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

	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

	V4FMADDSS (DI), [X0-X3], K5, X3                    // 62f27f0d9b1f or 62f27f2d9b1f or 62f27f4d9b1f
	V4FMADDSS 7(AX), [X10-X13], K5, X3                 // 62f22f0d9b9807000000 or 62f22f2d9b9807000000 or 62f22f4d9b9807000000
	V4FMADDSS (DI), [X10-X13], K5, X3                  // 62f22f0d9b1f or 62f22f2d9b1f or 62f22f4d9b1f
	V4FMADDSS 7(AX), [X20-X23], K5, X3                 // 62f25f059b9807000000 or 62f25f259b9807000000 or 62f25f459b9807000000

	VPMADD52HUQ -7(DI)(R8*4), X11, K1, X18             // 62a2a509b59487f9ffffff
	VPMADD52HUQ X7, X31, K1, X18                       // 62e28501b5d7
	VPMADD52HUQ X0, X31, K1, X18                       // 62e28501b5d0
	VPMADD52HUQ 17(SP)(BP*2), X31, K1, X18             // 62e28501b5946c11000000
	VPMADD52HUQ -7(DI)(R8*4), X31, K1, X18             // 62a28501b59487f9ffffff

	VPDPWSSD 7(SI)(DI*4), Z6, K3, Z22                  // 62e24d4b52b4be07000000
	VPDPWSSD -7(DI)(R8*2), Z6, K3, Z22                 // 62a24d4b52b447f9ffffff
	VPDPWSSD Z25, Z8, K3, Z22                          // 62823d4b52f1
	VPDPWSSD Z12, Z8, K3, Z22                          // 62c23d4b52f4
	VPDPWSSD 7(SI)(DI*4), Z8, K3, Z22                  // 62e23d4b52b4be07000000

	VPCOMPRESSW X12, K5, 17(SP)(BP*2)                  // 6272fd0d63a46c11000000
	VPCOMPRESSW X20, K5, -7(DI)(R8*4)                  // 62a2fd0d63a487f9ffffff
	VPCOMPRESSW X16, K5, -7(DI)(R8*4)                  // 62a2fd0d638487f9ffffff
	VPCOMPRESSW X12, K5, -7(DI)(R8*4)                  // 6232fd0d63a487f9ffffff
	VPCOMPRESSW Y18, K7, Y14                           // 62c2fd2f63d6

	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

	ADDL -2147483648(AX), AX // 038000000080
	// Make sure MOV CR/DR continues to work after changing its movtabs.
	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

	VGF2P8AFFINEQB $82, -7(DI)(R8*4), X21, K1, X15     // 6233d501cebc87f9ffffff52
	VGF2P8AFFINEQB $82, X22, X0, K1, X15               // 6233fd09cefe52
	VGF2P8AFFINEQB $82, X7, X0, K1, X15                // 6273fd09ceff52
	VGF2P8AFFINEQB $82, X19, X0, K1, X15               // 6233fd09cefb52
	VGF2P8AFFINEQB $82, 17(SP)(BP*2), X0, K1, X15      // 6273fd09cebc6c1100000052

	VPAVGB 7(SI)(DI*4), Z18, K4, Z11                   // 62716d44e09cbe07000000 or 6271ed44e09cbe07000000
	VPAVGB -7(DI)(R8*2), Z18, K4, Z11                  // 62316d44e09c47f9ffffff or 6231ed44e09c47f9ffffff
	VPAVGB Z2, Z24, K4, Z11                            // 62713d44e0da or 6271bd44e0da
	VPAVGB Z21, Z24, K4, Z11                           // 62313d44e0dd or 6231bd44e0dd