6. genextract 工具
Genextract 从机器描述文件输出 insn-extract.c 。这个文件定义了,从指定的模式中,提取操作数的函数。
346 int
347 main (int argc, char **argv) in genextract.c
348 {
349 rtx desc;
350 int i;
351 struct extraction *p;
352 struct code_ptr *link;
353 const char *name;
354
355 progname = "genextract";
356
357 if (argc <= 1)
358 fatal ("no input file name");
359
360 if (init_md_reader_args (argc, argv) != SUCCESS_EXIT_CODE)
361 return (FATAL_EXIT_CODE);
362
363 /* Assign sequential codes to all entries in the machine description
364 i n parallel with the tables in insn-output.c. */
365
366 insn_code_number = 0;
367
368 printf ("/* Generated automatically by the program `genextract'/n/
369 from the machine description file `md'. *//n/n");
370
371 printf ("#include /"config.h/"/n");
372 printf ("#include /"system.h/"/n");
373 printf ("#include /"coretypes.h/"/n");
374 printf ("#include /"tm.h/"/n");
375 printf ("#include /"rtl.h/"/n");
376 printf ("#include /"insn-config.h/"/n");
377 printf ("#include /"recog.h/"/n");
378 printf ("#include /"toplev.h/"/n/n");
379
380 /* This variable exists only so it can be the "location"
381 of any missing operand whose numbers are skipped by a given pattern. */
382 printf ("static rtx junk ATTRIBUTE_UNUSED;/n");
383
384 printf ("void/ninsn_extract (rtx insn)/n");
385 printf ("{/n");
386 printf (" rtx *ro = recog_data.operand;/n");
387 printf (" rtx **ro_loc = recog_data.operand_loc;/n");
388 printf (" rtx pat = PATTERN (insn);/n");
389 printf (" int i ATTRIBUTE_UNUSED;/n/n");
390 #ifdef ENABLE_CHECKING
391 printf (" memset (ro, 0xab, sizeof (*ro) * MAX_RECOG_OPERANDS);/n");
392 printf (" memset (ro_loc, 0xab, sizeof (*ro_loc) * MAX_RECOG_OPERANDS);/n");
393 #endif
394 printf (" switch (INSN_CODE (insn))/n");
395 printf (" {/n");
396 printf (" case -1:/n");
397 printf (" fatal_insn_not_found (insn);/n/n");
398
399 /* Read the machine description. */
400
401 while (1)
402 {
403 int line_no;
404
405 desc = read_md_rtx (&line_no, &insn_code_number );
406 if (desc == NULL)
407 break ;
408
409 if (GET_CODE (desc) == DEFINE_INSN)
410 {
411 record_insn_name (insn_code_number , XSTR (desc, 0));
412 gen_insn (desc);
413 }
414
415 else if (GET_CODE (desc) == DEFINE_PEEPHOLE)
416 {
417 struct code_ptr *link = xmalloc (sizeof (struct code_ptr));
418
419 link->insn_code = insn_code_number ;
420 link->next = peepholes ;
421 peepholes = link;
422 }
423 }
再一次, init_md_reader_args 读入机器描述文件,构建 rtx 对象,并把它们放入链表中;然后函数 read_md_rtx 一个个地读入 rtx 对象。
在上面的 415 行, define_peephole 模式现在不建议使用了,而是期望使用 define_peephole2 模式。这两个模式定义了一个方式来找出窥孔优化的机会。对于 define_peephole ,因为其输出模板给出了结果,因此它的处理非常直接。
至于 define_insn ,在 411 行,该模式的名字,以 insn_code_number 为索引,被记录在 insn_name_ptr 中。这里我们还是使用下面的模式作为例子。
467 (define_insn "cmpdi_ccno_1_rex64" in i386.md
468 [(set (reg 17)
469 (compare (match_operand:DI 0 "nonimmediate_operand" "r,?mr")
470 (match_operand:DI 1 "const0_operand" "n,n")))]
471 "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)"
472 "@
473 test{q}/t{%0, %0|%0, %0}
474 cmp{q}/t{%1, %0|%0, %1}"
475 [(set_attr "type" "test,icmp")
476 (set_attr "length_immediate" "0,1")
477 (set_attr "mode" "DI")])
图 34 : genextract - define_insn 的例子
103 static void
104 gen_insn (rtx insn) in genextract.c
105 {
106 int i;
107 struct extraction *p;
108 struct code_ptr *link;
109
110 op_count = 0;
111 dup_count = 0;
112
113 /* No operands seen so far in this pattern. */
114 memset (oplocs , 0, sizeof oplocs );
115
116 /* Walk the insn's pattern, remembering at all times the path
117 down to the walking point. */
118
119 if (XVECLEN (insn, 1) == 1)
120 walk_rtx (XVECEXP (insn, 1, 0), "");
121 else
122 for (i = XVECLEN (insn, 1) - 1; i >= 0; i--)
123 {
124 char path[2];
125
126 path[0] = 'a' + i;
127 path[1] = 0;
128
129 walk_rtx (XVECEXP (insn, 1, i), path);
130 }
walk_rtx 就像我们之前看到的处理其它模式的函数那样,递归来处理 define_insn 模式。
182 static void
183 walk_rtx (rtx x, const char *path)
184 {
185 RTX_CODE code;
186 int i;
187 int len;
188 const char *fmt;
189 int depth = strlen (path);
190 char *newpath;
191
192 if (x == 0)
193 return ;
194
195 code = GET_CODE (x);
196
197 switch (code)
198 {
199 case PC:
200 case CC0:
201 case CONST_INT:
202 case SYMBOL_REF:
203 return ;
204
205 case MATCH_OPERAND:
206 case MATCH_SCRATCH:
207 oplocs [XINT (x, 0)] = xstrdup (path);
208 op_count = MAX (op_count , XINT (x, 0) + 1);
209 break ;
210
211 case MATCH_DUP:
212 duplocs [dup_count ] = xstrdup (path);
213 dupnums [dup_count ] = XINT (x, 0);
214 dup_count ++;
215 break ;
216
217 case MATCH_PAR_DUP:
218 case MATCH_OP_DUP:
219 duplocs [dup_count ] = xstrdup (path);
220 dupnums [dup_count ] = XINT (x, 0);
221 dup_count ++;
222
223 newpath = xmalloc (depth + 2);
224 strcpy (newpath, path);
225 newpath[depth + 1] = 0;
226
227 for (i = XVECLEN (x, 1) - 1; i >= 0; i--)
228 {
229 newpath[depth] = (code == MATCH_OP_DUP ? '0' : 'a') + i;
230 walk_rtx (XVECEXP (x, 1, i), newpath);
231 }
232 free (newpath);
233 return ;
234
235 case MATCH_OPERATOR:
236 oplocs [XINT (x, 0)] = xstrdup (path);
237 op_count = MAX (op_count , XINT (x, 0) + 1);
238
239 newpath = xmalloc (depth + 2);
240 strcpy (newpath, path);
241 newpath[depth + 1] = 0;
242
243 for (i = XVECLEN (x, 2) - 1; i >= 0; i--)
244 {
245 newpath[depth] = '0' + i;
246 walk_rtx (XVECEXP (x, 2, i), newpath);
247 }
248 free (newpath);
249 return ;
250
251 case MATCH_PARALLEL:
252 oplocs [XINT (x, 0)] = xstrdup (path);
253 op_count = MAX (op_count , XINT (x, 0) + 1);
254
255 newpath = xmalloc (depth + 2);
256 strcpy (newpath, path);
257 newpath[depth + 1] = 0;
258
259 for (i = XVECLEN (x, 2) - 1; i >= 0; i--)
260 {
261 newpath[depth] = 'a' + i;
262 walk_rtx (XVECEXP (x, 2, i), newpath);
263 }
264 free (newpath);
265 return ;
266
267 case ADDRESS:
268 walk_rtx (XEXP (x, 0), path);
269 return ;
270
271 default :
272 break ;
273 }
274
275 newpath = xmalloc (depth + 2);
276 strcpy (newpath, path);
277 newpath[depth + 1] = 0;
278
279 fmt = GET_RTX_FORMAT (code);
280 len = GET_RTX_LENGTH (code);
281 for (i = 0; i < len; i++)
282 {
283 if (fmt[i] == 'e' || fmt[i] == 'u')
284 {
285 newpath[depth] = '0' + i;
286 walk_rtx (XEXP (x, i), newpath);
287 }
288 else if (fmt[i] == 'E')
289 {
290 int j;
291 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
292 {
293 newpath[depth] = 'a' + j;
294 walk_rtx (XVECEXP (x, i, j), newpath);
295 }
296 }
297 }
298 free (newpath);
299 }
注意到 rtl 编码 SET 的模板具有格式“ ee ”, REG 则具有格式“ i00 ”, COMPARE 的格式也是“ ee ”。对于我们的例子, walk_rtx 相当简单。当它返回时, oplocs 指向‘ 0 ’及‘ 1 ’, op_count 是 2 。回到 gen_insn , extraction 及 code_ptr 具有如下的定义:
43 struct extraction in genextract.c
44 {
45 int op_count;
46 char *oplocs[MAX_RECOG_OPERANDS];
47 int dup_count;
48 char *duplocs[MAX_DUP_OPERANDS];
49 int dupnums[MAX_DUP_OPERANDS];
50 struct code_ptr *insns;
51 struct extraction *next;
52 };
56 struct code_ptr
57 {
58 int insn_code;
59 struct code_ptr *next;
60 };
其中, 45 行的 op_count 记录指令中较大的指令号; 47 行的 dup_count 则是在指令中出现的 MATCH_DUP 的次数,从 0 开始计数。 Duplocs 及 dupnums 对应于每个 MATCH_DUP ,其中 duplocs 记录了 MATCH_DUP 的操作数的行号, dupnums 记录了 MATCH_DUP 的操作数个数。宏 MAX_DUP_OPERANDS 及 MAX_RECOG_OPERANDS 是由工具 genconfig 根据机器描述产生的配置的一部分。
对于具有相似形式的模式,它们将共享同一个 extraction ,而在 50 行的 insns 域将保存它们的指令编码号的记录,所有的 extraction 将通过 51 行的 next 域链接在一起。
gen_insn (continued)
132 link = xmalloc (sizeof (struct code_ptr));
133 link->insn_code = insn_code_number ;
134
135 /* See if we find something that already had this extraction method. */
136
137 for (p = extractions ; p; p = p->next)
138 {
139 if (p->op_count != op_count || p->dup_count != dup_count )
140 continue ;
141
142 for (i = 0; i < op_count ; i++)
143 if (p->oplocs[i] != oplocs [i]
144 && ! (p->oplocs[i] != 0 && oplocs [i] != 0
145 && ! strcmp (p->oplocs[i], oplocs [i])))
146 break ;
147
148 if (i != op_count )
149 continue ;
150
151 for (i = 0; i < dup_count ; i++)
152 if (p->dupnums[i] != dupnums [i]
153 || strcmp (p->duplocs[i], duplocs [i]))
154 break ;
155
156 if (i != dup_count )
157 continue ;
158
159 /* This extraction is the same as ours. Just link us in. */
160 link->next = p->insns;
161 p->insns = link;
162 return ;
163 }
164
165 /* Otherwise, make a new extraction method. */
166
167 p = xmalloc (sizeof (struct extraction));
168 p->op_count = op_count ;
169 p->dup_count = dup_count ;
170 p->next = extractions ;
171 extractions = p;
172 p->insns = link;
173 link->next = 0;
174
175 for (i = 0; i < op_count ; i++)
176 p->oplocs[i] = oplocs [i];
177
178 for (i = 0; i < dup_count ; i++)
179 p->dupnums[i] = dupnums [i], p->duplocs[i] = duplocs [i];
180 }
在 gen_insn 的余下部分 , 它把 walk_rtx 的结果链接入 extractions 链表。回到 main ,它将处理这个数据。
main (continued)
425 /* Write out code to handle peepholes and the insn_codes that it should
426 be called for. */
427 if (peepholes )
428 {
429 for (link = peepholes ; link; link = link->next)
430 printf (" case %d:/n", link->insn_code);
431
432 /* The vector in the insn says how many operands it has.
433 And all it contains are operands. In fact, the vector was
434 created just for the sake of this function. We need to set the
435 location of the operands for sake of simplifications after
436 extraction, like eliminating subregs. */
437 printf (" for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)/n");
438 printf (" ro[i] = *(ro_loc[i] = &XVECEXP (pat, 0, i));/n");
439 printf (" break;/n/n");
440 }
441
442 /* Write out all the ways to extract insn operands. */
443 for (p = extractions ; p; p = p->next)
444 {
445 for (link = p->insns; link; link = link->next)
446 {
447 i = link->insn_code;
448 name = get_insn_name (i);
449 if (name)
450 printf (" case %d: /* %s *//n", i, name);
451 else
452 printf (" case %d:/n", i);
453 }
454
455 for (i = 0; i < p->op_count; i++)
456 {
457 if (p->oplocs[i] == 0)
458 {
459 printf (" ro[%d] = const0_rtx;/n", i);
460 printf (" ro_loc[%d] = &junk;/n", i);
461 }
462 else
463 {
464 printf (" ro[%d] = *(ro_loc[%d] = &", i, i);
465 print_path (p->oplocs[i]);
466 printf (");/n");
467 }
468 }
469
470 for (i = 0; i < p->dup_count; i++)
471 {
472 printf (" recog_data.dup_loc[%d] = &", i);
473 print_path (p->duplocs[i]);
474 printf (";/n");
475 printf (" recog_data.dup_num[%d] = %d;/n", i, p->dupnums[i]);
476 }
477
478 printf (" break;/n/n");
479 }
480
481 /* This should never be reached. Note that we would also reach this abort
482 if we tried to extract something whose INSN_CODE was a DEFINE_EXPAND or
483 DEFINE_SPLIT, but that is correct. */
484 printf (" default:/n abort ();/n");
485
486 printf (" }/n}/n");
487
488 fflush (stdout);
489 return (ferror (stdout) != 0 ? FATAL_EXIT_CODE : SUCCESS_EXIT_CODE);
490 }
剩下的代码相当简单,但在实际的输出代码中,我们的例子出现在 i386 32 位系统中,它在 read_md_rtx 中被丢弃,因为 1002 行的 maybe_eval_c_test 显示其条件可以证明总是 false 。对于真实的输出,看起来将是如下:
1 /* Generated automatically by the program `genextract'
2 from the machine description file `md'. */
3
4 #include "config.h"
5 #include "system.h"
6 #include "coretypes.h"
7 #include "tm.h"
8 #include "rtl.h"
9 #include "insn-config.h"
10 #include "recog.h"
11 #include "toplev.h"
12
13 static rtx junk ATTRIBUTE_UNUSED;
14 void
15 insn_extract (rtx insn)
16 {
17 rtx *ro = recog_data.operand;
18 rtx **ro_loc = recog_data.operand_loc;
19 rtx pat = PATTERN (insn);
20 int i ATTRIBUTE_UNUSED;
21
22 memset (ro, 0xab, sizeof (*ro) * MAX_RECOG_OPERANDS);
23 memset (ro_loc, 0xab, sizeof (*ro_loc) * MAX_RECOG_OPERANDS);
24 switch (INSN_CODE (insn))
25 {
26 case -1:
27 fatal_insn_not_found (insn);
28
29 case 819: /* monitor */
30 ro[0] = *(ro_loc[0] = &XVECEXP (pat, 0, 0));
31 ro[1] = *(ro_loc[1] = &XVECEXP (pat, 0, 1));
32 ro[2] = *(ro_loc[2] = &XVECEXP (pat, 0, 2));
33 break;
34 …
35 default:
36 gcc_unreachable ();
37 }
38 }
7. genpreds 工具
Genpreds 从机器描述文件输出文件 tm-preds.h 。这个文件声明了由 define_insn 等使用的断言函数。注意到断言函数被定义在文件 `arch`.c 中。对于不同的架构,该工具输出匹配的 tm-preds.h 。
1 int
2 main (void) in genpreds.c
3 {
4 puts ("/* Generated automatically by the program `genpreds'. *//n");
5 puts ("#ifndef GCC_TM_PREDS_H");
6 puts ("#define GCC_TM_PREDS_H/n");
7
8 output_predicate_decls ();
9
10 puts ("#endif /* GCC_TM_PREDS_H */");
11
12 if (ferror (stdout) || fflush (stdout) || fclose (stdout))
13 return FATAL_EXIT_CODE;
14
15 return SUCCESS_EXIT_CODE;
16 }
一个简单的 main 函数。
33 static void
34 output_predicate_decls (void) in genpreds.c
35 {
36 #ifdef PREDICATE_CODES
37 static const struct {
38 const char *const name;
39 const RTX_CODE codes[NUM_RTX_CODE];
40 } predicate[] = {
42 };
43 size_t i;
44
45 puts ("#ifdef RTX_CODE/n");
46 for (i = 0; i < ARRAY_SIZE (predicate); i++)
47 printf ("extern int %s (rtx, enum machine_mode);/n",
48 predicate[i].name);
49 puts ("/n#endif /* RTX_CODE *//n");
50 #endif
51 }
PREDICATE_CODES 被定义在目标系统中。对于 i386 系统 , 它是如下。
2916 #define PREDICATE_CODES / in i386.h
2917 {"x86_64_immediate_operand", {CONST_INT, SUBREG, REG, /
2918 SYMBOL_REF, LABEL_REF, CONST}}, /
2919 {"x86_64_nonmemory_operand", {CONST_INT, SUBREG, REG, /
2920 SYMBOL_REF, LABEL_REF, CONST}}, /
2921 {"x86_64_movabs_operand", {CONST_INT, SUBREG, REG, /
2922 SYMBOL_REF, LABEL_REF, CONST}}, /
2923 {"x86_64_szext_nonmemory_operand", {CONST_INT, SUBREG, REG, /
2924 SYMBOL_REF, LABEL_REF, CONST}}, /
2925 {"x86_64_general_operand", {CONST_INT, SUBREG, REG, MEM, /
2926 SYMBOL_REF, LABEL_REF, CONST}}, /
2927 {"x86_64_szext_general_operand", {CONST_INT, SUBREG, REG, MEM, /
2928 SYMBOL_REF, LABEL_REF, CONST}}, /
2929 {"x86_64_zext_immediate_operand", {CONST_INT, CONST_DOUBLE, CONST, /
2930 SYMBOL_REF, LABEL_REF}}, /
2931 {"shiftdi_operand", {SUBREG, REG, MEM}}, /
2932 {"const_int_1_31_operand", {CONST_INT}}, /
2933 {"symbolic_operand", {SYMBOL_REF, LABEL_REF, CONST}}, /
2934 {"aligned_operand", {CONST_INT, CONST_DOUBLE, CONST, SYMBOL_REF, /
2935 LABEL_REF, SUBREG, REG, MEM}}, /
2936 {"pic_symbolic_operand", {CONST}}, /
2937 {"call_insn_operand", {REG, SUBREG, MEM, SYMBOL_REF}}, /
2938 {"sibcall_insn_operand", {REG, SUBREG, SYMBOL_REF}}, /
2939 {"constant_call_address_operand", {SYMBOL_REF, CONST}}, /
2940 {"const0_operand", {CONST_INT, CONST_DOUBLE}}, /
2941 {"const1_operand", {CONST_INT}}, /
2942 {"const248_operand", {CONST_INT}}, /
2943 {"const_0_to_3_operand", {CONST_INT}}, /
2944 {"const_0_to_7_operand", {CONST_INT}}, /
2945 {"const_0_to_15_operand", {CONST_INT}}, /
2946 {"const_0_to_255_operand", {CONST_INT}}, /
2947 {"incdec_operand", {CONST_INT}}, /
2948 {"mmx_reg_operand", {REG}}, /
2949 {"reg_no_sp_operand", {SUBREG, REG}}, /
2950 {"general_no_elim_operand", {CONST_INT, CONST_DOUBLE, CONST, /
2951 SYMBOL_REF, LABEL_REF, SUBREG, REG, MEM}}, /
2952 {"nonmemory_no_elim_operand", {CONST_INT, REG, SUBREG}}, /
2953 {"index_register_operand", {SUBREG, REG}}, /
2954 {"flags_reg_operand", {REG}}, /
2955 {"q_regs_operand", {SUBREG, REG}}, /
2956 {"non_q_regs_operand", {SUBREG, REG}}, /
2957 {"fcmov_comparison_operator", {EQ, NE, LTU, GTU, LEU, GEU, UNORDERED, /
2958 ORDERED, LT, UNLT, GT, UNGT, LE, UNLE, /
2959 GE, UNGE, LTGT, UNEQ}}, /
2960 {"sse_comparison_operator", {EQ, LT, LE, UNORDERED, NE, UNGE, UNGT, /
2961 ORDERED, UNEQ, UNLT, UNLE, LTGT, GE, GT /
2962 }}, /
2963 {"ix86_comparison_operator", {EQ, NE, LE, LT, GE, GT, LEU, LTU, GEU, /
2964 GTU, UNORDERED, ORDERED, UNLE, UNLT, /
2965 UNGE, UNGT, LTGT, UNEQ }}, /
2966 {"ix86_carry_flag_operator", {LTU, LT, UNLT, GT, UNGT, LE, UNLE, /
2967 GE, UNGE, LTGT, UNEQ}}, /
2968 {"cmp_fp_expander_operand", {CONST_DOUBLE, SUBREG, REG, MEM}}, /
2969 {"ext_register_operand", {SUBREG, REG}}, /
2970 {"binary_fp_operator", {PLUS, MINUS, MULT, DIV}}, /
2971 {"mult_operator", {MULT}}, /
2972 {"div_operator", {DIV}}, /
2973 {"arith_or_logical_operator", {PLUS, MULT, AND, IOR, XOR, SMIN, SMAX, /
2974 UMIN, UMAX, COMPARE, MINUS, DIV, MOD, /
2975 UDIV, UMOD, ASHIFT, ROTATE, ASHIFTRT, /
2976 LSHIFTRT, ROTATERT}}, /
2977 {"promotable_binary_operator", {PLUS, MULT, AND, IOR, XOR, ASHIFT}}, /
2978 {"memory_displacement_operand", {MEM}}, /
2979 {"cmpsi_operand", {CONST_INT, CONST_DOUBLE, CONST, SYMBOL_REF, /
2980 LABEL_REF, SUBREG, REG, MEM, AND}}, /
2981 {"long_memory_operand", {MEM}}, /
2982 {"tls_symbolic_operand", {SYMBOL_REF}}, /
2983 {"global_dynamic_symbolic_operand", {SYMBOL_REF}}, /
2984 {"local_dynamic_symbolic_operand", {SYMBOL_REF}}, /
2985 {"initial_exec_symbolic_operand", {SYMBOL_REF}}, /
2986 {"local_exec_symbolic_operand", {SYMBOL_REF}}, /
2987 {"any_fp_register_operand", {REG}}, /
2988 {"register_and_not_any_fp_reg_operand", {REG}}, /
2989 {"fp_register_operand", {REG}}, /
2990 {"register_and_not_fp_reg_operand", {REG}}, /
2991 {"zero_extended_scalar_load_operand", {MEM}}, /
2992 {"vector_move_operand", {CONST_VECTOR, SUBREG, REG, MEM}}, /
2993 {"no_seg_address_operand", {CONST_INT, CONST_DOUBLE, CONST, SYMBOL_REF, /
2994 LABEL_REF, SUBREG, REG, MEM, PLUS, MULT}}, /
2995 {"compare_operator", {COMPARE}},
以上的函数声明将被输出。在 4.3 版本及之后,这些定义也将从一个新的机器描述文件—— predicate.md 中产生。