xref: /freebsd/sys/powerpc/booke/spe.c (revision 42249ef2)
1 /*-
2  * Copyright (C) 1996 Wolfgang Solfrank.
3  * Copyright (C) 1996 TooLs GmbH.
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  * 3. All advertising materials mentioning features or use of this software
15  *    must display the following acknowledgement:
16  *	This product includes software developed by TooLs GmbH.
17  * 4. The name of TooLs GmbH may not be used to endorse or promote products
18  *    derived from this software without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
21  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
22  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
23  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
24  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
25  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
26  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
27  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
28  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
29  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30  *
31  *	$NetBSD: fpu.c,v 1.5 2001/07/22 11:29:46 wiz Exp $
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include <sys/param.h>
38 #include <sys/proc.h>
39 #include <sys/systm.h>
40 #include <sys/limits.h>
41 
42 #include <machine/altivec.h>
43 #include <machine/fpu.h>
44 #include <machine/ieeefp.h>
45 #include <machine/pcb.h>
46 #include <machine/psl.h>
47 
48 #include <powerpc/fpu/fpu_arith.h>
49 #include <powerpc/fpu/fpu_emu.h>
50 #include <powerpc/fpu/fpu_extern.h>
51 
52 void spe_handle_fpdata(struct trapframe *);
53 void spe_handle_fpround(struct trapframe *);
54 static int spe_emu_instr(uint32_t, struct fpemu *, struct fpn **, uint32_t *);
55 
56 static void
57 save_vec_int(struct thread *td)
58 {
59 	int	msr;
60 	struct	pcb *pcb;
61 
62 	pcb = td->td_pcb;
63 
64 	/*
65 	 * Temporarily re-enable the vector unit during the save
66 	 */
67 	msr = mfmsr();
68 	mtmsr(msr | PSL_VEC);
69 
70 	/*
71 	 * Save the vector registers and SPEFSCR to the PCB
72 	 */
73 #define EVSTDW(n)   __asm ("evstdw %1,0(%0)" \
74 		:: "b"(pcb->pcb_vec.vr[n]), "n"(n));
75 	EVSTDW(0);	EVSTDW(1);	EVSTDW(2);	EVSTDW(3);
76 	EVSTDW(4);	EVSTDW(5);	EVSTDW(6);	EVSTDW(7);
77 	EVSTDW(8);	EVSTDW(9);	EVSTDW(10);	EVSTDW(11);
78 	EVSTDW(12);	EVSTDW(13);	EVSTDW(14);	EVSTDW(15);
79 	EVSTDW(16);	EVSTDW(17);	EVSTDW(18);	EVSTDW(19);
80 	EVSTDW(20);	EVSTDW(21);	EVSTDW(22);	EVSTDW(23);
81 	EVSTDW(24);	EVSTDW(25);	EVSTDW(26);	EVSTDW(27);
82 	EVSTDW(28);	EVSTDW(29);	EVSTDW(30);	EVSTDW(31);
83 #undef EVSTDW
84 
85 	__asm ( "evxor 0,0,0\n"
86 		"evaddumiaaw 0,0\n"
87 		"evstdd 0,0(%0)" :: "b"(&pcb->pcb_vec.spare[0]));
88 	pcb->pcb_vec.vscr = mfspr(SPR_SPEFSCR);
89 
90 	/*
91 	 * Disable vector unit again
92 	 */
93 	isync();
94 	mtmsr(msr);
95 
96 }
97 
98 void
99 enable_vec(struct thread *td)
100 {
101 	int	msr;
102 	struct	pcb *pcb;
103 	struct	trapframe *tf;
104 
105 	pcb = td->td_pcb;
106 	tf = trapframe(td);
107 
108 	/*
109 	 * Save the thread's SPE CPU number, and set the CPU's current
110 	 * vector thread
111 	 */
112 	td->td_pcb->pcb_veccpu = PCPU_GET(cpuid);
113 	PCPU_SET(vecthread, td);
114 
115 	/*
116 	 * Enable the vector unit for when the thread returns from the
117 	 * exception. If this is the first time the unit has been used by
118 	 * the thread, initialise the vector registers and VSCR to 0, and
119 	 * set the flag to indicate that the vector unit is in use.
120 	 */
121 	tf->srr1 |= PSL_VEC;
122 	if (!(pcb->pcb_flags & PCB_VEC)) {
123 		memset(&pcb->pcb_vec, 0, sizeof pcb->pcb_vec);
124 		pcb->pcb_flags |= PCB_VEC;
125 		pcb->pcb_vec.vscr = mfspr(SPR_SPEFSCR);
126 	}
127 
128 	/*
129 	 * Temporarily enable the vector unit so the registers
130 	 * can be restored.
131 	 */
132 	msr = mfmsr();
133 	mtmsr(msr | PSL_VEC);
134 
135 	/* Restore SPEFSCR and ACC.  Use %r0 as the scratch for ACC. */
136 	mtspr(SPR_SPEFSCR, pcb->pcb_vec.vscr);
137 	__asm __volatile("evldd 0, 0(%0); evmra 0,0\n"
138 	    :: "b"(&pcb->pcb_vec.spare[0]));
139 
140 	/*
141 	 * The lower half of each register will be restored on trap return.  Use
142 	 * %r0 as a scratch register, and restore it last.
143 	 */
144 #define	EVLDW(n)   __asm __volatile("evldw 0, 0(%0); evmergehilo "#n",0,"#n \
145 	    :: "b"(&pcb->pcb_vec.vr[n]));
146 	EVLDW(1);	EVLDW(2);	EVLDW(3);	EVLDW(4);
147 	EVLDW(5);	EVLDW(6);	EVLDW(7);	EVLDW(8);
148 	EVLDW(9);	EVLDW(10);	EVLDW(11);	EVLDW(12);
149 	EVLDW(13);	EVLDW(14);	EVLDW(15);	EVLDW(16);
150 	EVLDW(17);	EVLDW(18);	EVLDW(19);	EVLDW(20);
151 	EVLDW(21);	EVLDW(22);	EVLDW(23);	EVLDW(24);
152 	EVLDW(25);	EVLDW(26);	EVLDW(27);	EVLDW(28);
153 	EVLDW(29);	EVLDW(30);	EVLDW(31);	EVLDW(0);
154 #undef EVLDW
155 
156 	isync();
157 	mtmsr(msr);
158 }
159 
160 void
161 save_vec(struct thread *td)
162 {
163 	struct pcb *pcb;
164 
165 	pcb = td->td_pcb;
166 
167 	save_vec_int(td);
168 
169 	/*
170 	 * Clear the current vec thread and pcb's CPU id
171 	 * XXX should this be left clear to allow lazy save/restore ?
172 	 */
173 	pcb->pcb_veccpu = INT_MAX;
174 	PCPU_SET(vecthread, NULL);
175 }
176 
177 /*
178  * Save SPE state without dropping ownership.  This will only save state if
179  * the current vector-thread is `td'.
180  */
181 void
182 save_vec_nodrop(struct thread *td)
183 {
184 	struct thread *vtd;
185 
186 	vtd = PCPU_GET(vecthread);
187 	if (td != vtd) {
188 		return;
189 	}
190 
191 	save_vec_int(td);
192 }
193 
194 
195 #define	SPE_INST_MASK	0x31f
196 #define	EADD	0x200
197 #define	ESUB	0x201
198 #define	EABS	0x204
199 #define	ENABS	0x205
200 #define	ENEG	0x206
201 #define	EMUL	0x208
202 #define	EDIV	0x209
203 #define	ECMPGT	0x20c
204 #define	ECMPLT	0x20d
205 #define	ECMPEQ	0x20e
206 #define	ECFUI	0x210
207 #define	ECFSI	0x211
208 #define	ECTUI	0x214
209 #define	ECTSI	0x215
210 #define	ECTUF	0x216
211 #define	ECTSF	0x217
212 #define	ECTUIZ	0x218
213 #define	ECTSIZ	0x21a
214 
215 #define	SPE		0x4
216 #define	SPFP		0x6
217 #define	DPFP		0x7
218 
219 #define	SPE_OPC		4
220 #define	OPC_SHIFT	26
221 
222 #define	EVFSADD		0x280
223 #define	EVFSSUB		0x281
224 #define	EVFSABS		0x284
225 #define	EVFSNABS	0x285
226 #define	EVFSNEG		0x286
227 #define	EVFSMUL		0x288
228 #define	EVFSDIV		0x289
229 #define	EVFSCMPGT	0x28c
230 #define	EVFSCMPLT	0x28d
231 #define	EVFSCMPEQ	0x28e
232 #define	EVFSCFUI	0x290
233 #define	EVFSCFSI	0x291
234 #define	EVFSCTUI	0x294
235 #define	EVFSCTSI	0x295
236 #define	EVFSCTUF	0x296
237 #define	EVFSCTSF	0x297
238 #define	EVFSCTUIZ	0x298
239 #define	EVFSCTSIZ	0x29a
240 
241 #define	EFSADD		0x2c0
242 #define	EFSSUB		0x2c1
243 #define	EFSABS		0x2c4
244 #define	EFSNABS		0x2c5
245 #define	EFSNEG		0x2c6
246 #define	EFSMUL		0x2c8
247 #define	EFSDIV		0x2c9
248 #define	EFSCMPGT	0x2cc
249 #define	EFSCMPLT	0x2cd
250 #define	EFSCMPEQ	0x2ce
251 #define	EFSCFD		0x2cf
252 #define	EFSCFUI		0x2d0
253 #define	EFSCFSI		0x2d1
254 #define	EFSCTUI		0x2d4
255 #define	EFSCTSI		0x2d5
256 #define	EFSCTUF		0x2d6
257 #define	EFSCTSF		0x2d7
258 #define	EFSCTUIZ	0x2d8
259 #define	EFSCTSIZ	0x2da
260 
261 #define	EFDADD		0x2e0
262 #define	EFDSUB		0x2e1
263 #define	EFDABS		0x2e4
264 #define	EFDNABS		0x2e5
265 #define	EFDNEG		0x2e6
266 #define	EFDMUL		0x2e8
267 #define	EFDDIV		0x2e9
268 #define	EFDCMPGT	0x2ec
269 #define	EFDCMPLT	0x2ed
270 #define	EFDCMPEQ	0x2ee
271 #define	EFDCFS		0x2ef
272 #define	EFDCFUI		0x2f0
273 #define	EFDCFSI		0x2f1
274 #define	EFDCTUI		0x2f4
275 #define	EFDCTSI		0x2f5
276 #define	EFDCTUF		0x2f6
277 #define	EFDCTSF		0x2f7
278 #define	EFDCTUIZ	0x2f8
279 #define	EFDCTSIZ	0x2fa
280 
281 enum {
282 	NONE,
283 	SINGLE,
284 	DOUBLE,
285 	VECTOR,
286 };
287 
288 static uint32_t fpscr_to_spefscr(uint32_t fpscr)
289 {
290 	uint32_t spefscr;
291 
292 	spefscr = 0;
293 
294 	if (fpscr & FPSCR_VX)
295 		spefscr |= SPEFSCR_FINV;
296 	if (fpscr & FPSCR_OX)
297 		spefscr |= SPEFSCR_FOVF;
298 	if (fpscr & FPSCR_UX)
299 		spefscr |= SPEFSCR_FUNF;
300 	if (fpscr & FPSCR_ZX)
301 		spefscr |= SPEFSCR_FDBZ;
302 	if (fpscr & FPSCR_XX)
303 		spefscr |= SPEFSCR_FX;
304 
305 	return (spefscr);
306 }
307 
308 /* Sign is 0 for unsigned, 1 for signed. */
309 static int
310 spe_to_int(struct fpemu *fpemu, struct fpn *fpn, uint32_t *val, int sign)
311 {
312 	uint32_t res[2];
313 
314 	res[0] = fpu_ftox(fpemu, fpn, res);
315 	if (res[0] != UINT_MAX && res[0] != 0)
316 		fpemu->fe_cx |= FPSCR_OX;
317 	else if (sign == 0 && res[0] != 0)
318 		fpemu->fe_cx |= FPSCR_UX;
319 	else
320 		*val = res[1];
321 
322 	return (0);
323 }
324 
325 /* Masked instruction */
326 /*
327  * For compare instructions, returns 1 if success, 0 if not.  For all others,
328  * returns -1, or -2 if no result needs recorded.
329  */
330 static int
331 spe_emu_instr(uint32_t instr, struct fpemu *fpemu,
332     struct fpn **result, uint32_t *iresult)
333 {
334 	switch (instr & SPE_INST_MASK) {
335 	case EABS:
336 	case ENABS:
337 	case ENEG:
338 		/* Taken care of elsewhere. */
339 		break;
340 	case ECTUIZ:
341 		fpemu->fe_cx &= ~FPSCR_RN;
342 		fpemu->fe_cx |= FP_RZ;
343 	case ECTUI:
344 		spe_to_int(fpemu, &fpemu->fe_f2, iresult, 0);
345 		return (-1);
346 	case ECTSIZ:
347 		fpemu->fe_cx &= ~FPSCR_RN;
348 		fpemu->fe_cx |= FP_RZ;
349 	case ECTSI:
350 		spe_to_int(fpemu, &fpemu->fe_f2, iresult, 1);
351 		return (-1);
352 	case EADD:
353 		*result = fpu_add(fpemu);
354 		break;
355 	case ESUB:
356 		*result = fpu_sub(fpemu);
357 		break;
358 	case EMUL:
359 		*result = fpu_mul(fpemu);
360 		break;
361 	case EDIV:
362 		*result = fpu_div(fpemu);
363 		break;
364 	case ECMPGT:
365 		fpu_compare(fpemu, 0);
366 		if (fpemu->fe_cx & FPSCR_FG)
367 			return (1);
368 		return (0);
369 	case ECMPLT:
370 		fpu_compare(fpemu, 0);
371 		if (fpemu->fe_cx & FPSCR_FL)
372 			return (1);
373 		return (0);
374 	case ECMPEQ:
375 		fpu_compare(fpemu, 0);
376 		if (fpemu->fe_cx & FPSCR_FE)
377 			return (1);
378 		return (0);
379 	default:
380 		printf("Unknown instruction %x\n", instr);
381 	}
382 
383 	return (-1);
384 }
385 
386 static int
387 spe_explode(struct fpemu *fe, struct fpn *fp, uint32_t type,
388     uint32_t hi, uint32_t lo)
389 {
390 	uint32_t s;
391 
392 	fp->fp_sign = hi >> 31;
393 	fp->fp_sticky = 0;
394 	switch (type) {
395 	case SINGLE:
396 		s = fpu_stof(fp, hi);
397 		break;
398 
399 	case DOUBLE:
400 		s = fpu_dtof(fp, hi, lo);
401 		break;
402 	}
403 
404 	if (s == FPC_QNAN && (fp->fp_mant[0] & FP_QUIETBIT) == 0) {
405 		/*
406 		 * Input is a signalling NaN.  All operations that return
407 		 * an input NaN operand put it through a ``NaN conversion'',
408 		 * which basically just means ``turn on the quiet bit''.
409 		 * We do this here so that all NaNs internally look quiet
410 		 * (we can tell signalling ones by their class).
411 		 */
412 		fp->fp_mant[0] |= FP_QUIETBIT;
413 		fe->fe_cx = FPSCR_VXSNAN;	/* assert invalid operand */
414 		s = FPC_SNAN;
415 	}
416 	fp->fp_class = s;
417 
418 	return (0);
419 }
420 
421 /*
422  * Save the high word of a 64-bit GPR for manipulation in the exception handler.
423  */
424 static uint32_t
425 spe_save_reg_high(int reg)
426 {
427 	uint32_t vec[2];
428 #define EVSTDW(n)   case n: __asm __volatile ("evstdw %1,0(%0)" \
429 		:: "b"(vec), "n"(n)); break;
430 	switch (reg) {
431 	EVSTDW(0);	EVSTDW(1);	EVSTDW(2);	EVSTDW(3);
432 	EVSTDW(4);	EVSTDW(5);	EVSTDW(6);	EVSTDW(7);
433 	EVSTDW(8);	EVSTDW(9);	EVSTDW(10);	EVSTDW(11);
434 	EVSTDW(12);	EVSTDW(13);	EVSTDW(14);	EVSTDW(15);
435 	EVSTDW(16);	EVSTDW(17);	EVSTDW(18);	EVSTDW(19);
436 	EVSTDW(20);	EVSTDW(21);	EVSTDW(22);	EVSTDW(23);
437 	EVSTDW(24);	EVSTDW(25);	EVSTDW(26);	EVSTDW(27);
438 	EVSTDW(28);	EVSTDW(29);	EVSTDW(30);	EVSTDW(31);
439 	}
440 #undef EVSTDW
441 
442 	return (vec[0]);
443 }
444 
445 /*
446  * Load the given value into the high word of the requested register.
447  */
448 static void
449 spe_load_reg_high(int reg, uint32_t val)
450 {
451 #define	EVLDW(n)   case n: __asm __volatile("evmergelo "#n",%0,"#n \
452 	    :: "r"(val)); break;
453 	switch (reg) {
454 	EVLDW(1);	EVLDW(2);	EVLDW(3);	EVLDW(4);
455 	EVLDW(5);	EVLDW(6);	EVLDW(7);	EVLDW(8);
456 	EVLDW(9);	EVLDW(10);	EVLDW(11);	EVLDW(12);
457 	EVLDW(13);	EVLDW(14);	EVLDW(15);	EVLDW(16);
458 	EVLDW(17);	EVLDW(18);	EVLDW(19);	EVLDW(20);
459 	EVLDW(21);	EVLDW(22);	EVLDW(23);	EVLDW(24);
460 	EVLDW(25);	EVLDW(26);	EVLDW(27);	EVLDW(28);
461 	EVLDW(29);	EVLDW(30);	EVLDW(31);	EVLDW(0);
462 	}
463 #undef EVLDW
464 
465 }
466 
467 void
468 spe_handle_fpdata(struct trapframe *frame)
469 {
470 	struct fpemu fpemu;
471 	struct fpn *result;
472 	uint32_t instr, instr_sec_op;
473 	uint32_t cr_shift, ra, rb, rd, src;
474 	uint32_t high, low, res, tmp; /* For vector operations. */
475 	uint32_t spefscr = 0;
476 	uint32_t ftod_res[2];
477 	int width; /* Single, Double, Vector, Integer */
478 	int err;
479 	uint32_t msr;
480 
481 	err = fueword32((void *)frame->srr0, &instr);
482 
483 	if (err != 0)
484 		return;
485 		/* Fault. */;
486 
487 	if ((instr >> OPC_SHIFT) != SPE_OPC)
488 		return;
489 
490 	msr = mfmsr();
491 	/*
492 	 * 'cr' field is the upper 3 bits of rd.  Magically, since a) rd is 5
493 	 * bits, b) each 'cr' field is 4 bits, and c) Only the 'GT' bit is
494 	 * modified for most compare operations, the full value of rd can be
495 	 * used as a shift value.
496 	 */
497 	rd = (instr >> 21) & 0x1f;
498 	ra = (instr >> 16) & 0x1f;
499 	rb = (instr >> 11) & 0x1f;
500 	src = (instr >> 5) & 0x7;
501 	cr_shift = 28 - (rd & 0x1f);
502 
503 	instr_sec_op = (instr & 0x7ff);
504 
505 	memset(&fpemu, 0, sizeof(fpemu));
506 
507 	width = NONE;
508 	switch (src) {
509 	case SPE:
510 		mtmsr(msr | PSL_VEC);
511 		switch (instr_sec_op) {
512 		case EVFSABS:
513 			high = spe_save_reg_high(ra) & ~(1U << 31);
514 			frame->fixreg[rd] = frame->fixreg[ra] & ~(1U << 31);
515 			spe_load_reg_high(rd, high);
516 			break;
517 		case EVFSNABS:
518 			high = spe_save_reg_high(ra) | (1U << 31);
519 			frame->fixreg[rd] = frame->fixreg[ra] | (1U << 31);
520 			spe_load_reg_high(rd, high);
521 			break;
522 		case EVFSNEG:
523 			high = spe_save_reg_high(ra) ^ (1U << 31);
524 			frame->fixreg[rd] = frame->fixreg[ra] ^ (1U << 31);
525 			spe_load_reg_high(rd, high);
526 			break;
527 		default:
528 			/* High word */
529 			spe_explode(&fpemu, &fpemu.fe_f1, SINGLE,
530 			    spe_save_reg_high(ra), 0);
531 			spe_explode(&fpemu, &fpemu.fe_f2, SINGLE,
532 			    spe_save_reg_high(rb), 0);
533 			high = spe_emu_instr(instr_sec_op, &fpemu, &result,
534 			    &tmp);
535 
536 			if (high < 0)
537 				spe_load_reg_high(rd, tmp);
538 
539 			spefscr = fpscr_to_spefscr(fpemu.fe_cx) << 16;
540 			/* Clear the fpemu to start over on the lower bits. */
541 			memset(&fpemu, 0, sizeof(fpemu));
542 
543 			/* Now low word */
544 			spe_explode(&fpemu, &fpemu.fe_f1, SINGLE,
545 			    frame->fixreg[ra], 0);
546 			spe_explode(&fpemu, &fpemu.fe_f2, SINGLE,
547 			    frame->fixreg[rb], 0);
548 			spefscr |= fpscr_to_spefscr(fpemu.fe_cx);
549 			low = spe_emu_instr(instr_sec_op, &fpemu, &result,
550 			    &frame->fixreg[rd]);
551 			if (instr_sec_op == EVFSCMPEQ ||
552 			    instr_sec_op == EVFSCMPGT ||
553 			    instr_sec_op == EVFSCMPLT) {
554 				res = (high << 3) | (low << 2) |
555 				    ((high | low) << 1) | (high & low);
556 				width = NONE;
557 			} else
558 				width = VECTOR;
559 			break;
560 		}
561 		goto end;
562 
563 	case SPFP:
564 		switch (instr_sec_op) {
565 		case EFSABS:
566 			frame->fixreg[rd] = frame->fixreg[ra] & ~(1U << 31);
567 			break;
568 		case EFSNABS:
569 			frame->fixreg[rd] = frame->fixreg[ra] | (1U << 31);
570 			break;
571 		case EFSNEG:
572 			frame->fixreg[rd] = frame->fixreg[ra] ^ (1U << 31);
573 			break;
574 		case EFSCFD:
575 			mtmsr(msr | PSL_VEC);
576 			spe_explode(&fpemu, &fpemu.fe_f3, DOUBLE,
577 			    spe_save_reg_high(rb), frame->fixreg[rb]);
578 			result = &fpemu.fe_f3;
579 			width = SINGLE;
580 			break;
581 		default:
582 			spe_explode(&fpemu, &fpemu.fe_f1, SINGLE,
583 			    frame->fixreg[ra], 0);
584 			spe_explode(&fpemu, &fpemu.fe_f2, SINGLE,
585 			    frame->fixreg[rb], 0);
586 			width = SINGLE;
587 		}
588 		break;
589 	case DPFP:
590 		mtmsr(msr | PSL_VEC);
591 		switch (instr_sec_op) {
592 		case EFDABS:
593 			high = spe_save_reg_high(ra) & ~(1U << 31);
594 			frame->fixreg[rd] = frame->fixreg[ra];
595 			spe_load_reg_high(rd, high);
596 			break;
597 		case EFDNABS:
598 			high = spe_save_reg_high(ra) | (1U << 31);
599 			frame->fixreg[rd] = frame->fixreg[ra];
600 			spe_load_reg_high(rd, high);
601 			break;
602 		case EFDNEG:
603 			high = spe_save_reg_high(ra) ^ (1U << 31);
604 			frame->fixreg[rd] = frame->fixreg[ra];
605 			spe_load_reg_high(rd, high);
606 			break;
607 		case EFDCFS:
608 			spe_explode(&fpemu, &fpemu.fe_f3, SINGLE,
609 			    frame->fixreg[rb], 0);
610 			result = &fpemu.fe_f3;
611 			width = DOUBLE;
612 			break;
613 		default:
614 			spe_explode(&fpemu, &fpemu.fe_f1, DOUBLE,
615 			    spe_save_reg_high(ra), frame->fixreg[ra]);
616 			spe_explode(&fpemu, &fpemu.fe_f2, DOUBLE,
617 			    spe_save_reg_high(rb), frame->fixreg[rb]);
618 			width = DOUBLE;
619 		}
620 		break;
621 	}
622 	switch (instr_sec_op) {
623 	case EFDCFS:
624 	case EFSCFD:
625 		/* Already handled. */
626 		break;
627 	default:
628 		res = spe_emu_instr(instr_sec_op, &fpemu, &result,
629 		    &frame->fixreg[rd]);
630 		if (res != -1)
631 			res <<= 2;
632 		break;
633 	}
634 
635 	switch (instr_sec_op & SPE_INST_MASK) {
636 	case ECMPEQ:
637 	case ECMPGT:
638 	case ECMPLT:
639 		frame->cr &= ~(0xf << cr_shift);
640 		frame->cr |= (res << cr_shift);
641 		break;
642 	case ECTUI:
643 	case ECTUIZ:
644 	case ECTSI:
645 	case ECTSIZ:
646 		break;
647 	default:
648 		switch (width) {
649 		case NONE:
650 		case VECTOR:
651 			break;
652 		case SINGLE:
653 			frame->fixreg[rd] = fpu_ftos(&fpemu, result);
654 			break;
655 		case DOUBLE:
656 			spe_load_reg_high(rd, fpu_ftod(&fpemu, result, ftod_res));
657 			frame->fixreg[rd] = ftod_res[1];
658 			break;
659 		default:
660 			panic("Unknown storage width %d", width);
661 			break;
662 		}
663 	}
664 
665 end:
666 	spefscr |= (mfspr(SPR_SPEFSCR) & ~SPEFSCR_FINVS);
667 	mtspr(SPR_SPEFSCR, spefscr);
668 	frame->srr0 += 4;
669 	mtmsr(msr);
670 
671 	return;
672 }
673 
674 void
675 spe_handle_fpround(struct trapframe *frame)
676 {
677 
678 	/*
679 	 * Punt fpround exceptions for now.  This leaves the truncated result in
680 	 * the register.  We'll deal with overflow/underflow later.
681 	 */
682 	return;
683 }
684