1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * arch/sparc/math-emu/math.c
4 *
5 * Copyright (C) 1998 Peter Maydell (pmaydell@chiark.greenend.org.uk)
6 * Copyright (C) 1997, 1999 Jakub Jelinek (jj@ultra.linux.cz)
7 * Copyright (C) 1999 David S. Miller (davem@redhat.com)
8 *
9 * This is a good place to start if you're trying to understand the
10 * emulation code, because it's pretty simple. What we do is
11 * essentially analyse the instruction to work out what the operation
12 * is and which registers are involved. We then execute the appropriate
13 * FXXXX function. [The floating point queue introduces a minor wrinkle;
14 * see below...]
15 * The fxxxxx.c files each emulate a single insn. They look relatively
16 * simple because the complexity is hidden away in an unholy tangle
17 * of preprocessor macros.
18 *
19 * The first layer of macros is single.h, double.h, quad.h. Generally
20 * these files define macros for working with floating point numbers
21 * of the three IEEE formats. FP_ADD_D(R,A,B) is for adding doubles,
22 * for instance. These macros are usually defined as calls to more
23 * generic macros (in this case _FP_ADD(D,2,R,X,Y) where the number
24 * of machine words required to store the given IEEE format is passed
25 * as a parameter. [double.h and co check the number of bits in a word
26 * and define FP_ADD_D & co appropriately].
27 * The generic macros are defined in op-common.h. This is where all
28 * the grotty stuff like handling NaNs is coded. To handle the possible
29 * word sizes macros in op-common.h use macros like _FP_FRAC_SLL_##wc()
30 * where wc is the 'number of machine words' parameter (here 2).
31 * These are defined in the third layer of macros: op-1.h, op-2.h
32 * and op-4.h. These handle operations on floating point numbers composed
33 * of 1,2 and 4 machine words respectively. [For example, on sparc64
34 * doubles are one machine word so macros in double.h eventually use
35 * constructs in op-1.h, but on sparc32 they use op-2.h definitions.]
36 * soft-fp.h is on the same level as op-common.h, and defines some
37 * macros which are independent of both word size and FP format.
38 * Finally, sfp-machine.h is the machine dependent part of the
39 * code: it defines the word size and what type a word is. It also
40 * defines how _FP_MUL_MEAT_t() maps to _FP_MUL_MEAT_n_* : op-n.h
41 * provide several possible flavours of multiply algorithm, most
42 * of which require that you supply some form of asm or C primitive to
43 * do the actual multiply. (such asm primitives should be defined
44 * in sfp-machine.h too). udivmodti4.c is the same sort of thing.
45 *
46 * There may be some errors here because I'm working from a
47 * SPARC architecture manual V9, and what I really want is V8...
48 * Also, the insns which can generate exceptions seem to be a
49 * greater subset of the FPops than for V9 (for example, FCMPED
50 * has to be emulated on V8). So I think I'm going to have
51 * to emulate them all just to be on the safe side...
52 *
53 * Emulation routines originate from soft-fp package, which is
54 * part of glibc and has appropriate copyrights in it (allegedly).
55 *
56 * NB: on sparc int == long == 4 bytes, long long == 8 bytes.
57 * Most bits of the kernel seem to go for long rather than int,
58 * so we follow that practice...
59 */
60
61 /* TODO:
62 * fpsave() saves the FP queue but fpload() doesn't reload it.
63 * Therefore when we context switch or change FPU ownership
64 * we have to check to see if the queue had anything in it and
65 * emulate it if it did. This is going to be a pain.
66 */
67
68 #include <linux/types.h>
69 #include <linux/sched.h>
70 #include <linux/mm.h>
71 #include <linux/perf_event.h>
72 #include <linux/uaccess.h>
73
74 #include "sfp-util_32.h"
75 #include <math-emu/soft-fp.h>
76 #include <math-emu/single.h>
77 #include <math-emu/double.h>
78 #include <math-emu/quad.h>
79
80 #define FLOATFUNC(x) extern int x(void *,void *,void *)
81
82 /* The Vn labels indicate what version of the SPARC architecture gas thinks
83 * each insn is. This is from the binutils source :->
84 */
85 /* quadword instructions */
86 #define FSQRTQ 0x02b /* v8 */
87 #define FADDQ 0x043 /* v8 */
88 #define FSUBQ 0x047 /* v8 */
89 #define FMULQ 0x04b /* v8 */
90 #define FDIVQ 0x04f /* v8 */
91 #define FDMULQ 0x06e /* v8 */
92 #define FQTOS 0x0c7 /* v8 */
93 #define FQTOD 0x0cb /* v8 */
94 #define FITOQ 0x0cc /* v8 */
95 #define FSTOQ 0x0cd /* v8 */
96 #define FDTOQ 0x0ce /* v8 */
97 #define FQTOI 0x0d3 /* v8 */
98 #define FCMPQ 0x053 /* v8 */
99 #define FCMPEQ 0x057 /* v8 */
100 /* single/double instructions (subnormal): should all work */
101 #define FSQRTS 0x029 /* v7 */
102 #define FSQRTD 0x02a /* v7 */
103 #define FADDS 0x041 /* v6 */
104 #define FADDD 0x042 /* v6 */
105 #define FSUBS 0x045 /* v6 */
106 #define FSUBD 0x046 /* v6 */
107 #define FMULS 0x049 /* v6 */
108 #define FMULD 0x04a /* v6 */
109 #define FDIVS 0x04d /* v6 */
110 #define FDIVD 0x04e /* v6 */
111 #define FSMULD 0x069 /* v6 */
112 #define FDTOS 0x0c6 /* v6 */
113 #define FSTOD 0x0c9 /* v6 */
114 #define FSTOI 0x0d1 /* v6 */
115 #define FDTOI 0x0d2 /* v6 */
116 #define FABSS 0x009 /* v6 */
117 #define FCMPS 0x051 /* v6 */
118 #define FCMPES 0x055 /* v6 */
119 #define FCMPD 0x052 /* v6 */
120 #define FCMPED 0x056 /* v6 */
121 #define FMOVS 0x001 /* v6 */
122 #define FNEGS 0x005 /* v6 */
123 #define FITOS 0x0c4 /* v6 */
124 #define FITOD 0x0c8 /* v6 */
125
126 #define FSR_TEM_SHIFT 23UL
127 #define FSR_TEM_MASK (0x1fUL << FSR_TEM_SHIFT)
128 #define FSR_AEXC_SHIFT 5UL
129 #define FSR_AEXC_MASK (0x1fUL << FSR_AEXC_SHIFT)
130 #define FSR_CEXC_SHIFT 0UL
131 #define FSR_CEXC_MASK (0x1fUL << FSR_CEXC_SHIFT)
132
133 static int do_one_mathemu(u32 insn, unsigned long *fsr, unsigned long *fregs);
134
135 /* Unlike the Sparc64 version (which has a struct fpustate), we
136 * pass the taskstruct corresponding to the task which currently owns the
137 * FPU. This is partly because we don't have the fpustate struct and
138 * partly because the task owning the FPU isn't always current (as is
139 * the case for the Sparc64 port). This is probably SMP-related...
140 * This function returns 1 if all queued insns were emulated successfully.
141 * The test for unimplemented FPop in kernel mode has been moved into
142 * kernel/traps.c for simplicity.
143 */
144 int do_mathemu(struct pt_regs *regs, struct task_struct *fpt)
145 {
146 /* regs->pc isn't necessarily the PC at which the offending insn is sitting.
147 * The FPU maintains a queue of FPops which cause traps.
148 * When it hits an instruction that requires that the trapped op succeeded
149 * (usually because it reads a reg. that the trapped op wrote) then it
150 * causes this exception. We need to emulate all the insns on the queue
151 * and then allow the op to proceed.
152 * This code should also handle the case where the trap was precise,
153 * in which case the queue length is zero and regs->pc points at the
154 * single FPop to be emulated. (this case is untested, though :->)
155 * You'll need this case if you want to be able to emulate all FPops
156 * because the FPU either doesn't exist or has been software-disabled.
157 * [The UltraSPARC makes FP a precise trap; this isn't as stupid as it
158 * might sound because the Ultra does funky things with a superscalar
159 * architecture.]
160 */
161
162 /* You wouldn't believe how often I typed 'ftp' when I meant 'fpt' :-> */
163
164 int i;
165 int retcode = 0; /* assume all succeed */
166 unsigned long insn;
167
168 perf_sw_event(PERF_COUNT_SW_EMULATION_FAULTS, 1, regs, 0);
169
170 #ifdef DEBUG_MATHEMU
171 printk("In do_mathemu()... pc is %08lx\n", regs->pc);
172 printk("fpqdepth is %ld\n", fpt->thread.fpqdepth);
173 for (i = 0; i < fpt->thread.fpqdepth; i++)
174 printk("%d: %08lx at %08lx\n", i, fpt->thread.fpqueue[i].insn,
175 (unsigned long)fpt->thread.fpqueue[i].insn_addr);
176 #endif
177
178 if (fpt->thread.fpqdepth == 0) { /* no queue, guilty insn is at regs->pc */
179 #ifdef DEBUG_MATHEMU
180 printk("precise trap at %08lx\n", regs->pc);
181 #endif
182 if (!get_user(insn, (u32 __user *) regs->pc)) {
183 retcode = do_one_mathemu(insn, &fpt->thread.fsr, fpt->thread.float_regs);
184 if (retcode) {
185 /* in this case we need to fix up PC & nPC */
186 regs->pc = regs->npc;
187 regs->npc += 4;
188 }
189 }
190 return retcode;
191 }
192
193 /* Normal case: need to empty the queue... */
194 for (i = 0; i < fpt->thread.fpqdepth; i++) {
195 retcode = do_one_mathemu(fpt->thread.fpqueue[i].insn, &(fpt->thread.fsr), fpt->thread.float_regs);
196 if (!retcode) /* insn failed, no point doing any more */
197 break;
198 }
199 /* Now empty the queue and clear the queue_not_empty flag */
200 if (retcode)
201 fpt->thread.fsr &= ~(0x3000 | FSR_CEXC_MASK);
202 else
203 fpt->thread.fsr &= ~0x3000;
204 fpt->thread.fpqdepth = 0;
205
206 return retcode;
207 }
208
209 /* All routines returning an exception to raise should detect
210 * such exceptions _before_ rounding to be consistent with
211 * the behavior of the hardware in the implemented cases
212 * (and thus with the recommendations in the V9 architecture
213 * manual).
214 *
215 * We return 0 if a SIGFPE should be sent, 1 otherwise.
216 */
217 static inline int record_exception(unsigned long *pfsr, int eflag)
218 {
219 unsigned long fsr = *pfsr;
220 int would_trap;
221
222 /* Determine if this exception would have generated a trap. */
223 would_trap = (fsr & ((long)eflag << FSR_TEM_SHIFT)) != 0UL;
224
225 /* If trapping, we only want to signal one bit. */
226 if (would_trap != 0) {
227 eflag &= ((fsr & FSR_TEM_MASK) >> FSR_TEM_SHIFT);
228 if ((eflag & (eflag - 1)) != 0) {
229 if (eflag & FP_EX_INVALID)
230 eflag = FP_EX_INVALID;
231 else if (eflag & FP_EX_OVERFLOW)
232 eflag = FP_EX_OVERFLOW;
233 else if (eflag & FP_EX_UNDERFLOW)
234 eflag = FP_EX_UNDERFLOW;
235 else if (eflag & FP_EX_DIVZERO)
236 eflag = FP_EX_DIVZERO;
237 else if (eflag & FP_EX_INEXACT)
238 eflag = FP_EX_INEXACT;
239 }
240 }
241
242 /* Set CEXC, here is the rule:
243 *
244 * In general all FPU ops will set one and only one
245 * bit in the CEXC field, this is always the case
246 * when the IEEE exception trap is enabled in TEM.
247 */
248 fsr &= ~(FSR_CEXC_MASK);
249 fsr |= ((long)eflag << FSR_CEXC_SHIFT);
250
251 /* Set the AEXC field, rule is:
252 *
253 * If a trap would not be generated, the
254 * CEXC just generated is OR'd into the
255 * existing value of AEXC.
256 */
257 if (would_trap == 0)
258 fsr |= ((long)eflag << FSR_AEXC_SHIFT);
259
260 /* If trapping, indicate fault trap type IEEE. */
261 if (would_trap != 0)
262 fsr |= (1UL << 14);
263
264 *pfsr = fsr;
265
266 return (would_trap ? 0 : 1);
267 }
268
269 typedef union {
270 u32 s;
271 u64 d;
272 u64 q[2];
273 } *argp;
274
275 static int do_one_mathemu(u32 insn, unsigned long *pfsr, unsigned long *fregs)
276 {
277 /* Emulate the given insn, updating fsr and fregs appropriately. */
278 int type = 0;
279 /* r is rd, b is rs2 and a is rs1. The *u arg tells
280 whether the argument should be packed/unpacked (0 - do not unpack/pack, 1 - unpack/pack)
281 non-u args tells the size of the argument (0 - no argument, 1 - single, 2 - double, 3 - quad */
282 #define TYPE(dummy, r, ru, b, bu, a, au) type = (au << 2) | (a << 0) | (bu << 5) | (b << 3) | (ru << 8) | (r << 6)
283 int freg;
284 argp rs1 = NULL, rs2 = NULL, rd = NULL;
285 FP_DECL_EX;
286 FP_DECL_S(SA); FP_DECL_S(SB); FP_DECL_S(SR);
287 FP_DECL_D(DA); FP_DECL_D(DB); FP_DECL_D(DR);
288 FP_DECL_Q(QA); FP_DECL_Q(QB); FP_DECL_Q(QR);
289 int IR;
290 long fsr;
291
292 #ifdef DEBUG_MATHEMU
293 printk("In do_mathemu(), emulating %08lx\n", insn);
294 #endif
295
296 if ((insn & 0xc1f80000) == 0x81a00000) /* FPOP1 */ {
297 switch ((insn >> 5) & 0x1ff) {
298 case FSQRTQ: TYPE(3,3,1,3,1,0,0); break;
299 case FADDQ:
300 case FSUBQ:
301 case FMULQ:
302 case FDIVQ: TYPE(3,3,1,3,1,3,1); break;
303 case FDMULQ: TYPE(3,3,1,2,1,2,1); break;
304 case FQTOS: TYPE(3,1,1,3,1,0,0); break;
305 case FQTOD: TYPE(3,2,1,3,1,0,0); break;
306 case FITOQ: TYPE(3,3,1,1,0,0,0); break;
307 case FSTOQ: TYPE(3,3,1,1,1,0,0); break;
308 case FDTOQ: TYPE(3,3,1,2,1,0,0); break;
309 case FQTOI: TYPE(3,1,0,3,1,0,0); break;
310 case FSQRTS: TYPE(2,1,1,1,1,0,0); break;
311 case FSQRTD: TYPE(2,2,1,2,1,0,0); break;
312 case FADDD:
313 case FSUBD:
314 case FMULD:
315 case FDIVD: TYPE(2,2,1,2,1,2,1); break;
316 case FADDS:
317 case FSUBS:
318 case FMULS:
319 case FDIVS: TYPE(2,1,1,1,1,1,1); break;
320 case FSMULD: TYPE(2,2,1,1,1,1,1); break;
321 case FDTOS: TYPE(2,1,1,2,1,0,0); break;
322 case FSTOD: TYPE(2,2,1,1,1,0,0); break;
323 case FSTOI: TYPE(2,1,0,1,1,0,0); break;
324 case FDTOI: TYPE(2,1,0,2,1,0,0); break;
325 case FITOS: TYPE(2,1,1,1,0,0,0); break;
326 case FITOD: TYPE(2,2,1,1,0,0,0); break;
327 case FMOVS:
328 case FABSS:
329 case FNEGS: TYPE(2,1,0,1,0,0,0); break;
330 }
331 } else if ((insn & 0xc1f80000) == 0x81a80000) /* FPOP2 */ {
332 switch ((insn >> 5) & 0x1ff) {
333 case FCMPS: TYPE(3,0,0,1,1,1,1); break;
334 case FCMPES: TYPE(3,0,0,1,1,1,1); break;
335 case FCMPD: TYPE(3,0,0,2,1,2,1); break;
336 case FCMPED: TYPE(3,0,0,2,1,2,1); break;
337 case FCMPQ: TYPE(3,0,0,3,1,3,1); break;
338 case FCMPEQ: TYPE(3,0,0,3,1,3,1); break;
339 }
340 }
341
342 if (!type) { /* oops, didn't recognise that FPop */
343 #ifdef DEBUG_MATHEMU
344 printk("attempt to emulate unrecognised FPop!\n");
345 #endif
346 return 0;
347 }
348
349 /* Decode the registers to be used */
350 freg = (*pfsr >> 14) & 0xf;
351
352 *pfsr &= ~0x1c000; /* clear the traptype bits */
353
354 freg = ((insn >> 14) & 0x1f);
355 switch (type & 0x3) { /* is rs1 single, double or quad? */
356 case 3:
357 if (freg & 3) { /* quadwords must have bits 4&5 of the */
358 /* encoded reg. number set to zero. */
359 *pfsr |= (6 << 14);
360 return 0; /* simulate invalid_fp_register exception */
361 }
362 /* fall through */
363 case 2:
364 if (freg & 1) { /* doublewords must have bit 5 zeroed */
365 *pfsr |= (6 << 14);
366 return 0;
367 }
368 }
369 rs1 = (argp)&fregs[freg];
370 switch (type & 0x7) {
371 case 7: FP_UNPACK_QP (QA, rs1); break;
372 case 6: FP_UNPACK_DP (DA, rs1); break;
373 case 5: FP_UNPACK_SP (SA, rs1); break;
374 }
375 freg = (insn & 0x1f);
376 switch ((type >> 3) & 0x3) { /* same again for rs2 */
377 case 3:
378 if (freg & 3) { /* quadwords must have bits 4&5 of the */
379 /* encoded reg. number set to zero. */
380 *pfsr |= (6 << 14);
381 return 0; /* simulate invalid_fp_register exception */
382 }
383 /* fall through */
384 case 2:
385 if (freg & 1) { /* doublewords must have bit 5 zeroed */
386 *pfsr |= (6 << 14);
387 return 0;
388 }
389 }
390 rs2 = (argp)&fregs[freg];
391 switch ((type >> 3) & 0x7) {
392 case 7: FP_UNPACK_QP (QB, rs2); break;
393 case 6: FP_UNPACK_DP (DB, rs2); break;
394 case 5: FP_UNPACK_SP (SB, rs2); break;
395 }
396 freg = ((insn >> 25) & 0x1f);
397 switch ((type >> 6) & 0x3) { /* and finally rd. This one's a bit different */
398 case 0: /* dest is fcc. (this must be FCMPQ or FCMPEQ) */
399 if (freg) { /* V8 has only one set of condition codes, so */
400 /* anything but 0 in the rd field is an error */
401 *pfsr |= (6 << 14); /* (should probably flag as invalid opcode */
402 return 0; /* but SIGFPE will do :-> ) */
403 }
404 break;
405 case 3:
406 if (freg & 3) { /* quadwords must have bits 4&5 of the */
407 /* encoded reg. number set to zero. */
408 *pfsr |= (6 << 14);
409 return 0; /* simulate invalid_fp_register exception */
410 }
411 /* fall through */
412 case 2:
413 if (freg & 1) { /* doublewords must have bit 5 zeroed */
414 *pfsr |= (6 << 14);
415 return 0;
416 }
417 /* fall through */
418 case 1:
419 rd = (void *)&fregs[freg];
420 break;
421 }
422 #ifdef DEBUG_MATHEMU
423 printk("executing insn...\n");
424 #endif
425 /* do the Right Thing */
426 switch ((insn >> 5) & 0x1ff) {
427 /* + */
428 case FADDS: FP_ADD_S (SR, SA, SB); break;
429 case FADDD: FP_ADD_D (DR, DA, DB); break;
430 case FADDQ: FP_ADD_Q (QR, QA, QB); break;
431 /* - */
432 case FSUBS: FP_SUB_S (SR, SA, SB); break;
433 case FSUBD: FP_SUB_D (DR, DA, DB); break;
434 case FSUBQ: FP_SUB_Q (QR, QA, QB); break;
435 /* * */
436 case FMULS: FP_MUL_S (SR, SA, SB); break;
437 case FSMULD: FP_CONV (D, S, 2, 1, DA, SA);
438 FP_CONV (D, S, 2, 1, DB, SB);
439 case FMULD: FP_MUL_D (DR, DA, DB); break;
440 case FDMULQ: FP_CONV (Q, D, 4, 2, QA, DA);
441 FP_CONV (Q, D, 4, 2, QB, DB);
442 case FMULQ: FP_MUL_Q (QR, QA, QB); break;
443 /* / */
444 case FDIVS: FP_DIV_S (SR, SA, SB); break;
445 case FDIVD: FP_DIV_D (DR, DA, DB); break;
446 case FDIVQ: FP_DIV_Q (QR, QA, QB); break;
447 /* sqrt */
448 case FSQRTS: FP_SQRT_S (SR, SB); break;
449 case FSQRTD: FP_SQRT_D (DR, DB); break;
450 case FSQRTQ: FP_SQRT_Q (QR, QB); break;
451 /* mov */
452 case FMOVS: rd->s = rs2->s; break;
453 case FABSS: rd->s = rs2->s & 0x7fffffff; break;
454 case FNEGS: rd->s = rs2->s ^ 0x80000000; break;
455 /* float to int */
456 case FSTOI: FP_TO_INT_S (IR, SB, 32, 1); break;
457 case FDTOI: FP_TO_INT_D (IR, DB, 32, 1); break;
458 case FQTOI: FP_TO_INT_Q (IR, QB, 32, 1); break;
459 /* int to float */
460 case FITOS: IR = rs2->s; FP_FROM_INT_S (SR, IR, 32, int); break;
461 case FITOD: IR = rs2->s; FP_FROM_INT_D (DR, IR, 32, int); break;
462 case FITOQ: IR = rs2->s; FP_FROM_INT_Q (QR, IR, 32, int); break;
463 /* float to float */
464 case FSTOD: FP_CONV (D, S, 2, 1, DR, SB); break;
465 case FSTOQ: FP_CONV (Q, S, 4, 1, QR, SB); break;
466 case FDTOQ: FP_CONV (Q, D, 4, 2, QR, DB); break;
467 case FDTOS: FP_CONV (S, D, 1, 2, SR, DB); break;
468 case FQTOS: FP_CONV (S, Q, 1, 4, SR, QB); break;
469 case FQTOD: FP_CONV (D, Q, 2, 4, DR, QB); break;
470 /* comparison */
471 case FCMPS:
472 case FCMPES:
473 FP_CMP_S(IR, SB, SA, 3);
474 if (IR == 3 &&
475 (((insn >> 5) & 0x1ff) == FCMPES ||
476 FP_ISSIGNAN_S(SA) ||
477 FP_ISSIGNAN_S(SB)))
478 FP_SET_EXCEPTION (FP_EX_INVALID);
479 break;
480 case FCMPD:
481 case FCMPED:
482 FP_CMP_D(IR, DB, DA, 3);
483 if (IR == 3 &&
484 (((insn >> 5) & 0x1ff) == FCMPED ||
485 FP_ISSIGNAN_D(DA) ||
486 FP_ISSIGNAN_D(DB)))
487 FP_SET_EXCEPTION (FP_EX_INVALID);
488 break;
489 case FCMPQ:
490 case FCMPEQ:
491 FP_CMP_Q(IR, QB, QA, 3);
492 if (IR == 3 &&
493 (((insn >> 5) & 0x1ff) == FCMPEQ ||
494 FP_ISSIGNAN_Q(QA) ||
495 FP_ISSIGNAN_Q(QB)))
496 FP_SET_EXCEPTION (FP_EX_INVALID);
497 }
498 if (!FP_INHIBIT_RESULTS) {
499 switch ((type >> 6) & 0x7) {
500 case 0: fsr = *pfsr;
501 if (IR == -1) IR = 2;
502 /* fcc is always fcc0 */
503 fsr &= ~0xc00; fsr |= (IR << 10);
504 *pfsr = fsr;
505 break;
506 case 1: rd->s = IR; break;
507 case 5: FP_PACK_SP (rd, SR); break;
508 case 6: FP_PACK_DP (rd, DR); break;
509 case 7: FP_PACK_QP (rd, QR); break;
510 }
511 }
512 if (_fex == 0)
513 return 1; /* success! */
514 return record_exception(pfsr, _fex);
515 }