/*
* generate an addressable node in res, containing the value of n.
* n is an array index, and might be any size; res width is <= 32-bit.
* returns Prog* to patch to panic call.
*/
func igenindex(n *gc.Node, res *gc.Node, bounded bool) *obj.Prog {
if !gc.Is64(n.Type) {
if n.Addable && (gc.Simtype[n.Etype] == gc.TUINT32 || gc.Simtype[n.Etype] == gc.TINT32) {
// nothing to do.
*res = *n
} else {
gc.Tempname(res, gc.Types[gc.TUINT32])
gc.Cgen(n, res)
}
return nil
}
var tmp gc.Node
gc.Tempname(&tmp, gc.Types[gc.TINT64])
gc.Cgen(n, &tmp)
var lo gc.Node
var hi gc.Node
split64(&tmp, &lo, &hi)
gc.Tempname(res, gc.Types[gc.TUINT32])
gmove(&lo, res)
if bounded {
splitclean()
return nil
}
var zero gc.Node
gc.Nodconst(&zero, gc.Types[gc.TINT32], 0)
gins(x86.ACMPL, &hi, &zero)
splitclean()
return gc.Gbranch(x86.AJNE, nil, +1)
}
/*
* generate byte multiply:
* res = nl * nr
* there is no 2-operand byte multiply instruction so
* we do a full-width multiplication and truncate afterwards.
*/
func cgen_bmul(op gc.Op, nl *gc.Node, nr *gc.Node, res *gc.Node) bool {
if optoas(op, nl.Type) != x86.AIMULB {
return false
}
// copy from byte to full registers
t := gc.Types[gc.TUINT32]
if gc.Issigned[nl.Type.Etype] {
t = gc.Types[gc.TINT32]
}
// largest ullman on left.
if nl.Ullman < nr.Ullman {
nl, nr = nr, nl
}
var nt gc.Node
gc.Tempname(&nt, nl.Type)
gc.Cgen(nl, &nt)
var n1 gc.Node
gc.Regalloc(&n1, t, res)
gc.Cgen(nr, &n1)
var n2 gc.Node
gc.Regalloc(&n2, t, nil)
gmove(&nt, &n2)
a := optoas(op, t)
gins(a, &n2, &n1)
gc.Regfree(&n2)
gmove(&n1, res)
gc.Regfree(&n1)
return true
}
/*
* generate array index into res.
* n might be any size; res is 32-bit.
* returns Prog* to patch to panic call.
*/
func cgenindex(n *gc.Node, res *gc.Node, bounded bool) *obj.Prog {
if !gc.Is64(n.Type) {
gc.Cgen(n, res)
return nil
}
var tmp gc.Node
gc.Tempname(&tmp, gc.Types[gc.TINT64])
gc.Cgen(n, &tmp)
var lo gc.Node
var hi gc.Node
split64(&tmp, &lo, &hi)
gmove(&lo, res)
if bounded {
splitclean()
return nil
}
var n1 gc.Node
gc.Regalloc(&n1, gc.Types[gc.TINT32], nil)
var n2 gc.Node
gc.Regalloc(&n2, gc.Types[gc.TINT32], nil)
var zero gc.Node
gc.Nodconst(&zero, gc.Types[gc.TINT32], 0)
gmove(&hi, &n1)
gmove(&zero, &n2)
gins(arm.ACMP, &n1, &n2)
gc.Regfree(&n2)
gc.Regfree(&n1)
splitclean()
return gc.Gbranch(arm.ABNE, nil, -1)
}
func savex(dr int, x *gc.Node, oldx *gc.Node, res *gc.Node, t *gc.Type) {
r := gc.GetReg(dr)
gc.Nodreg(x, gc.Types[gc.TINT32], dr)
// save current ax and dx if they are live
// and not the destination
*oldx = gc.Node{}
if r > 0 && !gc.Samereg(x, res) {
gc.Tempname(oldx, gc.Types[gc.TINT32])
gmove(x, oldx)
}
gc.Regalloc(x, t, x)
}
/*
* generate floating-point operation.
*/
func cgen_float(n *gc.Node, res *gc.Node) {
nl := n.Left
switch n.Op {
case gc.OEQ,
gc.ONE,
gc.OLT,
gc.OLE,
gc.OGE:
p1 := gc.Gbranch(obj.AJMP, nil, 0)
p2 := gc.Pc
gmove(gc.Nodbool(true), res)
p3 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
gc.Bgen(n, true, 0, p2)
gmove(gc.Nodbool(false), res)
gc.Patch(p3, gc.Pc)
return
case gc.OPLUS:
gc.Cgen(nl, res)
return
case gc.OCONV:
if gc.Eqtype(n.Type, nl.Type) || gc.Noconv(n.Type, nl.Type) {
gc.Cgen(nl, res)
return
}
var n2 gc.Node
gc.Tempname(&n2, n.Type)
var n1 gc.Node
gc.Mgen(nl, &n1, res)
gmove(&n1, &n2)
gmove(&n2, res)
gc.Mfree(&n1)
return
}
if gc.Thearch.Use387 {
cgen_float387(n, res)
} else {
cgen_floatsse(n, res)
}
}
/*
* generate high multiply:
* res = (nl*nr) >> width
*/
func cgen_hmul(nl *gc.Node, nr *gc.Node, res *gc.Node) {
var n1 gc.Node
var n2 gc.Node
t := nl.Type
a := optoas(gc.OHMUL, t)
// gen nl in n1.
gc.Tempname(&n1, t)
gc.Cgen(nl, &n1)
// gen nr in n2.
gc.Regalloc(&n2, t, res)
gc.Cgen(nr, &n2)
var ax, oldax, dx, olddx gc.Node
savex(x86.REG_AX, &ax, &oldax, res, gc.Types[gc.TUINT32])
savex(x86.REG_DX, &dx, &olddx, res, gc.Types[gc.TUINT32])
gmove(&n2, &ax)
gins(a, &n1, nil)
gc.Regfree(&n2)
if t.Width == 1 {
// byte multiply behaves differently.
var byteAH, byteDX gc.Node
gc.Nodreg(&byteAH, t, x86.REG_AH)
gc.Nodreg(&byteDX, t, x86.REG_DX)
gmove(&byteAH, &byteDX)
}
gmove(&dx, res)
restx(&ax, &oldax)
restx(&dx, &olddx)
}
func memname(n *gc.Node, t *gc.Type) {
gc.Tempname(n, t)
n.Sym = gc.Lookup("." + n.Sym.Name[1:]) // keep optimizer from registerizing
n.Orig.Sym = n.Sym
}
func floatmove_sse(f *gc.Node, t *gc.Node) {
var r1 gc.Node
var cvt *gc.Type
var a int
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
switch uint32(ft)<<16 | uint32(tt) {
// should not happen
default:
gc.Fatalf("gmove %v -> %v", f, t)
return
// convert via int32.
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT16,
gc.TFLOAT32<<16 | gc.TINT8,
gc.TFLOAT32<<16 | gc.TUINT16,
gc.TFLOAT32<<16 | gc.TUINT8,
gc.TFLOAT64<<16 | gc.TINT16,
gc.TFLOAT64<<16 | gc.TINT8,
gc.TFLOAT64<<16 | gc.TUINT16,
gc.TFLOAT64<<16 | gc.TUINT8:
cvt = gc.Types[gc.TINT32]
goto hard
// convert via int64.
case gc.TFLOAT32<<16 | gc.TUINT32,
gc.TFLOAT64<<16 | gc.TUINT32:
cvt = gc.Types[gc.TINT64]
goto hardmem
case gc.TFLOAT32<<16 | gc.TINT32:
a = x86.ACVTTSS2SL
goto rdst
case gc.TFLOAT64<<16 | gc.TINT32:
a = x86.ACVTTSD2SL
goto rdst
// convert via int32 memory
/*
* integer to float
*/
case gc.TINT8<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TINT16<<16 | gc.TFLOAT32,
gc.TINT16<<16 | gc.TFLOAT64,
gc.TUINT16<<16 | gc.TFLOAT32,
gc.TUINT16<<16 | gc.TFLOAT64,
gc.TUINT8<<16 | gc.TFLOAT32,
gc.TUINT8<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT32]
goto hard
// convert via int64 memory
case gc.TUINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT64]
goto hardmem
case gc.TINT32<<16 | gc.TFLOAT32:
a = x86.ACVTSL2SS
goto rdst
case gc.TINT32<<16 | gc.TFLOAT64:
a = x86.ACVTSL2SD
goto rdst
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32:
a = x86.AMOVSS
case gc.TFLOAT64<<16 | gc.TFLOAT64:
a = x86.AMOVSD
case gc.TFLOAT32<<16 | gc.TFLOAT64:
a = x86.ACVTSS2SD
goto rdst
case gc.TFLOAT64<<16 | gc.TFLOAT32:
a = x86.ACVTSD2SS
goto rdst
}
gins(a, f, t)
return
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
// requires memory intermediate
hardmem:
gc.Tempname(&r1, cvt)
gmove(f, &r1)
gmove(&r1, t)
return
// requires register destination
rdst:
gc.Regalloc(&r1, t.Type, t)
gins(a, f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
}
func floatmove_387(f *gc.Node, t *gc.Node) {
var r1 gc.Node
var a int
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
cvt := t.Type
switch uint32(ft)<<16 | uint32(tt) {
default:
goto fatal
/*
* float to integer
*/
case gc.TFLOAT32<<16 | gc.TINT16,
gc.TFLOAT32<<16 | gc.TINT32,
gc.TFLOAT32<<16 | gc.TINT64,
gc.TFLOAT64<<16 | gc.TINT16,
gc.TFLOAT64<<16 | gc.TINT32,
gc.TFLOAT64<<16 | gc.TINT64:
if t.Op == gc.OREGISTER {
goto hardmem
}
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[ft], x86.REG_F0)
if f.Op != gc.OREGISTER {
if ft == gc.TFLOAT32 {
gins(x86.AFMOVF, f, &r1)
} else {
gins(x86.AFMOVD, f, &r1)
}
}
// set round to zero mode during conversion
var t1 gc.Node
memname(&t1, gc.Types[gc.TUINT16])
var t2 gc.Node
memname(&t2, gc.Types[gc.TUINT16])
gins(x86.AFSTCW, nil, &t1)
gins(x86.AMOVW, ncon(0xf7f), &t2)
gins(x86.AFLDCW, &t2, nil)
if tt == gc.TINT16 {
gins(x86.AFMOVWP, &r1, t)
} else if tt == gc.TINT32 {
gins(x86.AFMOVLP, &r1, t)
} else {
gins(x86.AFMOVVP, &r1, t)
}
gins(x86.AFLDCW, &t1, nil)
return
// convert via int32.
case gc.TFLOAT32<<16 | gc.TINT8,
gc.TFLOAT32<<16 | gc.TUINT16,
gc.TFLOAT32<<16 | gc.TUINT8,
gc.TFLOAT64<<16 | gc.TINT8,
gc.TFLOAT64<<16 | gc.TUINT16,
gc.TFLOAT64<<16 | gc.TUINT8:
var t1 gc.Node
gc.Tempname(&t1, gc.Types[gc.TINT32])
gmove(f, &t1)
switch tt {
default:
gc.Fatalf("gmove %v", t)
case gc.TINT8:
gins(x86.ACMPL, &t1, ncon(-0x80&(1<<32-1)))
p1 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TINT32]), nil, -1)
gins(x86.ACMPL, &t1, ncon(0x7f))
p2 := gc.Gbranch(optoas(gc.OGT, gc.Types[gc.TINT32]), nil, -1)
p3 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
gc.Patch(p2, gc.Pc)
gmove(ncon(-0x80&(1<<32-1)), &t1)
gc.Patch(p3, gc.Pc)
gmove(&t1, t)
case gc.TUINT8:
gins(x86.ATESTL, ncon(0xffffff00), &t1)
p1 := gc.Gbranch(x86.AJEQ, nil, +1)
gins(x86.AMOVL, ncon(0), &t1)
gc.Patch(p1, gc.Pc)
gmove(&t1, t)
case gc.TUINT16:
gins(x86.ATESTL, ncon(0xffff0000), &t1)
p1 := gc.Gbranch(x86.AJEQ, nil, +1)
gins(x86.AMOVL, ncon(0), &t1)
gc.Patch(p1, gc.Pc)
gmove(&t1, t)
}
return
// convert via int64.
case gc.TFLOAT32<<16 | gc.TUINT32,
gc.TFLOAT64<<16 | gc.TUINT32:
cvt = gc.Types[gc.TINT64]
goto hardmem
/*
* integer to float
*/
case gc.TINT16<<16 | gc.TFLOAT32,
gc.TINT16<<16 | gc.TFLOAT64,
gc.TINT32<<16 | gc.TFLOAT32,
gc.TINT32<<16 | gc.TFLOAT64,
gc.TINT64<<16 | gc.TFLOAT32,
gc.TINT64<<16 | gc.TFLOAT64:
if t.Op != gc.OREGISTER {
goto hard
}
if f.Op == gc.OREGISTER {
cvt = f.Type
goto hardmem
}
switch ft {
case gc.TINT16:
a = x86.AFMOVW
case gc.TINT32:
a = x86.AFMOVL
default:
a = x86.AFMOVV
}
// convert via int32 memory
case gc.TINT8<<16 | gc.TFLOAT32,
gc.TINT8<<16 | gc.TFLOAT64,
gc.TUINT16<<16 | gc.TFLOAT32,
gc.TUINT16<<16 | gc.TFLOAT64,
gc.TUINT8<<16 | gc.TFLOAT32,
gc.TUINT8<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT32]
goto hardmem
// convert via int64 memory
case gc.TUINT32<<16 | gc.TFLOAT32,
gc.TUINT32<<16 | gc.TFLOAT64:
cvt = gc.Types[gc.TINT64]
goto hardmem
// The way the code generator uses floating-point
// registers, a move from F0 to F0 is intended as a no-op.
// On the x86, it's not: it pushes a second copy of F0
// on the floating point stack. So toss it away here.
// Also, F0 is the *only* register we ever evaluate
// into, so we should only see register/register as F0/F0.
/*
* float to float
*/
case gc.TFLOAT32<<16 | gc.TFLOAT32,
gc.TFLOAT64<<16 | gc.TFLOAT64:
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
if f.Op == gc.OREGISTER && t.Op == gc.OREGISTER {
if f.Reg != x86.REG_F0 || t.Reg != x86.REG_F0 {
goto fatal
}
return
}
a = x86.AFMOVF
if ft == gc.TFLOAT64 {
a = x86.AFMOVD
}
if gc.Ismem(t) {
if f.Op != gc.OREGISTER || f.Reg != x86.REG_F0 {
gc.Fatalf("gmove %v", f)
}
a = x86.AFMOVFP
if ft == gc.TFLOAT64 {
a = x86.AFMOVDP
}
}
case gc.TFLOAT32<<16 | gc.TFLOAT64:
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
if f.Op == gc.OREGISTER && t.Op == gc.OREGISTER {
if f.Reg != x86.REG_F0 || t.Reg != x86.REG_F0 {
goto fatal
}
return
}
if f.Op == gc.OREGISTER {
gins(x86.AFMOVDP, f, t)
} else {
gins(x86.AFMOVF, f, t)
}
return
case gc.TFLOAT64<<16 | gc.TFLOAT32:
if gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
if f.Op == gc.OREGISTER && t.Op == gc.OREGISTER {
var r1 gc.Node
gc.Tempname(&r1, gc.Types[gc.TFLOAT32])
gins(x86.AFMOVFP, f, &r1)
gins(x86.AFMOVF, &r1, t)
return
}
if f.Op == gc.OREGISTER {
gins(x86.AFMOVFP, f, t)
} else {
gins(x86.AFMOVD, f, t)
}
return
}
gins(a, f, t)
return
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
// requires memory intermediate
hardmem:
gc.Tempname(&r1, cvt)
gmove(f, &r1)
gmove(&r1, t)
return
// should not happen
fatal:
gc.Fatalf("gmove %v -> %v", gc.Nconv(f, obj.FmtLong), gc.Nconv(t, obj.FmtLong))
return
}
func floatmove(f *gc.Node, t *gc.Node) {
var r1 gc.Node
ft := gc.Simsimtype(f.Type)
tt := gc.Simsimtype(t.Type)
cvt := t.Type
// cannot have two floating point memory operands.
if gc.Isfloat[ft] && gc.Isfloat[tt] && gc.Ismem(f) && gc.Ismem(t) {
goto hard
}
// convert constant to desired type
if f.Op == gc.OLITERAL {
var con gc.Node
f.Convconst(&con, t.Type)
f = &con
ft = gc.Simsimtype(con.Type)
// some constants can't move directly to memory.
if gc.Ismem(t) {
// float constants come from memory.
if gc.Isfloat[tt] {
goto hard
}
}
}
// value -> value copy, only one memory operand.
// figure out the instruction to use.
// break out of switch for one-instruction gins.
// goto rdst for "destination must be register".
// goto hard for "convert to cvt type first".
// otherwise handle and return.
switch uint32(ft)<<16 | uint32(tt) {
default:
if gc.Thearch.Use387 {
floatmove_387(f, t)
} else {
floatmove_sse(f, t)
}
return
// float to very long integer.
case gc.TFLOAT32<<16 | gc.TINT64,
gc.TFLOAT64<<16 | gc.TINT64:
if f.Op == gc.OREGISTER {
cvt = f.Type
goto hardmem
}
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[ft], x86.REG_F0)
if ft == gc.TFLOAT32 {
gins(x86.AFMOVF, f, &r1)
} else {
gins(x86.AFMOVD, f, &r1)
}
// set round to zero mode during conversion
var t1 gc.Node
memname(&t1, gc.Types[gc.TUINT16])
var t2 gc.Node
memname(&t2, gc.Types[gc.TUINT16])
gins(x86.AFSTCW, nil, &t1)
gins(x86.AMOVW, ncon(0xf7f), &t2)
gins(x86.AFLDCW, &t2, nil)
if tt == gc.TINT16 {
gins(x86.AFMOVWP, &r1, t)
} else if tt == gc.TINT32 {
gins(x86.AFMOVLP, &r1, t)
} else {
gins(x86.AFMOVVP, &r1, t)
}
gins(x86.AFLDCW, &t1, nil)
return
case gc.TFLOAT32<<16 | gc.TUINT64,
gc.TFLOAT64<<16 | gc.TUINT64:
if !gc.Ismem(f) {
cvt = f.Type
goto hardmem
}
bignodes()
var f0 gc.Node
gc.Nodreg(&f0, gc.Types[ft], x86.REG_F0)
var f1 gc.Node
gc.Nodreg(&f1, gc.Types[ft], x86.REG_F0+1)
var ax gc.Node
gc.Nodreg(&ax, gc.Types[gc.TUINT16], x86.REG_AX)
if ft == gc.TFLOAT32 {
gins(x86.AFMOVF, f, &f0)
} else {
gins(x86.AFMOVD, f, &f0)
}
// if 0 > v { answer = 0 }
gins(x86.AFMOVD, &zerof, &f0)
gins(x86.AFUCOMP, &f0, &f1)
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
p1 := gc.Gbranch(optoas(gc.OGT, gc.Types[tt]), nil, 0)
// if 1<<64 <= v { answer = 0 too }
gins(x86.AFMOVD, &two64f, &f0)
gins(x86.AFUCOMP, &f0, &f1)
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
p2 := gc.Gbranch(optoas(gc.OGT, gc.Types[tt]), nil, 0)
gc.Patch(p1, gc.Pc)
gins(x86.AFMOVVP, &f0, t) // don't care about t, but will pop the stack
var thi gc.Node
var tlo gc.Node
split64(t, &tlo, &thi)
gins(x86.AMOVL, ncon(0), &tlo)
gins(x86.AMOVL, ncon(0), &thi)
splitclean()
p1 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p2, gc.Pc)
// in range; algorithm is:
// if small enough, use native float64 -> int64 conversion.
// otherwise, subtract 2^63, convert, and add it back.
// set round to zero mode during conversion
var t1 gc.Node
memname(&t1, gc.Types[gc.TUINT16])
var t2 gc.Node
memname(&t2, gc.Types[gc.TUINT16])
gins(x86.AFSTCW, nil, &t1)
gins(x86.AMOVW, ncon(0xf7f), &t2)
gins(x86.AFLDCW, &t2, nil)
// actual work
gins(x86.AFMOVD, &two63f, &f0)
gins(x86.AFUCOMP, &f0, &f1)
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
p2 = gc.Gbranch(optoas(gc.OLE, gc.Types[tt]), nil, 0)
gins(x86.AFMOVVP, &f0, t)
p3 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p2, gc.Pc)
gins(x86.AFMOVD, &two63f, &f0)
gins(x86.AFSUBDP, &f0, &f1)
gins(x86.AFMOVVP, &f0, t)
split64(t, &tlo, &thi)
gins(x86.AXORL, ncon(0x80000000), &thi) // + 2^63
gc.Patch(p3, gc.Pc)
splitclean()
// restore rounding mode
gins(x86.AFLDCW, &t1, nil)
gc.Patch(p1, gc.Pc)
return
/*
* integer to float
*/
case gc.TINT64<<16 | gc.TFLOAT32,
gc.TINT64<<16 | gc.TFLOAT64:
if t.Op == gc.OREGISTER {
goto hardmem
}
var f0 gc.Node
gc.Nodreg(&f0, t.Type, x86.REG_F0)
gins(x86.AFMOVV, f, &f0)
if tt == gc.TFLOAT32 {
gins(x86.AFMOVFP, &f0, t)
} else {
gins(x86.AFMOVDP, &f0, t)
}
return
// algorithm is:
// if small enough, use native int64 -> float64 conversion.
// otherwise, halve (rounding to odd?), convert, and double.
case gc.TUINT64<<16 | gc.TFLOAT32,
gc.TUINT64<<16 | gc.TFLOAT64:
var ax gc.Node
gc.Nodreg(&ax, gc.Types[gc.TUINT32], x86.REG_AX)
var dx gc.Node
gc.Nodreg(&dx, gc.Types[gc.TUINT32], x86.REG_DX)
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TUINT32], x86.REG_CX)
var t1 gc.Node
gc.Tempname(&t1, f.Type)
var tlo gc.Node
var thi gc.Node
split64(&t1, &tlo, &thi)
gmove(f, &t1)
gins(x86.ACMPL, &thi, ncon(0))
p1 := gc.Gbranch(x86.AJLT, nil, 0)
// native
var r1 gc.Node
gc.Nodreg(&r1, gc.Types[tt], x86.REG_F0)
gins(x86.AFMOVV, &t1, &r1)
if tt == gc.TFLOAT32 {
gins(x86.AFMOVFP, &r1, t)
} else {
gins(x86.AFMOVDP, &r1, t)
}
p2 := gc.Gbranch(obj.AJMP, nil, 0)
// simulated
gc.Patch(p1, gc.Pc)
gmove(&tlo, &ax)
gmove(&thi, &dx)
p1 = gins(x86.ASHRL, ncon(1), &ax)
p1.From.Index = x86.REG_DX // double-width shift DX -> AX
p1.From.Scale = 0
gins(x86.AMOVL, ncon(0), &cx)
gins(x86.ASETCC, nil, &cx)
gins(x86.AORL, &cx, &ax)
gins(x86.ASHRL, ncon(1), &dx)
gmove(&dx, &thi)
gmove(&ax, &tlo)
gc.Nodreg(&r1, gc.Types[tt], x86.REG_F0)
var r2 gc.Node
gc.Nodreg(&r2, gc.Types[tt], x86.REG_F0+1)
gins(x86.AFMOVV, &t1, &r1)
gins(x86.AFMOVD, &r1, &r1)
gins(x86.AFADDDP, &r1, &r2)
if tt == gc.TFLOAT32 {
gins(x86.AFMOVFP, &r1, t)
} else {
gins(x86.AFMOVDP, &r1, t)
}
gc.Patch(p2, gc.Pc)
splitclean()
return
}
// requires register intermediate
hard:
gc.Regalloc(&r1, cvt, t)
gmove(f, &r1)
gmove(&r1, t)
gc.Regfree(&r1)
return
// requires memory intermediate
hardmem:
gc.Tempname(&r1, cvt)
gmove(f, &r1)
gmove(&r1, t)
return
}
func blockcopy(n, res *gc.Node, osrc, odst, w int64) {
var dst gc.Node
gc.Nodreg(&dst, gc.Types[gc.Tptr], x86.REG_DI)
var src gc.Node
gc.Nodreg(&src, gc.Types[gc.Tptr], x86.REG_SI)
var tsrc gc.Node
gc.Tempname(&tsrc, gc.Types[gc.Tptr])
var tdst gc.Node
gc.Tempname(&tdst, gc.Types[gc.Tptr])
if !n.Addable {
gc.Agen(n, &tsrc)
}
if !res.Addable {
gc.Agen(res, &tdst)
}
if n.Addable {
gc.Agen(n, &src)
} else {
gmove(&tsrc, &src)
}
if res.Op == gc.ONAME {
gc.Gvardef(res)
}
if res.Addable {
gc.Agen(res, &dst)
} else {
gmove(&tdst, &dst)
}
c := int32(w % 4) // bytes
q := int32(w / 4) // doublewords
// if we are copying forward on the stack and
// the src and dst overlap, then reverse direction
if osrc < odst && int64(odst) < int64(osrc)+w {
// reverse direction
gins(x86.ASTD, nil, nil) // set direction flag
if c > 0 {
gconreg(x86.AADDL, w-1, x86.REG_SI)
gconreg(x86.AADDL, w-1, x86.REG_DI)
gconreg(x86.AMOVL, int64(c), x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSB, nil, nil) // MOVB *(SI)-,*(DI)-
}
if q > 0 {
if c > 0 {
gconreg(x86.AADDL, -3, x86.REG_SI)
gconreg(x86.AADDL, -3, x86.REG_DI)
} else {
gconreg(x86.AADDL, w-4, x86.REG_SI)
gconreg(x86.AADDL, w-4, x86.REG_DI)
}
gconreg(x86.AMOVL, int64(q), x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSL, nil, nil) // MOVL *(SI)-,*(DI)-
}
// we leave with the flag clear
gins(x86.ACLD, nil, nil)
} else {
gins(x86.ACLD, nil, nil) // paranoia. TODO(rsc): remove?
// normal direction
if q > 128 || (q >= 4 && gc.Nacl) {
gconreg(x86.AMOVL, int64(q), x86.REG_CX)
gins(x86.AREP, nil, nil) // repeat
gins(x86.AMOVSL, nil, nil) // MOVL *(SI)+,*(DI)+
} else if q >= 4 {
p := gins(obj.ADUFFCOPY, nil, nil)
p.To.Type = obj.TYPE_ADDR
p.To.Sym = gc.Linksym(gc.Pkglookup("duffcopy", gc.Runtimepkg))
// 10 and 128 = magic constants: see ../../runtime/asm_386.s
p.To.Offset = 10 * (128 - int64(q))
} else if !gc.Nacl && c == 0 {
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TINT32], x86.REG_CX)
// We don't need the MOVSL side-effect of updating SI and DI,
// and issuing a sequence of MOVLs directly is faster.
src.Op = gc.OINDREG
dst.Op = gc.OINDREG
for q > 0 {
gmove(&src, &cx) // MOVL x+(SI),CX
gmove(&cx, &dst) // MOVL CX,x+(DI)
src.Xoffset += 4
dst.Xoffset += 4
q--
}
} else {
for q > 0 {
gins(x86.AMOVSL, nil, nil) // MOVL *(SI)+,*(DI)+
q--
}
}
for c > 0 {
gins(x86.AMOVSB, nil, nil) // MOVB *(SI)+,*(DI)+
c--
}
}
}
/*
* 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 shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op gc.Op, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Type.Width > 4 {
gc.Fatalf("cgen_shift %v", nl.Type)
}
w := int(nl.Type.Width * 8)
if op == gc.OLROT {
v := nr.Int()
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
if w == 32 {
gc.Cgen(nl, &n1)
gshift(arm.AMOVW, &n1, arm.SHIFT_RR, int32(w)-int32(v), &n1)
} else {
var n2 gc.Node
gc.Regalloc(&n2, nl.Type, nil)
gc.Cgen(nl, &n2)
gshift(arm.AMOVW, &n2, arm.SHIFT_LL, int32(v), &n1)
gshift(arm.AORR, &n2, arm.SHIFT_LR, int32(w)-int32(v), &n1)
gc.Regfree(&n2)
// Ensure sign/zero-extended result.
gins(optoas(gc.OAS, nl.Type), &n1, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
if nr.Op == gc.OLITERAL {
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gc.Cgen(nl, &n1)
sc := uint64(nr.Int())
if sc == 0 {
} else // nothing to do
if sc >= uint64(nl.Type.Width*8) {
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gshift(arm.AMOVW, &n1, arm.SHIFT_AR, int32(w), &n1)
} else {
gins(arm.AEOR, &n1, &n1)
}
} else {
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gshift(arm.AMOVW, &n1, arm.SHIFT_AR, int32(sc), &n1)
} else if op == gc.ORSH {
gshift(arm.AMOVW, &n1, arm.SHIFT_LR, int32(sc), &n1) // OLSH
} else {
gshift(arm.AMOVW, &n1, arm.SHIFT_LL, int32(sc), &n1)
}
}
if w < 32 && op == gc.OLSH {
gins(optoas(gc.OAS, nl.Type), &n1, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
tr := nr.Type
var t gc.Node
var n1 gc.Node
var n2 gc.Node
var n3 gc.Node
if tr.Width > 4 {
var nt gc.Node
gc.Tempname(&nt, nr.Type)
if nl.Ullman >= nr.Ullman {
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
gc.Cgen(nr, &nt)
n1 = nt
} else {
gc.Cgen(nr, &nt)
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
}
var hi gc.Node
var lo gc.Node
split64(&nt, &lo, &hi)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], nil)
gc.Regalloc(&n3, gc.Types[gc.TUINT32], nil)
gmove(&lo, &n1)
gmove(&hi, &n3)
splitclean()
gins(arm.ATST, &n3, nil)
gc.Nodconst(&t, gc.Types[gc.TUINT32], int64(w))
p1 := gins(arm.AMOVW, &t, &n1)
p1.Scond = arm.C_SCOND_NE
tr = gc.Types[gc.TUINT32]
gc.Regfree(&n3)
} else {
if nl.Ullman >= nr.Ullman {
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
gc.Regalloc(&n1, nr.Type, nil)
gc.Cgen(nr, &n1)
} else {
gc.Regalloc(&n1, nr.Type, nil)
gc.Cgen(nr, &n1)
gc.Regalloc(&n2, nl.Type, res)
gc.Cgen(nl, &n2)
}
}
// test for shift being 0
gins(arm.ATST, &n1, nil)
p3 := gc.Gbranch(arm.ABEQ, nil, -1)
// test and fix up large shifts
// TODO: if(!bounded), don't emit some of this.
gc.Regalloc(&n3, tr, nil)
gc.Nodconst(&t, gc.Types[gc.TUINT32], int64(w))
gmove(&t, &n3)
gins(arm.ACMP, &n1, &n3)
if op == gc.ORSH {
var p1 *obj.Prog
var p2 *obj.Prog
if gc.Issigned[nl.Type.Etype] {
p1 = gshift(arm.AMOVW, &n2, arm.SHIFT_AR, int32(w)-1, &n2)
p2 = gregshift(arm.AMOVW, &n2, arm.SHIFT_AR, &n1, &n2)
} else {
p1 = gins(arm.AEOR, &n2, &n2)
p2 = gregshift(arm.AMOVW, &n2, arm.SHIFT_LR, &n1, &n2)
}
p1.Scond = arm.C_SCOND_HS
p2.Scond = arm.C_SCOND_LO
} else {
p1 := gins(arm.AEOR, &n2, &n2)
p2 := gregshift(arm.AMOVW, &n2, arm.SHIFT_LL, &n1, &n2)
p1.Scond = arm.C_SCOND_HS
p2.Scond = arm.C_SCOND_LO
}
gc.Regfree(&n3)
gc.Patch(p3, gc.Pc)
// Left-shift of smaller word must be sign/zero-extended.
if w < 32 && op == gc.OLSH {
gins(optoas(gc.OAS, nl.Type), &n2, &n2)
}
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
func bgen_float(n *gc.Node, wantTrue bool, likely int, to *obj.Prog) {
nl := n.Left
nr := n.Right
op := n.Op
if !wantTrue {
// brcom is not valid on floats when NaN is involved.
p1 := gc.Gbranch(obj.AJMP, nil, 0)
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// No need to avoid re-genning ninit.
bgen_float(n, true, -likely, p2)
gc.Patch(gc.Gbranch(obj.AJMP, nil, 0), to)
gc.Patch(p2, gc.Pc)
return
}
if gc.Thearch.Use387 {
op = gc.Brrev(op) // because the args are stacked
if op == gc.OGE || op == gc.OGT {
// only < and <= work right with NaN; reverse if needed
nl, nr = nr, nl
op = gc.Brrev(op)
}
var ax, n2, tmp gc.Node
gc.Nodreg(&tmp, nr.Type, x86.REG_F0)
gc.Nodreg(&n2, nr.Type, x86.REG_F0+1)
gc.Nodreg(&ax, gc.Types[gc.TUINT16], x86.REG_AX)
if gc.Simsimtype(nr.Type) == gc.TFLOAT64 {
if nl.Ullman > nr.Ullman {
gc.Cgen(nl, &tmp)
gc.Cgen(nr, &tmp)
gins(x86.AFXCHD, &tmp, &n2)
} else {
gc.Cgen(nr, &tmp)
gc.Cgen(nl, &tmp)
}
gins(x86.AFUCOMPP, &tmp, &n2)
} else {
// TODO(rsc): The moves back and forth to memory
// here are for truncating the value to 32 bits.
// This handles 32-bit comparison but presumably
// all the other ops have the same problem.
// We need to figure out what the right general
// solution is, besides telling people to use float64.
var t1 gc.Node
gc.Tempname(&t1, gc.Types[gc.TFLOAT32])
var t2 gc.Node
gc.Tempname(&t2, gc.Types[gc.TFLOAT32])
gc.Cgen(nr, &t1)
gc.Cgen(nl, &t2)
gmove(&t2, &tmp)
gins(x86.AFCOMFP, &t1, &tmp)
}
gins(x86.AFSTSW, nil, &ax)
gins(x86.ASAHF, nil, nil)
} else {
// Not 387
if !nl.Addable {
nl = gc.CgenTemp(nl)
}
if !nr.Addable {
nr = gc.CgenTemp(nr)
}
var n2 gc.Node
gc.Regalloc(&n2, nr.Type, nil)
gmove(nr, &n2)
nr = &n2
if nl.Op != gc.OREGISTER {
var n3 gc.Node
gc.Regalloc(&n3, nl.Type, nil)
gmove(nl, &n3)
nl = &n3
}
if op == gc.OGE || op == gc.OGT {
// only < and <= work right with NopN; reverse if needed
nl, nr = nr, nl
op = gc.Brrev(op)
}
gins(foptoas(gc.OCMP, nr.Type, 0), nl, nr)
if nl.Op == gc.OREGISTER {
gc.Regfree(nl)
}
gc.Regfree(nr)
}
switch op {
case gc.OEQ:
// neither NE nor P
p1 := gc.Gbranch(x86.AJNE, nil, -likely)
p2 := gc.Gbranch(x86.AJPS, nil, -likely)
gc.Patch(gc.Gbranch(obj.AJMP, nil, 0), to)
gc.Patch(p1, gc.Pc)
gc.Patch(p2, gc.Pc)
case gc.ONE:
// either NE or P
gc.Patch(gc.Gbranch(x86.AJNE, nil, likely), to)
gc.Patch(gc.Gbranch(x86.AJPS, nil, likely), to)
default:
gc.Patch(gc.Gbranch(optoas(op, nr.Type), nil, likely), to)
}
}
func cgen_floatsse(n *gc.Node, res *gc.Node) {
var a int
nl := n.Left
nr := n.Right
switch n.Op {
default:
gc.Dump("cgen_floatsse", n)
gc.Fatalf("cgen_floatsse %v", gc.Oconv(int(n.Op), 0))
return
case gc.OMINUS,
gc.OCOM:
nr = gc.Nodintconst(-1)
gc.Convlit(&nr, n.Type)
a = foptoas(gc.OMUL, nl.Type, 0)
goto sbop
// symmetric binary
case gc.OADD,
gc.OMUL:
a = foptoas(n.Op, nl.Type, 0)
goto sbop
// asymmetric binary
case gc.OSUB,
gc.OMOD,
gc.ODIV:
a = foptoas(n.Op, nl.Type, 0)
goto abop
}
sbop: // symmetric binary
if nl.Ullman < nr.Ullman || nl.Op == gc.OLITERAL {
nl, nr = nr, nl
}
abop: // asymmetric binary
if nl.Ullman >= nr.Ullman {
var nt gc.Node
gc.Tempname(&nt, nl.Type)
gc.Cgen(nl, &nt)
var n2 gc.Node
gc.Mgen(nr, &n2, nil)
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gmove(&nt, &n1)
gins(a, &n2, &n1)
gmove(&n1, res)
gc.Regfree(&n1)
gc.Mfree(&n2)
} else {
var n2 gc.Node
gc.Regalloc(&n2, nr.Type, res)
gc.Cgen(nr, &n2)
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, nil)
gc.Cgen(nl, &n1)
gins(a, &n2, &n1)
gc.Regfree(&n2)
gmove(&n1, res)
gc.Regfree(&n1)
}
return
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op gc.Op, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
if nl.Type.Width > 4 {
gc.Fatalf("cgen_shift %v", nl.Type)
}
w := int(nl.Type.Width * 8)
a := optoas(op, nl.Type)
if nr.Op == gc.OLITERAL {
var n2 gc.Node
gc.Tempname(&n2, nl.Type)
gc.Cgen(nl, &n2)
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gmove(&n2, &n1)
sc := uint64(nr.Int())
if sc >= uint64(nl.Type.Width*8) {
// large shift gets 2 shifts by width-1
gins(a, ncon(uint32(w)-1), &n1)
gins(a, ncon(uint32(w)-1), &n1)
} else {
gins(a, nr, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
var oldcx gc.Node
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TUINT32], x86.REG_CX)
if gc.GetReg(x86.REG_CX) > 1 && !gc.Samereg(&cx, res) {
gc.Tempname(&oldcx, gc.Types[gc.TUINT32])
gmove(&cx, &oldcx)
}
var n1 gc.Node
var nt gc.Node
if nr.Type.Width > 4 {
gc.Tempname(&nt, nr.Type)
n1 = nt
} else {
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, nr.Type, &n1) // to hold the shift type in CX
}
var n2 gc.Node
if gc.Samereg(&cx, res) {
gc.Regalloc(&n2, nl.Type, nil)
} else {
gc.Regalloc(&n2, nl.Type, res)
}
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &n2)
gc.Cgen(nr, &n1)
} else {
gc.Cgen(nr, &n1)
gc.Cgen(nl, &n2)
}
// test and fix up large shifts
if bounded {
if nr.Type.Width > 4 {
// delayed reg alloc
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], &n1) // to hold the shift type in CX
var lo gc.Node
var hi gc.Node
split64(&nt, &lo, &hi)
gmove(&lo, &n1)
splitclean()
}
} else {
var p1 *obj.Prog
if nr.Type.Width > 4 {
// delayed reg alloc
gc.Nodreg(&n1, gc.Types[gc.TUINT32], x86.REG_CX)
gc.Regalloc(&n1, gc.Types[gc.TUINT32], &n1) // to hold the shift type in CX
var lo gc.Node
var hi gc.Node
split64(&nt, &lo, &hi)
gmove(&lo, &n1)
gins(optoas(gc.OCMP, gc.Types[gc.TUINT32]), &hi, ncon(0))
p2 := gc.Gbranch(optoas(gc.ONE, gc.Types[gc.TUINT32]), nil, +1)
gins(optoas(gc.OCMP, gc.Types[gc.TUINT32]), &n1, ncon(uint32(w)))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
splitclean()
gc.Patch(p2, gc.Pc)
} else {
gins(optoas(gc.OCMP, nr.Type), &n1, ncon(uint32(w)))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
}
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gins(a, ncon(uint32(w)-1), &n2)
} else {
gmove(ncon(0), &n2)
}
gc.Patch(p1, gc.Pc)
}
gins(a, &n1, &n2)
if oldcx.Op != 0 {
gmove(&oldcx, &cx)
}
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
func igenindex(n *gc.Node, res *gc.Node, bounded bool) *obj.Prog {
gc.Tempname(res, n.Type)
return cgenindex(n, res, bounded)
}
/*
* 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))
}
switch n.Op {
default:
gc.Fatalf("cgen64 %v", gc.Oconv(int(n.Op), 0))
case gc.OMINUS:
gc.Cgen(n.Left, res)
var hi1 gc.Node
var lo1 gc.Node
split64(res, &lo1, &hi1)
gins(x86.ANEGL, nil, &lo1)
gins(x86.AADCL, ncon(0), &hi1)
gins(x86.ANEGL, nil, &hi1)
splitclean()
return
case gc.OCOM:
gc.Cgen(n.Left, res)
var lo1 gc.Node
var hi1 gc.Node
split64(res, &lo1, &hi1)
gins(x86.ANOTL, nil, &lo1)
gins(x86.ANOTL, nil, &hi1)
splitclean()
return
// binary operators.
// common setup below.
case gc.OADD,
gc.OSUB,
gc.OMUL,
gc.OLROT,
gc.OLSH,
gc.ORSH,
gc.OAND,
gc.OOR,
gc.OXOR:
break
}
l := n.Left
r := n.Right
if !l.Addable {
var t1 gc.Node
gc.Tempname(&t1, l.Type)
gc.Cgen(l, &t1)
l = &t1
}
if r != nil && !r.Addable {
var t2 gc.Node
gc.Tempname(&t2, r.Type)
gc.Cgen(r, &t2)
r = &t2
}
var ax gc.Node
gc.Nodreg(&ax, gc.Types[gc.TINT32], x86.REG_AX)
var cx gc.Node
gc.Nodreg(&cx, gc.Types[gc.TINT32], x86.REG_CX)
var dx gc.Node
gc.Nodreg(&dx, gc.Types[gc.TINT32], x86.REG_DX)
// Setup for binary operation.
var hi1 gc.Node
var lo1 gc.Node
split64(l, &lo1, &hi1)
var lo2 gc.Node
var hi2 gc.Node
if gc.Is64(r.Type) {
split64(r, &lo2, &hi2)
}
// Do op. Leave result in DX:AX.
switch n.Op {
// TODO: Constants
case gc.OADD:
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(x86.AADDL, &lo2, &ax)
gins(x86.AADCL, &hi2, &dx)
// TODO: Constants.
case gc.OSUB:
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(x86.ASUBL, &lo2, &ax)
gins(x86.ASBBL, &hi2, &dx)
case gc.OMUL:
// let's call the next three EX, FX and GX
var ex, fx, gx gc.Node
gc.Regalloc(&ex, gc.Types[gc.TPTR32], nil)
gc.Regalloc(&fx, gc.Types[gc.TPTR32], nil)
gc.Regalloc(&gx, gc.Types[gc.TPTR32], nil)
// load args into DX:AX and EX:GX.
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(x86.AMOVL, &lo2, &gx)
gins(x86.AMOVL, &hi2, &ex)
// if DX and EX are zero, use 32 x 32 -> 64 unsigned multiply.
gins(x86.AMOVL, &dx, &fx)
gins(x86.AORL, &ex, &fx)
p1 := gc.Gbranch(x86.AJNE, nil, 0)
gins(x86.AMULL, &gx, nil) // implicit &ax
p2 := gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// full 64x64 -> 64, from 32x32 -> 64.
gins(x86.AIMULL, &gx, &dx)
gins(x86.AMOVL, &ax, &fx)
gins(x86.AIMULL, &ex, &fx)
gins(x86.AADDL, &dx, &fx)
gins(x86.AMOVL, &gx, &dx)
gins(x86.AMULL, &dx, nil) // implicit &ax
gins(x86.AADDL, &fx, &dx)
gc.Patch(p2, gc.Pc)
gc.Regfree(&ex)
gc.Regfree(&fx)
gc.Regfree(&gx)
// We only rotate by a constant c in [0,64).
// if c >= 32:
// lo, hi = hi, lo
// c -= 32
// if c == 0:
// no-op
// else:
// t = hi
// shld hi:lo, c
// shld lo:t, c
case gc.OLROT:
v := uint64(r.Int())
if v >= 32 {
// reverse during load to do the first 32 bits of rotate
v -= 32
gins(x86.AMOVL, &lo1, &dx)
gins(x86.AMOVL, &hi1, &ax)
} else {
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
}
if v == 0 {
} else // done
{
gins(x86.AMOVL, &dx, &cx)
p1 := gins(x86.ASHLL, ncon(uint32(v)), &dx)
p1.From.Index = x86.REG_AX // double-width shift
p1.From.Scale = 0
p1 = gins(x86.ASHLL, ncon(uint32(v)), &ax)
p1.From.Index = x86.REG_CX // double-width shift
p1.From.Scale = 0
}
case gc.OLSH:
if r.Op == gc.OLITERAL {
v := uint64(r.Int())
if v >= 64 {
if gc.Is64(r.Type) {
splitclean()
}
splitclean()
split64(res, &lo2, &hi2)
gins(x86.AMOVL, ncon(0), &lo2)
gins(x86.AMOVL, ncon(0), &hi2)
splitclean()
return
}
if v >= 32 {
if gc.Is64(r.Type) {
splitclean()
}
split64(res, &lo2, &hi2)
gmove(&lo1, &hi2)
if v > 32 {
gins(x86.ASHLL, ncon(uint32(v-32)), &hi2)
}
gins(x86.AMOVL, ncon(0), &lo2)
splitclean()
splitclean()
return
}
// general shift
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
p1 := gins(x86.ASHLL, ncon(uint32(v)), &dx)
p1.From.Index = x86.REG_AX // double-width shift
p1.From.Scale = 0
gins(x86.ASHLL, ncon(uint32(v)), &ax)
break
}
// load value into DX:AX.
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
// load shift value into register.
// if high bits are set, zero value.
var p1 *obj.Prog
if gc.Is64(r.Type) {
gins(x86.ACMPL, &hi2, ncon(0))
p1 = gc.Gbranch(x86.AJNE, nil, +1)
gins(x86.AMOVL, &lo2, &cx)
} else {
cx.Type = gc.Types[gc.TUINT32]
gmove(r, &cx)
}
// if shift count is >=64, zero value
gins(x86.ACMPL, &cx, ncon(64))
p2 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
if p1 != nil {
gc.Patch(p1, gc.Pc)
}
gins(x86.AXORL, &dx, &dx)
gins(x86.AXORL, &ax, &ax)
gc.Patch(p2, gc.Pc)
// if shift count is >= 32, zero low.
gins(x86.ACMPL, &cx, ncon(32))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
gins(x86.AMOVL, &ax, &dx)
gins(x86.ASHLL, &cx, &dx) // SHLL only uses bottom 5 bits of count
gins(x86.AXORL, &ax, &ax)
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// general shift
p1 = gins(x86.ASHLL, &cx, &dx)
p1.From.Index = x86.REG_AX // double-width shift
p1.From.Scale = 0
gins(x86.ASHLL, &cx, &ax)
gc.Patch(p2, gc.Pc)
case gc.ORSH:
if r.Op == gc.OLITERAL {
v := uint64(r.Int())
if v >= 64 {
if gc.Is64(r.Type) {
splitclean()
}
splitclean()
split64(res, &lo2, &hi2)
if hi1.Type.Etype == gc.TINT32 {
gmove(&hi1, &lo2)
gins(x86.ASARL, ncon(31), &lo2)
gmove(&hi1, &hi2)
gins(x86.ASARL, ncon(31), &hi2)
} else {
gins(x86.AMOVL, ncon(0), &lo2)
gins(x86.AMOVL, ncon(0), &hi2)
}
splitclean()
return
}
if v >= 32 {
if gc.Is64(r.Type) {
splitclean()
}
split64(res, &lo2, &hi2)
gmove(&hi1, &lo2)
if v > 32 {
gins(optoas(gc.ORSH, hi1.Type), ncon(uint32(v-32)), &lo2)
}
if hi1.Type.Etype == gc.TINT32 {
gmove(&hi1, &hi2)
gins(x86.ASARL, ncon(31), &hi2)
} else {
gins(x86.AMOVL, ncon(0), &hi2)
}
splitclean()
splitclean()
return
}
// general shift
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
p1 := gins(x86.ASHRL, ncon(uint32(v)), &ax)
p1.From.Index = x86.REG_DX // double-width shift
p1.From.Scale = 0
gins(optoas(gc.ORSH, hi1.Type), ncon(uint32(v)), &dx)
break
}
// load value into DX:AX.
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
// load shift value into register.
// if high bits are set, zero value.
var p1 *obj.Prog
if gc.Is64(r.Type) {
gins(x86.ACMPL, &hi2, ncon(0))
p1 = gc.Gbranch(x86.AJNE, nil, +1)
gins(x86.AMOVL, &lo2, &cx)
} else {
cx.Type = gc.Types[gc.TUINT32]
gmove(r, &cx)
}
// if shift count is >=64, zero or sign-extend value
gins(x86.ACMPL, &cx, ncon(64))
p2 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
if p1 != nil {
gc.Patch(p1, gc.Pc)
}
if hi1.Type.Etype == gc.TINT32 {
gins(x86.ASARL, ncon(31), &dx)
gins(x86.AMOVL, &dx, &ax)
} else {
gins(x86.AXORL, &dx, &dx)
gins(x86.AXORL, &ax, &ax)
}
gc.Patch(p2, gc.Pc)
// if shift count is >= 32, sign-extend hi.
gins(x86.ACMPL, &cx, ncon(32))
p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
gins(x86.AMOVL, &dx, &ax)
if hi1.Type.Etype == gc.TINT32 {
gins(x86.ASARL, &cx, &ax) // SARL only uses bottom 5 bits of count
gins(x86.ASARL, ncon(31), &dx)
} else {
gins(x86.ASHRL, &cx, &ax)
gins(x86.AXORL, &dx, &dx)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
// general shift
p1 = gins(x86.ASHRL, &cx, &ax)
p1.From.Index = x86.REG_DX // double-width shift
p1.From.Scale = 0
gins(optoas(gc.ORSH, hi1.Type), &cx, &dx)
gc.Patch(p2, gc.Pc)
// make constant the right side (it usually is anyway).
case gc.OXOR,
gc.OAND,
gc.OOR:
if lo1.Op == gc.OLITERAL {
nswap(&lo1, &lo2)
nswap(&hi1, &hi2)
}
if lo2.Op == gc.OLITERAL {
// special cases for constants.
lv := uint32(lo2.Int())
hv := uint32(hi2.Int())
splitclean() // right side
split64(res, &lo2, &hi2)
switch n.Op {
case gc.OXOR:
gmove(&lo1, &lo2)
gmove(&hi1, &hi2)
switch lv {
case 0:
break
case 0xffffffff:
gins(x86.ANOTL, nil, &lo2)
default:
gins(x86.AXORL, ncon(lv), &lo2)
}
switch hv {
case 0:
break
case 0xffffffff:
gins(x86.ANOTL, nil, &hi2)
default:
gins(x86.AXORL, ncon(hv), &hi2)
}
case gc.OAND:
switch lv {
case 0:
gins(x86.AMOVL, ncon(0), &lo2)
default:
gmove(&lo1, &lo2)
if lv != 0xffffffff {
gins(x86.AANDL, ncon(lv), &lo2)
}
}
switch hv {
case 0:
gins(x86.AMOVL, ncon(0), &hi2)
default:
gmove(&hi1, &hi2)
if hv != 0xffffffff {
gins(x86.AANDL, ncon(hv), &hi2)
}
}
case gc.OOR:
switch lv {
case 0:
gmove(&lo1, &lo2)
case 0xffffffff:
gins(x86.AMOVL, ncon(0xffffffff), &lo2)
default:
gmove(&lo1, &lo2)
gins(x86.AORL, ncon(lv), &lo2)
}
switch hv {
case 0:
gmove(&hi1, &hi2)
case 0xffffffff:
gins(x86.AMOVL, ncon(0xffffffff), &hi2)
default:
gmove(&hi1, &hi2)
gins(x86.AORL, ncon(hv), &hi2)
}
}
splitclean()
splitclean()
return
}
gins(x86.AMOVL, &lo1, &ax)
gins(x86.AMOVL, &hi1, &dx)
gins(optoas(n.Op, lo1.Type), &lo2, &ax)
gins(optoas(n.Op, lo1.Type), &hi2, &dx)
}
if gc.Is64(r.Type) {
splitclean()
}
splitclean()
split64(res, &lo1, &hi1)
gins(x86.AMOVL, &ax, &lo1)
gins(x86.AMOVL, &dx, &hi1)
splitclean()
}
/*
* generate shift according to op, one of:
* res = nl << nr
* res = nl >> nr
*/
func cgen_shift(op gc.Op, bounded bool, nl *gc.Node, nr *gc.Node, res *gc.Node) {
a := int(optoas(op, nl.Type))
if nr.Op == gc.OLITERAL {
var n1 gc.Node
gc.Regalloc(&n1, nl.Type, res)
gc.Cgen(nl, &n1)
sc := uint64(nr.Int())
if sc >= uint64(nl.Type.Width*8) {
// large shift gets 2 shifts by width-1
var n3 gc.Node
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n1)
gins(a, &n3, &n1)
} else {
gins(a, nr, &n1)
}
gmove(&n1, res)
gc.Regfree(&n1)
return
}
if nl.Ullman >= gc.UINF {
var n4 gc.Node
gc.Tempname(&n4, nl.Type)
gc.Cgen(nl, &n4)
nl = &n4
}
if nr.Ullman >= gc.UINF {
var n5 gc.Node
gc.Tempname(&n5, nr.Type)
gc.Cgen(nr, &n5)
nr = &n5
}
// Allow either uint32 or uint64 as shift type,
// to avoid unnecessary conversion from uint32 to uint64
// just to do the comparison.
tcount := gc.Types[gc.Simtype[nr.Type.Etype]]
if tcount.Etype < gc.TUINT32 {
tcount = gc.Types[gc.TUINT32]
}
var n1 gc.Node
gc.Regalloc(&n1, nr.Type, nil) // to hold the shift type in CX
var n3 gc.Node
gc.Regalloc(&n3, tcount, &n1) // to clear high bits of CX
var n2 gc.Node
gc.Regalloc(&n2, nl.Type, res)
if nl.Ullman >= nr.Ullman {
gc.Cgen(nl, &n2)
gc.Cgen(nr, &n1)
gmove(&n1, &n3)
} else {
gc.Cgen(nr, &n1)
gmove(&n1, &n3)
gc.Cgen(nl, &n2)
}
gc.Regfree(&n3)
// test and fix up large shifts
if !bounded {
gc.Nodconst(&n3, tcount, nl.Type.Width*8)
gcmp(optoas(gc.OCMP, tcount), &n1, &n3)
p1 := (*obj.Prog)(gc.Gbranch(optoas(gc.OLT, tcount), nil, +1))
if op == gc.ORSH && gc.Issigned[nl.Type.Etype] {
gc.Nodconst(&n3, gc.Types[gc.TUINT32], nl.Type.Width*8-1)
gins(a, &n3, &n2)
} else {
gc.Nodconst(&n3, nl.Type, 0)
gmove(&n3, &n2)
}
gc.Patch(p1, gc.Pc)
}
gins(a, &n1, &n2)
gmove(&n2, res)
gc.Regfree(&n1)
gc.Regfree(&n2)
}
/*
* generate division.
* caller must set:
* ax = allocated AX register
* dx = allocated DX register
* generates one of:
* res = nl / nr
* res = nl % nr
* according to op.
*/
func dodiv(op gc.Op, nl *gc.Node, nr *gc.Node, res *gc.Node, ax *gc.Node, dx *gc.Node) {
// Have to be careful about handling
// most negative int divided by -1 correctly.
// The hardware will trap.
// Also the byte divide instruction needs AH,
// which we otherwise don't have to deal with.
// Easiest way to avoid for int8, int16: use int32.
// For int32 and int64, use explicit test.
// Could use int64 hw for int32.
t := nl.Type
t0 := t
check := false
if gc.Issigned[t.Etype] {
check = true
if gc.Isconst(nl, gc.CTINT) && nl.Int() != -1<<uint64(t.Width*8-1) {
check = false
} else if gc.Isconst(nr, gc.CTINT) && nr.Int() != -1 {
check = false
}
}
if t.Width < 4 {
if gc.Issigned[t.Etype] {
t = gc.Types[gc.TINT32]
} else {
t = gc.Types[gc.TUINT32]
}
check = false
}
var t1 gc.Node
gc.Tempname(&t1, t)
var t2 gc.Node
gc.Tempname(&t2, t)
if t0 != t {
var t3 gc.Node
gc.Tempname(&t3, t0)
var t4 gc.Node
gc.Tempname(&t4, t0)
gc.Cgen(nl, &t3)
gc.Cgen(nr, &t4)
// Convert.
gmove(&t3, &t1)
gmove(&t4, &t2)
} else {
gc.Cgen(nl, &t1)
gc.Cgen(nr, &t2)
}
var n1 gc.Node
if !gc.Samereg(ax, res) && !gc.Samereg(dx, res) {
gc.Regalloc(&n1, t, res)
} else {
gc.Regalloc(&n1, t, nil)
}
gmove(&t2, &n1)
gmove(&t1, ax)
var p2 *obj.Prog
var n4 gc.Node
if gc.Nacl {
// Native Client does not relay the divide-by-zero trap
// to the executing program, so we must insert a check
// for ourselves.
gc.Nodconst(&n4, t, 0)
gins(optoas(gc.OCMP, t), &n1, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if panicdiv == nil {
panicdiv = gc.Sysfunc("panicdivide")
}
gc.Ginscall(panicdiv, -1)
gc.Patch(p1, gc.Pc)
}
if check {
gc.Nodconst(&n4, t, -1)
gins(optoas(gc.OCMP, t), &n1, &n4)
p1 := gc.Gbranch(optoas(gc.ONE, t), nil, +1)
if op == gc.ODIV {
// a / (-1) is -a.
gins(optoas(gc.OMINUS, t), nil, ax)
gmove(ax, res)
} else {
// a % (-1) is 0.
gc.Nodconst(&n4, t, 0)
gmove(&n4, res)
}
p2 = gc.Gbranch(obj.AJMP, nil, 0)
gc.Patch(p1, gc.Pc)
}
if !gc.Issigned[t.Etype] {
var nz gc.Node
gc.Nodconst(&nz, t, 0)
gmove(&nz, dx)
} else {
gins(optoas(gc.OEXTEND, t), nil, nil)
}
gins(optoas(op, t), &n1, nil)
gc.Regfree(&n1)
if op == gc.ODIV {
gmove(ax, res)
} else {
gmove(dx, res)
}
if check {
gc.Patch(p2, gc.Pc)
}
}