Golang Prog.As方法代碼示例

func stacksplitPost(ctxt *obj.Link, p *obj.Prog, pPre *obj.Prog, pPreempt *obj.Prog) *obj.Prog {

	// MOVD	LR, R5
	p = obj.Appendp(ctxt, p)
	pPre.Pcond = p
	p.As = AMOVD
	p.From.Type = obj.TYPE_REG
	p.From.Reg = REG_LR
	p.To.Type = obj.TYPE_REG
	p.To.Reg = REG_R5
	if pPreempt != nil {
		pPreempt.Pcond = p
	}

	// BL	runtime.morestack(SB)
	p = obj.Appendp(ctxt, p)

	p.As = ABL
	p.To.Type = obj.TYPE_BRANCH
	if ctxt.Cursym.Cfunc {
		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)
	}

	// BR	start
	p = obj.Appendp(ctxt, p)

	p.As = ABR
	p.To.Type = obj.TYPE_BRANCH
	p.Pcond = ctxt.Cursym.Text.Link
	return p
}

// 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(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 = 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 Prog(as obj.As) *obj.Prog {
	var p *obj.Prog

	if as == obj.AGLOBL {
		if ddumped {
			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 && Debug['K'] != 0 {
		Warn("prog: line 0")
	}

	p.As = as
	p.Lineno = lineno
	return p
}

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)
}

/*
// instruction scheduling
	if(debug['Q'] == 0)
		return;

	curtext = nil;
	q = nil;	// p - 1
	q1 = firstp;	// top of block
	o = 0;		// count of instructions
	for(p = firstp; p != nil; p = p1) {
		p1 = p->link;
		o++;
		if(p->mark & NOSCHED){
			if(q1 != p){
				sched(q1, q);
			}
			for(; p != nil; p = p->link){
				if(!(p->mark & NOSCHED))
					break;
				q = p;
			}
			p1 = p;
			q1 = p;
			o = 0;
			continue;
		}
		if(p->mark & (LABEL|SYNC)) {
			if(q1 != p)
				sched(q1, q);
			q1 = p;
			o = 1;
		}
		if(p->mark & (BRANCH|SYNC)) {
			sched(q1, p);
			q1 = p1;
			o = 0;
		}
		if(o >= NSCHED) {
			sched(q1, p);
			q1 = p1;
			o = 0;
		}
		q = p;
	}
*/
func stacksplitPre(ctxt *obj.Link, p *obj.Prog, framesize int32) (*obj.Prog, *obj.Prog) {
	var q *obj.Prog

	// MOVD	g_stackguard(g), R3
	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 {
		p.From.Offset = 3 * int64(ctxt.Arch.PtrSize) // G.stackguard1
	}
	p.To.Type = obj.TYPE_REG
	p.To.Reg = REG_R3

	q = nil
	if framesize <= obj.StackSmall {
		// small stack: SP < stackguard
		//	CMP	stackguard, SP

		//p.To.Type = obj.TYPE_REG
		//p.To.Reg = REGSP

		// q1: BLT	done

		p = obj.Appendp(ctxt, p)
		//q1 = p
		p.From.Type = obj.TYPE_REG
		p.From.Reg = REG_R3
		p.Reg = REGSP
		p.As = ACMPUBGE
		p.To.Type = obj.TYPE_BRANCH
		//p = obj.Appendp(ctxt, p)

		//p.As = ACMPU
		//p.From.Type = obj.TYPE_REG
		//p.From.Reg = REG_R3
		//p.To.Type = obj.TYPE_REG
		//p.To.Reg = REGSP

		//p = obj.Appendp(ctxt, p)
		//p.As = ABGE
		//p.To.Type = obj.TYPE_BRANCH

	} else if framesize <= obj.StackBig {
		// large stack: SP-framesize < stackguard-StackSmall
		//	ADD $-framesize, SP, R4
		//	CMP stackguard, R4
		p = obj.Appendp(ctxt, p)

		p.As = AADD
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = int64(-framesize)
		p.Reg = REGSP
//.........這裏部分代碼省略.........

func Clearp(p *obj.Prog) {
	obj.Nopout(p)
	p.As = obj.AEND
	p.Pc = int64(pcloc)
	pcloc++
}

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 {
		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 {
	}
//.........這裏部分代碼省略.........

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 p1 *obj.Prog
	var p2 *obj.Prog
	var q2 *obj.Prog
	for p := cursym.Text; p != nil; p = p.Link {
		o := 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
				}
//.........這裏部分代碼省略.........

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 {
//.........這裏部分代碼省略.........

func xfol(ctxt *obj.Link, p *obj.Prog, last **obj.Prog) {
	var q *obj.Prog
	var r *obj.Prog
	var b obj.As

	for p != nil {
		a := p.As
		if a == ABR {
			q = p.Pcond
			if (p.Mark&NOSCHED != 0) || q != nil && (q.Mark&NOSCHED != 0) {
				p.Mark |= FOLL
				(*last).Link = p
				*last = p
				(*last).Pc = pc_cnt
				pc_cnt += 1
				p = p.Link
				xfol(ctxt, p, last)
				p = q
				if p != nil && p.Mark&FOLL == 0 {
					continue
				}
				return
			}

			if q != nil {
				p.Mark |= FOLL
				p = q
				if p.Mark&FOLL == 0 {
					continue
				}
			}
		}

		if p.Mark&FOLL != 0 {
			q = p
			for i := 0; i < 4; i, q = i+1, q.Link {
				if q == *last || (q.Mark&NOSCHED != 0) {
					break
				}
				b = 0 /* set */
				a = q.As
				if a == obj.ANOP {
					i--
					continue
				}
				if a != ABR && a != obj.ARET {
					if q.Pcond == nil || (q.Pcond.Mark&FOLL != 0) {
						continue
					}
					b = relinv(a)
					if b == 0 {
						continue
					}
				}

				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
						(*last).Pc = pc_cnt
						pc_cnt += 1
						continue
					}

					(*last).Link = r
					*last = r
					(*last).Pc = pc_cnt
					pc_cnt += 1
					if a == ABR || a == obj.ARET {
						return
					}
					r.As = b
					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 = ABR
			q = ctxt.NewProg()
			q.As = a
			q.Lineno = p.Lineno
			q.To.Type = obj.TYPE_BRANCH
			q.To.Offset = p.Pc
			q.Pcond = p
			p = q
		}
//.........這裏部分代碼省略.........

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
//.........這裏部分代碼省略.........

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 && ctxt.Framepointer_enabled && autoffset > 0 && p.From3.Offset&obj.NOFRAME == 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)
//.........這裏部分代碼省略.........

// 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 obj.As
	var 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 [email protected], $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 [email protected], Rx
		// $MOV $sym+<off>, Rx becomes $MOV [email protected], Rx; $LEA <off>(Rx), Rx
		// On 386 only, more complicated things like PUSHL $sym become $MOV [email protected], 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 = lea
			q.From.Type = obj.TYPE_MEM
			q.From.Reg = p.To.Reg
			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
//.........這裏部分代碼省略.........

// 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 := 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 arithmetic 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
				}
			}
		}
	}
}

// fuseMultiple merges memory loads and stores into load multiple and
// store multiple operations.
//
// Looks for this pattern (sequence of loads or stores):
// 	MOVD	R1, 0(R15)
//	MOVD	R2, 8(R15)
//	MOVD	R3, 16(R15)
// Replaces with:
//	STMG	R1, R3, 0(R15)
func fuseMultiple(r *gc.Flow) int {
	n := 0
	var fused *obj.Prog
	for ; r != nil; r = r.Link {
		// If there is a branch into the instruction stream then
		// we can't fuse into previous instructions.
		if gc.Uniqp(r) == nil {
			fused = nil
		}

		p := r.Prog

		isStore := isGPR(&p.From) && isBDMem(&p.To)
		isLoad := isGPR(&p.To) && isBDMem(&p.From)

		// are we a candidate?
		size := int64(0)
		switch p.As {
		default:
			fused = nil
			continue
		case obj.ANOP:
			// skip over nops
			continue
		case s390x.AMOVW, s390x.AMOVWZ:
			size = 4
			// TODO(mundaym): 32-bit load multiple is currently not supported
			// as it requires sign/zero extension.
			if !isStore {
				fused = nil
				continue
			}
		case s390x.AMOVD:
			size = 8
			if !isLoad && !isStore {
				fused = nil
				continue
			}
		}

		// If we merge two loads/stores with different source/destination Nodes
		// then we will lose a reference the second Node which means that the
		// compiler might mark the Node as unused and free its slot on the stack.
		// TODO(mundaym): allow this by adding a dummy reference to the Node.
		if fused == nil ||
			fused.From.Node != p.From.Node ||
			fused.From.Type != p.From.Type ||
			fused.To.Node != p.To.Node ||
			fused.To.Type != p.To.Type {
			fused = p
			continue
		}

		// check two addresses
		ca := func(a, b *obj.Addr, offset int64) bool {
			return a.Reg == b.Reg && a.Offset+offset == b.Offset &&
				a.Sym == b.Sym && a.Name == b.Name
		}

		switch fused.As {
		default:
			fused = p
		case s390x.AMOVW, s390x.AMOVWZ:
			if size == 4 && fused.From.Reg+1 == p.From.Reg && ca(&fused.To, &p.To, 4) {
				fused.As = s390x.ASTMY
				fused.Reg = p.From.Reg
				excise(r)
				n++
			} else {
				fused = p
			}
		case s390x.AMOVD:
			if size == 8 && fused.From.Reg+1 == p.From.Reg && ca(&fused.To, &p.To, 8) {
				fused.As = s390x.ASTMG
				fused.Reg = p.From.Reg
				excise(r)
				n++
			} else if size == 8 && fused.To.Reg+1 == p.To.Reg && ca(&fused.From, &p.From, 8) {
				fused.As = s390x.ALMG
				fused.Reg = fused.To.Reg
				fused.To.Reg = p.To.Reg
				excise(r)
				n++
			} else {
				fused = p
			}
		case s390x.ASTMG, s390x.ASTMY:
			if (fused.As == s390x.ASTMY && size != 4) ||
				(fused.As == s390x.ASTMG && size != 8) {
				fused = p
				continue
//.........這裏部分代碼省略.........