1 | /* Passes for transactional memory support. |
2 | Copyright (C) 2008-2023 Free Software Foundation, Inc. |
3 | Contributed by Richard Henderson <rth@redhat.com> |
4 | and Aldy Hernandez <aldyh@redhat.com>. |
5 | |
6 | This file is part of GCC. |
7 | |
8 | GCC is free software; you can redistribute it and/or modify it under |
9 | the terms of the GNU General Public License as published by the Free |
10 | Software Foundation; either version 3, or (at your option) any later |
11 | version. |
12 | |
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
16 | for more details. |
17 | |
18 | You should have received a copy of the GNU General Public License |
19 | along with GCC; see the file COPYING3. If not see |
20 | <http://www.gnu.org/licenses/>. */ |
21 | |
22 | #include "config.h" |
23 | #include "system.h" |
24 | #include "coretypes.h" |
25 | #include "backend.h" |
26 | #include "target.h" |
27 | #include "rtl.h" |
28 | #include "tree.h" |
29 | #include "gimple.h" |
30 | #include "cfghooks.h" |
31 | #include "tree-pass.h" |
32 | #include "ssa.h" |
33 | #include "cgraph.h" |
34 | #include "gimple-pretty-print.h" |
35 | #include "diagnostic-core.h" |
36 | #include "fold-const.h" |
37 | #include "tree-eh.h" |
38 | #include "calls.h" |
39 | #include "gimplify.h" |
40 | #include "gimple-iterator.h" |
41 | #include "gimplify-me.h" |
42 | #include "gimple-walk.h" |
43 | #include "tree-cfg.h" |
44 | #include "tree-into-ssa.h" |
45 | #include "tree-inline.h" |
46 | #include "demangle.h" |
47 | #include "output.h" |
48 | #include "trans-mem.h" |
49 | #include "langhooks.h" |
50 | #include "cfgloop.h" |
51 | #include "tree-ssa-address.h" |
52 | #include "stringpool.h" |
53 | #include "attribs.h" |
54 | #include "alloc-pool.h" |
55 | #include "symbol-summary.h" |
56 | #include "symtab-thunks.h" |
57 | |
58 | #define A_RUNINSTRUMENTEDCODE 0x0001 |
59 | #define A_RUNUNINSTRUMENTEDCODE 0x0002 |
60 | #define A_SAVELIVEVARIABLES 0x0004 |
61 | #define A_RESTORELIVEVARIABLES 0x0008 |
62 | #define A_ABORTTRANSACTION 0x0010 |
63 | |
64 | #define AR_USERABORT 0x0001 |
65 | #define AR_USERRETRY 0x0002 |
66 | #define AR_TMCONFLICT 0x0004 |
67 | #define AR_EXCEPTIONBLOCKABORT 0x0008 |
68 | #define AR_OUTERABORT 0x0010 |
69 | |
70 | #define MODE_SERIALIRREVOCABLE 0x0000 |
71 | |
72 | |
73 | /* The representation of a transaction changes several times during the |
74 | lowering process. In the beginning, in the front-end we have the |
75 | GENERIC tree TRANSACTION_EXPR. For example, |
76 | |
77 | __transaction { |
78 | local++; |
79 | if (++global == 10) |
80 | __tm_abort; |
81 | } |
82 | |
83 | During initial gimplification (gimplify.cc) the TRANSACTION_EXPR node is |
84 | trivially replaced with a GIMPLE_TRANSACTION node. |
85 | |
86 | During pass_lower_tm, we examine the body of transactions looking |
87 | for aborts. Transactions that do not contain an abort may be |
88 | merged into an outer transaction. We also add a TRY-FINALLY node |
89 | to arrange for the transaction to be committed on any exit. |
90 | |
91 | [??? Think about how this arrangement affects throw-with-commit |
92 | and throw-with-abort operations. In this case we want the TRY to |
93 | handle gotos, but not to catch any exceptions because the transaction |
94 | will already be closed.] |
95 | |
96 | GIMPLE_TRANSACTION [label=NULL] { |
97 | try { |
98 | local = local + 1; |
99 | t0 = global; |
100 | t1 = t0 + 1; |
101 | global = t1; |
102 | if (t1 == 10) |
103 | __builtin___tm_abort (); |
104 | } finally { |
105 | __builtin___tm_commit (); |
106 | } |
107 | } |
108 | |
109 | During pass_lower_eh, we create EH regions for the transactions, |
110 | intermixed with the regular EH stuff. This gives us a nice persistent |
111 | mapping (all the way through rtl) from transactional memory operation |
112 | back to the transaction, which allows us to get the abnormal edges |
113 | correct to model transaction aborts and restarts: |
114 | |
115 | GIMPLE_TRANSACTION [label=over] |
116 | local = local + 1; |
117 | t0 = global; |
118 | t1 = t0 + 1; |
119 | global = t1; |
120 | if (t1 == 10) |
121 | __builtin___tm_abort (); |
122 | __builtin___tm_commit (); |
123 | over: |
124 | |
125 | This is the end of all_lowering_passes, and so is what is present |
126 | during the IPA passes, and through all of the optimization passes. |
127 | |
128 | During pass_ipa_tm, we examine all GIMPLE_TRANSACTION blocks in all |
129 | functions and mark functions for cloning. |
130 | |
131 | At the end of gimple optimization, before exiting SSA form, |
132 | pass_tm_edges replaces statements that perform transactional |
133 | memory operations with the appropriate TM builtins, and swap |
134 | out function calls with their transactional clones. At this |
135 | point we introduce the abnormal transaction restart edges and |
136 | complete lowering of the GIMPLE_TRANSACTION node. |
137 | |
138 | x = __builtin___tm_start (MAY_ABORT); |
139 | eh_label: |
140 | if (x & abort_transaction) |
141 | goto over; |
142 | local = local + 1; |
143 | t0 = __builtin___tm_load (global); |
144 | t1 = t0 + 1; |
145 | __builtin___tm_store (&global, t1); |
146 | if (t1 == 10) |
147 | __builtin___tm_abort (); |
148 | __builtin___tm_commit (); |
149 | over: |
150 | */ |
151 | |
152 | static void *expand_regions (struct tm_region *, |
153 | void *(*callback)(struct tm_region *, void *), |
154 | void *, bool); |
155 | |
156 | |
157 | /* Return the attributes we want to examine for X, or NULL if it's not |
158 | something we examine. We look at function types, but allow pointers |
159 | to function types and function decls and peek through. */ |
160 | |
161 | static tree |
162 | get_attrs_for (const_tree x) |
163 | { |
164 | if (x == NULL_TREE) |
165 | return NULL_TREE; |
166 | |
167 | switch (TREE_CODE (x)) |
168 | { |
169 | case FUNCTION_DECL: |
170 | return TYPE_ATTRIBUTES (TREE_TYPE (x)); |
171 | |
172 | default: |
173 | if (TYPE_P (x)) |
174 | return NULL_TREE; |
175 | x = TREE_TYPE (x); |
176 | if (TREE_CODE (x) != POINTER_TYPE) |
177 | return NULL_TREE; |
178 | /* FALLTHRU */ |
179 | |
180 | case POINTER_TYPE: |
181 | x = TREE_TYPE (x); |
182 | if (TREE_CODE (x) != FUNCTION_TYPE && TREE_CODE (x) != METHOD_TYPE) |
183 | return NULL_TREE; |
184 | /* FALLTHRU */ |
185 | |
186 | case FUNCTION_TYPE: |
187 | case METHOD_TYPE: |
188 | return TYPE_ATTRIBUTES (x); |
189 | } |
190 | } |
191 | |
192 | /* Return true if X has been marked TM_PURE. */ |
193 | |
194 | bool |
195 | is_tm_pure (const_tree x) |
196 | { |
197 | unsigned flags; |
198 | |
199 | switch (TREE_CODE (x)) |
200 | { |
201 | case FUNCTION_DECL: |
202 | case FUNCTION_TYPE: |
203 | case METHOD_TYPE: |
204 | break; |
205 | |
206 | default: |
207 | if (TYPE_P (x)) |
208 | return false; |
209 | x = TREE_TYPE (x); |
210 | if (TREE_CODE (x) != POINTER_TYPE) |
211 | return false; |
212 | /* FALLTHRU */ |
213 | |
214 | case POINTER_TYPE: |
215 | x = TREE_TYPE (x); |
216 | if (TREE_CODE (x) != FUNCTION_TYPE && TREE_CODE (x) != METHOD_TYPE) |
217 | return false; |
218 | break; |
219 | } |
220 | |
221 | flags = flags_from_decl_or_type (x); |
222 | return (flags & ECF_TM_PURE) != 0; |
223 | } |
224 | |
225 | /* Return true if X has been marked TM_IRREVOCABLE. */ |
226 | |
227 | static bool |
228 | is_tm_irrevocable (tree x) |
229 | { |
230 | tree attrs = get_attrs_for (x); |
231 | |
232 | if (attrs && lookup_attribute (attr_name: "transaction_unsafe" , list: attrs)) |
233 | return true; |
234 | |
235 | /* A call to the irrevocable builtin is by definition, |
236 | irrevocable. */ |
237 | if (TREE_CODE (x) == ADDR_EXPR) |
238 | x = TREE_OPERAND (x, 0); |
239 | if (TREE_CODE (x) == FUNCTION_DECL |
240 | && fndecl_built_in_p (node: x, name1: BUILT_IN_TM_IRREVOCABLE)) |
241 | return true; |
242 | |
243 | return false; |
244 | } |
245 | |
246 | /* Return true if X has been marked TM_SAFE. */ |
247 | |
248 | bool |
249 | is_tm_safe (const_tree x) |
250 | { |
251 | if (flag_tm) |
252 | { |
253 | tree attrs = get_attrs_for (x); |
254 | if (attrs) |
255 | { |
256 | if (lookup_attribute (attr_name: "transaction_safe" , list: attrs)) |
257 | return true; |
258 | if (lookup_attribute (attr_name: "transaction_may_cancel_outer" , list: attrs)) |
259 | return true; |
260 | } |
261 | } |
262 | return false; |
263 | } |
264 | |
265 | /* Return true if CALL is const, or tm_pure. */ |
266 | |
267 | static bool |
268 | is_tm_pure_call (gimple *call) |
269 | { |
270 | return (gimple_call_flags (call) & (ECF_CONST | ECF_TM_PURE)) != 0; |
271 | } |
272 | |
273 | /* Return true if X has been marked TM_CALLABLE. */ |
274 | |
275 | static bool |
276 | is_tm_callable (tree x) |
277 | { |
278 | tree attrs = get_attrs_for (x); |
279 | if (attrs) |
280 | { |
281 | if (lookup_attribute (attr_name: "transaction_callable" , list: attrs)) |
282 | return true; |
283 | if (lookup_attribute (attr_name: "transaction_safe" , list: attrs)) |
284 | return true; |
285 | if (lookup_attribute (attr_name: "transaction_may_cancel_outer" , list: attrs)) |
286 | return true; |
287 | } |
288 | return false; |
289 | } |
290 | |
291 | /* Return true if X has been marked TRANSACTION_MAY_CANCEL_OUTER. */ |
292 | |
293 | bool |
294 | is_tm_may_cancel_outer (tree x) |
295 | { |
296 | tree attrs = get_attrs_for (x); |
297 | if (attrs) |
298 | return lookup_attribute (attr_name: "transaction_may_cancel_outer" , list: attrs) != NULL; |
299 | return false; |
300 | } |
301 | |
302 | /* Return true for built in functions that "end" a transaction. */ |
303 | |
304 | bool |
305 | is_tm_ending_fndecl (tree fndecl) |
306 | { |
307 | if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) |
308 | switch (DECL_FUNCTION_CODE (decl: fndecl)) |
309 | { |
310 | case BUILT_IN_TM_COMMIT: |
311 | case BUILT_IN_TM_COMMIT_EH: |
312 | case BUILT_IN_TM_ABORT: |
313 | case BUILT_IN_TM_IRREVOCABLE: |
314 | return true; |
315 | default: |
316 | break; |
317 | } |
318 | |
319 | return false; |
320 | } |
321 | |
322 | /* Return true if STMT is a built in function call that "ends" a |
323 | transaction. */ |
324 | |
325 | bool |
326 | is_tm_ending (gimple *stmt) |
327 | { |
328 | tree fndecl; |
329 | |
330 | if (gimple_code (g: stmt) != GIMPLE_CALL) |
331 | return false; |
332 | |
333 | fndecl = gimple_call_fndecl (gs: stmt); |
334 | return (fndecl != NULL_TREE |
335 | && is_tm_ending_fndecl (fndecl)); |
336 | } |
337 | |
338 | /* Return true if STMT is a TM load. */ |
339 | |
340 | static bool |
341 | is_tm_load (gimple *stmt) |
342 | { |
343 | tree fndecl; |
344 | |
345 | if (gimple_code (g: stmt) != GIMPLE_CALL) |
346 | return false; |
347 | |
348 | fndecl = gimple_call_fndecl (gs: stmt); |
349 | return (fndecl |
350 | && fndecl_built_in_p (node: fndecl, klass: BUILT_IN_NORMAL) |
351 | && BUILTIN_TM_LOAD_P (DECL_FUNCTION_CODE (fndecl))); |
352 | } |
353 | |
354 | /* Same as above, but for simple TM loads, that is, not the |
355 | after-write, after-read, etc optimized variants. */ |
356 | |
357 | static bool |
358 | is_tm_simple_load (gimple *stmt) |
359 | { |
360 | tree fndecl; |
361 | |
362 | if (gimple_code (g: stmt) != GIMPLE_CALL) |
363 | return false; |
364 | |
365 | fndecl = gimple_call_fndecl (gs: stmt); |
366 | if (fndecl && fndecl_built_in_p (node: fndecl, klass: BUILT_IN_NORMAL)) |
367 | { |
368 | enum built_in_function fcode = DECL_FUNCTION_CODE (decl: fndecl); |
369 | return (fcode == BUILT_IN_TM_LOAD_1 |
370 | || fcode == BUILT_IN_TM_LOAD_2 |
371 | || fcode == BUILT_IN_TM_LOAD_4 |
372 | || fcode == BUILT_IN_TM_LOAD_8 |
373 | || fcode == BUILT_IN_TM_LOAD_FLOAT |
374 | || fcode == BUILT_IN_TM_LOAD_DOUBLE |
375 | || fcode == BUILT_IN_TM_LOAD_LDOUBLE |
376 | || fcode == BUILT_IN_TM_LOAD_M64 |
377 | || fcode == BUILT_IN_TM_LOAD_M128 |
378 | || fcode == BUILT_IN_TM_LOAD_M256); |
379 | } |
380 | return false; |
381 | } |
382 | |
383 | /* Return true if STMT is a TM store. */ |
384 | |
385 | static bool |
386 | is_tm_store (gimple *stmt) |
387 | { |
388 | tree fndecl; |
389 | |
390 | if (gimple_code (g: stmt) != GIMPLE_CALL) |
391 | return false; |
392 | |
393 | fndecl = gimple_call_fndecl (gs: stmt); |
394 | return (fndecl |
395 | && fndecl_built_in_p (node: fndecl, klass: BUILT_IN_NORMAL) |
396 | && BUILTIN_TM_STORE_P (DECL_FUNCTION_CODE (fndecl))); |
397 | } |
398 | |
399 | /* Same as above, but for simple TM stores, that is, not the |
400 | after-write, after-read, etc optimized variants. */ |
401 | |
402 | static bool |
403 | is_tm_simple_store (gimple *stmt) |
404 | { |
405 | tree fndecl; |
406 | |
407 | if (gimple_code (g: stmt) != GIMPLE_CALL) |
408 | return false; |
409 | |
410 | fndecl = gimple_call_fndecl (gs: stmt); |
411 | if (fndecl |
412 | && fndecl_built_in_p (node: fndecl, klass: BUILT_IN_NORMAL)) |
413 | { |
414 | enum built_in_function fcode = DECL_FUNCTION_CODE (decl: fndecl); |
415 | return (fcode == BUILT_IN_TM_STORE_1 |
416 | || fcode == BUILT_IN_TM_STORE_2 |
417 | || fcode == BUILT_IN_TM_STORE_4 |
418 | || fcode == BUILT_IN_TM_STORE_8 |
419 | || fcode == BUILT_IN_TM_STORE_FLOAT |
420 | || fcode == BUILT_IN_TM_STORE_DOUBLE |
421 | || fcode == BUILT_IN_TM_STORE_LDOUBLE |
422 | || fcode == BUILT_IN_TM_STORE_M64 |
423 | || fcode == BUILT_IN_TM_STORE_M128 |
424 | || fcode == BUILT_IN_TM_STORE_M256); |
425 | } |
426 | return false; |
427 | } |
428 | |
429 | /* Return true if FNDECL is BUILT_IN_TM_ABORT. */ |
430 | |
431 | static bool |
432 | is_tm_abort (tree fndecl) |
433 | { |
434 | return (fndecl && fndecl_built_in_p (node: fndecl, name1: BUILT_IN_TM_ABORT)); |
435 | } |
436 | |
437 | /* Build a GENERIC tree for a user abort. This is called by front ends |
438 | while transforming the __tm_abort statement. */ |
439 | |
440 | tree |
441 | build_tm_abort_call (location_t loc, bool is_outer) |
442 | { |
443 | return build_call_expr_loc (loc, builtin_decl_explicit (fncode: BUILT_IN_TM_ABORT), 1, |
444 | build_int_cst (integer_type_node, |
445 | AR_USERABORT |
446 | | (is_outer ? AR_OUTERABORT : 0))); |
447 | } |
448 | |
449 | /* Map for arbitrary function replacement under TM, as created |
450 | by the tm_wrap attribute. */ |
451 | |
452 | struct tm_wrapper_hasher : ggc_cache_ptr_hash<tree_map> |
453 | { |
454 | static inline hashval_t hash (tree_map *m) { return m->hash; } |
455 | static inline bool |
456 | equal (tree_map *a, tree_map *b) |
457 | { |
458 | return a->base.from == b->base.from; |
459 | } |
460 | |
461 | static int |
462 | keep_cache_entry (tree_map *&m) |
463 | { |
464 | return ggc_marked_p (m->base.from); |
465 | } |
466 | }; |
467 | |
468 | static GTY((cache)) hash_table<tm_wrapper_hasher> *tm_wrap_map; |
469 | |
470 | void |
471 | record_tm_replacement (tree from, tree to) |
472 | { |
473 | struct tree_map **slot, *h; |
474 | |
475 | /* Do not inline wrapper functions that will get replaced in the TM |
476 | pass. |
477 | |
478 | Suppose you have foo() that will get replaced into tmfoo(). Make |
479 | sure the inliner doesn't try to outsmart us and inline foo() |
480 | before we get a chance to do the TM replacement. */ |
481 | DECL_UNINLINABLE (from) = 1; |
482 | |
483 | if (tm_wrap_map == NULL) |
484 | tm_wrap_map = hash_table<tm_wrapper_hasher>::create_ggc (n: 32); |
485 | |
486 | h = ggc_alloc<tree_map> (); |
487 | h->hash = htab_hash_pointer (from); |
488 | h->base.from = from; |
489 | h->to = to; |
490 | |
491 | slot = tm_wrap_map->find_slot_with_hash (comparable: h, hash: h->hash, insert: INSERT); |
492 | *slot = h; |
493 | } |
494 | |
495 | /* Return a TM-aware replacement function for DECL. */ |
496 | |
497 | static tree |
498 | find_tm_replacement_function (tree fndecl) |
499 | { |
500 | if (tm_wrap_map) |
501 | { |
502 | struct tree_map *h, in; |
503 | |
504 | in.base.from = fndecl; |
505 | in.hash = htab_hash_pointer (fndecl); |
506 | h = tm_wrap_map->find_with_hash (comparable: &in, hash: in.hash); |
507 | if (h) |
508 | return h->to; |
509 | } |
510 | |
511 | /* ??? We may well want TM versions of most of the common <string.h> |
512 | functions. For now, we've already these two defined. */ |
513 | /* Adjust expand_call_tm() attributes as necessary for the cases |
514 | handled here: */ |
515 | if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) |
516 | switch (DECL_FUNCTION_CODE (decl: fndecl)) |
517 | { |
518 | case BUILT_IN_MEMCPY: |
519 | return builtin_decl_explicit (fncode: BUILT_IN_TM_MEMCPY); |
520 | case BUILT_IN_MEMMOVE: |
521 | return builtin_decl_explicit (fncode: BUILT_IN_TM_MEMMOVE); |
522 | case BUILT_IN_MEMSET: |
523 | return builtin_decl_explicit (fncode: BUILT_IN_TM_MEMSET); |
524 | default: |
525 | return NULL; |
526 | } |
527 | |
528 | return NULL; |
529 | } |
530 | |
531 | /* When appropriate, record TM replacement for memory allocation functions. |
532 | |
533 | FROM is the FNDECL to wrap. */ |
534 | void |
535 | tm_malloc_replacement (tree from) |
536 | { |
537 | const char *str; |
538 | tree to; |
539 | |
540 | if (TREE_CODE (from) != FUNCTION_DECL) |
541 | return; |
542 | |
543 | /* If we have a previous replacement, the user must be explicitly |
544 | wrapping malloc/calloc/free. They better know what they're |
545 | doing... */ |
546 | if (find_tm_replacement_function (fndecl: from)) |
547 | return; |
548 | |
549 | str = IDENTIFIER_POINTER (DECL_NAME (from)); |
550 | |
551 | if (!strcmp (s1: str, s2: "malloc" )) |
552 | to = builtin_decl_explicit (fncode: BUILT_IN_TM_MALLOC); |
553 | else if (!strcmp (s1: str, s2: "calloc" )) |
554 | to = builtin_decl_explicit (fncode: BUILT_IN_TM_CALLOC); |
555 | else if (!strcmp (s1: str, s2: "free" )) |
556 | to = builtin_decl_explicit (fncode: BUILT_IN_TM_FREE); |
557 | else |
558 | return; |
559 | |
560 | TREE_NOTHROW (to) = 0; |
561 | |
562 | record_tm_replacement (from, to); |
563 | } |
564 | |
565 | /* Diagnostics for tm_safe functions/regions. Called by the front end |
566 | once we've lowered the function to high-gimple. */ |
567 | |
568 | /* Subroutine of diagnose_tm_safe_errors, called through walk_gimple_seq. |
569 | Process exactly one statement. WI->INFO is set to non-null when in |
570 | the context of a tm_safe function, and null for a __transaction block. */ |
571 | |
572 | #define DIAG_TM_OUTER 1 |
573 | #define DIAG_TM_SAFE 2 |
574 | #define DIAG_TM_RELAXED 4 |
575 | |
576 | struct diagnose_tm |
577 | { |
578 | unsigned int summary_flags : 8; |
579 | unsigned int block_flags : 8; |
580 | unsigned int func_flags : 8; |
581 | unsigned int saw_volatile : 1; |
582 | gimple *stmt; |
583 | }; |
584 | |
585 | /* Return true if T is a volatile lvalue of some kind. */ |
586 | |
587 | static bool |
588 | volatile_lvalue_p (tree t) |
589 | { |
590 | return ((SSA_VAR_P (t) || REFERENCE_CLASS_P (t)) |
591 | && TREE_THIS_VOLATILE (TREE_TYPE (t))); |
592 | } |
593 | |
594 | /* Tree callback function for diagnose_tm pass. */ |
595 | |
596 | static tree |
597 | diagnose_tm_1_op (tree *tp, int *walk_subtrees, void *data) |
598 | { |
599 | struct walk_stmt_info *wi = (struct walk_stmt_info *) data; |
600 | struct diagnose_tm *d = (struct diagnose_tm *) wi->info; |
601 | |
602 | if (TYPE_P (*tp)) |
603 | *walk_subtrees = false; |
604 | else if (volatile_lvalue_p (t: *tp) |
605 | && !d->saw_volatile) |
606 | { |
607 | d->saw_volatile = 1; |
608 | if (d->block_flags & DIAG_TM_SAFE) |
609 | error_at (gimple_location (g: d->stmt), |
610 | "invalid use of volatile lvalue inside transaction" ); |
611 | else if (d->func_flags & DIAG_TM_SAFE) |
612 | error_at (gimple_location (g: d->stmt), |
613 | "invalid use of volatile lvalue inside %<transaction_safe%> " |
614 | "function" ); |
615 | } |
616 | |
617 | return NULL_TREE; |
618 | } |
619 | |
620 | static inline bool |
621 | is_tm_safe_or_pure (const_tree x) |
622 | { |
623 | return is_tm_safe (x) || is_tm_pure (x); |
624 | } |
625 | |
626 | static tree |
627 | diagnose_tm_1 (gimple_stmt_iterator *gsi, bool *handled_ops_p, |
628 | struct walk_stmt_info *wi) |
629 | { |
630 | gimple *stmt = gsi_stmt (i: *gsi); |
631 | struct diagnose_tm *d = (struct diagnose_tm *) wi->info; |
632 | |
633 | /* Save stmt for use in leaf analysis. */ |
634 | d->stmt = stmt; |
635 | |
636 | switch (gimple_code (g: stmt)) |
637 | { |
638 | case GIMPLE_CALL: |
639 | { |
640 | if (gimple_call_internal_p (gs: stmt)) |
641 | break; |
642 | |
643 | tree fn = gimple_call_fn (gs: stmt); |
644 | |
645 | if ((d->summary_flags & DIAG_TM_OUTER) == 0 |
646 | && is_tm_may_cancel_outer (x: fn)) |
647 | error_at (gimple_location (g: stmt), |
648 | "%<transaction_may_cancel_outer%> function call not within" |
649 | " outer transaction or %<transaction_may_cancel_outer%>" ); |
650 | |
651 | if (d->summary_flags & DIAG_TM_SAFE) |
652 | { |
653 | bool is_safe, direct_call_p; |
654 | tree replacement; |
655 | |
656 | if (TREE_CODE (fn) == ADDR_EXPR |
657 | && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL) |
658 | { |
659 | direct_call_p = true; |
660 | replacement = TREE_OPERAND (fn, 0); |
661 | replacement = find_tm_replacement_function (fndecl: replacement); |
662 | if (replacement) |
663 | fn = replacement; |
664 | } |
665 | else |
666 | { |
667 | direct_call_p = false; |
668 | replacement = NULL_TREE; |
669 | } |
670 | |
671 | if (is_tm_safe_or_pure (x: fn)) |
672 | is_safe = true; |
673 | else if (is_tm_callable (x: fn) || is_tm_irrevocable (x: fn)) |
674 | { |
675 | /* A function explicitly marked transaction_callable as |
676 | opposed to transaction_safe is being defined to be |
677 | unsafe as part of its ABI, regardless of its contents. */ |
678 | is_safe = false; |
679 | } |
680 | else if (direct_call_p) |
681 | { |
682 | if (IS_TYPE_OR_DECL_P (fn) |
683 | && flags_from_decl_or_type (fn) & ECF_TM_BUILTIN) |
684 | is_safe = true; |
685 | else if (replacement) |
686 | { |
687 | /* ??? At present we've been considering replacements |
688 | merely transaction_callable, and therefore might |
689 | enter irrevocable. The tm_wrap attribute has not |
690 | yet made it into the new language spec. */ |
691 | is_safe = false; |
692 | } |
693 | else |
694 | { |
695 | /* ??? Diagnostics for unmarked direct calls moved into |
696 | the IPA pass. Section 3.2 of the spec details how |
697 | functions not marked should be considered "implicitly |
698 | safe" based on having examined the function body. */ |
699 | is_safe = true; |
700 | } |
701 | } |
702 | else |
703 | { |
704 | /* An unmarked indirect call. Consider it unsafe even |
705 | though optimization may yet figure out how to inline. */ |
706 | is_safe = false; |
707 | } |
708 | |
709 | if (!is_safe) |
710 | { |
711 | if (TREE_CODE (fn) == ADDR_EXPR) |
712 | fn = TREE_OPERAND (fn, 0); |
713 | if (d->block_flags & DIAG_TM_SAFE) |
714 | { |
715 | if (direct_call_p) |
716 | error_at (gimple_location (g: stmt), |
717 | "unsafe function call %qD within " |
718 | "atomic transaction" , fn); |
719 | else |
720 | { |
721 | if ((!DECL_P (fn) || DECL_NAME (fn)) |
722 | && TREE_CODE (fn) != SSA_NAME) |
723 | error_at (gimple_location (g: stmt), |
724 | "unsafe function call %qE within " |
725 | "atomic transaction" , fn); |
726 | else |
727 | error_at (gimple_location (g: stmt), |
728 | "unsafe indirect function call within " |
729 | "atomic transaction" ); |
730 | } |
731 | } |
732 | else |
733 | { |
734 | if (direct_call_p) |
735 | error_at (gimple_location (g: stmt), |
736 | "unsafe function call %qD within " |
737 | "%<transaction_safe%> function" , fn); |
738 | else |
739 | { |
740 | if ((!DECL_P (fn) || DECL_NAME (fn)) |
741 | && TREE_CODE (fn) != SSA_NAME) |
742 | error_at (gimple_location (g: stmt), |
743 | "unsafe function call %qE within " |
744 | "%<transaction_safe%> function" , fn); |
745 | else |
746 | error_at (gimple_location (g: stmt), |
747 | "unsafe indirect function call within " |
748 | "%<transaction_safe%> function" ); |
749 | } |
750 | } |
751 | } |
752 | } |
753 | } |
754 | break; |
755 | |
756 | case GIMPLE_ASM: |
757 | /* ??? We ought to come up with a way to add attributes to |
758 | asm statements, and then add "transaction_safe" to it. |
759 | Either that or get the language spec to resurrect __tm_waiver. */ |
760 | if (d->block_flags & DIAG_TM_SAFE) |
761 | error_at (gimple_location (g: stmt), |
762 | "%<asm%> not allowed in atomic transaction" ); |
763 | else if (d->func_flags & DIAG_TM_SAFE) |
764 | error_at (gimple_location (g: stmt), |
765 | "%<asm%> not allowed in %<transaction_safe%> function" ); |
766 | break; |
767 | |
768 | case GIMPLE_TRANSACTION: |
769 | { |
770 | gtransaction *trans_stmt = as_a <gtransaction *> (p: stmt); |
771 | unsigned char inner_flags = DIAG_TM_SAFE; |
772 | |
773 | if (gimple_transaction_subcode (transaction_stmt: trans_stmt) & GTMA_IS_RELAXED) |
774 | { |
775 | if (d->block_flags & DIAG_TM_SAFE) |
776 | error_at (gimple_location (g: stmt), |
777 | "relaxed transaction in atomic transaction" ); |
778 | else if (d->func_flags & DIAG_TM_SAFE) |
779 | error_at (gimple_location (g: stmt), |
780 | "relaxed transaction in %<transaction_safe%> function" ); |
781 | inner_flags = DIAG_TM_RELAXED; |
782 | } |
783 | else if (gimple_transaction_subcode (transaction_stmt: trans_stmt) & GTMA_IS_OUTER) |
784 | { |
785 | if (d->block_flags) |
786 | error_at (gimple_location (g: stmt), |
787 | "outer transaction in transaction" ); |
788 | else if (d->func_flags & DIAG_TM_OUTER) |
789 | error_at (gimple_location (g: stmt), |
790 | "outer transaction in " |
791 | "%<transaction_may_cancel_outer%> function" ); |
792 | else if (d->func_flags & DIAG_TM_SAFE) |
793 | error_at (gimple_location (g: stmt), |
794 | "outer transaction in %<transaction_safe%> function" ); |
795 | inner_flags |= DIAG_TM_OUTER; |
796 | } |
797 | |
798 | *handled_ops_p = true; |
799 | if (gimple_transaction_body (transaction_stmt: trans_stmt)) |
800 | { |
801 | struct walk_stmt_info wi_inner; |
802 | struct diagnose_tm d_inner; |
803 | |
804 | memset (s: &d_inner, c: 0, n: sizeof (d_inner)); |
805 | d_inner.func_flags = d->func_flags; |
806 | d_inner.block_flags = d->block_flags | inner_flags; |
807 | d_inner.summary_flags = d_inner.func_flags | d_inner.block_flags; |
808 | |
809 | memset (s: &wi_inner, c: 0, n: sizeof (wi_inner)); |
810 | wi_inner.info = &d_inner; |
811 | |
812 | walk_gimple_seq (gimple_transaction_body (transaction_stmt: trans_stmt), |
813 | diagnose_tm_1, diagnose_tm_1_op, &wi_inner); |
814 | } |
815 | } |
816 | break; |
817 | |
818 | default: |
819 | break; |
820 | } |
821 | |
822 | return NULL_TREE; |
823 | } |
824 | |
825 | static unsigned int |
826 | diagnose_tm_blocks (void) |
827 | { |
828 | struct walk_stmt_info wi; |
829 | struct diagnose_tm d; |
830 | |
831 | memset (s: &d, c: 0, n: sizeof (d)); |
832 | if (is_tm_may_cancel_outer (x: current_function_decl)) |
833 | d.func_flags = DIAG_TM_OUTER | DIAG_TM_SAFE; |
834 | else if (is_tm_safe (x: current_function_decl)) |
835 | d.func_flags = DIAG_TM_SAFE; |
836 | d.summary_flags = d.func_flags; |
837 | |
838 | memset (s: &wi, c: 0, n: sizeof (wi)); |
839 | wi.info = &d; |
840 | |
841 | walk_gimple_seq (gimple_body (current_function_decl), |
842 | diagnose_tm_1, diagnose_tm_1_op, &wi); |
843 | |
844 | return 0; |
845 | } |
846 | |
847 | namespace { |
848 | |
849 | const pass_data pass_data_diagnose_tm_blocks = |
850 | { |
851 | .type: GIMPLE_PASS, /* type */ |
852 | .name: "*diagnose_tm_blocks" , /* name */ |
853 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
854 | .tv_id: TV_TRANS_MEM, /* tv_id */ |
855 | PROP_gimple_any, /* properties_required */ |
856 | .properties_provided: 0, /* properties_provided */ |
857 | .properties_destroyed: 0, /* properties_destroyed */ |
858 | .todo_flags_start: 0, /* todo_flags_start */ |
859 | .todo_flags_finish: 0, /* todo_flags_finish */ |
860 | }; |
861 | |
862 | class pass_diagnose_tm_blocks : public gimple_opt_pass |
863 | { |
864 | public: |
865 | pass_diagnose_tm_blocks (gcc::context *ctxt) |
866 | : gimple_opt_pass (pass_data_diagnose_tm_blocks, ctxt) |
867 | {} |
868 | |
869 | /* opt_pass methods: */ |
870 | bool gate (function *) final override { return flag_tm; } |
871 | unsigned int execute (function *) final override |
872 | { |
873 | return diagnose_tm_blocks (); |
874 | } |
875 | |
876 | }; // class pass_diagnose_tm_blocks |
877 | |
878 | } // anon namespace |
879 | |
880 | gimple_opt_pass * |
881 | make_pass_diagnose_tm_blocks (gcc::context *ctxt) |
882 | { |
883 | return new pass_diagnose_tm_blocks (ctxt); |
884 | } |
885 | |
886 | /* Instead of instrumenting thread private memory, we save the |
887 | addresses in a log which we later use to save/restore the addresses |
888 | upon transaction start/restart. |
889 | |
890 | The log is keyed by address, where each element contains individual |
891 | statements among different code paths that perform the store. |
892 | |
893 | This log is later used to generate either plain save/restore of the |
894 | addresses upon transaction start/restart, or calls to the ITM_L* |
895 | logging functions. |
896 | |
897 | So for something like: |
898 | |
899 | struct large { int x[1000]; }; |
900 | struct large lala = { 0 }; |
901 | __transaction { |
902 | lala.x[i] = 123; |
903 | ... |
904 | } |
905 | |
906 | We can either save/restore: |
907 | |
908 | lala = { 0 }; |
909 | trxn = _ITM_startTransaction (); |
910 | if (trxn & a_saveLiveVariables) |
911 | tmp_lala1 = lala.x[i]; |
912 | else if (a & a_restoreLiveVariables) |
913 | lala.x[i] = tmp_lala1; |
914 | |
915 | or use the logging functions: |
916 | |
917 | lala = { 0 }; |
918 | trxn = _ITM_startTransaction (); |
919 | _ITM_LU4 (&lala.x[i]); |
920 | |
921 | Obviously, if we use _ITM_L* to log, we prefer to call _ITM_L* as |
922 | far up the dominator tree to shadow all of the writes to a given |
923 | location (thus reducing the total number of logging calls), but not |
924 | so high as to be called on a path that does not perform a |
925 | write. */ |
926 | |
927 | /* One individual log entry. We may have multiple statements for the |
928 | same location if neither dominate each other (on different |
929 | execution paths). */ |
930 | struct tm_log_entry |
931 | { |
932 | /* Address to save. */ |
933 | tree addr; |
934 | /* Entry block for the transaction this address occurs in. */ |
935 | basic_block entry_block; |
936 | /* Dominating statements the store occurs in. */ |
937 | vec<gimple *> stmts; |
938 | /* Initially, while we are building the log, we place a nonzero |
939 | value here to mean that this address *will* be saved with a |
940 | save/restore sequence. Later, when generating the save sequence |
941 | we place the SSA temp generated here. */ |
942 | tree save_var; |
943 | }; |
944 | |
945 | |
946 | /* Log entry hashtable helpers. */ |
947 | |
948 | struct log_entry_hasher : pointer_hash <tm_log_entry> |
949 | { |
950 | static inline hashval_t hash (const tm_log_entry *); |
951 | static inline bool equal (const tm_log_entry *, const tm_log_entry *); |
952 | static inline void remove (tm_log_entry *); |
953 | }; |
954 | |
955 | /* Htab support. Return hash value for a `tm_log_entry'. */ |
956 | inline hashval_t |
957 | log_entry_hasher::hash (const tm_log_entry *log) |
958 | { |
959 | return iterative_hash_expr (tree: log->addr, seed: 0); |
960 | } |
961 | |
962 | /* Htab support. Return true if two log entries are the same. */ |
963 | inline bool |
964 | log_entry_hasher::equal (const tm_log_entry *log1, const tm_log_entry *log2) |
965 | { |
966 | /* FIXME: |
967 | |
968 | rth: I suggest that we get rid of the component refs etc. |
969 | I.e. resolve the reference to base + offset. |
970 | |
971 | We may need to actually finish a merge with mainline for this, |
972 | since we'd like to be presented with Richi's MEM_REF_EXPRs more |
973 | often than not. But in the meantime your tm_log_entry could save |
974 | the results of get_inner_reference. |
975 | |
976 | See: g++.dg/tm/pr46653.C |
977 | */ |
978 | |
979 | /* Special case plain equality because operand_equal_p() below will |
980 | return FALSE if the addresses are equal but they have |
981 | side-effects (e.g. a volatile address). */ |
982 | if (log1->addr == log2->addr) |
983 | return true; |
984 | |
985 | return operand_equal_p (log1->addr, log2->addr, flags: 0); |
986 | } |
987 | |
988 | /* Htab support. Free one tm_log_entry. */ |
989 | inline void |
990 | log_entry_hasher::remove (tm_log_entry *lp) |
991 | { |
992 | lp->stmts.release (); |
993 | free (ptr: lp); |
994 | } |
995 | |
996 | |
997 | /* The actual log. */ |
998 | static hash_table<log_entry_hasher> *tm_log; |
999 | |
1000 | /* Addresses to log with a save/restore sequence. These should be in |
1001 | dominator order. */ |
1002 | static vec<tree> tm_log_save_addresses; |
1003 | |
1004 | enum thread_memory_type |
1005 | { |
1006 | mem_non_local = 0, |
1007 | mem_thread_local, |
1008 | mem_transaction_local, |
1009 | mem_max |
1010 | }; |
1011 | |
1012 | struct tm_new_mem_map |
1013 | { |
1014 | /* SSA_NAME being dereferenced. */ |
1015 | tree val; |
1016 | enum thread_memory_type local_new_memory; |
1017 | }; |
1018 | |
1019 | /* Hashtable helpers. */ |
1020 | |
1021 | struct tm_mem_map_hasher : free_ptr_hash <tm_new_mem_map> |
1022 | { |
1023 | static inline hashval_t hash (const tm_new_mem_map *); |
1024 | static inline bool equal (const tm_new_mem_map *, const tm_new_mem_map *); |
1025 | }; |
1026 | |
1027 | inline hashval_t |
1028 | tm_mem_map_hasher::hash (const tm_new_mem_map *v) |
1029 | { |
1030 | return (intptr_t)v->val >> 4; |
1031 | } |
1032 | |
1033 | inline bool |
1034 | tm_mem_map_hasher::equal (const tm_new_mem_map *v, const tm_new_mem_map *c) |
1035 | { |
1036 | return v->val == c->val; |
1037 | } |
1038 | |
1039 | /* Map for an SSA_NAME originally pointing to a non aliased new piece |
1040 | of memory (malloc, alloc, etc). */ |
1041 | static hash_table<tm_mem_map_hasher> *tm_new_mem_hash; |
1042 | |
1043 | /* Initialize logging data structures. */ |
1044 | static void |
1045 | tm_log_init (void) |
1046 | { |
1047 | tm_log = new hash_table<log_entry_hasher> (10); |
1048 | tm_new_mem_hash = new hash_table<tm_mem_map_hasher> (5); |
1049 | tm_log_save_addresses.create (nelems: 5); |
1050 | } |
1051 | |
1052 | /* Free logging data structures. */ |
1053 | static void |
1054 | tm_log_delete (void) |
1055 | { |
1056 | delete tm_log; |
1057 | tm_log = NULL; |
1058 | delete tm_new_mem_hash; |
1059 | tm_new_mem_hash = NULL; |
1060 | tm_log_save_addresses.release (); |
1061 | } |
1062 | |
1063 | /* Return true if MEM is a transaction invariant memory for the TM |
1064 | region starting at REGION_ENTRY_BLOCK. */ |
1065 | static bool |
1066 | transaction_invariant_address_p (const_tree mem, basic_block region_entry_block) |
1067 | { |
1068 | if ((INDIRECT_REF_P (mem) || TREE_CODE (mem) == MEM_REF) |
1069 | && TREE_CODE (TREE_OPERAND (mem, 0)) == SSA_NAME) |
1070 | { |
1071 | basic_block def_bb; |
1072 | |
1073 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (TREE_OPERAND (mem, 0))); |
1074 | return def_bb != region_entry_block |
1075 | && dominated_by_p (CDI_DOMINATORS, region_entry_block, def_bb); |
1076 | } |
1077 | |
1078 | mem = strip_invariant_refs (mem); |
1079 | return mem && (CONSTANT_CLASS_P (mem) || decl_address_invariant_p (mem)); |
1080 | } |
1081 | |
1082 | /* Given an address ADDR in STMT, find it in the memory log or add it, |
1083 | making sure to keep only the addresses highest in the dominator |
1084 | tree. |
1085 | |
1086 | ENTRY_BLOCK is the entry_block for the transaction. |
1087 | |
1088 | If we find the address in the log, make sure it's either the same |
1089 | address, or an equivalent one that dominates ADDR. |
1090 | |
1091 | If we find the address, but neither ADDR dominates the found |
1092 | address, nor the found one dominates ADDR, we're on different |
1093 | execution paths. Add it. |
1094 | |
1095 | If known, ENTRY_BLOCK is the entry block for the region, otherwise |
1096 | NULL. */ |
1097 | static void |
1098 | tm_log_add (basic_block entry_block, tree addr, gimple *stmt) |
1099 | { |
1100 | tm_log_entry **slot; |
1101 | struct tm_log_entry l, *lp; |
1102 | |
1103 | l.addr = addr; |
1104 | slot = tm_log->find_slot (value: &l, insert: INSERT); |
1105 | if (!*slot) |
1106 | { |
1107 | tree type = TREE_TYPE (addr); |
1108 | |
1109 | lp = XNEW (struct tm_log_entry); |
1110 | lp->addr = addr; |
1111 | *slot = lp; |
1112 | |
1113 | /* Small invariant addresses can be handled as save/restores. */ |
1114 | if (entry_block |
1115 | && transaction_invariant_address_p (mem: lp->addr, region_entry_block: entry_block) |
1116 | && TYPE_SIZE_UNIT (type) != NULL |
1117 | && tree_fits_uhwi_p (TYPE_SIZE_UNIT (type)) |
1118 | && ((HOST_WIDE_INT) tree_to_uhwi (TYPE_SIZE_UNIT (type)) |
1119 | < param_tm_max_aggregate_size) |
1120 | /* We must be able to copy this type normally. I.e., no |
1121 | special constructors and the like. */ |
1122 | && !TREE_ADDRESSABLE (type)) |
1123 | { |
1124 | lp->save_var = create_tmp_reg (TREE_TYPE (lp->addr), "tm_save" ); |
1125 | lp->stmts.create (nelems: 0); |
1126 | lp->entry_block = entry_block; |
1127 | /* Save addresses separately in dominator order so we don't |
1128 | get confused by overlapping addresses in the save/restore |
1129 | sequence. */ |
1130 | tm_log_save_addresses.safe_push (obj: lp->addr); |
1131 | } |
1132 | else |
1133 | { |
1134 | /* Use the logging functions. */ |
1135 | lp->stmts.create (nelems: 5); |
1136 | lp->stmts.quick_push (obj: stmt); |
1137 | lp->save_var = NULL; |
1138 | } |
1139 | } |
1140 | else |
1141 | { |
1142 | size_t i; |
1143 | gimple *oldstmt; |
1144 | |
1145 | lp = *slot; |
1146 | |
1147 | /* If we're generating a save/restore sequence, we don't care |
1148 | about statements. */ |
1149 | if (lp->save_var) |
1150 | return; |
1151 | |
1152 | for (i = 0; lp->stmts.iterate (ix: i, ptr: &oldstmt); ++i) |
1153 | { |
1154 | if (stmt == oldstmt) |
1155 | return; |
1156 | /* We already have a store to the same address, higher up the |
1157 | dominator tree. Nothing to do. */ |
1158 | if (dominated_by_p (CDI_DOMINATORS, |
1159 | gimple_bb (g: stmt), gimple_bb (g: oldstmt))) |
1160 | return; |
1161 | /* We should be processing blocks in dominator tree order. */ |
1162 | gcc_assert (!dominated_by_p (CDI_DOMINATORS, |
1163 | gimple_bb (oldstmt), gimple_bb (stmt))); |
1164 | } |
1165 | /* Store is on a different code path. */ |
1166 | lp->stmts.safe_push (obj: stmt); |
1167 | } |
1168 | } |
1169 | |
1170 | /* Gimplify the address of a TARGET_MEM_REF. Return the SSA_NAME |
1171 | result, insert the new statements before GSI. */ |
1172 | |
1173 | static tree |
1174 | gimplify_addr (gimple_stmt_iterator *gsi, tree x) |
1175 | { |
1176 | if (TREE_CODE (x) == TARGET_MEM_REF) |
1177 | x = tree_mem_ref_addr (build_pointer_type (TREE_TYPE (x)), x); |
1178 | else |
1179 | x = build_fold_addr_expr (x); |
1180 | return force_gimple_operand_gsi (gsi, x, true, NULL, true, GSI_SAME_STMT); |
1181 | } |
1182 | |
1183 | /* Instrument one address with the logging functions. |
1184 | ADDR is the address to save. |
1185 | STMT is the statement before which to place it. */ |
1186 | static void |
1187 | tm_log_emit_stmt (tree addr, gimple *stmt) |
1188 | { |
1189 | tree type = TREE_TYPE (addr); |
1190 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); |
1191 | gimple *log; |
1192 | enum built_in_function code = BUILT_IN_TM_LOG; |
1193 | |
1194 | if (type == float_type_node) |
1195 | code = BUILT_IN_TM_LOG_FLOAT; |
1196 | else if (type == double_type_node) |
1197 | code = BUILT_IN_TM_LOG_DOUBLE; |
1198 | else if (type == long_double_type_node) |
1199 | code = BUILT_IN_TM_LOG_LDOUBLE; |
1200 | else if (TYPE_SIZE (type) != NULL |
1201 | && tree_fits_uhwi_p (TYPE_SIZE (type))) |
1202 | { |
1203 | unsigned HOST_WIDE_INT type_size = tree_to_uhwi (TYPE_SIZE (type)); |
1204 | |
1205 | if (TREE_CODE (type) == VECTOR_TYPE) |
1206 | { |
1207 | switch (type_size) |
1208 | { |
1209 | case 64: |
1210 | code = BUILT_IN_TM_LOG_M64; |
1211 | break; |
1212 | case 128: |
1213 | code = BUILT_IN_TM_LOG_M128; |
1214 | break; |
1215 | case 256: |
1216 | code = BUILT_IN_TM_LOG_M256; |
1217 | break; |
1218 | default: |
1219 | goto unhandled_vec; |
1220 | } |
1221 | if (!builtin_decl_explicit_p (fncode: code)) |
1222 | goto unhandled_vec; |
1223 | } |
1224 | else |
1225 | { |
1226 | unhandled_vec: |
1227 | switch (type_size) |
1228 | { |
1229 | case 8: |
1230 | code = BUILT_IN_TM_LOG_1; |
1231 | break; |
1232 | case 16: |
1233 | code = BUILT_IN_TM_LOG_2; |
1234 | break; |
1235 | case 32: |
1236 | code = BUILT_IN_TM_LOG_4; |
1237 | break; |
1238 | case 64: |
1239 | code = BUILT_IN_TM_LOG_8; |
1240 | break; |
1241 | } |
1242 | } |
1243 | } |
1244 | |
1245 | if (code != BUILT_IN_TM_LOG && !builtin_decl_explicit_p (fncode: code)) |
1246 | code = BUILT_IN_TM_LOG; |
1247 | tree decl = builtin_decl_explicit (fncode: code); |
1248 | |
1249 | addr = gimplify_addr (gsi: &gsi, x: addr); |
1250 | if (code == BUILT_IN_TM_LOG) |
1251 | log = gimple_build_call (decl, 2, addr, TYPE_SIZE_UNIT (type)); |
1252 | else |
1253 | log = gimple_build_call (decl, 1, addr); |
1254 | gsi_insert_before (&gsi, log, GSI_SAME_STMT); |
1255 | } |
1256 | |
1257 | /* Go through the log and instrument address that must be instrumented |
1258 | with the logging functions. Leave the save/restore addresses for |
1259 | later. */ |
1260 | static void |
1261 | tm_log_emit (void) |
1262 | { |
1263 | hash_table<log_entry_hasher>::iterator hi; |
1264 | struct tm_log_entry *lp; |
1265 | |
1266 | FOR_EACH_HASH_TABLE_ELEMENT (*tm_log, lp, tm_log_entry_t, hi) |
1267 | { |
1268 | size_t i; |
1269 | gimple *stmt; |
1270 | |
1271 | if (dump_file) |
1272 | { |
1273 | fprintf (stream: dump_file, format: "TM thread private mem logging: " ); |
1274 | print_generic_expr (dump_file, lp->addr); |
1275 | fprintf (stream: dump_file, format: "\n" ); |
1276 | } |
1277 | |
1278 | if (lp->save_var) |
1279 | { |
1280 | if (dump_file) |
1281 | fprintf (stream: dump_file, format: "DUMPING to variable\n" ); |
1282 | continue; |
1283 | } |
1284 | else |
1285 | { |
1286 | if (dump_file) |
1287 | fprintf (stream: dump_file, format: "DUMPING with logging functions\n" ); |
1288 | for (i = 0; lp->stmts.iterate (ix: i, ptr: &stmt); ++i) |
1289 | tm_log_emit_stmt (addr: lp->addr, stmt); |
1290 | } |
1291 | } |
1292 | } |
1293 | |
1294 | /* Emit the save sequence for the corresponding addresses in the log. |
1295 | ENTRY_BLOCK is the entry block for the transaction. |
1296 | BB is the basic block to insert the code in. */ |
1297 | static void |
1298 | tm_log_emit_saves (basic_block entry_block, basic_block bb) |
1299 | { |
1300 | size_t i; |
1301 | gimple_stmt_iterator gsi = gsi_last_bb (bb); |
1302 | gimple *stmt; |
1303 | struct tm_log_entry l, *lp; |
1304 | |
1305 | for (i = 0; i < tm_log_save_addresses.length (); ++i) |
1306 | { |
1307 | l.addr = tm_log_save_addresses[i]; |
1308 | lp = *(tm_log->find_slot (value: &l, insert: NO_INSERT)); |
1309 | gcc_assert (lp->save_var != NULL); |
1310 | |
1311 | /* We only care about variables in the current transaction. */ |
1312 | if (lp->entry_block != entry_block) |
1313 | continue; |
1314 | |
1315 | stmt = gimple_build_assign (lp->save_var, unshare_expr (lp->addr)); |
1316 | |
1317 | /* Make sure we can create an SSA_NAME for this type. For |
1318 | instance, aggregates aren't allowed, in which case the system |
1319 | will create a VOP for us and everything will just work. */ |
1320 | if (is_gimple_reg_type (TREE_TYPE (lp->save_var))) |
1321 | { |
1322 | lp->save_var = make_ssa_name (var: lp->save_var, stmt); |
1323 | gimple_assign_set_lhs (gs: stmt, lhs: lp->save_var); |
1324 | } |
1325 | |
1326 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); |
1327 | } |
1328 | } |
1329 | |
1330 | /* Emit the restore sequence for the corresponding addresses in the log. |
1331 | ENTRY_BLOCK is the entry block for the transaction. |
1332 | BB is the basic block to insert the code in. */ |
1333 | static void |
1334 | tm_log_emit_restores (basic_block entry_block, basic_block bb) |
1335 | { |
1336 | int i; |
1337 | struct tm_log_entry l, *lp; |
1338 | gimple_stmt_iterator gsi; |
1339 | gimple *stmt; |
1340 | |
1341 | for (i = tm_log_save_addresses.length () - 1; i >= 0; i--) |
1342 | { |
1343 | l.addr = tm_log_save_addresses[i]; |
1344 | lp = *(tm_log->find_slot (value: &l, insert: NO_INSERT)); |
1345 | gcc_assert (lp->save_var != NULL); |
1346 | |
1347 | /* We only care about variables in the current transaction. */ |
1348 | if (lp->entry_block != entry_block) |
1349 | continue; |
1350 | |
1351 | /* Restores are in LIFO order from the saves in case we have |
1352 | overlaps. */ |
1353 | gsi = gsi_start_bb (bb); |
1354 | |
1355 | stmt = gimple_build_assign (unshare_expr (lp->addr), lp->save_var); |
1356 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
1357 | } |
1358 | } |
1359 | |
1360 | |
1361 | static tree lower_sequence_tm (gimple_stmt_iterator *, bool *, |
1362 | struct walk_stmt_info *); |
1363 | static tree lower_sequence_no_tm (gimple_stmt_iterator *, bool *, |
1364 | struct walk_stmt_info *); |
1365 | |
1366 | /* Evaluate an address X being dereferenced and determine if it |
1367 | originally points to a non aliased new chunk of memory (malloc, |
1368 | alloca, etc). |
1369 | |
1370 | Return MEM_THREAD_LOCAL if it points to a thread-local address. |
1371 | Return MEM_TRANSACTION_LOCAL if it points to a transaction-local address. |
1372 | Return MEM_NON_LOCAL otherwise. |
1373 | |
1374 | ENTRY_BLOCK is the entry block to the transaction containing the |
1375 | dereference of X. */ |
1376 | static enum thread_memory_type |
1377 | thread_private_new_memory (basic_block entry_block, tree x) |
1378 | { |
1379 | gimple *stmt = NULL; |
1380 | enum tree_code code; |
1381 | tm_new_mem_map **slot; |
1382 | tm_new_mem_map elt, *elt_p; |
1383 | tree val = x; |
1384 | enum thread_memory_type retval = mem_transaction_local; |
1385 | |
1386 | if (!entry_block |
1387 | || TREE_CODE (x) != SSA_NAME |
1388 | /* Possible uninitialized use, or a function argument. In |
1389 | either case, we don't care. */ |
1390 | || SSA_NAME_IS_DEFAULT_DEF (x)) |
1391 | return mem_non_local; |
1392 | |
1393 | /* Look in cache first. */ |
1394 | elt.val = x; |
1395 | slot = tm_new_mem_hash->find_slot (value: &elt, insert: INSERT); |
1396 | elt_p = *slot; |
1397 | if (elt_p) |
1398 | return elt_p->local_new_memory; |
1399 | |
1400 | /* Optimistically assume the memory is transaction local during |
1401 | processing. This catches recursion into this variable. */ |
1402 | *slot = elt_p = XNEW (tm_new_mem_map); |
1403 | elt_p->val = val; |
1404 | elt_p->local_new_memory = mem_transaction_local; |
1405 | |
1406 | /* Search DEF chain to find the original definition of this address. */ |
1407 | do |
1408 | { |
1409 | if (ptr_deref_may_alias_global_p (x, true)) |
1410 | { |
1411 | /* Address escapes. This is not thread-private. */ |
1412 | retval = mem_non_local; |
1413 | goto new_memory_ret; |
1414 | } |
1415 | |
1416 | stmt = SSA_NAME_DEF_STMT (x); |
1417 | |
1418 | /* If the malloc call is outside the transaction, this is |
1419 | thread-local. */ |
1420 | if (retval != mem_thread_local |
1421 | && !dominated_by_p (CDI_DOMINATORS, gimple_bb (g: stmt), entry_block)) |
1422 | retval = mem_thread_local; |
1423 | |
1424 | if (is_gimple_assign (gs: stmt)) |
1425 | { |
1426 | code = gimple_assign_rhs_code (gs: stmt); |
1427 | /* x = foo ==> foo */ |
1428 | if (code == SSA_NAME) |
1429 | x = gimple_assign_rhs1 (gs: stmt); |
1430 | /* x = foo + n ==> foo */ |
1431 | else if (code == POINTER_PLUS_EXPR) |
1432 | x = gimple_assign_rhs1 (gs: stmt); |
1433 | /* x = (cast*) foo ==> foo */ |
1434 | else if (code == VIEW_CONVERT_EXPR || CONVERT_EXPR_CODE_P (code)) |
1435 | x = gimple_assign_rhs1 (gs: stmt); |
1436 | /* x = c ? op1 : op2 == > op1 or op2 just like a PHI */ |
1437 | else if (code == COND_EXPR) |
1438 | { |
1439 | tree op1 = gimple_assign_rhs2 (gs: stmt); |
1440 | tree op2 = gimple_assign_rhs3 (gs: stmt); |
1441 | enum thread_memory_type mem; |
1442 | retval = thread_private_new_memory (entry_block, x: op1); |
1443 | if (retval == mem_non_local) |
1444 | goto new_memory_ret; |
1445 | mem = thread_private_new_memory (entry_block, x: op2); |
1446 | retval = MIN (retval, mem); |
1447 | goto new_memory_ret; |
1448 | } |
1449 | else |
1450 | { |
1451 | retval = mem_non_local; |
1452 | goto new_memory_ret; |
1453 | } |
1454 | } |
1455 | else |
1456 | { |
1457 | if (gimple_code (g: stmt) == GIMPLE_PHI) |
1458 | { |
1459 | unsigned int i; |
1460 | enum thread_memory_type mem; |
1461 | tree phi_result = gimple_phi_result (gs: stmt); |
1462 | |
1463 | /* If any of the ancestors are non-local, we are sure to |
1464 | be non-local. Otherwise we can avoid doing anything |
1465 | and inherit what has already been generated. */ |
1466 | retval = mem_max; |
1467 | for (i = 0; i < gimple_phi_num_args (gs: stmt); ++i) |
1468 | { |
1469 | tree op = PHI_ARG_DEF (stmt, i); |
1470 | |
1471 | /* Exclude self-assignment. */ |
1472 | if (phi_result == op) |
1473 | continue; |
1474 | |
1475 | mem = thread_private_new_memory (entry_block, x: op); |
1476 | if (mem == mem_non_local) |
1477 | { |
1478 | retval = mem; |
1479 | goto new_memory_ret; |
1480 | } |
1481 | retval = MIN (retval, mem); |
1482 | } |
1483 | goto new_memory_ret; |
1484 | } |
1485 | break; |
1486 | } |
1487 | } |
1488 | while (TREE_CODE (x) == SSA_NAME); |
1489 | |
1490 | if (stmt && is_gimple_call (gs: stmt) && gimple_call_flags (stmt) & ECF_MALLOC) |
1491 | /* Thread-local or transaction-local. */ |
1492 | ; |
1493 | else |
1494 | retval = mem_non_local; |
1495 | |
1496 | new_memory_ret: |
1497 | elt_p->local_new_memory = retval; |
1498 | return retval; |
1499 | } |
1500 | |
1501 | /* Determine whether X has to be instrumented using a read |
1502 | or write barrier. |
1503 | |
1504 | ENTRY_BLOCK is the entry block for the region where stmt resides |
1505 | in. NULL if unknown. |
1506 | |
1507 | STMT is the statement in which X occurs in. It is used for thread |
1508 | private memory instrumentation. If no TPM instrumentation is |
1509 | desired, STMT should be null. */ |
1510 | static bool |
1511 | requires_barrier (basic_block entry_block, tree x, gimple *stmt) |
1512 | { |
1513 | tree orig = x; |
1514 | while (handled_component_p (t: x)) |
1515 | x = TREE_OPERAND (x, 0); |
1516 | |
1517 | switch (TREE_CODE (x)) |
1518 | { |
1519 | case INDIRECT_REF: |
1520 | case MEM_REF: |
1521 | { |
1522 | enum thread_memory_type ret; |
1523 | |
1524 | ret = thread_private_new_memory (entry_block, TREE_OPERAND (x, 0)); |
1525 | if (ret == mem_non_local) |
1526 | return true; |
1527 | if (stmt && ret == mem_thread_local) |
1528 | /* ?? Should we pass `orig', or the INDIRECT_REF X. ?? */ |
1529 | tm_log_add (entry_block, addr: orig, stmt); |
1530 | |
1531 | /* Transaction-locals require nothing at all. For malloc, a |
1532 | transaction restart frees the memory and we reallocate. |
1533 | For alloca, the stack pointer gets reset by the retry and |
1534 | we reallocate. */ |
1535 | return false; |
1536 | } |
1537 | |
1538 | case TARGET_MEM_REF: |
1539 | if (TREE_CODE (TMR_BASE (x)) != ADDR_EXPR) |
1540 | return true; |
1541 | x = TREE_OPERAND (TMR_BASE (x), 0); |
1542 | if (TREE_CODE (x) == PARM_DECL) |
1543 | return false; |
1544 | gcc_assert (VAR_P (x)); |
1545 | /* FALLTHRU */ |
1546 | |
1547 | case PARM_DECL: |
1548 | case RESULT_DECL: |
1549 | case VAR_DECL: |
1550 | if (DECL_BY_REFERENCE (x)) |
1551 | { |
1552 | /* ??? This value is a pointer, but aggregate_value_p has been |
1553 | jigged to return true which confuses needs_to_live_in_memory. |
1554 | This ought to be cleaned up generically. |
1555 | |
1556 | FIXME: Verify this still happens after the next mainline |
1557 | merge. Testcase ie g++.dg/tm/pr47554.C. |
1558 | */ |
1559 | return false; |
1560 | } |
1561 | |
1562 | if (is_global_var (t: x)) |
1563 | return !TREE_READONLY (x); |
1564 | if (/* FIXME: This condition should actually go below in the |
1565 | tm_log_add() call, however is_call_clobbered() depends on |
1566 | aliasing info which is not available during |
1567 | gimplification. Since requires_barrier() gets called |
1568 | during lower_sequence_tm/gimplification, leave the call |
1569 | to needs_to_live_in_memory until we eliminate |
1570 | lower_sequence_tm altogether. */ |
1571 | needs_to_live_in_memory (x)) |
1572 | return true; |
1573 | else |
1574 | { |
1575 | /* For local memory that doesn't escape (aka thread private |
1576 | memory), we can either save the value at the beginning of |
1577 | the transaction and restore on restart, or call a tm |
1578 | function to dynamically save and restore on restart |
1579 | (ITM_L*). */ |
1580 | if (stmt) |
1581 | tm_log_add (entry_block, addr: orig, stmt); |
1582 | return false; |
1583 | } |
1584 | |
1585 | default: |
1586 | return false; |
1587 | } |
1588 | } |
1589 | |
1590 | /* Mark the GIMPLE_ASSIGN statement as appropriate for being inside |
1591 | a transaction region. */ |
1592 | |
1593 | static void |
1594 | examine_assign_tm (unsigned *state, gimple_stmt_iterator *gsi) |
1595 | { |
1596 | gimple *stmt = gsi_stmt (i: *gsi); |
1597 | |
1598 | if (requires_barrier (/*entry_block=*/NULL, x: gimple_assign_rhs1 (gs: stmt), NULL)) |
1599 | *state |= GTMA_HAVE_LOAD; |
1600 | if (requires_barrier (/*entry_block=*/NULL, x: gimple_assign_lhs (gs: stmt), NULL)) |
1601 | *state |= GTMA_HAVE_STORE; |
1602 | } |
1603 | |
1604 | /* Mark a GIMPLE_CALL as appropriate for being inside a transaction. */ |
1605 | |
1606 | static void |
1607 | examine_call_tm (unsigned *state, gimple_stmt_iterator *gsi) |
1608 | { |
1609 | gimple *stmt = gsi_stmt (i: *gsi); |
1610 | tree fn; |
1611 | |
1612 | if (is_tm_pure_call (call: stmt)) |
1613 | return; |
1614 | |
1615 | /* Check if this call is a transaction abort. */ |
1616 | fn = gimple_call_fndecl (gs: stmt); |
1617 | if (is_tm_abort (fndecl: fn)) |
1618 | *state |= GTMA_HAVE_ABORT; |
1619 | |
1620 | /* Note that something may happen. */ |
1621 | *state |= GTMA_HAVE_LOAD | GTMA_HAVE_STORE; |
1622 | } |
1623 | |
1624 | /* Iterate through the statements in the sequence, moving labels |
1625 | (and thus edges) of transactions from "label_norm" to "label_uninst". */ |
1626 | |
1627 | static tree |
1628 | make_tm_uninst (gimple_stmt_iterator *gsi, bool *handled_ops_p, |
1629 | struct walk_stmt_info *) |
1630 | { |
1631 | gimple *stmt = gsi_stmt (i: *gsi); |
1632 | |
1633 | if (gtransaction *txn = dyn_cast <gtransaction *> (p: stmt)) |
1634 | { |
1635 | *handled_ops_p = true; |
1636 | txn->label_uninst = txn->label_norm; |
1637 | txn->label_norm = NULL; |
1638 | } |
1639 | else |
1640 | *handled_ops_p = !gimple_has_substatements (g: stmt); |
1641 | |
1642 | return NULL_TREE; |
1643 | } |
1644 | |
1645 | /* Lower a GIMPLE_TRANSACTION statement. */ |
1646 | |
1647 | static void |
1648 | lower_transaction (gimple_stmt_iterator *gsi, struct walk_stmt_info *wi) |
1649 | { |
1650 | gimple *g; |
1651 | gtransaction *stmt = as_a <gtransaction *> (p: gsi_stmt (i: *gsi)); |
1652 | unsigned int *outer_state = (unsigned int *) wi->info; |
1653 | unsigned int this_state = 0; |
1654 | struct walk_stmt_info this_wi; |
1655 | |
1656 | /* First, lower the body. The scanning that we do inside gives |
1657 | us some idea of what we're dealing with. */ |
1658 | memset (s: &this_wi, c: 0, n: sizeof (this_wi)); |
1659 | this_wi.info = (void *) &this_state; |
1660 | walk_gimple_seq_mod (gimple_transaction_body_ptr (transaction_stmt: stmt), |
1661 | lower_sequence_tm, NULL, &this_wi); |
1662 | |
1663 | /* If there was absolutely nothing transaction related inside the |
1664 | transaction, we may elide it. Likewise if this is a nested |
1665 | transaction and does not contain an abort. */ |
1666 | if (this_state == 0 |
1667 | || (!(this_state & GTMA_HAVE_ABORT) && outer_state != NULL)) |
1668 | { |
1669 | if (outer_state) |
1670 | *outer_state |= this_state; |
1671 | |
1672 | gsi_insert_seq_before (gsi, gimple_transaction_body (transaction_stmt: stmt), |
1673 | GSI_SAME_STMT); |
1674 | gimple_transaction_set_body (transaction_stmt: stmt, NULL); |
1675 | |
1676 | gsi_remove (gsi, true); |
1677 | wi->removed_stmt = true; |
1678 | return; |
1679 | } |
1680 | |
1681 | /* Wrap the body of the transaction in a try-finally node so that |
1682 | the commit call is always properly called. */ |
1683 | g = gimple_build_call (builtin_decl_explicit (fncode: BUILT_IN_TM_COMMIT), 0); |
1684 | if (flag_exceptions) |
1685 | { |
1686 | tree ptr; |
1687 | gimple_seq n_seq, e_seq; |
1688 | |
1689 | n_seq = gimple_seq_alloc_with_stmt (stmt: g); |
1690 | e_seq = NULL; |
1691 | |
1692 | g = gimple_build_call (builtin_decl_explicit (fncode: BUILT_IN_EH_POINTER), |
1693 | 1, integer_zero_node); |
1694 | ptr = create_tmp_var (ptr_type_node); |
1695 | gimple_call_set_lhs (gs: g, lhs: ptr); |
1696 | gimple_seq_add_stmt (&e_seq, g); |
1697 | |
1698 | g = gimple_build_call (builtin_decl_explicit (fncode: BUILT_IN_TM_COMMIT_EH), |
1699 | 1, ptr); |
1700 | gimple_seq_add_stmt (&e_seq, g); |
1701 | |
1702 | g = gimple_build_eh_else (n_seq, e_seq); |
1703 | } |
1704 | |
1705 | g = gimple_build_try (gimple_transaction_body (transaction_stmt: stmt), |
1706 | gimple_seq_alloc_with_stmt (stmt: g), GIMPLE_TRY_FINALLY); |
1707 | |
1708 | /* For a (potentially) outer transaction, create two paths. */ |
1709 | gimple_seq uninst = NULL; |
1710 | if (outer_state == NULL) |
1711 | { |
1712 | uninst = copy_gimple_seq_and_replace_locals (seq: g); |
1713 | /* In the uninstrumented copy, reset inner transactions to have only |
1714 | an uninstrumented code path. */ |
1715 | memset (s: &this_wi, c: 0, n: sizeof (this_wi)); |
1716 | walk_gimple_seq (uninst, make_tm_uninst, NULL, &this_wi); |
1717 | } |
1718 | |
1719 | tree label1 = create_artificial_label (UNKNOWN_LOCATION); |
1720 | gsi_insert_after (gsi, gimple_build_label (label: label1), GSI_CONTINUE_LINKING); |
1721 | gsi_insert_after (gsi, g, GSI_CONTINUE_LINKING); |
1722 | gimple_transaction_set_label_norm (transaction_stmt: stmt, label: label1); |
1723 | |
1724 | /* If the transaction calls abort or if this is an outer transaction, |
1725 | add an "over" label afterwards. */ |
1726 | tree label3 = NULL; |
1727 | if ((this_state & GTMA_HAVE_ABORT) |
1728 | || outer_state == NULL |
1729 | || (gimple_transaction_subcode (transaction_stmt: stmt) & GTMA_IS_OUTER)) |
1730 | { |
1731 | label3 = create_artificial_label (UNKNOWN_LOCATION); |
1732 | gimple_transaction_set_label_over (transaction_stmt: stmt, label: label3); |
1733 | } |
1734 | |
1735 | if (uninst != NULL) |
1736 | { |
1737 | gsi_insert_after (gsi, gimple_build_goto (dest: label3), GSI_CONTINUE_LINKING); |
1738 | |
1739 | tree label2 = create_artificial_label (UNKNOWN_LOCATION); |
1740 | gsi_insert_after (gsi, gimple_build_label (label: label2), GSI_CONTINUE_LINKING); |
1741 | gsi_insert_seq_after (gsi, uninst, GSI_CONTINUE_LINKING); |
1742 | gimple_transaction_set_label_uninst (transaction_stmt: stmt, label: label2); |
1743 | } |
1744 | |
1745 | if (label3 != NULL) |
1746 | gsi_insert_after (gsi, gimple_build_label (label: label3), GSI_CONTINUE_LINKING); |
1747 | |
1748 | gimple_transaction_set_body (transaction_stmt: stmt, NULL); |
1749 | |
1750 | /* Record the set of operations found for use later. */ |
1751 | this_state |= gimple_transaction_subcode (transaction_stmt: stmt) & GTMA_DECLARATION_MASK; |
1752 | gimple_transaction_set_subcode (transaction_stmt: stmt, subcode: this_state); |
1753 | } |
1754 | |
1755 | /* Iterate through the statements in the sequence, lowering them all |
1756 | as appropriate for being in a transaction. */ |
1757 | |
1758 | static tree |
1759 | lower_sequence_tm (gimple_stmt_iterator *gsi, bool *handled_ops_p, |
1760 | struct walk_stmt_info *wi) |
1761 | { |
1762 | unsigned int *state = (unsigned int *) wi->info; |
1763 | gimple *stmt = gsi_stmt (i: *gsi); |
1764 | |
1765 | *handled_ops_p = true; |
1766 | switch (gimple_code (g: stmt)) |
1767 | { |
1768 | case GIMPLE_ASSIGN: |
1769 | /* Only memory reads/writes need to be instrumented. */ |
1770 | if (gimple_assign_single_p (gs: stmt)) |
1771 | examine_assign_tm (state, gsi); |
1772 | break; |
1773 | |
1774 | case GIMPLE_CALL: |
1775 | examine_call_tm (state, gsi); |
1776 | break; |
1777 | |
1778 | case GIMPLE_ASM: |
1779 | *state |= GTMA_MAY_ENTER_IRREVOCABLE; |
1780 | break; |
1781 | |
1782 | case GIMPLE_TRANSACTION: |
1783 | lower_transaction (gsi, wi); |
1784 | break; |
1785 | |
1786 | default: |
1787 | *handled_ops_p = !gimple_has_substatements (g: stmt); |
1788 | break; |
1789 | } |
1790 | |
1791 | return NULL_TREE; |
1792 | } |
1793 | |
1794 | /* Iterate through the statements in the sequence, lowering them all |
1795 | as appropriate for being outside of a transaction. */ |
1796 | |
1797 | static tree |
1798 | lower_sequence_no_tm (gimple_stmt_iterator *gsi, bool *handled_ops_p, |
1799 | struct walk_stmt_info * wi) |
1800 | { |
1801 | gimple *stmt = gsi_stmt (i: *gsi); |
1802 | |
1803 | if (gimple_code (g: stmt) == GIMPLE_TRANSACTION) |
1804 | { |
1805 | *handled_ops_p = true; |
1806 | lower_transaction (gsi, wi); |
1807 | } |
1808 | else |
1809 | *handled_ops_p = !gimple_has_substatements (g: stmt); |
1810 | |
1811 | return NULL_TREE; |
1812 | } |
1813 | |
1814 | /* Main entry point for flattening GIMPLE_TRANSACTION constructs. After |
1815 | this, GIMPLE_TRANSACTION nodes still exist, but the nested body has |
1816 | been moved out, and all the data required for constructing a proper |
1817 | CFG has been recorded. */ |
1818 | |
1819 | static unsigned int |
1820 | execute_lower_tm (void) |
1821 | { |
1822 | struct walk_stmt_info wi; |
1823 | gimple_seq body; |
1824 | |
1825 | /* Transactional clones aren't created until a later pass. */ |
1826 | gcc_assert (!decl_is_tm_clone (current_function_decl)); |
1827 | |
1828 | body = gimple_body (current_function_decl); |
1829 | memset (s: &wi, c: 0, n: sizeof (wi)); |
1830 | walk_gimple_seq_mod (&body, lower_sequence_no_tm, NULL, &wi); |
1831 | gimple_set_body (current_function_decl, body); |
1832 | |
1833 | return 0; |
1834 | } |
1835 | |
1836 | namespace { |
1837 | |
1838 | const pass_data pass_data_lower_tm = |
1839 | { |
1840 | .type: GIMPLE_PASS, /* type */ |
1841 | .name: "tmlower" , /* name */ |
1842 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
1843 | .tv_id: TV_TRANS_MEM, /* tv_id */ |
1844 | PROP_gimple_lcf, /* properties_required */ |
1845 | .properties_provided: 0, /* properties_provided */ |
1846 | .properties_destroyed: 0, /* properties_destroyed */ |
1847 | .todo_flags_start: 0, /* todo_flags_start */ |
1848 | .todo_flags_finish: 0, /* todo_flags_finish */ |
1849 | }; |
1850 | |
1851 | class pass_lower_tm : public gimple_opt_pass |
1852 | { |
1853 | public: |
1854 | pass_lower_tm (gcc::context *ctxt) |
1855 | : gimple_opt_pass (pass_data_lower_tm, ctxt) |
1856 | {} |
1857 | |
1858 | /* opt_pass methods: */ |
1859 | bool gate (function *) final override { return flag_tm; } |
1860 | unsigned int execute (function *) final override |
1861 | { |
1862 | return execute_lower_tm (); |
1863 | } |
1864 | |
1865 | }; // class pass_lower_tm |
1866 | |
1867 | } // anon namespace |
1868 | |
1869 | gimple_opt_pass * |
1870 | make_pass_lower_tm (gcc::context *ctxt) |
1871 | { |
1872 | return new pass_lower_tm (ctxt); |
1873 | } |
1874 | |
1875 | /* Collect region information for each transaction. */ |
1876 | |
1877 | struct tm_region |
1878 | { |
1879 | public: |
1880 | |
1881 | /* The field "transaction_stmt" is initially a gtransaction *, |
1882 | but eventually gets lowered to a gcall *(to BUILT_IN_TM_START). |
1883 | |
1884 | Helper method to get it as a gtransaction *, with code-checking |
1885 | in a checked-build. */ |
1886 | |
1887 | gtransaction * |
1888 | get_transaction_stmt () const |
1889 | { |
1890 | return as_a <gtransaction *> (p: transaction_stmt); |
1891 | } |
1892 | |
1893 | public: |
1894 | |
1895 | /* Link to the next unnested transaction. */ |
1896 | struct tm_region *next; |
1897 | |
1898 | /* Link to the next inner transaction. */ |
1899 | struct tm_region *inner; |
1900 | |
1901 | /* Link to the next outer transaction. */ |
1902 | struct tm_region *outer; |
1903 | |
1904 | /* The GIMPLE_TRANSACTION statement beginning this transaction. |
1905 | After TM_MARK, this gets replaced by a call to |
1906 | BUILT_IN_TM_START. |
1907 | Hence this will be either a gtransaction *or a gcall *. */ |
1908 | gimple *transaction_stmt; |
1909 | |
1910 | /* After TM_MARK expands the GIMPLE_TRANSACTION into a call to |
1911 | BUILT_IN_TM_START, this field is true if the transaction is an |
1912 | outer transaction. */ |
1913 | bool original_transaction_was_outer; |
1914 | |
1915 | /* Return value from BUILT_IN_TM_START. */ |
1916 | tree tm_state; |
1917 | |
1918 | /* The entry block to this region. This will always be the first |
1919 | block of the body of the transaction. */ |
1920 | basic_block entry_block; |
1921 | |
1922 | /* The first block after an expanded call to _ITM_beginTransaction. */ |
1923 | basic_block restart_block; |
1924 | |
1925 | /* The set of all blocks that end the region; NULL if only EXIT_BLOCK. |
1926 | These blocks are still a part of the region (i.e., the border is |
1927 | inclusive). Note that this set is only complete for paths in the CFG |
1928 | starting at ENTRY_BLOCK, and that there is no exit block recorded for |
1929 | the edge to the "over" label. */ |
1930 | bitmap exit_blocks; |
1931 | |
1932 | /* The set of all blocks that have an TM_IRREVOCABLE call. */ |
1933 | bitmap irr_blocks; |
1934 | }; |
1935 | |
1936 | /* True if there are pending edge statements to be committed for the |
1937 | current function being scanned in the tmmark pass. */ |
1938 | bool pending_edge_inserts_p; |
1939 | |
1940 | static struct tm_region *all_tm_regions; |
1941 | static bitmap_obstack tm_obstack; |
1942 | |
1943 | |
1944 | /* A subroutine of tm_region_init. Record the existence of the |
1945 | GIMPLE_TRANSACTION statement in a tree of tm_region elements. */ |
1946 | |
1947 | static struct tm_region * |
1948 | tm_region_init_0 (struct tm_region *outer, basic_block bb, |
1949 | gtransaction *stmt) |
1950 | { |
1951 | struct tm_region *region; |
1952 | |
1953 | region = (struct tm_region *) |
1954 | obstack_alloc (&tm_obstack.obstack, sizeof (struct tm_region)); |
1955 | |
1956 | if (outer) |
1957 | { |
1958 | region->next = outer->inner; |
1959 | outer->inner = region; |
1960 | } |
1961 | else |
1962 | { |
1963 | region->next = all_tm_regions; |
1964 | all_tm_regions = region; |
1965 | } |
1966 | region->inner = NULL; |
1967 | region->outer = outer; |
1968 | |
1969 | region->transaction_stmt = stmt; |
1970 | region->original_transaction_was_outer = false; |
1971 | region->tm_state = NULL; |
1972 | |
1973 | /* There are either one or two edges out of the block containing |
1974 | the GIMPLE_TRANSACTION, one to the actual region and one to the |
1975 | "over" label if the region contains an abort. The former will |
1976 | always be the one marked FALLTHRU. */ |
1977 | region->entry_block = FALLTHRU_EDGE (bb)->dest; |
1978 | |
1979 | region->exit_blocks = BITMAP_ALLOC (obstack: &tm_obstack); |
1980 | region->irr_blocks = BITMAP_ALLOC (obstack: &tm_obstack); |
1981 | |
1982 | return region; |
1983 | } |
1984 | |
1985 | /* A subroutine of tm_region_init. Record all the exit and |
1986 | irrevocable blocks in BB into the region's exit_blocks and |
1987 | irr_blocks bitmaps. Returns the new region being scanned. */ |
1988 | |
1989 | static struct tm_region * |
1990 | tm_region_init_1 (struct tm_region *region, basic_block bb) |
1991 | { |
1992 | gimple_stmt_iterator gsi; |
1993 | gimple *g; |
1994 | |
1995 | if (!region |
1996 | || (!region->irr_blocks && !region->exit_blocks)) |
1997 | return region; |
1998 | |
1999 | /* Check to see if this is the end of a region by seeing if it |
2000 | contains a call to __builtin_tm_commit{,_eh}. Note that the |
2001 | outermost region for DECL_IS_TM_CLONE need not collect this. */ |
2002 | for (gsi = gsi_last_bb (bb); !gsi_end_p (i: gsi); gsi_prev (i: &gsi)) |
2003 | { |
2004 | g = gsi_stmt (i: gsi); |
2005 | if (gimple_code (g) == GIMPLE_CALL) |
2006 | { |
2007 | tree fn = gimple_call_fndecl (gs: g); |
2008 | if (fn && fndecl_built_in_p (node: fn, klass: BUILT_IN_NORMAL)) |
2009 | { |
2010 | if ((DECL_FUNCTION_CODE (decl: fn) == BUILT_IN_TM_COMMIT |
2011 | || DECL_FUNCTION_CODE (decl: fn) == BUILT_IN_TM_COMMIT_EH) |
2012 | && region->exit_blocks) |
2013 | { |
2014 | bitmap_set_bit (region->exit_blocks, bb->index); |
2015 | region = region->outer; |
2016 | break; |
2017 | } |
2018 | if (DECL_FUNCTION_CODE (decl: fn) == BUILT_IN_TM_IRREVOCABLE) |
2019 | bitmap_set_bit (region->irr_blocks, bb->index); |
2020 | } |
2021 | } |
2022 | } |
2023 | return region; |
2024 | } |
2025 | |
2026 | /* Collect all of the transaction regions within the current function |
2027 | and record them in ALL_TM_REGIONS. The REGION parameter may specify |
2028 | an "outermost" region for use by tm clones. */ |
2029 | |
2030 | static void |
2031 | tm_region_init (struct tm_region *region) |
2032 | { |
2033 | gimple *g; |
2034 | edge_iterator ei; |
2035 | edge e; |
2036 | basic_block bb; |
2037 | auto_vec<basic_block> queue; |
2038 | bitmap visited_blocks = BITMAP_ALLOC (NULL); |
2039 | struct tm_region *old_region; |
2040 | auto_vec<tm_region *> bb_regions; |
2041 | |
2042 | /* We could store this information in bb->aux, but we may get called |
2043 | through get_all_tm_blocks() from another pass that may be already |
2044 | using bb->aux. */ |
2045 | bb_regions.safe_grow_cleared (last_basic_block_for_fn (cfun), exact: true); |
2046 | |
2047 | all_tm_regions = region; |
2048 | bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
2049 | queue.safe_push (obj: bb); |
2050 | bitmap_set_bit (visited_blocks, bb->index); |
2051 | bb_regions[bb->index] = region; |
2052 | |
2053 | do |
2054 | { |
2055 | bb = queue.pop (); |
2056 | region = bb_regions[bb->index]; |
2057 | bb_regions[bb->index] = NULL; |
2058 | |
2059 | /* Record exit and irrevocable blocks. */ |
2060 | region = tm_region_init_1 (region, bb); |
2061 | |
2062 | /* Check for the last statement in the block beginning a new region. */ |
2063 | g = last_nondebug_stmt (bb); |
2064 | old_region = region; |
2065 | if (g) |
2066 | if (gtransaction *trans_stmt = dyn_cast <gtransaction *> (p: g)) |
2067 | region = tm_region_init_0 (outer: region, bb, stmt: trans_stmt); |
2068 | |
2069 | /* Process subsequent blocks. */ |
2070 | FOR_EACH_EDGE (e, ei, bb->succs) |
2071 | if (!bitmap_bit_p (visited_blocks, e->dest->index)) |
2072 | { |
2073 | bitmap_set_bit (visited_blocks, e->dest->index); |
2074 | queue.safe_push (obj: e->dest); |
2075 | |
2076 | /* If the current block started a new region, make sure that only |
2077 | the entry block of the new region is associated with this region. |
2078 | Other successors are still part of the old region. */ |
2079 | if (old_region != region && e->dest != region->entry_block) |
2080 | bb_regions[e->dest->index] = old_region; |
2081 | else |
2082 | bb_regions[e->dest->index] = region; |
2083 | } |
2084 | } |
2085 | while (!queue.is_empty ()); |
2086 | BITMAP_FREE (visited_blocks); |
2087 | } |
2088 | |
2089 | /* The "gate" function for all transactional memory expansion and optimization |
2090 | passes. We collect region information for each top-level transaction, and |
2091 | if we don't find any, we skip all of the TM passes. Each region will have |
2092 | all of the exit blocks recorded, and the originating statement. */ |
2093 | |
2094 | static bool |
2095 | gate_tm_init (void) |
2096 | { |
2097 | if (!flag_tm) |
2098 | return false; |
2099 | |
2100 | calculate_dominance_info (CDI_DOMINATORS); |
2101 | bitmap_obstack_initialize (&tm_obstack); |
2102 | |
2103 | /* If the function is a TM_CLONE, then the entire function is the region. */ |
2104 | if (decl_is_tm_clone (fndecl: current_function_decl)) |
2105 | { |
2106 | struct tm_region *region = (struct tm_region *) |
2107 | obstack_alloc (&tm_obstack.obstack, sizeof (struct tm_region)); |
2108 | memset (s: region, c: 0, n: sizeof (*region)); |
2109 | region->entry_block = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
2110 | /* For a clone, the entire function is the region. But even if |
2111 | we don't need to record any exit blocks, we may need to |
2112 | record irrevocable blocks. */ |
2113 | region->irr_blocks = BITMAP_ALLOC (obstack: &tm_obstack); |
2114 | |
2115 | tm_region_init (region); |
2116 | } |
2117 | else |
2118 | { |
2119 | tm_region_init (NULL); |
2120 | |
2121 | /* If we didn't find any regions, cleanup and skip the whole tree |
2122 | of tm-related optimizations. */ |
2123 | if (all_tm_regions == NULL) |
2124 | { |
2125 | bitmap_obstack_release (&tm_obstack); |
2126 | return false; |
2127 | } |
2128 | } |
2129 | |
2130 | return true; |
2131 | } |
2132 | |
2133 | namespace { |
2134 | |
2135 | const pass_data pass_data_tm_init = |
2136 | { |
2137 | .type: GIMPLE_PASS, /* type */ |
2138 | .name: "*tminit" , /* name */ |
2139 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
2140 | .tv_id: TV_TRANS_MEM, /* tv_id */ |
2141 | .properties_required: ( PROP_ssa | PROP_cfg ), /* properties_required */ |
2142 | .properties_provided: 0, /* properties_provided */ |
2143 | .properties_destroyed: 0, /* properties_destroyed */ |
2144 | .todo_flags_start: 0, /* todo_flags_start */ |
2145 | .todo_flags_finish: 0, /* todo_flags_finish */ |
2146 | }; |
2147 | |
2148 | class pass_tm_init : public gimple_opt_pass |
2149 | { |
2150 | public: |
2151 | pass_tm_init (gcc::context *ctxt) |
2152 | : gimple_opt_pass (pass_data_tm_init, ctxt) |
2153 | {} |
2154 | |
2155 | /* opt_pass methods: */ |
2156 | bool gate (function *) final override { return gate_tm_init (); } |
2157 | |
2158 | }; // class pass_tm_init |
2159 | |
2160 | } // anon namespace |
2161 | |
2162 | gimple_opt_pass * |
2163 | make_pass_tm_init (gcc::context *ctxt) |
2164 | { |
2165 | return new pass_tm_init (ctxt); |
2166 | } |
2167 | |
2168 | /* Add FLAGS to the GIMPLE_TRANSACTION subcode for the transaction region |
2169 | represented by STATE. */ |
2170 | |
2171 | static inline void |
2172 | transaction_subcode_ior (struct tm_region *region, unsigned flags) |
2173 | { |
2174 | if (region && region->transaction_stmt) |
2175 | { |
2176 | gtransaction *transaction_stmt = region->get_transaction_stmt (); |
2177 | flags |= gimple_transaction_subcode (transaction_stmt); |
2178 | gimple_transaction_set_subcode (transaction_stmt, subcode: flags); |
2179 | } |
2180 | } |
2181 | |
2182 | /* Construct a memory load in a transactional context. Return the |
2183 | gimple statement performing the load, or NULL if there is no |
2184 | TM_LOAD builtin of the appropriate size to do the load. |
2185 | |
2186 | LOC is the location to use for the new statement(s). */ |
2187 | |
2188 | static gcall * |
2189 | build_tm_load (location_t loc, tree lhs, tree rhs, gimple_stmt_iterator *gsi) |
2190 | { |
2191 | tree t, type = TREE_TYPE (rhs); |
2192 | gcall *gcall; |
2193 | |
2194 | built_in_function code; |
2195 | if (type == float_type_node) |
2196 | code = BUILT_IN_TM_LOAD_FLOAT; |
2197 | else if (type == double_type_node) |
2198 | code = BUILT_IN_TM_LOAD_DOUBLE; |
2199 | else if (type == long_double_type_node) |
2200 | code = BUILT_IN_TM_LOAD_LDOUBLE; |
2201 | else |
2202 | { |
2203 | if (TYPE_SIZE (type) == NULL || !tree_fits_uhwi_p (TYPE_SIZE (type))) |
2204 | return NULL; |
2205 | unsigned HOST_WIDE_INT type_size = tree_to_uhwi (TYPE_SIZE (type)); |
2206 | |
2207 | if (TREE_CODE (type) == VECTOR_TYPE) |
2208 | { |
2209 | switch (type_size) |
2210 | { |
2211 | case 64: |
2212 | code = BUILT_IN_TM_LOAD_M64; |
2213 | break; |
2214 | case 128: |
2215 | code = BUILT_IN_TM_LOAD_M128; |
2216 | break; |
2217 | case 256: |
2218 | code = BUILT_IN_TM_LOAD_M256; |
2219 | break; |
2220 | default: |
2221 | goto unhandled_vec; |
2222 | } |
2223 | if (!builtin_decl_explicit_p (fncode: code)) |
2224 | goto unhandled_vec; |
2225 | } |
2226 | else |
2227 | { |
2228 | unhandled_vec: |
2229 | switch (type_size) |
2230 | { |
2231 | case 8: |
2232 | code = BUILT_IN_TM_LOAD_1; |
2233 | break; |
2234 | case 16: |
2235 | code = BUILT_IN_TM_LOAD_2; |
2236 | break; |
2237 | case 32: |
2238 | code = BUILT_IN_TM_LOAD_4; |
2239 | break; |
2240 | case 64: |
2241 | code = BUILT_IN_TM_LOAD_8; |
2242 | break; |
2243 | default: |
2244 | return NULL; |
2245 | } |
2246 | } |
2247 | } |
2248 | |
2249 | tree decl = builtin_decl_explicit (fncode: code); |
2250 | gcc_assert (decl); |
2251 | |
2252 | t = gimplify_addr (gsi, x: rhs); |
2253 | gcall = gimple_build_call (decl, 1, t); |
2254 | gimple_set_location (g: gcall, location: loc); |
2255 | |
2256 | t = TREE_TYPE (TREE_TYPE (decl)); |
2257 | if (useless_type_conversion_p (type, t)) |
2258 | { |
2259 | gimple_call_set_lhs (gs: gcall, lhs); |
2260 | gsi_insert_before (gsi, gcall, GSI_SAME_STMT); |
2261 | } |
2262 | else |
2263 | { |
2264 | gimple *g; |
2265 | tree temp; |
2266 | |
2267 | temp = create_tmp_reg (t); |
2268 | gimple_call_set_lhs (gs: gcall, lhs: temp); |
2269 | gsi_insert_before (gsi, gcall, GSI_SAME_STMT); |
2270 | |
2271 | t = fold_build1 (VIEW_CONVERT_EXPR, type, temp); |
2272 | g = gimple_build_assign (lhs, t); |
2273 | gsi_insert_before (gsi, g, GSI_SAME_STMT); |
2274 | } |
2275 | |
2276 | return gcall; |
2277 | } |
2278 | |
2279 | |
2280 | /* Similarly for storing TYPE in a transactional context. */ |
2281 | |
2282 | static gcall * |
2283 | build_tm_store (location_t loc, tree lhs, tree rhs, gimple_stmt_iterator *gsi) |
2284 | { |
2285 | tree t, fn, type = TREE_TYPE (rhs), simple_type; |
2286 | gcall *gcall; |
2287 | |
2288 | built_in_function code; |
2289 | if (type == float_type_node) |
2290 | code = BUILT_IN_TM_STORE_FLOAT; |
2291 | else if (type == double_type_node) |
2292 | code = BUILT_IN_TM_STORE_DOUBLE; |
2293 | else if (type == long_double_type_node) |
2294 | code = BUILT_IN_TM_STORE_LDOUBLE; |
2295 | else |
2296 | { |
2297 | if (TYPE_SIZE (type) == NULL || !tree_fits_uhwi_p (TYPE_SIZE (type))) |
2298 | return NULL; |
2299 | unsigned HOST_WIDE_INT type_size = tree_to_uhwi (TYPE_SIZE (type)); |
2300 | |
2301 | if (TREE_CODE (type) == VECTOR_TYPE) |
2302 | { |
2303 | switch (type_size) |
2304 | { |
2305 | case 64: |
2306 | code = BUILT_IN_TM_STORE_M64; |
2307 | break; |
2308 | case 128: |
2309 | code = BUILT_IN_TM_STORE_M128; |
2310 | break; |
2311 | case 256: |
2312 | code = BUILT_IN_TM_STORE_M256; |
2313 | break; |
2314 | default: |
2315 | goto unhandled_vec; |
2316 | } |
2317 | if (!builtin_decl_explicit_p (fncode: code)) |
2318 | goto unhandled_vec; |
2319 | } |
2320 | else |
2321 | { |
2322 | unhandled_vec: |
2323 | switch (type_size) |
2324 | { |
2325 | case 8: |
2326 | code = BUILT_IN_TM_STORE_1; |
2327 | break; |
2328 | case 16: |
2329 | code = BUILT_IN_TM_STORE_2; |
2330 | break; |
2331 | case 32: |
2332 | code = BUILT_IN_TM_STORE_4; |
2333 | break; |
2334 | case 64: |
2335 | code = BUILT_IN_TM_STORE_8; |
2336 | break; |
2337 | default: |
2338 | return NULL; |
2339 | } |
2340 | } |
2341 | } |
2342 | |
2343 | fn = builtin_decl_explicit (fncode: code); |
2344 | gcc_assert (fn); |
2345 | |
2346 | simple_type = TREE_VALUE (TREE_CHAIN (TYPE_ARG_TYPES (TREE_TYPE (fn)))); |
2347 | |
2348 | if (TREE_CODE (rhs) == CONSTRUCTOR) |
2349 | { |
2350 | /* Handle the easy initialization to zero. */ |
2351 | if (!CONSTRUCTOR_ELTS (rhs)) |
2352 | rhs = build_int_cst (simple_type, 0); |
2353 | else |
2354 | { |
2355 | /* ...otherwise punt to the caller and probably use |
2356 | BUILT_IN_TM_MEMMOVE, because we can't wrap a |
2357 | VIEW_CONVERT_EXPR around a CONSTRUCTOR (below) and produce |
2358 | valid gimple. */ |
2359 | return NULL; |
2360 | } |
2361 | } |
2362 | else if (!useless_type_conversion_p (simple_type, type)) |
2363 | { |
2364 | gimple *g; |
2365 | tree temp; |
2366 | |
2367 | temp = create_tmp_reg (simple_type); |
2368 | t = fold_build1 (VIEW_CONVERT_EXPR, simple_type, rhs); |
2369 | g = gimple_build_assign (temp, t); |
2370 | gimple_set_location (g, location: loc); |
2371 | gsi_insert_before (gsi, g, GSI_SAME_STMT); |
2372 | |
2373 | rhs = temp; |
2374 | } |
2375 | |
2376 | t = gimplify_addr (gsi, x: lhs); |
2377 | gcall = gimple_build_call (fn, 2, t, rhs); |
2378 | gimple_set_location (g: gcall, location: loc); |
2379 | gsi_insert_before (gsi, gcall, GSI_SAME_STMT); |
2380 | |
2381 | return gcall; |
2382 | } |
2383 | |
2384 | |
2385 | /* Expand an assignment statement into transactional builtins. */ |
2386 | |
2387 | static void |
2388 | expand_assign_tm (struct tm_region *region, gimple_stmt_iterator *gsi) |
2389 | { |
2390 | gimple *stmt = gsi_stmt (i: *gsi); |
2391 | location_t loc = gimple_location (g: stmt); |
2392 | tree lhs = gimple_assign_lhs (gs: stmt); |
2393 | tree rhs = gimple_assign_rhs1 (gs: stmt); |
2394 | bool store_p = requires_barrier (entry_block: region->entry_block, x: lhs, NULL); |
2395 | bool load_p = requires_barrier (entry_block: region->entry_block, x: rhs, NULL); |
2396 | gimple *gcall = NULL; |
2397 | |
2398 | if (!load_p && !store_p) |
2399 | { |
2400 | /* Add thread private addresses to log if applicable. */ |
2401 | requires_barrier (entry_block: region->entry_block, x: lhs, stmt); |
2402 | gsi_next (i: gsi); |
2403 | return; |
2404 | } |
2405 | |
2406 | if (load_p) |
2407 | transaction_subcode_ior (region, GTMA_HAVE_LOAD); |
2408 | if (store_p) |
2409 | transaction_subcode_ior (region, GTMA_HAVE_STORE); |
2410 | |
2411 | // Remove original load/store statement. |
2412 | gsi_remove (gsi, true); |
2413 | |
2414 | // Attempt to use a simple load/store helper function. |
2415 | if (load_p && !store_p) |
2416 | gcall = build_tm_load (loc, lhs, rhs, gsi); |
2417 | else if (store_p && !load_p) |
2418 | gcall = build_tm_store (loc, lhs, rhs, gsi); |
2419 | |
2420 | // If gcall has not been set, then we do not have a simple helper |
2421 | // function available for the type. This may be true of larger |
2422 | // structures, vectors, and non-standard float types. |
2423 | if (!gcall) |
2424 | { |
2425 | tree lhs_addr, rhs_addr, ltmp = NULL, copy_fn; |
2426 | |
2427 | // If this is a type that we couldn't handle above, but it's |
2428 | // in a register, we must spill it to memory for the copy. |
2429 | if (is_gimple_reg (lhs)) |
2430 | { |
2431 | ltmp = create_tmp_var (TREE_TYPE (lhs)); |
2432 | lhs_addr = build_fold_addr_expr (ltmp); |
2433 | } |
2434 | else |
2435 | lhs_addr = gimplify_addr (gsi, x: lhs); |
2436 | if (is_gimple_reg (rhs)) |
2437 | { |
2438 | tree rtmp = create_tmp_var (TREE_TYPE (rhs)); |
2439 | TREE_ADDRESSABLE (rtmp) = 1; |
2440 | rhs_addr = build_fold_addr_expr (rtmp); |
2441 | gcall = gimple_build_assign (rtmp, rhs); |
2442 | gsi_insert_before (gsi, gcall, GSI_SAME_STMT); |
2443 | } |
2444 | else |
2445 | rhs_addr = gimplify_addr (gsi, x: rhs); |
2446 | |
2447 | // Choose the appropriate memory transfer function. |
2448 | if (load_p && store_p) |
2449 | { |
2450 | // ??? Figure out if there's any possible overlap between |
2451 | // the LHS and the RHS and if not, use MEMCPY. |
2452 | copy_fn = builtin_decl_explicit (fncode: BUILT_IN_TM_MEMMOVE); |
2453 | } |
2454 | else if (load_p) |
2455 | { |
2456 | // Note that the store is non-transactional and cannot overlap. |
2457 | copy_fn = builtin_decl_explicit (fncode: BUILT_IN_TM_MEMCPY_RTWN); |
2458 | } |
2459 | else |
2460 | { |
2461 | // Note that the load is non-transactional and cannot overlap. |
2462 | copy_fn = builtin_decl_explicit (fncode: BUILT_IN_TM_MEMCPY_RNWT); |
2463 | } |
2464 | |
2465 | gcall = gimple_build_call (copy_fn, 3, lhs_addr, rhs_addr, |
2466 | TYPE_SIZE_UNIT (TREE_TYPE (lhs))); |
2467 | gimple_set_location (g: gcall, location: loc); |
2468 | gsi_insert_before (gsi, gcall, GSI_SAME_STMT); |
2469 | |
2470 | if (ltmp) |
2471 | { |
2472 | gcall = gimple_build_assign (lhs, ltmp); |
2473 | gsi_insert_before (gsi, gcall, GSI_SAME_STMT); |
2474 | } |
2475 | } |
2476 | |
2477 | // Now that we have the load/store in its instrumented form, add |
2478 | // thread private addresses to the log if applicable. |
2479 | if (!store_p) |
2480 | requires_barrier (entry_block: region->entry_block, x: lhs, stmt: gcall); |
2481 | } |
2482 | |
2483 | |
2484 | /* Expand a call statement as appropriate for a transaction. That is, |
2485 | either verify that the call does not affect the transaction, or |
2486 | redirect the call to a clone that handles transactions, or change |
2487 | the transaction state to IRREVOCABLE. Return true if the call is |
2488 | one of the builtins that end a transaction. */ |
2489 | |
2490 | static bool |
2491 | expand_call_tm (struct tm_region *region, |
2492 | gimple_stmt_iterator *gsi) |
2493 | { |
2494 | gcall *stmt = as_a <gcall *> (p: gsi_stmt (i: *gsi)); |
2495 | tree lhs = gimple_call_lhs (gs: stmt); |
2496 | tree fn_decl; |
2497 | struct cgraph_node *node; |
2498 | bool retval = false; |
2499 | |
2500 | fn_decl = gimple_call_fndecl (gs: stmt); |
2501 | |
2502 | if (fn_decl == builtin_decl_explicit (fncode: BUILT_IN_TM_MEMCPY) |
2503 | || fn_decl == builtin_decl_explicit (fncode: BUILT_IN_TM_MEMMOVE)) |
2504 | transaction_subcode_ior (region, GTMA_HAVE_STORE | GTMA_HAVE_LOAD); |
2505 | if (fn_decl == builtin_decl_explicit (fncode: BUILT_IN_TM_MEMSET)) |
2506 | transaction_subcode_ior (region, GTMA_HAVE_STORE); |
2507 | |
2508 | if (is_tm_pure_call (call: stmt)) |
2509 | return false; |
2510 | |
2511 | if (fn_decl) |
2512 | retval = is_tm_ending_fndecl (fndecl: fn_decl); |
2513 | if (!retval) |
2514 | { |
2515 | /* Assume all non-const/pure calls write to memory, except |
2516 | transaction ending builtins. */ |
2517 | transaction_subcode_ior (region, GTMA_HAVE_STORE); |
2518 | } |
2519 | |
2520 | /* For indirect calls, we already generated a call into the runtime. */ |
2521 | if (!fn_decl) |
2522 | { |
2523 | tree fn = gimple_call_fn (gs: stmt); |
2524 | |
2525 | /* We are guaranteed never to go irrevocable on a safe or pure |
2526 | call, and the pure call was handled above. */ |
2527 | if (is_tm_safe (x: fn)) |
2528 | return false; |
2529 | else |
2530 | transaction_subcode_ior (region, GTMA_MAY_ENTER_IRREVOCABLE); |
2531 | |
2532 | return false; |
2533 | } |
2534 | |
2535 | node = cgraph_node::get (decl: fn_decl); |
2536 | /* All calls should have cgraph here. */ |
2537 | if (!node) |
2538 | { |
2539 | /* We can have a nodeless call here if some pass after IPA-tm |
2540 | added uninstrumented calls. For example, loop distribution |
2541 | can transform certain loop constructs into __builtin_mem* |
2542 | calls. In this case, see if we have a suitable TM |
2543 | replacement and fill in the gaps. */ |
2544 | gcc_assert (DECL_BUILT_IN_CLASS (fn_decl) == BUILT_IN_NORMAL); |
2545 | enum built_in_function code = DECL_FUNCTION_CODE (decl: fn_decl); |
2546 | gcc_assert (code == BUILT_IN_MEMCPY |
2547 | || code == BUILT_IN_MEMMOVE |
2548 | || code == BUILT_IN_MEMSET); |
2549 | |
2550 | tree repl = find_tm_replacement_function (fndecl: fn_decl); |
2551 | if (repl) |
2552 | { |
2553 | gimple_call_set_fndecl (gs: stmt, decl: repl); |
2554 | update_stmt (s: stmt); |
2555 | node = cgraph_node::create (decl: repl); |
2556 | node->tm_may_enter_irr = false; |
2557 | return expand_call_tm (region, gsi); |
2558 | } |
2559 | gcc_unreachable (); |
2560 | } |
2561 | if (node->tm_may_enter_irr) |
2562 | transaction_subcode_ior (region, GTMA_MAY_ENTER_IRREVOCABLE); |
2563 | |
2564 | if (is_tm_abort (fndecl: fn_decl)) |
2565 | { |
2566 | transaction_subcode_ior (region, GTMA_HAVE_ABORT); |
2567 | return true; |
2568 | } |
2569 | |
2570 | /* Instrument the store if needed. |
2571 | |
2572 | If the assignment happens inside the function call (return slot |
2573 | optimization), there is no instrumentation to be done, since |
2574 | the callee should have done the right thing. */ |
2575 | if (lhs && requires_barrier (entry_block: region->entry_block, x: lhs, stmt) |
2576 | && !gimple_call_return_slot_opt_p (s: stmt)) |
2577 | { |
2578 | tree tmp = create_tmp_reg (TREE_TYPE (lhs)); |
2579 | location_t loc = gimple_location (g: stmt); |
2580 | edge fallthru_edge = NULL; |
2581 | gassign *assign_stmt; |
2582 | |
2583 | /* Remember if the call was going to throw. */ |
2584 | if (stmt_can_throw_internal (cfun, stmt)) |
2585 | { |
2586 | edge_iterator ei; |
2587 | edge e; |
2588 | basic_block bb = gimple_bb (g: stmt); |
2589 | |
2590 | FOR_EACH_EDGE (e, ei, bb->succs) |
2591 | if (e->flags & EDGE_FALLTHRU) |
2592 | { |
2593 | fallthru_edge = e; |
2594 | break; |
2595 | } |
2596 | } |
2597 | |
2598 | gimple_call_set_lhs (gs: stmt, lhs: tmp); |
2599 | update_stmt (s: stmt); |
2600 | assign_stmt = gimple_build_assign (lhs, tmp); |
2601 | gimple_set_location (g: assign_stmt, location: loc); |
2602 | |
2603 | /* We cannot throw in the middle of a BB. If the call was going |
2604 | to throw, place the instrumentation on the fallthru edge, so |
2605 | the call remains the last statement in the block. */ |
2606 | if (fallthru_edge) |
2607 | { |
2608 | gimple_seq fallthru_seq = gimple_seq_alloc_with_stmt (stmt: assign_stmt); |
2609 | gimple_stmt_iterator fallthru_gsi = gsi_start (seq&: fallthru_seq); |
2610 | expand_assign_tm (region, gsi: &fallthru_gsi); |
2611 | gsi_insert_seq_on_edge (fallthru_edge, fallthru_seq); |
2612 | pending_edge_inserts_p = true; |
2613 | } |
2614 | else |
2615 | { |
2616 | gsi_insert_after (gsi, assign_stmt, GSI_CONTINUE_LINKING); |
2617 | expand_assign_tm (region, gsi); |
2618 | } |
2619 | |
2620 | transaction_subcode_ior (region, GTMA_HAVE_STORE); |
2621 | } |
2622 | |
2623 | return retval; |
2624 | } |
2625 | |
2626 | |
2627 | /* Expand all statements in BB as appropriate for being inside |
2628 | a transaction. */ |
2629 | |
2630 | static void |
2631 | expand_block_tm (struct tm_region *region, basic_block bb) |
2632 | { |
2633 | gimple_stmt_iterator gsi; |
2634 | |
2635 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); ) |
2636 | { |
2637 | gimple *stmt = gsi_stmt (i: gsi); |
2638 | switch (gimple_code (g: stmt)) |
2639 | { |
2640 | case GIMPLE_ASSIGN: |
2641 | /* Only memory reads/writes need to be instrumented. */ |
2642 | if (gimple_assign_single_p (gs: stmt) |
2643 | && !gimple_clobber_p (s: stmt)) |
2644 | { |
2645 | expand_assign_tm (region, gsi: &gsi); |
2646 | continue; |
2647 | } |
2648 | break; |
2649 | |
2650 | case GIMPLE_CALL: |
2651 | if (expand_call_tm (region, gsi: &gsi)) |
2652 | return; |
2653 | break; |
2654 | |
2655 | case GIMPLE_ASM: |
2656 | gcc_unreachable (); |
2657 | |
2658 | default: |
2659 | break; |
2660 | } |
2661 | if (!gsi_end_p (i: gsi)) |
2662 | gsi_next (i: &gsi); |
2663 | } |
2664 | } |
2665 | |
2666 | /* Return the list of basic-blocks in REGION. |
2667 | |
2668 | STOP_AT_IRREVOCABLE_P is true if caller is uninterested in blocks |
2669 | following a TM_IRREVOCABLE call. |
2670 | |
2671 | INCLUDE_UNINSTRUMENTED_P is TRUE if we should include the |
2672 | uninstrumented code path blocks in the list of basic blocks |
2673 | returned, false otherwise. */ |
2674 | |
2675 | static vec<basic_block> |
2676 | get_tm_region_blocks (basic_block entry_block, |
2677 | bitmap exit_blocks, |
2678 | bitmap irr_blocks, |
2679 | bitmap all_region_blocks, |
2680 | bool stop_at_irrevocable_p, |
2681 | bool include_uninstrumented_p = true) |
2682 | { |
2683 | vec<basic_block> bbs = vNULL; |
2684 | unsigned i; |
2685 | edge e; |
2686 | edge_iterator ei; |
2687 | bitmap visited_blocks = BITMAP_ALLOC (NULL); |
2688 | |
2689 | i = 0; |
2690 | bbs.safe_push (obj: entry_block); |
2691 | bitmap_set_bit (visited_blocks, entry_block->index); |
2692 | |
2693 | do |
2694 | { |
2695 | basic_block bb = bbs[i++]; |
2696 | |
2697 | if (exit_blocks && |
2698 | bitmap_bit_p (exit_blocks, bb->index)) |
2699 | continue; |
2700 | |
2701 | if (stop_at_irrevocable_p |
2702 | && irr_blocks |
2703 | && bitmap_bit_p (irr_blocks, bb->index)) |
2704 | continue; |
2705 | |
2706 | FOR_EACH_EDGE (e, ei, bb->succs) |
2707 | if ((include_uninstrumented_p |
2708 | || !(e->flags & EDGE_TM_UNINSTRUMENTED)) |
2709 | && !bitmap_bit_p (visited_blocks, e->dest->index)) |
2710 | { |
2711 | bitmap_set_bit (visited_blocks, e->dest->index); |
2712 | bbs.safe_push (obj: e->dest); |
2713 | } |
2714 | } |
2715 | while (i < bbs.length ()); |
2716 | |
2717 | if (all_region_blocks) |
2718 | bitmap_ior_into (all_region_blocks, visited_blocks); |
2719 | |
2720 | BITMAP_FREE (visited_blocks); |
2721 | return bbs; |
2722 | } |
2723 | |
2724 | // Callback data for collect_bb2reg. |
2725 | struct bb2reg_stuff |
2726 | { |
2727 | vec<tm_region *> *bb2reg; |
2728 | bool include_uninstrumented_p; |
2729 | }; |
2730 | |
2731 | // Callback for expand_regions, collect innermost region data for each bb. |
2732 | static void * |
2733 | collect_bb2reg (struct tm_region *region, void *data) |
2734 | { |
2735 | struct bb2reg_stuff *stuff = (struct bb2reg_stuff *)data; |
2736 | vec<tm_region *> *bb2reg = stuff->bb2reg; |
2737 | vec<basic_block> queue; |
2738 | unsigned int i; |
2739 | basic_block bb; |
2740 | |
2741 | queue = get_tm_region_blocks (entry_block: region->entry_block, |
2742 | exit_blocks: region->exit_blocks, |
2743 | irr_blocks: region->irr_blocks, |
2744 | NULL, |
2745 | /*stop_at_irr_p=*/stop_at_irrevocable_p: true, |
2746 | include_uninstrumented_p: stuff->include_uninstrumented_p); |
2747 | |
2748 | // We expect expand_region to perform a post-order traversal of the region |
2749 | // tree. Therefore the last region seen for any bb is the innermost. |
2750 | FOR_EACH_VEC_ELT (queue, i, bb) |
2751 | (*bb2reg)[bb->index] = region; |
2752 | |
2753 | queue.release (); |
2754 | return NULL; |
2755 | } |
2756 | |
2757 | // Returns a vector, indexed by BB->INDEX, of the innermost tm_region to |
2758 | // which a basic block belongs. Note that we only consider the instrumented |
2759 | // code paths for the region; the uninstrumented code paths are ignored if |
2760 | // INCLUDE_UNINSTRUMENTED_P is false. |
2761 | // |
2762 | // ??? This data is very similar to the bb_regions array that is collected |
2763 | // during tm_region_init. Or, rather, this data is similar to what could |
2764 | // be used within tm_region_init. The actual computation in tm_region_init |
2765 | // begins and ends with bb_regions entirely full of NULL pointers, due to |
2766 | // the way in which pointers are swapped in and out of the array. |
2767 | // |
2768 | // ??? Our callers expect that blocks are not shared between transactions. |
2769 | // When the optimizers get too smart, and blocks are shared, then during |
2770 | // the tm_mark phase we'll add log entries to only one of the two transactions, |
2771 | // and in the tm_edge phase we'll add edges to the CFG that create invalid |
2772 | // cycles. The symptom being SSA defs that do not dominate their uses. |
2773 | // Note that the optimizers were locally correct with their transformation, |
2774 | // as we have no info within the program that suggests that the blocks cannot |
2775 | // be shared. |
2776 | // |
2777 | // ??? There is currently a hack inside tree-ssa-pre.cc to work around the |
2778 | // only known instance of this block sharing. |
2779 | |
2780 | static vec<tm_region *> |
2781 | get_bb_regions_instrumented (bool traverse_clones, |
2782 | bool include_uninstrumented_p) |
2783 | { |
2784 | unsigned n = last_basic_block_for_fn (cfun); |
2785 | struct bb2reg_stuff stuff; |
2786 | vec<tm_region *> ret; |
2787 | |
2788 | ret.create (nelems: n); |
2789 | ret.safe_grow_cleared (len: n, exact: true); |
2790 | stuff.bb2reg = &ret; |
2791 | stuff.include_uninstrumented_p = include_uninstrumented_p; |
2792 | expand_regions (all_tm_regions, callback: collect_bb2reg, &stuff, traverse_clones); |
2793 | |
2794 | return ret; |
2795 | } |
2796 | |
2797 | /* Set the IN_TRANSACTION for all gimple statements that appear in a |
2798 | transaction. */ |
2799 | |
2800 | void |
2801 | compute_transaction_bits (void) |
2802 | { |
2803 | struct tm_region *region; |
2804 | vec<basic_block> queue; |
2805 | unsigned int i; |
2806 | basic_block bb; |
2807 | |
2808 | /* ?? Perhaps we need to abstract gate_tm_init further, because we |
2809 | certainly don't need it to calculate CDI_DOMINATOR info. */ |
2810 | gate_tm_init (); |
2811 | |
2812 | FOR_EACH_BB_FN (bb, cfun) |
2813 | bb->flags &= ~BB_IN_TRANSACTION; |
2814 | |
2815 | for (region = all_tm_regions; region; region = region->next) |
2816 | { |
2817 | queue = get_tm_region_blocks (entry_block: region->entry_block, |
2818 | exit_blocks: region->exit_blocks, |
2819 | irr_blocks: region->irr_blocks, |
2820 | NULL, |
2821 | /*stop_at_irr_p=*/stop_at_irrevocable_p: true); |
2822 | for (i = 0; queue.iterate (ix: i, ptr: &bb); ++i) |
2823 | bb->flags |= BB_IN_TRANSACTION; |
2824 | queue.release (); |
2825 | } |
2826 | |
2827 | if (all_tm_regions) |
2828 | bitmap_obstack_release (&tm_obstack); |
2829 | } |
2830 | |
2831 | /* Replace the GIMPLE_TRANSACTION in this region with the corresponding |
2832 | call to BUILT_IN_TM_START. */ |
2833 | |
2834 | static void * |
2835 | expand_transaction (struct tm_region *region, void *data ATTRIBUTE_UNUSED) |
2836 | { |
2837 | tree tm_start = builtin_decl_explicit (fncode: BUILT_IN_TM_START); |
2838 | basic_block transaction_bb = gimple_bb (g: region->transaction_stmt); |
2839 | tree tm_state = region->tm_state; |
2840 | tree tm_state_type = TREE_TYPE (tm_state); |
2841 | edge abort_edge = NULL; |
2842 | edge inst_edge = NULL; |
2843 | edge uninst_edge = NULL; |
2844 | edge fallthru_edge = NULL; |
2845 | |
2846 | // Identify the various successors of the transaction start. |
2847 | { |
2848 | edge_iterator i; |
2849 | edge e; |
2850 | FOR_EACH_EDGE (e, i, transaction_bb->succs) |
2851 | { |
2852 | if (e->flags & EDGE_TM_ABORT) |
2853 | abort_edge = e; |
2854 | else if (e->flags & EDGE_TM_UNINSTRUMENTED) |
2855 | uninst_edge = e; |
2856 | else |
2857 | inst_edge = e; |
2858 | if (e->flags & EDGE_FALLTHRU) |
2859 | fallthru_edge = e; |
2860 | } |
2861 | } |
2862 | |
2863 | /* ??? There are plenty of bits here we're not computing. */ |
2864 | { |
2865 | int subcode = gimple_transaction_subcode (transaction_stmt: region->get_transaction_stmt ()); |
2866 | int flags = 0; |
2867 | if (subcode & GTMA_DOES_GO_IRREVOCABLE) |
2868 | flags |= PR_DOESGOIRREVOCABLE; |
2869 | if ((subcode & GTMA_MAY_ENTER_IRREVOCABLE) == 0) |
2870 | flags |= PR_HASNOIRREVOCABLE; |
2871 | /* If the transaction does not have an abort in lexical scope and is not |
2872 | marked as an outer transaction, then it will never abort. */ |
2873 | if ((subcode & GTMA_HAVE_ABORT) == 0 && (subcode & GTMA_IS_OUTER) == 0) |
2874 | flags |= PR_HASNOABORT; |
2875 | if ((subcode & GTMA_HAVE_STORE) == 0) |
2876 | flags |= PR_READONLY; |
2877 | if (inst_edge && !(subcode & GTMA_HAS_NO_INSTRUMENTATION)) |
2878 | flags |= PR_INSTRUMENTEDCODE; |
2879 | if (uninst_edge) |
2880 | flags |= PR_UNINSTRUMENTEDCODE; |
2881 | if (subcode & GTMA_IS_OUTER) |
2882 | region->original_transaction_was_outer = true; |
2883 | tree t = build_int_cst (tm_state_type, flags); |
2884 | gcall *call = gimple_build_call (tm_start, 1, t); |
2885 | gimple_call_set_lhs (gs: call, lhs: tm_state); |
2886 | gimple_set_location (g: call, location: gimple_location (g: region->transaction_stmt)); |
2887 | |
2888 | // Replace the GIMPLE_TRANSACTION with the call to BUILT_IN_TM_START. |
2889 | gimple_stmt_iterator gsi = gsi_last_bb (bb: transaction_bb); |
2890 | gcc_assert (gsi_stmt (gsi) == region->transaction_stmt); |
2891 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); |
2892 | gsi_remove (&gsi, true); |
2893 | region->transaction_stmt = call; |
2894 | } |
2895 | |
2896 | // Generate log saves. |
2897 | if (!tm_log_save_addresses.is_empty ()) |
2898 | tm_log_emit_saves (entry_block: region->entry_block, bb: transaction_bb); |
2899 | |
2900 | // In the beginning, we've no tests to perform on transaction restart. |
2901 | // Note that after this point, transaction_bb becomes the "most recent |
2902 | // block containing tests for the transaction". |
2903 | region->restart_block = region->entry_block; |
2904 | |
2905 | // Generate log restores. |
2906 | if (!tm_log_save_addresses.is_empty ()) |
2907 | { |
2908 | basic_block test_bb = create_empty_bb (transaction_bb); |
2909 | basic_block code_bb = create_empty_bb (test_bb); |
2910 | basic_block join_bb = create_empty_bb (code_bb); |
2911 | add_bb_to_loop (test_bb, transaction_bb->loop_father); |
2912 | add_bb_to_loop (code_bb, transaction_bb->loop_father); |
2913 | add_bb_to_loop (join_bb, transaction_bb->loop_father); |
2914 | if (region->restart_block == region->entry_block) |
2915 | region->restart_block = test_bb; |
2916 | |
2917 | tree t1 = create_tmp_reg (tm_state_type); |
2918 | tree t2 = build_int_cst (tm_state_type, A_RESTORELIVEVARIABLES); |
2919 | gimple *stmt = gimple_build_assign (t1, BIT_AND_EXPR, tm_state, t2); |
2920 | gimple_stmt_iterator gsi = gsi_last_bb (bb: test_bb); |
2921 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
2922 | |
2923 | t2 = build_int_cst (tm_state_type, 0); |
2924 | stmt = gimple_build_cond (NE_EXPR, t1, t2, NULL, NULL); |
2925 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
2926 | |
2927 | tm_log_emit_restores (entry_block: region->entry_block, bb: code_bb); |
2928 | |
2929 | edge ei = make_edge (transaction_bb, test_bb, EDGE_FALLTHRU); |
2930 | edge et = make_edge (test_bb, code_bb, EDGE_TRUE_VALUE); |
2931 | edge ef = make_edge (test_bb, join_bb, EDGE_FALSE_VALUE); |
2932 | redirect_edge_pred (fallthru_edge, join_bb); |
2933 | |
2934 | join_bb->count = test_bb->count = transaction_bb->count; |
2935 | |
2936 | ei->probability = profile_probability::always (); |
2937 | et->probability = profile_probability::likely (); |
2938 | ef->probability = profile_probability::unlikely (); |
2939 | |
2940 | code_bb->count = et->count (); |
2941 | |
2942 | transaction_bb = join_bb; |
2943 | } |
2944 | |
2945 | // If we have an ABORT edge, create a test to perform the abort. |
2946 | if (abort_edge) |
2947 | { |
2948 | basic_block test_bb = create_empty_bb (transaction_bb); |
2949 | add_bb_to_loop (test_bb, transaction_bb->loop_father); |
2950 | if (region->restart_block == region->entry_block) |
2951 | region->restart_block = test_bb; |
2952 | |
2953 | tree t1 = create_tmp_reg (tm_state_type); |
2954 | tree t2 = build_int_cst (tm_state_type, A_ABORTTRANSACTION); |
2955 | gimple *stmt = gimple_build_assign (t1, BIT_AND_EXPR, tm_state, t2); |
2956 | gimple_stmt_iterator gsi = gsi_last_bb (bb: test_bb); |
2957 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
2958 | |
2959 | t2 = build_int_cst (tm_state_type, 0); |
2960 | stmt = gimple_build_cond (NE_EXPR, t1, t2, NULL, NULL); |
2961 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
2962 | |
2963 | edge ei = make_edge (transaction_bb, test_bb, EDGE_FALLTHRU); |
2964 | test_bb->count = transaction_bb->count; |
2965 | ei->probability = profile_probability::always (); |
2966 | |
2967 | // Not abort edge. If both are live, chose one at random as we'll |
2968 | // we'll be fixing that up below. |
2969 | redirect_edge_pred (fallthru_edge, test_bb); |
2970 | fallthru_edge->flags = EDGE_FALSE_VALUE; |
2971 | fallthru_edge->probability = profile_probability::very_likely (); |
2972 | |
2973 | // Abort/over edge. |
2974 | redirect_edge_pred (abort_edge, test_bb); |
2975 | abort_edge->flags = EDGE_TRUE_VALUE; |
2976 | abort_edge->probability = profile_probability::unlikely (); |
2977 | |
2978 | transaction_bb = test_bb; |
2979 | } |
2980 | |
2981 | // If we have both instrumented and uninstrumented code paths, select one. |
2982 | if (inst_edge && uninst_edge) |
2983 | { |
2984 | basic_block test_bb = create_empty_bb (transaction_bb); |
2985 | add_bb_to_loop (test_bb, transaction_bb->loop_father); |
2986 | if (region->restart_block == region->entry_block) |
2987 | region->restart_block = test_bb; |
2988 | |
2989 | tree t1 = create_tmp_reg (tm_state_type); |
2990 | tree t2 = build_int_cst (tm_state_type, A_RUNUNINSTRUMENTEDCODE); |
2991 | |
2992 | gimple *stmt = gimple_build_assign (t1, BIT_AND_EXPR, tm_state, t2); |
2993 | gimple_stmt_iterator gsi = gsi_last_bb (bb: test_bb); |
2994 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
2995 | |
2996 | t2 = build_int_cst (tm_state_type, 0); |
2997 | stmt = gimple_build_cond (NE_EXPR, t1, t2, NULL, NULL); |
2998 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); |
2999 | |
3000 | // Create the edge into test_bb first, as we want to copy values |
3001 | // out of the fallthru edge. |
3002 | edge e = make_edge (transaction_bb, test_bb, fallthru_edge->flags); |
3003 | e->probability = fallthru_edge->probability; |
3004 | test_bb->count = fallthru_edge->count (); |
3005 | |
3006 | // Now update the edges to the inst/uninist implementations. |
3007 | // For now assume that the paths are equally likely. When using HTM, |
3008 | // we'll try the uninst path first and fallback to inst path if htm |
3009 | // buffers are exceeded. Without HTM we start with the inst path and |
3010 | // use the uninst path when falling back to serial mode. |
3011 | redirect_edge_pred (inst_edge, test_bb); |
3012 | inst_edge->flags = EDGE_FALSE_VALUE; |
3013 | inst_edge->probability = profile_probability::even (); |
3014 | |
3015 | redirect_edge_pred (uninst_edge, test_bb); |
3016 | uninst_edge->flags = EDGE_TRUE_VALUE; |
3017 | uninst_edge->probability = profile_probability::even (); |
3018 | } |
3019 | |
3020 | // If we have no previous special cases, and we have PHIs at the beginning |
3021 | // of the atomic region, this means we have a loop at the beginning of the |
3022 | // atomic region that shares the first block. This can cause problems with |
3023 | // the transaction restart abnormal edges to be added in the tm_edges pass. |
3024 | // Solve this by adding a new empty block to receive the abnormal edges. |
3025 | if (region->restart_block == region->entry_block |
3026 | && phi_nodes (bb: region->entry_block)) |
3027 | { |
3028 | basic_block empty_bb = create_empty_bb (transaction_bb); |
3029 | region->restart_block = empty_bb; |
3030 | add_bb_to_loop (empty_bb, transaction_bb->loop_father); |
3031 | |
3032 | redirect_edge_pred (fallthru_edge, empty_bb); |
3033 | make_edge (transaction_bb, empty_bb, EDGE_FALLTHRU); |
3034 | } |
3035 | |
3036 | return NULL; |
3037 | } |
3038 | |
3039 | /* Generate the temporary to be used for the return value of |
3040 | BUILT_IN_TM_START. */ |
3041 | |
3042 | static void * |
3043 | generate_tm_state (struct tm_region *region, void *data ATTRIBUTE_UNUSED) |
3044 | { |
3045 | tree tm_start = builtin_decl_explicit (fncode: BUILT_IN_TM_START); |
3046 | region->tm_state = |
3047 | create_tmp_reg (TREE_TYPE (TREE_TYPE (tm_start)), "tm_state" ); |
3048 | |
3049 | // Reset the subcode, post optimizations. We'll fill this in |
3050 | // again as we process blocks. |
3051 | if (region->exit_blocks) |
3052 | { |
3053 | gtransaction *transaction_stmt = region->get_transaction_stmt (); |
3054 | unsigned int subcode = gimple_transaction_subcode (transaction_stmt); |
3055 | |
3056 | if (subcode & GTMA_DOES_GO_IRREVOCABLE) |
3057 | subcode &= (GTMA_DECLARATION_MASK | GTMA_DOES_GO_IRREVOCABLE |
3058 | | GTMA_MAY_ENTER_IRREVOCABLE |
3059 | | GTMA_HAS_NO_INSTRUMENTATION); |
3060 | else |
3061 | subcode &= GTMA_DECLARATION_MASK; |
3062 | gimple_transaction_set_subcode (transaction_stmt, subcode); |
3063 | } |
3064 | |
3065 | return NULL; |
3066 | } |
3067 | |
3068 | // Propagate flags from inner transactions outwards. |
3069 | static void |
3070 | propagate_tm_flags_out (struct tm_region *region) |
3071 | { |
3072 | if (region == NULL) |
3073 | return; |
3074 | propagate_tm_flags_out (region: region->inner); |
3075 | |
3076 | if (region->outer && region->outer->transaction_stmt) |
3077 | { |
3078 | unsigned s |
3079 | = gimple_transaction_subcode (transaction_stmt: region->get_transaction_stmt ()); |
3080 | s &= (GTMA_HAVE_ABORT | GTMA_HAVE_LOAD | GTMA_HAVE_STORE |
3081 | | GTMA_MAY_ENTER_IRREVOCABLE); |
3082 | s |= gimple_transaction_subcode (transaction_stmt: region->outer->get_transaction_stmt ()); |
3083 | gimple_transaction_set_subcode (transaction_stmt: region->outer->get_transaction_stmt (), |
3084 | subcode: s); |
3085 | } |
3086 | |
3087 | propagate_tm_flags_out (region: region->next); |
3088 | } |
3089 | |
3090 | /* Entry point to the MARK phase of TM expansion. Here we replace |
3091 | transactional memory statements with calls to builtins, and function |
3092 | calls with their transactional clones (if available). But we don't |
3093 | yet lower GIMPLE_TRANSACTION or add the transaction restart back-edges. */ |
3094 | |
3095 | static unsigned int |
3096 | execute_tm_mark (void) |
3097 | { |
3098 | pending_edge_inserts_p = false; |
3099 | |
3100 | expand_regions (all_tm_regions, callback: generate_tm_state, NULL, |
3101 | /*traverse_clones=*/true); |
3102 | |
3103 | tm_log_init (); |
3104 | |
3105 | vec<tm_region *> bb_regions |
3106 | = get_bb_regions_instrumented (/*traverse_clones=*/true, |
3107 | /*include_uninstrumented_p=*/false); |
3108 | struct tm_region *r; |
3109 | unsigned i; |
3110 | |
3111 | // Expand memory operations into calls into the runtime. |
3112 | // This collects log entries as well. |
3113 | FOR_EACH_VEC_ELT (bb_regions, i, r) |
3114 | { |
3115 | if (r != NULL) |
3116 | { |
3117 | if (r->transaction_stmt) |
3118 | { |
3119 | unsigned sub |
3120 | = gimple_transaction_subcode (transaction_stmt: r->get_transaction_stmt ()); |
3121 | |
3122 | /* If we're sure to go irrevocable, there won't be |
3123 | anything to expand, since the run-time will go |
3124 | irrevocable right away. */ |
3125 | if (sub & GTMA_DOES_GO_IRREVOCABLE |
3126 | && sub & GTMA_MAY_ENTER_IRREVOCABLE) |
3127 | continue; |
3128 | } |
3129 | expand_block_tm (region: r, BASIC_BLOCK_FOR_FN (cfun, i)); |
3130 | } |
3131 | } |
3132 | |
3133 | bb_regions.release (); |
3134 | |
3135 | // Propagate flags from inner transactions outwards. |
3136 | propagate_tm_flags_out (region: all_tm_regions); |
3137 | |
3138 | // Expand GIMPLE_TRANSACTIONs into calls into the runtime. |
3139 | expand_regions (all_tm_regions, callback: expand_transaction, NULL, |
3140 | /*traverse_clones=*/false); |
3141 | |
3142 | tm_log_emit (); |
3143 | tm_log_delete (); |
3144 | |
3145 | if (pending_edge_inserts_p) |
3146 | gsi_commit_edge_inserts (); |
3147 | free_dominance_info (CDI_DOMINATORS); |
3148 | return 0; |
3149 | } |
3150 | |
3151 | namespace { |
3152 | |
3153 | const pass_data pass_data_tm_mark = |
3154 | { |
3155 | .type: GIMPLE_PASS, /* type */ |
3156 | .name: "tmmark" , /* name */ |
3157 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
3158 | .tv_id: TV_TRANS_MEM, /* tv_id */ |
3159 | .properties_required: ( PROP_ssa | PROP_cfg ), /* properties_required */ |
3160 | .properties_provided: 0, /* properties_provided */ |
3161 | .properties_destroyed: 0, /* properties_destroyed */ |
3162 | .todo_flags_start: 0, /* todo_flags_start */ |
3163 | TODO_update_ssa, /* todo_flags_finish */ |
3164 | }; |
3165 | |
3166 | class pass_tm_mark : public gimple_opt_pass |
3167 | { |
3168 | public: |
3169 | pass_tm_mark (gcc::context *ctxt) |
3170 | : gimple_opt_pass (pass_data_tm_mark, ctxt) |
3171 | {} |
3172 | |
3173 | /* opt_pass methods: */ |
3174 | unsigned int execute (function *) final override |
3175 | { |
3176 | return execute_tm_mark (); |
3177 | } |
3178 | |
3179 | }; // class pass_tm_mark |
3180 | |
3181 | } // anon namespace |
3182 | |
3183 | gimple_opt_pass * |
3184 | make_pass_tm_mark (gcc::context *ctxt) |
3185 | { |
3186 | return new pass_tm_mark (ctxt); |
3187 | } |
3188 | |
3189 | |
3190 | /* Create an abnormal edge from STMT at iter, splitting the block |
3191 | as necessary. Adjust *PNEXT as needed for the split block. */ |
3192 | |
3193 | static inline void |
3194 | split_bb_make_tm_edge (gimple *stmt, basic_block dest_bb, |
3195 | gimple_stmt_iterator iter, gimple_stmt_iterator *pnext) |
3196 | { |
3197 | basic_block bb = gimple_bb (g: stmt); |
3198 | if (!gsi_one_before_end_p (i: iter)) |
3199 | { |
3200 | edge e = split_block (bb, stmt); |
3201 | *pnext = gsi_start_bb (bb: e->dest); |
3202 | } |
3203 | edge e = make_edge (bb, dest_bb, EDGE_ABNORMAL); |
3204 | if (e) |
3205 | e->probability = profile_probability::guessed_never (); |
3206 | |
3207 | // Record the need for the edge for the benefit of the rtl passes. |
3208 | if (cfun->gimple_df->tm_restart == NULL) |
3209 | cfun->gimple_df->tm_restart |
3210 | = hash_table<tm_restart_hasher>::create_ggc (n: 31); |
3211 | |
3212 | struct tm_restart_node dummy; |
3213 | dummy.stmt = stmt; |
3214 | dummy.label_or_list = gimple_block_label (dest_bb); |
3215 | |
3216 | tm_restart_node **slot = cfun->gimple_df->tm_restart->find_slot (value: &dummy, |
3217 | insert: INSERT); |
3218 | struct tm_restart_node *n = *slot; |
3219 | if (n == NULL) |
3220 | { |
3221 | *slot = n = ggc_alloc<tm_restart_node> (); |
3222 | *n = dummy; |
3223 | } |
3224 | else |
3225 | { |
3226 | tree old = n->label_or_list; |
3227 | if (TREE_CODE (old) == LABEL_DECL) |
3228 | old = tree_cons (NULL, old, NULL); |
3229 | n->label_or_list = tree_cons (NULL, dummy.label_or_list, old); |
3230 | } |
3231 | } |
3232 | |
3233 | /* Split block BB as necessary for every builtin function we added, and |
3234 | wire up the abnormal back edges implied by the transaction restart. */ |
3235 | |
3236 | static void |
3237 | expand_block_edges (struct tm_region *const region, basic_block bb) |
3238 | { |
3239 | gimple_stmt_iterator gsi, next_gsi; |
3240 | |
3241 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi = next_gsi) |
3242 | { |
3243 | gimple *stmt = gsi_stmt (i: gsi); |
3244 | gcall *call_stmt; |
3245 | |
3246 | next_gsi = gsi; |
3247 | gsi_next (i: &next_gsi); |
3248 | |
3249 | // ??? Shouldn't we split for any non-pure, non-irrevocable function? |
3250 | call_stmt = dyn_cast <gcall *> (p: stmt); |
3251 | if ((!call_stmt) |
3252 | || (gimple_call_flags (call_stmt) & ECF_TM_BUILTIN) == 0) |
3253 | continue; |
3254 | |
3255 | if (gimple_call_builtin_p (call_stmt, BUILT_IN_TM_ABORT)) |
3256 | { |
3257 | // If we have a ``_transaction_cancel [[outer]]'', there is only |
3258 | // one abnormal edge: to the transaction marked OUTER. |
3259 | // All compiler-generated instances of BUILT_IN_TM_ABORT have a |
3260 | // constant argument, which we can examine here. Users invoking |
3261 | // TM_ABORT directly get what they deserve. |
3262 | tree arg = gimple_call_arg (gs: call_stmt, index: 0); |
3263 | if (TREE_CODE (arg) == INTEGER_CST |
3264 | && (TREE_INT_CST_LOW (arg) & AR_OUTERABORT) != 0 |
3265 | && !decl_is_tm_clone (fndecl: current_function_decl)) |
3266 | { |
3267 | // Find the GTMA_IS_OUTER transaction. |
3268 | for (struct tm_region *o = region; o; o = o->outer) |
3269 | if (o->original_transaction_was_outer) |
3270 | { |
3271 | split_bb_make_tm_edge (stmt: call_stmt, dest_bb: o->restart_block, |
3272 | iter: gsi, pnext: &next_gsi); |
3273 | break; |
3274 | } |
3275 | |
3276 | // Otherwise, the front-end should have semantically checked |
3277 | // outer aborts, but in either case the target region is not |
3278 | // within this function. |
3279 | continue; |
3280 | } |
3281 | |
3282 | // Non-outer, TM aborts have an abnormal edge to the inner-most |
3283 | // transaction, the one being aborted; |
3284 | split_bb_make_tm_edge (stmt: call_stmt, dest_bb: region->restart_block, iter: gsi, |
3285 | pnext: &next_gsi); |
3286 | } |
3287 | |
3288 | // All TM builtins have an abnormal edge to the outer-most transaction. |
3289 | // We never restart inner transactions. For tm clones, we know a-priori |
3290 | // that the outer-most transaction is outside the function. |
3291 | if (decl_is_tm_clone (fndecl: current_function_decl)) |
3292 | continue; |
3293 | |
3294 | if (cfun->gimple_df->tm_restart == NULL) |
3295 | cfun->gimple_df->tm_restart |
3296 | = hash_table<tm_restart_hasher>::create_ggc (n: 31); |
3297 | |
3298 | // All TM builtins have an abnormal edge to the outer-most transaction. |
3299 | // We never restart inner transactions. |
3300 | for (struct tm_region *o = region; o; o = o->outer) |
3301 | if (!o->outer) |
3302 | { |
3303 | split_bb_make_tm_edge (stmt: call_stmt, dest_bb: o->restart_block, iter: gsi, pnext: &next_gsi); |
3304 | break; |
3305 | } |
3306 | |
3307 | // Delete any tail-call annotation that may have been added. |
3308 | // The tail-call pass may have mis-identified the commit as being |
3309 | // a candidate because we had not yet added this restart edge. |
3310 | gimple_call_set_tail (s: call_stmt, tail_p: false); |
3311 | } |
3312 | } |
3313 | |
3314 | /* Entry point to the final expansion of transactional nodes. */ |
3315 | |
3316 | namespace { |
3317 | |
3318 | const pass_data pass_data_tm_edges = |
3319 | { |
3320 | .type: GIMPLE_PASS, /* type */ |
3321 | .name: "tmedge" , /* name */ |
3322 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
3323 | .tv_id: TV_TRANS_MEM, /* tv_id */ |
3324 | .properties_required: ( PROP_ssa | PROP_cfg ), /* properties_required */ |
3325 | .properties_provided: 0, /* properties_provided */ |
3326 | .properties_destroyed: 0, /* properties_destroyed */ |
3327 | .todo_flags_start: 0, /* todo_flags_start */ |
3328 | TODO_update_ssa, /* todo_flags_finish */ |
3329 | }; |
3330 | |
3331 | class pass_tm_edges : public gimple_opt_pass |
3332 | { |
3333 | public: |
3334 | pass_tm_edges (gcc::context *ctxt) |
3335 | : gimple_opt_pass (pass_data_tm_edges, ctxt) |
3336 | {} |
3337 | |
3338 | /* opt_pass methods: */ |
3339 | unsigned int execute (function *) final override; |
3340 | |
3341 | }; // class pass_tm_edges |
3342 | |
3343 | unsigned int |
3344 | pass_tm_edges::execute (function *fun) |
3345 | { |
3346 | vec<tm_region *> bb_regions |
3347 | = get_bb_regions_instrumented (/*traverse_clones=*/false, |
3348 | /*include_uninstrumented_p=*/true); |
3349 | struct tm_region *r; |
3350 | unsigned i; |
3351 | |
3352 | FOR_EACH_VEC_ELT (bb_regions, i, r) |
3353 | if (r != NULL) |
3354 | expand_block_edges (region: r, BASIC_BLOCK_FOR_FN (fun, i)); |
3355 | |
3356 | bb_regions.release (); |
3357 | |
3358 | /* We've got to release the dominance info now, to indicate that it |
3359 | must be rebuilt completely. Otherwise we'll crash trying to update |
3360 | the SSA web in the TODO section following this pass. */ |
3361 | free_dominance_info (CDI_DOMINATORS); |
3362 | /* We'ge also wrecked loops badly with inserting of abnormal edges. */ |
3363 | loops_state_set (flags: LOOPS_NEED_FIXUP); |
3364 | bitmap_obstack_release (&tm_obstack); |
3365 | all_tm_regions = NULL; |
3366 | |
3367 | return 0; |
3368 | } |
3369 | |
3370 | } // anon namespace |
3371 | |
3372 | gimple_opt_pass * |
3373 | make_pass_tm_edges (gcc::context *ctxt) |
3374 | { |
3375 | return new pass_tm_edges (ctxt); |
3376 | } |
3377 | |
3378 | /* Helper function for expand_regions. Expand REGION and recurse to |
3379 | the inner region. Call CALLBACK on each region. CALLBACK returns |
3380 | NULL to continue the traversal, otherwise a non-null value which |
3381 | this function will return as well. TRAVERSE_CLONES is true if we |
3382 | should traverse transactional clones. */ |
3383 | |
3384 | static void * |
3385 | expand_regions_1 (struct tm_region *region, |
3386 | void *(*callback)(struct tm_region *, void *), |
3387 | void *data, |
3388 | bool traverse_clones) |
3389 | { |
3390 | void *retval = NULL; |
3391 | if (region->exit_blocks |
3392 | || (traverse_clones && decl_is_tm_clone (fndecl: current_function_decl))) |
3393 | { |
3394 | retval = callback (region, data); |
3395 | if (retval) |
3396 | return retval; |
3397 | } |
3398 | if (region->inner) |
3399 | { |
3400 | retval = expand_regions (region->inner, callback, data, traverse_clones); |
3401 | if (retval) |
3402 | return retval; |
3403 | } |
3404 | return retval; |
3405 | } |
3406 | |
3407 | /* Traverse the regions enclosed and including REGION. Execute |
3408 | CALLBACK for each region, passing DATA. CALLBACK returns NULL to |
3409 | continue the traversal, otherwise a non-null value which this |
3410 | function will return as well. TRAVERSE_CLONES is true if we should |
3411 | traverse transactional clones. */ |
3412 | |
3413 | static void * |
3414 | expand_regions (struct tm_region *region, |
3415 | void *(*callback)(struct tm_region *, void *), |
3416 | void *data, |
3417 | bool traverse_clones) |
3418 | { |
3419 | void *retval = NULL; |
3420 | while (region) |
3421 | { |
3422 | retval = expand_regions_1 (region, callback, data, traverse_clones); |
3423 | if (retval) |
3424 | return retval; |
3425 | region = region->next; |
3426 | } |
3427 | return retval; |
3428 | } |
3429 | |
3430 | |
3431 | /* A unique TM memory operation. */ |
3432 | struct tm_memop |
3433 | { |
3434 | /* Unique ID that all memory operations to the same location have. */ |
3435 | unsigned int value_id; |
3436 | /* Address of load/store. */ |
3437 | tree addr; |
3438 | }; |
3439 | |
3440 | /* TM memory operation hashtable helpers. */ |
3441 | |
3442 | struct tm_memop_hasher : free_ptr_hash <tm_memop> |
3443 | { |
3444 | static inline hashval_t hash (const tm_memop *); |
3445 | static inline bool equal (const tm_memop *, const tm_memop *); |
3446 | }; |
3447 | |
3448 | /* Htab support. Return a hash value for a `tm_memop'. */ |
3449 | inline hashval_t |
3450 | tm_memop_hasher::hash (const tm_memop *mem) |
3451 | { |
3452 | tree addr = mem->addr; |
3453 | /* We drill down to the SSA_NAME/DECL for the hash, but equality is |
3454 | actually done with operand_equal_p (see tm_memop_eq). */ |
3455 | if (TREE_CODE (addr) == ADDR_EXPR) |
3456 | addr = TREE_OPERAND (addr, 0); |
3457 | return iterative_hash_expr (tree: addr, seed: 0); |
3458 | } |
3459 | |
3460 | /* Htab support. Return true if two tm_memop's are the same. */ |
3461 | inline bool |
3462 | tm_memop_hasher::equal (const tm_memop *mem1, const tm_memop *mem2) |
3463 | { |
3464 | return operand_equal_p (mem1->addr, mem2->addr, flags: 0); |
3465 | } |
3466 | |
3467 | /* Sets for solving data flow equations in the memory optimization pass. */ |
3468 | struct tm_memopt_bitmaps |
3469 | { |
3470 | /* Stores available to this BB upon entry. Basically, stores that |
3471 | dominate this BB. */ |
3472 | bitmap store_avail_in; |
3473 | /* Stores available at the end of this BB. */ |
3474 | bitmap store_avail_out; |
3475 | bitmap store_antic_in; |
3476 | bitmap store_antic_out; |
3477 | /* Reads available to this BB upon entry. Basically, reads that |
3478 | dominate this BB. */ |
3479 | bitmap read_avail_in; |
3480 | /* Reads available at the end of this BB. */ |
3481 | bitmap read_avail_out; |
3482 | /* Reads performed in this BB. */ |
3483 | bitmap read_local; |
3484 | /* Writes performed in this BB. */ |
3485 | bitmap store_local; |
3486 | |
3487 | /* Temporary storage for pass. */ |
3488 | /* Is the current BB in the worklist? */ |
3489 | bool avail_in_worklist_p; |
3490 | /* Have we visited this BB? */ |
3491 | bool visited_p; |
3492 | }; |
3493 | |
3494 | static bitmap_obstack tm_memopt_obstack; |
3495 | |
3496 | /* Unique counter for TM loads and stores. Loads and stores of the |
3497 | same address get the same ID. */ |
3498 | static unsigned int tm_memopt_value_id; |
3499 | static hash_table<tm_memop_hasher> *tm_memopt_value_numbers; |
3500 | |
3501 | #define STORE_AVAIL_IN(BB) \ |
3502 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->store_avail_in |
3503 | #define STORE_AVAIL_OUT(BB) \ |
3504 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->store_avail_out |
3505 | #define STORE_ANTIC_IN(BB) \ |
3506 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->store_antic_in |
3507 | #define STORE_ANTIC_OUT(BB) \ |
3508 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->store_antic_out |
3509 | #define READ_AVAIL_IN(BB) \ |
3510 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->read_avail_in |
3511 | #define READ_AVAIL_OUT(BB) \ |
3512 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->read_avail_out |
3513 | #define READ_LOCAL(BB) \ |
3514 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->read_local |
3515 | #define STORE_LOCAL(BB) \ |
3516 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->store_local |
3517 | #define AVAIL_IN_WORKLIST_P(BB) \ |
3518 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->avail_in_worklist_p |
3519 | #define BB_VISITED_P(BB) \ |
3520 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->visited_p |
3521 | |
3522 | /* Given a TM load/store in STMT, return the value number for the address |
3523 | it accesses. */ |
3524 | |
3525 | static unsigned int |
3526 | tm_memopt_value_number (gimple *stmt, enum insert_option op) |
3527 | { |
3528 | struct tm_memop tmpmem, *mem; |
3529 | tm_memop **slot; |
3530 | |
3531 | gcc_assert (is_tm_load (stmt) || is_tm_store (stmt)); |
3532 | tmpmem.addr = gimple_call_arg (gs: stmt, index: 0); |
3533 | slot = tm_memopt_value_numbers->find_slot (value: &tmpmem, insert: op); |
3534 | if (*slot) |
3535 | mem = *slot; |
3536 | else if (op == INSERT) |
3537 | { |
3538 | mem = XNEW (struct tm_memop); |
3539 | *slot = mem; |
3540 | mem->value_id = tm_memopt_value_id++; |
3541 | mem->addr = tmpmem.addr; |
3542 | } |
3543 | else |
3544 | gcc_unreachable (); |
3545 | return mem->value_id; |
3546 | } |
3547 | |
3548 | /* Accumulate TM memory operations in BB into STORE_LOCAL and READ_LOCAL. */ |
3549 | |
3550 | static void |
3551 | tm_memopt_accumulate_memops (basic_block bb) |
3552 | { |
3553 | gimple_stmt_iterator gsi; |
3554 | |
3555 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
3556 | { |
3557 | gimple *stmt = gsi_stmt (i: gsi); |
3558 | bitmap bits; |
3559 | unsigned int loc; |
3560 | |
3561 | if (is_tm_store (stmt)) |
3562 | bits = STORE_LOCAL (bb); |
3563 | else if (is_tm_load (stmt)) |
3564 | bits = READ_LOCAL (bb); |
3565 | else |
3566 | continue; |
3567 | |
3568 | loc = tm_memopt_value_number (stmt, op: INSERT); |
3569 | bitmap_set_bit (bits, loc); |
3570 | if (dump_file) |
3571 | { |
3572 | fprintf (stream: dump_file, format: "TM memopt (%s): value num=%d, BB=%d, addr=" , |
3573 | is_tm_load (stmt) ? "LOAD" : "STORE" , loc, |
3574 | gimple_bb (g: stmt)->index); |
3575 | print_generic_expr (dump_file, gimple_call_arg (gs: stmt, index: 0)); |
3576 | fprintf (stream: dump_file, format: "\n" ); |
3577 | } |
3578 | } |
3579 | } |
3580 | |
3581 | /* Prettily dump one of the memopt sets. BITS is the bitmap to dump. */ |
3582 | |
3583 | static void |
3584 | dump_tm_memopt_set (const char *set_name, bitmap bits) |
3585 | { |
3586 | unsigned i; |
3587 | bitmap_iterator bi; |
3588 | const char *comma = "" ; |
3589 | |
3590 | fprintf (stream: dump_file, format: "TM memopt: %s: [" , set_name); |
3591 | EXECUTE_IF_SET_IN_BITMAP (bits, 0, i, bi) |
3592 | { |
3593 | hash_table<tm_memop_hasher>::iterator hi; |
3594 | struct tm_memop *mem = NULL; |
3595 | |
3596 | /* Yeah, yeah, yeah. Whatever. This is just for debugging. */ |
3597 | FOR_EACH_HASH_TABLE_ELEMENT (*tm_memopt_value_numbers, mem, tm_memop_t, hi) |
3598 | if (mem->value_id == i) |
3599 | break; |
3600 | gcc_assert (mem->value_id == i); |
3601 | fprintf (stream: dump_file, format: "%s" , comma); |
3602 | comma = ", " ; |
3603 | print_generic_expr (dump_file, mem->addr); |
3604 | } |
3605 | fprintf (stream: dump_file, format: "]\n" ); |
3606 | } |
3607 | |
3608 | /* Prettily dump all of the memopt sets in BLOCKS. */ |
3609 | |
3610 | static void |
3611 | dump_tm_memopt_sets (vec<basic_block> blocks) |
3612 | { |
3613 | size_t i; |
3614 | basic_block bb; |
3615 | |
3616 | for (i = 0; blocks.iterate (ix: i, ptr: &bb); ++i) |
3617 | { |
3618 | fprintf (stream: dump_file, format: "------------BB %d---------\n" , bb->index); |
3619 | dump_tm_memopt_set (set_name: "STORE_LOCAL" , STORE_LOCAL (bb)); |
3620 | dump_tm_memopt_set (set_name: "READ_LOCAL" , READ_LOCAL (bb)); |
3621 | dump_tm_memopt_set (set_name: "STORE_AVAIL_IN" , STORE_AVAIL_IN (bb)); |
3622 | dump_tm_memopt_set (set_name: "STORE_AVAIL_OUT" , STORE_AVAIL_OUT (bb)); |
3623 | dump_tm_memopt_set (set_name: "READ_AVAIL_IN" , READ_AVAIL_IN (bb)); |
3624 | dump_tm_memopt_set (set_name: "READ_AVAIL_OUT" , READ_AVAIL_OUT (bb)); |
3625 | } |
3626 | } |
3627 | |
3628 | /* Compute {STORE,READ}_AVAIL_IN for the basic block BB. */ |
3629 | |
3630 | static void |
3631 | tm_memopt_compute_avin (basic_block bb) |
3632 | { |
3633 | edge e; |
3634 | unsigned ix; |
3635 | |
3636 | /* Seed with the AVOUT of any predecessor. */ |
3637 | for (ix = 0; ix < EDGE_COUNT (bb->preds); ix++) |
3638 | { |
3639 | e = EDGE_PRED (bb, ix); |
3640 | /* Make sure we have already visited this BB, and is thus |
3641 | initialized. |
3642 | |
3643 | If e->src->aux is NULL, this predecessor is actually on an |
3644 | enclosing transaction. We only care about the current |
3645 | transaction, so ignore it. */ |
3646 | if (e->src->aux && BB_VISITED_P (e->src)) |
3647 | { |
3648 | bitmap_copy (STORE_AVAIL_IN (bb), STORE_AVAIL_OUT (e->src)); |
3649 | bitmap_copy (READ_AVAIL_IN (bb), READ_AVAIL_OUT (e->src)); |
3650 | break; |
3651 | } |
3652 | } |
3653 | |
3654 | for (; ix < EDGE_COUNT (bb->preds); ix++) |
3655 | { |
3656 | e = EDGE_PRED (bb, ix); |
3657 | if (e->src->aux && BB_VISITED_P (e->src)) |
3658 | { |
3659 | bitmap_and_into (STORE_AVAIL_IN (bb), STORE_AVAIL_OUT (e->src)); |
3660 | bitmap_and_into (READ_AVAIL_IN (bb), READ_AVAIL_OUT (e->src)); |
3661 | } |
3662 | } |
3663 | |
3664 | BB_VISITED_P (bb) = true; |
3665 | } |
3666 | |
3667 | /* Compute the STORE_ANTIC_IN for the basic block BB. */ |
3668 | |
3669 | static void |
3670 | tm_memopt_compute_antin (basic_block bb) |
3671 | { |
3672 | edge e; |
3673 | unsigned ix; |
3674 | |
3675 | /* Seed with the ANTIC_OUT of any successor. */ |
3676 | for (ix = 0; ix < EDGE_COUNT (bb->succs); ix++) |
3677 | { |
3678 | e = EDGE_SUCC (bb, ix); |
3679 | /* Make sure we have already visited this BB, and is thus |
3680 | initialized. */ |
3681 | if (BB_VISITED_P (e->dest)) |
3682 | { |
3683 | bitmap_copy (STORE_ANTIC_IN (bb), STORE_ANTIC_OUT (e->dest)); |
3684 | break; |
3685 | } |
3686 | } |
3687 | |
3688 | for (; ix < EDGE_COUNT (bb->succs); ix++) |
3689 | { |
3690 | e = EDGE_SUCC (bb, ix); |
3691 | if (BB_VISITED_P (e->dest)) |
3692 | bitmap_and_into (STORE_ANTIC_IN (bb), STORE_ANTIC_OUT (e->dest)); |
3693 | } |
3694 | |
3695 | BB_VISITED_P (bb) = true; |
3696 | } |
3697 | |
3698 | /* Compute the AVAIL sets for every basic block in BLOCKS. |
3699 | |
3700 | We compute {STORE,READ}_AVAIL_{OUT,IN} as follows: |
3701 | |
3702 | AVAIL_OUT[bb] = union (AVAIL_IN[bb], LOCAL[bb]) |
3703 | AVAIL_IN[bb] = intersect (AVAIL_OUT[predecessors]) |
3704 | |
3705 | This is basically what we do in lcm's compute_available(), but here |
3706 | we calculate two sets of sets (one for STOREs and one for READs), |
3707 | and we work on a region instead of the entire CFG. |
3708 | |
3709 | REGION is the TM region. |
3710 | BLOCKS are the basic blocks in the region. */ |
3711 | |
3712 | static void |
3713 | tm_memopt_compute_available (struct tm_region *region, |
3714 | vec<basic_block> blocks) |
3715 | { |
3716 | edge e; |
3717 | basic_block *worklist, *qin, *qout, *qend, bb; |
3718 | unsigned int qlen, i; |
3719 | edge_iterator ei; |
3720 | bool changed; |
3721 | |
3722 | /* Allocate a worklist array/queue. Entries are only added to the |
3723 | list if they were not already on the list. So the size is |
3724 | bounded by the number of basic blocks in the region. */ |
3725 | gcc_assert (!blocks.is_empty ()); |
3726 | qlen = blocks.length () - 1; |
3727 | qin = qout = worklist = XNEWVEC (basic_block, qlen); |
3728 | |
3729 | /* Put every block in the region on the worklist. */ |
3730 | for (i = 0; blocks.iterate (ix: i, ptr: &bb); ++i) |
3731 | { |
3732 | /* Seed AVAIL_OUT with the LOCAL set. */ |
3733 | bitmap_ior_into (STORE_AVAIL_OUT (bb), STORE_LOCAL (bb)); |
3734 | bitmap_ior_into (READ_AVAIL_OUT (bb), READ_LOCAL (bb)); |
3735 | |
3736 | AVAIL_IN_WORKLIST_P (bb) = true; |
3737 | /* No need to insert the entry block, since it has an AVIN of |
3738 | null, and an AVOUT that has already been seeded in. */ |
3739 | if (bb != region->entry_block) |
3740 | *qin++ = bb; |
3741 | } |
3742 | |
3743 | /* The entry block has been initialized with the local sets. */ |
3744 | BB_VISITED_P (region->entry_block) = true; |
3745 | |
3746 | qin = worklist; |
3747 | qend = &worklist[qlen]; |
3748 | |
3749 | /* Iterate until the worklist is empty. */ |
3750 | while (qlen) |
3751 | { |
3752 | /* Take the first entry off the worklist. */ |
3753 | bb = *qout++; |
3754 | qlen--; |
3755 | |
3756 | if (qout >= qend) |
3757 | qout = worklist; |
3758 | |
3759 | /* This block can be added to the worklist again if necessary. */ |
3760 | AVAIL_IN_WORKLIST_P (bb) = false; |
3761 | tm_memopt_compute_avin (bb); |
3762 | |
3763 | /* Note: We do not add the LOCAL sets here because we already |
3764 | seeded the AVAIL_OUT sets with them. */ |
3765 | changed = bitmap_ior_into (STORE_AVAIL_OUT (bb), STORE_AVAIL_IN (bb)); |
3766 | changed |= bitmap_ior_into (READ_AVAIL_OUT (bb), READ_AVAIL_IN (bb)); |
3767 | if (changed |
3768 | && (region->exit_blocks == NULL |
3769 | || !bitmap_bit_p (region->exit_blocks, bb->index))) |
3770 | /* If the out state of this block changed, then we need to add |
3771 | its successors to the worklist if they are not already in. */ |
3772 | FOR_EACH_EDGE (e, ei, bb->succs) |
3773 | if (!AVAIL_IN_WORKLIST_P (e->dest) |
3774 | && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
3775 | { |
3776 | *qin++ = e->dest; |
3777 | AVAIL_IN_WORKLIST_P (e->dest) = true; |
3778 | qlen++; |
3779 | |
3780 | if (qin >= qend) |
3781 | qin = worklist; |
3782 | } |
3783 | } |
3784 | |
3785 | free (ptr: worklist); |
3786 | |
3787 | if (dump_file) |
3788 | dump_tm_memopt_sets (blocks); |
3789 | } |
3790 | |
3791 | /* Compute ANTIC sets for every basic block in BLOCKS. |
3792 | |
3793 | We compute STORE_ANTIC_OUT as follows: |
3794 | |
3795 | STORE_ANTIC_OUT[bb] = union(STORE_ANTIC_IN[bb], STORE_LOCAL[bb]) |
3796 | STORE_ANTIC_IN[bb] = intersect(STORE_ANTIC_OUT[successors]) |
3797 | |
3798 | REGION is the TM region. |
3799 | BLOCKS are the basic blocks in the region. */ |
3800 | |
3801 | static void |
3802 | tm_memopt_compute_antic (struct tm_region *region, |
3803 | vec<basic_block> blocks) |
3804 | { |
3805 | edge e; |
3806 | basic_block *worklist, *qin, *qout, *qend, bb; |
3807 | unsigned int qlen; |
3808 | int i; |
3809 | edge_iterator ei; |
3810 | |
3811 | /* Allocate a worklist array/queue. Entries are only added to the |
3812 | list if they were not already on the list. So the size is |
3813 | bounded by the number of basic blocks in the region. */ |
3814 | qin = qout = worklist = XNEWVEC (basic_block, blocks.length ()); |
3815 | |
3816 | for (qlen = 0, i = blocks.length () - 1; i >= 0; --i) |
3817 | { |
3818 | bb = blocks[i]; |
3819 | |
3820 | /* Seed ANTIC_OUT with the LOCAL set. */ |
3821 | bitmap_ior_into (STORE_ANTIC_OUT (bb), STORE_LOCAL (bb)); |
3822 | |
3823 | /* Put every block in the region on the worklist. */ |
3824 | AVAIL_IN_WORKLIST_P (bb) = true; |
3825 | /* No need to insert exit blocks, since their ANTIC_IN is NULL, |
3826 | and their ANTIC_OUT has already been seeded in. */ |
3827 | if (region->exit_blocks |
3828 | && !bitmap_bit_p (region->exit_blocks, bb->index)) |
3829 | { |
3830 | qlen++; |
3831 | *qin++ = bb; |
3832 | } |
3833 | } |
3834 | |
3835 | /* The exit blocks have been initialized with the local sets. */ |
3836 | if (region->exit_blocks) |
3837 | { |
3838 | unsigned int i; |
3839 | bitmap_iterator bi; |
3840 | EXECUTE_IF_SET_IN_BITMAP (region->exit_blocks, 0, i, bi) |
3841 | BB_VISITED_P (BASIC_BLOCK_FOR_FN (cfun, i)) = true; |
3842 | } |
3843 | |
3844 | qin = worklist; |
3845 | qend = &worklist[qlen]; |
3846 | |
3847 | /* Iterate until the worklist is empty. */ |
3848 | while (qlen) |
3849 | { |
3850 | /* Take the first entry off the worklist. */ |
3851 | bb = *qout++; |
3852 | qlen--; |
3853 | |
3854 | if (qout >= qend) |
3855 | qout = worklist; |
3856 | |
3857 | /* This block can be added to the worklist again if necessary. */ |
3858 | AVAIL_IN_WORKLIST_P (bb) = false; |
3859 | tm_memopt_compute_antin (bb); |
3860 | |
3861 | /* Note: We do not add the LOCAL sets here because we already |
3862 | seeded the ANTIC_OUT sets with them. */ |
3863 | if (bitmap_ior_into (STORE_ANTIC_OUT (bb), STORE_ANTIC_IN (bb)) |
3864 | && bb != region->entry_block) |
3865 | /* If the out state of this block changed, then we need to add |
3866 | its predecessors to the worklist if they are not already in. */ |
3867 | FOR_EACH_EDGE (e, ei, bb->preds) |
3868 | if (!AVAIL_IN_WORKLIST_P (e->src)) |
3869 | { |
3870 | *qin++ = e->src; |
3871 | AVAIL_IN_WORKLIST_P (e->src) = true; |
3872 | qlen++; |
3873 | |
3874 | if (qin >= qend) |
3875 | qin = worklist; |
3876 | } |
3877 | } |
3878 | |
3879 | free (ptr: worklist); |
3880 | |
3881 | if (dump_file) |
3882 | dump_tm_memopt_sets (blocks); |
3883 | } |
3884 | |
3885 | /* Offsets of load variants from TM_LOAD. For example, |
3886 | BUILT_IN_TM_LOAD_RAR* is an offset of 1 from BUILT_IN_TM_LOAD*. |
3887 | See gtm-builtins.def. */ |
3888 | #define TRANSFORM_RAR 1 |
3889 | #define TRANSFORM_RAW 2 |
3890 | #define TRANSFORM_RFW 3 |
3891 | /* Offsets of store variants from TM_STORE. */ |
3892 | #define TRANSFORM_WAR 1 |
3893 | #define TRANSFORM_WAW 2 |
3894 | |
3895 | /* Inform about a load/store optimization. */ |
3896 | |
3897 | static void |
3898 | dump_tm_memopt_transform (gimple *stmt) |
3899 | { |
3900 | if (dump_file) |
3901 | { |
3902 | fprintf (stream: dump_file, format: "TM memopt: transforming: " ); |
3903 | print_gimple_stmt (dump_file, stmt, 0); |
3904 | fprintf (stream: dump_file, format: "\n" ); |
3905 | } |
3906 | } |
3907 | |
3908 | /* Perform a read/write optimization. Replaces the TM builtin in STMT |
3909 | by a builtin that is OFFSET entries down in the builtins table in |
3910 | gtm-builtins.def. */ |
3911 | |
3912 | static void |
3913 | tm_memopt_transform_stmt (unsigned int offset, |
3914 | gcall *stmt, |
3915 | gimple_stmt_iterator *gsi) |
3916 | { |
3917 | tree fn = gimple_call_fn (gs: stmt); |
3918 | gcc_assert (TREE_CODE (fn) == ADDR_EXPR); |
3919 | TREE_OPERAND (fn, 0) |
3920 | = builtin_decl_explicit (fncode: (enum built_in_function) |
3921 | (DECL_FUNCTION_CODE (TREE_OPERAND (fn, 0)) |
3922 | + offset)); |
3923 | gimple_call_set_fn (gs: stmt, fn); |
3924 | gsi_replace (gsi, stmt, true); |
3925 | dump_tm_memopt_transform (stmt); |
3926 | } |
3927 | |
3928 | /* Perform the actual TM memory optimization transformations in the |
3929 | basic blocks in BLOCKS. */ |
3930 | |
3931 | static void |
3932 | tm_memopt_transform_blocks (vec<basic_block> blocks) |
3933 | { |
3934 | size_t i; |
3935 | basic_block bb; |
3936 | gimple_stmt_iterator gsi; |
3937 | |
3938 | for (i = 0; blocks.iterate (ix: i, ptr: &bb); ++i) |
3939 | { |
3940 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
3941 | { |
3942 | gimple *stmt = gsi_stmt (i: gsi); |
3943 | bitmap read_avail = READ_AVAIL_IN (bb); |
3944 | bitmap store_avail = STORE_AVAIL_IN (bb); |
3945 | bitmap store_antic = STORE_ANTIC_OUT (bb); |
3946 | unsigned int loc; |
3947 | |
3948 | if (is_tm_simple_load (stmt)) |
3949 | { |
3950 | gcall *call_stmt = as_a <gcall *> (p: stmt); |
3951 | loc = tm_memopt_value_number (stmt, op: NO_INSERT); |
3952 | if (store_avail && bitmap_bit_p (store_avail, loc)) |
3953 | tm_memopt_transform_stmt (TRANSFORM_RAW, stmt: call_stmt, gsi: &gsi); |
3954 | else if (store_antic && bitmap_bit_p (store_antic, loc)) |
3955 | { |
3956 | tm_memopt_transform_stmt (TRANSFORM_RFW, stmt: call_stmt, gsi: &gsi); |
3957 | bitmap_set_bit (store_avail, loc); |
3958 | } |
3959 | else if (read_avail && bitmap_bit_p (read_avail, loc)) |
3960 | tm_memopt_transform_stmt (TRANSFORM_RAR, stmt: call_stmt, gsi: &gsi); |
3961 | else |
3962 | bitmap_set_bit (read_avail, loc); |
3963 | } |
3964 | else if (is_tm_simple_store (stmt)) |
3965 | { |
3966 | gcall *call_stmt = as_a <gcall *> (p: stmt); |
3967 | loc = tm_memopt_value_number (stmt, op: NO_INSERT); |
3968 | if (store_avail && bitmap_bit_p (store_avail, loc)) |
3969 | tm_memopt_transform_stmt (TRANSFORM_WAW, stmt: call_stmt, gsi: &gsi); |
3970 | else |
3971 | { |
3972 | if (read_avail && bitmap_bit_p (read_avail, loc)) |
3973 | tm_memopt_transform_stmt (TRANSFORM_WAR, stmt: call_stmt, gsi: &gsi); |
3974 | bitmap_set_bit (store_avail, loc); |
3975 | } |
3976 | } |
3977 | } |
3978 | } |
3979 | } |
3980 | |
3981 | /* Return a new set of bitmaps for a BB. */ |
3982 | |
3983 | static struct tm_memopt_bitmaps * |
3984 | tm_memopt_init_sets (void) |
3985 | { |
3986 | struct tm_memopt_bitmaps *b |
3987 | = XOBNEW (&tm_memopt_obstack.obstack, struct tm_memopt_bitmaps); |
3988 | b->store_avail_in = BITMAP_ALLOC (obstack: &tm_memopt_obstack); |
3989 | b->store_avail_out = BITMAP_ALLOC (obstack: &tm_memopt_obstack); |
3990 | b->store_antic_in = BITMAP_ALLOC (obstack: &tm_memopt_obstack); |
3991 | b->store_antic_out = BITMAP_ALLOC (obstack: &tm_memopt_obstack); |
3992 | b->store_avail_out = BITMAP_ALLOC (obstack: &tm_memopt_obstack); |
3993 | b->read_avail_in = BITMAP_ALLOC (obstack: &tm_memopt_obstack); |
3994 | b->read_avail_out = BITMAP_ALLOC (obstack: &tm_memopt_obstack); |
3995 | b->read_local = BITMAP_ALLOC (obstack: &tm_memopt_obstack); |
3996 | b->store_local = BITMAP_ALLOC (obstack: &tm_memopt_obstack); |
3997 | return b; |
3998 | } |
3999 | |
4000 | /* Free sets computed for each BB. */ |
4001 | |
4002 | static void |
4003 | tm_memopt_free_sets (vec<basic_block> blocks) |
4004 | { |
4005 | size_t i; |
4006 | basic_block bb; |
4007 | |
4008 | for (i = 0; blocks.iterate (ix: i, ptr: &bb); ++i) |
4009 | bb->aux = NULL; |
4010 | } |
4011 | |
4012 | /* Clear the visited bit for every basic block in BLOCKS. */ |
4013 | |
4014 | static void |
4015 | tm_memopt_clear_visited (vec<basic_block> blocks) |
4016 | { |
4017 | size_t i; |
4018 | basic_block bb; |
4019 | |
4020 | for (i = 0; blocks.iterate (ix: i, ptr: &bb); ++i) |
4021 | BB_VISITED_P (bb) = false; |
4022 | } |
4023 | |
4024 | /* Replace TM load/stores with hints for the runtime. We handle |
4025 | things like read-after-write, write-after-read, read-after-read, |
4026 | read-for-write, etc. */ |
4027 | |
4028 | static unsigned int |
4029 | execute_tm_memopt (void) |
4030 | { |
4031 | struct tm_region *region; |
4032 | vec<basic_block> bbs; |
4033 | |
4034 | tm_memopt_value_id = 0; |
4035 | tm_memopt_value_numbers = new hash_table<tm_memop_hasher> (10); |
4036 | |
4037 | for (region = all_tm_regions; region; region = region->next) |
4038 | { |
4039 | /* All the TM stores/loads in the current region. */ |
4040 | size_t i; |
4041 | basic_block bb; |
4042 | |
4043 | bitmap_obstack_initialize (&tm_memopt_obstack); |
4044 | |
4045 | /* Save all BBs for the current region. */ |
4046 | bbs = get_tm_region_blocks (entry_block: region->entry_block, |
4047 | exit_blocks: region->exit_blocks, |
4048 | irr_blocks: region->irr_blocks, |
4049 | NULL, |
4050 | stop_at_irrevocable_p: false); |
4051 | |
4052 | /* Collect all the memory operations. */ |
4053 | for (i = 0; bbs.iterate (ix: i, ptr: &bb); ++i) |
4054 | { |
4055 | bb->aux = tm_memopt_init_sets (); |
4056 | tm_memopt_accumulate_memops (bb); |
4057 | } |
4058 | |
4059 | /* Solve data flow equations and transform each block accordingly. */ |
4060 | tm_memopt_clear_visited (blocks: bbs); |
4061 | tm_memopt_compute_available (region, blocks: bbs); |
4062 | tm_memopt_clear_visited (blocks: bbs); |
4063 | tm_memopt_compute_antic (region, blocks: bbs); |
4064 | tm_memopt_transform_blocks (blocks: bbs); |
4065 | |
4066 | tm_memopt_free_sets (blocks: bbs); |
4067 | bbs.release (); |
4068 | bitmap_obstack_release (&tm_memopt_obstack); |
4069 | tm_memopt_value_numbers->empty (); |
4070 | } |
4071 | |
4072 | delete tm_memopt_value_numbers; |
4073 | tm_memopt_value_numbers = NULL; |
4074 | return 0; |
4075 | } |
4076 | |
4077 | namespace { |
4078 | |
4079 | const pass_data pass_data_tm_memopt = |
4080 | { |
4081 | .type: GIMPLE_PASS, /* type */ |
4082 | .name: "tmmemopt" , /* name */ |
4083 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
4084 | .tv_id: TV_TRANS_MEM, /* tv_id */ |
4085 | .properties_required: ( PROP_ssa | PROP_cfg ), /* properties_required */ |
4086 | .properties_provided: 0, /* properties_provided */ |
4087 | .properties_destroyed: 0, /* properties_destroyed */ |
4088 | .todo_flags_start: 0, /* todo_flags_start */ |
4089 | .todo_flags_finish: 0, /* todo_flags_finish */ |
4090 | }; |
4091 | |
4092 | class pass_tm_memopt : public gimple_opt_pass |
4093 | { |
4094 | public: |
4095 | pass_tm_memopt (gcc::context *ctxt) |
4096 | : gimple_opt_pass (pass_data_tm_memopt, ctxt) |
4097 | {} |
4098 | |
4099 | /* opt_pass methods: */ |
4100 | bool gate (function *) final override { return flag_tm && optimize > 0; } |
4101 | unsigned int execute (function *) final override |
4102 | { |
4103 | return execute_tm_memopt (); |
4104 | } |
4105 | |
4106 | }; // class pass_tm_memopt |
4107 | |
4108 | } // anon namespace |
4109 | |
4110 | gimple_opt_pass * |
4111 | make_pass_tm_memopt (gcc::context *ctxt) |
4112 | { |
4113 | return new pass_tm_memopt (ctxt); |
4114 | } |
4115 | |
4116 | |
4117 | /* Interprocedual analysis for the creation of transactional clones. |
4118 | The aim of this pass is to find which functions are referenced in |
4119 | a non-irrevocable transaction context, and for those over which |
4120 | we have control (or user directive), create a version of the |
4121 | function which uses only the transactional interface to reference |
4122 | protected memories. This analysis proceeds in several steps: |
4123 | |
4124 | (1) Collect the set of all possible transactional clones: |
4125 | |
4126 | (a) For all local public functions marked tm_callable, push |
4127 | it onto the tm_callee queue. |
4128 | |
4129 | (b) For all local functions, scan for calls in transaction blocks. |
4130 | Push the caller and callee onto the tm_caller and tm_callee |
4131 | queues. Count the number of callers for each callee. |
4132 | |
4133 | (c) For each local function on the callee list, assume we will |
4134 | create a transactional clone. Push *all* calls onto the |
4135 | callee queues; count the number of clone callers separately |
4136 | to the number of original callers. |
4137 | |
4138 | (2) Propagate irrevocable status up the dominator tree: |
4139 | |
4140 | (a) Any external function on the callee list that is not marked |
4141 | tm_callable is irrevocable. Push all callers of such onto |
4142 | a worklist. |
4143 | |
4144 | (b) For each function on the worklist, mark each block that |
4145 | contains an irrevocable call. Use the AND operator to |
4146 | propagate that mark up the dominator tree. |
4147 | |
4148 | (c) If we reach the entry block for a possible transactional |
4149 | clone, then the transactional clone is irrevocable, and |
4150 | we should not create the clone after all. Push all |
4151 | callers onto the worklist. |
4152 | |
4153 | (d) Place tm_irrevocable calls at the beginning of the relevant |
4154 | blocks. Special case here is the entry block for the entire |
4155 | transaction region; there we mark it GTMA_DOES_GO_IRREVOCABLE for |
4156 | the library to begin the region in serial mode. Decrement |
4157 | the call count for all callees in the irrevocable region. |
4158 | |
4159 | (3) Create the transactional clones: |
4160 | |
4161 | Any tm_callee that still has a non-zero call count is cloned. |
4162 | */ |
4163 | |
4164 | /* This structure is stored in the AUX field of each cgraph_node. */ |
4165 | struct tm_ipa_cg_data |
4166 | { |
4167 | /* The clone of the function that got created. */ |
4168 | struct cgraph_node *clone; |
4169 | |
4170 | /* The tm regions in the normal function. */ |
4171 | struct tm_region *all_tm_regions; |
4172 | |
4173 | /* The blocks of the normal/clone functions that contain irrevocable |
4174 | calls, or blocks that are post-dominated by irrevocable calls. */ |
4175 | bitmap irrevocable_blocks_normal; |
4176 | bitmap irrevocable_blocks_clone; |
4177 | |
4178 | /* The blocks of the normal function that are involved in transactions. */ |
4179 | bitmap transaction_blocks_normal; |
4180 | |
4181 | /* The number of callers to the transactional clone of this function |
4182 | from normal and transactional clones respectively. */ |
4183 | unsigned tm_callers_normal; |
4184 | unsigned tm_callers_clone; |
4185 | |
4186 | /* True if all calls to this function's transactional clone |
4187 | are irrevocable. Also automatically true if the function |
4188 | has no transactional clone. */ |
4189 | bool is_irrevocable; |
4190 | |
4191 | /* Flags indicating the presence of this function in various queues. */ |
4192 | bool in_callee_queue; |
4193 | bool in_worklist; |
4194 | |
4195 | /* Flags indicating the kind of scan desired while in the worklist. */ |
4196 | bool want_irr_scan_normal; |
4197 | }; |
4198 | |
4199 | typedef vec<cgraph_node *> cgraph_node_queue; |
4200 | |
4201 | /* Return the ipa data associated with NODE, allocating zeroed memory |
4202 | if necessary. TRAVERSE_ALIASES is true if we must traverse aliases |
4203 | and set *NODE accordingly. */ |
4204 | |
4205 | static struct tm_ipa_cg_data * |
4206 | get_cg_data (struct cgraph_node **node, bool traverse_aliases) |
4207 | { |
4208 | struct tm_ipa_cg_data *d; |
4209 | |
4210 | if (traverse_aliases && (*node)->alias) |
4211 | *node = (*node)->get_alias_target (); |
4212 | |
4213 | d = (struct tm_ipa_cg_data *) (*node)->aux; |
4214 | |
4215 | if (d == NULL) |
4216 | { |
4217 | d = (struct tm_ipa_cg_data *) |
4218 | obstack_alloc (&tm_obstack.obstack, sizeof (*d)); |
4219 | (*node)->aux = (void *) d; |
4220 | memset (s: d, c: 0, n: sizeof (*d)); |
4221 | } |
4222 | |
4223 | return d; |
4224 | } |
4225 | |
4226 | /* Add NODE to the end of QUEUE, unless IN_QUEUE_P indicates that |
4227 | it is already present. */ |
4228 | |
4229 | static void |
4230 | maybe_push_queue (struct cgraph_node *node, |
4231 | cgraph_node_queue *queue_p, bool *in_queue_p) |
4232 | { |
4233 | if (!*in_queue_p) |
4234 | { |
4235 | *in_queue_p = true; |
4236 | queue_p->safe_push (obj: node); |
4237 | } |
4238 | } |
4239 | |
4240 | /* A subroutine of ipa_tm_scan_calls_transaction and ipa_tm_scan_calls_clone. |
4241 | Queue all callees within block BB. */ |
4242 | |
4243 | static void |
4244 | ipa_tm_scan_calls_block (cgraph_node_queue *callees_p, |
4245 | basic_block bb, bool for_clone) |
4246 | { |
4247 | gimple_stmt_iterator gsi; |
4248 | |
4249 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
4250 | { |
4251 | gimple *stmt = gsi_stmt (i: gsi); |
4252 | if (is_gimple_call (gs: stmt) && !is_tm_pure_call (call: stmt)) |
4253 | { |
4254 | tree fndecl = gimple_call_fndecl (gs: stmt); |
4255 | if (fndecl) |
4256 | { |
4257 | struct tm_ipa_cg_data *d; |
4258 | unsigned *pcallers; |
4259 | struct cgraph_node *node; |
4260 | |
4261 | if (is_tm_ending_fndecl (fndecl)) |
4262 | continue; |
4263 | if (find_tm_replacement_function (fndecl)) |
4264 | continue; |
4265 | |
4266 | node = cgraph_node::get (decl: fndecl); |
4267 | gcc_assert (node != NULL); |
4268 | d = get_cg_data (node: &node, traverse_aliases: true); |
4269 | |
4270 | pcallers = (for_clone ? &d->tm_callers_clone |
4271 | : &d->tm_callers_normal); |
4272 | *pcallers += 1; |
4273 | |
4274 | maybe_push_queue (node, queue_p: callees_p, in_queue_p: &d->in_callee_queue); |
4275 | } |
4276 | } |
4277 | } |
4278 | } |
4279 | |
4280 | /* Scan all calls in NODE that are within a transaction region, |
4281 | and push the resulting nodes into the callee queue. */ |
4282 | |
4283 | static void |
4284 | ipa_tm_scan_calls_transaction (struct tm_ipa_cg_data *d, |
4285 | cgraph_node_queue *callees_p) |
4286 | { |
4287 | d->transaction_blocks_normal = BITMAP_ALLOC (obstack: &tm_obstack); |
4288 | d->all_tm_regions = all_tm_regions; |
4289 | |
4290 | for (tm_region *r = all_tm_regions; r; r = r->next) |
4291 | { |
4292 | vec<basic_block> bbs; |
4293 | basic_block bb; |
4294 | unsigned i; |
4295 | |
4296 | bbs = get_tm_region_blocks (entry_block: r->entry_block, exit_blocks: r->exit_blocks, NULL, |
4297 | all_region_blocks: d->transaction_blocks_normal, stop_at_irrevocable_p: false, include_uninstrumented_p: false); |
4298 | |
4299 | FOR_EACH_VEC_ELT (bbs, i, bb) |
4300 | ipa_tm_scan_calls_block (callees_p, bb, for_clone: false); |
4301 | |
4302 | bbs.release (); |
4303 | } |
4304 | } |
4305 | |
4306 | /* Scan all calls in NODE as if this is the transactional clone, |
4307 | and push the destinations into the callee queue. */ |
4308 | |
4309 | static void |
4310 | ipa_tm_scan_calls_clone (struct cgraph_node *node, |
4311 | cgraph_node_queue *callees_p) |
4312 | { |
4313 | struct function *fn = DECL_STRUCT_FUNCTION (node->decl); |
4314 | basic_block bb; |
4315 | |
4316 | FOR_EACH_BB_FN (bb, fn) |
4317 | ipa_tm_scan_calls_block (callees_p, bb, for_clone: true); |
4318 | } |
4319 | |
4320 | /* The function NODE has been detected to be irrevocable. Push all |
4321 | of its callers onto WORKLIST for the purpose of re-scanning them. */ |
4322 | |
4323 | static void |
4324 | ipa_tm_note_irrevocable (struct cgraph_node *node, |
4325 | cgraph_node_queue *worklist_p) |
4326 | { |
4327 | struct tm_ipa_cg_data *d = get_cg_data (node: &node, traverse_aliases: true); |
4328 | struct cgraph_edge *e; |
4329 | |
4330 | d->is_irrevocable = true; |
4331 | |
4332 | for (e = node->callers; e ; e = e->next_caller) |
4333 | { |
4334 | basic_block bb; |
4335 | struct cgraph_node *caller; |
4336 | |
4337 | /* Don't examine recursive calls. */ |
4338 | if (e->caller == node) |
4339 | continue; |
4340 | /* Even if we think we can go irrevocable, believe the user |
4341 | above all. */ |
4342 | if (is_tm_safe_or_pure (x: e->caller->decl)) |
4343 | continue; |
4344 | |
4345 | caller = e->caller; |
4346 | d = get_cg_data (node: &caller, traverse_aliases: true); |
4347 | |
4348 | /* Check if the callee is in a transactional region. If so, |
4349 | schedule the function for normal re-scan as well. */ |
4350 | bb = gimple_bb (g: e->call_stmt); |
4351 | gcc_assert (bb != NULL); |
4352 | if (d->transaction_blocks_normal |
4353 | && bitmap_bit_p (d->transaction_blocks_normal, bb->index)) |
4354 | d->want_irr_scan_normal = true; |
4355 | |
4356 | maybe_push_queue (node: caller, queue_p: worklist_p, in_queue_p: &d->in_worklist); |
4357 | } |
4358 | } |
4359 | |
4360 | /* A subroutine of ipa_tm_scan_irr_blocks; return true iff any statement |
4361 | within the block is irrevocable. */ |
4362 | |
4363 | static bool |
4364 | ipa_tm_scan_irr_block (basic_block bb) |
4365 | { |
4366 | gimple_stmt_iterator gsi; |
4367 | tree fn; |
4368 | |
4369 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
4370 | { |
4371 | gimple *stmt = gsi_stmt (i: gsi); |
4372 | switch (gimple_code (g: stmt)) |
4373 | { |
4374 | case GIMPLE_ASSIGN: |
4375 | if (gimple_assign_single_p (gs: stmt)) |
4376 | { |
4377 | tree lhs = gimple_assign_lhs (gs: stmt); |
4378 | tree rhs = gimple_assign_rhs1 (gs: stmt); |
4379 | if (volatile_lvalue_p (t: lhs) || volatile_lvalue_p (t: rhs)) |
4380 | return true; |
4381 | } |
4382 | break; |
4383 | |
4384 | case GIMPLE_CALL: |
4385 | { |
4386 | tree lhs = gimple_call_lhs (gs: stmt); |
4387 | if (lhs && volatile_lvalue_p (t: lhs)) |
4388 | return true; |
4389 | |
4390 | if (is_tm_pure_call (call: stmt)) |
4391 | break; |
4392 | |
4393 | fn = gimple_call_fn (gs: stmt); |
4394 | |
4395 | /* Functions with the attribute are by definition irrevocable. */ |
4396 | if (is_tm_irrevocable (x: fn)) |
4397 | return true; |
4398 | |
4399 | /* For direct function calls, go ahead and check for replacement |
4400 | functions, or transitive irrevocable functions. For indirect |
4401 | functions, we'll ask the runtime. */ |
4402 | if (TREE_CODE (fn) == ADDR_EXPR) |
4403 | { |
4404 | struct tm_ipa_cg_data *d; |
4405 | struct cgraph_node *node; |
4406 | |
4407 | fn = TREE_OPERAND (fn, 0); |
4408 | if (is_tm_ending_fndecl (fndecl: fn)) |
4409 | break; |
4410 | if (find_tm_replacement_function (fndecl: fn)) |
4411 | break; |
4412 | |
4413 | node = cgraph_node::get (decl: fn); |
4414 | d = get_cg_data (node: &node, traverse_aliases: true); |
4415 | |
4416 | /* Return true if irrevocable, but above all, believe |
4417 | the user. */ |
4418 | if (d->is_irrevocable |
4419 | && !is_tm_safe_or_pure (x: fn)) |
4420 | return true; |
4421 | } |
4422 | break; |
4423 | } |
4424 | |
4425 | case GIMPLE_ASM: |
4426 | /* ??? The Approved Method of indicating that an inline |
4427 | assembly statement is not relevant to the transaction |
4428 | is to wrap it in a __tm_waiver block. This is not |
4429 | yet implemented, so we can't check for it. */ |
4430 | if (is_tm_safe (x: current_function_decl)) |
4431 | error_at (gimple_location (g: stmt), |
4432 | "%<asm%> not allowed in %<transaction_safe%> function" ); |
4433 | return true; |
4434 | |
4435 | default: |
4436 | break; |
4437 | } |
4438 | } |
4439 | |
4440 | return false; |
4441 | } |
4442 | |
4443 | /* For each of the blocks seeded witin PQUEUE, walk the CFG looking |
4444 | for new irrevocable blocks, marking them in NEW_IRR. Don't bother |
4445 | scanning past OLD_IRR or EXIT_BLOCKS. */ |
4446 | |
4447 | static bool |
4448 | ipa_tm_scan_irr_blocks (vec<basic_block> *pqueue, bitmap new_irr, |
4449 | bitmap old_irr, bitmap exit_blocks) |
4450 | { |
4451 | bool any_new_irr = false; |
4452 | edge e; |
4453 | edge_iterator ei; |
4454 | bitmap visited_blocks = BITMAP_ALLOC (NULL); |
4455 | |
4456 | do |
4457 | { |
4458 | basic_block bb = pqueue->pop (); |
4459 | |
4460 | /* Don't re-scan blocks we know already are irrevocable. */ |
4461 | if (old_irr && bitmap_bit_p (old_irr, bb->index)) |
4462 | continue; |
4463 | |
4464 | if (ipa_tm_scan_irr_block (bb)) |
4465 | { |
4466 | bitmap_set_bit (new_irr, bb->index); |
4467 | any_new_irr = true; |
4468 | } |
4469 | else if (exit_blocks == NULL || !bitmap_bit_p (exit_blocks, bb->index)) |
4470 | { |
4471 | FOR_EACH_EDGE (e, ei, bb->succs) |
4472 | if (!bitmap_bit_p (visited_blocks, e->dest->index)) |
4473 | { |
4474 | bitmap_set_bit (visited_blocks, e->dest->index); |
4475 | pqueue->safe_push (obj: e->dest); |
4476 | } |
4477 | } |
4478 | } |
4479 | while (!pqueue->is_empty ()); |
4480 | |
4481 | BITMAP_FREE (visited_blocks); |
4482 | |
4483 | return any_new_irr; |
4484 | } |
4485 | |
4486 | /* Propagate the irrevocable property both up and down the dominator tree. |
4487 | BB is the current block being scanned; EXIT_BLOCKS are the edges of the |
4488 | TM regions; OLD_IRR are the results of a previous scan of the dominator |
4489 | tree which has been fully propagated; NEW_IRR is the set of new blocks |
4490 | which are gaining the irrevocable property during the current scan. */ |
4491 | |
4492 | static void |
4493 | ipa_tm_propagate_irr (basic_block entry_block, bitmap new_irr, |
4494 | bitmap old_irr, bitmap exit_blocks) |
4495 | { |
4496 | vec<basic_block> bbs; |
4497 | bitmap all_region_blocks; |
4498 | |
4499 | /* If this block is in the old set, no need to rescan. */ |
4500 | if (old_irr && bitmap_bit_p (old_irr, entry_block->index)) |
4501 | return; |
4502 | |
4503 | all_region_blocks = BITMAP_ALLOC (obstack: &tm_obstack); |
4504 | bbs = get_tm_region_blocks (entry_block, exit_blocks, NULL, |
4505 | all_region_blocks, stop_at_irrevocable_p: false); |
4506 | do |
4507 | { |
4508 | basic_block bb = bbs.pop (); |
4509 | bool this_irr = bitmap_bit_p (new_irr, bb->index); |
4510 | bool all_son_irr = false; |
4511 | edge_iterator ei; |
4512 | edge e; |
4513 | |
4514 | /* Propagate up. If my children are, I am too, but we must have |
4515 | at least one child that is. */ |
4516 | if (!this_irr) |
4517 | { |
4518 | FOR_EACH_EDGE (e, ei, bb->succs) |
4519 | { |
4520 | if (!bitmap_bit_p (new_irr, e->dest->index)) |
4521 | { |
4522 | all_son_irr = false; |
4523 | break; |
4524 | } |
4525 | else |
4526 | all_son_irr = true; |
4527 | } |
4528 | if (all_son_irr) |
4529 | { |
4530 | /* Add block to new_irr if it hasn't already been processed. */ |
4531 | if (!old_irr || !bitmap_bit_p (old_irr, bb->index)) |
4532 | { |
4533 | bitmap_set_bit (new_irr, bb->index); |
4534 | this_irr = true; |
4535 | } |
4536 | } |
4537 | } |
4538 | |
4539 | /* Propagate down to everyone we immediately dominate. */ |
4540 | if (this_irr) |
4541 | { |
4542 | basic_block son; |
4543 | for (son = first_dom_son (CDI_DOMINATORS, bb); |
4544 | son; |
4545 | son = next_dom_son (CDI_DOMINATORS, son)) |
4546 | { |
4547 | /* Make sure block is actually in a TM region, and it |
4548 | isn't already in old_irr. */ |
4549 | if ((!old_irr || !bitmap_bit_p (old_irr, son->index)) |
4550 | && bitmap_bit_p (all_region_blocks, son->index)) |
4551 | bitmap_set_bit (new_irr, son->index); |
4552 | } |
4553 | } |
4554 | } |
4555 | while (!bbs.is_empty ()); |
4556 | |
4557 | BITMAP_FREE (all_region_blocks); |
4558 | bbs.release (); |
4559 | } |
4560 | |
4561 | static void |
4562 | ipa_tm_decrement_clone_counts (basic_block bb, bool for_clone) |
4563 | { |
4564 | gimple_stmt_iterator gsi; |
4565 | |
4566 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
4567 | { |
4568 | gimple *stmt = gsi_stmt (i: gsi); |
4569 | if (is_gimple_call (gs: stmt) && !is_tm_pure_call (call: stmt)) |
4570 | { |
4571 | tree fndecl = gimple_call_fndecl (gs: stmt); |
4572 | if (fndecl) |
4573 | { |
4574 | struct tm_ipa_cg_data *d; |
4575 | unsigned *pcallers; |
4576 | struct cgraph_node *tnode; |
4577 | |
4578 | if (is_tm_ending_fndecl (fndecl)) |
4579 | continue; |
4580 | if (find_tm_replacement_function (fndecl)) |
4581 | continue; |
4582 | |
4583 | tnode = cgraph_node::get (decl: fndecl); |
4584 | d = get_cg_data (node: &tnode, traverse_aliases: true); |
4585 | |
4586 | pcallers = (for_clone ? &d->tm_callers_clone |
4587 | : &d->tm_callers_normal); |
4588 | |
4589 | gcc_assert (*pcallers > 0); |
4590 | *pcallers -= 1; |
4591 | } |
4592 | } |
4593 | } |
4594 | } |
4595 | |
4596 | /* (Re-)Scan the transaction blocks in NODE for calls to irrevocable functions, |
4597 | as well as other irrevocable actions such as inline assembly. Mark all |
4598 | such blocks as irrevocable and decrement the number of calls to |
4599 | transactional clones. Return true if, for the transactional clone, the |
4600 | entire function is irrevocable. */ |
4601 | |
4602 | static bool |
4603 | ipa_tm_scan_irr_function (struct cgraph_node *node, bool for_clone) |
4604 | { |
4605 | struct tm_ipa_cg_data *d; |
4606 | bitmap new_irr, old_irr; |
4607 | bool ret = false; |
4608 | |
4609 | /* Builtin operators (operator new, and such). */ |
4610 | if (DECL_STRUCT_FUNCTION (node->decl) == NULL |
4611 | || DECL_STRUCT_FUNCTION (node->decl)->cfg == NULL) |
4612 | return false; |
4613 | |
4614 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); |
4615 | calculate_dominance_info (CDI_DOMINATORS); |
4616 | |
4617 | d = get_cg_data (node: &node, traverse_aliases: true); |
4618 | auto_vec<basic_block, 10> queue; |
4619 | new_irr = BITMAP_ALLOC (obstack: &tm_obstack); |
4620 | |
4621 | /* Scan each tm region, propagating irrevocable status through the tree. */ |
4622 | if (for_clone) |
4623 | { |
4624 | old_irr = d->irrevocable_blocks_clone; |
4625 | queue.quick_push (obj: single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun))); |
4626 | if (ipa_tm_scan_irr_blocks (pqueue: &queue, new_irr, old_irr, NULL)) |
4627 | { |
4628 | ipa_tm_propagate_irr (entry_block: single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)), |
4629 | new_irr, |
4630 | old_irr, NULL); |
4631 | ret = bitmap_bit_p (new_irr, |
4632 | single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun))->index); |
4633 | } |
4634 | } |
4635 | else |
4636 | { |
4637 | struct tm_region *region; |
4638 | |
4639 | old_irr = d->irrevocable_blocks_normal; |
4640 | for (region = d->all_tm_regions; region; region = region->next) |
4641 | { |
4642 | queue.quick_push (obj: region->entry_block); |
4643 | if (ipa_tm_scan_irr_blocks (pqueue: &queue, new_irr, old_irr, |
4644 | exit_blocks: region->exit_blocks)) |
4645 | ipa_tm_propagate_irr (entry_block: region->entry_block, new_irr, old_irr, |
4646 | exit_blocks: region->exit_blocks); |
4647 | } |
4648 | } |
4649 | |
4650 | /* If we found any new irrevocable blocks, reduce the call count for |
4651 | transactional clones within the irrevocable blocks. Save the new |
4652 | set of irrevocable blocks for next time. */ |
4653 | if (!bitmap_empty_p (map: new_irr)) |
4654 | { |
4655 | bitmap_iterator bmi; |
4656 | unsigned i; |
4657 | |
4658 | EXECUTE_IF_SET_IN_BITMAP (new_irr, 0, i, bmi) |
4659 | ipa_tm_decrement_clone_counts (BASIC_BLOCK_FOR_FN (cfun, i), |
4660 | for_clone); |
4661 | |
4662 | if (old_irr) |
4663 | { |
4664 | bitmap_ior_into (old_irr, new_irr); |
4665 | BITMAP_FREE (new_irr); |
4666 | } |
4667 | else if (for_clone) |
4668 | d->irrevocable_blocks_clone = new_irr; |
4669 | else |
4670 | d->irrevocable_blocks_normal = new_irr; |
4671 | |
4672 | if (dump_file && new_irr) |
4673 | { |
4674 | const char *dname; |
4675 | bitmap_iterator bmi; |
4676 | unsigned i; |
4677 | |
4678 | dname = lang_hooks.decl_printable_name (current_function_decl, 2); |
4679 | EXECUTE_IF_SET_IN_BITMAP (new_irr, 0, i, bmi) |
4680 | fprintf (stream: dump_file, format: "%s: bb %d goes irrevocable\n" , dname, i); |
4681 | } |
4682 | } |
4683 | else |
4684 | BITMAP_FREE (new_irr); |
4685 | |
4686 | pop_cfun (); |
4687 | |
4688 | return ret; |
4689 | } |
4690 | |
4691 | /* Return true if, for the transactional clone of NODE, any call |
4692 | may enter irrevocable mode. */ |
4693 | |
4694 | static bool |
4695 | ipa_tm_mayenterirr_function (struct cgraph_node *node) |
4696 | { |
4697 | struct tm_ipa_cg_data *d; |
4698 | tree decl; |
4699 | unsigned flags; |
4700 | |
4701 | d = get_cg_data (node: &node, traverse_aliases: true); |
4702 | decl = node->decl; |
4703 | flags = flags_from_decl_or_type (decl); |
4704 | |
4705 | /* Handle some TM builtins. Ordinarily these aren't actually generated |
4706 | at this point, but handling these functions when written in by the |
4707 | user makes it easier to build unit tests. */ |
4708 | if (flags & ECF_TM_BUILTIN) |
4709 | return false; |
4710 | |
4711 | /* Filter out all functions that are marked. */ |
4712 | if (flags & ECF_TM_PURE) |
4713 | return false; |
4714 | if (is_tm_safe (x: decl)) |
4715 | return false; |
4716 | if (is_tm_irrevocable (x: decl)) |
4717 | return true; |
4718 | if (is_tm_callable (x: decl)) |
4719 | return true; |
4720 | if (find_tm_replacement_function (fndecl: decl)) |
4721 | return true; |
4722 | |
4723 | /* If we aren't seeing the final version of the function we don't |
4724 | know what it will contain at runtime. */ |
4725 | if (node->get_availability () < AVAIL_AVAILABLE) |
4726 | return true; |
4727 | |
4728 | /* If the function must go irrevocable, then of course true. */ |
4729 | if (d->is_irrevocable) |
4730 | return true; |
4731 | |
4732 | /* If there are any blocks marked irrevocable, then the function |
4733 | as a whole may enter irrevocable. */ |
4734 | if (d->irrevocable_blocks_clone) |
4735 | return true; |
4736 | |
4737 | /* We may have previously marked this function as tm_may_enter_irr; |
4738 | see pass_diagnose_tm_blocks. */ |
4739 | if (node->tm_may_enter_irr) |
4740 | return true; |
4741 | |
4742 | /* Recurse on the main body for aliases. In general, this will |
4743 | result in one of the bits above being set so that we will not |
4744 | have to recurse next time. */ |
4745 | if (node->alias) |
4746 | return ipa_tm_mayenterirr_function |
4747 | (node: cgraph_node::get (decl: thunk_info::get (node)->alias)); |
4748 | |
4749 | /* What remains is unmarked local functions without items that force |
4750 | the function to go irrevocable. */ |
4751 | return false; |
4752 | } |
4753 | |
4754 | /* Diagnose calls from transaction_safe functions to unmarked |
4755 | functions that are determined to not be safe. */ |
4756 | |
4757 | static void |
4758 | ipa_tm_diagnose_tm_safe (struct cgraph_node *node) |
4759 | { |
4760 | struct cgraph_edge *e; |
4761 | |
4762 | for (e = node->callees; e ; e = e->next_callee) |
4763 | if (!is_tm_callable (x: e->callee->decl) |
4764 | && e->callee->tm_may_enter_irr) |
4765 | error_at (gimple_location (g: e->call_stmt), |
4766 | "unsafe function call %qD within " |
4767 | "%<transaction_safe%> function" , e->callee->decl); |
4768 | } |
4769 | |
4770 | /* Diagnose call from atomic transactions to unmarked functions |
4771 | that are determined to not be safe. */ |
4772 | |
4773 | static void |
4774 | ipa_tm_diagnose_transaction (struct cgraph_node *node, |
4775 | struct tm_region *all_tm_regions) |
4776 | { |
4777 | struct tm_region *r; |
4778 | |
4779 | for (r = all_tm_regions; r ; r = r->next) |
4780 | if (gimple_transaction_subcode (transaction_stmt: r->get_transaction_stmt ()) |
4781 | & GTMA_IS_RELAXED) |
4782 | { |
4783 | /* Atomic transactions can be nested inside relaxed. */ |
4784 | if (r->inner) |
4785 | ipa_tm_diagnose_transaction (node, all_tm_regions: r->inner); |
4786 | } |
4787 | else |
4788 | { |
4789 | vec<basic_block> bbs; |
4790 | gimple_stmt_iterator gsi; |
4791 | basic_block bb; |
4792 | size_t i; |
4793 | |
4794 | bbs = get_tm_region_blocks (entry_block: r->entry_block, exit_blocks: r->exit_blocks, |
4795 | irr_blocks: r->irr_blocks, NULL, stop_at_irrevocable_p: false); |
4796 | |
4797 | for (i = 0; bbs.iterate (ix: i, ptr: &bb); ++i) |
4798 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
4799 | { |
4800 | gimple *stmt = gsi_stmt (i: gsi); |
4801 | tree fndecl; |
4802 | |
4803 | if (gimple_code (g: stmt) == GIMPLE_ASM) |
4804 | { |
4805 | error_at (gimple_location (g: stmt), |
4806 | "%<asm%> not allowed in atomic transaction" ); |
4807 | continue; |
4808 | } |
4809 | |
4810 | if (!is_gimple_call (gs: stmt)) |
4811 | continue; |
4812 | fndecl = gimple_call_fndecl (gs: stmt); |
4813 | |
4814 | /* Indirect function calls have been diagnosed already. */ |
4815 | if (!fndecl) |
4816 | continue; |
4817 | |
4818 | /* Stop at the end of the transaction. */ |
4819 | if (is_tm_ending_fndecl (fndecl)) |
4820 | { |
4821 | if (bitmap_bit_p (r->exit_blocks, bb->index)) |
4822 | break; |
4823 | continue; |
4824 | } |
4825 | |
4826 | /* Marked functions have been diagnosed already. */ |
4827 | if (is_tm_pure_call (call: stmt)) |
4828 | continue; |
4829 | if (is_tm_callable (x: fndecl)) |
4830 | continue; |
4831 | |
4832 | if (cgraph_node::local_info_node (decl: fndecl)->tm_may_enter_irr) |
4833 | error_at (gimple_location (g: stmt), |
4834 | "unsafe function call %qD within " |
4835 | "atomic transaction" , fndecl); |
4836 | } |
4837 | |
4838 | bbs.release (); |
4839 | } |
4840 | } |
4841 | |
4842 | /* Return a transactional mangled name for the DECL_ASSEMBLER_NAME in |
4843 | OLD_DECL. The returned value is a freshly malloced pointer that |
4844 | should be freed by the caller. */ |
4845 | |
4846 | static tree |
4847 | tm_mangle (tree old_asm_id) |
4848 | { |
4849 | const char *old_asm_name; |
4850 | char *tm_name; |
4851 | void *alloc = NULL; |
4852 | struct demangle_component *dc; |
4853 | tree new_asm_id; |
4854 | |
4855 | /* Determine if the symbol is already a valid C++ mangled name. Do this |
4856 | even for C, which might be interfacing with C++ code via appropriately |
4857 | ugly identifiers. */ |
4858 | /* ??? We could probably do just as well checking for "_Z" and be done. */ |
4859 | old_asm_name = IDENTIFIER_POINTER (old_asm_id); |
4860 | dc = cplus_demangle_v3_components (mangled: old_asm_name, DMGL_NO_OPTS, mem: &alloc); |
4861 | |
4862 | if (dc == NULL) |
4863 | { |
4864 | char length[12]; |
4865 | |
4866 | do_unencoded: |
4867 | sprintf (s: length, format: "%u" , IDENTIFIER_LENGTH (old_asm_id)); |
4868 | tm_name = concat ("_ZGTt" , length, old_asm_name, NULL); |
4869 | } |
4870 | else |
4871 | { |
4872 | old_asm_name += 2; /* Skip _Z */ |
4873 | |
4874 | switch (dc->type) |
4875 | { |
4876 | case DEMANGLE_COMPONENT_TRANSACTION_CLONE: |
4877 | case DEMANGLE_COMPONENT_NONTRANSACTION_CLONE: |
4878 | /* Don't play silly games, you! */ |
4879 | goto do_unencoded; |
4880 | |
4881 | case DEMANGLE_COMPONENT_HIDDEN_ALIAS: |
4882 | /* I'd really like to know if we can ever be passed one of |
4883 | these from the C++ front end. The Logical Thing would |
4884 | seem that hidden-alias should be outer-most, so that we |
4885 | get hidden-alias of a transaction-clone and not vice-versa. */ |
4886 | old_asm_name += 2; |
4887 | break; |
4888 | |
4889 | default: |
4890 | break; |
4891 | } |
4892 | |
4893 | tm_name = concat ("_ZGTt" , old_asm_name, NULL); |
4894 | } |
4895 | free (ptr: alloc); |
4896 | |
4897 | new_asm_id = get_identifier (tm_name); |
4898 | free (ptr: tm_name); |
4899 | |
4900 | return new_asm_id; |
4901 | } |
4902 | |
4903 | static inline void |
4904 | ipa_tm_mark_force_output_node (struct cgraph_node *node) |
4905 | { |
4906 | node->mark_force_output (); |
4907 | node->analyzed = true; |
4908 | } |
4909 | |
4910 | static inline void |
4911 | ipa_tm_mark_forced_by_abi_node (struct cgraph_node *node) |
4912 | { |
4913 | node->forced_by_abi = true; |
4914 | node->analyzed = true; |
4915 | } |
4916 | |
4917 | /* Callback data for ipa_tm_create_version_alias. */ |
4918 | struct create_version_alias_info |
4919 | { |
4920 | struct cgraph_node *old_node; |
4921 | tree new_decl; |
4922 | }; |
4923 | |
4924 | /* A subroutine of ipa_tm_create_version, called via |
4925 | cgraph_for_node_and_aliases. Create new tm clones for each of |
4926 | the existing aliases. */ |
4927 | static bool |
4928 | ipa_tm_create_version_alias (struct cgraph_node *node, void *data) |
4929 | { |
4930 | struct create_version_alias_info *info |
4931 | = (struct create_version_alias_info *)data; |
4932 | tree old_decl, new_decl, tm_name; |
4933 | struct cgraph_node *new_node; |
4934 | |
4935 | if (!node->cpp_implicit_alias) |
4936 | return false; |
4937 | |
4938 | old_decl = node->decl; |
4939 | tm_name = tm_mangle (DECL_ASSEMBLER_NAME (old_decl)); |
4940 | new_decl = build_decl (DECL_SOURCE_LOCATION (old_decl), |
4941 | TREE_CODE (old_decl), tm_name, |
4942 | TREE_TYPE (old_decl)); |
4943 | |
4944 | SET_DECL_ASSEMBLER_NAME (new_decl, tm_name); |
4945 | SET_DECL_RTL (new_decl, NULL); |
4946 | |
4947 | /* Based loosely on C++'s make_alias_for(). */ |
4948 | TREE_PUBLIC (new_decl) = TREE_PUBLIC (old_decl); |
4949 | DECL_CONTEXT (new_decl) = DECL_CONTEXT (old_decl); |
4950 | DECL_LANG_SPECIFIC (new_decl) = DECL_LANG_SPECIFIC (old_decl); |
4951 | TREE_READONLY (new_decl) = TREE_READONLY (old_decl); |
4952 | DECL_EXTERNAL (new_decl) = 0; |
4953 | DECL_ARTIFICIAL (new_decl) = 1; |
4954 | TREE_ADDRESSABLE (new_decl) = 1; |
4955 | TREE_USED (new_decl) = 1; |
4956 | TREE_SYMBOL_REFERENCED (tm_name) = 1; |
4957 | |
4958 | /* Perform the same remapping to the comdat group. */ |
4959 | if (DECL_ONE_ONLY (new_decl)) |
4960 | varpool_node::get (decl: new_decl)->set_comdat_group |
4961 | (tm_mangle (old_asm_id: decl_comdat_group_id (old_decl))); |
4962 | |
4963 | new_node = cgraph_node::create_same_body_alias (alias: new_decl, decl: info->new_decl); |
4964 | new_node->tm_clone = true; |
4965 | new_node->externally_visible = info->old_node->externally_visible; |
4966 | new_node->no_reorder = info->old_node->no_reorder; |
4967 | /* ?? Do not traverse aliases here. */ |
4968 | get_cg_data (node: &node, traverse_aliases: false)->clone = new_node; |
4969 | |
4970 | record_tm_clone_pair (old_decl, new_decl); |
4971 | |
4972 | if (info->old_node->force_output |
4973 | || info->old_node->ref_list.first_referring ()) |
4974 | ipa_tm_mark_force_output_node (node: new_node); |
4975 | if (info->old_node->forced_by_abi) |
4976 | ipa_tm_mark_forced_by_abi_node (node: new_node); |
4977 | return false; |
4978 | } |
4979 | |
4980 | /* Create a copy of the function (possibly declaration only) of OLD_NODE, |
4981 | appropriate for the transactional clone. */ |
4982 | |
4983 | static void |
4984 | ipa_tm_create_version (struct cgraph_node *old_node) |
4985 | { |
4986 | tree new_decl, old_decl, tm_name; |
4987 | struct cgraph_node *new_node; |
4988 | |
4989 | old_decl = old_node->decl; |
4990 | new_decl = copy_node (old_decl); |
4991 | |
4992 | /* DECL_ASSEMBLER_NAME needs to be set before we call |
4993 | cgraph_copy_node_for_versioning below, because cgraph_node will |
4994 | fill the assembler_name_hash. */ |
4995 | tm_name = tm_mangle (DECL_ASSEMBLER_NAME (old_decl)); |
4996 | SET_DECL_ASSEMBLER_NAME (new_decl, tm_name); |
4997 | SET_DECL_RTL (new_decl, NULL); |
4998 | TREE_SYMBOL_REFERENCED (tm_name) = 1; |
4999 | |
5000 | /* Perform the same remapping to the comdat group. */ |
5001 | if (DECL_ONE_ONLY (new_decl)) |
5002 | varpool_node::get (decl: new_decl)->set_comdat_group |
5003 | (tm_mangle (DECL_COMDAT_GROUP (old_decl))); |
5004 | |
5005 | gcc_assert (!old_node->ipa_transforms_to_apply.exists ()); |
5006 | new_node = old_node->create_version_clone (new_decl, redirect_callers: vNULL, NULL); |
5007 | new_node->local = false; |
5008 | new_node->externally_visible = old_node->externally_visible; |
5009 | new_node->lowered = true; |
5010 | new_node->tm_clone = 1; |
5011 | if (!old_node->implicit_section) |
5012 | new_node->set_section (*old_node); |
5013 | get_cg_data (node: &old_node, traverse_aliases: true)->clone = new_node; |
5014 | |
5015 | if (old_node->get_availability () >= AVAIL_INTERPOSABLE) |
5016 | { |
5017 | /* Remap extern inline to static inline. */ |
5018 | /* ??? Is it worth trying to use make_decl_one_only? */ |
5019 | if (DECL_DECLARED_INLINE_P (new_decl) && DECL_EXTERNAL (new_decl)) |
5020 | { |
5021 | DECL_EXTERNAL (new_decl) = 0; |
5022 | TREE_PUBLIC (new_decl) = 0; |
5023 | DECL_WEAK (new_decl) = 0; |
5024 | } |
5025 | |
5026 | tree_function_versioning (old_decl, new_decl, |
5027 | NULL, NULL, false, NULL, NULL); |
5028 | } |
5029 | |
5030 | record_tm_clone_pair (old_decl, new_decl); |
5031 | |
5032 | symtab->call_cgraph_insertion_hooks (node: new_node); |
5033 | if (old_node->force_output |
5034 | || old_node->ref_list.first_referring ()) |
5035 | ipa_tm_mark_force_output_node (node: new_node); |
5036 | if (old_node->forced_by_abi) |
5037 | ipa_tm_mark_forced_by_abi_node (node: new_node); |
5038 | |
5039 | /* Do the same thing, but for any aliases of the original node. */ |
5040 | { |
5041 | struct create_version_alias_info data; |
5042 | data.old_node = old_node; |
5043 | data.new_decl = new_decl; |
5044 | old_node->call_for_symbol_thunks_and_aliases (callback: ipa_tm_create_version_alias, |
5045 | data: &data, include_overwritable: true); |
5046 | } |
5047 | } |
5048 | |
5049 | /* Construct a call to TM_IRREVOCABLE and insert it at the beginning of BB. */ |
5050 | |
5051 | static void |
5052 | ipa_tm_insert_irr_call (struct cgraph_node *node, struct tm_region *region, |
5053 | basic_block bb) |
5054 | { |
5055 | gimple_stmt_iterator gsi; |
5056 | gcall *g; |
5057 | |
5058 | transaction_subcode_ior (region, GTMA_MAY_ENTER_IRREVOCABLE); |
5059 | |
5060 | g = gimple_build_call (builtin_decl_explicit (fncode: BUILT_IN_TM_IRREVOCABLE), |
5061 | 1, build_int_cst (NULL_TREE, MODE_SERIALIRREVOCABLE)); |
5062 | |
5063 | split_block_after_labels (bb); |
5064 | gsi = gsi_after_labels (bb); |
5065 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); |
5066 | |
5067 | node->create_edge (callee: cgraph_node::get_create |
5068 | (builtin_decl_explicit (fncode: BUILT_IN_TM_IRREVOCABLE)), |
5069 | call_stmt: g, count: gimple_bb (g)->count); |
5070 | } |
5071 | |
5072 | /* Construct a call to TM_GETTMCLONE and insert it before GSI. */ |
5073 | |
5074 | static bool |
5075 | ipa_tm_insert_gettmclone_call (struct cgraph_node *node, |
5076 | struct tm_region *region, |
5077 | gimple_stmt_iterator *gsi, gcall *stmt) |
5078 | { |
5079 | tree gettm_fn, ret, old_fn, callfn; |
5080 | gcall *g; |
5081 | gassign *g2; |
5082 | bool safe; |
5083 | |
5084 | old_fn = gimple_call_fn (gs: stmt); |
5085 | |
5086 | if (TREE_CODE (old_fn) == ADDR_EXPR) |
5087 | { |
5088 | tree fndecl = TREE_OPERAND (old_fn, 0); |
5089 | tree clone = get_tm_clone_pair (fndecl); |
5090 | |
5091 | /* By transforming the call into a TM_GETTMCLONE, we are |
5092 | technically taking the address of the original function and |
5093 | its clone. Explain this so inlining will know this function |
5094 | is needed. */ |
5095 | cgraph_node::get (decl: fndecl)->mark_address_taken () ; |
5096 | if (clone) |
5097 | cgraph_node::get (decl: clone)->mark_address_taken (); |
5098 | } |
5099 | |
5100 | safe = is_tm_safe (TREE_TYPE (old_fn)); |
5101 | gettm_fn = builtin_decl_explicit (fncode: safe ? BUILT_IN_TM_GETTMCLONE_SAFE |
5102 | : BUILT_IN_TM_GETTMCLONE_IRR); |
5103 | ret = create_tmp_var (ptr_type_node); |
5104 | |
5105 | if (!safe) |
5106 | transaction_subcode_ior (region, GTMA_MAY_ENTER_IRREVOCABLE); |
5107 | |
5108 | /* Discard OBJ_TYPE_REF, since we weren't able to fold it. */ |
5109 | if (TREE_CODE (old_fn) == OBJ_TYPE_REF) |
5110 | old_fn = OBJ_TYPE_REF_EXPR (old_fn); |
5111 | |
5112 | g = gimple_build_call (gettm_fn, 1, old_fn); |
5113 | ret = make_ssa_name (var: ret, stmt: g); |
5114 | gimple_call_set_lhs (gs: g, lhs: ret); |
5115 | |
5116 | gsi_insert_before (gsi, g, GSI_SAME_STMT); |
5117 | |
5118 | node->create_edge (callee: cgraph_node::get_create (gettm_fn), call_stmt: g, count: gimple_bb (g)->count); |
5119 | |
5120 | /* Cast return value from tm_gettmclone* into appropriate function |
5121 | pointer. */ |
5122 | callfn = create_tmp_var (TREE_TYPE (old_fn)); |
5123 | g2 = gimple_build_assign (callfn, |
5124 | fold_build1 (NOP_EXPR, TREE_TYPE (callfn), ret)); |
5125 | callfn = make_ssa_name (var: callfn, stmt: g2); |
5126 | gimple_assign_set_lhs (gs: g2, lhs: callfn); |
5127 | gsi_insert_before (gsi, g2, GSI_SAME_STMT); |
5128 | |
5129 | /* ??? This is a hack to preserve the NOTHROW bit on the call, |
5130 | which we would have derived from the decl. Failure to save |
5131 | this bit means we might have to split the basic block. */ |
5132 | if (gimple_call_nothrow_p (s: stmt)) |
5133 | gimple_call_set_nothrow (s: stmt, nothrow_p: true); |
5134 | |
5135 | gimple_call_set_fn (gs: stmt, fn: callfn); |
5136 | |
5137 | /* Discarding OBJ_TYPE_REF above may produce incompatible LHS and RHS |
5138 | for a call statement. Fix it. */ |
5139 | { |
5140 | tree lhs = gimple_call_lhs (gs: stmt); |
5141 | tree rettype = TREE_TYPE (gimple_call_fntype (stmt)); |
5142 | if (lhs |
5143 | && !useless_type_conversion_p (TREE_TYPE (lhs), rettype)) |
5144 | { |
5145 | tree temp; |
5146 | |
5147 | temp = create_tmp_reg (rettype); |
5148 | gimple_call_set_lhs (gs: stmt, lhs: temp); |
5149 | |
5150 | g2 = gimple_build_assign (lhs, |
5151 | fold_build1 (VIEW_CONVERT_EXPR, |
5152 | TREE_TYPE (lhs), temp)); |
5153 | gsi_insert_after (gsi, g2, GSI_SAME_STMT); |
5154 | } |
5155 | } |
5156 | |
5157 | update_stmt (s: stmt); |
5158 | cgraph_edge *e = cgraph_node::get (decl: current_function_decl)->get_edge (call_stmt: stmt); |
5159 | if (e && e->indirect_info) |
5160 | e->indirect_info->polymorphic = false; |
5161 | |
5162 | return true; |
5163 | } |
5164 | |
5165 | /* Helper function for ipa_tm_transform_calls*. Given a call |
5166 | statement in GSI which resides inside transaction REGION, redirect |
5167 | the call to either its wrapper function, or its clone. */ |
5168 | |
5169 | static void |
5170 | ipa_tm_transform_calls_redirect (struct cgraph_node *node, |
5171 | struct tm_region *region, |
5172 | gimple_stmt_iterator *gsi, |
5173 | bool *need_ssa_rename_p) |
5174 | { |
5175 | gcall *stmt = as_a <gcall *> (p: gsi_stmt (i: *gsi)); |
5176 | struct cgraph_node *new_node; |
5177 | struct cgraph_edge *e = node->get_edge (call_stmt: stmt); |
5178 | tree fndecl = gimple_call_fndecl (gs: stmt); |
5179 | |
5180 | /* For indirect calls, pass the address through the runtime. */ |
5181 | if (fndecl == NULL) |
5182 | { |
5183 | *need_ssa_rename_p |= |
5184 | ipa_tm_insert_gettmclone_call (node, region, gsi, stmt); |
5185 | return; |
5186 | } |
5187 | |
5188 | /* Handle some TM builtins. Ordinarily these aren't actually generated |
5189 | at this point, but handling these functions when written in by the |
5190 | user makes it easier to build unit tests. */ |
5191 | if (flags_from_decl_or_type (fndecl) & ECF_TM_BUILTIN) |
5192 | return; |
5193 | |
5194 | /* Fixup recursive calls inside clones. */ |
5195 | /* ??? Why did cgraph_copy_node_for_versioning update the call edges |
5196 | for recursion but not update the call statements themselves? */ |
5197 | if (e->caller == e->callee && decl_is_tm_clone (fndecl: current_function_decl)) |
5198 | { |
5199 | gimple_call_set_fndecl (gs: stmt, decl: current_function_decl); |
5200 | return; |
5201 | } |
5202 | |
5203 | /* If there is a replacement, use it. */ |
5204 | fndecl = find_tm_replacement_function (fndecl); |
5205 | if (fndecl) |
5206 | { |
5207 | new_node = cgraph_node::get_create (fndecl); |
5208 | |
5209 | /* ??? Mark all transaction_wrap functions tm_may_enter_irr. |
5210 | |
5211 | We can't do this earlier in record_tm_replacement because |
5212 | cgraph_remove_unreachable_nodes is called before we inject |
5213 | references to the node. Further, we can't do this in some |
5214 | nice central place in ipa_tm_execute because we don't have |
5215 | the exact list of wrapper functions that would be used. |
5216 | Marking more wrappers than necessary results in the creation |
5217 | of unnecessary cgraph_nodes, which can cause some of the |
5218 | other IPA passes to crash. |
5219 | |
5220 | We do need to mark these nodes so that we get the proper |
5221 | result in expand_call_tm. */ |
5222 | /* ??? This seems broken. How is it that we're marking the |
5223 | CALLEE as may_enter_irr? Surely we should be marking the |
5224 | CALLER. Also note that find_tm_replacement_function also |
5225 | contains mappings into the TM runtime, e.g. memcpy. These |
5226 | we know won't go irrevocable. */ |
5227 | new_node->tm_may_enter_irr = 1; |
5228 | } |
5229 | else |
5230 | { |
5231 | struct tm_ipa_cg_data *d; |
5232 | struct cgraph_node *tnode = e->callee; |
5233 | |
5234 | d = get_cg_data (node: &tnode, traverse_aliases: true); |
5235 | new_node = d->clone; |
5236 | |
5237 | /* As we've already skipped pure calls and appropriate builtins, |
5238 | and we've already marked irrevocable blocks, if we can't come |
5239 | up with a static replacement, then ask the runtime. */ |
5240 | if (new_node == NULL) |
5241 | { |
5242 | *need_ssa_rename_p |= |
5243 | ipa_tm_insert_gettmclone_call (node, region, gsi, stmt); |
5244 | return; |
5245 | } |
5246 | |
5247 | fndecl = new_node->decl; |
5248 | } |
5249 | |
5250 | e->redirect_callee (n: new_node); |
5251 | gimple_call_set_fndecl (gs: stmt, decl: fndecl); |
5252 | } |
5253 | |
5254 | /* Helper function for ipa_tm_transform_calls. For a given BB, |
5255 | install calls to tm_irrevocable when IRR_BLOCKS are reached, |
5256 | redirect other calls to the generated transactional clone. */ |
5257 | |
5258 | static bool |
5259 | ipa_tm_transform_calls_1 (struct cgraph_node *node, struct tm_region *region, |
5260 | basic_block bb, bitmap irr_blocks) |
5261 | { |
5262 | gimple_stmt_iterator gsi; |
5263 | bool need_ssa_rename = false; |
5264 | |
5265 | if (irr_blocks && bitmap_bit_p (irr_blocks, bb->index)) |
5266 | { |
5267 | ipa_tm_insert_irr_call (node, region, bb); |
5268 | return true; |
5269 | } |
5270 | |
5271 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
5272 | { |
5273 | gimple *stmt = gsi_stmt (i: gsi); |
5274 | |
5275 | if (!is_gimple_call (gs: stmt)) |
5276 | continue; |
5277 | if (is_tm_pure_call (call: stmt)) |
5278 | continue; |
5279 | |
5280 | /* Redirect edges to the appropriate replacement or clone. */ |
5281 | ipa_tm_transform_calls_redirect (node, region, gsi: &gsi, need_ssa_rename_p: &need_ssa_rename); |
5282 | } |
5283 | |
5284 | return need_ssa_rename; |
5285 | } |
5286 | |
5287 | /* Walk the CFG for REGION, beginning at BB. Install calls to |
5288 | tm_irrevocable when IRR_BLOCKS are reached, redirect other calls to |
5289 | the generated transactional clone. */ |
5290 | |
5291 | static bool |
5292 | ipa_tm_transform_calls (struct cgraph_node *node, struct tm_region *region, |
5293 | basic_block bb, bitmap irr_blocks) |
5294 | { |
5295 | bool need_ssa_rename = false; |
5296 | edge e; |
5297 | edge_iterator ei; |
5298 | auto_vec<basic_block> queue; |
5299 | bitmap visited_blocks = BITMAP_ALLOC (NULL); |
5300 | |
5301 | queue.safe_push (obj: bb); |
5302 | do |
5303 | { |
5304 | bb = queue.pop (); |
5305 | |
5306 | need_ssa_rename |= |
5307 | ipa_tm_transform_calls_1 (node, region, bb, irr_blocks); |
5308 | |
5309 | if (irr_blocks && bitmap_bit_p (irr_blocks, bb->index)) |
5310 | continue; |
5311 | |
5312 | if (region && bitmap_bit_p (region->exit_blocks, bb->index)) |
5313 | continue; |
5314 | |
5315 | FOR_EACH_EDGE (e, ei, bb->succs) |
5316 | if (!bitmap_bit_p (visited_blocks, e->dest->index)) |
5317 | { |
5318 | bitmap_set_bit (visited_blocks, e->dest->index); |
5319 | queue.safe_push (obj: e->dest); |
5320 | } |
5321 | } |
5322 | while (!queue.is_empty ()); |
5323 | |
5324 | BITMAP_FREE (visited_blocks); |
5325 | |
5326 | return need_ssa_rename; |
5327 | } |
5328 | |
5329 | /* Transform the calls within the TM regions within NODE. */ |
5330 | |
5331 | static void |
5332 | ipa_tm_transform_transaction (struct cgraph_node *node) |
5333 | { |
5334 | struct tm_ipa_cg_data *d; |
5335 | struct tm_region *region; |
5336 | bool need_ssa_rename = false; |
5337 | |
5338 | d = get_cg_data (node: &node, traverse_aliases: true); |
5339 | |
5340 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); |
5341 | calculate_dominance_info (CDI_DOMINATORS); |
5342 | |
5343 | for (region = d->all_tm_regions; region; region = region->next) |
5344 | { |
5345 | /* If we're sure to go irrevocable, don't transform anything. */ |
5346 | if (d->irrevocable_blocks_normal |
5347 | && bitmap_bit_p (d->irrevocable_blocks_normal, |
5348 | region->entry_block->index)) |
5349 | { |
5350 | transaction_subcode_ior (region, GTMA_DOES_GO_IRREVOCABLE |
5351 | | GTMA_MAY_ENTER_IRREVOCABLE |
5352 | | GTMA_HAS_NO_INSTRUMENTATION); |
5353 | continue; |
5354 | } |
5355 | |
5356 | need_ssa_rename |= |
5357 | ipa_tm_transform_calls (node, region, bb: region->entry_block, |
5358 | irr_blocks: d->irrevocable_blocks_normal); |
5359 | } |
5360 | |
5361 | if (need_ssa_rename) |
5362 | update_ssa (TODO_update_ssa_only_virtuals); |
5363 | |
5364 | pop_cfun (); |
5365 | } |
5366 | |
5367 | /* Transform the calls within the transactional clone of NODE. */ |
5368 | |
5369 | static void |
5370 | ipa_tm_transform_clone (struct cgraph_node *node) |
5371 | { |
5372 | struct tm_ipa_cg_data *d; |
5373 | bool need_ssa_rename; |
5374 | |
5375 | d = get_cg_data (node: &node, traverse_aliases: true); |
5376 | |
5377 | /* If this function makes no calls and has no irrevocable blocks, |
5378 | then there's nothing to do. */ |
5379 | /* ??? Remove non-aborting top-level transactions. */ |
5380 | if (!node->callees && !node->indirect_calls && !d->irrevocable_blocks_clone) |
5381 | return; |
5382 | |
5383 | push_cfun (DECL_STRUCT_FUNCTION (d->clone->decl)); |
5384 | calculate_dominance_info (CDI_DOMINATORS); |
5385 | |
5386 | need_ssa_rename = |
5387 | ipa_tm_transform_calls (node: d->clone, NULL, |
5388 | bb: single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)), |
5389 | irr_blocks: d->irrevocable_blocks_clone); |
5390 | |
5391 | if (need_ssa_rename) |
5392 | update_ssa (TODO_update_ssa_only_virtuals); |
5393 | |
5394 | pop_cfun (); |
5395 | } |
5396 | |
5397 | /* Main entry point for the transactional memory IPA pass. */ |
5398 | |
5399 | static unsigned int |
5400 | ipa_tm_execute (void) |
5401 | { |
5402 | cgraph_node_queue tm_callees = cgraph_node_queue (); |
5403 | /* List of functions that will go irrevocable. */ |
5404 | cgraph_node_queue irr_worklist = cgraph_node_queue (); |
5405 | |
5406 | struct cgraph_node *node; |
5407 | struct tm_ipa_cg_data *d; |
5408 | enum availability a; |
5409 | unsigned int i; |
5410 | |
5411 | cgraph_node::checking_verify_cgraph_nodes (); |
5412 | |
5413 | bitmap_obstack_initialize (&tm_obstack); |
5414 | initialize_original_copy_tables (); |
5415 | |
5416 | /* For all local functions marked tm_callable, queue them. */ |
5417 | FOR_EACH_DEFINED_FUNCTION (node) |
5418 | if (is_tm_callable (x: node->decl) |
5419 | && node->get_availability () >= AVAIL_INTERPOSABLE) |
5420 | { |
5421 | d = get_cg_data (node: &node, traverse_aliases: true); |
5422 | maybe_push_queue (node, queue_p: &tm_callees, in_queue_p: &d->in_callee_queue); |
5423 | } |
5424 | |
5425 | /* For all local reachable functions... */ |
5426 | FOR_EACH_DEFINED_FUNCTION (node) |
5427 | if (node->lowered |
5428 | && node->get_availability () >= AVAIL_INTERPOSABLE) |
5429 | { |
5430 | /* ... marked tm_pure, record that fact for the runtime by |
5431 | indicating that the pure function is its own tm_callable. |
5432 | No need to do this if the function's address can't be taken. */ |
5433 | if (is_tm_pure (x: node->decl)) |
5434 | { |
5435 | if (!node->local) |
5436 | record_tm_clone_pair (node->decl, node->decl); |
5437 | continue; |
5438 | } |
5439 | |
5440 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); |
5441 | calculate_dominance_info (CDI_DOMINATORS); |
5442 | |
5443 | tm_region_init (NULL); |
5444 | if (all_tm_regions) |
5445 | { |
5446 | d = get_cg_data (node: &node, traverse_aliases: true); |
5447 | |
5448 | /* Scan for calls that are in each transaction, and |
5449 | generate the uninstrumented code path. */ |
5450 | ipa_tm_scan_calls_transaction (d, callees_p: &tm_callees); |
5451 | |
5452 | /* Put it in the worklist so we can scan the function |
5453 | later (ipa_tm_scan_irr_function) and mark the |
5454 | irrevocable blocks. */ |
5455 | maybe_push_queue (node, queue_p: &irr_worklist, in_queue_p: &d->in_worklist); |
5456 | d->want_irr_scan_normal = true; |
5457 | } |
5458 | |
5459 | pop_cfun (); |
5460 | } |
5461 | |
5462 | /* For every local function on the callee list, scan as if we will be |
5463 | creating a transactional clone, queueing all new functions we find |
5464 | along the way. */ |
5465 | for (i = 0; i < tm_callees.length (); ++i) |
5466 | { |
5467 | node = tm_callees[i]; |
5468 | a = node->get_availability (); |
5469 | d = get_cg_data (node: &node, traverse_aliases: true); |
5470 | |
5471 | /* Put it in the worklist so we can scan the function later |
5472 | (ipa_tm_scan_irr_function) and mark the irrevocable |
5473 | blocks. */ |
5474 | maybe_push_queue (node, queue_p: &irr_worklist, in_queue_p: &d->in_worklist); |
5475 | |
5476 | /* Some callees cannot be arbitrarily cloned. These will always be |
5477 | irrevocable. Mark these now, so that we need not scan them. */ |
5478 | if (is_tm_irrevocable (x: node->decl)) |
5479 | ipa_tm_note_irrevocable (node, worklist_p: &irr_worklist); |
5480 | else if (a <= AVAIL_NOT_AVAILABLE |
5481 | && !is_tm_safe_or_pure (x: node->decl)) |
5482 | ipa_tm_note_irrevocable (node, worklist_p: &irr_worklist); |
5483 | else if (a >= AVAIL_INTERPOSABLE) |
5484 | { |
5485 | if (!tree_versionable_function_p (node->decl)) |
5486 | ipa_tm_note_irrevocable (node, worklist_p: &irr_worklist); |
5487 | else if (!d->is_irrevocable) |
5488 | { |
5489 | /* If this is an alias, make sure its base is queued as well. |
5490 | we need not scan the callees now, as the base will do. */ |
5491 | if (node->alias) |
5492 | { |
5493 | node = cgraph_node::get (decl: thunk_info::get (node)->alias); |
5494 | d = get_cg_data (node: &node, traverse_aliases: true); |
5495 | maybe_push_queue (node, queue_p: &tm_callees, in_queue_p: &d->in_callee_queue); |
5496 | continue; |
5497 | } |
5498 | |
5499 | /* Add all nodes called by this function into |
5500 | tm_callees as well. */ |
5501 | ipa_tm_scan_calls_clone (node, callees_p: &tm_callees); |
5502 | } |
5503 | } |
5504 | } |
5505 | |
5506 | /* Iterate scans until no more work to be done. Prefer not to use |
5507 | vec::pop because the worklist tends to follow a breadth-first |
5508 | search of the callgraph, which should allow convergance with a |
5509 | minimum number of scans. But we also don't want the worklist |
5510 | array to grow without bound, so we shift the array up periodically. */ |
5511 | for (i = 0; i < irr_worklist.length (); ++i) |
5512 | { |
5513 | if (i > 256 && i == irr_worklist.length () / 8) |
5514 | { |
5515 | irr_worklist.block_remove (ix: 0, len: i); |
5516 | i = 0; |
5517 | } |
5518 | |
5519 | node = irr_worklist[i]; |
5520 | d = get_cg_data (node: &node, traverse_aliases: true); |
5521 | d->in_worklist = false; |
5522 | |
5523 | if (d->want_irr_scan_normal) |
5524 | { |
5525 | d->want_irr_scan_normal = false; |
5526 | ipa_tm_scan_irr_function (node, for_clone: false); |
5527 | } |
5528 | if (d->in_callee_queue && ipa_tm_scan_irr_function (node, for_clone: true)) |
5529 | ipa_tm_note_irrevocable (node, worklist_p: &irr_worklist); |
5530 | } |
5531 | |
5532 | /* For every function on the callee list, collect the tm_may_enter_irr |
5533 | bit on the node. */ |
5534 | irr_worklist.truncate (size: 0); |
5535 | for (i = 0; i < tm_callees.length (); ++i) |
5536 | { |
5537 | node = tm_callees[i]; |
5538 | if (ipa_tm_mayenterirr_function (node)) |
5539 | { |
5540 | d = get_cg_data (node: &node, traverse_aliases: true); |
5541 | gcc_assert (d->in_worklist == false); |
5542 | maybe_push_queue (node, queue_p: &irr_worklist, in_queue_p: &d->in_worklist); |
5543 | } |
5544 | } |
5545 | |
5546 | /* Propagate the tm_may_enter_irr bit to callers until stable. */ |
5547 | for (i = 0; i < irr_worklist.length (); ++i) |
5548 | { |
5549 | struct cgraph_node *caller; |
5550 | struct cgraph_edge *e; |
5551 | struct ipa_ref *ref; |
5552 | |
5553 | if (i > 256 && i == irr_worklist.length () / 8) |
5554 | { |
5555 | irr_worklist.block_remove (ix: 0, len: i); |
5556 | i = 0; |
5557 | } |
5558 | |
5559 | node = irr_worklist[i]; |
5560 | d = get_cg_data (node: &node, traverse_aliases: true); |
5561 | d->in_worklist = false; |
5562 | node->tm_may_enter_irr = true; |
5563 | |
5564 | /* Propagate back to normal callers. */ |
5565 | for (e = node->callers; e ; e = e->next_caller) |
5566 | { |
5567 | caller = e->caller; |
5568 | if (!is_tm_safe_or_pure (x: caller->decl) |
5569 | && !caller->tm_may_enter_irr) |
5570 | { |
5571 | d = get_cg_data (node: &caller, traverse_aliases: true); |
5572 | maybe_push_queue (node: caller, queue_p: &irr_worklist, in_queue_p: &d->in_worklist); |
5573 | } |
5574 | } |
5575 | |
5576 | /* Propagate back to referring aliases as well. */ |
5577 | FOR_EACH_ALIAS (node, ref) |
5578 | { |
5579 | caller = dyn_cast<cgraph_node *> (p: ref->referring); |
5580 | if (!caller->tm_may_enter_irr) |
5581 | { |
5582 | /* ?? Do not traverse aliases here. */ |
5583 | d = get_cg_data (node: &caller, traverse_aliases: false); |
5584 | maybe_push_queue (node: caller, queue_p: &irr_worklist, in_queue_p: &d->in_worklist); |
5585 | } |
5586 | } |
5587 | } |
5588 | |
5589 | /* Now validate all tm_safe functions, and all atomic regions in |
5590 | other functions. */ |
5591 | FOR_EACH_DEFINED_FUNCTION (node) |
5592 | if (node->lowered |
5593 | && node->get_availability () >= AVAIL_INTERPOSABLE) |
5594 | { |
5595 | d = get_cg_data (node: &node, traverse_aliases: true); |
5596 | if (is_tm_safe (x: node->decl)) |
5597 | ipa_tm_diagnose_tm_safe (node); |
5598 | else if (d->all_tm_regions) |
5599 | ipa_tm_diagnose_transaction (node, all_tm_regions: d->all_tm_regions); |
5600 | } |
5601 | |
5602 | /* Create clones. Do those that are not irrevocable and have a |
5603 | positive call count. Do those publicly visible functions that |
5604 | the user directed us to clone. */ |
5605 | for (i = 0; i < tm_callees.length (); ++i) |
5606 | { |
5607 | bool doit = false; |
5608 | |
5609 | node = tm_callees[i]; |
5610 | if (node->cpp_implicit_alias) |
5611 | continue; |
5612 | |
5613 | a = node->get_availability (); |
5614 | d = get_cg_data (node: &node, traverse_aliases: true); |
5615 | |
5616 | if (a <= AVAIL_NOT_AVAILABLE) |
5617 | doit = is_tm_callable (x: node->decl); |
5618 | else if (a <= AVAIL_AVAILABLE && is_tm_callable (x: node->decl)) |
5619 | doit = true; |
5620 | else if (!d->is_irrevocable |
5621 | && d->tm_callers_normal + d->tm_callers_clone > 0) |
5622 | doit = true; |
5623 | |
5624 | if (doit) |
5625 | ipa_tm_create_version (old_node: node); |
5626 | } |
5627 | |
5628 | /* Redirect calls to the new clones, and insert irrevocable marks. */ |
5629 | for (i = 0; i < tm_callees.length (); ++i) |
5630 | { |
5631 | node = tm_callees[i]; |
5632 | if (node->analyzed) |
5633 | { |
5634 | d = get_cg_data (node: &node, traverse_aliases: true); |
5635 | if (d->clone) |
5636 | ipa_tm_transform_clone (node); |
5637 | } |
5638 | } |
5639 | FOR_EACH_DEFINED_FUNCTION (node) |
5640 | if (node->lowered |
5641 | && node->get_availability () >= AVAIL_INTERPOSABLE) |
5642 | { |
5643 | d = get_cg_data (node: &node, traverse_aliases: true); |
5644 | if (d->all_tm_regions) |
5645 | ipa_tm_transform_transaction (node); |
5646 | } |
5647 | |
5648 | /* Free and clear all data structures. */ |
5649 | tm_callees.release (); |
5650 | irr_worklist.release (); |
5651 | bitmap_obstack_release (&tm_obstack); |
5652 | free_original_copy_tables (); |
5653 | |
5654 | FOR_EACH_FUNCTION (node) |
5655 | node->aux = NULL; |
5656 | |
5657 | cgraph_node::checking_verify_cgraph_nodes (); |
5658 | |
5659 | return 0; |
5660 | } |
5661 | |
5662 | namespace { |
5663 | |
5664 | const pass_data pass_data_ipa_tm = |
5665 | { |
5666 | .type: SIMPLE_IPA_PASS, /* type */ |
5667 | .name: "tmipa" , /* name */ |
5668 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
5669 | .tv_id: TV_TRANS_MEM, /* tv_id */ |
5670 | .properties_required: ( PROP_ssa | PROP_cfg ), /* properties_required */ |
5671 | .properties_provided: 0, /* properties_provided */ |
5672 | .properties_destroyed: 0, /* properties_destroyed */ |
5673 | .todo_flags_start: 0, /* todo_flags_start */ |
5674 | .todo_flags_finish: 0, /* todo_flags_finish */ |
5675 | }; |
5676 | |
5677 | class pass_ipa_tm : public simple_ipa_opt_pass |
5678 | { |
5679 | public: |
5680 | pass_ipa_tm (gcc::context *ctxt) |
5681 | : simple_ipa_opt_pass (pass_data_ipa_tm, ctxt) |
5682 | {} |
5683 | |
5684 | /* opt_pass methods: */ |
5685 | bool gate (function *) final override { return flag_tm; } |
5686 | unsigned int execute (function *) final override { return ipa_tm_execute (); } |
5687 | |
5688 | }; // class pass_ipa_tm |
5689 | |
5690 | } // anon namespace |
5691 | |
5692 | simple_ipa_opt_pass * |
5693 | make_pass_ipa_tm (gcc::context *ctxt) |
5694 | { |
5695 | return new pass_ipa_tm (ctxt); |
5696 | } |
5697 | |
5698 | #include "gt-trans-mem.h" |
5699 | |