1 #include <linux/module.h>
2 #include <linux/types.h>
3 #include <linux/kernel.h>
4 #include <linux/sched.h>
6 #include <asm/uaccess.h>
9 #include <math-emu/soft-fp.h>
10 #include <math-emu/single.h>
11 #include <math-emu/double.h>
29 #define FOP_FNC_ADDx 0
30 #define FOP_FNC_CVTQL 0
31 #define FOP_FNC_SUBx 1
32 #define FOP_FNC_MULx 2
33 #define FOP_FNC_DIVx 3
34 #define FOP_FNC_CMPxUN 4
35 #define FOP_FNC_CMPxEQ 5
36 #define FOP_FNC_CMPxLT 6
37 #define FOP_FNC_CMPxLE 7
38 #define FOP_FNC_SQRTx 11
39 #define FOP_FNC_CVTxS 12
40 #define FOP_FNC_CVTxT 14
41 #define FOP_FNC_CVTxQ 15
43 #define MISC_TRAPB 0x0000
44 #define MISC_EXCB 0x0400
46 extern unsigned long alpha_read_fp_reg (unsigned long reg
);
47 extern void alpha_write_fp_reg (unsigned long reg
, unsigned long val
);
48 extern unsigned long alpha_read_fp_reg_s (unsigned long reg
);
49 extern void alpha_write_fp_reg_s (unsigned long reg
, unsigned long val
);
54 MODULE_DESCRIPTION("FP Software completion module");
56 extern long (*alpha_fp_emul_imprecise
)(struct pt_regs
*, unsigned long);
57 extern long (*alpha_fp_emul
) (unsigned long pc
);
59 static long (*save_emul_imprecise
)(struct pt_regs
*, unsigned long);
60 static long (*save_emul
) (unsigned long pc
);
62 long do_alpha_fp_emul_imprecise(struct pt_regs
*, unsigned long);
63 long do_alpha_fp_emul(unsigned long);
67 save_emul_imprecise
= alpha_fp_emul_imprecise
;
68 save_emul
= alpha_fp_emul
;
69 alpha_fp_emul_imprecise
= do_alpha_fp_emul_imprecise
;
70 alpha_fp_emul
= do_alpha_fp_emul
;
74 void cleanup_module(void)
76 alpha_fp_emul_imprecise
= save_emul_imprecise
;
77 alpha_fp_emul
= save_emul
;
80 #undef alpha_fp_emul_imprecise
81 #define alpha_fp_emul_imprecise do_alpha_fp_emul_imprecise
83 #define alpha_fp_emul do_alpha_fp_emul
89 * Emulate the floating point instruction at address PC. Returns -1 if the
90 * instruction to be emulated is illegal (such as with the opDEC trap), else
91 * the SI_CODE for a SIGFPE signal, else 0 if everything's ok.
93 * Notice that the kernel does not and cannot use FP regs. This is good
94 * because it means that instead of saving/restoring all fp regs, we simply
95 * stick the result of the operation into the appropriate register.
98 alpha_fp_emul (unsigned long pc
)
101 FP_DECL_S(SA
); FP_DECL_S(SB
); FP_DECL_S(SR
);
102 FP_DECL_D(DA
); FP_DECL_D(DB
); FP_DECL_D(DR
);
104 unsigned long fa
, fb
, fc
, func
, mode
, src
;
105 unsigned long res
, va
, vb
, vc
, swcr
, fpcr
;
109 get_user(insn
, (__u32 __user
*)pc
);
110 fc
= (insn
>> 0) & 0x1f; /* destination register */
111 fb
= (insn
>> 16) & 0x1f;
112 fa
= (insn
>> 21) & 0x1f;
113 func
= (insn
>> 5) & 0xf;
114 src
= (insn
>> 9) & 0x3;
115 mode
= (insn
>> 11) & 0x3;
118 swcr
= swcr_update_status(current_thread_info()->ieee_state
, fpcr
);
121 /* Dynamic -- get rounding mode from fpcr. */
122 mode
= (fpcr
>> FPCR_DYN_SHIFT
) & 3;
127 va
= alpha_read_fp_reg_s(fa
);
128 vb
= alpha_read_fp_reg_s(fb
);
130 FP_UNPACK_SP(SA
, &va
);
131 FP_UNPACK_SP(SB
, &vb
);
135 FP_SUB_S(SR
, SA
, SB
);
139 FP_ADD_S(SR
, SA
, SB
);
143 FP_MUL_S(SR
, SA
, SB
);
147 FP_DIV_S(SR
, SA
, SB
);
157 va
= alpha_read_fp_reg(fa
);
158 vb
= alpha_read_fp_reg(fb
);
160 if ((func
& ~3) == FOP_FNC_CMPxUN
) {
161 FP_UNPACK_RAW_DP(DA
, &va
);
162 FP_UNPACK_RAW_DP(DB
, &vb
);
163 if (!DA_e
&& !_FP_FRAC_ZEROP_1(DA
)) {
164 FP_SET_EXCEPTION(FP_EX_DENORM
);
166 _FP_FRAC_SET_1(DA
, _FP_ZEROFRAC_1
);
168 if (!DB_e
&& !_FP_FRAC_ZEROP_1(DB
)) {
169 FP_SET_EXCEPTION(FP_EX_DENORM
);
171 _FP_FRAC_SET_1(DB
, _FP_ZEROFRAC_1
);
173 FP_CMP_D(res
, DA
, DB
, 3);
174 vc
= 0x4000000000000000UL
;
175 /* CMPTEQ, CMPTUN don't trap on QNaN,
176 while CMPTLT and CMPTLE do */
180 || FP_ISSIGNAN_D(DB
))) {
181 FP_SET_EXCEPTION(FP_EX_INVALID
);
184 case FOP_FNC_CMPxUN
: if (res
!= 3) vc
= 0; break;
185 case FOP_FNC_CMPxEQ
: if (res
) vc
= 0; break;
186 case FOP_FNC_CMPxLT
: if (res
!= -1) vc
= 0; break;
187 case FOP_FNC_CMPxLE
: if ((long)res
> 0) vc
= 0; break;
192 FP_UNPACK_DP(DA
, &va
);
193 FP_UNPACK_DP(DB
, &vb
);
197 FP_SUB_D(DR
, DA
, DB
);
201 FP_ADD_D(DR
, DA
, DB
);
205 FP_MUL_D(DR
, DA
, DB
);
209 FP_DIV_D(DR
, DA
, DB
);
217 /* It is irritating that DEC encoded CVTST with
218 SRC == T_floating. It is also interesting that
219 the bit used to tell the two apart is /U... */
221 FP_CONV(S
,D
,1,1,SR
,DB
);
224 vb
= alpha_read_fp_reg_s(fb
);
225 FP_UNPACK_SP(SB
, &vb
);
228 DR_e
= DB_e
+ (1024 - 128);
229 DR_f
= SB_f
<< (52 - 23);
234 if (DB_c
== FP_CLS_NAN
235 && (_FP_FRAC_HIGH_RAW_D(DB
) & _FP_QNANBIT_D
)) {
236 /* AAHB Table B-2 says QNaN should not trigger INV */
239 FP_TO_INT_ROUND_D(vc
, DB
, 64, 2);
245 vb
= alpha_read_fp_reg(fb
);
249 /* Notice: We can get here only due to an integer
250 overflow. Such overflows are reported as invalid
251 ops. We return the result the hw would have
253 vc
= ((vb
& 0xc0000000) << 32 | /* sign and msb */
254 (vb
& 0x3fffffff) << 29); /* rest of the int */
255 FP_SET_EXCEPTION (FP_EX_INVALID
);
259 FP_FROM_INT_S(SR
, ((long)vb
), 64, long);
263 FP_FROM_INT_D(DR
, ((long)vb
), 64, long);
272 if ((_fex
& FP_EX_UNDERFLOW
) && (swcr
& IEEE_MAP_UMZ
))
274 alpha_write_fp_reg_s(fc
, vc
);
279 if ((_fex
& FP_EX_UNDERFLOW
) && (swcr
& IEEE_MAP_UMZ
))
282 alpha_write_fp_reg(fc
, vc
);
286 * Take the appropriate action for each possible
287 * floating-point result:
289 * - Set the appropriate bits in the FPCR
290 * - If the specified exception is enabled in the FPCR,
291 * return. The caller (entArith) will dispatch
292 * the appropriate signal to the translated program.
294 * In addition, properly track the exception state in software
295 * as described in the Alpha Architecture Handbook section 4.7.7.3.
299 /* Record exceptions in software control word. */
300 swcr
|= (_fex
<< IEEE_STATUS_TO_EXCSUM_SHIFT
);
301 current_thread_info()->ieee_state
302 |= (_fex
<< IEEE_STATUS_TO_EXCSUM_SHIFT
);
304 /* Update hardware control register. */
305 fpcr
&= (~FPCR_MASK
| FPCR_DYN_MASK
);
306 fpcr
|= ieee_swcr_to_fpcr(swcr
);
309 /* Do we generate a signal? */
310 _fex
= _fex
& swcr
& IEEE_TRAP_ENABLE_MASK
;
313 if (_fex
& IEEE_TRAP_ENABLE_DNO
) si_code
= FPE_FLTUND
;
314 if (_fex
& IEEE_TRAP_ENABLE_INE
) si_code
= FPE_FLTRES
;
315 if (_fex
& IEEE_TRAP_ENABLE_UNF
) si_code
= FPE_FLTUND
;
316 if (_fex
& IEEE_TRAP_ENABLE_OVF
) si_code
= FPE_FLTOVF
;
317 if (_fex
& IEEE_TRAP_ENABLE_DZE
) si_code
= FPE_FLTDIV
;
318 if (_fex
& IEEE_TRAP_ENABLE_INV
) si_code
= FPE_FLTINV
;
324 /* We used to write the destination register here, but DEC FORTRAN
325 requires that the result *always* be written... so we do the write
326 immediately after the operations above. */
331 printk(KERN_ERR
"alpha_fp_emul: Invalid FP insn %#x at %#lx\n",
337 alpha_fp_emul_imprecise (struct pt_regs
*regs
, unsigned long write_mask
)
339 unsigned long trigger_pc
= regs
->pc
- 4;
340 unsigned long insn
, opcode
, rc
, si_code
= 0;
343 * Turn off the bits corresponding to registers that are the
344 * target of instructions that set bits in the exception
345 * summary register. We have some slack doing this because a
346 * register that is the target of a trapping instruction can
347 * be written at most once in the trap shadow.
349 * Branches, jumps, TRAPBs, EXCBs and calls to PALcode all
350 * bound the trap shadow, so we need not look any further than
351 * up to the first occurrence of such an instruction.
354 get_user(insn
, (__u32 __user
*)(trigger_pc
));
361 case 0x30 ... 0x3f: /* branches */
365 switch (insn
& 0xffff) {
379 write_mask
&= ~(1UL << rc
);
386 write_mask
&= ~(1UL << (rc
+ 32));
390 /* Re-execute insns in the trap-shadow. */
391 regs
->pc
= trigger_pc
+ 4;
392 si_code
= alpha_fp_emul(trigger_pc
);