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 iropt --- optimizations intertwined with IR construction.
23 * @author Christian Schaefer, Goetz Lindenmaier, Michael Beck
33 #include "irgraph_t.h"
34 #include "iredges_t.h"
41 #include "dbginfo_t.h"
42 #include "iropt_dbg.h"
48 #include "opt_confirms.h"
49 #include "opt_polymorphy.h"
53 /* Make types visible to allow most efficient access */
57 * Return the value of a Constant.
59 static tarval *computed_value_Const(ir_node *n) {
60 return get_Const_tarval(n);
61 } /* computed_value_Const */
64 * Return the value of a 'sizeof', 'alignof' or 'offsetof' SymConst.
66 static tarval *computed_value_SymConst(ir_node *n) {
70 switch (get_SymConst_kind(n)) {
71 case symconst_type_size:
72 type = get_SymConst_type(n);
73 if (get_type_state(type) == layout_fixed)
74 return new_tarval_from_long(get_type_size_bytes(type), get_irn_mode(n));
76 case symconst_type_align:
77 type = get_SymConst_type(n);
78 if (get_type_state(type) == layout_fixed)
79 return new_tarval_from_long(get_type_alignment_bytes(type), get_irn_mode(n));
81 case symconst_ofs_ent:
82 ent = get_SymConst_entity(n);
83 type = get_entity_owner(ent);
84 if (get_type_state(type) == layout_fixed)
85 return new_tarval_from_long(get_entity_offset(ent), get_irn_mode(n));
91 } /* computed_value_SymConst */
94 * Return the value of an Add.
96 static tarval *computed_value_Add(ir_node *n) {
97 ir_node *a = get_Add_left(n);
98 ir_node *b = get_Add_right(n);
100 tarval *ta = value_of(a);
101 tarval *tb = value_of(b);
103 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
104 return tarval_add(ta, tb);
107 } /* computed_value_Add */
110 * Return the value of a Sub.
111 * Special case: a - a
113 static tarval *computed_value_Sub(ir_node *n) {
114 ir_node *a = get_Sub_left(n);
115 ir_node *b = get_Sub_right(n);
120 if (a == b && !is_Bad(a))
121 return get_mode_null(get_irn_mode(n));
126 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b)))
127 return tarval_sub(ta, tb);
130 } /* computed_value_Sub */
133 * Return the value of a Carry.
134 * Special : a op 0, 0 op b
136 static tarval *computed_value_Carry(ir_node *n) {
137 ir_node *a = get_binop_left(n);
138 ir_node *b = get_binop_right(n);
139 ir_mode *m = get_irn_mode(n);
141 tarval *ta = value_of(a);
142 tarval *tb = value_of(b);
144 if ((ta != tarval_bad) && (tb != tarval_bad)) {
146 return tarval_carry() ? get_mode_one(m) : get_mode_null(m);
148 if ( (classify_tarval(ta) == TV_CLASSIFY_NULL)
149 || (classify_tarval(tb) == TV_CLASSIFY_NULL))
150 return get_mode_null(m);
153 } /* computed_value_Carry */
156 * Return the value of a Borrow.
159 static tarval *computed_value_Borrow(ir_node *n) {
160 ir_node *a = get_binop_left(n);
161 ir_node *b = get_binop_right(n);
162 ir_mode *m = get_irn_mode(n);
164 tarval *ta = value_of(a);
165 tarval *tb = value_of(b);
167 if ((ta != tarval_bad) && (tb != tarval_bad)) {
168 return tarval_cmp(ta, tb) == pn_Cmp_Lt ? get_mode_one(m) : get_mode_null(m);
169 } else if (classify_tarval(ta) == TV_CLASSIFY_NULL) {
170 return get_mode_null(m);
173 } /* computed_value_Borrow */
176 * Return the value of an unary Minus.
178 static tarval *computed_value_Minus(ir_node *n) {
179 ir_node *a = get_Minus_op(n);
180 tarval *ta = value_of(a);
182 if ((ta != tarval_bad) && mode_is_signed(get_irn_mode(a)))
183 return tarval_neg(ta);
186 } /* computed_value_Minus */
189 * Return the value of a Mul.
191 static tarval *computed_value_Mul(ir_node *n) {
192 ir_node *a = get_Mul_left(n);
193 ir_node *b = get_Mul_right(n);
195 tarval *ta = value_of(a);
196 tarval *tb = value_of(b);
198 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
199 return tarval_mul(ta, tb);
201 /* a*0 = 0 or 0*b = 0:
202 calls computed_value recursive and returns the 0 with proper
204 if ((ta != tarval_bad) && (ta == get_mode_null(get_tarval_mode(ta))))
206 if ((tb != tarval_bad) && (tb == get_mode_null(get_tarval_mode(tb))))
210 } /* computed_value_Mul */
213 * Return the value of a floating point Quot.
215 static tarval *computed_value_Quot(ir_node *n) {
216 ir_node *a = get_Quot_left(n);
217 ir_node *b = get_Quot_right(n);
219 tarval *ta = value_of(a);
220 tarval *tb = value_of(b);
222 /* This was missing in original implementation. Why? */
223 if ((ta != tarval_bad) && (tb != tarval_bad) && (get_irn_mode(a) == get_irn_mode(b))) {
224 if (tb != get_mode_null(get_tarval_mode(tb))) /* div by zero: return tarval_bad */
225 return tarval_quo(ta, tb);
228 } /* computed_value_Quot */
231 * Calculate the value of an integer Div of two nodes.
232 * Special case: 0 / b
234 static tarval *do_computed_value_Div(ir_node *a, ir_node *b) {
235 tarval *ta = value_of(a);
236 tarval *tb = value_of(b);
238 /* Compute c1 / c2 or 0 / a, a != 0 */
239 if (ta != tarval_bad) {
240 if ((tb != tarval_bad) && (tb != get_mode_null(get_irn_mode(b)))) /* div by zero: return tarval_bad */
241 return tarval_div(ta, tb);
242 else if (ta == get_mode_null(get_tarval_mode(ta))) /* 0 / b == 0 */
246 } /* do_computed_value_Div */
249 * Return the value of an integer Div.
251 static tarval *computed_value_Div(ir_node *n) {
252 return do_computed_value_Div(get_Div_left(n), get_Div_right(n));
253 } /* computed_value_Div */
256 * Calculate the value of an integer Mod of two nodes.
257 * Special case: a % 1
259 static tarval *do_computed_value_Mod(ir_node *a, ir_node *b) {
260 tarval *ta = value_of(a);
261 tarval *tb = value_of(b);
263 /* Compute c1 % c2 or a % 1 */
264 if (tb != tarval_bad) {
265 if ((ta != tarval_bad) && (tb != get_mode_null(get_tarval_mode(tb)))) /* div by zero: return tarval_bad */
266 return tarval_mod(ta, tb);
267 else if (tb == get_mode_one(get_tarval_mode(tb))) /* x mod 1 == 0 */
268 return get_mode_null(get_irn_mode(a));
271 } /* do_computed_value_Mod */
274 * Return the value of an integer Mod.
276 static tarval *computed_value_Mod(ir_node *n) {
277 return do_computed_value_Mod(get_Mod_left(n), get_Mod_right(n));
278 } /* computed_value_Mod */
281 * Return the value of an Abs.
283 static tarval *computed_value_Abs(ir_node *n) {
284 ir_node *a = get_Abs_op(n);
285 tarval *ta = value_of(a);
287 if (ta != tarval_bad)
288 return tarval_abs(ta);
291 } /* computed_value_Abs */
294 * Return the value of an And.
295 * Special case: a & 0, 0 & b
297 static tarval *computed_value_And(ir_node *n) {
298 ir_node *a = get_And_left(n);
299 ir_node *b = get_And_right(n);
301 tarval *ta = value_of(a);
302 tarval *tb = value_of(b);
304 if ((ta != tarval_bad) && (tb != tarval_bad)) {
305 return tarval_and (ta, tb);
309 if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_NULL)
310 || (classify_tarval ((v = tb)) == TV_CLASSIFY_NULL)) {
315 } /* computed_value_And */
318 * Return the value of an Or.
319 * Special case: a | 1...1, 1...1 | b
321 static tarval *computed_value_Or(ir_node *n) {
322 ir_node *a = get_Or_left(n);
323 ir_node *b = get_Or_right(n);
325 tarval *ta = value_of(a);
326 tarval *tb = value_of(b);
328 if ((ta != tarval_bad) && (tb != tarval_bad)) {
329 return tarval_or (ta, tb);
332 if ( (classify_tarval ((v = ta)) == TV_CLASSIFY_ALL_ONE)
333 || (classify_tarval ((v = tb)) == TV_CLASSIFY_ALL_ONE)) {
338 } /* computed_value_Or */
341 * Return the value of an Eor.
343 static tarval *computed_value_Eor(ir_node *n) {
344 ir_node *a = get_Eor_left(n);
345 ir_node *b = get_Eor_right(n);
350 return get_mode_null(get_irn_mode(n));
355 if ((ta != tarval_bad) && (tb != tarval_bad)) {
356 return tarval_eor (ta, tb);
359 } /* computed_value_Eor */
362 * Return the value of a Not.
364 static tarval *computed_value_Not(ir_node *n) {
365 ir_node *a = get_Not_op(n);
366 tarval *ta = value_of(a);
368 if (ta != tarval_bad)
369 return tarval_not(ta);
372 } /* computed_value_Not */
375 * Return the value of a Shl.
377 static tarval *computed_value_Shl(ir_node *n) {
378 ir_node *a = get_Shl_left(n);
379 ir_node *b = get_Shl_right(n);
381 tarval *ta = value_of(a);
382 tarval *tb = value_of(b);
384 if ((ta != tarval_bad) && (tb != tarval_bad)) {
385 return tarval_shl (ta, tb);
388 } /* computed_value_Shl */
391 * Return the value of a Shr.
393 static tarval *computed_value_Shr(ir_node *n) {
394 ir_node *a = get_Shr_left(n);
395 ir_node *b = get_Shr_right(n);
397 tarval *ta = value_of(a);
398 tarval *tb = value_of(b);
400 if ((ta != tarval_bad) && (tb != tarval_bad)) {
401 return tarval_shr (ta, tb);
404 } /* computed_value_Shr */
407 * Return the value of a Shrs.
409 static tarval *computed_value_Shrs(ir_node *n) {
410 ir_node *a = get_Shrs_left(n);
411 ir_node *b = get_Shrs_right(n);
413 tarval *ta = value_of(a);
414 tarval *tb = value_of(b);
416 if ((ta != tarval_bad) && (tb != tarval_bad)) {
417 return tarval_shrs (ta, tb);
420 } /* computed_value_Shrs */
423 * Return the value of a Rot.
425 static tarval *computed_value_Rot(ir_node *n) {
426 ir_node *a = get_Rot_left(n);
427 ir_node *b = get_Rot_right(n);
429 tarval *ta = value_of(a);
430 tarval *tb = value_of(b);
432 if ((ta != tarval_bad) && (tb != tarval_bad)) {
433 return tarval_rot (ta, tb);
436 } /* computed_value_Rot */
439 * Return the value of a Conv.
441 static tarval *computed_value_Conv(ir_node *n) {
442 ir_node *a = get_Conv_op(n);
443 tarval *ta = value_of(a);
445 if (ta != tarval_bad)
446 return tarval_convert_to(ta, get_irn_mode(n));
449 } /* computed_value_Conv */
452 * Return the value of a Proj(Cmp).
454 * This performs a first step of unreachable code elimination.
455 * Proj can not be computed, but folding a Cmp above the Proj here is
456 * not as wasteful as folding a Cmp into a Tuple of 16 Consts of which
458 * There are several case where we can evaluate a Cmp node, see later.
460 static tarval *computed_value_Proj_Cmp(ir_node *n) {
461 ir_node *a = get_Proj_pred(n);
462 ir_node *aa = get_Cmp_left(a);
463 ir_node *ab = get_Cmp_right(a);
464 long proj_nr = get_Proj_proj(n);
467 * BEWARE: a == a is NOT always True for floating Point values, as
468 * NaN != NaN is defined, so we must check this here.
471 !mode_is_float(get_irn_mode(aa)) || proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Gt)
474 /* This is a trick with the bits used for encoding the Cmp
475 Proj numbers, the following statement is not the same:
476 return new_tarval_from_long (proj_nr == pn_Cmp_Eq, mode_b) */
477 return new_tarval_from_long (proj_nr & pn_Cmp_Eq, mode_b);
480 tarval *taa = value_of(aa);
481 tarval *tab = value_of(ab);
482 ir_mode *mode = get_irn_mode(aa);
485 * The predecessors of Cmp are target values. We can evaluate
488 if ((taa != tarval_bad) && (tab != tarval_bad)) {
489 /* strange checks... */
490 pn_Cmp flags = tarval_cmp(taa, tab);
491 if (flags != pn_Cmp_False) {
492 return new_tarval_from_long (proj_nr & flags, mode_b);
495 /* for integer values, we can check against MIN/MAX */
496 else if (mode_is_int(mode)) {
497 /* MIN <=/> x. This results in true/false. */
498 if (taa == get_mode_min(mode)) {
499 /* a compare with the MIN value */
500 if (proj_nr == pn_Cmp_Le)
501 return get_tarval_b_true();
502 else if (proj_nr == pn_Cmp_Gt)
503 return get_tarval_b_false();
505 /* x >=/< MIN. This results in true/false. */
507 if (tab == get_mode_min(mode)) {
508 /* a compare with the MIN value */
509 if (proj_nr == pn_Cmp_Ge)
510 return get_tarval_b_true();
511 else if (proj_nr == pn_Cmp_Lt)
512 return get_tarval_b_false();
514 /* MAX >=/< x. This results in true/false. */
515 else if (taa == get_mode_max(mode)) {
516 if (proj_nr == pn_Cmp_Ge)
517 return get_tarval_b_true();
518 else if (proj_nr == pn_Cmp_Lt)
519 return get_tarval_b_false();
521 /* x <=/> MAX. This results in true/false. */
522 else if (tab == get_mode_max(mode)) {
523 if (proj_nr == pn_Cmp_Le)
524 return get_tarval_b_true();
525 else if (proj_nr == pn_Cmp_Gt)
526 return get_tarval_b_false();
530 * The predecessors are Allocs or (void*)(0) constants. Allocs never
531 * return NULL, they raise an exception. Therefore we can predict
535 ir_node *aaa = skip_Id(skip_Proj(aa));
536 ir_node *aba = skip_Id(skip_Proj(ab));
538 if ( ( (/* aa is ProjP and aaa is Alloc */
539 (get_irn_op(aa) == op_Proj)
540 && (mode_is_reference(get_irn_mode(aa)))
541 && (get_irn_op(aaa) == op_Alloc))
542 && ( (/* ab is NULL */
543 (get_irn_op(ab) == op_Const)
544 && (mode_is_reference(get_irn_mode(ab)))
545 && (get_Const_tarval(ab) == get_mode_null(get_irn_mode(ab))))
546 || (/* ab is other Alloc */
547 (get_irn_op(ab) == op_Proj)
548 && (mode_is_reference(get_irn_mode(ab)))
549 && (get_irn_op(aba) == op_Alloc)
551 || (/* aa is NULL and aba is Alloc */
552 (get_irn_op(aa) == op_Const)
553 && (mode_is_reference(get_irn_mode(aa)))
554 && (get_Const_tarval(aa) == get_mode_null(get_irn_mode(aa)))
555 && (get_irn_op(ab) == op_Proj)
556 && (mode_is_reference(get_irn_mode(ab)))
557 && (get_irn_op(aba) == op_Alloc)))
559 return new_tarval_from_long(proj_nr & pn_Cmp_Ne, mode_b);
562 return computed_value_Cmp_Confirm(a, aa, ab, proj_nr);
563 } /* computed_value_Proj_Cmp */
566 * Return the value of a Proj, handle Proj(Cmp), Proj(Div), Proj(Mod),
567 * Proj(DivMod) and Proj(Quot).
569 static tarval *computed_value_Proj(ir_node *n) {
570 ir_node *a = get_Proj_pred(n);
573 switch (get_irn_opcode(a)) {
575 return computed_value_Proj_Cmp(n);
578 /* compute either the Div or the Mod part */
579 proj_nr = get_Proj_proj(n);
580 if (proj_nr == pn_DivMod_res_div)
581 return do_computed_value_Div(get_DivMod_left(a), get_DivMod_right(a));
582 else if (proj_nr == pn_DivMod_res_mod)
583 return do_computed_value_Mod(get_DivMod_left(a), get_DivMod_right(a));
587 if (get_Proj_proj(n) == pn_Div_res)
588 return computed_value(a);
592 if (get_Proj_proj(n) == pn_Mod_res)
593 return computed_value(a);
597 if (get_Proj_proj(n) == pn_Quot_res)
598 return computed_value(a);
605 } /* computed_value_Proj */
608 * Calculate the value of a Mux: can be evaluated, if the
609 * sel and the right input are known.
611 static tarval *computed_value_Mux(ir_node *n) {
612 ir_node *sel = get_Mux_sel(n);
613 tarval *ts = value_of(sel);
615 if (ts == get_tarval_b_true()) {
616 ir_node *v = get_Mux_true(n);
619 else if (ts == get_tarval_b_false()) {
620 ir_node *v = get_Mux_false(n);
624 } /* computed_value_Mux */
627 * Calculate the value of a Psi: can be evaluated, if a condition is true
628 * and all previous conditions are false. If all conditions are false
629 * we evaluate to the default one.
631 static tarval *computed_value_Psi(ir_node *n) {
633 return computed_value_Mux(n);
635 } /* computed_value_Psi */
638 * Calculate the value of a Confirm: can be evaluated,
639 * if it has the form Confirm(x, '=', Const).
641 static tarval *computed_value_Confirm(ir_node *n) {
642 return get_Confirm_cmp(n) == pn_Cmp_Eq ?
643 value_of(get_Confirm_bound(n)) : tarval_bad;
644 } /* computed_value_Confirm */
647 * If the parameter n can be computed, return its value, else tarval_bad.
648 * Performs constant folding.
650 * @param n The node this should be evaluated
652 tarval *computed_value(ir_node *n) {
653 if (n->op->ops.computed_value)
654 return n->op->ops.computed_value(n);
656 } /* computed_value */
659 * Set the default computed_value evaluator in an ir_op_ops.
661 * @param code the opcode for the default operation
662 * @param ops the operations initialized
667 static ir_op_ops *firm_set_default_computed_value(ir_opcode code, ir_op_ops *ops)
671 ops->computed_value = computed_value_##a; \
706 } /* firm_set_default_computed_value */
709 * Returns a equivalent block for another block.
710 * If the block has only one predecessor, this is
711 * the equivalent one. If the only predecessor of a block is
712 * the block itself, this is a dead block.
714 * If both predecessors of a block are the branches of a binary
715 * Cond, the equivalent block is Cond's block.
717 * If all predecessors of a block are bad or lies in a dead
718 * block, the current block is dead as well.
720 * Note, that blocks are NEVER turned into Bad's, instead
721 * the dead_block flag is set. So, never test for is_Bad(block),
722 * always use is_dead_Block(block).
724 static ir_node *equivalent_node_Block(ir_node *n)
727 int n_preds = get_Block_n_cfgpreds(n);
729 /* The Block constructor does not call optimize, but mature_immBlock
730 calls the optimization. */
731 assert(get_Block_matured(n));
733 /* Straightening: a single entry Block following a single exit Block
734 can be merged, if it is not the Start block. */
735 /* !!! Beware, all Phi-nodes of n must have been optimized away.
736 This should be true, as the block is matured before optimize is called.
737 But what about Phi-cycles with the Phi0/Id that could not be resolved?
738 Remaining Phi nodes are just Ids. */
739 if ((n_preds == 1) && (get_irn_op(get_Block_cfgpred(n, 0)) == op_Jmp)) {
740 ir_node *predblock = get_nodes_block(get_Block_cfgpred(n, 0));
741 if (predblock == oldn) {
742 /* Jmp jumps into the block it is in -- deal self cycle. */
743 n = set_Block_dead(n);
744 DBG_OPT_DEAD_BLOCK(oldn, n);
745 } else if (get_opt_control_flow_straightening()) {
747 DBG_OPT_STG(oldn, n);
749 } else if ((n_preds == 1) &&
750 (get_irn_op(skip_Proj(get_Block_cfgpred(n, 0))) == op_Cond)) {
751 ir_node *predblock = get_Block_cfgpred_block(n, 0);
752 if (predblock == oldn) {
753 /* Jmp jumps into the block it is in -- deal self cycle. */
754 n = set_Block_dead(n);
755 DBG_OPT_DEAD_BLOCK(oldn, n);
757 } else if ((n_preds == 2) &&
758 (get_opt_control_flow_weak_simplification())) {
759 /* Test whether Cond jumps twice to this block
760 * The more general case which more than 2 predecessors is handles
761 * in optimize_cf(), we handle only this special case for speed here.
763 ir_node *a = get_Block_cfgpred(n, 0);
764 ir_node *b = get_Block_cfgpred(n, 1);
766 if ((get_irn_op(a) == op_Proj) &&
767 (get_irn_op(b) == op_Proj) &&
768 (get_Proj_pred(a) == get_Proj_pred(b)) &&
769 (get_irn_op(get_Proj_pred(a)) == op_Cond) &&
770 (get_irn_mode(get_Cond_selector(get_Proj_pred(a))) == mode_b)) {
771 /* Also a single entry Block following a single exit Block. Phis have
772 twice the same operand and will be optimized away. */
773 n = get_nodes_block(get_Proj_pred(a));
774 DBG_OPT_IFSIM1(oldn, a, b, n);
776 } else if (get_opt_unreachable_code() &&
777 (n != get_irg_start_block(current_ir_graph)) &&
778 (n != get_irg_end_block(current_ir_graph)) ) {
781 /* If all inputs are dead, this block is dead too, except if it is
782 the start or end block. This is one step of unreachable code
784 for (i = get_Block_n_cfgpreds(n) - 1; i >= 0; --i) {
785 ir_node *pred = get_Block_cfgpred(n, i);
788 if (is_Bad(pred)) continue;
789 pred_blk = get_nodes_block(skip_Proj(pred));
791 if (is_Block_dead(pred_blk)) continue;
794 /* really found a living input */
799 n = set_Block_dead(n);
800 DBG_OPT_DEAD_BLOCK(oldn, n);
805 } /* equivalent_node_Block */
808 * Returns a equivalent node for a Jmp, a Bad :-)
809 * Of course this only happens if the Block of the Jmp is dead.
811 static ir_node *equivalent_node_Jmp(ir_node *n) {
812 /* unreachable code elimination */
813 if (is_Block_dead(get_nodes_block(n)))
817 } /* equivalent_node_Jmp */
819 /** Raise is handled in the same way as Jmp. */
820 #define equivalent_node_Raise equivalent_node_Jmp
823 /* We do not evaluate Cond here as we replace it by a new node, a Jmp.
824 See transform_node_Proj_Cond(). */
827 * Optimize operations that are commutative and have neutral 0,
828 * so a op 0 = 0 op a = a.
830 static ir_node *equivalent_node_neutral_zero(ir_node *n)
834 ir_node *a = get_binop_left(n);
835 ir_node *b = get_binop_right(n);
840 /* After running compute_node there is only one constant predecessor.
841 Find this predecessors value and remember the other node: */
842 if ((tv = value_of(a)) != tarval_bad) {
844 } else if ((tv = value_of(b)) != tarval_bad) {
849 /* If this predecessors constant value is zero, the operation is
850 * unnecessary. Remove it.
852 * Beware: If n is a Add, the mode of on and n might be different
853 * which happens in this rare construction: NULL + 3.
854 * Then, a Conv would be needed which we cannot include here.
856 if (classify_tarval (tv) == TV_CLASSIFY_NULL) {
857 if (get_irn_mode(on) == get_irn_mode(n)) {
860 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
865 } /* equivalent_node_neutral_zero */
868 * Eor is commutative and has neutral 0.
870 #define equivalent_node_Eor equivalent_node_neutral_zero
873 * Optimize a - 0 and (a - x) + x (for modes with wrap-around).
875 * The second one looks strange, but this construct
876 * is used heavily in the LCC sources :-).
878 * Beware: The Mode of an Add may be different than the mode of its
879 * predecessors, so we could not return a predecessors in all cases.
881 static ir_node *equivalent_node_Add(ir_node *n) {
883 ir_node *left, *right;
884 ir_mode *mode = get_irn_mode(n);
886 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
887 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
890 n = equivalent_node_neutral_zero(n);
894 left = get_Add_left(n);
895 right = get_Add_right(n);
897 if (get_irn_op(left) == op_Sub) {
898 if (get_Sub_right(left) == right) {
901 n = get_Sub_left(left);
902 if (mode == get_irn_mode(n)) {
903 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
908 if (get_irn_op(right) == op_Sub) {
909 if (get_Sub_right(right) == left) {
912 n = get_Sub_left(right);
913 if (mode == get_irn_mode(n)) {
914 DBG_OPT_ALGSIM1(oldn, left, right, n, FS_OPT_ADD_SUB);
920 } /* equivalent_node_Add */
923 * optimize operations that are not commutative but have neutral 0 on left,
926 static ir_node *equivalent_node_left_zero(ir_node *n) {
929 ir_node *a = get_binop_left(n);
930 ir_node *b = get_binop_right(n);
932 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
935 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
938 } /* equivalent_node_left_zero */
940 #define equivalent_node_Shl equivalent_node_left_zero
941 #define equivalent_node_Shr equivalent_node_left_zero
942 #define equivalent_node_Shrs equivalent_node_left_zero
943 #define equivalent_node_Rot equivalent_node_left_zero
946 * Optimize a - 0 and (a + x) - x (for modes with wrap-around).
948 * The second one looks strange, but this construct
949 * is used heavily in the LCC sources :-).
951 * Beware: The Mode of a Sub may be different than the mode of its
952 * predecessors, so we could not return a predecessors in all cases.
954 static ir_node *equivalent_node_Sub(ir_node *n) {
957 ir_mode *mode = get_irn_mode(n);
959 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
960 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
963 b = get_Sub_right(n);
965 /* Beware: modes might be different */
966 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
967 ir_node *a = get_Sub_left(n);
968 if (mode == get_irn_mode(a)) {
971 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
975 } /* equivalent_node_Sub */
979 * Optimize an "idempotent unary op", ie op(op(n)) = n.
982 * -(-a) == a, but might overflow two times.
983 * We handle it anyway here but the better way would be a
984 * flag. This would be needed for Pascal for instance.
986 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
988 ir_node *pred = get_unop_op(n);
990 /* optimize symmetric unop */
991 if (get_irn_op(pred) == get_irn_op(n)) {
992 n = get_unop_op(pred);
993 DBG_OPT_ALGSIM2(oldn, pred, n, FS_OPT_IDEM_UNARY);
996 } /* equivalent_node_idempotent_unop */
998 /** Optimize Not(Not(x)) == x. */
999 #define equivalent_node_Not equivalent_node_idempotent_unop
1001 /** -(-x) == x ??? Is this possible or can --x raise an
1002 out of bounds exception if min =! max? */
1003 #define equivalent_node_Minus equivalent_node_idempotent_unop
1006 * Optimize a * 1 = 1 * a = a.
1008 static ir_node *equivalent_node_Mul(ir_node *n) {
1010 ir_node *a = get_Mul_left(n);
1011 ir_node *b = get_Mul_right(n);
1013 /* Mul is commutative and has again an other neutral element. */
1014 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
1016 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1017 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
1019 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1022 } /* equivalent_node_Mul */
1025 * Optimize a / 1 = a.
1027 static ir_node *equivalent_node_Div(ir_node *n) {
1028 ir_node *a = get_Div_left(n);
1029 ir_node *b = get_Div_right(n);
1031 /* Div is not commutative. */
1032 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1033 /* Turn Div into a tuple (mem, bad, a) */
1034 ir_node *mem = get_Div_mem(n);
1035 ir_node *blk = get_irn_n(n, -1);
1036 turn_into_tuple(n, pn_Div_max);
1037 set_Tuple_pred(n, pn_Div_M, mem);
1038 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
1039 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1040 set_Tuple_pred(n, pn_Div_res, a);
1043 } /* equivalent_node_Div */
1046 * Optimize a / 1.0 = a.
1048 static ir_node *equivalent_node_Quot(ir_node *n) {
1049 ir_node *a = get_Quot_left(n);
1050 ir_node *b = get_Quot_right(n);
1052 /* Div is not commutative. */
1053 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* Quot(x, 1) == x */
1054 /* Turn Quot into a tuple (mem, jmp, bad, a) */
1055 ir_node *mem = get_Quot_mem(n);
1056 ir_node *blk = get_irn_n(n, -1);
1057 turn_into_tuple(n, pn_Quot_max);
1058 set_Tuple_pred(n, pn_Quot_M, mem);
1059 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
1060 set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */
1061 set_Tuple_pred(n, pn_Quot_res, a);
1064 } /* equivalent_node_Quot */
1067 * Optimize a / 1 = a.
1069 static ir_node *equivalent_node_DivMod(ir_node *n) {
1070 ir_node *b = get_DivMod_right(n);
1072 /* Div is not commutative. */
1073 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1074 /* Turn DivMod into a tuple (mem, jmp, bad, a, 0) */
1075 ir_node *a = get_DivMod_left(n);
1076 ir_node *mem = get_Div_mem(n);
1077 ir_node *blk = get_irn_n(n, -1);
1078 ir_mode *mode = get_DivMod_resmode(n);
1080 turn_into_tuple(n, pn_DivMod_max);
1081 set_Tuple_pred(n, pn_DivMod_M, mem);
1082 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
1083 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1084 set_Tuple_pred(n, pn_DivMod_res_div, a);
1085 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1088 } /* equivalent_node_DivMod */
1091 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1093 static ir_node *equivalent_node_Or(ir_node *n) {
1096 ir_node *a = get_Or_left(n);
1097 ir_node *b = get_Or_right(n);
1100 n = a; /* Or has it's own neutral element */
1101 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1102 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1104 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1105 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1107 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1111 } /* equivalent_node_Or */
1114 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1116 static ir_node *equivalent_node_And(ir_node *n) {
1119 ir_node *a = get_And_left(n);
1120 ir_node *b = get_And_right(n);
1123 n = a; /* And has it's own neutral element */
1124 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1125 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1127 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1128 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1130 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1133 } /* equivalent_node_And */
1136 * Try to remove useless Conv's:
1138 static ir_node *equivalent_node_Conv(ir_node *n) {
1140 ir_node *a = get_Conv_op(n);
1143 ir_mode *n_mode = get_irn_mode(n);
1144 ir_mode *a_mode = get_irn_mode(a);
1146 if (n_mode == a_mode) { /* No Conv necessary */
1147 /* leave strict floating point Conv's */
1148 if (get_Conv_strict(n))
1151 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1152 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1156 n_mode = get_irn_mode(n);
1157 b_mode = get_irn_mode(b);
1159 if (n_mode == b_mode) {
1160 if (n_mode == mode_b) {
1161 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1162 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1163 } else if (mode_is_int(n_mode)) {
1164 if (smaller_mode(b_mode, a_mode)){
1165 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1166 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1172 } /* equivalent_node_Conv */
1175 * A Cast may be removed if the type of the previous node
1176 * is already the type of the Cast.
1178 static ir_node *equivalent_node_Cast(ir_node *n) {
1180 ir_node *pred = get_Cast_op(n);
1182 if (get_irn_type(pred) == get_Cast_type(n)) {
1184 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1187 } /* equivalent_node_Cast */
1190 * Several optimizations:
1191 * - no Phi in start block.
1192 * - remove Id operators that are inputs to Phi
1193 * - fold Phi-nodes, iff they have only one predecessor except
1196 static ir_node *equivalent_node_Phi(ir_node *n) {
1200 ir_node *block = NULL; /* to shutup gcc */
1201 ir_node *first_val = NULL; /* to shutup gcc */
1203 if (!get_opt_normalize()) return n;
1205 n_preds = get_Phi_n_preds(n);
1207 block = get_nodes_block(n);
1208 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1209 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1210 if ((is_Block_dead(block)) || /* Control dead */
1211 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1212 return new_Bad(); /* in the Start Block. */
1214 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1216 /* If the Block has a Bad pred, we also have one. */
1217 for (i = 0; i < n_preds; ++i)
1218 if (is_Bad(get_Block_cfgpred(block, i)))
1219 set_Phi_pred(n, i, new_Bad());
1221 /* Find first non-self-referencing input */
1222 for (i = 0; i < n_preds; ++i) {
1223 first_val = get_Phi_pred(n, i);
1224 if ( (first_val != n) /* not self pointer */
1226 && (! is_Bad(first_val))
1228 ) { /* value not dead */
1229 break; /* then found first value. */
1234 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1238 /* search for rest of inputs, determine if any of these
1239 are non-self-referencing */
1240 while (++i < n_preds) {
1241 ir_node *scnd_val = get_Phi_pred(n, i);
1242 if ( (scnd_val != n)
1243 && (scnd_val != first_val)
1245 && (! is_Bad(scnd_val))
1253 /* Fold, if no multiple distinct non-self-referencing inputs */
1255 DBG_OPT_PHI(oldn, n);
1258 } /* equivalent_node_Phi */
1261 * Several optimizations:
1262 * - no Sync in start block.
1263 * - fold Sync-nodes, iff they have only one predecessor except
1266 static ir_node *equivalent_node_Sync(ir_node *n) {
1270 ir_node *first_val = NULL; /* to shutup gcc */
1272 if (!get_opt_normalize()) return n;
1274 n_preds = get_Sync_n_preds(n);
1276 /* Find first non-self-referencing input */
1277 for (i = 0; i < n_preds; ++i) {
1278 first_val = get_Sync_pred(n, i);
1279 if ((first_val != n) /* not self pointer */ &&
1280 (! is_Bad(first_val))
1281 ) { /* value not dead */
1282 break; /* then found first value. */
1287 /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
1290 /* search the rest of inputs, determine if any of these
1291 are non-self-referencing */
1292 while (++i < n_preds) {
1293 ir_node *scnd_val = get_Sync_pred(n, i);
1294 if ((scnd_val != n) &&
1295 (scnd_val != first_val) &&
1296 (! is_Bad(scnd_val))
1302 /* Fold, if no multiple distinct non-self-referencing inputs */
1304 DBG_OPT_SYNC(oldn, n);
1307 } /* equivalent_node_Sync */
1310 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1311 * ProjX(Load) and ProjX(Store).
1313 static ir_node *equivalent_node_Proj(ir_node *proj) {
1314 ir_node *oldn = proj;
1315 ir_node *a = get_Proj_pred(proj);
1317 if (get_irn_op(a) == op_Tuple) {
1318 /* Remove the Tuple/Proj combination. */
1319 if ( get_Proj_proj(proj) <= get_Tuple_n_preds(a) ) {
1320 proj = get_Tuple_pred(a, get_Proj_proj(proj));
1321 DBG_OPT_TUPLE(oldn, a, proj);
1323 /* This should not happen! */
1324 assert(! "found a Proj with higher number than Tuple predecessors");
1327 } else if (get_irn_mode(proj) == mode_X) {
1328 if (is_Block_dead(get_nodes_block(skip_Proj(proj)))) {
1329 /* Remove dead control flow -- early gigo(). */
1331 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1332 ir_op *op = get_irn_op(a);
1334 if (op == op_Load) {
1335 /* get the Load address */
1336 ir_node *addr = get_Load_ptr(a);
1337 ir_node *blk = get_irn_n(a, -1);
1340 if (value_not_null(addr, &confirm)) {
1341 if (confirm == NULL) {
1342 /* this node may float if it did not depend on a Confirm */
1343 set_irn_pinned(a, op_pin_state_floats);
1345 if (get_Proj_proj(proj) == pn_Load_X_except) {
1346 DBG_OPT_EXC_REM(proj);
1349 return new_r_Jmp(current_ir_graph, blk);
1351 } else if (op == op_Store) {
1352 /* get the load/store address */
1353 ir_node *addr = get_Store_ptr(a);
1354 ir_node *blk = get_irn_n(a, -1);
1357 if (value_not_null(addr, &confirm)) {
1358 if (confirm == NULL) {
1359 /* this node may float if it did not depend on a Confirm */
1360 set_irn_pinned(a, op_pin_state_floats);
1362 if (get_Proj_proj(proj) == pn_Store_X_except) {
1363 DBG_OPT_EXC_REM(proj);
1366 return new_r_Jmp(current_ir_graph, blk);
1373 } /* equivalent_node_Proj */
1378 static ir_node *equivalent_node_Id(ir_node *n) {
1383 } while (get_irn_op(n) == op_Id);
1385 DBG_OPT_ID(oldn, n);
1387 } /* equivalent_node_Id */
1392 static ir_node *equivalent_node_Mux(ir_node *n)
1394 ir_node *oldn = n, *sel = get_Mux_sel(n);
1395 tarval *ts = value_of(sel);
1397 /* Mux(true, f, t) == t */
1398 if (ts == tarval_b_true) {
1399 n = get_Mux_true(n);
1400 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1402 /* Mux(false, f, t) == f */
1403 else if (ts == tarval_b_false) {
1404 n = get_Mux_false(n);
1405 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1407 /* Mux(v, x, x) == x */
1408 else if (get_Mux_false(n) == get_Mux_true(n)) {
1409 n = get_Mux_true(n);
1410 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1412 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1413 ir_node *cmp = get_Proj_pred(sel);
1414 long proj_nr = get_Proj_proj(sel);
1415 ir_node *b = get_Mux_false(n);
1416 ir_node *a = get_Mux_true(n);
1419 * Note: normalization puts the constant on the right site,
1420 * so we check only one case.
1422 * Note further that these optimization work even for floating point
1423 * with NaN's because -NaN == NaN.
1424 * However, if +0 and -0 is handled differently, we cannot use the first one.
1426 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1427 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1428 /* Mux(a CMP 0, X, a) */
1429 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1430 /* Mux(a CMP 0, -a, a) */
1431 if (proj_nr == pn_Cmp_Eq) {
1432 /* Mux(a == 0, -a, a) ==> -a */
1434 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1435 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1436 /* Mux(a != 0, -a, a) ==> a */
1438 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1440 } else if (classify_Const(b) == CNST_NULL) {
1441 /* Mux(a CMP 0, 0, a) */
1442 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1443 /* Mux(a != 0, 0, a) ==> a */
1445 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1446 } else if (proj_nr == pn_Cmp_Eq) {
1447 /* Mux(a == 0, 0, a) ==> 0 */
1449 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1456 } /* equivalent_node_Mux */
1459 * Returns a equivalent node of a Psi: if a condition is true
1460 * and all previous conditions are false we know its value.
1461 * If all conditions are false its value is the default one.
1463 static ir_node *equivalent_node_Psi(ir_node *n) {
1465 return equivalent_node_Mux(n);
1467 } /* equivalent_node_Psi */
1470 * Optimize -a CMP -b into b CMP a.
1471 * This works only for for modes where unary Minus
1473 * Note that two-complement integers can Overflow
1474 * so it will NOT work.
1476 * For == and != can be handled in Proj(Cmp)
1478 static ir_node *equivalent_node_Cmp(ir_node *n) {
1479 ir_node *left = get_Cmp_left(n);
1480 ir_node *right = get_Cmp_right(n);
1482 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1483 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1484 left = get_Minus_op(left);
1485 right = get_Minus_op(right);
1486 set_Cmp_left(n, right);
1487 set_Cmp_right(n, left);
1490 } /* equivalent_node_Cmp */
1493 * Remove Confirm nodes if setting is on.
1494 * Replace Confirms(x, '=', Constlike) by Constlike.
1496 static ir_node *equivalent_node_Confirm(ir_node *n) {
1497 ir_node *pred = get_Confirm_value(n);
1498 pn_Cmp pnc = get_Confirm_cmp(n);
1500 if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) {
1502 * rare case: two identical Confirms one after another,
1503 * replace the second one with the first.
1507 if (pnc == pn_Cmp_Eq) {
1508 ir_node *bound = get_Confirm_bound(n);
1511 * Optimize a rare case:
1512 * Confirm(x, '=', Constlike) ==> Constlike
1514 if (is_irn_constlike(bound)) {
1515 DBG_OPT_CONFIRM(n, bound);
1519 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1523 * Optimize CopyB(mem, x, x) into a Nop.
1525 static ir_node *equivalent_node_CopyB(ir_node *n) {
1526 ir_node *a = get_CopyB_dst(n);
1527 ir_node *b = get_CopyB_src(n);
1530 /* Turn CopyB into a tuple (mem, jmp, bad, bad) */
1531 ir_node *mem = get_CopyB_mem(n);
1532 ir_node *blk = get_nodes_block(n);
1533 turn_into_tuple(n, pn_CopyB_max);
1534 set_Tuple_pred(n, pn_CopyB_M, mem);
1535 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1536 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1537 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1540 } /* equivalent_node_CopyB */
1543 * Optimize Bounds(idx, idx, upper) into idx.
1545 static ir_node *equivalent_node_Bound(ir_node *n) {
1546 ir_node *idx = get_Bound_index(n);
1547 ir_node *lower = get_Bound_lower(n);
1550 /* By definition lower < upper, so if idx == lower -->
1551 lower <= idx && idx < upper */
1553 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1556 ir_node *pred = skip_Proj(idx);
1558 if (get_irn_op(pred) == op_Bound) {
1560 * idx was Bounds_check previously, it is still valid if
1561 * lower <= pred_lower && pred_upper <= upper.
1563 ir_node *upper = get_Bound_upper(n);
1564 if (get_Bound_lower(pred) == lower &&
1565 get_Bound_upper(pred) == upper) {
1567 * One could expect that we simply return the previous
1568 * Bound here. However, this would be wrong, as we could
1569 * add an exception Proj to a new location then.
1570 * So, we must turn in into a tuple.
1577 /* Turn Bound into a tuple (mem, jmp, bad, idx) */
1578 ir_node *mem = get_Bound_mem(n);
1579 ir_node *blk = get_nodes_block(n);
1580 turn_into_tuple(n, pn_Bound_max);
1581 set_Tuple_pred(n, pn_Bound_M, mem);
1582 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1583 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1584 set_Tuple_pred(n, pn_Bound_res, idx);
1587 } /* equivalent_node_Bound */
1590 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1591 * perform no actual computation, as, e.g., the Id nodes. It does not create
1592 * new nodes. It is therefore safe to free n if the node returned is not n.
1593 * If a node returns a Tuple we can not just skip it. If the size of the
1594 * in array fits, we transform n into a tuple (e.g., Div).
1596 ir_node *equivalent_node(ir_node *n) {
1597 if (n->op->ops.equivalent_node)
1598 return n->op->ops.equivalent_node(n);
1600 } /* equivalent_node */
1603 * Sets the default equivalent node operation for an ir_op_ops.
1605 * @param code the opcode for the default operation
1606 * @param ops the operations initialized
1611 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1615 ops->equivalent_node = equivalent_node_##a; \
1655 } /* firm_set_default_equivalent_node */
1658 * Do node specific optimizations of nodes predecessors.
1660 static void optimize_preds(ir_node *n) {
1661 switch (get_irn_opcode(n)) {
1663 case iro_Cmp: { /* We don't want Cast as input to Cmp. */
1664 ir_node *a = get_Cmp_left(n), *b = get_Cmp_right(n);
1666 if (get_irn_op(a) == op_Cast) {
1670 if (get_irn_op(b) == op_Cast) {
1672 set_Cmp_right(n, b);
1679 } /* optimize_preds */
1682 * Returns non-zero if a node is a Phi node
1683 * with all predecessors constant.
1685 static int is_const_Phi(ir_node *n) {
1690 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1691 if (! is_Const(get_irn_n(n, i)))
1694 } /* is_const_Phi */
1697 * Apply an evaluator on a binop with a constant operators (and one Phi).
1699 * @param phi the Phi node
1700 * @param other the other operand
1701 * @param eval an evaluator function
1702 * @param left if non-zero, other is the left operand, else the right
1704 * @return a new Phi node if the conversion was successful, NULL else
1706 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), int left) {
1712 int i, n = get_irn_arity(phi);
1714 NEW_ARR_A(void *, res, n);
1716 for (i = 0; i < n; ++i) {
1717 pred = get_irn_n(phi, i);
1718 tv = get_Const_tarval(pred);
1719 tv = eval(other, tv);
1721 if (tv == tarval_bad) {
1722 /* folding failed, bad */
1728 for (i = 0; i < n; ++i) {
1729 pred = get_irn_n(phi, i);
1730 tv = get_Const_tarval(pred);
1731 tv = eval(tv, other);
1733 if (tv == tarval_bad) {
1734 /* folding failed, bad */
1740 mode = get_irn_mode(phi);
1741 irg = current_ir_graph;
1742 for (i = 0; i < n; ++i) {
1743 pred = get_irn_n(phi, i);
1744 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1745 mode, res[i], get_Const_type(pred));
1747 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1748 } /* apply_binop_on_phi */
1751 * Apply an evaluator on a unop with a constant operator (a Phi).
1753 * @param phi the Phi node
1754 * @param eval an evaluator function
1756 * @return a new Phi node if the conversion was successful, NULL else
1758 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1764 int i, n = get_irn_arity(phi);
1766 NEW_ARR_A(void *, res, n);
1767 for (i = 0; i < n; ++i) {
1768 pred = get_irn_n(phi, i);
1769 tv = get_Const_tarval(pred);
1772 if (tv == tarval_bad) {
1773 /* folding failed, bad */
1778 mode = get_irn_mode(phi);
1779 irg = current_ir_graph;
1780 for (i = 0; i < n; ++i) {
1781 pred = get_irn_n(phi, i);
1782 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1783 mode, res[i], get_Const_type(pred));
1785 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1786 } /* apply_unop_on_phi */
1789 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1790 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1791 * If possible, remove the Conv's.
1793 static ir_node *transform_node_AddSub(ir_node *n) {
1794 ir_mode *mode = get_irn_mode(n);
1796 if (mode_is_reference(mode)) {
1797 ir_node *left = get_binop_left(n);
1798 ir_node *right = get_binop_right(n);
1799 int ref_bits = get_mode_size_bits(mode);
1801 if (get_irn_op(left) == op_Conv) {
1802 ir_mode *mode = get_irn_mode(left);
1803 int bits = get_mode_size_bits(mode);
1805 if (ref_bits == bits &&
1806 mode_is_int(mode) &&
1807 get_mode_arithmetic(mode) == irma_twos_complement) {
1808 ir_node *pre = get_Conv_op(left);
1809 ir_mode *pre_mode = get_irn_mode(pre);
1811 if (mode_is_int(pre_mode) &&
1812 get_mode_size_bits(pre_mode) == bits &&
1813 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1814 /* ok, this conv just changes to sign, moreover the calculation
1815 * is done with same number of bits as our address mode, so
1816 * we can ignore the conv as address calculation can be viewed
1817 * as either signed or unsigned
1819 set_binop_left(n, pre);
1824 if (get_irn_op(right) == op_Conv) {
1825 ir_mode *mode = get_irn_mode(right);
1826 int bits = get_mode_size_bits(mode);
1828 if (ref_bits == bits &&
1829 mode_is_int(mode) &&
1830 get_mode_arithmetic(mode) == irma_twos_complement) {
1831 ir_node *pre = get_Conv_op(right);
1832 ir_mode *pre_mode = get_irn_mode(pre);
1834 if (mode_is_int(pre_mode) &&
1835 get_mode_size_bits(pre_mode) == bits &&
1836 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1837 /* ok, this conv just changes to sign, moreover the calculation
1838 * is done with same number of bits as our address mode, so
1839 * we can ignore the conv as address calculation can be viewed
1840 * as either signed or unsigned
1842 set_binop_right(n, pre);
1848 } /* transform_node_AddSub */
1850 #define HANDLE_BINOP_PHI(op,a,b,c) \
1852 if (is_Const(b) && is_const_Phi(a)) { \
1853 /* check for Op(Phi, Const) */ \
1854 c = apply_binop_on_phi(a, get_Const_tarval(b), op, 0); \
1856 else if (is_Const(a) && is_const_Phi(b)) { \
1857 /* check for Op(Const, Phi) */ \
1858 c = apply_binop_on_phi(b, get_Const_tarval(a), op, 1); \
1861 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1865 #define HANDLE_UNOP_PHI(op,a,c) \
1867 if (is_const_Phi(a)) { \
1868 /* check for Op(Phi) */ \
1869 c = apply_unop_on_phi(a, op); \
1872 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1878 * Do the AddSub optimization, then Transform
1879 * Constant folding on Phi
1880 * Add(a,a) -> Mul(a, 2)
1881 * Add(Mul(a, x), a) -> Mul(a, x+1)
1882 * if the mode is integer or float.
1883 * Transform Add(a,-b) into Sub(a,b).
1884 * Reassociation might fold this further.
1886 static ir_node *transform_node_Add(ir_node *n) {
1888 ir_node *a, *b, *c, *oldn = n;
1890 n = transform_node_AddSub(n);
1892 a = get_Add_left(n);
1893 b = get_Add_right(n);
1895 HANDLE_BINOP_PHI(tarval_add, a,b,c);
1897 mode = get_irn_mode(n);
1899 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1900 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1903 if (mode_is_num(mode)) {
1905 ir_node *block = get_irn_n(n, -1);
1908 get_irn_dbg_info(n),
1912 new_r_Const_long(current_ir_graph, block, mode, 2),
1914 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1915 } else if (get_irn_op(a) == op_Minus) {
1917 get_irn_dbg_info(n),
1923 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1924 } else if (get_irn_op(b) == op_Minus) {
1926 get_irn_dbg_info(n),
1932 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1934 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1935 else if (!get_opt_reassociation() && get_irn_op(a) == op_Mul) {
1936 ir_node *ma = get_Mul_left(a);
1937 ir_node *mb = get_Mul_right(a);
1940 ir_node *blk = get_irn_n(n, -1);
1942 get_irn_dbg_info(n), current_ir_graph, blk,
1945 get_irn_dbg_info(n), current_ir_graph, blk,
1947 new_r_Const_long(current_ir_graph, blk, mode, 1),
1950 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1951 } else if (b == mb) {
1952 ir_node *blk = get_irn_n(n, -1);
1954 get_irn_dbg_info(n), current_ir_graph, blk,
1957 get_irn_dbg_info(n), current_ir_graph, blk,
1959 new_r_Const_long(current_ir_graph, blk, mode, 1),
1962 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1965 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1966 else if (!get_opt_reassociation() && get_irn_op(b) == op_Mul) {
1967 ir_node *ma = get_Mul_left(b);
1968 ir_node *mb = get_Mul_right(b);
1971 ir_node *blk = get_irn_n(n, -1);
1973 get_irn_dbg_info(n), current_ir_graph, blk,
1976 get_irn_dbg_info(n), current_ir_graph, blk,
1978 new_r_Const_long(current_ir_graph, blk, mode, 1),
1981 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1982 } else if (a == mb) {
1983 ir_node *blk = get_irn_n(n, -1);
1985 get_irn_dbg_info(n), current_ir_graph, blk,
1988 get_irn_dbg_info(n), current_ir_graph, blk,
1990 new_r_Const_long(current_ir_graph, blk, mode, 1),
1993 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1996 /* Here we rely on constants be on the RIGHT side */
1997 else if (is_Not(a) && classify_Const(b) == CNST_ONE) {
1999 ir_node *op = get_Not_op(a);
2000 ir_node *blk = get_irn_n(n, -1);
2001 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, mode);
2002 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_PLUS_1);
2006 } /* transform_node_Add */
2009 * Do the AddSub optimization, then Transform
2010 * Constant folding on Phi
2011 * Sub(0,a) -> Minus(a)
2012 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2013 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2015 static ir_node *transform_node_Sub(ir_node *n) {
2020 n = transform_node_AddSub(n);
2022 a = get_Sub_left(n);
2023 b = get_Sub_right(n);
2025 mode = get_irn_mode(n);
2028 HANDLE_BINOP_PHI(tarval_sub, a,b,c);
2030 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2031 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2035 if (mode_wrap_around(mode)) {
2036 ir_node *left = get_Add_left(a);
2037 ir_node *right = get_Add_right(a);
2039 /* FIXME: Does the Conv's word only for two complement or generally? */
2041 if (mode != get_irn_mode(right)) {
2042 /* This Sub is an effective Cast */
2043 right = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), right, mode);
2046 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2047 } else if (right == b) {
2048 if (mode != get_irn_mode(left)) {
2049 /* This Sub is an effective Cast */
2050 left = new_r_Conv(get_irn_irg(n), get_irn_n(n, -1), left, mode);
2053 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
2056 } else if (mode_is_int(mode) && is_Conv(a) && is_Conv(b)) {
2057 ir_mode *mode = get_irn_mode(a);
2059 if (mode == get_irn_mode(b)) {
2065 /* check if it's allowed to skip the conv */
2066 ma = get_irn_mode(a);
2067 mb = get_irn_mode(b);
2069 if (mode_is_reference(ma) && mode_is_reference(mb)) {
2070 /* SubInt(ConvInt(aP), ConvInt(bP)) -> SubInt(aP,bP) */
2072 set_Sub_right(n, b);
2078 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2079 else if (mode_is_num(mode) && mode == get_irn_mode(a) && (classify_Const(a) == CNST_NULL)) {
2081 get_irn_dbg_info(n),
2086 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2088 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2089 else if (get_opt_reassociation() && get_irn_op(a) == op_Mul) {
2090 ir_node *ma = get_Mul_left(a);
2091 ir_node *mb = get_Mul_right(a);
2094 ir_node *blk = get_irn_n(n, -1);
2096 get_irn_dbg_info(n),
2097 current_ir_graph, blk,
2100 get_irn_dbg_info(n),
2101 current_ir_graph, blk,
2103 new_r_Const_long(current_ir_graph, blk, mode, 1),
2106 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2107 } else if (mb == b) {
2108 ir_node *blk = get_irn_n(n, -1);
2110 get_irn_dbg_info(n),
2111 current_ir_graph, blk,
2114 get_irn_dbg_info(n),
2115 current_ir_graph, blk,
2117 new_r_Const_long(current_ir_graph, blk, mode, 1),
2120 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2122 } else if (get_irn_op(a) == op_Sub) {
2123 ir_node *x = get_Sub_left(a);
2124 ir_node *y = get_Sub_right(a);
2125 ir_node *blk = get_irn_n(n, -1);
2126 ir_mode *m_b = get_irn_mode(b);
2127 ir_mode *m_y = get_irn_mode(y);
2130 /* Determine the right mode for the Add. */
2133 else if (mode_is_reference(m_b))
2135 else if (mode_is_reference(m_y))
2139 * Both modes are different but none is reference,
2140 * happens for instance in SubP(SubP(P, Iu), Is).
2141 * We have two possibilities here: Cast or ignore.
2142 * Currently we ignore this case.
2147 add = new_r_Add(current_ir_graph, blk, y, b, mode);
2150 set_Sub_right(n, add);
2151 DBG_OPT_ALGSIM0(n, n, FS_OPT_SUB_SUB_X_Y_Z);
2154 } /* transform_node_Sub */
2157 * Transform Mul(a,-1) into -a.
2158 * Do constant evaluation of Phi nodes.
2159 * Do architecture dependent optimizations on Mul nodes
2161 static ir_node *transform_node_Mul(ir_node *n) {
2162 ir_node *c, *oldn = n;
2163 ir_node *a = get_Mul_left(n);
2164 ir_node *b = get_Mul_right(n);
2167 HANDLE_BINOP_PHI(tarval_mul, a,b,c);
2169 mode = get_irn_mode(n);
2170 if (mode_is_signed(mode)) {
2173 if (value_of(a) == get_mode_minus_one(mode))
2175 else if (value_of(b) == get_mode_minus_one(mode))
2178 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2179 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2183 return arch_dep_replace_mul_with_shifts(n);
2184 } /* transform_node_Mul */
2187 * Transform a Div Node.
2189 static ir_node *transform_node_Div(ir_node *n) {
2190 tarval *tv = value_of(n);
2193 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2195 if (tv != tarval_bad) {
2196 value = new_Const(get_tarval_mode(tv), tv);
2198 DBG_OPT_CSTEVAL(n, value);
2199 } else /* Try architecture dependent optimization */
2200 value = arch_dep_replace_div_by_const(n);
2203 /* Turn Div into a tuple (mem, jmp, bad, value) */
2204 ir_node *mem = get_Div_mem(n);
2205 ir_node *blk = get_irn_n(n, -1);
2207 turn_into_tuple(n, pn_Div_max);
2208 set_Tuple_pred(n, pn_Div_M, mem);
2209 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2210 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2211 set_Tuple_pred(n, pn_Div_res, value);
2214 } /* transform_node_Div */
2217 * Transform a Mod node.
2219 static ir_node *transform_node_Mod(ir_node *n) {
2220 tarval *tv = value_of(n);
2223 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
2225 if (tv != tarval_bad) {
2226 value = new_Const(get_tarval_mode(tv), tv);
2228 DBG_OPT_CSTEVAL(n, value);
2229 } else /* Try architecture dependent optimization */
2230 value = arch_dep_replace_mod_by_const(n);
2233 /* Turn Mod into a tuple (mem, jmp, bad, value) */
2234 ir_node *mem = get_Mod_mem(n);
2235 ir_node *blk = get_irn_n(n, -1);
2237 turn_into_tuple(n, pn_Mod_max);
2238 set_Tuple_pred(n, pn_Mod_M, mem);
2239 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2240 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2241 set_Tuple_pred(n, pn_Mod_res, value);
2244 } /* transform_node_Mod */
2247 * Transform a DivMod node.
2249 static ir_node *transform_node_DivMod(ir_node *n) {
2252 ir_node *a = get_DivMod_left(n);
2253 ir_node *b = get_DivMod_right(n);
2254 ir_mode *mode = get_irn_mode(a);
2255 tarval *ta = value_of(a);
2256 tarval *tb = value_of(b);
2258 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
2261 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2263 if (tb != tarval_bad) {
2264 if (tb == get_mode_one(get_tarval_mode(tb))) {
2265 b = new_Const (mode, get_mode_null(mode));
2268 DBG_OPT_CSTEVAL(n, b);
2269 } else if (ta != tarval_bad) {
2270 tarval *resa, *resb;
2271 resa = tarval_div (ta, tb);
2272 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2273 Jmp for X result!? */
2274 resb = tarval_mod (ta, tb);
2275 if (resb == tarval_bad) return n; /* Causes exception! */
2276 a = new_Const (mode, resa);
2277 b = new_Const (mode, resb);
2280 DBG_OPT_CSTEVAL(n, a);
2281 DBG_OPT_CSTEVAL(n, b);
2282 } else { /* Try architecture dependent optimization */
2283 arch_dep_replace_divmod_by_const(&a, &b, n);
2284 evaluated = a != NULL;
2286 } else if (ta == get_mode_null(mode)) {
2287 /* 0 / non-Const = 0 */
2292 if (evaluated) { /* replace by tuple */
2293 ir_node *mem = get_DivMod_mem(n);
2294 ir_node *blk = get_irn_n(n, -1);
2295 turn_into_tuple(n, pn_DivMod_max);
2296 set_Tuple_pred(n, pn_DivMod_M, mem);
2297 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2298 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2299 set_Tuple_pred(n, pn_DivMod_res_div, a);
2300 set_Tuple_pred(n, pn_DivMod_res_mod, b);
2304 } /* transform_node_DivMod */
2307 * Optimize Abs(x) into x if x is Confirmed >= 0
2308 * Optimize Abs(x) into -x if x is Confirmed <= 0
2310 static ir_node *transform_node_Abs(ir_node *n) {
2312 ir_node *a = get_Abs_op(n);
2313 value_classify_sign sign = classify_value_sign(a);
2315 if (sign == value_classified_negative) {
2316 ir_mode *mode = get_irn_mode(n);
2319 * We can replace the Abs by -x here.
2320 * We even could add a new Confirm here.
2322 * Note that -x would create a new node, so we could
2323 * not run it in the equivalent_node() context.
2325 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2326 get_irn_n(n, -1), a, mode);
2328 DBG_OPT_CONFIRM(oldn, n);
2329 } else if (sign == value_classified_positive) {
2330 /* n is positive, Abs is not needed */
2333 DBG_OPT_CONFIRM(oldn, n);
2337 } /* transform_node_Abs */
2340 * Transform a Cond node.
2342 * Replace the Cond by a Jmp if it branches on a constant
2345 static ir_node *transform_node_Cond(ir_node *n) {
2348 ir_node *a = get_Cond_selector(n);
2349 tarval *ta = value_of(a);
2351 /* we need block info which is not available in floating irgs */
2352 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2355 if ((ta != tarval_bad) &&
2356 (get_irn_mode(a) == mode_b) &&
2357 (get_opt_unreachable_code())) {
2358 /* It's a boolean Cond, branching on a boolean constant.
2359 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2360 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
2361 turn_into_tuple(n, pn_Cond_max);
2362 if (ta == tarval_b_true) {
2363 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2364 set_Tuple_pred(n, pn_Cond_true, jmp);
2366 set_Tuple_pred(n, pn_Cond_false, jmp);
2367 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2369 /* We might generate an endless loop, so keep it alive. */
2370 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
2373 } /* transform_node_Cond */
2375 typedef ir_node* (*recursive_transform) (ir_node *n);
2378 * makes use of distributive laws for and, or, eor
2379 * and(a OP c, b OP c) -> and(a, b) OP c
2381 static ir_node *transform_bitwise_distributive(ir_node *n,
2382 recursive_transform trans_func)
2385 ir_node *a = get_binop_left(n);
2386 ir_node *b = get_binop_right(n);
2387 ir_op *op = get_irn_op(a);
2388 ir_op *op_root = get_irn_op(n);
2390 if(op != get_irn_op(b))
2393 if (op == op_Conv) {
2394 ir_node *a_op = get_Conv_op(a);
2395 ir_node *b_op = get_Conv_op(b);
2396 ir_mode *a_mode = get_irn_mode(a_op);
2397 ir_mode *b_mode = get_irn_mode(b_op);
2398 if(a_mode == b_mode && (mode_is_int(a_mode) || a_mode == mode_b)) {
2399 ir_node *blk = get_irn_n(n, -1);
2402 set_binop_left(n, a_op);
2403 set_binop_right(n, b_op);
2404 set_irn_mode(n, a_mode);
2406 n = new_r_Conv(current_ir_graph, blk, n, get_irn_mode(oldn));
2408 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2414 /* nothing to gain here */
2418 if (op == op_Shrs || op == op_Shr || op == op_Shl
2419 || op == op_And || op == op_Or || op == op_Eor) {
2420 ir_node *a_left = get_binop_left(a);
2421 ir_node *a_right = get_binop_right(a);
2422 ir_node *b_left = get_binop_left(b);
2423 ir_node *b_right = get_binop_right(b);
2427 if (is_op_commutative(op)) {
2428 if (a_left == b_left) {
2432 } else if(a_left == b_right) {
2436 } else if(a_right == b_left) {
2442 if(a_right == b_right) {
2449 /* (a sop c) & (b sop c) => (a & b) sop c */
2450 ir_node *blk = get_irn_n(n, -1);
2452 ir_node *new_n = exact_copy(n);
2453 set_binop_left(new_n, op1);
2454 set_binop_right(new_n, op2);
2455 new_n = trans_func(new_n);
2457 if(op_root == op_Eor && op == op_Or) {
2458 dbg_info *dbgi = get_irn_dbg_info(n);
2459 ir_graph *irg = current_ir_graph;
2460 ir_mode *mode = get_irn_mode(c);
2462 c = new_rd_Not(dbgi, irg, blk, c, mode);
2463 n = new_rd_And(dbgi, irg, blk, new_n, c, mode);
2466 set_irn_n(n, -1, blk);
2467 set_binop_left(n, new_n);
2468 set_binop_right(n, c);
2472 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_SHIFT_AND);
2483 static ir_node *transform_node_And(ir_node *n) {
2485 ir_node *a = get_And_left(n);
2486 ir_node *b = get_And_right(n);
2488 HANDLE_BINOP_PHI(tarval_and, a,b,c);
2490 n = transform_bitwise_distributive(n, transform_node_And);
2493 } /* transform_node_And */
2498 static ir_node *transform_node_Eor(ir_node *n) {
2499 ir_node *c, *oldn = n;
2500 ir_node *a = get_Eor_left(n);
2501 ir_node *b = get_Eor_right(n);
2502 ir_mode *mode = get_irn_mode(n);
2504 HANDLE_BINOP_PHI(tarval_eor, a,b,c);
2508 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
2509 mode, get_mode_null(mode));
2510 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
2511 } else if ((mode == mode_b)
2512 && (get_irn_op(a) == op_Proj)
2513 && (get_irn_mode(a) == mode_b)
2514 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
2515 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2516 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
2517 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2518 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
2520 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
2521 } else if ((mode == mode_b)
2522 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
2523 /* The Eor is a Not. Replace it by a Not. */
2524 /* ????!!!Extend to bitfield 1111111. */
2525 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
2527 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2529 n = transform_bitwise_distributive(n, transform_node_Eor);
2533 } /* transform_node_Eor */
2538 static ir_node *transform_node_Not(ir_node *n) {
2539 ir_node *c, *oldn = n;
2540 ir_node *a = get_Not_op(n);
2541 ir_op *op_a = get_irn_op(a);
2543 HANDLE_UNOP_PHI(tarval_not,a,c);
2545 /* check for a boolean Not */
2546 if ( (get_irn_mode(n) == mode_b)
2547 && (op_a == op_Proj)
2548 && (get_irn_mode(a) == mode_b)
2549 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2550 /* We negate a Cmp. The Cmp has the negated result anyways! */
2551 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2552 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
2553 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
2556 if (op_a == op_Sub && classify_Const(get_Sub_right(a)) == CNST_ONE) {
2558 ir_node *op = get_Sub_left(a);
2559 ir_node *blk = get_irn_n(n, -1);
2560 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, blk, op, get_irn_mode(n));
2561 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_MINUS_1);
2564 } /* transform_node_Not */
2567 * Transform a Minus.
2571 static ir_node *transform_node_Minus(ir_node *n) {
2572 ir_node *c, *oldn = n;
2573 ir_node *a = get_Minus_op(n);
2575 HANDLE_UNOP_PHI(tarval_neg,a,c);
2579 ir_node *op = get_Not_op(a);
2580 ir_mode *mode = get_irn_mode(op);
2581 tarval *tv = get_mode_one(mode);
2582 ir_node *blk = get_irn_n(n, -1);
2583 ir_node *c = new_r_Const(current_ir_graph, blk, mode, tv);
2584 n = new_rd_Add(get_irn_dbg_info(n), current_ir_graph, blk, op, c, mode);
2585 DBG_OPT_ALGSIM2(oldn, a, n, FS_OPT_MINUS_NOT);
2589 } /* transform_node_Minus */
2592 * Transform a Cast_type(Const) into a new Const_type
2594 static ir_node *transform_node_Cast(ir_node *n) {
2596 ir_node *pred = get_Cast_op(n);
2597 ir_type *tp = get_irn_type(n);
2599 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
2600 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
2601 get_Const_tarval(pred), tp);
2602 DBG_OPT_CSTEVAL(oldn, n);
2603 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
2604 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
2605 get_SymConst_kind(pred), tp);
2606 DBG_OPT_CSTEVAL(oldn, n);
2610 } /* transform_node_Cast */
2613 * Transform a Proj(Div) with a non-zero value.
2614 * Removes the exceptions and routes the memory to the NoMem node.
2616 static ir_node *transform_node_Proj_Div(ir_node *proj) {
2617 ir_node *div = get_Proj_pred(proj);
2618 ir_node *b = get_Div_right(div);
2619 ir_node *confirm, *res, *new_mem;
2622 if (value_not_zero(b, &confirm)) {
2623 /* div(x, y) && y != 0 */
2624 proj_nr = get_Proj_proj(proj);
2626 case pn_Div_X_regular:
2627 return new_r_Jmp(current_ir_graph, get_irn_n(div, -1));
2629 case pn_Div_X_except:
2630 /* we found an exception handler, remove it */
2631 DBG_OPT_EXC_REM(proj);
2635 res = get_Div_mem(div);
2636 new_mem = get_irg_no_mem(current_ir_graph);
2639 /* This node can only float up to the Confirm block */
2640 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2642 set_irn_pinned(div, op_pin_state_floats);
2643 /* this is a Div without exception, we can remove the memory edge */
2644 set_Div_mem(div, new_mem);
2649 } /* transform_node_Proj_Div */
2652 * Transform a Proj(Mod) with a non-zero value.
2653 * Removes the exceptions and routes the memory to the NoMem node.
2655 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
2656 ir_node *mod = get_Proj_pred(proj);
2657 ir_node *b = get_Mod_right(mod);
2658 ir_node *confirm, *res, *new_mem;
2661 if (value_not_zero(b, &confirm)) {
2662 /* mod(x, y) && y != 0 */
2663 proj_nr = get_Proj_proj(proj);
2667 case pn_Mod_X_regular:
2668 return new_r_Jmp(current_ir_graph, get_irn_n(mod, -1));
2670 case pn_Mod_X_except:
2671 /* we found an exception handler, remove it */
2672 DBG_OPT_EXC_REM(proj);
2676 res = get_Mod_mem(mod);
2677 new_mem = get_irg_no_mem(current_ir_graph);
2680 /* This node can only float up to the Confirm block */
2681 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2683 set_irn_pinned(mod, op_pin_state_floats);
2684 /* this is a Mod without exception, we can remove the memory edge */
2685 set_Mod_mem(mod, get_irg_no_mem(current_ir_graph));
2688 if (get_Mod_left(mod) == b) {
2689 /* a % a = 0 if a != 0 */
2690 ir_mode *mode = get_irn_mode(proj);
2691 ir_node *res = new_Const(mode, get_mode_null(mode));
2693 DBG_OPT_CSTEVAL(mod, res);
2699 } /* transform_node_Proj_Mod */
2702 * Transform a Proj(DivMod) with a non-zero value.
2703 * Removes the exceptions and routes the memory to the NoMem node.
2705 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
2706 ir_node *divmod = get_Proj_pred(proj);
2707 ir_node *b = get_DivMod_right(divmod);
2708 ir_node *confirm, *res, *new_mem;
2711 if (value_not_zero(b, &confirm)) {
2712 /* DivMod(x, y) && y != 0 */
2713 proj_nr = get_Proj_proj(proj);
2717 case pn_DivMod_X_regular:
2718 return new_r_Jmp(current_ir_graph, get_irn_n(divmod, -1));
2720 case pn_DivMod_X_except:
2721 /* we found an exception handler, remove it */
2722 DBG_OPT_EXC_REM(proj);
2726 res = get_DivMod_mem(divmod);
2727 new_mem = get_irg_no_mem(current_ir_graph);
2730 /* This node can only float up to the Confirm block */
2731 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2733 set_irn_pinned(divmod, op_pin_state_floats);
2734 /* this is a DivMod without exception, we can remove the memory edge */
2735 set_DivMod_mem(divmod, get_irg_no_mem(current_ir_graph));
2738 case pn_DivMod_res_mod:
2739 if (get_DivMod_left(divmod) == b) {
2740 /* a % a = 0 if a != 0 */
2741 ir_mode *mode = get_irn_mode(proj);
2742 ir_node *res = new_Const(mode, get_mode_null(mode));
2744 DBG_OPT_CSTEVAL(divmod, res);
2750 } /* transform_node_Proj_DivMod */
2753 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2755 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
2756 if (get_opt_unreachable_code()) {
2757 ir_node *n = get_Proj_pred(proj);
2758 ir_node *b = get_Cond_selector(n);
2760 if (mode_is_int(get_irn_mode(b))) {
2761 tarval *tb = value_of(b);
2763 if (tb != tarval_bad) {
2764 /* we have a constant switch */
2765 long num = get_Proj_proj(proj);
2767 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2768 if (get_tarval_long(tb) == num) {
2769 /* Do NOT create a jump here, or we will have 2 control flow ops
2770 * in a block. This case is optimized away in optimize_cf(). */
2773 /* this case will NEVER be taken, kill it */
2781 } /* transform_node_Proj_Cond */
2784 * Normalizes and optimizes Cmp nodes.
2786 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
2787 if (get_opt_reassociation()) {
2788 ir_node *n = get_Proj_pred(proj);
2789 ir_node *left = get_Cmp_left(n);
2790 ir_node *right = get_Cmp_right(n);
2794 ir_mode *mode = NULL;
2795 long proj_nr = get_Proj_proj(proj);
2798 * First step: normalize the compare op
2799 * by placing the constant on the right site
2800 * or moving the lower address node to the left.
2801 * We ignore the case that both are constants
2802 * this case should be optimized away.
2804 if (get_irn_op(right) == op_Const) {
2806 } else if (get_irn_op(left) == op_Const) {
2811 proj_nr = get_inversed_pnc(proj_nr);
2813 } else if (get_irn_idx(left) > get_irn_idx(right)) {
2819 proj_nr = get_inversed_pnc(proj_nr);
2824 * Second step: Try to reduce the magnitude
2825 * of a constant. This may help to generate better code
2826 * later and may help to normalize more compares.
2827 * Of course this is only possible for integer values.
2830 mode = get_irn_mode(c);
2831 tv = get_Const_tarval(c);
2833 if (tv != tarval_bad) {
2834 /* the following optimization is possible on modes without Overflow
2835 * on Unary Minus or on == and !=:
2836 * -a CMP c ==> a swap(CMP) -c
2838 * Beware: for two-complement Overflow may occur, so only == and != can
2839 * be optimized, see this:
2840 * -MININT < 0 =/=> MININT > 0 !!!
2842 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2843 (!mode_overflow_on_unary_Minus(mode) ||
2844 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2845 left = get_Minus_op(left);
2846 tv = tarval_neg(tv);
2848 if (tv != tarval_bad) {
2849 proj_nr = get_inversed_pnc(proj_nr);
2854 /* for integer modes, we have more */
2855 if (mode_is_int(mode)) {
2856 /* Ne includes Unordered which is not possible on integers.
2857 * However, frontends often use this wrong, so fix it here */
2858 if (proj_nr & pn_Cmp_Uo) {
2859 proj_nr &= ~pn_Cmp_Uo;
2860 set_Proj_proj(proj, proj_nr);
2863 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2864 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2865 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2866 tv = tarval_sub(tv, get_mode_one(mode));
2868 if (tv != tarval_bad) {
2869 proj_nr ^= pn_Cmp_Eq;
2873 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2874 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2875 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2876 tv = tarval_add(tv, get_mode_one(mode));
2878 if (tv != tarval_bad) {
2879 proj_nr ^= pn_Cmp_Eq;
2884 /* the following reassociations work only for == and != */
2885 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2887 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2888 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2889 right = get_Sub_right(left);
2890 left = get_Sub_left(left);
2892 tv = value_of(right);
2893 if (tv != tarval_bad) {
2898 if (tv != tarval_bad) {
2899 ir_op *op = get_irn_op(left);
2901 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2903 ir_node *c1 = get_Sub_right(left);
2904 tarval *tv2 = value_of(c1);
2906 if (tv2 != tarval_bad) {
2907 tv2 = tarval_add(tv, value_of(c1));
2909 if (tv2 != tarval_bad) {
2910 left = get_Sub_left(left);
2916 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2917 else if (op == op_Add) {
2918 ir_node *a_l = get_Add_left(left);
2919 ir_node *a_r = get_Add_right(left);
2923 if (get_irn_op(a_l) == op_Const) {
2925 tv2 = value_of(a_l);
2928 tv2 = value_of(a_r);
2931 if (tv2 != tarval_bad) {
2932 tv2 = tarval_sub(tv, tv2);
2934 if (tv2 != tarval_bad) {
2941 /* -a == c ==> a == -c, -a != c ==> a != -c */
2942 else if (op == op_Minus) {
2943 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2945 if (tv2 != tarval_bad) {
2946 left = get_Minus_op(left);
2953 /* the following reassociations work only for <= */
2954 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2955 if (tv != tarval_bad) {
2956 ir_op *op = get_irn_op(left);
2958 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2966 * optimization for AND:
2968 * And(x, C) == C ==> And(x, C) != 0
2969 * And(x, C) != C ==> And(X, C) == 0
2971 * if C is a single Bit constant.
2973 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) &&
2974 (get_irn_op(left) == op_And)) {
2975 if (is_single_bit_tarval(tv)) {
2976 /* check for Constant's match. We have check hare the tarvals,
2977 because our const might be changed */
2978 ir_node *la = get_And_left(left);
2979 ir_node *ra = get_And_right(left);
2980 if ((is_Const(la) && get_Const_tarval(la) == tv) ||
2981 (is_Const(ra) && get_Const_tarval(ra) == tv)) {
2982 /* fine: do the transformation */
2983 tv = get_mode_null(get_tarval_mode(tv));
2984 proj_nr ^= pn_Cmp_Leg;
2989 } /* tarval != bad */
2993 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
2995 if (changed & 2) /* need a new Const */
2996 right = new_Const(mode, tv);
2998 /* create a new compare */
2999 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
3002 set_Proj_pred(proj, n);
3003 set_Proj_proj(proj, proj_nr);
3007 } /* transform_node_Proj_Cmp */
3010 * Does all optimizations on nodes that must be done on it's Proj's
3011 * because of creating new nodes.
3013 static ir_node *transform_node_Proj(ir_node *proj) {
3014 ir_node *n = get_Proj_pred(proj);
3016 switch (get_irn_opcode(n)) {
3018 return transform_node_Proj_Div(proj);
3021 return transform_node_Proj_Mod(proj);
3024 return transform_node_Proj_DivMod(proj);
3027 return transform_node_Proj_Cond(proj);
3030 return transform_node_Proj_Cmp(proj);
3033 /* should not happen, but if it does will be optimized away */
3034 return equivalent_node_Proj(proj);
3040 } /* transform_node_Proj */
3043 * Move Confirms down through Phi nodes.
3045 static ir_node *transform_node_Phi(ir_node *phi) {
3047 ir_mode *mode = get_irn_mode(phi);
3049 if (mode_is_reference(mode)) {
3050 n = get_irn_arity(phi);
3052 /* Beware of Phi0 */
3054 ir_node *pred = get_irn_n(phi, 0);
3055 ir_node *bound, *new_Phi, *block, **in;
3058 if (! is_Confirm(pred))
3061 bound = get_Confirm_bound(pred);
3062 pnc = get_Confirm_cmp(pred);
3064 NEW_ARR_A(ir_node *, in, n);
3065 in[0] = get_Confirm_value(pred);
3067 for (i = 1; i < n; ++i) {
3068 pred = get_irn_n(phi, i);
3070 if (! is_Confirm(pred) ||
3071 get_Confirm_bound(pred) != bound ||
3072 get_Confirm_cmp(pred) != pnc)
3074 in[i] = get_Confirm_value(pred);
3076 /* move the Confirm nodes "behind" the Phi */
3077 block = get_irn_n(phi, -1);
3078 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
3079 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
3083 } /* transform_node_Phi */
3086 * Returns the operands of a commutative bin-op, if one operand is
3087 * a const, it is returned as the second one.
3089 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
3090 ir_node *op_a = get_binop_left(binop);
3091 ir_node *op_b = get_binop_right(binop);
3093 assert(is_op_commutative(get_irn_op(binop)));
3095 if (get_irn_op(op_a) == op_Const) {
3102 } /* get_comm_Binop_Ops */
3105 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
3106 * Such pattern may arise in bitfield stores.
3108 * value c4 value c4 & c2
3109 * AND c3 AND c1 | c3
3114 static ir_node *transform_node_Or_bf_store(ir_node *or) {
3117 ir_node *and_l, *c3;
3118 ir_node *value, *c4;
3119 ir_node *new_and, *new_const, *block;
3120 ir_mode *mode = get_irn_mode(or);
3122 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
3124 get_comm_Binop_Ops(or, &and, &c1);
3125 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
3128 get_comm_Binop_Ops(and, &or_l, &c2);
3129 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
3132 get_comm_Binop_Ops(or_l, &and_l, &c3);
3133 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
3136 get_comm_Binop_Ops(and_l, &value, &c4);
3137 if (get_irn_op(c4) != op_Const)
3140 /* ok, found the pattern, check for conditions */
3141 assert(mode == get_irn_mode(and));
3142 assert(mode == get_irn_mode(or_l));
3143 assert(mode == get_irn_mode(and_l));
3145 tv1 = get_Const_tarval(c1);
3146 tv2 = get_Const_tarval(c2);
3147 tv3 = get_Const_tarval(c3);
3148 tv4 = get_Const_tarval(c4);
3150 tv = tarval_or(tv4, tv2);
3151 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
3152 /* have at least one 0 at the same bit position */
3156 n_tv4 = tarval_not(tv4);
3157 if (tv3 != tarval_and(tv3, n_tv4)) {
3158 /* bit in the or_mask is outside the and_mask */
3162 n_tv2 = tarval_not(tv2);
3163 if (tv1 != tarval_and(tv1, n_tv2)) {
3164 /* bit in the or_mask is outside the and_mask */
3168 /* ok, all conditions met */
3169 block = get_irn_n(or, -1);
3171 new_and = new_r_And(current_ir_graph, block,
3172 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
3174 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
3176 set_Or_left(or, new_and);
3177 set_Or_right(or, new_const);
3179 /* check for more */
3180 return transform_node_Or_bf_store(or);
3181 } /* transform_node_Or_bf_store */
3184 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
3186 static ir_node *transform_node_Or_Rot(ir_node *or) {
3187 ir_mode *mode = get_irn_mode(or);
3188 ir_node *shl, *shr, *block;
3189 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
3192 if (! mode_is_int(mode))
3195 shl = get_binop_left(or);
3196 shr = get_binop_right(or);
3198 if (get_irn_op(shl) == op_Shr) {
3199 if (get_irn_op(shr) != op_Shl)
3205 } else if (get_irn_op(shl) != op_Shl) {
3207 } else if (get_irn_op(shr) != op_Shr) {
3210 x = get_Shl_left(shl);
3211 if (x != get_Shr_left(shr))
3214 c1 = get_Shl_right(shl);
3215 c2 = get_Shr_right(shr);
3216 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
3217 tv1 = get_Const_tarval(c1);
3218 if (! tarval_is_long(tv1))
3221 tv2 = get_Const_tarval(c2);
3222 if (! tarval_is_long(tv2))
3225 if (get_tarval_long(tv1) + get_tarval_long(tv2)
3226 != get_mode_size_bits(mode))
3229 /* yet, condition met */
3230 block = get_irn_n(or, -1);
3232 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
3234 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
3236 } else if (get_irn_op(c1) == op_Sub) {
3240 if (get_Sub_right(sub) != v)
3243 c1 = get_Sub_left(sub);
3244 if (get_irn_op(c1) != op_Const)
3247 tv1 = get_Const_tarval(c1);
3248 if (! tarval_is_long(tv1))
3251 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3254 /* yet, condition met */
3255 block = get_nodes_block(or);
3257 /* a Rot right is not supported, so use a rot left */
3258 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
3260 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3262 } else if (get_irn_op(c2) == op_Sub) {
3266 c1 = get_Sub_left(sub);
3267 if (get_irn_op(c1) != op_Const)
3270 tv1 = get_Const_tarval(c1);
3271 if (! tarval_is_long(tv1))
3274 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3277 /* yet, condition met */
3278 block = get_irn_n(or, -1);
3281 n = new_r_Rot(current_ir_graph, block, x, v, mode);
3283 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3288 } /* transform_node_Or_Rot */
3293 static ir_node *transform_node_Or(ir_node *n) {
3294 ir_node *c, *oldn = n;
3295 ir_node *a = get_Or_left(n);
3296 ir_node *b = get_Or_right(n);
3298 HANDLE_BINOP_PHI(tarval_or, a,b,c);
3300 n = transform_node_Or_bf_store(n);
3301 n = transform_node_Or_Rot(n);
3305 n = transform_bitwise_distributive(n, transform_node_Or);
3308 } /* transform_node_Or */
3312 static ir_node *transform_node(ir_node *n);
3315 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
3317 * Should be moved to reassociation?
3319 static ir_node *transform_node_shift(ir_node *n) {
3320 ir_node *left, *right;
3321 tarval *tv1, *tv2, *res;
3323 int modulo_shf, flag;
3325 left = get_binop_left(n);
3327 /* different operations */
3328 if (get_irn_op(left) != get_irn_op(n))
3331 right = get_binop_right(n);
3332 tv1 = value_of(right);
3333 if (tv1 == tarval_bad)
3336 tv2 = value_of(get_binop_right(left));
3337 if (tv2 == tarval_bad)
3340 res = tarval_add(tv1, tv2);
3342 /* beware: a simple replacement works only, if res < modulo shift */
3343 mode = get_irn_mode(n);
3347 modulo_shf = get_mode_modulo_shift(mode);
3348 if (modulo_shf > 0) {
3349 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
3351 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
3357 /* ok, we can replace it */
3358 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
3360 in[0] = get_binop_left(left);
3361 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
3363 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
3365 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
3367 return transform_node(irn);
3370 } /* transform_node_shift */
3375 static ir_node *transform_node_Shr(ir_node *n) {
3376 ir_node *c, *oldn = n;
3377 ir_node *a = get_Shr_left(n);
3378 ir_node *b = get_Shr_right(n);
3380 HANDLE_BINOP_PHI(tarval_shr, a, b, c);
3381 return transform_node_shift(n);
3382 } /* transform_node_Shr */
3387 static ir_node *transform_node_Shrs(ir_node *n) {
3388 ir_node *c, *oldn = n;
3389 ir_node *a = get_Shrs_left(n);
3390 ir_node *b = get_Shrs_right(n);
3392 HANDLE_BINOP_PHI(tarval_shrs, a, b, c);
3393 return transform_node_shift(n);
3394 } /* transform_node_Shrs */
3399 static ir_node *transform_node_Shl(ir_node *n) {
3400 ir_node *c, *oldn = n;
3401 ir_node *a = get_Shl_left(n);
3402 ir_node *b = get_Shl_right(n);
3404 HANDLE_BINOP_PHI(tarval_shl, a, b, c);
3405 return transform_node_shift(n);
3406 } /* transform_node_Shl */
3409 * Remove dead blocks and nodes in dead blocks
3410 * in keep alive list. We do not generate a new End node.
3412 static ir_node *transform_node_End(ir_node *n) {
3413 int i, j, n_keepalives = get_End_n_keepalives(n);
3416 NEW_ARR_A(ir_node *, in, n_keepalives);
3418 for (i = j = 0; i < n_keepalives; ++i) {
3419 ir_node *ka = get_End_keepalive(n, i);
3421 if (! is_Block_dead(ka)) {
3425 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
3428 /* FIXME: beabi need to keep a Proj(M) */
3429 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
3432 if (j != n_keepalives)
3433 set_End_keepalives(n, j, in);
3435 } /* transform_node_End */
3438 * Optimize a Mux into some simpler cases.
3440 static ir_node *transform_node_Mux(ir_node *n) {
3441 ir_node *oldn = n, *sel = get_Mux_sel(n);
3442 ir_mode *mode = get_irn_mode(n);
3444 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
3445 ir_node *cmp = get_Proj_pred(sel);
3446 long proj_nr = get_Proj_proj(sel);
3447 ir_node *f = get_Mux_false(n);
3448 ir_node *t = get_Mux_true(n);
3450 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
3451 ir_node *block = get_irn_n(n, -1);
3454 * Note: normalization puts the constant on the right site,
3455 * so we check only one case.
3457 * Note further that these optimization work even for floating point
3458 * with NaN's because -NaN == NaN.
3459 * However, if +0 and -0 is handled differently, we cannot use the first one.
3461 if (get_irn_op(f) == op_Minus &&
3462 get_Minus_op(f) == t &&
3463 get_Cmp_left(cmp) == t) {
3465 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3466 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
3467 n = new_rd_Abs(get_irn_dbg_info(n),
3471 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3473 } else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3474 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
3475 n = new_rd_Abs(get_irn_dbg_info(n),
3479 n = new_rd_Minus(get_irn_dbg_info(n),
3484 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3487 } else if (get_irn_op(t) == op_Minus &&
3488 get_Minus_op(t) == f &&
3489 get_Cmp_left(cmp) == f) {
3491 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3492 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
3493 n = new_rd_Abs(get_irn_dbg_info(n),
3497 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3499 } else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3500 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
3501 n = new_rd_Abs(get_irn_dbg_info(n),
3505 n = new_rd_Minus(get_irn_dbg_info(n),
3510 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3515 if (mode_is_int(mode) && mode_is_signed(mode) &&
3516 get_mode_arithmetic(mode) == irma_twos_complement) {
3517 ir_node *x = get_Cmp_left(cmp);
3519 /* the following optimization works only with signed integer two-complement mode */
3521 if (mode == get_irn_mode(x)) {
3523 * FIXME: this restriction is two rigid, as it would still
3524 * work if mode(x) = Hs and mode == Is, but at least it removes
3527 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
3528 classify_Const(t) == CNST_ALL_ONE &&
3529 classify_Const(f) == CNST_NULL) {
3531 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
3535 n = new_rd_Shrs(get_irn_dbg_info(n),
3536 current_ir_graph, block, x,
3537 new_r_Const_long(current_ir_graph, block, mode_Iu,
3538 get_mode_size_bits(mode) - 1),
3540 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3542 } else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
3543 classify_Const(t) == CNST_ONE &&
3544 classify_Const(f) == CNST_NULL) {
3546 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
3550 n = new_rd_Shr(get_irn_dbg_info(n),
3551 current_ir_graph, block,
3552 new_r_Minus(current_ir_graph, block, x, mode),
3553 new_r_Const_long(current_ir_graph, block, mode_Iu,
3554 get_mode_size_bits(mode) - 1),
3556 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3563 return arch_transform_node_Mux(n);
3564 } /* transform_node_Mux */
3567 * Optimize a Psi into some simpler cases.
3569 static ir_node *transform_node_Psi(ir_node *n) {
3571 return transform_node_Mux(n);
3574 } /* transform_node_Psi */
3577 * Tries several [inplace] [optimizing] transformations and returns an
3578 * equivalent node. The difference to equivalent_node() is that these
3579 * transformations _do_ generate new nodes, and thus the old node must
3580 * not be freed even if the equivalent node isn't the old one.
3582 static ir_node *transform_node(ir_node *n) {
3586 * Transform_node is the only "optimizing transformation" that might
3587 * return a node with a different opcode. We iterate HERE until fixpoint
3588 * to get the final result.
3592 if (n->op->ops.transform_node)
3593 n = n->op->ops.transform_node(n);
3594 } while (oldn != n);
3597 } /* transform_node */
3600 * Sets the default transform node operation for an ir_op_ops.
3602 * @param code the opcode for the default operation
3603 * @param ops the operations initialized
3608 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
3612 ops->transform_node = transform_node_##a; \
3645 } /* firm_set_default_transform_node */
3648 /* **************** Common Subexpression Elimination **************** */
3650 /** The size of the hash table used, should estimate the number of nodes
3652 #define N_IR_NODES 512
3654 /** Compares the attributes of two Const nodes. */
3655 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
3656 return (get_Const_tarval(a) != get_Const_tarval(b))
3657 || (get_Const_type(a) != get_Const_type(b));
3658 } /* node_cmp_attr_Const */
3660 /** Compares the attributes of two Proj nodes. */
3661 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
3662 return get_irn_proj_attr(a) != get_irn_proj_attr(b);
3663 } /* node_cmp_attr_Proj */
3665 /** Compares the attributes of two Filter nodes. */
3666 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
3667 return get_Filter_proj(a) != get_Filter_proj(b);
3668 } /* node_cmp_attr_Filter */
3670 /** Compares the attributes of two Alloc nodes. */
3671 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
3672 const alloc_attr *pa = get_irn_alloc_attr(a);
3673 const alloc_attr *pb = get_irn_alloc_attr(b);
3674 return (pa->where != pb->where) || (pa->type != pb->type);
3675 } /* node_cmp_attr_Alloc */
3677 /** Compares the attributes of two Free nodes. */
3678 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
3679 const free_attr *pa = get_irn_free_attr(a);
3680 const free_attr *pb = get_irn_free_attr(b);
3681 return (pa->where != pb->where) || (pa->type != pb->type);
3682 } /* node_cmp_attr_Free */
3684 /** Compares the attributes of two SymConst nodes. */
3685 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
3686 const symconst_attr *pa = get_irn_symconst_attr(a);
3687 const symconst_attr *pb = get_irn_symconst_attr(b);
3688 return (pa->num != pb->num)
3689 || (pa->sym.type_p != pb->sym.type_p)
3690 || (pa->tp != pb->tp);
3691 } /* node_cmp_attr_SymConst */
3693 /** Compares the attributes of two Call nodes. */
3694 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
3695 return (get_irn_call_attr(a) != get_irn_call_attr(b));
3696 } /* node_cmp_attr_Call */
3698 /** Compares the attributes of two Sel nodes. */
3699 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
3700 const ir_entity *a_ent = get_Sel_entity(a);
3701 const ir_entity *b_ent = get_Sel_entity(b);
3703 (a_ent->kind != b_ent->kind) ||
3704 (a_ent->name != b_ent->name) ||
3705 (a_ent->owner != b_ent->owner) ||
3706 (a_ent->ld_name != b_ent->ld_name) ||
3707 (a_ent->type != b_ent->type);
3708 } /* node_cmp_attr_Sel */
3710 /** Compares the attributes of two Phi nodes. */
3711 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
3712 /* we can only enter this function if both nodes have the same number of inputs,
3713 hence it is enough to check if one of them is a Phi0 */
3715 /* check the Phi0 attribute */
3716 return get_irn_phi0_attr(a) != get_irn_phi0_attr(b);
3719 } /* node_cmp_attr_Phi */
3721 /** Compares the attributes of two Conv nodes. */
3722 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
3723 return get_Conv_strict(a) != get_Conv_strict(b);
3724 } /* node_cmp_attr_Conv */
3726 /** Compares the attributes of two Cast nodes. */
3727 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
3728 return get_Cast_type(a) != get_Cast_type(b);
3729 } /* node_cmp_attr_Cast */
3731 /** Compares the attributes of two Load nodes. */
3732 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
3733 if (get_Load_volatility(a) == volatility_is_volatile ||
3734 get_Load_volatility(b) == volatility_is_volatile)
3735 /* NEVER do CSE on volatile Loads */
3738 return get_Load_mode(a) != get_Load_mode(b);
3739 } /* node_cmp_attr_Load */
3741 /** Compares the attributes of two Store nodes. */
3742 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
3743 /* NEVER do CSE on volatile Stores */
3744 return (get_Store_volatility(a) == volatility_is_volatile ||
3745 get_Store_volatility(b) == volatility_is_volatile);
3746 } /* node_cmp_attr_Store */
3748 /** Compares the attributes of two Confirm nodes. */
3749 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
3750 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
3751 } /* node_cmp_attr_Confirm */
3753 /** Compares the attributes of two ASM nodes. */
3754 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
3756 const ir_asm_constraint *ca;
3757 const ir_asm_constraint *cb;
3760 if (get_ASM_text(a) != get_ASM_text(b))
3763 /* Should we really check the constraints here? Should be better, but is strange. */
3764 n = get_ASM_n_input_constraints(a);
3765 if (n != get_ASM_n_input_constraints(b))
3768 ca = get_ASM_input_constraints(a);
3769 cb = get_ASM_input_constraints(b);
3770 for (i = 0; i < n; ++i) {
3771 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
3775 n = get_ASM_n_output_constraints(a);
3776 if (n != get_ASM_n_output_constraints(b))
3779 ca = get_ASM_output_constraints(a);
3780 cb = get_ASM_output_constraints(b);
3781 for (i = 0; i < n; ++i) {
3782 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
3786 n = get_ASM_n_clobbers(a);
3787 if (n != get_ASM_n_clobbers(b))
3790 cla = get_ASM_clobbers(a);
3791 clb = get_ASM_clobbers(b);
3792 for (i = 0; i < n; ++i) {
3793 if (cla[i] != clb[i])
3797 } /* node_cmp_attr_ASM */
3800 * Set the default node attribute compare operation for an ir_op_ops.
3802 * @param code the opcode for the default operation
3803 * @param ops the operations initialized
3808 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
3812 ops->node_cmp_attr = node_cmp_attr_##a; \
3837 } /* firm_set_default_node_cmp_attr */
3840 * Compare function for two nodes in the hash table. Gets two
3841 * nodes as parameters. Returns 0 if the nodes are a cse.
3843 int identities_cmp(const void *elt, const void *key) {
3850 if (a == b) return 0;
3852 if ((get_irn_op(a) != get_irn_op(b)) ||
3853 (get_irn_mode(a) != get_irn_mode(b))) return 1;
3855 /* compare if a's in and b's in are of equal length */
3856 irn_arity_a = get_irn_intra_arity (a);
3857 if (irn_arity_a != get_irn_intra_arity(b))
3860 /* for block-local cse and op_pin_state_pinned nodes: */
3861 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
3862 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
3866 /* compare a->in[0..ins] with b->in[0..ins] */
3867 for (i = 0; i < irn_arity_a; i++)
3868 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
3872 * here, we already now that the nodes are identical except their
3875 if (a->op->ops.node_cmp_attr)
3876 return a->op->ops.node_cmp_attr(a, b);
3879 } /* identities_cmp */
3882 * Calculate a hash value of a node.
3884 unsigned ir_node_hash(ir_node *node) {
3888 if (node->op == op_Const) {
3889 /* special value for const, as they only differ in their tarval. */
3890 h = HASH_PTR(node->attr.con.tv);
3891 h = 9*h + HASH_PTR(get_irn_mode(node));
3892 } else if (node->op == op_SymConst) {
3893 /* special value for const, as they only differ in their symbol. */
3894 h = HASH_PTR(node->attr.symc.sym.type_p);
3895 h = 9*h + HASH_PTR(get_irn_mode(node));
3898 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
3899 h = irn_arity = get_irn_intra_arity(node);
3901 /* consider all in nodes... except the block if not a control flow. */
3902 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
3903 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
3907 h = 9*h + HASH_PTR(get_irn_mode(node));
3909 h = 9*h + HASH_PTR(get_irn_op(node));
3913 } /* ir_node_hash */
3915 pset *new_identities(void) {
3916 return new_pset(identities_cmp, N_IR_NODES);
3917 } /* new_identities */
3919 void del_identities(pset *value_table) {
3920 del_pset(value_table);
3921 } /* del_identities */
3924 * Return the canonical node computing the same value as n.
3926 * @param value_table The value table
3927 * @param n The node to lookup
3929 * Looks up the node in a hash table.
3931 * For Const nodes this is performed in the constructor, too. Const
3932 * nodes are extremely time critical because of their frequent use in
3933 * constant string arrays.
3935 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
3938 if (!value_table) return n;
3940 if (get_opt_reassociation()) {
3941 if (is_op_commutative(get_irn_op(n))) {
3942 ir_node *l = get_binop_left(n);
3943 ir_node *r = get_binop_right(n);
3945 /* for commutative operators perform a OP b == b OP a */
3946 if (get_irn_idx(l) > get_irn_idx(r)) {
3947 set_binop_left(n, r);
3948 set_binop_right(n, l);
3953 o = pset_find(value_table, n, ir_node_hash(n));
3962 * During construction we set the op_pin_state_pinned flag in the graph right when the
3963 * optimization is performed. The flag turning on procedure global cse could
3964 * be changed between two allocations. This way we are safe.
3966 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
3969 n = identify(value_table, n);
3970 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3971 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3973 } /* identify_cons */
3976 * Return the canonical node computing the same value as n.
3977 * Looks up the node in a hash table, enters it in the table
3978 * if it isn't there yet.
3980 ir_node *identify_remember(pset *value_table, ir_node *n) {
3983 if (!value_table) return n;
3985 if (get_opt_reassociation()) {
3986 if (is_op_commutative(get_irn_op(n))) {
3987 ir_node *l = get_binop_left(n);
3988 ir_node *r = get_binop_right(n);
3989 int l_idx = get_irn_idx(l);
3990 int r_idx = get_irn_idx(r);
3992 /* For commutative operators perform a OP b == b OP a but keep
3993 constants on the RIGHT side. This helps greatly in some optimizations.
3994 Moreover we use the idx number to make the form deterministic. */
3995 if (is_irn_constlike(l))
3997 if (is_irn_constlike(r))
3999 if (l_idx < r_idx) {
4000 set_binop_left(n, r);
4001 set_binop_right(n, l);
4006 /* lookup or insert in hash table with given hash key. */
4007 o = pset_insert(value_table, n, ir_node_hash(n));
4014 } /* identify_remember */
4016 /* Add a node to the identities value table. */
4017 void add_identities(pset *value_table, ir_node *node) {
4018 if (get_opt_cse() && is_no_Block(node))
4019 identify_remember(value_table, node);
4020 } /* add_identities */
4022 /* Visit each node in the value table of a graph. */
4023 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
4025 ir_graph *rem = current_ir_graph;
4027 current_ir_graph = irg;
4028 foreach_pset(irg->value_table, node)
4030 current_ir_graph = rem;
4031 } /* visit_all_identities */
4034 * Garbage in, garbage out. If a node has a dead input, i.e., the
4035 * Bad node is input to the node, return the Bad node.
4037 static INLINE ir_node *gigo(ir_node *node) {
4039 ir_op *op = get_irn_op(node);
4041 /* remove garbage blocks by looking at control flow that leaves the block
4042 and replacing the control flow by Bad. */
4043 if (get_irn_mode(node) == mode_X) {
4044 ir_node *block = get_nodes_block(skip_Proj(node));
4046 /* Don't optimize nodes in immature blocks. */
4047 if (!get_Block_matured(block)) return node;
4048 /* Don't optimize End, may have Bads. */
4049 if (op == op_End) return node;
4051 if (is_Block(block)) {
4052 irn_arity = get_irn_arity(block);
4053 for (i = 0; i < irn_arity; i++) {
4054 if (!is_Bad(get_irn_n(block, i)))
4057 if (i == irn_arity) {
4058 ir_graph *irg = get_irn_irg(block);
4059 /* the start block is never dead */
4060 if (block != get_irg_start_block(irg)
4061 && block != get_irg_end_block(irg))
4067 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
4068 blocks predecessors is dead. */
4069 if (op != op_Block && op != op_Phi && op != op_Tuple) {
4070 irn_arity = get_irn_arity(node);
4073 * Beware: we can only read the block of a non-floating node.
4075 if (is_irn_pinned_in_irg(node) &&
4076 is_Block_dead(get_nodes_block(node)))
4079 for (i = 0; i < irn_arity; i++) {
4080 ir_node *pred = get_irn_n(node, i);
4085 /* Propagating Unknowns here seems to be a bad idea, because
4086 sometimes we need a node as a input and did not want that
4088 However, it might be useful to move this into a later phase
4089 (if you think that optimizing such code is useful). */
4090 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
4091 return new_Unknown(get_irn_mode(node));
4096 /* With this code we violate the agreement that local_optimize
4097 only leaves Bads in Block, Phi and Tuple nodes. */
4098 /* If Block has only Bads as predecessors it's garbage. */
4099 /* If Phi has only Bads as predecessors it's garbage. */
4100 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
4101 irn_arity = get_irn_arity(node);
4102 for (i = 0; i < irn_arity; i++) {
4103 if (!is_Bad(get_irn_n(node, i))) break;
4105 if (i == irn_arity) node = new_Bad();
4112 * These optimizations deallocate nodes from the obstack.
4113 * It can only be called if it is guaranteed that no other nodes
4114 * reference this one, i.e., right after construction of a node.
4116 * @param n The node to optimize
4118 * current_ir_graph must be set to the graph of the node!
4120 ir_node *optimize_node(ir_node *n) {
4123 ir_opcode iro = get_irn_opcode(n);
4125 /* Always optimize Phi nodes: part of the construction. */
4126 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
4128 /* constant expression evaluation / constant folding */
4129 if (get_opt_constant_folding()) {
4130 /* neither constants nor Tuple values can be evaluated */
4131 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
4132 /* try to evaluate */
4133 tv = computed_value(n);
4134 if (tv != tarval_bad) {
4136 ir_type *old_tp = get_irn_type(n);
4137 int i, arity = get_irn_arity(n);
4141 * Try to recover the type of the new expression.
4143 for (i = 0; i < arity && !old_tp; ++i)
4144 old_tp = get_irn_type(get_irn_n(n, i));
4147 * we MUST copy the node here temporary, because it's still needed
4148 * for DBG_OPT_CSTEVAL
4150 node_size = offsetof(ir_node, attr) + n->op->attr_size;
4151 oldn = alloca(node_size);
4153 memcpy(oldn, n, node_size);
4154 CLONE_ARR_A(ir_node *, oldn->in, n->in);
4156 /* ARG, copy the in array, we need it for statistics */
4157 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
4159 /* note the inplace edges module */
4160 edges_node_deleted(n, current_ir_graph);
4162 /* evaluation was successful -- replace the node. */
4163 irg_kill_node(current_ir_graph, n);
4164 nw = new_Const(get_tarval_mode (tv), tv);
4166 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
4167 set_Const_type(nw, old_tp);
4168 DBG_OPT_CSTEVAL(oldn, nw);
4174 /* remove unnecessary nodes */
4175 if (get_opt_constant_folding() ||
4176 (iro == iro_Phi) || /* always optimize these nodes. */
4178 (iro == iro_Proj) ||
4179 (iro == iro_Block) ) /* Flags tested local. */
4180 n = equivalent_node(n);
4182 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
4184 /* Common Subexpression Elimination.
4186 * Checks whether n is already available.
4187 * The block input is used to distinguish different subexpressions. Right
4188 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
4189 * subexpressions within a block.
4192 n = identify_cons(current_ir_graph->value_table, n);
4195 edges_node_deleted(oldn, current_ir_graph);
4197 /* We found an existing, better node, so we can deallocate the old node. */
4198 irg_kill_node(current_ir_graph, oldn);
4202 /* Some more constant expression evaluation that does not allow to
4204 iro = get_irn_opcode(n);
4205 if (get_opt_constant_folding() ||
4206 (iro == iro_Cond) ||
4207 (iro == iro_Proj)) /* Flags tested local. */
4208 n = transform_node(n);
4210 /* Remove nodes with dead (Bad) input.
4211 Run always for transformation induced Bads. */
4214 /* Now we have a legal, useful node. Enter it in hash table for CSE */
4215 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
4216 n = identify_remember(current_ir_graph->value_table, n);
4220 } /* optimize_node */
4224 * These optimizations never deallocate nodes (in place). This can cause dead
4225 * nodes lying on the obstack. Remove these by a dead node elimination,
4226 * i.e., a copying garbage collection.
4228 ir_node *optimize_in_place_2(ir_node *n) {
4231 ir_opcode iro = get_irn_opcode(n);
4233 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
4235 /* constant expression evaluation / constant folding */
4236 if (get_opt_constant_folding()) {
4237 /* neither constants nor Tuple values can be evaluated */
4238 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
4239 /* try to evaluate */
4240 tv = computed_value(n);
4241 if (tv != tarval_bad) {
4242 /* evaluation was successful -- replace the node. */
4243 ir_type *old_tp = get_irn_type(n);
4244 int i, arity = get_irn_arity(n);
4247 * Try to recover the type of the new expression.
4249 for (i = 0; i < arity && !old_tp; ++i)
4250 old_tp = get_irn_type(get_irn_n(n, i));
4252 n = new_Const(get_tarval_mode(tv), tv);
4254 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
4255 set_Const_type(n, old_tp);
4257 DBG_OPT_CSTEVAL(oldn, n);
4263 /* remove unnecessary nodes */
4264 if (get_opt_constant_folding() ||
4265 (iro == iro_Phi) || /* always optimize these nodes. */
4266 (iro == iro_Id) || /* ... */
4267 (iro == iro_Proj) || /* ... */
4268 (iro == iro_Block) ) /* Flags tested local. */
4269 n = equivalent_node(n);
4271 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
4273 /** common subexpression elimination **/
4274 /* Checks whether n is already available. */
4275 /* The block input is used to distinguish different subexpressions. Right
4276 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
4277 subexpressions within a block. */
4278 if (get_opt_cse()) {
4279 n = identify(current_ir_graph->value_table, n);
4282 /* Some more constant expression evaluation. */
4283 iro = get_irn_opcode(n);
4284 if (get_opt_constant_folding() ||
4285 (iro == iro_Cond) ||
4286 (iro == iro_Proj)) /* Flags tested local. */
4287 n = transform_node(n);
4289 /* Remove nodes with dead (Bad) input.
4290 Run always for transformation induced Bads. */
4293 /* Now we can verify the node, as it has no dead inputs any more. */
4296 /* Now we have a legal, useful node. Enter it in hash table for cse.
4297 Blocks should be unique anyways. (Except the successor of start:
4298 is cse with the start block!) */
4299 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
4300 n = identify_remember(current_ir_graph->value_table, n);
4303 } /* optimize_in_place_2 */
4306 * Wrapper for external use, set proper status bits after optimization.
4308 ir_node *optimize_in_place(ir_node *n) {
4309 /* Handle graph state */
4310 assert(get_irg_phase_state(current_ir_graph) != phase_building);
4312 if (get_opt_global_cse())
4313 set_irg_pinned(current_ir_graph, op_pin_state_floats);
4314 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
4315 set_irg_outs_inconsistent(current_ir_graph);
4317 /* FIXME: Maybe we could also test whether optimizing the node can
4318 change the control graph. */
4319 set_irg_doms_inconsistent(current_ir_graph);
4320 return optimize_in_place_2(n);
4321 } /* optimize_in_place */
4324 * Sets the default operation for an ir_ops.
4326 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
4327 ops = firm_set_default_computed_value(code, ops);
4328 ops = firm_set_default_equivalent_node(code, ops);
4329 ops = firm_set_default_transform_node(code, ops);
4330 ops = firm_set_default_node_cmp_attr(code, ops);
4331 ops = firm_set_default_get_type(code, ops);
4332 ops = firm_set_default_get_type_attr(code, ops);
4333 ops = firm_set_default_get_entity_attr(code, ops);
4336 } /* firm_set_default_operations */