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 Available Optimisations of libFirm.
25 #ifndef FIRM_IROPTIMIZE_H
26 #define FIRM_IROPTIMIZE_H
28 #include "firm_types.h"
31 * Control flow optimization.
33 * Removes empty blocks doing if simplifications and loop simplifications.
34 * A block is empty if it contains only a Jmp node and Phi nodes.
35 * Merges single entry single exit blocks with their predecessor
36 * and propagates dead control flow by calling equivalent_node().
37 * Independent of compiler flag it removes Tuples from cf edges,
38 * Bad predecessors from Blocks and Phis, and unnecessary predecessors of End.
40 * @bug So far destroys backedge information.
41 * @bug Chokes on Id nodes if called in a certain order with other
42 * optimizations. Call local_optimize_graph() before to remove
45 void optimize_cf(ir_graph *irg);
48 * Creates an ir_graph pass for optimize_cf().
50 * @param name the name of this pass or NULL
51 * @param verify should this pass be verified?
52 * @param dump should this pass result be dumped?
54 * @return the newly created ir_graph pass
56 ir_graph_pass_t *optimize_cf_pass(const char *name, int verify, int dump);
59 * Perform path-sensitive jump threading on the given graph.
61 * @param irg the graph
63 void opt_jumpthreading(ir_graph* irg);
66 * Creates an ir_graph pass for opt_jumpthreading().
68 * @param name the name of this pass or NULL
70 * @return the newly created ir_graph pass
72 ir_graph_pass_t *opt_jumpthreading_pass(const char *name);
75 * Try to simplify boolean expression in the given ir graph.
76 * eg. x < 5 && x < 6 becomes x < 5
78 * @param irg the graph
80 void opt_bool(ir_graph *irg);
83 * Creates an ir_graph pass for opt_bool().
85 * @param name the name of this pass or NULL
87 * @return the newly created ir_graph pass
89 ir_graph_pass_t *opt_bool_pass(const char *name);
92 * Try to reduce the number of conv nodes in the given ir graph.
94 * @param irg the graph
96 * @return non-zero if the optimization could be applied, 0 else
98 int conv_opt(ir_graph *irg);
101 * Creates an ir_graph pass for conv_opt().
103 * @param name the name of this pass or NULL
105 * @return the newly created ir_graph pass
107 ir_graph_pass_t *conv_opt_pass(const char *name);
110 * Do the scalar replacement optimization.
111 * Make a date flow analyze and split the
114 * @param irg the graph which should be optimized
116 void data_flow_scalar_replacement_opt(ir_graph *irg);
119 * A callback that checks whether a entity is an allocation
122 typedef int (*check_alloc_entity_func)(ir_entity *ent);
125 * Do simple and fast escape analysis for one graph.
127 * @param irg the graph
128 * @param callback a callback function to check whether a
129 * given entity is a allocation call
131 void escape_enalysis_irg(ir_graph *irg, check_alloc_entity_func callback);
134 * Do simple and fast escape analysis for all graphs.
136 * This optimization implements a simple and fast but inexact
137 * escape analysis. Some addresses might be marked as 'escaped' even
139 * The advantage is a low memory footprint and fast speed.
141 * @param run_scalar_replace if this flag in non-zero, scalar
142 * replacement optimization is run on graphs with removed
144 * @param callback a callback function to check whether a
145 * given entity is a allocation call
147 * This optimization removes allocation which are not used (rare) and replace
148 * allocation that can be proved dead at the end of the graph which stack variables.
150 * The creation of stack variable allows scalar replacement to be run only
151 * on those graphs that have been changed.
153 * This is most effective on Java where no other stack variables exists.
155 void escape_analysis(int run_scalar_replace, check_alloc_entity_func callback);
158 * Optimize function calls by handling const functions.
160 * This optimization first detects all "const functions", i.e.,
161 * IR graphs that neither read nor write memory (and hence did
162 * not create exceptions, as these use memory in Firm).
164 * The result of calls to such functions depends only on its
165 * arguments, hence those calls are no more pinned.
167 * This is a rather strong criteria, so do not expect that a
168 * lot of functions will be found. Moreover, all of them might
169 * already be inlined if inlining is activated.
170 * Anyway, it might be good for handling builtin's or pseudo-graphs,
171 * even if the later read/write memory (but we know how).
173 * This optimizations read the irg_const_function property of
174 * entities and and sets the irg_const_function property of
177 * If callee information is valid, we also optimize polymorphic Calls.
179 * @param force_run if non-zero, an optimization run is started even
180 * if no const function graph was detected.
181 * Else calls are only optimized if at least one
182 * const function graph was detected.
183 * @param callback a callback function to check whether a
184 * given entity is a allocation call
186 * If the frontend created external entities with the irg_const_function
187 * property set, the force_run parameter should be set, else
190 * @note This optimization destroys the link fields of nodes.
192 void optimize_funccalls(int force_run, check_alloc_entity_func callback);
195 * Creates an ir_prog pass for optimize_funccalls().
197 * @param name the name of this pass or NULL
198 * @param force_run if non-zero, an optimization run is started even
199 * if no const function graph was detected.
200 * Else calls are only optimized if at least one
201 * const function graph was detected.
202 * @param callback a callback function to check whether a
203 * given entity is a allocation call
205 * @return the newly created ir_prog pass
207 ir_prog_pass_t *optimize_funccalls_pass(
209 int force_run, check_alloc_entity_func callback);
212 * Does Partial Redundancy Elimination combined with
213 * Global Value Numbering.
214 * Can be used to replace place_code() completely.
216 * Based on VanDrunen and Hosking 2004.
218 * @param irg the graph
220 void do_gvn_pre(ir_graph *irg);
223 * Creates an ir_graph pass for do_gvn_pre().
225 * @param dump should this pass result be dumped?
227 * @return the newly created ir_graph pass
229 ir_graph_pass_t *do_gvn_pre_pass(const char *name);
232 * This function is called to evaluate, if a mux can build
233 * of the current architecture.
234 * If it returns non-zero, a mux is created, else the code
236 * @param sel A selector of a Cond.
237 * @param phi_list List of Phi nodes about to be converted (linked via get_Phi_next() field)
238 * @param i First data predecessor involved in if conversion
239 * @param j Second data predecessor involved in if conversion
241 typedef int (*arch_allow_ifconv_func)(ir_node *sel, ir_node* phi_list, int i, int j);
244 * The parameters structure.
246 struct ir_settings_if_conv_t {
247 int max_depth; /**< The maximum depth up to which expressions
248 are examined when it has to be decided if they
249 can be placed into another block. */
250 arch_allow_ifconv_func allow_ifconv; /**< Evaluator function, if not set all possible Psi
251 nodes will be created. */
255 * Perform If conversion on a graph.
257 * @param irg The graph.
258 * @param params The parameters for the if conversion.
260 * Cannot handle blocks with Bad control predecessors, so call it after control
263 void opt_if_conv(ir_graph *irg, const ir_settings_if_conv_t *params);
266 * Creates an ir_graph pass for opt_if_conv().
268 * @param name the name of this pass or NULL
269 * @param params The parameters for the if conversion.
271 * @return the newly created ir_graph pass
273 ir_graph_pass_t *opt_if_conv_pass(
274 const char *name, const ir_settings_if_conv_t *params);
276 void opt_sync(ir_graph *irg);
279 * Creates an ir_graph pass for opt_sync().
281 * @param name the name of this pass or NULL
283 * @return the newly created ir_graph pass
285 ir_graph_pass_t *opt_sync_pass(const char *name);
288 * Check if we can replace the load by a given const from
289 * the const code irg.
291 * @param load the load to replace
292 * @param c the constant
294 * @return in the modes match or can be transformed using a reinterpret cast
295 * returns a copy of the constant (possibly Conv'ed) on the
298 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c);
301 * Load/Store optimization.
303 * Removes redundant non-volatile Loads and Stores.
304 * May introduce Bad nodes if exceptional control flow
305 * is removed. The following cases are optimized:
307 * Load without result: A Load which has only a memory use
310 * Load after Store: A Load after a Store is removed, if
311 * the Load doesn't have an exception handler OR is in
312 * the same block as the Store.
314 * Load after Load: A Load after a Load is removed, if the
315 * Load doesn't have an exception handler OR is in the
316 * same block as the previous Load.
318 * Store before Store: A Store immediately before another
319 * Store in the same block is removed, if the Store doesn't
320 * have an exception handler.
322 * Store after Load: A Store after a Load is removed, if the
323 * Store doesn't have an exception handler.
325 * @return non-zero if the optimization could be applied, 0 else
327 int optimize_load_store(ir_graph *irg);
330 * Creates an ir_graph pass for optimize_load_store().
332 * @param name the name of this pass or NULL
334 * @return the newly created ir_graph pass
336 ir_graph_pass_t *optimize_load_store_pass(const char *name);
339 * New experimental alternative to optimize_load_store.
340 * Based on a dataflow analysis, so load/stores are moved out of loops
343 int opt_ldst(ir_graph *irg);
346 * Creates an ir_graph pass for opt_ldst().
348 * @param name the name of this pass or NULL
350 * @return the newly created ir_graph pass
352 ir_graph_pass_t *opt_ldst_pass(const char *name);
355 * Do Loop unrolling in the given graph.
357 void optimize_loop_unrolling(ir_graph *irg);
360 * Creates an ir_graph pass for optimize_loop_unrolling().
362 * @param name the name of this pass or NULL
364 * @return the newly created ir_graph pass
366 ir_graph_pass_t *optimize_loop_unrolling_pass(const char *name);
369 * Optimize the frame type of an irg by removing
370 * never touched entities.
372 * @param irg The graph whose frame type will be optimized
374 * This function did not change the graph, only it's frame type.
375 * The layout state of the frame type will be set to layout_undefined
376 * if entities were removed.
378 void opt_frame_irg(ir_graph *irg);
381 * Creates an ir_graph pass for opt_frame_irg().
383 * @param name the name of this pass or NULL
385 * @return the newly created ir_graph pass
387 ir_graph_pass_t *opt_frame_irg_pass(const char *name);
389 /** Possible flags for the Operator Scalar Replacement. */
390 typedef enum osr_flags {
391 osr_flag_none = 0, /**< no additional flags */
392 osr_flag_lftr_with_ov_check = 1, /**< do linear function test replacement
393 only if no overflow can occur. */
394 osr_flag_ignore_x86_shift = 2, /**< ignore Multiplications by 2, 4, 8 */
395 osr_flag_keep_reg_pressure = 4 /**< do NOT increase register pressure by introducing new
396 induction variables. */
399 /* FirmJNI cannot handle identical enum values... */
401 /** default setting */
402 #define osr_flag_default osr_flag_lftr_with_ov_check
405 * Do the Operator Scalar Replacement optimization and linear
406 * function test replacement for loop control.
407 * Can be switched off using the set_opt_strength_red() flag.
408 * In that case, only remove_phi_cycles() is executed.
410 * @param irg the graph which should be optimized
411 * @param flags set of osr_flags
413 * The linear function replacement test is controlled by the flags.
414 * If the osr_flag_lftr_with_ov_check is set, the replacement is only
415 * done if do overflow can occur.
416 * Otherwise it is ALWAYS done which might be insecure.
420 * for (i = 0; i < 100; ++i)
422 * might be replaced by
424 * for (i = 0; i < 400; i += 4)
428 * for (i = 0; i < 0x7FFFFFFF; ++i)
430 * will not be replaced by
432 * for (i = 0; i < 0xFFFFFFFC; i += 4)
434 * because of overflow.
438 * for (i = 0; i <= 0xF; ++i)
440 * will NOT be transformed into
442 * for (i = 0xFFFFFFF0; i <= 0xFFFFFFFF; ++i)
444 * although here is no direct overflow. The OV occurs when the ++i
445 * is executed (and would created an endless loop here!).
447 * For the same reason, a loop
449 * for (i = 0; i <= 9; i += x)
451 * will NOT be transformed because we cannot estimate whether an overflow
452 * might happen adding x.
454 * Note that i < a + 400 is also not possible with the current implementation
455 * although this might be allowed by other compilers...
457 * Note further that tests for equality can be handled some simpler (but are not
460 * This algorithm destroys the link field of nodes.
462 void opt_osr(ir_graph *irg, unsigned flags);
465 * Creates an ir_graph pass for remove_phi_cycles().
467 * @param name the name of this pass or NULL
468 * @param flags set of osr_flags
470 * @return the newly created ir_graph pass
472 ir_graph_pass_t *opt_osr_pass(const char *name, unsigned flags);
475 * Removes useless Phi cycles, i.e cycles of Phi nodes with only one
477 * This is automatically done in opt_osr(), so there is no need to call it
480 * @param irg the graph which should be optimized
482 * This algorithm destroys the link field of nodes.
484 void remove_phi_cycles(ir_graph *irg);
487 * Creates an ir_graph pass for remove_phi_cycles().
489 * @param name the name of this pass or NULL
491 * @return the newly created ir_graph pass
493 ir_graph_pass_t *remove_phi_cycles_pass(const char *name);
496 /** A default threshold. */
497 #define DEFAULT_CLONE_THRESHOLD 300
500 * Do procedure cloning. Evaluate a heuristic weight for every
501 * Call(..., Const, ...). If the weight is bigger than threshold,
502 * clone the entity and fix the calls.
504 * @param threshold the threshold for cloning
506 * The threshold is an estimation of how many instructions are saved
507 * when executing a cloned method. If threshold is 0.0, every possible
510 void proc_cloning(float threshold);
515 * Applies Reassociation rules to integer expressions.
516 * Beware: Works only if integer overflow might be ignored, as for C, Java
517 * and for address expression.
518 * Works only if Constant folding is activated.
520 * Uses loop information to detect loop-invariant (ie contant
521 * inside the loop) values.
523 * See Muchnik 12.3.1 Algebraic Simplification and Reassociation of
524 * Addressing Expressions.
526 * @return non-zero if the optimization could be applied, 0 else
528 int optimize_reassociation(ir_graph *irg);
531 * Creates an ir_graph pass for optimize_reassociation().
533 * @param name the name of this pass or NULL
535 * @return the newly created ir_graph pass
537 ir_graph_pass_t *optimize_reassociation_pass(const char *name);
540 * Normalize the Returns of a graph by creating a new End block
541 * with One Return(Phi).
542 * This is the preferred input for the if-conversion.
544 * In pseudocode, it means:
551 * is transformed into
559 void normalize_one_return(ir_graph *irg);
562 * Creates an ir_graph pass for normalize_one_return().
564 * @param name the name of this pass or NULL
566 * @return the newly created ir_graph pass
568 ir_graph_pass_t *normalize_one_return_pass(const char *name);
571 * Normalize the Returns of a graph by moving
572 * the Returns upwards as much as possible.
573 * This might be preferred for code generation.
575 * In pseudocode, it means:
583 * is transformed into
590 void normalize_n_returns(ir_graph *irg);
593 * Creates an ir_graph pass for normalize_n_returns().
595 * @param name the name of this pass or NULL
597 * @return the newly created ir_graph pass
599 ir_graph_pass_t *normalize_n_returns_pass(const char *name);
602 * Do the scalar replacement optimization.
603 * Replace local compound entities (like structures and arrays)
604 * with atomic values if possible. Does not handle classes yet.
606 * @param irg the graph which should be optimized
608 * @return non-zero, if at least one entity was replaced
610 int scalar_replacement_opt(ir_graph *irg);
613 * Creates an ir_graph pass for scalar_replacement_opt().
615 * @param name the name of this pass or NULL
617 * @return the newly created ir_graph pass
619 ir_graph_pass_t *scalar_replacement_opt_pass(const char *name);
621 /** Performs strength reduction for the passed graph. */
622 void reduce_strength(ir_graph *irg);
625 * Optimizes tail-recursion calls by converting them into loops.
626 * Depends on the flag opt_tail_recursion.
627 * Currently supports the following forms:
629 * - return x + func();
630 * - return func() - x;
631 * - return x * func();
634 * Does not work for Calls that use the exception stuff.
636 * @param irg the graph to be optimized
638 * @return non-zero if the optimization could be applied, 0 else
640 int opt_tail_rec_irg(ir_graph *irg);
643 * Creates an ir_graph pass for opt_tail_rec_irg().
645 * @param name the name of this pass or NULL
647 * @return the newly created ir_graph pass
649 ir_graph_pass_t *opt_tail_rec_irg_pass(const char *name);
652 * Optimize tail-recursion calls for all IR-Graphs.
653 * Can currently handle:
654 * - direct return value, i.e. return func().
655 * - additive return value, i.e. return x +/- func()
656 * - multiplicative return value, i.e. return x * func() or return -func()
658 * The current implementation must be run before optimize_funccalls(),
659 * because it expects the memory edges pointing to calls, which might be
660 * removed by optimize_funccalls().
662 void opt_tail_recursion(void);
665 * Creates an ir_prog pass for opt_tail_recursion().
667 * @param name the name of this pass or NULL
669 * @return the newly created ir_prog pass
671 ir_prog_pass_t *opt_tail_recursion_pass(const char *name);
673 /** This is the type for a method, that returns a pointer type to
674 * tp. This is needed in the normalization. */
675 typedef ir_type *(*gen_pointer_type_to_func)(ir_type *tp);
677 /** Insert Casts so that class type casts conform exactly with the type hierarchy.
679 * Formulated in Java, this achieves the following:
681 * For a class hierarchy
683 * class B extends A {}
684 * class C extends B {}
685 * we transforms a cast
690 * The algorithm works for Casts with class types, but also for Casts
691 * with all pointer types that point (over several indirections,
692 * i.e. ***A) to a class type. Normalizes all graphs. Computes type
693 * information (@see irtypeinfo.h) if not available.
694 * Invalidates trout information as new casts are generated.
696 * @param gppt_fct A function that returns a pointer type that points
697 * to the type given as argument. If this parameter is NULL, a default
698 * function is used that either uses trout information or performs a O(n)
699 * search to find an existing pointer type. If it can not find a type,
700 * generates a pointer type with mode_P_mach and suffix "cc_ptr_tp".
702 void normalize_irp_class_casts(gen_pointer_type_to_func gppt_fct);
704 /** Insert Casts so that class type casts conform exactly with the type hierarchy
707 * For more details see normalize_irp_class_casts().
709 * This transformation requires that type information is computed. @see irtypeinfo.h.
711 void normalize_irg_class_casts(ir_graph *irg, gen_pointer_type_to_func gppt_fct);
713 /** Optimize casting between class types.
716 * class B extends A { }
717 * class C extends B {}
718 * Performs the following transformations:
719 * C c = (C)(B)(A)(B)new C() --> C c = (C)(B)newC() --> C c = new C()
720 * (Optimizing downcasts as A a = (A)(B)(new A()) --> A a = new A() can
721 * be suppressed by setting the flag opt_suppress_downcast_optimization.
722 * Downcasting A to B might cause an exception. It is not clear
723 * whether this is modeled by the Firm Cast node, as it has no exception
725 * If there is inh_m() that overwrites m() in B:
726 * ((A) new B()).m() --> (new B()).inh_m()
727 * Phi((A)x, (A)y) --> (A) Phi (x, y) if (A) is an upcast.
729 * Computes type information if not available. @see irtypeinfo.h.
730 * Typeinformation is valid after optimization.
731 * Invalidates trout information.
733 void optimize_class_casts(void);
736 * CLiff Click's combo algorithm from "Combining Analyses, combining Optimizations".
738 * Does conditional constant propagation, unreachable code elimination and optimistic
739 * global value numbering at once.
741 * @param irg the graph to run on
743 void combo(ir_graph *irg);
746 * Creates an ir_graph pass for combo.
748 * @param name the name of this pass or NULL
750 * @return the newly created ir_graph pass
752 ir_graph_pass_t *combo_pass(const char *name);
754 /** Inlines all small methods at call sites where the called address comes
755 * from a SymConst node that references the entity representing the called
758 * The size argument is a rough measure for the code size of the method:
759 * Methods where the obstack containing the firm graph is smaller than
760 * size are inlined. Further only a limited number of calls are inlined.
761 * If the method contains more than 1024 inlineable calls none will be
763 * Inlining is only performed if flags `optimize' and `inlineing' are set.
764 * The graph may not be in state phase_building.
765 * It is recommended to call local_optimize_graph() after inlining as this
766 * function leaves a set of obscure Tuple nodes, e.g. a Proj-Tuple-Jmp
767 * combination as control flow operation.
769 void inline_small_irgs(ir_graph *irg, int size);
772 /** Inlineing with a different heuristic than inline_small_irgs().
774 * Inlines leave functions. If inlinening creates new leave
775 * function inlines these, too. (If g calls f, and f calls leave h,
776 * h is first inlined in f and then f in g.)
778 * Then inlines all small functions (this is not recursive).
780 * For a heuristic this inlineing uses firm node counts. It does
781 * not count auxiliary nodes as Proj, Tuple, End, Start, Id, Sync.
782 * If the ignore_runtime flag is set, calls to functions marked with the
783 * mtp_property_runtime property are ignored.
785 * @param maxsize Do not inline any calls if a method has more than
786 * maxsize firm nodes. It may reach this limit by
788 * @param leavesize Inline leave functions if they have less than leavesize
790 * @param size Inline all function smaller than size.
791 * @param ignore_runtime count a function only calling runtime functions as
794 void inline_leave_functions(unsigned maxsize, unsigned leavesize,
795 unsigned size, int ignore_runtime);
798 * Heuristic inliner. Calculates a benefice value for every call and inlines
799 * those calls with a value higher than the threshold.
801 * @param maxsize Do not inline any calls if a method has more than
802 * maxsize firm nodes. It may reach this limit by
804 * @param threshold inlining threshold
806 void inline_functions(unsigned maxsize, int inline_threshold);
809 * Combines congruent blocks into one.
811 * @param irg The IR-graph to optimize.
813 * @return non-zero if the graph was transformed
815 int shape_blocks(ir_graph *irg);