WebSVN – planix.SVN – Blame – /os/branches/feature_unix/sys/src/libsec/amd64/aesni.s

Rev	Author	Line No.	Line
26	7u83	1	`#define AESOP(o,r1,r2) \`
		2	`BYTE $0x66; \`
		3	`BYTE $0x0F; \`
		4	`BYTE $0x38; \`
		5	`BYTE $(o); \`
		6	`BYTE $(0xC0 \| r2<<3 \| r1)`
		7
		8	`#define AESIMC(r1,r2) AESOP(0xDB,r1,r2)`
		9	`#define AESENC(r1,r2) AESOP(0xDC,r1,r2)`
		10	`#define AESENCLAST(r1,r2) AESOP(0xDD,r1,r2)`
		11	`#define AESDEC(r1,r2) AESOP(0xDE,r1,r2)`
		12	`#define AESDECLAST(r1,r2) AESOP(0xDF,r1,r2)`
		13
		14	`#define AESKEYGENASSIST(i,r1,r2) \`
		15	`BYTE $0x66; \`
		16	`BYTE $0x0F; \`
		17	`BYTE $0x3A; \`
		18	`BYTE $0xDF; \`
		19	`BYTE $(0xC0 \| r2<<3 \| r1); \`
		20	`BYTE $(i)`
		21
		22	`TEXT aesni_init(SB), 0, $0`
		23	`MOVL $1, AX`
		24	`CPUID`
		25	`XORL AX, AX`
		26	`ANDL $(1<<25), CX`
		27	`JZ _ret`
		28
		29	`/* override aes function pointers */`
		30	`MOVQ $AESencrypt<>(SB), AX`
		31	`MOVQ AX, aes_encrypt(SB)`
		32	`MOVQ $AESdecrypt<>(SB), AX`
		33	`MOVQ AX, aes_decrypt(SB)`
		34
		35	`/* return setup function pointer */`
		36	`MOVQ $AESsetup<>(SB), AX`
		37	`_ret:`
		38	`RET`
		39
		40	`TEXT AESencrypt<>(SB), 0, $0`
		41	`MOVL Nr+8(FP), CX`
		42	`MOVQ pt+16(FP), SI`
		43	`MOVQ ct+24(FP), DI`
		44	`MOVO (RARG), X0`
		45	`MOVOU (SI), X7`
		46	`ADDQ $16, RARG`
		47	`PXOR X7, X0`
		48	`CMPL CX, $12`
		49	`JLT erounds10`
		50	`JEQ erounds12`
		51	`erounds14:`
		52	`MOVO 0(RARG), X1`
		53	`MOVO 16(RARG), X2`
		54	`ADDQ $32, RARG`
		55	`AESENC(1, 0)`
		56	`AESENC(2, 0)`
		57	`erounds12:`
		58	`MOVO 0(RARG), X3`
		59	`MOVO 16(RARG), X4`
		60	`ADDQ $32, RARG`
		61	`AESENC(3, 0)`
		62	`AESENC(4, 0)`
		63	`erounds10:`
		64	`MOVO 0(RARG), X1`
		65	`MOVO 16(RARG), X2`
		66	`MOVO 32(RARG), X3`
		67	`MOVO 48(RARG), X4`
		68	`MOVO 64(RARG), X5`
		69	`MOVO 80(RARG), X6`
		70	`MOVO 96(RARG), X7`
		71	`AESENC(1, 0)`
		72	`MOVO 112(RARG), X1`
		73	`AESENC(2, 0)`
		74	`MOVO 128(RARG), X2`
		75	`AESENC(3, 0)`
		76	`MOVO 144(RARG), X3`
		77	`AESENC(4, 0)`
		78	`AESENC(5, 0)`
		79	`AESENC(6, 0)`
		80	`AESENC(7, 0)`
		81
		82	`AESENC(1, 0)`
		83	`AESENC(2, 0)`
		84	`AESENCLAST(3, 0)`
		85	`MOVOU X0, (DI)`
		86	`RET`
		87
		88	`TEXT AESdecrypt<>(SB), 0, $0`
		89	`MOVL Nr+8(FP), CX`
		90	`MOVQ ct+16(FP), SI`
		91	`MOVQ pt+24(FP), DI`
		92	`MOVO (RARG), X0`
		93	`MOVOU (SI), X7`
		94	`ADDQ $16, RARG`
		95	`PXOR X7, X0`
		96	`CMPL CX, $12`
		97	`JLT drounds10`
		98	`JEQ drounds12`
		99	`drounds14:`
		100	`MOVO 0(RARG), X1`
		101	`MOVO 16(RARG), X2`
		102	`ADDQ $32, RARG`
		103	`AESDEC(1, 0)`
		104	`AESDEC(2, 0)`
		105	`drounds12:`
		106	`MOVO 0(RARG), X3`
		107	`MOVO 16(RARG), X4`
		108	`ADDQ $32, RARG`
		109	`AESDEC(3, 0)`
		110	`AESDEC(4, 0)`
		111	`drounds10:`
		112	`MOVO 0(RARG), X1`
		113	`MOVO 16(RARG), X2`
		114	`MOVO 32(RARG), X3`
		115	`MOVO 48(RARG), X4`
		116	`MOVO 64(RARG), X5`
		117	`MOVO 80(RARG), X6`
		118	`MOVO 96(RARG), X7`
		119	`AESDEC(1, 0)`
		120	`MOVO 112(RARG), X1`
		121	`AESDEC(2, 0)`
		122	`MOVO 128(RARG), X2`
		123	`AESDEC(3, 0)`
		124	`MOVO 144(RARG), X3`
		125	`AESDEC(4, 0)`
		126	`AESDEC(5, 0)`
		127	`AESDEC(6, 0)`
		128	`AESDEC(7, 0)`
		129
		130	`AESDEC(1, 0)`
		131	`AESDEC(2, 0)`
		132	`AESDECLAST(3, 0)`
		133	`MOVOU X0, (DI)`
		134	`RET`
		135
		136	`TEXT AESsetup<>(SB), 0, $16`
		137	`MOVQ RARG, erk+0(FP)`
		138	`MOVQ key+16(FP), DX`
		139	`MOVL nkey+24(FP), BX`
		140	`MOVQ DX, 8(SP)`
		141	`CMPL BX, $32`
		142	`JEQ esetup256`
		143	`CMPL BX, $24`
		144	`JEQ esetup192`
		145	`CMPL BX, $16`
		146	`JEQ esetup128`
		147	`XORL AX, AX`
		148	`RET`
		149	`esetup256:`
		150	`CALL setupEnc256<>(SB)`
		151	`JMP dsetup`
		152	`esetup192:`
		153	`CALL setupEnc192<>(SB)`
		154	`JMP dsetup`
		155	`esetup128:`
		156	`CALL setupEnc128<>(SB)`
		157	`dsetup:`
		158	`MOVQ erk+0(FP), SI`
		159	`MOVQ drk+8(FP), DI`
		160
		161	`MOVL AX, BX`
		162	`SHLL $4, BX`
		163	`ADDQ BX, SI`
		164
		165	`MOVO (SI), X0`
		166	`MOVO X0, (DI)`
		167
		168	`MOVO -16(SI), X1`
		169	`MOVO -32(SI), X2`
		170	`MOVO -48(SI), X3`
		171	`MOVO -64(SI), X4`
		172	`AESIMC(1, 1)`
		173	`AESIMC(2, 2)`
		174	`AESIMC(3, 3)`
		175	`AESIMC(4, 4)`
		176	`MOVO X1, 16(DI)`
		177	`MOVO X2, 32(DI)`
		178	`MOVO X3, 48(DI)`
		179	`MOVO X4, 64(DI)`
		180
		181	`MOVO -80(SI), X1`
		182	`MOVO -96(SI), X2`
		183	`MOVO -112(SI), X3`
		184	`MOVO -128(SI), X4`
		185	`AESIMC(1, 1)`
		186	`AESIMC(2, 2)`
		187	`AESIMC(3, 3)`
		188	`AESIMC(4, 4)`
		189	`MOVO X1, 80(DI)`
		190	`MOVO X2, 96(DI)`
		191	`MOVO X3, 112(DI)`
		192	`MOVO X4, 128(DI)`
		193
		194	`MOVO -144(SI), X1`
		195	`AESIMC(1, 1)`
		196	`MOVO X1, 144(DI)`
		197
		198	`CMPL AX, $10`
		199	`JEQ dsetupend`
		200
		201	`MOVO -160(SI), X1`
		202	`MOVO -176(SI), X2`
		203	`AESIMC(1, 1)`
		204	`AESIMC(2, 2)`
		205	`MOVO X1, 160(DI)`
		206	`MOVO X2, 176(DI)`
		207
		208	`CMPL AX, $12`
		209	`JEQ dsetupend`
		210
		211	`MOVO -192(SI), X1`
		212	`MOVO -208(SI), X2`
		213	`AESIMC(1, 1)`
		214	`AESIMC(2, 2)`
		215	`MOVO X1, 192(DI)`
		216	`MOVO X2, 208(DI)`
		217	`dsetupend:`
		218	`SUBQ BX, SI`
		219	`ADDQ BX, DI`
		220	`MOVO (SI), X0`
		221	`MOVO X0, (DI)`
		222	`RET`
		223
		224	`TEXT setupEnc128<>(SB), 0, $0`
		225	`MOVQ key+8(FP), SI`
		226	`MOVOU (SI), X1`
		227	`MOVO X1, (RARG)`
		228	`AESKEYGENASSIST(0x01, 1, 0)`
		229	`CALL rk128<>(SB)`
		230	`MOVO X1, 16(RARG)`
		231	`AESKEYGENASSIST(0x02, 1, 0)`
		232	`CALL rk128<>(SB)`
		233	`MOVO X1, 32(RARG)`
		234	`AESKEYGENASSIST(0x04, 1, 0)`
		235	`CALL rk128<>(SB)`
		236	`MOVO X1, 48(RARG)`
		237	`AESKEYGENASSIST(0x08, 1, 0)`
		238	`CALL rk128<>(SB)`
		239	`MOVO X1, 64(RARG)`
		240	`AESKEYGENASSIST(0x10, 1, 0)`
		241	`CALL rk128<>(SB)`
		242	`MOVO X1, 80(RARG)`
		243	`AESKEYGENASSIST(0x20, 1, 0)`
		244	`CALL rk128<>(SB)`
		245	`MOVO X1, 96(RARG)`
		246	`AESKEYGENASSIST(0x40, 1, 0)`
		247	`CALL rk128<>(SB)`
		248	`MOVO X1, 112(RARG)`
		249	`AESKEYGENASSIST(0x80, 1, 0)`
		250	`CALL rk128<>(SB)`
		251	`MOVO X1, 128(RARG)`
		252	`AESKEYGENASSIST(0x1b, 1, 0)`
		253	`CALL rk128<>(SB)`
		254	`MOVO X1, 144(RARG)`
		255	`AESKEYGENASSIST(0x36, 1, 0)`
		256	`CALL rk128<>(SB)`
		257	`MOVO X1, 160(RARG)`
		258	`MOVL $10, AX`
		259	`RET`
		260	`TEXT rk128<>(SB), 0, $0`
		261	`PSHUFL $0xff, X0, X0`
		262	`MOVO X1, X2`
		263	`PSLLO $4, X2`
		264	`PXOR X2, X1`
		265	`PSLLO $4, X2`
		266	`PXOR X2, X1`
		267	`PSLLO $4, X2`
		268	`PXOR X2, X1`
		269	`PXOR X0, X1`
		270	`RET`
		271
		272	`TEXT setupEnc192<>(SB), 0, $0`
		273	`MOVQ key+8(FP), SI`
		274	`MOVOU (SI), X1`
		275	`MOVOU 16(SI), X2`
		276	`MOVO X1, (RARG)`
		277	`MOVO X2, X5`
		278	`AESKEYGENASSIST(0x01, 2, 0)`
		279	`CALL rk192<>(SB)`
		280	`SHUFPD $0, X1, X5`
		281	`MOVO X5, 16(RARG)`
		282	`MOVO X1, X6`
		283	`SHUFPD $1, X2, X6`
		284	`MOVO X6, 32(RARG)`
		285	`AESKEYGENASSIST(0x02, 2, 0)`
		286	`CALL rk192<>(SB)`
		287	`MOVO X1, 48(RARG)`
		288	`MOVO X2, X5`
		289	`AESKEYGENASSIST(0x04, 2, 0)`
		290	`CALL rk192<>(SB)`
		291	`SHUFPD $0, X1, X5`
		292	`MOVO X5, 64(RARG)`
		293	`MOVO X1, X6`
		294	`SHUFPD $1, X2, X6`
		295	`MOVO X6, 80(RARG)`
		296	`AESKEYGENASSIST(0x08, 2, 0)`
		297	`CALL rk192<>(SB)`
		298	`MOVO X1, 96(RARG)`
		299	`MOVO X2, X5`
		300	`AESKEYGENASSIST(0x10, 2, 0)`
		301	`CALL rk192<>(SB)`
		302	`SHUFPD $0, X1, X5`
		303	`MOVO X5, 112(RARG)`
		304	`MOVO X1, X6`
		305	`SHUFPD $1, X2, X6`
		306	`MOVO X6, 128(RARG)`
		307	`AESKEYGENASSIST(0x20, 2, 0)`
		308	`CALL rk192<>(SB)`
		309	`MOVO X1, 144(RARG)`
		310	`MOVO X2, X5`
		311	`AESKEYGENASSIST(0x40, 2, 0)`
		312	`CALL rk192<>(SB)`
		313	`SHUFPD $0, X1, X5`
		314	`MOVO X5, 160(RARG)`
		315	`MOVO X1, X6`
		316	`SHUFPD $1, X2, X6`
		317	`MOVO X6, 176(RARG)`
		318	`AESKEYGENASSIST(0x80, 2, 0)`
		319	`CALL rk192<>(SB)`
		320	`MOVO X1, 192(RARG)`
		321	`MOVL $12, AX`
		322	`RET`
		323	`TEXT rk192<>(SB), 0, $0`
		324	`PSHUFL $0x55, X0, X0`
		325	`MOVOU X1, X4`
		326	`PSLLO $4, X4`
		327	`PXOR X4, X1`
		328	`PSLLO $4, X4`
		329	`PXOR X4, X1`
		330	`PSLLO $4, X4`
		331	`PXOR X4, X1`
		332	`PXOR X0, X1`
		333	`PSHUFL $0xff, X1, X0`
		334	`MOVOU X2, X4`
		335	`PSLLO $4, X4`
		336	`PXOR X4, X2`
		337	`PXOR X0, X2`
		338	`RET`
		339
		340	`TEXT setupEnc256<>(SB), 0, $0`
		341	`MOVQ key+8(FP), SI`
		342	`MOVOU (SI), X1`
		343	`MOVOU 16(SI), X2`
		344	`MOVO X1, (RARG)`
		345	`MOVO X2, 16(RARG)`
		346	`AESKEYGENASSIST(0x01, 2, 0)`
		347	`CALL rk256_a<>(SB)`
		348	`MOVO X1, 32(RARG)`
		349	`AESKEYGENASSIST(0x00, 1, 0)`
		350	`CALL rk256_b<>(SB)`
		351	`MOVO X2, 48(RARG)`
		352	`AESKEYGENASSIST(0x02, 2, 0)`
		353	`CALL rk256_a<>(SB)`
		354	`MOVO X1, 64(RARG)`
		355	`AESKEYGENASSIST(0x00, 1, 0)`
		356	`CALL rk256_b<>(SB)`
		357	`MOVO X2, 80(RARG)`
		358	`AESKEYGENASSIST(0x04, 2, 0)`
		359	`CALL rk256_a<>(SB)`
		360	`MOVO X1, 96(RARG)`
		361	`AESKEYGENASSIST(0x00, 1, 0)`
		362	`CALL rk256_b<>(SB)`
		363	`MOVO X2, 112(RARG)`
		364	`AESKEYGENASSIST(0x08, 2, 0)`
		365	`CALL rk256_a<>(SB)`
		366	`MOVO X1, 128(RARG)`
		367	`AESKEYGENASSIST(0x00, 1, 0)`
		368	`CALL rk256_b<>(SB)`
		369	`MOVO X2, 144(RARG)`
		370	`AESKEYGENASSIST(0x10, 2, 0)`
		371	`CALL rk256_a<>(SB)`
		372	`MOVO X1, 160(RARG)`
		373	`AESKEYGENASSIST(0x00, 1, 0)`
		374	`CALL rk256_b<>(SB)`
		375	`MOVO X2, 176(RARG)`
		376	`AESKEYGENASSIST(0x20, 2, 0)`
		377	`CALL rk256_a<>(SB)`
		378	`MOVO X1, 192(RARG)`
		379	`AESKEYGENASSIST(0x00, 1, 0)`
		380	`CALL rk256_b<>(SB)`
		381	`MOVO X2, 208(RARG)`
		382	`AESKEYGENASSIST(0x40, 2, 0)`
		383	`CALL rk256_a<>(SB)`
		384	`MOVO X1, 224(RARG)`
		385	`MOVL $14, AX`
		386	`RET`
		387	`TEXT rk256_a<>(SB), 0, $0`
		388	`PSHUFL $0xff, X0, X0`
		389	`MOVO X1, X4`
		390	`PSLLO $4, X4`
		391	`PXOR X4, X1`
		392	`PSLLO $4, X4`
		393	`PXOR X4, X1`
		394	`PSLLO $4, X4`
		395	`PXOR X4, X1`
		396	`PXOR X0, X1`
		397	`RET`
		398	`TEXT rk256_b<>(SB), 0, $0`
		399	`PSHUFL $0xaa, X0, X0`
		400	`MOVO X2, X4`
		401	`PSLLO $4, X4`
		402	`PXOR X4, X2`
		403	`PSLLO $4, X4`
		404	`PXOR X4, X2`
		405	`PSLLO $4, X4`
		406	`PXOR X4, X2`
		407	`PXOR X0, X2`
		408	`RET`

Subversion Repositories planix.SVN

(root)/os/branches/feature_unix/sys/src/libsec/amd64/aesni.s – Rev 60