162306a36Sopenharmony_ci/* SPDX-License-Identifier: GPL-2.0-or-later */
262306a36Sopenharmony_ci/*
362306a36Sopenharmony_ci * Twofish Cipher 3-way parallel algorithm (x86_64)
462306a36Sopenharmony_ci *
562306a36Sopenharmony_ci * Copyright (C) 2011 Jussi Kivilinna <jussi.kivilinna@mbnet.fi>
662306a36Sopenharmony_ci */
762306a36Sopenharmony_ci
862306a36Sopenharmony_ci#include <linux/linkage.h>
962306a36Sopenharmony_ci
1062306a36Sopenharmony_ci.file "twofish-x86_64-asm-3way.S"
1162306a36Sopenharmony_ci.text
1262306a36Sopenharmony_ci
1362306a36Sopenharmony_ci/* structure of crypto context */
1462306a36Sopenharmony_ci#define s0	0
1562306a36Sopenharmony_ci#define s1	1024
1662306a36Sopenharmony_ci#define s2	2048
1762306a36Sopenharmony_ci#define s3	3072
1862306a36Sopenharmony_ci#define w	4096
1962306a36Sopenharmony_ci#define k	4128
2062306a36Sopenharmony_ci
2162306a36Sopenharmony_ci/**********************************************************************
2262306a36Sopenharmony_ci  3-way twofish
2362306a36Sopenharmony_ci **********************************************************************/
2462306a36Sopenharmony_ci#define CTX %rdi
2562306a36Sopenharmony_ci#define RIO %rdx
2662306a36Sopenharmony_ci
2762306a36Sopenharmony_ci#define RAB0 %rax
2862306a36Sopenharmony_ci#define RAB1 %rbx
2962306a36Sopenharmony_ci#define RAB2 %rcx
3062306a36Sopenharmony_ci
3162306a36Sopenharmony_ci#define RAB0d %eax
3262306a36Sopenharmony_ci#define RAB1d %ebx
3362306a36Sopenharmony_ci#define RAB2d %ecx
3462306a36Sopenharmony_ci
3562306a36Sopenharmony_ci#define RAB0bh %ah
3662306a36Sopenharmony_ci#define RAB1bh %bh
3762306a36Sopenharmony_ci#define RAB2bh %ch
3862306a36Sopenharmony_ci
3962306a36Sopenharmony_ci#define RAB0bl %al
4062306a36Sopenharmony_ci#define RAB1bl %bl
4162306a36Sopenharmony_ci#define RAB2bl %cl
4262306a36Sopenharmony_ci
4362306a36Sopenharmony_ci#define CD0 0x0(%rsp)
4462306a36Sopenharmony_ci#define CD1 0x8(%rsp)
4562306a36Sopenharmony_ci#define CD2 0x10(%rsp)
4662306a36Sopenharmony_ci
4762306a36Sopenharmony_ci# used only before/after all rounds
4862306a36Sopenharmony_ci#define RCD0 %r8
4962306a36Sopenharmony_ci#define RCD1 %r9
5062306a36Sopenharmony_ci#define RCD2 %r10
5162306a36Sopenharmony_ci
5262306a36Sopenharmony_ci# used only during rounds
5362306a36Sopenharmony_ci#define RX0 %r8
5462306a36Sopenharmony_ci#define RX1 %r9
5562306a36Sopenharmony_ci#define RX2 %r10
5662306a36Sopenharmony_ci
5762306a36Sopenharmony_ci#define RX0d %r8d
5862306a36Sopenharmony_ci#define RX1d %r9d
5962306a36Sopenharmony_ci#define RX2d %r10d
6062306a36Sopenharmony_ci
6162306a36Sopenharmony_ci#define RY0 %r11
6262306a36Sopenharmony_ci#define RY1 %r12
6362306a36Sopenharmony_ci#define RY2 %r13
6462306a36Sopenharmony_ci
6562306a36Sopenharmony_ci#define RY0d %r11d
6662306a36Sopenharmony_ci#define RY1d %r12d
6762306a36Sopenharmony_ci#define RY2d %r13d
6862306a36Sopenharmony_ci
6962306a36Sopenharmony_ci#define RT0 %rdx
7062306a36Sopenharmony_ci#define RT1 %rsi
7162306a36Sopenharmony_ci
7262306a36Sopenharmony_ci#define RT0d %edx
7362306a36Sopenharmony_ci#define RT1d %esi
7462306a36Sopenharmony_ci
7562306a36Sopenharmony_ci#define RT1bl %sil
7662306a36Sopenharmony_ci
7762306a36Sopenharmony_ci#define do16bit_ror(rot, op1, op2, T0, T1, tmp1, tmp2, ab, dst) \
7862306a36Sopenharmony_ci	movzbl ab ## bl,		tmp2 ## d; \
7962306a36Sopenharmony_ci	movzbl ab ## bh,		tmp1 ## d; \
8062306a36Sopenharmony_ci	rorq $(rot),			ab; \
8162306a36Sopenharmony_ci	op1##l T0(CTX, tmp2, 4),	dst ## d; \
8262306a36Sopenharmony_ci	op2##l T1(CTX, tmp1, 4),	dst ## d;
8362306a36Sopenharmony_ci
8462306a36Sopenharmony_ci#define swap_ab_with_cd(ab, cd, tmp)	\
8562306a36Sopenharmony_ci	movq cd, tmp;			\
8662306a36Sopenharmony_ci	movq ab, cd;			\
8762306a36Sopenharmony_ci	movq tmp, ab;
8862306a36Sopenharmony_ci
8962306a36Sopenharmony_ci/*
9062306a36Sopenharmony_ci * Combined G1 & G2 function. Reordered with help of rotates to have moves
9162306a36Sopenharmony_ci * at beginning.
9262306a36Sopenharmony_ci */
9362306a36Sopenharmony_ci#define g1g2_3(ab, cd, Tx0, Tx1, Tx2, Tx3, Ty0, Ty1, Ty2, Ty3, x, y) \
9462306a36Sopenharmony_ci	/* G1,1 && G2,1 */ \
9562306a36Sopenharmony_ci	do16bit_ror(32, mov, xor, Tx0, Tx1, RT0, x ## 0, ab ## 0, x ## 0); \
9662306a36Sopenharmony_ci	do16bit_ror(48, mov, xor, Ty1, Ty2, RT0, y ## 0, ab ## 0, y ## 0); \
9762306a36Sopenharmony_ci	\
9862306a36Sopenharmony_ci	do16bit_ror(32, mov, xor, Tx0, Tx1, RT0, x ## 1, ab ## 1, x ## 1); \
9962306a36Sopenharmony_ci	do16bit_ror(48, mov, xor, Ty1, Ty2, RT0, y ## 1, ab ## 1, y ## 1); \
10062306a36Sopenharmony_ci	\
10162306a36Sopenharmony_ci	do16bit_ror(32, mov, xor, Tx0, Tx1, RT0, x ## 2, ab ## 2, x ## 2); \
10262306a36Sopenharmony_ci	do16bit_ror(48, mov, xor, Ty1, Ty2, RT0, y ## 2, ab ## 2, y ## 2); \
10362306a36Sopenharmony_ci	\
10462306a36Sopenharmony_ci	/* G1,2 && G2,2 */ \
10562306a36Sopenharmony_ci	do16bit_ror(32, xor, xor, Tx2, Tx3, RT0, RT1, ab ## 0, x ## 0); \
10662306a36Sopenharmony_ci	do16bit_ror(16, xor, xor, Ty3, Ty0, RT0, RT1, ab ## 0, y ## 0); \
10762306a36Sopenharmony_ci	swap_ab_with_cd(ab ## 0, cd ## 0, RT0); \
10862306a36Sopenharmony_ci	\
10962306a36Sopenharmony_ci	do16bit_ror(32, xor, xor, Tx2, Tx3, RT0, RT1, ab ## 1, x ## 1); \
11062306a36Sopenharmony_ci	do16bit_ror(16, xor, xor, Ty3, Ty0, RT0, RT1, ab ## 1, y ## 1); \
11162306a36Sopenharmony_ci	swap_ab_with_cd(ab ## 1, cd ## 1, RT0); \
11262306a36Sopenharmony_ci	\
11362306a36Sopenharmony_ci	do16bit_ror(32, xor, xor, Tx2, Tx3, RT0, RT1, ab ## 2, x ## 2); \
11462306a36Sopenharmony_ci	do16bit_ror(16, xor, xor, Ty3, Ty0, RT0, RT1, ab ## 2, y ## 2); \
11562306a36Sopenharmony_ci	swap_ab_with_cd(ab ## 2, cd ## 2, RT0);
11662306a36Sopenharmony_ci
11762306a36Sopenharmony_ci#define enc_round_end(ab, x, y, n) \
11862306a36Sopenharmony_ci	addl y ## d,			x ## d; \
11962306a36Sopenharmony_ci	addl x ## d,			y ## d; \
12062306a36Sopenharmony_ci	addl k+4*(2*(n))(CTX),		x ## d; \
12162306a36Sopenharmony_ci	xorl ab ## d,			x ## d; \
12262306a36Sopenharmony_ci	addl k+4*(2*(n)+1)(CTX),	y ## d; \
12362306a36Sopenharmony_ci	shrq $32,			ab; \
12462306a36Sopenharmony_ci	roll $1,			ab ## d; \
12562306a36Sopenharmony_ci	xorl y ## d,			ab ## d; \
12662306a36Sopenharmony_ci	shlq $32,			ab; \
12762306a36Sopenharmony_ci	rorl $1,			x ## d; \
12862306a36Sopenharmony_ci	orq x,				ab;
12962306a36Sopenharmony_ci
13062306a36Sopenharmony_ci#define dec_round_end(ba, x, y, n) \
13162306a36Sopenharmony_ci	addl y ## d,			x ## d; \
13262306a36Sopenharmony_ci	addl x ## d,			y ## d; \
13362306a36Sopenharmony_ci	addl k+4*(2*(n))(CTX),		x ## d; \
13462306a36Sopenharmony_ci	addl k+4*(2*(n)+1)(CTX),	y ## d; \
13562306a36Sopenharmony_ci	xorl ba ## d,			y ## d; \
13662306a36Sopenharmony_ci	shrq $32,			ba; \
13762306a36Sopenharmony_ci	roll $1,			ba ## d; \
13862306a36Sopenharmony_ci	xorl x ## d,			ba ## d; \
13962306a36Sopenharmony_ci	shlq $32,			ba; \
14062306a36Sopenharmony_ci	rorl $1,			y ## d; \
14162306a36Sopenharmony_ci	orq y,				ba;
14262306a36Sopenharmony_ci
14362306a36Sopenharmony_ci#define encrypt_round3(ab, cd, n) \
14462306a36Sopenharmony_ci	g1g2_3(ab, cd, s0, s1, s2, s3, s0, s1, s2, s3, RX, RY); \
14562306a36Sopenharmony_ci	\
14662306a36Sopenharmony_ci	enc_round_end(ab ## 0, RX0, RY0, n); \
14762306a36Sopenharmony_ci	enc_round_end(ab ## 1, RX1, RY1, n); \
14862306a36Sopenharmony_ci	enc_round_end(ab ## 2, RX2, RY2, n);
14962306a36Sopenharmony_ci
15062306a36Sopenharmony_ci#define decrypt_round3(ba, dc, n) \
15162306a36Sopenharmony_ci	g1g2_3(ba, dc, s1, s2, s3, s0, s3, s0, s1, s2, RY, RX); \
15262306a36Sopenharmony_ci	\
15362306a36Sopenharmony_ci	dec_round_end(ba ## 0, RX0, RY0, n); \
15462306a36Sopenharmony_ci	dec_round_end(ba ## 1, RX1, RY1, n); \
15562306a36Sopenharmony_ci	dec_round_end(ba ## 2, RX2, RY2, n);
15662306a36Sopenharmony_ci
15762306a36Sopenharmony_ci#define encrypt_cycle3(ab, cd, n) \
15862306a36Sopenharmony_ci	encrypt_round3(ab, cd, n*2); \
15962306a36Sopenharmony_ci	encrypt_round3(ab, cd, (n*2)+1);
16062306a36Sopenharmony_ci
16162306a36Sopenharmony_ci#define decrypt_cycle3(ba, dc, n) \
16262306a36Sopenharmony_ci	decrypt_round3(ba, dc, (n*2)+1); \
16362306a36Sopenharmony_ci	decrypt_round3(ba, dc, (n*2));
16462306a36Sopenharmony_ci
16562306a36Sopenharmony_ci#define push_cd()	\
16662306a36Sopenharmony_ci	pushq RCD2;	\
16762306a36Sopenharmony_ci	pushq RCD1;	\
16862306a36Sopenharmony_ci	pushq RCD0;
16962306a36Sopenharmony_ci
17062306a36Sopenharmony_ci#define pop_cd()	\
17162306a36Sopenharmony_ci	popq RCD0;	\
17262306a36Sopenharmony_ci	popq RCD1;	\
17362306a36Sopenharmony_ci	popq RCD2;
17462306a36Sopenharmony_ci
17562306a36Sopenharmony_ci#define inpack3(in, n, xy, m) \
17662306a36Sopenharmony_ci	movq 4*(n)(in),			xy ## 0; \
17762306a36Sopenharmony_ci	xorq w+4*m(CTX),		xy ## 0; \
17862306a36Sopenharmony_ci	\
17962306a36Sopenharmony_ci	movq 4*(4+(n))(in),		xy ## 1; \
18062306a36Sopenharmony_ci	xorq w+4*m(CTX),		xy ## 1; \
18162306a36Sopenharmony_ci	\
18262306a36Sopenharmony_ci	movq 4*(8+(n))(in),		xy ## 2; \
18362306a36Sopenharmony_ci	xorq w+4*m(CTX),		xy ## 2;
18462306a36Sopenharmony_ci
18562306a36Sopenharmony_ci#define outunpack3(op, out, n, xy, m) \
18662306a36Sopenharmony_ci	xorq w+4*m(CTX),		xy ## 0; \
18762306a36Sopenharmony_ci	op ## q xy ## 0,		4*(n)(out); \
18862306a36Sopenharmony_ci	\
18962306a36Sopenharmony_ci	xorq w+4*m(CTX),		xy ## 1; \
19062306a36Sopenharmony_ci	op ## q xy ## 1,		4*(4+(n))(out); \
19162306a36Sopenharmony_ci	\
19262306a36Sopenharmony_ci	xorq w+4*m(CTX),		xy ## 2; \
19362306a36Sopenharmony_ci	op ## q xy ## 2,		4*(8+(n))(out);
19462306a36Sopenharmony_ci
19562306a36Sopenharmony_ci#define inpack_enc3() \
19662306a36Sopenharmony_ci	inpack3(RIO, 0, RAB, 0); \
19762306a36Sopenharmony_ci	inpack3(RIO, 2, RCD, 2);
19862306a36Sopenharmony_ci
19962306a36Sopenharmony_ci#define outunpack_enc3(op) \
20062306a36Sopenharmony_ci	outunpack3(op, RIO, 2, RAB, 6); \
20162306a36Sopenharmony_ci	outunpack3(op, RIO, 0, RCD, 4);
20262306a36Sopenharmony_ci
20362306a36Sopenharmony_ci#define inpack_dec3() \
20462306a36Sopenharmony_ci	inpack3(RIO, 0, RAB, 4); \
20562306a36Sopenharmony_ci	rorq $32,			RAB0; \
20662306a36Sopenharmony_ci	rorq $32,			RAB1; \
20762306a36Sopenharmony_ci	rorq $32,			RAB2; \
20862306a36Sopenharmony_ci	inpack3(RIO, 2, RCD, 6); \
20962306a36Sopenharmony_ci	rorq $32,			RCD0; \
21062306a36Sopenharmony_ci	rorq $32,			RCD1; \
21162306a36Sopenharmony_ci	rorq $32,			RCD2;
21262306a36Sopenharmony_ci
21362306a36Sopenharmony_ci#define outunpack_dec3() \
21462306a36Sopenharmony_ci	rorq $32,			RCD0; \
21562306a36Sopenharmony_ci	rorq $32,			RCD1; \
21662306a36Sopenharmony_ci	rorq $32,			RCD2; \
21762306a36Sopenharmony_ci	outunpack3(mov, RIO, 0, RCD, 0); \
21862306a36Sopenharmony_ci	rorq $32,			RAB0; \
21962306a36Sopenharmony_ci	rorq $32,			RAB1; \
22062306a36Sopenharmony_ci	rorq $32,			RAB2; \
22162306a36Sopenharmony_ci	outunpack3(mov, RIO, 2, RAB, 2);
22262306a36Sopenharmony_ci
22362306a36Sopenharmony_ciSYM_FUNC_START(__twofish_enc_blk_3way)
22462306a36Sopenharmony_ci	/* input:
22562306a36Sopenharmony_ci	 *	%rdi: ctx, CTX
22662306a36Sopenharmony_ci	 *	%rsi: dst
22762306a36Sopenharmony_ci	 *	%rdx: src, RIO
22862306a36Sopenharmony_ci	 *	%rcx: bool, if true: xor output
22962306a36Sopenharmony_ci	 */
23062306a36Sopenharmony_ci	pushq %r13;
23162306a36Sopenharmony_ci	pushq %r12;
23262306a36Sopenharmony_ci	pushq %rbx;
23362306a36Sopenharmony_ci
23462306a36Sopenharmony_ci	pushq %rcx; /* bool xor */
23562306a36Sopenharmony_ci	pushq %rsi; /* dst */
23662306a36Sopenharmony_ci
23762306a36Sopenharmony_ci	inpack_enc3();
23862306a36Sopenharmony_ci
23962306a36Sopenharmony_ci	push_cd();
24062306a36Sopenharmony_ci	encrypt_cycle3(RAB, CD, 0);
24162306a36Sopenharmony_ci	encrypt_cycle3(RAB, CD, 1);
24262306a36Sopenharmony_ci	encrypt_cycle3(RAB, CD, 2);
24362306a36Sopenharmony_ci	encrypt_cycle3(RAB, CD, 3);
24462306a36Sopenharmony_ci	encrypt_cycle3(RAB, CD, 4);
24562306a36Sopenharmony_ci	encrypt_cycle3(RAB, CD, 5);
24662306a36Sopenharmony_ci	encrypt_cycle3(RAB, CD, 6);
24762306a36Sopenharmony_ci	encrypt_cycle3(RAB, CD, 7);
24862306a36Sopenharmony_ci	pop_cd();
24962306a36Sopenharmony_ci
25062306a36Sopenharmony_ci	popq RIO; /* dst */
25162306a36Sopenharmony_ci	popq RT1; /* bool xor */
25262306a36Sopenharmony_ci
25362306a36Sopenharmony_ci	testb RT1bl, RT1bl;
25462306a36Sopenharmony_ci	jnz .L__enc_xor3;
25562306a36Sopenharmony_ci
25662306a36Sopenharmony_ci	outunpack_enc3(mov);
25762306a36Sopenharmony_ci
25862306a36Sopenharmony_ci	popq %rbx;
25962306a36Sopenharmony_ci	popq %r12;
26062306a36Sopenharmony_ci	popq %r13;
26162306a36Sopenharmony_ci	RET;
26262306a36Sopenharmony_ci
26362306a36Sopenharmony_ci.L__enc_xor3:
26462306a36Sopenharmony_ci	outunpack_enc3(xor);
26562306a36Sopenharmony_ci
26662306a36Sopenharmony_ci	popq %rbx;
26762306a36Sopenharmony_ci	popq %r12;
26862306a36Sopenharmony_ci	popq %r13;
26962306a36Sopenharmony_ci	RET;
27062306a36Sopenharmony_ciSYM_FUNC_END(__twofish_enc_blk_3way)
27162306a36Sopenharmony_ci
27262306a36Sopenharmony_ciSYM_FUNC_START(twofish_dec_blk_3way)
27362306a36Sopenharmony_ci	/* input:
27462306a36Sopenharmony_ci	 *	%rdi: ctx, CTX
27562306a36Sopenharmony_ci	 *	%rsi: dst
27662306a36Sopenharmony_ci	 *	%rdx: src, RIO
27762306a36Sopenharmony_ci	 */
27862306a36Sopenharmony_ci	pushq %r13;
27962306a36Sopenharmony_ci	pushq %r12;
28062306a36Sopenharmony_ci	pushq %rbx;
28162306a36Sopenharmony_ci
28262306a36Sopenharmony_ci	pushq %rsi; /* dst */
28362306a36Sopenharmony_ci
28462306a36Sopenharmony_ci	inpack_dec3();
28562306a36Sopenharmony_ci
28662306a36Sopenharmony_ci	push_cd();
28762306a36Sopenharmony_ci	decrypt_cycle3(RAB, CD, 7);
28862306a36Sopenharmony_ci	decrypt_cycle3(RAB, CD, 6);
28962306a36Sopenharmony_ci	decrypt_cycle3(RAB, CD, 5);
29062306a36Sopenharmony_ci	decrypt_cycle3(RAB, CD, 4);
29162306a36Sopenharmony_ci	decrypt_cycle3(RAB, CD, 3);
29262306a36Sopenharmony_ci	decrypt_cycle3(RAB, CD, 2);
29362306a36Sopenharmony_ci	decrypt_cycle3(RAB, CD, 1);
29462306a36Sopenharmony_ci	decrypt_cycle3(RAB, CD, 0);
29562306a36Sopenharmony_ci	pop_cd();
29662306a36Sopenharmony_ci
29762306a36Sopenharmony_ci	popq RIO; /* dst */
29862306a36Sopenharmony_ci
29962306a36Sopenharmony_ci	outunpack_dec3();
30062306a36Sopenharmony_ci
30162306a36Sopenharmony_ci	popq %rbx;
30262306a36Sopenharmony_ci	popq %r12;
30362306a36Sopenharmony_ci	popq %r13;
30462306a36Sopenharmony_ci	RET;
30562306a36Sopenharmony_ciSYM_FUNC_END(twofish_dec_blk_3way)
306