2 * arch/powerpc/math-emu/math_efp.c
4 * Copyright (C) 2006-2008, 2010 Freescale Semiconductor, Inc.
6 * Author: Ebony Zhu, <ebony.zhu@freescale.com>
7 * Yu Liu, <yu.liu@freescale.com>
9 * Derived from arch/alpha/math-emu/math.c
10 * arch/powerpc/math-emu/math.c
13 * This file is the exception handler to make E500 SPE instructions
14 * fully comply with IEEE-754 floating point standard.
16 * This program is free software; you can redistribute it and/or
17 * modify it under the terms of the GNU General Public License
18 * as published by the Free Software Foundation; either version
19 * 2 of the License, or (at your option) any later version.
22 #include <linux/types.h>
24 #include <asm/uaccess.h>
27 #define FP_EX_BOOKE_E500_SPE
28 #include <asm/sfp-machine.h>
30 #include <math-emu/soft-fp.h>
31 #include <math-emu/single.h>
32 #include <math-emu/double.h>
47 #define EFSCMPGT 0x2cc
48 #define EFSCMPLT 0x2cd
49 #define EFSCMPEQ 0x2ce
56 #define EFSCTUIZ 0x2d8
57 #define EFSCTSIZ 0x2da
62 #define EVFSNABS 0x285
66 #define EVFSCMPGT 0x28c
67 #define EVFSCMPLT 0x28d
68 #define EVFSCMPEQ 0x28e
69 #define EVFSCTUI 0x294
70 #define EVFSCTSI 0x295
71 #define EVFSCTUF 0x296
72 #define EVFSCTSF 0x297
73 #define EVFSCTUIZ 0x298
74 #define EVFSCTSIZ 0x29a
83 #define EFDCTUIDZ 0x2ea
84 #define EFDCTSIDZ 0x2eb
85 #define EFDCMPGT 0x2ec
86 #define EFDCMPLT 0x2ed
87 #define EFDCMPEQ 0x2ee
93 #define EFDCTUIZ 0x2f8
94 #define EFDCTSIZ 0x2fa
102 #define SIGN_BIT_S (1UL << 31)
103 #define SIGN_BIT_D (1ULL << 63)
104 #define FP_EX_MASK (FP_EX_INEXACT | FP_EX_INVALID | FP_EX_DIVZERO | \
105 FP_EX_UNDERFLOW | FP_EX_OVERFLOW)
107 static int have_e500_cpu_a005_erratum
;
114 static unsigned long insn_type(unsigned long speinsn
)
116 unsigned long ret
= NOTYPE
;
118 switch (speinsn
& 0x7ff) {
119 case EFSABS
: ret
= XA
; break;
120 case EFSADD
: ret
= AB
; break;
121 case EFSCFD
: ret
= XB
; break;
122 case EFSCMPEQ
: ret
= XCR
; break;
123 case EFSCMPGT
: ret
= XCR
; break;
124 case EFSCMPLT
: ret
= XCR
; break;
125 case EFSCTSF
: ret
= XB
; break;
126 case EFSCTSI
: ret
= XB
; break;
127 case EFSCTSIZ
: ret
= XB
; break;
128 case EFSCTUF
: ret
= XB
; break;
129 case EFSCTUI
: ret
= XB
; break;
130 case EFSCTUIZ
: ret
= XB
; break;
131 case EFSDIV
: ret
= AB
; break;
132 case EFSMUL
: ret
= AB
; break;
133 case EFSNABS
: ret
= XA
; break;
134 case EFSNEG
: ret
= XA
; break;
135 case EFSSUB
: ret
= AB
; break;
136 case EFSCFSI
: ret
= XB
; break;
138 case EVFSABS
: ret
= XA
; break;
139 case EVFSADD
: ret
= AB
; break;
140 case EVFSCMPEQ
: ret
= XCR
; break;
141 case EVFSCMPGT
: ret
= XCR
; break;
142 case EVFSCMPLT
: ret
= XCR
; break;
143 case EVFSCTSF
: ret
= XB
; break;
144 case EVFSCTSI
: ret
= XB
; break;
145 case EVFSCTSIZ
: ret
= XB
; break;
146 case EVFSCTUF
: ret
= XB
; break;
147 case EVFSCTUI
: ret
= XB
; break;
148 case EVFSCTUIZ
: ret
= XB
; break;
149 case EVFSDIV
: ret
= AB
; break;
150 case EVFSMUL
: ret
= AB
; break;
151 case EVFSNABS
: ret
= XA
; break;
152 case EVFSNEG
: ret
= XA
; break;
153 case EVFSSUB
: ret
= AB
; break;
155 case EFDABS
: ret
= XA
; break;
156 case EFDADD
: ret
= AB
; break;
157 case EFDCFS
: ret
= XB
; break;
158 case EFDCMPEQ
: ret
= XCR
; break;
159 case EFDCMPGT
: ret
= XCR
; break;
160 case EFDCMPLT
: ret
= XCR
; break;
161 case EFDCTSF
: ret
= XB
; break;
162 case EFDCTSI
: ret
= XB
; break;
163 case EFDCTSIDZ
: ret
= XB
; break;
164 case EFDCTSIZ
: ret
= XB
; break;
165 case EFDCTUF
: ret
= XB
; break;
166 case EFDCTUI
: ret
= XB
; break;
167 case EFDCTUIDZ
: ret
= XB
; break;
168 case EFDCTUIZ
: ret
= XB
; break;
169 case EFDDIV
: ret
= AB
; break;
170 case EFDMUL
: ret
= AB
; break;
171 case EFDNABS
: ret
= XA
; break;
172 case EFDNEG
: ret
= XA
; break;
173 case EFDSUB
: ret
= AB
; break;
179 int do_spe_mathemu(struct pt_regs
*regs
)
184 unsigned long type
, func
, fc
, fa
, fb
, src
, speinsn
;
185 union dw_union vc
, va
, vb
;
187 if (get_user(speinsn
, (unsigned int __user
*) regs
->nip
))
189 if ((speinsn
>> 26) != EFAPU
)
190 return -EINVAL
; /* not an spe instruction */
192 type
= insn_type(speinsn
);
196 func
= speinsn
& 0x7ff;
197 fc
= (speinsn
>> 21) & 0x1f;
198 fa
= (speinsn
>> 16) & 0x1f;
199 fb
= (speinsn
>> 11) & 0x1f;
200 src
= (speinsn
>> 5) & 0x7;
202 vc
.wp
[0] = current
->thread
.evr
[fc
];
203 vc
.wp
[1] = regs
->gpr
[fc
];
204 va
.wp
[0] = current
->thread
.evr
[fa
];
205 va
.wp
[1] = regs
->gpr
[fa
];
206 vb
.wp
[0] = current
->thread
.evr
[fb
];
207 vb
.wp
[1] = regs
->gpr
[fb
];
209 __FPU_FPSCR
= mfspr(SPRN_SPEFSCR
);
211 pr_debug("speinsn:%08lx spefscr:%08lx\n", speinsn
, __FPU_FPSCR
);
212 pr_debug("vc: %08x %08x\n", vc
.wp
[0], vc
.wp
[1]);
213 pr_debug("va: %08x %08x\n", va
.wp
[0], va
.wp
[1]);
214 pr_debug("vb: %08x %08x\n", vb
.wp
[0], vb
.wp
[1]);
218 FP_DECL_S(SA
); FP_DECL_S(SB
); FP_DECL_S(SR
);
223 FP_UNPACK_SP(SA
, va
.wp
+ 1);
225 FP_UNPACK_SP(SB
, vb
.wp
+ 1);
228 FP_UNPACK_SP(SA
, va
.wp
+ 1);
232 pr_debug("SA: %ld %08lx %ld (%ld)\n", SA_s
, SA_f
, SA_e
, SA_c
);
233 pr_debug("SB: %ld %08lx %ld (%ld)\n", SB_s
, SB_f
, SB_e
, SB_c
);
237 vc
.wp
[1] = va
.wp
[1] & ~SIGN_BIT_S
;
241 vc
.wp
[1] = va
.wp
[1] | SIGN_BIT_S
;
245 vc
.wp
[1] = va
.wp
[1] ^ SIGN_BIT_S
;
249 FP_ADD_S(SR
, SA
, SB
);
253 FP_SUB_S(SR
, SA
, SB
);
257 FP_MUL_S(SR
, SA
, SB
);
261 FP_DIV_S(SR
, SA
, SB
);
278 if (!((vb
.wp
[1] >> 23) == 0xff && ((vb
.wp
[1] & 0x7fffff) > 0))) {
280 if (((vb
.wp
[1] >> 23) & 0xff) == 0) {
283 } else if ((vb
.wp
[1] >> 31) == 0) {
284 /* positive normal */
285 vc
.wp
[1] = (func
== EFSCTSF
) ?
286 0x7fffffff : 0xffffffff;
287 } else { /* negative normal */
288 vc
.wp
[1] = (func
== EFSCTSF
) ?
291 } else { /* rB is NaN */
299 FP_UNPACK_DP(DB
, vb
.dp
);
301 pr_debug("DB: %ld %08lx %08lx %ld (%ld)\n",
302 DB_s
, DB_f1
, DB_f0
, DB_e
, DB_c
);
304 FP_CONV(S
, D
, 1, 2, SR
, DB
);
315 _FP_ROUND_ZERO(1, SB
);
317 FP_TO_INT_S(vc
.wp
[1], SB
, 32,
318 (((func
& 0x3) != 0) || SB_s
));
327 pr_debug("SR: %ld %08lx %ld (%ld)\n", SR_s
, SR_f
, SR_e
, SR_c
);
329 FP_PACK_SP(vc
.wp
+ 1, SR
);
333 FP_CMP_S(IR
, SA
, SB
, 3);
334 if (IR
== 3 && (FP_ISSIGNAN_S(SA
) || FP_ISSIGNAN_S(SB
)))
335 FP_SET_EXCEPTION(FP_EX_INVALID
);
345 FP_DECL_D(DA
); FP_DECL_D(DB
); FP_DECL_D(DR
);
350 FP_UNPACK_DP(DA
, va
.dp
);
352 FP_UNPACK_DP(DB
, vb
.dp
);
355 FP_UNPACK_DP(DA
, va
.dp
);
359 pr_debug("DA: %ld %08lx %08lx %ld (%ld)\n",
360 DA_s
, DA_f1
, DA_f0
, DA_e
, DA_c
);
361 pr_debug("DB: %ld %08lx %08lx %ld (%ld)\n",
362 DB_s
, DB_f1
, DB_f0
, DB_e
, DB_c
);
366 vc
.dp
[0] = va
.dp
[0] & ~SIGN_BIT_D
;
370 vc
.dp
[0] = va
.dp
[0] | SIGN_BIT_D
;
374 vc
.dp
[0] = va
.dp
[0] ^ SIGN_BIT_D
;
378 FP_ADD_D(DR
, DA
, DB
);
382 FP_SUB_D(DR
, DA
, DB
);
386 FP_MUL_D(DR
, DA
, DB
);
390 FP_DIV_D(DR
, DA
, DB
);
407 if (!((vb
.wp
[0] >> 20) == 0x7ff &&
408 ((vb
.wp
[0] & 0xfffff) > 0 || (vb
.wp
[1] > 0)))) {
410 if (((vb
.wp
[0] >> 20) & 0x7ff) == 0) {
413 } else if ((vb
.wp
[0] >> 31) == 0) {
414 /* positive normal */
415 vc
.wp
[1] = (func
== EFDCTSF
) ?
416 0x7fffffff : 0xffffffff;
417 } else { /* negative normal */
418 vc
.wp
[1] = (func
== EFDCTSF
) ?
429 FP_UNPACK_SP(SB
, vb
.wp
+ 1);
431 pr_debug("SB: %ld %08lx %ld (%ld)\n",
432 SB_s
, SB_f
, SB_e
, SB_c
);
434 FP_CONV(D
, S
, 2, 1, DR
, SB
);
440 _FP_ROUND_ZERO(2, DB
);
441 FP_TO_INT_D(vc
.dp
[0], DB
, 64, ((func
& 0x1) == 0));
451 _FP_ROUND_ZERO(2, DB
);
453 FP_TO_INT_D(vc
.wp
[1], DB
, 32,
454 (((func
& 0x3) != 0) || DB_s
));
463 pr_debug("DR: %ld %08lx %08lx %ld (%ld)\n",
464 DR_s
, DR_f1
, DR_f0
, DR_e
, DR_c
);
466 FP_PACK_DP(vc
.dp
, DR
);
470 FP_CMP_D(IR
, DA
, DB
, 3);
471 if (IR
== 3 && (FP_ISSIGNAN_D(DA
) || FP_ISSIGNAN_D(DB
)))
472 FP_SET_EXCEPTION(FP_EX_INVALID
);
483 FP_DECL_S(SA0
); FP_DECL_S(SB0
); FP_DECL_S(SR0
);
484 FP_DECL_S(SA1
); FP_DECL_S(SB1
); FP_DECL_S(SR1
);
490 FP_UNPACK_SP(SA0
, va
.wp
);
491 FP_UNPACK_SP(SA1
, va
.wp
+ 1);
493 FP_UNPACK_SP(SB0
, vb
.wp
);
494 FP_UNPACK_SP(SB1
, vb
.wp
+ 1);
497 FP_UNPACK_SP(SA0
, va
.wp
);
498 FP_UNPACK_SP(SA1
, va
.wp
+ 1);
502 pr_debug("SA0: %ld %08lx %ld (%ld)\n",
503 SA0_s
, SA0_f
, SA0_e
, SA0_c
);
504 pr_debug("SA1: %ld %08lx %ld (%ld)\n",
505 SA1_s
, SA1_f
, SA1_e
, SA1_c
);
506 pr_debug("SB0: %ld %08lx %ld (%ld)\n",
507 SB0_s
, SB0_f
, SB0_e
, SB0_c
);
508 pr_debug("SB1: %ld %08lx %ld (%ld)\n",
509 SB1_s
, SB1_f
, SB1_e
, SB1_c
);
513 vc
.wp
[0] = va
.wp
[0] & ~SIGN_BIT_S
;
514 vc
.wp
[1] = va
.wp
[1] & ~SIGN_BIT_S
;
518 vc
.wp
[0] = va
.wp
[0] | SIGN_BIT_S
;
519 vc
.wp
[1] = va
.wp
[1] | SIGN_BIT_S
;
523 vc
.wp
[0] = va
.wp
[0] ^ SIGN_BIT_S
;
524 vc
.wp
[1] = va
.wp
[1] ^ SIGN_BIT_S
;
528 FP_ADD_S(SR0
, SA0
, SB0
);
529 FP_ADD_S(SR1
, SA1
, SB1
);
533 FP_SUB_S(SR0
, SA0
, SB0
);
534 FP_SUB_S(SR1
, SA1
, SB1
);
538 FP_MUL_S(SR0
, SA0
, SB0
);
539 FP_MUL_S(SR1
, SA1
, SB1
);
543 FP_DIV_S(SR0
, SA0
, SB0
);
544 FP_DIV_S(SR1
, SA1
, SB1
);
560 __asm__
__volatile__ ("mtspr 512, %4\n"
563 : "=r" (vc
.wp
[0]), "=r" (vc
.wp
[1])
564 : "r" (vb
.wp
[0]), "r" (vb
.wp
[1]), "r" (0));
568 __asm__
__volatile__ ("mtspr 512, %4\n"
571 : "=r" (vc
.wp
[0]), "=r" (vc
.wp
[1])
572 : "r" (vb
.wp
[0]), "r" (vb
.wp
[1]), "r" (0));
583 _FP_ROUND_ZERO(1, SB0
);
584 _FP_ROUND_ZERO(1, SB1
);
586 FP_TO_INT_S(vc
.wp
[0], SB0
, 32,
587 (((func
& 0x3) != 0) || SB0_s
));
588 FP_TO_INT_S(vc
.wp
[1], SB1
, 32,
589 (((func
& 0x3) != 0) || SB1_s
));
598 pr_debug("SR0: %ld %08lx %ld (%ld)\n",
599 SR0_s
, SR0_f
, SR0_e
, SR0_c
);
600 pr_debug("SR1: %ld %08lx %ld (%ld)\n",
601 SR1_s
, SR1_f
, SR1_e
, SR1_c
);
603 FP_PACK_SP(vc
.wp
, SR0
);
604 FP_PACK_SP(vc
.wp
+ 1, SR1
);
611 FP_CMP_S(IR0
, SA0
, SB0
, 3);
612 FP_CMP_S(IR1
, SA1
, SB1
, 3);
613 if (IR0
== 3 && (FP_ISSIGNAN_S(SA0
) || FP_ISSIGNAN_S(SB0
)))
614 FP_SET_EXCEPTION(FP_EX_INVALID
);
615 if (IR1
== 3 && (FP_ISSIGNAN_S(SA1
) || FP_ISSIGNAN_S(SB1
)))
616 FP_SET_EXCEPTION(FP_EX_INVALID
);
617 ch
= (IR0
== cmp
) ? 1 : 0;
618 cl
= (IR1
== cmp
) ? 1 : 0;
619 IR
= (ch
<< 3) | (cl
<< 2) | ((ch
| cl
) << 1) |
629 regs
->ccr
&= ~(15 << ((7 - ((speinsn
>> 23) & 0x7)) << 2));
630 regs
->ccr
|= (IR
<< ((7 - ((speinsn
>> 23) & 0x7)) << 2));
633 __FPU_FPSCR
&= ~FP_EX_MASK
;
634 __FPU_FPSCR
|= (FP_CUR_EXCEPTIONS
& FP_EX_MASK
);
635 mtspr(SPRN_SPEFSCR
, __FPU_FPSCR
);
637 current
->thread
.evr
[fc
] = vc
.wp
[0];
638 regs
->gpr
[fc
] = vc
.wp
[1];
640 pr_debug("ccr = %08lx\n", regs
->ccr
);
641 pr_debug("cur exceptions = %08x spefscr = %08lx\n",
642 FP_CUR_EXCEPTIONS
, __FPU_FPSCR
);
643 pr_debug("vc: %08x %08x\n", vc
.wp
[0], vc
.wp
[1]);
644 pr_debug("va: %08x %08x\n", va
.wp
[0], va
.wp
[1]);
645 pr_debug("vb: %08x %08x\n", vb
.wp
[0], vb
.wp
[1]);
650 if (have_e500_cpu_a005_erratum
) {
651 /* according to e500 cpu a005 erratum, reissue efp inst */
653 pr_debug("re-issue efp inst: %08lx\n", speinsn
);
657 printk(KERN_ERR
"\nOoops! IEEE-754 compliance handler encountered un-supported instruction.\ninst code: %08lx\n", speinsn
);
661 int speround_handler(struct pt_regs
*regs
)
665 unsigned long speinsn
, type
, fc
;
667 if (get_user(speinsn
, (unsigned int __user
*) regs
->nip
))
669 if ((speinsn
>> 26) != 4)
670 return -EINVAL
; /* not an spe instruction */
672 type
= insn_type(speinsn
& 0x7ff);
673 if (type
== XCR
) return -ENOSYS
;
675 __FPU_FPSCR
= mfspr(SPRN_SPEFSCR
);
676 pr_debug("speinsn:%08lx spefscr:%08lx\n", speinsn
, __FPU_FPSCR
);
678 /* No need to round if the result is exact */
679 if (!(__FPU_FPSCR
& FP_EX_INEXACT
))
682 fc
= (speinsn
>> 21) & 0x1f;
683 s_lo
= regs
->gpr
[fc
] & SIGN_BIT_S
;
684 s_hi
= current
->thread
.evr
[fc
] & SIGN_BIT_S
;
685 fgpr
.wp
[0] = current
->thread
.evr
[fc
];
686 fgpr
.wp
[1] = regs
->gpr
[fc
];
688 pr_debug("round fgpr: %08x %08x\n", fgpr
.wp
[0], fgpr
.wp
[1]);
690 switch ((speinsn
>> 5) & 0x7) {
691 /* Since SPE instructions on E500 core can handle round to nearest
692 * and round toward zero with IEEE-754 complied, we just need
693 * to handle round toward +Inf and round toward -Inf by software.
696 if ((FP_ROUNDMODE
) == FP_RND_PINF
) {
697 if (!s_lo
) fgpr
.wp
[1]++; /* Z > 0, choose Z1 */
698 } else { /* round to -Inf */
699 if (s_lo
) fgpr
.wp
[1]++; /* Z < 0, choose Z2 */
704 if (FP_ROUNDMODE
== FP_RND_PINF
) {
705 if (!s_hi
) fgpr
.dp
[0]++; /* Z > 0, choose Z1 */
706 } else { /* round to -Inf */
707 if (s_hi
) fgpr
.dp
[0]++; /* Z < 0, choose Z2 */
712 if (FP_ROUNDMODE
== FP_RND_PINF
) {
713 if (!s_lo
) fgpr
.wp
[1]++; /* Z_low > 0, choose Z1 */
714 if (!s_hi
) fgpr
.wp
[0]++; /* Z_high word > 0, choose Z1 */
715 } else { /* round to -Inf */
716 if (s_lo
) fgpr
.wp
[1]++; /* Z_low < 0, choose Z2 */
717 if (s_hi
) fgpr
.wp
[0]++; /* Z_high < 0, choose Z2 */
725 current
->thread
.evr
[fc
] = fgpr
.wp
[0];
726 regs
->gpr
[fc
] = fgpr
.wp
[1];
728 pr_debug(" to fgpr: %08x %08x\n", fgpr
.wp
[0], fgpr
.wp
[1]);
733 int __init
spe_mathemu_init(void)
737 pvr
= mfspr(SPRN_PVR
);
739 if ((PVR_VER(pvr
) == PVR_VER_E500V1
) ||
740 (PVR_VER(pvr
) == PVR_VER_E500V2
)) {
745 * E500 revision below 1.1, 2.3, 3.1, 4.1, 5.1
746 * need cpu a005 errata workaround
751 have_e500_cpu_a005_erratum
= 1;
755 have_e500_cpu_a005_erratum
= 1;
761 have_e500_cpu_a005_erratum
= 1;
771 module_init(spe_mathemu_init
);