8.2. 程序入口
为了显示的目的,在本节中我们使用以下的例子。
278 (define_delay (and (eq_attr "type" "branch") in mips.md
279 (eq (symbol_ref "TARGET_MIPS16") (const_int 0)))
280 [(eq_attr "can_delay" "yes")
281 (nil)
282 (and (eq_attr "branch_likely" "yes")
283 (eq_attr "can_delay" "yes"))])
在 init_md_reader_args 处理之后,上面的模式将被作为以下的 rtx 对象载入内存。
图 35 : genattr - define_delay 的例子
43 (define_function_unit "k6_alux" 1 0 in k6.md
44 (and (eq_attr "cpu" "k6")
45 (eq_attr "type" "ishift,ishift1,rotate,rotate1,alu1,negnot,cld"))
46 1 1)
在 init_md_reader_args 处理之后,上面的模式将被作为以下的 rtx 对象载入内存。
图 36 : genattr - define_function_unit 的例子
135 (define_insn_reservation "pent_mul" 11 in pentium.md
136 (and (eq_attr "cpu" "pentium")
137 (eq_attr "type" "imul"))
138 "pentium-np*11")
在 init_md_reader_args 处理之后,上面的模式将被作为以下的 rtx 对象载入内存。
图 37 : genattr - define_insn_reservation 的例子
好,现在我们可以开始了。
175 int
176 main (int argc, char **argv) in genattr.c
177 {
178 rtx desc;
179 int have_delay = 0;
180 int have_annul_true = 0;
181 int have_annul_false = 0;
182 int num_insn_reservations = 0;
183 int num_units = 0;
184 struct range all_simultaneity, all_multiplicity;
185 struct range all_ready_cost, all_issue_delay, all_blockage;
186 struct function_unit *units = 0, *unit;
187 int i;
188
189 init_range (&all_multiplicity);
190 init_range (&all_simultaneity);
191 init_range (&all_ready_cost);
192 init_range (&all_issue_delay);
193 init_range (&all_blockage);
194
195 progname = "genattr";
196
197 if (argc <= 1)
198 fatal ("no input file name");
199
200 if (init_md_reader_args (argc, argv) != SUCCESS_EXIT_CODE)
201 return (FATAL_EXIT_CODE);
202
203 puts ("/* Generated automatically by the program `genattr'");
204 puts (" from the machine description file `md'. *//n");
205 puts ("#ifndef GCC_INSN_ATTR_H");
206 puts ("#define GCC_INSN_ATTR_H/n");
207
208 /* For compatibility, define the attribute `alternative', which is just
209 a reference to the variable `which_alternative'. */
210
211 puts ("#define HAVE_ATTR_alternative");
212 puts ("#define get_attr_alternative(insn) which_alternative");
213
214 /* Read the machine description. */
215
216 while (1)
217 {
218 int line_no, insn_code_number;
219
220 desc = read_md_rtx (&line_no, &insn_code_number);
221 if (desc == NULL)
222 break ;
223
224 if (GET_CODE (desc) == DEFINE_ATTR)
225 gen_attr (desc);
226
227 else if (GET_CODE (desc) == DEFINE_DELAY)
228 {
229 if (! have_delay)
230 {
231 printf ("#define DELAY_SLOTS/n");
232 printf ("extern int num_delay_slots (rtx);/n");
233 printf ("extern int eligible_for_delay (rtx, int, rtx, int);/n/n");
234 printf ("extern int const_num_delay_slots (rtx);/n/n");
235 have_delay = 1;
236 }
237
238 for (i = 0; i < XVECLEN (desc, 1); i += 3)
239 {
240 if (XVECEXP (desc, 1, i + 1) && ! have_annul_true)
241 {
242 printf ("#define ANNUL_IFTRUE_SLOTS/n");
243 printf ("extern int eligible_for_annul_true (rtx, int, rtx, int);/n");
244 have_annul_true = 1;
245 }
246
247 if (XVECEXP (desc, 1, i + 2) && ! have_annul_false)
248 {
249 printf ("#define ANNUL_IFFALSE_SLOTS/n");
250 printf ("extern int eligible_for_annul_false (rtx, int, rtx, int);/n");
251 have_annul_false = 1;
252 }
253 }
254 }
上面,在 225 行, define_attr 总是作为枚举值来处理。它为工具定义了常量。而至于 define_delay ,相关函数的声明被输出。注意到对于每个时间槽,将用到 3 个项。这里在我们的例子中, have_annul_true 将是 0 ,而 have_annual_false 将是 1 。
main (continued)
256 else if (GET_CODE (desc) == DEFINE_FUNCTION_UNIT)
257 {
258 const char *name = XSTR (desc, 0);
259 int multiplicity = XINT (desc, 1);
260 int simultaneity = XINT (desc, 2);
261 int ready_cost = MAX (XINT (desc, 4), 1);
262 int issue_delay = MAX (XINT (desc, 5), 1);
263 int issueexp_p = (XVEC (desc, 6) != 0);
264
265 for (unit = units; unit; unit = unit->next)
266 if (strcmp (unit->name, name) == 0)
267 break ;
268
269 if (unit == 0)
270 {
271 unit = xmalloc (sizeof (struct function_unit ));
272 unit->name = xstrdup (name);
273 unit->multiplicity = multiplicity;
274 unit->simultaneity = simultaneity;
275 unit->ready_cost.min = unit->ready_cost.max = ready_cost;
276 unit->issue_delay.min = unit->issue_delay.max = issue_delay;
277 unit->next = units;
278 units = unit;
279 num_units++;
280
281 extend_range (&all_multiplicity, multiplicity, multiplicity);
282 extend_range (&all_simultaneity, simultaneity, simultaneity);
283 }
284 else if (unit->multiplicity != multiplicity
285 || unit->simultaneity != simultaneity)
286 fatal ("Differing specifications given for `%s' function unit",
287 unit->name);
288
289 extend_range (&unit->ready_cost, ready_cost, ready_cost);
290 extend_range (&unit->issue_delay,
291 issueexp_p ? 1 : issue_delay, issue_delay);
292 extend_range (&all_ready_cost,
293 unit->ready_cost.min, unit->ready_cost.max);
294 extend_range (&all_issue_delay,
295 unit->issue_delay.min, unit->issue_delay.max);
296 }
297 else if (GET_CODE (desc) == DEFINE_INSN_RESERVATION)
298 num_insn_reservations++;
299 }
300
301 if (num_units > 0 || num_insn_reservations > 0)
302 {
303 if (num_units > 0)
304 printf ("#define TRADITIONAL_PIPELINE_INTERFACE 1/n");
305
306 if (num_insn_reservations > 0)
307 printf ("#define DFA_PIPELINE_INTERFACE 1/n");
308
309 /* Compute the range of blockage cost values. See genattrtab.c
310 for the derivation. BLOCKAGE (E,C) when SIMULTANEITY is zero is
311
312 MAX (ISSUE-DELAY (E,C),
313 READY-COST (E) - (READY-COST (C) - 1))
314
315 and otherwise
316
317 MAX (ISSUE-DELAY (E,C),
318 READY-COST (E) - (READY-COST (C) - 1),
319 READY-COST (E) - FILL-TIME) */
320
321 for (unit = units; unit; unit = unit->next)
322 {
323 struct range blockage;
324
325 blockage = unit->issue_delay;
326 blockage.max = MAX (unit->ready_cost.max
327 - (unit->ready_cost.min - 1),
328 blockage.max);
329 blockage.min = MAX (1, blockage.min);
330
331 if (unit->simultaneity != 0)
332 {
333 int fill_time = ((unit->simultaneity - 1)
334 * unit->issue_delay.min);
335 blockage.min = MAX (unit->ready_cost.min - fill_time,
336 blockage.min);
337 blockage.max = MAX (unit->ready_cost.max - fill_time,
338 blockage.max);
339 }
340 extend_range (&all_blockage, blockage.min, blockage.max);
341 }
结构体 function_unit 收集 define_function_uint 模式的内容。看到 define_function_uint 模式必须具有唯一的名字,正如上面的 286 行所示。
44 struct function_unit in genattr.c
45 {
46 char *name; /* Function unit name. */
47 struct function_unit *next; /* Next function unit. */
48 int multiplicity; /* Number of units of this type. */
49 int simultaneity; /* Maximum number of simultaneous insns
50 on this function unit or 0 if unlimited. */
51 struct range ready_cost; /* Range of ready cost values. */
52 struct range issue_delay; /* Range of issue delay values. */
53 };
62 static void
63 extend_range (struct range *range, int min, int max)
64 {
65 if (range->min > min) range->min = min;
66 if (range->max < max) range->max = max;
67 }
在 genattr.c 中对 define_function_uint 及 define_insn_reservation 的处理都相当简单 , 只是保存所有数据的取值范围。然后下面的代码产生输出。
main (continued)
343 write_units (num_units, &all_multiplicity, &all_simultaneity,
344 &all_ready_cost, &all_issue_delay, &all_blockage);
345
346 /* Output interface for pipeline hazards recognition based on
347 DFA (deterministic finite state automata. */
348 printf ("/n/* DFA based pipeline interface. */");
349 printf ("/n#ifndef AUTOMATON_ALTS/n");
350 printf ("#define AUTOMATON_ALTS 0/n");
351 printf ("#endif/n/n");
352 printf ("/n#ifndef AUTOMATON_STATE_ALTS/n");
353 printf ("#define AUTOMATON_STATE_ALTS 0/n");
354 printf ("#endif/n/n");
355 printf ("#ifndef CPU_UNITS_QUERY/n");
356 printf ("#define CPU_UNITS_QUERY 0/n");
357 printf ("#endif/n/n");
358 /* Interface itself: */
359 printf ("extern int max_dfa_issue_rate;/n/n");
360 printf ("/* The following macro value is calculated from the/n");
361 printf (" automaton based pipeline description and is equal to/n");
362 printf (" maximal number of all insns described in constructions/n");
363 printf (" `define_insn_reservation' which can be issued on the/n");
364 printf (" same processor cycle. *//n");
365 printf ("#define MAX_DFA_ISSUE_RATE max_dfa_issue_rate/n/n");
366 printf ("/* Insn latency time defined in define_insn_reservation. *//n");
367 printf ("extern int insn_default_latency (rtx);/n/n");
368 printf ("/* Return nonzero if there is a bypass for given insn/n");
369 printf (" which is a data producer. *//n");
370 printf ("extern int bypass_p (rtx);/n/n");
371 printf ("/* Insn latency time on data consumed by the 2nd insn./n");
372 printf (" Use the function if bypass_p returns nonzero for/n");
373 printf (" the 1st insn. *//n");
374 printf ("extern int insn_latency (rtx, rtx);/n/n");
375 printf ("/n#if AUTOMATON_ALTS/n");
376 printf ("/* The following function returns number of alternative/n");
377 printf (" reservations of given insn. It may be used for better/n");
378 printf (" insns scheduling heuristics. *//n");
379 printf ("extern int insn_alts (rtx);/n/n");
380 printf ("#endif/n/n");
381 printf ("/* Maximal possible number of insns waiting results being/n");
382 printf (" produced by insns whose execution is not finished. *//n");
383 printf ("extern int max_insn_queue_index;/n/n");
384 printf ("/* Pointer to data describing current state of DFA. *//n");
385 printf ("typedef void *state_t;/n/n");
386 printf ("/* Size of the data in bytes. *//n");
387 printf ("extern int state_size (void);/n/n");
388 printf ("/* Initiate given DFA state, i.e. Set up the state/n");
389 printf (" as all functional units were not reserved. *//n");
390 printf ("extern void state_reset (state_t);/n");
391 printf ("/* The following function returns negative value if given/n");
392 printf (" insn can be issued in processor state described by given/n");
393 printf (" DFA state. In this case, the DFA state is changed to/n");
394 printf (" reflect the current and future reservations by given/n");
395 printf (" insn. Otherwise the function returns minimal time/n");
396 printf (" delay to issue the insn. This delay may be zero/n");
397 printf (" for superscalar or VLIW processors. If the second/n");
398 printf (" parameter is NULL the function changes given DFA state/n");
399 printf (" as new processor cycle started. *//n");
400 printf ("extern int state_transition (state_t, rtx);/n");
401 printf ("/n#if AUTOMATON_STATE_ALTS/n");
402 printf ("/* The following function returns number of possible/n");
403 printf (" alternative reservations of given insn in given/n");
404 printf (" DFA state. It may be used for better insns scheduling/n");
405 printf (" heuristics. By default the function is defined if/n");
406 printf (" macro AUTOMATON_STATE_ALTS is defined because its/n");
407 printf (" implementation may require much memory. *//n");
408 printf ("extern int state_alts (state_t, rtx);/n");
409 printf ("#endif/n/n");
410 printf ("extern int min_issue_delay (state_t, rtx);/n");
411 printf ("/* The following function returns nonzero if no one insn/n");
412 printf (" can be issued in current DFA state. *//n");
413 printf ("extern int state_dead_lock_p (state_t);/n");
414 printf ("/* The function returns minimal delay of issue of the 2nd/n");
415 printf (" insn after issuing the 1st insn in given DFA state./n");
416 printf (" The 1st insn should be issued in given state (i.e./n");
417 printf (" state_transition should return negative value for/n");
418 printf (" the insn and the state). Data dependencies between/n");
419 printf (" the insns are ignored by the function. *//n");
420 printf
421 ("extern int min_insn_conflict_delay (state_t, rtx, rtx);/n");
422 printf ("/* The following function outputs reservations for given/n");
423 printf (" insn as they are described in the corresponding/n");
424 printf (" define_insn_reservation. *//n");
425 printf ("extern void print_reservation (FILE *, rtx);/n");
426 printf ("/n#if CPU_UNITS_QUERY/n");
427 printf ("/* The following function returns code of functional unit/n");
428 printf (" with given name (see define_cpu_unit). *//n");
429 printf ("extern int get_cpu_unit_code (const char *);/n");
430 printf ("/* The following function returns nonzero if functional/n");
431 printf (" unit with given code is currently reserved in given/n");
432 printf (" DFA state. *//n");
433 printf ("extern int cpu_unit_reservation_p (state_t, int);/n");
434 printf ("#endif/n/n");
435 printf ("/* Clean insn code cache. It should be called if there/n");
436 printf (" is a chance that condition value in a/n");
437 printf (" define_insn_reservation will be changed after/n");
438 printf (" last call of dfa_start. *//n");
439 printf ("extern void dfa_clean_insn_cache (void);/n/n");
440 printf ("/* Initiate and finish work with DFA. They should be/n");
441 printf (" called as the first and the last interface/n");
442 printf (" functions. *//n");
443 printf ("extern void dfa_start (void);/n");
444 printf ("extern void dfa_finish (void);/n");
445 }
446 else
447 {
448 /* Otherwise we do no scheduling, but we need these typedefs
449 i n order to avoid uglifying other code with more ifdefs. */
450 printf ("typedef void *state_t;/n/n");
451 }
452
453 /* Output flag masks for use by reorg.
454
455 Flags are used to hold branch direction and prediction information
456 for use by eligible_for_... */
457 printf("/n#define ATTR_FLAG_forward/t0x1/n");
458 printf("#define ATTR_FLAG_backward/t0x2/n");
459 printf("#define ATTR_FLAG_likely/t0x4/n");
460 printf("#define ATTR_FLAG_very_likely/t0x8/n");
461 printf("#define ATTR_FLAG_unlikely/t0x10/n");
462 printf("#define ATTR_FLAG_very_unlikely/t0x20/n");
463
464 puts("/n#endif /* GCC_INSN_ATTR_H */");
465
466 if (ferror (stdout) || fflush (stdout) || fclose (stdout))
467 return FATAL_EXIT_CODE;
468
469 return SUCCESS_EXIT_CODE;
470 }
write_uints 也是一个简单函数。
129 write_units (int num_units, struct range *multiplicity, struct range *simultaneity,
130 struct range *ready_cost, struct range *issue_delay,
131 struct range *blockage)
132 {
133 int i, q_size;
134
135 printf ("#define INSN_SCHEDULING/n/n");
136 printf ("extern int result_ready_cost (rtx);/n");
137 printf ("extern int function_units_used (rtx);/n/n");
138 printf ("extern const struct function_unit_desc/n");
139 printf ("{/n");
140 printf (" const char *const name;/n");
141 printf (" const int bitmask;/n");
142 printf (" const int multiplicity;/n");
143 printf (" const int simultaneity;/n");
144 printf (" const int default_cost;/n");
145 printf (" const int max_issue_delay;/n");
146 printf (" int (*const ready_cost_function) (rtx);/n");
147 printf (" int (*const conflict_cost_function) (rtx, rtx);/n");
148 printf (" const int max_blockage;/n");
149 printf (" unsigned int (*const blockage_range_function) (rtx);/n");
150 printf (" int (*const blockage_function) (rtx, rtx);/n");
151 printf ("} function_units[];/n/n");
152 printf ("#define FUNCTION_UNITS_SIZE %d/n", num_units);
153 printf ("#define MIN_MULTIPLICITY %d/n", multiplicity->min);
154 printf ("#define MAX_MULTIPLICITY %d/n", multiplicity->max);
155 printf ("#define MIN_SIMULTANEITY %d/n", simultaneity->min);
156 printf ("#define MAX_SIMULTANEITY %d/n", simultaneity->max);
157 printf ("#define MIN_READY_COST %d/n", ready_cost->min);
158 printf ("#define MAX_READY_COST %d/n", ready_cost->max);
159 printf ("#define MIN_ISSUE_DELAY %d/n", issue_delay->min);
160 printf ("#define MAX_ISSUE_DELAY %d/n", issue_delay->max);
161 printf ("#define MIN_BLOCKAGE %d/n", blockage->min);
162 printf ("#define MAX_BLOCKAGE %d/n", blockage->max);
163 for (i = 0; (1 << i) < blockage->max; i++)
164 ;
165 printf ("#define BLOCKAGE_BITS %d/n", i + 1);
166
167 /* INSN_QUEUE_SIZE is a power of two larger than MAX_BLOCKAGE and
168 MAX_READY_COST. This is the longest time an insn may be queued. */
169 i = MAX (blockage->max, ready_cost->max);
170 for (q_size = 1; q_size <= i; q_size <<= 1)
171 ;
172 printf ("#define INSN_QUEUE_SIZE %d/n", q_size);
173 }
Insn-attr.h 实际上类似于下面要产生的 insn-attrtab.c 的头文件。