x86-mmx.s

UNIX、linux密码的破密程序源代码实现

	pxor pnot,%mm6; \
	movq %mm5,%mm7; \
	movq %mm1,S6_a2; \
	movq %mm4,%mm2; \
	movq %mm3,S6_a4; \
	pxor %mm1,%mm7; \
	pxor pnot,%mm1; \
	pxor %mm6,%mm7; \
	movq %mm6,S6_x2; \
	pxor %mm0,%mm7; \
	pand %mm5,%mm2; \
	movq %mm4,%mm6; \
	movq %mm1,S6_x1; \
	movq %mm5,%mm3; \
	pand S6_a2,%mm3; \
	pand %mm7,%mm6; \
	movq %mm0,S6_a1; \
	por %mm2,%mm1; \
	movq %mm2,S6_x6; \
	pand %mm6,%mm0; \
	movq %mm3,S6_x15; \
	pxor %mm0,%mm1; \
	movq S6_a4,%mm0; \
	movq %mm4,%mm2; \
	movq %mm6,S6_x8; \
	pand %mm1,%mm0; \
	movq %mm7,S6_x5; \
	pxor %mm3,%mm2; \
	movq S6_x2,%mm6; \
	pxor %mm7,%mm0; \
	movq S6_a1,%mm7; \
	pxor %mm5,%mm1; \
	movq %mm2,S6_x16; \
	pand %mm7,%mm2; \
	movq S6_a4,%mm3; \
	pxor %mm2,%mm6; \
	pxor S6_a2,%mm2; \
	pand %mm7,%mm1; \
	por %mm6,%mm3; \
	pxor %mm5,%mm6; \
	pxor %mm3,%mm1; \
	pand %mm6,%mm7; \
	pand S6_a3,%mm1; \
	pand %mm4,%mm6; \
	movq S6_x6,%mm3; \
	pxor %mm1,%mm0; \
	pxor out2,%mm0; \
	por %mm2,%mm3; \
	pand S6_a4,%mm3; \
	pxor %mm7,%mm4; \
	movq S6_x5,%mm1; \
	pxor %mm3,%mm4; \
	pxor pnot,%mm2; \
	por %mm4,%mm5; \
	movq %mm0,out2; \
	movq %mm5,%mm3; \
	pandn S6_a4,%mm3; \
	pxor %mm6,%mm1; \
	movq S6_x6,%mm0; \
	pxor %mm2,%mm3; \
	por S6_a4,%mm1; \
	pxor %mm3,%mm0; \
	pand S6_a3,%mm3; \
	pxor %mm1,%mm0; \
	por S6_x5,%mm6; \
	movq %mm7,%mm1; \
	pxor S6_x15,%mm7; \
	pxor %mm3,%mm4; \
	movq S6_a4,%mm3; \
	pxor %mm5,%mm7; \
	pand S6_x8,%mm5; \
	por %mm3,%mm7; \
	pxor S6_x6,%mm6; \
	por %mm3,%mm5; \
	por S6_x16,%mm1; \
	pxor %mm6,%mm5; \
	pxor S6_x1,%mm1; \
	movq S6_a3,%mm3; \
	pxor %mm1,%mm7; \
	pxor out4,%mm4; \
	por %mm3,%mm7; \
	pand %mm1,%mm2; \
	pxor out1,%mm0; \
	por %mm3,%mm2; \
	pxor %mm7,%mm0; \
	pxor %mm5,%mm2; \
	movq %mm4,out4; \
	pxor out3,%mm2; \
	movq %mm0,out1; \
	movq %mm2,out3

#define S7_out1				%mm7
#define S7_out2				%mm1
#define S7_out3				%mm3
#define S7_out4				%mm0

#define S7_a1				tmp_at(1)
#define S7_a2				tmp_at(2)
#define S7_a4				tmp_at(3)
#define S7_a6				tmp_at(4)
#define S7_x6				tmp_at(5)
#define S7_x7				tmp_at(6)
#define S7_x8				tmp_at(7)
#define S7_x11				tmp_at(8)
#define S7_x13				tmp_at(9)
#define S7_x15				tmp_at(10)
#define S7_x25				tmp_at(11)
#define S7_x26				tmp_at(12)

#define S7(out1, out2, out3, out4, extra) \
	movq %mm0,S7_a1; \
	movq %mm1,%mm6; \
	extra; \
	movq %mm1,S7_a2; \
	movq %mm3,%mm7; \
	movq %mm5,S7_a6; \
	pand %mm3,%mm6; \
	movq %mm3,S7_a4; \
	pxor %mm4,%mm6; \
	pxor pnot,%mm4; \
	pand %mm6,%mm7; \
	pand %mm4,%mm3; \
	movq %mm1,%mm5; \
	pxor %mm2,%mm6; \
	pxor %mm7,%mm5; \
	movq %mm7,S7_x6; \
	por %mm1,%mm4; \
	por %mm3,%mm1; \
	pxor %mm6,%mm7; \
	movq %mm5,S7_x7; \
	pand %mm2,%mm4; \
	pand %mm2,%mm5; \
	por %mm7,%mm3; \
	movq %mm1,S7_x13; \
	pxor %mm5,%mm0; \
	por S7_a6,%mm0; \
	pxor %mm4,%mm1; \
	movq %mm4,S7_x15; \
	pxor %mm6,%mm0; \
	movq %mm5,S7_x8; \
	movq %mm3,%mm4; \
	movq S7_a6,%mm6; \
	movq %mm0,%mm5; \
	pxor S7_x6,%mm5; \
	por %mm6,%mm4; \
	movq %mm7,S7_x25; \
	por %mm6,%mm5; \
	movq S7_a1,%mm7; \
	pxor %mm1,%mm5; \
	movq %mm3,S7_x26; \
	pand %mm5,%mm7; \
	movq %mm0,S7_x11; \
	pxor %mm0,%mm7; \
	movq S7_a4,%mm3; \
	movq %mm7,%mm0; \
	por S7_a2,%mm0; \
	pand %mm3,%mm1; \
	pand S7_x13,%mm3; \
	por S7_x7,%mm2; \
	pxor S7_x6,%mm0; \
	pxor %mm3,%mm2; \
	movq S7_a2,%mm3; \
	movq %mm0,%mm6; \
	pxor pnot,%mm3; \
	pxor S7_x15,%mm6; \
	por %mm3,%mm1; \
	pand S7_x26,%mm0; \
	pxor %mm6,%mm4; \
	pand S7_a6,%mm0; \
	por %mm3,%mm6; \
	por S7_a6,%mm6; \
	pand %mm5,%mm3; \
	pand S7_a6,%mm1; \
	pxor %mm3,%mm0; \
	por S7_a1,%mm0; \
	pxor %mm6,%mm2; \
	pxor S7_x11,%mm1; \
	pxor %mm4,%mm0; \
	movq S7_a1,%mm4; \
	pxor %mm2,%mm5; \
	movq S7_a4,%mm6; \
	por %mm2,%mm4; \
	pxor S7_x25,%mm6; \
	pxor %mm4,%mm1; \
	movq S7_a6,%mm4; \
	pand %mm1,%mm6; \
	movq S7_x6,%mm3; \
	pand %mm4,%mm6; \
	pxor S7_x15,%mm3; \
	pxor %mm5,%mm6; \
	pxor S7_x8,%mm2; \
	por %mm4,%mm3; \
	por S7_a1,%mm6; \
	pxor %mm2,%mm3; \
	pxor out1,%mm7; \
	pxor %mm6,%mm3; \
	pxor out2,%mm1; \
	movq %mm7,out1; \
	pxor out3,%mm3; \
	movq %mm1,out2; \
	pxor out4,%mm0; \
	movq %mm3,out3; \
	movq %mm0,out4

#define S8_out1				%mm6
#define S8_out2				%mm2
#define S8_out3				%mm5
#define S8_out4				%mm1

#define S8_a1				tmp_at(1)
#define S8_a2				tmp_at(2)
#define S8_a4				tmp_at(3)
#define S8_a5				tmp_at(4)
#define S8_a6				tmp_at(5)
#define S8_x14				tmp_at(6)
#define S8_x22				tmp_at(7)
#define S8_x33				tmp_at(8)

#define S8(out1, out2, out3, out4, extra) \
	movq %mm0,S8_a1; \
	extra; \
	movq %mm2,%mm6; \
	pxor pnot,%mm0; \
	movq %mm2,%mm7; \
	movq %mm3,S8_a4; \
	por %mm0,%mm7; \
	pxor pnot,%mm3; \
	pxor %mm0,%mm6; \
	movq %mm5,S8_a6; \
	movq %mm4,%mm5; \
	movq %mm1,S8_a2; \
	movq %mm7,%mm1; \
	movq %mm4,S8_a5; \
	pxor %mm3,%mm7; \
	por %mm6,%mm5; \
	por %mm7,%mm0; \
	pand %mm4,%mm1; \
	pandn %mm0,%mm2; \
	por %mm7,%mm4; \
	pxor %mm1,%mm2; \
	movq %mm5,S8_x22; \
	pand %mm3,%mm5; \
	por S8_a2,%mm2; \
	pxor %mm4,%mm7; \
	pxor %mm0,%mm3; \
	movq %mm4,%mm1; \
	pxor S8_x22,%mm7; \
	pxor %mm3,%mm1; \
	pxor %mm6,%mm4; \
	pxor %mm5,%mm2; \
	pxor S8_a1,%mm5; \
	pand %mm3,%mm6; \
	movq %mm1,S8_x14; \
	pand %mm4,%mm5; \
	movq %mm7,S8_x33; \
	movq %mm0,%mm1; \
	pand S8_a5,%mm3; \
	movq %mm0,%mm7; \
	pand S8_a5,%mm1; \
	pxor %mm3,%mm7; \
	pand S8_a2,%mm7; \
	pxor %mm1,%mm6; \
	movq S8_a6,%mm1; \
	pxor %mm4,%mm7; \
	por S8_a2,%mm6; \
	pandn %mm0,%mm4; \
	pxor S8_x14,%mm6; \
	pand %mm2,%mm1; \
	pxor S8_a1,%mm3; \
	pxor %mm6,%mm2; \
	por S8_a6,%mm6; \
	pxor %mm7,%mm1; \
	pxor S8_x22,%mm3; \
	pxor %mm7,%mm6; \
	por S8_a2,%mm4; \
	pand S8_a2,%mm5; \
	pxor %mm4,%mm3; \
	movq S8_a1,%mm4; \
	pand S8_x33,%mm4; \
	por S8_a4,%mm7; \
	pxor %mm4,%mm0; \
	pand S8_a2,%mm7; \
	pxor %mm0,%mm5; \
	movq S8_a6,%mm4; \
	por %mm0,%mm2; \
	pxor S8_x33,%mm7; \
	por %mm4,%mm5; \
	pxor out1,%mm6; \
	pand %mm4,%mm2; \
	pxor out4,%mm1; \
	pxor %mm7,%mm5; \
	pxor %mm3,%mm2; \
	pxor out3,%mm5; \
	movq %mm6,out1; \
	pxor out2,%mm2; \
	movq %mm1,out4; \
	movq %mm5,out3; \
	movq %mm2,out2

#define zero				%mm0

#define DES_bs_clear_block_8(i) \
	movq zero,B(i); \
	movq zero,B(i + 1); \
	movq zero,B(i + 2); \
	movq zero,B(i + 3); \
	movq zero,B(i + 4); \
	movq zero,B(i + 5); \
	movq zero,B(i + 6); \
	movq zero,B(i + 7)

#define DES_bs_clear_block \
	DES_bs_clear_block_8(0); \
	DES_bs_clear_block_8(8); \
	DES_bs_clear_block_8(16); \
	DES_bs_clear_block_8(24); \
	DES_bs_clear_block_8(32); \
	DES_bs_clear_block_8(40); \
	DES_bs_clear_block_8(48); \
	DES_bs_clear_block_8(56)

#define k_ptr				%edx
#define K(i)				(i)*8(k_ptr)
#define k(i)				(i)*4(k_ptr)

#define a6_xor_ptr			%esi
#define a6_p				pxor (a6_xor_ptr),a6
#define a6_v(i)				pxor K(i),a6

#define tmp1				%ecx
#define tmp2				a6_xor_ptr

#define xor_E(i) \
	movl E(i),tmp1; \
	movq K(i),a1; \
	movl E(i + 1),tmp2; \
	movq K(i + 1),a2; \
	pxor (tmp1),a1; \
	pxor (tmp2),a2; \
	movl E(i + 2),tmp1; \
	movq K(i + 2),a3; \
	movl E(i + 3),tmp2; \
	movq K(i + 3),a4; \
	pxor (tmp1),a3; \
	pxor (tmp2),a4; \
	movl E(i + 4),tmp1; \
	movq K(i + 4),a5; \
	movl E(i + 5),a6_xor_ptr; \
	movq K(i + 5),a6; \
	pxor (tmp1),a5

#define xor_B(b1, k1, b2, k2, b3, k3, b4, k4, b5, k5, b6) \
	movq B(b1),a1; \
	movq B(b2),a2; \
	pxor K(k1),a1; \
	movq B(b3),a3; \
	pxor K(k2),a2; \
	movq B(b4),a4; \
	pxor K(k3),a3; \
	movq B(b5),a5; \
	pxor K(k4),a4; \
	movq B(b6),a6; \
	pxor K(k5),a5

#define xor_B_KS_p(b1, k1, b2, k2, b3, k3, b4, k4, b5, k5, b6, k6) \
	movl k(k1),tmp1; \
	movl k(k2),tmp2; \
	movq B(b1),a1; \
	movq B(b2),a2; \
	pxor (tmp1),a1; \
	movl k(k3),tmp1; \
	pxor (tmp2),a2; \
	movl k(k4),tmp2; \
	movq B(b3),a3; \
	movq B(b4),a4; \
	pxor (tmp1),a3; \
	movl k(k5),tmp1; \
	pxor (tmp2),a4; \
	movq B(b5),a5; \
	movl k(k6),a6_xor_ptr; \
	movq B(b6),a6; \
	pxor (tmp1),a5

.data

DO_ALIGN(3)
mm_ones:
.long -1
.long -1

.text

DO_ALIGN(5)
.globl DES_bs_init_asm
DES_bs_init_asm:
	movq mm_ones,%mm0
	movq %mm0,pnot
	ret

#define rounds_and_swapped		%ebp
#define iterations			%eax

DO_ALIGN(5)
.globl DES_bs_crypt
DES_bs_crypt:
	movl 4(%esp),iterations
	pxor zero,zero
	pushl %ebp
	pushl %esi
	movl $DES_bs_all_KS_v,k_ptr
	DES_bs_clear_block
	movl $8,rounds_and_swapped
DES_bs_crypt_start:
	xor_E(0)
	S1(B(40), B(48), B(54), B(62), a6_p)
	xor_E(6)
	S2(B(44), B(59), B(33), B(49), a6_p)
	xor_E(12)
	S3(B(55), B(47), B(61), B(37), a6_p)
	xor_E(18)
	S4(B(57), B(51), B(41), B(32), a6_p)
	xor_E(24)
	S5(B(39), B(45), B(56), B(34), a6_p)
	xor_E(30)
	S6(B(35), B(60), B(42), B(50), a6_p)
	xor_E(36)
	S7(B(63), B(43), B(53), B(38), a6_p)
	xor_E(42)
	S8(B(36), B(58), B(46), B(52), a6_p)
	cmpl $0x100,rounds_and_swapped
	je DES_bs_crypt_next
DES_bs_crypt_swap:
	xor_E(48)
	S1(B(8), B(16), B(22), B(30), a6_p)
	xor_E(54)
	S2(B(12), B(27), B(1), B(17), a6_p)
	xor_E(60)
	S3(B(23), B(15), B(29), B(5), a6_p)
	xor_E(66)
	S4(B(25), B(19), B(9), B(0), a6_p)
	xor_E(72)
	S5(B(7), B(13), B(24), B(2), a6_p)
	xor_E(78)
	S6(B(3), B(28), B(10), B(18), a6_p)
	xor_E(84)
	S7(B(31), B(11), B(21), B(6), a6_p)
	xor_E(90)
	addl $96*8,k_ptr
	S8(B(4), B(26), B(14), B(20), a6_p)
	decl rounds_and_swapped
	jnz DES_bs_crypt_start
	subl $0x300*8+48*8,k_ptr
	movl $0x108,rounds_and_swapped
	decl iterations
	jnz DES_bs_crypt_swap
	popl %esi
	popl %ebp
#ifdef EMMS
	emms
#endif
	ret
DES_bs_crypt_next:
	subl $0x300*8-48*8,k_ptr
	movl $8,rounds_and_swapped
	decl iterations
	jnz DES_bs_crypt_start
	popl %esi
	popl %ebp
#ifdef EMMS
	emms
#endif
	ret

DO_ALIGN(5)
.globl DES_bs_crypt_25
DES_bs_crypt_25:
	pxor zero,zero
	pushl %ebp
	pushl %esi
	movl $DES_bs_all_KS_v,k_ptr
	DES_bs_clear_block
	movl $8,rounds_and_swapped
	movl $25,iterations
DES_bs_crypt_25_start:
	xor_E(0)
	S1(B(40), B(48), B(54), B(62), a6_p)
	xor_E(6)
	S2(B(44), B(59), B(33), B(49), a6_p)
	xor_B(7, 12, 8, 13, 9, 14, 10, 15, 11, 16, 12)
	S3(B(55), B(47), B(61), B(37), a6_v(17))
	xor_B(11, 18, 12, 19, 13, 20, 14, 21, 15, 22, 16)
	S4(B(57), B(51), B(41), B(32), a6_v(23))
	xor_E(24)
	S5(B(39), B(45), B(56), B(34), a6_p)
	xor_E(30)
	S6(B(35), B(60), B(42), B(50), a6_p)
	xor_B(23, 36, 24, 37, 25, 38, 26, 39, 27, 40, 28)
	S7(B(63), B(43), B(53), B(38), a6_v(41))
	xor_B(27, 42, 28, 43, 29, 44, 30, 45, 31, 46, 0)
	S8(B(36), B(58), B(46), B(52), a6_v(47))
	cmpl $0x100,rounds_and_swapped
	je DES_bs_crypt_25_next
DES_bs_crypt_25_swap:
	xor_E(48)
	S1(B(8), B(16), B(22), B(30), a6_p)
	xor_E(54)
	S2(B(12), B(27), B(1), B(17), a6_p)
	xor_B(39, 60, 40, 61, 41, 62, 42, 63, 43, 64, 44)
	S3(B(23), B(15), B(29), B(5), a6_v(65))
	xor_B(43, 66, 44, 67, 45, 68, 46, 69, 47, 70, 48)
	S4(B(25), B(19), B(9), B(0), a6_v(71))
	xor_E(72)
	S5(B(7), B(13), B(24), B(2), a6_p)
	xor_E(78)
	S6(B(3), B(28), B(10), B(18), a6_p)
	xor_B(55, 84, 56, 85, 57, 86, 58, 87, 59, 88, 60)
	S7(B(31), B(11), B(21), B(6), a6_v(89))
	xor_B(59, 90, 60, 91, 61, 92, 62, 93, 63, 94, 32)
	S8(B(4), B(26), B(14), B(20), a6_v(95))
	addl $96*8,k_ptr
	decl rounds_and_swapped
	jnz DES_bs_crypt_25_start
	subl $0x300*8+48*8,k_ptr
	movl $0x108,rounds_and_swapped
	decl iterations
	jnz DES_bs_crypt_25_swap
	popl %esi
	popl %ebp
#ifdef EMMS
	emms
#endif
	ret
DES_bs_crypt_25_next:
	subl $0x300*8-48*8,k_ptr
	movl $8,rounds_and_swapped
	decl iterations
	jmp DES_bs_crypt_25_start

#define ones				%mm1

#define rounds				%eax

DO_ALIGN(5)
.globl DES_bs_crypt_LM
DES_bs_crypt_LM:
	pxor zero,zero
	pushl %esi
	movq pnot,ones
	movl $DES_bs_all_KS_p,k_ptr
	movq zero,B(0)
	movq zero,B(1)
	movq zero,B(2)
	movq zero,B(3)
	movq zero,B(4)
	movq zero,B(5)
	movq zero,B(6)
	movq zero,B(7)
	movq ones,B(8)
	movq ones,B(9)
	movq ones,B(10)
	movq zero,B(11)
	movq ones,B(12)
	movq zero,B(13)
	movq zero,B(14)
	movq zero,B(15)
	movq zero,B(16)
	movq zero,B(17)
	movq zero,B(18)
	movq zero,B(19)
	movq zero,B(20)
	movq zero,B(21)
	movq zero,B(22)
	movq ones,B(23)
	movq zero,B(24)
	movq zero,B(25)
	movq ones,B(26)
	movq zero,B(27)
	movq zero,B(28)
	movq ones,B(29)
	movq ones,B(30)
	movq ones,B(31)
	movq zero,B(32)
	movq zero,B(33)
	movq zero,B(34)
	movq ones,B(35)
	movq zero,B(36)
	movq ones,B(37)
	movq ones,B(38)
	movq ones,B(39)
	movq zero,B(40)
	movq zero,B(41)
	movq zero,B(42)
	movq zero,B(43)
	movq zero,B(44)
	movq ones,B(45)
	movq zero,B(46)
	movq zero,B(47)
	movq ones,B(48)
	movq ones,B(49)
	movq zero,B(50)
	movq zero,B(51)
	movq zero,B(52)
	movq zero,B(53)
	movq ones,B(54)
	movq zero,B(55)
	movq ones,B(56)
	movq zero,B(57)
	movq ones,B(58)
	movq zero,B(59)
	movq ones,B(60)
	movq ones,B(61)
	movq ones,B(62)
	movq ones,B(63)
	movl $8,rounds
DES_bs_crypt_LM_loop:
	xor_B_KS_p(31, 0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5)
	S1(B(40), B(48), B(54), B(62), a6_p)
	xor_B_KS_p(3, 6, 4, 7, 5, 8, 6, 9, 7, 10, 8, 11)
	S2(B(44), B(59), B(33), B(49), a6_p)
	xor_B_KS_p(7, 12, 8, 13, 9, 14, 10, 15, 11, 16, 12, 17)
	S3(B(55), B(47), B(61), B(37), a6_p)
	xor_B_KS_p(11, 18, 12, 19, 13, 20, 14, 21, 15, 22, 16, 23)
	S4(B(57), B(51), B(41), B(32), a6_p)
	xor_B_KS_p(15, 24, 16, 25, 17, 26, 18, 27, 19, 28, 20, 29)
	S5(B(39), B(45), B(56), B(34), a6_p)
	xor_B_KS_p(19, 30, 20, 31, 21, 32, 22, 33, 23, 34, 24, 35)
	S6(B(35), B(60), B(42), B(50), a6_p)
	xor_B_KS_p(23, 36, 24, 37, 25, 38, 26, 39, 27, 40, 28, 41)
	S7(B(63), B(43), B(53), B(38), a6_p)
	xor_B_KS_p(27, 42, 28, 43, 29, 44, 30, 45, 31, 46, 0, 47)
	S8(B(36), B(58), B(46), B(52), a6_p)
	xor_B_KS_p(63, 48, 32, 49, 33, 50, 34, 51, 35, 52, 36, 53)
	S1(B(8), B(16), B(22), B(30), a6_p)
	xor_B_KS_p(35, 54, 36, 55, 37, 56, 38, 57, 39, 58, 40, 59)
	S2(B(12), B(27), B(1), B(17), a6_p)
	xor_B_KS_p(39, 60, 40, 61, 41, 62, 42, 63, 43, 64, 44, 65)
	S3(B(23), B(15), B(29), B(5), a6_p)
	xor_B_KS_p(43, 66, 44, 67, 45, 68, 46, 69, 47, 70, 48, 71)
	S4(B(25), B(19), B(9), B(0), a6_p)
	xor_B_KS_p(47, 72, 48, 73, 49, 74, 50, 75, 51, 76, 52, 77)
	S5(B(7), B(13), B(24), B(2), a6_p)
	xor_B_KS_p(51, 78, 52, 79, 53, 80, 54, 81, 55, 82, 56, 83)
	S6(B(3), B(28), B(10), B(18), a6_p)
	xor_B_KS_p(55, 84, 56, 85, 57, 86, 58, 87, 59, 88, 60, 89)
	S7(B(31), B(11), B(21), B(6), a6_p)
	xor_B_KS_p(59, 90, 60, 91, 61, 92, 62, 93, 63, 94, 32, 95)
	addl $96*4,k_ptr
	S8(B(4), B(26), B(14), B(20), a6_p)
	decl rounds
	jnz DES_bs_crypt_LM_loop
	popl %esi
#ifdef EMMS
	emms
#endif
	ret