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))
964 b = get_Sub_right(n);
966 /* Beware: modes might be different */
967 if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
968 if (mode == get_irn_mode(a)) {
971 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_0);
973 } else if (get_irn_op(a) == op_Add) {
974 if (mode_wrap_around(mode)) {
975 ir_node *left = get_Add_left(a);
976 ir_node *right = get_Add_right(a);
979 if (mode == get_irn_mode(right)) {
981 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
983 } else if (right == b) {
984 if (mode == get_irn_mode(left)) {
986 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_ADD_SUB);
992 } /* equivalent_node_Sub */
996 * Optimize an "idempotent unary op", ie op(op(n)) = n.
999 * -(-a) == a, but might overflow two times.
1000 * We handle it anyway here but the better way would be a
1001 * flag. This would be needed for Pascal for instance.
1003 static ir_node *equivalent_node_idempotent_unop(ir_node *n) {
1005 ir_node *pred = get_unop_op(n);
1007 /* optimize symmetric unop */
1008 if (get_irn_op(pred) == get_irn_op(n)) {
1009 n = get_unop_op(pred);
1010 DBG_OPT_ALGSIM2(oldn, pred, n);
1013 } /* equivalent_node_idempotent_unop */
1015 /** Optimize Not(Not(x)) == x. */
1016 #define equivalent_node_Not equivalent_node_idempotent_unop
1018 /** --x == x ??? Is this possible or can --x raise an
1019 out of bounds exception if min =! max? */
1020 #define equivalent_node_Minus equivalent_node_idempotent_unop
1023 * Optimize a * 1 = 1 * a = a.
1025 static ir_node *equivalent_node_Mul(ir_node *n) {
1027 ir_node *a = get_Mul_left(n);
1028 ir_node *b = get_Mul_right(n);
1030 /* Mul is commutative and has again an other neutral element. */
1031 if (classify_tarval(value_of(a)) == TV_CLASSIFY_ONE) {
1033 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1034 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) {
1036 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_NEUTRAL_1);
1039 } /* equivalent_node_Mul */
1042 * Optimize a / 1 = a.
1044 static ir_node *equivalent_node_Div(ir_node *n) {
1045 ir_node *a = get_Div_left(n);
1046 ir_node *b = get_Div_right(n);
1048 /* Div is not commutative. */
1049 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1050 /* Turn Div into a tuple (mem, bad, a) */
1051 ir_node *mem = get_Div_mem(n);
1052 ir_node *blk = get_nodes_block(n);
1053 turn_into_tuple(n, pn_Div_max);
1054 set_Tuple_pred(n, pn_Div_M, mem);
1055 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
1056 set_Tuple_pred(n, pn_Div_X_except, new_Bad()); /* no exception */
1057 set_Tuple_pred(n, pn_Div_res, a);
1060 } /* equivalent_node_Div */
1063 * Optimize a / 1.0 = a.
1065 static ir_node *equivalent_node_Quot(ir_node *n) {
1066 ir_node *a = get_Quot_left(n);
1067 ir_node *b = get_Quot_right(n);
1069 /* Div is not commutative. */
1070 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* Quot(x, 1) == x */
1071 /* Turn Quot into a tuple (mem, bad, a) */
1072 ir_node *mem = get_Quot_mem(n);
1073 ir_node *blk = get_nodes_block(n);
1074 turn_into_tuple(n, pn_Quot_max);
1075 set_Tuple_pred(n, pn_Quot_M, mem);
1076 set_Tuple_pred(n, pn_Quot_X_regular, new_r_Jmp(current_ir_graph, blk));
1077 set_Tuple_pred(n, pn_Quot_X_except, new_Bad()); /* no exception */
1078 set_Tuple_pred(n, pn_Quot_res, a);
1081 } /* equivalent_node_Quot */
1084 * Optimize a / 1 = a.
1086 static ir_node *equivalent_node_DivMod(ir_node *n) {
1087 ir_node *b = get_DivMod_right(n);
1089 /* Div is not commutative. */
1090 if (classify_tarval(value_of(b)) == TV_CLASSIFY_ONE) { /* div(x, 1) == x */
1091 /* Turn DivMod into a tuple (mem, bad, a, 0) */
1092 ir_node *a = get_DivMod_left(n);
1093 ir_node *mem = get_Div_mem(n);
1094 ir_node *blk = get_nodes_block(n);
1095 ir_mode *mode = get_DivMod_resmode(n);
1097 turn_into_tuple(n, pn_DivMod_max);
1098 set_Tuple_pred(n, pn_DivMod_M, mem);
1099 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
1100 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
1101 set_Tuple_pred(n, pn_DivMod_res_div, a);
1102 set_Tuple_pred(n, pn_DivMod_res_mod, new_Const(mode, get_mode_null(mode)));
1105 } /* equivalent_node_DivMod */
1108 * Use algebraic simplification a | a = a | 0 = 0 | a = a.
1110 static ir_node *equivalent_node_Or(ir_node *n) {
1113 ir_node *a = get_Or_left(n);
1114 ir_node *b = get_Or_right(n);
1117 n = a; /* Or has it's own neutral element */
1118 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_OR);
1119 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_NULL) {
1121 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1122 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_NULL) {
1124 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_OR);
1128 } /* equivalent_node_Or */
1131 * Optimize a & 0b1...1 = 0b1...1 & a = a & a = a.
1133 static ir_node *equivalent_node_And(ir_node *n) {
1136 ir_node *a = get_And_left(n);
1137 ir_node *b = get_And_right(n);
1140 n = a; /* And has it's own neutral element */
1141 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_AND);
1142 } else if (classify_tarval(value_of(a)) == TV_CLASSIFY_ALL_ONE) {
1144 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1145 } else if (classify_tarval(value_of(b)) == TV_CLASSIFY_ALL_ONE) {
1147 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_AND);
1150 } /* equivalent_node_And */
1153 * Try to remove useless Conv's:
1155 static ir_node *equivalent_node_Conv(ir_node *n) {
1157 ir_node *a = get_Conv_op(n);
1160 ir_mode *n_mode = get_irn_mode(n);
1161 ir_mode *a_mode = get_irn_mode(a);
1163 if (n_mode == a_mode) { /* No Conv necessary */
1164 /* leave strict floating point Conv's */
1165 if (get_Conv_strict(n))
1168 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CONV);
1169 } else if (get_irn_op(a) == op_Conv) { /* Conv(Conv(b)) */
1173 n_mode = get_irn_mode(n);
1174 b_mode = get_irn_mode(b);
1176 if (n_mode == b_mode) {
1177 if (n_mode == mode_b) {
1178 n = b; /* Convb(Conv*(xxxb(...))) == xxxb(...) */
1179 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1180 } else if (mode_is_int(n_mode) || mode_is_character(n_mode)) {
1181 if (smaller_mode(b_mode, a_mode)){
1182 n = b; /* ConvS(ConvL(xxxS(...))) == xxxS(...) */
1183 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_CONV);
1189 } /* equivalent_node_Conv */
1192 * A Cast may be removed if the type of the previous node
1193 * is already the type of the Cast.
1195 static ir_node *equivalent_node_Cast(ir_node *n) {
1197 ir_node *pred = get_Cast_op(n);
1199 if (get_irn_type(pred) == get_Cast_type(n)) {
1201 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_CAST);
1204 } /* equivalent_node_Cast */
1207 * Several optimizations:
1208 * - no Phi in start block.
1209 * - remove Id operators that are inputs to Phi
1210 * - fold Phi-nodes, iff they have only one predecessor except
1213 static ir_node *equivalent_node_Phi(ir_node *n) {
1217 ir_node *block = NULL; /* to shutup gcc */
1218 ir_node *first_val = NULL; /* to shutup gcc */
1220 if (!get_opt_normalize()) return n;
1222 n_preds = get_Phi_n_preds(n);
1224 block = get_nodes_block(n);
1225 /* @@@ fliegt 'raus, sollte aber doch immer wahr sein!!!
1226 assert(get_irn_arity(block) == n_preds && "phi in wrong block!"); */
1227 if ((is_Block_dead(block)) || /* Control dead */
1228 (block == get_irg_start_block(current_ir_graph))) /* There should be no Phi nodes */
1229 return new_Bad(); /* in the Start Block. */
1231 if (n_preds == 0) return n; /* Phi of dead Region without predecessors. */
1233 /* If the Block has a Bad pred, we also have one. */
1234 for (i = 0; i < n_preds; ++i)
1235 if (is_Bad(get_Block_cfgpred(block, i)))
1236 set_Phi_pred(n, i, new_Bad());
1238 /* Find first non-self-referencing input */
1239 for (i = 0; i < n_preds; ++i) {
1240 first_val = get_Phi_pred(n, i);
1241 if ( (first_val != n) /* not self pointer */
1243 && (! is_Bad(first_val))
1245 ) { /* value not dead */
1246 break; /* then found first value. */
1251 /* A totally Bad or self-referencing Phi (we didn't break the above loop) */
1255 /* search for rest of inputs, determine if any of these
1256 are non-self-referencing */
1257 while (++i < n_preds) {
1258 ir_node *scnd_val = get_Phi_pred(n, i);
1259 if ( (scnd_val != n)
1260 && (scnd_val != first_val)
1262 && (! is_Bad(scnd_val))
1270 /* Fold, if no multiple distinct non-self-referencing inputs */
1272 DBG_OPT_PHI(oldn, n);
1275 } /* equivalent_node_Phi */
1278 * Several optimizations:
1279 * - no Sync in start block.
1280 * - fold Sync-nodes, iff they have only one predecessor except
1283 static ir_node *equivalent_node_Sync(ir_node *n) {
1287 ir_node *first_val = NULL; /* to shutup gcc */
1289 if (!get_opt_normalize()) return n;
1291 n_preds = get_Sync_n_preds(n);
1293 /* Find first non-self-referencing input */
1294 for (i = 0; i < n_preds; ++i) {
1295 first_val = get_Sync_pred(n, i);
1296 if ((first_val != n) /* not self pointer */ &&
1297 (! is_Bad(first_val))
1298 ) { /* value not dead */
1299 break; /* then found first value. */
1304 /* A totally Bad or self-referencing Sync (we didn't break the above loop) */
1307 /* search the rest of inputs, determine if any of these
1308 are non-self-referencing */
1309 while (++i < n_preds) {
1310 ir_node *scnd_val = get_Sync_pred(n, i);
1311 if ((scnd_val != n) &&
1312 (scnd_val != first_val) &&
1313 (! is_Bad(scnd_val))
1319 /* Fold, if no multiple distinct non-self-referencing inputs */
1321 DBG_OPT_SYNC(oldn, n);
1324 } /* equivalent_node_Sync */
1327 * Optimize Proj(Tuple) and gigo() for ProjX in Bad block,
1328 * ProjX(Load) and ProjX(Store).
1330 static ir_node *equivalent_node_Proj(ir_node *n) {
1332 ir_node *a = get_Proj_pred(n);
1334 if ( get_irn_op(a) == op_Tuple) {
1335 /* Remove the Tuple/Proj combination. */
1336 if ( get_Proj_proj(n) <= get_Tuple_n_preds(a) ) {
1337 n = get_Tuple_pred(a, get_Proj_proj(n));
1338 DBG_OPT_TUPLE(oldn, a, n);
1340 assert(0); /* This should not happen! */
1343 } else if (get_irn_mode(n) == mode_X) {
1344 if (is_Block_dead(get_nodes_block(skip_Proj(n)))) {
1345 /* Remove dead control flow -- early gigo(). */
1347 } else if (get_opt_ldst_only_null_ptr_exceptions()) {
1348 ir_op *op = get_irn_op(a);
1350 if (op == op_Load || op == op_Store) {
1351 /* get the load/store address */
1352 ir_node *addr = get_irn_n(a, 1);
1355 if (value_not_null(addr, &confirm)) {
1356 if (confirm == NULL) {
1357 /* this node may float if it did not depend on a Confirm */
1358 set_irn_pinned(a, op_pin_state_floats);
1368 } /* equivalent_node_Proj */
1373 static ir_node *equivalent_node_Id(ir_node *n) {
1378 } while (get_irn_op(n) == op_Id);
1380 DBG_OPT_ID(oldn, n);
1382 } /* equivalent_node_Id */
1387 static ir_node *equivalent_node_Mux(ir_node *n)
1389 ir_node *oldn = n, *sel = get_Mux_sel(n);
1390 tarval *ts = value_of(sel);
1392 /* Mux(true, f, t) == t */
1393 if (ts == tarval_b_true) {
1394 n = get_Mux_true(n);
1395 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1397 /* Mux(false, f, t) == f */
1398 else if (ts == tarval_b_false) {
1399 n = get_Mux_false(n);
1400 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_C);
1402 /* Mux(v, x, x) == x */
1403 else if (get_Mux_false(n) == get_Mux_true(n)) {
1404 n = get_Mux_true(n);
1405 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_EQ);
1407 else if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(get_irn_mode(n))) {
1408 ir_node *cmp = get_Proj_pred(sel);
1409 long proj_nr = get_Proj_proj(sel);
1410 ir_node *b = get_Mux_false(n);
1411 ir_node *a = get_Mux_true(n);
1414 * Note: normalization puts the constant on the right site,
1415 * so we check only one case.
1417 * Note further that these optimization work even for floating point
1418 * with NaN's because -NaN == NaN.
1419 * However, if +0 and -0 is handled differently, we cannot use the first one.
1421 if (get_irn_op(cmp) == op_Cmp && get_Cmp_left(cmp) == a) {
1422 if (classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
1423 /* Mux(a CMP 0, X, a) */
1424 if (get_irn_op(b) == op_Minus && get_Minus_op(b) == a) {
1425 /* Mux(a CMP 0, -a, a) */
1426 if (proj_nr == pn_Cmp_Eq) {
1427 /* Mux(a == 0, -a, a) ==> -a */
1429 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1430 } else if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1431 /* Mux(a != 0, -a, a) ==> a */
1433 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1435 } else if (classify_Const(b) == CNST_NULL) {
1436 /* Mux(a CMP 0, 0, a) */
1437 if (proj_nr == pn_Cmp_Lg || proj_nr == pn_Cmp_Ne) {
1438 /* Mux(a != 0, 0, a) ==> a */
1440 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1441 } else if (proj_nr == pn_Cmp_Eq) {
1442 /* Mux(a == 0, 0, a) ==> 0 */
1444 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_MUX_TRANSFORM);
1451 } /* equivalent_node_Mux */
1454 * Returns a equivalent node of a Psi: if a condition is true
1455 * and all previous conditions are false we know its value.
1456 * If all conditions are false its value is the default one.
1458 static ir_node *equivalent_node_Psi(ir_node *n) {
1460 return equivalent_node_Mux(n);
1462 } /* equivalent_node_Psi */
1465 * Optimize -a CMP -b into b CMP a.
1466 * This works only for for modes where unary Minus
1468 * Note that two-complement integers can Overflow
1469 * so it will NOT work.
1471 * For == and != can be handled in Proj(Cmp)
1473 static ir_node *equivalent_node_Cmp(ir_node *n) {
1474 ir_node *left = get_Cmp_left(n);
1475 ir_node *right = get_Cmp_right(n);
1477 if (get_irn_op(left) == op_Minus && get_irn_op(right) == op_Minus &&
1478 !mode_overflow_on_unary_Minus(get_irn_mode(left))) {
1479 left = get_Minus_op(left);
1480 right = get_Minus_op(right);
1481 set_Cmp_left(n, right);
1482 set_Cmp_right(n, left);
1485 } /* equivalent_node_Cmp */
1488 * Remove Confirm nodes if setting is on.
1489 * Replace Confirms(x, '=', Constlike) by Constlike.
1491 static ir_node *equivalent_node_Confirm(ir_node *n) {
1492 ir_node *pred = get_Confirm_value(n);
1493 pn_Cmp pnc = get_Confirm_cmp(n);
1495 if (get_irn_op(pred) == op_Confirm && pnc == get_Confirm_cmp(pred)) {
1497 * rare case: two identical Confirms one after another,
1498 * replace the second one with the first.
1502 if (pnc == pn_Cmp_Eq) {
1503 ir_node *bound = get_Confirm_bound(n);
1506 * Optimize a rare case:
1507 * Confirm(x, '=', Constlike) ==> Constlike
1509 if (is_irn_constlike(bound)) {
1510 DBG_OPT_CONFIRM(n, bound);
1514 return get_opt_remove_confirm() ? get_Confirm_value(n) : n;
1518 * Optimize CopyB(mem, x, x) into a Nop.
1520 static ir_node *equivalent_node_CopyB(ir_node *n) {
1521 ir_node *a = get_CopyB_dst(n);
1522 ir_node *b = get_CopyB_src(n);
1525 /* Turn CopyB into a tuple (mem, bad, bad) */
1526 ir_node *mem = get_CopyB_mem(n);
1527 ir_node *blk = get_nodes_block(n);
1528 turn_into_tuple(n, pn_CopyB_max);
1529 set_Tuple_pred(n, pn_CopyB_M, mem);
1530 set_Tuple_pred(n, pn_CopyB_X_regular, new_r_Jmp(current_ir_graph, blk));
1531 set_Tuple_pred(n, pn_CopyB_X_except, new_Bad()); /* no exception */
1532 set_Tuple_pred(n, pn_CopyB_M_except, new_Bad());
1535 } /* equivalent_node_CopyB */
1538 * Optimize Bounds(idx, idx, upper) into idx.
1540 static ir_node *equivalent_node_Bound(ir_node *n) {
1541 ir_node *idx = get_Bound_index(n);
1542 ir_node *lower = get_Bound_lower(n);
1545 /* By definition lower < upper, so if idx == lower -->
1546 lower <= idx && idx < upper */
1548 /* Turn Bound into a tuple (mem, bad, idx) */
1551 ir_node *pred = skip_Proj(idx);
1553 if (get_irn_op(pred) == op_Bound) {
1555 * idx was Bounds_check previously, it is still valid if
1556 * lower <= pred_lower && pred_upper <= upper.
1558 ir_node *upper = get_Bound_upper(n);
1559 if (get_Bound_lower(pred) == lower &&
1560 get_Bound_upper(pred) == upper) {
1562 * One could expect that we simply return the previous
1563 * Bound here. However, this would be wrong, as we could
1564 * add an exception Proj to a new location than.
1565 * So, we must turn in into a tuple
1572 /* Turn Bound into a tuple (mem, bad, idx) */
1573 ir_node *mem = get_Bound_mem(n);
1574 ir_node *blk = get_nodes_block(n);
1575 turn_into_tuple(n, pn_Bound_max);
1576 set_Tuple_pred(n, pn_Bound_M, mem);
1577 set_Tuple_pred(n, pn_Bound_X_regular, new_r_Jmp(current_ir_graph, blk)); /* no exception */
1578 set_Tuple_pred(n, pn_Bound_X_except, new_Bad()); /* no exception */
1579 set_Tuple_pred(n, pn_Bound_res, idx);
1582 } /* equivalent_node_Bound */
1585 * equivalent_node() returns a node equivalent to input n. It skips all nodes that
1586 * perform no actual computation, as, e.g., the Id nodes. It does not create
1587 * new nodes. It is therefore safe to free n if the node returned is not n.
1588 * If a node returns a Tuple we can not just skip it. If the size of the
1589 * in array fits, we transform n into a tuple (e.g., Div).
1591 ir_node *equivalent_node(ir_node *n) {
1592 if (n->op->ops.equivalent_node)
1593 return n->op->ops.equivalent_node(n);
1595 } /* equivalent_node */
1598 * Sets the default equivalent node operation for an ir_op_ops.
1600 * @param code the opcode for the default operation
1601 * @param ops the operations initialized
1606 static ir_op_ops *firm_set_default_equivalent_node(ir_opcode code, ir_op_ops *ops)
1610 ops->equivalent_node = equivalent_node_##a; \
1650 } /* firm_set_default_equivalent_node */
1653 * Do node specific optimizations of nodes predecessors.
1655 static void optimize_preds(ir_node *n) {
1656 switch (get_irn_opcode(n)) {
1658 case iro_Cmp: { /* We don't want Cast as input to Cmp. */
1659 ir_node *a = get_Cmp_left(n), *b = get_Cmp_right(n);
1661 if (get_irn_op(a) == op_Cast) {
1665 if (get_irn_op(b) == op_Cast) {
1667 set_Cmp_right(n, b);
1674 } /* optimize_preds */
1677 * Returns non-zero if a node is a Phi node
1678 * with all predecessors constant.
1680 static int is_const_Phi(ir_node *n) {
1685 for (i = get_irn_arity(n) - 1; i >= 0; --i)
1686 if (! is_Const(get_irn_n(n, i)))
1689 } /* is_const_Phi */
1692 * Apply an evaluator on a binop with a constant operators (and one Phi).
1694 * @param phi the Phi node
1695 * @param other the other operand
1696 * @param eval an evaluator function
1697 * @param left if non-zero, other is the left operand, else the right
1699 * @return a new Phi node if the conversion was successful, NULL else
1701 static ir_node *apply_binop_on_phi(ir_node *phi, tarval *other, tarval *(*eval)(tarval *, tarval *), int left) {
1707 int i, n = get_irn_arity(phi);
1709 NEW_ARR_A(void *, res, n);
1711 for (i = 0; i < n; ++i) {
1712 pred = get_irn_n(phi, i);
1713 tv = get_Const_tarval(pred);
1714 tv = eval(other, tv);
1716 if (tv == tarval_bad) {
1717 /* folding failed, bad */
1723 for (i = 0; i < n; ++i) {
1724 pred = get_irn_n(phi, i);
1725 tv = get_Const_tarval(pred);
1726 tv = eval(tv, other);
1728 if (tv == tarval_bad) {
1729 /* folding failed, bad */
1735 mode = get_irn_mode(phi);
1736 irg = current_ir_graph;
1737 for (i = 0; i < n; ++i) {
1738 pred = get_irn_n(phi, i);
1739 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1740 mode, res[i], get_Const_type(pred));
1742 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1743 } /* apply_binop_on_phi */
1746 * Apply an evaluator on a unop with a constant operator (a Phi).
1748 * @param phi the Phi node
1749 * @param eval an evaluator function
1751 * @return a new Phi node if the conversion was successful, NULL else
1753 static ir_node *apply_unop_on_phi(ir_node *phi, tarval *(*eval)(tarval *)) {
1759 int i, n = get_irn_arity(phi);
1761 NEW_ARR_A(void *, res, n);
1762 for (i = 0; i < n; ++i) {
1763 pred = get_irn_n(phi, i);
1764 tv = get_Const_tarval(pred);
1767 if (tv == tarval_bad) {
1768 /* folding failed, bad */
1773 mode = get_irn_mode(phi);
1774 irg = current_ir_graph;
1775 for (i = 0; i < n; ++i) {
1776 pred = get_irn_n(phi, i);
1777 res[i] = new_r_Const_type(irg, get_irg_start_block(irg),
1778 mode, res[i], get_Const_type(pred));
1780 return new_r_Phi(irg, get_nodes_block(phi), n, (ir_node **)res, mode);
1781 } /* apply_unop_on_phi */
1784 * Transform AddP(P, ConvIs(Iu)), AddP(P, ConvIu(Is)) and
1785 * SubP(P, ConvIs(Iu)), SubP(P, ConvIu(Is)).
1786 * If possible, remove the Conv's.
1788 static ir_node *transform_node_AddSub(ir_node *n) {
1789 ir_mode *mode = get_irn_mode(n);
1791 if (mode_is_reference(mode)) {
1792 ir_node *left = get_binop_left(n);
1793 ir_node *right = get_binop_right(n);
1794 int ref_bits = get_mode_size_bits(mode);
1796 if (get_irn_op(left) == op_Conv) {
1797 ir_mode *mode = get_irn_mode(left);
1798 int bits = get_mode_size_bits(mode);
1800 if (ref_bits == bits &&
1801 mode_is_int(mode) &&
1802 get_mode_arithmetic(mode) == irma_twos_complement) {
1803 ir_node *pre = get_Conv_op(left);
1804 ir_mode *pre_mode = get_irn_mode(pre);
1806 if (mode_is_int(pre_mode) &&
1807 get_mode_size_bits(pre_mode) == bits &&
1808 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1809 /* ok, this conv just changes to sign, moreover the calculation
1810 * is done with same number of bits as our address mode, so
1811 * we can ignore the conv as address calculation can be viewed
1812 * as either signed or unsigned
1814 set_binop_left(n, pre);
1819 if (get_irn_op(right) == op_Conv) {
1820 ir_mode *mode = get_irn_mode(right);
1821 int bits = get_mode_size_bits(mode);
1823 if (ref_bits == bits &&
1824 mode_is_int(mode) &&
1825 get_mode_arithmetic(mode) == irma_twos_complement) {
1826 ir_node *pre = get_Conv_op(right);
1827 ir_mode *pre_mode = get_irn_mode(pre);
1829 if (mode_is_int(pre_mode) &&
1830 get_mode_size_bits(pre_mode) == bits &&
1831 get_mode_arithmetic(pre_mode) == irma_twos_complement) {
1832 /* ok, this conv just changes to sign, moreover the calculation
1833 * is done with same number of bits as our address mode, so
1834 * we can ignore the conv as address calculation can be viewed
1835 * as either signed or unsigned
1837 set_binop_right(n, pre);
1843 } /* transform_node_AddSub */
1845 #define HANDLE_BINOP_PHI(op,a,b,c) \
1847 if (is_Const(b) && is_const_Phi(a)) { \
1848 /* check for Op(Phi, Const) */ \
1849 c = apply_binop_on_phi(a, get_Const_tarval(b), op, 0); \
1851 else if (is_Const(a) && is_const_Phi(b)) { \
1852 /* check for Op(Const, Phi) */ \
1853 c = apply_binop_on_phi(b, get_Const_tarval(a), op, 1); \
1856 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1860 #define HANDLE_UNOP_PHI(op,a,c) \
1862 if (is_const_Phi(a)) { \
1863 /* check for Op(Phi) */ \
1864 c = apply_unop_on_phi(a, op); \
1867 DBG_OPT_ALGSIM0(oldn, c, FS_OPT_CONST_PHI); \
1873 * Do the AddSub optimization, then Transform
1874 * Constant folding on Phi
1875 * Add(a,a) -> Mul(a, 2)
1876 * Add(Mul(a, x), a) -> Mul(a, x+1)
1877 * if the mode is integer or float.
1878 * Transform Add(a,-b) into Sub(a,b).
1879 * Reassociation might fold this further.
1881 static ir_node *transform_node_Add(ir_node *n) {
1883 ir_node *a, *b, *c, *oldn = n;
1885 n = transform_node_AddSub(n);
1887 a = get_Add_left(n);
1888 b = get_Add_right(n);
1890 HANDLE_BINOP_PHI(tarval_add, a,b,c);
1892 mode = get_irn_mode(n);
1894 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
1895 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
1898 if (mode_is_num(mode)) {
1900 ir_node *block = get_irn_n(n, -1);
1903 get_irn_dbg_info(n),
1907 new_r_Const_long(current_ir_graph, block, mode, 2),
1909 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_A);
1910 } else if (get_irn_op(a) == op_Minus) {
1912 get_irn_dbg_info(n),
1918 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1919 } else if (get_irn_op(b) == op_Minus) {
1921 get_irn_dbg_info(n),
1927 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_A_MINUS_B);
1929 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1930 else if (!get_opt_reassociation() && get_irn_op(a) == op_Mul) {
1931 ir_node *ma = get_Mul_left(a);
1932 ir_node *mb = get_Mul_right(a);
1935 ir_node *blk = get_irn_n(n, -1);
1937 get_irn_dbg_info(n), current_ir_graph, blk,
1940 get_irn_dbg_info(n), current_ir_graph, blk,
1942 new_r_Const_long(current_ir_graph, blk, mode, 1),
1945 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1946 } else if (b == mb) {
1947 ir_node *blk = get_irn_n(n, -1);
1949 get_irn_dbg_info(n), current_ir_graph, blk,
1952 get_irn_dbg_info(n), current_ir_graph, blk,
1954 new_r_Const_long(current_ir_graph, blk, mode, 1),
1957 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1960 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
1961 else if (!get_opt_reassociation() && get_irn_op(b) == op_Mul) {
1962 ir_node *ma = get_Mul_left(b);
1963 ir_node *mb = get_Mul_right(b);
1966 ir_node *blk = get_irn_n(n, -1);
1968 get_irn_dbg_info(n), current_ir_graph, blk,
1971 get_irn_dbg_info(n), current_ir_graph, blk,
1973 new_r_Const_long(current_ir_graph, blk, mode, 1),
1976 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1977 } else if (a == mb) {
1978 ir_node *blk = get_irn_n(n, -1);
1980 get_irn_dbg_info(n), current_ir_graph, blk,
1983 get_irn_dbg_info(n), current_ir_graph, blk,
1985 new_r_Const_long(current_ir_graph, blk, mode, 1),
1988 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_ADD_MUL_A_X_A);
1993 } /* transform_node_Add */
1996 * Do the AddSub optimization, then Transform
1997 * Constant folding on Phi
1998 * Sub(0,a) -> Minus(a)
1999 * Sub(Mul(a, x), a) -> Mul(a, x-1)
2000 * Sub(Sub(x, y), b) -> Sub(x, Add(y,b))
2002 static ir_node *transform_node_Sub(ir_node *n) {
2007 n = transform_node_AddSub(n);
2009 a = get_Sub_left(n);
2010 b = get_Sub_right(n);
2012 HANDLE_BINOP_PHI(tarval_sub, a,b,c);
2014 mode = get_irn_mode(n);
2016 /* for FP these optimizations are only allowed if fp_strict_algebraic is disabled */
2017 if (mode_is_float(mode) && (get_irg_fp_model(current_ir_graph) & fp_strict_algebraic))
2020 /* Beware of Sub(P, P) which cannot be optimized into a simple Minus ... */
2021 if (mode_is_num(mode) && mode == get_irn_mode(a) && (classify_Const(a) == CNST_NULL)) {
2023 get_irn_dbg_info(n),
2028 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_0_A);
2030 /* do NOT execute this code if reassociation is enabled, it does the inverse! */
2031 else if (get_opt_reassociation() && get_irn_op(a) == op_Mul) {
2032 ir_node *ma = get_Mul_left(a);
2033 ir_node *mb = get_Mul_right(a);
2036 ir_node *blk = get_irn_n(n, -1);
2038 get_irn_dbg_info(n),
2039 current_ir_graph, blk,
2042 get_irn_dbg_info(n),
2043 current_ir_graph, blk,
2045 new_r_Const_long(current_ir_graph, blk, mode, 1),
2048 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2049 } else if (mb == b) {
2050 ir_node *blk = get_irn_n(n, -1);
2052 get_irn_dbg_info(n),
2053 current_ir_graph, blk,
2056 get_irn_dbg_info(n),
2057 current_ir_graph, blk,
2059 new_r_Const_long(current_ir_graph, blk, mode, 1),
2062 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_SUB_MUL_A_X_A);
2064 } else if (get_irn_op(a) == op_Sub) {
2065 ir_node *x = get_Sub_left(a);
2066 ir_node *y = get_Sub_right(a);
2067 ir_node *blk = get_irn_n(n, -1);
2068 ir_mode *m_b = get_irn_mode(b);
2069 ir_mode *m_y = get_irn_mode(y);
2072 /* Determine the right mode for the Add. */
2075 else if (mode_is_reference(m_b))
2077 else if (mode_is_reference(m_y))
2081 * Both modes are different but none is reference,
2082 * happens for instance in SubP(SubP(P, Iu), Is).
2083 * We have two possibilities here: Cast or ignore.
2084 * Currently we ignore this case.
2089 add = new_r_Add(current_ir_graph, blk, y, b, mode);
2092 set_Sub_right(n, add);
2093 DBG_OPT_ALGSIM0(n, n, FS_OPT_SUB_SUB_X_Y_Z);
2097 } /* transform_node_Sub */
2100 * Transform Mul(a,-1) into -a.
2101 * Do constant evaluation of Phi nodes.
2102 * Do architecture dependent optimizations on Mul nodes
2104 static ir_node *transform_node_Mul(ir_node *n) {
2105 ir_node *c, *oldn = n;
2106 ir_node *a = get_Mul_left(n);
2107 ir_node *b = get_Mul_right(n);
2110 HANDLE_BINOP_PHI(tarval_mul, a,b,c);
2112 mode = get_irn_mode(n);
2113 if (mode_is_signed(mode)) {
2116 if (value_of(a) == get_mode_minus_one(mode))
2118 else if (value_of(b) == get_mode_minus_one(mode))
2121 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1), r, mode);
2122 DBG_OPT_ALGSIM1(oldn, a, b, n, FS_OPT_MUL_MINUS_1);
2126 return arch_dep_replace_mul_with_shifts(n);
2127 } /* transform_node_Mul */
2130 * Transform a Div Node.
2132 static ir_node *transform_node_Div(ir_node *n) {
2133 tarval *tv = value_of(n);
2136 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2138 if (tv != tarval_bad) {
2139 value = new_Const(get_tarval_mode(tv), tv);
2141 DBG_OPT_CSTEVAL(n, value);
2142 } else /* Try architecture dependent optimization */
2143 value = arch_dep_replace_div_by_const(n);
2146 /* Turn Div into a tuple (mem, bad, value) */
2147 ir_node *mem = get_Div_mem(n);
2148 ir_node *blk = get_nodes_block(n);
2150 turn_into_tuple(n, pn_Div_max);
2151 set_Tuple_pred(n, pn_Div_M, mem);
2152 set_Tuple_pred(n, pn_Div_X_regular, new_r_Jmp(current_ir_graph, blk));
2153 set_Tuple_pred(n, pn_Div_X_except, new_Bad());
2154 set_Tuple_pred(n, pn_Div_res, value);
2157 } /* transform_node_Div */
2160 * Transform a Mod node.
2162 static ir_node *transform_node_Mod(ir_node *n) {
2163 tarval *tv = value_of(n);
2166 /* BEWARE: it is NOT possible to optimize a%a to 0, as this may cause a exception */
2168 if (tv != tarval_bad) {
2169 value = new_Const(get_tarval_mode(tv), tv);
2171 DBG_OPT_CSTEVAL(n, value);
2172 } else /* Try architecture dependent optimization */
2173 value = arch_dep_replace_mod_by_const(n);
2176 /* Turn Mod into a tuple (mem, bad, value) */
2177 ir_node *mem = get_Mod_mem(n);
2178 ir_node *blk = get_nodes_block(n);
2180 turn_into_tuple(n, pn_Mod_max);
2181 set_Tuple_pred(n, pn_Mod_M, mem);
2182 set_Tuple_pred(n, pn_Mod_X_regular, new_r_Jmp(current_ir_graph, blk));
2183 set_Tuple_pred(n, pn_Mod_X_except, new_Bad());
2184 set_Tuple_pred(n, pn_Mod_res, value);
2187 } /* transform_node_Mod */
2190 * Transform a DivMod node.
2192 static ir_node *transform_node_DivMod(ir_node *n) {
2195 ir_node *a = get_DivMod_left(n);
2196 ir_node *b = get_DivMod_right(n);
2197 ir_mode *mode = get_irn_mode(a);
2198 tarval *ta = value_of(a);
2199 tarval *tb = value_of(b);
2201 if (!(mode_is_int(mode) && mode_is_int(get_irn_mode(b))))
2204 /* BEWARE: it is NOT possible to optimize a/a to 1, as this may cause a exception */
2206 if (tb != tarval_bad) {
2207 if (tb == get_mode_one(get_tarval_mode(tb))) {
2208 b = new_Const (mode, get_mode_null(mode));
2211 DBG_OPT_CSTEVAL(n, b);
2212 } else if (ta != tarval_bad) {
2213 tarval *resa, *resb;
2214 resa = tarval_div (ta, tb);
2215 if (resa == tarval_bad) return n; /* Causes exception!!! Model by replacing through
2216 Jmp for X result!? */
2217 resb = tarval_mod (ta, tb);
2218 if (resb == tarval_bad) return n; /* Causes exception! */
2219 a = new_Const (mode, resa);
2220 b = new_Const (mode, resb);
2223 DBG_OPT_CSTEVAL(n, a);
2224 DBG_OPT_CSTEVAL(n, b);
2225 } else { /* Try architecture dependent optimization */
2226 arch_dep_replace_divmod_by_const(&a, &b, n);
2227 evaluated = a != NULL;
2229 } else if (ta == get_mode_null(mode)) {
2230 /* 0 / non-Const = 0 */
2235 if (evaluated) { /* replace by tuple */
2236 ir_node *mem = get_DivMod_mem(n);
2237 ir_node *blk = get_nodes_block(n);
2238 turn_into_tuple(n, pn_DivMod_max);
2239 set_Tuple_pred(n, pn_DivMod_M, mem);
2240 set_Tuple_pred(n, pn_DivMod_X_regular, new_r_Jmp(current_ir_graph, blk));
2241 set_Tuple_pred(n, pn_DivMod_X_except, new_Bad()); /* no exception */
2242 set_Tuple_pred(n, pn_DivMod_res_div, a);
2243 set_Tuple_pred(n, pn_DivMod_res_mod, b);
2247 } /* transform_node_DivMod */
2250 * Optimize Abs(x) into x if x is Confirmed >= 0
2251 * Optimize Abs(x) into -x if x is Confirmed <= 0
2253 static ir_node *transform_node_Abs(ir_node *n) {
2255 ir_node *a = get_Abs_op(n);
2256 value_classify_sign sign = classify_value_sign(a);
2258 if (sign == value_classified_negative) {
2259 ir_mode *mode = get_irn_mode(n);
2262 * We can replace the Abs by -x here.
2263 * We even could add a new Confirm here.
2265 * Note that -x would create a new node, so we could
2266 * not run it in the equivalent_node() context.
2268 n = new_rd_Minus(get_irn_dbg_info(n), current_ir_graph,
2269 get_irn_n(n, -1), a, mode);
2271 DBG_OPT_CONFIRM(oldn, n);
2272 } else if (sign == value_classified_positive) {
2273 /* n is positive, Abs is not needed */
2276 DBG_OPT_CONFIRM(oldn, n);
2280 } /* transform_node_Abs */
2283 * Transform a Cond node.
2285 * Replace the Cond by a Jmp if it branches on a constant
2288 static ir_node *transform_node_Cond(ir_node *n) {
2291 ir_node *a = get_Cond_selector(n);
2292 tarval *ta = value_of(a);
2294 /* we need block info which is not available in floating irgs */
2295 if (get_irg_pinned(current_ir_graph) == op_pin_state_floats)
2298 if ((ta != tarval_bad) &&
2299 (get_irn_mode(a) == mode_b) &&
2300 (get_opt_unreachable_code())) {
2301 /* It's a boolean Cond, branching on a boolean constant.
2302 Replace it by a tuple (Bad, Jmp) or (Jmp, Bad) */
2303 jmp = new_r_Jmp(current_ir_graph, get_nodes_block(n));
2304 turn_into_tuple(n, pn_Cond_max);
2305 if (ta == tarval_b_true) {
2306 set_Tuple_pred(n, pn_Cond_false, new_Bad());
2307 set_Tuple_pred(n, pn_Cond_true, jmp);
2309 set_Tuple_pred(n, pn_Cond_false, jmp);
2310 set_Tuple_pred(n, pn_Cond_true, new_Bad());
2312 /* We might generate an endless loop, so keep it alive. */
2313 add_End_keepalive(get_irg_end(current_ir_graph), get_nodes_block(n));
2316 } /* transform_node_Cond */
2321 static ir_node *transform_node_And(ir_node *n) {
2322 ir_node *c, *oldn = n;
2323 ir_node *a = get_And_left(n);
2324 ir_node *b = get_And_right(n);
2326 HANDLE_BINOP_PHI(tarval_and, a,b,c);
2328 } /* transform_node_And */
2333 static ir_node *transform_node_Eor(ir_node *n) {
2334 ir_node *c, *oldn = n;
2335 ir_node *a = get_Eor_left(n);
2336 ir_node *b = get_Eor_right(n);
2337 ir_mode *mode = get_irn_mode(n);
2339 HANDLE_BINOP_PHI(tarval_eor, a,b,c);
2343 n = new_rd_Const(get_irn_dbg_info(n), current_ir_graph, get_irn_n(n, -1),
2344 mode, get_mode_null(mode));
2345 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_A_A);
2346 } else if ((mode == mode_b)
2347 && (get_irn_op(a) == op_Proj)
2348 && (get_irn_mode(a) == mode_b)
2349 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)
2350 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2351 /* The Eor negates a Cmp. The Cmp has the negated result anyways! */
2352 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2353 mode_b, get_negated_pnc(get_Proj_proj(a), mode));
2355 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT_BOOL);
2356 } else if ((mode == mode_b)
2357 && (classify_tarval (value_of(b)) == TV_CLASSIFY_ONE)) {
2358 /* The Eor is a Not. Replace it by a Not. */
2359 /* ????!!!Extend to bitfield 1111111. */
2360 n = new_r_Not(current_ir_graph, get_irn_n(n, -1), a, mode_b);
2362 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_EOR_TO_NOT);
2366 } /* transform_node_Eor */
2371 static ir_node *transform_node_Not(ir_node *n) {
2372 ir_node *c, *oldn = n;
2373 ir_node *a = get_Not_op(n);
2375 HANDLE_UNOP_PHI(tarval_not,a,c);
2377 /* check for a boolean Not */
2378 if ( (get_irn_mode(n) == mode_b)
2379 && (get_irn_op(a) == op_Proj)
2380 && (get_irn_mode(a) == mode_b)
2381 && (get_irn_op(get_Proj_pred(a)) == op_Cmp)) {
2382 /* We negate a Cmp. The Cmp has the negated result anyways! */
2383 n = new_r_Proj(current_ir_graph, get_irn_n(n, -1), get_Proj_pred(a),
2384 mode_b, get_negated_pnc(get_Proj_proj(a), mode_b));
2385 DBG_OPT_ALGSIM0(oldn, n, FS_OPT_NOT_CMP);
2388 } /* transform_node_Not */
2391 * Transform a Minus.
2393 static ir_node *transform_node_Minus(ir_node *n) {
2394 ir_node *c, *oldn = n;
2395 ir_node *a = get_Minus_op(n);
2397 HANDLE_UNOP_PHI(tarval_neg,a,c);
2399 } /* transform_node_Minus */
2402 * Transform a Cast_type(Const) into a new Const_type
2404 static ir_node *transform_node_Cast(ir_node *n) {
2406 ir_node *pred = get_Cast_op(n);
2407 ir_type *tp = get_irn_type(n);
2409 if (get_irn_op(pred) == op_Const && get_Const_type(pred) != tp) {
2410 n = new_rd_Const_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_irn_mode(pred),
2411 get_Const_tarval(pred), tp);
2412 DBG_OPT_CSTEVAL(oldn, n);
2413 } else if ((get_irn_op(pred) == op_SymConst) && (get_SymConst_value_type(pred) != tp)) {
2414 n = new_rd_SymConst_type(NULL, current_ir_graph, get_irn_n(pred, -1), get_SymConst_symbol(pred),
2415 get_SymConst_kind(pred), tp);
2416 DBG_OPT_CSTEVAL(oldn, n);
2420 } /* transform_node_Cast */
2423 * Transform a Proj(Div) with a non-zero value.
2424 * Removes the exceptions and routes the memory to the NoMem node.
2426 static ir_node *transform_node_Proj_Div(ir_node *proj) {
2427 ir_node *n = get_Proj_pred(proj);
2428 ir_node *b = get_Div_right(n);
2432 if (value_not_zero(b, &confirm)) {
2433 /* div(x, y) && y != 0 */
2434 proj_nr = get_Proj_proj(proj);
2435 if (proj_nr == pn_Div_X_except) {
2436 /* we found an exception handler, remove it */
2437 DBG_OPT_EXC_REM(proj);
2439 } else if (proj_nr == pn_Div_M) {
2440 ir_node *res = get_Div_mem(n);
2441 ir_node *new_mem = get_irg_no_mem(current_ir_graph);
2444 /* This node can only float up to the Confirm block */
2445 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2447 set_irn_pinned(n, op_pin_state_floats);
2448 /* this is a Div without exception, we can remove the memory edge */
2449 set_Div_mem(n, new_mem);
2454 } /* transform_node_Proj_Div */
2457 * Transform a Proj(Mod) with a non-zero value.
2458 * Removes the exceptions and routes the memory to the NoMem node.
2460 static ir_node *transform_node_Proj_Mod(ir_node *proj) {
2461 ir_node *n = get_Proj_pred(proj);
2462 ir_node *b = get_Mod_right(n);
2466 if (value_not_zero(b, &confirm)) {
2467 /* mod(x, y) && y != 0 */
2468 proj_nr = get_Proj_proj(proj);
2470 if (proj_nr == pn_Mod_X_except) {
2471 /* we found an exception handler, remove it */
2472 DBG_OPT_EXC_REM(proj);
2474 } else if (proj_nr == pn_Mod_M) {
2475 ir_node *res = get_Mod_mem(n);
2476 ir_node *new_mem = get_irg_no_mem(current_ir_graph);
2479 /* This node can only float up to the Confirm block */
2480 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2482 set_irn_pinned(n, op_pin_state_floats);
2483 /* this is a Mod without exception, we can remove the memory edge */
2484 set_Mod_mem(n, get_irg_no_mem(current_ir_graph));
2486 } else if (proj_nr == pn_Mod_res && get_Mod_left(n) == b) {
2487 /* a % a = 0 if a != 0 */
2488 ir_mode *mode = get_irn_mode(proj);
2489 ir_node *res = new_Const(mode, get_mode_null(mode));
2491 DBG_OPT_CSTEVAL(n, res);
2496 } /* transform_node_Proj_Mod */
2499 * Transform a Proj(DivMod) with a non-zero value.
2500 * Removes the exceptions and routes the memory to the NoMem node.
2502 static ir_node *transform_node_Proj_DivMod(ir_node *proj) {
2503 ir_node *n = get_Proj_pred(proj);
2504 ir_node *b = get_DivMod_right(n);
2508 if (value_not_zero(b, &confirm)) {
2509 /* DivMod(x, y) && y != 0 */
2510 proj_nr = get_Proj_proj(proj);
2512 if (proj_nr == pn_DivMod_X_except) {
2513 /* we found an exception handler, remove it */
2514 DBG_OPT_EXC_REM(proj);
2516 } else if (proj_nr == pn_DivMod_M) {
2517 ir_node *res = get_DivMod_mem(n);
2518 ir_node *new_mem = get_irg_no_mem(current_ir_graph);
2521 /* This node can only float up to the Confirm block */
2522 new_mem = new_r_Pin(current_ir_graph, get_nodes_block(confirm), new_mem);
2524 set_irn_pinned(n, op_pin_state_floats);
2525 /* this is a DivMod without exception, we can remove the memory edge */
2526 set_DivMod_mem(n, get_irg_no_mem(current_ir_graph));
2528 } else if (proj_nr == pn_DivMod_res_mod && get_DivMod_left(n) == b) {
2529 /* a % a = 0 if a != 0 */
2530 ir_mode *mode = get_irn_mode(proj);
2531 ir_node *res = new_Const(mode, get_mode_null(mode));
2533 DBG_OPT_CSTEVAL(n, res);
2538 } /* transform_node_Proj_DivMod */
2541 * Optimizes jump tables (CondIs or CondIu) by removing all impossible cases.
2543 static ir_node *transform_node_Proj_Cond(ir_node *proj) {
2544 if (get_opt_unreachable_code()) {
2545 ir_node *n = get_Proj_pred(proj);
2546 ir_node *b = get_Cond_selector(n);
2548 if (mode_is_int(get_irn_mode(b))) {
2549 tarval *tb = value_of(b);
2551 if (tb != tarval_bad) {
2552 /* we have a constant switch */
2553 long num = get_Proj_proj(proj);
2555 if (num != get_Cond_defaultProj(n)) { /* we cannot optimize default Proj's yet */
2556 if (get_tarval_long(tb) == num) {
2557 /* Do NOT create a jump here, or we will have 2 control flow ops
2558 * in a block. This case is optimized away in optimize_cf(). */
2561 /* this case will NEVER be taken, kill it */
2569 } /* transform_node_Proj_Cond */
2572 * Normalizes and optimizes Cmp nodes.
2574 static ir_node *transform_node_Proj_Cmp(ir_node *proj) {
2575 if (get_opt_reassociation()) {
2576 ir_node *n = get_Proj_pred(proj);
2577 ir_node *left = get_Cmp_left(n);
2578 ir_node *right = get_Cmp_right(n);
2582 ir_mode *mode = NULL;
2583 long proj_nr = get_Proj_proj(proj);
2586 * First step: normalize the compare op
2587 * by placing the constant on the right site
2588 * or moving the lower address node to the left.
2589 * We ignore the case that both are constants
2590 * this case should be optimized away.
2592 if (get_irn_op(right) == op_Const) {
2594 } else if (get_irn_op(left) == op_Const) {
2599 proj_nr = get_inversed_pnc(proj_nr);
2601 } else if (get_irn_idx(left) > get_irn_idx(right)) {
2607 proj_nr = get_inversed_pnc(proj_nr);
2612 * Second step: Try to reduce the magnitude
2613 * of a constant. This may help to generate better code
2614 * later and may help to normalize more compares.
2615 * Of course this is only possible for integer values.
2618 mode = get_irn_mode(c);
2619 tv = get_Const_tarval(c);
2621 if (tv != tarval_bad) {
2622 /* the following optimization is possible on modes without Overflow
2623 * on Unary Minus or on == and !=:
2624 * -a CMP c ==> a swap(CMP) -c
2626 * Beware: for two-complement Overflow may occur, so only == and != can
2627 * be optimized, see this:
2628 * -MININT < 0 =/=> MININT > 0 !!!
2630 if (get_opt_constant_folding() && get_irn_op(left) == op_Minus &&
2631 (!mode_overflow_on_unary_Minus(mode) ||
2632 (mode_is_int(mode) && (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg)))) {
2633 left = get_Minus_op(left);
2634 tv = tarval_sub(get_mode_null(mode), tv);
2636 proj_nr = get_inversed_pnc(proj_nr);
2640 /* for integer modes, we have more */
2641 if (mode_is_int(mode)) {
2642 /* Ne includes Unordered which is not possible on integers.
2643 * However, frontends often use this wrong, so fix it here */
2644 if (proj_nr & pn_Cmp_Uo) {
2645 proj_nr &= ~pn_Cmp_Uo;
2646 set_Proj_proj(proj, proj_nr);
2649 /* c > 0 : a < c ==> a <= (c-1) a >= c ==> a > (c-1) */
2650 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Ge) &&
2651 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Gt) {
2652 tv = tarval_sub(tv, get_mode_one(mode));
2654 proj_nr ^= pn_Cmp_Eq;
2657 /* c < 0 : a > c ==> a >= (c+1) a <= c ==> a < (c+1) */
2658 else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Le) &&
2659 tarval_cmp(tv, get_mode_null(mode)) == pn_Cmp_Lt) {
2660 tv = tarval_add(tv, get_mode_one(mode));
2662 proj_nr ^= pn_Cmp_Eq;
2666 /* the following reassociations work only for == and != */
2667 if (proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) {
2669 /* a-b == 0 ==> a == b, a-b != 0 ==> a != b */
2670 if (classify_tarval(tv) == TV_CLASSIFY_NULL && get_irn_op(left) == op_Sub) {
2671 right = get_Sub_right(left);
2672 left = get_Sub_left(left);
2674 tv = value_of(right);
2678 if (tv != tarval_bad) {
2679 ir_op *op = get_irn_op(left);
2681 /* a-c1 == c2 ==> a == c2+c1, a-c1 != c2 ==> a != c2+c1 */
2683 ir_node *c1 = get_Sub_right(left);
2684 tarval *tv2 = value_of(c1);
2686 if (tv2 != tarval_bad) {
2687 tv2 = tarval_add(tv, value_of(c1));
2689 if (tv2 != tarval_bad) {
2690 left = get_Sub_left(left);
2696 /* a+c1 == c2 ==> a == c2-c1, a+c1 != c2 ==> a != c2-c1 */
2697 else if (op == op_Add) {
2698 ir_node *a_l = get_Add_left(left);
2699 ir_node *a_r = get_Add_right(left);
2703 if (get_irn_op(a_l) == op_Const) {
2705 tv2 = value_of(a_l);
2708 tv2 = value_of(a_r);
2711 if (tv2 != tarval_bad) {
2712 tv2 = tarval_sub(tv, tv2);
2714 if (tv2 != tarval_bad) {
2721 /* -a == c ==> a == -c, -a != c ==> a != -c */
2722 else if (op == op_Minus) {
2723 tarval *tv2 = tarval_sub(get_mode_null(mode), tv);
2725 if (tv2 != tarval_bad) {
2726 left = get_Minus_op(left);
2733 /* the following reassociations work only for <= */
2734 else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
2735 if (tv != tarval_bad) {
2736 ir_op *op = get_irn_op(left);
2738 /* c >= 0 : Abs(a) <= c ==> (unsigned)(a + c) <= 2*c */
2746 * optimization for AND:
2748 * And(x, C) == C ==> And(x, C) != 0
2749 * And(x, C) != C ==> And(X, C) == 0
2751 * if C is a single Bit constant.
2753 if ((proj_nr == pn_Cmp_Eq || proj_nr == pn_Cmp_Lg) &&
2754 (get_irn_op(left) == op_And)) {
2755 if (is_single_bit_tarval(tv)) {
2756 /* check for Constant's match. We have check hare the tarvals,
2757 because our const might be changed */
2758 ir_node *la = get_And_left(left);
2759 ir_node *ra = get_And_right(left);
2760 if ((is_Const(la) && get_Const_tarval(la) == tv) ||
2761 (is_Const(ra) && get_Const_tarval(ra) == tv)) {
2762 /* fine: do the transformation */
2763 tv = get_mode_null(get_tarval_mode(tv));
2764 proj_nr ^= pn_Cmp_Leg;
2769 } /* tarval != bad */
2773 ir_node *block = get_irn_n(n, -1); /* Beware of get_nodes_Block() */
2775 if (changed & 2) /* need a new Const */
2776 right = new_Const(mode, tv);
2778 /* create a new compare */
2779 n = new_rd_Cmp(get_irn_dbg_info(n), current_ir_graph, block,
2782 set_Proj_pred(proj, n);
2783 set_Proj_proj(proj, proj_nr);
2787 } /* transform_node_Proj_Cmp */
2790 * Does all optimizations on nodes that must be done on it's Proj's
2791 * because of creating new nodes.
2793 static ir_node *transform_node_Proj(ir_node *proj) {
2794 ir_node *n = get_Proj_pred(proj);
2796 switch (get_irn_opcode(n)) {
2798 return transform_node_Proj_Div(proj);
2801 return transform_node_Proj_Mod(proj);
2804 return transform_node_Proj_DivMod(proj);
2807 return transform_node_Proj_Cond(proj);
2810 return transform_node_Proj_Cmp(proj);
2813 /* should not happen, but if it does will be optimized away */
2814 return equivalent_node_Proj(proj);
2820 } /* transform_node_Proj */
2823 * Move Confirms down through Phi nodes.
2825 static ir_node *transform_node_Phi(ir_node *phi) {
2827 ir_mode *mode = get_irn_mode(phi);
2829 if (mode_is_reference(mode)) {
2830 n = get_irn_arity(phi);
2832 /* Beware of Phi0 */
2834 ir_node *pred = get_irn_n(phi, 0);
2835 ir_node *bound, *new_Phi, *block, **in;
2838 if (! is_Confirm(pred))
2841 bound = get_Confirm_bound(pred);
2842 pnc = get_Confirm_cmp(pred);
2844 NEW_ARR_A(ir_node *, in, n);
2845 in[0] = get_Confirm_value(pred);
2847 for (i = 1; i < n; ++i) {
2848 pred = get_irn_n(phi, i);
2850 if (! is_Confirm(pred) ||
2851 get_Confirm_bound(pred) != bound ||
2852 get_Confirm_cmp(pred) != pnc)
2854 in[i] = get_Confirm_value(pred);
2856 /* move the Confirm nodes "behind" the Phi */
2857 block = get_irn_n(phi, -1);
2858 new_Phi = new_r_Phi(current_ir_graph, block, n, in, get_irn_mode(phi));
2859 return new_r_Confirm(current_ir_graph, block, new_Phi, bound, pnc);
2863 } /* transform_node_Phi */
2866 * Returns the operands of a commutative bin-op, if one operand is
2867 * a const, it is returned as the second one.
2869 static void get_comm_Binop_Ops(ir_node *binop, ir_node **a, ir_node **c) {
2870 ir_node *op_a = get_binop_left(binop);
2871 ir_node *op_b = get_binop_right(binop);
2873 assert(is_op_commutative(get_irn_op(binop)));
2875 if (get_irn_op(op_a) == op_Const) {
2882 } /* get_comm_Binop_Ops */
2885 * Optimize a Or(And(Or(And(v,c4),c3),c2),c1) pattern if possible.
2886 * Such pattern may arise in bitfield stores.
2888 * value c4 value c4 & c2
2889 * AND c3 AND c1 | c3
2894 static ir_node *transform_node_Or_bf_store(ir_node *or) {
2897 ir_node *and_l, *c3;
2898 ir_node *value, *c4;
2899 ir_node *new_and, *new_const, *block;
2900 ir_mode *mode = get_irn_mode(or);
2902 tarval *tv1, *tv2, *tv3, *tv4, *tv, *n_tv4, *n_tv2;
2904 get_comm_Binop_Ops(or, &and, &c1);
2905 if ((get_irn_op(c1) != op_Const) || (get_irn_op(and) != op_And))
2908 get_comm_Binop_Ops(and, &or_l, &c2);
2909 if ((get_irn_op(c2) != op_Const) || (get_irn_op(or_l) != op_Or))
2912 get_comm_Binop_Ops(or_l, &and_l, &c3);
2913 if ((get_irn_op(c3) != op_Const) || (get_irn_op(and_l) != op_And))
2916 get_comm_Binop_Ops(and_l, &value, &c4);
2917 if (get_irn_op(c4) != op_Const)
2920 /* ok, found the pattern, check for conditions */
2921 assert(mode == get_irn_mode(and));
2922 assert(mode == get_irn_mode(or_l));
2923 assert(mode == get_irn_mode(and_l));
2925 tv1 = get_Const_tarval(c1);
2926 tv2 = get_Const_tarval(c2);
2927 tv3 = get_Const_tarval(c3);
2928 tv4 = get_Const_tarval(c4);
2930 tv = tarval_or(tv4, tv2);
2931 if (classify_tarval(tv) != TV_CLASSIFY_ALL_ONE) {
2932 /* have at least one 0 at the same bit position */
2936 n_tv4 = tarval_not(tv4);
2937 if (tv3 != tarval_and(tv3, n_tv4)) {
2938 /* bit in the or_mask is outside the and_mask */
2942 n_tv2 = tarval_not(tv2);
2943 if (tv1 != tarval_and(tv1, n_tv2)) {
2944 /* bit in the or_mask is outside the and_mask */
2948 /* ok, all conditions met */
2949 block = get_irn_n(or, -1);
2951 new_and = new_r_And(current_ir_graph, block,
2952 value, new_r_Const(current_ir_graph, block, mode, tarval_and(tv4, tv2)), mode);
2954 new_const = new_r_Const(current_ir_graph, block, mode, tarval_or(tv3, tv1));
2956 set_Or_left(or, new_and);
2957 set_Or_right(or, new_const);
2959 /* check for more */
2960 return transform_node_Or_bf_store(or);
2961 } /* transform_node_Or_bf_store */
2964 * Optimize an Or(shl(x, c), shr(x, bits - c)) into a Rot
2966 static ir_node *transform_node_Or_Rot(ir_node *or) {
2967 ir_mode *mode = get_irn_mode(or);
2968 ir_node *shl, *shr, *block;
2969 ir_node *irn, *x, *c1, *c2, *v, *sub, *n;
2972 if (! mode_is_int(mode))
2975 shl = get_binop_left(or);
2976 shr = get_binop_right(or);
2978 if (get_irn_op(shl) == op_Shr) {
2979 if (get_irn_op(shr) != op_Shl)
2985 } else if (get_irn_op(shl) != op_Shl) {
2987 } else if (get_irn_op(shr) != op_Shr) {
2990 x = get_Shl_left(shl);
2991 if (x != get_Shr_left(shr))
2994 c1 = get_Shl_right(shl);
2995 c2 = get_Shr_right(shr);
2996 if (get_irn_op(c1) == op_Const && get_irn_op(c2) == op_Const) {
2997 tv1 = get_Const_tarval(c1);
2998 if (! tarval_is_long(tv1))
3001 tv2 = get_Const_tarval(c2);
3002 if (! tarval_is_long(tv2))
3005 if (get_tarval_long(tv1) + get_tarval_long(tv2)
3006 != get_mode_size_bits(mode))
3009 /* yet, condition met */
3010 block = get_irn_n(or, -1);
3012 n = new_r_Rot(current_ir_graph, block, x, c1, mode);
3014 DBG_OPT_ALGSIM1(or, shl, shr, n, FS_OPT_OR_SHFT_TO_ROT);
3016 } else if (get_irn_op(c1) == op_Sub) {
3020 if (get_Sub_right(sub) != v)
3023 c1 = get_Sub_left(sub);
3024 if (get_irn_op(c1) != op_Const)
3027 tv1 = get_Const_tarval(c1);
3028 if (! tarval_is_long(tv1))
3031 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3034 /* yet, condition met */
3035 block = get_nodes_block(or);
3037 /* a Rot right is not supported, so use a rot left */
3038 n = new_r_Rot(current_ir_graph, block, x, sub, mode);
3040 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3042 } else if (get_irn_op(c2) == op_Sub) {
3046 c1 = get_Sub_left(sub);
3047 if (get_irn_op(c1) != op_Const)
3050 tv1 = get_Const_tarval(c1);
3051 if (! tarval_is_long(tv1))
3054 if (get_tarval_long(tv1) != get_mode_size_bits(mode))
3057 /* yet, condition met */
3058 block = get_irn_n(or, -1);
3061 n = new_r_Rot(current_ir_graph, block, x, v, mode);
3063 DBG_OPT_ALGSIM0(or, n, FS_OPT_OR_SHFT_TO_ROT);
3068 } /* transform_node_Or_Rot */
3073 static ir_node *transform_node_Or(ir_node *n) {
3074 ir_node *c, *oldn = n;
3075 ir_node *a = get_Or_left(n);
3076 ir_node *b = get_Or_right(n);
3078 HANDLE_BINOP_PHI(tarval_or, a,b,c);
3080 n = transform_node_Or_bf_store(n);
3081 n = transform_node_Or_Rot(n);
3084 } /* transform_node_Or */
3088 static ir_node *transform_node(ir_node *n);
3091 * Optimize (a >> c1) >> c2), works for Shr, Shrs, Shl.
3093 * Should be moved to reassociation?
3095 static ir_node *transform_node_shift(ir_node *n) {
3096 ir_node *left, *right;
3097 tarval *tv1, *tv2, *res;
3099 int modulo_shf, flag;
3101 left = get_binop_left(n);
3103 /* different operations */
3104 if (get_irn_op(left) != get_irn_op(n))
3107 right = get_binop_right(n);
3108 tv1 = value_of(right);
3109 if (tv1 == tarval_bad)
3112 tv2 = value_of(get_binop_right(left));
3113 if (tv2 == tarval_bad)
3116 res = tarval_add(tv1, tv2);
3118 /* beware: a simple replacement works only, if res < modulo shift */
3119 mode = get_irn_mode(n);
3123 modulo_shf = get_mode_modulo_shift(mode);
3124 if (modulo_shf > 0) {
3125 tarval *modulo = new_tarval_from_long(modulo_shf, get_tarval_mode(res));
3127 if (tarval_cmp(res, modulo) & pn_Cmp_Lt)
3133 /* ok, we can replace it */
3134 ir_node *in[2], *irn, *block = get_irn_n(n, -1);
3136 in[0] = get_binop_left(left);
3137 in[1] = new_r_Const(current_ir_graph, block, get_tarval_mode(res), res);
3139 irn = new_ir_node(NULL, current_ir_graph, block, get_irn_op(n), mode, 2, in);
3141 DBG_OPT_ALGSIM0(n, irn, FS_OPT_REASSOC_SHIFT);
3143 return transform_node(irn);
3146 } /* transform_node_shift */
3151 static ir_node *transform_node_Shr(ir_node *n) {
3152 ir_node *c, *oldn = n;
3153 ir_node *a = get_Shr_left(n);
3154 ir_node *b = get_Shr_right(n);
3156 HANDLE_BINOP_PHI(tarval_shr, a, b, c);
3157 return transform_node_shift(n);
3158 } /* transform_node_Shr */
3163 static ir_node *transform_node_Shrs(ir_node *n) {
3164 ir_node *c, *oldn = n;
3165 ir_node *a = get_Shrs_left(n);
3166 ir_node *b = get_Shrs_right(n);
3168 HANDLE_BINOP_PHI(tarval_shrs, a, b, c);
3169 return transform_node_shift(n);
3170 } /* transform_node_Shrs */
3175 static ir_node *transform_node_Shl(ir_node *n) {
3176 ir_node *c, *oldn = n;
3177 ir_node *a = get_Shl_left(n);
3178 ir_node *b = get_Shl_right(n);
3180 HANDLE_BINOP_PHI(tarval_shl, a, b, c);
3181 return transform_node_shift(n);
3182 } /* transform_node_Shl */
3185 * Remove dead blocks and nodes in dead blocks
3186 * in keep alive list. We do not generate a new End node.
3188 static ir_node *transform_node_End(ir_node *n) {
3189 int i, j, n_keepalives = get_End_n_keepalives(n);
3192 NEW_ARR_A(ir_node *, in, n_keepalives);
3194 for (i = j = 0; i < n_keepalives; ++i) {
3195 ir_node *ka = get_End_keepalive(n, i);
3197 if (! is_Block_dead(ka)) {
3201 } else if (is_irn_pinned_in_irg(ka) && is_Block_dead(get_nodes_block(ka))) {
3204 /* FIXME: beabi need to keep a Proj(M) */
3205 if (is_Phi(ka) || is_irn_keep(ka) || is_Proj(ka))
3208 if (j != n_keepalives)
3209 set_End_keepalives(n, j, in);
3211 } /* transform_node_End */
3214 * Optimize a Mux into some simpler cases.
3216 static ir_node *transform_node_Mux(ir_node *n) {
3217 ir_node *oldn = n, *sel = get_Mux_sel(n);
3218 ir_mode *mode = get_irn_mode(n);
3220 if (get_irn_op(sel) == op_Proj && !mode_honor_signed_zeros(mode)) {
3221 ir_node *cmp = get_Proj_pred(sel);
3222 long proj_nr = get_Proj_proj(sel);
3223 ir_node *f = get_Mux_false(n);
3224 ir_node *t = get_Mux_true(n);
3226 if (get_irn_op(cmp) == op_Cmp && classify_Const(get_Cmp_right(cmp)) == CNST_NULL) {
3227 ir_node *block = get_irn_n(n, -1);
3230 * Note: normalization puts the constant on the right site,
3231 * so we check only one case.
3233 * Note further that these optimization work even for floating point
3234 * with NaN's because -NaN == NaN.
3235 * However, if +0 and -0 is handled differently, we cannot use the first one.
3237 if (get_irn_op(f) == op_Minus &&
3238 get_Minus_op(f) == t &&
3239 get_Cmp_left(cmp) == t) {
3241 if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3242 /* Mux(a >=/> 0, -a, a) ==> Abs(a) */
3243 n = new_rd_Abs(get_irn_dbg_info(n),
3247 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3249 } else if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3250 /* Mux(a <=/< 0, -a, a) ==> Minus(Abs(a)) */
3251 n = new_rd_Abs(get_irn_dbg_info(n),
3255 n = new_rd_Minus(get_irn_dbg_info(n),
3260 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3263 } else if (get_irn_op(t) == op_Minus &&
3264 get_Minus_op(t) == f &&
3265 get_Cmp_left(cmp) == f) {
3267 if (proj_nr == pn_Cmp_Le || proj_nr == pn_Cmp_Lt) {
3268 /* Mux(a <=/< 0, a, -a) ==> Abs(a) */
3269 n = new_rd_Abs(get_irn_dbg_info(n),
3273 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3275 } else if (proj_nr == pn_Cmp_Ge || proj_nr == pn_Cmp_Gt) {
3276 /* Mux(a >=/> 0, a, -a) ==> Minus(Abs(a)) */
3277 n = new_rd_Abs(get_irn_dbg_info(n),
3281 n = new_rd_Minus(get_irn_dbg_info(n),
3286 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_ABS);
3291 if (mode_is_int(mode) && mode_is_signed(mode) &&
3292 get_mode_arithmetic(mode) == irma_twos_complement) {
3293 ir_node *x = get_Cmp_left(cmp);
3295 /* the following optimization works only with signed integer two-complement mode */
3297 if (mode == get_irn_mode(x)) {
3299 * FIXME: this restriction is two rigid, as it would still
3300 * work if mode(x) = Hs and mode == Is, but at least it removes
3303 if ((proj_nr == pn_Cmp_Lt || proj_nr == pn_Cmp_Le) &&
3304 classify_Const(t) == CNST_ALL_ONE &&
3305 classify_Const(f) == CNST_NULL) {
3307 * Mux(x:T </<= 0, 0, -1) -> Shrs(x, sizeof_bits(T) - 1)
3311 n = new_rd_Shrs(get_irn_dbg_info(n),
3312 current_ir_graph, block, x,
3313 new_r_Const_long(current_ir_graph, block, mode_Iu,
3314 get_mode_size_bits(mode) - 1),
3316 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3318 } else if ((proj_nr == pn_Cmp_Gt || proj_nr == pn_Cmp_Ge) &&
3319 classify_Const(t) == CNST_ONE &&
3320 classify_Const(f) == CNST_NULL) {
3322 * Mux(x:T >/>= 0, 0, 1) -> Shr(-x, sizeof_bits(T) - 1)
3326 n = new_rd_Shr(get_irn_dbg_info(n),
3327 current_ir_graph, block,
3328 new_r_Minus(current_ir_graph, block, x, mode),
3329 new_r_Const_long(current_ir_graph, block, mode_Iu,
3330 get_mode_size_bits(mode) - 1),
3332 DBG_OPT_ALGSIM1(oldn, cmp, sel, n, FS_OPT_MUX_TO_SHR);
3339 return arch_transform_node_Mux(n);
3340 } /* transform_node_Mux */
3343 * Optimize a Psi into some simpler cases.
3345 static ir_node *transform_node_Psi(ir_node *n) {
3347 return transform_node_Mux(n);
3350 } /* transform_node_Psi */
3353 * Tries several [inplace] [optimizing] transformations and returns an
3354 * equivalent node. The difference to equivalent_node() is that these
3355 * transformations _do_ generate new nodes, and thus the old node must
3356 * not be freed even if the equivalent node isn't the old one.
3358 static ir_node *transform_node(ir_node *n) {
3359 if (n->op->ops.transform_node)
3360 n = n->op->ops.transform_node(n);
3362 } /* transform_node */
3365 * Sets the default transform node operation for an ir_op_ops.
3367 * @param code the opcode for the default operation
3368 * @param ops the operations initialized
3373 static ir_op_ops *firm_set_default_transform_node(ir_opcode code, ir_op_ops *ops)
3377 ops->transform_node = transform_node_##a; \
3410 } /* firm_set_default_transform_node */
3413 /* **************** Common Subexpression Elimination **************** */
3415 /** The size of the hash table used, should estimate the number of nodes
3417 #define N_IR_NODES 512
3419 /** Compares the attributes of two Const nodes. */
3420 static int node_cmp_attr_Const(ir_node *a, ir_node *b) {
3421 return (get_Const_tarval(a) != get_Const_tarval(b))
3422 || (get_Const_type(a) != get_Const_type(b));
3423 } /* node_cmp_attr_Const */
3425 /** Compares the attributes of two Proj nodes. */
3426 static int node_cmp_attr_Proj(ir_node *a, ir_node *b) {
3427 return get_irn_proj_attr (a) != get_irn_proj_attr (b);
3428 } /* node_cmp_attr_Proj */
3430 /** Compares the attributes of two Filter nodes. */
3431 static int node_cmp_attr_Filter(ir_node *a, ir_node *b) {
3432 return get_Filter_proj(a) != get_Filter_proj(b);
3433 } /* node_cmp_attr_Filter */
3435 /** Compares the attributes of two Alloc nodes. */
3436 static int node_cmp_attr_Alloc(ir_node *a, ir_node *b) {
3437 return (get_irn_alloc_attr(a).where != get_irn_alloc_attr(b).where)
3438 || (get_irn_alloc_attr(a).type != get_irn_alloc_attr(b).type);
3439 } /* node_cmp_attr_Alloc */
3441 /** Compares the attributes of two Free nodes. */
3442 static int node_cmp_attr_Free(ir_node *a, ir_node *b) {
3443 return (get_irn_free_attr(a).where != get_irn_free_attr(b).where)
3444 || (get_irn_free_attr(a).type != get_irn_free_attr(b).type);
3445 } /* node_cmp_attr_Free */
3447 /** Compares the attributes of two SymConst nodes. */
3448 static int node_cmp_attr_SymConst(ir_node *a, ir_node *b) {
3449 return (get_irn_symconst_attr(a).num != get_irn_symconst_attr(b).num)
3450 || (get_irn_symconst_attr(a).sym.type_p != get_irn_symconst_attr(b).sym.type_p)
3451 || (get_irn_symconst_attr(a).tp != get_irn_symconst_attr(b).tp);
3452 } /* node_cmp_attr_SymConst */
3454 /** Compares the attributes of two Call nodes. */
3455 static int node_cmp_attr_Call(ir_node *a, ir_node *b) {
3456 return (get_irn_call_attr(a) != get_irn_call_attr(b));
3457 } /* node_cmp_attr_Call */
3459 /** Compares the attributes of two Sel nodes. */
3460 static int node_cmp_attr_Sel(ir_node *a, ir_node *b) {
3461 return (get_irn_sel_attr(a).ent->kind != get_irn_sel_attr(b).ent->kind)
3462 || (get_irn_sel_attr(a).ent->name != get_irn_sel_attr(b).ent->name)
3463 || (get_irn_sel_attr(a).ent->owner != get_irn_sel_attr(b).ent->owner)
3464 || (get_irn_sel_attr(a).ent->ld_name != get_irn_sel_attr(b).ent->ld_name)
3465 || (get_irn_sel_attr(a).ent->type != get_irn_sel_attr(b).ent->type);
3466 } /* node_cmp_attr_Sel */
3468 /** Compares the attributes of two Phi nodes. */
3469 static int node_cmp_attr_Phi(ir_node *a, ir_node *b) {
3470 /* we can only enter this function if both nodes have the same number of inputs,
3471 hence it is enough to check if one of them is a Phi0 */
3473 /* check the Phi0 attribute */
3474 return get_irn_phi0_attr(a) != get_irn_phi0_attr(b);
3477 } /* node_cmp_attr_Phi */
3479 /** Compares the attributes of two Conv nodes. */
3480 static int node_cmp_attr_Conv(ir_node *a, ir_node *b) {
3481 return get_Conv_strict(a) != get_Conv_strict(b);
3482 } /* node_cmp_attr_Conv */
3484 /** Compares the attributes of two Cast nodes. */
3485 static int node_cmp_attr_Cast(ir_node *a, ir_node *b) {
3486 return get_Cast_type(a) != get_Cast_type(b);
3487 } /* node_cmp_attr_Cast */
3489 /** Compares the attributes of two Load nodes. */
3490 static int node_cmp_attr_Load(ir_node *a, ir_node *b) {
3491 if (get_Load_volatility(a) == volatility_is_volatile ||
3492 get_Load_volatility(b) == volatility_is_volatile)
3493 /* NEVER do CSE on volatile Loads */
3496 return get_Load_mode(a) != get_Load_mode(b);
3497 } /* node_cmp_attr_Load */
3499 /** Compares the attributes of two Store nodes. */
3500 static int node_cmp_attr_Store(ir_node *a, ir_node *b) {
3501 /* NEVER do CSE on volatile Stores */
3502 return (get_Store_volatility(a) == volatility_is_volatile ||
3503 get_Store_volatility(b) == volatility_is_volatile);
3504 } /* node_cmp_attr_Store */
3506 /** Compares the attributes of two Confirm nodes. */
3507 static int node_cmp_attr_Confirm(ir_node *a, ir_node *b) {
3508 return (get_Confirm_cmp(a) != get_Confirm_cmp(b));
3509 } /* node_cmp_attr_Confirm */
3511 /** Compares the attributes of two ASM nodes. */
3512 static int node_cmp_attr_ASM(ir_node *a, ir_node *b) {
3514 const ir_asm_constraint *ca;
3515 const ir_asm_constraint *cb;
3518 if (get_ASM_text(a) != get_ASM_text(b));
3521 /* Should we really check the constraints here? Should be better, but is strange. */
3522 n = get_ASM_n_input_constraints(a);
3523 if (n != get_ASM_n_input_constraints(b))
3526 ca = get_ASM_input_constraints(a);
3527 cb = get_ASM_input_constraints(b);
3528 for (i = 0; i < n; ++i) {
3529 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
3533 n = get_ASM_n_output_constraints(a);
3534 if (n != get_ASM_n_output_constraints(b))
3537 ca = get_ASM_output_constraints(a);
3538 cb = get_ASM_output_constraints(b);
3539 for (i = 0; i < n; ++i) {
3540 if (ca[i].pos != cb[i].pos || ca[i].constraint != cb[i].constraint)
3544 n = get_ASM_n_clobbers(a);
3545 if (n != get_ASM_n_clobbers(b))
3548 cla = get_ASM_clobbers(a);
3549 clb = get_ASM_clobbers(b);
3550 for (i = 0; i < n; ++i) {
3551 if (cla[i] != clb[i])
3555 } /* node_cmp_attr_ASM */
3558 * Set the default node attribute compare operation for an ir_op_ops.
3560 * @param code the opcode for the default operation
3561 * @param ops the operations initialized
3566 static ir_op_ops *firm_set_default_node_cmp_attr(ir_opcode code, ir_op_ops *ops)
3570 ops->node_cmp_attr = node_cmp_attr_##a; \
3595 } /* firm_set_default_node_cmp_attr */
3598 * Compare function for two nodes in the hash table. Gets two
3599 * nodes as parameters. Returns 0 if the nodes are a cse.
3601 int identities_cmp(const void *elt, const void *key) {
3608 if (a == b) return 0;
3610 if ((get_irn_op(a) != get_irn_op(b)) ||
3611 (get_irn_mode(a) != get_irn_mode(b))) return 1;
3613 /* compare if a's in and b's in are of equal length */
3614 irn_arity_a = get_irn_intra_arity (a);
3615 if (irn_arity_a != get_irn_intra_arity(b))
3618 /* for block-local cse and op_pin_state_pinned nodes: */
3619 if (!get_opt_global_cse() || (get_irn_pinned(a) == op_pin_state_pinned)) {
3620 if (get_irn_intra_n(a, -1) != get_irn_intra_n(b, -1))
3624 /* compare a->in[0..ins] with b->in[0..ins] */
3625 for (i = 0; i < irn_arity_a; i++)
3626 if (get_irn_intra_n(a, i) != get_irn_intra_n(b, i))
3630 * here, we already now that the nodes are identical except their
3633 if (a->op->ops.node_cmp_attr)
3634 return a->op->ops.node_cmp_attr(a, b);
3637 } /* identities_cmp */
3640 * Calculate a hash value of a node.
3642 unsigned ir_node_hash(ir_node *node) {
3646 if (node->op == op_Const) {
3647 /* special value for const, as they only differ in their tarval. */
3648 h = HASH_PTR(node->attr.con.tv);
3649 h = 9*h + HASH_PTR(get_irn_mode(node));
3650 } else if (node->op == op_SymConst) {
3651 /* special value for const, as they only differ in their symbol. */
3652 h = HASH_PTR(node->attr.symc.sym.type_p);
3653 h = 9*h + HASH_PTR(get_irn_mode(node));
3656 /* hash table value = 9*(9*(9*(9*(9*arity+in[0])+in[1])+ ...)+mode)+code */
3657 h = irn_arity = get_irn_intra_arity(node);
3659 /* consider all in nodes... except the block if not a control flow. */
3660 for (i = is_cfop(node) ? -1 : 0; i < irn_arity; i++) {
3661 h = 9*h + HASH_PTR(get_irn_intra_n(node, i));
3665 h = 9*h + HASH_PTR(get_irn_mode(node));
3667 h = 9*h + HASH_PTR(get_irn_op(node));
3671 } /* ir_node_hash */
3673 pset *new_identities(void) {
3674 return new_pset(identities_cmp, N_IR_NODES);
3675 } /* new_identities */
3677 void del_identities(pset *value_table) {
3678 del_pset(value_table);
3679 } /* del_identities */
3682 * Return the canonical node computing the same value as n.
3684 * @param value_table The value table
3685 * @param n The node to lookup
3687 * Looks up the node in a hash table.
3689 * For Const nodes this is performed in the constructor, too. Const
3690 * nodes are extremely time critical because of their frequent use in
3691 * constant string arrays.
3693 static INLINE ir_node *identify(pset *value_table, ir_node *n) {
3696 if (!value_table) return n;
3698 if (get_opt_reassociation()) {
3699 if (is_op_commutative(get_irn_op(n))) {
3700 ir_node *l = get_binop_left(n);
3701 ir_node *r = get_binop_right(n);
3703 /* for commutative operators perform a OP b == b OP a */
3704 if (get_irn_idx(l) > get_irn_idx(r)) {
3705 set_binop_left(n, r);
3706 set_binop_right(n, l);
3711 o = pset_find(value_table, n, ir_node_hash(n));
3720 * During construction we set the op_pin_state_pinned flag in the graph right when the
3721 * optimization is performed. The flag turning on procedure global cse could
3722 * be changed between two allocations. This way we are safe.
3724 static INLINE ir_node *identify_cons(pset *value_table, ir_node *n) {
3727 n = identify(value_table, n);
3728 if (get_irn_n(old, -1) != get_irn_n(n, -1))
3729 set_irg_pinned(current_ir_graph, op_pin_state_floats);
3731 } /* identify_cons */
3734 * Return the canonical node computing the same value as n.
3735 * Looks up the node in a hash table, enters it in the table
3736 * if it isn't there yet.
3738 ir_node *identify_remember(pset *value_table, ir_node *n) {
3741 if (!value_table) return n;
3743 if (get_opt_reassociation()) {
3744 if (is_op_commutative(get_irn_op(n))) {
3745 ir_node *l = get_binop_left(n);
3746 ir_node *r = get_binop_right(n);
3747 int l_idx = get_irn_idx(l);
3748 int r_idx = get_irn_idx(r);
3750 /* For commutative operators perform a OP b == b OP a but keep
3751 constants on the RIGHT side. This helps greatly in some optimizations.
3752 Moreover we use the idx number to make the form deterministic. */
3753 if (is_irn_constlike(l))
3755 if (is_irn_constlike(r))
3757 if (l_idx < r_idx) {
3758 set_binop_left(n, r);
3759 set_binop_right(n, l);
3764 /* lookup or insert in hash table with given hash key. */
3765 o = pset_insert(value_table, n, ir_node_hash(n));
3772 } /* identify_remember */
3774 /* Add a node to the identities value table. */
3775 void add_identities(pset *value_table, ir_node *node) {
3776 if (get_opt_cse() && is_no_Block(node))
3777 identify_remember(value_table, node);
3778 } /* add_identities */
3780 /* Visit each node in the value table of a graph. */
3781 void visit_all_identities(ir_graph *irg, irg_walk_func visit, void *env) {
3783 ir_graph *rem = current_ir_graph;
3785 current_ir_graph = irg;
3786 foreach_pset(irg->value_table, node)
3788 current_ir_graph = rem;
3789 } /* visit_all_identities */
3792 * Garbage in, garbage out. If a node has a dead input, i.e., the
3793 * Bad node is input to the node, return the Bad node.
3795 static INLINE ir_node *gigo(ir_node *node) {
3797 ir_op *op = get_irn_op(node);
3799 /* remove garbage blocks by looking at control flow that leaves the block
3800 and replacing the control flow by Bad. */
3801 if (get_irn_mode(node) == mode_X) {
3802 ir_node *block = get_nodes_block(skip_Proj(node));
3804 /* Don't optimize nodes in immature blocks. */
3805 if (!get_Block_matured(block)) return node;
3806 /* Don't optimize End, may have Bads. */
3807 if (op == op_End) return node;
3809 if (is_Block(block)) {
3810 irn_arity = get_irn_arity(block);
3811 for (i = 0; i < irn_arity; i++) {
3812 if (!is_Bad(get_irn_n(block, i)))
3815 if (i == irn_arity) {
3816 ir_graph *irg = get_irn_irg(block);
3817 /* the start block is never dead */
3818 if (block != get_irg_start_block(irg)
3819 && block != get_irg_end_block(irg))
3825 /* Blocks, Phis and Tuples may have dead inputs, e.g., if one of the
3826 blocks predecessors is dead. */
3827 if (op != op_Block && op != op_Phi && op != op_Tuple) {
3828 irn_arity = get_irn_arity(node);
3831 * Beware: we can only read the block of a non-floating node.
3833 if (is_irn_pinned_in_irg(node) &&
3834 is_Block_dead(get_nodes_block(node)))
3837 for (i = 0; i < irn_arity; i++) {
3838 ir_node *pred = get_irn_n(node, i);
3843 /* Propagating Unknowns here seems to be a bad idea, because
3844 sometimes we need a node as a input and did not want that
3846 However, it might be useful to move this into a later phase
3847 (if you think that optimizing such code is useful). */
3848 if (is_Unknown(pred) && mode_is_data(get_irn_mode(node)))
3849 return new_Unknown(get_irn_mode(node));
3854 /* With this code we violate the agreement that local_optimize
3855 only leaves Bads in Block, Phi and Tuple nodes. */
3856 /* If Block has only Bads as predecessors it's garbage. */
3857 /* If Phi has only Bads as predecessors it's garbage. */
3858 if ((op == op_Block && get_Block_matured(node)) || op == op_Phi) {
3859 irn_arity = get_irn_arity(node);
3860 for (i = 0; i < irn_arity; i++) {
3861 if (!is_Bad(get_irn_n(node, i))) break;
3863 if (i == irn_arity) node = new_Bad();
3870 * These optimizations deallocate nodes from the obstack.
3871 * It can only be called if it is guaranteed that no other nodes
3872 * reference this one, i.e., right after construction of a node.
3874 * @param n The node to optimize
3876 * current_ir_graph must be set to the graph of the node!
3878 ir_node *optimize_node(ir_node *n) {
3881 ir_opcode iro = get_irn_opcode(n);
3883 /* Always optimize Phi nodes: part of the construction. */
3884 if ((!get_opt_optimize()) && (iro != iro_Phi)) return n;
3886 /* constant expression evaluation / constant folding */
3887 if (get_opt_constant_folding()) {
3888 /* neither constants nor Tuple values can be evaluated */
3889 if (iro != iro_Const && (get_irn_mode(n) != mode_T)) {
3890 /* try to evaluate */
3891 tv = computed_value(n);
3892 if (tv != tarval_bad) {
3894 ir_type *old_tp = get_irn_type(n);
3895 int i, arity = get_irn_arity(n);
3899 * Try to recover the type of the new expression.
3901 for (i = 0; i < arity && !old_tp; ++i)
3902 old_tp = get_irn_type(get_irn_n(n, i));
3905 * we MUST copy the node here temporary, because it's still needed
3906 * for DBG_OPT_CSTEVAL
3908 node_size = offsetof(ir_node, attr) + n->op->attr_size;
3909 oldn = alloca(node_size);
3911 memcpy(oldn, n, node_size);
3912 CLONE_ARR_A(ir_node *, oldn->in, n->in);
3914 /* ARG, copy the in array, we need it for statistics */
3915 memcpy(oldn->in, n->in, ARR_LEN(n->in) * sizeof(n->in[0]));
3917 /* note the inplace edges module */
3918 edges_node_deleted(n, current_ir_graph);
3920 /* evaluation was successful -- replace the node. */
3921 irg_kill_node(current_ir_graph, n);
3922 nw = new_Const(get_tarval_mode (tv), tv);
3924 if (old_tp && get_type_mode(old_tp) == get_tarval_mode (tv))
3925 set_Const_type(nw, old_tp);
3926 DBG_OPT_CSTEVAL(oldn, nw);
3932 /* remove unnecessary nodes */
3933 if (get_opt_constant_folding() ||
3934 (iro == iro_Phi) || /* always optimize these nodes. */
3936 (iro == iro_Proj) ||
3937 (iro == iro_Block) ) /* Flags tested local. */
3938 n = equivalent_node(n);
3940 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
3942 /* Common Subexpression Elimination.
3944 * Checks whether n is already available.
3945 * The block input is used to distinguish different subexpressions. Right
3946 * now all nodes are op_pin_state_pinned to blocks, i.e., the CSE only finds common
3947 * subexpressions within a block.
3950 n = identify_cons(current_ir_graph->value_table, n);
3953 edges_node_deleted(oldn, current_ir_graph);
3955 /* We found an existing, better node, so we can deallocate the old node. */
3956 irg_kill_node(current_ir_graph, oldn);
3960 /* Some more constant expression evaluation that does not allow to
3962 iro = get_irn_opcode(n);
3963 if (get_opt_constant_folding() ||
3964 (iro == iro_Cond) ||
3965 (iro == iro_Proj)) /* Flags tested local. */
3966 n = transform_node(n);
3968 /* Remove nodes with dead (Bad) input.
3969 Run always for transformation induced Bads. */
3972 /* Now we have a legal, useful node. Enter it in hash table for CSE */
3973 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block)) {
3974 n = identify_remember(current_ir_graph->value_table, n);
3978 } /* optimize_node */
3982 * These optimizations never deallocate nodes (in place). This can cause dead
3983 * nodes lying on the obstack. Remove these by a dead node elimination,
3984 * i.e., a copying garbage collection.
3986 ir_node *optimize_in_place_2(ir_node *n) {
3989 ir_opcode iro = get_irn_opcode(n);
3991 if (!get_opt_optimize() && (get_irn_op(n) != op_Phi)) return n;
3993 /* constant expression evaluation / constant folding */
3994 if (get_opt_constant_folding()) {
3995 /* neither constants nor Tuple values can be evaluated */
3996 if (iro != iro_Const && get_irn_mode(n) != mode_T) {
3997 /* try to evaluate */
3998 tv = computed_value(n);
3999 if (tv != tarval_bad) {
4000 /* evaluation was successful -- replace the node. */
4001 ir_type *old_tp = get_irn_type(n);
4002 int i, arity = get_irn_arity(n);
4005 * Try to recover the type of the new expression.
4007 for (i = 0; i < arity && !old_tp; ++i)
4008 old_tp = get_irn_type(get_irn_n(n, i));
4010 n = new_Const(get_tarval_mode(tv), tv);
4012 if (old_tp && get_type_mode(old_tp) == get_tarval_mode(tv))
4013 set_Const_type(n, old_tp);
4015 DBG_OPT_CSTEVAL(oldn, n);
4021 /* remove unnecessary nodes */
4022 if (get_opt_constant_folding() ||
4023 (iro == iro_Phi) || /* always optimize these nodes. */
4024 (iro == iro_Id) || /* ... */
4025 (iro == iro_Proj) || /* ... */
4026 (iro == iro_Block) ) /* Flags tested local. */
4027 n = equivalent_node(n);
4029 optimize_preds(n); /* do node specific optimizations of nodes predecessors. */
4031 /** common subexpression elimination **/
4032 /* Checks whether n is already available. */
4033 /* The block input is used to distinguish different subexpressions. Right
4034 now all nodes are op_pin_state_pinned to blocks, i.e., the cse only finds common
4035 subexpressions within a block. */
4036 if (get_opt_cse()) {
4037 n = identify(current_ir_graph->value_table, n);
4040 /* Some more constant expression evaluation. */
4041 iro = get_irn_opcode(n);
4042 if (get_opt_constant_folding() ||
4043 (iro == iro_Cond) ||
4044 (iro == iro_Proj)) /* Flags tested local. */
4045 n = transform_node(n);
4047 /* Remove nodes with dead (Bad) input.
4048 Run always for transformation induced Bads. */
4051 /* Now we can verify the node, as it has no dead inputs any more. */
4054 /* Now we have a legal, useful node. Enter it in hash table for cse.
4055 Blocks should be unique anyways. (Except the successor of start:
4056 is cse with the start block!) */
4057 if (get_opt_cse() && (get_irn_opcode(n) != iro_Block))
4058 n = identify_remember(current_ir_graph->value_table, n);
4061 } /* optimize_in_place_2 */
4064 * Wrapper for external use, set proper status bits after optimization.
4066 ir_node *optimize_in_place(ir_node *n) {
4067 /* Handle graph state */
4068 assert(get_irg_phase_state(current_ir_graph) != phase_building);
4070 if (get_opt_global_cse())
4071 set_irg_pinned(current_ir_graph, op_pin_state_floats);
4072 if (get_irg_outs_state(current_ir_graph) == outs_consistent)
4073 set_irg_outs_inconsistent(current_ir_graph);
4075 /* FIXME: Maybe we could also test whether optimizing the node can
4076 change the control graph. */
4077 set_irg_doms_inconsistent(current_ir_graph);
4078 return optimize_in_place_2(n);
4079 } /* optimize_in_place */
4082 * Sets the default operation for an ir_ops.
4084 ir_op_ops *firm_set_default_operations(ir_opcode code, ir_op_ops *ops) {
4085 ops = firm_set_default_computed_value(code, ops);
4086 ops = firm_set_default_equivalent_node(code, ops);
4087 ops = firm_set_default_transform_node(code, ops);
4088 ops = firm_set_default_node_cmp_attr(code, ops);
4089 ops = firm_set_default_get_type(code, ops);
4090 ops = firm_set_default_get_type_attr(code, ops);
4091 ops = firm_set_default_get_entity_attr(code, ops);
4094 } /* firm_set_default_operations */