xref: /freebsd/sys/powerpc/fpu/fpu_emu.c (revision 23541160bb3e58f5deb04a299eda60fc80b731bc)
1 /*	$NetBSD: fpu_emu.c,v 1.14 2005/12/11 12:18:42 christos Exp $ */
2 
3 /*
4  * Copyright 2001 Wasabi Systems, Inc.
5  * All rights reserved.
6  *
7  * Written by Eduardo Horvath and Simon Burge for Wasabi Systems, Inc.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. All advertising materials mentioning features or use of this software
18  *    must display the following acknowledgement:
19  *      This product includes software developed for the NetBSD Project by
20  *      Wasabi Systems, Inc.
21  * 4. The name of Wasabi Systems, Inc. may not be used to endorse
22  *    or promote products derived from this software without specific prior
23  *    written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY WASABI SYSTEMS, INC. ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
27  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
28  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL WASABI SYSTEMS, INC
29  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
30  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
31  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
32  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
33  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
34  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
35  * POSSIBILITY OF SUCH DAMAGE.
36  */
37 
38 /*
39  * Copyright (c) 1992, 1993
40  *	The Regents of the University of California.  All rights reserved.
41  *
42  * This software was developed by the Computer Systems Engineering group
43  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
44  * contributed to Berkeley.
45  *
46  * All advertising materials mentioning features or use of this software
47  * must display the following acknowledgement:
48  *	This product includes software developed by the University of
49  *	California, Lawrence Berkeley Laboratory.
50  *
51  * Redistribution and use in source and binary forms, with or without
52  * modification, are permitted provided that the following conditions
53  * are met:
54  * 1. Redistributions of source code must retain the above copyright
55  *    notice, this list of conditions and the following disclaimer.
56  * 2. Redistributions in binary form must reproduce the above copyright
57  *    notice, this list of conditions and the following disclaimer in the
58  *    documentation and/or other materials provided with the distribution.
59  * 3. Neither the name of the University nor the names of its contributors
60  *    may be used to endorse or promote products derived from this software
61  *    without specific prior written permission.
62  *
63  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
64  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
65  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
66  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
67  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
68  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
69  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
70  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
71  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
72  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
73  * SUCH DAMAGE.
74  *
75  *	@(#)fpu.c	8.1 (Berkeley) 6/11/93
76  */
77 
78 #include <sys/cdefs.h>
79 __FBSDID("$FreeBSD$");
80 
81 #include "opt_ddb.h"
82 
83 #include <sys/param.h>
84 #include <sys/systm.h>
85 #include <sys/kdb.h>
86 #include <sys/kernel.h>
87 #include <sys/proc.h>
88 #include <sys/sysctl.h>
89 #include <sys/signal.h>
90 #include <sys/syslog.h>
91 #include <sys/signalvar.h>
92 
93 #include <machine/fpu.h>
94 #include <machine/reg.h>
95 
96 #include <powerpc/fpu/fpu_emu.h>
97 #include <powerpc/fpu/fpu_extern.h>
98 #include <powerpc/fpu/fpu_instr.h>
99 
100 static SYSCTL_NODE(_hw, OID_AUTO, fpu_emu, CTLFLAG_RW, 0, "FPU emulator");
101 
102 #define	FPU_EMU_EVCNT_DECL(name)					\
103 static u_int fpu_emu_evcnt_##name;					\
104 SYSCTL_INT(_hw_fpu_emu, OID_AUTO, evcnt_##name, CTLFLAG_RD,		\
105     &fpu_emu_evcnt_##name, 0, "")
106 
107 #define	FPU_EMU_EVCNT_INCR(name)	fpu_emu_evcnt_##name++
108 
109 FPU_EMU_EVCNT_DECL(stfiwx);
110 FPU_EMU_EVCNT_DECL(fpstore);
111 FPU_EMU_EVCNT_DECL(fpload);
112 FPU_EMU_EVCNT_DECL(fcmpu);
113 FPU_EMU_EVCNT_DECL(frsp);
114 FPU_EMU_EVCNT_DECL(fctiw);
115 FPU_EMU_EVCNT_DECL(fcmpo);
116 FPU_EMU_EVCNT_DECL(mtfsb1);
117 FPU_EMU_EVCNT_DECL(fnegabs);
118 FPU_EMU_EVCNT_DECL(mcrfs);
119 FPU_EMU_EVCNT_DECL(mtfsb0);
120 FPU_EMU_EVCNT_DECL(fmr);
121 FPU_EMU_EVCNT_DECL(mtfsfi);
122 FPU_EMU_EVCNT_DECL(fnabs);
123 FPU_EMU_EVCNT_DECL(fabs);
124 FPU_EMU_EVCNT_DECL(mffs);
125 FPU_EMU_EVCNT_DECL(mtfsf);
126 FPU_EMU_EVCNT_DECL(fctid);
127 FPU_EMU_EVCNT_DECL(fcfid);
128 FPU_EMU_EVCNT_DECL(fdiv);
129 FPU_EMU_EVCNT_DECL(fsub);
130 FPU_EMU_EVCNT_DECL(fadd);
131 FPU_EMU_EVCNT_DECL(fsqrt);
132 FPU_EMU_EVCNT_DECL(fsel);
133 FPU_EMU_EVCNT_DECL(fpres);
134 FPU_EMU_EVCNT_DECL(fmul);
135 FPU_EMU_EVCNT_DECL(frsqrte);
136 FPU_EMU_EVCNT_DECL(fmulsub);
137 FPU_EMU_EVCNT_DECL(fmuladd);
138 FPU_EMU_EVCNT_DECL(fnmsub);
139 FPU_EMU_EVCNT_DECL(fnmadd);
140 
141 /* FPSR exception masks */
142 #define FPSR_EX_MSK	(FPSCR_VX|FPSCR_OX|FPSCR_UX|FPSCR_ZX|		\
143 			FPSCR_XX|FPSCR_VXSNAN|FPSCR_VXISI|FPSCR_VXIDI|	\
144 			FPSCR_VXZDZ|FPSCR_VXIMZ|FPSCR_VXVC|FPSCR_VXSOFT|\
145 			FPSCR_VXSQRT|FPSCR_VXCVI)
146 #define	FPSR_EX		(FPSCR_VE|FPSCR_OE|FPSCR_UE|FPSCR_ZE|FPSCR_XE)
147 #define	FPSR_EXOP	(FPSR_EX_MSK&(~FPSR_EX))
148 
149 int fpe_debug = 0;
150 
151 #ifdef DEBUG
152 vm_offset_t opc_disasm(vm_offset_t, int);
153 
154 /*
155  * Dump a `fpn' structure.
156  */
157 void
158 fpu_dumpfpn(struct fpn *fp)
159 {
160 	static const char *class[] = {
161 		"SNAN", "QNAN", "ZERO", "NUM", "INF"
162 	};
163 
164 	printf("%s %c.%x %x %x %xE%d", class[fp->fp_class + 2],
165 		fp->fp_sign ? '-' : ' ',
166 		fp->fp_mant[0],	fp->fp_mant[1],
167 		fp->fp_mant[2], fp->fp_mant[3],
168 		fp->fp_exp);
169 }
170 #endif
171 
172 /*
173  * fpu_execute returns the following error numbers (0 = no error):
174  */
175 #define	FPE		1	/* take a floating point exception */
176 #define	NOTFPU		2	/* not an FPU instruction */
177 #define	FAULT		3
178 
179 
180 /*
181  * Emulate a floating-point instruction.
182  * Return zero for success, else signal number.
183  * (Typically: zero, SIGFPE, SIGILL, SIGSEGV)
184  */
185 int
186 fpu_emulate(struct trapframe *frame, struct fpu *fpf)
187 {
188 	static union instr insn;
189 	static struct fpemu fe;
190 	static int lastill = 0;
191 	int sig;
192 
193 	/* initialize insn.is_datasize to tell it is *not* initialized */
194 	fe.fe_fpstate = fpf;
195 	fe.fe_cx = 0;
196 
197 	/* always set this (to avoid a warning) */
198 
199 	if (copyin((void *) (frame->srr0), &insn.i_int, sizeof (insn.i_int))) {
200 #ifdef DEBUG
201 		printf("fpu_emulate: fault reading opcode\n");
202 #endif
203 		return SIGSEGV;
204 	}
205 
206 	DPRINTF(FPE_EX, ("fpu_emulate: emulating insn %x at %p\n",
207 	    insn.i_int, (void *)frame->srr0));
208 
209 
210 	if ((insn.i_any.i_opcd == OPC_TWI) ||
211 	    ((insn.i_any.i_opcd == OPC_integer_31) &&
212 	    (insn.i_x.i_xo == OPC31_TW))) {
213 		/* Check for the two trap insns. */
214 		DPRINTF(FPE_EX, ("fpu_emulate: SIGTRAP\n"));
215 		return (SIGTRAP);
216 	}
217 	sig = 0;
218 	switch (fpu_execute(frame, &fe, &insn)) {
219 	case 0:
220 		DPRINTF(FPE_EX, ("fpu_emulate: success\n"));
221 		frame->srr0 += 4;
222 		break;
223 
224 	case FPE:
225 		DPRINTF(FPE_EX, ("fpu_emulate: SIGFPE\n"));
226 		sig = SIGFPE;
227 		break;
228 
229 	case FAULT:
230 		DPRINTF(FPE_EX, ("fpu_emulate: SIGSEGV\n"));
231 		sig = SIGSEGV;
232 		break;
233 
234 	case NOTFPU:
235 	default:
236 		DPRINTF(FPE_EX, ("fpu_emulate: SIGILL\n"));
237 #ifdef DEBUG
238 		if (fpe_debug & FPE_EX) {
239 			printf("fpu_emulate:  illegal insn %x at %p:",
240 			insn.i_int, (void *) (frame->srr0));
241 			opc_disasm(frame->srr0, insn.i_int);
242 		}
243 #endif
244 		/*
245 		* XXXX retry an illegal insn once due to cache issues.
246 		*/
247 		if (lastill == frame->srr0) {
248 			sig = SIGILL;
249 #ifdef DEBUG
250 			if (fpe_debug & FPE_EX)
251 				kdb_enter(KDB_WHY_UNSET, "illegal instruction");
252 #endif
253 		}
254 		lastill = frame->srr0;
255 		break;
256 	}
257 
258 	return (sig);
259 }
260 
261 /*
262  * Execute an FPU instruction (one that runs entirely in the FPU; not
263  * FBfcc or STF, for instance).  On return, fe->fe_fs->fs_fsr will be
264  * modified to reflect the setting the hardware would have left.
265  *
266  * Note that we do not catch all illegal opcodes, so you can, for instance,
267  * multiply two integers this way.
268  */
269 int
270 fpu_execute(struct trapframe *tf, struct fpemu *fe, union instr *insn)
271 {
272 	struct fpn *fp;
273 	union instr instr = *insn;
274 	int *a;
275 	vm_offset_t addr;
276 	int ra, rb, rc, rt, type, mask, fsr, cx, bf, setcr;
277 	unsigned int cond;
278 	struct fpu *fs;
279 
280 	/* Setup work. */
281 	fp = NULL;
282 	fs = fe->fe_fpstate;
283 	fe->fe_fpscr = ((int *)&fs->fpscr)[1];
284 
285 	/*
286 	 * On PowerPC all floating point values are stored in registers
287 	 * as doubles, even when used for single precision operations.
288 	 */
289 	type = FTYPE_DBL;
290 	cond = instr.i_any.i_rc;
291 	setcr = 0;
292 	bf = 0;	/* XXX gcc */
293 
294 #if defined(DDB) && defined(DEBUG)
295 	if (fpe_debug & FPE_EX) {
296 		vm_offset_t loc = tf->srr0;
297 
298 		printf("Trying to emulate: %p ", (void *)loc);
299 		opc_disasm(loc, instr.i_int);
300 	}
301 #endif
302 
303 	/*
304 	 * `Decode' and execute instruction.
305 	 */
306 
307 	if ((instr.i_any.i_opcd >= OPC_LFS && instr.i_any.i_opcd <= OPC_STFDU) ||
308 	    instr.i_any.i_opcd == OPC_integer_31) {
309 		/*
310 		 * Handle load/store insns:
311 		 *
312 		 * Convert to/from single if needed, calculate addr,
313 		 * and update index reg if needed.
314 		 */
315 		double buf;
316 		size_t size = sizeof(float);
317 		int store, update;
318 
319 		cond = 0; /* ld/st never set condition codes */
320 
321 
322 		if (instr.i_any.i_opcd == OPC_integer_31) {
323 			if (instr.i_x.i_xo == OPC31_STFIWX) {
324 				FPU_EMU_EVCNT_INCR(stfiwx);
325 
326 				/* Store as integer */
327 				ra = instr.i_x.i_ra;
328 				rb = instr.i_x.i_rb;
329 				DPRINTF(FPE_INSN,
330 					("reg %d has %jx reg %d has %jx\n",
331 					ra, (uintmax_t)tf->fixreg[ra], rb,
332 					(uintmax_t)tf->fixreg[rb]));
333 
334 				addr = tf->fixreg[rb];
335 				if (ra != 0)
336 					addr += tf->fixreg[ra];
337 				rt = instr.i_x.i_rt;
338 				a = (int *)&fs->fpr[rt].fpr;
339 				DPRINTF(FPE_INSN,
340 					("fpu_execute: Store INT %x at %p\n",
341 						a[1], (void *)addr));
342 				if (copyout(&a[1], (void *)addr, sizeof(int)))
343 					return (FAULT);
344 				return (0);
345 			}
346 
347 			if ((instr.i_x.i_xo & OPC31_FPMASK) != OPC31_FPOP)
348 				/* Not an indexed FP load/store op */
349 				return (NOTFPU);
350 
351 			store = (instr.i_x.i_xo & 0x80);
352 			if (instr.i_x.i_xo & 0x40)
353 				size = sizeof(double);
354 			else
355 				type = FTYPE_SNG;
356 			update = (instr.i_x.i_xo & 0x20);
357 
358 			/* calculate EA of load/store */
359 			ra = instr.i_x.i_ra;
360 			rb = instr.i_x.i_rb;
361 			DPRINTF(FPE_INSN, ("reg %d has %jx reg %d has %jx\n",
362 				ra, (uintmax_t)tf->fixreg[ra], rb,
363 				(uintmax_t)tf->fixreg[rb]));
364 			addr = tf->fixreg[rb];
365 			if (ra != 0)
366 				addr += tf->fixreg[ra];
367 			rt = instr.i_x.i_rt;
368 		} else {
369 			store = instr.i_d.i_opcd & 0x4;
370 			if (instr.i_d.i_opcd & 0x2)
371 				size = sizeof(double);
372 			else
373 				type = FTYPE_SNG;
374 			update = instr.i_d.i_opcd & 0x1;
375 
376 			/* calculate EA of load/store */
377 			ra = instr.i_d.i_ra;
378 			addr = instr.i_d.i_d;
379 			DPRINTF(FPE_INSN, ("reg %d has %jx displ %jx\n",
380 				ra, (uintmax_t)tf->fixreg[ra],
381 				(uintmax_t)addr));
382 			if (ra != 0)
383 				addr += tf->fixreg[ra];
384 			rt = instr.i_d.i_rt;
385 		}
386 
387 		if (update && ra == 0)
388 			return (NOTFPU);
389 
390 		if (store) {
391 			/* Store */
392 			FPU_EMU_EVCNT_INCR(fpstore);
393 			if (type != FTYPE_DBL) {
394 				DPRINTF(FPE_INSN,
395 					("fpu_execute: Store SNG at %p\n",
396 						(void *)addr));
397 				fpu_explode(fe, fp = &fe->fe_f1, FTYPE_DBL, rt);
398 				fpu_implode(fe, fp, type, (void *)&buf);
399 				if (copyout(&buf, (void *)addr, size))
400 					return (FAULT);
401 			} else {
402 				DPRINTF(FPE_INSN,
403 					("fpu_execute: Store DBL at %p\n",
404 						(void *)addr));
405 				if (copyout(&fs->fpr[rt].fpr, (void *)addr,
406 				    size))
407 					return (FAULT);
408 			}
409 		} else {
410 			/* Load */
411 			FPU_EMU_EVCNT_INCR(fpload);
412 			DPRINTF(FPE_INSN, ("fpu_execute: Load from %p\n",
413 				(void *)addr));
414 			if (copyin((const void *)addr, &fs->fpr[rt].fpr,
415 			    size))
416 				return (FAULT);
417 			if (type != FTYPE_DBL) {
418 				fpu_explode(fe, fp = &fe->fe_f1, type, rt);
419 				fpu_implode(fe, fp, FTYPE_DBL,
420 					(u_int *)&fs->fpr[rt].fpr);
421 			}
422 		}
423 		if (update)
424 			tf->fixreg[ra] = addr;
425 		/* Complete. */
426 		return (0);
427 #ifdef notyet
428 	} else if (instr.i_any.i_opcd == OPC_load_st_62) {
429 		/* These are 64-bit extensions */
430 		return (NOTFPU);
431 #endif
432 	} else if (instr.i_any.i_opcd == OPC_sp_fp_59 ||
433 		instr.i_any.i_opcd == OPC_dp_fp_63) {
434 
435 
436 		if (instr.i_any.i_opcd == OPC_dp_fp_63 &&
437 		    !(instr.i_a.i_xo & OPC63M_MASK)) {
438 			/* Format X */
439 			rt = instr.i_x.i_rt;
440 			ra = instr.i_x.i_ra;
441 			rb = instr.i_x.i_rb;
442 
443 
444 			/* One of the special opcodes.... */
445 			switch (instr.i_x.i_xo) {
446 			case	OPC63_FCMPU:
447 				FPU_EMU_EVCNT_INCR(fcmpu);
448 				DPRINTF(FPE_INSN, ("fpu_execute: FCMPU\n"));
449 				rt >>= 2;
450 				fpu_explode(fe, &fe->fe_f1, type, ra);
451 				fpu_explode(fe, &fe->fe_f2, type, rb);
452 				fpu_compare(fe, 0);
453 				/* Make sure we do the condition regs. */
454 				cond = 0;
455 				/* N.B.: i_rs is already left shifted by two. */
456 				bf = instr.i_x.i_rs & 0xfc;
457 				setcr = 1;
458 				break;
459 
460 			case	OPC63_FRSP:
461 				/*
462 				 * Convert to single:
463 				 *
464 				 * PowerPC uses this to round a double
465 				 * precision value to single precision,
466 				 * but values in registers are always
467 				 * stored in double precision format.
468 				 */
469 				FPU_EMU_EVCNT_INCR(frsp);
470 				DPRINTF(FPE_INSN, ("fpu_execute: FRSP\n"));
471 				fpu_explode(fe, fp = &fe->fe_f1, FTYPE_DBL, rb);
472 				fpu_implode(fe, fp, FTYPE_SNG,
473 					(u_int *)&fs->fpr[rt].fpr);
474 				fpu_explode(fe, fp = &fe->fe_f1, FTYPE_SNG, rt);
475 				type = FTYPE_DBL;
476 				break;
477 			case	OPC63_FCTIW:
478 			case	OPC63_FCTIWZ:
479 				FPU_EMU_EVCNT_INCR(fctiw);
480 				DPRINTF(FPE_INSN, ("fpu_execute: FCTIW\n"));
481 				fpu_explode(fe, fp = &fe->fe_f1, type, rb);
482 				type = FTYPE_INT;
483 				break;
484 			case	OPC63_FCMPO:
485 				FPU_EMU_EVCNT_INCR(fcmpo);
486 				DPRINTF(FPE_INSN, ("fpu_execute: FCMPO\n"));
487 				rt >>= 2;
488 				fpu_explode(fe, &fe->fe_f1, type, ra);
489 				fpu_explode(fe, &fe->fe_f2, type, rb);
490 				fpu_compare(fe, 1);
491 				/* Make sure we do the condition regs. */
492 				cond = 0;
493 				/* N.B.: i_rs is already left shifted by two. */
494 				bf = instr.i_x.i_rs & 0xfc;
495 				setcr = 1;
496 				break;
497 			case	OPC63_MTFSB1:
498 				FPU_EMU_EVCNT_INCR(mtfsb1);
499 				DPRINTF(FPE_INSN, ("fpu_execute: MTFSB1\n"));
500 				fe->fe_fpscr |=
501 					(~(FPSCR_VX|FPSR_EX) & (1<<(31-rt)));
502 				break;
503 			case	OPC63_FNEG:
504 				FPU_EMU_EVCNT_INCR(fnegabs);
505 				DPRINTF(FPE_INSN, ("fpu_execute: FNEGABS\n"));
506 				memcpy(&fs->fpr[rt].fpr, &fs->fpr[rb].fpr,
507 					sizeof(double));
508 				a = (int *)&fs->fpr[rt].fpr;
509 				*a ^= (1U << 31);
510 				break;
511 			case	OPC63_MCRFS:
512 				FPU_EMU_EVCNT_INCR(mcrfs);
513 				DPRINTF(FPE_INSN, ("fpu_execute: MCRFS\n"));
514 				cond = 0;
515 				rt &= 0x1c;
516 				ra &= 0x1c;
517 				/* Extract the bits we want */
518 				mask = (fe->fe_fpscr >> (28 - ra)) & 0xf;
519 				/* Clear the bits we copied. */
520 				fe->fe_cx =
521 					(FPSR_EX_MSK | (0xf << (28 - ra)));
522 				fe->fe_fpscr &= fe->fe_cx;
523 				/* Now shove them in the right part of cr */
524 				tf->cr &= ~(0xf << (28 - rt));
525 				tf->cr |= (mask << (28 - rt));
526 				break;
527 			case	OPC63_MTFSB0:
528 				FPU_EMU_EVCNT_INCR(mtfsb0);
529 				DPRINTF(FPE_INSN, ("fpu_execute: MTFSB0\n"));
530 				fe->fe_fpscr &=
531 					((FPSCR_VX|FPSR_EX) & ~(1<<(31-rt)));
532 				break;
533 			case	OPC63_FMR:
534 				FPU_EMU_EVCNT_INCR(fmr);
535 				DPRINTF(FPE_INSN, ("fpu_execute: FMR\n"));
536 				memcpy(&fs->fpr[rt].fpr, &fs->fpr[rb].fpr,
537 					sizeof(double));
538 				break;
539 			case	OPC63_MTFSFI:
540 				FPU_EMU_EVCNT_INCR(mtfsfi);
541 				DPRINTF(FPE_INSN, ("fpu_execute: MTFSFI\n"));
542 				rb >>= 1;
543 				rt &= 0x1c; /* Already left-shifted 4 */
544 				fe->fe_cx = rb << (28 - rt);
545 				mask = 0xf<<(28 - rt);
546 				fe->fe_fpscr = (fe->fe_fpscr & ~mask) |
547 					fe->fe_cx;
548 /* XXX weird stuff about OX, FX, FEX, and VX should be handled */
549 				break;
550 			case	OPC63_FNABS:
551 				FPU_EMU_EVCNT_INCR(fnabs);
552 				DPRINTF(FPE_INSN, ("fpu_execute: FABS\n"));
553 				memcpy(&fs->fpr[rt].fpr, &fs->fpr[rb].fpr,
554 					sizeof(double));
555 				a = (int *)&fs->fpr[rt].fpr;
556 				*a |= (1U << 31);
557 				break;
558 			case	OPC63_FABS:
559 				FPU_EMU_EVCNT_INCR(fabs);
560 				DPRINTF(FPE_INSN, ("fpu_execute: FABS\n"));
561 				memcpy(&fs->fpr[rt].fpr, &fs->fpr[rb].fpr,
562 					sizeof(double));
563 				a = (int *)&fs->fpr[rt].fpr;
564 				*a &= ~(1U << 31);
565 				break;
566 			case	OPC63_MFFS:
567 				FPU_EMU_EVCNT_INCR(mffs);
568 				DPRINTF(FPE_INSN, ("fpu_execute: MFFS\n"));
569 				memcpy(&fs->fpr[rt].fpr, &fs->fpscr,
570 					sizeof(fs->fpscr));
571 				break;
572 			case	OPC63_MTFSF:
573 				FPU_EMU_EVCNT_INCR(mtfsf);
574 				DPRINTF(FPE_INSN, ("fpu_execute: MTFSF\n"));
575 				if ((rt = instr.i_xfl.i_flm) == -1)
576 					mask = -1;
577 				else {
578 					mask = 0;
579 					/* Convert 1 bit -> 4 bits */
580 					for (ra = 0; ra < 8; ra ++)
581 						if (rt & (1<<ra))
582 							mask |= (0xf<<(4*ra));
583 				}
584 				a = (int *)&fs->fpr[rt].fpr;
585 				fe->fe_cx = mask & a[1];
586 				fe->fe_fpscr = (fe->fe_fpscr&~mask) |
587 					(fe->fe_cx);
588 /* XXX weird stuff about OX, FX, FEX, and VX should be handled */
589 				break;
590 			case	OPC63_FCTID:
591 			case	OPC63_FCTIDZ:
592 				FPU_EMU_EVCNT_INCR(fctid);
593 				DPRINTF(FPE_INSN, ("fpu_execute: FCTID\n"));
594 				fpu_explode(fe, fp = &fe->fe_f1, type, rb);
595 				type = FTYPE_LNG;
596 				break;
597 			case	OPC63_FCFID:
598 				FPU_EMU_EVCNT_INCR(fcfid);
599 				DPRINTF(FPE_INSN, ("fpu_execute: FCFID\n"));
600 				type = FTYPE_LNG;
601 				fpu_explode(fe, fp = &fe->fe_f1, type, rb);
602 				type = FTYPE_DBL;
603 				break;
604 			default:
605 				return (NOTFPU);
606 				break;
607 			}
608 		} else {
609 			/* Format A */
610 			rt = instr.i_a.i_frt;
611 			ra = instr.i_a.i_fra;
612 			rb = instr.i_a.i_frb;
613 			rc = instr.i_a.i_frc;
614 
615 			/*
616 			 * All arithmetic operations work on registers, which
617 			 * are stored as doubles.
618 			 */
619 			type = FTYPE_DBL;
620 			switch ((unsigned int)instr.i_a.i_xo) {
621 			case	OPC59_FDIVS:
622 				FPU_EMU_EVCNT_INCR(fdiv);
623 				DPRINTF(FPE_INSN, ("fpu_execute: FDIV\n"));
624 				fpu_explode(fe, &fe->fe_f1, type, ra);
625 				fpu_explode(fe, &fe->fe_f2, type, rb);
626 				fp = fpu_div(fe);
627 				break;
628 			case	OPC59_FSUBS:
629 				FPU_EMU_EVCNT_INCR(fsub);
630 				DPRINTF(FPE_INSN, ("fpu_execute: FSUB\n"));
631 				fpu_explode(fe, &fe->fe_f1, type, ra);
632 				fpu_explode(fe, &fe->fe_f2, type, rb);
633 				fp = fpu_sub(fe);
634 				break;
635 			case	OPC59_FADDS:
636 				FPU_EMU_EVCNT_INCR(fadd);
637 				DPRINTF(FPE_INSN, ("fpu_execute: FADD\n"));
638 				fpu_explode(fe, &fe->fe_f1, type, ra);
639 				fpu_explode(fe, &fe->fe_f2, type, rb);
640 				fp = fpu_add(fe);
641 				break;
642 			case	OPC59_FSQRTS:
643 				FPU_EMU_EVCNT_INCR(fsqrt);
644 				DPRINTF(FPE_INSN, ("fpu_execute: FSQRT\n"));
645 				fpu_explode(fe, &fe->fe_f1, type, rb);
646 				fp = fpu_sqrt(fe);
647 				break;
648 			case	OPC63M_FSEL:
649 				FPU_EMU_EVCNT_INCR(fsel);
650 				DPRINTF(FPE_INSN, ("fpu_execute: FSEL\n"));
651 				a = (int *)&fe->fe_fpstate->fpr[ra].fpr;
652 				if ((*a & 0x80000000) && (*a & 0x7fffffff))
653 					/* fra < 0 */
654 					rc = rb;
655 				DPRINTF(FPE_INSN, ("f%d => f%d\n", rc, rt));
656 				memcpy(&fs->fpr[rt].fpr, &fs->fpr[rc].fpr,
657 					sizeof(double));
658 				break;
659 			case	OPC59_FRES:
660 				FPU_EMU_EVCNT_INCR(fpres);
661 				DPRINTF(FPE_INSN, ("fpu_execute: FPRES\n"));
662 				fpu_explode(fe, &fe->fe_f1, type, rb);
663 				fp = fpu_sqrt(fe);
664 				/* now we've gotta overwrite the dest reg */
665 				*((int *)&fe->fe_fpstate->fpr[rt].fpr) = 1;
666 				fpu_explode(fe, &fe->fe_f1, FTYPE_INT, rt);
667 				fpu_div(fe);
668 				break;
669 			case	OPC59_FMULS:
670 				FPU_EMU_EVCNT_INCR(fmul);
671 				DPRINTF(FPE_INSN, ("fpu_execute: FMUL\n"));
672 				fpu_explode(fe, &fe->fe_f1, type, ra);
673 				fpu_explode(fe, &fe->fe_f2, type, rc);
674 				fp = fpu_mul(fe);
675 				break;
676 			case	OPC63M_FRSQRTE:
677 				/* Reciprocal sqrt() estimate */
678 				FPU_EMU_EVCNT_INCR(frsqrte);
679 				DPRINTF(FPE_INSN, ("fpu_execute: FRSQRTE\n"));
680 				fpu_explode(fe, &fe->fe_f1, type, rb);
681 				fp = fpu_sqrt(fe);
682 				fe->fe_f2 = *fp;
683 				/* now we've gotta overwrite the dest reg */
684 				*((int *)&fe->fe_fpstate->fpr[rt].fpr) = 1;
685 				fpu_explode(fe, &fe->fe_f1, FTYPE_INT, rt);
686 				fpu_div(fe);
687 				break;
688 			case	OPC59_FMSUBS:
689 				FPU_EMU_EVCNT_INCR(fmulsub);
690 				DPRINTF(FPE_INSN, ("fpu_execute: FMULSUB\n"));
691 				fpu_explode(fe, &fe->fe_f1, type, ra);
692 				fpu_explode(fe, &fe->fe_f2, type, rc);
693 				fp = fpu_mul(fe);
694 				fe->fe_f1 = *fp;
695 				fpu_explode(fe, &fe->fe_f2, type, rb);
696 				fp = fpu_sub(fe);
697 				break;
698 			case	OPC59_FMADDS:
699 				FPU_EMU_EVCNT_INCR(fmuladd);
700 				DPRINTF(FPE_INSN, ("fpu_execute: FMULADD\n"));
701 				fpu_explode(fe, &fe->fe_f1, type, ra);
702 				fpu_explode(fe, &fe->fe_f2, type, rc);
703 				fp = fpu_mul(fe);
704 				fe->fe_f1 = *fp;
705 				fpu_explode(fe, &fe->fe_f2, type, rb);
706 				fp = fpu_add(fe);
707 				break;
708 			case	OPC59_FNMSUBS:
709 				FPU_EMU_EVCNT_INCR(fnmsub);
710 				DPRINTF(FPE_INSN, ("fpu_execute: FNMSUB\n"));
711 				fpu_explode(fe, &fe->fe_f1, type, ra);
712 				fpu_explode(fe, &fe->fe_f2, type, rc);
713 				fp = fpu_mul(fe);
714 				fe->fe_f1 = *fp;
715 				fpu_explode(fe, &fe->fe_f2, type, rb);
716 				fp = fpu_sub(fe);
717 				/* Negate */
718 				fp->fp_sign ^= 1;
719 				break;
720 			case	OPC59_FNMADDS:
721 				FPU_EMU_EVCNT_INCR(fnmadd);
722 				DPRINTF(FPE_INSN, ("fpu_execute: FNMADD\n"));
723 				fpu_explode(fe, &fe->fe_f1, type, ra);
724 				fpu_explode(fe, &fe->fe_f2, type, rc);
725 				fp = fpu_mul(fe);
726 				fe->fe_f1 = *fp;
727 				fpu_explode(fe, &fe->fe_f2, type, rb);
728 				fp = fpu_add(fe);
729 				/* Negate */
730 				fp->fp_sign ^= 1;
731 				break;
732 			default:
733 				return (NOTFPU);
734 				break;
735 			}
736 
737 			/* If the instruction was single precision, round */
738 			if (!(instr.i_any.i_opcd & 0x4)) {
739 				fpu_implode(fe, fp, FTYPE_SNG,
740 					(u_int *)&fs->fpr[rt].fpr);
741 				fpu_explode(fe, fp = &fe->fe_f1, FTYPE_SNG, rt);
742 			}
743 		}
744 	} else {
745 		return (NOTFPU);
746 	}
747 
748 	/*
749 	 * ALU operation is complete.  Collapse the result and then check
750 	 * for exceptions.  If we got any, and they are enabled, do not
751 	 * alter the destination register, just stop with an exception.
752 	 * Otherwise set new current exceptions and accrue.
753 	 */
754 	if (fp)
755 		fpu_implode(fe, fp, type, (u_int *)&fs->fpr[rt].fpr);
756 	cx = fe->fe_cx;
757 	fsr = fe->fe_fpscr;
758 	if (cx != 0) {
759 		fsr &= ~FPSCR_FX;
760 		if ((cx^fsr)&FPSR_EX_MSK)
761 			fsr |= FPSCR_FX;
762 		mask = fsr & FPSR_EX;
763 		mask <<= (25-3);
764 		if (cx & mask)
765 			fsr |= FPSCR_FEX;
766 		if (cx & FPSCR_FPRF) {
767 			/* Need to replace CC */
768 			fsr &= ~FPSCR_FPRF;
769 		}
770 		if (cx & (FPSR_EXOP))
771 			fsr |= FPSCR_VX;
772 		fsr |= cx;
773 		DPRINTF(FPE_INSN, ("fpu_execute: cx %x, fsr %x\n", cx, fsr));
774 	}
775 
776 	if (cond) {
777 		cond = fsr & 0xf0000000;
778 		/* Isolate condition codes */
779 		cond >>= 28;
780 		/* Move fpu condition codes to cr[1] */
781 		tf->cr &= (0x0f000000);
782 		tf->cr |= (cond<<24);
783 		DPRINTF(FPE_INSN, ("fpu_execute: cr[1] <= %x\n", cond));
784 	}
785 
786 	if (setcr) {
787 		cond = fsr & FPSCR_FPCC;
788 		/* Isolate condition codes */
789 		cond <<= 16;
790 		/* Move fpu condition codes to cr[1] */
791 		tf->cr &= ~(0xf0000000>>bf);
792 		tf->cr |= (cond>>bf);
793 		DPRINTF(FPE_INSN, ("fpu_execute: cr[%d] (cr=%jx) <= %x\n",
794 			bf/4, (uintmax_t)tf->cr, cond));
795 	}
796 
797 	((int *)&fs->fpscr)[1] = fsr;
798 	if (fsr & FPSCR_FEX)
799 		return(FPE);
800 	return (0);	/* success */
801 }
802