1 /*
2 * Copyright 2016-2024 The OpenSSL Project Authors. All Rights Reserved.
3 *
4 * Licensed under the Apache License 2.0 (the "License"). You may not use
5 * this file except in compliance with the License. You can obtain a copy
6 * in the file LICENSE in the source distribution or at
7 * https://www.openssl.org/source/license.html
8 */
9
10 /*
11 * This module is meant to be used as template for non-x87 floating-
12 * point assembly modules. The template itself is x86_64-specific
13 * though, as it was debugged on x86_64. So that implementer would
14 * have to recognize platform-specific parts, UxTOy and inline asm,
15 * and act accordingly.
16 *
17 * Huh? x86_64-specific code as template for non-x87? Note seven, which
18 * is not a typo, but reference to 80-bit precision. This module on the
19 * other hand relies on 64-bit precision operations, which are default
20 * for x86_64 code. And since we are at it, just for sense of it,
21 * large-block performance in cycles per processed byte for *this* code
22 * is:
23 * gcc-4.8 icc-15.0 clang-3.4(*)
24 *
25 * Westmere 4.96 5.09 4.37
26 * Sandy Bridge 4.95 4.90 4.17
27 * Haswell 4.92 4.87 3.78
28 * Bulldozer 4.67 4.49 4.68
29 * VIA Nano 7.07 7.05 5.98
30 * Silvermont 10.6 9.61 12.6
31 *
32 * (*) clang managed to discover parallelism and deployed SIMD;
33 *
34 * And for range of other platforms with unspecified gcc versions:
35 *
36 * Freescale e300 12.5
37 * PPC74x0 10.8
38 * POWER6 4.92
39 * POWER7 4.50
40 * POWER8 4.10
41 *
42 * z10 11.2
43 * z196+ 7.30
44 *
45 * UltraSPARC III 16.0
46 * SPARC T4 16.1
47 */
48
49 #if !(defined(__GNUC__) && __GNUC__>=2)
50 # error "this is gcc-specific template"
51 #endif
52
53 #include <stdlib.h>
54
55 typedef unsigned char u8;
56 typedef unsigned int u32;
57 typedef unsigned long long u64;
58 typedef union { double d; u64 u; } elem64;
59
60 #define TWO(p) ((double)(1ULL<<(p)))
61 #define TWO0 TWO(0)
62 #define TWO32 TWO(32)
63 #define TWO64 (TWO32*TWO(32))
64 #define TWO96 (TWO64*TWO(32))
65 #define TWO130 (TWO96*TWO(34))
66
67 #define EXP(p) ((1023ULL+(p))<<52)
68
69 #if defined(__x86_64__) || (defined(__PPC__) && defined(__LITTLE_ENDIAN__))
70 # define U8TOU32(p) (*(const u32 *)(p))
71 # define U32TO8(p,v) (*(u32 *)(p) = (v))
72 #elif defined(__PPC__) || defined(__POWERPC__)
73 # define U8TOU32(p) ({u32 ret; asm ("lwbrx %0,0,%1":"=r"(ret):"b"(p)); ret; })
74 # define U32TO8(p,v) asm ("stwbrx %0,0,%1"::"r"(v),"b"(p):"memory")
75 #elif defined(__s390x__)
76 # define U8TOU32(p) ({u32 ret; asm ("lrv %0,%1":"=d"(ret):"m"(*(u32 *)(p))); ret; })
77 # define U32TO8(p,v) asm ("strv %1,%0":"=m"(*(u32 *)(p)):"d"(v))
78 #endif
79
80 #ifndef U8TOU32
81 # define U8TOU32(p) ((u32)(p)[0] | (u32)(p)[1]<<8 | \
82 (u32)(p)[2]<<16 | (u32)(p)[3]<<24 )
83 #endif
84 #ifndef U32TO8
85 # define U32TO8(p,v) ((p)[0] = (u8)(v), (p)[1] = (u8)((v)>>8), \
86 (p)[2] = (u8)((v)>>16), (p)[3] = (u8)((v)>>24) )
87 #endif
88
89 typedef struct {
90 elem64 h[4];
91 double r[8];
92 double s[6];
93 } poly1305_internal;
94
95 /* "round toward zero (truncate), mask all exceptions" */
96 #if defined(__x86_64__)
97 static const u32 mxcsr = 0x7f80;
98 #elif defined(__PPC__) || defined(__POWERPC__)
99 static const u64 one = 1;
100 #elif defined(__s390x__)
101 static const u32 fpc = 1;
102 #elif defined(__sparc__)
103 static const u64 fsr = 1ULL<<30;
104 #elif defined(__mips__)
105 static const u32 fcsr = 1;
106 #else
107 #error "unrecognized platform"
108 #endif
109
poly1305_init(void * ctx,const unsigned char key[16])110 int poly1305_init(void *ctx, const unsigned char key[16])
111 {
112 poly1305_internal *st = (poly1305_internal *) ctx;
113 elem64 r0, r1, r2, r3;
114
115 /* h = 0, biased */
116 #if 0
117 st->h[0].d = TWO(52)*TWO0;
118 st->h[1].d = TWO(52)*TWO32;
119 st->h[2].d = TWO(52)*TWO64;
120 st->h[3].d = TWO(52)*TWO96;
121 #else
122 st->h[0].u = EXP(52+0);
123 st->h[1].u = EXP(52+32);
124 st->h[2].u = EXP(52+64);
125 st->h[3].u = EXP(52+96);
126 #endif
127
128 if (key) {
129 /*
130 * set "truncate" rounding mode
131 */
132 #if defined(__x86_64__)
133 u32 mxcsr_orig;
134
135 asm volatile ("stmxcsr %0":"=m"(mxcsr_orig));
136 asm volatile ("ldmxcsr %0"::"m"(mxcsr));
137 #elif defined(__PPC__) || defined(__POWERPC__)
138 double fpscr_orig, fpscr = *(double *)&one;
139
140 asm volatile ("mffs %0":"=f"(fpscr_orig));
141 asm volatile ("mtfsf 255,%0"::"f"(fpscr));
142 #elif defined(__s390x__)
143 u32 fpc_orig;
144
145 asm volatile ("stfpc %0":"=m"(fpc_orig));
146 asm volatile ("lfpc %0"::"m"(fpc));
147 #elif defined(__sparc__)
148 u64 fsr_orig;
149
150 asm volatile ("stx %%fsr,%0":"=m"(fsr_orig));
151 asm volatile ("ldx %0,%%fsr"::"m"(fsr));
152 #elif defined(__mips__)
153 u32 fcsr_orig;
154
155 asm volatile ("cfc1 %0,$31":"=r"(fcsr_orig));
156 asm volatile ("ctc1 %0,$31"::"r"(fcsr));
157 #endif
158
159 /* r &= 0xffffffc0ffffffc0ffffffc0fffffff */
160 r0.u = EXP(52+0) | (U8TOU32(&key[0]) & 0x0fffffff);
161 r1.u = EXP(52+32) | (U8TOU32(&key[4]) & 0x0ffffffc);
162 r2.u = EXP(52+64) | (U8TOU32(&key[8]) & 0x0ffffffc);
163 r3.u = EXP(52+96) | (U8TOU32(&key[12]) & 0x0ffffffc);
164
165 st->r[0] = r0.d - TWO(52)*TWO0;
166 st->r[2] = r1.d - TWO(52)*TWO32;
167 st->r[4] = r2.d - TWO(52)*TWO64;
168 st->r[6] = r3.d - TWO(52)*TWO96;
169
170 st->s[0] = st->r[2] * (5.0/TWO130);
171 st->s[2] = st->r[4] * (5.0/TWO130);
172 st->s[4] = st->r[6] * (5.0/TWO130);
173
174 /*
175 * base 2^32 -> base 2^16
176 */
177 st->r[1] = (st->r[0] + TWO(52)*TWO(16)*TWO0) -
178 TWO(52)*TWO(16)*TWO0;
179 st->r[0] -= st->r[1];
180
181 st->r[3] = (st->r[2] + TWO(52)*TWO(16)*TWO32) -
182 TWO(52)*TWO(16)*TWO32;
183 st->r[2] -= st->r[3];
184
185 st->r[5] = (st->r[4] + TWO(52)*TWO(16)*TWO64) -
186 TWO(52)*TWO(16)*TWO64;
187 st->r[4] -= st->r[5];
188
189 st->r[7] = (st->r[6] + TWO(52)*TWO(16)*TWO96) -
190 TWO(52)*TWO(16)*TWO96;
191 st->r[6] -= st->r[7];
192
193 st->s[1] = (st->s[0] + TWO(52)*TWO(16)*TWO0/TWO96) -
194 TWO(52)*TWO(16)*TWO0/TWO96;
195 st->s[0] -= st->s[1];
196
197 st->s[3] = (st->s[2] + TWO(52)*TWO(16)*TWO32/TWO96) -
198 TWO(52)*TWO(16)*TWO32/TWO96;
199 st->s[2] -= st->s[3];
200
201 st->s[5] = (st->s[4] + TWO(52)*TWO(16)*TWO64/TWO96) -
202 TWO(52)*TWO(16)*TWO64/TWO96;
203 st->s[4] -= st->s[5];
204
205 /*
206 * restore original FPU control register
207 */
208 #if defined(__x86_64__)
209 asm volatile ("ldmxcsr %0"::"m"(mxcsr_orig));
210 #elif defined(__PPC__) || defined(__POWERPC__)
211 asm volatile ("mtfsf 255,%0"::"f"(fpscr_orig));
212 #elif defined(__s390x__)
213 asm volatile ("lfpc %0"::"m"(fpc_orig));
214 #elif defined(__sparc__)
215 asm volatile ("ldx %0,%%fsr"::"m"(fsr_orig));
216 #elif defined(__mips__)
217 asm volatile ("ctc1 %0,$31"::"r"(fcsr_orig));
218 #endif
219 }
220
221 return 0;
222 }
223
poly1305_blocks(void * ctx,const unsigned char * inp,size_t len,int padbit)224 void poly1305_blocks(void *ctx, const unsigned char *inp, size_t len,
225 int padbit)
226 {
227 poly1305_internal *st = (poly1305_internal *)ctx;
228 elem64 in0, in1, in2, in3;
229 u64 pad = (u64)padbit<<32;
230
231 double x0, x1, x2, x3;
232 double h0lo, h0hi, h1lo, h1hi, h2lo, h2hi, h3lo, h3hi;
233 double c0lo, c0hi, c1lo, c1hi, c2lo, c2hi, c3lo, c3hi;
234
235 const double r0lo = st->r[0];
236 const double r0hi = st->r[1];
237 const double r1lo = st->r[2];
238 const double r1hi = st->r[3];
239 const double r2lo = st->r[4];
240 const double r2hi = st->r[5];
241 const double r3lo = st->r[6];
242 const double r3hi = st->r[7];
243
244 const double s1lo = st->s[0];
245 const double s1hi = st->s[1];
246 const double s2lo = st->s[2];
247 const double s2hi = st->s[3];
248 const double s3lo = st->s[4];
249 const double s3hi = st->s[5];
250
251 /*
252 * set "truncate" rounding mode
253 */
254 #if defined(__x86_64__)
255 u32 mxcsr_orig;
256
257 asm volatile ("stmxcsr %0":"=m"(mxcsr_orig));
258 asm volatile ("ldmxcsr %0"::"m"(mxcsr));
259 #elif defined(__PPC__) || defined(__POWERPC__)
260 double fpscr_orig, fpscr = *(double *)&one;
261
262 asm volatile ("mffs %0":"=f"(fpscr_orig));
263 asm volatile ("mtfsf 255,%0"::"f"(fpscr));
264 #elif defined(__s390x__)
265 u32 fpc_orig;
266
267 asm volatile ("stfpc %0":"=m"(fpc_orig));
268 asm volatile ("lfpc %0"::"m"(fpc));
269 #elif defined(__sparc__)
270 u64 fsr_orig;
271
272 asm volatile ("stx %%fsr,%0":"=m"(fsr_orig));
273 asm volatile ("ldx %0,%%fsr"::"m"(fsr));
274 #elif defined(__mips__)
275 u32 fcsr_orig;
276
277 asm volatile ("cfc1 %0,$31":"=r"(fcsr_orig));
278 asm volatile ("ctc1 %0,$31"::"r"(fcsr));
279 #endif
280
281 /*
282 * load base 2^32 and de-bias
283 */
284 h0lo = st->h[0].d - TWO(52)*TWO0;
285 h1lo = st->h[1].d - TWO(52)*TWO32;
286 h2lo = st->h[2].d - TWO(52)*TWO64;
287 h3lo = st->h[3].d - TWO(52)*TWO96;
288
289 #ifdef __clang__
290 h0hi = 0;
291 h1hi = 0;
292 h2hi = 0;
293 h3hi = 0;
294 #else
295 in0.u = EXP(52+0) | U8TOU32(&inp[0]);
296 in1.u = EXP(52+32) | U8TOU32(&inp[4]);
297 in2.u = EXP(52+64) | U8TOU32(&inp[8]);
298 in3.u = EXP(52+96) | U8TOU32(&inp[12]) | pad;
299
300 x0 = in0.d - TWO(52)*TWO0;
301 x1 = in1.d - TWO(52)*TWO32;
302 x2 = in2.d - TWO(52)*TWO64;
303 x3 = in3.d - TWO(52)*TWO96;
304
305 x0 += h0lo;
306 x1 += h1lo;
307 x2 += h2lo;
308 x3 += h3lo;
309
310 goto fast_entry;
311 #endif
312
313 do {
314 in0.u = EXP(52+0) | U8TOU32(&inp[0]);
315 in1.u = EXP(52+32) | U8TOU32(&inp[4]);
316 in2.u = EXP(52+64) | U8TOU32(&inp[8]);
317 in3.u = EXP(52+96) | U8TOU32(&inp[12]) | pad;
318
319 x0 = in0.d - TWO(52)*TWO0;
320 x1 = in1.d - TWO(52)*TWO32;
321 x2 = in2.d - TWO(52)*TWO64;
322 x3 = in3.d - TWO(52)*TWO96;
323
324 /*
325 * note that there are multiple ways to accumulate input, e.g.
326 * one can as well accumulate to h0lo-h1lo-h1hi-h2hi...
327 */
328 h0lo += x0;
329 h0hi += x1;
330 h2lo += x2;
331 h2hi += x3;
332
333 /*
334 * carries that cross 32n-bit (and 130-bit) boundaries
335 */
336 c0lo = (h0lo + TWO(52)*TWO32) - TWO(52)*TWO32;
337 c1lo = (h1lo + TWO(52)*TWO64) - TWO(52)*TWO64;
338 c2lo = (h2lo + TWO(52)*TWO96) - TWO(52)*TWO96;
339 c3lo = (h3lo + TWO(52)*TWO130) - TWO(52)*TWO130;
340
341 c0hi = (h0hi + TWO(52)*TWO32) - TWO(52)*TWO32;
342 c1hi = (h1hi + TWO(52)*TWO64) - TWO(52)*TWO64;
343 c2hi = (h2hi + TWO(52)*TWO96) - TWO(52)*TWO96;
344 c3hi = (h3hi + TWO(52)*TWO130) - TWO(52)*TWO130;
345
346 /*
347 * base 2^48 -> base 2^32 with last reduction step
348 */
349 x1 = (h1lo - c1lo) + c0lo;
350 x2 = (h2lo - c2lo) + c1lo;
351 x3 = (h3lo - c3lo) + c2lo;
352 x0 = (h0lo - c0lo) + c3lo * (5.0/TWO130);
353
354 x1 += (h1hi - c1hi) + c0hi;
355 x2 += (h2hi - c2hi) + c1hi;
356 x3 += (h3hi - c3hi) + c2hi;
357 x0 += (h0hi - c0hi) + c3hi * (5.0/TWO130);
358
359 #ifndef __clang__
360 fast_entry:
361 #endif
362 /*
363 * base 2^32 * base 2^16 = base 2^48
364 */
365 h0lo = s3lo * x1 + s2lo * x2 + s1lo * x3 + r0lo * x0;
366 h1lo = r0lo * x1 + s3lo * x2 + s2lo * x3 + r1lo * x0;
367 h2lo = r1lo * x1 + r0lo * x2 + s3lo * x3 + r2lo * x0;
368 h3lo = r2lo * x1 + r1lo * x2 + r0lo * x3 + r3lo * x0;
369
370 h0hi = s3hi * x1 + s2hi * x2 + s1hi * x3 + r0hi * x0;
371 h1hi = r0hi * x1 + s3hi * x2 + s2hi * x3 + r1hi * x0;
372 h2hi = r1hi * x1 + r0hi * x2 + s3hi * x3 + r2hi * x0;
373 h3hi = r2hi * x1 + r1hi * x2 + r0hi * x3 + r3hi * x0;
374
375 inp += 16;
376 len -= 16;
377
378 } while (len >= 16);
379
380 /*
381 * carries that cross 32n-bit (and 130-bit) boundaries
382 */
383 c0lo = (h0lo + TWO(52)*TWO32) - TWO(52)*TWO32;
384 c1lo = (h1lo + TWO(52)*TWO64) - TWO(52)*TWO64;
385 c2lo = (h2lo + TWO(52)*TWO96) - TWO(52)*TWO96;
386 c3lo = (h3lo + TWO(52)*TWO130) - TWO(52)*TWO130;
387
388 c0hi = (h0hi + TWO(52)*TWO32) - TWO(52)*TWO32;
389 c1hi = (h1hi + TWO(52)*TWO64) - TWO(52)*TWO64;
390 c2hi = (h2hi + TWO(52)*TWO96) - TWO(52)*TWO96;
391 c3hi = (h3hi + TWO(52)*TWO130) - TWO(52)*TWO130;
392
393 /*
394 * base 2^48 -> base 2^32 with last reduction step
395 */
396 x1 = (h1lo - c1lo) + c0lo;
397 x2 = (h2lo - c2lo) + c1lo;
398 x3 = (h3lo - c3lo) + c2lo;
399 x0 = (h0lo - c0lo) + c3lo * (5.0/TWO130);
400
401 x1 += (h1hi - c1hi) + c0hi;
402 x2 += (h2hi - c2hi) + c1hi;
403 x3 += (h3hi - c3hi) + c2hi;
404 x0 += (h0hi - c0hi) + c3hi * (5.0/TWO130);
405
406 /*
407 * store base 2^32, with bias
408 */
409 st->h[1].d = x1 + TWO(52)*TWO32;
410 st->h[2].d = x2 + TWO(52)*TWO64;
411 st->h[3].d = x3 + TWO(52)*TWO96;
412 st->h[0].d = x0 + TWO(52)*TWO0;
413
414 /*
415 * restore original FPU control register
416 */
417 #if defined(__x86_64__)
418 asm volatile ("ldmxcsr %0"::"m"(mxcsr_orig));
419 #elif defined(__PPC__) || defined(__POWERPC__)
420 asm volatile ("mtfsf 255,%0"::"f"(fpscr_orig));
421 #elif defined(__s390x__)
422 asm volatile ("lfpc %0"::"m"(fpc_orig));
423 #elif defined(__sparc__)
424 asm volatile ("ldx %0,%%fsr"::"m"(fsr_orig));
425 #elif defined(__mips__)
426 asm volatile ("ctc1 %0,$31"::"r"(fcsr_orig));
427 #endif
428 }
429
poly1305_emit(void * ctx,unsigned char mac[16],const u32 nonce[4])430 void poly1305_emit(void *ctx, unsigned char mac[16], const u32 nonce[4])
431 {
432 poly1305_internal *st = (poly1305_internal *) ctx;
433 u64 h0, h1, h2, h3, h4;
434 u32 g0, g1, g2, g3, g4;
435 u64 t;
436 u32 mask;
437
438 /*
439 * thanks to bias masking exponent gives integer result
440 */
441 h0 = st->h[0].u & 0x000fffffffffffffULL;
442 h1 = st->h[1].u & 0x000fffffffffffffULL;
443 h2 = st->h[2].u & 0x000fffffffffffffULL;
444 h3 = st->h[3].u & 0x000fffffffffffffULL;
445
446 /*
447 * can be partially reduced, so reduce...
448 */
449 h4 = h3>>32; h3 &= 0xffffffffU;
450 g4 = h4&-4;
451 h4 &= 3;
452 g4 += g4>>2;
453
454 h0 += g4;
455 h1 += h0>>32; h0 &= 0xffffffffU;
456 h2 += h1>>32; h1 &= 0xffffffffU;
457 h3 += h2>>32; h2 &= 0xffffffffU;
458
459 /* compute h + -p */
460 g0 = (u32)(t = h0 + 5);
461 g1 = (u32)(t = h1 + (t >> 32));
462 g2 = (u32)(t = h2 + (t >> 32));
463 g3 = (u32)(t = h3 + (t >> 32));
464 g4 = h4 + (u32)(t >> 32);
465
466 /* if there was carry, select g0-g3 */
467 mask = 0 - (g4 >> 2);
468 g0 &= mask;
469 g1 &= mask;
470 g2 &= mask;
471 g3 &= mask;
472 mask = ~mask;
473 g0 |= (h0 & mask);
474 g1 |= (h1 & mask);
475 g2 |= (h2 & mask);
476 g3 |= (h3 & mask);
477
478 /* mac = (h + nonce) % (2^128) */
479 g0 = (u32)(t = (u64)g0 + nonce[0]);
480 g1 = (u32)(t = (u64)g1 + (t >> 32) + nonce[1]);
481 g2 = (u32)(t = (u64)g2 + (t >> 32) + nonce[2]);
482 g3 = (u32)(t = (u64)g3 + (t >> 32) + nonce[3]);
483
484 U32TO8(mac + 0, g0);
485 U32TO8(mac + 4, g1);
486 U32TO8(mac + 8, g2);
487 U32TO8(mac + 12, g3);
488 }
489
