Exemple #1
0
/*
 * generate an addressable node in res, containing the value of n.
 * n is an array index, and might be any size; res width is <= 32-bit.
 * returns Prog* to patch to panic call.
 */
func igenindex(n *gc.Node, res *gc.Node, bounded bool) *obj.Prog {
	if !gc.Is64(n.Type) {
		if n.Addable && (gc.Simtype[n.Etype] == gc.TUINT32 || gc.Simtype[n.Etype] == gc.TINT32) {
			// nothing to do.
			*res = *n
		} else {
			gc.Tempname(res, gc.Types[gc.TUINT32])
			gc.Cgen(n, res)
		}

		return nil
	}

	var tmp gc.Node
	gc.Tempname(&tmp, gc.Types[gc.TINT64])
	gc.Cgen(n, &tmp)
	var lo gc.Node
	var hi gc.Node
	split64(&tmp, &lo, &hi)
	gc.Tempname(res, gc.Types[gc.TUINT32])
	gmove(&lo, res)
	if bounded {
		splitclean()
		return nil
	}

	var zero gc.Node
	gc.Nodconst(&zero, gc.Types[gc.TINT32], 0)
	gins(x86.ACMPL, &hi, &zero)
	splitclean()
	return gc.Gbranch(x86.AJNE, nil, +1)
}
Exemple #2
0
/*
 * generate array index into res.
 * n might be any size; res is 32-bit.
 * returns Prog* to patch to panic call.
 */
func cgenindex(n *gc.Node, res *gc.Node, bounded bool) *obj.Prog {
	if !gc.Is64(n.Type) {
		gc.Cgen(n, res)
		return nil
	}

	var tmp gc.Node
	gc.Tempname(&tmp, gc.Types[gc.TINT64])
	gc.Cgen(n, &tmp)
	var lo gc.Node
	var hi gc.Node
	split64(&tmp, &lo, &hi)
	gmove(&lo, res)
	if bounded {
		splitclean()
		return nil
	}

	var n1 gc.Node
	gc.Regalloc(&n1, gc.Types[gc.TINT32], nil)
	var n2 gc.Node
	gc.Regalloc(&n2, gc.Types[gc.TINT32], nil)
	var zero gc.Node
	gc.Nodconst(&zero, gc.Types[gc.TINT32], 0)
	gmove(&hi, &n1)
	gmove(&zero, &n2)
	gins(arm.ACMP, &n1, &n2)
	gc.Regfree(&n2)
	gc.Regfree(&n1)
	splitclean()
	return gc.Gbranch(arm.ABNE, nil, -1)
}
Exemple #3
0
/*
 * n is a 64-bit value.  fill in lo and hi to refer to its 32-bit halves.
 */
func split64(n *gc.Node, lo *gc.Node, hi *gc.Node) {
	if !gc.Is64(n.Type) {
		gc.Fatalf("split64 %v", n.Type)
	}

	if nsclean >= len(sclean) {
		gc.Fatalf("split64 clean")
	}
	sclean[nsclean].Op = gc.OEMPTY
	nsclean++
	switch n.Op {
	default:
		switch n.Op {
		default:
			var n1 gc.Node
			if !dotaddable(n, &n1) {
				gc.Igen(n, &n1, nil)
				sclean[nsclean-1] = n1
			}

			n = &n1

		case gc.ONAME:
			if n.Class == gc.PPARAMREF {
				var n1 gc.Node
				gc.Cgen(n.Name.Heapaddr, &n1)
				sclean[nsclean-1] = n1
				n = &n1
			}

			// nothing
		case gc.OINDREG:
			break
		}

		*lo = *n
		*hi = *n
		lo.Type = gc.Types[gc.TUINT32]
		if n.Type.Etype == gc.TINT64 {
			hi.Type = gc.Types[gc.TINT32]
		} else {
			hi.Type = gc.Types[gc.TUINT32]
		}
		hi.Xoffset += 4

	case gc.OLITERAL:
		var n1 gc.Node
		n.Convconst(&n1, n.Type)
		i := n1.Int()
		gc.Nodconst(lo, gc.Types[gc.TUINT32], int64(uint32(i)))
		i >>= 32
		if n.Type.Etype == gc.TINT64 {
			gc.Nodconst(hi, gc.Types[gc.TINT32], int64(int32(i)))
		} else {
			gc.Nodconst(hi, gc.Types[gc.TUINT32], int64(uint32(i)))
		}
	}
}
Exemple #4
0
/*
 * generate division according to op, one of:
 *	res = nl / nr
 *	res = nl % nr
 */
func cgen_div(op gc.Op, nl *gc.Node, nr *gc.Node, res *gc.Node) {
	if gc.Is64(nl.Type) {
		gc.Fatalf("cgen_div %v", nl.Type)
	}

	var t *gc.Type
	if gc.Issigned[nl.Type.Etype] {
		t = gc.Types[gc.TINT32]
	} else {
		t = gc.Types[gc.TUINT32]
	}
	var ax gc.Node
	var oldax gc.Node
	savex(x86.REG_AX, &ax, &oldax, res, t)
	var olddx gc.Node
	var dx gc.Node
	savex(x86.REG_DX, &dx, &olddx, res, t)
	dodiv(op, nl, nr, res, &ax, &dx)
	restx(&dx, &olddx)
	restx(&ax, &oldax)
}
Exemple #5
0
/*
 * attempt to generate 64-bit
 *	res = n
 * return 1 on success, 0 if op not handled.
 */
func cgen64(n *gc.Node, res *gc.Node) {
	if res.Op != gc.OINDREG && res.Op != gc.ONAME {
		gc.Dump("n", n)
		gc.Dump("res", res)
		gc.Fatalf("cgen64 %v of %v", gc.Oconv(int(n.Op), 0), gc.Oconv(int(res.Op), 0))
	}

	switch n.Op {
	default:
		gc.Fatalf("cgen64 %v", gc.Oconv(int(n.Op), 0))

	case gc.OMINUS:
		gc.Cgen(n.Left, res)
		var hi1 gc.Node
		var lo1 gc.Node
		split64(res, &lo1, &hi1)
		gins(x86.ANEGL, nil, &lo1)
		gins(x86.AADCL, ncon(0), &hi1)
		gins(x86.ANEGL, nil, &hi1)
		splitclean()
		return

	case gc.OCOM:
		gc.Cgen(n.Left, res)
		var lo1 gc.Node
		var hi1 gc.Node
		split64(res, &lo1, &hi1)
		gins(x86.ANOTL, nil, &lo1)
		gins(x86.ANOTL, nil, &hi1)
		splitclean()
		return

		// binary operators.
	// common setup below.
	case gc.OADD,
		gc.OSUB,
		gc.OMUL,
		gc.OLROT,
		gc.OLSH,
		gc.ORSH,
		gc.OAND,
		gc.OOR,
		gc.OXOR:
		break
	}

	l := n.Left
	r := n.Right
	if !l.Addable {
		var t1 gc.Node
		gc.Tempname(&t1, l.Type)
		gc.Cgen(l, &t1)
		l = &t1
	}

	if r != nil && !r.Addable {
		var t2 gc.Node
		gc.Tempname(&t2, r.Type)
		gc.Cgen(r, &t2)
		r = &t2
	}

	var ax gc.Node
	gc.Nodreg(&ax, gc.Types[gc.TINT32], x86.REG_AX)
	var cx gc.Node
	gc.Nodreg(&cx, gc.Types[gc.TINT32], x86.REG_CX)
	var dx gc.Node
	gc.Nodreg(&dx, gc.Types[gc.TINT32], x86.REG_DX)

	// Setup for binary operation.
	var hi1 gc.Node
	var lo1 gc.Node
	split64(l, &lo1, &hi1)

	var lo2 gc.Node
	var hi2 gc.Node
	if gc.Is64(r.Type) {
		split64(r, &lo2, &hi2)
	}

	// Do op. Leave result in DX:AX.
	switch n.Op {
	// TODO: Constants
	case gc.OADD:
		gins(x86.AMOVL, &lo1, &ax)

		gins(x86.AMOVL, &hi1, &dx)
		gins(x86.AADDL, &lo2, &ax)
		gins(x86.AADCL, &hi2, &dx)

		// TODO: Constants.
	case gc.OSUB:
		gins(x86.AMOVL, &lo1, &ax)

		gins(x86.AMOVL, &hi1, &dx)
		gins(x86.ASUBL, &lo2, &ax)
		gins(x86.ASBBL, &hi2, &dx)

	case gc.OMUL:
		// let's call the next three EX, FX and GX
		var ex, fx, gx gc.Node
		gc.Regalloc(&ex, gc.Types[gc.TPTR32], nil)
		gc.Regalloc(&fx, gc.Types[gc.TPTR32], nil)
		gc.Regalloc(&gx, gc.Types[gc.TPTR32], nil)

		// load args into DX:AX and EX:GX.
		gins(x86.AMOVL, &lo1, &ax)

		gins(x86.AMOVL, &hi1, &dx)
		gins(x86.AMOVL, &lo2, &gx)
		gins(x86.AMOVL, &hi2, &ex)

		// if DX and EX are zero, use 32 x 32 -> 64 unsigned multiply.
		gins(x86.AMOVL, &dx, &fx)

		gins(x86.AORL, &ex, &fx)
		p1 := gc.Gbranch(x86.AJNE, nil, 0)
		gins(x86.AMULL, &gx, nil) // implicit &ax
		p2 := gc.Gbranch(obj.AJMP, nil, 0)
		gc.Patch(p1, gc.Pc)

		// full 64x64 -> 64, from 32x32 -> 64.
		gins(x86.AIMULL, &gx, &dx)

		gins(x86.AMOVL, &ax, &fx)
		gins(x86.AIMULL, &ex, &fx)
		gins(x86.AADDL, &dx, &fx)
		gins(x86.AMOVL, &gx, &dx)
		gins(x86.AMULL, &dx, nil) // implicit &ax
		gins(x86.AADDL, &fx, &dx)
		gc.Patch(p2, gc.Pc)

		gc.Regfree(&ex)
		gc.Regfree(&fx)
		gc.Regfree(&gx)

	// We only rotate by a constant c in [0,64).
	// if c >= 32:
	//	lo, hi = hi, lo
	//	c -= 32
	// if c == 0:
	//	no-op
	// else:
	//	t = hi
	//	shld hi:lo, c
	//	shld lo:t, c
	case gc.OLROT:
		v := uint64(r.Int())

		if v >= 32 {
			// reverse during load to do the first 32 bits of rotate
			v -= 32

			gins(x86.AMOVL, &lo1, &dx)
			gins(x86.AMOVL, &hi1, &ax)
		} else {
			gins(x86.AMOVL, &lo1, &ax)
			gins(x86.AMOVL, &hi1, &dx)
		}

		if v == 0 {
		} else // done
		{
			gins(x86.AMOVL, &dx, &cx)
			p1 := gins(x86.ASHLL, ncon(uint32(v)), &dx)
			p1.From.Index = x86.REG_AX // double-width shift
			p1.From.Scale = 0
			p1 = gins(x86.ASHLL, ncon(uint32(v)), &ax)
			p1.From.Index = x86.REG_CX // double-width shift
			p1.From.Scale = 0
		}

	case gc.OLSH:
		if r.Op == gc.OLITERAL {
			v := uint64(r.Int())
			if v >= 64 {
				if gc.Is64(r.Type) {
					splitclean()
				}
				splitclean()
				split64(res, &lo2, &hi2)
				gins(x86.AMOVL, ncon(0), &lo2)
				gins(x86.AMOVL, ncon(0), &hi2)
				splitclean()
				return
			}

			if v >= 32 {
				if gc.Is64(r.Type) {
					splitclean()
				}
				split64(res, &lo2, &hi2)
				gmove(&lo1, &hi2)
				if v > 32 {
					gins(x86.ASHLL, ncon(uint32(v-32)), &hi2)
				}

				gins(x86.AMOVL, ncon(0), &lo2)
				splitclean()
				splitclean()
				return
			}

			// general shift
			gins(x86.AMOVL, &lo1, &ax)

			gins(x86.AMOVL, &hi1, &dx)
			p1 := gins(x86.ASHLL, ncon(uint32(v)), &dx)
			p1.From.Index = x86.REG_AX // double-width shift
			p1.From.Scale = 0
			gins(x86.ASHLL, ncon(uint32(v)), &ax)
			break
		}

		// load value into DX:AX.
		gins(x86.AMOVL, &lo1, &ax)

		gins(x86.AMOVL, &hi1, &dx)

		// load shift value into register.
		// if high bits are set, zero value.
		var p1 *obj.Prog

		if gc.Is64(r.Type) {
			gins(x86.ACMPL, &hi2, ncon(0))
			p1 = gc.Gbranch(x86.AJNE, nil, +1)
			gins(x86.AMOVL, &lo2, &cx)
		} else {
			cx.Type = gc.Types[gc.TUINT32]
			gmove(r, &cx)
		}

		// if shift count is >=64, zero value
		gins(x86.ACMPL, &cx, ncon(64))

		p2 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
		if p1 != nil {
			gc.Patch(p1, gc.Pc)
		}
		gins(x86.AXORL, &dx, &dx)
		gins(x86.AXORL, &ax, &ax)
		gc.Patch(p2, gc.Pc)

		// if shift count is >= 32, zero low.
		gins(x86.ACMPL, &cx, ncon(32))

		p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
		gins(x86.AMOVL, &ax, &dx)
		gins(x86.ASHLL, &cx, &dx) // SHLL only uses bottom 5 bits of count
		gins(x86.AXORL, &ax, &ax)
		p2 = gc.Gbranch(obj.AJMP, nil, 0)
		gc.Patch(p1, gc.Pc)

		// general shift
		p1 = gins(x86.ASHLL, &cx, &dx)

		p1.From.Index = x86.REG_AX // double-width shift
		p1.From.Scale = 0
		gins(x86.ASHLL, &cx, &ax)
		gc.Patch(p2, gc.Pc)

	case gc.ORSH:
		if r.Op == gc.OLITERAL {
			v := uint64(r.Int())
			if v >= 64 {
				if gc.Is64(r.Type) {
					splitclean()
				}
				splitclean()
				split64(res, &lo2, &hi2)
				if hi1.Type.Etype == gc.TINT32 {
					gmove(&hi1, &lo2)
					gins(x86.ASARL, ncon(31), &lo2)
					gmove(&hi1, &hi2)
					gins(x86.ASARL, ncon(31), &hi2)
				} else {
					gins(x86.AMOVL, ncon(0), &lo2)
					gins(x86.AMOVL, ncon(0), &hi2)
				}

				splitclean()
				return
			}

			if v >= 32 {
				if gc.Is64(r.Type) {
					splitclean()
				}
				split64(res, &lo2, &hi2)
				gmove(&hi1, &lo2)
				if v > 32 {
					gins(optoas(gc.ORSH, hi1.Type), ncon(uint32(v-32)), &lo2)
				}
				if hi1.Type.Etype == gc.TINT32 {
					gmove(&hi1, &hi2)
					gins(x86.ASARL, ncon(31), &hi2)
				} else {
					gins(x86.AMOVL, ncon(0), &hi2)
				}
				splitclean()
				splitclean()
				return
			}

			// general shift
			gins(x86.AMOVL, &lo1, &ax)

			gins(x86.AMOVL, &hi1, &dx)
			p1 := gins(x86.ASHRL, ncon(uint32(v)), &ax)
			p1.From.Index = x86.REG_DX // double-width shift
			p1.From.Scale = 0
			gins(optoas(gc.ORSH, hi1.Type), ncon(uint32(v)), &dx)
			break
		}

		// load value into DX:AX.
		gins(x86.AMOVL, &lo1, &ax)

		gins(x86.AMOVL, &hi1, &dx)

		// load shift value into register.
		// if high bits are set, zero value.
		var p1 *obj.Prog

		if gc.Is64(r.Type) {
			gins(x86.ACMPL, &hi2, ncon(0))
			p1 = gc.Gbranch(x86.AJNE, nil, +1)
			gins(x86.AMOVL, &lo2, &cx)
		} else {
			cx.Type = gc.Types[gc.TUINT32]
			gmove(r, &cx)
		}

		// if shift count is >=64, zero or sign-extend value
		gins(x86.ACMPL, &cx, ncon(64))

		p2 := gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
		if p1 != nil {
			gc.Patch(p1, gc.Pc)
		}
		if hi1.Type.Etype == gc.TINT32 {
			gins(x86.ASARL, ncon(31), &dx)
			gins(x86.AMOVL, &dx, &ax)
		} else {
			gins(x86.AXORL, &dx, &dx)
			gins(x86.AXORL, &ax, &ax)
		}

		gc.Patch(p2, gc.Pc)

		// if shift count is >= 32, sign-extend hi.
		gins(x86.ACMPL, &cx, ncon(32))

		p1 = gc.Gbranch(optoas(gc.OLT, gc.Types[gc.TUINT32]), nil, +1)
		gins(x86.AMOVL, &dx, &ax)
		if hi1.Type.Etype == gc.TINT32 {
			gins(x86.ASARL, &cx, &ax) // SARL only uses bottom 5 bits of count
			gins(x86.ASARL, ncon(31), &dx)
		} else {
			gins(x86.ASHRL, &cx, &ax)
			gins(x86.AXORL, &dx, &dx)
		}

		p2 = gc.Gbranch(obj.AJMP, nil, 0)
		gc.Patch(p1, gc.Pc)

		// general shift
		p1 = gins(x86.ASHRL, &cx, &ax)

		p1.From.Index = x86.REG_DX // double-width shift
		p1.From.Scale = 0
		gins(optoas(gc.ORSH, hi1.Type), &cx, &dx)
		gc.Patch(p2, gc.Pc)

		// make constant the right side (it usually is anyway).
	case gc.OXOR,
		gc.OAND,
		gc.OOR:
		if lo1.Op == gc.OLITERAL {
			nswap(&lo1, &lo2)
			nswap(&hi1, &hi2)
		}

		if lo2.Op == gc.OLITERAL {
			// special cases for constants.
			lv := uint32(lo2.Int())
			hv := uint32(hi2.Int())
			splitclean() // right side
			split64(res, &lo2, &hi2)
			switch n.Op {
			case gc.OXOR:
				gmove(&lo1, &lo2)
				gmove(&hi1, &hi2)
				switch lv {
				case 0:
					break

				case 0xffffffff:
					gins(x86.ANOTL, nil, &lo2)

				default:
					gins(x86.AXORL, ncon(lv), &lo2)
				}

				switch hv {
				case 0:
					break

				case 0xffffffff:
					gins(x86.ANOTL, nil, &hi2)

				default:
					gins(x86.AXORL, ncon(hv), &hi2)
				}

			case gc.OAND:
				switch lv {
				case 0:
					gins(x86.AMOVL, ncon(0), &lo2)

				default:
					gmove(&lo1, &lo2)
					if lv != 0xffffffff {
						gins(x86.AANDL, ncon(lv), &lo2)
					}
				}

				switch hv {
				case 0:
					gins(x86.AMOVL, ncon(0), &hi2)

				default:
					gmove(&hi1, &hi2)
					if hv != 0xffffffff {
						gins(x86.AANDL, ncon(hv), &hi2)
					}
				}

			case gc.OOR:
				switch lv {
				case 0:
					gmove(&lo1, &lo2)

				case 0xffffffff:
					gins(x86.AMOVL, ncon(0xffffffff), &lo2)

				default:
					gmove(&lo1, &lo2)
					gins(x86.AORL, ncon(lv), &lo2)
				}

				switch hv {
				case 0:
					gmove(&hi1, &hi2)

				case 0xffffffff:
					gins(x86.AMOVL, ncon(0xffffffff), &hi2)

				default:
					gmove(&hi1, &hi2)
					gins(x86.AORL, ncon(hv), &hi2)
				}
			}

			splitclean()
			splitclean()
			return
		}

		gins(x86.AMOVL, &lo1, &ax)
		gins(x86.AMOVL, &hi1, &dx)
		gins(optoas(n.Op, lo1.Type), &lo2, &ax)
		gins(optoas(n.Op, lo1.Type), &hi2, &dx)
	}

	if gc.Is64(r.Type) {
		splitclean()
	}
	splitclean()

	split64(res, &lo1, &hi1)
	gins(x86.AMOVL, &ax, &lo1)
	gins(x86.AMOVL, &dx, &hi1)
	splitclean()
}
Exemple #6
0
func gmove(f *gc.Node, t *gc.Node) {
	if gc.Debug['M'] != 0 {
		fmt.Printf("gmove %v -> %v\n", f, t)
	}

	ft := gc.Simsimtype(f.Type)
	tt := gc.Simsimtype(t.Type)
	cvt := t.Type

	if gc.Iscomplex[ft] || gc.Iscomplex[tt] {
		gc.Complexmove(f, t)
		return
	}

	if gc.Isfloat[ft] || gc.Isfloat[tt] {
		floatmove(f, t)
		return
	}

	// cannot have two integer memory operands;
	// except 64-bit, which always copies via registers anyway.
	var r1 gc.Node
	var a int
	if gc.Isint[ft] && gc.Isint[tt] && !gc.Is64(f.Type) && !gc.Is64(t.Type) && gc.Ismem(f) && gc.Ismem(t) {
		goto hard
	}

	// convert constant to desired type
	if f.Op == gc.OLITERAL {
		var con gc.Node
		f.Convconst(&con, t.Type)
		f = &con
		ft = gc.Simsimtype(con.Type)
	}

	// value -> value copy, only one memory operand.
	// figure out the instruction to use.
	// break out of switch for one-instruction gins.
	// goto rdst for "destination must be register".
	// goto hard for "convert to cvt type first".
	// otherwise handle and return.

	switch uint32(ft)<<16 | uint32(tt) {
	default:
		// should not happen
		gc.Fatalf("gmove %v -> %v", f, t)
		return

		/*
		 * integer copy and truncate
		 */
	case gc.TINT8<<16 | gc.TINT8, // same size
		gc.TINT8<<16 | gc.TUINT8,
		gc.TUINT8<<16 | gc.TINT8,
		gc.TUINT8<<16 | gc.TUINT8:
		a = x86.AMOVB

	case gc.TINT16<<16 | gc.TINT8, // truncate
		gc.TUINT16<<16 | gc.TINT8,
		gc.TINT32<<16 | gc.TINT8,
		gc.TUINT32<<16 | gc.TINT8,
		gc.TINT16<<16 | gc.TUINT8,
		gc.TUINT16<<16 | gc.TUINT8,
		gc.TINT32<<16 | gc.TUINT8,
		gc.TUINT32<<16 | gc.TUINT8:
		a = x86.AMOVB

		goto rsrc

	case gc.TINT64<<16 | gc.TINT8, // truncate low word
		gc.TUINT64<<16 | gc.TINT8,
		gc.TINT64<<16 | gc.TUINT8,
		gc.TUINT64<<16 | gc.TUINT8:
		var flo gc.Node
		var fhi gc.Node
		split64(f, &flo, &fhi)

		var r1 gc.Node
		gc.Nodreg(&r1, t.Type, x86.REG_AX)
		gmove(&flo, &r1)
		gins(x86.AMOVB, &r1, t)
		splitclean()
		return

	case gc.TINT16<<16 | gc.TINT16, // same size
		gc.TINT16<<16 | gc.TUINT16,
		gc.TUINT16<<16 | gc.TINT16,
		gc.TUINT16<<16 | gc.TUINT16:
		a = x86.AMOVW

	case gc.TINT32<<16 | gc.TINT16, // truncate
		gc.TUINT32<<16 | gc.TINT16,
		gc.TINT32<<16 | gc.TUINT16,
		gc.TUINT32<<16 | gc.TUINT16:
		a = x86.AMOVW

		goto rsrc

	case gc.TINT64<<16 | gc.TINT16, // truncate low word
		gc.TUINT64<<16 | gc.TINT16,
		gc.TINT64<<16 | gc.TUINT16,
		gc.TUINT64<<16 | gc.TUINT16:
		var flo gc.Node
		var fhi gc.Node
		split64(f, &flo, &fhi)

		var r1 gc.Node
		gc.Nodreg(&r1, t.Type, x86.REG_AX)
		gmove(&flo, &r1)
		gins(x86.AMOVW, &r1, t)
		splitclean()
		return

	case gc.TINT32<<16 | gc.TINT32, // same size
		gc.TINT32<<16 | gc.TUINT32,
		gc.TUINT32<<16 | gc.TINT32,
		gc.TUINT32<<16 | gc.TUINT32:
		a = x86.AMOVL

	case gc.TINT64<<16 | gc.TINT32, // truncate
		gc.TUINT64<<16 | gc.TINT32,
		gc.TINT64<<16 | gc.TUINT32,
		gc.TUINT64<<16 | gc.TUINT32:
		var fhi gc.Node
		var flo gc.Node
		split64(f, &flo, &fhi)

		var r1 gc.Node
		gc.Nodreg(&r1, t.Type, x86.REG_AX)
		gmove(&flo, &r1)
		gins(x86.AMOVL, &r1, t)
		splitclean()
		return

	case gc.TINT64<<16 | gc.TINT64, // same size
		gc.TINT64<<16 | gc.TUINT64,
		gc.TUINT64<<16 | gc.TINT64,
		gc.TUINT64<<16 | gc.TUINT64:
		var fhi gc.Node
		var flo gc.Node
		split64(f, &flo, &fhi)

		var tlo gc.Node
		var thi gc.Node
		split64(t, &tlo, &thi)
		if f.Op == gc.OLITERAL {
			gins(x86.AMOVL, &flo, &tlo)
			gins(x86.AMOVL, &fhi, &thi)
		} else {
			// Implementation of conversion-free x = y for int64 or uint64 x.
			// This is generated by the code that copies small values out of closures,
			// and that code has DX live, so avoid DX and just use AX twice.
			var r1 gc.Node
			gc.Nodreg(&r1, gc.Types[gc.TUINT32], x86.REG_AX)
			gins(x86.AMOVL, &flo, &r1)
			gins(x86.AMOVL, &r1, &tlo)
			gins(x86.AMOVL, &fhi, &r1)
			gins(x86.AMOVL, &r1, &thi)
		}

		splitclean()
		splitclean()
		return

		/*
		 * integer up-conversions
		 */
	case gc.TINT8<<16 | gc.TINT16, // sign extend int8
		gc.TINT8<<16 | gc.TUINT16:
		a = x86.AMOVBWSX

		goto rdst

	case gc.TINT8<<16 | gc.TINT32,
		gc.TINT8<<16 | gc.TUINT32:
		a = x86.AMOVBLSX
		goto rdst

	case gc.TINT8<<16 | gc.TINT64, // convert via int32
		gc.TINT8<<16 | gc.TUINT64:
		cvt = gc.Types[gc.TINT32]

		goto hard

	case gc.TUINT8<<16 | gc.TINT16, // zero extend uint8
		gc.TUINT8<<16 | gc.TUINT16:
		a = x86.AMOVBWZX

		goto rdst

	case gc.TUINT8<<16 | gc.TINT32,
		gc.TUINT8<<16 | gc.TUINT32:
		a = x86.AMOVBLZX
		goto rdst

	case gc.TUINT8<<16 | gc.TINT64, // convert via uint32
		gc.TUINT8<<16 | gc.TUINT64:
		cvt = gc.Types[gc.TUINT32]

		goto hard

	case gc.TINT16<<16 | gc.TINT32, // sign extend int16
		gc.TINT16<<16 | gc.TUINT32:
		a = x86.AMOVWLSX

		goto rdst

	case gc.TINT16<<16 | gc.TINT64, // convert via int32
		gc.TINT16<<16 | gc.TUINT64:
		cvt = gc.Types[gc.TINT32]

		goto hard

	case gc.TUINT16<<16 | gc.TINT32, // zero extend uint16
		gc.TUINT16<<16 | gc.TUINT32:
		a = x86.AMOVWLZX

		goto rdst

	case gc.TUINT16<<16 | gc.TINT64, // convert via uint32
		gc.TUINT16<<16 | gc.TUINT64:
		cvt = gc.Types[gc.TUINT32]

		goto hard

	case gc.TINT32<<16 | gc.TINT64, // sign extend int32
		gc.TINT32<<16 | gc.TUINT64:
		var thi gc.Node
		var tlo gc.Node
		split64(t, &tlo, &thi)

		var flo gc.Node
		gc.Nodreg(&flo, tlo.Type, x86.REG_AX)
		var fhi gc.Node
		gc.Nodreg(&fhi, thi.Type, x86.REG_DX)
		gmove(f, &flo)
		gins(x86.ACDQ, nil, nil)
		gins(x86.AMOVL, &flo, &tlo)
		gins(x86.AMOVL, &fhi, &thi)
		splitclean()
		return

	case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
		gc.TUINT32<<16 | gc.TUINT64:
		var tlo gc.Node
		var thi gc.Node
		split64(t, &tlo, &thi)

		gmove(f, &tlo)
		gins(x86.AMOVL, ncon(0), &thi)
		splitclean()
		return
	}

	gins(a, f, t)
	return

	// requires register source
rsrc:
	gc.Regalloc(&r1, f.Type, t)

	gmove(f, &r1)
	gins(a, &r1, t)
	gc.Regfree(&r1)
	return

	// requires register destination
rdst:
	{
		gc.Regalloc(&r1, t.Type, t)

		gins(a, f, &r1)
		gmove(&r1, t)
		gc.Regfree(&r1)
		return
	}

	// requires register intermediate
hard:
	gc.Regalloc(&r1, cvt, t)

	gmove(f, &r1)
	gmove(&r1, t)
	gc.Regfree(&r1)
	return
}
Exemple #7
0
func gmove(f *gc.Node, t *gc.Node) {
	if gc.Debug['M'] != 0 {
		fmt.Printf("gmove %v -> %v\n", f, t)
	}

	ft := gc.Simsimtype(f.Type)
	tt := gc.Simsimtype(t.Type)
	cvt := t.Type

	if gc.Iscomplex[ft] || gc.Iscomplex[tt] {
		gc.Complexmove(f, t)
		return
	}

	// cannot have two memory operands;
	// except 64-bit, which always copies via registers anyway.
	var a int
	var r1 gc.Node
	if !gc.Is64(f.Type) && !gc.Is64(t.Type) && gc.Ismem(f) && gc.Ismem(t) {
		goto hard
	}

	// convert constant to desired type
	if f.Op == gc.OLITERAL {
		var con gc.Node
		switch tt {
		default:
			f.Convconst(&con, t.Type)

		case gc.TINT16,
			gc.TINT8:
			var con gc.Node
			f.Convconst(&con, gc.Types[gc.TINT32])
			var r1 gc.Node
			gc.Regalloc(&r1, con.Type, t)
			gins(arm.AMOVW, &con, &r1)
			gmove(&r1, t)
			gc.Regfree(&r1)
			return

		case gc.TUINT16,
			gc.TUINT8:
			var con gc.Node
			f.Convconst(&con, gc.Types[gc.TUINT32])
			var r1 gc.Node
			gc.Regalloc(&r1, con.Type, t)
			gins(arm.AMOVW, &con, &r1)
			gmove(&r1, t)
			gc.Regfree(&r1)
			return
		}

		f = &con
		ft = gc.Simsimtype(con.Type)

		// constants can't move directly to memory
		if gc.Ismem(t) && !gc.Is64(t.Type) {
			goto hard
		}
	}

	// value -> value copy, only one memory operand.
	// figure out the instruction to use.
	// break out of switch for one-instruction gins.
	// goto rdst for "destination must be register".
	// goto hard for "convert to cvt type first".
	// otherwise handle and return.

	switch uint32(ft)<<16 | uint32(tt) {
	default:
		// should not happen
		gc.Fatal("gmove %v -> %v", f, t)
		return

		/*
		 * integer copy and truncate
		 */
	case gc.TINT8<<16 | gc.TINT8: // same size
		if !gc.Ismem(f) {
			a = arm.AMOVB
			break
		}
		fallthrough

	case gc.TUINT8<<16 | gc.TINT8,
		gc.TINT16<<16 | gc.TINT8, // truncate
		gc.TUINT16<<16 | gc.TINT8,
		gc.TINT32<<16 | gc.TINT8,
		gc.TUINT32<<16 | gc.TINT8:
		a = arm.AMOVBS

	case gc.TUINT8<<16 | gc.TUINT8:
		if !gc.Ismem(f) {
			a = arm.AMOVB
			break
		}
		fallthrough

	case gc.TINT8<<16 | gc.TUINT8,
		gc.TINT16<<16 | gc.TUINT8,
		gc.TUINT16<<16 | gc.TUINT8,
		gc.TINT32<<16 | gc.TUINT8,
		gc.TUINT32<<16 | gc.TUINT8:
		a = arm.AMOVBU

	case gc.TINT64<<16 | gc.TINT8, // truncate low word
		gc.TUINT64<<16 | gc.TINT8:
		a = arm.AMOVBS

		goto trunc64

	case gc.TINT64<<16 | gc.TUINT8,
		gc.TUINT64<<16 | gc.TUINT8:
		a = arm.AMOVBU
		goto trunc64

	case gc.TINT16<<16 | gc.TINT16: // same size
		if !gc.Ismem(f) {
			a = arm.AMOVH
			break
		}
		fallthrough

	case gc.TUINT16<<16 | gc.TINT16,
		gc.TINT32<<16 | gc.TINT16, // truncate
		gc.TUINT32<<16 | gc.TINT16:
		a = arm.AMOVHS

	case gc.TUINT16<<16 | gc.TUINT16:
		if !gc.Ismem(f) {
			a = arm.AMOVH
			break
		}
		fallthrough

	case gc.TINT16<<16 | gc.TUINT16,
		gc.TINT32<<16 | gc.TUINT16,
		gc.TUINT32<<16 | gc.TUINT16:
		a = arm.AMOVHU

	case gc.TINT64<<16 | gc.TINT16, // truncate low word
		gc.TUINT64<<16 | gc.TINT16:
		a = arm.AMOVHS

		goto trunc64

	case gc.TINT64<<16 | gc.TUINT16,
		gc.TUINT64<<16 | gc.TUINT16:
		a = arm.AMOVHU
		goto trunc64

	case gc.TINT32<<16 | gc.TINT32, // same size
		gc.TINT32<<16 | gc.TUINT32,
		gc.TUINT32<<16 | gc.TINT32,
		gc.TUINT32<<16 | gc.TUINT32:
		a = arm.AMOVW

	case gc.TINT64<<16 | gc.TINT32, // truncate
		gc.TUINT64<<16 | gc.TINT32,
		gc.TINT64<<16 | gc.TUINT32,
		gc.TUINT64<<16 | gc.TUINT32:
		var flo gc.Node
		var fhi gc.Node
		split64(f, &flo, &fhi)

		var r1 gc.Node
		gc.Regalloc(&r1, t.Type, nil)
		gins(arm.AMOVW, &flo, &r1)
		gins(arm.AMOVW, &r1, t)
		gc.Regfree(&r1)
		splitclean()
		return

	case gc.TINT64<<16 | gc.TINT64, // same size
		gc.TINT64<<16 | gc.TUINT64,
		gc.TUINT64<<16 | gc.TINT64,
		gc.TUINT64<<16 | gc.TUINT64:
		var fhi gc.Node
		var flo gc.Node
		split64(f, &flo, &fhi)

		var tlo gc.Node
		var thi gc.Node
		split64(t, &tlo, &thi)
		var r1 gc.Node
		gc.Regalloc(&r1, flo.Type, nil)
		var r2 gc.Node
		gc.Regalloc(&r2, fhi.Type, nil)
		gins(arm.AMOVW, &flo, &r1)
		gins(arm.AMOVW, &fhi, &r2)
		gins(arm.AMOVW, &r1, &tlo)
		gins(arm.AMOVW, &r2, &thi)
		gc.Regfree(&r1)
		gc.Regfree(&r2)
		splitclean()
		splitclean()
		return

		/*
		 * integer up-conversions
		 */
	case gc.TINT8<<16 | gc.TINT16, // sign extend int8
		gc.TINT8<<16 | gc.TUINT16,
		gc.TINT8<<16 | gc.TINT32,
		gc.TINT8<<16 | gc.TUINT32:
		a = arm.AMOVBS

		goto rdst

	case gc.TINT8<<16 | gc.TINT64, // convert via int32
		gc.TINT8<<16 | gc.TUINT64:
		cvt = gc.Types[gc.TINT32]

		goto hard

	case gc.TUINT8<<16 | gc.TINT16, // zero extend uint8
		gc.TUINT8<<16 | gc.TUINT16,
		gc.TUINT8<<16 | gc.TINT32,
		gc.TUINT8<<16 | gc.TUINT32:
		a = arm.AMOVBU

		goto rdst

	case gc.TUINT8<<16 | gc.TINT64, // convert via uint32
		gc.TUINT8<<16 | gc.TUINT64:
		cvt = gc.Types[gc.TUINT32]

		goto hard

	case gc.TINT16<<16 | gc.TINT32, // sign extend int16
		gc.TINT16<<16 | gc.TUINT32:
		a = arm.AMOVHS

		goto rdst

	case gc.TINT16<<16 | gc.TINT64, // convert via int32
		gc.TINT16<<16 | gc.TUINT64:
		cvt = gc.Types[gc.TINT32]

		goto hard

	case gc.TUINT16<<16 | gc.TINT32, // zero extend uint16
		gc.TUINT16<<16 | gc.TUINT32:
		a = arm.AMOVHU

		goto rdst

	case gc.TUINT16<<16 | gc.TINT64, // convert via uint32
		gc.TUINT16<<16 | gc.TUINT64:
		cvt = gc.Types[gc.TUINT32]

		goto hard

	case gc.TINT32<<16 | gc.TINT64, // sign extend int32
		gc.TINT32<<16 | gc.TUINT64:
		var tlo gc.Node
		var thi gc.Node
		split64(t, &tlo, &thi)

		var r1 gc.Node
		gc.Regalloc(&r1, tlo.Type, nil)
		var r2 gc.Node
		gc.Regalloc(&r2, thi.Type, nil)
		gmove(f, &r1)
		p1 := gins(arm.AMOVW, &r1, &r2)
		p1.From.Type = obj.TYPE_SHIFT
		p1.From.Offset = 2<<5 | 31<<7 | int64(r1.Reg)&15 // r1->31
		p1.From.Reg = 0

		//print("gmove: %v\n", p1);
		gins(arm.AMOVW, &r1, &tlo)

		gins(arm.AMOVW, &r2, &thi)
		gc.Regfree(&r1)
		gc.Regfree(&r2)
		splitclean()
		return

	case gc.TUINT32<<16 | gc.TINT64, // zero extend uint32
		gc.TUINT32<<16 | gc.TUINT64:
		var thi gc.Node
		var tlo gc.Node
		split64(t, &tlo, &thi)

		gmove(f, &tlo)
		var r1 gc.Node
		gc.Regalloc(&r1, thi.Type, nil)
		gins(arm.AMOVW, ncon(0), &r1)
		gins(arm.AMOVW, &r1, &thi)
		gc.Regfree(&r1)
		splitclean()
		return

		//	case CASE(TFLOAT64, TUINT64):
	/*
	* float to integer
	 */
	case gc.TFLOAT32<<16 | gc.TINT8,
		gc.TFLOAT32<<16 | gc.TUINT8,
		gc.TFLOAT32<<16 | gc.TINT16,
		gc.TFLOAT32<<16 | gc.TUINT16,
		gc.TFLOAT32<<16 | gc.TINT32,
		gc.TFLOAT32<<16 | gc.TUINT32,

		//	case CASE(TFLOAT32, TUINT64):

		gc.TFLOAT64<<16 | gc.TINT8,
		gc.TFLOAT64<<16 | gc.TUINT8,
		gc.TFLOAT64<<16 | gc.TINT16,
		gc.TFLOAT64<<16 | gc.TUINT16,
		gc.TFLOAT64<<16 | gc.TINT32,
		gc.TFLOAT64<<16 | gc.TUINT32:
		fa := arm.AMOVF

		a := arm.AMOVFW
		if ft == gc.TFLOAT64 {
			fa = arm.AMOVD
			a = arm.AMOVDW
		}

		ta := arm.AMOVW
		switch tt {
		case gc.TINT8:
			ta = arm.AMOVBS

		case gc.TUINT8:
			ta = arm.AMOVBU

		case gc.TINT16:
			ta = arm.AMOVHS

		case gc.TUINT16:
			ta = arm.AMOVHU
		}

		var r1 gc.Node
		gc.Regalloc(&r1, gc.Types[ft], f)
		var r2 gc.Node
		gc.Regalloc(&r2, gc.Types[tt], t)
		gins(fa, f, &r1)        // load to fpu
		p1 := gins(a, &r1, &r1) // convert to w
		switch tt {
		case gc.TUINT8,
			gc.TUINT16,
			gc.TUINT32:
			p1.Scond |= arm.C_UBIT
		}

		gins(arm.AMOVW, &r1, &r2) // copy to cpu
		gins(ta, &r2, t)          // store
		gc.Regfree(&r1)
		gc.Regfree(&r2)
		return

		/*
		 * integer to float
		 */
	case gc.TINT8<<16 | gc.TFLOAT32,
		gc.TUINT8<<16 | gc.TFLOAT32,
		gc.TINT16<<16 | gc.TFLOAT32,
		gc.TUINT16<<16 | gc.TFLOAT32,
		gc.TINT32<<16 | gc.TFLOAT32,
		gc.TUINT32<<16 | gc.TFLOAT32,
		gc.TINT8<<16 | gc.TFLOAT64,
		gc.TUINT8<<16 | gc.TFLOAT64,
		gc.TINT16<<16 | gc.TFLOAT64,
		gc.TUINT16<<16 | gc.TFLOAT64,
		gc.TINT32<<16 | gc.TFLOAT64,
		gc.TUINT32<<16 | gc.TFLOAT64:
		fa := arm.AMOVW

		switch ft {
		case gc.TINT8:
			fa = arm.AMOVBS

		case gc.TUINT8:
			fa = arm.AMOVBU

		case gc.TINT16:
			fa = arm.AMOVHS

		case gc.TUINT16:
			fa = arm.AMOVHU
		}

		a := arm.AMOVWF
		ta := arm.AMOVF
		if tt == gc.TFLOAT64 {
			a = arm.AMOVWD
			ta = arm.AMOVD
		}

		var r1 gc.Node
		gc.Regalloc(&r1, gc.Types[ft], f)
		var r2 gc.Node
		gc.Regalloc(&r2, gc.Types[tt], t)
		gins(fa, f, &r1)          // load to cpu
		gins(arm.AMOVW, &r1, &r2) // copy to fpu
		p1 := gins(a, &r2, &r2)   // convert
		switch ft {
		case gc.TUINT8,
			gc.TUINT16,
			gc.TUINT32:
			p1.Scond |= arm.C_UBIT
		}

		gins(ta, &r2, t) // store
		gc.Regfree(&r1)
		gc.Regfree(&r2)
		return

	case gc.TUINT64<<16 | gc.TFLOAT32,
		gc.TUINT64<<16 | gc.TFLOAT64:
		gc.Fatal("gmove UINT64, TFLOAT not implemented")
		return

		/*
		 * float to float
		 */
	case gc.TFLOAT32<<16 | gc.TFLOAT32:
		a = arm.AMOVF

	case gc.TFLOAT64<<16 | gc.TFLOAT64:
		a = arm.AMOVD

	case gc.TFLOAT32<<16 | gc.TFLOAT64:
		var r1 gc.Node
		gc.Regalloc(&r1, gc.Types[gc.TFLOAT64], t)
		gins(arm.AMOVF, f, &r1)
		gins(arm.AMOVFD, &r1, &r1)
		gins(arm.AMOVD, &r1, t)
		gc.Regfree(&r1)
		return

	case gc.TFLOAT64<<16 | gc.TFLOAT32:
		var r1 gc.Node
		gc.Regalloc(&r1, gc.Types[gc.TFLOAT64], t)
		gins(arm.AMOVD, f, &r1)
		gins(arm.AMOVDF, &r1, &r1)
		gins(arm.AMOVF, &r1, t)
		gc.Regfree(&r1)
		return
	}

	gins(a, f, t)
	return

	// TODO(kaib): we almost always require a register dest anyway, this can probably be
	// removed.
	// requires register destination
rdst:
	{
		gc.Regalloc(&r1, t.Type, t)

		gins(a, f, &r1)
		gmove(&r1, t)
		gc.Regfree(&r1)
		return
	}

	// requires register intermediate
hard:
	gc.Regalloc(&r1, cvt, t)

	gmove(f, &r1)
	gmove(&r1, t)
	gc.Regfree(&r1)
	return

	// truncate 64 bit integer
trunc64:
	var fhi gc.Node
	var flo gc.Node
	split64(f, &flo, &fhi)

	gc.Regalloc(&r1, t.Type, nil)
	gins(a, &flo, &r1)
	gins(a, &r1, t)
	gc.Regfree(&r1)
	splitclean()
	return
}
Exemple #8
0
/*
 * attempt to generate 64-bit
 *	res = n
 * return 1 on success, 0 if op not handled.
 */
func cgen64(n *gc.Node, res *gc.Node) {
	if res.Op != gc.OINDREG && res.Op != gc.ONAME {
		gc.Dump("n", n)
		gc.Dump("res", res)
		gc.Fatal("cgen64 %v of %v", gc.Oconv(int(n.Op), 0), gc.Oconv(int(res.Op), 0))
	}

	l := n.Left
	var t1 gc.Node
	if !l.Addable {
		gc.Tempname(&t1, l.Type)
		gc.Cgen(l, &t1)
		l = &t1
	}

	var hi1 gc.Node
	var lo1 gc.Node
	split64(l, &lo1, &hi1)
	switch n.Op {
	default:
		gc.Fatal("cgen64 %v", gc.Oconv(int(n.Op), 0))

	case gc.OMINUS:
		var lo2 gc.Node
		var hi2 gc.Node
		split64(res, &lo2, &hi2)

		gc.Regalloc(&t1, lo1.Type, nil)
		var al gc.Node
		gc.Regalloc(&al, lo1.Type, nil)
		var ah gc.Node
		gc.Regalloc(&ah, hi1.Type, nil)

		gins(arm.AMOVW, &lo1, &al)
		gins(arm.AMOVW, &hi1, &ah)

		gmove(ncon(0), &t1)
		p1 := gins(arm.ASUB, &al, &t1)
		p1.Scond |= arm.C_SBIT
		gins(arm.AMOVW, &t1, &lo2)

		gmove(ncon(0), &t1)
		gins(arm.ASBC, &ah, &t1)
		gins(arm.AMOVW, &t1, &hi2)

		gc.Regfree(&t1)
		gc.Regfree(&al)
		gc.Regfree(&ah)
		splitclean()
		splitclean()
		return

	case gc.OCOM:
		gc.Regalloc(&t1, lo1.Type, nil)
		gmove(ncon(^uint32(0)), &t1)

		var lo2 gc.Node
		var hi2 gc.Node
		split64(res, &lo2, &hi2)
		var n1 gc.Node
		gc.Regalloc(&n1, lo1.Type, nil)

		gins(arm.AMOVW, &lo1, &n1)
		gins(arm.AEOR, &t1, &n1)
		gins(arm.AMOVW, &n1, &lo2)

		gins(arm.AMOVW, &hi1, &n1)
		gins(arm.AEOR, &t1, &n1)
		gins(arm.AMOVW, &n1, &hi2)

		gc.Regfree(&t1)
		gc.Regfree(&n1)
		splitclean()
		splitclean()
		return

		// binary operators.
	// common setup below.
	case gc.OADD,
		gc.OSUB,
		gc.OMUL,
		gc.OLSH,
		gc.ORSH,
		gc.OAND,
		gc.OOR,
		gc.OXOR,
		gc.OLROT:
		break
	}

	// setup for binary operators
	r := n.Right

	if r != nil && !r.Addable {
		var t2 gc.Node
		gc.Tempname(&t2, r.Type)
		gc.Cgen(r, &t2)
		r = &t2
	}

	var hi2 gc.Node
	var lo2 gc.Node
	if gc.Is64(r.Type) {
		split64(r, &lo2, &hi2)
	}

	var al gc.Node
	gc.Regalloc(&al, lo1.Type, nil)
	var ah gc.Node
	gc.Regalloc(&ah, hi1.Type, nil)

	// Do op.  Leave result in ah:al.
	switch n.Op {
	default:
		gc.Fatal("cgen64: not implemented: %v\n", n)

		// TODO: Constants
	case gc.OADD:
		var bl gc.Node
		gc.Regalloc(&bl, gc.Types[gc.TPTR32], nil)

		var bh gc.Node
		gc.Regalloc(&bh, gc.Types[gc.TPTR32], nil)
		gins(arm.AMOVW, &hi1, &ah)
		gins(arm.AMOVW, &lo1, &al)
		gins(arm.AMOVW, &hi2, &bh)
		gins(arm.AMOVW, &lo2, &bl)
		p1 := gins(arm.AADD, &bl, &al)
		p1.Scond |= arm.C_SBIT
		gins(arm.AADC, &bh, &ah)
		gc.Regfree(&bl)
		gc.Regfree(&bh)

		// TODO: Constants.
	case gc.OSUB:
		var bl gc.Node
		gc.Regalloc(&bl, gc.Types[gc.TPTR32], nil)

		var bh gc.Node
		gc.Regalloc(&bh, gc.Types[gc.TPTR32], nil)
		gins(arm.AMOVW, &lo1, &al)
		gins(arm.AMOVW, &hi1, &ah)
		gins(arm.AMOVW, &lo2, &bl)
		gins(arm.AMOVW, &hi2, &bh)
		p1 := gins(arm.ASUB, &bl, &al)
		p1.Scond |= arm.C_SBIT
		gins(arm.ASBC, &bh, &ah)
		gc.Regfree(&bl)
		gc.Regfree(&bh)

		// TODO(kaib): this can be done with 4 regs and does not need 6
	case gc.OMUL:
		var bl gc.Node
		gc.Regalloc(&bl, gc.Types[gc.TPTR32], nil)

		var bh gc.Node
		gc.Regalloc(&bh, gc.Types[gc.TPTR32], nil)
		var cl gc.Node
		gc.Regalloc(&cl, gc.Types[gc.TPTR32], nil)
		var ch gc.Node
		gc.Regalloc(&ch, gc.Types[gc.TPTR32], nil)

		// load args into bh:bl and bh:bl.
		gins(arm.AMOVW, &hi1, &bh)

		gins(arm.AMOVW, &lo1, &bl)
		gins(arm.AMOVW, &hi2, &ch)
		gins(arm.AMOVW, &lo2, &cl)

		// bl * cl -> ah al
		p1 := gins(arm.AMULLU, nil, nil)

		p1.From.Type = obj.TYPE_REG
		p1.From.Reg = bl.Reg
		p1.Reg = cl.Reg
		p1.To.Type = obj.TYPE_REGREG
		p1.To.Reg = ah.Reg
		p1.To.Offset = int64(al.Reg)

		//print("%v\n", p1);

		// bl * ch + ah -> ah
		p1 = gins(arm.AMULA, nil, nil)

		p1.From.Type = obj.TYPE_REG
		p1.From.Reg = bl.Reg
		p1.Reg = ch.Reg
		p1.To.Type = obj.TYPE_REGREG2
		p1.To.Reg = ah.Reg
		p1.To.Offset = int64(ah.Reg)

		//print("%v\n", p1);

		// bh * cl + ah -> ah
		p1 = gins(arm.AMULA, nil, nil)

		p1.From.Type = obj.TYPE_REG
		p1.From.Reg = bh.Reg
		p1.Reg = cl.Reg
		p1.To.Type = obj.TYPE_REGREG2
		p1.To.Reg = ah.Reg
		p1.To.Offset = int64(ah.Reg)

		//print("%v\n", p1);

		gc.Regfree(&bh)

		gc.Regfree(&bl)
		gc.Regfree(&ch)
		gc.Regfree(&cl)

		// We only rotate by a constant c in [0,64).
	// if c >= 32:
	//	lo, hi = hi, lo
	//	c -= 32
	// if c == 0:
	//	no-op
	// else:
	//	t = hi
	//	shld hi:lo, c
	//	shld lo:t, c
	case gc.OLROT:
		v := uint64(r.Int())

		var bl gc.Node
		gc.Regalloc(&bl, lo1.Type, nil)
		var bh gc.Node
		gc.Regalloc(&bh, hi1.Type, nil)
		if v >= 32 {
			// reverse during load to do the first 32 bits of rotate
			v -= 32

			gins(arm.AMOVW, &hi1, &bl)
			gins(arm.AMOVW, &lo1, &bh)
		} else {
			gins(arm.AMOVW, &hi1, &bh)
			gins(arm.AMOVW, &lo1, &bl)
		}

		if v == 0 {
			gins(arm.AMOVW, &bh, &ah)
			gins(arm.AMOVW, &bl, &al)
		} else {
			// rotate by 1 <= v <= 31
			//	MOVW	bl<<v, al
			//	MOVW	bh<<v, ah
			//	OR		bl>>(32-v), ah
			//	OR		bh>>(32-v), al
			gshift(arm.AMOVW, &bl, arm.SHIFT_LL, int32(v), &al)

			gshift(arm.AMOVW, &bh, arm.SHIFT_LL, int32(v), &ah)
			gshift(arm.AORR, &bl, arm.SHIFT_LR, int32(32-v), &ah)
			gshift(arm.AORR, &bh, arm.SHIFT_LR, int32(32-v), &al)
		}

		gc.Regfree(&bl)
		gc.Regfree(&bh)

	case gc.OLSH:
		var bl gc.Node
		gc.Regalloc(&bl, lo1.Type, nil)
		var bh gc.Node
		gc.Regalloc(&bh, hi1.Type, nil)
		gins(arm.AMOVW, &hi1, &bh)
		gins(arm.AMOVW, &lo1, &bl)

		var p6 *obj.Prog
		var s gc.Node
		var n1 gc.Node
		var creg gc.Node
		var p1 *obj.Prog
		var p2 *obj.Prog
		var p3 *obj.Prog
		var p4 *obj.Prog
		var p5 *obj.Prog
		if r.Op == gc.OLITERAL {
			v := uint64(r.Int())
			if v >= 64 {
				// TODO(kaib): replace with gins(AMOVW, nodintconst(0), &al)
				// here and below (verify it optimizes to EOR)
				gins(arm.AEOR, &al, &al)

				gins(arm.AEOR, &ah, &ah)
			} else if v > 32 {
				gins(arm.AEOR, &al, &al)

				//	MOVW	bl<<(v-32), ah
				gshift(arm.AMOVW, &bl, arm.SHIFT_LL, int32(v-32), &ah)
			} else if v == 32 {
				gins(arm.AEOR, &al, &al)
				gins(arm.AMOVW, &bl, &ah)
			} else if v > 0 {
				//	MOVW	bl<<v, al
				gshift(arm.AMOVW, &bl, arm.SHIFT_LL, int32(v), &al)

				//	MOVW	bh<<v, ah
				gshift(arm.AMOVW, &bh, arm.SHIFT_LL, int32(v), &ah)

				//	OR		bl>>(32-v), ah
				gshift(arm.AORR, &bl, arm.SHIFT_LR, int32(32-v), &ah)
			} else {
				gins(arm.AMOVW, &bl, &al)
				gins(arm.AMOVW, &bh, &ah)
			}

			goto olsh_break
		}

		gc.Regalloc(&s, gc.Types[gc.TUINT32], nil)
		gc.Regalloc(&creg, gc.Types[gc.TUINT32], nil)
		if gc.Is64(r.Type) {
			// shift is >= 1<<32
			var cl gc.Node
			var ch gc.Node
			split64(r, &cl, &ch)

			gmove(&ch, &s)
			gins(arm.ATST, &s, nil)
			p6 = gc.Gbranch(arm.ABNE, nil, 0)
			gmove(&cl, &s)
			splitclean()
		} else {
			gmove(r, &s)
			p6 = nil
		}

		gins(arm.ATST, &s, nil)

		// shift == 0
		p1 = gins(arm.AMOVW, &bl, &al)

		p1.Scond = arm.C_SCOND_EQ
		p1 = gins(arm.AMOVW, &bh, &ah)
		p1.Scond = arm.C_SCOND_EQ
		p2 = gc.Gbranch(arm.ABEQ, nil, 0)

		// shift is < 32
		gc.Nodconst(&n1, gc.Types[gc.TUINT32], 32)

		gmove(&n1, &creg)
		gins(arm.ACMP, &s, &creg)

		//	MOVW.LO		bl<<s, al
		p1 = gregshift(arm.AMOVW, &bl, arm.SHIFT_LL, &s, &al)

		p1.Scond = arm.C_SCOND_LO

		//	MOVW.LO		bh<<s, ah
		p1 = gregshift(arm.AMOVW, &bh, arm.SHIFT_LL, &s, &ah)

		p1.Scond = arm.C_SCOND_LO

		//	SUB.LO		s, creg
		p1 = gins(arm.ASUB, &s, &creg)

		p1.Scond = arm.C_SCOND_LO

		//	OR.LO		bl>>creg, ah
		p1 = gregshift(arm.AORR, &bl, arm.SHIFT_LR, &creg, &ah)

		p1.Scond = arm.C_SCOND_LO

		//	BLO	end
		p3 = gc.Gbranch(arm.ABLO, nil, 0)

		// shift == 32
		p1 = gins(arm.AEOR, &al, &al)

		p1.Scond = arm.C_SCOND_EQ
		p1 = gins(arm.AMOVW, &bl, &ah)
		p1.Scond = arm.C_SCOND_EQ
		p4 = gc.Gbranch(arm.ABEQ, nil, 0)

		// shift is < 64
		gc.Nodconst(&n1, gc.Types[gc.TUINT32], 64)

		gmove(&n1, &creg)
		gins(arm.ACMP, &s, &creg)

		//	EOR.LO	al, al
		p1 = gins(arm.AEOR, &al, &al)

		p1.Scond = arm.C_SCOND_LO

		//	MOVW.LO		creg>>1, creg
		p1 = gshift(arm.AMOVW, &creg, arm.SHIFT_LR, 1, &creg)

		p1.Scond = arm.C_SCOND_LO

		//	SUB.LO		creg, s
		p1 = gins(arm.ASUB, &creg, &s)

		p1.Scond = arm.C_SCOND_LO

		//	MOVW	bl<<s, ah
		p1 = gregshift(arm.AMOVW, &bl, arm.SHIFT_LL, &s, &ah)

		p1.Scond = arm.C_SCOND_LO

		p5 = gc.Gbranch(arm.ABLO, nil, 0)

		// shift >= 64
		if p6 != nil {
			gc.Patch(p6, gc.Pc)
		}
		gins(arm.AEOR, &al, &al)
		gins(arm.AEOR, &ah, &ah)

		gc.Patch(p2, gc.Pc)
		gc.Patch(p3, gc.Pc)
		gc.Patch(p4, gc.Pc)
		gc.Patch(p5, gc.Pc)
		gc.Regfree(&s)
		gc.Regfree(&creg)

	olsh_break:
		gc.Regfree(&bl)
		gc.Regfree(&bh)

	case gc.ORSH:
		var bl gc.Node
		gc.Regalloc(&bl, lo1.Type, nil)
		var bh gc.Node
		gc.Regalloc(&bh, hi1.Type, nil)
		gins(arm.AMOVW, &hi1, &bh)
		gins(arm.AMOVW, &lo1, &bl)

		var p4 *obj.Prog
		var p5 *obj.Prog
		var n1 gc.Node
		var p6 *obj.Prog
		var s gc.Node
		var p1 *obj.Prog
		var p2 *obj.Prog
		var creg gc.Node
		var p3 *obj.Prog
		if r.Op == gc.OLITERAL {
			v := uint64(r.Int())
			if v >= 64 {
				if bh.Type.Etype == gc.TINT32 {
					//	MOVW	bh->31, al
					gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &al)

					//	MOVW	bh->31, ah
					gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
				} else {
					gins(arm.AEOR, &al, &al)
					gins(arm.AEOR, &ah, &ah)
				}
			} else if v > 32 {
				if bh.Type.Etype == gc.TINT32 {
					//	MOVW	bh->(v-32), al
					gshift(arm.AMOVW, &bh, arm.SHIFT_AR, int32(v-32), &al)

					//	MOVW	bh->31, ah
					gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
				} else {
					//	MOVW	bh>>(v-32), al
					gshift(arm.AMOVW, &bh, arm.SHIFT_LR, int32(v-32), &al)

					gins(arm.AEOR, &ah, &ah)
				}
			} else if v == 32 {
				gins(arm.AMOVW, &bh, &al)
				if bh.Type.Etype == gc.TINT32 {
					//	MOVW	bh->31, ah
					gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
				} else {
					gins(arm.AEOR, &ah, &ah)
				}
			} else if v > 0 {
				//	MOVW	bl>>v, al
				gshift(arm.AMOVW, &bl, arm.SHIFT_LR, int32(v), &al)

				//	OR		bh<<(32-v), al
				gshift(arm.AORR, &bh, arm.SHIFT_LL, int32(32-v), &al)

				if bh.Type.Etype == gc.TINT32 {
					//	MOVW	bh->v, ah
					gshift(arm.AMOVW, &bh, arm.SHIFT_AR, int32(v), &ah)
				} else {
					//	MOVW	bh>>v, ah
					gshift(arm.AMOVW, &bh, arm.SHIFT_LR, int32(v), &ah)
				}
			} else {
				gins(arm.AMOVW, &bl, &al)
				gins(arm.AMOVW, &bh, &ah)
			}

			goto orsh_break
		}

		gc.Regalloc(&s, gc.Types[gc.TUINT32], nil)
		gc.Regalloc(&creg, gc.Types[gc.TUINT32], nil)
		if gc.Is64(r.Type) {
			// shift is >= 1<<32
			var ch gc.Node
			var cl gc.Node
			split64(r, &cl, &ch)

			gmove(&ch, &s)
			gins(arm.ATST, &s, nil)
			var p1 *obj.Prog
			if bh.Type.Etype == gc.TINT32 {
				p1 = gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
			} else {
				p1 = gins(arm.AEOR, &ah, &ah)
			}
			p1.Scond = arm.C_SCOND_NE
			p6 = gc.Gbranch(arm.ABNE, nil, 0)
			gmove(&cl, &s)
			splitclean()
		} else {
			gmove(r, &s)
			p6 = nil
		}

		gins(arm.ATST, &s, nil)

		// shift == 0
		p1 = gins(arm.AMOVW, &bl, &al)

		p1.Scond = arm.C_SCOND_EQ
		p1 = gins(arm.AMOVW, &bh, &ah)
		p1.Scond = arm.C_SCOND_EQ
		p2 = gc.Gbranch(arm.ABEQ, nil, 0)

		// check if shift is < 32
		gc.Nodconst(&n1, gc.Types[gc.TUINT32], 32)

		gmove(&n1, &creg)
		gins(arm.ACMP, &s, &creg)

		//	MOVW.LO		bl>>s, al
		p1 = gregshift(arm.AMOVW, &bl, arm.SHIFT_LR, &s, &al)

		p1.Scond = arm.C_SCOND_LO

		//	SUB.LO		s,creg
		p1 = gins(arm.ASUB, &s, &creg)

		p1.Scond = arm.C_SCOND_LO

		//	OR.LO		bh<<(32-s), al
		p1 = gregshift(arm.AORR, &bh, arm.SHIFT_LL, &creg, &al)

		p1.Scond = arm.C_SCOND_LO

		if bh.Type.Etype == gc.TINT32 {
			//	MOVW	bh->s, ah
			p1 = gregshift(arm.AMOVW, &bh, arm.SHIFT_AR, &s, &ah)
		} else {
			//	MOVW	bh>>s, ah
			p1 = gregshift(arm.AMOVW, &bh, arm.SHIFT_LR, &s, &ah)
		}

		p1.Scond = arm.C_SCOND_LO

		//	BLO	end
		p3 = gc.Gbranch(arm.ABLO, nil, 0)

		// shift == 32
		p1 = gins(arm.AMOVW, &bh, &al)

		p1.Scond = arm.C_SCOND_EQ
		if bh.Type.Etype == gc.TINT32 {
			gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &ah)
		} else {
			gins(arm.AEOR, &ah, &ah)
		}
		p4 = gc.Gbranch(arm.ABEQ, nil, 0)

		// check if shift is < 64
		gc.Nodconst(&n1, gc.Types[gc.TUINT32], 64)

		gmove(&n1, &creg)
		gins(arm.ACMP, &s, &creg)

		//	MOVW.LO		creg>>1, creg
		p1 = gshift(arm.AMOVW, &creg, arm.SHIFT_LR, 1, &creg)

		p1.Scond = arm.C_SCOND_LO

		//	SUB.LO		creg, s
		p1 = gins(arm.ASUB, &creg, &s)

		p1.Scond = arm.C_SCOND_LO

		if bh.Type.Etype == gc.TINT32 {
			//	MOVW	bh->(s-32), al
			p1 := gregshift(arm.AMOVW, &bh, arm.SHIFT_AR, &s, &al)

			p1.Scond = arm.C_SCOND_LO
		} else {
			//	MOVW	bh>>(v-32), al
			p1 := gregshift(arm.AMOVW, &bh, arm.SHIFT_LR, &s, &al)

			p1.Scond = arm.C_SCOND_LO
		}

		//	BLO	end
		p5 = gc.Gbranch(arm.ABLO, nil, 0)

		// s >= 64
		if p6 != nil {
			gc.Patch(p6, gc.Pc)
		}
		if bh.Type.Etype == gc.TINT32 {
			//	MOVW	bh->31, al
			gshift(arm.AMOVW, &bh, arm.SHIFT_AR, 31, &al)
		} else {
			gins(arm.AEOR, &al, &al)
		}

		gc.Patch(p2, gc.Pc)
		gc.Patch(p3, gc.Pc)
		gc.Patch(p4, gc.Pc)
		gc.Patch(p5, gc.Pc)
		gc.Regfree(&s)
		gc.Regfree(&creg)

	orsh_break:
		gc.Regfree(&bl)
		gc.Regfree(&bh)

		// TODO(kaib): literal optimizations
	// make constant the right side (it usually is anyway).
	//		if(lo1.op == OLITERAL) {
	//			nswap(&lo1, &lo2);
	//			nswap(&hi1, &hi2);
	//		}
	//		if(lo2.op == OLITERAL) {
	//			// special cases for constants.
	//			lv = mpgetfix(lo2.val.u.xval);
	//			hv = mpgetfix(hi2.val.u.xval);
	//			splitclean();	// right side
	//			split64(res, &lo2, &hi2);
	//			switch(n->op) {
	//			case OXOR:
	//				gmove(&lo1, &lo2);
	//				gmove(&hi1, &hi2);
	//				switch(lv) {
	//				case 0:
	//					break;
	//				case 0xffffffffu:
	//					gins(ANOTL, N, &lo2);
	//					break;
	//				default:
	//					gins(AXORL, ncon(lv), &lo2);
	//					break;
	//				}
	//				switch(hv) {
	//				case 0:
	//					break;
	//				case 0xffffffffu:
	//					gins(ANOTL, N, &hi2);
	//					break;
	//				default:
	//					gins(AXORL, ncon(hv), &hi2);
	//					break;
	//				}
	//				break;

	//			case OAND:
	//				switch(lv) {
	//				case 0:
	//					gins(AMOVL, ncon(0), &lo2);
	//					break;
	//				default:
	//					gmove(&lo1, &lo2);
	//					if(lv != 0xffffffffu)
	//						gins(AANDL, ncon(lv), &lo2);
	//					break;
	//				}
	//				switch(hv) {
	//				case 0:
	//					gins(AMOVL, ncon(0), &hi2);
	//					break;
	//				default:
	//					gmove(&hi1, &hi2);
	//					if(hv != 0xffffffffu)
	//						gins(AANDL, ncon(hv), &hi2);
	//					break;
	//				}
	//				break;

	//			case OOR:
	//				switch(lv) {
	//				case 0:
	//					gmove(&lo1, &lo2);
	//					break;
	//				case 0xffffffffu:
	//					gins(AMOVL, ncon(0xffffffffu), &lo2);
	//					break;
	//				default:
	//					gmove(&lo1, &lo2);
	//					gins(AORL, ncon(lv), &lo2);
	//					break;
	//				}
	//				switch(hv) {
	//				case 0:
	//					gmove(&hi1, &hi2);
	//					break;
	//				case 0xffffffffu:
	//					gins(AMOVL, ncon(0xffffffffu), &hi2);
	//					break;
	//				default:
	//					gmove(&hi1, &hi2);
	//					gins(AORL, ncon(hv), &hi2);
	//					break;
	//				}
	//				break;
	//			}
	//			splitclean();
	//			splitclean();
	//			goto out;
	//		}
	case gc.OXOR,
		gc.OAND,
		gc.OOR:
		var n1 gc.Node
		gc.Regalloc(&n1, lo1.Type, nil)

		gins(arm.AMOVW, &lo1, &al)
		gins(arm.AMOVW, &hi1, &ah)
		gins(arm.AMOVW, &lo2, &n1)
		gins(optoas(int(n.Op), lo1.Type), &n1, &al)
		gins(arm.AMOVW, &hi2, &n1)
		gins(optoas(int(n.Op), lo1.Type), &n1, &ah)
		gc.Regfree(&n1)
	}

	if gc.Is64(r.Type) {
		splitclean()
	}
	splitclean()

	split64(res, &lo1, &hi1)
	gins(arm.AMOVW, &al, &lo1)
	gins(arm.AMOVW, &ah, &hi1)
	splitclean()

	//out:
	gc.Regfree(&al)

	gc.Regfree(&ah)
}