func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if cnt < int64(4*gc.Widthptr) {
for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
p = appendpp(p, ppc64.AMOVD, obj.TYPE_REG, ppc64.REGZERO, 0, obj.TYPE_MEM, ppc64.REGSP, 8+frame+lo+i)
}
} else if cnt <= int64(128*gc.Widthptr) {
p = appendpp(p, ppc64.AADD, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, ppc64.REGRT1, 0)
p.Reg = ppc64.REGSP
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
f := gc.Sysfunc("duffzero")
gc.Naddr(&p.To, f)
gc.Afunclit(&p.To, f)
p.To.Offset = 4 * (128 - cnt/int64(gc.Widthptr))
} else {
p = appendpp(p, ppc64.AMOVD, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, ppc64.REGTMP, 0)
p = appendpp(p, ppc64.AADD, obj.TYPE_REG, ppc64.REGTMP, 0, obj.TYPE_REG, ppc64.REGRT1, 0)
p.Reg = ppc64.REGSP
p = appendpp(p, ppc64.AMOVD, obj.TYPE_CONST, 0, cnt, obj.TYPE_REG, ppc64.REGTMP, 0)
p = appendpp(p, ppc64.AADD, obj.TYPE_REG, ppc64.REGTMP, 0, obj.TYPE_REG, ppc64.REGRT2, 0)
p.Reg = ppc64.REGRT1
p = appendpp(p, ppc64.AMOVDU, obj.TYPE_REG, ppc64.REGZERO, 0, obj.TYPE_MEM, ppc64.REGRT1, int64(gc.Widthptr))
p1 := p
p = appendpp(p, ppc64.ACMP, obj.TYPE_REG, ppc64.REGRT1, 0, obj.TYPE_REG, ppc64.REGRT2, 0)
p = appendpp(p, ppc64.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
gc.Patch(p, p1)
}
return p
}
/*
* generate floating-point operation.
*/
func cgen_float(n *gc.Node, res *gc.Node) {
nl := n.Left
switch n.Op {
case gc.OEQ,
gc.ONE,
gc.OLT,
gc.OLE,
gc.OGE:
p1 := gc.Gbranch(obj.AJMP, nil, 0)
p2 := gc.Pc
gmove(gc.Nodbool(true), res)
p3 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
gc.Bgen(n, true, 0, p2)
gmove(gc.Nodbool(false), res)
gc.Patch(p3, gc.Pc)
return
case gc.OPLUS:
gc.Cgen(nl, res)
return
case gc.OCONV:
if gc.Eqtype(n.Type, nl.Type) || gc.Noconv(n.Type, nl.Type) {
gc.Cgen(nl, res)
return
}
var n2 gc.Node
gc.Tempname(&n2, n.Type)
var n1 gc.Node
gc.Mgen(nl, &n1, res)
gmove(&n1, &n2)
gmove(&n2, res)
gc.Mfree(&n1)
return
}
if gc.Thearch.Use387 {
cgen_float387(n, res)
} else {
cgen_floatsse(n, res)
}
}
func gencmp0(n *gc.Node, t *gc.Type, o int, likely int, to *obj.Prog) {
var n1 gc.Node
gc.Regalloc(&n1, t, nil)
gc.Cgen(n, &n1)
a := optoas(gc.OCMP, t)
if a != arm.ACMP {
var n2 gc.Node
gc.Nodconst(&n2, t, 0)
var n3 gc.Node
gc.Regalloc(&n3, t, nil)
gmove(&n2, &n3)
gins(a, &n1, &n3)
gc.Regfree(&n3)
} else {
gins(arm.ATST, &n1, nil)
}
a = optoas(o, t)
gc.Patch(gc.Gbranch(a, t, likely), to)
gc.Regfree(&n1)
}
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64, r0 *uint32) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if *r0 == 0 {
p = appendpp(p, arm.AMOVW, obj.TYPE_CONST, 0, 0, obj.TYPE_REG, arm.REG_R0, 0)
*r0 = 1
}
if cnt < int64(4*gc.Widthptr) {
for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
p = appendpp(p, arm.AMOVW, obj.TYPE_REG, arm.REG_R0, 0, obj.TYPE_MEM, arm.REGSP, int32(4+frame+lo+i))
}
} else if !gc.Nacl && (cnt <= int64(128*gc.Widthptr)) {
p = appendpp(p, arm.AADD, obj.TYPE_CONST, 0, int32(4+frame+lo), obj.TYPE_REG, arm.REG_R1, 0)
p.Reg = arm.REGSP
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
f := gc.Sysfunc("duffzero")
gc.Naddr(&p.To, f)
gc.Afunclit(&p.To, f)
p.To.Offset = 4 * (128 - cnt/int64(gc.Widthptr))
} else {
p = appendpp(p, arm.AADD, obj.TYPE_CONST, 0, int32(4+frame+lo), obj.TYPE_REG, arm.REG_R1, 0)
p.Reg = arm.REGSP
p = appendpp(p, arm.AADD, obj.TYPE_CONST, 0, int32(cnt), obj.TYPE_REG, arm.REG_R2, 0)
p.Reg = arm.REG_R1
p = appendpp(p, arm.AMOVW, obj.TYPE_REG, arm.REG_R0, 0, obj.TYPE_MEM, arm.REG_R1, 4)
p1 := p
p.Scond |= arm.C_PBIT
p = appendpp(p, arm.ACMP, obj.TYPE_REG, arm.REG_R1, 0, obj.TYPE_NONE, 0, 0)
p.Reg = arm.REG_R2
p = appendpp(p, arm.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
gc.Patch(p, p1)
}
return p
}
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if cnt < int64(4*gc.Widthptr) {
for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGSP, 8+frame+lo+i)
}
} else if cnt <= int64(128*gc.Widthptr) && !darwin { // darwin ld64 cannot handle BR26 reloc with non-zero addend
p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGSP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
p = appendpp(p, arm64.AADD, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, arm64.REGRT1, 0)
p.Reg = arm64.REGRT1
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
f := gc.Sysfunc("duffzero")
gc.Naddr(&p.To, f)
gc.Afunclit(&p.To, f)
p.To.Offset = 4 * (128 - cnt/int64(gc.Widthptr))
} else {
p = appendpp(p, arm64.AMOVD, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, arm64.REGTMP, 0)
p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGSP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
p = appendpp(p, arm64.AADD, obj.TYPE_REG, arm64.REGTMP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
p.Reg = arm64.REGRT1
p = appendpp(p, arm64.AMOVD, obj.TYPE_CONST, 0, cnt, obj.TYPE_REG, arm64.REGTMP, 0)
p = appendpp(p, arm64.AADD, obj.TYPE_REG, arm64.REGTMP, 0, obj.TYPE_REG, arm64.REGRT2, 0)
p.Reg = arm64.REGRT1
p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGRT1, int64(gc.Widthptr))
p.Scond = arm64.C_XPRE
p1 := p
p = appendpp(p, arm64.ACMP, obj.TYPE_REG, arm64.REGRT1, 0, obj.TYPE_NONE, 0, 0)
p.Reg = arm64.REGRT2
p = appendpp(p, arm64.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
gc.Patch(p, p1)
}
return p
}
/*
* block copy:
* memmove(&ns, &n, w);
*/
func sgen(n *gc.Node, ns *gc.Node, w int64) {
var res *gc.Node = ns
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 dst gc.Node
regalloc(&dst, gc.Types[gc.Tptr], nil)
agen(res, &dst)
agen(n, &dst)
regfree(&dst)
return
}
// determine alignment.
// want to avoid unaligned access, so have to use
// smaller operations for less aligned types.
// for example moving [4]byte must use 4 MOVB not 1 MOVW.
align := int(n.Type.Align)
var op int
switch align {
default:
gc.Fatal("sgen: invalid alignment %d for %v", align, gc.Tconv(n.Type, 0))
case 1:
op = ppc64.AMOVBU
case 2:
op = ppc64.AMOVHU
case 4:
op = ppc64.AMOVWZU // there is no lwau, only lwaux
case 8:
op = ppc64.AMOVDU
}
if w%int64(align) != 0 {
gc.Fatal("sgen: unaligned size %d (align=%d) for %v", w, align, gc.Tconv(n.Type, 0))
}
c := int32(w / int64(align))
// offset on the stack
osrc := int32(stkof(n))
odst := int32(stkof(res))
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, res, w)
return
}
if osrc%int32(align) != 0 || odst%int32(align) != 0 {
gc.Fatal("sgen: unaligned offset src %d or dst %d (align %d)", osrc, odst, align)
}
// if we are copying forward on the stack and
// the src and dst overlap, then reverse direction
dir := align
if osrc < odst && int64(odst) < int64(osrc)+w {
dir = -dir
}
var dst gc.Node
var src gc.Node
if n.Ullman >= res.Ullman {
agenr(n, &dst, res) // temporarily use dst
regalloc(&src, gc.Types[gc.Tptr], nil)
gins(ppc64.AMOVD, &dst, &src)
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
agen(res, &dst)
} else {
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
agenr(res, &dst, res)
agenr(n, &src, nil)
}
var tmp gc.Node
regalloc(&tmp, gc.Types[gc.Tptr], nil)
// set up end marker
var nend gc.Node
// move src and dest to the end of block if necessary
if dir < 0 {
if c >= 4 {
regalloc(&nend, gc.Types[gc.Tptr], nil)
gins(ppc64.AMOVD, &src, &nend)
}
p := gins(ppc64.AADD, nil, &src)
p.From.Type = obj.TYPE_CONST
p.From.Offset = w
p = gins(ppc64.AADD, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = w
} else {
p := gins(ppc64.AADD, nil, &src)
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(-dir)
p = gins(ppc64.AADD, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(-dir)
if c >= 4 {
regalloc(&nend, gc.Types[gc.Tptr], nil)
p := gins(ppc64.AMOVD, &src, &nend)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w
}
}
// move
// TODO: enable duffcopy for larger copies.
if c >= 4 {
p := gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
ploop := p
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
p = gins(ppc64.ACMP, &src, &nend)
gc.Patch(gc.Gbranch(ppc64.ABNE, nil, 0), ploop)
regfree(&nend)
} else {
// TODO(austin): Instead of generating ADD $-8,R8; ADD
// $-8,R7; n*(MOVDU 8(R8),R9; MOVDU R9,8(R7);) just
// generate the offsets directly and eliminate the
// ADDs. That will produce shorter, more
// pipeline-able code.
var p *obj.Prog
for {
tmp14 := c
c--
if tmp14 <= 0 {
break
}
p = gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
}
}
regfree(&dst)
regfree(&src)
regfree(&tmp)
}
/*
* 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
}
/*
* generate move:
* t = f
* hard part is conversions.
*/
func gmove(f *gc.Node, t *gc.Node) {
if gc.Debug['M'] != 0 {
fmt.Printf("gmove %v -> %v\n", gc.Nconv(f, obj.FmtLong), gc.Nconv(t, obj.FmtLong))
}
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
cvt := t.Type
if gc.Iscomplex[ft] || gc.Iscomplex[tt] {
gc.Complexmove(f, t)
return
}
// cannot have two memory operands
var a int
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
f.Convconst(&con, t.Type)
f = &con
ft = tt // so big switch will choose a simple mov
// some constants can't move directly to memory.
if gc.Ismem(t) {
// float constants come from memory.
if gc.Isfloat[tt] {
goto hard
}
// 64-bit immediates are really 32-bit sign-extended
// unless moving into a register.
if gc.Isint[tt] {
if i := con.Int(); int64(int32(i)) != i {
goto hard
}
}
}
}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
gc.Fatal("gmove %v -> %v", gc.Tconv(f.Type, obj.FmtLong), gc.Tconv(t.Type, obj.FmtLong))
/*
* integer copy and truncate
*/
case gc.TINT8<<16 | gc.TINT8, // same size
gc.TINT8<<16 | gc.TUINT8,
gc.TUINT8<<16 | gc.TINT8,
gc.TUINT8<<16 | gc.TUINT8,
gc.TINT16<<16 | gc.TINT8,
// truncate
gc.TUINT16<<16 | gc.TINT8,
gc.TINT32<<16 | gc.TINT8,
gc.TUINT32<<16 | gc.TINT8,
gc.TINT64<<16 | gc.TINT8,
gc.TUINT64<<16 | gc.TINT8,
gc.TINT16<<16 | gc.TUINT8,
gc.TUINT16<<16 | gc.TUINT8,
gc.TINT32<<16 | gc.TUINT8,
gc.TUINT32<<16 | gc.TUINT8,
gc.TINT64<<16 | gc.TUINT8,
gc.TUINT64<<16 | gc.TUINT8:
a = x86.AMOVB
case gc.TINT16<<16 | gc.TINT16, // same size
gc.TINT16<<16 | gc.TUINT16,
gc.TUINT16<<16 | gc.TINT16,
gc.TUINT16<<16 | gc.TUINT16,
gc.TINT32<<16 | gc.TINT16,
// truncate
gc.TUINT32<<16 | gc.TINT16,
gc.TINT64<<16 | gc.TINT16,
gc.TUINT64<<16 | gc.TINT16,
gc.TINT32<<16 | gc.TUINT16,
gc.TUINT32<<16 | gc.TUINT16,
gc.TINT64<<16 | gc.TUINT16,
gc.TUINT64<<16 | gc.TUINT16:
a = x86.AMOVW
case gc.TINT32<<16 | gc.TINT32, // same size
gc.TINT32<<16 | gc.TUINT32,
gc.TUINT32<<16 | gc.TINT32,
gc.TUINT32<<16 | gc.TUINT32:
a = x86.AMOVL
case gc.TINT64<<16 | gc.TINT32, // truncate
gc.TUINT64<<16 | gc.TINT32,
gc.TINT64<<16 | gc.TUINT32,
gc.TUINT64<<16 | gc.TUINT32:
a = x86.AMOVQL
case gc.TINT64<<16 | gc.TINT64, // same size
gc.TINT64<<16 | gc.TUINT64,
gc.TUINT64<<16 | gc.TINT64,
gc.TUINT64<<16 | gc.TUINT64:
a = x86.AMOVQ
/*
* integer up-conversions
*/
case gc.TINT8<<16 | gc.TINT16, // sign extend int8
gc.TINT8<<16 | gc.TUINT16:
a = x86.AMOVBWSX
goto rdst
case gc.TINT8<<16 | gc.TINT32,
gc.TINT8<<16 | gc.TUINT32:
a = x86.AMOVBLSX
goto rdst
case gc.TINT8<<16 | gc.TINT64,
gc.TINT8<<16 | gc.TUINT64:
a = x86.AMOVBQSX
goto rdst
case gc.TUINT8<<16 | gc.TINT16, // zero extend uint8
gc.TUINT8<<16 | gc.TUINT16:
a = x86.AMOVBWZX
goto rdst
case gc.TUINT8<<16 | gc.TINT32,
gc.TUINT8<<16 | gc.TUINT32:
a = x86.AMOVBLZX
goto rdst
case gc.TUINT8<<16 | gc.TINT64,
gc.TUINT8<<16 | gc.TUINT64:
a = x86.AMOVBQZX
goto rdst
case gc.TINT16<<16 | gc.TINT32, // sign extend int16
gc.TINT16<<16 | gc.TUINT32:
a = x86.AMOVWLSX
goto rdst
case gc.TINT16<<16 | gc.TINT64,
gc.TINT16<<16 | gc.TUINT64:
a = x86.AMOVWQSX
goto rdst
case gc.TUINT16<<16 | gc.TINT32, // zero extend uint16
gc.TUINT16<<16 | gc.TUINT32:
a = x86.AMOVWLZX
goto rdst
case gc.TUINT16<<16 | gc.TINT64,
gc.TUINT16<<16 | gc.TUINT64:
a = x86.AMOVWQZX
goto rdst
case gc.TINT32<<16 | gc.TINT64, // sign extend int32
gc.TINT32<<16 | gc.TUINT64:
a = x86.AMOVLQSX
goto rdst
// AMOVL into a register zeros the top of the register,
// so this is not always necessary, but if we rely on AMOVL
// the optimizer is almost certain to screw with us.
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
a = x86.AMOVLQZX
goto rdst
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT32:
a = x86.ACVTTSS2SL
goto rdst
case gc.TFLOAT64<<16 | gc.TINT32:
a = x86.ACVTTSD2SL
goto rdst
case gc.TFLOAT32<<16 | gc.TINT64:
a = x86.ACVTTSS2SQ
goto rdst
case gc.TFLOAT64<<16 | gc.TINT64:
a = x86.ACVTTSD2SQ
goto rdst
// convert via int32.
case gc.TFLOAT32<<16 | gc.TINT16,
gc.TFLOAT32<<16 | gc.TINT8,
gc.TFLOAT32<<16 | gc.TUINT16,
gc.TFLOAT32<<16 | gc.TUINT8,
gc.TFLOAT64<<16 | gc.TINT16,
gc.TFLOAT64<<16 | gc.TINT8,
gc.TFLOAT64<<16 | gc.TUINT16,
gc.TFLOAT64<<16 | gc.TUINT8:
cvt = gc.Types[gc.TINT32]
goto hard
// convert via int64.
case gc.TFLOAT32<<16 | gc.TUINT32,
gc.TFLOAT64<<16 | gc.TUINT32:
cvt = gc.Types[gc.TINT64]
goto hard
// algorithm is:
// if small enough, use native float64 -> int64 conversion.
// otherwise, subtract 2^63, convert, and add it back.
case gc.TFLOAT32<<16 | gc.TUINT64,
gc.TFLOAT64<<16 | gc.TUINT64:
a := x86.ACVTTSS2SQ
if ft == gc.TFLOAT64 {
a = x86.ACVTTSD2SQ
}
bignodes()
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[ft], nil)
var r2 gc.Node
gc.Regalloc(&r2, gc.Types[tt], t)
var r3 gc.Node
gc.Regalloc(&r3, gc.Types[ft], nil)
var r4 gc.Node
gc.Regalloc(&r4, gc.Types[tt], nil)
gins(optoas(gc.OAS, f.Type), f, &r1)
gins(optoas(gc.OCMP, f.Type), &bigf, &r1)
p1 := gc.Gbranch(optoas(gc.OLE, f.Type), nil, +1)
gins(a, &r1, &r2)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
gins(optoas(gc.OAS, f.Type), &bigf, &r3)
gins(optoas(gc.OSUB, f.Type), &r3, &r1)
gins(a, &r1, &r2)
gins(x86.AMOVQ, &bigi, &r4)
gins(x86.AXORQ, &r4, &r2)
gc.Patch(p2, gc.Pc)
gmove(&r2, t)
gc.Regfree(&r4)
gc.Regfree(&r3)
gc.Regfree(&r2)
gc.Regfree(&r1)
return
/*
* integer to float
*/
case gc.TINT32<<16 | gc.TFLOAT32:
a = x86.ACVTSL2SS
goto rdst
case gc.TINT32<<16 | gc.TFLOAT64:
a = x86.ACVTSL2SD
goto rdst
case gc.TINT64<<16 | gc.TFLOAT32:
a = x86.ACVTSQ2SS
goto rdst
case gc.TINT64<<16 | gc.TFLOAT64:
a = x86.ACVTSQ2SD
goto rdst
// convert via int32
case gc.TINT16<<16 | gc.TFLOAT32,
gc.TINT16<<16 | gc.TFLOAT64,
gc.TINT8<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TUINT16<<16 | gc.TFLOAT32,
gc.TUINT16<<16 | gc.TFLOAT64,
gc.TUINT8<<16 | gc.TFLOAT32,
gc.TUINT8<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT32]
goto hard
// convert via int64.
case gc.TUINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT64]
goto hard
// algorithm is:
// if small enough, use native int64 -> uint64 conversion.
// otherwise, halve (rounding to odd?), convert, and double.
case gc.TUINT64<<16 | gc.TFLOAT32,
gc.TUINT64<<16 | gc.TFLOAT64:
a := x86.ACVTSQ2SS
if tt == gc.TFLOAT64 {
a = x86.ACVTSQ2SD
}
var zero gc.Node
gc.Nodconst(&zero, gc.Types[gc.TUINT64], 0)
var one gc.Node
gc.Nodconst(&one, gc.Types[gc.TUINT64], 1)
var r1 gc.Node
gc.Regalloc(&r1, f.Type, f)
var r2 gc.Node
gc.Regalloc(&r2, t.Type, t)
var r3 gc.Node
gc.Regalloc(&r3, f.Type, nil)
var r4 gc.Node
gc.Regalloc(&r4, f.Type, nil)
gmove(f, &r1)
gins(x86.ACMPQ, &r1, &zero)
p1 := gc.Gbranch(x86.AJLT, nil, +1)
gins(a, &r1, &r2)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
gmove(&r1, &r3)
gins(x86.ASHRQ, &one, &r3)
gmove(&r1, &r4)
gins(x86.AANDL, &one, &r4)
gins(x86.AORQ, &r4, &r3)
gins(a, &r3, &r2)
gins(optoas(gc.OADD, t.Type), &r2, &r2)
gc.Patch(p2, gc.Pc)
gmove(&r2, t)
gc.Regfree(&r4)
gc.Regfree(&r3)
gc.Regfree(&r2)
gc.Regfree(&r1)
return
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32:
a = x86.AMOVSS
case gc.TFLOAT64<<16 | gc.TFLOAT64:
a = x86.AMOVSD
case gc.TFLOAT32<<16 | gc.TFLOAT64:
a = x86.ACVTSS2SD
goto rdst
case gc.TFLOAT64<<16 | gc.TFLOAT32:
a = x86.ACVTSD2SS
goto rdst
}
gins(a, f, t)
return
// requires register destination
rdst:
{
var r1 gc.Node
gc.Regalloc(&r1, t.Type, t)
gins(a, f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
// requires register intermediate
hard:
var r1 gc.Node
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
/*
* 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
}
/*
* generate:
* call f
* proc=-1 normal call but no return
* proc=0 normal call
* proc=1 goroutine run in new proc
* proc=2 defer call save away stack
* proc=3 normal call to C pointer (not Go func value)
*/
func ginscall(f *gc.Node, proc int) {
if f.Type != nil {
extra := int32(0)
if proc == 1 || proc == 2 {
extra = 2 * int32(gc.Widthptr)
}
gc.Setmaxarg(f.Type, extra)
}
switch proc {
default:
gc.Fatal("ginscall: bad proc %d", proc)
case 0, // normal call
-1: // normal call but no return
if f.Op == gc.ONAME && f.Class == gc.PFUNC {
if f == gc.Deferreturn {
// Deferred calls will appear to be returning to
// the CALL deferreturn(SB) that we are about to emit.
// However, the stack trace code will show the line
// of the instruction byte before the return PC.
// To avoid that being an unrelated instruction,
// insert a ppc64 NOP that we will have the right line number.
// The ppc64 NOP is really or r0, r0, r0; use that description
// because the NOP pseudo-instruction would be removed by
// the linker.
var reg gc.Node
gc.Nodreg(®, gc.Types[gc.TINT], ppc64.REG_R0)
gins(ppc64.AOR, ®, ®)
}
p := gins(ppc64.ABL, nil, f)
gc.Afunclit(&p.To, f)
if proc == -1 || gc.Noreturn(p) {
gins(obj.AUNDEF, nil, nil)
}
break
}
var reg gc.Node
gc.Nodreg(®, gc.Types[gc.Tptr], ppc64.REGCTXT)
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[gc.Tptr], ppc64.REG_R3)
gmove(f, ®)
reg.Op = gc.OINDREG
gmove(®, &r1)
reg.Op = gc.OREGISTER
ginsBL(®, &r1)
case 3: // normal call of c function pointer
ginsBL(nil, f)
case 1, // call in new proc (go)
2: // deferred call (defer)
var con gc.Node
gc.Nodconst(&con, gc.Types[gc.TINT64], int64(gc.Argsize(f.Type)))
var reg gc.Node
gc.Nodreg(®, gc.Types[gc.TINT64], ppc64.REG_R3)
var reg2 gc.Node
gc.Nodreg(®2, gc.Types[gc.TINT64], ppc64.REG_R4)
gmove(f, ®)
gmove(&con, ®2)
p := gins(ppc64.AMOVW, ®2, nil)
p.To.Type = obj.TYPE_MEM
p.To.Reg = ppc64.REGSP
p.To.Offset = 8
p = gins(ppc64.AMOVD, ®, nil)
p.To.Type = obj.TYPE_MEM
p.To.Reg = ppc64.REGSP
p.To.Offset = 16
if proc == 1 {
ginscall(gc.Newproc, 0)
} else {
if gc.Hasdefer == 0 {
gc.Fatal("hasdefer=0 but has defer")
}
ginscall(gc.Deferproc, 0)
}
if proc == 2 {
gc.Nodreg(®, gc.Types[gc.TINT64], ppc64.REG_R3)
p := gins(ppc64.ACMP, ®, nil)
p.To.Type = obj.TYPE_REG
p.To.Reg = ppc64.REGZERO
p = gc.Gbranch(ppc64.ABEQ, nil, +1)
cgen_ret(nil)
gc.Patch(p, gc.Pc)
}
}
}
/*
* generate:
* call f
* proc=-1 normal call but no return
* proc=0 normal call
* proc=1 goroutine run in new proc
* proc=2 defer call save away stack
* proc=3 normal call to C pointer (not Go func value)
*/
func ginscall(f *gc.Node, proc int) {
if f.Type != nil {
extra := int32(0)
if proc == 1 || proc == 2 {
extra = 2 * int32(gc.Widthptr)
}
gc.Setmaxarg(f.Type, extra)
}
switch proc {
default:
gc.Fatal("ginscall: bad proc %d", proc)
case 0, // normal call
-1: // normal call but no return
if f.Op == gc.ONAME && f.Class == gc.PFUNC {
if f == gc.Deferreturn {
// Deferred calls will appear to be returning to
// the BL deferreturn(SB) that we are about to emit.
// However, the stack trace code will show the line
// of the instruction before that return PC.
// To avoid that instruction being an unrelated instruction,
// insert a NOP so that we will have the right line number.
// ARM NOP 0x00000000 is really AND.EQ R0, R0, R0.
// Use the latter form because the NOP pseudo-instruction
// would be removed by the linker.
var r gc.Node
gc.Nodreg(&r, gc.Types[gc.TINT], arm.REG_R0)
p := gins(arm.AAND, &r, &r)
p.Scond = arm.C_SCOND_EQ
}
p := gins(arm.ABL, nil, f)
gc.Afunclit(&p.To, f)
if proc == -1 || gc.Noreturn(p) {
gins(obj.AUNDEF, nil, nil)
}
break
}
var r gc.Node
gc.Nodreg(&r, gc.Types[gc.Tptr], arm.REG_R7)
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[gc.Tptr], arm.REG_R1)
gmove(f, &r)
r.Op = gc.OINDREG
gmove(&r, &r1)
r.Op = gc.OREGISTER
r1.Op = gc.OINDREG
gins(arm.ABL, &r, &r1)
case 3: // normal call of c function pointer
gins(arm.ABL, nil, f)
case 1, // call in new proc (go)
2: // deferred call (defer)
var r gc.Node
regalloc(&r, gc.Types[gc.Tptr], nil)
var con gc.Node
gc.Nodconst(&con, gc.Types[gc.TINT32], int64(gc.Argsize(f.Type)))
gins(arm.AMOVW, &con, &r)
p := gins(arm.AMOVW, &r, nil)
p.To.Type = obj.TYPE_MEM
p.To.Reg = arm.REGSP
p.To.Offset = 4
gins(arm.AMOVW, f, &r)
p = gins(arm.AMOVW, &r, nil)
p.To.Type = obj.TYPE_MEM
p.To.Reg = arm.REGSP
p.To.Offset = 8
regfree(&r)
if proc == 1 {
ginscall(gc.Newproc, 0)
} else {
ginscall(gc.Deferproc, 0)
}
if proc == 2 {
gc.Nodconst(&con, gc.Types[gc.TINT32], 0)
p := gins(arm.ACMP, &con, nil)
p.Reg = arm.REG_R0
p = gc.Gbranch(arm.ABEQ, nil, +1)
cgen_ret(nil)
gc.Patch(p, gc.Pc)
}
}
}
/*
* generate division.
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op int, nl *gc.Node, nr *gc.Node, res *gc.Node) {
// Have to be careful about handling
// most negative int divided by -1 correctly.
// The hardware will trap.
// Also the byte divide instruction needs AH,
// which we otherwise don't have to deal with.
// Easiest way to avoid for int8, int16: use int32.
// For int32 and int64, use explicit test.
// Could use int64 hw for int32.
t := nl.Type
t0 := t
check := 0
if gc.Issigned[t.Etype] {
check = 1
if gc.Isconst(nl, gc.CTINT) && gc.Mpgetfix(nl.Val.U.Xval) != -(1<<uint64(t.Width*8-1)) {
check = 0
} else if gc.Isconst(nr, gc.CTINT) && gc.Mpgetfix(nr.Val.U.Xval) != -1 {
check = 0
}
}
if t.Width < 4 {
if gc.Issigned[t.Etype] {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
check = 0
}
a := optoas(op, t)
var n3 gc.Node
gc.Regalloc(&n3, t0, nil)
var ax gc.Node
var oldax gc.Node
if nl.Ullman >= nr.Ullman {
savex(x86.REG_AX, &ax, &oldax, res, t0)
gc.Cgen(nl, &ax)
gc.Regalloc(&ax, t0, &ax) // mark ax live during cgen
gc.Cgen(nr, &n3)
gc.Regfree(&ax)
} else {
gc.Cgen(nr, &n3)
savex(x86.REG_AX, &ax, &oldax, res, t0)
gc.Cgen(nl, &ax)
}
if t != t0 {
// Convert
ax1 := ax
n31 := n3
ax.Type = t
n3.Type = t
gmove(&ax1, &ax)
gmove(&n31, &n3)
}
var n4 gc.Node
if gc.Nacl {
// Native Client does not relay the divide-by-zero trap
// to the executing program, so we must insert a check
// for ourselves.
gc.Nodconst(&n4, t, 0)
gins(optoas(gc.OCMP, t), &n3, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if panicdiv == nil {
panicdiv = gc.Sysfunc("panicdivide")
}
gc.Ginscall(panicdiv, -1)
gc.Patch(p1, gc.Pc)
}
var p2 *obj.Prog
if check != 0 {
gc.Nodconst(&n4, t, -1)
gins(optoas(gc.OCMP, t), &n3, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if op == gc.ODIV {
// a / (-1) is -a.
gins(optoas(gc.OMINUS, t), nil, &ax)
gmove(&ax, res)
} else {
// a % (-1) is 0.
gc.Nodconst(&n4, t, 0)
gmove(&n4, res)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
}
var olddx gc.Node
var dx gc.Node
savex(x86.REG_DX, &dx, &olddx, res, t)
if !gc.Issigned[t.Etype] {
gc.Nodconst(&n4, t, 0)
gmove(&n4, &dx)
} else {
gins(optoas(gc.OEXTEND, t), nil, nil)
}
gins(a, &n3, nil)
gc.Regfree(&n3)
if op == gc.ODIV {
gmove(&ax, res)
} else {
gmove(&dx, res)
}
restx(&dx, &olddx)
if check != 0 {
gc.Patch(p2, gc.Pc)
}
restx(&ax, &oldax)
}
/*
* generate division.
* caller must set:
* ax = allocated AX register
* dx = allocated DX register
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op int, nl *gc.Node, nr *gc.Node, res *gc.Node, ax *gc.Node, dx *gc.Node) {
// Have to be careful about handling
// most negative int divided by -1 correctly.
// The hardware will trap.
// Also the byte divide instruction needs AH,
// which we otherwise don't have to deal with.
// Easiest way to avoid for int8, int16: use int32.
// For int32 and int64, use explicit test.
// Could use int64 hw for int32.
t := nl.Type
t0 := t
check := 0
if gc.Issigned[t.Etype] {
check = 1
if gc.Isconst(nl, gc.CTINT) && gc.Mpgetfix(nl.Val.U.Xval) != -1<<uint64(t.Width*8-1) {
check = 0
} else if gc.Isconst(nr, gc.CTINT) && gc.Mpgetfix(nr.Val.U.Xval) != -1 {
check = 0
}
}
if t.Width < 4 {
if gc.Issigned[t.Etype] {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
check = 0
}
var t1 gc.Node
gc.Tempname(&t1, t)
var t2 gc.Node
gc.Tempname(&t2, t)
if t0 != t {
var t3 gc.Node
gc.Tempname(&t3, t0)
var t4 gc.Node
gc.Tempname(&t4, t0)
gc.Cgen(nl, &t3)
gc.Cgen(nr, &t4)
// Convert.
gmove(&t3, &t1)
gmove(&t4, &t2)
} else {
gc.Cgen(nl, &t1)
gc.Cgen(nr, &t2)
}
var n1 gc.Node
if !gc.Samereg(ax, res) && !gc.Samereg(dx, res) {
gc.Regalloc(&n1, t, res)
} else {
gc.Regalloc(&n1, t, nil)
}
gmove(&t2, &n1)
gmove(&t1, ax)
var p2 *obj.Prog
var n4 gc.Node
if gc.Nacl {
// Native Client does not relay the divide-by-zero trap
// to the executing program, so we must insert a check
// for ourselves.
gc.Nodconst(&n4, t, 0)
gins(optoas(gc.OCMP, t), &n1, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if panicdiv == nil {
panicdiv = gc.Sysfunc("panicdivide")
}
gc.Ginscall(panicdiv, -1)
gc.Patch(p1, gc.Pc)
}
if check != 0 {
gc.Nodconst(&n4, t, -1)
gins(optoas(gc.OCMP, t), &n1, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if op == gc.ODIV {
// a / (-1) is -a.
gins(optoas(gc.OMINUS, t), nil, ax)
gmove(ax, res)
} else {
// a % (-1) is 0.
gc.Nodconst(&n4, t, 0)
gmove(&n4, res)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
}
if !gc.Issigned[t.Etype] {
var nz gc.Node
gc.Nodconst(&nz, t, 0)
gmove(&nz, dx)
} else {
gins(optoas(gc.OEXTEND, t), nil, nil)
}
gins(optoas(op, t), &n1, nil)
gc.Regfree(&n1)
if op == gc.ODIV {
gmove(ax, res)
} else {
gmove(dx, res)
}
if check != 0 {
gc.Patch(p2, gc.Pc)
}
}
/*
* 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)
}
}
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.Reg != x86.REG_F0 || t.Reg != x86.REG_F0 {
goto fatal
}
return
}
a = x86.AFMOVF
if ft == gc.TFLOAT64 {
a = x86.AFMOVD
}
if gc.Ismem(t) {
if f.Op != gc.OREGISTER || f.Reg != x86.REG_F0 {
gc.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.Reg != x86.REG_F0 || t.Reg != x86.REG_F0 {
goto fatal
}
return
}
if f.Op == gc.OREGISTER {
gins(x86.AFMOVDP, f, t)
} else {
gins(x86.AFMOVF, f, t)
}
return
case gc.TFLOAT64<<16 | gc.TFLOAT32:
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
if f.Op == gc.OREGISTER && t.Op == gc.OREGISTER {
var r1 gc.Node
gc.Tempname(&r1, gc.Types[gc.TFLOAT32])
gins(x86.AFMOVFP, f, &r1)
gins(x86.AFMOVF, &r1, t)
return
}
if f.Op == gc.OREGISTER {
gins(x86.AFMOVFP, f, t)
} else {
gins(x86.AFMOVD, f, t)
}
return
}
gins(a, f, t)
return
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
// requires memory intermediate
hardmem:
gc.Tempname(&r1, cvt)
gmove(f, &r1)
gmove(&r1, t)
return
// should not happen
fatal:
gc.Fatal("gmove %v -> %v", gc.Nconv(f, obj.FmtLong), gc.Nconv(t, obj.FmtLong))
return
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
gc.Dump("\nclearfat", nl)
}
w := uint32(nl.Type.Width)
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := w % 4 // bytes
q := w / 4 // quads
var r0 gc.Node
r0.Op = gc.OREGISTER
r0.Reg = arm.REG_R0
var r1 gc.Node
r1.Op = gc.OREGISTER
r1.Reg = arm.REG_R1
var dst gc.Node
gc.Regalloc(&dst, gc.Types[gc.Tptr], &r1)
gc.Agen(nl, &dst)
var nc gc.Node
gc.Nodconst(&nc, gc.Types[gc.TUINT32], 0)
var nz gc.Node
gc.Regalloc(&nz, gc.Types[gc.TUINT32], &r0)
gc.Cgen(&nc, &nz)
if q > 128 {
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p := gins(arm.AMOVW, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(q) * 4
p = gins(arm.AMOVW, &nz, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 4
p.Scond |= arm.C_PBIT
pl := p
p = gins(arm.ACMP, &dst, nil)
raddr(&end, p)
gc.Patch(gc.Gbranch(arm.ABNE, nil, 0), pl)
gc.Regfree(&end)
} else if q >= 4 && !gc.Nacl {
f := gc.Sysfunc("duffzero")
p := gins(obj.ADUFFZERO, nil, f)
gc.Afunclit(&p.To, f)
// 4 and 128 = magic constants: see ../../runtime/asm_arm.s
p.To.Offset = 4 * (128 - int64(q))
} else {
var p *obj.Prog
for q > 0 {
p = gins(arm.AMOVW, &nz, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 4
p.Scond |= arm.C_PBIT
//print("1. %P\n", p);
q--
}
}
var p *obj.Prog
for c > 0 {
p = gins(arm.AMOVB, &nz, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 1
p.Scond |= arm.C_PBIT
//print("2. %P\n", p);
c--
}
gc.Regfree(&dst)
gc.Regfree(&nz)
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op int, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Type.Width > 4 {
gc.Fatal("cgen_shift %v", nl.Type)
}
w := int(nl.Type.Width * 8)
if op == gc.OLROT {
v := nr.Int()
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
if w == 32 {
gc.Cgen(nl, &n1)
gshift(arm.AMOVW, &n1, arm.SHIFT_RR, int32(w)-int32(v), &n1)
} else {
var n2 gc.Node
gc.Regalloc(&n2, nl.Type, nil)
gc.Cgen(nl, &n2)
gshift(arm.AMOVW, &n2, arm.SHIFT_LL, int32(v), &n1)
gshift(arm.AORR, &n2, arm.SHIFT_LR, int32(w)-int32(v), &n1)
gc.Regfree(&n2)
// Ensure sign/zero-extended result.
gins(optoas(gc.OAS, nl.Type), &n1, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
if nr.Op == gc.OLITERAL {
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gc.Cgen(nl, &n1)
sc := uint64(nr.Int())
if sc == 0 {
} else // nothing to do
if sc >= uint64(nl.Type.Width*8) {
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gshift(arm.AMOVW, &n1, arm.SHIFT_AR, int32(w), &n1)
} else {
gins(arm.AEOR, &n1, &n1)
}
} else {
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gshift(arm.AMOVW, &n1, arm.SHIFT_AR, int32(sc), &n1)
} else if op == gc.ORSH {
gshift(arm.AMOVW, &n1, arm.SHIFT_LR, int32(sc), &n1) // OLSH
} else {
gshift(arm.AMOVW, &n1, arm.SHIFT_LL, int32(sc), &n1)
}
}
if w < 32 && op == gc.OLSH {
gins(optoas(gc.OAS, nl.Type), &n1, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
tr := nr.Type
var t gc.Node
var n1 gc.Node
var n2 gc.Node
var n3 gc.Node
if tr.Width > 4 {
var nt gc.Node
gc.Tempname(&nt, nr.Type)
if nl.Ullman >= nr.Ullman {
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
gc.Cgen(nr, &nt)
n1 = nt
} else {
gc.Cgen(nr, &nt)
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
}
var hi gc.Node
var lo gc.Node
split64(&nt, &lo, &hi)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], nil)
gc.Regalloc(&n3, gc.Types[gc.TUINT32], nil)
gmove(&lo, &n1)
gmove(&hi, &n3)
splitclean()
gins(arm.ATST, &n3, nil)
gc.Nodconst(&t, gc.Types[gc.TUINT32], int64(w))
p1 := gins(arm.AMOVW, &t, &n1)
p1.Scond = arm.C_SCOND_NE
tr = gc.Types[gc.TUINT32]
gc.Regfree(&n3)
} else {
if nl.Ullman >= nr.Ullman {
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
gc.Regalloc(&n1, nr.Type, nil)
gc.Cgen(nr, &n1)
} else {
gc.Regalloc(&n1, nr.Type, nil)
gc.Cgen(nr, &n1)
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
}
}
// test for shift being 0
gins(arm.ATST, &n1, nil)
p3 := gc.Gbranch(arm.ABEQ, nil, -1)
// test and fix up large shifts
// TODO: if(!bounded), don't emit some of this.
gc.Regalloc(&n3, tr, nil)
gc.Nodconst(&t, gc.Types[gc.TUINT32], int64(w))
gmove(&t, &n3)
gins(arm.ACMP, &n1, &n3)
if op == gc.ORSH {
var p1 *obj.Prog
var p2 *obj.Prog
if gc.Issigned[nl.Type.Etype] {
p1 = gshift(arm.AMOVW, &n2, arm.SHIFT_AR, int32(w)-1, &n2)
p2 = gregshift(arm.AMOVW, &n2, arm.SHIFT_AR, &n1, &n2)
} else {
p1 = gins(arm.AEOR, &n2, &n2)
p2 = gregshift(arm.AMOVW, &n2, arm.SHIFT_LR, &n1, &n2)
}
p1.Scond = arm.C_SCOND_HS
p2.Scond = arm.C_SCOND_LO
} else {
p1 := gins(arm.AEOR, &n2, &n2)
p2 := gregshift(arm.AMOVW, &n2, arm.SHIFT_LL, &n1, &n2)
p1.Scond = arm.C_SCOND_HS
p2.Scond = arm.C_SCOND_LO
}
gc.Regfree(&n3)
gc.Patch(p3, gc.Pc)
// Left-shift of smaller word must be sign/zero-extended.
if w < 32 && op == gc.OLSH {
gins(optoas(gc.OAS, nl.Type), &n2, &n2)
}
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
/*
* generate comparison of nl, nr, both 64-bit.
* nl is memory; nr is constant or memory.
*/
func cmp64(nl *gc.Node, nr *gc.Node, op int, likely int, to *obj.Prog) {
var lo1 gc.Node
var hi1 gc.Node
var lo2 gc.Node
var hi2 gc.Node
var rr gc.Node
split64(nl, &lo1, &hi1)
split64(nr, &lo2, &hi2)
// compare most significant word;
// if they differ, we're done.
t := hi1.Type
if nl.Op == gc.OLITERAL || nr.Op == gc.OLITERAL {
gins(x86.ACMPL, &hi1, &hi2)
} else {
gc.Regalloc(&rr, gc.Types[gc.TINT32], nil)
gins(x86.AMOVL, &hi1, &rr)
gins(x86.ACMPL, &rr, &hi2)
gc.Regfree(&rr)
}
var br *obj.Prog
switch op {
default:
gc.Fatal("cmp64 %v %v", gc.Oconv(int(op), 0), gc.Tconv(t, 0))
// cmp hi
// jne L
// cmp lo
// jeq to
// L:
case gc.OEQ:
br = gc.Gbranch(x86.AJNE, nil, -likely)
// cmp hi
// jne to
// cmp lo
// jne to
case gc.ONE:
gc.Patch(gc.Gbranch(x86.AJNE, nil, likely), to)
// cmp hi
// jgt to
// jlt L
// cmp lo
// jge to (or jgt to)
// L:
case gc.OGE,
gc.OGT:
gc.Patch(gc.Gbranch(optoas(gc.OGT, t), nil, likely), to)
br = gc.Gbranch(optoas(gc.OLT, t), nil, -likely)
// cmp hi
// jlt to
// jgt L
// cmp lo
// jle to (or jlt to)
// L:
case gc.OLE,
gc.OLT:
gc.Patch(gc.Gbranch(optoas(gc.OLT, t), nil, likely), to)
br = gc.Gbranch(optoas(gc.OGT, t), nil, -likely)
}
// compare least significant word
t = lo1.Type
if nl.Op == gc.OLITERAL || nr.Op == gc.OLITERAL {
gins(x86.ACMPL, &lo1, &lo2)
} else {
gc.Regalloc(&rr, gc.Types[gc.TINT32], nil)
gins(x86.AMOVL, &lo1, &rr)
gins(x86.ACMPL, &rr, &lo2)
gc.Regfree(&rr)
}
// jump again
gc.Patch(gc.Gbranch(optoas(op, t), nil, likely), to)
// point first branch down here if appropriate
if br != nil {
gc.Patch(br, gc.Pc)
}
splitclean()
splitclean()
}
/*
* 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 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 := int(optoas(op, nl.Type))
if nr.Op == gc.OLITERAL {
var n1 gc.Node
regalloc(&n1, nl.Type, res)
cgen(nl, &n1)
sc := uint64(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)
regfree(&n1)
return
}
if nl.Ullman >= gc.UINF {
var n4 gc.Node
gc.Tempname(&n4, nl.Type)
cgen(nl, &n4)
nl = &n4
}
if nr.Ullman >= gc.UINF {
var n5 gc.Node
gc.Tempname(&n5, nr.Type)
cgen(nr, &n5)
nr = &n5
}
// Allow either uint32 or uint64 as shift type,
// to avoid unnecessary conversion from uint32 to uint64
// just to do the comparison.
tcount := gc.Types[gc.Simtype[nr.Type.Etype]]
if tcount.Etype < gc.TUINT32 {
tcount = gc.Types[gc.TUINT32]
}
var n1 gc.Node
regalloc(&n1, nr.Type, nil) // to hold the shift type in CX
var n3 gc.Node
regalloc(&n3, tcount, &n1) // to clear high bits of CX
var n2 gc.Node
regalloc(&n2, nl.Type, res)
if nl.Ullman >= nr.Ullman {
cgen(nl, &n2)
cgen(nr, &n1)
gmove(&n1, &n3)
} else {
cgen(nr, &n1)
gmove(&n1, &n3)
cgen(nl, &n2)
}
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 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, tcount), nil, +1))
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n2)
} else {
gc.Nodconst(&n3, nl.Type, 0)
gmove(&n3, &n2)
}
gc.Patch(p1, gc.Pc)
}
gins(a, &n1, &n2)
gmove(&n2, res)
regfree(&n1)
regfree(&n2)
}
func floatmove(f *gc.Node, t *gc.Node) {
var r1 gc.Node
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
cvt := t.Type
// cannot have two floating point memory operands.
if gc.Isfloat[ft] && gc.Isfloat[tt] && gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
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.Thearch.Use387 {
floatmove_387(f, t)
} else {
floatmove_sse(f, t)
}
return
// float to very long integer.
case gc.TFLOAT32<<16 | gc.TINT64,
gc.TFLOAT64<<16 | gc.TINT64:
if f.Op == gc.OREGISTER {
cvt = f.Type
goto hardmem
}
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[ft], x86.REG_F0)
if ft == gc.TFLOAT32 {
gins(x86.AFMOVF, f, &r1)
} else {
gins(x86.AFMOVD, f, &r1)
}
// set round to zero mode during conversion
var t1 gc.Node
memname(&t1, gc.Types[gc.TUINT16])
var t2 gc.Node
memname(&t2, gc.Types[gc.TUINT16])
gins(x86.AFSTCW, nil, &t1)
gins(x86.AMOVW, ncon(0xf7f), &t2)
gins(x86.AFLDCW, &t2, nil)
if tt == gc.TINT16 {
gins(x86.AFMOVWP, &r1, t)
} else if tt == gc.TINT32 {
gins(x86.AFMOVLP, &r1, t)
} else {
gins(x86.AFMOVVP, &r1, t)
}
gins(x86.AFLDCW, &t1, nil)
return
case gc.TFLOAT32<<16 | gc.TUINT64,
gc.TFLOAT64<<16 | gc.TUINT64:
if !gc.Ismem(f) {
cvt = f.Type
goto hardmem
}
bignodes()
var f0 gc.Node
gc.Nodreg(&f0, gc.Types[ft], x86.REG_F0)
var f1 gc.Node
gc.Nodreg(&f1, gc.Types[ft], x86.REG_F0+1)
var ax gc.Node
gc.Nodreg(&ax, gc.Types[gc.TUINT16], x86.REG_AX)
if ft == gc.TFLOAT32 {
gins(x86.AFMOVF, f, &f0)
} else {
gins(x86.AFMOVD, f, &f0)
}
// if 0 > v { answer = 0 }
gins(x86.AFMOVD, &zerof, &f0)
gins(x86.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:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
// requires memory intermediate
hardmem:
gc.Tempname(&r1, cvt)
gmove(f, &r1)
gmove(&r1, t)
return
}
func blockcopy(n, res *gc.Node, osrc, odst, w int64) {
// determine alignment.
// want to avoid unaligned access, so have to use
// smaller operations for less aligned types.
// for example moving [4]byte must use 4 MOVB not 1 MOVW.
align := int(n.Type.Align)
var op int
switch align {
default:
gc.Fatal("sgen: invalid alignment %d for %v", align, n.Type)
case 1:
op = arm.AMOVB
case 2:
op = arm.AMOVH
case 4:
op = arm.AMOVW
}
if w%int64(align) != 0 {
gc.Fatal("sgen: unaligned size %d (align=%d) for %v", w, align, n.Type)
}
c := int32(w / int64(align))
if osrc%int64(align) != 0 || odst%int64(align) != 0 {
gc.Fatal("sgen: unaligned offset src %d or dst %d (align %d)", osrc, odst, align)
}
// if we are copying forward on the stack and
// the src and dst overlap, then reverse direction
dir := align
if osrc < odst && int64(odst) < int64(osrc)+w {
dir = -dir
}
if op == arm.AMOVW && !gc.Nacl && dir > 0 && c >= 4 && c <= 128 {
var r0 gc.Node
r0.Op = gc.OREGISTER
r0.Reg = arm.REG_R0
var r1 gc.Node
r1.Op = gc.OREGISTER
r1.Reg = arm.REG_R0 + 1
var r2 gc.Node
r2.Op = gc.OREGISTER
r2.Reg = arm.REG_R0 + 2
var src gc.Node
gc.Regalloc(&src, gc.Types[gc.Tptr], &r1)
var dst gc.Node
gc.Regalloc(&dst, gc.Types[gc.Tptr], &r2)
if n.Ullman >= res.Ullman {
// eval n first
gc.Agen(n, &src)
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agen(res, &dst)
} else {
// eval res first
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agen(res, &dst)
gc.Agen(n, &src)
}
var tmp gc.Node
gc.Regalloc(&tmp, gc.Types[gc.Tptr], &r0)
f := gc.Sysfunc("duffcopy")
p := gins(obj.ADUFFCOPY, nil, f)
gc.Afunclit(&p.To, f)
// 8 and 128 = magic constants: see ../../runtime/asm_arm.s
p.To.Offset = 8 * (128 - int64(c))
gc.Regfree(&tmp)
gc.Regfree(&src)
gc.Regfree(&dst)
return
}
var dst gc.Node
var src gc.Node
if n.Ullman >= res.Ullman {
gc.Agenr(n, &dst, res) // temporarily use dst
gc.Regalloc(&src, gc.Types[gc.Tptr], nil)
gins(arm.AMOVW, &dst, &src)
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agen(res, &dst)
} else {
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agenr(res, &dst, res)
gc.Agenr(n, &src, nil)
}
var tmp gc.Node
gc.Regalloc(&tmp, gc.Types[gc.TUINT32], nil)
// set up end marker
var nend gc.Node
if c >= 4 {
gc.Regalloc(&nend, gc.Types[gc.TUINT32], nil)
p := gins(arm.AMOVW, &src, &nend)
p.From.Type = obj.TYPE_ADDR
if dir < 0 {
p.From.Offset = int64(dir)
} else {
p.From.Offset = w
}
}
// move src and dest to the end of block if necessary
if dir < 0 {
p := gins(arm.AMOVW, &src, &src)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w + int64(dir)
p = gins(arm.AMOVW, &dst, &dst)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w + int64(dir)
}
// move
if c >= 4 {
p := gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
p.Scond |= arm.C_PBIT
ploop := p
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
p.Scond |= arm.C_PBIT
p = gins(arm.ACMP, &src, nil)
raddr(&nend, p)
gc.Patch(gc.Gbranch(arm.ABNE, nil, 0), ploop)
gc.Regfree(&nend)
} else {
var p *obj.Prog
for {
tmp14 := c
c--
if tmp14 <= 0 {
break
}
p = gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
p.Scond |= arm.C_PBIT
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
p.Scond |= arm.C_PBIT
}
}
gc.Regfree(&dst)
gc.Regfree(&src)
gc.Regfree(&tmp)
}
/*
* generate move:
* t = f
* hard part is conversions.
*/
func gmove(f *gc.Node, t *gc.Node) {
if gc.Debug['M'] != 0 {
fmt.Printf("gmove %v -> %v\n", gc.Nconv(f, obj.FmtLong), gc.Nconv(t, obj.FmtLong))
}
ft := int(gc.Simsimtype(f.Type))
tt := int(gc.Simsimtype(t.Type))
cvt := (*gc.Type)(t.Type)
if gc.Iscomplex[ft] || gc.Iscomplex[tt] {
gc.Complexmove(f, t)
return
}
// cannot have two memory operands
var r2 gc.Node
var r1 gc.Node
var a int
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
switch tt {
default:
gc.Convconst(&con, t.Type, &f.Val)
case gc.TINT32,
gc.TINT16,
gc.TINT8:
var con gc.Node
gc.Convconst(&con, gc.Types[gc.TINT64], &f.Val)
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(ppc64.AMOVD, &con, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
case gc.TUINT32,
gc.TUINT16,
gc.TUINT8:
var con gc.Node
gc.Convconst(&con, gc.Types[gc.TUINT64], &f.Val)
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(ppc64.AMOVD, &con, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
f = &con
ft = tt // so big switch will choose a simple mov
// constants can't move directly to memory.
if gc.Ismem(t) {
goto hard
}
}
// float constants come from memory.
//if(isfloat[tt])
// goto hard;
// 64-bit immediates are also from memory.
//if(isint[tt])
// goto hard;
//// 64-bit immediates are really 32-bit sign-extended
//// unless moving into a register.
//if(isint[tt]) {
// if(mpcmpfixfix(con.val.u.xval, minintval[TINT32]) < 0)
// goto hard;
// if(mpcmpfixfix(con.val.u.xval, maxintval[TINT32]) > 0)
// goto hard;
//}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
gc.Fatal("gmove %v -> %v", gc.Tconv(f.Type, obj.FmtLong), gc.Tconv(t.Type, obj.FmtLong))
/*
* integer copy and truncate
*/
case gc.TINT8<<16 | gc.TINT8, // same size
gc.TUINT8<<16 | gc.TINT8,
gc.TINT16<<16 | gc.TINT8,
// truncate
gc.TUINT16<<16 | gc.TINT8,
gc.TINT32<<16 | gc.TINT8,
gc.TUINT32<<16 | gc.TINT8,
gc.TINT64<<16 | gc.TINT8,
gc.TUINT64<<16 | gc.TINT8:
a = ppc64.AMOVB
case gc.TINT8<<16 | gc.TUINT8, // same size
gc.TUINT8<<16 | gc.TUINT8,
gc.TINT16<<16 | gc.TUINT8,
// truncate
gc.TUINT16<<16 | gc.TUINT8,
gc.TINT32<<16 | gc.TUINT8,
gc.TUINT32<<16 | gc.TUINT8,
gc.TINT64<<16 | gc.TUINT8,
gc.TUINT64<<16 | gc.TUINT8:
a = ppc64.AMOVBZ
case gc.TINT16<<16 | gc.TINT16, // same size
gc.TUINT16<<16 | gc.TINT16,
gc.TINT32<<16 | gc.TINT16,
// truncate
gc.TUINT32<<16 | gc.TINT16,
gc.TINT64<<16 | gc.TINT16,
gc.TUINT64<<16 | gc.TINT16:
a = ppc64.AMOVH
case gc.TINT16<<16 | gc.TUINT16, // same size
gc.TUINT16<<16 | gc.TUINT16,
gc.TINT32<<16 | gc.TUINT16,
// truncate
gc.TUINT32<<16 | gc.TUINT16,
gc.TINT64<<16 | gc.TUINT16,
gc.TUINT64<<16 | gc.TUINT16:
a = ppc64.AMOVHZ
case gc.TINT32<<16 | gc.TINT32, // same size
gc.TUINT32<<16 | gc.TINT32,
gc.TINT64<<16 | gc.TINT32,
// truncate
gc.TUINT64<<16 | gc.TINT32:
a = ppc64.AMOVW
case gc.TINT32<<16 | gc.TUINT32, // same size
gc.TUINT32<<16 | gc.TUINT32,
gc.TINT64<<16 | gc.TUINT32,
gc.TUINT64<<16 | gc.TUINT32:
a = ppc64.AMOVWZ
case gc.TINT64<<16 | gc.TINT64, // same size
gc.TINT64<<16 | gc.TUINT64,
gc.TUINT64<<16 | gc.TINT64,
gc.TUINT64<<16 | gc.TUINT64:
a = ppc64.AMOVD
/*
* integer up-conversions
*/
case gc.TINT8<<16 | gc.TINT16, // sign extend int8
gc.TINT8<<16 | gc.TUINT16,
gc.TINT8<<16 | gc.TINT32,
gc.TINT8<<16 | gc.TUINT32,
gc.TINT8<<16 | gc.TINT64,
gc.TINT8<<16 | gc.TUINT64:
a = ppc64.AMOVB
goto rdst
case gc.TUINT8<<16 | gc.TINT16, // zero extend uint8
gc.TUINT8<<16 | gc.TUINT16,
gc.TUINT8<<16 | gc.TINT32,
gc.TUINT8<<16 | gc.TUINT32,
gc.TUINT8<<16 | gc.TINT64,
gc.TUINT8<<16 | gc.TUINT64:
a = ppc64.AMOVBZ
goto rdst
case gc.TINT16<<16 | gc.TINT32, // sign extend int16
gc.TINT16<<16 | gc.TUINT32,
gc.TINT16<<16 | gc.TINT64,
gc.TINT16<<16 | gc.TUINT64:
a = ppc64.AMOVH
goto rdst
case gc.TUINT16<<16 | gc.TINT32, // zero extend uint16
gc.TUINT16<<16 | gc.TUINT32,
gc.TUINT16<<16 | gc.TINT64,
gc.TUINT16<<16 | gc.TUINT64:
a = ppc64.AMOVHZ
goto rdst
case gc.TINT32<<16 | gc.TINT64, // sign extend int32
gc.TINT32<<16 | gc.TUINT64:
a = ppc64.AMOVW
goto rdst
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
a = ppc64.AMOVWZ
goto rdst
//warn("gmove: convert float to int not implemented: %N -> %N\n", f, t);
//return;
// algorithm is:
// if small enough, use native float64 -> int64 conversion.
// otherwise, subtract 2^63, convert, and add it back.
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT32,
gc.TFLOAT64<<16 | gc.TINT32,
gc.TFLOAT32<<16 | gc.TINT64,
gc.TFLOAT64<<16 | gc.TINT64,
gc.TFLOAT32<<16 | gc.TINT16,
gc.TFLOAT32<<16 | gc.TINT8,
gc.TFLOAT32<<16 | gc.TUINT16,
gc.TFLOAT32<<16 | gc.TUINT8,
gc.TFLOAT64<<16 | gc.TINT16,
gc.TFLOAT64<<16 | gc.TINT8,
gc.TFLOAT64<<16 | gc.TUINT16,
gc.TFLOAT64<<16 | gc.TUINT8,
gc.TFLOAT32<<16 | gc.TUINT32,
gc.TFLOAT64<<16 | gc.TUINT32,
gc.TFLOAT32<<16 | gc.TUINT64,
gc.TFLOAT64<<16 | gc.TUINT64:
bignodes()
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[ft], f)
gmove(f, &r1)
if tt == gc.TUINT64 {
gc.Regalloc(&r2, gc.Types[gc.TFLOAT64], nil)
gmove(&bigf, &r2)
gins(ppc64.AFCMPU, &r1, &r2)
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TFLOAT64]), nil, +1))
gins(ppc64.AFSUB, &r2, &r1)
gc.Patch(p1, gc.Pc)
gc.Regfree(&r2)
}
gc.Regalloc(&r2, gc.Types[gc.TFLOAT64], nil)
var r3 gc.Node
gc.Regalloc(&r3, gc.Types[gc.TINT64], t)
gins(ppc64.AFCTIDZ, &r1, &r2)
p1 := (*obj.Prog)(gins(ppc64.AFMOVD, &r2, nil))
p1.To.Type = obj.TYPE_MEM
p1.To.Reg = ppc64.REGSP
p1.To.Offset = -8
p1 = gins(ppc64.AMOVD, nil, &r3)
p1.From.Type = obj.TYPE_MEM
p1.From.Reg = ppc64.REGSP
p1.From.Offset = -8
gc.Regfree(&r2)
gc.Regfree(&r1)
if tt == gc.TUINT64 {
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TFLOAT64]), nil, +1)) // use CR0 here again
gc.Nodreg(&r1, gc.Types[gc.TINT64], ppc64.REGTMP)
gins(ppc64.AMOVD, &bigi, &r1)
gins(ppc64.AADD, &r1, &r3)
gc.Patch(p1, gc.Pc)
}
gmove(&r3, t)
gc.Regfree(&r3)
return
//warn("gmove: convert int to float not implemented: %N -> %N\n", f, t);
//return;
// algorithm is:
// if small enough, use native int64 -> uint64 conversion.
// otherwise, halve (rounding to odd?), convert, and double.
/*
* integer to float
*/
case gc.TINT32<<16 | gc.TFLOAT32,
gc.TINT32<<16 | gc.TFLOAT64,
gc.TINT64<<16 | gc.TFLOAT32,
gc.TINT64<<16 | gc.TFLOAT64,
gc.TINT16<<16 | gc.TFLOAT32,
gc.TINT16<<16 | gc.TFLOAT64,
gc.TINT8<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TUINT16<<16 | gc.TFLOAT32,
gc.TUINT16<<16 | gc.TFLOAT64,
gc.TUINT8<<16 | gc.TFLOAT32,
gc.TUINT8<<16 | gc.TFLOAT64,
gc.TUINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT64,
gc.TUINT64<<16 | gc.TFLOAT32,
gc.TUINT64<<16 | gc.TFLOAT64:
bignodes()
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[gc.TINT64], nil)
gmove(f, &r1)
if ft == gc.TUINT64 {
gc.Nodreg(&r2, gc.Types[gc.TUINT64], ppc64.REGTMP)
gmove(&bigi, &r2)
gins(ppc64.ACMPU, &r1, &r2)
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT64]), nil, +1))
p2 := (*obj.Prog)(gins(ppc64.ASRD, nil, &r1))
p2.From.Type = obj.TYPE_CONST
p2.From.Offset = 1
gc.Patch(p1, gc.Pc)
}
gc.Regalloc(&r2, gc.Types[gc.TFLOAT64], t)
p1 := (*obj.Prog)(gins(ppc64.AMOVD, &r1, nil))
p1.To.Type = obj.TYPE_MEM
p1.To.Reg = ppc64.REGSP
p1.To.Offset = -8
p1 = gins(ppc64.AFMOVD, nil, &r2)
p1.From.Type = obj.TYPE_MEM
p1.From.Reg = ppc64.REGSP
p1.From.Offset = -8
gins(ppc64.AFCFID, &r2, &r2)
gc.Regfree(&r1)
if ft == gc.TUINT64 {
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT64]), nil, +1)) // use CR0 here again
gc.Nodreg(&r1, gc.Types[gc.TFLOAT64], ppc64.FREGTWO)
gins(ppc64.AFMUL, &r1, &r2)
gc.Patch(p1, gc.Pc)
}
gmove(&r2, t)
gc.Regfree(&r2)
return
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32:
a = ppc64.AFMOVS
case gc.TFLOAT64<<16 | gc.TFLOAT64:
a = ppc64.AFMOVD
case gc.TFLOAT32<<16 | gc.TFLOAT64:
a = ppc64.AFMOVS
goto rdst
case gc.TFLOAT64<<16 | gc.TFLOAT32:
a = ppc64.AFRSP
goto rdst
}
gins(a, f, t)
return
// requires register destination
rdst:
{
gc.Regalloc(&r1, t.Type, t)
gins(a, f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
func blockcopy(n, res *gc.Node, osrc, odst, w int64) {
// determine alignment.
// want to avoid unaligned access, so have to use
// smaller operations for less aligned types.
// for example moving [4]byte must use 4 MOVB not 1 MOVW.
align := int(n.Type.Align)
var op int
switch align {
default:
gc.Fatal("sgen: invalid alignment %d for %v", align, n.Type)
case 1:
op = ppc64.AMOVBU
case 2:
op = ppc64.AMOVHU
case 4:
op = ppc64.AMOVWZU // there is no lwau, only lwaux
case 8:
op = ppc64.AMOVDU
}
if w%int64(align) != 0 {
gc.Fatal("sgen: unaligned size %d (align=%d) for %v", w, align, n.Type)
}
c := int32(w / int64(align))
// if we are copying forward on the stack and
// the src and dst overlap, then reverse direction
dir := align
if osrc < odst && int64(odst) < int64(osrc)+w {
dir = -dir
}
var dst gc.Node
var src gc.Node
if n.Ullman >= res.Ullman {
gc.Agenr(n, &dst, res) // temporarily use dst
gc.Regalloc(&src, gc.Types[gc.Tptr], nil)
gins(ppc64.AMOVD, &dst, &src)
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agen(res, &dst)
} else {
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
gc.Agenr(res, &dst, res)
gc.Agenr(n, &src, nil)
}
var tmp gc.Node
gc.Regalloc(&tmp, gc.Types[gc.Tptr], nil)
// set up end marker
var nend gc.Node
// move src and dest to the end of block if necessary
if dir < 0 {
if c >= 4 {
gc.Regalloc(&nend, gc.Types[gc.Tptr], nil)
gins(ppc64.AMOVD, &src, &nend)
}
p := gins(ppc64.AADD, nil, &src)
p.From.Type = obj.TYPE_CONST
p.From.Offset = w
p = gins(ppc64.AADD, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = w
} else {
p := gins(ppc64.AADD, nil, &src)
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(-dir)
p = gins(ppc64.AADD, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(-dir)
if c >= 4 {
gc.Regalloc(&nend, gc.Types[gc.Tptr], nil)
p := gins(ppc64.AMOVD, &src, &nend)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = w
}
}
// move
// TODO: enable duffcopy for larger copies.
if c >= 4 {
p := gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
ploop := p
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
p = gins(ppc64.ACMP, &src, &nend)
gc.Patch(gc.Gbranch(ppc64.ABNE, nil, 0), ploop)
gc.Regfree(&nend)
} else {
// TODO(austin): Instead of generating ADD $-8,R8; ADD
// $-8,R7; n*(MOVDU 8(R8),R9; MOVDU R9,8(R7);) just
// generate the offsets directly and eliminate the
// ADDs. That will produce shorter, more
// pipeline-able code.
var p *obj.Prog
for {
tmp14 := c
c--
if tmp14 <= 0 {
break
}
p = gins(op, &src, &tmp)
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(dir)
p = gins(op, &tmp, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(dir)
}
}
gc.Regfree(&dst)
gc.Regfree(&src)
gc.Regfree(&tmp)
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op int, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
a := optoas(op, nl.Type)
if nr.Op == gc.OLITERAL {
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gc.Cgen(nl, &n1)
sc := uint64(gc.Mpgetfix(nr.Val.U.Xval))
if sc >= uint64(nl.Type.Width*8) {
// large shift gets 2 shifts by width-1
var n3 gc.Node
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n1)
gins(a, &n3, &n1)
} else {
gins(a, nr, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
if nl.Ullman >= gc.UINF {
var n4 gc.Node
gc.Tempname(&n4, nl.Type)
gc.Cgen(nl, &n4)
nl = &n4
}
if nr.Ullman >= gc.UINF {
var n5 gc.Node
gc.Tempname(&n5, nr.Type)
gc.Cgen(nr, &n5)
nr = &n5
}
rcx := int(reg[x86.REG_CX])
var n1 gc.Node
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
// Allow either uint32 or uint64 as shift type,
// to avoid unnecessary conversion from uint32 to uint64
// just to do the comparison.
tcount := gc.Types[gc.Simtype[nr.Type.Etype]]
if tcount.Etype < gc.TUINT32 {
tcount = gc.Types[gc.TUINT32]
}
gc.Regalloc(&n1, nr.Type, &n1) // to hold the shift type in CX
var n3 gc.Node
gc.Regalloc(&n3, tcount, &n1) // to clear high bits of CX
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TUINT64], x86.REG_CX)
var oldcx gc.Node
if rcx > 0 && !gc.Samereg(&cx, res) {
gc.Regalloc(&oldcx, gc.Types[gc.TUINT64], nil)
gmove(&cx, &oldcx)
}
cx.Type = tcount
var n2 gc.Node
if gc.Samereg(&cx, res) {
gc.Regalloc(&n2, nl.Type, nil)
} else {
gc.Regalloc(&n2, nl.Type, res)
}
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &n2)
gc.Cgen(nr, &n1)
gmove(&n1, &n3)
} else {
gc.Cgen(nr, &n1)
gmove(&n1, &n3)
gc.Cgen(nl, &n2)
}
gc.Regfree(&n3)
// test and fix up large shifts
if !bounded {
gc.Nodconst(&n3, tcount, nl.Type.Width*8)
gins(optoas(gc.OCMP, tcount), &n1, &n3)
p1 := gc.Gbranch(optoas(gc.OLT, tcount), nil, +1)
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n2)
} else {
gc.Nodconst(&n3, nl.Type, 0)
gmove(&n3, &n2)
}
gc.Patch(p1, gc.Pc)
}
gins(a, &n1, &n2)
if oldcx.Op != 0 {
cx.Type = gc.Types[gc.TUINT64]
gmove(&oldcx, &cx)
gc.Regfree(&oldcx)
}
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
/*
* generate:
* call f
* proc=-1 normal call but no return
* proc=0 normal call
* proc=1 goroutine run in new proc
* proc=2 defer call save away stack
* proc=3 normal call to C pointer (not Go func value)
*/
func ginscall(f *gc.Node, proc int) {
if f.Type != nil {
extra := int32(0)
if proc == 1 || proc == 2 {
extra = 2 * int32(gc.Widthptr)
}
gc.Setmaxarg(f.Type, extra)
}
switch proc {
default:
gc.Fatal("ginscall: bad proc %d", proc)
case 0, // normal call
-1: // normal call but no return
if f.Op == gc.ONAME && f.Class == gc.PFUNC {
if f == gc.Deferreturn {
// Deferred calls will appear to be returning to
// the CALL deferreturn(SB) that we are about to emit.
// However, the stack trace code will show the line
// of the instruction byte before the return PC.
// To avoid that being an unrelated instruction,
// insert an x86 NOP that we will have the right line number.
// x86 NOP 0x90 is really XCHG AX, AX; use that description
// because the NOP pseudo-instruction will be removed by
// the linker.
var reg gc.Node
gc.Nodreg(®, gc.Types[gc.TINT], x86.REG_AX)
gins(x86.AXCHGL, ®, ®)
}
p := gins(obj.ACALL, nil, f)
gc.Afunclit(&p.To, f)
if proc == -1 || gc.Noreturn(p) {
gins(obj.AUNDEF, nil, nil)
}
break
}
var reg gc.Node
gc.Nodreg(®, gc.Types[gc.Tptr], x86.REG_DX)
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[gc.Tptr], x86.REG_BX)
gmove(f, ®)
reg.Op = gc.OINDREG
gmove(®, &r1)
reg.Op = gc.OREGISTER
gins(obj.ACALL, ®, &r1)
case 3: // normal call of c function pointer
gins(obj.ACALL, nil, f)
case 1, // call in new proc (go)
2: // deferred call (defer)
var stk gc.Node
stk.Op = gc.OINDREG
stk.Val.U.Reg = x86.REG_SP
stk.Xoffset = 0
// size of arguments at 0(SP)
var con gc.Node
gc.Nodconst(&con, gc.Types[gc.TINT32], int64(gc.Argsize(f.Type)))
gins(x86.AMOVL, &con, &stk)
// FuncVal* at 4(SP)
stk.Xoffset = int64(gc.Widthptr)
gins(x86.AMOVL, f, &stk)
if proc == 1 {
ginscall(gc.Newproc, 0)
} else {
ginscall(gc.Deferproc, 0)
}
if proc == 2 {
var reg gc.Node
gc.Nodreg(®, gc.Types[gc.TINT32], x86.REG_AX)
gins(x86.ATESTL, ®, ®)
p := gc.Gbranch(x86.AJEQ, nil, +1)
cgen_ret(nil)
gc.Patch(p, gc.Pc)
}
}
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op int, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Type.Width > 4 {
gc.Fatal("cgen_shift %v", gc.Tconv(nl.Type, 0))
}
w := int(nl.Type.Width * 8)
a := optoas(op, nl.Type)
if nr.Op == gc.OLITERAL {
var n2 gc.Node
gc.Tempname(&n2, nl.Type)
gc.Cgen(nl, &n2)
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gmove(&n2, &n1)
sc := uint64(gc.Mpgetfix(nr.Val.U.Xval))
if sc >= uint64(nl.Type.Width*8) {
// large shift gets 2 shifts by width-1
gins(a, ncon(uint32(w)-1), &n1)
gins(a, ncon(uint32(w)-1), &n1)
} else {
gins(a, nr, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
var oldcx gc.Node
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TUINT32], x86.REG_CX)
if reg[x86.REG_CX] > 1 && !gc.Samereg(&cx, res) {
gc.Tempname(&oldcx, gc.Types[gc.TUINT32])
gmove(&cx, &oldcx)
}
var n1 gc.Node
var nt gc.Node
if nr.Type.Width > 4 {
gc.Tempname(&nt, nr.Type)
n1 = nt
} else {
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, nr.Type, &n1) // to hold the shift type in CX
}
var n2 gc.Node
if gc.Samereg(&cx, res) {
gc.Regalloc(&n2, nl.Type, nil)
} else {
gc.Regalloc(&n2, nl.Type, res)
}
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &n2)
gc.Cgen(nr, &n1)
} else {
gc.Cgen(nr, &n1)
gc.Cgen(nl, &n2)
}
// test and fix up large shifts
if bounded {
if nr.Type.Width > 4 {
// delayed reg alloc
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], &n1) // to hold the shift type in CX
var lo gc.Node
var hi gc.Node
split64(&nt, &lo, &hi)
gmove(&lo, &n1)
splitclean()
}
} else {
var p1 *obj.Prog
if nr.Type.Width > 4 {
// delayed reg alloc
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], &n1) // to hold the shift type in CX
var lo gc.Node
var hi gc.Node
split64(&nt, &lo, &hi)
gmove(&lo, &n1)
gins(optoas(gc.OCMP, gc.Types[gc.TUINT32]), &hi, ncon(0))
p2 := gc.Gbranch(optoas(gc.ONE, gc.Types[gc.TUINT32]), nil, +1)
gins(optoas(gc.OCMP, gc.Types[gc.TUINT32]), &n1, ncon(uint32(w)))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
splitclean()
gc.Patch(p2, gc.Pc)
} else {
gins(optoas(gc.OCMP, nr.Type), &n1, ncon(uint32(w)))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
}
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gins(a, ncon(uint32(w)-1), &n2)
} else {
gmove(ncon(0), &n2)
}
gc.Patch(p1, gc.Pc)
}
gins(a, &n1, &n2)
if oldcx.Op != 0 {
gmove(&oldcx, &cx)
}
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
/*
* 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)
}
}
func bgen_float(n *gc.Node, true_ int, likely int, to *obj.Prog) {
nl := n.Left
nr := n.Right
a := int(n.Op)
if true_ == 0 {
// brcom is not valid on floats when NaN is involved.
p1 := gc.Gbranch(obj.AJMP, nil, 0)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// No need to avoid re-genning ninit.
bgen_float(n, 1, -likely, p2)
gc.Patch(gc.Gbranch(obj.AJMP, nil, 0), to)
gc.Patch(p2, gc.Pc)
return
}
var tmp gc.Node
var et int
var n2 gc.Node
var ax gc.Node
if !gc.Thearch.Use387 {
if nl.Addable == 0 {
var n1 gc.Node
gc.Tempname(&n1, nl.Type)
gc.Cgen(nl, &n1)
nl = &n1
}
if nr.Addable == 0 {
var tmp gc.Node
gc.Tempname(&tmp, nr.Type)
gc.Cgen(nr, &tmp)
nr = &tmp
}
var n2 gc.Node
gc.Regalloc(&n2, nr.Type, nil)
gmove(nr, &n2)
nr = &n2
if nl.Op != gc.OREGISTER {
var n3 gc.Node
gc.Regalloc(&n3, nl.Type, nil)
gmove(nl, &n3)
nl = &n3
}
if a == gc.OGE || a == gc.OGT {
// only < and <= work right with NaN; reverse if needed
r := nr
nr = nl
nl = r
a = gc.Brrev(a)
}
gins(foptoas(gc.OCMP, nr.Type, 0), nl, nr)
if nl.Op == gc.OREGISTER {
gc.Regfree(nl)
}
gc.Regfree(nr)
goto ret
} else {
goto x87
}
x87:
a = gc.Brrev(a) // because the args are stacked
if a == gc.OGE || a == gc.OGT {
// only < and <= work right with NaN; reverse if needed
r := nr
nr = nl
nl = r
a = gc.Brrev(a)
}
gc.Nodreg(&tmp, nr.Type, x86.REG_F0)
gc.Nodreg(&n2, nr.Type, x86.REG_F0+1)
gc.Nodreg(&ax, gc.Types[gc.TUINT16], x86.REG_AX)
et = gc.Simsimtype(nr.Type)
if et == gc.TFLOAT64 {
if nl.Ullman > nr.Ullman {
gc.Cgen(nl, &tmp)
gc.Cgen(nr, &tmp)
gins(x86.AFXCHD, &tmp, &n2)
} else {
gc.Cgen(nr, &tmp)
gc.Cgen(nl, &tmp)
}
gins(x86.AFUCOMIP, &tmp, &n2)
gins(x86.AFMOVDP, &tmp, &tmp) // annoying pop but still better than STSW+SAHF
} else {
// TODO(rsc): The moves back and forth to memory
// here are for truncating the value to 32 bits.
// This handles 32-bit comparison but presumably
// all the other ops have the same problem.
// We need to figure out what the right general
// solution is, besides telling people to use float64.
var t1 gc.Node
gc.Tempname(&t1, gc.Types[gc.TFLOAT32])
var t2 gc.Node
gc.Tempname(&t2, gc.Types[gc.TFLOAT32])
gc.Cgen(nr, &t1)
gc.Cgen(nl, &t2)
gmove(&t2, &tmp)
gins(x86.AFCOMFP, &t1, &tmp)
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
}
goto ret
ret:
if a == gc.OEQ {
// neither NE nor P
p1 := gc.Gbranch(x86.AJNE, nil, -likely)
p2 := gc.Gbranch(x86.AJPS, nil, -likely)
gc.Patch(gc.Gbranch(obj.AJMP, nil, 0), to)
gc.Patch(p1, gc.Pc)
gc.Patch(p2, gc.Pc)
} else if a == gc.ONE {
// either NE or P
gc.Patch(gc.Gbranch(x86.AJNE, nil, likely), to)
gc.Patch(gc.Gbranch(x86.AJPS, nil, likely), to)
} else {
gc.Patch(gc.Gbranch(optoas(a, nr.Type), nil, likely), to)
}
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
fmt.Printf("clearfat %v (%v, size: %d)\n", gc.Nconv(nl, 0), gc.Tconv(nl.Type, 0), nl.Type.Width)
}
w := uint64(uint64(nl.Type.Width))
// Avoid taking the address for simple enough types.
if gc.Componentgen(nil, nl) {
return
}
c := uint64(w % 8) // bytes
q := uint64(w / 8) // dwords
if gc.Reginuse(ppc64.REGRT1) {
gc.Fatal("%v in use during clearfat", obj.Rconv(ppc64.REGRT1))
}
var r0 gc.Node
gc.Nodreg(&r0, gc.Types[gc.TUINT64], ppc64.REGZERO)
var dst gc.Node
gc.Nodreg(&dst, gc.Types[gc.Tptr], ppc64.REGRT1)
gc.Regrealloc(&dst)
gc.Agen(nl, &dst)
var boff uint64
if q > 128 {
p := gins(ppc64.ASUB, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p = gins(ppc64.AMOVD, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(q * 8)
p = gins(ppc64.AMOVDU, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 8
pl := (*obj.Prog)(p)
p = gins(ppc64.ACMP, &dst, &end)
gc.Patch(gc.Gbranch(ppc64.ABNE, nil, 0), pl)
gc.Regfree(&end)
// The loop leaves R3 on the last zeroed dword
boff = 8
} else if q >= 4 {
p := gins(ppc64.ASUB, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
f := (*gc.Node)(gc.Sysfunc("duffzero"))
p = gins(obj.ADUFFZERO, nil, f)
gc.Afunclit(&p.To, f)
// 4 and 128 = magic constants: see ../../runtime/asm_ppc64x.s
p.To.Offset = int64(4 * (128 - q))
// duffzero leaves R3 on the last zeroed dword
boff = 8
} else {
var p *obj.Prog
for t := uint64(0); t < q; t++ {
p = gins(ppc64.AMOVD, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(8 * t)
}
boff = 8 * q
}
var p *obj.Prog
for t := uint64(0); t < c; t++ {
p = gins(ppc64.AMOVB, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(t + boff)
}
gc.Regfree(&dst)
}