1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * arch/sparc64/math-emu/math.c
4 *
5 * Copyright (C) 1997,1999 Jakub Jelinek (jj@ultra.linux.cz)
6 * Copyright (C) 1999 David S. Miller (davem@redhat.com)
7 *
8 * Emulation routines originate from soft-fp package, which is part
9 * of glibc and has appropriate copyrights in it.
10 */
11
12 #include <linux/types.h>
13 #include <linux/sched.h>
14 #include <linux/errno.h>
15 #include <linux/perf_event.h>
16
17 #include <asm/fpumacro.h>
18 #include <asm/ptrace.h>
19 #include <linux/uaccess.h>
20 #include <asm/cacheflush.h>
21
22 #include "sfp-util_64.h"
23 #include <math-emu/soft-fp.h>
24 #include <math-emu/single.h>
25 #include <math-emu/double.h>
26 #include <math-emu/quad.h>
27
28 /* QUAD - ftt == 3 */
29 #define FMOVQ 0x003
30 #define FNEGQ 0x007
31 #define FABSQ 0x00b
32 #define FSQRTQ 0x02b
33 #define FADDQ 0x043
34 #define FSUBQ 0x047
35 #define FMULQ 0x04b
36 #define FDIVQ 0x04f
37 #define FDMULQ 0x06e
38 #define FQTOX 0x083
39 #define FXTOQ 0x08c
40 #define FQTOS 0x0c7
41 #define FQTOD 0x0cb
42 #define FITOQ 0x0cc
43 #define FSTOQ 0x0cd
44 #define FDTOQ 0x0ce
45 #define FQTOI 0x0d3
46 /* SUBNORMAL - ftt == 2 */
47 #define FSQRTS 0x029
48 #define FSQRTD 0x02a
49 #define FADDS 0x041
50 #define FADDD 0x042
51 #define FSUBS 0x045
52 #define FSUBD 0x046
53 #define FMULS 0x049
54 #define FMULD 0x04a
55 #define FDIVS 0x04d
56 #define FDIVD 0x04e
57 #define FSMULD 0x069
58 #define FSTOX 0x081
59 #define FDTOX 0x082
60 #define FDTOS 0x0c6
61 #define FSTOD 0x0c9
62 #define FSTOI 0x0d1
63 #define FDTOI 0x0d2
64 #define FXTOS 0x084 /* Only Ultra-III generates this. */
65 #define FXTOD 0x088 /* Only Ultra-III generates this. */
66 #if 0 /* Optimized inline in sparc64/kernel/entry.S */
67 #define FITOS 0x0c4 /* Only Ultra-III generates this. */
68 #endif
69 #define FITOD 0x0c8 /* Only Ultra-III generates this. */
70 /* FPOP2 */
71 #define FCMPQ 0x053
72 #define FCMPEQ 0x057
73 #define FMOVQ0 0x003
74 #define FMOVQ1 0x043
75 #define FMOVQ2 0x083
76 #define FMOVQ3 0x0c3
77 #define FMOVQI 0x103
78 #define FMOVQX 0x183
79 #define FMOVQZ 0x027
80 #define FMOVQLE 0x047
81 #define FMOVQLZ 0x067
82 #define FMOVQNZ 0x0a7
83 #define FMOVQGZ 0x0c7
84 #define FMOVQGE 0x0e7
85
86 #define FSR_TEM_SHIFT 23UL
87 #define FSR_TEM_MASK (0x1fUL << FSR_TEM_SHIFT)
88 #define FSR_AEXC_SHIFT 5UL
89 #define FSR_AEXC_MASK (0x1fUL << FSR_AEXC_SHIFT)
90 #define FSR_CEXC_SHIFT 0UL
91 #define FSR_CEXC_MASK (0x1fUL << FSR_CEXC_SHIFT)
92
93 /* All routines returning an exception to raise should detect
94 * such exceptions _before_ rounding to be consistent with
95 * the behavior of the hardware in the implemented cases
96 * (and thus with the recommendations in the V9 architecture
97 * manual).
98 *
99 * We return 0 if a SIGFPE should be sent, 1 otherwise.
100 */
record_exception(struct pt_regs * regs,int eflag)101 static inline int record_exception(struct pt_regs *regs, int eflag)
102 {
103 u64 fsr = current_thread_info()->xfsr[0];
104 int would_trap;
105
106 /* Determine if this exception would have generated a trap. */
107 would_trap = (fsr & ((long)eflag << FSR_TEM_SHIFT)) != 0UL;
108
109 /* If trapping, we only want to signal one bit. */
110 if(would_trap != 0) {
111 eflag &= ((fsr & FSR_TEM_MASK) >> FSR_TEM_SHIFT);
112 if((eflag & (eflag - 1)) != 0) {
113 if(eflag & FP_EX_INVALID)
114 eflag = FP_EX_INVALID;
115 else if(eflag & FP_EX_OVERFLOW)
116 eflag = FP_EX_OVERFLOW;
117 else if(eflag & FP_EX_UNDERFLOW)
118 eflag = FP_EX_UNDERFLOW;
119 else if(eflag & FP_EX_DIVZERO)
120 eflag = FP_EX_DIVZERO;
121 else if(eflag & FP_EX_INEXACT)
122 eflag = FP_EX_INEXACT;
123 }
124 }
125
126 /* Set CEXC, here is the rule:
127 *
128 * In general all FPU ops will set one and only one
129 * bit in the CEXC field, this is always the case
130 * when the IEEE exception trap is enabled in TEM.
131 */
132 fsr &= ~(FSR_CEXC_MASK);
133 fsr |= ((long)eflag << FSR_CEXC_SHIFT);
134
135 /* Set the AEXC field, rule is:
136 *
137 * If a trap would not be generated, the
138 * CEXC just generated is OR'd into the
139 * existing value of AEXC.
140 */
141 if(would_trap == 0)
142 fsr |= ((long)eflag << FSR_AEXC_SHIFT);
143
144 /* If trapping, indicate fault trap type IEEE. */
145 if(would_trap != 0)
146 fsr |= (1UL << 14);
147
148 current_thread_info()->xfsr[0] = fsr;
149
150 /* If we will not trap, advance the program counter over
151 * the instruction being handled.
152 */
153 if(would_trap == 0) {
154 regs->tpc = regs->tnpc;
155 regs->tnpc += 4;
156 }
157
158 return (would_trap ? 0 : 1);
159 }
160
161 typedef union {
162 u32 s;
163 u64 d;
164 u64 q[2];
165 } *argp;
166
do_mathemu(struct pt_regs * regs,struct fpustate * f,bool illegal_insn_trap)167 int do_mathemu(struct pt_regs *regs, struct fpustate *f, bool illegal_insn_trap)
168 {
169 unsigned long pc = regs->tpc;
170 unsigned long tstate = regs->tstate;
171 u32 insn = 0;
172 int type = 0;
173 /* ftt tells which ftt it may happen in, r is rd, b is rs2 and a is rs1. The *u arg tells
174 whether the argument should be packed/unpacked (0 - do not unpack/pack, 1 - unpack/pack)
175 non-u args tells the size of the argument (0 - no argument, 1 - single, 2 - double, 3 - quad */
176 #define TYPE(ftt, r, ru, b, bu, a, au) type = (au << 2) | (a << 0) | (bu << 5) | (b << 3) | (ru << 8) | (r << 6) | (ftt << 9)
177 int freg;
178 static u64 zero[2] = { 0L, 0L };
179 int flags;
180 FP_DECL_EX;
181 FP_DECL_S(SA); FP_DECL_S(SB); FP_DECL_S(SR);
182 FP_DECL_D(DA); FP_DECL_D(DB); FP_DECL_D(DR);
183 FP_DECL_Q(QA); FP_DECL_Q(QB); FP_DECL_Q(QR);
184 int IR;
185 long XR, xfsr;
186
187 if (tstate & TSTATE_PRIV)
188 die_if_kernel("unfinished/unimplemented FPop from kernel", regs);
189 perf_sw_event(PERF_COUNT_SW_EMULATION_FAULTS, 1, regs, 0);
190 if (test_thread_flag(TIF_32BIT))
191 pc = (u32)pc;
192 if (get_user(insn, (u32 __user *) pc) != -EFAULT) {
193 if ((insn & 0xc1f80000) == 0x81a00000) /* FPOP1 */ {
194 switch ((insn >> 5) & 0x1ff) {
195 /* QUAD - ftt == 3 */
196 case FMOVQ:
197 case FNEGQ:
198 case FABSQ: TYPE(3,3,0,3,0,0,0); break;
199 case FSQRTQ: TYPE(3,3,1,3,1,0,0); break;
200 case FADDQ:
201 case FSUBQ:
202 case FMULQ:
203 case FDIVQ: TYPE(3,3,1,3,1,3,1); break;
204 case FDMULQ: TYPE(3,3,1,2,1,2,1); break;
205 case FQTOX: TYPE(3,2,0,3,1,0,0); break;
206 case FXTOQ: TYPE(3,3,1,2,0,0,0); break;
207 case FQTOS: TYPE(3,1,1,3,1,0,0); break;
208 case FQTOD: TYPE(3,2,1,3,1,0,0); break;
209 case FITOQ: TYPE(3,3,1,1,0,0,0); break;
210 case FSTOQ: TYPE(3,3,1,1,1,0,0); break;
211 case FDTOQ: TYPE(3,3,1,2,1,0,0); break;
212 case FQTOI: TYPE(3,1,0,3,1,0,0); break;
213
214 /* We can get either unimplemented or unfinished
215 * for these cases. Pre-Niagara systems generate
216 * unfinished fpop for SUBNORMAL cases, and Niagara
217 * always gives unimplemented fpop for fsqrt{s,d}.
218 */
219 case FSQRTS: {
220 unsigned long x = current_thread_info()->xfsr[0];
221
222 x = (x >> 14) & 0x7;
223 TYPE(x,1,1,1,1,0,0);
224 break;
225 }
226
227 case FSQRTD: {
228 unsigned long x = current_thread_info()->xfsr[0];
229
230 x = (x >> 14) & 0x7;
231 TYPE(x,2,1,2,1,0,0);
232 break;
233 }
234
235 /* SUBNORMAL - ftt == 2 */
236 case FADDD:
237 case FSUBD:
238 case FMULD:
239 case FDIVD: TYPE(2,2,1,2,1,2,1); break;
240 case FADDS:
241 case FSUBS:
242 case FMULS:
243 case FDIVS: TYPE(2,1,1,1,1,1,1); break;
244 case FSMULD: TYPE(2,2,1,1,1,1,1); break;
245 case FSTOX: TYPE(2,2,0,1,1,0,0); break;
246 case FDTOX: TYPE(2,2,0,2,1,0,0); break;
247 case FDTOS: TYPE(2,1,1,2,1,0,0); break;
248 case FSTOD: TYPE(2,2,1,1,1,0,0); break;
249 case FSTOI: TYPE(2,1,0,1,1,0,0); break;
250 case FDTOI: TYPE(2,1,0,2,1,0,0); break;
251
252 /* Only Ultra-III generates these */
253 case FXTOS: TYPE(2,1,1,2,0,0,0); break;
254 case FXTOD: TYPE(2,2,1,2,0,0,0); break;
255 #if 0 /* Optimized inline in sparc64/kernel/entry.S */
256 case FITOS: TYPE(2,1,1,1,0,0,0); break;
257 #endif
258 case FITOD: TYPE(2,2,1,1,0,0,0); break;
259 }
260 }
261 else if ((insn & 0xc1f80000) == 0x81a80000) /* FPOP2 */ {
262 IR = 2;
263 switch ((insn >> 5) & 0x1ff) {
264 case FCMPQ: TYPE(3,0,0,3,1,3,1); break;
265 case FCMPEQ: TYPE(3,0,0,3,1,3,1); break;
266 /* Now the conditional fmovq support */
267 case FMOVQ0:
268 case FMOVQ1:
269 case FMOVQ2:
270 case FMOVQ3:
271 /* fmovq %fccX, %fY, %fZ */
272 if (!((insn >> 11) & 3))
273 XR = current_thread_info()->xfsr[0] >> 10;
274 else
275 XR = current_thread_info()->xfsr[0] >> (30 + ((insn >> 10) & 0x6));
276 XR &= 3;
277 IR = 0;
278 switch ((insn >> 14) & 0x7) {
279 /* case 0: IR = 0; break; */ /* Never */
280 case 1: if (XR) IR = 1; break; /* Not Equal */
281 case 2: if (XR == 1 || XR == 2) IR = 1; break; /* Less or Greater */
282 case 3: if (XR & 1) IR = 1; break; /* Unordered or Less */
283 case 4: if (XR == 1) IR = 1; break; /* Less */
284 case 5: if (XR & 2) IR = 1; break; /* Unordered or Greater */
285 case 6: if (XR == 2) IR = 1; break; /* Greater */
286 case 7: if (XR == 3) IR = 1; break; /* Unordered */
287 }
288 if ((insn >> 14) & 8)
289 IR ^= 1;
290 break;
291 case FMOVQI:
292 case FMOVQX:
293 /* fmovq %[ix]cc, %fY, %fZ */
294 XR = regs->tstate >> 32;
295 if ((insn >> 5) & 0x80)
296 XR >>= 4;
297 XR &= 0xf;
298 IR = 0;
299 freg = ((XR >> 2) ^ XR) & 2;
300 switch ((insn >> 14) & 0x7) {
301 /* case 0: IR = 0; break; */ /* Never */
302 case 1: if (XR & 4) IR = 1; break; /* Equal */
303 case 2: if ((XR & 4) || freg) IR = 1; break; /* Less or Equal */
304 case 3: if (freg) IR = 1; break; /* Less */
305 case 4: if (XR & 5) IR = 1; break; /* Less or Equal Unsigned */
306 case 5: if (XR & 1) IR = 1; break; /* Carry Set */
307 case 6: if (XR & 8) IR = 1; break; /* Negative */
308 case 7: if (XR & 2) IR = 1; break; /* Overflow Set */
309 }
310 if ((insn >> 14) & 8)
311 IR ^= 1;
312 break;
313 case FMOVQZ:
314 case FMOVQLE:
315 case FMOVQLZ:
316 case FMOVQNZ:
317 case FMOVQGZ:
318 case FMOVQGE:
319 freg = (insn >> 14) & 0x1f;
320 if (!freg)
321 XR = 0;
322 else if (freg < 16)
323 XR = regs->u_regs[freg];
324 else if (!test_thread_64bit_stack(regs->u_regs[UREG_FP])) {
325 struct reg_window32 __user *win32;
326 flushw_user ();
327 win32 = (struct reg_window32 __user *)((unsigned long)((u32)regs->u_regs[UREG_FP]));
328 get_user(XR, &win32->locals[freg - 16]);
329 } else {
330 struct reg_window __user *win;
331 flushw_user ();
332 win = (struct reg_window __user *)(regs->u_regs[UREG_FP] + STACK_BIAS);
333 get_user(XR, &win->locals[freg - 16]);
334 }
335 IR = 0;
336 switch ((insn >> 10) & 3) {
337 case 1: if (!XR) IR = 1; break; /* Register Zero */
338 case 2: if (XR <= 0) IR = 1; break; /* Register Less Than or Equal to Zero */
339 case 3: if (XR < 0) IR = 1; break; /* Register Less Than Zero */
340 }
341 if ((insn >> 10) & 4)
342 IR ^= 1;
343 break;
344 }
345 if (IR == 0) {
346 /* The fmov test was false. Do a nop instead */
347 current_thread_info()->xfsr[0] &= ~(FSR_CEXC_MASK);
348 regs->tpc = regs->tnpc;
349 regs->tnpc += 4;
350 return 1;
351 } else if (IR == 1) {
352 /* Change the instruction into plain fmovq */
353 insn = (insn & 0x3e00001f) | 0x81a00060;
354 TYPE(3,3,0,3,0,0,0);
355 }
356 }
357 }
358 if (type) {
359 argp rs1 = NULL, rs2 = NULL, rd = NULL;
360
361 /* Starting with UltraSPARC-T2, the cpu does not set the FP Trap
362 * Type field in the %fsr to unimplemented_FPop. Nor does it
363 * use the fp_exception_other trap. Instead it signals an
364 * illegal instruction and leaves the FP trap type field of
365 * the %fsr unchanged.
366 */
367 if (!illegal_insn_trap) {
368 int ftt = (current_thread_info()->xfsr[0] >> 14) & 0x7;
369 if (ftt != (type >> 9))
370 goto err;
371 }
372 current_thread_info()->xfsr[0] &= ~0x1c000;
373 freg = ((insn >> 14) & 0x1f);
374 switch (type & 0x3) {
375 case 3: if (freg & 2) {
376 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */;
377 goto err;
378 }
379 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e);
380 case 1: rs1 = (argp)&f->regs[freg];
381 flags = (freg < 32) ? FPRS_DL : FPRS_DU;
382 if (!(current_thread_info()->fpsaved[0] & flags))
383 rs1 = (argp)&zero;
384 break;
385 }
386 switch (type & 0x7) {
387 case 7: FP_UNPACK_QP (QA, rs1); break;
388 case 6: FP_UNPACK_DP (DA, rs1); break;
389 case 5: FP_UNPACK_SP (SA, rs1); break;
390 }
391 freg = (insn & 0x1f);
392 switch ((type >> 3) & 0x3) {
393 case 3: if (freg & 2) {
394 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */;
395 goto err;
396 }
397 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e);
398 case 1: rs2 = (argp)&f->regs[freg];
399 flags = (freg < 32) ? FPRS_DL : FPRS_DU;
400 if (!(current_thread_info()->fpsaved[0] & flags))
401 rs2 = (argp)&zero;
402 break;
403 }
404 switch ((type >> 3) & 0x7) {
405 case 7: FP_UNPACK_QP (QB, rs2); break;
406 case 6: FP_UNPACK_DP (DB, rs2); break;
407 case 5: FP_UNPACK_SP (SB, rs2); break;
408 }
409 freg = ((insn >> 25) & 0x1f);
410 switch ((type >> 6) & 0x3) {
411 case 3: if (freg & 2) {
412 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */;
413 goto err;
414 }
415 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e);
416 case 1: rd = (argp)&f->regs[freg];
417 flags = (freg < 32) ? FPRS_DL : FPRS_DU;
418 if (!(current_thread_info()->fpsaved[0] & FPRS_FEF)) {
419 current_thread_info()->fpsaved[0] = FPRS_FEF;
420 current_thread_info()->gsr[0] = 0;
421 }
422 if (!(current_thread_info()->fpsaved[0] & flags)) {
423 if (freg < 32)
424 memset(f->regs, 0, 32*sizeof(u32));
425 else
426 memset(f->regs+32, 0, 32*sizeof(u32));
427 }
428 current_thread_info()->fpsaved[0] |= flags;
429 break;
430 }
431 switch ((insn >> 5) & 0x1ff) {
432 /* + */
433 case FADDS: FP_ADD_S (SR, SA, SB); break;
434 case FADDD: FP_ADD_D (DR, DA, DB); break;
435 case FADDQ: FP_ADD_Q (QR, QA, QB); break;
436 /* - */
437 case FSUBS: FP_SUB_S (SR, SA, SB); break;
438 case FSUBD: FP_SUB_D (DR, DA, DB); break;
439 case FSUBQ: FP_SUB_Q (QR, QA, QB); break;
440 /* * */
441 case FMULS: FP_MUL_S (SR, SA, SB); break;
442 case FSMULD: FP_CONV (D, S, 1, 1, DA, SA);
443 FP_CONV (D, S, 1, 1, DB, SB);
444 case FMULD: FP_MUL_D (DR, DA, DB); break;
445 case FDMULQ: FP_CONV (Q, D, 2, 1, QA, DA);
446 FP_CONV (Q, D, 2, 1, QB, DB);
447 case FMULQ: FP_MUL_Q (QR, QA, QB); break;
448 /* / */
449 case FDIVS: FP_DIV_S (SR, SA, SB); break;
450 case FDIVD: FP_DIV_D (DR, DA, DB); break;
451 case FDIVQ: FP_DIV_Q (QR, QA, QB); break;
452 /* sqrt */
453 case FSQRTS: FP_SQRT_S (SR, SB); break;
454 case FSQRTD: FP_SQRT_D (DR, DB); break;
455 case FSQRTQ: FP_SQRT_Q (QR, QB); break;
456 /* mov */
457 case FMOVQ: rd->q[0] = rs2->q[0]; rd->q[1] = rs2->q[1]; break;
458 case FABSQ: rd->q[0] = rs2->q[0] & 0x7fffffffffffffffUL; rd->q[1] = rs2->q[1]; break;
459 case FNEGQ: rd->q[0] = rs2->q[0] ^ 0x8000000000000000UL; rd->q[1] = rs2->q[1]; break;
460 /* float to int */
461 case FSTOI: FP_TO_INT_S (IR, SB, 32, 1); break;
462 case FDTOI: FP_TO_INT_D (IR, DB, 32, 1); break;
463 case FQTOI: FP_TO_INT_Q (IR, QB, 32, 1); break;
464 case FSTOX: FP_TO_INT_S (XR, SB, 64, 1); break;
465 case FDTOX: FP_TO_INT_D (XR, DB, 64, 1); break;
466 case FQTOX: FP_TO_INT_Q (XR, QB, 64, 1); break;
467 /* int to float */
468 case FITOQ: IR = rs2->s; FP_FROM_INT_Q (QR, IR, 32, int); break;
469 case FXTOQ: XR = rs2->d; FP_FROM_INT_Q (QR, XR, 64, long); break;
470 /* Only Ultra-III generates these */
471 case FXTOS: XR = rs2->d; FP_FROM_INT_S (SR, XR, 64, long); break;
472 case FXTOD: XR = rs2->d; FP_FROM_INT_D (DR, XR, 64, long); break;
473 #if 0 /* Optimized inline in sparc64/kernel/entry.S */
474 case FITOS: IR = rs2->s; FP_FROM_INT_S (SR, IR, 32, int); break;
475 #endif
476 case FITOD: IR = rs2->s; FP_FROM_INT_D (DR, IR, 32, int); break;
477 /* float to float */
478 case FSTOD: FP_CONV (D, S, 1, 1, DR, SB); break;
479 case FSTOQ: FP_CONV (Q, S, 2, 1, QR, SB); break;
480 case FDTOQ: FP_CONV (Q, D, 2, 1, QR, DB); break;
481 case FDTOS: FP_CONV (S, D, 1, 1, SR, DB); break;
482 case FQTOS: FP_CONV (S, Q, 1, 2, SR, QB); break;
483 case FQTOD: FP_CONV (D, Q, 1, 2, DR, QB); break;
484 /* comparison */
485 case FCMPQ:
486 case FCMPEQ:
487 FP_CMP_Q(XR, QB, QA, 3);
488 if (XR == 3 &&
489 (((insn >> 5) & 0x1ff) == FCMPEQ ||
490 FP_ISSIGNAN_Q(QA) ||
491 FP_ISSIGNAN_Q(QB)))
492 FP_SET_EXCEPTION (FP_EX_INVALID);
493 }
494 if (!FP_INHIBIT_RESULTS) {
495 switch ((type >> 6) & 0x7) {
496 case 0: xfsr = current_thread_info()->xfsr[0];
497 if (XR == -1) XR = 2;
498 switch (freg & 3) {
499 /* fcc0, 1, 2, 3 */
500 case 0: xfsr &= ~0xc00; xfsr |= (XR << 10); break;
501 case 1: xfsr &= ~0x300000000UL; xfsr |= (XR << 32); break;
502 case 2: xfsr &= ~0xc00000000UL; xfsr |= (XR << 34); break;
503 case 3: xfsr &= ~0x3000000000UL; xfsr |= (XR << 36); break;
504 }
505 current_thread_info()->xfsr[0] = xfsr;
506 break;
507 case 1: rd->s = IR; break;
508 case 2: rd->d = XR; break;
509 case 5: FP_PACK_SP (rd, SR); break;
510 case 6: FP_PACK_DP (rd, DR); break;
511 case 7: FP_PACK_QP (rd, QR); break;
512 }
513 }
514
515 if(_fex != 0)
516 return record_exception(regs, _fex);
517
518 /* Success and no exceptions detected. */
519 current_thread_info()->xfsr[0] &= ~(FSR_CEXC_MASK);
520 regs->tpc = regs->tnpc;
521 regs->tnpc += 4;
522 return 1;
523 }
524 err: return 0;
525 }
526