
tl.Type = t
tr.Type = t
gmove(&tl2, &tl)
gmove(&tr2, &tr)
}
// Handle divide-by-zero panic.
p1 := gins(optoas(gc.OCMP, t), &tr, nil)
p1.To.Type = obj.TYPE_REG
p1.To.Reg = s390x.REGZERO
p1 = gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if panicdiv == nil {
panicdiv = gc.Sysfunc("panicdivide")
}
gc.Ginscall(panicdiv, -1)
gc.Patch(p1, gc.Pc)
var p2 *obj.Prog
if check != 0 {
var nm1 gc.Node
gc.Nodconst(&nm1, t, -1)
gins(optoas(gc.OCMP, t), &tr, &nm1)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if op == gc.ODIV {
// a / (-1) is -a.
gins(optoas(gc.OMINUS, t), nil, &tl)
gmove(&tl, res)
} else {
// a % (-1) is 0.
var nz gc.Node
gc.Nodconst(&nz, t, 0)
gmove(&nz, res)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
}
p1 = gins(a, &tr, &tl)
if op == gc.ODIV {
gc.Regfree(&tr)
gmove(&tl, res)
} else {
// A%B = A-(A/B*B)
var tm gc.Node
gc.Regalloc(&tm, t, nil)
// patch div to use the 3 register form
// TODO(minux): add gins3?
p1.Reg = p1.To.Reg
p1.To.Reg = tm.Reg
gins(optoas(gc.OMUL, t), &tr, &tm)
gc.Regfree(&tr)
gins(optoas(gc.OSUB, t), &tm, &tl)
gc.Regfree(&tm)
gmove(&tl, res)
}
gc.Regfree(&tl)
if check != 0 {
gc.Patch(p2, gc.Pc)
}
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
fmt.Printf("clearfat %v (%v, size: %d)\n", nl, nl.Type, nl.Type.Width)
}
w := uint64(nl.Type.Width)
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := w % 8 // bytes
q := w / 8 // dwords
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 := 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.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)
}
}
// clearfat clears (i.e. replaces with zeros) the value pointed to by nl.
func clearfat(nl *gc.Node) {
if gc.Debug['g'] != 0 {
fmt.Printf("clearfat %v (%v, size: %d)\n", nl, nl.Type, nl.Type.Width)
}
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
var dst gc.Node
gc.Regalloc(&dst, gc.Types[gc.Tptr], nil)
gc.Agen(nl, &dst)
var boff int64
w := nl.Type.Width
if w > clearLoopCutoff {
// Generate a loop clearing 256 bytes per iteration using XCs.
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p := gins(s390x.AMOVD, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w - (w % 256)
p = gins(s390x.AXC, &dst, &dst)
p.From.Type = obj.TYPE_MEM
p.From.Offset = 0
p.To.Type = obj.TYPE_MEM
p.To.Offset = 0
p.From3 = new(obj.Addr)
p.From3.Offset = 256
p.From3.Type = obj.TYPE_CONST
pl := p
ginscon(s390x.AADD, 256, &dst)
gins(s390x.ACMP, &dst, &end)
gc.Patch(gc.Gbranch(s390x.ABNE, nil, 0), pl)
gc.Regfree(&end)
w = w % 256
}
// Generate instructions to clear the remaining memory.
for w > 0 {
n := w
// Can clear at most 256 bytes per instruction.
if n > 256 {
n = 256
}
switch n {
// Handle very small clears using moves.
case 8, 4, 2, 1:
ins := s390x.AMOVB
switch n {
case 8:
ins = s390x.AMOVD
case 4:
ins = s390x.AMOVW
case 2:
ins = s390x.AMOVH
}
p := gins(ins, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 0
p.To.Type = obj.TYPE_MEM
p.To.Offset = boff
// Handle clears that would require multiple moves with a XC.
default:
p := gins(s390x.AXC, &dst, &dst)
p.From.Type = obj.TYPE_MEM
p.From.Offset = boff
p.To.Type = obj.TYPE_MEM
p.To.Offset = boff
p.From3 = new(obj.Addr)
p.From3.Offset = n
p.From3.Type = obj.TYPE_CONST
}
boff += n
w -= n
}
gc.Regfree(&dst)
}
/*
* attempt to generate 64-bit
* res = n
* return 1 on success, 0 if op not handled.
*/
func cgen64(n *gc.Node, res *gc.Node) {
if res.Op != gc.OINDREG && res.Op != gc.ONAME {
gc.Dump("n", n)
gc.Dump("res", res)
gc.Fatalf("cgen64 %v of %v", n.Op, res.Op)
}
switch n.Op {
default:
gc.Fatalf("cgen64 %v", n.Op)
case gc.OMINUS:
gc.Cgen(n.Left, res)
var hi1 gc.Node
var lo1 gc.Node
split64(res, &lo1, &hi1)
gins(x86.ANEGL, nil, &lo1)
gins(x86.AADCL, ncon(0), &hi1)
gins(x86.ANEGL, nil, &hi1)
splitclean()
return
case gc.OCOM:
gc.Cgen(n.Left, res)
var lo1 gc.Node
var hi1 gc.Node
split64(res, &lo1, &hi1)
gins(x86.ANOTL, nil, &lo1)
gins(x86.ANOTL, nil, &hi1)
splitclean()
return
// binary operators.
// common setup below.
case gc.OADD,
gc.OSUB,
gc.OMUL,
gc.OLROT,
gc.OLSH,
gc.ORSH,
gc.OAND,
gc.OOR,
gc.OXOR:
break
}
l := n.Left
r := n.Right
if !l.Addable {
var t1 gc.Node
gc.Tempname(&t1, l.Type)
gc.Cgen(l, &t1)
l = &t1
}
if r != nil && !r.Addable {
var t2 gc.Node
gc.Tempname(&t2, r.Type)
gc.Cgen(r, &t2)
r = &t2
}
var ax gc.Node
gc.Nodreg(&ax, gc.Types[gc.TINT32], x86.REG_AX)
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TINT32], x86.REG_CX)
var dx gc.Node
gc.Nodreg(&dx, gc.Types[gc.TINT32], x86.REG_DX)
// Setup for binary operation.
var hi1 gc.Node
var lo1 gc.Node
split64(l, &lo1, &hi1)
var lo2 gc.Node
var hi2 gc.Node
if gc.Is64(r.Type) {
split64(r, &lo2, &hi2)
}
// Do op. Leave result in DX:AX.
switch n.Op {
// TODO: Constants
case gc.OADD:
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(x86.AADDL, &lo2, &ax)
gins(x86.AADCL, &hi2, &dx)
// TODO: Constants.
case gc.OSUB:
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(x86.ASUBL, &lo2, &ax)
gins(x86.ASBBL, &hi2, &dx)
case gc.OMUL:
// let's call the next three EX, FX and GX
var ex, fx, gx gc.Node
gc.Regalloc(&ex, gc.Types[gc.TPTR32], nil)
gc.Regalloc(&fx, gc.Types[gc.TPTR32], nil)
gc.Regalloc(&gx, gc.Types[gc.TPTR32], nil)
// load args into DX:AX and EX:GX.
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(x86.AMOVL, &lo2, &gx)
gins(x86.AMOVL, &hi2, &ex)
// if DX and EX are zero, use 32 x 32 -> 64 unsigned multiply.
gins(x86.AMOVL, &dx, &fx)
gins(x86.AORL, &ex, &fx)
p1 := gc.Gbranch(x86.AJNE, nil, 0)
gins(x86.AMULL, &gx, nil) // implicit &ax
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// full 64x64 -> 64, from 32x32 -> 64.
gins(x86.AIMULL, &gx, &dx)
gins(x86.AMOVL, &ax, &fx)
gins(x86.AIMULL, &ex, &fx)
gins(x86.AADDL, &dx, &fx)
gins(x86.AMOVL, &gx, &dx)
gins(x86.AMULL, &dx, nil) // implicit &ax
gins(x86.AADDL, &fx, &dx)
gc.Patch(p2, gc.Pc)
gc.Regfree(&ex)
gc.Regfree(&fx)
gc.Regfree(&gx)
// We only rotate by a constant c in [0,64).
// if c >= 32:
// lo, hi = hi, lo
// c -= 32
// if c == 0:
// no-op
// else:
// t = hi
// shld hi:lo, c
// shld lo:t, c
case gc.OLROT:
v := uint64(r.Int64())
if v >= 32 {
// reverse during load to do the first 32 bits of rotate
v -= 32
gins(x86.AMOVL, &lo1, &dx)
gins(x86.AMOVL, &hi1, &ax)
} else {
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
}
if v == 0 {
} else // done
{
gins(x86.AMOVL, &dx, &cx)
p1 := gins(x86.ASHLL, ncon(uint32(v)), &dx)
p1.From.Index = x86.REG_AX // double-width shift
p1.From.Scale = 0
p1 = gins(x86.ASHLL, ncon(uint32(v)), &ax)
p1.From.Index = x86.REG_CX // double-width shift
p1.From.Scale = 0
}
case gc.OLSH:
if r.Op == gc.OLITERAL {
v := uint64(r.Int64())
if v >= 64 {
if gc.Is64(r.Type) {
splitclean()
}
splitclean()
split64(res, &lo2, &hi2)
gins(x86.AMOVL, ncon(0), &lo2)
gins(x86.AMOVL, ncon(0), &hi2)
splitclean()
return
}
if v >= 32 {
if gc.Is64(r.Type) {
splitclean()
}
split64(res, &lo2, &hi2)
gmove(&lo1, &hi2)
if v > 32 {
gins(x86.ASHLL, ncon(uint32(v-32)), &hi2)
}
gins(x86.AMOVL, ncon(0), &lo2)
splitclean()
splitclean()
return
}
// general shift
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
p1 := gins(x86.ASHLL, ncon(uint32(v)), &dx)
p1.From.Index = x86.REG_AX // double-width shift
p1.From.Scale = 0
gins(x86.ASHLL, ncon(uint32(v)), &ax)
break
}
// load value into DX:AX.
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
// load shift value into register.
// if high bits are set, zero value.
var p1 *obj.Prog
if gc.Is64(r.Type) {
gins(x86.ACMPL, &hi2, ncon(0))
p1 = gc.Gbranch(x86.AJNE, nil, +1)
gins(x86.AMOVL, &lo2, &cx)
} else {
cx.Type = gc.Types[gc.TUINT32]
gmove(r, &cx)
}
// if shift count is >=64, zero value
gins(x86.ACMPL, &cx, ncon(64))
p2 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
if p1 != nil {
gc.Patch(p1, gc.Pc)
}
gins(x86.AXORL, &dx, &dx)
gins(x86.AXORL, &ax, &ax)
gc.Patch(p2, gc.Pc)
// if shift count is >= 32, zero low.
gins(x86.ACMPL, &cx, ncon(32))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
gins(x86.AMOVL, &ax, &dx)
gins(x86.ASHLL, &cx, &dx) // SHLL only uses bottom 5 bits of count
gins(x86.AXORL, &ax, &ax)
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// general shift
p1 = gins(x86.ASHLL, &cx, &dx)
p1.From.Index = x86.REG_AX // double-width shift
p1.From.Scale = 0
gins(x86.ASHLL, &cx, &ax)
gc.Patch(p2, gc.Pc)
case gc.ORSH:
if r.Op == gc.OLITERAL {
v := uint64(r.Int64())
if v >= 64 {
if gc.Is64(r.Type) {
splitclean()
}
splitclean()
split64(res, &lo2, &hi2)
if hi1.Type.Etype == gc.TINT32 {
gmove(&hi1, &lo2)
gins(x86.ASARL, ncon(31), &lo2)
gmove(&hi1, &hi2)
gins(x86.ASARL, ncon(31), &hi2)
} else {
gins(x86.AMOVL, ncon(0), &lo2)
gins(x86.AMOVL, ncon(0), &hi2)
}
splitclean()
return
}
if v >= 32 {
if gc.Is64(r.Type) {
splitclean()
}
split64(res, &lo2, &hi2)
gmove(&hi1, &lo2)
if v > 32 {
gins(optoas(gc.ORSH, hi1.Type), ncon(uint32(v-32)), &lo2)
}
if hi1.Type.Etype == gc.TINT32 {
gmove(&hi1, &hi2)
gins(x86.ASARL, ncon(31), &hi2)
} else {
gins(x86.AMOVL, ncon(0), &hi2)
}
splitclean()
splitclean()
return
}
// general shift
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
p1 := gins(x86.ASHRL, ncon(uint32(v)), &ax)
p1.From.Index = x86.REG_DX // double-width shift
p1.From.Scale = 0
gins(optoas(gc.ORSH, hi1.Type), ncon(uint32(v)), &dx)
break
}
// load value into DX:AX.
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
// load shift value into register.
// if high bits are set, zero value.
var p1 *obj.Prog
if gc.Is64(r.Type) {
gins(x86.ACMPL, &hi2, ncon(0))
p1 = gc.Gbranch(x86.AJNE, nil, +1)
gins(x86.AMOVL, &lo2, &cx)
} else {
cx.Type = gc.Types[gc.TUINT32]
gmove(r, &cx)
}
// if shift count is >=64, zero or sign-extend value
gins(x86.ACMPL, &cx, ncon(64))
p2 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
if p1 != nil {
gc.Patch(p1, gc.Pc)
}
if hi1.Type.Etype == gc.TINT32 {
gins(x86.ASARL, ncon(31), &dx)
gins(x86.AMOVL, &dx, &ax)
} else {
gins(x86.AXORL, &dx, &dx)
gins(x86.AXORL, &ax, &ax)
}
gc.Patch(p2, gc.Pc)
// if shift count is >= 32, sign-extend hi.
gins(x86.ACMPL, &cx, ncon(32))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
gins(x86.AMOVL, &dx, &ax)
if hi1.Type.Etype == gc.TINT32 {
gins(x86.ASARL, &cx, &ax) // SARL only uses bottom 5 bits of count
gins(x86.ASARL, ncon(31), &dx)
} else {
gins(x86.ASHRL, &cx, &ax)
gins(x86.AXORL, &dx, &dx)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// general shift
p1 = gins(x86.ASHRL, &cx, &ax)
p1.From.Index = x86.REG_DX // double-width shift
p1.From.Scale = 0
gins(optoas(gc.ORSH, hi1.Type), &cx, &dx)
gc.Patch(p2, gc.Pc)
// make constant the right side (it usually is anyway).
case gc.OXOR,
gc.OAND,
gc.OOR:
if lo1.Op == gc.OLITERAL {
nswap(&lo1, &lo2)
nswap(&hi1, &hi2)
}
if lo2.Op == gc.OLITERAL {
// special cases for constants.
lv := uint32(lo2.Int64())
hv := uint32(hi2.Int64())
splitclean() // right side
split64(res, &lo2, &hi2)
switch n.Op {
case gc.OXOR:
gmove(&lo1, &lo2)
gmove(&hi1, &hi2)
switch lv {
case 0:
break
case 0xffffffff:
gins(x86.ANOTL, nil, &lo2)
default:
gins(x86.AXORL, ncon(lv), &lo2)
}
switch hv {
case 0:
break
case 0xffffffff:
gins(x86.ANOTL, nil, &hi2)
default:
gins(x86.AXORL, ncon(hv), &hi2)
}
case gc.OAND:
switch lv {
case 0:
gins(x86.AMOVL, ncon(0), &lo2)
default:
gmove(&lo1, &lo2)
if lv != 0xffffffff {
gins(x86.AANDL, ncon(lv), &lo2)
}
}
switch hv {
case 0:
gins(x86.AMOVL, ncon(0), &hi2)
default:
gmove(&hi1, &hi2)
if hv != 0xffffffff {
gins(x86.AANDL, ncon(hv), &hi2)
}
}
case gc.OOR:
switch lv {
case 0:
gmove(&lo1, &lo2)
case 0xffffffff:
gins(x86.AMOVL, ncon(0xffffffff), &lo2)
default:
gmove(&lo1, &lo2)
gins(x86.AORL, ncon(lv), &lo2)
}
switch hv {
case 0:
gmove(&hi1, &hi2)
case 0xffffffff:
gins(x86.AMOVL, ncon(0xffffffff), &hi2)
default:
gmove(&hi1, &hi2)
gins(x86.AORL, ncon(hv), &hi2)
}
}
splitclean()
splitclean()
return
}
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(optoas(n.Op, lo1.Type), &lo2, &ax)
gins(optoas(n.Op, lo1.Type), &hi2, &dx)
}
if gc.Is64(r.Type) {
splitclean()
}
splitclean()
split64(res, &lo1, &hi1)
gins(x86.AMOVL, &ax, &lo1)
gins(x86.AMOVL, &dx, &hi1)
splitclean()
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
fmt.Printf("clearfat %v (%v, size: %d)\n", nl, nl.Type, nl.Type.Width)
}
w := uint64(nl.Type.Width)
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := w % 8 // bytes
q := w / 8 // dwords
if gc.Reginuse(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 := 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.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)
}
// blockcopy copies w bytes from &n to &res
func blockcopy(n, res *gc.Node, osrc, odst, w int64) {
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(s390x.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)
}
defer gc.Regfree(&src)
defer gc.Regfree(&dst)
var tmp gc.Node
gc.Regalloc(&tmp, gc.Types[gc.Tptr], nil)
defer gc.Regfree(&tmp)
offset := int64(0)
dir := _FORWARDS
if osrc < odst && odst < osrc+w {
// Reverse. Can't use MVC, fall back onto basic moves.
dir = _BACKWARDS
const copiesPerIter = 2
if w >= 8*copiesPerIter {
cnt := w - (w % (8 * copiesPerIter))
ginscon(s390x.AADD, w, &src)
ginscon(s390x.AADD, w, &dst)
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p := gins(s390x.ASUB, nil, &end)
p.From.Type = obj.TYPE_CONST
p.From.Offset = cnt
p.Reg = src.Reg
var label *obj.Prog
for i := 0; i < copiesPerIter; i++ {
offset := int64(-8 * (i + 1))
p := gins(s390x.AMOVD, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = offset
if i == 0 {
label = p
}
p = gins(s390x.AMOVD, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = offset
}
ginscon(s390x.ASUB, 8*copiesPerIter, &src)
ginscon(s390x.ASUB, 8*copiesPerIter, &dst)
gins(s390x.ACMP, &src, &end)
gc.Patch(gc.Gbranch(s390x.ABNE, nil, 0), label)
gc.Regfree(&end)
w -= cnt
} else {
offset = w
}
}
if dir == _FORWARDS && w > 1024 {
// Loop over MVCs
cnt := w - (w % 256)
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
add := gins(s390x.AADD, nil, &end)
add.From.Type = obj.TYPE_CONST
add.From.Offset = cnt
add.Reg = src.Reg
mvc := gins(s390x.AMVC, &src, &dst)
mvc.From.Type = obj.TYPE_MEM
mvc.From.Offset = 0
mvc.To.Type = obj.TYPE_MEM
mvc.To.Offset = 0
mvc.From3 = new(obj.Addr)
mvc.From3.Type = obj.TYPE_CONST
mvc.From3.Offset = 256
ginscon(s390x.AADD, 256, &src)
ginscon(s390x.AADD, 256, &dst)
gins(s390x.ACMP, &src, &end)
gc.Patch(gc.Gbranch(s390x.ABNE, nil, 0), mvc)
gc.Regfree(&end)
w -= cnt
}
for w > 0 {
cnt := w
// If in reverse we can only do 8, 4, 2 or 1 bytes at a time.
if dir == _BACKWARDS {
switch {
case cnt >= 8:
cnt = 8
case cnt >= 4:
cnt = 4
case cnt >= 2:
cnt = 2
}
} else if cnt > 256 {
cnt = 256
}
switch cnt {
case 8, 4, 2, 1:
op := s390x.AMOVB
switch cnt {
case 8:
op = s390x.AMOVD
case 4:
op = s390x.AMOVW
case 2:
op = s390x.AMOVH
}
load := gins(op, &src, &tmp)
load.From.Type = obj.TYPE_MEM
load.From.Offset = offset
store := gins(op, &tmp, &dst)
store.To.Type = obj.TYPE_MEM
store.To.Offset = offset
if dir == _BACKWARDS {
load.From.Offset -= cnt
store.To.Offset -= cnt
}
default:
p := gins(s390x.AMVC, &src, &dst)
p.From.Type = obj.TYPE_MEM
p.From.Offset = offset
p.To.Type = obj.TYPE_MEM
p.To.Offset = offset
p.From3 = new(obj.Addr)
p.From3.Type = obj.TYPE_CONST
p.From3.Offset = cnt
}
switch dir {
case _FORWARDS:
offset += cnt
case _BACKWARDS:
offset -= cnt
}
w -= cnt
}
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
fmt.Printf("clearfat %v (%v, size: %d)\n", nl, nl.Type, nl.Type.Width)
}
w := uint64(nl.Type.Width)
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := w % 8 // bytes
q := w / 8 // dwords
if gc.Reginuse(mips.REGRT1) {
gc.Fatalf("%v in use during clearfat", obj.Rconv(mips.REGRT1))
}
var r0 gc.Node
gc.Nodreg(&r0, gc.Types[gc.TUINT64], mips.REGZERO)
var dst gc.Node
gc.Nodreg(&dst, gc.Types[gc.Tptr], mips.REGRT1)
gc.Regrealloc(&dst)
gc.Agen(nl, &dst)
var boff uint64
if q > 128 {
p := gins(mips.ASUBV, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p = gins(mips.AMOVV, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(q * 8)
p = gins(mips.AMOVV, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 8
pl := p
p = gins(mips.AADDV, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
gc.Patch(ginsbranch(mips.ABNE, nil, &dst, &end, 0), pl)
gc.Regfree(&end)
// The loop leaves R1 on the last zeroed dword
boff = 8
// TODO(dfc): https://golang.org/issue/12108
// If DUFFZERO is used inside a tail call (see genwrapper) it will
// overwrite the link register.
} else if false && q >= 4 {
p := gins(mips.ASUBV, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
f := gc.Sysfunc("duffzero")
p = gins(obj.ADUFFZERO, nil, f)
gc.Afunclit(&p.To, f)
// 8 and 128 = magic constants: see ../../runtime/asm_mips64x.s
p.To.Offset = int64(8 * (128 - q))
// duffzero leaves R1 on the last zeroed dword
boff = 8
} else {
var p *obj.Prog
for t := uint64(0); t < q; t++ {
p = gins(mips.AMOVV, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(8 * t)
}
boff = 8 * q
}
var p *obj.Prog
for t := uint64(0); t < c; t++ {
p = gins(mips.AMOVB, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(t + boff)
}
gc.Regfree(&dst)
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op gc.Op, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
a := optoas(op, nl.Type)
if nr.Op == gc.OLITERAL {
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gc.Cgen(nl, &n1)
sc := uint64(nr.Int64())
if sc >= uint64(nl.Type.Width*8) {
// large shift gets 2 shifts by width-1
var n3 gc.Node
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n1)
gins(a, &n3, &n1)
} else {
gins(a, nr, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
if nl.Ullman >= gc.UINF {
var n4 gc.Node
gc.Tempname(&n4, nl.Type)
gc.Cgen(nl, &n4)
nl = &n4
}
if nr.Ullman >= gc.UINF {
var n5 gc.Node
gc.Tempname(&n5, nr.Type)
gc.Cgen(nr, &n5)
nr = &n5
}
// Allow either uint32 or uint64 as shift type,
// to avoid unnecessary conversion from uint32 to uint64
// just to do the comparison.
tcount := gc.Types[gc.Simtype[nr.Type.Etype]]
if tcount.Etype < gc.TUINT32 {
tcount = gc.Types[gc.TUINT32]
}
var n1 gc.Node
gc.Regalloc(&n1, nr.Type, nil) // to hold the shift type in CX
var n3 gc.Node
gc.Regalloc(&n3, tcount, &n1) // to clear high bits of CX
var n2 gc.Node
gc.Regalloc(&n2, nl.Type, res)
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &n2)
gc.Cgen(nr, &n1)
gmove(&n1, &n3)
} else {
gc.Cgen(nr, &n1)
gmove(&n1, &n3)
gc.Cgen(nl, &n2)
}
gc.Regfree(&n3)
// test and fix up large shifts
if !bounded {
gc.Nodconst(&n3, tcount, nl.Type.Width*8)
gins(optoas(gc.OCMP, tcount), &n1, &n3)
p1 := gc.Gbranch(optoas(gc.OLT, tcount), nil, 1)
if op == gc.ORSH && nl.Type.IsSigned() {
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n2)
} else {
gc.Nodconst(&n3, nl.Type, 0)
gmove(&n3, &n2)
}
gc.Patch(p1, gc.Pc)
}
gins(a, &n1, &n2)
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
func blockcopy(n, res *gc.Node, osrc, odst, w int64) {
// determine alignment.
// want to avoid unaligned access, so have to use
// smaller operations for less aligned types.
// for example moving [4]byte must use 4 MOVB not 1 MOVW.
align := int(n.Type.Align)
var op obj.As
switch align {
default:
gc.Fatalf("sgen: invalid alignment %d for %v", align, n.Type)
case 1:
op = 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 && odst < 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)
}
// zerorange clears the stack in the given range.
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
// Adjust the frame to account for LR.
frame += gc.Ctxt.FixedFrameSize()
offset := frame + lo
reg := int16(s390x.REGSP)
// If the offset cannot fit in a 12-bit unsigned displacement then we
// need to create a copy of the stack pointer that we can adjust.
// We also need to do this if we are going to loop.
if offset < 0 || offset > 4096-clearLoopCutoff || cnt > clearLoopCutoff {
p = appendpp(p, s390x.AADD, obj.TYPE_CONST, 0, offset, obj.TYPE_REG, s390x.REGRT1, 0)
p.Reg = int16(s390x.REGSP)
reg = s390x.REGRT1
offset = 0
}
// Generate a loop of large clears.
if cnt > clearLoopCutoff {
n := cnt - (cnt % 256)
end := int16(s390x.REGRT2)
p = appendpp(p, s390x.AADD, obj.TYPE_CONST, 0, offset+n, obj.TYPE_REG, end, 0)
p.Reg = reg
p = appendpp(p, s390x.AXC, obj.TYPE_MEM, reg, offset, obj.TYPE_MEM, reg, offset)
p.From3 = new(obj.Addr)
p.From3.Type = obj.TYPE_CONST
p.From3.Offset = 256
pl := p
p = appendpp(p, s390x.AADD, obj.TYPE_CONST, 0, 256, obj.TYPE_REG, reg, 0)
p = appendpp(p, s390x.ACMP, obj.TYPE_REG, reg, 0, obj.TYPE_REG, end, 0)
p = appendpp(p, s390x.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
gc.Patch(p, pl)
cnt -= n
}
// Generate remaining clear instructions without a loop.
for cnt > 0 {
n := cnt
// Can clear at most 256 bytes per instruction.
if n > 256 {
n = 256
}
switch n {
// Handle very small clears with move instructions.
case 8, 4, 2, 1:
ins := s390x.AMOVB
switch n {
case 8:
ins = s390x.AMOVD
case 4:
ins = s390x.AMOVW
case 2:
ins = s390x.AMOVH
}
p = appendpp(p, ins, obj.TYPE_CONST, 0, 0, obj.TYPE_MEM, reg, offset)
// Handle clears that would require multiple move instructions with XC.
default:
p = appendpp(p, s390x.AXC, obj.TYPE_MEM, reg, offset, obj.TYPE_MEM, reg, offset)
p.From3 = new(obj.Addr)
p.From3.Type = obj.TYPE_CONST
p.From3.Offset = n
}
cnt -= n
offset += n
}
return p
}
/*
* generate move:
* t = f
* hard part is conversions.
*/
func gmove(f *gc.Node, t *gc.Node) {
if gc.Debug['M'] != 0 {
fmt.Printf("gmove %v -> %v\n", gc.Nconv(f, gc.FmtLong), gc.Nconv(t, gc.FmtLong))
}
ft := int(gc.Simsimtype(f.Type))
tt := int(gc.Simsimtype(t.Type))
cvt := t.Type
if gc.Iscomplex[ft] || gc.Iscomplex[tt] {
gc.Complexmove(f, t)
return
}
// cannot have two memory operands
var r2 gc.Node
var r1 gc.Node
var a obj.As
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
switch tt {
default:
f.Convconst(&con, t.Type)
case gc.TINT32,
gc.TINT16,
gc.TINT8:
var con gc.Node
f.Convconst(&con, gc.Types[gc.TINT64])
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(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, gc.FmtLong), gc.Tconv(t.Type, gc.FmtLong))
/*
* integer copy and truncate
*/
case gc.TINT8<<16 | gc.TINT8, // same size
gc.TUINT8<<16 | gc.TINT8,
gc.TINT16<<16 | gc.TINT8,
// truncate
gc.TUINT16<<16 | gc.TINT8,
gc.TINT32<<16 | gc.TINT8,
gc.TUINT32<<16 | gc.TINT8,
gc.TINT64<<16 | gc.TINT8,
gc.TUINT64<<16 | gc.TINT8:
a = 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 := 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 := 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 := 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
/*
* 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()
// The algorithm is:
// if small enough, use native int64 -> float64 conversion,
// otherwise halve (x -> (x>>1)|(x&1)), convert, and double.
// Note: could use FCFIDU instead if target supports it.
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 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT64]), nil, +1)
var r3 gc.Node
gc.Regalloc(&r3, gc.Types[gc.TUINT64], nil)
p2 := gins(ppc64.AANDCC, nil, &r3) // andi.
p2.Reg = r1.Reg
p2.From.Type = obj.TYPE_CONST
p2.From.Offset = 1
p3 := gins(ppc64.ASRD, nil, &r1)
p3.From.Type = obj.TYPE_CONST
p3.From.Offset = 1
gins(ppc64.AOR, &r3, &r1)
gc.Regfree(&r3)
gc.Patch(p1, gc.Pc)
}
gc.Regalloc(&r2, gc.Types[gc.TFLOAT64], t)
p1 := 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 := 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
}