func outgcode(a int, g1 *obj.Addr, reg int, g2, g3 *obj.Addr) {
var p *obj.Prog
var pl *obj.Plist
if asm.Pass == 1 {
goto out
}
p = asm.Ctxt.NewProg()
p.As = int16(a)
p.Lineno = stmtline
if nosched != 0 {
p.Mark |= ppc64.NOSCHED
}
p.From = *g1
p.Reg = int16(reg)
p.From3 = *g2
p.To = *g3
p.Pc = int64(asm.PC)
if lastpc == nil {
pl = obj.Linknewplist(asm.Ctxt)
pl.Firstpc = p
} else {
lastpc.Link = p
}
lastpc = p
out:
if a != obj.AGLOBL && a != obj.ADATA {
asm.PC++
}
}
func outcode(a int, g2 *Addr2) {
var p *obj.Prog
var pl *obj.Plist
if asm.Pass == 1 {
goto out
}
p = new(obj.Prog)
*p = obj.Prog{}
p.Ctxt = asm.Ctxt
p.As = int16(a)
p.Lineno = stmtline
p.From = g2.from
if g2.from3.Type != 0 {
p.From3 = new(obj.Addr)
*p.From3 = g2.from3
}
p.To = g2.to
p.Pc = int64(asm.PC)
if lastpc == nil {
pl = obj.Linknewplist(asm.Ctxt)
pl.Firstpc = p
} else {
lastpc.Link = p
}
lastpc = p
out:
if a != obj.AGLOBL && a != obj.ADATA {
asm.PC++
}
}
func compile(fn *Node) {
if Newproc == nil {
Newproc = Sysfunc("newproc")
Deferproc = Sysfunc("deferproc")
Deferreturn = Sysfunc("deferreturn")
Panicindex = Sysfunc("panicindex")
panicslice = Sysfunc("panicslice")
throwreturn = Sysfunc("throwreturn")
}
lno := setlineno(fn)
Curfn = fn
dowidth(Curfn.Type)
var oldstksize int64
var nod1 Node
var ptxt *obj.Prog
var pl *obj.Plist
var p *obj.Prog
var n *Node
var nam *Node
var gcargs *Sym
var gclocals *Sym
if fn.Nbody == nil {
if pure_go != 0 || strings.HasPrefix(fn.Func.Nname.Sym.Name, "init.") {
Yyerror("missing function body for %q", fn.Func.Nname.Sym.Name)
goto ret
}
if Debug['A'] != 0 {
goto ret
}
emitptrargsmap()
goto ret
}
saveerrors()
// set up domain for labels
clearlabels()
if Curfn.Type.Outnamed {
// add clearing of the output parameters
var save Iter
t := Structfirst(&save, Getoutarg(Curfn.Type))
for t != nil {
if t.Nname != nil {
n = Nod(OAS, t.Nname, nil)
typecheck(&n, Etop)
Curfn.Nbody = concat(list1(n), Curfn.Nbody)
}
t = structnext(&save)
}
}
order(Curfn)
if nerrors != 0 {
goto ret
}
hasdefer = false
walk(Curfn)
if nerrors != 0 {
goto ret
}
if instrumenting {
instrument(Curfn)
}
if nerrors != 0 {
goto ret
}
continpc = nil
breakpc = nil
pl = newplist()
pl.Name = Linksym(Curfn.Func.Nname.Sym)
setlineno(Curfn)
Nodconst(&nod1, Types[TINT32], 0)
nam = Curfn.Func.Nname
if isblank(nam) {
nam = nil
}
ptxt = Thearch.Gins(obj.ATEXT, nam, &nod1)
Afunclit(&ptxt.From, Curfn.Func.Nname)
ptxt.From3 = new(obj.Addr)
if fn.Func.Dupok {
ptxt.From3.Offset |= obj.DUPOK
}
if fn.Func.Wrapper {
ptxt.From3.Offset |= obj.WRAPPER
}
if fn.Func.Needctxt {
ptxt.From3.Offset |= obj.NEEDCTXT
}
if fn.Func.Nosplit {
ptxt.From3.Offset |= obj.NOSPLIT
}
if fn.Func.Systemstack {
ptxt.From.Sym.Cfunc = 1
}
// Clumsy but important.
// See test/recover.go for test cases and src/reflect/value.go
// for the actual functions being considered.
if myimportpath != "" && myimportpath == "reflect" {
if Curfn.Func.Nname.Sym.Name == "callReflect" || Curfn.Func.Nname.Sym.Name == "callMethod" {
ptxt.From3.Offset |= obj.WRAPPER
}
}
ginit()
gcargs = makefuncdatasym("gcargs·%d", obj.FUNCDATA_ArgsPointerMaps)
gclocals = makefuncdatasym("gclocals·%d", obj.FUNCDATA_LocalsPointerMaps)
for _, t := range Curfn.Func.Fieldtrack {
gtrack(tracksym(t))
}
for l := fn.Func.Dcl; l != nil; l = l.Next {
n = l.N
if n.Op != ONAME { // might be OTYPE or OLITERAL
continue
}
switch n.Class {
case PAUTO, PPARAM, PPARAMOUT:
Nodconst(&nod1, Types[TUINTPTR], l.N.Type.Width)
p = Thearch.Gins(obj.ATYPE, l.N, &nod1)
p.From.Gotype = Linksym(ngotype(l.N))
}
}
Genlist(Curfn.Func.Enter)
Genlist(Curfn.Nbody)
gclean()
checklabels()
if nerrors != 0 {
goto ret
}
if Curfn.Func.Endlineno != 0 {
lineno = Curfn.Func.Endlineno
}
if Curfn.Type.Outtuple != 0 {
Ginscall(throwreturn, 0)
}
ginit()
// TODO: Determine when the final cgen_ret can be omitted. Perhaps always?
cgen_ret(nil)
if hasdefer {
// deferreturn pretends to have one uintptr argument.
// Reserve space for it so stack scanner is happy.
if Maxarg < int64(Widthptr) {
Maxarg = int64(Widthptr)
}
}
gclean()
if nerrors != 0 {
goto ret
}
Pc.As = obj.ARET // overwrite AEND
Pc.Lineno = lineno
fixjmp(ptxt)
if Debug['N'] == 0 || Debug['R'] != 0 || Debug['P'] != 0 {
regopt(ptxt)
nilopt(ptxt)
}
Thearch.Expandchecks(ptxt)
oldstksize = Stksize
allocauto(ptxt)
if false {
fmt.Printf("allocauto: %d to %d\n", oldstksize, int64(Stksize))
}
setlineno(Curfn)
if int64(Stksize)+Maxarg > 1<<31 {
Yyerror("stack frame too large (>2GB)")
goto ret
}
// Emit garbage collection symbols.
liveness(Curfn, ptxt, gcargs, gclocals)
gcsymdup(gcargs)
gcsymdup(gclocals)
Thearch.Defframe(ptxt)
if Debug['f'] != 0 {
frame(0)
}
// Remove leftover instrumentation from the instruction stream.
removevardef(ptxt)
ret:
lineno = lno
}
// zerorange clears the stack in the given range.
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
cnt := hi - lo
if cnt == 0 {
return p
}
// Adjust the frame to account for LR.
frame += gc.Ctxt.FixedFrameSize()
offset := frame + lo
reg := int16(s390x.REGSP)
// If the offset cannot fit in a 12-bit unsigned displacement then we
// need to create a copy of the stack pointer that we can adjust.
// We also need to do this if we are going to loop.
if offset < 0 || offset > 4096-clearLoopCutoff || cnt > clearLoopCutoff {
p = appendpp(p, s390x.AADD, obj.TYPE_CONST, 0, offset, obj.TYPE_REG, s390x.REGRT1, 0)
p.Reg = int16(s390x.REGSP)
reg = s390x.REGRT1
offset = 0
}
// Generate a loop of large clears.
if cnt > clearLoopCutoff {
n := cnt - (cnt % 256)
end := int16(s390x.REGRT2)
p = appendpp(p, s390x.AADD, obj.TYPE_CONST, 0, offset+n, obj.TYPE_REG, end, 0)
p.Reg = reg
p = appendpp(p, s390x.AXC, obj.TYPE_MEM, reg, offset, obj.TYPE_MEM, reg, offset)
p.From3 = new(obj.Addr)
p.From3.Type = obj.TYPE_CONST
p.From3.Offset = 256
pl := p
p = appendpp(p, s390x.AADD, obj.TYPE_CONST, 0, 256, obj.TYPE_REG, reg, 0)
p = appendpp(p, s390x.ACMP, obj.TYPE_REG, reg, 0, obj.TYPE_REG, end, 0)
p = appendpp(p, s390x.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
gc.Patch(p, pl)
cnt -= n
}
// Generate remaining clear instructions without a loop.
for cnt > 0 {
n := cnt
// Can clear at most 256 bytes per instruction.
if n > 256 {
n = 256
}
switch n {
// Handle very small clears with move instructions.
case 8, 4, 2, 1:
ins := s390x.AMOVB
switch n {
case 8:
ins = s390x.AMOVD
case 4:
ins = s390x.AMOVW
case 2:
ins = s390x.AMOVH
}
p = appendpp(p, ins, obj.TYPE_CONST, 0, 0, obj.TYPE_MEM, reg, offset)
// Handle clears that would require multiple move instructions with XC.
default:
p = appendpp(p, s390x.AXC, obj.TYPE_MEM, reg, offset, obj.TYPE_MEM, reg, offset)
p.From3 = new(obj.Addr)
p.From3.Type = obj.TYPE_CONST
p.From3.Offset = n
}
cnt -= n
offset += n
}
return p
}
func peep(firstp *obj.Prog) {
g := (*gc.Graph)(gc.Flowstart(firstp, nil))
if g == nil {
return
}
gactive = 0
var p *obj.Prog
if false {
// constant propagation
// find MOV $con,R followed by
// another MOV $con,R without
// setting R in the interim
for r := (*gc.Flow)(g.Start); r != nil; r = r.Link {
p = r.Prog
switch p.As {
case s390x.AMOVB,
s390x.AMOVW,
s390x.AMOVD:
if regtyp(&p.To) {
if p.From.Type == obj.TYPE_CONST || p.From.Type == obj.TYPE_FCONST {
conprop(r)
}
}
}
}
}
var r *gc.Flow
var t int
loop1:
// if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
// gc.Dumpit("loop1", g.Start, 0)
// }
t = 0
for r = g.Start; r != nil; r = r.Link {
p = r.Prog
// TODO(austin) Handle smaller moves. arm and amd64
// distinguish between moves that moves that *must*
// sign/zero extend and moves that don't care so they
// can eliminate moves that don't care without
// breaking moves that do care. This might let us
// simplify or remove the next peep loop, too.
if p.As == s390x.AMOVD || p.As == s390x.AFMOVD {
if regtyp(&p.To) {
// Try to eliminate reg->reg moves
if regtyp(&p.From) {
if p.From.Type == p.To.Type {
if copyprop(r) {
excise(r)
t++
} else if subprop(r) && copyprop(r) {
excise(r)
t++
}
}
}
// Convert uses to $0 to uses of R0 and
// propagate R0
if regzer(&p.From) != 0 {
if p.To.Type == obj.TYPE_REG {
p.From.Type = obj.TYPE_REG
p.From.Reg = s390x.REGZERO
if copyprop(r) {
excise(r)
t++
} else if subprop(r) && copyprop(r) {
excise(r)
t++
}
}
}
}
}
}
if t != 0 {
goto loop1
}
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
gc.Dumpit("pass7 copyprop", g.Start, 0)
}
/*
* look for MOVB x,R; MOVB R,R (for small MOVs not handled above)
*/
var p1 *obj.Prog
var r1 *gc.Flow
for r := (*gc.Flow)(g.Start); r != nil; r = r.Link {
p = r.Prog
switch p.As {
default:
continue
case s390x.AMOVH,
s390x.AMOVHZ,
s390x.AMOVB,
s390x.AMOVBZ,
s390x.AMOVW,
s390x.AMOVWZ:
if p.To.Type != obj.TYPE_REG {
continue
}
}
r1 = r.Link
if r1 == nil {
continue
}
p1 = r1.Prog
if p1.As != p.As {
continue
}
if p1.From.Type != obj.TYPE_REG || p1.From.Reg != p.To.Reg {
continue
}
if p1.To.Type != obj.TYPE_REG || p1.To.Reg != p.To.Reg {
continue
}
excise(r1)
}
if gc.Debug['P'] > 1 {
goto ret /* allow following code improvement to be suppressed */
}
if gc.Debug['p'] == 0 {
// load pipelining
// push any load from memory as early as possible
// to give it time to complete before use.
for r := (*gc.Flow)(g.Start); r != nil; r = r.Link {
p = r.Prog
switch p.As {
case s390x.AMOVB,
s390x.AMOVW,
s390x.AMOVD:
if regtyp(&p.To) && !regconsttyp(&p.From) {
pushback(r)
}
}
}
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
gc.Dumpit("pass8 push load as early as possible", g.Start, 0)
}
}
/*
* look for OP a, b, c; MOV c, d; -> OP a, b, d;
*/
for r := (*gc.Flow)(g.Start); r != nil; r = r.Link {
if (gc.Debugmergeopmv != -1) && (mergeopmv_cnt >= gc.Debugmergeopmv) {
break
}
p = r.Prog
switch p.As {
case s390x.AADD,
s390x.AADDC,
s390x.AADDME,
s390x.AADDE,
s390x.AADDZE,
s390x.AAND,
s390x.AANDN,
s390x.ADIVW,
s390x.ADIVWU,
s390x.ADIVD,
s390x.ADIVDU,
s390x.AMULLW,
s390x.AMULHD,
s390x.AMULHDU,
s390x.AMULLD,
s390x.ANAND,
s390x.ANOR,
s390x.AOR,
s390x.AORN,
s390x.AREM,
s390x.AREMU,
s390x.AREMD,
s390x.AREMDU,
s390x.ARLWMI,
s390x.ARLWNM,
s390x.ASLW,
s390x.ASRAW,
s390x.ASRW,
s390x.ASLD,
s390x.ASRAD,
s390x.ASRD,
s390x.ASUB,
s390x.ASUBC,
s390x.ASUBME,
s390x.ASUBE,
s390x.ASUBZE,
s390x.AXOR:
if p.To.Type != obj.TYPE_REG {
continue
}
if p.Reg == 0 { // Only for 3 ops instruction
continue
}
default:
continue
}
r1 := r.Link
for ; r1 != nil; r1 = r1.Link {
if r1.Prog.As != obj.ANOP {
break
}
}
if r1 == nil {
continue
}
p1 := r1.Prog
switch p1.As {
case s390x.AMOVD,
s390x.AMOVW, s390x.AMOVWZ,
s390x.AMOVH, s390x.AMOVHZ,
s390x.AMOVB, s390x.AMOVBZ:
if p1.To.Type != obj.TYPE_REG {
continue
}
default:
continue
}
if p1.From.Type != obj.TYPE_REG || p1.From.Reg != p.To.Reg {
continue
}
if trymergeopmv(r1) {
p.To = p1.To
excise(r1)
mergeopmv_cnt += 1
}
}
if gc.Debug['v'] != 0 {
gc.Dumpit("Merge operation and move", g.Start, 0)
}
/*
* look for CMP x, y; Branch -> Compare and branch
*/
if gc.Debugcnb == 0 {
goto ret
}
for r := (*gc.Flow)(g.Start); r != nil; r = r.Link {
if (gc.Debugcnb != -1) && (cnb_cnt >= gc.Debugcnb) {
break
}
p = r.Prog
r1 = gc.Uniqs(r)
if r1 == nil {
continue
}
p1 = r1.Prog
switch p.As {
case s390x.ACMP:
switch p1.As {
case s390x.ABCL, s390x.ABC:
continue
case s390x.ABEQ:
t = s390x.ACMPBEQ
case s390x.ABGE:
t = s390x.ACMPBGE
case s390x.ABGT:
t = s390x.ACMPBGT
case s390x.ABLE:
t = s390x.ACMPBLE
case s390x.ABLT:
t = s390x.ACMPBLT
case s390x.ABNE:
t = s390x.ACMPBNE
default:
continue
}
case s390x.ACMPU:
switch p1.As {
case s390x.ABCL, s390x.ABC:
continue
case s390x.ABEQ:
t = s390x.ACMPUBEQ
case s390x.ABGE:
t = s390x.ACMPUBGE
case s390x.ABGT:
t = s390x.ACMPUBGT
case s390x.ABLE:
t = s390x.ACMPUBLE
case s390x.ABLT:
t = s390x.ACMPUBLT
case s390x.ABNE:
t = s390x.ACMPUBNE
default:
continue
}
case s390x.ACMPW, s390x.ACMPWU:
continue
default:
continue
}
if gc.Debug['D'] != 0 {
fmt.Printf("cnb %v; %v -> ", p, p1)
}
if p1.To.Sym != nil {
continue
}
if p.To.Type == obj.TYPE_REG {
p1.As = int16(t)
p1.From = p.From
p1.Reg = p.To.Reg
p1.From3 = nil
} else if p.To.Type == obj.TYPE_CONST {
switch p.As {
case s390x.ACMP, s390x.ACMPW:
if (p.To.Offset < -(1 << 7)) || (p.To.Offset >= ((1 << 7) - 1)) {
continue
}
case s390x.ACMPU, s390x.ACMPWU:
if p.To.Offset >= (1 << 8) {
continue
}
default:
}
p1.As = int16(t)
p1.From = p.From
p1.Reg = 0
p1.From3 = new(obj.Addr)
*(p1.From3) = p.To
} else {
continue
}
if gc.Debug['D'] != 0 {
fmt.Printf("%v\n", p1)
}
cnb_cnt += 1
excise(r)
}
if gc.Debug['v'] != 0 {
gc.Dumpit("compare and branch", g.Start, 0)
}
ret:
gc.Flowend(g)
}