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"
33 * Control flow optimization.
35 * Removes empty blocks doing if simplifications and loop simplifications.
36 * A block is empty if it contains only a Jmp node and Phi nodes.
37 * Merges single entry single exit blocks with their predecessor
38 * and propagates dead control flow by calling equivalent_node().
39 * Independent of compiler flag it removes Tuples from cf edges,
40 * Bad predecessors from Blocks and Phis, and unnecessary predecessors of End.
41 * Destroys backedge information.
43 * @bug Chokes on Id nodes if called in a certain order with other
44 * optimizations. Call local_optimize_graph() before to remove
47 FIRM_API void optimize_cf(ir_graph *irg);
50 * Creates an ir_graph pass for optimize_cf().
52 * @param name the name of this pass or NULL
54 * @return the newly created ir_graph pass
56 FIRM_API ir_graph_pass_t *optimize_cf_pass(const char *name);
59 * Perform path-sensitive jump threading on the given graph.
61 * @param irg the graph
63 FIRM_API 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 FIRM_API ir_graph_pass_t *opt_jumpthreading_pass(const char *name);
75 * Creates an ir_graph pass for opt_loopunroll().
77 * @param name the name of this pass or NULL
79 * @return the newly created ir_graph pass
81 FIRM_API ir_graph_pass_t *opt_loopunroll_pass(const char *name);
85 * Try to simplify boolean expression in the given ir graph.
86 * eg. x < 5 && x < 6 becomes x < 5
88 * @param irg the graph
90 FIRM_API void opt_bool(ir_graph *irg);
93 * Creates an ir_graph pass for opt_bool().
95 * @param name the name of this pass or NULL
97 * @return the newly created ir_graph pass
99 FIRM_API ir_graph_pass_t *opt_bool_pass(const char *name);
102 * Try to reduce the number of conv nodes in the given ir graph.
104 * @param irg the graph
106 * @return non-zero if the optimization could be applied, 0 else
108 FIRM_API int conv_opt(ir_graph *irg);
111 * Creates an ir_graph pass for conv_opt().
113 * @param name the name of this pass or NULL
115 * @return the newly created ir_graph pass
117 FIRM_API ir_graph_pass_t *conv_opt_pass(const char *name);
120 * Do the scalar replacement optimization.
121 * Make a date flow analyze and split the
124 * @param irg the graph which should be optimized
126 FIRM_API void data_flow_scalar_replacement_opt(ir_graph *irg);
129 * A callback that checks whether a entity is an allocation
132 typedef int (*check_alloc_entity_func)(ir_entity *ent);
135 * Do simple and fast escape analysis for one graph.
137 * @param irg the graph
138 * @param callback a callback function to check whether a
139 * given entity is a allocation call
141 FIRM_API void escape_enalysis_irg(ir_graph *irg,
142 check_alloc_entity_func callback);
145 * Do simple and fast escape analysis for all graphs.
147 * This optimization implements a simple and fast but inexact
148 * escape analysis. Some addresses might be marked as 'escaped' even
150 * The advantage is a low memory footprint and fast speed.
152 * @param run_scalar_replace if this flag in non-zero, scalar
153 * replacement optimization is run on graphs with removed
155 * @param callback a callback function to check whether a
156 * given entity is a allocation call
158 * This optimization removes allocation which are not used (rare) and replace
159 * allocation that can be proved dead at the end of the graph which stack variables.
161 * The creation of stack variable allows scalar replacement to be run only
162 * on those graphs that have been changed.
164 * This is most effective on Java where no other stack variables exists.
166 FIRM_API void escape_analysis(int run_scalar_replace,
167 check_alloc_entity_func callback);
170 * Optimize function calls by handling const functions.
172 * This optimization first detects all "const functions", i.e.,
173 * IR graphs that neither read nor write memory (and hence did
174 * not create exceptions, as these use memory in Firm).
176 * The result of calls to such functions depends only on its
177 * arguments, hence those calls are no more pinned.
179 * This is a rather strong criteria, so do not expect that a
180 * lot of functions will be found. Moreover, all of them might
181 * already be inlined if inlining is activated.
182 * Anyway, it might be good for handling builtin's
183 * even if the later read/write memory (but we know how).
185 * This optimizations read the irg_const_function property of
186 * entities and and sets the irg_const_function property of
189 * If callee information is valid, we also optimize polymorphic Calls.
191 * @param force_run if non-zero, an optimization run is started even
192 * if no const function graph was detected.
193 * Else calls are only optimized if at least one
194 * const function graph was detected.
195 * @param callback a callback function to check whether a
196 * given entity is a allocation call
198 * If the frontend created external entities with the irg_const_function
199 * property set, the force_run parameter should be set, else
202 * @note This optimization destroys the link fields of nodes.
204 FIRM_API void optimize_funccalls(int force_run,
205 check_alloc_entity_func callback);
208 * Creates an ir_prog pass for optimize_funccalls().
210 * @param name the name of this pass or NULL
211 * @param force_run if non-zero, an optimization run is started even
212 * if no const function graph was detected.
213 * Else calls are only optimized if at least one
214 * const function graph was detected.
215 * @param callback a callback function to check whether a
216 * given entity is a allocation call
218 * @return the newly created ir_prog pass
220 FIRM_API ir_prog_pass_t *optimize_funccalls_pass(const char *name,
222 check_alloc_entity_func callback);
225 * Does Partial Redundancy Elimination combined with
226 * Global Value Numbering.
227 * Can be used to replace place_code() completely.
229 * Based on VanDrunen and Hosking 2004.
231 * @param irg the graph
233 FIRM_API void do_gvn_pre(ir_graph *irg);
236 * Creates an ir_graph pass for do_gvn_pre().
238 * @param name the name of this pass or NULL
240 * @return the newly created ir_graph pass
242 FIRM_API ir_graph_pass_t *do_gvn_pre_pass(const char *name);
245 * This function is called to evaluate, if a
246 * mux(@p sel, @p mux_false, @p mux_true) should be built for the current
248 * If it returns non-zero, a mux is created, else the code
250 * @param sel A selector of a Cond.
251 * @param phi_list phi node to be converted
252 * @param i First data predecessor involved in if conversion
253 * @param j Second data predecessor involved in if conversion
255 typedef int (*arch_allow_ifconv_func)(ir_node *sel, ir_node *mux_false,
259 * Perform If conversion on a graph.
261 * @param irg The graph.
263 * Cannot handle blocks with Bad control predecessors, so call it after control
266 FIRM_API void opt_if_conv(ir_graph *irg);
269 * Creates an ir_graph pass for opt_if_conv().
271 * @param name the name of this pass or NULL
273 * @return the newly created ir_graph pass
275 FIRM_API ir_graph_pass_t *opt_if_conv_pass(const char *name);
278 * Tries to reduce dependencies for memory nodes where possible by parllelizing
279 * them and synchronising with Sync nodes
280 * @param irg the graph where memory operations should be parallelised
282 FIRM_API void opt_parallelize_mem(ir_graph *irg);
285 * Creates an ir_graph pass for opt_sync().
287 * @param name the name of this pass or NULL
289 * @return the newly created ir_graph pass
291 FIRM_API ir_graph_pass_t *opt_parallelize_mem_pass(const char *name);
294 * Check if we can replace the load by a given const from
295 * the const code irg.
297 * @param load the load to replace
298 * @param c the constant
300 * @return in the modes match or can be transformed using a reinterpret cast
301 * returns a copy of the constant (possibly Conv'ed) on the
304 FIRM_API ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c);
307 * Load/Store optimization.
309 * Removes redundant non-volatile Loads and Stores.
310 * May introduce Bad nodes if exceptional control flow
311 * is removed. The following cases are optimized:
313 * Load without result: A Load which has only a memory use
316 * Load after Store: A Load after a Store is removed, if
317 * the Load doesn't have an exception handler OR is in
318 * the same block as the Store.
320 * Load after Load: A Load after a Load is removed, if the
321 * Load doesn't have an exception handler OR is in the
322 * same block as the previous Load.
324 * Store before Store: A Store immediately before another
325 * Store in the same block is removed, if the Store doesn't
326 * have an exception handler.
328 * Store after Load: A Store after a Load is removed, if the
329 * Store doesn't have an exception handler.
331 * @return non-zero if the optimization could be applied, 0 else
333 FIRM_API int optimize_load_store(ir_graph *irg);
336 * Creates an ir_graph pass for optimize_load_store().
338 * @param name the name of this pass or NULL
340 * @return the newly created ir_graph pass
342 FIRM_API ir_graph_pass_t *optimize_load_store_pass(const char *name);
345 * New experimental alternative to optimize_load_store.
346 * Based on a dataflow analysis, so load/stores are moved out of loops
349 FIRM_API int opt_ldst(ir_graph *irg);
352 * Creates an ir_graph pass for opt_ldst().
354 * @param name the name of this pass or NULL
356 * @return the newly created ir_graph pass
358 FIRM_API ir_graph_pass_t *opt_ldst_pass(const char *name);
361 * Optimize loops by peeling or unrolling them if beneficial.
363 * @param irg The graph whose loops will be processed
365 * This function did not change the graph, only it's frame type.
366 * The layout state of the frame type will be set to layout_undefined
367 * if entities were removed.
369 FIRM_API void loop_optimization(ir_graph *irg);
372 * Optimize the frame type of an irg by removing
373 * never touched entities.
375 * @param irg The graph whose frame type will be optimized
377 * This function did not change the graph, only it's frame type.
378 * The layout state of the frame type will be set to layout_undefined
379 * if entities were removed.
381 FIRM_API void opt_frame_irg(ir_graph *irg);
384 * Creates an ir_graph pass for opt_frame_irg().
386 * @param name the name of this pass or NULL
388 * @return the newly created ir_graph pass
390 FIRM_API ir_graph_pass_t *opt_frame_irg_pass(const char *name);
392 /** Possible flags for the Operator Scalar Replacement. */
393 typedef enum osr_flags {
394 osr_flag_none = 0, /**< no additional flags */
395 osr_flag_lftr_with_ov_check = 1, /**< do linear function test replacement
396 only if no overflow can occur. */
397 osr_flag_ignore_x86_shift = 2, /**< ignore Multiplications by 2, 4, 8 */
398 osr_flag_keep_reg_pressure = 4 /**< do NOT increase register pressure by introducing new
399 induction variables. */
402 /* FirmJNI cannot handle identical enum values... */
404 /** default setting */
405 #define osr_flag_default osr_flag_lftr_with_ov_check
408 * Do the Operator Scalar Replacement optimization and linear
409 * function test replacement for loop control.
410 * Can be switched off using the set_opt_strength_red() flag.
411 * In that case, only remove_phi_cycles() is executed.
413 * @param irg the graph which should be optimized
414 * @param flags set of osr_flags
416 * The linear function replacement test is controlled by the flags.
417 * If the osr_flag_lftr_with_ov_check is set, the replacement is only
418 * done if do overflow can occur.
419 * Otherwise it is ALWAYS done which might be insecure.
423 * for (i = 0; i < 100; ++i)
425 * might be replaced by
427 * for (i = 0; i < 400; i += 4)
431 * for (i = 0; i < 0x7FFFFFFF; ++i)
433 * will not be replaced by
435 * for (i = 0; i < 0xFFFFFFFC; i += 4)
437 * because of overflow.
441 * for (i = 0; i <= 0xF; ++i)
443 * will NOT be transformed into
445 * for (i = 0xFFFFFFF0; i <= 0xFFFFFFFF; ++i)
447 * although here is no direct overflow. The OV occurs when the ++i
448 * is executed (and would created an endless loop here!).
450 * For the same reason, a loop
452 * for (i = 0; i <= 9; i += x)
454 * will NOT be transformed because we cannot estimate whether an overflow
455 * might happen adding x.
457 * Note that i < a + 400 is also not possible with the current implementation
458 * although this might be allowed by other compilers...
460 * Note further that tests for equality can be handled some simpler (but are not
463 * This algorithm destroys the link field of nodes.
465 FIRM_API void opt_osr(ir_graph *irg, unsigned flags);
468 * Creates an ir_graph pass for remove_phi_cycles().
470 * @param name the name of this pass or NULL
471 * @param flags set of osr_flags
473 * @return the newly created ir_graph pass
475 FIRM_API ir_graph_pass_t *opt_osr_pass(const char *name, unsigned flags);
478 * Removes useless Phi cycles, i.e cycles of Phi nodes with only one
480 * This is automatically done in opt_osr(), so there is no need to call it
483 * @param irg the graph which should be optimized
485 * This algorithm destroys the link field of nodes.
487 FIRM_API void remove_phi_cycles(ir_graph *irg);
490 * Creates an ir_graph pass for remove_phi_cycles().
492 * @param name the name of this pass or NULL
494 * @return the newly created ir_graph pass
496 FIRM_API ir_graph_pass_t *remove_phi_cycles_pass(const char *name);
499 /** A default threshold. */
500 #define DEFAULT_CLONE_THRESHOLD 20
503 * Do procedure cloning. Evaluate a heuristic weight for every
504 * Call(..., Const, ...). If the weight is bigger than threshold,
505 * clone the entity and fix the calls.
507 * @param threshold the threshold for cloning
509 * The threshold is an estimation of how many instructions are saved
510 * when executing a cloned method. If threshold is 0.0, every possible
513 FIRM_API void proc_cloning(float threshold);
516 * Creates an ir_prog pass for proc_cloning().
518 * @param name the name of this pass or NULL
519 * @param threshold the threshold for cloning
521 * @return the newly created ir_prog pass
523 FIRM_API ir_prog_pass_t *proc_cloning_pass(const char *name, float threshold);
528 * Applies Reassociation rules to integer expressions.
529 * Beware: Works only if integer overflow might be ignored, as for C, Java
530 * and for address expression.
531 * Works only if Constant folding is activated.
533 * Uses loop information to detect loop-invariant (ie contant
534 * inside the loop) values.
536 * See Muchnik 12.3.1 Algebraic Simplification and Reassociation of
537 * Addressing Expressions.
539 * @return non-zero if the optimization could be applied, 0 else
541 FIRM_API int optimize_reassociation(ir_graph *irg);
544 * Creates an ir_graph pass for optimize_reassociation().
546 * @param name the name of this pass or NULL
548 * @return the newly created ir_graph pass
550 FIRM_API ir_graph_pass_t *optimize_reassociation_pass(const char *name);
553 * Normalize the Returns of a graph by creating a new End block
554 * with One Return(Phi).
555 * This is the preferred input for the if-conversion.
557 * In pseudocode, it means:
564 * is transformed into
572 FIRM_API void normalize_one_return(ir_graph *irg);
575 * Creates an ir_graph pass for normalize_one_return().
577 * @param name the name of this pass or NULL
579 * @return the newly created ir_graph pass
581 FIRM_API ir_graph_pass_t *normalize_one_return_pass(const char *name);
584 * Normalize the Returns of a graph by moving
585 * the Returns upwards as much as possible.
586 * This might be preferred for code generation.
588 * In pseudocode, it means:
596 * is transformed into
603 FIRM_API void normalize_n_returns(ir_graph *irg);
606 * Creates an ir_graph pass for normalize_n_returns().
608 * @param name the name of this pass or NULL
610 * @return the newly created ir_graph pass
612 FIRM_API ir_graph_pass_t *normalize_n_returns_pass(const char *name);
615 * Do the scalar replacement optimization.
616 * Replace local compound entities (like structures and arrays)
617 * with atomic values if possible. Does not handle classes yet.
619 * @param irg the graph which should be optimized
621 * @return non-zero, if at least one entity was replaced
623 FIRM_API int scalar_replacement_opt(ir_graph *irg);
626 * Creates an ir_graph pass for scalar_replacement_opt().
628 * @param name the name of this pass or NULL
630 * @return the newly created ir_graph pass
632 FIRM_API ir_graph_pass_t *scalar_replacement_opt_pass(const char *name);
634 /** Performs strength reduction for the passed graph. */
635 FIRM_API void reduce_strength(ir_graph *irg);
638 * Optimizes tail-recursion calls by converting them into loops.
639 * Depends on the flag opt_tail_recursion.
640 * Currently supports the following forms:
642 * - return x + func();
643 * - return func() - x;
644 * - return x * func();
647 * Does not work for Calls that use the exception stuff.
649 * @param irg the graph to be optimized
651 * @return non-zero if the optimization could be applied, 0 else
653 FIRM_API int opt_tail_rec_irg(ir_graph *irg);
656 * Creates an ir_graph pass for opt_tail_rec_irg().
658 * @param name the name of this pass or NULL
660 * @return the newly created ir_graph pass
662 FIRM_API ir_graph_pass_t *opt_tail_rec_irg_pass(const char *name);
665 * Optimize tail-recursion calls for all IR-Graphs.
666 * Can currently handle:
667 * - direct return value, i.e. return func().
668 * - additive return value, i.e. return x +/- func()
669 * - multiplicative return value, i.e. return x * func() or return -func()
671 * The current implementation must be run before optimize_funccalls(),
672 * because it expects the memory edges pointing to calls, which might be
673 * removed by optimize_funccalls().
675 FIRM_API void opt_tail_recursion(void);
678 * Creates an ir_prog pass for opt_tail_recursion().
680 * @param name the name of this pass or NULL
682 * @return the newly created ir_prog pass
684 FIRM_API ir_prog_pass_t *opt_tail_recursion_pass(const char *name);
686 /** This is the type for a method, that returns a pointer type to
687 * tp. This is needed in the normalization. */
688 typedef ir_type *(*gen_pointer_type_to_func)(ir_type *tp);
690 /** Insert Casts so that class type casts conform exactly with the type hierarchy.
692 * Formulated in Java, this achieves the following:
694 * For a class hierarchy
696 * class B extends A {}
697 * class C extends B {}
698 * we transforms a cast
703 * The algorithm works for Casts with class types, but also for Casts
704 * with all pointer types that point (over several indirections,
705 * i.e. ***A) to a class type. Normalizes all graphs. Computes type
706 * information (@see irtypeinfo.h) if not available.
707 * Invalidates trout information as new casts are generated.
709 * @param gppt_fct A function that returns a pointer type that points
710 * to the type given as argument. If this parameter is NULL, a default
711 * function is used that either uses trout information or performs a O(n)
712 * search to find an existing pointer type. If it can not find a type,
713 * generates a pointer type with mode_P_mach and suffix "cc_ptr_tp".
715 FIRM_API void normalize_irp_class_casts(gen_pointer_type_to_func gppt_fct);
717 /** Insert Casts so that class type casts conform exactly with the type hierarchy
720 * For more details see normalize_irp_class_casts().
722 * This transformation requires that type information is computed. @see irtypeinfo.h.
724 FIRM_API void normalize_irg_class_casts(ir_graph *irg,
725 gen_pointer_type_to_func gppt_fct);
727 /** Optimize casting between class types.
730 * class B extends A { }
731 * class C extends B {}
732 * Performs the following transformations:
733 * C c = (C)(B)(A)(B)new C() --> C c = (C)(B)newC() --> C c = new C()
734 * (Optimizing downcasts as A a = (A)(B)(new A()) --> A a = new A() can
735 * be suppressed by setting the flag opt_suppress_downcast_optimization.
736 * Downcasting A to B might cause an exception. It is not clear
737 * whether this is modeled by the Firm Cast node, as it has no exception
739 * If there is inh_m() that overwrites m() in B:
740 * ((A) new B()).m() --> (new B()).inh_m()
741 * Phi((A)x, (A)y) --> (A) Phi (x, y) if (A) is an upcast.
743 * Computes type information if not available. @see irtypeinfo.h.
744 * Typeinformation is valid after optimization.
745 * Invalidates trout information.
747 FIRM_API void optimize_class_casts(void);
750 * CLiff Click's combo algorithm from
751 * "Combining Analyses, combining Optimizations".
753 * Does conditional constant propagation, unreachable code elimination and
754 * optimistic global value numbering at once.
756 * @param irg the graph to run on
758 FIRM_API void combo(ir_graph *irg);
761 * Creates an ir_graph pass for combo.
763 * @param name the name of this pass or NULL
765 * @return the newly created ir_graph pass
767 FIRM_API ir_graph_pass_t *combo_pass(const char *name);
770 * Inlines all small methods at call sites where the called address comes
771 * from a SymConst node that references the entity representing the called
774 * @param irg the graph
775 * @param size maximum function size
777 * The size argument is a rough measure for the code size of the method:
778 * Methods where the obstack containing the firm graph is smaller than
779 * size are inlined. Further only a limited number of calls are inlined.
780 * If the method contains more than 1024 inlineable calls none will be
782 * Inlining is only performed if flags `optimize' and `inlineing' are set.
783 * The graph may not be in state phase_building.
784 * It is recommended to call local_optimize_graph() after inlining as this
785 * function leaves a set of obscure Tuple nodes, e.g. a Proj-Tuple-Jmp
786 * combination as control flow operation.
788 FIRM_API void inline_small_irgs(ir_graph *irg, int size);
791 * Creates an ir_graph pass for inline_small_irgs().
793 * @param name the name of this pass or NULL
794 * @param size maximum function size
796 * @return the newly created ir_graph pass
798 FIRM_API ir_graph_pass_t *inline_small_irgs_pass(const char *name, int size);
801 * Inlineing with a different heuristic than inline_small_irgs().
803 * Inlines leave functions. If inlinening creates new leave
804 * function inlines these, too. (If g calls f, and f calls leave h,
805 * h is first inlined in f and then f in g.)
807 * Then inlines all small functions (this is not recursive).
809 * For a heuristic this inlineing uses firm node counts. It does
810 * not count auxiliary nodes as Proj, Tuple, End, Start, Id, Sync.
811 * If the ignore_runtime flag is set, calls to functions marked with the
812 * mtp_property_runtime property are ignored.
814 * @param maxsize Do not inline any calls if a method has more than
815 * maxsize firm nodes. It may reach this limit by
817 * @param leavesize Inline leave functions if they have less than leavesize
819 * @param size Inline all function smaller than size.
820 * @param ignore_runtime count a function only calling runtime functions as
823 FIRM_API void inline_leave_functions(unsigned maxsize, unsigned leavesize,
824 unsigned size, int ignore_runtime);
827 * Creates an ir_prog pass for inline_leave_functions().
829 * @param name the name of this pass or NULL
830 * @param maxsize Do not inline any calls if a method has more than
831 * maxsize firm nodes. It may reach this limit by
833 * @param leavesize Inline leave functions if they have less than leavesize
835 * @param size Inline all function smaller than size.
836 * @param ignore_runtime count a function only calling runtime functions as
839 * @return the newly created ir_prog pass
841 FIRM_API ir_prog_pass_t *inline_leave_functions_pass(const char *name,
842 unsigned maxsize, unsigned leavesize, unsigned size,
845 typedef void (*opt_ptr)(ir_graph *irg);
848 * Heuristic inliner. Calculates a benefice value for every call and inlines
849 * those calls with a value higher than the threshold.
851 * @param maxsize Do not inline any calls if a method has more than
852 * maxsize firm nodes. It may reach this limit by
854 * @param inline_threshold inlining threshold
855 * @param after_inline_opt optimizations performed immediately after inlining
858 FIRM_API void inline_functions(unsigned maxsize, int inline_threshold,
859 opt_ptr after_inline_opt);
862 * Creates an ir_prog pass for inline_functions().
864 * @param name the name of this pass or NULL
865 * @param maxsize Do not inline any calls if a method has more than
866 * maxsize firm nodes. It may reach this limit by
868 * @param inline_threshold inlining threshold
869 * @param after_inline_opt a function that is called after inlining a
870 * procedure. You should run fast local optimisations
871 * here which cleanup the graph before further
874 * @return the newly created ir_prog pass
876 FIRM_API ir_prog_pass_t *inline_functions_pass(const char *name,
877 unsigned maxsize, int inline_threshold, opt_ptr after_inline_opt);
880 * Combines congruent blocks into one.
882 * @param irg The IR-graph to optimize.
884 * @return non-zero if the graph was transformed
886 FIRM_API int shape_blocks(ir_graph *irg);
889 * Creates an ir_graph pass for shape_blocks().
891 * @param name the name of this pass or NULL
893 * @return the newly created ir_graph pass
895 FIRM_API ir_graph_pass_t *shape_blocks_pass(const char *name);
898 * Perform loop inversion on a given graph.
899 * Loop inversion transforms a head controlled loop (like while(...) {} and
900 * for(...) {}) into a foot controlled loop (do {} while(...)).
902 FIRM_API void do_loop_inversion(ir_graph *irg);
905 * Perform loop unrolling on a given graph.
906 * Loop unrolling multiplies the number loop completely by a number found
907 * through a heuristic.
909 FIRM_API void do_loop_unrolling(ir_graph *irg);
912 * Perform loop peeling on a given graph.
914 FIRM_API void do_loop_peeling(ir_graph *irg);
917 * Creates an ir_graph pass for loop inversion.
919 * @param name the name of this pass or NULL
921 * @return the newly created ir_graph pass
923 FIRM_API ir_graph_pass_t *loop_inversion_pass(const char *name);
926 * Creates an ir_graph pass for loop unrolling.
928 * @param name the name of this pass or NULL
930 * @return the newly created ir_graph pass
932 FIRM_API ir_graph_pass_t *loop_unroll_pass(const char *name);
935 * Creates an ir_graph pass for loop peeling.
937 * @param name the name of this pass or NULL
939 * @return the newly created ir_graph pass
941 FIRM_API ir_graph_pass_t *loop_peeling_pass(const char *name);
943 typedef ir_type *(*get_Alloc_func)(ir_node *n);
944 /** Set a new get_Alloc_func and returns the old one. */
945 FIRM_API get_Alloc_func firm_set_Alloc_func(get_Alloc_func newf);
948 * Creates an ir_graph pass for set_vrp_data()
950 * @param name The name of this pass or NULL
952 * @return the newly created ir_graph pass
954 FIRM_API ir_graph_pass_t *set_vrp_pass(const char *name);
957 * Removes all entities which are unused.
959 * Unused entities have ir_visibility_local and are not used directly or
960 * indirectly through entities/code visible outside the compilation unit.
961 * This is usually conservative than gc_irgs, but does not respect properties
962 * of object-oriented programs.
964 FIRM_API void garbage_collect_entities(void);
966 /** Pass for garbage_collect_entities */
967 FIRM_API ir_prog_pass_t *garbage_collect_entities_pass(const char *name);
970 * Performs dead node elimination by copying the ir graph to a new obstack.
972 * The major intention of this pass is to free memory occupied by
973 * dead nodes and outdated analyzes information. Further this
974 * function removes Bad predecessors from Blocks and the corresponding
975 * inputs to Phi nodes. This opens optimization potential for other
976 * optimizations. Further this phase reduces dead Block<->Jmp
977 * self-cycles to Bad nodes.
979 * Dead_node_elimination is only performed if options `optimize' and
980 * `opt_dead_node_elimination' are set. The graph may
981 * not be in state phase_building. The outs datasturcture is freed,
982 * the outs state set to outs_none. Backedge information is conserved.
983 * Removes old attributes of nodes. Sets link field to NULL.
984 * Callee information must be freed (irg_callee_info_none).
986 * @param irg The graph to be optimized.
988 FIRM_API void dead_node_elimination(ir_graph *irg);
991 * Creates an ir_graph pass for dead_node_elimination().
993 * @param name the name of this pass or NULL
995 * @return the newly created ir_graph pass
997 FIRM_API ir_graph_pass_t *dead_node_elimination_pass(const char *name);
1000 * Inlines a method at the given call site.
1002 * Removes the call node and splits the basic block the call node
1003 * belongs to. Inserts a copy of the called graph between these nodes.
1004 * Assumes that call is a Call node in current_ir_graph and that
1005 * the type in the Call nodes type attribute is the same as the
1006 * type of the called graph.
1007 * Further it assumes that all Phi nodes in a block of current_ir_graph
1008 * are assembled in a "link" list in the link field of the corresponding
1009 * block nodes. Further assumes that all Proj nodes are in a "link" list
1010 * in the nodes producing the tuple. (This is only an optical feature
1011 * for the graph.) Conserves this feature for the old
1012 * nodes of the graph. This precondition can be established by a call to
1013 * collect_phisprojs(), see irgmod.h.
1014 * As dead_node_elimination this function reduces dead Block<->Jmp
1015 * self-cycles to Bad nodes.
1017 * Called_graph must be unequal to current_ir_graph. Will not inline
1018 * if they are equal.
1019 * Sets visited masterflag in current_ir_graph to the max of the flag in
1020 * current and called graph.
1021 * Assumes that both, the called and the calling graph are in state
1022 * "op_pin_state_pinned".
1023 * It is recommended to call local_optimize_graph() after inlining as this
1024 * function leaves a set of obscure Tuple nodes, e.g. a Proj-Tuple-Jmp
1025 * combination as control flow operation.
1027 * @param call the call node that should be inlined
1028 * @param called_graph the IR-graph that is called at call
1030 * @return zero if method could not be inlined (recursion for instance),
1031 * non-zero if all went ok
1033 FIRM_API int inline_method(ir_node *call, ir_graph *called_graph);
1038 * Pins all floating nodes to a block where they
1039 * will be executed only if needed. Depends on the flag opt_global_cse.
1040 * Graph may not be in phase_building. Does not schedule control dead
1041 * code. Uses dominator information which it computes if the irg is not
1042 * in state dom_consistent. Destroys the out information as it moves nodes
1043 * to other blocks. Optimizes Tuples in Control edges.
1045 * Call remove_critical_cf_edges() before place_code(). This normalizes
1046 * the control flow graph so that for all operations a basic block exists
1047 * where they can be optimally placed.
1049 FIRM_API void place_code(ir_graph *irg);
1052 * Creates an ir_graph pass for place_code().
1053 * This pass enables GCSE, runs optimize_graph_df() and finally
1056 * @param name the name of this pass or NULL
1058 * @return the newly created ir_graph pass
1060 FIRM_API ir_graph_pass_t *place_code_pass(const char *name);
1063 * Determine information about the values of nodes and perform simplications
1064 * using this information. This optimization performs a data-flow analysis to
1065 * find the minimal fixpoint.
1067 FIRM_API void fixpoint_vrp(ir_graph*);
1070 * Creates an ir_graph pass for fixpoint_vrp().
1071 * This pass dDetermines information about the values of nodes
1072 * and perform simplications using this information.
1073 * This optimization performs a data-flow analysis to
1074 * find the minimal fixpoint.
1076 * @param name the name of this pass or NULL
1078 * @return the newly created ir_graph pass
1080 FIRM_API ir_graph_pass_t *fixpoint_vrp_irg_pass(const char *name);
1082 /** Needed for MSVC to suppress warnings because it doest NOT handle const right. */
1083 typedef const ir_node *ir_node_cnst_ptr;
1086 * Check, if the value of a node is != 0.
1088 * This is a often needed case, so we handle here Confirm
1091 * @param n a node representing the value
1092 * @param confirm if n is confirmed to be != 0, returns
1093 * the the Confirm-node, else NULL
1095 FIRM_API int value_not_zero(const ir_node *n, ir_node_cnst_ptr *confirm);
1098 * Check, if the value of a node cannot represent a NULL pointer.
1100 * - If option sel_based_null_check_elim is enabled, all
1101 * Sel nodes can be skipped.
1102 * - A SymConst(entity) is NEVER a NULL pointer
1103 * - A Const != NULL is NEVER a NULL pointer
1104 * - Confirms are evaluated
1106 * @param n a node representing the value
1107 * @param confirm if n is confirmed to be != NULL, returns
1108 * the the Confirm-node, else NULL
1110 FIRM_API int value_not_null(const ir_node *n, ir_node_cnst_ptr *confirm);
1113 * Possible return values of value_classify().
1115 typedef enum ir_value_classify_sign {
1116 value_classified_unknown = 0, /**< could not classify */
1117 value_classified_positive = 1, /**< value is positive, i.e. >= 0 */
1118 value_classified_negative = -1 /**< value is negative, i.e. <= 0 if
1119 no signed zero exists or < 0 else */
1120 } ir_value_classify_sign;
1123 * Check, if the value of a node can be confirmed >= 0 or <= 0,
1124 * If the mode of the value did not honor signed zeros, else
1125 * check for >= 0 or < 0.
1127 * @param n a node representing the value
1129 FIRM_API ir_value_classify_sign classify_value_sign(ir_node *n);
1132 * Return the value of a Cmp if one or both predecessors
1133 * are Confirm nodes.
1135 * @param cmp the compare node that will be evaluated
1136 * @param left the left operand of the Cmp
1137 * @param right the right operand of the Cmp
1138 * @param pnc the compare relation
1140 FIRM_API ir_tarval *computed_value_Cmp_Confirm(
1141 ir_node *cmp, ir_node *left, ir_node *right, pn_Cmp pnc);