VPADDQ Z6, Z2, K3, Z5                              // 62f1ed4bd4ee
	VPADDQ Z14, Z2, K3, Z5                             // 62d1ed4bd4ee
	VPADDQ 99(R15)(R15*2), Z2, K3, Z5                  // 6291ed4bd4ac7f63000000
	VPADDQ -7(DI), Z2, K3, Z5                          // 62f1ed4bd4aff9ffffff
	VPADDQ Z6, Z2, K3, Z23                             // 62e1ed4bd4fe

	VPADDQ (BX), Y15, Y2            // c4e105d413 or c585d413
	VPADDQ (R11), Y15, Y2           // c4c105d413
	VPADDQ Y2, Y15, Y2              // c4e105d4d2 or c585d4d2
	VPADDQ Y11, Y15, Y2             // c4c105d4d3
	VPADDQ (BX), Y15, Y11           // c46105d41b or c505d41b
	VPADDQ (R11), Y15, Y11          // c44105d41b
	VPADDQ Y2, Y15, Y11             // c46105d4da or c505d4da

BP VMOVDQU (DI), Y4 VPSHUFB Y9, Y4, Y4 VMOVDQU 32(DI), Y5 VPSHUFB Y9, Y5, Y5 VMOVDQU 64(DI), Y6 VPSHUFB Y9, Y6, Y6 VMOVDQU 96(DI), Y7 VPSHUFB Y9, Y7, Y7 MOVQ DI, 40(SP) MOVQ $0x00000004, 32(SP) loop1: VPADDQ (BP), Y4, Y0 VMOVDQU Y0, (SP) VPERM2F128 $0x03, Y6, Y7, Y0 VPALIGNR $0x08, Y6, Y0, Y0 VPADDQ Y4, Y0, Y0 VPERM2F128 $0x03, Y4, Y5, Y1 VPALIGNR $0x08, Y4, Y1, Y1 VPSRLQ $0x01, Y1, Y2 VPSLLQ $0x3f, Y1, Y3 VPOR Y2, Y3, Y3 VPSRLQ $0x07, Y1, Y8 MOVQ AX, DI RORXQ $0x29, DX, R13 RORXQ $0x12, DX, R14 ADDQ...

BP VMOVDQU (DI), Y4 VPSHUFB Y9, Y4, Y4 VMOVDQU 32(DI), Y5 VPSHUFB Y9, Y5, Y5 VMOVDQU 64(DI), Y6 VPSHUFB Y9, Y6, Y6 VMOVDQU 96(DI), Y7 VPSHUFB Y9, Y7, Y7 MOVQ DI, 40(SP) MOVQ $0x00000004, 32(SP) loop1: VPADDQ (BP), Y4, Y0 VMOVDQU Y0, (SP) VPERM2F128 $0x03, Y6, Y7, Y0 VPALIGNR $0x08, Y6, Y0, Y0 VPADDQ Y4, Y0, Y0 VPERM2F128 $0x03, Y4, Y5, Y1 VPALIGNR $0x08, Y4, Y1, Y1 VPSRLQ $0x01, Y1, Y2 VPSLLQ $0x3f, Y1, Y3 VPOR Y2, Y3, Y3 VPSRLQ $0x07, Y1, Y8 MOVQ AX, DI RORXQ $0x29, DX, R13 RORXQ $0x12, DX, R14 ADDQ...