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;
138 * walker, collects all Load/Store/Proj nodes
140 * walks form Start -> End
142 static void collect_nodes(ir_node *node, void *env)
145 ldst_info_t *ldst_info;
146 walk_env_t *wenv = env;
148 if (get_irn_op(node) == op_Proj) {
149 pred = get_Proj_pred(node);
151 if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) {
152 ldst_info = get_ldst_info(pred, wenv);
154 wenv->changes |= update_projs(ldst_info, node);
157 else if (get_irn_op(node) == op_Block) { /* check, if it's an exception block */
160 for (i = 0, n = get_Block_n_cfgpreds(node); i < n; ++i) {
162 block_info_t *bl_info;
164 pred = skip_Proj(get_Block_cfgpred(node, i));
166 /* ignore Bad predecessors, they will be removed later */
170 pred_block = get_nodes_block(pred);
171 bl_info = get_block_info(pred_block, wenv);
173 if (is_fragile_op(pred))
174 bl_info->flags |= BLOCK_HAS_EXC;
175 else if (is_forking_op(pred))
176 bl_info->flags |= BLOCK_HAS_COND;
178 if (get_irn_op(pred) == op_Load || get_irn_op(pred) == op_Store) {
179 ldst_info = get_ldst_info(pred, wenv);
181 wenv->changes |= update_exc(ldst_info, node, i);
190 static int optimize_load(ir_node *load)
192 ldst_info_t *info = get_irn_link(load);
193 ir_mode *load_mode = get_Load_mode(load);
194 ir_node *pred, *mem, *ptr;
197 if (get_Load_volatility(load) == volatility_is_volatile)
201 * BEWARE: one might think that checking the modes is useless, because
202 * if the pointers are identical, they refer to the same object.
203 * This is only true in strong typed languages, not is C were the following
204 * is possible a = *(type1 *)p; b = *(type2 *)p ...
207 ptr = get_Load_ptr(load);
208 mem = get_Load_mem(load);
209 pred = skip_Proj(mem);
211 if (! info->projs[pn_Load_res] && ! info->projs[pn_Load_X_except]) {
212 /* a Load which value is neither used nor exception checked, remove it */
213 exchange(info->projs[pn_Load_M], mem);
216 else if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
217 get_irn_mode(get_Store_value(pred)) == load_mode) {
219 * a load immediately after a store -- a read after write.
220 * We may remove the Load, if it does not have an exception handler
221 * OR they are in the same block. In the latter case the Load cannot
222 * throw an exception when the previous Store was quiet.
224 if (! info->projs[pn_Load_X_except] || get_nodes_block(load) == get_nodes_block(pred)) {
225 DBG_OPT_RAW(pred, load);
226 exchange( info->projs[pn_Load_res], get_Store_value(pred) );
228 if (info->projs[pn_Load_M])
229 exchange(info->projs[pn_Load_M], mem);
232 if (info->projs[pn_Load_X_except])
233 exchange( info->projs[pn_Load_X_except], new_Bad());
237 else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
238 get_Load_mode(pred) == load_mode) {
240 * a load immediately after a load -- a read after read.
241 * We may remove the second Load, if it does not have an exception handler
242 * OR they are in the same block. In the later case the Load cannot
243 * throw an exception when the previous Load was quiet.
245 if (! info->projs[pn_Load_X_except] || get_nodes_block(load) == get_nodes_block(pred)) {
246 ldst_info_t *pred_info = get_irn_link(pred);
248 DBG_OPT_RAR(pred, load);
250 if (pred_info->projs[pn_Load_res]) {
251 /* we need a data proj from the previous load for this optimization */
252 exchange( info->projs[pn_Load_res], pred_info->projs[pn_Load_res] );
253 if (info->projs[pn_Load_M])
254 exchange(info->projs[pn_Load_M], mem);
257 if (info->projs[pn_Load_res]) {
258 set_Proj_pred(info->projs[pn_Load_res], pred);
259 set_nodes_block(info->projs[pn_Load_res], get_nodes_block(pred));
261 if (info->projs[pn_Load_M]) {
262 /* Actually, this if should not be necessary. Construct the Loads
264 exchange(info->projs[pn_Load_M], mem);
269 if (info->projs[pn_Load_X_except])
270 exchange(info->projs[pn_Load_X_except], new_Bad());
281 static int optimize_store(ir_node *store)
283 ldst_info_t *info = get_irn_link(store);
284 ir_node *pred, *mem, *ptr, *value, *block;
286 ldst_info_t *pred_info;
289 if (get_Store_volatility(store) == volatility_is_volatile)
293 * BEWARE: one might think that checking the modes is useless, because
294 * if the pointers are identical, they refer to the same object.
295 * This is only true in strong typed languages, not is C were the following
296 * is possible *(type1 *)p = a; *(type2 *)p = b ...
299 block = get_nodes_block(store);
300 ptr = get_Store_ptr(store);
301 mem = get_Store_mem(store);
302 value = get_Store_value(store);
303 pred = skip_Proj(mem);
304 mode = get_irn_mode(value);
306 pred_info = get_irn_link(pred);
308 if (get_irn_op(pred) == op_Store && get_Store_ptr(pred) == ptr &&
309 get_nodes_block(pred) == block && get_irn_mode(get_Store_value(pred)) == mode) {
311 * a store immediately after a store in the same block -- a write after write.
312 * We may remove the first Store, if it does not have an exception handler.
314 * TODO: What, if both have the same exception handler ???
316 if (get_Store_volatility(pred) != volatility_is_volatile && !pred_info->projs[pn_Store_X_except]) {
317 DBG_OPT_WAW(pred, store);
318 exchange( pred_info->projs[pn_Store_M], get_Store_mem(pred) );
322 else if (get_irn_op(pred) == op_Load && get_Load_ptr(pred) == ptr &&
323 value == pred_info->projs[pn_Load_res]) {
325 * a store a value immediately after a load -- a write after read.
326 * We may remove the second Store, if it does not have an exception handler.
328 if (! info->projs[pn_Store_X_except]) {
329 DBG_OPT_WAR(pred, store);
330 exchange( info->projs[pn_Store_M], mem );
338 * walker, optimizes Phi after Stores:
339 * Does the following optimization:
341 * val1 val2 val3 val1 val2 val3
349 * This removes the number of stores and allows for predicated execution.
350 * Moves Stores back to the end of a function which may be bad
352 * Is only allowed if the predecessor blocks have only one successor.
354 static int optimize_phi(ir_node *phi, void *env)
356 walk_env_t *wenv = env;
358 ir_node *store, *ptr, *block, *phiM, *phiD, *exc, *projM;
360 ir_node **inM, **inD;
364 block_info_t *bl_info;
366 /* Must be a memory Phi */
367 if (get_irn_mode(phi) != mode_M)
370 n = get_Phi_n_preds(phi);
374 store = skip_Proj(get_Phi_pred(phi, 0));
375 if (get_irn_op(store) != op_Store)
378 /* abort on bad blocks */
379 if (is_Bad(get_nodes_block(store)))
382 /* check if the block has only one output */
383 bl_info = get_irn_link(get_nodes_block(store));
387 /* this is the address of the store */
388 ptr = get_Store_ptr(store);
389 mode = get_irn_mode(get_Store_value(store));
390 info = get_irn_link(store);
391 exc = info->exc_block;
393 for (i = 1; i < n; ++i) {
394 ir_node *pred = skip_Proj(get_Phi_pred(phi, i));
396 if (get_irn_op(pred) != op_Store)
399 if (mode != get_irn_mode(get_Store_value(pred)) || ptr != get_Store_ptr(pred))
402 info = get_irn_link(pred);
404 /* check, if all stores have the same exception flow */
405 if (exc != info->exc_block)
408 /* abort on bad blocks */
409 if (is_Bad(get_nodes_block(store)))
412 /* check if the block has only one output */
413 bl_info = get_irn_link(get_nodes_block(store));
419 * ok, when we are here, we found all predecessors of a Phi that
420 * are Stores to the same address. That means whatever we do before
421 * we enter the block of the Phi, we do a Store.
422 * So, we can move the store to the current block:
424 * val1 val2 val3 val1 val2 val3
426 * | Str | | Str | | Str | \ | /
432 * Is only allowed if the predecessor blocks have only one successor.
435 /* first step: collect all inputs */
436 NEW_ARR_A(ir_node *, inM, n);
437 NEW_ARR_A(ir_node *, inD, n);
438 NEW_ARR_A(int, idx, n);
440 for (i = 0; i < n; ++i) {
441 ir_node *pred = skip_Proj(get_Phi_pred(phi, i));
442 info = get_irn_link(pred);
444 inM[i] = get_Store_mem(pred);
445 inD[i] = get_Store_value(pred);
446 idx[i] = info->exc_idx;
448 block = get_nodes_block(phi);
450 /* second step: create a new memory Phi */
451 phiM = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inM, mode_M);
453 /* third step: create a new data Phi */
454 phiD = new_rd_Phi(get_irn_dbg_info(phi), current_ir_graph, block, n, inD, mode);
456 /* fourth step: create the Store */
457 store = new_rd_Store(db, current_ir_graph, block, phiM, ptr, phiD);
458 projM = new_rd_Proj(NULL, current_ir_graph, block, store, mode_M, pn_Store_M);
460 info = get_ldst_info(store, wenv);
461 info->projs[pn_Store_M] = projM;
463 /* fifths step: repair exception flow */
465 ir_node *projX = new_rd_Proj(NULL, current_ir_graph, block, store, mode_X, pn_Store_X_except);
467 info->projs[pn_Store_X_except] = projX;
468 info->exc_block = exc;
469 info->exc_idx = idx[0];
471 for (i = 0; i < n; ++i) {
472 set_Block_cfgpred(exc, idx[i], projX);
476 /* the exception block should be optimized as some inputs are identical now */
480 /* sixt step: replace old Phi */
481 exchange(phi, projM);
487 * walker, collects all Load/Store/Proj nodes
489 static void do_load_store_optimize(ir_node *n, void *env)
491 walk_env_t *wenv = env;
493 switch (get_irn_opcode(n)) {
496 wenv->changes |= optimize_load(n);
500 wenv->changes |= optimize_store(n);
504 wenv->changes |= optimize_phi(n, env);
512 * do the load store optimization
514 void optimize_load_store(ir_graph *irg)
518 assert(get_irg_phase_state(irg) != phase_building);
520 if (!get_opt_redundant_LoadStore())
523 obstack_init(&env.obst);
526 /* init the links, then collect Loads/Stores/Proj's in lists */
527 irg_walk_graph(irg, init_links, collect_nodes, &env);
529 /* now we have collected enough information, optimize */
530 irg_walk_graph(irg, NULL, do_load_store_optimize, &env);
532 obstack_free(&env.obst, NULL);
534 /* Handle graph state */
536 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
537 set_irg_outs_inconsistent(current_ir_graph);
539 /* is this really needed: Yes, as exception block may get bad but this might be tested */
540 if (get_irg_dom_state(current_ir_graph) == dom_consistent)
541 set_irg_dom_inconsistent(current_ir_graph);