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 * Perform partial conditional evaluation on the given graph.
50 * @param irg the graph
52 void opt_cond_eval(ir_graph* irg);
55 * Try to simplify boolean expression in the given ir graph.
56 * eg. x < 5 && x < 6 becomes x < 5
58 * @param irg the graph
60 void opt_bool(ir_graph *irg);
63 * Try to reduce the number of conv nodes in the given ir graph.
65 * @param irg the graph
67 void conv_opt(ir_graph *irg);
70 * Do the scalar replacement optimization.
71 * Make a date flow analyze and split the
74 * @param irg the graph which should be optimized
76 void data_flow_scalar_replacement_opt(ir_graph *irg);
79 * A callback that checks whether a entity is an allocation
82 typedef int (*check_alloc_entity_func)(ir_entity *ent);
85 * Do simple and fast escape analysis for one graph.
87 * @param irg the graph
88 * @param callback a callback function to check whether a
89 * given entity is a allocation call
91 void escape_enalysis_irg(ir_graph *irg, check_alloc_entity_func callback);
94 * Do simple and fast escape analysis for all graphs.
96 * This optimization implements a simple and fast but inexact
97 * escape analysis. Some addresses might be marked as 'escaped' even
99 * The advantage is a low memory footprint and fast speed.
101 * @param run_scalar_replace if this flag in non-zero, scalar
102 * replacement optimization is run on graphs with removed
104 * @param callback a callback function to check whether a
105 * given entity is a allocation call
107 * This optimization removes allocation which are not used (rare) and replace
108 * allocation that can be proved dead at the end of the graph which stack variables.
110 * The creation of stack variable allows scalar replacement to be run only
111 * on those graphs that have been changed.
113 * This is most effective on Java where no other stack variables exists.
115 void escape_analysis(int run_scalar_replace, check_alloc_entity_func callback);
118 * Optimize function calls by handling const functions.
120 * This optimization first detects all "const functions", i.e.,
121 * IR graphs that neither read nor write memory (and hence did
122 * not create exceptions, as these use memory in Firm).
124 * The result of calls to such functions depends only on its
125 * arguments, hence those calls are no more pinned.
127 * This is a rather strong criteria, so do not expect that a
128 * lot of functions will be found. Moreover, all of them might
129 * already be inlined if inlining is activated.
130 * Anyway, it might be good for handling builtin's or pseudo-graphs,
131 * even if the later read/write memory (but we know how).
133 * This optimizations read the irg_const_function property of
134 * entities and and sets the irg_const_function property of
137 * If callee information is valid, we also optimize polymorphic Calls.
139 * @param force_run if non-zero, an optimization run is started even
140 * if no const function graph was detected.
141 * Else calls are only optimized if at least one
142 * const function graph was detected.
143 * @param callback a callback function to check whether a
144 * given entity is a allocation call
146 * If the frontend created external entities with the irg_const_function
147 * property set, the force_run parameter should be set, else
150 * @note This optimization destroys the link fields of nodes.
152 void optimize_funccalls(int force_run, check_alloc_entity_func callback);
155 * Does Partial Redundancy Elimination combined with
156 * Global Value Numbering.
157 * Can be used to replace place_code() completely.
159 * Based on VanDrunen and Hosking 2004.
161 * @param irg the graph
163 void do_gvn_pre(ir_graph *irg);
166 * This function is called to evaluate, if a mux can build
167 * of the current architecture.
168 * If it returns non-zero, a mux is created, else the code
170 * @param sel A selector of a Cond.
171 * @param phi_list List of Phi nodes about to be converted (linked via get_Phi_next() field)
172 * @param i First data predecessor involved in if conversion
173 * @param j Second data predecessor involved in if conversion
175 typedef int (*arch_allow_ifconv_func)(ir_node *sel, ir_node* phi_list, int i, int j);
178 * The parameters structure.
180 struct ir_settings_if_conv_t {
181 int max_depth; /**< The maximum depth up to which expressions
182 are examined when it has to be decided if they
183 can be placed into another block. */
184 arch_allow_ifconv_func allow_ifconv; /**< Evaluator function, if not set all possible Psi
185 nodes will be created. */
189 * Perform If conversion on a graph.
191 * @param irg The graph.
192 * @param params The parameters for the if conversion.
194 * Cannot handle blocks with Bad control predecessors, so call it after control
197 void opt_if_conv(ir_graph *irg, const ir_settings_if_conv_t *params);
199 void opt_sync(ir_graph *irg);
202 * Check if we can replace the load by a given const from
203 * the const code irg.
205 * @param load the load to replace
206 * @param c the constant
208 * @return in the modes match or can be transformed using a reinterpret cast
209 * returns a copy of the constant (possibly Conv'ed) on the
212 ir_node *can_replace_load_by_const(const ir_node *load, ir_node *c);
215 * Load/Store optimization.
217 * Removes redundant non-volatile Loads and Stores.
218 * May introduce Bad nodes if exceptional control flow
219 * is removed. The following cases are optimized:
221 * Load without result: A Load which has only a memory use
224 * Load after Store: A Load after a Store is removed, if
225 * the Load doesn't have an exception handler OR is in
226 * the same block as the Store.
228 * Load after Load: A Load after a Load is removed, if the
229 * Load doesn't have an exception handler OR is in the
230 * same block as the previous Load.
232 * Store before Store: A Store immediately before another
233 * Store in the same block is removed, if the Store doesn't
234 * have an exception handler.
236 * Store after Load: A Store after a Load is removed, if the
237 * Store doesn't have an exception handler.
239 void optimize_load_store(ir_graph *irg);
242 * Do Loop unrolling in the given graph.
244 void optimize_loop_unrolling(ir_graph *irg);
247 * Optimize the frame type of an irg by removing
248 * never touched entities.
250 * @param irg The graph whose frame type will be optimized
252 * This function did not change the graph, only it's frame type.
253 * The layout state of the frame type will be set to layout_undefined
254 * if entities were removed.
256 void opt_frame_irg(ir_graph *irg);
258 /** Possible flags for the Operator Scalar Replacement. */
259 typedef enum osr_flags {
260 osr_flag_none = 0, /**< no additional flags */
261 osr_flag_lftr_with_ov_check = 1, /**< do linear function test replacement
262 only if no overflow can occur. */
263 osr_flag_ignore_x86_shift = 2, /**< ignore Multiplications by 2, 4, 8 */
264 osr_flag_keep_reg_pressure = 4 /**< do NOT increase register pressure by introducing new
265 induction variables. */
268 /* FirmJNI cannot handle identical enum values... */
270 /** default setting */
271 #define osr_flag_default osr_flag_lftr_with_ov_check
274 * Do the Operator Scalar Replacement optimization and linear
275 * function test replacement for loop control.
276 * Can be switched off using the set_opt_strength_red() flag.
277 * In that case, only remove_phi_cycles() is executed.
279 * @param irg the graph which should be optimized
280 * @param flags set of osr_flags
282 * The linear function replacement test is controlled by the flags.
283 * If the osr_flag_lftr_with_ov_check is set, the replacement is only
284 * done if do overflow can occur.
285 * Otherwise it is ALWAYS done which might be insecure.
289 * for (i = 0; i < 100; ++i)
291 * might be replaced by
293 * for (i = 0; i < 400; i += 4)
297 * for (i = 0; i < 0x7FFFFFFF; ++i)
299 * will not be replaced by
301 * for (i = 0; i < 0xFFFFFFFC; i += 4)
303 * because of overflow.
307 * for (i = 0; i <= 0xF; ++i)
309 * will NOT be transformed into
311 * for (i = 0xFFFFFFF0; i <= 0xFFFFFFFF; ++i)
313 * although here is no direct overflow. The OV occurs when the ++i
314 * is executed (and would created an endless loop here!).
316 * For the same reason, a loop
318 * for (i = 0; i <= 9; i += x)
320 * will NOT be transformed because we cannot estimate whether an overflow
321 * might happen adding x.
323 * Note that i < a + 400 is also not possible with the current implementation
324 * although this might be allowed by other compilers...
326 * Note further that tests for equality can be handled some simpler (but are not
329 * This algorithm destroys the link field of nodes.
331 void opt_osr(ir_graph *irg, unsigned flags);
334 * Removes useless Phi cycles, i.e cycles of Phi nodes with only one
336 * This is automatically done in opt_osr(), so there is no need to call it
339 * @param irg the graph which should be optimized
341 * This algorithm destroys the link field of nodes.
343 void remove_phi_cycles(ir_graph *irg);
345 /** A default threshold. */
346 #define DEFAULT_CLONE_THRESHOLD 300
349 * Do procedure cloning. Evaluate a heuristic weight for every
350 * Call(..., Const, ...). If the weight is bigger than threshold,
351 * clone the entity and fix the calls.
353 * @param threshold the threshold for cloning
355 * The threshold is an estimation of how many instructions are saved
356 * when executing a cloned method. If threshold is 0.0, every possible
359 void proc_cloning(float threshold);
364 * Applies Reassociation rules to integer expressions.
365 * Beware: Works only if integer overflow might be ignored, as for C, Java
366 * and for address expression.
367 * Works only if Constant folding is activated.
369 * Uses loop information to detect loop-invariant (ie contant
370 * inside the loop) values.
372 * See Muchnik 12.3.1 Algebraic Simplification and Reassociation of
373 * Addressing Expressions.
377 void optimize_reassociation(ir_graph *irg);
380 * Normalize the Returns of a graph by creating a new End block
381 * with One Return(Phi).
382 * This is the preferred input for the if-conversion.
384 * In pseudocode, it means:
391 * is transformed into
399 void normalize_one_return(ir_graph *irg);
402 * Normalize the Returns of a graph by moving
403 * the Returns upwards as much as possible.
404 * This might be preferred for code generation.
406 * In pseudocode, it means:
414 * is transformed into
421 void normalize_n_returns(ir_graph *irg);
424 * Do the scalar replacement optimization.
425 * Replace local compound entities (like structures and arrays)
426 * with atomic values if possible. Does not handle classes yet.
428 * @param irg the graph which should be optimized
430 * @return non-zero, if at least one entity was replaced
432 int scalar_replacement_opt(ir_graph *irg);
434 /** Performs strength reduction for the passed graph. */
435 void reduce_strength(ir_graph *irg);
438 * Optimizes tail-recursion calls by converting them into loops.
439 * Depends on the flag opt_tail_recursion.
440 * Currently supports the following forms:
442 * - return x + func();
443 * - return func() - x;
444 * - return x * func();
447 * Does not work for Calls that use the exception stuff.
449 * @param irg the graph to be optimized
451 * @return non-zero if the optimization could be applied, 0 else
453 int opt_tail_rec_irg(ir_graph *irg);
456 * Optimize tail-recursion calls for all IR-Graphs.
457 * Can currently handle:
458 * - direct return value, i.e. return func().
459 * - additive return value, i.e. return x +/- func()
460 * - multiplicative return value, i.e. return x * func() or return -func()
462 * The current implementation must be run before optimize_funccalls(),
463 * because it expects the memory edges pointing to calls, which might be
464 * removed by optimize_funccalls().
466 void opt_tail_recursion(void);
468 /** This is the type for a method, that returns a pointer type to
469 * tp. This is needed in the normalization. */
470 typedef ir_type *(*gen_pointer_type_to_func)(ir_type *tp);
472 /** Insert Casts so that class type casts conform exactly with the type hierarchy.
474 * Formulated in Java, this achieves the following:
476 * For a class hierarchy
478 * class B extends A {}
479 * class C extends B {}
480 * we transforms a cast
485 * The algorithm works for Casts with class types, but also for Casts
486 * with all pointer types that point (over several indirections,
487 * i.e. ***A) to a class type. Normalizes all graphs. Computes type
488 * information (@see irtypeinfo.h) if not available.
489 * Invalidates trout information as new casts are generated.
491 * @param gppt_fct A function that returns a pointer type that points
492 * to the type given as argument. If this parameter is NULL, a default
493 * function is used that either uses trout information or performs a O(n)
494 * search to find an existing pointer type. If it can not find a type,
495 * generates a pointer type with mode_P_mach and suffix "cc_ptr_tp".
497 void normalize_irp_class_casts(gen_pointer_type_to_func gppt_fct);
500 /** Insert Casts so that class type casts conform exactly with the type hierarchy
503 * For more details see normalize_irp_class_casts().
505 * This transformation requires that type information is computed. @see irtypeinfo.h.
507 void normalize_irg_class_casts(ir_graph *irg, gen_pointer_type_to_func gppt_fct);
510 /** Optimize casting between class types.
513 * class B extends A { }
514 * class C extends B {}
515 * Performs the following transformations:
516 * C c = (C)(B)(A)(B)new C() --> C c = (C)(B)newC() --> C c = new C()
517 * (Optimizing downcasts as A a = (A)(B)(new A()) --> A a = new A() can
518 * be suppressed by setting the flag opt_suppress_downcast_optimization.
519 * Downcasting A to B might cause an exception. It is not clear
520 * whether this is modeled by the Firm Cast node, as it has no exception
522 * If there is inh_m() that overwrites m() in B:
523 * ((A) new B()).m() --> (new B()).inh_m()
524 * Phi((A)x, (A)y) --> (A) Phi (x, y) if (A) is an upcast.
526 * Computes type information if not available. @see irtypeinfo.h.
527 * Typeinformation is valid after optimization.
528 * Invalidates trout information.
530 void optimize_class_casts(void);
533 * CLiff Click's combo algorithm from "Combining Analyses, combining Optimizations".
535 * Does conditional constant propagation, unreachable code elimination and optimistic
536 * global value numbering at once.
538 * @param irg the graph to run on
540 void combo(ir_graph *irg);
542 /** Inlines all small methods at call sites where the called address comes
543 * from a SymConst node that references the entity representing the called
546 * The size argument is a rough measure for the code size of the method:
547 * Methods where the obstack containing the firm graph is smaller than
548 * size are inlined. Further only a limited number of calls are inlined.
549 * If the method contains more than 1024 inlineable calls none will be
551 * Inlining is only performed if flags `optimize' and `inlineing' are set.
552 * The graph may not be in state phase_building.
553 * It is recommended to call local_optimize_graph() after inlining as this
554 * function leaves a set of obscure Tuple nodes, e.g. a Proj-Tuple-Jmp
555 * combination as control flow operation.
557 void inline_small_irgs(ir_graph *irg, int size);
560 /** Inlineing with a different heuristic than inline_small_irgs().
562 * Inlines leave functions. If inlinening creates new leave
563 * function inlines these, too. (If g calls f, and f calls leave h,
564 * h is first inlined in f and then f in g.)
566 * Then inlines all small functions (this is not recursive).
568 * For a heuristic this inlineing uses firm node counts. It does
569 * not count auxiliary nodes as Proj, Tuple, End, Start, Id, Sync.
570 * If the ignore_runtime flag is set, calls to functions marked with the
571 * mtp_property_runtime property are ignored.
573 * @param maxsize Do not inline any calls if a method has more than
574 * maxsize firm nodes. It may reach this limit by
576 * @param leavesize Inline leave functions if they have less than leavesize
578 * @param size Inline all function smaller than size.
579 * @param ignore_runtime count a function only calling runtime functions as
582 void inline_leave_functions(unsigned maxsize, unsigned leavesize,
583 unsigned size, int ignore_runtime);
586 * Heuristic inliner. Calculates a benefice value for every call and inlines
587 * those calls with a value higher than the threshold.
589 * @param maxsize Do not inline any calls if a method has more than
590 * maxsize firm nodes. It may reach this limit by
592 * @param threshold inlining threshold
594 void inline_functions(unsigned maxsize, int inline_threshold);