func conprop(r0 *gc.Flow) {
var p *obj.Prog
var t int
p0 := (*obj.Prog)(r0.Prog)
v0 := (*obj.Addr)(&p0.To)
r := (*gc.Flow)(r0)
loop:
r = gc.Uniqs(r)
if r == nil || r == r0 {
return
}
if gc.Uniqp(r) == nil {
return
}
p = r.Prog
t = copyu(p, v0, nil)
switch t {
case 0, // miss
1: // use
goto loop
case 2, // rar
4: // use and set
break
case 3: // set
if p.As == p0.As {
if p.From.Type == p0.From.Type {
if p.From.Reg == p0.From.Reg {
if p.From.Node == p0.From.Node {
if p.From.Offset == p0.From.Offset {
if p.From.Scale == p0.From.Scale {
if p.From.Type == obj.TYPE_FCONST && p.From.Val.(float64) == p0.From.Val.(float64) {
if p.From.Index == p0.From.Index {
excise(r)
goto loop
}
}
}
}
}
}
}
}
}
}
func rnops(r *gc.Flow) *gc.Flow {
if r != nil {
var p *obj.Prog
var r1 *gc.Flow
for {
p = r.Prog
if p.As != obj.ANOP || p.From.Type != obj.TYPE_NONE || p.To.Type != obj.TYPE_NONE {
break
}
r1 = gc.Uniqs(r)
if r1 == nil {
break
}
r = r1
}
}
return r
}
func peep(firstp *obj.Prog) {
g := (*gc.Graph)(gc.Flowstart(firstp, nil))
if g == nil {
return
}
gactive = 0
// byte, word arithmetic elimination.
elimshortmov(g)
// constant propagation
// find MOV $con,R followed by
// another MOV $con,R without
// setting R in the interim
var p *obj.Prog
for r := (*gc.Flow)(g.Start); r != nil; r = r.Link {
p = r.Prog
switch p.As {
case x86.ALEAL,
x86.ALEAQ:
if regtyp(&p.To) {
if p.From.Sym != nil {
if p.From.Index == x86.REG_NONE {
conprop(r)
}
}
}
case x86.AMOVB,
x86.AMOVW,
x86.AMOVL,
x86.AMOVQ,
x86.AMOVSS,
x86.AMOVSD:
if regtyp(&p.To) {
if p.From.Type == obj.TYPE_CONST || p.From.Type == obj.TYPE_FCONST {
conprop(r)
}
}
}
}
var r *gc.Flow
var r1 *gc.Flow
var p1 *obj.Prog
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
switch p.As {
case x86.AMOVL,
x86.AMOVQ,
x86.AMOVSS,
x86.AMOVSD:
if regtyp(&p.To) {
if regtyp(&p.From) {
if copyprop(g, r) {
excise(r)
t++
} else if subprop(r) && copyprop(g, r) {
excise(r)
t++
}
}
}
case x86.AMOVBLZX,
x86.AMOVWLZX,
x86.AMOVBLSX,
x86.AMOVWLSX:
if regtyp(&p.To) {
r1 = rnops(gc.Uniqs(r))
if r1 != nil {
p1 = r1.Prog
if p.As == p1.As && p.To.Type == p1.From.Type && p.To.Reg == p1.From.Reg {
p1.As = x86.AMOVL
t++
}
}
}
case x86.AMOVBQSX,
x86.AMOVBQZX,
x86.AMOVWQSX,
x86.AMOVWQZX,
x86.AMOVLQSX,
x86.AMOVLQZX,
x86.AMOVQL:
if regtyp(&p.To) {
r1 = rnops(gc.Uniqs(r))
if r1 != nil {
p1 = r1.Prog
if p.As == p1.As && p.To.Type == p1.From.Type && p.To.Reg == p1.From.Reg {
p1.As = x86.AMOVQ
t++
}
}
}
case x86.AADDL,
x86.AADDQ,
x86.AADDW:
if p.From.Type != obj.TYPE_CONST || needc(p.Link) {
break
}
if p.From.Offset == -1 {
if p.As == x86.AADDQ {
p.As = x86.ADECQ
} else if p.As == x86.AADDL {
p.As = x86.ADECL
} else {
p.As = x86.ADECW
}
p.From = obj.Addr{}
break
}
if p.From.Offset == 1 {
if p.As == x86.AADDQ {
p.As = x86.AINCQ
} else if p.As == x86.AADDL {
p.As = x86.AINCL
} else {
p.As = x86.AINCW
}
p.From = obj.Addr{}
break
}
case x86.ASUBL,
x86.ASUBQ,
x86.ASUBW:
if p.From.Type != obj.TYPE_CONST || needc(p.Link) {
break
}
if p.From.Offset == -1 {
if p.As == x86.ASUBQ {
p.As = x86.AINCQ
} else if p.As == x86.ASUBL {
p.As = x86.AINCL
} else {
p.As = x86.AINCW
}
p.From = obj.Addr{}
break
}
if p.From.Offset == 1 {
if p.As == x86.ASUBQ {
p.As = x86.ADECQ
} else if p.As == x86.ASUBL {
p.As = x86.ADECL
} else {
p.As = x86.ADECW
}
p.From = obj.Addr{}
break
}
}
}
if t != 0 {
goto loop1
}
// MOVLQZX removal.
// The MOVLQZX exists to avoid being confused for a
// MOVL that is just copying 32-bit data around during
// copyprop. Now that copyprop is done, remov MOVLQZX R1, R2
// if it is dominated by an earlier ADDL/MOVL/etc into R1 that
// will have already cleared the high bits.
//
// MOVSD removal.
// We never use packed registers, so a MOVSD between registers
// can be replaced by MOVAPD, which moves the pair of float64s
// instead of just the lower one. We only use the lower one, but
// the processor can do better if we do moves using both.
for r := (*gc.Flow)(g.Start); r != nil; r = r.Link {
p = r.Prog
if p.As == x86.AMOVLQZX {
if regtyp(&p.From) {
if p.From.Type == p.To.Type && p.From.Reg == p.To.Reg {
if prevl(r, int(p.From.Reg)) {
excise(r)
}
}
}
}
if p.As == x86.AMOVSD {
if regtyp(&p.From) {
if regtyp(&p.To) {
p.As = x86.AMOVAPD
}
}
}
}
// 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 x86.AMOVB,
x86.AMOVW,
x86.AMOVL,
x86.AMOVQ,
x86.AMOVLQZX:
if regtyp(&p.To) && !regconsttyp(&p.From) {
pushback(r)
}
}
}
gc.Flowend(g)
}
/*
* the idea is to substitute
* one register for another
* from one MOV to another
* MOV a, R0
* ADD b, R0 / no use of R1
* MOV R0, R1
* would be converted to
* MOV a, R1
* ADD b, R1
* MOV R1, R0
* hopefully, then the former or latter MOV
* will be eliminated by copy propagation.
*/
func subprop(r0 *gc.Flow) bool {
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("subprop %v\n", r0.Prog)
}
p := (*obj.Prog)(r0.Prog)
v1 := (*obj.Addr)(&p.From)
if !regtyp(v1) {
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("\tnot regtype %v; return 0\n", gc.Ctxt.Dconv(v1))
}
return false
}
v2 := (*obj.Addr)(&p.To)
if !regtyp(v2) {
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("\tnot regtype %v; return 0\n", gc.Ctxt.Dconv(v2))
}
return false
}
for r := gc.Uniqp(r0); r != nil; r = gc.Uniqp(r) {
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("\t? %v\n", r.Prog)
}
if gc.Uniqs(r) == nil {
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("\tno unique successor\n")
}
break
}
p = r.Prog
if p.As == obj.AVARDEF || p.As == obj.AVARKILL {
continue
}
if p.Info.Flags&gc.Call != 0 {
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("\tfound %v; return 0\n", p)
}
return false
}
if p.Info.Reguse|p.Info.Regset != 0 {
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("\tfound %v; return 0\n", p)
}
return false
}
if (p.Info.Flags&gc.Move != 0) && (p.Info.Flags&(gc.SizeL|gc.SizeQ|gc.SizeF|gc.SizeD) != 0) && p.To.Type == v1.Type && p.To.Reg == v1.Reg {
copysub(&p.To, v1, v2, 1)
if gc.Debug['P'] != 0 {
fmt.Printf("gotit: %v->%v\n%v", gc.Ctxt.Dconv(v1), gc.Ctxt.Dconv(v2), r.Prog)
if p.From.Type == v2.Type && p.From.Reg == v2.Reg {
fmt.Printf(" excise")
}
fmt.Printf("\n")
}
for r = gc.Uniqs(r); r != r0; r = gc.Uniqs(r) {
p = r.Prog
copysub(&p.From, v1, v2, 1)
copysub(&p.To, v1, v2, 1)
if gc.Debug['P'] != 0 {
fmt.Printf("%v\n", r.Prog)
}
}
t := int(int(v1.Reg))
v1.Reg = v2.Reg
v2.Reg = int16(t)
if gc.Debug['P'] != 0 {
fmt.Printf("%v last\n", r.Prog)
}
return true
}
if copyau(&p.From, v2) || copyau(&p.To, v2) {
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("\tcopyau %v failed\n", gc.Ctxt.Dconv(v2))
}
break
}
if copysub(&p.From, v1, v2, 0) != 0 || copysub(&p.To, v1, v2, 0) != 0 {
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("\tcopysub failed\n")
}
break
}
}
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("\tran off end; return 0\n")
}
return false
}
func pushback(r0 *gc.Flow) {
var r *gc.Flow
var p *obj.Prog
var b *gc.Flow
p0 := (*obj.Prog)(r0.Prog)
for r = gc.Uniqp(r0); r != nil && gc.Uniqs(r) != nil; r = gc.Uniqp(r) {
p = r.Prog
if p.As != obj.ANOP {
if !regconsttyp(&p.From) || !regtyp(&p.To) {
break
}
if copyu(p, &p0.To, nil) != 0 || copyu(p0, &p.To, nil) != 0 {
break
}
}
if p.As == obj.ACALL {
break
}
b = r
}
if b == nil {
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("no pushback: %v\n", r0.Prog)
if r != nil {
fmt.Printf("\t%v [%v]\n", r.Prog, gc.Uniqs(r) != nil)
}
}
return
}
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("pushback\n")
for r := (*gc.Flow)(b); ; r = r.Link {
fmt.Printf("\t%v\n", r.Prog)
if r == r0 {
break
}
}
}
t := obj.Prog(*r0.Prog)
for r = gc.Uniqp(r0); ; r = gc.Uniqp(r) {
p0 = r.Link.Prog
p = r.Prog
p0.As = p.As
p0.Lineno = p.Lineno
p0.From = p.From
p0.To = p.To
if r == b {
break
}
}
p0 = r.Prog
p0.As = t.As
p0.Lineno = t.Lineno
p0.From = t.From
p0.To = t.To
if gc.Debug['P'] != 0 && gc.Debug['v'] != 0 {
fmt.Printf("\tafter\n")
for r := (*gc.Flow)(b); ; r = r.Link {
fmt.Printf("\t%v\n", r.Prog)
if r == r0 {
break
}
}
}
}
