aes-ppc.pl

OpenSSL 0.9.8k 最新版OpenSSL

	xor	$s3,$s3,$t3
	mtctr	$acc00
.align	4
Lenc_loop:
	rlwinm	$acc00,$s0,`32-24+3`,21,28
	rlwinm	$acc01,$s1,`32-24+3`,21,28
	lwz	$t0,0($key)
	lwz	$t1,4($key)
	rlwinm	$acc02,$s2,`32-24+3`,21,28
	rlwinm	$acc03,$s3,`32-24+3`,21,28
	lwz	$t2,8($key)
	lwz	$t3,12($key)
	rlwinm	$acc04,$s1,`32-16+3`,21,28
	rlwinm	$acc05,$s2,`32-16+3`,21,28
	lwzx	$acc00,$Tbl0,$acc00
	lwzx	$acc01,$Tbl0,$acc01
	rlwinm	$acc06,$s3,`32-16+3`,21,28
	rlwinm	$acc07,$s0,`32-16+3`,21,28
	lwzx	$acc02,$Tbl0,$acc02
	lwzx	$acc03,$Tbl0,$acc03
	rlwinm	$acc08,$s2,`32-8+3`,21,28
	rlwinm	$acc09,$s3,`32-8+3`,21,28
	lwzx	$acc04,$Tbl1,$acc04
	lwzx	$acc05,$Tbl1,$acc05
	rlwinm	$acc10,$s0,`32-8+3`,21,28
	rlwinm	$acc11,$s1,`32-8+3`,21,28
	lwzx	$acc06,$Tbl1,$acc06
	lwzx	$acc07,$Tbl1,$acc07
	rlwinm	$acc12,$s3,`0+3`,21,28
	rlwinm	$acc13,$s0,`0+3`,21,28
	lwzx	$acc08,$Tbl2,$acc08
	lwzx	$acc09,$Tbl2,$acc09
	rlwinm	$acc14,$s1,`0+3`,21,28
	rlwinm	$acc15,$s2,`0+3`,21,28
	lwzx	$acc10,$Tbl2,$acc10
	lwzx	$acc11,$Tbl2,$acc11
	xor	$t0,$t0,$acc00
	xor	$t1,$t1,$acc01
	lwzx	$acc12,$Tbl3,$acc12
	lwzx	$acc13,$Tbl3,$acc13
	xor	$t2,$t2,$acc02
	xor	$t3,$t3,$acc03
	lwzx	$acc14,$Tbl3,$acc14
	lwzx	$acc15,$Tbl3,$acc15
	xor	$t0,$t0,$acc04
	xor	$t1,$t1,$acc05
	xor	$t2,$t2,$acc06
	xor	$t3,$t3,$acc07
	xor	$t0,$t0,$acc08
	xor	$t1,$t1,$acc09
	xor	$t2,$t2,$acc10
	xor	$t3,$t3,$acc11
	xor	$s0,$t0,$acc12
	xor	$s1,$t1,$acc13
	xor	$s2,$t2,$acc14
	xor	$s3,$t3,$acc15
	addi	$key,$key,16
	bdnz-	Lenc_loop

	addi	$Tbl2,$Tbl0,2048
	nop
	lwz	$acc08,`2048+0`($Tbl0)	! prefetch Te4
	lwz	$acc09,`2048+32`($Tbl0)
	lwz	$acc10,`2048+64`($Tbl0)
	lwz	$acc11,`2048+96`($Tbl0)
	lwz	$acc08,`2048+128`($Tbl0)
	lwz	$acc09,`2048+160`($Tbl0)
	lwz	$acc10,`2048+192`($Tbl0)
	lwz	$acc11,`2048+224`($Tbl0)
	rlwinm	$acc00,$s0,`32-24`,24,31
	rlwinm	$acc01,$s1,`32-24`,24,31
	lwz	$t0,0($key)
	lwz	$t1,4($key)
	rlwinm	$acc02,$s2,`32-24`,24,31
	rlwinm	$acc03,$s3,`32-24`,24,31
	lwz	$t2,8($key)
	lwz	$t3,12($key)
	rlwinm	$acc04,$s1,`32-16`,24,31
	rlwinm	$acc05,$s2,`32-16`,24,31
	lbzx	$acc00,$Tbl2,$acc00
	lbzx	$acc01,$Tbl2,$acc01
	rlwinm	$acc06,$s3,`32-16`,24,31
	rlwinm	$acc07,$s0,`32-16`,24,31
	lbzx	$acc02,$Tbl2,$acc02
	lbzx	$acc03,$Tbl2,$acc03
	rlwinm	$acc08,$s2,`32-8`,24,31
	rlwinm	$acc09,$s3,`32-8`,24,31
	lbzx	$acc04,$Tbl2,$acc04
	lbzx	$acc05,$Tbl2,$acc05
	rlwinm	$acc10,$s0,`32-8`,24,31
	rlwinm	$acc11,$s1,`32-8`,24,31
	lbzx	$acc06,$Tbl2,$acc06
	lbzx	$acc07,$Tbl2,$acc07
	rlwinm	$acc12,$s3,`0`,24,31
	rlwinm	$acc13,$s0,`0`,24,31
	lbzx	$acc08,$Tbl2,$acc08
	lbzx	$acc09,$Tbl2,$acc09
	rlwinm	$acc14,$s1,`0`,24,31
	rlwinm	$acc15,$s2,`0`,24,31
	lbzx	$acc10,$Tbl2,$acc10
	lbzx	$acc11,$Tbl2,$acc11
	rlwinm	$s0,$acc00,24,0,7
	rlwinm	$s1,$acc01,24,0,7
	lbzx	$acc12,$Tbl2,$acc12
	lbzx	$acc13,$Tbl2,$acc13
	rlwinm	$s2,$acc02,24,0,7
	rlwinm	$s3,$acc03,24,0,7
	lbzx	$acc14,$Tbl2,$acc14
	lbzx	$acc15,$Tbl2,$acc15
	rlwimi	$s0,$acc04,16,8,15
	rlwimi	$s1,$acc05,16,8,15
	rlwimi	$s2,$acc06,16,8,15
	rlwimi	$s3,$acc07,16,8,15
	rlwimi	$s0,$acc08,8,16,23
	rlwimi	$s1,$acc09,8,16,23
	rlwimi	$s2,$acc10,8,16,23
	rlwimi	$s3,$acc11,8,16,23
	or	$s0,$s0,$acc12
	or	$s1,$s1,$acc13
	or	$s2,$s2,$acc14
	or	$s3,$s3,$acc15
	xor	$s0,$s0,$t0
	xor	$s1,$s1,$t1
	xor	$s2,$s2,$t2
	xor	$s3,$s3,$t3
	blr

.align	4
Lppc_AES_encrypt_compact:
	lwz	$acc00,240($key)
	lwz	$t0,0($key)
	lwz	$t1,4($key)
	lwz	$t2,8($key)
	lwz	$t3,12($key)
	addi	$Tbl1,$Tbl0,2048
	lis	$mask80,0x8080
	lis	$mask1b,0x1b1b
	addi	$key,$key,16
	ori	$mask80,$mask80,0x8080
	ori	$mask1b,$mask1b,0x1b1b
	mtctr	$acc00
.align	4
Lenc_compact_loop:
	xor	$s0,$s0,$t0
	xor	$s1,$s1,$t1
	xor	$s2,$s2,$t2
	xor	$s3,$s3,$t3
	rlwinm	$acc00,$s0,`32-24`,24,31
	rlwinm	$acc01,$s1,`32-24`,24,31
	rlwinm	$acc02,$s2,`32-24`,24,31
	rlwinm	$acc03,$s3,`32-24`,24,31
	lbzx	$acc00,$Tbl1,$acc00
	lbzx	$acc01,$Tbl1,$acc01
	rlwinm	$acc04,$s1,`32-16`,24,31
	rlwinm	$acc05,$s2,`32-16`,24,31
	lbzx	$acc02,$Tbl1,$acc02
	lbzx	$acc03,$Tbl1,$acc03
	rlwinm	$acc06,$s3,`32-16`,24,31
	rlwinm	$acc07,$s0,`32-16`,24,31
	lbzx	$acc04,$Tbl1,$acc04
	lbzx	$acc05,$Tbl1,$acc05
	rlwinm	$acc08,$s2,`32-8`,24,31
	rlwinm	$acc09,$s3,`32-8`,24,31
	lbzx	$acc06,$Tbl1,$acc06
	lbzx	$acc07,$Tbl1,$acc07
	rlwinm	$acc10,$s0,`32-8`,24,31
	rlwinm	$acc11,$s1,`32-8`,24,31
	lbzx	$acc08,$Tbl1,$acc08
	lbzx	$acc09,$Tbl1,$acc09
	rlwinm	$acc12,$s3,`0`,24,31
	rlwinm	$acc13,$s0,`0`,24,31
	lbzx	$acc10,$Tbl1,$acc10
	lbzx	$acc11,$Tbl1,$acc11
	rlwinm	$acc14,$s1,`0`,24,31
	rlwinm	$acc15,$s2,`0`,24,31
	lbzx	$acc12,$Tbl1,$acc12
	lbzx	$acc13,$Tbl1,$acc13
	rlwinm	$s0,$acc00,24,0,7
	rlwinm	$s1,$acc01,24,0,7
	lbzx	$acc14,$Tbl1,$acc14
	lbzx	$acc15,$Tbl1,$acc15
	rlwinm	$s2,$acc02,24,0,7
	rlwinm	$s3,$acc03,24,0,7
	rlwimi	$s0,$acc04,16,8,15
	rlwimi	$s1,$acc05,16,8,15
	rlwimi	$s2,$acc06,16,8,15
	rlwimi	$s3,$acc07,16,8,15
	rlwimi	$s0,$acc08,8,16,23
	rlwimi	$s1,$acc09,8,16,23
	rlwimi	$s2,$acc10,8,16,23
	rlwimi	$s3,$acc11,8,16,23
	lwz	$t0,0($key)
	lwz	$t1,4($key)
	or	$s0,$s0,$acc12
	or	$s1,$s1,$acc13
	lwz	$t2,8($key)
	lwz	$t3,12($key)
	or	$s2,$s2,$acc14
	or	$s3,$s3,$acc15

	addi	$key,$key,16
	bdz	Lenc_compact_done

	and	$acc00,$s0,$mask80	# r1=r0&0x80808080
	and	$acc01,$s1,$mask80
	and	$acc02,$s2,$mask80
	and	$acc03,$s3,$mask80
	srwi	$acc04,$acc00,7		# r1>>7
	srwi	$acc05,$acc01,7
	srwi	$acc06,$acc02,7
	srwi	$acc07,$acc03,7
	andc	$acc08,$s0,$mask80	# r0&0x7f7f7f7f
	andc	$acc09,$s1,$mask80
	andc	$acc10,$s2,$mask80
	andc	$acc11,$s3,$mask80
	sub	$acc00,$acc00,$acc04	# r1-(r1>>7)
	sub	$acc01,$acc01,$acc05
	sub	$acc02,$acc02,$acc06
	sub	$acc03,$acc03,$acc07
	add	$acc08,$acc08,$acc08	# (r0&0x7f7f7f7f)<<1
	add	$acc09,$acc09,$acc09
	add	$acc10,$acc10,$acc10
	add	$acc11,$acc11,$acc11
	and	$acc00,$acc00,$mask1b	# (r1-(r1>>7))&0x1b1b1b1b
	and	$acc01,$acc01,$mask1b
	and	$acc02,$acc02,$mask1b
	and	$acc03,$acc03,$mask1b
	xor	$acc00,$acc00,$acc08	# r2
	xor	$acc01,$acc01,$acc09
	xor	$acc02,$acc02,$acc10
	xor	$acc03,$acc03,$acc11

	rotlwi	$acc12,$s0,16		# ROTATE(r0,16)
	rotlwi	$acc13,$s1,16
	rotlwi	$acc14,$s2,16
	rotlwi	$acc15,$s3,16
	xor	$s0,$s0,$acc00		# r0^r2
	xor	$s1,$s1,$acc01
	xor	$s2,$s2,$acc02
	xor	$s3,$s3,$acc03
	rotrwi	$s0,$s0,24		# ROTATE(r2^r0,24)
	rotrwi	$s1,$s1,24
	rotrwi	$s2,$s2,24
	rotrwi	$s3,$s3,24
	xor	$s0,$s0,$acc00		# ROTATE(r2^r0,24)^r2
	xor	$s1,$s1,$acc01
	xor	$s2,$s2,$acc02
	xor	$s3,$s3,$acc03
	rotlwi	$acc08,$acc12,8		# ROTATE(r0,24)
	rotlwi	$acc09,$acc13,8
	rotlwi	$acc10,$acc14,8
	rotlwi	$acc11,$acc15,8
	xor	$s0,$s0,$acc12		#
	xor	$s1,$s1,$acc13
	xor	$s2,$s2,$acc14
	xor	$s3,$s3,$acc15
	xor	$s0,$s0,$acc08		#
	xor	$s1,$s1,$acc09
	xor	$s2,$s2,$acc10
	xor	$s3,$s3,$acc11

	b	Lenc_compact_loop
.align	4
Lenc_compact_done:
	xor	$s0,$s0,$t0
	xor	$s1,$s1,$t1
	xor	$s2,$s2,$t2
	xor	$s3,$s3,$t3
	blr

.globl	.AES_decrypt
.align	7
.AES_decrypt:
	mflr	r0
	$STU	$sp,-$FRAME($sp)

	$PUSH	r0,`$FRAME-$SIZE_T*21`($sp)
	$PUSH	$toc,`$FRAME-$SIZE_T*20`($sp)
	$PUSH	r13,`$FRAME-$SIZE_T*19`($sp)
	$PUSH	r14,`$FRAME-$SIZE_T*18`($sp)
	$PUSH	r15,`$FRAME-$SIZE_T*17`($sp)
	$PUSH	r16,`$FRAME-$SIZE_T*16`($sp)
	$PUSH	r17,`$FRAME-$SIZE_T*15`($sp)
	$PUSH	r18,`$FRAME-$SIZE_T*14`($sp)
	$PUSH	r19,`$FRAME-$SIZE_T*13`($sp)
	$PUSH	r20,`$FRAME-$SIZE_T*12`($sp)
	$PUSH	r21,`$FRAME-$SIZE_T*11`($sp)
	$PUSH	r22,`$FRAME-$SIZE_T*10`($sp)
	$PUSH	r23,`$FRAME-$SIZE_T*9`($sp)
	$PUSH	r24,`$FRAME-$SIZE_T*8`($sp)
	$PUSH	r25,`$FRAME-$SIZE_T*7`($sp)
	$PUSH	r26,`$FRAME-$SIZE_T*6`($sp)
	$PUSH	r27,`$FRAME-$SIZE_T*5`($sp)
	$PUSH	r28,`$FRAME-$SIZE_T*4`($sp)
	$PUSH	r29,`$FRAME-$SIZE_T*3`($sp)
	$PUSH	r30,`$FRAME-$SIZE_T*2`($sp)
	$PUSH	r31,`$FRAME-$SIZE_T*1`($sp)

	lwz	$s0,0($inp)
	lwz	$s1,4($inp)
	lwz	$s2,8($inp)
	lwz	$s3,12($inp)
	bl	LAES_Td
	bl	Lppc_AES_decrypt_compact
	stw	$s0,0($out)
	stw	$s1,4($out)
	stw	$s2,8($out)
	stw	$s3,12($out)

	$POP	r0,`$FRAME-$SIZE_T*21`($sp)
	$POP	$toc,`$FRAME-$SIZE_T*20`($sp)
	$POP	r13,`$FRAME-$SIZE_T*19`($sp)
	$POP	r14,`$FRAME-$SIZE_T*18`($sp)
	$POP	r15,`$FRAME-$SIZE_T*17`($sp)
	$POP	r16,`$FRAME-$SIZE_T*16`($sp)
	$POP	r17,`$FRAME-$SIZE_T*15`($sp)
	$POP	r18,`$FRAME-$SIZE_T*14`($sp)
	$POP	r19,`$FRAME-$SIZE_T*13`($sp)
	$POP	r20,`$FRAME-$SIZE_T*12`($sp)
	$POP	r21,`$FRAME-$SIZE_T*11`($sp)
	$POP	r22,`$FRAME-$SIZE_T*10`($sp)
	$POP	r23,`$FRAME-$SIZE_T*9`($sp)
	$POP	r24,`$FRAME-$SIZE_T*8`($sp)
	$POP	r25,`$FRAME-$SIZE_T*7`($sp)
	$POP	r26,`$FRAME-$SIZE_T*6`($sp)
	$POP	r27,`$FRAME-$SIZE_T*5`($sp)
	$POP	r28,`$FRAME-$SIZE_T*4`($sp)
	$POP	r29,`$FRAME-$SIZE_T*3`($sp)
	$POP	r30,`$FRAME-$SIZE_T*2`($sp)
	$POP	r31,`$FRAME-$SIZE_T*1`($sp)
	mtlr	r0
	addi	$sp,$sp,$FRAME
	blr

.align	4
Lppc_AES_decrypt:
	lwz	$acc00,240($key)
	lwz	$t0,0($key)
	lwz	$t1,4($key)
	lwz	$t2,8($key)
	lwz	$t3,12($key)
	addi	$Tbl1,$Tbl0,3
	addi	$Tbl2,$Tbl0,2
	addi	$Tbl3,$Tbl0,1
	addi	$acc00,$acc00,-1
	addi	$key,$key,16
	xor	$s0,$s0,$t0
	xor	$s1,$s1,$t1
	xor	$s2,$s2,$t2
	xor	$s3,$s3,$t3
	mtctr	$acc00
.align	4
Ldec_loop:
	rlwinm	$acc00,$s0,`32-24+3`,21,28
	rlwinm	$acc01,$s1,`32-24+3`,21,28
	lwz	$t0,0($key)
	lwz	$t1,4($key)
	rlwinm	$acc02,$s2,`32-24+3`,21,28
	rlwinm	$acc03,$s3,`32-24+3`,21,28
	lwz	$t2,8($key)
	lwz	$t3,12($key)
	rlwinm	$acc04,$s3,`32-16+3`,21,28
	rlwinm	$acc05,$s0,`32-16+3`,21,28
	lwzx	$acc00,$Tbl0,$acc00
	lwzx	$acc01,$Tbl0,$acc01
	rlwinm	$acc06,$s1,`32-16+3`,21,28
	rlwinm	$acc07,$s2,`32-16+3`,21,28
	lwzx	$acc02,$Tbl0,$acc02
	lwzx	$acc03,$Tbl0,$acc03
	rlwinm	$acc08,$s2,`32-8+3`,21,28
	rlwinm	$acc09,$s3,`32-8+3`,21,28
	lwzx	$acc04,$Tbl1,$acc04
	lwzx	$acc05,$Tbl1,$acc05
	rlwinm	$acc10,$s0,`32-8+3`,21,28
	rlwinm	$acc11,$s1,`32-8+3`,21,28
	lwzx	$acc06,$Tbl1,$acc06
	lwzx	$acc07,$Tbl1,$acc07
	rlwinm	$acc12,$s1,`0+3`,21,28
	rlwinm	$acc13,$s2,`0+3`,21,28
	lwzx	$acc08,$Tbl2,$acc08
	lwzx	$acc09,$Tbl2,$acc09
	rlwinm	$acc14,$s3,`0+3`,21,28
	rlwinm	$acc15,$s0,`0+3`,21,28
	lwzx	$acc10,$Tbl2,$acc10
	lwzx	$acc11,$Tbl2,$acc11
	xor	$t0,$t0,$acc00
	xor	$t1,$t1,$acc01
	lwzx	$acc12,$Tbl3,$acc12
	lwzx	$acc13,$Tbl3,$acc13
	xor	$t2,$t2,$acc02
	xor	$t3,$t3,$acc03
	lwzx	$acc14,$Tbl3,$acc14
	lwzx	$acc15,$Tbl3,$acc15