VPERMB -7(CX)(DX*8), Y31, K7, Y22                  // 62e205278db4d1f9ffffff
	VPERMB Y5, Y5, K7, Y22                             // 62e2552f8df5
	VPERMB Y19, Y5, K7, Y22                            // 62a2552f8df3
	VPERMB Y31, Y5, K7, Y22                            // 6282552f8df7
	VPERMB 17(SP)(BP*1), Y5, K7, Y22                   // 62e2552f8db42c11000000

	VBROADCASTF64X2 15(DX)(BX*1), K1, Z14              // 6272fd491ab41a0f000000
	VBROADCASTF64X2 -7(CX)(DX*2), K1, Z14              // 6272fd491ab451f9ffffff
	VBROADCASTF64X2 15(DX)(BX*1), K1, Z28              // 6262fd491aa41a0f000000
	VBROADCASTF64X2 -7(CX)(DX*2), K1, Z28              // 6262fd491aa451f9ffffff
	VBROADCASTI32X2 X14, K1, X19                       // 62c27d0959de

32(BX), AX MULQ SI ADDQ 32(CX), AX ADCQ $0x00, DX ADDQ DI, AX ADCQ $0x00, DX MOVQ DX, DI MOVQ AX, 32(CX) // Iteration 5 MOVQ 40(BX), AX MULQ SI ADDQ 40(CX), AX ADCQ $0x00, DX ADDQ DI, AX ADCQ $0x00, DX MOVQ DX, DI MOVQ AX, 40(CX) // Iteration 6 MOVQ 48(BX), AX MULQ SI ADDQ 48(CX), AX ADCQ $0x00, DX ADDQ DI, AX ADCQ $0x00, DX MOVQ DX, DI MOVQ AX, 48(CX) // Iteration 7 MOVQ 56(BX), AX MULQ SI ADDQ 56(CX), AX ADCQ $0x00, DX ADDQ DI, AX ADCQ $0x00, DX MOVQ DX, DI MOVQ AX, 56(CX) // Iteration 8 MOVQ 64(BX),...

	VSCATTERPF0DPD K5, (SP)(Y4*2)                      // 62f2fd4dc62c64
	VSCATTERPF0DPD K5, (DX)(Y10*4)                     // 62b2fd4dc62c92
	VSCATTERPF0DPS K3, (DX)(Z10*4)                     // 62b27d4bc62c92
	VSCATTERPF0DPS K3, (AX)(Z4*1)                      // 62f27d4bc62c20
	VSCATTERPF0DPS K3, (SP)(Z4*2)                      // 62f27d4bc62c64
	VSCATTERPF0QPD K4, (DX)(Z10*4)                     // 62b2fd4cc72c92

	VGF2P8AFFINEINVQB $47, (DX), Z14, K2, Z28          // 62638d4acf222f
	VGF2P8AFFINEINVQB $47, Z3, Z28, K2, Z28            // 62639d42cfe32f
	VGF2P8AFFINEINVQB $47, Z12, Z28, K2, Z28           // 62439d42cfe42f
	VGF2P8AFFINEINVQB $47, 99(R15)(R15*1), Z28, K2, Z28 // 62039d42cfa43f630000002f
	VGF2P8AFFINEINVQB $47, (DX), Z28, K2, Z28          // 62639d42cf222f

	VAESDEC (DX), Y0, Y23                              // 62e27d28de3a or 62e2fd28de3a
	VAESDEC Z27, Z3, Z11                               // 62126548dedb or 6212e548dedb
	VAESDEC Z15, Z3, Z11                               // 62526548dedf or 6252e548dedf
	VAESDEC 99(R15)(R15*1), Z3, Z11                    // 62126548de9c3f63000000 or 6212e548de9c3f63000000

	VP4DPWSSD 15(DX)(BX*8), [Z2-Z5], K4, Z17           // 62e26f4c528cda0f000000
	VP4DPWSSD 7(SI)(DI*1), [Z12-Z15], K4, Z17          // 62e21f4c528c3e07000000
	VP4DPWSSD 15(DX)(BX*8), [Z12-Z15], K4, Z17         // 62e21f4c528cda0f000000
	VP4DPWSSD 7(SI)(DI*1), [Z22-Z25], K4, Z17          // 62e24f44528c3e07000000
	VP4DPWSSD 15(DX)(BX*8), [Z22-Z25], K4, Z17         // 62e24f44528cda0f000000

	VPLZCNTD 15(DX)(BX*1), K3, X17                     // 62e27d0b448c1a0f000000
	VPLZCNTD -7(CX)(DX*2), K3, X17                     // 62e27d0b448c51f9ffffff
	VPLZCNTD X3, K3, X15                               // 62727d0b44fb
	VPLZCNTD X26, K3, X15                              // 62127d0b44fa
	VPLZCNTD X23, K3, X15                              // 62327d0b44ff

	V4FNMADDPS 99(R15)(R15*1), [Z1-Z4], K3, Z15        // 6212774baabc3f63000000
	V4FNMADDPS (DX), [Z1-Z4], K3, Z15                  // 6272774baa3a
	V4FNMADDPS 99(R15)(R15*1), [Z11-Z14], K3, Z15      // 6212274baabc3f63000000
	V4FNMADDPS (DX), [Z11-Z14], K3, Z15                // 6272274baa3a
	V4FNMADDPS 99(R15)(R15*1), [Z21-Z24], K3, Z15      // 62125743aabc3f63000000
	V4FNMADDPS (DX), [Z21-Z24], K3, Z15                // 62725743aa3a

// LTYPEM spec6	{ outcode($1, &$2); }
	MOVL	AX, R11
	MOVL	$4, R11
//	MOVL	AX, 0(AX):DS // no longer works - did it ever?

// LTYPEI spec7	{ outcode($1, &$2); }
	IMULB	DX
	IMULW	DX, BX
	IMULL	R11, R12
	IMULQ	foo+4(SB), R11

// LTYPEXC spec8	{ outcode($1, &$2); }
	CMPPD	X1, X2, 4
	CMPPD	foo+4(SB), X2, 4

// LTYPEX spec9	{ outcode($1, &$2); }
	PINSRW	$4, AX, X2