// 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 := 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")
}
// check is
// CMP arg, $0
// JNE 2(PC) (likely)
// MOV AX, 0
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 = int16(cmpptr)
p.To.Type = obj.TYPE_CONST
p.To.Offset = 0
p1.As = x86.AJNE
p1.From.Type = obj.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.
// if possible, since we know arg is 0, use 0(arg),
// which will be shorter to encode than plain 0.
p2.As = x86.AMOVL
p2.From.Type = obj.TYPE_REG
p2.From.Reg = x86.REG_AX
if regtyp(&p.From) {
p2.To.Type = obj.TYPE_MEM
p2.To.Reg = p.From.Reg
} else {
p2.To.Type = obj.TYPE_MEM
p2.To.Reg = x86.REG_NONE
}
p2.To.Offset = 0
}
}
func rewriteToPcrel(ctxt *obj.Link, p *obj.Prog) {
// RegTo2 is set on the instructions we insert here so they don't get
// processed twice.
if p.RegTo2 != 0 {
return
}
if p.As == obj.ATEXT || p.As == obj.AFUNCDATA || p.As == obj.ACALL || p.As == obj.ARET || p.As == obj.AJMP {
return
}
// Any Prog (aside from the above special cases) with an Addr with Name ==
// NAME_EXTERN, NAME_STATIC or NAME_GOTREF has a CALL __x86.get_pc_thunk.cx
// inserted before it.
isName := func(a *obj.Addr) bool {
if a.Sym == nil || (a.Type != obj.TYPE_MEM && a.Type != obj.TYPE_ADDR) || a.Reg != 0 {
return false
}
if a.Sym.Type == obj.STLSBSS {
return false
}
return a.Name == obj.NAME_EXTERN || a.Name == obj.NAME_STATIC || a.Name == obj.NAME_GOTREF
}
if isName(&p.From) && p.From.Type == obj.TYPE_ADDR {
// Handle things like "MOVL $sym, (SP)" or "PUSHL $sym" by rewriting
// to "MOVL $sym, CX; MOVL CX, (SP)" or "MOVL $sym, CX; PUSHL CX"
// respectively.
if p.To.Type != obj.TYPE_REG {
q := obj.Appendp(ctxt, p)
q.As = p.As
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_CX
q.To = p.To
p.As = AMOVL
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_CX
p.To.Sym = nil
p.To.Name = obj.NAME_NONE
}
}
if !isName(&p.From) && !isName(&p.To) && (p.From3 == nil || !isName(p.From3)) {
return
}
q := obj.Appendp(ctxt, p)
q.RegTo2 = 1
r := obj.Appendp(ctxt, q)
r.RegTo2 = 1
q.As = obj.ACALL
q.To.Sym = obj.Linklookup(ctxt, "__x86.get_pc_thunk.cx", 0)
q.To.Type = obj.TYPE_MEM
q.To.Name = obj.NAME_EXTERN
q.To.Sym.Local = true
r.As = p.As
r.Scond = p.Scond
r.From = p.From
r.From3 = p.From3
r.Reg = p.Reg
r.To = p.To
obj.Nopout(p)
}
func Prog(as int) *obj.Prog {
var p *obj.Prog
if as == obj.ADATA || as == obj.AGLOBL {
if ddumped != 0 {
Fatalf("already dumped data")
}
if dpc == nil {
dpc = Ctxt.NewProg()
dfirst = dpc
}
p = dpc
dpc = Ctxt.NewProg()
p.Link = dpc
} else {
p = Pc
Pc = Ctxt.NewProg()
Clearp(Pc)
p.Link = Pc
}
if lineno == 0 {
if Debug['K'] != 0 {
Warn("prog: line 0")
}
}
p.As = int16(as)
p.Lineno = lineno
return p
}
// ARMConditionCodes handles the special condition code situation for the ARM.
// It returns a boolean to indicate success; failure means cond was unrecognized.
func ARMConditionCodes(prog *obj.Prog, cond string) bool {
if cond == "" {
return true
}
bits, ok := ParseARMCondition(cond)
if !ok {
return false
}
/* hack to make B.NE etc. work: turn it into the corresponding conditional */
if prog.As == arm.AB {
prog.As = int16(bcode[(bits^arm.C_SCOND_XOR)&0xf])
bits = (bits &^ 0xf) | arm.C_SCOND_NONE
}
prog.Scond = bits
return true
}
// 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)
}
}
// 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
}
}
func peep(firstp *obj.Prog) {
g := 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 := g.Start; r != nil; r = r.Link {
p = r.Prog
switch p.As {
case x86.ALEAL:
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.AMOVSS,
x86.AMOVSD:
if regtyp(&p.To) {
if p.From.Type == obj.TYPE_CONST || p.From.Type == obj.TYPE_FCONST {
conprop(r)
}
}
}
}
var r1 *gc.Flow
var p1 *obj.Prog
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
switch p.As {
case x86.AMOVL,
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.AADDL,
x86.AADDW:
if p.From.Type != obj.TYPE_CONST || needc(p.Link) {
break
}
if p.From.Offset == -1 {
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.AADDL {
p.As = x86.AINCL
} else {
p.As = x86.AINCW
}
p.From = obj.Addr{}
break
}
case x86.ASUBL,
x86.ASUBW:
if p.From.Type != obj.TYPE_CONST || needc(p.Link) {
break
}
if p.From.Offset == -1 {
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.ASUBL {
p.As = x86.ADECL
} else {
p.As = x86.ADECW
}
p.From = obj.Addr{}
break
}
}
}
if t != 0 {
goto loop1
}
// 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 := g.Start; r != nil; r = r.Link {
p = r.Prog
if p.As == x86.AMOVSD {
if regtyp(&p.From) {
if regtyp(&p.To) {
p.As = x86.AMOVAPD
}
}
}
}
gc.Flowend(g)
}
// Rewrite p, if necessary, to access global data via the global offset table.
func rewriteToUseGot(ctxt *obj.Link, p *obj.Prog) {
var add, lea, mov, reg int16
if p.Mode == 64 {
add = AADDQ
lea = ALEAQ
mov = AMOVQ
reg = REG_R15
} else {
add = AADDL
lea = ALEAL
mov = AMOVL
reg = REG_CX
}
if p.As == obj.ADUFFCOPY || p.As == obj.ADUFFZERO {
// ADUFFxxx $offset
// becomes
// $MOV runtime.duffxxx@GOT, $reg
// $ADD $offset, $reg
// CALL $reg
var sym *obj.LSym
if p.As == obj.ADUFFZERO {
sym = obj.Linklookup(ctxt, "runtime.duffzero", 0)
} else {
sym = obj.Linklookup(ctxt, "runtime.duffcopy", 0)
}
offset := p.To.Offset
p.As = mov
p.From.Type = obj.TYPE_MEM
p.From.Name = obj.NAME_GOTREF
p.From.Sym = sym
p.To.Type = obj.TYPE_REG
p.To.Reg = reg
p.To.Offset = 0
p.To.Sym = nil
p1 := obj.Appendp(ctxt, p)
p1.As = add
p1.From.Type = obj.TYPE_CONST
p1.From.Offset = offset
p1.To.Type = obj.TYPE_REG
p1.To.Reg = reg
p2 := obj.Appendp(ctxt, p1)
p2.As = obj.ACALL
p2.To.Type = obj.TYPE_REG
p2.To.Reg = reg
}
// We only care about global data: NAME_EXTERN means a global
// symbol in the Go sense, and p.Sym.Local is true for a few
// internally defined symbols.
if p.As == lea && p.From.Type == obj.TYPE_MEM && p.From.Name == obj.NAME_EXTERN && !p.From.Sym.Local {
// $LEA sym, Rx becomes $MOV $sym, Rx which will be rewritten below
p.As = mov
p.From.Type = obj.TYPE_ADDR
}
if p.From.Type == obj.TYPE_ADDR && p.From.Name == obj.NAME_EXTERN && !p.From.Sym.Local {
// $MOV $sym, Rx becomes $MOV sym@GOT, Rx
// $MOV $sym+<off>, Rx becomes $MOV sym@GOT, Rx; $ADD <off>, Rx
// On 386 only, more complicated things like PUSHL $sym become $MOV sym@GOT, CX; PUSHL CX
cmplxdest := false
pAs := p.As
var dest obj.Addr
if p.To.Type != obj.TYPE_REG || pAs != mov {
if p.Mode == 64 {
ctxt.Diag("do not know how to handle LEA-type insn to non-register in %v with -dynlink", p)
}
cmplxdest = true
dest = p.To
p.As = mov
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_CX
p.To.Sym = nil
p.To.Name = obj.NAME_NONE
}
p.From.Type = obj.TYPE_MEM
p.From.Name = obj.NAME_GOTREF
q := p
if p.From.Offset != 0 {
q = obj.Appendp(ctxt, p)
q.As = add
q.From.Type = obj.TYPE_CONST
q.From.Offset = p.From.Offset
q.To = p.To
p.From.Offset = 0
}
if cmplxdest {
q = obj.Appendp(ctxt, q)
q.As = pAs
q.To = dest
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_CX
}
}
if p.From3 != nil && p.From3.Name == obj.NAME_EXTERN {
ctxt.Diag("don't know how to handle %v with -dynlink", p)
}
var source *obj.Addr
// MOVx sym, Ry becomes $MOV sym@GOT, R15; MOVx (R15), Ry
// MOVx Ry, sym becomes $MOV sym@GOT, R15; MOVx Ry, (R15)
// An addition may be inserted between the two MOVs if there is an offset.
if p.From.Name == obj.NAME_EXTERN && !p.From.Sym.Local {
if p.To.Name == obj.NAME_EXTERN && !p.To.Sym.Local {
ctxt.Diag("cannot handle NAME_EXTERN on both sides in %v with -dynlink", p)
}
source = &p.From
} else if p.To.Name == obj.NAME_EXTERN && !p.To.Sym.Local {
source = &p.To
} else {
return
}
if p.As == obj.ACALL {
// When dynlinking on 386, almost any call might end up being a call
// to a PLT, so make sure the GOT pointer is loaded into BX.
// RegTo2 is set on the replacement call insn to stop it being
// processed when it is in turn passed to progedit.
if p.Mode == 64 || (p.To.Sym != nil && p.To.Sym.Local) || p.RegTo2 != 0 {
return
}
p1 := obj.Appendp(ctxt, p)
p2 := obj.Appendp(ctxt, p1)
p1.As = ALEAL
p1.From.Type = obj.TYPE_MEM
p1.From.Name = obj.NAME_STATIC
p1.From.Sym = obj.Linklookup(ctxt, "_GLOBAL_OFFSET_TABLE_", 0)
p1.To.Type = obj.TYPE_REG
p1.To.Reg = REG_BX
p2.As = p.As
p2.Scond = p.Scond
p2.From = p.From
p2.From3 = p.From3
p2.Reg = p.Reg
p2.To = p.To
// p.To.Type was set to TYPE_BRANCH above, but that makes checkaddr
// in ../pass.go complain, so set it back to TYPE_MEM here, until p2
// itself gets passed to progedit.
p2.To.Type = obj.TYPE_MEM
p2.RegTo2 = 1
obj.Nopout(p)
return
}
if p.As == obj.ATEXT || p.As == obj.AFUNCDATA || p.As == obj.ARET || p.As == obj.AJMP {
return
}
if source.Type != obj.TYPE_MEM {
ctxt.Diag("don't know how to handle %v with -dynlink", p)
}
p1 := obj.Appendp(ctxt, p)
p2 := obj.Appendp(ctxt, p1)
p1.As = mov
p1.From.Type = obj.TYPE_MEM
p1.From.Sym = source.Sym
p1.From.Name = obj.NAME_GOTREF
p1.To.Type = obj.TYPE_REG
p1.To.Reg = reg
p2.As = p.As
p2.From = p.From
p2.To = p.To
if p.From.Name == obj.NAME_EXTERN {
p2.From.Reg = reg
p2.From.Name = obj.NAME_NONE
p2.From.Sym = nil
} else if p.To.Name == obj.NAME_EXTERN {
p2.To.Reg = reg
p2.To.Name = obj.NAME_NONE
p2.To.Sym = nil
} else {
return
}
obj.Nopout(p)
}
func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
ctxt.Cursym = cursym
if cursym.Text == nil || cursym.Text.Link == nil {
return
}
p := cursym.Text
textstksiz := p.To.Offset
aoffset := int32(textstksiz)
cursym.Args = p.To.Val.(int32)
cursym.Locals = int32(textstksiz)
/*
* find leaf subroutines
* strip NOPs
* expand RET
*/
ctxt.Bso.Flush()
q := (*obj.Prog)(nil)
var q1 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
switch p.As {
case obj.ATEXT:
p.Mark |= LEAF
case obj.ARET:
break
case obj.ANOP:
q1 = p.Link
q.Link = q1 /* q is non-nop */
q1.Mark |= p.Mark
continue
case ABL,
obj.ADUFFZERO,
obj.ADUFFCOPY:
cursym.Text.Mark &^= LEAF
fallthrough
case ACBNZ,
ACBZ,
ACBNZW,
ACBZW,
ATBZ,
ATBNZ,
AB,
ABEQ,
ABNE,
ABCS,
ABHS,
ABCC,
ABLO,
ABMI,
ABPL,
ABVS,
ABVC,
ABHI,
ABLS,
ABGE,
ABLT,
ABGT,
ABLE,
AADR, /* strange */
AADRP:
q1 = p.Pcond
if q1 != nil {
for q1.As == obj.ANOP {
q1 = q1.Link
p.Pcond = q1
}
}
break
}
q = p
}
var o int
var q2 *obj.Prog
var retjmp *obj.LSym
for p := cursym.Text; p != nil; p = p.Link {
o = int(p.As)
switch o {
case obj.ATEXT:
cursym.Text = p
if textstksiz < 0 {
ctxt.Autosize = 0
} else {
ctxt.Autosize = int32(textstksiz + 8)
}
if (cursym.Text.Mark&LEAF != 0) && ctxt.Autosize <= 8 {
ctxt.Autosize = 0
} else if ctxt.Autosize&(16-1) != 0 {
// The frame includes an LR.
// If the frame size is 8, it's only an LR,
// so there's no potential for breaking references to
// local variables by growing the frame size,
// because there are no local variables.
// But otherwise, if there is a non-empty locals section,
// the author of the code is responsible for making sure
// that the frame size is 8 mod 16.
if ctxt.Autosize == 8 {
ctxt.Autosize += 8
cursym.Locals += 8
} else {
ctxt.Diag("%v: unaligned frame size %d - must be 8 mod 16 (or 0)", p, ctxt.Autosize-8)
}
}
p.To.Offset = int64(ctxt.Autosize) - 8
if ctxt.Autosize == 0 && !(cursym.Text.Mark&LEAF != 0) {
if ctxt.Debugvlog != 0 {
fmt.Fprintf(ctxt.Bso, "save suppressed in: %s\n", cursym.Text.From.Sym.Name)
}
ctxt.Bso.Flush()
cursym.Text.Mark |= LEAF
}
if !(p.From3.Offset&obj.NOSPLIT != 0) {
p = stacksplit(ctxt, p, ctxt.Autosize) // emit split check
}
aoffset = ctxt.Autosize
if aoffset > 0xF0 {
aoffset = 0xF0
}
if cursym.Text.Mark&LEAF != 0 {
cursym.Leaf = 1
if ctxt.Autosize == 0 {
break
}
aoffset = 0
}
q = p
if ctxt.Autosize > aoffset {
q = ctxt.NewProg()
q.As = ASUB
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(ctxt.Autosize) - int64(aoffset)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Spadj = int32(q.From.Offset)
q.Link = p.Link
p.Link = q
if cursym.Text.Mark&LEAF != 0 {
break
}
}
q1 = ctxt.NewProg()
q1.As = AMOVD
q1.Lineno = p.Lineno
q1.From.Type = obj.TYPE_REG
q1.From.Reg = REGLINK
q1.To.Type = obj.TYPE_MEM
q1.Scond = C_XPRE
q1.To.Offset = int64(-aoffset)
q1.To.Reg = REGSP
q1.Link = q.Link
q1.Spadj = aoffset
q.Link = q1
if cursym.Text.From3.Offset&obj.WRAPPER != 0 {
// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
//
// MOV g_panic(g), R1
// CMP ZR, R1
// BEQ end
// MOV panic_argp(R1), R2
// ADD $(autosize+8), RSP, R3
// CMP R2, R3
// BNE end
// ADD $8, RSP, R4
// MOVD R4, panic_argp(R1)
// end:
// NOP
//
// The NOP is needed to give the jumps somewhere to land.
// It is a liblink NOP, not a ARM64 NOP: it encodes to 0 instruction bytes.
q = q1
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_MEM
q.From.Reg = REGG
q.From.Offset = 4 * int64(ctxt.Arch.Ptrsize) // G.panic
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R1
q = obj.Appendp(ctxt, q)
q.As = ACMP
q.From.Type = obj.TYPE_REG
q.From.Reg = REGZERO
q.Reg = REG_R1
q = obj.Appendp(ctxt, q)
q.As = ABEQ
q.To.Type = obj.TYPE_BRANCH
q1 = q
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_MEM
q.From.Reg = REG_R1
q.From.Offset = 0 // Panic.argp
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R2
q = obj.Appendp(ctxt, q)
q.As = AADD
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(ctxt.Autosize) + 8
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R3
q = obj.Appendp(ctxt, q)
q.As = ACMP
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R2
q.Reg = REG_R3
q = obj.Appendp(ctxt, q)
q.As = ABNE
q.To.Type = obj.TYPE_BRANCH
q2 = q
q = obj.Appendp(ctxt, q)
q.As = AADD
q.From.Type = obj.TYPE_CONST
q.From.Offset = 8
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R4
q = obj.Appendp(ctxt, q)
q.As = AMOVD
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R4
q.To.Type = obj.TYPE_MEM
q.To.Reg = REG_R1
q.To.Offset = 0 // Panic.argp
q = obj.Appendp(ctxt, q)
q.As = obj.ANOP
q1.Pcond = q
q2.Pcond = q
}
case obj.ARET:
nocache(p)
if p.From.Type == obj.TYPE_CONST {
ctxt.Diag("using BECOME (%v) is not supported!", p)
break
}
retjmp = p.To.Sym
p.To = obj.Addr{}
if cursym.Text.Mark&LEAF != 0 {
if ctxt.Autosize != 0 {
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(ctxt.Autosize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REGSP
p.Spadj = -ctxt.Autosize
}
} else {
/* want write-back pre-indexed SP+autosize -> SP, loading REGLINK*/
aoffset = ctxt.Autosize
if aoffset > 0xF0 {
aoffset = 0xF0
}
p.As = AMOVD
p.From.Type = obj.TYPE_MEM
p.Scond = C_XPOST
p.From.Offset = int64(aoffset)
p.From.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REGLINK
p.Spadj = -aoffset
if ctxt.Autosize > aoffset {
q = ctxt.NewProg()
q.As = AADD
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(ctxt.Autosize) - int64(aoffset)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Link = p.Link
q.Spadj = int32(-q.From.Offset)
q.Lineno = p.Lineno
p.Link = q
p = q
}
}
if p.As != obj.ARET {
q = ctxt.NewProg()
q.Lineno = p.Lineno
q.Link = p.Link
p.Link = q
p = q
}
if retjmp != nil { // retjmp
p.As = AB
p.To.Type = obj.TYPE_BRANCH
p.To.Sym = retjmp
p.Spadj = +ctxt.Autosize
break
}
p.As = obj.ARET
p.To.Type = obj.TYPE_MEM
p.To.Offset = 0
p.To.Reg = REGLINK
p.Spadj = +ctxt.Autosize
case AADD, ASUB:
if p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP && p.From.Type == obj.TYPE_CONST {
if p.As == AADD {
p.Spadj = int32(-p.From.Offset)
} else {
p.Spadj = int32(+p.From.Offset)
}
}
break
}
}
}
func span0(ctxt *obj.Link, cursym *obj.LSym) {
p := cursym.Text
if p == nil || p.Link == nil { // handle external functions and ELF section symbols
return
}
ctxt.Cursym = cursym
ctxt.Autosize = int32(p.To.Offset + 8)
if oprange[AOR&obj.AMask].start == nil {
buildop(ctxt)
}
c := int64(0)
p.Pc = c
var m int
var o *Optab
for p = p.Link; p != nil; p = p.Link {
ctxt.Curp = p
p.Pc = c
o = oplook(ctxt, p)
m = int(o.size)
if m == 0 {
if p.As != obj.ANOP && p.As != obj.AFUNCDATA && p.As != obj.APCDATA && p.As != obj.AUSEFIELD {
ctxt.Diag("zero-width instruction\n%v", p)
}
continue
}
c += int64(m)
}
cursym.Size = c
/*
* if any procedure is large enough to
* generate a large SBRA branch, then
* generate extra passes putting branches
* around jmps to fix. this is rare.
*/
bflag := 1
var otxt int64
var q *obj.Prog
for bflag != 0 {
if ctxt.Debugvlog != 0 {
fmt.Fprintf(ctxt.Bso, "%5.2f span1\n", obj.Cputime())
}
bflag = 0
c = 0
for p = cursym.Text.Link; p != nil; p = p.Link {
p.Pc = c
o = oplook(ctxt, p)
// very large conditional branches
if o.type_ == 6 && p.Pcond != nil {
otxt = p.Pcond.Pc - c
if otxt < -(1<<17)+10 || otxt >= (1<<17)-10 {
q = ctxt.NewProg()
q.Link = p.Link
p.Link = q
q.As = AJMP
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_BRANCH
q.Pcond = p.Pcond
p.Pcond = q
q = ctxt.NewProg()
q.Link = p.Link
p.Link = q
q.As = AJMP
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_BRANCH
q.Pcond = q.Link.Link
addnop(ctxt, p.Link)
addnop(ctxt, p)
bflag = 1
}
}
m = int(o.size)
if m == 0 {
if p.As != obj.ANOP && p.As != obj.AFUNCDATA && p.As != obj.APCDATA && p.As != obj.AUSEFIELD {
ctxt.Diag("zero-width instruction\n%v", p)
}
continue
}
c += int64(m)
}
cursym.Size = c
}
c += -c & (FuncAlign - 1)
cursym.Size = c
/*
* lay out the code, emitting code and data relocations.
*/
obj.Symgrow(ctxt, cursym, cursym.Size)
bp := cursym.P
var i int32
var out [4]uint32
for p := cursym.Text.Link; p != nil; p = p.Link {
ctxt.Pc = p.Pc
ctxt.Curp = p
o = oplook(ctxt, p)
if int(o.size) > 4*len(out) {
log.Fatalf("out array in span0 is too small, need at least %d for %v", o.size/4, p)
}
asmout(ctxt, p, o, out[:])
for i = 0; i < int32(o.size/4); i++ {
ctxt.Arch.ByteOrder.PutUint32(bp, out[i])
bp = bp[4:]
}
}
}
func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
autosize := int32(0)
ctxt.Cursym = cursym
if cursym.Text == nil || cursym.Text.Link == nil {
return
}
softfloat(ctxt, cursym)
p := cursym.Text
autoffset := int32(p.To.Offset)
if autoffset < 0 {
autoffset = 0
}
cursym.Locals = autoffset
cursym.Args = p.To.Val.(int32)
/*
* find leaf subroutines
* strip NOPs
* expand RET
* expand BECOME pseudo
*/
var q1 *obj.Prog
var q *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
switch p.As {
case obj.ATEXT:
p.Mark |= LEAF
case obj.ARET:
break
case ADIV, ADIVU, AMOD, AMODU:
q = p
if ctxt.Sym_div == nil {
initdiv(ctxt)
}
cursym.Text.Mark &^= LEAF
continue
case obj.ANOP:
q1 = p.Link
q.Link = q1 /* q is non-nop */
if q1 != nil {
q1.Mark |= p.Mark
}
continue
case ABL,
ABX,
obj.ADUFFZERO,
obj.ADUFFCOPY:
cursym.Text.Mark &^= LEAF
fallthrough
case AB,
ABEQ,
ABNE,
ABCS,
ABHS,
ABCC,
ABLO,
ABMI,
ABPL,
ABVS,
ABVC,
ABHI,
ABLS,
ABGE,
ABLT,
ABGT,
ABLE:
q1 = p.Pcond
if q1 != nil {
for q1.As == obj.ANOP {
q1 = q1.Link
p.Pcond = q1
}
}
}
q = p
}
var o int
var p1 *obj.Prog
var p2 *obj.Prog
var q2 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
o = int(p.As)
switch o {
case obj.ATEXT:
autosize = int32(p.To.Offset + 4)
if autosize <= 4 {
if cursym.Text.Mark&LEAF != 0 {
p.To.Offset = -4
autosize = 0
}
}
if autosize == 0 && cursym.Text.Mark&LEAF == 0 {
if ctxt.Debugvlog != 0 {
fmt.Fprintf(ctxt.Bso, "save suppressed in: %s\n", cursym.Name)
ctxt.Bso.Flush()
}
cursym.Text.Mark |= LEAF
}
if cursym.Text.Mark&LEAF != 0 {
cursym.Leaf = 1
if autosize == 0 {
break
}
}
if p.From3.Offset&obj.NOSPLIT == 0 {
p = stacksplit(ctxt, p, autosize) // emit split check
}
// MOVW.W R14,$-autosize(SP)
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.Scond |= C_WBIT
p.From.Type = obj.TYPE_REG
p.From.Reg = REGLINK
p.To.Type = obj.TYPE_MEM
p.To.Offset = int64(-autosize)
p.To.Reg = REGSP
p.Spadj = autosize
if cursym.Text.From3.Offset&obj.WRAPPER != 0 {
// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
//
// MOVW g_panic(g), R1
// CMP $0, R1
// B.EQ end
// MOVW panic_argp(R1), R2
// ADD $(autosize+4), R13, R3
// CMP R2, R3
// B.NE end
// ADD $4, R13, R4
// MOVW R4, panic_argp(R1)
// end:
// NOP
//
// The NOP is needed to give the jumps somewhere to land.
// It is a liblink NOP, not an ARM NOP: it encodes to 0 instruction bytes.
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGG
p.From.Offset = 4 * int64(ctxt.Arch.Ptrsize) // G.panic
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_CONST
p.From.Offset = 0
p.Reg = REG_R1
p = obj.Appendp(ctxt, p)
p.As = ABEQ
p.To.Type = obj.TYPE_BRANCH
p1 = p
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_R1
p.From.Offset = 0 // Panic.argp
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(autosize) + 4
p.Reg = REG_R13
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R3
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R2
p.Reg = REG_R3
p = obj.Appendp(ctxt, p)
p.As = ABNE
p.To.Type = obj.TYPE_BRANCH
p2 = p
p = obj.Appendp(ctxt, p)
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = 4
p.Reg = REG_R13
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R4
p.To.Type = obj.TYPE_MEM
p.To.Reg = REG_R1
p.To.Offset = 0 // Panic.argp
p = obj.Appendp(ctxt, p)
p.As = obj.ANOP
p1.Pcond = p
p2.Pcond = p
}
case obj.ARET:
obj.Nocache(p)
if cursym.Text.Mark&LEAF != 0 {
if autosize == 0 {
p.As = AB
p.From = obj.Addr{}
if p.To.Sym != nil { // retjmp
p.To.Type = obj.TYPE_BRANCH
} else {
p.To.Type = obj.TYPE_MEM
p.To.Offset = 0
p.To.Reg = REGLINK
}
break
}
}
p.As = AMOVW
p.Scond |= C_PBIT
p.From.Type = obj.TYPE_MEM
p.From.Offset = int64(autosize)
p.From.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REGPC
// If there are instructions following
// this ARET, they come from a branch
// with the same stackframe, so no spadj.
if p.To.Sym != nil { // retjmp
p.To.Reg = REGLINK
q2 = obj.Appendp(ctxt, p)
q2.As = AB
q2.To.Type = obj.TYPE_BRANCH
q2.To.Sym = p.To.Sym
p.To.Sym = nil
p = q2
}
case AADD:
if p.From.Type == obj.TYPE_CONST && p.From.Reg == 0 && p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP {
p.Spadj = int32(-p.From.Offset)
}
case ASUB:
if p.From.Type == obj.TYPE_CONST && p.From.Reg == 0 && p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP {
p.Spadj = int32(p.From.Offset)
}
case ADIV, ADIVU, AMOD, AMODU:
if cursym.Text.From3.Offset&obj.NOSPLIT != 0 {
ctxt.Diag("cannot divide in NOSPLIT function")
}
if ctxt.Debugdivmod != 0 {
break
}
if p.From.Type != obj.TYPE_REG {
break
}
if p.To.Type != obj.TYPE_REG {
break
}
// Make copy because we overwrite p below.
q1 := *p
if q1.Reg == REGTMP || q1.Reg == 0 && q1.To.Reg == REGTMP {
ctxt.Diag("div already using REGTMP: %v", p)
}
/* MOV m(g),REGTMP */
p.As = AMOVW
p.Lineno = q1.Lineno
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGG
p.From.Offset = 6 * 4 // offset of g.m
p.Reg = 0
p.To.Type = obj.TYPE_REG
p.To.Reg = REGTMP
/* MOV a,m_divmod(REGTMP) */
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.Lineno = q1.Lineno
p.From.Type = obj.TYPE_REG
p.From.Reg = q1.From.Reg
p.To.Type = obj.TYPE_MEM
p.To.Reg = REGTMP
p.To.Offset = 8 * 4 // offset of m.divmod
/* MOV b,REGTMP */
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.Lineno = q1.Lineno
p.From.Type = obj.TYPE_REG
p.From.Reg = q1.Reg
if q1.Reg == 0 {
p.From.Reg = q1.To.Reg
}
p.To.Type = obj.TYPE_REG
p.To.Reg = REGTMP
p.To.Offset = 0
/* CALL appropriate */
p = obj.Appendp(ctxt, p)
p.As = ABL
p.Lineno = q1.Lineno
p.To.Type = obj.TYPE_BRANCH
switch o {
case ADIV:
p.To.Sym = ctxt.Sym_div
case ADIVU:
p.To.Sym = ctxt.Sym_divu
case AMOD:
p.To.Sym = ctxt.Sym_mod
case AMODU:
p.To.Sym = ctxt.Sym_modu
}
/* MOV REGTMP, b */
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.Lineno = q1.Lineno
p.From.Type = obj.TYPE_REG
p.From.Reg = REGTMP
p.From.Offset = 0
p.To.Type = obj.TYPE_REG
p.To.Reg = q1.To.Reg
case AMOVW:
if (p.Scond&C_WBIT != 0) && p.To.Type == obj.TYPE_MEM && p.To.Reg == REGSP {
p.Spadj = int32(-p.To.Offset)
}
if (p.Scond&C_PBIT != 0) && p.From.Type == obj.TYPE_MEM && p.From.Reg == REGSP && p.To.Reg != REGPC {
p.Spadj = int32(-p.From.Offset)
}
if p.From.Type == obj.TYPE_ADDR && p.From.Reg == REGSP && p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP {
p.Spadj = int32(-p.From.Offset)
}
}
}
}
/*
* ASLL x,y,w
* .. (not use w, not set x y w)
* AXXX w,a,b (a != w)
* .. (not use w)
* (set w)
* ----------- changed to
* ..
* AXXX (x<<y),a,b
* ..
*/
func shiftprop(r *gc.Flow) bool {
p := (*obj.Prog)(r.Prog)
if p.To.Type != obj.TYPE_REG {
if gc.Debug['P'] != 0 {
fmt.Printf("\tBOTCH: result not reg; FAILURE\n")
}
return false
}
n := int(int(p.To.Reg))
a := obj.Addr(obj.Addr{})
if p.Reg != 0 && p.Reg != p.To.Reg {
a.Type = obj.TYPE_REG
a.Reg = p.Reg
}
if gc.Debug['P'] != 0 {
fmt.Printf("shiftprop\n%v", p)
}
r1 := (*gc.Flow)(r)
var p1 *obj.Prog
for {
/* find first use of shift result; abort if shift operands or result are changed */
r1 = gc.Uniqs(r1)
if r1 == nil {
if gc.Debug['P'] != 0 {
fmt.Printf("\tbranch; FAILURE\n")
}
return false
}
if gc.Uniqp(r1) == nil {
if gc.Debug['P'] != 0 {
fmt.Printf("\tmerge; FAILURE\n")
}
return false
}
p1 = r1.Prog
if gc.Debug['P'] != 0 {
fmt.Printf("\n%v", p1)
}
switch copyu(p1, &p.To, nil) {
case 0: /* not used or set */
if (p.From.Type == obj.TYPE_REG && copyu(p1, &p.From, nil) > 1) || (a.Type == obj.TYPE_REG && copyu(p1, &a, nil) > 1) {
if gc.Debug['P'] != 0 {
fmt.Printf("\targs modified; FAILURE\n")
}
return false
}
continue
case 3: /* set, not used */
{
if gc.Debug['P'] != 0 {
fmt.Printf("\tBOTCH: noref; FAILURE\n")
}
return false
}
}
break
}
/* check whether substitution can be done */
switch p1.As {
default:
if gc.Debug['P'] != 0 {
fmt.Printf("\tnon-dpi; FAILURE\n")
}
return false
case arm.AAND,
arm.AEOR,
arm.AADD,
arm.AADC,
arm.AORR,
arm.ASUB,
arm.ASBC,
arm.ARSB,
arm.ARSC:
if int(p1.Reg) == n || (p1.Reg == 0 && p1.To.Type == obj.TYPE_REG && int(p1.To.Reg) == n) {
if p1.From.Type != obj.TYPE_REG {
if gc.Debug['P'] != 0 {
fmt.Printf("\tcan't swap; FAILURE\n")
}
return false
}
p1.Reg = p1.From.Reg
p1.From.Reg = int16(n)
switch p1.As {
case arm.ASUB:
p1.As = arm.ARSB
case arm.ARSB:
p1.As = arm.ASUB
case arm.ASBC:
p1.As = arm.ARSC
case arm.ARSC:
p1.As = arm.ASBC
}
if gc.Debug['P'] != 0 {
fmt.Printf("\t=>%v", p1)
}
}
fallthrough
case arm.ABIC,
arm.ATST,
arm.ACMP,
arm.ACMN:
if int(p1.Reg) == n {
if gc.Debug['P'] != 0 {
fmt.Printf("\tcan't swap; FAILURE\n")
}
return false
}
if p1.Reg == 0 && int(p1.To.Reg) == n {
if gc.Debug['P'] != 0 {
fmt.Printf("\tshift result used twice; FAILURE\n")
}
return false
}
// case AMVN:
if p1.From.Type == obj.TYPE_SHIFT {
if gc.Debug['P'] != 0 {
fmt.Printf("\tshift result used in shift; FAILURE\n")
}
return false
}
if p1.From.Type != obj.TYPE_REG || int(p1.From.Reg) != n {
if gc.Debug['P'] != 0 {
fmt.Printf("\tBOTCH: where is it used?; FAILURE\n")
}
return false
}
}
/* check whether shift result is used subsequently */
p2 := (*obj.Prog)(p1)
if int(p1.To.Reg) != n {
var p1 *obj.Prog
for {
r1 = gc.Uniqs(r1)
if r1 == nil {
if gc.Debug['P'] != 0 {
fmt.Printf("\tinconclusive; FAILURE\n")
}
return false
}
p1 = r1.Prog
if gc.Debug['P'] != 0 {
fmt.Printf("\n%v", p1)
}
switch copyu(p1, &p.To, nil) {
case 0: /* not used or set */
continue
case 3: /* set, not used */
break
default: /* used */
if gc.Debug['P'] != 0 {
fmt.Printf("\treused; FAILURE\n")
}
return false
}
break
}
}
/* make the substitution */
p2.From.Reg = 0
o := int(int(p.Reg))
if o == 0 {
o = int(p.To.Reg)
}
o &= 15
switch p.From.Type {
case obj.TYPE_CONST:
o |= int((p.From.Offset & 0x1f) << 7)
case obj.TYPE_REG:
o |= 1<<4 | (int(p.From.Reg)&15)<<8
}
switch p.As {
case arm.ASLL:
o |= 0 << 5
case arm.ASRL:
o |= 1 << 5
case arm.ASRA:
o |= 2 << 5
}
p2.From = obj.Addr{}
p2.From.Type = obj.TYPE_SHIFT
p2.From.Offset = int64(o)
if gc.Debug['P'] != 0 {
fmt.Printf("\t=>%v\tSUCCEED\n", p2)
}
return true
}
func Clearp(p *obj.Prog) {
obj.Nopout(p)
p.As = obj.AEND
p.Pc = int64(pcloc)
pcloc++
}
func sched(ctxt *obj.Link, p0, pe *obj.Prog) {
var sch [NSCHED]Sch
/*
* build side structure
*/
s := sch[:]
for p := p0; ; p = p.Link {
s[0].p = *p
markregused(ctxt, &s[0])
if p == pe {
break
}
s = s[1:]
}
se := s
for i := cap(sch) - cap(se); i >= 0; i-- {
s = sch[i:]
if s[0].p.Mark&DELAY == 0 {
continue
}
if -cap(s) < -cap(se) {
if !conflict(&s[0], &s[1]) {
continue
}
}
var t []Sch
var j int
for j = cap(sch) - cap(s) - 1; j >= 0; j-- {
t = sch[j:]
if t[0].comp {
if s[0].p.Mark&BRANCH != 0 {
goto no2
}
}
if t[0].p.Mark&DELAY != 0 {
if -cap(s) >= -cap(se) || conflict(&t[0], &s[1]) {
goto no2
}
}
for u := t[1:]; -cap(u) <= -cap(s); u = u[1:] {
if depend(ctxt, &u[0], &t[0]) {
goto no2
}
}
goto out2
no2:
}
if s[0].p.Mark&BRANCH != 0 {
s[0].nop = 1
}
continue
out2:
// t[0] is the instruction being moved to fill the delay
stmp := t[0]
copy(t[:i-j], t[1:i-j+1])
s[0] = stmp
if t[i-j-1].p.Mark&BRANCH != 0 {
// t[i-j] is being put into a branch delay slot
// combine its Spadj with the branch instruction
t[i-j-1].p.Spadj += t[i-j].p.Spadj
t[i-j].p.Spadj = 0
}
i--
}
/*
* put it all back
*/
var p *obj.Prog
var q *obj.Prog
for s, p = sch[:], p0; -cap(s) <= -cap(se); s, p = s[1:], q {
q = p.Link
if q != s[0].p.Link {
*p = s[0].p
p.Link = q
}
for s[0].nop != 0 {
s[0].nop--
addnop(ctxt, p)
}
}
}
func markregused(ctxt *obj.Link, s *Sch) {
p := &s.p
s.comp = compound(ctxt, p)
s.nop = 0
if s.comp {
s.set.ireg |= 1 << (REGTMP - REG_R0)
s.used.ireg |= 1 << (REGTMP - REG_R0)
}
ar := 0 /* dest is really reference */
ad := 0 /* source/dest is really address */
ld := 0 /* opcode is load instruction */
sz := 20 /* size of load/store for overlap computation */
/*
* flags based on opcode
*/
switch p.As {
case obj.ATEXT:
ctxt.Autosize = int32(p.To.Offset + 8)
ad = 1
case AJAL:
c := p.Reg
if c == 0 {
c = REGLINK
}
s.set.ireg |= 1 << uint(c-REG_R0)
ar = 1
ad = 1
case ABGEZAL,
ABLTZAL:
s.set.ireg |= 1 << (REGLINK - REG_R0)
fallthrough
case ABEQ,
ABGEZ,
ABGTZ,
ABLEZ,
ABLTZ,
ABNE:
ar = 1
ad = 1
case ABFPT,
ABFPF:
ad = 1
s.used.cc |= E_FCR
case ACMPEQD,
ACMPEQF,
ACMPGED,
ACMPGEF,
ACMPGTD,
ACMPGTF:
ar = 1
s.set.cc |= E_FCR
p.Mark |= FCMP
case AJMP:
ar = 1
ad = 1
case AMOVB,
AMOVBU:
sz = 1
ld = 1
case AMOVH,
AMOVHU:
sz = 2
ld = 1
case AMOVF,
AMOVW,
AMOVWL,
AMOVWR:
sz = 4
ld = 1
case AMOVD,
AMOVV,
AMOVVL,
AMOVVR:
sz = 8
ld = 1
case ADIV,
ADIVU,
AMUL,
AMULU,
AREM,
AREMU,
ADIVV,
ADIVVU,
AMULV,
AMULVU,
AREMV,
AREMVU:
s.set.cc = E_HILO
fallthrough
case AADD,
AADDU,
AADDV,
AADDVU,
AAND,
ANOR,
AOR,
ASGT,
ASGTU,
ASLL,
ASRA,
ASRL,
ASLLV,
ASRAV,
ASRLV,
ASUB,
ASUBU,
ASUBV,
ASUBVU,
AXOR,
AADDD,
AADDF,
AADDW,
ASUBD,
ASUBF,
ASUBW,
AMULF,
AMULD,
AMULW,
ADIVF,
ADIVD,
ADIVW:
if p.Reg == 0 {
if p.To.Type == obj.TYPE_REG {
p.Reg = p.To.Reg
}
//if(p->reg == NREG)
// print("botch %P\n", p);
}
}
/*
* flags based on 'to' field
*/
c := int(p.To.Class)
if c == 0 {
c = aclass(ctxt, &p.To) + 1
p.To.Class = int8(c)
}
c--
switch c {
default:
fmt.Printf("unknown class %d %v\n", c, p)
case C_ZCON,
C_SCON,
C_ADD0CON,
C_AND0CON,
C_ADDCON,
C_ANDCON,
C_UCON,
C_LCON,
C_NONE,
C_SBRA,
C_LBRA,
C_ADDR,
C_TEXTSIZE:
break
case C_HI,
C_LO:
s.set.cc |= E_HILO
case C_FCREG:
s.set.cc |= E_FCR
case C_MREG:
s.set.cc |= E_MCR
case C_ZOREG,
C_SOREG,
C_LOREG:
c = int(p.To.Reg)
s.used.ireg |= 1 << uint(c-REG_R0)
if ad != 0 {
break
}
s.size = uint8(sz)
s.soffset = regoff(ctxt, &p.To)
m := uint32(ANYMEM)
if c == REGSB {
m = E_MEMSB
}
if c == REGSP {
m = E_MEMSP
}
if ar != 0 {
s.used.cc |= m
} else {
s.set.cc |= m
}
case C_SACON,
C_LACON:
s.used.ireg |= 1 << (REGSP - REG_R0)
case C_SECON,
C_LECON:
s.used.ireg |= 1 << (REGSB - REG_R0)
case C_REG:
if ar != 0 {
s.used.ireg |= 1 << uint(p.To.Reg-REG_R0)
} else {
s.set.ireg |= 1 << uint(p.To.Reg-REG_R0)
}
case C_FREG:
if ar != 0 {
s.used.freg |= 1 << uint(p.To.Reg-REG_F0)
} else {
s.set.freg |= 1 << uint(p.To.Reg-REG_F0)
}
if ld != 0 && p.From.Type == obj.TYPE_REG {
p.Mark |= LOAD
}
case C_SAUTO,
C_LAUTO:
s.used.ireg |= 1 << (REGSP - REG_R0)
if ad != 0 {
break
}
s.size = uint8(sz)
s.soffset = regoff(ctxt, &p.To)
if ar != 0 {
s.used.cc |= E_MEMSP
} else {
s.set.cc |= E_MEMSP
}
case C_SEXT,
C_LEXT:
s.used.ireg |= 1 << (REGSB - REG_R0)
if ad != 0 {
break
}
s.size = uint8(sz)
s.soffset = regoff(ctxt, &p.To)
if ar != 0 {
s.used.cc |= E_MEMSB
} else {
s.set.cc |= E_MEMSB
}
}
/*
* flags based on 'from' field
*/
c = int(p.From.Class)
if c == 0 {
c = aclass(ctxt, &p.From) + 1
p.From.Class = int8(c)
}
c--
switch c {
default:
fmt.Printf("unknown class %d %v\n", c, p)
case C_ZCON,
C_SCON,
C_ADD0CON,
C_AND0CON,
C_ADDCON,
C_ANDCON,
C_UCON,
C_LCON,
C_NONE,
C_SBRA,
C_LBRA,
C_ADDR,
C_TEXTSIZE:
break
case C_HI,
C_LO:
s.used.cc |= E_HILO
case C_FCREG:
s.used.cc |= E_FCR
case C_MREG:
s.used.cc |= E_MCR
case C_ZOREG,
C_SOREG,
C_LOREG:
c = int(p.From.Reg)
s.used.ireg |= 1 << uint(c-REG_R0)
if ld != 0 {
p.Mark |= LOAD
}
s.size = uint8(sz)
s.soffset = regoff(ctxt, &p.From)
m := uint32(ANYMEM)
if c == REGSB {
m = E_MEMSB
}
if c == REGSP {
m = E_MEMSP
}
s.used.cc |= m
case C_SACON,
C_LACON:
c = int(p.From.Reg)
if c == 0 {
c = REGSP
}
s.used.ireg |= 1 << uint(c-REG_R0)
case C_SECON,
C_LECON:
s.used.ireg |= 1 << (REGSB - REG_R0)
case C_REG:
s.used.ireg |= 1 << uint(p.From.Reg-REG_R0)
case C_FREG:
s.used.freg |= 1 << uint(p.From.Reg-REG_F0)
if ld != 0 && p.To.Type == obj.TYPE_REG {
p.Mark |= LOAD
}
case C_SAUTO,
C_LAUTO:
s.used.ireg |= 1 << (REGSP - REG_R0)
if ld != 0 {
p.Mark |= LOAD
}
if ad != 0 {
break
}
s.size = uint8(sz)
s.soffset = regoff(ctxt, &p.From)
s.used.cc |= E_MEMSP
case C_SEXT:
case C_LEXT:
s.used.ireg |= 1 << (REGSB - REG_R0)
if ld != 0 {
p.Mark |= LOAD
}
if ad != 0 {
break
}
s.size = uint8(sz)
s.soffset = regoff(ctxt, &p.From)
s.used.cc |= E_MEMSB
}
c = int(p.Reg)
if c != 0 {
if REG_F0 <= c && c <= REG_F31 {
s.used.freg |= 1 << uint(c-REG_F0)
} else {
s.used.ireg |= 1 << uint(c-REG_R0)
}
}
s.set.ireg &^= (1 << (REGZERO - REG_R0)) /* R0 cant be set */
}
/*
* test to see if 2 instrictions can be
* interchanged without changing semantics
*/
func depend(ctxt *obj.Link, sa, sb *Sch) bool {
if sa.set.ireg&(sb.set.ireg|sb.used.ireg) != 0 {
return true
}
if sb.set.ireg&sa.used.ireg != 0 {
return true
}
if sa.set.freg&(sb.set.freg|sb.used.freg) != 0 {
return true
}
if sb.set.freg&sa.used.freg != 0 {
return true
}
/*
* special case.
* loads from same address cannot pass.
* this is for hardware fifo's and the like
*/
if sa.used.cc&sb.used.cc&E_MEM != 0 {
if sa.p.Reg == sb.p.Reg {
if regoff(ctxt, &sa.p.From) == regoff(ctxt, &sb.p.From) {
return true
}
}
}
x := (sa.set.cc & (sb.set.cc | sb.used.cc)) | (sb.set.cc & sa.used.cc)
if x != 0 {
/*
* allow SB and SP to pass each other.
* allow SB to pass SB iff doffsets are ok
* anything else conflicts
*/
if x != E_MEMSP && x != E_MEMSB {
return true
}
x = sa.set.cc | sb.set.cc | sa.used.cc | sb.used.cc
if x&E_MEM != 0 {
return true
}
if offoverlap(sa, sb) {
return true
}
}
return false
}
func offoverlap(sa, sb *Sch) bool {
if sa.soffset < sb.soffset {
if sa.soffset+int32(sa.size) > sb.soffset {
return true
}
return false
}
if sb.soffset+int32(sb.size) > sa.soffset {
return true
}
return false
}
/*
* test 2 adjacent instructions
* and find out if inserted instructions
* are desired to prevent stalls.
*/
func conflict(sa, sb *Sch) bool {
if sa.set.ireg&sb.used.ireg != 0 {
return true
}
if sa.set.freg&sb.used.freg != 0 {
return true
}
if sa.set.cc&sb.used.cc != 0 {
return true
}
return false
}
func compound(ctxt *obj.Link, p *obj.Prog) bool {
o := oplook(ctxt, p)
if o.size != 4 {
return true
}
if p.To.Type == obj.TYPE_REG && p.To.Reg == REGSB {
return true
}
return false
}
func follow(ctxt *obj.Link, s *obj.LSym) {
ctxt.Cursym = s
firstp := ctxt.NewProg()
lastp := firstp
xfol(ctxt, s.Text, &lastp)
lastp.Link = nil
s.Text = firstp.Link
}
func xfol(ctxt *obj.Link, p *obj.Prog, last **obj.Prog) {
var q *obj.Prog
var r *obj.Prog
var a int
var i int
loop:
if p == nil {
return
}
a = int(p.As)
if a == AJMP {
q = p.Pcond
if (p.Mark&NOSCHED != 0) || q != nil && (q.Mark&NOSCHED != 0) {
p.Mark |= FOLL
(*last).Link = p
*last = p
p = p.Link
xfol(ctxt, p, last)
p = q
if p != nil && p.Mark&FOLL == 0 {
goto loop
}
return
}
if q != nil {
p.Mark |= FOLL
p = q
if p.Mark&FOLL == 0 {
goto loop
}
}
}
if p.Mark&FOLL != 0 {
i = 0
q = p
for ; i < 4; i, q = i+1, q.Link {
if q == *last || (q.Mark&NOSCHED != 0) {
break
}
a = int(q.As)
if a == obj.ANOP {
i--
continue
}
if a == AJMP || a == ARET || a == ARFE {
goto copy
}
if q.Pcond == nil || (q.Pcond.Mark&FOLL != 0) {
continue
}
if a != ABEQ && a != ABNE {
continue
}
copy:
for {
r = ctxt.NewProg()
*r = *p
if r.Mark&FOLL == 0 {
fmt.Printf("cant happen 1\n")
}
r.Mark |= FOLL
if p != q {
p = p.Link
(*last).Link = r
*last = r
continue
}
(*last).Link = r
*last = r
if a == AJMP || a == ARET || a == ARFE {
return
}
r.As = ABNE
if a == ABNE {
r.As = ABEQ
}
r.Pcond = p.Link
r.Link = p.Pcond
if r.Link.Mark&FOLL == 0 {
xfol(ctxt, r.Link, last)
}
if r.Pcond.Mark&FOLL == 0 {
fmt.Printf("cant happen 2\n")
}
return
}
}
a = AJMP
q = ctxt.NewProg()
q.As = int16(a)
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_BRANCH
q.To.Offset = p.Pc
q.Pcond = p
p = q
}
p.Mark |= FOLL
(*last).Link = p
*last = p
if a == AJMP || a == ARET || a == ARFE {
if p.Mark&NOSCHED != 0 {
p = p.Link
goto loop
}
return
}
if p.Pcond != nil {
if a != AJAL && p.Link != nil {
xfol(ctxt, p.Link, last)
p = p.Pcond
if p == nil || (p.Mark&FOLL != 0) {
return
}
goto loop
}
}
p = p.Link
goto loop
}
var Linkmips64 = obj.LinkArch{
ByteOrder: binary.BigEndian,
Name: "mips64",
Thechar: '0',
Preprocess: preprocess,
Assemble: span0,
Follow: follow,
Progedit: progedit,
Minlc: 4,
Ptrsize: 8,
Regsize: 8,
}
var Linkmips64le = obj.LinkArch{
ByteOrder: binary.LittleEndian,
Name: "mips64le",
Thechar: '0',
Preprocess: preprocess,
Assemble: span0,
Follow: follow,
Progedit: progedit,
Minlc: 4,
Ptrsize: 8,
Regsize: 8,
}
// movb elimination.
// movb is simulated by the linker
// when a register other than ax, bx, cx, dx
// is used, so rewrite to other instructions
// when possible. a movb into a register
// can smash the entire 32-bit register without
// causing any trouble.
//
// TODO: Using the Q forms here instead of the L forms
// seems unnecessary, and it makes the instructions longer.
func elimshortmov(g *gc.Graph) {
var p *obj.Prog
for r := (*gc.Flow)(g.Start); r != nil; r = r.Link {
p = r.Prog
if regtyp(&p.To) {
switch p.As {
case x86.AINCB,
x86.AINCW:
p.As = x86.AINCQ
case x86.ADECB,
x86.ADECW:
p.As = x86.ADECQ
case x86.ANEGB,
x86.ANEGW:
p.As = x86.ANEGQ
case x86.ANOTB,
x86.ANOTW:
p.As = x86.ANOTQ
}
if regtyp(&p.From) || p.From.Type == obj.TYPE_CONST {
// move or artihmetic into partial register.
// from another register or constant can be movl.
// we don't switch to 64-bit arithmetic if it can
// change how the carry bit is set (and the carry bit is needed).
switch p.As {
case x86.AMOVB,
x86.AMOVW:
p.As = x86.AMOVQ
case x86.AADDB,
x86.AADDW:
if !needc(p.Link) {
p.As = x86.AADDQ
}
case x86.ASUBB,
x86.ASUBW:
if !needc(p.Link) {
p.As = x86.ASUBQ
}
case x86.AMULB,
x86.AMULW:
p.As = x86.AMULQ
case x86.AIMULB,
x86.AIMULW:
p.As = x86.AIMULQ
case x86.AANDB,
x86.AANDW:
p.As = x86.AANDQ
case x86.AORB,
x86.AORW:
p.As = x86.AORQ
case x86.AXORB,
x86.AXORW:
p.As = x86.AXORQ
case x86.ASHLB,
x86.ASHLW:
p.As = x86.ASHLQ
}
} else if p.From.Type != obj.TYPE_REG {
// explicit zero extension, but don't
// do that if source is a byte register
// (only AH can occur and it's forbidden).
switch p.As {
case x86.AMOVB:
p.As = x86.AMOVBQZX
case x86.AMOVW:
p.As = x86.AMOVWQZX
}
}
}
}
}
func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32) *obj.Prog {
// MOVV g_stackguard(g), R1
p = obj.Appendp(ctxt, p)
p.As = AMOVV
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGG
p.From.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0
if ctxt.Cursym.Cfunc != 0 {
p.From.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1
}
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
var q *obj.Prog
if framesize <= obj.StackSmall {
// small stack: SP < stackguard
// AGTU SP, stackguard, R1
p = obj.Appendp(ctxt, p)
p.As = ASGTU
p.From.Type = obj.TYPE_REG
p.From.Reg = REGSP
p.Reg = REG_R1
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
} else if framesize <= obj.StackBig {
// large stack: SP-framesize < stackguard-StackSmall
// ADDV $-framesize, SP, R2
// SGTU R2, stackguard, R1
p = obj.Appendp(ctxt, p)
p.As = AADDV
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(-framesize)
p.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = ASGTU
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R2
p.Reg = REG_R1
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
} else {
// Such a large stack we need to protect against wraparound.
// If SP is close to zero:
// SP-stackguard+StackGuard <= framesize + (StackGuard-StackSmall)
// The +StackGuard on both sides is required to keep the left side positive:
// SP is allowed to be slightly below stackguard. See stack.h.
//
// Preemption sets stackguard to StackPreempt, a very large value.
// That breaks the math above, so we have to check for that explicitly.
// // stackguard is R1
// MOVV $StackPreempt, R2
// BEQ R1, R2, label-of-call-to-morestack
// ADDV $StackGuard, SP, R2
// SUBVU R1, R2
// MOVV $(framesize+(StackGuard-StackSmall)), R1
// SGTU R2, R1, R1
p = obj.Appendp(ctxt, p)
p.As = AMOVV
p.From.Type = obj.TYPE_CONST
p.From.Offset = obj.StackPreempt
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
q = p
p.As = ABEQ
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.Reg = REG_R2
p.To.Type = obj.TYPE_BRANCH
p.Mark |= BRANCH
p = obj.Appendp(ctxt, p)
p.As = AADDV
p.From.Type = obj.TYPE_CONST
p.From.Offset = obj.StackGuard
p.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = ASUBVU
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = AMOVV
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(framesize) + obj.StackGuard - obj.StackSmall
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
p = obj.Appendp(ctxt, p)
p.As = ASGTU
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R2
p.Reg = REG_R1
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
}
// q1: BNE R1, done
p = obj.Appendp(ctxt, p)
q1 := p
p.As = ABNE
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.To.Type = obj.TYPE_BRANCH
p.Mark |= BRANCH
// MOVV LINK, R3
p = obj.Appendp(ctxt, p)
p.As = AMOVV
p.From.Type = obj.TYPE_REG
p.From.Reg = REGLINK
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R3
if q != nil {
q.Pcond = p
p.Mark |= LABEL
}
// JAL runtime.morestack(SB)
p = obj.Appendp(ctxt, p)
p.As = AJAL
p.To.Type = obj.TYPE_BRANCH
if ctxt.Cursym.Cfunc != 0 {
p.To.Sym = obj.Linklookup(ctxt, "runtime.morestackc", 0)
} else if ctxt.Cursym.Text.From3.Offset&obj.NEEDCTXT == 0 {
p.To.Sym = obj.Linklookup(ctxt, "runtime.morestack_noctxt", 0)
} else {
p.To.Sym = obj.Linklookup(ctxt, "runtime.morestack", 0)
}
p.Mark |= BRANCH
// JMP start
p = obj.Appendp(ctxt, p)
p.As = AJMP
p.To.Type = obj.TYPE_BRANCH
p.Pcond = ctxt.Cursym.Text.Link
p.Mark |= BRANCH
// placeholder for q1's jump target
p = obj.Appendp(ctxt, p)
p.As = obj.ANOP // zero-width place holder
q1.Pcond = p
return p
}
func progedit(ctxt *obj.Link, p *obj.Prog) {
p.From.Class = 0
p.To.Class = 0
// Rewrite JMP/JAL to symbol as TYPE_BRANCH.
switch p.As {
case AJMP,
AJAL,
ARET,
obj.ADUFFZERO,
obj.ADUFFCOPY:
if p.To.Sym != nil {
p.To.Type = obj.TYPE_BRANCH
}
}
// Rewrite float constants to values stored in memory.
switch p.As {
case AMOVF:
if p.From.Type == obj.TYPE_FCONST {
f32 := float32(p.From.Val.(float64))
i32 := math.Float32bits(f32)
literal := fmt.Sprintf("$f32.%08x", i32)
s := obj.Linklookup(ctxt, literal, 0)
s.Size = 4
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
case AMOVD:
if p.From.Type == obj.TYPE_FCONST {
i64 := math.Float64bits(p.From.Val.(float64))
literal := fmt.Sprintf("$f64.%016x", i64)
s := obj.Linklookup(ctxt, literal, 0)
s.Size = 8
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
// Put >32-bit constants in memory and load them
case AMOVV:
if p.From.Type == obj.TYPE_CONST && p.From.Name == obj.NAME_NONE && p.From.Reg == 0 && int64(int32(p.From.Offset)) != p.From.Offset {
literal := fmt.Sprintf("$i64.%016x", uint64(p.From.Offset))
s := obj.Linklookup(ctxt, literal, 0)
s.Size = 8
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
}
// Rewrite SUB constants into ADD.
switch p.As {
case ASUB:
if p.From.Type == obj.TYPE_CONST {
p.From.Offset = -p.From.Offset
p.As = AADD
}
case ASUBU:
if p.From.Type == obj.TYPE_CONST {
p.From.Offset = -p.From.Offset
p.As = AADDU
}
case ASUBV:
if p.From.Type == obj.TYPE_CONST {
p.From.Offset = -p.From.Offset
p.As = AADDV
}
case ASUBVU:
if p.From.Type == obj.TYPE_CONST {
p.From.Offset = -p.From.Offset
p.As = AADDVU
}
}
}
func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
// TODO(minux): add morestack short-cuts with small fixed frame-size.
ctxt.Cursym = cursym
// a switch for enabling/disabling instruction scheduling
nosched := true
if cursym.Text == nil || cursym.Text.Link == nil {
return
}
p := cursym.Text
textstksiz := p.To.Offset
cursym.Args = p.To.Val.(int32)
cursym.Locals = int32(textstksiz)
/*
* find leaf subroutines
* strip NOPs
* expand RET
* expand BECOME pseudo
*/
if ctxt.Debugvlog != 0 {
fmt.Fprintf(ctxt.Bso, "%5.2f noops\n", obj.Cputime())
}
ctxt.Bso.Flush()
var q *obj.Prog
var q1 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
switch p.As {
/* too hard, just leave alone */
case obj.ATEXT:
q = p
p.Mark |= LABEL | LEAF | SYNC
if p.Link != nil {
p.Link.Mark |= LABEL
}
/* too hard, just leave alone */
case AMOVW,
AMOVV:
q = p
if p.To.Type == obj.TYPE_REG && p.To.Reg >= REG_SPECIAL {
p.Mark |= LABEL | SYNC
break
}
if p.From.Type == obj.TYPE_REG && p.From.Reg >= REG_SPECIAL {
p.Mark |= LABEL | SYNC
}
/* too hard, just leave alone */
case ASYSCALL,
AWORD,
ATLBWR,
ATLBWI,
ATLBP,
ATLBR:
q = p
p.Mark |= LABEL | SYNC
case ANOR:
q = p
if p.To.Type == obj.TYPE_REG {
if p.To.Reg == REGZERO {
p.Mark |= LABEL | SYNC
}
}
case ABGEZAL,
ABLTZAL,
AJAL,
obj.ADUFFZERO,
obj.ADUFFCOPY:
cursym.Text.Mark &^= LEAF
fallthrough
case AJMP,
ABEQ,
ABGEZ,
ABGTZ,
ABLEZ,
ABLTZ,
ABNE,
ABFPT, ABFPF:
if p.As == ABFPT || p.As == ABFPF {
// We don't treat ABFPT and ABFPF as branches here,
// so that we will always fill nop (0x0) in their
// delay slot during assembly.
// This is to workaround a kernel FPU emulator bug
// where it uses the user stack to simulate the
// instruction in the delay slot if it's not 0x0,
// and somehow that leads to SIGSEGV when the kernel
// jump to the stack.
p.Mark |= SYNC
} else {
p.Mark |= BRANCH
}
q = p
q1 = p.Pcond
if q1 != nil {
for q1.As == obj.ANOP {
q1 = q1.Link
p.Pcond = q1
}
if q1.Mark&LEAF == 0 {
q1.Mark |= LABEL
}
}
//else {
// p.Mark |= LABEL
//}
q1 = p.Link
if q1 != nil {
q1.Mark |= LABEL
}
continue
case ARET:
q = p
if p.Link != nil {
p.Link.Mark |= LABEL
}
continue
case obj.ANOP:
q1 = p.Link
q.Link = q1 /* q is non-nop */
q1.Mark |= p.Mark
continue
default:
q = p
continue
}
}
autosize := int32(0)
var o int
var p1 *obj.Prog
var p2 *obj.Prog
for p := cursym.Text; p != nil; p = p.Link {
o = int(p.As)
switch o {
case obj.ATEXT:
autosize = int32(textstksiz + 8)
if (p.Mark&LEAF != 0) && autosize <= 8 {
autosize = 0
} else if autosize&4 != 0 {
autosize += 4
}
p.To.Offset = int64(autosize) - 8
if p.From3.Offset&obj.NOSPLIT == 0 {
p = stacksplit(ctxt, p, autosize) // emit split check
}
q = p
if autosize != 0 {
q = obj.Appendp(ctxt, p)
q.As = AADDV
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(-autosize)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Spadj = +autosize
} else if cursym.Text.Mark&LEAF == 0 {
if cursym.Text.From3.Offset&obj.NOSPLIT != 0 {
if ctxt.Debugvlog != 0 {
fmt.Fprintf(ctxt.Bso, "save suppressed in: %s\n", cursym.Name)
ctxt.Bso.Flush()
}
cursym.Text.Mark |= LEAF
}
}
if cursym.Text.Mark&LEAF != 0 {
cursym.Leaf = 1
break
}
q = obj.Appendp(ctxt, q)
q.As = AMOVV
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_REG
q.From.Reg = REGLINK
q.To.Type = obj.TYPE_MEM
q.To.Offset = int64(0)
q.To.Reg = REGSP
if cursym.Text.From3.Offset&obj.WRAPPER != 0 {
// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
//
// MOVV g_panic(g), R1
// BEQ R1, end
// MOVV panic_argp(R1), R2
// ADDV $(autosize+8), R29, R3
// BNE R2, R3, end
// ADDV $8, R29, R2
// MOVV R2, panic_argp(R1)
// end:
// NOP
//
// The NOP is needed to give the jumps somewhere to land.
// It is a liblink NOP, not an mips NOP: it encodes to 0 instruction bytes.
q = obj.Appendp(ctxt, q)
q.As = AMOVV
q.From.Type = obj.TYPE_MEM
q.From.Reg = REGG
q.From.Offset = 4 * int64(ctxt.Arch.Ptrsize) // G.panic
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R1
q = obj.Appendp(ctxt, q)
q.As = ABEQ
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R1
q.To.Type = obj.TYPE_BRANCH
q.Mark |= BRANCH
p1 = q
q = obj.Appendp(ctxt, q)
q.As = AMOVV
q.From.Type = obj.TYPE_MEM
q.From.Reg = REG_R1
q.From.Offset = 0 // Panic.argp
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R2
q = obj.Appendp(ctxt, q)
q.As = AADDV
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(autosize) + 8
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R3
q = obj.Appendp(ctxt, q)
q.As = ABNE
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R2
q.Reg = REG_R3
q.To.Type = obj.TYPE_BRANCH
q.Mark |= BRANCH
p2 = q
q = obj.Appendp(ctxt, q)
q.As = AADDV
q.From.Type = obj.TYPE_CONST
q.From.Offset = 8
q.Reg = REGSP
q.To.Type = obj.TYPE_REG
q.To.Reg = REG_R2
q = obj.Appendp(ctxt, q)
q.As = AMOVV
q.From.Type = obj.TYPE_REG
q.From.Reg = REG_R2
q.To.Type = obj.TYPE_MEM
q.To.Reg = REG_R1
q.To.Offset = 0 // Panic.argp
q = obj.Appendp(ctxt, q)
q.As = obj.ANOP
p1.Pcond = q
p2.Pcond = q
}
case ARET:
if p.From.Type == obj.TYPE_CONST {
ctxt.Diag("using BECOME (%v) is not supported!", p)
break
}
if p.To.Sym != nil { // retjmp
p.As = AJMP
p.To.Type = obj.TYPE_BRANCH
break
}
if cursym.Text.Mark&LEAF != 0 {
if autosize == 0 {
p.As = AJMP
p.From = obj.Addr{}
p.To.Type = obj.TYPE_MEM
p.To.Offset = 0
p.To.Reg = REGLINK
p.Mark |= BRANCH
break
}
p.As = AADDV
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(autosize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REGSP
p.Spadj = -autosize
q = ctxt.NewProg()
q.As = AJMP
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_MEM
q.To.Offset = 0
q.To.Reg = REGLINK
q.Mark |= BRANCH
q.Spadj = +autosize
q.Link = p.Link
p.Link = q
break
}
p.As = AMOVV
p.From.Type = obj.TYPE_MEM
p.From.Offset = 0
p.From.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R4
if false {
// Debug bad returns
q = ctxt.NewProg()
q.As = AMOVV
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_MEM
q.From.Offset = 0
q.From.Reg = REG_R4
q.To.Type = obj.TYPE_REG
q.To.Reg = REGTMP
q.Link = p.Link
p.Link = q
p = q
}
if autosize != 0 {
q = ctxt.NewProg()
q.As = AADDV
q.Lineno = p.Lineno
q.From.Type = obj.TYPE_CONST
q.From.Offset = int64(autosize)
q.To.Type = obj.TYPE_REG
q.To.Reg = REGSP
q.Spadj = -autosize
q.Link = p.Link
p.Link = q
}
q1 = ctxt.NewProg()
q1.As = AJMP
q1.Lineno = p.Lineno
q1.To.Type = obj.TYPE_MEM
q1.To.Offset = 0
q1.To.Reg = REG_R4
q1.Mark |= BRANCH
q1.Spadj = +autosize
q1.Link = q.Link
q.Link = q1
case AADDV,
AADDVU:
if p.To.Type == obj.TYPE_REG && p.To.Reg == REGSP && p.From.Type == obj.TYPE_CONST {
p.Spadj = int32(-p.From.Offset)
}
}
}
if nosched {
// if we don't do instruction scheduling, simply add
// NOP after each branch instruction.
for p = cursym.Text; p != nil; p = p.Link {
if p.Mark&BRANCH != 0 {
addnop(ctxt, p)
}
}
return
}
// instruction scheduling
q = nil // p - 1
q1 = cursym.Text // top of block
o = 0 // count of instructions
for p = cursym.Text; p != nil; p = p1 {
p1 = p.Link
o++
if p.Mark&NOSCHED != 0 {
if q1 != p {
sched(ctxt, q1, q)
}
for ; p != nil; p = p.Link {
if p.Mark&NOSCHED == 0 {
break
}
q = p
}
p1 = p
q1 = p
o = 0
continue
}
if p.Mark&(LABEL|SYNC) != 0 {
if q1 != p {
sched(ctxt, q1, q)
}
q1 = p
o = 1
}
if p.Mark&(BRANCH|SYNC) != 0 {
sched(ctxt, q1, p)
q1 = p1
o = 0
}
if o >= NSCHED {
sched(ctxt, q1, p)
q1 = p1
o = 0
}
q = p
}
}
func preprocess(ctxt *obj.Link, cursym *obj.LSym) {
if ctxt.Headtype == obj.Hplan9 && ctxt.Plan9privates == nil {
ctxt.Plan9privates = obj.Linklookup(ctxt, "_privates", 0)
}
ctxt.Cursym = cursym
if cursym.Text == nil || cursym.Text.Link == nil {
return
}
p := cursym.Text
autoffset := int32(p.To.Offset)
if autoffset < 0 {
autoffset = 0
}
var bpsize int
if p.Mode == 64 && obj.Framepointer_enabled != 0 && autoffset > 0 {
// Make room for to save a base pointer. If autoffset == 0,
// this might do something special like a tail jump to
// another function, so in that case we omit this.
bpsize = ctxt.Arch.Ptrsize
autoffset += int32(bpsize)
p.To.Offset += int64(bpsize)
} else {
bpsize = 0
}
textarg := int64(p.To.Val.(int32))
cursym.Args = int32(textarg)
cursym.Locals = int32(p.To.Offset)
// TODO(rsc): Remove.
if p.Mode == 32 && cursym.Locals < 0 {
cursym.Locals = 0
}
// TODO(rsc): Remove 'p.Mode == 64 &&'.
if p.Mode == 64 && autoffset < obj.StackSmall && p.From3Offset()&obj.NOSPLIT == 0 {
for q := p; q != nil; q = q.Link {
if q.As == obj.ACALL {
goto noleaf
}
if (q.As == obj.ADUFFCOPY || q.As == obj.ADUFFZERO) && autoffset >= obj.StackSmall-8 {
goto noleaf
}
}
p.From3.Offset |= obj.NOSPLIT
noleaf:
}
if p.From3Offset()&obj.NOSPLIT == 0 || p.From3Offset()&obj.WRAPPER != 0 {
p = obj.Appendp(ctxt, p)
p = load_g_cx(ctxt, p) // load g into CX
}
if cursym.Text.From3Offset()&obj.NOSPLIT == 0 {
p = stacksplit(ctxt, p, autoffset, int32(textarg)) // emit split check
}
if autoffset != 0 {
if autoffset%int32(ctxt.Arch.Regsize) != 0 {
ctxt.Diag("unaligned stack size %d", autoffset)
}
p = obj.Appendp(ctxt, p)
p.As = AADJSP
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(autoffset)
p.Spadj = autoffset
} else {
// zero-byte stack adjustment.
// Insert a fake non-zero adjustment so that stkcheck can
// recognize the end of the stack-splitting prolog.
p = obj.Appendp(ctxt, p)
p.As = obj.ANOP
p.Spadj = int32(-ctxt.Arch.Ptrsize)
p = obj.Appendp(ctxt, p)
p.As = obj.ANOP
p.Spadj = int32(ctxt.Arch.Ptrsize)
}
deltasp := autoffset
if bpsize > 0 {
// Save caller's BP
p = obj.Appendp(ctxt, p)
p.As = AMOVQ
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_BP
p.To.Type = obj.TYPE_MEM
p.To.Reg = REG_SP
p.To.Scale = 1
p.To.Offset = int64(autoffset) - int64(bpsize)
// Move current frame to BP
p = obj.Appendp(ctxt, p)
p.As = ALEAQ
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_SP
p.From.Scale = 1
p.From.Offset = int64(autoffset) - int64(bpsize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_BP
}
if cursym.Text.From3Offset()&obj.WRAPPER != 0 {
// if(g->panic != nil && g->panic->argp == FP) g->panic->argp = bottom-of-frame
//
// MOVQ g_panic(CX), BX
// TESTQ BX, BX
// JEQ end
// LEAQ (autoffset+8)(SP), DI
// CMPQ panic_argp(BX), DI
// JNE end
// MOVQ SP, panic_argp(BX)
// end:
// NOP
//
// The NOP is needed to give the jumps somewhere to land.
// It is a liblink NOP, not an x86 NOP: it encodes to 0 instruction bytes.
p = obj.Appendp(ctxt, p)
p.As = AMOVQ
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_CX
p.From.Offset = 4 * int64(ctxt.Arch.Ptrsize) // G.panic
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_BX
if ctxt.Headtype == obj.Hnacl && p.Mode == 64 {
p.As = AMOVL
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_R15
p.From.Scale = 1
p.From.Index = REG_CX
}
if p.Mode == 32 {
p.As = AMOVL
}
p = obj.Appendp(ctxt, p)
p.As = ATESTQ
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_BX
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_BX
if ctxt.Headtype == obj.Hnacl || p.Mode == 32 {
p.As = ATESTL
}
p = obj.Appendp(ctxt, p)
p.As = AJEQ
p.To.Type = obj.TYPE_BRANCH
p1 := p
p = obj.Appendp(ctxt, p)
p.As = ALEAQ
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_SP
p.From.Offset = int64(autoffset) + int64(ctxt.Arch.Regsize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_DI
if ctxt.Headtype == obj.Hnacl || p.Mode == 32 {
p.As = ALEAL
}
p = obj.Appendp(ctxt, p)
p.As = ACMPQ
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_BX
p.From.Offset = 0 // Panic.argp
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_DI
if ctxt.Headtype == obj.Hnacl && p.Mode == 64 {
p.As = ACMPL
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_R15
p.From.Scale = 1
p.From.Index = REG_BX
}
if p.Mode == 32 {
p.As = ACMPL
}
p = obj.Appendp(ctxt, p)
p.As = AJNE
p.To.Type = obj.TYPE_BRANCH
p2 := p
p = obj.Appendp(ctxt, p)
p.As = AMOVQ
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_SP
p.To.Type = obj.TYPE_MEM
p.To.Reg = REG_BX
p.To.Offset = 0 // Panic.argp
if ctxt.Headtype == obj.Hnacl && p.Mode == 64 {
p.As = AMOVL
p.To.Type = obj.TYPE_MEM
p.To.Reg = REG_R15
p.To.Scale = 1
p.To.Index = REG_BX
}
if p.Mode == 32 {
p.As = AMOVL
}
p = obj.Appendp(ctxt, p)
p.As = obj.ANOP
p1.Pcond = p
p2.Pcond = p
}
var a int
var pcsize int
for ; p != nil; p = p.Link {
pcsize = int(p.Mode) / 8
a = int(p.From.Name)
if a == obj.NAME_AUTO {
p.From.Offset += int64(deltasp) - int64(bpsize)
}
if a == obj.NAME_PARAM {
p.From.Offset += int64(deltasp) + int64(pcsize)
}
if p.From3 != nil {
a = int(p.From3.Name)
if a == obj.NAME_AUTO {
p.From3.Offset += int64(deltasp) - int64(bpsize)
}
if a == obj.NAME_PARAM {
p.From3.Offset += int64(deltasp) + int64(pcsize)
}
}
a = int(p.To.Name)
if a == obj.NAME_AUTO {
p.To.Offset += int64(deltasp) - int64(bpsize)
}
if a == obj.NAME_PARAM {
p.To.Offset += int64(deltasp) + int64(pcsize)
}
switch p.As {
default:
continue
case APUSHL, APUSHFL:
deltasp += 4
p.Spadj = 4
continue
case APUSHQ, APUSHFQ:
deltasp += 8
p.Spadj = 8
continue
case APUSHW, APUSHFW:
deltasp += 2
p.Spadj = 2
continue
case APOPL, APOPFL:
deltasp -= 4
p.Spadj = -4
continue
case APOPQ, APOPFQ:
deltasp -= 8
p.Spadj = -8
continue
case APOPW, APOPFW:
deltasp -= 2
p.Spadj = -2
continue
case obj.ARET:
break
}
if autoffset != deltasp {
ctxt.Diag("unbalanced PUSH/POP")
}
if autoffset != 0 {
if bpsize > 0 {
// Restore caller's BP
p.As = AMOVQ
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_SP
p.From.Scale = 1
p.From.Offset = int64(autoffset) - int64(bpsize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_BP
p = obj.Appendp(ctxt, p)
}
p.As = AADJSP
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(-autoffset)
p.Spadj = -autoffset
p = obj.Appendp(ctxt, p)
p.As = obj.ARET
// If there are instructions following
// this ARET, they come from a branch
// with the same stackframe, so undo
// the cleanup.
p.Spadj = +autoffset
}
if p.To.Sym != nil { // retjmp
p.As = obj.AJMP
}
}
}
func indir_cx(ctxt *obj.Link, p *obj.Prog, a *obj.Addr) {
if ctxt.Headtype == obj.Hnacl && p.Mode == 64 {
a.Type = obj.TYPE_MEM
a.Reg = REG_R15
a.Index = REG_CX
a.Scale = 1
return
}
a.Type = obj.TYPE_MEM
a.Reg = REG_CX
}
// Append code to p to load g into cx.
// Overwrites p with the first instruction (no first appendp).
// Overwriting p is unusual but it lets use this in both the
// prologue (caller must call appendp first) and in the epilogue.
// Returns last new instruction.
func load_g_cx(ctxt *obj.Link, p *obj.Prog) *obj.Prog {
p.As = AMOVQ
if ctxt.Arch.Ptrsize == 4 {
p.As = AMOVL
}
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_TLS
p.From.Offset = 0
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_CX
next := p.Link
progedit(ctxt, p)
for p.Link != next {
p = p.Link
}
if p.From.Index == REG_TLS {
p.From.Scale = 2
}
return p
}
// Append code to p to check for stack split.
// Appends to (does not overwrite) p.
// Assumes g is in CX.
// Returns last new instruction.
func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32, textarg int32) *obj.Prog {
cmp := ACMPQ
lea := ALEAQ
mov := AMOVQ
sub := ASUBQ
if ctxt.Headtype == obj.Hnacl || p.Mode == 32 {
cmp = ACMPL
lea = ALEAL
mov = AMOVL
sub = ASUBL
}
var q1 *obj.Prog
if framesize <= obj.StackSmall {
// small stack: SP <= stackguard
// CMPQ SP, stackguard
p = obj.Appendp(ctxt, p)
p.As = int16(cmp)
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_SP
indir_cx(ctxt, p, &p.To)
p.To.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0
if ctxt.Cursym.Cfunc != 0 {
p.To.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1
}
} else if framesize <= obj.StackBig {
// large stack: SP-framesize <= stackguard-StackSmall
// LEAQ -xxx(SP), AX
// CMPQ AX, stackguard
p = obj.Appendp(ctxt, p)
p.As = int16(lea)
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_SP
p.From.Offset = -(int64(framesize) - obj.StackSmall)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_AX
p = obj.Appendp(ctxt, p)
p.As = int16(cmp)
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_AX
indir_cx(ctxt, p, &p.To)
p.To.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0
if ctxt.Cursym.Cfunc != 0 {
p.To.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1
}
} else {
// Such a large stack we need to protect against wraparound.
// If SP is close to zero:
// SP-stackguard+StackGuard <= framesize + (StackGuard-StackSmall)
// The +StackGuard on both sides is required to keep the left side positive:
// SP is allowed to be slightly below stackguard. See stack.h.
//
// Preemption sets stackguard to StackPreempt, a very large value.
// That breaks the math above, so we have to check for that explicitly.
// MOVQ stackguard, CX
// CMPQ CX, $StackPreempt
// JEQ label-of-call-to-morestack
// LEAQ StackGuard(SP), AX
// SUBQ CX, AX
// CMPQ AX, $(framesize+(StackGuard-StackSmall))
p = obj.Appendp(ctxt, p)
p.As = int16(mov)
indir_cx(ctxt, p, &p.From)
p.From.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0
if ctxt.Cursym.Cfunc != 0 {
p.From.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1
}
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_SI
p = obj.Appendp(ctxt, p)
p.As = int16(cmp)
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_SI
p.To.Type = obj.TYPE_CONST
p.To.Offset = obj.StackPreempt
if p.Mode == 32 {
p.To.Offset = int64(uint32(obj.StackPreempt & (1<<32 - 1)))
}
p = obj.Appendp(ctxt, p)
p.As = AJEQ
p.To.Type = obj.TYPE_BRANCH
q1 = p
p = obj.Appendp(ctxt, p)
p.As = int16(lea)
p.From.Type = obj.TYPE_MEM
p.From.Reg = REG_SP
p.From.Offset = obj.StackGuard
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_AX
p = obj.Appendp(ctxt, p)
p.As = int16(sub)
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_SI
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_AX
p = obj.Appendp(ctxt, p)
p.As = int16(cmp)
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_AX
p.To.Type = obj.TYPE_CONST
p.To.Offset = int64(framesize) + (obj.StackGuard - obj.StackSmall)
}
// common
jls := obj.Appendp(ctxt, p)
jls.As = AJLS
jls.To.Type = obj.TYPE_BRANCH
var last *obj.Prog
for last = ctxt.Cursym.Text; last.Link != nil; last = last.Link {
}
spfix := obj.Appendp(ctxt, last)
spfix.As = obj.ANOP
spfix.Spadj = -framesize
call := obj.Appendp(ctxt, spfix)
call.Lineno = ctxt.Cursym.Text.Lineno
call.Mode = ctxt.Cursym.Text.Mode
call.As = obj.ACALL
call.To.Type = obj.TYPE_BRANCH
morestack := "runtime.morestack"
switch {
case ctxt.Cursym.Cfunc != 0:
morestack = "runtime.morestackc"
case ctxt.Cursym.Text.From3Offset()&obj.NEEDCTXT == 0:
morestack = "runtime.morestack_noctxt"
}
call.To.Sym = obj.Linklookup(ctxt, morestack, 0)
jmp := obj.Appendp(ctxt, call)
jmp.As = obj.AJMP
jmp.To.Type = obj.TYPE_BRANCH
jmp.Pcond = ctxt.Cursym.Text.Link
jmp.Spadj = +framesize
jls.Pcond = call
if q1 != nil {
q1.Pcond = call
}
return jls
}
func follow(ctxt *obj.Link, s *obj.LSym) {
ctxt.Cursym = s
firstp := ctxt.NewProg()
lastp := firstp
xfol(ctxt, s.Text, &lastp)
lastp.Link = nil
s.Text = firstp.Link
}
func nofollow(a int) bool {
switch a {
case obj.AJMP,
obj.ARET,
AIRETL,
AIRETQ,
AIRETW,
ARETFL,
ARETFQ,
ARETFW,
obj.AUNDEF:
return true
}
return false
}
func pushpop(a int) bool {
switch a {
case APUSHL,
APUSHFL,
APUSHQ,
APUSHFQ,
APUSHW,
APUSHFW,
APOPL,
APOPFL,
APOPQ,
APOPFQ,
APOPW,
APOPFW:
return true
}
return false
}
func relinv(a int16) int16 {
switch a {
case AJEQ:
return AJNE
case AJNE:
return AJEQ
case AJLE:
return AJGT
case AJLS:
return AJHI
case AJLT:
return AJGE
case AJMI:
return AJPL
case AJGE:
return AJLT
case AJPL:
return AJMI
case AJGT:
return AJLE
case AJHI:
return AJLS
case AJCS:
return AJCC
case AJCC:
return AJCS
case AJPS:
return AJPC
case AJPC:
return AJPS
case AJOS:
return AJOC
case AJOC:
return AJOS
}
log.Fatalf("unknown relation: %s", obj.Aconv(int(a)))
return 0
}
func xfol(ctxt *obj.Link, p *obj.Prog, last **obj.Prog) {
var q *obj.Prog
var i int
var a int
loop:
if p == nil {
return
}
if p.As == obj.AJMP {
q = p.Pcond
if q != nil && q.As != obj.ATEXT {
/* mark instruction as done and continue layout at target of jump */
p.Mark = 1
p = q
if p.Mark == 0 {
goto loop
}
}
}
if p.Mark != 0 {
/*
* p goes here, but already used it elsewhere.
* copy up to 4 instructions or else branch to other copy.
*/
i = 0
q = p
for ; i < 4; i, q = i+1, q.Link {
if q == nil {
break
}
if q == *last {
break
}
a = int(q.As)
if a == obj.ANOP {
i--
continue
}
if nofollow(a) || pushpop(a) {
break // NOTE(rsc): arm does goto copy
}
if q.Pcond == nil || q.Pcond.Mark != 0 {
continue
}
if a == obj.ACALL || a == ALOOP {
continue
}
for {
if p.As == obj.ANOP {
p = p.Link
continue
}
q = obj.Copyp(ctxt, p)
p = p.Link
q.Mark = 1
(*last).Link = q
*last = q
if int(q.As) != a || q.Pcond == nil || q.Pcond.Mark != 0 {
continue
}
q.As = relinv(q.As)
p = q.Pcond
q.Pcond = q.Link
q.Link = p
xfol(ctxt, q.Link, last)
p = q.Link
if p.Mark != 0 {
return
}
goto loop
/* */
}
}
q = ctxt.NewProg()
q.As = obj.AJMP
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_BRANCH
q.To.Offset = p.Pc
q.Pcond = p
p = q
}
/* emit p */
p.Mark = 1
(*last).Link = p
*last = p
a = int(p.As)
/* continue loop with what comes after p */
if nofollow(a) {
return
}
if p.Pcond != nil && a != obj.ACALL {
/*
* some kind of conditional branch.
* recurse to follow one path.
* continue loop on the other.
*/
q = obj.Brchain(ctxt, p.Pcond)
if q != nil {
p.Pcond = q
}
q = obj.Brchain(ctxt, p.Link)
if q != nil {
p.Link = q
}
if p.From.Type == obj.TYPE_CONST {
if p.From.Offset == 1 {
/*
* expect conditional jump to be taken.
* rewrite so that's the fall-through case.
*/
p.As = relinv(int16(a))
q = p.Link
p.Link = p.Pcond
p.Pcond = q
}
} else {
q = p.Link
if q.Mark != 0 {
if a != ALOOP {
p.As = relinv(int16(a))
p.Link = p.Pcond
p.Pcond = q
}
}
}
xfol(ctxt, p.Link, last)
if p.Pcond.Mark != 0 {
return
}
p = p.Pcond
goto loop
}
p = p.Link
goto loop
}
var unaryDst = map[int]bool{
ABSWAPL: true,
ABSWAPQ: true,
ACMPXCHG8B: true,
ADECB: true,
ADECL: true,
ADECQ: true,
ADECW: true,
AINCB: true,
AINCL: true,
AINCQ: true,
AINCW: true,
ANEGB: true,
ANEGL: true,
ANEGQ: true,
ANEGW: true,
ANOTB: true,
ANOTL: true,
ANOTQ: true,
ANOTW: true,
APOPL: true,
APOPQ: true,
APOPW: true,
ASETCC: true,
ASETCS: true,
ASETEQ: true,
ASETGE: true,
ASETGT: true,
ASETHI: true,
ASETLE: true,
ASETLS: true,
ASETLT: true,
ASETMI: true,
ASETNE: true,
ASETOC: true,
ASETOS: true,
ASETPC: true,
ASETPL: true,
ASETPS: true,
AFFREE: true,
AFLDENV: true,
AFSAVE: true,
AFSTCW: true,
AFSTENV: true,
AFSTSW: true,
AFXSAVE: true,
AFXSAVE64: true,
ASTMXCSR: true,
}
var Linkamd64 = obj.LinkArch{
ByteOrder: binary.LittleEndian,
Name: "amd64",
Thechar: '6',
Preprocess: preprocess,
Assemble: span6,
Follow: follow,
Progedit: progedit,
UnaryDst: unaryDst,
Minlc: 1,
Ptrsize: 8,
Regsize: 8,
}
var Linkamd64p32 = obj.LinkArch{
ByteOrder: binary.LittleEndian,
Name: "amd64p32",
Thechar: '6',
Preprocess: preprocess,
Assemble: span6,
Follow: follow,
Progedit: progedit,
UnaryDst: unaryDst,
Minlc: 1,
Ptrsize: 4,
Regsize: 8,
}
var Link386 = obj.LinkArch{
ByteOrder: binary.LittleEndian,
Name: "386",
Thechar: '8',
Preprocess: preprocess,
Assemble: span6,
Follow: follow,
Progedit: progedit,
UnaryDst: unaryDst,
Minlc: 1,
Ptrsize: 4,
Regsize: 4,
}
// 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 progedit(ctxt *obj.Link, p *obj.Prog) {
// Maintain information about code generation mode.
if ctxt.Mode == 0 {
ctxt.Mode = ctxt.Arch.Regsize * 8
}
p.Mode = int8(ctxt.Mode)
switch p.As {
case AMODE:
if p.From.Type == obj.TYPE_CONST || (p.From.Type == obj.TYPE_MEM && p.From.Reg == REG_NONE) {
switch int(p.From.Offset) {
case 16, 32, 64:
ctxt.Mode = int(p.From.Offset)
}
}
obj.Nopout(p)
}
// Thread-local storage references use the TLS pseudo-register.
// As a register, TLS refers to the thread-local storage base, and it
// can only be loaded into another register:
//
// MOVQ TLS, AX
//
// An offset from the thread-local storage base is written off(reg)(TLS*1).
// Semantically it is off(reg), but the (TLS*1) annotation marks this as
// indexing from the loaded TLS base. This emits a relocation so that
// if the linker needs to adjust the offset, it can. For example:
//
// MOVQ TLS, AX
// MOVQ 0(AX)(TLS*1), CX // load g into CX
//
// On systems that support direct access to the TLS memory, this
// pair of instructions can be reduced to a direct TLS memory reference:
//
// MOVQ 0(TLS), CX // load g into CX
//
// The 2-instruction and 1-instruction forms correspond to the two code
// sequences for loading a TLS variable in the local exec model given in "ELF
// Handling For Thread-Local Storage".
//
// We apply this rewrite on systems that support the 1-instruction form.
// The decision is made using only the operating system and the -shared flag,
// not the link mode. If some link modes on a particular operating system
// require the 2-instruction form, then all builds for that operating system
// will use the 2-instruction form, so that the link mode decision can be
// delayed to link time.
//
// In this way, all supported systems use identical instructions to
// access TLS, and they are rewritten appropriately first here in
// liblink and then finally using relocations in the linker.
//
// When -shared is passed, we leave the code in the 2-instruction form but
// assemble (and relocate) them in different ways to generate the initial
// exec code sequence. It's a bit of a fluke that this is possible without
// rewriting the instructions more comprehensively, and it only does because
// we only support a single TLS variable (g).
if canuse1insntls(ctxt) {
// Reduce 2-instruction sequence to 1-instruction sequence.
// Sequences like
// MOVQ TLS, BX
// ... off(BX)(TLS*1) ...
// become
// NOP
// ... off(TLS) ...
//
// TODO(rsc): Remove the Hsolaris special case. It exists only to
// guarantee we are producing byte-identical binaries as before this code.
// But it should be unnecessary.
if (p.As == AMOVQ || p.As == AMOVL) && p.From.Type == obj.TYPE_REG && p.From.Reg == REG_TLS && p.To.Type == obj.TYPE_REG && REG_AX <= p.To.Reg && p.To.Reg <= REG_R15 && ctxt.Headtype != obj.Hsolaris {
obj.Nopout(p)
}
if p.From.Type == obj.TYPE_MEM && p.From.Index == REG_TLS && REG_AX <= p.From.Reg && p.From.Reg <= REG_R15 {
p.From.Reg = REG_TLS
p.From.Scale = 0
p.From.Index = REG_NONE
}
if p.To.Type == obj.TYPE_MEM && p.To.Index == REG_TLS && REG_AX <= p.To.Reg && p.To.Reg <= REG_R15 {
p.To.Reg = REG_TLS
p.To.Scale = 0
p.To.Index = REG_NONE
}
} else {
// load_g_cx, below, always inserts the 1-instruction sequence. Rewrite it
// as the 2-instruction sequence if necessary.
// MOVQ 0(TLS), BX
// becomes
// MOVQ TLS, BX
// MOVQ 0(BX)(TLS*1), BX
if (p.As == AMOVQ || p.As == AMOVL) && p.From.Type == obj.TYPE_MEM && p.From.Reg == REG_TLS && p.To.Type == obj.TYPE_REG && REG_AX <= p.To.Reg && p.To.Reg <= REG_R15 {
q := obj.Appendp(ctxt, p)
q.As = p.As
q.From = p.From
q.From.Type = obj.TYPE_MEM
q.From.Reg = p.To.Reg
q.From.Index = REG_TLS
q.From.Scale = 2 // TODO: use 1
q.To = p.To
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_TLS
p.From.Index = REG_NONE
p.From.Offset = 0
}
}
// TODO: Remove.
if ctxt.Headtype == obj.Hwindows && p.Mode == 64 || ctxt.Headtype == obj.Hplan9 {
if p.From.Scale == 1 && p.From.Index == REG_TLS {
p.From.Scale = 2
}
if p.To.Scale == 1 && p.To.Index == REG_TLS {
p.To.Scale = 2
}
}
// Rewrite 0 to $0 in 3rd argment to CMPPS etc.
// That's what the tables expect.
switch p.As {
case ACMPPD, ACMPPS, ACMPSD, ACMPSS:
if p.To.Type == obj.TYPE_MEM && p.To.Name == obj.NAME_NONE && p.To.Reg == REG_NONE && p.To.Index == REG_NONE && p.To.Sym == nil {
p.To.Type = obj.TYPE_CONST
}
}
// Rewrite CALL/JMP/RET to symbol as TYPE_BRANCH.
switch p.As {
case obj.ACALL, obj.AJMP, obj.ARET:
if p.To.Type == obj.TYPE_MEM && (p.To.Name == obj.NAME_EXTERN || p.To.Name == obj.NAME_STATIC) && p.To.Sym != nil {
p.To.Type = obj.TYPE_BRANCH
}
}
// Rewrite MOVL/MOVQ $XXX(FP/SP) as LEAL/LEAQ.
if p.From.Type == obj.TYPE_ADDR && (ctxt.Arch.Thechar == '6' || p.From.Name != obj.NAME_EXTERN && p.From.Name != obj.NAME_STATIC) {
switch p.As {
case AMOVL:
p.As = ALEAL
p.From.Type = obj.TYPE_MEM
case AMOVQ:
p.As = ALEAQ
p.From.Type = obj.TYPE_MEM
}
}
if ctxt.Headtype == obj.Hnacl && p.Mode == 64 {
if p.From3 != nil {
nacladdr(ctxt, p, p.From3)
}
nacladdr(ctxt, p, &p.From)
nacladdr(ctxt, p, &p.To)
}
// Rewrite float constants to values stored in memory.
switch p.As {
// Convert AMOVSS $(0), Xx to AXORPS Xx, Xx
case AMOVSS:
if p.From.Type == obj.TYPE_FCONST {
if p.From.Val.(float64) == 0 {
if p.To.Type == obj.TYPE_REG && REG_X0 <= p.To.Reg && p.To.Reg <= REG_X15 {
p.As = AXORPS
p.From = p.To
break
}
}
}
fallthrough
case AFMOVF,
AFADDF,
AFSUBF,
AFSUBRF,
AFMULF,
AFDIVF,
AFDIVRF,
AFCOMF,
AFCOMFP,
AADDSS,
ASUBSS,
AMULSS,
ADIVSS,
ACOMISS,
AUCOMISS:
if p.From.Type == obj.TYPE_FCONST {
f32 := float32(p.From.Val.(float64))
i32 := math.Float32bits(f32)
literal := fmt.Sprintf("$f32.%08x", i32)
s := obj.Linklookup(ctxt, literal, 0)
p.From.Type = obj.TYPE_MEM
p.From.Name = obj.NAME_EXTERN
p.From.Sym = s
p.From.Sym.Local = true
p.From.Offset = 0
}
case AMOVSD:
// Convert AMOVSD $(0), Xx to AXORPS Xx, Xx
if p.From.Type == obj.TYPE_FCONST {
if p.From.Val.(float64) == 0 {
if p.To.Type == obj.TYPE_REG && REG_X0 <= p.To.Reg && p.To.Reg <= REG_X15 {
p.As = AXORPS
p.From = p.To
break
}
}
}
fallthrough
case AFMOVD,
AFADDD,
AFSUBD,
AFSUBRD,
AFMULD,
AFDIVD,
AFDIVRD,
AFCOMD,
AFCOMDP,
AADDSD,
ASUBSD,
AMULSD,
ADIVSD,
ACOMISD,
AUCOMISD:
if p.From.Type == obj.TYPE_FCONST {
i64 := math.Float64bits(p.From.Val.(float64))
literal := fmt.Sprintf("$f64.%016x", i64)
s := obj.Linklookup(ctxt, literal, 0)
p.From.Type = obj.TYPE_MEM
p.From.Name = obj.NAME_EXTERN
p.From.Sym = s
p.From.Sym.Local = true
p.From.Offset = 0
}
}
if ctxt.Flag_dynlink {
rewriteToUseGot(ctxt, p)
}
if ctxt.Flag_shared != 0 && p.Mode == 32 {
rewriteToPcrel(ctxt, p)
}
}
func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32) *obj.Prog {
// MOVW g_stackguard(g), R1
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGG
p.From.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0
if ctxt.Cursym.Cfunc != 0 {
p.From.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1
}
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
if framesize <= obj.StackSmall {
// small stack: SP < stackguard
// CMP stackguard, SP
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.Reg = REGSP
} else if framesize <= obj.StackBig {
// large stack: SP-framesize < stackguard-StackSmall
// MOVW $-framesize(SP), R2
// CMP stackguard, R2
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_ADDR
p.From.Reg = REGSP
p.From.Offset = int64(-framesize)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.Reg = REG_R2
} else {
// Such a large stack we need to protect against wraparound
// if SP is close to zero.
// SP-stackguard+StackGuard < framesize + (StackGuard-StackSmall)
// The +StackGuard on both sides is required to keep the left side positive:
// SP is allowed to be slightly below stackguard. See stack.h.
// CMP $StackPreempt, R1
// MOVW.NE $StackGuard(SP), R2
// SUB.NE R1, R2
// MOVW.NE $(framesize+(StackGuard-StackSmall)), R3
// CMP.NE R3, R2
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(uint32(obj.StackPreempt & (1<<32 - 1)))
p.Reg = REG_R1
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_ADDR
p.From.Reg = REGSP
p.From.Offset = obj.StackGuard
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p.Scond = C_SCOND_NE
p = obj.Appendp(ctxt, p)
p.As = ASUB
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p.Scond = C_SCOND_NE
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_ADDR
p.From.Offset = int64(framesize) + (obj.StackGuard - obj.StackSmall)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R3
p.Scond = C_SCOND_NE
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R3
p.Reg = REG_R2
p.Scond = C_SCOND_NE
}
// BLS call-to-morestack
bls := obj.Appendp(ctxt, p)
bls.As = ABLS
bls.To.Type = obj.TYPE_BRANCH
var last *obj.Prog
for last = ctxt.Cursym.Text; last.Link != nil; last = last.Link {
}
spfix := obj.Appendp(ctxt, last)
spfix.As = obj.ANOP
spfix.Spadj = -framesize
// MOVW LR, R3
movw := obj.Appendp(ctxt, spfix)
movw.As = AMOVW
movw.From.Type = obj.TYPE_REG
movw.From.Reg = REGLINK
movw.To.Type = obj.TYPE_REG
movw.To.Reg = REG_R3
bls.Pcond = movw
// BL runtime.morestack
call := obj.Appendp(ctxt, movw)
call.As = obj.ACALL
call.To.Type = obj.TYPE_BRANCH
morestack := "runtime.morestack"
switch {
case ctxt.Cursym.Cfunc != 0:
morestack = "runtime.morestackc"
case ctxt.Cursym.Text.From3.Offset&obj.NEEDCTXT == 0:
morestack = "runtime.morestack_noctxt"
}
call.To.Sym = obj.Linklookup(ctxt, morestack, 0)
// B start
b := obj.Appendp(ctxt, call)
b.As = obj.AJMP
b.To.Type = obj.TYPE_BRANCH
b.Pcond = ctxt.Cursym.Text.Link
b.Spadj = +framesize
return bls
}
// UNUSED
func peep(firstp *obj.Prog) {
g := (*gc.Graph)(gc.Flowstart(firstp, nil))
if g == nil {
return
}
gactive = 0
var r *gc.Flow
var p *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 {
/*
* elide shift into TYPE_SHIFT operand of subsequent instruction
*/
// if(shiftprop(r)) {
// excise(r);
// t++;
// break;
// }
case arm.ASLL,
arm.ASRL,
arm.ASRA:
break
case arm.AMOVB,
arm.AMOVH,
arm.AMOVW,
arm.AMOVF,
arm.AMOVD:
if regtyp(&p.From) {
if p.From.Type == p.To.Type && isfloatreg(&p.From) == isfloatreg(&p.To) {
if p.Scond == arm.C_SCOND_NONE {
if copyprop(g, r) {
excise(r)
t++
break
}
if subprop(r) && copyprop(g, r) {
excise(r)
t++
break
}
}
}
}
case arm.AMOVHS,
arm.AMOVHU,
arm.AMOVBS,
arm.AMOVBU:
if p.From.Type == obj.TYPE_REG {
if shortprop(r) {
t++
}
}
}
}
/*
if(p->scond == C_SCOND_NONE)
if(regtyp(&p->to))
if(isdconst(&p->from)) {
constprop(&p->from, &p->to, r->s1);
}
break;
*/
if t != 0 {
goto loop1
}
for r := (*gc.Flow)(g.Start); r != nil; r = r.Link {
p = r.Prog
switch p.As {
/*
* EOR -1,x,y => MVN x,y
*/
case arm.AEOR:
if isdconst(&p.From) && p.From.Offset == -1 {
p.As = arm.AMVN
p.From.Type = obj.TYPE_REG
if p.Reg != 0 {
p.From.Reg = p.Reg
} else {
p.From.Reg = p.To.Reg
}
p.Reg = 0
}
}
}
for r := (*gc.Flow)(g.Start); r != nil; r = r.Link {
p = r.Prog
switch p.As {
case arm.AMOVW,
arm.AMOVB,
arm.AMOVBS,
arm.AMOVBU:
if p.From.Type == obj.TYPE_MEM && p.From.Offset == 0 {
xtramodes(g, r, &p.From)
} else if p.To.Type == obj.TYPE_MEM && p.To.Offset == 0 {
xtramodes(g, r, &p.To)
} else {
continue
}
}
}
// case ACMP:
// /*
// * elide CMP $0,x if calculation of x can set condition codes
// */
// if(isdconst(&p->from) || p->from.offset != 0)
// continue;
// r2 = r->s1;
// if(r2 == nil)
// continue;
// t = r2->prog->as;
// switch(t) {
// default:
// continue;
// case ABEQ:
// case ABNE:
// case ABMI:
// case ABPL:
// break;
// case ABGE:
// t = ABPL;
// break;
// case ABLT:
// t = ABMI;
// break;
// case ABHI:
// t = ABNE;
// break;
// case ABLS:
// t = ABEQ;
// break;
// }
// r1 = r;
// do
// r1 = uniqp(r1);
// while (r1 != nil && r1->prog->as == ANOP);
// if(r1 == nil)
// continue;
// p1 = r1->prog;
// if(p1->to.type != TYPE_REG)
// continue;
// if(p1->to.reg != p->reg)
// if(!(p1->as == AMOVW && p1->from.type == TYPE_REG && p1->from.reg == p->reg))
// continue;
//
// switch(p1->as) {
// default:
// continue;
// case AMOVW:
// if(p1->from.type != TYPE_REG)
// continue;
// case AAND:
// case AEOR:
// case AORR:
// case ABIC:
// case AMVN:
// case ASUB:
// case ARSB:
// case AADD:
// case AADC:
// case ASBC:
// case ARSC:
// break;
// }
// p1->scond |= C_SBIT;
// r2->prog->as = t;
// excise(r);
// continue;
// predicate(g);
gc.Flowend(g)
}
// movb elimination.
// movb is simulated by the linker
// when a register other than ax, bx, cx, dx
// is used, so rewrite to other instructions
// when possible. a movb into a register
// can smash the entire 64-bit register without
// causing any trouble.
func elimshortmov(g *gc.Graph) {
var p *obj.Prog
for r := g.Start; r != nil; r = r.Link {
p = r.Prog
if regtyp(&p.To) {
switch p.As {
case x86.AINCB,
x86.AINCW:
p.As = x86.AINCL
case x86.ADECB,
x86.ADECW:
p.As = x86.ADECL
case x86.ANEGB,
x86.ANEGW:
p.As = x86.ANEGL
case x86.ANOTB,
x86.ANOTW:
p.As = x86.ANOTL
}
if regtyp(&p.From) || p.From.Type == obj.TYPE_CONST {
// move or artihmetic into partial register.
// from another register or constant can be movl.
// we don't switch to 32-bit arithmetic if it can
// change how the carry bit is set (and the carry bit is needed).
switch p.As {
case x86.AMOVB,
x86.AMOVW:
p.As = x86.AMOVL
case x86.AADDB,
x86.AADDW:
if !needc(p.Link) {
p.As = x86.AADDL
}
case x86.ASUBB,
x86.ASUBW:
if !needc(p.Link) {
p.As = x86.ASUBL
}
case x86.AMULB,
x86.AMULW:
p.As = x86.AMULL
case x86.AIMULB,
x86.AIMULW:
p.As = x86.AIMULL
case x86.AANDB,
x86.AANDW:
p.As = x86.AANDL
case x86.AORB,
x86.AORW:
p.As = x86.AORL
case x86.AXORB,
x86.AXORW:
p.As = x86.AXORL
case x86.ASHLB,
x86.ASHLW:
p.As = x86.ASHLL
}
} else {
// explicit zero extension
switch p.As {
case x86.AMOVB:
p.As = x86.AMOVBLZX
case x86.AMOVW:
p.As = x86.AMOVWLZX
}
}
}
}
}
// Rewrite p, if necessary, to access global data via the global offset table.
func rewriteToUseGot(ctxt *obj.Link, p *obj.Prog) {
if p.As == obj.ADUFFCOPY || p.As == obj.ADUFFZERO {
// ADUFFxxx $offset
// becomes
// MOVW runtime.duffxxx@GOT, R9
// ADD $offset, R9
// CALL (R9)
var sym *obj.LSym
if p.As == obj.ADUFFZERO {
sym = obj.Linklookup(ctxt, "runtime.duffzero", 0)
} else {
sym = obj.Linklookup(ctxt, "runtime.duffcopy", 0)
}
offset := p.To.Offset
p.As = AMOVW
p.From.Type = obj.TYPE_MEM
p.From.Name = obj.NAME_GOTREF
p.From.Sym = sym
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R9
p.To.Name = obj.NAME_NONE
p.To.Offset = 0
p.To.Sym = nil
p1 := obj.Appendp(ctxt, p)
p1.As = AADD
p1.From.Type = obj.TYPE_CONST
p1.From.Offset = offset
p1.To.Type = obj.TYPE_REG
p1.To.Reg = REG_R9
p2 := obj.Appendp(ctxt, p1)
p2.As = obj.ACALL
p2.To.Type = obj.TYPE_MEM
p2.To.Reg = REG_R9
return
}
// We only care about global data: NAME_EXTERN means a global
// symbol in the Go sense, and p.Sym.Local is true for a few
// internally defined symbols.
if p.From.Type == obj.TYPE_ADDR && p.From.Name == obj.NAME_EXTERN && !p.From.Sym.Local {
// MOVW $sym, Rx becomes MOVW sym@GOT, Rx
// MOVW $sym+<off>, Rx becomes MOVW sym@GOT, Rx; ADD <off>, Rx
if p.As != AMOVW {
ctxt.Diag("do not know how to handle TYPE_ADDR in %v with -dynlink", p)
}
if p.To.Type != obj.TYPE_REG {
ctxt.Diag("do not know how to handle LEAQ-type insn to non-register in %v with -dynlink", p)
}
p.From.Type = obj.TYPE_MEM
p.From.Name = obj.NAME_GOTREF
if p.From.Offset != 0 {
q := obj.Appendp(ctxt, p)
q.As = AADD
q.From.Type = obj.TYPE_CONST
q.From.Offset = p.From.Offset
q.To = p.To
p.From.Offset = 0
}
}
if p.From3 != nil && p.From3.Name == obj.NAME_EXTERN {
ctxt.Diag("don't know how to handle %v with -dynlink", p)
}
var source *obj.Addr
// MOVx sym, Ry becomes MOVW sym@GOT, R9; MOVx (R9), Ry
// MOVx Ry, sym becomes MOVW sym@GOT, R9; MOVx Ry, (R9)
// An addition may be inserted between the two MOVs if there is an offset.
if p.From.Name == obj.NAME_EXTERN && !p.From.Sym.Local {
if p.To.Name == obj.NAME_EXTERN && !p.To.Sym.Local {
ctxt.Diag("cannot handle NAME_EXTERN on both sides in %v with -dynlink", p)
}
source = &p.From
} else if p.To.Name == obj.NAME_EXTERN && !p.To.Sym.Local {
source = &p.To
} else {
return
}
if p.As == obj.ATEXT || p.As == obj.AFUNCDATA || p.As == obj.ACALL || p.As == obj.ARET || p.As == obj.AJMP {
return
}
if source.Sym.Type == obj.STLSBSS {
return
}
if source.Type != obj.TYPE_MEM {
ctxt.Diag("don't know how to handle %v with -dynlink", p)
}
p1 := obj.Appendp(ctxt, p)
p2 := obj.Appendp(ctxt, p1)
p1.As = AMOVW
p1.From.Type = obj.TYPE_MEM
p1.From.Sym = source.Sym
p1.From.Name = obj.NAME_GOTREF
p1.To.Type = obj.TYPE_REG
p1.To.Reg = REG_R9
p2.As = p.As
p2.From = p.From
p2.To = p.To
if p.From.Name == obj.NAME_EXTERN {
p2.From.Reg = REG_R9
p2.From.Name = obj.NAME_NONE
p2.From.Sym = nil
} else if p.To.Name == obj.NAME_EXTERN {
p2.To.Reg = REG_R9
p2.To.Name = obj.NAME_NONE
p2.To.Sym = nil
} else {
return
}
obj.Nopout(p)
}
func stacksplit(ctxt *obj.Link, p *obj.Prog, framesize int32) *obj.Prog {
// MOV g_stackguard(g), R1
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_MEM
p.From.Reg = REGG
p.From.Offset = 2 * int64(ctxt.Arch.Ptrsize) // G.stackguard0
if ctxt.Cursym.Cfunc != 0 {
p.From.Offset = 3 * int64(ctxt.Arch.Ptrsize) // G.stackguard1
}
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R1
q := (*obj.Prog)(nil)
if framesize <= obj.StackSmall {
// small stack: SP < stackguard
// MOV SP, R2
// CMP stackguard, R2
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_REG
p.From.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.Reg = REG_R2
} else if framesize <= obj.StackBig {
// large stack: SP-framesize < stackguard-StackSmall
// SUB $framesize, SP, R2
// CMP stackguard, R2
p = obj.Appendp(ctxt, p)
p.As = ASUB
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(framesize)
p.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.Reg = REG_R2
} else {
// Such a large stack we need to protect against wraparound
// if SP is close to zero.
// SP-stackguard+StackGuard < framesize + (StackGuard-StackSmall)
// The +StackGuard on both sides is required to keep the left side positive:
// SP is allowed to be slightly below stackguard. See stack.h.
// CMP $StackPreempt, R1
// BEQ label_of_call_to_morestack
// ADD $StackGuard, SP, R2
// SUB R1, R2
// MOV $(framesize+(StackGuard-StackSmall)), R3
// CMP R3, R2
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_CONST
p.From.Offset = obj.StackPreempt
p.Reg = REG_R1
p = obj.Appendp(ctxt, p)
q = p
p.As = ABEQ
p.To.Type = obj.TYPE_BRANCH
p = obj.Appendp(ctxt, p)
p.As = AADD
p.From.Type = obj.TYPE_CONST
p.From.Offset = obj.StackGuard
p.Reg = REGSP
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = ASUB
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R1
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R2
p = obj.Appendp(ctxt, p)
p.As = AMOVD
p.From.Type = obj.TYPE_CONST
p.From.Offset = int64(framesize) + (obj.StackGuard - obj.StackSmall)
p.To.Type = obj.TYPE_REG
p.To.Reg = REG_R3
p = obj.Appendp(ctxt, p)
p.As = ACMP
p.From.Type = obj.TYPE_REG
p.From.Reg = REG_R3
p.Reg = REG_R2
}
// BLS do-morestack
bls := obj.Appendp(ctxt, p)
bls.As = ABLS
bls.To.Type = obj.TYPE_BRANCH
var last *obj.Prog
for last = ctxt.Cursym.Text; last.Link != nil; last = last.Link {
}
spfix := obj.Appendp(ctxt, last)
spfix.As = obj.ANOP
spfix.Spadj = -framesize
// MOV LR, R3
movlr := obj.Appendp(ctxt, spfix)
movlr.As = AMOVD
movlr.From.Type = obj.TYPE_REG
movlr.From.Reg = REGLINK
movlr.To.Type = obj.TYPE_REG
movlr.To.Reg = REG_R3
if q != nil {
q.Pcond = movlr
}
bls.Pcond = movlr
debug := movlr
if false {
debug = obj.Appendp(ctxt, debug)
debug.As = AMOVD
debug.From.Type = obj.TYPE_CONST
debug.From.Offset = int64(framesize)
debug.To.Type = obj.TYPE_REG
debug.To.Reg = REGTMP
}
// BL runtime.morestack(SB)
call := obj.Appendp(ctxt, debug)
call.As = ABL
call.To.Type = obj.TYPE_BRANCH
morestack := "runtime.morestack"
switch {
case ctxt.Cursym.Cfunc != 0:
morestack = "runtime.morestackc"
case ctxt.Cursym.Text.From3.Offset&obj.NEEDCTXT == 0:
morestack = "runtime.morestack_noctxt"
}
call.To.Sym = obj.Linklookup(ctxt, morestack, 0)
// B start
jmp := obj.Appendp(ctxt, call)
jmp.As = AB
jmp.To.Type = obj.TYPE_BRANCH
jmp.Pcond = ctxt.Cursym.Text.Link
jmp.Spadj = +framesize
// placeholder for bls's jump target
// p = obj.Appendp(ctxt, p)
// p.As = obj.ANOP
return bls
}
func progedit(ctxt *obj.Link, p *obj.Prog) {
p.From.Class = 0
p.To.Class = 0
// Rewrite B/BL to symbol as TYPE_BRANCH.
switch p.As {
case AB,
ABL,
obj.ADUFFZERO,
obj.ADUFFCOPY:
if p.To.Type == obj.TYPE_MEM && (p.To.Name == obj.NAME_EXTERN || p.To.Name == obj.NAME_STATIC) && p.To.Sym != nil {
p.To.Type = obj.TYPE_BRANCH
}
}
// Replace TLS register fetches on older ARM procesors.
switch p.As {
// Treat MRC 15, 0, <reg>, C13, C0, 3 specially.
case AMRC:
if p.To.Offset&0xffff0fff == 0xee1d0f70 {
// Because the instruction might be rewriten to a BL which returns in R0
// the register must be zero.
if p.To.Offset&0xf000 != 0 {
ctxt.Diag("%v: TLS MRC instruction must write to R0 as it might get translated into a BL instruction", p.Line())
}
if ctxt.Goarm < 7 {
// Replace it with BL runtime.read_tls_fallback(SB) for ARM CPUs that lack the tls extension.
if progedit_tlsfallback == nil {
progedit_tlsfallback = obj.Linklookup(ctxt, "runtime.read_tls_fallback", 0)
}
// MOVW LR, R11
p.As = AMOVW
p.From.Type = obj.TYPE_REG
p.From.Reg = REGLINK
p.To.Type = obj.TYPE_REG
p.To.Reg = REGTMP
// BL runtime.read_tls_fallback(SB)
p = obj.Appendp(ctxt, p)
p.As = ABL
p.To.Type = obj.TYPE_BRANCH
p.To.Sym = progedit_tlsfallback
p.To.Offset = 0
// MOVW R11, LR
p = obj.Appendp(ctxt, p)
p.As = AMOVW
p.From.Type = obj.TYPE_REG
p.From.Reg = REGTMP
p.To.Type = obj.TYPE_REG
p.To.Reg = REGLINK
break
}
}
// Otherwise, MRC/MCR instructions need no further treatment.
p.As = AWORD
}
// Rewrite float constants to values stored in memory.
switch p.As {
case AMOVF:
if p.From.Type == obj.TYPE_FCONST && chipfloat5(ctxt, p.From.Val.(float64)) < 0 && (chipzero5(ctxt, p.From.Val.(float64)) < 0 || p.Scond&C_SCOND != C_SCOND_NONE) {
f32 := float32(p.From.Val.(float64))
i32 := math.Float32bits(f32)
literal := fmt.Sprintf("$f32.%08x", i32)
s := obj.Linklookup(ctxt, literal, 0)
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
case AMOVD:
if p.From.Type == obj.TYPE_FCONST && chipfloat5(ctxt, p.From.Val.(float64)) < 0 && (chipzero5(ctxt, p.From.Val.(float64)) < 0 || p.Scond&C_SCOND != C_SCOND_NONE) {
i64 := math.Float64bits(p.From.Val.(float64))
literal := fmt.Sprintf("$f64.%016x", i64)
s := obj.Linklookup(ctxt, literal, 0)
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
}
if ctxt.Flag_dynlink {
rewriteToUseGot(ctxt, p)
}
}
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)
}
func xfol(ctxt *obj.Link, p *obj.Prog, last **obj.Prog) {
var q *obj.Prog
var r *obj.Prog
var a int
var i int
loop:
if p == nil {
return
}
a = int(p.As)
if a == AB {
q = p.Pcond
if q != nil && q.As != obj.ATEXT {
p.Mark |= FOLL
p = q
if p.Mark&FOLL == 0 {
goto loop
}
}
}
if p.Mark&FOLL != 0 {
i = 0
q = p
for ; i < 4; i, q = i+1, q.Link {
if q == *last || q == nil {
break
}
a = int(q.As)
if a == obj.ANOP {
i--
continue
}
if a == AB || (a == obj.ARET && q.Scond == C_SCOND_NONE) || a == ARFE || a == obj.AUNDEF {
goto copy
}
if q.Pcond == nil || (q.Pcond.Mark&FOLL != 0) {
continue
}
if a != ABEQ && a != ABNE {
continue
}
copy:
for {
r = ctxt.NewProg()
*r = *p
if r.Mark&FOLL == 0 {
fmt.Printf("can't happen 1\n")
}
r.Mark |= FOLL
if p != q {
p = p.Link
(*last).Link = r
*last = r
continue
}
(*last).Link = r
*last = r
if a == AB || (a == obj.ARET && q.Scond == C_SCOND_NONE) || a == ARFE || a == obj.AUNDEF {
return
}
r.As = ABNE
if a == ABNE {
r.As = ABEQ
}
r.Pcond = p.Link
r.Link = p.Pcond
if r.Link.Mark&FOLL == 0 {
xfol(ctxt, r.Link, last)
}
if r.Pcond.Mark&FOLL == 0 {
fmt.Printf("can't happen 2\n")
}
return
}
}
a = AB
q = ctxt.NewProg()
q.As = int16(a)
q.Lineno = p.Lineno
q.To.Type = obj.TYPE_BRANCH
q.To.Offset = p.Pc
q.Pcond = p
p = q
}
p.Mark |= FOLL
(*last).Link = p
*last = p
if a == AB || (a == obj.ARET && p.Scond == C_SCOND_NONE) || a == ARFE || a == obj.AUNDEF {
return
}
if p.Pcond != nil {
if a != ABL && a != ABX && p.Link != nil {
q = obj.Brchain(ctxt, p.Link)
if a != obj.ATEXT {
if q != nil && (q.Mark&FOLL != 0) {
p.As = int16(relinv(a))
p.Link = p.Pcond
p.Pcond = q
}
}
xfol(ctxt, p.Link, last)
q = obj.Brchain(ctxt, p.Pcond)
if q == nil {
q = p.Pcond
}
if q.Mark&FOLL != 0 {
p.Pcond = q
return
}
p = q
goto loop
}
}
p = p.Link
goto loop
}
func progedit(ctxt *obj.Link, p *obj.Prog) {
p.From.Class = 0
p.To.Class = 0
// $0 results in C_ZCON, which matches both C_REG and various
// C_xCON, however the C_REG cases in asmout don't expect a
// constant, so they will use the register fields and assemble
// a R0. To prevent that, rewrite $0 as ZR.
if p.From.Type == obj.TYPE_CONST && p.From.Offset == 0 {
p.From.Type = obj.TYPE_REG
p.From.Reg = REGZERO
}
if p.To.Type == obj.TYPE_CONST && p.To.Offset == 0 {
p.To.Type = obj.TYPE_REG
p.To.Reg = REGZERO
}
// Rewrite BR/BL to symbol as TYPE_BRANCH.
switch p.As {
case AB,
ABL,
obj.ARET,
obj.ADUFFZERO,
obj.ADUFFCOPY:
if p.To.Sym != nil {
p.To.Type = obj.TYPE_BRANCH
}
break
}
// Rewrite float constants to values stored in memory.
switch p.As {
case AFMOVS:
if p.From.Type == obj.TYPE_FCONST {
f32 := float32(p.From.Val.(float64))
i32 := math.Float32bits(f32)
literal := fmt.Sprintf("$f32.%08x", uint32(i32))
s := obj.Linklookup(ctxt, literal, 0)
s.Size = 4
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Sym.Local = true
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
case AFMOVD:
if p.From.Type == obj.TYPE_FCONST {
i64 := math.Float64bits(p.From.Val.(float64))
literal := fmt.Sprintf("$f64.%016x", uint64(i64))
s := obj.Linklookup(ctxt, literal, 0)
s.Size = 8
p.From.Type = obj.TYPE_MEM
p.From.Sym = s
p.From.Sym.Local = true
p.From.Name = obj.NAME_EXTERN
p.From.Offset = 0
}
break
}
// Rewrite negative immediates as positive immediates with
// complementary instruction.
switch p.As {
case AADD,
AADDW,
ASUB,
ASUBW,
ACMP,
ACMPW,
ACMN,
ACMNW:
if p.From.Type == obj.NAME_EXTERN && p.From.Offset < 0 {
p.From.Offset = -p.From.Offset
p.As = complements[p.As]
}
break
}
if ctxt.Flag_dynlink {
rewriteToUseGot(ctxt, p)
}
}
