2 * Copyright (C) 1995-2007 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 Reassociation
23 * @author Michael Beck
32 #include "irgraph_t.h"
36 #include "iropt_dbg.h"
39 #include "reassoc_t.h"
45 DEBUG_ONLY(static firm_dbg_module_t *dbg;)
47 typedef struct _walker_t {
48 int changes; /**< set, if a reassociation take place */
49 waitq *wq; /**< a wait queue */
53 NO_CONSTANT = 0, /**< node is not constant */
54 REAL_CONSTANT = 1, /**< node is a Const that is suitable for constant folding */
55 REGION_CONST = 4 /**< node is a constant expression in the current context,
56 use 4 here to simplify implementation of get_comm_Binop_ops() */
60 * returns whether a node is constant ie is a constant or
61 * is loop invariant (called region constant)
63 * @param n the node to be checked for constant
64 * @param block a block that might be in a loop
66 static const_class_t get_const_class(ir_node *n, ir_node *block)
68 ir_op *op = get_irn_op(n);
73 /* although SymConst's are of course real constant, we cannot
74 fold them, so handle them like region constants */
75 if (op == op_SymConst)
79 * Beware: Bad nodes are always loop-invariant, but
80 * cannot handled in later code, so filter them here.
82 if (! is_Bad(n) && is_loop_invariant(n, block))
86 } /* get_const_class */
89 * returns the operands of a commutative bin-op, if one operand is
90 * a region constant, it is returned as the second one.
92 * Beware: Real constants must be returned with higher priority than
93 * region constants, because they might be folded.
95 static void get_comm_Binop_ops(ir_node *binop, ir_node **a, ir_node **c)
97 ir_node *op_a = get_binop_left(binop);
98 ir_node *op_b = get_binop_right(binop);
99 ir_node *block = get_nodes_block(binop);
100 int class_a = get_const_class(op_a, block);
101 int class_b = get_const_class(op_b, block);
103 assert(is_op_commutative(get_irn_op(binop)));
105 switch (class_a + 2*class_b) {
106 case REAL_CONSTANT + 2*REAL_CONSTANT:
107 /* if both are constants, one might be a
108 * pointer constant like NULL, return the other
110 if (mode_is_reference(get_irn_mode(op_a))) {
118 case REAL_CONSTANT + 2*NO_CONSTANT:
119 case REAL_CONSTANT + 2*REGION_CONST:
120 case REGION_CONST + 2*NO_CONSTANT:
129 } /* get_comm_Binop_ops */
132 * reassociate a Sub: x - c = x + (-c)
134 static int reassoc_Sub(ir_node **in)
137 ir_node *right = get_Sub_right(n);
138 ir_mode *rmode = get_irn_mode(right);
141 /* cannot handle SubIs(P, P) */
142 if (mode_is_reference(rmode))
145 block = get_nodes_block(n);
148 * convert x - c => x + (-c)
150 if (get_const_class(right, block) == REAL_CONSTANT) {
151 ir_node *left = get_Sub_left(n);
156 switch (get_const_class(left, block)) {
158 irn = optimize_in_place(n);
168 /* already constant, nothing to do */
171 mode = get_irn_mode(n);
172 dbi = get_irn_dbg_info(n);
174 /* Beware of SubP(P, Is) */
175 irn = new_rd_Minus(dbi, current_ir_graph, block, right, rmode);
176 irn = new_rd_Add(dbi, current_ir_graph, block, left, irn, get_irn_mode(n));
178 DBG((dbg, LEVEL_5, "Applied: %n - %n => %n + (-%n)\n",
179 get_Sub_left(n), right, get_Sub_left(n), right));
192 /** Retrieve a mode from the operands. We need this, because
193 * Add and Sub are allowed to operate on (P, Is)
195 static ir_mode *get_mode_from_ops(ir_node *op1, ir_node *op2)
199 m1 = get_irn_mode(op1);
200 if (mode_is_reference(m1))
203 m2 = get_irn_mode(op2);
204 if (mode_is_reference(m2))
210 } /* get_mode_from_ops */
213 * reassociate a commutative Binop
215 * BEWARE: this rule leads to a potential loop, if
216 * two operands are region constants and the third is a
217 * constant, so avoid this situation.
219 static int reassoc_commutative(ir_node **node)
222 ir_op *op = get_irn_op(n);
223 ir_node *block = get_nodes_block(n);
226 get_comm_Binop_ops(n, &t1, &c1);
228 if (get_irn_op(t1) == op) {
230 const_class_t c_c1, c_c2, c_t2;
232 get_comm_Binop_ops(t1, &t2, &c2);
234 /* do not optimize Bad nodes, will fail later */
238 c_c1 = get_const_class(c1, block);
239 c_c2 = get_const_class(c2, block);
240 c_t2 = get_const_class(t2, block);
242 if ( ((c_c1 > NO_CONSTANT) & (c_t2 > NO_CONSTANT)) &&
243 ((((c_c1 ^ c_c2 ^ c_t2) & REGION_CONST) == 0) || ((c_c1 & c_c2 & c_t2) == REGION_CONST)) ) {
244 /* All three are constant and either all are constant expressions or two of them are:
245 * then applying this rule would lead into a cycle
247 * Note that if t2 is a constant so is c2 hence we save one test.
252 if ((c_c1 != NO_CONSTANT) /* & (c_c2 != NO_CONSTANT) */) {
253 /* handles rules R7, R8, R9, R10:
254 * convert c1 .OP. (c2 .OP. x) => x .OP. (c1 .OP. c2)
256 ir_node *irn, *in[2];
257 ir_mode *mode, *mode_c1 = get_irn_mode(c1), *mode_c2 = get_irn_mode(c2);
259 /* It might happen, that c1 and c2 have different modes, for instance Is and Iu.
262 if (mode_c1 != mode_c2) {
263 if (mode_is_int(mode_c1) && mode_is_int(mode_c2)) {
264 /* get the bigger one */
265 if (get_mode_size_bits(mode_c1) > get_mode_size_bits(mode_c2))
266 c2 = new_r_Conv(current_ir_graph, block, c2, mode_c1);
267 else if (get_mode_size_bits(mode_c1) < get_mode_size_bits(mode_c2))
268 c1 = new_r_Conv(current_ir_graph, block, c1, mode_c2);
270 /* Try to cast the real const */
271 if (c_c1 == REAL_CONSTANT)
272 c1 = new_r_Conv(current_ir_graph, block, c1, mode_c2);
274 c2 = new_r_Conv(current_ir_graph, block, c2, mode_c1);
282 mode = get_mode_from_ops(in[0], in[1]);
283 in[1] = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in));
286 mode = get_mode_from_ops(in[0], in[1]);
287 irn = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in));
289 DBG((dbg, LEVEL_5, "Applied: %n .%s. (%n .%s. %n) => %n .%s. (%n .%s. %n)\n",
290 c1, get_irn_opname(n), c2, get_irn_opname(n), t2,
291 t2, get_irn_opname(n), c1, get_irn_opname(n), c2));
293 * In some rare cases it can really happen that we get the same node back.
294 * This might be happen in dead loops, were the Phi nodes are already gone away.
305 } /* reassoc_commutative */
307 #define reassoc_Add reassoc_commutative
308 #define reassoc_And reassoc_commutative
309 #define reassoc_Or reassoc_commutative
310 #define reassoc_Eor reassoc_commutative
313 * Reassociate using commutative law for Mul and distributive law for Mul and Add/Sub:
315 static int reassoc_Mul(ir_node **node)
318 ir_node *add_sub, *c;
321 if (reassoc_commutative(&n))
324 get_comm_Binop_ops(n, &add_sub, &c);
325 op = get_irn_op(add_sub);
327 /* handles rules R11, R12, R13, R14, R15, R16, R17, R18, R19, R20 */
328 if (op == op_Add || op == op_Sub) {
329 ir_mode *mode = get_irn_mode(n);
330 ir_node *irn, *block, *t1, *t2, *in[2];
332 block = get_nodes_block(n);
333 t1 = get_binop_left(add_sub);
334 t2 = get_binop_right(add_sub);
336 /* we can only multiplication rules on integer arithmetic */
337 if (mode_is_int(get_irn_mode(t1)) && mode_is_int(get_irn_mode(t2))) {
338 in[0] = new_rd_Mul(NULL, current_ir_graph, block, c, t1, mode);
339 in[1] = new_rd_Mul(NULL, current_ir_graph, block, c, t2, mode);
341 mode = get_mode_from_ops(in[0], in[1]);
342 irn = optimize_node(new_ir_node(NULL, current_ir_graph, block, op, mode, 2, in));
344 /* In some cases it might happen that the new irn is equal the old one, for
346 * (x - 1) * y == x * y - y
347 * will be transformed back by simpler optimization
348 * We could switch simple optimizations off, but this only happens iff y
349 * is a loop-invariant expression and that it is not clear if the new form
351 * So, we let the old one.
354 DBG((dbg, LEVEL_5, "Applied: (%n .%s. %n) %n %n => (%n %n %n) .%s. (%n %n %n)\n",
355 t1, get_op_name(op), t2, n, c, t1, n, c, get_op_name(op), t2, n, c));
367 * The walker for the reassociation.
369 static void wq_walker(ir_node *n, void *env)
371 walker_t *wenv = env;
373 set_irn_link(n, NULL);
374 if (is_no_Block(n)) {
375 ir_node *blk = get_nodes_block(n);
377 if (is_Block_dead(blk) || get_Block_dom_depth(blk) < 0) {
378 /* We are in a dead block, do not optimize or we may fall into an endless
379 loop. We check this here instead of requiring that all dead blocks are removed
380 which or cf_opt do not guarantee yet. */
383 waitq_put(wenv->wq, n);
384 set_irn_link(n, wenv->wq);
389 * The walker for the reassociation.
391 static void do_reassociation(walker_t *wenv)
397 while (! waitq_empty(wenv->wq)) {
398 n = waitq_get(wenv->wq);
399 set_irn_link(n, NULL);
401 blk = get_nodes_block(n);
402 if (is_Block_dead(blk) || get_Block_dom_depth(blk) < 0) {
403 /* We are in a dead block, do not optimize or we may fall into an endless
404 loop. We check this here instead of requiring that all dead blocks are removed
405 which or cf_opt do not guarantee yet. */
412 /* reassociation must run until a fixpoint is reached. */
415 ir_op *op = get_irn_op(n);
416 ir_mode *mode = get_irn_mode(n);
420 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
421 if (mode_is_float(mode) && get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)
424 if (op->ops.reassociate) {
425 res = op->ops.reassociate(&n);
432 wenv->changes |= changed;
435 for (i = get_irn_arity(n) - 1; i >= 0; --i) {
436 ir_node *pred = get_irn_n(n, i);
438 if (get_irn_link(pred) != wenv->wq) {
439 waitq_put(wenv->wq, pred);
440 set_irn_link(pred, wenv->wq);
445 } /* do_reassociation */
448 * Returns the earliest were a,b are available.
449 * Note that we know that we know that a, b both dominate
450 * the block of the previous operation, so one must dominate the other.
452 static ir_node *earliest_block(ir_node *a, ir_node *b) {
453 ir_node *blk_a = get_nodes_block(a);
454 ir_node *blk_b = get_nodes_block(b);
456 /* if blk_a != blk_b, one must dominate the other */
457 if (block_dominates(blk_a, blk_b))
465 * Move Constants towards the root.
467 static int move_consts_up(ir_node **node) {
469 ir_op *op = get_irn_op(n);
470 ir_node *l, *r, *a, *b, *c, *blk, *irn, *in[2];
472 dbg_info *dbg = get_irn_dbg_info(n);
474 l = get_binop_left(n);
475 r = get_binop_right(n);
476 if (get_irn_op(l) == op && !is_irn_constlike(r)) {
477 a = get_binop_left(l);
478 b = get_binop_right(l);
480 if (is_irn_constlike(a)) {
483 blk = get_nodes_block(l);
484 dbg = dbg == get_irn_dbg_info(l) ? dbg : NULL;
486 } else if (is_irn_constlike(b)) {
489 blk = get_nodes_block(l);
490 dbg = dbg == get_irn_dbg_info(l) ? dbg : NULL;
493 } else if (get_irn_op(r) == op && !is_irn_constlike(l)) {
494 a = get_binop_left(r);
495 b = get_binop_right(r);
497 if (is_irn_constlike(a)) {
500 blk = get_nodes_block(r);
501 dbg = dbg == get_irn_dbg_info(r) ? dbg : NULL;
503 } else if (is_irn_constlike(b)) {
506 blk = get_nodes_block(r);
507 dbg = dbg == get_irn_dbg_info(r) ? dbg : NULL;
514 /* check if a+b can be calculted in the same block is the old instruction */
515 if (! block_dominates(get_nodes_block(a), blk))
517 if (! block_dominates(get_nodes_block(b), blk))
523 mode = get_mode_from_ops(in[0], in[1]);
524 in[0] = optimize_node(new_ir_node(dbg, current_ir_graph, blk, op, mode, 2, in));
527 mode = get_mode_from_ops(in[0], in[1]);
528 irn = optimize_node(new_ir_node(dbg, current_ir_graph, blk, op, mode, 2, in));
533 } /* mve_consts_up */
536 * Apply distributive Law for Mul and Add/Sub
538 static int reverse_rule_distributive(ir_node **node) {
540 ir_node *left = get_binop_left(n);
541 ir_node *right = get_binop_right(n);
543 ir_node *a, *b, *irn;
547 if (! is_Mul(left) || !is_Mul(right))
550 x = get_Mul_left(left);
552 if (x == get_Mul_left(right)) {
553 /* (x * a) +/- (x * b) */
554 a = get_Mul_right(left);
555 b = get_Mul_right(right);
557 } else if (x == get_Mul_right(right)) {
558 /* (x * a) +/- (b * x) */
559 a = get_Mul_right(left);
560 b = get_Mul_left(right);
564 x = get_Mul_right(left);
566 if (x == get_Mul_right(right)) {
567 /* (a * x) +/- (b * x) */
568 a = get_Mul_left(left);
569 b = get_Mul_left(right);
571 } else if (x == get_Mul_left(right)) {
572 /* (a * x) +/- (x * b) */
573 a = get_Mul_left(left);
574 b = get_Mul_right(right);
580 blk = earliest_block(a, b);
582 dbg = get_irn_dbg_info(n);
583 mode = get_irn_mode(n);
586 irn = new_rd_Add(dbg, current_ir_graph, blk, a, b, mode);
588 irn = new_rd_Sub(dbg, current_ir_graph, blk, a, b, mode);
590 blk = earliest_block(irn, x);
591 irn = new_rd_Mul(dbg, current_ir_graph, blk, irn, x, mode);
596 } /* reverse_rule_distributive */
599 * Apply the rules in reverse order, removing code that was not collapsed
601 static void reverse_rules(ir_node *node, void *env) {
602 walker_t *wenv = env;
603 ir_mode *mode = get_irn_mode(node);
606 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
607 if (mode_is_float(mode) && get_irg_fp_model(current_ir_graph) & fp_strict_algebraic)
611 ir_op *op = get_irn_op(node);
614 if (is_op_commutative(op)) {
615 /* no need to recurse here */
616 wenv->changes |= move_consts_up(&node);
618 if (op == op_Add || op == op_Sub) {
619 wenv->changes |= res = reverse_rule_distributive(&node);
625 * do the reassociation
627 void optimize_reassociation(ir_graph *irg)
630 irg_loopinfo_state state;
632 assert(get_irg_phase_state(irg) != phase_building);
633 assert(get_irg_pinned(irg) != op_pin_state_floats &&
634 "Reassociation needs pinned graph to work properly");
636 /* reassociation needs constant folding */
637 if (!get_opt_reassociation() || !get_opt_constant_folding())
640 /* we use dominance to detect dead blocks */
644 * Calculate loop info, so we could identify loop-invariant
645 * code and threat it like a constant.
646 * We only need control flow loops here but can handle generic
647 * INTRA info as well.
649 state = get_irg_loopinfo_state(irg);
650 if ((state & loopinfo_inter) ||
651 (state & (loopinfo_constructed | loopinfo_valid)) != (loopinfo_constructed | loopinfo_valid))
652 construct_cf_backedges(irg);
655 env.wq = new_waitq();
657 /* now we have collected enough information, optimize */
658 irg_walk_graph(irg, NULL, wq_walker, &env);
659 do_reassociation(&env);
661 /* reverse those rules that do not result in collapsed constants */
662 irg_walk_graph(irg, NULL, reverse_rules, &env);
664 /* Handle graph state */
666 set_irg_outs_inconsistent(irg);
667 set_irg_loopinfo_inconsistent(irg);
671 } /* optimize_reassociation */
673 /* Sets the default reassociation operation for an ir_op_ops. */
674 ir_op_ops *firm_set_default_reassoc(ir_opcode code, ir_op_ops *ops)
676 #define CASE(a) case iro_##a: ops->reassociate = reassoc_##a; break
691 } /* firm_set_default_reassoc */
693 /* initialize the reassociation by adding operations to some opcodes */
694 void firm_init_reassociation(void)
696 FIRM_DBG_REGISTER(dbg, "firm.opt.reassoc");
697 } /* firm_init_reassociation */