/*
* n is a 64-bit value. fill in lo and hi to refer to its 32-bit halves.
*/
func split64(n *gc.Node, lo *gc.Node, hi *gc.Node) {
if !gc.Is64(n.Type) {
gc.Fatal("split64 %v", gc.Tconv(n.Type, 0))
}
if nsclean >= len(sclean) {
gc.Fatal("split64 clean")
}
sclean[nsclean].Op = gc.OEMPTY
nsclean++
switch n.Op {
default:
switch n.Op {
default:
var n1 gc.Node
if !dotaddable(n, &n1) {
gc.Igen(n, &n1, nil)
sclean[nsclean-1] = n1
}
n = &n1
case gc.ONAME:
if n.Class == gc.PPARAMREF {
var n1 gc.Node
gc.Cgen(n.Heapaddr, &n1)
sclean[nsclean-1] = n1
n = &n1
}
// nothing
case gc.OINDREG:
break
}
*lo = *n
*hi = *n
lo.Type = gc.Types[gc.TUINT32]
if n.Type.Etype == gc.TINT64 {
hi.Type = gc.Types[gc.TINT32]
} else {
hi.Type = gc.Types[gc.TUINT32]
}
hi.Xoffset += 4
case gc.OLITERAL:
var n1 gc.Node
gc.Convconst(&n1, n.Type, &n.Val)
i := gc.Mpgetfix(n1.Val.U.Xval)
gc.Nodconst(lo, gc.Types[gc.TUINT32], int64(uint32(i)))
i >>= 32
if n.Type.Etype == gc.TINT64 {
gc.Nodconst(hi, gc.Types[gc.TINT32], int64(int32(i)))
} else {
gc.Nodconst(hi, gc.Types[gc.TUINT32], int64(uint32(i)))
}
}
}
/*
* call to n has already been generated.
* generate:
* res = &return value from call.
*/
func cgen_aret(n *gc.Node, res *gc.Node) {
t := n.Left.Type
if gc.Isptr[t.Etype] {
t = t.Type
}
var flist gc.Iter
fp := gc.Structfirst(&flist, gc.Getoutarg(t))
if fp == nil {
gc.Fatal("cgen_aret: nil")
}
var nod1 gc.Node
nod1.Op = gc.OINDREG
nod1.Val.U.Reg = ppc64.REGSP
nod1.Addable = 1
nod1.Xoffset = fp.Width + int64(gc.Widthptr) // +widthptr: saved lr at 0(SP)
nod1.Type = fp.Type
if res.Op != gc.OREGISTER {
var nod2 gc.Node
regalloc(&nod2, gc.Types[gc.Tptr], res)
agen(&nod1, &nod2)
gins(ppc64.AMOVD, &nod2, res)
regfree(&nod2)
} else {
agen(&nod1, res)
}
}
/*
* call to n has already been generated.
* generate:
* res = &return value from call.
*/
func cgen_aret(n *gc.Node, res *gc.Node) {
t := n.Left.Type
if gc.Isptr[t.Etype] {
t = t.Type
}
var flist gc.Iter
fp := gc.Structfirst(&flist, gc.Getoutarg(t))
if fp == nil {
gc.Fatal("cgen_aret: nil")
}
var nod1 gc.Node
nod1.Op = gc.OINDREG
nod1.Val.U.Reg = x86.REG_SP
nod1.Addable = 1
nod1.Xoffset = fp.Width
nod1.Type = fp.Type
if res.Op != gc.OREGISTER {
var nod2 gc.Node
regalloc(&nod2, gc.Types[gc.Tptr], res)
gins(x86.ALEAL, &nod1, &nod2)
gins(x86.AMOVL, &nod2, res)
regfree(&nod2)
} else {
gins(x86.ALEAL, &nod1, res)
}
}
func restx(x *gc.Node, oldx *gc.Node) {
if oldx.Op != 0 {
x.Type = gc.Types[gc.TINT64]
reg[x.Reg] = uint8(oldx.Ostk)
gmove(oldx, x)
gc.Regfree(oldx)
}
}
/*
* register dr is one of the special ones (AX, CX, DI, SI, etc.).
* we need to use it. if it is already allocated as a temporary
* (r > 1; can only happen if a routine like sgen passed a
* special as cgen's res and then cgen used regalloc to reuse
* it as its own temporary), then move it for now to another
* register. caller must call restx to move it back.
* the move is not necessary if dr == res, because res is
* known to be dead.
*/
func savex(dr int, x *gc.Node, oldx *gc.Node, res *gc.Node, t *gc.Type) {
r := int(reg[dr])
// save current ax and dx if they are live
// and not the destination
*oldx = gc.Node{}
gc.Nodreg(x, t, dr)
if r > 1 && !gc.Samereg(x, res) {
gc.Regalloc(oldx, gc.Types[gc.TINT64], nil)
x.Type = gc.Types[gc.TINT64]
gmove(x, oldx)
x.Type = t
oldx.Ostk = int32(r) // squirrel away old r value
reg[dr] = 1
}
}
func restx(x *gc.Node, oldx *gc.Node) {
gc.Regfree(x)
if oldx.Op != 0 {
x.Type = gc.Types[gc.TINT32]
gmove(oldx, x)
}
}
/*
* generate function call;
* proc=0 normal call
* proc=1 goroutine run in new proc
* proc=2 defer call save away stack
*/
func cgen_call(n *gc.Node, proc int) {
if n == nil {
return
}
var afun gc.Node
if n.Left.Ullman >= gc.UINF {
// if name involves a fn call
// precompute the address of the fn
gc.Tempname(&afun, gc.Types[gc.Tptr])
cgen(n.Left, &afun)
}
gc.Genlist(n.List) // assign the args
t := n.Left.Type
// call tempname pointer
if n.Left.Ullman >= gc.UINF {
var nod gc.Node
regalloc(&nod, gc.Types[gc.Tptr], nil)
gc.Cgen_as(&nod, &afun)
nod.Type = t
ginscall(&nod, proc)
regfree(&nod)
return
}
// call pointer
if n.Left.Op != gc.ONAME || n.Left.Class != gc.PFUNC {
var nod gc.Node
regalloc(&nod, gc.Types[gc.Tptr], nil)
gc.Cgen_as(&nod, n.Left)
nod.Type = t
ginscall(&nod, proc)
regfree(&nod)
return
}
// call direct
n.Left.Method = 1
ginscall(n.Left, proc)
}
func dotaddable(n *gc.Node, n1 *gc.Node) bool {
if n.Op != gc.ODOT {
return false
}
var oary [10]int64
var nn *gc.Node
o := gc.Dotoffset(n, oary[:], &nn)
if nn != nil && nn.Addable && o == 1 && oary[0] >= 0 {
*n1 = *nn
n1.Type = n.Type
n1.Xoffset += oary[0]
return true
}
return false
}
/*
* call to n has already been generated.
* generate:
* res = return value from call.
*/
func cgen_callret(n *gc.Node, res *gc.Node) {
t := n.Left.Type
if t.Etype == gc.TPTR32 || t.Etype == gc.TPTR64 {
t = t.Type
}
var flist gc.Iter
fp := gc.Structfirst(&flist, gc.Getoutarg(t))
if fp == nil {
gc.Fatal("cgen_callret: nil")
}
var nod gc.Node
nod.Op = gc.OINDREG
nod.Val.U.Reg = x86.REG_SP
nod.Addable = 1
nod.Xoffset = fp.Width
nod.Type = fp.Type
gc.Cgen_as(res, &nod)
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op int, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
a := optoas(op, nl.Type)
if nr.Op == gc.OLITERAL {
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gc.Cgen(nl, &n1)
sc := uint64(gc.Mpgetfix(nr.Val.U.Xval))
if sc >= uint64(nl.Type.Width*8) {
// large shift gets 2 shifts by width-1
var n3 gc.Node
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n1)
gins(a, &n3, &n1)
} else {
gins(a, nr, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
if nl.Ullman >= gc.UINF {
var n4 gc.Node
gc.Tempname(&n4, nl.Type)
gc.Cgen(nl, &n4)
nl = &n4
}
if nr.Ullman >= gc.UINF {
var n5 gc.Node
gc.Tempname(&n5, nr.Type)
gc.Cgen(nr, &n5)
nr = &n5
}
rcx := int(reg[x86.REG_CX])
var n1 gc.Node
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
// Allow either uint32 or uint64 as shift type,
// to avoid unnecessary conversion from uint32 to uint64
// just to do the comparison.
tcount := gc.Types[gc.Simtype[nr.Type.Etype]]
if tcount.Etype < gc.TUINT32 {
tcount = gc.Types[gc.TUINT32]
}
gc.Regalloc(&n1, nr.Type, &n1) // to hold the shift type in CX
var n3 gc.Node
gc.Regalloc(&n3, tcount, &n1) // to clear high bits of CX
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TUINT64], x86.REG_CX)
var oldcx gc.Node
if rcx > 0 && !gc.Samereg(&cx, res) {
gc.Regalloc(&oldcx, gc.Types[gc.TUINT64], nil)
gmove(&cx, &oldcx)
}
cx.Type = tcount
var n2 gc.Node
if gc.Samereg(&cx, res) {
gc.Regalloc(&n2, nl.Type, nil)
} else {
gc.Regalloc(&n2, nl.Type, res)
}
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &n2)
gc.Cgen(nr, &n1)
gmove(&n1, &n3)
} else {
gc.Cgen(nr, &n1)
gmove(&n1, &n3)
gc.Cgen(nl, &n2)
}
gc.Regfree(&n3)
// test and fix up large shifts
if !bounded {
gc.Nodconst(&n3, tcount, nl.Type.Width*8)
gins(optoas(gc.OCMP, tcount), &n1, &n3)
p1 := gc.Gbranch(optoas(gc.OLT, tcount), nil, +1)
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n2)
} else {
gc.Nodconst(&n3, nl.Type, 0)
gmove(&n3, &n2)
}
gc.Patch(p1, gc.Pc)
}
gins(a, &n1, &n2)
if oldcx.Op != 0 {
cx.Type = gc.Types[gc.TUINT64]
gmove(&oldcx, &cx)
gc.Regfree(&oldcx)
}
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
func stackcopy(n, ns *gc.Node, osrc, odst, w int64) {
var noddi gc.Node
gc.Nodreg(&noddi, gc.Types[gc.Tptr], x86.REG_DI)
var nodsi gc.Node
gc.Nodreg(&nodsi, gc.Types[gc.Tptr], x86.REG_SI)
var nodl gc.Node
var nodr gc.Node
if n.Ullman >= ns.Ullman {
gc.Agenr(n, &nodr, &nodsi)
if ns.Op == gc.ONAME {
gc.Gvardef(ns)
}
gc.Agenr(ns, &nodl, &noddi)
} else {
if ns.Op == gc.ONAME {
gc.Gvardef(ns)
}
gc.Agenr(ns, &nodl, &noddi)
gc.Agenr(n, &nodr, &nodsi)
}
if nodl.Val.U.Reg != x86.REG_DI {
gmove(&nodl, &noddi)
}
if nodr.Val.U.Reg != x86.REG_SI {
gmove(&nodr, &nodsi)
}
gc.Regfree(&nodl)
gc.Regfree(&nodr)
c := w % 8 // bytes
q := w / 8 // quads
var oldcx gc.Node
var cx gc.Node
savex(x86.REG_CX, &cx, &oldcx, nil, gc.Types[gc.TINT64])
// if we are copying forward on the stack and
// the src and dst overlap, then reverse direction
if osrc < odst && odst < osrc+w {
// reverse direction
gins(x86.ASTD, nil, nil) // set direction flag
if c > 0 {
gconreg(addptr, w-1, x86.REG_SI)
gconreg(addptr, w-1, x86.REG_DI)
gconreg(movptr, c, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSB, nil, nil) // MOVB *(SI)-,*(DI)-
}
if q > 0 {
if c > 0 {
gconreg(addptr, -7, x86.REG_SI)
gconreg(addptr, -7, x86.REG_DI)
} else {
gconreg(addptr, w-8, x86.REG_SI)
gconreg(addptr, w-8, x86.REG_DI)
}
gconreg(movptr, q, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)-,*(DI)-
}
// we leave with the flag clear
gins(x86.ACLD, nil, nil)
} else {
// normal direction
if q > 128 || (gc.Nacl && q >= 4) {
gconreg(movptr, q, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)+,*(DI)+
} else if q >= 4 {
p := gins(obj.ADUFFCOPY, nil, nil)
p.To.Type = obj.TYPE_ADDR
p.To.Sym = gc.Linksym(gc.Pkglookup("duffcopy", gc.Runtimepkg))
// 14 and 128 = magic constants: see ../../runtime/asm_amd64.s
p.To.Offset = 14 * (128 - q)
} else if !gc.Nacl && c == 0 {
// We don't need the MOVSQ side-effect of updating SI and DI,
// and issuing a sequence of MOVQs directly is faster.
nodsi.Op = gc.OINDREG
noddi.Op = gc.OINDREG
for q > 0 {
gmove(&nodsi, &cx) // MOVQ x+(SI),CX
gmove(&cx, &noddi) // MOVQ CX,x+(DI)
nodsi.Xoffset += 8
noddi.Xoffset += 8
q--
}
} else {
for q > 0 {
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)+,*(DI)+
q--
}
}
// copy the remaining c bytes
if w < 4 || c <= 1 || (odst < osrc && osrc < odst+w) {
for c > 0 {
gins(x86.AMOVSB, nil, nil) // MOVB *(SI)+,*(DI)+
c--
}
} else if w < 8 || c <= 4 {
nodsi.Op = gc.OINDREG
noddi.Op = gc.OINDREG
cx.Type = gc.Types[gc.TINT32]
nodsi.Type = gc.Types[gc.TINT32]
noddi.Type = gc.Types[gc.TINT32]
if c > 4 {
nodsi.Xoffset = 0
noddi.Xoffset = 0
gmove(&nodsi, &cx)
gmove(&cx, &noddi)
}
nodsi.Xoffset = c - 4
noddi.Xoffset = c - 4
gmove(&nodsi, &cx)
gmove(&cx, &noddi)
} else {
nodsi.Op = gc.OINDREG
noddi.Op = gc.OINDREG
cx.Type = gc.Types[gc.TINT64]
nodsi.Type = gc.Types[gc.TINT64]
noddi.Type = gc.Types[gc.TINT64]
nodsi.Xoffset = c - 8
noddi.Xoffset = c - 8
gmove(&nodsi, &cx)
gmove(&cx, &noddi)
}
}
restx(&cx, &oldcx)
}
/*
* allocate a register (reusing res if possible) and generate
* a = &n
* The caller must call regfree(a).
* The generated code checks that the result is not nil.
*/
func agenr(n *gc.Node, a *gc.Node, res *gc.Node) {
if gc.Debug['g'] != 0 {
gc.Dump("agenr-n", n)
}
nl := n.Left
nr := n.Right
switch n.Op {
case gc.ODOT,
gc.ODOTPTR,
gc.OCALLFUNC,
gc.OCALLMETH,
gc.OCALLINTER:
var n1 gc.Node
igen(n, &n1, res)
regalloc(a, gc.Types[gc.Tptr], &n1)
agen(&n1, a)
regfree(&n1)
case gc.OIND:
cgenr(n.Left, a, res)
gc.Cgen_checknil(a)
case gc.OINDEX:
var p2 *obj.Prog // to be patched to panicindex.
w := uint32(n.Type.Width)
//bounded = debug['B'] || n->bounded;
var n3 gc.Node
var n1 gc.Node
if nr.Addable != 0 {
var tmp gc.Node
if !gc.Isconst(nr, gc.CTINT) {
gc.Tempname(&tmp, gc.Types[gc.TINT64])
}
if !gc.Isconst(nl, gc.CTSTR) {
agenr(nl, &n3, res)
}
if !gc.Isconst(nr, gc.CTINT) {
cgen(nr, &tmp)
regalloc(&n1, tmp.Type, nil)
gmove(&tmp, &n1)
}
} else if nl.Addable != 0 {
if !gc.Isconst(nr, gc.CTINT) {
var tmp gc.Node
gc.Tempname(&tmp, gc.Types[gc.TINT64])
cgen(nr, &tmp)
regalloc(&n1, tmp.Type, nil)
gmove(&tmp, &n1)
}
if !gc.Isconst(nl, gc.CTSTR) {
agenr(nl, &n3, res)
}
} else {
var tmp gc.Node
gc.Tempname(&tmp, gc.Types[gc.TINT64])
cgen(nr, &tmp)
nr = &tmp
if !gc.Isconst(nl, gc.CTSTR) {
agenr(nl, &n3, res)
}
regalloc(&n1, tmp.Type, nil)
gins(optoas(gc.OAS, tmp.Type), &tmp, &n1)
}
// &a is in &n3 (allocated in res)
// i is in &n1 (if not constant)
// w is width
// constant index
if gc.Isconst(nr, gc.CTINT) {
if gc.Isconst(nl, gc.CTSTR) {
gc.Fatal("constant string constant index")
}
v := uint64(gc.Mpgetfix(nr.Val.U.Xval))
if gc.Isslice(nl.Type) || nl.Type.Etype == gc.TSTRING {
if gc.Debug['B'] == 0 && !n.Bounded {
n1 = n3
n1.Op = gc.OINDREG
n1.Type = gc.Types[gc.Tptr]
n1.Xoffset = int64(gc.Array_nel)
var n4 gc.Node
regalloc(&n4, n1.Type, nil)
gmove(&n1, &n4)
ginscon2(optoas(gc.OCMP, gc.Types[gc.TUINT64]), &n4, int64(v))
regfree(&n4)
p1 := gc.Gbranch(optoas(gc.OGT, gc.Types[gc.TUINT64]), nil, +1)
ginscall(gc.Panicindex, 0)
gc.Patch(p1, gc.Pc)
}
n1 = n3
n1.Op = gc.OINDREG
n1.Type = gc.Types[gc.Tptr]
n1.Xoffset = int64(gc.Array_array)
gmove(&n1, &n3)
}
if v*uint64(w) != 0 {
ginscon(optoas(gc.OADD, gc.Types[gc.Tptr]), int64(v*uint64(w)), &n3)
}
*a = n3
break
}
var n2 gc.Node
regalloc(&n2, gc.Types[gc.TINT64], &n1) // i
gmove(&n1, &n2)
regfree(&n1)
var n4 gc.Node
if gc.Debug['B'] == 0 && !n.Bounded {
// check bounds
if gc.Isconst(nl, gc.CTSTR) {
gc.Nodconst(&n4, gc.Types[gc.TUINT64], int64(len(nl.Val.U.Sval)))
} else if gc.Isslice(nl.Type) || nl.Type.Etype == gc.TSTRING {
n1 = n3
n1.Op = gc.OINDREG
n1.Type = gc.Types[gc.Tptr]
n1.Xoffset = int64(gc.Array_nel)
regalloc(&n4, gc.Types[gc.TUINT64], nil)
gmove(&n1, &n4)
} else {
if nl.Type.Bound < (1<<15)-1 {
gc.Nodconst(&n4, gc.Types[gc.TUINT64], nl.Type.Bound)
} else {
regalloc(&n4, gc.Types[gc.TUINT64], nil)
p1 := gins(ppc64.AMOVD, nil, &n4)
p1.From.Type = obj.TYPE_CONST
p1.From.Offset = nl.Type.Bound
}
}
gins(optoas(gc.OCMP, gc.Types[gc.TUINT64]), &n2, &n4)
if n4.Op == gc.OREGISTER {
regfree(&n4)
}
p1 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT64]), nil, +1)
if p2 != nil {
gc.Patch(p2, gc.Pc)
}
ginscall(gc.Panicindex, 0)
gc.Patch(p1, gc.Pc)
}
if gc.Isconst(nl, gc.CTSTR) {
regalloc(&n3, gc.Types[gc.Tptr], res)
p1 := gins(ppc64.AMOVD, nil, &n3)
gc.Datastring(nl.Val.U.Sval, &p1.From)
p1.From.Type = obj.TYPE_ADDR
} else if gc.Isslice(nl.Type) || nl.Type.Etype == gc.TSTRING {
n1 = n3
n1.Op = gc.OINDREG
n1.Type = gc.Types[gc.Tptr]
n1.Xoffset = int64(gc.Array_array)
gmove(&n1, &n3)
}
if w == 0 {
} else // nothing to do
if w == 1 {
/* w already scaled */
gins(optoas(gc.OADD, gc.Types[gc.Tptr]), &n2, &n3)
/* else if(w == 2 || w == 4 || w == 8) {
// TODO(minux): scale using shift
} */
} else {
regalloc(&n4, gc.Types[gc.TUINT64], nil)
gc.Nodconst(&n1, gc.Types[gc.TUINT64], int64(w))
gmove(&n1, &n4)
gins(optoas(gc.OMUL, gc.Types[gc.TUINT64]), &n4, &n2)
gins(optoas(gc.OADD, gc.Types[gc.Tptr]), &n2, &n3)
regfree(&n4)
}
*a = n3
regfree(&n2)
default:
regalloc(a, gc.Types[gc.Tptr], res)
agen(n, a)
}
}
/*
* generate:
* res = n;
* simplifies and calls gmove.
*/
func cgen(n *gc.Node, res *gc.Node) {
//print("cgen %N(%d) -> %N(%d)\n", n, n->addable, res, res->addable);
if gc.Debug['g'] != 0 {
gc.Dump("\ncgen-n", n)
gc.Dump("cgen-res", res)
}
if n == nil || n.Type == nil {
return
}
if res == nil || res.Type == nil {
gc.Fatal("cgen: res nil")
}
for n.Op == gc.OCONVNOP {
n = n.Left
}
switch n.Op {
case gc.OSLICE,
gc.OSLICEARR,
gc.OSLICESTR,
gc.OSLICE3,
gc.OSLICE3ARR:
if res.Op != gc.ONAME || res.Addable == 0 {
var n1 gc.Node
gc.Tempname(&n1, n.Type)
gc.Cgen_slice(n, &n1)
cgen(&n1, res)
} else {
gc.Cgen_slice(n, res)
}
return
case gc.OEFACE:
if res.Op != gc.ONAME || res.Addable == 0 {
var n1 gc.Node
gc.Tempname(&n1, n.Type)
gc.Cgen_eface(n, &n1)
cgen(&n1, res)
} else {
gc.Cgen_eface(n, res)
}
return
}
if n.Ullman >= gc.UINF {
if n.Op == gc.OINDREG {
gc.Fatal("cgen: this is going to misscompile")
}
if res.Ullman >= gc.UINF {
var n1 gc.Node
gc.Tempname(&n1, n.Type)
cgen(n, &n1)
cgen(&n1, res)
return
}
}
if gc.Isfat(n.Type) {
if n.Type.Width < 0 {
gc.Fatal("forgot to compute width for %v", gc.Tconv(n.Type, 0))
}
sgen(n, res, n.Type.Width)
return
}
if res.Addable == 0 {
if n.Ullman > res.Ullman {
var n1 gc.Node
regalloc(&n1, n.Type, res)
cgen(n, &n1)
if n1.Ullman > res.Ullman {
gc.Dump("n1", &n1)
gc.Dump("res", res)
gc.Fatal("loop in cgen")
}
cgen(&n1, res)
regfree(&n1)
return
}
var f int
if res.Ullman >= gc.UINF {
goto gen
}
if gc.Complexop(n, res) {
gc.Complexgen(n, res)
return
}
f = 1 // gen thru register
switch n.Op {
case gc.OLITERAL:
if gc.Smallintconst(n) {
f = 0
}
case gc.OREGISTER:
f = 0
}
if !gc.Iscomplex[n.Type.Etype] {
a := optoas(gc.OAS, res.Type)
var addr obj.Addr
if sudoaddable(a, res, &addr) {
var p1 *obj.Prog
if f != 0 {
var n2 gc.Node
regalloc(&n2, res.Type, nil)
cgen(n, &n2)
p1 = gins(a, &n2, nil)
regfree(&n2)
} else {
p1 = gins(a, n, nil)
}
p1.To = addr
if gc.Debug['g'] != 0 {
fmt.Printf("%v [ignore previous line]\n", p1)
}
sudoclean()
return
}
}
gen:
var n1 gc.Node
igen(res, &n1, nil)
cgen(n, &n1)
regfree(&n1)
return
}
// update addressability for string, slice
// can't do in walk because n->left->addable
// changes if n->left is an escaping local variable.
switch n.Op {
case gc.OSPTR,
gc.OLEN:
if gc.Isslice(n.Left.Type) || gc.Istype(n.Left.Type, gc.TSTRING) {
n.Addable = n.Left.Addable
}
case gc.OCAP:
if gc.Isslice(n.Left.Type) {
n.Addable = n.Left.Addable
}
case gc.OITAB:
n.Addable = n.Left.Addable
}
if gc.Complexop(n, res) {
gc.Complexgen(n, res)
return
}
// if both are addressable, move
if n.Addable != 0 {
if n.Op == gc.OREGISTER || res.Op == gc.OREGISTER {
gmove(n, res)
} else {
var n1 gc.Node
regalloc(&n1, n.Type, nil)
gmove(n, &n1)
cgen(&n1, res)
regfree(&n1)
}
return
}
nl := n.Left
nr := n.Right
if nl != nil && nl.Ullman >= gc.UINF {
if nr != nil && nr.Ullman >= gc.UINF {
var n1 gc.Node
gc.Tempname(&n1, nl.Type)
cgen(nl, &n1)
n2 := *n
n2.Left = &n1
cgen(&n2, res)
return
}
}
if !gc.Iscomplex[n.Type.Etype] {
a := optoas(gc.OAS, n.Type)
var addr obj.Addr
if sudoaddable(a, n, &addr) {
if res.Op == gc.OREGISTER {
p1 := gins(a, nil, res)
p1.From = addr
} else {
var n2 gc.Node
regalloc(&n2, n.Type, nil)
p1 := gins(a, nil, &n2)
p1.From = addr
gins(a, &n2, res)
regfree(&n2)
}
sudoclean()
return
}
}
// TODO(minux): we shouldn't reverse FP comparisons, but then we need to synthesize
// OGE, OLE, and ONE ourselves.
// if(nl != N && isfloat[n->type->etype] && isfloat[nl->type->etype]) goto flt;
var a int
switch n.Op {
default:
gc.Dump("cgen", n)
gc.Fatal("cgen: unknown op %v", gc.Nconv(n, obj.FmtShort|obj.FmtSign))
// these call bgen to get a bool value
case gc.OOROR,
gc.OANDAND,
gc.OEQ,
gc.ONE,
gc.OLT,
gc.OLE,
gc.OGE,
gc.OGT,
gc.ONOT:
p1 := gc.Gbranch(ppc64.ABR, nil, 0)
p2 := gc.Pc
gmove(gc.Nodbool(true), res)
p3 := gc.Gbranch(ppc64.ABR, nil, 0)
gc.Patch(p1, gc.Pc)
bgen(n, true, 0, p2)
gmove(gc.Nodbool(false), res)
gc.Patch(p3, gc.Pc)
return
case gc.OPLUS:
cgen(nl, res)
return
// unary
case gc.OCOM:
a := optoas(gc.OXOR, nl.Type)
var n1 gc.Node
regalloc(&n1, nl.Type, nil)
cgen(nl, &n1)
var n2 gc.Node
gc.Nodconst(&n2, nl.Type, -1)
gins(a, &n2, &n1)
gmove(&n1, res)
regfree(&n1)
return
case gc.OMINUS:
if gc.Isfloat[nl.Type.Etype] {
nr = gc.Nodintconst(-1)
gc.Convlit(&nr, n.Type)
a = optoas(gc.OMUL, nl.Type)
goto sbop
}
a := optoas(int(n.Op), nl.Type)
// unary
var n1 gc.Node
regalloc(&n1, nl.Type, res)
cgen(nl, &n1)
gins(a, nil, &n1)
gmove(&n1, res)
regfree(&n1)
return
// symmetric binary
case gc.OAND,
gc.OOR,
gc.OXOR,
gc.OADD,
gc.OMUL:
a = optoas(int(n.Op), nl.Type)
goto sbop
// asymmetric binary
case gc.OSUB:
a = optoas(int(n.Op), nl.Type)
goto abop
case gc.OHMUL:
cgen_hmul(nl, nr, res)
case gc.OCONV:
if n.Type.Width > nl.Type.Width {
// If loading from memory, do conversion during load,
// so as to avoid use of 8-bit register in, say, int(*byteptr).
switch nl.Op {
case gc.ODOT,
gc.ODOTPTR,
gc.OINDEX,
gc.OIND,
gc.ONAME:
var n1 gc.Node
igen(nl, &n1, res)
var n2 gc.Node
regalloc(&n2, n.Type, res)
gmove(&n1, &n2)
gmove(&n2, res)
regfree(&n2)
regfree(&n1)
return
}
}
var n1 gc.Node
regalloc(&n1, nl.Type, res)
var n2 gc.Node
regalloc(&n2, n.Type, &n1)
cgen(nl, &n1)
// if we do the conversion n1 -> n2 here
// reusing the register, then gmove won't
// have to allocate its own register.
gmove(&n1, &n2)
gmove(&n2, res)
regfree(&n2)
regfree(&n1)
case gc.ODOT,
gc.ODOTPTR,
gc.OINDEX,
gc.OIND,
gc.ONAME: // PHEAP or PPARAMREF var
var n1 gc.Node
igen(n, &n1, res)
gmove(&n1, res)
regfree(&n1)
// interface table is first word of interface value
case gc.OITAB:
var n1 gc.Node
igen(nl, &n1, res)
n1.Type = n.Type
gmove(&n1, res)
regfree(&n1)
// pointer is the first word of string or slice.
case gc.OSPTR:
if gc.Isconst(nl, gc.CTSTR) {
var n1 gc.Node
regalloc(&n1, gc.Types[gc.Tptr], res)
p1 := gins(ppc64.AMOVD, nil, &n1)
gc.Datastring(nl.Val.U.Sval, &p1.From)
gmove(&n1, res)
regfree(&n1)
break
}
var n1 gc.Node
igen(nl, &n1, res)
n1.Type = n.Type
gmove(&n1, res)
regfree(&n1)
case gc.OLEN:
if gc.Istype(nl.Type, gc.TMAP) || gc.Istype(nl.Type, gc.TCHAN) {
// map and chan have len in the first int-sized word.
// a zero pointer means zero length
var n1 gc.Node
regalloc(&n1, gc.Types[gc.Tptr], res)
cgen(nl, &n1)
var n2 gc.Node
gc.Nodconst(&n2, gc.Types[gc.Tptr], 0)
gins(optoas(gc.OCMP, gc.Types[gc.Tptr]), &n1, &n2)
p1 := gc.Gbranch(optoas(gc.OEQ, gc.Types[gc.Tptr]), nil, 0)
n2 = n1
n2.Op = gc.OINDREG
n2.Type = gc.Types[gc.Simtype[gc.TINT]]
gmove(&n2, &n1)
gc.Patch(p1, gc.Pc)
gmove(&n1, res)
regfree(&n1)
break
}
if gc.Istype(nl.Type, gc.TSTRING) || gc.Isslice(nl.Type) {
// both slice and string have len one pointer into the struct.
// a zero pointer means zero length
var n1 gc.Node
igen(nl, &n1, res)
n1.Type = gc.Types[gc.Simtype[gc.TUINT]]
n1.Xoffset += int64(gc.Array_nel)
gmove(&n1, res)
regfree(&n1)
break
}
gc.Fatal("cgen: OLEN: unknown type %v", gc.Tconv(nl.Type, obj.FmtLong))
case gc.OCAP:
if gc.Istype(nl.Type, gc.TCHAN) {
// chan has cap in the second int-sized word.
// a zero pointer means zero length
var n1 gc.Node
regalloc(&n1, gc.Types[gc.Tptr], res)
cgen(nl, &n1)
var n2 gc.Node
gc.Nodconst(&n2, gc.Types[gc.Tptr], 0)
gins(optoas(gc.OCMP, gc.Types[gc.Tptr]), &n1, &n2)
p1 := gc.Gbranch(optoas(gc.OEQ, gc.Types[gc.Tptr]), nil, 0)
n2 = n1
n2.Op = gc.OINDREG
n2.Xoffset = int64(gc.Widthint)
n2.Type = gc.Types[gc.Simtype[gc.TINT]]
gmove(&n2, &n1)
gc.Patch(p1, gc.Pc)
gmove(&n1, res)
regfree(&n1)
break
}
if gc.Isslice(nl.Type) {
var n1 gc.Node
igen(nl, &n1, res)
n1.Type = gc.Types[gc.Simtype[gc.TUINT]]
n1.Xoffset += int64(gc.Array_cap)
gmove(&n1, res)
regfree(&n1)
break
}
gc.Fatal("cgen: OCAP: unknown type %v", gc.Tconv(nl.Type, obj.FmtLong))
case gc.OADDR:
if n.Bounded { // let race detector avoid nil checks
gc.Disable_checknil++
}
agen(nl, res)
if n.Bounded {
gc.Disable_checknil--
}
case gc.OCALLMETH:
gc.Cgen_callmeth(n, 0)
cgen_callret(n, res)
case gc.OCALLINTER:
cgen_callinter(n, res, 0)
cgen_callret(n, res)
case gc.OCALLFUNC:
cgen_call(n, 0)
cgen_callret(n, res)
case gc.OMOD,
gc.ODIV:
if gc.Isfloat[n.Type.Etype] {
a = optoas(int(n.Op), nl.Type)
goto abop
}
if nl.Ullman >= nr.Ullman {
var n1 gc.Node
regalloc(&n1, nl.Type, res)
cgen(nl, &n1)
cgen_div(int(n.Op), &n1, nr, res)
regfree(&n1)
} else {
var n2 gc.Node
if !gc.Smallintconst(nr) {
regalloc(&n2, nr.Type, res)
cgen(nr, &n2)
} else {
n2 = *nr
}
cgen_div(int(n.Op), nl, &n2, res)
if n2.Op != gc.OLITERAL {
regfree(&n2)
}
}
case gc.OLSH,
gc.ORSH,
gc.OLROT:
cgen_shift(int(n.Op), n.Bounded, nl, nr, res)
}
return
/*
* put simplest on right - we'll generate into left
* and then adjust it using the computation of right.
* constants and variables have the same ullman
* count, so look for constants specially.
*
* an integer constant we can use as an immediate
* is simpler than a variable - we can use the immediate
* in the adjustment instruction directly - so it goes
* on the right.
*
* other constants, like big integers or floating point
* constants, require a mov into a register, so those
* might as well go on the left, so we can reuse that
* register for the computation.
*/
sbop: // symmetric binary
if nl.Ullman < nr.Ullman || (nl.Ullman == nr.Ullman && (gc.Smallintconst(nl) || (nr.Op == gc.OLITERAL && !gc.Smallintconst(nr)))) {
r := nl
nl = nr
nr = r
}
abop: // asymmetric binary
var n1 gc.Node
var n2 gc.Node
if nl.Ullman >= nr.Ullman {
regalloc(&n1, nl.Type, res)
cgen(nl, &n1)
/*
* This generates smaller code - it avoids a MOV - but it's
* easily 10% slower due to not being able to
* optimize/manipulate the move.
* To see, run: go test -bench . crypto/md5
* with and without.
*
if(sudoaddable(a, nr, &addr)) {
p1 = gins(a, N, &n1);
p1->from = addr;
gmove(&n1, res);
sudoclean();
regfree(&n1);
goto ret;
}
*
*/
// TODO(minux): enable using constants directly in certain instructions.
//if(smallintconst(nr))
// n2 = *nr;
//else {
regalloc(&n2, nr.Type, nil)
cgen(nr, &n2)
} else //}
{
//if(smallintconst(nr))
// n2 = *nr;
//else {
regalloc(&n2, nr.Type, res)
cgen(nr, &n2)
//}
regalloc(&n1, nl.Type, nil)
cgen(nl, &n1)
}
gins(a, &n2, &n1)
// Normalize result for types smaller than word.
if n.Type.Width < int64(gc.Widthreg) {
switch n.Op {
case gc.OADD,
gc.OSUB,
gc.OMUL,
gc.OLSH:
gins(optoas(gc.OAS, n.Type), &n1, &n1)
}
}
gmove(&n1, res)
regfree(&n1)
if n2.Op != gc.OLITERAL {
regfree(&n2)
}
return
}
/*
* copy a composite value by moving its individual components.
* Slices, strings and interfaces are supported.
* Small structs or arrays with elements of basic type are
* also supported.
* nr is N when assigning a zero value.
* return 1 if can do, 0 if can't.
*/
func componentgen(nr *gc.Node, nl *gc.Node) bool {
var nodl gc.Node
var nodr gc.Node
freel := 0
freer := 0
switch nl.Type.Etype {
default:
goto no
case gc.TARRAY:
t := nl.Type
// Slices are ok.
if gc.Isslice(t) {
break
}
// Small arrays are ok.
if t.Bound > 0 && t.Bound <= 3 && !gc.Isfat(t.Type) {
break
}
goto no
// Small structs with non-fat types are ok.
// Zero-sized structs are treated separately elsewhere.
case gc.TSTRUCT:
fldcount := int64(0)
for t := nl.Type.Type; t != nil; t = t.Down {
if gc.Isfat(t.Type) {
goto no
}
if t.Etype != gc.TFIELD {
gc.Fatal("componentgen: not a TFIELD: %v", gc.Tconv(t, obj.FmtLong))
}
fldcount++
}
if fldcount == 0 || fldcount > 4 {
goto no
}
case gc.TSTRING,
gc.TINTER:
break
}
nodl = *nl
if !cadable(nl) {
if nr != nil && !cadable(nr) {
goto no
}
igen(nl, &nodl, nil)
freel = 1
}
if nr != nil {
nodr = *nr
if !cadable(nr) {
igen(nr, &nodr, nil)
freer = 1
}
} else {
// When zeroing, prepare a register containing zero.
var tmp gc.Node
gc.Nodconst(&tmp, nl.Type, 0)
regalloc(&nodr, gc.Types[gc.TUINT], nil)
gmove(&tmp, &nodr)
freer = 1
}
// nl and nr are 'cadable' which basically means they are names (variables) now.
// If they are the same variable, don't generate any code, because the
// VARDEF we generate will mark the old value as dead incorrectly.
// (And also the assignments are useless.)
if nr != nil && nl.Op == gc.ONAME && nr.Op == gc.ONAME && nl == nr {
goto yes
}
switch nl.Type.Etype {
// componentgen for arrays.
case gc.TARRAY:
if nl.Op == gc.ONAME {
gc.Gvardef(nl)
}
t := nl.Type
if !gc.Isslice(t) {
nodl.Type = t.Type
nodr.Type = nodl.Type
for fldcount := int64(0); fldcount < t.Bound; fldcount++ {
if nr == nil {
gc.Clearslim(&nodl)
} else {
gmove(&nodr, &nodl)
}
nodl.Xoffset += t.Type.Width
nodr.Xoffset += t.Type.Width
}
goto yes
}
// componentgen for slices.
nodl.Xoffset += int64(gc.Array_array)
nodl.Type = gc.Ptrto(nl.Type.Type)
if nr != nil {
nodr.Xoffset += int64(gc.Array_array)
nodr.Type = nodl.Type
}
gmove(&nodr, &nodl)
nodl.Xoffset += int64(gc.Array_nel) - int64(gc.Array_array)
nodl.Type = gc.Types[gc.Simtype[gc.TUINT]]
if nr != nil {
nodr.Xoffset += int64(gc.Array_nel) - int64(gc.Array_array)
nodr.Type = nodl.Type
}
gmove(&nodr, &nodl)
nodl.Xoffset += int64(gc.Array_cap) - int64(gc.Array_nel)
nodl.Type = gc.Types[gc.Simtype[gc.TUINT]]
if nr != nil {
nodr.Xoffset += int64(gc.Array_cap) - int64(gc.Array_nel)
nodr.Type = nodl.Type
}
gmove(&nodr, &nodl)
goto yes
case gc.TSTRING:
if nl.Op == gc.ONAME {
gc.Gvardef(nl)
}
nodl.Xoffset += int64(gc.Array_array)
nodl.Type = gc.Ptrto(gc.Types[gc.TUINT8])
if nr != nil {
nodr.Xoffset += int64(gc.Array_array)
nodr.Type = nodl.Type
}
gmove(&nodr, &nodl)
nodl.Xoffset += int64(gc.Array_nel) - int64(gc.Array_array)
nodl.Type = gc.Types[gc.Simtype[gc.TUINT]]
if nr != nil {
nodr.Xoffset += int64(gc.Array_nel) - int64(gc.Array_array)
nodr.Type = nodl.Type
}
gmove(&nodr, &nodl)
goto yes
case gc.TINTER:
if nl.Op == gc.ONAME {
gc.Gvardef(nl)
}
nodl.Xoffset += int64(gc.Array_array)
nodl.Type = gc.Ptrto(gc.Types[gc.TUINT8])
if nr != nil {
nodr.Xoffset += int64(gc.Array_array)
nodr.Type = nodl.Type
}
gmove(&nodr, &nodl)
nodl.Xoffset += int64(gc.Array_nel) - int64(gc.Array_array)
nodl.Type = gc.Ptrto(gc.Types[gc.TUINT8])
if nr != nil {
nodr.Xoffset += int64(gc.Array_nel) - int64(gc.Array_array)
nodr.Type = nodl.Type
}
gmove(&nodr, &nodl)
goto yes
case gc.TSTRUCT:
if nl.Op == gc.ONAME {
gc.Gvardef(nl)
}
loffset := nodl.Xoffset
roffset := nodr.Xoffset
// funarg structs may not begin at offset zero.
if nl.Type.Etype == gc.TSTRUCT && nl.Type.Funarg != 0 && nl.Type.Type != nil {
loffset -= nl.Type.Type.Width
}
if nr != nil && nr.Type.Etype == gc.TSTRUCT && nr.Type.Funarg != 0 && nr.Type.Type != nil {
roffset -= nr.Type.Type.Width
}
for t := nl.Type.Type; t != nil; t = t.Down {
nodl.Xoffset = loffset + t.Width
nodl.Type = t.Type
if nr == nil {
gc.Clearslim(&nodl)
} else {
nodr.Xoffset = roffset + t.Width
nodr.Type = nodl.Type
gmove(&nodr, &nodl)
}
}
goto yes
}
no:
if freer != 0 {
regfree(&nodr)
}
if freel != 0 {
regfree(&nodl)
}
return false
yes:
if freer != 0 {
regfree(&nodr)
}
if freel != 0 {
regfree(&nodl)
}
return true
}
/*
* generate:
* if(n == true) goto to;
*/
func bgen(n *gc.Node, true_ bool, likely int, to *obj.Prog) {
if gc.Debug['g'] != 0 {
gc.Dump("\nbgen", n)
}
if n == nil {
n = gc.Nodbool(true)
}
if n.Ninit != nil {
gc.Genlist(n.Ninit)
}
if n.Type == nil {
gc.Convlit(&n, gc.Types[gc.TBOOL])
if n.Type == nil {
return
}
}
et := int(n.Type.Etype)
if et != gc.TBOOL {
gc.Yyerror("cgen: bad type %v for %v", gc.Tconv(n.Type, 0), gc.Oconv(int(n.Op), 0))
gc.Patch(gins(obj.AEND, nil, nil), to)
return
}
var nr *gc.Node
for n.Op == gc.OCONVNOP {
n = n.Left
if n.Ninit != nil {
gc.Genlist(n.Ninit)
}
}
var nl *gc.Node
switch n.Op {
default:
var n1 gc.Node
regalloc(&n1, n.Type, nil)
cgen(n, &n1)
var n2 gc.Node
gc.Nodconst(&n2, n.Type, 0)
gins(optoas(gc.OCMP, n.Type), &n1, &n2)
a := ppc64.ABNE
if !true_ {
a = ppc64.ABEQ
}
gc.Patch(gc.Gbranch(a, n.Type, likely), to)
regfree(&n1)
return
// need to ask if it is bool?
case gc.OLITERAL:
if !true_ == (n.Val.U.Bval == 0) {
gc.Patch(gc.Gbranch(ppc64.ABR, nil, likely), to)
}
return
case gc.OANDAND,
gc.OOROR:
if (n.Op == gc.OANDAND) == true_ {
p1 := gc.Gbranch(obj.AJMP, nil, 0)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
bgen(n.Left, !true_, -likely, p2)
bgen(n.Right, !true_, -likely, p2)
p1 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, to)
gc.Patch(p2, gc.Pc)
} else {
bgen(n.Left, true_, likely, to)
bgen(n.Right, true_, likely, to)
}
return
case gc.OEQ,
gc.ONE,
gc.OLT,
gc.OGT,
gc.OLE,
gc.OGE:
nr = n.Right
if nr == nil || nr.Type == nil {
return
}
fallthrough
case gc.ONOT: // unary
nl = n.Left
if nl == nil || nl.Type == nil {
return
}
}
switch n.Op {
case gc.ONOT:
bgen(nl, !true_, likely, to)
return
case gc.OEQ,
gc.ONE,
gc.OLT,
gc.OGT,
gc.OLE,
gc.OGE:
a := int(n.Op)
if !true_ {
if gc.Isfloat[nr.Type.Etype] {
// brcom is not valid on floats when NaN is involved.
p1 := gc.Gbranch(ppc64.ABR, nil, 0)
p2 := gc.Gbranch(ppc64.ABR, nil, 0)
gc.Patch(p1, gc.Pc)
ll := n.Ninit // avoid re-genning ninit
n.Ninit = nil
bgen(n, true, -likely, p2)
n.Ninit = ll
gc.Patch(gc.Gbranch(ppc64.ABR, nil, 0), to)
gc.Patch(p2, gc.Pc)
return
}
a = gc.Brcom(a)
true_ = !true_
}
// make simplest on right
if nl.Op == gc.OLITERAL || (nl.Ullman < nr.Ullman && nl.Ullman < gc.UINF) {
a = gc.Brrev(a)
r := nl
nl = nr
nr = r
}
if gc.Isslice(nl.Type) {
// front end should only leave cmp to literal nil
if (a != gc.OEQ && a != gc.ONE) || nr.Op != gc.OLITERAL {
gc.Yyerror("illegal slice comparison")
break
}
a = optoas(a, gc.Types[gc.Tptr])
var n1 gc.Node
igen(nl, &n1, nil)
n1.Xoffset += int64(gc.Array_array)
n1.Type = gc.Types[gc.Tptr]
var tmp gc.Node
gc.Nodconst(&tmp, gc.Types[gc.Tptr], 0)
var n2 gc.Node
regalloc(&n2, gc.Types[gc.Tptr], &n1)
gmove(&n1, &n2)
gins(optoas(gc.OCMP, gc.Types[gc.Tptr]), &n2, &tmp)
regfree(&n2)
gc.Patch(gc.Gbranch(a, gc.Types[gc.Tptr], likely), to)
regfree(&n1)
break
}
if gc.Isinter(nl.Type) {
// front end should only leave cmp to literal nil
if (a != gc.OEQ && a != gc.ONE) || nr.Op != gc.OLITERAL {
gc.Yyerror("illegal interface comparison")
break
}
a = optoas(a, gc.Types[gc.Tptr])
var n1 gc.Node
igen(nl, &n1, nil)
n1.Type = gc.Types[gc.Tptr]
var tmp gc.Node
gc.Nodconst(&tmp, gc.Types[gc.Tptr], 0)
var n2 gc.Node
regalloc(&n2, gc.Types[gc.Tptr], &n1)
gmove(&n1, &n2)
gins(optoas(gc.OCMP, gc.Types[gc.Tptr]), &n2, &tmp)
regfree(&n2)
gc.Patch(gc.Gbranch(a, gc.Types[gc.Tptr], likely), to)
regfree(&n1)
break
}
if gc.Iscomplex[nl.Type.Etype] {
gc.Complexbool(a, nl, nr, true_, likely, to)
break
}
var n1 gc.Node
var n2 gc.Node
if nr.Ullman >= gc.UINF {
regalloc(&n1, nl.Type, nil)
cgen(nl, &n1)
var tmp gc.Node
gc.Tempname(&tmp, nl.Type)
gmove(&n1, &tmp)
regfree(&n1)
regalloc(&n2, nr.Type, nil)
cgen(nr, &n2)
regalloc(&n1, nl.Type, nil)
cgen(&tmp, &n1)
goto cmp
}
regalloc(&n1, nl.Type, nil)
cgen(nl, &n1)
// TODO(minux): cmpi does accept 16-bit signed immediate as p->to.
// and cmpli accepts 16-bit unsigned immediate.
//if(smallintconst(nr)) {
// gins(optoas(OCMP, nr->type), &n1, nr);
// patch(gbranch(optoas(a, nr->type), nr->type, likely), to);
// regfree(&n1);
// break;
//}
regalloc(&n2, nr.Type, nil)
cgen(nr, &n2)
cmp:
l := &n1
r := &n2
gins(optoas(gc.OCMP, nr.Type), l, r)
if gc.Isfloat[nr.Type.Etype] && (a == gc.OLE || a == gc.OGE) {
// To get NaN right, must rewrite x <= y into separate x < y or x = y.
switch a {
case gc.OLE:
a = gc.OLT
case gc.OGE:
a = gc.OGT
}
gc.Patch(gc.Gbranch(optoas(a, nr.Type), nr.Type, likely), to)
gc.Patch(gc.Gbranch(optoas(gc.OEQ, nr.Type), nr.Type, likely), to)
} else {
gc.Patch(gc.Gbranch(optoas(a, nr.Type), nr.Type, likely), to)
}
regfree(&n1)
regfree(&n2)
}
return
}
/*
* allocate a register (reusing res if possible) and generate
* a = &n
* The caller must call regfree(a).
* The generated code checks that the result is not nil.
*/
func agenr(n *gc.Node, a *gc.Node, res *gc.Node) {
if gc.Debug['g'] != 0 {
gc.Dump("\nagenr-n", n)
}
nl := n.Left
nr := n.Right
switch n.Op {
case gc.ODOT,
gc.ODOTPTR,
gc.OCALLFUNC,
gc.OCALLMETH,
gc.OCALLINTER:
var n1 gc.Node
igen(n, &n1, res)
regalloc(a, gc.Types[gc.Tptr], &n1)
agen(&n1, a)
regfree(&n1)
case gc.OIND:
cgenr(n.Left, a, res)
gc.Cgen_checknil(a)
case gc.OINDEX:
freelen := 0
w := uint64(n.Type.Width)
// Generate the non-addressable child first.
var n3 gc.Node
var nlen gc.Node
var tmp gc.Node
var n1 gc.Node
if nr.Addable != 0 {
goto irad
}
if nl.Addable != 0 {
cgenr(nr, &n1, nil)
if !gc.Isconst(nl, gc.CTSTR) {
if gc.Isfixedarray(nl.Type) {
agenr(nl, &n3, res)
} else {
igen(nl, &nlen, res)
freelen = 1
nlen.Type = gc.Types[gc.Tptr]
nlen.Xoffset += int64(gc.Array_array)
regalloc(&n3, gc.Types[gc.Tptr], res)
gmove(&nlen, &n3)
nlen.Type = gc.Types[gc.Simtype[gc.TUINT]]
nlen.Xoffset += int64(gc.Array_nel) - int64(gc.Array_array)
}
}
goto index
}
gc.Tempname(&tmp, nr.Type)
cgen(nr, &tmp)
nr = &tmp
irad:
if !gc.Isconst(nl, gc.CTSTR) {
if gc.Isfixedarray(nl.Type) {
agenr(nl, &n3, res)
} else {
if nl.Addable == 0 {
// igen will need an addressable node.
var tmp2 gc.Node
gc.Tempname(&tmp2, nl.Type)
cgen(nl, &tmp2)
nl = &tmp2
}
igen(nl, &nlen, res)
freelen = 1
nlen.Type = gc.Types[gc.Tptr]
nlen.Xoffset += int64(gc.Array_array)
regalloc(&n3, gc.Types[gc.Tptr], res)
gmove(&nlen, &n3)
nlen.Type = gc.Types[gc.Simtype[gc.TUINT]]
nlen.Xoffset += int64(gc.Array_nel) - int64(gc.Array_array)
}
}
if !gc.Isconst(nr, gc.CTINT) {
cgenr(nr, &n1, nil)
}
goto index
// &a is in &n3 (allocated in res)
// i is in &n1 (if not constant)
// len(a) is in nlen (if needed)
// w is width
// constant index
index:
if gc.Isconst(nr, gc.CTINT) {
if gc.Isconst(nl, gc.CTSTR) {
gc.Fatal("constant string constant index") // front end should handle
}
v := uint64(gc.Mpgetfix(nr.Val.U.Xval))
if gc.Isslice(nl.Type) || nl.Type.Etype == gc.TSTRING {
if gc.Debug['B'] == 0 && !n.Bounded {
var n2 gc.Node
gc.Nodconst(&n2, gc.Types[gc.Simtype[gc.TUINT]], int64(v))
if gc.Smallintconst(nr) {
gins(optoas(gc.OCMP, gc.Types[gc.Simtype[gc.TUINT]]), &nlen, &n2)
} else {
regalloc(&tmp, gc.Types[gc.Simtype[gc.TUINT]], nil)
gmove(&n2, &tmp)
gins(optoas(gc.OCMP, gc.Types[gc.Simtype[gc.TUINT]]), &nlen, &tmp)
regfree(&tmp)
}
p1 := gc.Gbranch(optoas(gc.OGT, gc.Types[gc.Simtype[gc.TUINT]]), nil, +1)
ginscall(gc.Panicindex, -1)
gc.Patch(p1, gc.Pc)
}
regfree(&nlen)
}
if v*w != 0 {
ginscon(optoas(gc.OADD, gc.Types[gc.Tptr]), int64(v*w), &n3)
}
*a = n3
break
}
// type of the index
t := gc.Types[gc.TUINT64]
if gc.Issigned[n1.Type.Etype] {
t = gc.Types[gc.TINT64]
}
var n2 gc.Node
regalloc(&n2, t, &n1) // i
gmove(&n1, &n2)
regfree(&n1)
if gc.Debug['B'] == 0 && !n.Bounded {
// check bounds
t = gc.Types[gc.Simtype[gc.TUINT]]
if gc.Is64(nr.Type) {
t = gc.Types[gc.TUINT64]
}
if gc.Isconst(nl, gc.CTSTR) {
gc.Nodconst(&nlen, t, int64(len(nl.Val.U.Sval)))
} else if gc.Isslice(nl.Type) || nl.Type.Etype == gc.TSTRING {
if gc.Is64(nr.Type) {
var n5 gc.Node
regalloc(&n5, t, nil)
gmove(&nlen, &n5)
regfree(&nlen)
nlen = n5
}
} else {
gc.Nodconst(&nlen, t, nl.Type.Bound)
if !gc.Smallintconst(&nlen) {
var n5 gc.Node
regalloc(&n5, t, nil)
gmove(&nlen, &n5)
nlen = n5
freelen = 1
}
}
gins(optoas(gc.OCMP, t), &n2, &nlen)
p1 := gc.Gbranch(optoas(gc.OLT, t), nil, +1)
ginscall(gc.Panicindex, -1)
gc.Patch(p1, gc.Pc)
}
if gc.Isconst(nl, gc.CTSTR) {
regalloc(&n3, gc.Types[gc.Tptr], res)
p1 := gins(x86.ALEAQ, nil, &n3)
gc.Datastring(nl.Val.U.Sval, &p1.From)
gins(x86.AADDQ, &n2, &n3)
goto indexdone
}
if w == 0 {
} else // nothing to do
if w == 1 || w == 2 || w == 4 || w == 8 {
p1 := gins(x86.ALEAQ, &n2, &n3)
p1.From.Type = obj.TYPE_MEM
p1.From.Scale = int16(w)
p1.From.Index = p1.From.Reg
p1.From.Reg = p1.To.Reg
} else {
ginscon(optoas(gc.OMUL, t), int64(w), &n2)
gins(optoas(gc.OADD, gc.Types[gc.Tptr]), &n2, &n3)
}
indexdone:
*a = n3
regfree(&n2)
if freelen != 0 {
regfree(&nlen)
}
default:
regalloc(a, gc.Types[gc.Tptr], res)
agen(n, a)
}
}
func anyregalloc() bool {
var j int
for i := x86.REG_AX; i <= x86.REG_DI; i++ {
if reg[i] == 0 {
goto ok
}
for j = 0; j < len(resvd); j++ {
if resvd[j] == i {
goto ok
}
}
return true
ok:
}
for i := x86.REG_X0; i <= x86.REG_X7; i++ {
if reg[i] != 0 {
return true
}
}
return false
}
/*
* allocate register of type t, leave in n.
* if o != N, o is desired fixed register.
* caller must regfree(n).
*/
func regalloc(n *gc.Node, t *gc.Type, o *gc.Node) {
if t == nil {
gc.Fatal("regalloc: t nil")
}
et := int(gc.Simtype[t.Etype])
var i int
switch et {
case gc.TINT64,
gc.TUINT64:
gc.Fatal("regalloc64")
case gc.TINT8,
gc.TUINT8,
gc.TINT16,
gc.TUINT16,
gc.TINT32,
gc.TUINT32,
gc.TPTR32,
gc.TPTR64,
gc.TBOOL:
if o != nil && o.Op == gc.OREGISTER {
i = int(o.Val.U.Reg)
if i >= x86.REG_AX && i <= x86.REG_DI {
goto out
}
}
for i = x86.REG_AX; i <= x86.REG_DI; i++ {
if reg[i] == 0 {
goto out
}
}
fmt.Printf("registers allocated at\n")
for i := x86.REG_AX; i <= x86.REG_DI; i++ {
fmt.Printf("\t%v\t%#x\n", obj.Rconv(i), regpc[i])
}
gc.Fatal("out of fixed registers")
goto err
case gc.TFLOAT32,
gc.TFLOAT64:
if gc.Use_sse == 0 {
i = x86.REG_F0
goto out
}
if o != nil && o.Op == gc.OREGISTER {
i = int(o.Val.U.Reg)
if i >= x86.REG_X0 && i <= x86.REG_X7 {
goto out
}
}
for i = x86.REG_X0; i <= x86.REG_X7; i++ {
if reg[i] == 0 {
goto out
}
}
fmt.Printf("registers allocated at\n")
for i := x86.REG_X0; i <= x86.REG_X7; i++ {
fmt.Printf("\t%v\t%#x\n", obj.Rconv(i), regpc[i])
}
gc.Fatal("out of floating registers")
}
gc.Yyerror("regalloc: unknown type %v", gc.Tconv(t, 0))
err:
gc.Nodreg(n, t, 0)
return
out:
if i == x86.REG_SP {
fmt.Printf("alloc SP\n")
}
if reg[i] == 0 {
regpc[i] = uint32(obj.Getcallerpc(&n))
if i == x86.REG_AX || i == x86.REG_CX || i == x86.REG_DX || i == x86.REG_SP {
gc.Dump("regalloc-o", o)
gc.Fatal("regalloc %v", obj.Rconv(i))
}
}
reg[i]++
gc.Nodreg(n, t, i)
}
func regfree(n *gc.Node) {
if n.Op == gc.ONAME {
return
}
if n.Op != gc.OREGISTER && n.Op != gc.OINDREG {
gc.Fatal("regfree: not a register")
}
i := int(n.Val.U.Reg)
if i == x86.REG_SP {
return
}
if i < 0 || i >= len(reg) {
gc.Fatal("regfree: reg out of range")
}
if reg[i] <= 0 {
gc.Fatal("regfree: reg not allocated")
}
reg[i]--
if reg[i] == 0 && (i == x86.REG_AX || i == x86.REG_CX || i == x86.REG_DX || i == x86.REG_SP) {
gc.Fatal("regfree %v", obj.Rconv(i))
}
}
/*
* generate
* as $c, reg
*/
func gconreg(as int, c int64, reg int) {
var n1 gc.Node
var n2 gc.Node
gc.Nodconst(&n1, gc.Types[gc.TINT64], c)
gc.Nodreg(&n2, gc.Types[gc.TINT64], reg)
gins(as, &n1, &n2)
}
/*
* swap node contents
*/
func nswap(a *gc.Node, b *gc.Node) {
t := *a
*a = *b
*b = t
}
/*
* return constant i node.
* overwritten by next call, but useful in calls to gins.
*/
var ncon_n gc.Node
func ncon(i uint32) *gc.Node {
if ncon_n.Type == nil {
gc.Nodconst(&ncon_n, gc.Types[gc.TUINT32], 0)
}
gc.Mpmovecfix(ncon_n.Val.U.Xval, int64(i))
return &ncon_n
}
var sclean [10]gc.Node
var nsclean int
/*
* n is a 64-bit value. fill in lo and hi to refer to its 32-bit halves.
*/
func split64(n *gc.Node, lo *gc.Node, hi *gc.Node) {
if !gc.Is64(n.Type) {
gc.Fatal("split64 %v", gc.Tconv(n.Type, 0))
}
if nsclean >= len(sclean) {
gc.Fatal("split64 clean")
}
sclean[nsclean].Op = gc.OEMPTY
nsclean++
switch n.Op {
default:
switch n.Op {
default:
var n1 gc.Node
if !dotaddable(n, &n1) {
igen(n, &n1, nil)
sclean[nsclean-1] = n1
}
n = &n1
case gc.ONAME:
if n.Class == gc.PPARAMREF {
var n1 gc.Node
cgen(n.Heapaddr, &n1)
sclean[nsclean-1] = n1
n = &n1
}
// nothing
case gc.OINDREG:
break
}
*lo = *n
*hi = *n
lo.Type = gc.Types[gc.TUINT32]
if n.Type.Etype == gc.TINT64 {
hi.Type = gc.Types[gc.TINT32]
} else {
hi.Type = gc.Types[gc.TUINT32]
}
hi.Xoffset += 4
case gc.OLITERAL:
var n1 gc.Node
gc.Convconst(&n1, n.Type, &n.Val)
i := gc.Mpgetfix(n1.Val.U.Xval)
gc.Nodconst(lo, gc.Types[gc.TUINT32], int64(uint32(i)))
i >>= 32
if n.Type.Etype == gc.TINT64 {
gc.Nodconst(hi, gc.Types[gc.TINT32], int64(int32(i)))
} else {
gc.Nodconst(hi, gc.Types[gc.TUINT32], int64(uint32(i)))
}
}
}
func splitclean() {
if nsclean <= 0 {
gc.Fatal("splitclean")
}
nsclean--
if sclean[nsclean].Op != gc.OEMPTY {
regfree(&sclean[nsclean])
}
}
/*
* set up nodes representing fp constants
*/
var zerof gc.Node
var two64f gc.Node
var two63f gc.Node
var bignodes_did int
func bignodes() {
if bignodes_did != 0 {
return
}
bignodes_did = 1
two64f = *ncon(0)
two64f.Type = gc.Types[gc.TFLOAT64]
two64f.Val.Ctype = gc.CTFLT
two64f.Val.U.Fval = new(gc.Mpflt)
gc.Mpmovecflt(two64f.Val.U.Fval, 18446744073709551616.)
two63f = two64f
two63f.Val.U.Fval = new(gc.Mpflt)
gc.Mpmovecflt(two63f.Val.U.Fval, 9223372036854775808.)
zerof = two64f
zerof.Val.U.Fval = new(gc.Mpflt)
gc.Mpmovecflt(zerof.Val.U.Fval, 0)
}
func memname(n *gc.Node, t *gc.Type) {
gc.Tempname(n, t)
n.Sym = gc.Lookup("." + n.Sym.Name[1:]) // keep optimizer from registerizing
n.Orig.Sym = n.Sym
}
func gmove(f *gc.Node, t *gc.Node) {
if gc.Debug['M'] != 0 {
fmt.Printf("gmove %v -> %v\n", gc.Nconv(f, 0), gc.Nconv(t, 0))
}
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
cvt := t.Type
if gc.Iscomplex[ft] || gc.Iscomplex[tt] {
gc.Complexmove(f, t)
return
}
if gc.Isfloat[ft] || gc.Isfloat[tt] {
floatmove(f, t)
return
}
// cannot have two integer memory operands;
// except 64-bit, which always copies via registers anyway.
var r1 gc.Node
var a int
if gc.Isint[ft] && gc.Isint[tt] && !gc.Is64(f.Type) && !gc.Is64(t.Type) && gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
gc.Convconst(&con, t.Type, &f.Val)
f = &con
ft = gc.Simsimtype(con.Type)
}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
// should not happen
gc.Fatal("gmove %v -> %v", gc.Nconv(f, 0), gc.Nconv(t, 0))
return
/*
* integer copy and truncate
*/
case gc.TINT8<<16 | gc.TINT8, // same size
gc.TINT8<<16 | gc.TUINT8,
gc.TUINT8<<16 | gc.TINT8,
gc.TUINT8<<16 | gc.TUINT8:
a = x86.AMOVB
case gc.TINT16<<16 | gc.TINT8, // truncate
gc.TUINT16<<16 | gc.TINT8,
gc.TINT32<<16 | gc.TINT8,
gc.TUINT32<<16 | gc.TINT8,
gc.TINT16<<16 | gc.TUINT8,
gc.TUINT16<<16 | gc.TUINT8,
gc.TINT32<<16 | gc.TUINT8,
gc.TUINT32<<16 | gc.TUINT8:
a = x86.AMOVB
goto rsrc
case gc.TINT64<<16 | gc.TINT8, // truncate low word
gc.TUINT64<<16 | gc.TINT8,
gc.TINT64<<16 | gc.TUINT8,
gc.TUINT64<<16 | gc.TUINT8:
var flo gc.Node
var fhi gc.Node
split64(f, &flo, &fhi)
var r1 gc.Node
gc.Nodreg(&r1, t.Type, x86.REG_AX)
gmove(&flo, &r1)
gins(x86.AMOVB, &r1, t)
splitclean()
return
case gc.TINT16<<16 | gc.TINT16, // same size
gc.TINT16<<16 | gc.TUINT16,
gc.TUINT16<<16 | gc.TINT16,
gc.TUINT16<<16 | gc.TUINT16:
a = x86.AMOVW
case gc.TINT32<<16 | gc.TINT16, // truncate
gc.TUINT32<<16 | gc.TINT16,
gc.TINT32<<16 | gc.TUINT16,
gc.TUINT32<<16 | gc.TUINT16:
a = x86.AMOVW
goto rsrc
case gc.TINT64<<16 | gc.TINT16, // truncate low word
gc.TUINT64<<16 | gc.TINT16,
gc.TINT64<<16 | gc.TUINT16,
gc.TUINT64<<16 | gc.TUINT16:
var flo gc.Node
var fhi gc.Node
split64(f, &flo, &fhi)
var r1 gc.Node
gc.Nodreg(&r1, t.Type, x86.REG_AX)
gmove(&flo, &r1)
gins(x86.AMOVW, &r1, t)
splitclean()
return
case gc.TINT32<<16 | gc.TINT32, // same size
gc.TINT32<<16 | gc.TUINT32,
gc.TUINT32<<16 | gc.TINT32,
gc.TUINT32<<16 | gc.TUINT32:
a = x86.AMOVL
case gc.TINT64<<16 | gc.TINT32, // truncate
gc.TUINT64<<16 | gc.TINT32,
gc.TINT64<<16 | gc.TUINT32,
gc.TUINT64<<16 | gc.TUINT32:
var fhi gc.Node
var flo gc.Node
split64(f, &flo, &fhi)
var r1 gc.Node
gc.Nodreg(&r1, t.Type, x86.REG_AX)
gmove(&flo, &r1)
gins(x86.AMOVL, &r1, t)
splitclean()
return
case gc.TINT64<<16 | gc.TINT64, // same size
gc.TINT64<<16 | gc.TUINT64,
gc.TUINT64<<16 | gc.TINT64,
gc.TUINT64<<16 | gc.TUINT64:
var fhi gc.Node
var flo gc.Node
split64(f, &flo, &fhi)
var tlo gc.Node
var thi gc.Node
split64(t, &tlo, &thi)
if f.Op == gc.OLITERAL {
gins(x86.AMOVL, &flo, &tlo)
gins(x86.AMOVL, &fhi, &thi)
} else {
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[gc.TUINT32], x86.REG_AX)
var r2 gc.Node
gc.Nodreg(&r2, gc.Types[gc.TUINT32], x86.REG_DX)
gins(x86.AMOVL, &flo, &r1)
gins(x86.AMOVL, &fhi, &r2)
gins(x86.AMOVL, &r1, &tlo)
gins(x86.AMOVL, &r2, &thi)
}
splitclean()
splitclean()
return
/*
* integer up-conversions
*/
case gc.TINT8<<16 | gc.TINT16, // sign extend int8
gc.TINT8<<16 | gc.TUINT16:
a = x86.AMOVBWSX
goto rdst
case gc.TINT8<<16 | gc.TINT32,
gc.TINT8<<16 | gc.TUINT32:
a = x86.AMOVBLSX
goto rdst
case gc.TINT8<<16 | gc.TINT64, // convert via int32
gc.TINT8<<16 | gc.TUINT64:
cvt = gc.Types[gc.TINT32]
goto hard
case gc.TUINT8<<16 | gc.TINT16, // zero extend uint8
gc.TUINT8<<16 | gc.TUINT16:
a = x86.AMOVBWZX
goto rdst
case gc.TUINT8<<16 | gc.TINT32,
gc.TUINT8<<16 | gc.TUINT32:
a = x86.AMOVBLZX
goto rdst
case gc.TUINT8<<16 | gc.TINT64, // convert via uint32
gc.TUINT8<<16 | gc.TUINT64:
cvt = gc.Types[gc.TUINT32]
goto hard
case gc.TINT16<<16 | gc.TINT32, // sign extend int16
gc.TINT16<<16 | gc.TUINT32:
a = x86.AMOVWLSX
goto rdst
case gc.TINT16<<16 | gc.TINT64, // convert via int32
gc.TINT16<<16 | gc.TUINT64:
cvt = gc.Types[gc.TINT32]
goto hard
case gc.TUINT16<<16 | gc.TINT32, // zero extend uint16
gc.TUINT16<<16 | gc.TUINT32:
a = x86.AMOVWLZX
goto rdst
case gc.TUINT16<<16 | gc.TINT64, // convert via uint32
gc.TUINT16<<16 | gc.TUINT64:
cvt = gc.Types[gc.TUINT32]
goto hard
case gc.TINT32<<16 | gc.TINT64, // sign extend int32
gc.TINT32<<16 | gc.TUINT64:
var thi gc.Node
var tlo gc.Node
split64(t, &tlo, &thi)
var flo gc.Node
gc.Nodreg(&flo, tlo.Type, x86.REG_AX)
var fhi gc.Node
gc.Nodreg(&fhi, thi.Type, x86.REG_DX)
gmove(f, &flo)
gins(x86.ACDQ, nil, nil)
gins(x86.AMOVL, &flo, &tlo)
gins(x86.AMOVL, &fhi, &thi)
splitclean()
return
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
var tlo gc.Node
var thi gc.Node
split64(t, &tlo, &thi)
gmove(f, &tlo)
gins(x86.AMOVL, ncon(0), &thi)
splitclean()
return
}
gins(a, f, t)
return
// requires register source
rsrc:
regalloc(&r1, f.Type, t)
gmove(f, &r1)
gins(a, &r1, t)
regfree(&r1)
return
// requires register destination
rdst:
{
regalloc(&r1, t.Type, t)
gins(a, f, &r1)
gmove(&r1, t)
regfree(&r1)
return
}
// requires register intermediate
hard:
regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
regfree(&r1)
return
}
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
gc.Convconst(&con, t.Type, &f.Val)
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.Use_sse != 0 {
floatmove_sse(f, t)
} else {
floatmove_387(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.AFUCOMIP, &f0, &f1)
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.AFUCOMIP, &f0, &f1)
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.AFUCOMIP, &f0, &f1)
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:
regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
regfree(&r1)
return
// requires memory intermediate
hardmem:
gc.Tempname(&r1, cvt)
gmove(f, &r1)
gmove(&r1, t)
return
}
func floatmove_387(f *gc.Node, t *gc.Node) {
var r1 gc.Node
var a int
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.Fatal("gmove %v", gc.Nconv(t, 0))
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.Val.U.Reg != x86.REG_F0 || t.Val.U.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.Val.U.Reg != x86.REG_F0 {
gc.Fatal("gmove %v", gc.Nconv(f, 0))
}
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.Val.U.Reg != x86.REG_F0 || t.Val.U.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:
regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
regfree(&r1)
return
// requires memory intermediate
hardmem:
gc.Tempname(&r1, cvt)
gmove(f, &r1)
gmove(&r1, t)
return
// should not happen
fatal:
gc.Fatal("gmove %v -> %v", gc.Nconv(f, obj.FmtLong), gc.Nconv(t, obj.FmtLong))
return
}
func floatmove_sse(f *gc.Node, t *gc.Node) {
var r1 gc.Node
var cvt *gc.Type
var a int
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
switch uint32(ft)<<16 | uint32(tt) {
// should not happen
default:
gc.Fatal("gmove %v -> %v", gc.Nconv(f, 0), gc.Nconv(t, 0))
return
// convert via int32.
/*
* float to integer
*/
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 hardmem
case gc.TFLOAT32<<16 | gc.TINT32:
a = x86.ACVTTSS2SL
goto rdst
case gc.TFLOAT64<<16 | gc.TINT32:
a = x86.ACVTTSD2SL
goto rdst
// convert via int32 memory
/*
* integer to float
*/
case gc.TINT8<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TINT16<<16 | gc.TFLOAT32,
gc.TINT16<<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 memory
case gc.TUINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT64]
goto hardmem
case gc.TINT32<<16 | gc.TFLOAT32:
a = x86.ACVTSL2SS
goto rdst
case gc.TINT32<<16 | gc.TFLOAT64:
a = x86.ACVTSL2SD
goto rdst
/*
* 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 intermediate
hard:
regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
regfree(&r1)
return
// requires memory intermediate
hardmem:
gc.Tempname(&r1, cvt)
gmove(f, &r1)
gmove(&r1, t)
return
// requires register destination
rdst:
regalloc(&r1, t.Type, t)
gins(a, f, &r1)
gmove(&r1, t)
regfree(&r1)
return
}
func samaddr(f *gc.Node, t *gc.Node) bool {
if f.Op != t.Op {
return false
}
switch f.Op {
case gc.OREGISTER:
if f.Val.U.Reg != t.Val.U.Reg {
break
}
return true
}
return false
}
/*
* generate one instruction:
* as f, t
*/
func gins(as int, f *gc.Node, t *gc.Node) *obj.Prog {
if as == x86.AFMOVF && f != nil && f.Op == gc.OREGISTER && t != nil && t.Op == gc.OREGISTER {
gc.Fatal("gins MOVF reg, reg")
}
if as == x86.ACVTSD2SS && f != nil && f.Op == gc.OLITERAL {
gc.Fatal("gins CVTSD2SS const")
}
if as == x86.AMOVSD && t != nil && t.Op == gc.OREGISTER && t.Val.U.Reg == x86.REG_F0 {
gc.Fatal("gins MOVSD into F0")
}
switch as {
case x86.AMOVB,
x86.AMOVW,
x86.AMOVL:
if f != nil && t != nil && samaddr(f, t) {
return nil
}
case x86.ALEAL:
if f != nil && gc.Isconst(f, gc.CTNIL) {
gc.Fatal("gins LEAL nil %v", gc.Tconv(f.Type, 0))
}
}
var af obj.Addr
var at obj.Addr
if f != nil {
af = gc.Naddr(f)
}
if t != nil {
at = gc.Naddr(t)
}
p := gc.Prog(as)
if f != nil {
p.From = af
}
if t != nil {
p.To = at
}
if gc.Debug['g'] != 0 {
fmt.Printf("%v\n", p)
}
w := 0
switch as {
case x86.AMOVB:
w = 1
case x86.AMOVW:
w = 2
case x86.AMOVL:
w = 4
}
if true && w != 0 && f != nil && (af.Width > int64(w) || at.Width > int64(w)) {
gc.Dump("bad width from:", f)
gc.Dump("bad width to:", t)
gc.Fatal("bad width: %v (%d, %d)\n", p, af.Width, at.Width)
}
if p.To.Type == obj.TYPE_ADDR && w > 0 {
gc.Fatal("bad use of addr: %v", p)
}
return p
}
func dotaddable(n *gc.Node, n1 *gc.Node) bool {
if n.Op != gc.ODOT {
return false
}
var oary [10]int64
var nn *gc.Node
o := gc.Dotoffset(n, oary[:], &nn)
if nn != nil && nn.Addable != 0 && o == 1 && oary[0] >= 0 {
*n1 = *nn
n1.Type = n.Type
n1.Xoffset += oary[0]
return true
}
return false
}
func sudoclean() {
}
func sudoaddable(as int, n *gc.Node, a *obj.Addr) bool {
*a = obj.Addr{}
return false
}
func anyregalloc() bool {
var j int
for i := int(0); i < len(reg); i++ {
if reg[i] == 0 {
goto ok
}
for j = 0; j < len(resvd); j++ {
if resvd[j] == i {
goto ok
}
}
return true
ok:
}
return false
}
/*
* allocate register of type t, leave in n.
* if o != N, o is desired fixed register.
* caller must regfree(n).
*/
func regalloc(n *gc.Node, t *gc.Type, o *gc.Node) {
if t == nil {
gc.Fatal("regalloc: t nil")
}
et := int(int(gc.Simtype[t.Etype]))
if gc.Debug['r'] != 0 {
fixfree := int(0)
fltfree := int(0)
for i := int(ppc64.REG_R0); i < ppc64.REG_F31; i++ {
if reg[i-ppc64.REG_R0] == 0 {
if i < ppc64.REG_F0 {
fixfree++
} else {
fltfree++
}
}
}
fmt.Printf("regalloc fix %d flt %d free\n", fixfree, fltfree)
}
var i int
switch et {
case gc.TINT8,
gc.TUINT8,
gc.TINT16,
gc.TUINT16,
gc.TINT32,
gc.TUINT32,
gc.TINT64,
gc.TUINT64,
gc.TPTR32,
gc.TPTR64,
gc.TBOOL:
if o != nil && o.Op == gc.OREGISTER {
i = int(o.Val.U.Reg)
if i >= ppc64.REGMIN && i <= ppc64.REGMAX {
goto out
}
}
for i = ppc64.REGMIN; i <= ppc64.REGMAX; i++ {
if reg[i-ppc64.REG_R0] == 0 {
regpc[i-ppc64.REG_R0] = uint32(obj.Getcallerpc(&n))
goto out
}
}
gc.Flusherrors()
for i := int(ppc64.REG_R0); i < ppc64.REG_R0+ppc64.NREG; i++ {
fmt.Printf("R%d %p\n", i, regpc[i-ppc64.REG_R0])
}
gc.Fatal("out of fixed registers")
case gc.TFLOAT32,
gc.TFLOAT64:
if o != nil && o.Op == gc.OREGISTER {
i = int(o.Val.U.Reg)
if i >= ppc64.FREGMIN && i <= ppc64.FREGMAX {
goto out
}
}
for i = ppc64.FREGMIN; i <= ppc64.FREGMAX; i++ {
if reg[i-ppc64.REG_R0] == 0 {
regpc[i-ppc64.REG_R0] = uint32(obj.Getcallerpc(&n))
goto out
}
}
gc.Flusherrors()
for i := int(ppc64.REG_F0); i < ppc64.REG_F0+ppc64.NREG; i++ {
fmt.Printf("F%d %p\n", i, regpc[i-ppc64.REG_R0])
}
gc.Fatal("out of floating registers")
case gc.TCOMPLEX64,
gc.TCOMPLEX128:
gc.Tempname(n, t)
return
}
gc.Fatal("regalloc: unknown type %v", gc.Tconv(t, 0))
return
out:
reg[i-ppc64.REG_R0]++
gc.Nodreg(n, t, i)
}
func regfree(n *gc.Node) {
if n.Op == gc.ONAME {
return
}
if n.Op != gc.OREGISTER && n.Op != gc.OINDREG {
gc.Fatal("regfree: not a register")
}
i := int(int(n.Val.U.Reg) - ppc64.REG_R0)
if i == ppc64.REGSP-ppc64.REG_R0 {
return
}
if i < 0 || i >= len(reg) {
gc.Fatal("regfree: reg out of range")
}
if reg[i] <= 0 {
gc.Fatal("regfree: reg not allocated")
}
reg[i]--
if reg[i] == 0 {
regpc[i] = 0
}
}
/*
* generate
* as $c, n
*/
func ginscon(as int, c int64, n2 *gc.Node) {
var n1 gc.Node
gc.Nodconst(&n1, gc.Types[gc.TINT64], c)
if as != ppc64.AMOVD && (c < -ppc64.BIG || c > ppc64.BIG) {
// cannot have more than 16-bit of immediate in ADD, etc.
// instead, MOV into register first.
var ntmp gc.Node
regalloc(&ntmp, gc.Types[gc.TINT64], nil)
gins(ppc64.AMOVD, &n1, &ntmp)
gins(as, &ntmp, n2)
regfree(&ntmp)
return
}
gins(as, &n1, n2)
}
/*
* generate
* as n, $c (CMP/CMPU)
*/
func ginscon2(as int, n2 *gc.Node, c int64) {
var n1 gc.Node
gc.Nodconst(&n1, gc.Types[gc.TINT64], c)
switch as {
default:
gc.Fatal("ginscon2")
case ppc64.ACMP:
if -ppc64.BIG <= c && c <= ppc64.BIG {
gins(as, n2, &n1)
return
}
case ppc64.ACMPU:
if 0 <= c && c <= 2*ppc64.BIG {
gins(as, n2, &n1)
return
}
}
// MOV n1 into register first
var ntmp gc.Node
regalloc(&ntmp, gc.Types[gc.TINT64], nil)
gins(ppc64.AMOVD, &n1, &ntmp)
gins(as, n2, &ntmp)
regfree(&ntmp)
}
/*
* set up nodes representing 2^63
*/
var bigi gc.Node
var bigf gc.Node
var bignodes_did int
func bignodes() {
if bignodes_did != 0 {
return
}
bignodes_did = 1
gc.Nodconst(&bigi, gc.Types[gc.TUINT64], 1)
gc.Mpshiftfix(bigi.Val.U.Xval, 63)
bigf = bigi
bigf.Type = gc.Types[gc.TFLOAT64]
bigf.Val.Ctype = gc.CTFLT
bigf.Val.U.Fval = new(gc.Mpflt)
gc.Mpmovefixflt(bigf.Val.U.Fval, bigi.Val.U.Xval)
}
/*
* generate move:
* t = f
* hard part is conversions.
*/
func gmove(f *gc.Node, t *gc.Node) {
if gc.Debug['M'] != 0 {
fmt.Printf("gmove %v -> %v\n", gc.Nconv(f, obj.FmtLong), gc.Nconv(t, obj.FmtLong))
}
ft := int(gc.Simsimtype(f.Type))
tt := int(gc.Simsimtype(t.Type))
cvt := (*gc.Type)(t.Type)
if gc.Iscomplex[ft] || gc.Iscomplex[tt] {
gc.Complexmove(f, t)
return
}
// cannot have two memory operands
var r2 gc.Node
var r1 gc.Node
var a int
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
switch tt {
default:
gc.Convconst(&con, t.Type, &f.Val)
case gc.TINT32,
gc.TINT16,
gc.TINT8:
var con gc.Node
gc.Convconst(&con, gc.Types[gc.TINT64], &f.Val)
var r1 gc.Node
regalloc(&r1, con.Type, t)
gins(ppc64.AMOVD, &con, &r1)
gmove(&r1, t)
regfree(&r1)
return
case gc.TUINT32,
gc.TUINT16,
gc.TUINT8:
var con gc.Node
gc.Convconst(&con, gc.Types[gc.TUINT64], &f.Val)
var r1 gc.Node
regalloc(&r1, con.Type, t)
gins(ppc64.AMOVD, &con, &r1)
gmove(&r1, t)
regfree(&r1)
return
}
f = &con
ft = tt // so big switch will choose a simple mov
// constants can't move directly to memory.
if gc.Ismem(t) {
goto hard
}
}
// float constants come from memory.
//if(isfloat[tt])
// goto hard;
// 64-bit immediates are also from memory.
//if(isint[tt])
// goto hard;
//// 64-bit immediates are really 32-bit sign-extended
//// unless moving into a register.
//if(isint[tt]) {
// if(mpcmpfixfix(con.val.u.xval, minintval[TINT32]) < 0)
// goto hard;
// if(mpcmpfixfix(con.val.u.xval, maxintval[TINT32]) > 0)
// goto hard;
//}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
gc.Fatal("gmove %v -> %v", gc.Tconv(f.Type, obj.FmtLong), gc.Tconv(t.Type, obj.FmtLong))
/*
* integer copy and truncate
*/
case gc.TINT8<<16 | gc.TINT8, // same size
gc.TUINT8<<16 | gc.TINT8,
gc.TINT16<<16 | gc.TINT8,
// truncate
gc.TUINT16<<16 | gc.TINT8,
gc.TINT32<<16 | gc.TINT8,
gc.TUINT32<<16 | gc.TINT8,
gc.TINT64<<16 | gc.TINT8,
gc.TUINT64<<16 | gc.TINT8:
a = ppc64.AMOVB
case gc.TINT8<<16 | gc.TUINT8, // same size
gc.TUINT8<<16 | gc.TUINT8,
gc.TINT16<<16 | gc.TUINT8,
// truncate
gc.TUINT16<<16 | gc.TUINT8,
gc.TINT32<<16 | gc.TUINT8,
gc.TUINT32<<16 | gc.TUINT8,
gc.TINT64<<16 | gc.TUINT8,
gc.TUINT64<<16 | gc.TUINT8:
a = ppc64.AMOVBZ
case gc.TINT16<<16 | gc.TINT16, // same size
gc.TUINT16<<16 | gc.TINT16,
gc.TINT32<<16 | gc.TINT16,
// truncate
gc.TUINT32<<16 | gc.TINT16,
gc.TINT64<<16 | gc.TINT16,
gc.TUINT64<<16 | gc.TINT16:
a = ppc64.AMOVH
case gc.TINT16<<16 | gc.TUINT16, // same size
gc.TUINT16<<16 | gc.TUINT16,
gc.TINT32<<16 | gc.TUINT16,
// truncate
gc.TUINT32<<16 | gc.TUINT16,
gc.TINT64<<16 | gc.TUINT16,
gc.TUINT64<<16 | gc.TUINT16:
a = ppc64.AMOVHZ
case gc.TINT32<<16 | gc.TINT32, // same size
gc.TUINT32<<16 | gc.TINT32,
gc.TINT64<<16 | gc.TINT32,
// truncate
gc.TUINT64<<16 | gc.TINT32:
a = ppc64.AMOVW
case gc.TINT32<<16 | gc.TUINT32, // same size
gc.TUINT32<<16 | gc.TUINT32,
gc.TINT64<<16 | gc.TUINT32,
gc.TUINT64<<16 | gc.TUINT32:
a = ppc64.AMOVWZ
case gc.TINT64<<16 | gc.TINT64, // same size
gc.TINT64<<16 | gc.TUINT64,
gc.TUINT64<<16 | gc.TINT64,
gc.TUINT64<<16 | gc.TUINT64:
a = ppc64.AMOVD
/*
* integer up-conversions
*/
case gc.TINT8<<16 | gc.TINT16, // sign extend int8
gc.TINT8<<16 | gc.TUINT16,
gc.TINT8<<16 | gc.TINT32,
gc.TINT8<<16 | gc.TUINT32,
gc.TINT8<<16 | gc.TINT64,
gc.TINT8<<16 | gc.TUINT64:
a = ppc64.AMOVB
goto rdst
case gc.TUINT8<<16 | gc.TINT16, // zero extend uint8
gc.TUINT8<<16 | gc.TUINT16,
gc.TUINT8<<16 | gc.TINT32,
gc.TUINT8<<16 | gc.TUINT32,
gc.TUINT8<<16 | gc.TINT64,
gc.TUINT8<<16 | gc.TUINT64:
a = ppc64.AMOVBZ
goto rdst
case gc.TINT16<<16 | gc.TINT32, // sign extend int16
gc.TINT16<<16 | gc.TUINT32,
gc.TINT16<<16 | gc.TINT64,
gc.TINT16<<16 | gc.TUINT64:
a = ppc64.AMOVH
goto rdst
case gc.TUINT16<<16 | gc.TINT32, // zero extend uint16
gc.TUINT16<<16 | gc.TUINT32,
gc.TUINT16<<16 | gc.TINT64,
gc.TUINT16<<16 | gc.TUINT64:
a = ppc64.AMOVHZ
goto rdst
case gc.TINT32<<16 | gc.TINT64, // sign extend int32
gc.TINT32<<16 | gc.TUINT64:
a = ppc64.AMOVW
goto rdst
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
a = ppc64.AMOVWZ
goto rdst
//warn("gmove: convert float to int not implemented: %N -> %N\n", f, t);
//return;
// algorithm is:
// if small enough, use native float64 -> int64 conversion.
// otherwise, subtract 2^63, convert, and add it back.
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT32,
gc.TFLOAT64<<16 | gc.TINT32,
gc.TFLOAT32<<16 | gc.TINT64,
gc.TFLOAT64<<16 | gc.TINT64,
gc.TFLOAT32<<16 | gc.TINT16,
gc.TFLOAT32<<16 | gc.TINT8,
gc.TFLOAT32<<16 | gc.TUINT16,
gc.TFLOAT32<<16 | gc.TUINT8,
gc.TFLOAT64<<16 | gc.TINT16,
gc.TFLOAT64<<16 | gc.TINT8,
gc.TFLOAT64<<16 | gc.TUINT16,
gc.TFLOAT64<<16 | gc.TUINT8,
gc.TFLOAT32<<16 | gc.TUINT32,
gc.TFLOAT64<<16 | gc.TUINT32,
gc.TFLOAT32<<16 | gc.TUINT64,
gc.TFLOAT64<<16 | gc.TUINT64:
bignodes()
var r1 gc.Node
regalloc(&r1, gc.Types[ft], f)
gmove(f, &r1)
if tt == gc.TUINT64 {
regalloc(&r2, gc.Types[gc.TFLOAT64], nil)
gmove(&bigf, &r2)
gins(ppc64.AFCMPU, &r1, &r2)
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TFLOAT64]), nil, +1))
gins(ppc64.AFSUB, &r2, &r1)
gc.Patch(p1, gc.Pc)
regfree(&r2)
}
regalloc(&r2, gc.Types[gc.TFLOAT64], nil)
var r3 gc.Node
regalloc(&r3, gc.Types[gc.TINT64], t)
gins(ppc64.AFCTIDZ, &r1, &r2)
p1 := (*obj.Prog)(gins(ppc64.AFMOVD, &r2, nil))
p1.To.Type = obj.TYPE_MEM
p1.To.Reg = ppc64.REGSP
p1.To.Offset = -8
p1 = gins(ppc64.AMOVD, nil, &r3)
p1.From.Type = obj.TYPE_MEM
p1.From.Reg = ppc64.REGSP
p1.From.Offset = -8
regfree(&r2)
regfree(&r1)
if tt == gc.TUINT64 {
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TFLOAT64]), nil, +1)) // use CR0 here again
gc.Nodreg(&r1, gc.Types[gc.TINT64], ppc64.REGTMP)
gins(ppc64.AMOVD, &bigi, &r1)
gins(ppc64.AADD, &r1, &r3)
gc.Patch(p1, gc.Pc)
}
gmove(&r3, t)
regfree(&r3)
return
//warn("gmove: convert int to float not implemented: %N -> %N\n", f, t);
//return;
// algorithm is:
// if small enough, use native int64 -> uint64 conversion.
// otherwise, halve (rounding to odd?), convert, and double.
/*
* integer to float
*/
case gc.TINT32<<16 | gc.TFLOAT32,
gc.TINT32<<16 | gc.TFLOAT64,
gc.TINT64<<16 | gc.TFLOAT32,
gc.TINT64<<16 | gc.TFLOAT64,
gc.TINT16<<16 | gc.TFLOAT32,
gc.TINT16<<16 | gc.TFLOAT64,
gc.TINT8<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TUINT16<<16 | gc.TFLOAT32,
gc.TUINT16<<16 | gc.TFLOAT64,
gc.TUINT8<<16 | gc.TFLOAT32,
gc.TUINT8<<16 | gc.TFLOAT64,
gc.TUINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT64,
gc.TUINT64<<16 | gc.TFLOAT32,
gc.TUINT64<<16 | gc.TFLOAT64:
bignodes()
var r1 gc.Node
regalloc(&r1, gc.Types[gc.TINT64], nil)
gmove(f, &r1)
if ft == gc.TUINT64 {
gc.Nodreg(&r2, gc.Types[gc.TUINT64], ppc64.REGTMP)
gmove(&bigi, &r2)
gins(ppc64.ACMPU, &r1, &r2)
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT64]), nil, +1))
p2 := (*obj.Prog)(gins(ppc64.ASRD, nil, &r1))
p2.From.Type = obj.TYPE_CONST
p2.From.Offset = 1
gc.Patch(p1, gc.Pc)
}
regalloc(&r2, gc.Types[gc.TFLOAT64], t)
p1 := (*obj.Prog)(gins(ppc64.AMOVD, &r1, nil))
p1.To.Type = obj.TYPE_MEM
p1.To.Reg = ppc64.REGSP
p1.To.Offset = -8
p1 = gins(ppc64.AFMOVD, nil, &r2)
p1.From.Type = obj.TYPE_MEM
p1.From.Reg = ppc64.REGSP
p1.From.Offset = -8
gins(ppc64.AFCFID, &r2, &r2)
regfree(&r1)
if ft == gc.TUINT64 {
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT64]), nil, +1)) // use CR0 here again
gc.Nodreg(&r1, gc.Types[gc.TFLOAT64], ppc64.FREGTWO)
gins(ppc64.AFMUL, &r1, &r2)
gc.Patch(p1, gc.Pc)
}
gmove(&r2, t)
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:
{
regalloc(&r1, t.Type, t)
gins(a, f, &r1)
gmove(&r1, t)
regfree(&r1)
return
}
// requires register intermediate
hard:
regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
regfree(&r1)
return
}
/*
* generate one instruction:
* as f, t
*/
func gins(as int, f *gc.Node, t *gc.Node) *obj.Prog {
// TODO(austin): Add self-move test like in 6g (but be careful
// of truncation moves)
af := obj.Addr(obj.Addr{})
at := obj.Addr(obj.Addr{})
if f != nil {
af = gc.Naddr(f)
}
if t != nil {
at = gc.Naddr(t)
}
p := (*obj.Prog)(gc.Prog(as))
if f != nil {
p.From = af
}
if t != nil {
p.To = at
}
if gc.Debug['g'] != 0 {
fmt.Printf("%v\n", p)
}
w := int32(0)
switch as {
case ppc64.AMOVB,
ppc64.AMOVBU,
ppc64.AMOVBZ,
ppc64.AMOVBZU:
w = 1
case ppc64.AMOVH,
ppc64.AMOVHU,
ppc64.AMOVHZ,
ppc64.AMOVHZU:
w = 2
case ppc64.AMOVW,
ppc64.AMOVWU,
ppc64.AMOVWZ,
ppc64.AMOVWZU:
w = 4
case ppc64.AMOVD,
ppc64.AMOVDU:
if af.Type == obj.TYPE_CONST || af.Type == obj.TYPE_ADDR {
break
}
w = 8
}
if w != 0 && ((f != nil && af.Width < int64(w)) || (t != nil && at.Type != obj.TYPE_REG && at.Width > int64(w))) {
gc.Dump("f", f)
gc.Dump("t", t)
gc.Fatal("bad width: %v (%d, %d)\n", p, af.Width, at.Width)
}
return p
}
func fixlargeoffset(n *gc.Node) {
if n == nil {
return
}
if n.Op != gc.OINDREG {
return
}
if n.Val.U.Reg == ppc64.REGSP { // stack offset cannot be large
return
}
if n.Xoffset != int64(int32(n.Xoffset)) {
// TODO(minux): offset too large, move into R31 and add to R31 instead.
// this is used only in test/fixedbugs/issue6036.go.
gc.Fatal("offset too large: %v", gc.Nconv(n, 0))
a := gc.Node(*n)
a.Op = gc.OREGISTER
a.Type = gc.Types[gc.Tptr]
a.Xoffset = 0
gc.Cgen_checknil(&a)
ginscon(optoas(gc.OADD, gc.Types[gc.Tptr]), n.Xoffset, &a)
n.Xoffset = 0
}
}
/*
* return Axxx for Oxxx on type t.
*/
func optoas(op int, t *gc.Type) int {
if t == nil {
gc.Fatal("optoas: t is nil")
}
a := int(obj.AXXX)
switch uint32(op)<<16 | uint32(gc.Simtype[t.Etype]) {
default:
gc.Fatal("optoas: no entry for op=%v type=%v", gc.Oconv(int(op), 0), gc.Tconv(t, 0))
case gc.OEQ<<16 | gc.TBOOL,
gc.OEQ<<16 | gc.TINT8,
gc.OEQ<<16 | gc.TUINT8,
gc.OEQ<<16 | gc.TINT16,
gc.OEQ<<16 | gc.TUINT16,
gc.OEQ<<16 | gc.TINT32,
gc.OEQ<<16 | gc.TUINT32,
gc.OEQ<<16 | gc.TINT64,
gc.OEQ<<16 | gc.TUINT64,
gc.OEQ<<16 | gc.TPTR32,
gc.OEQ<<16 | gc.TPTR64,
gc.OEQ<<16 | gc.TFLOAT32,
gc.OEQ<<16 | gc.TFLOAT64:
a = ppc64.ABEQ
case gc.ONE<<16 | gc.TBOOL,
gc.ONE<<16 | gc.TINT8,
gc.ONE<<16 | gc.TUINT8,
gc.ONE<<16 | gc.TINT16,
gc.ONE<<16 | gc.TUINT16,
gc.ONE<<16 | gc.TINT32,
gc.ONE<<16 | gc.TUINT32,
gc.ONE<<16 | gc.TINT64,
gc.ONE<<16 | gc.TUINT64,
gc.ONE<<16 | gc.TPTR32,
gc.ONE<<16 | gc.TPTR64,
gc.ONE<<16 | gc.TFLOAT32,
gc.ONE<<16 | gc.TFLOAT64:
a = ppc64.ABNE
case gc.OLT<<16 | gc.TINT8, // ACMP
gc.OLT<<16 | gc.TINT16,
gc.OLT<<16 | gc.TINT32,
gc.OLT<<16 | gc.TINT64,
gc.OLT<<16 | gc.TUINT8,
// ACMPU
gc.OLT<<16 | gc.TUINT16,
gc.OLT<<16 | gc.TUINT32,
gc.OLT<<16 | gc.TUINT64,
gc.OLT<<16 | gc.TFLOAT32,
// AFCMPU
gc.OLT<<16 | gc.TFLOAT64:
a = ppc64.ABLT
case gc.OLE<<16 | gc.TINT8, // ACMP
gc.OLE<<16 | gc.TINT16,
gc.OLE<<16 | gc.TINT32,
gc.OLE<<16 | gc.TINT64,
gc.OLE<<16 | gc.TUINT8,
// ACMPU
gc.OLE<<16 | gc.TUINT16,
gc.OLE<<16 | gc.TUINT32,
gc.OLE<<16 | gc.TUINT64,
gc.OLE<<16 | gc.TFLOAT32,
// AFCMPU
gc.OLE<<16 | gc.TFLOAT64:
a = ppc64.ABLE
case gc.OGT<<16 | gc.TINT8,
gc.OGT<<16 | gc.TINT16,
gc.OGT<<16 | gc.TINT32,
gc.OGT<<16 | gc.TINT64,
gc.OGT<<16 | gc.TUINT8,
gc.OGT<<16 | gc.TUINT16,
gc.OGT<<16 | gc.TUINT32,
gc.OGT<<16 | gc.TUINT64,
gc.OGT<<16 | gc.TFLOAT32,
gc.OGT<<16 | gc.TFLOAT64:
a = ppc64.ABGT
case gc.OGE<<16 | gc.TINT8,
gc.OGE<<16 | gc.TINT16,
gc.OGE<<16 | gc.TINT32,
gc.OGE<<16 | gc.TINT64,
gc.OGE<<16 | gc.TUINT8,
gc.OGE<<16 | gc.TUINT16,
gc.OGE<<16 | gc.TUINT32,
gc.OGE<<16 | gc.TUINT64,
gc.OGE<<16 | gc.TFLOAT32,
gc.OGE<<16 | gc.TFLOAT64:
a = ppc64.ABGE
case gc.OCMP<<16 | gc.TBOOL,
gc.OCMP<<16 | gc.TINT8,
gc.OCMP<<16 | gc.TINT16,
gc.OCMP<<16 | gc.TINT32,
gc.OCMP<<16 | gc.TPTR32,
gc.OCMP<<16 | gc.TINT64:
a = ppc64.ACMP
case gc.OCMP<<16 | gc.TUINT8,
gc.OCMP<<16 | gc.TUINT16,
gc.OCMP<<16 | gc.TUINT32,
gc.OCMP<<16 | gc.TUINT64,
gc.OCMP<<16 | gc.TPTR64:
a = ppc64.ACMPU
case gc.OCMP<<16 | gc.TFLOAT32,
gc.OCMP<<16 | gc.TFLOAT64:
a = ppc64.AFCMPU
case gc.OAS<<16 | gc.TBOOL,
gc.OAS<<16 | gc.TINT8:
a = ppc64.AMOVB
case gc.OAS<<16 | gc.TUINT8:
a = ppc64.AMOVBZ
case gc.OAS<<16 | gc.TINT16:
a = ppc64.AMOVH
case gc.OAS<<16 | gc.TUINT16:
a = ppc64.AMOVHZ
case gc.OAS<<16 | gc.TINT32:
a = ppc64.AMOVW
case gc.OAS<<16 | gc.TUINT32,
gc.OAS<<16 | gc.TPTR32:
a = ppc64.AMOVWZ
case gc.OAS<<16 | gc.TINT64,
gc.OAS<<16 | gc.TUINT64,
gc.OAS<<16 | gc.TPTR64:
a = ppc64.AMOVD
case gc.OAS<<16 | gc.TFLOAT32:
a = ppc64.AFMOVS
case gc.OAS<<16 | gc.TFLOAT64:
a = ppc64.AFMOVD
case gc.OADD<<16 | gc.TINT8,
gc.OADD<<16 | gc.TUINT8,
gc.OADD<<16 | gc.TINT16,
gc.OADD<<16 | gc.TUINT16,
gc.OADD<<16 | gc.TINT32,
gc.OADD<<16 | gc.TUINT32,
gc.OADD<<16 | gc.TPTR32,
gc.OADD<<16 | gc.TINT64,
gc.OADD<<16 | gc.TUINT64,
gc.OADD<<16 | gc.TPTR64:
a = ppc64.AADD
case gc.OADD<<16 | gc.TFLOAT32:
a = ppc64.AFADDS
case gc.OADD<<16 | gc.TFLOAT64:
a = ppc64.AFADD
case gc.OSUB<<16 | gc.TINT8,
gc.OSUB<<16 | gc.TUINT8,
gc.OSUB<<16 | gc.TINT16,
gc.OSUB<<16 | gc.TUINT16,
gc.OSUB<<16 | gc.TINT32,
gc.OSUB<<16 | gc.TUINT32,
gc.OSUB<<16 | gc.TPTR32,
gc.OSUB<<16 | gc.TINT64,
gc.OSUB<<16 | gc.TUINT64,
gc.OSUB<<16 | gc.TPTR64:
a = ppc64.ASUB
case gc.OSUB<<16 | gc.TFLOAT32:
a = ppc64.AFSUBS
case gc.OSUB<<16 | gc.TFLOAT64:
a = ppc64.AFSUB
case gc.OMINUS<<16 | gc.TINT8,
gc.OMINUS<<16 | gc.TUINT8,
gc.OMINUS<<16 | gc.TINT16,
gc.OMINUS<<16 | gc.TUINT16,
gc.OMINUS<<16 | gc.TINT32,
gc.OMINUS<<16 | gc.TUINT32,
gc.OMINUS<<16 | gc.TPTR32,
gc.OMINUS<<16 | gc.TINT64,
gc.OMINUS<<16 | gc.TUINT64,
gc.OMINUS<<16 | gc.TPTR64:
a = ppc64.ANEG
case gc.OAND<<16 | gc.TINT8,
gc.OAND<<16 | gc.TUINT8,
gc.OAND<<16 | gc.TINT16,
gc.OAND<<16 | gc.TUINT16,
gc.OAND<<16 | gc.TINT32,
gc.OAND<<16 | gc.TUINT32,
gc.OAND<<16 | gc.TPTR32,
gc.OAND<<16 | gc.TINT64,
gc.OAND<<16 | gc.TUINT64,
gc.OAND<<16 | gc.TPTR64:
a = ppc64.AAND
case gc.OOR<<16 | gc.TINT8,
gc.OOR<<16 | gc.TUINT8,
gc.OOR<<16 | gc.TINT16,
gc.OOR<<16 | gc.TUINT16,
gc.OOR<<16 | gc.TINT32,
gc.OOR<<16 | gc.TUINT32,
gc.OOR<<16 | gc.TPTR32,
gc.OOR<<16 | gc.TINT64,
gc.OOR<<16 | gc.TUINT64,
gc.OOR<<16 | gc.TPTR64:
a = ppc64.AOR
case gc.OXOR<<16 | gc.TINT8,
gc.OXOR<<16 | gc.TUINT8,
gc.OXOR<<16 | gc.TINT16,
gc.OXOR<<16 | gc.TUINT16,
gc.OXOR<<16 | gc.TINT32,
gc.OXOR<<16 | gc.TUINT32,
gc.OXOR<<16 | gc.TPTR32,
gc.OXOR<<16 | gc.TINT64,
gc.OXOR<<16 | gc.TUINT64,
gc.OXOR<<16 | gc.TPTR64:
a = ppc64.AXOR
// TODO(minux): handle rotates
//case CASE(OLROT, TINT8):
//case CASE(OLROT, TUINT8):
//case CASE(OLROT, TINT16):
//case CASE(OLROT, TUINT16):
//case CASE(OLROT, TINT32):
//case CASE(OLROT, TUINT32):
//case CASE(OLROT, TPTR32):
//case CASE(OLROT, TINT64):
//case CASE(OLROT, TUINT64):
//case CASE(OLROT, TPTR64):
// a = 0//???; RLDC?
// break;
case gc.OLSH<<16 | gc.TINT8,
gc.OLSH<<16 | gc.TUINT8,
gc.OLSH<<16 | gc.TINT16,
gc.OLSH<<16 | gc.TUINT16,
gc.OLSH<<16 | gc.TINT32,
gc.OLSH<<16 | gc.TUINT32,
gc.OLSH<<16 | gc.TPTR32,
gc.OLSH<<16 | gc.TINT64,
gc.OLSH<<16 | gc.TUINT64,
gc.OLSH<<16 | gc.TPTR64:
a = ppc64.ASLD
case gc.ORSH<<16 | gc.TUINT8,
gc.ORSH<<16 | gc.TUINT16,
gc.ORSH<<16 | gc.TUINT32,
gc.ORSH<<16 | gc.TPTR32,
gc.ORSH<<16 | gc.TUINT64,
gc.ORSH<<16 | gc.TPTR64:
a = ppc64.ASRD
case gc.ORSH<<16 | gc.TINT8,
gc.ORSH<<16 | gc.TINT16,
gc.ORSH<<16 | gc.TINT32,
gc.ORSH<<16 | gc.TINT64:
a = ppc64.ASRAD
// TODO(minux): handle rotates
//case CASE(ORROTC, TINT8):
//case CASE(ORROTC, TUINT8):
//case CASE(ORROTC, TINT16):
//case CASE(ORROTC, TUINT16):
//case CASE(ORROTC, TINT32):
//case CASE(ORROTC, TUINT32):
//case CASE(ORROTC, TINT64):
//case CASE(ORROTC, TUINT64):
// a = 0//??? RLDC??
// break;
case gc.OHMUL<<16 | gc.TINT64:
a = ppc64.AMULHD
case gc.OHMUL<<16 | gc.TUINT64,
gc.OHMUL<<16 | gc.TPTR64:
a = ppc64.AMULHDU
case gc.OMUL<<16 | gc.TINT8,
gc.OMUL<<16 | gc.TINT16,
gc.OMUL<<16 | gc.TINT32,
gc.OMUL<<16 | gc.TINT64:
a = ppc64.AMULLD
case gc.OMUL<<16 | gc.TUINT8,
gc.OMUL<<16 | gc.TUINT16,
gc.OMUL<<16 | gc.TUINT32,
gc.OMUL<<16 | gc.TPTR32,
// don't use word multiply, the high 32-bit are undefined.
// fallthrough
gc.OMUL<<16 | gc.TUINT64,
gc.OMUL<<16 | gc.TPTR64:
a = ppc64.AMULLD
// for 64-bit multiplies, signedness doesn't matter.
case gc.OMUL<<16 | gc.TFLOAT32:
a = ppc64.AFMULS
case gc.OMUL<<16 | gc.TFLOAT64:
a = ppc64.AFMUL
case gc.ODIV<<16 | gc.TINT8,
gc.ODIV<<16 | gc.TINT16,
gc.ODIV<<16 | gc.TINT32,
gc.ODIV<<16 | gc.TINT64:
a = ppc64.ADIVD
case gc.ODIV<<16 | gc.TUINT8,
gc.ODIV<<16 | gc.TUINT16,
gc.ODIV<<16 | gc.TUINT32,
gc.ODIV<<16 | gc.TPTR32,
gc.ODIV<<16 | gc.TUINT64,
gc.ODIV<<16 | gc.TPTR64:
a = ppc64.ADIVDU
case gc.ODIV<<16 | gc.TFLOAT32:
a = ppc64.AFDIVS
case gc.ODIV<<16 | gc.TFLOAT64:
a = ppc64.AFDIV
}
return a
}
const (
ODynam = 1 << 0
OAddable = 1 << 1
)
func xgen(n *gc.Node, a *gc.Node, o int) bool {
// TODO(minux)
return -1 != 0 /*TypeKind(100016)*/
}
func sudoclean() {
return
}
/*
* generate code to compute address of n,
* a reference to a (perhaps nested) field inside
* an array or struct.
* return 0 on failure, 1 on success.
* on success, leaves usable address in a.
*
* caller is responsible for calling sudoclean
* after successful sudoaddable,
* to release the register used for a.
*/
func sudoaddable(as int, n *gc.Node, a *obj.Addr) bool {
// TODO(minux)
*a = obj.Addr{}
return false
}
/*
* n is call to interface method.
* generate res = n.
*/
func cgen_callinter(n *gc.Node, res *gc.Node, proc int) {
i := n.Left
if i.Op != gc.ODOTINTER {
gc.Fatal("cgen_callinter: not ODOTINTER %v", gc.Oconv(int(i.Op), 0))
}
f := i.Right // field
if f.Op != gc.ONAME {
gc.Fatal("cgen_callinter: not ONAME %v", gc.Oconv(int(f.Op), 0))
}
i = i.Left // interface
// Release res register during genlist and cgen,
// which might have their own function calls.
r := -1
if res != nil && (res.Op == gc.OREGISTER || res.Op == gc.OINDREG) {
r = int(res.Val.U.Reg)
reg[r]--
}
if i.Addable == 0 {
var tmpi gc.Node
gc.Tempname(&tmpi, i.Type)
cgen(i, &tmpi)
i = &tmpi
}
gc.Genlist(n.List) // args
if r >= 0 {
reg[r]++
}
var nodr gc.Node
regalloc(&nodr, gc.Types[gc.Tptr], res)
var nodo gc.Node
regalloc(&nodo, gc.Types[gc.Tptr], &nodr)
nodo.Op = gc.OINDREG
agen(i, &nodr) // REG = &inter
var nodsp gc.Node
gc.Nodindreg(&nodsp, gc.Types[gc.Tptr], arm.REGSP)
nodsp.Xoffset = int64(gc.Widthptr)
if proc != 0 {
nodsp.Xoffset += 2 * int64(gc.Widthptr) // leave room for size & fn
}
nodo.Xoffset += int64(gc.Widthptr)
cgen(&nodo, &nodsp) // {4 or 12}(SP) = 4(REG) -- i.data
nodo.Xoffset -= int64(gc.Widthptr)
cgen(&nodo, &nodr) // REG = 0(REG) -- i.tab
gc.Cgen_checknil(&nodr) // in case offset is huge
nodo.Xoffset = n.Left.Xoffset + 3*int64(gc.Widthptr) + 8
if proc == 0 {
// plain call: use direct c function pointer - more efficient
cgen(&nodo, &nodr) // REG = 20+offset(REG) -- i.tab->fun[f]
nodr.Op = gc.OINDREG
proc = 3
} else {
// go/defer. generate go func value.
p := gins(arm.AMOVW, &nodo, &nodr)
p.From.Type = obj.TYPE_ADDR // REG = &(20+offset(REG)) -- i.tab->fun[f]
}
nodr.Type = n.Left.Type
ginscall(&nodr, proc)
regfree(&nodr)
regfree(&nodo)
}
/*
* n is call to interface method.
* generate res = n.
*/
func cgen_callinter(n *gc.Node, res *gc.Node, proc int) {
i := n.Left
if i.Op != gc.ODOTINTER {
gc.Fatal("cgen_callinter: not ODOTINTER %v", gc.Oconv(int(i.Op), 0))
}
f := i.Right // field
if f.Op != gc.ONAME {
gc.Fatal("cgen_callinter: not ONAME %v", gc.Oconv(int(f.Op), 0))
}
i = i.Left // interface
if i.Addable == 0 {
var tmpi gc.Node
gc.Tempname(&tmpi, i.Type)
cgen(i, &tmpi)
i = &tmpi
}
gc.Genlist(n.List) // assign the args
// i is now addable, prepare an indirected
// register to hold its address.
var nodi gc.Node
igen(i, &nodi, res) // REG = &inter
var nodsp gc.Node
gc.Nodindreg(&nodsp, gc.Types[gc.Tptr], x86.REG_SP)
nodsp.Xoffset = 0
if proc != 0 {
nodsp.Xoffset += 2 * int64(gc.Widthptr) // leave room for size & fn
}
nodi.Type = gc.Types[gc.Tptr]
nodi.Xoffset += int64(gc.Widthptr)
cgen(&nodi, &nodsp) // {0 or 8}(SP) = 4(REG) -- i.data
var nodo gc.Node
regalloc(&nodo, gc.Types[gc.Tptr], res)
nodi.Type = gc.Types[gc.Tptr]
nodi.Xoffset -= int64(gc.Widthptr)
cgen(&nodi, &nodo) // REG = 0(REG) -- i.tab
regfree(&nodi)
var nodr gc.Node
regalloc(&nodr, gc.Types[gc.Tptr], &nodo)
if n.Left.Xoffset == gc.BADWIDTH {
gc.Fatal("cgen_callinter: badwidth")
}
gc.Cgen_checknil(&nodo)
nodo.Op = gc.OINDREG
nodo.Xoffset = n.Left.Xoffset + 3*int64(gc.Widthptr) + 8
if proc == 0 {
// plain call: use direct c function pointer - more efficient
cgen(&nodo, &nodr) // REG = 20+offset(REG) -- i.tab->fun[f]
proc = 3
} else {
// go/defer. generate go func value.
gins(x86.ALEAL, &nodo, &nodr) // REG = &(20+offset(REG)) -- i.tab->fun[f]
}
nodr.Type = n.Left.Type
ginscall(&nodr, proc)
regfree(&nodr)
regfree(&nodo)
}
/*
* 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.Fatal("cgen64 %v of %v", gc.Oconv(int(n.Op), 0), gc.Oconv(int(res.Op), 0))
}
switch n.Op {
default:
gc.Fatal("cgen64 %v", gc.Oconv(int(n.Op), 0))
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)
// let's call the next two EX and FX.
case gc.OMUL:
var ex gc.Node
gc.Regalloc(&ex, gc.Types[gc.TPTR32], nil)
var fx gc.Node
gc.Regalloc(&fx, gc.Types[gc.TPTR32], nil)
// load args into DX:AX and EX:CX.
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(x86.AMOVL, &lo2, &cx)
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, &cx, nil) // implicit &ax
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// full 64x64 -> 64, from 32x32 -> 64.
gins(x86.AIMULL, &cx, &dx)
gins(x86.AMOVL, &ax, &fx)
gins(x86.AIMULL, &ex, &fx)
gins(x86.AADDL, &dx, &fx)
gins(x86.AMOVL, &cx, &dx)
gins(x86.AMULL, &dx, nil) // implicit &ax
gins(x86.AADDL, &fx, &dx)
gc.Patch(p2, gc.Pc)
gc.Regfree(&ex)
gc.Regfree(&fx)
// 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(gc.Mpgetfix(r.Val.U.Xval))
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(gc.Mpgetfix(r.Val.U.Xval))
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(gc.Mpgetfix(r.Val.U.Xval))
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(gc.Mpgetfix(lo2.Val.U.Xval))
hv := uint32(gc.Mpgetfix(hi2.Val.U.Xval))
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(int(n.Op), lo1.Type), &lo2, &ax)
gins(optoas(int(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()
}
/*
* generate:
* if(n == true) goto to;
*/
func bgen(n *gc.Node, true_ bool, likely int, to *obj.Prog) {
if gc.Debug['g'] != 0 {
gc.Dump("\nbgen", n)
}
if n == nil {
n = gc.Nodbool(true)
}
if n.Ninit != nil {
gc.Genlist(n.Ninit)
}
if n.Type == nil {
gc.Convlit(&n, gc.Types[gc.TBOOL])
if n.Type == nil {
return
}
}
et := int(n.Type.Etype)
if et != gc.TBOOL {
gc.Yyerror("cgen: bad type %v for %v", gc.Tconv(n.Type, 0), gc.Oconv(int(n.Op), 0))
gc.Patch(gins(obj.AEND, nil, nil), to)
return
}
for n.Op == gc.OCONVNOP {
n = n.Left
if n.Ninit != nil {
gc.Genlist(n.Ninit)
}
}
var nl *gc.Node
var nr *gc.Node
switch n.Op {
default:
goto def
// need to ask if it is bool?
case gc.OLITERAL:
if !true_ == (n.Val.U.Bval == 0) {
gc.Patch(gc.Gbranch(obj.AJMP, nil, likely), to)
}
return
case gc.ONAME:
if n.Addable == 0 {
goto def
}
var n1 gc.Node
gc.Nodconst(&n1, n.Type, 0)
gins(optoas(gc.OCMP, n.Type), n, &n1)
a := x86.AJNE
if !true_ {
a = x86.AJEQ
}
gc.Patch(gc.Gbranch(a, n.Type, likely), to)
return
case gc.OANDAND,
gc.OOROR:
if (n.Op == gc.OANDAND) == true_ {
p1 := gc.Gbranch(obj.AJMP, nil, 0)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
bgen(n.Left, !true_, -likely, p2)
bgen(n.Right, !true_, -likely, p2)
p1 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, to)
gc.Patch(p2, gc.Pc)
} else {
bgen(n.Left, true_, likely, to)
bgen(n.Right, true_, likely, to)
}
return
case gc.OEQ,
gc.ONE,
gc.OLT,
gc.OGT,
gc.OLE,
gc.OGE:
nr = n.Right
if nr == nil || nr.Type == nil {
return
}
fallthrough
case gc.ONOT: // unary
nl = n.Left
if nl == nil || nl.Type == nil {
return
}
}
switch n.Op {
case gc.ONOT:
bgen(nl, !true_, likely, to)
return
case gc.OEQ,
gc.ONE,
gc.OLT,
gc.OGT,
gc.OLE,
gc.OGE:
a := int(n.Op)
if !true_ {
if gc.Isfloat[nr.Type.Etype] {
// 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)
ll := n.Ninit // avoid re-genning ninit
n.Ninit = nil
bgen(n, true, -likely, p2)
n.Ninit = ll
gc.Patch(gc.Gbranch(obj.AJMP, nil, 0), to)
gc.Patch(p2, gc.Pc)
return
}
a = gc.Brcom(a)
true_ = !true_
}
// make simplest on right
if nl.Op == gc.OLITERAL || (nl.Ullman < nr.Ullman && nl.Ullman < gc.UINF) {
a = gc.Brrev(a)
r := nl
nl = nr
nr = r
}
if gc.Isslice(nl.Type) {
// front end should only leave cmp to literal nil
if (a != gc.OEQ && a != gc.ONE) || nr.Op != gc.OLITERAL {
gc.Yyerror("illegal slice comparison")
break
}
a = optoas(a, gc.Types[gc.Tptr])
var n1 gc.Node
igen(nl, &n1, nil)
n1.Xoffset += int64(gc.Array_array)
n1.Type = gc.Types[gc.Tptr]
var tmp gc.Node
gc.Nodconst(&tmp, gc.Types[gc.Tptr], 0)
gins(optoas(gc.OCMP, gc.Types[gc.Tptr]), &n1, &tmp)
gc.Patch(gc.Gbranch(a, gc.Types[gc.Tptr], likely), to)
regfree(&n1)
break
}
if gc.Isinter(nl.Type) {
// front end should only leave cmp to literal nil
if (a != gc.OEQ && a != gc.ONE) || nr.Op != gc.OLITERAL {
gc.Yyerror("illegal interface comparison")
break
}
a = optoas(a, gc.Types[gc.Tptr])
var n1 gc.Node
igen(nl, &n1, nil)
n1.Type = gc.Types[gc.Tptr]
var tmp gc.Node
gc.Nodconst(&tmp, gc.Types[gc.Tptr], 0)
gins(optoas(gc.OCMP, gc.Types[gc.Tptr]), &n1, &tmp)
gc.Patch(gc.Gbranch(a, gc.Types[gc.Tptr], likely), to)
regfree(&n1)
break
}
if gc.Iscomplex[nl.Type.Etype] {
gc.Complexbool(a, nl, nr, true_, likely, to)
break
}
var n2 gc.Node
var n1 gc.Node
if nr.Ullman >= gc.UINF {
regalloc(&n1, nl.Type, nil)
cgen(nl, &n1)
var tmp gc.Node
gc.Tempname(&tmp, nl.Type)
gmove(&n1, &tmp)
regfree(&n1)
regalloc(&n2, nr.Type, nil)
cgen(nr, &n2)
regalloc(&n1, nl.Type, nil)
cgen(&tmp, &n1)
goto cmp
}
regalloc(&n1, nl.Type, nil)
cgen(nl, &n1)
if gc.Smallintconst(nr) {
gins(optoas(gc.OCMP, nr.Type), &n1, nr)
gc.Patch(gc.Gbranch(optoas(a, nr.Type), nr.Type, likely), to)
regfree(&n1)
break
}
regalloc(&n2, nr.Type, nil)
cgen(nr, &n2)
// only < and <= work right with NaN; reverse if needed
cmp:
l := &n1
r := &n2
if gc.Isfloat[nl.Type.Etype] && (a == gc.OGT || a == gc.OGE) {
l = &n2
r = &n1
a = gc.Brrev(a)
}
gins(optoas(gc.OCMP, nr.Type), l, r)
if gc.Isfloat[nr.Type.Etype] && (n.Op == gc.OEQ || n.Op == gc.ONE) {
if n.Op == gc.OEQ {
// neither NE nor P
p1 := gc.Gbranch(x86.AJNE, nil, -likely)
p2 := gc.Gbranch(x86.AJPS, nil, -likely)
gc.Patch(gc.Gbranch(obj.AJMP, nil, 0), to)
gc.Patch(p1, gc.Pc)
gc.Patch(p2, gc.Pc)
} else {
// either NE or P
gc.Patch(gc.Gbranch(x86.AJNE, nil, likely), to)
gc.Patch(gc.Gbranch(x86.AJPS, nil, likely), to)
}
} else {
gc.Patch(gc.Gbranch(optoas(a, nr.Type), nr.Type, likely), to)
}
regfree(&n1)
regfree(&n2)
}
return
def:
var n1 gc.Node
regalloc(&n1, n.Type, nil)
cgen(n, &n1)
var n2 gc.Node
gc.Nodconst(&n2, n.Type, 0)
gins(optoas(gc.OCMP, n.Type), &n1, &n2)
a := x86.AJNE
if !true_ {
a = x86.AJEQ
}
gc.Patch(gc.Gbranch(a, n.Type, likely), to)
regfree(&n1)
return
}
/*
* generate division.
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op int, nl *gc.Node, nr *gc.Node, res *gc.Node) {
// Have to be careful about handling
// most negative int divided by -1 correctly.
// The hardware will trap.
// Also the byte divide instruction needs AH,
// which we otherwise don't have to deal with.
// Easiest way to avoid for int8, int16: use int32.
// For int32 and int64, use explicit test.
// Could use int64 hw for int32.
t := nl.Type
t0 := t
check := 0
if gc.Issigned[t.Etype] {
check = 1
if gc.Isconst(nl, gc.CTINT) && gc.Mpgetfix(nl.Val.U.Xval) != -(1<<uint64(t.Width*8-1)) {
check = 0
} else if gc.Isconst(nr, gc.CTINT) && gc.Mpgetfix(nr.Val.U.Xval) != -1 {
check = 0
}
}
if t.Width < 4 {
if gc.Issigned[t.Etype] {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
check = 0
}
a := optoas(op, t)
var n3 gc.Node
gc.Regalloc(&n3, t0, nil)
var ax gc.Node
var oldax gc.Node
if nl.Ullman >= nr.Ullman {
savex(x86.REG_AX, &ax, &oldax, res, t0)
gc.Cgen(nl, &ax)
gc.Regalloc(&ax, t0, &ax) // mark ax live during cgen
gc.Cgen(nr, &n3)
gc.Regfree(&ax)
} else {
gc.Cgen(nr, &n3)
savex(x86.REG_AX, &ax, &oldax, res, t0)
gc.Cgen(nl, &ax)
}
if t != t0 {
// Convert
ax1 := ax
n31 := n3
ax.Type = t
n3.Type = t
gmove(&ax1, &ax)
gmove(&n31, &n3)
}
var n4 gc.Node
if gc.Nacl {
// Native Client does not relay the divide-by-zero trap
// to the executing program, so we must insert a check
// for ourselves.
gc.Nodconst(&n4, t, 0)
gins(optoas(gc.OCMP, t), &n3, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if panicdiv == nil {
panicdiv = gc.Sysfunc("panicdivide")
}
gc.Ginscall(panicdiv, -1)
gc.Patch(p1, gc.Pc)
}
var p2 *obj.Prog
if check != 0 {
gc.Nodconst(&n4, t, -1)
gins(optoas(gc.OCMP, t), &n3, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if op == gc.ODIV {
// a / (-1) is -a.
gins(optoas(gc.OMINUS, t), nil, &ax)
gmove(&ax, res)
} else {
// a % (-1) is 0.
gc.Nodconst(&n4, t, 0)
gmove(&n4, res)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
}
var olddx gc.Node
var dx gc.Node
savex(x86.REG_DX, &dx, &olddx, res, t)
if !gc.Issigned[t.Etype] {
gc.Nodconst(&n4, t, 0)
gmove(&n4, &dx)
} else {
gins(optoas(gc.OEXTEND, t), nil, nil)
}
gins(a, &n3, nil)
gc.Regfree(&n3)
if op == gc.ODIV {
gmove(&ax, res)
} else {
gmove(&dx, res)
}
restx(&dx, &olddx)
if check != 0 {
gc.Patch(p2, gc.Pc)
}
restx(&ax, &oldax)
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
gc.Dump("\nclearfat", nl)
}
w := uint32(nl.Type.Width)
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := w % 4 // bytes
q := w / 4 // quads
if q < 4 {
// Write sequence of MOV 0, off(base) instead of using STOSL.
// The hope is that although the code will be slightly longer,
// the MOVs will have no dependencies and pipeline better
// than the unrolled STOSL loop.
// NOTE: Must use agen, not igen, so that optimizer sees address
// being taken. We are not writing on field boundaries.
var n1 gc.Node
gc.Regalloc(&n1, gc.Types[gc.Tptr], nil)
gc.Agen(nl, &n1)
n1.Op = gc.OINDREG
var z gc.Node
gc.Nodconst(&z, gc.Types[gc.TUINT64], 0)
for {
tmp14 := q
q--
if tmp14 <= 0 {
break
}
n1.Type = z.Type
gins(x86.AMOVL, &z, &n1)
n1.Xoffset += 4
}
gc.Nodconst(&z, gc.Types[gc.TUINT8], 0)
for {
tmp15 := c
c--
if tmp15 <= 0 {
break
}
n1.Type = z.Type
gins(x86.AMOVB, &z, &n1)
n1.Xoffset++
}
gc.Regfree(&n1)
return
}
var n1 gc.Node
gc.Nodreg(&n1, gc.Types[gc.Tptr], x86.REG_DI)
gc.Agen(nl, &n1)
gconreg(x86.AMOVL, 0, x86.REG_AX)
if q > 128 || (q >= 4 && gc.Nacl) {
gconreg(x86.AMOVL, int64(q), x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.ASTOSL, nil, nil) // STOL AL,*(DI)+
} else if q >= 4 {
p := gins(obj.ADUFFZERO, nil, nil)
p.To.Type = obj.TYPE_ADDR
p.To.Sym = gc.Linksym(gc.Pkglookup("duffzero", gc.Runtimepkg))
// 1 and 128 = magic constants: see ../../runtime/asm_386.s
p.To.Offset = 1 * (128 - int64(q))
} else {
for q > 0 {
gins(x86.ASTOSL, nil, nil) // STOL AL,*(DI)+
q--
}
}
for c > 0 {
gins(x86.ASTOSB, nil, nil) // STOB AL,*(DI)+
c--
}
}
/*
* generate:
* newreg = &n;
*
* caller must regfree(a).
* The generated code checks that the result is not nil.
*/
func agenr(n *gc.Node, a *gc.Node, res *gc.Node) {
if gc.Debug['g'] != 0 {
gc.Dump("agenr-n", n)
}
nl := n.Left
nr := n.Right
switch n.Op {
case gc.ODOT,
gc.ODOTPTR,
gc.OCALLFUNC,
gc.OCALLMETH,
gc.OCALLINTER:
var n1 gc.Node
igen(n, &n1, res)
regalloc(a, gc.Types[gc.Tptr], &n1)
agen(&n1, a)
regfree(&n1)
case gc.OIND:
cgenr(n.Left, a, res)
gc.Cgen_checknil(a)
case gc.OINDEX:
var p2 *obj.Prog // to be patched to panicindex.
w := uint32(n.Type.Width)
bounded := gc.Debug['B'] != 0 || n.Bounded
var n1 gc.Node
var n3 gc.Node
if nr.Addable != 0 {
var tmp gc.Node
if !gc.Isconst(nr, gc.CTINT) {
gc.Tempname(&tmp, gc.Types[gc.TINT32])
}
if !gc.Isconst(nl, gc.CTSTR) {
agenr(nl, &n3, res)
}
if !gc.Isconst(nr, gc.CTINT) {
p2 = cgenindex(nr, &tmp, bounded)
regalloc(&n1, tmp.Type, nil)
gmove(&tmp, &n1)
}
} else if nl.Addable != 0 {
if !gc.Isconst(nr, gc.CTINT) {
var tmp gc.Node
gc.Tempname(&tmp, gc.Types[gc.TINT32])
p2 = cgenindex(nr, &tmp, bounded)
regalloc(&n1, tmp.Type, nil)
gmove(&tmp, &n1)
}
if !gc.Isconst(nl, gc.CTSTR) {
agenr(nl, &n3, res)
}
} else {
var tmp gc.Node
gc.Tempname(&tmp, gc.Types[gc.TINT32])
p2 = cgenindex(nr, &tmp, bounded)
nr = &tmp
if !gc.Isconst(nl, gc.CTSTR) {
agenr(nl, &n3, res)
}
regalloc(&n1, tmp.Type, nil)
gins(optoas(gc.OAS, tmp.Type), &tmp, &n1)
}
// &a is in &n3 (allocated in res)
// i is in &n1 (if not constant)
// w is width
// constant index
if gc.Isconst(nr, gc.CTINT) {
if gc.Isconst(nl, gc.CTSTR) {
gc.Fatal("constant string constant index")
}
v := uint64(gc.Mpgetfix(nr.Val.U.Xval))
var n2 gc.Node
if gc.Isslice(nl.Type) || nl.Type.Etype == gc.TSTRING {
if gc.Debug['B'] == 0 && !n.Bounded {
n1 = n3
n1.Op = gc.OINDREG
n1.Type = gc.Types[gc.Tptr]
n1.Xoffset = int64(gc.Array_nel)
var n4 gc.Node
regalloc(&n4, n1.Type, nil)
gmove(&n1, &n4)
gc.Nodconst(&n2, gc.Types[gc.TUINT32], int64(v))
gcmp(optoas(gc.OCMP, gc.Types[gc.TUINT32]), &n4, &n2)
regfree(&n4)
p1 := gc.Gbranch(optoas(gc.OGT, gc.Types[gc.TUINT32]), nil, +1)
ginscall(gc.Panicindex, 0)
gc.Patch(p1, gc.Pc)
}
n1 = n3
n1.Op = gc.OINDREG
n1.Type = gc.Types[gc.Tptr]
n1.Xoffset = int64(gc.Array_array)
gmove(&n1, &n3)
}
gc.Nodconst(&n2, gc.Types[gc.Tptr], int64(v*uint64(w)))
gins(optoas(gc.OADD, gc.Types[gc.Tptr]), &n2, &n3)
*a = n3
break
}
var n2 gc.Node
regalloc(&n2, gc.Types[gc.TINT32], &n1) // i
gmove(&n1, &n2)
regfree(&n1)
var n4 gc.Node
if gc.Debug['B'] == 0 && !n.Bounded {
// check bounds
if gc.Isconst(nl, gc.CTSTR) {
gc.Nodconst(&n4, gc.Types[gc.TUINT32], int64(len(nl.Val.U.Sval)))
} else if gc.Isslice(nl.Type) || nl.Type.Etype == gc.TSTRING {
n1 = n3
n1.Op = gc.OINDREG
n1.Type = gc.Types[gc.Tptr]
n1.Xoffset = int64(gc.Array_nel)
regalloc(&n4, gc.Types[gc.TUINT32], nil)
gmove(&n1, &n4)
} else {
gc.Nodconst(&n4, gc.Types[gc.TUINT32], nl.Type.Bound)
}
gcmp(optoas(gc.OCMP, gc.Types[gc.TUINT32]), &n2, &n4)
if n4.Op == gc.OREGISTER {
regfree(&n4)
}
p1 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
if p2 != nil {
gc.Patch(p2, gc.Pc)
}
ginscall(gc.Panicindex, 0)
gc.Patch(p1, gc.Pc)
}
if gc.Isconst(nl, gc.CTSTR) {
regalloc(&n3, gc.Types[gc.Tptr], res)
p1 := gins(arm.AMOVW, nil, &n3)
gc.Datastring(nl.Val.U.Sval, &p1.From)
p1.From.Type = obj.TYPE_ADDR
} else if gc.Isslice(nl.Type) || nl.Type.Etype == gc.TSTRING {
n1 = n3
n1.Op = gc.OINDREG
n1.Type = gc.Types[gc.Tptr]
n1.Xoffset = int64(gc.Array_array)
gmove(&n1, &n3)
}
if w == 0 {
} else // nothing to do
if w == 1 || w == 2 || w == 4 || w == 8 {
n4 = gc.Node{}
n4.Op = gc.OADDR
n4.Left = &n2
cgen(&n4, &n3)
if w == 1 {
gins(arm.AADD, &n2, &n3)
} else if w == 2 {
gshift(arm.AADD, &n2, arm.SHIFT_LL, 1, &n3)
} else if w == 4 {
gshift(arm.AADD, &n2, arm.SHIFT_LL, 2, &n3)
} else if w == 8 {
gshift(arm.AADD, &n2, arm.SHIFT_LL, 3, &n3)
}
} else {
regalloc(&n4, gc.Types[gc.TUINT32], nil)
gc.Nodconst(&n1, gc.Types[gc.TUINT32], int64(w))
gmove(&n1, &n4)
gins(optoas(gc.OMUL, gc.Types[gc.TUINT32]), &n4, &n2)
gins(optoas(gc.OADD, gc.Types[gc.Tptr]), &n2, &n3)
regfree(&n4)
}
*a = n3
regfree(&n2)
default:
regalloc(a, gc.Types[gc.Tptr], res)
agen(n, a)
}
}
/*
* block copy:
* memmove(&ns, &n, w);
*/
func sgen(n *gc.Node, ns *gc.Node, w int64) {
if gc.Debug['g'] != 0 {
fmt.Printf("\nsgen w=%d\n", w)
gc.Dump("r", n)
gc.Dump("res", ns)
}
if n.Ullman >= gc.UINF && ns.Ullman >= gc.UINF {
gc.Fatal("sgen UINF")
}
if w < 0 {
gc.Fatal("sgen copy %d", w)
}
// If copying .args, that's all the results, so record definition sites
// for them for the liveness analysis.
if ns.Op == gc.ONAME && ns.Sym.Name == ".args" {
for l := gc.Curfn.Dcl; l != nil; l = l.Next {
if l.N.Class == gc.PPARAMOUT {
gc.Gvardef(l.N)
}
}
}
// Avoid taking the address for simple enough types.
if componentgen(n, ns) {
return
}
if w == 0 {
// evaluate side effects only
var nodr gc.Node
regalloc(&nodr, gc.Types[gc.Tptr], nil)
agen(ns, &nodr)
agen(n, &nodr)
regfree(&nodr)
return
}
// offset on the stack
osrc := stkof(n)
odst := stkof(ns)
if osrc != -1000 && odst != -1000 && (osrc == 1000 || odst == 1000) {
// osrc and odst both on stack, and at least one is in
// an unknown position. Could generate code to test
// for forward/backward copy, but instead just copy
// to a temporary location first.
var tmp gc.Node
gc.Tempname(&tmp, n.Type)
sgen(n, &tmp, w)
sgen(&tmp, ns, w)
return
}
var noddi gc.Node
gc.Nodreg(&noddi, gc.Types[gc.Tptr], x86.REG_DI)
var nodsi gc.Node
gc.Nodreg(&nodsi, gc.Types[gc.Tptr], x86.REG_SI)
var nodl gc.Node
var nodr gc.Node
if n.Ullman >= ns.Ullman {
agenr(n, &nodr, &nodsi)
if ns.Op == gc.ONAME {
gc.Gvardef(ns)
}
agenr(ns, &nodl, &noddi)
} else {
if ns.Op == gc.ONAME {
gc.Gvardef(ns)
}
agenr(ns, &nodl, &noddi)
agenr(n, &nodr, &nodsi)
}
if nodl.Val.U.Reg != x86.REG_DI {
gmove(&nodl, &noddi)
}
if nodr.Val.U.Reg != x86.REG_SI {
gmove(&nodr, &nodsi)
}
regfree(&nodl)
regfree(&nodr)
c := w % 8 // bytes
q := w / 8 // quads
var oldcx gc.Node
var cx gc.Node
savex(x86.REG_CX, &cx, &oldcx, nil, gc.Types[gc.TINT64])
// if we are copying forward on the stack and
// the src and dst overlap, then reverse direction
if osrc < odst && odst < osrc+w {
// reverse direction
gins(x86.ASTD, nil, nil) // set direction flag
if c > 0 {
gconreg(addptr, w-1, x86.REG_SI)
gconreg(addptr, w-1, x86.REG_DI)
gconreg(movptr, c, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSB, nil, nil) // MOVB *(SI)-,*(DI)-
}
if q > 0 {
if c > 0 {
gconreg(addptr, -7, x86.REG_SI)
gconreg(addptr, -7, x86.REG_DI)
} else {
gconreg(addptr, w-8, x86.REG_SI)
gconreg(addptr, w-8, x86.REG_DI)
}
gconreg(movptr, q, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)-,*(DI)-
}
// we leave with the flag clear
gins(x86.ACLD, nil, nil)
} else {
// normal direction
if q > 128 || (gc.Nacl && q >= 4) {
gconreg(movptr, q, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)+,*(DI)+
} else if q >= 4 {
p := gins(obj.ADUFFCOPY, nil, nil)
p.To.Type = obj.TYPE_ADDR
p.To.Sym = gc.Linksym(gc.Pkglookup("duffcopy", gc.Runtimepkg))
// 14 and 128 = magic constants: see ../../runtime/asm_amd64.s
p.To.Offset = 14 * (128 - q)
} else if !gc.Nacl && c == 0 {
// We don't need the MOVSQ side-effect of updating SI and DI,
// and issuing a sequence of MOVQs directly is faster.
nodsi.Op = gc.OINDREG
noddi.Op = gc.OINDREG
for q > 0 {
gmove(&nodsi, &cx) // MOVQ x+(SI),CX
gmove(&cx, &noddi) // MOVQ CX,x+(DI)
nodsi.Xoffset += 8
noddi.Xoffset += 8
q--
}
} else {
for q > 0 {
gins(x86.AMOVSQ, nil, nil) // MOVQ *(SI)+,*(DI)+
q--
}
}
// copy the remaining c bytes
if w < 4 || c <= 1 || (odst < osrc && osrc < odst+w) {
for c > 0 {
gins(x86.AMOVSB, nil, nil) // MOVB *(SI)+,*(DI)+
c--
}
} else if w < 8 || c <= 4 {
nodsi.Op = gc.OINDREG
noddi.Op = gc.OINDREG
cx.Type = gc.Types[gc.TINT32]
nodsi.Type = gc.Types[gc.TINT32]
noddi.Type = gc.Types[gc.TINT32]
if c > 4 {
nodsi.Xoffset = 0
noddi.Xoffset = 0
gmove(&nodsi, &cx)
gmove(&cx, &noddi)
}
nodsi.Xoffset = c - 4
noddi.Xoffset = c - 4
gmove(&nodsi, &cx)
gmove(&cx, &noddi)
} else {
nodsi.Op = gc.OINDREG
noddi.Op = gc.OINDREG
cx.Type = gc.Types[gc.TINT64]
nodsi.Type = gc.Types[gc.TINT64]
noddi.Type = gc.Types[gc.TINT64]
nodsi.Xoffset = c - 8
noddi.Xoffset = c - 8
gmove(&nodsi, &cx)
gmove(&cx, &noddi)
}
}
restx(&cx, &oldcx)
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
gc.Dump("\nclearfat", nl)
}
w := nl.Type.Width
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := w % 8 // bytes
q := w / 8 // quads
if q < 4 {
// Write sequence of MOV 0, off(base) instead of using STOSQ.
// The hope is that although the code will be slightly longer,
// the MOVs will have no dependencies and pipeline better
// than the unrolled STOSQ loop.
// NOTE: Must use agen, not igen, so that optimizer sees address
// being taken. We are not writing on field boundaries.
var n1 gc.Node
gc.Agenr(nl, &n1, nil)
n1.Op = gc.OINDREG
var z gc.Node
gc.Nodconst(&z, gc.Types[gc.TUINT64], 0)
for {
tmp14 := q
q--
if tmp14 <= 0 {
break
}
n1.Type = z.Type
gins(x86.AMOVQ, &z, &n1)
n1.Xoffset += 8
}
if c >= 4 {
gc.Nodconst(&z, gc.Types[gc.TUINT32], 0)
n1.Type = z.Type
gins(x86.AMOVL, &z, &n1)
n1.Xoffset += 4
c -= 4
}
gc.Nodconst(&z, gc.Types[gc.TUINT8], 0)
for {
tmp15 := c
c--
if tmp15 <= 0 {
break
}
n1.Type = z.Type
gins(x86.AMOVB, &z, &n1)
n1.Xoffset++
}
gc.Regfree(&n1)
return
}
var oldn1 gc.Node
var n1 gc.Node
savex(x86.REG_DI, &n1, &oldn1, nil, gc.Types[gc.Tptr])
gc.Agen(nl, &n1)
var ax gc.Node
var oldax gc.Node
savex(x86.REG_AX, &ax, &oldax, nil, gc.Types[gc.Tptr])
gconreg(x86.AMOVL, 0, x86.REG_AX)
if q > 128 || gc.Nacl {
gconreg(movptr, q, x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.ASTOSQ, nil, nil) // STOQ AL,*(DI)+
} else {
if di := dzDI(q); di != 0 {
gconreg(addptr, di, x86.REG_DI)
}
p := gins(obj.ADUFFZERO, nil, nil)
p.To.Type = obj.TYPE_ADDR
p.To.Sym = gc.Linksym(gc.Pkglookup("duffzero", gc.Runtimepkg))
p.To.Offset = dzOff(q)
}
z := ax
di := n1
if w >= 8 && c >= 4 {
di.Op = gc.OINDREG
z.Type = gc.Types[gc.TINT64]
di.Type = z.Type
p := gins(x86.AMOVQ, &z, &di)
p.To.Scale = 1
p.To.Offset = c - 8
} else if c >= 4 {
di.Op = gc.OINDREG
z.Type = gc.Types[gc.TINT32]
di.Type = z.Type
gins(x86.AMOVL, &z, &di)
if c > 4 {
p := gins(x86.AMOVL, &z, &di)
p.To.Scale = 1
p.To.Offset = c - 4
}
} else {
for c > 0 {
gins(x86.ASTOSB, nil, nil) // STOB AL,*(DI)+
c--
}
}
restx(&n1, &oldn1)
restx(&ax, &oldax)
}
/*
* generate division.
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op int, nl *gc.Node, nr *gc.Node, res *gc.Node) {
// Have to be careful about handling
// most negative int divided by -1 correctly.
// The hardware will 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 gc.Issigned[t.Etype] {
check = 1
if gc.Isconst(nl, gc.CTINT) && gc.Mpgetfix(nl.Val.U.Xval) != -(1<<uint64(t.Width*8-1)) {
check = 0
} else if gc.Isconst(nr, gc.CTINT) && gc.Mpgetfix(nr.Val.U.Xval) != -1 {
check = 0
}
}
if t.Width < 8 {
if gc.Issigned[t.Etype] {
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 = ppc64.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.Val.U.Reg
gins(optoas(gc.OMUL, t), &tr, &tm)
gc.Regfree(&tr)
gins(optoas(gc.OSUB, t), &tm, &tl)
gc.Regfree(&tm)
gmove(&tl, res)
}
gc.Regfree(&tl)
if check != 0 {
gc.Patch(p2, gc.Pc)
}
}
/*
* generate:
* newreg = &n;
* res = newreg
*
* on exit, a has been changed to be *newreg.
* caller must regfree(a).
* The generated code checks that the result is not *nil.
*/
func igen(n *gc.Node, a *gc.Node, res *gc.Node) {
if gc.Debug['g'] != 0 {
gc.Dump("\nigen-n", n)
}
switch n.Op {
case gc.ONAME:
if (n.Class&gc.PHEAP != 0) || n.Class == gc.PPARAMREF {
break
}
*a = *n
return
// Increase the refcount of the register so that igen's caller
// has to call regfree.
case gc.OINDREG:
if n.Val.U.Reg != ppc64.REGSP {
reg[n.Val.U.Reg]++
}
*a = *n
return
case gc.ODOT:
igen(n.Left, a, res)
a.Xoffset += n.Xoffset
a.Type = n.Type
fixlargeoffset(a)
return
case gc.ODOTPTR:
cgenr(n.Left, a, res)
gc.Cgen_checknil(a)
a.Op = gc.OINDREG
a.Xoffset += n.Xoffset
a.Type = n.Type
fixlargeoffset(a)
return
case gc.OCALLFUNC,
gc.OCALLMETH,
gc.OCALLINTER:
switch n.Op {
case gc.OCALLFUNC:
cgen_call(n, 0)
case gc.OCALLMETH:
gc.Cgen_callmeth(n, 0)
case gc.OCALLINTER:
cgen_callinter(n, nil, 0)
}
var flist gc.Iter
fp := gc.Structfirst(&flist, gc.Getoutarg(n.Left.Type))
*a = gc.Node{}
a.Op = gc.OINDREG
a.Val.U.Reg = ppc64.REGSP
a.Addable = 1
a.Xoffset = fp.Width + int64(gc.Widthptr) // +widthptr: saved lr at 0(SP)
a.Type = n.Type
return
// Index of fixed-size array by constant can
// put the offset in the addressing.
// Could do the same for slice except that we need
// to use the real index for the bounds checking.
case gc.OINDEX:
if gc.Isfixedarray(n.Left.Type) || (gc.Isptr[n.Left.Type.Etype] && gc.Isfixedarray(n.Left.Left.Type)) {
if gc.Isconst(n.Right, gc.CTINT) {
// Compute &a.
if !gc.Isptr[n.Left.Type.Etype] {
igen(n.Left, a, res)
} else {
var n1 gc.Node
igen(n.Left, &n1, res)
gc.Cgen_checknil(&n1)
regalloc(a, gc.Types[gc.Tptr], res)
gmove(&n1, a)
regfree(&n1)
a.Op = gc.OINDREG
}
// Compute &a[i] as &a + i*width.
a.Type = n.Type
a.Xoffset += gc.Mpgetfix(n.Right.Val.U.Xval) * n.Type.Width
fixlargeoffset(a)
return
}
}
}
agenr(n, a, res)
a.Op = gc.OINDREG
a.Type = n.Type
}
/*
* generate:
* newreg = &n;
* res = newreg
*
* on exit, a has been changed to be *newreg.
* caller must regfree(a).
* The generated code checks that the result is not *nil.
*/
func igen(n *gc.Node, a *gc.Node, res *gc.Node) {
if gc.Debug['g'] != 0 {
gc.Dump("\nigen-n", n)
}
switch n.Op {
case gc.ONAME:
if (n.Class&gc.PHEAP != 0) || n.Class == gc.PPARAMREF {
break
}
*a = *n
return
// Increase the refcount of the register so that igen's caller
// has to call regfree.
case gc.OINDREG:
if n.Val.U.Reg != arm.REGSP {
reg[n.Val.U.Reg]++
}
*a = *n
return
case gc.ODOT:
igen(n.Left, a, res)
a.Xoffset += n.Xoffset
a.Type = n.Type
return
case gc.ODOTPTR:
if n.Left.Addable != 0 || n.Left.Op == gc.OCALLFUNC || n.Left.Op == gc.OCALLMETH || n.Left.Op == gc.OCALLINTER {
// igen-able nodes.
var n1 gc.Node
igen(n.Left, &n1, res)
regalloc(a, gc.Types[gc.Tptr], &n1)
gmove(&n1, a)
regfree(&n1)
} else {
regalloc(a, gc.Types[gc.Tptr], res)
cgen(n.Left, a)
}
gc.Cgen_checknil(a)
a.Op = gc.OINDREG
a.Xoffset = n.Xoffset
a.Type = n.Type
return
// Release res so that it is available for cgen_call.
// Pick it up again after the call.
case gc.OCALLMETH,
gc.OCALLFUNC,
gc.OCALLINTER:
r := -1
if n.Ullman >= gc.UINF {
if res != nil && (res.Op == gc.OREGISTER || res.Op == gc.OINDREG) {
r = int(res.Val.U.Reg)
reg[r]--
}
}
switch n.Op {
case gc.OCALLMETH:
gc.Cgen_callmeth(n, 0)
case gc.OCALLFUNC:
cgen_call(n, 0)
case gc.OCALLINTER:
cgen_callinter(n, nil, 0)
}
if r >= 0 {
reg[r]++
}
regalloc(a, gc.Types[gc.Tptr], res)
cgen_aret(n, a)
a.Op = gc.OINDREG
a.Type = n.Type
return
}
agenr(n, a, res)
a.Op = gc.OINDREG
a.Type = n.Type
}