Commit | Line | Data |
---|---|---|
ffd652c3 ILT |
1 | /* ehopt.c--optimize gcc exception frame information. |
2 | Copyright (C) 1998 Free Software Foundation, Inc. | |
3 | Written by Ian Lance Taylor <ian@cygnus.com>. | |
4 | ||
5 | This file is part of GAS, the GNU Assembler. | |
6 | ||
7 | GAS is free software; you can redistribute it and/or modify | |
8 | it under the terms of the GNU General Public License as published by | |
9 | the Free Software Foundation; either version 2, or (at your option) | |
10 | any later version. | |
11 | ||
12 | GAS is distributed in the hope that it will be useful, | |
13 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
15 | GNU General Public License for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GAS; see the file COPYING. If not, write to the Free | |
19 | Software Foundation, 59 Temple Place - Suite 330, Boston, MA | |
20 | 02111-1307, USA. */ | |
21 | ||
22 | #include "as.h" | |
23 | #include "subsegs.h" | |
24 | ||
25 | /* We include this ELF file, even though we may not be assembling for | |
26 | ELF, since the exception frame information is always in a format | |
27 | derived from DWARF. */ | |
28 | ||
29 | #include "elf/dwarf2.h" | |
30 | ||
31 | /* Try to optimize gcc 2.8 exception frame information. | |
32 | ||
33 | Exception frame information is emitted for every function in the | |
34 | .eh_frame section. Simple information for a function with no | |
35 | exceptions looks like this: | |
36 | ||
37 | __FRAME_BEGIN__: | |
38 | .4byte .LLCIE1 / Length of Common Information Entry | |
39 | .LSCIE1: | |
40 | .4byte 0x0 / CIE Identifier Tag | |
41 | .byte 0x1 / CIE Version | |
42 | .byte 0x0 / CIE Augmentation (none) | |
43 | .byte 0x1 / ULEB128 0x1 (CIE Code Alignment Factor) | |
44 | .byte 0x7c / SLEB128 -4 (CIE Data Alignment Factor) | |
45 | .byte 0x8 / CIE RA Column | |
46 | .byte 0xc / DW_CFA_def_cfa | |
47 | .byte 0x4 / ULEB128 0x4 | |
48 | .byte 0x4 / ULEB128 0x4 | |
49 | .byte 0x88 / DW_CFA_offset, column 0x8 | |
50 | .byte 0x1 / ULEB128 0x1 | |
51 | .align 4 | |
52 | .LECIE1: | |
53 | .set .LLCIE1,.LECIE1-.LSCIE1 / CIE Length Symbol | |
54 | .4byte .LLFDE1 / FDE Length | |
55 | .LSFDE1: | |
56 | .4byte .LSFDE1-__FRAME_BEGIN__ / FDE CIE offset | |
57 | .4byte .LFB1 / FDE initial location | |
58 | .4byte .LFE1-.LFB1 / FDE address range | |
59 | .byte 0x4 / DW_CFA_advance_loc4 | |
60 | .4byte .LCFI0-.LFB1 | |
61 | .byte 0xe / DW_CFA_def_cfa_offset | |
62 | .byte 0x8 / ULEB128 0x8 | |
63 | .byte 0x85 / DW_CFA_offset, column 0x5 | |
64 | .byte 0x2 / ULEB128 0x2 | |
65 | .byte 0x4 / DW_CFA_advance_loc4 | |
66 | .4byte .LCFI1-.LCFI0 | |
67 | .byte 0xd / DW_CFA_def_cfa_register | |
68 | .byte 0x5 / ULEB128 0x5 | |
69 | .byte 0x4 / DW_CFA_advance_loc4 | |
70 | .4byte .LCFI2-.LCFI1 | |
71 | .byte 0x2e / DW_CFA_GNU_args_size | |
72 | .byte 0x4 / ULEB128 0x4 | |
73 | .byte 0x4 / DW_CFA_advance_loc4 | |
74 | .4byte .LCFI3-.LCFI2 | |
75 | .byte 0x2e / DW_CFA_GNU_args_size | |
76 | .byte 0x0 / ULEB128 0x0 | |
77 | .align 4 | |
78 | .LEFDE1: | |
79 | .set .LLFDE1,.LEFDE1-.LSFDE1 / FDE Length Symbol | |
80 | ||
81 | The immediate issue we can address in the assembler is the | |
82 | DW_CFA_advance_loc4 followed by a four byte value. The value is | |
83 | the difference of two addresses in the function. Since gcc does | |
84 | not know this value, it always uses four bytes. We will know the | |
85 | value at the end of assembly, so we can do better. */ | |
86 | ||
87 | static int eh_frame_code_alignment PARAMS ((void)); | |
88 | ||
89 | /* Get the code alignment factor from the CIE. */ | |
90 | ||
91 | static int | |
92 | eh_frame_code_alignment () | |
93 | { | |
94 | static int code_alignment; | |
95 | segT current_seg; | |
96 | subsegT current_subseg; | |
97 | fragS *f; | |
1430b6ed | 98 | fixS *fix; |
ffd652c3 | 99 | int offset; |
1430b6ed | 100 | int eh_state; |
ffd652c3 ILT |
101 | |
102 | if (code_alignment != 0) | |
103 | return code_alignment; | |
104 | ||
105 | /* We should find the CIE at the start of the .eh_frame section. */ | |
106 | ||
107 | current_seg = now_seg; | |
108 | current_subseg = now_subseg; | |
109 | subseg_new (".eh_frame", 0); | |
110 | f = seg_info (now_seg)->frchainP->frch_root; | |
1430b6ed | 111 | fix = seg_info (now_seg)->frchainP->fix_root; |
ffd652c3 ILT |
112 | subseg_set (current_seg, current_subseg); |
113 | ||
114 | /* Look through the frags of the section to find the code alignment. */ | |
115 | ||
116 | /* First make sure that the CIE Identifier Tag is 0. */ | |
117 | ||
118 | offset = 4; | |
119 | while (f != NULL && offset >= f->fr_fix) | |
120 | { | |
121 | offset -= f->fr_fix; | |
122 | f = f->fr_next; | |
123 | } | |
124 | if (f == NULL | |
125 | || f->fr_fix - offset < 4 | |
126 | || f->fr_literal[offset] != 0 | |
127 | || f->fr_literal[offset + 1] != 0 | |
128 | || f->fr_literal[offset + 2] != 0 | |
129 | || f->fr_literal[offset + 3] != 0) | |
130 | { | |
131 | code_alignment = -1; | |
132 | return -1; | |
133 | } | |
134 | ||
135 | /* Next make sure the CIE version number is 1. */ | |
136 | ||
137 | offset += 4; | |
138 | while (f != NULL && offset >= f->fr_fix) | |
139 | { | |
140 | offset -= f->fr_fix; | |
141 | f = f->fr_next; | |
142 | } | |
143 | if (f == NULL | |
144 | || f->fr_fix - offset < 1 | |
145 | || f->fr_literal[offset] != 1) | |
146 | { | |
147 | code_alignment = -1; | |
148 | return -1; | |
149 | } | |
150 | ||
151 | /* Skip the augmentation (a null terminated string). */ | |
152 | ||
153 | ++offset; | |
1430b6ed | 154 | eh_state = 0; |
ffd652c3 ILT |
155 | while (1) |
156 | { | |
157 | while (f != NULL && offset >= f->fr_fix) | |
158 | { | |
159 | offset -= f->fr_fix; | |
160 | f = f->fr_next; | |
161 | } | |
162 | if (f == NULL) | |
163 | { | |
164 | code_alignment = -1; | |
165 | return -1; | |
166 | } | |
167 | while (offset < f->fr_fix && f->fr_literal[offset] != '\0') | |
1430b6ed ILT |
168 | { |
169 | switch (eh_state) | |
170 | { | |
171 | case 0: | |
172 | if (f->fr_literal[offset] == 'e') | |
173 | eh_state = 1; | |
174 | break; | |
175 | case 1: | |
176 | if (f->fr_literal[offset] == 'h') | |
177 | eh_state = 2; | |
178 | break; | |
179 | default: | |
180 | eh_state = 3; | |
181 | break; | |
182 | } | |
183 | ++offset; | |
184 | } | |
ffd652c3 ILT |
185 | if (offset < f->fr_fix) |
186 | break; | |
187 | } | |
188 | ++offset; | |
189 | while (f != NULL && offset >= f->fr_fix) | |
190 | { | |
191 | offset -= f->fr_fix; | |
192 | f = f->fr_next; | |
193 | } | |
194 | if (f == NULL) | |
195 | { | |
196 | code_alignment = -1; | |
197 | return -1; | |
198 | } | |
199 | ||
1430b6ed ILT |
200 | /* If the augmentation field is "eh", then we have to skip a |
201 | pointer. Unfortunately, we don't know how large it is. We find | |
202 | out by looking for a matching fixup. */ | |
203 | if (eh_state == 2) | |
204 | { | |
205 | while (fix != NULL | |
206 | && (fix->fx_frag != f || fix->fx_where != offset)) | |
207 | fix = fix->fx_next; | |
208 | if (fix == NULL) | |
209 | offset += 4; | |
210 | else | |
211 | offset += fix->fx_size; | |
212 | while (f != NULL && offset >= f->fr_fix) | |
213 | { | |
214 | offset -= f->fr_fix; | |
215 | f = f->fr_next; | |
216 | } | |
217 | if (f == NULL) | |
218 | { | |
219 | code_alignment = -1; | |
220 | return -1; | |
221 | } | |
222 | } | |
223 | ||
ffd652c3 ILT |
224 | /* We're now at the code alignment factor, which is a ULEB128. If |
225 | it isn't a single byte, forget it. */ | |
226 | ||
227 | code_alignment = f->fr_literal[offset] & 0xff; | |
228 | if ((code_alignment & 0x80) != 0 || code_alignment == 0) | |
229 | { | |
230 | code_alignment = -1; | |
231 | return -1; | |
232 | } | |
233 | ||
234 | return code_alignment; | |
235 | } | |
236 | ||
237 | /* This function is called from emit_expr. It looks for cases which | |
238 | we can optimize. | |
239 | ||
240 | Rather than try to parse all this information as we read it, we | |
241 | look for a single byte DW_CFA_advance_loc4 followed by a 4 byte | |
242 | difference. We turn that into a rs_cfa_advance frag, and handle | |
243 | those frags at the end of the assembly. If the gcc output changes | |
244 | somewhat, this optimization may stop working. | |
245 | ||
246 | This function returns non-zero if it handled the expression and | |
247 | emit_expr should not do anything, or zero otherwise. It can also | |
248 | change *EXP and *PNBYTES. */ | |
249 | ||
250 | int | |
251 | check_eh_frame (exp, pnbytes) | |
252 | expressionS *exp; | |
253 | unsigned int *pnbytes; | |
254 | { | |
255 | static int saw_advance_loc4; | |
256 | static fragS *loc4_frag; | |
257 | static int loc4_fix; | |
258 | ||
259 | if (flag_traditional_format) | |
260 | { | |
261 | /* Don't optimize. */ | |
262 | } | |
263 | else if (strcmp (segment_name (now_seg), ".eh_frame") != 0) | |
264 | saw_advance_loc4 = 0; | |
265 | else if (*pnbytes == 1 | |
266 | && exp->X_op == O_constant | |
267 | && exp->X_add_number == DW_CFA_advance_loc4) | |
268 | { | |
269 | /* This might be a DW_CFA_advance_loc4. Record the frag and the | |
270 | position within the frag, so that we can change it later. */ | |
271 | saw_advance_loc4 = 1; | |
272 | frag_grow (1); | |
273 | loc4_frag = frag_now; | |
274 | loc4_fix = frag_now_fix (); | |
275 | } | |
276 | else if (saw_advance_loc4 | |
277 | && *pnbytes == 4 | |
278 | && exp->X_op == O_constant) | |
279 | { | |
280 | int ca; | |
281 | ||
282 | /* This is a case which we can optimize. The two symbols being | |
283 | subtracted were in the same frag and the expression was | |
284 | reduced to a constant. We can do the optimization entirely | |
285 | in this function. */ | |
286 | ||
287 | saw_advance_loc4 = 0; | |
288 | ||
289 | ca = eh_frame_code_alignment (); | |
290 | if (ca < 0) | |
291 | { | |
292 | /* Don't optimize. */ | |
293 | } | |
294 | else if (exp->X_add_number % ca == 0 | |
1430b6ed | 295 | && exp->X_add_number / ca < 0x40) |
ffd652c3 ILT |
296 | { |
297 | loc4_frag->fr_literal[loc4_fix] | |
298 | = DW_CFA_advance_loc | (exp->X_add_number / ca); | |
299 | /* No more bytes needed. */ | |
300 | return 1; | |
301 | } | |
302 | else if (exp->X_add_number < 0x100) | |
303 | { | |
304 | loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc1; | |
305 | *pnbytes = 1; | |
306 | } | |
307 | else if (exp->X_add_number < 0x10000) | |
308 | { | |
309 | loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc2; | |
310 | *pnbytes = 2; | |
311 | } | |
312 | } | |
313 | else if (saw_advance_loc4 | |
314 | && *pnbytes == 4 | |
315 | && exp->X_op == O_subtract) | |
316 | { | |
317 | ||
318 | /* This is a case we can optimize. The expression was not | |
319 | reduced, so we can not finish the optimization until the end | |
320 | of the assembly. We set up a variant frag which we handle | |
321 | later. */ | |
322 | ||
323 | saw_advance_loc4 = 0; | |
324 | ||
325 | frag_var (rs_cfa, 4, 0, 0, make_expr_symbol (exp), | |
326 | loc4_fix, (char *) loc4_frag); | |
1430b6ed ILT |
327 | |
328 | return 1; | |
ffd652c3 ILT |
329 | } |
330 | else | |
331 | saw_advance_loc4 = 0; | |
332 | ||
333 | return 0; | |
334 | } | |
335 | ||
336 | /* The function estimates the size of a rs_cfa variant frag based on | |
337 | the current values of the symbols. It is called before the | |
338 | relaxation loop. We set fr_subtype to the expected length. */ | |
339 | ||
340 | int | |
341 | eh_frame_estimate_size_before_relax (frag) | |
342 | fragS *frag; | |
343 | { | |
344 | int ca; | |
345 | offsetT diff; | |
346 | int ret; | |
347 | ||
348 | ca = eh_frame_code_alignment (); | |
349 | diff = resolve_symbol_value (frag->fr_symbol, 0); | |
350 | ||
351 | if (ca < 0) | |
352 | ret = 4; | |
353 | else if (diff % ca == 0 && diff / ca < 0x40) | |
354 | ret = 0; | |
355 | else if (diff < 0x100) | |
356 | ret = 1; | |
357 | else if (diff < 0x10000) | |
358 | ret = 2; | |
359 | else | |
360 | ret = 4; | |
361 | ||
362 | frag->fr_subtype = ret; | |
363 | ||
364 | return ret; | |
365 | } | |
366 | ||
367 | /* This function relaxes a rs_cfa variant frag based on the current | |
368 | values of the symbols. fr_subtype is the current length of the | |
369 | frag. This returns the change in frag length. */ | |
370 | ||
371 | int | |
372 | eh_frame_relax_frag (frag) | |
373 | fragS *frag; | |
374 | { | |
375 | int oldsize, newsize; | |
376 | ||
377 | oldsize = frag->fr_subtype; | |
378 | newsize = eh_frame_estimate_size_before_relax (frag); | |
379 | return newsize - oldsize; | |
380 | } | |
381 | ||
382 | /* This function converts a rs_cfa variant frag into a normal fill | |
383 | frag. This is called after all relaxation has been done. | |
384 | fr_subtype will be the desired length of the frag. */ | |
385 | ||
386 | void | |
387 | eh_frame_convert_frag (frag) | |
388 | fragS *frag; | |
389 | { | |
390 | offsetT diff; | |
391 | fragS *loc4_frag; | |
392 | int loc4_fix; | |
393 | ||
394 | loc4_frag = (fragS *) frag->fr_opcode; | |
395 | loc4_fix = (int) frag->fr_offset; | |
396 | ||
397 | diff = resolve_symbol_value (frag->fr_symbol, 1); | |
398 | ||
399 | if (frag->fr_subtype == 0) | |
400 | { | |
401 | int ca; | |
402 | ||
403 | ca = eh_frame_code_alignment (); | |
404 | assert (ca > 0 && diff % ca == 0 && diff / ca < 0x40); | |
405 | loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc | (diff / ca); | |
406 | } | |
407 | else if (frag->fr_subtype == 1) | |
408 | { | |
409 | assert (diff < 0x100); | |
410 | loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc1; | |
411 | frag->fr_literal[frag->fr_fix] = diff; | |
412 | } | |
413 | else if (frag->fr_subtype == 2) | |
414 | { | |
415 | assert (diff < 0x10000); | |
416 | loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc2; | |
417 | md_number_to_chars (frag->fr_literal + frag->fr_fix, diff, 2); | |
418 | } | |
419 | else | |
420 | md_number_to_chars (frag->fr_literal + frag->fr_fix, diff, 4); | |
421 | ||
422 | frag->fr_fix += frag->fr_subtype; | |
423 | frag->fr_type = rs_fill; | |
1430b6ed | 424 | frag->fr_offset = 0; |
ffd652c3 | 425 | } |