xref: /openbmc/linux/arch/alpha/math-emu/math.c (revision d4d016ca)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/module.h>
3 #include <linux/types.h>
4 #include <linux/kernel.h>
5 #include <linux/sched.h>
6 #include <asm/ptrace.h>
7 
8 #include <linux/uaccess.h>
9 
10 #include "sfp-util.h"
11 #include <math-emu/soft-fp.h>
12 #include <math-emu/single.h>
13 #include <math-emu/double.h>
14 
15 #define	OPC_PAL		0x00
16 #define OPC_INTA	0x10
17 #define OPC_INTL	0x11
18 #define OPC_INTS	0x12
19 #define OPC_INTM	0x13
20 #define OPC_FLTC	0x14
21 #define OPC_FLTV	0x15
22 #define OPC_FLTI	0x16
23 #define OPC_FLTL	0x17
24 #define OPC_MISC	0x18
25 #define	OPC_JSR		0x1a
26 
27 #define FOP_SRC_S	0
28 #define FOP_SRC_T	2
29 #define FOP_SRC_Q	3
30 
31 #define FOP_FNC_ADDx	0
32 #define FOP_FNC_CVTQL	0
33 #define FOP_FNC_SUBx	1
34 #define FOP_FNC_MULx	2
35 #define FOP_FNC_DIVx	3
36 #define FOP_FNC_CMPxUN	4
37 #define FOP_FNC_CMPxEQ	5
38 #define FOP_FNC_CMPxLT	6
39 #define FOP_FNC_CMPxLE	7
40 #define FOP_FNC_SQRTx	11
41 #define FOP_FNC_CVTxS	12
42 #define FOP_FNC_CVTxT	14
43 #define FOP_FNC_CVTxQ	15
44 
45 #define MISC_TRAPB	0x0000
46 #define MISC_EXCB	0x0400
47 
48 extern unsigned long alpha_read_fp_reg (unsigned long reg);
49 extern void alpha_write_fp_reg (unsigned long reg, unsigned long val);
50 extern unsigned long alpha_read_fp_reg_s (unsigned long reg);
51 extern void alpha_write_fp_reg_s (unsigned long reg, unsigned long val);
52 
53 
54 #ifdef MODULE
55 
56 MODULE_DESCRIPTION("FP Software completion module");
57 MODULE_LICENSE("GPL v2");
58 
59 extern long (*alpha_fp_emul_imprecise)(struct pt_regs *, unsigned long);
60 extern long (*alpha_fp_emul) (unsigned long pc);
61 
62 static long (*save_emul_imprecise)(struct pt_regs *, unsigned long);
63 static long (*save_emul) (unsigned long pc);
64 
65 long do_alpha_fp_emul_imprecise(struct pt_regs *, unsigned long);
66 long do_alpha_fp_emul(unsigned long);
67 
alpha_fp_emul_init_module(void)68 static int alpha_fp_emul_init_module(void)
69 {
70 	save_emul_imprecise = alpha_fp_emul_imprecise;
71 	save_emul = alpha_fp_emul;
72 	alpha_fp_emul_imprecise = do_alpha_fp_emul_imprecise;
73 	alpha_fp_emul = do_alpha_fp_emul;
74 	return 0;
75 }
76 module_init(alpha_fp_emul_init_module);
77 
alpha_fp_emul_cleanup_module(void)78 static void alpha_fp_emul_cleanup_module(void)
79 {
80 	alpha_fp_emul_imprecise = save_emul_imprecise;
81 	alpha_fp_emul = save_emul;
82 }
83 module_exit(alpha_fp_emul_cleanup_module);
84 
85 #undef  alpha_fp_emul_imprecise
86 #define alpha_fp_emul_imprecise		do_alpha_fp_emul_imprecise
87 #undef  alpha_fp_emul
88 #define alpha_fp_emul			do_alpha_fp_emul
89 
90 #endif /* MODULE */
91 
92 
93 /*
94  * Emulate the floating point instruction at address PC.  Returns -1 if the
95  * instruction to be emulated is illegal (such as with the opDEC trap), else
96  * the SI_CODE for a SIGFPE signal, else 0 if everything's ok.
97  *
98  * Notice that the kernel does not and cannot use FP regs.  This is good
99  * because it means that instead of saving/restoring all fp regs, we simply
100  * stick the result of the operation into the appropriate register.
101  */
102 long
alpha_fp_emul(unsigned long pc)103 alpha_fp_emul (unsigned long pc)
104 {
105 	FP_DECL_EX;
106 	FP_DECL_S(SA); FP_DECL_S(SB); FP_DECL_S(SR);
107 	FP_DECL_D(DA); FP_DECL_D(DB); FP_DECL_D(DR);
108 
109 	unsigned long fa, fb, fc, func, mode, src;
110 	unsigned long res, va, vb, vc, swcr, fpcr;
111 	__u32 insn;
112 	long si_code;
113 
114 	get_user(insn, (__u32 __user *)pc);
115 	fc     = (insn >>  0) & 0x1f;	/* destination register */
116 	fb     = (insn >> 16) & 0x1f;
117 	fa     = (insn >> 21) & 0x1f;
118 	func   = (insn >>  5) & 0xf;
119 	src    = (insn >>  9) & 0x3;
120 	mode   = (insn >> 11) & 0x3;
121 
122 	fpcr = rdfpcr();
123 	swcr = swcr_update_status(current_thread_info()->ieee_state, fpcr);
124 
125 	if (mode == 3) {
126 		/* Dynamic -- get rounding mode from fpcr.  */
127 		mode = (fpcr >> FPCR_DYN_SHIFT) & 3;
128 	}
129 
130 	switch (src) {
131 	case FOP_SRC_S:
132 		va = alpha_read_fp_reg_s(fa);
133 		vb = alpha_read_fp_reg_s(fb);
134 
135 		FP_UNPACK_SP(SA, &va);
136 		FP_UNPACK_SP(SB, &vb);
137 
138 		switch (func) {
139 		case FOP_FNC_SUBx:
140 			FP_SUB_S(SR, SA, SB);
141 			goto pack_s;
142 
143 		case FOP_FNC_ADDx:
144 			FP_ADD_S(SR, SA, SB);
145 			goto pack_s;
146 
147 		case FOP_FNC_MULx:
148 			FP_MUL_S(SR, SA, SB);
149 			goto pack_s;
150 
151 		case FOP_FNC_DIVx:
152 			FP_DIV_S(SR, SA, SB);
153 			goto pack_s;
154 
155 		case FOP_FNC_SQRTx:
156 			FP_SQRT_S(SR, SB);
157 			goto pack_s;
158 		}
159 		goto bad_insn;
160 
161 	case FOP_SRC_T:
162 		va = alpha_read_fp_reg(fa);
163 		vb = alpha_read_fp_reg(fb);
164 
165 		if ((func & ~3) == FOP_FNC_CMPxUN) {
166 			FP_UNPACK_RAW_DP(DA, &va);
167 			FP_UNPACK_RAW_DP(DB, &vb);
168 			if (!DA_e && !_FP_FRAC_ZEROP_1(DA)) {
169 				FP_SET_EXCEPTION(FP_EX_DENORM);
170 				if (FP_DENORM_ZERO)
171 					_FP_FRAC_SET_1(DA, _FP_ZEROFRAC_1);
172 			}
173 			if (!DB_e && !_FP_FRAC_ZEROP_1(DB)) {
174 				FP_SET_EXCEPTION(FP_EX_DENORM);
175 				if (FP_DENORM_ZERO)
176 					_FP_FRAC_SET_1(DB, _FP_ZEROFRAC_1);
177 			}
178 			FP_CMP_D(res, DA, DB, 3);
179 			vc = 0x4000000000000000UL;
180 			/* CMPTEQ, CMPTUN don't trap on QNaN,
181 			   while CMPTLT and CMPTLE do */
182 			if (res == 3
183 			    && ((func & 3) >= 2
184 				|| FP_ISSIGNAN_D(DA)
185 				|| FP_ISSIGNAN_D(DB))) {
186 				FP_SET_EXCEPTION(FP_EX_INVALID);
187 			}
188 			switch (func) {
189 			case FOP_FNC_CMPxUN: if (res != 3) vc = 0; break;
190 			case FOP_FNC_CMPxEQ: if (res) vc = 0; break;
191 			case FOP_FNC_CMPxLT: if (res != -1) vc = 0; break;
192 			case FOP_FNC_CMPxLE: if ((long)res > 0) vc = 0; break;
193 			}
194 			goto done_d;
195 		}
196 
197 		FP_UNPACK_DP(DA, &va);
198 		FP_UNPACK_DP(DB, &vb);
199 
200 		switch (func) {
201 		case FOP_FNC_SUBx:
202 			FP_SUB_D(DR, DA, DB);
203 			goto pack_d;
204 
205 		case FOP_FNC_ADDx:
206 			FP_ADD_D(DR, DA, DB);
207 			goto pack_d;
208 
209 		case FOP_FNC_MULx:
210 			FP_MUL_D(DR, DA, DB);
211 			goto pack_d;
212 
213 		case FOP_FNC_DIVx:
214 			FP_DIV_D(DR, DA, DB);
215 			goto pack_d;
216 
217 		case FOP_FNC_SQRTx:
218 			FP_SQRT_D(DR, DB);
219 			goto pack_d;
220 
221 		case FOP_FNC_CVTxS:
222 			/* It is irritating that DEC encoded CVTST with
223 			   SRC == T_floating.  It is also interesting that
224 			   the bit used to tell the two apart is /U... */
225 			if (insn & 0x2000) {
226 				FP_CONV(S,D,1,1,SR,DB);
227 				goto pack_s;
228 			} else {
229 				vb = alpha_read_fp_reg_s(fb);
230 				FP_UNPACK_SP(SB, &vb);
231 				DR_c = DB_c;
232 				DR_s = DB_s;
233 				DR_e = DB_e + (1024 - 128);
234 				DR_f = SB_f << (52 - 23);
235 				goto pack_d;
236 			}
237 
238 		case FOP_FNC_CVTxQ:
239 			if (DB_c == FP_CLS_NAN
240 			    && (_FP_FRAC_HIGH_RAW_D(DB) & _FP_QNANBIT_D)) {
241 			  /* AAHB Table B-2 says QNaN should not trigger INV */
242 				vc = 0;
243 			} else
244 				FP_TO_INT_ROUND_D(vc, DB, 64, 2);
245 			goto done_d;
246 		}
247 		goto bad_insn;
248 
249 	case FOP_SRC_Q:
250 		vb = alpha_read_fp_reg(fb);
251 
252 		switch (func) {
253 		case FOP_FNC_CVTQL:
254 			/* Notice: We can get here only due to an integer
255 			   overflow.  Such overflows are reported as invalid
256 			   ops.  We return the result the hw would have
257 			   computed.  */
258 			vc = ((vb & 0xc0000000) << 32 |	/* sign and msb */
259 			      (vb & 0x3fffffff) << 29);	/* rest of the int */
260 			FP_SET_EXCEPTION (FP_EX_INVALID);
261 			goto done_d;
262 
263 		case FOP_FNC_CVTxS:
264 			FP_FROM_INT_S(SR, ((long)vb), 64, long);
265 			goto pack_s;
266 
267 		case FOP_FNC_CVTxT:
268 			FP_FROM_INT_D(DR, ((long)vb), 64, long);
269 			goto pack_d;
270 		}
271 		goto bad_insn;
272 	}
273 	goto bad_insn;
274 
275 pack_s:
276 	FP_PACK_SP(&vc, SR);
277 	if ((_fex & FP_EX_UNDERFLOW) && (swcr & IEEE_MAP_UMZ))
278 		vc = 0;
279 	alpha_write_fp_reg_s(fc, vc);
280 	goto done;
281 
282 pack_d:
283 	FP_PACK_DP(&vc, DR);
284 	if ((_fex & FP_EX_UNDERFLOW) && (swcr & IEEE_MAP_UMZ))
285 		vc = 0;
286 done_d:
287 	alpha_write_fp_reg(fc, vc);
288 	goto done;
289 
290 	/*
291 	 * Take the appropriate action for each possible
292 	 * floating-point result:
293 	 *
294 	 *	- Set the appropriate bits in the FPCR
295 	 *	- If the specified exception is enabled in the FPCR,
296 	 *	  return.  The caller (entArith) will dispatch
297 	 *	  the appropriate signal to the translated program.
298 	 *
299 	 * In addition, properly track the exception state in software
300 	 * as described in the Alpha Architecture Handbook section 4.7.7.3.
301 	 */
302 done:
303 	if (_fex) {
304 		/* Record exceptions in software control word.  */
305 		swcr |= (_fex << IEEE_STATUS_TO_EXCSUM_SHIFT);
306 		current_thread_info()->ieee_state
307 		  |= (_fex << IEEE_STATUS_TO_EXCSUM_SHIFT);
308 
309 		/* Update hardware control register.  */
310 		fpcr &= (~FPCR_MASK | FPCR_DYN_MASK);
311 		fpcr |= ieee_swcr_to_fpcr(swcr);
312 		wrfpcr(fpcr);
313 
314 		/* Do we generate a signal?  */
315 		_fex = _fex & swcr & IEEE_TRAP_ENABLE_MASK;
316 		si_code = 0;
317 		if (_fex) {
318 			if (_fex & IEEE_TRAP_ENABLE_DNO) si_code = FPE_FLTUND;
319 			if (_fex & IEEE_TRAP_ENABLE_INE) si_code = FPE_FLTRES;
320 			if (_fex & IEEE_TRAP_ENABLE_UNF) si_code = FPE_FLTUND;
321 			if (_fex & IEEE_TRAP_ENABLE_OVF) si_code = FPE_FLTOVF;
322 			if (_fex & IEEE_TRAP_ENABLE_DZE) si_code = FPE_FLTDIV;
323 			if (_fex & IEEE_TRAP_ENABLE_INV) si_code = FPE_FLTINV;
324 		}
325 
326 		return si_code;
327 	}
328 
329 	/* We used to write the destination register here, but DEC FORTRAN
330 	   requires that the result *always* be written... so we do the write
331 	   immediately after the operations above.  */
332 
333 	return 0;
334 
335 bad_insn:
336 	printk(KERN_ERR "alpha_fp_emul: Invalid FP insn %#x at %#lx\n",
337 	       insn, pc);
338 	return -1;
339 }
340 
341 long
alpha_fp_emul_imprecise(struct pt_regs * regs,unsigned long write_mask)342 alpha_fp_emul_imprecise (struct pt_regs *regs, unsigned long write_mask)
343 {
344 	unsigned long trigger_pc = regs->pc - 4;
345 	unsigned long insn, opcode, rc, si_code = 0;
346 
347 	/*
348 	 * Turn off the bits corresponding to registers that are the
349 	 * target of instructions that set bits in the exception
350 	 * summary register.  We have some slack doing this because a
351 	 * register that is the target of a trapping instruction can
352 	 * be written at most once in the trap shadow.
353 	 *
354 	 * Branches, jumps, TRAPBs, EXCBs and calls to PALcode all
355 	 * bound the trap shadow, so we need not look any further than
356 	 * up to the first occurrence of such an instruction.
357 	 */
358 	while (write_mask) {
359 		get_user(insn, (__u32 __user *)(trigger_pc));
360 		opcode = insn >> 26;
361 		rc = insn & 0x1f;
362 
363 		switch (opcode) {
364 		      case OPC_PAL:
365 		      case OPC_JSR:
366 		      case 0x30 ... 0x3f:	/* branches */
367 			goto egress;
368 
369 		      case OPC_MISC:
370 			switch (insn & 0xffff) {
371 			      case MISC_TRAPB:
372 			      case MISC_EXCB:
373 				goto egress;
374 
375 			      default:
376 				break;
377 			}
378 			break;
379 
380 		      case OPC_INTA:
381 		      case OPC_INTL:
382 		      case OPC_INTS:
383 		      case OPC_INTM:
384 			write_mask &= ~(1UL << rc);
385 			break;
386 
387 		      case OPC_FLTC:
388 		      case OPC_FLTV:
389 		      case OPC_FLTI:
390 		      case OPC_FLTL:
391 			write_mask &= ~(1UL << (rc + 32));
392 			break;
393 		}
394 		if (!write_mask) {
395 			/* Re-execute insns in the trap-shadow.  */
396 			regs->pc = trigger_pc + 4;
397 			si_code = alpha_fp_emul(trigger_pc);
398 			goto egress;
399 		}
400 		trigger_pc -= 4;
401 	}
402 
403 egress:
404 	return si_code;
405 }
406