2 * Copyright (C) 1995-2008 University of Karlsruhe. All right reserved.
4 * This file is part of libFirm.
6 * This file may be distributed and/or modified under the terms of the
7 * GNU General Public License version 2 as published by the Free Software
8 * Foundation and appearing in the file LICENSE.GPL included in the
9 * packaging of this file.
11 * Licensees holding valid libFirm Professional Edition licenses may use
12 * this file in accordance with the libFirm Commercial License.
13 * Agreement provided with the Software.
15 * This file is provided AS IS with NO WARRANTY OF ANY KIND, INCLUDING THE
16 * WARRANTY OF DESIGN, MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * @brief parallelizing Load/Store optimisation
23 * @author Christoph Mallon
30 #include "iroptimize.h"
41 #include "irnodeset.h"
46 #define OPTIMISE_LOAD_AFTER_LOAD
49 #define UNIMPLEMENTED abort();
52 DEBUG_ONLY(static firm_dbg_module_t *dbg);
55 static struct obstack obst;
56 static size_t count_addrs;
57 static ir_node** addrs;
60 static void AddressCollector(ir_node* node, void* env)
62 ir_nodeset_t* addrs_set = env;
65 addr = get_Load_ptr(node);
66 } else if (is_Store(node)) {
67 addr = get_Store_ptr(node);
71 ir_nodeset_insert(addrs_set, addr);
75 /* Collects all unique addresses used by load and store nodes of a graph and
76 * puts them into an array for later use */
77 static void CollectAddresses(ir_graph* irg)
79 ir_nodeset_t addrs_set;
81 ir_nodeset_init(&addrs_set);
82 irg_walk_graph(irg, AddressCollector, NULL, &addrs_set);
84 count_addrs = ir_nodeset_size(&addrs_set);
85 DB((dbg, LEVEL_1, "===> %+F uses %u unique addresses\n", irg, (unsigned int)count_addrs));
86 if (count_addrs != 0) {
87 ir_nodeset_iterator_t addr_iter;
90 addrs = NEW_ARR_D(ir_node*, &obst, count_addrs);
91 ir_nodeset_iterator_init(&addr_iter, &addrs_set);
92 for (i = 0; i < count_addrs; i++) {
93 ir_node* addr = ir_nodeset_iterator_next(&addr_iter);
95 set_irn_link(addr, (void *)i);
97 DB((dbg, LEVEL_2, "===> Collected unique symbolic address %+F\n", addr));
103 static void AliasSetAdder(ir_node* block, void* env)
105 ir_nodeset_t* alias_set;
109 alias_set = NEW_ARR_D(ir_nodeset_t, &obst, count_addrs);
110 for (i = 0; i < count_addrs; i++) {
111 ir_nodeset_init(&alias_set[i]);
113 set_irn_link(block, alias_set);
117 static void SetStartAddressesTop(ir_graph* irg)
119 ir_node* initial_mem;
120 ir_node* start_block;
121 ir_nodeset_t* start_addrs;
124 initial_mem = get_irg_initial_mem(irg);
125 start_block = get_irg_start_block(irg);
126 start_addrs = get_irn_link(start_block);
127 for (i = 0; i < count_addrs; i++) {
128 ir_nodeset_insert(&start_addrs[i], initial_mem);
130 mark_Block_block_visited(start_block);
134 static void AliasSetDestroyer(ir_node* block, void* env)
136 ir_nodeset_t* alias_set = get_irn_link(block);
140 for (i = 0; i < count_addrs; i++) {
141 ir_nodeset_destroy(&alias_set[i]);
146 static ir_alias_relation AliasTest(ir_graph* irg, ir_node* addr, ir_mode* mode, ir_node* other)
151 if (is_Proj(other)) other = get_Proj_pred(other);
153 if (is_Load(other)) {
154 other_addr = get_Load_ptr(other);
155 } else if (is_Store(other)) {
156 other_addr = get_Store_ptr(other);
161 other_mode = get_irn_mode(other);
162 return get_alias_relation(irg, addr, mode, other_addr, other_mode);
166 static ir_node* GenerateSync(ir_graph* irg, ir_node* block, ir_nodeset_t* after_set)
168 size_t set_size = ir_nodeset_size(after_set);
169 ir_nodeset_iterator_t iter;
171 assert(set_size != 0);
173 ir_nodeset_iterator_init(&iter, after_set);
175 return ir_nodeset_iterator_next(&iter);
180 NEW_ARR_A(ir_node*, in, set_size);
181 for (i = 0; i < set_size; i++) {
182 in[i] = ir_nodeset_iterator_next(&iter);
184 return new_r_Sync(irg, block, set_size, in);
189 static ir_node** unfinished_phis;
192 static void PlaceMemPhis(ir_graph* irg, ir_node* block, ir_node* phi)
195 size_t block_n_preds = get_Block_n_cfgpreds(block);
196 ir_nodeset_t* thissets;
201 thissets = get_irn_link(block);
202 NEW_ARR_A(ir_node*, in, block_n_preds);
203 for (j = 0; j < count_addrs; j++) {
206 for (i = 0; i < block_n_preds; i++) {
207 ir_node* pred_block = get_nodes_block(get_Phi_pred(phi, i)); // TODO get_Block_cfgpred_block(block, i);
208 ir_nodeset_t* predsets = get_irn_link(pred_block);
209 size_t predset_size = ir_nodeset_size(&predsets[j]);
211 if (predset_size == 0) {
212 in[i] = new_r_Unknown(irg, mode_M);
215 in[i] = GenerateSync(irg, pred_block, &predsets[j]);
218 new_phi = new_r_Phi(irg, block, block_n_preds, in, mode_M);
220 set_irn_link(new_phi, unfinished_phis[j]);
221 unfinished_phis[j] = new_phi;
223 ir_nodeset_insert(&thissets[j], new_phi);
228 static int WalkMem(ir_graph* irg, ir_node* node, ir_node* last_block);
231 static void WalkMemPhi(ir_graph* irg, ir_node* block, ir_node* phi)
233 size_t n = get_Phi_n_preds(phi);
236 for (i = 0; i < n; i++) {
237 WalkMem(irg, get_Phi_pred(phi, i), block);
240 PlaceMemPhis(irg, block, phi);
241 exchange(phi, new_Bad());
245 static void PlaceLoad(ir_graph* irg, ir_node* block, ir_node* load, ir_node* memory)
247 ir_node* addr = get_Load_ptr(load);
248 size_t addr_idx = (size_t)get_irn_link(addr);
249 ir_nodeset_t* interfere_sets = get_irn_link(block);
250 ir_nodeset_t* interfere_set = &interfere_sets[addr_idx];
251 size_t size = ir_nodeset_size(interfere_set);
252 ir_nodeset_iterator_t interfere_iter;
256 ir_nodeset_iterator_init(&interfere_iter, interfere_set);
258 ir_node* after = ir_nodeset_iterator_next(&interfere_iter);
259 assert(!is_Proj(after) || !is_Load(get_Proj_pred(after)));
260 DB((dbg, LEVEL_3, "===> %+F must be executed after %+F\n", load, after));
261 set_Load_mem(load, after);
268 NEW_ARR_A(ir_node*, after_set, size);
270 while ((mem = ir_nodeset_iterator_next(&interfere_iter)) != NULL) {
272 ir_node* pred = get_Proj_pred(mem);
274 #ifdef OPTIMISE_LOAD_AFTER_LOAD
275 if (get_Load_ptr(pred) == addr && get_Load_mode(pred) == get_Load_mode(load)) {
276 exchange(load, pred);
283 DB((dbg, LEVEL_3, "===> %+F must be executed after %+F\n", load, mem));
284 after_set[i++] = mem;
288 after = after_set[0];
290 after = new_r_Sync(irg, block, i, after_set);
292 set_Load_mem(load, after);
295 for (i = 0; i < count_addrs; i++) {
296 ir_mode* mode = get_Load_mode(load);
297 ir_node* other_addr = addrs[i];
298 ir_mode* other_mode = mode; // XXX second mode is nonsense
299 ir_alias_relation rel = get_alias_relation(irg, addr, mode, other_addr, other_mode);
301 DB((dbg, LEVEL_3, "===> Testing for alias between %+F and %+F. Relation is %d\n", addr, other_addr, rel));
302 if (rel == no_alias) {
305 DB((dbg, LEVEL_3, "===> %+F potentially aliases address %+F\n", load, other_addr));
307 ir_nodeset_insert(&interfere_sets[i], memory);
312 static void PlaceStore(ir_graph* irg, ir_node* block, ir_node* store, ir_node* memory)
314 ir_node* addr = get_Store_ptr(store);
315 size_t addr_idx = (size_t)get_irn_link(addr);
316 ir_nodeset_t* interfere_sets = get_irn_link(block);
317 ir_nodeset_t* interfere_set = &interfere_sets[addr_idx];
321 after = GenerateSync(irg, block, interfere_set);
322 set_Store_mem(store, after);
324 for (i = 0; i < count_addrs; i++) {
325 ir_nodeset_iterator_t interfere_iter;
326 ir_mode* mode = get_irn_mode(get_Store_value(store));
327 ir_node* other_addr = addrs[i];
328 ir_mode* other_mode = mode; // XXX second mode is nonsense
329 ir_alias_relation rel = get_alias_relation(irg, addr, mode, other_addr, other_mode);
332 DB((dbg, LEVEL_3, "===> Testing for alias between %+F and %+F. Relation is %d\n", addr, other_addr, rel));
333 if (rel == no_alias) {
336 DB((dbg, LEVEL_3, "===> %+F potentially aliases address %+F\n", store, other_addr));
338 ir_nodeset_iterator_init(&interfere_iter, &interfere_sets[i]);
339 while ((other_node = ir_nodeset_iterator_next(&interfere_iter)) != NULL) {
340 if (AliasTest(irg, addr, mode, other_node) != no_alias) {
341 DB((dbg, LEVEL_3, "===> Removing %+F from execute-after set of %+F due to %+F\n", other_node, addrs[i], store));
342 ir_nodeset_remove_iterator(&interfere_sets[i], &interfere_iter);
346 ir_nodeset_insert(&interfere_sets[i], memory);
351 static int WalkMem(ir_graph* irg, ir_node* node, ir_node* last_block)
353 int block_change = 0;
354 ir_node* block = get_nodes_block(node);
356 ir_node* memory = node;
357 ir_nodeset_t* addr_sets;
359 if (block != last_block) {
360 DB((dbg, LEVEL_3, "===> Changing block from %+F to %+F\n", last_block, block));
362 if (Block_not_block_visited(block)) {
363 mark_Block_block_visited(block);
365 DB((dbg, LEVEL_2, "===> Hit already visited block at %+F\n", node));
371 if (is_Proj(node)) node = get_Proj_pred(node);
374 WalkMemPhi(irg, block, node);
376 } else if (is_Sync(node)) {
378 } else if (is_Return(node)) {
379 pred = get_Return_mem(node);
381 pred = get_fragile_op_mem(node);
384 if (WalkMem(irg, pred, block)) {
385 // There was a block change
386 size_t block_arity = get_Block_n_cfgpreds(block);
388 DB((dbg, LEVEL_3, "===> There is a block change before %+F\n", node));
389 if (block_arity == 1) {
390 // Just one predecessor, inherit its alias sets
391 ir_node* pred_block = get_nodes_block(pred);
392 ir_nodeset_t* predsets = get_irn_link(pred_block);
393 ir_nodeset_t* thissets = get_irn_link(block);
396 DB((dbg, LEVEL_3, "===> Copying the only predecessor's address sets\n"));
398 if (ir_nodeset_size(&predsets[0]) == 0) {
401 DB((dbg, LEVEL_3, "===> The predecessor was not finished yet\n"));
402 assert(!Block_not_block_visited(pred_block));
404 unknown = new_r_Unknown(irg, mode_M);
405 for (i = 0; i < count_addrs; i++) {
406 ir_node* phi_unk = new_r_Phi(irg, block, 1, &unknown, mode_M);
407 DB((dbg, LEVEL_3, "===> Placing unfinished %+F for %+F in %+F\n", phi_unk, addrs[i], block));
408 set_irn_link(phi_unk, unfinished_phis[i]);
409 unfinished_phis[i] = phi_unk;
410 ir_nodeset_insert(&thissets[i], phi_unk);
413 for (i = 0; i < count_addrs; i++) {
414 ir_nodeset_iterator_t prediter;
417 ir_nodeset_iterator_init(&prediter, &predsets[i]);
418 while ((addr = ir_nodeset_iterator_next(&prediter)) != NULL) {
419 ir_nodeset_insert(&thissets[i], addr);
426 DB((dbg, LEVEL_3, "===> Detotalising %+F\n", node));
428 addr_sets = get_irn_link(block);
431 PlaceLoad(irg, block, node, memory);
432 } else if (is_Store(node)) {
433 PlaceStore(irg, block, node, memory);
435 ir_nodeset_t sync_set;
439 DB((dbg, LEVEL_3, "===> Fallback: %+F aliases everything\n", node));
441 ir_nodeset_init(&sync_set);
442 for (i = 0; i < count_addrs; i++) {
443 ir_nodeset_iterator_t iter;
446 ir_nodeset_iterator_init(&iter, &addr_sets[i]);
447 while ((mem = ir_nodeset_iterator_next(&iter)) != NULL) {
448 ir_nodeset_insert(&sync_set, mem);
452 after = GenerateSync(irg, block, &sync_set);
453 set_irn_n(node, 0, after); // XXX unnice way to set the memory input
455 for (i = 0; i < count_addrs; i++) {
456 ir_nodeset_iterator_t iter;
457 ir_nodeset_iterator_init(&iter, &addr_sets[i]);
458 while (ir_nodeset_iterator_next(&iter) != NULL) {
459 ir_nodeset_remove_iterator(&addr_sets[i], &iter);
461 ir_nodeset_insert(&addr_sets[i], memory);
469 static void FinalisePhis(ir_graph* irg)
473 for (i = 0; i < count_addrs; i++) {
477 for (phi = unfinished_phis[i]; phi != NULL; phi = next_phi) {
478 ir_node* block = get_nodes_block(phi);
479 size_t block_n_preds = get_Block_n_cfgpreds(block);
481 next_phi = get_irn_link(phi);
483 DB((dbg, LEVEL_4, "===> Finialising phi %+F in %+F\n", phi, block));
485 if (block_n_preds == 1) {
486 ir_node* pred_block = get_Block_cfgpred_block(block, 0);
487 ir_nodeset_t* pred_sets = get_irn_link(pred_block);
488 ir_node* after = GenerateSync(irg, pred_block, &pred_sets[i]);
490 assert(is_Unknown(get_Phi_pred(phi, 0)));
491 exchange(phi, after);
496 NEW_ARR_A(ir_node*, in, block_n_preds);
497 for (j = 0; j < block_n_preds; j++) {
498 ir_node* pred_block = get_Block_cfgpred_block(block, j);
499 ir_nodeset_t* pred_sets = get_irn_link(pred_block);
501 if (is_Unknown(get_Phi_pred(phi, j))) {
502 set_Phi_pred(phi, j, GenerateSync(irg, pred_block, &pred_sets[i]));
511 static void Detotalise(ir_graph* irg)
513 ir_node* end_block = get_irg_end_block(irg);
514 size_t npreds = get_Block_n_cfgpreds(end_block);
517 unfinished_phis = xmalloc(sizeof(*unfinished_phis) * count_addrs);
518 for (i = 0; i < count_addrs; i++) {
519 unfinished_phis[i] = NULL;
522 for (i = 0; i < npreds; i++) {
523 ir_node* pred = get_Block_cfgpred(end_block, i);
524 assert(is_Return(pred));
525 DB((dbg, LEVEL_2, "===> Starting memory walk at %+F\n", pred));
526 WalkMem(irg, pred, NULL);
530 xfree(unfinished_phis);
534 static void AddSyncPreds(ir_nodeset_t* preds, ir_node* sync)
536 size_t n = get_Sync_n_preds(sync);
539 for (i = 0; i < n; i++) {
540 ir_node* pred = get_Sync_pred(sync, i);
542 AddSyncPreds(preds, pred);
544 ir_nodeset_insert(preds, pred);
550 static void NormaliseSync(ir_node* node, void* env)
553 ir_nodeset_iterator_t iter;
559 if (!is_Sync(node)) return;
561 ir_nodeset_init(&preds);
562 AddSyncPreds(&preds, node);
564 count_preds = ir_nodeset_size(&preds);
565 if (count_preds != (unsigned)get_Sync_n_preds(node)) {
566 NEW_ARR_A(ir_node*, in, count_preds);
567 ir_nodeset_iterator_init(&iter, &preds);
568 for (i = 0; i < count_preds; i++) {
569 ir_node* pred = ir_nodeset_iterator_next(&iter);
570 assert(pred != NULL);
573 set_irn_in(node, count_preds, in);
576 ir_nodeset_destroy(&preds);
580 void opt_ldst2(ir_graph* irg)
582 FIRM_DBG_REGISTER(dbg, "firm.opt.ldst2");
583 DB((dbg, LEVEL_1, "===> Performing load/store optimisation on %+F\n", irg));
585 normalize_one_return(irg);
586 dump_ir_block_graph(irg, "-prefluffig");
590 if (1 /* XXX */ || get_opt_alias_analysis()) {
591 assure_irg_address_taken_computed(irg);
592 assure_irp_globals_address_taken_computed();
596 CollectAddresses(irg);
597 if (count_addrs == 0) return;
599 irg_block_walk_graph(irg, AliasSetAdder, NULL, NULL);
600 inc_irg_block_visited(irg);
601 SetStartAddressesTop(irg);
603 dump_ir_block_graph(irg, "-fluffig");
605 irg_block_walk_graph(irg, AliasSetDestroyer, NULL, NULL);
606 obstack_free(&obst, NULL);
608 normalize_proj_nodes(irg);
609 irg_walk_graph(irg, NormaliseSync, NULL, NULL);
610 optimize_graph_df(irg);
611 irg_walk_graph(irg, NormaliseSync, NULL, NULL);
612 dump_ir_block_graph(irg, "-postfluffig");