1 /*	$OpenBSD: fpu.c,v 1.11 2003/06/02 23:27:54 millert Exp $	*/
2 /*	$NetBSD: fpu.c,v 1.6 1997/07/29 10:09:51 fair Exp $	*/
3 
4 /*
5  * Copyright (c) 1992, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  *
8  * This software was developed by the Computer Systems Engineering group
9  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
10  * contributed to Berkeley.
11  *
12  * All advertising materials mentioning features or use of this software
13  * must display the following acknowledgement:
14  *	This product includes software developed by the University of
15  *	California, Lawrence Berkeley Laboratory.
16  *
17  * Redistribution and use in source and binary forms, with or without
18  * modification, are permitted provided that the following conditions
19  * are met:
20  * 1. Redistributions of source code must retain the above copyright
21  *    notice, this list of conditions and the following disclaimer.
22  * 2. Redistributions in binary form must reproduce the above copyright
23  *    notice, this list of conditions and the following disclaimer in the
24  *    documentation and/or other materials provided with the distribution.
25  * 3. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	@(#)fpu.c	8.1 (Berkeley) 6/11/93
42  */
43 
44 #include <sys/param.h>
45 #include <sys/proc.h>
46 #include <sys/signal.h>
47 #include <sys/systm.h>
48 #include <sys/syslog.h>
49 #include <sys/signalvar.h>
50 
51 #include <machine/instr.h>
52 #include <machine/reg.h>
53 
54 #include <sparc/fpu/fpu_emu.h>
55 #include <sparc/fpu/fpu_extern.h>
56 
57 /*
58  * fpu_execute returns the following error numbers (0 = no error):
59  */
60 #define	FPE		1	/* take a floating point exception */
61 #define	NOTFPU		2	/* not an FPU instruction */
62 
63 /*
64  * Translate current exceptions into `first' exception.  The
65  * bits go the wrong way for ffs() (0x10 is most important, etc).
66  * There are only 5, so do it the obvious way.
67  */
68 #define	X1(x) x
69 #define	X2(x) x,x
70 #define	X4(x) x,x,x,x
71 #define	X8(x) X4(x),X4(x)
72 #define	X16(x) X8(x),X8(x)
73 
74 static char cx_to_trapx[] = {
75 	X1(FSR_NX),
76 	X2(FSR_DZ),
77 	X4(FSR_UF),
78 	X8(FSR_OF),
79 	X16(FSR_NV)
80 };
81 static u_char fpu_codes[] = {
82 	X1(FPE_FLTINEX_TRAP),
83 	X2(FPE_FLTDIV_TRAP),
84 	X4(FPE_FLTUND_TRAP),
85 	X8(FPE_FLTOVF_TRAP),
86 	X16(FPE_FLTOPERR_TRAP)
87 };
88 
89 static int fpu_types[] = {
90 	FPE_FLTRES,
91 	FPE_FLTDIV,
92 	FPE_FLTUND,
93 	FPE_FLTOVF,
94 	FPE_FLTINV,
95 };
96 
97 /*
98  * The FPU gave us an exception.  Clean up the mess.  Note that the
99  * fp queue can only have FPops in it, never load/store FP registers
100  * nor FBfcc instructions.  Experiments with `crashme' prove that
101  * unknown FPops do enter the queue, however.
102  */
103 void
fpu_cleanup(p,fs)104 fpu_cleanup(p, fs)
105 	register struct proc *p;
106 	register struct fpstate *fs;
107 {
108 	register int i, fsr = fs->fs_fsr, error;
109 	union instr instr;
110 	union sigval sv;
111 	struct fpemu fe;
112 
113 	sv.sival_int = p->p_md.md_tf->tf_pc;  /* XXX only approximate */
114 
115 	switch ((fsr >> FSR_FTT_SHIFT) & FSR_FTT_MASK) {
116 
117 	case FSR_TT_NONE:
118 #if 0
119 		/* XXX I'm not sure how we get here, but ignoring the trap */
120 		/* XXX seems to work in my limited tests		   */
121 		/* XXX More research to be done =)			   */
122 		panic("fpu_cleanup 1"); /* ??? */
123 #else
124 		printf("fpu_cleanup 1\n");
125 #endif
126 		break;
127 
128 	case FSR_TT_IEEE:
129 		if ((i = fsr & FSR_CX) == 0)
130 			panic("fpu ieee trap, but no exception");
131 		trapsignal(p, SIGFPE, fpu_codes[i - 1], fpu_types[i - i], sv);
132 		break;		/* XXX should return, but queue remains */
133 
134 	case FSR_TT_UNFIN:
135 	case FSR_TT_UNIMP:
136 		if (fs->fs_qsize == 0)
137 			panic("fpu_cleanup 2");
138 		break;
139 
140 	case FSR_TT_SEQ:
141 		panic("fpu sequence error");
142 		/* NOTREACHED */
143 
144 	case FSR_TT_HWERR:
145 		log(LOG_ERR, "fpu hardware error (%s[%d])\n",
146 		    p->p_comm, p->p_pid);
147 		uprintf("%s[%d]: fpu hardware error\n", p->p_comm, p->p_pid);
148 		trapsignal(p, SIGFPE, -1, FPE_FLTINV, sv);	/* ??? */
149 		goto out;
150 
151 	default:
152 		printf("fsr=0x%x\n", fsr);
153 		panic("fpu error");
154 	}
155 
156 	/* emulate the instructions left in the queue */
157 	fe.fe_fpstate = fs;
158 	for (i = 0; i < fs->fs_qsize; i++) {
159 		instr.i_int = fs->fs_queue[i].fq_instr;
160 		if (instr.i_any.i_op != IOP_reg ||
161 		    (instr.i_op3.i_op3 != IOP3_FPop1 &&
162 		     instr.i_op3.i_op3 != IOP3_FPop2))
163 			panic("bogus fpu queue");
164 		error = fpu_execute(&fe, instr);
165 		switch (error) {
166 
167 		case 0:
168 			continue;
169 
170 		case FPE:
171 			trapsignal(p, SIGFPE,
172 			    fpu_codes[(fs->fs_fsr & FSR_CX) - 1],
173 			    fpu_types[(fs->fs_fsr & FSR_CX) - 1], sv);
174 			break;
175 
176 		case NOTFPU:
177 			trapsignal(p, SIGILL, 0, ILL_COPROC, sv);
178 			break;
179 
180 		default:
181 			panic("fpu_cleanup 3");
182 			/* NOTREACHED */
183 		}
184 		/* XXX should stop here, but queue remains */
185 	}
186 out:
187 	fs->fs_qsize = 0;
188 }
189 
190 #ifdef notyet
191 /*
192  * If we have no FPU at all (are there any machines like this out
193  * there!?) we have to emulate each instruction, and we need a pointer
194  * to the trapframe so that we can step over them and do FBfcc's.
195  * We know the `queue' is empty, though; we just want to emulate
196  * the instruction at tf->tf_pc.
197  */
198 fpu_emulate(p, tf, fs)
199 	struct proc *p;
200 	register struct trapframe *tf;
201 	register struct fpstate *fs;
202 {
203 
204 	do {
205 		fetch instr from pc
206 		decode
207 		if (integer instr) {
208 			/*
209 			 * We do this here, rather than earlier, to avoid
210 			 * losing even more badly than usual.
211 			 */
212 			if (p->p_addr->u_pcb.pcb_uw) {
213 				write_user_windows();
214 				if (rwindow_save(p))
215 					sigexit(p, SIGILL);
216 			}
217 			if (loadstore) {
218 				do_it;
219 				pc = npc, npc += 4
220 			} else if (fbfcc) {
221 				do_annul_stuff;
222 			} else
223 				return;
224 		} else if (fpu instr) {
225 			fe.fe_fsr = fs->fs_fsr &= ~FSR_CX;
226 			error = fpu_execute(&fe, fs, instr);
227 			switch (error) {
228 				etc;
229 			}
230 		} else
231 			return;
232 		if (want to reschedule)
233 			return;
234 	} while (error == 0);
235 }
236 #endif
237 
238 /*
239  * Execute an FPU instruction (one that runs entirely in the FPU; not
240  * FBfcc or STF, for instance).  On return, fe->fe_fs->fs_fsr will be
241  * modified to reflect the setting the hardware would have left.
242  *
243  * Note that we do not catch all illegal opcodes, so you can, for instance,
244  * multiply two integers this way.
245  */
246 int
fpu_execute(fe,instr)247 fpu_execute(fe, instr)
248 	register struct fpemu *fe;
249 	union instr instr;
250 {
251 	register struct fpn *fp;
252 	register int opf, rs1, rs2, rd, type, mask, fsr, cx;
253 	register struct fpstate *fs;
254 	u_int space[4];
255 
256 	/*
257 	 * `Decode' and execute instruction.  Start with no exceptions.
258 	 * The type of any i_opf opcode is in the bottom two bits, so we
259 	 * squish them out here.
260 	 */
261 	opf = instr.i_opf.i_opf;
262 	type = opf & 3;
263 	mask = "\0\0\1\3"[type];
264 	rs1 = instr.i_opf.i_rs1 & ~mask;
265 	rs2 = instr.i_opf.i_rs2 & ~mask;
266 	rd = instr.i_opf.i_rd & ~mask;
267 #ifdef notdef
268 	if ((rs1 | rs2 | rd) & mask)
269 		return (BADREG);
270 #endif
271 	fs = fe->fe_fpstate;
272 	fe->fe_fsr = fs->fs_fsr & ~FSR_CX;
273 	fe->fe_cx = 0;
274 	switch (opf >>= 2) {
275 
276 	default:
277 		return (NOTFPU);
278 
279 	case FMOV >> 2:		/* these should all be pretty obvious */
280 		rs1 = fs->fs_regs[rs2];
281 		goto mov;
282 
283 	case FNEG >> 2:
284 		rs1 = fs->fs_regs[rs2] ^ (1 << 31);
285 		goto mov;
286 
287 	case FABS >> 2:
288 		rs1 = fs->fs_regs[rs2] & ~(1 << 31);
289 	mov:
290 		fs->fs_regs[rd] = rs1;
291 		fs->fs_fsr = fe->fe_fsr;
292 		return (0);	/* success */
293 
294 	case FSQRT >> 2:
295 		fpu_explode(fe, &fe->fe_f1, type, rs2);
296 		fp = fpu_sqrt(fe);
297 		break;
298 
299 	case FADD >> 2:
300 		fpu_explode(fe, &fe->fe_f1, type, rs1);
301 		fpu_explode(fe, &fe->fe_f2, type, rs2);
302 		fp = fpu_add(fe);
303 		break;
304 
305 	case FSUB >> 2:
306 		fpu_explode(fe, &fe->fe_f1, type, rs1);
307 		fpu_explode(fe, &fe->fe_f2, type, rs2);
308 		fp = fpu_sub(fe);
309 		break;
310 
311 	case FMUL >> 2:
312 		fpu_explode(fe, &fe->fe_f1, type, rs1);
313 		fpu_explode(fe, &fe->fe_f2, type, rs2);
314 		fp = fpu_mul(fe);
315 		break;
316 
317 	case FDIV >> 2:
318 		fpu_explode(fe, &fe->fe_f1, type, rs1);
319 		fpu_explode(fe, &fe->fe_f2, type, rs2);
320 		fp = fpu_div(fe);
321 		break;
322 
323 	case FCMP >> 2:
324 		fpu_explode(fe, &fe->fe_f1, type, rs1);
325 		fpu_explode(fe, &fe->fe_f2, type, rs2);
326 		fpu_compare(fe, 0);
327 		goto cmpdone;
328 
329 	case FCMPE >> 2:
330 		fpu_explode(fe, &fe->fe_f1, type, rs1);
331 		fpu_explode(fe, &fe->fe_f2, type, rs2);
332 		fpu_compare(fe, 1);
333 	cmpdone:
334 		/*
335 		 * The only possible exception here is NV; catch it
336 		 * early and get out, as there is no result register.
337 		 */
338 		cx = fe->fe_cx;
339 		fsr = fe->fe_fsr | (cx << FSR_CX_SHIFT);
340 		if (cx != 0) {
341 			if (fsr & (FSR_NV << FSR_TEM_SHIFT)) {
342 				fs->fs_fsr = (fsr & ~FSR_FTT) |
343 				    (FSR_TT_IEEE << FSR_FTT_SHIFT);
344 				return (FPE);
345 			}
346 			fsr |= FSR_NV << FSR_AX_SHIFT;
347 		}
348 		fs->fs_fsr = fsr;
349 		return (0);
350 
351 	case FSMULD >> 2:
352 	case FDMULX >> 2:
353 		if (type == FTYPE_EXT)
354 			return (NOTFPU);
355 		fpu_explode(fe, &fe->fe_f1, type, rs1);
356 		fpu_explode(fe, &fe->fe_f2, type, rs2);
357 		type++;	/* single to double, or double to quad */
358 		fp = fpu_mul(fe);
359 		break;
360 
361 	case FTOI >> 2:
362 	case FTOS >> 2:
363 		rd = instr.i_opf.i_rd;
364 		goto fto;
365 	case FTOD >> 2:
366 		rd = instr.i_opf.i_rd & (~1);
367 		goto fto;
368 	case FTOX >> 2:
369 		rd = instr.i_opf.i_rd & (~3);
370 
371 fto:
372 		fpu_explode(fe, fp = &fe->fe_f1, type, rs2);
373 		type = opf & 3;	/* sneaky; depends on instruction encoding */
374 		break;
375 	}
376 
377 	/*
378 	 * ALU operation is complete.  Collapse the result and then check
379 	 * for exceptions.  If we got any, and they are enabled, do not
380 	 * alter the destination register, just stop with an exception.
381 	 * Otherwise set new current exceptions and accrue.
382 	 */
383 	fpu_implode(fe, fp, type, space);
384 	cx = fe->fe_cx;
385 	fsr = fe->fe_fsr;
386 	if (cx != 0) {
387 		mask = (fsr >> FSR_TEM_SHIFT) & FSR_TEM_MASK;
388 		if (cx & mask) {
389 			/* not accrued??? */
390 			fs->fs_fsr = (fsr & ~FSR_FTT) |
391 			    (FSR_TT_IEEE << FSR_FTT_SHIFT) |
392 			    (cx_to_trapx[(cx & mask) - 1] << FSR_CX_SHIFT);
393 			return (FPE);
394 		}
395 		fsr |= (cx << FSR_CX_SHIFT) | (cx << FSR_AX_SHIFT);
396 	}
397 	fs->fs_fsr = fsr;
398 	fs->fs_regs[rd] = space[0];
399 	if (type >= FTYPE_DBL) {
400 		fs->fs_regs[rd + 1] = space[1];
401 		if (type > FTYPE_DBL) {
402 			fs->fs_regs[rd + 2] = space[2];
403 			fs->fs_regs[rd + 3] = space[3];
404 		}
405 	}
406 	return (0);	/* success */
407 }
408