blob: 56d2c44747b8fcf3fd69324319b36ebd4b7ebce7 [file] [log] [blame]
Adrian Bunkb00dc832008-05-19 16:52:27 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * arch/sparc64/math-emu/math.c
3 *
4 * Copyright (C) 1997,1999 Jakub Jelinek (jj@ultra.linux.cz)
5 * Copyright (C) 1999 David S. Miller (davem@redhat.com)
6 *
7 * Emulation routines originate from soft-fp package, which is part
8 * of glibc and has appropriate copyrights in it.
9 */
10
11#include <linux/types.h>
12#include <linux/sched.h>
13#include <linux/errno.h>
David S. Miller121dd5f2009-12-11 01:07:53 -080014#include <linux/perf_event.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070015
16#include <asm/fpumacro.h>
17#include <asm/ptrace.h>
18#include <asm/uaccess.h>
19
Sam Ravnborg5115f392008-11-16 20:07:11 -080020#include "sfp-util_64.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070021#include <math-emu/soft-fp.h>
22#include <math-emu/single.h>
23#include <math-emu/double.h>
24#include <math-emu/quad.h>
25
26/* QUAD - ftt == 3 */
27#define FMOVQ 0x003
28#define FNEGQ 0x007
29#define FABSQ 0x00b
30#define FSQRTQ 0x02b
31#define FADDQ 0x043
32#define FSUBQ 0x047
33#define FMULQ 0x04b
34#define FDIVQ 0x04f
35#define FDMULQ 0x06e
36#define FQTOX 0x083
37#define FXTOQ 0x08c
38#define FQTOS 0x0c7
39#define FQTOD 0x0cb
40#define FITOQ 0x0cc
41#define FSTOQ 0x0cd
42#define FDTOQ 0x0ce
43#define FQTOI 0x0d3
44/* SUBNORMAL - ftt == 2 */
45#define FSQRTS 0x029
46#define FSQRTD 0x02a
47#define FADDS 0x041
48#define FADDD 0x042
49#define FSUBS 0x045
50#define FSUBD 0x046
51#define FMULS 0x049
52#define FMULD 0x04a
53#define FDIVS 0x04d
54#define FDIVD 0x04e
55#define FSMULD 0x069
56#define FSTOX 0x081
57#define FDTOX 0x082
58#define FDTOS 0x0c6
59#define FSTOD 0x0c9
60#define FSTOI 0x0d1
61#define FDTOI 0x0d2
62#define FXTOS 0x084 /* Only Ultra-III generates this. */
63#define FXTOD 0x088 /* Only Ultra-III generates this. */
64#if 0 /* Optimized inline in sparc64/kernel/entry.S */
65#define FITOS 0x0c4 /* Only Ultra-III generates this. */
66#endif
67#define FITOD 0x0c8 /* Only Ultra-III generates this. */
68/* FPOP2 */
69#define FCMPQ 0x053
70#define FCMPEQ 0x057
71#define FMOVQ0 0x003
72#define FMOVQ1 0x043
73#define FMOVQ2 0x083
74#define FMOVQ3 0x0c3
75#define FMOVQI 0x103
76#define FMOVQX 0x183
77#define FMOVQZ 0x027
78#define FMOVQLE 0x047
79#define FMOVQLZ 0x067
80#define FMOVQNZ 0x0a7
81#define FMOVQGZ 0x0c7
82#define FMOVQGE 0x0e7
83
84#define FSR_TEM_SHIFT 23UL
85#define FSR_TEM_MASK (0x1fUL << FSR_TEM_SHIFT)
86#define FSR_AEXC_SHIFT 5UL
87#define FSR_AEXC_MASK (0x1fUL << FSR_AEXC_SHIFT)
88#define FSR_CEXC_SHIFT 0UL
89#define FSR_CEXC_MASK (0x1fUL << FSR_CEXC_SHIFT)
90
91/* All routines returning an exception to raise should detect
92 * such exceptions _before_ rounding to be consistent with
93 * the behavior of the hardware in the implemented cases
94 * (and thus with the recommendations in the V9 architecture
95 * manual).
96 *
97 * We return 0 if a SIGFPE should be sent, 1 otherwise.
98 */
99static inline int record_exception(struct pt_regs *regs, int eflag)
100{
101 u64 fsr = current_thread_info()->xfsr[0];
102 int would_trap;
103
104 /* Determine if this exception would have generated a trap. */
105 would_trap = (fsr & ((long)eflag << FSR_TEM_SHIFT)) != 0UL;
106
107 /* If trapping, we only want to signal one bit. */
108 if(would_trap != 0) {
109 eflag &= ((fsr & FSR_TEM_MASK) >> FSR_TEM_SHIFT);
110 if((eflag & (eflag - 1)) != 0) {
111 if(eflag & FP_EX_INVALID)
112 eflag = FP_EX_INVALID;
113 else if(eflag & FP_EX_OVERFLOW)
114 eflag = FP_EX_OVERFLOW;
115 else if(eflag & FP_EX_UNDERFLOW)
116 eflag = FP_EX_UNDERFLOW;
117 else if(eflag & FP_EX_DIVZERO)
118 eflag = FP_EX_DIVZERO;
119 else if(eflag & FP_EX_INEXACT)
120 eflag = FP_EX_INEXACT;
121 }
122 }
123
124 /* Set CEXC, here is the rule:
125 *
126 * In general all FPU ops will set one and only one
127 * bit in the CEXC field, this is always the case
128 * when the IEEE exception trap is enabled in TEM.
129 */
130 fsr &= ~(FSR_CEXC_MASK);
131 fsr |= ((long)eflag << FSR_CEXC_SHIFT);
132
133 /* Set the AEXC field, rule is:
134 *
135 * If a trap would not be generated, the
136 * CEXC just generated is OR'd into the
137 * existing value of AEXC.
138 */
139 if(would_trap == 0)
140 fsr |= ((long)eflag << FSR_AEXC_SHIFT);
141
142 /* If trapping, indicate fault trap type IEEE. */
143 if(would_trap != 0)
144 fsr |= (1UL << 14);
145
146 current_thread_info()->xfsr[0] = fsr;
147
148 /* If we will not trap, advance the program counter over
149 * the instruction being handled.
150 */
151 if(would_trap == 0) {
152 regs->tpc = regs->tnpc;
153 regs->tnpc += 4;
154 }
155
156 return (would_trap ? 0 : 1);
157}
158
159typedef union {
160 u32 s;
161 u64 d;
162 u64 q[2];
163} *argp;
164
165int do_mathemu(struct pt_regs *regs, struct fpustate *f)
166{
167 unsigned long pc = regs->tpc;
168 unsigned long tstate = regs->tstate;
169 u32 insn = 0;
170 int type = 0;
171 /* ftt tells which ftt it may happen in, r is rd, b is rs2 and a is rs1. The *u arg tells
172 whether the argument should be packed/unpacked (0 - do not unpack/pack, 1 - unpack/pack)
173 non-u args tells the size of the argument (0 - no argument, 1 - single, 2 - double, 3 - quad */
174#define TYPE(ftt, r, ru, b, bu, a, au) type = (au << 2) | (a << 0) | (bu << 5) | (b << 3) | (ru << 8) | (r << 6) | (ftt << 9)
175 int freg;
176 static u64 zero[2] = { 0L, 0L };
177 int flags;
178 FP_DECL_EX;
179 FP_DECL_S(SA); FP_DECL_S(SB); FP_DECL_S(SR);
180 FP_DECL_D(DA); FP_DECL_D(DB); FP_DECL_D(DR);
181 FP_DECL_Q(QA); FP_DECL_Q(QB); FP_DECL_Q(QR);
182 int IR;
183 long XR, xfsr;
184
185 if (tstate & TSTATE_PRIV)
186 die_if_kernel("unfinished/unimplemented FPop from kernel", regs);
David S. Miller121dd5f2009-12-11 01:07:53 -0800187 perf_sw_event(PERF_COUNT_SW_EMULATION_FAULTS, 1, 0, regs, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188 if (test_thread_flag(TIF_32BIT))
189 pc = (u32)pc;
190 if (get_user(insn, (u32 __user *) pc) != -EFAULT) {
191 if ((insn & 0xc1f80000) == 0x81a00000) /* FPOP1 */ {
192 switch ((insn >> 5) & 0x1ff) {
193 /* QUAD - ftt == 3 */
194 case FMOVQ:
195 case FNEGQ:
196 case FABSQ: TYPE(3,3,0,3,0,0,0); break;
197 case FSQRTQ: TYPE(3,3,1,3,1,0,0); break;
198 case FADDQ:
199 case FSUBQ:
200 case FMULQ:
201 case FDIVQ: TYPE(3,3,1,3,1,3,1); break;
202 case FDMULQ: TYPE(3,3,1,2,1,2,1); break;
203 case FQTOX: TYPE(3,2,0,3,1,0,0); break;
204 case FXTOQ: TYPE(3,3,1,2,0,0,0); break;
205 case FQTOS: TYPE(3,1,1,3,1,0,0); break;
206 case FQTOD: TYPE(3,2,1,3,1,0,0); break;
207 case FITOQ: TYPE(3,3,1,1,0,0,0); break;
208 case FSTOQ: TYPE(3,3,1,1,1,0,0); break;
209 case FDTOQ: TYPE(3,3,1,2,1,0,0); break;
210 case FQTOI: TYPE(3,1,0,3,1,0,0); break;
David S. Miller4e74ae82006-02-20 16:02:24 -0800211
212 /* We can get either unimplemented or unfinished
213 * for these cases. Pre-Niagara systems generate
214 * unfinished fpop for SUBNORMAL cases, and Niagara
215 * always gives unimplemented fpop for fsqrt{s,d}.
216 */
217 case FSQRTS: {
218 unsigned long x = current_thread_info()->xfsr[0];
219
220 x = (x >> 14) & 0xf;
221 TYPE(x,1,1,1,1,0,0);
David S. Miller4e74ae82006-02-20 16:02:24 -0800222 break;
223 }
224
225 case FSQRTD: {
226 unsigned long x = current_thread_info()->xfsr[0];
227
228 x = (x >> 14) & 0xf;
229 TYPE(x,2,1,2,1,0,0);
230 break;
231 }
232
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 /* SUBNORMAL - ftt == 2 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 case FADDD:
235 case FSUBD:
236 case FMULD:
237 case FDIVD: TYPE(2,2,1,2,1,2,1); break;
238 case FADDS:
239 case FSUBS:
240 case FMULS:
241 case FDIVS: TYPE(2,1,1,1,1,1,1); break;
242 case FSMULD: TYPE(2,2,1,1,1,1,1); break;
243 case FSTOX: TYPE(2,2,0,1,1,0,0); break;
244 case FDTOX: TYPE(2,2,0,2,1,0,0); break;
245 case FDTOS: TYPE(2,1,1,2,1,0,0); break;
246 case FSTOD: TYPE(2,2,1,1,1,0,0); break;
247 case FSTOI: TYPE(2,1,0,1,1,0,0); break;
248 case FDTOI: TYPE(2,1,0,2,1,0,0); break;
249
250 /* Only Ultra-III generates these */
251 case FXTOS: TYPE(2,1,1,2,0,0,0); break;
252 case FXTOD: TYPE(2,2,1,2,0,0,0); break;
253#if 0 /* Optimized inline in sparc64/kernel/entry.S */
254 case FITOS: TYPE(2,1,1,1,0,0,0); break;
255#endif
256 case FITOD: TYPE(2,2,1,1,0,0,0); break;
257 }
258 }
259 else if ((insn & 0xc1f80000) == 0x81a80000) /* FPOP2 */ {
260 IR = 2;
261 switch ((insn >> 5) & 0x1ff) {
262 case FCMPQ: TYPE(3,0,0,3,1,3,1); break;
263 case FCMPEQ: TYPE(3,0,0,3,1,3,1); break;
264 /* Now the conditional fmovq support */
265 case FMOVQ0:
266 case FMOVQ1:
267 case FMOVQ2:
268 case FMOVQ3:
269 /* fmovq %fccX, %fY, %fZ */
270 if (!((insn >> 11) & 3))
271 XR = current_thread_info()->xfsr[0] >> 10;
272 else
273 XR = current_thread_info()->xfsr[0] >> (30 + ((insn >> 10) & 0x6));
274 XR &= 3;
275 IR = 0;
276 switch ((insn >> 14) & 0x7) {
277 /* case 0: IR = 0; break; */ /* Never */
278 case 1: if (XR) IR = 1; break; /* Not Equal */
279 case 2: if (XR == 1 || XR == 2) IR = 1; break; /* Less or Greater */
280 case 3: if (XR & 1) IR = 1; break; /* Unordered or Less */
281 case 4: if (XR == 1) IR = 1; break; /* Less */
282 case 5: if (XR & 2) IR = 1; break; /* Unordered or Greater */
283 case 6: if (XR == 2) IR = 1; break; /* Greater */
284 case 7: if (XR == 3) IR = 1; break; /* Unordered */
285 }
286 if ((insn >> 14) & 8)
287 IR ^= 1;
288 break;
289 case FMOVQI:
290 case FMOVQX:
291 /* fmovq %[ix]cc, %fY, %fZ */
292 XR = regs->tstate >> 32;
293 if ((insn >> 5) & 0x80)
294 XR >>= 4;
295 XR &= 0xf;
296 IR = 0;
297 freg = ((XR >> 2) ^ XR) & 2;
298 switch ((insn >> 14) & 0x7) {
299 /* case 0: IR = 0; break; */ /* Never */
300 case 1: if (XR & 4) IR = 1; break; /* Equal */
301 case 2: if ((XR & 4) || freg) IR = 1; break; /* Less or Equal */
302 case 3: if (freg) IR = 1; break; /* Less */
303 case 4: if (XR & 5) IR = 1; break; /* Less or Equal Unsigned */
304 case 5: if (XR & 1) IR = 1; break; /* Carry Set */
305 case 6: if (XR & 8) IR = 1; break; /* Negative */
306 case 7: if (XR & 2) IR = 1; break; /* Overflow Set */
307 }
308 if ((insn >> 14) & 8)
309 IR ^= 1;
310 break;
311 case FMOVQZ:
312 case FMOVQLE:
313 case FMOVQLZ:
314 case FMOVQNZ:
315 case FMOVQGZ:
316 case FMOVQGE:
317 freg = (insn >> 14) & 0x1f;
318 if (!freg)
319 XR = 0;
320 else if (freg < 16)
321 XR = regs->u_regs[freg];
322 else if (test_thread_flag(TIF_32BIT)) {
323 struct reg_window32 __user *win32;
324 flushw_user ();
325 win32 = (struct reg_window32 __user *)((unsigned long)((u32)regs->u_regs[UREG_FP]));
326 get_user(XR, &win32->locals[freg - 16]);
327 } else {
328 struct reg_window __user *win;
329 flushw_user ();
330 win = (struct reg_window __user *)(regs->u_regs[UREG_FP] + STACK_BIAS);
331 get_user(XR, &win->locals[freg - 16]);
332 }
333 IR = 0;
334 switch ((insn >> 10) & 3) {
335 case 1: if (!XR) IR = 1; break; /* Register Zero */
336 case 2: if (XR <= 0) IR = 1; break; /* Register Less Than or Equal to Zero */
337 case 3: if (XR < 0) IR = 1; break; /* Register Less Than Zero */
338 }
339 if ((insn >> 10) & 4)
340 IR ^= 1;
341 break;
342 }
343 if (IR == 0) {
344 /* The fmov test was false. Do a nop instead */
345 current_thread_info()->xfsr[0] &= ~(FSR_CEXC_MASK);
346 regs->tpc = regs->tnpc;
347 regs->tnpc += 4;
348 return 1;
349 } else if (IR == 1) {
350 /* Change the instruction into plain fmovq */
351 insn = (insn & 0x3e00001f) | 0x81a00060;
352 TYPE(3,3,0,3,0,0,0);
353 }
354 }
355 }
356 if (type) {
357 argp rs1 = NULL, rs2 = NULL, rd = NULL;
358
359 freg = (current_thread_info()->xfsr[0] >> 14) & 0xf;
360 if (freg != (type >> 9))
361 goto err;
362 current_thread_info()->xfsr[0] &= ~0x1c000;
363 freg = ((insn >> 14) & 0x1f);
364 switch (type & 0x3) {
365 case 3: if (freg & 2) {
366 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */;
367 goto err;
368 }
369 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e);
370 case 1: rs1 = (argp)&f->regs[freg];
371 flags = (freg < 32) ? FPRS_DL : FPRS_DU;
372 if (!(current_thread_info()->fpsaved[0] & flags))
373 rs1 = (argp)&zero;
374 break;
375 }
376 switch (type & 0x7) {
377 case 7: FP_UNPACK_QP (QA, rs1); break;
378 case 6: FP_UNPACK_DP (DA, rs1); break;
379 case 5: FP_UNPACK_SP (SA, rs1); break;
380 }
381 freg = (insn & 0x1f);
382 switch ((type >> 3) & 0x3) {
383 case 3: if (freg & 2) {
384 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */;
385 goto err;
386 }
387 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e);
388 case 1: rs2 = (argp)&f->regs[freg];
389 flags = (freg < 32) ? FPRS_DL : FPRS_DU;
390 if (!(current_thread_info()->fpsaved[0] & flags))
391 rs2 = (argp)&zero;
392 break;
393 }
394 switch ((type >> 3) & 0x7) {
395 case 7: FP_UNPACK_QP (QB, rs2); break;
396 case 6: FP_UNPACK_DP (DB, rs2); break;
397 case 5: FP_UNPACK_SP (SB, rs2); break;
398 }
399 freg = ((insn >> 25) & 0x1f);
400 switch ((type >> 6) & 0x3) {
401 case 3: if (freg & 2) {
402 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */;
403 goto err;
404 }
405 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e);
406 case 1: rd = (argp)&f->regs[freg];
407 flags = (freg < 32) ? FPRS_DL : FPRS_DU;
408 if (!(current_thread_info()->fpsaved[0] & FPRS_FEF)) {
409 current_thread_info()->fpsaved[0] = FPRS_FEF;
410 current_thread_info()->gsr[0] = 0;
411 }
412 if (!(current_thread_info()->fpsaved[0] & flags)) {
413 if (freg < 32)
414 memset(f->regs, 0, 32*sizeof(u32));
415 else
416 memset(f->regs+32, 0, 32*sizeof(u32));
417 }
418 current_thread_info()->fpsaved[0] |= flags;
419 break;
420 }
421 switch ((insn >> 5) & 0x1ff) {
422 /* + */
423 case FADDS: FP_ADD_S (SR, SA, SB); break;
424 case FADDD: FP_ADD_D (DR, DA, DB); break;
425 case FADDQ: FP_ADD_Q (QR, QA, QB); break;
426 /* - */
427 case FSUBS: FP_SUB_S (SR, SA, SB); break;
428 case FSUBD: FP_SUB_D (DR, DA, DB); break;
429 case FSUBQ: FP_SUB_Q (QR, QA, QB); break;
430 /* * */
431 case FMULS: FP_MUL_S (SR, SA, SB); break;
432 case FSMULD: FP_CONV (D, S, 1, 1, DA, SA);
433 FP_CONV (D, S, 1, 1, DB, SB);
434 case FMULD: FP_MUL_D (DR, DA, DB); break;
435 case FDMULQ: FP_CONV (Q, D, 2, 1, QA, DA);
436 FP_CONV (Q, D, 2, 1, QB, DB);
437 case FMULQ: FP_MUL_Q (QR, QA, QB); break;
438 /* / */
439 case FDIVS: FP_DIV_S (SR, SA, SB); break;
440 case FDIVD: FP_DIV_D (DR, DA, DB); break;
441 case FDIVQ: FP_DIV_Q (QR, QA, QB); break;
442 /* sqrt */
443 case FSQRTS: FP_SQRT_S (SR, SB); break;
444 case FSQRTD: FP_SQRT_D (DR, DB); break;
445 case FSQRTQ: FP_SQRT_Q (QR, QB); break;
446 /* mov */
447 case FMOVQ: rd->q[0] = rs2->q[0]; rd->q[1] = rs2->q[1]; break;
448 case FABSQ: rd->q[0] = rs2->q[0] & 0x7fffffffffffffffUL; rd->q[1] = rs2->q[1]; break;
449 case FNEGQ: rd->q[0] = rs2->q[0] ^ 0x8000000000000000UL; rd->q[1] = rs2->q[1]; break;
450 /* float to int */
451 case FSTOI: FP_TO_INT_S (IR, SB, 32, 1); break;
452 case FDTOI: FP_TO_INT_D (IR, DB, 32, 1); break;
453 case FQTOI: FP_TO_INT_Q (IR, QB, 32, 1); break;
454 case FSTOX: FP_TO_INT_S (XR, SB, 64, 1); break;
455 case FDTOX: FP_TO_INT_D (XR, DB, 64, 1); break;
456 case FQTOX: FP_TO_INT_Q (XR, QB, 64, 1); break;
457 /* int to float */
458 case FITOQ: IR = rs2->s; FP_FROM_INT_Q (QR, IR, 32, int); break;
459 case FXTOQ: XR = rs2->d; FP_FROM_INT_Q (QR, XR, 64, long); break;
460 /* Only Ultra-III generates these */
461 case FXTOS: XR = rs2->d; FP_FROM_INT_S (SR, XR, 64, long); break;
462 case FXTOD: XR = rs2->d; FP_FROM_INT_D (DR, XR, 64, long); break;
463#if 0 /* Optimized inline in sparc64/kernel/entry.S */
464 case FITOS: IR = rs2->s; FP_FROM_INT_S (SR, IR, 32, int); break;
465#endif
466 case FITOD: IR = rs2->s; FP_FROM_INT_D (DR, IR, 32, int); break;
467 /* float to float */
468 case FSTOD: FP_CONV (D, S, 1, 1, DR, SB); break;
469 case FSTOQ: FP_CONV (Q, S, 2, 1, QR, SB); break;
470 case FDTOQ: FP_CONV (Q, D, 2, 1, QR, DB); break;
471 case FDTOS: FP_CONV (S, D, 1, 1, SR, DB); break;
472 case FQTOS: FP_CONV (S, Q, 1, 2, SR, QB); break;
473 case FQTOD: FP_CONV (D, Q, 1, 2, DR, QB); break;
474 /* comparison */
475 case FCMPQ:
476 case FCMPEQ:
477 FP_CMP_Q(XR, QB, QA, 3);
478 if (XR == 3 &&
479 (((insn >> 5) & 0x1ff) == FCMPEQ ||
480 FP_ISSIGNAN_Q(QA) ||
481 FP_ISSIGNAN_Q(QB)))
482 FP_SET_EXCEPTION (FP_EX_INVALID);
483 }
484 if (!FP_INHIBIT_RESULTS) {
485 switch ((type >> 6) & 0x7) {
486 case 0: xfsr = current_thread_info()->xfsr[0];
487 if (XR == -1) XR = 2;
488 switch (freg & 3) {
489 /* fcc0, 1, 2, 3 */
490 case 0: xfsr &= ~0xc00; xfsr |= (XR << 10); break;
491 case 1: xfsr &= ~0x300000000UL; xfsr |= (XR << 32); break;
492 case 2: xfsr &= ~0xc00000000UL; xfsr |= (XR << 34); break;
493 case 3: xfsr &= ~0x3000000000UL; xfsr |= (XR << 36); break;
494 }
495 current_thread_info()->xfsr[0] = xfsr;
496 break;
497 case 1: rd->s = IR; break;
498 case 2: rd->d = XR; break;
499 case 5: FP_PACK_SP (rd, SR); break;
500 case 6: FP_PACK_DP (rd, DR); break;
501 case 7: FP_PACK_QP (rd, QR); break;
502 }
503 }
504
505 if(_fex != 0)
506 return record_exception(regs, _fex);
507
508 /* Success and no exceptions detected. */
509 current_thread_info()->xfsr[0] &= ~(FSR_CEXC_MASK);
510 regs->tpc = regs->tnpc;
511 regs->tnpc += 4;
512 return 1;
513 }
514err: return 0;
515}