aboutsummaryrefslogtreecommitdiffstats
path: root/arch/sh/kernel/cpu/sh4/fpu.c
diff options
context:
space:
mode:
Diffstat (limited to 'arch/sh/kernel/cpu/sh4/fpu.c')
-rw-r--r--arch/sh/kernel/cpu/sh4/fpu.c159
1 files changed, 39 insertions, 120 deletions
diff --git a/arch/sh/kernel/cpu/sh4/fpu.c b/arch/sh/kernel/cpu/sh4/fpu.c
index e97857aec8a0..447482d7f65e 100644
--- a/arch/sh/kernel/cpu/sh4/fpu.c
+++ b/arch/sh/kernel/cpu/sh4/fpu.c
@@ -85,14 +85,14 @@ void save_fpu(struct task_struct *tsk)
85 "fmov.s fr1, @-%0\n\t" 85 "fmov.s fr1, @-%0\n\t"
86 "fmov.s fr0, @-%0\n\t" 86 "fmov.s fr0, @-%0\n\t"
87 "lds %3, fpscr\n\t":"=r" (dummy) 87 "lds %3, fpscr\n\t":"=r" (dummy)
88 :"0"((char *)(&tsk->thread.fpu.hard.status)), 88 :"0"((char *)(&tsk->thread.xstate->hardfpu.status)),
89 "r"(FPSCR_RCHG), "r"(FPSCR_INIT) 89 "r"(FPSCR_RCHG), "r"(FPSCR_INIT)
90 :"memory"); 90 :"memory");
91 91
92 disable_fpu(); 92 disable_fpu();
93} 93}
94 94
95static void restore_fpu(struct task_struct *tsk) 95void restore_fpu(struct task_struct *tsk)
96{ 96{
97 unsigned long dummy; 97 unsigned long dummy;
98 98
@@ -135,62 +135,11 @@ static void restore_fpu(struct task_struct *tsk)
135 "lds.l @%0+, fpscr\n\t" 135 "lds.l @%0+, fpscr\n\t"
136 "lds.l @%0+, fpul\n\t" 136 "lds.l @%0+, fpul\n\t"
137 :"=r" (dummy) 137 :"=r" (dummy)
138 :"0"(&tsk->thread.fpu), "r"(FPSCR_RCHG) 138 :"0" (tsk->thread.xstate), "r" (FPSCR_RCHG)
139 :"memory"); 139 :"memory");
140 disable_fpu(); 140 disable_fpu();
141} 141}
142 142
143/*
144 * Load the FPU with signalling NANS. This bit pattern we're using
145 * has the property that no matter wether considered as single or as
146 * double precision represents signaling NANS.
147 */
148
149static void fpu_init(void)
150{
151 enable_fpu();
152 asm volatile ( "lds %0, fpul\n\t"
153 "lds %1, fpscr\n\t"
154 "fsts fpul, fr0\n\t"
155 "fsts fpul, fr1\n\t"
156 "fsts fpul, fr2\n\t"
157 "fsts fpul, fr3\n\t"
158 "fsts fpul, fr4\n\t"
159 "fsts fpul, fr5\n\t"
160 "fsts fpul, fr6\n\t"
161 "fsts fpul, fr7\n\t"
162 "fsts fpul, fr8\n\t"
163 "fsts fpul, fr9\n\t"
164 "fsts fpul, fr10\n\t"
165 "fsts fpul, fr11\n\t"
166 "fsts fpul, fr12\n\t"
167 "fsts fpul, fr13\n\t"
168 "fsts fpul, fr14\n\t"
169 "fsts fpul, fr15\n\t"
170 "frchg\n\t"
171 "fsts fpul, fr0\n\t"
172 "fsts fpul, fr1\n\t"
173 "fsts fpul, fr2\n\t"
174 "fsts fpul, fr3\n\t"
175 "fsts fpul, fr4\n\t"
176 "fsts fpul, fr5\n\t"
177 "fsts fpul, fr6\n\t"
178 "fsts fpul, fr7\n\t"
179 "fsts fpul, fr8\n\t"
180 "fsts fpul, fr9\n\t"
181 "fsts fpul, fr10\n\t"
182 "fsts fpul, fr11\n\t"
183 "fsts fpul, fr12\n\t"
184 "fsts fpul, fr13\n\t"
185 "fsts fpul, fr14\n\t"
186 "fsts fpul, fr15\n\t"
187 "frchg\n\t"
188 "lds %2, fpscr\n\t"
189 : /* no output */
190 :"r" (0), "r"(FPSCR_RCHG), "r"(FPSCR_INIT));
191 disable_fpu();
192}
193
194/** 143/**
195 * denormal_to_double - Given denormalized float number, 144 * denormal_to_double - Given denormalized float number,
196 * store double float 145 * store double float
@@ -282,9 +231,9 @@ static int ieee_fpe_handler(struct pt_regs *regs)
282 /* fcnvsd */ 231 /* fcnvsd */
283 struct task_struct *tsk = current; 232 struct task_struct *tsk = current;
284 233
285 if ((tsk->thread.fpu.hard.fpscr & FPSCR_CAUSE_ERROR)) 234 if ((tsk->thread.xstate->hardfpu.fpscr & FPSCR_CAUSE_ERROR))
286 /* FPU error */ 235 /* FPU error */
287 denormal_to_double(&tsk->thread.fpu.hard, 236 denormal_to_double(&tsk->thread.xstate->hardfpu,
288 (finsn >> 8) & 0xf); 237 (finsn >> 8) & 0xf);
289 else 238 else
290 return 0; 239 return 0;
@@ -300,9 +249,9 @@ static int ieee_fpe_handler(struct pt_regs *regs)
300 249
301 n = (finsn >> 8) & 0xf; 250 n = (finsn >> 8) & 0xf;
302 m = (finsn >> 4) & 0xf; 251 m = (finsn >> 4) & 0xf;
303 hx = tsk->thread.fpu.hard.fp_regs[n]; 252 hx = tsk->thread.xstate->hardfpu.fp_regs[n];
304 hy = tsk->thread.fpu.hard.fp_regs[m]; 253 hy = tsk->thread.xstate->hardfpu.fp_regs[m];
305 fpscr = tsk->thread.fpu.hard.fpscr; 254 fpscr = tsk->thread.xstate->hardfpu.fpscr;
306 prec = fpscr & FPSCR_DBL_PRECISION; 255 prec = fpscr & FPSCR_DBL_PRECISION;
307 256
308 if ((fpscr & FPSCR_CAUSE_ERROR) 257 if ((fpscr & FPSCR_CAUSE_ERROR)
@@ -312,18 +261,18 @@ static int ieee_fpe_handler(struct pt_regs *regs)
312 261
313 /* FPU error because of denormal (doubles) */ 262 /* FPU error because of denormal (doubles) */
314 llx = ((long long)hx << 32) 263 llx = ((long long)hx << 32)
315 | tsk->thread.fpu.hard.fp_regs[n + 1]; 264 | tsk->thread.xstate->hardfpu.fp_regs[n + 1];
316 lly = ((long long)hy << 32) 265 lly = ((long long)hy << 32)
317 | tsk->thread.fpu.hard.fp_regs[m + 1]; 266 | tsk->thread.xstate->hardfpu.fp_regs[m + 1];
318 llx = float64_mul(llx, lly); 267 llx = float64_mul(llx, lly);
319 tsk->thread.fpu.hard.fp_regs[n] = llx >> 32; 268 tsk->thread.xstate->hardfpu.fp_regs[n] = llx >> 32;
320 tsk->thread.fpu.hard.fp_regs[n + 1] = llx & 0xffffffff; 269 tsk->thread.xstate->hardfpu.fp_regs[n + 1] = llx & 0xffffffff;
321 } else if ((fpscr & FPSCR_CAUSE_ERROR) 270 } else if ((fpscr & FPSCR_CAUSE_ERROR)
322 && (!prec && ((hx & 0x7fffffff) < 0x00800000 271 && (!prec && ((hx & 0x7fffffff) < 0x00800000
323 || (hy & 0x7fffffff) < 0x00800000))) { 272 || (hy & 0x7fffffff) < 0x00800000))) {
324 /* FPU error because of denormal (floats) */ 273 /* FPU error because of denormal (floats) */
325 hx = float32_mul(hx, hy); 274 hx = float32_mul(hx, hy);
326 tsk->thread.fpu.hard.fp_regs[n] = hx; 275 tsk->thread.xstate->hardfpu.fp_regs[n] = hx;
327 } else 276 } else
328 return 0; 277 return 0;
329 278
@@ -338,9 +287,9 @@ static int ieee_fpe_handler(struct pt_regs *regs)
338 287
339 n = (finsn >> 8) & 0xf; 288 n = (finsn >> 8) & 0xf;
340 m = (finsn >> 4) & 0xf; 289 m = (finsn >> 4) & 0xf;
341 hx = tsk->thread.fpu.hard.fp_regs[n]; 290 hx = tsk->thread.xstate->hardfpu.fp_regs[n];
342 hy = tsk->thread.fpu.hard.fp_regs[m]; 291 hy = tsk->thread.xstate->hardfpu.fp_regs[m];
343 fpscr = tsk->thread.fpu.hard.fpscr; 292 fpscr = tsk->thread.xstate->hardfpu.fpscr;
344 prec = fpscr & FPSCR_DBL_PRECISION; 293 prec = fpscr & FPSCR_DBL_PRECISION;
345 294
346 if ((fpscr & FPSCR_CAUSE_ERROR) 295 if ((fpscr & FPSCR_CAUSE_ERROR)
@@ -350,15 +299,15 @@ static int ieee_fpe_handler(struct pt_regs *regs)
350 299
351 /* FPU error because of denormal (doubles) */ 300 /* FPU error because of denormal (doubles) */
352 llx = ((long long)hx << 32) 301 llx = ((long long)hx << 32)
353 | tsk->thread.fpu.hard.fp_regs[n + 1]; 302 | tsk->thread.xstate->hardfpu.fp_regs[n + 1];
354 lly = ((long long)hy << 32) 303 lly = ((long long)hy << 32)
355 | tsk->thread.fpu.hard.fp_regs[m + 1]; 304 | tsk->thread.xstate->hardfpu.fp_regs[m + 1];
356 if ((finsn & 0xf00f) == 0xf000) 305 if ((finsn & 0xf00f) == 0xf000)
357 llx = float64_add(llx, lly); 306 llx = float64_add(llx, lly);
358 else 307 else
359 llx = float64_sub(llx, lly); 308 llx = float64_sub(llx, lly);
360 tsk->thread.fpu.hard.fp_regs[n] = llx >> 32; 309 tsk->thread.xstate->hardfpu.fp_regs[n] = llx >> 32;
361 tsk->thread.fpu.hard.fp_regs[n + 1] = llx & 0xffffffff; 310 tsk->thread.xstate->hardfpu.fp_regs[n + 1] = llx & 0xffffffff;
362 } else if ((fpscr & FPSCR_CAUSE_ERROR) 311 } else if ((fpscr & FPSCR_CAUSE_ERROR)
363 && (!prec && ((hx & 0x7fffffff) < 0x00800000 312 && (!prec && ((hx & 0x7fffffff) < 0x00800000
364 || (hy & 0x7fffffff) < 0x00800000))) { 313 || (hy & 0x7fffffff) < 0x00800000))) {
@@ -367,7 +316,7 @@ static int ieee_fpe_handler(struct pt_regs *regs)
367 hx = float32_add(hx, hy); 316 hx = float32_add(hx, hy);
368 else 317 else
369 hx = float32_sub(hx, hy); 318 hx = float32_sub(hx, hy);
370 tsk->thread.fpu.hard.fp_regs[n] = hx; 319 tsk->thread.xstate->hardfpu.fp_regs[n] = hx;
371 } else 320 } else
372 return 0; 321 return 0;
373 322
@@ -382,9 +331,9 @@ static int ieee_fpe_handler(struct pt_regs *regs)
382 331
383 n = (finsn >> 8) & 0xf; 332 n = (finsn >> 8) & 0xf;
384 m = (finsn >> 4) & 0xf; 333 m = (finsn >> 4) & 0xf;
385 hx = tsk->thread.fpu.hard.fp_regs[n]; 334 hx = tsk->thread.xstate->hardfpu.fp_regs[n];
386 hy = tsk->thread.fpu.hard.fp_regs[m]; 335 hy = tsk->thread.xstate->hardfpu.fp_regs[m];
387 fpscr = tsk->thread.fpu.hard.fpscr; 336 fpscr = tsk->thread.xstate->hardfpu.fpscr;
388 prec = fpscr & FPSCR_DBL_PRECISION; 337 prec = fpscr & FPSCR_DBL_PRECISION;
389 338
390 if ((fpscr & FPSCR_CAUSE_ERROR) 339 if ((fpscr & FPSCR_CAUSE_ERROR)
@@ -394,20 +343,20 @@ static int ieee_fpe_handler(struct pt_regs *regs)
394 343
395 /* FPU error because of denormal (doubles) */ 344 /* FPU error because of denormal (doubles) */
396 llx = ((long long)hx << 32) 345 llx = ((long long)hx << 32)
397 | tsk->thread.fpu.hard.fp_regs[n + 1]; 346 | tsk->thread.xstate->hardfpu.fp_regs[n + 1];
398 lly = ((long long)hy << 32) 347 lly = ((long long)hy << 32)
399 | tsk->thread.fpu.hard.fp_regs[m + 1]; 348 | tsk->thread.xstate->hardfpu.fp_regs[m + 1];
400 349
401 llx = float64_div(llx, lly); 350 llx = float64_div(llx, lly);
402 351
403 tsk->thread.fpu.hard.fp_regs[n] = llx >> 32; 352 tsk->thread.xstate->hardfpu.fp_regs[n] = llx >> 32;
404 tsk->thread.fpu.hard.fp_regs[n + 1] = llx & 0xffffffff; 353 tsk->thread.xstate->hardfpu.fp_regs[n + 1] = llx & 0xffffffff;
405 } else if ((fpscr & FPSCR_CAUSE_ERROR) 354 } else if ((fpscr & FPSCR_CAUSE_ERROR)
406 && (!prec && ((hx & 0x7fffffff) < 0x00800000 355 && (!prec && ((hx & 0x7fffffff) < 0x00800000
407 || (hy & 0x7fffffff) < 0x00800000))) { 356 || (hy & 0x7fffffff) < 0x00800000))) {
408 /* FPU error because of denormal (floats) */ 357 /* FPU error because of denormal (floats) */
409 hx = float32_div(hx, hy); 358 hx = float32_div(hx, hy);
410 tsk->thread.fpu.hard.fp_regs[n] = hx; 359 tsk->thread.xstate->hardfpu.fp_regs[n] = hx;
411 } else 360 } else
412 return 0; 361 return 0;
413 362
@@ -420,17 +369,17 @@ static int ieee_fpe_handler(struct pt_regs *regs)
420 unsigned int hx; 369 unsigned int hx;
421 370
422 m = (finsn >> 8) & 0x7; 371 m = (finsn >> 8) & 0x7;
423 hx = tsk->thread.fpu.hard.fp_regs[m]; 372 hx = tsk->thread.xstate->hardfpu.fp_regs[m];
424 373
425 if ((tsk->thread.fpu.hard.fpscr & FPSCR_CAUSE_ERROR) 374 if ((tsk->thread.xstate->hardfpu.fpscr & FPSCR_CAUSE_ERROR)
426 && ((hx & 0x7fffffff) < 0x00100000)) { 375 && ((hx & 0x7fffffff) < 0x00100000)) {
427 /* subnormal double to float conversion */ 376 /* subnormal double to float conversion */
428 long long llx; 377 long long llx;
429 378
430 llx = ((long long)tsk->thread.fpu.hard.fp_regs[m] << 32) 379 llx = ((long long)tsk->thread.xstate->hardfpu.fp_regs[m] << 32)
431 | tsk->thread.fpu.hard.fp_regs[m + 1]; 380 | tsk->thread.xstate->hardfpu.fp_regs[m + 1];
432 381
433 tsk->thread.fpu.hard.fpul = float64_to_float32(llx); 382 tsk->thread.xstate->hardfpu.fpul = float64_to_float32(llx);
434 } else 383 } else
435 return 0; 384 return 0;
436 385
@@ -449,7 +398,7 @@ void float_raise(unsigned int flags)
449int float_rounding_mode(void) 398int float_rounding_mode(void)
450{ 399{
451 struct task_struct *tsk = current; 400 struct task_struct *tsk = current;
452 int roundingMode = FPSCR_ROUNDING_MODE(tsk->thread.fpu.hard.fpscr); 401 int roundingMode = FPSCR_ROUNDING_MODE(tsk->thread.xstate->hardfpu.fpscr);
453 return roundingMode; 402 return roundingMode;
454} 403}
455 404
@@ -461,16 +410,16 @@ BUILD_TRAP_HANDLER(fpu_error)
461 __unlazy_fpu(tsk, regs); 410 __unlazy_fpu(tsk, regs);
462 fpu_exception_flags = 0; 411 fpu_exception_flags = 0;
463 if (ieee_fpe_handler(regs)) { 412 if (ieee_fpe_handler(regs)) {
464 tsk->thread.fpu.hard.fpscr &= 413 tsk->thread.xstate->hardfpu.fpscr &=
465 ~(FPSCR_CAUSE_MASK | FPSCR_FLAG_MASK); 414 ~(FPSCR_CAUSE_MASK | FPSCR_FLAG_MASK);
466 tsk->thread.fpu.hard.fpscr |= fpu_exception_flags; 415 tsk->thread.xstate->hardfpu.fpscr |= fpu_exception_flags;
467 /* Set the FPSCR flag as well as cause bits - simply 416 /* Set the FPSCR flag as well as cause bits - simply
468 * replicate the cause */ 417 * replicate the cause */
469 tsk->thread.fpu.hard.fpscr |= (fpu_exception_flags >> 10); 418 tsk->thread.xstate->hardfpu.fpscr |= (fpu_exception_flags >> 10);
470 grab_fpu(regs); 419 grab_fpu(regs);
471 restore_fpu(tsk); 420 restore_fpu(tsk);
472 task_thread_info(tsk)->status |= TS_USEDFPU; 421 task_thread_info(tsk)->status |= TS_USEDFPU;
473 if ((((tsk->thread.fpu.hard.fpscr & FPSCR_ENABLE_MASK) >> 7) & 422 if ((((tsk->thread.xstate->hardfpu.fpscr & FPSCR_ENABLE_MASK) >> 7) &
474 (fpu_exception_flags >> 2)) == 0) { 423 (fpu_exception_flags >> 2)) == 0) {
475 return; 424 return;
476 } 425 }
@@ -478,33 +427,3 @@ BUILD_TRAP_HANDLER(fpu_error)
478 427
479 force_sig(SIGFPE, tsk); 428 force_sig(SIGFPE, tsk);
480} 429}
481
482void fpu_state_restore(struct pt_regs *regs)
483{
484 struct task_struct *tsk = current;
485
486 grab_fpu(regs);
487 if (unlikely(!user_mode(regs))) {
488 printk(KERN_ERR "BUG: FPU is used in kernel mode.\n");
489 BUG();
490 return;
491 }
492
493 if (likely(used_math())) {
494 /* Using the FPU again. */
495 restore_fpu(tsk);
496 } else {
497 /* First time FPU user. */
498 fpu_init();
499 set_used_math();
500 }
501 task_thread_info(tsk)->status |= TS_USEDFPU;
502 tsk->fpu_counter++;
503}
504
505BUILD_TRAP_HANDLER(fpu_state_restore)
506{
507 TRAP_HANDLER_DECL;
508
509 fpu_state_restore(regs);
510}
generated by cgit v1.2.2 (git 2.25.0) at 2025-03-03 06:02:27 -0500