Exemple #1
0
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
	cnt := hi - lo
	if cnt == 0 {
		return p
	}
	if cnt < int64(4*gc.Widthptr) {
		for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
			p = appendpp(p, ppc64.AMOVD, obj.TYPE_REG, ppc64.REGZERO, 0, obj.TYPE_MEM, ppc64.REGSP, 8+frame+lo+i)
		}
	} else if cnt <= int64(128*gc.Widthptr) {
		p = appendpp(p, ppc64.AADD, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, ppc64.REGRT1, 0)
		p.Reg = ppc64.REGSP
		p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
		f := gc.Sysfunc("duffzero")
		gc.Naddr(&p.To, f)
		gc.Afunclit(&p.To, f)
		p.To.Offset = 4 * (128 - cnt/int64(gc.Widthptr))
	} else {
		p = appendpp(p, ppc64.AMOVD, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, ppc64.REGTMP, 0)
		p = appendpp(p, ppc64.AADD, obj.TYPE_REG, ppc64.REGTMP, 0, obj.TYPE_REG, ppc64.REGRT1, 0)
		p.Reg = ppc64.REGSP
		p = appendpp(p, ppc64.AMOVD, obj.TYPE_CONST, 0, cnt, obj.TYPE_REG, ppc64.REGTMP, 0)
		p = appendpp(p, ppc64.AADD, obj.TYPE_REG, ppc64.REGTMP, 0, obj.TYPE_REG, ppc64.REGRT2, 0)
		p.Reg = ppc64.REGRT1
		p = appendpp(p, ppc64.AMOVDU, obj.TYPE_REG, ppc64.REGZERO, 0, obj.TYPE_MEM, ppc64.REGRT1, int64(gc.Widthptr))
		p1 := p
		p = appendpp(p, ppc64.ACMP, obj.TYPE_REG, ppc64.REGRT1, 0, obj.TYPE_REG, ppc64.REGRT2, 0)
		p = appendpp(p, ppc64.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
		gc.Patch(p, p1)
	}

	return p
}
Exemple #2
0
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
	cnt := hi - lo
	if cnt == 0 {
		return p
	}
	if cnt < int64(4*gc.Widthptr) {
		for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
			p = appendpp(p, mips.AMOVV, obj.TYPE_REG, mips.REGZERO, 0, obj.TYPE_MEM, mips.REGSP, 8+frame+lo+i)
		}
		// TODO(dfc): https://golang.org/issue/12108
		// If DUFFZERO is used inside a tail call (see genwrapper) it will
		// overwrite the link register.
	} else if false && cnt <= int64(128*gc.Widthptr) {
		p = appendpp(p, mips.AADDV, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, mips.REGRT1, 0)
		p.Reg = mips.REGSP
		p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
		f := gc.Sysfunc("duffzero")
		gc.Naddr(&p.To, f)
		gc.Afunclit(&p.To, f)
		p.To.Offset = 8 * (128 - cnt/int64(gc.Widthptr))
	} else {
		//	ADDV	$(8+frame+lo-8), SP, r1
		//	ADDV	$cnt, r1, r2
		// loop:
		//	MOVV	R0, (Widthptr)r1
		//	ADDV	$Widthptr, r1
		//	BNE		r1, r2, loop
		p = appendpp(p, mips.AADDV, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, mips.REGRT1, 0)
		p.Reg = mips.REGSP
		p = appendpp(p, mips.AADDV, obj.TYPE_CONST, 0, cnt, obj.TYPE_REG, mips.REGRT2, 0)
		p.Reg = mips.REGRT1
		p = appendpp(p, mips.AMOVV, obj.TYPE_REG, mips.REGZERO, 0, obj.TYPE_MEM, mips.REGRT1, int64(gc.Widthptr))
		p1 := p
		p = appendpp(p, mips.AADDV, obj.TYPE_CONST, 0, int64(gc.Widthptr), obj.TYPE_REG, mips.REGRT1, 0)
		p = appendpp(p, mips.ABNE, obj.TYPE_REG, mips.REGRT1, 0, obj.TYPE_BRANCH, 0, 0)
		p.Reg = mips.REGRT2
		gc.Patch(p, p1)
	}

	return p
}
Exemple #3
0
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64, r0 *uint32) *obj.Prog {
	cnt := hi - lo
	if cnt == 0 {
		return p
	}
	if *r0 == 0 {
		p = appendpp(p, arm.AMOVW, obj.TYPE_CONST, 0, 0, obj.TYPE_REG, arm.REG_R0, 0)
		*r0 = 1
	}

	if cnt < int64(4*gc.Widthptr) {
		for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
			p = appendpp(p, arm.AMOVW, obj.TYPE_REG, arm.REG_R0, 0, obj.TYPE_MEM, arm.REGSP, int32(4+frame+lo+i))
		}
	} else if !gc.Nacl && (cnt <= int64(128*gc.Widthptr)) {
		p = appendpp(p, arm.AADD, obj.TYPE_CONST, 0, int32(4+frame+lo), obj.TYPE_REG, arm.REG_R1, 0)
		p.Reg = arm.REGSP
		p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
		f := gc.Sysfunc("duffzero")
		gc.Naddr(&p.To, f)
		gc.Afunclit(&p.To, f)
		p.To.Offset = 4 * (128 - cnt/int64(gc.Widthptr))
	} else {
		p = appendpp(p, arm.AADD, obj.TYPE_CONST, 0, int32(4+frame+lo), obj.TYPE_REG, arm.REG_R1, 0)
		p.Reg = arm.REGSP
		p = appendpp(p, arm.AADD, obj.TYPE_CONST, 0, int32(cnt), obj.TYPE_REG, arm.REG_R2, 0)
		p.Reg = arm.REG_R1
		p = appendpp(p, arm.AMOVW, obj.TYPE_REG, arm.REG_R0, 0, obj.TYPE_MEM, arm.REG_R1, 4)
		p1 := p
		p.Scond |= arm.C_PBIT
		p = appendpp(p, arm.ACMP, obj.TYPE_REG, arm.REG_R1, 0, obj.TYPE_NONE, 0, 0)
		p.Reg = arm.REG_R2
		p = appendpp(p, arm.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
		gc.Patch(p, p1)
	}

	return p
}
Exemple #4
0
func zerorange(p *obj.Prog, frame int64, lo int64, hi int64) *obj.Prog {
	cnt := hi - lo
	if cnt == 0 {
		return p
	}
	if cnt < int64(4*gc.Widthptr) {
		for i := int64(0); i < cnt; i += int64(gc.Widthptr) {
			p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGSP, 8+frame+lo+i)
		}
	} else if cnt <= int64(128*gc.Widthptr) && !darwin { // darwin ld64 cannot handle BR26 reloc with non-zero addend
		p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGSP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
		p = appendpp(p, arm64.AADD, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, arm64.REGRT1, 0)
		p.Reg = arm64.REGRT1
		p = appendpp(p, obj.ADUFFZERO, obj.TYPE_NONE, 0, 0, obj.TYPE_MEM, 0, 0)
		f := gc.Sysfunc("duffzero")
		gc.Naddr(&p.To, f)
		gc.Afunclit(&p.To, f)
		p.To.Offset = 4 * (128 - cnt/int64(gc.Widthptr))
	} else {
		p = appendpp(p, arm64.AMOVD, obj.TYPE_CONST, 0, 8+frame+lo-8, obj.TYPE_REG, arm64.REGTMP, 0)
		p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGSP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
		p = appendpp(p, arm64.AADD, obj.TYPE_REG, arm64.REGTMP, 0, obj.TYPE_REG, arm64.REGRT1, 0)
		p.Reg = arm64.REGRT1
		p = appendpp(p, arm64.AMOVD, obj.TYPE_CONST, 0, cnt, obj.TYPE_REG, arm64.REGTMP, 0)
		p = appendpp(p, arm64.AADD, obj.TYPE_REG, arm64.REGTMP, 0, obj.TYPE_REG, arm64.REGRT2, 0)
		p.Reg = arm64.REGRT1
		p = appendpp(p, arm64.AMOVD, obj.TYPE_REG, arm64.REGZERO, 0, obj.TYPE_MEM, arm64.REGRT1, int64(gc.Widthptr))
		p.Scond = arm64.C_XPRE
		p1 := p
		p = appendpp(p, arm64.ACMP, obj.TYPE_REG, arm64.REGRT1, 0, obj.TYPE_NONE, 0, 0)
		p.Reg = arm64.REGRT2
		p = appendpp(p, arm64.ABNE, obj.TYPE_NONE, 0, 0, obj.TYPE_BRANCH, 0, 0)
		gc.Patch(p, p1)
	}

	return p
}
Exemple #5
0
func clearfat(nl *gc.Node) {
	/* clear a fat object */
	if gc.Debug['g'] != 0 {
		fmt.Printf("clearfat %v (%v, size: %d)\n", nl, nl.Type, nl.Type.Width)
	}

	w := uint64(uint64(nl.Type.Width))

	// Avoid taking the address for simple enough types.
	if gc.Componentgen(nil, nl) {
		return
	}

	c := uint64(w % 8) // bytes
	q := uint64(w / 8) // dwords

	if gc.Reginuse(ppc64.REGRT1) {
		gc.Fatal("%v in use during clearfat", obj.Rconv(ppc64.REGRT1))
	}

	var r0 gc.Node
	gc.Nodreg(&r0, gc.Types[gc.TUINT64], ppc64.REGZERO)
	var dst gc.Node
	gc.Nodreg(&dst, gc.Types[gc.Tptr], ppc64.REGRT1)
	gc.Regrealloc(&dst)
	gc.Agen(nl, &dst)

	var boff uint64
	if q > 128 {
		p := gins(ppc64.ASUB, nil, &dst)
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = 8

		var end gc.Node
		gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
		p = gins(ppc64.AMOVD, &dst, &end)
		p.From.Type = obj.TYPE_ADDR
		p.From.Offset = int64(q * 8)

		p = gins(ppc64.AMOVDU, &r0, &dst)
		p.To.Type = obj.TYPE_MEM
		p.To.Offset = 8
		pl := (*obj.Prog)(p)

		p = gins(ppc64.ACMP, &dst, &end)
		gc.Patch(gc.Gbranch(ppc64.ABNE, nil, 0), pl)

		gc.Regfree(&end)

		// The loop leaves R3 on the last zeroed dword
		boff = 8
	} else if q >= 4 {
		p := gins(ppc64.ASUB, nil, &dst)
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = 8
		f := (*gc.Node)(gc.Sysfunc("duffzero"))
		p = gins(obj.ADUFFZERO, nil, f)
		gc.Afunclit(&p.To, f)

		// 4 and 128 = magic constants: see ../../runtime/asm_ppc64x.s
		p.To.Offset = int64(4 * (128 - q))

		// duffzero leaves R3 on the last zeroed dword
		boff = 8
	} else {
		var p *obj.Prog
		for t := uint64(0); t < q; t++ {
			p = gins(ppc64.AMOVD, &r0, &dst)
			p.To.Type = obj.TYPE_MEM
			p.To.Offset = int64(8 * t)
		}

		boff = 8 * q
	}

	var p *obj.Prog
	for t := uint64(0); t < c; t++ {
		p = gins(ppc64.AMOVB, &r0, &dst)
		p.To.Type = obj.TYPE_MEM
		p.To.Offset = int64(t + boff)
	}

	gc.Regfree(&dst)
}
Exemple #6
0
func blockcopy(n, res *gc.Node, osrc, odst, w int64) {
	// determine alignment.
	// want to avoid unaligned access, so have to use
	// smaller operations for less aligned types.
	// for example moving [4]byte must use 4 MOVB not 1 MOVW.
	align := int(n.Type.Align)

	var op int
	switch align {
	default:
		gc.Fatalf("sgen: invalid alignment %d for %v", align, n.Type)

	case 1:
		op = arm.AMOVB

	case 2:
		op = arm.AMOVH

	case 4:
		op = arm.AMOVW
	}

	if w%int64(align) != 0 {
		gc.Fatalf("sgen: unaligned size %d (align=%d) for %v", w, align, n.Type)
	}
	c := int32(w / int64(align))

	if osrc%int64(align) != 0 || odst%int64(align) != 0 {
		gc.Fatalf("sgen: unaligned offset src %d or dst %d (align %d)", osrc, odst, align)
	}

	// if we are copying forward on the stack and
	// the src and dst overlap, then reverse direction
	dir := align
	if osrc < odst && int64(odst) < int64(osrc)+w {
		dir = -dir
	}

	if op == arm.AMOVW && !gc.Nacl && dir > 0 && c >= 4 && c <= 128 {
		var r0 gc.Node
		r0.Op = gc.OREGISTER
		r0.Reg = arm.REG_R0
		var r1 gc.Node
		r1.Op = gc.OREGISTER
		r1.Reg = arm.REG_R0 + 1
		var r2 gc.Node
		r2.Op = gc.OREGISTER
		r2.Reg = arm.REG_R0 + 2

		var src gc.Node
		gc.Regalloc(&src, gc.Types[gc.Tptr], &r1)
		var dst gc.Node
		gc.Regalloc(&dst, gc.Types[gc.Tptr], &r2)
		if n.Ullman >= res.Ullman {
			// eval n first
			gc.Agen(n, &src)

			if res.Op == gc.ONAME {
				gc.Gvardef(res)
			}
			gc.Agen(res, &dst)
		} else {
			// eval res first
			if res.Op == gc.ONAME {
				gc.Gvardef(res)
			}
			gc.Agen(res, &dst)
			gc.Agen(n, &src)
		}

		var tmp gc.Node
		gc.Regalloc(&tmp, gc.Types[gc.Tptr], &r0)
		f := gc.Sysfunc("duffcopy")
		p := gins(obj.ADUFFCOPY, nil, f)
		gc.Afunclit(&p.To, f)

		// 8 and 128 = magic constants: see ../../runtime/asm_arm.s
		p.To.Offset = 8 * (128 - int64(c))

		gc.Regfree(&tmp)
		gc.Regfree(&src)
		gc.Regfree(&dst)
		return
	}

	var dst gc.Node
	var src gc.Node
	if n.Ullman >= res.Ullman {
		gc.Agenr(n, &dst, res) // temporarily use dst
		gc.Regalloc(&src, gc.Types[gc.Tptr], nil)
		gins(arm.AMOVW, &dst, &src)
		if res.Op == gc.ONAME {
			gc.Gvardef(res)
		}
		gc.Agen(res, &dst)
	} else {
		if res.Op == gc.ONAME {
			gc.Gvardef(res)
		}
		gc.Agenr(res, &dst, res)
		gc.Agenr(n, &src, nil)
	}

	var tmp gc.Node
	gc.Regalloc(&tmp, gc.Types[gc.TUINT32], nil)

	// set up end marker
	var nend gc.Node

	if c >= 4 {
		gc.Regalloc(&nend, gc.Types[gc.TUINT32], nil)

		p := gins(arm.AMOVW, &src, &nend)
		p.From.Type = obj.TYPE_ADDR
		if dir < 0 {
			p.From.Offset = int64(dir)
		} else {
			p.From.Offset = w
		}
	}

	// move src and dest to the end of block if necessary
	if dir < 0 {
		p := gins(arm.AMOVW, &src, &src)
		p.From.Type = obj.TYPE_ADDR
		p.From.Offset = w + int64(dir)

		p = gins(arm.AMOVW, &dst, &dst)
		p.From.Type = obj.TYPE_ADDR
		p.From.Offset = w + int64(dir)
	}

	// move
	if c >= 4 {
		p := gins(op, &src, &tmp)
		p.From.Type = obj.TYPE_MEM
		p.From.Offset = int64(dir)
		p.Scond |= arm.C_PBIT
		ploop := p

		p = gins(op, &tmp, &dst)
		p.To.Type = obj.TYPE_MEM
		p.To.Offset = int64(dir)
		p.Scond |= arm.C_PBIT

		p = gins(arm.ACMP, &src, nil)
		raddr(&nend, p)

		gc.Patch(gc.Gbranch(arm.ABNE, nil, 0), ploop)
		gc.Regfree(&nend)
	} else {
		var p *obj.Prog
		for {
			tmp14 := c
			c--
			if tmp14 <= 0 {
				break
			}
			p = gins(op, &src, &tmp)
			p.From.Type = obj.TYPE_MEM
			p.From.Offset = int64(dir)
			p.Scond |= arm.C_PBIT

			p = gins(op, &tmp, &dst)
			p.To.Type = obj.TYPE_MEM
			p.To.Offset = int64(dir)
			p.Scond |= arm.C_PBIT
		}
	}

	gc.Regfree(&dst)
	gc.Regfree(&src)
	gc.Regfree(&tmp)
}
Exemple #7
0
func clearfat(nl *gc.Node) {
	/* clear a fat object */
	if gc.Debug['g'] != 0 {
		fmt.Printf("clearfat %v (%v, size: %d)\n", nl, nl.Type, nl.Type.Width)
	}

	w := uint64(uint64(nl.Type.Width))

	// Avoid taking the address for simple enough types.
	if gc.Componentgen(nil, nl) {
		return
	}

	c := uint64(w % 8) // bytes
	q := uint64(w / 8) // dwords

	var r0 gc.Node
	gc.Nodreg(&r0, gc.Types[gc.TUINT64], arm64.REGZERO)
	var dst gc.Node

	// REGRT1 is reserved on arm64, see arm64/gsubr.go.
	gc.Nodreg(&dst, gc.Types[gc.Tptr], arm64.REGRT1)
	gc.Agen(nl, &dst)

	var boff uint64
	if q > 128 {
		p := gins(arm64.ASUB, nil, &dst)
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = 8

		var end gc.Node
		gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
		p = gins(arm64.AMOVD, &dst, &end)
		p.From.Type = obj.TYPE_ADDR
		p.From.Offset = int64(q * 8)

		p = gins(arm64.AMOVD, &r0, &dst)
		p.To.Type = obj.TYPE_MEM
		p.To.Offset = 8
		p.Scond = arm64.C_XPRE
		pl := (*obj.Prog)(p)

		p = gcmp(arm64.ACMP, &dst, &end)
		gc.Patch(gc.Gbranch(arm64.ABNE, nil, 0), pl)

		gc.Regfree(&end)

		// The loop leaves R16 on the last zeroed dword
		boff = 8
	} else if q >= 4 && !darwin { // darwin ld64 cannot handle BR26 reloc with non-zero addend
		p := gins(arm64.ASUB, nil, &dst)
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = 8
		f := (*gc.Node)(gc.Sysfunc("duffzero"))
		p = gins(obj.ADUFFZERO, nil, f)
		gc.Afunclit(&p.To, f)

		// 4 and 128 = magic constants: see ../../runtime/asm_arm64x.s
		p.To.Offset = int64(4 * (128 - q))

		// duffzero leaves R16 on the last zeroed dword
		boff = 8
	} else {
		var p *obj.Prog
		for t := uint64(0); t < q; t++ {
			p = gins(arm64.AMOVD, &r0, &dst)
			p.To.Type = obj.TYPE_MEM
			p.To.Offset = int64(8 * t)
		}

		boff = 8 * q
	}

	var p *obj.Prog
	for t := uint64(0); t < c; t++ {
		p = gins(arm64.AMOVB, &r0, &dst)
		p.To.Type = obj.TYPE_MEM
		p.To.Offset = int64(t + boff)
	}
}
Exemple #8
0
func clearfat(nl *gc.Node) {
	/* clear a fat object */
	if gc.Debug['g'] != 0 {
		fmt.Printf("clearfat %v (%v, size: %d)\n", nl, nl.Type, nl.Type.Width)
	}

	w := uint64(nl.Type.Width)

	// Avoid taking the address for simple enough types.
	if gc.Componentgen(nil, nl) {
		return
	}

	c := w % 8 // bytes
	q := w / 8 // dwords

	if gc.Reginuse(mips.REGRT1) {
		gc.Fatalf("%v in use during clearfat", obj.Rconv(mips.REGRT1))
	}

	var r0 gc.Node
	gc.Nodreg(&r0, gc.Types[gc.TUINT64], mips.REGZERO)
	var dst gc.Node
	gc.Nodreg(&dst, gc.Types[gc.Tptr], mips.REGRT1)
	gc.Regrealloc(&dst)
	gc.Agen(nl, &dst)

	var boff uint64
	if q > 128 {
		p := gins(mips.ASUBV, nil, &dst)
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = 8

		var end gc.Node
		gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
		p = gins(mips.AMOVV, &dst, &end)
		p.From.Type = obj.TYPE_ADDR
		p.From.Offset = int64(q * 8)

		p = gins(mips.AMOVV, &r0, &dst)
		p.To.Type = obj.TYPE_MEM
		p.To.Offset = 8
		pl := p

		p = gins(mips.AADDV, nil, &dst)
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = 8

		gc.Patch(ginsbranch(mips.ABNE, nil, &dst, &end, 0), pl)

		gc.Regfree(&end)

		// The loop leaves R1 on the last zeroed dword
		boff = 8
		// TODO(dfc): https://golang.org/issue/12108
		// If DUFFZERO is used inside a tail call (see genwrapper) it will
		// overwrite the link register.
	} else if false && q >= 4 {
		p := gins(mips.ASUBV, nil, &dst)
		p.From.Type = obj.TYPE_CONST
		p.From.Offset = 8
		f := gc.Sysfunc("duffzero")
		p = gins(obj.ADUFFZERO, nil, f)
		gc.Afunclit(&p.To, f)

		// 8 and 128 = magic constants: see ../../runtime/asm_mips64x.s
		p.To.Offset = int64(8 * (128 - q))

		// duffzero leaves R1 on the last zeroed dword
		boff = 8
	} else {
		var p *obj.Prog
		for t := uint64(0); t < q; t++ {
			p = gins(mips.AMOVV, &r0, &dst)
			p.To.Type = obj.TYPE_MEM
			p.To.Offset = int64(8 * t)
		}

		boff = 8 * q
	}

	var p *obj.Prog
	for t := uint64(0); t < c; t++ {
		p = gins(mips.AMOVB, &r0, &dst)
		p.To.Type = obj.TYPE_MEM
		p.To.Offset = int64(t + boff)
	}

	gc.Regfree(&dst)
}
Exemple #9
0
func clearfat(nl *gc.Node) {
	/* clear a fat object */
	if gc.Debug['g'] != 0 {
		gc.Dump("\nclearfat", nl)
	}

	w := uint32(nl.Type.Width)

	// Avoid taking the address for simple enough types.
	if gc.Componentgen(nil, nl) {
		return
	}

	c := w % 4 // bytes
	q := w / 4 // quads

	var r0 gc.Node
	r0.Op = gc.OREGISTER

	r0.Reg = arm.REG_R0
	var r1 gc.Node
	r1.Op = gc.OREGISTER
	r1.Reg = arm.REG_R1
	var dst gc.Node
	gc.Regalloc(&dst, gc.Types[gc.Tptr], &r1)
	gc.Agen(nl, &dst)
	var nc gc.Node
	gc.Nodconst(&nc, gc.Types[gc.TUINT32], 0)
	var nz gc.Node
	gc.Regalloc(&nz, gc.Types[gc.TUINT32], &r0)
	gc.Cgen(&nc, &nz)

	if q > 128 {
		var end gc.Node
		gc.Regalloc(&end, gc.Types[gc.Tptr], nil)
		p := gins(arm.AMOVW, &dst, &end)
		p.From.Type = obj.TYPE_ADDR
		p.From.Offset = int64(q) * 4

		p = gins(arm.AMOVW, &nz, &dst)
		p.To.Type = obj.TYPE_MEM
		p.To.Offset = 4
		p.Scond |= arm.C_PBIT
		pl := p

		p = gins(arm.ACMP, &dst, nil)
		raddr(&end, p)
		gc.Patch(gc.Gbranch(arm.ABNE, nil, 0), pl)

		gc.Regfree(&end)
	} else if q >= 4 && !gc.Nacl {
		f := gc.Sysfunc("duffzero")
		p := gins(obj.ADUFFZERO, nil, f)
		gc.Afunclit(&p.To, f)

		// 4 and 128 = magic constants: see ../../runtime/asm_arm.s
		p.To.Offset = 4 * (128 - int64(q))
	} else {
		var p *obj.Prog
		for q > 0 {
			p = gins(arm.AMOVW, &nz, &dst)
			p.To.Type = obj.TYPE_MEM
			p.To.Offset = 4
			p.Scond |= arm.C_PBIT

			//print("1. %v\n", p);
			q--
		}
	}

	var p *obj.Prog
	for c > 0 {
		p = gins(arm.AMOVB, &nz, &dst)
		p.To.Type = obj.TYPE_MEM
		p.To.Offset = 1
		p.Scond |= arm.C_PBIT

		//print("2. %v\n", p);
		c--
	}

	gc.Regfree(&dst)
	gc.Regfree(&nz)
}