Linux 2.6.13-rc4
[linux-2.6/next.git] / arch / sparc64 / math-emu / math.c
blob2ae05cd7b773fd2a45b3343968cd5ee072d5ea8e
1 /* $Id: math.c,v 1.11 1999/12/20 05:02:25 davem Exp $
2 * arch/sparc64/math-emu/math.c
4 * Copyright (C) 1997,1999 Jakub Jelinek (jj@ultra.linux.cz)
5 * Copyright (C) 1999 David S. Miller (davem@redhat.com)
7 * Emulation routines originate from soft-fp package, which is part
8 * of glibc and has appropriate copyrights in it.
9 */
11 #include <linux/types.h>
12 #include <linux/sched.h>
13 #include <linux/errno.h>
15 #include <asm/fpumacro.h>
16 #include <asm/ptrace.h>
17 #include <asm/uaccess.h>
19 #include "sfp-util.h"
20 #include <math-emu/soft-fp.h>
21 #include <math-emu/single.h>
22 #include <math-emu/double.h>
23 #include <math-emu/quad.h>
25 /* QUAD - ftt == 3 */
26 #define FMOVQ 0x003
27 #define FNEGQ 0x007
28 #define FABSQ 0x00b
29 #define FSQRTQ 0x02b
30 #define FADDQ 0x043
31 #define FSUBQ 0x047
32 #define FMULQ 0x04b
33 #define FDIVQ 0x04f
34 #define FDMULQ 0x06e
35 #define FQTOX 0x083
36 #define FXTOQ 0x08c
37 #define FQTOS 0x0c7
38 #define FQTOD 0x0cb
39 #define FITOQ 0x0cc
40 #define FSTOQ 0x0cd
41 #define FDTOQ 0x0ce
42 #define FQTOI 0x0d3
43 /* SUBNORMAL - ftt == 2 */
44 #define FSQRTS 0x029
45 #define FSQRTD 0x02a
46 #define FADDS 0x041
47 #define FADDD 0x042
48 #define FSUBS 0x045
49 #define FSUBD 0x046
50 #define FMULS 0x049
51 #define FMULD 0x04a
52 #define FDIVS 0x04d
53 #define FDIVD 0x04e
54 #define FSMULD 0x069
55 #define FSTOX 0x081
56 #define FDTOX 0x082
57 #define FDTOS 0x0c6
58 #define FSTOD 0x0c9
59 #define FSTOI 0x0d1
60 #define FDTOI 0x0d2
61 #define FXTOS 0x084 /* Only Ultra-III generates this. */
62 #define FXTOD 0x088 /* Only Ultra-III generates this. */
63 #if 0 /* Optimized inline in sparc64/kernel/entry.S */
64 #define FITOS 0x0c4 /* Only Ultra-III generates this. */
65 #endif
66 #define FITOD 0x0c8 /* Only Ultra-III generates this. */
67 /* FPOP2 */
68 #define FCMPQ 0x053
69 #define FCMPEQ 0x057
70 #define FMOVQ0 0x003
71 #define FMOVQ1 0x043
72 #define FMOVQ2 0x083
73 #define FMOVQ3 0x0c3
74 #define FMOVQI 0x103
75 #define FMOVQX 0x183
76 #define FMOVQZ 0x027
77 #define FMOVQLE 0x047
78 #define FMOVQLZ 0x067
79 #define FMOVQNZ 0x0a7
80 #define FMOVQGZ 0x0c7
81 #define FMOVQGE 0x0e7
83 #define FSR_TEM_SHIFT 23UL
84 #define FSR_TEM_MASK (0x1fUL << FSR_TEM_SHIFT)
85 #define FSR_AEXC_SHIFT 5UL
86 #define FSR_AEXC_MASK (0x1fUL << FSR_AEXC_SHIFT)
87 #define FSR_CEXC_SHIFT 0UL
88 #define FSR_CEXC_MASK (0x1fUL << FSR_CEXC_SHIFT)
90 /* All routines returning an exception to raise should detect
91 * such exceptions _before_ rounding to be consistent with
92 * the behavior of the hardware in the implemented cases
93 * (and thus with the recommendations in the V9 architecture
94 * manual).
96 * We return 0 if a SIGFPE should be sent, 1 otherwise.
98 static inline int record_exception(struct pt_regs *regs, int eflag)
100 u64 fsr = current_thread_info()->xfsr[0];
101 int would_trap;
103 /* Determine if this exception would have generated a trap. */
104 would_trap = (fsr & ((long)eflag << FSR_TEM_SHIFT)) != 0UL;
106 /* If trapping, we only want to signal one bit. */
107 if(would_trap != 0) {
108 eflag &= ((fsr & FSR_TEM_MASK) >> FSR_TEM_SHIFT);
109 if((eflag & (eflag - 1)) != 0) {
110 if(eflag & FP_EX_INVALID)
111 eflag = FP_EX_INVALID;
112 else if(eflag & FP_EX_OVERFLOW)
113 eflag = FP_EX_OVERFLOW;
114 else if(eflag & FP_EX_UNDERFLOW)
115 eflag = FP_EX_UNDERFLOW;
116 else if(eflag & FP_EX_DIVZERO)
117 eflag = FP_EX_DIVZERO;
118 else if(eflag & FP_EX_INEXACT)
119 eflag = FP_EX_INEXACT;
123 /* Set CEXC, here is the rule:
125 * In general all FPU ops will set one and only one
126 * bit in the CEXC field, this is always the case
127 * when the IEEE exception trap is enabled in TEM.
129 fsr &= ~(FSR_CEXC_MASK);
130 fsr |= ((long)eflag << FSR_CEXC_SHIFT);
132 /* Set the AEXC field, rule is:
134 * If a trap would not be generated, the
135 * CEXC just generated is OR'd into the
136 * existing value of AEXC.
138 if(would_trap == 0)
139 fsr |= ((long)eflag << FSR_AEXC_SHIFT);
141 /* If trapping, indicate fault trap type IEEE. */
142 if(would_trap != 0)
143 fsr |= (1UL << 14);
145 current_thread_info()->xfsr[0] = fsr;
147 /* If we will not trap, advance the program counter over
148 * the instruction being handled.
150 if(would_trap == 0) {
151 regs->tpc = regs->tnpc;
152 regs->tnpc += 4;
155 return (would_trap ? 0 : 1);
158 typedef union {
159 u32 s;
160 u64 d;
161 u64 q[2];
162 } *argp;
164 int do_mathemu(struct pt_regs *regs, struct fpustate *f)
166 unsigned long pc = regs->tpc;
167 unsigned long tstate = regs->tstate;
168 u32 insn = 0;
169 int type = 0;
170 /* ftt tells which ftt it may happen in, r is rd, b is rs2 and a is rs1. The *u arg tells
171 whether the argument should be packed/unpacked (0 - do not unpack/pack, 1 - unpack/pack)
172 non-u args tells the size of the argument (0 - no argument, 1 - single, 2 - double, 3 - quad */
173 #define TYPE(ftt, r, ru, b, bu, a, au) type = (au << 2) | (a << 0) | (bu << 5) | (b << 3) | (ru << 8) | (r << 6) | (ftt << 9)
174 int freg;
175 static u64 zero[2] = { 0L, 0L };
176 int flags;
177 FP_DECL_EX;
178 FP_DECL_S(SA); FP_DECL_S(SB); FP_DECL_S(SR);
179 FP_DECL_D(DA); FP_DECL_D(DB); FP_DECL_D(DR);
180 FP_DECL_Q(QA); FP_DECL_Q(QB); FP_DECL_Q(QR);
181 int IR;
182 long XR, xfsr;
184 if (tstate & TSTATE_PRIV)
185 die_if_kernel("unfinished/unimplemented FPop from kernel", regs);
186 if (test_thread_flag(TIF_32BIT))
187 pc = (u32)pc;
188 if (get_user(insn, (u32 __user *) pc) != -EFAULT) {
189 if ((insn & 0xc1f80000) == 0x81a00000) /* FPOP1 */ {
190 switch ((insn >> 5) & 0x1ff) {
191 /* QUAD - ftt == 3 */
192 case FMOVQ:
193 case FNEGQ:
194 case FABSQ: TYPE(3,3,0,3,0,0,0); break;
195 case FSQRTQ: TYPE(3,3,1,3,1,0,0); break;
196 case FADDQ:
197 case FSUBQ:
198 case FMULQ:
199 case FDIVQ: TYPE(3,3,1,3,1,3,1); break;
200 case FDMULQ: TYPE(3,3,1,2,1,2,1); break;
201 case FQTOX: TYPE(3,2,0,3,1,0,0); break;
202 case FXTOQ: TYPE(3,3,1,2,0,0,0); break;
203 case FQTOS: TYPE(3,1,1,3,1,0,0); break;
204 case FQTOD: TYPE(3,2,1,3,1,0,0); break;
205 case FITOQ: TYPE(3,3,1,1,0,0,0); break;
206 case FSTOQ: TYPE(3,3,1,1,1,0,0); break;
207 case FDTOQ: TYPE(3,3,1,2,1,0,0); break;
208 case FQTOI: TYPE(3,1,0,3,1,0,0); break;
209 /* SUBNORMAL - ftt == 2 */
210 case FSQRTS: TYPE(2,1,1,1,1,0,0); break;
211 case FSQRTD: TYPE(2,2,1,2,1,0,0); break;
212 case FADDD:
213 case FSUBD:
214 case FMULD:
215 case FDIVD: TYPE(2,2,1,2,1,2,1); break;
216 case FADDS:
217 case FSUBS:
218 case FMULS:
219 case FDIVS: TYPE(2,1,1,1,1,1,1); break;
220 case FSMULD: TYPE(2,2,1,1,1,1,1); break;
221 case FSTOX: TYPE(2,2,0,1,1,0,0); break;
222 case FDTOX: TYPE(2,2,0,2,1,0,0); break;
223 case FDTOS: TYPE(2,1,1,2,1,0,0); break;
224 case FSTOD: TYPE(2,2,1,1,1,0,0); break;
225 case FSTOI: TYPE(2,1,0,1,1,0,0); break;
226 case FDTOI: TYPE(2,1,0,2,1,0,0); break;
228 /* Only Ultra-III generates these */
229 case FXTOS: TYPE(2,1,1,2,0,0,0); break;
230 case FXTOD: TYPE(2,2,1,2,0,0,0); break;
231 #if 0 /* Optimized inline in sparc64/kernel/entry.S */
232 case FITOS: TYPE(2,1,1,1,0,0,0); break;
233 #endif
234 case FITOD: TYPE(2,2,1,1,0,0,0); break;
237 else if ((insn & 0xc1f80000) == 0x81a80000) /* FPOP2 */ {
238 IR = 2;
239 switch ((insn >> 5) & 0x1ff) {
240 case FCMPQ: TYPE(3,0,0,3,1,3,1); break;
241 case FCMPEQ: TYPE(3,0,0,3,1,3,1); break;
242 /* Now the conditional fmovq support */
243 case FMOVQ0:
244 case FMOVQ1:
245 case FMOVQ2:
246 case FMOVQ3:
247 /* fmovq %fccX, %fY, %fZ */
248 if (!((insn >> 11) & 3))
249 XR = current_thread_info()->xfsr[0] >> 10;
250 else
251 XR = current_thread_info()->xfsr[0] >> (30 + ((insn >> 10) & 0x6));
252 XR &= 3;
253 IR = 0;
254 switch ((insn >> 14) & 0x7) {
255 /* case 0: IR = 0; break; */ /* Never */
256 case 1: if (XR) IR = 1; break; /* Not Equal */
257 case 2: if (XR == 1 || XR == 2) IR = 1; break; /* Less or Greater */
258 case 3: if (XR & 1) IR = 1; break; /* Unordered or Less */
259 case 4: if (XR == 1) IR = 1; break; /* Less */
260 case 5: if (XR & 2) IR = 1; break; /* Unordered or Greater */
261 case 6: if (XR == 2) IR = 1; break; /* Greater */
262 case 7: if (XR == 3) IR = 1; break; /* Unordered */
264 if ((insn >> 14) & 8)
265 IR ^= 1;
266 break;
267 case FMOVQI:
268 case FMOVQX:
269 /* fmovq %[ix]cc, %fY, %fZ */
270 XR = regs->tstate >> 32;
271 if ((insn >> 5) & 0x80)
272 XR >>= 4;
273 XR &= 0xf;
274 IR = 0;
275 freg = ((XR >> 2) ^ XR) & 2;
276 switch ((insn >> 14) & 0x7) {
277 /* case 0: IR = 0; break; */ /* Never */
278 case 1: if (XR & 4) IR = 1; break; /* Equal */
279 case 2: if ((XR & 4) || freg) IR = 1; break; /* Less or Equal */
280 case 3: if (freg) IR = 1; break; /* Less */
281 case 4: if (XR & 5) IR = 1; break; /* Less or Equal Unsigned */
282 case 5: if (XR & 1) IR = 1; break; /* Carry Set */
283 case 6: if (XR & 8) IR = 1; break; /* Negative */
284 case 7: if (XR & 2) IR = 1; break; /* Overflow Set */
286 if ((insn >> 14) & 8)
287 IR ^= 1;
288 break;
289 case FMOVQZ:
290 case FMOVQLE:
291 case FMOVQLZ:
292 case FMOVQNZ:
293 case FMOVQGZ:
294 case FMOVQGE:
295 freg = (insn >> 14) & 0x1f;
296 if (!freg)
297 XR = 0;
298 else if (freg < 16)
299 XR = regs->u_regs[freg];
300 else if (test_thread_flag(TIF_32BIT)) {
301 struct reg_window32 __user *win32;
302 flushw_user ();
303 win32 = (struct reg_window32 __user *)((unsigned long)((u32)regs->u_regs[UREG_FP]));
304 get_user(XR, &win32->locals[freg - 16]);
305 } else {
306 struct reg_window __user *win;
307 flushw_user ();
308 win = (struct reg_window __user *)(regs->u_regs[UREG_FP] + STACK_BIAS);
309 get_user(XR, &win->locals[freg - 16]);
311 IR = 0;
312 switch ((insn >> 10) & 3) {
313 case 1: if (!XR) IR = 1; break; /* Register Zero */
314 case 2: if (XR <= 0) IR = 1; break; /* Register Less Than or Equal to Zero */
315 case 3: if (XR < 0) IR = 1; break; /* Register Less Than Zero */
317 if ((insn >> 10) & 4)
318 IR ^= 1;
319 break;
321 if (IR == 0) {
322 /* The fmov test was false. Do a nop instead */
323 current_thread_info()->xfsr[0] &= ~(FSR_CEXC_MASK);
324 regs->tpc = regs->tnpc;
325 regs->tnpc += 4;
326 return 1;
327 } else if (IR == 1) {
328 /* Change the instruction into plain fmovq */
329 insn = (insn & 0x3e00001f) | 0x81a00060;
330 TYPE(3,3,0,3,0,0,0);
334 if (type) {
335 argp rs1 = NULL, rs2 = NULL, rd = NULL;
337 freg = (current_thread_info()->xfsr[0] >> 14) & 0xf;
338 if (freg != (type >> 9))
339 goto err;
340 current_thread_info()->xfsr[0] &= ~0x1c000;
341 freg = ((insn >> 14) & 0x1f);
342 switch (type & 0x3) {
343 case 3: if (freg & 2) {
344 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */;
345 goto err;
347 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e);
348 case 1: rs1 = (argp)&f->regs[freg];
349 flags = (freg < 32) ? FPRS_DL : FPRS_DU;
350 if (!(current_thread_info()->fpsaved[0] & flags))
351 rs1 = (argp)&zero;
352 break;
354 switch (type & 0x7) {
355 case 7: FP_UNPACK_QP (QA, rs1); break;
356 case 6: FP_UNPACK_DP (DA, rs1); break;
357 case 5: FP_UNPACK_SP (SA, rs1); break;
359 freg = (insn & 0x1f);
360 switch ((type >> 3) & 0x3) {
361 case 3: if (freg & 2) {
362 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */;
363 goto err;
365 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e);
366 case 1: rs2 = (argp)&f->regs[freg];
367 flags = (freg < 32) ? FPRS_DL : FPRS_DU;
368 if (!(current_thread_info()->fpsaved[0] & flags))
369 rs2 = (argp)&zero;
370 break;
372 switch ((type >> 3) & 0x7) {
373 case 7: FP_UNPACK_QP (QB, rs2); break;
374 case 6: FP_UNPACK_DP (DB, rs2); break;
375 case 5: FP_UNPACK_SP (SB, rs2); break;
377 freg = ((insn >> 25) & 0x1f);
378 switch ((type >> 6) & 0x3) {
379 case 3: if (freg & 2) {
380 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */;
381 goto err;
383 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e);
384 case 1: rd = (argp)&f->regs[freg];
385 flags = (freg < 32) ? FPRS_DL : FPRS_DU;
386 if (!(current_thread_info()->fpsaved[0] & FPRS_FEF)) {
387 current_thread_info()->fpsaved[0] = FPRS_FEF;
388 current_thread_info()->gsr[0] = 0;
390 if (!(current_thread_info()->fpsaved[0] & flags)) {
391 if (freg < 32)
392 memset(f->regs, 0, 32*sizeof(u32));
393 else
394 memset(f->regs+32, 0, 32*sizeof(u32));
396 current_thread_info()->fpsaved[0] |= flags;
397 break;
399 switch ((insn >> 5) & 0x1ff) {
400 /* + */
401 case FADDS: FP_ADD_S (SR, SA, SB); break;
402 case FADDD: FP_ADD_D (DR, DA, DB); break;
403 case FADDQ: FP_ADD_Q (QR, QA, QB); break;
404 /* - */
405 case FSUBS: FP_SUB_S (SR, SA, SB); break;
406 case FSUBD: FP_SUB_D (DR, DA, DB); break;
407 case FSUBQ: FP_SUB_Q (QR, QA, QB); break;
408 /* * */
409 case FMULS: FP_MUL_S (SR, SA, SB); break;
410 case FSMULD: FP_CONV (D, S, 1, 1, DA, SA);
411 FP_CONV (D, S, 1, 1, DB, SB);
412 case FMULD: FP_MUL_D (DR, DA, DB); break;
413 case FDMULQ: FP_CONV (Q, D, 2, 1, QA, DA);
414 FP_CONV (Q, D, 2, 1, QB, DB);
415 case FMULQ: FP_MUL_Q (QR, QA, QB); break;
416 /* / */
417 case FDIVS: FP_DIV_S (SR, SA, SB); break;
418 case FDIVD: FP_DIV_D (DR, DA, DB); break;
419 case FDIVQ: FP_DIV_Q (QR, QA, QB); break;
420 /* sqrt */
421 case FSQRTS: FP_SQRT_S (SR, SB); break;
422 case FSQRTD: FP_SQRT_D (DR, DB); break;
423 case FSQRTQ: FP_SQRT_Q (QR, QB); break;
424 /* mov */
425 case FMOVQ: rd->q[0] = rs2->q[0]; rd->q[1] = rs2->q[1]; break;
426 case FABSQ: rd->q[0] = rs2->q[0] & 0x7fffffffffffffffUL; rd->q[1] = rs2->q[1]; break;
427 case FNEGQ: rd->q[0] = rs2->q[0] ^ 0x8000000000000000UL; rd->q[1] = rs2->q[1]; break;
428 /* float to int */
429 case FSTOI: FP_TO_INT_S (IR, SB, 32, 1); break;
430 case FDTOI: FP_TO_INT_D (IR, DB, 32, 1); break;
431 case FQTOI: FP_TO_INT_Q (IR, QB, 32, 1); break;
432 case FSTOX: FP_TO_INT_S (XR, SB, 64, 1); break;
433 case FDTOX: FP_TO_INT_D (XR, DB, 64, 1); break;
434 case FQTOX: FP_TO_INT_Q (XR, QB, 64, 1); break;
435 /* int to float */
436 case FITOQ: IR = rs2->s; FP_FROM_INT_Q (QR, IR, 32, int); break;
437 case FXTOQ: XR = rs2->d; FP_FROM_INT_Q (QR, XR, 64, long); break;
438 /* Only Ultra-III generates these */
439 case FXTOS: XR = rs2->d; FP_FROM_INT_S (SR, XR, 64, long); break;
440 case FXTOD: XR = rs2->d; FP_FROM_INT_D (DR, XR, 64, long); break;
441 #if 0 /* Optimized inline in sparc64/kernel/entry.S */
442 case FITOS: IR = rs2->s; FP_FROM_INT_S (SR, IR, 32, int); break;
443 #endif
444 case FITOD: IR = rs2->s; FP_FROM_INT_D (DR, IR, 32, int); break;
445 /* float to float */
446 case FSTOD: FP_CONV (D, S, 1, 1, DR, SB); break;
447 case FSTOQ: FP_CONV (Q, S, 2, 1, QR, SB); break;
448 case FDTOQ: FP_CONV (Q, D, 2, 1, QR, DB); break;
449 case FDTOS: FP_CONV (S, D, 1, 1, SR, DB); break;
450 case FQTOS: FP_CONV (S, Q, 1, 2, SR, QB); break;
451 case FQTOD: FP_CONV (D, Q, 1, 2, DR, QB); break;
452 /* comparison */
453 case FCMPQ:
454 case FCMPEQ:
455 FP_CMP_Q(XR, QB, QA, 3);
456 if (XR == 3 &&
457 (((insn >> 5) & 0x1ff) == FCMPEQ ||
458 FP_ISSIGNAN_Q(QA) ||
459 FP_ISSIGNAN_Q(QB)))
460 FP_SET_EXCEPTION (FP_EX_INVALID);
462 if (!FP_INHIBIT_RESULTS) {
463 switch ((type >> 6) & 0x7) {
464 case 0: xfsr = current_thread_info()->xfsr[0];
465 if (XR == -1) XR = 2;
466 switch (freg & 3) {
467 /* fcc0, 1, 2, 3 */
468 case 0: xfsr &= ~0xc00; xfsr |= (XR << 10); break;
469 case 1: xfsr &= ~0x300000000UL; xfsr |= (XR << 32); break;
470 case 2: xfsr &= ~0xc00000000UL; xfsr |= (XR << 34); break;
471 case 3: xfsr &= ~0x3000000000UL; xfsr |= (XR << 36); break;
473 current_thread_info()->xfsr[0] = xfsr;
474 break;
475 case 1: rd->s = IR; break;
476 case 2: rd->d = XR; break;
477 case 5: FP_PACK_SP (rd, SR); break;
478 case 6: FP_PACK_DP (rd, DR); break;
479 case 7: FP_PACK_QP (rd, QR); break;
483 if(_fex != 0)
484 return record_exception(regs, _fex);
486 /* Success and no exceptions detected. */
487 current_thread_info()->xfsr[0] &= ~(FSR_CEXC_MASK);
488 regs->tpc = regs->tnpc;
489 regs->tnpc += 4;
490 return 1;
492 err: return 0;