1 // gcm_simd.cpp - written and placed in the public domain by
2 // Jeffrey Walton, Uri Blumenthal and Marcel Raad.
3 // Original x86 CLMUL by Wei Dai. ARM and POWER8
4 // PMULL and VMULL by JW, UB and MR.
5 //
6 // This source file uses intrinsics to gain access to SSE4.2 and
7 // ARMv8a CRC-32 and CRC-32C instructions. A separate source file
8 // is needed because additional CXXFLAGS are required to enable
9 // the appropriate instructions sets in some build configurations.
10
11 #include "pch.h"
12 #include "config.h"
13 #include "misc.h"
14
15 #if defined(CRYPTOPP_DISABLE_GCM_ASM)
16 # undef CRYPTOPP_X86_ASM_AVAILABLE
17 # undef CRYPTOPP_X32_ASM_AVAILABLE
18 # undef CRYPTOPP_X64_ASM_AVAILABLE
19 # undef CRYPTOPP_SSE2_ASM_AVAILABLE
20 #endif
21
22 #if (CRYPTOPP_SSE2_INTRIN_AVAILABLE)
23 # include <emmintrin.h>
24 # include <xmmintrin.h>
25 #endif
26
27 #if (CRYPTOPP_CLMUL_AVAILABLE)
28 # include <tmmintrin.h>
29 # include <wmmintrin.h>
30 #endif
31
32 #if (CRYPTOPP_ARM_NEON_HEADER)
33 # include <arm_neon.h>
34 #endif
35
36 #if (CRYPTOPP_ARM_ACLE_HEADER)
37 # include <stdint.h>
38 # include <arm_acle.h>
39 #endif
40
41 #if defined(CRYPTOPP_ARM_PMULL_AVAILABLE)
42 # include "arm_simd.h"
43 #endif
44
45 #if defined(CRYPTOPP_ALTIVEC_AVAILABLE)
46 # include "ppc_simd.h"
47 #endif
48
49 #ifdef CRYPTOPP_GNU_STYLE_INLINE_ASSEMBLY
50 # include <signal.h>
51 # include <setjmp.h>
52 #endif
53
54 #ifndef EXCEPTION_EXECUTE_HANDLER
55 # define EXCEPTION_EXECUTE_HANDLER 1
56 #endif
57
58 // Clang intrinsic casts, http://bugs.llvm.org/show_bug.cgi?id=20670
59 #define M128_CAST(x) ((__m128i *)(void *)(x))
60 #define CONST_M128_CAST(x) ((const __m128i *)(const void *)(x))
61 #define UINT64_CAST(x) ((uint64_t *)(void *)(x))
62 #define CONST_UINT64_CAST(x) ((const uint64_t *)(const void *)(x))
63
64 // Squash MS LNK4221 and libtool warnings
65 extern const char GCM_SIMD_FNAME[] = __FILE__;
66
NAMESPACE_BEGIN(CryptoPP)67 NAMESPACE_BEGIN(CryptoPP)
68
69 // ************************* Feature Probes ************************* //
70
71 #ifdef CRYPTOPP_GNU_STYLE_INLINE_ASSEMBLY
72 extern "C" {
73 typedef void (*SigHandler)(int);
74
75 static jmp_buf s_jmpSIGILL;
76 static void SigIllHandler(int)
77 {
78 longjmp(s_jmpSIGILL, 1);
79 }
80 }
81 #endif // Not CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY
82
83 #if (CRYPTOPP_BOOL_ARM32 || CRYPTOPP_BOOL_ARMV8)
CPU_ProbePMULL()84 bool CPU_ProbePMULL()
85 {
86 #if defined(CRYPTOPP_NO_CPU_FEATURE_PROBES)
87 return false;
88 #elif (CRYPTOPP_ARM_PMULL_AVAILABLE)
89 # if defined(CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY)
90 volatile bool result = true;
91 __try
92 {
93 // Linaro is missing a lot of pmull gear. Also see http://github.com/weidai11/cryptopp/issues/233.
94 const uint64_t wa1[]={0,0x9090909090909090}, wb1[]={0,0xb0b0b0b0b0b0b0b0};
95 const uint64x2_t a1=vld1q_u64(wa1), b1=vld1q_u64(wb1);
96
97 const uint8_t wa2[]={0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,
98 0xa0,0xa0,0xa0,0xa0,0xa0,0xa0,0xa0,0xa0},
99 wb2[]={0xc0,0xc0,0xc0,0xc0,0xc0,0xc0,0xc0,0xc0,
100 0xe0,0xe0,0xe0,0xe0,0xe0,0xe0,0xe0,0xe0};
101 const uint8x16_t a2=vld1q_u8(wa2), b2=vld1q_u8(wb2);
102
103 const uint64x2_t r1 = PMULL_00(a1, b1);
104 const uint64x2_t r2 = PMULL_11(vreinterpretq_u64_u8(a2),
105 vreinterpretq_u64_u8(b2));
106
107 result = !!(vgetq_lane_u64(r1,0) == 0x5300530053005300 &&
108 vgetq_lane_u64(r1,1) == 0x5300530053005300 &&
109 vgetq_lane_u64(r2,0) == 0x6c006c006c006c00 &&
110 vgetq_lane_u64(r2,1) == 0x6c006c006c006c00);
111 }
112 __except (EXCEPTION_EXECUTE_HANDLER)
113 {
114 return false;
115 }
116 return result;
117 # else
118
119 // longjmp and clobber warnings. Volatile is required.
120 volatile bool result = true;
121
122 volatile SigHandler oldHandler = signal(SIGILL, SigIllHandler);
123 if (oldHandler == SIG_ERR)
124 return false;
125
126 volatile sigset_t oldMask;
127 if (sigprocmask(0, NULLPTR, (sigset_t*)&oldMask))
128 {
129 signal(SIGILL, oldHandler);
130 return false;
131 }
132
133 if (setjmp(s_jmpSIGILL))
134 result = false;
135 else
136 {
137 // Linaro is missing a lot of pmull gear. Also see http://github.com/weidai11/cryptopp/issues/233.
138 const uint64_t wa1[]={0,0x9090909090909090}, wb1[]={0,0xb0b0b0b0b0b0b0b0};
139 const uint64x2_t a1=vld1q_u64(wa1), b1=vld1q_u64(wb1);
140
141 const uint8_t wa2[]={0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,
142 0xa0,0xa0,0xa0,0xa0,0xa0,0xa0,0xa0,0xa0},
143 wb2[]={0xc0,0xc0,0xc0,0xc0,0xc0,0xc0,0xc0,0xc0,
144 0xe0,0xe0,0xe0,0xe0,0xe0,0xe0,0xe0,0xe0};
145 const uint8x16_t a2=vld1q_u8(wa2), b2=vld1q_u8(wb2);
146
147 const uint64x2_t r1 = PMULL_00(a1, b1);
148 const uint64x2_t r2 = PMULL_11(vreinterpretq_u64_u8(a2),
149 vreinterpretq_u64_u8(b2));
150
151 result = !!(vgetq_lane_u64(r1,0) == 0x5300530053005300 &&
152 vgetq_lane_u64(r1,1) == 0x5300530053005300 &&
153 vgetq_lane_u64(r2,0) == 0x6c006c006c006c00 &&
154 vgetq_lane_u64(r2,1) == 0x6c006c006c006c00);
155 }
156
157 sigprocmask(SIG_SETMASK, (sigset_t*)&oldMask, NULLPTR);
158 signal(SIGILL, oldHandler);
159 return result;
160 # endif
161 #else
162 return false;
163 #endif // CRYPTOPP_ARM_PMULL_AVAILABLE
164 }
165 #endif // ARM32 or ARM64
166
167 #if (CRYPTOPP_BOOL_PPC32 || CRYPTOPP_BOOL_PPC64)
CPU_ProbePMULL()168 bool CPU_ProbePMULL()
169 {
170 #if defined(CRYPTOPP_NO_CPU_FEATURE_PROBES)
171 return false;
172 #elif (CRYPTOPP_POWER8_VMULL_AVAILABLE)
173 // longjmp and clobber warnings. Volatile is required.
174 volatile bool result = true;
175
176 volatile SigHandler oldHandler = signal(SIGILL, SigIllHandler);
177 if (oldHandler == SIG_ERR)
178 return false;
179
180 volatile sigset_t oldMask;
181 if (sigprocmask(0, NULLPTR, (sigset_t*)&oldMask))
182 {
183 signal(SIGILL, oldHandler);
184 return false;
185 }
186
187 if (setjmp(s_jmpSIGILL))
188 result = false;
189 else
190 {
191 const uint64_t wa1[]={0,W64LIT(0x9090909090909090)},
192 wb1[]={0,W64LIT(0xb0b0b0b0b0b0b0b0)};
193 const uint64x2_p a1=VecLoad(wa1), b1=VecLoad(wb1);
194
195 const uint8_t wa2[]={0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,
196 0xa0,0xa0,0xa0,0xa0,0xa0,0xa0,0xa0,0xa0},
197 wb2[]={0xc0,0xc0,0xc0,0xc0,0xc0,0xc0,0xc0,0xc0,
198 0xe0,0xe0,0xe0,0xe0,0xe0,0xe0,0xe0,0xe0};
199 const uint32x4_p a2=VecLoad(wa2), b2=VecLoad(wb2);
200
201 const uint64x2_p r1 = VecIntelMultiply11(a1, b1);
202 const uint64x2_p r2 = VecIntelMultiply11((uint64x2_p)a2, (uint64x2_p)b2);
203
204 const uint64_t wc1[]={W64LIT(0x5300530053005300), W64LIT(0x5300530053005300)},
205 wc2[]={W64LIT(0x6c006c006c006c00), W64LIT(0x6c006c006c006c00)};
206 const uint64x2_p c1=VecLoad(wc1), c2=VecLoad(wc2);
207
208 result = !!(VecEqual(r1, c1) && VecEqual(r2, c2));
209 }
210
211 sigprocmask(SIG_SETMASK, (sigset_t*)&oldMask, NULLPTR);
212 signal(SIGILL, oldHandler);
213 return result;
214 #else
215 return false;
216 #endif // CRYPTOPP_POWER8_VMULL_AVAILABLE
217 }
218 #endif // PPC32 or PPC64
219
220 // *************************** ARM NEON *************************** //
221
222 #if CRYPTOPP_ARM_NEON_AVAILABLE
GCM_Xor16_NEON(byte * a,const byte * b,const byte * c)223 void GCM_Xor16_NEON(byte *a, const byte *b, const byte *c)
224 {
225 vst1q_u8(a, veorq_u8(vld1q_u8(b), vld1q_u8(c)));
226 }
227 #endif // CRYPTOPP_ARM_NEON_AVAILABLE
228
229 #if CRYPTOPP_ARM_PMULL_AVAILABLE
230
231 // Swaps high and low 64-bit words
SwapWords(const uint64x2_t & data)232 inline uint64x2_t SwapWords(const uint64x2_t& data)
233 {
234 return (uint64x2_t)vcombine_u64(
235 vget_high_u64(data), vget_low_u64(data));
236 }
237
GCM_Reduce_PMULL(uint64x2_t c0,uint64x2_t c1,uint64x2_t c2,const uint64x2_t & r)238 uint64x2_t GCM_Reduce_PMULL(uint64x2_t c0, uint64x2_t c1, uint64x2_t c2, const uint64x2_t &r)
239 {
240 c1 = veorq_u64(c1, VEXT_U8<8>(vdupq_n_u64(0), c0));
241 c1 = veorq_u64(c1, PMULL_01(c0, r));
242 c0 = VEXT_U8<8>(c0, vdupq_n_u64(0));
243 c0 = vshlq_n_u64(veorq_u64(c0, c1), 1);
244 c0 = PMULL_00(c0, r);
245 c2 = veorq_u64(c2, c0);
246 c2 = veorq_u64(c2, VEXT_U8<8>(c1, vdupq_n_u64(0)));
247 c1 = vshrq_n_u64(vcombine_u64(vget_low_u64(c1), vget_low_u64(c2)), 63);
248 c2 = vshlq_n_u64(c2, 1);
249
250 return veorq_u64(c2, c1);
251 }
252
GCM_Multiply_PMULL(const uint64x2_t & x,const uint64x2_t & h,const uint64x2_t & r)253 uint64x2_t GCM_Multiply_PMULL(const uint64x2_t &x, const uint64x2_t &h, const uint64x2_t &r)
254 {
255 const uint64x2_t c0 = PMULL_00(x, h);
256 const uint64x2_t c1 = veorq_u64(PMULL_10(x, h), PMULL_01(x, h));
257 const uint64x2_t c2 = PMULL_11(x, h);
258
259 return GCM_Reduce_PMULL(c0, c1, c2, r);
260 }
261
GCM_SetKeyWithoutResync_PMULL(const byte * hashKey,byte * mulTable,unsigned int tableSize)262 void GCM_SetKeyWithoutResync_PMULL(const byte *hashKey, byte *mulTable, unsigned int tableSize)
263 {
264 const uint64x2_t r = {0xe100000000000000ull, 0xc200000000000000ull};
265 const uint64x2_t t = vreinterpretq_u64_u8(vrev64q_u8(vld1q_u8(hashKey)));
266 const uint64x2_t h0 = vextq_u64(t, t, 1);
267
268 uint64x2_t h = h0;
269 unsigned int i;
270 for (i=0; i<tableSize-32; i+=32)
271 {
272 const uint64x2_t h1 = GCM_Multiply_PMULL(h, h0, r);
273 vst1_u64(UINT64_CAST(mulTable+i), vget_low_u64(h));
274 vst1q_u64(UINT64_CAST(mulTable+i+16), h1);
275 vst1q_u64(UINT64_CAST(mulTable+i+8), h);
276 vst1_u64(UINT64_CAST(mulTable+i+8), vget_low_u64(h1));
277 h = GCM_Multiply_PMULL(h1, h0, r);
278 }
279
280 const uint64x2_t h1 = GCM_Multiply_PMULL(h, h0, r);
281 vst1_u64(UINT64_CAST(mulTable+i), vget_low_u64(h));
282 vst1q_u64(UINT64_CAST(mulTable+i+16), h1);
283 vst1q_u64(UINT64_CAST(mulTable+i+8), h);
284 vst1_u64(UINT64_CAST(mulTable+i+8), vget_low_u64(h1));
285 }
286
GCM_AuthenticateBlocks_PMULL(const byte * data,size_t len,const byte * mtable,byte * hbuffer)287 size_t GCM_AuthenticateBlocks_PMULL(const byte *data, size_t len, const byte *mtable, byte *hbuffer)
288 {
289 const uint64x2_t r = {0xe100000000000000ull, 0xc200000000000000ull};
290 uint64x2_t x = vreinterpretq_u64_u8(vld1q_u8(hbuffer));
291
292 while (len >= 16)
293 {
294 size_t i=0, s = UnsignedMin(len/16U, 8U);
295 uint64x2_t d1, d2 = vreinterpretq_u64_u8(vrev64q_u8(vld1q_u8(data+(s-1)*16U)));
296 uint64x2_t c0 = vdupq_n_u64(0);
297 uint64x2_t c1 = vdupq_n_u64(0);
298 uint64x2_t c2 = vdupq_n_u64(0);
299
300 while (true)
301 {
302 const uint64x2_t h0 = vld1q_u64(CONST_UINT64_CAST(mtable+(i+0)*16));
303 const uint64x2_t h1 = vld1q_u64(CONST_UINT64_CAST(mtable+(i+1)*16));
304 const uint64x2_t h2 = veorq_u64(h0, h1);
305
306 if (++i == s)
307 {
308 const uint64x2_t t1 = vreinterpretq_u64_u8(vrev64q_u8(vld1q_u8(data)));
309 d1 = veorq_u64(vextq_u64(t1, t1, 1), x);
310 c0 = veorq_u64(c0, PMULL_00(d1, h0));
311 c2 = veorq_u64(c2, PMULL_10(d1, h1));
312 d1 = veorq_u64(d1, SwapWords(d1));
313 c1 = veorq_u64(c1, PMULL_00(d1, h2));
314
315 break;
316 }
317
318 d1 = vreinterpretq_u64_u8(vrev64q_u8(vld1q_u8(data+(s-i)*16-8)));
319 c0 = veorq_u64(c0, PMULL_10(d2, h0));
320 c2 = veorq_u64(c2, PMULL_10(d1, h1));
321 d2 = veorq_u64(d2, d1);
322 c1 = veorq_u64(c1, PMULL_10(d2, h2));
323
324 if (++i == s)
325 {
326 const uint64x2_t t2 = vreinterpretq_u64_u8(vrev64q_u8(vld1q_u8(data)));
327 d1 = veorq_u64(vextq_u64(t2, t2, 1), x);
328 c0 = veorq_u64(c0, PMULL_01(d1, h0));
329 c2 = veorq_u64(c2, PMULL_11(d1, h1));
330 d1 = veorq_u64(d1, SwapWords(d1));
331 c1 = veorq_u64(c1, PMULL_01(d1, h2));
332
333 break;
334 }
335
336 const uint64x2_t t3 = vreinterpretq_u64_u8(vrev64q_u8(vld1q_u8(data+(s-i)*16-8)));
337 d2 = vextq_u64(t3, t3, 1);
338 c0 = veorq_u64(c0, PMULL_01(d1, h0));
339 c2 = veorq_u64(c2, PMULL_01(d2, h1));
340 d1 = veorq_u64(d1, d2);
341 c1 = veorq_u64(c1, PMULL_01(d1, h2));
342 }
343 data += s*16;
344 len -= s*16;
345
346 c1 = veorq_u64(veorq_u64(c1, c0), c2);
347 x = GCM_Reduce_PMULL(c0, c1, c2, r);
348 }
349
350 vst1q_u64(UINT64_CAST(hbuffer), x);
351 return len;
352 }
353
GCM_ReverseHashBufferIfNeeded_PMULL(byte * hashBuffer)354 void GCM_ReverseHashBufferIfNeeded_PMULL(byte *hashBuffer)
355 {
356 if (GetNativeByteOrder() != BIG_ENDIAN_ORDER)
357 {
358 const uint8x16_t x = vrev64q_u8(vld1q_u8(hashBuffer));
359 vst1q_u8(hashBuffer, vextq_u8(x, x, 8));
360 }
361 }
362 #endif // CRYPTOPP_ARM_PMULL_AVAILABLE
363
364 // ***************************** SSE ***************************** //
365
366 #if CRYPTOPP_SSE2_INTRIN_AVAILABLE || CRYPTOPP_SSE2_ASM_AVAILABLE
367 // SunCC 5.10-5.11 compiler crash. Move GCM_Xor16_SSE2 out-of-line, and place in
368 // a source file with a SSE architecture switch. Also see GH #226 and GH #284.
GCM_Xor16_SSE2(byte * a,const byte * b,const byte * c)369 void GCM_Xor16_SSE2(byte *a, const byte *b, const byte *c)
370 {
371 # if CRYPTOPP_SSE2_ASM_AVAILABLE && defined(__GNUC__)
372 asm ("movdqa %1, %%xmm0; pxor %2, %%xmm0; movdqa %%xmm0, %0;"
373 : "=m" (a[0]) : "m"(b[0]), "m"(c[0]));
374 # else // CRYPTOPP_SSE2_INTRIN_AVAILABLE
375 _mm_store_si128(M128_CAST(a), _mm_xor_si128(
376 _mm_load_si128(CONST_M128_CAST(b)),
377 _mm_load_si128(CONST_M128_CAST(c))));
378 # endif
379 }
380 #endif // CRYPTOPP_SSE2_ASM_AVAILABLE
381
382 #if CRYPTOPP_CLMUL_AVAILABLE
383
384 #if 0
385 // preserved for testing
386 void gcm_gf_mult(const unsigned char *a, const unsigned char *b, unsigned char *c)
387 {
388 word64 Z0=0, Z1=0, V0, V1;
389
390 typedef BlockGetAndPut<word64, BigEndian> Block;
391 Block::Get(a)(V0)(V1);
392
393 for (int i=0; i<16; i++)
394 {
395 for (int j=0x80; j!=0; j>>=1)
396 {
397 int x = b[i] & j;
398 Z0 ^= x ? V0 : 0;
399 Z1 ^= x ? V1 : 0;
400 x = (int)V1 & 1;
401 V1 = (V1>>1) | (V0<<63);
402 V0 = (V0>>1) ^ (x ? W64LIT(0xe1) << 56 : 0);
403 }
404 }
405 Block::Put(NULLPTR, c)(Z0)(Z1);
406 }
407
408 __m128i _mm_clmulepi64_si128(const __m128i &a, const __m128i &b, int i)
409 {
410 word64 A[1] = {ByteReverse(((word64*)&a)[i&1])};
411 word64 B[1] = {ByteReverse(((word64*)&b)[i>>4])};
412
413 PolynomialMod2 pa((byte *)A, 8);
414 PolynomialMod2 pb((byte *)B, 8);
415 PolynomialMod2 c = pa*pb;
416
417 __m128i output;
418 for (int i=0; i<16; i++)
419 ((byte *)&output)[i] = c.GetByte(i);
420 return output;
421 }
422 #endif // Testing
423
424 // Swaps high and low 64-bit words
SwapWords(const __m128i & val)425 inline __m128i SwapWords(const __m128i& val)
426 {
427 return _mm_shuffle_epi32(val, _MM_SHUFFLE(1, 0, 3, 2));
428 }
429
430 // SunCC 5.11-5.15 compiler crash. Make the function inline
431 // and parameters non-const. Also see GH #188 and GH #224.
GCM_Reduce_CLMUL(__m128i c0,__m128i c1,__m128i c2,const __m128i & r)432 inline __m128i GCM_Reduce_CLMUL(__m128i c0, __m128i c1, __m128i c2, const __m128i& r)
433 {
434 /*
435 The polynomial to be reduced is c0 * x^128 + c1 * x^64 + c2. c0t below refers to the most
436 significant half of c0 as a polynomial, which, due to GCM's bit reflection, are in the
437 rightmost bit positions, and the lowest byte addresses.
438
439 c1 ^= c0t * 0xc200000000000000
440 c2t ^= c0t
441 t = shift (c1t ^ c0b) left 1 bit
442 c2 ^= t * 0xe100000000000000
443 c2t ^= c1b
444 shift c2 left 1 bit and xor in lowest bit of c1t
445 */
446 c1 = _mm_xor_si128(c1, _mm_slli_si128(c0, 8));
447 c1 = _mm_xor_si128(c1, _mm_clmulepi64_si128(c0, r, 0x10));
448 c0 = _mm_xor_si128(c1, _mm_srli_si128(c0, 8));
449 c0 = _mm_slli_epi64(c0, 1);
450 c0 = _mm_clmulepi64_si128(c0, r, 0);
451 c2 = _mm_xor_si128(c2, c0);
452 c2 = _mm_xor_si128(c2, _mm_srli_si128(c1, 8));
453 c1 = _mm_unpacklo_epi64(c1, c2);
454 c1 = _mm_srli_epi64(c1, 63);
455 c2 = _mm_slli_epi64(c2, 1);
456 return _mm_xor_si128(c2, c1);
457 }
458
459 // SunCC 5.13-5.14 compiler crash. Don't make the function inline.
460 // This is in contrast to GCM_Reduce_CLMUL, which must be inline.
GCM_Multiply_CLMUL(const __m128i & x,const __m128i & h,const __m128i & r)461 __m128i GCM_Multiply_CLMUL(const __m128i &x, const __m128i &h, const __m128i &r)
462 {
463 const __m128i c0 = _mm_clmulepi64_si128(x,h,0);
464 const __m128i c1 = _mm_xor_si128(_mm_clmulepi64_si128(x,h,1), _mm_clmulepi64_si128(x,h,0x10));
465 const __m128i c2 = _mm_clmulepi64_si128(x,h,0x11);
466
467 return GCM_Reduce_CLMUL(c0, c1, c2, r);
468 }
469
GCM_SetKeyWithoutResync_CLMUL(const byte * hashKey,byte * mulTable,unsigned int tableSize)470 void GCM_SetKeyWithoutResync_CLMUL(const byte *hashKey, byte *mulTable, unsigned int tableSize)
471 {
472 const __m128i r = _mm_set_epi32(0xc2000000, 0x00000000, 0xe1000000, 0x00000000);
473 const __m128i m = _mm_set_epi32(0x00010203, 0x04050607, 0x08090a0b, 0x0c0d0e0f);
474 __m128i h0 = _mm_shuffle_epi8(_mm_load_si128(CONST_M128_CAST(hashKey)), m), h = h0;
475
476 unsigned int i;
477 for (i=0; i<tableSize-32; i+=32)
478 {
479 const __m128i h1 = GCM_Multiply_CLMUL(h, h0, r);
480 _mm_storel_epi64(M128_CAST(mulTable+i), h);
481 _mm_storeu_si128(M128_CAST(mulTable+i+16), h1);
482 _mm_storeu_si128(M128_CAST(mulTable+i+8), h);
483 _mm_storel_epi64(M128_CAST(mulTable+i+8), h1);
484 h = GCM_Multiply_CLMUL(h1, h0, r);
485 }
486
487 const __m128i h1 = GCM_Multiply_CLMUL(h, h0, r);
488 _mm_storel_epi64(M128_CAST(mulTable+i), h);
489 _mm_storeu_si128(M128_CAST(mulTable+i+16), h1);
490 _mm_storeu_si128(M128_CAST(mulTable+i+8), h);
491 _mm_storel_epi64(M128_CAST(mulTable+i+8), h1);
492 }
493
GCM_AuthenticateBlocks_CLMUL(const byte * data,size_t len,const byte * mtable,byte * hbuffer)494 size_t GCM_AuthenticateBlocks_CLMUL(const byte *data, size_t len, const byte *mtable, byte *hbuffer)
495 {
496 const __m128i r = _mm_set_epi32(0xc2000000, 0x00000000, 0xe1000000, 0x00000000);
497 const __m128i m1 = _mm_set_epi32(0x00010203, 0x04050607, 0x08090a0b, 0x0c0d0e0f);
498 const __m128i m2 = _mm_set_epi32(0x08090a0b, 0x0c0d0e0f, 0x00010203, 0x04050607);
499 __m128i x = _mm_load_si128(M128_CAST(hbuffer));
500
501 while (len >= 16)
502 {
503 size_t i=0, s = UnsignedMin(len/16, 8U);
504 __m128i d1 = _mm_loadu_si128(CONST_M128_CAST(data+(s-1)*16));
505 __m128i d2 = _mm_shuffle_epi8(d1, m2);
506 __m128i c0 = _mm_setzero_si128();
507 __m128i c1 = _mm_setzero_si128();
508 __m128i c2 = _mm_setzero_si128();
509
510 while (true)
511 {
512 const __m128i h0 = _mm_load_si128(CONST_M128_CAST(mtable+(i+0)*16));
513 const __m128i h1 = _mm_load_si128(CONST_M128_CAST(mtable+(i+1)*16));
514 const __m128i h2 = _mm_xor_si128(h0, h1);
515
516 if (++i == s)
517 {
518 d1 = _mm_shuffle_epi8(_mm_loadu_si128(CONST_M128_CAST(data)), m1);
519 d1 = _mm_xor_si128(d1, x);
520 c0 = _mm_xor_si128(c0, _mm_clmulepi64_si128(d1, h0, 0));
521 c2 = _mm_xor_si128(c2, _mm_clmulepi64_si128(d1, h1, 1));
522 d1 = _mm_xor_si128(d1, SwapWords(d1));
523 c1 = _mm_xor_si128(c1, _mm_clmulepi64_si128(d1, h2, 0));
524 break;
525 }
526
527 d1 = _mm_shuffle_epi8(_mm_loadu_si128(CONST_M128_CAST(data+(s-i)*16-8)), m2);
528 c0 = _mm_xor_si128(c0, _mm_clmulepi64_si128(d2, h0, 1));
529 c2 = _mm_xor_si128(c2, _mm_clmulepi64_si128(d1, h1, 1));
530 d2 = _mm_xor_si128(d2, d1);
531 c1 = _mm_xor_si128(c1, _mm_clmulepi64_si128(d2, h2, 1));
532
533 if (++i == s)
534 {
535 d1 = _mm_shuffle_epi8(_mm_loadu_si128(CONST_M128_CAST(data)), m1);
536 d1 = _mm_xor_si128(d1, x);
537 c0 = _mm_xor_si128(c0, _mm_clmulepi64_si128(d1, h0, 0x10));
538 c2 = _mm_xor_si128(c2, _mm_clmulepi64_si128(d1, h1, 0x11));
539 d1 = _mm_xor_si128(d1, SwapWords(d1));
540 c1 = _mm_xor_si128(c1, _mm_clmulepi64_si128(d1, h2, 0x10));
541 break;
542 }
543
544 d2 = _mm_shuffle_epi8(_mm_loadu_si128(CONST_M128_CAST(data+(s-i)*16-8)), m1);
545 c0 = _mm_xor_si128(c0, _mm_clmulepi64_si128(d1, h0, 0x10));
546 c2 = _mm_xor_si128(c2, _mm_clmulepi64_si128(d2, h1, 0x10));
547 d1 = _mm_xor_si128(d1, d2);
548 c1 = _mm_xor_si128(c1, _mm_clmulepi64_si128(d1, h2, 0x10));
549 }
550 data += s*16;
551 len -= s*16;
552
553 c1 = _mm_xor_si128(_mm_xor_si128(c1, c0), c2);
554 x = GCM_Reduce_CLMUL(c0, c1, c2, r);
555 }
556
557 _mm_store_si128(M128_CAST(hbuffer), x);
558 return len;
559 }
560
GCM_ReverseHashBufferIfNeeded_CLMUL(byte * hashBuffer)561 void GCM_ReverseHashBufferIfNeeded_CLMUL(byte *hashBuffer)
562 {
563 // SSSE3 instruction, but only used with CLMUL
564 const __m128i mask = _mm_set_epi32(0x00010203, 0x04050607, 0x08090a0b, 0x0c0d0e0f);
565 _mm_storeu_si128(M128_CAST(hashBuffer), _mm_shuffle_epi8(
566 _mm_loadu_si128(CONST_M128_CAST(hashBuffer)), mask));
567 }
568 #endif // CRYPTOPP_CLMUL_AVAILABLE
569
570 // ***************************** POWER8 ***************************** //
571
572 #if CRYPTOPP_POWER8_AVAILABLE
GCM_Xor16_POWER8(byte * a,const byte * b,const byte * c)573 void GCM_Xor16_POWER8(byte *a, const byte *b, const byte *c)
574 {
575 VecStore(VecXor(VecLoad(b), VecLoad(c)), a);
576 }
577 #endif // CRYPTOPP_POWER8_AVAILABLE
578
579 #if CRYPTOPP_POWER8_VMULL_AVAILABLE
580
GCM_Reduce_VMULL(uint64x2_p c0,uint64x2_p c1,uint64x2_p c2,uint64x2_p r)581 uint64x2_p GCM_Reduce_VMULL(uint64x2_p c0, uint64x2_p c1, uint64x2_p c2, uint64x2_p r)
582 {
583 const uint64x2_p m1 = {1,1}, m63 = {63,63};
584
585 c1 = VecXor(c1, VecShiftRightOctet<8>(c0));
586 c1 = VecXor(c1, VecIntelMultiply10(c0, r));
587 c0 = VecXor(c1, VecShiftLeftOctet<8>(c0));
588 c0 = VecIntelMultiply00(vec_sl(c0, m1), r);
589 c2 = VecXor(c2, c0);
590 c2 = VecXor(c2, VecShiftLeftOctet<8>(c1));
591 c1 = vec_sr(vec_mergeh(c1, c2), m63);
592 c2 = vec_sl(c2, m1);
593
594 return VecXor(c2, c1);
595 }
596
GCM_Multiply_VMULL(uint64x2_p x,uint64x2_p h,uint64x2_p r)597 inline uint64x2_p GCM_Multiply_VMULL(uint64x2_p x, uint64x2_p h, uint64x2_p r)
598 {
599 const uint64x2_p c0 = VecIntelMultiply00(x, h);
600 const uint64x2_p c1 = VecXor(VecIntelMultiply01(x, h), VecIntelMultiply10(x, h));
601 const uint64x2_p c2 = VecIntelMultiply11(x, h);
602
603 return GCM_Reduce_VMULL(c0, c1, c2, r);
604 }
605
LoadHashKey(const byte * hashKey)606 inline uint64x2_p LoadHashKey(const byte *hashKey)
607 {
608 #if (CRYPTOPP_BIG_ENDIAN)
609 const uint64x2_p key = (uint64x2_p)VecLoad(hashKey);
610 const uint8x16_p mask = {8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7};
611 return VecPermute(key, key, mask);
612 #else
613 const uint64x2_p key = (uint64x2_p)VecLoad(hashKey);
614 const uint8x16_p mask = {15,14,13,12, 11,10,9,8, 7,6,5,4, 3,2,1,0};
615 return VecPermute(key, key, mask);
616 #endif
617 }
618
GCM_SetKeyWithoutResync_VMULL(const byte * hashKey,byte * mulTable,unsigned int tableSize)619 void GCM_SetKeyWithoutResync_VMULL(const byte *hashKey, byte *mulTable, unsigned int tableSize)
620 {
621 const uint64x2_p r = {0xe100000000000000ull, 0xc200000000000000ull};
622 uint64x2_p h = LoadHashKey(hashKey), h0 = h;
623
624 unsigned int i;
625 uint64_t temp[2];
626
627 for (i=0; i<tableSize-32; i+=32)
628 {
629 const uint64x2_p h1 = GCM_Multiply_VMULL(h, h0, r);
630 VecStore(h, (byte*)temp);
631 std::memcpy(mulTable+i, temp+0, 8);
632 VecStore(h1, mulTable+i+16);
633 VecStore(h, mulTable+i+8);
634 VecStore(h1, (byte*)temp);
635 std::memcpy(mulTable+i+8, temp+0, 8);
636 h = GCM_Multiply_VMULL(h1, h0, r);
637 }
638
639 const uint64x2_p h1 = GCM_Multiply_VMULL(h, h0, r);
640 VecStore(h, (byte*)temp);
641 std::memcpy(mulTable+i, temp+0, 8);
642 VecStore(h1, mulTable+i+16);
643 VecStore(h, mulTable+i+8);
644 VecStore(h1, (byte*)temp);
645 std::memcpy(mulTable+i+8, temp+0, 8);
646 }
647
648 // Swaps high and low 64-bit words
649 template <class T>
SwapWords(const T & data)650 inline T SwapWords(const T& data)
651 {
652 return (T)VecRotateLeftOctet<8>(data);
653 }
654
LoadBuffer1(const byte * dataBuffer)655 inline uint64x2_p LoadBuffer1(const byte *dataBuffer)
656 {
657 #if (CRYPTOPP_BIG_ENDIAN)
658 return (uint64x2_p)VecLoad(dataBuffer);
659 #else
660 const uint64x2_p data = (uint64x2_p)VecLoad(dataBuffer);
661 const uint8x16_p mask = {7,6,5,4, 3,2,1,0, 15,14,13,12, 11,10,9,8};
662 return VecPermute(data, data, mask);
663 #endif
664 }
665
LoadBuffer2(const byte * dataBuffer)666 inline uint64x2_p LoadBuffer2(const byte *dataBuffer)
667 {
668 #if (CRYPTOPP_BIG_ENDIAN)
669 return (uint64x2_p)SwapWords(VecLoadBE(dataBuffer));
670 #else
671 return (uint64x2_p)VecLoadBE(dataBuffer);
672 #endif
673 }
674
GCM_AuthenticateBlocks_VMULL(const byte * data,size_t len,const byte * mtable,byte * hbuffer)675 size_t GCM_AuthenticateBlocks_VMULL(const byte *data, size_t len, const byte *mtable, byte *hbuffer)
676 {
677 const uint64x2_p r = {0xe100000000000000ull, 0xc200000000000000ull};
678 uint64x2_p x = (uint64x2_p)VecLoad(hbuffer);
679
680 while (len >= 16)
681 {
682 size_t i=0, s = UnsignedMin(len/16, 8U);
683 uint64x2_p d1, d2 = LoadBuffer1(data+(s-1)*16);
684 uint64x2_p c0 = {0}, c1 = {0}, c2 = {0};
685
686 while (true)
687 {
688 const uint64x2_p h0 = (uint64x2_p)VecLoad(mtable+(i+0)*16);
689 const uint64x2_p h1 = (uint64x2_p)VecLoad(mtable+(i+1)*16);
690 const uint64x2_p h2 = (uint64x2_p)VecXor(h0, h1);
691
692 if (++i == s)
693 {
694 d1 = LoadBuffer2(data);
695 d1 = VecXor(d1, x);
696 c0 = VecXor(c0, VecIntelMultiply00(d1, h0));
697 c2 = VecXor(c2, VecIntelMultiply01(d1, h1));
698 d1 = VecXor(d1, SwapWords(d1));
699 c1 = VecXor(c1, VecIntelMultiply00(d1, h2));
700 break;
701 }
702
703 d1 = LoadBuffer1(data+(s-i)*16-8);
704 c0 = VecXor(c0, VecIntelMultiply01(d2, h0));
705 c2 = VecXor(c2, VecIntelMultiply01(d1, h1));
706 d2 = VecXor(d2, d1);
707 c1 = VecXor(c1, VecIntelMultiply01(d2, h2));
708
709 if (++i == s)
710 {
711 d1 = LoadBuffer2(data);
712 d1 = VecXor(d1, x);
713 c0 = VecXor(c0, VecIntelMultiply10(d1, h0));
714 c2 = VecXor(c2, VecIntelMultiply11(d1, h1));
715 d1 = VecXor(d1, SwapWords(d1));
716 c1 = VecXor(c1, VecIntelMultiply10(d1, h2));
717 break;
718 }
719
720 d2 = LoadBuffer2(data+(s-i)*16-8);
721 c0 = VecXor(c0, VecIntelMultiply10(d1, h0));
722 c2 = VecXor(c2, VecIntelMultiply10(d2, h1));
723 d1 = VecXor(d1, d2);
724 c1 = VecXor(c1, VecIntelMultiply10(d1, h2));
725 }
726 data += s*16;
727 len -= s*16;
728
729 c1 = VecXor(VecXor(c1, c0), c2);
730 x = GCM_Reduce_VMULL(c0, c1, c2, r);
731 }
732
733 VecStore(x, hbuffer);
734 return len;
735 }
736
GCM_ReverseHashBufferIfNeeded_VMULL(byte * hashBuffer)737 void GCM_ReverseHashBufferIfNeeded_VMULL(byte *hashBuffer)
738 {
739 const uint64x2_p mask = {0x08090a0b0c0d0e0full, 0x0001020304050607ull};
740 VecStore(VecPermute(VecLoad(hashBuffer), mask), hashBuffer);
741 }
742 #endif // CRYPTOPP_POWER8_VMULL_AVAILABLE
743
744 NAMESPACE_END
745