func progedit(ctxt *liblink.Link, p *liblink.Prog) {
var literal string
var s *liblink.LSym
var q *liblink.Prog
// See obj6.c for discussion of TLS.
if canuselocaltls(ctxt) {
// Reduce TLS initial exec model to TLS local exec model.
// Sequences like
// MOVL TLS, BX
// ... off(BX)(TLS*1) ...
// become
// NOP
// ... off(TLS) ...
if p.As == AMOVL && p.From.Typ == D_TLS && D_AX <= p.To.Typ && p.To.Typ <= D_DI {
p.As = ANOP
p.From.Typ = D_NONE
p.To.Typ = D_NONE
}
if p.From.Index == D_TLS && D_INDIR+D_AX <= p.From.Typ && p.From.Typ <= D_INDIR+D_DI {
p.From.Typ = D_INDIR + D_TLS
p.From.Scale = 0
p.From.Index = D_NONE
}
if p.To.Index == D_TLS && D_INDIR+D_AX <= p.To.Typ && p.To.Typ <= D_INDIR+D_DI {
p.To.Typ = D_INDIR + D_TLS
p.To.Scale = 0
p.To.Index = D_NONE
}
} else {
// As a courtesy to the C compilers, rewrite TLS local exec load as TLS initial exec load.
// The instruction
// MOVL off(TLS), BX
// becomes the sequence
// MOVL TLS, BX
// MOVL off(BX)(TLS*1), BX
// This allows the C compilers to emit references to m and g using the direct off(TLS) form.
if p.As == AMOVL && p.From.Typ == D_INDIR+D_TLS && D_AX <= p.To.Typ && p.To.Typ <= D_DI {
q = liblink.Appendp(ctxt, p)
q.As = p.As
q.From = p.From
q.From.Typ = D_INDIR + p.To.Typ
q.From.Index = D_TLS
q.From.Scale = 2 // TODO: use 1
q.To = p.To
p.From.Typ = D_TLS
p.From.Index = D_NONE
p.From.Offset = 0
}
}
// TODO: Remove.
if ctxt.Headtype == liblink.Hplan9 {
if p.From.Scale == 1 && p.From.Index == D_TLS {
p.From.Scale = 2
}
if p.To.Scale == 1 && p.To.Index == D_TLS {
p.To.Scale = 2
}
}
// Rewrite CALL/JMP/RET to symbol as D_BRANCH.
switch p.As {
case ACALL,
AJMP,
ARET:
if (p.To.Typ == D_EXTERN || p.To.Typ == D_STATIC) && p.To.Sym != nil {
p.To.Typ = D_BRANCH
}
break
}
// Rewrite float constants to values stored in memory.
switch p.As {
case AFMOVF,
AFADDF,
AFSUBF,
AFSUBRF,
AFMULF,
AFDIVF,
AFDIVRF,
AFCOMF,
AFCOMFP,
AMOVSS,
AADDSS,
ASUBSS,
AMULSS,
ADIVSS,
ACOMISS,
AUCOMISS:
if p.From.Typ == D_FCONST {
var i32 uint32
var f32 float32
f32 = float32(p.From.U.Dval)
i32 = math.Float32bits(f32)
literal = fmt.Sprintf("$f32.%08x", uint32(i32))
s = liblink.Linklookup(ctxt, literal, 0)
if s.Typ == 0 {
s.Typ = liblink.SRODATA
liblink.Adduint32(ctxt, s, i32)
s.Reachable = 0
}
p.From.Typ = D_EXTERN
p.From.Sym = s
p.From.Offset = 0
}
case AFMOVD,
AFADDD,
AFSUBD,
AFSUBRD,
AFMULD,
AFDIVD,
AFDIVRD,
AFCOMD,
AFCOMDP,
AMOVSD,
AADDSD,
ASUBSD,
AMULSD,
ADIVSD,
ACOMISD,
AUCOMISD:
if p.From.Typ == D_FCONST {
var i64 uint64
i64 = math.Float64bits(p.From.U.Dval)
literal = fmt.Sprintf("$f64.%016x", uint64(i64))
s = liblink.Linklookup(ctxt, literal, 0)
if s.Typ == 0 {
s.Typ = liblink.SRODATA
liblink.Adduint64(ctxt, s, i64)
s.Reachable = 0
}
p.From.Typ = D_EXTERN
p.From.Sym = s
p.From.Offset = 0
}
break
}
}
func progedit(ctxt *liblink.Link, p *liblink.Prog) {
var literal string
var s *liblink.LSym
var q *liblink.Prog
// 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 8(AX)(TLS*1), CX // load m 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 8(TLS), CX // load m into CX
//
// The 2-instruction and 1-instruction forms correspond roughly to
// ELF TLS initial exec mode and ELF TLS local exec mode, respectively.
//
// We applies this rewrite on systems that support the 1-instruction form.
// The decision is made using only the operating system (and probably
// the -shared flag, eventually), 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.
if canuselocaltls(ctxt) {
// Reduce TLS initial exec model to TLS local exec model.
// 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.Typ == D_TLS && D_AX <= p.To.Typ && p.To.Typ <= D_R15 && ctxt.Headtype != liblink.Hsolaris {
nopout(p)
}
if p.From.Index == D_TLS && D_INDIR+D_AX <= p.From.Typ && p.From.Typ <= D_INDIR+D_R15 {
p.From.Typ = D_INDIR + D_TLS
p.From.Scale = 0
p.From.Index = D_NONE
}
if p.To.Index == D_TLS && D_INDIR+D_AX <= p.To.Typ && p.To.Typ <= D_INDIR+D_R15 {
p.To.Typ = D_INDIR + D_TLS
p.To.Scale = 0
p.To.Index = D_NONE
}
} else {
// As a courtesy to the C compilers, rewrite TLS local exec load as TLS initial exec load.
// The instruction
// MOVQ off(TLS), BX
// becomes the sequence
// MOVQ TLS, BX
// MOVQ off(BX)(TLS*1), BX
// This allows the C compilers to emit references to m and g using the direct off(TLS) form.
if (p.As == AMOVQ || p.As == AMOVL) && p.From.Typ == D_INDIR+D_TLS && D_AX <= p.To.Typ && p.To.Typ <= D_R15 {
q = liblink.Appendp(ctxt, p)
q.As = p.As
q.From = p.From
q.From.Typ = D_INDIR + p.To.Typ
q.From.Index = D_TLS
q.From.Scale = 2 // TODO: use 1
q.To = p.To
p.From.Typ = D_TLS
p.From.Index = D_NONE
p.From.Offset = 0
}
}
// TODO: Remove.
if ctxt.Headtype == liblink.Hwindows || ctxt.Headtype == liblink.Hplan9 {
if p.From.Scale == 1 && p.From.Index == D_TLS {
p.From.Scale = 2
}
if p.To.Scale == 1 && p.To.Index == D_TLS {
p.To.Scale = 2
}
}
if ctxt.Headtype == liblink.Hnacl {
nacladdr(ctxt, p, &p.From)
nacladdr(ctxt, p, &p.To)
}
// Maintain information about code generation mode.
if ctxt.Mode == 0 {
ctxt.Mode = 64
}
p.Mode = ctxt.Mode
switch p.As {
case AMODE:
if p.From.Typ == D_CONST || p.From.Typ == D_INDIR+D_NONE {
switch int(p.From.Offset) {
case 16,
32,
64:
ctxt.Mode = int(p.From.Offset)
break
}
}
nopout(p)
break
}
// Rewrite CALL/JMP/RET to symbol as D_BRANCH.
switch p.As {
case ACALL,
AJMP,
ARET:
if (p.To.Typ == D_EXTERN || p.To.Typ == D_STATIC) && p.To.Sym != nil {
p.To.Typ = D_BRANCH
}
break
}
// Rewrite float constants to values stored in memory.
switch p.As {
case AFMOVF,
AFADDF,
AFSUBF,
AFSUBRF,
AFMULF,
AFDIVF,
AFDIVRF,
AFCOMF,
AFCOMFP,
AMOVSS,
AADDSS,
ASUBSS,
AMULSS,
ADIVSS,
ACOMISS,
AUCOMISS:
if p.From.Typ == D_FCONST {
var i32 uint32
var f32 float32
f32 = float32(p.From.U.Dval)
i32 = math.Float32bits(f32)
literal = fmt.Sprintf("$f32.%08x", uint32(i32))
s = liblink.Linklookup(ctxt, literal, 0)
if s.Typ == 0 {
s.Typ = liblink.SRODATA
liblink.Adduint32(ctxt, s, i32)
s.Reachable = 0
}
p.From.Typ = D_EXTERN
p.From.Sym = s
p.From.Offset = 0
}
case AFMOVD,
AFADDD,
AFSUBD,
AFSUBRD,
AFMULD,
AFDIVD,
AFDIVRD,
AFCOMD,
AFCOMDP,
AMOVSD,
AADDSD,
ASUBSD,
AMULSD,
ADIVSD,
ACOMISD,
AUCOMISD:
if p.From.Typ == D_FCONST {
var i64 uint64
i64 = math.Float64bits(p.From.U.Dval)
literal = fmt.Sprintf("$f64.%016x", uint64(i64))
s = liblink.Linklookup(ctxt, literal, 0)
if s.Typ == 0 {
s.Typ = liblink.SRODATA
liblink.Adduint64(ctxt, s, i64)
s.Reachable = 0
}
p.From.Typ = D_EXTERN
p.From.Sym = s
p.From.Offset = 0
}
break
}
}
func addstacksplit(ctxt *liblink.Link, cursym *liblink.LSym) {
var p *liblink.Prog
var pl *liblink.Prog
var q *liblink.Prog
var q1 *liblink.Prog
var q2 *liblink.Prog
var o int
var autosize int64
var autoffset int64
autosize = 0
if ctxt.Symmorestack[0] == nil {
ctxt.Symmorestack[0] = liblink.Linklookup(ctxt, "runtime.morestack", 0)
ctxt.Symmorestack[1] = liblink.Linklookup(ctxt, "runtime.morestack_noctxt", 0)
}
q = nil
ctxt.Cursym = cursym
if cursym.Text == nil || cursym.Text.Link == nil {
return
}
softfloat(ctxt, cursym)
p = cursym.Text
autoffset = p.To.Offset
if autoffset < 0 {
autoffset = 0
}
cursym.Locals = autoffset
cursym.Args = p.To.Offset2
if ctxt.Debugzerostack != 0 {
if autoffset != 0 && p.Reg&liblink.NOSPLIT == 0 {
// MOVW $4(R13), R1
p = liblink.Appendp(ctxt, p)
p.As = AMOVW
p.From.Typ = D_CONST
p.From.Reg = 13
p.From.Offset = 4
p.To.Typ = D_REG
p.To.Reg = 1
// MOVW $n(R13), R2
p = liblink.Appendp(ctxt, p)
p.As = AMOVW
p.From.Typ = D_CONST
p.From.Reg = 13
p.From.Offset = 4 + autoffset
p.To.Typ = D_REG
p.To.Reg = 2
// MOVW $0, R3
p = liblink.Appendp(ctxt, p)
p.As = AMOVW
p.From.Typ = D_CONST
p.From.Offset = 0
p.To.Typ = D_REG
p.To.Reg = 3
// L:
// MOVW.nil R3, 0(R1) +4
// CMP R1, R2
// BNE L
pl = liblink.Appendp(ctxt, p)
p = pl
p.As = AMOVW
p.From.Typ = D_REG
p.From.Reg = 3
p.To.Typ = D_OREG
p.To.Reg = 1
p.To.Offset = 4
p.Scond |= C_PBIT
p = liblink.Appendp(ctxt, p)
p.As = ACMP
p.From.Typ = D_REG
p.From.Reg = 1
p.Reg = 2
p = liblink.Appendp(ctxt, p)
p.As = ABNE
p.To.Typ = D_BRANCH
p.Pcond = pl
}
}
/*
* find leaf subroutines
* strip NOPs
* expand RET
* expand BECOME pseudo
*/
for p = cursym.Text; p != nil; p = p.Link {
switch p.As {
case ACASE:
if ctxt.Flag_shared != 0 {
linkcase(p)
}
case ATEXT:
p.Mark |= LEAF
case ARET:
break
case ADIV,
ADIVU,
AMOD,
AMODU:
q = p
if ctxt.Sym_div == nil {
initdiv(ctxt)
}
cursym.Text.Mark &^= LEAF
continue
case ANOP:
q1 = p.Link
q.Link = q1 /* q is non-nop */
if q1 != nil {
q1.Mark |= p.Mark
}
continue
case ABL,
ABX,
ADUFFZERO,
ADUFFCOPY:
cursym.Text.Mark &^= LEAF
fallthrough
case ABCASE,
AB,
ABEQ,
ABNE,
ABCS,
ABHS,
ABCC,
ABLO,
ABMI,
ABPL,
ABVS,
ABVC,
ABHI,
ABLS,
ABGE,
ABLT,
ABGT,
ABLE:
q1 = p.Pcond
if q1 != nil {
for q1.As == ANOP {
q1 = q1.Link
p.Pcond = q1
}
}
break
}
q = p
}
for p = cursym.Text; p != nil; p = p.Link {
o = p.As
switch o {
case ATEXT:
autosize = p.To.Offset + 4
if autosize <= 4 {
if cursym.Text.Mark&LEAF != 0 {
p.To.Offset = -4
autosize = 0
}
}
if autosize == 0 && cursym.Text.Mark&LEAF == 0 {
if ctxt.Debugvlog != 0 {
fmt.Fprintf(ctxt.Bso, "save suppressed in: %s\n", cursym.Name)
liblink.Bflush(ctxt.Bso)
}
cursym.Text.Mark |= LEAF
}
if cursym.Text.Mark&LEAF != 0 {
cursym.Leaf = 1
if autosize == 0 {
break
}
}
if p.Reg&liblink.NOSPLIT == 0 {
p = stacksplit(ctxt, p, autosize, bool2int(cursym.Text.Reg&liblink.NEEDCTXT == 0)) // emit split check
}
// MOVW.W R14,$-autosize(SP)
p = liblink.Appendp(ctxt, p)
p.As = AMOVW
p.Scond |= C_WBIT
p.From.Typ = D_REG
p.From.Reg = REGLINK
p.To.Typ = D_OREG
p.To.Offset = -autosize
p.To.Reg = REGSP
p.Spadj = autosize
if cursym.Text.Reg&liblink.WRAPPER != 0 {
// g->panicwrap += autosize;
// MOVW panicwrap_offset(g), R3
// ADD $autosize, R3
// MOVW R3 panicwrap_offset(g)
p = liblink.Appendp(ctxt, p)
p.As = AMOVW
p.From.Typ = D_OREG
p.From.Reg = REGG
p.From.Offset = 2 * ctxt.Arch.Ptrsize
p.To.Typ = D_REG
p.To.Reg = 3
p = liblink.Appendp(ctxt, p)
p.As = AADD
p.From.Typ = D_CONST
p.From.Offset = autosize
p.To.Typ = D_REG
p.To.Reg = 3
p = liblink.Appendp(ctxt, p)
p.As = AMOVW
p.From.Typ = D_REG
p.From.Reg = 3
p.To.Typ = D_OREG
p.To.Reg = REGG
p.To.Offset = 2 * ctxt.Arch.Ptrsize
}
case ARET:
nocache_obj5(p)
if cursym.Text.Mark&LEAF != 0 {
if autosize == 0 {
p.As = AB
p.From = zprg_obj5.From
if p.To.Sym != nil { // retjmp
p.To.Typ = D_BRANCH
} else {
p.To.Typ = D_OREG
p.To.Offset = 0
p.To.Reg = REGLINK
}
break
}
}
if cursym.Text.Reg&liblink.WRAPPER != 0 {
var scond int
// Preserve original RET's cond, to allow RET.EQ
// in the implementation of reflect.call.
scond = p.Scond
p.Scond = C_SCOND_NONE
// g->panicwrap -= autosize;
// MOVW panicwrap_offset(g), R3
// SUB $autosize, R3
// MOVW R3 panicwrap_offset(g)
p.As = AMOVW
p.From.Typ = D_OREG
p.From.Reg = REGG
p.From.Offset = 2 * ctxt.Arch.Ptrsize
p.To.Typ = D_REG
p.To.Reg = 3
p = liblink.Appendp(ctxt, p)
p.As = ASUB
p.From.Typ = D_CONST
p.From.Offset = autosize
p.To.Typ = D_REG
p.To.Reg = 3
p = liblink.Appendp(ctxt, p)
p.As = AMOVW
p.From.Typ = D_REG
p.From.Reg = 3
p.To.Typ = D_OREG
p.To.Reg = REGG
p.To.Offset = 2 * ctxt.Arch.Ptrsize
p = liblink.Appendp(ctxt, p)
p.Scond = scond
}
p.As = AMOVW
p.Scond |= C_PBIT
p.From.Typ = D_OREG
p.From.Offset = autosize
p.From.Reg = REGSP
p.To.Typ = D_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 = liblink.Appendp(ctxt, p)
q2.As = AB
q2.To.Typ = D_BRANCH
q2.To.Sym = p.To.Sym
p.To.Sym = nil
p = q2
}
case AADD:
if p.From.Typ == D_CONST && p.From.Reg == NREG && p.To.Typ == D_REG && p.To.Reg == REGSP {
p.Spadj = -p.From.Offset
}
case ASUB:
if p.From.Typ == D_CONST && p.From.Reg == NREG && p.To.Typ == D_REG && p.To.Reg == REGSP {
p.Spadj = p.From.Offset
}
case ADIV,
ADIVU,
AMOD,
AMODU:
if ctxt.Debugdivmod != 0 {
break
}
if p.From.Typ != D_REG {
break
}
if p.To.Typ != D_REG {
break
}
q1 = p
/* MOV a,4(SP) */
p = liblink.Appendp(ctxt, p)
p.As = AMOVW
p.Lineno = q1.Lineno
p.From.Typ = D_REG
p.From.Reg = q1.From.Reg
p.To.Typ = D_OREG
p.To.Reg = REGSP
p.To.Offset = 4
/* MOV b,REGTMP */
p = liblink.Appendp(ctxt, p)
p.As = AMOVW
p.Lineno = q1.Lineno
p.From.Typ = D_REG
p.From.Reg = q1.Reg
if q1.Reg == NREG {
p.From.Reg = q1.To.Reg
}
p.To.Typ = D_REG
p.To.Reg = REGTMP
p.To.Offset = 0
/* CALL appropriate */
p = liblink.Appendp(ctxt, p)
p.As = ABL
p.Lineno = q1.Lineno
p.To.Typ = D_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
break
}
/* MOV REGTMP, b */
p = liblink.Appendp(ctxt, p)
p.As = AMOVW
p.Lineno = q1.Lineno
p.From.Typ = D_REG
p.From.Reg = REGTMP
p.From.Offset = 0
p.To.Typ = D_REG
p.To.Reg = q1.To.Reg
/* ADD $8,SP */
p = liblink.Appendp(ctxt, p)
p.As = AADD
p.Lineno = q1.Lineno
p.From.Typ = D_CONST
p.From.Reg = NREG
p.From.Offset = 8
p.Reg = NREG
p.To.Typ = D_REG
p.To.Reg = REGSP
p.Spadj = -8
/* Keep saved LR at 0(SP) after SP change. */
/* MOVW 0(SP), REGTMP; MOVW REGTMP, -8!(SP) */
/* TODO: Remove SP adjustments; see issue 6699. */
q1.As = AMOVW
q1.From.Typ = D_OREG
q1.From.Reg = REGSP
q1.From.Offset = 0
q1.Reg = NREG
q1.To.Typ = D_REG
q1.To.Reg = REGTMP
/* SUB $8,SP */
q1 = liblink.Appendp(ctxt, q1)
q1.As = AMOVW
q1.From.Typ = D_REG
q1.From.Reg = REGTMP
q1.Reg = NREG
q1.To.Typ = D_OREG
q1.To.Reg = REGSP
q1.To.Offset = -8
q1.Scond |= C_WBIT
q1.Spadj = 8
case AMOVW:
if (p.Scond&C_WBIT != 0) && p.To.Typ == D_OREG && p.To.Reg == REGSP {
p.Spadj = -p.To.Offset
}
if (p.Scond&C_PBIT != 0) && p.From.Typ == D_OREG && p.From.Reg == REGSP && p.To.Reg != REGPC {
p.Spadj = -p.From.Offset
}
if p.From.Typ == D_CONST && p.From.Reg == REGSP && p.To.Typ == D_REG && p.To.Reg == REGSP {
p.Spadj = -p.From.Offset
}
break
}
}
}