U3:	// n >= 0
	// regular loop body unrolled 4x
	MOVQ 0(R8)(SI*8), R11
	MOVQ 8(R8)(SI*8), R12
	MOVQ 16(R8)(SI*8), R13
	MOVQ 24(R8)(SI*8), R14
	ADDQ CX, R11
	ADCQ $0, R12
	ADCQ $0, R13
	ADCQ $0, R14
	SBBQ CX, CX		// save CF
	NEGQ CX
	MOVQ R11, 0(R10)(SI*8)
	MOVQ R12, 8(R10)(SI*8)
	MOVQ R13, 16(R10)(SI*8)
	MOVQ R14, 24(R10)(SI*8)

	ADDQ $4, SI		// i += 4
	SUBQ $4, DI		// n -= 4

// Registers are cyclically rotated DX -> AX -> DI -> SI -> BX -> CX
#define CALC_0 \
	MOVL SI, BX \ // Precalculating first round
	RORXL $2, SI, SI \
	ANDNL AX, BX, BP \
	ANDL DI, BX \
	XORL BP, BX \
	CALC_F1_PRE(0x0,CX,BX,DI,DX) \
	PRECALC_0(0x80) \
	CALC_F1_POST(CX,SI,DX)

#define CALC_1 \
	CALC_F1_PRE(0x4,DX,CX,SI,AX) \
	PRECALC_1(0x80) \
	CALC_F1_POST(DX,BX,AX)

#define CALC_2 \

// func mulAddVWW(z, x []Word, y, r Word) (c Word)
TEXT ·mulAddVWW(SB),NOSPLIT,$0
	MOVL z+0(FP), DI
	MOVL x+12(FP), SI
	MOVL y+24(FP), BP
	MOVL r+28(FP), CX	// c = r
	MOVL z_len+4(FP), BX
	LEAL (DI)(BX*4), DI
	LEAL (SI)(BX*4), SI
	NEGL BX			// i = -n
	JMP E5

L5:	MOVL (SI)(BX*4), AX
	MULL BP
	ADDL CX, AX
	ADCL $0, DX
	MOVL AX, (DI)(BX*4)
	MOVL DX, CX
	ADDL $1, BX		// i++

	LEAQ	0(BX*8), CX
	NEGQ	CX

	CMPB	SI, $0xf8
	JA	si_high

	// load at SI won't cross a page boundary.
	MOVQ	(SI), SI
	JMP	si_finish
si_high:
	// address ends in 11111xxx. Load up to bytes we want, move to correct position.
	MOVQ	-8(SI)(BX*1), SI
	SHRQ	CX, SI
si_finish:

	// same for DI.
	CMPB	DI, $0xf8
	JA	di_high
	MOVQ	(DI), DI
	JMP	di_finish
di_high:
	MOVQ	-8(DI)(BX*1), DI

	offsets := []struct {
		name string
		got  uintptr
		want int
	}{
		{"Signo", unsafe.Offsetof(si.Signo), ofSigno},
		{"Errno", unsafe.Offsetof(si.Errno), ofErrno},
		{"Code", unsafe.Offsetof(si.Code), ofCode},
		{"Pid", unsafe.Offsetof(si.Pid), ofPid},
		{"Uid", unsafe.Offsetof(si.Uid), ofUid},
		{"Status", unsafe.Offsetof(si.Status), ofStatus},
	}

	for _, tc := range offsets {
		if int(tc.got) != tc.want {

				// not enough padding
				return si, 0, CorruptInputError(len(src))
			}
			if rune(src[si]) != enc.padChar {
				// incorrect padding
				return si, 0, CorruptInputError(si - 1)
			}

			si++
		}

		// skip over newlines
		for si < len(src) && (src[si] == '\n' || src[si] == '\r') {
			si++
		}
		if si < len(src) {
			// trailing garbage
			err = CorruptInputError(si)
		}
		dlen = j
		break

	MOVQ b+16(FP), BX

	// r0 = a0×b0
	MOVQ (CX), AX
	MULQ (BX)
	MOVQ AX, DI
	MOVQ DX, SI

	// r0 += 19×a1×b4
	MOVQ   8(CX), AX
	IMUL3Q $0x13, AX, AX
	MULQ   32(BX)
	ADDQ   AX, DI
	ADCQ   DX, SI

	// r0 += 19×a2×b3
	MOVQ   16(CX), AX
	IMUL3Q $0x13, AX, AX
	MULQ   24(BX)
	ADDQ   AX, DI
	ADCQ   DX, SI

	// r0 += 19×a3×b2
	MOVQ   24(CX), AX
	IMUL3Q $0x13, AX, AX
	MULQ   16(BX)
	ADDQ   AX, DI

// The start of the data is SI.
ssesuccess:
	SUBQ SI, DI	// Compute offset of chunk within data.
	ADDQ DX, DI	// Add offset of byte within chunk.
	MOVQ DI, (R8)
	RET

// handle for lengths < 16
small:
	TESTQ	BX, BX
	JEQ	failure

	// Check if we'll load across a page boundary.
	LEAQ	16(SI), AX
	TESTW	$0xff0, AX
	JEQ	endofpage

	MOVOU	(SI), X1 // Load data

	var m TraceMap

	var wg sync.WaitGroup
	for i := range 3 {
		wg.Add(1)
		go func(i int) {
			defer wg.Done()

			si := strconv.Itoa(i)
			var d = [...]string{
				"a" + si,
				"b" + si,
				"aa" + si,
				"ab" + si,
				"ba" + si,
				"bb" + si,
			}
			ids := make([]uint64, 0, len(d))
			for _, s := range d {
				id, inserted := m.PutString(s)
				if !inserted {

	MOVL	dig+0(FP),	BP
	MOVL	p+4(FP),	SI
	MOVL	p_len+8(FP), DX
	SHRL	$6,		DX
	SHLL	$6,		DX

	LEAL	(SI)(DX*1),	DI
	MOVL	(0*4)(BP),	AX
	MOVL	(1*4)(BP),	BX
	MOVL	(2*4)(BP),	CX
	MOVL	(3*4)(BP),	DX

	CMPL	SI,		DI
	JEQ	end

	MOVL	DI,		16(SP)

loop:
	MOVL	AX,		0(SP)
	MOVL	BX,		4(SP)
	MOVL	CX,		8(SP)
	MOVL	DX,		12(SP)

	MOVL	(0*4)(SI),	DI
	MOVL	DX,		BP