xref: /freebsd/sys/powerpc/fpu/fpu_emu.c (revision 66df505066f51e6d8411b966765d828817f88971)
1 /*	$NetBSD: fpu_emu.c,v 1.14 2005/12/11 12:18:42 christos Exp $ */
2 
3 /*-
4  * SPDX-License-Identifier: BSD-4-Clause
5  *
6  * Copyright 2001 Wasabi Systems, Inc.
7  * All rights reserved.
8  *
9  * Written by Eduardo Horvath and Simon Burge for Wasabi Systems, Inc.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  * 3. All advertising materials mentioning features or use of this software
20  *    must display the following acknowledgement:
21  *      This product includes software developed for the NetBSD Project by
22  *      Wasabi Systems, Inc.
23  * 4. The name of Wasabi Systems, Inc. may not be used to endorse
24  *    or promote products derived from this software without specific prior
25  *    written permission.
26  *
27  * THIS SOFTWARE IS PROVIDED BY WASABI SYSTEMS, INC. ``AS IS'' AND
28  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
29  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
30  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL WASABI SYSTEMS, INC
31  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
32  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
33  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
34  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
35  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
36  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
37  * POSSIBILITY OF SUCH DAMAGE.
38  */
39 
40 /*
41  * Copyright (c) 1992, 1993
42  *	The Regents of the University of California.  All rights reserved.
43  *
44  * This software was developed by the Computer Systems Engineering group
45  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
46  * contributed to Berkeley.
47  *
48  * All advertising materials mentioning features or use of this software
49  * must display the following acknowledgement:
50  *	This product includes software developed by the University of
51  *	California, Lawrence Berkeley Laboratory.
52  *
53  * Redistribution and use in source and binary forms, with or without
54  * modification, are permitted provided that the following conditions
55  * are met:
56  * 1. Redistributions of source code must retain the above copyright
57  *    notice, this list of conditions and the following disclaimer.
58  * 2. Redistributions in binary form must reproduce the above copyright
59  *    notice, this list of conditions and the following disclaimer in the
60  *    documentation and/or other materials provided with the distribution.
61  * 3. Neither the name of the University nor the names of its contributors
62  *    may be used to endorse or promote products derived from this software
63  *    without specific prior written permission.
64  *
65  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
66  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
67  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
68  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
69  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
70  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
71  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
72  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
73  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
74  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
75  * SUCH DAMAGE.
76  *
77  *	@(#)fpu.c	8.1 (Berkeley) 6/11/93
78  */
79 
80 #include <sys/cdefs.h>
81 __FBSDID("$FreeBSD$");
82 
83 #include "opt_ddb.h"
84 
85 #include <sys/param.h>
86 #include <sys/systm.h>
87 #include <sys/kdb.h>
88 #include <sys/kernel.h>
89 #include <sys/proc.h>
90 #include <sys/sysctl.h>
91 #include <sys/signal.h>
92 #include <sys/syslog.h>
93 #include <sys/signalvar.h>
94 
95 #include <machine/fpu.h>
96 
97 #include <powerpc/fpu/fpu_emu.h>
98 #include <powerpc/fpu/fpu_extern.h>
99 #include <powerpc/fpu/fpu_instr.h>
100 
101 static SYSCTL_NODE(_hw, OID_AUTO, fpu_emu, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
102     "FPU emulator");
103 
104 #define	FPU_EMU_EVCNT_DECL(name)					\
105 static u_int fpu_emu_evcnt_##name;					\
106 SYSCTL_INT(_hw_fpu_emu, OID_AUTO, evcnt_##name, CTLFLAG_RD,		\
107     &fpu_emu_evcnt_##name, 0, "")
108 
109 #define	FPU_EMU_EVCNT_INCR(name)	fpu_emu_evcnt_##name++
110 
111 FPU_EMU_EVCNT_DECL(stfiwx);
112 FPU_EMU_EVCNT_DECL(fpstore);
113 FPU_EMU_EVCNT_DECL(fpload);
114 FPU_EMU_EVCNT_DECL(fcmpu);
115 FPU_EMU_EVCNT_DECL(frsp);
116 FPU_EMU_EVCNT_DECL(fctiw);
117 FPU_EMU_EVCNT_DECL(fcmpo);
118 FPU_EMU_EVCNT_DECL(mtfsb1);
119 FPU_EMU_EVCNT_DECL(fnegabs);
120 FPU_EMU_EVCNT_DECL(mcrfs);
121 FPU_EMU_EVCNT_DECL(mtfsb0);
122 FPU_EMU_EVCNT_DECL(fmr);
123 FPU_EMU_EVCNT_DECL(mtfsfi);
124 FPU_EMU_EVCNT_DECL(fnabs);
125 FPU_EMU_EVCNT_DECL(fabs);
126 FPU_EMU_EVCNT_DECL(mffs);
127 FPU_EMU_EVCNT_DECL(mtfsf);
128 FPU_EMU_EVCNT_DECL(fctid);
129 FPU_EMU_EVCNT_DECL(fcfid);
130 FPU_EMU_EVCNT_DECL(fdiv);
131 FPU_EMU_EVCNT_DECL(fsub);
132 FPU_EMU_EVCNT_DECL(fadd);
133 FPU_EMU_EVCNT_DECL(fsqrt);
134 FPU_EMU_EVCNT_DECL(fsel);
135 FPU_EMU_EVCNT_DECL(fpres);
136 FPU_EMU_EVCNT_DECL(fmul);
137 FPU_EMU_EVCNT_DECL(frsqrte);
138 FPU_EMU_EVCNT_DECL(fmulsub);
139 FPU_EMU_EVCNT_DECL(fmuladd);
140 FPU_EMU_EVCNT_DECL(fnmsub);
141 FPU_EMU_EVCNT_DECL(fnmadd);
142 
143 /* FPSR exception masks */
144 #define FPSR_EX_MSK	(FPSCR_VX|FPSCR_OX|FPSCR_UX|FPSCR_ZX|		\
145 			FPSCR_XX|FPSCR_VXSNAN|FPSCR_VXISI|FPSCR_VXIDI|	\
146 			FPSCR_VXZDZ|FPSCR_VXIMZ|FPSCR_VXVC|FPSCR_VXSOFT|\
147 			FPSCR_VXSQRT|FPSCR_VXCVI)
148 #define	FPSR_EX		(FPSCR_VE|FPSCR_OE|FPSCR_UE|FPSCR_ZE|FPSCR_XE)
149 #define	FPSR_EXOP	(FPSR_EX_MSK&(~FPSR_EX))
150 
151 int fpe_debug = 0;
152 
153 #ifdef DEBUG
154 vm_offset_t opc_disasm(vm_offset_t, int);
155 
156 /*
157  * Dump a `fpn' structure.
158  */
159 void
160 fpu_dumpfpn(struct fpn *fp)
161 {
162 	static const char *class[] = {
163 		"SNAN", "QNAN", "ZERO", "NUM", "INF"
164 	};
165 
166 	printf("%s %c.%x %x %x %xE%d", class[fp->fp_class + 2],
167 		fp->fp_sign ? '-' : ' ',
168 		fp->fp_mant[0],	fp->fp_mant[1],
169 		fp->fp_mant[2], fp->fp_mant[3],
170 		fp->fp_exp);
171 }
172 #endif
173 
174 /*
175  * fpu_execute returns the following error numbers (0 = no error):
176  */
177 #define	FPE		1	/* take a floating point exception */
178 #define	NOTFPU		2	/* not an FPU instruction */
179 #define	FAULT		3
180 
181 /*
182  * Emulate a floating-point instruction.
183  * Return zero for success, else signal number.
184  * (Typically: zero, SIGFPE, SIGILL, SIGSEGV)
185  */
186 int
187 fpu_emulate(struct trapframe *frame, struct fpu *fpf)
188 {
189 	union instr insn;
190 	struct fpemu fe;
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 	if ((insn.i_any.i_opcd == OPC_TWI) ||
210 	    ((insn.i_any.i_opcd == OPC_integer_31) &&
211 	    (insn.i_x.i_xo == OPC31_TW))) {
212 		/* Check for the two trap insns. */
213 		DPRINTF(FPE_EX, ("fpu_emulate: SIGTRAP\n"));
214 		return (SIGTRAP);
215 	}
216 	sig = 0;
217 	switch (fpu_execute(frame, &fe, &insn)) {
218 	case 0:
219 		DPRINTF(FPE_EX, ("fpu_emulate: success\n"));
220 		frame->srr0 += 4;
221 		break;
222 
223 	case FPE:
224 		DPRINTF(FPE_EX, ("fpu_emulate: SIGFPE\n"));
225 		sig = SIGFPE;
226 		break;
227 
228 	case FAULT:
229 		DPRINTF(FPE_EX, ("fpu_emulate: SIGSEGV\n"));
230 		sig = SIGSEGV;
231 		break;
232 
233 	case NOTFPU:
234 	default:
235 		DPRINTF(FPE_EX, ("fpu_emulate: SIGILL\n"));
236 #ifdef DEBUG
237 		if (fpe_debug & FPE_EX) {
238 			printf("fpu_emulate:  illegal insn %x at %p:",
239 			insn.i_int, (void *) (frame->srr0));
240 			opc_disasm(frame->srr0, insn.i_int);
241 		}
242 #endif
243 		sig = SIGILL;
244 #ifdef DEBUG
245 		if (fpe_debug & FPE_EX)
246 			kdb_enter(KDB_WHY_UNSET, "illegal instruction");
247 #endif
248 		break;
249 	}
250 
251 	return (sig);
252 }
253 
254 /*
255  * Execute an FPU instruction (one that runs entirely in the FPU; not
256  * FBfcc or STF, for instance).  On return, fe->fe_fs->fs_fsr will be
257  * modified to reflect the setting the hardware would have left.
258  *
259  * Note that we do not catch all illegal opcodes, so you can, for instance,
260  * multiply two integers this way.
261  */
262 int
263 fpu_execute(struct trapframe *tf, struct fpemu *fe, union instr *insn)
264 {
265 	struct fpn *fp;
266 	union instr instr = *insn;
267 	int *a;
268 	vm_offset_t addr;
269 	int ra, rb, rc, rt, type, mask, fsr, cx, bf, setcr;
270 	unsigned int cond;
271 	struct fpu *fs;
272 
273 	/* Setup work. */
274 	fp = NULL;
275 	fs = fe->fe_fpstate;
276 	fe->fe_fpscr = ((int *)&fs->fpscr)[1];
277 
278 	/*
279 	 * On PowerPC all floating point values are stored in registers
280 	 * as doubles, even when used for single precision operations.
281 	 */
282 	type = FTYPE_DBL;
283 	cond = instr.i_any.i_rc;
284 	setcr = 0;
285 	bf = 0;	/* XXX gcc */
286 
287 #if defined(DDB) && defined(DEBUG)
288 	if (fpe_debug & FPE_EX) {
289 		vm_offset_t loc = tf->srr0;
290 
291 		printf("Trying to emulate: %p ", (void *)loc);
292 		opc_disasm(loc, instr.i_int);
293 	}
294 #endif
295 
296 	/*
297 	 * `Decode' and execute instruction.
298 	 */
299 
300 	if ((instr.i_any.i_opcd >= OPC_LFS && instr.i_any.i_opcd <= OPC_STFDU) ||
301 	    instr.i_any.i_opcd == OPC_integer_31) {
302 		/*
303 		 * Handle load/store insns:
304 		 *
305 		 * Convert to/from single if needed, calculate addr,
306 		 * and update index reg if needed.
307 		 */
308 		double buf;
309 		size_t size = sizeof(float);
310 		int store, update;
311 
312 		cond = 0; /* ld/st never set condition codes */
313 
314 		if (instr.i_any.i_opcd == OPC_integer_31) {
315 			if (instr.i_x.i_xo == OPC31_STFIWX) {
316 				FPU_EMU_EVCNT_INCR(stfiwx);
317 
318 				/* Store as integer */
319 				ra = instr.i_x.i_ra;
320 				rb = instr.i_x.i_rb;
321 				DPRINTF(FPE_INSN,
322 					("reg %d has %jx reg %d has %jx\n",
323 					ra, (uintmax_t)tf->fixreg[ra], rb,
324 					(uintmax_t)tf->fixreg[rb]));
325 
326 				addr = tf->fixreg[rb];
327 				if (ra != 0)
328 					addr += tf->fixreg[ra];
329 				rt = instr.i_x.i_rt;
330 				a = (int *)&fs->fpr[rt].fpr;
331 				DPRINTF(FPE_INSN,
332 					("fpu_execute: Store INT %x at %p\n",
333 						a[1], (void *)addr));
334 				if (copyout(&a[1], (void *)addr, sizeof(int)))
335 					return (FAULT);
336 				return (0);
337 			}
338 
339 			if ((instr.i_x.i_xo & OPC31_FPMASK) != OPC31_FPOP)
340 				/* Not an indexed FP load/store op */
341 				return (NOTFPU);
342 
343 			store = (instr.i_x.i_xo & 0x80);
344 			if (instr.i_x.i_xo & 0x40)
345 				size = sizeof(double);
346 			else
347 				type = FTYPE_SNG;
348 			update = (instr.i_x.i_xo & 0x20);
349 
350 			/* calculate EA of load/store */
351 			ra = instr.i_x.i_ra;
352 			rb = instr.i_x.i_rb;
353 			DPRINTF(FPE_INSN, ("reg %d has %jx reg %d has %jx\n",
354 				ra, (uintmax_t)tf->fixreg[ra], rb,
355 				(uintmax_t)tf->fixreg[rb]));
356 			addr = tf->fixreg[rb];
357 			if (ra != 0)
358 				addr += tf->fixreg[ra];
359 			rt = instr.i_x.i_rt;
360 		} else {
361 			store = instr.i_d.i_opcd & 0x4;
362 			if (instr.i_d.i_opcd & 0x2)
363 				size = sizeof(double);
364 			else
365 				type = FTYPE_SNG;
366 			update = instr.i_d.i_opcd & 0x1;
367 
368 			/* calculate EA of load/store */
369 			ra = instr.i_d.i_ra;
370 			addr = instr.i_d.i_d;
371 			DPRINTF(FPE_INSN, ("reg %d has %jx displ %jx\n",
372 				ra, (uintmax_t)tf->fixreg[ra],
373 				(uintmax_t)addr));
374 			if (ra != 0)
375 				addr += tf->fixreg[ra];
376 			rt = instr.i_d.i_rt;
377 		}
378 
379 		if (update && ra == 0)
380 			return (NOTFPU);
381 
382 		if (store) {
383 			/* Store */
384 			FPU_EMU_EVCNT_INCR(fpstore);
385 			if (type != FTYPE_DBL) {
386 				DPRINTF(FPE_INSN,
387 					("fpu_execute: Store SNG at %p\n",
388 						(void *)addr));
389 				fpu_explode(fe, fp = &fe->fe_f1, FTYPE_DBL, rt);
390 				fpu_implode(fe, fp, type, (void *)&buf);
391 				if (copyout(&buf, (void *)addr, size))
392 					return (FAULT);
393 			} else {
394 				DPRINTF(FPE_INSN,
395 					("fpu_execute: Store DBL at %p\n",
396 						(void *)addr));
397 				if (copyout(&fs->fpr[rt].fpr, (void *)addr,
398 				    size))
399 					return (FAULT);
400 			}
401 		} else {
402 			/* Load */
403 			FPU_EMU_EVCNT_INCR(fpload);
404 			DPRINTF(FPE_INSN, ("fpu_execute: Load from %p\n",
405 				(void *)addr));
406 			if (copyin((const void *)addr, &fs->fpr[rt].fpr,
407 			    size))
408 				return (FAULT);
409 			if (type != FTYPE_DBL) {
410 				fpu_explode(fe, fp = &fe->fe_f1, type, rt);
411 				fpu_implode(fe, fp, FTYPE_DBL,
412 					(u_int *)&fs->fpr[rt].fpr);
413 			}
414 		}
415 		if (update)
416 			tf->fixreg[ra] = addr;
417 		/* Complete. */
418 		return (0);
419 #ifdef notyet
420 	} else if (instr.i_any.i_opcd == OPC_load_st_62) {
421 		/* These are 64-bit extensions */
422 		return (NOTFPU);
423 #endif
424 	} else if (instr.i_any.i_opcd == OPC_sp_fp_59 ||
425 		instr.i_any.i_opcd == OPC_dp_fp_63) {
426 		if (instr.i_any.i_opcd == OPC_dp_fp_63 &&
427 		    !(instr.i_a.i_xo & OPC63M_MASK)) {
428 			/* Format X */
429 			rt = instr.i_x.i_rt;
430 			ra = instr.i_x.i_ra;
431 			rb = instr.i_x.i_rb;
432 
433 			/* One of the special opcodes.... */
434 			switch (instr.i_x.i_xo) {
435 			case	OPC63_FCMPU:
436 				FPU_EMU_EVCNT_INCR(fcmpu);
437 				DPRINTF(FPE_INSN, ("fpu_execute: FCMPU\n"));
438 				rt >>= 2;
439 				fpu_explode(fe, &fe->fe_f1, type, ra);
440 				fpu_explode(fe, &fe->fe_f2, type, rb);
441 				fpu_compare(fe, 0);
442 				/* Make sure we do the condition regs. */
443 				cond = 0;
444 				/* N.B.: i_rs is already left shifted by two. */
445 				bf = instr.i_x.i_rs & 0xfc;
446 				setcr = 1;
447 				break;
448 
449 			case	OPC63_FRSP:
450 				/*
451 				 * Convert to single:
452 				 *
453 				 * PowerPC uses this to round a double
454 				 * precision value to single precision,
455 				 * but values in registers are always
456 				 * stored in double precision format.
457 				 */
458 				FPU_EMU_EVCNT_INCR(frsp);
459 				DPRINTF(FPE_INSN, ("fpu_execute: FRSP\n"));
460 				fpu_explode(fe, fp = &fe->fe_f1, FTYPE_DBL, rb);
461 				fpu_implode(fe, fp, FTYPE_SNG,
462 					(u_int *)&fs->fpr[rt].fpr);
463 				fpu_explode(fe, fp = &fe->fe_f1, FTYPE_SNG, rt);
464 				type = FTYPE_DBL;
465 				break;
466 			case	OPC63_FCTIW:
467 			case	OPC63_FCTIWZ:
468 				FPU_EMU_EVCNT_INCR(fctiw);
469 				DPRINTF(FPE_INSN, ("fpu_execute: FCTIW\n"));
470 				fpu_explode(fe, fp = &fe->fe_f1, type, rb);
471 				type = FTYPE_INT;
472 				break;
473 			case	OPC63_FCMPO:
474 				FPU_EMU_EVCNT_INCR(fcmpo);
475 				DPRINTF(FPE_INSN, ("fpu_execute: FCMPO\n"));
476 				rt >>= 2;
477 				fpu_explode(fe, &fe->fe_f1, type, ra);
478 				fpu_explode(fe, &fe->fe_f2, type, rb);
479 				fpu_compare(fe, 1);
480 				/* Make sure we do the condition regs. */
481 				cond = 0;
482 				/* N.B.: i_rs is already left shifted by two. */
483 				bf = instr.i_x.i_rs & 0xfc;
484 				setcr = 1;
485 				break;
486 			case	OPC63_MTFSB1:
487 				FPU_EMU_EVCNT_INCR(mtfsb1);
488 				DPRINTF(FPE_INSN, ("fpu_execute: MTFSB1\n"));
489 				fe->fe_fpscr |=
490 					(~(FPSCR_VX|FPSR_EX) & (1<<(31-rt)));
491 				break;
492 			case	OPC63_FNEG:
493 				FPU_EMU_EVCNT_INCR(fnegabs);
494 				DPRINTF(FPE_INSN, ("fpu_execute: FNEGABS\n"));
495 				memcpy(&fs->fpr[rt].fpr, &fs->fpr[rb].fpr,
496 					sizeof(double));
497 				a = (int *)&fs->fpr[rt].fpr;
498 				*a ^= (1U << 31);
499 				break;
500 			case	OPC63_MCRFS:
501 				FPU_EMU_EVCNT_INCR(mcrfs);
502 				DPRINTF(FPE_INSN, ("fpu_execute: MCRFS\n"));
503 				cond = 0;
504 				rt &= 0x1c;
505 				ra &= 0x1c;
506 				/* Extract the bits we want */
507 				mask = (fe->fe_fpscr >> (28 - ra)) & 0xf;
508 				/* Clear the bits we copied. */
509 				fe->fe_cx =
510 					(FPSR_EX_MSK | (0xf << (28 - ra)));
511 				fe->fe_fpscr &= fe->fe_cx;
512 				/* Now shove them in the right part of cr */
513 				tf->cr &= ~(0xf << (28 - rt));
514 				tf->cr |= (mask << (28 - rt));
515 				break;
516 			case	OPC63_MTFSB0:
517 				FPU_EMU_EVCNT_INCR(mtfsb0);
518 				DPRINTF(FPE_INSN, ("fpu_execute: MTFSB0\n"));
519 				fe->fe_fpscr &=
520 					((FPSCR_VX|FPSR_EX) & ~(1<<(31-rt)));
521 				break;
522 			case	OPC63_FMR:
523 				FPU_EMU_EVCNT_INCR(fmr);
524 				DPRINTF(FPE_INSN, ("fpu_execute: FMR\n"));
525 				memcpy(&fs->fpr[rt].fpr, &fs->fpr[rb].fpr,
526 					sizeof(double));
527 				break;
528 			case	OPC63_MTFSFI:
529 				FPU_EMU_EVCNT_INCR(mtfsfi);
530 				DPRINTF(FPE_INSN, ("fpu_execute: MTFSFI\n"));
531 				rb >>= 1;
532 				rt &= 0x1c; /* Already left-shifted 4 */
533 				fe->fe_cx = rb << (28 - rt);
534 				mask = 0xf<<(28 - rt);
535 				fe->fe_fpscr = (fe->fe_fpscr & ~mask) |
536 					fe->fe_cx;
537 /* XXX weird stuff about OX, FX, FEX, and VX should be handled */
538 				break;
539 			case	OPC63_FNABS:
540 				FPU_EMU_EVCNT_INCR(fnabs);
541 				DPRINTF(FPE_INSN, ("fpu_execute: FABS\n"));
542 				memcpy(&fs->fpr[rt].fpr, &fs->fpr[rb].fpr,
543 					sizeof(double));
544 				a = (int *)&fs->fpr[rt].fpr;
545 				*a |= (1U << 31);
546 				break;
547 			case	OPC63_FABS:
548 				FPU_EMU_EVCNT_INCR(fabs);
549 				DPRINTF(FPE_INSN, ("fpu_execute: FABS\n"));
550 				memcpy(&fs->fpr[rt].fpr, &fs->fpr[rb].fpr,
551 					sizeof(double));
552 				a = (int *)&fs->fpr[rt].fpr;
553 				*a &= ~(1U << 31);
554 				break;
555 			case	OPC63_MFFS:
556 				FPU_EMU_EVCNT_INCR(mffs);
557 				DPRINTF(FPE_INSN, ("fpu_execute: MFFS\n"));
558 				memcpy(&fs->fpr[rt].fpr, &fs->fpscr,
559 					sizeof(fs->fpscr));
560 				break;
561 			case	OPC63_MTFSF:
562 				FPU_EMU_EVCNT_INCR(mtfsf);
563 				DPRINTF(FPE_INSN, ("fpu_execute: MTFSF\n"));
564 				if ((rt = instr.i_xfl.i_flm) == -1)
565 					mask = -1;
566 				else {
567 					mask = 0;
568 					/* Convert 1 bit -> 4 bits */
569 					for (ra = 0; ra < 8; ra ++)
570 						if (rt & (1<<ra))
571 							mask |= (0xf<<(4*ra));
572 				}
573 				a = (int *)&fs->fpr[rt].fpr;
574 				fe->fe_cx = mask & a[1];
575 				fe->fe_fpscr = (fe->fe_fpscr&~mask) |
576 					(fe->fe_cx);
577 /* XXX weird stuff about OX, FX, FEX, and VX should be handled */
578 				break;
579 			case	OPC63_FCTID:
580 			case	OPC63_FCTIDZ:
581 				FPU_EMU_EVCNT_INCR(fctid);
582 				DPRINTF(FPE_INSN, ("fpu_execute: FCTID\n"));
583 				fpu_explode(fe, fp = &fe->fe_f1, type, rb);
584 				type = FTYPE_LNG;
585 				break;
586 			case	OPC63_FCFID:
587 				FPU_EMU_EVCNT_INCR(fcfid);
588 				DPRINTF(FPE_INSN, ("fpu_execute: FCFID\n"));
589 				type = FTYPE_LNG;
590 				fpu_explode(fe, fp = &fe->fe_f1, type, rb);
591 				type = FTYPE_DBL;
592 				break;
593 			default:
594 				return (NOTFPU);
595 				break;
596 			}
597 		} else {
598 			/* Format A */
599 			rt = instr.i_a.i_frt;
600 			ra = instr.i_a.i_fra;
601 			rb = instr.i_a.i_frb;
602 			rc = instr.i_a.i_frc;
603 
604 			/*
605 			 * All arithmetic operations work on registers, which
606 			 * are stored as doubles.
607 			 */
608 			type = FTYPE_DBL;
609 			switch ((unsigned int)instr.i_a.i_xo) {
610 			case	OPC59_FDIVS:
611 				FPU_EMU_EVCNT_INCR(fdiv);
612 				DPRINTF(FPE_INSN, ("fpu_execute: FDIV\n"));
613 				fpu_explode(fe, &fe->fe_f1, type, ra);
614 				fpu_explode(fe, &fe->fe_f2, type, rb);
615 				fp = fpu_div(fe);
616 				break;
617 			case	OPC59_FSUBS:
618 				FPU_EMU_EVCNT_INCR(fsub);
619 				DPRINTF(FPE_INSN, ("fpu_execute: FSUB\n"));
620 				fpu_explode(fe, &fe->fe_f1, type, ra);
621 				fpu_explode(fe, &fe->fe_f2, type, rb);
622 				fp = fpu_sub(fe);
623 				break;
624 			case	OPC59_FADDS:
625 				FPU_EMU_EVCNT_INCR(fadd);
626 				DPRINTF(FPE_INSN, ("fpu_execute: FADD\n"));
627 				fpu_explode(fe, &fe->fe_f1, type, ra);
628 				fpu_explode(fe, &fe->fe_f2, type, rb);
629 				fp = fpu_add(fe);
630 				break;
631 			case	OPC59_FSQRTS:
632 				FPU_EMU_EVCNT_INCR(fsqrt);
633 				DPRINTF(FPE_INSN, ("fpu_execute: FSQRT\n"));
634 				fpu_explode(fe, &fe->fe_f1, type, rb);
635 				fp = fpu_sqrt(fe);
636 				break;
637 			case	OPC63M_FSEL:
638 				FPU_EMU_EVCNT_INCR(fsel);
639 				DPRINTF(FPE_INSN, ("fpu_execute: FSEL\n"));
640 				a = (int *)&fe->fe_fpstate->fpr[ra].fpr;
641 				if ((*a & 0x80000000) && (*a & 0x7fffffff))
642 					/* fra < 0 */
643 					rc = rb;
644 				DPRINTF(FPE_INSN, ("f%d => f%d\n", rc, rt));
645 				memcpy(&fs->fpr[rt].fpr, &fs->fpr[rc].fpr,
646 					sizeof(double));
647 				break;
648 			case	OPC59_FRES:
649 				FPU_EMU_EVCNT_INCR(fpres);
650 				DPRINTF(FPE_INSN, ("fpu_execute: FPRES\n"));
651 				fpu_explode(fe, &fe->fe_f1, type, rb);
652 				fp = fpu_sqrt(fe);
653 				/* now we've gotta overwrite the dest reg */
654 				*((int *)&fe->fe_fpstate->fpr[rt].fpr) = 1;
655 				fpu_explode(fe, &fe->fe_f1, FTYPE_INT, rt);
656 				fpu_div(fe);
657 				break;
658 			case	OPC59_FMULS:
659 				FPU_EMU_EVCNT_INCR(fmul);
660 				DPRINTF(FPE_INSN, ("fpu_execute: FMUL\n"));
661 				fpu_explode(fe, &fe->fe_f1, type, ra);
662 				fpu_explode(fe, &fe->fe_f2, type, rc);
663 				fp = fpu_mul(fe);
664 				break;
665 			case	OPC63M_FRSQRTE:
666 				/* Reciprocal sqrt() estimate */
667 				FPU_EMU_EVCNT_INCR(frsqrte);
668 				DPRINTF(FPE_INSN, ("fpu_execute: FRSQRTE\n"));
669 				fpu_explode(fe, &fe->fe_f1, type, rb);
670 				fp = fpu_sqrt(fe);
671 				fe->fe_f2 = *fp;
672 				/* now we've gotta overwrite the dest reg */
673 				*((int *)&fe->fe_fpstate->fpr[rt].fpr) = 1;
674 				fpu_explode(fe, &fe->fe_f1, FTYPE_INT, rt);
675 				fpu_div(fe);
676 				break;
677 			case	OPC59_FMSUBS:
678 				FPU_EMU_EVCNT_INCR(fmulsub);
679 				DPRINTF(FPE_INSN, ("fpu_execute: FMULSUB\n"));
680 				fpu_explode(fe, &fe->fe_f1, type, ra);
681 				fpu_explode(fe, &fe->fe_f2, type, rc);
682 				fp = fpu_mul(fe);
683 				fe->fe_f1 = *fp;
684 				fpu_explode(fe, &fe->fe_f2, type, rb);
685 				fp = fpu_sub(fe);
686 				break;
687 			case	OPC59_FMADDS:
688 				FPU_EMU_EVCNT_INCR(fmuladd);
689 				DPRINTF(FPE_INSN, ("fpu_execute: FMULADD\n"));
690 				fpu_explode(fe, &fe->fe_f1, type, ra);
691 				fpu_explode(fe, &fe->fe_f2, type, rc);
692 				fp = fpu_mul(fe);
693 				fe->fe_f1 = *fp;
694 				fpu_explode(fe, &fe->fe_f2, type, rb);
695 				fp = fpu_add(fe);
696 				break;
697 			case	OPC59_FNMSUBS:
698 				FPU_EMU_EVCNT_INCR(fnmsub);
699 				DPRINTF(FPE_INSN, ("fpu_execute: FNMSUB\n"));
700 				fpu_explode(fe, &fe->fe_f1, type, ra);
701 				fpu_explode(fe, &fe->fe_f2, type, rc);
702 				fp = fpu_mul(fe);
703 				fe->fe_f1 = *fp;
704 				fpu_explode(fe, &fe->fe_f2, type, rb);
705 				fp = fpu_sub(fe);
706 				/* Negate */
707 				fp->fp_sign ^= 1;
708 				break;
709 			case	OPC59_FNMADDS:
710 				FPU_EMU_EVCNT_INCR(fnmadd);
711 				DPRINTF(FPE_INSN, ("fpu_execute: FNMADD\n"));
712 				fpu_explode(fe, &fe->fe_f1, type, ra);
713 				fpu_explode(fe, &fe->fe_f2, type, rc);
714 				fp = fpu_mul(fe);
715 				fe->fe_f1 = *fp;
716 				fpu_explode(fe, &fe->fe_f2, type, rb);
717 				fp = fpu_add(fe);
718 				/* Negate */
719 				fp->fp_sign ^= 1;
720 				break;
721 			default:
722 				return (NOTFPU);
723 				break;
724 			}
725 
726 			/* If the instruction was single precision, round */
727 			if (!(instr.i_any.i_opcd & 0x4)) {
728 				fpu_implode(fe, fp, FTYPE_SNG,
729 					(u_int *)&fs->fpr[rt].fpr);
730 				fpu_explode(fe, fp = &fe->fe_f1, FTYPE_SNG, rt);
731 			}
732 		}
733 	} else {
734 		return (NOTFPU);
735 	}
736 
737 	/*
738 	 * ALU operation is complete.  Collapse the result and then check
739 	 * for exceptions.  If we got any, and they are enabled, do not
740 	 * alter the destination register, just stop with an exception.
741 	 * Otherwise set new current exceptions and accrue.
742 	 */
743 	if (fp)
744 		fpu_implode(fe, fp, type, (u_int *)&fs->fpr[rt].fpr);
745 	cx = fe->fe_cx;
746 	fsr = fe->fe_fpscr;
747 	if (cx != 0) {
748 		fsr &= ~FPSCR_FX;
749 		if ((cx^fsr)&FPSR_EX_MSK)
750 			fsr |= FPSCR_FX;
751 		mask = fsr & FPSR_EX;
752 		mask <<= (25-3);
753 		if (cx & mask)
754 			fsr |= FPSCR_FEX;
755 		if (cx & FPSCR_FPRF) {
756 			/* Need to replace CC */
757 			fsr &= ~FPSCR_FPRF;
758 		}
759 		if (cx & (FPSR_EXOP))
760 			fsr |= FPSCR_VX;
761 		fsr |= cx;
762 		DPRINTF(FPE_INSN, ("fpu_execute: cx %x, fsr %x\n", cx, fsr));
763 	}
764 
765 	if (cond) {
766 		cond = fsr & 0xf0000000;
767 		/* Isolate condition codes */
768 		cond >>= 28;
769 		/* Move fpu condition codes to cr[1] */
770 		tf->cr &= (0x0f000000);
771 		tf->cr |= (cond<<24);
772 		DPRINTF(FPE_INSN, ("fpu_execute: cr[1] <= %x\n", cond));
773 	}
774 
775 	if (setcr) {
776 		cond = fsr & FPSCR_FPCC;
777 		/* Isolate condition codes */
778 		cond <<= 16;
779 		/* Move fpu condition codes to cr[1] */
780 		tf->cr &= ~(0xf0000000>>bf);
781 		tf->cr |= (cond>>bf);
782 		DPRINTF(FPE_INSN, ("fpu_execute: cr[%d] (cr=%jx) <= %x\n",
783 			bf/4, (uintmax_t)tf->cr, cond));
784 	}
785 
786 	((int *)&fs->fpscr)[1] = fsr;
787 	if (fsr & FPSCR_FEX)
788 		return(FPE);
789 	return (0);	/* success */
790 }
791