3 * File name: ir/opt/ldstopt.c
4 * Purpose: load store optimizations
8 * Copyright: (c) 1998-2004 Universit
\81ät Karlsruhe
9 * Licence: This file protected by GPL - GNU GENERAL PUBLIC LICENSE.
11 # include "irnode_t.h"
12 # include "irgraph_t.h"
13 # include "irmode_t.h"
15 # include "ircons_t.h"
20 # include "dbginfo_t.h"
21 # include "iropt_dbg.h"
22 # include "irflag_t.h"
24 # include "firmstat.h"
27 #define IMAX(a,b) ((a) > (b) ? (a) : (b))
29 #define MAX_PROJ IMAX(pn_Load_max, pn_Store_max)
34 typedef struct _walk_env_t {
35 struct obstack obst; /**< list of all stores */
42 typedef struct _ldst_info_t {
43 ir_node *projs[MAX_PROJ]; /**< list of Proj's of this node */
44 ir_node *exc_block; /**< the exception block if available */
45 int exc_idx; /**< predecessor index in the exception block */
49 * flags for control flow
52 BLOCK_HAS_COND = 1, /**< Block has conditional control flow */
53 BLOCK_HAS_EXC = 2 /**< Block has exceptionl control flow */
59 typedef struct _block_info_t {
60 unsigned flags; /**< flags for the block */
64 * walker, clears all links first
66 static void init_links(ir_node *n, void *env)
68 set_irn_link(n, NULL);
72 * get the Load/Store info of a node
74 static ldst_info_t *get_ldst_info(ir_node *node, walk_env_t *env)
76 ldst_info_t *info = get_irn_link(node);
79 info = obstack_alloc(&env->obst, sizeof(*info));
81 memset(info, 0, sizeof(*info));
83 set_irn_link(node, info);
89 * get the Block info of a node
91 static block_info_t *get_block_info(ir_node *node, walk_env_t *env)
93 block_info_t *info = get_irn_link(node);
96 info = obstack_alloc(&env->obst, sizeof(*info));
98 memset(info, 0, sizeof(*info));
100 set_irn_link(node, info);
106 * update the projection info for a Load/Store
108 static int update_projs(ldst_info_t *info, ir_node *proj)
110 long nr = get_Proj_proj(proj);
112 assert(0 <= nr && nr <= MAX_PROJ && "Wrong proj from LoadStore");
114 if (info->projs[nr]) {
115 /* there is already one, do CSE */
116 exchange(proj, info->projs[nr]);
120 info->projs[nr] = proj;
126 * update the exception block info for a Load/Store
128 static int update_exc(ldst_info_t *info, ir_node *block, int pos)
130 assert(info->exc_block == NULL && "more than one exception block found");
132 info->exc_block = block;
137 #define get_irn_out_n(node) (unsigned)get_irn_link(node)
138 #define set_irn_out_n(node, n) set_irn_link(adr, (void *)(n))
141 * walker, collects all Load/Store/Proj nodes
143 * walks form Start -> End
145 static void collect_nodes(ir_node *node, void *env)
148 ldst_info_t *ldst_info;
149 walk_env_t *wenv = env;
151 if (get_irn_op(node) == op_Proj) {
155 pred = get_Proj_pred(node);
156 op = get_irn_op(pred);
159 adr = get_Load_ptr(pred);
160 ldst_info = get_ldst_info(pred, wenv);
162 wenv->changes |= update_projs(ldst_info, node);
164 set_irn_out_n(adr, get_irn_out_n(adr) + 1);
166 else if (op == op_Store) {
167 adr = get_Store_ptr(pred);
168 ldst_info = get_ldst_info(pred, wenv);
170 wenv->changes |= update_projs(ldst_info, node);
172 set_irn_out_n(adr, get_irn_out_n(adr) + 1);
175 else if (get_irn_op(node) == op_Block) { /* check, if it's an exception block */
178 for (i = 0, n = get_Block_n_cfgpreds(node); i < n; ++i) {
180 block_info_t *bl_info;
182 pred = skip_Proj(get_Block_cfgpred(node, i));
184 /* ignore Bad predecessors, they will be removed later */
188 pred_block = get_nodes_block(pred);
189 bl_info = get_block_info(pred_block, wenv);
191 if (is_fragile_op(pred))
192 bl_info->flags |= BLOCK_HAS_EXC;
193 else if (is_forking_op(pred))
194 bl_info->flags |= BLOCK_HAS_COND;
196 if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) {
197 ldst_info = get_ldst_info(pred, wenv);
199 wenv->changes |= update_exc(ldst_info, node, i);
208 static int optimize_load(ir_node *load)
210 ldst_info_t *info = get_irn_link(load);
211 ir_mode *load_mode = get_Load_mode(load);
212 ir_node *pred, *mem, *ptr;
215 /* the address of the load to be optimized */
216 ptr = get_Load_ptr(load);
219 * Check if we can remove the exception form a Load:
220 * this can be done, if the address is from an Sel(Alloc) and
221 * the Sel type is a subtype of the alloc type.
223 * This optimizes some often used OO constructs,
224 * like x = new O; x->t;
226 if (info->projs[pn_Load_X_except]) {
227 if (get_irn_op(ptr) == op_Sel) {
228 ir_node *mem = get_Sel_mem(ptr);
230 if (get_irn_op(mem) == op_Alloc) {
231 /* ok, check the types */
232 entity *ent = get_Sel_entity(ptr);
233 type *s_type = get_entity_type(ent);
234 type *a_type = get_Alloc_type(mem);
236 if (is_subclass_of(s_type, a_type)) {
237 /* ok, condition met: there can't be an exception because
238 * alloc guarantees that enough memory was allocated */
240 exchange(info->projs[pn_Load_X_except], new_Bad());
241 info->projs[pn_Load_X_except] = NULL;
245 else if (get_irn_op(ptr) == op_Alloc) {
246 /* simple case: a direct load after an Alloc. Firm Alloc throw
247 * an exception in case of out-of-memory. So, there is no way for an
248 * exception in this load.
249 * This code is constructed by the "exception lowering" in the Jack compiler.
251 exchange(info->projs[pn_Load_X_except], new_Bad());
252 info->projs[pn_Load_X_except] = NULL;
256 /* do NOT touch volatile loads for now */
257 if (get_Load_volatility(load) == volatility_is_volatile)
260 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
261 /* a Load which value is neither used nor exception checked, remove it */
262 mem = get_Load_mem(load);
263 exchange(info->projs[pn_Load_M], mem);
268 /* the mem of the Load. Must still be returned after optimization */
269 mem = get_Load_mem(load);
271 if ((get_irn_opcode(ptr) == iro_SymConst) && (get_SymConst_kind(ptr) == symconst_addr_ent)) {
272 entity *ent = get_SymConst_entity(ptr);
274 assert(mode_is_reference(get_irn_mode(ptr)));
276 if ((allocation_static == get_entity_allocation(ent)) &&
277 (visibility_external_allocated != get_entity_visibility(ent))) {
278 /* a static allocation that is not external: there should be NO exception
281 /* no exception, clear the info field as it might be checked later again */
282 if (info->projs[pn_Load_X_except]) {
283 exchange(info->projs[pn_Load_X_except], new_Bad());
284 info->projs[pn_Load_X_except] = NULL;
287 if (variability_constant == get_entity_variability(ent)
288 && is_atomic_entity(ent)) { /* Might not be atomic after
289 lowering of Sels. In this
290 case we could also load, but
291 it's more complicated. */
292 /* more simpler case: we load the content of a constant value:
293 * replace it by the constant itself
297 if (info->projs[pn_Load_M])
298 exchange(info->projs[pn_Load_M], mem);
301 if (info->projs[pn_Load_res])
302 exchange(info->projs[pn_Load_res], copy_const_value(get_atomic_ent_value(ent)));
307 /* we changed the irg, but try further */
312 /* Check, if the address of this load is used more than once.
313 * If not, this load cannot be removed in any case. */
314 if (get_irn_out_n(ptr) <= 1)
317 /* follow the memory chain as long as there are only Loads */
318 for (pred = skip_Proj(mem); ; pred = skip_Proj(get_Load_mem(pred))) {
321 * BEWARE: one might think that checking the modes is useless, because
322 * if the pointers are identical, they refer to the same object.
323 * This is only true in strong typed languages, not is C were the following
324 * is possible a = *(type1 *)p; b = *(type2 *)p ...
327 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
328 get_irn_mode(get_Store_value(pred)) == load_mode) {
329 ldst_info_t *pred_info = get_irn_link(pred);
332 * a Load immediately after a Store -- a read after write.
333 * We may remove the Load, if both Load & Store does not have an exception handler
334 * OR they are in the same block. In the latter case the Load cannot
335 * throw an exception when the previous Store was quiet.
337 * Why we need to check for Store Exc? If the Store cannot be executed (ROM)
338 * the exception handler might simply jump into the load block :-(
339 * We could make it a little bit better if we would know that the exception
340 * handler of the Store jumps directly to the end...
342 if ((!pred_info->projs[pn_Store_X_except] && !info->projs[pn_Load_X_except]) ||
343 get_nodes_block(load) == get_nodes_block(pred)) {
344 DBG_OPT_RAW(pred, load);
345 exchange( info->projs[pn_Load_res], get_Store_value(pred) );
347 if (info->projs[pn_Load_M])
348 exchange(info->projs[pn_Load_M], mem);
351 if (info->projs[pn_Load_X_except])
352 exchange( info->projs[pn_Load_X_except], new_Bad());
356 else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
357 get_Load_mode(pred) == load_mode) {
359 * a Load after a Load -- a read after read.
360 * We may remove the second Load, if it does not have an exception handler
361 * OR they are in the same block. In the later case the Load cannot
362 * throw an exception when the previous Load was quiet.
364 * Here, there is no need to check if the previos Load has an exception hander because
365 * they would have exact the same exception...
367 if (! info->projs[pn_Load_X_except] || get_nodes_block(load) == get_nodes_block(pred)) {
368 ldst_info_t *pred_info = get_irn_link(pred);
370 DBG_OPT_RAR(pred, load);
372 if (pred_info->projs[pn_Load_res]) {
373 /* we need a data proj from the previous load for this optimization */
374 exchange( info->projs[pn_Load_res], pred_info->projs[pn_Load_res] );
375 if (info->projs[pn_Load_M])
376 exchange(info->projs[pn_Load_M], mem);
379 if (info->projs[pn_Load_res]) {
380 set_Proj_pred(info->projs[pn_Load_res], pred);
381 set_nodes_block(info->projs[pn_Load_res], get_nodes_block(pred));
383 if (info->projs[pn_Load_M]) {
384 /* Actually, this if should not be necessary. Construct the Loads
386 exchange(info->projs[pn_Load_M], mem);
391 if (info->projs[pn_Load_X_except])
392 exchange(info->projs[pn_Load_X_except], new_Bad());
398 /* follow only Load chains */
399 if (get_irn_op(pred) != op_Load)
408 static int optimize_store(ir_node *store)
410 ldst_info_t *info = get_irn_link(store);
411 ir_node *pred, *mem, *ptr, *value, *block;
415 if (get_Store_volatility(store) == volatility_is_volatile)
419 * BEWARE: one might think that checking the modes is useless, because
420 * if the pointers are identical, they refer to the same object.
421 * This is only true in strong typed languages, not is C were the following
422 * is possible *(type1 *)p = a; *(type2 *)p = b ...
425 ptr = get_Store_ptr(store);
427 /* Check, if the address of this load is used more than once.
428 * If not, this load cannot be removed in any case. */
429 if (get_irn_out_n(ptr) <= 1)
432 block = get_nodes_block(store);
433 mem = get_Store_mem(store);
434 value = get_Store_value(store);
435 mode = get_irn_mode(value);
437 /* follow the memory chain as long as there are only Loads */
438 for (pred = skip_Proj(mem); ; pred = skip_Proj(get_Load_mem(pred))) {
439 ldst_info_t *pred_info = get_irn_link(pred);
441 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
442 get_nodes_block(pred) == block && get_irn_mode(get_Store_value(pred)) == mode) {
444 * a Store after a Store in the same block -- a write after write.
445 * We may remove the first Store, if it does not have an exception handler.
447 * TODO: What, if both have the same exception handler ???
449 if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
450 DBG_OPT_WAW(pred, store);
451 exchange( pred_info->projs[pn_Store_M], get_Store_mem(pred) );
455 else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
456 value == pred_info->projs[pn_Load_res]) {
458 * a Store of a value after a Load -- a write after read.
459 * We may remove the second Store, if it does not have an exception handler.
461 if (! info->projs[pn_Store_X_except]) {
462 DBG_OPT_WAR(pred, store);
463 exchange( info->projs[pn_Store_M], mem );
468 /* follow only Load chains */
469 if (get_irn_op(pred) != op_Load)
476 * walker, optimizes Phi after Stores:
477 * Does the following optimization:
479 * val1 val2 val3 val1 val2 val3
487 * This removes the number of stores and allows for predicated execution.
488 * Moves Stores back to the end of a function which may be bad
490 * Is only allowed if the predecessor blocks have only one successor.
492 static int optimize_phi(ir_node *phi, void *env)
494 walk_env_t *wenv = env;
496 ir_node *store, *ptr, *block, *phiM, *phiD, *exc, *projM;
498 ir_node **inM, **inD;
502 block_info_t *bl_info;
504 /* Must be a memory Phi */
505 if (get_irn_mode(phi) != mode_M)
508 n = get_Phi_n_preds(phi);
512 store = skip_Proj(get_Phi_pred(phi, 0));
513 if (get_irn_op(store) != op_Store)
516 /* abort on bad blocks */
517 if (is_Bad(get_nodes_block(store)))
520 /* check if the block has only one output */
521 bl_info = get_irn_link(get_nodes_block(store));
525 /* this is the address of the store */
526 ptr = get_Store_ptr(store);
527 mode = get_irn_mode(get_Store_value(store));
528 info = get_irn_link(store);
529 exc = info->exc_block;
531 for (i = 1; i < n; ++i) {
532 ir_node *pred = skip_Proj(get_Phi_pred(phi, i));
534 if (get_irn_op(pred) != op_Store)
537 if (mode != get_irn_mode(get_Store_value(pred)) || ptr != get_Store_ptr(pred))
540 info = get_irn_link(pred);
542 /* check, if all stores have the same exception flow */
543 if (exc != info->exc_block)
546 /* abort on bad blocks */
547 if (is_Bad(get_nodes_block(store)))
550 /* check if the block has only one output */
551 bl_info = get_irn_link(get_nodes_block(store));
557 * ok, when we are here, we found all predecessors of a Phi that
558 * are Stores to the same address. That means whatever we do before
559 * we enter the block of the Phi, we do a Store.
560 * So, we can move the store to the current block:
562 * val1 val2 val3 val1 val2 val3
564 * | Str | | Str | | Str | \ | /
570 * Is only allowed if the predecessor blocks have only one successor.
573 /* first step: collect all inputs */
574 NEW_ARR_A(ir_node *, inM, n);
575 NEW_ARR_A(ir_node *, inD, n);
576 NEW_ARR_A(int, idx, n);
578 for (i = 0; i < n; ++i) {
579 ir_node *pred = skip_Proj(get_Phi_pred(phi, i));
580 info = get_irn_link(pred);
582 inM[i] = get_Store_mem(pred);
583 inD[i] = get_Store_value(pred);
584 idx[i] = info->exc_idx;
586 block = get_nodes_block(phi);
588 /* second step: create a new memory Phi */
589 phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
591 /* third step: create a new data Phi */
592 phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
594 /* fourth step: create the Store */
595 store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
596 projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
598 info = get_ldst_info(store, wenv);
599 info->projs[pn_Store_M] = projM;
601 /* fifths step: repair exception flow */
603 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
605 info->projs[pn_Store_X_except] = projX;
606 info->exc_block = exc;
607 info->exc_idx = idx[0];
609 for (i = 0; i < n; ++i) {
610 set_Block_cfgpred(exc, idx[i], projX);
614 /* the exception block should be optimized as some inputs are identical now */
618 /* sixt step: replace old Phi */
619 exchange(phi, projM);
625 * walker, collects all Load/Store/Proj nodes
627 static void do_load_store_optimize(ir_node *n, void *env)
629 walk_env_t *wenv = env;
631 switch (get_irn_opcode(n)) {
634 wenv->changes |= optimize_load(n);
638 wenv->changes |= optimize_store(n);
642 wenv->changes |= optimize_phi(n, env);
650 * do the load store optimization
652 void optimize_load_store(ir_graph *irg)
656 assert(get_irg_phase_state(irg) != phase_building);
658 if (!get_opt_redundant_LoadStore())
661 obstack_init(&env.obst);
664 /* init the links, then collect Loads/Stores/Proj's in lists */
665 irg_walk_graph(irg, init_links, collect_nodes, &env);
667 /* now we have collected enough information, optimize */
668 irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
670 obstack_free(&env.obst, NULL);
672 /* Handle graph state */
674 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
675 set_irg_outs_inconsistent(current_ir_graph);
677 /* is this really needed: Yes, as exception block may get bad but this might be tested */
678 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
679 set_irg_dom_inconsistent(current_ir_graph);