1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
|
/*
* arch/sparc/math-emu/math.c
*
* Copyright (C) 1998 Peter Maydell (pmaydell@chiark.greenend.org.uk)
* Based on the sparc64 code by Jakub Jelinek.
*
* This is a good place to start if you're trying to understand the
* emulation code, because it's pretty simple. What we do is
* essentially analyse the instruction to work out what the operation
* is and which registers are involved. We then execute the appropriate
* FXXXX function. [The floating point queue introduces a minor wrinkle;
* see below...]
* The fxxxxx.c files each emulate a single insn. They look relatively
* simple because the complexity is hidden away in an unholy tangle
* of preprocessor macros.
*
* WARNING : don't look at the macro definitions unless you
* absolutely have to! They're extremely ugly, rather complicated
* and a single line in an fxxxx.c file can expand to the equivalent
* of 30 lines or more of C. Of course, any error in those 30 lines
* is reported by the compiler as an error in the single line with the
* macro usage...
* Question: should we replace them with inline functions?
*
* The first layer of macros is single.h, double.h, quad.h. Generally
* these files define macros for working with floating point numbers
* of the three IEEE formats. FP_ADD_D(R,A,B) is for adding doubles,
* for instance. These macros are usually defined as calls to more
* generic macros (in this case _FP_ADD(D,2,R,X,Y) where the number
* of machine words required to store the given IEEE format is passed
* as a parameter. [double.h and co check the number of bits in a word
* and define FP_ADD_D & co appropriately].
* The generic macros are defined in op-common.h. This is where all
* the grotty stuff like handling NaNs is coded. To handle the possible
* word sizes macros in op-common.h use macros like _FP_FRAC_SLL_##wc()
* where wc is the 'number of machine words' parameter (here 2).
* These are defined in the third layer of macros: op-1.h, op-2.h
* and op-4.h. These handle operations on floating point numbers composed
* of 1,2 and 4 machine words respectively. [For example, on sparc64
* doubles are one machine word so macros in double.h eventually use
* constructs in op-1.h, but on sparc32 they use op-2.h definitions.]
* soft-fp.h is on the same level as op-common.h, and defines some
* macros which are independent of both word size and FP format.
* Finally, sfp-machine.h is the machine dependent part of the
* code: it defines the word size and what type a word is. It also
* defines how _FP_MUL_MEAT_t() maps to _FP_MUL_MEAT_n_* : op-n.h
* provide several possible flavours of multiply algorithm, most
* of which require that you supply some form of asm or C primitive to
* do the actual multiply. (such asm primitives should be defined
* in sfp-machine.h too). udivmodti4.c is the same sort of thing.
*
* There may be some errors here because I'm working from a
* SPARC architecture manual V9, and what I really want is V8...
* Also, the insns which can generate exceptions seem to be a
* greater subset of the FPops than for V9 (for example, FCMPED
* has to be emulated on V8). So I think I'm going to have
* to emulate them all just to be on the safe side...
*
* Emulation routines originate from soft-fp package, which is
* part of glibc and has appropriate copyrights in it (allegedly).
*
* NB: on sparc int == long == 4 bytes, long long == 8 bytes.
* Most bits of the kernel seem to go for long rather than int,
* so we follow that practice...
*/
/* WISHLIST:
*
* + Replace all the macros with inline functions. These should
* have the same effect but be much easier to work with.
*
* + Emulate the IEEE exception flags. We don't currently do this
* because a) it would require significant alterations to
* the emulation macros [see the comments about _FP_NEG()
* in op-common.c and note that we'd need to invent a convention
* for passing in the flags to FXXXX fns and returning them] and
* b) SPARClinux doesn't let users access the flags anyway
* [contrast Solaris, which allows you to examine, clear or set
* the flags, and request that exceptions cause SIGFPE
* [which you then set up a signal handler for, obviously...]].
* Erm, (b) may quite possibly be garbage. %fsr is user-writable
* so you don't need a syscall. There may or may not be library
* support.
*
* + Emulation of FMULQ, FDIVQ, FSQRTQ, FDMULQ needs to be
* written!
*
* + reindent code to conform to Linux kernel standard :->
*
* + work out whether all the compile-time warnings are bogus
*
* + check that conversion to/from integers works
*
* + check with the SPARC architecture manual to see if we resolve
* the implementation-dependent bits of the IEEE spec in the
* same manner as the hardware.
*
* + more test cases for the test script always welcome!
*
* + illegal opcodes currently cause SIGFPEs. We should arrange
* to tell the traps.c code to SIGILL instead. Currently,
* everywhere that we return 0 should cause SIGILL, I think.
* SIGFPE should only be caused if we set an IEEE exception bit
* and the relevant trap bit is also set. (this means that
* traps.c should do this; also it should handle the case of
* IEEE exception generated directly by the hardware.)
* Should illegal_fp_register (which is a flavour of fp exception)
* cause SIGFPE or SIGILL?
*
* + the test script needs to be extended to handle the quadword
* and comparison insns.
*
* + _FP_DIV_MEAT_2_udiv_64() appears to work but it should be
* checked by somebody who understands the algorithm :->
*
* + fpsave() saves the FP queue but fpload() doesn't reload it.
* Therefore when we context switch or change FPU ownership
* we have to check to see if the queue had anything in it and
* emulate it if it did. This is going to be a pain.
*/
#include <linux/types.h>
#include <linux/sched.h>
#include <linux/mm.h>
#include <asm/uaccess.h>
#include "soft-fp.h"
#define FLOATFUNC(x) extern int x(void *,void *,void *)
/* Current status: we don't properly emulate the difficult quadword
* insns (MUL, DIV, SQRT).
* There are also some ops involving the FP registers which we don't
* emulate: the branch on FP condition flags and the load/store to
* FP regs or FSR. I'm assuming that these will never generate traps
* (not unreasonable if there's an FPU at all; comments in the NetBSD
* kernel source agree on this point). If we wanted to allow
* purely software-emulation of the FPU with FPU totally disabled
* or non-existent, we'd have to emulate these as well. We'd also
* need to alter the fp_disabled trap handler to call the math-emu
* code appropriately. The structure of do_one_mathemu() is also
* inappropriate for these ops (as it has no way to alter the pc,
* for a start) and it might be better to special-case them in do_mathemu().
* Oh, and you'd need to alter the traps.c code so it didn't try to
* fpsave() and fpload(). If there's genuinely no FPU then there's
* probably bits of kernel stuff that just won't work anyway...
*/
/* The Vn labels indicate what version of the SPARC architecture gas thinks
* each insn is. This is from the binutils source :->
*/
/* quadword instructions */
FLOATFUNC(FSQRTQ); /* v8 NYI */
FLOATFUNC(FADDQ); /* v8 */
FLOATFUNC(FSUBQ); /* v8 */
FLOATFUNC(FMULQ); /* v8 NYI */
FLOATFUNC(FDIVQ); /* v8 NYI */
FLOATFUNC(FDMULQ); /* v8 NYI */
FLOATFUNC(FQTOS); /* v8 */
FLOATFUNC(FQTOD); /* v8 */
FLOATFUNC(FITOQ); /* v8 */
FLOATFUNC(FSTOQ); /* v8 */
FLOATFUNC(FDTOQ); /* v8 */
FLOATFUNC(FQTOI); /* v8 */
FLOATFUNC(FCMPQ); /* v8 */
FLOATFUNC(FCMPEQ); /* v8 */
/* single/double instructions (subnormal): should all work */
FLOATFUNC(FSQRTS); /* v7 */
FLOATFUNC(FSQRTD); /* v7 */
FLOATFUNC(FADDS); /* v6 */
FLOATFUNC(FADDD); /* v6 */
FLOATFUNC(FSUBS); /* v6 */
FLOATFUNC(FSUBD); /* v6 */
FLOATFUNC(FMULS); /* v6 */
FLOATFUNC(FMULD); /* v6 */
FLOATFUNC(FDIVS); /* v6 */
FLOATFUNC(FDIVD); /* v6 */
FLOATFUNC(FSMULD); /* v8 */
FLOATFUNC(FDTOS); /* v6 */
FLOATFUNC(FSTOD); /* v6 */
FLOATFUNC(FSTOI); /* v6 */
FLOATFUNC(FDTOI); /* v6 */
FLOATFUNC(FABSS); /* v6 */
FLOATFUNC(FCMPS); /* v6 */
FLOATFUNC(FCMPES); /* v6 */
FLOATFUNC(FCMPD); /* v6 */
FLOATFUNC(FCMPED); /* v6 */
FLOATFUNC(FMOVS); /* v6 */
FLOATFUNC(FNEGS); /* v6 */
FLOATFUNC(FITOS); /* v6 */
FLOATFUNC(FITOD); /* v6 */
#define FSR_TEM_SHIFT 23UL
#define FSR_TEM_MASK (0x1fUL << FSR_TEM_SHIFT)
#define FSR_AEXC_SHIFT 5UL
#define FSR_AEXC_MASK (0x1fUL << FSR_AEXC_SHIFT)
#define FSR_CEXC_SHIFT 0UL
#define FSR_CEXC_MASK (0x1fUL << FSR_CEXC_SHIFT)
static int do_one_mathemu(u32 insn, unsigned long *fsr, unsigned long *fregs);
/* Unlike the Sparc64 version (which has a struct fpustate), we
* pass the taskstruct corresponding to the task which currently owns the
* FPU. This is partly because we don't have the fpustate struct and
* partly because the task owning the FPU isn't always current (as is
* the case for the Sparc64 port). This is probably SMP-related...
* This function returns 1 if all queued insns were emulated successfully.
* The test for unimplemented FPop in kernel mode has been moved into
* kernel/traps.c for simplicity.
*/
int do_mathemu(struct pt_regs *regs, struct task_struct *fpt)
{
/* regs->pc isn't necessarily the PC at which the offending insn is sitting.
* The FPU maintains a queue of FPops which cause traps.
* When it hits an instruction that requires that the trapped op succeeded
* (usually because it reads a reg. that the trapped op wrote) then it
* causes this exception. We need to emulate all the insns on the queue
* and then allow the op to proceed.
* This code should also handle the case where the trap was precise,
* in which case the queue length is zero and regs->pc points at the
* single FPop to be emulated. (this case is untested, though :->)
* You'll need this case if you want to be able to emulate all FPops
* because the FPU either doesn't exist or has been software-disabled.
* [The UltraSPARC makes FP a precise trap; this isn't as stupid as it
* might sound because the Ultra does funky things with a superscalar
* architecture.]
*/
/* You wouldn't believe how often I typed 'ftp' when I meant 'fpt' :-> */
int i;
int retcode = 0; /* assume all succeed */
unsigned long insn;
#ifdef DEBUG_MATHEMU
printk("In do_mathemu()... pc is %08lx\n", regs->pc);
printk("fpqdepth is %ld\n",fpt->tss.fpqdepth);
for (i = 0; i < fpt->tss.fpqdepth; i++)
printk("%d: %08lx at %08lx\n",i,fpt->tss.fpqueue[i].insn, (unsigned long)fpt->tss.fpqueue[i].insn_addr);
#endif
if (fpt->tss.fpqdepth == 0) { /* no queue, guilty insn is at regs->pc */
#ifdef DEBUG_MATHEMU
printk("precise trap at %08lx\n", regs->pc);
#endif
if (!get_user(insn, (u32 *)regs->pc)) {
retcode = do_one_mathemu(insn, &fpt->tss.fsr, fpt->tss.float_regs);
if (retcode) {
/* in this case we need to fix up PC & nPC */
regs->pc = regs->npc;
regs->npc += 4;
}
}
return retcode;
}
/* Normal case: need to empty the queue... */
for (i = 0; i < fpt->tss.fpqdepth; i++)
{
retcode = do_one_mathemu(fpt->tss.fpqueue[i].insn, &(fpt->tss.fsr), fpt->tss.float_regs);
if (!retcode) /* insn failed, no point doing any more */
break;
}
/* Now empty the queue and clear the queue_not_empty flag */
if(retcode)
fpt->tss.fsr &= ~(0x3000 | FSR_CEXC_MASK);
else
fpt->tss.fsr &= ~0x3000;
fpt->tss.fpqdepth = 0;
return retcode;
}
/* All routines returning an exception to raise should detect
* such exceptions _before_ rounding to be consistant with
* the behavior of the hardware in the implemented cases
* (and thus with the recommendations in the V9 architecture
* manual).
*
* We return 0 if a SIGFPE should be sent, 1 otherwise.
*/
static int record_exception(unsigned long *pfsr, int eflag)
{
unsigned long fsr = *pfsr;
int would_trap;
/* Determine if this exception would have generated a trap. */
would_trap = (fsr & ((long)eflag << FSR_TEM_SHIFT)) != 0UL;
/* If trapping, we only want to signal one bit. */
if(would_trap != 0) {
eflag &= ((fsr & FSR_TEM_MASK) >> FSR_TEM_SHIFT);
if((eflag & (eflag - 1)) != 0) {
if(eflag & EFLAG_INVALID)
eflag = EFLAG_INVALID;
else if(eflag & EFLAG_DIVZERO)
eflag = EFLAG_DIVZERO;
else if(eflag & EFLAG_INEXACT)
eflag = EFLAG_INEXACT;
}
}
/* Set CEXC, here are the rules:
*
* 1) In general all FPU ops will set one and only one
* bit in the CEXC field, this is always the case
* when the IEEE exception trap is enabled in TEM.
*
* 2) As a special case, if an overflow or underflow
* is being signalled, AND the trap is not enabled
* in TEM, then the inexact field shall also be set.
*/
fsr &= ~(FSR_CEXC_MASK);
if(would_trap ||
(eflag & (EFLAG_OVERFLOW | EFLAG_UNDERFLOW)) == 0) {
fsr |= ((long)eflag << FSR_CEXC_SHIFT);
} else {
fsr |= (((long)eflag << FSR_CEXC_SHIFT) |
(EFLAG_INEXACT << FSR_CEXC_SHIFT));
}
/* Set the AEXC field, rules are:
*
* 1) If a trap would not be generated, the
* CEXC just generated is OR'd into the
* existing value of AEXC.
*
* 2) When a trap is generated, AEXC is cleared.
*/
if(would_trap == 0)
fsr |= ((long)eflag << FSR_AEXC_SHIFT);
else
fsr &= ~(FSR_AEXC_MASK);
/* If trapping, indicate fault trap type IEEE. */
if(would_trap != 0)
fsr |= (1UL << 14);
*pfsr = fsr;
return (would_trap ? 0 : 1);
}
static int do_one_mathemu(u32 insn, unsigned long *fsr, unsigned long *fregs)
{
/* Emulate the given insn, updating fsr and fregs appropriately. */
int type = 0;
/* 01 is single, 10 is double, 11 is quad,
* 000011 is rs1, 001100 is rs2, 110000 is rd (00 in rd is fcc)
* 111100000000 tells which ftt that may happen in
* (this field not used on sparc32 code, as we can't
* extract trap type info for ops on the FP queue)
*/
int freg, eflag;
int (*func)(void *,void *,void *) = NULL;
void *rs1 = NULL, *rs2 = NULL, *rd = NULL;
#ifdef DEBUG_MATHEMU
printk("In do_mathemu(), emulating %08lx\n", insn);
#endif
if ((insn & 0xc1f80000) == 0x81a00000) /* FPOP1 */ {
switch ((insn >> 5) & 0x1ff) {
/* QUAD - ftt == 3 */
case 0x001: type = 0x314; func = FMOVS; break;
case 0x005: type = 0x314; func = FNEGS; break;
case 0x009: type = 0x314; func = FABSS; break;
case 0x02b: type = 0x33c; func = FSQRTQ; break;
case 0x043: type = 0x33f; func = FADDQ; break;
case 0x047: type = 0x33f; func = FSUBQ; break;
case 0x04b: type = 0x33f; func = FMULQ; break;
case 0x04f: type = 0x33f; func = FDIVQ; break;
case 0x06e: type = 0x33a; func = FDMULQ; break;
case 0x0c7: type = 0x31c; func = FQTOS; break;
case 0x0cb: type = 0x32c; func = FQTOD; break;
case 0x0cc: type = 0x334; func = FITOQ; break;
case 0x0cd: type = 0x334; func = FSTOQ; break;
case 0x0ce: type = 0x338; func = FDTOQ; break;
case 0x0d3: type = 0x31c; func = FQTOI; break;
/* SUBNORMAL - ftt == 2 */
case 0x029: type = 0x214; func = FSQRTS; break;
case 0x02a: type = 0x228; func = FSQRTD; break;
case 0x041: type = 0x215; func = FADDS; break;
case 0x042: type = 0x22a; func = FADDD; break;
case 0x045: type = 0x215; func = FSUBS; break;
case 0x046: type = 0x22a; func = FSUBD; break;
case 0x049: type = 0x215; func = FMULS; break;
case 0x04a: type = 0x22a; func = FMULD; break;
case 0x04d: type = 0x215; func = FDIVS; break;
case 0x04e: type = 0x22a; func = FDIVD; break;
case 0x069: type = 0x225; func = FSMULD; break;
case 0x0c6: type = 0x218; func = FDTOS; break;
case 0x0c9: type = 0x224; func = FSTOD; break;
case 0x0d1: type = 0x214; func = FSTOI; break;
case 0x0d2: type = 0x218; func = FDTOI; break;
default:
#ifdef DEBUG_MATHEMU
printk("unknown FPop1: %03lx\n",(insn>>5)&0x1ff);
#endif
}
}
else if ((insn & 0xc1f80000) == 0x81a80000) /* FPOP2 */ {
switch ((insn >> 5) & 0x1ff) {
case 0x051: type = 0x305; func = FCMPS; break;
case 0x052: type = 0x30a; func = FCMPD; break;
case 0x053: type = 0x30f; func = FCMPQ; break;
case 0x055: type = 0x305; func = FCMPES; break;
case 0x056: type = 0x30a; func = FCMPED; break;
case 0x057: type = 0x30f; func = FCMPEQ; break;
default:
#ifdef DEBUG_MATHEMU
printk("unknown FPop2: %03lx\n",(insn>>5)&0x1ff);
#endif
}
}
if (!type) { /* oops, didn't recognise that FPop */
printk("attempt to emulate unrecognised FPop!\n");
return 0;
}
/* Decode the registers to be used */
freg = (*fsr >> 14) & 0xf;
*fsr &= ~0x1c000; /* clear the traptype bits */
freg = ((insn >> 14) & 0x1f);
switch (type & 0x3) /* is rs1 single, double or quad? */
{
case 3:
if (freg & 3) /* quadwords must have bits 4&5 of the */
{ /* encoded reg. number set to zero. */
*fsr |= (6 << 14);
return 0; /* simulate invalid_fp_register exception */
}
/* fall through */
case 2:
if (freg & 1) /* doublewords must have bit 5 zeroed */
{
*fsr |= (6 << 14);
return 0;
}
}
rs1 = (void *)&fregs[freg];
freg = (insn & 0x1f);
switch ((type >> 2) & 0x3)
{ /* same again for rs2 */
case 3:
if (freg & 3) /* quadwords must have bits 4&5 of the */
{ /* encoded reg. number set to zero. */
*fsr |= (6 << 14);
return 0; /* simulate invalid_fp_register exception */
}
/* fall through */
case 2:
if (freg & 1) /* doublewords must have bit 5 zeroed */
{
*fsr |= (6 << 14);
return 0;
}
}
rs2 = (void *)&fregs[freg];
freg = ((insn >> 25) & 0x1f);
switch ((type >> 4) & 0x3) /* and finally rd. This one's a bit different */
{
case 0: /* dest is fcc. (this must be FCMPQ or FCMPEQ) */
if (freg) /* V8 has only one set of condition codes, so */
{ /* anything but 0 in the rd field is an error */
*fsr |= (6 << 14); /* (should probably flag as invalid opcode */
return 0; /* but SIGFPE will do :-> ) */
}
rd = (void *)(fsr); /* FCMPQ and FCMPEQ are special and only */
break; /* set bits they're supposed to :-> */
case 3:
if (freg & 3) /* quadwords must have bits 4&5 of the */
{ /* encoded reg. number set to zero. */
*fsr |= (6 << 14);
return 0; /* simulate invalid_fp_register exception */
}
/* fall through */
case 2:
if (freg & 1) /* doublewords must have bit 5 zeroed */
{
*fsr |= (6 << 14);
return 0;
}
/* fall through */
case 1:
rd = (void *)&fregs[freg];
break;
}
#ifdef DEBUG_MATHEMU
printk("executing insn...\n");
#endif
eflag = func(rd, rs2, rs1); /* do the Right Thing */
if(eflag == 0)
return 1; /* success! */
return record_exception(fsr, eflag);
}
|