]>
Commit | Line | Data |
---|---|---|
f8861f5d | 1 | /* tc-bpf.c -- Assembler for the Linux eBPF. |
d87bef3a | 2 | Copyright (C) 2019-2023 Free Software Foundation, Inc. |
f8861f5d JM |
3 | Contributed by Oracle, Inc. |
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 3, 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 | |
19 | the Free Software Foundation, 51 Franklin Street - Fifth Floor, | |
20 | Boston, MA 02110-1301, USA. */ | |
21 | ||
22 | #include "as.h" | |
23 | #include "subsegs.h" | |
24 | #include "symcat.h" | |
d218e7fe | 25 | #include "opcode/bpf.h" |
f8861f5d JM |
26 | #include "elf/common.h" |
27 | #include "elf/bpf.h" | |
28 | #include "dwarf2dbg.h" | |
d218e7fe | 29 | #include "libiberty.h" |
ff5a51b3 | 30 | #include <ctype.h> |
f8861f5d | 31 | |
d218e7fe JM |
32 | /* Data structure representing a parsed BPF instruction. */ |
33 | ||
34 | struct bpf_insn | |
35 | { | |
c2ca88d7 | 36 | enum bpf_insn_id id; |
d218e7fe JM |
37 | int size; /* Instruction size in bytes. */ |
38 | bpf_insn_word opcode; | |
39 | uint8_t dst; | |
40 | uint8_t src; | |
41 | expressionS offset16; | |
42 | expressionS imm32; | |
43 | expressionS imm64; | |
44 | expressionS disp16; | |
45 | expressionS disp32; | |
46 | ||
47 | unsigned int has_dst : 1; | |
48 | unsigned int has_src : 1; | |
49 | unsigned int has_offset16 : 1; | |
50 | unsigned int has_disp16 : 1; | |
51 | unsigned int has_disp32 : 1; | |
52 | unsigned int has_imm32 : 1; | |
53 | unsigned int has_imm64 : 1; | |
249d4715 JM |
54 | |
55 | unsigned int is_relaxable : 1; | |
56 | expressionS *relaxed_exp; | |
d218e7fe JM |
57 | }; |
58 | ||
9d1e07bd | 59 | const char comment_chars[] = ";#"; |
ff5a51b3 | 60 | const char line_comment_chars[] = "#"; |
f8861f5d JM |
61 | const char line_separator_chars[] = "`"; |
62 | const char EXP_CHARS[] = "eE"; | |
63 | const char FLT_CHARS[] = "fFdD"; | |
64 | ||
1802aae8 JM |
65 | /* Like s_lcomm_internal in gas/read.c but the alignment string |
66 | is allowed to be optional. */ | |
67 | ||
68 | static symbolS * | |
69 | pe_lcomm_internal (int needs_align, symbolS *symbolP, addressT size) | |
70 | { | |
71 | addressT align = 0; | |
72 | ||
73 | SKIP_WHITESPACE (); | |
74 | ||
75 | if (needs_align | |
76 | && *input_line_pointer == ',') | |
77 | { | |
78 | align = parse_align (needs_align - 1); | |
79 | ||
80 | if (align == (addressT) -1) | |
81 | return NULL; | |
82 | } | |
83 | else | |
84 | { | |
85 | if (size >= 8) | |
86 | align = 3; | |
87 | else if (size >= 4) | |
88 | align = 2; | |
89 | else if (size >= 2) | |
90 | align = 1; | |
91 | else | |
92 | align = 0; | |
93 | } | |
94 | ||
95 | bss_alloc (symbolP, size, align); | |
96 | return symbolP; | |
97 | } | |
98 | ||
99 | static void | |
100 | pe_lcomm (int needs_align) | |
101 | { | |
102 | s_comm_internal (needs_align * 2, pe_lcomm_internal); | |
103 | } | |
104 | ||
f8861f5d JM |
105 | /* The target specific pseudo-ops which we support. */ |
106 | const pseudo_typeS md_pseudo_table[] = | |
107 | { | |
e0b989a6 JM |
108 | { "half", cons, 2 }, |
109 | { "word", cons, 4 }, | |
110 | { "dword", cons, 8 }, | |
1802aae8 | 111 | { "lcomm", pe_lcomm, 1 }, |
d0044bac | 112 | { NULL, NULL, 0 } |
f8861f5d JM |
113 | }; |
114 | ||
1802aae8 JM |
115 | \f |
116 | ||
f8861f5d JM |
117 | /* Command-line options processing. */ |
118 | ||
119 | enum options | |
120 | { | |
121 | OPTION_LITTLE_ENDIAN = OPTION_MD_BASE, | |
4449c81a | 122 | OPTION_BIG_ENDIAN, |
d218e7fe JM |
123 | OPTION_XBPF, |
124 | OPTION_DIALECT, | |
125 | OPTION_ISA_SPEC, | |
249d4715 | 126 | OPTION_NO_RELAX, |
f8861f5d JM |
127 | }; |
128 | ||
129 | struct option md_longopts[] = | |
130 | { | |
131 | { "EL", no_argument, NULL, OPTION_LITTLE_ENDIAN }, | |
132 | { "EB", no_argument, NULL, OPTION_BIG_ENDIAN }, | |
4449c81a | 133 | { "mxbpf", no_argument, NULL, OPTION_XBPF }, |
d218e7fe JM |
134 | { "mdialect", required_argument, NULL, OPTION_DIALECT}, |
135 | { "misa-spec", required_argument, NULL, OPTION_ISA_SPEC}, | |
249d4715 | 136 | { "mno-relax", no_argument, NULL, OPTION_NO_RELAX}, |
f8861f5d JM |
137 | { NULL, no_argument, NULL, 0 }, |
138 | }; | |
139 | ||
140 | size_t md_longopts_size = sizeof (md_longopts); | |
141 | ||
142 | const char * md_shortopts = ""; | |
143 | ||
d218e7fe JM |
144 | /* BPF supports little-endian and big-endian variants. The following |
145 | global records what endianness to use. It can be configured using | |
146 | command-line options. It defaults to the host endianness | |
147 | initialized in md_begin. */ | |
f8861f5d | 148 | |
f8861f5d | 149 | static int set_target_endian = 0; |
d218e7fe JM |
150 | extern int target_big_endian; |
151 | ||
249d4715 JM |
152 | /* Whether to relax branch instructions. Default is yes. Can be |
153 | changed using the -mno-relax command line option. */ | |
154 | ||
155 | static int do_relax = 1; | |
156 | ||
d218e7fe JM |
157 | /* The ISA specification can be one of BPF_V1, BPF_V2, BPF_V3, BPF_V4 |
158 | or BPF_XPBF. The ISA spec to use can be configured using | |
159 | command-line options. It defaults to the latest BPF spec. */ | |
160 | ||
161 | static int isa_spec = BPF_V4; | |
f8861f5d | 162 | |
d218e7fe JM |
163 | /* The assembler supports two different dialects: "normal" syntax and |
164 | "pseudoc" syntax. The dialect to use can be configured using | |
165 | command-line options. */ | |
4449c81a | 166 | |
d218e7fe JM |
167 | enum target_asm_dialect |
168 | { | |
169 | DIALECT_NORMAL, | |
170 | DIALECT_PSEUDOC | |
171 | }; | |
172 | ||
173 | static int asm_dialect = DIALECT_NORMAL; | |
4449c81a | 174 | |
f8861f5d | 175 | int |
d218e7fe | 176 | md_parse_option (int c, const char * arg) |
f8861f5d JM |
177 | { |
178 | switch (c) | |
179 | { | |
180 | case OPTION_BIG_ENDIAN: | |
181 | set_target_endian = 1; | |
182 | target_big_endian = 1; | |
183 | break; | |
184 | case OPTION_LITTLE_ENDIAN: | |
d218e7fe | 185 | set_target_endian = 0; |
f8861f5d JM |
186 | target_big_endian = 0; |
187 | break; | |
d218e7fe JM |
188 | case OPTION_DIALECT: |
189 | if (strcmp (arg, "normal") == 0) | |
190 | asm_dialect = DIALECT_NORMAL; | |
191 | else if (strcmp (arg, "pseudoc") == 0) | |
192 | asm_dialect = DIALECT_PSEUDOC; | |
193 | else | |
194 | as_fatal (_("-mdialect=%s is not valid. Expected normal or pseudoc"), | |
195 | arg); | |
196 | break; | |
197 | case OPTION_ISA_SPEC: | |
198 | if (strcmp (arg, "v1") == 0) | |
199 | isa_spec = BPF_V1; | |
200 | else if (strcmp (arg, "v2") == 0) | |
201 | isa_spec = BPF_V2; | |
202 | else if (strcmp (arg, "v3") == 0) | |
203 | isa_spec = BPF_V3; | |
204 | else if (strcmp (arg, "v4") == 0) | |
205 | isa_spec = BPF_V4; | |
206 | else if (strcmp (arg, "xbpf") == 0) | |
207 | isa_spec = BPF_XBPF; | |
208 | else | |
209 | as_fatal (_("-misa-spec=%s is not valid. Expected v1, v2, v3, v4 o xbpf"), | |
210 | arg); | |
211 | break; | |
4449c81a | 212 | case OPTION_XBPF: |
d218e7fe JM |
213 | /* This is an alias for -misa-spec=xbpf. */ |
214 | isa_spec = BPF_XBPF; | |
4449c81a | 215 | break; |
249d4715 JM |
216 | case OPTION_NO_RELAX: |
217 | do_relax = 0; | |
218 | break; | |
f8861f5d JM |
219 | default: |
220 | return 0; | |
221 | } | |
222 | ||
223 | return 1; | |
224 | } | |
225 | ||
226 | void | |
227 | md_show_usage (FILE * stream) | |
228 | { | |
229 | fprintf (stream, _("\nBPF options:\n")); | |
230 | fprintf (stream, _("\ | |
d218e7fe JM |
231 | BPF options:\n\ |
232 | -EL generate code for a little endian machine\n\ | |
233 | -EB generate code for a big endian machine\n\ | |
234 | -mdialect=DIALECT set the assembly dialect (normal, pseudoc)\n\ | |
235 | -misa-spec set the BPF ISA spec (v1, v2, v3, v4, xbpf)\n\ | |
236 | -mxbpf alias for -misa-spec=xbpf\n")); | |
f8861f5d JM |
237 | } |
238 | ||
239 | \f | |
d218e7fe JM |
240 | /* This function is called once, at assembler startup time. This |
241 | should set up all the tables, etc that the MD part of the assembler | |
242 | needs. */ | |
ff5a51b3 | 243 | |
f8861f5d JM |
244 | void |
245 | md_begin (void) | |
246 | { | |
f8861f5d JM |
247 | /* If not specified in the command line, use the host |
248 | endianness. */ | |
249 | if (!set_target_endian) | |
250 | { | |
251 | #ifdef WORDS_BIGENDIAN | |
252 | target_big_endian = 1; | |
253 | #else | |
254 | target_big_endian = 0; | |
255 | #endif | |
256 | } | |
257 | ||
ff5a51b3 GM |
258 | /* Ensure that lines can begin with '*' in BPF store pseudoc instruction. */ |
259 | lex_type['*'] |= LEX_BEGIN_NAME; | |
260 | ||
f8861f5d JM |
261 | /* Set the machine type. */ |
262 | bfd_default_set_arch_mach (stdoutput, bfd_arch_bpf, bfd_mach_bpf); | |
263 | } | |
264 | ||
d218e7fe JM |
265 | /* Round up a section size to the appropriate boundary. */ |
266 | ||
f8861f5d JM |
267 | valueT |
268 | md_section_align (segT segment, valueT size) | |
269 | { | |
fd361982 | 270 | int align = bfd_section_alignment (segment); |
f8861f5d JM |
271 | |
272 | return ((size + (1 << align) - 1) & -(1 << align)); | |
273 | } | |
274 | ||
275 | \f | |
276 | /* Functions concerning relocs. */ | |
277 | ||
278 | /* The location from which a PC relative jump should be calculated, | |
279 | given a PC relative reloc. */ | |
280 | ||
281 | long | |
282 | md_pcrel_from_section (fixS *fixP, segT sec) | |
283 | { | |
284 | if (fixP->fx_addsy != (symbolS *) NULL | |
285 | && (! S_IS_DEFINED (fixP->fx_addsy) | |
286 | || (S_GET_SEGMENT (fixP->fx_addsy) != sec) | |
287 | || S_IS_EXTERNAL (fixP->fx_addsy) | |
288 | || S_IS_WEAK (fixP->fx_addsy))) | |
289 | { | |
290 | /* The symbol is undefined (or is defined but not in this section). | |
291 | Let the linker figure it out. */ | |
292 | return 0; | |
293 | } | |
294 | ||
295 | return fixP->fx_where + fixP->fx_frag->fr_address; | |
296 | } | |
297 | ||
298 | /* Write a value out to the object file, using the appropriate endianness. */ | |
299 | ||
300 | void | |
301 | md_number_to_chars (char * buf, valueT val, int n) | |
302 | { | |
303 | if (target_big_endian) | |
304 | number_to_chars_bigendian (buf, val, n); | |
305 | else | |
306 | number_to_chars_littleendian (buf, val, n); | |
307 | } | |
308 | ||
309 | arelent * | |
d218e7fe | 310 | tc_gen_reloc (asection *sec ATTRIBUTE_UNUSED, fixS *fixP) |
f8861f5d | 311 | { |
d218e7fe JM |
312 | bfd_reloc_code_real_type r_type = fixP->fx_r_type; |
313 | arelent *reloc; | |
f8861f5d | 314 | |
d218e7fe | 315 | reloc = XNEW (arelent); |
f8861f5d | 316 | |
d218e7fe JM |
317 | if (fixP->fx_pcrel) |
318 | { | |
319 | r_type = (r_type == BFD_RELOC_8 ? BFD_RELOC_8_PCREL | |
320 | : r_type == BFD_RELOC_16 ? BFD_RELOC_16_PCREL | |
321 | : r_type == BFD_RELOC_24 ? BFD_RELOC_24_PCREL | |
322 | : r_type == BFD_RELOC_32 ? BFD_RELOC_32_PCREL | |
323 | : r_type == BFD_RELOC_64 ? BFD_RELOC_64_PCREL | |
324 | : r_type); | |
325 | } | |
326 | ||
327 | reloc->howto = bfd_reloc_type_lookup (stdoutput, r_type); | |
328 | ||
329 | if (reloc->howto == (reloc_howto_type *) NULL) | |
f8861f5d | 330 | { |
d218e7fe JM |
331 | as_bad_where (fixP->fx_file, fixP->fx_line, |
332 | _("relocation is not supported")); | |
333 | return NULL; | |
f8861f5d | 334 | } |
d218e7fe JM |
335 | |
336 | //XXX gas_assert (!fixP->fx_pcrel == !reloc->howto->pc_relative); | |
337 | ||
338 | reloc->sym_ptr_ptr = XNEW (asymbol *); | |
339 | *reloc->sym_ptr_ptr = symbol_get_bfdsym (fixP->fx_addsy); | |
340 | ||
341 | /* Use fx_offset for these cases. */ | |
342 | if (fixP->fx_r_type == BFD_RELOC_VTABLE_ENTRY | |
343 | || fixP->fx_r_type == BFD_RELOC_VTABLE_INHERIT) | |
344 | reloc->addend = fixP->fx_offset; | |
345 | else | |
346 | reloc->addend = fixP->fx_addnumber; | |
347 | ||
348 | reloc->address = fixP->fx_frag->fr_address + fixP->fx_where; | |
349 | return reloc; | |
f8861f5d | 350 | } |
d218e7fe | 351 | |
f8861f5d | 352 | \f |
249d4715 JM |
353 | /* Relaxations supported by this assembler. */ |
354 | ||
355 | #define RELAX_BRANCH_ENCODE(uncond, constant, length) \ | |
356 | ((relax_substateT) \ | |
357 | (0xc0000000 \ | |
358 | | ((uncond) ? 1 : 0) \ | |
359 | | ((constant) ? 2 : 0) \ | |
360 | | ((length) << 2))) | |
361 | ||
362 | #define RELAX_BRANCH_P(i) (((i) & 0xf0000000) == 0xc0000000) | |
363 | #define RELAX_BRANCH_LENGTH(i) (((i) >> 2) & 0xff) | |
364 | #define RELAX_BRANCH_CONST(i) (((i) & 2) != 0) | |
365 | #define RELAX_BRANCH_UNCOND(i) (((i) & 1) != 0) | |
366 | ||
367 | ||
368 | /* Compute the length of a branch seuqence, and adjust the stored | |
369 | length accordingly. If FRAG is NULL, the worst-case length is | |
370 | returned. */ | |
371 | ||
372 | static unsigned | |
373 | relaxed_branch_length (fragS *fragp, asection *sec, int update) | |
374 | { | |
375 | int length, uncond; | |
376 | ||
377 | if (!fragp) | |
378 | return 8 * 3; | |
379 | ||
380 | uncond = RELAX_BRANCH_UNCOND (fragp->fr_subtype); | |
381 | length = RELAX_BRANCH_LENGTH (fragp->fr_subtype); | |
382 | ||
383 | if (uncond) | |
384 | /* Length is the same for both JA and JAL. */ | |
385 | length = 8; | |
386 | else | |
387 | { | |
388 | if (RELAX_BRANCH_CONST (fragp->fr_subtype)) | |
389 | { | |
390 | int64_t val = fragp->fr_offset; | |
391 | ||
392 | if (val < -32768 || val > 32767) | |
393 | length = 8 * 3; | |
394 | else | |
395 | length = 8; | |
396 | } | |
397 | else if (fragp->fr_symbol != NULL | |
398 | && S_IS_DEFINED (fragp->fr_symbol) | |
399 | && !S_IS_WEAK (fragp->fr_symbol) | |
400 | && sec == S_GET_SEGMENT (fragp->fr_symbol)) | |
401 | { | |
402 | offsetT val = S_GET_VALUE (fragp->fr_symbol) + fragp->fr_offset; | |
403 | ||
404 | /* Convert to 64-bit words, minus one. */ | |
405 | val = (val - 8) / 8; | |
406 | ||
407 | /* See if it fits in the signed 16-bits field. */ | |
408 | if (val < -32768 || val > 32767) | |
409 | length = 8 * 3; | |
410 | else | |
411 | length = 8; | |
412 | } | |
413 | else | |
414 | /* Use short version, and let the linker relax instead, if | |
415 | appropriate and if supported. */ | |
416 | length = 8; | |
417 | } | |
418 | ||
419 | if (update) | |
420 | fragp->fr_subtype = RELAX_BRANCH_ENCODE (uncond, | |
421 | RELAX_BRANCH_CONST (fragp->fr_subtype), | |
422 | length); | |
423 | ||
424 | return length; | |
425 | } | |
426 | ||
427 | /* Estimate the size of a variant frag before relaxing. */ | |
428 | ||
429 | int | |
430 | md_estimate_size_before_relax (fragS *fragp, asection *sec) | |
431 | { | |
432 | return (fragp->fr_var = relaxed_branch_length (fragp, sec, true)); | |
433 | } | |
434 | ||
435 | /* Read a BPF instruction word from BUF. */ | |
436 | ||
437 | static uint64_t | |
438 | read_insn_word (bfd_byte *buf) | |
439 | { | |
440 | return bfd_getb64 (buf); | |
441 | } | |
442 | ||
443 | /* Write the given signed 16-bit value in the given BUFFER using the | |
444 | target endianness. */ | |
445 | ||
446 | static void | |
447 | encode_int16 (int16_t value, char *buffer) | |
448 | { | |
449 | uint16_t val = value; | |
450 | ||
451 | if (target_big_endian) | |
452 | { | |
453 | buffer[0] = (val >> 8) & 0xff; | |
454 | buffer[1] = val & 0xff; | |
455 | } | |
456 | else | |
457 | { | |
458 | buffer[1] = (val >> 8) & 0xff; | |
459 | buffer[0] = val & 0xff; | |
460 | } | |
461 | } | |
462 | ||
463 | /* Write the given signed 32-bit value in the given BUFFER using the | |
464 | target endianness. */ | |
465 | ||
466 | static void | |
467 | encode_int32 (int32_t value, char *buffer) | |
468 | { | |
469 | uint32_t val = value; | |
470 | ||
471 | if (target_big_endian) | |
472 | { | |
473 | buffer[0] = (val >> 24) & 0xff; | |
474 | buffer[1] = (val >> 16) & 0xff; | |
475 | buffer[2] = (val >> 8) & 0xff; | |
476 | buffer[3] = val & 0xff; | |
477 | } | |
478 | else | |
479 | { | |
480 | buffer[3] = (val >> 24) & 0xff; | |
481 | buffer[2] = (val >> 16) & 0xff; | |
482 | buffer[1] = (val >> 8) & 0xff; | |
483 | buffer[0] = value & 0xff; | |
484 | } | |
485 | } | |
486 | ||
487 | /* Write a BPF instruction to BUF. */ | |
488 | ||
489 | static void | |
490 | write_insn_bytes (bfd_byte *buf, char *bytes) | |
491 | { | |
492 | int i; | |
493 | ||
494 | for (i = 0; i < 8; ++i) | |
495 | md_number_to_chars ((char *) buf + i, (valueT) bytes[i], 1); | |
496 | } | |
497 | ||
f8861f5d JM |
498 | /* *FRAGP has been relaxed to its final size, and now needs to have |
499 | the bytes inside it modified to conform to the new size. | |
500 | ||
501 | Called after relaxation is finished. | |
502 | fragP->fr_type == rs_machine_dependent. | |
503 | fragP->fr_subtype is the subtype of what the address relaxed to. */ | |
504 | ||
505 | void | |
506 | md_convert_frag (bfd *abfd ATTRIBUTE_UNUSED, | |
507 | segT sec ATTRIBUTE_UNUSED, | |
249d4715 | 508 | fragS *fragp ATTRIBUTE_UNUSED) |
f8861f5d | 509 | { |
249d4715 JM |
510 | bfd_byte *buf = (bfd_byte *) fragp->fr_literal + fragp->fr_fix; |
511 | expressionS exp; | |
512 | fixS *fixp; | |
513 | bpf_insn_word word; | |
514 | int disp_is_known = 0; | |
515 | int64_t disp_to_target = 0; | |
f8861f5d | 516 | |
249d4715 JM |
517 | uint64_t code; |
518 | ||
519 | gas_assert (RELAX_BRANCH_P (fragp->fr_subtype)); | |
520 | ||
521 | /* Expression to be used in any resulting relocation in the relaxed | |
522 | instructions. */ | |
523 | exp.X_op = O_symbol; | |
524 | exp.X_add_symbol = fragp->fr_symbol; | |
525 | exp.X_add_number = fragp->fr_offset; | |
526 | ||
527 | gas_assert (fragp->fr_var == RELAX_BRANCH_LENGTH (fragp->fr_subtype)); | |
528 | ||
529 | /* Read an instruction word from the instruction to be relaxed, and | |
530 | get the code. */ | |
531 | word = read_insn_word (buf); | |
532 | code = (word >> 60) & 0xf; | |
533 | ||
534 | /* Determine whether the 16-bit displacement to the target is known | |
535 | at this point. */ | |
536 | if (RELAX_BRANCH_CONST (fragp->fr_subtype)) | |
537 | { | |
538 | /* XXX this loses the 32-bit value if the constant was | |
539 | overflown! */ | |
540 | disp_to_target = fragp->fr_offset; | |
541 | disp_is_known = 1; | |
542 | } | |
543 | else if (fragp->fr_symbol != NULL | |
544 | && S_IS_DEFINED (fragp->fr_symbol) | |
545 | && !S_IS_WEAK (fragp->fr_symbol) | |
546 | && sec == S_GET_SEGMENT (fragp->fr_symbol)) | |
547 | { | |
548 | offsetT val = S_GET_VALUE (fragp->fr_symbol) + fragp->fr_offset; | |
549 | /* Convert to 64-bit blocks minus one. */ | |
550 | disp_to_target = (val - 8) / 8; | |
551 | disp_is_known = 1; | |
552 | } | |
553 | ||
554 | /* Now relax particular jump instructions. */ | |
555 | if (code == BPF_CODE_JA) | |
556 | { | |
557 | /* Unconditional jump. | |
558 | JA d16 -> JAL d32 */ | |
559 | ||
560 | gas_assert (RELAX_BRANCH_UNCOND (fragp->fr_subtype)); | |
561 | ||
562 | if (disp_is_known) | |
563 | { | |
564 | if (disp_to_target >= -32768 && disp_to_target <= 32767) | |
565 | { | |
566 | /* 16-bit disp is known and in range. Install a fixup | |
567 | for the disp16 if the branch value is not constant. | |
568 | This will be resolved by the assembler and units | |
569 | converted. */ | |
570 | ||
571 | if (!RELAX_BRANCH_CONST (fragp->fr_subtype)) | |
572 | { | |
573 | /* Install fixup for the JA. */ | |
574 | reloc_howto_type *reloc_howto | |
575 | = bfd_reloc_type_lookup (stdoutput, BFD_RELOC_BPF_DISP16); | |
576 | if (!reloc_howto) | |
577 | abort(); | |
578 | ||
579 | fixp = fix_new_exp (fragp, buf - (bfd_byte *) fragp->fr_literal, | |
580 | bfd_get_reloc_size (reloc_howto), | |
581 | &exp, | |
582 | reloc_howto->pc_relative, | |
583 | BFD_RELOC_BPF_DISP16); | |
584 | fixp->fx_file = fragp->fr_file; | |
585 | fixp->fx_line = fragp->fr_line; | |
586 | } | |
587 | } | |
588 | else | |
589 | { | |
590 | /* 16-bit disp is known and not in range. Turn the JA | |
591 | into a JAL with a 32-bit displacement. */ | |
592 | char bytes[8]; | |
593 | ||
594 | bytes[0] = ((BPF_CLASS_JMP32|BPF_CODE_JA|BPF_SRC_K) >> 56) & 0xff; | |
595 | bytes[1] = (word >> 48) & 0xff; | |
596 | bytes[2] = 0; /* disp16 high */ | |
597 | bytes[3] = 0; /* disp16 lo */ | |
598 | encode_int32 ((int32_t) disp_to_target, bytes + 4); | |
599 | ||
600 | write_insn_bytes (buf, bytes); | |
601 | } | |
602 | } | |
603 | else | |
604 | { | |
605 | /* The displacement to the target is not known. Do not | |
606 | relax. The linker will maybe do it if it chooses to. */ | |
607 | ||
608 | reloc_howto_type *reloc_howto = NULL; | |
609 | ||
610 | gas_assert (!RELAX_BRANCH_CONST (fragp->fr_subtype)); | |
611 | ||
612 | /* Install fixup for the JA. */ | |
613 | reloc_howto = bfd_reloc_type_lookup (stdoutput, BFD_RELOC_BPF_DISP16); | |
614 | if (!reloc_howto) | |
615 | abort (); | |
616 | ||
617 | fixp = fix_new_exp (fragp, buf - (bfd_byte *) fragp->fr_literal, | |
618 | bfd_get_reloc_size (reloc_howto), | |
619 | &exp, | |
620 | reloc_howto->pc_relative, | |
621 | BFD_RELOC_BPF_DISP16); | |
622 | fixp->fx_file = fragp->fr_file; | |
623 | fixp->fx_line = fragp->fr_line; | |
624 | } | |
625 | ||
626 | buf += 8; | |
627 | } | |
628 | else | |
629 | { | |
630 | /* Conditional jump. | |
631 | JXX d16 -> JXX +1; JA +1; JAL d32 */ | |
632 | ||
633 | gas_assert (!RELAX_BRANCH_UNCOND (fragp->fr_subtype)); | |
634 | ||
635 | if (disp_is_known) | |
636 | { | |
637 | if (disp_to_target >= -32768 && disp_to_target <= 32767) | |
638 | { | |
639 | /* 16-bit disp is known and in range. Install a fixup | |
640 | for the disp16 if the branch value is not constant. | |
641 | This will be resolved by the assembler and units | |
642 | converted. */ | |
643 | ||
644 | if (!RELAX_BRANCH_CONST (fragp->fr_subtype)) | |
645 | { | |
646 | /* Install fixup for the branch. */ | |
647 | reloc_howto_type *reloc_howto | |
648 | = bfd_reloc_type_lookup (stdoutput, BFD_RELOC_BPF_DISP16); | |
649 | if (!reloc_howto) | |
650 | abort(); | |
651 | ||
652 | fixp = fix_new_exp (fragp, buf - (bfd_byte *) fragp->fr_literal, | |
653 | bfd_get_reloc_size (reloc_howto), | |
654 | &exp, | |
655 | reloc_howto->pc_relative, | |
656 | BFD_RELOC_BPF_DISP16); | |
657 | fixp->fx_file = fragp->fr_file; | |
658 | fixp->fx_line = fragp->fr_line; | |
659 | } | |
660 | ||
661 | buf += 8; | |
662 | } | |
663 | else | |
664 | { | |
665 | /* 16-bit disp is known and not in range. Turn the JXX | |
666 | into a sequence JXX +1; JA +1; JAL d32. */ | |
667 | ||
668 | char bytes[8]; | |
669 | ||
670 | /* First, set the 16-bit offset in the current | |
671 | instruction to 1. */ | |
672 | ||
673 | if (target_big_endian) | |
674 | bfd_putb16 (1, buf + 2); | |
675 | else | |
676 | bfd_putl16 (1, buf + 2); | |
677 | buf += 8; | |
678 | ||
679 | /* Then, write the JA + 1 */ | |
680 | ||
681 | bytes[0] = 0x05; /* JA */ | |
682 | bytes[1] = 0x0; | |
683 | encode_int16 (1, bytes + 2); | |
684 | bytes[4] = 0x0; | |
685 | bytes[5] = 0x0; | |
686 | bytes[6] = 0x0; | |
687 | bytes[7] = 0x0; | |
688 | write_insn_bytes (buf, bytes); | |
689 | buf += 8; | |
690 | ||
691 | /* Finally, write the JAL to the target. */ | |
692 | ||
693 | bytes[0] = ((BPF_CLASS_JMP32|BPF_CODE_JA|BPF_SRC_K) >> 56) & 0xff; | |
694 | bytes[1] = 0; | |
695 | bytes[2] = 0; | |
696 | bytes[3] = 0; | |
697 | encode_int32 ((int32_t) disp_to_target, bytes + 4); | |
698 | write_insn_bytes (buf, bytes); | |
699 | buf += 8; | |
700 | } | |
701 | } | |
702 | else | |
703 | { | |
704 | /* The displacement to the target is not known. Do not | |
705 | relax. The linker will maybe do it if it chooses to. */ | |
706 | ||
707 | reloc_howto_type *reloc_howto = NULL; | |
708 | ||
709 | gas_assert (!RELAX_BRANCH_CONST (fragp->fr_subtype)); | |
710 | ||
711 | /* Install fixup for the conditional jump. */ | |
712 | reloc_howto = bfd_reloc_type_lookup (stdoutput, BFD_RELOC_BPF_DISP16); | |
713 | if (!reloc_howto) | |
714 | abort (); | |
715 | ||
716 | fixp = fix_new_exp (fragp, buf - (bfd_byte *) fragp->fr_literal, | |
717 | bfd_get_reloc_size (reloc_howto), | |
718 | &exp, | |
719 | reloc_howto->pc_relative, | |
720 | BFD_RELOC_BPF_DISP16); | |
721 | fixp->fx_file = fragp->fr_file; | |
722 | fixp->fx_line = fragp->fr_line; | |
723 | buf += 8; | |
724 | } | |
725 | } | |
726 | ||
727 | gas_assert (buf == (bfd_byte *)fragp->fr_literal | |
728 | + fragp->fr_fix + fragp->fr_var); | |
729 | ||
730 | fragp->fr_fix += fragp->fr_var; | |
f8861f5d JM |
731 | } |
732 | ||
733 | \f | |
d218e7fe JM |
734 | /* Apply a fixS (fixup of an instruction or data that we didn't have |
735 | enough info to complete immediately) to the data in a frag. */ | |
736 | ||
f8861f5d | 737 | void |
d218e7fe | 738 | md_apply_fix (fixS *fixP, valueT *valP, segT seg ATTRIBUTE_UNUSED) |
f8861f5d | 739 | { |
d218e7fe JM |
740 | char *where = fixP->fx_frag->fr_literal + fixP->fx_where; |
741 | ||
742 | switch (fixP->fx_r_type) | |
f8861f5d | 743 | { |
d218e7fe JM |
744 | case BFD_RELOC_BPF_DISP16: |
745 | /* Convert from bytes to number of 64-bit words to the target, | |
746 | minus one. */ | |
747 | *valP = (((long) (*valP)) - 8) / 8; | |
748 | break; | |
c2ca88d7 | 749 | case BFD_RELOC_BPF_DISPCALL32: |
d218e7fe | 750 | case BFD_RELOC_BPF_DISP32: |
d218e7fe JM |
751 | /* Convert from bytes to number of 64-bit words to the target, |
752 | minus one. */ | |
753 | *valP = (((long) (*valP)) - 8) / 8; | |
c2ca88d7 JM |
754 | |
755 | if (fixP->fx_r_type == BFD_RELOC_BPF_DISPCALL32) | |
756 | { | |
757 | /* eBPF supports two kind of CALL instructions: the so | |
758 | called pseudo calls ("bpf to bpf") and external calls | |
759 | ("bpf to kernel"). | |
760 | ||
761 | Both kind of calls use the same instruction (CALL). | |
762 | However, external calls are constructed by passing a | |
763 | constant argument to the instruction, whereas pseudo | |
764 | calls result from expressions involving symbols. In | |
765 | practice, instructions requiring a fixup are interpreted | |
766 | as pseudo-calls. If we are executing this code, this is | |
767 | a pseudo call. | |
768 | ||
769 | The kernel expects for pseudo-calls to be annotated by | |
770 | having BPF_PSEUDO_CALL in the SRC field of the | |
771 | instruction. But beware the infamous nibble-swapping of | |
772 | eBPF and take endianness into account here. | |
773 | ||
774 | Note that the CALL instruction has only one operand, so | |
775 | this code is executed only once per instruction. */ | |
776 | md_number_to_chars (where + 1, target_big_endian ? 0x01 : 0x10, 1); | |
777 | } | |
d218e7fe JM |
778 | break; |
779 | case BFD_RELOC_16_PCREL: | |
780 | /* Convert from bytes to number of 64-bit words to the target, | |
781 | minus one. */ | |
782 | *valP = (((long) (*valP)) - 8) / 8; | |
783 | break; | |
784 | default: | |
785 | break; | |
786 | } | |
f8861f5d | 787 | |
d218e7fe JM |
788 | if (fixP->fx_addsy == (symbolS *) NULL) |
789 | fixP->fx_done = 1; | |
790 | ||
791 | if (fixP->fx_done) | |
792 | { | |
793 | /* We're finished with this fixup. Install it because | |
794 | bfd_install_relocation won't be called to do it. */ | |
795 | switch (fixP->fx_r_type) | |
796 | { | |
797 | case BFD_RELOC_8: | |
798 | md_number_to_chars (where, *valP, 1); | |
799 | break; | |
800 | case BFD_RELOC_16: | |
801 | md_number_to_chars (where, *valP, 2); | |
802 | break; | |
803 | case BFD_RELOC_32: | |
804 | md_number_to_chars (where, *valP, 4); | |
805 | break; | |
806 | case BFD_RELOC_64: | |
807 | md_number_to_chars (where, *valP, 8); | |
808 | break; | |
809 | case BFD_RELOC_BPF_DISP16: | |
810 | md_number_to_chars (where + 2, (uint16_t) *valP, 2); | |
f8861f5d | 811 | break; |
d218e7fe | 812 | case BFD_RELOC_BPF_DISP32: |
c2ca88d7 | 813 | case BFD_RELOC_BPF_DISPCALL32: |
d218e7fe | 814 | md_number_to_chars (where + 4, (uint32_t) *valP, 4); |
f8861f5d | 815 | break; |
d218e7fe JM |
816 | case BFD_RELOC_16_PCREL: |
817 | md_number_to_chars (where + 2, (uint32_t) *valP, 2); | |
818 | break; | |
819 | default: | |
820 | as_bad_where (fixP->fx_file, fixP->fx_line, | |
821 | _("internal error: can't install fix for reloc type %d (`%s')"), | |
822 | fixP->fx_r_type, bfd_get_reloc_code_name (fixP->fx_r_type)); | |
823 | break; | |
824 | } | |
f8861f5d JM |
825 | } |
826 | ||
d218e7fe JM |
827 | /* Tuck `value' away for use by tc_gen_reloc. |
828 | See the comment describing fx_addnumber in write.h. | |
829 | This field is misnamed (or misused :-). */ | |
830 | fixP->fx_addnumber = *valP; | |
f8861f5d JM |
831 | } |
832 | ||
249d4715 JM |
833 | \f |
834 | /* Instruction writing routines. */ | |
835 | ||
836 | /* Encode a BPF instruction in the given buffer BYTES. Non-constant | |
837 | immediates are encoded as zeroes. */ | |
838 | ||
839 | static void | |
840 | encode_insn (struct bpf_insn *insn, char *bytes) | |
841 | { | |
842 | uint8_t src, dst; | |
843 | ||
844 | /* Zero all the bytes. */ | |
845 | memset (bytes, 0, 16); | |
846 | ||
847 | /* First encode the opcodes. Note that we have to handle the | |
848 | endianness groups of the BPF instructions: 8 | 4 | 4 | 16 | | |
849 | 32. */ | |
850 | if (target_big_endian) | |
851 | { | |
852 | /* code */ | |
853 | bytes[0] = (insn->opcode >> 56) & 0xff; | |
854 | /* regs */ | |
855 | bytes[1] = (insn->opcode >> 48) & 0xff; | |
856 | /* offset16 */ | |
857 | bytes[2] = (insn->opcode >> 40) & 0xff; | |
858 | bytes[3] = (insn->opcode >> 32) & 0xff; | |
859 | /* imm32 */ | |
860 | bytes[4] = (insn->opcode >> 24) & 0xff; | |
861 | bytes[5] = (insn->opcode >> 16) & 0xff; | |
862 | bytes[6] = (insn->opcode >> 8) & 0xff; | |
863 | bytes[7] = insn->opcode & 0xff; | |
864 | } | |
865 | else | |
866 | { | |
867 | /* code */ | |
868 | bytes[0] = (insn->opcode >> 56) & 0xff; | |
869 | /* regs */ | |
870 | bytes[1] = (((((insn->opcode >> 48) & 0xff) & 0xf) << 4) | |
871 | | (((insn->opcode >> 48) & 0xff) & 0xf)); | |
872 | /* offset16 */ | |
873 | bytes[3] = (insn->opcode >> 40) & 0xff; | |
874 | bytes[2] = (insn->opcode >> 32) & 0xff; | |
875 | /* imm32 */ | |
876 | bytes[7] = (insn->opcode >> 24) & 0xff; | |
877 | bytes[6] = (insn->opcode >> 16) & 0xff; | |
878 | bytes[5] = (insn->opcode >> 8) & 0xff; | |
879 | bytes[4] = insn->opcode & 0xff; | |
880 | } | |
881 | ||
882 | /* Now the registers. */ | |
883 | src = insn->has_src ? insn->src : 0; | |
884 | dst = insn->has_dst ? insn->dst : 0; | |
885 | ||
886 | if (target_big_endian) | |
887 | bytes[1] = ((dst & 0xf) << 4) | (src & 0xf); | |
888 | else | |
889 | bytes[1] = ((src & 0xf) << 4) | (dst & 0xf); | |
890 | ||
891 | /* Now the immediates that are known to be constant. */ | |
892 | ||
893 | if (insn->has_imm32 && insn->imm32.X_op == O_constant) | |
894 | encode_int32 (insn->imm32.X_add_number, bytes + 4); | |
895 | ||
896 | if (insn->has_disp32 && insn->disp32.X_op == O_constant) | |
897 | encode_int32 (insn->disp32.X_add_number, bytes + 4); | |
898 | ||
899 | if (insn->has_offset16 && insn->offset16.X_op == O_constant) | |
900 | encode_int16 (insn->offset16.X_add_number, bytes + 2); | |
901 | ||
902 | if (insn->has_disp16 && insn->disp16.X_op == O_constant) | |
903 | encode_int16 (insn->disp16.X_add_number, bytes + 2); | |
904 | ||
905 | if (insn->has_imm64 && insn->imm64.X_op == O_constant) | |
906 | { | |
907 | uint64_t imm64 = insn->imm64.X_add_number; | |
908 | ||
909 | if (target_big_endian) | |
910 | { | |
911 | bytes[12] = (imm64 >> 56) & 0xff; | |
912 | bytes[13] = (imm64 >> 48) & 0xff; | |
913 | bytes[14] = (imm64 >> 40) & 0xff; | |
914 | bytes[15] = (imm64 >> 32) & 0xff; | |
915 | bytes[4] = (imm64 >> 24) & 0xff; | |
916 | bytes[5] = (imm64 >> 16) & 0xff; | |
917 | bytes[6] = (imm64 >> 8) & 0xff; | |
918 | bytes[7] = imm64 & 0xff; | |
919 | } | |
920 | else | |
921 | { | |
922 | bytes[15] = (imm64 >> 56) & 0xff; | |
923 | bytes[14] = (imm64 >> 48) & 0xff; | |
924 | bytes[13] = (imm64 >> 40) & 0xff; | |
925 | bytes[12] = (imm64 >> 32) & 0xff; | |
926 | bytes[7] = (imm64 >> 24) & 0xff; | |
927 | bytes[6] = (imm64 >> 16) & 0xff; | |
928 | bytes[5] = (imm64 >> 8) & 0xff; | |
929 | bytes[4] = imm64 & 0xff; | |
930 | } | |
931 | } | |
932 | } | |
933 | ||
934 | /* Install the fixups in INSN in their proper location in the | |
935 | specified FRAG at the location pointed by WHERE. */ | |
936 | ||
937 | static void | |
938 | install_insn_fixups (struct bpf_insn *insn, fragS *frag, long where) | |
939 | { | |
940 | if (insn->has_imm64) | |
941 | { | |
942 | switch (insn->imm64.X_op) | |
943 | { | |
944 | case O_symbol: | |
945 | case O_subtract: | |
946 | case O_add: | |
947 | { | |
948 | reloc_howto_type *reloc_howto; | |
949 | int size; | |
950 | ||
951 | reloc_howto = bfd_reloc_type_lookup (stdoutput, BFD_RELOC_BPF_64); | |
952 | if (!reloc_howto) | |
953 | abort (); | |
954 | ||
955 | size = bfd_get_reloc_size (reloc_howto); | |
956 | ||
957 | fix_new_exp (frag, where, | |
958 | size, &insn->imm64, reloc_howto->pc_relative, | |
959 | BFD_RELOC_BPF_64); | |
960 | break; | |
961 | } | |
962 | case O_constant: | |
963 | /* Already handled in encode_insn. */ | |
964 | break; | |
965 | default: | |
966 | abort (); | |
967 | } | |
968 | } | |
969 | ||
970 | if (insn->has_imm32) | |
971 | { | |
972 | switch (insn->imm32.X_op) | |
973 | { | |
974 | case O_symbol: | |
975 | case O_subtract: | |
976 | case O_add: | |
977 | case O_uminus: | |
978 | { | |
979 | reloc_howto_type *reloc_howto; | |
980 | int size; | |
981 | ||
982 | reloc_howto = bfd_reloc_type_lookup (stdoutput, BFD_RELOC_32); | |
983 | if (!reloc_howto) | |
984 | abort (); | |
985 | ||
986 | size = bfd_get_reloc_size (reloc_howto); | |
987 | ||
988 | fix_new_exp (frag, where + 4, | |
989 | size, &insn->imm32, reloc_howto->pc_relative, | |
990 | BFD_RELOC_32); | |
991 | break; | |
992 | } | |
993 | case O_constant: | |
994 | /* Already handled in encode_insn. */ | |
995 | break; | |
996 | default: | |
997 | abort (); | |
998 | } | |
999 | } | |
1000 | ||
1001 | if (insn->has_disp32) | |
1002 | { | |
1003 | switch (insn->disp32.X_op) | |
1004 | { | |
1005 | case O_symbol: | |
1006 | case O_subtract: | |
1007 | case O_add: | |
1008 | { | |
1009 | reloc_howto_type *reloc_howto; | |
1010 | int size; | |
1011 | unsigned int bfd_reloc | |
1012 | = (insn->id == BPF_INSN_CALL | |
1013 | ? BFD_RELOC_BPF_DISPCALL32 | |
1014 | : BFD_RELOC_BPF_DISP32); | |
1015 | ||
1016 | reloc_howto = bfd_reloc_type_lookup (stdoutput, bfd_reloc); | |
1017 | if (!reloc_howto) | |
1018 | abort (); | |
1019 | ||
1020 | size = bfd_get_reloc_size (reloc_howto); | |
1021 | ||
1022 | fix_new_exp (frag, where, | |
1023 | size, &insn->disp32, reloc_howto->pc_relative, | |
1024 | bfd_reloc); | |
1025 | break; | |
1026 | } | |
1027 | case O_constant: | |
1028 | /* Already handled in encode_insn. */ | |
1029 | break; | |
1030 | default: | |
1031 | abort (); | |
1032 | } | |
1033 | } | |
1034 | ||
1035 | if (insn->has_offset16) | |
1036 | { | |
1037 | switch (insn->offset16.X_op) | |
1038 | { | |
1039 | case O_symbol: | |
1040 | case O_subtract: | |
1041 | case O_add: | |
1042 | { | |
1043 | reloc_howto_type *reloc_howto; | |
1044 | int size; | |
1045 | ||
1046 | /* XXX we really need a new pc-rel offset in bytes | |
1047 | relocation for this. */ | |
1048 | reloc_howto = bfd_reloc_type_lookup (stdoutput, BFD_RELOC_BPF_DISP16); | |
1049 | if (!reloc_howto) | |
1050 | abort (); | |
1051 | ||
1052 | size = bfd_get_reloc_size (reloc_howto); | |
1053 | ||
1054 | fix_new_exp (frag, where, | |
1055 | size, &insn->offset16, reloc_howto->pc_relative, | |
1056 | BFD_RELOC_BPF_DISP16); | |
1057 | break; | |
1058 | } | |
1059 | case O_constant: | |
1060 | /* Already handled in encode_insn. */ | |
1061 | break; | |
1062 | default: | |
1063 | abort (); | |
1064 | } | |
1065 | } | |
1066 | ||
1067 | if (insn->has_disp16) | |
1068 | { | |
1069 | switch (insn->disp16.X_op) | |
1070 | { | |
1071 | case O_symbol: | |
1072 | case O_subtract: | |
1073 | case O_add: | |
1074 | { | |
1075 | reloc_howto_type *reloc_howto; | |
1076 | int size; | |
1077 | ||
1078 | reloc_howto = bfd_reloc_type_lookup (stdoutput, BFD_RELOC_BPF_DISP16); | |
1079 | if (!reloc_howto) | |
1080 | abort (); | |
1081 | ||
1082 | size = bfd_get_reloc_size (reloc_howto); | |
1083 | ||
1084 | fix_new_exp (frag, where, | |
1085 | size, &insn->disp16, reloc_howto->pc_relative, | |
1086 | BFD_RELOC_BPF_DISP16); | |
1087 | break; | |
1088 | } | |
1089 | case O_constant: | |
1090 | /* Already handled in encode_insn. */ | |
1091 | break; | |
1092 | default: | |
1093 | abort (); | |
1094 | } | |
1095 | } | |
1096 | ||
1097 | } | |
1098 | ||
1099 | /* Add a new insn to the list of instructions. */ | |
1100 | ||
1101 | static void | |
1102 | add_fixed_insn (struct bpf_insn *insn) | |
1103 | { | |
1104 | char *this_frag = frag_more (insn->size); | |
1105 | char bytes[16]; | |
1106 | int i; | |
1107 | ||
1108 | /* First encode the known parts of the instruction, including | |
1109 | opcodes and constant immediates, and write them to the frag. */ | |
1110 | encode_insn (insn, bytes); | |
1111 | for (i = 0; i < insn->size; ++i) | |
1112 | md_number_to_chars (this_frag + i, (valueT) bytes[i], 1); | |
1113 | ||
1114 | /* Now install the instruction fixups. */ | |
1115 | install_insn_fixups (insn, frag_now, | |
1116 | this_frag - frag_now->fr_literal); | |
1117 | } | |
1118 | ||
1119 | /* Add a new relaxable to the list of instructions. */ | |
1120 | ||
1121 | static void | |
1122 | add_relaxed_insn (struct bpf_insn *insn, expressionS *exp) | |
1123 | { | |
1124 | char bytes[16]; | |
1125 | int i; | |
1126 | char *this_frag; | |
1127 | unsigned worst_case = relaxed_branch_length (NULL, NULL, 0); | |
1128 | unsigned best_case = insn->size; | |
1129 | ||
1130 | /* We only support relaxing branches, for the moment. */ | |
1131 | relax_substateT subtype | |
1132 | = RELAX_BRANCH_ENCODE (insn->id == BPF_INSN_JAR, | |
1133 | exp->X_op == O_constant, | |
1134 | worst_case); | |
1135 | ||
1136 | frag_grow (worst_case); | |
1137 | this_frag = frag_more (0); | |
1138 | ||
1139 | /* First encode the known parts of the instruction, including | |
1140 | opcodes and constant immediates, and write them to the frag. */ | |
1141 | encode_insn (insn, bytes); | |
1142 | for (i = 0; i < insn->size; ++i) | |
1143 | md_number_to_chars (this_frag + i, (valueT) bytes[i], 1); | |
1144 | ||
1145 | /* Note that instruction fixups will be applied once the frag is | |
1146 | relaxed, in md_convert_frag. */ | |
1147 | frag_var (rs_machine_dependent, | |
1148 | worst_case, best_case, | |
1149 | subtype, exp->X_add_symbol, exp->X_add_number /* offset */, | |
1150 | NULL); | |
1151 | } | |
1152 | ||
1153 | \f | |
d218e7fe JM |
1154 | /* Parse an operand expression. Returns the first character that is |
1155 | not part of the expression, or NULL in case of parse error. | |
ff5a51b3 | 1156 | |
d218e7fe | 1157 | See md_operand below to see how exp_parse_failed is used. */ |
ff5a51b3 | 1158 | |
d218e7fe | 1159 | static int exp_parse_failed = 0; |
ff5a51b3 GM |
1160 | |
1161 | static char * | |
d218e7fe | 1162 | parse_expression (char *s, expressionS *exp) |
ff5a51b3 | 1163 | { |
d218e7fe JM |
1164 | char *saved_input_line_pointer = input_line_pointer; |
1165 | char *saved_s = s; | |
ff5a51b3 | 1166 | |
d218e7fe JM |
1167 | exp_parse_failed = 0; |
1168 | input_line_pointer = s; | |
1169 | expression (exp); | |
1170 | s = input_line_pointer; | |
1171 | input_line_pointer = saved_input_line_pointer; | |
ff5a51b3 | 1172 | |
d218e7fe JM |
1173 | switch (exp->X_op == O_absent || exp_parse_failed) |
1174 | return NULL; | |
ff5a51b3 | 1175 | |
d218e7fe JM |
1176 | /* The expression parser may consume trailing whitespaces. We have |
1177 | to undo that since the instruction templates may be expecting | |
1178 | these whitespaces. */ | |
1179 | { | |
1180 | char *p; | |
1181 | for (p = s - 1; p >= saved_s && *p == ' '; --p) | |
1182 | --s; | |
1183 | } | |
ff5a51b3 | 1184 | |
d218e7fe | 1185 | return s; |
ff5a51b3 GM |
1186 | } |
1187 | ||
d218e7fe JM |
1188 | /* Parse a BPF register name and return the corresponding register |
1189 | number. Return NULL in case of parse error, or a pointer to the | |
1190 | first character in S that is not part of the register name. */ | |
ff5a51b3 | 1191 | |
d218e7fe JM |
1192 | static char * |
1193 | parse_bpf_register (char *s, char rw, uint8_t *regno) | |
ff5a51b3 | 1194 | { |
d218e7fe | 1195 | if (asm_dialect == DIALECT_NORMAL) |
ff5a51b3 | 1196 | { |
d218e7fe JM |
1197 | rw = 'r'; |
1198 | if (*s != '%') | |
1199 | return NULL; | |
1200 | s += 1; | |
ff5a51b3 | 1201 | |
d218e7fe JM |
1202 | if (*s == 'f' && *(s + 1) == 'p') |
1203 | { | |
1204 | *regno = 10; | |
1205 | s += 2; | |
1206 | return s; | |
1207 | } | |
1208 | } | |
ff5a51b3 | 1209 | |
d218e7fe JM |
1210 | if (*s != rw) |
1211 | return NULL; | |
1212 | s += 1; | |
ff5a51b3 | 1213 | |
d218e7fe JM |
1214 | if (*s == '1') |
1215 | { | |
1216 | if (*(s + 1) == '0') | |
1217 | { | |
1218 | *regno = 10; | |
1219 | s += 2; | |
1220 | } | |
1221 | else | |
1222 | { | |
1223 | *regno = 1; | |
1224 | s += 1; | |
1225 | } | |
1226 | } | |
1227 | else if (*s >= '0' && *s <= '9') | |
1228 | { | |
1229 | *regno = *s - '0'; | |
1230 | s += 1; | |
1231 | } | |
ff5a51b3 | 1232 | |
d218e7fe | 1233 | return s; |
ff5a51b3 GM |
1234 | } |
1235 | ||
d218e7fe | 1236 | /* Collect a parse error message. */ |
ff5a51b3 | 1237 | |
d218e7fe JM |
1238 | static int partial_match_length = 0; |
1239 | static char *errmsg = NULL; | |
ff5a51b3 | 1240 | |
d218e7fe JM |
1241 | static void |
1242 | parse_error (int length, const char *fmt, ...) | |
ff5a51b3 | 1243 | { |
d218e7fe | 1244 | if (length > partial_match_length) |
ff5a51b3 | 1245 | { |
d218e7fe | 1246 | va_list args; |
ff5a51b3 | 1247 | |
d218e7fe JM |
1248 | free (errmsg); |
1249 | va_start (args, fmt); | |
1250 | errmsg = xvasprintf (fmt, args); | |
1251 | va_end (args); | |
1252 | partial_match_length = length; | |
ff5a51b3 | 1253 | } |
ff5a51b3 GM |
1254 | } |
1255 | ||
d218e7fe JM |
1256 | /* Assemble a machine instruction in STR and emit the frags/bytes it |
1257 | assembles to. */ | |
ff5a51b3 | 1258 | |
d218e7fe JM |
1259 | void |
1260 | md_assemble (char *str ATTRIBUTE_UNUSED) | |
ff5a51b3 | 1261 | { |
d218e7fe JM |
1262 | /* There are two different syntaxes that can be used to write BPF |
1263 | instructions. One is very conventional and like any other | |
1264 | assembly language where each instruction is conformed by an | |
1265 | instruction mnemonic followed by its operands. This is what we | |
1266 | call the "normal" syntax. The other syntax tries to look like C | |
1267 | statements. We have to support both syntaxes in this assembler. | |
1268 | ||
1269 | One of the many nuisances introduced by this eccentricity is that | |
1270 | in the pseudo-c syntax it is not possible to hash the opcodes | |
1271 | table by instruction mnemonic, because there is none. So we have | |
1272 | no other choice than to try to parse all instruction opcodes | |
1273 | until one matches. This is slow. | |
1274 | ||
1275 | Another problem is that emitting detailed diagnostics becomes | |
1276 | tricky, since the lack of mnemonic means it is not clear what | |
1277 | instruction was intended by the user, and we cannot emit | |
1278 | diagnostics for every attempted template. So if an instruction | |
1279 | is not parsed, we report the diagnostic corresponding to the | |
1280 | partially parsed instruction that was matched further. */ | |
1281 | ||
1282 | unsigned int idx = 0; | |
1283 | struct bpf_insn insn; | |
1284 | const struct bpf_opcode *opcode; | |
1285 | ||
1286 | /* Initialize the global diagnostic variables. See the parse_error | |
1287 | function above. */ | |
1288 | partial_match_length = 0; | |
1289 | errmsg = NULL; | |
1290 | ||
1291 | #define PARSE_ERROR(...) parse_error (s - str, __VA_ARGS__) | |
1292 | ||
1293 | while ((opcode = bpf_get_opcode (idx++)) != NULL) | |
ff5a51b3 | 1294 | { |
d218e7fe JM |
1295 | const char *p; |
1296 | char *s; | |
1297 | const char *template | |
1298 | = (asm_dialect == DIALECT_PSEUDOC ? opcode->pseudoc : opcode->normal); | |
1299 | ||
1300 | /* Do not try to match opcodes with a higher version than the | |
1301 | selected ISA spec. */ | |
1302 | if (opcode->version > isa_spec) | |
1303 | continue; | |
1304 | ||
1305 | memset (&insn, 0, sizeof (struct bpf_insn)); | |
1306 | insn.size = 8; | |
1307 | for (s = str, p = template; *p != '\0';) | |
1308 | { | |
1309 | if (*p == ' ') | |
1310 | { | |
1311 | /* Expect zero or more spaces. */ | |
1312 | while (*s != '\0' && (*s == ' ' || *s == '\t')) | |
1313 | s += 1; | |
1314 | p += 1; | |
1315 | } | |
1316 | else if (*p == '%') | |
1317 | { | |
1318 | if (*(p + 1) == '%') | |
1319 | { | |
1320 | if (*s != '%') | |
1321 | { | |
1322 | PARSE_ERROR ("expected '%%'"); | |
1323 | break; | |
1324 | } | |
1325 | p += 2; | |
1326 | s += 1; | |
1327 | } | |
1328 | else if (*(p + 1) == 'w') | |
1329 | { | |
1330 | /* Expect zero or more spaces. */ | |
1331 | while (*s != '\0' && (*s == ' ' || *s == '\t')) | |
1332 | s += 1; | |
1333 | p += 2; | |
1334 | } | |
1335 | else if (*(p + 1) == 'W') | |
1336 | { | |
1337 | /* Expect one or more spaces. */ | |
1338 | if (*s != ' ' && *s != '\t') | |
1339 | { | |
1340 | PARSE_ERROR ("expected white space, got '%s'", | |
1341 | s); | |
1342 | break; | |
1343 | } | |
1344 | while (*s != '\0' && (*s == ' ' || *s == '\t')) | |
1345 | s += 1; | |
1346 | p += 2; | |
1347 | } | |
1348 | else if (strncmp (p, "%dr", 3) == 0) | |
1349 | { | |
1350 | uint8_t regno; | |
1351 | char *news = parse_bpf_register (s, 'r', ®no); | |
1352 | ||
1353 | if (news == NULL || (insn.has_dst && regno != insn.dst)) | |
1354 | { | |
1355 | if (news != NULL) | |
1356 | PARSE_ERROR ("expected register r%d, got r%d", | |
1357 | insn.dst, regno); | |
1358 | else | |
1359 | PARSE_ERROR ("expected register name, got '%s'", s); | |
1360 | break; | |
1361 | } | |
1362 | s = news; | |
1363 | insn.dst = regno; | |
1364 | insn.has_dst = 1; | |
1365 | p += 3; | |
1366 | } | |
1367 | else if (strncmp (p, "%sr", 3) == 0) | |
1368 | { | |
1369 | uint8_t regno; | |
1370 | char *news = parse_bpf_register (s, 'r', ®no); | |
1371 | ||
1372 | if (news == NULL || (insn.has_src && regno != insn.src)) | |
1373 | { | |
1374 | if (news != NULL) | |
1375 | PARSE_ERROR ("expected register r%d, got r%d", | |
1376 | insn.dst, regno); | |
1377 | else | |
1378 | PARSE_ERROR ("expected register name, got '%s'", s); | |
1379 | break; | |
1380 | } | |
1381 | s = news; | |
1382 | insn.src = regno; | |
1383 | insn.has_src = 1; | |
1384 | p += 3; | |
1385 | } | |
1386 | else if (strncmp (p, "%dw", 3) == 0) | |
1387 | { | |
1388 | uint8_t regno; | |
1389 | char *news = parse_bpf_register (s, 'w', ®no); | |
1390 | ||
1391 | if (news == NULL || (insn.has_dst && regno != insn.dst)) | |
1392 | { | |
1393 | if (news != NULL) | |
1394 | PARSE_ERROR ("expected register r%d, got r%d", | |
1395 | insn.dst, regno); | |
1396 | else | |
1397 | PARSE_ERROR ("expected register name, got '%s'", s); | |
1398 | break; | |
1399 | } | |
1400 | s = news; | |
1401 | insn.dst = regno; | |
1402 | insn.has_dst = 1; | |
1403 | p += 3; | |
1404 | } | |
1405 | else if (strncmp (p, "%sw", 3) == 0) | |
1406 | { | |
1407 | uint8_t regno; | |
1408 | char *news = parse_bpf_register (s, 'w', ®no); | |
1409 | ||
1410 | if (news == NULL || (insn.has_src && regno != insn.src)) | |
1411 | { | |
1412 | if (news != NULL) | |
1413 | PARSE_ERROR ("expected register r%d, got r%d", | |
1414 | insn.dst, regno); | |
1415 | else | |
1416 | PARSE_ERROR ("expected register name, got '%s'", s); | |
1417 | break; | |
1418 | } | |
1419 | s = news; | |
1420 | insn.src = regno; | |
1421 | insn.has_src = 1; | |
1422 | p += 3; | |
1423 | } | |
1424 | else if (strncmp (p, "%i32", 4) == 0 | |
1425 | || strncmp (p, "%I32", 4) == 0) | |
1426 | { | |
1427 | if (p[1] == 'I') | |
1428 | { | |
1429 | while (*s == ' ' || *s == '\t') | |
1430 | s += 1; | |
1431 | if (*s != '+' && *s != '-') | |
1432 | { | |
1433 | PARSE_ERROR ("expected `+' or `-', got `%c'", *s); | |
1434 | break; | |
1435 | } | |
1436 | } | |
1437 | ||
1438 | s = parse_expression (s, &insn.imm32); | |
1439 | if (s == NULL) | |
1440 | { | |
1441 | PARSE_ERROR ("expected signed 32-bit immediate"); | |
1442 | break; | |
1443 | } | |
1444 | insn.has_imm32 = 1; | |
1445 | p += 4; | |
1446 | } | |
1447 | else if (strncmp (p, "%o16", 4) == 0) | |
1448 | { | |
1449 | while (*s == ' ' || *s == '\t') | |
1450 | s += 1; | |
1451 | if (*s != '+' && *s != '-') | |
1452 | { | |
1453 | PARSE_ERROR ("expected `+' or `-', got `%c'", *s); | |
1454 | break; | |
1455 | } | |
1456 | ||
1457 | s = parse_expression (s, &insn.offset16); | |
1458 | if (s == NULL) | |
1459 | { | |
1460 | PARSE_ERROR ("expected signed 16-bit offset"); | |
1461 | break; | |
1462 | } | |
1463 | insn.has_offset16 = 1; | |
1464 | p += 4; | |
1465 | } | |
1466 | else if (strncmp (p, "%d16", 4) == 0) | |
1467 | { | |
1468 | s = parse_expression (s, &insn.disp16); | |
1469 | if (s == NULL) | |
1470 | { | |
1471 | PARSE_ERROR ("expected signed 16-bit displacement"); | |
1472 | break; | |
1473 | } | |
1474 | insn.has_disp16 = 1; | |
249d4715 | 1475 | insn.is_relaxable = 1; |
d218e7fe JM |
1476 | p += 4; |
1477 | } | |
1478 | else if (strncmp (p, "%d32", 4) == 0) | |
1479 | { | |
1480 | s = parse_expression (s, &insn.disp32); | |
1481 | if (s == NULL) | |
1482 | { | |
1483 | PARSE_ERROR ("expected signed 32-bit displacement"); | |
1484 | break; | |
1485 | } | |
1486 | insn.has_disp32 = 1; | |
1487 | p += 4; | |
1488 | } | |
1489 | else if (strncmp (p, "%i64", 4) == 0) | |
1490 | { | |
1491 | s = parse_expression (s, &insn.imm64); | |
1492 | if (s == NULL) | |
1493 | { | |
1494 | PARSE_ERROR ("expected signed 64-bit immediate"); | |
1495 | break; | |
1496 | } | |
1497 | insn.has_imm64 = 1; | |
1498 | insn.size = 16; | |
1499 | p += 4; | |
1500 | } | |
1501 | else | |
1502 | as_fatal (_("invalid %%-tag in BPF opcode '%s'\n"), template); | |
1503 | } | |
1504 | else | |
1505 | { | |
1506 | /* Match a literal character. */ | |
1507 | if (*s != *p) | |
1508 | { | |
1509 | if (*s == '\0') | |
1510 | PARSE_ERROR ("expected '%c'", *p); | |
1511 | else if (*s == '%') | |
1512 | { | |
1513 | /* This is to workaround a bug in as_bad. */ | |
1514 | char tmp[3]; | |
1515 | ||
1516 | tmp[0] = '%'; | |
1517 | tmp[1] = '%'; | |
1518 | tmp[2] = '\0'; | |
1519 | ||
1520 | PARSE_ERROR ("expected '%c', got '%s'", *p, tmp); | |
1521 | } | |
1522 | else | |
1523 | PARSE_ERROR ("expected '%c', got '%c'", *p, *s); | |
1524 | break; | |
1525 | } | |
1526 | p += 1; | |
1527 | s += 1; | |
1528 | } | |
1529 | } | |
ff5a51b3 | 1530 | |
d218e7fe JM |
1531 | if (*p == '\0') |
1532 | { | |
1533 | /* Allow white spaces at the end of the line. */ | |
1534 | while (*s != '\0' && (*s == ' ' || *s == '\t')) | |
1535 | s += 1; | |
1536 | if (*s == '\0') | |
1537 | /* We parsed an instruction successfully. */ | |
1538 | break; | |
1539 | PARSE_ERROR ("extra junk at end of line"); | |
1540 | } | |
1541 | } | |
ff5a51b3 | 1542 | |
d218e7fe JM |
1543 | if (opcode == NULL) |
1544 | { | |
1545 | as_bad (_("unrecognized instruction `%s'"), str); | |
1546 | if (errmsg != NULL) | |
1547 | { | |
fafcbd14 | 1548 | as_bad ("%s", errmsg); |
d218e7fe JM |
1549 | free (errmsg); |
1550 | } | |
ff5a51b3 | 1551 | |
d218e7fe JM |
1552 | return; |
1553 | } | |
c2ca88d7 | 1554 | insn.id = opcode->id; |
d218e7fe | 1555 | insn.opcode = opcode->opcode; |
ff5a51b3 | 1556 | |
d218e7fe | 1557 | #undef PARSE_ERROR |
ff5a51b3 | 1558 | |
d218e7fe | 1559 | /* Generate the frags and fixups for the parsed instruction. */ |
249d4715 JM |
1560 | if (do_relax && insn.is_relaxable) |
1561 | { | |
1562 | expressionS *relaxable_exp = NULL; | |
d218e7fe | 1563 | |
249d4715 JM |
1564 | if (insn.has_disp16) |
1565 | relaxable_exp = &insn.disp16; | |
1566 | else | |
1567 | abort (); | |
d218e7fe | 1568 | |
249d4715 JM |
1569 | add_relaxed_insn (&insn, relaxable_exp); |
1570 | } | |
1571 | else | |
1572 | add_fixed_insn (&insn); | |
ff5a51b3 | 1573 | |
d218e7fe JM |
1574 | /* Emit DWARF2 debugging information. */ |
1575 | dwarf2_emit_insn (insn.size); | |
ff5a51b3 GM |
1576 | } |
1577 | ||
d218e7fe | 1578 | /* Parse an operand that is machine-specific. */ |
f8861f5d JM |
1579 | |
1580 | void | |
1581 | md_operand (expressionS *expressionP) | |
1582 | { | |
d218e7fe JM |
1583 | /* If this hook is invoked it means GAS failed to parse a generic |
1584 | expression. We should inhibit the as_bad in expr.c, so we can fail | |
1585 | while parsing instruction alternatives. To do that, we change the | |
1586 | expression to not have an O_absent. But then we also need to set | |
1587 | exp_parse_failed to parse_expression above does the right thing. */ | |
1588 | ++input_line_pointer; | |
1589 | expressionP->X_op = O_constant; | |
1590 | expressionP->X_add_number = 0; | |
1591 | exp_parse_failed = 1; | |
f8861f5d JM |
1592 | } |
1593 | ||
f8861f5d JM |
1594 | symbolS * |
1595 | md_undefined_symbol (char *name ATTRIBUTE_UNUSED) | |
1596 | { | |
1597 | return NULL; | |
1598 | } | |
1599 | ||
1600 | \f | |
1601 | /* Turn a string in input_line_pointer into a floating point constant | |
1602 | of type TYPE, and store the appropriate bytes in *LITP. The number | |
1603 | of LITTLENUMS emitted is stored in *SIZEP. An error message is | |
1604 | returned, or NULL on OK. */ | |
1605 | ||
1606 | const char * | |
1607 | md_atof (int type, char *litP, int *sizeP) | |
1608 | { | |
5b7c81bd | 1609 | return ieee_md_atof (type, litP, sizeP, false); |
f8861f5d | 1610 | } |
d218e7fe JM |
1611 | |
1612 | \f | |
1613 | /* Determine whether the equal sign in the given string corresponds to | |
1614 | a BPF instruction, i.e. when it is not to be considered a symbol | |
1615 | assignment. */ | |
1616 | ||
1617 | bool | |
1618 | bpf_tc_equal_in_insn (int c ATTRIBUTE_UNUSED, char *str ATTRIBUTE_UNUSED) | |
1619 | { | |
1620 | uint8_t regno; | |
1621 | ||
1622 | /* Only pseudo-c instructions can have equal signs, and of these, | |
1623 | all that could be confused with a symbol assignment all start | |
1624 | with a register name. */ | |
1625 | if (asm_dialect == DIALECT_PSEUDOC) | |
1626 | { | |
1627 | char *w = parse_bpf_register (str, 'w', ®no); | |
1628 | char *r = parse_bpf_register (str, 'r', ®no); | |
1629 | ||
1630 | if ((w != NULL && *w == '\0') | |
1631 | || (r != NULL && *r == '\0')) | |
1632 | return 1; | |
1633 | } | |
1634 | ||
1635 | return 0; | |
1636 | } |