Context Navigation

source: git/src-cryptopp/salsa.cpp

Last change on this file was e230cb0, checked in by David Stainton <dstainton415@…>, at 2016-10-12T13:27:29Z
Add cryptopp from tag CRYPTOPP_5_6_5
Property mode set to `100644`
File size: 23.5 KB

Line
1	// salsa.cpp - written and placed in the public domain by Wei Dai
2
3	// use "cl /EP /P /DCRYPTOPP_GENERATE_X64_MASM salsa.cpp" to generate MASM code
4
5	#include "pch.h"
6	#include "config.h"
7
8	#ifndef CRYPTOPP_GENERATE_X64_MASM
9
10	#include "salsa.h"
11	#include "argnames.h"
12	#include "misc.h"
13	#include "cpu.h"
14
15	#if CRYPTOPP_MSC_VERSION
16	# pragma warning(disable: 4702 4740)
17	#endif
18
19	// TODO: work around GCC 4.8+ issue with SSE2 ASM until the exact details are known
20	// and fix is released. Duplicate with "valgrind ./cryptest.exe tv salsa"
21	// Clang due to "Inline assembly operands don't work with .intel_syntax"
22	// https://llvm.org/bugs/show_bug.cgi?id=24232
23	#if defined(CRYPTOPP_DISABLE_SALSA_ASM)
24	# undef CRYPTOPP_X86_ASM_AVAILABLE
25	# undef CRYPTOPP_X32_ASM_AVAILABLE
26	# undef CRYPTOPP_X64_ASM_AVAILABLE
27	# undef CRYPTOPP_BOOL_SSE2_ASM_AVAILABLE
28	# undef CRYPTOPP_BOOL_SSSE3_ASM_AVAILABLE
29	# define CRYPTOPP_BOOL_SSE2_ASM_AVAILABLE 0
30	# define CRYPTOPP_BOOL_SSSE3_ASM_AVAILABLE 0
31	#endif
32
33	NAMESPACE_BEGIN(CryptoPP)
34
35	#if CRYPTOPP_DEBUG && !defined(CRYPTOPP_DOXYGEN_PROCESSING)
36	void Salsa20_TestInstantiations()
37	{
38	Salsa20::Encryption x1;
39	XSalsa20::Encryption x2;
40	}
41	#endif
42
43	void Salsa20_Policy::CipherSetKey(const NameValuePairs &params, const byte *key, size_t length)
44	{
45	m_rounds = params.GetIntValueWithDefault(Name::Rounds(), 20);
46
47	if (!(m_rounds == 8 \|\| m_rounds == 12 \|\| m_rounds == 20))
48	throw InvalidRounds(Salsa20::StaticAlgorithmName(), m_rounds);
49
50	// m_state is reordered for SSE2
51	GetBlock<word32, LittleEndian> get1(key);
52	get1(m_state[13])(m_state[10])(m_state[7])(m_state[4]);
53	GetBlock<word32, LittleEndian> get2(key + length - 16);
54	get2(m_state[15])(m_state[12])(m_state[9])(m_state[6]);
55
56	// "expand 16-byte k" or "expand 32-byte k"
57	m_state[0] = 0x61707865;
58	m_state[1] = (length == 16) ? 0x3120646e : 0x3320646e;
59	m_state[2] = (length == 16) ? 0x79622d36 : 0x79622d32;
60	m_state[3] = 0x6b206574;
61	}
62
63	void Salsa20_Policy::CipherResynchronize(byte keystreamBuffer, const byte IV, size_t length)
64	{
65	CRYPTOPP_UNUSED(keystreamBuffer), CRYPTOPP_UNUSED(length);
66	CRYPTOPP_ASSERT(length==8);
67
68	GetBlock<word32, LittleEndian> get(IV);
69	get(m_state[14])(m_state[11]);
70	m_state[8] = m_state[5] = 0;
71	}
72
73	void Salsa20_Policy::SeekToIteration(lword iterationCount)
74	{
75	m_state[8] = (word32)iterationCount;
76	m_state[5] = (word32)SafeRightShift<32>(iterationCount);
77	}
78
79	#if (CRYPTOPP_BOOL_X86 \|\| CRYPTOPP_BOOL_X32 \|\| CRYPTOPP_BOOL_X64) && !defined(CRYPTOPP_DISABLE_SALSA_ASM)
80	unsigned int Salsa20_Policy::GetAlignment() const
81	{
82	#if CRYPTOPP_BOOL_SSE2_ASM_AVAILABLE
83	if (HasSSE2())
84	return 16;
85	else
86	#endif
87	return GetAlignmentOf<word32>();
88	}
89
90	unsigned int Salsa20_Policy::GetOptimalBlockSize() const
91	{
92	#if CRYPTOPP_BOOL_SSE2_ASM_AVAILABLE
93	if (HasSSE2())
94	return 4*BYTES_PER_ITERATION;
95	else
96	#endif
97	return BYTES_PER_ITERATION;
98	}
99	#endif
100
101	#ifdef CRYPTOPP_X64_MASM_AVAILABLE
102	extern "C" {
103	void Salsa20_OperateKeystream(byte output, const byte input, size_t iterationCount, int rounds, void *state);
104	}
105	#endif
106
107	#if CRYPTOPP_MSC_VERSION
108	# pragma warning(disable: 4731) // frame pointer register 'ebp' modified by inline assembly code
109	#endif
110
111	void Salsa20_Policy::OperateKeystream(KeystreamOperation operation, byte output, const byte input, size_t iterationCount)
112	{
113	#endif // #ifdef CRYPTOPP_GENERATE_X64_MASM
114
115	#ifdef CRYPTOPP_X64_MASM_AVAILABLE
116	Salsa20_OperateKeystream(output, input, iterationCount, m_rounds, m_state.data());
117	return;
118	#endif
119
120	#if CRYPTOPP_BOOL_SSE2_ASM_AVAILABLE
121	#ifdef CRYPTOPP_GENERATE_X64_MASM
122	ALIGN 8
123	Salsa20_OperateKeystream PROC FRAME
124	mov r10, [rsp + 5*8] ; state
125	alloc_stack(1016 + 3216 + 8)
126	save_xmm128 xmm6, 0200h
127	save_xmm128 xmm7, 0210h
128	save_xmm128 xmm8, 0220h
129	save_xmm128 xmm9, 0230h
130	save_xmm128 xmm10, 0240h
131	save_xmm128 xmm11, 0250h
132	save_xmm128 xmm12, 0260h
133	save_xmm128 xmm13, 0270h
134	save_xmm128 xmm14, 0280h
135	save_xmm128 xmm15, 0290h
136	.endprolog
137
138	#define REG_output rcx
139	#define REG_input rdx
140	#define REG_iterationCount r8
141	#define REG_state r10
142	#define REG_rounds e9d
143	#define REG_roundsLeft eax
144	#define REG_temp32 r11d
145	#define REG_temp r11
146	#define SSE2_WORKSPACE rsp
147	#else
148	if (HasSSE2())
149	{
150	#if CRYPTOPP_BOOL_X64
151	#define REG_output %1
152	#define REG_input %0
153	#define REG_iterationCount %2
154	#define REG_state %4 /* constant */
155	#define REG_rounds %3 /* constant */
156	#define REG_roundsLeft eax
157	#define REG_temp32 edx
158	#define REG_temp rdx
159	#define SSE2_WORKSPACE %5 /* constant */
160
161	CRYPTOPP_ALIGN_DATA(16) byte workspace[16*32];
162	#else
163	#define REG_output edi
164	#define REG_input eax
165	#define REG_iterationCount ecx
166	#define REG_state esi
167	#define REG_rounds edx
168	#define REG_roundsLeft ebx
169	#define REG_temp32 ebp
170	#define REG_temp ebp
171	#define SSE2_WORKSPACE esp + WORD_SZ
172	#endif
173
174	#ifdef __GNUC__
175	__asm__ __volatile__
176	(
177	INTEL_NOPREFIX
178	AS_PUSH_IF86( bx)
179	#else
180	void *s = m_state.data();
181	word32 r = m_rounds;
182
183	AS2( mov REG_iterationCount, iterationCount)
184	AS2( mov REG_input, input)
185	AS2( mov REG_output, output)
186	AS2( mov REG_state, s)
187	AS2( mov REG_rounds, r)
188	#endif
189	#endif // #ifndef CRYPTOPP_GENERATE_X64_MASM
190
191	AS_PUSH_IF86( bp)
192	AS2( cmp REG_iterationCount, 4)
193	ASJ( jl, 5, f)
194
195	#if CRYPTOPP_BOOL_X86
196	AS2( mov ebx, esp)
197	AS2( and esp, -16)
198	AS2( sub esp, 32*16)
199	AS1( push ebx)
200	#endif
201
202	#define SSE2_EXPAND_S(i, j) \
203	ASS( pshufd xmm4, xmm##i, j, j, j, j) \
204	AS2( movdqa [SSE2_WORKSPACE + (i4+j)16 + 256], xmm4)
205
206	AS2( movdqa xmm0, [REG_state + 0*16])
207	AS2( movdqa xmm1, [REG_state + 1*16])
208	AS2( movdqa xmm2, [REG_state + 2*16])
209	AS2( movdqa xmm3, [REG_state + 3*16])
210	SSE2_EXPAND_S(0, 0)
211	SSE2_EXPAND_S(0, 1)
212	SSE2_EXPAND_S(0, 2)
213	SSE2_EXPAND_S(0, 3)
214	SSE2_EXPAND_S(1, 0)
215	SSE2_EXPAND_S(1, 2)
216	SSE2_EXPAND_S(1, 3)
217	SSE2_EXPAND_S(2, 1)
218	SSE2_EXPAND_S(2, 2)
219	SSE2_EXPAND_S(2, 3)
220	SSE2_EXPAND_S(3, 0)
221	SSE2_EXPAND_S(3, 1)
222	SSE2_EXPAND_S(3, 2)
223	SSE2_EXPAND_S(3, 3)
224
225	#define SSE2_EXPAND_S85(i) \
226	AS2( mov dword ptr [SSE2_WORKSPACE + 816 + i4 + 256], REG_roundsLeft) \
227	AS2( mov dword ptr [SSE2_WORKSPACE + 516 + i4 + 256], REG_temp32) \
228	AS2( add REG_roundsLeft, 1) \
229	AS2( adc REG_temp32, 0)
230
231	ASL(1)
232	AS2( mov REG_roundsLeft, dword ptr [REG_state + 8*4])
233	AS2( mov REG_temp32, dword ptr [REG_state + 5*4])
234	SSE2_EXPAND_S85(0)
235	SSE2_EXPAND_S85(1)
236	SSE2_EXPAND_S85(2)
237	SSE2_EXPAND_S85(3)
238	AS2( mov dword ptr [REG_state + 8*4], REG_roundsLeft)
239	AS2( mov dword ptr [REG_state + 5*4], REG_temp32)
240
241	#define SSE2_QUARTER_ROUND(a, b, d, i) \
242	AS2( movdqa xmm4, xmm##d) \
243	AS2( paddd xmm4, xmm##a) \
244	AS2( movdqa xmm5, xmm4) \
245	AS2( pslld xmm4, i) \
246	AS2( psrld xmm5, 32-i) \
247	AS2( pxor xmm##b, xmm4) \
248	AS2( pxor xmm##b, xmm5)
249
250	#define L01(A,B,C,D,a,b,c,d,i) AS2( movdqa xmm##A, [SSE2_WORKSPACE + d16 + i256]) /* y3 */
251	#define L02(A,B,C,D,a,b,c,d,i) AS2( movdqa xmm##C, [SSE2_WORKSPACE + a16 + i256]) /* y0 */
252	#define L03(A,B,C,D,a,b,c,d,i) AS2( paddd xmm##A, xmm##C) /* y0+y3 */
253	#define L04(A,B,C,D,a,b,c,d,i) AS2( movdqa xmm##B, xmm##A)
254	#define L05(A,B,C,D,a,b,c,d,i) AS2( pslld xmm##A, 7)
255	#define L06(A,B,C,D,a,b,c,d,i) AS2( psrld xmm##B, 32-7)
256	#define L07(A,B,C,D,a,b,c,d,i) AS2( pxor xmm##A, [SSE2_WORKSPACE + b16 + i256])
257	#define L08(A,B,C,D,a,b,c,d,i) AS2( pxor xmm##A, xmm##B) /* z1 */
258	#define L09(A,B,C,D,a,b,c,d,i) AS2( movdqa [SSE2_WORKSPACE + b*16], xmm##A)
259	#define L10(A,B,C,D,a,b,c,d,i) AS2( movdqa xmm##B, xmm##A)
260	#define L11(A,B,C,D,a,b,c,d,i) AS2( paddd xmm##A, xmm##C) /* z1+y0 */
261	#define L12(A,B,C,D,a,b,c,d,i) AS2( movdqa xmm##D, xmm##A)
262	#define L13(A,B,C,D,a,b,c,d,i) AS2( pslld xmm##A, 9)
263	#define L14(A,B,C,D,a,b,c,d,i) AS2( psrld xmm##D, 32-9)
264	#define L15(A,B,C,D,a,b,c,d,i) AS2( pxor xmm##A, [SSE2_WORKSPACE + c16 + i256])
265	#define L16(A,B,C,D,a,b,c,d,i) AS2( pxor xmm##A, xmm##D) /* z2 */
266	#define L17(A,B,C,D,a,b,c,d,i) AS2( movdqa [SSE2_WORKSPACE + c*16], xmm##A)
267	#define L18(A,B,C,D,a,b,c,d,i) AS2( movdqa xmm##D, xmm##A)
268	#define L19(A,B,C,D,a,b,c,d,i) AS2( paddd xmm##A, xmm##B) /* z2+z1 */
269	#define L20(A,B,C,D,a,b,c,d,i) AS2( movdqa xmm##B, xmm##A)
270	#define L21(A,B,C,D,a,b,c,d,i) AS2( pslld xmm##A, 13)
271	#define L22(A,B,C,D,a,b,c,d,i) AS2( psrld xmm##B, 32-13)
272	#define L23(A,B,C,D,a,b,c,d,i) AS2( pxor xmm##A, [SSE2_WORKSPACE + d16 + i256])
273	#define L24(A,B,C,D,a,b,c,d,i) AS2( pxor xmm##A, xmm##B) /* z3 */
274	#define L25(A,B,C,D,a,b,c,d,i) AS2( movdqa [SSE2_WORKSPACE + d*16], xmm##A)
275	#define L26(A,B,C,D,a,b,c,d,i) AS2( paddd xmm##A, xmm##D) /* z3+z2 */
276	#define L27(A,B,C,D,a,b,c,d,i) AS2( movdqa xmm##D, xmm##A)
277	#define L28(A,B,C,D,a,b,c,d,i) AS2( pslld xmm##A, 18)
278	#define L29(A,B,C,D,a,b,c,d,i) AS2( psrld xmm##D, 32-18)
279	#define L30(A,B,C,D,a,b,c,d,i) AS2( pxor xmm##A, xmm##C) /* xor y0 */
280	#define L31(A,B,C,D,a,b,c,d,i) AS2( pxor xmm##A, xmm##D) /* z0 */
281	#define L32(A,B,C,D,a,b,c,d,i) AS2( movdqa [SSE2_WORKSPACE + a*16], xmm##A)
282
283	#define SSE2_QUARTER_ROUND_X8(i, a, b, c, d, e, f, g, h) \
284	L01(0,1,2,3, a,b,c,d, i) L01(4,5,6,7, e,f,g,h, i) \
285	L02(0,1,2,3, a,b,c,d, i) L02(4,5,6,7, e,f,g,h, i) \
286	L03(0,1,2,3, a,b,c,d, i) L03(4,5,6,7, e,f,g,h, i) \
287	L04(0,1,2,3, a,b,c,d, i) L04(4,5,6,7, e,f,g,h, i) \
288	L05(0,1,2,3, a,b,c,d, i) L05(4,5,6,7, e,f,g,h, i) \
289	L06(0,1,2,3, a,b,c,d, i) L06(4,5,6,7, e,f,g,h, i) \
290	L07(0,1,2,3, a,b,c,d, i) L07(4,5,6,7, e,f,g,h, i) \
291	L08(0,1,2,3, a,b,c,d, i) L08(4,5,6,7, e,f,g,h, i) \
292	L09(0,1,2,3, a,b,c,d, i) L09(4,5,6,7, e,f,g,h, i) \
293	L10(0,1,2,3, a,b,c,d, i) L10(4,5,6,7, e,f,g,h, i) \
294	L11(0,1,2,3, a,b,c,d, i) L11(4,5,6,7, e,f,g,h, i) \
295	L12(0,1,2,3, a,b,c,d, i) L12(4,5,6,7, e,f,g,h, i) \
296	L13(0,1,2,3, a,b,c,d, i) L13(4,5,6,7, e,f,g,h, i) \
297	L14(0,1,2,3, a,b,c,d, i) L14(4,5,6,7, e,f,g,h, i) \
298	L15(0,1,2,3, a,b,c,d, i) L15(4,5,6,7, e,f,g,h, i) \
299	L16(0,1,2,3, a,b,c,d, i) L16(4,5,6,7, e,f,g,h, i) \
300	L17(0,1,2,3, a,b,c,d, i) L17(4,5,6,7, e,f,g,h, i) \
301	L18(0,1,2,3, a,b,c,d, i) L18(4,5,6,7, e,f,g,h, i) \
302	L19(0,1,2,3, a,b,c,d, i) L19(4,5,6,7, e,f,g,h, i) \
303	L20(0,1,2,3, a,b,c,d, i) L20(4,5,6,7, e,f,g,h, i) \
304	L21(0,1,2,3, a,b,c,d, i) L21(4,5,6,7, e,f,g,h, i) \
305	L22(0,1,2,3, a,b,c,d, i) L22(4,5,6,7, e,f,g,h, i) \
306	L23(0,1,2,3, a,b,c,d, i) L23(4,5,6,7, e,f,g,h, i) \
307	L24(0,1,2,3, a,b,c,d, i) L24(4,5,6,7, e,f,g,h, i) \
308	L25(0,1,2,3, a,b,c,d, i) L25(4,5,6,7, e,f,g,h, i) \
309	L26(0,1,2,3, a,b,c,d, i) L26(4,5,6,7, e,f,g,h, i) \
310	L27(0,1,2,3, a,b,c,d, i) L27(4,5,6,7, e,f,g,h, i) \
311	L28(0,1,2,3, a,b,c,d, i) L28(4,5,6,7, e,f,g,h, i) \
312	L29(0,1,2,3, a,b,c,d, i) L29(4,5,6,7, e,f,g,h, i) \
313	L30(0,1,2,3, a,b,c,d, i) L30(4,5,6,7, e,f,g,h, i) \
314	L31(0,1,2,3, a,b,c,d, i) L31(4,5,6,7, e,f,g,h, i) \
315	L32(0,1,2,3, a,b,c,d, i) L32(4,5,6,7, e,f,g,h, i)
316
317	#define SSE2_QUARTER_ROUND_X16(i, a, b, c, d, e, f, g, h, A, B, C, D, E, F, G, H) \
318	L01(0,1,2,3, a,b,c,d, i) L01(4,5,6,7, e,f,g,h, i) L01(8,9,10,11, A,B,C,D, i) L01(12,13,14,15, E,F,G,H, i) \
319	L02(0,1,2,3, a,b,c,d, i) L02(4,5,6,7, e,f,g,h, i) L02(8,9,10,11, A,B,C,D, i) L02(12,13,14,15, E,F,G,H, i) \
320	L03(0,1,2,3, a,b,c,d, i) L03(4,5,6,7, e,f,g,h, i) L03(8,9,10,11, A,B,C,D, i) L03(12,13,14,15, E,F,G,H, i) \
321	L04(0,1,2,3, a,b,c,d, i) L04(4,5,6,7, e,f,g,h, i) L04(8,9,10,11, A,B,C,D, i) L04(12,13,14,15, E,F,G,H, i) \
322	L05(0,1,2,3, a,b,c,d, i) L05(4,5,6,7, e,f,g,h, i) L05(8,9,10,11, A,B,C,D, i) L05(12,13,14,15, E,F,G,H, i) \
323	L06(0,1,2,3, a,b,c,d, i) L06(4,5,6,7, e,f,g,h, i) L06(8,9,10,11, A,B,C,D, i) L06(12,13,14,15, E,F,G,H, i) \
324	L07(0,1,2,3, a,b,c,d, i) L07(4,5,6,7, e,f,g,h, i) L07(8,9,10,11, A,B,C,D, i) L07(12,13,14,15, E,F,G,H, i) \
325	L08(0,1,2,3, a,b,c,d, i) L08(4,5,6,7, e,f,g,h, i) L08(8,9,10,11, A,B,C,D, i) L08(12,13,14,15, E,F,G,H, i) \
326	L09(0,1,2,3, a,b,c,d, i) L09(4,5,6,7, e,f,g,h, i) L09(8,9,10,11, A,B,C,D, i) L09(12,13,14,15, E,F,G,H, i) \
327	L10(0,1,2,3, a,b,c,d, i) L10(4,5,6,7, e,f,g,h, i) L10(8,9,10,11, A,B,C,D, i) L10(12,13,14,15, E,F,G,H, i) \
328	L11(0,1,2,3, a,b,c,d, i) L11(4,5,6,7, e,f,g,h, i) L11(8,9,10,11, A,B,C,D, i) L11(12,13,14,15, E,F,G,H, i) \
329	L12(0,1,2,3, a,b,c,d, i) L12(4,5,6,7, e,f,g,h, i) L12(8,9,10,11, A,B,C,D, i) L12(12,13,14,15, E,F,G,H, i) \
330	L13(0,1,2,3, a,b,c,d, i) L13(4,5,6,7, e,f,g,h, i) L13(8,9,10,11, A,B,C,D, i) L13(12,13,14,15, E,F,G,H, i) \
331	L14(0,1,2,3, a,b,c,d, i) L14(4,5,6,7, e,f,g,h, i) L14(8,9,10,11, A,B,C,D, i) L14(12,13,14,15, E,F,G,H, i) \
332	L15(0,1,2,3, a,b,c,d, i) L15(4,5,6,7, e,f,g,h, i) L15(8,9,10,11, A,B,C,D, i) L15(12,13,14,15, E,F,G,H, i) \
333	L16(0,1,2,3, a,b,c,d, i) L16(4,5,6,7, e,f,g,h, i) L16(8,9,10,11, A,B,C,D, i) L16(12,13,14,15, E,F,G,H, i) \
334	L17(0,1,2,3, a,b,c,d, i) L17(4,5,6,7, e,f,g,h, i) L17(8,9,10,11, A,B,C,D, i) L17(12,13,14,15, E,F,G,H, i) \
335	L18(0,1,2,3, a,b,c,d, i) L18(4,5,6,7, e,f,g,h, i) L18(8,9,10,11, A,B,C,D, i) L18(12,13,14,15, E,F,G,H, i) \
336	L19(0,1,2,3, a,b,c,d, i) L19(4,5,6,7, e,f,g,h, i) L19(8,9,10,11, A,B,C,D, i) L19(12,13,14,15, E,F,G,H, i) \
337	L20(0,1,2,3, a,b,c,d, i) L20(4,5,6,7, e,f,g,h, i) L20(8,9,10,11, A,B,C,D, i) L20(12,13,14,15, E,F,G,H, i) \
338	L21(0,1,2,3, a,b,c,d, i) L21(4,5,6,7, e,f,g,h, i) L21(8,9,10,11, A,B,C,D, i) L21(12,13,14,15, E,F,G,H, i) \
339	L22(0,1,2,3, a,b,c,d, i) L22(4,5,6,7, e,f,g,h, i) L22(8,9,10,11, A,B,C,D, i) L22(12,13,14,15, E,F,G,H, i) \
340	L23(0,1,2,3, a,b,c,d, i) L23(4,5,6,7, e,f,g,h, i) L23(8,9,10,11, A,B,C,D, i) L23(12,13,14,15, E,F,G,H, i) \
341	L24(0,1,2,3, a,b,c,d, i) L24(4,5,6,7, e,f,g,h, i) L24(8,9,10,11, A,B,C,D, i) L24(12,13,14,15, E,F,G,H, i) \
342	L25(0,1,2,3, a,b,c,d, i) L25(4,5,6,7, e,f,g,h, i) L25(8,9,10,11, A,B,C,D, i) L25(12,13,14,15, E,F,G,H, i) \
343	L26(0,1,2,3, a,b,c,d, i) L26(4,5,6,7, e,f,g,h, i) L26(8,9,10,11, A,B,C,D, i) L26(12,13,14,15, E,F,G,H, i) \
344	L27(0,1,2,3, a,b,c,d, i) L27(4,5,6,7, e,f,g,h, i) L27(8,9,10,11, A,B,C,D, i) L27(12,13,14,15, E,F,G,H, i) \
345	L28(0,1,2,3, a,b,c,d, i) L28(4,5,6,7, e,f,g,h, i) L28(8,9,10,11, A,B,C,D, i) L28(12,13,14,15, E,F,G,H, i) \
346	L29(0,1,2,3, a,b,c,d, i) L29(4,5,6,7, e,f,g,h, i) L29(8,9,10,11, A,B,C,D, i) L29(12,13,14,15, E,F,G,H, i) \
347	L30(0,1,2,3, a,b,c,d, i) L30(4,5,6,7, e,f,g,h, i) L30(8,9,10,11, A,B,C,D, i) L30(12,13,14,15, E,F,G,H, i) \
348	L31(0,1,2,3, a,b,c,d, i) L31(4,5,6,7, e,f,g,h, i) L31(8,9,10,11, A,B,C,D, i) L31(12,13,14,15, E,F,G,H, i) \
349	L32(0,1,2,3, a,b,c,d, i) L32(4,5,6,7, e,f,g,h, i) L32(8,9,10,11, A,B,C,D, i) L32(12,13,14,15, E,F,G,H, i)
350
351	#if CRYPTOPP_BOOL_X64
352	SSE2_QUARTER_ROUND_X16(1, 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15)
353	#else
354	SSE2_QUARTER_ROUND_X8(1, 2, 6, 10, 14, 3, 7, 11, 15)
355	SSE2_QUARTER_ROUND_X8(1, 0, 4, 8, 12, 1, 5, 9, 13)
356	#endif
357	AS2( mov REG_roundsLeft, REG_rounds)
358	ASJ( jmp, 2, f)
359
360	ASL(SSE2_Salsa_Output)
361	AS2( movdqa xmm0, xmm4)
362	AS2( punpckldq xmm4, xmm5)
363	AS2( movdqa xmm1, xmm6)
364	AS2( punpckldq xmm6, xmm7)
365	AS2( movdqa xmm2, xmm4)
366	AS2( punpcklqdq xmm4, xmm6) // e
367	AS2( punpckhqdq xmm2, xmm6) // f
368	AS2( punpckhdq xmm0, xmm5)
369	AS2( punpckhdq xmm1, xmm7)
370	AS2( movdqa xmm6, xmm0)
371	AS2( punpcklqdq xmm0, xmm1) // g
372	AS2( punpckhqdq xmm6, xmm1) // h
373	AS_XMM_OUTPUT4(SSE2_Salsa_Output_A, REG_input, REG_output, 4, 2, 0, 6, 1, 0, 4, 8, 12, 1)
374	AS1( ret)
375
376	ASL(6)
377	#if CRYPTOPP_BOOL_X64
378	SSE2_QUARTER_ROUND_X16(0, 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15)
379	ASL(2)
380	SSE2_QUARTER_ROUND_X16(0, 0, 13, 10, 7, 1, 14, 11, 4, 2, 15, 8, 5, 3, 12, 9, 6)
381	#else
382	SSE2_QUARTER_ROUND_X8(0, 2, 6, 10, 14, 3, 7, 11, 15)
383	SSE2_QUARTER_ROUND_X8(0, 0, 4, 8, 12, 1, 5, 9, 13)
384	ASL(2)
385	SSE2_QUARTER_ROUND_X8(0, 2, 15, 8, 5, 3, 12, 9, 6)
386	SSE2_QUARTER_ROUND_X8(0, 0, 13, 10, 7, 1, 14, 11, 4)
387	#endif
388	AS2( sub REG_roundsLeft, 2)
389	ASJ( jnz, 6, b)
390
391	#define SSE2_OUTPUT_4(a, b, c, d) \
392	AS2( movdqa xmm4, [SSE2_WORKSPACE + a*16 + 256])\
393	AS2( paddd xmm4, [SSE2_WORKSPACE + a*16])\
394	AS2( movdqa xmm5, [SSE2_WORKSPACE + b*16 + 256])\
395	AS2( paddd xmm5, [SSE2_WORKSPACE + b*16])\
396	AS2( movdqa xmm6, [SSE2_WORKSPACE + c*16 + 256])\
397	AS2( paddd xmm6, [SSE2_WORKSPACE + c*16])\
398	AS2( movdqa xmm7, [SSE2_WORKSPACE + d*16 + 256])\
399	AS2( paddd xmm7, [SSE2_WORKSPACE + d*16])\
400	ASC( call, SSE2_Salsa_Output)
401
402	SSE2_OUTPUT_4(0, 13, 10, 7)
403	SSE2_OUTPUT_4(4, 1, 14, 11)
404	SSE2_OUTPUT_4(8, 5, 2, 15)
405	SSE2_OUTPUT_4(12, 9, 6, 3)
406	AS2( test REG_input, REG_input)
407	ASJ( jz, 9, f)
408	AS2( add REG_input, 12*16)
409	ASL(9)
410	AS2( add REG_output, 12*16)
411	AS2( sub REG_iterationCount, 4)
412	AS2( cmp REG_iterationCount, 4)
413	ASJ( jge, 1, b)
414	AS_POP_IF86( sp)
415
416	ASL(5)
417	AS2( sub REG_iterationCount, 1)
418	ASJ( jl, 4, f)
419	AS2( movdqa xmm0, [REG_state + 0*16])
420	AS2( movdqa xmm1, [REG_state + 1*16])
421	AS2( movdqa xmm2, [REG_state + 2*16])
422	AS2( movdqa xmm3, [REG_state + 3*16])
423	AS2( mov REG_roundsLeft, REG_rounds)
424
425	ASL(0)
426	SSE2_QUARTER_ROUND(0, 1, 3, 7)
427	SSE2_QUARTER_ROUND(1, 2, 0, 9)
428	SSE2_QUARTER_ROUND(2, 3, 1, 13)
429	SSE2_QUARTER_ROUND(3, 0, 2, 18)
430	ASS( pshufd xmm1, xmm1, 2, 1, 0, 3)
431	ASS( pshufd xmm2, xmm2, 1, 0, 3, 2)
432	ASS( pshufd xmm3, xmm3, 0, 3, 2, 1)
433	SSE2_QUARTER_ROUND(0, 3, 1, 7)
434	SSE2_QUARTER_ROUND(3, 2, 0, 9)
435	SSE2_QUARTER_ROUND(2, 1, 3, 13)
436	SSE2_QUARTER_ROUND(1, 0, 2, 18)
437	ASS( pshufd xmm1, xmm1, 0, 3, 2, 1)
438	ASS( pshufd xmm2, xmm2, 1, 0, 3, 2)
439	ASS( pshufd xmm3, xmm3, 2, 1, 0, 3)
440	AS2( sub REG_roundsLeft, 2)
441	ASJ( jnz, 0, b)
442
443	AS2( paddd xmm0, [REG_state + 0*16])
444	AS2( paddd xmm1, [REG_state + 1*16])
445	AS2( paddd xmm2, [REG_state + 2*16])
446	AS2( paddd xmm3, [REG_state + 3*16])
447
448	AS2( add dword ptr [REG_state + 8*4], 1)
449	AS2( adc dword ptr [REG_state + 5*4], 0)
450
451	AS2( pcmpeqb xmm6, xmm6) // all ones
452	AS2( psrlq xmm6, 32) // lo32 mask
453	ASS( pshufd xmm7, xmm6, 0, 1, 2, 3) // hi32 mask
454	AS2( movdqa xmm4, xmm0)
455	AS2( movdqa xmm5, xmm3)
456	AS2( pand xmm0, xmm7)
457	AS2( pand xmm4, xmm6)
458	AS2( pand xmm3, xmm6)
459	AS2( pand xmm5, xmm7)
460	AS2( por xmm4, xmm5) // 0,13,2,15
461	AS2( movdqa xmm5, xmm1)
462	AS2( pand xmm1, xmm7)
463	AS2( pand xmm5, xmm6)
464	AS2( por xmm0, xmm5) // 4,1,6,3
465	AS2( pand xmm6, xmm2)
466	AS2( pand xmm2, xmm7)
467	AS2( por xmm1, xmm6) // 8,5,10,7
468	AS2( por xmm2, xmm3) // 12,9,14,11
469
470	AS2( movdqa xmm5, xmm4)
471	AS2( movdqa xmm6, xmm0)
472	AS3( shufpd xmm4, xmm1, 2) // 0,13,10,7
473	AS3( shufpd xmm0, xmm2, 2) // 4,1,14,11
474	AS3( shufpd xmm1, xmm5, 2) // 8,5,2,15
475	AS3( shufpd xmm2, xmm6, 2) // 12,9,6,3
476
477	// output keystream
478	AS_XMM_OUTPUT4(SSE2_Salsa_Output_B, REG_input, REG_output, 4, 0, 1, 2, 3, 0, 1, 2, 3, 4)
479	ASJ( jmp, 5, b)
480	ASL(4)
481
482	AS_POP_IF86( bp)
483	#ifdef __GNUC__
484	AS_POP_IF86( bx)
485	ATT_PREFIX
486	#if CRYPTOPP_BOOL_X64
487	: "+r" (input), "+r" (output), "+r" (iterationCount)
488	: "r" (m_rounds), "r" (m_state.m_ptr), "r" (workspace)
489	: "%eax", "%rdx", "memory", "cc", "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7", "%xmm8", "%xmm9", "%xmm10", "%xmm11", "%xmm12", "%xmm13", "%xmm14", "%xmm15"
490	#else
491	: "+a" (input), "+D" (output), "+c" (iterationCount)
492	: "d" (m_rounds), "S" (m_state.m_ptr)
493	: "memory", "cc"
494	#endif
495	);
496	#endif
497	#ifdef CRYPTOPP_GENERATE_X64_MASM
498	movdqa xmm6, [rsp + 0200h]
499	movdqa xmm7, [rsp + 0210h]
500	movdqa xmm8, [rsp + 0220h]
501	movdqa xmm9, [rsp + 0230h]
502	movdqa xmm10, [rsp + 0240h]
503	movdqa xmm11, [rsp + 0250h]
504	movdqa xmm12, [rsp + 0260h]
505	movdqa xmm13, [rsp + 0270h]
506	movdqa xmm14, [rsp + 0280h]
507	movdqa xmm15, [rsp + 0290h]
508	add rsp, 1016 + 3216 + 8
509	ret
510	Salsa20_OperateKeystream ENDP
511	#else
512	}
513	else
514	#endif
515	#endif
516	#ifndef CRYPTOPP_GENERATE_X64_MASM
517	{
518	word32 x0, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14, x15;
519
520	while (iterationCount--)
521	{
522	x0 = m_state[0]; x1 = m_state[1]; x2 = m_state[2]; x3 = m_state[3];
523	x4 = m_state[4]; x5 = m_state[5]; x6 = m_state[6]; x7 = m_state[7];
524	x8 = m_state[8]; x9 = m_state[9]; x10 = m_state[10]; x11 = m_state[11];
525	x12 = m_state[12]; x13 = m_state[13]; x14 = m_state[14]; x15 = m_state[15];
526
527	for (int i=m_rounds; i>0; i-=2)
528	{
529	#define QUARTER_ROUND(a, b, c, d) \
530	b = b ^ rotlFixed(a + d, 7); \
531	c = c ^ rotlFixed(b + a, 9); \
532	d = d ^ rotlFixed(c + b, 13); \
533	a = a ^ rotlFixed(d + c, 18);
534
535	QUARTER_ROUND(x0, x4, x8, x12)
536	QUARTER_ROUND(x1, x5, x9, x13)
537	QUARTER_ROUND(x2, x6, x10, x14)
538	QUARTER_ROUND(x3, x7, x11, x15)
539
540	QUARTER_ROUND(x0, x13, x10, x7)
541	QUARTER_ROUND(x1, x14, x11, x4)
542	QUARTER_ROUND(x2, x15, x8, x5)
543	QUARTER_ROUND(x3, x12, x9, x6)
544	}
545
546	#define SALSA_OUTPUT(x) {\
547	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 0, x0 + m_state[0]);\
548	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 1, x13 + m_state[13]);\
549	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 2, x10 + m_state[10]);\
550	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 3, x7 + m_state[7]);\
551	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 4, x4 + m_state[4]);\
552	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 5, x1 + m_state[1]);\
553	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 6, x14 + m_state[14]);\
554	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 7, x11 + m_state[11]);\
555	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 8, x8 + m_state[8]);\
556	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 9, x5 + m_state[5]);\
557	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 10, x2 + m_state[2]);\
558	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 11, x15 + m_state[15]);\
559	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 12, x12 + m_state[12]);\
560	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 13, x9 + m_state[9]);\
561	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 14, x6 + m_state[6]);\
562	CRYPTOPP_KEYSTREAM_OUTPUT_WORD(x, LITTLE_ENDIAN_ORDER, 15, x3 + m_state[3]);}
563
564	#ifndef CRYPTOPP_DOXYGEN_PROCESSING
565	CRYPTOPP_KEYSTREAM_OUTPUT_SWITCH(SALSA_OUTPUT, BYTES_PER_ITERATION);
566	#endif
567
568	if (++m_state[8] == 0)
569	++m_state[5];
570	}
571	}
572	} // see comment above if an internal compiler error occurs here
573
574	void XSalsa20_Policy::CipherSetKey(const NameValuePairs &params, const byte *key, size_t length)
575	{
576	m_rounds = params.GetIntValueWithDefault(Name::Rounds(), 20);
577
578	if (!(m_rounds == 8 \|\| m_rounds == 12 \|\| m_rounds == 20))
579	throw InvalidRounds(XSalsa20::StaticAlgorithmName(), m_rounds);
580
581	GetUserKey(LITTLE_ENDIAN_ORDER, m_key.begin(), m_key.size(), key, length);
582	if (length == 16)
583	memcpy(m_key.begin()+4, m_key.begin(), 16);
584
585	// "expand 32-byte k"
586	m_state[0] = 0x61707865;
587	m_state[1] = 0x3320646e;
588	m_state[2] = 0x79622d32;
589	m_state[3] = 0x6b206574;
590	}
591
592	void XSalsa20_Policy::CipherResynchronize(byte keystreamBuffer, const byte IV, size_t length)
593	{
594	CRYPTOPP_UNUSED(keystreamBuffer), CRYPTOPP_UNUSED(length);
595	CRYPTOPP_ASSERT(length==24);
596
597	word32 x0, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14, x15;
598
599	GetBlock<word32, LittleEndian> get(IV);
600	get(x14)(x11)(x8)(x5)(m_state[14])(m_state[11]);
601
602	x13 = m_key[0]; x10 = m_key[1]; x7 = m_key[2]; x4 = m_key[3];
603	x15 = m_key[4]; x12 = m_key[5]; x9 = m_key[6]; x6 = m_key[7];
604	x0 = m_state[0]; x1 = m_state[1]; x2 = m_state[2]; x3 = m_state[3];
605
606	for (int i=m_rounds; i>0; i-=2)
607	{
608	QUARTER_ROUND(x0, x4, x8, x12)
609	QUARTER_ROUND(x1, x5, x9, x13)
610	QUARTER_ROUND(x2, x6, x10, x14)
611	QUARTER_ROUND(x3, x7, x11, x15)
612
613	QUARTER_ROUND(x0, x13, x10, x7)
614	QUARTER_ROUND(x1, x14, x11, x4)
615	QUARTER_ROUND(x2, x15, x8, x5)
616	QUARTER_ROUND(x3, x12, x9, x6)
617	}
618
619	m_state[13] = x0; m_state[10] = x1; m_state[7] = x2; m_state[4] = x3;
620	m_state[15] = x14; m_state[12] = x11; m_state[9] = x8; m_state[6] = x5;
621	m_state[8] = m_state[5] = 0;
622	}
623
624	NAMESPACE_END
625
626	#endif // #ifndef CRYPTOPP_GENERATE_X64_MASM

Note: See TracBrowser for help on using the repository browser.

Download in other formats: