/*
* 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.Fatalf("split64 %v", n.Type)
}
if nsclean >= len(sclean) {
gc.Fatalf("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.Fatalf("needzero class %d", n.Class)
}
if n.Type.Width%int64(gc.Widthptr) != 0 || n.Xoffset%int64(gc.Widthptr) != 0 || n.Type.Width == 0 {
gc.Fatalf("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.Fatalf("bad in raddr: %v", gc.Oconv(int(n.Op), 0))
} else {
gc.Fatalf("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 contain ambiguously live variables
// so that garbage collector only sees initialized values
// when it looks for pointers.
p := ptxt
hi := int64(0)
lo := hi
r0 := 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.Fatalf("needzero class %d", n.Class)
}
if n.Type.Width%int64(gc.Widthptr) != 0 || n.Xoffset%int64(gc.Widthptr) != 0 || n.Type.Width == 0 {
gc.Fatalf("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 = gc.Rnd(n.Xoffset, int64(gc.Widthptr))
continue
}
// zero old range
p = zerorange(p, int64(frame), lo, hi, &r0)
// set new range
hi = n.Xoffset + n.Type.Width
lo = n.Xoffset
}
// zero final range
zerorange(p, int64(frame), lo, hi, &r0)
}
/*
* 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.Fatalf("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.Fatalf("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 proginfo(p *obj.Prog) {
info := &p.Info
*info = progtable[p.As]
if info.Flags == 0 {
gc.Fatalf("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 splitclean() {
if nsclean <= 0 {
gc.Fatalf("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.Fatalf("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.Fatalf("as2variant: instruction %v is not a variant of itself", obj.Aconv(as))
return 0
}
/*
* 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 {
nl, nr = nr, nl
}
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:
gins3(optoas(gc.OMUL, t), &n2, &n1, nil)
var lo gc.Node
gc.Nodreg(&lo, gc.Types[gc.TUINT64], mips.REG_LO)
gins(mips.AMOVV, &lo, &n1)
p := (*obj.Prog)(gins(mips.ASRAV, nil, &n1))
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(w)
case gc.TUINT8,
gc.TUINT16,
gc.TUINT32:
gins3(optoas(gc.OMUL, t), &n2, &n1, nil)
var lo gc.Node
gc.Nodreg(&lo, gc.Types[gc.TUINT64], mips.REG_LO)
gins(mips.AMOVV, &lo, &n1)
p := (*obj.Prog)(gins(mips.ASRLV, nil, &n1))
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(w)
case gc.TINT64,
gc.TUINT64:
if gc.Issigned[t.Etype] {
gins3(mips.AMULV, &n2, &n1, nil)
} else {
gins3(mips.AMULVU, &n2, &n1, nil)
}
var hi gc.Node
gc.Nodreg(&hi, gc.Types[gc.TUINT64], mips.REG_HI)
gins(mips.AMOVV, &hi, &n1)
default:
gc.Fatalf("cgen_hmul %v", t)
}
gc.Cgen(&n1, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
/*
* 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.Fatalf("use of byte register")
}
if x86.REG_AL <= v.Reg && v.Reg <= x86.REG_R15B {
gc.Fatalf("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 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.Fatalf("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
}
/*
* generate high multiply
* res = (nl * nr) >> wordsize
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Ullman < nr.Ullman {
nl, nr = nr, nl
}
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.Fatalf("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.Fatalf("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.Fatalf("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 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 {
nl, nr = nr, nl
}
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(arm64.AASR, 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(arm64.ALSR, nil, &n1))
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(w)
case gc.TINT64,
gc.TUINT64:
if gc.Issigned[t.Etype] {
gins(arm64.ASMULH, &n2, &n1)
} else {
gins(arm64.AUMULH, &n2, &n1)
}
default:
gc.Fatalf("cgen_hmul %v", t)
}
gc.Cgen(&n1, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
/*
* generate division according to op, one of:
* res = nl / nr
* res = nl % nr
*/
func cgen_div(op gc.Op, nl *gc.Node, nr *gc.Node, res *gc.Node) {
if gc.Is64(nl.Type) {
gc.Fatalf("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.Fatalf("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.Fatalf("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.Fatalf("jmptoset: no entry for %v", gc.Oconv(jmp, 0))
panic("unreachable")
}
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64, ax *uint32, x0 *uint32) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
if cnt%int64(gc.Widthreg) != 0 {
// should only happen with nacl
if cnt%int64(gc.Widthptr) != 0 {
gc.Fatalf("zerorange count not a multiple of widthptr %d", cnt)
}
if *ax == 0 {
p = appendpp(p, x86.AMOVQ, obj.TYPE_CONST, 0, 0, obj.TYPE_REG, x86.REG_AX, 0)
*ax = 1
}
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 == 8 {
if *ax == 0 {
p = appendpp(p, x86.AMOVQ, obj.TYPE_CONST, 0, 0, obj.TYPE_REG, x86.REG_AX, 0)
*ax = 1
}
p = appendpp(p, x86.AMOVQ, obj.TYPE_REG, x86.REG_AX, 0, obj.TYPE_MEM, x86.REG_SP, frame+lo)
} else if cnt <= int64(8*gc.Widthreg) {
if *x0 == 0 {
p = appendpp(p, x86.AXORPS, obj.TYPE_REG, x86.REG_X0, 0, obj.TYPE_REG, x86.REG_X0, 0)
*x0 = 1
}
for i := int64(0); i < cnt/16; i++ {
p = appendpp(p, x86.AMOVUPS, obj.TYPE_REG, x86.REG_X0, 0, obj.TYPE_MEM, x86.REG_SP, frame+lo+i*16)
}
if cnt%16 != 0 {
p = appendpp(p, x86.AMOVUPS, obj.TYPE_REG, x86.REG_X0, 0, obj.TYPE_MEM, x86.REG_SP, frame+lo+cnt-int64(16))
}
} else if !gc.Nacl && (cnt <= int64(128*gc.Widthreg)) {
if *x0 == 0 {
p = appendpp(p, x86.AXORPS, obj.TYPE_REG, x86.REG_X0, 0, obj.TYPE_REG, x86.REG_X0, 0)
*x0 = 1
}
p = appendpp(p, leaptr, obj.TYPE_MEM, x86.REG_SP, frame+lo+dzDI(cnt), obj.TYPE_REG, x86.REG_DI, 0)
p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_ADDR, 0, dzOff(cnt))
p.To.Sym = gc.Linksym(gc.Pkglookup("duffzero", gc.Runtimepkg))
if cnt%16 != 0 {
p = appendpp(p, x86.AMOVUPS, obj.TYPE_REG, x86.REG_X0, 0, obj.TYPE_MEM, x86.REG_DI, -int64(8))
}
} else {
if *ax == 0 {
p = appendpp(p, x86.AMOVQ, obj.TYPE_CONST, 0, 0, obj.TYPE_REG, x86.REG_AX, 0)
*ax = 1
}
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
}
/*
* return Axxx for Oxxx on type t.
*/
func optoas(op gc.Op, t *gc.Type) int {
if t == nil {
gc.Fatalf("optoas: t is nil")
}
// avoid constant conversions in switches below
const (
OMINUS_ = uint32(gc.OMINUS) << 16
OLSH_ = uint32(gc.OLSH) << 16
ORSH_ = uint32(gc.ORSH) << 16
OADD_ = uint32(gc.OADD) << 16
OSUB_ = uint32(gc.OSUB) << 16
OMUL_ = uint32(gc.OMUL) << 16
ODIV_ = uint32(gc.ODIV) << 16
OOR_ = uint32(gc.OOR) << 16
OAND_ = uint32(gc.OAND) << 16
OXOR_ = uint32(gc.OXOR) << 16
OEQ_ = uint32(gc.OEQ) << 16
ONE_ = uint32(gc.ONE) << 16
OLT_ = uint32(gc.OLT) << 16
OLE_ = uint32(gc.OLE) << 16
OGE_ = uint32(gc.OGE) << 16
OGT_ = uint32(gc.OGT) << 16
OCMP_ = uint32(gc.OCMP) << 16
OAS_ = uint32(gc.OAS) << 16
OHMUL_ = uint32(gc.OHMUL) << 16
)
a := int(obj.AXXX)
switch uint32(op)<<16 | uint32(gc.Simtype[t.Etype]) {
default:
gc.Fatalf("optoas: no entry for op=%v type=%v", gc.Oconv(int(op), 0), t)
case OEQ_ | gc.TBOOL,
OEQ_ | gc.TINT8,
OEQ_ | gc.TUINT8,
OEQ_ | gc.TINT16,
OEQ_ | gc.TUINT16,
OEQ_ | gc.TINT32,
OEQ_ | gc.TUINT32,
OEQ_ | gc.TINT64,
OEQ_ | gc.TUINT64,
OEQ_ | gc.TPTR32,
OEQ_ | gc.TPTR64,
OEQ_ | gc.TFLOAT32,
OEQ_ | gc.TFLOAT64:
a = ppc64.ABEQ
case ONE_ | gc.TBOOL,
ONE_ | gc.TINT8,
ONE_ | gc.TUINT8,
ONE_ | gc.TINT16,
ONE_ | gc.TUINT16,
ONE_ | gc.TINT32,
ONE_ | gc.TUINT32,
ONE_ | gc.TINT64,
ONE_ | gc.TUINT64,
ONE_ | gc.TPTR32,
ONE_ | gc.TPTR64,
ONE_ | gc.TFLOAT32,
ONE_ | gc.TFLOAT64:
a = ppc64.ABNE
case OLT_ | gc.TINT8, // ACMP
OLT_ | gc.TINT16,
OLT_ | gc.TINT32,
OLT_ | gc.TINT64,
OLT_ | gc.TUINT8,
// ACMPU
OLT_ | gc.TUINT16,
OLT_ | gc.TUINT32,
OLT_ | gc.TUINT64,
OLT_ | gc.TFLOAT32,
// AFCMPU
OLT_ | gc.TFLOAT64:
a = ppc64.ABLT
case OLE_ | gc.TINT8, // ACMP
OLE_ | gc.TINT16,
OLE_ | gc.TINT32,
OLE_ | gc.TINT64,
OLE_ | gc.TUINT8,
// ACMPU
OLE_ | gc.TUINT16,
OLE_ | gc.TUINT32,
OLE_ | gc.TUINT64:
// No OLE for floats, because it mishandles NaN.
// Front end must reverse comparison or use OLT and OEQ together.
a = ppc64.ABLE
case OGT_ | gc.TINT8,
OGT_ | gc.TINT16,
OGT_ | gc.TINT32,
OGT_ | gc.TINT64,
OGT_ | gc.TUINT8,
OGT_ | gc.TUINT16,
OGT_ | gc.TUINT32,
OGT_ | gc.TUINT64,
OGT_ | gc.TFLOAT32,
OGT_ | gc.TFLOAT64:
a = ppc64.ABGT
case OGE_ | gc.TINT8,
OGE_ | gc.TINT16,
OGE_ | gc.TINT32,
OGE_ | gc.TINT64,
OGE_ | gc.TUINT8,
OGE_ | gc.TUINT16,
OGE_ | gc.TUINT32,
OGE_ | gc.TUINT64:
// No OGE for floats, because it mishandles NaN.
// Front end must reverse comparison or use OLT and OEQ together.
a = ppc64.ABGE
case OCMP_ | gc.TBOOL,
OCMP_ | gc.TINT8,
OCMP_ | gc.TINT16,
OCMP_ | gc.TINT32,
OCMP_ | gc.TPTR32,
OCMP_ | gc.TINT64:
a = ppc64.ACMP
case OCMP_ | gc.TUINT8,
OCMP_ | gc.TUINT16,
OCMP_ | gc.TUINT32,
OCMP_ | gc.TUINT64,
OCMP_ | gc.TPTR64:
a = ppc64.ACMPU
case OCMP_ | gc.TFLOAT32,
OCMP_ | gc.TFLOAT64:
a = ppc64.AFCMPU
case OAS_ | gc.TBOOL,
OAS_ | gc.TINT8:
a = ppc64.AMOVB
case OAS_ | gc.TUINT8:
a = ppc64.AMOVBZ
case OAS_ | gc.TINT16:
a = ppc64.AMOVH
case OAS_ | gc.TUINT16:
a = ppc64.AMOVHZ
case OAS_ | gc.TINT32:
a = ppc64.AMOVW
case OAS_ | gc.TUINT32,
OAS_ | gc.TPTR32:
a = ppc64.AMOVWZ
case OAS_ | gc.TINT64,
OAS_ | gc.TUINT64,
OAS_ | gc.TPTR64:
a = ppc64.AMOVD
case OAS_ | gc.TFLOAT32:
a = ppc64.AFMOVS
case OAS_ | gc.TFLOAT64:
a = ppc64.AFMOVD
case OADD_ | gc.TINT8,
OADD_ | gc.TUINT8,
OADD_ | gc.TINT16,
OADD_ | gc.TUINT16,
OADD_ | gc.TINT32,
OADD_ | gc.TUINT32,
OADD_ | gc.TPTR32,
OADD_ | gc.TINT64,
OADD_ | gc.TUINT64,
OADD_ | gc.TPTR64:
a = ppc64.AADD
case OADD_ | gc.TFLOAT32:
a = ppc64.AFADDS
case OADD_ | gc.TFLOAT64:
a = ppc64.AFADD
case OSUB_ | gc.TINT8,
OSUB_ | gc.TUINT8,
OSUB_ | gc.TINT16,
OSUB_ | gc.TUINT16,
OSUB_ | gc.TINT32,
OSUB_ | gc.TUINT32,
OSUB_ | gc.TPTR32,
OSUB_ | gc.TINT64,
OSUB_ | gc.TUINT64,
OSUB_ | gc.TPTR64:
a = ppc64.ASUB
case OSUB_ | gc.TFLOAT32:
a = ppc64.AFSUBS
case OSUB_ | gc.TFLOAT64:
a = ppc64.AFSUB
case OMINUS_ | gc.TINT8,
OMINUS_ | gc.TUINT8,
OMINUS_ | gc.TINT16,
OMINUS_ | gc.TUINT16,
OMINUS_ | gc.TINT32,
OMINUS_ | gc.TUINT32,
OMINUS_ | gc.TPTR32,
OMINUS_ | gc.TINT64,
OMINUS_ | gc.TUINT64,
OMINUS_ | gc.TPTR64:
a = ppc64.ANEG
case OAND_ | gc.TINT8,
OAND_ | gc.TUINT8,
OAND_ | gc.TINT16,
OAND_ | gc.TUINT16,
OAND_ | gc.TINT32,
OAND_ | gc.TUINT32,
OAND_ | gc.TPTR32,
OAND_ | gc.TINT64,
OAND_ | gc.TUINT64,
OAND_ | gc.TPTR64:
a = ppc64.AAND
case OOR_ | gc.TINT8,
OOR_ | gc.TUINT8,
OOR_ | gc.TINT16,
OOR_ | gc.TUINT16,
OOR_ | gc.TINT32,
OOR_ | gc.TUINT32,
OOR_ | gc.TPTR32,
OOR_ | gc.TINT64,
OOR_ | gc.TUINT64,
OOR_ | gc.TPTR64:
a = ppc64.AOR
case OXOR_ | gc.TINT8,
OXOR_ | gc.TUINT8,
OXOR_ | gc.TINT16,
OXOR_ | gc.TUINT16,
OXOR_ | gc.TINT32,
OXOR_ | gc.TUINT32,
OXOR_ | gc.TPTR32,
OXOR_ | gc.TINT64,
OXOR_ | gc.TUINT64,
OXOR_ | gc.TPTR64:
a = ppc64.AXOR
// TODO(minux): handle rotates
//case CASE(OLROT, TINT8):
//case CASE(OLROT, TUINT8):
//case CASE(OLROT, TINT16):
//case CASE(OLROT, TUINT16):
//case CASE(OLROT, TINT32):
//case CASE(OLROT, TUINT32):
//case CASE(OLROT, TPTR32):
//case CASE(OLROT, TINT64):
//case CASE(OLROT, TUINT64):
//case CASE(OLROT, TPTR64):
// a = 0//???; RLDC?
// break;
case OLSH_ | gc.TINT8,
OLSH_ | gc.TUINT8,
OLSH_ | gc.TINT16,
OLSH_ | gc.TUINT16,
OLSH_ | gc.TINT32,
OLSH_ | gc.TUINT32,
OLSH_ | gc.TPTR32,
OLSH_ | gc.TINT64,
OLSH_ | gc.TUINT64,
OLSH_ | gc.TPTR64:
a = ppc64.ASLD
case ORSH_ | gc.TUINT8,
ORSH_ | gc.TUINT16,
ORSH_ | gc.TUINT32,
ORSH_ | gc.TPTR32,
ORSH_ | gc.TUINT64,
ORSH_ | gc.TPTR64:
a = ppc64.ASRD
case ORSH_ | gc.TINT8,
ORSH_ | gc.TINT16,
ORSH_ | gc.TINT32,
ORSH_ | gc.TINT64:
a = ppc64.ASRAD
// TODO(minux): handle rotates
//case CASE(ORROTC, TINT8):
//case CASE(ORROTC, TUINT8):
//case CASE(ORROTC, TINT16):
//case CASE(ORROTC, TUINT16):
//case CASE(ORROTC, TINT32):
//case CASE(ORROTC, TUINT32):
//case CASE(ORROTC, TINT64):
//case CASE(ORROTC, TUINT64):
// a = 0//??? RLDC??
// break;
case OHMUL_ | gc.TINT64:
a = ppc64.AMULHD
case OHMUL_ | gc.TUINT64,
OHMUL_ | gc.TPTR64:
a = ppc64.AMULHDU
case OMUL_ | gc.TINT8,
OMUL_ | gc.TINT16,
OMUL_ | gc.TINT32,
OMUL_ | gc.TINT64:
a = ppc64.AMULLD
case OMUL_ | gc.TUINT8,
OMUL_ | gc.TUINT16,
OMUL_ | gc.TUINT32,
OMUL_ | gc.TPTR32,
// don't use word multiply, the high 32-bit are undefined.
OMUL_ | gc.TUINT64,
OMUL_ | gc.TPTR64:
// for 64-bit multiplies, signedness doesn't matter.
a = ppc64.AMULLD
case OMUL_ | gc.TFLOAT32:
a = ppc64.AFMULS
case OMUL_ | gc.TFLOAT64:
a = ppc64.AFMUL
case ODIV_ | gc.TINT8,
ODIV_ | gc.TINT16,
ODIV_ | gc.TINT32,
ODIV_ | gc.TINT64:
a = ppc64.ADIVD
case ODIV_ | gc.TUINT8,
ODIV_ | gc.TUINT16,
ODIV_ | gc.TUINT32,
ODIV_ | gc.TPTR32,
ODIV_ | gc.TUINT64,
ODIV_ | gc.TPTR64:
a = ppc64.ADIVDU
case ODIV_ | gc.TFLOAT32:
a = ppc64.AFDIVS
case ODIV_ | gc.TFLOAT64:
a = ppc64.AFDIV
}
return a
}
/*
* generate one instruction:
* as f, t
*/
func rawgins(as int, f *gc.Node, t *gc.Node) *obj.Prog {
// TODO(austin): Add self-move test like in 6g (but be careful
// of truncation moves)
p := gc.Prog(as)
gc.Naddr(&p.From, f)
gc.Naddr(&p.To, t)
switch as {
case obj.ACALL:
if p.To.Type == obj.TYPE_REG && p.To.Reg != ppc64.REG_CTR {
// Allow front end to emit CALL REG, and rewrite into MOV REG, CTR; CALL CTR.
if gc.Ctxt.Flag_shared != 0 {
// Make sure function pointer is in R12 as well when
// compiling Go into PIC.
// TODO(mwhudson): it would obviously be better to
// change the register allocation to put the value in
// R12 already, but I don't know how to do that.
q := gc.Prog(as)
q.As = ppc64.AMOVD
q.From = p.To
q.To.Type = obj.TYPE_REG
q.To.Reg = ppc64.REG_R12
}
pp := gc.Prog(as)
pp.From = p.From
pp.To.Type = obj.TYPE_REG
pp.To.Reg = ppc64.REG_CTR
p.As = ppc64.AMOVD
p.From = p.To
p.To.Type = obj.TYPE_REG
p.To.Reg = ppc64.REG_CTR
if gc.Ctxt.Flag_shared != 0 {
// When compiling Go into PIC, the function we just
// called via pointer might have been implemented in
// a separate module and so overwritten the TOC
// pointer in R2; reload it.
q := gc.Prog(ppc64.AMOVD)
q.From.Type = obj.TYPE_MEM
q.From.Offset = 24
q.From.Reg = ppc64.REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = ppc64.REG_R2
}
if gc.Debug['g'] != 0 {
fmt.Printf("%v\n", p)
fmt.Printf("%v\n", pp)
}
return pp
}
// Bad things the front end has done to us. Crash to find call stack.
case ppc64.AAND, ppc64.AMULLD:
if p.From.Type == obj.TYPE_CONST {
gc.Debug['h'] = 1
gc.Fatalf("bad inst: %v", p)
}
case ppc64.ACMP, ppc64.ACMPU:
if p.From.Type == obj.TYPE_MEM || p.To.Type == obj.TYPE_MEM {
gc.Debug['h'] = 1
gc.Fatalf("bad inst: %v", p)
}
}
if gc.Debug['g'] != 0 {
fmt.Printf("%v\n", p)
}
w := int32(0)
switch as {
case ppc64.AMOVB,
ppc64.AMOVBU,
ppc64.AMOVBZ,
ppc64.AMOVBZU:
w = 1
case ppc64.AMOVH,
ppc64.AMOVHU,
ppc64.AMOVHZ,
ppc64.AMOVHZU:
w = 2
case ppc64.AMOVW,
ppc64.AMOVWU,
ppc64.AMOVWZ,
ppc64.AMOVWZU:
w = 4
case ppc64.AMOVD,
ppc64.AMOVDU:
if p.From.Type == obj.TYPE_CONST || p.From.Type == obj.TYPE_ADDR {
break
}
w = 8
}
if w != 0 && ((f != nil && p.From.Width < int64(w)) || (t != nil && p.To.Type != obj.TYPE_REG && p.To.Width > int64(w))) {
gc.Dump("f", f)
gc.Dump("t", t)
gc.Fatalf("bad width: %v (%d, %d)\n", p, p.From.Width, p.To.Width)
}
return p
}
/*
* generate move:
* t = f
* hard part is conversions.
*/
func gmove(f *gc.Node, t *gc.Node) {
if gc.Debug['M'] != 0 {
fmt.Printf("gmove %v -> %v\n", gc.Nconv(f, 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:
f.Convconst(&con, t.Type)
case gc.TINT32,
gc.TINT16,
gc.TINT8:
var con gc.Node
f.Convconst(&con, gc.Types[gc.TINT64])
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(ppc64.AMOVD, &con, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
case gc.TUINT32,
gc.TUINT16,
gc.TUINT8:
var con gc.Node
f.Convconst(&con, gc.Types[gc.TUINT64])
var r1 gc.Node
gc.Regalloc(&r1, con.Type, t)
gins(ppc64.AMOVD, &con, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
f = &con
ft = tt // so big switch will choose a simple mov
// constants can't move directly to memory.
if gc.Ismem(t) {
goto hard
}
}
// float constants come from memory.
//if(isfloat[tt])
// goto hard;
// 64-bit immediates are also from memory.
//if(isint[tt])
// goto hard;
//// 64-bit immediates are really 32-bit sign-extended
//// unless moving into a register.
//if(isint[tt]) {
// if(mpcmpfixfix(con.val.u.xval, minintval[TINT32]) < 0)
// goto hard;
// if(mpcmpfixfix(con.val.u.xval, maxintval[TINT32]) > 0)
// goto hard;
//}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
gc.Fatalf("gmove %v -> %v", gc.Tconv(f.Type, 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 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.Fatalf("%v in use during clearfat", obj.Rconv(ppc64.REGRT1))
}
var r0 gc.Node
gc.Nodreg(&r0, gc.Types[gc.TUINT64], ppc64.REGZERO)
var dst gc.Node
gc.Nodreg(&dst, gc.Types[gc.Tptr], ppc64.REGRT1)
gc.Regrealloc(&dst)
gc.Agen(nl, &dst)
var boff uint64
if q > 128 {
p := gins(ppc64.ASUB, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p = gins(ppc64.AMOVD, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(q * 8)
p = gins(ppc64.AMOVDU, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 8
pl := (*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.Fatalf("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
}
}
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(mips.REGRT1) {
gc.Fatalf("%v in use during clearfat", obj.Rconv(mips.REGRT1))
}
var r0 gc.Node
gc.Nodreg(&r0, gc.Types[gc.TUINT64], mips.REGZERO)
var dst gc.Node
gc.Nodreg(&dst, gc.Types[gc.Tptr], mips.REGRT1)
gc.Regrealloc(&dst)
gc.Agen(nl, &dst)
var boff uint64
if q > 128 {
p := gins(mips.ASUBV, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
var end gc.Node
gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
p = gins(mips.AMOVV, &dst, &end)
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(q * 8)
p = gins(mips.AMOVV, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = 8
pl := (*obj.Prog)(p)
p = gins(mips.AADDV, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
gc.Patch(ginsbranch(mips.ABNE, nil, &dst, &end, 0), pl)
gc.Regfree(&end)
// The loop leaves R1 on the last zeroed dword
boff = 8
// TODO(dfc): https://golang.org/issue/12108
// If DUFFZERO is used inside a tail call (see genwrapper) it will
// overwrite the link register.
} else if false && q >= 4 {
p := gins(mips.ASUBV, nil, &dst)
p.From.Type = obj.TYPE_CONST
p.From.Offset = 8
f := (*gc.Node)(gc.Sysfunc("duffzero"))
p = gins(obj.ADUFFZERO, nil, f)
gc.Afunclit(&p.To, f)
// 8 and 128 = magic constants: see ../../runtime/asm_mips64x.s
p.To.Offset = int64(8 * (128 - q))
// duffzero leaves R1 on the last zeroed dword
boff = 8
} else {
var p *obj.Prog
for t := uint64(0); t < q; t++ {
p = gins(mips.AMOVV, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(8 * t)
}
boff = 8 * q
}
var p *obj.Prog
for t := uint64(0); t < c; t++ {
p = gins(mips.AMOVB, &r0, &dst)
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(t + boff)
}
gc.Regfree(&dst)
}
/*
* attempt to generate 64-bit
* res = n
* return 1 on success, 0 if op not handled.
*/
func cgen64(n *gc.Node, res *gc.Node) {
if res.Op != gc.OINDREG && res.Op != gc.ONAME {
gc.Dump("n", n)
gc.Dump("res", res)
gc.Fatalf("cgen64 %v of %v", gc.Oconv(int(n.Op), 0), gc.Oconv(int(res.Op), 0))
}
l := n.Left
var t1 gc.Node
if !l.Addable {
gc.Tempname(&t1, l.Type)
gc.Cgen(l, &t1)
l = &t1
}
var hi1 gc.Node
var lo1 gc.Node
split64(l, &lo1, &hi1)
switch n.Op {
default:
gc.Fatalf("cgen64 %v", gc.Oconv(int(n.Op), 0))
case gc.OMINUS:
var lo2 gc.Node
var hi2 gc.Node
split64(res, &lo2, &hi2)
gc.Regalloc(&t1, lo1.Type, nil)
var al gc.Node
gc.Regalloc(&al, lo1.Type, nil)
var ah gc.Node
gc.Regalloc(&ah, hi1.Type, nil)
gins(arm.AMOVW, &lo1, &al)
gins(arm.AMOVW, &hi1, &ah)
gmove(ncon(0), &t1)
p1 := gins(arm.ASUB, &al, &t1)
p1.Scond |= arm.C_SBIT
gins(arm.AMOVW, &t1, &lo2)
gmove(ncon(0), &t1)
gins(arm.ASBC, &ah, &t1)
gins(arm.AMOVW, &t1, &hi2)
gc.Regfree(&t1)
gc.Regfree(&al)
gc.Regfree(&ah)
splitclean()
splitclean()
return
case gc.OCOM:
gc.Regalloc(&t1, lo1.Type, nil)
gmove(ncon(^uint32(0)), &t1)
var lo2 gc.Node
var hi2 gc.Node
split64(res, &lo2, &hi2)
var n1 gc.Node
gc.Regalloc(&n1, lo1.Type, nil)
gins(arm.AMOVW, &lo1, &n1)
gins(arm.AEOR, &t1, &n1)
gins(arm.AMOVW, &n1, &lo2)
gins(arm.AMOVW, &hi1, &n1)
gins(arm.AEOR, &t1, &n1)
gins(arm.AMOVW, &n1, &hi2)
gc.Regfree(&t1)
gc.Regfree(&n1)
splitclean()
splitclean()
return
// binary operators.
// common setup below.
case gc.OADD,
gc.OSUB,
gc.OMUL,
gc.OLSH,
gc.ORSH,
gc.OAND,
gc.OOR,
gc.OXOR,
gc.OLROT:
break
}
// setup for binary operators
r := n.Right
if r != nil && !r.Addable {
var t2 gc.Node
gc.Tempname(&t2, r.Type)
gc.Cgen(r, &t2)
r = &t2
}
var hi2 gc.Node
var lo2 gc.Node
if gc.Is64(r.Type) {
split64(r, &lo2, &hi2)
}
var al gc.Node
gc.Regalloc(&al, lo1.Type, nil)
var ah gc.Node
gc.Regalloc(&ah, hi1.Type, nil)
// Do op. Leave result in ah:al.
switch n.Op {
default:
gc.Fatalf("cgen64: not implemented: %v\n", n)
// TODO: Constants
case gc.OADD:
var bl gc.Node
gc.Regalloc(&bl, gc.Types[gc.TPTR32], nil)
var bh gc.Node
gc.Regalloc(&bh, gc.Types[gc.TPTR32], nil)
gins(arm.AMOVW, &hi1, &ah)
gins(arm.AMOVW, &lo1, &al)
gins(arm.AMOVW, &hi2, &bh)
gins(arm.AMOVW, &lo2, &bl)
p1 := gins(arm.AADD, &bl, &al)
p1.Scond |= arm.C_SBIT
gins(arm.AADC, &bh, &ah)
gc.Regfree(&bl)
gc.Regfree(&bh)
// TODO: Constants.
case gc.OSUB:
var bl gc.Node
gc.Regalloc(&bl, gc.Types[gc.TPTR32], nil)
var bh gc.Node
gc.Regalloc(&bh, gc.Types[gc.TPTR32], nil)
gins(arm.AMOVW, &lo1, &al)
gins(arm.AMOVW, &hi1, &ah)
gins(arm.AMOVW, &lo2, &bl)
gins(arm.AMOVW, &hi2, &bh)
p1 := gins(arm.ASUB, &bl, &al)
p1.Scond |= arm.C_SBIT
gins(arm.ASBC, &bh, &ah)
gc.Regfree(&bl)
gc.Regfree(&bh)
// TODO(kaib): this can be done with 4 regs and does not need 6
case gc.OMUL:
var bl gc.Node
gc.Regalloc(&bl, gc.Types[gc.TPTR32], nil)
var bh gc.Node
gc.Regalloc(&bh, gc.Types[gc.TPTR32], nil)
var cl gc.Node
gc.Regalloc(&cl, gc.Types[gc.TPTR32], nil)
var ch gc.Node
gc.Regalloc(&ch, gc.Types[gc.TPTR32], nil)
// load args into bh:bl and bh:bl.
gins(arm.AMOVW, &hi1, &bh)
gins(arm.AMOVW, &lo1, &bl)
gins(arm.AMOVW, &hi2, &ch)
gins(arm.AMOVW, &lo2, &cl)
// bl * cl -> ah al
p1 := gins(arm.AMULLU, nil, nil)
p1.From.Type = obj.TYPE_REG
p1.From.Reg = bl.Reg
p1.Reg = cl.Reg
p1.To.Type = obj.TYPE_REGREG
p1.To.Reg = ah.Reg
p1.To.Offset = int64(al.Reg)
//print("%v\n", p1);
// bl * ch + ah -> ah
p1 = gins(arm.AMULA, nil, nil)
p1.From.Type = obj.TYPE_REG
p1.From.Reg = bl.Reg
p1.Reg = ch.Reg
p1.To.Type = obj.TYPE_REGREG2
p1.To.Reg = ah.Reg
p1.To.Offset = int64(ah.Reg)
//print("%v\n", p1);
// bh * cl + ah -> ah
p1 = gins(arm.AMULA, nil, nil)
p1.From.Type = obj.TYPE_REG
p1.From.Reg = bh.Reg
p1.Reg = cl.Reg
p1.To.Type = obj.TYPE_REGREG2
p1.To.Reg = ah.Reg
p1.To.Offset = int64(ah.Reg)
//print("%v\n", p1);
gc.Regfree(&bh)
gc.Regfree(&bl)
gc.Regfree(&ch)
gc.Regfree(&cl)
// We only rotate by a constant c in [0,64).
// if c >= 32:
// lo, hi = hi, lo
// c -= 32
// if c == 0:
// no-op
// else:
// t = hi
// shld hi:lo, c
// shld lo:t, c
case gc.OLROT:
v := uint64(r.Int())
var bl gc.Node
gc.Regalloc(&bl, lo1.Type, nil)
var bh gc.Node
gc.Regalloc(&bh, hi1.Type, nil)
if v >= 32 {
// reverse during load to do the first 32 bits of rotate
v -= 32
gins(arm.AMOVW, &hi1, &bl)
gins(arm.AMOVW, &lo1, &bh)
} else {
gins(arm.AMOVW, &hi1, &bh)
gins(arm.AMOVW, &lo1, &bl)
}
if v == 0 {
gins(arm.AMOVW, &bh, &ah)
gins(arm.AMOVW, &bl, &al)
} else {
// rotate by 1 <= v <= 31
// MOVW bl<<v, al
// MOVW bh<<v, ah
// OR bl>>(32-v), ah
// OR bh>>(32-v), al
gshift(arm.AMOVW, &bl, arm.SHIFT_LL, int32(v), &al)
gshift(arm.AMOVW, &bh, arm.SHIFT_LL, int32(v), &ah)
gshift(arm.AORR, &bl, arm.SHIFT_LR, int32(32-v), &ah)
gshift(arm.AORR, &bh, arm.SHIFT_LR, int32(32-v), &al)
}
gc.Regfree(&bl)
gc.Regfree(&bh)
case gc.OLSH:
var bl gc.Node
gc.Regalloc(&bl, lo1.Type, nil)
var bh gc.Node
gc.Regalloc(&bh, hi1.Type, nil)
gins(arm.AMOVW, &hi1, &bh)
gins(arm.AMOVW, &lo1, &bl)
var p6 *obj.Prog
var s gc.Node
var n1 gc.Node
var creg gc.Node
var p1 *obj.Prog
var p2 *obj.Prog
var p3 *obj.Prog
var p4 *obj.Prog
var p5 *obj.Prog
if r.Op == gc.OLITERAL {
v := uint64(r.Int())
if v >= 64 {
// TODO(kaib): replace with gins(AMOVW, nodintconst(0), &al)
// here and below (verify it optimizes to EOR)
gins(arm.AEOR, &al, &al)
gins(arm.AEOR, &ah, &ah)
} else if v > 32 {
gins(arm.AEOR, &al, &al)
// MOVW bl<<(v-32), ah
gshift(arm.AMOVW, &bl, arm.SHIFT_LL, int32(v-32), &ah)
} else if v == 32 {
gins(arm.AEOR, &al, &al)
gins(arm.AMOVW, &bl, &ah)
} else if v > 0 {
// MOVW bl<<v, al
gshift(arm.AMOVW, &bl, arm.SHIFT_LL, int32(v), &al)
// MOVW bh<<v, ah
gshift(arm.AMOVW, &bh, arm.SHIFT_LL, int32(v), &ah)
// OR bl>>(32-v), ah
gshift(arm.AORR, &bl, arm.SHIFT_LR, int32(32-v), &ah)
} else {
gins(arm.AMOVW, &bl, &al)
gins(arm.AMOVW, &bh, &ah)
}
goto olsh_break
}
gc.Regalloc(&s, gc.Types[gc.TUINT32], nil)
gc.Regalloc(&creg, gc.Types[gc.TUINT32], nil)
if gc.Is64(r.Type) {
// shift is >= 1<<32
var cl gc.Node
var ch gc.Node
split64(r, &cl, &ch)
gmove(&ch, &s)
gins(arm.ATST, &s, nil)
p6 = gc.Gbranch(arm.ABNE, nil, 0)
gmove(&cl, &s)
splitclean()
} else {
gmove(r, &s)
p6 = nil
}
gins(arm.ATST, &s, nil)
// shift == 0
p1 = gins(arm.AMOVW, &bl, &al)
p1.Scond = arm.C_SCOND_EQ
p1 = gins(arm.AMOVW, &bh, &ah)
p1.Scond = arm.C_SCOND_EQ
p2 = gc.Gbranch(arm.ABEQ, nil, 0)
// shift is < 32
gc.Nodconst(&n1, gc.Types[gc.TUINT32], 32)
gmove(&n1, &creg)
gins(arm.ACMP, &s, &creg)
// MOVW.LO bl<<s, al
p1 = gregshift(arm.AMOVW, &bl, arm.SHIFT_LL, &s, &al)
p1.Scond = arm.C_SCOND_LO
// MOVW.LO bh<<s, ah
p1 = gregshift(arm.AMOVW, &bh, arm.SHIFT_LL, &s, &ah)
p1.Scond = arm.C_SCOND_LO
// SUB.LO s, creg
p1 = gins(arm.ASUB, &s, &creg)
p1.Scond = arm.C_SCOND_LO
// OR.LO bl>>creg, ah
p1 = gregshift(arm.AORR, &bl, arm.SHIFT_LR, &creg, &ah)
p1.Scond = arm.C_SCOND_LO
// BLO end
p3 = gc.Gbranch(arm.ABLO, nil, 0)
// shift == 32
p1 = gins(arm.AEOR, &al, &al)
p1.Scond = arm.C_SCOND_EQ
p1 = gins(arm.AMOVW, &bl, &ah)
p1.Scond = arm.C_SCOND_EQ
p4 = gc.Gbranch(arm.ABEQ, nil, 0)
// shift is < 64
gc.Nodconst(&n1, gc.Types[gc.TUINT32], 64)
gmove(&n1, &creg)
gins(arm.ACMP, &s, &creg)
// EOR.LO al, al
p1 = gins(arm.AEOR, &al, &al)
p1.Scond = arm.C_SCOND_LO
// MOVW.LO creg>>1, creg
p1 = gshift(arm.AMOVW, &creg, arm.SHIFT_LR, 1, &creg)
p1.Scond = arm.C_SCOND_LO
// SUB.LO creg, s
p1 = gins(arm.ASUB, &creg, &s)
p1.Scond = arm.C_SCOND_LO
// MOVW bl<<s, ah
p1 = gregshift(arm.AMOVW, &bl, arm.SHIFT_LL, &s, &ah)
p1.Scond = arm.C_SCOND_LO
p5 = gc.Gbranch(arm.ABLO, nil, 0)
// shift >= 64
if p6 != nil {
gc.Patch(p6, gc.Pc)
}
gins(arm.AEOR, &al, &al)
gins(arm.AEOR, &ah, &ah)
gc.Patch(p2, gc.Pc)
gc.Patch(p3, gc.Pc)
gc.Patch(p4, gc.Pc)
gc.Patch(p5, gc.Pc)
gc.Regfree(&s)
gc.Regfree(&creg)
olsh_break:
gc.Regfree(&bl)
gc.Regfree(&bh)
case gc.ORSH:
var bl gc.Node
gc.Regalloc(&bl, lo1.Type, nil)
var bh gc.Node
gc.Regalloc(&bh, hi1.Type, nil)
gins(arm.AMOVW, &hi1, &bh)
gins(arm.AMOVW, &lo1, &bl)
var p4 *obj.Prog
var p5 *obj.Prog
var n1 gc.Node
var p6 *obj.Prog
var s gc.Node
var p1 *obj.Prog
var p2 *obj.Prog
var creg gc.Node
var p3 *obj.Prog
if r.Op == gc.OLITERAL {
v := uint64(r.Int())
if v >= 64 {
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->31, al
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &al)
// MOVW bh->31, ah
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
} else {
gins(arm.AEOR, &al, &al)
gins(arm.AEOR, &ah, &ah)
}
} else if v > 32 {
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->(v-32), al
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, int32(v-32), &al)
// MOVW bh->31, ah
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
} else {
// MOVW bh>>(v-32), al
gshift(arm.AMOVW, &bh, arm.SHIFT_LR, int32(v-32), &al)
gins(arm.AEOR, &ah, &ah)
}
} else if v == 32 {
gins(arm.AMOVW, &bh, &al)
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->31, ah
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
} else {
gins(arm.AEOR, &ah, &ah)
}
} else if v > 0 {
// MOVW bl>>v, al
gshift(arm.AMOVW, &bl, arm.SHIFT_LR, int32(v), &al)
// OR bh<<(32-v), al
gshift(arm.AORR, &bh, arm.SHIFT_LL, int32(32-v), &al)
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->v, ah
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, int32(v), &ah)
} else {
// MOVW bh>>v, ah
gshift(arm.AMOVW, &bh, arm.SHIFT_LR, int32(v), &ah)
}
} else {
gins(arm.AMOVW, &bl, &al)
gins(arm.AMOVW, &bh, &ah)
}
goto orsh_break
}
gc.Regalloc(&s, gc.Types[gc.TUINT32], nil)
gc.Regalloc(&creg, gc.Types[gc.TUINT32], nil)
if gc.Is64(r.Type) {
// shift is >= 1<<32
var ch gc.Node
var cl gc.Node
split64(r, &cl, &ch)
gmove(&ch, &s)
gins(arm.ATST, &s, nil)
var p1 *obj.Prog
if bh.Type.Etype == gc.TINT32 {
p1 = gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
} else {
p1 = gins(arm.AEOR, &ah, &ah)
}
p1.Scond = arm.C_SCOND_NE
p6 = gc.Gbranch(arm.ABNE, nil, 0)
gmove(&cl, &s)
splitclean()
} else {
gmove(r, &s)
p6 = nil
}
gins(arm.ATST, &s, nil)
// shift == 0
p1 = gins(arm.AMOVW, &bl, &al)
p1.Scond = arm.C_SCOND_EQ
p1 = gins(arm.AMOVW, &bh, &ah)
p1.Scond = arm.C_SCOND_EQ
p2 = gc.Gbranch(arm.ABEQ, nil, 0)
// check if shift is < 32
gc.Nodconst(&n1, gc.Types[gc.TUINT32], 32)
gmove(&n1, &creg)
gins(arm.ACMP, &s, &creg)
// MOVW.LO bl>>s, al
p1 = gregshift(arm.AMOVW, &bl, arm.SHIFT_LR, &s, &al)
p1.Scond = arm.C_SCOND_LO
// SUB.LO s,creg
p1 = gins(arm.ASUB, &s, &creg)
p1.Scond = arm.C_SCOND_LO
// OR.LO bh<<(32-s), al
p1 = gregshift(arm.AORR, &bh, arm.SHIFT_LL, &creg, &al)
p1.Scond = arm.C_SCOND_LO
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->s, ah
p1 = gregshift(arm.AMOVW, &bh, arm.SHIFT_AR, &s, &ah)
} else {
// MOVW bh>>s, ah
p1 = gregshift(arm.AMOVW, &bh, arm.SHIFT_LR, &s, &ah)
}
p1.Scond = arm.C_SCOND_LO
// BLO end
p3 = gc.Gbranch(arm.ABLO, nil, 0)
// shift == 32
p1 = gins(arm.AMOVW, &bh, &al)
p1.Scond = arm.C_SCOND_EQ
if bh.Type.Etype == gc.TINT32 {
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
} else {
gins(arm.AEOR, &ah, &ah)
}
p4 = gc.Gbranch(arm.ABEQ, nil, 0)
// check if shift is < 64
gc.Nodconst(&n1, gc.Types[gc.TUINT32], 64)
gmove(&n1, &creg)
gins(arm.ACMP, &s, &creg)
// MOVW.LO creg>>1, creg
p1 = gshift(arm.AMOVW, &creg, arm.SHIFT_LR, 1, &creg)
p1.Scond = arm.C_SCOND_LO
// SUB.LO creg, s
p1 = gins(arm.ASUB, &creg, &s)
p1.Scond = arm.C_SCOND_LO
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->(s-32), al
p1 := gregshift(arm.AMOVW, &bh, arm.SHIFT_AR, &s, &al)
p1.Scond = arm.C_SCOND_LO
} else {
// MOVW bh>>(v-32), al
p1 := gregshift(arm.AMOVW, &bh, arm.SHIFT_LR, &s, &al)
p1.Scond = arm.C_SCOND_LO
}
// BLO end
p5 = gc.Gbranch(arm.ABLO, nil, 0)
// s >= 64
if p6 != nil {
gc.Patch(p6, gc.Pc)
}
if bh.Type.Etype == gc.TINT32 {
// MOVW bh->31, al
gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &al)
} else {
gins(arm.AEOR, &al, &al)
}
gc.Patch(p2, gc.Pc)
gc.Patch(p3, gc.Pc)
gc.Patch(p4, gc.Pc)
gc.Patch(p5, gc.Pc)
gc.Regfree(&s)
gc.Regfree(&creg)
orsh_break:
gc.Regfree(&bl)
gc.Regfree(&bh)
// TODO(kaib): literal optimizations
// make constant the right side (it usually is anyway).
// if(lo1.op == OLITERAL) {
// nswap(&lo1, &lo2);
// nswap(&hi1, &hi2);
// }
// if(lo2.op == OLITERAL) {
// // special cases for constants.
// lv = mpgetfix(lo2.val.u.xval);
// hv = mpgetfix(hi2.val.u.xval);
// splitclean(); // right side
// split64(res, &lo2, &hi2);
// switch(n->op) {
// case OXOR:
// gmove(&lo1, &lo2);
// gmove(&hi1, &hi2);
// switch(lv) {
// case 0:
// break;
// case 0xffffffffu:
// gins(ANOTL, N, &lo2);
// break;
// default:
// gins(AXORL, ncon(lv), &lo2);
// break;
// }
// switch(hv) {
// case 0:
// break;
// case 0xffffffffu:
// gins(ANOTL, N, &hi2);
// break;
// default:
// gins(AXORL, ncon(hv), &hi2);
// break;
// }
// break;
// case OAND:
// switch(lv) {
// case 0:
// gins(AMOVL, ncon(0), &lo2);
// break;
// default:
// gmove(&lo1, &lo2);
// if(lv != 0xffffffffu)
// gins(AANDL, ncon(lv), &lo2);
// break;
// }
// switch(hv) {
// case 0:
// gins(AMOVL, ncon(0), &hi2);
// break;
// default:
// gmove(&hi1, &hi2);
// if(hv != 0xffffffffu)
// gins(AANDL, ncon(hv), &hi2);
// break;
// }
// break;
// case OOR:
// switch(lv) {
// case 0:
// gmove(&lo1, &lo2);
// break;
// case 0xffffffffu:
// gins(AMOVL, ncon(0xffffffffu), &lo2);
// break;
// default:
// gmove(&lo1, &lo2);
// gins(AORL, ncon(lv), &lo2);
// break;
// }
// switch(hv) {
// case 0:
// gmove(&hi1, &hi2);
// break;
// case 0xffffffffu:
// gins(AMOVL, ncon(0xffffffffu), &hi2);
// break;
// default:
// gmove(&hi1, &hi2);
// gins(AORL, ncon(hv), &hi2);
// break;
// }
// break;
// }
// splitclean();
// splitclean();
// goto out;
// }
case gc.OXOR,
gc.OAND,
gc.OOR:
var n1 gc.Node
gc.Regalloc(&n1, lo1.Type, nil)
gins(arm.AMOVW, &lo1, &al)
gins(arm.AMOVW, &hi1, &ah)
gins(arm.AMOVW, &lo2, &n1)
gins(optoas(n.Op, lo1.Type), &n1, &al)
gins(arm.AMOVW, &hi2, &n1)
gins(optoas(n.Op, lo1.Type), &n1, &ah)
gc.Regfree(&n1)
}
if gc.Is64(r.Type) {
splitclean()
}
splitclean()
split64(res, &lo1, &hi1)
gins(arm.AMOVW, &al, &lo1)
gins(arm.AMOVW, &ah, &hi1)
splitclean()
//out:
gc.Regfree(&al)
gc.Regfree(&ah)
}
/*
* generate comparison of nl, nr, both 64-bit.
* nl is memory; nr is constant or memory.
*/
func cmp64(nl *gc.Node, nr *gc.Node, op gc.Op, likely int, to *obj.Prog) {
var lo1 gc.Node
var hi1 gc.Node
var lo2 gc.Node
var hi2 gc.Node
var r1 gc.Node
var r2 gc.Node
split64(nl, &lo1, &hi1)
split64(nr, &lo2, &hi2)
// compare most significant word;
// if they differ, we're done.
t := hi1.Type
gc.Regalloc(&r1, gc.Types[gc.TINT32], nil)
gc.Regalloc(&r2, gc.Types[gc.TINT32], nil)
gins(arm.AMOVW, &hi1, &r1)
gins(arm.AMOVW, &hi2, &r2)
gins(arm.ACMP, &r1, &r2)
gc.Regfree(&r1)
gc.Regfree(&r2)
var br *obj.Prog
switch op {
default:
gc.Fatalf("cmp64 %v %v", gc.Oconv(int(op), 0), t)
// cmp hi
// bne L
// cmp lo
// beq to
// L:
case gc.OEQ:
br = gc.Gbranch(arm.ABNE, nil, -likely)
// cmp hi
// bne to
// cmp lo
// bne to
case gc.ONE:
gc.Patch(gc.Gbranch(arm.ABNE, nil, likely), to)
// cmp hi
// bgt to
// blt L
// cmp lo
// bge to (or bgt 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
// blt to
// bgt L
// cmp lo
// ble 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
gc.Regalloc(&r1, gc.Types[gc.TINT32], nil)
gc.Regalloc(&r2, gc.Types[gc.TINT32], nil)
gins(arm.AMOVW, &lo1, &r1)
gins(arm.AMOVW, &lo2, &r2)
gins(arm.ACMP, &r1, &r2)
gc.Regfree(&r1)
gc.Regfree(&r2)
// 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()
}
