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 Optimization of function calls.
23 * @author Michael Beck
31 #include "irgraph_t.h"
34 #include "dbginfo_t.h"
38 #include "iredges_t.h"
40 #include "iroptimize.h"
41 #include "analyze_irg_args.h"
43 #include "raw_bitset.h"
46 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
49 * The walker environment for updating function calls.
51 typedef struct env_t {
52 unsigned n_calls_SymConst;
54 ir_node *float_const_call_list; /**< The list of all floating const function calls that will be changed. */
55 ir_node *nonfloat_const_call_list; /**< The list of all non-floating const function calls that will be changed. */
56 ir_node *pure_call_list; /**< The list of all pure function calls that will be changed. */
57 ir_node *nothrow_call_list; /**< The list of all nothrow function calls that will be changed. */
58 ir_node *proj_list; /**< The list of all potential Proj nodes that must be fixed. */
61 /** If non-null, evaluates entities for being a heap alloc. */
62 static check_alloc_entity_func is_alloc_entity = NULL;
64 /** Ready IRG's are marked in the ready set. */
65 static unsigned *ready_set;
67 /** IRG's that are in progress are marked here. */
68 static unsigned *busy_set;
71 * We misuse the mtp_property_inherited flag as temporary here.
72 * The is ok, as we cannot set or get it anyway using the
73 * get_addtional_properties API.
75 #define mtp_temporary mtp_property_inherited
78 * Walker: Collect all calls to const and pure functions
79 * to lists. Collect all Proj(Call) nodes into a Proj list.
81 static void collect_const_and_pure_calls(ir_node *node, void *env)
86 unsigned and_prop, or_prop, prop;
91 /* set the link to NULL for all non-const/pure calls */
92 set_irn_link(call, NULL);
93 ptr = get_Call_ptr(call);
95 ent = get_Global_entity(ptr);
97 prop = get_entity_additional_properties(ent);
98 if ((prop & (mtp_property_const|mtp_property_pure)) == 0)
100 ++ctx->n_calls_SymConst;
101 } else if (get_opt_closed_world() &&
103 get_irg_callee_info_state(current_ir_graph) == irg_callee_info_consistent) {
104 /* If all possible callees are const functions, we can remove the memory edge. */
105 int i, n_callees = get_Call_n_callees(call);
106 if (n_callees == 0) {
107 /* This is kind of strange: dying code or a Call that will raise an exception
108 when executed as there is no implementation to call. So better not
113 /* note that const function are a subset of pure ones */
114 and_prop = mtp_property_const | mtp_property_pure;
116 for (i = 0; i < n_callees; ++i) {
117 ent = get_Call_callee(call, i);
118 if (ent == unknown_entity) {
119 /* we don't know which entity is called here */
122 prop = get_entity_additional_properties(ent);
125 if (and_prop == mtp_no_property)
128 prop = and_prop | (or_prop & mtp_property_has_loop);
133 /* ok, if we get here we found a call to a const or a pure function */
134 if (prop & mtp_property_pure) {
135 set_irn_link(call, ctx->pure_call_list);
136 ctx->pure_call_list = call;
138 if (prop & mtp_property_has_loop) {
139 set_irn_link(call, ctx->nonfloat_const_call_list);
140 ctx->nonfloat_const_call_list = call;
142 set_irn_link(call, ctx->float_const_call_list);
143 ctx->float_const_call_list = call;
146 } else if (is_Proj(node)) {
148 * Collect all memory and exception Proj's from
151 call = get_Proj_pred(node);
155 /* collect the Proj's in the Proj list */
156 switch (get_Proj_proj(node)) {
158 case pn_Call_X_except:
159 case pn_Call_X_regular:
160 set_irn_link(node, ctx->proj_list);
161 ctx->proj_list = node;
167 } /* collect_const_and_pure_calls */
170 * Fix the list of collected Calls.
172 * @param irg the graph that contained calls to pure functions
175 static void fix_const_call_lists(ir_graph *irg, env_t *ctx)
177 ir_node *call, *next, *mem, *proj;
179 ir_graph *rem = current_ir_graph;
181 current_ir_graph = irg;
183 /* First step: fix all calls by removing their memory input and let
185 * The original memory input is preserved in their link fields. */
186 for (call = ctx->float_const_call_list; call != NULL; call = next) {
187 next = get_irn_link(call);
188 mem = get_Call_mem(call);
190 set_irn_link(call, mem);
191 set_Call_mem(call, get_irg_no_mem(irg));
194 * Unfortunately we cannot simply set the node to 'float'.
195 * There is a reason for that:
197 * - The call might be inside a loop/if that is NOT entered
198 * and calls a endless function. Setting the call to float
199 * would allow to move it out from the loop/if causing this
200 * function be called even if the loop/if is not entered ...
202 * This could be fixed using post-dominators for calls and Pin nodes
203 * but need some more analyzes to ensure that a call that potential
204 * never returns is not executed before some code that generates
205 * observable states...
208 /* finally, this call can float */
209 set_irn_pinned(call, op_pin_state_floats);
210 hook_func_call(irg, call);
213 /* Last step: fix all Proj's */
214 for (proj = ctx->proj_list; proj != NULL; proj = next) {
215 next = get_irn_link(proj);
216 call = get_Proj_pred(proj);
217 mem = get_irn_link(call);
219 /* beware of calls in the pure call list */
220 if (!mem || is_Call(mem))
222 assert(get_irn_mode(mem) == mode_M);
224 switch (get_Proj_proj(proj)) {
226 /* in dead code there might be cycles where proj == mem */
231 case pn_Call_X_except:
233 exchange(proj, get_irg_bad(irg));
235 case pn_Call_X_regular: {
236 ir_node *block = get_nodes_block(call);
238 exchange(proj, new_r_Jmp(block));
246 /* changes were done ... */
247 set_irg_outs_inconsistent(irg);
248 set_irg_loopinfo_state(irg, loopinfo_cf_inconsistent);
251 /* ... including exception edges */
252 set_irg_doms_inconsistent(irg);
254 current_ir_graph = rem;
255 } /* fix_const_call_list */
258 * Walker: Collect all calls to nothrow functions
259 * to lists. Collect all Proj(Call) nodes into a Proj list.
261 static void collect_nothrow_calls(ir_node *node, void *env)
271 /* set the link to NULL for all non-const/pure calls */
272 set_irn_link(call, NULL);
273 ptr = get_Call_ptr(call);
274 if (is_Global(ptr)) {
275 ent = get_Global_entity(ptr);
277 prop = get_entity_additional_properties(ent);
278 if ((prop & mtp_property_nothrow) == 0)
280 ++ctx->n_calls_SymConst;
281 } else if (get_opt_closed_world() &&
283 get_irg_callee_info_state(current_ir_graph) == irg_callee_info_consistent) {
284 /* If all possible callees are nothrow functions, we can remove the exception edge. */
285 int i, n_callees = get_Call_n_callees(call);
286 if (n_callees == 0) {
287 /* This is kind of strange: dying code or a Call that will raise an exception
288 when executed as there is no implementation to call. So better not
293 /* note that const function are a subset of pure ones */
294 prop = mtp_property_nothrow;
295 for (i = 0; i < n_callees; ++i) {
296 ent = get_Call_callee(call, i);
297 if (ent == unknown_entity) {
298 /* we don't know which entity is called here */
301 prop &= get_entity_additional_properties(ent);
302 if (prop == mtp_no_property)
309 /* ok, if we get here we found a call to a nothrow function */
310 set_irn_link(call, ctx->nothrow_call_list);
311 ctx->nothrow_call_list = call;
312 } else if (is_Proj(node)) {
314 * Collect all memory and exception Proj's from
317 call = get_Proj_pred(node);
321 /* collect the Proj's in the Proj list */
322 switch (get_Proj_proj(node)) {
324 case pn_Call_X_except:
325 case pn_Call_X_regular:
326 set_irn_link(node, ctx->proj_list);
327 ctx->proj_list = node;
333 } /* collect_nothrow_calls */
336 * Fix the list of collected nothrow Calls.
338 * @param irg the graph that contained calls to pure functions
339 * @param call_list the list of all call sites of const functions
340 * @param proj_list the list of all memory/exception Proj's of this call sites
342 static void fix_nothrow_call_list(ir_graph *irg, ir_node *call_list, ir_node *proj_list)
344 ir_node *call, *next, *proj;
346 ir_graph *rem = current_ir_graph;
348 current_ir_graph = irg;
350 /* First step: go through the list of calls and mark them. */
351 for (call = call_list; call; call = next) {
352 next = get_irn_link(call);
354 /* current_ir_graph is in memory anyway, so it's a good marker */
355 set_irn_link(call, ¤t_ir_graph);
356 hook_func_call(irg, call);
359 /* Second step: Remove all exception Proj's */
360 for (proj = proj_list; proj; proj = next) {
361 next = get_irn_link(proj);
362 call = get_Proj_pred(proj);
364 /* handle only marked calls */
365 if (get_irn_link(call) != ¤t_ir_graph)
368 /* kill any exception flow */
369 switch (get_Proj_proj(proj)) {
370 case pn_Call_X_except:
372 exchange(proj, get_irg_bad(irg));
374 case pn_Call_X_regular: {
375 ir_node *block = get_nodes_block(call);
377 exchange(proj, new_r_Jmp(block));
385 /* changes were done ... */
386 set_irg_outs_inconsistent(irg);
387 set_irg_loopinfo_state(irg, loopinfo_cf_inconsistent);
390 /* ... including exception edges */
391 set_irg_doms_inconsistent(irg);
393 current_ir_graph = rem;
394 } /* fix_nothrow_call_list */
397 #define SET_IRG_READY(irg) rbitset_set(ready_set, get_irg_idx(irg))
398 #define IS_IRG_READY(irg) rbitset_is_set(ready_set, get_irg_idx(irg))
399 #define SET_IRG_BUSY(irg) rbitset_set(busy_set, get_irg_idx(irg))
400 #define CLEAR_IRG_BUSY(irg) rbitset_clear(busy_set, get_irg_idx(irg))
401 #define IS_IRG_BUSY(irg) rbitset_is_set(busy_set, get_irg_idx(irg))
404 static unsigned check_const_or_pure_function(ir_graph *irg, int top);
407 * Calculate the bigger property of two. Handle the temporary flag right.
409 static unsigned max_property(unsigned a, unsigned b)
411 unsigned r, t = (a | b) & mtp_temporary;
415 if (a == mtp_no_property || b == mtp_no_property)
416 return mtp_no_property;
422 * Follow the memory chain starting at node and determine
425 * @return mtp_property_const if only calls of const functions are detected
426 * mtp_property_pure if only Loads and const/pure calls detected
427 * mtp_no_property else
429 static unsigned _follow_mem(ir_node *node)
431 unsigned m, mode = mtp_property_const;
436 if (mode == mtp_no_property)
437 return mtp_no_property;
439 if (irn_visited_else_mark(node))
442 switch (get_irn_opcode(node)) {
444 node = get_Proj_pred(node);
453 /* do a dfs search */
454 for (i = get_irn_arity(node) - 1; i >= 0; --i) {
455 m = _follow_mem(get_irn_n(node, i));
456 mode = max_property(mode, m);
457 if (mode == mtp_no_property)
458 return mtp_no_property;
463 /* Beware volatile Loads are NOT allowed in pure functions. */
464 if (get_Load_volatility(node) == volatility_is_volatile)
465 return mtp_no_property;
466 mode = max_property(mode, mtp_property_pure);
467 node = get_Load_mem(node);
471 /* A call is only tolerable if its either constant or pure. */
472 ptr = get_Call_ptr(node);
473 if (is_SymConst_addr_ent(ptr)) {
474 ir_entity *ent = get_SymConst_entity(ptr);
475 ir_graph *irg = get_entity_irg(ent);
477 if (irg == current_ir_graph) {
478 /* A self-recursive call. The property did not depend on this call. */
479 } else if (irg == NULL) {
480 m = get_entity_additional_properties(ent) & (mtp_property_const|mtp_property_pure);
481 mode = max_property(mode, m);
482 } else if (irg != NULL) {
483 /* we have a graph, analyze it. */
484 m = check_const_or_pure_function(irg, /*top=*/0);
485 mode = max_property(mode, m);
488 return mtp_no_property;
489 node = get_Call_mem(node);
493 return mtp_no_property;
499 * Follow the memory chain starting at node and determine
502 * @return mtp_property_const if only calls of const functions are detected
503 * mtp_property_pure if only Loads and const/pure calls detected
504 * mtp_no_property else
506 static unsigned follow_mem(ir_node *node, unsigned mode)
510 m = _follow_mem(node);
511 return max_property(mode, m);
515 * Check if a graph represents a const or a pure function.
517 * @param irg the graph to check
518 * @param top if set, this is the top call
520 static unsigned check_const_or_pure_function(ir_graph *irg, int top)
522 ir_node *end, *endbl;
524 unsigned prop = get_irg_additional_properties(irg);
525 ir_graph *rem = current_ir_graph;
527 if (prop & mtp_property_const) {
528 /* already marked as a const function */
529 return mtp_property_const;
531 if (prop & mtp_property_pure) {
532 /* already marked as a pure function */
533 return mtp_property_pure;
536 if (IS_IRG_READY(irg)) {
537 /* already checked */
538 return mtp_no_property;
540 if (IS_IRG_BUSY(irg)) {
541 /* we are still evaluate this method. Be optimistic,
542 return the best possible so far but mark the result as temporary. */
543 return mtp_temporary | mtp_property_const;
547 end = get_irg_end(irg);
548 endbl = get_nodes_block(end);
549 prop = mtp_property_const;
551 current_ir_graph = irg;
553 ir_reserve_resources(irg, IR_RESOURCE_IRN_VISITED);
554 inc_irg_visited(irg);
555 /* mark the initial mem: recursion of follow_mem() stops here */
556 mark_irn_visited(get_irg_initial_mem(irg));
558 /* visit every Return */
559 for (j = get_Block_n_cfgpreds(endbl) - 1; j >= 0; --j) {
560 ir_node *node = get_Block_cfgpred(endbl, j);
561 ir_opcode code = get_irn_opcode(node);
564 /* Bad nodes usually do NOT produce anything, so it's ok */
568 if (code == iro_Return) {
569 mem = get_Return_mem(node);
571 /* Bad nodes usually do NOT produce anything, so it's ok */
575 if (mem != get_irg_initial_mem(irg))
576 prop = max_property(prop, follow_mem(mem, prop));
578 /* Exception found. Cannot be const or pure. */
579 prop = mtp_no_property;
582 if (prop == mtp_no_property)
586 if (prop != mtp_no_property) {
587 /* check, if a keep-alive exists */
588 for (j = get_End_n_keepalives(end) - 1; j >= 0; --j) {
589 ir_node *kept = get_End_keepalive(end, j);
591 if (is_Block(kept)) {
592 prop = mtp_no_property;
596 if (mode_M != get_irn_mode(kept))
599 prop = max_property(prop, follow_mem(kept, prop));
600 if (prop == mtp_no_property)
605 if (prop != mtp_no_property) {
606 if (top || (prop & mtp_temporary) == 0) {
607 /* We use the temporary flag here to mark optimistic result.
608 Set the property only if we are sure that it does NOT base on
609 temporary results OR if we are at top-level. */
610 set_irg_additional_property(irg, prop & ~mtp_temporary);
617 ir_free_resources(irg, IR_RESOURCE_IRN_VISITED);
618 current_ir_graph = rem;
620 } /* check_const_or_pure_function */
623 * Handle calls to const functions.
627 static void handle_const_Calls(env_t *ctx)
631 ctx->n_calls_SymConst = 0;
632 ctx->n_calls_Sel = 0;
634 /* all calls of const functions can be transformed */
635 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
636 ir_graph *irg = get_irp_irg(i);
638 ctx->float_const_call_list = NULL;
639 ctx->nonfloat_const_call_list = NULL;
640 ctx->pure_call_list = NULL;
641 ctx->proj_list = NULL;
643 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
644 irg_walk_graph(irg, NULL, collect_const_and_pure_calls, ctx);
646 if (ctx->float_const_call_list != NULL)
647 fix_const_call_lists(irg, ctx);
648 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
650 } /* handle_const_Calls */
653 * Handle calls to nothrow functions.
657 static void handle_nothrow_Calls(env_t *ctx)
661 ctx->n_calls_SymConst = 0;
662 ctx->n_calls_Sel = 0;
664 /* all calls of const functions can be transformed */
665 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
666 ir_graph *irg = get_irp_irg(i);
668 ctx->nothrow_call_list = NULL;
669 ctx->proj_list = NULL;
671 ir_reserve_resources(irg, IR_RESOURCE_IRN_LINK);
672 irg_walk_graph(irg, NULL, collect_nothrow_calls, ctx);
674 if (ctx->nothrow_call_list)
675 fix_nothrow_call_list(irg, ctx->nothrow_call_list, ctx->proj_list);
676 ir_free_resources(irg, IR_RESOURCE_IRN_LINK);
681 * Check, whether a given node represents a return value of
682 * a malloc like function (ie, new heap allocated memory).
684 * @param node the node to check
686 static int is_malloc_call_result(const ir_node *node)
688 if (is_Alloc(node) && get_Alloc_where(node) == heap_alloc) {
689 /* Firm style high-level allocation */
692 if (is_alloc_entity != NULL && is_Call(node)) {
693 ir_node *ptr = get_Call_ptr(node);
695 if (is_Global(ptr)) {
696 ir_entity *ent = get_Global_entity(ptr);
697 return is_alloc_entity(ent);
701 } /* is_malloc_call_result */
704 * Update a property depending on a call property.
706 static unsigned update_property(unsigned orig_prop, unsigned call_prop)
708 unsigned t = (orig_prop | call_prop) & mtp_temporary;
709 unsigned r = orig_prop & call_prop;
711 } /** update_property */
714 * Check if a node is stored.
716 static int is_stored(const ir_node *n)
718 const ir_edge_t *edge;
721 foreach_out_edge(n, edge) {
722 const ir_node *succ = get_edge_src_irn(edge);
724 switch (get_irn_opcode(succ)) {
731 if (get_Store_value(succ) == n)
733 /* ok if its only the address input */
742 ptr = get_Call_ptr(succ);
743 if (is_Global(ptr)) {
744 ir_entity *ent = get_Global_entity(ptr);
747 /* we know the called entity */
748 for (i = get_Call_n_params(succ) - 1; i >= 0; --i) {
749 if (get_Call_param(succ, i) == n) {
750 /* n is the i'th param of the call */
751 if (get_method_param_access(ent, i) & ptr_access_store) {
752 /* n is store in ent */
758 /* unknown call address */
763 /* bad, potential alias */
771 * Check that the return value of an irg is not stored anywhere.
773 * return ~mtp_property_malloc if return values are stored, ~0 else
775 static unsigned check_stored_result(ir_graph *irg)
777 ir_node *end_blk = get_irg_end_block(irg);
780 int old_edges = edges_assure_kind(irg, EDGE_KIND_NORMAL);
782 for (i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) {
783 ir_node *pred = get_Block_cfgpred(end_blk, i);
785 if (! is_Return(pred))
787 for (j = get_Return_n_ress(pred) - 1; j >= 0; --j) {
788 const ir_node *irn = get_Return_res(pred, j);
790 if (is_stored(irn)) {
791 /* bad, might create an alias */
792 res = ~mtp_property_malloc;
799 edges_deactivate_kind(irg, EDGE_KIND_NORMAL);
801 } /* check_stored_result */
804 * Check if a graph represents a nothrow or a malloc function.
806 * @param irg the graph to check
807 * @param top if set, this is the top call
809 static unsigned check_nothrow_or_malloc(ir_graph *irg, int top)
811 ir_node *end_blk = get_irg_end_block(irg);
815 unsigned curr_prop = mtp_property_malloc | mtp_property_nothrow;
817 if (IS_IRG_READY(irg)) {
818 /* already checked */
819 return get_irg_additional_properties(irg);
821 if (IS_IRG_BUSY(irg)) {
822 /* we are still evaluate this method. Be optimistic,
823 return the best possible so far but mark the result as temporary. */
824 return mtp_temporary | mtp_property_malloc | mtp_property_nothrow;
828 ent = get_irg_entity(irg);
829 mtp = get_entity_type(ent);
831 if (get_method_n_ress(mtp) <= 0)
832 curr_prop &= ~mtp_property_malloc;
834 for (i = get_Block_n_cfgpreds(end_blk) - 1; i >= 0; --i) {
835 ir_node *pred = get_Block_cfgpred(end_blk, i);
837 if (is_Return(pred)) {
838 if (curr_prop & mtp_property_malloc) {
839 /* check, if malloc is called here */
840 for (j = get_Return_n_ress(pred) - 1; j >= 0; --j) {
841 ir_node *res = get_Return_res(pred, j);
843 /* skip Confirms and Casts */
844 res = skip_HighLevel_ops(res);
847 res = get_Proj_pred(res);
848 if (is_malloc_call_result(res)) {
849 /* ok, this is a malloc */
850 } else if (is_Call(res)) {
851 ir_node *ptr = get_Call_ptr(res);
853 if (is_Global(ptr)) {
855 ir_entity *ent = get_Global_entity(ptr);
856 ir_graph *callee = get_entity_irg(ent);
859 /* A self-recursive call. The property did not depend on this call. */
860 } else if (callee != NULL) {
861 unsigned prop = check_nothrow_or_malloc(callee, /*top=*/0);
862 curr_prop = update_property(curr_prop, prop);
864 curr_prop = update_property(curr_prop, get_entity_additional_properties(ent));
866 } else if (get_opt_closed_world() &&
868 get_irg_callee_info_state(irg) == irg_callee_info_consistent) {
869 /* check if all possible callees are malloc functions. */
870 int i, n_callees = get_Call_n_callees(res);
871 if (n_callees == 0) {
872 /* This is kind of strange: dying code or a Call that will raise an exception
873 when executed as there is no implementation to call. So better not
875 curr_prop &= ~mtp_property_malloc;
879 for (i = 0; i < n_callees; ++i) {
880 ir_entity *ent = get_Call_callee(res, i);
881 if (ent == unknown_entity) {
882 /* we don't know which entity is called here */
883 curr_prop &= ~mtp_property_malloc;
886 if ((get_entity_additional_properties(ent) & mtp_property_malloc) == 0) {
887 curr_prop &= ~mtp_property_malloc;
891 /* if we pass the for cycle, malloc is still ok */
894 curr_prop &= ~mtp_property_malloc;
897 /* unknown return value */
898 curr_prop &= ~mtp_property_malloc;
902 } else if (curr_prop & mtp_property_nothrow) {
903 /* exception flow detected */
904 pred = skip_Proj(pred);
907 ir_node *ptr = get_Call_ptr(pred);
909 if (is_Global(ptr)) {
911 ir_entity *ent = get_Global_entity(ptr);
912 ir_graph *callee = get_entity_irg(ent);
915 /* A self-recursive call. The property did not depend on this call. */
916 } else if (callee != NULL) {
917 /* Note: we check here for nothrow only, so do NOT reset the malloc property */
918 unsigned prop = check_nothrow_or_malloc(callee, /*top=*/0) | mtp_property_malloc;
919 curr_prop = update_property(curr_prop, prop);
921 if ((get_entity_additional_properties(ent) & mtp_property_nothrow) == 0)
922 curr_prop &= ~mtp_property_nothrow;
924 } else if (get_opt_closed_world() &&
926 get_irg_callee_info_state(irg) == irg_callee_info_consistent) {
927 /* check if all possible callees are nothrow functions. */
928 int i, n_callees = get_Call_n_callees(pred);
929 if (n_callees == 0) {
930 /* This is kind of strange: dying code or a Call that will raise an exception
931 when executed as there is no implementation to call. So better not
933 curr_prop &= ~mtp_property_nothrow;
937 for (i = 0; i < n_callees; ++i) {
938 ir_entity *ent = get_Call_callee(pred, i);
939 if (ent == unknown_entity) {
940 /* we don't know which entity is called here */
941 curr_prop &= ~mtp_property_nothrow;
944 if ((get_entity_additional_properties(ent) & mtp_property_nothrow) == 0) {
945 curr_prop &= ~mtp_property_nothrow;
949 /* if we pass the for cycle, nothrow is still ok */
952 curr_prop &= ~mtp_property_nothrow;
955 /* real exception flow possible. */
956 curr_prop &= ~mtp_property_nothrow;
959 if ((curr_prop & ~mtp_temporary) == mtp_no_property) {
960 /* no need to search further */
965 if (curr_prop & mtp_property_malloc) {
967 * Note that the malloc property means not only return newly allocated
968 * memory, but also that this memory is ALIAS FREE.
969 * To ensure that, we do NOT allow that the returned memory is somewhere
972 curr_prop &= check_stored_result(irg);
975 if (curr_prop != mtp_no_property) {
976 if (top || (curr_prop & mtp_temporary) == 0) {
977 /* We use the temporary flag here to mark an optimistic result.
978 Set the property only if we are sure that it does NOT base on
979 temporary results OR if we are at top-level. */
980 set_irg_additional_property(irg, curr_prop & ~mtp_temporary);
988 } /* check_nothrow_or_malloc */
991 * When a function was detected as "const", it might be moved out of loops.
992 * This might be dangerous if the graph can contain endless loops.
994 static void check_for_possible_endless_loops(ir_graph *irg)
999 root_loop = get_irg_loop(irg);
1000 if (root_loop->flags & loop_outer_loop)
1001 set_irg_additional_property(irg, mtp_property_has_loop);
1005 * optimize function calls by handling const functions
1007 void optimize_funccalls(int force_run, check_alloc_entity_func callback)
1010 unsigned num_const = 0;
1011 unsigned num_pure = 0;
1012 unsigned num_nothrow = 0;
1013 unsigned num_malloc = 0;
1015 is_alloc_entity = callback;
1017 /* prepare: mark all graphs as not analyzed */
1018 last_idx = get_irp_last_idx();
1019 ready_set = rbitset_malloc(last_idx);
1020 busy_set = rbitset_malloc(last_idx);
1022 /* first step: detect, which functions are nothrow or malloc */
1023 DB((dbg, LEVEL_2, "Detecting nothrow and malloc properties ...\n"));
1024 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1025 ir_graph *irg = get_irp_irg(i);
1026 unsigned prop = check_nothrow_or_malloc(irg, /*top=*/1);
1028 if (prop & mtp_property_nothrow) {
1030 DB((dbg, LEVEL_2, "%+F has the nothrow property\n", irg));
1031 } else if (prop & mtp_property_malloc) {
1033 DB((dbg, LEVEL_2, "%+F has the malloc property\n", irg));
1037 /* second step: remove exception edges: this must be done before the
1038 detection of const and pure functions take place. */
1039 if (force_run || num_nothrow > 0) {
1042 handle_nothrow_Calls(&ctx);
1043 DB((dbg, LEVEL_1, "Detected %u nothrow graphs, %u malloc graphs.\n", num_nothrow, num_malloc));
1044 DB((dbg, LEVEL_1, "Optimizes %u(SymConst) + %u(Sel) calls to nothrow functions.\n",
1045 ctx.n_calls_SymConst, ctx.n_calls_Sel));
1047 DB((dbg, LEVEL_1, "No graphs without side effects detected\n"));
1050 rbitset_clear_all(ready_set, last_idx);
1051 rbitset_clear_all(busy_set, last_idx);
1053 /* third step: detect, which functions are const or pure */
1054 DB((dbg, LEVEL_2, "Detecting const and pure properties ...\n"));
1055 for (i = get_irp_n_irgs() - 1; i >= 0; --i) {
1056 ir_graph *irg = get_irp_irg(i);
1057 unsigned prop = check_const_or_pure_function(irg, /*top=*/1);
1059 if (prop & mtp_property_const) {
1061 DB((dbg, LEVEL_2, "%+F has the const property\n", irg));
1062 check_for_possible_endless_loops(irg);
1063 } else if (prop & mtp_property_pure) {
1065 DB((dbg, LEVEL_2, "%+F has the pure property\n", irg));
1069 if (force_run || num_const > 0) {
1072 handle_const_Calls(&ctx);
1073 DB((dbg, LEVEL_1, "Detected %u const graphs, %u pure graphs.\n", num_const, num_pure));
1074 DB((dbg, LEVEL_1, "Optimizes %u(SymConst) + %u(Sel) calls to const functions.\n",
1075 ctx.n_calls_SymConst, ctx.n_calls_Sel));
1077 DB((dbg, LEVEL_1, "No graphs without side effects detected\n"));
1081 } /* optimize_funccalls */
1083 /* initialize the funccall optimization */
1084 void firm_init_funccalls(void)
1086 FIRM_DBG_REGISTER(dbg, "firm.opt.funccalls");
1087 } /* firm_init_funccalls */
1090 ir_prog_pass_t pass;
1092 check_alloc_entity_func callback;
1096 * Wrapper for running optimize_funccalls() as an ir_prog pass.
1098 static int pass_wrapper(ir_prog *irp, void *context)
1100 struct pass_t *pass = context;
1103 optimize_funccalls(pass->force_run, pass->callback);
1105 } /* pass_wrapper */
1107 /* Creates an ir_prog pass for optimize_funccalls. */
1108 ir_prog_pass_t *optimize_funccalls_pass(
1110 int force_run, check_alloc_entity_func callback)
1112 struct pass_t *pass = XMALLOCZ(struct pass_t);
1114 pass->force_run = force_run;
1115 pass->callback = callback;
1117 return def_prog_pass_constructor(
1118 &pass->pass, name ? name : "funccall", pass_wrapper);
1119 } /* optimize_funccalls_pass */