JCS	2(PC)
	RETFL	$4

// Was bug: LOOP is a branch instruction.
	JCS	2(PC)
loop:
	LOOP	loop // LOOP

	// Intel pseudonyms for our own renamings.
	PADDD	M2, M1 // PADDL M2, M1
	MOVDQ2Q	X1, M1 // MOVQ X1, M1
	MOVNTDQ	X1, (AX)	// MOVNTO X1, (AX)
	MOVOA	(AX), X1	// MOVO (AX), X1

// Tests for SP indexed addresses.
	MOVQ	foo(SP)(AX*1), BX		// 488b1c04

	PADDL (BX), X2                          // 660ffe13
	PADDL (R11), X2                         // 66410ffe13
	PADDL X2, X2                            // 660ffed2
	PADDL X11, X2                           // 66410ffed3
	PADDL (BX), X11                         // 66440ffe1b
	PADDL (R11), X11                        // 66450ffe1b
	PADDL X2, X11                           // 66440ffeda

X14, X14 PXOR X0, X0 PXOR X1, X1 PXOR X2, X2 PXOR X3, X3 MOVQ $0x00000010, AX MOVOU X15, X13 loop_select_base: MOVOU X13, X12 PADDL X15, X13 PCMPEQL X14, X12 MOVOU (DI), X4 MOVOU 16(DI), X5 MOVOU 32(DI), X6 MOVOU 48(DI), X7 MOVOU 64(DI), X8 MOVOU 80(DI), X9 MOVOU 96(DI), X10 MOVOU 112(DI), X11 ADDQ $0x80, DI PAND X12, X4 PAND X12, X5 PAND X12, X6 PAND X12, X7 MOVOU X13, X12 PADDL X15, X13 PCMPEQL X14, X12 PAND X12, X8 PAND X12, X9 PAND X12, X10 PAND X12, X11 PXOR X4, X0 PXOR X5, X1 PXOR X6, X2 PXOR X7,...