xref: /linux/tools/perf/util/annotate-arch/annotate-x86.c (revision d35b0d5877109ecca106cc3835d4d23ac2cdc33c)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <string.h>
3 #include <linux/compiler.h>
4 #include <assert.h>
5 #include <inttypes.h>
6 #include "../annotate-data.h"
7 #include "../debug.h"
8 #include "../disasm.h"
9 #include "../dso.h"
10 #include "../map.h"
11 #include "../string2.h" // strstarts
12 #include "../symbol.h"
13 
14 /*
15  * x86 instruction nmemonic table to parse disasm lines for annotate.
16  * This table is searched twice - one for exact match and another for
17  * match without a size suffix (b, w, l, q) in case of AT&T syntax.
18  *
19  * So this table should not have entries with the suffix unless it's
20  * a complete different instruction than ones without the suffix.
21  */
22 static const struct ins x86__instructions[] = {
23 	{ .name = "adc",	.ops = &mov_ops,  },
24 	{ .name = "add",	.ops = &mov_ops,  },
25 	{ .name = "addsd",	.ops = &mov_ops,  },
26 	{ .name = "and",	.ops = &mov_ops,  },
27 	{ .name = "andpd",	.ops = &mov_ops,  },
28 	{ .name = "andps",	.ops = &mov_ops,  },
29 	{ .name = "bsr",	.ops = &mov_ops,  },
30 	{ .name = "bt",		.ops = &mov_ops,  },
31 	{ .name = "btr",	.ops = &mov_ops,  },
32 	{ .name = "bts",	.ops = &mov_ops,  },
33 	{ .name = "call",	.ops = &call_ops, },
34 	{ .name = "cmovae",	.ops = &mov_ops,  },
35 	{ .name = "cmovbe",	.ops = &mov_ops,  },
36 	{ .name = "cmove",	.ops = &mov_ops,  },
37 	{ .name = "cmp",	.ops = &mov_ops,  },
38 	{ .name = "cmpxch",	.ops = &mov_ops,  },
39 	{ .name = "cmpxchg",	.ops = &mov_ops,  },
40 	{ .name = "cs",		.ops = &mov_ops,  },
41 	{ .name = "dec",	.ops = &dec_ops,  },
42 	{ .name = "divsd",	.ops = &mov_ops,  },
43 	{ .name = "divss",	.ops = &mov_ops,  },
44 	{ .name = "gs",		.ops = &mov_ops,  },
45 	{ .name = "imul",	.ops = &mov_ops,  },
46 	{ .name = "inc",	.ops = &dec_ops,  },
47 	{ .name = "ja",		.ops = &jump_ops, },
48 	{ .name = "jae",	.ops = &jump_ops, },
49 	{ .name = "jb",		.ops = &jump_ops, },
50 	{ .name = "jbe",	.ops = &jump_ops, },
51 	{ .name = "jc",		.ops = &jump_ops, },
52 	{ .name = "jcxz",	.ops = &jump_ops, },
53 	{ .name = "je",		.ops = &jump_ops, },
54 	{ .name = "jecxz",	.ops = &jump_ops, },
55 	{ .name = "jg",		.ops = &jump_ops, },
56 	{ .name = "jge",	.ops = &jump_ops, },
57 	{ .name = "jl",		.ops = &jump_ops, },
58 	{ .name = "jle",	.ops = &jump_ops, },
59 	{ .name = "jmp",	.ops = &jump_ops, },
60 	{ .name = "jna",	.ops = &jump_ops, },
61 	{ .name = "jnae",	.ops = &jump_ops, },
62 	{ .name = "jnb",	.ops = &jump_ops, },
63 	{ .name = "jnbe",	.ops = &jump_ops, },
64 	{ .name = "jnc",	.ops = &jump_ops, },
65 	{ .name = "jne",	.ops = &jump_ops, },
66 	{ .name = "jng",	.ops = &jump_ops, },
67 	{ .name = "jnge",	.ops = &jump_ops, },
68 	{ .name = "jnl",	.ops = &jump_ops, },
69 	{ .name = "jnle",	.ops = &jump_ops, },
70 	{ .name = "jno",	.ops = &jump_ops, },
71 	{ .name = "jnp",	.ops = &jump_ops, },
72 	{ .name = "jns",	.ops = &jump_ops, },
73 	{ .name = "jnz",	.ops = &jump_ops, },
74 	{ .name = "jo",		.ops = &jump_ops, },
75 	{ .name = "jp",		.ops = &jump_ops, },
76 	{ .name = "jpe",	.ops = &jump_ops, },
77 	{ .name = "jpo",	.ops = &jump_ops, },
78 	{ .name = "jrcxz",	.ops = &jump_ops, },
79 	{ .name = "js",		.ops = &jump_ops, },
80 	{ .name = "jz",		.ops = &jump_ops, },
81 	{ .name = "lea",	.ops = &mov_ops,  },
82 	{ .name = "lock",	.ops = &lock_ops, },
83 	{ .name = "mov",	.ops = &mov_ops,  },
84 	{ .name = "movapd",	.ops = &mov_ops,  },
85 	{ .name = "movaps",	.ops = &mov_ops,  },
86 	{ .name = "movdqa",	.ops = &mov_ops,  },
87 	{ .name = "movdqu",	.ops = &mov_ops,  },
88 	{ .name = "movsb",	.ops = &mov_ops,  },
89 	{ .name = "movsd",	.ops = &mov_ops,  },
90 	{ .name = "movsl",	.ops = &mov_ops,  },
91 	{ .name = "movss",	.ops = &mov_ops,  },
92 	{ .name = "movsw",	.ops = &mov_ops,  },
93 	{ .name = "movupd",	.ops = &mov_ops,  },
94 	{ .name = "movups",	.ops = &mov_ops,  },
95 	{ .name = "movzb",	.ops = &mov_ops,  },
96 	{ .name = "movzl",	.ops = &mov_ops,  },
97 	{ .name = "movzw",	.ops = &mov_ops,  },
98 	{ .name = "mulsd",	.ops = &mov_ops,  },
99 	{ .name = "mulss",	.ops = &mov_ops,  },
100 	{ .name = "nop",	.ops = &nop_ops,  },
101 	{ .name = "or",		.ops = &mov_ops,  },
102 	{ .name = "orps",	.ops = &mov_ops,  },
103 	{ .name = "paddq",	.ops = &mov_ops,  },
104 	{ .name = "pand",	.ops = &mov_ops,  },
105 	{ .name = "pcmpeqb",	.ops = &mov_ops,  },
106 	{ .name = "por",	.ops = &mov_ops,  },
107 	{ .name = "rcl",	.ops = &mov_ops,  },
108 	{ .name = "ret",	.ops = &ret_ops,  },
109 	{ .name = "sbb",	.ops = &mov_ops,  },
110 	{ .name = "sete",	.ops = &mov_ops,  },
111 	{ .name = "sub",	.ops = &mov_ops,  },
112 	{ .name = "subsd",	.ops = &mov_ops,  },
113 	{ .name = "test",	.ops = &mov_ops,  },
114 	{ .name = "tzcnt",	.ops = &mov_ops,  },
115 	{ .name = "ucomisd",	.ops = &mov_ops,  },
116 	{ .name = "ucomiss",	.ops = &mov_ops,  },
117 	{ .name = "vaddsd",	.ops = &mov_ops,  },
118 	{ .name = "vandpd",	.ops = &mov_ops,  },
119 	{ .name = "vmovdqa",	.ops = &mov_ops,  },
120 	{ .name = "vmovq",	.ops = &mov_ops,  },
121 	{ .name = "vmovsd",	.ops = &mov_ops,  },
122 	{ .name = "vmulsd",	.ops = &mov_ops,  },
123 	{ .name = "vorpd",	.ops = &mov_ops,  },
124 	{ .name = "vsubsd",	.ops = &mov_ops,  },
125 	{ .name = "vucomisd",	.ops = &mov_ops,  },
126 	{ .name = "xadd",	.ops = &mov_ops,  },
127 	{ .name = "xbegin",	.ops = &jump_ops, },
128 	{ .name = "xchg",	.ops = &mov_ops,  },
129 	{ .name = "xor",	.ops = &mov_ops, },
130 	{ .name = "xorpd",	.ops = &mov_ops, },
131 	{ .name = "xorps",	.ops = &mov_ops, },
132 };
133 
134 static bool amd__ins_is_fused(const struct arch *arch, const char *ins1,
135 			      const char *ins2)
136 {
137 	if (strstr(ins2, "jmp"))
138 		return false;
139 
140 	/* Family >= 15h supports cmp/test + branch fusion */
141 	if (arch->family >= 0x15 && (strstarts(ins1, "test") ||
142 	    (strstarts(ins1, "cmp") && !strstr(ins1, "xchg")))) {
143 		return true;
144 	}
145 
146 	/* Family >= 19h supports some ALU + branch fusion */
147 	if (arch->family >= 0x19 && (strstarts(ins1, "add") ||
148 	    strstarts(ins1, "sub") || strstarts(ins1, "and") ||
149 	    strstarts(ins1, "inc") || strstarts(ins1, "dec") ||
150 	    strstarts(ins1, "or") || strstarts(ins1, "xor"))) {
151 		return true;
152 	}
153 
154 	return false;
155 }
156 
157 static bool intel__ins_is_fused(const struct arch *arch, const char *ins1,
158 				const char *ins2)
159 {
160 	if (arch->family != 6 || arch->model < 0x1e || strstr(ins2, "jmp"))
161 		return false;
162 
163 	if (arch->model == 0x1e) {
164 		/* Nehalem */
165 		if ((strstr(ins1, "cmp") && !strstr(ins1, "xchg")) ||
166 		     strstr(ins1, "test")) {
167 			return true;
168 		}
169 	} else {
170 		/* Newer platform */
171 		if ((strstr(ins1, "cmp") && !strstr(ins1, "xchg")) ||
172 		     strstr(ins1, "test") ||
173 		     strstr(ins1, "add") ||
174 		     strstr(ins1, "sub") ||
175 		     strstr(ins1, "and") ||
176 		     strstr(ins1, "inc") ||
177 		     strstr(ins1, "dec")) {
178 			return true;
179 		}
180 	}
181 
182 	return false;
183 }
184 
185 static int x86__cpuid_parse(struct arch *arch, const char *cpuid)
186 {
187 	unsigned int family, model, stepping;
188 	int ret;
189 
190 	/*
191 	 * cpuid = "GenuineIntel,family,model,stepping"
192 	 */
193 	ret = sscanf(cpuid, "%*[^,],%u,%u,%u", &family, &model, &stepping);
194 	if (ret == 3) {
195 		arch->family = family;
196 		arch->model = model;
197 		arch->ins_is_fused = strstarts(cpuid, "AuthenticAMD") ?
198 					amd__ins_is_fused :
199 					intel__ins_is_fused;
200 		return 0;
201 	}
202 
203 	return -1;
204 }
205 
206 #ifdef HAVE_LIBDW_SUPPORT
207 static void invalidate_reg_state(struct type_state_reg *reg)
208 {
209 	reg->kind = TSR_KIND_INVALID;
210 	reg->ok = false;
211 	reg->copied_from = -1;
212 }
213 
214 static void update_insn_state_x86(struct type_state *state,
215 				  struct data_loc_info *dloc, Dwarf_Die *cu_die,
216 				  struct disasm_line *dl)
217 {
218 	struct annotated_insn_loc loc;
219 	struct annotated_op_loc *src = &loc.ops[INSN_OP_SOURCE];
220 	struct annotated_op_loc *dst = &loc.ops[INSN_OP_TARGET];
221 	struct type_state_reg *tsr;
222 	Dwarf_Die type_die;
223 	u32 insn_offset = dl->al.offset;
224 	int fbreg = dloc->fbreg;
225 	int fboff = 0;
226 
227 	if (annotate_get_insn_location(dloc->arch, dl, &loc) < 0)
228 		return;
229 
230 	if (ins__is_call(&dl->ins)) {
231 		struct symbol *func = dl->ops.target.sym;
232 		const char *call_name;
233 
234 		/* Try to resolve the call target name */
235 		if (func)
236 			call_name = func->name;
237 		else
238 			call_name = dl->ops.target.name;
239 
240 		/* __fentry__ will preserve all registers */
241 		if (call_name && !strcmp(call_name, "__fentry__"))
242 			return;
243 
244 		if (call_name)
245 			pr_debug_dtp("call [%x] %s\n", insn_offset, call_name);
246 		else
247 			pr_debug_dtp("call [%x] <unknown>\n", insn_offset);
248 
249 		/* Invalidate caller-saved registers after call (ABI requirement) */
250 		for (unsigned i = 0; i < ARRAY_SIZE(state->regs); i++) {
251 			if (state->regs[i].caller_saved)
252 				invalidate_reg_state(&state->regs[i]);
253 		}
254 
255 		/* Update register with the return type (if any) */
256 		if (call_name && die_find_func_rettype(cu_die, call_name, &type_die)) {
257 			tsr = &state->regs[state->ret_reg];
258 			tsr->type = type_die;
259 			tsr->kind = TSR_KIND_TYPE;
260 			tsr->offset = 0;
261 			tsr->ok = true;
262 
263 			pr_debug_dtp("call [%x] return -> reg%d",
264 				     insn_offset, state->ret_reg);
265 			pr_debug_type_name(&type_die, tsr->kind);
266 		}
267 		return;
268 	}
269 
270 	if (!strncmp(dl->ins.name, "add", 3)) {
271 		u64 imm_value = -1ULL;
272 		int offset;
273 		const char *var_name = NULL;
274 		struct map_symbol *ms = dloc->ms;
275 		u64 ip = ms->sym->start + dl->al.offset;
276 
277 		if (!has_reg_type(state, dst->reg1))
278 			return;
279 
280 		tsr = &state->regs[dst->reg1];
281 		tsr->copied_from = -1;
282 
283 		if (src->imm)
284 			imm_value = src->offset;
285 		else if (has_reg_type(state, src->reg1) &&
286 			 state->regs[src->reg1].kind == TSR_KIND_CONST)
287 			imm_value = state->regs[src->reg1].imm_value;
288 		else if (src->reg1 == DWARF_REG_PC) {
289 			u64 var_addr = annotate_calc_pcrel(dloc->ms, ip,
290 							   src->offset, dl);
291 
292 			if (get_global_var_info(dloc, var_addr,
293 						&var_name, &offset) &&
294 			    !strcmp(var_name, "this_cpu_off") &&
295 			    tsr->kind == TSR_KIND_CONST) {
296 				tsr->kind = TSR_KIND_PERCPU_BASE;
297 				tsr->offset = 0;
298 				tsr->ok = true;
299 				imm_value = tsr->imm_value;
300 			}
301 		}
302 		else
303 			return;
304 
305 		/* Ignore add to non-pointer or non-const types */
306 		if (tsr->kind == TSR_KIND_POINTER ||
307 		    (dwarf_tag(&tsr->type) == DW_TAG_pointer_type &&
308 		     src->reg1 != DWARF_REG_PC && tsr->kind == TSR_KIND_TYPE && !dst->mem_ref)) {
309 			tsr->offset += imm_value;
310 			pr_debug_dtp("add [%x] offset %#"PRIx64" to reg%d",
311 				     insn_offset, imm_value, dst->reg1);
312 			pr_debug_type_name(&tsr->type, tsr->kind);
313 		}
314 
315 		if (tsr->kind == TSR_KIND_CONST)
316 			tsr->imm_value += imm_value;
317 
318 		if (tsr->kind != TSR_KIND_PERCPU_BASE)
319 			return;
320 
321 		if (get_global_var_type(cu_die, dloc, ip, imm_value, &offset,
322 					&type_die) && offset == 0) {
323 			/*
324 			 * This is not a pointer type, but it should be treated
325 			 * as a pointer.
326 			 */
327 			tsr->type = type_die;
328 			tsr->kind = TSR_KIND_PERCPU_POINTER;
329 			tsr->offset = 0;
330 			tsr->ok = true;
331 
332 			pr_debug_dtp("add [%x] percpu %#"PRIx64" -> reg%d",
333 				     insn_offset, imm_value, dst->reg1);
334 			pr_debug_type_name(&tsr->type, tsr->kind);
335 		}
336 		return;
337 	}
338 
339 	if (!strncmp(dl->ins.name, "sub", 3)) {
340 		u64 imm_value = -1ULL;
341 
342 		if (!has_reg_type(state, dst->reg1))
343 			return;
344 
345 		tsr = &state->regs[dst->reg1];
346 		tsr->copied_from = -1;
347 
348 		if (src->imm)
349 			imm_value = src->offset;
350 		else if (has_reg_type(state, src->reg1) &&
351 			 state->regs[src->reg1].kind == TSR_KIND_CONST)
352 			imm_value = state->regs[src->reg1].imm_value;
353 
354 		if (tsr->kind == TSR_KIND_POINTER ||
355 		    (dwarf_tag(&tsr->type) == DW_TAG_pointer_type &&
356 		     src->reg1 != DWARF_REG_PC && tsr->kind == TSR_KIND_TYPE && !dst->mem_ref)) {
357 			tsr->offset -= imm_value;
358 			pr_debug_dtp("sub [%x] offset %#"PRIx64" to reg%d",
359 				     insn_offset, imm_value, dst->reg1);
360 			pr_debug_type_name(&tsr->type, tsr->kind);
361 		}
362 
363 		if (tsr->kind == TSR_KIND_CONST)
364 			tsr->imm_value -= imm_value;
365 
366 		return;
367 	}
368 
369 	if (!strncmp(dl->ins.name, "lea", 3)) {
370 		int sreg = src->reg1;
371 		struct type_state_reg src_tsr;
372 
373 		if (!has_reg_type(state, sreg) ||
374 		    !has_reg_type(state, dst->reg1) ||
375 		    !src->mem_ref)
376 			return;
377 
378 		src_tsr = state->regs[sreg];
379 		tsr = &state->regs[dst->reg1];
380 
381 		invalidate_reg_state(tsr);
382 
383 		/* Case 1: Based on stack pointer or frame pointer */
384 		if (sreg == fbreg || sreg == state->stack_reg) {
385 			struct type_state_stack *stack;
386 			int offset = src->offset - fboff;
387 
388 			stack = find_stack_state(state, offset);
389 			if (!stack)
390 				return;
391 
392 			tsr->type = stack->type;
393 			tsr->kind = TSR_KIND_POINTER;
394 			tsr->offset = offset - stack->offset;
395 			tsr->ok = true;
396 
397 			if (sreg == fbreg) {
398 				pr_debug_dtp("lea [%x] address of -%#x(stack) -> reg%d",
399 					     insn_offset, -src->offset, dst->reg1);
400 			} else {
401 				pr_debug_dtp("lea [%x] address of %#x(reg%d) -> reg%d",
402 					     insn_offset, src->offset, sreg, dst->reg1);
403 			}
404 
405 			pr_debug_type_name(&tsr->type, tsr->kind);
406 		}
407 		/* Case 2: Based on a register holding a typed pointer */
408 		else if (src_tsr.ok && (src_tsr.kind == TSR_KIND_POINTER ||
409 			 (dwarf_tag(&src_tsr.type) == DW_TAG_pointer_type &&
410 			  src_tsr.kind == TSR_KIND_TYPE))) {
411 
412 			if (src_tsr.kind == TSR_KIND_TYPE &&
413 			    __die_get_real_type(&state->regs[sreg].type, &type_die) == NULL)
414 				return;
415 
416 			if (src_tsr.kind == TSR_KIND_POINTER)
417 				type_die = state->regs[sreg].type;
418 
419 			/* Check if the target type has a member at the new offset */
420 			if (die_get_member_type(&type_die,
421 						src->offset + src_tsr.offset, &type_die) == NULL)
422 				return;
423 
424 			tsr->type = src_tsr.type;
425 			tsr->kind = src_tsr.kind;
426 			tsr->offset = src->offset + src_tsr.offset;
427 			tsr->ok = true;
428 
429 			pr_debug_dtp("lea [%x] address of %s%#x(reg%d) -> reg%d",
430 						insn_offset, src->offset < 0 ? "-" : "",
431 						abs(src->offset), sreg, dst->reg1);
432 
433 			pr_debug_type_name(&tsr->type, tsr->kind);
434 		}
435 		return;
436 	}
437 
438 	/* Invalidate register states for other ops which may change pointers */
439 	if (has_reg_type(state, dst->reg1) && !dst->mem_ref &&
440 	    dwarf_tag(&state->regs[dst->reg1].type) == DW_TAG_pointer_type) {
441 		if (!strncmp(dl->ins.name, "imul", 4) || !strncmp(dl->ins.name, "mul", 3) ||
442 		    !strncmp(dl->ins.name, "idiv", 4) || !strncmp(dl->ins.name, "div", 3) ||
443 		    !strncmp(dl->ins.name, "shl", 3)  || !strncmp(dl->ins.name, "shr", 3) ||
444 		    !strncmp(dl->ins.name, "sar", 3)  || !strncmp(dl->ins.name, "and", 3) ||
445 		    !strncmp(dl->ins.name, "or", 2)   || !strncmp(dl->ins.name, "neg", 3) ||
446 		    !strncmp(dl->ins.name, "inc", 3)  || !strncmp(dl->ins.name, "dec", 3)) {
447 			pr_debug_dtp("%s [%x] invalidate reg%d\n",
448 						dl->ins.name, insn_offset, dst->reg1);
449 			invalidate_reg_state(&state->regs[dst->reg1]);
450 			return;
451 		}
452 
453 		if (!strncmp(dl->ins.name, "xor", 3) && dst->reg1 == src->reg1) {
454 			/* xor reg, reg clears the register */
455 			pr_debug_dtp("xor [%x] clear reg%d\n",
456 				     insn_offset, dst->reg1);
457 
458 			state->regs[dst->reg1].kind = TSR_KIND_CONST;
459 			state->regs[dst->reg1].imm_value = 0;
460 			state->regs[dst->reg1].ok = true;
461 			state->regs[dst->reg1].copied_from = -1;
462 			return;
463 		}
464 	}
465 
466 	if (strncmp(dl->ins.name, "mov", 3))
467 		return;
468 
469 	if (dloc->fb_cfa) {
470 		u64 ip = dloc->ms->sym->start + dl->al.offset;
471 		u64 pc = map__rip_2objdump(dloc->ms->map, ip);
472 
473 		if (die_get_cfa(dloc->di->dbg, pc, &fbreg, &fboff) < 0)
474 			fbreg = -1;
475 	}
476 
477 	/* Case 1. register to register or segment:offset to register transfers */
478 	if (!src->mem_ref && !dst->mem_ref) {
479 		if (!has_reg_type(state, dst->reg1))
480 			return;
481 
482 		tsr = &state->regs[dst->reg1];
483 		tsr->copied_from = -1;
484 
485 		if (dso__kernel(map__dso(dloc->ms->map)) &&
486 		    src->segment == INSN_SEG_X86_GS && src->imm) {
487 			u64 ip = dloc->ms->sym->start + dl->al.offset;
488 			u64 var_addr;
489 			int offset;
490 
491 			/*
492 			 * In kernel, %gs points to a per-cpu region for the
493 			 * current CPU.  Access with a constant offset should
494 			 * be treated as a global variable access.
495 			 */
496 			var_addr = src->offset;
497 
498 			if (var_addr == 40) {
499 				tsr->kind = TSR_KIND_CANARY;
500 				tsr->offset = 0;
501 				tsr->ok = true;
502 
503 				pr_debug_dtp("mov [%x] stack canary -> reg%d\n",
504 					     insn_offset, dst->reg1);
505 				return;
506 			}
507 
508 			if (!get_global_var_type(cu_die, dloc, ip, var_addr,
509 						 &offset, &type_die) ||
510 			    !die_get_member_type(&type_die, offset, &type_die)) {
511 				invalidate_reg_state(tsr);
512 				return;
513 			}
514 
515 			tsr->type = type_die;
516 			tsr->kind = TSR_KIND_TYPE;
517 			tsr->offset = 0;
518 			tsr->ok = true;
519 
520 			pr_debug_dtp("mov [%x] this-cpu addr=%#"PRIx64" -> reg%d",
521 				     insn_offset, var_addr, dst->reg1);
522 			pr_debug_type_name(&tsr->type, tsr->kind);
523 			return;
524 		}
525 
526 		if (src->imm) {
527 			tsr->kind = TSR_KIND_CONST;
528 			tsr->imm_value = src->offset;
529 			tsr->offset = 0;
530 			tsr->ok = true;
531 
532 			pr_debug_dtp("mov [%x] imm=%#x -> reg%d\n",
533 				     insn_offset, tsr->imm_value, dst->reg1);
534 			return;
535 		}
536 
537 		if (!has_reg_type(state, src->reg1) ||
538 		    !state->regs[src->reg1].ok) {
539 			invalidate_reg_state(tsr);
540 			return;
541 		}
542 
543 		tsr->type = state->regs[src->reg1].type;
544 		tsr->kind = state->regs[src->reg1].kind;
545 		tsr->imm_value = state->regs[src->reg1].imm_value;
546 		tsr->offset = state->regs[src->reg1].offset;
547 		tsr->ok = true;
548 
549 		/* To copy back the variable type later (hopefully) */
550 		if (tsr->kind == TSR_KIND_TYPE || tsr->kind == TSR_KIND_POINTER)
551 			tsr->copied_from = src->reg1;
552 
553 		pr_debug_dtp("mov [%x] reg%d -> reg%d",
554 			     insn_offset, src->reg1, dst->reg1);
555 		pr_debug_type_name(&tsr->type, tsr->kind);
556 	}
557 	/* Case 2. memory to register transers */
558 	if (src->mem_ref && !dst->mem_ref) {
559 		int sreg = src->reg1;
560 
561 		if (!has_reg_type(state, dst->reg1))
562 			return;
563 
564 		tsr = &state->regs[dst->reg1];
565 		tsr->copied_from = -1;
566 
567 retry:
568 		/* Check stack variables with offset */
569 		if (sreg == fbreg || sreg == state->stack_reg) {
570 			struct type_state_stack *stack;
571 			int offset = src->offset - fboff;
572 
573 			stack = find_stack_state(state, offset);
574 			if (stack == NULL) {
575 				invalidate_reg_state(tsr);
576 				return;
577 			} else if (!stack->compound) {
578 				tsr->type = stack->type;
579 				tsr->kind = stack->kind;
580 				tsr->offset = stack->ptr_offset;
581 				tsr->ok = true;
582 			} else if (die_get_member_type(&stack->type,
583 						       offset - stack->offset,
584 						       &type_die)) {
585 				tsr->type = type_die;
586 				tsr->kind = TSR_KIND_TYPE;
587 				tsr->offset = 0;
588 				tsr->ok = true;
589 			} else {
590 				invalidate_reg_state(tsr);
591 				return;
592 			}
593 
594 			if (sreg == fbreg) {
595 				pr_debug_dtp("mov [%x] -%#x(stack) -> reg%d",
596 					     insn_offset, -offset, dst->reg1);
597 			} else {
598 				pr_debug_dtp("mov [%x] %#x(reg%d) -> reg%d",
599 					     insn_offset, offset, sreg, dst->reg1);
600 			}
601 			pr_debug_type_name(&tsr->type, tsr->kind);
602 		}
603 		/* And then dereference the pointer if it has one */
604 		else if (has_reg_type(state, sreg) && state->regs[sreg].ok &&
605 			 state->regs[sreg].kind == TSR_KIND_TYPE &&
606 			 die_deref_ptr_type(&state->regs[sreg].type,
607 					    src->offset + state->regs[sreg].offset, &type_die)) {
608 			tsr->type = type_die;
609 			tsr->kind = TSR_KIND_TYPE;
610 			tsr->offset = 0;
611 			tsr->ok = true;
612 
613 			pr_debug_dtp("mov [%x] %#x(reg%d) -> reg%d",
614 				     insn_offset, src->offset, sreg, dst->reg1);
615 			pr_debug_type_name(&tsr->type, tsr->kind);
616 		}
617 		/* Handle dereference of TSR_KIND_POINTER registers */
618 		else if (has_reg_type(state, sreg) && state->regs[sreg].ok &&
619 			 state->regs[sreg].kind == TSR_KIND_POINTER &&
620 			 die_get_member_type(&state->regs[sreg].type,
621 					     src->offset + state->regs[sreg].offset, &type_die)) {
622 			tsr->type = state->regs[sreg].type;
623 			tsr->kind = TSR_KIND_TYPE;
624 			tsr->offset = src->offset + state->regs[sreg].offset;
625 			tsr->ok = true;
626 
627 			pr_debug_dtp("mov [%x] addr %#x(reg%d) -> reg%d",
628 				     insn_offset, src->offset, sreg, dst->reg1);
629 			pr_debug_type_name(&tsr->type, tsr->kind);
630 		}
631 		/* Or check if it's a global variable */
632 		else if (sreg == DWARF_REG_PC) {
633 			struct map_symbol *ms = dloc->ms;
634 			u64 ip = ms->sym->start + dl->al.offset;
635 			u64 addr;
636 			int offset;
637 
638 			addr = annotate_calc_pcrel(ms, ip, src->offset, dl);
639 
640 			if (!get_global_var_type(cu_die, dloc, ip, addr, &offset,
641 						 &type_die) ||
642 			    !die_get_member_type(&type_die, offset, &type_die)) {
643 				invalidate_reg_state(tsr);
644 				return;
645 			}
646 
647 			tsr->type = type_die;
648 			tsr->kind = TSR_KIND_TYPE;
649 			tsr->offset = 0;
650 			tsr->ok = true;
651 
652 			pr_debug_dtp("mov [%x] global addr=%"PRIx64" -> reg%d",
653 				     insn_offset, addr, dst->reg1);
654 			pr_debug_type_name(&type_die, tsr->kind);
655 		}
656 		/* And check percpu access with base register */
657 		else if (has_reg_type(state, sreg) &&
658 			 state->regs[sreg].kind == TSR_KIND_PERCPU_BASE) {
659 			u64 ip = dloc->ms->sym->start + dl->al.offset;
660 			u64 var_addr = src->offset;
661 			int offset;
662 
663 			if (src->multi_regs) {
664 				int reg2 = (sreg == src->reg1) ? src->reg2 : src->reg1;
665 
666 				if (has_reg_type(state, reg2) && state->regs[reg2].ok &&
667 				    state->regs[reg2].kind == TSR_KIND_CONST)
668 					var_addr += state->regs[reg2].imm_value;
669 			}
670 
671 			/*
672 			 * In kernel, %gs points to a per-cpu region for the
673 			 * current CPU.  Access with a constant offset should
674 			 * be treated as a global variable access.
675 			 */
676 			if (get_global_var_type(cu_die, dloc, ip, var_addr,
677 						&offset, &type_die) &&
678 			    die_get_member_type(&type_die, offset, &type_die)) {
679 				tsr->type = type_die;
680 				tsr->kind = TSR_KIND_TYPE;
681 				tsr->offset = 0;
682 				tsr->ok = true;
683 
684 				if (src->multi_regs) {
685 					pr_debug_dtp("mov [%x] percpu %#x(reg%d,reg%d) -> reg%d",
686 						     insn_offset, src->offset, src->reg1,
687 						     src->reg2, dst->reg1);
688 				} else {
689 					pr_debug_dtp("mov [%x] percpu %#x(reg%d) -> reg%d",
690 						     insn_offset, src->offset, sreg, dst->reg1);
691 				}
692 				pr_debug_type_name(&tsr->type, tsr->kind);
693 			} else {
694 				invalidate_reg_state(tsr);
695 			}
696 		}
697 		/* And then dereference the calculated pointer if it has one */
698 		else if (has_reg_type(state, sreg) && state->regs[sreg].ok &&
699 			 state->regs[sreg].kind == TSR_KIND_PERCPU_POINTER &&
700 			 die_get_member_type(&state->regs[sreg].type,
701 					     src->offset, &type_die)) {
702 			tsr->type = type_die;
703 			tsr->kind = TSR_KIND_TYPE;
704 			tsr->offset = 0;
705 			tsr->ok = true;
706 
707 			pr_debug_dtp("mov [%x] pointer %#x(reg%d) -> reg%d",
708 				     insn_offset, src->offset, sreg, dst->reg1);
709 			pr_debug_type_name(&tsr->type, tsr->kind);
710 		}
711 		/* Or try another register if any */
712 		else if (src->multi_regs && sreg == src->reg1 &&
713 			 src->reg1 != src->reg2) {
714 			sreg = src->reg2;
715 			goto retry;
716 		}
717 		else {
718 			int offset;
719 			const char *var_name = NULL;
720 
721 			/* it might be per-cpu variable (in kernel) access */
722 			if (src->offset < 0) {
723 				if (get_global_var_info(dloc, (s64)src->offset,
724 							&var_name, &offset) &&
725 				    !strcmp(var_name, "__per_cpu_offset")) {
726 					tsr->kind = TSR_KIND_PERCPU_BASE;
727 					tsr->offset = 0;
728 					tsr->ok = true;
729 
730 					pr_debug_dtp("mov [%x] percpu base reg%d\n",
731 						     insn_offset, dst->reg1);
732 					return;
733 				}
734 			}
735 
736 			invalidate_reg_state(tsr);
737 		}
738 	}
739 	/* Case 3. register to memory transfers */
740 	if (!src->mem_ref && dst->mem_ref) {
741 		if (!has_reg_type(state, src->reg1) ||
742 		    !state->regs[src->reg1].ok)
743 			return;
744 
745 		/* Check stack variables with offset */
746 		if (dst->reg1 == fbreg || dst->reg1 == state->stack_reg) {
747 			struct type_state_stack *stack;
748 			int offset = dst->offset - fboff;
749 
750 			tsr = &state->regs[src->reg1];
751 
752 			stack = find_stack_state(state, offset);
753 			if (stack) {
754 				/*
755 				 * The source register is likely to hold a type
756 				 * of member if it's a compound type.  Do not
757 				 * update the stack variable type since we can
758 				 * get the member type later by using the
759 				 * die_get_member_type().
760 				 */
761 				if (!stack->compound)
762 					set_stack_state(stack, offset, tsr->kind,
763 							&tsr->type, tsr->offset);
764 			} else {
765 				findnew_stack_state(state, offset, tsr->kind,
766 						    &tsr->type, tsr->offset);
767 			}
768 
769 			if (dst->reg1 == fbreg) {
770 				pr_debug_dtp("mov [%x] reg%d -> -%#x(stack)",
771 					     insn_offset, src->reg1, -offset);
772 			} else {
773 				pr_debug_dtp("mov [%x] reg%d -> %#x(reg%d)",
774 					     insn_offset, src->reg1, offset, dst->reg1);
775 			}
776 			if (tsr->offset != 0) {
777 				pr_debug_dtp(" reg%d offset %#x ->",
778 					src->reg1, tsr->offset);
779 			}
780 
781 			pr_debug_type_name(&tsr->type, tsr->kind);
782 		}
783 		/*
784 		 * Ignore other transfers since it'd set a value in a struct
785 		 * and won't change the type.
786 		 */
787 	}
788 	/* Case 4. memory to memory transfers (not handled for now) */
789 }
790 #endif
791 
792 const struct arch *arch__new_x86(const struct e_machine_and_e_flags *id, const char *cpuid)
793 {
794 	struct arch *arch = zalloc(sizeof(*arch));
795 
796 	if (!arch)
797 		return NULL;
798 
799 	arch->name = "x86";
800 	arch->id = *id;
801 	if (cpuid) {
802 		if (x86__cpuid_parse(arch, cpuid)) {
803 			errno = SYMBOL_ANNOTATE_ERRNO__ARCH_INIT_CPUID_PARSING;
804 			return NULL;
805 		}
806 	}
807 	arch->instructions = x86__instructions;
808 	arch->nr_instructions = ARRAY_SIZE(x86__instructions);
809 #ifndef NDEBUG
810 	{
811 		static bool sorted_check;
812 
813 		if (!sorted_check) {
814 			for (size_t i = 0; i < arch->nr_instructions - 1; i++) {
815 				assert(strcmp(arch->instructions[i].name,
816 					      arch->instructions[i + 1].name) <= 0);
817 			}
818 			sorted_check = true;
819 		}
820 	}
821 #endif
822 	arch->sorted_instructions = true;
823 	arch->objdump.comment_char = '#';
824 	arch->objdump.register_char = '%';
825 	arch->objdump.memory_ref_char = '(';
826 	arch->objdump.imm_char = '$';
827 	arch->insn_suffix = "bwlq";
828 #ifdef HAVE_LIBDW_SUPPORT
829 	arch->update_insn_state = update_insn_state_x86;
830 #endif
831 	return arch;
832 }
833