]>
Commit | Line | Data |
---|---|---|
0a35513e | 1 | /* Passes for transactional memory support. |
99dee823 | 2 | Copyright (C) 2008-2021 Free Software Foundation, Inc. |
d4c61094 AH |
3 | Contributed by Richard Henderson <rth@redhat.com> |
4 | and Aldy Hernandez <aldyh@redhat.com>. | |
0a35513e AH |
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" | |
c7131fb2 | 25 | #include "backend.h" |
957060b5 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
40e23961 | 28 | #include "tree.h" |
c7131fb2 | 29 | #include "gimple.h" |
957060b5 AM |
30 | #include "cfghooks.h" |
31 | #include "tree-pass.h" | |
c7131fb2 | 32 | #include "ssa.h" |
957060b5 AM |
33 | #include "cgraph.h" |
34 | #include "gimple-pretty-print.h" | |
35 | #include "diagnostic-core.h" | |
40e23961 | 36 | #include "fold-const.h" |
2fb9a547 | 37 | #include "tree-eh.h" |
d8a2d370 | 38 | #include "calls.h" |
45b0be94 | 39 | #include "gimplify.h" |
5be5c238 | 40 | #include "gimple-iterator.h" |
18f429e2 | 41 | #include "gimplify-me.h" |
5be5c238 | 42 | #include "gimple-walk.h" |
442b4905 | 43 | #include "tree-cfg.h" |
442b4905 | 44 | #include "tree-into-ssa.h" |
0a35513e | 45 | #include "tree-inline.h" |
0a35513e AH |
46 | #include "demangle.h" |
47 | #include "output.h" | |
48 | #include "trans-mem.h" | |
0a35513e | 49 | #include "langhooks.h" |
7d776ee2 | 50 | #include "cfgloop.h" |
4484a35a | 51 | #include "tree-ssa-address.h" |
314e6352 ML |
52 | #include "stringpool.h" |
53 | #include "attribs.h" | |
67f3791f JH |
54 | #include "alloc-pool.h" |
55 | #include "symbol-summary.h" | |
56 | #include "symtab-thunks.h" | |
0a35513e | 57 | |
0a35513e AH |
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.c) 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 | ||
398b1daa AH |
152 | static void *expand_regions (struct tm_region *, |
153 | void *(*callback)(struct tm_region *, void *), | |
b5e10eac | 154 | void *, bool); |
398b1daa | 155 | |
0a35513e AH |
156 | \f |
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 | { | |
59b8926c MP |
164 | if (x == NULL_TREE) |
165 | return NULL_TREE; | |
166 | ||
0a35513e AH |
167 | switch (TREE_CODE (x)) |
168 | { | |
169 | case FUNCTION_DECL: | |
170 | return TYPE_ATTRIBUTES (TREE_TYPE (x)); | |
0a35513e AH |
171 | |
172 | default: | |
173 | if (TYPE_P (x)) | |
59b8926c | 174 | return NULL_TREE; |
0a35513e AH |
175 | x = TREE_TYPE (x); |
176 | if (TREE_CODE (x) != POINTER_TYPE) | |
59b8926c | 177 | return NULL_TREE; |
0a35513e AH |
178 | /* FALLTHRU */ |
179 | ||
180 | case POINTER_TYPE: | |
181 | x = TREE_TYPE (x); | |
182 | if (TREE_CODE (x) != FUNCTION_TYPE && TREE_CODE (x) != METHOD_TYPE) | |
59b8926c | 183 | return NULL_TREE; |
0a35513e AH |
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 ("transaction_unsafe", 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 | |
3d78e008 | 240 | && fndecl_built_in_p (x, BUILT_IN_TM_IRREVOCABLE)) |
0a35513e AH |
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 ("transaction_safe", attrs)) | |
257 | return true; | |
258 | if (lookup_attribute ("transaction_may_cancel_outer", 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 | |
355fe088 | 268 | is_tm_pure_call (gimple *call) |
0a35513e | 269 | { |
0e02fb26 | 270 | return (gimple_call_flags (call) & (ECF_CONST | ECF_TM_PURE)) != 0; |
0a35513e AH |
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 ("transaction_callable", attrs)) | |
282 | return true; | |
283 | if (lookup_attribute ("transaction_safe", attrs)) | |
284 | return true; | |
285 | if (lookup_attribute ("transaction_may_cancel_outer", 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 ("transaction_may_cancel_outer", 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 (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 | ||
d5ae1c25 TV |
322 | /* Return true if STMT is a built in function call that "ends" a |
323 | transaction. */ | |
324 | ||
325 | bool | |
355fe088 | 326 | is_tm_ending (gimple *stmt) |
d5ae1c25 TV |
327 | { |
328 | tree fndecl; | |
329 | ||
330 | if (gimple_code (stmt) != GIMPLE_CALL) | |
331 | return false; | |
332 | ||
333 | fndecl = gimple_call_fndecl (stmt); | |
334 | return (fndecl != NULL_TREE | |
335 | && is_tm_ending_fndecl (fndecl)); | |
336 | } | |
337 | ||
0a35513e AH |
338 | /* Return true if STMT is a TM load. */ |
339 | ||
340 | static bool | |
355fe088 | 341 | is_tm_load (gimple *stmt) |
0a35513e AH |
342 | { |
343 | tree fndecl; | |
344 | ||
345 | if (gimple_code (stmt) != GIMPLE_CALL) | |
346 | return false; | |
347 | ||
348 | fndecl = gimple_call_fndecl (stmt); | |
3d78e008 ML |
349 | return (fndecl |
350 | && fndecl_built_in_p (fndecl, BUILT_IN_NORMAL) | |
0a35513e AH |
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 | |
355fe088 | 358 | is_tm_simple_load (gimple *stmt) |
0a35513e AH |
359 | { |
360 | tree fndecl; | |
361 | ||
362 | if (gimple_code (stmt) != GIMPLE_CALL) | |
363 | return false; | |
364 | ||
365 | fndecl = gimple_call_fndecl (stmt); | |
3d78e008 | 366 | if (fndecl && fndecl_built_in_p (fndecl, BUILT_IN_NORMAL)) |
0a35513e AH |
367 | { |
368 | enum built_in_function fcode = DECL_FUNCTION_CODE (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 | |
355fe088 | 386 | is_tm_store (gimple *stmt) |
0a35513e AH |
387 | { |
388 | tree fndecl; | |
389 | ||
390 | if (gimple_code (stmt) != GIMPLE_CALL) | |
391 | return false; | |
392 | ||
393 | fndecl = gimple_call_fndecl (stmt); | |
3d78e008 ML |
394 | return (fndecl |
395 | && fndecl_built_in_p (fndecl, BUILT_IN_NORMAL) | |
0a35513e AH |
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 | |
355fe088 | 403 | is_tm_simple_store (gimple *stmt) |
0a35513e AH |
404 | { |
405 | tree fndecl; | |
406 | ||
407 | if (gimple_code (stmt) != GIMPLE_CALL) | |
408 | return false; | |
409 | ||
410 | fndecl = gimple_call_fndecl (stmt); | |
3d78e008 ML |
411 | if (fndecl |
412 | && fndecl_built_in_p (fndecl, BUILT_IN_NORMAL)) | |
0a35513e AH |
413 | { |
414 | enum built_in_function fcode = DECL_FUNCTION_CODE (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 | { | |
3d78e008 | 434 | return (fndecl && fndecl_built_in_p (fndecl, BUILT_IN_TM_ABORT)); |
0a35513e AH |
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 (BUILT_IN_TM_ABORT), 1, | |
444 | build_int_cst (integer_type_node, | |
445 | AR_USERABORT | |
446 | | (is_outer ? AR_OUTERABORT : 0))); | |
447 | } | |
0a35513e | 448 | \f |
5764ee3c | 449 | /* Map for arbitrary function replacement under TM, as created |
0a35513e AH |
450 | by the tm_wrap attribute. */ |
451 | ||
6c907cff | 452 | struct tm_wrapper_hasher : ggc_cache_ptr_hash<tree_map> |
d242408f TS |
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 | ||
08ec2754 RS |
461 | static int |
462 | keep_cache_entry (tree_map *&m) | |
463 | { | |
464 | return ggc_marked_p (m->base.from); | |
465 | } | |
d242408f TS |
466 | }; |
467 | ||
468 | static GTY((cache)) hash_table<tm_wrapper_hasher> *tm_wrap_map; | |
0a35513e AH |
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) | |
d242408f | 484 | tm_wrap_map = hash_table<tm_wrapper_hasher>::create_ggc (32); |
0a35513e | 485 | |
766090c2 | 486 | h = ggc_alloc<tree_map> (); |
0a35513e AH |
487 | h->hash = htab_hash_pointer (from); |
488 | h->base.from = from; | |
489 | h->to = to; | |
490 | ||
d242408f | 491 | slot = tm_wrap_map->find_slot_with_hash (h, h->hash, INSERT); |
0a35513e AH |
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); | |
d242408f | 506 | h = tm_wrap_map->find_with_hash (&in, in.hash); |
0a35513e AH |
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 (fndecl)) | |
517 | { | |
518 | case BUILT_IN_MEMCPY: | |
519 | return builtin_decl_explicit (BUILT_IN_TM_MEMCPY); | |
520 | case BUILT_IN_MEMMOVE: | |
521 | return builtin_decl_explicit (BUILT_IN_TM_MEMMOVE); | |
522 | case BUILT_IN_MEMSET: | |
523 | return builtin_decl_explicit (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 (from)) | |
547 | return; | |
548 | ||
549 | str = IDENTIFIER_POINTER (DECL_NAME (from)); | |
550 | ||
551 | if (!strcmp (str, "malloc")) | |
552 | to = builtin_decl_explicit (BUILT_IN_TM_MALLOC); | |
553 | else if (!strcmp (str, "calloc")) | |
554 | to = builtin_decl_explicit (BUILT_IN_TM_CALLOC); | |
555 | else if (!strcmp (str, "free")) | |
556 | to = builtin_decl_explicit (BUILT_IN_TM_FREE); | |
557 | else | |
558 | return; | |
559 | ||
560 | TREE_NOTHROW (to) = 0; | |
561 | ||
562 | record_tm_replacement (from, to); | |
563 | } | |
564 | \f | |
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; | |
0a35513e | 581 | unsigned int saw_volatile : 1; |
355fe088 | 582 | gimple *stmt; |
0a35513e AH |
583 | }; |
584 | ||
1d4afc04 | 585 | /* Return true if T is a volatile lvalue of some kind. */ |
a3770d3b AH |
586 | |
587 | static bool | |
1d4afc04 | 588 | volatile_lvalue_p (tree t) |
a3770d3b | 589 | { |
1d4afc04 | 590 | return ((SSA_VAR_P (t) || REFERENCE_CLASS_P (t)) |
a3770d3b AH |
591 | && TREE_THIS_VOLATILE (TREE_TYPE (t))); |
592 | } | |
593 | ||
0a35513e AH |
594 | /* Tree callback function for diagnose_tm pass. */ |
595 | ||
596 | static tree | |
1d4afc04 | 597 | diagnose_tm_1_op (tree *tp, int *walk_subtrees, void *data) |
0a35513e AH |
598 | { |
599 | struct walk_stmt_info *wi = (struct walk_stmt_info *) data; | |
600 | struct diagnose_tm *d = (struct diagnose_tm *) wi->info; | |
0a35513e | 601 | |
1d4afc04 JM |
602 | if (TYPE_P (*tp)) |
603 | *walk_subtrees = false; | |
604 | else if (volatile_lvalue_p (*tp) | |
605 | && !d->saw_volatile) | |
0a35513e AH |
606 | { |
607 | d->saw_volatile = 1; | |
1d4afc04 JM |
608 | if (d->block_flags & DIAG_TM_SAFE) |
609 | error_at (gimple_location (d->stmt), | |
610 | "invalid use of volatile lvalue inside transaction"); | |
611 | else if (d->func_flags & DIAG_TM_SAFE) | |
612 | error_at (gimple_location (d->stmt), | |
aa326bfb | 613 | "invalid use of volatile lvalue inside %<transaction_safe%> " |
1d4afc04 | 614 | "function"); |
0a35513e AH |
615 | } |
616 | ||
617 | return NULL_TREE; | |
618 | } | |
619 | ||
862d0b35 DN |
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 | ||
0a35513e AH |
626 | static tree |
627 | diagnose_tm_1 (gimple_stmt_iterator *gsi, bool *handled_ops_p, | |
628 | struct walk_stmt_info *wi) | |
629 | { | |
355fe088 | 630 | gimple *stmt = gsi_stmt (*gsi); |
0a35513e AH |
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 (stmt)) | |
637 | { | |
638 | case GIMPLE_CALL: | |
639 | { | |
640 | tree fn = gimple_call_fn (stmt); | |
641 | ||
642 | if ((d->summary_flags & DIAG_TM_OUTER) == 0 | |
643 | && is_tm_may_cancel_outer (fn)) | |
644 | error_at (gimple_location (stmt), | |
645 | "%<transaction_may_cancel_outer%> function call not within" | |
646 | " outer transaction or %<transaction_may_cancel_outer%>"); | |
647 | ||
648 | if (d->summary_flags & DIAG_TM_SAFE) | |
649 | { | |
650 | bool is_safe, direct_call_p; | |
651 | tree replacement; | |
652 | ||
653 | if (TREE_CODE (fn) == ADDR_EXPR | |
654 | && TREE_CODE (TREE_OPERAND (fn, 0)) == FUNCTION_DECL) | |
655 | { | |
656 | direct_call_p = true; | |
657 | replacement = TREE_OPERAND (fn, 0); | |
658 | replacement = find_tm_replacement_function (replacement); | |
659 | if (replacement) | |
660 | fn = replacement; | |
661 | } | |
662 | else | |
663 | { | |
664 | direct_call_p = false; | |
665 | replacement = NULL_TREE; | |
666 | } | |
667 | ||
668 | if (is_tm_safe_or_pure (fn)) | |
669 | is_safe = true; | |
670 | else if (is_tm_callable (fn) || is_tm_irrevocable (fn)) | |
671 | { | |
672 | /* A function explicitly marked transaction_callable as | |
673 | opposed to transaction_safe is being defined to be | |
674 | unsafe as part of its ABI, regardless of its contents. */ | |
675 | is_safe = false; | |
676 | } | |
677 | else if (direct_call_p) | |
678 | { | |
17fc8d6f AH |
679 | if (IS_TYPE_OR_DECL_P (fn) |
680 | && flags_from_decl_or_type (fn) & ECF_TM_BUILTIN) | |
0a35513e AH |
681 | is_safe = true; |
682 | else if (replacement) | |
683 | { | |
684 | /* ??? At present we've been considering replacements | |
685 | merely transaction_callable, and therefore might | |
686 | enter irrevocable. The tm_wrap attribute has not | |
687 | yet made it into the new language spec. */ | |
688 | is_safe = false; | |
689 | } | |
690 | else | |
691 | { | |
692 | /* ??? Diagnostics for unmarked direct calls moved into | |
693 | the IPA pass. Section 3.2 of the spec details how | |
694 | functions not marked should be considered "implicitly | |
695 | safe" based on having examined the function body. */ | |
696 | is_safe = true; | |
697 | } | |
698 | } | |
699 | else | |
700 | { | |
701 | /* An unmarked indirect call. Consider it unsafe even | |
702 | though optimization may yet figure out how to inline. */ | |
703 | is_safe = false; | |
704 | } | |
705 | ||
706 | if (!is_safe) | |
707 | { | |
708 | if (TREE_CODE (fn) == ADDR_EXPR) | |
709 | fn = TREE_OPERAND (fn, 0); | |
710 | if (d->block_flags & DIAG_TM_SAFE) | |
cd6baa16 TR |
711 | { |
712 | if (direct_call_p) | |
713 | error_at (gimple_location (stmt), | |
714 | "unsafe function call %qD within " | |
715 | "atomic transaction", fn); | |
716 | else | |
3a54c456 | 717 | { |
381cdae4 RB |
718 | if ((!DECL_P (fn) || DECL_NAME (fn)) |
719 | && TREE_CODE (fn) != SSA_NAME) | |
3a54c456 AH |
720 | error_at (gimple_location (stmt), |
721 | "unsafe function call %qE within " | |
722 | "atomic transaction", fn); | |
723 | else | |
724 | error_at (gimple_location (stmt), | |
725 | "unsafe indirect function call within " | |
726 | "atomic transaction"); | |
727 | } | |
cd6baa16 | 728 | } |
0a35513e | 729 | else |
cd6baa16 TR |
730 | { |
731 | if (direct_call_p) | |
732 | error_at (gimple_location (stmt), | |
733 | "unsafe function call %qD within " | |
734 | "%<transaction_safe%> function", fn); | |
735 | else | |
3a54c456 | 736 | { |
381cdae4 RB |
737 | if ((!DECL_P (fn) || DECL_NAME (fn)) |
738 | && TREE_CODE (fn) != SSA_NAME) | |
3a54c456 AH |
739 | error_at (gimple_location (stmt), |
740 | "unsafe function call %qE within " | |
741 | "%<transaction_safe%> function", fn); | |
742 | else | |
743 | error_at (gimple_location (stmt), | |
744 | "unsafe indirect function call within " | |
745 | "%<transaction_safe%> function"); | |
746 | } | |
cd6baa16 | 747 | } |
0a35513e AH |
748 | } |
749 | } | |
750 | } | |
751 | break; | |
752 | ||
753 | case GIMPLE_ASM: | |
754 | /* ??? We ought to come up with a way to add attributes to | |
755 | asm statements, and then add "transaction_safe" to it. | |
756 | Either that or get the language spec to resurrect __tm_waiver. */ | |
757 | if (d->block_flags & DIAG_TM_SAFE) | |
758 | error_at (gimple_location (stmt), | |
a9c697b8 | 759 | "%<asm%> not allowed in atomic transaction"); |
0a35513e AH |
760 | else if (d->func_flags & DIAG_TM_SAFE) |
761 | error_at (gimple_location (stmt), | |
a9c697b8 | 762 | "%<asm%> not allowed in %<transaction_safe%> function"); |
0a35513e AH |
763 | break; |
764 | ||
765 | case GIMPLE_TRANSACTION: | |
766 | { | |
538dd0b7 | 767 | gtransaction *trans_stmt = as_a <gtransaction *> (stmt); |
0a35513e AH |
768 | unsigned char inner_flags = DIAG_TM_SAFE; |
769 | ||
538dd0b7 | 770 | if (gimple_transaction_subcode (trans_stmt) & GTMA_IS_RELAXED) |
0a35513e AH |
771 | { |
772 | if (d->block_flags & DIAG_TM_SAFE) | |
773 | error_at (gimple_location (stmt), | |
774 | "relaxed transaction in atomic transaction"); | |
775 | else if (d->func_flags & DIAG_TM_SAFE) | |
776 | error_at (gimple_location (stmt), | |
777 | "relaxed transaction in %<transaction_safe%> function"); | |
0a35513e AH |
778 | inner_flags = DIAG_TM_RELAXED; |
779 | } | |
538dd0b7 | 780 | else if (gimple_transaction_subcode (trans_stmt) & GTMA_IS_OUTER) |
0a35513e AH |
781 | { |
782 | if (d->block_flags) | |
783 | error_at (gimple_location (stmt), | |
784 | "outer transaction in transaction"); | |
785 | else if (d->func_flags & DIAG_TM_OUTER) | |
786 | error_at (gimple_location (stmt), | |
787 | "outer transaction in " | |
788 | "%<transaction_may_cancel_outer%> function"); | |
789 | else if (d->func_flags & DIAG_TM_SAFE) | |
790 | error_at (gimple_location (stmt), | |
791 | "outer transaction in %<transaction_safe%> function"); | |
0a35513e AH |
792 | inner_flags |= DIAG_TM_OUTER; |
793 | } | |
794 | ||
795 | *handled_ops_p = true; | |
538dd0b7 | 796 | if (gimple_transaction_body (trans_stmt)) |
0a35513e AH |
797 | { |
798 | struct walk_stmt_info wi_inner; | |
799 | struct diagnose_tm d_inner; | |
800 | ||
801 | memset (&d_inner, 0, sizeof (d_inner)); | |
802 | d_inner.func_flags = d->func_flags; | |
803 | d_inner.block_flags = d->block_flags | inner_flags; | |
804 | d_inner.summary_flags = d_inner.func_flags | d_inner.block_flags; | |
805 | ||
806 | memset (&wi_inner, 0, sizeof (wi_inner)); | |
807 | wi_inner.info = &d_inner; | |
808 | ||
538dd0b7 | 809 | walk_gimple_seq (gimple_transaction_body (trans_stmt), |
0a35513e | 810 | diagnose_tm_1, diagnose_tm_1_op, &wi_inner); |
0a35513e AH |
811 | } |
812 | } | |
813 | break; | |
814 | ||
815 | default: | |
816 | break; | |
817 | } | |
818 | ||
819 | return NULL_TREE; | |
820 | } | |
821 | ||
822 | static unsigned int | |
823 | diagnose_tm_blocks (void) | |
824 | { | |
825 | struct walk_stmt_info wi; | |
826 | struct diagnose_tm d; | |
827 | ||
828 | memset (&d, 0, sizeof (d)); | |
829 | if (is_tm_may_cancel_outer (current_function_decl)) | |
830 | d.func_flags = DIAG_TM_OUTER | DIAG_TM_SAFE; | |
831 | else if (is_tm_safe (current_function_decl)) | |
832 | d.func_flags = DIAG_TM_SAFE; | |
833 | d.summary_flags = d.func_flags; | |
834 | ||
835 | memset (&wi, 0, sizeof (wi)); | |
836 | wi.info = &d; | |
837 | ||
838 | walk_gimple_seq (gimple_body (current_function_decl), | |
839 | diagnose_tm_1, diagnose_tm_1_op, &wi); | |
840 | ||
0a35513e AH |
841 | return 0; |
842 | } | |
843 | ||
17795822 TS |
844 | namespace { |
845 | ||
846 | const pass_data pass_data_diagnose_tm_blocks = | |
27a4cd48 DM |
847 | { |
848 | GIMPLE_PASS, /* type */ | |
849 | "*diagnose_tm_blocks", /* name */ | |
850 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
851 | TV_TRANS_MEM, /* tv_id */ |
852 | PROP_gimple_any, /* properties_required */ | |
853 | 0, /* properties_provided */ | |
854 | 0, /* properties_destroyed */ | |
855 | 0, /* todo_flags_start */ | |
856 | 0, /* todo_flags_finish */ | |
0a35513e | 857 | }; |
27a4cd48 | 858 | |
17795822 | 859 | class pass_diagnose_tm_blocks : public gimple_opt_pass |
27a4cd48 DM |
860 | { |
861 | public: | |
c3284718 RS |
862 | pass_diagnose_tm_blocks (gcc::context *ctxt) |
863 | : gimple_opt_pass (pass_data_diagnose_tm_blocks, ctxt) | |
27a4cd48 DM |
864 | {} |
865 | ||
866 | /* opt_pass methods: */ | |
1a3d085c | 867 | virtual bool gate (function *) { return flag_tm; } |
be55bfe6 | 868 | virtual unsigned int execute (function *) { return diagnose_tm_blocks (); } |
27a4cd48 DM |
869 | |
870 | }; // class pass_diagnose_tm_blocks | |
871 | ||
17795822 TS |
872 | } // anon namespace |
873 | ||
27a4cd48 DM |
874 | gimple_opt_pass * |
875 | make_pass_diagnose_tm_blocks (gcc::context *ctxt) | |
876 | { | |
877 | return new pass_diagnose_tm_blocks (ctxt); | |
878 | } | |
0a35513e AH |
879 | \f |
880 | /* Instead of instrumenting thread private memory, we save the | |
881 | addresses in a log which we later use to save/restore the addresses | |
882 | upon transaction start/restart. | |
883 | ||
884 | The log is keyed by address, where each element contains individual | |
885 | statements among different code paths that perform the store. | |
886 | ||
887 | This log is later used to generate either plain save/restore of the | |
888 | addresses upon transaction start/restart, or calls to the ITM_L* | |
889 | logging functions. | |
890 | ||
891 | So for something like: | |
892 | ||
893 | struct large { int x[1000]; }; | |
894 | struct large lala = { 0 }; | |
895 | __transaction { | |
896 | lala.x[i] = 123; | |
897 | ... | |
898 | } | |
899 | ||
900 | We can either save/restore: | |
901 | ||
902 | lala = { 0 }; | |
903 | trxn = _ITM_startTransaction (); | |
904 | if (trxn & a_saveLiveVariables) | |
905 | tmp_lala1 = lala.x[i]; | |
906 | else if (a & a_restoreLiveVariables) | |
907 | lala.x[i] = tmp_lala1; | |
908 | ||
909 | or use the logging functions: | |
910 | ||
911 | lala = { 0 }; | |
912 | trxn = _ITM_startTransaction (); | |
913 | _ITM_LU4 (&lala.x[i]); | |
914 | ||
915 | Obviously, if we use _ITM_L* to log, we prefer to call _ITM_L* as | |
916 | far up the dominator tree to shadow all of the writes to a given | |
917 | location (thus reducing the total number of logging calls), but not | |
918 | so high as to be called on a path that does not perform a | |
919 | write. */ | |
920 | ||
921 | /* One individual log entry. We may have multiple statements for the | |
922 | same location if neither dominate each other (on different | |
923 | execution paths). */ | |
526ceb68 | 924 | struct tm_log_entry |
0a35513e AH |
925 | { |
926 | /* Address to save. */ | |
927 | tree addr; | |
928 | /* Entry block for the transaction this address occurs in. */ | |
929 | basic_block entry_block; | |
930 | /* Dominating statements the store occurs in. */ | |
355fe088 | 931 | vec<gimple *> stmts; |
0a35513e AH |
932 | /* Initially, while we are building the log, we place a nonzero |
933 | value here to mean that this address *will* be saved with a | |
934 | save/restore sequence. Later, when generating the save sequence | |
935 | we place the SSA temp generated here. */ | |
936 | tree save_var; | |
526ceb68 | 937 | }; |
0a35513e | 938 | |
0a35513e | 939 | |
4a8fb1a1 | 940 | /* Log entry hashtable helpers. */ |
0a35513e | 941 | |
7edd9b15 | 942 | struct log_entry_hasher : pointer_hash <tm_log_entry> |
0a35513e | 943 | { |
67f58944 TS |
944 | static inline hashval_t hash (const tm_log_entry *); |
945 | static inline bool equal (const tm_log_entry *, const tm_log_entry *); | |
946 | static inline void remove (tm_log_entry *); | |
4a8fb1a1 | 947 | }; |
0a35513e AH |
948 | |
949 | /* Htab support. Return hash value for a `tm_log_entry'. */ | |
4a8fb1a1 | 950 | inline hashval_t |
67f58944 | 951 | log_entry_hasher::hash (const tm_log_entry *log) |
0a35513e | 952 | { |
0a35513e AH |
953 | return iterative_hash_expr (log->addr, 0); |
954 | } | |
955 | ||
956 | /* Htab support. Return true if two log entries are the same. */ | |
4a8fb1a1 | 957 | inline bool |
67f58944 | 958 | log_entry_hasher::equal (const tm_log_entry *log1, const tm_log_entry *log2) |
0a35513e | 959 | { |
0a35513e AH |
960 | /* FIXME: |
961 | ||
962 | rth: I suggest that we get rid of the component refs etc. | |
963 | I.e. resolve the reference to base + offset. | |
964 | ||
965 | We may need to actually finish a merge with mainline for this, | |
966 | since we'd like to be presented with Richi's MEM_REF_EXPRs more | |
967 | often than not. But in the meantime your tm_log_entry could save | |
968 | the results of get_inner_reference. | |
969 | ||
970 | See: g++.dg/tm/pr46653.C | |
971 | */ | |
972 | ||
973 | /* Special case plain equality because operand_equal_p() below will | |
974 | return FALSE if the addresses are equal but they have | |
975 | side-effects (e.g. a volatile address). */ | |
976 | if (log1->addr == log2->addr) | |
977 | return true; | |
978 | ||
979 | return operand_equal_p (log1->addr, log2->addr, 0); | |
980 | } | |
981 | ||
982 | /* Htab support. Free one tm_log_entry. */ | |
4a8fb1a1 | 983 | inline void |
67f58944 | 984 | log_entry_hasher::remove (tm_log_entry *lp) |
0a35513e | 985 | { |
9771b263 | 986 | lp->stmts.release (); |
0a35513e AH |
987 | free (lp); |
988 | } | |
989 | ||
4a8fb1a1 LC |
990 | |
991 | /* The actual log. */ | |
c203e8a7 | 992 | static hash_table<log_entry_hasher> *tm_log; |
4a8fb1a1 LC |
993 | |
994 | /* Addresses to log with a save/restore sequence. These should be in | |
995 | dominator order. */ | |
996 | static vec<tree> tm_log_save_addresses; | |
997 | ||
998 | enum thread_memory_type | |
999 | { | |
1000 | mem_non_local = 0, | |
1001 | mem_thread_local, | |
1002 | mem_transaction_local, | |
1003 | mem_max | |
1004 | }; | |
1005 | ||
526ceb68 | 1006 | struct tm_new_mem_map |
4a8fb1a1 LC |
1007 | { |
1008 | /* SSA_NAME being dereferenced. */ | |
1009 | tree val; | |
1010 | enum thread_memory_type local_new_memory; | |
526ceb68 | 1011 | }; |
4a8fb1a1 LC |
1012 | |
1013 | /* Hashtable helpers. */ | |
1014 | ||
526ceb68 | 1015 | struct tm_mem_map_hasher : free_ptr_hash <tm_new_mem_map> |
4a8fb1a1 | 1016 | { |
526ceb68 TS |
1017 | static inline hashval_t hash (const tm_new_mem_map *); |
1018 | static inline bool equal (const tm_new_mem_map *, const tm_new_mem_map *); | |
4a8fb1a1 LC |
1019 | }; |
1020 | ||
1021 | inline hashval_t | |
526ceb68 | 1022 | tm_mem_map_hasher::hash (const tm_new_mem_map *v) |
4a8fb1a1 LC |
1023 | { |
1024 | return (intptr_t)v->val >> 4; | |
1025 | } | |
1026 | ||
1027 | inline bool | |
526ceb68 | 1028 | tm_mem_map_hasher::equal (const tm_new_mem_map *v, const tm_new_mem_map *c) |
4a8fb1a1 LC |
1029 | { |
1030 | return v->val == c->val; | |
1031 | } | |
1032 | ||
1033 | /* Map for an SSA_NAME originally pointing to a non aliased new piece | |
1034 | of memory (malloc, alloc, etc). */ | |
c203e8a7 | 1035 | static hash_table<tm_mem_map_hasher> *tm_new_mem_hash; |
4a8fb1a1 | 1036 | |
0a35513e AH |
1037 | /* Initialize logging data structures. */ |
1038 | static void | |
1039 | tm_log_init (void) | |
1040 | { | |
c203e8a7 TS |
1041 | tm_log = new hash_table<log_entry_hasher> (10); |
1042 | tm_new_mem_hash = new hash_table<tm_mem_map_hasher> (5); | |
9771b263 | 1043 | tm_log_save_addresses.create (5); |
0a35513e AH |
1044 | } |
1045 | ||
1046 | /* Free logging data structures. */ | |
1047 | static void | |
1048 | tm_log_delete (void) | |
1049 | { | |
c203e8a7 TS |
1050 | delete tm_log; |
1051 | tm_log = NULL; | |
1052 | delete tm_new_mem_hash; | |
1053 | tm_new_mem_hash = NULL; | |
9771b263 | 1054 | tm_log_save_addresses.release (); |
0a35513e AH |
1055 | } |
1056 | ||
1057 | /* Return true if MEM is a transaction invariant memory for the TM | |
1058 | region starting at REGION_ENTRY_BLOCK. */ | |
1059 | static bool | |
1060 | transaction_invariant_address_p (const_tree mem, basic_block region_entry_block) | |
1061 | { | |
1062 | if ((TREE_CODE (mem) == INDIRECT_REF || TREE_CODE (mem) == MEM_REF) | |
1063 | && TREE_CODE (TREE_OPERAND (mem, 0)) == SSA_NAME) | |
1064 | { | |
1065 | basic_block def_bb; | |
1066 | ||
1067 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (TREE_OPERAND (mem, 0))); | |
1068 | return def_bb != region_entry_block | |
1069 | && dominated_by_p (CDI_DOMINATORS, region_entry_block, def_bb); | |
1070 | } | |
1071 | ||
1072 | mem = strip_invariant_refs (mem); | |
1073 | return mem && (CONSTANT_CLASS_P (mem) || decl_address_invariant_p (mem)); | |
1074 | } | |
1075 | ||
1076 | /* Given an address ADDR in STMT, find it in the memory log or add it, | |
1077 | making sure to keep only the addresses highest in the dominator | |
1078 | tree. | |
1079 | ||
1080 | ENTRY_BLOCK is the entry_block for the transaction. | |
1081 | ||
1082 | If we find the address in the log, make sure it's either the same | |
1083 | address, or an equivalent one that dominates ADDR. | |
1084 | ||
1085 | If we find the address, but neither ADDR dominates the found | |
1086 | address, nor the found one dominates ADDR, we're on different | |
1087 | execution paths. Add it. | |
1088 | ||
1089 | If known, ENTRY_BLOCK is the entry block for the region, otherwise | |
1090 | NULL. */ | |
1091 | static void | |
355fe088 | 1092 | tm_log_add (basic_block entry_block, tree addr, gimple *stmt) |
0a35513e | 1093 | { |
4a8fb1a1 | 1094 | tm_log_entry **slot; |
0a35513e AH |
1095 | struct tm_log_entry l, *lp; |
1096 | ||
1097 | l.addr = addr; | |
c203e8a7 | 1098 | slot = tm_log->find_slot (&l, INSERT); |
0a35513e AH |
1099 | if (!*slot) |
1100 | { | |
1101 | tree type = TREE_TYPE (addr); | |
1102 | ||
1103 | lp = XNEW (struct tm_log_entry); | |
1104 | lp->addr = addr; | |
1105 | *slot = lp; | |
1106 | ||
1107 | /* Small invariant addresses can be handled as save/restores. */ | |
1108 | if (entry_block | |
1109 | && transaction_invariant_address_p (lp->addr, entry_block) | |
1110 | && TYPE_SIZE_UNIT (type) != NULL | |
cc269bb6 | 1111 | && tree_fits_uhwi_p (TYPE_SIZE_UNIT (type)) |
7d362f6c | 1112 | && ((HOST_WIDE_INT) tree_to_uhwi (TYPE_SIZE_UNIT (type)) |
028d4092 | 1113 | < param_tm_max_aggregate_size) |
0a35513e AH |
1114 | /* We must be able to copy this type normally. I.e., no |
1115 | special constructors and the like. */ | |
1116 | && !TREE_ADDRESSABLE (type)) | |
1117 | { | |
2091795a | 1118 | lp->save_var = create_tmp_reg (TREE_TYPE (lp->addr), "tm_save"); |
9771b263 | 1119 | lp->stmts.create (0); |
0a35513e AH |
1120 | lp->entry_block = entry_block; |
1121 | /* Save addresses separately in dominator order so we don't | |
1122 | get confused by overlapping addresses in the save/restore | |
1123 | sequence. */ | |
9771b263 | 1124 | tm_log_save_addresses.safe_push (lp->addr); |
0a35513e AH |
1125 | } |
1126 | else | |
1127 | { | |
1128 | /* Use the logging functions. */ | |
9771b263 DN |
1129 | lp->stmts.create (5); |
1130 | lp->stmts.quick_push (stmt); | |
0a35513e AH |
1131 | lp->save_var = NULL; |
1132 | } | |
1133 | } | |
1134 | else | |
1135 | { | |
1136 | size_t i; | |
355fe088 | 1137 | gimple *oldstmt; |
0a35513e | 1138 | |
4a8fb1a1 | 1139 | lp = *slot; |
0a35513e AH |
1140 | |
1141 | /* If we're generating a save/restore sequence, we don't care | |
1142 | about statements. */ | |
1143 | if (lp->save_var) | |
1144 | return; | |
1145 | ||
9771b263 | 1146 | for (i = 0; lp->stmts.iterate (i, &oldstmt); ++i) |
0a35513e AH |
1147 | { |
1148 | if (stmt == oldstmt) | |
1149 | return; | |
1150 | /* We already have a store to the same address, higher up the | |
1151 | dominator tree. Nothing to do. */ | |
1152 | if (dominated_by_p (CDI_DOMINATORS, | |
1153 | gimple_bb (stmt), gimple_bb (oldstmt))) | |
1154 | return; | |
1155 | /* We should be processing blocks in dominator tree order. */ | |
1156 | gcc_assert (!dominated_by_p (CDI_DOMINATORS, | |
1157 | gimple_bb (oldstmt), gimple_bb (stmt))); | |
1158 | } | |
1159 | /* Store is on a different code path. */ | |
9771b263 | 1160 | lp->stmts.safe_push (stmt); |
0a35513e AH |
1161 | } |
1162 | } | |
1163 | ||
1164 | /* Gimplify the address of a TARGET_MEM_REF. Return the SSA_NAME | |
1165 | result, insert the new statements before GSI. */ | |
1166 | ||
1167 | static tree | |
1168 | gimplify_addr (gimple_stmt_iterator *gsi, tree x) | |
1169 | { | |
1170 | if (TREE_CODE (x) == TARGET_MEM_REF) | |
1171 | x = tree_mem_ref_addr (build_pointer_type (TREE_TYPE (x)), x); | |
1172 | else | |
1173 | x = build_fold_addr_expr (x); | |
1174 | return force_gimple_operand_gsi (gsi, x, true, NULL, true, GSI_SAME_STMT); | |
1175 | } | |
1176 | ||
1177 | /* Instrument one address with the logging functions. | |
1178 | ADDR is the address to save. | |
1179 | STMT is the statement before which to place it. */ | |
1180 | static void | |
355fe088 | 1181 | tm_log_emit_stmt (tree addr, gimple *stmt) |
0a35513e AH |
1182 | { |
1183 | tree type = TREE_TYPE (addr); | |
0a35513e | 1184 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); |
355fe088 | 1185 | gimple *log; |
0a35513e AH |
1186 | enum built_in_function code = BUILT_IN_TM_LOG; |
1187 | ||
1188 | if (type == float_type_node) | |
1189 | code = BUILT_IN_TM_LOG_FLOAT; | |
1190 | else if (type == double_type_node) | |
1191 | code = BUILT_IN_TM_LOG_DOUBLE; | |
1192 | else if (type == long_double_type_node) | |
1193 | code = BUILT_IN_TM_LOG_LDOUBLE; | |
4c868789 RH |
1194 | else if (TYPE_SIZE (type) != NULL |
1195 | && tree_fits_uhwi_p (TYPE_SIZE (type))) | |
0a35513e | 1196 | { |
4c868789 RH |
1197 | unsigned HOST_WIDE_INT type_size = tree_to_uhwi (TYPE_SIZE (type)); |
1198 | ||
1199 | if (TREE_CODE (type) == VECTOR_TYPE) | |
0a35513e | 1200 | { |
4c868789 | 1201 | switch (type_size) |
0a35513e | 1202 | { |
4c868789 RH |
1203 | case 64: |
1204 | code = BUILT_IN_TM_LOG_M64; | |
1205 | break; | |
1206 | case 128: | |
1207 | code = BUILT_IN_TM_LOG_M128; | |
1208 | break; | |
1209 | case 256: | |
1210 | code = BUILT_IN_TM_LOG_M256; | |
1211 | break; | |
1212 | default: | |
1213 | goto unhandled_vec; | |
1214 | } | |
1215 | if (!builtin_decl_explicit_p (code)) | |
1216 | goto unhandled_vec; | |
1217 | } | |
1218 | else | |
1219 | { | |
1220 | unhandled_vec: | |
1221 | switch (type_size) | |
1222 | { | |
47c147b0 | 1223 | case 8: |
4c868789 RH |
1224 | code = BUILT_IN_TM_LOG_1; |
1225 | break; | |
47c147b0 | 1226 | case 16: |
4c868789 RH |
1227 | code = BUILT_IN_TM_LOG_2; |
1228 | break; | |
47c147b0 | 1229 | case 32: |
4c868789 RH |
1230 | code = BUILT_IN_TM_LOG_4; |
1231 | break; | |
47c147b0 | 1232 | case 64: |
4c868789 RH |
1233 | code = BUILT_IN_TM_LOG_8; |
1234 | break; | |
0a35513e | 1235 | } |
0a35513e AH |
1236 | } |
1237 | } | |
1238 | ||
4c868789 RH |
1239 | if (code != BUILT_IN_TM_LOG && !builtin_decl_explicit_p (code)) |
1240 | code = BUILT_IN_TM_LOG; | |
1241 | tree decl = builtin_decl_explicit (code); | |
1242 | ||
0a35513e AH |
1243 | addr = gimplify_addr (&gsi, addr); |
1244 | if (code == BUILT_IN_TM_LOG) | |
4c868789 | 1245 | log = gimple_build_call (decl, 2, addr, TYPE_SIZE_UNIT (type)); |
0a35513e | 1246 | else |
4c868789 | 1247 | log = gimple_build_call (decl, 1, addr); |
0a35513e AH |
1248 | gsi_insert_before (&gsi, log, GSI_SAME_STMT); |
1249 | } | |
1250 | ||
1251 | /* Go through the log and instrument address that must be instrumented | |
1252 | with the logging functions. Leave the save/restore addresses for | |
1253 | later. */ | |
1254 | static void | |
1255 | tm_log_emit (void) | |
1256 | { | |
c203e8a7 | 1257 | hash_table<log_entry_hasher>::iterator hi; |
0a35513e AH |
1258 | struct tm_log_entry *lp; |
1259 | ||
c203e8a7 | 1260 | FOR_EACH_HASH_TABLE_ELEMENT (*tm_log, lp, tm_log_entry_t, hi) |
0a35513e AH |
1261 | { |
1262 | size_t i; | |
355fe088 | 1263 | gimple *stmt; |
0a35513e AH |
1264 | |
1265 | if (dump_file) | |
1266 | { | |
1267 | fprintf (dump_file, "TM thread private mem logging: "); | |
ef6cb4c7 | 1268 | print_generic_expr (dump_file, lp->addr); |
0a35513e AH |
1269 | fprintf (dump_file, "\n"); |
1270 | } | |
1271 | ||
1272 | if (lp->save_var) | |
1273 | { | |
1274 | if (dump_file) | |
1275 | fprintf (dump_file, "DUMPING to variable\n"); | |
1276 | continue; | |
1277 | } | |
1278 | else | |
1279 | { | |
1280 | if (dump_file) | |
1281 | fprintf (dump_file, "DUMPING with logging functions\n"); | |
9771b263 | 1282 | for (i = 0; lp->stmts.iterate (i, &stmt); ++i) |
0a35513e AH |
1283 | tm_log_emit_stmt (lp->addr, stmt); |
1284 | } | |
1285 | } | |
1286 | } | |
1287 | ||
1288 | /* Emit the save sequence for the corresponding addresses in the log. | |
1289 | ENTRY_BLOCK is the entry block for the transaction. | |
1290 | BB is the basic block to insert the code in. */ | |
1291 | static void | |
1292 | tm_log_emit_saves (basic_block entry_block, basic_block bb) | |
1293 | { | |
1294 | size_t i; | |
1295 | gimple_stmt_iterator gsi = gsi_last_bb (bb); | |
355fe088 | 1296 | gimple *stmt; |
0a35513e AH |
1297 | struct tm_log_entry l, *lp; |
1298 | ||
9771b263 | 1299 | for (i = 0; i < tm_log_save_addresses.length (); ++i) |
0a35513e | 1300 | { |
9771b263 | 1301 | l.addr = tm_log_save_addresses[i]; |
c203e8a7 | 1302 | lp = *(tm_log->find_slot (&l, NO_INSERT)); |
0a35513e AH |
1303 | gcc_assert (lp->save_var != NULL); |
1304 | ||
1305 | /* We only care about variables in the current transaction. */ | |
1306 | if (lp->entry_block != entry_block) | |
1307 | continue; | |
1308 | ||
1309 | stmt = gimple_build_assign (lp->save_var, unshare_expr (lp->addr)); | |
1310 | ||
1311 | /* Make sure we can create an SSA_NAME for this type. For | |
1312 | instance, aggregates aren't allowed, in which case the system | |
1313 | will create a VOP for us and everything will just work. */ | |
1314 | if (is_gimple_reg_type (TREE_TYPE (lp->save_var))) | |
1315 | { | |
1316 | lp->save_var = make_ssa_name (lp->save_var, stmt); | |
1317 | gimple_assign_set_lhs (stmt, lp->save_var); | |
1318 | } | |
1319 | ||
1320 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
1321 | } | |
1322 | } | |
1323 | ||
1324 | /* Emit the restore sequence for the corresponding addresses in the log. | |
1325 | ENTRY_BLOCK is the entry block for the transaction. | |
1326 | BB is the basic block to insert the code in. */ | |
1327 | static void | |
1328 | tm_log_emit_restores (basic_block entry_block, basic_block bb) | |
1329 | { | |
1330 | int i; | |
1331 | struct tm_log_entry l, *lp; | |
1332 | gimple_stmt_iterator gsi; | |
355fe088 | 1333 | gimple *stmt; |
0a35513e | 1334 | |
9771b263 | 1335 | for (i = tm_log_save_addresses.length () - 1; i >= 0; i--) |
0a35513e | 1336 | { |
9771b263 | 1337 | l.addr = tm_log_save_addresses[i]; |
c203e8a7 | 1338 | lp = *(tm_log->find_slot (&l, NO_INSERT)); |
0a35513e AH |
1339 | gcc_assert (lp->save_var != NULL); |
1340 | ||
1341 | /* We only care about variables in the current transaction. */ | |
1342 | if (lp->entry_block != entry_block) | |
1343 | continue; | |
1344 | ||
1345 | /* Restores are in LIFO order from the saves in case we have | |
1346 | overlaps. */ | |
1347 | gsi = gsi_start_bb (bb); | |
1348 | ||
1349 | stmt = gimple_build_assign (unshare_expr (lp->addr), lp->save_var); | |
1350 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
1351 | } | |
1352 | } | |
1353 | ||
0a35513e AH |
1354 | \f |
1355 | static tree lower_sequence_tm (gimple_stmt_iterator *, bool *, | |
1356 | struct walk_stmt_info *); | |
1357 | static tree lower_sequence_no_tm (gimple_stmt_iterator *, bool *, | |
1358 | struct walk_stmt_info *); | |
1359 | ||
1360 | /* Evaluate an address X being dereferenced and determine if it | |
1361 | originally points to a non aliased new chunk of memory (malloc, | |
1362 | alloca, etc). | |
1363 | ||
1364 | Return MEM_THREAD_LOCAL if it points to a thread-local address. | |
1365 | Return MEM_TRANSACTION_LOCAL if it points to a transaction-local address. | |
1366 | Return MEM_NON_LOCAL otherwise. | |
1367 | ||
1368 | ENTRY_BLOCK is the entry block to the transaction containing the | |
1369 | dereference of X. */ | |
1370 | static enum thread_memory_type | |
1371 | thread_private_new_memory (basic_block entry_block, tree x) | |
1372 | { | |
355fe088 | 1373 | gimple *stmt = NULL; |
0a35513e | 1374 | enum tree_code code; |
526ceb68 TS |
1375 | tm_new_mem_map **slot; |
1376 | tm_new_mem_map elt, *elt_p; | |
0a35513e AH |
1377 | tree val = x; |
1378 | enum thread_memory_type retval = mem_transaction_local; | |
1379 | ||
1380 | if (!entry_block | |
1381 | || TREE_CODE (x) != SSA_NAME | |
1382 | /* Possible uninitialized use, or a function argument. In | |
1383 | either case, we don't care. */ | |
1384 | || SSA_NAME_IS_DEFAULT_DEF (x)) | |
1385 | return mem_non_local; | |
1386 | ||
1387 | /* Look in cache first. */ | |
1388 | elt.val = x; | |
c203e8a7 | 1389 | slot = tm_new_mem_hash->find_slot (&elt, INSERT); |
4a8fb1a1 | 1390 | elt_p = *slot; |
0a35513e AH |
1391 | if (elt_p) |
1392 | return elt_p->local_new_memory; | |
1393 | ||
1394 | /* Optimistically assume the memory is transaction local during | |
1395 | processing. This catches recursion into this variable. */ | |
526ceb68 | 1396 | *slot = elt_p = XNEW (tm_new_mem_map); |
0a35513e AH |
1397 | elt_p->val = val; |
1398 | elt_p->local_new_memory = mem_transaction_local; | |
1399 | ||
1400 | /* Search DEF chain to find the original definition of this address. */ | |
1401 | do | |
1402 | { | |
1403 | if (ptr_deref_may_alias_global_p (x)) | |
1404 | { | |
1405 | /* Address escapes. This is not thread-private. */ | |
1406 | retval = mem_non_local; | |
1407 | goto new_memory_ret; | |
1408 | } | |
1409 | ||
1410 | stmt = SSA_NAME_DEF_STMT (x); | |
1411 | ||
1412 | /* If the malloc call is outside the transaction, this is | |
1413 | thread-local. */ | |
1414 | if (retval != mem_thread_local | |
1415 | && !dominated_by_p (CDI_DOMINATORS, gimple_bb (stmt), entry_block)) | |
1416 | retval = mem_thread_local; | |
1417 | ||
1418 | if (is_gimple_assign (stmt)) | |
1419 | { | |
1420 | code = gimple_assign_rhs_code (stmt); | |
1421 | /* x = foo ==> foo */ | |
1422 | if (code == SSA_NAME) | |
1423 | x = gimple_assign_rhs1 (stmt); | |
1424 | /* x = foo + n ==> foo */ | |
1425 | else if (code == POINTER_PLUS_EXPR) | |
1426 | x = gimple_assign_rhs1 (stmt); | |
1427 | /* x = (cast*) foo ==> foo */ | |
625a9766 | 1428 | else if (code == VIEW_CONVERT_EXPR || CONVERT_EXPR_CODE_P (code)) |
0a35513e | 1429 | x = gimple_assign_rhs1 (stmt); |
01a723fa AP |
1430 | /* x = c ? op1 : op2 == > op1 or op2 just like a PHI */ |
1431 | else if (code == COND_EXPR) | |
1432 | { | |
1433 | tree op1 = gimple_assign_rhs2 (stmt); | |
1434 | tree op2 = gimple_assign_rhs3 (stmt); | |
1435 | enum thread_memory_type mem; | |
1436 | retval = thread_private_new_memory (entry_block, op1); | |
1437 | if (retval == mem_non_local) | |
1438 | goto new_memory_ret; | |
1439 | mem = thread_private_new_memory (entry_block, op2); | |
1440 | retval = MIN (retval, mem); | |
1441 | goto new_memory_ret; | |
1442 | } | |
0a35513e AH |
1443 | else |
1444 | { | |
1445 | retval = mem_non_local; | |
1446 | goto new_memory_ret; | |
1447 | } | |
1448 | } | |
1449 | else | |
1450 | { | |
1451 | if (gimple_code (stmt) == GIMPLE_PHI) | |
1452 | { | |
1453 | unsigned int i; | |
1454 | enum thread_memory_type mem; | |
1455 | tree phi_result = gimple_phi_result (stmt); | |
1456 | ||
1457 | /* If any of the ancestors are non-local, we are sure to | |
1458 | be non-local. Otherwise we can avoid doing anything | |
1459 | and inherit what has already been generated. */ | |
1460 | retval = mem_max; | |
1461 | for (i = 0; i < gimple_phi_num_args (stmt); ++i) | |
1462 | { | |
1463 | tree op = PHI_ARG_DEF (stmt, i); | |
1464 | ||
1465 | /* Exclude self-assignment. */ | |
1466 | if (phi_result == op) | |
1467 | continue; | |
1468 | ||
1469 | mem = thread_private_new_memory (entry_block, op); | |
1470 | if (mem == mem_non_local) | |
1471 | { | |
1472 | retval = mem; | |
1473 | goto new_memory_ret; | |
1474 | } | |
1475 | retval = MIN (retval, mem); | |
1476 | } | |
1477 | goto new_memory_ret; | |
1478 | } | |
1479 | break; | |
1480 | } | |
1481 | } | |
1482 | while (TREE_CODE (x) == SSA_NAME); | |
1483 | ||
1484 | if (stmt && is_gimple_call (stmt) && gimple_call_flags (stmt) & ECF_MALLOC) | |
1485 | /* Thread-local or transaction-local. */ | |
1486 | ; | |
1487 | else | |
1488 | retval = mem_non_local; | |
1489 | ||
1490 | new_memory_ret: | |
1491 | elt_p->local_new_memory = retval; | |
1492 | return retval; | |
1493 | } | |
1494 | ||
1495 | /* Determine whether X has to be instrumented using a read | |
1496 | or write barrier. | |
1497 | ||
1498 | ENTRY_BLOCK is the entry block for the region where stmt resides | |
1499 | in. NULL if unknown. | |
1500 | ||
1501 | STMT is the statement in which X occurs in. It is used for thread | |
1502 | private memory instrumentation. If no TPM instrumentation is | |
1503 | desired, STMT should be null. */ | |
1504 | static bool | |
355fe088 | 1505 | requires_barrier (basic_block entry_block, tree x, gimple *stmt) |
0a35513e AH |
1506 | { |
1507 | tree orig = x; | |
1508 | while (handled_component_p (x)) | |
1509 | x = TREE_OPERAND (x, 0); | |
1510 | ||
1511 | switch (TREE_CODE (x)) | |
1512 | { | |
1513 | case INDIRECT_REF: | |
1514 | case MEM_REF: | |
1515 | { | |
1516 | enum thread_memory_type ret; | |
1517 | ||
1518 | ret = thread_private_new_memory (entry_block, TREE_OPERAND (x, 0)); | |
1519 | if (ret == mem_non_local) | |
1520 | return true; | |
1521 | if (stmt && ret == mem_thread_local) | |
1522 | /* ?? Should we pass `orig', or the INDIRECT_REF X. ?? */ | |
1523 | tm_log_add (entry_block, orig, stmt); | |
1524 | ||
1525 | /* Transaction-locals require nothing at all. For malloc, a | |
1526 | transaction restart frees the memory and we reallocate. | |
1527 | For alloca, the stack pointer gets reset by the retry and | |
1528 | we reallocate. */ | |
1529 | return false; | |
1530 | } | |
1531 | ||
1532 | case TARGET_MEM_REF: | |
1533 | if (TREE_CODE (TMR_BASE (x)) != ADDR_EXPR) | |
1534 | return true; | |
1535 | x = TREE_OPERAND (TMR_BASE (x), 0); | |
1536 | if (TREE_CODE (x) == PARM_DECL) | |
1537 | return false; | |
8813a647 | 1538 | gcc_assert (VAR_P (x)); |
0a35513e AH |
1539 | /* FALLTHRU */ |
1540 | ||
1541 | case PARM_DECL: | |
1542 | case RESULT_DECL: | |
1543 | case VAR_DECL: | |
1544 | if (DECL_BY_REFERENCE (x)) | |
1545 | { | |
1546 | /* ??? This value is a pointer, but aggregate_value_p has been | |
1547 | jigged to return true which confuses needs_to_live_in_memory. | |
1548 | This ought to be cleaned up generically. | |
1549 | ||
1550 | FIXME: Verify this still happens after the next mainline | |
1551 | merge. Testcase ie g++.dg/tm/pr47554.C. | |
1552 | */ | |
1553 | return false; | |
1554 | } | |
1555 | ||
1556 | if (is_global_var (x)) | |
3be18e47 | 1557 | return !TREE_READONLY (x); |
0a35513e AH |
1558 | if (/* FIXME: This condition should actually go below in the |
1559 | tm_log_add() call, however is_call_clobbered() depends on | |
1560 | aliasing info which is not available during | |
1561 | gimplification. Since requires_barrier() gets called | |
1562 | during lower_sequence_tm/gimplification, leave the call | |
1563 | to needs_to_live_in_memory until we eliminate | |
1564 | lower_sequence_tm altogether. */ | |
63c0efdb | 1565 | needs_to_live_in_memory (x)) |
0a35513e | 1566 | return true; |
3be18e47 RH |
1567 | else |
1568 | { | |
1569 | /* For local memory that doesn't escape (aka thread private | |
1570 | memory), we can either save the value at the beginning of | |
1571 | the transaction and restore on restart, or call a tm | |
1572 | function to dynamically save and restore on restart | |
1573 | (ITM_L*). */ | |
1574 | if (stmt) | |
1575 | tm_log_add (entry_block, orig, stmt); | |
1576 | return false; | |
1577 | } | |
0a35513e AH |
1578 | |
1579 | default: | |
1580 | return false; | |
1581 | } | |
1582 | } | |
1583 | ||
1584 | /* Mark the GIMPLE_ASSIGN statement as appropriate for being inside | |
1585 | a transaction region. */ | |
1586 | ||
1587 | static void | |
1588 | examine_assign_tm (unsigned *state, gimple_stmt_iterator *gsi) | |
1589 | { | |
355fe088 | 1590 | gimple *stmt = gsi_stmt (*gsi); |
0a35513e AH |
1591 | |
1592 | if (requires_barrier (/*entry_block=*/NULL, gimple_assign_rhs1 (stmt), NULL)) | |
1593 | *state |= GTMA_HAVE_LOAD; | |
1594 | if (requires_barrier (/*entry_block=*/NULL, gimple_assign_lhs (stmt), NULL)) | |
1595 | *state |= GTMA_HAVE_STORE; | |
1596 | } | |
1597 | ||
1598 | /* Mark a GIMPLE_CALL as appropriate for being inside a transaction. */ | |
1599 | ||
1600 | static void | |
1601 | examine_call_tm (unsigned *state, gimple_stmt_iterator *gsi) | |
1602 | { | |
355fe088 | 1603 | gimple *stmt = gsi_stmt (*gsi); |
0a35513e AH |
1604 | tree fn; |
1605 | ||
1606 | if (is_tm_pure_call (stmt)) | |
1607 | return; | |
1608 | ||
1609 | /* Check if this call is a transaction abort. */ | |
1610 | fn = gimple_call_fndecl (stmt); | |
1611 | if (is_tm_abort (fn)) | |
1612 | *state |= GTMA_HAVE_ABORT; | |
1613 | ||
1614 | /* Note that something may happen. */ | |
1615 | *state |= GTMA_HAVE_LOAD | GTMA_HAVE_STORE; | |
1616 | } | |
1617 | ||
7c11b0fe RH |
1618 | /* Iterate through the statements in the sequence, moving labels |
1619 | (and thus edges) of transactions from "label_norm" to "label_uninst". */ | |
1620 | ||
1621 | static tree | |
1622 | make_tm_uninst (gimple_stmt_iterator *gsi, bool *handled_ops_p, | |
1623 | struct walk_stmt_info *) | |
1624 | { | |
1625 | gimple *stmt = gsi_stmt (*gsi); | |
1626 | ||
1627 | if (gtransaction *txn = dyn_cast <gtransaction *> (stmt)) | |
1628 | { | |
1629 | *handled_ops_p = true; | |
1630 | txn->label_uninst = txn->label_norm; | |
1631 | txn->label_norm = NULL; | |
1632 | } | |
1633 | else | |
1634 | *handled_ops_p = !gimple_has_substatements (stmt); | |
1635 | ||
1636 | return NULL_TREE; | |
1637 | } | |
1638 | ||
0a35513e AH |
1639 | /* Lower a GIMPLE_TRANSACTION statement. */ |
1640 | ||
1641 | static void | |
1642 | lower_transaction (gimple_stmt_iterator *gsi, struct walk_stmt_info *wi) | |
1643 | { | |
355fe088 | 1644 | gimple *g; |
538dd0b7 | 1645 | gtransaction *stmt = as_a <gtransaction *> (gsi_stmt (*gsi)); |
0a35513e AH |
1646 | unsigned int *outer_state = (unsigned int *) wi->info; |
1647 | unsigned int this_state = 0; | |
1648 | struct walk_stmt_info this_wi; | |
1649 | ||
1650 | /* First, lower the body. The scanning that we do inside gives | |
1651 | us some idea of what we're dealing with. */ | |
1652 | memset (&this_wi, 0, sizeof (this_wi)); | |
1653 | this_wi.info = (void *) &this_state; | |
355a7673 MM |
1654 | walk_gimple_seq_mod (gimple_transaction_body_ptr (stmt), |
1655 | lower_sequence_tm, NULL, &this_wi); | |
0a35513e AH |
1656 | |
1657 | /* If there was absolutely nothing transaction related inside the | |
1658 | transaction, we may elide it. Likewise if this is a nested | |
1659 | transaction and does not contain an abort. */ | |
1660 | if (this_state == 0 | |
1661 | || (!(this_state & GTMA_HAVE_ABORT) && outer_state != NULL)) | |
1662 | { | |
1663 | if (outer_state) | |
1664 | *outer_state |= this_state; | |
1665 | ||
1666 | gsi_insert_seq_before (gsi, gimple_transaction_body (stmt), | |
1667 | GSI_SAME_STMT); | |
1668 | gimple_transaction_set_body (stmt, NULL); | |
1669 | ||
1670 | gsi_remove (gsi, true); | |
1671 | wi->removed_stmt = true; | |
1672 | return; | |
1673 | } | |
1674 | ||
1675 | /* Wrap the body of the transaction in a try-finally node so that | |
1676 | the commit call is always properly called. */ | |
1677 | g = gimple_build_call (builtin_decl_explicit (BUILT_IN_TM_COMMIT), 0); | |
1678 | if (flag_exceptions) | |
1679 | { | |
1680 | tree ptr; | |
1681 | gimple_seq n_seq, e_seq; | |
1682 | ||
1683 | n_seq = gimple_seq_alloc_with_stmt (g); | |
355a7673 | 1684 | e_seq = NULL; |
0a35513e AH |
1685 | |
1686 | g = gimple_build_call (builtin_decl_explicit (BUILT_IN_EH_POINTER), | |
1687 | 1, integer_zero_node); | |
b731b390 | 1688 | ptr = create_tmp_var (ptr_type_node); |
0a35513e AH |
1689 | gimple_call_set_lhs (g, ptr); |
1690 | gimple_seq_add_stmt (&e_seq, g); | |
1691 | ||
1692 | g = gimple_build_call (builtin_decl_explicit (BUILT_IN_TM_COMMIT_EH), | |
1693 | 1, ptr); | |
1694 | gimple_seq_add_stmt (&e_seq, g); | |
1695 | ||
1696 | g = gimple_build_eh_else (n_seq, e_seq); | |
1697 | } | |
1698 | ||
1699 | g = gimple_build_try (gimple_transaction_body (stmt), | |
1700 | gimple_seq_alloc_with_stmt (g), GIMPLE_TRY_FINALLY); | |
0a35513e | 1701 | |
7c11b0fe RH |
1702 | /* For a (potentially) outer transaction, create two paths. */ |
1703 | gimple_seq uninst = NULL; | |
1704 | if (outer_state == NULL) | |
1705 | { | |
1706 | uninst = copy_gimple_seq_and_replace_locals (g); | |
1707 | /* In the uninstrumented copy, reset inner transactions to have only | |
1708 | an uninstrumented code path. */ | |
1709 | memset (&this_wi, 0, sizeof (this_wi)); | |
1710 | walk_gimple_seq (uninst, make_tm_uninst, NULL, &this_wi); | |
1711 | } | |
1712 | ||
1713 | tree label1 = create_artificial_label (UNKNOWN_LOCATION); | |
1714 | gsi_insert_after (gsi, gimple_build_label (label1), GSI_CONTINUE_LINKING); | |
1715 | gsi_insert_after (gsi, g, GSI_CONTINUE_LINKING); | |
1716 | gimple_transaction_set_label_norm (stmt, label1); | |
0a35513e AH |
1717 | |
1718 | /* If the transaction calls abort or if this is an outer transaction, | |
1719 | add an "over" label afterwards. */ | |
7c11b0fe RH |
1720 | tree label3 = NULL; |
1721 | if ((this_state & GTMA_HAVE_ABORT) | |
1722 | || outer_state == NULL | |
c3284718 | 1723 | || (gimple_transaction_subcode (stmt) & GTMA_IS_OUTER)) |
0a35513e | 1724 | { |
7c11b0fe RH |
1725 | label3 = create_artificial_label (UNKNOWN_LOCATION); |
1726 | gimple_transaction_set_label_over (stmt, label3); | |
0a35513e AH |
1727 | } |
1728 | ||
7c11b0fe RH |
1729 | if (uninst != NULL) |
1730 | { | |
1731 | gsi_insert_after (gsi, gimple_build_goto (label3), GSI_CONTINUE_LINKING); | |
1732 | ||
1733 | tree label2 = create_artificial_label (UNKNOWN_LOCATION); | |
1734 | gsi_insert_after (gsi, gimple_build_label (label2), GSI_CONTINUE_LINKING); | |
1735 | gsi_insert_seq_after (gsi, uninst, GSI_CONTINUE_LINKING); | |
1736 | gimple_transaction_set_label_uninst (stmt, label2); | |
1737 | } | |
1738 | ||
1739 | if (label3 != NULL) | |
1740 | gsi_insert_after (gsi, gimple_build_label (label3), GSI_CONTINUE_LINKING); | |
1741 | ||
1742 | gimple_transaction_set_body (stmt, NULL); | |
1743 | ||
0a35513e AH |
1744 | /* Record the set of operations found for use later. */ |
1745 | this_state |= gimple_transaction_subcode (stmt) & GTMA_DECLARATION_MASK; | |
1746 | gimple_transaction_set_subcode (stmt, this_state); | |
1747 | } | |
1748 | ||
1749 | /* Iterate through the statements in the sequence, lowering them all | |
1750 | as appropriate for being in a transaction. */ | |
1751 | ||
1752 | static tree | |
1753 | lower_sequence_tm (gimple_stmt_iterator *gsi, bool *handled_ops_p, | |
1754 | struct walk_stmt_info *wi) | |
1755 | { | |
1756 | unsigned int *state = (unsigned int *) wi->info; | |
355fe088 | 1757 | gimple *stmt = gsi_stmt (*gsi); |
0a35513e AH |
1758 | |
1759 | *handled_ops_p = true; | |
1760 | switch (gimple_code (stmt)) | |
1761 | { | |
1762 | case GIMPLE_ASSIGN: | |
1763 | /* Only memory reads/writes need to be instrumented. */ | |
1764 | if (gimple_assign_single_p (stmt)) | |
1765 | examine_assign_tm (state, gsi); | |
1766 | break; | |
1767 | ||
1768 | case GIMPLE_CALL: | |
1769 | examine_call_tm (state, gsi); | |
1770 | break; | |
1771 | ||
1772 | case GIMPLE_ASM: | |
1773 | *state |= GTMA_MAY_ENTER_IRREVOCABLE; | |
1774 | break; | |
1775 | ||
1776 | case GIMPLE_TRANSACTION: | |
1777 | lower_transaction (gsi, wi); | |
1778 | break; | |
1779 | ||
1780 | default: | |
1781 | *handled_ops_p = !gimple_has_substatements (stmt); | |
1782 | break; | |
1783 | } | |
1784 | ||
1785 | return NULL_TREE; | |
1786 | } | |
1787 | ||
1788 | /* Iterate through the statements in the sequence, lowering them all | |
1789 | as appropriate for being outside of a transaction. */ | |
1790 | ||
1791 | static tree | |
1792 | lower_sequence_no_tm (gimple_stmt_iterator *gsi, bool *handled_ops_p, | |
1793 | struct walk_stmt_info * wi) | |
1794 | { | |
355fe088 | 1795 | gimple *stmt = gsi_stmt (*gsi); |
0a35513e AH |
1796 | |
1797 | if (gimple_code (stmt) == GIMPLE_TRANSACTION) | |
1798 | { | |
1799 | *handled_ops_p = true; | |
1800 | lower_transaction (gsi, wi); | |
1801 | } | |
1802 | else | |
1803 | *handled_ops_p = !gimple_has_substatements (stmt); | |
1804 | ||
1805 | return NULL_TREE; | |
1806 | } | |
1807 | ||
1808 | /* Main entry point for flattening GIMPLE_TRANSACTION constructs. After | |
1809 | this, GIMPLE_TRANSACTION nodes still exist, but the nested body has | |
1810 | been moved out, and all the data required for constructing a proper | |
1811 | CFG has been recorded. */ | |
1812 | ||
1813 | static unsigned int | |
1814 | execute_lower_tm (void) | |
1815 | { | |
1816 | struct walk_stmt_info wi; | |
355a7673 | 1817 | gimple_seq body; |
0a35513e AH |
1818 | |
1819 | /* Transactional clones aren't created until a later pass. */ | |
1820 | gcc_assert (!decl_is_tm_clone (current_function_decl)); | |
1821 | ||
355a7673 | 1822 | body = gimple_body (current_function_decl); |
0a35513e | 1823 | memset (&wi, 0, sizeof (wi)); |
355a7673 MM |
1824 | walk_gimple_seq_mod (&body, lower_sequence_no_tm, NULL, &wi); |
1825 | gimple_set_body (current_function_decl, body); | |
0a35513e AH |
1826 | |
1827 | return 0; | |
1828 | } | |
1829 | ||
17795822 TS |
1830 | namespace { |
1831 | ||
1832 | const pass_data pass_data_lower_tm = | |
27a4cd48 DM |
1833 | { |
1834 | GIMPLE_PASS, /* type */ | |
1835 | "tmlower", /* name */ | |
1836 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
1837 | TV_TRANS_MEM, /* tv_id */ |
1838 | PROP_gimple_lcf, /* properties_required */ | |
1839 | 0, /* properties_provided */ | |
1840 | 0, /* properties_destroyed */ | |
1841 | 0, /* todo_flags_start */ | |
1842 | 0, /* todo_flags_finish */ | |
0a35513e | 1843 | }; |
27a4cd48 | 1844 | |
17795822 | 1845 | class pass_lower_tm : public gimple_opt_pass |
27a4cd48 DM |
1846 | { |
1847 | public: | |
c3284718 RS |
1848 | pass_lower_tm (gcc::context *ctxt) |
1849 | : gimple_opt_pass (pass_data_lower_tm, ctxt) | |
27a4cd48 DM |
1850 | {} |
1851 | ||
1852 | /* opt_pass methods: */ | |
1a3d085c | 1853 | virtual bool gate (function *) { return flag_tm; } |
be55bfe6 | 1854 | virtual unsigned int execute (function *) { return execute_lower_tm (); } |
27a4cd48 DM |
1855 | |
1856 | }; // class pass_lower_tm | |
1857 | ||
17795822 TS |
1858 | } // anon namespace |
1859 | ||
27a4cd48 DM |
1860 | gimple_opt_pass * |
1861 | make_pass_lower_tm (gcc::context *ctxt) | |
1862 | { | |
1863 | return new pass_lower_tm (ctxt); | |
1864 | } | |
0a35513e AH |
1865 | \f |
1866 | /* Collect region information for each transaction. */ | |
1867 | ||
1868 | struct tm_region | |
1869 | { | |
538dd0b7 DM |
1870 | public: |
1871 | ||
1872 | /* The field "transaction_stmt" is initially a gtransaction *, | |
1873 | but eventually gets lowered to a gcall *(to BUILT_IN_TM_START). | |
1874 | ||
1875 | Helper method to get it as a gtransaction *, with code-checking | |
1876 | in a checked-build. */ | |
1877 | ||
1878 | gtransaction * | |
1879 | get_transaction_stmt () const | |
1880 | { | |
1881 | return as_a <gtransaction *> (transaction_stmt); | |
1882 | } | |
1883 | ||
1884 | public: | |
1885 | ||
0a35513e AH |
1886 | /* Link to the next unnested transaction. */ |
1887 | struct tm_region *next; | |
1888 | ||
1889 | /* Link to the next inner transaction. */ | |
1890 | struct tm_region *inner; | |
1891 | ||
1892 | /* Link to the next outer transaction. */ | |
1893 | struct tm_region *outer; | |
1894 | ||
398b1daa AH |
1895 | /* The GIMPLE_TRANSACTION statement beginning this transaction. |
1896 | After TM_MARK, this gets replaced by a call to | |
538dd0b7 DM |
1897 | BUILT_IN_TM_START. |
1898 | Hence this will be either a gtransaction *or a gcall *. */ | |
355fe088 | 1899 | gimple *transaction_stmt; |
0a35513e | 1900 | |
398b1daa AH |
1901 | /* After TM_MARK expands the GIMPLE_TRANSACTION into a call to |
1902 | BUILT_IN_TM_START, this field is true if the transaction is an | |
1903 | outer transaction. */ | |
1904 | bool original_transaction_was_outer; | |
1905 | ||
1906 | /* Return value from BUILT_IN_TM_START. */ | |
1907 | tree tm_state; | |
1908 | ||
1909 | /* The entry block to this region. This will always be the first | |
1910 | block of the body of the transaction. */ | |
0a35513e AH |
1911 | basic_block entry_block; |
1912 | ||
398b1daa AH |
1913 | /* The first block after an expanded call to _ITM_beginTransaction. */ |
1914 | basic_block restart_block; | |
1915 | ||
0a35513e AH |
1916 | /* The set of all blocks that end the region; NULL if only EXIT_BLOCK. |
1917 | These blocks are still a part of the region (i.e., the border is | |
1918 | inclusive). Note that this set is only complete for paths in the CFG | |
1919 | starting at ENTRY_BLOCK, and that there is no exit block recorded for | |
1920 | the edge to the "over" label. */ | |
1921 | bitmap exit_blocks; | |
1922 | ||
1923 | /* The set of all blocks that have an TM_IRREVOCABLE call. */ | |
1924 | bitmap irr_blocks; | |
1925 | }; | |
1926 | ||
1927 | /* True if there are pending edge statements to be committed for the | |
1928 | current function being scanned in the tmmark pass. */ | |
1929 | bool pending_edge_inserts_p; | |
1930 | ||
1931 | static struct tm_region *all_tm_regions; | |
1932 | static bitmap_obstack tm_obstack; | |
1933 | ||
1934 | ||
073a8998 | 1935 | /* A subroutine of tm_region_init. Record the existence of the |
0a35513e AH |
1936 | GIMPLE_TRANSACTION statement in a tree of tm_region elements. */ |
1937 | ||
1938 | static struct tm_region * | |
538dd0b7 DM |
1939 | tm_region_init_0 (struct tm_region *outer, basic_block bb, |
1940 | gtransaction *stmt) | |
0a35513e AH |
1941 | { |
1942 | struct tm_region *region; | |
1943 | ||
1944 | region = (struct tm_region *) | |
1945 | obstack_alloc (&tm_obstack.obstack, sizeof (struct tm_region)); | |
1946 | ||
1947 | if (outer) | |
1948 | { | |
1949 | region->next = outer->inner; | |
1950 | outer->inner = region; | |
1951 | } | |
1952 | else | |
1953 | { | |
1954 | region->next = all_tm_regions; | |
1955 | all_tm_regions = region; | |
1956 | } | |
1957 | region->inner = NULL; | |
1958 | region->outer = outer; | |
1959 | ||
1960 | region->transaction_stmt = stmt; | |
398b1daa AH |
1961 | region->original_transaction_was_outer = false; |
1962 | region->tm_state = NULL; | |
0a35513e AH |
1963 | |
1964 | /* There are either one or two edges out of the block containing | |
1965 | the GIMPLE_TRANSACTION, one to the actual region and one to the | |
1966 | "over" label if the region contains an abort. The former will | |
1967 | always be the one marked FALLTHRU. */ | |
1968 | region->entry_block = FALLTHRU_EDGE (bb)->dest; | |
1969 | ||
1970 | region->exit_blocks = BITMAP_ALLOC (&tm_obstack); | |
1971 | region->irr_blocks = BITMAP_ALLOC (&tm_obstack); | |
1972 | ||
1973 | return region; | |
1974 | } | |
1975 | ||
1976 | /* A subroutine of tm_region_init. Record all the exit and | |
1977 | irrevocable blocks in BB into the region's exit_blocks and | |
1978 | irr_blocks bitmaps. Returns the new region being scanned. */ | |
1979 | ||
1980 | static struct tm_region * | |
1981 | tm_region_init_1 (struct tm_region *region, basic_block bb) | |
1982 | { | |
1983 | gimple_stmt_iterator gsi; | |
355fe088 | 1984 | gimple *g; |
0a35513e AH |
1985 | |
1986 | if (!region | |
1987 | || (!region->irr_blocks && !region->exit_blocks)) | |
1988 | return region; | |
1989 | ||
1990 | /* Check to see if this is the end of a region by seeing if it | |
1991 | contains a call to __builtin_tm_commit{,_eh}. Note that the | |
1992 | outermost region for DECL_IS_TM_CLONE need not collect this. */ | |
1993 | for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi)) | |
1994 | { | |
1995 | g = gsi_stmt (gsi); | |
1996 | if (gimple_code (g) == GIMPLE_CALL) | |
1997 | { | |
1998 | tree fn = gimple_call_fndecl (g); | |
3d78e008 | 1999 | if (fn && fndecl_built_in_p (fn, BUILT_IN_NORMAL)) |
0a35513e AH |
2000 | { |
2001 | if ((DECL_FUNCTION_CODE (fn) == BUILT_IN_TM_COMMIT | |
2002 | || DECL_FUNCTION_CODE (fn) == BUILT_IN_TM_COMMIT_EH) | |
2003 | && region->exit_blocks) | |
2004 | { | |
2005 | bitmap_set_bit (region->exit_blocks, bb->index); | |
2006 | region = region->outer; | |
2007 | break; | |
2008 | } | |
2009 | if (DECL_FUNCTION_CODE (fn) == BUILT_IN_TM_IRREVOCABLE) | |
2010 | bitmap_set_bit (region->irr_blocks, bb->index); | |
2011 | } | |
2012 | } | |
2013 | } | |
2014 | return region; | |
2015 | } | |
2016 | ||
2017 | /* Collect all of the transaction regions within the current function | |
2018 | and record them in ALL_TM_REGIONS. The REGION parameter may specify | |
2019 | an "outermost" region for use by tm clones. */ | |
2020 | ||
2021 | static void | |
2022 | tm_region_init (struct tm_region *region) | |
2023 | { | |
355fe088 | 2024 | gimple *g; |
0a35513e AH |
2025 | edge_iterator ei; |
2026 | edge e; | |
2027 | basic_block bb; | |
ef062b13 | 2028 | auto_vec<basic_block> queue; |
0a35513e AH |
2029 | bitmap visited_blocks = BITMAP_ALLOC (NULL); |
2030 | struct tm_region *old_region; | |
526ceb68 | 2031 | auto_vec<tm_region *> bb_regions; |
0a35513e | 2032 | |
19c0d7df AH |
2033 | /* We could store this information in bb->aux, but we may get called |
2034 | through get_all_tm_blocks() from another pass that may be already | |
2035 | using bb->aux. */ | |
cb3874dc | 2036 | bb_regions.safe_grow_cleared (last_basic_block_for_fn (cfun), true); |
19c0d7df | 2037 | |
ffc445f4 RH |
2038 | all_tm_regions = region; |
2039 | bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); | |
9771b263 | 2040 | queue.safe_push (bb); |
ffc445f4 | 2041 | bitmap_set_bit (visited_blocks, bb->index); |
9771b263 | 2042 | bb_regions[bb->index] = region; |
ffc445f4 | 2043 | |
0a35513e AH |
2044 | do |
2045 | { | |
9771b263 DN |
2046 | bb = queue.pop (); |
2047 | region = bb_regions[bb->index]; | |
2048 | bb_regions[bb->index] = NULL; | |
0a35513e AH |
2049 | |
2050 | /* Record exit and irrevocable blocks. */ | |
2051 | region = tm_region_init_1 (region, bb); | |
2052 | ||
2053 | /* Check for the last statement in the block beginning a new region. */ | |
2054 | g = last_stmt (bb); | |
2055 | old_region = region; | |
538dd0b7 DM |
2056 | if (g) |
2057 | if (gtransaction *trans_stmt = dyn_cast <gtransaction *> (g)) | |
2058 | region = tm_region_init_0 (region, bb, trans_stmt); | |
0a35513e AH |
2059 | |
2060 | /* Process subsequent blocks. */ | |
2061 | FOR_EACH_EDGE (e, ei, bb->succs) | |
2062 | if (!bitmap_bit_p (visited_blocks, e->dest->index)) | |
2063 | { | |
2064 | bitmap_set_bit (visited_blocks, e->dest->index); | |
9771b263 | 2065 | queue.safe_push (e->dest); |
0a35513e AH |
2066 | |
2067 | /* If the current block started a new region, make sure that only | |
2068 | the entry block of the new region is associated with this region. | |
2069 | Other successors are still part of the old region. */ | |
2070 | if (old_region != region && e->dest != region->entry_block) | |
9771b263 | 2071 | bb_regions[e->dest->index] = old_region; |
0a35513e | 2072 | else |
9771b263 | 2073 | bb_regions[e->dest->index] = region; |
0a35513e AH |
2074 | } |
2075 | } | |
9771b263 | 2076 | while (!queue.is_empty ()); |
0a35513e AH |
2077 | BITMAP_FREE (visited_blocks); |
2078 | } | |
2079 | ||
2080 | /* The "gate" function for all transactional memory expansion and optimization | |
2081 | passes. We collect region information for each top-level transaction, and | |
2082 | if we don't find any, we skip all of the TM passes. Each region will have | |
2083 | all of the exit blocks recorded, and the originating statement. */ | |
2084 | ||
2085 | static bool | |
2086 | gate_tm_init (void) | |
2087 | { | |
2088 | if (!flag_tm) | |
2089 | return false; | |
2090 | ||
2091 | calculate_dominance_info (CDI_DOMINATORS); | |
2092 | bitmap_obstack_initialize (&tm_obstack); | |
2093 | ||
2094 | /* If the function is a TM_CLONE, then the entire function is the region. */ | |
2095 | if (decl_is_tm_clone (current_function_decl)) | |
2096 | { | |
2097 | struct tm_region *region = (struct tm_region *) | |
2098 | obstack_alloc (&tm_obstack.obstack, sizeof (struct tm_region)); | |
2099 | memset (region, 0, sizeof (*region)); | |
fefa31b5 | 2100 | region->entry_block = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
0a35513e AH |
2101 | /* For a clone, the entire function is the region. But even if |
2102 | we don't need to record any exit blocks, we may need to | |
2103 | record irrevocable blocks. */ | |
2104 | region->irr_blocks = BITMAP_ALLOC (&tm_obstack); | |
2105 | ||
2106 | tm_region_init (region); | |
2107 | } | |
2108 | else | |
2109 | { | |
2110 | tm_region_init (NULL); | |
2111 | ||
2112 | /* If we didn't find any regions, cleanup and skip the whole tree | |
2113 | of tm-related optimizations. */ | |
2114 | if (all_tm_regions == NULL) | |
2115 | { | |
2116 | bitmap_obstack_release (&tm_obstack); | |
2117 | return false; | |
2118 | } | |
2119 | } | |
2120 | ||
2121 | return true; | |
2122 | } | |
2123 | ||
17795822 TS |
2124 | namespace { |
2125 | ||
2126 | const pass_data pass_data_tm_init = | |
27a4cd48 DM |
2127 | { |
2128 | GIMPLE_PASS, /* type */ | |
2129 | "*tminit", /* name */ | |
2130 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
2131 | TV_TRANS_MEM, /* tv_id */ |
2132 | ( PROP_ssa | PROP_cfg ), /* properties_required */ | |
2133 | 0, /* properties_provided */ | |
2134 | 0, /* properties_destroyed */ | |
2135 | 0, /* todo_flags_start */ | |
2136 | 0, /* todo_flags_finish */ | |
0a35513e | 2137 | }; |
27a4cd48 | 2138 | |
17795822 | 2139 | class pass_tm_init : public gimple_opt_pass |
27a4cd48 DM |
2140 | { |
2141 | public: | |
c3284718 RS |
2142 | pass_tm_init (gcc::context *ctxt) |
2143 | : gimple_opt_pass (pass_data_tm_init, ctxt) | |
27a4cd48 DM |
2144 | {} |
2145 | ||
2146 | /* opt_pass methods: */ | |
1a3d085c | 2147 | virtual bool gate (function *) { return gate_tm_init (); } |
27a4cd48 DM |
2148 | |
2149 | }; // class pass_tm_init | |
2150 | ||
17795822 TS |
2151 | } // anon namespace |
2152 | ||
27a4cd48 DM |
2153 | gimple_opt_pass * |
2154 | make_pass_tm_init (gcc::context *ctxt) | |
2155 | { | |
2156 | return new pass_tm_init (ctxt); | |
2157 | } | |
0a35513e AH |
2158 | \f |
2159 | /* Add FLAGS to the GIMPLE_TRANSACTION subcode for the transaction region | |
2160 | represented by STATE. */ | |
2161 | ||
2162 | static inline void | |
2163 | transaction_subcode_ior (struct tm_region *region, unsigned flags) | |
2164 | { | |
2165 | if (region && region->transaction_stmt) | |
2166 | { | |
538dd0b7 DM |
2167 | gtransaction *transaction_stmt = region->get_transaction_stmt (); |
2168 | flags |= gimple_transaction_subcode (transaction_stmt); | |
2169 | gimple_transaction_set_subcode (transaction_stmt, flags); | |
0a35513e AH |
2170 | } |
2171 | } | |
2172 | ||
2173 | /* Construct a memory load in a transactional context. Return the | |
2174 | gimple statement performing the load, or NULL if there is no | |
2175 | TM_LOAD builtin of the appropriate size to do the load. | |
2176 | ||
2177 | LOC is the location to use for the new statement(s). */ | |
2178 | ||
538dd0b7 | 2179 | static gcall * |
0a35513e AH |
2180 | build_tm_load (location_t loc, tree lhs, tree rhs, gimple_stmt_iterator *gsi) |
2181 | { | |
4c868789 | 2182 | tree t, type = TREE_TYPE (rhs); |
538dd0b7 | 2183 | gcall *gcall; |
0a35513e | 2184 | |
4c868789 | 2185 | built_in_function code; |
0a35513e AH |
2186 | if (type == float_type_node) |
2187 | code = BUILT_IN_TM_LOAD_FLOAT; | |
2188 | else if (type == double_type_node) | |
2189 | code = BUILT_IN_TM_LOAD_DOUBLE; | |
2190 | else if (type == long_double_type_node) | |
2191 | code = BUILT_IN_TM_LOAD_LDOUBLE; | |
4c868789 | 2192 | else |
0a35513e | 2193 | { |
4c868789 RH |
2194 | if (TYPE_SIZE (type) == NULL || !tree_fits_uhwi_p (TYPE_SIZE (type))) |
2195 | return NULL; | |
2196 | unsigned HOST_WIDE_INT type_size = tree_to_uhwi (TYPE_SIZE (type)); | |
2197 | ||
2198 | if (TREE_CODE (type) == VECTOR_TYPE) | |
0a35513e | 2199 | { |
4c868789 RH |
2200 | switch (type_size) |
2201 | { | |
2202 | case 64: | |
2203 | code = BUILT_IN_TM_LOAD_M64; | |
2204 | break; | |
2205 | case 128: | |
2206 | code = BUILT_IN_TM_LOAD_M128; | |
2207 | break; | |
2208 | case 256: | |
2209 | code = BUILT_IN_TM_LOAD_M256; | |
2210 | break; | |
2211 | default: | |
2212 | goto unhandled_vec; | |
2213 | } | |
2214 | if (!builtin_decl_explicit_p (code)) | |
2215 | goto unhandled_vec; | |
2216 | } | |
2217 | else | |
2218 | { | |
2219 | unhandled_vec: | |
2220 | switch (type_size) | |
2221 | { | |
2222 | case 8: | |
2223 | code = BUILT_IN_TM_LOAD_1; | |
2224 | break; | |
2225 | case 16: | |
2226 | code = BUILT_IN_TM_LOAD_2; | |
2227 | break; | |
2228 | case 32: | |
2229 | code = BUILT_IN_TM_LOAD_4; | |
2230 | break; | |
2231 | case 64: | |
2232 | code = BUILT_IN_TM_LOAD_8; | |
2233 | break; | |
2234 | default: | |
2235 | return NULL; | |
2236 | } | |
0a35513e AH |
2237 | } |
2238 | } | |
2239 | ||
4c868789 RH |
2240 | tree decl = builtin_decl_explicit (code); |
2241 | gcc_assert (decl); | |
0a35513e AH |
2242 | |
2243 | t = gimplify_addr (gsi, rhs); | |
2244 | gcall = gimple_build_call (decl, 1, t); | |
2245 | gimple_set_location (gcall, loc); | |
2246 | ||
2247 | t = TREE_TYPE (TREE_TYPE (decl)); | |
2248 | if (useless_type_conversion_p (type, t)) | |
2249 | { | |
2250 | gimple_call_set_lhs (gcall, lhs); | |
2251 | gsi_insert_before (gsi, gcall, GSI_SAME_STMT); | |
2252 | } | |
2253 | else | |
2254 | { | |
355fe088 | 2255 | gimple *g; |
0a35513e AH |
2256 | tree temp; |
2257 | ||
b731b390 | 2258 | temp = create_tmp_reg (t); |
0a35513e AH |
2259 | gimple_call_set_lhs (gcall, temp); |
2260 | gsi_insert_before (gsi, gcall, GSI_SAME_STMT); | |
2261 | ||
2262 | t = fold_build1 (VIEW_CONVERT_EXPR, type, temp); | |
2263 | g = gimple_build_assign (lhs, t); | |
2264 | gsi_insert_before (gsi, g, GSI_SAME_STMT); | |
2265 | } | |
2266 | ||
2267 | return gcall; | |
2268 | } | |
2269 | ||
2270 | ||
2271 | /* Similarly for storing TYPE in a transactional context. */ | |
2272 | ||
538dd0b7 | 2273 | static gcall * |
0a35513e AH |
2274 | build_tm_store (location_t loc, tree lhs, tree rhs, gimple_stmt_iterator *gsi) |
2275 | { | |
0a35513e | 2276 | tree t, fn, type = TREE_TYPE (rhs), simple_type; |
538dd0b7 | 2277 | gcall *gcall; |
0a35513e | 2278 | |
4c868789 | 2279 | built_in_function code; |
0a35513e AH |
2280 | if (type == float_type_node) |
2281 | code = BUILT_IN_TM_STORE_FLOAT; | |
2282 | else if (type == double_type_node) | |
2283 | code = BUILT_IN_TM_STORE_DOUBLE; | |
2284 | else if (type == long_double_type_node) | |
2285 | code = BUILT_IN_TM_STORE_LDOUBLE; | |
4c868789 | 2286 | else |
0a35513e | 2287 | { |
4c868789 RH |
2288 | if (TYPE_SIZE (type) == NULL || !tree_fits_uhwi_p (TYPE_SIZE (type))) |
2289 | return NULL; | |
2290 | unsigned HOST_WIDE_INT type_size = tree_to_uhwi (TYPE_SIZE (type)); | |
2291 | ||
2292 | if (TREE_CODE (type) == VECTOR_TYPE) | |
0a35513e | 2293 | { |
4c868789 RH |
2294 | switch (type_size) |
2295 | { | |
2296 | case 64: | |
2297 | code = BUILT_IN_TM_STORE_M64; | |
2298 | break; | |
2299 | case 128: | |
2300 | code = BUILT_IN_TM_STORE_M128; | |
2301 | break; | |
2302 | case 256: | |
2303 | code = BUILT_IN_TM_STORE_M256; | |
2304 | break; | |
2305 | default: | |
2306 | goto unhandled_vec; | |
2307 | } | |
2308 | if (!builtin_decl_explicit_p (code)) | |
2309 | goto unhandled_vec; | |
2310 | } | |
2311 | else | |
2312 | { | |
2313 | unhandled_vec: | |
2314 | switch (type_size) | |
2315 | { | |
2316 | case 8: | |
2317 | code = BUILT_IN_TM_STORE_1; | |
2318 | break; | |
2319 | case 16: | |
2320 | code = BUILT_IN_TM_STORE_2; | |
2321 | break; | |
2322 | case 32: | |
2323 | code = BUILT_IN_TM_STORE_4; | |
2324 | break; | |
2325 | case 64: | |
2326 | code = BUILT_IN_TM_STORE_8; | |
2327 | break; | |
2328 | default: | |
2329 | return NULL; | |
2330 | } | |
0a35513e AH |
2331 | } |
2332 | } | |
2333 | ||
4c868789 RH |
2334 | fn = builtin_decl_explicit (code); |
2335 | gcc_assert (fn); | |
0a35513e AH |
2336 | |
2337 | simple_type = TREE_VALUE (TREE_CHAIN (TYPE_ARG_TYPES (TREE_TYPE (fn)))); | |
2338 | ||
2339 | if (TREE_CODE (rhs) == CONSTRUCTOR) | |
2340 | { | |
2341 | /* Handle the easy initialization to zero. */ | |
9771b263 | 2342 | if (!CONSTRUCTOR_ELTS (rhs)) |
0a35513e AH |
2343 | rhs = build_int_cst (simple_type, 0); |
2344 | else | |
2345 | { | |
2346 | /* ...otherwise punt to the caller and probably use | |
2347 | BUILT_IN_TM_MEMMOVE, because we can't wrap a | |
2348 | VIEW_CONVERT_EXPR around a CONSTRUCTOR (below) and produce | |
2349 | valid gimple. */ | |
2350 | return NULL; | |
2351 | } | |
2352 | } | |
2353 | else if (!useless_type_conversion_p (simple_type, type)) | |
2354 | { | |
355fe088 | 2355 | gimple *g; |
0a35513e AH |
2356 | tree temp; |
2357 | ||
b731b390 | 2358 | temp = create_tmp_reg (simple_type); |
0a35513e AH |
2359 | t = fold_build1 (VIEW_CONVERT_EXPR, simple_type, rhs); |
2360 | g = gimple_build_assign (temp, t); | |
2361 | gimple_set_location (g, loc); | |
2362 | gsi_insert_before (gsi, g, GSI_SAME_STMT); | |
2363 | ||
2364 | rhs = temp; | |
2365 | } | |
2366 | ||
2367 | t = gimplify_addr (gsi, lhs); | |
2368 | gcall = gimple_build_call (fn, 2, t, rhs); | |
2369 | gimple_set_location (gcall, loc); | |
2370 | gsi_insert_before (gsi, gcall, GSI_SAME_STMT); | |
2371 | ||
2372 | return gcall; | |
2373 | } | |
2374 | ||
2375 | ||
2376 | /* Expand an assignment statement into transactional builtins. */ | |
2377 | ||
2378 | static void | |
2379 | expand_assign_tm (struct tm_region *region, gimple_stmt_iterator *gsi) | |
2380 | { | |
355fe088 | 2381 | gimple *stmt = gsi_stmt (*gsi); |
0a35513e AH |
2382 | location_t loc = gimple_location (stmt); |
2383 | tree lhs = gimple_assign_lhs (stmt); | |
2384 | tree rhs = gimple_assign_rhs1 (stmt); | |
2385 | bool store_p = requires_barrier (region->entry_block, lhs, NULL); | |
2386 | bool load_p = requires_barrier (region->entry_block, rhs, NULL); | |
355fe088 | 2387 | gimple *gcall = NULL; |
0a35513e AH |
2388 | |
2389 | if (!load_p && !store_p) | |
2390 | { | |
2391 | /* Add thread private addresses to log if applicable. */ | |
2392 | requires_barrier (region->entry_block, lhs, stmt); | |
2393 | gsi_next (gsi); | |
2394 | return; | |
2395 | } | |
2396 | ||
4c868789 RH |
2397 | if (load_p) |
2398 | transaction_subcode_ior (region, GTMA_HAVE_LOAD); | |
2399 | if (store_p) | |
2400 | transaction_subcode_ior (region, GTMA_HAVE_STORE); | |
2401 | ||
398b1daa | 2402 | // Remove original load/store statement. |
0a35513e AH |
2403 | gsi_remove (gsi, true); |
2404 | ||
4c868789 | 2405 | // Attempt to use a simple load/store helper function. |
0a35513e | 2406 | if (load_p && !store_p) |
4c868789 | 2407 | gcall = build_tm_load (loc, lhs, rhs, gsi); |
0a35513e | 2408 | else if (store_p && !load_p) |
4c868789 RH |
2409 | gcall = build_tm_store (loc, lhs, rhs, gsi); |
2410 | ||
2411 | // If gcall has not been set, then we do not have a simple helper | |
2412 | // function available for the type. This may be true of larger | |
2413 | // structures, vectors, and non-standard float types. | |
0a35513e AH |
2414 | if (!gcall) |
2415 | { | |
4c868789 | 2416 | tree lhs_addr, rhs_addr, ltmp = NULL, copy_fn; |
0a35513e | 2417 | |
4c868789 RH |
2418 | // If this is a type that we couldn't handle above, but it's |
2419 | // in a register, we must spill it to memory for the copy. | |
2420 | if (is_gimple_reg (lhs)) | |
2421 | { | |
2422 | ltmp = create_tmp_var (TREE_TYPE (lhs)); | |
2423 | lhs_addr = build_fold_addr_expr (ltmp); | |
2424 | } | |
2425 | else | |
2426 | lhs_addr = gimplify_addr (gsi, lhs); | |
2427 | if (is_gimple_reg (rhs)) | |
2428 | { | |
2429 | tree rtmp = create_tmp_var (TREE_TYPE (rhs)); | |
6d9ef062 | 2430 | TREE_ADDRESSABLE (rtmp) = 1; |
4c868789 RH |
2431 | rhs_addr = build_fold_addr_expr (rtmp); |
2432 | gcall = gimple_build_assign (rtmp, rhs); | |
2433 | gsi_insert_before (gsi, gcall, GSI_SAME_STMT); | |
2434 | } | |
2435 | else | |
2436 | rhs_addr = gimplify_addr (gsi, rhs); | |
713b8dfb | 2437 | |
4c868789 RH |
2438 | // Choose the appropriate memory transfer function. |
2439 | if (load_p && store_p) | |
713b8dfb | 2440 | { |
4c868789 RH |
2441 | // ??? Figure out if there's any possible overlap between |
2442 | // the LHS and the RHS and if not, use MEMCPY. | |
2443 | copy_fn = builtin_decl_explicit (BUILT_IN_TM_MEMMOVE); | |
2444 | } | |
2445 | else if (load_p) | |
2446 | { | |
2447 | // Note that the store is non-transactional and cannot overlap. | |
2448 | copy_fn = builtin_decl_explicit (BUILT_IN_TM_MEMCPY_RTWN); | |
713b8dfb AH |
2449 | } |
2450 | else | |
2451 | { | |
4c868789 RH |
2452 | // Note that the load is non-transactional and cannot overlap. |
2453 | copy_fn = builtin_decl_explicit (BUILT_IN_TM_MEMCPY_RNWT); | |
713b8dfb | 2454 | } |
4c868789 RH |
2455 | |
2456 | gcall = gimple_build_call (copy_fn, 3, lhs_addr, rhs_addr, | |
0a35513e AH |
2457 | TYPE_SIZE_UNIT (TREE_TYPE (lhs))); |
2458 | gimple_set_location (gcall, loc); | |
2459 | gsi_insert_before (gsi, gcall, GSI_SAME_STMT); | |
713b8dfb | 2460 | |
4c868789 | 2461 | if (ltmp) |
713b8dfb | 2462 | { |
4c868789 | 2463 | gcall = gimple_build_assign (lhs, ltmp); |
713b8dfb AH |
2464 | gsi_insert_before (gsi, gcall, GSI_SAME_STMT); |
2465 | } | |
0a35513e AH |
2466 | } |
2467 | ||
4c868789 RH |
2468 | // Now that we have the load/store in its instrumented form, add |
2469 | // thread private addresses to the log if applicable. | |
0a35513e AH |
2470 | if (!store_p) |
2471 | requires_barrier (region->entry_block, lhs, gcall); | |
0a35513e AH |
2472 | } |
2473 | ||
2474 | ||
2475 | /* Expand a call statement as appropriate for a transaction. That is, | |
2476 | either verify that the call does not affect the transaction, or | |
2477 | redirect the call to a clone that handles transactions, or change | |
2478 | the transaction state to IRREVOCABLE. Return true if the call is | |
2479 | one of the builtins that end a transaction. */ | |
2480 | ||
2481 | static bool | |
2482 | expand_call_tm (struct tm_region *region, | |
2483 | gimple_stmt_iterator *gsi) | |
2484 | { | |
538dd0b7 | 2485 | gcall *stmt = as_a <gcall *> (gsi_stmt (*gsi)); |
0a35513e AH |
2486 | tree lhs = gimple_call_lhs (stmt); |
2487 | tree fn_decl; | |
2488 | struct cgraph_node *node; | |
2489 | bool retval = false; | |
2490 | ||
2491 | fn_decl = gimple_call_fndecl (stmt); | |
2492 | ||
2493 | if (fn_decl == builtin_decl_explicit (BUILT_IN_TM_MEMCPY) | |
2494 | || fn_decl == builtin_decl_explicit (BUILT_IN_TM_MEMMOVE)) | |
2495 | transaction_subcode_ior (region, GTMA_HAVE_STORE | GTMA_HAVE_LOAD); | |
2496 | if (fn_decl == builtin_decl_explicit (BUILT_IN_TM_MEMSET)) | |
2497 | transaction_subcode_ior (region, GTMA_HAVE_STORE); | |
2498 | ||
2499 | if (is_tm_pure_call (stmt)) | |
2500 | return false; | |
2501 | ||
2502 | if (fn_decl) | |
2503 | retval = is_tm_ending_fndecl (fn_decl); | |
2504 | if (!retval) | |
2505 | { | |
2506 | /* Assume all non-const/pure calls write to memory, except | |
2507 | transaction ending builtins. */ | |
2508 | transaction_subcode_ior (region, GTMA_HAVE_STORE); | |
2509 | } | |
2510 | ||
2511 | /* For indirect calls, we already generated a call into the runtime. */ | |
2512 | if (!fn_decl) | |
2513 | { | |
2514 | tree fn = gimple_call_fn (stmt); | |
2515 | ||
2516 | /* We are guaranteed never to go irrevocable on a safe or pure | |
2517 | call, and the pure call was handled above. */ | |
2518 | if (is_tm_safe (fn)) | |
2519 | return false; | |
2520 | else | |
2521 | transaction_subcode_ior (region, GTMA_MAY_ENTER_IRREVOCABLE); | |
2522 | ||
2523 | return false; | |
2524 | } | |
2525 | ||
d52f5295 | 2526 | node = cgraph_node::get (fn_decl); |
91cad09b AH |
2527 | /* All calls should have cgraph here. */ |
2528 | if (!node) | |
2529 | { | |
2530 | /* We can have a nodeless call here if some pass after IPA-tm | |
2531 | added uninstrumented calls. For example, loop distribution | |
2532 | can transform certain loop constructs into __builtin_mem* | |
2533 | calls. In this case, see if we have a suitable TM | |
2534 | replacement and fill in the gaps. */ | |
2535 | gcc_assert (DECL_BUILT_IN_CLASS (fn_decl) == BUILT_IN_NORMAL); | |
2536 | enum built_in_function code = DECL_FUNCTION_CODE (fn_decl); | |
2537 | gcc_assert (code == BUILT_IN_MEMCPY | |
2538 | || code == BUILT_IN_MEMMOVE | |
2539 | || code == BUILT_IN_MEMSET); | |
2540 | ||
2541 | tree repl = find_tm_replacement_function (fn_decl); | |
2542 | if (repl) | |
2543 | { | |
2544 | gimple_call_set_fndecl (stmt, repl); | |
2545 | update_stmt (stmt); | |
d52f5295 | 2546 | node = cgraph_node::create (repl); |
87f94429 | 2547 | node->tm_may_enter_irr = false; |
91cad09b AH |
2548 | return expand_call_tm (region, gsi); |
2549 | } | |
2550 | gcc_unreachable (); | |
2551 | } | |
87f94429 | 2552 | if (node->tm_may_enter_irr) |
0a35513e AH |
2553 | transaction_subcode_ior (region, GTMA_MAY_ENTER_IRREVOCABLE); |
2554 | ||
2555 | if (is_tm_abort (fn_decl)) | |
2556 | { | |
2557 | transaction_subcode_ior (region, GTMA_HAVE_ABORT); | |
2558 | return true; | |
2559 | } | |
2560 | ||
2561 | /* Instrument the store if needed. | |
2562 | ||
2563 | If the assignment happens inside the function call (return slot | |
2564 | optimization), there is no instrumentation to be done, since | |
2565 | the callee should have done the right thing. */ | |
2566 | if (lhs && requires_barrier (region->entry_block, lhs, stmt) | |
2567 | && !gimple_call_return_slot_opt_p (stmt)) | |
2568 | { | |
b731b390 | 2569 | tree tmp = create_tmp_reg (TREE_TYPE (lhs)); |
0a35513e AH |
2570 | location_t loc = gimple_location (stmt); |
2571 | edge fallthru_edge = NULL; | |
538dd0b7 | 2572 | gassign *assign_stmt; |
0a35513e AH |
2573 | |
2574 | /* Remember if the call was going to throw. */ | |
36bbc05d | 2575 | if (stmt_can_throw_internal (cfun, stmt)) |
0a35513e AH |
2576 | { |
2577 | edge_iterator ei; | |
2578 | edge e; | |
2579 | basic_block bb = gimple_bb (stmt); | |
2580 | ||
2581 | FOR_EACH_EDGE (e, ei, bb->succs) | |
2582 | if (e->flags & EDGE_FALLTHRU) | |
2583 | { | |
2584 | fallthru_edge = e; | |
2585 | break; | |
2586 | } | |
2587 | } | |
2588 | ||
2589 | gimple_call_set_lhs (stmt, tmp); | |
2590 | update_stmt (stmt); | |
538dd0b7 DM |
2591 | assign_stmt = gimple_build_assign (lhs, tmp); |
2592 | gimple_set_location (assign_stmt, loc); | |
0a35513e AH |
2593 | |
2594 | /* We cannot throw in the middle of a BB. If the call was going | |
2595 | to throw, place the instrumentation on the fallthru edge, so | |
2596 | the call remains the last statement in the block. */ | |
2597 | if (fallthru_edge) | |
2598 | { | |
538dd0b7 | 2599 | gimple_seq fallthru_seq = gimple_seq_alloc_with_stmt (assign_stmt); |
0a35513e AH |
2600 | gimple_stmt_iterator fallthru_gsi = gsi_start (fallthru_seq); |
2601 | expand_assign_tm (region, &fallthru_gsi); | |
2602 | gsi_insert_seq_on_edge (fallthru_edge, fallthru_seq); | |
2603 | pending_edge_inserts_p = true; | |
2604 | } | |
2605 | else | |
2606 | { | |
538dd0b7 | 2607 | gsi_insert_after (gsi, assign_stmt, GSI_CONTINUE_LINKING); |
0a35513e AH |
2608 | expand_assign_tm (region, gsi); |
2609 | } | |
2610 | ||
2611 | transaction_subcode_ior (region, GTMA_HAVE_STORE); | |
2612 | } | |
2613 | ||
2614 | return retval; | |
2615 | } | |
2616 | ||
2617 | ||
2618 | /* Expand all statements in BB as appropriate for being inside | |
2619 | a transaction. */ | |
2620 | ||
2621 | static void | |
2622 | expand_block_tm (struct tm_region *region, basic_block bb) | |
2623 | { | |
2624 | gimple_stmt_iterator gsi; | |
2625 | ||
2626 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); ) | |
2627 | { | |
355fe088 | 2628 | gimple *stmt = gsi_stmt (gsi); |
0a35513e AH |
2629 | switch (gimple_code (stmt)) |
2630 | { | |
2631 | case GIMPLE_ASSIGN: | |
2632 | /* Only memory reads/writes need to be instrumented. */ | |
56cb44d4 MM |
2633 | if (gimple_assign_single_p (stmt) |
2634 | && !gimple_clobber_p (stmt)) | |
0a35513e AH |
2635 | { |
2636 | expand_assign_tm (region, &gsi); | |
2637 | continue; | |
2638 | } | |
2639 | break; | |
2640 | ||
2641 | case GIMPLE_CALL: | |
2642 | if (expand_call_tm (region, &gsi)) | |
2643 | return; | |
2644 | break; | |
2645 | ||
2646 | case GIMPLE_ASM: | |
2647 | gcc_unreachable (); | |
2648 | ||
2649 | default: | |
2650 | break; | |
2651 | } | |
2652 | if (!gsi_end_p (gsi)) | |
2653 | gsi_next (&gsi); | |
2654 | } | |
2655 | } | |
2656 | ||
2657 | /* Return the list of basic-blocks in REGION. | |
2658 | ||
2659 | STOP_AT_IRREVOCABLE_P is true if caller is uninterested in blocks | |
0ed3d24e AH |
2660 | following a TM_IRREVOCABLE call. |
2661 | ||
2662 | INCLUDE_UNINSTRUMENTED_P is TRUE if we should include the | |
2663 | uninstrumented code path blocks in the list of basic blocks | |
2664 | returned, false otherwise. */ | |
0a35513e | 2665 | |
9771b263 | 2666 | static vec<basic_block> |
0a35513e AH |
2667 | get_tm_region_blocks (basic_block entry_block, |
2668 | bitmap exit_blocks, | |
2669 | bitmap irr_blocks, | |
2670 | bitmap all_region_blocks, | |
0ed3d24e AH |
2671 | bool stop_at_irrevocable_p, |
2672 | bool include_uninstrumented_p = true) | |
0a35513e | 2673 | { |
6e1aa848 | 2674 | vec<basic_block> bbs = vNULL; |
0a35513e AH |
2675 | unsigned i; |
2676 | edge e; | |
2677 | edge_iterator ei; | |
2678 | bitmap visited_blocks = BITMAP_ALLOC (NULL); | |
2679 | ||
2680 | i = 0; | |
9771b263 | 2681 | bbs.safe_push (entry_block); |
0a35513e AH |
2682 | bitmap_set_bit (visited_blocks, entry_block->index); |
2683 | ||
2684 | do | |
2685 | { | |
9771b263 | 2686 | basic_block bb = bbs[i++]; |
0a35513e AH |
2687 | |
2688 | if (exit_blocks && | |
2689 | bitmap_bit_p (exit_blocks, bb->index)) | |
2690 | continue; | |
2691 | ||
2692 | if (stop_at_irrevocable_p | |
2693 | && irr_blocks | |
2694 | && bitmap_bit_p (irr_blocks, bb->index)) | |
2695 | continue; | |
2696 | ||
2697 | FOR_EACH_EDGE (e, ei, bb->succs) | |
0ed3d24e AH |
2698 | if ((include_uninstrumented_p |
2699 | || !(e->flags & EDGE_TM_UNINSTRUMENTED)) | |
2700 | && !bitmap_bit_p (visited_blocks, e->dest->index)) | |
0a35513e AH |
2701 | { |
2702 | bitmap_set_bit (visited_blocks, e->dest->index); | |
9771b263 | 2703 | bbs.safe_push (e->dest); |
0a35513e AH |
2704 | } |
2705 | } | |
9771b263 | 2706 | while (i < bbs.length ()); |
0a35513e AH |
2707 | |
2708 | if (all_region_blocks) | |
2709 | bitmap_ior_into (all_region_blocks, visited_blocks); | |
2710 | ||
2711 | BITMAP_FREE (visited_blocks); | |
2712 | return bbs; | |
2713 | } | |
2714 | ||
0ed3d24e AH |
2715 | // Callback data for collect_bb2reg. |
2716 | struct bb2reg_stuff | |
2717 | { | |
526ceb68 | 2718 | vec<tm_region *> *bb2reg; |
0ed3d24e AH |
2719 | bool include_uninstrumented_p; |
2720 | }; | |
2721 | ||
398b1daa AH |
2722 | // Callback for expand_regions, collect innermost region data for each bb. |
2723 | static void * | |
2724 | collect_bb2reg (struct tm_region *region, void *data) | |
2725 | { | |
0ed3d24e | 2726 | struct bb2reg_stuff *stuff = (struct bb2reg_stuff *)data; |
526ceb68 | 2727 | vec<tm_region *> *bb2reg = stuff->bb2reg; |
9771b263 | 2728 | vec<basic_block> queue; |
398b1daa AH |
2729 | unsigned int i; |
2730 | basic_block bb; | |
2731 | ||
2732 | queue = get_tm_region_blocks (region->entry_block, | |
2733 | region->exit_blocks, | |
2734 | region->irr_blocks, | |
2735 | NULL, | |
0ed3d24e AH |
2736 | /*stop_at_irr_p=*/true, |
2737 | stuff->include_uninstrumented_p); | |
398b1daa AH |
2738 | |
2739 | // We expect expand_region to perform a post-order traversal of the region | |
2740 | // tree. Therefore the last region seen for any bb is the innermost. | |
9771b263 DN |
2741 | FOR_EACH_VEC_ELT (queue, i, bb) |
2742 | (*bb2reg)[bb->index] = region; | |
398b1daa | 2743 | |
9771b263 | 2744 | queue.release (); |
398b1daa AH |
2745 | return NULL; |
2746 | } | |
2747 | ||
2748 | // Returns a vector, indexed by BB->INDEX, of the innermost tm_region to | |
2749 | // which a basic block belongs. Note that we only consider the instrumented | |
0ed3d24e AH |
2750 | // code paths for the region; the uninstrumented code paths are ignored if |
2751 | // INCLUDE_UNINSTRUMENTED_P is false. | |
398b1daa AH |
2752 | // |
2753 | // ??? This data is very similar to the bb_regions array that is collected | |
2754 | // during tm_region_init. Or, rather, this data is similar to what could | |
2755 | // be used within tm_region_init. The actual computation in tm_region_init | |
2756 | // begins and ends with bb_regions entirely full of NULL pointers, due to | |
2757 | // the way in which pointers are swapped in and out of the array. | |
2758 | // | |
2759 | // ??? Our callers expect that blocks are not shared between transactions. | |
2760 | // When the optimizers get too smart, and blocks are shared, then during | |
2761 | // the tm_mark phase we'll add log entries to only one of the two transactions, | |
2762 | // and in the tm_edge phase we'll add edges to the CFG that create invalid | |
2763 | // cycles. The symptom being SSA defs that do not dominate their uses. | |
2764 | // Note that the optimizers were locally correct with their transformation, | |
2765 | // as we have no info within the program that suggests that the blocks cannot | |
2766 | // be shared. | |
2767 | // | |
2768 | // ??? There is currently a hack inside tree-ssa-pre.c to work around the | |
2769 | // only known instance of this block sharing. | |
2770 | ||
526ceb68 | 2771 | static vec<tm_region *> |
0ed3d24e AH |
2772 | get_bb_regions_instrumented (bool traverse_clones, |
2773 | bool include_uninstrumented_p) | |
398b1daa | 2774 | { |
8b1c6fd7 | 2775 | unsigned n = last_basic_block_for_fn (cfun); |
0ed3d24e | 2776 | struct bb2reg_stuff stuff; |
526ceb68 | 2777 | vec<tm_region *> ret; |
398b1daa | 2778 | |
9771b263 | 2779 | ret.create (n); |
cb3874dc | 2780 | ret.safe_grow_cleared (n, true); |
0ed3d24e AH |
2781 | stuff.bb2reg = &ret; |
2782 | stuff.include_uninstrumented_p = include_uninstrumented_p; | |
2783 | expand_regions (all_tm_regions, collect_bb2reg, &stuff, traverse_clones); | |
398b1daa AH |
2784 | |
2785 | return ret; | |
2786 | } | |
2787 | ||
19c0d7df AH |
2788 | /* Set the IN_TRANSACTION for all gimple statements that appear in a |
2789 | transaction. */ | |
2790 | ||
2791 | void | |
2792 | compute_transaction_bits (void) | |
2793 | { | |
2794 | struct tm_region *region; | |
9771b263 | 2795 | vec<basic_block> queue; |
19c0d7df | 2796 | unsigned int i; |
19c0d7df AH |
2797 | basic_block bb; |
2798 | ||
2799 | /* ?? Perhaps we need to abstract gate_tm_init further, because we | |
2800 | certainly don't need it to calculate CDI_DOMINATOR info. */ | |
2801 | gate_tm_init (); | |
2802 | ||
11cd3bed | 2803 | FOR_EACH_BB_FN (bb, cfun) |
83e1a7f0 AH |
2804 | bb->flags &= ~BB_IN_TRANSACTION; |
2805 | ||
19c0d7df AH |
2806 | for (region = all_tm_regions; region; region = region->next) |
2807 | { | |
2808 | queue = get_tm_region_blocks (region->entry_block, | |
2809 | region->exit_blocks, | |
2810 | region->irr_blocks, | |
2811 | NULL, | |
2812 | /*stop_at_irr_p=*/true); | |
9771b263 | 2813 | for (i = 0; queue.iterate (i, &bb); ++i) |
83e1a7f0 | 2814 | bb->flags |= BB_IN_TRANSACTION; |
9771b263 | 2815 | queue.release (); |
19c0d7df AH |
2816 | } |
2817 | ||
2818 | if (all_tm_regions) | |
2819 | bitmap_obstack_release (&tm_obstack); | |
2820 | } | |
2821 | ||
398b1daa AH |
2822 | /* Replace the GIMPLE_TRANSACTION in this region with the corresponding |
2823 | call to BUILT_IN_TM_START. */ | |
2824 | ||
2825 | static void * | |
2826 | expand_transaction (struct tm_region *region, void *data ATTRIBUTE_UNUSED) | |
2827 | { | |
2828 | tree tm_start = builtin_decl_explicit (BUILT_IN_TM_START); | |
2829 | basic_block transaction_bb = gimple_bb (region->transaction_stmt); | |
2830 | tree tm_state = region->tm_state; | |
2831 | tree tm_state_type = TREE_TYPE (tm_state); | |
2832 | edge abort_edge = NULL; | |
2833 | edge inst_edge = NULL; | |
2834 | edge uninst_edge = NULL; | |
2835 | edge fallthru_edge = NULL; | |
2836 | ||
2837 | // Identify the various successors of the transaction start. | |
2838 | { | |
2839 | edge_iterator i; | |
2840 | edge e; | |
2841 | FOR_EACH_EDGE (e, i, transaction_bb->succs) | |
2842 | { | |
2843 | if (e->flags & EDGE_TM_ABORT) | |
2844 | abort_edge = e; | |
2845 | else if (e->flags & EDGE_TM_UNINSTRUMENTED) | |
2846 | uninst_edge = e; | |
2847 | else | |
2848 | inst_edge = e; | |
2849 | if (e->flags & EDGE_FALLTHRU) | |
2850 | fallthru_edge = e; | |
2851 | } | |
2852 | } | |
2853 | ||
2854 | /* ??? There are plenty of bits here we're not computing. */ | |
2855 | { | |
538dd0b7 | 2856 | int subcode = gimple_transaction_subcode (region->get_transaction_stmt ()); |
398b1daa AH |
2857 | int flags = 0; |
2858 | if (subcode & GTMA_DOES_GO_IRREVOCABLE) | |
2859 | flags |= PR_DOESGOIRREVOCABLE; | |
2860 | if ((subcode & GTMA_MAY_ENTER_IRREVOCABLE) == 0) | |
2861 | flags |= PR_HASNOIRREVOCABLE; | |
2862 | /* If the transaction does not have an abort in lexical scope and is not | |
2863 | marked as an outer transaction, then it will never abort. */ | |
2864 | if ((subcode & GTMA_HAVE_ABORT) == 0 && (subcode & GTMA_IS_OUTER) == 0) | |
2865 | flags |= PR_HASNOABORT; | |
2866 | if ((subcode & GTMA_HAVE_STORE) == 0) | |
2867 | flags |= PR_READONLY; | |
b7a78683 | 2868 | if (inst_edge && !(subcode & GTMA_HAS_NO_INSTRUMENTATION)) |
398b1daa AH |
2869 | flags |= PR_INSTRUMENTEDCODE; |
2870 | if (uninst_edge) | |
2871 | flags |= PR_UNINSTRUMENTEDCODE; | |
2872 | if (subcode & GTMA_IS_OUTER) | |
2873 | region->original_transaction_was_outer = true; | |
2874 | tree t = build_int_cst (tm_state_type, flags); | |
538dd0b7 | 2875 | gcall *call = gimple_build_call (tm_start, 1, t); |
398b1daa AH |
2876 | gimple_call_set_lhs (call, tm_state); |
2877 | gimple_set_location (call, gimple_location (region->transaction_stmt)); | |
2878 | ||
2879 | // Replace the GIMPLE_TRANSACTION with the call to BUILT_IN_TM_START. | |
2880 | gimple_stmt_iterator gsi = gsi_last_bb (transaction_bb); | |
2881 | gcc_assert (gsi_stmt (gsi) == region->transaction_stmt); | |
2882 | gsi_insert_before (&gsi, call, GSI_SAME_STMT); | |
2883 | gsi_remove (&gsi, true); | |
2884 | region->transaction_stmt = call; | |
2885 | } | |
2886 | ||
2887 | // Generate log saves. | |
9771b263 | 2888 | if (!tm_log_save_addresses.is_empty ()) |
398b1daa AH |
2889 | tm_log_emit_saves (region->entry_block, transaction_bb); |
2890 | ||
2891 | // In the beginning, we've no tests to perform on transaction restart. | |
2892 | // Note that after this point, transaction_bb becomes the "most recent | |
2893 | // block containing tests for the transaction". | |
2894 | region->restart_block = region->entry_block; | |
2895 | ||
2896 | // Generate log restores. | |
9771b263 | 2897 | if (!tm_log_save_addresses.is_empty ()) |
398b1daa AH |
2898 | { |
2899 | basic_block test_bb = create_empty_bb (transaction_bb); | |
2900 | basic_block code_bb = create_empty_bb (test_bb); | |
2901 | basic_block join_bb = create_empty_bb (code_bb); | |
726338f4 RB |
2902 | add_bb_to_loop (test_bb, transaction_bb->loop_father); |
2903 | add_bb_to_loop (code_bb, transaction_bb->loop_father); | |
2904 | add_bb_to_loop (join_bb, transaction_bb->loop_father); | |
398b1daa AH |
2905 | if (region->restart_block == region->entry_block) |
2906 | region->restart_block = test_bb; | |
2907 | ||
b731b390 | 2908 | tree t1 = create_tmp_reg (tm_state_type); |
398b1daa | 2909 | tree t2 = build_int_cst (tm_state_type, A_RESTORELIVEVARIABLES); |
355fe088 | 2910 | gimple *stmt = gimple_build_assign (t1, BIT_AND_EXPR, tm_state, t2); |
398b1daa AH |
2911 | gimple_stmt_iterator gsi = gsi_last_bb (test_bb); |
2912 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
2913 | ||
2914 | t2 = build_int_cst (tm_state_type, 0); | |
2915 | stmt = gimple_build_cond (NE_EXPR, t1, t2, NULL, NULL); | |
2916 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
2917 | ||
2918 | tm_log_emit_restores (region->entry_block, code_bb); | |
2919 | ||
2920 | edge ei = make_edge (transaction_bb, test_bb, EDGE_FALLTHRU); | |
2921 | edge et = make_edge (test_bb, code_bb, EDGE_TRUE_VALUE); | |
2922 | edge ef = make_edge (test_bb, join_bb, EDGE_FALSE_VALUE); | |
2923 | redirect_edge_pred (fallthru_edge, join_bb); | |
2924 | ||
398b1daa AH |
2925 | join_bb->count = test_bb->count = transaction_bb->count; |
2926 | ||
357067f2 JH |
2927 | ei->probability = profile_probability::always (); |
2928 | et->probability = profile_probability::likely (); | |
2929 | ef->probability = profile_probability::unlikely (); | |
398b1daa | 2930 | |
ef30ab83 | 2931 | code_bb->count = et->count (); |
398b1daa AH |
2932 | |
2933 | transaction_bb = join_bb; | |
2934 | } | |
2935 | ||
2936 | // If we have an ABORT edge, create a test to perform the abort. | |
2937 | if (abort_edge) | |
2938 | { | |
2939 | basic_block test_bb = create_empty_bb (transaction_bb); | |
726338f4 | 2940 | add_bb_to_loop (test_bb, transaction_bb->loop_father); |
398b1daa AH |
2941 | if (region->restart_block == region->entry_block) |
2942 | region->restart_block = test_bb; | |
2943 | ||
b731b390 | 2944 | tree t1 = create_tmp_reg (tm_state_type); |
398b1daa | 2945 | tree t2 = build_int_cst (tm_state_type, A_ABORTTRANSACTION); |
355fe088 | 2946 | gimple *stmt = gimple_build_assign (t1, BIT_AND_EXPR, tm_state, t2); |
398b1daa AH |
2947 | gimple_stmt_iterator gsi = gsi_last_bb (test_bb); |
2948 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
2949 | ||
2950 | t2 = build_int_cst (tm_state_type, 0); | |
2951 | stmt = gimple_build_cond (NE_EXPR, t1, t2, NULL, NULL); | |
2952 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
2953 | ||
2954 | edge ei = make_edge (transaction_bb, test_bb, EDGE_FALLTHRU); | |
398b1daa | 2955 | test_bb->count = transaction_bb->count; |
357067f2 | 2956 | ei->probability = profile_probability::always (); |
398b1daa AH |
2957 | |
2958 | // Not abort edge. If both are live, chose one at random as we'll | |
2959 | // we'll be fixing that up below. | |
2960 | redirect_edge_pred (fallthru_edge, test_bb); | |
2961 | fallthru_edge->flags = EDGE_FALSE_VALUE; | |
357067f2 | 2962 | fallthru_edge->probability = profile_probability::very_likely (); |
398b1daa AH |
2963 | |
2964 | // Abort/over edge. | |
2965 | redirect_edge_pred (abort_edge, test_bb); | |
2966 | abort_edge->flags = EDGE_TRUE_VALUE; | |
357067f2 | 2967 | abort_edge->probability = profile_probability::unlikely (); |
398b1daa AH |
2968 | |
2969 | transaction_bb = test_bb; | |
2970 | } | |
2971 | ||
2972 | // If we have both instrumented and uninstrumented code paths, select one. | |
2973 | if (inst_edge && uninst_edge) | |
2974 | { | |
2975 | basic_block test_bb = create_empty_bb (transaction_bb); | |
726338f4 | 2976 | add_bb_to_loop (test_bb, transaction_bb->loop_father); |
398b1daa AH |
2977 | if (region->restart_block == region->entry_block) |
2978 | region->restart_block = test_bb; | |
2979 | ||
b731b390 | 2980 | tree t1 = create_tmp_reg (tm_state_type); |
398b1daa AH |
2981 | tree t2 = build_int_cst (tm_state_type, A_RUNUNINSTRUMENTEDCODE); |
2982 | ||
355fe088 | 2983 | gimple *stmt = gimple_build_assign (t1, BIT_AND_EXPR, tm_state, t2); |
398b1daa AH |
2984 | gimple_stmt_iterator gsi = gsi_last_bb (test_bb); |
2985 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
2986 | ||
2987 | t2 = build_int_cst (tm_state_type, 0); | |
2988 | stmt = gimple_build_cond (NE_EXPR, t1, t2, NULL, NULL); | |
2989 | gsi_insert_after (&gsi, stmt, GSI_CONTINUE_LINKING); | |
2990 | ||
2991 | // Create the edge into test_bb first, as we want to copy values | |
2992 | // out of the fallthru edge. | |
2993 | edge e = make_edge (transaction_bb, test_bb, fallthru_edge->flags); | |
2994 | e->probability = fallthru_edge->probability; | |
ef30ab83 | 2995 | test_bb->count = fallthru_edge->count (); |
398b1daa AH |
2996 | |
2997 | // Now update the edges to the inst/uninist implementations. | |
2998 | // For now assume that the paths are equally likely. When using HTM, | |
2999 | // we'll try the uninst path first and fallback to inst path if htm | |
3000 | // buffers are exceeded. Without HTM we start with the inst path and | |
3001 | // use the uninst path when falling back to serial mode. | |
3002 | redirect_edge_pred (inst_edge, test_bb); | |
3003 | inst_edge->flags = EDGE_FALSE_VALUE; | |
357067f2 | 3004 | inst_edge->probability = profile_probability::even (); |
398b1daa AH |
3005 | |
3006 | redirect_edge_pred (uninst_edge, test_bb); | |
3007 | uninst_edge->flags = EDGE_TRUE_VALUE; | |
357067f2 | 3008 | uninst_edge->probability = profile_probability::even (); |
398b1daa AH |
3009 | } |
3010 | ||
3011 | // If we have no previous special cases, and we have PHIs at the beginning | |
3012 | // of the atomic region, this means we have a loop at the beginning of the | |
3013 | // atomic region that shares the first block. This can cause problems with | |
3014 | // the transaction restart abnormal edges to be added in the tm_edges pass. | |
3015 | // Solve this by adding a new empty block to receive the abnormal edges. | |
3016 | if (region->restart_block == region->entry_block | |
3017 | && phi_nodes (region->entry_block)) | |
3018 | { | |
3019 | basic_block empty_bb = create_empty_bb (transaction_bb); | |
3020 | region->restart_block = empty_bb; | |
726338f4 | 3021 | add_bb_to_loop (empty_bb, transaction_bb->loop_father); |
398b1daa AH |
3022 | |
3023 | redirect_edge_pred (fallthru_edge, empty_bb); | |
3024 | make_edge (transaction_bb, empty_bb, EDGE_FALLTHRU); | |
3025 | } | |
3026 | ||
3027 | return NULL; | |
3028 | } | |
3029 | ||
3030 | /* Generate the temporary to be used for the return value of | |
3031 | BUILT_IN_TM_START. */ | |
3032 | ||
3033 | static void * | |
3034 | generate_tm_state (struct tm_region *region, void *data ATTRIBUTE_UNUSED) | |
3035 | { | |
3036 | tree tm_start = builtin_decl_explicit (BUILT_IN_TM_START); | |
3037 | region->tm_state = | |
3038 | create_tmp_reg (TREE_TYPE (TREE_TYPE (tm_start)), "tm_state"); | |
3039 | ||
3040 | // Reset the subcode, post optimizations. We'll fill this in | |
3041 | // again as we process blocks. | |
3042 | if (region->exit_blocks) | |
3043 | { | |
538dd0b7 DM |
3044 | gtransaction *transaction_stmt = region->get_transaction_stmt (); |
3045 | unsigned int subcode = gimple_transaction_subcode (transaction_stmt); | |
398b1daa AH |
3046 | |
3047 | if (subcode & GTMA_DOES_GO_IRREVOCABLE) | |
3048 | subcode &= (GTMA_DECLARATION_MASK | GTMA_DOES_GO_IRREVOCABLE | |
b7a78683 AH |
3049 | | GTMA_MAY_ENTER_IRREVOCABLE |
3050 | | GTMA_HAS_NO_INSTRUMENTATION); | |
398b1daa AH |
3051 | else |
3052 | subcode &= GTMA_DECLARATION_MASK; | |
538dd0b7 | 3053 | gimple_transaction_set_subcode (transaction_stmt, subcode); |
398b1daa AH |
3054 | } |
3055 | ||
3056 | return NULL; | |
3057 | } | |
3058 | ||
3059 | // Propagate flags from inner transactions outwards. | |
3060 | static void | |
3061 | propagate_tm_flags_out (struct tm_region *region) | |
3062 | { | |
3063 | if (region == NULL) | |
3064 | return; | |
3065 | propagate_tm_flags_out (region->inner); | |
3066 | ||
3067 | if (region->outer && region->outer->transaction_stmt) | |
3068 | { | |
538dd0b7 DM |
3069 | unsigned s |
3070 | = gimple_transaction_subcode (region->get_transaction_stmt ()); | |
398b1daa AH |
3071 | s &= (GTMA_HAVE_ABORT | GTMA_HAVE_LOAD | GTMA_HAVE_STORE |
3072 | | GTMA_MAY_ENTER_IRREVOCABLE); | |
538dd0b7 DM |
3073 | s |= gimple_transaction_subcode (region->outer->get_transaction_stmt ()); |
3074 | gimple_transaction_set_subcode (region->outer->get_transaction_stmt (), | |
3075 | s); | |
398b1daa AH |
3076 | } |
3077 | ||
3078 | propagate_tm_flags_out (region->next); | |
3079 | } | |
3080 | ||
0a35513e AH |
3081 | /* Entry point to the MARK phase of TM expansion. Here we replace |
3082 | transactional memory statements with calls to builtins, and function | |
3083 | calls with their transactional clones (if available). But we don't | |
3084 | yet lower GIMPLE_TRANSACTION or add the transaction restart back-edges. */ | |
3085 | ||
3086 | static unsigned int | |
3087 | execute_tm_mark (void) | |
3088 | { | |
0a35513e AH |
3089 | pending_edge_inserts_p = false; |
3090 | ||
b5e10eac AH |
3091 | expand_regions (all_tm_regions, generate_tm_state, NULL, |
3092 | /*traverse_clones=*/true); | |
0a35513e | 3093 | |
398b1daa | 3094 | tm_log_init (); |
0a35513e | 3095 | |
526ceb68 | 3096 | vec<tm_region *> bb_regions |
0ed3d24e AH |
3097 | = get_bb_regions_instrumented (/*traverse_clones=*/true, |
3098 | /*include_uninstrumented_p=*/false); | |
398b1daa AH |
3099 | struct tm_region *r; |
3100 | unsigned i; | |
0a35513e | 3101 | |
398b1daa AH |
3102 | // Expand memory operations into calls into the runtime. |
3103 | // This collects log entries as well. | |
9771b263 | 3104 | FOR_EACH_VEC_ELT (bb_regions, i, r) |
6aad4455 AH |
3105 | { |
3106 | if (r != NULL) | |
3107 | { | |
3108 | if (r->transaction_stmt) | |
3109 | { | |
538dd0b7 DM |
3110 | unsigned sub |
3111 | = gimple_transaction_subcode (r->get_transaction_stmt ()); | |
6aad4455 AH |
3112 | |
3113 | /* If we're sure to go irrevocable, there won't be | |
3114 | anything to expand, since the run-time will go | |
3115 | irrevocable right away. */ | |
3116 | if (sub & GTMA_DOES_GO_IRREVOCABLE | |
3117 | && sub & GTMA_MAY_ENTER_IRREVOCABLE) | |
3118 | continue; | |
3119 | } | |
06e28de2 | 3120 | expand_block_tm (r, BASIC_BLOCK_FOR_FN (cfun, i)); |
6aad4455 AH |
3121 | } |
3122 | } | |
398b1daa | 3123 | |
639498a8 AH |
3124 | bb_regions.release (); |
3125 | ||
398b1daa AH |
3126 | // Propagate flags from inner transactions outwards. |
3127 | propagate_tm_flags_out (all_tm_regions); | |
3128 | ||
3129 | // Expand GIMPLE_TRANSACTIONs into calls into the runtime. | |
b5e10eac AH |
3130 | expand_regions (all_tm_regions, expand_transaction, NULL, |
3131 | /*traverse_clones=*/false); | |
398b1daa AH |
3132 | |
3133 | tm_log_emit (); | |
3134 | tm_log_delete (); | |
0a35513e AH |
3135 | |
3136 | if (pending_edge_inserts_p) | |
3137 | gsi_commit_edge_inserts (); | |
398b1daa | 3138 | free_dominance_info (CDI_DOMINATORS); |
0a35513e AH |
3139 | return 0; |
3140 | } | |
3141 | ||
17795822 TS |
3142 | namespace { |
3143 | ||
3144 | const pass_data pass_data_tm_mark = | |
27a4cd48 DM |
3145 | { |
3146 | GIMPLE_PASS, /* type */ | |
3147 | "tmmark", /* name */ | |
3148 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3149 | TV_TRANS_MEM, /* tv_id */ |
3150 | ( PROP_ssa | PROP_cfg ), /* properties_required */ | |
3151 | 0, /* properties_provided */ | |
3152 | 0, /* properties_destroyed */ | |
3153 | 0, /* todo_flags_start */ | |
3bea341f | 3154 | TODO_update_ssa, /* todo_flags_finish */ |
0a35513e | 3155 | }; |
27a4cd48 | 3156 | |
17795822 | 3157 | class pass_tm_mark : public gimple_opt_pass |
27a4cd48 DM |
3158 | { |
3159 | public: | |
c3284718 RS |
3160 | pass_tm_mark (gcc::context *ctxt) |
3161 | : gimple_opt_pass (pass_data_tm_mark, ctxt) | |
27a4cd48 DM |
3162 | {} |
3163 | ||
3164 | /* opt_pass methods: */ | |
be55bfe6 | 3165 | virtual unsigned int execute (function *) { return execute_tm_mark (); } |
27a4cd48 DM |
3166 | |
3167 | }; // class pass_tm_mark | |
3168 | ||
17795822 TS |
3169 | } // anon namespace |
3170 | ||
27a4cd48 DM |
3171 | gimple_opt_pass * |
3172 | make_pass_tm_mark (gcc::context *ctxt) | |
3173 | { | |
3174 | return new pass_tm_mark (ctxt); | |
3175 | } | |
0a35513e | 3176 | \f |
398b1daa AH |
3177 | |
3178 | /* Create an abnormal edge from STMT at iter, splitting the block | |
3179 | as necessary. Adjust *PNEXT as needed for the split block. */ | |
0a35513e AH |
3180 | |
3181 | static inline void | |
355fe088 | 3182 | split_bb_make_tm_edge (gimple *stmt, basic_block dest_bb, |
398b1daa | 3183 | gimple_stmt_iterator iter, gimple_stmt_iterator *pnext) |
0a35513e | 3184 | { |
398b1daa AH |
3185 | basic_block bb = gimple_bb (stmt); |
3186 | if (!gsi_one_before_end_p (iter)) | |
3187 | { | |
3188 | edge e = split_block (bb, stmt); | |
3189 | *pnext = gsi_start_bb (e->dest); | |
3190 | } | |
e39df546 JH |
3191 | edge e = make_edge (bb, dest_bb, EDGE_ABNORMAL); |
3192 | if (e) | |
ef30ab83 | 3193 | e->probability = profile_probability::guessed_never (); |
0a35513e | 3194 | |
398b1daa | 3195 | // Record the need for the edge for the benefit of the rtl passes. |
0a35513e | 3196 | if (cfun->gimple_df->tm_restart == NULL) |
50979347 TS |
3197 | cfun->gimple_df->tm_restart |
3198 | = hash_table<tm_restart_hasher>::create_ggc (31); | |
0a35513e | 3199 | |
398b1daa | 3200 | struct tm_restart_node dummy; |
0a35513e | 3201 | dummy.stmt = stmt; |
398b1daa AH |
3202 | dummy.label_or_list = gimple_block_label (dest_bb); |
3203 | ||
50979347 TS |
3204 | tm_restart_node **slot = cfun->gimple_df->tm_restart->find_slot (&dummy, |
3205 | INSERT); | |
3206 | struct tm_restart_node *n = *slot; | |
0a35513e AH |
3207 | if (n == NULL) |
3208 | { | |
766090c2 | 3209 | n = ggc_alloc<tm_restart_node> (); |
0a35513e AH |
3210 | *n = dummy; |
3211 | } | |
3212 | else | |
3213 | { | |
3214 | tree old = n->label_or_list; | |
3215 | if (TREE_CODE (old) == LABEL_DECL) | |
398b1daa | 3216 | old = tree_cons (NULL, old, NULL); |
0a35513e AH |
3217 | n->label_or_list = tree_cons (NULL, dummy.label_or_list, old); |
3218 | } | |
0a35513e AH |
3219 | } |
3220 | ||
0a35513e AH |
3221 | /* Split block BB as necessary for every builtin function we added, and |
3222 | wire up the abnormal back edges implied by the transaction restart. */ | |
3223 | ||
3224 | static void | |
398b1daa | 3225 | expand_block_edges (struct tm_region *const region, basic_block bb) |
0a35513e | 3226 | { |
398b1daa | 3227 | gimple_stmt_iterator gsi, next_gsi; |
0a35513e | 3228 | |
398b1daa | 3229 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi = next_gsi) |
0a35513e | 3230 | { |
355fe088 | 3231 | gimple *stmt = gsi_stmt (gsi); |
538dd0b7 | 3232 | gcall *call_stmt; |
0a35513e | 3233 | |
398b1daa AH |
3234 | next_gsi = gsi; |
3235 | gsi_next (&next_gsi); | |
3236 | ||
3237 | // ??? Shouldn't we split for any non-pure, non-irrevocable function? | |
538dd0b7 DM |
3238 | call_stmt = dyn_cast <gcall *> (stmt); |
3239 | if ((!call_stmt) | |
3240 | || (gimple_call_flags (call_stmt) & ECF_TM_BUILTIN) == 0) | |
398b1daa AH |
3241 | continue; |
3242 | ||
cb1180d5 | 3243 | if (gimple_call_builtin_p (call_stmt, BUILT_IN_TM_ABORT)) |
0a35513e | 3244 | { |
398b1daa AH |
3245 | // If we have a ``_transaction_cancel [[outer]]'', there is only |
3246 | // one abnormal edge: to the transaction marked OUTER. | |
3247 | // All compiler-generated instances of BUILT_IN_TM_ABORT have a | |
3248 | // constant argument, which we can examine here. Users invoking | |
3249 | // TM_ABORT directly get what they deserve. | |
538dd0b7 | 3250 | tree arg = gimple_call_arg (call_stmt, 0); |
398b1daa AH |
3251 | if (TREE_CODE (arg) == INTEGER_CST |
3252 | && (TREE_INT_CST_LOW (arg) & AR_OUTERABORT) != 0 | |
3253 | && !decl_is_tm_clone (current_function_decl)) | |
0a35513e | 3254 | { |
398b1daa AH |
3255 | // Find the GTMA_IS_OUTER transaction. |
3256 | for (struct tm_region *o = region; o; o = o->outer) | |
3257 | if (o->original_transaction_was_outer) | |
3258 | { | |
538dd0b7 | 3259 | split_bb_make_tm_edge (call_stmt, o->restart_block, |
398b1daa AH |
3260 | gsi, &next_gsi); |
3261 | break; | |
3262 | } | |
3263 | ||
3264 | // Otherwise, the front-end should have semantically checked | |
3265 | // outer aborts, but in either case the target region is not | |
3266 | // within this function. | |
3267 | continue; | |
0a35513e AH |
3268 | } |
3269 | ||
398b1daa AH |
3270 | // Non-outer, TM aborts have an abnormal edge to the inner-most |
3271 | // transaction, the one being aborted; | |
538dd0b7 DM |
3272 | split_bb_make_tm_edge (call_stmt, region->restart_block, gsi, |
3273 | &next_gsi); | |
0a35513e AH |
3274 | } |
3275 | ||
398b1daa AH |
3276 | // All TM builtins have an abnormal edge to the outer-most transaction. |
3277 | // We never restart inner transactions. For tm clones, we know a-priori | |
3278 | // that the outer-most transaction is outside the function. | |
3279 | if (decl_is_tm_clone (current_function_decl)) | |
3280 | continue; | |
0a35513e | 3281 | |
398b1daa AH |
3282 | if (cfun->gimple_df->tm_restart == NULL) |
3283 | cfun->gimple_df->tm_restart | |
50979347 | 3284 | = hash_table<tm_restart_hasher>::create_ggc (31); |
0a35513e | 3285 | |
398b1daa AH |
3286 | // All TM builtins have an abnormal edge to the outer-most transaction. |
3287 | // We never restart inner transactions. | |
3288 | for (struct tm_region *o = region; o; o = o->outer) | |
3289 | if (!o->outer) | |
3290 | { | |
538dd0b7 | 3291 | split_bb_make_tm_edge (call_stmt, o->restart_block, gsi, &next_gsi); |
398b1daa AH |
3292 | break; |
3293 | } | |
0a35513e | 3294 | |
398b1daa AH |
3295 | // Delete any tail-call annotation that may have been added. |
3296 | // The tail-call pass may have mis-identified the commit as being | |
3297 | // a candidate because we had not yet added this restart edge. | |
538dd0b7 | 3298 | gimple_call_set_tail (call_stmt, false); |
0a35513e AH |
3299 | } |
3300 | } | |
3301 | ||
3302 | /* Entry point to the final expansion of transactional nodes. */ | |
3303 | ||
17795822 TS |
3304 | namespace { |
3305 | ||
3306 | const pass_data pass_data_tm_edges = | |
27a4cd48 DM |
3307 | { |
3308 | GIMPLE_PASS, /* type */ | |
3309 | "tmedge", /* name */ | |
3310 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
3311 | TV_TRANS_MEM, /* tv_id */ |
3312 | ( PROP_ssa | PROP_cfg ), /* properties_required */ | |
3313 | 0, /* properties_provided */ | |
3314 | 0, /* properties_destroyed */ | |
3315 | 0, /* todo_flags_start */ | |
3bea341f | 3316 | TODO_update_ssa, /* todo_flags_finish */ |
0a35513e | 3317 | }; |
27a4cd48 | 3318 | |
17795822 | 3319 | class pass_tm_edges : public gimple_opt_pass |
27a4cd48 DM |
3320 | { |
3321 | public: | |
c3284718 RS |
3322 | pass_tm_edges (gcc::context *ctxt) |
3323 | : gimple_opt_pass (pass_data_tm_edges, ctxt) | |
27a4cd48 DM |
3324 | {} |
3325 | ||
3326 | /* opt_pass methods: */ | |
be55bfe6 | 3327 | virtual unsigned int execute (function *); |
27a4cd48 DM |
3328 | |
3329 | }; // class pass_tm_edges | |
3330 | ||
be55bfe6 TS |
3331 | unsigned int |
3332 | pass_tm_edges::execute (function *fun) | |
3333 | { | |
526ceb68 | 3334 | vec<tm_region *> bb_regions |
be55bfe6 TS |
3335 | = get_bb_regions_instrumented (/*traverse_clones=*/false, |
3336 | /*include_uninstrumented_p=*/true); | |
3337 | struct tm_region *r; | |
3338 | unsigned i; | |
3339 | ||
3340 | FOR_EACH_VEC_ELT (bb_regions, i, r) | |
3341 | if (r != NULL) | |
3342 | expand_block_edges (r, BASIC_BLOCK_FOR_FN (fun, i)); | |
3343 | ||
3344 | bb_regions.release (); | |
3345 | ||
3346 | /* We've got to release the dominance info now, to indicate that it | |
3347 | must be rebuilt completely. Otherwise we'll crash trying to update | |
3348 | the SSA web in the TODO section following this pass. */ | |
3349 | free_dominance_info (CDI_DOMINATORS); | |
687aed9c RB |
3350 | /* We'ge also wrecked loops badly with inserting of abnormal edges. */ |
3351 | loops_state_set (LOOPS_NEED_FIXUP); | |
be55bfe6 TS |
3352 | bitmap_obstack_release (&tm_obstack); |
3353 | all_tm_regions = NULL; | |
3354 | ||
3355 | return 0; | |
3356 | } | |
3357 | ||
17795822 TS |
3358 | } // anon namespace |
3359 | ||
27a4cd48 DM |
3360 | gimple_opt_pass * |
3361 | make_pass_tm_edges (gcc::context *ctxt) | |
3362 | { | |
3363 | return new pass_tm_edges (ctxt); | |
3364 | } | |
398b1daa AH |
3365 | \f |
3366 | /* Helper function for expand_regions. Expand REGION and recurse to | |
3367 | the inner region. Call CALLBACK on each region. CALLBACK returns | |
3368 | NULL to continue the traversal, otherwise a non-null value which | |
b5e10eac AH |
3369 | this function will return as well. TRAVERSE_CLONES is true if we |
3370 | should traverse transactional clones. */ | |
398b1daa AH |
3371 | |
3372 | static void * | |
3373 | expand_regions_1 (struct tm_region *region, | |
3374 | void *(*callback)(struct tm_region *, void *), | |
b5e10eac AH |
3375 | void *data, |
3376 | bool traverse_clones) | |
398b1daa AH |
3377 | { |
3378 | void *retval = NULL; | |
b5e10eac AH |
3379 | if (region->exit_blocks |
3380 | || (traverse_clones && decl_is_tm_clone (current_function_decl))) | |
398b1daa AH |
3381 | { |
3382 | retval = callback (region, data); | |
3383 | if (retval) | |
3384 | return retval; | |
3385 | } | |
3386 | if (region->inner) | |
3387 | { | |
b5e10eac | 3388 | retval = expand_regions (region->inner, callback, data, traverse_clones); |
398b1daa AH |
3389 | if (retval) |
3390 | return retval; | |
3391 | } | |
3392 | return retval; | |
3393 | } | |
3394 | ||
3395 | /* Traverse the regions enclosed and including REGION. Execute | |
3396 | CALLBACK for each region, passing DATA. CALLBACK returns NULL to | |
3397 | continue the traversal, otherwise a non-null value which this | |
b5e10eac AH |
3398 | function will return as well. TRAVERSE_CLONES is true if we should |
3399 | traverse transactional clones. */ | |
398b1daa AH |
3400 | |
3401 | static void * | |
3402 | expand_regions (struct tm_region *region, | |
3403 | void *(*callback)(struct tm_region *, void *), | |
b5e10eac AH |
3404 | void *data, |
3405 | bool traverse_clones) | |
398b1daa AH |
3406 | { |
3407 | void *retval = NULL; | |
3408 | while (region) | |
3409 | { | |
b5e10eac | 3410 | retval = expand_regions_1 (region, callback, data, traverse_clones); |
398b1daa AH |
3411 | if (retval) |
3412 | return retval; | |
3413 | region = region->next; | |
3414 | } | |
3415 | return retval; | |
3416 | } | |
3417 | ||
0a35513e AH |
3418 | \f |
3419 | /* A unique TM memory operation. */ | |
526ceb68 | 3420 | struct tm_memop |
0a35513e AH |
3421 | { |
3422 | /* Unique ID that all memory operations to the same location have. */ | |
3423 | unsigned int value_id; | |
3424 | /* Address of load/store. */ | |
3425 | tree addr; | |
526ceb68 | 3426 | }; |
0a35513e | 3427 | |
4a8fb1a1 LC |
3428 | /* TM memory operation hashtable helpers. */ |
3429 | ||
95fbe13e | 3430 | struct tm_memop_hasher : free_ptr_hash <tm_memop> |
4a8fb1a1 | 3431 | { |
67f58944 TS |
3432 | static inline hashval_t hash (const tm_memop *); |
3433 | static inline bool equal (const tm_memop *, const tm_memop *); | |
4a8fb1a1 LC |
3434 | }; |
3435 | ||
3436 | /* Htab support. Return a hash value for a `tm_memop'. */ | |
3437 | inline hashval_t | |
67f58944 | 3438 | tm_memop_hasher::hash (const tm_memop *mem) |
4a8fb1a1 LC |
3439 | { |
3440 | tree addr = mem->addr; | |
3441 | /* We drill down to the SSA_NAME/DECL for the hash, but equality is | |
3442 | actually done with operand_equal_p (see tm_memop_eq). */ | |
3443 | if (TREE_CODE (addr) == ADDR_EXPR) | |
3444 | addr = TREE_OPERAND (addr, 0); | |
3445 | return iterative_hash_expr (addr, 0); | |
3446 | } | |
3447 | ||
3448 | /* Htab support. Return true if two tm_memop's are the same. */ | |
3449 | inline bool | |
67f58944 | 3450 | tm_memop_hasher::equal (const tm_memop *mem1, const tm_memop *mem2) |
4a8fb1a1 LC |
3451 | { |
3452 | return operand_equal_p (mem1->addr, mem2->addr, 0); | |
3453 | } | |
3454 | ||
0a35513e AH |
3455 | /* Sets for solving data flow equations in the memory optimization pass. */ |
3456 | struct tm_memopt_bitmaps | |
3457 | { | |
3458 | /* Stores available to this BB upon entry. Basically, stores that | |
3459 | dominate this BB. */ | |
3460 | bitmap store_avail_in; | |
3461 | /* Stores available at the end of this BB. */ | |
3462 | bitmap store_avail_out; | |
3463 | bitmap store_antic_in; | |
3464 | bitmap store_antic_out; | |
3465 | /* Reads available to this BB upon entry. Basically, reads that | |
3466 | dominate this BB. */ | |
3467 | bitmap read_avail_in; | |
3468 | /* Reads available at the end of this BB. */ | |
3469 | bitmap read_avail_out; | |
3470 | /* Reads performed in this BB. */ | |
3471 | bitmap read_local; | |
3472 | /* Writes performed in this BB. */ | |
3473 | bitmap store_local; | |
3474 | ||
3475 | /* Temporary storage for pass. */ | |
3476 | /* Is the current BB in the worklist? */ | |
3477 | bool avail_in_worklist_p; | |
3478 | /* Have we visited this BB? */ | |
3479 | bool visited_p; | |
3480 | }; | |
3481 | ||
3482 | static bitmap_obstack tm_memopt_obstack; | |
3483 | ||
3484 | /* Unique counter for TM loads and stores. Loads and stores of the | |
3485 | same address get the same ID. */ | |
3486 | static unsigned int tm_memopt_value_id; | |
c203e8a7 | 3487 | static hash_table<tm_memop_hasher> *tm_memopt_value_numbers; |
0a35513e AH |
3488 | |
3489 | #define STORE_AVAIL_IN(BB) \ | |
3490 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->store_avail_in | |
3491 | #define STORE_AVAIL_OUT(BB) \ | |
3492 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->store_avail_out | |
3493 | #define STORE_ANTIC_IN(BB) \ | |
3494 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->store_antic_in | |
3495 | #define STORE_ANTIC_OUT(BB) \ | |
3496 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->store_antic_out | |
3497 | #define READ_AVAIL_IN(BB) \ | |
3498 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->read_avail_in | |
3499 | #define READ_AVAIL_OUT(BB) \ | |
3500 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->read_avail_out | |
3501 | #define READ_LOCAL(BB) \ | |
3502 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->read_local | |
3503 | #define STORE_LOCAL(BB) \ | |
3504 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->store_local | |
3505 | #define AVAIL_IN_WORKLIST_P(BB) \ | |
3506 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->avail_in_worklist_p | |
3507 | #define BB_VISITED_P(BB) \ | |
3508 | ((struct tm_memopt_bitmaps *) ((BB)->aux))->visited_p | |
3509 | ||
0a35513e AH |
3510 | /* Given a TM load/store in STMT, return the value number for the address |
3511 | it accesses. */ | |
3512 | ||
3513 | static unsigned int | |
355fe088 | 3514 | tm_memopt_value_number (gimple *stmt, enum insert_option op) |
0a35513e AH |
3515 | { |
3516 | struct tm_memop tmpmem, *mem; | |
4a8fb1a1 | 3517 | tm_memop **slot; |
0a35513e AH |
3518 | |
3519 | gcc_assert (is_tm_load (stmt) || is_tm_store (stmt)); | |
3520 | tmpmem.addr = gimple_call_arg (stmt, 0); | |
c203e8a7 | 3521 | slot = tm_memopt_value_numbers->find_slot (&tmpmem, op); |
0a35513e | 3522 | if (*slot) |
4a8fb1a1 | 3523 | mem = *slot; |
0a35513e AH |
3524 | else if (op == INSERT) |
3525 | { | |
3526 | mem = XNEW (struct tm_memop); | |
3527 | *slot = mem; | |
3528 | mem->value_id = tm_memopt_value_id++; | |
3529 | mem->addr = tmpmem.addr; | |
3530 | } | |
3531 | else | |
3532 | gcc_unreachable (); | |
3533 | return mem->value_id; | |
3534 | } | |
3535 | ||
3536 | /* Accumulate TM memory operations in BB into STORE_LOCAL and READ_LOCAL. */ | |
3537 | ||
3538 | static void | |
3539 | tm_memopt_accumulate_memops (basic_block bb) | |
3540 | { | |
3541 | gimple_stmt_iterator gsi; | |
3542 | ||
3543 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3544 | { | |
355fe088 | 3545 | gimple *stmt = gsi_stmt (gsi); |
0a35513e AH |
3546 | bitmap bits; |
3547 | unsigned int loc; | |
3548 | ||
3549 | if (is_tm_store (stmt)) | |
3550 | bits = STORE_LOCAL (bb); | |
3551 | else if (is_tm_load (stmt)) | |
3552 | bits = READ_LOCAL (bb); | |
3553 | else | |
3554 | continue; | |
3555 | ||
3556 | loc = tm_memopt_value_number (stmt, INSERT); | |
3557 | bitmap_set_bit (bits, loc); | |
3558 | if (dump_file) | |
3559 | { | |
3560 | fprintf (dump_file, "TM memopt (%s): value num=%d, BB=%d, addr=", | |
3561 | is_tm_load (stmt) ? "LOAD" : "STORE", loc, | |
3562 | gimple_bb (stmt)->index); | |
ef6cb4c7 | 3563 | print_generic_expr (dump_file, gimple_call_arg (stmt, 0)); |
0a35513e AH |
3564 | fprintf (dump_file, "\n"); |
3565 | } | |
3566 | } | |
3567 | } | |
3568 | ||
3569 | /* Prettily dump one of the memopt sets. BITS is the bitmap to dump. */ | |
3570 | ||
3571 | static void | |
3572 | dump_tm_memopt_set (const char *set_name, bitmap bits) | |
3573 | { | |
3574 | unsigned i; | |
3575 | bitmap_iterator bi; | |
3576 | const char *comma = ""; | |
3577 | ||
3578 | fprintf (dump_file, "TM memopt: %s: [", set_name); | |
3579 | EXECUTE_IF_SET_IN_BITMAP (bits, 0, i, bi) | |
3580 | { | |
c203e8a7 | 3581 | hash_table<tm_memop_hasher>::iterator hi; |
4a8fb1a1 | 3582 | struct tm_memop *mem = NULL; |
0a35513e AH |
3583 | |
3584 | /* Yeah, yeah, yeah. Whatever. This is just for debugging. */ | |
c203e8a7 | 3585 | FOR_EACH_HASH_TABLE_ELEMENT (*tm_memopt_value_numbers, mem, tm_memop_t, hi) |
0a35513e AH |
3586 | if (mem->value_id == i) |
3587 | break; | |
3588 | gcc_assert (mem->value_id == i); | |
3589 | fprintf (dump_file, "%s", comma); | |
3590 | comma = ", "; | |
ef6cb4c7 | 3591 | print_generic_expr (dump_file, mem->addr); |
0a35513e AH |
3592 | } |
3593 | fprintf (dump_file, "]\n"); | |
3594 | } | |
3595 | ||
3596 | /* Prettily dump all of the memopt sets in BLOCKS. */ | |
3597 | ||
3598 | static void | |
9771b263 | 3599 | dump_tm_memopt_sets (vec<basic_block> blocks) |
0a35513e AH |
3600 | { |
3601 | size_t i; | |
3602 | basic_block bb; | |
3603 | ||
9771b263 | 3604 | for (i = 0; blocks.iterate (i, &bb); ++i) |
0a35513e AH |
3605 | { |
3606 | fprintf (dump_file, "------------BB %d---------\n", bb->index); | |
3607 | dump_tm_memopt_set ("STORE_LOCAL", STORE_LOCAL (bb)); | |
3608 | dump_tm_memopt_set ("READ_LOCAL", READ_LOCAL (bb)); | |
3609 | dump_tm_memopt_set ("STORE_AVAIL_IN", STORE_AVAIL_IN (bb)); | |
3610 | dump_tm_memopt_set ("STORE_AVAIL_OUT", STORE_AVAIL_OUT (bb)); | |
3611 | dump_tm_memopt_set ("READ_AVAIL_IN", READ_AVAIL_IN (bb)); | |
3612 | dump_tm_memopt_set ("READ_AVAIL_OUT", READ_AVAIL_OUT (bb)); | |
3613 | } | |
3614 | } | |
3615 | ||
3616 | /* Compute {STORE,READ}_AVAIL_IN for the basic block BB. */ | |
3617 | ||
3618 | static void | |
3619 | tm_memopt_compute_avin (basic_block bb) | |
3620 | { | |
3621 | edge e; | |
3622 | unsigned ix; | |
3623 | ||
3624 | /* Seed with the AVOUT of any predecessor. */ | |
3625 | for (ix = 0; ix < EDGE_COUNT (bb->preds); ix++) | |
3626 | { | |
3627 | e = EDGE_PRED (bb, ix); | |
3628 | /* Make sure we have already visited this BB, and is thus | |
3629 | initialized. | |
3630 | ||
3631 | If e->src->aux is NULL, this predecessor is actually on an | |
3632 | enclosing transaction. We only care about the current | |
3633 | transaction, so ignore it. */ | |
3634 | if (e->src->aux && BB_VISITED_P (e->src)) | |
3635 | { | |
3636 | bitmap_copy (STORE_AVAIL_IN (bb), STORE_AVAIL_OUT (e->src)); | |
3637 | bitmap_copy (READ_AVAIL_IN (bb), READ_AVAIL_OUT (e->src)); | |
3638 | break; | |
3639 | } | |
3640 | } | |
3641 | ||
3642 | for (; ix < EDGE_COUNT (bb->preds); ix++) | |
3643 | { | |
3644 | e = EDGE_PRED (bb, ix); | |
3645 | if (e->src->aux && BB_VISITED_P (e->src)) | |
3646 | { | |
3647 | bitmap_and_into (STORE_AVAIL_IN (bb), STORE_AVAIL_OUT (e->src)); | |
3648 | bitmap_and_into (READ_AVAIL_IN (bb), READ_AVAIL_OUT (e->src)); | |
3649 | } | |
3650 | } | |
3651 | ||
3652 | BB_VISITED_P (bb) = true; | |
3653 | } | |
3654 | ||
3655 | /* Compute the STORE_ANTIC_IN for the basic block BB. */ | |
3656 | ||
3657 | static void | |
3658 | tm_memopt_compute_antin (basic_block bb) | |
3659 | { | |
3660 | edge e; | |
3661 | unsigned ix; | |
3662 | ||
3663 | /* Seed with the ANTIC_OUT of any successor. */ | |
3664 | for (ix = 0; ix < EDGE_COUNT (bb->succs); ix++) | |
3665 | { | |
3666 | e = EDGE_SUCC (bb, ix); | |
3667 | /* Make sure we have already visited this BB, and is thus | |
3668 | initialized. */ | |
3669 | if (BB_VISITED_P (e->dest)) | |
3670 | { | |
3671 | bitmap_copy (STORE_ANTIC_IN (bb), STORE_ANTIC_OUT (e->dest)); | |
3672 | break; | |
3673 | } | |
3674 | } | |
3675 | ||
3676 | for (; ix < EDGE_COUNT (bb->succs); ix++) | |
3677 | { | |
3678 | e = EDGE_SUCC (bb, ix); | |
3679 | if (BB_VISITED_P (e->dest)) | |
3680 | bitmap_and_into (STORE_ANTIC_IN (bb), STORE_ANTIC_OUT (e->dest)); | |
3681 | } | |
3682 | ||
3683 | BB_VISITED_P (bb) = true; | |
3684 | } | |
3685 | ||
3686 | /* Compute the AVAIL sets for every basic block in BLOCKS. | |
3687 | ||
3688 | We compute {STORE,READ}_AVAIL_{OUT,IN} as follows: | |
3689 | ||
3690 | AVAIL_OUT[bb] = union (AVAIL_IN[bb], LOCAL[bb]) | |
3691 | AVAIL_IN[bb] = intersect (AVAIL_OUT[predecessors]) | |
3692 | ||
3693 | This is basically what we do in lcm's compute_available(), but here | |
3694 | we calculate two sets of sets (one for STOREs and one for READs), | |
3695 | and we work on a region instead of the entire CFG. | |
3696 | ||
3697 | REGION is the TM region. | |
3698 | BLOCKS are the basic blocks in the region. */ | |
3699 | ||
3700 | static void | |
3701 | tm_memopt_compute_available (struct tm_region *region, | |
9771b263 | 3702 | vec<basic_block> blocks) |
0a35513e AH |
3703 | { |
3704 | edge e; | |
3705 | basic_block *worklist, *qin, *qout, *qend, bb; | |
3706 | unsigned int qlen, i; | |
3707 | edge_iterator ei; | |
3708 | bool changed; | |
3709 | ||
3710 | /* Allocate a worklist array/queue. Entries are only added to the | |
3711 | list if they were not already on the list. So the size is | |
3712 | bounded by the number of basic blocks in the region. */ | |
9f95c526 | 3713 | gcc_assert (!blocks.is_empty ()); |
9771b263 | 3714 | qlen = blocks.length () - 1; |
9f95c526 | 3715 | qin = qout = worklist = XNEWVEC (basic_block, qlen); |
0a35513e AH |
3716 | |
3717 | /* Put every block in the region on the worklist. */ | |
9771b263 | 3718 | for (i = 0; blocks.iterate (i, &bb); ++i) |
0a35513e AH |
3719 | { |
3720 | /* Seed AVAIL_OUT with the LOCAL set. */ | |
3721 | bitmap_ior_into (STORE_AVAIL_OUT (bb), STORE_LOCAL (bb)); | |
3722 | bitmap_ior_into (READ_AVAIL_OUT (bb), READ_LOCAL (bb)); | |
3723 | ||
3724 | AVAIL_IN_WORKLIST_P (bb) = true; | |
3725 | /* No need to insert the entry block, since it has an AVIN of | |
3726 | null, and an AVOUT that has already been seeded in. */ | |
3727 | if (bb != region->entry_block) | |
3728 | *qin++ = bb; | |
3729 | } | |
3730 | ||
3731 | /* The entry block has been initialized with the local sets. */ | |
3732 | BB_VISITED_P (region->entry_block) = true; | |
3733 | ||
3734 | qin = worklist; | |
3735 | qend = &worklist[qlen]; | |
3736 | ||
3737 | /* Iterate until the worklist is empty. */ | |
3738 | while (qlen) | |
3739 | { | |
3740 | /* Take the first entry off the worklist. */ | |
3741 | bb = *qout++; | |
3742 | qlen--; | |
3743 | ||
3744 | if (qout >= qend) | |
3745 | qout = worklist; | |
3746 | ||
3747 | /* This block can be added to the worklist again if necessary. */ | |
3748 | AVAIL_IN_WORKLIST_P (bb) = false; | |
3749 | tm_memopt_compute_avin (bb); | |
3750 | ||
3751 | /* Note: We do not add the LOCAL sets here because we already | |
3752 | seeded the AVAIL_OUT sets with them. */ | |
3753 | changed = bitmap_ior_into (STORE_AVAIL_OUT (bb), STORE_AVAIL_IN (bb)); | |
3754 | changed |= bitmap_ior_into (READ_AVAIL_OUT (bb), READ_AVAIL_IN (bb)); | |
3755 | if (changed | |
3756 | && (region->exit_blocks == NULL | |
3757 | || !bitmap_bit_p (region->exit_blocks, bb->index))) | |
3758 | /* If the out state of this block changed, then we need to add | |
3759 | its successors to the worklist if they are not already in. */ | |
3760 | FOR_EACH_EDGE (e, ei, bb->succs) | |
fefa31b5 DM |
3761 | if (!AVAIL_IN_WORKLIST_P (e->dest) |
3762 | && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)) | |
0a35513e AH |
3763 | { |
3764 | *qin++ = e->dest; | |
3765 | AVAIL_IN_WORKLIST_P (e->dest) = true; | |
3766 | qlen++; | |
3767 | ||
3768 | if (qin >= qend) | |
3769 | qin = worklist; | |
3770 | } | |
3771 | } | |
3772 | ||
3773 | free (worklist); | |
3774 | ||
3775 | if (dump_file) | |
3776 | dump_tm_memopt_sets (blocks); | |
3777 | } | |
3778 | ||
3779 | /* Compute ANTIC sets for every basic block in BLOCKS. | |
3780 | ||
3781 | We compute STORE_ANTIC_OUT as follows: | |
3782 | ||
3783 | STORE_ANTIC_OUT[bb] = union(STORE_ANTIC_IN[bb], STORE_LOCAL[bb]) | |
3784 | STORE_ANTIC_IN[bb] = intersect(STORE_ANTIC_OUT[successors]) | |
3785 | ||
3786 | REGION is the TM region. | |
3787 | BLOCKS are the basic blocks in the region. */ | |
3788 | ||
3789 | static void | |
3790 | tm_memopt_compute_antic (struct tm_region *region, | |
9771b263 | 3791 | vec<basic_block> blocks) |
0a35513e AH |
3792 | { |
3793 | edge e; | |
3794 | basic_block *worklist, *qin, *qout, *qend, bb; | |
3795 | unsigned int qlen; | |
3796 | int i; | |
3797 | edge_iterator ei; | |
3798 | ||
3799 | /* Allocate a worklist array/queue. Entries are only added to the | |
3800 | list if they were not already on the list. So the size is | |
3801 | bounded by the number of basic blocks in the region. */ | |
9771b263 | 3802 | qin = qout = worklist = XNEWVEC (basic_block, blocks.length ()); |
0a35513e | 3803 | |
9771b263 | 3804 | for (qlen = 0, i = blocks.length () - 1; i >= 0; --i) |
0a35513e | 3805 | { |
9771b263 | 3806 | bb = blocks[i]; |
0a35513e AH |
3807 | |
3808 | /* Seed ANTIC_OUT with the LOCAL set. */ | |
3809 | bitmap_ior_into (STORE_ANTIC_OUT (bb), STORE_LOCAL (bb)); | |
3810 | ||
3811 | /* Put every block in the region on the worklist. */ | |
3812 | AVAIL_IN_WORKLIST_P (bb) = true; | |
3813 | /* No need to insert exit blocks, since their ANTIC_IN is NULL, | |
3814 | and their ANTIC_OUT has already been seeded in. */ | |
3815 | if (region->exit_blocks | |
3816 | && !bitmap_bit_p (region->exit_blocks, bb->index)) | |
3817 | { | |
3818 | qlen++; | |
3819 | *qin++ = bb; | |
3820 | } | |
3821 | } | |
3822 | ||
3823 | /* The exit blocks have been initialized with the local sets. */ | |
3824 | if (region->exit_blocks) | |
3825 | { | |
3826 | unsigned int i; | |
3827 | bitmap_iterator bi; | |
3828 | EXECUTE_IF_SET_IN_BITMAP (region->exit_blocks, 0, i, bi) | |
06e28de2 | 3829 | BB_VISITED_P (BASIC_BLOCK_FOR_FN (cfun, i)) = true; |
0a35513e AH |
3830 | } |
3831 | ||
3832 | qin = worklist; | |
3833 | qend = &worklist[qlen]; | |
3834 | ||
3835 | /* Iterate until the worklist is empty. */ | |
3836 | while (qlen) | |
3837 | { | |
3838 | /* Take the first entry off the worklist. */ | |
3839 | bb = *qout++; | |
3840 | qlen--; | |
3841 | ||
3842 | if (qout >= qend) | |
3843 | qout = worklist; | |
3844 | ||
3845 | /* This block can be added to the worklist again if necessary. */ | |
3846 | AVAIL_IN_WORKLIST_P (bb) = false; | |
3847 | tm_memopt_compute_antin (bb); | |
3848 | ||
3849 | /* Note: We do not add the LOCAL sets here because we already | |
3850 | seeded the ANTIC_OUT sets with them. */ | |
3851 | if (bitmap_ior_into (STORE_ANTIC_OUT (bb), STORE_ANTIC_IN (bb)) | |
3852 | && bb != region->entry_block) | |
3853 | /* If the out state of this block changed, then we need to add | |
3854 | its predecessors to the worklist if they are not already in. */ | |
3855 | FOR_EACH_EDGE (e, ei, bb->preds) | |
3856 | if (!AVAIL_IN_WORKLIST_P (e->src)) | |
3857 | { | |
3858 | *qin++ = e->src; | |
3859 | AVAIL_IN_WORKLIST_P (e->src) = true; | |
3860 | qlen++; | |
3861 | ||
3862 | if (qin >= qend) | |
3863 | qin = worklist; | |
3864 | } | |
3865 | } | |
3866 | ||
3867 | free (worklist); | |
3868 | ||
3869 | if (dump_file) | |
3870 | dump_tm_memopt_sets (blocks); | |
3871 | } | |
3872 | ||
3873 | /* Offsets of load variants from TM_LOAD. For example, | |
3874 | BUILT_IN_TM_LOAD_RAR* is an offset of 1 from BUILT_IN_TM_LOAD*. | |
3875 | See gtm-builtins.def. */ | |
3876 | #define TRANSFORM_RAR 1 | |
3877 | #define TRANSFORM_RAW 2 | |
3878 | #define TRANSFORM_RFW 3 | |
3879 | /* Offsets of store variants from TM_STORE. */ | |
3880 | #define TRANSFORM_WAR 1 | |
3881 | #define TRANSFORM_WAW 2 | |
3882 | ||
3883 | /* Inform about a load/store optimization. */ | |
3884 | ||
3885 | static void | |
355fe088 | 3886 | dump_tm_memopt_transform (gimple *stmt) |
0a35513e AH |
3887 | { |
3888 | if (dump_file) | |
3889 | { | |
3890 | fprintf (dump_file, "TM memopt: transforming: "); | |
ef6cb4c7 | 3891 | print_gimple_stmt (dump_file, stmt, 0); |
0a35513e AH |
3892 | fprintf (dump_file, "\n"); |
3893 | } | |
3894 | } | |
3895 | ||
3896 | /* Perform a read/write optimization. Replaces the TM builtin in STMT | |
3897 | by a builtin that is OFFSET entries down in the builtins table in | |
3898 | gtm-builtins.def. */ | |
3899 | ||
3900 | static void | |
3901 | tm_memopt_transform_stmt (unsigned int offset, | |
538dd0b7 | 3902 | gcall *stmt, |
0a35513e AH |
3903 | gimple_stmt_iterator *gsi) |
3904 | { | |
3905 | tree fn = gimple_call_fn (stmt); | |
3906 | gcc_assert (TREE_CODE (fn) == ADDR_EXPR); | |
3907 | TREE_OPERAND (fn, 0) | |
3908 | = builtin_decl_explicit ((enum built_in_function) | |
3909 | (DECL_FUNCTION_CODE (TREE_OPERAND (fn, 0)) | |
3910 | + offset)); | |
3911 | gimple_call_set_fn (stmt, fn); | |
3912 | gsi_replace (gsi, stmt, true); | |
3913 | dump_tm_memopt_transform (stmt); | |
3914 | } | |
3915 | ||
3916 | /* Perform the actual TM memory optimization transformations in the | |
3917 | basic blocks in BLOCKS. */ | |
3918 | ||
3919 | static void | |
9771b263 | 3920 | tm_memopt_transform_blocks (vec<basic_block> blocks) |
0a35513e AH |
3921 | { |
3922 | size_t i; | |
3923 | basic_block bb; | |
3924 | gimple_stmt_iterator gsi; | |
3925 | ||
9771b263 | 3926 | for (i = 0; blocks.iterate (i, &bb); ++i) |
0a35513e AH |
3927 | { |
3928 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3929 | { | |
355fe088 | 3930 | gimple *stmt = gsi_stmt (gsi); |
0a35513e AH |
3931 | bitmap read_avail = READ_AVAIL_IN (bb); |
3932 | bitmap store_avail = STORE_AVAIL_IN (bb); | |
3933 | bitmap store_antic = STORE_ANTIC_OUT (bb); | |
3934 | unsigned int loc; | |
3935 | ||
3936 | if (is_tm_simple_load (stmt)) | |
3937 | { | |
538dd0b7 | 3938 | gcall *call_stmt = as_a <gcall *> (stmt); |
0a35513e AH |
3939 | loc = tm_memopt_value_number (stmt, NO_INSERT); |
3940 | if (store_avail && bitmap_bit_p (store_avail, loc)) | |
538dd0b7 | 3941 | tm_memopt_transform_stmt (TRANSFORM_RAW, call_stmt, &gsi); |
0a35513e AH |
3942 | else if (store_antic && bitmap_bit_p (store_antic, loc)) |
3943 | { | |
538dd0b7 | 3944 | tm_memopt_transform_stmt (TRANSFORM_RFW, call_stmt, &gsi); |
0a35513e AH |
3945 | bitmap_set_bit (store_avail, loc); |
3946 | } | |
3947 | else if (read_avail && bitmap_bit_p (read_avail, loc)) | |
538dd0b7 | 3948 | tm_memopt_transform_stmt (TRANSFORM_RAR, call_stmt, &gsi); |
0a35513e AH |
3949 | else |
3950 | bitmap_set_bit (read_avail, loc); | |
3951 | } | |
3952 | else if (is_tm_simple_store (stmt)) | |
3953 | { | |
538dd0b7 | 3954 | gcall *call_stmt = as_a <gcall *> (stmt); |
0a35513e AH |
3955 | loc = tm_memopt_value_number (stmt, NO_INSERT); |
3956 | if (store_avail && bitmap_bit_p (store_avail, loc)) | |
538dd0b7 | 3957 | tm_memopt_transform_stmt (TRANSFORM_WAW, call_stmt, &gsi); |
0a35513e AH |
3958 | else |
3959 | { | |
3960 | if (read_avail && bitmap_bit_p (read_avail, loc)) | |
538dd0b7 | 3961 | tm_memopt_transform_stmt (TRANSFORM_WAR, call_stmt, &gsi); |
0a35513e AH |
3962 | bitmap_set_bit (store_avail, loc); |
3963 | } | |
3964 | } | |
3965 | } | |
3966 | } | |
3967 | } | |
3968 | ||
3969 | /* Return a new set of bitmaps for a BB. */ | |
3970 | ||
3971 | static struct tm_memopt_bitmaps * | |
3972 | tm_memopt_init_sets (void) | |
3973 | { | |
3974 | struct tm_memopt_bitmaps *b | |
3975 | = XOBNEW (&tm_memopt_obstack.obstack, struct tm_memopt_bitmaps); | |
3976 | b->store_avail_in = BITMAP_ALLOC (&tm_memopt_obstack); | |
3977 | b->store_avail_out = BITMAP_ALLOC (&tm_memopt_obstack); | |
3978 | b->store_antic_in = BITMAP_ALLOC (&tm_memopt_obstack); | |
3979 | b->store_antic_out = BITMAP_ALLOC (&tm_memopt_obstack); | |
3980 | b->store_avail_out = BITMAP_ALLOC (&tm_memopt_obstack); | |
3981 | b->read_avail_in = BITMAP_ALLOC (&tm_memopt_obstack); | |
3982 | b->read_avail_out = BITMAP_ALLOC (&tm_memopt_obstack); | |
3983 | b->read_local = BITMAP_ALLOC (&tm_memopt_obstack); | |
3984 | b->store_local = BITMAP_ALLOC (&tm_memopt_obstack); | |
3985 | return b; | |
3986 | } | |
3987 | ||
3988 | /* Free sets computed for each BB. */ | |
3989 | ||
3990 | static void | |
9771b263 | 3991 | tm_memopt_free_sets (vec<basic_block> blocks) |
0a35513e AH |
3992 | { |
3993 | size_t i; | |
3994 | basic_block bb; | |
3995 | ||
9771b263 | 3996 | for (i = 0; blocks.iterate (i, &bb); ++i) |
0a35513e AH |
3997 | bb->aux = NULL; |
3998 | } | |
3999 | ||
4000 | /* Clear the visited bit for every basic block in BLOCKS. */ | |
4001 | ||
4002 | static void | |
9771b263 | 4003 | tm_memopt_clear_visited (vec<basic_block> blocks) |
0a35513e AH |
4004 | { |
4005 | size_t i; | |
4006 | basic_block bb; | |
4007 | ||
9771b263 | 4008 | for (i = 0; blocks.iterate (i, &bb); ++i) |
0a35513e AH |
4009 | BB_VISITED_P (bb) = false; |
4010 | } | |
4011 | ||
4012 | /* Replace TM load/stores with hints for the runtime. We handle | |
4013 | things like read-after-write, write-after-read, read-after-read, | |
4014 | read-for-write, etc. */ | |
4015 | ||
4016 | static unsigned int | |
4017 | execute_tm_memopt (void) | |
4018 | { | |
4019 | struct tm_region *region; | |
9771b263 | 4020 | vec<basic_block> bbs; |
0a35513e AH |
4021 | |
4022 | tm_memopt_value_id = 0; | |
c203e8a7 | 4023 | tm_memopt_value_numbers = new hash_table<tm_memop_hasher> (10); |
0a35513e AH |
4024 | |
4025 | for (region = all_tm_regions; region; region = region->next) | |
4026 | { | |
4027 | /* All the TM stores/loads in the current region. */ | |
4028 | size_t i; | |
4029 | basic_block bb; | |
4030 | ||
4031 | bitmap_obstack_initialize (&tm_memopt_obstack); | |
4032 | ||
4033 | /* Save all BBs for the current region. */ | |
4034 | bbs = get_tm_region_blocks (region->entry_block, | |
4035 | region->exit_blocks, | |
4036 | region->irr_blocks, | |
4037 | NULL, | |
4038 | false); | |
4039 | ||
4040 | /* Collect all the memory operations. */ | |
9771b263 | 4041 | for (i = 0; bbs.iterate (i, &bb); ++i) |
0a35513e AH |
4042 | { |
4043 | bb->aux = tm_memopt_init_sets (); | |
4044 | tm_memopt_accumulate_memops (bb); | |
4045 | } | |
4046 | ||
4047 | /* Solve data flow equations and transform each block accordingly. */ | |
4048 | tm_memopt_clear_visited (bbs); | |
4049 | tm_memopt_compute_available (region, bbs); | |
4050 | tm_memopt_clear_visited (bbs); | |
4051 | tm_memopt_compute_antic (region, bbs); | |
4052 | tm_memopt_transform_blocks (bbs); | |
4053 | ||
4054 | tm_memopt_free_sets (bbs); | |
9771b263 | 4055 | bbs.release (); |
0a35513e | 4056 | bitmap_obstack_release (&tm_memopt_obstack); |
c203e8a7 | 4057 | tm_memopt_value_numbers->empty (); |
0a35513e AH |
4058 | } |
4059 | ||
c203e8a7 TS |
4060 | delete tm_memopt_value_numbers; |
4061 | tm_memopt_value_numbers = NULL; | |
0a35513e AH |
4062 | return 0; |
4063 | } | |
4064 | ||
17795822 TS |
4065 | namespace { |
4066 | ||
4067 | const pass_data pass_data_tm_memopt = | |
27a4cd48 DM |
4068 | { |
4069 | GIMPLE_PASS, /* type */ | |
4070 | "tmmemopt", /* name */ | |
4071 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
4072 | TV_TRANS_MEM, /* tv_id */ |
4073 | ( PROP_ssa | PROP_cfg ), /* properties_required */ | |
4074 | 0, /* properties_provided */ | |
4075 | 0, /* properties_destroyed */ | |
4076 | 0, /* todo_flags_start */ | |
4077 | 0, /* todo_flags_finish */ | |
0a35513e AH |
4078 | }; |
4079 | ||
17795822 | 4080 | class pass_tm_memopt : public gimple_opt_pass |
27a4cd48 DM |
4081 | { |
4082 | public: | |
c3284718 RS |
4083 | pass_tm_memopt (gcc::context *ctxt) |
4084 | : gimple_opt_pass (pass_data_tm_memopt, ctxt) | |
27a4cd48 DM |
4085 | {} |
4086 | ||
4087 | /* opt_pass methods: */ | |
1a3d085c | 4088 | virtual bool gate (function *) { return flag_tm && optimize > 0; } |
be55bfe6 | 4089 | virtual unsigned int execute (function *) { return execute_tm_memopt (); } |
27a4cd48 DM |
4090 | |
4091 | }; // class pass_tm_memopt | |
4092 | ||
17795822 TS |
4093 | } // anon namespace |
4094 | ||
27a4cd48 DM |
4095 | gimple_opt_pass * |
4096 | make_pass_tm_memopt (gcc::context *ctxt) | |
4097 | { | |
4098 | return new pass_tm_memopt (ctxt); | |
4099 | } | |
4100 | ||
0a35513e AH |
4101 | \f |
4102 | /* Interprocedual analysis for the creation of transactional clones. | |
4103 | The aim of this pass is to find which functions are referenced in | |
4104 | a non-irrevocable transaction context, and for those over which | |
4105 | we have control (or user directive), create a version of the | |
4106 | function which uses only the transactional interface to reference | |
4107 | protected memories. This analysis proceeds in several steps: | |
4108 | ||
4109 | (1) Collect the set of all possible transactional clones: | |
4110 | ||
4111 | (a) For all local public functions marked tm_callable, push | |
4112 | it onto the tm_callee queue. | |
4113 | ||
4114 | (b) For all local functions, scan for calls in transaction blocks. | |
4115 | Push the caller and callee onto the tm_caller and tm_callee | |
4116 | queues. Count the number of callers for each callee. | |
4117 | ||
4118 | (c) For each local function on the callee list, assume we will | |
4119 | create a transactional clone. Push *all* calls onto the | |
4120 | callee queues; count the number of clone callers separately | |
4121 | to the number of original callers. | |
4122 | ||
4123 | (2) Propagate irrevocable status up the dominator tree: | |
4124 | ||
4125 | (a) Any external function on the callee list that is not marked | |
4126 | tm_callable is irrevocable. Push all callers of such onto | |
4127 | a worklist. | |
4128 | ||
4129 | (b) For each function on the worklist, mark each block that | |
4130 | contains an irrevocable call. Use the AND operator to | |
4131 | propagate that mark up the dominator tree. | |
4132 | ||
4133 | (c) If we reach the entry block for a possible transactional | |
4134 | clone, then the transactional clone is irrevocable, and | |
4135 | we should not create the clone after all. Push all | |
4136 | callers onto the worklist. | |
4137 | ||
4138 | (d) Place tm_irrevocable calls at the beginning of the relevant | |
4139 | blocks. Special case here is the entry block for the entire | |
4140 | transaction region; there we mark it GTMA_DOES_GO_IRREVOCABLE for | |
4141 | the library to begin the region in serial mode. Decrement | |
4142 | the call count for all callees in the irrevocable region. | |
4143 | ||
4144 | (3) Create the transactional clones: | |
4145 | ||
4146 | Any tm_callee that still has a non-zero call count is cloned. | |
4147 | */ | |
4148 | ||
4149 | /* This structure is stored in the AUX field of each cgraph_node. */ | |
4150 | struct tm_ipa_cg_data | |
4151 | { | |
4152 | /* The clone of the function that got created. */ | |
4153 | struct cgraph_node *clone; | |
4154 | ||
4155 | /* The tm regions in the normal function. */ | |
4156 | struct tm_region *all_tm_regions; | |
4157 | ||
4158 | /* The blocks of the normal/clone functions that contain irrevocable | |
4159 | calls, or blocks that are post-dominated by irrevocable calls. */ | |
4160 | bitmap irrevocable_blocks_normal; | |
4161 | bitmap irrevocable_blocks_clone; | |
4162 | ||
4163 | /* The blocks of the normal function that are involved in transactions. */ | |
4164 | bitmap transaction_blocks_normal; | |
4165 | ||
4166 | /* The number of callers to the transactional clone of this function | |
4167 | from normal and transactional clones respectively. */ | |
4168 | unsigned tm_callers_normal; | |
4169 | unsigned tm_callers_clone; | |
4170 | ||
4171 | /* True if all calls to this function's transactional clone | |
4172 | are irrevocable. Also automatically true if the function | |
4173 | has no transactional clone. */ | |
4174 | bool is_irrevocable; | |
4175 | ||
4176 | /* Flags indicating the presence of this function in various queues. */ | |
4177 | bool in_callee_queue; | |
4178 | bool in_worklist; | |
4179 | ||
4180 | /* Flags indicating the kind of scan desired while in the worklist. */ | |
4181 | bool want_irr_scan_normal; | |
4182 | }; | |
4183 | ||
d52f5295 | 4184 | typedef vec<cgraph_node *> cgraph_node_queue; |
0a35513e AH |
4185 | |
4186 | /* Return the ipa data associated with NODE, allocating zeroed memory | |
594ec92f AH |
4187 | if necessary. TRAVERSE_ALIASES is true if we must traverse aliases |
4188 | and set *NODE accordingly. */ | |
0a35513e AH |
4189 | |
4190 | static struct tm_ipa_cg_data * | |
594ec92f | 4191 | get_cg_data (struct cgraph_node **node, bool traverse_aliases) |
0a35513e | 4192 | { |
594ec92f AH |
4193 | struct tm_ipa_cg_data *d; |
4194 | ||
67348ccc | 4195 | if (traverse_aliases && (*node)->alias) |
d52f5295 | 4196 | *node = (*node)->get_alias_target (); |
594ec92f | 4197 | |
67348ccc | 4198 | d = (struct tm_ipa_cg_data *) (*node)->aux; |
0a35513e AH |
4199 | |
4200 | if (d == NULL) | |
4201 | { | |
4202 | d = (struct tm_ipa_cg_data *) | |
4203 | obstack_alloc (&tm_obstack.obstack, sizeof (*d)); | |
67348ccc | 4204 | (*node)->aux = (void *) d; |
0a35513e AH |
4205 | memset (d, 0, sizeof (*d)); |
4206 | } | |
4207 | ||
4208 | return d; | |
4209 | } | |
4210 | ||
4211 | /* Add NODE to the end of QUEUE, unless IN_QUEUE_P indicates that | |
4212 | it is already present. */ | |
4213 | ||
4214 | static void | |
4215 | maybe_push_queue (struct cgraph_node *node, | |
4216 | cgraph_node_queue *queue_p, bool *in_queue_p) | |
4217 | { | |
4218 | if (!*in_queue_p) | |
4219 | { | |
4220 | *in_queue_p = true; | |
9771b263 | 4221 | queue_p->safe_push (node); |
0a35513e AH |
4222 | } |
4223 | } | |
4224 | ||
4225 | /* A subroutine of ipa_tm_scan_calls_transaction and ipa_tm_scan_calls_clone. | |
4226 | Queue all callees within block BB. */ | |
4227 | ||
4228 | static void | |
4229 | ipa_tm_scan_calls_block (cgraph_node_queue *callees_p, | |
4230 | basic_block bb, bool for_clone) | |
4231 | { | |
4232 | gimple_stmt_iterator gsi; | |
4233 | ||
4234 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
4235 | { | |
355fe088 | 4236 | gimple *stmt = gsi_stmt (gsi); |
0a35513e AH |
4237 | if (is_gimple_call (stmt) && !is_tm_pure_call (stmt)) |
4238 | { | |
4239 | tree fndecl = gimple_call_fndecl (stmt); | |
4240 | if (fndecl) | |
4241 | { | |
4242 | struct tm_ipa_cg_data *d; | |
4243 | unsigned *pcallers; | |
4244 | struct cgraph_node *node; | |
4245 | ||
4246 | if (is_tm_ending_fndecl (fndecl)) | |
4247 | continue; | |
4248 | if (find_tm_replacement_function (fndecl)) | |
4249 | continue; | |
4250 | ||
d52f5295 | 4251 | node = cgraph_node::get (fndecl); |
0a35513e | 4252 | gcc_assert (node != NULL); |
594ec92f | 4253 | d = get_cg_data (&node, true); |
0a35513e AH |
4254 | |
4255 | pcallers = (for_clone ? &d->tm_callers_clone | |
4256 | : &d->tm_callers_normal); | |
4257 | *pcallers += 1; | |
4258 | ||
4259 | maybe_push_queue (node, callees_p, &d->in_callee_queue); | |
4260 | } | |
4261 | } | |
4262 | } | |
4263 | } | |
4264 | ||
4265 | /* Scan all calls in NODE that are within a transaction region, | |
4266 | and push the resulting nodes into the callee queue. */ | |
4267 | ||
4268 | static void | |
4269 | ipa_tm_scan_calls_transaction (struct tm_ipa_cg_data *d, | |
4270 | cgraph_node_queue *callees_p) | |
4271 | { | |
0a35513e AH |
4272 | d->transaction_blocks_normal = BITMAP_ALLOC (&tm_obstack); |
4273 | d->all_tm_regions = all_tm_regions; | |
4274 | ||
7c11b0fe | 4275 | for (tm_region *r = all_tm_regions; r; r = r->next) |
0a35513e | 4276 | { |
9771b263 | 4277 | vec<basic_block> bbs; |
0a35513e AH |
4278 | basic_block bb; |
4279 | unsigned i; | |
4280 | ||
4281 | bbs = get_tm_region_blocks (r->entry_block, r->exit_blocks, NULL, | |
7c11b0fe | 4282 | d->transaction_blocks_normal, false, false); |
398b1daa | 4283 | |
9771b263 | 4284 | FOR_EACH_VEC_ELT (bbs, i, bb) |
0a35513e AH |
4285 | ipa_tm_scan_calls_block (callees_p, bb, false); |
4286 | ||
9771b263 | 4287 | bbs.release (); |
0a35513e AH |
4288 | } |
4289 | } | |
4290 | ||
4291 | /* Scan all calls in NODE as if this is the transactional clone, | |
4292 | and push the destinations into the callee queue. */ | |
4293 | ||
4294 | static void | |
4295 | ipa_tm_scan_calls_clone (struct cgraph_node *node, | |
4296 | cgraph_node_queue *callees_p) | |
4297 | { | |
67348ccc | 4298 | struct function *fn = DECL_STRUCT_FUNCTION (node->decl); |
0a35513e AH |
4299 | basic_block bb; |
4300 | ||
4301 | FOR_EACH_BB_FN (bb, fn) | |
4302 | ipa_tm_scan_calls_block (callees_p, bb, true); | |
4303 | } | |
4304 | ||
4305 | /* The function NODE has been detected to be irrevocable. Push all | |
4306 | of its callers onto WORKLIST for the purpose of re-scanning them. */ | |
4307 | ||
4308 | static void | |
4309 | ipa_tm_note_irrevocable (struct cgraph_node *node, | |
4310 | cgraph_node_queue *worklist_p) | |
4311 | { | |
594ec92f | 4312 | struct tm_ipa_cg_data *d = get_cg_data (&node, true); |
0a35513e AH |
4313 | struct cgraph_edge *e; |
4314 | ||
4315 | d->is_irrevocable = true; | |
4316 | ||
4317 | for (e = node->callers; e ; e = e->next_caller) | |
4318 | { | |
4319 | basic_block bb; | |
594ec92f | 4320 | struct cgraph_node *caller; |
0a35513e AH |
4321 | |
4322 | /* Don't examine recursive calls. */ | |
4323 | if (e->caller == node) | |
4324 | continue; | |
4325 | /* Even if we think we can go irrevocable, believe the user | |
4326 | above all. */ | |
67348ccc | 4327 | if (is_tm_safe_or_pure (e->caller->decl)) |
0a35513e AH |
4328 | continue; |
4329 | ||
594ec92f AH |
4330 | caller = e->caller; |
4331 | d = get_cg_data (&caller, true); | |
0a35513e AH |
4332 | |
4333 | /* Check if the callee is in a transactional region. If so, | |
4334 | schedule the function for normal re-scan as well. */ | |
4335 | bb = gimple_bb (e->call_stmt); | |
4336 | gcc_assert (bb != NULL); | |
4337 | if (d->transaction_blocks_normal | |
4338 | && bitmap_bit_p (d->transaction_blocks_normal, bb->index)) | |
4339 | d->want_irr_scan_normal = true; | |
4340 | ||
594ec92f | 4341 | maybe_push_queue (caller, worklist_p, &d->in_worklist); |
0a35513e AH |
4342 | } |
4343 | } | |
4344 | ||
4345 | /* A subroutine of ipa_tm_scan_irr_blocks; return true iff any statement | |
4346 | within the block is irrevocable. */ | |
4347 | ||
4348 | static bool | |
4349 | ipa_tm_scan_irr_block (basic_block bb) | |
4350 | { | |
4351 | gimple_stmt_iterator gsi; | |
4352 | tree fn; | |
4353 | ||
4354 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
4355 | { | |
355fe088 | 4356 | gimple *stmt = gsi_stmt (gsi); |
0a35513e AH |
4357 | switch (gimple_code (stmt)) |
4358 | { | |
a3770d3b AH |
4359 | case GIMPLE_ASSIGN: |
4360 | if (gimple_assign_single_p (stmt)) | |
4361 | { | |
4362 | tree lhs = gimple_assign_lhs (stmt); | |
4363 | tree rhs = gimple_assign_rhs1 (stmt); | |
1d4afc04 | 4364 | if (volatile_lvalue_p (lhs) || volatile_lvalue_p (rhs)) |
a3770d3b AH |
4365 | return true; |
4366 | } | |
4367 | break; | |
4368 | ||
0a35513e | 4369 | case GIMPLE_CALL: |
a3770d3b AH |
4370 | { |
4371 | tree lhs = gimple_call_lhs (stmt); | |
1d4afc04 | 4372 | if (lhs && volatile_lvalue_p (lhs)) |
a3770d3b | 4373 | return true; |
0a35513e | 4374 | |
a3770d3b AH |
4375 | if (is_tm_pure_call (stmt)) |
4376 | break; | |
0a35513e | 4377 | |
a3770d3b | 4378 | fn = gimple_call_fn (stmt); |
0a35513e | 4379 | |
a3770d3b AH |
4380 | /* Functions with the attribute are by definition irrevocable. */ |
4381 | if (is_tm_irrevocable (fn)) | |
4382 | return true; | |
0a35513e | 4383 | |
a3770d3b AH |
4384 | /* For direct function calls, go ahead and check for replacement |
4385 | functions, or transitive irrevocable functions. For indirect | |
4386 | functions, we'll ask the runtime. */ | |
4387 | if (TREE_CODE (fn) == ADDR_EXPR) | |
4388 | { | |
4389 | struct tm_ipa_cg_data *d; | |
4390 | struct cgraph_node *node; | |
0a35513e | 4391 | |
a3770d3b AH |
4392 | fn = TREE_OPERAND (fn, 0); |
4393 | if (is_tm_ending_fndecl (fn)) | |
4394 | break; | |
4395 | if (find_tm_replacement_function (fn)) | |
4396 | break; | |
80fd8eba | 4397 | |
d52f5295 | 4398 | node = cgraph_node::get (fn); |
a3770d3b AH |
4399 | d = get_cg_data (&node, true); |
4400 | ||
4401 | /* Return true if irrevocable, but above all, believe | |
4402 | the user. */ | |
4403 | if (d->is_irrevocable | |
4404 | && !is_tm_safe_or_pure (fn)) | |
4405 | return true; | |
4406 | } | |
4407 | break; | |
4408 | } | |
0a35513e AH |
4409 | |
4410 | case GIMPLE_ASM: | |
4411 | /* ??? The Approved Method of indicating that an inline | |
4412 | assembly statement is not relevant to the transaction | |
4413 | is to wrap it in a __tm_waiver block. This is not | |
4414 | yet implemented, so we can't check for it. */ | |
a4d031c7 AH |
4415 | if (is_tm_safe (current_function_decl)) |
4416 | { | |
4417 | tree t = build1 (NOP_EXPR, void_type_node, size_zero_node); | |
4418 | SET_EXPR_LOCATION (t, gimple_location (stmt)); | |
a9c697b8 MS |
4419 | error ("%K%<asm%> not allowed in %<transaction_safe%> function", |
4420 | t); | |
a4d031c7 | 4421 | } |
0a35513e AH |
4422 | return true; |
4423 | ||
4424 | default: | |
4425 | break; | |
4426 | } | |
4427 | } | |
4428 | ||
4429 | return false; | |
4430 | } | |
4431 | ||
4432 | /* For each of the blocks seeded witin PQUEUE, walk the CFG looking | |
4433 | for new irrevocable blocks, marking them in NEW_IRR. Don't bother | |
4434 | scanning past OLD_IRR or EXIT_BLOCKS. */ | |
4435 | ||
4436 | static bool | |
9771b263 | 4437 | ipa_tm_scan_irr_blocks (vec<basic_block> *pqueue, bitmap new_irr, |
0a35513e AH |
4438 | bitmap old_irr, bitmap exit_blocks) |
4439 | { | |
4440 | bool any_new_irr = false; | |
4441 | edge e; | |
4442 | edge_iterator ei; | |
4443 | bitmap visited_blocks = BITMAP_ALLOC (NULL); | |
4444 | ||
4445 | do | |
4446 | { | |
9771b263 | 4447 | basic_block bb = pqueue->pop (); |
0a35513e AH |
4448 | |
4449 | /* Don't re-scan blocks we know already are irrevocable. */ | |
4450 | if (old_irr && bitmap_bit_p (old_irr, bb->index)) | |
4451 | continue; | |
4452 | ||
4453 | if (ipa_tm_scan_irr_block (bb)) | |
4454 | { | |
4455 | bitmap_set_bit (new_irr, bb->index); | |
4456 | any_new_irr = true; | |
4457 | } | |
4458 | else if (exit_blocks == NULL || !bitmap_bit_p (exit_blocks, bb->index)) | |
4459 | { | |
4460 | FOR_EACH_EDGE (e, ei, bb->succs) | |
4461 | if (!bitmap_bit_p (visited_blocks, e->dest->index)) | |
4462 | { | |
4463 | bitmap_set_bit (visited_blocks, e->dest->index); | |
9771b263 | 4464 | pqueue->safe_push (e->dest); |
0a35513e AH |
4465 | } |
4466 | } | |
4467 | } | |
9771b263 | 4468 | while (!pqueue->is_empty ()); |
0a35513e AH |
4469 | |
4470 | BITMAP_FREE (visited_blocks); | |
4471 | ||
4472 | return any_new_irr; | |
4473 | } | |
4474 | ||
4475 | /* Propagate the irrevocable property both up and down the dominator tree. | |
4476 | BB is the current block being scanned; EXIT_BLOCKS are the edges of the | |
4477 | TM regions; OLD_IRR are the results of a previous scan of the dominator | |
4478 | tree which has been fully propagated; NEW_IRR is the set of new blocks | |
4479 | which are gaining the irrevocable property during the current scan. */ | |
4480 | ||
4481 | static void | |
4482 | ipa_tm_propagate_irr (basic_block entry_block, bitmap new_irr, | |
4483 | bitmap old_irr, bitmap exit_blocks) | |
4484 | { | |
9771b263 | 4485 | vec<basic_block> bbs; |
0a35513e AH |
4486 | bitmap all_region_blocks; |
4487 | ||
4488 | /* If this block is in the old set, no need to rescan. */ | |
4489 | if (old_irr && bitmap_bit_p (old_irr, entry_block->index)) | |
4490 | return; | |
4491 | ||
4492 | all_region_blocks = BITMAP_ALLOC (&tm_obstack); | |
4493 | bbs = get_tm_region_blocks (entry_block, exit_blocks, NULL, | |
4494 | all_region_blocks, false); | |
4495 | do | |
4496 | { | |
9771b263 | 4497 | basic_block bb = bbs.pop (); |
0a35513e AH |
4498 | bool this_irr = bitmap_bit_p (new_irr, bb->index); |
4499 | bool all_son_irr = false; | |
4500 | edge_iterator ei; | |
4501 | edge e; | |
4502 | ||
4503 | /* Propagate up. If my children are, I am too, but we must have | |
4504 | at least one child that is. */ | |
4505 | if (!this_irr) | |
4506 | { | |
4507 | FOR_EACH_EDGE (e, ei, bb->succs) | |
4508 | { | |
4509 | if (!bitmap_bit_p (new_irr, e->dest->index)) | |
4510 | { | |
4511 | all_son_irr = false; | |
4512 | break; | |
4513 | } | |
4514 | else | |
4515 | all_son_irr = true; | |
4516 | } | |
4517 | if (all_son_irr) | |
4518 | { | |
4519 | /* Add block to new_irr if it hasn't already been processed. */ | |
4520 | if (!old_irr || !bitmap_bit_p (old_irr, bb->index)) | |
4521 | { | |
4522 | bitmap_set_bit (new_irr, bb->index); | |
4523 | this_irr = true; | |
4524 | } | |
4525 | } | |
4526 | } | |
4527 | ||
4528 | /* Propagate down to everyone we immediately dominate. */ | |
4529 | if (this_irr) | |
4530 | { | |
4531 | basic_block son; | |
4532 | for (son = first_dom_son (CDI_DOMINATORS, bb); | |
4533 | son; | |
4534 | son = next_dom_son (CDI_DOMINATORS, son)) | |
4535 | { | |
4536 | /* Make sure block is actually in a TM region, and it | |
4537 | isn't already in old_irr. */ | |
4538 | if ((!old_irr || !bitmap_bit_p (old_irr, son->index)) | |
4539 | && bitmap_bit_p (all_region_blocks, son->index)) | |
4540 | bitmap_set_bit (new_irr, son->index); | |
4541 | } | |
4542 | } | |
4543 | } | |
9771b263 | 4544 | while (!bbs.is_empty ()); |
0a35513e AH |
4545 | |
4546 | BITMAP_FREE (all_region_blocks); | |
9771b263 | 4547 | bbs.release (); |
0a35513e AH |
4548 | } |
4549 | ||
4550 | static void | |
4551 | ipa_tm_decrement_clone_counts (basic_block bb, bool for_clone) | |
4552 | { | |
4553 | gimple_stmt_iterator gsi; | |
4554 | ||
4555 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
4556 | { | |
355fe088 | 4557 | gimple *stmt = gsi_stmt (gsi); |
0a35513e AH |
4558 | if (is_gimple_call (stmt) && !is_tm_pure_call (stmt)) |
4559 | { | |
4560 | tree fndecl = gimple_call_fndecl (stmt); | |
4561 | if (fndecl) | |
4562 | { | |
4563 | struct tm_ipa_cg_data *d; | |
4564 | unsigned *pcallers; | |
594ec92f | 4565 | struct cgraph_node *tnode; |
0a35513e AH |
4566 | |
4567 | if (is_tm_ending_fndecl (fndecl)) | |
4568 | continue; | |
4569 | if (find_tm_replacement_function (fndecl)) | |
4570 | continue; | |
4571 | ||
d52f5295 | 4572 | tnode = cgraph_node::get (fndecl); |
594ec92f AH |
4573 | d = get_cg_data (&tnode, true); |
4574 | ||
0a35513e AH |
4575 | pcallers = (for_clone ? &d->tm_callers_clone |
4576 | : &d->tm_callers_normal); | |
4577 | ||
4578 | gcc_assert (*pcallers > 0); | |
4579 | *pcallers -= 1; | |
4580 | } | |
4581 | } | |
4582 | } | |
4583 | } | |
4584 | ||
4585 | /* (Re-)Scan the transaction blocks in NODE for calls to irrevocable functions, | |
4586 | as well as other irrevocable actions such as inline assembly. Mark all | |
4587 | such blocks as irrevocable and decrement the number of calls to | |
4588 | transactional clones. Return true if, for the transactional clone, the | |
4589 | entire function is irrevocable. */ | |
4590 | ||
4591 | static bool | |
4592 | ipa_tm_scan_irr_function (struct cgraph_node *node, bool for_clone) | |
4593 | { | |
4594 | struct tm_ipa_cg_data *d; | |
4595 | bitmap new_irr, old_irr; | |
0a35513e AH |
4596 | bool ret = false; |
4597 | ||
80fd8eba | 4598 | /* Builtin operators (operator new, and such). */ |
67348ccc DM |
4599 | if (DECL_STRUCT_FUNCTION (node->decl) == NULL |
4600 | || DECL_STRUCT_FUNCTION (node->decl)->cfg == NULL) | |
80fd8eba AH |
4601 | return false; |
4602 | ||
67348ccc | 4603 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); |
0a35513e AH |
4604 | calculate_dominance_info (CDI_DOMINATORS); |
4605 | ||
594ec92f | 4606 | d = get_cg_data (&node, true); |
00f96dc9 | 4607 | auto_vec<basic_block, 10> queue; |
0a35513e AH |
4608 | new_irr = BITMAP_ALLOC (&tm_obstack); |
4609 | ||
4610 | /* Scan each tm region, propagating irrevocable status through the tree. */ | |
4611 | if (for_clone) | |
4612 | { | |
4613 | old_irr = d->irrevocable_blocks_clone; | |
fefa31b5 | 4614 | queue.quick_push (single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun))); |
0a35513e AH |
4615 | if (ipa_tm_scan_irr_blocks (&queue, new_irr, old_irr, NULL)) |
4616 | { | |
fefa31b5 DM |
4617 | ipa_tm_propagate_irr (single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)), |
4618 | new_irr, | |
0a35513e | 4619 | old_irr, NULL); |
fefa31b5 DM |
4620 | ret = bitmap_bit_p (new_irr, |
4621 | single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun))->index); | |
0a35513e AH |
4622 | } |
4623 | } | |
4624 | else | |
4625 | { | |
4626 | struct tm_region *region; | |
4627 | ||
4628 | old_irr = d->irrevocable_blocks_normal; | |
4629 | for (region = d->all_tm_regions; region; region = region->next) | |
4630 | { | |
9771b263 | 4631 | queue.quick_push (region->entry_block); |
0a35513e AH |
4632 | if (ipa_tm_scan_irr_blocks (&queue, new_irr, old_irr, |
4633 | region->exit_blocks)) | |
4634 | ipa_tm_propagate_irr (region->entry_block, new_irr, old_irr, | |
4635 | region->exit_blocks); | |
4636 | } | |
4637 | } | |
4638 | ||
4639 | /* If we found any new irrevocable blocks, reduce the call count for | |
4640 | transactional clones within the irrevocable blocks. Save the new | |
4641 | set of irrevocable blocks for next time. */ | |
4642 | if (!bitmap_empty_p (new_irr)) | |
4643 | { | |
4644 | bitmap_iterator bmi; | |
4645 | unsigned i; | |
4646 | ||
4647 | EXECUTE_IF_SET_IN_BITMAP (new_irr, 0, i, bmi) | |
06e28de2 DM |
4648 | ipa_tm_decrement_clone_counts (BASIC_BLOCK_FOR_FN (cfun, i), |
4649 | for_clone); | |
0a35513e AH |
4650 | |
4651 | if (old_irr) | |
4652 | { | |
4653 | bitmap_ior_into (old_irr, new_irr); | |
4654 | BITMAP_FREE (new_irr); | |
4655 | } | |
4656 | else if (for_clone) | |
4657 | d->irrevocable_blocks_clone = new_irr; | |
4658 | else | |
4659 | d->irrevocable_blocks_normal = new_irr; | |
4660 | ||
4661 | if (dump_file && new_irr) | |
4662 | { | |
4663 | const char *dname; | |
4664 | bitmap_iterator bmi; | |
4665 | unsigned i; | |
4666 | ||
4667 | dname = lang_hooks.decl_printable_name (current_function_decl, 2); | |
4668 | EXECUTE_IF_SET_IN_BITMAP (new_irr, 0, i, bmi) | |
4669 | fprintf (dump_file, "%s: bb %d goes irrevocable\n", dname, i); | |
4670 | } | |
4671 | } | |
4672 | else | |
4673 | BITMAP_FREE (new_irr); | |
4674 | ||
0a35513e | 4675 | pop_cfun (); |
0a35513e AH |
4676 | |
4677 | return ret; | |
4678 | } | |
4679 | ||
4680 | /* Return true if, for the transactional clone of NODE, any call | |
4681 | may enter irrevocable mode. */ | |
4682 | ||
4683 | static bool | |
4684 | ipa_tm_mayenterirr_function (struct cgraph_node *node) | |
4685 | { | |
594ec92f AH |
4686 | struct tm_ipa_cg_data *d; |
4687 | tree decl; | |
4688 | unsigned flags; | |
4689 | ||
4690 | d = get_cg_data (&node, true); | |
67348ccc | 4691 | decl = node->decl; |
594ec92f | 4692 | flags = flags_from_decl_or_type (decl); |
0a35513e AH |
4693 | |
4694 | /* Handle some TM builtins. Ordinarily these aren't actually generated | |
4695 | at this point, but handling these functions when written in by the | |
4696 | user makes it easier to build unit tests. */ | |
4697 | if (flags & ECF_TM_BUILTIN) | |
4698 | return false; | |
4699 | ||
4700 | /* Filter out all functions that are marked. */ | |
4701 | if (flags & ECF_TM_PURE) | |
4702 | return false; | |
4703 | if (is_tm_safe (decl)) | |
4704 | return false; | |
4705 | if (is_tm_irrevocable (decl)) | |
4706 | return true; | |
4707 | if (is_tm_callable (decl)) | |
4708 | return true; | |
4709 | if (find_tm_replacement_function (decl)) | |
4710 | return true; | |
4711 | ||
4712 | /* If we aren't seeing the final version of the function we don't | |
4713 | know what it will contain at runtime. */ | |
d52f5295 | 4714 | if (node->get_availability () < AVAIL_AVAILABLE) |
0a35513e AH |
4715 | return true; |
4716 | ||
4717 | /* If the function must go irrevocable, then of course true. */ | |
4718 | if (d->is_irrevocable) | |
4719 | return true; | |
4720 | ||
4721 | /* If there are any blocks marked irrevocable, then the function | |
4722 | as a whole may enter irrevocable. */ | |
4723 | if (d->irrevocable_blocks_clone) | |
4724 | return true; | |
4725 | ||
4726 | /* We may have previously marked this function as tm_may_enter_irr; | |
4727 | see pass_diagnose_tm_blocks. */ | |
87f94429 | 4728 | if (node->tm_may_enter_irr) |
0a35513e AH |
4729 | return true; |
4730 | ||
4731 | /* Recurse on the main body for aliases. In general, this will | |
4732 | result in one of the bits above being set so that we will not | |
4733 | have to recurse next time. */ | |
67348ccc | 4734 | if (node->alias) |
67f3791f JH |
4735 | return ipa_tm_mayenterirr_function |
4736 | (cgraph_node::get (thunk_info::get (node)->alias)); | |
0a35513e AH |
4737 | |
4738 | /* What remains is unmarked local functions without items that force | |
4739 | the function to go irrevocable. */ | |
4740 | return false; | |
4741 | } | |
4742 | ||
4743 | /* Diagnose calls from transaction_safe functions to unmarked | |
4744 | functions that are determined to not be safe. */ | |
4745 | ||
4746 | static void | |
4747 | ipa_tm_diagnose_tm_safe (struct cgraph_node *node) | |
4748 | { | |
4749 | struct cgraph_edge *e; | |
4750 | ||
4751 | for (e = node->callees; e ; e = e->next_callee) | |
67348ccc | 4752 | if (!is_tm_callable (e->callee->decl) |
87f94429 | 4753 | && e->callee->tm_may_enter_irr) |
0a35513e AH |
4754 | error_at (gimple_location (e->call_stmt), |
4755 | "unsafe function call %qD within " | |
67348ccc | 4756 | "%<transaction_safe%> function", e->callee->decl); |
0a35513e AH |
4757 | } |
4758 | ||
4759 | /* Diagnose call from atomic transactions to unmarked functions | |
4760 | that are determined to not be safe. */ | |
4761 | ||
4762 | static void | |
4763 | ipa_tm_diagnose_transaction (struct cgraph_node *node, | |
4764 | struct tm_region *all_tm_regions) | |
4765 | { | |
4766 | struct tm_region *r; | |
4767 | ||
4768 | for (r = all_tm_regions; r ; r = r->next) | |
538dd0b7 DM |
4769 | if (gimple_transaction_subcode (r->get_transaction_stmt ()) |
4770 | & GTMA_IS_RELAXED) | |
0a35513e AH |
4771 | { |
4772 | /* Atomic transactions can be nested inside relaxed. */ | |
4773 | if (r->inner) | |
4774 | ipa_tm_diagnose_transaction (node, r->inner); | |
4775 | } | |
4776 | else | |
4777 | { | |
9771b263 | 4778 | vec<basic_block> bbs; |
0a35513e AH |
4779 | gimple_stmt_iterator gsi; |
4780 | basic_block bb; | |
4781 | size_t i; | |
4782 | ||
4783 | bbs = get_tm_region_blocks (r->entry_block, r->exit_blocks, | |
4784 | r->irr_blocks, NULL, false); | |
4785 | ||
9771b263 | 4786 | for (i = 0; bbs.iterate (i, &bb); ++i) |
0a35513e AH |
4787 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
4788 | { | |
355fe088 | 4789 | gimple *stmt = gsi_stmt (gsi); |
0a35513e AH |
4790 | tree fndecl; |
4791 | ||
4792 | if (gimple_code (stmt) == GIMPLE_ASM) | |
4793 | { | |
4794 | error_at (gimple_location (stmt), | |
a9c697b8 | 4795 | "%<asm%> not allowed in atomic transaction"); |
0a35513e AH |
4796 | continue; |
4797 | } | |
4798 | ||
4799 | if (!is_gimple_call (stmt)) | |
4800 | continue; | |
4801 | fndecl = gimple_call_fndecl (stmt); | |
4802 | ||
4803 | /* Indirect function calls have been diagnosed already. */ | |
4804 | if (!fndecl) | |
4805 | continue; | |
4806 | ||
4807 | /* Stop at the end of the transaction. */ | |
4808 | if (is_tm_ending_fndecl (fndecl)) | |
4809 | { | |
4810 | if (bitmap_bit_p (r->exit_blocks, bb->index)) | |
4811 | break; | |
4812 | continue; | |
4813 | } | |
4814 | ||
4815 | /* Marked functions have been diagnosed already. */ | |
4816 | if (is_tm_pure_call (stmt)) | |
4817 | continue; | |
4818 | if (is_tm_callable (fndecl)) | |
4819 | continue; | |
4820 | ||
87f94429 | 4821 | if (cgraph_node::local_info_node (fndecl)->tm_may_enter_irr) |
0a35513e AH |
4822 | error_at (gimple_location (stmt), |
4823 | "unsafe function call %qD within " | |
4824 | "atomic transaction", fndecl); | |
4825 | } | |
4826 | ||
9771b263 | 4827 | bbs.release (); |
0a35513e AH |
4828 | } |
4829 | } | |
4830 | ||
4831 | /* Return a transactional mangled name for the DECL_ASSEMBLER_NAME in | |
4832 | OLD_DECL. The returned value is a freshly malloced pointer that | |
4833 | should be freed by the caller. */ | |
4834 | ||
4835 | static tree | |
4836 | tm_mangle (tree old_asm_id) | |
4837 | { | |
4838 | const char *old_asm_name; | |
4839 | char *tm_name; | |
4840 | void *alloc = NULL; | |
4841 | struct demangle_component *dc; | |
4842 | tree new_asm_id; | |
4843 | ||
4844 | /* Determine if the symbol is already a valid C++ mangled name. Do this | |
4845 | even for C, which might be interfacing with C++ code via appropriately | |
4846 | ugly identifiers. */ | |
4847 | /* ??? We could probably do just as well checking for "_Z" and be done. */ | |
4848 | old_asm_name = IDENTIFIER_POINTER (old_asm_id); | |
4849 | dc = cplus_demangle_v3_components (old_asm_name, DMGL_NO_OPTS, &alloc); | |
4850 | ||
4851 | if (dc == NULL) | |
4852 | { | |
40ce7fa6 | 4853 | char length[12]; |
0a35513e AH |
4854 | |
4855 | do_unencoded: | |
4856 | sprintf (length, "%u", IDENTIFIER_LENGTH (old_asm_id)); | |
4857 | tm_name = concat ("_ZGTt", length, old_asm_name, NULL); | |
4858 | } | |
4859 | else | |
4860 | { | |
4861 | old_asm_name += 2; /* Skip _Z */ | |
4862 | ||
4863 | switch (dc->type) | |
4864 | { | |
4865 | case DEMANGLE_COMPONENT_TRANSACTION_CLONE: | |
4866 | case DEMANGLE_COMPONENT_NONTRANSACTION_CLONE: | |
4867 | /* Don't play silly games, you! */ | |
4868 | goto do_unencoded; | |
4869 | ||
4870 | case DEMANGLE_COMPONENT_HIDDEN_ALIAS: | |
4871 | /* I'd really like to know if we can ever be passed one of | |
4872 | these from the C++ front end. The Logical Thing would | |
4873 | seem that hidden-alias should be outer-most, so that we | |
4874 | get hidden-alias of a transaction-clone and not vice-versa. */ | |
4875 | old_asm_name += 2; | |
4876 | break; | |
4877 | ||
4878 | default: | |
4879 | break; | |
4880 | } | |
4881 | ||
4882 | tm_name = concat ("_ZGTt", old_asm_name, NULL); | |
4883 | } | |
4884 | free (alloc); | |
4885 | ||
4886 | new_asm_id = get_identifier (tm_name); | |
4887 | free (tm_name); | |
4888 | ||
4889 | return new_asm_id; | |
4890 | } | |
4891 | ||
4892 | static inline void | |
ead84f73 | 4893 | ipa_tm_mark_force_output_node (struct cgraph_node *node) |
0a35513e | 4894 | { |
d52f5295 | 4895 | node->mark_force_output (); |
67348ccc | 4896 | node->analyzed = true; |
0a35513e AH |
4897 | } |
4898 | ||
edb983b2 JH |
4899 | static inline void |
4900 | ipa_tm_mark_forced_by_abi_node (struct cgraph_node *node) | |
4901 | { | |
67348ccc DM |
4902 | node->forced_by_abi = true; |
4903 | node->analyzed = true; | |
edb983b2 JH |
4904 | } |
4905 | ||
0a35513e AH |
4906 | /* Callback data for ipa_tm_create_version_alias. */ |
4907 | struct create_version_alias_info | |
4908 | { | |
4909 | struct cgraph_node *old_node; | |
4910 | tree new_decl; | |
4911 | }; | |
4912 | ||
e617b554 | 4913 | /* A subroutine of ipa_tm_create_version, called via |
0a35513e AH |
4914 | cgraph_for_node_and_aliases. Create new tm clones for each of |
4915 | the existing aliases. */ | |
4916 | static bool | |
4917 | ipa_tm_create_version_alias (struct cgraph_node *node, void *data) | |
4918 | { | |
4919 | struct create_version_alias_info *info | |
4920 | = (struct create_version_alias_info *)data; | |
4921 | tree old_decl, new_decl, tm_name; | |
4922 | struct cgraph_node *new_node; | |
4923 | ||
67348ccc | 4924 | if (!node->cpp_implicit_alias) |
0a35513e AH |
4925 | return false; |
4926 | ||
67348ccc | 4927 | old_decl = node->decl; |
0a35513e AH |
4928 | tm_name = tm_mangle (DECL_ASSEMBLER_NAME (old_decl)); |
4929 | new_decl = build_decl (DECL_SOURCE_LOCATION (old_decl), | |
4930 | TREE_CODE (old_decl), tm_name, | |
4931 | TREE_TYPE (old_decl)); | |
4932 | ||
4933 | SET_DECL_ASSEMBLER_NAME (new_decl, tm_name); | |
4934 | SET_DECL_RTL (new_decl, NULL); | |
4935 | ||
4936 | /* Based loosely on C++'s make_alias_for(). */ | |
4937 | TREE_PUBLIC (new_decl) = TREE_PUBLIC (old_decl); | |
54c39478 AH |
4938 | DECL_CONTEXT (new_decl) = DECL_CONTEXT (old_decl); |
4939 | DECL_LANG_SPECIFIC (new_decl) = DECL_LANG_SPECIFIC (old_decl); | |
0a35513e AH |
4940 | TREE_READONLY (new_decl) = TREE_READONLY (old_decl); |
4941 | DECL_EXTERNAL (new_decl) = 0; | |
4942 | DECL_ARTIFICIAL (new_decl) = 1; | |
4943 | TREE_ADDRESSABLE (new_decl) = 1; | |
4944 | TREE_USED (new_decl) = 1; | |
4945 | TREE_SYMBOL_REFERENCED (tm_name) = 1; | |
4946 | ||
4947 | /* Perform the same remapping to the comdat group. */ | |
43042ea7 | 4948 | if (DECL_ONE_ONLY (new_decl)) |
9041d2e6 ML |
4949 | varpool_node::get (new_decl)->set_comdat_group |
4950 | (tm_mangle (decl_comdat_group_id (old_decl))); | |
0a35513e | 4951 | |
d52f5295 | 4952 | new_node = cgraph_node::create_same_body_alias (new_decl, info->new_decl); |
0a35513e | 4953 | new_node->tm_clone = true; |
67348ccc | 4954 | new_node->externally_visible = info->old_node->externally_visible; |
7861b648 | 4955 | new_node->no_reorder = info->old_node->no_reorder; |
594ec92f AH |
4956 | /* ?? Do not traverse aliases here. */ |
4957 | get_cg_data (&node, false)->clone = new_node; | |
0a35513e AH |
4958 | |
4959 | record_tm_clone_pair (old_decl, new_decl); | |
4960 | ||
67348ccc | 4961 | if (info->old_node->force_output |
d122681a | 4962 | || info->old_node->ref_list.first_referring ()) |
ead84f73 | 4963 | ipa_tm_mark_force_output_node (new_node); |
67348ccc | 4964 | if (info->old_node->forced_by_abi) |
edb983b2 | 4965 | ipa_tm_mark_forced_by_abi_node (new_node); |
0a35513e AH |
4966 | return false; |
4967 | } | |
4968 | ||
4969 | /* Create a copy of the function (possibly declaration only) of OLD_NODE, | |
4970 | appropriate for the transactional clone. */ | |
4971 | ||
4972 | static void | |
4973 | ipa_tm_create_version (struct cgraph_node *old_node) | |
4974 | { | |
4975 | tree new_decl, old_decl, tm_name; | |
4976 | struct cgraph_node *new_node; | |
4977 | ||
67348ccc | 4978 | old_decl = old_node->decl; |
0a35513e AH |
4979 | new_decl = copy_node (old_decl); |
4980 | ||
4981 | /* DECL_ASSEMBLER_NAME needs to be set before we call | |
4982 | cgraph_copy_node_for_versioning below, because cgraph_node will | |
4983 | fill the assembler_name_hash. */ | |
4984 | tm_name = tm_mangle (DECL_ASSEMBLER_NAME (old_decl)); | |
4985 | SET_DECL_ASSEMBLER_NAME (new_decl, tm_name); | |
4986 | SET_DECL_RTL (new_decl, NULL); | |
4987 | TREE_SYMBOL_REFERENCED (tm_name) = 1; | |
4988 | ||
4989 | /* Perform the same remapping to the comdat group. */ | |
43042ea7 | 4990 | if (DECL_ONE_ONLY (new_decl)) |
9041d2e6 ML |
4991 | varpool_node::get (new_decl)->set_comdat_group |
4992 | (tm_mangle (DECL_COMDAT_GROUP (old_decl))); | |
0a35513e | 4993 | |
ca860d03 | 4994 | gcc_assert (!old_node->ipa_transforms_to_apply.exists ()); |
d52f5295 | 4995 | new_node = old_node->create_version_clone (new_decl, vNULL, NULL); |
87f94429 | 4996 | new_node->local = false; |
67348ccc | 4997 | new_node->externally_visible = old_node->externally_visible; |
0a35513e AH |
4998 | new_node->lowered = true; |
4999 | new_node->tm_clone = 1; | |
4ab26ee0 | 5000 | if (!old_node->implicit_section) |
8b9a92f7 | 5001 | new_node->set_section (*old_node); |
594ec92f | 5002 | get_cg_data (&old_node, true)->clone = new_node; |
0a35513e | 5003 | |
d52f5295 | 5004 | if (old_node->get_availability () >= AVAIL_INTERPOSABLE) |
0a35513e AH |
5005 | { |
5006 | /* Remap extern inline to static inline. */ | |
5007 | /* ??? Is it worth trying to use make_decl_one_only? */ | |
5008 | if (DECL_DECLARED_INLINE_P (new_decl) && DECL_EXTERNAL (new_decl)) | |
5009 | { | |
5010 | DECL_EXTERNAL (new_decl) = 0; | |
5011 | TREE_PUBLIC (new_decl) = 0; | |
e0b22991 | 5012 | DECL_WEAK (new_decl) = 0; |
0a35513e AH |
5013 | } |
5014 | ||
9771b263 | 5015 | tree_function_versioning (old_decl, new_decl, |
ff6686d2 | 5016 | NULL, NULL, false, NULL, NULL); |
0a35513e AH |
5017 | } |
5018 | ||
5019 | record_tm_clone_pair (old_decl, new_decl); | |
5020 | ||
3dafb85c | 5021 | symtab->call_cgraph_insertion_hooks (new_node); |
67348ccc | 5022 | if (old_node->force_output |
d122681a | 5023 | || old_node->ref_list.first_referring ()) |
ead84f73 | 5024 | ipa_tm_mark_force_output_node (new_node); |
67348ccc | 5025 | if (old_node->forced_by_abi) |
edb983b2 | 5026 | ipa_tm_mark_forced_by_abi_node (new_node); |
0a35513e AH |
5027 | |
5028 | /* Do the same thing, but for any aliases of the original node. */ | |
5029 | { | |
5030 | struct create_version_alias_info data; | |
5031 | data.old_node = old_node; | |
5032 | data.new_decl = new_decl; | |
d52f5295 ML |
5033 | old_node->call_for_symbol_thunks_and_aliases (ipa_tm_create_version_alias, |
5034 | &data, true); | |
0a35513e AH |
5035 | } |
5036 | } | |
5037 | ||
5038 | /* Construct a call to TM_IRREVOCABLE and insert it at the beginning of BB. */ | |
5039 | ||
5040 | static void | |
5041 | ipa_tm_insert_irr_call (struct cgraph_node *node, struct tm_region *region, | |
5042 | basic_block bb) | |
5043 | { | |
5044 | gimple_stmt_iterator gsi; | |
538dd0b7 | 5045 | gcall *g; |
0a35513e AH |
5046 | |
5047 | transaction_subcode_ior (region, GTMA_MAY_ENTER_IRREVOCABLE); | |
5048 | ||
5049 | g = gimple_build_call (builtin_decl_explicit (BUILT_IN_TM_IRREVOCABLE), | |
5050 | 1, build_int_cst (NULL_TREE, MODE_SERIALIRREVOCABLE)); | |
5051 | ||
5052 | split_block_after_labels (bb); | |
5053 | gsi = gsi_after_labels (bb); | |
5054 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
5055 | ||
d52f5295 ML |
5056 | node->create_edge (cgraph_node::get_create |
5057 | (builtin_decl_explicit (BUILT_IN_TM_IRREVOCABLE)), | |
1bad9c18 | 5058 | g, gimple_bb (g)->count); |
0a35513e AH |
5059 | } |
5060 | ||
5061 | /* Construct a call to TM_GETTMCLONE and insert it before GSI. */ | |
5062 | ||
5063 | static bool | |
5064 | ipa_tm_insert_gettmclone_call (struct cgraph_node *node, | |
5065 | struct tm_region *region, | |
538dd0b7 | 5066 | gimple_stmt_iterator *gsi, gcall *stmt) |
0a35513e AH |
5067 | { |
5068 | tree gettm_fn, ret, old_fn, callfn; | |
538dd0b7 DM |
5069 | gcall *g; |
5070 | gassign *g2; | |
0a35513e AH |
5071 | bool safe; |
5072 | ||
5073 | old_fn = gimple_call_fn (stmt); | |
5074 | ||
5075 | if (TREE_CODE (old_fn) == ADDR_EXPR) | |
5076 | { | |
5077 | tree fndecl = TREE_OPERAND (old_fn, 0); | |
5078 | tree clone = get_tm_clone_pair (fndecl); | |
5079 | ||
5080 | /* By transforming the call into a TM_GETTMCLONE, we are | |
5081 | technically taking the address of the original function and | |
5082 | its clone. Explain this so inlining will know this function | |
5083 | is needed. */ | |
d52f5295 | 5084 | cgraph_node::get (fndecl)->mark_address_taken () ; |
0a35513e | 5085 | if (clone) |
d52f5295 | 5086 | cgraph_node::get (clone)->mark_address_taken (); |
0a35513e AH |
5087 | } |
5088 | ||
5089 | safe = is_tm_safe (TREE_TYPE (old_fn)); | |
5090 | gettm_fn = builtin_decl_explicit (safe ? BUILT_IN_TM_GETTMCLONE_SAFE | |
5091 | : BUILT_IN_TM_GETTMCLONE_IRR); | |
b731b390 | 5092 | ret = create_tmp_var (ptr_type_node); |
0a35513e AH |
5093 | |
5094 | if (!safe) | |
5095 | transaction_subcode_ior (region, GTMA_MAY_ENTER_IRREVOCABLE); | |
5096 | ||
5097 | /* Discard OBJ_TYPE_REF, since we weren't able to fold it. */ | |
5098 | if (TREE_CODE (old_fn) == OBJ_TYPE_REF) | |
5099 | old_fn = OBJ_TYPE_REF_EXPR (old_fn); | |
5100 | ||
5101 | g = gimple_build_call (gettm_fn, 1, old_fn); | |
5102 | ret = make_ssa_name (ret, g); | |
5103 | gimple_call_set_lhs (g, ret); | |
5104 | ||
5105 | gsi_insert_before (gsi, g, GSI_SAME_STMT); | |
5106 | ||
1bad9c18 | 5107 | node->create_edge (cgraph_node::get_create (gettm_fn), g, gimple_bb (g)->count); |
0a35513e AH |
5108 | |
5109 | /* Cast return value from tm_gettmclone* into appropriate function | |
5110 | pointer. */ | |
b731b390 | 5111 | callfn = create_tmp_var (TREE_TYPE (old_fn)); |
0a35513e AH |
5112 | g2 = gimple_build_assign (callfn, |
5113 | fold_build1 (NOP_EXPR, TREE_TYPE (callfn), ret)); | |
5114 | callfn = make_ssa_name (callfn, g2); | |
5115 | gimple_assign_set_lhs (g2, callfn); | |
5116 | gsi_insert_before (gsi, g2, GSI_SAME_STMT); | |
5117 | ||
5118 | /* ??? This is a hack to preserve the NOTHROW bit on the call, | |
5119 | which we would have derived from the decl. Failure to save | |
5120 | this bit means we might have to split the basic block. */ | |
5121 | if (gimple_call_nothrow_p (stmt)) | |
5122 | gimple_call_set_nothrow (stmt, true); | |
5123 | ||
5124 | gimple_call_set_fn (stmt, callfn); | |
5125 | ||
5126 | /* Discarding OBJ_TYPE_REF above may produce incompatible LHS and RHS | |
5127 | for a call statement. Fix it. */ | |
5128 | { | |
5129 | tree lhs = gimple_call_lhs (stmt); | |
5130 | tree rettype = TREE_TYPE (gimple_call_fntype (stmt)); | |
5131 | if (lhs | |
5132 | && !useless_type_conversion_p (TREE_TYPE (lhs), rettype)) | |
5133 | { | |
5134 | tree temp; | |
5135 | ||
b731b390 | 5136 | temp = create_tmp_reg (rettype); |
0a35513e AH |
5137 | gimple_call_set_lhs (stmt, temp); |
5138 | ||
5139 | g2 = gimple_build_assign (lhs, | |
5140 | fold_build1 (VIEW_CONVERT_EXPR, | |
5141 | TREE_TYPE (lhs), temp)); | |
5142 | gsi_insert_after (gsi, g2, GSI_SAME_STMT); | |
5143 | } | |
5144 | } | |
5145 | ||
5146 | update_stmt (stmt); | |
7d0aa05b JH |
5147 | cgraph_edge *e = cgraph_node::get (current_function_decl)->get_edge (stmt); |
5148 | if (e && e->indirect_info) | |
5149 | e->indirect_info->polymorphic = false; | |
0a35513e AH |
5150 | |
5151 | return true; | |
5152 | } | |
5153 | ||
5154 | /* Helper function for ipa_tm_transform_calls*. Given a call | |
5155 | statement in GSI which resides inside transaction REGION, redirect | |
5156 | the call to either its wrapper function, or its clone. */ | |
5157 | ||
5158 | static void | |
5159 | ipa_tm_transform_calls_redirect (struct cgraph_node *node, | |
5160 | struct tm_region *region, | |
5161 | gimple_stmt_iterator *gsi, | |
5162 | bool *need_ssa_rename_p) | |
5163 | { | |
538dd0b7 | 5164 | gcall *stmt = as_a <gcall *> (gsi_stmt (*gsi)); |
0a35513e | 5165 | struct cgraph_node *new_node; |
d52f5295 | 5166 | struct cgraph_edge *e = node->get_edge (stmt); |
0a35513e AH |
5167 | tree fndecl = gimple_call_fndecl (stmt); |
5168 | ||
5169 | /* For indirect calls, pass the address through the runtime. */ | |
5170 | if (fndecl == NULL) | |
5171 | { | |
5172 | *need_ssa_rename_p |= | |
5173 | ipa_tm_insert_gettmclone_call (node, region, gsi, stmt); | |
5174 | return; | |
5175 | } | |
5176 | ||
5177 | /* Handle some TM builtins. Ordinarily these aren't actually generated | |
5178 | at this point, but handling these functions when written in by the | |
5179 | user makes it easier to build unit tests. */ | |
5180 | if (flags_from_decl_or_type (fndecl) & ECF_TM_BUILTIN) | |
5181 | return; | |
5182 | ||
5183 | /* Fixup recursive calls inside clones. */ | |
5184 | /* ??? Why did cgraph_copy_node_for_versioning update the call edges | |
5185 | for recursion but not update the call statements themselves? */ | |
5186 | if (e->caller == e->callee && decl_is_tm_clone (current_function_decl)) | |
5187 | { | |
5188 | gimple_call_set_fndecl (stmt, current_function_decl); | |
5189 | return; | |
5190 | } | |
5191 | ||
5192 | /* If there is a replacement, use it. */ | |
5193 | fndecl = find_tm_replacement_function (fndecl); | |
5194 | if (fndecl) | |
5195 | { | |
d52f5295 | 5196 | new_node = cgraph_node::get_create (fndecl); |
0a35513e AH |
5197 | |
5198 | /* ??? Mark all transaction_wrap functions tm_may_enter_irr. | |
5199 | ||
5200 | We can't do this earlier in record_tm_replacement because | |
5201 | cgraph_remove_unreachable_nodes is called before we inject | |
5202 | references to the node. Further, we can't do this in some | |
5203 | nice central place in ipa_tm_execute because we don't have | |
5204 | the exact list of wrapper functions that would be used. | |
5205 | Marking more wrappers than necessary results in the creation | |
5206 | of unnecessary cgraph_nodes, which can cause some of the | |
5207 | other IPA passes to crash. | |
5208 | ||
5209 | We do need to mark these nodes so that we get the proper | |
5210 | result in expand_call_tm. */ | |
5211 | /* ??? This seems broken. How is it that we're marking the | |
5212 | CALLEE as may_enter_irr? Surely we should be marking the | |
5213 | CALLER. Also note that find_tm_replacement_function also | |
5214 | contains mappings into the TM runtime, e.g. memcpy. These | |
5215 | we know won't go irrevocable. */ | |
87f94429 | 5216 | new_node->tm_may_enter_irr = 1; |
0a35513e AH |
5217 | } |
5218 | else | |
5219 | { | |
594ec92f AH |
5220 | struct tm_ipa_cg_data *d; |
5221 | struct cgraph_node *tnode = e->callee; | |
5222 | ||
5223 | d = get_cg_data (&tnode, true); | |
0a35513e AH |
5224 | new_node = d->clone; |
5225 | ||
5226 | /* As we've already skipped pure calls and appropriate builtins, | |
5227 | and we've already marked irrevocable blocks, if we can't come | |
5228 | up with a static replacement, then ask the runtime. */ | |
5229 | if (new_node == NULL) | |
5230 | { | |
5231 | *need_ssa_rename_p |= | |
5232 | ipa_tm_insert_gettmclone_call (node, region, gsi, stmt); | |
0a35513e AH |
5233 | return; |
5234 | } | |
5235 | ||
67348ccc | 5236 | fndecl = new_node->decl; |
0a35513e AH |
5237 | } |
5238 | ||
3dafb85c | 5239 | e->redirect_callee (new_node); |
0a35513e AH |
5240 | gimple_call_set_fndecl (stmt, fndecl); |
5241 | } | |
5242 | ||
5243 | /* Helper function for ipa_tm_transform_calls. For a given BB, | |
5244 | install calls to tm_irrevocable when IRR_BLOCKS are reached, | |
5245 | redirect other calls to the generated transactional clone. */ | |
5246 | ||
5247 | static bool | |
5248 | ipa_tm_transform_calls_1 (struct cgraph_node *node, struct tm_region *region, | |
5249 | basic_block bb, bitmap irr_blocks) | |
5250 | { | |
5251 | gimple_stmt_iterator gsi; | |
5252 | bool need_ssa_rename = false; | |
5253 | ||
5254 | if (irr_blocks && bitmap_bit_p (irr_blocks, bb->index)) | |
5255 | { | |
5256 | ipa_tm_insert_irr_call (node, region, bb); | |
5257 | return true; | |
5258 | } | |
5259 | ||
5260 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
5261 | { | |
355fe088 | 5262 | gimple *stmt = gsi_stmt (gsi); |
0a35513e AH |
5263 | |
5264 | if (!is_gimple_call (stmt)) | |
5265 | continue; | |
5266 | if (is_tm_pure_call (stmt)) | |
5267 | continue; | |
5268 | ||
5269 | /* Redirect edges to the appropriate replacement or clone. */ | |
5270 | ipa_tm_transform_calls_redirect (node, region, &gsi, &need_ssa_rename); | |
5271 | } | |
5272 | ||
5273 | return need_ssa_rename; | |
5274 | } | |
5275 | ||
5276 | /* Walk the CFG for REGION, beginning at BB. Install calls to | |
5277 | tm_irrevocable when IRR_BLOCKS are reached, redirect other calls to | |
5278 | the generated transactional clone. */ | |
5279 | ||
5280 | static bool | |
5281 | ipa_tm_transform_calls (struct cgraph_node *node, struct tm_region *region, | |
5282 | basic_block bb, bitmap irr_blocks) | |
5283 | { | |
5284 | bool need_ssa_rename = false; | |
5285 | edge e; | |
5286 | edge_iterator ei; | |
ef062b13 | 5287 | auto_vec<basic_block> queue; |
0a35513e AH |
5288 | bitmap visited_blocks = BITMAP_ALLOC (NULL); |
5289 | ||
9771b263 | 5290 | queue.safe_push (bb); |
0a35513e AH |
5291 | do |
5292 | { | |
9771b263 | 5293 | bb = queue.pop (); |
0a35513e AH |
5294 | |
5295 | need_ssa_rename |= | |
5296 | ipa_tm_transform_calls_1 (node, region, bb, irr_blocks); | |
5297 | ||
5298 | if (irr_blocks && bitmap_bit_p (irr_blocks, bb->index)) | |
5299 | continue; | |
5300 | ||
5301 | if (region && bitmap_bit_p (region->exit_blocks, bb->index)) | |
5302 | continue; | |
5303 | ||
5304 | FOR_EACH_EDGE (e, ei, bb->succs) | |
5305 | if (!bitmap_bit_p (visited_blocks, e->dest->index)) | |
5306 | { | |
5307 | bitmap_set_bit (visited_blocks, e->dest->index); | |
9771b263 | 5308 | queue.safe_push (e->dest); |
0a35513e AH |
5309 | } |
5310 | } | |
9771b263 | 5311 | while (!queue.is_empty ()); |
0a35513e | 5312 | |
0a35513e AH |
5313 | BITMAP_FREE (visited_blocks); |
5314 | ||
5315 | return need_ssa_rename; | |
5316 | } | |
5317 | ||
5318 | /* Transform the calls within the TM regions within NODE. */ | |
5319 | ||
5320 | static void | |
5321 | ipa_tm_transform_transaction (struct cgraph_node *node) | |
5322 | { | |
594ec92f | 5323 | struct tm_ipa_cg_data *d; |
0a35513e AH |
5324 | struct tm_region *region; |
5325 | bool need_ssa_rename = false; | |
5326 | ||
594ec92f AH |
5327 | d = get_cg_data (&node, true); |
5328 | ||
67348ccc | 5329 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); |
0a35513e AH |
5330 | calculate_dominance_info (CDI_DOMINATORS); |
5331 | ||
5332 | for (region = d->all_tm_regions; region; region = region->next) | |
5333 | { | |
5334 | /* If we're sure to go irrevocable, don't transform anything. */ | |
5335 | if (d->irrevocable_blocks_normal | |
5336 | && bitmap_bit_p (d->irrevocable_blocks_normal, | |
5337 | region->entry_block->index)) | |
5338 | { | |
b7a78683 AH |
5339 | transaction_subcode_ior (region, GTMA_DOES_GO_IRREVOCABLE |
5340 | | GTMA_MAY_ENTER_IRREVOCABLE | |
5341 | | GTMA_HAS_NO_INSTRUMENTATION); | |
0a35513e AH |
5342 | continue; |
5343 | } | |
5344 | ||
5345 | need_ssa_rename |= | |
5346 | ipa_tm_transform_calls (node, region, region->entry_block, | |
5347 | d->irrevocable_blocks_normal); | |
5348 | } | |
5349 | ||
5350 | if (need_ssa_rename) | |
5351 | update_ssa (TODO_update_ssa_only_virtuals); | |
5352 | ||
5353 | pop_cfun (); | |
0a35513e AH |
5354 | } |
5355 | ||
5356 | /* Transform the calls within the transactional clone of NODE. */ | |
5357 | ||
5358 | static void | |
5359 | ipa_tm_transform_clone (struct cgraph_node *node) | |
5360 | { | |
594ec92f | 5361 | struct tm_ipa_cg_data *d; |
0a35513e AH |
5362 | bool need_ssa_rename; |
5363 | ||
594ec92f AH |
5364 | d = get_cg_data (&node, true); |
5365 | ||
0a35513e AH |
5366 | /* If this function makes no calls and has no irrevocable blocks, |
5367 | then there's nothing to do. */ | |
5368 | /* ??? Remove non-aborting top-level transactions. */ | |
8730965e | 5369 | if (!node->callees && !node->indirect_calls && !d->irrevocable_blocks_clone) |
0a35513e AH |
5370 | return; |
5371 | ||
67348ccc | 5372 | push_cfun (DECL_STRUCT_FUNCTION (d->clone->decl)); |
0a35513e AH |
5373 | calculate_dominance_info (CDI_DOMINATORS); |
5374 | ||
5375 | need_ssa_rename = | |
fefa31b5 DM |
5376 | ipa_tm_transform_calls (d->clone, NULL, |
5377 | single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)), | |
0a35513e AH |
5378 | d->irrevocable_blocks_clone); |
5379 | ||
5380 | if (need_ssa_rename) | |
5381 | update_ssa (TODO_update_ssa_only_virtuals); | |
5382 | ||
5383 | pop_cfun (); | |
0a35513e AH |
5384 | } |
5385 | ||
5386 | /* Main entry point for the transactional memory IPA pass. */ | |
5387 | ||
5388 | static unsigned int | |
5389 | ipa_tm_execute (void) | |
5390 | { | |
c3284718 | 5391 | cgraph_node_queue tm_callees = cgraph_node_queue (); |
0a35513e | 5392 | /* List of functions that will go irrevocable. */ |
c3284718 | 5393 | cgraph_node_queue irr_worklist = cgraph_node_queue (); |
0a35513e AH |
5394 | |
5395 | struct cgraph_node *node; | |
5396 | struct tm_ipa_cg_data *d; | |
5397 | enum availability a; | |
5398 | unsigned int i; | |
5399 | ||
b2b29377 | 5400 | cgraph_node::checking_verify_cgraph_nodes (); |
0a35513e AH |
5401 | |
5402 | bitmap_obstack_initialize (&tm_obstack); | |
398b1daa | 5403 | initialize_original_copy_tables (); |
0a35513e AH |
5404 | |
5405 | /* For all local functions marked tm_callable, queue them. */ | |
65c70e6b | 5406 | FOR_EACH_DEFINED_FUNCTION (node) |
67348ccc | 5407 | if (is_tm_callable (node->decl) |
d52f5295 | 5408 | && node->get_availability () >= AVAIL_INTERPOSABLE) |
0a35513e | 5409 | { |
594ec92f | 5410 | d = get_cg_data (&node, true); |
0a35513e AH |
5411 | maybe_push_queue (node, &tm_callees, &d->in_callee_queue); |
5412 | } | |
5413 | ||
5414 | /* For all local reachable functions... */ | |
65c70e6b | 5415 | FOR_EACH_DEFINED_FUNCTION (node) |
93a18a70 | 5416 | if (node->lowered |
d52f5295 | 5417 | && node->get_availability () >= AVAIL_INTERPOSABLE) |
0a35513e AH |
5418 | { |
5419 | /* ... marked tm_pure, record that fact for the runtime by | |
5420 | indicating that the pure function is its own tm_callable. | |
5421 | No need to do this if the function's address can't be taken. */ | |
67348ccc | 5422 | if (is_tm_pure (node->decl)) |
0a35513e | 5423 | { |
87f94429 | 5424 | if (!node->local) |
67348ccc | 5425 | record_tm_clone_pair (node->decl, node->decl); |
0a35513e AH |
5426 | continue; |
5427 | } | |
5428 | ||
67348ccc | 5429 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); |
0a35513e AH |
5430 | calculate_dominance_info (CDI_DOMINATORS); |
5431 | ||
5432 | tm_region_init (NULL); | |
5433 | if (all_tm_regions) | |
5434 | { | |
594ec92f | 5435 | d = get_cg_data (&node, true); |
0a35513e | 5436 | |
398b1daa AH |
5437 | /* Scan for calls that are in each transaction, and |
5438 | generate the uninstrumented code path. */ | |
0a35513e AH |
5439 | ipa_tm_scan_calls_transaction (d, &tm_callees); |
5440 | ||
80fd8eba AH |
5441 | /* Put it in the worklist so we can scan the function |
5442 | later (ipa_tm_scan_irr_function) and mark the | |
5443 | irrevocable blocks. */ | |
5444 | maybe_push_queue (node, &irr_worklist, &d->in_worklist); | |
5445 | d->want_irr_scan_normal = true; | |
0a35513e AH |
5446 | } |
5447 | ||
5448 | pop_cfun (); | |
0a35513e AH |
5449 | } |
5450 | ||
5451 | /* For every local function on the callee list, scan as if we will be | |
5452 | creating a transactional clone, queueing all new functions we find | |
5453 | along the way. */ | |
9771b263 | 5454 | for (i = 0; i < tm_callees.length (); ++i) |
0a35513e | 5455 | { |
9771b263 | 5456 | node = tm_callees[i]; |
d52f5295 | 5457 | a = node->get_availability (); |
594ec92f | 5458 | d = get_cg_data (&node, true); |
0a35513e | 5459 | |
80fd8eba AH |
5460 | /* Put it in the worklist so we can scan the function later |
5461 | (ipa_tm_scan_irr_function) and mark the irrevocable | |
5462 | blocks. */ | |
5463 | maybe_push_queue (node, &irr_worklist, &d->in_worklist); | |
0a35513e AH |
5464 | |
5465 | /* Some callees cannot be arbitrarily cloned. These will always be | |
5466 | irrevocable. Mark these now, so that we need not scan them. */ | |
67348ccc | 5467 | if (is_tm_irrevocable (node->decl)) |
0a35513e AH |
5468 | ipa_tm_note_irrevocable (node, &irr_worklist); |
5469 | else if (a <= AVAIL_NOT_AVAILABLE | |
67348ccc | 5470 | && !is_tm_safe_or_pure (node->decl)) |
0a35513e | 5471 | ipa_tm_note_irrevocable (node, &irr_worklist); |
d52f5295 | 5472 | else if (a >= AVAIL_INTERPOSABLE) |
0a35513e | 5473 | { |
67348ccc | 5474 | if (!tree_versionable_function_p (node->decl)) |
0a35513e AH |
5475 | ipa_tm_note_irrevocable (node, &irr_worklist); |
5476 | else if (!d->is_irrevocable) | |
5477 | { | |
5478 | /* If this is an alias, make sure its base is queued as well. | |
5479 | we need not scan the callees now, as the base will do. */ | |
67348ccc | 5480 | if (node->alias) |
0a35513e | 5481 | { |
67f3791f | 5482 | node = cgraph_node::get (thunk_info::get (node)->alias); |
594ec92f | 5483 | d = get_cg_data (&node, true); |
0a35513e AH |
5484 | maybe_push_queue (node, &tm_callees, &d->in_callee_queue); |
5485 | continue; | |
5486 | } | |
5487 | ||
5488 | /* Add all nodes called by this function into | |
5489 | tm_callees as well. */ | |
5490 | ipa_tm_scan_calls_clone (node, &tm_callees); | |
5491 | } | |
5492 | } | |
5493 | } | |
5494 | ||
5495 | /* Iterate scans until no more work to be done. Prefer not to use | |
9771b263 | 5496 | vec::pop because the worklist tends to follow a breadth-first |
0a35513e AH |
5497 | search of the callgraph, which should allow convergance with a |
5498 | minimum number of scans. But we also don't want the worklist | |
5499 | array to grow without bound, so we shift the array up periodically. */ | |
9771b263 | 5500 | for (i = 0; i < irr_worklist.length (); ++i) |
0a35513e | 5501 | { |
9771b263 | 5502 | if (i > 256 && i == irr_worklist.length () / 8) |
0a35513e | 5503 | { |
9771b263 | 5504 | irr_worklist.block_remove (0, i); |
0a35513e AH |
5505 | i = 0; |
5506 | } | |
5507 | ||
9771b263 | 5508 | node = irr_worklist[i]; |
594ec92f | 5509 | d = get_cg_data (&node, true); |
0a35513e AH |
5510 | d->in_worklist = false; |
5511 | ||
5512 | if (d->want_irr_scan_normal) | |
5513 | { | |
5514 | d->want_irr_scan_normal = false; | |
5515 | ipa_tm_scan_irr_function (node, false); | |
5516 | } | |
5517 | if (d->in_callee_queue && ipa_tm_scan_irr_function (node, true)) | |
5518 | ipa_tm_note_irrevocable (node, &irr_worklist); | |
5519 | } | |
5520 | ||
5521 | /* For every function on the callee list, collect the tm_may_enter_irr | |
5522 | bit on the node. */ | |
9771b263 DN |
5523 | irr_worklist.truncate (0); |
5524 | for (i = 0; i < tm_callees.length (); ++i) | |
0a35513e | 5525 | { |
9771b263 | 5526 | node = tm_callees[i]; |
0a35513e AH |
5527 | if (ipa_tm_mayenterirr_function (node)) |
5528 | { | |
594ec92f | 5529 | d = get_cg_data (&node, true); |
0a35513e AH |
5530 | gcc_assert (d->in_worklist == false); |
5531 | maybe_push_queue (node, &irr_worklist, &d->in_worklist); | |
5532 | } | |
5533 | } | |
5534 | ||
5535 | /* Propagate the tm_may_enter_irr bit to callers until stable. */ | |
9771b263 | 5536 | for (i = 0; i < irr_worklist.length (); ++i) |
0a35513e AH |
5537 | { |
5538 | struct cgraph_node *caller; | |
5539 | struct cgraph_edge *e; | |
e55637b7 | 5540 | struct ipa_ref *ref; |
0a35513e | 5541 | |
9771b263 | 5542 | if (i > 256 && i == irr_worklist.length () / 8) |
0a35513e | 5543 | { |
9771b263 | 5544 | irr_worklist.block_remove (0, i); |
0a35513e AH |
5545 | i = 0; |
5546 | } | |
5547 | ||
9771b263 | 5548 | node = irr_worklist[i]; |
594ec92f | 5549 | d = get_cg_data (&node, true); |
0a35513e | 5550 | d->in_worklist = false; |
87f94429 | 5551 | node->tm_may_enter_irr = true; |
0a35513e AH |
5552 | |
5553 | /* Propagate back to normal callers. */ | |
5554 | for (e = node->callers; e ; e = e->next_caller) | |
5555 | { | |
5556 | caller = e->caller; | |
67348ccc | 5557 | if (!is_tm_safe_or_pure (caller->decl) |
87f94429 | 5558 | && !caller->tm_may_enter_irr) |
0a35513e | 5559 | { |
594ec92f | 5560 | d = get_cg_data (&caller, true); |
0a35513e AH |
5561 | maybe_push_queue (caller, &irr_worklist, &d->in_worklist); |
5562 | } | |
5563 | } | |
5564 | ||
5565 | /* Propagate back to referring aliases as well. */ | |
e55637b7 | 5566 | FOR_EACH_ALIAS (node, ref) |
0a35513e | 5567 | { |
d52f5295 | 5568 | caller = dyn_cast<cgraph_node *> (ref->referring); |
87f94429 | 5569 | if (!caller->tm_may_enter_irr) |
0a35513e | 5570 | { |
594ec92f AH |
5571 | /* ?? Do not traverse aliases here. */ |
5572 | d = get_cg_data (&caller, false); | |
0a35513e AH |
5573 | maybe_push_queue (caller, &irr_worklist, &d->in_worklist); |
5574 | } | |
5575 | } | |
5576 | } | |
5577 | ||
5578 | /* Now validate all tm_safe functions, and all atomic regions in | |
5579 | other functions. */ | |
65c70e6b | 5580 | FOR_EACH_DEFINED_FUNCTION (node) |
93a18a70 | 5581 | if (node->lowered |
d52f5295 | 5582 | && node->get_availability () >= AVAIL_INTERPOSABLE) |
0a35513e | 5583 | { |
594ec92f | 5584 | d = get_cg_data (&node, true); |
67348ccc | 5585 | if (is_tm_safe (node->decl)) |
0a35513e AH |
5586 | ipa_tm_diagnose_tm_safe (node); |
5587 | else if (d->all_tm_regions) | |
5588 | ipa_tm_diagnose_transaction (node, d->all_tm_regions); | |
5589 | } | |
5590 | ||
5591 | /* Create clones. Do those that are not irrevocable and have a | |
5592 | positive call count. Do those publicly visible functions that | |
5593 | the user directed us to clone. */ | |
9771b263 | 5594 | for (i = 0; i < tm_callees.length (); ++i) |
0a35513e AH |
5595 | { |
5596 | bool doit = false; | |
5597 | ||
9771b263 | 5598 | node = tm_callees[i]; |
67348ccc | 5599 | if (node->cpp_implicit_alias) |
0a35513e AH |
5600 | continue; |
5601 | ||
d52f5295 | 5602 | a = node->get_availability (); |
594ec92f | 5603 | d = get_cg_data (&node, true); |
0a35513e AH |
5604 | |
5605 | if (a <= AVAIL_NOT_AVAILABLE) | |
67348ccc DM |
5606 | doit = is_tm_callable (node->decl); |
5607 | else if (a <= AVAIL_AVAILABLE && is_tm_callable (node->decl)) | |
0a35513e AH |
5608 | doit = true; |
5609 | else if (!d->is_irrevocable | |
5610 | && d->tm_callers_normal + d->tm_callers_clone > 0) | |
5611 | doit = true; | |
5612 | ||
5613 | if (doit) | |
5614 | ipa_tm_create_version (node); | |
5615 | } | |
5616 | ||
5617 | /* Redirect calls to the new clones, and insert irrevocable marks. */ | |
9771b263 | 5618 | for (i = 0; i < tm_callees.length (); ++i) |
0a35513e | 5619 | { |
9771b263 | 5620 | node = tm_callees[i]; |
67348ccc | 5621 | if (node->analyzed) |
0a35513e | 5622 | { |
594ec92f | 5623 | d = get_cg_data (&node, true); |
0a35513e AH |
5624 | if (d->clone) |
5625 | ipa_tm_transform_clone (node); | |
5626 | } | |
5627 | } | |
65c70e6b | 5628 | FOR_EACH_DEFINED_FUNCTION (node) |
93a18a70 | 5629 | if (node->lowered |
d52f5295 | 5630 | && node->get_availability () >= AVAIL_INTERPOSABLE) |
0a35513e | 5631 | { |
594ec92f | 5632 | d = get_cg_data (&node, true); |
0a35513e AH |
5633 | if (d->all_tm_regions) |
5634 | ipa_tm_transform_transaction (node); | |
5635 | } | |
5636 | ||
5637 | /* Free and clear all data structures. */ | |
9771b263 DN |
5638 | tm_callees.release (); |
5639 | irr_worklist.release (); | |
0a35513e | 5640 | bitmap_obstack_release (&tm_obstack); |
398b1daa | 5641 | free_original_copy_tables (); |
0a35513e | 5642 | |
65c70e6b | 5643 | FOR_EACH_FUNCTION (node) |
67348ccc | 5644 | node->aux = NULL; |
0a35513e | 5645 | |
b2b29377 | 5646 | cgraph_node::checking_verify_cgraph_nodes (); |
0a35513e AH |
5647 | |
5648 | return 0; | |
5649 | } | |
5650 | ||
17795822 TS |
5651 | namespace { |
5652 | ||
5653 | const pass_data pass_data_ipa_tm = | |
27a4cd48 DM |
5654 | { |
5655 | SIMPLE_IPA_PASS, /* type */ | |
5656 | "tmipa", /* name */ | |
5657 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
5658 | TV_TRANS_MEM, /* tv_id */ |
5659 | ( PROP_ssa | PROP_cfg ), /* properties_required */ | |
5660 | 0, /* properties_provided */ | |
5661 | 0, /* properties_destroyed */ | |
5662 | 0, /* todo_flags_start */ | |
5663 | 0, /* todo_flags_finish */ | |
0a35513e AH |
5664 | }; |
5665 | ||
17795822 | 5666 | class pass_ipa_tm : public simple_ipa_opt_pass |
27a4cd48 DM |
5667 | { |
5668 | public: | |
c3284718 RS |
5669 | pass_ipa_tm (gcc::context *ctxt) |
5670 | : simple_ipa_opt_pass (pass_data_ipa_tm, ctxt) | |
27a4cd48 DM |
5671 | {} |
5672 | ||
5673 | /* opt_pass methods: */ | |
1a3d085c | 5674 | virtual bool gate (function *) { return flag_tm; } |
be55bfe6 | 5675 | virtual unsigned int execute (function *) { return ipa_tm_execute (); } |
27a4cd48 DM |
5676 | |
5677 | }; // class pass_ipa_tm | |
5678 | ||
17795822 TS |
5679 | } // anon namespace |
5680 | ||
27a4cd48 DM |
5681 | simple_ipa_opt_pass * |
5682 | make_pass_ipa_tm (gcc::context *ctxt) | |
5683 | { | |
5684 | return new pass_ipa_tm (ctxt); | |
5685 | } | |
5686 | ||
0a35513e | 5687 | #include "gt-trans-mem.h" |