1 #include <linux/module.h>
2 #include <linux/types.h>
3 #include <linux/kernel.h>
4 #include <linux/sched.h>
5 #include <asm/ptrace.h>
7 #include <linux/uaccess.h>
10 #include <math-emu/soft-fp.h>
11 #include <math-emu/single.h>
12 #include <math-emu/double.h>
30 #define FOP_FNC_ADDx 0
31 #define FOP_FNC_CVTQL 0
32 #define FOP_FNC_SUBx 1
33 #define FOP_FNC_MULx 2
34 #define FOP_FNC_DIVx 3
35 #define FOP_FNC_CMPxUN 4
36 #define FOP_FNC_CMPxEQ 5
37 #define FOP_FNC_CMPxLT 6
38 #define FOP_FNC_CMPxLE 7
39 #define FOP_FNC_SQRTx 11
40 #define FOP_FNC_CVTxS 12
41 #define FOP_FNC_CVTxT 14
42 #define FOP_FNC_CVTxQ 15
44 #define MISC_TRAPB 0x0000
45 #define MISC_EXCB 0x0400
47 extern unsigned long alpha_read_fp_reg (unsigned long reg
);
48 extern void alpha_write_fp_reg (unsigned long reg
, unsigned long val
);
49 extern unsigned long alpha_read_fp_reg_s (unsigned long reg
);
50 extern void alpha_write_fp_reg_s (unsigned long reg
, unsigned long val
);
55 MODULE_DESCRIPTION("FP Software completion module");
56 MODULE_LICENSE("GPL v2");
58 extern long (*alpha_fp_emul_imprecise
)(struct pt_regs
*, unsigned long);
59 extern long (*alpha_fp_emul
) (unsigned long pc
);
61 static long (*save_emul_imprecise
)(struct pt_regs
*, unsigned long);
62 static long (*save_emul
) (unsigned long pc
);
64 long do_alpha_fp_emul_imprecise(struct pt_regs
*, unsigned long);
65 long do_alpha_fp_emul(unsigned long);
69 save_emul_imprecise
= alpha_fp_emul_imprecise
;
70 save_emul
= alpha_fp_emul
;
71 alpha_fp_emul_imprecise
= do_alpha_fp_emul_imprecise
;
72 alpha_fp_emul
= do_alpha_fp_emul
;
76 void cleanup_module(void)
78 alpha_fp_emul_imprecise
= save_emul_imprecise
;
79 alpha_fp_emul
= save_emul
;
82 #undef alpha_fp_emul_imprecise
83 #define alpha_fp_emul_imprecise do_alpha_fp_emul_imprecise
85 #define alpha_fp_emul do_alpha_fp_emul
91 * Emulate the floating point instruction at address PC. Returns -1 if the
92 * instruction to be emulated is illegal (such as with the opDEC trap), else
93 * the SI_CODE for a SIGFPE signal, else 0 if everything's ok.
95 * Notice that the kernel does not and cannot use FP regs. This is good
96 * because it means that instead of saving/restoring all fp regs, we simply
97 * stick the result of the operation into the appropriate register.
100 alpha_fp_emul (unsigned long pc
)
103 FP_DECL_S(SA
); FP_DECL_S(SB
); FP_DECL_S(SR
);
104 FP_DECL_D(DA
); FP_DECL_D(DB
); FP_DECL_D(DR
);
106 unsigned long fa
, fb
, fc
, func
, mode
, src
;
107 unsigned long res
, va
, vb
, vc
, swcr
, fpcr
;
111 get_user(insn
, (__u32 __user
*)pc
);
112 fc
= (insn
>> 0) & 0x1f; /* destination register */
113 fb
= (insn
>> 16) & 0x1f;
114 fa
= (insn
>> 21) & 0x1f;
115 func
= (insn
>> 5) & 0xf;
116 src
= (insn
>> 9) & 0x3;
117 mode
= (insn
>> 11) & 0x3;
120 swcr
= swcr_update_status(current_thread_info()->ieee_state
, fpcr
);
123 /* Dynamic -- get rounding mode from fpcr. */
124 mode
= (fpcr
>> FPCR_DYN_SHIFT
) & 3;
129 va
= alpha_read_fp_reg_s(fa
);
130 vb
= alpha_read_fp_reg_s(fb
);
132 FP_UNPACK_SP(SA
, &va
);
133 FP_UNPACK_SP(SB
, &vb
);
137 FP_SUB_S(SR
, SA
, SB
);
141 FP_ADD_S(SR
, SA
, SB
);
145 FP_MUL_S(SR
, SA
, SB
);
149 FP_DIV_S(SR
, SA
, SB
);
159 va
= alpha_read_fp_reg(fa
);
160 vb
= alpha_read_fp_reg(fb
);
162 if ((func
& ~3) == FOP_FNC_CMPxUN
) {
163 FP_UNPACK_RAW_DP(DA
, &va
);
164 FP_UNPACK_RAW_DP(DB
, &vb
);
165 if (!DA_e
&& !_FP_FRAC_ZEROP_1(DA
)) {
166 FP_SET_EXCEPTION(FP_EX_DENORM
);
168 _FP_FRAC_SET_1(DA
, _FP_ZEROFRAC_1
);
170 if (!DB_e
&& !_FP_FRAC_ZEROP_1(DB
)) {
171 FP_SET_EXCEPTION(FP_EX_DENORM
);
173 _FP_FRAC_SET_1(DB
, _FP_ZEROFRAC_1
);
175 FP_CMP_D(res
, DA
, DB
, 3);
176 vc
= 0x4000000000000000UL
;
177 /* CMPTEQ, CMPTUN don't trap on QNaN,
178 while CMPTLT and CMPTLE do */
182 || FP_ISSIGNAN_D(DB
))) {
183 FP_SET_EXCEPTION(FP_EX_INVALID
);
186 case FOP_FNC_CMPxUN
: if (res
!= 3) vc
= 0; break;
187 case FOP_FNC_CMPxEQ
: if (res
) vc
= 0; break;
188 case FOP_FNC_CMPxLT
: if (res
!= -1) vc
= 0; break;
189 case FOP_FNC_CMPxLE
: if ((long)res
> 0) vc
= 0; break;
194 FP_UNPACK_DP(DA
, &va
);
195 FP_UNPACK_DP(DB
, &vb
);
199 FP_SUB_D(DR
, DA
, DB
);
203 FP_ADD_D(DR
, DA
, DB
);
207 FP_MUL_D(DR
, DA
, DB
);
211 FP_DIV_D(DR
, DA
, DB
);
219 /* It is irritating that DEC encoded CVTST with
220 SRC == T_floating. It is also interesting that
221 the bit used to tell the two apart is /U... */
223 FP_CONV(S
,D
,1,1,SR
,DB
);
226 vb
= alpha_read_fp_reg_s(fb
);
227 FP_UNPACK_SP(SB
, &vb
);
230 DR_e
= DB_e
+ (1024 - 128);
231 DR_f
= SB_f
<< (52 - 23);
236 if (DB_c
== FP_CLS_NAN
237 && (_FP_FRAC_HIGH_RAW_D(DB
) & _FP_QNANBIT_D
)) {
238 /* AAHB Table B-2 says QNaN should not trigger INV */
241 FP_TO_INT_ROUND_D(vc
, DB
, 64, 2);
247 vb
= alpha_read_fp_reg(fb
);
251 /* Notice: We can get here only due to an integer
252 overflow. Such overflows are reported as invalid
253 ops. We return the result the hw would have
255 vc
= ((vb
& 0xc0000000) << 32 | /* sign and msb */
256 (vb
& 0x3fffffff) << 29); /* rest of the int */
257 FP_SET_EXCEPTION (FP_EX_INVALID
);
261 FP_FROM_INT_S(SR
, ((long)vb
), 64, long);
265 FP_FROM_INT_D(DR
, ((long)vb
), 64, long);
274 if ((_fex
& FP_EX_UNDERFLOW
) && (swcr
& IEEE_MAP_UMZ
))
276 alpha_write_fp_reg_s(fc
, vc
);
281 if ((_fex
& FP_EX_UNDERFLOW
) && (swcr
& IEEE_MAP_UMZ
))
284 alpha_write_fp_reg(fc
, vc
);
288 * Take the appropriate action for each possible
289 * floating-point result:
291 * - Set the appropriate bits in the FPCR
292 * - If the specified exception is enabled in the FPCR,
293 * return. The caller (entArith) will dispatch
294 * the appropriate signal to the translated program.
296 * In addition, properly track the exception state in software
297 * as described in the Alpha Architecture Handbook section 4.7.7.3.
301 /* Record exceptions in software control word. */
302 swcr
|= (_fex
<< IEEE_STATUS_TO_EXCSUM_SHIFT
);
303 current_thread_info()->ieee_state
304 |= (_fex
<< IEEE_STATUS_TO_EXCSUM_SHIFT
);
306 /* Update hardware control register. */
307 fpcr
&= (~FPCR_MASK
| FPCR_DYN_MASK
);
308 fpcr
|= ieee_swcr_to_fpcr(swcr
);
311 /* Do we generate a signal? */
312 _fex
= _fex
& swcr
& IEEE_TRAP_ENABLE_MASK
;
315 if (_fex
& IEEE_TRAP_ENABLE_DNO
) si_code
= FPE_FLTUND
;
316 if (_fex
& IEEE_TRAP_ENABLE_INE
) si_code
= FPE_FLTRES
;
317 if (_fex
& IEEE_TRAP_ENABLE_UNF
) si_code
= FPE_FLTUND
;
318 if (_fex
& IEEE_TRAP_ENABLE_OVF
) si_code
= FPE_FLTOVF
;
319 if (_fex
& IEEE_TRAP_ENABLE_DZE
) si_code
= FPE_FLTDIV
;
320 if (_fex
& IEEE_TRAP_ENABLE_INV
) si_code
= FPE_FLTINV
;
326 /* We used to write the destination register here, but DEC FORTRAN
327 requires that the result *always* be written... so we do the write
328 immediately after the operations above. */
333 printk(KERN_ERR
"alpha_fp_emul: Invalid FP insn %#x at %#lx\n",
339 alpha_fp_emul_imprecise (struct pt_regs
*regs
, unsigned long write_mask
)
341 unsigned long trigger_pc
= regs
->pc
- 4;
342 unsigned long insn
, opcode
, rc
, si_code
= 0;
345 * Turn off the bits corresponding to registers that are the
346 * target of instructions that set bits in the exception
347 * summary register. We have some slack doing this because a
348 * register that is the target of a trapping instruction can
349 * be written at most once in the trap shadow.
351 * Branches, jumps, TRAPBs, EXCBs and calls to PALcode all
352 * bound the trap shadow, so we need not look any further than
353 * up to the first occurrence of such an instruction.
356 get_user(insn
, (__u32 __user
*)(trigger_pc
));
363 case 0x30 ... 0x3f: /* branches */
367 switch (insn
& 0xffff) {
381 write_mask
&= ~(1UL << rc
);
388 write_mask
&= ~(1UL << (rc
+ 32));
392 /* Re-execute insns in the trap-shadow. */
393 regs
->pc
= trigger_pc
+ 4;
394 si_code
= alpha_fp_emul(trigger_pc
);