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
52 * @return the newly created ir_graph pass
54 ir_graph_pass_t *optimize_cf_pass(const char *name);
57 * Perform path-sensitive jump threading on the given graph.
59 * @param irg the graph
61 void opt_jumpthreading(ir_graph* irg);
64 * Creates an ir_graph pass for opt_jumpthreading().
66 * @param name the name of this pass or NULL
68 * @return the newly created ir_graph pass
70 ir_graph_pass_t *opt_jumpthreading_pass(const char *name);
73 * Try to simplify boolean expression in the given ir graph.
74 * eg. x < 5 && x < 6 becomes x < 5
76 * @param irg the graph
78 void opt_bool(ir_graph *irg);
81 * Creates an ir_graph pass for opt_bool().
83 * @param name the name of this pass or NULL
85 * @return the newly created ir_graph pass
87 ir_graph_pass_t *opt_bool_pass(const char *name);
90 * Try to reduce the number of conv nodes in the given ir graph.
92 * @param irg the graph
94 * @return non-zero if the optimization could be applied, 0 else
96 int conv_opt(ir_graph *irg);
99 * Creates an ir_graph pass for conv_opt().
101 * @param name the name of this pass or NULL
103 * @return the newly created ir_graph pass
105 ir_graph_pass_t *conv_opt_pass(const char *name);
108 * Do the scalar replacement optimization.
109 * Make a date flow analyze and split the
112 * @param irg the graph which should be optimized
114 void data_flow_scalar_replacement_opt(ir_graph *irg);
117 * A callback that checks whether a entity is an allocation
120 typedef int (*check_alloc_entity_func)(ir_entity *ent);
123 * Do simple and fast escape analysis for one graph.
125 * @param irg the graph
126 * @param callback a callback function to check whether a
127 * given entity is a allocation call
129 void escape_enalysis_irg(ir_graph *irg, check_alloc_entity_func callback);
132 * Do simple and fast escape analysis for all graphs.
134 * This optimization implements a simple and fast but inexact
135 * escape analysis. Some addresses might be marked as 'escaped' even
137 * The advantage is a low memory footprint and fast speed.
139 * @param run_scalar_replace if this flag in non-zero, scalar
140 * replacement optimization is run on graphs with removed
142 * @param callback a callback function to check whether a
143 * given entity is a allocation call
145 * This optimization removes allocation which are not used (rare) and replace
146 * allocation that can be proved dead at the end of the graph which stack variables.
148 * The creation of stack variable allows scalar replacement to be run only
149 * on those graphs that have been changed.
151 * This is most effective on Java where no other stack variables exists.
153 void escape_analysis(int run_scalar_replace, check_alloc_entity_func callback);
156 * Optimize function calls by handling const functions.
158 * This optimization first detects all "const functions", i.e.,
159 * IR graphs that neither read nor write memory (and hence did
160 * not create exceptions, as these use memory in Firm).
162 * The result of calls to such functions depends only on its
163 * arguments, hence those calls are no more pinned.
165 * This is a rather strong criteria, so do not expect that a
166 * lot of functions will be found. Moreover, all of them might
167 * already be inlined if inlining is activated.
168 * Anyway, it might be good for handling builtin's or pseudo-graphs,
169 * even if the later read/write memory (but we know how).
171 * This optimizations read the irg_const_function property of
172 * entities and and sets the irg_const_function property of
175 * If callee information is valid, we also optimize polymorphic Calls.
177 * @param force_run if non-zero, an optimization run is started even
178 * if no const function graph was detected.
179 * Else calls are only optimized if at least one
180 * const function graph was detected.
181 * @param callback a callback function to check whether a
182 * given entity is a allocation call
184 * If the frontend created external entities with the irg_const_function
185 * property set, the force_run parameter should be set, else
188 * @note This optimization destroys the link fields of nodes.
190 void optimize_funccalls(int force_run, check_alloc_entity_func callback);
193 * Creates an ir_prog pass for optimize_funccalls().
195 * @param name the name of this pass or NULL
196 * @param force_run if non-zero, an optimization run is started even
197 * if no const function graph was detected.
198 * Else calls are only optimized if at least one
199 * const function graph was detected.
200 * @param callback a callback function to check whether a
201 * given entity is a allocation call
203 * @return the newly created ir_prog pass
205 ir_prog_pass_t *optimize_funccalls_pass(
207 int force_run, check_alloc_entity_func callback);
210 * Does Partial Redundancy Elimination combined with
211 * Global Value Numbering.
212 * Can be used to replace place_code() completely.
214 * Based on VanDrunen and Hosking 2004.
216 * @param irg the graph
218 void do_gvn_pre(ir_graph *irg);
221 * Creates an ir_graph pass for do_gvn_pre().
223 * @param name the name of this pass or NULL
225 * @return the newly created ir_graph pass
227 ir_graph_pass_t *do_gvn_pre_pass(const char *name);
230 * This function is called to evaluate, if a mux can build
231 * of the current architecture.
232 * If it returns non-zero, a mux is created, else the code
234 * @param sel A selector of a Cond.
235 * @param phi_list List of Phi nodes about to be converted (linked via get_Phi_next() field)
236 * @param i First data predecessor involved in if conversion
237 * @param j Second data predecessor involved in if conversion
239 typedef int (*arch_allow_ifconv_func)(ir_node *sel, ir_node* phi_list, int i, int j);
242 * The parameters structure.
244 struct ir_settings_if_conv_t {
245 int max_depth; /**< The maximum depth up to which expressions
246 are examined when it has to be decided if they
247 can be placed into another block. */
248 arch_allow_ifconv_func allow_ifconv; /**< Evaluator function, if not set all possible Psi
249 nodes will be created. */
253 * Perform If conversion on a graph.
255 * @param irg The graph.
256 * @param params The parameters for the if conversion.
258 * Cannot handle blocks with Bad control predecessors, so call it after control
261 void opt_if_conv(ir_graph *irg, const ir_settings_if_conv_t *params);
264 * Creates an ir_graph pass for opt_if_conv().
266 * @param name the name of this pass or NULL
267 * @param params The parameters for the if conversion.
269 * @return the newly created ir_graph pass
271 ir_graph_pass_t *opt_if_conv_pass(
272 const char *name, const ir_settings_if_conv_t *params);
274 void opt_sync(ir_graph *irg);
277 * Creates an ir_graph pass for opt_sync().
279 * @param name the name of this pass or NULL
281 * @return the newly created ir_graph pass
283 ir_graph_pass_t *opt_sync_pass(const char *name);
286 * Check if we can replace the load by a given const from
287 * the const code irg.
289 * @param load the load to replace
290 * @param c the constant
292 * @return in the modes match or can be transformed using a reinterpret cast
293 * returns a copy of the constant (possibly Conv'ed) on the
296 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c);
299 * Load/Store optimization.
301 * Removes redundant non-volatile Loads and Stores.
302 * May introduce Bad nodes if exceptional control flow
303 * is removed. The following cases are optimized:
305 * Load without result: A Load which has only a memory use
308 * Load after Store: A Load after a Store is removed, if
309 * the Load doesn't have an exception handler OR is in
310 * the same block as the Store.
312 * Load after Load: A Load after a Load is removed, if the
313 * Load doesn't have an exception handler OR is in the
314 * same block as the previous Load.
316 * Store before Store: A Store immediately before another
317 * Store in the same block is removed, if the Store doesn't
318 * have an exception handler.
320 * Store after Load: A Store after a Load is removed, if the
321 * Store doesn't have an exception handler.
323 * @return non-zero if the optimization could be applied, 0 else
325 int optimize_load_store(ir_graph *irg);
328 * Creates an ir_graph pass for optimize_load_store().
330 * @param name the name of this pass or NULL
332 * @return the newly created ir_graph pass
334 ir_graph_pass_t *optimize_load_store_pass(const char *name);
337 * New experimental alternative to optimize_load_store.
338 * Based on a dataflow analysis, so load/stores are moved out of loops
341 int opt_ldst(ir_graph *irg);
344 * Creates an ir_graph pass for opt_ldst().
346 * @param name the name of this pass or NULL
348 * @return the newly created ir_graph pass
350 ir_graph_pass_t *opt_ldst_pass(const char *name);
353 * Do Loop unrolling in the given graph.
355 void optimize_loop_unrolling(ir_graph *irg);
358 * Creates an ir_graph pass for optimize_loop_unrolling().
360 * @param name the name of this pass or NULL
362 * @return the newly created ir_graph pass
364 ir_graph_pass_t *optimize_loop_unrolling_pass(const char *name);
367 * Optimize the frame type of an irg by removing
368 * never touched entities.
370 * @param irg The graph whose frame type will be optimized
372 * This function did not change the graph, only it's frame type.
373 * The layout state of the frame type will be set to layout_undefined
374 * if entities were removed.
376 void opt_frame_irg(ir_graph *irg);
379 * Creates an ir_graph pass for opt_frame_irg().
381 * @param name the name of this pass or NULL
383 * @return the newly created ir_graph pass
385 ir_graph_pass_t *opt_frame_irg_pass(const char *name);
387 /** Possible flags for the Operator Scalar Replacement. */
388 typedef enum osr_flags {
389 osr_flag_none = 0, /**< no additional flags */
390 osr_flag_lftr_with_ov_check = 1, /**< do linear function test replacement
391 only if no overflow can occur. */
392 osr_flag_ignore_x86_shift = 2, /**< ignore Multiplications by 2, 4, 8 */
393 osr_flag_keep_reg_pressure = 4 /**< do NOT increase register pressure by introducing new
394 induction variables. */
397 /* FirmJNI cannot handle identical enum values... */
399 /** default setting */
400 #define osr_flag_default osr_flag_lftr_with_ov_check
403 * Do the Operator Scalar Replacement optimization and linear
404 * function test replacement for loop control.
405 * Can be switched off using the set_opt_strength_red() flag.
406 * In that case, only remove_phi_cycles() is executed.
408 * @param irg the graph which should be optimized
409 * @param flags set of osr_flags
411 * The linear function replacement test is controlled by the flags.
412 * If the osr_flag_lftr_with_ov_check is set, the replacement is only
413 * done if do overflow can occur.
414 * Otherwise it is ALWAYS done which might be insecure.
418 * for (i = 0; i < 100; ++i)
420 * might be replaced by
422 * for (i = 0; i < 400; i += 4)
426 * for (i = 0; i < 0x7FFFFFFF; ++i)
428 * will not be replaced by
430 * for (i = 0; i < 0xFFFFFFFC; i += 4)
432 * because of overflow.
436 * for (i = 0; i <= 0xF; ++i)
438 * will NOT be transformed into
440 * for (i = 0xFFFFFFF0; i <= 0xFFFFFFFF; ++i)
442 * although here is no direct overflow. The OV occurs when the ++i
443 * is executed (and would created an endless loop here!).
445 * For the same reason, a loop
447 * for (i = 0; i <= 9; i += x)
449 * will NOT be transformed because we cannot estimate whether an overflow
450 * might happen adding x.
452 * Note that i < a + 400 is also not possible with the current implementation
453 * although this might be allowed by other compilers...
455 * Note further that tests for equality can be handled some simpler (but are not
458 * This algorithm destroys the link field of nodes.
460 void opt_osr(ir_graph *irg, unsigned flags);
463 * Creates an ir_graph pass for remove_phi_cycles().
465 * @param name the name of this pass or NULL
466 * @param flags set of osr_flags
468 * @return the newly created ir_graph pass
470 ir_graph_pass_t *opt_osr_pass(const char *name, unsigned flags);
473 * Removes useless Phi cycles, i.e cycles of Phi nodes with only one
475 * This is automatically done in opt_osr(), so there is no need to call it
478 * @param irg the graph which should be optimized
480 * This algorithm destroys the link field of nodes.
482 void remove_phi_cycles(ir_graph *irg);
485 * Creates an ir_graph pass for remove_phi_cycles().
487 * @param name the name of this pass or NULL
489 * @return the newly created ir_graph pass
491 ir_graph_pass_t *remove_phi_cycles_pass(const char *name);
494 /** A default threshold. */
495 #define DEFAULT_CLONE_THRESHOLD 300
498 * Do procedure cloning. Evaluate a heuristic weight for every
499 * Call(..., Const, ...). If the weight is bigger than threshold,
500 * clone the entity and fix the calls.
502 * @param threshold the threshold for cloning
504 * The threshold is an estimation of how many instructions are saved
505 * when executing a cloned method. If threshold is 0.0, every possible
508 void proc_cloning(float threshold);
511 * Creates an ir_prog pass for proc_cloning().
513 * @param name the name of this pass or NULL
514 * @param threshold the threshold for cloning
516 * @return the newly created ir_prog pass
518 ir_prog_pass_t *proc_cloning_pass(const char *name, float threshold);
523 * Applies Reassociation rules to integer expressions.
524 * Beware: Works only if integer overflow might be ignored, as for C, Java
525 * and for address expression.
526 * Works only if Constant folding is activated.
528 * Uses loop information to detect loop-invariant (ie contant
529 * inside the loop) values.
531 * See Muchnik 12.3.1 Algebraic Simplification and Reassociation of
532 * Addressing Expressions.
534 * @return non-zero if the optimization could be applied, 0 else
536 int optimize_reassociation(ir_graph *irg);
539 * Creates an ir_graph pass for optimize_reassociation().
541 * @param name the name of this pass or NULL
543 * @return the newly created ir_graph pass
545 ir_graph_pass_t *optimize_reassociation_pass(const char *name);
548 * Normalize the Returns of a graph by creating a new End block
549 * with One Return(Phi).
550 * This is the preferred input for the if-conversion.
552 * In pseudocode, it means:
559 * is transformed into
567 void normalize_one_return(ir_graph *irg);
570 * Creates an ir_graph pass for normalize_one_return().
572 * @param name the name of this pass or NULL
574 * @return the newly created ir_graph pass
576 ir_graph_pass_t *normalize_one_return_pass(const char *name);
579 * Normalize the Returns of a graph by moving
580 * the Returns upwards as much as possible.
581 * This might be preferred for code generation.
583 * In pseudocode, it means:
591 * is transformed into
598 void normalize_n_returns(ir_graph *irg);
601 * Creates an ir_graph pass for normalize_n_returns().
603 * @param name the name of this pass or NULL
605 * @return the newly created ir_graph pass
607 ir_graph_pass_t *normalize_n_returns_pass(const char *name);
610 * Do the scalar replacement optimization.
611 * Replace local compound entities (like structures and arrays)
612 * with atomic values if possible. Does not handle classes yet.
614 * @param irg the graph which should be optimized
616 * @return non-zero, if at least one entity was replaced
618 int scalar_replacement_opt(ir_graph *irg);
621 * Creates an ir_graph pass for scalar_replacement_opt().
623 * @param name the name of this pass or NULL
625 * @return the newly created ir_graph pass
627 ir_graph_pass_t *scalar_replacement_opt_pass(const char *name);
629 /** Performs strength reduction for the passed graph. */
630 void reduce_strength(ir_graph *irg);
633 * Optimizes tail-recursion calls by converting them into loops.
634 * Depends on the flag opt_tail_recursion.
635 * Currently supports the following forms:
637 * - return x + func();
638 * - return func() - x;
639 * - return x * func();
642 * Does not work for Calls that use the exception stuff.
644 * @param irg the graph to be optimized
646 * @return non-zero if the optimization could be applied, 0 else
648 int opt_tail_rec_irg(ir_graph *irg);
651 * Creates an ir_graph pass for opt_tail_rec_irg().
653 * @param name the name of this pass or NULL
655 * @return the newly created ir_graph pass
657 ir_graph_pass_t *opt_tail_rec_irg_pass(const char *name);
660 * Optimize tail-recursion calls for all IR-Graphs.
661 * Can currently handle:
662 * - direct return value, i.e. return func().
663 * - additive return value, i.e. return x +/- func()
664 * - multiplicative return value, i.e. return x * func() or return -func()
666 * The current implementation must be run before optimize_funccalls(),
667 * because it expects the memory edges pointing to calls, which might be
668 * removed by optimize_funccalls().
670 void opt_tail_recursion(void);
673 * Creates an ir_prog pass for opt_tail_recursion().
675 * @param name the name of this pass or NULL
677 * @return the newly created ir_prog pass
679 ir_prog_pass_t *opt_tail_recursion_pass(const char *name);
681 /** This is the type for a method, that returns a pointer type to
682 * tp. This is needed in the normalization. */
683 typedef ir_type *(*gen_pointer_type_to_func)(ir_type *tp);
685 /** Insert Casts so that class type casts conform exactly with the type hierarchy.
687 * Formulated in Java, this achieves the following:
689 * For a class hierarchy
691 * class B extends A {}
692 * class C extends B {}
693 * we transforms a cast
698 * The algorithm works for Casts with class types, but also for Casts
699 * with all pointer types that point (over several indirections,
700 * i.e. ***A) to a class type. Normalizes all graphs. Computes type
701 * information (@see irtypeinfo.h) if not available.
702 * Invalidates trout information as new casts are generated.
704 * @param gppt_fct A function that returns a pointer type that points
705 * to the type given as argument. If this parameter is NULL, a default
706 * function is used that either uses trout information or performs a O(n)
707 * search to find an existing pointer type. If it can not find a type,
708 * generates a pointer type with mode_P_mach and suffix "cc_ptr_tp".
710 void normalize_irp_class_casts(gen_pointer_type_to_func gppt_fct);
712 /** Insert Casts so that class type casts conform exactly with the type hierarchy
715 * For more details see normalize_irp_class_casts().
717 * This transformation requires that type information is computed. @see irtypeinfo.h.
719 void normalize_irg_class_casts(ir_graph *irg, gen_pointer_type_to_func gppt_fct);
721 /** Optimize casting between class types.
724 * class B extends A { }
725 * class C extends B {}
726 * Performs the following transformations:
727 * C c = (C)(B)(A)(B)new C() --> C c = (C)(B)newC() --> C c = new C()
728 * (Optimizing downcasts as A a = (A)(B)(new A()) --> A a = new A() can
729 * be suppressed by setting the flag opt_suppress_downcast_optimization.
730 * Downcasting A to B might cause an exception. It is not clear
731 * whether this is modeled by the Firm Cast node, as it has no exception
733 * If there is inh_m() that overwrites m() in B:
734 * ((A) new B()).m() --> (new B()).inh_m()
735 * Phi((A)x, (A)y) --> (A) Phi (x, y) if (A) is an upcast.
737 * Computes type information if not available. @see irtypeinfo.h.
738 * Typeinformation is valid after optimization.
739 * Invalidates trout information.
741 void optimize_class_casts(void);
744 * CLiff Click's combo algorithm from "Combining Analyses, combining Optimizations".
746 * Does conditional constant propagation, unreachable code elimination and optimistic
747 * global value numbering at once.
749 * @param irg the graph to run on
751 void combo(ir_graph *irg);
754 * Creates an ir_graph pass for combo.
756 * @param name the name of this pass or NULL
758 * @return the newly created ir_graph pass
760 ir_graph_pass_t *combo_pass(const char *name);
763 * Inlines all small methods at call sites where the called address comes
764 * from a SymConst node that references the entity representing the called
767 * @param irg the graph
768 * @param size maximum function size
770 * The size argument is a rough measure for the code size of the method:
771 * Methods where the obstack containing the firm graph is smaller than
772 * size are inlined. Further only a limited number of calls are inlined.
773 * If the method contains more than 1024 inlineable calls none will be
775 * Inlining is only performed if flags `optimize' and `inlineing' are set.
776 * The graph may not be in state phase_building.
777 * It is recommended to call local_optimize_graph() after inlining as this
778 * function leaves a set of obscure Tuple nodes, e.g. a Proj-Tuple-Jmp
779 * combination as control flow operation.
781 void inline_small_irgs(ir_graph *irg, int size);
784 * Creates an ir_graph pass for inline_small_irgs().
786 * @param name the name of this pass or NULL
787 * @param size maximum function size
789 * @return the newly created ir_graph pass
791 ir_graph_pass_t *inline_small_irgs_pass(const char *name, int size);
794 * Inlineing with a different heuristic than inline_small_irgs().
796 * Inlines leave functions. If inlinening creates new leave
797 * function inlines these, too. (If g calls f, and f calls leave h,
798 * h is first inlined in f and then f in g.)
800 * Then inlines all small functions (this is not recursive).
802 * For a heuristic this inlineing uses firm node counts. It does
803 * not count auxiliary nodes as Proj, Tuple, End, Start, Id, Sync.
804 * If the ignore_runtime flag is set, calls to functions marked with the
805 * mtp_property_runtime property are ignored.
807 * @param maxsize Do not inline any calls if a method has more than
808 * maxsize firm nodes. It may reach this limit by
810 * @param leavesize Inline leave functions if they have less than leavesize
812 * @param size Inline all function smaller than size.
813 * @param ignore_runtime count a function only calling runtime functions as
816 void inline_leave_functions(unsigned maxsize, unsigned leavesize,
817 unsigned size, int ignore_runtime);
820 * Creates an ir_prog pass for inline_leave_functions().
822 * @param name the name of this pass or NULL
823 * @param maxsize Do not inline any calls if a method has more than
824 * maxsize firm nodes. It may reach this limit by
826 * @param leavesize Inline leave functions if they have less than leavesize
828 * @param size Inline all function smaller than size.
829 * @param ignore_runtime count a function only calling runtime functions as
832 * @return the newly created ir_prog pass
834 ir_prog_pass_t *inline_leave_functions_pass(
835 const char *name, unsigned maxsize, unsigned leavesize,
836 unsigned size, int ignore_runtime);
839 * Heuristic inliner. Calculates a benefice value for every call and inlines
840 * those calls with a value higher than the threshold.
842 * @param maxsize Do not inline any calls if a method has more than
843 * maxsize firm nodes. It may reach this limit by
845 * @param inline_threshold inlining threshold
847 void inline_functions(unsigned maxsize, int inline_threshold);
850 * Creates an ir_prog pass for inline_functions().
852 * @param name the name of this pass or NULL
853 * @param maxsize Do not inline any calls if a method has more than
854 * maxsize firm nodes. It may reach this limit by
856 * @param inline_threshold inlining threshold
858 * @return the newly created ir_prog pass
860 ir_prog_pass_t *inline_functions_pass(
861 const char *name, unsigned maxsize, int inline_threshold);
864 * Combines congruent blocks into one.
866 * @param irg The IR-graph to optimize.
868 * @return non-zero if the graph was transformed
870 int shape_blocks(ir_graph *irg);
873 * Creates an ir_graph pass for shape_blocks().
875 * @param name the name of this pass or NULL
877 * @return the newly created ir_graph pass
879 ir_graph_pass_t *shape_blocks_pass(const char *name);