VMOVNTPD Y26, (AX)                                 // 6261fd282b10
	VMOVNTPD Y26, 7(SI)                                // 6261fd282b9607000000
	VMOVNTPD Z7, (SI)                                  // 62f1fd482b3e
	VMOVNTPD Z13, (SI)                                 // 6271fd482b2e
	VMOVNTPD Z7, 7(SI)(DI*2)                           // 62f1fd482bbc7e07000000
	VMOVNTPD Z13, 7(SI)(DI*2)                          // 6271fd482bac7e07000000

	VPOPCNTD Z31, K3, Z22                              // 62827d4b55f7
	VPOPCNTD 7(SI)(DI*4), K3, Z22                      // 62e27d4b55b4be07000000
	VPOPCNTD -7(DI)(R8*2), K3, Z22                     // 62a27d4b55b447f9ffffff
	VPOPCNTD Z2, K3, Z7                                // 62f27d4b55fa
	VPOPCNTD Z31, K3, Z7                               // 62927d4b55ff
	VPOPCNTD 7(SI)(DI*4), K3, Z7                       // 62f27d4b55bcbe07000000

	VPDPBUSD (SI), Z6, K3, Z23                         // 62e24d4b503e
	VPDPBUSD 7(SI)(DI*2), Z6, K3, Z23                  // 62e24d4b50bc7e07000000
	VPDPBUSD Z8, Z23, K3, Z5                           // 62d2454350e8
	VPDPBUSD Z28, Z23, K3, Z5                          // 6292454350ec
	VPDPBUSD (SI), Z23, K3, Z5                         // 62f24543502e

	V4FMADDPS -17(BP)(SI*4), [Z0-Z3], K2, Z0           // 62f27f4a9a84b5efffffff
	V4FMADDPS 17(SP), [Z10-Z13], K2, Z0                // 62f22f4a9a842411000000
	V4FMADDPS -17(BP)(SI*4), [Z10-Z13], K2, Z0         // 62f22f4a9a84b5efffffff
	V4FMADDPS 17(SP), [Z20-Z23], K2, Z0                // 62f25f429a842411000000
	V4FMADDPS -17(BP)(SI*4), [Z20-Z23], K2, Z0         // 62f25f429a84b5efffffff

	VPERMB 7(SI)(DI*4), X20, K1, X23                   // 62e25d018dbcbe07000000
	VPERMB -7(DI)(R8*2), X20, K1, X23                  // 62a25d018dbc47f9ffffff
	VPERMB X26, X2, K1, X23                            // 62826d098dfa
	VPERMB X19, X2, K1, X23                            // 62a26d098dfb
	VPERMB X0, X2, K1, X23                             // 62e26d098df8

	VPCLMULQDQ $97, 7(SI)(DI*1), Z0, Z12               // 62737d4844a43e0700000061 or 6273fd4844a43e0700000061
	VPCLMULQDQ $97, 15(DX)(BX*8), Z0, Z12              // 62737d4844a4da0f00000061 or 6273fd4844a4da0f00000061
	VPCLMULQDQ $97, Z9, Z26, Z12                       // 62532d4044e161 or 6253ad4044e161
	VPCLMULQDQ $97, Z3, Z26, Z12                       // 62732d4044e361 or 6273ad4044e361

TEXT asmtest_avx512_4vnniw(SB), NOSPLIT, $0
	VP4DPWSSD 7(SI)(DI*1), [Z2-Z5], K4, Z17            // 62e26f4c528c3e07000000
	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

	VAESDECLAST -17(BP)(SI*8), Y11, Y24                // 62622528df84f5efffffff or 6262a528df84f5efffffff
	VAESDECLAST (R15), Y11, Y24                        // 62422528df07 or 6242a528df07
	VAESDECLAST Z8, Z23, Z23                           // 62c24540dff8 or 62c2c540dff8
	VAESDECLAST Z28, Z23, Z23                          // 62824540dffc or 6282c540dffc

	MOVOA	(AX), X1	// MOVO (AX), X1

// Tests for SP indexed addresses.
	MOVQ	foo(SP)(AX*1), BX		// 488b1c04
	MOVQ	foo+32(SP)(CX*2), DX		// 488b544c20
	MOVQ	foo+32323(SP)(R8*4), R9		// 4e8b8c84437e0000
	MOVL	foo(SP)(SI*8), DI		// 8b3cf4
	MOVL	foo+32(SP)(R10*1), R11		// 468b5c1420
	MOVL	foo+32323(SP)(R12*2), R13	// 468bac64437e0000
	MOVW	foo(SP)(AX*4), R8		// 66448b0484
	MOVW	foo+32(SP)(R9*8), CX		// 66428b4ccc20

	MOVL CR0, AX // 0f20c0
	MOVL CR0, DX // 0f20c2
	MOVL CR4, DI // 0f20e7
	MOVL AX, CR0 // 0f22c0
	MOVL DX, CR0 // 0f22c2
	MOVL DI, CR4 // 0f22e7
	MOVL DR0, AX // 0f21c0
	MOVL DR6, DX // 0f21f2
	MOVL DR7, SI // 0f21fe
	// Test other movtab entries.
	PUSHL SS // 16
	PUSHL FS // 0fa0
	POPL FS  // 0fa1
	POPL SS  // 17

	RDPID AX                                // f30fc7f8

	// End of tests.