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.
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>
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>
43 /* SUBNORMAL - ftt == 2 */
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. */
66 #define FITOD 0x0c8 /* Only Ultra-III generates this. */
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
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];
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.
139 fsr |= ((long)eflag << FSR_AEXC_SHIFT);
141 /* If trapping, indicate fault trap type IEEE. */
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;
155 return (would_trap ? 0 : 1);
164 int do_mathemu(struct pt_regs *regs, struct fpustate *f)
166 unsigned long pc = regs->tpc;
167 unsigned long tstate = regs->tstate;
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)
175 static u64 zero[2] = { 0L, 0L };
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);
184 if (tstate & TSTATE_PRIV)
185 die_if_kernel("unfinished/unimplemented FPop from kernel", regs);
186 if (test_thread_flag(TIF_32BIT))
188 if (get_user(insn, (u32 __user *) pc) != -EFAULT) {
189 if ((insn & 0xc1f80000) == 0x81a00000) /* FPOP1 */ {
190 switch ((insn >> 5) & 0x1ff) {
191 /* QUAD - ftt == 3 */
194 case FABSQ: TYPE(3,3,0,3,0,0,0); break;
195 case FSQRTQ: TYPE(3,3,1,3,1,0,0); break;
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;
210 /* We can get either unimplemented or unfinished
211 * for these cases. Pre-Niagara systems generate
212 * unfinished fpop for SUBNORMAL cases, and Niagara
213 * always gives unimplemented fpop for fsqrt{s,d}.
216 unsigned long x = current_thread_info()->xfsr[0];
224 unsigned long x = current_thread_info()->xfsr[0];
231 /* SUBNORMAL - ftt == 2 */
235 case FDIVD: TYPE(2,2,1,2,1,2,1); break;
239 case FDIVS: TYPE(2,1,1,1,1,1,1); break;
240 case FSMULD: TYPE(2,2,1,1,1,1,1); break;
241 case FSTOX: TYPE(2,2,0,1,1,0,0); break;
242 case FDTOX: TYPE(2,2,0,2,1,0,0); break;
243 case FDTOS: TYPE(2,1,1,2,1,0,0); break;
244 case FSTOD: TYPE(2,2,1,1,1,0,0); break;
245 case FSTOI: TYPE(2,1,0,1,1,0,0); break;
246 case FDTOI: TYPE(2,1,0,2,1,0,0); break;
248 /* Only Ultra-III generates these */
249 case FXTOS: TYPE(2,1,1,2,0,0,0); break;
250 case FXTOD: TYPE(2,2,1,2,0,0,0); break;
251 #if 0 /* Optimized inline in sparc64/kernel/entry.S */
252 case FITOS: TYPE(2,1,1,1,0,0,0); break;
254 case FITOD: TYPE(2,2,1,1,0,0,0); break;
257 else if ((insn & 0xc1f80000) == 0x81a80000) /* FPOP2 */ {
259 switch ((insn >> 5) & 0x1ff) {
260 case FCMPQ: TYPE(3,0,0,3,1,3,1); break;
261 case FCMPEQ: TYPE(3,0,0,3,1,3,1); break;
262 /* Now the conditional fmovq support */
267 /* fmovq %fccX, %fY, %fZ */
268 if (!((insn >> 11) & 3))
269 XR = current_thread_info()->xfsr[0] >> 10;
271 XR = current_thread_info()->xfsr[0] >> (30 + ((insn >> 10) & 0x6));
274 switch ((insn >> 14) & 0x7) {
275 /* case 0: IR = 0; break; */ /* Never */
276 case 1: if (XR) IR = 1; break; /* Not Equal */
277 case 2: if (XR == 1 || XR == 2) IR = 1; break; /* Less or Greater */
278 case 3: if (XR & 1) IR = 1; break; /* Unordered or Less */
279 case 4: if (XR == 1) IR = 1; break; /* Less */
280 case 5: if (XR & 2) IR = 1; break; /* Unordered or Greater */
281 case 6: if (XR == 2) IR = 1; break; /* Greater */
282 case 7: if (XR == 3) IR = 1; break; /* Unordered */
284 if ((insn >> 14) & 8)
289 /* fmovq %[ix]cc, %fY, %fZ */
290 XR = regs->tstate >> 32;
291 if ((insn >> 5) & 0x80)
295 freg = ((XR >> 2) ^ XR) & 2;
296 switch ((insn >> 14) & 0x7) {
297 /* case 0: IR = 0; break; */ /* Never */
298 case 1: if (XR & 4) IR = 1; break; /* Equal */
299 case 2: if ((XR & 4) || freg) IR = 1; break; /* Less or Equal */
300 case 3: if (freg) IR = 1; break; /* Less */
301 case 4: if (XR & 5) IR = 1; break; /* Less or Equal Unsigned */
302 case 5: if (XR & 1) IR = 1; break; /* Carry Set */
303 case 6: if (XR & 8) IR = 1; break; /* Negative */
304 case 7: if (XR & 2) IR = 1; break; /* Overflow Set */
306 if ((insn >> 14) & 8)
315 freg = (insn >> 14) & 0x1f;
319 XR = regs->u_regs[freg];
320 else if (test_thread_flag(TIF_32BIT)) {
321 struct reg_window32 __user *win32;
323 win32 = (struct reg_window32 __user *)((unsigned long)((u32)regs->u_regs[UREG_FP]));
324 get_user(XR, &win32->locals[freg - 16]);
326 struct reg_window __user *win;
328 win = (struct reg_window __user *)(regs->u_regs[UREG_FP] + STACK_BIAS);
329 get_user(XR, &win->locals[freg - 16]);
332 switch ((insn >> 10) & 3) {
333 case 1: if (!XR) IR = 1; break; /* Register Zero */
334 case 2: if (XR <= 0) IR = 1; break; /* Register Less Than or Equal to Zero */
335 case 3: if (XR < 0) IR = 1; break; /* Register Less Than Zero */
337 if ((insn >> 10) & 4)
342 /* The fmov test was false. Do a nop instead */
343 current_thread_info()->xfsr[0] &= ~(FSR_CEXC_MASK);
344 regs->tpc = regs->tnpc;
347 } else if (IR == 1) {
348 /* Change the instruction into plain fmovq */
349 insn = (insn & 0x3e00001f) | 0x81a00060;
355 argp rs1 = NULL, rs2 = NULL, rd = NULL;
357 freg = (current_thread_info()->xfsr[0] >> 14) & 0xf;
358 if (freg != (type >> 9))
360 current_thread_info()->xfsr[0] &= ~0x1c000;
361 freg = ((insn >> 14) & 0x1f);
362 switch (type & 0x3) {
363 case 3: if (freg & 2) {
364 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */;
367 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e);
368 case 1: rs1 = (argp)&f->regs[freg];
369 flags = (freg < 32) ? FPRS_DL : FPRS_DU;
370 if (!(current_thread_info()->fpsaved[0] & flags))
374 switch (type & 0x7) {
375 case 7: FP_UNPACK_QP (QA, rs1); break;
376 case 6: FP_UNPACK_DP (DA, rs1); break;
377 case 5: FP_UNPACK_SP (SA, rs1); break;
379 freg = (insn & 0x1f);
380 switch ((type >> 3) & 0x3) {
381 case 3: if (freg & 2) {
382 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */;
385 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e);
386 case 1: rs2 = (argp)&f->regs[freg];
387 flags = (freg < 32) ? FPRS_DL : FPRS_DU;
388 if (!(current_thread_info()->fpsaved[0] & flags))
392 switch ((type >> 3) & 0x7) {
393 case 7: FP_UNPACK_QP (QB, rs2); break;
394 case 6: FP_UNPACK_DP (DB, rs2); break;
395 case 5: FP_UNPACK_SP (SB, rs2); break;
397 freg = ((insn >> 25) & 0x1f);
398 switch ((type >> 6) & 0x3) {
399 case 3: if (freg & 2) {
400 current_thread_info()->xfsr[0] |= (6 << 14) /* invalid_fp_register */;
403 case 2: freg = ((freg & 1) << 5) | (freg & 0x1e);
404 case 1: rd = (argp)&f->regs[freg];
405 flags = (freg < 32) ? FPRS_DL : FPRS_DU;
406 if (!(current_thread_info()->fpsaved[0] & FPRS_FEF)) {
407 current_thread_info()->fpsaved[0] = FPRS_FEF;
408 current_thread_info()->gsr[0] = 0;
410 if (!(current_thread_info()->fpsaved[0] & flags)) {
412 memset(f->regs, 0, 32*sizeof(u32));
414 memset(f->regs+32, 0, 32*sizeof(u32));
416 current_thread_info()->fpsaved[0] |= flags;
419 switch ((insn >> 5) & 0x1ff) {
421 case FADDS: FP_ADD_S (SR, SA, SB); break;
422 case FADDD: FP_ADD_D (DR, DA, DB); break;
423 case FADDQ: FP_ADD_Q (QR, QA, QB); break;
425 case FSUBS: FP_SUB_S (SR, SA, SB); break;
426 case FSUBD: FP_SUB_D (DR, DA, DB); break;
427 case FSUBQ: FP_SUB_Q (QR, QA, QB); break;
429 case FMULS: FP_MUL_S (SR, SA, SB); break;
430 case FSMULD: FP_CONV (D, S, 1, 1, DA, SA);
431 FP_CONV (D, S, 1, 1, DB, SB);
432 case FMULD: FP_MUL_D (DR, DA, DB); break;
433 case FDMULQ: FP_CONV (Q, D, 2, 1, QA, DA);
434 FP_CONV (Q, D, 2, 1, QB, DB);
435 case FMULQ: FP_MUL_Q (QR, QA, QB); break;
437 case FDIVS: FP_DIV_S (SR, SA, SB); break;
438 case FDIVD: FP_DIV_D (DR, DA, DB); break;
439 case FDIVQ: FP_DIV_Q (QR, QA, QB); break;
441 case FSQRTS: FP_SQRT_S (SR, SB); break;
442 case FSQRTD: FP_SQRT_D (DR, DB); break;
443 case FSQRTQ: FP_SQRT_Q (QR, QB); break;
445 case FMOVQ: rd->q[0] = rs2->q[0]; rd->q[1] = rs2->q[1]; break;
446 case FABSQ: rd->q[0] = rs2->q[0] & 0x7fffffffffffffffUL; rd->q[1] = rs2->q[1]; break;
447 case FNEGQ: rd->q[0] = rs2->q[0] ^ 0x8000000000000000UL; rd->q[1] = rs2->q[1]; break;
449 case FSTOI: FP_TO_INT_S (IR, SB, 32, 1); break;
450 case FDTOI: FP_TO_INT_D (IR, DB, 32, 1); break;
451 case FQTOI: FP_TO_INT_Q (IR, QB, 32, 1); break;
452 case FSTOX: FP_TO_INT_S (XR, SB, 64, 1); break;
453 case FDTOX: FP_TO_INT_D (XR, DB, 64, 1); break;
454 case FQTOX: FP_TO_INT_Q (XR, QB, 64, 1); break;
456 case FITOQ: IR = rs2->s; FP_FROM_INT_Q (QR, IR, 32, int); break;
457 case FXTOQ: XR = rs2->d; FP_FROM_INT_Q (QR, XR, 64, long); break;
458 /* Only Ultra-III generates these */
459 case FXTOS: XR = rs2->d; FP_FROM_INT_S (SR, XR, 64, long); break;
460 case FXTOD: XR = rs2->d; FP_FROM_INT_D (DR, XR, 64, long); break;
461 #if 0 /* Optimized inline in sparc64/kernel/entry.S */
462 case FITOS: IR = rs2->s; FP_FROM_INT_S (SR, IR, 32, int); break;
464 case FITOD: IR = rs2->s; FP_FROM_INT_D (DR, IR, 32, int); break;
466 case FSTOD: FP_CONV (D, S, 1, 1, DR, SB); break;
467 case FSTOQ: FP_CONV (Q, S, 2, 1, QR, SB); break;
468 case FDTOQ: FP_CONV (Q, D, 2, 1, QR, DB); break;
469 case FDTOS: FP_CONV (S, D, 1, 1, SR, DB); break;
470 case FQTOS: FP_CONV (S, Q, 1, 2, SR, QB); break;
471 case FQTOD: FP_CONV (D, Q, 1, 2, DR, QB); break;
475 FP_CMP_Q(XR, QB, QA, 3);
477 (((insn >> 5) & 0x1ff) == FCMPEQ ||
480 FP_SET_EXCEPTION (FP_EX_INVALID);
482 if (!FP_INHIBIT_RESULTS) {
483 switch ((type >> 6) & 0x7) {
484 case 0: xfsr = current_thread_info()->xfsr[0];
485 if (XR == -1) XR = 2;
488 case 0: xfsr &= ~0xc00; xfsr |= (XR << 10); break;
489 case 1: xfsr &= ~0x300000000UL; xfsr |= (XR << 32); break;
490 case 2: xfsr &= ~0xc00000000UL; xfsr |= (XR << 34); break;
491 case 3: xfsr &= ~0x3000000000UL; xfsr |= (XR << 36); break;
493 current_thread_info()->xfsr[0] = xfsr;
495 case 1: rd->s = IR; break;
496 case 2: rd->d = XR; break;
497 case 5: FP_PACK_SP (rd, SR); break;
498 case 6: FP_PACK_DP (rd, DR); break;
499 case 7: FP_PACK_QP (rd, QR); break;
504 return record_exception(regs, _fex);
506 /* Success and no exceptions detected. */
507 current_thread_info()->xfsr[0] &= ~(FSR_CEXC_MASK);
508 regs->tpc = regs->tnpc;