/*
* n is a 64-bit value. fill in lo and hi to refer to its 32-bit halves.
*/
func split64(n *gc.Node, lo *gc.Node, hi *gc.Node) {
if !gc.Is64(n.Type) {
gc.Fatal("split64 %v", n.Type)
}
if nsclean >= len(sclean) {
gc.Fatal("split64 clean")
}
sclean[nsclean].Op = gc.OEMPTY
nsclean++
switch n.Op {
default:
switch n.Op {
default:
var n1 gc.Node
if !dotaddable(n, &n1) {
gc.Igen(n, &n1, nil)
sclean[nsclean-1] = n1
}
n = &n1
case gc.ONAME:
if n.Class == gc.PPARAMREF {
var n1 gc.Node
gc.Cgen(n.Name.Heapaddr, &n1)
sclean[nsclean-1] = n1
n = &n1
}
// nothing
case gc.OINDREG:
break
}
*lo = *n
*hi = *n
lo.Type = gc.Types[gc.TUINT32]
if n.Type.Etype == gc.TINT64 {
hi.Type = gc.Types[gc.TINT32]
} else {
hi.Type = gc.Types[gc.TUINT32]
}
hi.Xoffset += 4
case gc.OLITERAL:
var n1 gc.Node
n.Convconst(&n1, n.Type)
i := n1.Int()
gc.Nodconst(lo, gc.Types[gc.TUINT32], int64(uint32(i)))
i >>= 32
if n.Type.Etype == gc.TINT64 {
gc.Nodconst(hi, gc.Types[gc.TINT32], int64(int32(i)))
} else {
gc.Nodconst(hi, gc.Types[gc.TUINT32], int64(uint32(i)))
}
}
}
func defframe(ptxt *obj.Prog) {
var n *gc.Node
// fill in argument size, stack size
ptxt.To.Type = obj.TYPE_TEXTSIZE
ptxt.To.Val = int32(gc.Rnd(gc.Curfn.Type.Argwid, int64(gc.Widthptr)))
frame := uint32(gc.Rnd(gc.Stksize+gc.Maxarg, int64(gc.Widthreg)))
// arm64 requires that the frame size (not counting saved LR)
// be empty or be 8 mod 16. If not, pad it.
if frame != 0 && frame%16 != 8 {
frame += 8
}
ptxt.To.Offset = int64(frame)
// insert code to zero ambiguously live variables
// so that the garbage collector only sees initialized values
// when it looks for pointers.
p := ptxt
hi := int64(0)
lo := hi
// iterate through declarations - they are sorted in decreasing xoffset order.
for l := gc.Curfn.Func.Dcl; l != nil; l = l.Next {
n = l.N
if !n.Name.Needzero {
continue
}
if n.Class != gc.PAUTO {
gc.Fatal("needzero class %d", n.Class)
}
if n.Type.Width%int64(gc.Widthptr) != 0 || n.Xoffset%int64(gc.Widthptr) != 0 || n.Type.Width == 0 {
gc.Fatal("var %v has size %d offset %d", gc.Nconv(n, obj.FmtLong), int(n.Type.Width), int(n.Xoffset))
}
if lo != hi && n.Xoffset+n.Type.Width >= lo-int64(2*gc.Widthreg) {
// merge with range we already have
lo = n.Xoffset
continue
}
// zero old range
p = zerorange(p, int64(frame), lo, hi)
// set new range
hi = n.Xoffset + n.Type.Width
lo = n.Xoffset
}
// zero final range
zerorange(p, int64(frame), lo, hi)
}
/*
* insert n into reg slot of p
*/
func raddr(n *gc.Node, p *obj.Prog) {
var a obj.Addr
gc.Naddr(&a, n)
if a.Type != obj.TYPE_REG {
if n != nil {
gc.Fatal("bad in raddr: %v", gc.Oconv(int(n.Op), 0))
} else {
gc.Fatal("bad in raddr: <null>")
}
p.Reg = 0
} else {
p.Reg = a.Reg
}
}
func defframe(ptxt *obj.Prog) {
var n *gc.Node
// fill in argument size, stack size
ptxt.To.Type = obj.TYPE_TEXTSIZE
ptxt.To.Val = int32(gc.Rnd(gc.Curfn.Type.Argwid, int64(gc.Widthptr)))
frame := uint32(gc.Rnd(gc.Stksize+gc.Maxarg, int64(gc.Widthreg)))
ptxt.To.Offset = int64(frame)
// insert code to zero ambiguously live variables
// so that the garbage collector only sees initialized values
// when it looks for pointers.
p := ptxt
hi := int64(0)
lo := hi
ax := uint32(0)
for l := gc.Curfn.Func.Dcl; l != nil; l = l.Next {
n = l.N
if !n.Name.Needzero {
continue
}
if n.Class != gc.PAUTO {
gc.Fatal("needzero class %d", n.Class)
}
if n.Type.Width%int64(gc.Widthptr) != 0 || n.Xoffset%int64(gc.Widthptr) != 0 || n.Type.Width == 0 {
gc.Fatal("var %v has size %d offset %d", gc.Nconv(n, obj.FmtLong), int(n.Type.Width), int(n.Xoffset))
}
if lo != hi && n.Xoffset+n.Type.Width == lo-int64(2*gc.Widthptr) {
// merge with range we already have
lo = n.Xoffset
continue
}
// zero old range
p = zerorange(p, int64(frame), lo, hi, &ax)
// set new range
hi = n.Xoffset + n.Type.Width
lo = n.Xoffset
}
// zero final range
zerorange(p, int64(frame), lo, hi, &ax)
}
/*
* generate
* as n, $c (CMP/CMPU)
*/
func ginscon2(as int, n2 *gc.Node, c int64) {
var n1 gc.Node
gc.Nodconst(&n1, gc.Types[gc.TINT64], c)
switch as {
default:
gc.Fatal("ginscon2")
case ppc64.ACMP:
if -ppc64.BIG <= c && c <= ppc64.BIG {
rawgins(as, n2, &n1)
return
}
case ppc64.ACMPU:
if 0 <= c && c <= 2*ppc64.BIG {
rawgins(as, n2, &n1)
return
}
}
// MOV n1 into register first
var ntmp gc.Node
gc.Regalloc(&ntmp, gc.Types[gc.TINT64], nil)
rawgins(ppc64.AMOVD, &n1, &ntmp)
rawgins(as, n2, &ntmp)
gc.Regfree(&ntmp)
}
func proginfo(p *obj.Prog) {
info := &p.Info
*info = progtable[p.As]
if info.Flags == 0 {
gc.Fatal("unknown instruction %v", p)
}
if (info.Flags&gc.ShiftCX != 0) && p.From.Type != obj.TYPE_CONST {
info.Reguse |= CX
}
if info.Flags&gc.ImulAXDX != 0 {
if p.To.Type == obj.TYPE_NONE {
info.Reguse |= AX
info.Regset |= AX | DX
} else {
info.Flags |= RightRdwr
}
}
// Addressing makes some registers used.
if p.From.Type == obj.TYPE_MEM && p.From.Name == obj.NAME_NONE {
info.Regindex |= RtoB(int(p.From.Reg))
}
if p.From.Index != x86.REG_NONE {
info.Regindex |= RtoB(int(p.From.Index))
}
if p.To.Type == obj.TYPE_MEM && p.To.Name == obj.NAME_NONE {
info.Regindex |= RtoB(int(p.To.Reg))
}
if p.To.Index != x86.REG_NONE {
info.Regindex |= RtoB(int(p.To.Index))
}
}
func proginfo(p *obj.Prog) {
info := &p.Info
*info = progtable[p.As]
if info.Flags == 0 {
gc.Fatal("unknown instruction %v", p)
}
if p.From.Type == obj.TYPE_ADDR && p.From.Sym != nil && (info.Flags&gc.LeftRead != 0) {
info.Flags &^= gc.LeftRead
info.Flags |= gc.LeftAddr
}
if (info.Flags&gc.RegRead != 0) && p.Reg == 0 {
info.Flags &^= gc.RegRead
info.Flags |= gc.CanRegRead | gc.RightRead
}
if (p.Scond&arm.C_SCOND != arm.C_SCOND_NONE) && (info.Flags&gc.RightWrite != 0) {
info.Flags |= gc.RightRead
}
switch p.As {
case arm.ADIV,
arm.ADIVU,
arm.AMOD,
arm.AMODU:
info.Regset |= RtoB(arm.REG_R12)
}
}
func splitclean() {
if nsclean <= 0 {
gc.Fatal("splitclean")
}
nsclean--
if sclean[nsclean].Op != gc.OEMPTY {
gc.Regfree(&sclean[nsclean])
}
}
func gcmp(as int, lhs *gc.Node, rhs *gc.Node) *obj.Prog {
if lhs.Op != gc.OREGISTER {
gc.Fatal("bad operands to gcmp: %v %v", gc.Oconv(int(lhs.Op), 0), gc.Oconv(int(rhs.Op), 0))
}
p := rawgins(as, rhs, nil)
raddr(lhs, p)
return p
}
// as2variant returns the variant (V_*) flags of instruction as.
func as2variant(as int) int {
initvariants()
for i := int(0); i < len(varianttable[as]); i++ {
if varianttable[as][i] == as {
return i
}
}
gc.Fatal("as2variant: instruction %v is not a variant of itself", obj.Aconv(as))
return 0
}
/* generate a constant shift
* arm encodes a shift by 32 as 0, thus asking for 0 shift is illegal.
*/
func gshift(as int, lhs *gc.Node, stype int32, sval int32, rhs *gc.Node) *obj.Prog {
if sval <= 0 || sval > 32 {
gc.Fatal("bad shift value: %d", sval)
}
sval = sval & 0x1f
p := gins(as, nil, rhs)
p.From.Type = obj.TYPE_SHIFT
p.From.Offset = int64(stype) | int64(sval)<<7 | int64(lhs.Reg)&15
return p
}
/*
* direct reference,
* could be set/use depending on
* semantics
*/
func copyas(a *obj.Addr, v *obj.Addr) bool {
if x86.REG_AL <= a.Reg && a.Reg <= x86.REG_R15B {
gc.Fatal("use of byte register")
}
if x86.REG_AL <= v.Reg && v.Reg <= x86.REG_R15B {
gc.Fatal("use of byte register")
}
if a.Type != v.Type || a.Name != v.Name || a.Reg != v.Reg {
return false
}
if regtyp(v) {
return true
}
if v.Type == obj.TYPE_MEM && (v.Name == obj.NAME_AUTO || v.Name == obj.NAME_PARAM) {
if v.Offset == a.Offset {
return true
}
}
return false
}
/*
* generate high multiply
* res = (nl * nr) >> wordsize
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Ullman < nr.Ullman {
tmp := nl
nl = nr
nr = tmp
}
t := nl.Type
w := int(t.Width * 8)
var n1 gc.Node
gc.Regalloc(&n1, t, res)
gc.Cgen(nl, &n1)
var n2 gc.Node
gc.Regalloc(&n2, t, nil)
gc.Cgen(nr, &n2)
switch gc.Simtype[t.Etype] {
case gc.TINT8,
gc.TINT16:
gins(optoas(gc.OMUL, t), &n2, &n1)
gshift(arm.AMOVW, &n1, arm.SHIFT_AR, int32(w), &n1)
case gc.TUINT8,
gc.TUINT16:
gins(optoas(gc.OMUL, t), &n2, &n1)
gshift(arm.AMOVW, &n1, arm.SHIFT_LR, int32(w), &n1)
// perform a long multiplication.
case gc.TINT32,
gc.TUINT32:
var p *obj.Prog
if gc.Issigned[t.Etype] {
p = gins(arm.AMULL, &n2, nil)
} else {
p = gins(arm.AMULLU, &n2, nil)
}
// n2 * n1 -> (n1 n2)
p.Reg = n1.Reg
p.To.Type = obj.TYPE_REGREG
p.To.Reg = n1.Reg
p.To.Offset = int64(n2.Reg)
default:
gc.Fatal("cgen_hmul %v", t)
}
gc.Cgen(&n1, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
/*
* generate high multiply:
* res = (nl*nr) >> width
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
// largest ullman on left.
if nl.Ullman < nr.Ullman {
tmp := (*gc.Node)(nl)
nl = nr
nr = tmp
}
t := (*gc.Type)(nl.Type)
w := int(int(t.Width * 8))
var n1 gc.Node
gc.Cgenr(nl, &n1, res)
var n2 gc.Node
gc.Cgenr(nr, &n2, nil)
switch gc.Simtype[t.Etype] {
case gc.TINT8,
gc.TINT16,
gc.TINT32:
gins(optoas(gc.OMUL, t), &n2, &n1)
p := (*obj.Prog)(gins(ppc64.ASRAD, nil, &n1))
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(w)
case gc.TUINT8,
gc.TUINT16,
gc.TUINT32:
gins(optoas(gc.OMUL, t), &n2, &n1)
p := (*obj.Prog)(gins(ppc64.ASRD, nil, &n1))
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(w)
case gc.TINT64,
gc.TUINT64:
if gc.Issigned[t.Etype] {
gins(ppc64.AMULHD, &n2, &n1)
} else {
gins(ppc64.AMULHDU, &n2, &n1)
}
default:
gc.Fatal("cgen_hmul %v", t)
}
gc.Cgen(&n1, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
func proginfo(p *obj.Prog) {
info := &p.Info
*info = progtable[p.As]
if info.Flags == 0 {
gc.Fatal("unknown instruction %v", p)
}
if (info.Flags&gc.ShiftCX != 0) && p.From.Type != obj.TYPE_CONST {
info.Reguse |= CX
}
if info.Flags&gc.ImulAXDX != 0 {
if p.To.Type == obj.TYPE_NONE {
info.Reguse |= AX
info.Regset |= AX | DX
} else {
info.Flags |= RightRdwr
}
}
// Addressing makes some registers used.
if p.From.Type == obj.TYPE_MEM && p.From.Name == obj.NAME_NONE {
info.Regindex |= RtoB(int(p.From.Reg))
}
if p.From.Index != x86.REG_NONE {
info.Regindex |= RtoB(int(p.From.Index))
}
if p.To.Type == obj.TYPE_MEM && p.To.Name == obj.NAME_NONE {
info.Regindex |= RtoB(int(p.To.Reg))
}
if p.To.Index != x86.REG_NONE {
info.Regindex |= RtoB(int(p.To.Index))
}
if gc.Ctxt.Flag_dynlink {
// When -dynlink is passed, many operations on external names (and
// also calling duffzero/duffcopy) use R15 as a scratch register.
if p.As == x86.ALEAQ || info.Flags == gc.Pseudo || p.As == obj.ACALL || p.As == obj.ARET || p.As == obj.AJMP {
return
}
if p.As == obj.ADUFFZERO || p.As == obj.ADUFFCOPY || (p.From.Name == obj.NAME_EXTERN && !p.From.Sym.Local) || (p.To.Name == obj.NAME_EXTERN && !p.To.Sym.Local) {
info.Reguse |= R15
info.Regset |= R15
return
}
}
}
func proginfo(p *obj.Prog) {
initproginfo()
info := &p.Info
*info = progtable[p.As]
if info.Flags == 0 {
gc.Fatal("proginfo: unknown instruction %v", p)
}
if (info.Flags&gc.RegRead != 0) && p.Reg == 0 {
info.Flags &^= gc.RegRead
info.Flags |= gc.RightRead /*CanRegRead |*/
}
if (p.From.Type == obj.TYPE_MEM || p.From.Type == obj.TYPE_ADDR) && p.From.Reg != 0 {
info.Regindex |= RtoB(int(p.From.Reg))
if info.Flags&gc.PostInc != 0 {
info.Regset |= RtoB(int(p.From.Reg))
}
}
if (p.To.Type == obj.TYPE_MEM || p.To.Type == obj.TYPE_ADDR) && p.To.Reg != 0 {
info.Regindex |= RtoB(int(p.To.Reg))
if info.Flags&gc.PostInc != 0 {
info.Regset |= RtoB(int(p.To.Reg))
}
}
if p.From.Type == obj.TYPE_ADDR && p.From.Sym != nil && (info.Flags&gc.LeftRead != 0) {
info.Flags &^= gc.LeftRead
info.Flags |= gc.LeftAddr
}
if p.As == obj.ADUFFZERO {
info.Reguse |= 1<<0 | RtoB(ppc64.REG_R3)
info.Regset |= RtoB(ppc64.REG_R3)
}
if p.As == obj.ADUFFCOPY {
// TODO(austin) Revisit when duffcopy is implemented
info.Reguse |= RtoB(ppc64.REG_R3) | RtoB(ppc64.REG_R4) | RtoB(ppc64.REG_R5)
info.Regset |= RtoB(ppc64.REG_R3) | RtoB(ppc64.REG_R4)
}
}
/*
* generate division according to op, one of:
* res = nl / nr
* res = nl % nr
*/
func cgen_div(op int, nl *gc.Node, nr *gc.Node, res *gc.Node) {
if gc.Is64(nl.Type) {
gc.Fatal("cgen_div %v", nl.Type)
}
var t *gc.Type
if gc.Issigned[nl.Type.Etype] {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
var ax gc.Node
var oldax gc.Node
savex(x86.REG_AX, &ax, &oldax, res, t)
var olddx gc.Node
var dx gc.Node
savex(x86.REG_DX, &dx, &olddx, res, t)
dodiv(op, nl, nr, res, &ax, &dx)
restx(&dx, &olddx)
restx(&ax, &oldax)
}
// Called after regopt and peep have run.
// Expand CHECKNIL pseudo-op into actual nil pointer check.
func expandchecks(firstp *obj.Prog) {
var p1 *obj.Prog
for p := (*obj.Prog)(firstp); p != nil; p = p.Link {
if gc.Debug_checknil != 0 && gc.Ctxt.Debugvlog != 0 {
fmt.Printf("expandchecks: %v\n", p)
}
if p.As != obj.ACHECKNIL {
continue
}
if gc.Debug_checknil != 0 && p.Lineno > 1 { // p->lineno==1 in generated wrappers
gc.Warnl(int(p.Lineno), "generated nil check")
}
if p.From.Type != obj.TYPE_REG {
gc.Fatal("invalid nil check %v\n", p)
}
// check is
// CBNZ arg, 2(PC)
// MOVD ZR, 0(arg)
p1 = gc.Ctxt.NewProg()
gc.Clearp(p1)
p1.Link = p.Link
p.Link = p1
p1.Lineno = p.Lineno
p1.Pc = 9999
p.As = arm64.ACBNZ
p.To.Type = obj.TYPE_BRANCH
p.To.Val = p1.Link
// crash by write to memory address 0.
p1.As = arm64.AMOVD
p1.From.Type = obj.TYPE_REG
p1.From.Reg = arm64.REGZERO
p1.To.Type = obj.TYPE_MEM
p1.To.Reg = p.From.Reg
p1.To.Offset = 0
}
}
// Called after regopt and peep have run.
// Expand CHECKNIL pseudo-op into actual nil pointer check.
func expandchecks(firstp *obj.Prog) {
var reg int
var p1 *obj.Prog
for p := firstp; p != nil; p = p.Link {
if p.As != obj.ACHECKNIL {
continue
}
if gc.Debug_checknil != 0 && p.Lineno > 1 { // p->lineno==1 in generated wrappers
gc.Warnl(int(p.Lineno), "generated nil check")
}
if p.From.Type != obj.TYPE_REG {
gc.Fatal("invalid nil check %v", p)
}
reg = int(p.From.Reg)
// check is
// CMP arg, $0
// MOV.EQ arg, 0(arg)
p1 = gc.Ctxt.NewProg()
gc.Clearp(p1)
p1.Link = p.Link
p.Link = p1
p1.Lineno = p.Lineno
p1.Pc = 9999
p1.As = arm.AMOVW
p1.From.Type = obj.TYPE_REG
p1.From.Reg = int16(reg)
p1.To.Type = obj.TYPE_MEM
p1.To.Reg = int16(reg)
p1.To.Offset = 0
p1.Scond = arm.C_SCOND_EQ
p.As = arm.ACMP
p.From.Type = obj.TYPE_CONST
p.From.Reg = 0
p.From.Offset = 0
p.Reg = int16(reg)
}
}
// jmptoset returns ASETxx for AJxx.
func jmptoset(jmp int) int {
switch jmp {
case x86.AJEQ:
return x86.ASETEQ
case x86.AJNE:
return x86.ASETNE
case x86.AJLT:
return x86.ASETLT
case x86.AJCS:
return x86.ASETCS
case x86.AJLE:
return x86.ASETLE
case x86.AJLS:
return x86.ASETLS
case x86.AJGT:
return x86.ASETGT
case x86.AJHI:
return x86.ASETHI
case x86.AJGE:
return x86.ASETGE
case x86.AJCC:
return x86.ASETCC
case x86.AJMI:
return x86.ASETMI
case x86.AJOC:
return x86.ASETOC
case x86.AJOS:
return x86.ASETOS
case x86.AJPC:
return x86.ASETPC
case x86.AJPL:
return x86.ASETPL
case x86.AJPS:
return x86.ASETPS
}
gc.Fatal("jmptoset: no entry for %v", gc.Oconv(jmp, 0))
panic("unreachable")
}
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64, ax *uint32) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if *ax == 0 {
p = appendpp(p, x86.AMOVQ, obj.TYPE_CONST, 0, 0, obj.TYPE_REG, x86.REG_AX, 0)
*ax = 1
}
if cnt%int64(gc.Widthreg) != 0 {
// should only happen with nacl
if cnt%int64(gc.Widthptr) != 0 {
gc.Fatal("zerorange count not a multiple of widthptr %d", cnt)
}
p = appendpp(p, x86.AMOVL, obj.TYPE_REG, x86.REG_AX, 0, obj.TYPE_MEM, x86.REG_SP, frame+lo)
lo += int64(gc.Widthptr)
cnt -= int64(gc.Widthptr)
}
if cnt <= int64(4*gc.Widthreg) {
for i := int64(0); i < cnt; i += int64(gc.Widthreg) {
p = appendpp(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_AX, 0, obj.TYPE_MEM, x86.REG_SP, frame+lo+i)
}
} else if !gc.Nacl && (cnt <= int64(128*gc.Widthreg)) {
q := cnt / int64(gc.Widthreg)
p = appendpp(p, leaptr, obj.TYPE_MEM, x86.REG_SP, frame+lo+dzDI(q), obj.TYPE_REG, x86.REG_DI, 0)
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_ADDR, 0, dzOff(q))
p.To.Sym = gc.Linksym(gc.Pkglookup("duffzero", gc.Runtimepkg))
} else {
p = appendpp(p, x86.AMOVQ, obj.TYPE_CONST, 0, cnt/int64(gc.Widthreg), obj.TYPE_REG, x86.REG_CX, 0)
p = appendpp(p, leaptr, obj.TYPE_MEM, x86.REG_SP, frame+lo, obj.TYPE_REG, x86.REG_DI, 0)
p = appendpp(p, x86.AREP, obj.TYPE_NONE, 0, 0, obj.TYPE_NONE, 0, 0)
p = appendpp(p, x86.ASTOSQ, obj.TYPE_NONE, 0, 0, obj.TYPE_NONE, 0, 0)
}
return p
}
func clearfat(nl *gc.Node) {
/* clear a fat object */
if gc.Debug['g'] != 0 {
fmt.Printf("clearfat %v (%v, size: %d)\n", nl, nl.Type, nl.Type.Width)
}
w := uint64(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)
}
// Called after regopt and peep have run.
// Expand CHECKNIL pseudo-op into actual nil pointer check.
func expandchecks(firstp *obj.Prog) {
var p1 *obj.Prog
var p2 *obj.Prog
for p := (*obj.Prog)(firstp); p != nil; p = p.Link {
if gc.Debug_checknil != 0 && gc.Ctxt.Debugvlog != 0 {
fmt.Printf("expandchecks: %v\n", p)
}
if p.As != obj.ACHECKNIL {
continue
}
if gc.Debug_checknil != 0 && p.Lineno > 1 { // p->lineno==1 in generated wrappers
gc.Warnl(int(p.Lineno), "generated nil check")
}
if p.From.Type != obj.TYPE_REG {
gc.Fatal("invalid nil check %v\n", p)
}
/*
// check is
// TD $4, R0, arg (R0 is always zero)
// eqv. to:
// tdeq r0, arg
// NOTE: this needs special runtime support to make SIGTRAP recoverable.
reg = p->from.reg;
p->as = ATD;
p->from = p->to = p->from3 = zprog.from;
p->from.type = TYPE_CONST;
p->from.offset = 4;
p->from.reg = 0;
p->reg = REGZERO;
p->to.type = TYPE_REG;
p->to.reg = reg;
*/
// check is
// CMP arg, R0
// BNE 2(PC) [likely]
// MOVD R0, 0(R0)
p1 = gc.Ctxt.NewProg()
p2 = gc.Ctxt.NewProg()
gc.Clearp(p1)
gc.Clearp(p2)
p1.Link = p2
p2.Link = p.Link
p.Link = p1
p1.Lineno = p.Lineno
p2.Lineno = p.Lineno
p1.Pc = 9999
p2.Pc = 9999
p.As = ppc64.ACMP
p.To.Type = obj.TYPE_REG
p.To.Reg = ppc64.REGZERO
p1.As = ppc64.ABNE
//p1->from.type = TYPE_CONST;
//p1->from.offset = 1; // likely
p1.To.Type = obj.TYPE_BRANCH
p1.To.Val = p2.Link
// crash by write to memory address 0.
p2.As = ppc64.AMOVD
p2.From.Type = obj.TYPE_REG
p2.From.Reg = ppc64.REGZERO
p2.To.Type = obj.TYPE_MEM
p2.To.Reg = ppc64.REGZERO
p2.To.Offset = 0
}
}
/*
* return Axxx for Oxxx on type t.
*/
func optoas(op int, t *gc.Type) int {
if t == nil {
gc.Fatal("optoas: t is nil")
}
a := obj.AXXX
switch uint32(op)<<16 | uint32(gc.Simtype[t.Etype]) {
default:
gc.Fatal("optoas: no entry %v-%v etype %v simtype %v", gc.Oconv(int(op), 0), t, gc.Types[t.Etype], gc.Types[gc.Simtype[t.Etype]])
/* case CASE(OADDR, TPTR32):
a = ALEAL;
break;
case CASE(OADDR, TPTR64):
a = ALEAQ;
break;
*/
// TODO(kaib): make sure the conditional branches work on all edge cases
case gc.OEQ<<16 | gc.TBOOL,
gc.OEQ<<16 | gc.TINT8,
gc.OEQ<<16 | gc.TUINT8,
gc.OEQ<<16 | gc.TINT16,
gc.OEQ<<16 | gc.TUINT16,
gc.OEQ<<16 | gc.TINT32,
gc.OEQ<<16 | gc.TUINT32,
gc.OEQ<<16 | gc.TINT64,
gc.OEQ<<16 | gc.TUINT64,
gc.OEQ<<16 | gc.TPTR32,
gc.OEQ<<16 | gc.TPTR64,
gc.OEQ<<16 | gc.TFLOAT32,
gc.OEQ<<16 | gc.TFLOAT64:
a = arm.ABEQ
case gc.ONE<<16 | gc.TBOOL,
gc.ONE<<16 | gc.TINT8,
gc.ONE<<16 | gc.TUINT8,
gc.ONE<<16 | gc.TINT16,
gc.ONE<<16 | gc.TUINT16,
gc.ONE<<16 | gc.TINT32,
gc.ONE<<16 | gc.TUINT32,
gc.ONE<<16 | gc.TINT64,
gc.ONE<<16 | gc.TUINT64,
gc.ONE<<16 | gc.TPTR32,
gc.ONE<<16 | gc.TPTR64,
gc.ONE<<16 | gc.TFLOAT32,
gc.ONE<<16 | gc.TFLOAT64:
a = arm.ABNE
case gc.OLT<<16 | gc.TINT8,
gc.OLT<<16 | gc.TINT16,
gc.OLT<<16 | gc.TINT32,
gc.OLT<<16 | gc.TINT64,
gc.OLT<<16 | gc.TFLOAT32,
gc.OLT<<16 | gc.TFLOAT64:
a = arm.ABLT
case gc.OLT<<16 | gc.TUINT8,
gc.OLT<<16 | gc.TUINT16,
gc.OLT<<16 | gc.TUINT32,
gc.OLT<<16 | gc.TUINT64:
a = arm.ABLO
case gc.OLE<<16 | gc.TINT8,
gc.OLE<<16 | gc.TINT16,
gc.OLE<<16 | gc.TINT32,
gc.OLE<<16 | gc.TINT64,
gc.OLE<<16 | gc.TFLOAT32,
gc.OLE<<16 | gc.TFLOAT64:
a = arm.ABLE
case gc.OLE<<16 | gc.TUINT8,
gc.OLE<<16 | gc.TUINT16,
gc.OLE<<16 | gc.TUINT32,
gc.OLE<<16 | gc.TUINT64:
a = arm.ABLS
case gc.OGT<<16 | gc.TINT8,
gc.OGT<<16 | gc.TINT16,
gc.OGT<<16 | gc.TINT32,
gc.OGT<<16 | gc.TINT64,
gc.OGT<<16 | gc.TFLOAT32,
gc.OGT<<16 | gc.TFLOAT64:
a = arm.ABGT
case gc.OGT<<16 | gc.TUINT8,
gc.OGT<<16 | gc.TUINT16,
gc.OGT<<16 | gc.TUINT32,
gc.OGT<<16 | gc.TUINT64:
a = arm.ABHI
case gc.OGE<<16 | gc.TINT8,
gc.OGE<<16 | gc.TINT16,
gc.OGE<<16 | gc.TINT32,
gc.OGE<<16 | gc.TINT64,
gc.OGE<<16 | gc.TFLOAT32,
gc.OGE<<16 | gc.TFLOAT64:
a = arm.ABGE
case gc.OGE<<16 | gc.TUINT8,
gc.OGE<<16 | gc.TUINT16,
gc.OGE<<16 | gc.TUINT32,
gc.OGE<<16 | gc.TUINT64:
a = arm.ABHS
case gc.OCMP<<16 | gc.TBOOL,
gc.OCMP<<16 | gc.TINT8,
gc.OCMP<<16 | gc.TUINT8,
gc.OCMP<<16 | gc.TINT16,
gc.OCMP<<16 | gc.TUINT16,
gc.OCMP<<16 | gc.TINT32,
gc.OCMP<<16 | gc.TUINT32,
gc.OCMP<<16 | gc.TPTR32:
a = arm.ACMP
case gc.OCMP<<16 | gc.TFLOAT32:
a = arm.ACMPF
case gc.OCMP<<16 | gc.TFLOAT64:
a = arm.ACMPD
case gc.OPS<<16 | gc.TFLOAT32,
gc.OPS<<16 | gc.TFLOAT64:
a = arm.ABVS
case gc.OAS<<16 | gc.TBOOL:
a = arm.AMOVB
case gc.OAS<<16 | gc.TINT8:
a = arm.AMOVBS
case gc.OAS<<16 | gc.TUINT8:
a = arm.AMOVBU
case gc.OAS<<16 | gc.TINT16:
a = arm.AMOVHS
case gc.OAS<<16 | gc.TUINT16:
a = arm.AMOVHU
case gc.OAS<<16 | gc.TINT32,
gc.OAS<<16 | gc.TUINT32,
gc.OAS<<16 | gc.TPTR32:
a = arm.AMOVW
case gc.OAS<<16 | gc.TFLOAT32:
a = arm.AMOVF
case gc.OAS<<16 | gc.TFLOAT64:
a = arm.AMOVD
case gc.OADD<<16 | gc.TINT8,
gc.OADD<<16 | gc.TUINT8,
gc.OADD<<16 | gc.TINT16,
gc.OADD<<16 | gc.TUINT16,
gc.OADD<<16 | gc.TINT32,
gc.OADD<<16 | gc.TUINT32,
gc.OADD<<16 | gc.TPTR32:
a = arm.AADD
case gc.OADD<<16 | gc.TFLOAT32:
a = arm.AADDF
case gc.OADD<<16 | gc.TFLOAT64:
a = arm.AADDD
case gc.OSUB<<16 | gc.TINT8,
gc.OSUB<<16 | gc.TUINT8,
gc.OSUB<<16 | gc.TINT16,
gc.OSUB<<16 | gc.TUINT16,
gc.OSUB<<16 | gc.TINT32,
gc.OSUB<<16 | gc.TUINT32,
gc.OSUB<<16 | gc.TPTR32:
a = arm.ASUB
case gc.OSUB<<16 | gc.TFLOAT32:
a = arm.ASUBF
case gc.OSUB<<16 | gc.TFLOAT64:
a = arm.ASUBD
case gc.OMINUS<<16 | gc.TINT8,
gc.OMINUS<<16 | gc.TUINT8,
gc.OMINUS<<16 | gc.TINT16,
gc.OMINUS<<16 | gc.TUINT16,
gc.OMINUS<<16 | gc.TINT32,
gc.OMINUS<<16 | gc.TUINT32,
gc.OMINUS<<16 | gc.TPTR32:
a = arm.ARSB
case gc.OAND<<16 | gc.TINT8,
gc.OAND<<16 | gc.TUINT8,
gc.OAND<<16 | gc.TINT16,
gc.OAND<<16 | gc.TUINT16,
gc.OAND<<16 | gc.TINT32,
gc.OAND<<16 | gc.TUINT32,
gc.OAND<<16 | gc.TPTR32:
a = arm.AAND
case gc.OOR<<16 | gc.TINT8,
gc.OOR<<16 | gc.TUINT8,
gc.OOR<<16 | gc.TINT16,
gc.OOR<<16 | gc.TUINT16,
gc.OOR<<16 | gc.TINT32,
gc.OOR<<16 | gc.TUINT32,
gc.OOR<<16 | gc.TPTR32:
a = arm.AORR
case gc.OXOR<<16 | gc.TINT8,
gc.OXOR<<16 | gc.TUINT8,
gc.OXOR<<16 | gc.TINT16,
gc.OXOR<<16 | gc.TUINT16,
gc.OXOR<<16 | gc.TINT32,
gc.OXOR<<16 | gc.TUINT32,
gc.OXOR<<16 | gc.TPTR32:
a = arm.AEOR
case gc.OLSH<<16 | gc.TINT8,
gc.OLSH<<16 | gc.TUINT8,
gc.OLSH<<16 | gc.TINT16,
gc.OLSH<<16 | gc.TUINT16,
gc.OLSH<<16 | gc.TINT32,
gc.OLSH<<16 | gc.TUINT32,
gc.OLSH<<16 | gc.TPTR32:
a = arm.ASLL
case gc.ORSH<<16 | gc.TUINT8,
gc.ORSH<<16 | gc.TUINT16,
gc.ORSH<<16 | gc.TUINT32,
gc.ORSH<<16 | gc.TPTR32:
a = arm.ASRL
case gc.ORSH<<16 | gc.TINT8,
gc.ORSH<<16 | gc.TINT16,
gc.ORSH<<16 | gc.TINT32:
a = arm.ASRA
case gc.OMUL<<16 | gc.TUINT8,
gc.OMUL<<16 | gc.TUINT16,
gc.OMUL<<16 | gc.TUINT32,
gc.OMUL<<16 | gc.TPTR32:
a = arm.AMULU
case gc.OMUL<<16 | gc.TINT8,
gc.OMUL<<16 | gc.TINT16,
gc.OMUL<<16 | gc.TINT32:
a = arm.AMUL
case gc.OMUL<<16 | gc.TFLOAT32:
a = arm.AMULF
case gc.OMUL<<16 | gc.TFLOAT64:
a = arm.AMULD
case gc.ODIV<<16 | gc.TUINT8,
gc.ODIV<<16 | gc.TUINT16,
gc.ODIV<<16 | gc.TUINT32,
gc.ODIV<<16 | gc.TPTR32:
a = arm.ADIVU
case gc.ODIV<<16 | gc.TINT8,
gc.ODIV<<16 | gc.TINT16,
gc.ODIV<<16 | gc.TINT32:
a = arm.ADIV
case gc.OMOD<<16 | gc.TUINT8,
gc.OMOD<<16 | gc.TUINT16,
gc.OMOD<<16 | gc.TUINT32,
gc.OMOD<<16 | gc.TPTR32:
a = arm.AMODU
case gc.OMOD<<16 | gc.TINT8,
gc.OMOD<<16 | gc.TINT16,
gc.OMOD<<16 | gc.TINT32:
a = arm.AMOD
// case CASE(OEXTEND, TINT16):
// a = ACWD;
// break;
// case CASE(OEXTEND, TINT32):
// a = ACDQ;
// break;
// case CASE(OEXTEND, TINT64):
// a = ACQO;
// break;
case gc.ODIV<<16 | gc.TFLOAT32:
a = arm.ADIVF
case gc.ODIV<<16 | gc.TFLOAT64:
a = arm.ADIVD
case gc.OSQRT<<16 | gc.TFLOAT64:
a = arm.ASQRTD
}
return a
}
/*
* generate one instruction:
* as f, t
*/
func gins(as int, f *gc.Node, t *gc.Node) *obj.Prog {
// Node nod;
// if(f != N && f->op == OINDEX) {
// gc.Regalloc(&nod, ®node, Z);
// v = constnode.vconst;
// gc.Cgen(f->right, &nod);
// constnode.vconst = v;
// idx.reg = nod.reg;
// gc.Regfree(&nod);
// }
// if(t != N && t->op == OINDEX) {
// gc.Regalloc(&nod, ®node, Z);
// v = constnode.vconst;
// gc.Cgen(t->right, &nod);
// constnode.vconst = v;
// idx.reg = nod.reg;
// gc.Regfree(&nod);
// }
if f != nil && f.Op == gc.OADDR && (as == x86.AMOVL || as == x86.AMOVQ) {
// Turn MOVL $xxx into LEAL xxx.
// These should be equivalent but most of the backend
// only expects to see LEAL, because that's what we had
// historically generated. Various hidden assumptions are baked in by now.
if as == x86.AMOVL {
as = x86.ALEAL
} else {
as = x86.ALEAQ
}
f = f.Left
}
switch as {
case x86.AMOVB,
x86.AMOVW,
x86.AMOVL,
x86.AMOVQ,
x86.AMOVSS,
x86.AMOVSD:
if f != nil && t != nil && samaddr(f, t) {
return nil
}
case x86.ALEAQ:
if f != nil && gc.Isconst(f, gc.CTNIL) {
gc.Fatal("gins LEAQ nil %v", f.Type)
}
}
p := gc.Prog(as)
gc.Naddr(&p.From, f)
gc.Naddr(&p.To, t)
if gc.Debug['g'] != 0 {
fmt.Printf("%v\n", p)
}
w := int32(0)
switch as {
case x86.AMOVB:
w = 1
case x86.AMOVW:
w = 2
case x86.AMOVL:
w = 4
case x86.AMOVQ:
w = 8
}
if w != 0 && ((f != nil && p.From.Width < int64(w)) || (t != nil && p.To.Width > int64(w))) {
gc.Dump("f", f)
gc.Dump("t", t)
gc.Fatal("bad width: %v (%d, %d)\n", p, p.From.Width, p.To.Width)
}
if p.To.Type == obj.TYPE_ADDR && w > 0 {
gc.Fatal("bad use of addr: %v", p)
}
return p
}
/*
* return Axxx for Oxxx on type t.
*/
func optoas(op int, t *gc.Type) int {
if t == nil {
gc.Fatal("optoas: t is nil")
}
a := obj.AXXX
switch uint32(op)<<16 | uint32(gc.Simtype[t.Etype]) {
default:
gc.Fatal("optoas: no entry %v-%v", gc.Oconv(int(op), 0), t)
case gc.OADDR<<16 | gc.TPTR32:
a = x86.ALEAL
case gc.OADDR<<16 | gc.TPTR64:
a = x86.ALEAQ
case gc.OEQ<<16 | gc.TBOOL,
gc.OEQ<<16 | gc.TINT8,
gc.OEQ<<16 | gc.TUINT8,
gc.OEQ<<16 | gc.TINT16,
gc.OEQ<<16 | gc.TUINT16,
gc.OEQ<<16 | gc.TINT32,
gc.OEQ<<16 | gc.TUINT32,
gc.OEQ<<16 | gc.TINT64,
gc.OEQ<<16 | gc.TUINT64,
gc.OEQ<<16 | gc.TPTR32,
gc.OEQ<<16 | gc.TPTR64,
gc.OEQ<<16 | gc.TFLOAT32,
gc.OEQ<<16 | gc.TFLOAT64:
a = x86.AJEQ
case gc.ONE<<16 | gc.TBOOL,
gc.ONE<<16 | gc.TINT8,
gc.ONE<<16 | gc.TUINT8,
gc.ONE<<16 | gc.TINT16,
gc.ONE<<16 | gc.TUINT16,
gc.ONE<<16 | gc.TINT32,
gc.ONE<<16 | gc.TUINT32,
gc.ONE<<16 | gc.TINT64,
gc.ONE<<16 | gc.TUINT64,
gc.ONE<<16 | gc.TPTR32,
gc.ONE<<16 | gc.TPTR64,
gc.ONE<<16 | gc.TFLOAT32,
gc.ONE<<16 | gc.TFLOAT64:
a = x86.AJNE
case gc.OPS<<16 | gc.TBOOL,
gc.OPS<<16 | gc.TINT8,
gc.OPS<<16 | gc.TUINT8,
gc.OPS<<16 | gc.TINT16,
gc.OPS<<16 | gc.TUINT16,
gc.OPS<<16 | gc.TINT32,
gc.OPS<<16 | gc.TUINT32,
gc.OPS<<16 | gc.TINT64,
gc.OPS<<16 | gc.TUINT64,
gc.OPS<<16 | gc.TPTR32,
gc.OPS<<16 | gc.TPTR64,
gc.OPS<<16 | gc.TFLOAT32,
gc.OPS<<16 | gc.TFLOAT64:
a = x86.AJPS
case gc.OPC<<16 | gc.TBOOL,
gc.OPC<<16 | gc.TINT8,
gc.OPC<<16 | gc.TUINT8,
gc.OPC<<16 | gc.TINT16,
gc.OPC<<16 | gc.TUINT16,
gc.OPC<<16 | gc.TINT32,
gc.OPC<<16 | gc.TUINT32,
gc.OPC<<16 | gc.TINT64,
gc.OPC<<16 | gc.TUINT64,
gc.OPC<<16 | gc.TPTR32,
gc.OPC<<16 | gc.TPTR64,
gc.OPC<<16 | gc.TFLOAT32,
gc.OPC<<16 | gc.TFLOAT64:
a = x86.AJPC
case gc.OLT<<16 | gc.TINT8,
gc.OLT<<16 | gc.TINT16,
gc.OLT<<16 | gc.TINT32,
gc.OLT<<16 | gc.TINT64:
a = x86.AJLT
case gc.OLT<<16 | gc.TUINT8,
gc.OLT<<16 | gc.TUINT16,
gc.OLT<<16 | gc.TUINT32,
gc.OLT<<16 | gc.TUINT64:
a = x86.AJCS
case gc.OLE<<16 | gc.TINT8,
gc.OLE<<16 | gc.TINT16,
gc.OLE<<16 | gc.TINT32,
gc.OLE<<16 | gc.TINT64:
a = x86.AJLE
case gc.OLE<<16 | gc.TUINT8,
gc.OLE<<16 | gc.TUINT16,
gc.OLE<<16 | gc.TUINT32,
gc.OLE<<16 | gc.TUINT64:
a = x86.AJLS
case gc.OGT<<16 | gc.TINT8,
gc.OGT<<16 | gc.TINT16,
gc.OGT<<16 | gc.TINT32,
gc.OGT<<16 | gc.TINT64:
a = x86.AJGT
case gc.OGT<<16 | gc.TUINT8,
gc.OGT<<16 | gc.TUINT16,
gc.OGT<<16 | gc.TUINT32,
gc.OGT<<16 | gc.TUINT64,
gc.OLT<<16 | gc.TFLOAT32,
gc.OLT<<16 | gc.TFLOAT64:
a = x86.AJHI
case gc.OGE<<16 | gc.TINT8,
gc.OGE<<16 | gc.TINT16,
gc.OGE<<16 | gc.TINT32,
gc.OGE<<16 | gc.TINT64:
a = x86.AJGE
case gc.OGE<<16 | gc.TUINT8,
gc.OGE<<16 | gc.TUINT16,
gc.OGE<<16 | gc.TUINT32,
gc.OGE<<16 | gc.TUINT64,
gc.OLE<<16 | gc.TFLOAT32,
gc.OLE<<16 | gc.TFLOAT64:
a = x86.AJCC
case gc.OCMP<<16 | gc.TBOOL,
gc.OCMP<<16 | gc.TINT8,
gc.OCMP<<16 | gc.TUINT8:
a = x86.ACMPB
case gc.OCMP<<16 | gc.TINT16,
gc.OCMP<<16 | gc.TUINT16:
a = x86.ACMPW
case gc.OCMP<<16 | gc.TINT32,
gc.OCMP<<16 | gc.TUINT32,
gc.OCMP<<16 | gc.TPTR32:
a = x86.ACMPL
case gc.OCMP<<16 | gc.TINT64,
gc.OCMP<<16 | gc.TUINT64,
gc.OCMP<<16 | gc.TPTR64:
a = x86.ACMPQ
case gc.OCMP<<16 | gc.TFLOAT32:
a = x86.AUCOMISS
case gc.OCMP<<16 | gc.TFLOAT64:
a = x86.AUCOMISD
case gc.OAS<<16 | gc.TBOOL,
gc.OAS<<16 | gc.TINT8,
gc.OAS<<16 | gc.TUINT8:
a = x86.AMOVB
case gc.OAS<<16 | gc.TINT16,
gc.OAS<<16 | gc.TUINT16:
a = x86.AMOVW
case gc.OAS<<16 | gc.TINT32,
gc.OAS<<16 | gc.TUINT32,
gc.OAS<<16 | gc.TPTR32:
a = x86.AMOVL
case gc.OAS<<16 | gc.TINT64,
gc.OAS<<16 | gc.TUINT64,
gc.OAS<<16 | gc.TPTR64:
a = x86.AMOVQ
case gc.OAS<<16 | gc.TFLOAT32:
a = x86.AMOVSS
case gc.OAS<<16 | gc.TFLOAT64:
a = x86.AMOVSD
case gc.OADD<<16 | gc.TINT8,
gc.OADD<<16 | gc.TUINT8:
a = x86.AADDB
case gc.OADD<<16 | gc.TINT16,
gc.OADD<<16 | gc.TUINT16:
a = x86.AADDW
case gc.OADD<<16 | gc.TINT32,
gc.OADD<<16 | gc.TUINT32,
gc.OADD<<16 | gc.TPTR32:
a = x86.AADDL
case gc.OADD<<16 | gc.TINT64,
gc.OADD<<16 | gc.TUINT64,
gc.OADD<<16 | gc.TPTR64:
a = x86.AADDQ
case gc.OADD<<16 | gc.TFLOAT32:
a = x86.AADDSS
case gc.OADD<<16 | gc.TFLOAT64:
a = x86.AADDSD
case gc.OSUB<<16 | gc.TINT8,
gc.OSUB<<16 | gc.TUINT8:
a = x86.ASUBB
case gc.OSUB<<16 | gc.TINT16,
gc.OSUB<<16 | gc.TUINT16:
a = x86.ASUBW
case gc.OSUB<<16 | gc.TINT32,
gc.OSUB<<16 | gc.TUINT32,
gc.OSUB<<16 | gc.TPTR32:
a = x86.ASUBL
case gc.OSUB<<16 | gc.TINT64,
gc.OSUB<<16 | gc.TUINT64,
gc.OSUB<<16 | gc.TPTR64:
a = x86.ASUBQ
case gc.OSUB<<16 | gc.TFLOAT32:
a = x86.ASUBSS
case gc.OSUB<<16 | gc.TFLOAT64:
a = x86.ASUBSD
case gc.OINC<<16 | gc.TINT8,
gc.OINC<<16 | gc.TUINT8:
a = x86.AINCB
case gc.OINC<<16 | gc.TINT16,
gc.OINC<<16 | gc.TUINT16:
a = x86.AINCW
case gc.OINC<<16 | gc.TINT32,
gc.OINC<<16 | gc.TUINT32,
gc.OINC<<16 | gc.TPTR32:
a = x86.AINCL
case gc.OINC<<16 | gc.TINT64,
gc.OINC<<16 | gc.TUINT64,
gc.OINC<<16 | gc.TPTR64:
a = x86.AINCQ
case gc.ODEC<<16 | gc.TINT8,
gc.ODEC<<16 | gc.TUINT8:
a = x86.ADECB
case gc.ODEC<<16 | gc.TINT16,
gc.ODEC<<16 | gc.TUINT16:
a = x86.ADECW
case gc.ODEC<<16 | gc.TINT32,
gc.ODEC<<16 | gc.TUINT32,
gc.ODEC<<16 | gc.TPTR32:
a = x86.ADECL
case gc.ODEC<<16 | gc.TINT64,
gc.ODEC<<16 | gc.TUINT64,
gc.ODEC<<16 | gc.TPTR64:
a = x86.ADECQ
case gc.OMINUS<<16 | gc.TINT8,
gc.OMINUS<<16 | gc.TUINT8:
a = x86.ANEGB
case gc.OMINUS<<16 | gc.TINT16,
gc.OMINUS<<16 | gc.TUINT16:
a = x86.ANEGW
case gc.OMINUS<<16 | gc.TINT32,
gc.OMINUS<<16 | gc.TUINT32,
gc.OMINUS<<16 | gc.TPTR32:
a = x86.ANEGL
case gc.OMINUS<<16 | gc.TINT64,
gc.OMINUS<<16 | gc.TUINT64,
gc.OMINUS<<16 | gc.TPTR64:
a = x86.ANEGQ
case gc.OAND<<16 | gc.TBOOL,
gc.OAND<<16 | gc.TINT8,
gc.OAND<<16 | gc.TUINT8:
a = x86.AANDB
case gc.OAND<<16 | gc.TINT16,
gc.OAND<<16 | gc.TUINT16:
a = x86.AANDW
case gc.OAND<<16 | gc.TINT32,
gc.OAND<<16 | gc.TUINT32,
gc.OAND<<16 | gc.TPTR32:
a = x86.AANDL
case gc.OAND<<16 | gc.TINT64,
gc.OAND<<16 | gc.TUINT64,
gc.OAND<<16 | gc.TPTR64:
a = x86.AANDQ
case gc.OOR<<16 | gc.TBOOL,
gc.OOR<<16 | gc.TINT8,
gc.OOR<<16 | gc.TUINT8:
a = x86.AORB
case gc.OOR<<16 | gc.TINT16,
gc.OOR<<16 | gc.TUINT16:
a = x86.AORW
case gc.OOR<<16 | gc.TINT32,
gc.OOR<<16 | gc.TUINT32,
gc.OOR<<16 | gc.TPTR32:
a = x86.AORL
case gc.OOR<<16 | gc.TINT64,
gc.OOR<<16 | gc.TUINT64,
gc.OOR<<16 | gc.TPTR64:
a = x86.AORQ
case gc.OXOR<<16 | gc.TINT8,
gc.OXOR<<16 | gc.TUINT8:
a = x86.AXORB
case gc.OXOR<<16 | gc.TINT16,
gc.OXOR<<16 | gc.TUINT16:
a = x86.AXORW
case gc.OXOR<<16 | gc.TINT32,
gc.OXOR<<16 | gc.TUINT32,
gc.OXOR<<16 | gc.TPTR32:
a = x86.AXORL
case gc.OXOR<<16 | gc.TINT64,
gc.OXOR<<16 | gc.TUINT64,
gc.OXOR<<16 | gc.TPTR64:
a = x86.AXORQ
case gc.OLROT<<16 | gc.TINT8,
gc.OLROT<<16 | gc.TUINT8:
a = x86.AROLB
case gc.OLROT<<16 | gc.TINT16,
gc.OLROT<<16 | gc.TUINT16:
a = x86.AROLW
case gc.OLROT<<16 | gc.TINT32,
gc.OLROT<<16 | gc.TUINT32,
gc.OLROT<<16 | gc.TPTR32:
a = x86.AROLL
case gc.OLROT<<16 | gc.TINT64,
gc.OLROT<<16 | gc.TUINT64,
gc.OLROT<<16 | gc.TPTR64:
a = x86.AROLQ
case gc.OLSH<<16 | gc.TINT8,
gc.OLSH<<16 | gc.TUINT8:
a = x86.ASHLB
case gc.OLSH<<16 | gc.TINT16,
gc.OLSH<<16 | gc.TUINT16:
a = x86.ASHLW
case gc.OLSH<<16 | gc.TINT32,
gc.OLSH<<16 | gc.TUINT32,
gc.OLSH<<16 | gc.TPTR32:
a = x86.ASHLL
case gc.OLSH<<16 | gc.TINT64,
gc.OLSH<<16 | gc.TUINT64,
gc.OLSH<<16 | gc.TPTR64:
a = x86.ASHLQ
case gc.ORSH<<16 | gc.TUINT8:
a = x86.ASHRB
case gc.ORSH<<16 | gc.TUINT16:
a = x86.ASHRW
case gc.ORSH<<16 | gc.TUINT32,
gc.ORSH<<16 | gc.TPTR32:
a = x86.ASHRL
case gc.ORSH<<16 | gc.TUINT64,
gc.ORSH<<16 | gc.TPTR64:
a = x86.ASHRQ
case gc.ORSH<<16 | gc.TINT8:
a = x86.ASARB
case gc.ORSH<<16 | gc.TINT16:
a = x86.ASARW
case gc.ORSH<<16 | gc.TINT32:
a = x86.ASARL
case gc.ORSH<<16 | gc.TINT64:
a = x86.ASARQ
case gc.ORROTC<<16 | gc.TINT8,
gc.ORROTC<<16 | gc.TUINT8:
a = x86.ARCRB
case gc.ORROTC<<16 | gc.TINT16,
gc.ORROTC<<16 | gc.TUINT16:
a = x86.ARCRW
case gc.ORROTC<<16 | gc.TINT32,
gc.ORROTC<<16 | gc.TUINT32:
a = x86.ARCRL
case gc.ORROTC<<16 | gc.TINT64,
gc.ORROTC<<16 | gc.TUINT64:
a = x86.ARCRQ
case gc.OHMUL<<16 | gc.TINT8,
gc.OMUL<<16 | gc.TINT8,
gc.OMUL<<16 | gc.TUINT8:
a = x86.AIMULB
case gc.OHMUL<<16 | gc.TINT16,
gc.OMUL<<16 | gc.TINT16,
gc.OMUL<<16 | gc.TUINT16:
a = x86.AIMULW
case gc.OHMUL<<16 | gc.TINT32,
gc.OMUL<<16 | gc.TINT32,
gc.OMUL<<16 | gc.TUINT32,
gc.OMUL<<16 | gc.TPTR32:
a = x86.AIMULL
case gc.OHMUL<<16 | gc.TINT64,
gc.OMUL<<16 | gc.TINT64,
gc.OMUL<<16 | gc.TUINT64,
gc.OMUL<<16 | gc.TPTR64:
a = x86.AIMULQ
case gc.OHMUL<<16 | gc.TUINT8:
a = x86.AMULB
case gc.OHMUL<<16 | gc.TUINT16:
a = x86.AMULW
case gc.OHMUL<<16 | gc.TUINT32,
gc.OHMUL<<16 | gc.TPTR32:
a = x86.AMULL
case gc.OHMUL<<16 | gc.TUINT64,
gc.OHMUL<<16 | gc.TPTR64:
a = x86.AMULQ
case gc.OMUL<<16 | gc.TFLOAT32:
a = x86.AMULSS
case gc.OMUL<<16 | gc.TFLOAT64:
a = x86.AMULSD
case gc.ODIV<<16 | gc.TINT8,
gc.OMOD<<16 | gc.TINT8:
a = x86.AIDIVB
case gc.ODIV<<16 | gc.TUINT8,
gc.OMOD<<16 | gc.TUINT8:
a = x86.ADIVB
case gc.ODIV<<16 | gc.TINT16,
gc.OMOD<<16 | gc.TINT16:
a = x86.AIDIVW
case gc.ODIV<<16 | gc.TUINT16,
gc.OMOD<<16 | gc.TUINT16:
a = x86.ADIVW
case gc.ODIV<<16 | gc.TINT32,
gc.OMOD<<16 | gc.TINT32:
a = x86.AIDIVL
case gc.ODIV<<16 | gc.TUINT32,
gc.ODIV<<16 | gc.TPTR32,
gc.OMOD<<16 | gc.TUINT32,
gc.OMOD<<16 | gc.TPTR32:
a = x86.ADIVL
case gc.ODIV<<16 | gc.TINT64,
gc.OMOD<<16 | gc.TINT64:
a = x86.AIDIVQ
case gc.ODIV<<16 | gc.TUINT64,
gc.ODIV<<16 | gc.TPTR64,
gc.OMOD<<16 | gc.TUINT64,
gc.OMOD<<16 | gc.TPTR64:
a = x86.ADIVQ
case gc.OEXTEND<<16 | gc.TINT16:
a = x86.ACWD
case gc.OEXTEND<<16 | gc.TINT32:
a = x86.ACDQ
case gc.OEXTEND<<16 | gc.TINT64:
a = x86.ACQO
case gc.ODIV<<16 | gc.TFLOAT32:
a = x86.ADIVSS
case gc.ODIV<<16 | gc.TFLOAT64:
a = x86.ADIVSD
case gc.OSQRT<<16 | gc.TFLOAT64:
a = x86.ASQRTSD
}
return a
}
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 code to compute address of n,
* a reference to a (perhaps nested) field inside
* an array or struct.
* return 0 on failure, 1 on success.
* on success, leaves usable address in a.
*
* caller is responsible for calling sudoclean
* after successful sudoaddable,
* to release the register used for a.
*/
func sudoaddable(as int, n *gc.Node, a *obj.Addr) bool {
if n.Type == nil {
return false
}
*a = obj.Addr{}
switch n.Op {
case gc.OLITERAL:
if !gc.Isconst(n, gc.CTINT) {
break
}
v := n.Int()
if v >= 32000 || v <= -32000 {
break
}
switch as {
default:
return false
case arm.AADD,
arm.ASUB,
arm.AAND,
arm.AORR,
arm.AEOR,
arm.AMOVB,
arm.AMOVBS,
arm.AMOVBU,
arm.AMOVH,
arm.AMOVHS,
arm.AMOVHU,
arm.AMOVW:
break
}
cleani += 2
reg := &clean[cleani-1]
reg1 := &clean[cleani-2]
reg.Op = gc.OEMPTY
reg1.Op = gc.OEMPTY
gc.Naddr(a, n)
return true
case gc.ODOT,
gc.ODOTPTR:
cleani += 2
reg := &clean[cleani-1]
reg1 := &clean[cleani-2]
reg.Op = gc.OEMPTY
reg1.Op = gc.OEMPTY
var nn *gc.Node
var oary [10]int64
o := gc.Dotoffset(n, oary[:], &nn)
if nn == nil {
sudoclean()
return false
}
if nn.Addable && o == 1 && oary[0] >= 0 {
// directly addressable set of DOTs
n1 := *nn
n1.Type = n.Type
n1.Xoffset += oary[0]
gc.Naddr(a, &n1)
return true
}
gc.Regalloc(reg, gc.Types[gc.Tptr], nil)
n1 := *reg
n1.Op = gc.OINDREG
if oary[0] >= 0 {
gc.Agen(nn, reg)
n1.Xoffset = oary[0]
} else {
gc.Cgen(nn, reg)
gc.Cgen_checknil(reg)
n1.Xoffset = -(oary[0] + 1)
}
for i := 1; i < o; i++ {
if oary[i] >= 0 {
gc.Fatal("can't happen")
}
gins(arm.AMOVW, &n1, reg)
gc.Cgen_checknil(reg)
n1.Xoffset = -(oary[i] + 1)
}
a.Type = obj.TYPE_NONE
a.Name = obj.NAME_NONE
n1.Type = n.Type
gc.Naddr(a, &n1)
return true
case gc.OINDEX:
return false
}
return false
}
/*
* generate one instruction:
* as f, t
*/
func gins(as int, f *gc.Node, t *gc.Node) *obj.Prog {
// Node nod;
// int32 v;
if f != nil && f.Op == gc.OINDEX {
gc.Fatal("gins OINDEX not implemented")
}
// gc.Regalloc(&nod, ®node, Z);
// v = constnode.vconst;
// gc.Cgen(f->right, &nod);
// constnode.vconst = v;
// idx.reg = nod.reg;
// gc.Regfree(&nod);
if t != nil && t.Op == gc.OINDEX {
gc.Fatal("gins OINDEX not implemented")
}
// gc.Regalloc(&nod, ®node, Z);
// v = constnode.vconst;
// gc.Cgen(t->right, &nod);
// constnode.vconst = v;
// idx.reg = nod.reg;
// gc.Regfree(&nod);
p := gc.Prog(as)
gc.Naddr(&p.From, f)
gc.Naddr(&p.To, t)
switch as {
case arm.ABL:
if p.To.Type == obj.TYPE_REG {
p.To.Type = obj.TYPE_MEM
}
case arm.ACMP, arm.ACMPF, arm.ACMPD:
if t != nil {
if f.Op != gc.OREGISTER {
/* generate a comparison
TODO(kaib): one of the args can actually be a small constant. relax the constraint and fix call sites.
*/
gc.Fatal("bad operands to gcmp")
}
p.From = p.To
p.To = obj.Addr{}
raddr(f, p)
}
case arm.AMULU:
if f != nil && f.Op != gc.OREGISTER {
gc.Fatal("bad operands to mul")
}
case arm.AMOVW:
if (p.From.Type == obj.TYPE_MEM || p.From.Type == obj.TYPE_ADDR || p.From.Type == obj.TYPE_CONST) && (p.To.Type == obj.TYPE_MEM || p.To.Type == obj.TYPE_ADDR) {
gc.Fatal("gins double memory")
}
case arm.AADD:
if p.To.Type == obj.TYPE_MEM {
gc.Fatal("gins arith to mem")
}
case arm.ARSB:
if p.From.Type == obj.TYPE_NONE {
gc.Fatal("rsb with no from")
}
}
if gc.Debug['g'] != 0 {
fmt.Printf("%v\n", p)
}
return p
}
func gmove(f *gc.Node, t *gc.Node) {
if gc.Debug['M'] != 0 {
fmt.Printf("gmove %v -> %v\n", f, t)
}
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;
// except 64-bit, which always copies via registers anyway.
var a int
var r1 gc.Node
if !gc.Is64(f.Type) && !gc.Is64(t.Type) && gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
switch tt {
default:
f.Convconst(&con, t.Type)
case gc.TINT16,
gc.TINT8:
var con gc.Node
f.Convconst(&con, gc.Types[gc.TINT32])
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(arm.AMOVW, &con, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
case gc.TUINT16,
gc.TUINT8:
var con gc.Node
f.Convconst(&con, gc.Types[gc.TUINT32])
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(arm.AMOVW, &con, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
f = &con
ft = gc.Simsimtype(con.Type)
// constants can't move directly to memory
if gc.Ismem(t) && !gc.Is64(t.Type) {
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:
// should not happen
gc.Fatal("gmove %v -> %v", f, t)
return
/*
* integer copy and truncate
*/
case gc.TINT8<<16 | gc.TINT8: // same size
if !gc.Ismem(f) {
a = arm.AMOVB
break
}
fallthrough
case 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:
a = arm.AMOVBS
case gc.TUINT8<<16 | gc.TUINT8:
if !gc.Ismem(f) {
a = arm.AMOVB
break
}
fallthrough
case gc.TINT8<<16 | gc.TUINT8,
gc.TINT16<<16 | gc.TUINT8,
gc.TUINT16<<16 | gc.TUINT8,
gc.TINT32<<16 | gc.TUINT8,
gc.TUINT32<<16 | gc.TUINT8:
a = arm.AMOVBU
case gc.TINT64<<16 | gc.TINT8, // truncate low word
gc.TUINT64<<16 | gc.TINT8:
a = arm.AMOVBS
goto trunc64
case gc.TINT64<<16 | gc.TUINT8,
gc.TUINT64<<16 | gc.TUINT8:
a = arm.AMOVBU
goto trunc64
case gc.TINT16<<16 | gc.TINT16: // same size
if !gc.Ismem(f) {
a = arm.AMOVH
break
}
fallthrough
case gc.TUINT16<<16 | gc.TINT16,
gc.TINT32<<16 | gc.TINT16, // truncate
gc.TUINT32<<16 | gc.TINT16:
a = arm.AMOVHS
case gc.TUINT16<<16 | gc.TUINT16:
if !gc.Ismem(f) {
a = arm.AMOVH
break
}
fallthrough
case gc.TINT16<<16 | gc.TUINT16,
gc.TINT32<<16 | gc.TUINT16,
gc.TUINT32<<16 | gc.TUINT16:
a = arm.AMOVHU
case gc.TINT64<<16 | gc.TINT16, // truncate low word
gc.TUINT64<<16 | gc.TINT16:
a = arm.AMOVHS
goto trunc64
case gc.TINT64<<16 | gc.TUINT16,
gc.TUINT64<<16 | gc.TUINT16:
a = arm.AMOVHU
goto trunc64
case gc.TINT32<<16 | gc.TINT32, // same size
gc.TINT32<<16 | gc.TUINT32,
gc.TUINT32<<16 | gc.TINT32,
gc.TUINT32<<16 | gc.TUINT32:
a = arm.AMOVW
case gc.TINT64<<16 | gc.TINT32, // truncate
gc.TUINT64<<16 | gc.TINT32,
gc.TINT64<<16 | gc.TUINT32,
gc.TUINT64<<16 | gc.TUINT32:
var flo gc.Node
var fhi gc.Node
split64(f, &flo, &fhi)
var r1 gc.Node
gc.Regalloc(&r1, t.Type, nil)
gins(arm.AMOVW, &flo, &r1)
gins(arm.AMOVW, &r1, t)
gc.Regfree(&r1)
splitclean()
return
case gc.TINT64<<16 | gc.TINT64, // same size
gc.TINT64<<16 | gc.TUINT64,
gc.TUINT64<<16 | gc.TINT64,
gc.TUINT64<<16 | gc.TUINT64:
var fhi gc.Node
var flo gc.Node
split64(f, &flo, &fhi)
var tlo gc.Node
var thi gc.Node
split64(t, &tlo, &thi)
var r1 gc.Node
gc.Regalloc(&r1, flo.Type, nil)
var r2 gc.Node
gc.Regalloc(&r2, fhi.Type, nil)
gins(arm.AMOVW, &flo, &r1)
gins(arm.AMOVW, &fhi, &r2)
gins(arm.AMOVW, &r1, &tlo)
gins(arm.AMOVW, &r2, &thi)
gc.Regfree(&r1)
gc.Regfree(&r2)
splitclean()
splitclean()
return
/*
* 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:
a = arm.AMOVBS
goto rdst
case gc.TINT8<<16 | gc.TINT64, // convert via int32
gc.TINT8<<16 | gc.TUINT64:
cvt = gc.Types[gc.TINT32]
goto hard
case gc.TUINT8<<16 | gc.TINT16, // zero extend uint8
gc.TUINT8<<16 | gc.TUINT16,
gc.TUINT8<<16 | gc.TINT32,
gc.TUINT8<<16 | gc.TUINT32:
a = arm.AMOVBU
goto rdst
case gc.TUINT8<<16 | gc.TINT64, // convert via uint32
gc.TUINT8<<16 | gc.TUINT64:
cvt = gc.Types[gc.TUINT32]
goto hard
case gc.TINT16<<16 | gc.TINT32, // sign extend int16
gc.TINT16<<16 | gc.TUINT32:
a = arm.AMOVHS
goto rdst
case gc.TINT16<<16 | gc.TINT64, // convert via int32
gc.TINT16<<16 | gc.TUINT64:
cvt = gc.Types[gc.TINT32]
goto hard
case gc.TUINT16<<16 | gc.TINT32, // zero extend uint16
gc.TUINT16<<16 | gc.TUINT32:
a = arm.AMOVHU
goto rdst
case gc.TUINT16<<16 | gc.TINT64, // convert via uint32
gc.TUINT16<<16 | gc.TUINT64:
cvt = gc.Types[gc.TUINT32]
goto hard
case gc.TINT32<<16 | gc.TINT64, // sign extend int32
gc.TINT32<<16 | gc.TUINT64:
var tlo gc.Node
var thi gc.Node
split64(t, &tlo, &thi)
var r1 gc.Node
gc.Regalloc(&r1, tlo.Type, nil)
var r2 gc.Node
gc.Regalloc(&r2, thi.Type, nil)
gmove(f, &r1)
p1 := gins(arm.AMOVW, &r1, &r2)
p1.From.Type = obj.TYPE_SHIFT
p1.From.Offset = 2<<5 | 31<<7 | int64(r1.Reg)&15 // r1->31
p1.From.Reg = 0
//print("gmove: %v\n", p1);
gins(arm.AMOVW, &r1, &tlo)
gins(arm.AMOVW, &r2, &thi)
gc.Regfree(&r1)
gc.Regfree(&r2)
splitclean()
return
case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
gc.TUINT32<<16 | gc.TUINT64:
var thi gc.Node
var tlo gc.Node
split64(t, &tlo, &thi)
gmove(f, &tlo)
var r1 gc.Node
gc.Regalloc(&r1, thi.Type, nil)
gins(arm.AMOVW, ncon(0), &r1)
gins(arm.AMOVW, &r1, &thi)
gc.Regfree(&r1)
splitclean()
return
// case CASE(TFLOAT64, TUINT64):
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT8,
gc.TFLOAT32<<16 | gc.TUINT8,
gc.TFLOAT32<<16 | gc.TINT16,
gc.TFLOAT32<<16 | gc.TUINT16,
gc.TFLOAT32<<16 | gc.TINT32,
gc.TFLOAT32<<16 | gc.TUINT32,
// case CASE(TFLOAT32, TUINT64):
gc.TFLOAT64<<16 | gc.TINT8,
gc.TFLOAT64<<16 | gc.TUINT8,
gc.TFLOAT64<<16 | gc.TINT16,
gc.TFLOAT64<<16 | gc.TUINT16,
gc.TFLOAT64<<16 | gc.TINT32,
gc.TFLOAT64<<16 | gc.TUINT32:
fa := arm.AMOVF
a := arm.AMOVFW
if ft == gc.TFLOAT64 {
fa = arm.AMOVD
a = arm.AMOVDW
}
ta := arm.AMOVW
switch tt {
case gc.TINT8:
ta = arm.AMOVBS
case gc.TUINT8:
ta = arm.AMOVBU
case gc.TINT16:
ta = arm.AMOVHS
case gc.TUINT16:
ta = arm.AMOVHU
}
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[ft], f)
var r2 gc.Node
gc.Regalloc(&r2, gc.Types[tt], t)
gins(fa, f, &r1) // load to fpu
p1 := gins(a, &r1, &r1) // convert to w
switch tt {
case gc.TUINT8,
gc.TUINT16,
gc.TUINT32:
p1.Scond |= arm.C_UBIT
}
gins(arm.AMOVW, &r1, &r2) // copy to cpu
gins(ta, &r2, t) // store
gc.Regfree(&r1)
gc.Regfree(&r2)
return
/*
* integer to float
*/
case gc.TINT8<<16 | gc.TFLOAT32,
gc.TUINT8<<16 | gc.TFLOAT32,
gc.TINT16<<16 | gc.TFLOAT32,
gc.TUINT16<<16 | gc.TFLOAT32,
gc.TINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TUINT8<<16 | gc.TFLOAT64,
gc.TINT16<<16 | gc.TFLOAT64,
gc.TUINT16<<16 | gc.TFLOAT64,
gc.TINT32<<16 | gc.TFLOAT64,
gc.TUINT32<<16 | gc.TFLOAT64:
fa := arm.AMOVW
switch ft {
case gc.TINT8:
fa = arm.AMOVBS
case gc.TUINT8:
fa = arm.AMOVBU
case gc.TINT16:
fa = arm.AMOVHS
case gc.TUINT16:
fa = arm.AMOVHU
}
a := arm.AMOVWF
ta := arm.AMOVF
if tt == gc.TFLOAT64 {
a = arm.AMOVWD
ta = arm.AMOVD
}
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[ft], f)
var r2 gc.Node
gc.Regalloc(&r2, gc.Types[tt], t)
gins(fa, f, &r1) // load to cpu
gins(arm.AMOVW, &r1, &r2) // copy to fpu
p1 := gins(a, &r2, &r2) // convert
switch ft {
case gc.TUINT8,
gc.TUINT16,
gc.TUINT32:
p1.Scond |= arm.C_UBIT
}
gins(ta, &r2, t) // store
gc.Regfree(&r1)
gc.Regfree(&r2)
return
case gc.TUINT64<<16 | gc.TFLOAT32,
gc.TUINT64<<16 | gc.TFLOAT64:
gc.Fatal("gmove UINT64, TFLOAT not implemented")
return
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32:
a = arm.AMOVF
case gc.TFLOAT64<<16 | gc.TFLOAT64:
a = arm.AMOVD
case gc.TFLOAT32<<16 | gc.TFLOAT64:
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[gc.TFLOAT64], t)
gins(arm.AMOVF, f, &r1)
gins(arm.AMOVFD, &r1, &r1)
gins(arm.AMOVD, &r1, t)
gc.Regfree(&r1)
return
case gc.TFLOAT64<<16 | gc.TFLOAT32:
var r1 gc.Node
gc.Regalloc(&r1, gc.Types[gc.TFLOAT64], t)
gins(arm.AMOVD, f, &r1)
gins(arm.AMOVDF, &r1, &r1)
gins(arm.AMOVF, &r1, t)
gc.Regfree(&r1)
return
}
gins(a, f, t)
return
// TODO(kaib): we almost always require a register dest anyway, this can probably be
// removed.
// 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
// truncate 64 bit integer
trunc64:
var fhi gc.Node
var flo gc.Node
split64(f, &flo, &fhi)
gc.Regalloc(&r1, t.Type, nil)
gins(a, &flo, &r1)
gins(a, &r1, t)
gc.Regfree(&r1)
splitclean()
return
}
